2961 lines
86 KiB
Plaintext
2961 lines
86 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<174>;
|
|
.reg .b16 %rs<11>;
|
|
.reg .f32 %f<1426>;
|
|
.reg .b32 %r<950>;
|
|
.reg .b64 %rd<303>;
|
|
|
|
|
|
mov.u64 %rd302, __local_depot0;
|
|
cvta.local.u64 %SP, %rd302;
|
|
ld.global.v2.u32 {%r297, %r298}, [pixelID];
|
|
ld.global.v4.u32 {%r300, %r301, %r302, %r303}, [rndOffset];
|
|
add.s32 %r305, %r300, %r297;
|
|
and.b32 %r306, %r305, 255;
|
|
add.s32 %r309, %r301, %r298;
|
|
and.b32 %r310, %r309, 255;
|
|
add.s32 %r312, %r302, %r297;
|
|
and.b32 %r313, %r312, 255;
|
|
add.s32 %r315, %r303, %r298;
|
|
and.b32 %r316, %r315, 255;
|
|
cvt.u64.u32 %rd71, %r306;
|
|
cvt.u64.u32 %rd72, %r310;
|
|
mov.u64 %rd87, rnd_seeds;
|
|
cvta.global.u64 %rd58, %rd87;
|
|
mov.u32 %r295, 2;
|
|
mov.u32 %r296, 3;
|
|
mov.u64 %rd293, 0;
|
|
// inline asm
|
|
call (%rd57), _rt_buffer_get_64, (%rd58, %r295, %r296, %rd71, %rd72, %rd293, %rd293);
|
|
// inline asm
|
|
ld.u8 %rs4, [%rd57];
|
|
cvt.rn.f32.u16 %f425, %rs4;
|
|
div.rn.f32 %f1, %f425, 0f437F0000;
|
|
// inline asm
|
|
call (%rd63), _rt_buffer_get_64, (%rd58, %r295, %r296, %rd71, %rd72, %rd293, %rd293);
|
|
// inline asm
|
|
ld.u8 %rs5, [%rd63+1];
|
|
cvt.rn.f32.u16 %f426, %rs5;
|
|
div.rn.f32 %f2, %f426, 0f437F0000;
|
|
// inline asm
|
|
call (%rd69), _rt_buffer_get_64, (%rd58, %r295, %r296, %rd71, %rd72, %rd293, %rd293);
|
|
// inline asm
|
|
ld.u8 %rs6, [%rd69+2];
|
|
cvt.rn.f32.u16 %f427, %rs6;
|
|
div.rn.f32 %f3, %f427, 0f437F0000;
|
|
cvt.u64.u32 %rd83, %r313;
|
|
cvt.u64.u32 %rd84, %r316;
|
|
// inline asm
|
|
call (%rd75), _rt_buffer_get_64, (%rd58, %r295, %r296, %rd83, %rd84, %rd293, %rd293);
|
|
// inline asm
|
|
ld.u8 %rs1, [%rd75];
|
|
// inline asm
|
|
call (%rd81), _rt_buffer_get_64, (%rd58, %r295, %r296, %rd83, %rd84, %rd293, %rd293);
|
|
// inline asm
|
|
ld.global.u32 %r317, [bounceFlag];
|
|
setp.eq.s32 %p3, %r317, 0;
|
|
@%p3 bra BB0_48;
|
|
|
|
ld.u8 %rs7, [%rd81+1];
|
|
cvt.rn.f32.u16 %f428, %rs7;
|
|
div.rn.f32 %f429, %f428, 0f437F0000;
|
|
cvt.rn.f32.u16 %f430, %rs1;
|
|
div.rn.f32 %f431, %f430, 0f437F0000;
|
|
ld.global.v2.u32 {%r322, %r323}, [pixelID];
|
|
cvt.u64.u32 %rd90, %r322;
|
|
cvt.u64.u32 %rd91, %r323;
|
|
mov.u64 %rd100, GBufferPos;
|
|
cvta.global.u64 %rd89, %rd100;
|
|
mov.u32 %r321, 12;
|
|
// inline asm
|
|
call (%rd88), _rt_buffer_get_64, (%rd89, %r295, %r321, %rd90, %rd91, %rd293, %rd293);
|
|
// inline asm
|
|
ld.f32 %f1335, [%rd88+8];
|
|
ld.f32 %f1334, [%rd88+4];
|
|
ld.f32 %f1333, [%rd88];
|
|
ld.global.v2.u32 {%r326, %r327}, [pixelID];
|
|
cvt.u64.u32 %rd96, %r326;
|
|
cvt.u64.u32 %rd97, %r327;
|
|
mov.u64 %rd101, GBufferNormal;
|
|
cvta.global.u64 %rd95, %rd101;
|
|
// inline asm
|
|
call (%rd94), _rt_buffer_get_64, (%rd95, %r295, %r321, %rd96, %rd97, %rd293, %rd293);
|
|
// inline asm
|
|
ld.f32 %f9, [%rd94+4];
|
|
ld.f32 %f8, [%rd94+8];
|
|
ld.f32 %f7, [%rd94];
|
|
abs.f32 %f432, %f7;
|
|
abs.f32 %f433, %f8;
|
|
setp.gt.f32 %p4, %f432, %f433;
|
|
neg.f32 %f434, %f9;
|
|
neg.f32 %f435, %f8;
|
|
selp.f32 %f436, %f434, 0f00000000, %p4;
|
|
selp.f32 %f437, %f7, %f435, %p4;
|
|
selp.f32 %f438, 0f00000000, %f9, %p4;
|
|
mul.f32 %f439, %f437, %f437;
|
|
fma.rn.f32 %f440, %f436, %f436, %f439;
|
|
fma.rn.f32 %f441, %f438, %f438, %f440;
|
|
sqrt.rn.f32 %f442, %f441;
|
|
rcp.rn.f32 %f443, %f442;
|
|
mul.f32 %f10, %f436, %f443;
|
|
mul.f32 %f11, %f437, %f443;
|
|
mul.f32 %f12, %f438, %f443;
|
|
sqrt.rn.f32 %f13, %f431;
|
|
mul.f32 %f1324, %f429, 0f40C90FDB;
|
|
add.u64 %rd102, %SP, 0;
|
|
cvta.to.local.u64 %rd2, %rd102;
|
|
abs.f32 %f15, %f1324;
|
|
setp.neu.f32 %p5, %f15, 0f7F800000;
|
|
mov.f32 %f1318, %f1324;
|
|
@%p5 bra BB0_3;
|
|
|
|
mov.f32 %f444, 0f00000000;
|
|
mul.rn.f32 %f1318, %f1324, %f444;
|
|
|
|
BB0_3:
|
|
mul.f32 %f445, %f1318, 0f3F22F983;
|
|
cvt.rni.s32.f32 %r879, %f445;
|
|
cvt.rn.f32.s32 %f446, %r879;
|
|
neg.f32 %f447, %f446;
|
|
mov.f32 %f448, 0f3FC90FDA;
|
|
fma.rn.f32 %f449, %f447, %f448, %f1318;
|
|
mov.f32 %f450, 0f33A22168;
|
|
fma.rn.f32 %f451, %f447, %f450, %f449;
|
|
mov.f32 %f452, 0f27C234C5;
|
|
fma.rn.f32 %f1319, %f447, %f452, %f451;
|
|
abs.f32 %f453, %f1318;
|
|
add.s64 %rd3, %rd2, 24;
|
|
setp.leu.f32 %p6, %f453, 0f47CE4780;
|
|
@%p6 bra BB0_14;
|
|
|
|
mov.b32 %r2, %f1318;
|
|
shr.u32 %r3, %r2, 23;
|
|
shl.b32 %r332, %r2, 8;
|
|
or.b32 %r4, %r332, -2147483648;
|
|
mov.u32 %r871, 0;
|
|
mov.u64 %rd286, __cudart_i2opi_f;
|
|
mov.u32 %r870, -6;
|
|
mov.u64 %rd287, %rd2;
|
|
|
|
BB0_5:
|
|
.pragma "nounroll";
|
|
ld.const.u32 %r335, [%rd286];
|
|
// inline asm
|
|
{
|
|
mad.lo.cc.u32 %r333, %r335, %r4, %r871;
|
|
madc.hi.u32 %r871, %r335, %r4, 0;
|
|
}
|
|
// inline asm
|
|
st.local.u32 [%rd287], %r333;
|
|
add.s64 %rd287, %rd287, 4;
|
|
add.s64 %rd286, %rd286, 4;
|
|
add.s32 %r870, %r870, 1;
|
|
setp.ne.s32 %p7, %r870, 0;
|
|
@%p7 bra BB0_5;
|
|
|
|
and.b32 %r338, %r3, 255;
|
|
add.s32 %r339, %r338, -128;
|
|
shr.u32 %r340, %r339, 5;
|
|
and.b32 %r9, %r2, -2147483648;
|
|
st.local.u32 [%rd3], %r871;
|
|
mov.u32 %r341, 6;
|
|
sub.s32 %r342, %r341, %r340;
|
|
mul.wide.s32 %rd104, %r342, 4;
|
|
add.s64 %rd8, %rd2, %rd104;
|
|
ld.local.u32 %r872, [%rd8];
|
|
ld.local.u32 %r873, [%rd8+-4];
|
|
and.b32 %r12, %r3, 31;
|
|
setp.eq.s32 %p8, %r12, 0;
|
|
@%p8 bra BB0_8;
|
|
|
|
mov.u32 %r343, 32;
|
|
sub.s32 %r344, %r343, %r12;
|
|
shr.u32 %r345, %r873, %r344;
|
|
shl.b32 %r346, %r872, %r12;
|
|
add.s32 %r872, %r345, %r346;
|
|
ld.local.u32 %r347, [%rd8+-8];
|
|
shr.u32 %r348, %r347, %r344;
|
|
shl.b32 %r349, %r873, %r12;
|
|
add.s32 %r873, %r348, %r349;
|
|
|
|
BB0_8:
|
|
shr.u32 %r350, %r873, 30;
|
|
shl.b32 %r351, %r872, 2;
|
|
add.s32 %r874, %r350, %r351;
|
|
shl.b32 %r18, %r873, 2;
|
|
shr.u32 %r352, %r874, 31;
|
|
shr.u32 %r353, %r872, 30;
|
|
add.s32 %r19, %r352, %r353;
|
|
setp.eq.s32 %p9, %r352, 0;
|
|
@%p9 bra BB0_9;
|
|
|
|
not.b32 %r354, %r874;
|
|
neg.s32 %r876, %r18;
|
|
setp.eq.s32 %p10, %r18, 0;
|
|
selp.u32 %r355, 1, 0, %p10;
|
|
add.s32 %r874, %r355, %r354;
|
|
xor.b32 %r875, %r9, -2147483648;
|
|
bra.uni BB0_11;
|
|
|
|
BB0_48:
|
|
ld.global.f32 %f1333, [cameraPos];
|
|
ld.global.f32 %f1334, [cameraPos+4];
|
|
ld.global.f32 %f1335, [cameraPos+8];
|
|
ld.global.v2.u32 {%r422, %r423}, [resolution];
|
|
cvt.rn.f32.u32 %f515, %r422;
|
|
cvt.rn.f32.u32 %f516, %r423;
|
|
div.rn.f32 %f517, %f515, %f516;
|
|
ld.global.v2.u32 {%r426, %r427}, [pixelID];
|
|
cvt.rn.f32.u32 %f518, %r426;
|
|
add.f32 %f519, %f1, %f518;
|
|
div.rn.f32 %f520, %f519, %f515;
|
|
cvt.rn.f32.u32 %f521, %r427;
|
|
add.f32 %f522, %f2, %f521;
|
|
div.rn.f32 %f523, %f522, %f516;
|
|
fma.rn.f32 %f524, %f520, 0f40000000, 0fBF800000;
|
|
fma.rn.f32 %f525, %f523, 0f40000000, 0fBF800000;
|
|
ld.global.f32 %f526, [cameraFov];
|
|
mul.f32 %f527, %f517, %f526;
|
|
ld.global.f32 %f528, [cameraRht];
|
|
mul.f32 %f529, %f524, %f528;
|
|
ld.global.f32 %f530, [cameraRht+4];
|
|
mul.f32 %f531, %f524, %f530;
|
|
ld.global.f32 %f532, [cameraRht+8];
|
|
mul.f32 %f533, %f524, %f532;
|
|
ld.global.f32 %f534, [cameraFwd];
|
|
fma.rn.f32 %f535, %f527, %f529, %f534;
|
|
ld.global.f32 %f536, [cameraFwd+4];
|
|
fma.rn.f32 %f537, %f527, %f531, %f536;
|
|
ld.global.f32 %f538, [cameraFwd+8];
|
|
fma.rn.f32 %f539, %f527, %f533, %f538;
|
|
ld.global.f32 %f540, [cameraUp];
|
|
mul.f32 %f541, %f525, %f540;
|
|
ld.global.f32 %f542, [cameraUp+4];
|
|
mul.f32 %f543, %f525, %f542;
|
|
ld.global.f32 %f544, [cameraUp+8];
|
|
mul.f32 %f545, %f525, %f544;
|
|
fma.rn.f32 %f546, %f526, %f541, %f535;
|
|
fma.rn.f32 %f547, %f526, %f543, %f537;
|
|
fma.rn.f32 %f548, %f526, %f545, %f539;
|
|
mul.f32 %f549, %f547, %f547;
|
|
fma.rn.f32 %f550, %f546, %f546, %f549;
|
|
fma.rn.f32 %f551, %f548, %f548, %f550;
|
|
sqrt.rn.f32 %f552, %f551;
|
|
rcp.rn.f32 %f553, %f552;
|
|
mul.f32 %f62, %f546, %f553;
|
|
mul.f32 %f63, %f547, %f553;
|
|
mul.f32 %f64, %f553, %f548;
|
|
bra.uni BB0_49;
|
|
|
|
BB0_9:
|
|
mov.u32 %r875, %r9;
|
|
mov.u32 %r876, %r18;
|
|
|
|
BB0_11:
|
|
clz.b32 %r878, %r874;
|
|
setp.eq.s32 %p11, %r878, 0;
|
|
shl.b32 %r356, %r874, %r878;
|
|
mov.u32 %r357, 32;
|
|
sub.s32 %r358, %r357, %r878;
|
|
shr.u32 %r359, %r876, %r358;
|
|
add.s32 %r360, %r359, %r356;
|
|
selp.b32 %r27, %r874, %r360, %p11;
|
|
mov.u32 %r361, -921707870;
|
|
mul.hi.u32 %r877, %r27, %r361;
|
|
setp.eq.s32 %p12, %r9, 0;
|
|
neg.s32 %r362, %r19;
|
|
selp.b32 %r879, %r19, %r362, %p12;
|
|
setp.lt.s32 %p13, %r877, 1;
|
|
@%p13 bra BB0_13;
|
|
|
|
mul.lo.s32 %r363, %r27, -921707870;
|
|
shr.u32 %r364, %r363, 31;
|
|
shl.b32 %r365, %r877, 1;
|
|
add.s32 %r877, %r364, %r365;
|
|
add.s32 %r878, %r878, 1;
|
|
|
|
BB0_13:
|
|
mov.u32 %r366, 126;
|
|
sub.s32 %r367, %r366, %r878;
|
|
shl.b32 %r368, %r367, 23;
|
|
add.s32 %r369, %r877, 1;
|
|
shr.u32 %r370, %r369, 7;
|
|
add.s32 %r371, %r370, 1;
|
|
shr.u32 %r372, %r371, 1;
|
|
add.s32 %r373, %r372, %r368;
|
|
or.b32 %r374, %r373, %r875;
|
|
mov.b32 %f1319, %r374;
|
|
|
|
BB0_14:
|
|
mul.rn.f32 %f21, %f1319, %f1319;
|
|
add.s32 %r35, %r879, 1;
|
|
and.b32 %r36, %r35, 1;
|
|
setp.eq.s32 %p14, %r36, 0;
|
|
@%p14 bra BB0_16;
|
|
|
|
mov.f32 %f454, 0fBAB6061A;
|
|
mov.f32 %f455, 0f37CCF5CE;
|
|
fma.rn.f32 %f1320, %f455, %f21, %f454;
|
|
bra.uni BB0_17;
|
|
|
|
BB0_16:
|
|
mov.f32 %f456, 0f3C08839E;
|
|
mov.f32 %f457, 0fB94CA1F9;
|
|
fma.rn.f32 %f1320, %f457, %f21, %f456;
|
|
|
|
BB0_17:
|
|
@%p14 bra BB0_19;
|
|
|
|
mov.f32 %f458, 0f3D2AAAA5;
|
|
fma.rn.f32 %f459, %f1320, %f21, %f458;
|
|
mov.f32 %f460, 0fBF000000;
|
|
fma.rn.f32 %f1321, %f459, %f21, %f460;
|
|
bra.uni BB0_20;
|
|
|
|
BB0_19:
|
|
mov.f32 %f461, 0fBE2AAAA3;
|
|
fma.rn.f32 %f462, %f1320, %f21, %f461;
|
|
mov.f32 %f463, 0f00000000;
|
|
fma.rn.f32 %f1321, %f462, %f21, %f463;
|
|
|
|
BB0_20:
|
|
fma.rn.f32 %f1322, %f1321, %f1319, %f1319;
|
|
@%p14 bra BB0_22;
|
|
|
|
mov.f32 %f464, 0f3F800000;
|
|
fma.rn.f32 %f1322, %f1321, %f21, %f464;
|
|
|
|
BB0_22:
|
|
and.b32 %r375, %r35, 2;
|
|
setp.eq.s32 %p17, %r375, 0;
|
|
@%p17 bra BB0_24;
|
|
|
|
mov.f32 %f465, 0f00000000;
|
|
mov.f32 %f466, 0fBF800000;
|
|
fma.rn.f32 %f1322, %f1322, %f466, %f465;
|
|
|
|
BB0_24:
|
|
@%p5 bra BB0_26;
|
|
|
|
mov.f32 %f467, 0f00000000;
|
|
mul.rn.f32 %f1324, %f1324, %f467;
|
|
|
|
BB0_26:
|
|
mul.f32 %f468, %f1324, 0f3F22F983;
|
|
