Files
PrinceOfGlory/Assets/Editor/x64/Bakery/lmPreviewAO.ptx
kridoo 6e91a0c7f0 111
2025-09-15 17:32:08 +08:00

3382 lines
99 KiB
Plaintext

//
// 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 accumBufferLerpAO[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 .f32 aoRadius;
.global .align 4 .f32 aoIntensity;
.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 4 .b8 _ZN21rti_internal_typeinfo8aoRadiusE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
.global .align 4 .b8 _ZN21rti_internal_typeinfo11aoIntensityE[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 8 .b8 _ZN21rti_internal_typename8aoRadiusE[6] = {102, 108, 111, 97, 116, 0};
.global .align 8 .b8 _ZN21rti_internal_typename11aoIntensityE[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 4 .u32 _ZN21rti_internal_typeenum8aoRadiusE = 4919;
.global .align 4 .u32 _ZN21rti_internal_typeenum11aoIntensityE = 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 _ZN21rti_internal_semantic8aoRadiusE[1];
.global .align 1 .b8 _ZN21rti_internal_semantic11aoIntensityE[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];
.global .align 1 .b8 _ZN23rti_internal_annotation8aoRadiusE[1];
.global .align 1 .b8 _ZN23rti_internal_annotation11aoIntensityE[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<225>;
.reg .b16 %rs<15>;
.reg .f32 %f<1717>;
.reg .b32 %r<980>;
.reg .b64 %rd<304>;
mov.u64 %rd303, __local_depot0;
cvta.local.u64 %SP, %rd303;
ld.global.v2.u32 {%r298, %r299}, [pixelID];
ld.global.v4.u32 {%r301, %r302, %r303, %r304}, [rndOffset];
add.s32 %r306, %r301, %r298;
and.b32 %r307, %r306, 255;
add.s32 %r310, %r302, %r299;
and.b32 %r311, %r310, 255;
add.s32 %r313, %r303, %r298;
and.b32 %r314, %r313, 255;
add.s32 %r316, %r304, %r299;
and.b32 %r317, %r316, 255;
cvt.u64.u32 %rd72, %r307;
cvt.u64.u32 %rd73, %r311;
mov.u64 %rd88, rnd_seeds;
cvta.global.u64 %rd59, %rd88;
mov.u32 %r296, 2;
mov.u32 %r297, 3;
mov.u64 %rd294, 0;
// inline asm
call (%rd58), _rt_buffer_get_64, (%rd59, %r296, %r297, %rd72, %rd73, %rd294, %rd294);
// inline asm
ld.u8 %rs4, [%rd58];
cvt.rn.f32.u16 %f467, %rs4;
div.rn.f32 %f1, %f467, 0f437F0000;
// inline asm
call (%rd64), _rt_buffer_get_64, (%rd59, %r296, %r297, %rd72, %rd73, %rd294, %rd294);
// inline asm
ld.u8 %rs5, [%rd64+1];
cvt.rn.f32.u16 %f468, %rs5;
div.rn.f32 %f2, %f468, 0f437F0000;
// inline asm
call (%rd70), _rt_buffer_get_64, (%rd59, %r296, %r297, %rd72, %rd73, %rd294, %rd294);
// inline asm
ld.u8 %rs6, [%rd70+2];
cvt.rn.f32.u16 %f469, %rs6;
div.rn.f32 %f3, %f469, 0f437F0000;
cvt.u64.u32 %rd84, %r314;
cvt.u64.u32 %rd85, %r317;
// inline asm
call (%rd76), _rt_buffer_get_64, (%rd59, %r296, %r297, %rd84, %rd85, %rd294, %rd294);
// inline asm
ld.u8 %rs1, [%rd76];
// inline asm
call (%rd82), _rt_buffer_get_64, (%rd59, %r296, %r297, %rd84, %rd85, %rd294, %rd294);
// inline asm
ld.global.u32 %r318, [bounceFlag];
setp.eq.s32 %p6, %r318, 0;
@%p6 bra BB0_48;
ld.u8 %rs7, [%rd82+1];
cvt.rn.f32.u16 %f470, %rs7;
div.rn.f32 %f471, %f470, 0f437F0000;
cvt.rn.f32.u16 %f472, %rs1;
div.rn.f32 %f473, %f472, 0f437F0000;
ld.global.v2.u32 {%r323, %r324}, [pixelID];
cvt.u64.u32 %rd91, %r323;
cvt.u64.u32 %rd92, %r324;
mov.u64 %rd101, GBufferPos;
cvta.global.u64 %rd90, %rd101;
mov.u32 %r322, 12;
// inline asm
call (%rd89), _rt_buffer_get_64, (%rd90, %r296, %r322, %rd91, %rd92, %rd294, %rd294);
// inline asm
ld.f32 %f1617, [%rd89+8];
ld.f32 %f1616, [%rd89+4];
ld.f32 %f1615, [%rd89];
ld.global.v2.u32 {%r327, %r328}, [pixelID];
cvt.u64.u32 %rd97, %r327;
cvt.u64.u32 %rd98, %r328;
mov.u64 %rd102, GBufferNormal;
cvta.global.u64 %rd96, %rd102;
// inline asm
call (%rd95), _rt_buffer_get_64, (%rd96, %r296, %r322, %rd97, %rd98, %rd294, %rd294);
// inline asm
ld.f32 %f9, [%rd95+4];
ld.f32 %f8, [%rd95+8];
ld.f32 %f7, [%rd95];
abs.f32 %f474, %f7;
abs.f32 %f475, %f8;
setp.gt.f32 %p7, %f474, %f475;
neg.f32 %f476, %f9;
neg.f32 %f477, %f8;
selp.f32 %f478, %f476, 0f00000000, %p7;
selp.f32 %f479, %f7, %f477, %p7;
selp.f32 %f480, 0f00000000, %f9, %p7;
mul.f32 %f481, %f479, %f479;
fma.rn.f32 %f482, %f478, %f478, %f481;
fma.rn.f32 %f483, %f480, %f480, %f482;
sqrt.rn.f32 %f484, %f483;
rcp.rn.f32 %f485, %f484;
mul.f32 %f10, %f478, %f485;
mul.f32 %f11, %f479, %f485;
mul.f32 %f12, %f480, %f485;
sqrt.rn.f32 %f13, %f473;
mul.f32 %f1606, %f471, 0f40C90FDB;
add.u64 %rd103, %SP, 32;
cvta.to.local.u64 %rd2, %rd103;
abs.f32 %f15, %f1606;
setp.neu.f32 %p8, %f15, 0f7F800000;
mov.f32 %f1600, %f1606;
@%p8 bra BB0_3;
mov.f32 %f486, 0f00000000;
mul.rn.f32 %f1600, %f1606, %f486;
BB0_3:
mul.f32 %f487, %f1600, 0f3F22F983;
cvt.rni.s32.f32 %r909, %f487;
cvt.rn.f32.s32 %f488, %r909;
neg.f32 %f489, %f488;
mov.f32 %f490, 0f3FC90FDA;
fma.rn.f32 %f491, %f489, %f490, %f1600;
mov.f32 %f492, 0f33A22168;
fma.rn.f32 %f493, %f489, %f492, %f491;
mov.f32 %f494, 0f27C234C5;
fma.rn.f32 %f1601, %f489, %f494, %f493;
abs.f32 %f495, %f1600;
add.s64 %rd3, %rd2, 24;
setp.leu.f32 %p9, %f495, 0f47CE4780;
@%p9 bra BB0_14;
mov.b32 %r2, %f1600;
shr.u32 %r3, %r2, 23;
shl.b32 %r333, %r2, 8;
or.b32 %r4, %r333, -2147483648;
mov.u32 %r901, 0;
mov.u64 %rd287, __cudart_i2opi_f;
mov.u32 %r900, -6;
mov.u64 %rd288, %rd2;
BB0_5:
.pragma "nounroll";
ld.const.u32 %r336, [%rd287];
// inline asm
{
mad.lo.cc.u32 %r334, %r336, %r4, %r901;
madc.hi.u32 %r901, %r336, %r4, 0;
}
// inline asm
st.local.u32 [%rd288], %r334;
add.s64 %rd288, %rd288, 4;
add.s64 %rd287, %rd287, 4;
add.s32 %r900, %r900, 1;
setp.ne.s32 %p10, %r900, 0;
@%p10 bra BB0_5;
and.b32 %r339, %r3, 255;
add.s32 %r340, %r339, -128;
shr.u32 %r341, %r340, 5;
and.b32 %r9, %r2, -2147483648;
st.local.u32 [%rd3], %r901;
mov.u32 %r342, 6;
sub.s32 %r343, %r342, %r341;
mul.wide.s32 %rd105, %r343, 4;
add.s64 %rd8, %rd2, %rd105;
ld.local.u32 %r902, [%rd8];
ld.local.u32 %r903, [%rd8+-4];
and.b32 %r12, %r3, 31;
setp.eq.s32 %p11, %r12, 0;
@%p11 bra BB0_8;
mov.u32 %r344, 32;
sub.s32 %r345, %r344, %r12;
shr.u32 %r346, %r903, %r345;
shl.b32 %r347, %r902, %r12;
add.s32 %r902, %r346, %r347;
ld.local.u32 %r348, [%rd8+-8];
shr.u32 %r349, %r348, %r345;
shl.b32 %r350, %r903, %r12;
add.s32 %r903, %r349, %r350;
BB0_8:
shr.u32 %r351, %r903, 30;
shl.b32 %r352, %r902, 2;
add.s32 %r904, %r351, %r352;
shl.b32 %r18, %r903, 2;
shr.u32 %r353, %r904, 31;
shr.u32 %r354, %r902, 30;
add.s32 %r19, %r353, %r354;
setp.eq.s32 %p12, %r353, 0;
@%p12 bra BB0_9;
not.b32 %r355, %r904;
neg.s32 %r906, %r18;
setp.eq.s32 %p13, %r18, 0;
selp.u32 %r356, 1, 0, %p13;
add.s32 %r904, %r356, %r355;
xor.b32 %r905, %r9, -2147483648;
bra.uni BB0_11;
BB0_48:
ld.global.f32 %f1615, [cameraPos];
ld.global.f32 %f1616, [cameraPos+4];
ld.global.f32 %f1617, [cameraPos+8];
ld.global.v2.u32 {%r423, %r424}, [resolution];
cvt.rn.f32.u32 %f557, %r423;
cvt.rn.f32.u32 %f558, %r424;
div.rn.f32 %f559, %f557, %f558;
ld.global.v2.u32 {%r427, %r428}, [pixelID];
cvt.rn.f32.u32 %f560, %r427;
add.f32 %f561, %f1, %f560;
div.rn.f32 %f562, %f561, %f557;
cvt.rn.f32.u32 %f563, %r428;
add.f32 %f564, %f2, %f563;
div.rn.f32 %f565, %f564, %f558;
fma.rn.f32 %f566, %f562, 0f40000000, 0fBF800000;
fma.rn.f32 %f567, %f565, 0f40000000, 0fBF800000;
ld.global.f32 %f568, [cameraFov];
mul.f32 %f569, %f559, %f568;
ld.global.f32 %f570, [cameraRht];
mul.f32 %f571, %f566, %f570;
ld.global.f32 %f572, [cameraRht+4];
mul.f32 %f573, %f566, %f572;
ld.global.f32 %f574, [cameraRht+8];
mul.f32 %f575, %f566, %f574;
ld.global.f32 %f576, [cameraFwd];
fma.rn.f32 %f577, %f569, %f571, %f576;
ld.global.f32 %f578, [cameraFwd+4];
fma.rn.f32 %f579, %f569, %f573, %f578;
ld.global.f32 %f580, [cameraFwd+8];
fma.rn.f32 %f581, %f569, %f575, %f580;
ld.global.f32 %f582, [cameraUp];
mul.f32 %f583, %f567, %f582;
ld.global.f32 %f584, [cameraUp+4];
mul.f32 %f585, %f567, %f584;
ld.global.f32 %f586, [cameraUp+8];
mul.f32 %f587, %f567, %f586;
fma.rn.f32 %f588, %f568, %f583, %f577;
fma.rn.f32 %f589, %f568, %f585, %f579;
fma.rn.f32 %f590, %f568, %f587, %f581;
mul.f32 %f591, %f589, %f589;
fma.rn.f32 %f592, %f588, %f588, %f591;
fma.rn.f32 %f593, %f590, %f590, %f592;
sqrt.rn.f32 %f594, %f593;
rcp.rn.f32 %f595, %f594;
mul.f32 %f62, %f588, %f595;
mul.f32 %f63, %f589, %f595;
mul.f32 %f64, %f595, %f590;
bra.uni BB0_49;
BB0_9:
mov.u32 %r905, %r9;
mov.u32 %r906, %r18;
BB0_11:
clz.b32 %r908, %r904;
setp.eq.s32 %p14, %r908, 0;
shl.b32 %r357, %r904, %r908;
mov.u32 %r358, 32;
sub.s32 %r359, %r358, %r908;
shr.u32 %r360, %r906, %r359;
add.s32 %r361, %r360, %r357;
selp.b32 %r27, %r904, %r361, %p14;
mov.u32 %r362, -921707870;
mul.hi.u32 %r907, %r27, %r362;
setp.eq.s32 %p15, %r9, 0;
neg.s32 %r363, %r19;
selp.b32 %r909, %r19, %r363, %p15;
setp.lt.s32 %p16, %r907, 1;
@%p16 bra BB0_13;
mul.lo.s32 %r364, %r27, -921707870;
shr.u32 %r365, %r364, 31;
shl.b32 %r366, %r907, 1;
add.s32 %r907, %r365, %r366;
add.s32 %r908, %r908, 1;
BB0_13:
mov.u32 %r367, 126;
sub.s32 %r368, %r367, %r908;
shl.b32 %r369, %r368, 23;
add.s32 %r370, %r907, 1;
shr.u32 %r371, %r370, 7;
add.s32 %r372, %r371, 1;
shr.u32 %r373, %r372, 1;
add.s32 %r374, %r373, %r369;
or.b32 %r375, %r374, %r905;
mov.b32 %f1601, %r375;
BB0_14:
mul.rn.f32 %f21, %f1601, %f1601;
add.s32 %r35, %r909, 1;
and.b32 %r36, %r35, 1;
setp.eq.s32 %p17, %r36, 0;
@%p17 bra BB0_16;
mov.f32 %f496, 0fBAB6061A;
mov.f32 %f497, 0f37CCF5CE;
fma.rn.f32 %f1602, %f497, %f21, %f496;
bra.uni BB0_17;
BB0_16:
mov.f32 %f498, 0f3C08839E;
mov.f32 %f499, 0fB94CA1F9;
fma.rn.f32 %f1602, %f499, %f21, %f498;
BB0_17:
@%p17 bra BB0_19;
mov.f32 %f500, 0f3D2AAAA5;
fma.rn.f32 %f501, %f1602, %f21, %f500;
mov.f32 %f502, 0fBF000000;
fma.rn.f32 %f1603, %f501, %f21, %f502;
bra.uni BB0_20;
BB0_19:
mov.f32 %f503, 0fBE2AAAA3;
fma.rn.f32 %f504, %f1602, %f21, %f503;
mov.f32 %f505, 0f00000000;
fma.rn.f32 %f1603, %f504, %f21, %f505;
BB0_20:
fma.rn.f32 %f1604, %f1603, %f1601, %f1601;
@%p17 bra BB0_22;
mov.f32 %f506, 0f3F800000;
fma.rn.f32 %f1604, %f1603, %f21, %f506;
BB0_22:
and.b32 %r376, %r35, 2;
setp.eq.s32 %p20, %r376, 0;
@%p20 bra BB0_24;
mov.f32 %f507, 0f00000000;
mov.f32 %f508, 0fBF800000;
fma.rn.f32 %f1604, %f1604, %f508, %f507;
BB0_24:
@%p8 bra BB0_26;
mov.f32 %f509, 0f00000000;
mul.rn.f32 %f1606, %f1606, %f509;
BB0_26:
mul.f32 %f510, %f1606, 0f3F22F983;
cvt.rni.s32.f32 %r919, %f510;
cvt.rn.f32.s32 %f511, %r919;
neg.f32 %f512, %f511;
fma.rn.f32 %f514, %f512, %f490, %f1606;
fma.rn.f32 %f516, %f512, %f492, %f514;
fma.rn.f32 %f1607, %f512, %f494, %f516;
abs.f32 %f518, %f1606;
setp.leu.f32 %p22, %f518, 0f47CE4780;
@%p22 bra BB0_37;
mov.b32 %r38, %f1606;
shr.u32 %r39, %r38, 23;
shl.b32 %r379, %r38, 8;
or.b32 %r40, %r379, -2147483648;
mov.u32 %r911, 0;
mov.u64 %rd289, __cudart_i2opi_f;
mov.u32 %r910, -6;
mov.u64 %rd290, %rd2;
BB0_28:
.pragma "nounroll";
