From eb84bb298d2b95aec7b2ae12cbf25ac64f25379a Mon Sep 17 00:00:00 2001 From: tylermurphy534 Date: Sun, 6 Nov 2022 15:12:42 -0500 Subject: move to self host --- .../Assets/Editor/x64/Bakery/lmTexAreaLightDir.ptx | 1928 ++++++++++++++++++++ 1 file changed, 1928 insertions(+) create mode 100644 VRCSDK3Worlds/Assets/Editor/x64/Bakery/lmTexAreaLightDir.ptx (limited to 'VRCSDK3Worlds/Assets/Editor/x64/Bakery/lmTexAreaLightDir.ptx') diff --git a/VRCSDK3Worlds/Assets/Editor/x64/Bakery/lmTexAreaLightDir.ptx b/VRCSDK3Worlds/Assets/Editor/x64/Bakery/lmTexAreaLightDir.ptx new file mode 100644 index 00000000..ce9e88ac --- /dev/null +++ b/VRCSDK3Worlds/Assets/Editor/x64/Bakery/lmTexAreaLightDir.ptx @@ -0,0 +1,1928 @@ +// +// 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_Mask[1]; +.global .align 1 .b8 image_Dir[1]; +.global .align 1 .b8 uvpos[1]; +.global .align 1 .b8 uvnormal[1]; +.global .align 1 .b8 lightMeshBuffer[1]; +.global .align 4 .u32 lightMeshBufferSize; +.global .align 4 .f32 lightInvCutoff; +.global .align 4 .f32 lightPointSize; +.global .align 4 .b8 lightColor[12]; +.global .align 1 .b8 rnd_seeds[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_typeinfo19lightMeshBufferSizeE[8] = {82, 97, 121, 0, 4, 0, 0, 0}; +.global .align 4 .b8 _ZN21rti_internal_typeinfo14lightInvCutoffE[8] = {82, 97, 121, 0, 4, 0, 0, 0}; +.global .align 4 .b8 _ZN21rti_internal_typeinfo14lightPointSizeE[8] = {82, 97, 121, 0, 4, 0, 0, 0}; +.global .align 4 .b8 _ZN21rti_internal_typeinfo10lightColorE[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 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 16 .b8 _ZN21rti_internal_typename19lightMeshBufferSizeE[13] = {117, 110, 115, 105, 103, 110, 101, 100, 32, 105, 110, 116, 0}; +.global .align 8 .b8 _ZN21rti_internal_typename14lightInvCutoffE[6] = {102, 108, 111, 97, 116, 0}; +.global .align 8 .b8 _ZN21rti_internal_typename14lightPointSizeE[6] = {102, 108, 111, 97, 116, 0}; +.global .align 8 .b8 _ZN21rti_internal_typename10lightColorE[7] = {102, 108, 111, 97, 116, 51, 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_typeenum19lightMeshBufferSizeE = 4919; +.global .align 4 .u32 _ZN21rti_internal_typeenum14lightInvCutoffE = 4919; +.global .align 4 .u32 _ZN21rti_internal_typeenum14lightPointSizeE = 4919; +.global .align 4 .u32 _ZN21rti_internal_typeenum10lightColorE = 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 1 .b8 _ZN21rti_internal_semantic19lightMeshBufferSizeE[1]; +.global .align 1 .b8 _ZN21rti_internal_semantic14lightInvCutoffE[1]; +.global .align 1 .b8 _ZN21rti_internal_semantic14lightPointSizeE[1]; +.global .align 1 .b8 _ZN21rti_internal_semantic10lightColorE[1]; +.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_annotation19lightMeshBufferSizeE[1]; +.global .align 1 .b8 _ZN23rti_internal_annotation14lightInvCutoffE[1]; +.global .align 1 .b8 _ZN23rti_internal_annotation14lightPointSizeE[1]; +.global .align 1 .b8 _ZN23rti_internal_annotation10lightColorE[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[44]; + .reg .b64 %SP; + .reg .b64 %SPL; + .reg .pred %p<145>; + .reg .b16 %rs<55>; + .reg .f32 %f<992>; + .reg .b32 %r<375>; + .reg .b64 %rd<171>; + + + mov.u64 %rd170, __local_depot0; + cvta.local.u64 %SP, %rd170; + ld.global.u32 %r1, [samples]; + ld.global.v2.u32 {%r103, %r104}, [pixelID]; + cvt.u64.u32 %rd23, %r103; + cvt.u64.u32 %rd24, %r104; + mov.u64 %rd27, uvnormal; + cvta.global.u64 %rd22, %rd27; + mov.u32 %r101, 2; + mov.u32 %r102, 4; + mov.u64 %rd26, 0; + // inline asm + call (%rd21), _rt_buffer_get_64, (%rd22, %r101, %r102, %rd23, %rd24, %rd26, %rd26); + // inline asm + ld.u32 %r2, [%rd21]; + shr.u32 %r107, %r2, 16; + cvt.u16.u32 %rs1, %r107; + and.b16 %rs3, %rs1, 255; + cvt.u16.u32 %rs4, %r2; + or.b16 %rs5, %rs4, %rs3; + setp.eq.s16 %p7, %rs5, 0; + mov.f32 %f909, 0f00000000; + mov.f32 %f910, %f909; + mov.f32 %f911, %f909; + @%p7 bra BB0_2; + + ld.u8 %rs6, [%rd21+1]; + and.b16 %rs8, %rs4, 255; + cvt.rn.f32.u16 %f224, %rs8; + div.rn.f32 %f225, %f224, 0f437F0000; + fma.rn.f32 %f226, %f225, 0f40000000, 0fBF800000; + cvt.rn.f32.u16 %f227, %rs6; + div.rn.f32 %f228, %f227, 0f437F0000; + fma.rn.f32 %f229, %f228, 0f40000000, 0fBF800000; + cvt.rn.f32.u16 %f230, %rs3; + div.rn.f32 %f231, %f230, 0f437F0000; + fma.rn.f32 %f232, %f231, 0f40000000, 0fBF800000; + mul.f32 %f233, %f229, %f229; + fma.rn.f32 %f234, %f226, %f226, %f233; + fma.rn.f32 %f235, %f232, %f232, %f234; + sqrt.rn.f32 %f236, %f235; + rcp.rn.f32 %f237, %f236; + mul.f32 %f909, %f226, %f237; + mul.f32 %f910, %f229, %f237; + mul.f32 %f911, %f232, %f237; + +BB0_2: + ld.global.v2.u32 {%r108, %r109}, [pixelID]; + ld.global.v2.u32 {%r111, %r112}, [tileInfo]; + add.s32 %r3, %r108, %r111; + add.s32 %r4, %r109, %r112; + setp.eq.f32 %p8, %f910, 0f00000000; + setp.eq.f32 %p9, %f909, 0f00000000; + and.pred %p10, %p9, %p8; + setp.eq.f32 %p11, %f911, 0f00000000; + and.pred %p12, %p10, %p11; + @%p12 bra BB0_131; + bra.uni BB0_3; + +BB0_131: + ld.global.u32 %r374, [imageEnabled]; + and.b32 %r316, %r374, 1; + setp.eq.b32 %p139, %r316, 1; + @!%p139 bra BB0_133; + bra.uni BB0_132; + +BB0_132: + cvt.u64.u32 %rd123, %r3; + cvt.u64.u32 %rd124, %r4; + mov.u64 %rd127, image; + cvta.global.u64 %rd122, %rd127; + mov.u64 %rd126, 0; + // inline asm + call (%rd121), _rt_buffer_get_64, (%rd122, %r101, %r102, %rd123, %rd124, %rd126, %rd126); + // inline asm + mov.u16 %rs36, 0; + st.v4.u8 [%rd121], {%rs36, %rs36, %rs36, %rs36}; + ld.global.u32 %r374, [imageEnabled]; + +BB0_133: + and.b32 %r319, %r374, 8; + setp.eq.s32 %p140, %r319, 0; + @%p140 bra BB0_135; + + cvt.u64.u32 %rd130, %r3; + cvt.u64.u32 %rd131, %r4; + mov.u64 %rd134, image_Mask; + cvta.global.u64 %rd129, %rd134; + mov.u64 %rd133, 0; + // inline asm + call (%rd128), _rt_buffer_get_64, (%rd129, %r101, %r101, %rd130, %rd131, %rd133, %rd133); + // inline asm + mov.f32 %f892, 0f00000000; + cvt.rzi.u32.f32 %r322, %f892; + cvt.u16.u32 %rs37, %r322; + mov.u16 %rs38, 0; + st.v2.u8 [%rd128], {%rs37, %rs38}; + ld.global.u32 %r374, [imageEnabled]; + +BB0_135: + and.b32 %r323, %r374, 4; + setp.eq.s32 %p141, %r323, 0; + @%p141 bra BB0_139; + + ld.global.u32 %r324, [additive]; + setp.eq.s32 %p142, %r324, 0; + cvt.u64.u32 %rd19, %r3; + cvt.u64.u32 %rd20, %r4; + @%p142 bra BB0_138; + + mov.u64 %rd147, image_HDR; + cvta.global.u64 %rd136, %rd147; + mov.u32 %r328, 8; + mov.u64 %rd146, 0; + // inline asm + call (%rd135), _rt_buffer_get_64, (%rd136, %r101, %r328, %rd19, %rd20, %rd146, %rd146); + // inline asm + ld.v4.u16 {%rs45, %rs46, %rs47, %rs48}, [%rd135]; + // inline