OracleStudio/PrinceOfGlory
5
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity
Files
de88d2d6a11e38f586c4560cee12978520412e61
PrinceOfGlory/Assets/Editor/x64/Bakery/lmPreviewHDR.ptx
kridoo 6e91a0c7f0 111
2025-09-15 17:32:08 +08:00

2467 lines
73 KiB
Plaintext
Raw Blame History

//
// 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 accumBufferAdd[1];
.global .align 1 .b8 accumBufferLerp[1];
.global .align 1 .b8 GBufferPos[1];
.global .align 1 .b8 GBufferNormal[1];
.global .align 1 .b8 GBufferAtten[1];
.global .align 1 .b8 colorOutput[1];
.global .align 1 .b8 rnd_seeds[1];
.global .align 1 .b8 globalLights[1];
.global .align 1 .b8 localLights[1];
.global .align 1 .b8 localLightIndices[1];
.global .align 4 .u32 tileCubemap;
.global .align 4 .b8 cameraRht[12];
.global .align 4 .b8 cameraUp[12];
.global .align 4 .b8 cameraFwd[12];
.global .align 4 .b8 cameraPos[12];
.global .align 4 .f32 cameraFov;
.global .align 4 .f32 cameraExp;
.global .align 4 .f32 integration;
.global .align 4 .u32 finalBounceFlag;
.global .align 4 .u32 bounceFlag;
.global .align 16 .b8 rndOffset[16];
.global .align 4 .u32 firstSkylight;
.global .align 4 .f32 backFaceWeight;
.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_typeinfo11tileCubemapE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
.global .align 4 .b8 _ZN21rti_internal_typeinfo9cameraRhtE[8] = {82, 97, 121, 0, 12, 0, 0, 0};
.global .align 4 .b8 _ZN21rti_internal_typeinfo8cameraUpE[8] = {82, 97, 121, 0, 12, 0, 0, 0};
.global .align 4 .b8 _ZN21rti_internal_typeinfo9cameraFwdE[8] = {82, 97, 121, 0, 12, 0, 0, 0};
.global .align 4 .b8 _ZN21rti_internal_typeinfo9cameraPosE[8] = {82, 97, 121, 0, 12, 0, 0, 0};
.global .align 4 .b8 _ZN21rti_internal_typeinfo9cameraFovE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
.global .align 4 .b8 _ZN21rti_internal_typeinfo9cameraExpE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
.global .align 4 .b8 _ZN21rti_internal_typeinfo11integrationE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
.global .align 4 .b8 _ZN21rti_internal_typeinfo15finalBounceFlagE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
.global .align 4 .b8 _ZN21rti_internal_typeinfo10bounceFlagE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
.global .align 4 .b8 _ZN21rti_internal_typeinfo9rndOffsetE[8] = {82, 97, 121, 0, 16, 0, 0, 0};
.global .align 4 .b8 _ZN21rti_internal_typeinfo13firstSkylightE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
.global .align 4 .b8 _ZN21rti_internal_typeinfo14backFaceWeightE[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 4 .b8 _ZN21rti_internal_typename11tileCubemapE[4] = {105, 110, 116, 0};
.global .align 8 .b8 _ZN21rti_internal_typename9cameraRhtE[7] = {102, 108, 111, 97, 116, 51, 0};
.global .align 8 .b8 _ZN21rti_internal_typename8cameraUpE[7] = {102, 108, 111, 97, 116, 51, 0};
.global .align 8 .b8 _ZN21rti_internal_typename9cameraFwdE[7] = {102, 108, 111, 97, 116, 51, 0};
.global .align 8 .b8 _ZN21rti_internal_typename9cameraPosE[7] = {102, 108, 111, 97, 116, 51, 0};
.global .align 8 .b8 _ZN21rti_internal_typename9cameraFovE[6] = {102, 108, 111, 97, 116, 0};
.global .align 8 .b8 _ZN21rti_internal_typename9cameraExpE[6] = {102, 108, 111, 97, 116, 0};
.global .align 8 .b8 _ZN21rti_internal_typename11integrationE[6] = {102, 108, 111, 97, 116, 0};
.global .align 4 .b8 _ZN21rti_internal_typename15finalBounceFlagE[4] = {105, 110, 116, 0};
.global .align 4 .b8 _ZN21rti_internal_typename10bounceFlagE[4] = {105, 110, 116, 0};
.global .align 8 .b8 _ZN21rti_internal_typename9rndOffsetE[6] = {117, 105, 110, 116, 52, 0};
.global .align 4 .b8 _ZN21rti_internal_typename13firstSkylightE[4] = {105, 110, 116, 0};
.global .align 8 .b8 _ZN21rti_internal_typename14backFaceWeightE[6] = {102, 108, 111, 97, 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_typeenum11tileCubemapE = 4919;
.global .align 4 .u32 _ZN21rti_internal_typeenum9cameraRhtE = 4919;
.global .align 4 .u32 _ZN21rti_internal_typeenum8cameraUpE = 4919;
.global .align 4 .u32 _ZN21rti_internal_typeenum9cameraFwdE = 4919;
.global .align 4 .u32 _ZN21rti_internal_typeenum9cameraPosE = 4919;
.global .align 4 .u32 _ZN21rti_internal_typeenum9cameraFovE = 4919;
.global .align 4 .u32 _ZN21rti_internal_typeenum9cameraExpE = 4919;
.global .align 4 .u32 _ZN21rti_internal_typeenum11integrationE = 4919;
.global .align 4 .u32 _ZN21rti_internal_typeenum15finalBounceFlagE = 4919;
.global .align 4 .u32 _ZN21rti_internal_typeenum10bounceFlagE = 4919;
.global .align 4 .u32 _ZN21rti_internal_typeenum9rndOffsetE = 4919;
.global .align 4 .u32 _ZN21rti_internal_typeenum13firstSkylightE = 4919;
.global .align 4 .u32 _ZN21rti_internal_typeenum14backFaceWeightE = 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_semantic11tileCubemapE[1];
.global .align 1 .b8 _ZN21rti_internal_semantic9cameraRhtE[1];
.global .align 1 .b8 _ZN21rti_internal_semantic8cameraUpE[1];
.global .align 1 .b8 _ZN21rti_internal_semantic9cameraFwdE[1];
.global .align 1 .b8 _ZN21rti_internal_semantic9cameraPosE[1];
.global .align 1 .b8 _ZN21rti_internal_semantic9cameraFovE[1];
.global .align 1 .b8 _ZN21rti_internal_semantic9cameraExpE[1];
.global .align 1 .b8 _ZN21rti_internal_semantic11integrationE[1];
.global .align 1 .b8 _ZN21rti_internal_semantic15finalBounceFlagE[1];
.global .align 1 .b8 _ZN21rti_internal_semantic10bounceFlagE[1];
.global .align 1 .b8 _ZN21rti_internal_semantic9rndOffsetE[1];
.global .align 1 .b8 _ZN21rti_internal_semantic13firstSkylightE[1];
.global .align 1 .b8 _ZN21rti_internal_semantic14backFaceWeightE[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_annotation11tileCubemapE[1];
.global .align 1 .b8 _ZN23rti_internal_annotation9cameraRhtE[1];
.global .align 1 .b8 _ZN23rti_internal_annotation8cameraUpE[1];
.global .align 1 .b8 _ZN23rti_internal_annotation9cameraFwdE[1];
.global .align 1 .b8 _ZN23rti_internal_annotation9cameraPosE[1];
.global .align 1 .b8 _ZN23rti_internal_annotation9cameraFovE[1];
.global .align 1 .b8 _ZN23rti_internal_annotation9cameraExpE[1];
.global .align 1 .b8 _ZN23rti_internal_annotation11integrationE[1];
.global .align 1 .b8 _ZN23rti_internal_annotation15finalBounceFlagE[1];
.global .align 1 .b8 _ZN23rti_internal_annotation10bounceFlagE[1];
.global .align 1 .b8 _ZN23rti_internal_annotation9rndOffsetE[1];
.global .align 1 .b8 _ZN23rti_internal_annotation13firstSkylightE[1];
.global .align 1 .b8 _ZN23rti_internal_annotation14backFaceWeightE[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[60];
.reg .b64 %SP;
.reg .b64 %SPL;
.reg .pred %p<146>;
.reg .b16 %rs<14>;
.reg .f32 %f<1246>;
.reg .b32 %r<747>;
.reg .b64 %rd<248>;
mov.u64 %rd247, __local_depot0;
cvta.local.u64 %SP, %rd247;
ld.global.v2.u32 {%r226, %r227}, [pixelID];
ld.global.v4.u32 {%r229, %r230, %r231, %r232}, [rndOffset];
add.s32 %r234, %r229, %r226;
and.b32 %r235, %r234, 255;
add.s32 %r238, %r230, %r227;
and.b32 %r239, %r238, 255;
add.s32 %r241, %r231, %r226;
and.b32 %r242, %r241, 255;
add.s32 %r244, %r232, %r227;
and.b32 %r245, %r244, 255;
cvt.u64.u32 %rd56, %r235;
cvt.u64.u32 %rd57, %r239;
mov.u64 %rd72, rnd_seeds;
cvta.global.u64 %rd43, %rd72;
mov.u32 %r224, 2;
mov.u32 %r225, 3;
mov.u64 %rd240, 0;
// inline asm
call (%rd42), _rt_buffer_get_64, (%rd43, %r224, %r225, %rd56, %rd57, %rd240, %rd240);
// inline asm
ld.u8 %rs4, [%rd42];
cvt.rn.f32.u16 %f374, %rs4;
div.rn.f32 %f1, %f374, 0f437F0000;
// inline asm
call (%rd48), _rt_buffer_get_64, (%rd43, %r224, %r225, %rd56, %rd57, %rd240, %rd240);
// inline asm
ld.u8 %rs5, [%rd48+1];
cvt.rn.f32.u16 %f375, %rs5;
div.rn.f32 %f2, %f375, 0f437F0000;
// inline asm
call (%rd54), _rt_buffer_get_64, (%rd43, %r224, %r225, %rd56, %rd57, %rd240, %rd240);
// inline asm
ld.u8 %rs6, [%rd54+2];
cvt.rn.f32.u16 %f376, %rs6;
div.rn.f32 %f3, %f376, 0f437F0000;
cvt.u64.u32 %rd68, %r242;
cvt.u64.u32 %rd69, %r245;
// inline asm
call (%rd60), _rt_buffer_get_64, (%rd43, %r224, %r225, %rd68, %rd69, %rd240, %rd240);
// inline asm
ld.u8 %rs1, [%rd60];
// inline asm
call (%rd66), _rt_buffer_get_64, (%rd43, %r224, %r225, %rd68, %rd69, %rd240, %rd240);
// inline asm
ld.global.u32 %r246, [bounceFlag];
setp.eq.s32 %p3, %r246, 0;
@%p3 bra BB0_48;
ld.u8 %rs7, [%rd66+1];
