diff options
Diffstat (limited to 'VRCSDK3Worlds/Assets/Editor/x64/Bakery/heightfieldTexLODSelect.ptx')
-rw-r--r-- | VRCSDK3Worlds/Assets/Editor/x64/Bakery/heightfieldTexLODSelect.ptx | 680 |
1 files changed, 680 insertions, 0 deletions
diff --git a/VRCSDK3Worlds/Assets/Editor/x64/Bakery/heightfieldTexLODSelect.ptx b/VRCSDK3Worlds/Assets/Editor/x64/Bakery/heightfieldTexLODSelect.ptx new file mode 100644 index 00000000..5f161955 --- /dev/null +++ b/VRCSDK3Worlds/Assets/Editor/x64/Bakery/heightfieldTexLODSelect.ptx @@ -0,0 +1,680 @@ +// +// 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 _Z9intersecti +.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 4 .b8 ray[36]; +.global .align 4 .u32 heightmap; +.global .align 4 .f32 heightmapMaxMip; +.global .align 4 .b8 heightmapMin[12]; +.global .align 4 .b8 heightmapMax[12]; +.global .align 4 .f32 heightmapLMID; +.global .align 16 .b8 heightmapOffsetScaleUV[16]; +.global .align 8 .b8 texCoords[8]; +.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_typeinfo3rayE[8] = {82, 97, 121, 0, 36, 0, 0, 0}; +.global .align 4 .b8 _ZN21rti_internal_typeinfo9heightmapE[8] = {82, 97, 121, 0, 4, 0, 0, 0}; +.global .align 4 .b8 _ZN21rti_internal_typeinfo15heightmapMaxMipE[8] = {82, 97, 121, 0, 4, 0, 0, 0}; +.global .align 4 .b8 _ZN21rti_internal_typeinfo12heightmapMinE[8] = {82, 97, 121, 0, 12, 0, 0, 0}; +.global .align 4 .b8 _ZN21rti_internal_typeinfo12heightmapMaxE[8] = {82, 97, 121, 0, 12, 0, 0, 0}; +.global .align 4 .b8 _ZN21rti_internal_typeinfo13heightmapLMIDE[8] = {82, 97, 121, 0, 4, 0, 0, 0}; +.global .align 4 .b8 _ZN21rti_internal_typeinfo22heightmapOffsetScaleUVE[8] = {82, 97, 121, 0, 16, 0, 0, 0}; +.global .align 4 .b8 _ZN21rti_internal_typeinfo9texCoordsE[8] = {82, 97, 121, 0, 8, 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 4 .b8 _ZN21rti_internal_typename3rayE[4] = {82, 97, 121, 0}; +.global .align 4 .b8 _ZN21rti_internal_typename9heightmapE[4] = {105, 110, 116, 0}; +.global .align 8 .b8 _ZN21rti_internal_typename15heightmapMaxMipE[6] = {102, 108, 111, 97, 116, 0}; +.global .align 8 .b8 _ZN21rti_internal_typename12heightmapMinE[7] = {102, 108, 111, 97, 116, 51, 0}; +.global .align 8 .b8 _ZN21rti_internal_typename12heightmapMaxE[7] = {102, 108, 111, 97, 116, 51, 0}; +.global .align 8 .b8 _ZN21rti_internal_typename13heightmapLMIDE[6] = {102, 108, 111, 97, 116, 0}; +.global .align 8 .b8 _ZN21rti_internal_typename22heightmapOffsetScaleUVE[7] = {102, 108, 111, 97, 116, 52, 0}; +.global .align 8 .b8 _ZN21rti_internal_typename9texCoordsE[7] = {102, 108, 111, 97, 116, 50, 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_typeenum3rayE = 4919; +.global .align 4 .u32 _ZN21rti_internal_typeenum9heightmapE = 4919; +.global .align 4 .u32 _ZN21rti_internal_typeenum15heightmapMaxMipE = 4919; +.global .align 