ld.const.u32 %r382, [%rd289];
// inline asm
{
mad.lo.cc.u32 %r380, %r382, %r40, %r911;
madc.hi.u32 %r911, %r382, %r40, 0;
}
// inline asm
st.local.u32 [%rd290], %r380;
add.s64 %rd290, %rd290, 4;
add.s64 %rd289, %rd289, 4;
add.s32 %r910, %r910, 1;
setp.ne.s32 %p23, %r910, 0;
@%p23 bra BB0_28;
and.b32 %r385, %r39, 255;
add.s32 %r386, %r385, -128;
shr.u32 %r387, %r386, 5;
and.b32 %r45, %r38, -2147483648;
st.local.u32 [%rd3], %r911;
mov.u32 %r388, 6;
sub.s32 %r389, %r388, %r387;
mul.wide.s32 %rd107, %r389, 4;
add.s64 %rd13, %rd2, %rd107;
ld.local.u32 %r912, [%rd13];
ld.local.u32 %r913, [%rd13+-4];
and.b32 %r48, %r39, 31;
setp.eq.s32 %p24, %r48, 0;
@%p24 bra BB0_31;
mov.u32 %r390, 32;
sub.s32 %r391, %r390, %r48;
shr.u32 %r392, %r913, %r391;
shl.b32 %r393, %r912, %r48;
add.s32 %r912, %r392, %r393;
ld.local.u32 %r394, [%rd13+-8];
shr.u32 %r395, %r394, %r391;
shl.b32 %r396, %r913, %r48;
add.s32 %r913, %r395, %r396;
BB0_31:
shr.u32 %r397, %r913, 30;
shl.b32 %r398, %r912, 2;
add.s32 %r914, %r397, %r398;
shl.b32 %r54, %r913, 2;
shr.u32 %r399, %r914, 31;
shr.u32 %r400, %r912, 30;
add.s32 %r55, %r399, %r400;
setp.eq.s32 %p25, %r399, 0;
@%p25 bra BB0_32;
not.b32 %r401, %r914;
neg.s32 %r916, %r54;
setp.eq.s32 %p26, %r54, 0;
selp.u32 %r402, 1, 0, %p26;
add.s32 %r914, %r402, %r401;
xor.b32 %r915, %r45, -2147483648;
bra.uni BB0_34;
BB0_32:
mov.u32 %r915, %r45;
mov.u32 %r916, %r54;
BB0_34:
clz.b32 %r918, %r914;
setp.eq.s32 %p27, %r918, 0;
shl.b32 %r403, %r914, %r918;
mov.u32 %r404, 32;
sub.s32 %r405, %r404, %r918;
shr.u32 %r406, %r916, %r405;
add.s32 %r407, %r406, %r403;
selp.b32 %r63, %r914, %r407, %p27;
mov.u32 %r408, -921707870;
mul.hi.u32 %r917, %r63, %r408;
setp.eq.s32 %p28, %r45, 0;
neg.s32 %r409, %r55;
selp.b32 %r919, %r55, %r409, %p28;
setp.lt.s32 %p29, %r917, 1;
@%p29 bra BB0_36;
mul.lo.s32 %r410, %r63, -921707870;
shr.u32 %r411, %r410, 31;
shl.b32 %r412, %r917, 1;
add.s32 %r917, %r411, %r412;
add.s32 %r918, %r918, 1;
BB0_36:
mov.u32 %r413, 126;
sub.s32 %r414, %r413, %r918;
shl.b32 %r415, %r414, 23;
add.s32 %r416, %r917, 1;
shr.u32 %r417, %r416, 7;
add.s32 %r418, %r417, 1;
shr.u32 %r419, %r418, 1;
add.s32 %r420, %r419, %r415;
or.b32 %r421, %r420, %r915;
mov.b32 %f1607, %r421;
BB0_37:
mul.rn.f32 %f38, %f1607, %f1607;
and.b32 %r71, %r919, 1;
setp.eq.s32 %p30, %r71, 0;
@%p30 bra BB0_39;
mov.f32 %f519, 0fBAB6061A;
mov.f32 %f520, 0f37CCF5CE;
fma.rn.f32 %f1608, %f520, %f38, %f519;
bra.uni BB0_40;
BB0_39:
mov.f32 %f521, 0f3C08839E;
mov.f32 %f522, 0fB94CA1F9;
fma.rn.f32 %f1608, %f522, %f38, %f521;
BB0_40:
@%p30 bra BB0_42;
mov.f32 %f523, 0f3D2AAAA5;
fma.rn.f32 %f524, %f1608, %f38, %f523;
mov.f32 %f525, 0fBF000000;
fma.rn.f32 %f1609, %f524, %f38, %f525;
bra.uni BB0_43;
BB0_42:
mov.f32 %f526, 0fBE2AAAA3;
fma.rn.f32 %f527, %f1608, %f38, %f526;
mov.f32 %f528, 0f00000000;
fma.rn.f32 %f1609, %f527, %f38, %f528;
BB0_43:
fma.rn.f32 %f1610, %f1609, %f1607, %f1607;
@%p30 bra BB0_45;
mov.f32 %f529, 0f3F800000;
fma.rn.f32 %f1610, %f1609, %f38, %f529;
BB0_45:
and.b32 %r422, %r919, 2;
setp.eq.s32 %p33, %r422, 0;
@%p33 bra BB0_47;
mov.f32 %f530, 0f00000000;
mov.f32 %f531, 0fBF800000;
fma.rn.f32 %f1610, %f1610, %f531, %f530;
BB0_47:
mul.f32 %f532, %f13, %f1604;
mul.f32 %f533, %f532, %f532;
mov.f32 %f534, 0f3F800000;
sub.f32 %f535, %f534, %f533;
mul.f32 %f536, %f13, %f1610;
mul.f32 %f537, %f536, %f536;
sub.f32 %f538, %f535, %f537;
mov.f32 %f539, 0f00000000;
max.f32 %f540, %f539, %f538;
sqrt.rn.f32 %f541, %f540;
mul.f32 %f542, %f10, %f536;
mul.f32 %f543, %f11, %f536;
mul.f32 %f544, %f12, %f536;
mul.f32 %f545, %f8, %f11;
mul.f32 %f546, %f9, %f12;
sub.f32 %f547, %f545, %f546;
fma.rn.f32 %f548, %f547, %f532, %f542;
mul.f32 %f549, %f7, %f12;
mul.f32 %f550, %f8, %f10;
sub.f32 %f551, %f549, %f550;
fma.rn.f32 %f552, %f551, %f532, %f543;
mul.f32 %f553, %f9, %f10;
mul.f32 %f554, %f7, %f11;
sub.f32 %f555, %f553, %f554;
fma.rn.f32 %f556, %f555, %f532, %f544;
fma.rn.f32 %f62, %f7, %f541, %f548;
fma.rn.f32 %f63, %f9, %f541, %f552;
fma.rn.f32 %f64, %f8, %f541, %f556;
BB0_49:
add.u64 %rd114, %SP, 0;
cvta.to.local.u64 %rd115, %rd114;
add.s64 %rd14, %rd115, 12;
mov.u32 %r433, -1082130432;
st.local.u32 [%rd115+12], %r433;
mov.u32 %r434, 0;
st.local.u32 [%rd115], %r434;
st.local.u32 [%rd115+4], %r434;
st.local.u32 [%rd115+8], %r434;
st.local.u32 [%rd115+16], %r434;
st.local.u32 [%rd115+20], %r434;
st.local.u32 [%rd115+24], %r434;
ld.global.v2.u32 {%r435, %r436}, [pixelID];
cvt.u64.u32 %rd110, %r435;
cvt.u64.u32 %rd111, %r436;
mov.u64 %rd116, GBufferAtten;
cvta.global.u64 %rd109, %rd116;
mov.u32 %r432, 12;
// inline asm
call (%rd108), _rt_buffer_get_64, (%rd109, %r296, %r432, %rd110, %rd111, %rd294, %rd294);
// inline asm
ld.f32 %f70, [%rd108+8];
ld.f32 %f68, [%rd108+4];
ld.f32 %f69, [%rd108];
add.f32 %f596, %f69, %f68;
add.f32 %f597, %f70, %f596;
ld.global.u32 %r439, [bounceFlag];
setp.ne.s32 %p34, %r439, 0;
setp.eq.f32 %p35, %f597, 0f00000000;
and.pred %p36, %p34, %p35;
@%p36 bra BB0_51;
ld.global.u32 %r440, [root];
mov.u32 %r441, 0;
mov.f32 %f604, 0f00000000;
mov.f32 %f605, 0f6C4ECB8F;
mov.u32 %r442, 28;
// inline asm
call _rt_trace_64, (%r440, %f1615, %f1616, %f1617, %f62, %f63, %f64, %r441, %f604, %f605, %rd114, %r442);
// inline asm
BB0_51:
ld.local.f32 %f71, [%rd14];
setp.ltu.f32 %p37, %f71, 0f00000000;
@%p37 bra BB0_235;
bra.uni BB0_52;
BB0_235:
ld.global.u32 %r796, [bounceFlag];
setp.eq.s32 %p163, %r796, 0;
selp.u16 %rs14, 1, 0, %p163;
mov.f32 %f1700, 0f3F800000;
mov.f32 %f1691, 0f00000000;
mov.f32 %f1692, %f1691;
mov.f32 %f1693, %f1691;
mov.f32 %f1694, %f1691;
mov.f32 %f1695, %f1691;
mov.f32 %f1696, %f1691;
mov.f32 %f1676, %f1691;
mov.f32 %f1677, %f1691;
mov.f32 %f1678, %f1691;
bra.uni BB0_236;
BB0_52:
fma.rn.f32 %f606, %f62, %f71, %f1615;
fma.rn.f32 %f607, %f63, %f71, %f1616;
fma.rn.f32 %f608, %f64, %f71, %f1617;
mul.f32 %f609, %f606, 0f3456BF95;
mul.f32 %f610, %f607, 0f3456BF95;
mul.f32 %f611, %f608, 0f3456BF95;
abs.f32 %f612, %f609;
abs.f32 %f613, %f610;
abs.f32 %f614, %f611;
max.f32 %f615, %f612, %f613;
max.f32 %f616, %f615, %f614;
mov.f32 %f617, 0f38D1B717;
max.f32 %f618, %f616, %f617;
ld.local.f32 %f619, [%rd14+-12];
ld.local.f32 %f620, [%rd14+-8];
ld.local.f32 %f621, [%rd14+-4];
fma.rn.f32 %f72, %f618, %f619, %f606;
fma.rn.f32 %f73, %f618, %f620, %f607;
fma.rn.f32 %f74, %f618, %f621, %f608;
ld.global.v2.u32 {%r447, %r448}, [pixelID];
cvt.u64.u32 %rd120, %r447;
cvt.u64.u32 %rd121, %r448;
mov.u64 %rd130, GBufferPos;
cvta.global.u64 %rd119, %rd130;
// inline asm
call (%rd118), _rt_buffer_get_64, (%rd119, %r296, %r432, %rd120, %rd121, %rd294, %rd294);
// inline asm
st.f32 [%rd118+8], %f74;
st.f32 [%rd118+4], %f73;
st.f32 [%rd118], %f72;
ld.global.v2.u32 {%r451, %r452}, [pixelID];
cvt.u64.u32 %rd126, %r451;
cvt.u64.u32 %rd127, %r452;
mov.u64 %rd131, GBufferNormal;
cvta.global.u64 %rd125, %rd131;
// inline asm
call (%rd124), _rt_buffer_get_64, (%rd125, %r296, %r432, %rd126, %rd127, %rd294, %rd294);
// inline asm
ld.local.f32 %f622, [%rd14+-12];
ld.local.f32 %f623, [%rd14+-8];
ld.local.f32 %f624, [%rd14+-4];
st.f32 [%rd124+8], %f624;
st.f32 [%rd124+4], %f623;
st.f32 [%rd124], %f622;
ld.local.f32 %f625, [%rd14+12];
ld.local.f32 %f626, [%rd14+8];
ld.local.f32 %f627, [%rd14+4];
setp.lt.f32 %p38, %f627, 0f00000000;
neg.f32 %f628, %f627;
selp.f32 %f629, %f628, %f627, %p38;
ld.global.f32 %f630, [backFaceWeight];
selp.f32 %f631, %f630, 0f3F800000, %p38;
cvt.rmi.f32.f32 %f632, %f629;
cvt.rmi.f32.f32 %f633, %f626;
cvt.rmi.f32.f32 %f634, %f625;
sub.f32 %f635, %f629, %f632;
sub.f32 %f636, %f626, %f633;
sub.f32 %f637, %f625, %f634;
mul.f32 %f1694, %f631, %f635;
mul.f32 %f1695, %f631, %f636;
mul.f32 %f1696, %f631, %f637;
mul.f32 %f1691, %f632, 0f3B808081;
mul.f32 %f1692, %f633, 0f3B808081;
mul.f32 %f1693, %f634, 0f3B808081;
ld.global.u32 %r455, [bounceFlag];
setp.lt.s32 %p39, %r455, 1;
@%p39 bra BB0_54;
mul.f32 %f1694, %f69, %f1694;
mul.f32 %f1695, %f68, %f1695;
mul.f32 %f1696, %f70, %f1696;
mul.f32 %f1691, %f69, %f1691;
mul.f32 %f1692, %f68, %f1692;
mul.f32 %f1693, %f70, %f1693;
BB0_54:
mov.u64 %rd137, globalLights;
cvta.global.u64 %rd136, %rd137;
mov.u32 %r456, 1;
mov.u32 %r457, 48;
// inline asm
call (%rd132, %rd133, %rd134, %rd135), _rt_buffer_get_size_64, (%rd136, %r456, %r457);
// inline asm
cvt.u32.u64 %r72, %rd132;
setp.eq.s32 %p40, %r72, 0;
mov.f32 %f1676, 0f00000000;
mov.f32 %f1677, %f1676;
mov.f32 %f1678, %f1676;
@%p40 bra BB0_162;
// inline asm
call (%rd138, %rd139, %rd140, %rd141), _rt_buffer_get_size_64, (%rd136, %r456, %r457);
// inline asm
cvt.u32.u64 %r462, %rd138;
cvt.rn.f32.u32 %f641, %r462;
mul.f32 %f642, %f3, %f641;
cvt.rzi.u32.f32 %r463, %f642;
add.s32 %r464, %r462, -1;
min.u32 %r465, %r463, %r464;
cvt.u64.u32 %rd145, %r465;
// inline asm
call (%rd143), _rt_buffer_get_64, (%rd136, %r456, %r457, %rd145, %rd294, %rd294, %rd294);
// inline asm
ld.v4.f32 {%f643, %f644, %f645, %f646}, [%rd143+32];
ld.v4.f32 {%f647, %f648, %f649, %f650}, [%rd143+16];
ld.v4.f32 {%f651, %f652, %f653, %f654}, [%rd143];
setp.lt.f32 %p41, %f651, 0f3F800000;
@%p41 bra BB0_109;
bra.uni BB0_56;
BB0_109:
fma.rn.f32 %f774, %f652, 0f40000000, 0fBF800000;
fma.rn.f32 %f775, %f653, 0f40000000, 0fBF800000;
abs.f32 %f776, %f774;
mov.f32 %f1658, 0f3F800000;
sub.f32 %f778, %f1658, %f776;
abs.f32 %f779, %f775;
sub.f32 %f780, %f778, %f779;
neg.f32 %f781, %f780;
cvt.sat.f32.f32 %f782, %f781;
setp.ltu.f32 %p74, %f774, 0f00000000;
neg.f32 %f783, %f782;
selp.f32 %f784, %f782, %f783, %p74;
add.f32 %f785, %f774, %f784;
setp.ltu.f32 %p75, %f775, 0f00000000;
selp.f32 %f786, %f782, %f783, %p75;
add.f32 %f787, %f775, %f786;
mul.f32 %f788, %f787, %f787;
fma.rn.f32 %f789, %f785, %f785, %f788;
fma.rn.f32 %f790, %f780, %f780, %f789;
sqrt.rn.f32 %f791, %f790;
rcp.rn.f32 %f792, %f791;
mul.f32 %f180, %f792, %f785;
mul.f32 %f181, %f792, %f787;
mul.f32 %f182, %f780, %f792;
ld.local.f32 %f793, [%rd14+-12];
mul.f32 %f794, %f793, %f180;
ld.local.f32 %f795, [%rd14+-8];
mul.f32 %f796, %f181, %f795;
neg.f32 %f797, %f796;
sub.f32 %f798, %f797, %f794;
ld.local.f32 %f799, [%rd14+-4];
mul.f32 %f800, %f182, %f799;
sub.f32 %f1645, %f798, %f800;
setp.gt.f32 %p76, %f1645, 0f00000000;
setp.neu.f32 %p77, %f654, 0f00000000;
and.pred %p78, %p76, %p77;
@!%p78 bra BB0_111;
bra.uni BB0_110;
BB0_110:
abs.f32 %f809, %f180;
abs.f32 %f810, %f182;
setp.gt.f32 %p79, %f809, %f810;
neg.f32 %f811, %f181;
selp.f32 %f812, %f811, 0f00000000, %p79;
neg.f32 %f813, %f182;
selp.f32 %f814, %f180, %f813, %p79;
selp.f32 %f815, 0f00000000, %f181, %p79;
mul.f32 %f816, %f814, %f814;
fma.rn.f32 %f817, %f812, %f812, %f816;
fma.rn.f32 %f818, %f815, %f815, %f817;
sqrt.rn.f32 %f819, %f818;
rcp.rn.f32 %f820, %f819;