asm + { cvt.f32.f16 %f893, %rs45;} + + // inline asm + // inline asm + { cvt.f32.f16 %f894, %rs46;} + + // inline asm + // inline asm + { cvt.f32.f16 %f895, %rs47;} + + // inline asm + // inline asm + call (%rd141), _rt_buffer_get_64, (%rd136, %r101, %r328, %rd19, %rd20, %rd146, %rd146); + // inline asm + add.f32 %f896, %f893, 0f00000000; + add.f32 %f897, %f894, 0f00000000; + add.f32 %f898, %f895, 0f00000000; + // inline asm + { cvt.rn.f16.f32 %rs44, %f898;} + + // inline asm + // inline asm + { cvt.rn.f16.f32 %rs43, %f897;} + + // inline asm + // inline asm + { cvt.rn.f16.f32 %rs42, %f896;} + + // inline asm + mov.u16 %rs49, 0; + st.v4.u16 [%rd141], {%rs42, %rs43, %rs44, %rs49}; + bra.uni BB0_139; + +BB0_3: + ld.global.v2.u32 {%r121, %r122}, [pixelID]; + cvt.u64.u32 %rd30, %r121; + cvt.u64.u32 %rd31, %r122; + mov.u64 %rd40, uvpos; + cvta.global.u64 %rd29, %rd40; + mov.u32 %r117, 12; + // inline asm + call (%rd28), _rt_buffer_get_64, (%rd29, %r101, %r117, %rd30, %rd31, %rd26, %rd26); + // inline asm + ld.f32 %f9, [%rd28+8]; + ld.f32 %f8, [%rd28+4]; + ld.f32 %f7, [%rd28]; + mul.f32 %f245, %f7, 0f3456BF95; + mul.f32 %f246, %f8, 0f3456BF95; + mul.f32 %f247, %f9, 0f3456BF95; + abs.f32 %f10, %f909; + div.rn.f32 %f248, %f245, %f10; + abs.f32 %f249, %f910; + div.rn.f32 %f250, %f246, %f249; + abs.f32 %f11, %f911; + div.rn.f32 %f251, %f247, %f11; + abs.f32 %f252, %f248; + abs.f32 %f253, %f250; + abs.f32 %f254, %f251; + mov.f32 %f255, 0f38D1B717; + max.f32 %f256, %f252, %f255; + max.f32 %f257, %f253, %f255; + max.f32 %f258, %f254, %f255; + fma.rn.f32 %f12, %f909, %f256, %f7; + fma.rn.f32 %f13, %f910, %f257, %f8; + fma.rn.f32 %f14, %f911, %f258, %f9; + ld.global.v2.u32 {%r125, %r126}, [pixelID]; + cvt.u64.u32 %rd36, %r125; + cvt.u64.u32 %rd37, %r126; + mov.u64 %rd41, rnd_seeds; + cvta.global.u64 %rd35, %rd41; + // inline asm + call (%rd34), _rt_buffer_get_64, (%rd35, %r101, %r102, %rd36, %rd37, %rd26, %rd26); + // inline asm + ld.u32 %r129, [%rd34]; + mad.lo.s32 %r5, %r129, 1664525, 1013904223; + ld.global.u32 %r130, [lightMeshBufferSize]; + setp.eq.s32 %p14, %r130, 0; + mov.pred %p13, 0; + mov.f32 %f20, 0f00000000; + mov.u32 %r7, 0; + @%p14 bra BB0_4; + + ld.global.f32 %f15, [lightPointSize]; + mul.f32 %f16, %f12, 0f3456BF95; + mul.f32 %f17, %f13, 0f3456BF95; + mul.f32 %f18, %f14, 0f3456BF95; + and.b32 %r133, %r5, 16777215; + cvt.rn.f32.u32 %f266, %r133; + mul.f32 %f267, %f266, 0fB3800000; + fma.rn.f32 %f19, %f267, 0f3F333333, 0f3F800000; + mov.f32 %f20, 0f00000000; + mov.u32 %r343, 0; + abs.f32 %f390, %f17; + abs.f32 %f391, %f16; + max.f32 %f392, %f391, %f390; + abs.f32 %f393, %f18; + max.f32 %f394, %f392, %f393; + mov.u32 %r7, %r343; + mov.f32 %f21, %f20; + mov.f32 %f22, %f20; + mov.f32 %f23, %f20; + mov.f32 %f940, %f20; + mov.f32 %f941, %f20; + mov.f32 %f942, %f20; + +BB0_6: + mul.lo.s32 %r8, %r343, 3; + cvt.s64.s32 %rd44, %r8; + mov.u64 %rd48, lightMeshBuffer; + cvta.global.u64 %rd43, %rd48; + mov.u32 %r134, 1; + // inline asm + call (%rd42), _rt_buffer_get_64, (%rd43, %r134, %r117, %rd44, %rd26, %rd26, %rd26); + // inline asm + ld.f32 %f268, [%rd42]; + sub.f32 %f269, %f268, %f7; + ld.f32 %f270, [%rd42+4]; + sub.f32 %f271, %f270, %f8; + ld.f32 %f272, [%rd42+8]; + sub.f32 %f273, %f272, %f9; + mul.f32 %f274, %f271, %f271; + fma.rn.f32 %f275, %f269, %f269, %f274; + fma.rn.f32 %f276, %f273, %f273, %f275; + sqrt.rn.f32 %f27, %f276; + rcp.rn.f32 %f277, %f27; + mul.f32 %f28, %f269, %f277; + mul.f32 %f29, %f271, %f277; + mul.f32 %f30, %f273, %f277; + mul.f32 %f278, %f910, %f29; + fma.rn.f32 %f279, %f909, %f28, %f278; + fma.rn.f32 %f31, %f911, %f30, %f279; + setp.leu.f32 %p15, %f31, 0f00000000; + @%p15 bra BB0_22; + + setp.ne.s32 %p17, %r1, 0; + mul.f32 %f280, %f27, %f27; + mul.f32 %f281, %f280, 0f40C90FDB; + div.rn.f32 %f282, %f15, %f281; + add.f32 %f32, %f282, %f282; + setp.gt.f32 %p18, %f32, %f19; + and.pred %p19, %p17, %p18; + mov.pred %p144, -1; + @%p19 bra BB0_24; + + ld.global.f32 %f285, [lightInvCutoff]; + mul.f32 %f33, %f27, %f285; + mov.f32 %f289, 0f40800000; + abs.f32 %f35, %f33; + setp.lt.f32 %p20, %f35, 0f00800000; + mul.f32 %f291, %f35, 0f4B800000; + selp.f32 %f292, 0fC3170000, 0fC2FE0000, %p20; + selp.f32 %f293, %f291, %f35, %p20; + mov.b32 %r136, %f293; + and.b32 %r137, %r136, 8388607; + or.b32 %r138, %r137, 1065353216; + mov.b32 %f294, %r138; + shr.u32 %r139, %r136, 23; + cvt.rn.f32.u32 %f295, %r139; + add.f32 %f296, %f292, %f295; + setp.gt.f32 %p21, %f294, 0f3FB504F3; + mul.f32 %f297, %f294, 0f3F000000; + add.f32 %f298, %f296, 0f3F800000; + selp.f32 %f299, %f297, %f294, %p21; + selp.f32 %f300, %f298, %f296, %p21; + add.f32 %f301, %f299, 0fBF800000; + add.f32 %f284, %f299, 0f3F800000; + // inline asm + rcp.approx.ftz.f32 %f283,%f284; + // inline asm + add.f32 %f302, %f301, %f301; + mul.f32 %f303, %f283, %f302; + mul.f32 %f304, %f303, %f303; + mov.f32 %f305, 0f3C4CAF63; + mov.f32 %f306, 0f3B18F0FE; + fma.rn.f32 %f307, %f306, %f304, %f305; + mov.f32 %f308, 0f3DAAAABD; + fma.rn.f32 %f309, %f307, %f304, %f308; + mul.rn.f32 %f310, %f309, %f304; + mul.rn.f32 %f311, %f310, %f303; + sub.f32 %f312, %f301, %f303; + neg.f32 %f313, %f303; + add.f32 %f314, %f312, %f312; + fma.rn.f32 %f315, %f313, %f301, %f314; + mul.rn.f32 %f316, %f283, %f315; + add.f32 %f317, %f311, %f303; + sub.f32 %f318, %f303, %f317; + add.f32 %f319, %f311, %f318; + add.f32 %f320, %f316, %f319; + add.f32 %f321, %f317, %f320; + sub.f32 %f322, %f317, %f321; + add.f32 %f323, %f320, %f322; + mov.f32 %f324, 0f3F317200; + mul.rn.f32 %f325, %f300, %f324; + mov.f32 %f326, 0f35BFBE8E; + mul.rn.f32 %f327, %f300, %f326; + add.f32 %f328, %f325, %f321; + sub.f32 %f329, %f325, %f328; + add.f32 %f330, %f321, %f329; + add.f32 %f331, %f323, %f330; + add.f32 %f332, %f327, %f331; + add.f32 %f333, %f328, %f332; + sub.f32 %f334, %f328, %f333; + add.f32 %f335, %f332, %f334; + mul.rn.f32 %f336, %f289, %f333; + neg.f32 %f337, %f336; + fma.rn.f32 %f338, %f289, %f333, %f337; + fma.rn.f32 %f339, %f289, %f335, %f338; + mov.f32 %f340, 0f00000000; + fma.rn.f32 %f341, %f340, %f333, %f339; + add.rn.f32 %f342, %f336, %f341; + neg.f32 %f343, %f342; + add.rn.f32 %f344, %f336, %f343; + add.rn.f32 %f345, %f344, %f341; + mov.b32 %r140, %f342; + setp.eq.s32 %p22, %r140, 1118925336; + add.s32 %r141, %r140, -1; + mov.b32 %f346, %r141; + add.f32 %f347, %f345, 0f37000000; + selp.f32 %f348, %f346, %f342, %p22; + selp.f32 %f36, %f347, %f345, %p22; + mul.f32 %f349, %f348, 0f3FB8AA3B; + cvt.rzi.f32.f32 %f350, %f349; + mov.f32 %f351, 0fBF317200; + fma.rn.f32 %f352, %f350, %f351, %f348; + mov.f32 %f353, 0fB5BFBE8E; + fma.rn.f32 %f354, %f350, %f353, %f352; + mul.f32 %f355, %f354, 0f3FB8AA3B; + ex2.approx.ftz.f32 %f356, %f355; + add.f32 %f357, %f350, 0f00000000; + ex2.approx.f32 %f358, %f357; + mul.f32 %f359, %f356, %f358; + setp.lt.f32 %p23, %f348, 0fC2D20000; + selp.f32 %f360, 0f00000000, %f359, %p23; + setp.gt.f32 %p24, %f348, 0f42D20000; + selp.f32 %f919, 0f7F800000, %f360, %p24; + setp.eq.f32 %p25, %f919, 0f7F800000; + @%p25 bra BB0_10; + + fma.rn.f32 %f919, %f919, %f36, %f919; + +BB0_10: + mov.f32 %f905, 0f40000000; + cvt.rzi.f32.f32 %f904, %f905; + add.f32 %f903, %f904, %f904; + mov.f32 %f902, 0f40800000; + sub.f32 %f901, %f902, %f903; + abs.f32 %f900, %f901; + setp.lt.f32 %p26, %f33, 0f00000000; + setp.eq.f32 %p27, %f900, 0f3F800000; + and.pred %p1, %p26, %p27; + mov.b32 %r142, %f919; + xor.b32 %r143, %r142, -2147483648; + mov.b32 %f361, %r143; + selp.f32 %f921, %f361, %f919, %p1; + setp.eq.f32 %p28, %f33, 0f00000000; + @%p28 bra BB0_13; + bra.uni BB0_11; + +BB0_13: + add.f32 %f364, %f33, %f33; + selp.f32 %f921, %f364, 0f00000000, %p27; + bra.uni BB0_14; + +BB0_11: + setp.geu.f32 %p29, %f33, 0f00000000; + @%p29 bra BB0_14; + + mov.f32 %f908, 0f40800000; + cvt.rzi.f32.f32 %f363, %f908; + setp.neu.f32 %p30, %f363, 0f40800000; + selp.f32 %f921, 0f7FFFFFFF, %f921, %p30; + +BB0_14: + add.f32 %f365, %f35, 0f40800000; + mov.b32 %r144, %f365; + setp.lt.s32 %p32, %r144, 2139095040; + @%p32 bra BB0_19; + + setp.gtu.f32 %p33, %f35, 0f7F800000; + @%p33 bra BB0_18; + bra.uni BB0_16; + +BB0_18: + add.f32 %f921, %f33, 0f40800000; + bra.uni BB0_19; + +BB0_16: + setp.neu.f32 %p34, %f35, 0f7F800000; + @%p34 bra BB0_19; + + selp.f32 %f921, 0fFF800000, 0f7F800000, %p1; + +BB0_19: + mov.u32 %r334, 1; + mov.u64 %rd163, lightMeshBuffer; + cvta.global.u64 %rd162, %rd163; + mul.lo.s32 %r333, %r343, 3; + mov.f32 %f366, 0f3F800000; + sub.f32 %f367, %f366, %f921; + setp.eq.f32 %p35, %f33, 0f3F800000; + selp.f32 %f368, 0f00000000, %f367, %p35; + cvt.sat.f32.f32 %f369, %f368; + mul.f32 %f370, %f32, %f369; + add.s32 %r147, %r333, 1; + cvt.s64.s32 %rd51, %r147; + // inline asm + call (%rd49), _rt_buffer_get_64, (%rd162, %r334, %r117, %rd51, %rd26, %rd26, %rd26); + // inline asm + ld.f32 %f371, [%rd49]; + mul.f32 %f372, %f28, %f371; + ld.f32 %f373, [%rd49+4]; + mul.f32 %f374, %f29, %f373; + neg.f32 %f375, %f374; + sub.f32 %f376, %f375, %f372; + ld.f32 %f377, [%rd49+8]; + mul.f32 %f378, %f30, %f377; + sub.f32 %f379, %f376, %f378; + cvt.sat.f32.f32 %f380, %f379; + mul.f32 %f47, %f370, %f380; + cvt.sat.f32.f32 %f48, %f31; + mul.f32 %f381, %f47, %f48; + setp.leu.f32 %p36, %f381, 0f3727C5AC; + @%p36 bra BB0_21; + + mov.u32 %r342, 1; + mov.u64 %rd165, lightMeshBuffer; + cvta.global.u64 %rd164, %rd165; + mul.lo.s32 %r341, %r343, 3; + add.u64 %rd56, %SP, 28; + cvta.to.local.u64 %rd63, %rd56; + max.f32 %f388, %f394, %f255; + sub.f32 %f389, %f27, %f388; + mov.u32 %r153, 1065353216; + st.local.u32 [%rd63], %r153; + ld.global.u32 %r148, [root]; + // inline asm + call _rt_trace_64, (%r148, %f12, %f13, %f14, %f28, %f29, %f30, %r342, %f388, %f389, %rd56, %r102); + // inline asm + add.s32 %r154, %r341, 2; + cvt.s64.s32 %rd59, %r154; + // inline asm + call (%rd57), _rt_buffer_get_64, (%rd164, %r342, %r117, %rd59, %rd26, %rd26, %rd26); + // inline asm + ld.f32 %f396, [%rd57]; + ld.local.f32 %f397, [%rd63]; + mul.f32 %f398, %f397, %f396; + ld.f32 %f399, [%rd57+4]; + mul.f32 %f400, %f397, %f399; + ld.f32 %f401, [%rd57+8]; + mul.f32 %f402, %f397, %f401; + mul.f32 %f403, %f47, %f398; + mul.f32 %f404, %f47, %f400; + mul.f32 %f405, %f47, %f402; + fma.rn.f32 %f23, %f48, %f403, %f23; + fma.rn.f32 %f22, %f48, %f404, %f22; + fma.rn.f32 %f21, %f48, %f405, %f21; + add.f32 %f20, %f20, %f397; + fma.rn.f32 %f942, %f28, %f397, %f942; + fma.rn.f32 %f941, %f29, %f397, %f941; + fma.rn.f32 %f940, %f30, %f397, %f940; + +BB0_21: + add.s32 %r7, %r7, 1; + +BB0_22: + ld.global.u32 %r155, [lightMeshBufferSize]; + add.s32 %r343, %r343, 1; + setp.lt.u32 %p38, %r343, %r155; + @%p38 bra BB0_6; + bra.uni BB0_23; + +BB0_4: + mov.f32 %f21, %f20; + mov.f32 %f22, %f20; + mov.f32 %f23, %f20; + mov.f32 %f940, %f20; + mov.f32 %f941, %f20; + mov.f32 %f942, %f20; + +BB0_23: + mov.pred %p144, %p13; + +BB0_24: + cvt.rn.f32.s32 %f406, %r7; + mov.f32 %f407, 0f3F800000; + max.f32 %f408, %f406, %f407; + rcp.rn.f32 %f409, %f408; + mul.f32 %f976, %f23, %f409; + mul.f32 %f977, %f22, %f409; + mul.f32 %f978, %f21, %f409; + div.rn.f32 %f979, %f20, %f408; + @!%p144 bra BB0_77; + bra.uni BB0_25; + +BB0_25: + abs.f32 %f907, %f911; + abs.f32 %f906, %f909; + setp.gt.f32 %p39, %f906, %f907; + neg.f32 %f416, %f910; + selp.f32 %f417, %f416, 0f00000000, %p39; + neg.f32 %f418, %f911; + selp.f32 %f419, %f909, %f418, %p39; + selp.f32 %f420, 0f00000000, %f910, %p39; + mul.f32 %f421, %f419, %f419; + fma.rn.f32 %f422, %f417, %f417, %f421; + fma.rn.f32 %f423, %f420, %f420, %f422; + sqrt.rn.f32 %f424, %f423; + rcp.rn.f32 %f425, %f424; + mul.f32 %f81, %f417, %f425; + mul.f32 %f82, %f419, %f425; + mul.f32 %f83, %f420, %f425; + mov.f32 %f415, 0f00000000; + setp.lt.s32 %p40, %r1, 1; + mov.f32 %f949, %f415; + mov.f32 %f950, %f415; + mov.f32 %f951, %f415; + mov.f32 %f940, %f415; + mov.f32 %f941, %f415; + mov.f32 %f942, %f415; + @%p40 bra BB0_76; + + mad.lo.s32 %r350, %r129, 1664525, 1013904223; + cvt.rn.f32.s32 %f432, %r1; + rcp.rn.f32 %f84, %f432; + add.u64 %rd65, %SP, 0; + cvta.to.local.u64 %rd2, %rd65; + mul.f32 %f85, %f12, 0f3456BF95; + mul.f32 %f86, %f13, 0f3456BF95; + mul.f32 %f87, %f14, 0f3456BF95; + add.u64 %rd66, %SP, 32; + cvta.to.local.u64 %rd3, %rd66; + mul.f32 %f433, %f909, %f82; + mul.f32 %f434, %f910, %f81; + sub.f32 %f88, %f434, %f433; + mul.f32 %f435, %f911, %f81; + mul.f32 %f436, %f909, %f83; + sub.f32 %f89, %f436, %f435; + mul.f32 %f437, %f910, %f83; + mul.f32 %f438, %f911, %f82; + sub.f32 %f90, %f438, %f437; + mov.f32 %f949, 0f00000000; + mov.u32 %r156, 0; + abs.f32 %f439, %f86; + abs.f32 %f440, %f85; + max.f32 %f441, %f440, %f439; + abs.f32 %f442, %f87; + max.f32 %f443, %f441, %f442; + mov.u32 %r347, %r156; + mov.f32 %f950, %f949; + mov.f32 %f951, %f949; + mov.f32 %f940, %f949; + mov.f32 %f941, %f949; + mov.f32 %f942, %f949; + +BB0_27: + cvt.rn.f32.s32 %f97, %r347; + max.f32 %f98, %f443, %f255; + mov.u32 %r349, %r156; + +BB0_28: + mad.lo.s32 %r158, %r350, 1664525, 1013904223; + and.b32 %r159, %r158, 16777215; + cvt.rn.f32.u32 %f445, %r159; + fma.rn.f32 %f446, %f445, 0f33800000, %f97; + mul.f32 %f105, %f84, %f446; + mad.lo.s32 %r350, %r158, 1664525, 1013904223; + and.b32 %r160, %r350, 16777215; + cvt.rn.f32.u32 %f447, %r160; + cvt.rn.f32.s32 %f448, %r349; + fma.rn.f32 %f449, %f447, 0f33800000, %f448; + mul.f32 %f450, %f84, %f449; + mul.f32 %f451, %f105, %f105; + sub.f32 %f453, %f407, %f451; + mov.f32 %f454, 0f00000000; + max.f32 %f455, %f454, %f453; + sqrt.rn.f32 %f106, %f455; + mul.f32 %f961, %f450, 0f40C90FDB; + abs.f32 %f108, %f961; + setp.neu.f32 %p41, %f108, 0f7F800000; + mov.f32 %f955, %f961; + @%p41 bra BB0_30; + + mul.rn.f32 %f955, %f961, %f454; + +BB0_30: + mul.f32 %f457, %f955, 0f3F22F983; + cvt.rni.s32.f32 %r360, %f457; + cvt.rn.f32.s32 %f458, %r360; + neg.f32 %f459, %f458; + mov.f32 %f460, 0f3FC90FDA; + fma.rn.f32 %f461, %f459, %f460, %f955; + mov.f32 %f462, 0f33A22168; + fma.rn.f32 %f463, %f459, %f462, %f461; + mov.f32 %f464, 0f27C234C5; + fma.rn.f32 %f956, %f459, %f464, %f463; + abs.f32 %f465, %f955; + setp.leu.f32 %p42, %f465, 0f47CE4780; + @%p42 bra BB0_41; + + mov.b32 %r19, %f955; + shr.u32 %r20, %r19, 23; + shl.b32 %r163, %r19, 8; + or.b32 %r21, %r163, -2147483648; + mov.u32 %r351, 0; + mov.u64 %rd167, 0; + mov.u64 %rd166, %rd2; + mov.u32 %r352, %r351; + +BB0_32: + .pragma "nounroll"; + shl.b64 %rd68, %rd167, 2; + mov.u64 %rd69, __cudart_i2opi_f; + add.s64 %rd70, %rd69, %rd68; + ld.const.u32 %r166, [%rd70]; + // inline asm + { + mad.lo.cc.u32 %r164, %r166, %r21, %r352; + madc.hi.u32 %r352, %r166, %r21, 0; + } + // inline asm + st.local.u32 [%rd166], %r164; + add.s32 %r351, %r351, 1; + cvt.s64.s32 %rd167, %r351; + mul.wide.s32 %rd71, %r351, 4; + add.s64 %rd166, %rd2, %rd71; + setp.ne.s32 %p43, %r351, 6; + @%p43 bra BB0_32; + + and.b32 %r169, %r20, 255; + add.s32 %r170, %r169, -128; + shr.u32 %r171, %r170, 5; + and.b32 %r26, %r19, -2147483648; + st.local.u32 [%rd2+24], %r352; + mov.u32 %r172, 6; + sub.s32 %r173, %r172, %r171; + mul.wide.s32 %rd72, %r173, 4; + add.s64 %rd9, %rd2, %rd72; + ld.local.u32 %r353, [%rd9]; + ld.local.u32 %r354, [%rd9+-4]; + and.b32 %r29, %r20, 31; + setp.eq.s32 %p44, %r29, 0; + @%p44 bra BB0_35; + + mov.u32 %r174, 32; + sub.s32 %r175, %r174, %r29; + shr.u32 %r176, %r354, %r175; + shl.b32 %r177, %r353, %r29; + add.s32 %r353, %r176, %r177; + ld.local.u32 %r178, [%rd9+-8]; + shr.u32 %r179, %r178, %r175; + shl.b32 %r180, %r354, %r29; + add.s32 %r354, %r179, %r180; + +BB0_35: + shr.u32 %r181, %r354, 30; + shl.b32 %r182, %r353, 2; + add.s32 %r355, %r181, %r182; + shl.b32 %r35, %r354, 2; + shr.u32 %r183, %r355, 31; + shr.u32 %r184, %r353, 30; + add.s32 %r36, %r183, %r184; + setp.eq.s32 %p45, %r183, 0; + @%p45 bra BB0_36; + bra.uni BB0_37; + +BB0_36: + mov.u32 %r356, %r26; + mov.u32 %r357, %r35; + bra.uni BB0_38; + +BB0_37: + not.b32 %r185, %r355; + neg.s32 %r357, %r35; + setp.eq.s32 %p46, %r35, 0; + selp.u32 %r186, 1, 0, %p46; + add.s32 %r355, %r186, %r185; + xor.b32 %r356, %r26, -2147483648; + +BB0_38: + clz.b32 %r359, %r355; + setp.eq.s32 %p47, %r359, 0; + shl.b32 %r187, %r355, %r359; + mov.u32 %r188, 32; + sub.s32 %r189, %r188, %r359; + shr.u32 %r190, %r357, %r189; + add.s32 %r191, %r190, %r187; + selp.b32 %r44, %r355, %r191, %p47; + mov.u32 %r192, -921707870; + mul.hi.u32 %r358, %r44, %r192; + setp.eq.s32 %p48, %r26, 0; + neg.s32 %r193, %r36; + selp.b32 %r360, %r36, %r193, %p48; + setp.lt.s32 %p49, %r358, 1; + @%p49 bra BB0_40; + + mul.lo.s32 %r194, %r44, -921707870; + shr.u32 %r195, %r194, 31; + shl.b32 %r196, %r358, 1; + add.s32 %r358, %r195, %r196; + add.s32 %r359, %r359, 1; + +BB0_40: + mov.u32 %r197, 126; + sub.s32 %r198, %r197, %r359; + shl.b32 %r199, %r198, 23; + add.s32 %r200, %r358, 1; + shr.u32 %r201, %r200, 7; + add.s32 %r202, %r201, 1; + shr.u32 %r203, %r202, 1; + add.s32 %r204, %r203, %r199; + or.b32 %r205, %r204, %r356; + mov.b32 %f956, %r205; + +BB0_41: + mul.rn.f32 %f114, %f956, %f956; + add.s32 %r52, %r360, 1; + and.b32 %r53, %r52, 1; + setp.eq.s32 %p50, %r53, 0; + @%p50 bra BB0_43; + bra.uni BB0_42; + +BB0_43: + mov.f32 %f468, 0f3C08839E; + mov.f32 %f469, 0fB94CA1F9; + fma.rn.f32 %f957, %f469, %f114, %f468; + bra.uni BB0_44; + +BB0_42: + mov.f32 %f466, 0fBAB6061A; + mov.f32 %f467, 0f37CCF5CE; + fma.rn.f32 %f957, %f467, %f114, %f466; + +BB0_44: + @%p50 bra BB0_46; + bra.uni BB0_45; + +BB0_46: + mov.f32 %f473, 0fBE2AAAA3; + fma.rn.f32 %f474, %f957, %f114, %f473; + fma.rn.f32 %f958, %f474, %f114, %f454; + bra.uni BB0_47; + +BB0_45: + mov.f32 %f470, 0f3D2AAAA5; + fma.rn.f32 %f471, %f957, %f114, %f470; + mov.f32 %f472, 0fBF000000; + fma.rn.f32 %f958, %f471, %f114, %f472; + +BB0_47: + fma.rn.f32 %f959, %f958, %f956, %f956; + @%p50 bra BB0_49; + + fma.rn.f32 %f959, %f958, %f114, %f407; + +BB0_49: + and.b32 %r206, %r52, 2; + setp.eq.s32 %p53, %r206, 0; + @%p53 bra BB0_51; + + mov.f32 %f478, 0fBF800000; + fma.rn.f32 %f959, %f959, %f478, %f454; + +BB0_51: + @%p41 bra BB0_53; + + mul.rn.f32 %f961, %f961, %f454; + +BB0_53: + mul.f32 %f480, %f961, 0f3F22F983; + cvt.rni.s32.f32 %r370, %f480; + cvt.rn.f32.s32 %f481, %r370; + neg.f32 %f482, %f481; + fma.rn.f32 %f484, %f482, %f460, %f961; + fma.rn.f32 %f486, %f482, %f462, %f484; + fma.rn.f32 %f962, %f482, %f464, %f486; + abs.f32 %f488, %f961; + setp.leu.f32 %p55, %f488, 0f47CE4780; + @%p55 bra BB0_64; + + mov.b32 %r55, %f961; + shr.u32 %r56, %r55, 23; + shl.b32 %r209, %r55, 8; + or.b32 %r57, %r209, -2147483648; + mov.u32 %r361, 0; + mov.u64 %rd168, %rd2; + mov.u64 %rd169, %rd26; + mov.u32 %r362, %r361; + +BB0_55: + .pragma "nounroll"; + shl.b64 %rd74, %rd169, 2; + mov.u64 %rd75, __cudart_i2opi_f; + add.s64 %rd76, %rd75, %rd74; + ld.const.u32 %r212, [%rd76]; + // inline asm + { + mad.lo.cc.u32 %r210, %r212, %r57, %r362; + madc.hi.u32 %r362, %r212, %r57, 0; + } + // inline asm + st.local.u32 [%rd168], %r210; + add.s32 %r361, %r361, 1; + cvt.s64.s32 %rd169, %r361; + mul.wide.s32 %rd77, %r361, 4; + add.s64 %rd168, %rd2, %rd77; + setp.ne.s32 %p56, %r361, 6; + @%p56 bra BB0_55; + + and.b32 %r215, %r56, 255; + add.s32 %r216, %r215, -128; + shr.u32 %r217, %r216, 5; + and.b32 %r62, %r55, -2147483648; + st.local.u32 [%rd2+24], %r362; + mov.u32 %r218, 6; + sub.s32 %r219, %r218, %r217; + mul.wide.s32 %rd78, %r219, 4; + add.s64 %rd15, %rd2, %rd78; + ld.local.u32 %r363, [%rd15]; + ld.local.u32 %r364, [%rd15+-4]; + and.b32 %r65, %r56, 31; + setp.eq.s32 %p57, %r65, 0; + @%p57 bra BB0_58; + + mov.u32 %r220, 32; + sub.s32 %r221, %r220, %r65; + shr.u32 %r222, %r364, %r221; + shl.b32 %r223, %r363, %r65; + add.s32 %r363, %r222, %r223; + ld.local.u32 %r224, [%rd15+-8]; + shr.u32 %r225, %r224, %r221; + shl.b32 %r226, %r364, %r65; + add.s32 %r364, %r225, %r226; + +BB0_58: + shr.u32 %r227, %r364, 30; + shl.b32 %r228, %r363, 2; + add.s32 %r365, %r227, %r228; + shl.b32 %r71, %r364, 2; + shr.u32 %r229, %r365, 31; + shr.u32 %r230, %r363, 30; + add.s32 %r72, %r229, %r230; + setp.eq.s32 %p58, %r229, 0; + @%p58 bra BB0_59; + bra.uni BB0_60; + +BB0_59: + mov.u32 %r366, %r62; + mov.u32 %r367, %r71; + bra.uni BB0_61; + +BB0_60: + not.b32 %r231, %r365; + neg.s32 %r367, %r71; + setp.eq.s32 %p59, %r71, 0; + selp.u32 %r232, 1, 0, %p59; + add.s32 %r365, %r232, %r231; + xor.b32 %r366, %r62, -2147483648; + +BB0_61: + clz.b32 %r369, %r365; + setp.eq.s32 %p60, %r369, 0; + shl.b32 %r233, %r365, %r369; + mov.u32 %r234, 32; + sub.s32 %r235, %r234, %r369; + shr.u32 %r236, %r367, %r235; + add.s32 %r237, %r236, %r233; + selp.b32 %r80, %r365, %r237, %p60; + mov.u32 %r238, -921707870; + mul.hi.u32 %r368, %r80, %r238; + setp.eq.s32 %p61, %r62, 0; + neg.s32 %r239, %r72; + selp.b32 %r370, %r72, %r239, %p61; + setp.lt.s32 %p62, %r368, 1; + @%p62 bra BB0_63; + + mul.lo.s32 %r240, %r80, -921707870; + shr.u32 %r241, %r240, 31; + shl.b32 %r242, %r368, 1; + add.s32 %r368, %r241, %r242; + add.s32 %r369, %r369, 1; + +BB0_63: + mov.u32 %r243, 126; + sub.s32 %r244, %r243, %r369; + shl.b32 %r245, %r244, 23; + add.s32 %r246, %r368, 1; + shr.u32 %r247, %r246, 7; + add.s32 %r248, %r247, 1; + shr.u32 %r249, %r248, 1; + add.s32 %r250, %r249, %r245; + or.b32 %r251, %r250, %r366; + mov.b32 %f962, %r251; + +BB0_64: + mul.rn.f32 %f131, %f962, %f962; + and.b32 %r88, %r370, 1; + setp.eq.s32 %p63, %r88, 0; + @%p63 bra BB0_66; + bra.uni BB0_65; + +BB0_66: + mov.f32 %f491, 0f3C08839E; + mov.f32 %f492, 0fB94CA1F9; + fma.rn.f32 %f963, %f492, %f131, %f491; + bra.uni BB0_67; + +BB0_65: + mov.f32 %f489, 0fBAB6061A; + mov.f32 %f490, 0f37CCF5CE; + fma.rn.f32 %f963, %f490, %f131, %f489; + +BB0_67: + @%p63 bra BB0_69; + bra.uni BB0_68; + +BB0_69: + mov.f32 %f496, 0fBE2AAAA3; + fma.rn.f32 %f497, %f963, %f131, %f496; + fma.rn.f32 %f964, %f497, %f131, %f454; + bra.uni BB0_70; + +BB0_68: + mov.f32 %f493, 0f3D2AAAA5; + fma.rn.f32 %f494, %f963, %f131, %f493; + mov.f32 %f495, 0fBF000000; + fma.rn.f32 %f964, %f494, %f131, %f495; + +BB0_70: + fma.rn.f32 %f965, %f964, %f962, %f962; + @%p63 bra BB0_72; + + fma.rn.f32 %f965, %f964, %f131, %f407; + +BB0_72: + and.b32 %r252, %r370, 2; + setp.eq.s32 %p66, %r252, 0; + @%p66 bra BB0_74; + + mov.f32 %f501, 0fBF800000; + fma.rn.f32 %f965, %f965, %f501, %f454; + +BB0_74: + mul.f32 %f510, %f106, %f959; + mul.f32 %f511, %f106, %f965; + mul.f32 %f512, %f81, %f511; + mul.f32 %f513, %f82, %f511; + mul.f32 %f514, %f83, %f511; + fma.rn.f32 %f515, %f90, %f510, %f512; + fma.rn.f32 %f516, %f89, %f510, %f513; + fma.rn.f32 %f517, %f88, %f510, %f514; + fma.rn.f32 %f505, %f909, %f105, %f515; + fma.rn.f32 %f506, %f910, %f105, %f516; + fma.rn.f32 %f507, %f911, %f105, %f517; + mov.u32 %r254, 0; + st.local.u32 [%rd3+8], %r254; + st.local.u32 [%rd3+4], %r254; + st.local.u32 [%rd3], %r254; + ld.global.u32 %r253, [root]; + mov.f32 %f509, 0f6C4ECB8F; + // inline asm + call _rt_trace_64, (%r253, %f12, %f13, %f14, %f505, %f506, %f507, %r254, %f98, %f509, %rd66, %r117); + // inline asm + ld.local.f32 %f518, [%rd3]; + max.f32 %f520, %f518, %f454; + ld.local.f32 %f521, [%rd3+4]; + max.f32 %f522, %f521, %f454; + ld.local.f32 %f523, [%rd3+8]; + max.f32 %f524, %f523, %f454; + setp.lt.f32 %p67, %f518, 0f00000000; + selp.f32 %f525, 0f00000000, 0f3F800000, %p67; + fma.rn.f32 %f942, %f505, %f525, %f942; + fma.rn.f32 %f941, %f506, %f525, %f941; + fma.rn.f32 %f940, %f507, %f525, %f940; + mul.f32 %f526, %f910, %f506; + fma.rn.f32 %f527, %f909, %f505, %f526; + fma.rn.f32 %f528, %f911, %f507, %f527; + cvt.sat.f32.f32 %f529, %f528; + fma.rn.f32 %f951, %f520, %f529, %f951; + fma.rn.f32 %f950, %f522, %f529, %f950; + fma.rn.f32 %f949, %f524, %f529, %f949; + add.s32 %r349, %r349, 1; + setp.lt.s32 %p68, %r349, %r1; + @%p68 bra BB0_28; + + add.s32 %r347, %r347, 1; + setp.lt.s32 %p69, %r347, %r1; + @%p69 bra BB0_27; + +BB0_76: + mul.lo.s32 %r256, %r1, %r1; + cvt.rn.f32.s32 %f530, %r256; + rcp.rn.f32 %f531, %f530; + mul.f32 %f532, %f951, %f531; + mul.f32 %f533, %f950, %f531; + mul.f32 %f534, %f949, %f531; + div.rn.f32 %f979, %f415, %f530; + fma.rn.f32 %f976, %f951, %f531, %f532; + fma.rn.f32 %f977, %f950, %f531, %f533; + fma.rn.f32 %f978, %f949, %f531, %f534; + +BB0_77: + ld.global.u32 %r372, [imageEnabled]; + and.b32 %r257, %r372, 8; + setp.eq.s32 %p70, %r257, 0; + @%p70 bra BB0_90; + + mov.u32 %r335, 2; + cvt.u64.u32 %rd82, %r3; + cvt.u64.u32 %rd83, %r4; + mov.u64 %rd86, image_Mask; + cvta.global.u64 %rd81, %rd86; + // inline asm + call (%rd80), _rt_buffer_get_64, (%rd81, %r335, %r335, %rd82, %rd83, %rd26, %rd26); + // inline asm + mov.f32 %f538, 0f3E68BA2E; + cvt.rzi.f32.f32 %f539, %f538; + fma.rn.f32 %f540, %f539, 0fC0000000, 0f3EE8BA2E; + abs.f32 %f166, %f540; + abs.f32 %f167, %f979; + setp.lt.f32 %p71, %f167, 0f00800000; + mul.f32 %f541, %f167, 0f4B800000; + selp.f32 %f542, 0fC3170000, 0fC2FE0000, %p71; + selp.f32 %f543, %f541, %f167, %p71; + mov.b32 %r260, %f543; + and.b32 %r261, %r260, 8388607; + or.b32 %r262, %r261, 1065353216; + mov.b32 %f544, %r262; + shr.u32 %r263, %r260, 23; + cvt.rn.f32.u32 %f545, %r263; + add.f32 %f546, %f542, %f545; + setp.gt.f32 %p72, %f544, 0f3FB504F3; + mul.f32 %f547, %f544, 0f3F000000; + add.f32 %f548, %f546, 0f3F800000; + selp.f32 %f549, %f547, %f544, %p72; + selp.f32 %f550, %f548, %f546, %p72; + add.f32 %f551, %f549, 0fBF800000; + add.f32 %f537, %f549, 0f3F800000; + // inline asm + rcp.approx.ftz.f32 %f536,%f537; + // inline asm + add.f32 %f552, %f551, %f551; + mul.f32 %f553, %f536, %f552; + mul.f32 %f554, %f553, %f553; + mov.f32 %f555, 0f3C4CAF63; + mov.f32 %f556, 0f3B18F0FE; + fma.rn.f32 %f557, %f556, %f554, %f555; + mov.f32 %f558, 0f3DAAAABD; + fma.rn.f32 %f559, %f557, %f554, %f558; + mul.rn.f32 %f560, %f559, %f554; + mul.rn.f32 %f561, %f560, %f553; + sub.f32 %f562, %f551, %f553; + neg.f32 %f563, %f553; + add.f32 %f564, %f562, %f562; + fma.rn.f32 %f565, %f563, %f551, %f564; + mul.rn.f32 %f566, %f536, %f565; + add.f32 %f567, %f561, %f553; + sub.f32 %f568, %f553, %f567; + add.f32 %f569, %f561, %f568; + add.f32 %f570, %f566, %f569; + add.f32 %f571, %f567, %f570; + sub.f32 %f572, %f567, %f571; + add.f32 %f573, %f570, %f572; + mov.f32 %f574, 0f3F317200; + mul.rn.f32 %f575, %f550, %f574; + mov.f32 %f576, 0f35BFBE8E; + mul.rn.f32 %f577, %f550, %f576; + add.f32 %f578, %f575, %f571; + sub.f32 %f579, %f575, %f578; + add.f32 %f580, %f571, %f579; + add.f32 %f581, %f573, %f580; + add.f32 %f582, %f577, %f581; + add.f32 %f583, %f578, %f582; + sub.f32 %f584, %f578, %f583; + add.f32 %f585, %f582, %f584; + mov.f32 %f586, 0f3EE8BA2E; + mul.rn.f32 %f587, %f586, %f583; + neg.f32 %f588, %f587; + fma.rn.f32 %f589, %f586, %f583, %f588; + fma.rn.f32 %f590, %f586, %f585, %f589; + mov.f32 %f591, 0f00000000; + fma.rn.f32 %f592, %f591, %f583, %f590; + add.rn.f32 %f593, %f587, %f592; + neg.f32 %f594, %f593; + add.rn.f32 %f595, %f587, %f594; + add.rn.f32 %f596, %f595, %f592; + mov.b32 %r264, %f593; + setp.eq.s32 %p73, %r264, 1118925336; + add.s32 %r265, %r264, -1; + mov.b32 %f597, %r265; + add.f32 %f598, %f596, 0f37000000; + selp.f32 %f599, %f597, %f593, %p73; + selp.f32 %f168, %f598, %f596, %p73; + mul.f32 %f600, %f599, 0f3FB8AA3B; + cvt.rzi.f32.f32 %f601, %f600; + mov.f32 %f602, 0fBF317200; + fma.rn.f32 %f603, %f601, %f602, %f599; + mov.f32 %f604, 0fB5BFBE8E; + fma.rn.f32 %f605, %f601, %f604, %f603; + mul.f32 %f606, %f605, 0f3FB8AA3B; + ex2.approx.ftz.f32 %f607, %f606; + add.f32 %f608, %f601, 0f00000000; + ex2.approx.f32 %f609, %f608; + mul.f32 %f610, %f607, %f609; + setp.lt.f32 %p74, %f599, 0fC2D20000; + selp.f32 %f611, 0f00000000, %f610, %p74; + setp.gt.f32 %p75, %f599, 0f42D20000; + selp.f32 %f980, 0f7F800000, %f611, %p75; + setp.eq.f32 %p76, %f980, 0f7F800000; + @%p76 bra BB0_80; + + fma.rn.f32 %f980, %f980, %f168, %f980; + +BB0_80: + setp.lt.f32 %p77, %f979, 0f00000000; + setp.eq.f32 %p78, %f166, 0f3F800000; + and.pred %p3, %p77, %p78; + mov.b32 %r266, %f980; + xor.b32 %r267, %r266, -2147483648; + mov.b32 %f612, %r267; + selp.f32 %f982, %f612, %f980, %p3; + setp.eq.f32 %p79, %f979, 0f00000000; + @%p79 bra BB0_83; + bra.uni BB0_81; + +BB0_83: + add.f32 %f615, %f979, %f979; + selp.f32 %f982, %f615, 0f00000000, %p78; + bra.uni BB0_84; + +BB0_138: + mov.u64 %rd154, image_HDR; + cvta.global.u64 %rd149, %rd154; + mov.u32 %r330, 8; + mov.u64 %rd153, 0; + // inline asm + call (%rd148), _rt_buffer_get_64, (%rd149, %r101, %r330, %rd19, %rd20, %rd153, %rd153); + // inline asm + mov.f32 %f899, 0f00000000; + // inline asm + { cvt.rn.f16.f32 %rs50, %f899;} + + // inline asm + mov.u16 %rs51, 0; + st.v4.u16 [%rd148], {%rs50, %rs50, %rs50, %rs51}; + +BB0_139: + ld.global.u8 %rs52, [imageEnabled]; + and.b16 %rs53, %rs52, 64; + setp.eq.s16 %p143, %rs53, 0; + @%p143 bra BB0_141; + + cvt.u64.u32 %rd157, %r3; + cvt.u64.u32 %rd158, %r4; + mov.u64 %rd161, image_Dir; + cvta.global.u64 %rd156, %rd161; + mov.u64 %rd160, 0; + // inline asm + call (%rd155), _rt_buffer_get_64, (%rd156, %r101, %r102, %rd157, %rd158, %rd160, %rd160); + // inline asm + mov.u16 %rs54, 0; + st.v4.u8 [%rd155], {%rs54, %rs54, %rs54, %rs54}; + bra.uni BB0_141; + +BB0_81: + setp.geu.f32 %p80, %f979, 0f00000000; + @%p80 bra BB0_84; + + cvt.rzi.f32.f32 %f614, %f586; + setp.neu.f32 %p81, %f614, 0f3EE8BA2E; + selp.f32 %f982, 0f7FFFFFFF, %f982, %p81; + +BB0_84: + add.f32 %f616, %f167, 0f3EE8BA2E; + mov.b32 %r268, %f616; + setp.lt.s32 %p83, %r268, 2139095040; + @%p83 bra BB0_89; + + setp.gtu.f32 %p84, %f167, 0f7F800000; + @%p84 bra BB0_88; + bra.uni BB0_86; + +BB0_88: + add.f32 %f982, %f979, 0f3EE8BA2E; + bra.uni BB0_89; + +BB0_86: + setp.neu.f32 %p85, %f167, 0f7F800000; + @%p85 bra BB0_89; + + selp.f32 %f982, 0fFF800000, 0f7F800000, %p3; + +BB0_89: + mul.f32 %f617, %f982, 0f437F0000; + setp.eq.f32 %p86, %f979, 0f3F800000; + selp.f32 %f618, 0f437F0000, %f617, %p86; + cvt.rzi.u32.f32 %r269, %f618; + cvt.u16.u32 %rs10, %r269; + mov.u16 %rs11, 255; + st.v2.u8 [%rd80], {%rs10, %rs11}; + ld.global.u32 %r372, [imageEnabled]; + +BB0_90: + ld.global.f32 %f619, [lightColor]; + mul.f32 %f179, %f976, %f619; + ld.global.f32 %f620, [lightColor+4]; + mul.f32 %f180, %f977, %f620; + ld.global.f32 %f621, [lightColor+8]; + mul.f32 %f181, %f978, %f621; + and.b32 %r270, %r372, 1; + setp.eq.b32 %p87, %r270, 1; + @!%p87 bra BB0_125; + bra.uni BB0_91; + +BB0_91: + mov.f32 %f624, 0f3E666666; + cvt.rzi.f32.f32 %f625, %f624; + fma.rn.f32 %f626, %f625, 0fC0000000, 0f3EE66666; + abs.f32 %f182, %f626; + abs.f32 %f183, %f179; + setp.lt.f32 %p88, %f183, 0f00800000; + mul.f32 %f627, %f183, 0f4B800000; + selp.f32 %f628, 0fC3170000, 0fC2FE0000, %p88; + selp.f32 %f629, %f627, %f183, %p88; + mov.b32 %r271, %f629; + and.b32 %r272, %r271, 8388607; + or.b32 %r273, %r272, 1065353216; + mov.b32 %f630, %r273; + shr.u32 %r274, %r271, 23; + cvt.rn.f32.u32 %f631, %r274; + add.f32 %f632, %f628, %f631; + setp.gt.f32 %p89, %f630, 0f3FB504F3; + mul.f32 %f633, %f630, 0f3F000000; + add.f32 %f634, %f632, 0f3F800000; + selp.f32 %f635, %f633, %f630, %p89; + selp.f32 %f636, %f634, %f632, %p89; + add.f32 %f637, %f635, 0fBF800000; + add.f32 %f623, %f635, 0f3F800000; + // inline asm + rcp.approx.ftz.f32 %f622,%f623; + // inline asm + add.f32 %f638, %f637, %f637; + mul.f32 %f639, %f622, %f638; + mul.f32 %f640, %f639, %f639; + mov.f32 %f641, 0f3C4CAF63; + mov.f32 %f642, 0f3B18F0FE; + fma.rn.f32 %f643, %f642, %f640, %f641; + mov.f32 %f644, 0f3DAAAABD; + fma.rn.f32 %f645, %f643, %f640, %f644; + mul.rn.f32 %f646, %f645, %f640; + mul.rn.f32 %f647, %f646, %f639; + sub.f32 %f648, %f637, %f639; + neg.f32 %f649, %f639; + add.f32 %f650, %f648, %f648; + fma.rn.f32 %f651, %f649, %f637, %f650; + mul.rn.f32 %f652, %f622, %f651; + add.f32 %f653, %f647, %f639; + sub.f32 %f654, %f639, %f653; + add.f32 %f655, %f647, %f654; + add.f32 %f656, %f652, %f655; + add.f32 %f657, %f653, %f656; + sub.f32 %f658, %f653, %f657; + add.f32 %f659, %f656, %f658; + mov.f32 %f660, 0f3F317200; + mul.rn.f32 %f661, %f636, %f660; + mov.f32 %f662, 0f35BFBE8E; + mul.rn.f32 %f663, %f636, %f662; + add.f32 %f664, %f661, %f657; + sub.f32 %f665, %f661, %f664; + add.f32 %f666, %f657, %f665; + add.f32 %f667, %f659, %f666; + add.f32 %f668, %f663, %f667; + add.f32 %f669, %f664, %f668; + sub.f32 %f670, %f664, %f669; + add.f32 %f671, %f668, %f670; + mov.f32 %f672, 0f3EE66666; + mul.rn.f32 %f673, %f672, %f669; + neg.f32 %f674, %f673; + fma.rn.f32 %f675, %f672, %f669, %f674; + fma.rn.f32 %f676, %f672, %f671, %f675; + mov.f32 %f677, 0f00000000; + fma.rn.f32 %f678, %f677, %f669, %f676; + add.rn.f32 %f679, %f673, %f678; + neg.f32 %f680, %f679; + add.rn.f32 %f681, %f673, %f680; + add.rn.f32 %f682, %f681, %f678; + mov.b32 %r275, %f679; + setp.eq.s32 %p90, %r275, 1118925336; + add.s32 %r276, %r275, -1; + mov.b32 %f683, %r276; + add.f32 %f684, %f682, 0f37000000; + selp.f32 %f685, %f683, %f679, %p90; + selp.f32 %f184, %f684, %f682, %p90; + mul.f32 %f686, %f685, 0f3FB8AA3B; + cvt.rzi.f32.f32 %f687, %f686; + mov.f32 %f688, 0fBF317200; + fma.rn.f32 %f689, %f687, %f688, %f685; + mov.f32 %f690, 0fB5BFBE8E; + fma.rn.f32 %f691, %f687, %f690, %f689; + mul.f32 %f692, %f691, 0f3FB8AA3B; + ex2.approx.ftz.f32 %f693, %f692; + add.f32 %f694, %f687, 0f00000000; + ex2.approx.f32 %f695, %f694; + mul.f32 %f696, %f693, %f695; + setp.lt.f32 %p91, %f685, 0fC2D20000; + selp.f32 %f697, 0f00000000, %f696, %p91; + setp.gt.f32 %p92, %f685, 0f42D20000; + selp.f32 %f983, 0f7F800000, %f697, %p92; + setp.eq.f32 %p93, %f983, 0f7F800000; + @%p93 bra BB0_93; + + fma.rn.f32 %f983, %f983, %f184, %f983; + +BB0_93: + setp.lt.f32 %p94, %f179, 0f00000000; + setp.eq.f32 %p95, %f182, 0f3F800000; + and.pred %p4, %p94, %p95; + mov.b32 %r277, %f983; + xor.b32 %r278, %r277, -2147483648; + mov.b32 %f698, %r278; + selp.f32 %f985, %f698, %f983, %p4; + setp.eq.f32 %p96, %f179, 0f00000000; + @%p96 bra BB0_96; + bra.uni BB0_94; + +BB0_96: + add.f32 %f701, %f179, %f179; + selp.f32 %f985, %f701, 0f00000000, %p95; + bra.uni BB0_97; + +BB0_94: + setp.geu.f32 %p97, %f179, 0f00000000; + @%p97 bra BB0_97; + + cvt.rzi.f32.f32 %f700, %f672; + setp.neu.f32 %p98, %f700, 0f3EE66666; + selp.f32 %f985, 0f7FFFFFFF, %f985, %p98; + +BB0_97: + add.f32 %f702, %f183, 0f3EE66666; + mov.b32 %r279, %f702; + setp.lt.s32 %p100, %r279, 2139095040; + @%p100 bra BB0_102; + + setp.gtu.f32 %p101, %f183, 0f7F800000; + @%p101 bra BB0_101; + bra.uni BB0_99; + +BB0_101: + add.f32 %f985, %f179, 0f3EE66666; + bra.uni BB0_102; + +BB0_99: + setp.neu.f32 %p102, %f183, 0f7F800000; + @%p102 bra BB0_102; + + selp.f32 %f985, 0fFF800000, 0f7F800000, %p4; + +BB0_102: + setp.eq.f32 %p103, %f179, 0f3F800000; + selp.f32 %f195, 0f3F800000, %f985, %p103; + abs.f32 %f196, %f180; + setp.lt.f32 %p104, %f196, 0f00800000; + mul.f32 %f705, %f196, 0f4B800000; + selp.f32 %f706, 0fC3170000, 0fC2FE0000, %p104; + selp.f32 %f707, %f705, %f196, %p104; + mov.b32 %r280, %f707; + and.b32 %r281, %r280, 8388607; + or.b32 %r282, %r281, 1065353216; + mov.b32 %f708, %r282; + shr.u32 %r283, %r280, 23; + cvt.rn.f32.u32 %f709, %r283; + add.f32 %f710, %f706, %f709; + setp.gt.f32 %p105, %f708, 0f3FB504F3; + mul.f32 %f711, %f708, 0f3F000000; + add.f32 %f712, %f710, 0f3F800000; + selp.f32 %f713, %f711, %f708, %p105; + selp.f32 %f714, %f712, %f710, %p105; + add.f32 %f715, %f713, 0fBF800000; + add.f32 %f704, %f713, 0f3F800000; + // inline asm + rcp.approx.ftz.f32 %f703,%f704; + // inline asm + add.f32 %f716, %f715, %f715; + mul.f32 %f717, %f703, %f716; + mul.f32 %f718, %f717, %f717; + fma.rn.f32 %f721, %f642, %f718, %f641; + fma.rn.f32 %f723, %f721, %f718, %f644; + mul.rn.f32 %f724, %f723, %f718; + mul.rn.f32 %f725, %f724, %f717; + sub.f32 %f726, %f715, %f717; + neg.f32 %f727, %f717; + add.f32 %f728, %f726, %f726; + fma.rn.f32 %f729, %f727, %f715, %f728; + mul.rn.f32 %f730, %f703, %f729; + add.f32 %f731, %f725, %f717; + sub.f32 %f732, %f717, %f731; + add.f32 %f733, %f725, %f732; + add.f32 %f734, %f730, %f733; + add.f32 %f735, %f731, %f734; + sub.f32 %f736, %f731, %f735; + add.f32 %f737, %f734, %f736; + mul.rn.f32 %f739, %f714, %f660; + mul.rn.f32 %f741, %f714, %f662; + add.f32 %f742, %f739, %f735; + sub.f32 %f743, %f739, %f742; + add.f32 %f744, %f735, %f743; + add.f32 %f745, %f737, %f744; + add.f32 %f746, %f741, %f745; + add.f32 %f747, %f742, %f746; + sub.f32 %f748, %f742, %f747; + add.f32 %f749, %f746, %f748; + mul.rn.f32 %f751, %f672, %f747; + neg.f32 %f752, %f751; + fma.rn.f32 %f753, %f672, %f747, %f752; + fma.rn.f32 %f754, %f672, %f749, %f753; + fma.rn.f32 %f756, %f677, %f747, %f754; + add.rn.f32 %f757, %f751, %f756; + neg.f32 %f758, %f757; + add.rn.f32 %f759, %f751, %f758; + add.rn.f32 %f760, %f759, %f756; + mov.b32 %r284, %f757; + setp.eq.s32 %p106, %r284, 1118925336; + add.s32 %r285, %r284, -1; + mov.b32 %f761, %r285; + add.f32 %f762, %f760, 0f37000000; + selp.f32 %f763, %f761, %f757, %p106; + selp.f32 %f197, %f762, %f760, %p106; + mul.f32 %f764, %f763, 0f3FB8AA3B; + cvt.rzi.f32.f32 %f765, %f764; + fma.rn.f32 %f767, %f765, %f688, %f763; + fma.rn.f32 %f769, %f765, %f690, %f767; + mul.f32 %f770, %f769, 0f3FB8AA3B; + ex2.approx.ftz.f32 %f771, %f770; + add.f32 %f772, %f765, 0f00000000; + ex2.approx.f32 %f773, %f772; + mul.f32 %f774, %f771, %f773; + setp.lt.f32 %p107, %f763, 0fC2D20000; + selp.f32 %f775, 0f00000000, %f774, %p107; + setp.gt.f32 %p108, %f763, 0f42D20000; + selp.f32 %f986, 0f7F800000, %f775, %p108; + setp.eq.f32 %p109, %f986, 0f7F800000; + @%p109 bra BB0_104; + + fma.rn.f32 %f986, %f986, %f197, %f986; + +BB0_104: + setp.lt.f32 %p110, %f180, 0f00000000; + and.pred %p5, %p110, %p95; + mov.b32 %r286, %f986; + xor.b32 %r287, %r286, -2147483648; + mov.b32 %f776, %r287; + selp.f32 %f988, %f776, %f986, %p5; + setp.eq.f32 %p112, %f180, 0f00000000; + @%p112 bra BB0_107; + bra.uni BB0_105; + +BB0_107: + add.f32 %f779, %f180, %f180; + selp.f32 %f988, %f779, 0f00000000, %p95; + bra.uni BB0_108; + +BB0_105: + setp.geu.f32 %p113, %f180, 0f00000000; + @%p113 bra BB0_108; + + cvt.rzi.f32.f32 %f778, %f672; + setp.neu.f32 %p114, %f778, 0f3EE66666; + selp.f32 %f988, 0f7FFFFFFF, %f988, %p114; + +BB0_108: + add.f32 %f780, %f196, 0f3EE66666; + mov.b32 %r288, %f780; + setp.lt.s32 %p116, %r288, 2139095040; + @%p116 bra BB0_113; + + setp.gtu.f32 %p117, %f196, 0f7F800000; + @%p117 bra BB0_112; + bra.uni BB0_110; + +BB0_112: + add.f32 %f988, %f180, 0f3EE66666; + bra.uni BB0_113; + +BB0_110: + setp.neu.f32 %p118, %f196, 0f7F800000; + @%p118 bra BB0_113; + + selp.f32 %f988, 0fFF800000, 0f7F800000, %p5; + +BB0_113: + setp.eq.f32 %p119, %f180, 0f3F800000; + selp.f32 %f208, 0f3F800000, %f988, %p119; + abs.f32 %f209, %f181; + setp.lt.f32 %p120, %f209, 0f00800000; + mul.f32 %f783, %f209, 0f4B800000; + selp.f32 %f784, 0fC3170000, 0fC2FE0000, %p120; + selp.f32 %f785, %f783, %f209, %p120; + mov.b32 %r289, %f785; + and.b32 %r290, %r289, 8388607; + or.b32 %r291, %r290, 1065353216; + mov.b32 %f786, %r291; + shr.u32 %r292, %r289, 23; + cvt.rn.f32.u32 %f787, %r292; + add.f32 %f788, %f784, %f787; + setp.gt.f32 %p121, %f786, 0f3FB504F3; + mul.f32 %f789, %f786, 0f3F000000; + add.f32 %f790, %f788, 0f3F800000; + selp.f32 %f791, %f789, %f786, %p121; + selp.f32 %f792, %f790, %f788, %p121; + add.f32 %f793, %f791, 0fBF800000; + add.f32 %f782, %f791, 0f3F800000; + // inline asm + rcp.approx.ftz.f32 %f781,%f782; + // inline asm + add.f32 %f794, %f793, %f793; + mul.f32 %f795, %f781, %f794; + mul.f32 %f796, %f795, %f795; + fma.rn.f32 %f799, %f642, %f796, %f641; + fma.rn.f32 %f801, %f799, %f796, %f644; + mul.rn.f32 %f802, %f801, %f796; + mul.rn.f32 %f803, %f802, %f795; + sub.f32 %f804, %f793, %f795; + neg.f32 %f805, %f795; + add.f32 %f806, %f804, %f804; + fma.rn.f32 %f807, %f805, %f793, %f806; + mul.rn.f32 %f808, %f781, %f807; + add.f32 %f809, %f803, %f795; + sub.f32 %f810, %f795, %f809; + add.f32 %f811, %f803, %f810; + add.f32 %f812, %f808, %f811; + add.f32 %f813, %f809, %f812; + sub.f32 %f814, %f809, %f813; + add.f32 %f815, %f812, %f814; + mul.rn.f32 %f817, %f792, %f660; + mul.rn.f32 %f819, %f792, %f662; + add.f32 %f820, %f817, %f813; + sub.f32 %f821, %f817, %f820; + add.f32 %f822, %f813, %f821; + add.f32 %f823, %f815, %f822; + add.f32 %f824, %f819, %f823; + add.f32 %f825, %f820, %f824; + sub.f32 %f826, %f820, %f825; + add.f32 %f827, %f824, %f826; + mul.rn.f32 %f829, %f672, %f825; + neg.f32 %f830, %f829; + fma.rn.f32 %f831, %f672, %f825, %f830; + fma.rn.f32 %f832, %f672, %f827, %f831; + fma.rn.f32 %f834, %f677, %f825, %f832; + add.rn.f32 %f835, %f829, %f834; + neg.f32 %f836, %f835; + add.rn.f32 %f837, %f829, %f836; + add.rn.f32 %f838, %f837, %f834; + mov.b32 %r293, %f835; + setp.eq.s32 %p122, %r293, 1118925336; + add.s32 %r294, %r293, -1; + mov.b32 %f839, %r294; + add.f32 %f840, %f838, 0f37000000; + selp.f32 %f841, %f839, %f835, %p122; + selp.f32 %f210, %f840, %f838, %p122; + mul.f32 %f842, %f841, 0f3FB8AA3B; + cvt.rzi.f32.f32 %f843, %f842; + fma.rn.f32 %f845, %f843, %f688, %f841; + fma.rn.f32 %f847, %f843, %f690, %f845; + mul.f32 %f848, %f847, 0f3FB8AA3B; + ex2.approx.ftz.f32 %f849, %f848; + add.f32 %f850, %f843, 0f00000000; + ex2.approx.f32 %f851, %f850; + mul.f32 %f852, %f849, %f851; + setp.lt.f32 %p123, %f841, 0fC2D20000; + selp.f32 %f853, 0f00000000, %f852, %p123; + setp.gt.f32 %p124, %f841, 0f42D20000; + selp.f32 %f989, 0f7F800000, %f853, %p124; + setp.eq.f32 %p125, %f989, 0f7F800000; + @%p125 bra BB0_115; + + fma.rn.f32 %f989, %f989, %f210, %f989; + +BB0_115: + setp.lt.f32 %p126, %f181, 0f00000000; + and.pred %p6, %p126, %p95; + mov.b32 %r295, %f989; + xor.b32 %r296, %r295, -2147483648; + mov.b32 %f854, %r296; + selp.f32 %f991, %f854, %f989, %p6; + setp.eq.f32 %p128, %f181, 0f00000000; + @%p128 bra BB0_118; + bra.uni BB0_116; + +BB0_118: + add.f32 %f857, %f181, %f181; + selp.f32 %f991, %f857, 0f00000000, %p95; + bra.uni BB0_119; + +BB0_116: + setp.geu.f32 %p129, %f181, 0f00000000; + @%p129 bra BB0_119; + + cvt.rzi.f32.f32 %f856, %f672; + setp.neu.f32 %p130, %f856, 0f3EE66666; + selp.f32 %f991, 0f7FFFFFFF, %f991, %p130; + +BB0_119: + add.f32 %f858, %f209, 0f3EE66666; + mov.b32 %r297, %f858; + setp.lt.s32 %p132, %r297, 2139095040; + @%p132 bra BB0_124; + + setp.gtu.f32 %p133, %f209, 0f7F800000; + @%p133 bra BB0_123; + bra.uni BB0_121; + +BB0_123: + add.f32 %f991, %f181, 0f3EE66666; + bra.uni BB0_124; + +BB0_121: + setp.neu.f32 %p134, %f209, 0f7F800000; + @%p134 bra BB0_124; + + selp.f32 %f991, 0fFF800000, 0f7F800000, %p6; + +BB0_124: + mov.u32 %r336, 2; + setp.eq.f32 %p135, %f181, 0f3F800000; + selp.f32 %f859, 0f3F800000, %f991, %p135; + cvt.u64.u32 %rd90, %r4; + cvt.u64.u32 %rd89, %r3; + mov.u64 %rd93, image; + cvta.global.u64 %rd88, %rd93; + // inline asm + call (%rd87), _rt_buffer_get_64, (%rd88, %r336, %r102, %rd89, %rd90, %rd26, %rd26); + // inline asm + cvt.sat.f32.f32 %f860, %f859; + mul.f32 %f861, %f860, 0f437FFD71; + cvt.rzi.u32.f32 %r300, %f861; + cvt.sat.f32.f32 %f862, %f208; + mul.f32 %f863, %f862, 0f437FFD71; + cvt.rzi.u32.f32 %r301, %f863; + cvt.sat.f32.f32 %f864, %f195; + mul.f32 %f865, %f864, 0f437FFD71; + cvt.rzi.u32.f32 %r302, %f865; + cvt.u16.u32 %rs12, %r300; + cvt.u16.u32 %rs13, %r302; + cvt.u16.u32 %rs14, %r301; + mov.u16 %rs15, 255; + st.v4.u8 [%rd87], {%rs12, %rs14, %rs13, %rs15}; + ld.global.u32 %r372, [imageEnabled]; + +BB0_125: + and.b32 %r303, %r372, 4; + setp.eq.s32 %p136, %r303, 0; + @%p136 bra BB0_129; + + ld.global.u32 %r304, [additive]; + setp.eq.s32 %p137, %r304, 0; + cvt.u64.u32 %rd17, %r3; + cvt.u64.u32 %rd18, %r4; + // inline asm + { cvt.rn.f16.f32 %rs16, %f407;} + + // inline asm + @%p137 bra BB0_128; + + mov.u32 %r337, 2; + mov.u64 %rd106, image_HDR; + cvta.global.u64 %rd95, %rd106; + mov.u32 %r308, 8; + // inline asm + call (%rd94), _rt_buffer_get_64, (%rd95, %r337, %r308, %rd17, %rd18, %rd26, %rd26); + // inline asm + ld.v4.u16 {%rs23, %rs24, %rs25, %rs26}, [%rd94]; + // inline asm + { cvt.f32.f16 %f867, %rs23;} + + // inline asm + // inline asm + { cvt.f32.f16 %f868, %rs24;} + + // inline asm + // inline asm + { cvt.f32.f16 %f869, %rs25;} + + // inline asm + // inline asm + call (%rd100), _rt_buffer_get_64, (%rd95, %r337, %r308, %rd17, %rd18, %rd26, %rd26); + // inline asm + add.f32 %f870, %f179, %f867; + add.f32 %f871, %f180, %f868; + add.f32 %f872, %f181, %f869; + // inline asm + { cvt.rn.f16.f32 %rs22, %f872;} + + // inline asm + // inline asm + { cvt.rn.f16.f32 %rs21, %f871;} + + // inline asm + // inline asm + { cvt.rn.f16.f32 %rs20, %f870;} + + // inline asm + st.v4.u16 [%rd100], {%rs20, %rs21, %rs22, %rs16}; + bra.uni BB0_129; + +BB0_128: + mov.u32 %r338, 2; + mov.u64 %rd113, image_HDR; + cvta.global.u64 %rd108, %rd113; + mov.u32 %r310, 8; + // inline asm + call (%rd107), _rt_buffer_get_64, (%rd108, %r338, %r310, %rd17, %rd18, %rd26, %rd26); + // inline asm + // inline asm + { cvt.rn.f16.f32 %rs29, %f181;} + + // inline asm + // inline asm + { cvt.rn.f16.f32 %rs28, %f180;} + + // inline asm + // inline asm + { cvt.rn.f16.f32 %rs27, %f179;} + + // inline asm + st.v4.u16 [%rd107], {%rs27, %rs28, %rs29, %rs16}; + +BB0_129: + ld.global.u8 %rs30, [imageEnabled]; + and.b16 %rs31, %rs30, 64; + setp.eq.s16 %p138, %rs31, 0; + @%p138 bra BB0_141; + + mov.u32 %r339, 2; + mul.f32 %f876, %f941, %f941; + fma.rn.f32 %f877, %f942, %f942, %f876; + fma.rn.f32 %f878, %f940, %f940, %f877; + sqrt.rn.f32 %f879, %f878; + rcp.rn.f32 %f880, %f879; + mul.f32 %f881, %f942, %f880; + mul.f32 %f882, %f941, %f880; + mul.f32 %f883, %f940, %f880; + cvt.u64.u32 %rd117, %r4; + cvt.u64.u32 %rd116, %r3; + mov.u64 %rd120, image_Dir; + cvta.global.u64 %rd115, %rd120; + // inline asm + call (%rd114), _rt_buffer_get_64, (%rd115, %r339, %r102, %rd116, %rd117, %rd26, %rd26); + // inline asm + fma.rn.f32 %f884, %f881, 0f3F000000, 0f3F000000; + mul.f32 %f885, %f884, 0f437F0000; + cvt.rzi.u32.f32 %r313, %f885; + fma.rn.f32 %f886, %f882, 0f3F000000, 0f3F000000; + mul.f32 %f887, %f886, 0f437F0000; + cvt.rzi.u32.f32 %r314, %f887; + fma.rn.f32 %f888, %f883, 0f3F000000, 0f3F000000; + mul.f32 %f889, %f888, 0f437F0000; + cvt.rzi.u32.f32 %r315, %f889; + cvt.u16.u32 %rs32, %r315; + cvt.u16.u32 %rs33, %r314; + cvt.u16.u32 %rs34, %r313; + mov.u16 %rs35, 255; + st.v4.u8 [%rd114], {%rs34, %rs33, %rs32, %rs35}; + +BB0_141: + ret; +} + + -- cgit v1.2.3-freya