cvt.rni.s32.f32 %r889, %f468;
|
|
cvt.rn.f32.s32 %f469, %r889;
|
|
neg.f32 %f470, %f469;
|
|
fma.rn.f32 %f472, %f470, %f448, %f1324;
|
|
fma.rn.f32 %f474, %f470, %f450, %f472;
|
|
fma.rn.f32 %f1325, %f470, %f452, %f474;
|
|
abs.f32 %f476, %f1324;
|
|
setp.leu.f32 %p19, %f476, 0f47CE4780;
|
|
@%p19 bra BB0_37;
|
|
|
|
mov.b32 %r38, %f1324;
|
|
shr.u32 %r39, %r38, 23;
|
|
shl.b32 %r378, %r38, 8;
|
|
or.b32 %r40, %r378, -2147483648;
|
|
mov.u32 %r881, 0;
|
|
mov.u64 %rd288, __cudart_i2opi_f;
|
|
mov.u32 %r880, -6;
|
|
mov.u64 %rd289, %rd2;
|
|
|
|
BB0_28:
|
|
.pragma "nounroll";
|
|
ld.const.u32 %r381, [%rd288];
|
|
// inline asm
|
|
{
|
|
mad.lo.cc.u32 %r379, %r381, %r40, %r881;
|
|
madc.hi.u32 %r881, %r381, %r40, 0;
|
|
}
|
|
// inline asm
|
|
st.local.u32 [%rd289], %r379;
|
|
add.s64 %rd289, %rd289, 4;
|
|
add.s64 %rd288, %rd288, 4;
|
|
add.s32 %r880, %r880, 1;
|
|
setp.ne.s32 %p20, %r880, 0;
|
|
@%p20 bra BB0_28;
|
|
|
|
and.b32 %r384, %r39, 255;
|
|
add.s32 %r385, %r384, -128;
|
|
shr.u32 %r386, %r385, 5;
|
|
and.b32 %r45, %r38, -2147483648;
|
|
st.local.u32 [%rd3], %r881;
|
|
mov.u32 %r387, 6;
|
|
sub.s32 %r388, %r387, %r386;
|
|
mul.wide.s32 %rd106, %r388, 4;
|
|
add.s64 %rd13, %rd2, %rd106;
|
|
ld.local.u32 %r882, [%rd13];
|
|
ld.local.u32 %r883, [%rd13+-4];
|
|
and.b32 %r48, %r39, 31;
|
|
setp.eq.s32 %p21, %r48, 0;
|
|
@%p21 bra BB0_31;
|
|
|
|
mov.u32 %r389, 32;
|
|
sub.s32 %r390, %r389, %r48;
|
|
shr.u32 %r391, %r883, %r390;
|
|
shl.b32 %r392, %r882, %r48;
|
|
add.s32 %r882, %r391, %r392;
|
|
ld.local.u32 %r393, [%rd13+-8];
|
|
shr.u32 %r394, %r393, %r390;
|
|
shl.b32 %r395, %r883, %r48;
|
|
add.s32 %r883, %r394, %r395;
|
|
|
|
BB0_31:
|
|
shr.u32 %r396, %r883, 30;
|
|
shl.b32 %r397, %r882, 2;
|
|
add.s32 %r884, %r396, %r397;
|
|
shl.b32 %r54, %r883, 2;
|
|
shr.u32 %r398, %r884, 31;
|
|
shr.u32 %r399, %r882, 30;
|
|
add.s32 %r55, %r398, %r399;
|
|
setp.eq.s32 %p22, %r398, 0;
|
|
@%p22 bra BB0_32;
|
|
|
|
not.b32 %r400, %r884;
|
|
neg.s32 %r886, %r54;
|
|
setp.eq.s32 %p23, %r54, 0;
|
|
selp.u32 %r401, 1, 0, %p23;
|
|
add.s32 %r884, %r401, %r400;
|
|
xor.b32 %r885, %r45, -2147483648;
|
|
bra.uni BB0_34;
|
|
|
|
BB0_32:
|
|
mov.u32 %r885, %r45;
|
|
mov.u32 %r886, %r54;
|
|
|
|
BB0_34:
|
|
clz.b32 %r888, %r884;
|
|
setp.eq.s32 %p24, %r888, 0;
|
|
shl.b32 %r402, %r884, %r888;
|
|
mov.u32 %r403, 32;
|
|
sub.s32 %r404, %r403, %r888;
|
|
shr.u32 %r405, %r886, %r404;
|
|
add.s32 %r406, %r405, %r402;
|
|
selp.b32 %r63, %r884, %r406, %p24;
|
|
mov.u32 %r407, -921707870;
|
|
mul.hi.u32 %r887, %r63, %r407;
|
|
setp.eq.s32 %p25, %r45, 0;
|
|
neg.s32 %r408, %r55;
|
|
selp.b32 %r889, %r55, %r408, %p25;
|
|
setp.lt.s32 %p26, %r887, 1;
|
|
@%p26 bra BB0_36;
|
|
|
|
mul.lo.s32 %r409, %r63, -921707870;
|
|
shr.u32 %r410, %r409, 31;
|
|
shl.b32 %r411, %r887, 1;
|
|
add.s32 %r887, %r410, %r411;
|
|
add.s32 %r888, %r888, 1;
|
|
|
|
BB0_36:
|
|
mov.u32 %r412, 126;
|
|
sub.s32 %r413, %r412, %r888;
|
|
shl.b32 %r414, %r413, 23;
|
|
add.s32 %r415, %r887, 1;
|
|
shr.u32 %r416, %r415, 7;
|
|
add.s32 %r417, %r416, 1;
|
|
shr.u32 %r418, %r417, 1;
|
|
add.s32 %r419, %r418, %r414;
|
|
or.b32 %r420, %r419, %r885;
|
|
mov.b32 %f1325, %r420;
|
|
|
|
BB0_37:
|
|
mul.rn.f32 %f38, %f1325, %f1325;
|
|
and.b32 %r71, %r889, 1;
|
|
setp.eq.s32 %p27, %r71, 0;
|
|
@%p27 bra BB0_39;
|
|
|
|
mov.f32 %f477, 0fBAB6061A;
|
|
mov.f32 %f478, 0f37CCF5CE;
|
|
fma.rn.f32 %f1326, %f478, %f38, %f477;
|
|
bra.uni BB0_40;
|
|
|
|
BB0_39:
|
|
mov.f32 %f479, 0f3C08839E;
|
|
mov.f32 %f480, 0fB94CA1F9;
|
|
fma.rn.f32 %f1326, %f480, %f38, %f479;
|
|
|
|
BB0_40:
|
|
@%p27 bra BB0_42;
|
|
|
|
mov.f32 %f481, 0f3D2AAAA5;
|
|
fma.rn.f32 %f482, %f1326, %f38, %f481;
|
|
mov.f32 %f483, 0fBF000000;
|
|
fma.rn.f32 %f1327, %f482, %f38, %f483;
|
|
bra.uni BB0_43;
|
|
|
|
BB0_42:
|
|
mov.f32 %f484, 0fBE2AAAA3;
|
|
fma.rn.f32 %f485, %f1326, %f38, %f484;
|
|
mov.f32 %f486, 0f00000000;
|
|
fma.rn.f32 %f1327, %f485, %f38, %f486;
|
|
|
|
BB0_43:
|
|
fma.rn.f32 %f1328, %f1327, %f1325, %f1325;
|
|
@%p27 bra BB0_45;
|
|
|
|
mov.f32 %f487, 0f3F800000;
|
|
fma.rn.f32 %f1328, %f1327, %f38, %f487;
|
|
|
|
BB0_45:
|
|
and.b32 %r421, %r889, 2;
|
|
setp.eq.s32 %p30, %r421, 0;
|
|
@%p30 bra BB0_47;
|
|
|
|
mov.f32 %f488, 0f00000000;
|
|
mov.f32 %f489, 0fBF800000;
|
|
fma.rn.f32 %f1328, %f1328, %f489, %f488;
|
|
|
|
BB0_47:
|
|
mul.f32 %f490, %f13, %f1322;
|
|
mul.f32 %f491, %f490, %f490;
|
|
mov.f32 %f492, 0f3F800000;
|
|
sub.f32 %f493, %f492, %f491;
|
|
mul.f32 %f494, %f13, %f1328;
|
|
mul.f32 %f495, %f494, %f494;
|
|
sub.f32 %f496, %f493, %f495;
|
|
mov.f32 %f497, 0f00000000;
|
|
max.f32 %f498, %f497, %f496;
|
|
sqrt.rn.f32 %f499, %f498;
|
|
mul.f32 %f500, %f10, %f494;
|
|
mul.f32 %f501, %f11, %f494;
|
|
mul.f32 %f502, %f12, %f494;
|
|
mul.f32 %f503, %f8, %f11;
|
|
mul.f32 %f504, %f9, %f12;
|
|
sub.f32 %f505, %f503, %f504;
|
|
fma.rn.f32 %f506, %f505, %f490, %f500;
|
|
mul.f32 %f507, %f7, %f12;
|
|
mul.f32 %f508, %f8, %f10;
|
|
sub.f32 %f509, %f507, %f508;
|
|
fma.rn.f32 %f510, %f509, %f490, %f501;
|
|
mul.f32 %f511, %f9, %f10;
|
|
mul.f32 %f512, %f7, %f11;
|
|
sub.f32 %f513, %f511, %f512;
|
|
fma.rn.f32 %f514, %f513, %f490, %f502;
|
|
fma.rn.f32 %f62, %f7, %f499, %f506;
|
|
fma.rn.f32 %f63, %f9, %f499, %f510;
|
|
fma.rn.f32 %f64, %f8, %f499, %f514;
|
|
|
|
BB0_49:
|
|
add.u64 %rd113, %SP, 32;
|
|
cvta.to.local.u64 %rd114, %rd113;
|
|
add.s64 %rd14, %rd114, 12;
|
|
mov.u32 %r432, -1082130432;
|
|
st.local.u32 [%rd114+12], %r432;
|
|
mov.u32 %r433, 0;
|
|
st.local.u32 [%rd114], %r433;
|
|
st.local.u32 [%rd114+4], %r433;
|
|
st.local.u32 [%rd114+8], %r433;
|
|
st.local.u32 [%rd114+16], %r433;
|
|
st.local.u32 [%rd114+20], %r433;
|
|
st.local.u32 [%rd114+24], %r433;
|
|
ld.global.v2.u32 {%r434, %r435}, [pixelID];
|
|
cvt.u64.u32 %rd109, %r434;
|
|
cvt.u64.u32 %rd110, %r435;
|
|
mov.u64 %rd115, GBufferAtten;
|
|
cvta.global.u64 %rd108, %rd115;
|
|
mov.u32 %r431, 12;
|
|
// inline asm
|
|
call (%rd107), _rt_buffer_get_64, (%rd108, %r295, %r431, %rd109, %rd110, %rd293, %rd293);
|
|
// inline asm
|
|
ld.f32 %f70, [%rd107+8];
|
|
ld.f32 %f68, [%rd107+4];
|
|
ld.f32 %f69, [%rd107];
|
|
add.f32 %f554, %f69, %f68;
|
|
add.f32 %f555, %f70, %f554;
|
|
ld.global.u32 %r438, [bounceFlag];
|
|
setp.ne.s32 %p31, %r438, 0;
|
|
setp.eq.f32 %p32, %f555, 0f00000000;
|
|
and.pred %p33, %p31, %p32;
|
|
@%p33 bra BB0_51;
|
|
|
|
ld.global.u32 %r439, [root];
|
|
mov.u32 %r440, 0;
|
|
mov.f32 %f562, 0f00000000;
|
|
mov.f32 %f563, 0f6C4ECB8F;
|
|
mov.u32 %r441, 28;
|
|
// inline asm
|
|
call _rt_trace_64, (%r439, %f1333, %f1334, %f1335, %f62, %f63, %f64, %r440, %f562, %f563, %rd113, %r441);
|
|
// inline asm
|
|
|
|
BB0_51:
|
|
ld.local.f32 %f71, [%rd14];
|
|
setp.ltu.f32 %p34, %f71, 0f00000000;
|
|
@%p34 bra BB0_235;
|
|
bra.uni BB0_52;
|
|
|
|
BB0_235:
|
|
ld.global.u32 %r795, [bounceFlag];
|
|
setp.eq.s32 %p160, %r795, 0;
|
|
selp.u16 %rs10, 1, 0, %p160;
|
|
mov.f32 %f1418, 0f3F800000;
|
|
mov.f32 %f1409, 0f00000000;
|
|
mov.f32 %f1410, %f1409;
|
|
mov.f32 %f1411, %f1409;
|
|
mov.f32 %f1412, %f1409;
|
|
mov.f32 %f1413, %f1409;
|
|
mov.f32 %f1414, %f1409;
|
|
mov.f32 %f1394, %f1409;
|
|
mov.f32 %f1395, %f1409;
|
|
mov.f32 %f1396, %f1409;
|
|
bra.uni BB0_236;
|
|
|
|
BB0_52:
|
|
fma.rn.f32 %f564, %f62, %f71, %f1333;
|
|
fma.rn.f32 %f565, %f63, %f71, %f1334;
|
|
fma.rn.f32 %f566, %f64, %f71, %f1335;
|
|
mul.f32 %f567, %f564, 0f3456BF95;
|
|
mul.f32 %f568, %f565, 0f3456BF95;
|
|
mul.f32 %f569, %f566, 0f3456BF95;
|
|
abs.f32 %f570, %f567;
|
|
abs.f32 %f571, %f568;
|
|
abs.f32 %f572, %f569;
|
|
max.f32 %f573, %f570, %f571;
|
|
max.f32 %f574, %f573, %f572;
|
|
mov.f32 %f575, 0f38D1B717;
|
|
max.f32 %f576, %f574, %f575;
|
|
ld.local.f32 %f577, [%rd14+-12];
|
|
ld.local.f32 %f578, [%rd14+-8];
|
|
ld.local.f32 %f579, [%rd14+-4];
|
|
fma.rn.f32 %f72, %f576, %f577, %f564;
|
|
fma.rn.f32 %f73, %f576, %f578, %f565;
|
|
fma.rn.f32 %f74, %f576, %f579, %f566;
|
|
ld.global.v2.u32 {%r446, %r447}, [pixelID];
|
|
cvt.u64.u32 %rd119, %r446;
|
|
cvt.u64.u32 %rd120, %r447;
|
|
mov.u64 %rd129, GBufferPos;
|
|
cvta.global.u64 %rd118, %rd129;
|
|
// inline asm
|
|
call (%rd117), _rt_buffer_get_64, (%rd118, %r295, %r431, %rd119, %rd120, %rd293, %rd293);
|
|
// inline asm
|
|
st.f32 [%rd117+8], %f74;
|
|
st.f32 [%rd117+4], %f73;
|
|
st.f32 [%rd117], %f72;
|
|
ld.global.v2.u32 {%r450, %r451}, [pixelID];
|
|
cvt.u64.u32 %rd125, %r450;
|
|
cvt.u64.u32 %rd126, %r451;
|
|
mov.u64 %rd130, GBufferNormal;
|
|
cvta.global.u64 %rd124, %rd130;
|
|
// inline asm
|
|
call (%rd123), _rt_buffer_get_64, (%rd124, %r295, %r431, %rd125, %rd126, %rd293, %rd293);
|
|
// inline asm
|
|
ld.local.f32 %f580, [%rd14+-12];
|
|
ld.local.f32 %f581, [%rd14+-8];
|
|
ld.local.f32 %f582, [%rd14+-4];
|
|
st.f32 [%rd123+8], %f582;
|
|
st.f32 [%rd123+4], %f581;
|
|
st.f32 [%rd123], %f580;
|
|
ld.local.f32 %f583, [%rd14+12];
|
|
ld.local.f32 %f584, [%rd14+8];
|
|
ld.local.f32 %f585, [%rd14+4];
|
|
setp.lt.f32 %p35, %f585, 0f00000000;
|
|
neg.f32 %f586, %f585;
|
|
selp.f32 %f587, %f586, %f585, %p35;
|
|
ld.global.f32 %f588, [backFaceWeight];
|
|
selp.f32 %f589, %f588, 0f3F800000, %p35;
|
|
cvt.rmi.f32.f32 %f590, %f587;
|
|
cvt.rmi.f32.f32 %f591, %f584;
|
|
cvt.rmi.f32.f32 %f592, %f583;
|
|
sub.f32 %f593, %f587, %f590;
|
|
sub.f32 %f594, %f584, %f591;
|
|
sub.f32 %f595, %f583, %f592;
|
|
mul.f32 %f1412, %f589, %f593;
|
|
mul.f32 %f1413, %f589, %f594;
|
|
mul.f32 %f1414, %f589, %f595;
|
|
mul.f32 %f1409, %f590, 0f3B808081;
|
|
mul.f32 %f1410, %f591, 0f3B808081;
|
|
mul.f32 %f1411, %f592, 0f3B808081;
|
|
ld.global.u32 %r454, [bounceFlag];
|
|
setp.lt.s32 %p36, %r454, 1;
|
|
@%p36 bra BB0_54;
|
|
|
|
mul.f32 %f1412, %f69, %f1412;
|
|
mul.f32 %f1413, %f68, %f1413;
|
|
mul.f32 %f1414, %f70, %f1414;
|
|
mul.f32 %f1409, %f69, %f1409;
|
|
mul.f32 %f1410, %f68, %f1410;
|
|
mul.f32 %f1411, %f70, %f1411;
|
|
|
|
BB0_54:
|
|
mov.u64 %rd136, globalLights;
|
|
cvta.global.u64 %rd135, %rd136;
|
|
mov.u32 %r455, 1;
|
|
mov.u32 %r456, 48;
|
|
// inline asm
|
|
call (%rd131, %rd132, %rd133, %rd134), _rt_buffer_get_size_64, (%rd135, %r455, %r456);
|
|
// inline asm
|
|
cvt.u32.u64 %r72, %rd131;
|
|
setp.eq.s32 %p37, %r72, 0;
|
|
mov.f32 %f1394, 0f00000000;
|
|
mov.f32 %f1395, %f1394;
|
|
mov.f32 %f1396, %f1394;
|
|