mul.f32 %f821, %f812, %f820;
mul.f32 %f822, %f814, %f820;
mul.f32 %f823, %f815, %f820;
mul.f32 %f824, %f182, %f822;
mul.f32 %f825, %f181, %f823;
sub.f32 %f826, %f824, %f825;
mul.f32 %f827, %f180, %f823;
mul.f32 %f828, %f182, %f821;
sub.f32 %f829, %f827, %f828;
mul.f32 %f830, %f181, %f821;
mul.f32 %f831, %f180, %f822;
sub.f32 %f832, %f830, %f831;
mul.f32 %f833, %f73, %f829;
fma.rn.f32 %f834, %f72, %f826, %f833;
fma.rn.f32 %f835, %f74, %f832, %f834;
fma.rn.f32 %f805, %f643, %f835, %f645;
mul.f32 %f836, %f73, %f822;
fma.rn.f32 %f837, %f72, %f821, %f836;
fma.rn.f32 %f838, %f74, %f823, %f837;
fma.rn.f32 %f806, %f644, %f838, %f646;
cvt.rzi.s32.f32 %r567, %f654;
mov.f32 %f808, 0f00000000;
// inline asm
call (%f801, %f802, %f803, %f804), _rt_texture_get_f_id, (%r567, %r296, %f805, %f806, %f808, %f808);
// inline asm
mul.f32 %f1645, %f1645, %f801;
BB0_111:
mov.f32 %f1659, 0f00000000;
setp.le.f32 %p80, %f1645, 0f00000000;
mov.f32 %f1660, %f1659;
mov.f32 %f1661, %f1659;
@%p80 bra BB0_161;
abs.f32 %f843, %f180;
abs.f32 %f844, %f182;
setp.gt.f32 %p81, %f843, %f844;
neg.f32 %f845, %f181;
selp.f32 %f846, %f845, 0f00000000, %p81;
neg.f32 %f847, %f182;
selp.f32 %f848, %f180, %f847, %p81;
selp.f32 %f849, 0f00000000, %f181, %p81;
mul.f32 %f850, %f848, %f848;
fma.rn.f32 %f851, %f846, %f846, %f850;
fma.rn.f32 %f852, %f849, %f849, %f851;
sqrt.rn.f32 %f853, %f852;
rcp.rn.f32 %f854, %f853;
mul.f32 %f186, %f846, %f854;
mul.f32 %f187, %f848, %f854;
mul.f32 %f188, %f849, %f854;
setp.ltu.f32 %p82, %f647, 0f00000000;
@%p82 bra BB0_160;
sqrt.rn.f32 %f189, %f1;
add.u64 %rd172, %SP, 32;
cvta.to.local.u64 %rd29, %rd172;
mul.f32 %f1652, %f2, 0f40C90FDB;
abs.f32 %f191, %f1652;
setp.neu.f32 %p83, %f191, 0f7F800000;
mov.f32 %f1646, %f1652;
@%p83 bra BB0_115;
mov.f32 %f855, 0f00000000;
mul.rn.f32 %f1646, %f1652, %f855;
BB0_115:
mul.f32 %f856, %f1646, 0f3F22F983;
cvt.rni.s32.f32 %r949, %f856;
cvt.rn.f32.s32 %f857, %r949;
neg.f32 %f858, %f857;
mov.f32 %f859, 0f3FC90FDA;
fma.rn.f32 %f860, %f858, %f859, %f1646;
mov.f32 %f861, 0f33A22168;
fma.rn.f32 %f862, %f858, %f861, %f860;
mov.f32 %f863, 0f27C234C5;
fma.rn.f32 %f1647, %f858, %f863, %f862;
abs.f32 %f864, %f1646;
add.s64 %rd30, %rd29, 24;
setp.leu.f32 %p84, %f864, 0f47CE4780;
@%p84 bra BB0_126;
mov.b32 %r144, %f1646;
shr.u32 %r145, %r144, 23;
shl.b32 %r571, %r144, 8;
or.b32 %r146, %r571, -2147483648;
mov.u32 %r940, 0;
mov.u64 %rd296, 0;
mov.u64 %rd295, %rd29;
mov.u32 %r941, %r940;
BB0_117:
.pragma "nounroll";
shl.b64 %rd174, %rd296, 2;
mov.u64 %rd175, __cudart_i2opi_f;
add.s64 %rd176, %rd175, %rd174;
ld.const.u32 %r574, [%rd176];
// inline asm
{
mad.lo.cc.u32 %r572, %r574, %r146, %r941;
madc.hi.u32 %r941, %r574, %r146, 0;
}
// inline asm
st.local.u32 [%rd295], %r572;
add.s32 %r940, %r940, 1;
cvt.s64.s32 %rd296, %r940;
mul.wide.s32 %rd177, %r940, 4;
add.s64 %rd295, %rd29, %rd177;
setp.ne.s32 %p85, %r940, 6;
@%p85 bra BB0_117;
and.b32 %r577, %r145, 255;
add.s32 %r578, %r577, -128;
shr.u32 %r579, %r578, 5;
and.b32 %r151, %r144, -2147483648;
st.local.u32 [%rd30], %r941;
mov.u32 %r580, 6;
sub.s32 %r581, %r580, %r579;
mul.wide.s32 %rd178, %r581, 4;
add.s64 %rd36, %rd29, %rd178;
ld.local.u32 %r942, [%rd36];
ld.local.u32 %r943, [%rd36+-4];
and.b32 %r154, %r145, 31;
setp.eq.s32 %p86, %r154, 0;
@%p86 bra BB0_120;
mov.u32 %r582, 32;
sub.s32 %r583, %r582, %r154;
shr.u32 %r584, %r943, %r583;
shl.b32 %r585, %r942, %r154;
add.s32 %r942, %r584, %r585;
ld.local.u32 %r586, [%rd36+-8];
shr.u32 %r587, %r586, %r583;
shl.b32 %r588, %r943, %r154;
add.s32 %r943, %r587, %r588;
BB0_120:
shr.u32 %r589, %r943, 30;
shl.b32 %r590, %r942, 2;
add.s32 %r944, %r589, %r590;
shl.b32 %r160, %r943, 2;
shr.u32 %r591, %r944, 31;
shr.u32 %r592, %r942, 30;
add.s32 %r161, %r591, %r592;
setp.eq.s32 %p87, %r591, 0;
@%p87 bra BB0_121;
not.b32 %r593, %r944;
neg.s32 %r946, %r160;
setp.eq.s32 %p88, %r160, 0;
selp.u32 %r594, 1, 0, %p88;
add.s32 %r944, %r594, %r593;
xor.b32 %r945, %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 %f655, %f112;
abs.f32 %f656, %f113;
setp.gt.f32 %p42, %f655, %f656;
neg.f32 %f657, %f114;
neg.f32 %f658, %f113;
selp.f32 %f659, %f657, 0f00000000, %p42;
selp.f32 %f660, %f112, %f658, %p42;
selp.f32 %f661, 0f00000000, %f114, %p42;
mul.f32 %f662, %f660, %f660;
fma.rn.f32 %f663, %f659, %f659, %f662;
fma.rn.f32 %f664, %f661, %f661, %f663;
sqrt.rn.f32 %f665, %f664;
rcp.rn.f32 %f666, %f665;
mul.f32 %f115, %f659, %f666;
mul.f32 %f116, %f660, %f666;
mul.f32 %f117, %f661, %f666;
sqrt.rn.f32 %f118, %f1;
mul.f32 %f1630, %f2, 0f40C90FDB;
abs.f32 %f120, %f1630;
setp.neu.f32 %p43, %f120, 0f7F800000;
mov.f32 %f1624, %f1630;
@%p43 bra BB0_58;
mov.f32 %f667, 0f00000000;
mul.rn.f32 %f1624, %f1630, %f667;
BB0_58:
mul.f32 %f668, %f1624, 0f3F22F983;
cvt.rni.s32.f32 %r929, %f668;
cvt.rn.f32.s32 %f669, %r929;
neg.f32 %f670, %f669;
mov.f32 %f671, 0f3FC90FDA;
fma.rn.f32 %f672, %f670, %f671, %f1624;
mov.f32 %f673, 0f33A22168;
fma.rn.f32 %f674, %f670, %f673, %f672;
mov.f32 %f675, 0f27C234C5;
fma.rn.f32 %f1625, %f670, %f675, %f674;
abs.f32 %f676, %f1624;
setp.leu.f32 %p44, %f676, 0f47CE4780;
@%p44 bra BB0_69;
mov.b32 %r74, %f1624;
shl.b32 %r468, %r74, 8;
or.b32 %r75, %r468, -2147483648;
add.u64 %rd151, %SP, 32;
cvta.to.local.u64 %rd15, %rd151;
mov.u32 %r920, 0;
mov.u64 %rd292, 0;
mov.u64 %rd291, %rd15;
mov.u32 %r921, %r920;
BB0_60:
.pragma "nounroll";
shl.b64 %rd152, %rd292, 2;
mov.u64 %rd153, __cudart_i2opi_f;
add.s64 %rd154, %rd153, %rd152;
ld.const.u32 %r471, [%rd154];
// inline asm
{
mad.lo.cc.u32 %r469, %r471, %r75, %r921;
madc.hi.u32 %r921, %r471, %r75, 0;
}
// inline asm
st.local.u32 [%rd291], %r469;
add.s32 %r920, %r920, 1;
cvt.s64.s32 %rd292, %r920;
mul.wide.s32 %rd157, %r920, 4;
add.s64 %rd291, %rd15, %rd157;
setp.ne.s32 %p45, %r920, 6;
@%p45 bra BB0_60;
bfe.u32 %r474, %r74, 23, 8;
add.s32 %r475, %r474, -128;
shr.u32 %r476, %r475, 5;
and.b32 %r80, %r74, -2147483648;
cvta.to.local.u64 %rd159, %rd151;
st.local.u32 [%rd159+24], %r921;
bfe.u32 %r81, %r74, 23, 5;
mov.u32 %r477, 6;
sub.s32 %r478, %r477, %r476;
mul.wide.s32 %rd160, %r478, 4;
add.s64 %rd20, %rd159, %rd160;
ld.local.u32 %r922, [%rd20];
ld.local.u32 %r923, [%rd20+-4];
setp.eq.s32 %p46, %r81, 0;
@%p46 bra BB0_63;
mov.u32 %r479, 32;
sub.s32 %r480, %r479, %r81;
shr.u32 %r481, %r923, %r480;
shl.b32 %r482, %r922, %r81;
add.s32 %r922, %r481, %r482;
ld.local.u32 %r483, [%rd20+-8];
shr.u32 %r484, %r483, %r480;
shl.b32 %r485, %r923, %r81;
add.s32 %r923, %r484, %r485;
BB0_63:
shr.u32 %r486, %r923, 30;
shl.b32 %r487, %r922, 2;
add.s32 %r924, %r486, %r487;
shl.b32 %r89, %r923, 2;
shr.u32 %r488, %r924, 31;
shr.u32 %r489, %r922, 30;
add.s32 %r90, %r488, %r489;
setp.eq.s32 %p47, %r488, 0;
@%p47 bra BB0_64;
not.b32 %r490, %r924;
neg.s32 %r926, %r89;
setp.eq.s32 %p48, %r89, 0;
selp.u32 %r491, 1, 0, %p48;
add.s32 %r924, %r491, %r490;
xor.b32 %r925, %r80, -2147483648;
bra.uni BB0_66;
BB0_64:
mov.u32 %r925, %r80;
mov.u32 %r926, %r89;
BB0_66:
clz.b32 %r928, %r924;
setp.eq.s32 %p49, %r928, 0;
shl.b32 %r492, %r924, %r928;
mov.u32 %r493, 32;
sub.s32 %r494, %r493, %r928;
shr.u32 %r495, %r926, %r494;
add.s32 %r496, %r495, %r492;
selp.b32 %r98, %r924, %r496, %p49;
mov.u32 %r497, -921707870;
mul.hi.u32 %r927, %r98, %r497;
setp.eq.s32 %p50, %r80, 0;
neg.s32 %r498, %r90;
selp.b32 %r929, %r90, %r498, %p50;
setp.lt.s32 %p51, %r927, 1;
@%p51 bra BB0_68;
mul.lo.s32 %r499, %r98, -921707870;
shr.u32 %r500, %r499, 31;
shl.b32 %r501, %r927, 1;
add.s32 %r927, %r500, %r501;
add.s32 %r928, %r928, 1;
BB0_68:
mov.u32 %r502, 126;
sub.s32 %r503, %r502, %r928;
shl.b32 %r504, %r503, 23;
add.s32 %r505, %r927, 1;
shr.u32 %r506, %r505, 7;
add.s32 %r507, %r506, 1;
shr.u32 %r508, %r507, 1;
add.s32 %r509, %r508, %r504;
or.b32 %r510, %r509, %r925;
mov.b32 %f1625, %r510;
BB0_69:
mul.rn.f32 %f126, %f1625, %f1625;
add.s32 %r106, %r929, 1;
and.b32 %r107, %r106, 1;
setp.eq.s32 %p52, %r107, 0;
@%p52 bra BB0_71;
mov.f32 %f677, 0fBAB6061A;
mov.f32 %f678, 0f37CCF5CE;
fma.rn.f32 %f1626, %f678, %f126, %f677;
bra.uni BB0_72;
BB0_71:
mov.f32 %f679, 0f3C08839E;
mov.f32 %f680, 0fB94CA1F9;
fma.rn.f32 %f1626, %f680, %f126, %f679;
BB0_72:
@%p52 bra BB0_74;
mov.f32 %f681, 0f3D2AAAA5;
fma.rn.f32 %f682, %f1626, %f126, %f681;
mov.f32 %f683, 0fBF000000;
fma.rn.f32 %f1627, %f682, %f126, %f683;
bra.uni BB0_75;
BB0_74:
mov.f32 %f684, 0fBE2AAAA3;
fma.rn.f32 %f685, %f1626, %f126, %f684;
mov.f32 %f686, 0f00000000;
fma.rn.f32 %f1627, %f685, %f126, %f686;
BB0_75:
fma.rn.f32 %f1628, %f1627, %f1625, %f1625;
@%p52 bra BB0_77;
mov.f32 %f687, 0f3F800000;
fma.rn.f32 %f1628, %f1627, %f126, %f687;
BB0_77:
and.b32 %r511, %r106, 2;
setp.eq.s32 %p55, %r511, 0;
@%p55 bra BB0_79;
mov.f32 %f688, 0f00000000;
mov.f32 %f689, 0fBF800000;
fma.rn.f32 %f1628, %f1628, %f689, %f688;
BB0_79:
@%p43 bra BB0_81;
mov.f32 %f690, 0f00000000;
mul.rn.f32 %f1630, %f1630, %f690;
BB0_81:
mul.f32 %f691, %f1630, 0f3F22F983;
cvt.rni.s32.f32 %r939, %f691;
cvt.rn.f32.s32 %f692, %r939;
neg.f32 %f693, %f692;
fma.rn.f32 %f695, %f693, %f671, %f1630;
fma.rn.f32 %f697, %f693, %f673, %f695;
fma.rn.f32 %f1631, %f693, %f675, %f697;
abs.f32 %f699, %f1630;
setp.leu.f32 %p57, %f699, 0f47CE4780;
@%p57 bra BB0_92;
mov.b32 %r109, %f1630;
shr.u32 %r110, %r109, 23;
shl.b32 %r514, %r109, 8;
or.b32 %r111, %r514, -2147483648;
add.u64 %rd162, %SP, 32;
cvta.to.local.u64 %rd21, %rd162;
mov.u64 %rd293, %rd21;
mov.u32 %r930, %r434;
mov.u32 %r931, %r434;
BB0_83:
.pragma "nounroll";
shl.b64 %rd163, %rd294, 2;
mov.u64 %rd164, __cudart_i2opi_f;
add.s64 %rd165, %rd164, %rd163;
ld.const.u32 %r517, [%rd165];
// inline asm
{
mad.lo.cc.u32 %r515, %r517, %r111, %r931;
madc.hi.u32 %r931, %r517, %r111, 0;
}
// inline asm
st.local.u32 [%rd293], %r515;
add.s32 %r930, %r930, 1;
cvt.s64.s32 %rd294, %r930;
mul.wide.s32 %rd166, %r930, 4;
add.s64 %rd293, %rd21, %rd166;
setp.ne.s32 %p58, %r930, 6;
@%p58 bra BB0_83;
and.b32 %r520, %r110, 255;
add.s32 %r521, %r520, -128;
shr.u32 %r522, %r521, 5;
and.b32 %r116, %r109, -2147483648;
cvta.to.local.u64 %rd168, %rd162;
st.local.u32 [%rd168+24], %r931;
mov.u32 %r523, 6;
sub.s32 %r524, %r523, %r522;
mul.wide.s32 %rd169, %r524, 4;
add.s64 %rd27, %rd168, %rd169;
ld.local.u32 %r932, [%rd27];
ld.local.u32 %r933, [%rd27+-4];
and.b32 %r119, %r110, 31;
setp.eq.s32 %p59, %r119, 0;
@%p59 bra BB0_86;
mov.u32 %r525, 32;
sub.s32 %r526, %r525, %r119;
shr.u32 %r527, %r933, %r526;
shl.b32 %r528, %r932, %r119;
add.s32 %r932, %r527, %r528;
ld.local.u32 %r529, [%rd27+-8];
shr.u32 %r530, %r529, %r526;
shl.b32 %r531, %r933, %r119;
add.s32 %r933, %r530, %r531;
BB0_86:
shr.u32 %r532, %r933, 30;
shl.b32 %r533, %r932, 2;
add.s32 %r934, %r532, %r533;
shl.b32 %r125, %r933, 2;