cvt.rn.f32.u16 %f377, %rs7;
div.rn.f32 %f378, %f377, 0f437F0000;
cvt.rn.f32.u16 %f379, %rs1;
div.rn.f32 %f380, %f379, 0f437F0000;
ld.global.v2.u32 {%r251, %r252}, [pixelID];
cvt.u64.u32 %rd75, %r251;
cvt.u64.u32 %rd76, %r252;
mov.u64 %rd85, GBufferPos;
cvta.global.u64 %rd74, %rd85;
mov.u32 %r250, 12;
// inline asm
call (%rd73), _rt_buffer_get_64, (%rd74, %r224, %r250, %rd75, %rd76, %rd240, %rd240);
// inline asm
ld.f32 %f1172, [%rd73+8];
ld.f32 %f1171, [%rd73+4];
ld.f32 %f1170, [%rd73];
ld.global.v2.u32 {%r255, %r256}, [pixelID];
cvt.u64.u32 %rd81, %r255;
cvt.u64.u32 %rd82, %r256;
mov.u64 %rd86, GBufferNormal;
cvta.global.u64 %rd80, %rd86;
// inline asm
call (%rd79), _rt_buffer_get_64, (%rd80, %r224, %r250, %rd81, %rd82, %rd240, %rd240);
// inline asm
ld.f32 %f9, [%rd79+4];
ld.f32 %f8, [%rd79+8];
ld.f32 %f7, [%rd79];
abs.f32 %f381, %f7;
abs.f32 %f382, %f8;
setp.gt.f32 %p4, %f381, %f382;
neg.f32 %f383, %f9;
neg.f32 %f384, %f8;
selp.f32 %f385, %f383, 0f00000000, %p4;
selp.f32 %f386, %f7, %f384, %p4;
selp.f32 %f387, 0f00000000, %f9, %p4;
mul.f32 %f388, %f386, %f386;
fma.rn.f32 %f389, %f385, %f385, %f388;
fma.rn.f32 %f390, %f387, %f387, %f389;
sqrt.rn.f32 %f391, %f390;
rcp.rn.f32 %f392, %f391;
mul.f32 %f10, %f385, %f392;
mul.f32 %f11, %f386, %f392;
mul.f32 %f12, %f387, %f392;
sqrt.rn.f32 %f13, %f380;
mul.f32 %f1161, %f378, 0f40C90FDB;
add.u64 %rd87, %SP, 0;
cvta.to.local.u64 %rd2, %rd87;
abs.f32 %f15, %f1161;
setp.neu.f32 %p5, %f15, 0f7F800000;
mov.f32 %f1155, %f1161;
@%p5 bra BB0_3;
mov.f32 %f393, 0f00000000;
mul.rn.f32 %f1155, %f1161, %f393;
BB0_3:
mul.f32 %f394, %f1155, 0f3F22F983;
cvt.rni.s32.f32 %r696, %f394;
cvt.rn.f32.s32 %f395, %r696;
neg.f32 %f396, %f395;
mov.f32 %f397, 0f3FC90FDA;
fma.rn.f32 %f398, %f396, %f397, %f1155;
mov.f32 %f399, 0f33A22168;
fma.rn.f32 %f400, %f396, %f399, %f398;
mov.f32 %f401, 0f27C234C5;
fma.rn.f32 %f1156, %f396, %f401, %f400;
abs.f32 %f402, %f1155;
add.s64 %rd3, %rd2, 24;
setp.leu.f32 %p6, %f402, 0f47CE4780;
@%p6 bra BB0_14;
mov.b32 %r2, %f1155;
shr.u32 %r3, %r2, 23;
shl.b32 %r261, %r2, 8;
or.b32 %r4, %r261, -2147483648;
mov.u32 %r688, 0;
mov.u64 %rd235, __cudart_i2opi_f;
mov.u32 %r687, -6;
mov.u64 %rd236, %rd2;
BB0_5:
.pragma "nounroll";
ld.const.u32 %r264, [%rd235];
// inline asm
{
mad.lo.cc.u32 %r262, %r264, %r4, %r688;
madc.hi.u32 %r688, %r264, %r4, 0;
}
// inline asm
st.local.u32 [%rd236], %r262;
add.s64 %rd236, %rd236, 4;
add.s64 %rd235, %rd235, 4;
add.s32 %r687, %r687, 1;
setp.ne.s32 %p7, %r687, 0;
@%p7 bra BB0_5;
and.b32 %r267, %r3, 255;
add.s32 %r268, %r267, -128;
shr.u32 %r269, %r268, 5;
and.b32 %r9, %r2, -2147483648;
st.local.u32 [%rd3], %r688;
mov.u32 %r270, 6;
sub.s32 %r271, %r270, %r269;
mul.wide.s32 %rd89, %r271, 4;
add.s64 %rd8, %rd2, %rd89;
ld.local.u32 %r689, [%rd8];
ld.local.u32 %r690, [%rd8+-4];
and.b32 %r12, %r3, 31;
setp.eq.s32 %p8, %r12, 0;
@%p8 bra BB0_8;
mov.u32 %r272, 32;
sub.s32 %r273, %r272, %r12;
shr.u32 %r274, %r690, %r273;
shl.b32 %r275, %r689, %r12;
add.s32 %r689, %r274, %r275;
ld.local.u32 %r276, [%rd8+-8];
shr.u32 %r277, %r276, %r273;
shl.b32 %r278, %r690, %r12;
add.s32 %r690, %r277, %r278;
BB0_8:
shr.u32 %r279, %r690, 30;
shl.b32 %r280, %r689, 2;
add.s32 %r691, %r279, %r280;
shl.b32 %r18, %r690, 2;
shr.u32 %r281, %r691, 31;
shr.u32 %r282, %r689, 30;
add.s32 %r19, %r281, %r282;
setp.eq.s32 %p9, %r281, 0;
@%p9 bra BB0_9;
not.b32 %r283, %r691;
neg.s32 %r693, %r18;
setp.eq.s32 %p10, %r18, 0;
selp.u32 %r284, 1, 0, %p10;
add.s32 %r691, %r284, %r283;
xor.b32 %r692, %r9, -2147483648;
bra.uni BB0_11;
BB0_48:
ld.global.f32 %f1170, [cameraPos];
ld.global.f32 %f1171, [cameraPos+4];
ld.global.f32 %f1172, [cameraPos+8];
ld.global.v2.u32 {%r351, %r352}, [resolution];
cvt.rn.f32.u32 %f464, %r351;
cvt.rn.f32.u32 %f465, %r352;
div.rn.f32 %f466, %f464, %f465;
ld.global.v2.u32 {%r355, %r356}, [pixelID];
cvt.rn.f32.u32 %f467, %r355;
add.f32 %f468, %f1, %f467;
div.rn.f32 %f469, %f468, %f464;
cvt.rn.f32.u32 %f470, %r356;
add.f32 %f471, %f2, %f470;
div.rn.f32 %f472, %f471, %f465;
fma.rn.f32 %f473, %f469, 0f40000000, 0fBF800000;
fma.rn.f32 %f474, %f472, 0f40000000, 0fBF800000;
ld.global.f32 %f475, [cameraFov];
mul.f32 %f476, %f466, %f475;
ld.global.f32 %f477, [cameraRht];
mul.f32 %f478, %f473, %f477;
ld.global.f32 %f479, [cameraRht+4];
mul.f32 %f480, %f473, %f479;
ld.global.f32 %f481, [cameraRht+8];
mul.f32 %f482, %f473, %f481;
ld.global.f32 %f483, [cameraFwd];
fma.rn.f32 %f484, %f476, %f478, %f483;
ld.global.f32 %f485, [cameraFwd+4];
fma.rn.f32 %f486, %f476, %f480, %f485;
ld.global.f32 %f487, [cameraFwd+8];
fma.rn.f32 %f488, %f476, %f482, %f487;
ld.global.f32 %f489, [cameraUp];
mul.f32 %f490, %f474, %f489;
ld.global.f32 %f491, [cameraUp+4];
mul.f32 %f492, %f474, %f491;
ld.global.f32 %f493, [cameraUp+8];
mul.f32 %f494, %f474, %f493;
fma.rn.f32 %f495, %f475, %f490, %f484;
fma.rn.f32 %f496, %f475, %f492, %f486;
fma.rn.f32 %f497, %f475, %f494, %f488;
mul.f32 %f498, %f496, %f496;
fma.rn.f32 %f499, %f495, %f495, %f498;
fma.rn.f32 %f500, %f497, %f497, %f499;
sqrt.rn.f32 %f501, %f500;
rcp.rn.f32 %f502, %f501;
mul.f32 %f62, %f495, %f502;
mul.f32 %f63, %f496, %f502;
mul.f32 %f64, %f502, %f497;
bra.uni BB0_49;
BB0_9:
mov.u32 %r692, %r9;
mov.u32 %r693, %r18;
BB0_11:
clz.b32 %r695, %r691;
setp.eq.s32 %p11, %r695, 0;
shl.b32 %r285, %r691, %r695;
mov.u32 %r286, 32;
sub.s32 %r287, %r286, %r695;
shr.u32 %r288, %r693, %r287;
add.s32 %r289, %r288, %r285;
selp.b32 %r27, %r691, %r289, %p11;
mov.u32 %r290, -921707870;
mul.hi.u32 %r694, %r27, %r290;
setp.eq.s32 %p12, %r9, 0;
neg.s32 %r291, %r19;
selp.b32 %r696, %r19, %r291, %p12;
setp.lt.s32 %p13, %r694, 1;
@%p13 bra BB0_13;
mul.lo.s32 %r292, %r27, -921707870;
shr.u32 %r293, %r292, 31;
shl.b32 %r294, %r694, 1;
add.s32 %r694, %r293, %r294;
add.s32 %r695, %r695, 1;
BB0_13:
mov.u32 %r295, 126;
sub.s32 %r296, %r295, %r695;
shl.b32 %r297, %r296, 23;
add.s32 %r298, %r694, 1;
shr.u32 %r299, %r298, 7;
add.s32 %r300, %r299, 1;
shr.u32 %r301, %r300, 1;
add.s32 %r302, %r301, %r297;
or.b32 %r303, %r302, %r692;
mov.b32 %f1156, %r303;
BB0_14:
mul.rn.f32 %f21, %f1156, %f1156;
add.s32 %r35, %r696, 1;
and.b32 %r36, %r35, 1;
setp.eq.s32 %p14, %r36, 0;
@%p14 bra BB0_16;
mov.f32 %f403, 0fBAB6061A;
mov.f32 %f404, 0f37CCF5CE;
fma.rn.f32 %f1157, %f404, %f21, %f403;
bra.uni BB0_17;
BB0_16:
mov.f32 %f405, 0f3C08839E;
mov.f32 %f406, 0fB94CA1F9;
fma.rn.f32 %f1157, %f406, %f21, %f405;
BB0_17:
@%p14 bra BB0_19;
mov.f32 %f407, 0f3D2AAAA5;
fma.rn.f32 %f408, %f1157, %f21, %f407;
mov.f32 %f409, 0fBF000000;
fma.rn.f32 %f1158, %f408, %f21, %f409;
bra.uni BB0_20;
BB0_19:
mov.f32 %f410, 0fBE2AAAA3;
fma.rn.f32 %f411, %f1157, %f21, %f410;
mov.f32 %f412, 0f00000000;
fma.rn.f32 %f1158, %f411, %f21, %f412;
BB0_20:
fma.rn.f32 %f1159, %f1158, %f1156, %f1156;
@%p14 bra BB0_22;
mov.f32 %f413, 0f3F800000;
fma.rn.f32 %f1159, %f1158, %f21, %f413;
BB0_22:
and.b32 %r304, %r35, 2;
setp.eq.s32 %p17, %r304, 0;
@%p17 bra BB0_24;
mov.f32 %f414, 0f00000000;
mov.f32 %f415, 0fBF800000;
fma.rn.f32 %f1159, %f1159, %f415, %f414;
BB0_24:
@%p5 bra BB0_26;
mov.f32 %f416, 0f00000000;
mul.rn.f32 %f1161, %f1161, %f416;
BB0_26:
mul.f32 %f417, %f1161, 0f3F22F983;
cvt.rni.s32.f32 %r706, %f417;
cvt.rn.f32.s32 %f418, %r706;
neg.f32 %f419, %f418;
fma.rn.f32 %f421, %f419, %f397, %f1161;
fma.rn.f32 %f423, %f419, %f399, %f421;
fma.rn.f32 %f1162, %f419, %f401, %f423;
abs.f32 %f425, %f1161;
setp.leu.f32 %p19, %f425, 0f47CE4780;
@%p19 bra BB0_37;
mov.b32 %r38, %f1161;
shr.u32 %r39, %r38, 23;
shl.b32 %r307, %r38, 8;
or.b32 %r40, %r307, -2147483648;
mov.u32 %r698, 0;
mov.u64 %rd237, __cudart_i2opi_f;
mov.u32 %r697, -6;
mov.u64 %rd238, %rd2;
BB0_28:
.pragma "nounroll";
ld.const.u32 %r310, [%rd237];
// inline asm
{
mad.lo.cc.u32 %r308, %r310, %r40, %r698;
madc.hi.u32 %r698, %r310, %r40, 0;
}
// inline asm
st.local.u32 [%rd238], %r308;
add.s64 %rd238, %rd238, 4;
add.s64 %rd237, %rd237, 4;
add.s32 %r697, %r697, 1;
setp.ne.s32 %p20, %r697, 0;
@%p20 bra BB0_28;
and.b32 %r313, %r39, 255;
add.s32 %r314, %r313, -128;
shr.u32 %r315, %r314, 5;
and.b32 %r45, %r38, -2147483648;
st.local.u32 [%rd3], %r698;
mov.u32 %r316, 6;
sub.s32 %r317, %r316, %r315;
mul.wide.s32 %rd91, %r317, 4;
add.s64 %rd13, %rd2, %rd91;
ld.local.u32 %r699, [%rd13];