@%p37 bra BB0_162;
|
|
|
|
// inline asm
|
|
call (%rd137, %rd138, %rd139, %rd140), _rt_buffer_get_size_64, (%rd135, %r455, %r456);
|
|
// inline asm
|
|
cvt.u32.u64 %r461, %rd137;
|
|
cvt.rn.f32.u32 %f599, %r461;
|
|
mul.f32 %f600, %f3, %f599;
|
|
cvt.rzi.u32.f32 %r462, %f600;
|
|
add.s32 %r463, %r461, -1;
|
|
min.u32 %r464, %r462, %r463;
|
|
cvt.u64.u32 %rd144, %r464;
|
|
// inline asm
|
|
call (%rd142), _rt_buffer_get_64, (%rd135, %r455, %r456, %rd144, %rd293, %rd293, %rd293);
|
|
// inline asm
|
|
ld.v4.f32 {%f601, %f602, %f603, %f604}, [%rd142+32];
|
|
ld.v4.f32 {%f605, %f606, %f607, %f608}, [%rd142+16];
|
|
ld.v4.f32 {%f609, %f610, %f611, %f612}, [%rd142];
|
|
setp.lt.f32 %p38, %f609, 0f3F800000;
|
|
@%p38 bra BB0_109;
|
|
bra.uni BB0_56;
|
|
|
|
BB0_109:
|
|
fma.rn.f32 %f732, %f610, 0f40000000, 0fBF800000;
|
|
fma.rn.f32 %f733, %f611, 0f40000000, 0fBF800000;
|
|
abs.f32 %f734, %f732;
|
|
mov.f32 %f1376, 0f3F800000;
|
|
sub.f32 %f736, %f1376, %f734;
|
|
abs.f32 %f737, %f733;
|
|
sub.f32 %f738, %f736, %f737;
|
|
neg.f32 %f739, %f738;
|
|
cvt.sat.f32.f32 %f740, %f739;
|
|
setp.ltu.f32 %p71, %f732, 0f00000000;
|
|
neg.f32 %f741, %f740;
|
|
selp.f32 %f742, %f740, %f741, %p71;
|
|
add.f32 %f743, %f732, %f742;
|
|
setp.ltu.f32 %p72, %f733, 0f00000000;
|
|
selp.f32 %f744, %f740, %f741, %p72;
|
|
add.f32 %f745, %f733, %f744;
|
|
mul.f32 %f746, %f745, %f745;
|
|
fma.rn.f32 %f747, %f743, %f743, %f746;
|
|
fma.rn.f32 %f748, %f738, %f738, %f747;
|
|
sqrt.rn.f32 %f749, %f748;
|
|
rcp.rn.f32 %f750, %f749;
|
|
mul.f32 %f180, %f750, %f743;
|
|
mul.f32 %f181, %f750, %f745;
|
|
mul.f32 %f182, %f738, %f750;
|
|
ld.local.f32 %f751, [%rd14+-12];
|
|
mul.f32 %f752, %f751, %f180;
|
|
ld.local.f32 %f753, [%rd14+-8];
|
|
mul.f32 %f754, %f181, %f753;
|
|
neg.f32 %f755, %f754;
|
|
sub.f32 %f756, %f755, %f752;
|
|
ld.local.f32 %f757, [%rd14+-4];
|
|
mul.f32 %f758, %f182, %f757;
|
|
sub.f32 %f1363, %f756, %f758;
|
|
setp.gt.f32 %p73, %f1363, 0f00000000;
|
|
setp.neu.f32 %p74, %f612, 0f00000000;
|
|
and.pred %p75, %p73, %p74;
|
|
@!%p75 bra BB0_111;
|
|
bra.uni BB0_110;
|
|
|
|
BB0_110:
|
|
abs.f32 %f767, %f180;
|
|
abs.f32 %f768, %f182;
|
|
setp.gt.f32 %p76, %f767, %f768;
|
|
neg.f32 %f769, %f181;
|
|
selp.f32 %f770, %f769, 0f00000000, %p76;
|
|
neg.f32 %f771, %f182;
|
|
selp.f32 %f772, %f180, %f771, %p76;
|
|
selp.f32 %f773, 0f00000000, %f181, %p76;
|
|
mul.f32 %f774, %f772, %f772;
|
|
fma.rn.f32 %f775, %f770, %f770, %f774;
|
|
fma.rn.f32 %f776, %f773, %f773, %f775;
|
|
sqrt.rn.f32 %f777, %f776;
|
|
rcp.rn.f32 %f778, %f777;
|
|
mul.f32 %f779, %f770, %f778;
|
|
mul.f32 %f780, %f772, %f778;
|
|
mul.f32 %f781, %f773, %f778;
|
|
mul.f32 %f782, %f182, %f780;
|
|
mul.f32 %f783, %f181, %f781;
|
|
sub.f32 %f784, %f782, %f783;
|
|
mul.f32 %f785, %f180, %f781;
|
|
mul.f32 %f786, %f182, %f779;
|
|
sub.f32 %f787, %f785, %f786;
|
|
mul.f32 %f788, %f181, %f779;
|
|
mul.f32 %f789, %f180, %f780;
|
|
sub.f32 %f790, %f788, %f789;
|
|
mul.f32 %f791, %f73, %f787;
|
|
fma.rn.f32 %f792, %f72, %f784, %f791;
|
|
fma.rn.f32 %f793, %f74, %f790, %f792;
|
|
fma.rn.f32 %f763, %f601, %f793, %f603;
|
|
mul.f32 %f794, %f73, %f780;
|
|
fma.rn.f32 %f795, %f72, %f779, %f794;
|
|
fma.rn.f32 %f796, %f74, %f781, %f795;
|
|
fma.rn.f32 %f764, %f602, %f796, %f604;
|
|
cvt.rzi.s32.f32 %r566, %f612;
|
|
mov.f32 %f766, 0f00000000;
|
|
// inline asm
|
|
call (%f759, %f760, %f761, %f762), _rt_texture_get_f_id, (%r566, %r295, %f763, %f764, %f766, %f766);
|
|
// inline asm
|
|
mul.f32 %f1363, %f1363, %f759;
|
|
|
|
BB0_111:
|
|
mov.f32 %f1377, 0f00000000;
|
|
setp.le.f32 %p77, %f1363, 0f00000000;
|
|
mov.f32 %f1378, %f1377;
|
|
mov.f32 %f1379, %f1377;
|
|
@%p77 bra BB0_161;
|
|
|
|
abs.f32 %f801, %f180;
|
|
abs.f32 %f802, %f182;
|
|
setp.gt.f32 %p78, %f801, %f802;
|
|
neg.f32 %f803, %f181;
|
|
selp.f32 %f804, %f803, 0f00000000, %p78;
|
|
neg.f32 %f805, %f182;
|
|
selp.f32 %f806, %f180, %f805, %p78;
|
|
selp.f32 %f807, 0f00000000, %f181, %p78;
|
|
mul.f32 %f808, %f806, %f806;
|
|
fma.rn.f32 %f809, %f804, %f804, %f808;
|
|
fma.rn.f32 %f810, %f807, %f807, %f809;
|
|
sqrt.rn.f32 %f811, %f810;
|
|
rcp.rn.f32 %f812, %f811;
|
|
mul.f32 %f186, %f804, %f812;
|
|
mul.f32 %f187, %f806, %f812;
|
|
mul.f32 %f188, %f807, %f812;
|
|
setp.ltu.f32 %p79, %f605, 0f00000000;
|
|
@%p79 bra BB0_160;
|
|
|
|
sqrt.rn.f32 %f189, %f1;
|
|
add.u64 %rd171, %SP, 0;
|
|
cvta.to.local.u64 %rd29, %rd171;
|
|
mul.f32 %f1370, %f2, 0f40C90FDB;
|
|
abs.f32 %f191, %f1370;
|
|
setp.neu.f32 %p80, %f191, 0f7F800000;
|
|
mov.f32 %f1364, %f1370;
|
|
@%p80 bra BB0_115;
|
|
|
|
mov.f32 %f813, 0f00000000;
|
|
mul.rn.f32 %f1364, %f1370, %f813;
|
|
|
|
BB0_115:
|
|
mul.f32 %f814, %f1364, 0f3F22F983;
|
|
cvt.rni.s32.f32 %r919, %f814;
|
|
cvt.rn.f32.s32 %f815, %r919;
|
|
neg.f32 %f816, %f815;
|
|
mov.f32 %f817, 0f3FC90FDA;
|
|
fma.rn.f32 %f818, %f816, %f817, %f1364;
|
|
mov.f32 %f819, 0f33A22168;
|
|
fma.rn.f32 %f820, %f816, %f819, %f818;
|
|
mov.f32 %f821, 0f27C234C5;
|
|
fma.rn.f32 %f1365, %f816, %f821, %f820;
|
|
abs.f32 %f822, %f1364;
|
|
add.s64 %rd30, %rd29, 24;
|
|
setp.leu.f32 %p81, %f822, 0f47CE4780;
|
|
@%p81 bra BB0_126;
|
|
|
|
mov.b32 %r144, %f1364;
|
|
shr.u32 %r145, %r144, 23;
|
|
shl.b32 %r570, %r144, 8;
|
|
or.b32 %r146, %r570, -2147483648;
|
|
mov.u32 %r910, 0;
|
|
mov.u64 %rd295, 0;
|
|
mov.u64 %rd294, %rd29;
|
|
mov.u32 %r911, %r910;
|
|
|
|
BB0_117:
|
|
.pragma "nounroll";
|
|
shl.b64 %rd173, %rd295, 2;
|
|
mov.u64 %rd174, __cudart_i2opi_f;
|
|
add.s64 %rd175, %rd174, %rd173;
|
|
ld.const.u32 %r573, [%rd175];
|
|
// inline asm
|
|
{
|
|
mad.lo.cc.u32 %r571, %r573, %r146, %r911;
|
|
madc.hi.u32 %r911, %r573, %r146, 0;
|
|
}
|
|
// inline asm
|
|
st.local.u32 [%rd294], %r571;
|
|
add.s32 %r910, %r910, 1;
|
|
cvt.s64.s32 %rd295, %r910;
|
|
mul.wide.s32 %rd176, %r910, 4;
|
|
add.s64 %rd294, %rd29, %rd176;
|
|
setp.ne.s32 %p82, %r910, 6;
|
|
@%p82 bra BB0_117;
|
|
|
|
and.b32 %r576, %r145, 255;
|
|
add.s32 %r577, %r576, -128;
|
|
shr.u32 %r578, %r577, 5;
|
|
and.b32 %r151, %r144, -2147483648;
|
|
st.local.u32 [%rd30], %r911;
|
|
mov.u32 %r579, 6;
|
|
sub.s32 %r580, %r579, %r578;
|
|
mul.wide.s32 %rd177, %r580, 4;
|
|
add.s64 %rd36, %rd29, %rd177;
|
|
ld.local.u32 %r912, [%rd36];
|
|
ld.local.u32 %r913, [%rd36+-4];
|
|
and.b32 %r154, %r145, 31;
|
|
setp.eq.s32 %p83, %r154, 0;
|
|
@%p83 bra BB0_120;
|
|
|
|
mov.u32 %r581, 32;
|
|
sub.s32 %r582, %r581, %r154;
|
|
shr.u32 %r583, %r913, %r582;
|
|
shl.b32 %r584, %r912, %r154;
|
|
add.s32 %r912, %r583, %r584;
|
|
ld.local.u32 %r585, [%rd36+-8];
|
|
shr.u32 %r586, %r585, %r582;
|
|
shl.b32 %r587, %r913, %r154;
|
|
add.s32 %r913, %r586, %r587;
|
|
|
|
BB0_120:
|
|
shr.u32 %r588, %r913, 30;
|
|
shl.b32 %r589, %r912, 2;
|
|
add.s32 %r914, %r588, %r589;
|
|
shl.b32 %r160, %r913, 2;
|
|
shr.u32 %r590, %r914, 31;
|
|
shr.u32 %r591, %r912, 30;
|
|
add.s32 %r161, %r590, %r591;
|
|
setp.eq.s32 %p84, %r590, 0;
|
|
@%p84 bra BB0_121;
|
|
|
|
not.b32 %r592, %r914;
|
|
neg.s32 %r916, %r160;
|
|
setp.eq.s32 %p85, %r160, 0;
|
|
selp.u32 %r593, 1, 0, %p85;
|
|
add.s32 %r914, %r593, %r592;
|
|
xor.b32 %r915, %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 %f613, %f112;
|
|
abs.f32 %f614, %f113;
|
|
setp.gt.f32 %p39, %f613, %f614;
|
|
neg.f32 %f615, %f114;
|
|
neg.f32 %f616, %f113;
|
|
selp.f32 %f617, %f615, 0f00000000, %p39;
|
|
selp.f32 %f618, %f112, %f616, %p39;
|
|
selp.f32 %f619, 0f00000000, %f114, %p39;
|
|
mul.f32 %f620, %f618, %f618;
|
|
fma.rn.f32 %f621, %f617, %f617, %f620;
|
|
fma.rn.f32 %f622, %f619, %f619, %f621;
|
|
sqrt.rn.f32 %f623, %f622;
|
|
rcp.rn.f32 %f624, %f623;
|
|
mul.f32 %f115, %f617, %f624;
|
|
mul.f32 %f116, %f618, %f624;
|
|
mul.f32 %f117, %f619, %f624;
|
|
sqrt.rn.f32 %f118, %f1;
|
|
mul.f32 %f1348, %f2, 0f40C90FDB;
|
|
abs.f32 %f120, %f1348;
|
|
setp.neu.f32 %p40, %f120, 0f7F800000;
|
|
mov.f32 %f1342, %f1348;
|
|
@%p40 bra BB0_58;
|
|
|
|
mov.f32 %f625, 0f00000000;
|
|
mul.rn.f32 %f1342, %f1348, %f625;
|
|
|
|
BB0_58:
|
|
mul.f32 %f626, %f1342, 0f3F22F983;
|
|
cvt.rni.s32.f32 %r899, %f626;
|
|
cvt.rn.f32.s32 %f627, %r899;
|
|
neg.f32 %f628, %f627;
|
|
mov.f32 %f629, 0f3FC90FDA;
|
|
fma.rn.f32 %f630, %f628, %f629, %f1342;
|
|
mov.f32 %f631, 0f33A22168;
|
|
fma.rn.f32 %f632, %f628, %f631, %f630;
|
|
mov.f32 %f633, 0f27C234C5;
|
|
fma.rn.f32 %f1343, %f628, %f633, %f632;
|
|
abs.f32 %f634, %f1342;
|
|
setp.leu.f32 %p41, %f634, 0f47CE4780;
|
|
@%p41 bra BB0_69;
|
|
|
|
mov.b32 %r74, %f1342;
|
|
shl.b32 %r467, %r74, 8;
|
|
or.b32 %r75, %r467, -2147483648;
|
|
add.u64 %rd150, %SP, 0;
|
|
cvta.to.local.u64 %rd15, %rd150;
|
|
mov.u32 %r890, 0;
|
|
mov.u64 %rd291, 0;
|
|
mov.u64 %rd290, %rd15;
|
|
mov.u32 %r891, %r890;
|
|
|
|
BB0_60:
|
|
.pragma "nounroll";
|
|
shl.b64 %rd151, %rd291, 2;
|
|
mov.u64 %rd152, __cudart_i2opi_f;
|
|
add.s64 %rd153, %rd152, %rd151;
|
|
ld.const.u32 %r470, [%rd153];
|
|
// inline asm
|
|
{
|
|
mad.lo.cc.u32 %r468, %r470, %r75, %r891;
|
|
madc.hi.u32 %r891, %r470, %r75, 0;
|
|
}
|
|
// inline asm
|
|
st.local.u32 [%rd290], %r468;
|
|
add.s32 %r890, %r890, 1;
|
|
cvt.s64.s32 %rd291, %r890;
|
|
mul.wide.s32 %rd156, %r890, 4;
|
|
add.s64 %rd290, %rd15, %rd156;
|
|
setp.ne.s32 %p42, %r890, 6;
|
|
@%p42 bra BB0_60;
|
|
|
|
bfe.u32 %r473, %r74, 23, 8;
|
|
add.s32 %r474, %r473, -128;
|
|
shr.u32 %r475, %r474, 5;
|
|
and.b32 %r80, %r74, -2147483648;
|
|
cvta.to.local.u64 %rd158, %rd150;
|
|
st.local.u32 [%rd158+24], %r891;
|
|
bfe.u32 %r81, %r74, 23, 5;
|
|
mov.u32 %r476, 6;
|
|
sub.s32 %r477, %r476, %r475;
|
|
mul.wide.s32 %rd159, %r477, 4;
|
|
add.s64 %rd20, %rd158, %rd159;
|
|
ld.local.u32 %r892, [%rd20];
|
|
ld.local.u32 %r893, [%rd20+-4];
|
|
setp.eq.s32 %p43, %r81, 0;
|
|
@%p43 bra BB0_63;
|
|
|
|
mov.u32 %r478, 32;
|
|
sub.s32 %r479, %r478, %r81;
|
|
shr.u32 %r480, %r893, %r479;
|
|
shl.b32 %r481, %r892, %r81;
|
|
add.s32 %r892, %r480, %r481;
|
|
ld.local.u32 %r482, [%rd20+-8];
|
|
shr.u32 %r483, %r482, %r479;
|
|
shl.b32 %r484, %r893, %r81;
|
|
add.s32 %r893, %r483, %r484;
|
|
|
|
BB0_63:
|
|
shr.u32 %r485, %r893, 30;
|
|
shl.b32 %r486, %r892, 2;
|
|
add.s32 %r894, %r485, %r486;
|
|
shl.b32 %r89, %r893, 2;
|
|
shr.u32 %r487, %r894, 31;
|
|
shr.u32 %r488, %r892, 30;
|
|
add.s32 %r90, %r487, %r488;
|
|
setp.eq.s32 %p44, %r487, 0;
|
|
@%p44 bra BB0_64;