4 .u32 _ZN21rti_internal_typeenum12heightmapMinE = 4919; +.global .align 4 .u32 _ZN21rti_internal_typeenum12heightmapMaxE = 4919; +.global .align 4 .u32 _ZN21rti_internal_typeenum13heightmapLMIDE = 4919; +.global .align 4 .u32 _ZN21rti_internal_typeenum22heightmapOffsetScaleUVE = 4919; +.global .align 4 .u32 _ZN21rti_internal_typeenum9texCoordsE = 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_semantic3rayE[13] = {114, 116, 67, 117, 114, 114, 101, 110, 116, 82, 97, 121, 0}; +.global .align 1 .b8 _ZN21rti_internal_semantic9heightmapE[1]; +.global .align 1 .b8 _ZN21rti_internal_semantic15heightmapMaxMipE[1]; +.global .align 1 .b8 _ZN21rti_internal_semantic12heightmapMinE[1]; +.global .align 1 .b8 _ZN21rti_internal_semantic12heightmapMaxE[1]; +.global .align 1 .b8 _ZN21rti_internal_semantic13heightmapLMIDE[1]; +.global .align 1 .b8 _ZN21rti_internal_semantic22heightmapOffsetScaleUVE[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 _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_annotation3rayE[1]; +.global .align 1 .b8 _ZN23rti_internal_annotation9heightmapE[1]; +.global .align 1 .b8 _ZN23rti_internal_annotation15heightmapMaxMipE[1]; +.global .align 1 .b8 _ZN23rti_internal_annotation12heightmapMinE[1]; +.global .align 1 .b8 _ZN23rti_internal_annotation12heightmapMaxE[1]; +.global .align 1 .b8 _ZN23rti_internal_annotation13heightmapLMIDE[1]; +.global .align 1 .b8 _ZN23rti_internal_annotation22heightmapOffsetScaleUVE[1]; +.global .align 1 .b8 _ZN23rti_internal_annotation9texCoordsE[1]; + +.visible .entry _Z9intersecti( + .param .u32 _Z9intersecti_param_0 +) +{ + .reg .pred %p<57>; + .reg .b16 %rs<9>; + .reg .f32 %f<457>; + .reg .b32 %r<36>; + + + ld.global.f32 %f4, [heightmapMin+8]; + ld.global.f32 %f3, [heightmapMin+4]; + ld.global.f32 %f2, [heightmapMin]; + ld.global.f32 %f1, [heightmapLMID]; + ld.global.f32 %f143, [heightmapMax+8]; + ld.global.f32 %f144, [heightmapMax+4]; + ld.global.f32 %f145, [heightmapMax]; + add.f32 %f146, %f2, %f145; + add.f32 %f147, %f3, %f144; + add.f32 %f148, %f4, %f143; + mul.f32 %f149, %f146, 0f3F000000; + mul.f32 %f150, %f147, 0f3F000000; + mul.f32 %f151, %f148, 0f3F000000; + sub.f32 %f5, %f145, %f2; + sub.f32 %f6, %f144, %f3; + sub.f32 %f7, %f143, %f4; + mul.f32 %f152, %f5, 0f3F000000; + mul.f32 %f153, %f6, 0f3F000000; + mul.f32 %f154, %f7, 0f3F000000; + ld.global.f32 %f8, [ray+12]; + rcp.rn.f32 %f155, %f8; + ld.global.f32 %f9, [ray+16]; + rcp.rn.f32 %f156, %f9; + ld.global.f32 %f10, [ray+20]; + rcp.rn.f32 %f157, %f10; + ld.global.f32 %f11, [ray]; + sub.f32 %f158, %f11, %f149; + ld.global.f32 %f12, [ray+4]; + sub.f32 %f159, %f12, %f150; + ld.global.f32 %f13, [ray+8]; + sub.f32 %f160, %f13, %f151; + mul.f32 %f161, %f155, %f158; + mul.f32 %f162, %f156, %f159; + mul.f32 %f163, %f157, %f160; + abs.f32 %f164, %f155; + abs.f32 %f165, %f156; + abs.f32 %f166, %f157; + mul.f32 %f167, %f152, %f164; + mul.f32 %f168, %f153, %f165; + mul.f32 %f169, %f154, %f166; + neg.f32 %f170, %f161; + neg.f32 %f171, %f162; + neg.f32 %f172, %f163; + sub.f32 %f173, %f170, %f167; + sub.f32 %f174, %f171, %f168; + sub.f32 %f175, %f172, %f169; + sub.f32 %f176, %f167, %f161; + sub.f32 %f177, %f168, %f162; + sub.f32 %f178, %f169, %f163; + max.f32 %f179, %f173, %f174; + max.f32 %f14, %f179, %f175; + min.f32 %f180, %f176, %f177; + min.f32 %f181, %f180, %f178; + setp.gt.f32 %p4, %f14, %f181; + setp.lt.f32 %p5, %f181, 0f00000000; + or.pred %p6, %p4, %p5; + @%p6 bra BB0_32; + + ld.global.v4.f32 {%f182, %f183, %f184, %f185}, [heightmapOffsetScaleUV]; + add.f32 %f186, %f14, 0f3C23D70A; + setp.lt.f32 %p7, %f186, 0f00000000; + selp.f32 %f18, 0f00000000, %f186, %p7; + fma.rn.f32 %f187, %f18, %f8, %f11; + fma.rn.f32 %f188, %f18, %f9, %f12; + fma.rn.f32 %f189, %f18, %f10, %f13; + sub.f32 %f190, %f187, %f2; + sub.f32 %f19, %f188, %f3; + sub.f32 %f191, %f189, %f4; + div.rn.f32 %f437, %f190, %f5; + div.rn.f32 %f439, %f191, %f7; + abs.f32 %f192, %f19; + sub.f32 %f22, %f12, %f3; + setp.geu.f32 %p8, %f192, 0f38D1B717; + @%p8 bra BB0_3; + + abs.f32 %f193, %f22; + setp.gt.f32 %p9, %f193, 0f38D1B717; + @%p9 bra BB0_30; + bra.uni BB0_3; + +BB0_30: + // inline asm + call (%r29), _rt_potential_intersection, (%f18); + // inline asm + setp.eq.s32 %p55, %r29, 0; + @%p55 bra BB0_32; + + mov.u32 %r31, 0; + st.global.u32 [normal+8], %r31; + mov.u32 %r32, 1065353216; + st.global.u32 [normal+4], %r32; + st.global.u32 [normal], %r31; + fma.rn.f32 %f405, %f184, %f437, %f182; + mul.f32 %f406, %f184, %f185; + mul.f32 %f407, %f5, %f7; + div.rn.f32 %f408, %f407, %f406; + fma.rn.f32 %f409, %f1, 0f41200000, %f405; + st.global.v2.f32 [texCoords], {%f409, %f408}; + // inline asm + call (%r30), _rt_report_intersection, (%r31); + // inline asm + bra.uni BB0_32; + +BB0_3: + sub.f32 %f195, %f11, %f2; + div.rn.f32 %f23, %f195, %f5; + div.rn.f32 %f24, %f22, %f6; + sub.f32 %f196, %f13, %f4; + div.rn.f32 %f25, %f196, %f7; + add.f32 %f197, %f437, %f437; + cvt.rmi.f32.f32 %f198, %f197; + mul.f32 %f440, %f198, 0f3F000000; + add.f32 %f199, %f439, %f439; + cvt.rmi.f32.f32 %f200, %f199; + mul.f32 %f441, %f200, 0f3F000000; + div.rn.f32 %f201, %f8, %f5; + div.rn.f32 %f202, %f9, %f6; + mul.f32 %f203, %f202, %f202; + fma.rn.f32 %f204, %f201, %f201, %f203; + div.rn.f32 %f205, %f10, %f7; + fma.rn.f32 %f206, %f205, %f205, %f204; + sqrt.rn.f32 %f207, %f206; + rcp.rn.f32 %f208, %f207; + mul.f32 %f28, %f201, %f208; + mul.f32 %f29, %f202, %f208; + mul.f32 %f30, %f205, %f208; + setp.gt.f32 %p10, %f28, 0f00000000; + mov.f32 %f455, 0f3F800000; + mov.f32 %f410, %f455; + @%p10 bra BB0_5; + + setp.eq.f32 %p11, %f28, 0f00000000; + selp.f32 %f410, 0f00000000, 