shr.u32 %r534, %r934, 31;
shr.u32 %r535, %r932, 30;
add.s32 %r126, %r534, %r535;
setp.eq.s32 %p60, %r534, 0;
@%p60 bra BB0_87;
not.b32 %r536, %r934;
neg.s32 %r936, %r125;
setp.eq.s32 %p61, %r125, 0;
selp.u32 %r537, 1, 0, %p61;
add.s32 %r934, %r537, %r536;
xor.b32 %r935, %r116, -2147483648;
bra.uni BB0_89;
BB0_87:
mov.u32 %r935, %r116;
mov.u32 %r936, %r125;
BB0_89:
clz.b32 %r938, %r934;
setp.eq.s32 %p62, %r938, 0;
shl.b32 %r538, %r934, %r938;
mov.u32 %r539, 32;
sub.s32 %r540, %r539, %r938;
shr.u32 %r541, %r936, %r540;
add.s32 %r542, %r541, %r538;
selp.b32 %r134, %r934, %r542, %p62;
mov.u32 %r543, -921707870;
mul.hi.u32 %r937, %r134, %r543;
setp.eq.s32 %p63, %r116, 0;
neg.s32 %r544, %r126;
selp.b32 %r939, %r126, %r544, %p63;
setp.lt.s32 %p64, %r937, 1;
@%p64 bra BB0_91;
mul.lo.s32 %r545, %r134, -921707870;
shr.u32 %r546, %r545, 31;
shl.b32 %r547, %r937, 1;
add.s32 %r937, %r546, %r547;
add.s32 %r938, %r938, 1;
BB0_91:
mov.u32 %r548, 126;
sub.s32 %r549, %r548, %r938;
shl.b32 %r550, %r549, 23;
add.s32 %r551, %r937, 1;
shr.u32 %r552, %r551, 7;
add.s32 %r553, %r552, 1;
shr.u32 %r554, %r553, 1;
add.s32 %r555, %r554, %r550;
or.b32 %r556, %r555, %r935;
mov.b32 %f1631, %r556;
BB0_92:
mul.rn.f32 %f143, %f1631, %f1631;
and.b32 %r142, %r939, 1;
setp.eq.s32 %p65, %r142, 0;
@%p65 bra BB0_94;
mov.f32 %f700, 0fBAB6061A;
mov.f32 %f701, 0f37CCF5CE;
fma.rn.f32 %f1632, %f701, %f143, %f700;
bra.uni BB0_95;
BB0_94:
mov.f32 %f702, 0f3C08839E;
mov.f32 %f703, 0fB94CA1F9;
fma.rn.f32 %f1632, %f703, %f143, %f702;
BB0_95:
@%p65 bra BB0_97;
mov.f32 %f704, 0f3D2AAAA5;
fma.rn.f32 %f705, %f1632, %f143, %f704;
mov.f32 %f706, 0fBF000000;
fma.rn.f32 %f1633, %f705, %f143, %f706;
bra.uni BB0_98;
BB0_97:
mov.f32 %f707, 0fBE2AAAA3;
fma.rn.f32 %f708, %f1632, %f143, %f707;
mov.f32 %f709, 0f00000000;
fma.rn.f32 %f1633, %f708, %f143, %f709;
BB0_98:
fma.rn.f32 %f1634, %f1633, %f1631, %f1631;
@%p65 bra BB0_100;
mov.f32 %f710, 0f3F800000;
fma.rn.f32 %f1634, %f1633, %f143, %f710;
BB0_100:
and.b32 %r557, %r939, 2;
setp.eq.s32 %p68, %r557, 0;
@%p68 bra BB0_102;
mov.f32 %f711, 0f00000000;
mov.f32 %f712, 0fBF800000;
fma.rn.f32 %f1634, %f1634, %f712, %f711;
BB0_102:
mul.f32 %f716, %f118, %f1628;
mul.f32 %f717, %f716, %f716;
mov.f32 %f1636, 0f3F800000;
sub.f32 %f719, %f1636, %f717;
mul.f32 %f720, %f118, %f1634;
mul.f32 %f721, %f720, %f720;
sub.f32 %f722, %f719, %f721;
mov.f32 %f1639, 0f00000000;
max.f32 %f723, %f1639, %f722;
sqrt.rn.f32 %f724, %f723;
mul.f32 %f725, %f115, %f720;
mul.f32 %f726, %f116, %f720;
mul.f32 %f727, %f117, %f720;
mul.f32 %f728, %f113, %f116;
mul.f32 %f729, %f114, %f117;
sub.f32 %f730, %f728, %f729;
fma.rn.f32 %f731, %f730, %f716, %f725;
mul.f32 %f732, %f112, %f117;
mul.f32 %f733, %f113, %f115;
sub.f32 %f734, %f732, %f733;
fma.rn.f32 %f735, %f734, %f716, %f726;
mul.f32 %f736, %f114, %f115;
mul.f32 %f737, %f112, %f116;
sub.f32 %f738, %f736, %f737;
fma.rn.f32 %f739, %f738, %f716, %f727;
fma.rn.f32 %f155, %f112, %f724, %f731;
fma.rn.f32 %f156, %f114, %f724, %f735;
fma.rn.f32 %f157, %f113, %f724, %f739;
add.u64 %rd170, %SP, 28;
cvta.to.local.u64 %rd28, %rd170;
st.local.u32 [%rd28], %r434;
setp.gt.f32 %p69, %f156, 0f00000000;
setp.eq.f32 %p70, %f647, 0f00000000;
or.pred %p71, %p70, %p69;
mov.f32 %f1640, %f1639;
mov.f32 %f1641, %f1639;
@!%p71 bra BB0_106;
bra.uni BB0_103;
BB0_103:
mov.u32 %r562, 1065353216;
st.local.u32 [%rd28], %r562;
mul.f32 %f751, %f72, 0f3456BF95;
abs.f32 %f752, %f751;
mul.f32 %f753, %f73, 0f3456BF95;
abs.f32 %f754, %f753;
mul.f32 %f755, %f74, 0f3456BF95;
abs.f32 %f756, %f755;
max.f32 %f757, %f752, %f754;
max.f32 %f758, %f757, %f756;
max.f32 %f746, %f758, %f617;
ld.global.u32 %r559, [root];
mov.f32 %f747, 0f6C4ECB8F;
mov.u32 %r561, 4;
// inline asm
call _rt_trace_64, (%r559, %f72, %f73, %f74, %f155, %f156, %f157, %r456, %f746, %f747, %rd170, %r561);
// inline asm
setp.leu.f32 %p72, %f652, 0f00000000;
mov.f32 %f1637, %f1636;
mov.f32 %f1638, %f1636;
@%p72 bra BB0_105;
cvt.rzi.s32.f32 %r563, %f653;
neg.f32 %f766, %f155;
neg.f32 %f765, %f156;
neg.f32 %f764, %f157;
mov.u32 %r564, 6;
// inline asm
call (%f1636, %f1637, %f1638, %f763), _rt_texture_get_base_id, (%r563, %r564, %f764, %f765, %f766, %r434);
// inline asm
BB0_105:
ld.local.f32 %f767, [%rd28];
fma.rn.f32 %f1639, %f1636, %f767, 0f00000000;
fma.rn.f32 %f1640, %f1637, %f767, 0f00000000;
fma.rn.f32 %f1641, %f1638, %f767, 0f00000000;
BB0_106:
ld.global.u32 %r566, [bounceFlag];
setp.lt.s32 %p73, %r566, 1;
@%p73 bra BB0_108;
add.f32 %f768, %f651, 0fBF800000;
mul.f32 %f769, %f768, 0f42C80000;
mul.f32 %f1639, %f769, %f1639;
mul.f32 %f1640, %f769, %f1640;
mul.f32 %f1641, %f769, %f1641;
BB0_108:
cvt.rn.f32.u32 %f770, %r72;
mul.f32 %f771, %f648, %f1639;
mul.f32 %f1676, %f770, %f771;
mul.f32 %f772, %f649, %f1640;
mul.f32 %f1677, %f770, %f772;
mul.f32 %f773, %f650, %f1641;
mul.f32 %f1678, %f770, %f773;
bra.uni BB0_162;
BB0_121:
mov.u32 %r945, %r151;
mov.u32 %r946, %r160;
BB0_123:
clz.b32 %r948, %r944;
setp.eq.s32 %p89, %r948, 0;
shl.b32 %r595, %r944, %r948;
mov.u32 %r596, 32;
sub.s32 %r597, %r596, %r948;
shr.u32 %r598, %r946, %r597;
add.s32 %r599, %r598, %r595;
selp.b32 %r169, %r944, %r599, %p89;
mov.u32 %r600, -921707870;
mul.hi.u32 %r947, %r169, %r600;
setp.eq.s32 %p90, %r151, 0;
neg.s32 %r601, %r161;
selp.b32 %r949, %r161, %r601, %p90;
setp.lt.s32 %p91, %r947, 1;
@%p91 bra BB0_125;
mul.lo.s32 %r602, %r169, -921707870;
shr.u32 %r603, %r602, 31;
shl.b32 %r604, %r947, 1;
add.s32 %r947, %r603, %r604;
add.s32 %r948, %r948, 1;
BB0_125:
mov.u32 %r605, 126;
sub.s32 %r606, %r605, %r948;
shl.b32 %r607, %r606, 23;
add.s32 %r608, %r947, 1;
shr.u32 %r609, %r608, 7;
add.s32 %r610, %r609, 1;
shr.u32 %r611, %r610, 1;
add.s32 %r612, %r611, %r607;
or.b32 %r613, %r612, %r945;
mov.b32 %f1647, %r613;
BB0_126:
mul.rn.f32 %f197, %f1647, %f1647;
add.s32 %r177, %r949, 1;
and.b32 %r178, %r177, 1;
setp.eq.s32 %p92, %r178, 0;
@%p92 bra BB0_128;
mov.f32 %f865, 0fBAB6061A;
mov.f32 %f866, 0f37CCF5CE;
fma.rn.f32 %f1648, %f866, %f197, %f865;
bra.uni BB0_129;
BB0_128:
mov.f32 %f867, 0f3C08839E;
mov.f32 %f868, 0fB94CA1F9;
fma.rn.f32 %f1648, %f868, %f197, %f867;
BB0_129:
@%p92 bra BB0_131;
mov.f32 %f869, 0f3D2AAAA5;
fma.rn.f32 %f870, %f1648, %f197, %f869;
mov.f32 %f871, 0fBF000000;
fma.rn.f32 %f1649, %f870, %f197, %f871;
bra.uni BB0_132;
BB0_131:
mov.f32 %f872, 0fBE2AAAA3;
fma.rn.f32 %f873, %f1648, %f197, %f872;
mov.f32 %f874, 0f00000000;
fma.rn.f32 %f1649, %f873, %f197, %f874;
BB0_132:
fma.rn.f32 %f1650, %f1649, %f1647, %f1647;
@%p92 bra BB0_134;
mov.f32 %f875, 0f3F800000;
fma.rn.f32 %f1650, %f1649, %f197, %f875;
BB0_134:
and.b32 %r614, %r177, 2;
setp.eq.s32 %p95, %r614, 0;
@%p95 bra BB0_136;
mov.f32 %f876, 0f00000000;
mov.f32 %f877, 0fBF800000;
fma.rn.f32 %f1650, %f1650, %f877, %f876;
BB0_136:
@%p83 bra BB0_138;
mov.f32 %f878, 0f00000000;
mul.rn.f32 %f1652, %f1652, %f878;
BB0_138:
mul.f32 %f879, %f1652, 0f3F22F983;
cvt.rni.s32.f32 %r959, %f879;
cvt.rn.f32.s32 %f880, %r959;
neg.f32 %f881, %f880;
fma.rn.f32 %f883, %f881, %f859, %f1652;
fma.rn.f32 %f885, %f881, %f861, %f883;
fma.rn.f32 %f1653, %f881, %f863, %f885;
abs.f32 %f887, %f1652;
setp.leu.f32 %p97, %f887, 0f47CE4780;
@%p97 bra BB0_149;
mov.b32 %r180, %f1652;
shr.u32 %r181, %r180, 23;
shl.b32 %r617, %r180, 8;
or.b32 %r182, %r617, -2147483648;
mov.u32 %r950, 0;
mov.u64 %rd298, 0;
mov.u64 %rd297, %rd29;
mov.u32 %r951, %r950;
BB0_140:
.pragma "nounroll";
shl.b64 %rd180, %rd298, 2;
mov.u64 %rd181, __cudart_i2opi_f;
add.s64 %rd182, %rd181, %rd180;
ld.const.u32 %r620, [%rd182];
// inline asm
{
mad.lo.cc.u32 %r618, %r620, %r182, %r951;
madc.hi.u32 %r951, %r620, %r182, 0;
}
// inline asm
st.local.u32 [%rd297], %r618;
add.s32 %r950, %r950, 1;
cvt.s64.s32 %rd298, %r950;
mul.wide.s32 %rd183, %r950, 4;
add.s64 %rd297, %rd29, %rd183;
setp.ne.s32 %p98, %r950, 6;
@%p98 bra BB0_140;
and.b32 %r623, %r181, 255;
add.s32 %r624, %r623, -128;
shr.u32 %r625, %r624, 5;
and.b32 %r187, %r180, -2147483648;
st.local.u32 [%rd30], %r951;
mov.u32 %r626, 6;
sub.s32 %r627, %r626, %r625;
mul.wide.s32 %rd184, %r627, 4;
add.s64 %rd42, %rd29, %rd184;
ld.local.u32 %r952, [%rd42];
ld.local.u32 %r953, [%rd42+-4];
and.b32 %r190, %r181, 31;
setp.eq.s32 %p99, %r190, 0;
@%p99 bra BB0_143;
mov.u32 %r628, 32;
sub.s32 %r629, %r628, %r190;
shr.u32 %r630, %r953, %r629;
shl.b32 %r631, %r952, %r190;
add.s32 %r952, %r630, %r631;
ld.local.u32 %r632, [%rd42+-8];
shr.u32 %r633, %r632, %r629;
shl.b32 %r634, %r953, %r190;
add.s32 %r953, %r633, %r634;
BB0_143:
shr.u32 %r635, %r953, 30;
shl.b32 %r636, %r952, 2;
add.s32 %r954, %r635, %r636;
shl.b32 %r196, %r953, 2;
shr.u32 %r637, %r954, 31;
shr.u32 %r638, %r952, 30;
add.s32 %r197, %r637, %r638;
setp.eq.s32 %p100, %r637, 0;
@%p100 bra BB0_144;
not.b32 %r639, %r954;
neg.s32 %r956, %r196;
setp.eq.s32 %p101, %r196, 0;
selp.u32 %r640, 1, 0, %p101;
add.s32 %r954, %r640, %r639;
xor.b32 %r955, %r187, -2147483648;
bra.uni BB0_146;
BB0_144:
mov.u32 %r955, %r187;
mov.u32 %r956, %r196;
BB0_146:
clz.b32 %r958, %r954;
setp.eq.s32 %p102, %r958, 0;
shl.b32 %r641, %r954, %r958;
mov.u32 %r642, 32;
sub.s32 %r643, %r642, %r958;
shr.u32 %r644, %r956, %r643;
add.s32 %r645, %r644, %r641;
selp.b32 %r205, %r954, %r645, %p102;
mov.u32 %r646, -921707870;
mul.hi.u32 %r957, %r205, %r646;
setp.eq.s32 %p103, %r187, 0;
neg.s32 %r647, %r197;
selp.b32 %r959, %r197, %r647, %p103;
setp.lt.s32 %p104, %r957, 1;
@%p104 bra BB0_148;
mul.lo.s32 %r648, %r205, -921707870;
shr.u32 %r649, %r648, 31;
shl.b32 %r650, %r957, 1;
add.s32 %r957, %r649, %r650;
add.s32 %r958, %r958, 1;
BB0_148:
mov.u32 %r651, 126;
sub.s32 %r652, %r651, %r958;
shl.b32 %r653, %r652, 23;
add.s32 %r654, %r957, 1;
shr.u32 %r655, %r654, 7;
add.s32 %r656, %r655, 1;
shr.u32 %r657, %r656, 1;
add.s32 %r658, %r657, %r653;
or.b32 %r659, %r658, %r955;
mov.b32 %f1653, %r659;
BB0_149:
mul.rn.f32 %f214, %f1653, %f1653;
and.b32 %r213, %r959, 1;
setp.eq.s32 %p105, %r213, 0;
@%p105 bra BB0_151;
mov.f32 %f888, 0fBAB6061A;
mov.f32 %f889, 0f37CCF5CE;
fma.rn.f32 %f1654, %f889, %f214, %f888;
bra.uni BB0_152;
BB0_151:
mov.f32 %f890, 0f3C08839E;
mov.f32 %f891, 0fB94CA1F9;
fma.rn.f32 %f1654, %f891, %f214, %f890;
BB0_152:
@%p105 bra BB0_154;
mov.f32 %f892, 0f3D2AAAA5;
fma.rn.f32 %f893, %f1654, %f214, %f892;
mov.f32 %f894, 0fBF000000;
fma.rn.f32 %f1655, %f893, %f214, %f894;
bra.uni BB0_155;
BB0_154:
mov.f32 %f895, 0fBE2AAAA3;
fma.rn.f32 %f896, %f1654, %f214, %f895;
mov.f32 %f897, 0f00000000;
fma.rn.f32 %f1655, %f896, %f214, %f897;
BB0_155:
fma.rn.f32 %f1656, %f1655, %f1653, %f1653;
@%p105 bra BB0_157;
mov.f32 %f898, 0f3F800000;
fma.rn.f32 %f1656, %f1655, %f214, %f898;
BB0_157:
and.b32 %r660, %r959, 2;
setp.eq.s32 %p108, %r660, 0;