|
|
|
|
not.b32 %r489, %r894;
|
|
neg.s32 %r896, %r89;
|
|
setp.eq.s32 %p45, %r89, 0;
|
|
selp.u32 %r490, 1, 0, %p45;
|
|
add.s32 %r894, %r490, %r489;
|
|
xor.b32 %r895, %r80, -2147483648;
|
|
bra.uni BB0_66;
|
|
|
|
BB0_64:
|
|
mov.u32 %r895, %r80;
|
|
mov.u32 %r896, %r89;
|
|
|
|
BB0_66:
|
|
clz.b32 %r898, %r894;
|
|
setp.eq.s32 %p46, %r898, 0;
|
|
shl.b32 %r491, %r894, %r898;
|
|
mov.u32 %r492, 32;
|
|
sub.s32 %r493, %r492, %r898;
|
|
shr.u32 %r494, %r896, %r493;
|
|
add.s32 %r495, %r494, %r491;
|
|
selp.b32 %r98, %r894, %r495, %p46;
|
|
mov.u32 %r496, -921707870;
|
|
mul.hi.u32 %r897, %r98, %r496;
|
|
setp.eq.s32 %p47, %r80, 0;
|
|
neg.s32 %r497, %r90;
|
|
selp.b32 %r899, %r90, %r497, %p47;
|
|
setp.lt.s32 %p48, %r897, 1;
|
|
@%p48 bra BB0_68;
|
|
|
|
mul.lo.s32 %r498, %r98, -921707870;
|
|
shr.u32 %r499, %r498, 31;
|
|
shl.b32 %r500, %r897, 1;
|
|
add.s32 %r897, %r499, %r500;
|
|
add.s32 %r898, %r898, 1;
|
|
|
|
BB0_68:
|
|
mov.u32 %r501, 126;
|
|
sub.s32 %r502, %r501, %r898;
|
|
shl.b32 %r503, %r502, 23;
|
|
add.s32 %r504, %r897, 1;
|
|
shr.u32 %r505, %r504, 7;
|
|
add.s32 %r506, %r505, 1;
|
|
shr.u32 %r507, %r506, 1;
|
|
add.s32 %r508, %r507, %r503;
|
|
or.b32 %r509, %r508, %r895;
|
|
mov.b32 %f1343, %r509;
|
|
|
|
BB0_69:
|
|
mul.rn.f32 %f126, %f1343, %f1343;
|
|
add.s32 %r106, %r899, 1;
|
|
and.b32 %r107, %r106, 1;
|
|
setp.eq.s32 %p49, %r107, 0;
|
|
@%p49 bra BB0_71;
|
|
|
|
mov.f32 %f635, 0fBAB6061A;
|
|
mov.f32 %f636, 0f37CCF5CE;
|
|
fma.rn.f32 %f1344, %f636, %f126, %f635;
|
|
bra.uni BB0_72;
|
|
|
|
BB0_71:
|
|
mov.f32 %f637, 0f3C08839E;
|
|
mov.f32 %f638, 0fB94CA1F9;
|
|
fma.rn.f32 %f1344, %f638, %f126, %f637;
|
|
|
|
BB0_72:
|
|
@%p49 bra BB0_74;
|
|
|
|
mov.f32 %f639, 0f3D2AAAA5;
|
|
fma.rn.f32 %f640, %f1344, %f126, %f639;
|
|
mov.f32 %f641, 0fBF000000;
|
|
fma.rn.f32 %f1345, %f640, %f126, %f641;
|
|
bra.uni BB0_75;
|
|
|
|
BB0_74:
|
|
mov.f32 %f642, 0fBE2AAAA3;
|
|
fma.rn.f32 %f643, %f1344, %f126, %f642;
|
|
mov.f32 %f644, 0f00000000;
|
|
fma.rn.f32 %f1345, %f643, %f126, %f644;
|
|
|
|
BB0_75:
|
|
fma.rn.f32 %f1346, %f1345, %f1343, %f1343;
|
|
@%p49 bra BB0_77;
|
|
|
|
mov.f32 %f645, 0f3F800000;
|
|
fma.rn.f32 %f1346, %f1345, %f126, %f645;
|
|
|
|
BB0_77:
|
|
and.b32 %r510, %r106, 2;
|
|
setp.eq.s32 %p52, %r510, 0;
|
|
@%p52 bra BB0_79;
|
|
|
|
mov.f32 %f646, 0f00000000;
|
|
mov.f32 %f647, 0fBF800000;
|
|
fma.rn.f32 %f1346, %f1346, %f647, %f646;
|
|
|
|
BB0_79:
|
|
@%p40 bra BB0_81;
|
|
|
|
mov.f32 %f648, 0f00000000;
|
|
mul.rn.f32 %f1348, %f1348, %f648;
|
|
|
|
BB0_81:
|
|
mul.f32 %f649, %f1348, 0f3F22F983;
|
|
cvt.rni.s32.f32 %r909, %f649;
|
|
cvt.rn.f32.s32 %f650, %r909;
|
|
neg.f32 %f651, %f650;
|
|
fma.rn.f32 %f653, %f651, %f629, %f1348;
|
|
fma.rn.f32 %f655, %f651, %f631, %f653;
|
|
fma.rn.f32 %f1349, %f651, %f633, %f655;
|
|
abs.f32 %f657, %f1348;
|
|
setp.leu.f32 %p54, %f657, 0f47CE4780;
|
|
@%p54 bra BB0_92;
|
|
|
|
mov.b32 %r109, %f1348;
|
|
shr.u32 %r110, %r109, 23;
|
|
shl.b32 %r513, %r109, 8;
|
|
or.b32 %r111, %r513, -2147483648;
|
|
add.u64 %rd161, %SP, 0;
|
|
cvta.to.local.u64 %rd21, %rd161;
|
|
mov.u64 %rd292, %rd21;
|
|
mov.u32 %r900, %r433;
|
|
mov.u32 %r901, %r433;
|
|
|
|
BB0_83:
|
|
.pragma "nounroll";
|
|
shl.b64 %rd162, %rd293, 2;
|
|
mov.u64 %rd163, __cudart_i2opi_f;
|
|
add.s64 %rd164, %rd163, %rd162;
|
|
ld.const.u32 %r516, [%rd164];
|
|
// inline asm
|
|
{
|
|
mad.lo.cc.u32 %r514, %r516, %r111, %r901;
|
|
madc.hi.u32 %r901, %r516, %r111, 0;
|
|
}
|
|
// inline asm
|
|
st.local.u32 [%rd292], %r514;
|
|
add.s32 %r900, %r900, 1;
|
|
cvt.s64.s32 %rd293, %r900;
|
|
mul.wide.s32 %rd165, %r900, 4;
|
|
add.s64 %rd292, %rd21, %rd165;
|
|
setp.ne.s32 %p55, %r900, 6;
|
|
@%p55 bra BB0_83;
|
|
|
|
and.b32 %r519, %r110, 255;
|
|
add.s32 %r520, %r519, -128;
|
|
shr.u32 %r521, %r520, 5;
|
|
and.b32 %r116, %r109, -2147483648;
|
|
cvta.to.local.u64 %rd167, %rd161;
|
|
st.local.u32 [%rd167+24], %r901;
|
|
mov.u32 %r522, 6;
|
|
sub.s32 %r523, %r522, %r521;
|
|
mul.wide.s32 %rd168, %r523, 4;
|
|
add.s64 %rd27, %rd167, %rd168;
|
|
ld.local.u32 %r902, [%rd27];
|
|
ld.local.u32 %r903, [%rd27+-4];
|
|
and.b32 %r119, %r110, 31;
|
|
setp.eq.s32 %p56, %r119, 0;
|
|
@%p56 bra BB0_86;
|
|
|
|
mov.u32 %r524, 32;
|
|
sub.s32 %r525, %r524, %r119;
|
|
shr.u32 %r526, %r903, %r525;
|
|
shl.b32 %r527, %r902, %r119;
|
|
add.s32 %r902, %r526, %r527;
|
|
ld.local.u32 %r528, [%rd27+-8];
|
|
shr.u32 %r529, %r528, %r525;
|
|
shl.b32 %r530, %r903, %r119;
|
|
add.s32 %r903, %r529, %r530;
|
|
|
|
BB0_86:
|
|
shr.u32 %r531, %r903, 30;
|
|
shl.b32 %r532, %r902, 2;
|
|
add.s32 %r904, %r531, %r532;
|
|
shl.b32 %r125, %r903, 2;
|
|
shr.u32 %r533, %r904, 31;
|
|
shr.u32 %r534, %r902, 30;
|
|
add.s32 %r126, %r533, %r534;
|
|
setp.eq.s32 %p57, %r533, 0;
|
|
@%p57 bra BB0_87;
|
|
|
|
not.b32 %r535, %r904;
|
|
neg.s32 %r906, %r125;
|
|
setp.eq.s32 %p58, %r125, 0;
|
|
selp.u32 %r536, 1, 0, %p58;
|
|
add.s32 %r904, %r536, %r535;
|
|
xor.b32 %r905, %r116, -2147483648;
|
|
bra.uni BB0_89;
|
|
|
|
BB0_87:
|
|
mov.u32 %r905, %r116;
|
|
mov.u32 %r906, %r125;
|
|
|
|
BB0_89:
|
|
clz.b32 %r908, %r904;
|
|
setp.eq.s32 %p59, %r908, 0;
|
|
shl.b32 %r537, %r904, %r908;
|
|
mov.u32 %r538, 32;
|
|
sub.s32 %r539, %r538, %r908;
|
|
shr.u32 %r540, %r906, %r539;
|
|
add.s32 %r541, %r540, %r537;
|
|
selp.b32 %r134, %r904, %r541, %p59;
|
|
mov.u32 %r542, -921707870;
|
|
mul.hi.u32 %r907, %r134, %r542;
|
|
setp.eq.s32 %p60, %r116, 0;
|
|
neg.s32 %r543, %r126;
|
|
selp.b32 %r909, %r126, %r543, %p60;
|
|
setp.lt.s32 %p61, %r907, 1;
|
|
@%p61 bra BB0_91;
|
|
|
|
mul.lo.s32 %r544, %r134, -921707870;
|
|
shr.u32 %r545, %r544, 31;
|
|
shl.b32 %r546, %r907, 1;
|
|
add.s32 %r907, %r545, %r546;
|
|
add.s32 %r908, %r908, 1;
|
|
|
|
BB0_91:
|
|
mov.u32 %r547, 126;
|
|
sub.s32 %r548, %r547, %r908;
|
|
shl.b32 %r549, %r548, 23;
|
|
add.s32 %r550, %r907, 1;
|
|
shr.u32 %r551, %r550, 7;
|
|
add.s32 %r552, %r551, 1;
|
|
shr.u32 %r553, %r552, 1;
|
|
add.s32 %r554, %r553, %r549;
|
|
or.b32 %r555, %r554, %r905;
|
|
mov.b32 %f1349, %r555;
|
|
|
|
BB0_92:
|
|
mul.rn.f32 %f143, %f1349, %f1349;
|
|
and.b32 %r142, %r909, 1;
|
|
setp.eq.s32 %p62, %r142, 0;
|
|
@%p62 bra BB0_94;
|
|
|
|
mov.f32 %f658, 0fBAB6061A;
|
|
mov.f32 %f659, 0f37CCF5CE;
|
|
fma.rn.f32 %f1350, %f659, %f143, %f658;
|
|
bra.uni BB0_95;
|
|
|
|
BB0_94:
|
|
mov.f32 %f660, 0f3C08839E;
|
|
mov.f32 %f661, 0fB94CA1F9;
|
|
fma.rn.f32 %f1350, %f661, %f143, %f660;
|
|
|
|
BB0_95:
|
|
@%p62 bra BB0_97;
|
|
|
|
mov.f32 %f662, 0f3D2AAAA5;
|
|
fma.rn.f32 %f663, %f1350, %f143, %f662;
|
|
mov.f32 %f664, 0fBF000000;
|
|
fma.rn.f32 %f1351, %f663, %f143, %f664;
|
|
bra.uni BB0_98;
|
|
|
|
BB0_97:
|
|
mov.f32 %f665, 0fBE2AAAA3;
|
|
fma.rn.f32 %f666, %f1350, %f143, %f665;
|
|
mov.f32 %f667, 0f00000000;
|
|
fma.rn.f32 %f1351, %f666, %f143, %f667;
|
|
|
|
BB0_98:
|
|
fma.rn.f32 %f1352, %f1351, %f1349, %f1349;
|
|
@%p62 bra BB0_100;
|
|
|
|
mov.f32 %f668, 0f3F800000;
|
|
fma.rn.f32 %f1352, %f1351, %f143, %f668;
|
|
|
|
BB0_100:
|
|
and.b32 %r556, %r909, 2;
|
|
setp.eq.s32 %p65, %r556, 0;
|
|
@%p65 bra BB0_102;
|
|
|
|
mov.f32 %f669, 0f00000000;
|
|
mov.f32 %f670, 0fBF800000;
|
|
fma.rn.f32 %f1352, %f1352, %f670, %f669;
|
|
|
|
BB0_102:
|
|
mul.f32 %f674, %f118, %f1346;
|
|
mul.f32 %f675, %f674, %f674;
|
|
mov.f32 %f1354, 0f3F800000;
|
|
sub.f32 %f677, %f1354, %f675;
|
|
mul.f32 %f678, %f118, %f1352;
|
|
mul.f32 %f679, %f678, %f678;
|
|
sub.f32 %f680, %f677, %f679;
|
|
mov.f32 %f1357, 0f00000000;
|
|
max.f32 %f681, %f1357, %f680;
|
|
sqrt.rn.f32 %f682, %f681;
|
|
mul.f32 %f683, %f115, %f678;
|
|
mul.f32 %f684, %f116, %f678;
|
|
mul.f32 %f685, %f117, %f678;
|
|
mul.f32 %f686, %f113, %f116;
|
|
mul.f32 %f687, %f114, %f117;
|
|
sub.f32 %f688, %f686, %f687;
|
|
fma.rn.f32 %f689, %f688, %f674, %f683;
|
|
mul.f32 %f690, %f112, %f117;
|
|
mul.f32 %f691, %f113, %f115;
|
|
sub.f32 %f692, %f690, %f691;
|
|
fma.rn.f32 %f693, %f692, %f674, %f684;
|
|
mul.f32 %f694, %f114, %f115;
|
|
mul.f32 %f695, %f112, %f116;
|
|
sub.f32 %f696, %f694, %f695;
|
|
fma.rn.f32 %f697, %f696, %f674, %f685;
|
|
fma.rn.f32 %f155, %f112, %f682, %f689;
|
|
fma.rn.f32 %f156, %f114, %f682, %f693;
|
|
fma.rn.f32 %f157, %f113, %f682, %f697;
|
|
add.u64 %rd169, %SP, 28;
|
|
cvta.to.local.u64 %rd28, %rd169;
|
|
st.local.u32 [%rd28], %r433;
|
|
setp.gt.f32 %p66, %f156, 0f00000000;
|
|
setp.eq.f32 %p67, %f605, 0f00000000;
|
|
or.pred %p68, %p67, %p66;
|
|
mov.f32 %f1358, %f1357;
|
|
mov.f32 %f1359, %f1357;
|
|
@!%p68 bra BB0_106;
|
|
bra.uni BB0_103;
|
|
|
|
BB0_103:
|
|
mov.u32 %r561, 1065353216;
|
|
st.local.u32 [%rd28], %r561;
|
|
mul.f32 %f709, %f72, 0f3456BF95;
|
|
abs.f32 %f710, %f709;
|
|
mul.f32 %f711, %f73, 0f3456BF95;
|
|
abs.f32 %f712, %f711;
|
|
mul.f32 %f713, %f74, 0f3456BF95;
|
|
abs.f32 %f714, %f713;
|
|
max.f32 %f715, %f710, %f712;
|
|
max.f32 %f716, %f715, %f714;
|
|
max.f32 %f704, %f716, %f575;
|
|
ld.global.u32 %r558, [root];
|
|
mov.f32 %f705, 0f6C4ECB8F;
|
|
mov.u32 %r560, 4;
|
|
// inline asm
|
|
call _rt_trace_64, (%r558, %f72, %f73, %f74, %f155, %f156, %f157, %r455, %f704, %f705, %rd169, %r560);
|
|
// inline asm
|
|
setp.leu.f32 %p69, %f610, 0f00000000;
|
|
mov.f32 %f1355, %f1354;
|
|
mov.f32 %f1356, %f1354;
|
|
@%p69 bra BB0_105;
|
|
|
|
cvt.rzi.s32.f32 %r562, %f611;
|
|
neg.f32 %f724, %f155;
|
|
neg.f32 %f723, %f156;
|
|
neg.f32 %f722, %f157;
|
|
mov.u32 %r563, 6;
|
|
// inline asm
|
|
call (%f1354, %f1355, %f1356, %f721), _rt_texture_get_base_id, (%r562, %r563, %f722, %f723, %f724, %r433);
|
|
// inline asm
|
|
|
|
BB0_105:
|
|
ld.local.f32 %f725, [%rd28];
|
|
fma.rn.f32 %f1357, %f1354, %f725, 0f00000000;
|
|
fma.rn.f32 %f1358, %f1355, %f725, 0f00000000;
|
|
fma.rn.f32 %f1359, %f1356, %f725, 0f00000000;
|
|
|
|
BB0_106:
|
|
ld.global.u32 %r565, [bounceFlag];
|
|
setp.lt.s32 %p70, %r565, 1;
|
|
@%p70 bra BB0_108;
|
|
|
|
add.f32 %f726, %f609, 0fBF800000;
|
|
mul.f32 %f727, %f726, 0f42C80000;
|
|
mul.f32 %f1357, %f727, %f1357;
|
|
mul.f32 %f1358, %f727, %f1358;
|
|
mul.f32 %f1359, %f727, %f1359;
|
|
|
|
BB0_108:
|
|
cvt.rn.f32.u32 %f728, %r72;
|
|
mul.f32 %f729, %f606, %f1357;
|
|
mul.f32 %f1394, %f728, %f729;
|
|
mul.f32 %f730, %f607, %f1358;
|
|
mul.f32 %f1395, %f728, %f730;
|
|
mul.f32 %f731, %f608, %f1359;
|
|
mul.f32 %f1396, %f728, %f731;
|
|