0fBF800000, %p11; + +BB0_5: + setp.gt.f32 %p12, %f30, 0f00000000; + mov.f32 %f411, %f455; + @%p12 bra BB0_7; + + setp.eq.f32 %p13, %f30, 0f00000000; + selp.f32 %f411, 0f00000000, 0fBF800000, %p13; + +BB0_7: + ld.global.f32 %f443, [heightmapMaxMip]; + abs.f32 %f215, %f30; + mov.f32 %f216, 0f38D1B717; + max.f32 %f36, %f215, %f216; + add.f32 %f217, %f440, 0f3F000000; + setp.lt.f32 %p14, %f410, 0f00000000; + selp.f32 %f218, %f440, %f217, %p14; + sub.f32 %f219, %f218, %f437; + abs.f32 %f220, %f219; + abs.f32 %f221, %f28; + max.f32 %f37, %f221, %f216; + div.rn.f32 %f445, %f220, %f37; + add.f32 %f222, %f441, 0f3F000000; + setp.lt.f32 %p15, %f411, 0f00000000; + selp.f32 %f223, %f441, %f222, %p15; + sub.f32 %f224, %f223, %f439; + abs.f32 %f225, %f224; + div.rn.f32 %f446, %f225, %f36; + rcp.rn.f32 %f226, %f28; + abs.f32 %f227, %f226; + mul.f32 %f447, %f227, 0f3F000000; + rcp.rn.f32 %f228, %f30; + abs.f32 %f229, %f228; + mul.f32 %f448, %f229, 0f3F000000; + mul.f32 %f449, %f410, 0f3F000000; + mul.f32 %f450, %f411, 0f3F000000; + rcp.rn.f32 %f44, %f29; + div.rn.f32 %f438, %f19, %f6; + mov.f32 %f456, 0f00000000; + mov.f32 %f442, 0f3F000000; + mov.f32 %f444, 0f40000000; + mov.u32 %r33, 0; + mov.u32 %r35, %r33; + mov.f32 %f454, %f456; + +BB0_8: + ld.global.u32 %r3, [heightmap]; + setp.eq.f32 %p16, %f443, 0f00000000; + @%p16 bra BB0_12; + bra.uni BB0_9; + +BB0_12: + mov.u32 %r22, 2; + mov.u32 %r23, 0; + mov.f32 %f272, 0f00000000; + // inline asm + call (%f241, %f242, %f243, %f244), _rt_texture_get_level_id, (%r3, %r22, %f440, %f441, %f272, %r23, %f272); + // inline asm + ld.global.u32 %r15, [heightmap]; + add.f32 %f261, %f440, %f442; + // inline asm + call (%f249, %f250, %f251, %f252), _rt_texture_get_level_id, (%r15, %r22, %f261, %f441, %f272, %r23, %f272); + // inline asm + ld.global.u32 %r18, [heightmap]; + add.f32 %f270, %f441, %f442; + // inline asm + call (%f257, %f258, %f259, %f260), _rt_texture_get_level_id, (%r18, %r22, %f261, %f270, %f272, %r23, %f272); + // inline asm + ld.global.u32 %r21, [heightmap]; + // inline asm + call (%f265, %f266, %f267, %f268), _rt_texture_get_level_id, (%r21, %r22, %f440, %f270, %f272, %r23, %f272); + // inline asm + sub.f32 %f273, %f261, %f440; + sub.f32 %f274, %f257, %f241; + sub.f32 %f275, %f270, %f441; + sub.f32 %f276, %f241, %f265; + sub.f32 %f277, %f441, %f270; + mul.f32 %f278, %f275, %f276; + mul.f32 %f279, %f277, %f274; + sub.f32 %f454, %f278, %f279; + mul.f32 %f71, %f277, %f273; + sub.f32 %f280, %f440, %f440; + mul.f32 %f281, %f280, %f275; + sub.f32 %f455, %f71, %f281; + mul.f32 %f282, %f280, %f274; + mul.f32 %f283, %f273, %f276; + sub.f32 %f456, %f282, %f283; + mul.f32 %f284, %f28, %f454; + fma.rn.f32 %f285, %f29, %f455, %f284; + fma.rn.f32 %f286, %f30, %f456, %f285; + rcp.rn.f32 %f287, %f286; + sub.f32 %f74, %f241, %f24; + sub.f32 %f288, %f440, %f23; + mul.f32 %f289, %f288, %f287; + mul.f32 %f290, %f74, %f287; + sub.f32 %f291, %f441, %f25; + mul.f32 %f292, %f291, %f287; + mul.f32 %f293, %f29, %f292; + mul.f32 %f294, %f30, %f290; + sub.f32 %f295, %f293, %f294; + mul.f32 %f296, %f30, %f289; + mul.f32 %f297, %f28, %f292; + sub.f32 %f298, %f296, %f297; + mul.f32 %f299, %f28, %f290; + mul.f32 %f300, %f29, %f289; + sub.f32 %f301, %f299, %f300; + mul.f32 %f302, %f276, %f298; + fma.rn.f32 %f303, %f280, %f295, %f302; + fma.rn.f32 %f304, %f277, %f301, %f303; + mul.f32 %f305, %f274, %f298; + fma.rn.f32 %f306, %f273, %f295, %f305; + fma.rn.f32 %f307, %f275, %f301, %f306; + mul.f32 %f308, %f289, %f454; + fma.rn.f32 %f309, %f455, %f290, %f308; + fma.rn.f32 %f75, %f292, %f456, %f309; + setp.gt.f32 %p22, %f75, 0f3456BF95; + setp.ge.f32 %p23, %f304, 0f00000000; + and.pred %p24, %p22, %p23; + setp.ge.f32 %p25, %f307, 0f00000000; + and.pred %p26, %p24, %p25; + add.f32 %f310, %f304, %f307; + setp.le.f32 %p27, %f310, 0f3F800000; + and.pred %p28, %p26, %p27; + @%p28 bra BB0_26; + + sub.f32 %f311, %f249, %f241; + sub.f32 %f313, %f440, %f261; + sub.f32 %f314, %f241, %f257; + sub.f32 %f315, %f441, %f441; + mul.f32 %f316, %f315, %f314; + mul.f32 %f319, %f277, %f311; + sub.f32 %f454, %f316, %f319; + mul.f32 %f320, %f315, %f313; + sub.f32 %f455, %f71, %f320; + mul.f32 %f321, %f313, %f311; + mul.f32 %f323, %f273, %f314; + sub.f32 %f456, %f321, %f323; + mul.f32 %f324, %f28, %f454; + fma.rn.f32 %f325, %f29, %f455, %f324; + fma.rn.f32 %f326, %f30, %f456, %f325; + rcp.rn.f32 %f327, %f326; + mul.f32 %f329, %f288, %f327; + mul.f32 %f330, %f74, %f327; + mul.f32 %f332, %f291, %f327; + mul.f32 %f333, %f29, %f332; + mul.f32 %f334, %f30, %f330; + sub.f32 %f335, %f333, %f334; + mul.f32 %f336, %f30, %f329; + mul.f32 %f337, %f28, %f332; + sub.f32 %f338, %f336, %f337; + mul.f32 %f339, %f28, %f330; + mul.f32 %f340, %f29, %f329; + sub.f32 %f341, %f339, %f340; + mul.f32 %f342, %f314, %f338; + fma.rn.f32 %f343, %f313, %f335, %f342; + fma.rn.f32 %f344, %f277, %f341, %f343; + mul.f32 %f345, %f311, %f338; + fma.rn.f32 %f346, %f273, %f335, %f345; + fma.rn.f32 %f347, %f315, %f341, %f346; + mul.f32 %f348, %f455, %f330; + fma.rn.f32 %f349, %f454, %f329, %f348; + fma.rn.f32 %f79, %f456, %f332, %f349; + setp.gt.f32 %p29, %f79, 0f3456BF95; + setp.ge.f32 %p30, %f344, 0f00000000; + and.pred %p31, %p29, %p30; + setp.ge.f32 %p32, %f347, 0f00000000; + and.pred %p33, %p31, %p32; + add.f32 %f350, %f344, %f347; + setp.le.f32 %p34, %f350, 0f3F800000; + and.pred %p35, %p33, %p34; + selp.u16 %rs7, 1, 0, %p35; + @%p35 bra BB0_25; + bra.uni BB0_14; + +BB0_9: + setp.lt.f32 %p17, %f29, 0f00000000; + mov.u32 %r10, 2; + mov.f32 %f236, 0f00000000; + mov.u32 %r11, 0; + // inline asm + call (%f230, %f231, %f232, %f233), _rt_texture_get_level_id, (%r3, %r10, %f440, %f441, %f236, %r11, %f443); + // inline asm + setp.gt.f32 %p56, %f230, %f438; + setp.leu.f32 %p18, %f230, %f438; + and.pred %p19, %p17, %p18; + @!