@%p108 bra BB0_159;
mov.f32 %f899, 0f00000000;
mov.f32 %f900, 0fBF800000;
fma.rn.f32 %f1656, %f1656, %f900, %f899;
BB0_159:
mul.f32 %f909, %f189, %f1650;
mul.f32 %f910, %f909, %f909;
mov.f32 %f911, 0f3F800000;
sub.f32 %f912, %f911, %f910;
mul.f32 %f913, %f189, %f1656;
mul.f32 %f914, %f913, %f913;
sub.f32 %f915, %f912, %f914;
mov.f32 %f916, 0f00000000;
max.f32 %f917, %f916, %f915;
sqrt.rn.f32 %f918, %f917;
mul.f32 %f919, %f186, %f913;
mul.f32 %f920, %f187, %f913;
mul.f32 %f921, %f188, %f913;
mul.f32 %f922, %f181, %f188;
mul.f32 %f923, %f182, %f187;
sub.f32 %f924, %f923, %f922;
fma.rn.f32 %f925, %f924, %f909, %f919;
mul.f32 %f926, %f182, %f186;
mul.f32 %f927, %f180, %f188;
sub.f32 %f928, %f927, %f926;
fma.rn.f32 %f929, %f928, %f909, %f920;
mul.f32 %f930, %f180, %f187;
mul.f32 %f931, %f181, %f186;
sub.f32 %f932, %f931, %f930;
fma.rn.f32 %f933, %f932, %f909, %f921;
fma.rn.f32 %f934, %f180, %f918, %f925;
fma.rn.f32 %f935, %f181, %f918, %f929;
fma.rn.f32 %f936, %f182, %f918, %f933;
add.f32 %f937, %f180, %f934;
add.f32 %f938, %f181, %f935;
add.f32 %f939, %f182, %f936;
mul.f32 %f940, %f647, %f937;
mul.f32 %f941, %f647, %f938;
mul.f32 %f942, %f647, %f939;
sub.f32 %f943, %f940, %f180;
sub.f32 %f944, %f941, %f181;
sub.f32 %f945, %f942, %f182;
mul.f32 %f946, %f944, %f944;
fma.rn.f32 %f947, %f943, %f943, %f946;
fma.rn.f32 %f948, %f945, %f945, %f947;
sqrt.rn.f32 %f949, %f948;
rcp.rn.f32 %f950, %f949;
mul.f32 %f904, %f950, %f943;
mul.f32 %f905, %f950, %f944;
mul.f32 %f906, %f950, %f945;
mul.f32 %f951, %f72, 0f3456BF95;
abs.f32 %f952, %f951;
mul.f32 %f953, %f73, 0f3456BF95;
abs.f32 %f954, %f953;
mul.f32 %f955, %f74, 0f3456BF95;
abs.f32 %f956, %f955;
max.f32 %f957, %f952, %f954;
max.f32 %f958, %f957, %f956;
max.f32 %f907, %f958, %f617;
add.u64 %rd185, %SP, 28;
cvta.to.local.u64 %rd186, %rd185;
mov.u32 %r664, 1065353216;
st.local.u32 [%rd186], %r664;
ld.global.u32 %r661, [root];
mov.f32 %f908, 0f6C4ECB8F;
mov.u32 %r663, 4;
// inline asm
call _rt_trace_64, (%r661, %f72, %f73, %f74, %f904, %f905, %f906, %r456, %f907, %f908, %rd185, %r663);
// inline asm
ld.local.f32 %f1658, [%rd186];
BB0_160:
ld.global.u32 %r665, [bounceFlag];
setp.gt.s32 %p109, %r665, 0;
mul.f32 %f960, %f651, 0f42C80000;
mul.f32 %f961, %f1645, %f1658;
mul.f32 %f962, %f960, %f961;
selp.f32 %f963, %f962, %f961, %p109;
mul.f32 %f1659, %f648, %f963;
mul.f32 %f1660, %f649, %f963;
mul.f32 %f1661, %f650, %f963;
BB0_161:
cvt.rn.f32.u32 %f964, %r72;
mul.f32 %f1676, %f964, %f1659;
mul.f32 %f1677, %f964, %f1660;
mul.f32 %f1678, %f964, %f1661;
BB0_162:
mov.u64 %rd192, localLights;
cvta.global.u64 %rd191, %rd192;
mov.u32 %r667, 96;
// inline asm
call (%rd187, %rd188, %rd189, %rd190), _rt_buffer_get_size_64, (%rd191, %r456, %r667);
// inline asm
cvt.u32.u64 %r668, %rd187;
setp.eq.s32 %p110, %r668, 0;
@%p110 bra BB0_188;
ld.global.f32 %f972, [cameraPos];
sub.f32 %f969, %f72, %f972;
ld.global.f32 %f973, [cameraPos+4];
sub.f32 %f970, %f73, %f973;
ld.global.f32 %f974, [cameraPos+8];
sub.f32 %f971, %f74, %f974;
ld.global.u32 %r669, [tileCubemap];
mov.u32 %r670, 6;
mov.u32 %r671, 0;
// inline asm
call (%f965, %f966, %f967, %f968), _rt_texture_get_base_id, (%r669, %r670, %f969, %f970, %f971, %r671);
// inline asm
mov.b32 %r214, %f966;
setp.eq.s32 %p111, %r214, 0;
@%p111 bra BB0_188;
mov.b32 %r676, %f965;
cvt.rn.f32.u32 %f244, %r214;
mul.f32 %f977, %f3, %f244;
cvt.rzi.u32.f32 %r677, %f977;
add.s32 %r678, %r214, -1;
min.u32 %r679, %r677, %r678;
add.s32 %r680, %r679, %r676;
cvt.u64.u32 %rd195, %r680;
mov.u64 %rd205, localLightIndices;
cvta.global.u64 %rd194, %rd205;
mov.u32 %r673, 4;
mov.u64 %rd204, 0;
// inline asm
call (%rd193), _rt_buffer_get_64, (%rd194, %r456, %r673, %rd195, %rd204, %rd204, %rd204);
// inline asm
ld.u32 %rd201, [%rd193];
// inline asm
call (%rd199), _rt_buffer_get_64, (%rd191, %r456, %r667, %rd201, %rd204, %rd204, %rd204);
// inline asm
ld.v4.f32 {%f978, %f979, %f980, %f981}, [%rd199+80];
ld.v4.f32 {%f982, %f983, %f984, %f985}, [%rd199+64];
ld.v4.f32 {%f986, %f987, %f988, %f989}, [%rd199+48];
ld.v4.f32 {%f1668, %f1669, %f1670, %f993}, [%rd199+32];
ld.v4.f32 {%f994, %f995, %f996, %f997}, [%rd199+16];
ld.v4.f32 {%f998, %f999, %f1000, %f1001}, [%rd199];
sub.f32 %f1002, %f999, %f72;
sub.f32 %f1003, %f1000, %f73;
sub.f32 %f1004, %f1001, %f74;
mul.f32 %f1005, %f1003, %f1003;
fma.rn.f32 %f1006, %f1002, %f1002, %f1005;
fma.rn.f32 %f1007, %f1004, %f1004, %f1006;
sqrt.rn.f32 %f272, %f1007;
rcp.rn.f32 %f1008, %f272;
mul.f32 %f273, %f1002, %f1008;
mul.f32 %f274, %f1003, %f1008;
mul.f32 %f275, %f1004, %f1008;
mul.f32 %f276, %f997, %f272;
mov.f32 %f1012, 0f40800000;
abs.f32 %f278, %f276;
setp.lt.f32 %p112, %f278, 0f00800000;
mul.f32 %f1014, %f278, 0f4B800000;
selp.f32 %f1015, 0fC3170000, 0fC2FE0000, %p112;
selp.f32 %f1016, %f1014, %f278, %p112;
mov.b32 %r681, %f1016;
and.b32 %r682, %r681, 8388607;
or.b32 %r683, %r682, 1065353216;
mov.b32 %f1017, %r683;
shr.u32 %r684, %r681, 23;
cvt.rn.f32.u32 %f1018, %r684;
add.f32 %f1019, %f1015, %f1018;
setp.gt.f32 %p113, %f1017, 0f3FB504F3;
mul.f32 %f1020, %f1017, 0f3F000000;
add.f32 %f1021, %f1019, 0f3F800000;
selp.f32 %f1022, %f1020, %f1017, %p113;
selp.f32 %f1023, %f1021, %f1019, %p113;
add.f32 %f1024, %f1022, 0fBF800000;
add.f32 %f976, %f1022, 0f3F800000;
// inline asm
rcp.approx.ftz.f32 %f975,%f976;
// inline asm
add.f32 %f1025, %f1024, %f1024;
mul.f32 %f1026, %f975, %f1025;
mul.f32 %f1027, %f1026, %f1026;
mov.f32 %f1028, 0f3C4CAF63;
mov.f32 %f1029, 0f3B18F0FE;
fma.rn.f32 %f1030, %f1029, %f1027, %f1028;
mov.f32 %f1031, 0f3DAAAABD;
fma.rn.f32 %f1032, %f1030, %f1027, %f1031;
mul.rn.f32 %f1033, %f1032, %f1027;
mul.rn.f32 %f1034, %f1033, %f1026;
sub.f32 %f1035, %f1024, %f1026;
neg.f32 %f1036, %f1026;
add.f32 %f1037, %f1035, %f1035;
fma.rn.f32 %f1038, %f1036, %f1024, %f1037;
mul.rn.f32 %f1039, %f975, %f1038;
add.f32 %f1040, %f1034, %f1026;
sub.f32 %f1041, %f1026, %f1040;
add.f32 %f1042, %f1034, %f1041;
add.f32 %f1043, %f1039, %f1042;
add.f32 %f1044, %f1040, %f1043;
sub.f32 %f1045, %f1040, %f1044;
add.f32 %f1046, %f1043, %f1045;
mov.f32 %f1047, 0f3F317200;
mul.rn.f32 %f1048, %f1023, %f1047;
mov.f32 %f1049, 0f35BFBE8E;
mul.rn.f32 %f1050, %f1023, %f1049;
add.f32 %f1051, %f1048, %f1044;
sub.f32 %f1052, %f1048, %f1051;
add.f32 %f1053, %f1044, %f1052;
add.f32 %f1054, %f1046, %f1053;
add.f32 %f1055, %f1050, %f1054;
add.f32 %f1056, %f1051, %f1055;
sub.f32 %f1057, %f1051, %f1056;
add.f32 %f1058, %f1055, %f1057;
mul.rn.f32 %f279, %f1012, %f1056;
neg.f32 %f1059, %f279;
fma.rn.f32 %f1060, %f1012, %f1056, %f1059;
fma.rn.f32 %f1061, %f1012, %f1058, %f1060;
mov.f32 %f1062, 0f00000000;
fma.rn.f32 %f280, %f1062, %f1056, %f1061;
add.rn.f32 %f281, %f279, %f280;
mov.b32 %r685, %f281;
setp.eq.s32 %p1, %r685, 1118925336;
add.s32 %r686, %r685, -1;
mov.b32 %f1063, %r686;
selp.f32 %f1064, %f1063, %f281, %p1;
mul.f32 %f1065, %f1064, 0f3FB8AA3B;
cvt.rzi.f32.f32 %f1066, %f1065;
mov.f32 %f1067, 0fBF317200;
fma.rn.f32 %f1068, %f1066, %f1067, %f1064;
mov.f32 %f1069, 0fB5BFBE8E;
fma.rn.f32 %f1070, %f1066, %f1069, %f1068;
mul.f32 %f1071, %f1070, 0f3FB8AA3B;
ex2.approx.ftz.f32 %f1072, %f1071;
add.f32 %f1073, %f1066, 0f00000000;
ex2.approx.f32 %f1074, %f1073;
mul.f32 %f1075, %f1072, %f1074;
setp.lt.f32 %p114, %f1064, 0fC2D20000;
selp.f32 %f1076, 0f00000000, %f1075, %p114;
setp.gt.f32 %p115, %f1064, 0f42D20000;
selp.f32 %f1665, 0f7F800000, %f1076, %p115;
setp.eq.f32 %p116, %f1665, 0f7F800000;
@%p116 bra BB0_166;
neg.f32 %f1077, %f281;
add.rn.f32 %f1078, %f279, %f1077;
add.rn.f32 %f1079, %f1078, %f280;
add.f32 %f1080, %f1079, 0f37000000;
selp.f32 %f1081, %f1080, %f1079, %p1;
fma.rn.f32 %f1665, %f1665, %f1081, %f1665;
BB0_166:
mov.f32 %f1595, 0f40000000;
cvt.rzi.f32.f32 %f1594, %f1595;
add.f32 %f1593, %f1594, %f1594;
mov.f32 %f1592, 0f40800000;
sub.f32 %f1591, %f1592, %f1593;
abs.f32 %f1590, %f1591;
setp.lt.f32 %p117, %f276, 0f00000000;
setp.eq.f32 %p118, %f1590, 0f3F800000;
and.pred %p2, %p117, %p118;
mov.b32 %r687, %f1665;
xor.b32 %r688, %r687, -2147483648;
mov.b32 %f1082, %r688;
selp.f32 %f1667, %f1082, %f1665, %p2;
setp.eq.f32 %p119, %f276, 0f00000000;
@%p119 bra BB0_169;
bra.uni BB0_167;
BB0_169:
add.f32 %f1085, %f276, %f276;
selp.f32 %f1667, %f1085, 0f00000000, %p118;
bra.uni BB0_170;
BB0_167:
setp.geu.f32 %p120, %f276, 0f00000000;
@%p120 bra BB0_170;
mov.f32 %f1599, 0f40800000;
cvt.rzi.f32.f32 %f1084, %f1599;
setp.neu.f32 %p121, %f1084, 0f40800000;
selp.f32 %f1667, 0f7FFFFFFF, %f1667, %p121;
BB0_170:
add.f32 %f1086, %f278, 0f40800000;
mov.b32 %r689, %f1086;
setp.lt.s32 %p123, %r689, 2139095040;
@%p123 bra BB0_175;
setp.gtu.f32 %p124, %f278, 0f7F800000;
@%p124 bra BB0_174;
bra.uni BB0_172;
BB0_174:
add.f32 %f1667, %f276, 0f40800000;
bra.uni BB0_175;
BB0_172:
setp.neu.f32 %p125, %f278, 0f7F800000;
@%p125 bra BB0_175;
selp.f32 %f1667, 0fFF800000, 0f7F800000, %p2;
BB0_175:
add.u64 %rd281, %SP, 0;
cvta.to.local.u64 %rd280, %rd281;
add.s64 %rd279, %rd280, 12;
mul.f32 %f1087, %f995, %f272;
mov.f32 %f1672, 0f3F800000;
sub.f32 %f1089, %f1672, %f1667;
setp.eq.f32 %p126, %f276, 0f3F800000;
selp.f32 %f1090, 0f00000000, %f1089, %p126;
cvt.sat.f32.f32 %f1091, %f1090;
fma.rn.f32 %f1092, %f1087, %f1087, %f996;
div.rn.f32 %f1671, %f1091, %f1092;
ld.local.f32 %f1093, [%rd279+-12];
ld.local.f32 %f1094, [%rd279+-8];
mul.f32 %f1095, %f274, %f1094;
fma.rn.f32 %f1096, %f273, %f1093, %f1095;
ld.local.f32 %f1097, [%rd279+-4];
fma.rn.f32 %f1098, %f275, %f1097, %f1096;
cvt.sat.f32.f32 %f308, %f1098;
setp.eq.f32 %p127, %f998, 0f3F800000;
@%p127 bra BB0_181;
bra.uni BB0_176;
BB0_181:
setp.leu.f32 %p131, %f993, 0f00000000;
@%p131 bra BB0_183;
mov.u32 %r899, 2;
mul.f32 %f1129, %f978, %f273;
mul.f32 %f1130, %f979, %f274;
neg.f32 %f1131, %f1130;
sub.f32 %f1132, %f1131, %f1129;
mul.f32 %f1133, %f980, %f275;
sub.f32 %f1134, %f1132, %f1133;
setp.gt.f32 %p132, %f1134, 0f00000000;
selp.f32 %f1135, 0f3F800000, 0f00000000, %p132;
mul.f32 %f1136, %f987, %f274;
fma.rn.f32 %f1137, %f986, %f273, %f1136;
mul.f32 %f1138, %f983, %f274;
fma.rn.f32 %f1139, %f982, %f273, %f1138;
fma.rn.f32 %f1140, %f988, %f275, %f1137;
fma.rn.f32 %f1141, %f984, %f275, %f1139;
fma.rn.f32 %f1142, %f989, %f1140, 0f3F000000;
mov.f32 %f1143, 0f3F800000;
sub.f32 %f1125, %f1143, %f1142;
fma.rn.f32 %f1126, %f989, %f1141, 0f3F000000;
cvt.rzi.s32.f32 %r693, %f993;
mov.f32 %f1128, 0f00000000;
// inline asm
call (%f1121, %f1122, %f1123, %f1124), _rt_texture_get_f_id, (%r693, %r899, %f1125, %f1126, %f1128, %f1128);
// inline asm
mul.f32 %f1144, %f1135, %f1121;
mul.f32 %f1145, %f1135, %f1122;
mul.f32 %f1146, %f1135, %f1123;
mul.f32 %f1668, %f1668, %f1144;
mul.f32 %f1669, %f1669, %f1145;