bra.uni BB0_162;
|
|
|
|
BB0_121:
|
|
mov.u32 %r915, %r151;
|
|
mov.u32 %r916, %r160;
|
|
|
|
BB0_123:
|
|
clz.b32 %r918, %r914;
|
|
setp.eq.s32 %p86, %r918, 0;
|
|
shl.b32 %r594, %r914, %r918;
|
|
mov.u32 %r595, 32;
|
|
sub.s32 %r596, %r595, %r918;
|
|
shr.u32 %r597, %r916, %r596;
|
|
add.s32 %r598, %r597, %r594;
|
|
selp.b32 %r169, %r914, %r598, %p86;
|
|
mov.u32 %r599, -921707870;
|
|
mul.hi.u32 %r917, %r169, %r599;
|
|
setp.eq.s32 %p87, %r151, 0;
|
|
neg.s32 %r600, %r161;
|
|
selp.b32 %r919, %r161, %r600, %p87;
|
|
setp.lt.s32 %p88, %r917, 1;
|
|
@%p88 bra BB0_125;
|
|
|
|
mul.lo.s32 %r601, %r169, -921707870;
|
|
shr.u32 %r602, %r601, 31;
|
|
shl.b32 %r603, %r917, 1;
|
|
add.s32 %r917, %r602, %r603;
|
|
add.s32 %r918, %r918, 1;
|
|
|
|
BB0_125:
|
|
mov.u32 %r604, 126;
|
|
sub.s32 %r605, %r604, %r918;
|
|
shl.b32 %r606, %r605, 23;
|
|
add.s32 %r607, %r917, 1;
|
|
shr.u32 %r608, %r607, 7;
|
|
add.s32 %r609, %r608, 1;
|
|
shr.u32 %r610, %r609, 1;
|
|
add.s32 %r611, %r610, %r606;
|
|
or.b32 %r612, %r611, %r915;
|
|
mov.b32 %f1365, %r612;
|
|
|
|
BB0_126:
|
|
mul.rn.f32 %f197, %f1365, %f1365;
|
|
add.s32 %r177, %r919, 1;
|
|
and.b32 %r178, %r177, 1;
|
|
setp.eq.s32 %p89, %r178, 0;
|
|
@%p89 bra BB0_128;
|
|
|
|
mov.f32 %f823, 0fBAB6061A;
|
|
mov.f32 %f824, 0f37CCF5CE;
|
|
fma.rn.f32 %f1366, %f824, %f197, %f823;
|
|
bra.uni BB0_129;
|
|
|
|
BB0_128:
|
|
mov.f32 %f825, 0f3C08839E;
|
|
mov.f32 %f826, 0fB94CA1F9;
|
|
fma.rn.f32 %f1366, %f826, %f197, %f825;
|
|
|
|
BB0_129:
|
|
@%p89 bra BB0_131;
|
|
|
|
mov.f32 %f827, 0f3D2AAAA5;
|
|
fma.rn.f32 %f828, %f1366, %f197, %f827;
|
|
mov.f32 %f829, 0fBF000000;
|
|
fma.rn.f32 %f1367, %f828, %f197, %f829;
|
|
bra.uni BB0_132;
|
|
|
|
BB0_131:
|
|
mov.f32 %f830, 0fBE2AAAA3;
|
|
fma.rn.f32 %f831, %f1366, %f197, %f830;
|
|
mov.f32 %f832, 0f00000000;
|
|
fma.rn.f32 %f1367, %f831, %f197, %f832;
|
|
|
|
BB0_132:
|
|
fma.rn.f32 %f1368, %f1367, %f1365, %f1365;
|
|
@%p89 bra BB0_134;
|
|
|
|
mov.f32 %f833, 0f3F800000;
|
|
fma.rn.f32 %f1368, %f1367, %f197, %f833;
|
|
|
|
BB0_134:
|
|
and.b32 %r613, %r177, 2;
|
|
setp.eq.s32 %p92, %r613, 0;
|
|
@%p92 bra BB0_136;
|
|
|
|
mov.f32 %f834, 0f00000000;
|
|
mov.f32 %f835, 0fBF800000;
|
|
fma.rn.f32 %f1368, %f1368, %f835, %f834;
|
|
|
|
BB0_136:
|
|
@%p80 bra BB0_138;
|
|
|
|
mov.f32 %f836, 0f00000000;
|
|
mul.rn.f32 %f1370, %f1370, %f836;
|
|
|
|
BB0_138:
|
|
mul.f32 %f837, %f1370, 0f3F22F983;
|
|
cvt.rni.s32.f32 %r929, %f837;
|
|
cvt.rn.f32.s32 %f838, %r929;
|
|
neg.f32 %f839, %f838;
|
|
fma.rn.f32 %f841, %f839, %f817, %f1370;
|
|
fma.rn.f32 %f843, %f839, %f819, %f841;
|
|
fma.rn.f32 %f1371, %f839, %f821, %f843;
|
|
abs.f32 %f845, %f1370;
|
|
setp.leu.f32 %p94, %f845, 0f47CE4780;
|
|
@%p94 bra BB0_149;
|
|
|
|
mov.b32 %r180, %f1370;
|
|
shr.u32 %r181, %r180, 23;
|
|
shl.b32 %r616, %r180, 8;
|
|
or.b32 %r182, %r616, -2147483648;
|
|
mov.u32 %r920, 0;
|
|
mov.u64 %rd297, 0;
|
|
mov.u64 %rd296, %rd29;
|
|
mov.u32 %r921, %r920;
|
|
|
|
BB0_140:
|
|
.pragma "nounroll";
|
|
shl.b64 %rd179, %rd297, 2;
|
|
mov.u64 %rd180, __cudart_i2opi_f;
|
|
add.s64 %rd181, %rd180, %rd179;
|
|
ld.const.u32 %r619, [%rd181];
|
|
// inline asm
|
|
{
|
|
mad.lo.cc.u32 %r617, %r619, %r182, %r921;
|
|
madc.hi.u32 %r921, %r619, %r182, 0;
|
|
}
|
|
// inline asm
|
|
st.local.u32 [%rd296], %r617;
|
|
add.s32 %r920, %r920, 1;
|
|
cvt.s64.s32 %rd297, %r920;
|
|
mul.wide.s32 %rd182, %r920, 4;
|
|
add.s64 %rd296, %rd29, %rd182;
|
|
setp.ne.s32 %p95, %r920, 6;
|
|
@%p95 bra BB0_140;
|
|
|
|
and.b32 %r622, %r181, 255;
|
|
add.s32 %r623, %r622, -128;
|
|
shr.u32 %r624, %r623, 5;
|
|
and.b32 %r187, %r180, -2147483648;
|
|
st.local.u32 [%rd30], %r921;
|
|
mov.u32 %r625, 6;
|
|
sub.s32 %r626, %r625, %r624;
|
|
mul.wide.s32 %rd183, %r626, 4;
|
|
add.s64 %rd42, %rd29, %rd183;
|
|
ld.local.u32 %r922, [%rd42];
|
|
ld.local.u32 %r923, [%rd42+-4];
|
|
and.b32 %r190, %r181, 31;
|
|
setp.eq.s32 %p96, %r190, 0;
|
|
@%p96 bra BB0_143;
|
|
|
|
mov.u32 %r627, 32;
|
|
sub.s32 %r628, %r627, %r190;
|
|
shr.u32 %r629, %r923, %r628;
|
|
shl.b32 %r630, %r922, %r190;
|
|
add.s32 %r922, %r629, %r630;
|
|
ld.local.u32 %r631, [%rd42+-8];
|
|
shr.u32 %r632, %r631, %r628;
|
|
shl.b32 %r633, %r923, %r190;
|
|
add.s32 %r923, %r632, %r633;
|
|
|
|
BB0_143:
|
|
shr.u32 %r634, %r923, 30;
|
|
shl.b32 %r635, %r922, 2;
|
|
add.s32 %r924, %r634, %r635;
|
|
shl.b32 %r196, %r923, 2;
|
|
shr.u32 %r636, %r924, 31;
|
|
shr.u32 %r637, %r922, 30;
|
|
add.s32 %r197, %r636, %r637;
|
|
setp.eq.s32 %p97, %r636, 0;
|
|
@%p97 bra BB0_144;
|
|
|
|
not.b32 %r638, %r924;
|
|
neg.s32 %r926, %r196;
|
|
setp.eq.s32 %p98, %r196, 0;
|
|
selp.u32 %r639, 1, 0, %p98;
|
|
add.s32 %r924, %r639, %r638;
|
|
xor.b32 %r925, %r187, -2147483648;
|
|
bra.uni BB0_146;
|
|
|
|
BB0_144:
|
|
mov.u32 %r925, %r187;
|
|
mov.u32 %r926, %r196;
|
|
|
|
BB0_146:
|
|
clz.b32 %r928, %r924;
|
|
setp.eq.s32 %p99, %r928, 0;
|
|
shl.b32 %r640, %r924, %r928;
|
|
mov.u32 %r641, 32;
|
|
sub.s32 %r642, %r641, %r928;
|
|
shr.u32 %r643, %r926, %r642;
|
|
add.s32 %r644, %r643, %r640;
|
|
selp.b32 %r205, %r924, %r644, %p99;
|
|
mov.u32 %r645, -921707870;
|
|
mul.hi.u32 %r927, %r205, %r645;
|
|
setp.eq.s32 %p100, %r187, 0;
|
|
neg.s32 %r646, %r197;
|
|
selp.b32 %r929, %r197, %r646, %p100;
|
|
setp.lt.s32 %p101, %r927, 1;
|
|
@%p101 bra BB0_148;
|
|
|
|
mul.lo.s32 %r647, %r205, -921707870;
|
|
shr.u32 %r648, %r647, 31;
|
|
shl.b32 %r649, %r927, 1;
|
|
add.s32 %r927, %r648, %r649;
|
|
add.s32 %r928, %r928, 1;
|
|
|
|
BB0_148:
|
|
mov.u32 %r650, 126;
|
|
sub.s32 %r651, %r650, %r928;
|
|
shl.b32 %r652, %r651, 23;
|
|
add.s32 %r653, %r927, 1;
|
|
shr.u32 %r654, %r653, 7;
|
|
add.s32 %r655, %r654, 1;
|
|
shr.u32 %r656, %r655, 1;
|
|
add.s32 %r657, %r656, %r652;
|
|
or.b32 %r658, %r657, %r925;
|
|
mov.b32 %f1371, %r658;
|
|
|
|
BB0_149:
|
|
mul.rn.f32 %f214, %f1371, %f1371;
|
|
and.b32 %r213, %r929, 1;
|
|
setp.eq.s32 %p102, %r213, 0;
|
|
@%p102 bra BB0_151;
|
|
|
|
mov.f32 %f846, 0fBAB6061A;
|
|
mov.f32 %f847, 0f37CCF5CE;
|
|
fma.rn.f32 %f1372, %f847, %f214, %f846;
|
|
bra.uni BB0_152;
|
|
|
|
BB0_151:
|
|
mov.f32 %f848, 0f3C08839E;
|
|
mov.f32 %f849, 0fB94CA1F9;
|
|
fma.rn.f32 %f1372, %f849, %f214, %f848;
|
|
|
|
BB0_152:
|
|
@%p102 bra BB0_154;
|
|
|
|
mov.f32 %f850, 0f3D2AAAA5;
|
|
fma.rn.f32 %f851, %f1372, %f214, %f850;
|
|
mov.f32 %f852, 0fBF000000;
|
|
fma.rn.f32 %f1373, %f851, %f214, %f852;
|
|
bra.uni BB0_155;
|
|
|
|
BB0_154:
|
|
mov.f32 %f853, 0fBE2AAAA3;
|
|
fma.rn.f32 %f854, %f1372, %f214, %f853;
|
|
mov.f32 %f855, 0f00000000;
|
|
fma.rn.f32 %f1373, %f854, %f214, %f855;
|
|
|
|
BB0_155:
|
|
fma.rn.f32 %f1374, %f1373, %f1371, %f1371;
|
|
@%p102 bra BB0_157;
|
|
|
|
mov.f32 %f856, 0f3F800000;
|
|
fma.rn.f32 %f1374, %f1373, %f214, %f856;
|
|
|
|
BB0_157:
|
|
and.b32 %r659, %r929, 2;
|
|
setp.eq.s32 %p105, %r659, 0;
|
|
@%p105 bra BB0_159;
|
|
|
|
mov.f32 %f857, 0f00000000;
|
|
mov.f32 %f858, 0fBF800000;
|
|
fma.rn.f32 %f1374, %f1374, %f858, %f857;
|
|
|
|
BB0_159:
|
|
mul.f32 %f867, %f189, %f1368;
|
|
mul.f32 %f868, %f867, %f867;
|
|
mov.f32 %f869, 0f3F800000;
|
|
sub.f32 %f870, %f869, %f868;
|
|
mul.f32 %f871, %f189, %f1374;
|
|
mul.f32 %f872, %f871, %f871;
|
|
sub.f32 %f873, %f870, %f872;
|
|
mov.f32 %f874, 0f00000000;
|
|
max.f32 %f875, %f874, %f873;
|
|
sqrt.rn.f32 %f876, %f875;
|
|
mul.f32 %f877, %f186, %f871;
|
|
mul.f32 %f878, %f187, %f871;
|
|
mul.f32 %f879, %f188, %f871;
|
|
mul.f32 %f880, %f181, %f188;
|
|
mul.f32 %f881, %f182, %f187;
|
|
sub.f32 %f882, %f881, %f880;
|
|
fma.rn.f32 %f883, %f882, %f867, %f877;
|
|
mul.f32 %f884, %f182, %f186;
|
|
mul.f32 %f885, %f180, %f188;
|
|
sub.f32 %f886, %f885, %f884;
|
|
fma.rn.f32 %f887, %f886, %f867, %f878;
|
|
mul.f32 %f888, %f180, %f187;
|
|
mul.f32 %f889, %f181, %f186;
|
|
sub.f32 %f890, %f889, %f888;
|
|
fma.rn.f32 %f891, %f890, %f867, %f879;
|
|
fma.rn.f32 %f892, %f180, %f876, %f883;
|
|
fma.rn.f32 %f893, %f181, %f876, %f887;
|
|
fma.rn.f32 %f894, %f182, %f876, %f891;
|
|
add.f32 %f895, %f180, %f892;
|
|
add.f32 %f896, %f181, %f893;
|
|
add.f32 %f897, %f182, %f894;
|
|
mul.f32 %f898, %f605, %f895;
|
|
mul.f32 %f899, %f605, %f896;
|
|
mul.f32 %f900, %f605, %f897;
|
|
sub.f32 %f901, %f898, %f180;
|
|
sub.f32 %f902, %f899, %f181;
|
|
sub.f32 %f903, %f900, %f182;
|
|
mul.f32 %f904, %f902, %f902;
|
|
fma.rn.f32 %f905, %f901, %f901, %f904;
|
|
fma.rn.f32 %f906, %f903, %f903, %f905;
|
|
sqrt.rn.f32 %f907, %f906;
|
|
rcp.rn.f32 %f908, %f907;
|
|
mul.f32 %f862, %f908, %f901;
|
|
mul.f32 %f863, %f908, %f902;
|
|
mul.f32 %f864, %f908, %f903;
|
|
mul.f32 %f909, %f72, 0f3456BF95;
|
|
abs.f32 %f910, %f909;
|
|
mul.f32 %f911, %f73, 0f3456BF95;
|
|
abs.f32 %f912, %f911;
|
|
mul.f32 %f913, %f74, 0f3456BF95;
|
|
abs.f32 %f914, %f913;
|
|
max.f32 %f915, %f910, %f912;
|
|
max.f32 %f916, %f915, %f914;
|
|
max.f32 %f865, %f916, %f575;
|
|
add.u64 %rd184, %SP, 28;
|
|
cvta.to.local.u64 %rd185, %rd184;
|
|
mov.u32 %r663, 1065353216;
|
|
st.local.u32 [%rd185], %r663;
|
|
ld.global.u32 %r660, [root];
|
|
mov.f32 %f866, 0f6C4ECB8F;
|
|
mov.u32 %r662, 4;
|
|
// inline asm
|
|
call _rt_trace_64, (%r660, %f72, %f73, %f74, %f862, %f863, %f864, %r455, %f865, %f866, %rd184, %r662);
|
|
// inline asm
|
|
ld.local.f32 %f1376, [%rd185];
|
|
|
|
BB0_160:
|
|
ld.global.u32 %r664, [bounceFlag];
|
|
setp.gt.s32 %p106, %r664, 0;
|
|
mul.f32 %f918, %f609, 0f42C80000;
|
|
mul.f32 %f919, %f1363, %f1376;
|
|
mul.f32 %f920, %f918, %f919;
|
|
selp.f32 %f921, %f920, %f919, %p106;
|
|
mul.f32 %f1377, %f606, %f921;
|
|
mul.f32 %f1378, %f607, %f921;
|
|
mul.f32 %f1379, %f608, %f921;
|
|
|
|
BB0_161:
|
|
cvt.rn.f32.u32 %f922, %r72;
|
|
mul.f32 %f1394, %f922, %f1377;
|
|
mul.f32 %f1395, %f922, %f1378;
|
|
mul.f32 %f1396, %f922, %f1379;
|
|
|
|
BB0_162:
|
|
mov.u64 %rd191, localLights;
|
|
cvta.global.u64 %rd190, %rd191;
|
|
mov.u32 %r666, 96;
|
|
// inline asm
|
|
call (%rd186, %rd187, %rd188, %rd189), _rt_buffer_get_size_64, (%rd190, %r455, %r666);
|
|
// inline asm
|
|
cvt.u32.u64 %r667, %rd186;
|
|
setp.eq.s32 %p107, %r667, 0;
|
|
@%p107 bra BB0_188;
|
|
|
|
ld.global.f32 %f930, [cameraPos];
|
|
sub.f32 %f927, %f72, %f930;
|
|
ld.global.f32 %f931, [cameraPos+4];
|
|
sub.f32 %f928, %f73, %f931;
|
|
ld.global.f32 %f932, [cameraPos+8];
|
|
sub.f32 %f929, %f74, %f932;
|
|