ld.local.u32 %r700, [%rd13+-4];
and.b32 %r48, %r39, 31;
setp.eq.s32 %p21, %r48, 0;
@%p21 bra BB0_31;
mov.u32 %r318, 32;
sub.s32 %r319, %r318, %r48;
shr.u32 %r320, %r700, %r319;
shl.b32 %r321, %r699, %r48;
add.s32 %r699, %r320, %r321;
ld.local.u32 %r322, [%rd13+-8];
shr.u32 %r323, %r322, %r319;
shl.b32 %r324, %r700, %r48;
add.s32 %r700, %r323, %r324;
BB0_31:
shr.u32 %r325, %r700, 30;
shl.b32 %r326, %r699, 2;
add.s32 %r701, %r325, %r326;
shl.b32 %r54, %r700, 2;
shr.u32 %r327, %r701, 31;
shr.u32 %r328, %r699, 30;
add.s32 %r55, %r327, %r328;
setp.eq.s32 %p22, %r327, 0;
@%p22 bra BB0_32;
not.b32 %r329, %r701;
neg.s32 %r703, %r54;
setp.eq.s32 %p23, %r54, 0;
selp.u32 %r330, 1, 0, %p23;
add.s32 %r701, %r330, %r329;
xor.b32 %r702, %r45, -2147483648;
bra.uni BB0_34;
BB0_32:
mov.u32 %r702, %r45;
mov.u32 %r703, %r54;
BB0_34:
clz.b32 %r705, %r701;
setp.eq.s32 %p24, %r705, 0;
shl.b32 %r331, %r701, %r705;
mov.u32 %r332, 32;
sub.s32 %r333, %r332, %r705;
shr.u32 %r334, %r703, %r333;
add.s32 %r335, %r334, %r331;
selp.b32 %r63, %r701, %r335, %p24;
mov.u32 %r336, -921707870;
mul.hi.u32 %r704, %r63, %r336;
setp.eq.s32 %p25, %r45, 0;
neg.s32 %r337, %r55;
selp.b32 %r706, %r55, %r337, %p25;
setp.lt.s32 %p26, %r704, 1;
@%p26 bra BB0_36;
mul.lo.s32 %r338, %r63, -921707870;
shr.u32 %r339, %r338, 31;
shl.b32 %r340, %r704, 1;
add.s32 %r704, %r339, %r340;
add.s32 %r705, %r705, 1;
BB0_36:
mov.u32 %r341, 126;
sub.s32 %r342, %r341, %r705;
shl.b32 %r343, %r342, 23;
add.s32 %r344, %r704, 1;
shr.u32 %r345, %r344, 7;
add.s32 %r346, %r345, 1;
shr.u32 %r347, %r346, 1;
add.s32 %r348, %r347, %r343;
or.b32 %r349, %r348, %r702;
mov.b32 %f1162, %r349;
BB0_37:
mul.rn.f32 %f38, %f1162, %f1162;
and.b32 %r71, %r706, 1;
setp.eq.s32 %p27, %r71, 0;
@%p27 bra BB0_39;
mov.f32 %f426, 0fBAB6061A;
mov.f32 %f427, 0f37CCF5CE;
fma.rn.f32 %f1163, %f427, %f38, %f426;
bra.uni BB0_40;
BB0_39:
mov.f32 %f428, 0f3C08839E;
mov.f32 %f429, 0fB94CA1F9;
fma.rn.f32 %f1163, %f429, %f38, %f428;
BB0_40:
@%p27 bra BB0_42;
mov.f32 %f430, 0f3D2AAAA5;
fma.rn.f32 %f431, %f1163, %f38, %f430;
mov.f32 %f432, 0fBF000000;
fma.rn.f32 %f1164, %f431, %f38, %f432;
bra.uni BB0_43;
BB0_42:
mov.f32 %f433, 0fBE2AAAA3;
fma.rn.f32 %f434, %f1163, %f38, %f433;
mov.f32 %f435, 0f00000000;
fma.rn.f32 %f1164, %f434, %f38, %f435;
BB0_43:
fma.rn.f32 %f1165, %f1164, %f1162, %f1162;
@%p27 bra BB0_45;
mov.f32 %f436, 0f3F800000;
fma.rn.f32 %f1165, %f1164, %f38, %f436;
BB0_45:
and.b32 %r350, %r706, 2;
setp.eq.s32 %p30, %r350, 0;
@%p30 bra BB0_47;
mov.f32 %f437, 0f00000000;
mov.f32 %f438, 0fBF800000;
fma.rn.f32 %f1165, %f1165, %f438, %f437;
BB0_47:
mul.f32 %f439, %f13, %f1159;
mul.f32 %f440, %f439, %f439;
mov.f32 %f441, 0f3F800000;
sub.f32 %f442, %f441, %f440;
mul.f32 %f443, %f13, %f1165;
mul.f32 %f444, %f443, %f443;
sub.f32 %f445, %f442, %f444;
mov.f32 %f446, 0f00000000;
max.f32 %f447, %f446, %f445;
sqrt.rn.f32 %f448, %f447;
mul.f32 %f449, %f10, %f443;
mul.f32 %f450, %f11, %f443;
mul.f32 %f451, %f12, %f443;
mul.f32 %f452, %f8, %f11;
mul.f32 %f453, %f9, %f12;
sub.f32 %f454, %f452, %f453;
fma.rn.f32 %f455, %f454, %f439, %f449;
mul.f32 %f456, %f7, %f12;
mul.f32 %f457, %f8, %f10;
sub.f32 %f458, %f456, %f457;
fma.rn.f32 %f459, %f458, %f439, %f450;
mul.f32 %f460, %f9, %f10;
mul.f32 %f461, %f7, %f11;
sub.f32 %f462, %f460, %f461;
fma.rn.f32 %f463, %f462, %f439, %f451;
fma.rn.f32 %f62, %f7, %f448, %f455;
fma.rn.f32 %f63, %f9, %f448, %f459;
fma.rn.f32 %f64, %f8, %f448, %f463;
BB0_49:
add.u64 %rd98, %SP, 32;
cvta.to.local.u64 %rd99, %rd98;
add.s64 %rd14, %rd99, 12;
mov.u32 %r361, -1082130432;
st.local.u32 [%rd99+12], %r361;
mov.u32 %r362, 0;
st.local.u32 [%rd99], %r362;
st.local.u32 [%rd99+4], %r362;
st.local.u32 [%rd99+8], %r362;
st.local.u32 [%rd99+16], %r362;
st.local.u32 [%rd99+20], %r362;
st.local.u32 [%rd99+24], %r362;
ld.global.v2.u32 {%r363, %r364}, [pixelID];
cvt.u64.u32 %rd94, %r363;
cvt.u64.u32 %rd95, %r364;
mov.u64 %rd100, GBufferAtten;
cvta.global.u64 %rd93, %rd100;
mov.u32 %r360, 12;
// inline asm
call (%rd92), _rt_buffer_get_64, (%rd93, %r224, %r360, %rd94, %rd95, %rd240, %rd240);
// inline asm
ld.f32 %f70, [%rd92+8];
ld.f32 %f68, [%rd92+4];
ld.f32 %f69, [%rd92];
add.f32 %f503, %f69, %f68;
add.f32 %f504, %f70, %f503;
ld.global.u32 %r367, [bounceFlag];
setp.ne.s32 %p31, %r367, 0;
setp.eq.f32 %p32, %f504, 0f00000000;
and.pred %p33, %p31, %p32;
@%p33 bra BB0_51;
ld.global.u32 %r368, [root];
mov.u32 %r369, 0;
mov.f32 %f511, 0f00000000;
mov.f32 %f512, 0f6C4ECB8F;
mov.u32 %r370, 28;
// inline asm
call _rt_trace_64, (%r368, %f1170, %f1171, %f1172, %f62, %f63, %f64, %r369, %f511, %f512, %rd98, %r370);
// inline asm
BB0_51:
ld.local.f32 %f71, [%rd14];
setp.ltu.f32 %p34, %f71, 0f00000000;
@%p34 bra BB0_188;
bra.uni BB0_52;
BB0_188:
ld.global.u32 %r628, [bounceFlag];
setp.eq.s32 %p133, %r628, 0;
selp.u16 %rs13, 1, 0, %p133;
mov.f32 %f1231, 0f00000000;
mov.f32 %f1232, %f1231;
mov.f32 %f1233, %f1231;
mov.f32 %f1234, %f1231;
mov.f32 %f1235, %f1231;
mov.f32 %f1236, %f1231;
mov.f32 %f1217, %f1231;
mov.f32 %f1218, %f1231;
mov.f32 %f1219, %f1231;
bra.uni BB0_189;
BB0_52:
fma.rn.f32 %f513, %f62, %f71, %f1170;
fma.rn.f32 %f514, %f63, %f71, %f1171;
fma.rn.f32 %f515, %f64, %f71, %f1172;
mul.f32 %f516, %f513, 0f3456BF95;
mul.f32 %f517, %f514, 0f3456BF95;
mul.f32 %f518, %f515, 0f3456BF95;
abs.f32 %f519, %f516;
abs.f32 %f520, %f517;
abs.f32 %f521, %f518;
max.f32 %f522, %f519, %f520;
max.f32 %f523, %f522, %f521;
mov.f32 %f524, 0f38D1B717;
max.f32 %f525, %f523, %f524;
ld.local.f32 %f526, [%rd14+-12];
ld.local.f32 %f527, [%rd14+-8];
ld.local.f32 %f528, [%rd14+-4];
fma.rn.f32 %f72, %f525, %f526, %f513;
fma.rn.f32 %f73, %f525, %f527, %f514;
fma.rn.f32 %f74, %f525, %f528, %f515;
ld.global.v2.u32 {%r375, %r376}, [pixelID];
cvt.u64.u32 %rd104, %r375;
cvt.u64.u32 %rd105, %r376;
mov.u64 %rd114, GBufferPos;
cvta.global.u64 %rd103, %rd114;
// inline asm
call (%rd102), _rt_buffer_get_64, (%rd103, %r224, %r360, %rd104, %rd105, %rd240, %rd240);
// inline asm
st.f32 [%rd102+8], %f74;
st.f32 [%rd102+4], %f73;
st.f32 [%rd102], %f72;
ld.global.v2.u32 {%r379, %r380}, [pixelID];
cvt.u64.u32 %rd110, %r379;
cvt.u64.u32 %rd111, %r380;
mov.u64 %rd115, GBufferNormal;
cvta.global.u64 %rd109, %rd115;
// inline asm
call (%rd108), _rt_buffer_get_64, (%rd109, %r224, %r360, %rd110, %rd111, %rd240, %rd240);
// inline asm
ld.local.f32 %f529, [%rd14+-12];
ld.local.f32 %f530, [%rd14+-8];
ld.local.f32 %f531, [%rd14+-4];
st.f32 [%rd108+8], %f531;
st.f32 [%rd108+4], %f530;
st.f32 [%rd108], %f529;
ld.local.f32 %f532, [%rd14+12];
ld.local.f32 %f533, [%rd14+8];
ld.local.f32 %f534, [%rd14+4];
setp.lt.f32 %p35, %f534, 0f00000000;
neg.f32 %f535, %f534;
selp.f32 %f536, %f535, %f534, %p35;
ld.global.f32 %f537, [backFaceWeight];
selp.f32 %f538, %f537, 0f3F800000, %p35;
cvt.rmi.f32.f32 %f539, %f536;
cvt.rmi.f32.f32 %f540, %f533;
cvt.rmi.f32.f32 %f541, %f532;
sub.f32 %f542, %f536, %f539;
sub.f32 %f543, %f533, %f540;
sub.f32 %f544, %f532, %f541;
mul.f32 %f1234, %f538, %f542;
mul.f32 %f1235, %f538, %f543;
mul.f32 %f1236, %f538, %f544;
mul.f32 %f1231, %f539, 0f3B808081;
mul.f32 %f1232, %f540, 0f3B808081;
mul.f32 %f1233, %f541, 0f3B808081;
ld.global.u32 %r383, [bounceFlag];
setp.lt.s32 %p36, %r383, 1;
@%p36 bra BB0_54;
mul.f32 %f1234, %f69, %f1234;
mul.f32 %f1235, %f68, %f1235;
mul.f32 %f1236, %f70, %f1236;
mul.f32 %f1231, %f69, %f1231;
mul.f32 %f1232, %f68, %f1232;
mul.f32 %f1233, %f70, %f1233;
BB0_54:
mov.u64 %rd121, globalLights;
cvta.global.u64 %rd120, %rd121;
mov.u32 %r384, 1;
mov.u32 %r385, 48;
// inline asm
call (%rd116, %rd117, %rd118, %rd119), _rt_buffer_get_size_64, (%rd120, %r384, %r385);
// inline asm
cvt.u32.u64 %r72, %rd116;
setp.eq.s32 %p37, %r72, 0;
mov.f32 %f1217, 0f00000000;
mov.f32 %f1218, %f1217;
mov.f32 %f1219, %f1217;
@%p37 bra BB0_162;
// inline asm
call (%rd122, %rd123, %rd124, %rd125), _rt_buffer_get_size_64, (%rd120, %r384, %r385);
// inline asm
cvt.u32.u64 %r390, %rd122;
cvt.rn.f32.u32 %f548, %r390;
mul.f32 %f549, %f3, %f548;
cvt.rzi.u32.f32 %r391, %f549;
add.s32 %r392, %r390, -1;
min.u32 %r393, %r391, %r392;
cvt.u64.u32 %rd129, %r393;
// inline asm
call (%rd127), _rt_buffer_get_64, (%rd120, %r384, %r385, %rd129, %rd240, %rd240, %rd240);
// inline asm
ld.v4.f32 {%f550, %f551, %f552, %f553}, [%rd127+32];