mul.f32 %f1670, %f1670, %f1146;
bra.uni BB0_183;
BB0_176:
setp.eq.f32 %p128, %f998, 0f40000000;
@%p128 bra BB0_179;
bra.uni BB0_177;
BB0_179:
setp.leu.f32 %p130, %f993, 0f00000000;
@%p130 bra BB0_183;
mov.u32 %r898, 0;
mov.u32 %r897, 6;
mul.f32 %f1115, %f987, %f274;
fma.rn.f32 %f1116, %f986, %f273, %f1115;
mul.f32 %f1117, %f983, %f274;
fma.rn.f32 %f1118, %f982, %f273, %f1117;
mul.f32 %f1119, %f979, %f274;
fma.rn.f32 %f1120, %f978, %f273, %f1119;
fma.rn.f32 %f1112, %f988, %f275, %f1116;
fma.rn.f32 %f1113, %f984, %f275, %f1118;
fma.rn.f32 %f1114, %f980, %f275, %f1120;
cvt.rzi.s32.f32 %r690, %f993;
// inline asm
call (%f1108, %f1109, %f1110, %f1111), _rt_texture_get_base_id, (%r690, %r897, %f1112, %f1113, %f1114, %r898);
// inline asm
mul.f32 %f1668, %f1668, %f1108;
mul.f32 %f1669, %f1669, %f1109;
mul.f32 %f1670, %f1670, %f1110;
bra.uni BB0_183;
BB0_177:
setp.neu.f32 %p129, %f998, 0f40800000;
@%p129 bra BB0_183;
mul.f32 %f1099, %f978, %f273;
mul.f32 %f1100, %f979, %f274;
neg.f32 %f1101, %f1100;
sub.f32 %f1102, %f1101, %f1099;
mul.f32 %f1103, %f980, %f275;
sub.f32 %f1104, %f1102, %f1103;
fma.rn.f32 %f1105, %f993, %f1104, %f989;
cvt.sat.f32.f32 %f1106, %f1105;
mul.f32 %f1107, %f1106, %f1106;
mul.f32 %f1671, %f1671, %f1107;
BB0_183:
mov.f32 %f1673, 0f00000000;
max.f32 %f1150, %f1668, %f1669;
max.f32 %f1151, %f1150, %f1670;
mul.f32 %f320, %f308, %f1671;
mul.f32 %f1152, %f320, %f1151;
setp.lt.f32 %p133, %f1152, 0f3727C5AC;
mov.f32 %f1674, %f1673;
mov.f32 %f1675, %f1673;
@%p133 bra BB0_187;
add.u64 %rd207, %SP, 32;
cvta.to.local.u64 %rd43, %rd207;
mov.u32 %r695, 1065353216;
st.local.u32 [%rd43], %r695;
setp.leu.f32 %p134, %f981, 0f00000000;
@%p134 bra BB0_186;
mov.u32 %r886, 4;
mov.f32 %f1597, 0f38D1B717;
mov.u32 %r885, 1;
fma.rn.f32 %f1162, %f1, 0f40000000, 0fBF800000;
fma.rn.f32 %f1163, %f2, 0f40000000, 0fBF800000;
fma.rn.f32 %f1164, %f3, 0f40000000, 0fBF800000;
fma.rn.f32 %f1165, %f1162, %f994, %f999;
fma.rn.f32 %f1166, %f1163, %f994, %f1000;
fma.rn.f32 %f1167, %f1164, %f994, %f1001;
sub.f32 %f1168, %f1165, %f72;
sub.f32 %f1169, %f1166, %f73;
sub.f32 %f1170, %f1167, %f74;
mul.f32 %f1171, %f1169, %f1169;
fma.rn.f32 %f1172, %f1168, %f1168, %f1171;
fma.rn.f32 %f1173, %f1170, %f1170, %f1172;
sqrt.rn.f32 %f1161, %f1173;
rcp.rn.f32 %f1174, %f1161;
mul.f32 %f1157, %f1168, %f1174;
mul.f32 %f1158, %f1169, %f1174;
mul.f32 %f1159, %f1170, %f1174;
mul.f32 %f1175, %f72, 0f3456BF95;
abs.f32 %f1176, %f1175;
mul.f32 %f1177, %f73, 0f3456BF95;
abs.f32 %f1178, %f1177;
mul.f32 %f1179, %f74, 0f3456BF95;
abs.f32 %f1180, %f1179;
max.f32 %f1181, %f1176, %f1178;
max.f32 %f1182, %f1181, %f1180;
max.f32 %f1160, %f1182, %f1597;
ld.global.u32 %r696, [root];
// inline asm
call _rt_trace_64, (%r696, %f72, %f73, %f74, %f1157, %f1158, %f1159, %r885, %f1160, %f1161, %rd207, %r886);
// inline asm
ld.local.f32 %f1672, [%rd43];
BB0_186:
ld.global.u32 %r699, [bounceFlag];
setp.gt.s32 %p135, %r699, 0;
mul.f32 %f1184, %f320, %f1672;
mul.f32 %f1185, %f985, %f1184;
selp.f32 %f1186, %f1185, %f1184, %p135;
mul.f32 %f1673, %f1668, %f1186;
mul.f32 %f1674, %f1669, %f1186;
mul.f32 %f1675, %f1670, %f1186;
BB0_187:
fma.rn.f32 %f1676, %f244, %f1673, %f1676;
fma.rn.f32 %f1677, %f244, %f1674, %f1677;
fma.rn.f32 %f1678, %f244, %f1675, %f1678;
BB0_188:
add.u64 %rd284, %SP, 0;
cvta.to.local.u64 %rd283, %rd284;
add.s64 %rd282, %rd283, 12;
ld.local.f32 %f337, [%rd282+-8];
ld.local.f32 %f336, [%rd282+-4];
ld.local.f32 %f335, [%rd282+-12];
abs.f32 %f1187, %f335;
abs.f32 %f1188, %f336;
setp.gt.f32 %p136, %f1187, %f1188;
neg.f32 %f1189, %f337;
neg.f32 %f1190, %f336;
selp.f32 %f1191, %f1189, 0f00000000, %p136;
selp.f32 %f1192, %f335, %f1190, %p136;
selp.f32 %f1193, 0f00000000, %f337, %p136;
mul.f32 %f1194, %f1192, %f1192;
fma.rn.f32 %f1195, %f1191, %f1191, %f1194;
fma.rn.f32 %f1196, %f1193, %f1193, %f1195;
sqrt.rn.f32 %f1197, %f1196;
rcp.rn.f32 %f1198, %f1197;
mul.f32 %f338, %f1191, %f1198;
mul.f32 %f339, %f1192, %f1198;
mul.f32 %f340, %f1193, %f1198;
sqrt.rn.f32 %f341, %f1;
add.u64 %rd209, %SP, 32;
cvta.to.local.u64 %rd44, %rd209;
mul.f32 %f1685, %f2, 0f40C90FDB;
abs.f32 %f343, %f1685;
setp.neu.f32 %p137, %f343, 0f7F800000;
mov.f32 %f1679, %f1685;
@%p137 bra BB0_190;
mov.f32 %f1199, 0f00000000;
mul.rn.f32 %f1679, %f1685, %f1199;
BB0_190:
mul.f32 %f1200, %f1679, 0f3F22F983;
cvt.rni.s32.f32 %r969, %f1200;
cvt.rn.f32.s32 %f1201, %r969;
neg.f32 %f1202, %f1201;
mov.f32 %f1203, 0f3FC90FDA;
fma.rn.f32 %f1204, %f1202, %f1203, %f1679;
mov.f32 %f1205, 0f33A22168;
fma.rn.f32 %f1206, %f1202, %f1205, %f1204;
mov.f32 %f1207, 0f27C234C5;
fma.rn.f32 %f1680, %f1202, %f1207, %f1206;
abs.f32 %f1208, %f1679;
add.s64 %rd45, %rd44, 24;
setp.leu.f32 %p138, %f1208, 0f47CE4780;
@%p138 bra BB0_201;
mov.b32 %r216, %f1679;
shr.u32 %r217, %r216, 23;
shl.b32 %r702, %r216, 8;
or.b32 %r218, %r702, -2147483648;
mov.u32 %r961, 0;
mov.u64 %rd299, __cudart_i2opi_f;
mov.u32 %r960, -6;
mov.u64 %rd300, %rd44;
BB0_192:
.pragma "nounroll";
ld.const.u32 %r705, [%rd299];
// inline asm
{
mad.lo.cc.u32 %r703, %r705, %r218, %r961;
madc.hi.u32 %r961, %r705, %r218, 0;
}
// inline asm
st.local.u32 [%rd300], %r703;
add.s64 %rd300, %rd300, 4;
add.s64 %rd299, %rd299, 4;
add.s32 %r960, %r960, 1;
setp.ne.s32 %p139, %r960, 0;
@%p139 bra BB0_192;
and.b32 %r708, %r217, 255;
add.s32 %r709, %r708, -128;
shr.u32 %r710, %r709, 5;
and.b32 %r223, %r216, -2147483648;
st.local.u32 [%rd45], %r961;
mov.u32 %r711, 6;
sub.s32 %r712, %r711, %r710;
mul.wide.s32 %rd211, %r712, 4;
add.s64 %rd50, %rd44, %rd211;
ld.local.u32 %r962, [%rd50];
ld.local.u32 %r963, [%rd50+-4];
and.b32 %r226, %r217, 31;
setp.eq.s32 %p140, %r226, 0;
@%p140 bra BB0_195;
mov.u32 %r713, 32;
sub.s32 %r714, %r713, %r226;
shr.u32 %r715, %r963, %r714;
shl.b32 %r716, %r962, %r226;
add.s32 %r962, %r715, %r716;
ld.local.u32 %r717, [%rd50+-8];
shr.u32 %r718, %r717, %r714;
shl.b32 %r719, %r963, %r226;
add.s32 %r963, %r718, %r719;
BB0_195:
shr.u32 %r720, %r963, 30;
shl.b32 %r721, %r962, 2;
add.s32 %r964, %r720, %r721;
shl.b32 %r232, %r963, 2;
shr.u32 %r722, %r964, 31;
shr.u32 %r723, %r962, 30;
add.s32 %r233, %r722, %r723;
setp.eq.s32 %p141, %r722, 0;
@%p141 bra BB0_196;
not.b32 %r724, %r964;
neg.s32 %r966, %r232;
setp.eq.s32 %p142, %r232, 0;
selp.u32 %r725, 1, 0, %p142;
add.s32 %r964, %r725, %r724;
xor.b32 %r965, %r223, -2147483648;
bra.uni BB0_198;
BB0_196:
mov.u32 %r965, %r223;
mov.u32 %r966, %r232;
BB0_198:
clz.b32 %r968, %r964;
setp.eq.s32 %p143, %r968, 0;
shl.b32 %r726, %r964, %r968;
mov.u32 %r727, 32;
sub.s32 %r728, %r727, %r968;
shr.u32 %r729, %r966, %r728;
add.s32 %r730, %r729, %r726;
selp.b32 %r241, %r964, %r730, %p143;
mov.u32 %r731, -921707870;
mul.hi.u32 %r967, %r241, %r731;
setp.eq.s32 %p144, %r223, 0;
neg.s32 %r732, %r233;
selp.b32 %r969, %r233, %r732, %p144;
setp.lt.s32 %p145, %r967, 1;
@%p145 bra BB0_200;
mul.lo.s32 %r733, %r241, -921707870;
shr.u32 %r734, %r733, 31;
shl.b32 %r735, %r967, 1;
add.s32 %r967, %r734, %r735;
add.s32 %r968, %r968, 1;
BB0_200:
mov.u32 %r736, 126;
sub.s32 %r737, %r736, %r968;
shl.b32 %r738, %r737, 23;
add.s32 %r739, %r967, 1;
shr.u32 %r740, %r739, 7;
add.s32 %r741, %r740, 1;
shr.u32 %r742, %r741, 1;
add.s32 %r743, %r742, %r738;
or.b32 %r744, %r743, %r965;
mov.b32 %f1680, %r744;
BB0_201:
mul.rn.f32 %f349, %f1680, %f1680;
add.s32 %r249, %r969, 1;
and.b32 %r250, %r249, 1;
setp.eq.s32 %p146, %r250, 0;
@%p146 bra BB0_203;
mov.f32 %f1209, 0fBAB6061A;
mov.f32 %f1210, 0f37CCF5CE;
fma.rn.f32 %f1681, %f1210, %f349, %f1209;
bra.uni BB0_204;
BB0_203:
mov.f32 %f1211, 0f3C08839E;
mov.f32 %f1212, 0fB94CA1F9;
fma.rn.f32 %f1681, %f1212, %f349, %f1211;
BB0_204:
@%p146 bra BB0_206;
mov.f32 %f1213, 0f3D2AAAA5;
fma.rn.f32 %f1214, %f1681, %f349, %f1213;
mov.f32 %f1215, 0fBF000000;
fma.rn.f32 %f1682, %f1214, %f349, %f1215;
bra.uni BB0_207;
BB0_206:
mov.f32 %f1216, 0fBE2AAAA3;
fma.rn.f32 %f1217, %f1681, %f349, %f1216;
mov.f32 %f1218, 0f00000000;
fma.rn.f32 %f1682, %f1217, %f349, %f1218;
BB0_207:
fma.rn.f32 %f1683, %f1682, %f1680, %f1680;
@%p146 bra BB0_209;
mov.f32 %f1219, 0f3F800000;
fma.rn.f32 %f1683, %f1682, %f349, %f1219;
BB0_209:
and.b32 %r745, %r249, 2;
setp.eq.s32 %p149, %r745, 0;
@%p149 bra BB0_211;
mov.f32 %f1220, 0f00000000;
mov.f32 %f1221, 0fBF800000;
fma.rn.f32 %f1683, %f1683, %f1221, %f1220;
BB0_211:
@%p137 bra BB0_213;
mov.f32 %f1222, 0f00000000;
mul.rn.f32 %f1685, %f1685, %f1222;
BB0_213:
mul.f32 %f1223, %f1685, 0f3F22F983;
cvt.rni.s32.f32 %r979, %f1223;
cvt.rn.f32.s32 %f1224, %r979;
neg.f32 %f1225, %f1224;
fma.rn.f32 %f1227, %f1225, %f1203, %f1685;
fma.rn.f32 %f1229, %f1225, %f1205, %f1227;
fma.rn.f32 %f1686, %f1225, %f1207, %f1229;
abs.f32 %f1231, %f1685;
setp.leu.f32 %p151, %f1231, 0f47CE4780;
@%p151 bra BB0_224;
mov.b32 %r252, %f1685;
shr.u32 %r253, %r252, 23;
shl.b32 %r748, %r252, 8;
or.b32 %r254, %r748, -2147483648;
mov.u32 %r971, 0;
mov.u64 %rd301, __cudart_i2opi_f;
mov.u32 %r970, -6;
mov.u64 %rd302, %rd44;
BB0_215:
.pragma "nounroll";
ld.const.u32 %r751, [%rd301];
// inline asm
{
mad.lo.cc.u32 %r749, %r751, %r254, %r971;
madc.hi.u32 %r971, %r751, %r254, 0;
}
// inline asm
st.local.u32 [%rd302], %r749;
add.s64 %rd302, %rd302, 4;
add.s64 %rd301, %rd301, 4;
add.s32 %r970, %r970, 1;
setp.ne.s32 %p152, %r970, 0;
@%p152 bra BB0_215;
and.b32 %r754, %r253, 255;
add.s32 %r755, %r754, -128;
shr.u32 %r756, %r755, 5;
and.b32 %r259, %r252, -2147483648;
st.local.u32 [%rd45], %r971;
mov.u32 %r757, 6;
sub.s32 %r758, %r757, %r756;
mul.wide.s32 %rd213, %r758, 4;
add.s64 %rd55, %rd44, %rd213;
ld.local.u32 %r972, [%rd55];
ld.local.u32 %r973, [%rd55+-4];
and.b32 %r262, %r253, 31;
setp.eq.s32 %p153, %r262, 0;
@%p153 bra BB0_218;
mov.u32 %r759, 32;
sub.s32 %r760, %r759, %r262;
shr.u32 %r761, %r973, %r760;
shl.b32 %r762, %r972, %r262;
add.s32 %r972, %r761, %r762;
ld.local.u32 %r763, [%rd55+-8];
shr.u32 %r764, %r763, %r760;
shl.b32 %r765, %r973, %r262;
add.s32 %r973, %r764, %r765;
BB0_218:
shr.u32 %r766, %r973, 30;
shl.b32 %r767, %r972, 2;
add.s32 %r974, %r766, %r767;
shl.b32 %r268, %r973, 2;
shr.u32 %r768, %r974, 31;
shr.u32 %r769, %r972, 30;
add.s32 %r269, %r768, %r769;
setp.eq.s32 %p154, %r768, 0;
@%p154 bra BB0_219;
not.b32 %r770, %r974;
neg.s32 %r976, %r268;
setp.eq.s32 %p155, %r268, 0;
selp.u32 %r771, 1, 0, %p155;
add.s32 %r974, %r771, %r770;
xor.b32 %r975, %r259, -2147483648;
bra.uni BB0_221;
BB0_219:
mov.u32 %r975, %r259;
mov.u32 %r976, %r268;
BB0_221:
clz.b32 %r978, %r974;
setp.eq.s32 %p156, %r978, 0;
shl.b32 %r772, %r974, %r978;
mov.u32 %r773, 32;
sub.s32 %r774, %r773, %r978;
shr.u32 %r775, %r976, %r774;
add.s32 %r776, %r775, %r772;
selp.b32 %r277, %r974, %r776, %p156;
mov.u32 %r777, -921707870;