%p19 bra BB0_11; + bra.uni BB0_10; + +BB0_10: + sub.f32 %f238, %f438, %f230; + mul.f32 %f239, %f44, %f238; + neg.f32 %f240, %f239; + setp.gt.f32 %p20, %f445, %f240; + setp.gt.f32 %p21, %f446, %f240; + and.pred %p56, %p20, %p21; + +BB0_11: + selp.u16 %rs7, 1, 0, %p56; + +BB0_14: + setp.eq.s16 %p36, %rs7, 0; + @%p36 bra BB0_16; + bra.uni BB0_15; + +BB0_16: + setp.eq.s32 %p39, %r35, 4; + @%p39 bra BB0_22; + bra.uni BB0_17; + +BB0_22: + add.f32 %f443, %f443, 0f3F800000; + mul.f32 %f444, %f444, 0f3F000000; + mul.f32 %f363, %f437, %f444; + cvt.rmi.f32.f32 %f364, %f363; + div.rn.f32 %f440, %f364, %f444; + mul.f32 %f365, %f439, %f444; + cvt.rmi.f32.f32 %f366, %f365; + div.rn.f32 %f441, %f366, %f444; + add.f32 %f442, %f442, %f442; + add.f32 %f367, %f442, %f440; + selp.f32 %f368, %f440, %f367, %p14; + sub.f32 %f369, %f368, %f437; + abs.f32 %f370, %f369; + div.rn.f32 %f445, %f370, %f37; + add.f32 %f371, %f442, %f441; + selp.f32 %f372, %f441, %f371, %p15; + sub.f32 %f373, %f372, %f439; + abs.f32 %f374, %f373; + div.rn.f32 %f446, %f374, %f36; + add.f32 %f447, %f447, %f447; + add.f32 %f448, %f448, %f448; + add.f32 %f449, %f449, %f449; + add.f32 %f450, %f450, %f450; + bra.uni BB0_23; + +BB0_15: + add.f32 %f443, %f443, 0fBF800000; + add.f32 %f444, %f444, %f444; + mul.f32 %f351, %f437, %f444; + cvt.rmi.f32.f32 %f352, %f351; + div.rn.f32 %f440, %f352, %f444; + mul.f32 %f353, %f439, %f444; + cvt.rmi.f32.f32 %f354, %f353; + div.rn.f32 %f441, %f354, %f444; + mul.f32 %f442, %f442, 0f3F000000; + add.f32 %f355, %f442, %f440; + selp.f32 %f356, %f440, %f355, %p14; + sub.f32 %f357, %f356, %f437; + abs.f32 %f358, %f357; + div.rn.f32 %f445, %f358, %f37; + add.f32 %f359, %f442, %f441; + selp.f32 %f360, %f441, %f359, %p15; + sub.f32 %f361, %f360, %f439; + abs.f32 %f362, %f361; + div.rn.f32 %f446, %f362, %f36; + mul.f32 %f447, %f447, 0f3F000000; + mul.f32 %f448, %f448, 0f3F000000; + mul.f32 %f449, %f449, 0f3F000000; + mul.f32 %f450, %f450, 0f3F000000; + +BB0_23: + mov.u32 %r35, 0; + bra.uni BB0_24; + +BB0_17: + setp.lt.f32 %p40, %f445, %f446; + @%p40 bra BB0_19; + bra.uni BB0_18; + +BB0_19: + sub.f32 %f436, %f446, %f445; + add.f32 %f440, %f440, %f449; + mov.f32 %f446, %f445; + mov.f32 %f445, %f447; + bra.uni BB0_20; + +BB0_18: + sub.f32 %f445, %f445, %f446; + add.f32 %f441, %f441, %f450; + mov.f32 %f436, %f448; + +BB0_20: + fma.rn.f32 %f439, %f30, %f446, %f439; + fma.rn.f32 %f438, %f29, %f446, %f438; + fma.rn.f32 %f437, %f28, %f446, %f437; + setp.ge.f32 %p41, %f437, 0f00000000; + setp.ge.f32 %p42, %f439, 0f00000000; + and.pred %p43, %p41, %p42; + setp.lt.f32 %p44, %f437, 0f3F800000; + and.pred %p45, %p43, %p44; + setp.lt.f32 %p46, %f439, 0f3F800000; + and.pred %p47, %p45, %p46; + setp.lt.f32 %p48, %f438, 0f3F800000; + and.pred %p49, %p47, %p48; + mov.u16 %rs8, 0; + @!