ld.v4.f32 {%f554, %f555, %f556, %f557}, [%rd127+16];
ld.v4.f32 {%f558, %f559, %f560, %f561}, [%rd127];
setp.lt.f32 %p38, %f558, 0f3F800000;
@%p38 bra BB0_109;
bra.uni BB0_56;
BB0_109:
fma.rn.f32 %f681, %f559, 0f40000000, 0fBF800000;
fma.rn.f32 %f682, %f560, 0f40000000, 0fBF800000;
abs.f32 %f683, %f681;
mov.f32 %f1213, 0f3F800000;
sub.f32 %f685, %f1213, %f683;
abs.f32 %f686, %f682;
sub.f32 %f687, %f685, %f686;
neg.f32 %f688, %f687;
cvt.sat.f32.f32 %f689, %f688;
setp.ltu.f32 %p71, %f681, 0f00000000;
neg.f32 %f690, %f689;
selp.f32 %f691, %f689, %f690, %p71;
add.f32 %f692, %f681, %f691;
setp.ltu.f32 %p72, %f682, 0f00000000;
selp.f32 %f693, %f689, %f690, %p72;
add.f32 %f694, %f682, %f693;
mul.f32 %f695, %f694, %f694;
fma.rn.f32 %f696, %f692, %f692, %f695;
fma.rn.f32 %f697, %f687, %f687, %f696;
sqrt.rn.f32 %f698, %f697;
rcp.rn.f32 %f699, %f698;
mul.f32 %f180, %f699, %f692;
mul.f32 %f181, %f699, %f694;
mul.f32 %f182, %f687, %f699;
ld.local.f32 %f700, [%rd14+-12];
mul.f32 %f701, %f700, %f180;
ld.local.f32 %f702, [%rd14+-8];
mul.f32 %f703, %f181, %f702;
neg.f32 %f704, %f703;
sub.f32 %f705, %f704, %f701;
ld.local.f32 %f706, [%rd14+-4];
mul.f32 %f707, %f182, %f706;
sub.f32 %f1200, %f705, %f707;
setp.gt.f32 %p73, %f1200, 0f00000000;
setp.neu.f32 %p74, %f561, 0f00000000;
and.pred %p75, %p73, %p74;
@!%p75 bra BB0_111;
bra.uni BB0_110;
BB0_110:
abs.f32 %f716, %f180;
abs.f32 %f717, %f182;
setp.gt.f32 %p76, %f716, %f717;
neg.f32 %f718, %f181;
selp.f32 %f719, %f718, 0f00000000, %p76;
neg.f32 %f720, %f182;
selp.f32 %f721, %f180, %f720, %p76;
selp.f32 %f722, 0f00000000, %f181, %p76;
mul.f32 %f723, %f721, %f721;
fma.rn.f32 %f724, %f719, %f719, %f723;
fma.rn.f32 %f725, %f722, %f722, %f724;
sqrt.rn.f32 %f726, %f725;
rcp.rn.f32 %f727, %f726;
mul.f32 %f728, %f719, %f727;
mul.f32 %f729, %f721, %f727;
mul.f32 %f730, %f722, %f727;
mul.f32 %f731, %f182, %f729;
mul.f32 %f732, %f181, %f730;
sub.f32 %f733, %f731, %f732;
mul.f32 %f734, %f180, %f730;
mul.f32 %f735, %f182, %f728;
sub.f32 %f736, %f734, %f735;
mul.f32 %f737, %f181, %f728;
mul.f32 %f738, %f180, %f729;
sub.f32 %f739, %f737, %f738;
mul.f32 %f740, %f73, %f736;
fma.rn.f32 %f741, %f72, %f733, %f740;
fma.rn.f32 %f742, %f74, %f739, %f741;
fma.rn.f32 %f712, %f550, %f742, %f552;
mul.f32 %f743, %f73, %f729;
fma.rn.f32 %f744, %f72, %f728, %f743;
fma.rn.f32 %f745, %f74, %f730, %f744;
fma.rn.f32 %f713, %f551, %f745, %f553;
cvt.rzi.s32.f32 %r495, %f561;
mov.f32 %f715, 0f00000000;
// inline asm
call (%f708, %f709, %f710, %f711), _rt_texture_get_f_id, (%r495, %r224, %f712, %f713, %f715, %f715);
// inline asm
mul.f32 %f1200, %f1200, %f708;
BB0_111:
mov.f32 %f1214, 0f00000000;
setp.le.f32 %p77, %f1200, 0f00000000;
mov.f32 %f1215, %f1214;
mov.f32 %f1216, %f1214;
@%p77 bra BB0_161;
abs.f32 %f750, %f180;
abs.f32 %f751, %f182;
setp.gt.f32 %p78, %f750, %f751;
neg.f32 %f752, %f181;
selp.f32 %f753, %f752, 0f00000000, %p78;
neg.f32 %f754, %f182;
selp.f32 %f755, %f180, %f754, %p78;
selp.f32 %f756, 0f00000000, %f181, %p78;
mul.f32 %f757, %f755, %f755;
fma.rn.f32 %f758, %f753, %f753, %f757;
fma.rn.f32 %f759, %f756, %f756, %f758;
sqrt.rn.f32 %f760, %f759;
rcp.rn.f32 %f761, %f760;
mul.f32 %f186, %f753, %f761;
mul.f32 %f187, %f755, %f761;
mul.f32 %f188, %f756, %f761;
setp.ltu.f32 %p79, %f554, 0f00000000;
@%p79 bra BB0_160;
sqrt.rn.f32 %f189, %f1;
add.u64 %rd152, %SP, 0;
cvta.to.local.u64 %rd28, %rd152;
mul.f32 %f1207, %f2, 0f40C90FDB;
abs.f32 %f191, %f1207;
setp.neu.f32 %p80, %f191, 0f7F800000;
mov.f32 %f1201, %f1207;
@%p80 bra BB0_115;
mov.f32 %f762, 0f00000000;
mul.rn.f32 %f1201, %f1207, %f762;
BB0_115:
mul.f32 %f763, %f1201, 0f3F22F983;
cvt.rni.s32.f32 %r736, %f763;
cvt.rn.f32.s32 %f764, %r736;
neg.f32 %f765, %f764;
mov.f32 %f766, 0f3FC90FDA;
fma.rn.f32 %f767, %f765, %f766, %f1201;
mov.f32 %f768, 0f33A22168;
fma.rn.f32 %f769, %f765, %f768, %f767;
mov.f32 %f770, 0f27C234C5;
fma.rn.f32 %f1202, %f765, %f770, %f769;
abs.f32 %f771, %f1201;
add.s64 %rd29, %rd28, 24;
setp.leu.f32 %p81, %f771, 0f47CE4780;
@%p81 bra BB0_126;
mov.b32 %r144, %f1201;
shr.u32 %r145, %r144, 23;
shl.b32 %r499, %r144, 8;
or.b32 %r146, %r499, -2147483648;
mov.u32 %r728, 0;
mov.u64 %rd243, __cudart_i2opi_f;
mov.u32 %r727, -6;
mov.u64 %rd244, %rd28;
BB0_117:
.pragma "nounroll";
ld.const.u32 %r502, [%rd243];
// inline asm
{
mad.lo.cc.u32 %r500, %r502, %r146, %r728;
madc.hi.u32 %r728, %r502, %r146, 0;
}
// inline asm
st.local.u32 [%rd244], %r500;
add.s64 %rd244, %rd244, 4;
add.s64 %rd243, %rd243, 4;
add.s32 %r727, %r727, 1;
setp.ne.s32 %p82, %r727, 0;
@%p82 bra BB0_117;
and.b32 %r505, %r145, 255;
add.s32 %r506, %r505, -128;
shr.u32 %r507, %r506, 5;
and.b32 %r151, %r144, -2147483648;
st.local.u32 [%rd29], %r728;
mov.u32 %r508, 6;
sub.s32 %r509, %r508, %r507;
mul.wide.s32 %rd154, %r509, 4;
add.s64 %rd34, %rd28, %rd154;
ld.local.u32 %r729, [%rd34];
ld.local.u32 %r730, [%rd34+-4];
and.b32 %r154, %r145, 31;
setp.eq.s32 %p83, %r154, 0;
@%p83 bra BB0_120;
mov.u32 %r510, 32;
sub.s32 %r511, %r510, %r154;
shr.u32 %r512, %r730, %r511;
shl.b32 %r513, %r729, %r154;
add.s32 %r729, %r512, %r513;
ld.local.u32 %r514, [%rd34+-8];
shr.u32 %r515, %r514, %r511;
shl.b32 %r516, %r730, %r154;
add.s32 %r730, %r515, %r516;
BB0_120:
shr.u32 %r517, %r730, 30;
shl.b32 %r518, %r729, 2;
add.s32 %r731, %r517, %r518;
shl.b32 %r160, %r730, 2;
shr.u32 %r519, %r731, 31;
shr.u32 %r520, %r729, 30;
add.s32 %r161, %r519, %r520;
setp.eq.s32 %p84, %r519, 0;
@%p84 bra BB0_121;
not.b32 %r521, %r731;
neg.s32 %r733, %r160;
setp.eq.s32 %p85, %r160, 0;
selp.u32 %r522, 1, 0, %p85;
add.s32 %r731, %r522, %r521;
xor.b32 %r732, %r151, -2147483648;
bra.uni BB0_123;
BB0_56:
ld.local.f32 %f114, [%rd14+-8];
ld.local.f32 %f113, [%rd14+-4];
ld.local.f32 %f112, [%rd14+-12];
abs.f32 %f562, %f112;
abs.f32 %f563, %f113;
setp.gt.f32 %p39, %f562, %f563;
neg.f32 %f564, %f114;
neg.f32 %f565, %f113;
selp.f32 %f566, %f564, 0f00000000, %p39;
selp.f32 %f567, %f112, %f565, %p39;
selp.f32 %f568, 0f00000000, %f114, %p39;
mul.f32 %f569, %f567, %f567;
fma.rn.f32 %f570, %f566, %f566, %f569;
fma.rn.f32 %f571, %f568, %f568, %f570;
sqrt.rn.f32 %f572, %f571;
rcp.rn.f32 %f573, %f572;
mul.f32 %f115, %f566, %f573;
mul.f32 %f116, %f567, %f573;
mul.f32 %f117, %f568, %f573;
sqrt.rn.f32 %f118, %f1;
mul.f32 %f1185, %f2, 0f40C90FDB;
abs.f32 %f120, %f1185;
setp.neu.f32 %p40, %f120, 0f7F800000;
mov.f32 %f1179, %f1185;
@%p40 bra BB0_58;
mov.f32 %f574, 0f00000000;
mul.rn.f32 %f1179, %f1185, %f574;
BB0_58:
mul.f32 %f575, %f1179, 0f3F22F983;
cvt.rni.s32.f32 %r716, %f575;
cvt.rn.f32.s32 %f576, %r716;
neg.f32 %f577, %f576;
mov.f32 %f578, 0f3FC90FDA;
fma.rn.f32 %f579, %f577, %f578, %f1179;
mov.f32 %f580, 0f33A22168;
fma.rn.f32 %f581, %f577, %f580, %f579;
mov.f32 %f582, 0f27C234C5;
fma.rn.f32 %f1180, %f577, %f582, %f581;
abs.f32 %f583, %f1179;
setp.leu.f32 %p41, %f583, 0f47CE4780;
@%p41 bra BB0_69;
mov.b32 %r74, %f1179;
shl.b32 %r396, %r74, 8;
or.b32 %r75, %r396, -2147483648;
add.u64 %rd135, %SP, 0;
cvta.to.local.u64 %rd15, %rd135;
mov.u32 %r707, 0;
mov.u64 %rd239, %rd15;
mov.u32 %r708, %r707;
BB0_60:
.pragma "nounroll";
shl.b64 %rd136, %rd240, 2;
mov.u64 %rd137, __cudart_i2opi_f;
add.s64 %rd138, %rd137, %rd136;
ld.const.u32 %r399, [%rd138];
// inline asm
{
mad.lo.cc.u32 %r397, %r399, %r75, %r708;
madc.hi.u32 %r708, %r399, %r75, 0;
}
// inline asm
st.local.u32 [%rd239], %r397;
add.s32 %r707, %r707, 1;
cvt.s64.s32 %rd240, %r707;
mul.wide.s32 %rd141, %r707, 4;
add.s64 %rd239, %rd15, %rd141;
setp.ne.s32 %p42, %r707, 6;
@%p42 bra BB0_60;
bfe.u32 %r402, %r74, 23, 8;
add.s32 %r403, %r402, -128;
shr.u32 %r404, %r403, 5;
and.b32 %r80, %r74, -2147483648;
cvta.to.local.u64 %rd143, %rd135;
st.local.u32 [%rd143+24], %r708;
bfe.u32 %r81, %r74, 23, 5;
mov.u32 %r405, 6;
sub.s32 %r406, %r405, %r404;
mul.wide.s32 %rd144, %r406, 4;
add.s64 %rd20, %rd143, %rd144;
ld.local.u32 %r709, [%rd20];
ld.local.u32 %r710, [%rd20+-4];
setp.eq.s32 %p43, %r81, 0;
@%p43 bra BB0_63;
mov.u32 %r407, 32;
sub.s32 %r408, %r407, %r81;
shr.u32 %r409, %r710, %r408;
shl.b32 %r410, %r709, %r81;
add.s32 %r709, %r409, %r410;
ld.local.u32 %r411, [%rd20+-8];
shr.u32 %r412, %r411, %r408;
shl.b32 %r413, %r710, %r81;
add.s32 %r710, %r412, %r413;
BB0_63:
shr.u32 %r414, %r710, 30;