mul.hi.u32 %r977, %r277, %r777;
setp.eq.s32 %p157, %r259, 0;
neg.s32 %r778, %r269;
selp.b32 %r979, %r269, %r778, %p157;
setp.lt.s32 %p158, %r977, 1;
@%p158 bra BB0_223;
mul.lo.s32 %r779, %r277, -921707870;
shr.u32 %r780, %r779, 31;
shl.b32 %r781, %r977, 1;
add.s32 %r977, %r780, %r781;
add.s32 %r978, %r978, 1;
BB0_223:
mov.u32 %r782, 126;
sub.s32 %r783, %r782, %r978;
shl.b32 %r784, %r783, 23;
add.s32 %r785, %r977, 1;
shr.u32 %r786, %r785, 7;
add.s32 %r787, %r786, 1;
shr.u32 %r788, %r787, 1;
add.s32 %r789, %r788, %r784;
or.b32 %r790, %r789, %r975;
mov.b32 %f1686, %r790;
BB0_224:
mul.rn.f32 %f366, %f1686, %f1686;
and.b32 %r285, %r979, 1;
setp.eq.s32 %p159, %r285, 0;
@%p159 bra BB0_226;
mov.f32 %f1232, 0fBAB6061A;
mov.f32 %f1233, 0f37CCF5CE;
fma.rn.f32 %f1687, %f1233, %f366, %f1232;
bra.uni BB0_227;
BB0_226:
mov.f32 %f1234, 0f3C08839E;
mov.f32 %f1235, 0fB94CA1F9;
fma.rn.f32 %f1687, %f1235, %f366, %f1234;
BB0_227:
@%p159 bra BB0_229;
mov.f32 %f1236, 0f3D2AAAA5;
fma.rn.f32 %f1237, %f1687, %f366, %f1236;
mov.f32 %f1238, 0fBF000000;
fma.rn.f32 %f1688, %f1237, %f366, %f1238;
bra.uni BB0_230;
BB0_229:
mov.f32 %f1239, 0fBE2AAAA3;
fma.rn.f32 %f1240, %f1687, %f366, %f1239;
mov.f32 %f1241, 0f00000000;
fma.rn.f32 %f1688, %f1240, %f366, %f1241;
BB0_230:
fma.rn.f32 %f1689, %f1688, %f1686, %f1686;
@%p159 bra BB0_232;
mov.f32 %f1242, 0f3F800000;
fma.rn.f32 %f1689, %f1688, %f366, %f1242;
BB0_232:
and.b32 %r791, %r979, 2;
setp.eq.s32 %p162, %r791, 0;
@%p162 bra BB0_234;
mov.f32 %f1243, 0f00000000;
mov.f32 %f1244, 0fBF800000;
fma.rn.f32 %f1689, %f1689, %f1244, %f1243;
BB0_234:
mov.f32 %f1598, 0f38D1B717;
mov.u32 %r887, 1;
mul.f32 %f1253, %f341, %f1683;
mul.f32 %f1254, %f1253, %f1253;
mov.f32 %f1255, 0f3F800000;
sub.f32 %f1256, %f1255, %f1254;
mul.f32 %f1257, %f341, %f1689;
mul.f32 %f1258, %f1257, %f1257;
sub.f32 %f1259, %f1256, %f1258;
mov.f32 %f1260, 0f00000000;
max.f32 %f1261, %f1260, %f1259;
sqrt.rn.f32 %f1262, %f1261;
mul.f32 %f1263, %f338, %f1257;
mul.f32 %f1264, %f339, %f1257;
mul.f32 %f1265, %f340, %f1257;
mul.f32 %f1266, %f336, %f339;
mul.f32 %f1267, %f337, %f340;
sub.f32 %f1268, %f1266, %f1267;
fma.rn.f32 %f1269, %f1268, %f1253, %f1263;
mul.f32 %f1270, %f335, %f340;
mul.f32 %f1271, %f336, %f338;
sub.f32 %f1272, %f1270, %f1271;
fma.rn.f32 %f1273, %f1272, %f1253, %f1264;
mul.f32 %f1274, %f337, %f338;
mul.f32 %f1275, %f335, %f339;
sub.f32 %f1276, %f1274, %f1275;
fma.rn.f32 %f1277, %f1276, %f1253, %f1265;
fma.rn.f32 %f1248, %f335, %f1262, %f1269;
fma.rn.f32 %f1249, %f337, %f1262, %f1273;
fma.rn.f32 %f1250, %f336, %f1262, %f1277;
add.u64 %rd214, %SP, 28;
cvta.to.local.u64 %rd215, %rd214;
mov.u32 %r795, 1065353216;
st.local.u32 [%rd215], %r795;
mul.f32 %f1278, %f72, 0f3456BF95;
abs.f32 %f1279, %f1278;
mul.f32 %f1280, %f73, 0f3456BF95;
abs.f32 %f1281, %f1280;
mul.f32 %f1282, %f74, 0f3456BF95;
abs.f32 %f1283, %f1282;
max.f32 %f1284, %f1279, %f1281;
max.f32 %f1285, %f1284, %f1283;
max.f32 %f1251, %f1285, %f1598;
ld.global.f32 %f1252, [aoRadius];
ld.global.u32 %r792, [root];
mov.u32 %r794, 4;
// inline asm
call _rt_trace_64, (%r792, %f72, %f73, %f74, %f1248, %f1249, %f1250, %r887, %f1251, %f1252, %rd214, %r794);
// inline asm
ld.local.f32 %f1700, [%rd215];
mov.u16 %rs14, 0;
BB0_236:
mov.u64 %rd286, GBufferAtten;
cvta.global.u64 %rd285, %rd286;
mov.u32 %r889, 12;
mov.u32 %r888, 2;
abs.f32 %f1297, %f1676;
setp.gtu.f32 %p164, %f1297, 0f7F800000;
selp.f32 %f1298, 0f00000000, %f1676, %p164;
abs.f32 %f1299, %f1677;
setp.gtu.f32 %p165, %f1299, 0f7F800000;
selp.f32 %f1300, 0f00000000, %f1677, %p165;
abs.f32 %f1301, %f1678;
setp.gtu.f32 %p166, %f1301, 0f7F800000;
selp.f32 %f1302, 0f00000000, %f1678, %p166;
fma.rn.f32 %f1704, %f1694, %f1298, %f1691;
fma.rn.f32 %f1705, %f1695, %f1300, %f1692;
fma.rn.f32 %f1706, %f1696, %f1302, %f1693;
ld.global.v2.u32 {%r801, %r802}, [pixelID];
cvt.u64.u32 %rd218, %r801;
cvt.u64.u32 %rd219, %r802;
mov.u64 %rd227, 0;
// inline asm
call (%rd216), _rt_buffer_get_64, (%rd285, %r888, %r889, %rd218, %rd219, %rd227, %rd227);
// inline asm
st.f32 [%rd216+8], %f1696;
st.f32 [%rd216+4], %f1695;
st.f32 [%rd216], %f1694;
ld.global.v2.u32 {%r805, %r806}, [pixelID];
cvt.u64.u32 %rd224, %r805;
cvt.u64.u32 %rd225, %r806;
mov.u64 %rd229, accumBufferAdd;
cvta.global.u64 %rd223, %rd229;
// inline asm
call (%rd222), _rt_buffer_get_64, (%rd223, %r888, %r889, %rd224, %rd225, %rd227, %rd227);
// inline asm
ld.global.u32 %r809, [bounceFlag];
setp.eq.s32 %p167, %r809, 0;
@%p167 bra BB0_238;
ld.f32 %f1303, [%rd222+8];
ld.f32 %f1304, [%rd222+4];
ld.f32 %f1305, [%rd222];
add.f32 %f1704, %f1704, %f1305;
add.f32 %f1705, %f1705, %f1304;
add.f32 %f1706, %f1706, %f1303;
bra.uni BB0_243;
BB0_238:
ld.global.u32 %r286, [firstSkylight];
setp.lt.s32 %p168, %r286, 0;
setp.eq.s16 %p169, %rs14, 0;
or.pred %p170, %p168, %p169;
@%p170 bra BB0_243;
cvt.s64.s32 %rd232, %r286;
mov.u64 %rd236, globalLights;
cvta.global.u64 %rd231, %rd236;
mov.u32 %r810, 1;
mov.u32 %r811, 48;
// inline asm
call (%rd230), _rt_buffer_get_64, (%rd231, %r810, %r811, %rd232, %rd227, %rd227, %rd227);
// inline asm
ld.v4.f32 {%f1310, %f1311, %f1312, %f1313}, [%rd230+16];
ld.v4.f32 {%f1314, %f1315, %f1316, %f1317}, [%rd230];
setp.neu.f32 %p171, %f1310, 0f00000000;
setp.leu.f32 %p172, %f63, 0f00000000;
and.pred %p173, %p171, %p172;
@%p173 bra BB0_243;
mov.f32 %f1701, 0f3F800000;
setp.leu.f32 %p174, %f1315, 0f00000000;
mov.f32 %f1702, %f1701;
mov.f32 %f1703, %f1701;
@%p174 bra BB0_242;
cvt.rzi.s32.f32 %r812, %f1316;
neg.f32 %f1327, %f62;
neg.f32 %f1326, %f63;
neg.f32 %f1325, %f64;
mov.u32 %r813, 6;
mov.u32 %r814, 0;
// inline asm
call (%f1703, %f1702, %f1701, %f1324), _rt_texture_get_base_id, (%r812, %r813, %f1325, %f1326, %f1327, %r814);
// inline asm
BB0_242:
mul.f32 %f1704, %f1311, %f1703;
mul.f32 %f1705, %f1312, %f1702;
mul.f32 %f1706, %f1313, %f1701;
BB0_243:
mov.u32 %r891, 12;
mov.u32 %r890, 2;
ld.global.v2.u32 {%r819, %r820}, [pixelID];
cvt.u64.u32 %rd239, %r819;
cvt.u64.u32 %rd240, %r820;
// inline asm
call (%rd237), _rt_buffer_get_64, (%rd223, %r890, %r891, %rd239, %rd240, %rd227, %rd227);
// inline asm
st.f32 [%rd237+8], %f1706;
st.f32 [%rd237+4], %f1705;
st.f32 [%rd237], %f1704;
ld.global.v2.u32 {%r823, %r824}, [pixelID];
cvt.u64.u32 %rd245, %r823;
cvt.u64.u32 %rd246, %r824;
mov.u64 %rd250, accumBufferLerp;
cvta.global.u64 %rd244, %rd250;
// inline asm
call (%rd243), _rt_buffer_get_64, (%rd244, %r890, %r891, %rd245, %rd246, %rd227, %rd227);
// inline asm
ld.f32 %f1328, [%rd243+8];
ld.f32 %f1329, [%rd243+4];
ld.f32 %f1330, [%rd243];
sub.f32 %f1331, %f1704, %f1330;
sub.f32 %f1332, %f1705, %f1329;
sub.f32 %f1333, %f1706, %f1328;
ld.global.f32 %f1334, [integration];
fma.rn.f32 %f419, %f1334, %f1331, %f1330;
fma.rn.f32 %f420, %f1334, %f1332, %f1329;
fma.rn.f32 %f421, %f1334, %f1333, %f1328;
ld.global.u32 %r827, [finalBounceFlag];
setp.eq.s32 %p175, %r827, 0;
@%p175 bra BB0_245;
mov.u32 %r893, 12;
mov.u32 %r892, 2;
ld.global.v2.u32 {%r830, %r831}, [pixelID];
cvt.u64.u32 %rd253, %r830;
cvt.u64.u32 %rd254, %r831;
// inline asm
call (%rd251), _rt_buffer_get_64, (%rd244, %r892, %r893, %rd253, %rd254, %rd227, %rd227);
// inline asm
st.f32 [%rd251+8], %f421;
st.f32 [%rd251+4], %f420;
st.f32 [%rd251], %f419;
BB0_245:
mov.u32 %r894, 2;
ld.global.v2.u32 {%r836, %r837}, [pixelID];
cvt.u64.u32 %rd260, %r836;
cvt.u64.u32 %rd261, %r837;
mov.u64 %rd264, accumBufferLerpAO;
cvta.global.u64 %rd259, %rd264;
mov.u32 %r835, 4;
// inline asm
call (%rd258), _rt_buffer_get_64, (%rd259, %r894, %r835, %rd260, %rd261, %rd227, %rd227);
// inline asm
ld.f32 %f1707, [%rd258];
ld.global.u32 %r840, [bounceFlag];
setp.ne.s32 %p176, %r840, 0;
@%p176 bra BB0_247;
mov.u32 %r896, 2;
ld.global.f32 %f1335, [integration];
sub.f32 %f1336, %f1700, %f1707;
fma.rn.f32 %f1707, %f1336, %f1335, %f1707;
ld.global.v2.u32 {%r843, %r844}, [pixelID];
cvt.u64.u32 %rd267, %r843;
cvt.u64.u32 %rd268, %r844;
// inline asm
call (%rd265), _rt_buffer_get_64, (%rd259, %r896, %r835, %rd267, %rd268, %rd227, %rd227);
// inline asm
st.f32 [%rd265], %f1707;
BB0_247:
ld.global.f32 %f1339, [aoIntensity];
mul.f32 %f1340, %f1339, 0f3F000000;
add.f32 %f1341, %f1707, 0fBF800000;
fma.rn.f32 %f1342, %f1341, %f1340, 0f3F800000;
cvt.sat.f32.f32 %f1343, %f1342;
mul.f32 %f1344, %f419, %f1343;
mul.f32 %f1345, %f420, %f1343;
mul.f32 %f1346, %f421, %f1343;
ld.global.v2.u32 {%r847, %r848}, [pixelID];
cvt.u64.u32 %rd57, %r847;
ld.global.u32 %r850, [resolution+4];
add.s32 %r851, %r850, -1;
sub.s32 %r287, %r851, %r848;
ld.global.f32 %f1347, [cameraExp];
mul.f32 %f425, %f1344, %f1347;
mul.f32 %f426, %f1345, %f1347;
mul.f32 %f427, %f1346, %f1347;
mov.f32 %f1348, 0f3E68BA2E;
cvt.rzi.f32.f32 %f1349, %f1348;
fma.rn.f32 %f1350, %f1349, 0fC0000000, 0f3EE8BA2E;
abs.f32 %f428, %f1350;
abs.f32 %f429, %f425;
setp.lt.f32 %p177, %f429, 0f00800000;
mul.f32 %f1351, %f429, 0f4B800000;
selp.f32 %f1352, 0fC3170000, 0fC2FE0000, %p177;
selp.f32 %f1353, %f1351, %f429, %p177;
mov.b32 %r853, %f1353;
and.b32 %r854, %r853, 8388607;
or.b32 %r855, %r854, 1065353216;
mov.b32 %f1354, %r855;
shr.u32 %r856, %r853, 23;
cvt.rn.f32.u32 %f1355, %r856;
add.f32 %f1356, %f1352, %f1355;
setp.gt.f32 %p178, %f1354, 0f3FB504F3;
mul.f32 %f1357, %f1354, 0f3F000000;
add.f32 %f1358, %f1356, 0f3F800000;
selp.f32 %f1359, %f1357, %f1354, %p178;
selp.f32 %f1360, %f1358, %f1356, %p178;
add.f32 %f1361, %f1359, 0fBF800000;
add.f32 %f1338, %f1359, 0f3F800000;
// inline asm
rcp.approx.ftz.f32 %f1337,%f1338;
// inline asm
add.f32 %f1362, %f1361, %f1361;
mul.f32 %f1363, %f1337, %f1362;
mul.f32 %f1364, %f1363, %f1363;
mov.f32 %f1365, 0f3C4CAF63;
mov.f32 %f1366, 0f3B18F0FE;
fma.rn.f32 %f1367, %f1366, %f1364, %f1365;
mov.f32 %f1368, 0f3DAAAABD;
fma.rn.f32 %f1369, %f1367, %f1364, %f1368;
mul.rn.f32 %f1370, %f1369, %f1364;
mul.rn.f32 %f1371, %f1370, %f1363;
sub.f32 %f1372, %f1361, %f1363;
neg.f32 %f1373, %f1363;
add.f32 %f1374, %f1372, %f1372;
fma.rn.f32 %f1375, %f1373, %f1361, %f1374;
mul.rn.f32 %f1376, %f1337, %f1375;
add.f32 %f1377, %f1371, %f1363;
sub.f32 %f1378, %f1363, %f1377;
add.f32 %f1379, %f1371, %f1378;
add.f32 %f1380, %f1376, %f1379;
add.f32 %f1381, %f1377, %f1380;
sub.f32 %f1382, %f1377, %f1381;
add.f32 %f1383, %f1380, %f1382;
mov.f32 %f1384, 0f3F317200;
mul.rn.f32 %f1385, %f1360, %f1384;
mov.f32 %f1386, 0f35BFBE8E;
mul.rn.f32 %f1387, %f1360, %f1386;
add.f32 %f1388, %f1385, %f1381;
sub.f32 %f1389, %f1385, %f1388;
add.f32 %f1390, %f1381, %f1389;
add.f32 %f1391, %f1383, %f1390;
add.f32 %f1392, %f1387, %f1391;
add.f32 %f1393, %f1388, %f1392;
sub.f32 %f1394, %f1388, %f1393;
add.f32 %f1395, %f1392, %f1394;
mov.f32 %f1396, 0f3EE8BA2E;
mul.rn.f32 %f1397, %f1396, %f1393;