%p49 bra BB0_27; + bra.uni BB0_21; + +BB0_21: + add.s32 %r35, %r35, 1; + mov.f32 %f446, %f436; + +BB0_24: + add.s32 %r33, %r33, 1; + setp.lt.s32 %p52, %r33, 512; + mov.u16 %rs8, %rs7; + @%p52 bra BB0_8; + bra.uni BB0_27; + +BB0_26: + fma.rn.f32 %f437, %f28, %f75, %f23; + fma.rn.f32 %f438, %f29, %f75, %f24; + fma.rn.f32 %f439, %f30, %f75, %f25; + mov.u16 %rs8, 1; + bra.uni BB0_27; + +BB0_25: + fma.rn.f32 %f437, %f28, %f79, %f23; + fma.rn.f32 %f438, %f29, %f79, %f24; + fma.rn.f32 %f439, %f30, %f79, %f25; + mov.u16 %rs8, %rs7; + +BB0_27: + setp.eq.s16 %p53, %rs8, 0; + @%p53 bra BB0_32; + + fma.rn.f32 %f376, %f5, %f437, %f2; + fma.rn.f32 %f377, %f6, %f438, %f3; + fma.rn.f32 %f378, %f7, %f439, %f4; + ld.global.f32 %f379, [ray]; + sub.f32 %f380, %f376, %f379; + ld.global.f32 %f381, [ray+4]; + sub.f32 %f382, %f377, %f381; + ld.global.f32 %f383, [ray+8]; + sub.f32 %f384, %f378, %f383; + mul.f32 %f385, %f382, %f382; + fma.rn.f32 %f386, %f380, %f380, %f385; + fma.rn.f32 %f387, %f384, %f384, %f386; + sqrt.rn.f32 %f375, %f387; + // inline asm + call (%r26), _rt_potential_intersection, (%f375); + // inline asm + setp.eq.s32 %p54, %r26, 0; + @%p54 bra BB0_32; + + mul.f32 %f388, %f455, %f455; + fma.rn.f32 %f389, %f454, %f454, %f388; + fma.rn.f32 %f390, %f456, %f456, %f389; + sqrt.rn.f32 %f391, %f390; + rcp.rn.f32 %f392, %f391; + mul.f32 %f393, %f454, %f392; + neg.f32 %f394, %f393; + mul.f32 %f395, %f455, %f392; + neg.f32 %f396, %f395; + mul.f32 %f397, %f456, %f392; + neg.f32 %f398, %f397; + st.global.f32 [normal], %f394; + st.global.f32 [normal+4], %f396; + st.global.f32 [normal+8], %f398; + fma.rn.f32 %f399, %f184, %f437, %f182; + mul.f32 %f400, %f184, %f185; + mul.f32 %f401, %f5, %f7; + div.rn.f32 %f402, %f401, %f400; + fma.rn.f32 %f403, %f1, 0f41200000, %f399; + st.global.v2.f32 [texCoords], {%f403, %f402}; + mov.u32 %r28, 0; + // inline asm + call (%r27), _rt_report_intersection, (%r28); + // inline asm + +BB0_32: + ret; +} + + // .globl _Z4bboxiPN5optix4AabbE +.visible .entry _Z4bboxiPN5optix4AabbE( + .param .u32 _Z4bboxiPN5optix4AabbE_param_0, + .param .u64 _Z4bboxiPN5optix4AabbE_param_1 +) +{ + .reg .f32 %f<7>; + .reg .b64 %rd<3>; + + + ld.param.u64 %rd1, [_Z4bboxiPN5optix4AabbE_param_1]; + cvta.to.global.u64 %rd2, %rd1; + ld.global.f32 %f1, [heightmapMin]; + ld.global.f32 %f2, [heightmapMin+4]; + ld.global.f32 %f3, [heightmapMin+8]; + st.global.f32 [%rd2+8], %f3; + st.global.f32 [%rd2+4], %f2; + st.global.f32 [%rd2], %f1; + ld.global.f32 %f4, [heightmapMax]; + ld.global.f32 %f5, [heightmapMax+4]; + ld.global.f32 %f6, [heightmapMax+8]; + st.global.f32 [%rd2+20], %f6; + st.global.f32 [%rd2+16], %f5; + st.global.f32 [%rd2+12], %f4; + ret; +} + + |