shl.b32 %r415, %r709, 2;
add.s32 %r711, %r414, %r415;
shl.b32 %r89, %r710, 2;
shr.u32 %r416, %r711, 31;
shr.u32 %r417, %r709, 30;
add.s32 %r90, %r416, %r417;
setp.eq.s32 %p44, %r416, 0;
@%p44 bra BB0_64;
not.b32 %r418, %r711;
neg.s32 %r713, %r89;
setp.eq.s32 %p45, %r89, 0;
selp.u32 %r419, 1, 0, %p45;
add.s32 %r711, %r419, %r418;
xor.b32 %r712, %r80, -2147483648;
bra.uni BB0_66;
BB0_64:
mov.u32 %r712, %r80;
mov.u32 %r713, %r89;
BB0_66:
clz.b32 %r715, %r711;
setp.eq.s32 %p46, %r715, 0;
shl.b32 %r420, %r711, %r715;
mov.u32 %r421, 32;
sub.s32 %r422, %r421, %r715;
shr.u32 %r423, %r713, %r422;
add.s32 %r424, %r423, %r420;
selp.b32 %r98, %r711, %r424, %p46;
mov.u32 %r425, -921707870;
mul.hi.u32 %r714, %r98, %r425;
setp.eq.s32 %p47, %r80, 0;
neg.s32 %r426, %r90;
selp.b32 %r716, %r90, %r426, %p47;
setp.lt.s32 %p48, %r714, 1;
@%p48 bra BB0_68;
mul.lo.s32 %r427, %r98, -921707870;
shr.u32 %r428, %r427, 31;
shl.b32 %r429, %r714, 1;
add.s32 %r714, %r428, %r429;
add.s32 %r715, %r715, 1;
BB0_68:
mov.u32 %r430, 126;
sub.s32 %r431, %r430, %r715;
shl.b32 %r432, %r431, 23;
add.s32 %r433, %r714, 1;
shr.u32 %r434, %r433, 7;
add.s32 %r435, %r434, 1;
shr.u32 %r436, %r435, 1;
add.s32 %r437, %r436, %r432;
or.b32 %r438, %r437, %r712;
mov.b32 %f1180, %r438;
BB0_69:
mul.rn.f32 %f126, %f1180, %f1180;
add.s32 %r106, %r716, 1;
and.b32 %r107, %r106, 1;
setp.eq.s32 %p49, %r107, 0;
@%p49 bra BB0_71;
mov.f32 %f584, 0fBAB6061A;
mov.f32 %f585, 0f37CCF5CE;
fma.rn.f32 %f1181, %f585, %f126, %f584;
bra.uni BB0_72;
BB0_71:
mov.f32 %f586, 0f3C08839E;
mov.f32 %f587, 0fB94CA1F9;
fma.rn.f32 %f1181, %f587, %f126, %f586;
BB0_72:
@%p49 bra BB0_74;
mov.f32 %f588, 0f3D2AAAA5;
fma.rn.f32 %f589, %f1181, %f126, %f588;
mov.f32 %f590, 0fBF000000;
fma.rn.f32 %f1182, %f589, %f126, %f590;
bra.uni BB0_75;
BB0_74:
mov.f32 %f591, 0fBE2AAAA3;
fma.rn.f32 %f592, %f1181, %f126, %f591;
mov.f32 %f593, 0f00000000;
fma.rn.f32 %f1182, %f592, %f126, %f593;
BB0_75:
fma.rn.f32 %f1183, %f1182, %f1180, %f1180;
@%p49 bra BB0_77;
mov.f32 %f594, 0f3F800000;
fma.rn.f32 %f1183, %f1182, %f126, %f594;
BB0_77:
and.b32 %r439, %r106, 2;
setp.eq.s32 %p52, %r439, 0;
@%p52 bra BB0_79;
mov.f32 %f595, 0f00000000;
mov.f32 %f596, 0fBF800000;
fma.rn.f32 %f1183, %f1183, %f596, %f595;
BB0_79:
@%p40 bra BB0_81;
mov.f32 %f597, 0f00000000;
mul.rn.f32 %f1185, %f1185, %f597;
BB0_81:
mul.f32 %f598, %f1185, 0f3F22F983;
cvt.rni.s32.f32 %r726, %f598;
cvt.rn.f32.s32 %f599, %r726;
neg.f32 %f600, %f599;
fma.rn.f32 %f602, %f600, %f578, %f1185;
fma.rn.f32 %f604, %f600, %f580, %f602;
fma.rn.f32 %f1186, %f600, %f582, %f604;
abs.f32 %f606, %f1185;
setp.leu.f32 %p54, %f606, 0f47CE4780;
@%p54 bra BB0_92;
mov.b32 %r109, %f1185;
shr.u32 %r110, %r109, 23;
shl.b32 %r442, %r109, 8;
or.b32 %r111, %r442, -2147483648;
add.u64 %rd146, %SP, 0;
cvta.to.local.u64 %rd242, %rd146;
mov.u64 %rd241, __cudart_i2opi_f;
mov.u32 %r717, -6;
mov.u32 %r718, %r362;
BB0_83:
.pragma "nounroll";
ld.const.u32 %r445, [%rd241];
// inline asm
{
mad.lo.cc.u32 %r443, %r445, %r111, %r718;
madc.hi.u32 %r718, %r445, %r111, 0;
}
// inline asm
st.local.u32 [%rd242], %r443;
add.s64 %rd242, %rd242, 4;
add.s64 %rd241, %rd241, 4;
add.s32 %r717, %r717, 1;
setp.ne.s32 %p55, %r717, 0;
@%p55 bra BB0_83;
and.b32 %r448, %r110, 255;
add.s32 %r449, %r448, -128;
shr.u32 %r450, %r449, 5;
and.b32 %r116, %r109, -2147483648;
cvta.to.local.u64 %rd148, %rd146;
st.local.u32 [%rd148+24], %r718;
mov.u32 %r451, 6;
sub.s32 %r452, %r451, %r450;
mul.wide.s32 %rd149, %r452, 4;
add.s64 %rd26, %rd148, %rd149;
ld.local.u32 %r719, [%rd26];
ld.local.u32 %r720, [%rd26+-4];
and.b32 %r119, %r110, 31;
setp.eq.s32 %p56, %r119, 0;
@%p56 bra BB0_86;
mov.u32 %r453, 32;
sub.s32 %r454, %r453, %r119;
shr.u32 %r455, %r720, %r454;
shl.b32 %r456, %r719, %r119;
add.s32 %r719, %r455, %r456;
ld.local.u32 %r457, [%rd26+-8];
shr.u32 %r458, %r457, %r454;
shl.b32 %r459, %r720, %r119;
add.s32 %r720, %r458, %r459;
BB0_86:
shr.u32 %r460, %r720, 30;
shl.b32 %r461, %r719, 2;
add.s32 %r721, %r460, %r461;
shl.b32 %r125, %r720, 2;
shr.u32 %r462, %r721, 31;
shr.u32 %r463, %r719, 30;
add.s32 %r126, %r462, %r463;
setp.eq.s32 %p57, %r462, 0;
@%p57 bra BB0_87;
not.b32 %r464, %r721;
neg.s32 %r723, %r125;
setp.eq.s32 %p58, %r125, 0;
selp.u32 %r465, 1, 0, %p58;
add.s32 %r721, %r465, %r464;
xor.b32 %r722, %r116, -2147483648;
bra.uni BB0_89;
BB0_87:
mov.u32 %r722, %r116;
mov.u32 %r723, %r125;
BB0_89:
clz.b32 %r725, %r721;
setp.eq.s32 %p59, %r725, 0;
shl.b32 %r466, %r721, %r725;
mov.u32 %r467, 32;
sub.s32 %r468, %r467, %r725;
shr.u32 %r469, %r723, %r468;
add.s32 %r470, %r469, %r466;
selp.b32 %r134, %r721, %r470, %p59;
mov.u32 %r471, -921707870;
mul.hi.u32 %r724, %r134, %r471;
setp.eq.s32 %p60, %r116, 0;
neg.s32 %r472, %r126;
selp.b32 %r726, %r126, %r472, %p60;
setp.lt.s32 %p61, %r724, 1;
@%p61 bra BB0_91;
mul.lo.s32 %r473, %r134, -921707870;
shr.u32 %r474, %r473, 31;
shl.b32 %r475, %r724, 1;
add.s32 %r724, %r474, %r475;
add.s32 %r725, %r725, 1;
BB0_91:
mov.u32 %r476, 126;
sub.s32 %r477, %r476, %r725;
shl.b32 %r478, %r477, 23;
add.s32 %r479, %r724, 1;
shr.u32 %r480, %r479, 7;
add.s32 %r481, %r480, 1;
shr.u32 %r482, %r481, 1;
add.s32 %r483, %r482, %r478;
or.b32 %r484, %r483, %r722;
mov.b32 %f1186, %r484;
BB0_92:
mul.rn.f32 %f143, %f1186, %f1186;
and.b32 %r142, %r726, 1;
setp.eq.s32 %p62, %r142, 0;
@%p62 bra BB0_94;
mov.f32 %f607, 0fBAB6061A;
mov.f32 %f608, 0f37CCF5CE;
fma.rn.f32 %f1187, %f608, %f143, %f607;
bra.uni BB0_95;
BB0_94:
mov.f32 %f609, 0f3C08839E;
mov.f32 %f610, 0fB94CA1F9;
fma.rn.f32 %f1187, %f610, %f143, %f609;
BB0_95:
@%p62 bra BB0_97;
mov.f32 %f611, 0f3D2AAAA5;
fma.rn.f32 %f612, %f1187, %f143, %f611;
mov.f32 %f613, 0fBF000000;
fma.rn.f32 %f1188, %f612, %f143, %f613;
bra.uni BB0_98;
BB0_97:
mov.f32 %f614, 0fBE2AAAA3;
fma.rn.f32 %f615, %f1187, %f143, %f614;
mov.f32 %f616, 0f00000000;
fma.rn.f32 %f1188, %f615, %f143, %f616;
BB0_98:
fma.rn.f32 %f1189, %f1188, %f1186, %f1186;
@%p62 bra BB0_100;
mov.f32 %f617, 0f3F800000;
fma.rn.f32 %f1189, %f1188, %f143, %f617;
BB0_100:
and.b32 %r485, %r726, 2;
setp.eq.s32 %p65, %r485, 0;
@%p65 bra BB0_102;
mov.f32 %f618, 0f00000000;
mov.f32 %f619, 0fBF800000;
fma.rn.f32 %f1189, %f1189, %f619, %f618;
BB0_102:
mul.f32 %f623, %f118, %f1183;
mul.f32 %f624, %f623, %f623;
mov.f32 %f1191, 0f3F800000;
sub.f32 %f626, %f1191, %f624;
mul.f32 %f627, %f118, %f1189;
mul.f32 %f628, %f627, %f627;
sub.f32 %f629, %f626, %f628;
mov.f32 %f1194, 0f00000000;
max.f32 %f630, %f1194, %f629;
sqrt.rn.f32 %f631, %f630;
mul.f32 %f632, %f115, %f627;
mul.f32 %f633, %f116, %f627;
mul.f32 %f634, %f117, %f627;
mul.f32 %f635, %f113, %f116;
mul.f32 %f636, %f114, %f117;
sub.f32 %f637, %f635, %f636;
fma.rn.f32 %f638, %f637, %f623, %f632;
mul.f32 %f639, %f112, %f117;
mul.f32 %f640, %f113, %f115;
sub.f32 %f641, %f639, %f640;
fma.rn.f32 %f642, %f641, %f623, %f633;
mul.f32 %f643, %f114, %f115;
mul.f32 %f644, %f112, %f116;
sub.f32 %f645, %f643, %f644;
fma.rn.f32 %f646, %f645, %f623, %f634;
fma.rn.f32 %f155, %f112, %f631, %f638;
fma.rn.f32 %f156, %f114, %f631, %f642;
fma.rn.f32 %f157, %f113, %f631, %f646;
add.u64 %rd150, %SP, 28;
cvta.to.local.u64 %rd27, %rd150;
st.local.u32 [%rd27], %r362;
setp.gt.f32 %p66, %f156, 0f00000000;
setp.eq.f32 %p67, %f554, 0f00000000;
or.pred %p68, %p67, %p66;
mov.f32 %f1195, %f1194;
mov.f32 %f1196, %f1194;
@!%p68 bra BB0_106;
bra.uni BB0_103;
BB0_103:
mov.u32 %r490, 1065353216;
st.local.u32 [%rd27], %r490;
mul.f32 %f658, %f72, 0f3456BF95;
abs.f32 %f659, %f658;
mul.f32 %f660, %f73, 0f3456BF95;
abs.f32 %f661, %f660;
mul.f32 %f662, %f74, 0f3456BF95;
abs.f32 %f663, %f662;
max.f32 %f664, %f659, %f661;
max.f32 %f665, %f664, %f663;
max.f32 %f653, %f665, %f524;
ld.global.u32 %r487, [root];
mov.f32 %f654, 0f6C4ECB8F;
mov.u32 %r489, 4;
// inline asm
call _rt_trace_64, (%r487, %f72, %f73, %f74, %f155, %f156, %f157, %r384, %f653, %f654, %rd150, %r489);
// inline asm
setp.leu.f32 %p69, %f559, 0f00000000;
mov.f32 %f1192, %f1191;
mov.f32 %f1193, %f1191;
@%p69 bra BB0_105;
cvt.rzi.s32.f32 %r491, %f560;
neg.f32 %f673, %f155;
neg.f32 %f672, %f156;
neg.f32 %f671, %f157;
mov.u32 %r492, 6;
// inline asm
call (%f1191, %f1192, %f1193, %f670), _rt_texture_get_base_id, (%r491, %r492, %f671, %f672, %f673, %r362);
// inline asm
BB0_105:
ld.local.f32 %f674, [%rd27];
fma.rn.f32 %f1194, %f1191, %f674, 0f00000000;
fma.rn.f32 %f1195, %f1192, %f674, 0f00000000;
fma.rn.f32 %f1196, %f1193, %f674, 0f00000000;
BB0_106:
ld.global.u32 %r494, [bounceFlag];
setp.lt.s32 %p70, %r494, 1;
@%p70 bra BB0_108;
add.f32 %f675, %f558, 0fBF800000;
mul.f32 %f676, %f675, 0f42C80000;
mul.f32 %f1194, %f676, %f1194;
mul.f32 %f1195, %f676, %f1195;
mul.f32 %f1196, %f676, %f1196;
BB0_108:
cvt.rn.f32.u32 %f677, %r72;
mul.f32 %f678, %f555, %f1194;
mul.f32 %f1217, %f677, %f678;
mul.f32 %f679, %f556, %f1195;
mul.f32 %f1218, %f677, %f679;
mul.f32 %f680, %f557, %f1196;
mul.f32 %f1219, %f677, %f680;
bra.uni BB0_162;
BB0_121:
mov.u32 %r732, %r151;
mov.u32 %r733, %r160;
BB0_123:
clz.b32 %r735, %r731;
setp.eq.s32 %p86, %r735, 0;
shl.b32 %r523, %r731, %r735;
mov.u32 %r524, 32;
sub.s32 %r525, %r524, %r735;
shr.u32 %r526, %r733, %r525;
add.s32 %r527, %r526, %r523;
selp.b32 %r169, %r731, %r527, %p86;
mov.u32 %r528, -921707870;
mul.hi.u32 %r734, %r169, %r528;
setp.eq.s32 %p87, %r151, 0;
neg.s32 %r529, %r161;
selp.b32 %r736, %r161, %r529, %p87;
setp.lt.s32 %p88, %r734, 1;
@%p88 bra BB0_125;
mul.lo.s32 %r530, %r169, -921707870;
shr.u32 %r531, %r530, 31;
shl.b32 %r532, %r734, 1;
add.s32 %r734, %r531, %r532;
add.s32 %r735, %r735, 1;
BB0_125:
mov.u32 %r533, 126;
sub.s32 %r534, %r533, %r735;
shl.b32 %r535, %r534, 23;
add.s32 %r536, %r734, 1;
shr.u32 %r537, %r536, 7;
add.s32 %r538, %r537, 1;
shr.u32 %r539, %r538, 1;
add.s32 %r540, %r539, %r535;
or.b32 %r541, %r540, %r732;
mov.b32 %f1202, %r541;
BB0_126:
mul.rn.f32 %f197, %f1202, %f1202;
add.s32 %r177, %r736, 1;
and.b32 %r178, %r177, 1;
setp.eq.s32 %p89, %r178, 0;
@%p89 bra BB0_128;
mov.f32 %f772, 0fBAB6061A;
mov.f32 %f773, 0f37CCF5CE;
fma.rn.f32 %f1203, %f773, %f197, %f772;
bra.uni BB0_129;
BB0_128:
mov.f32 %f774, 0f3C08839E;
mov.f32 %f775, 0fB94CA1F9;
fma.rn.f32 %f1203, %f775, %f197, %f774;
BB0_129:
@%p89 bra BB0_131;
mov.f32 %f776, 0f3D2AAAA5;
fma.rn.f32 %f777, %f1203, %f197, %f776;
mov.f32 %f778, 0fBF000000;
fma.rn.f32 %f1204, %f777, %f197, %f778;
bra.uni BB0_132;
BB0_131:
mov.f32 %f779, 0fBE2AAAA3;
fma.rn.f32 %f780, %f1203, %f197, %f779;
mov.f32 %f781, 0f00000000;
fma.rn.f32 %f1204, %f780, %f197, %f781;
BB0_132:
fma.rn.f32 %f1205, %f1204, %f1202, %f1202;
@%p89 bra BB0_134;
mov.f32 %f782, 0f3F800000;
fma.rn.f32 %f1205, %f1204, %f197, %f782;
BB0_134:
and.b32 %r542, %r177, 2;
setp.eq.s32 %p92, %r542, 0;
@%p92 bra BB0_136;
mov.f32 %f783, 0f00000000;
mov.f32 %f784, 0fBF800000;
fma.rn.f32 %f1205, %f1205, %f784, %f783;
BB0_136:
@%p80 bra BB0_138;
mov.f32 %f785, 0f00000000;
mul.rn.f32 %f1207, %f1207, %f785;
BB0_138:
mul.f32 %f786, %f1207, 0f3F22F983;
cvt.rni.s32.f32 %r746, %f786;
cvt.rn.f32.s32 %f787, %r746;
neg.f32 %f788, %f787;
fma.rn.f32 %f790, %f788, %f766, %f1207;
fma.rn.f32 %f792, %f788, %f768, %f790;
fma.rn.f32 %f1208, %f788, %f770, %f792;
abs.f32 %f794, %f1207;
setp.leu.f32 %p94, %f794, 0f47CE4780;
@%p94 bra BB0_149;
mov.b32 %r180, %f1207;
shr.u32 %r181, %r180, 23;
shl.b32 %r545, %r180, 8;
or.b32 %r182, %r545, -2147483648;
mov.u32 %r738, 0;
mov.u64 %rd245, __cudart_i2opi_f;
mov.u32 %r737, -6;
mov.u64 %rd246, %rd28;
BB0_140:
.pragma "nounroll";
ld.const.u32 %r548, [%rd245];
// inline asm
{
mad.lo.cc.u32 %r546, %r548, %r182, %r738;
madc.hi.u32 %r738, %r548, %r182, 0;
}
// inline asm
st.local.u32 [%rd246], %r546;
add.s64 %rd246, %rd246, 4;
add.s64 %rd245, %rd245, 4;
add.s32 %r737, %r737, 1;
setp.ne.s32 %p95, %r737, 0;
@%p95 bra BB0_140;
and.b32 %r551, %r181, 255;
add.s32 %r552, %r551, -128;
shr.u32 %r553, %r552, 5;
and.b32 %r187, %r180, -2147483648;
st.local.u32 [%rd29], %r738;
mov.u32 %r554, 6;
sub.s32 %r555, %r554, %r553;
mul.wide.s32 %rd156, %r555, 4;
add.s64 %rd39, %rd28, %rd156;
ld.local.u32 %r739, [%rd39];
ld.local.u32 %r740, [%rd39+-4];
and.b32 %r190, %r181, 31;
setp.eq.s32 %p96, %r190, 0;
@%p96 bra BB0_143;
mov.u32 %r556, 32;
sub.s32 %r557, %r556, %r190;
shr.u32 %r558, %r740, %r557;
shl.b32 %r559, %r739, %r190;
add.s32 %r739, %r558, %r559;
ld.local.u32 %r560, [%rd39+-8];
shr.u32 %r561, %r560, %r557;
shl.b32 %r562, %r740, %r190;
add.s32 %r740, %r561, %r562;
BB0_143:
shr.u32 %r563, %r740, 30;
shl.b32 %r564, %r739, 2;
add.s32 %r741, %r563, %r564;
shl.b32 %r196, %r740, 2;
shr.u32 %r565, %r741, 31;
shr.u32 %r566, %r739, 30;
add.s32 %r197, %r565, %r566;
setp.eq.s32 %p97, %r565, 0;
@%p97 bra BB0_144;
not.b32 %r567, %r741;
neg.s32 %r743, %r196;
setp.eq.s32 %p98, %r196, 0;
selp.u32 %r568, 1, 0, %p98;
add.s32 %r741, %r568, %r567;
xor.b32 %r742, %r187, -2147483648;
bra.uni BB0_146;
BB0_144:
mov.u32 %r742, %r187;
mov.u32 %r743, %r196;
BB0_146:
clz.b32 %r745, %r741;
setp.eq.s32 %p99, %r745, 0;
shl.b32 %r569, %r741, %r745;
mov.u32 %r570, 32;
sub.s32 %r571, %r570, %r745;
shr.u32 %r572, %r743, %r571;
add.s32 %r573, %r572, %r569;
selp.b32 %r205, %r741, %r573, %p99;
mov.u32 %r574, -921707870;
mul.hi.u32 %r744, %r205, %r574;
setp.eq.s32 %p100, %r187, 0;
neg.s32 %r575, %r197;
selp.b32 %r746, %r197, %r575, %p100;
setp.lt.s32 %p101, %r744, 1;
@%p101 bra BB0_148;
mul.lo.s32 %r576, %r205, -921707870;
shr.u32 %r577, %r576, 31;
shl.b32 %r578, %r744, 1;
add.s32 %r744, %r577, %r578;
add.s32 %r745, %r745, 1;
BB0_148:
mov.u32 %r579, 126;
sub.s32 %r580, %r579, %r745;
shl.b32 %r581, %r580, 23;
add.s32 %r582, %r744, 1;
shr.u32 %r583, %r582, 7;
add.s32 %r584, %r583, 1;
shr.u32 %r585, %r584, 1;
add.s32 %r586, %r585, %r581;
or.b32 %r587, %r586, %r742;
mov.b32 %f1208, %r587;
BB0_149:
mul.rn.f32 %f214, %f1208, %f1208;
and.b32 %r213, %r746, 1;
setp.eq.s32 %p102, %r213, 0;
@%p102 bra BB0_151;
mov.f32 %f795, 0fBAB6061A;
mov.f32 %f796, 0f37CCF5CE;
fma.rn.f32 %f1209, %f796, %f214, %f795;
bra.uni BB0_152;
BB0_151:
mov.f32 %f797, 0f3C08839E;
mov.f32 %f798, 0fB94CA1F9;
fma.rn.f32 %f1209, %f798, %f214, %f797;
BB0_152:
@%p102 bra BB0_154;
mov.f32 %f799, 0f3D2AAAA5;
fma.rn.f32 %f800, %f1209, %f214, %f799;
mov.f32 %f801, 0fBF000000;
fma.rn.f32 %f1210, %f800, %f214, %f801;
bra.uni BB0_155;
BB0_154:
mov.f32 %f802, 0fBE2AAAA3;
fma.rn.f32 %f803, %f1209, %f214, %f802;
mov.f32 %f804, 0f00000000;
fma.rn.f32 %f1210, %f803, %f214, %f804;
BB0_155:
fma.rn.f32 %f1211, %f1210, %f1208, %f1208;
@%p102 bra BB0_157;
mov.f32 %f805, 0f3F800000;
fma.rn.f32 %f1211, %f1210, %f214, %f805;
BB0_157:
and.b32 %r588, %r746, 2;
setp.eq.s32 %p105, %r588, 0;
@%p105 bra BB0_159;
mov.f32 %f806, 0f00000000;
mov.f32 %f807, 0fBF800000;
fma.rn.f32 %f1211, %f1211, %f807, %f806;
BB0_159:
mul.f32 %f816, %f189, %f1205;
mul.f32 %f817, %f816, %f816;
mov.f32 %f818, 0f3F800000;
sub.f32 %f819, %f818, %f817;
mul.f32 %f820, %f189, %f1211;
mul.f32 %f821, %f820, %f820;
sub.f32 %f822, %f819, %f821;
mov.f32 %f823, 0f00000000;
max.f32 %f824, %f823, %f822;
sqrt.rn.f32 %f825, %f824;
mul.f32 %f826, %f186, %f820;
mul.f32 %f827, %f187, %f820;
mul.f32 %f828, %f188, %f820;
mul.f32 %f829, %f181, %f188;
mul.f32 %f830, %f182, %f187;
sub.f32 %f831, %f830, %f829;
fma.rn.f32 %f832, %f831, %f816, %f826;
mul.f32 %f833, %f182, %f186;
mul.f32 %f834, %f180, %f188;
sub.f32 %f835, %f834, %f833;
fma.rn.f32 %f836, %f835, %f816, %f827;
mul.f32 %f837, %f180, %f187;
mul.f32 %f838, %f181, %f186;
sub.f32 %f839, %f838, %f837;
fma.rn.f32 %f840, %f839, %f816, %f828;
fma.rn.f32 %f841, %f180, %f825, %f832;
fma.rn.f32 %f842, %f181, %f825, %f836;
fma.rn.f32 %f843, %f182, %f825, %f840;
add.f32 %f844, %f180, %f841;
add.f32 %f845, %f181, %f842;
add.f32 %f846, %f182, %f843;
mul.f32 %f847, %f554, %f844;
mul.f32 %f848, %f554, %f845;
mul.f32 %f849, %f554, %f846;
sub.f32 %f850, %f847, %f180;
sub.f32 %f851, %f848, %f181;
sub.f32 %f852, %f849, %f182;
mul.f32 %f853, %f851, %f851;
fma.rn.f32 %f854, %f850, %f850, %f853;
fma.rn.f32 %f855, %f852, %f852, %f854;
sqrt.rn.f32 %f856, %f855;
rcp.rn.f32 %f857, %f856;
mul.f32 %f811, %f857, %f850;
mul.f32 %f812, %f857, %f851;
mul.f32 %f813, %f857, %f852;
mul.f32 %f858, %f72, 0f3456BF95;
abs.f32 %f859, %f858;
mul.f32 %f860, %f73, 0f3456BF95;
abs.f32 %f861, %f860;
mul.f32 %f862, %f74, 0f3456BF95;
abs.f32 %f863, %f862;
max.f32 %f864, %f859, %f861;
max.f32 %f865, %f864, %f863;
max.f32 %f814, %f865, %f524;
add.u64 %rd157, %SP, 28;
cvta.to.local.u64 %rd158, %rd157;
mov.u32 %r592, 1065353216;
st.local.u32 [%rd158], %r592;
ld.global.u32 %r589, [root];