ld.global.u32 %r668, [tileCubemap];
|
|
mov.u32 %r669, 6;
|
|
mov.u32 %r670, 0;
|
|
// inline asm
|
|
call (%f923, %f924, %f925, %f926), _rt_texture_get_base_id, (%r668, %r669, %f927, %f928, %f929, %r670);
|
|
// inline asm
|
|
mov.b32 %r214, %f924;
|
|
setp.eq.s32 %p108, %r214, 0;
|
|
@%p108 bra BB0_188;
|
|
|
|
mov.b32 %r675, %f923;
|
|
cvt.rn.f32.u32 %f244, %r214;
|
|
mul.f32 %f935, %f3, %f244;
|
|
cvt.rzi.u32.f32 %r676, %f935;
|
|
add.s32 %r677, %r214, -1;
|
|
min.u32 %r678, %r676, %r677;
|
|
add.s32 %r679, %r678, %r675;
|
|
cvt.u64.u32 %rd194, %r679;
|
|
mov.u64 %rd204, localLightIndices;
|
|
cvta.global.u64 %rd193, %rd204;
|
|
mov.u32 %r672, 4;
|
|
mov.u64 %rd203, 0;
|
|
// inline asm
|
|
call (%rd192), _rt_buffer_get_64, (%rd193, %r455, %r672, %rd194, %rd203, %rd203, %rd203);
|
|
// inline asm
|
|
ld.u32 %rd200, [%rd192];
|
|
// inline asm
|
|
call (%rd198), _rt_buffer_get_64, (%rd190, %r455, %r666, %rd200, %rd203, %rd203, %rd203);
|
|
// inline asm
|
|
ld.v4.f32 {%f936, %f937, %f938, %f939}, [%rd198+80];
|
|
ld.v4.f32 {%f940, %f941, %f942, %f943}, [%rd198+64];
|
|
ld.v4.f32 {%f944, %f945, %f946, %f947}, [%rd198+48];
|
|
ld.v4.f32 {%f1386, %f1387, %f1388, %f951}, [%rd198+32];
|
|
ld.v4.f32 {%f952, %f953, %f954, %f955}, [%rd198+16];
|
|
ld.v4.f32 {%f956, %f957, %f958, %f959}, [%rd198];
|
|
sub.f32 %f960, %f957, %f72;
|
|
sub.f32 %f961, %f958, %f73;
|
|
sub.f32 %f962, %f959, %f74;
|
|
mul.f32 %f963, %f961, %f961;
|
|
fma.rn.f32 %f964, %f960, %f960, %f963;
|
|
fma.rn.f32 %f965, %f962, %f962, %f964;
|
|
sqrt.rn.f32 %f272, %f965;
|
|
rcp.rn.f32 %f966, %f272;
|
|
mul.f32 %f273, %f960, %f966;
|
|
mul.f32 %f274, %f961, %f966;
|
|
mul.f32 %f275, %f962, %f966;
|
|
mul.f32 %f276, %f955, %f272;
|
|
mov.f32 %f970, 0f40800000;
|
|
abs.f32 %f278, %f276;
|
|
setp.lt.f32 %p109, %f278, 0f00800000;
|
|
mul.f32 %f972, %f278, 0f4B800000;
|
|
selp.f32 %f973, 0fC3170000, 0fC2FE0000, %p109;
|
|
selp.f32 %f974, %f972, %f278, %p109;
|
|
mov.b32 %r680, %f974;
|
|
and.b32 %r681, %r680, 8388607;
|
|
or.b32 %r682, %r681, 1065353216;
|
|
mov.b32 %f975, %r682;
|
|
shr.u32 %r683, %r680, 23;
|
|
cvt.rn.f32.u32 %f976, %r683;
|
|
add.f32 %f977, %f973, %f976;
|
|
setp.gt.f32 %p110, %f975, 0f3FB504F3;
|
|
mul.f32 %f978, %f975, 0f3F000000;
|
|
add.f32 %f979, %f977, 0f3F800000;
|
|
selp.f32 %f980, %f978, %f975, %p110;
|
|
selp.f32 %f981, %f979, %f977, %p110;
|
|
add.f32 %f982, %f980, 0fBF800000;
|
|
add.f32 %f934, %f980, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f933,%f934;
|
|
// inline asm
|
|
add.f32 %f983, %f982, %f982;
|
|
mul.f32 %f984, %f933, %f983;
|
|
mul.f32 %f985, %f984, %f984;
|
|
mov.f32 %f986, 0f3C4CAF63;
|
|
mov.f32 %f987, 0f3B18F0FE;
|
|
fma.rn.f32 %f988, %f987, %f985, %f986;
|
|
mov.f32 %f989, 0f3DAAAABD;
|
|
fma.rn.f32 %f990, %f988, %f985, %f989;
|
|
mul.rn.f32 %f991, %f990, %f985;
|
|
mul.rn.f32 %f992, %f991, %f984;
|
|
sub.f32 %f993, %f982, %f984;
|
|
neg.f32 %f994, %f984;
|
|
add.f32 %f995, %f993, %f993;
|
|
fma.rn.f32 %f996, %f994, %f982, %f995;
|
|
mul.rn.f32 %f997, %f933, %f996;
|
|
add.f32 %f998, %f992, %f984;
|
|
sub.f32 %f999, %f984, %f998;
|
|
add.f32 %f1000, %f992, %f999;
|
|
add.f32 %f1001, %f997, %f1000;
|
|
add.f32 %f1002, %f998, %f1001;
|
|
sub.f32 %f1003, %f998, %f1002;
|
|
add.f32 %f1004, %f1001, %f1003;
|
|
mov.f32 %f1005, 0f3F317200;
|
|
mul.rn.f32 %f1006, %f981, %f1005;
|
|
mov.f32 %f1007, 0f35BFBE8E;
|
|
mul.rn.f32 %f1008, %f981, %f1007;
|
|
add.f32 %f1009, %f1006, %f1002;
|
|
sub.f32 %f1010, %f1006, %f1009;
|
|
add.f32 %f1011, %f1002, %f1010;
|
|
add.f32 %f1012, %f1004, %f1011;
|
|
add.f32 %f1013, %f1008, %f1012;
|
|
add.f32 %f1014, %f1009, %f1013;
|
|
sub.f32 %f1015, %f1009, %f1014;
|
|
add.f32 %f1016, %f1013, %f1015;
|
|
mul.rn.f32 %f279, %f970, %f1014;
|
|
neg.f32 %f1017, %f279;
|
|
fma.rn.f32 %f1018, %f970, %f1014, %f1017;
|
|
fma.rn.f32 %f1019, %f970, %f1016, %f1018;
|
|
mov.f32 %f1020, 0f00000000;
|
|
fma.rn.f32 %f280, %f1020, %f1014, %f1019;
|
|
add.rn.f32 %f281, %f279, %f280;
|
|
mov.b32 %r684, %f281;
|
|
setp.eq.s32 %p1, %r684, 1118925336;
|
|
add.s32 %r685, %r684, -1;
|
|
mov.b32 %f1021, %r685;
|
|
selp.f32 %f1022, %f1021, %f281, %p1;
|
|
mul.f32 %f1023, %f1022, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f1024, %f1023;
|
|
mov.f32 %f1025, 0fBF317200;
|
|
fma.rn.f32 %f1026, %f1024, %f1025, %f1022;
|
|
mov.f32 %f1027, 0fB5BFBE8E;
|
|
fma.rn.f32 %f1028, %f1024, %f1027, %f1026;
|
|
mul.f32 %f1029, %f1028, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f1030, %f1029;
|
|
add.f32 %f1031, %f1024, 0f00000000;
|
|
ex2.approx.f32 %f1032, %f1031;
|
|
mul.f32 %f1033, %f1030, %f1032;
|
|
setp.lt.f32 %p111, %f1022, 0fC2D20000;
|
|
selp.f32 %f1034, 0f00000000, %f1033, %p111;
|
|
setp.gt.f32 %p112, %f1022, 0f42D20000;
|
|
selp.f32 %f1383, 0f7F800000, %f1034, %p112;
|
|
setp.eq.f32 %p113, %f1383, 0f7F800000;
|
|
@%p113 bra BB0_166;
|
|
|
|
neg.f32 %f1035, %f281;
|
|
add.rn.f32 %f1036, %f279, %f1035;
|
|
add.rn.f32 %f1037, %f1036, %f280;
|
|
add.f32 %f1038, %f1037, 0f37000000;
|
|
selp.f32 %f1039, %f1038, %f1037, %p1;
|
|
fma.rn.f32 %f1383, %f1383, %f1039, %f1383;
|
|
|
|
BB0_166:
|
|
mov.f32 %f1313, 0f40000000;
|
|
cvt.rzi.f32.f32 %f1312, %f1313;
|
|
add.f32 %f1311, %f1312, %f1312;
|
|
mov.f32 %f1310, 0f40800000;
|
|
sub.f32 %f1309, %f1310, %f1311;
|
|
abs.f32 %f1308, %f1309;
|
|
setp.lt.f32 %p114, %f276, 0f00000000;
|
|
setp.eq.f32 %p115, %f1308, 0f3F800000;
|
|
and.pred %p2, %p114, %p115;
|
|
mov.b32 %r686, %f1383;
|
|
xor.b32 %r687, %r686, -2147483648;
|
|
mov.b32 %f1040, %r687;
|
|
selp.f32 %f1385, %f1040, %f1383, %p2;
|
|
setp.eq.f32 %p116, %f276, 0f00000000;
|
|
@%p116 bra BB0_169;
|
|
bra.uni BB0_167;
|
|
|
|
BB0_169:
|
|
add.f32 %f1043, %f276, %f276;
|
|
selp.f32 %f1385, %f1043, 0f00000000, %p115;
|
|
bra.uni BB0_170;
|
|
|
|
BB0_167:
|
|
setp.geu.f32 %p117, %f276, 0f00000000;
|
|
@%p117 bra BB0_170;
|
|
|
|
mov.f32 %f1317, 0f40800000;
|
|
cvt.rzi.f32.f32 %f1042, %f1317;
|
|
setp.neu.f32 %p118, %f1042, 0f40800000;
|
|
selp.f32 %f1385, 0f7FFFFFFF, %f1385, %p118;
|
|
|
|
BB0_170:
|
|
add.f32 %f1044, %f278, 0f40800000;
|
|
mov.b32 %r688, %f1044;
|
|
setp.lt.s32 %p120, %r688, 2139095040;
|
|
@%p120 bra BB0_175;
|
|
|
|
setp.gtu.f32 %p121, %f278, 0f7F800000;
|
|
@%p121 bra BB0_174;
|
|
bra.uni BB0_172;
|
|
|
|
BB0_174:
|
|
add.f32 %f1385, %f276, 0f40800000;
|
|
bra.uni BB0_175;
|
|
|
|
BB0_172:
|
|
setp.neu.f32 %p122, %f278, 0f7F800000;
|
|
@%p122 bra BB0_175;
|
|
|
|
selp.f32 %f1385, 0fFF800000, 0f7F800000, %p2;
|
|
|
|
BB0_175:
|
|
add.u64 %rd280, %SP, 32;
|
|
cvta.to.local.u64 %rd279, %rd280;
|
|
add.s64 %rd278, %rd279, 12;
|
|
mul.f32 %f1045, %f953, %f272;
|
|
mov.f32 %f1390, 0f3F800000;
|
|
sub.f32 %f1047, %f1390, %f1385;
|
|
setp.eq.f32 %p123, %f276, 0f3F800000;
|
|
selp.f32 %f1048, 0f00000000, %f1047, %p123;
|
|
cvt.sat.f32.f32 %f1049, %f1048;
|
|
fma.rn.f32 %f1050, %f1045, %f1045, %f954;
|
|
div.rn.f32 %f1389, %f1049, %f1050;
|
|
ld.local.f32 %f1051, [%rd278+-12];
|
|
ld.local.f32 %f1052, [%rd278+-8];
|
|
mul.f32 %f1053, %f274, %f1052;
|
|
fma.rn.f32 %f1054, %f273, %f1051, %f1053;
|
|
ld.local.f32 %f1055, [%rd278+-4];
|
|
fma.rn.f32 %f1056, %f275, %f1055, %f1054;
|
|
cvt.sat.f32.f32 %f308, %f1056;
|
|
setp.eq.f32 %p124, %f956, 0f3F800000;
|
|
@%p124 bra BB0_181;
|
|
bra.uni BB0_176;
|
|
|
|
BB0_181:
|
|
setp.leu.f32 %p128, %f951, 0f00000000;
|
|
@%p128 bra BB0_183;
|
|
|
|
mov.u32 %r869, 2;
|
|
mul.f32 %f1087, %f936, %f273;
|
|
mul.f32 %f1088, %f937, %f274;
|
|
neg.f32 %f1089, %f1088;
|
|
sub.f32 %f1090, %f1089, %f1087;
|
|
mul.f32 %f1091, %f938, %f275;
|
|
sub.f32 %f1092, %f1090, %f1091;
|
|
setp.gt.f32 %p129, %f1092, 0f00000000;
|
|
selp.f32 %f1093, 0f3F800000, 0f00000000, %p129;
|
|
mul.f32 %f1094, %f945, %f274;
|
|
fma.rn.f32 %f1095, %f944, %f273, %f1094;
|
|
mul.f32 %f1096, %f941, %f274;
|
|
fma.rn.f32 %f1097, %f940, %f273, %f1096;
|
|
fma.rn.f32 %f1098, %f946, %f275, %f1095;
|
|
fma.rn.f32 %f1099, %f942, %f275, %f1097;
|
|
fma.rn.f32 %f1100, %f947, %f1098, 0f3F000000;
|
|
mov.f32 %f1101, 0f3F800000;
|
|
sub.f32 %f1083, %f1101, %f1100;
|
|
fma.rn.f32 %f1084, %f947, %f1099, 0f3F000000;
|
|
cvt.rzi.s32.f32 %r692, %f951;
|
|
mov.f32 %f1086, 0f00000000;
|
|
// inline asm
|
|
call (%f1079, %f1080, %f1081, %f1082), _rt_texture_get_f_id, (%r692, %r869, %f1083, %f1084, %f1086, %f1086);
|
|
// inline asm
|
|
mul.f32 %f1102, %f1093, %f1079;
|
|
mul.f32 %f1103, %f1093, %f1080;
|
|
mul.f32 %f1104, %f1093, %f1081;
|
|
mul.f32 %f1386, %f1386, %f1102;
|
|
mul.f32 %f1387, %f1387, %f1103;
|
|
mul.f32 %f1388, %f1388, %f1104;
|
|
bra.uni BB0_183;
|
|
|
|
BB0_176:
|
|
setp.eq.f32 %p125, %f956, 0f40000000;
|
|
@%p125 bra BB0_179;
|
|
bra.uni BB0_177;
|
|
|
|
BB0_179:
|
|
setp.leu.f32 %p127, %f951, 0f00000000;
|
|
@%p127 bra BB0_183;
|
|
|
|
mov.u32 %r868, 0;
|
|
mov.u32 %r867, 6;
|
|
mul.f32 %f1073, %f945, %f274;
|
|
fma.rn.f32 %f1074, %f944, %f273, %f1073;
|
|
mul.f32 %f1075, %f941, %f274;
|
|
fma.rn.f32 %f1076, %f940, %f273, %f1075;
|
|
mul.f32 %f1077, %f937, %f274;
|
|
fma.rn.f32 %f1078, %f936, %f273, %f1077;
|
|
fma.rn.f32 %f1070, %f946, %f275, %f1074;
|
|
fma.rn.f32 %f1071, %f942, %f275, %f1076;
|
|
fma.rn.f32 %f1072, %f938, %f275, %f1078;
|
|
cvt.rzi.s32.f32 %r689, %f951;
|
|
// inline asm
|
|
call (%f1066, %f1067, %f1068, %f1069), _rt_texture_get_base_id, (%r689, %r867, %f1070, %f1071, %f1072, %r868);
|
|
// inline asm
|
|
mul.f32 %f1386, %f1386, %f1066;
|
|
mul.f32 %f1387, %f1387, %f1067;
|
|
mul.f32 %f1388, %f1388, %f1068;
|
|
bra.uni BB0_183;
|
|
|
|
BB0_177:
|
|
setp.neu.f32 %p126, %f956, 0f40800000;
|
|
@%p126 bra BB0_183;
|
|
|
|
mul.f32 %f1057, %f936, %f273;
|
|
mul.f32 %f1058, %f937, %f274;
|
|
neg.f32 %f1059, %f1058;
|
|
sub.f32 %f1060, %f1059, %f1057;
|
|
mul.f32 %f1061, %f938, %f275;
|
|
sub.f32 %f1062, %f1060, %f1061;
|
|
fma.rn.f32 %f1063, %f951, %f1062, %f947;
|
|
cvt.sat.f32.f32 %f1064, %f1063;
|
|
mul.f32 %f1065, %f1064, %f1064;
|
|
mul.f32 %f1389, %f1389, %f1065;
|
|
|
|
BB0_183:
|
|
mov.f32 %f1391, 0f00000000;
|
|
max.f32 %f1108, %f1386, %f1387;
|
|
max.f32 %f1109, %f1108, %f1388;
|
|
mul.f32 %f320, %f308, %f1389;
|
|
mul.f32 %f1110, %f320, %f1109;
|
|
setp.lt.f32 %p130, %f1110, 0f3727C5AC;
|
|
mov.f32 %f1392, %f1391;
|
|
mov.f32 %f1393, %f1391;
|
|
@%p130 bra BB0_187;
|
|
|
|