neg.f32 %f1398, %f1397;
fma.rn.f32 %f1399, %f1396, %f1393, %f1398;
fma.rn.f32 %f1400, %f1396, %f1395, %f1399;
mov.f32 %f1401, 0f00000000;
fma.rn.f32 %f1402, %f1401, %f1393, %f1400;
add.rn.f32 %f1403, %f1397, %f1402;
neg.f32 %f1404, %f1403;
add.rn.f32 %f1405, %f1397, %f1404;
add.rn.f32 %f1406, %f1405, %f1402;
mov.b32 %r857, %f1403;
setp.eq.s32 %p179, %r857, 1118925336;
add.s32 %r858, %r857, -1;
mov.b32 %f1407, %r858;
add.f32 %f1408, %f1406, 0f37000000;
selp.f32 %f1409, %f1407, %f1403, %p179;
selp.f32 %f430, %f1408, %f1406, %p179;
mul.f32 %f1410, %f1409, 0f3FB8AA3B;
cvt.rzi.f32.f32 %f1411, %f1410;
mov.f32 %f1412, 0fBF317200;
fma.rn.f32 %f1413, %f1411, %f1412, %f1409;
mov.f32 %f1414, 0fB5BFBE8E;
fma.rn.f32 %f1415, %f1411, %f1414, %f1413;
mul.f32 %f1416, %f1415, 0f3FB8AA3B;
ex2.approx.ftz.f32 %f1417, %f1416;
add.f32 %f1418, %f1411, 0f00000000;
ex2.approx.f32 %f1419, %f1418;
mul.f32 %f1420, %f1417, %f1419;
setp.lt.f32 %p180, %f1409, 0fC2D20000;
selp.f32 %f1421, 0f00000000, %f1420, %p180;
setp.gt.f32 %p181, %f1409, 0f42D20000;
selp.f32 %f1708, 0f7F800000, %f1421, %p181;
setp.eq.f32 %p182, %f1708, 0f7F800000;
@%p182 bra BB0_249;
fma.rn.f32 %f1708, %f1708, %f430, %f1708;
BB0_249:
setp.lt.f32 %p183, %f425, 0f00000000;
setp.eq.f32 %p184, %f428, 0f3F800000;
and.pred %p3, %p183, %p184;
mov.b32 %r859, %f1708;
xor.b32 %r860, %r859, -2147483648;
mov.b32 %f1422, %r860;
selp.f32 %f1710, %f1422, %f1708, %p3;
setp.eq.f32 %p185, %f425, 0f00000000;
@%p185 bra BB0_252;
bra.uni BB0_250;
BB0_252:
add.f32 %f1425, %f425, %f425;
selp.f32 %f1710, %f1425, 0f00000000, %p184;
bra.uni BB0_253;
BB0_250:
setp.geu.f32 %p186, %f425, 0f00000000;
@%p186 bra BB0_253;
cvt.rzi.f32.f32 %f1424, %f1396;
setp.neu.f32 %p187, %f1424, 0f3EE8BA2E;
selp.f32 %f1710, 0f7FFFFFFF, %f1710, %p187;
BB0_253:
add.f32 %f1426, %f429, 0f3EE8BA2E;
mov.b32 %r861, %f1426;
setp.lt.s32 %p189, %r861, 2139095040;
@%p189 bra BB0_258;
setp.gtu.f32 %p190, %f429, 0f7F800000;
@%p190 bra BB0_257;
bra.uni BB0_255;
BB0_257:
add.f32 %f1710, %f425, 0f3EE8BA2E;
bra.uni BB0_258;
BB0_255:
setp.neu.f32 %p191, %f429, 0f7F800000;
@%p191 bra BB0_258;
selp.f32 %f1710, 0fFF800000, 0f7F800000, %p3;
BB0_258:
setp.eq.f32 %p192, %f425, 0f3F800000;
selp.f32 %f1429, 0f3F800000, %f1710, %p192;
cvt.sat.f32.f32 %f441, %f1429;
abs.f32 %f442, %f426;
setp.lt.f32 %p193, %f442, 0f00800000;
mul.f32 %f1430, %f442, 0f4B800000;
selp.f32 %f1431, 0fC3170000, 0fC2FE0000, %p193;
selp.f32 %f1432, %f1430, %f442, %p193;
mov.b32 %r862, %f1432;
and.b32 %r863, %r862, 8388607;
or.b32 %r864, %r863, 1065353216;
mov.b32 %f1433, %r864;
shr.u32 %r865, %r862, 23;
cvt.rn.f32.u32 %f1434, %r865;
add.f32 %f1435, %f1431, %f1434;
setp.gt.f32 %p194, %f1433, 0f3FB504F3;
mul.f32 %f1436, %f1433, 0f3F000000;
add.f32 %f1437, %f1435, 0f3F800000;
selp.f32 %f1438, %f1436, %f1433, %p194;
selp.f32 %f1439, %f1437, %f1435, %p194;
add.f32 %f1440, %f1438, 0fBF800000;
add.f32 %f1428, %f1438, 0f3F800000;
// inline asm
rcp.approx.ftz.f32 %f1427,%f1428;
// inline asm
add.f32 %f1441, %f1440, %f1440;
mul.f32 %f1442, %f1427, %f1441;
mul.f32 %f1443, %f1442, %f1442;
fma.rn.f32 %f1446, %f1366, %f1443, %f1365;
fma.rn.f32 %f1448, %f1446, %f1443, %f1368;
mul.rn.f32 %f1449, %f1448, %f1443;
mul.rn.f32 %f1450, %f1449, %f1442;
sub.f32 %f1451, %f1440, %f1442;
neg.f32 %f1452, %f1442;
add.f32 %f1453, %f1451, %f1451;
fma.rn.f32 %f1454, %f1452, %f1440, %f1453;
mul.rn.f32 %f1455, %f1427, %f1454;
add.f32 %f1456, %f1450, %f1442;
sub.f32 %f1457, %f1442, %f1456;
add.f32 %f1458, %f1450, %f1457;
add.f32 %f1459, %f1455, %f1458;
add.f32 %f1460, %f1456, %f1459;
sub.f32 %f1461, %f1456, %f1460;
add.f32 %f1462, %f1459, %f1461;
mul.rn.f32 %f1464, %f1439, %f1384;
mul.rn.f32 %f1466, %f1439, %f1386;
add.f32 %f1467, %f1464, %f1460;
sub.f32 %f1468, %f1464, %f1467;
add.f32 %f1469, %f1460, %f1468;
add.f32 %f1470, %f1462, %f1469;
add.f32 %f1471, %f1466, %f1470;
add.f32 %f1472, %f1467, %f1471;
sub.f32 %f1473, %f1467, %f1472;
add.f32 %f1474, %f1471, %f1473;
mul.rn.f32 %f1476, %f1396, %f1472;
neg.f32 %f1477, %f1476;
fma.rn.f32 %f1478, %f1396, %f1472, %f1477;
fma.rn.f32 %f1479, %f1396, %f1474, %f1478;
fma.rn.f32 %f1481, %f1401, %f1472, %f1479;
add.rn.f32 %f1482, %f1476, %f1481;
neg.f32 %f1483, %f1482;
add.rn.f32 %f1484, %f1476, %f1483;
add.rn.f32 %f1485, %f1484, %f1481;
mov.b32 %r866, %f1482;
setp.eq.s32 %p195, %r866, 1118925336;
add.s32 %r867, %r866, -1;
mov.b32 %f1486, %r867;
add.f32 %f1487, %f1485, 0f37000000;
selp.f32 %f1488, %f1486, %f1482, %p195;
selp.f32 %f443, %f1487, %f1485, %p195;
mul.f32 %f1489, %f1488, 0f3FB8AA3B;
cvt.rzi.f32.f32 %f1490, %f1489;
fma.rn.f32 %f1492, %f1490, %f1412, %f1488;
fma.rn.f32 %f1494, %f1490, %f1414, %f1492;
mul.f32 %f1495, %f1494, 0f3FB8AA3B;
ex2.approx.ftz.f32 %f1496, %f1495;
add.f32 %f1497, %f1490, 0f00000000;
ex2.approx.f32 %f1498, %f1497;
mul.f32 %f1499, %f1496, %f1498;
setp.lt.f32 %p196, %f1488, 0fC2D20000;
selp.f32 %f1500, 0f00000000, %f1499, %p196;
setp.gt.f32 %p197, %f1488, 0f42D20000;
selp.f32 %f1711, 0f7F800000, %f1500, %p197;
setp.eq.f32 %p198, %f1711, 0f7F800000;
@%p198 bra BB0_260;
fma.rn.f32 %f1711, %f1711, %f443, %f1711;
BB0_260:
setp.lt.f32 %p199, %f426, 0f00000000;
and.pred %p4, %p199, %p184;
mov.b32 %r868, %f1711;
xor.b32 %r869, %r868, -2147483648;
mov.b32 %f1501, %r869;
selp.f32 %f1713, %f1501, %f1711, %p4;
setp.eq.f32 %p201, %f426, 0f00000000;
@%p201 bra BB0_263;
bra.uni BB0_261;
BB0_263:
add.f32 %f1504, %f426, %f426;
selp.f32 %f1713, %f1504, 0f00000000, %p184;
bra.uni BB0_264;
BB0_261:
setp.geu.f32 %p202, %f426, 0f00000000;
@%p202 bra BB0_264;
cvt.rzi.f32.f32 %f1503, %f1396;
setp.neu.f32 %p203, %f1503, 0f3EE8BA2E;
selp.f32 %f1713, 0f7FFFFFFF, %f1713, %p203;
BB0_264:
add.f32 %f1505, %f442, 0f3EE8BA2E;
mov.b32 %r870, %f1505;
setp.lt.s32 %p205, %r870, 2139095040;
@%p205 bra BB0_269;
setp.gtu.f32 %p206, %f442, 0f7F800000;
@%p206 bra BB0_268;
bra.uni BB0_266;
BB0_268:
add.f32 %f1713, %f426, 0f3EE8BA2E;
bra.uni BB0_269;
BB0_266:
setp.neu.f32 %p207, %f442, 0f7F800000;
@%p207 bra BB0_269;
selp.f32 %f1713, 0fFF800000, 0f7F800000, %p4;
BB0_269:
setp.eq.f32 %p208, %f426, 0f3F800000;
selp.f32 %f1508, 0f3F800000, %f1713, %p208;
cvt.sat.f32.f32 %f454, %f1508;
abs.f32 %f455, %f427;
setp.lt.f32 %p209, %f455, 0f00800000;
mul.f32 %f1509, %f455, 0f4B800000;
selp.f32 %f1510, 0fC3170000, 0fC2FE0000, %p209;
selp.f32 %f1511, %f1509, %f455, %p209;
mov.b32 %r871, %f1511;
and.b32 %r872, %r871, 8388607;
or.b32 %r873, %r872, 1065353216;
mov.b32 %f1512, %r873;
shr.u32 %r874, %r871, 23;
cvt.rn.f32.u32 %f1513, %r874;
add.f32 %f1514, %f1510, %f1513;
setp.gt.f32 %p210, %f1512, 0f3FB504F3;
mul.f32 %f1515, %f1512, 0f3F000000;
add.f32 %f1516, %f1514, 0f3F800000;
selp.f32 %f1517, %f1515, %f1512, %p210;
selp.f32 %f1518, %f1516, %f1514, %p210;
add.f32 %f1519, %f1517, 0fBF800000;
add.f32 %f1507, %f1517, 0f3F800000;
// inline asm
rcp.approx.ftz.f32 %f1506,%f1507;
// inline asm
add.f32 %f1520, %f1519, %f1519;
mul.f32 %f1521, %f1506, %f1520;
mul.f32 %f1522, %f1521, %f1521;
fma.rn.f32 %f1525, %f1366, %f1522, %f1365;
fma.rn.f32 %f1527, %f1525, %f1522, %f1368;
mul.rn.f32 %f1528, %f1527, %f1522;
mul.rn.f32 %f1529, %f1528, %f1521;
sub.f32 %f1530, %f1519, %f1521;
neg.f32 %f1531, %f1521;
add.f32 %f1532, %f1530, %f1530;
fma.rn.f32 %f1533, %f1531, %f1519, %f1532;
mul.rn.f32 %f1534, %f1506, %f1533;
add.f32 %f1535, %f1529, %f1521;
sub.f32 %f1536, %f1521, %f1535;
add.f32 %f1537, %f1529, %f1536;
add.f32 %f1538, %f1534, %f1537;
add.f32 %f1539, %f1535, %f1538;
sub.f32 %f1540, %f1535, %f1539;
add.f32 %f1541, %f1538, %f1540;
mul.rn.f32 %f1543, %f1518, %f1384;
mul.rn.f32 %f1545, %f1518, %f1386;
add.f32 %f1546, %f1543, %f1539;
sub.f32 %f1547, %f1543, %f1546;
add.f32 %f1548, %f1539, %f1547;
add.f32 %f1549, %f1541, %f1548;
add.f32 %f1550, %f1545, %f1549;
add.f32 %f1551, %f1546, %f1550;
sub.f32 %f1552, %f1546, %f1551;
add.f32 %f1553, %f1550, %f1552;
mul.rn.f32 %f1555, %f1396, %f1551;
neg.f32 %f1556, %f1555;
fma.rn.f32 %f1557, %f1396, %f1551, %f1556;
fma.rn.f32 %f1558, %f1396, %f1553, %f1557;
fma.rn.f32 %f1560, %f1401, %f1551, %f1558;
add.rn.f32 %f1561, %f1555, %f1560;
neg.f32 %f1562, %f1561;
add.rn.f32 %f1563, %f1555, %f1562;
add.rn.f32 %f1564, %f1563, %f1560;
mov.b32 %r875, %f1561;
setp.eq.s32 %p211, %r875, 1118925336;
add.s32 %r876, %r875, -1;
mov.b32 %f1565, %r876;
add.f32 %f1566, %f1564, 0f37000000;
selp.f32 %f1567, %f1565, %f1561, %p211;
selp.f32 %f456, %f1566, %f1564, %p211;
mul.f32 %f1568, %f1567, 0f3FB8AA3B;
cvt.rzi.f32.f32 %f1569, %f1568;
fma.rn.f32 %f1571, %f1569, %f1412, %f1567;
fma.rn.f32 %f1573, %f1569, %f1414, %f1571;
mul.f32 %f1574, %f1573, 0f3FB8AA3B;
ex2.approx.ftz.f32 %f1575, %f1574;
add.f32 %f1576, %f1569, 0f00000000;
ex2.approx.f32 %f1577, %f1576;
mul.f32 %f1578, %f1575, %f1577;
setp.lt.f32 %p212, %f1567, 0fC2D20000;
selp.f32 %f1579, 0f00000000, %f1578, %p212;
setp.gt.f32 %p213, %f1567, 0f42D20000;
selp.f32 %f1714, 0f7F800000, %f1579, %p213;
setp.eq.f32 %p214, %f1714, 0f7F800000;
@%p214 bra BB0_271;
fma.rn.f32 %f1714, %f1714, %f456, %f1714;
BB0_271:
setp.lt.f32 %p215, %f427, 0f00000000;
and.pred %p5, %p215, %p184;
mov.b32 %r877, %f1714;
xor.b32 %r878, %r877, -2147483648;
mov.b32 %f1580, %r878;
selp.f32 %f1716, %f1580, %f1714, %p5;
setp.eq.f32 %p217, %f427, 0f00000000;
@%p217 bra BB0_274;
bra.uni BB0_272;
BB0_274:
add.f32 %f1583, %f427, %f427;
selp.f32 %f1716, %f1583, 0f00000000, %p184;
bra.uni BB0_275;
BB0_272:
setp.geu.f32 %p218, %f427, 0f00000000;
@%p218 bra BB0_275;
cvt.rzi.f32.f32 %f1582, %f1396;
setp.neu.f32 %p219, %f1582, 0f3EE8BA2E;
selp.f32 %f1716, 0f7FFFFFFF, %f1716, %p219;
BB0_275:
add.f32 %f1584, %f455, 0f3EE8BA2E;
mov.b32 %r879, %f1584;
setp.lt.s32 %p221, %r879, 2139095040;
@%p221 bra BB0_280;
setp.gtu.f32 %p222, %f455, 0f7F800000;
@%p222 bra BB0_279;
bra.uni BB0_277;
BB0_279:
add.f32 %f1716, %f427, 0f3EE8BA2E;
bra.uni BB0_280;
BB0_277:
setp.neu.f32 %p223, %f455, 0f7F800000;
@%p223 bra BB0_280;
selp.f32 %f1716, 0fFF800000, 0f7F800000, %p5;
BB0_280:
mov.u32 %r895, 2;
setp.eq.f32 %p224, %f427, 0f3F800000;
selp.f32 %f1585, 0f3F800000, %f1716, %p224;
cvt.sat.f32.f32 %f1586, %f1585;
cvt.u64.u32 %rd275, %r287;
mov.u64 %rd278, colorOutput;
cvta.global.u64 %rd273, %rd278;
// inline asm
call (%rd272), _rt_buffer_get_64, (%rd273, %r895, %r835, %rd57, %rd275, %rd227, %rd227);
// inline asm
mul.f32 %f1587, %f441, 0f437F0000;
cvt.rzi.u32.f32 %r882, %f1587;
mul.f32 %f1588, %f454, 0f437F0000;
cvt.rzi.u32.f32 %r883, %f1588;
mul.f32 %f1589, %f1586, 0f437F0000;
cvt.rzi.u32.f32 %r884, %f1589;
cvt.u16.u32 %rs10, %r884;
cvt.u16.u32 %rs11, %r883;
cvt.u16.u32 %rs12, %r882;
mov.u16 %rs13, 255;
st.v4.u8 [%rd272], {%rs12, %rs11, %rs10, %rs13};
ret;
}