mov.f32 %f815, 0f6C4ECB8F;
mov.u32 %r591, 4;
// inline asm
call _rt_trace_64, (%r589, %f72, %f73, %f74, %f811, %f812, %f813, %r384, %f814, %f815, %rd157, %r591);
// inline asm
ld.local.f32 %f1213, [%rd158];
BB0_160:
ld.global.u32 %r593, [bounceFlag];
setp.gt.s32 %p106, %r593, 0;
mul.f32 %f867, %f558, 0f42C80000;
mul.f32 %f868, %f1200, %f1213;
mul.f32 %f869, %f867, %f868;
selp.f32 %f870, %f869, %f868, %p106;
mul.f32 %f1214, %f555, %f870;
mul.f32 %f1215, %f556, %f870;
mul.f32 %f1216, %f557, %f870;
BB0_161:
cvt.rn.f32.u32 %f871, %r72;
mul.f32 %f1217, %f871, %f1214;
mul.f32 %f1218, %f871, %f1215;
mul.f32 %f1219, %f871, %f1216;
BB0_162:
mov.u64 %rd164, localLights;
cvta.global.u64 %rd163, %rd164;
mov.u32 %r595, 96;
// inline asm
call (%rd159, %rd160, %rd161, %rd162), _rt_buffer_get_size_64, (%rd163, %r384, %r595);
// inline asm
cvt.u32.u64 %r596, %rd159;
setp.eq.s32 %p107, %r596, 0;
mov.u16 %rs13, 0;
@%p107 bra BB0_189;
ld.global.f32 %f879, [cameraPos];
sub.f32 %f876, %f72, %f879;
ld.global.f32 %f880, [cameraPos+4];
sub.f32 %f877, %f73, %f880;
ld.global.f32 %f881, [cameraPos+8];
sub.f32 %f878, %f74, %f881;
ld.global.u32 %r597, [tileCubemap];
mov.u32 %r598, 6;
mov.u32 %r599, 0;
// inline asm
call (%f872, %f873, %f874, %f875), _rt_texture_get_base_id, (%r597, %r598, %f876, %f877, %f878, %r599);
// inline asm
mov.b32 %r214, %f873;
setp.eq.s32 %p108, %r214, 0;
@%p108 bra BB0_189;
mov.b32 %r604, %f872;
cvt.rn.f32.u32 %f244, %r214;
mul.f32 %f884, %f3, %f244;
cvt.rzi.u32.f32 %r605, %f884;
add.s32 %r606, %r214, -1;
min.u32 %r607, %r605, %r606;
add.s32 %r608, %r607, %r604;
cvt.u64.u32 %rd167, %r608;
mov.u64 %rd177, localLightIndices;
cvta.global.u64 %rd166, %rd177;
mov.u32 %r601, 4;
mov.u64 %rd176, 0;
// inline asm
call (%rd165), _rt_buffer_get_64, (%rd166, %r384, %r601, %rd167, %rd176, %rd176, %rd176);
// inline asm
ld.u32 %rd173, [%rd165];
// inline asm
call (%rd171), _rt_buffer_get_64, (%rd163, %r384, %r595, %rd173, %rd176, %rd176, %rd176);
// inline asm
ld.v4.f32 {%f885, %f886, %f887, %f888}, [%rd171+80];
ld.v4.f32 {%f889, %f890, %f891, %f892}, [%rd171+64];
ld.v4.f32 {%f893, %f894, %f895, %f896}, [%rd171+48];
ld.v4.f32 {%f1223, %f1224, %f1225, %f900}, [%rd171+32];
ld.v4.f32 {%f901, %f902, %f903, %f904}, [%rd171+16];
ld.v4.f32 {%f905, %f906, %f907, %f908}, [%rd171];
sub.f32 %f909, %f906, %f72;
sub.f32 %f910, %f907, %f73;
sub.f32 %f911, %f908, %f74;
mul.f32 %f912, %f910, %f910;
fma.rn.f32 %f913, %f909, %f909, %f912;
fma.rn.f32 %f914, %f911, %f911, %f913;
sqrt.rn.f32 %f272, %f914;
rcp.rn.f32 %f915, %f272;
mul.f32 %f273, %f909, %f915;
mul.f32 %f274, %f910, %f915;
mul.f32 %f275, %f911, %f915;
mul.f32 %f276, %f904, %f272;
mov.f32 %f919, 0f40800000;
abs.f32 %f278, %f276;
setp.lt.f32 %p109, %f278, 0f00800000;
mul.f32 %f921, %f278, 0f4B800000;
selp.f32 %f922, 0fC3170000, 0fC2FE0000, %p109;
selp.f32 %f923, %f921, %f278, %p109;
mov.b32 %r609, %f923;
and.b32 %r610, %r609, 8388607;
or.b32 %r611, %r610, 1065353216;
mov.b32 %f924, %r611;
shr.u32 %r612, %r609, 23;
cvt.rn.f32.u32 %f925, %r612;
add.f32 %f926, %f922, %f925;
setp.gt.f32 %p110, %f924, 0f3FB504F3;
mul.f32 %f927, %f924, 0f3F000000;
add.f32 %f928, %f926, 0f3F800000;
selp.f32 %f929, %f927, %f924, %p110;
selp.f32 %f930, %f928, %f926, %p110;
add.f32 %f931, %f929, 0fBF800000;
add.f32 %f883, %f929, 0f3F800000;
// inline asm
rcp.approx.ftz.f32 %f882,%f883;
// inline asm
add.f32 %f932, %f931, %f931;
mul.f32 %f933, %f882, %f932;
mul.f32 %f934, %f933, %f933;
mov.f32 %f935, 0f3C4CAF63;
mov.f32 %f936, 0f3B18F0FE;
fma.rn.f32 %f937, %f936, %f934, %f935;
mov.f32 %f938, 0f3DAAAABD;
fma.rn.f32 %f939, %f937, %f934, %f938;
mul.rn.f32 %f940, %f939, %f934;
mul.rn.f32 %f941, %f940, %f933;
sub.f32 %f942, %f931, %f933;
neg.f32 %f943, %f933;
add.f32 %f944, %f942, %f942;
fma.rn.f32 %f945, %f943, %f931, %f944;
mul.rn.f32 %f946, %f882, %f945;
add.f32 %f947, %f941, %f933;
sub.f32 %f948, %f933, %f947;
add.f32 %f949, %f941, %f948;
add.f32 %f950, %f946, %f949;
add.f32 %f951, %f947, %f950;
sub.f32 %f952, %f947, %f951;
add.f32 %f953, %f950, %f952;
mov.f32 %f954, 0f3F317200;
mul.rn.f32 %f955, %f930, %f954;
mov.f32 %f956, 0f35BFBE8E;
mul.rn.f32 %f957, %f930, %f956;
add.f32 %f958, %f955, %f951;
sub.f32 %f959, %f955, %f958;
add.f32 %f960, %f951, %f959;
add.f32 %f961, %f953, %f960;
add.f32 %f962, %f957, %f961;
add.f32 %f963, %f958, %f962;
sub.f32 %f964, %f958, %f963;
add.f32 %f965, %f962, %f964;
mul.rn.f32 %f279, %f919, %f963;
neg.f32 %f966, %f279;
fma.rn.f32 %f967, %f919, %f963, %f966;
fma.rn.f32 %f968, %f919, %f965, %f967;
mov.f32 %f969, 0f00000000;
fma.rn.f32 %f280, %f969, %f963, %f968;
add.rn.f32 %f281, %f279, %f280;
mov.b32 %r613, %f281;
setp.eq.s32 %p1, %r613, 1118925336;
add.s32 %r614, %r613, -1;
mov.b32 %f970, %r614;
selp.f32 %f971, %f970, %f281, %p1;
mul.f32 %f972, %f971, 0f3FB8AA3B;
cvt.rzi.f32.f32 %f973, %f972;
mov.f32 %f974, 0fBF317200;
fma.rn.f32 %f975, %f973, %f974, %f971;
mov.f32 %f976, 0fB5BFBE8E;
fma.rn.f32 %f977, %f973, %f976, %f975;
mul.f32 %f978, %f977, 0f3FB8AA3B;
ex2.approx.ftz.f32 %f979, %f978;
add.f32 %f980, %f973, 0f00000000;
ex2.approx.f32 %f981, %f980;
mul.f32 %f982, %f979, %f981;
setp.lt.f32 %p111, %f971, 0fC2D20000;
selp.f32 %f983, 0f00000000, %f982, %p111;
setp.gt.f32 %p112, %f971, 0f42D20000;
selp.f32 %f1220, 0f7F800000, %f983, %p112;
setp.eq.f32 %p113, %f1220, 0f7F800000;
@%p113 bra BB0_166;
neg.f32 %f984, %f281;
add.rn.f32 %f985, %f279, %f984;
add.rn.f32 %f986, %f985, %f280;
add.f32 %f987, %f986, 0f37000000;
selp.f32 %f988, %f987, %f986, %p1;
fma.rn.f32 %f1220, %f1220, %f988, %f1220;
BB0_166:
mov.f32 %f1151, 0f40000000;
cvt.rzi.f32.f32 %f1150, %f1151;
add.f32 %f1149, %f1150, %f1150;
mov.f32 %f1148, 0f40800000;
sub.f32 %f1147, %f1148, %f1149;
abs.f32 %f1146, %f1147;
setp.lt.f32 %p114, %f276, 0f00000000;
setp.eq.f32 %p115, %f1146, 0f3F800000;
and.pred %p2, %p114, %p115;
mov.b32 %r615, %f1220;
xor.b32 %r616, %r615, -2147483648;
mov.b32 %f989, %r616;
selp.f32 %f1222, %f989, %f1220, %p2;
setp.eq.f32 %p116, %f276, 0f00000000;
@%p116 bra BB0_169;
bra.uni BB0_167;
BB0_169:
add.f32 %f992, %f276, %f276;
selp.f32 %f1222, %f992, 0f00000000, %p115;
bra.uni BB0_170;
BB0_167:
setp.geu.f32 %p117, %f276, 0f00000000;
@%p117 bra BB0_170;
mov.f32 %f1154, 0f40800000;
cvt.rzi.f32.f32 %f991, %f1154;
setp.neu.f32 %p118, %f991, 0f40800000;
selp.f32 %f1222, 0f7FFFFFFF, %f1222, %p118;
BB0_170:
add.f32 %f993, %f278, 0f40800000;
mov.b32 %r617, %f993;
setp.lt.s32 %p120, %r617, 2139095040;
@%p120 bra BB0_175;
setp.gtu.f32 %p121, %f278, 0f7F800000;
@%p121 bra BB0_174;
bra.uni BB0_172;
BB0_174:
add.f32 %f1222, %f276, 0f40800000;
bra.uni BB0_175;
BB0_172:
setp.neu.f32 %p122, %f278, 0f7F800000;
@%p122 bra BB0_175;
selp.f32 %f1222, 0fFF800000, 0f7F800000, %p2;
BB0_175:
add.u64 %rd232, %SP, 32;
cvta.to.local.u64 %rd231, %rd232;
add.s64 %rd230, %rd231, 12;
mul.f32 %f994, %f902, %f272;
mov.f32 %f1227, 0f3F800000;
sub.f32 %f996, %f1227, %f1222;
setp.eq.f32 %p123, %f276, 0f3F800000;
selp.f32 %f997, 0f00000000, %f996, %p123;
cvt.sat.f32.f32 %f998, %f997;
fma.rn.f32 %f999, %f994, %f994, %f903;
div.rn.f32 %f1226, %f998, %f999;
ld.local.f32 %f1000, [%rd230+-12];
ld.local.f32 %f1001, [%rd230+-8];
mul.f32 %f1002, %f274, %f1001;
fma.rn.f32 %f1003, %f273, %f1000, %f1002;
ld.local.f32 %f1004, [%rd230+-4];
fma.rn.f32 %f1005, %f275, %f1004, %f1003;
cvt.sat.f32.f32 %f308, %f1005;
setp.eq.f32 %p124, %f905, 0f3F800000;
@%p124 bra BB0_181;
bra.uni BB0_176;
BB0_181:
setp.leu.f32 %p128, %f900, 0f00000000;
@%p128 bra BB0_183;
mov.u32 %r686, 2;
mul.f32 %f1036, %f885, %f273;
mul.f32 %f1037, %f886, %f274;
neg.f32 %f1038, %f1037;
sub.f32 %f1039, %f1038, %f1036;
mul.f32 %f1040, %f887, %f275;
sub.f32 %f1041, %f1039, %f1040;
setp.gt.f32 %p129, %f1041, 0f00000000;
selp.f32 %f1042, 0f3F800000, 0f00000000, %p129;
mul.f32 %f1043, %f894, %f274;
fma.rn.f32 %f1044, %f893, %f273, %f1043;
mul.f32 %f1045, %f890, %f274;
fma.rn.f32 %f1046, %f889, %f273, %f1045;
fma.rn.f32 %f1047, %f895, %f275, %f1044;
fma.rn.f32 %f1048, %f891, %f275, %f1046;
fma.rn.f32 %f1049, %f896, %f1047, 0f3F000000;
mov.f32 %f1050, 0f3F800000;
sub.f32 %f1032, %f1050, %f1049;
fma.rn.f32 %f1033, %f896, %f1048, 0f3F000000;
cvt.rzi.s32.f32 %r621, %f900;
mov.f32 %f1035, 0f00000000;
// inline asm