add.u64 %rd206, %SP, 0;
|
|
cvta.to.local.u64 %rd43, %rd206;
|
|
mov.u32 %r694, 1065353216;
|
|
st.local.u32 [%rd43], %r694;
|
|
setp.leu.f32 %p131, %f939, 0f00000000;
|
|
@%p131 bra BB0_186;
|
|
|
|
mov.u32 %r856, 4;
|
|
mov.f32 %f1315, 0f38D1B717;
|
|
mov.u32 %r855, 1;
|
|
fma.rn.f32 %f1120, %f1, 0f40000000, 0fBF800000;
|
|
fma.rn.f32 %f1121, %f2, 0f40000000, 0fBF800000;
|
|
fma.rn.f32 %f1122, %f3, 0f40000000, 0fBF800000;
|
|
fma.rn.f32 %f1123, %f1120, %f952, %f957;
|
|
fma.rn.f32 %f1124, %f1121, %f952, %f958;
|
|
fma.rn.f32 %f1125, %f1122, %f952, %f959;
|
|
sub.f32 %f1126, %f1123, %f72;
|
|
sub.f32 %f1127, %f1124, %f73;
|
|
sub.f32 %f1128, %f1125, %f74;
|
|
mul.f32 %f1129, %f1127, %f1127;
|
|
fma.rn.f32 %f1130, %f1126, %f1126, %f1129;
|
|
fma.rn.f32 %f1131, %f1128, %f1128, %f1130;
|
|
sqrt.rn.f32 %f1119, %f1131;
|
|
rcp.rn.f32 %f1132, %f1119;
|
|
mul.f32 %f1115, %f1126, %f1132;
|
|
mul.f32 %f1116, %f1127, %f1132;
|
|
mul.f32 %f1117, %f1128, %f1132;
|
|
mul.f32 %f1133, %f72, 0f3456BF95;
|
|
abs.f32 %f1134, %f1133;
|
|
mul.f32 %f1135, %f73, 0f3456BF95;
|
|
abs.f32 %f1136, %f1135;
|
|
mul.f32 %f1137, %f74, 0f3456BF95;
|
|
abs.f32 %f1138, %f1137;
|
|
max.f32 %f1139, %f1134, %f1136;
|
|
max.f32 %f1140, %f1139, %f1138;
|
|
max.f32 %f1118, %f1140, %f1315;
|
|
ld.global.u32 %r695, [root];
|
|
// inline asm
|
|
call _rt_trace_64, (%r695, %f72, %f73, %f74, %f1115, %f1116, %f1117, %r855, %f1118, %f1119, %rd206, %r856);
|
|
// inline asm
|
|
ld.local.f32 %f1390, [%rd43];
|
|
|
|
BB0_186:
|
|
ld.global.u32 %r698, [bounceFlag];
|
|
setp.gt.s32 %p132, %r698, 0;
|
|
mul.f32 %f1142, %f320, %f1390;
|
|
mul.f32 %f1143, %f943, %f1142;
|
|
selp.f32 %f1144, %f1143, %f1142, %p132;
|
|
mul.f32 %f1391, %f1386, %f1144;
|
|
mul.f32 %f1392, %f1387, %f1144;
|
|
mul.f32 %f1393, %f1388, %f1144;
|
|
|
|
BB0_187:
|
|
fma.rn.f32 %f1394, %f244, %f1391, %f1394;
|
|
fma.rn.f32 %f1395, %f244, %f1392, %f1395;
|
|
fma.rn.f32 %f1396, %f244, %f1393, %f1396;
|
|
|
|
BB0_188:
|
|
add.u64 %rd283, %SP, 32;
|
|
cvta.to.local.u64 %rd282, %rd283;
|
|
add.s64 %rd281, %rd282, 12;
|
|
ld.local.f32 %f337, [%rd281+-8];
|
|
ld.local.f32 %f336, [%rd281+-4];
|
|
ld.local.f32 %f335, [%rd281+-12];
|
|
abs.f32 %f1145, %f335;
|
|
abs.f32 %f1146, %f336;
|
|
setp.gt.f32 %p133, %f1145, %f1146;
|
|
neg.f32 %f1147, %f337;
|
|
neg.f32 %f1148, %f336;
|
|
selp.f32 %f1149, %f1147, 0f00000000, %p133;
|
|
selp.f32 %f1150, %f335, %f1148, %p133;
|
|
selp.f32 %f1151, 0f00000000, %f337, %p133;
|
|
mul.f32 %f1152, %f1150, %f1150;
|
|
fma.rn.f32 %f1153, %f1149, %f1149, %f1152;
|
|
fma.rn.f32 %f1154, %f1151, %f1151, %f1153;
|
|
sqrt.rn.f32 %f1155, %f1154;
|
|
rcp.rn.f32 %f1156, %f1155;
|
|
mul.f32 %f338, %f1149, %f1156;
|
|
mul.f32 %f339, %f1150, %f1156;
|
|
mul.f32 %f340, %f1151, %f1156;
|
|
sqrt.rn.f32 %f341, %f1;
|
|
add.u64 %rd208, %SP, 0;
|
|
cvta.to.local.u64 %rd44, %rd208;
|
|
mul.f32 %f1403, %f2, 0f40C90FDB;
|
|
abs.f32 %f343, %f1403;
|
|
setp.neu.f32 %p134, %f343, 0f7F800000;
|
|
mov.f32 %f1397, %f1403;
|
|
@%p134 bra BB0_190;
|
|
|
|
mov.f32 %f1157, 0f00000000;
|
|
mul.rn.f32 %f1397, %f1403, %f1157;
|
|
|
|
BB0_190:
|
|
mul.f32 %f1158, %f1397, 0f3F22F983;
|
|
cvt.rni.s32.f32 %r939, %f1158;
|
|
cvt.rn.f32.s32 %f1159, %r939;
|
|
neg.f32 %f1160, %f1159;
|
|
mov.f32 %f1161, 0f3FC90FDA;
|
|
fma.rn.f32 %f1162, %f1160, %f1161, %f1397;
|
|
mov.f32 %f1163, 0f33A22168;
|
|
fma.rn.f32 %f1164, %f1160, %f1163, %f1162;
|
|
mov.f32 %f1165, 0f27C234C5;
|
|
fma.rn.f32 %f1398, %f1160, %f1165, %f1164;
|
|
abs.f32 %f1166, %f1397;
|
|
add.s64 %rd45, %rd44, 24;
|
|
setp.leu.f32 %p135, %f1166, 0f47CE4780;
|
|
@%p135 bra BB0_201;
|
|
|
|
mov.b32 %r216, %f1397;
|
|
shr.u32 %r217, %r216, 23;
|
|
shl.b32 %r701, %r216, 8;
|
|
or.b32 %r218, %r701, -2147483648;
|
|
mov.u32 %r931, 0;
|
|
mov.u64 %rd298, __cudart_i2opi_f;
|
|
mov.u32 %r930, -6;
|
|
mov.u64 %rd299, %rd44;
|
|
|
|
BB0_192:
|
|
.pragma "nounroll";
|
|
ld.const.u32 %r704, [%rd298];
|
|
// inline asm
|
|
{
|
|
mad.lo.cc.u32 %r702, %r704, %r218, %r931;
|
|
madc.hi.u32 %r931, %r704, %r218, 0;
|
|
}
|
|
// inline asm
|
|
st.local.u32 [%rd299], %r702;
|
|
add.s64 %rd299, %rd299, 4;
|
|
add.s64 %rd298, %rd298, 4;
|
|
add.s32 %r930, %r930, 1;
|
|
setp.ne.s32 %p136, %r930, 0;
|
|
@%p136 bra BB0_192;
|
|
|
|
and.b32 %r707, %r217, 255;
|
|
add.s32 %r708, %r707, -128;
|
|
shr.u32 %r709, %r708, 5;
|
|
and.b32 %r223, %r216, -2147483648;
|
|
st.local.u32 [%rd45], %r931;
|
|
mov.u32 %r710, 6;
|
|
sub.s32 %r711, %r710, %r709;
|
|
mul.wide.s32 %rd210, %r711, 4;
|
|
add.s64 %rd50, %rd44, %rd210;
|
|
ld.local.u32 %r932, [%rd50];
|
|
ld.local.u32 %r933, [%rd50+-4];
|
|
and.b32 %r226, %r217, 31;
|
|
setp.eq.s32 %p137, %r226, 0;
|
|
@%p137 bra BB0_195;
|
|
|
|
mov.u32 %r712, 32;
|
|
sub.s32 %r713, %r712, %r226;
|
|
shr.u32 %r714, %r933, %r713;
|
|
shl.b32 %r715, %r932, %r226;
|
|
add.s32 %r932, %r714, %r715;
|
|
ld.local.u32 %r716, [%rd50+-8];
|
|
shr.u32 %r717, %r716, %r713;
|
|
shl.b32 %r718, %r933, %r226;
|
|
add.s32 %r933, %r717, %r718;
|
|
|
|
BB0_195:
|
|
shr.u32 %r719, %r933, 30;
|
|
shl.b32 %r720, %r932, 2;
|
|
add.s32 %r934, %r719, %r720;
|
|
shl.b32 %r232, %r933, 2;
|
|
shr.u32 %r721, %r934, 31;
|
|
shr.u32 %r722, %r932, 30;
|
|
add.s32 %r233, %r721, %r722;
|
|
setp.eq.s32 %p138, %r721, 0;
|
|
@%p138 bra BB0_196;
|
|
|
|
not.b32 %r723, %r934;
|
|
neg.s32 %r936, %r232;
|
|
setp.eq.s32 %p139, %r232, 0;
|
|
selp.u32 %r724, 1, 0, %p139;
|
|
add.s32 %r934, %r724, %r723;
|
|
xor.b32 %r935, %r223, -2147483648;
|
|
bra.uni BB0_198;
|
|
|
|
BB0_196:
|
|
mov.u32 %r935, %r223;
|
|
mov.u32 %r936, %r232;
|
|
|
|
BB0_198:
|
|
clz.b32 %r938, %r934;
|
|
setp.eq.s32 %p140, %r938, 0;
|
|
shl.b32 %r725, %r934, %r938;
|
|
mov.u32 %r726, 32;
|
|
sub.s32 %r727, %r726, %r938;
|
|
shr.u32 %r728, %r936, %r727;
|
|
add.s32 %r729, %r728, %r725;
|
|
selp.b32 %r241, %r934, %r729, %p140;
|
|
mov.u32 %r730, -921707870;
|
|
mul.hi.u32 %r937, %r241, %r730;
|
|
setp.eq.s32 %p141, %r223, 0;
|
|
neg.s32 %r731, %r233;
|
|
selp.b32 %r939, %r233, %r731, %p141;
|
|
setp.lt.s32 %p142, %r937, 1;
|
|
@%p142 bra BB0_200;
|
|
|
|
mul.lo.s32 %r732, %r241, -921707870;
|
|
shr.u32 %r733, %r732, 31;
|
|
shl.b32 %r734, %r937, 1;
|
|
add.s32 %r937, %r733, %r734;
|
|
add.s32 %r938, %r938, 1;
|
|
|
|
BB0_200:
|
|
mov.u32 %r735, 126;
|
|
sub.s32 %r736, %r735, %r938;
|
|
shl.b32 %r737, %r736, 23;
|
|
add.s32 %r738, %r937, 1;
|
|
shr.u32 %r739, %r738, 7;
|
|
add.s32 %r740, %r739, 1;
|
|
shr.u32 %r741, %r740, 1;
|
|
add.s32 %r742, %r741, %r737;
|
|
or.b32 %r743, %r742, %r935;
|
|
mov.b32 %f1398, %r743;
|
|
|
|
BB0_201:
|
|
mul.rn.f32 %f349, %f1398, %f1398;
|
|
add.s32 %r249, %r939, 1;
|
|
and.b32 %r250, %r249, 1;
|
|
setp.eq.s32 %p143, %r250, 0;
|
|
@%p143 bra BB0_203;
|
|
|
|
mov.f32 %f1167, 0fBAB6061A;
|
|
mov.f32 %f1168, 0f37CCF5CE;
|
|
fma.rn.f32 %f1399, %f1168, %f349, %f1167;
|
|
bra.uni BB0_204;
|
|
|
|
BB0_203:
|
|
mov.f32 %f1169, 0f3C08839E;
|
|
mov.f32 %f1170, 0fB94CA1F9;
|
|
fma.rn.f32 %f1399, %f1170, %f349, %f1169;
|
|
|
|
BB0_204:
|
|
@%p143 bra BB0_206;
|
|
|
|
mov.f32 %f1171, 0f3D2AAAA5;
|
|
fma.rn.f32 %f1172, %f1399, %f349, %f1171;
|
|
mov.f32 %f1173, 0fBF000000;
|
|
fma.rn.f32 %f1400, %f1172, %f349, %f1173;
|
|
bra.uni BB0_207;
|
|
|
|
BB0_206:
|
|
mov.f32 %f1174, 0fBE2AAAA3;
|
|
fma.rn.f32 %f1175, %f1399, %f349, %f1174;
|
|
mov.f32 %f1176, 0f00000000;
|
|
fma.rn.f32 %f1400, %f1175, %f349, %f1176;
|
|
|
|
BB0_207:
|
|
fma.rn.f32 %f1401, %f1400, %f1398, %f1398;
|
|
@%p143 bra BB0_209;
|
|
|
|
mov.f32 %f1177, 0f3F800000;
|
|
fma.rn.f32 %f1401, %f1400, %f349, %f1177;
|
|
|
|
BB0_209:
|
|
and.b32 %r744, %r249, 2;
|
|
setp.eq.s32 %p146, %r744, 0;
|
|
@%p146 bra BB0_211;
|
|
|
|
mov.f32 %f1178, 0f00000000;
|
|
mov.f32 %f1179, 0fBF800000;
|
|
fma.rn.f32 %f1401, %f1401, %f1179, %f1178;
|
|
|
|
BB0_211:
|
|
@%p134 bra BB0_213;
|
|
|
|
mov.f32 %f1180, 0f00000000;
|
|
mul.rn.f32 %f1403, %f1403, %f1180;
|
|
|
|
BB0_213:
|
|
mul.f32 %f1181, %f1403, 0f3F22F983;
|
|
cvt.rni.s32.f32 %r949, %f1181;
|
|
cvt.rn.f32.s32 %f1182, %r949;
|
|
neg.f32 %f1183, %f1182;
|
|
fma.rn.f32 %f1185, %f1183, %f1161, %f1403;
|
|
fma.rn.f32 %f1187, %f1183, %f1163, %f1185;
|
|
fma.rn.f32 %f1404, %f1183, %f1165, %f1187;
|
|
abs.f32 %f1189, %f1403;
|
|
setp.leu.f32 %p148, %f1189, 0f47CE4780;
|
|
@%p148 bra BB0_224;
|
|
|
|
mov.b32 %r252, %f1403;
|
|
shr.u32 %r253, %r252, 23;
|
|
shl.b32 %r747, %r252, 8;
|
|
or.b32 %r254, %r747, -2147483648;
|
|
mov.u32 %r941, 0;
|
|
mov.u64 %rd300, __cudart_i2opi_f;
|
|
mov.u32 %r940, -6;
|
|
mov.u64 %rd301, %rd44;
|
|
|
|
BB0_215:
|
|
.pragma "nounroll";
|
|
ld.const.u32 %r750, [%rd300];
|
|
// inline asm
|
|
{
|
|
mad.lo.cc.u32 %r748, %r750, %r254, %r941;
|
|
madc.hi.u32 %r941, %r750, %r254, 0;
|
|
}
|
|
// inline asm
|
|
st.local.u32 [%rd301], %r748;
|
|
add.s64 %rd301, %rd301, 4;
|
|
add.s64 %rd300, %rd300, 4;
|
|
add.s32 %r940, %r940, 1;
|
|
setp.ne.s32 %p149, %r940, 0;
|
|
@%p149 bra BB0_215;
|
|
|
|
and.b32 %r753, %r253, 255;
|
|
add.s32 %r754, %r753, -128;
|
|
shr.u32 %r755, %r754, 5;
|
|
and.b32 %r259, %r252, -2147483648;
|
|
st.local.u32 [%rd45], %r941;
|
|
mov.u32 %r756, 6;
|
|
sub.s32 %r757, %r756, %r755;
|
|
mul.wide.s32 %rd212, %r757, 4;
|
|
add.s64 %rd55, %rd44, %rd212;
|
|
ld.local.u32 %r942, [%rd55];
|
|
ld.local.u32 %r943, [%rd55+-4];
|
|
and.b32 %r262, %r253, 31;
|
|
setp.eq.s32 %p150, %r262, 0;
|
|
@%p150 bra BB0_218;
|
|
|
|
mov.u32 %r758, 32;
|
|
sub.s32 %r759, %r758, %r262;
|
|
shr.u32 %r760, %r943, %r759;
|
|
shl.b32 %r761, %r942, %r262;
|
|
add.s32 %r942, %r760, %r761;
|
|
ld.local.u32 %r762, [%rd55+-8];
|
|
shr.u32 %r763, %r762, %r759;
|
|
shl.b32 %r764, %r943, %r262;
|
|
add.s32 %r943, %r763, %r764;
|
|
|
|
BB0_218:
|
|
shr.u32 %r765, %r943, 30;
|
|
shl.b32 %r766, %r942, 2;
|
|
add.s32 %r944, %r765, %r766;
|
|
shl.b32 %r268, %r943, 2;
|
|
shr.u32 %r767, %r944, 31;
|
|
shr.u32 %r768, %r942, 30;
|
|
add.s32 %r269, %r767, %r768;
|
|
setp.eq.s32 %p151, %r767, 0;
|
|
@%p151 bra BB0_219;
|
|
|
|
not.b32 %r769, %r944;
|
|
neg.s32 %r946, %r268;
|
|
setp.eq.s32 %p152, %r268, 0;
|
|
selp.u32 %r770, 1, 0, %p152;
|
|
add.s32 %r944, %r770, %r769;
|
|
xor.b32 %r945, %r259, -2147483648;
|
|
bra.uni BB0_221;
|
|
|
|
BB0_219:
|
|
mov.u32 %r945, %r259;
|
|
mov.u32 %r946, %r268;
|
|
|
|
BB0_221:
|
|
clz.b32 %r948, %r944;