call (%f1028, %f1029, %f1030, %f1031), _rt_texture_get_f_id, (%r621, %r686, %f1032, %f1033, %f1035, %f1035);
// inline asm
mul.f32 %f1051, %f1042, %f1028;
mul.f32 %f1052, %f1042, %f1029;
mul.f32 %f1053, %f1042, %f1030;
mul.f32 %f1223, %f1223, %f1051;
mul.f32 %f1224, %f1224, %f1052;
mul.f32 %f1225, %f1225, %f1053;
bra.uni BB0_183;
BB0_176:
setp.eq.f32 %p125, %f905, 0f40000000;
@%p125 bra BB0_179;
bra.uni BB0_177;
BB0_179:
setp.leu.f32 %p127, %f900, 0f00000000;
@%p127 bra BB0_183;
mov.u32 %r685, 0;
mov.u32 %r684, 6;
mul.f32 %f1022, %f894, %f274;
fma.rn.f32 %f1023, %f893, %f273, %f1022;
mul.f32 %f1024, %f890, %f274;
fma.rn.f32 %f1025, %f889, %f273, %f1024;
mul.f32 %f1026, %f886, %f274;
fma.rn.f32 %f1027, %f885, %f273, %f1026;
fma.rn.f32 %f1019, %f895, %f275, %f1023;
fma.rn.f32 %f1020, %f891, %f275, %f1025;
fma.rn.f32 %f1021, %f887, %f275, %f1027;
cvt.rzi.s32.f32 %r618, %f900;
// inline asm
call (%f1015, %f1016, %f1017, %f1018), _rt_texture_get_base_id, (%r618, %r684, %f1019, %f1020, %f1021, %r685);
// inline asm
mul.f32 %f1223, %f1223, %f1015;
mul.f32 %f1224, %f1224, %f1016;
mul.f32 %f1225, %f1225, %f1017;
bra.uni BB0_183;
BB0_177:
setp.neu.f32 %p126, %f905, 0f40800000;
@%p126 bra BB0_183;
mul.f32 %f1006, %f885, %f273;
mul.f32 %f1007, %f886, %f274;
neg.f32 %f1008, %f1007;
sub.f32 %f1009, %f1008, %f1006;
mul.f32 %f1010, %f887, %f275;
sub.f32 %f1011, %f1009, %f1010;
fma.rn.f32 %f1012, %f900, %f1011, %f896;
cvt.sat.f32.f32 %f1013, %f1012;
mul.f32 %f1014, %f1013, %f1013;
mul.f32 %f1226, %f1226, %f1014;
BB0_183:
mov.f32 %f1228, 0f00000000;
max.f32 %f1057, %f1223, %f1224;
max.f32 %f1058, %f1057, %f1225;
mul.f32 %f320, %f308, %f1226;
mul.f32 %f1059, %f320, %f1058;
setp.lt.f32 %p130, %f1059, 0f3727C5AC;
mov.f32 %f1229, %f1228;
mov.f32 %f1230, %f1228;
@%p130 bra BB0_187;
add.u64 %rd179, %SP, 0;
cvta.to.local.u64 %rd40, %rd179;
mov.u32 %r623, 1065353216;
st.local.u32 [%rd40], %r623;
setp.leu.f32 %p131, %f888, 0f00000000;
@%p131 bra BB0_186;
mov.u32 %r676, 4;
mov.f32 %f1153, 0f38D1B717;
mov.u32 %r675, 1;
fma.rn.f32 %f1069, %f1, 0f40000000, 0fBF800000;
fma.rn.f32 %f1070, %f2, 0f40000000, 0fBF800000;
fma.rn.f32 %f1071, %f3, 0f40000000, 0fBF800000;
fma.rn.f32 %f1072, %f1069, %f901, %f906;
fma.rn.f32 %f1073, %f1070, %f901, %f907;
fma.rn.f32 %f1074, %f1071, %f901, %f908;
sub.f32 %f1075, %f1072, %f72;
sub.f32 %f1076, %f1073, %f73;
sub.f32 %f1077, %f1074, %f74;
mul.f32 %f1078, %f1076, %f1076;
fma.rn.f32 %f1079, %f1075, %f1075, %f1078;
fma.rn.f32 %f1080, %f1077, %f1077, %f1079;
sqrt.rn.f32 %f1068, %f1080;
rcp.rn.f32 %f1081, %f1068;
mul.f32 %f1064, %f1075, %f1081;
mul.f32 %f1065, %f1076, %f1081;
mul.f32 %f1066, %f1077, %f1081;
mul.f32 %f1082, %f72, 0f3456BF95;
abs.f32 %f1083, %f1082;
mul.f32 %f1084, %f73, 0f3456BF95;
abs.f32 %f1085, %f1084;
mul.f32 %f1086, %f74, 0f3456BF95;
abs.f32 %f1087, %f1086;
max.f32 %f1088, %f1083, %f1085;
max.f32 %f1089, %f1088, %f1087;
max.f32 %f1067, %f1089, %f1153;
ld.global.u32 %r624, [root];
// inline asm
call _rt_trace_64, (%r624, %f72, %f73, %f74, %f1064, %f1065, %f1066, %r675, %f1067, %f1068, %rd179, %r676);
// inline asm
ld.local.f32 %f1227, [%rd40];
BB0_186:
ld.global.u32 %r627, [bounceFlag];
setp.gt.s32 %p132, %r627, 0;
mul.f32 %f1091, %f320, %f1227;
mul.f32 %f1092, %f892, %f1091;
selp.f32 %f1093, %f1092, %f1091, %p132;
mul.f32 %f1228, %f1223, %f1093;
mul.f32 %f1229, %f1224, %f1093;
mul.f32 %f1230, %f1225, %f1093;
BB0_187:
mov.u16 %rs13, 0;
fma.rn.f32 %f1217, %f244, %f1228, %f1217;
fma.rn.f32 %f1218, %f244, %f1229, %f1218;
fma.rn.f32 %f1219, %f244, %f1230, %f1219;
BB0_189:
mov.u64 %rd234, GBufferAtten;
cvta.global.u64 %rd233, %rd234;
mov.u32 %r678, 12;
mov.u32 %r677, 2;
abs.f32 %f1103, %f1217;
setp.gtu.f32 %p134, %f1103, 0f7F800000;
selp.f32 %f1104, 0f00000000, %f1217, %p134;
abs.f32 %f1105, %f1218;
setp.gtu.f32 %p135, %f1105, 0f7F800000;
selp.f32 %f1106, 0f00000000, %f1218, %p135;
abs.f32 %f1107, %f1219;
setp.gtu.f32 %p136, %f1107, 0f7F800000;
selp.f32 %f1108, 0f00000000, %f1219, %p136;
fma.rn.f32 %f1243, %f1234, %f1104, %f1231;
fma.rn.f32 %f1244, %f1235, %f1106, %f1232;
fma.rn.f32 %f1245, %f1236, %f1108, %f1233;
ld.global.v2.u32 {%r633, %r634}, [pixelID];
cvt.u64.u32 %rd183, %r633;
cvt.u64.u32 %rd184, %r634;
mov.u64 %rd192, 0;
// inline asm
call (%rd181), _rt_buffer_get_64, (%rd233, %r677, %r678, %rd183, %rd184, %rd192, %rd192);
// inline asm
st.f32 [%rd181+8], %f1236;
st.f32 [%rd181+4], %f1235;
st.f32 [%rd181], %f1234;
ld.global.v2.u32 {%r637, %r638}, [pixelID];
cvt.u64.u32 %rd189, %r637;
cvt.u64.u32 %rd190, %r638;
mov.u64 %rd194, accumBufferAdd;
cvta.global.u64 %rd188, %rd194;
// inline asm
call (%rd187), _rt_buffer_get_64, (%rd188, %r677, %r678, %rd189, %rd190, %rd192, %rd192);
// inline asm
ld.global.u32 %r641, [bounceFlag];
setp.eq.s32 %p137, %r641, 0;
@%p137 bra BB0_191;
ld.f32 %f1109, [%rd187+8];
ld.f32 %f1110, [%rd187+4];
ld.f32 %f1111, [%rd187];
add.f32 %f1243, %f1243, %f1111;
add.f32 %f1244, %f1244, %f1110;
add.f32 %f1245, %f1245, %f1109;
bra.uni BB0_196;
BB0_191:
ld.global.u32 %r215, [firstSkylight];
setp.lt.s32 %p138, %r215, 0;
setp.eq.s16 %p139, %rs13, 0;
or.pred %p140, %p138, %p139;
@%p140 bra BB0_196;
cvt.s64.s32 %rd197, %r215;
mov.u64 %rd201, globalLights;
cvta.global.u64 %rd196, %rd201;
mov.u32 %r642, 1;
mov.u32 %r643, 48;
// inline asm
call (%rd195), _rt_buffer_get_64, (%rd196, %r642, %r643, %rd197, %rd192, %rd192, %rd192);
// inline asm
ld.v4.f32 {%f1116, %f1117, %f1118, %f1119}, [%rd195+16];
ld.v4.f32 {%f1120, %f1121, %f1122, %f1123}, [%rd195];
setp.neu.f32 %p141, %f1116, 0f00000000;
setp.leu.f32 %p142, %f63, 0f00000000;
and.pred %p143, %p141, %p142;
@%p143 bra BB0_196;
mov.f32 %f1240, 0f3F800000;
setp.leu.f32 %p144, %f1121, 0f00000000;
mov.f32 %f1241, %f1240;
mov.f32 %f1242, %f1240;
@%p144 bra BB0_195;
cvt.rzi.s32.f32 %r644, %f1122;
neg.f32 %f1133, %f62;
neg.f32 %f1132, %f63;
neg.f32 %f1131, %f64;
mov.u32 %r645, 6;
mov.u32 %r646, 0;
// inline asm
call (%f1242, %f1241, %f1240, %f1130), _rt_texture_get_base_id, (%r644, %r645, %f1131, %f1132, %f1133, %r646);
// inline asm
BB0_195:
mul.f32 %f1243, %f1117, %f1242;
mul.f32 %f1244, %f1118, %f1241;
mul.f32 %f1245, %f1119, %f1240;
BB0_196:
mov.u32 %r680, 12;
mov.u32 %r679, 2;
ld.global.v2.u32 {%r651, %r652}, [pixelID];
cvt.u64.u32 %rd204, %r651;
cvt.u64.u32 %rd205, %r652;
// inline asm
call (%rd202), _rt_buffer_get_64, (%rd188, %r679, %r680, %rd204, %rd205, %rd192, %rd192);
// inline asm
st.f32 [%rd202+8], %f1245;
st.f32 [%rd202+4], %f1244;
st.f32 [%rd202], %f1243;
ld.global.v2.u32 {%r655, %r656}, [pixelID];
cvt.u64.u32 %rd210, %r655;
cvt.u64.u32 %rd211, %r656;
mov.u64 %rd215, accumBufferLerp;
cvta.global.u64 %rd209, %rd215;
// inline asm
call (%rd208), _rt_buffer_get_64, (%rd209, %r679, %r680, %rd210, %rd211, %rd192, %rd192);
// inline asm
ld.f32 %f1134, [%rd208+8];
ld.f32 %f1135, [%rd208+4];
ld.f32 %f1136, [%rd208];
sub.f32 %f1137, %f1243, %f1136;
sub.f32 %f1138, %f1244, %f1135;
sub.f32 %f1139, %f1245, %f1134;
ld.global.f32 %f1140, [integration];
fma.rn.f32 %f371, %f1140, %f1137, %f1136;
fma.rn.f32 %f372, %f1140, %f1138, %f1135;
fma.rn.f32 %f373, %f1140, %f1139, %f1134;
ld.global.u32 %r659, [finalBounceFlag];
setp.eq.s32 %p145, %r659, 0;
@%p145 bra BB0_198;
mov.u32 %r682, 12;
mov.u32 %r681, 2;
ld.global.v2.u32 {%r662, %r663}, [pixelID];
cvt.u64.u32 %rd218, %r662;
cvt.u64.u32 %rd219, %r663;
// inline asm
call (%rd216), _rt_buffer_get_64, (%rd209, %r681, %r682, %rd218, %rd219, %rd192, %rd192);
// inline asm
st.f32 [%rd216+8], %f373;
st.f32 [%rd216+4], %f372;
st.f32 [%rd216], %f371;
BB0_198:
mov.u32 %r683, 2;
ld.global.u32 %r668, [resolution+4];
add.s32 %r669, %r668, -1;
ld.global.v2.u32 {%r670, %r671}, [pixelID];
sub.s32 %r674, %r669, %r671;
ld.global.f32 %f1141, [cameraExp];
cvt.u64.u32 %rd225, %r670;
cvt.u64.u32 %rd226, %r674;
mov.u64 %rd229, colorOutput;
cvta.global.u64 %rd224, %rd229;
mov.u32 %r667, 16;
// inline asm
call (%rd223), _rt_buffer_get_64, (%rd224, %r683, %r667, %rd225, %rd226, %rd192, %rd192);
// inline asm
mul.f32 %f1142, %f373, %f1141;
mul.f32 %f1143, %f372, %f1141;
mul.f32 %f1144, %f371, %f1141;
mov.f32 %f1145, 0f3F800000;
st.v4.f32 [%rd223], {%f1144, %f1143, %f1142, %f1145};
ret;
}
Reference in New Issue View Git Blame Copy Permalink