|
|
setp.eq.s32 %p153, %r948, 0;
|
|
shl.b32 %r771, %r944, %r948;
|
|
mov.u32 %r772, 32;
|
|
sub.s32 %r773, %r772, %r948;
|
|
shr.u32 %r774, %r946, %r773;
|
|
add.s32 %r775, %r774, %r771;
|
|
selp.b32 %r277, %r944, %r775, %p153;
|
|
mov.u32 %r776, -921707870;
|
|
mul.hi.u32 %r947, %r277, %r776;
|
|
setp.eq.s32 %p154, %r259, 0;
|
|
neg.s32 %r777, %r269;
|
|
selp.b32 %r949, %r269, %r777, %p154;
|
|
setp.lt.s32 %p155, %r947, 1;
|
|
@%p155 bra BB0_223;
|
|
|
|
mul.lo.s32 %r778, %r277, -921707870;
|
|
shr.u32 %r779, %r778, 31;
|
|
shl.b32 %r780, %r947, 1;
|
|
add.s32 %r947, %r779, %r780;
|
|
add.s32 %r948, %r948, 1;
|
|
|
|
BB0_223:
|
|
mov.u32 %r781, 126;
|
|
sub.s32 %r782, %r781, %r948;
|
|
shl.b32 %r783, %r782, 23;
|
|
add.s32 %r784, %r947, 1;
|
|
shr.u32 %r785, %r784, 7;
|
|
add.s32 %r786, %r785, 1;
|
|
shr.u32 %r787, %r786, 1;
|
|
add.s32 %r788, %r787, %r783;
|
|
or.b32 %r789, %r788, %r945;
|
|
mov.b32 %f1404, %r789;
|
|
|
|
BB0_224:
|
|
mul.rn.f32 %f366, %f1404, %f1404;
|
|
and.b32 %r285, %r949, 1;
|
|
setp.eq.s32 %p156, %r285, 0;
|
|
@%p156 bra BB0_226;
|
|
|
|
mov.f32 %f1190, 0fBAB6061A;
|
|
mov.f32 %f1191, 0f37CCF5CE;
|
|
fma.rn.f32 %f1405, %f1191, %f366, %f1190;
|
|
bra.uni BB0_227;
|
|
|
|
BB0_226:
|
|
mov.f32 %f1192, 0f3C08839E;
|
|
mov.f32 %f1193, 0fB94CA1F9;
|
|
fma.rn.f32 %f1405, %f1193, %f366, %f1192;
|
|
|
|
BB0_227:
|
|
@%p156 bra BB0_229;
|
|
|
|
mov.f32 %f1194, 0f3D2AAAA5;
|
|
fma.rn.f32 %f1195, %f1405, %f366, %f1194;
|
|
mov.f32 %f1196, 0fBF000000;
|
|
fma.rn.f32 %f1406, %f1195, %f366, %f1196;
|
|
bra.uni BB0_230;
|
|
|
|
BB0_229:
|
|
mov.f32 %f1197, 0fBE2AAAA3;
|
|
fma.rn.f32 %f1198, %f1405, %f366, %f1197;
|
|
mov.f32 %f1199, 0f00000000;
|
|
fma.rn.f32 %f1406, %f1198, %f366, %f1199;
|
|
|
|
BB0_230:
|
|
fma.rn.f32 %f1407, %f1406, %f1404, %f1404;
|
|
@%p156 bra BB0_232;
|
|
|
|
mov.f32 %f1200, 0f3F800000;
|
|
fma.rn.f32 %f1407, %f1406, %f366, %f1200;
|
|
|
|
BB0_232:
|
|
and.b32 %r790, %r949, 2;
|
|
setp.eq.s32 %p159, %r790, 0;
|
|
@%p159 bra BB0_234;
|
|
|
|
mov.f32 %f1201, 0f00000000;
|
|
mov.f32 %f1202, 0fBF800000;
|
|
fma.rn.f32 %f1407, %f1407, %f1202, %f1201;
|
|
|
|
BB0_234:
|
|
mov.f32 %f1316, 0f38D1B717;
|
|
mov.u32 %r857, 1;
|
|
mul.f32 %f1211, %f341, %f1401;
|
|
mul.f32 %f1212, %f1211, %f1211;
|
|
mov.f32 %f1213, 0f3F800000;
|
|
sub.f32 %f1214, %f1213, %f1212;
|
|
mul.f32 %f1215, %f341, %f1407;
|
|
mul.f32 %f1216, %f1215, %f1215;
|
|
sub.f32 %f1217, %f1214, %f1216;
|
|
mov.f32 %f1218, 0f00000000;
|
|
max.f32 %f1219, %f1218, %f1217;
|
|
sqrt.rn.f32 %f1220, %f1219;
|
|
mul.f32 %f1221, %f338, %f1215;
|
|
mul.f32 %f1222, %f339, %f1215;
|
|
mul.f32 %f1223, %f340, %f1215;
|
|
mul.f32 %f1224, %f336, %f339;
|
|
mul.f32 %f1225, %f337, %f340;
|
|
sub.f32 %f1226, %f1224, %f1225;
|
|
fma.rn.f32 %f1227, %f1226, %f1211, %f1221;
|
|
mul.f32 %f1228, %f335, %f340;
|
|
mul.f32 %f1229, %f336, %f338;
|
|
sub.f32 %f1230, %f1228, %f1229;
|
|
fma.rn.f32 %f1231, %f1230, %f1211, %f1222;
|
|
mul.f32 %f1232, %f337, %f338;
|
|
mul.f32 %f1233, %f335, %f339;
|
|
sub.f32 %f1234, %f1232, %f1233;
|
|
fma.rn.f32 %f1235, %f1234, %f1211, %f1223;
|
|
fma.rn.f32 %f1206, %f335, %f1220, %f1227;
|
|
fma.rn.f32 %f1207, %f337, %f1220, %f1231;
|
|
fma.rn.f32 %f1208, %f336, %f1220, %f1235;
|
|
add.u64 %rd213, %SP, 28;
|
|
cvta.to.local.u64 %rd214, %rd213;
|
|
mov.u32 %r794, 1065353216;
|
|
st.local.u32 [%rd214], %r794;
|
|
mul.f32 %f1236, %f72, 0f3456BF95;
|
|
abs.f32 %f1237, %f1236;
|
|
mul.f32 %f1238, %f73, 0f3456BF95;
|
|
abs.f32 %f1239, %f1238;
|
|
mul.f32 %f1240, %f74, 0f3456BF95;
|
|
abs.f32 %f1241, %f1240;
|
|
max.f32 %f1242, %f1237, %f1239;
|
|
max.f32 %f1243, %f1242, %f1241;
|
|
max.f32 %f1209, %f1243, %f1316;
|
|
ld.global.f32 %f1210, [aoRadius];
|
|
ld.global.u32 %r791, [root];
|
|
mov.u32 %r793, 4;
|
|
// inline asm
|
|
call _rt_trace_64, (%r791, %f72, %f73, %f74, %f1206, %f1207, %f1208, %r857, %f1209, %f1210, %rd213, %r793);
|
|
// inline asm
|
|
ld.local.f32 %f1418, [%rd214];
|
|
mov.u16 %rs10, 0;
|
|
|
|
BB0_236:
|
|
mov.u64 %rd285, GBufferAtten;
|
|
cvta.global.u64 %rd284, %rd285;
|
|
mov.u32 %r859, 12;
|
|
mov.u32 %r858, 2;
|
|
abs.f32 %f1255, %f1394;
|
|
setp.gtu.f32 %p161, %f1255, 0f7F800000;
|
|
selp.f32 %f1256, 0f00000000, %f1394, %p161;
|
|
abs.f32 %f1257, %f1395;
|
|
setp.gtu.f32 %p162, %f1257, 0f7F800000;
|
|
selp.f32 %f1258, 0f00000000, %f1395, %p162;
|
|
abs.f32 %f1259, %f1396;
|
|
setp.gtu.f32 %p163, %f1259, 0f7F800000;
|
|
selp.f32 %f1260, 0f00000000, %f1396, %p163;
|
|
fma.rn.f32 %f1422, %f1412, %f1256, %f1409;
|
|
fma.rn.f32 %f1423, %f1413, %f1258, %f1410;
|
|
fma.rn.f32 %f1424, %f1414, %f1260, %f1411;
|
|
ld.global.v2.u32 {%r800, %r801}, [pixelID];
|
|
cvt.u64.u32 %rd217, %r800;
|
|
cvt.u64.u32 %rd218, %r801;
|
|
mov.u64 %rd226, 0;
|
|
// inline asm
|
|
call (%rd215), _rt_buffer_get_64, (%rd284, %r858, %r859, %rd217, %rd218, %rd226, %rd226);
|
|
// inline asm
|
|
st.f32 [%rd215+8], %f1414;
|
|
st.f32 [%rd215+4], %f1413;
|
|
st.f32 [%rd215], %f1412;
|
|
ld.global.v2.u32 {%r804, %r805}, [pixelID];
|
|
cvt.u64.u32 %rd223, %r804;
|
|
cvt.u64.u32 %rd224, %r805;
|
|
mov.u64 %rd228, accumBufferAdd;
|
|
cvta.global.u64 %rd222, %rd228;
|
|
// inline asm
|
|
call (%rd221), _rt_buffer_get_64, (%rd222, %r858, %r859, %rd223, %rd224, %rd226, %rd226);
|
|
// inline asm
|
|
ld.global.u32 %r808, [bounceFlag];
|
|
setp.eq.s32 %p164, %r808, 0;
|
|
@%p164 bra BB0_238;
|
|
|
|
ld.f32 %f1261, [%rd221+8];
|
|
ld.f32 %f1262, [%rd221+4];
|
|
ld.f32 %f1263, [%rd221];
|
|
add.f32 %f1422, %f1422, %f1263;
|
|
add.f32 %f1423, %f1423, %f1262;
|
|
add.f32 %f1424, %f1424, %f1261;
|
|
bra.uni BB0_243;
|
|
|
|
BB0_238:
|
|
ld.global.u32 %r286, [firstSkylight];
|
|
setp.lt.s32 %p165, %r286, 0;
|
|
setp.eq.s16 %p166, %rs10, 0;
|
|
or.pred %p167, %p165, %p166;
|
|
@%p167 bra BB0_243;
|
|
|
|
cvt.s64.s32 %rd231, %r286;
|
|
mov.u64 %rd235, globalLights;
|
|
cvta.global.u64 %rd230, %rd235;
|
|
mov.u32 %r809, 1;
|
|
mov.u32 %r810, 48;
|
|
// inline asm
|
|
call (%rd229), _rt_buffer_get_64, (%rd230, %r809, %r810, %rd231, %rd226, %rd226, %rd226);
|
|
// inline asm
|
|
ld.v4.f32 {%f1268, %f1269, %f1270, %f1271}, [%rd229+16];
|
|
ld.v4.f32 {%f1272, %f1273, %f1274, %f1275}, [%rd229];
|
|
setp.neu.f32 %p168, %f1268, 0f00000000;
|
|
setp.leu.f32 %p169, %f63, 0f00000000;
|
|
and.pred %p170, %p168, %p169;
|
|
@%p170 bra BB0_243;
|
|
|
|
mov.f32 %f1419, 0f3F800000;
|
|
setp.leu.f32 %p171, %f1273, 0f00000000;
|
|
mov.f32 %f1420, %f1419;
|
|
mov.f32 %f1421, %f1419;
|
|
@%p171 bra BB0_242;
|
|
|
|
cvt.rzi.s32.f32 %r811, %f1274;
|
|
neg.f32 %f1285, %f62;
|
|
neg.f32 %f1284, %f63;
|
|
neg.f32 %f1283, %f64;
|
|
mov.u32 %r812, 6;
|
|
mov.u32 %r813, 0;
|
|
// inline asm
|
|
call (%f1421, %f1420, %f1419, %f1282), _rt_texture_get_base_id, (%r811, %r812, %f1283, %f1284, %f1285, %r813);
|
|
// inline asm
|
|
|
|
BB0_242:
|
|
mul.f32 %f1422, %f1269, %f1421;
|
|
mul.f32 %f1423, %f1270, %f1420;
|
|
mul.f32 %f1424, %f1271, %f1419;
|
|
|
|
BB0_243:
|
|
mov.u32 %r861, 12;
|
|
mov.u32 %r860, 2;
|
|
ld.global.v2.u32 {%r818, %r819}, [pixelID];
|
|
cvt.u64.u32 %rd238, %r818;
|
|
cvt.u64.u32 %rd239, %r819;
|
|
// inline asm
|
|
call (%rd236), _rt_buffer_get_64, (%rd222, %r860, %r861, %rd238, %rd239, %rd226, %rd226);
|
|
// inline asm
|
|
st.f32 [%rd236+8], %f1424;
|
|
st.f32 [%rd236+4], %f1423;
|
|
st.f32 [%rd236], %f1422;
|
|
ld.global.v2.u32 {%r822, %r823}, [pixelID];
|
|
cvt.u64.u32 %rd244, %r822;
|
|
cvt.u64.u32 %rd245, %r823;
|
|
mov.u64 %rd249, accumBufferLerp;
|
|
cvta.global.u64 %rd243, %rd249;
|
|
// inline asm
|
|
call (%rd242), _rt_buffer_get_64, (%rd243, %r860, %r861, %rd244, %rd245, %rd226, %rd226);
|
|
// inline asm
|
|
ld.f32 %f1286, [%rd242+8];
|
|
ld.f32 %f1287, [%rd242+4];
|
|
ld.f32 %f1288, [%rd242];
|
|
sub.f32 %f1289, %f1422, %f1288;
|
|
sub.f32 %f1290, %f1423, %f1287;
|
|
sub.f32 %f1291, %f1424, %f1286;
|
|
ld.global.f32 %f1292, [integration];
|
|
fma.rn.f32 %f419, %f1292, %f1289, %f1288;
|
|
fma.rn.f32 %f420, %f1292, %f1290, %f1287;
|
|
fma.rn.f32 %f421, %f1292, %f1291, %f1286;
|
|
ld.global.u32 %r826, [finalBounceFlag];
|
|
setp.eq.s32 %p172, %r826, 0;
|
|
@%p172 bra BB0_245;
|
|
|
|
mov.u32 %r863, 12;
|
|
mov.u32 %r862, 2;
|
|
ld.global.v2.u32 {%r829, %r830}, [pixelID];
|
|
cvt.u64.u32 %rd252, %r829;
|
|
cvt.u64.u32 %rd253, %r830;
|
|
// inline asm
|
|
call (%rd250), _rt_buffer_get_64, (%rd243, %r862, %r863, %rd252, %rd253, %rd226, %rd226);
|
|
// inline asm
|
|
st.f32 [%rd250+8], %f421;
|
|
st.f32 [%rd250+4], %f420;
|
|
st.f32 [%rd250], %f419;
|
|
|
|
BB0_245:
|
|
mov.u32 %r864, 2;
|
|
ld.global.v2.u32 {%r835, %r836}, [pixelID];
|
|
cvt.u64.u32 %rd259, %r835;
|
|
cvt.u64.u32 %rd260, %r836;
|
|
mov.u64 %rd263, accumBufferLerpAO;
|
|
cvta.global.u64 %rd258, %rd263;
|
|
mov.u32 %r834, 4;
|
|
// inline asm
|
|
call (%rd257), _rt_buffer_get_64, (%rd258, %r864, %r834, %rd259, %rd260, %rd226, %rd226);
|
|
// inline asm
|
|
ld.f32 %f1425, [%rd257];
|
|
ld.global.u32 %r839, [bounceFlag];
|
|
setp.ne.s32 %p173, %r839, 0;
|
|
@%p173 bra BB0_247;
|
|
|
|
mov.u32 %r866, 2;
|
|
ld.global.f32 %f1293, [integration];
|
|
sub.f32 %f1294, %f1418, %f1425;
|
|
fma.rn.f32 %f1425, %f1294, %f1293, %f1425;
|
|
ld.global.v2.u32 {%r842, %r843}, [pixelID];
|
|
cvt.u64.u32 %rd266, %r842;
|
|
cvt.u64.u32 %rd267, %r843;
|
|
// inline asm
|
|
call (%rd264), _rt_buffer_get_64, (%rd258, %r866, %r834, %rd266, %rd267, %rd226, %rd226);
|
|
// inline asm
|
|
st.f32 [%rd264], %f1425;
|
|
|
|
BB0_247:
|
|
mov.u32 %r865, 2;
|
|
ld.global.f32 %f1295, [aoIntensity];
|
|
mul.f32 %f1296, %f1295, 0f3F000000;
|
|
add.f32 %f1297, %f1425, 0fBF800000;
|
|
fma.rn.f32 %f1298, %f1297, %f1296, 0f3F800000;
|
|
cvt.sat.f32.f32 %f1299, %f1298;
|
|
mul.f32 %f1300, %f419, %f1299;
|
|
mul.f32 %f1301, %f420, %f1299;
|
|
mul.f32 %f1302, %f421, %f1299;
|
|
ld.global.u32 %r848, [resolution+4];
|
|
add.s32 %r849, %r848, -1;
|
|
ld.global.v2.u32 {%r850, %r851}, [pixelID];
|
|
sub.s32 %r854, %r849, %r851;
|
|
ld.global.f32 %f1303, [cameraExp];
|
|
cvt.u64.u32 %rd273, %r850;
|
|
cvt.u64.u32 %rd274, %r854;
|
|
mov.u64 %rd277, colorOutput;
|
|
cvta.global.u64 %rd272, %rd277;
|
|
mov.u32 %r847, 16;
|
|
// inline asm
|
|
call (%rd271), _rt_buffer_get_64, (%rd272, %r865, %r847, %rd273, %rd274, %rd226, %rd226);
|
|
// inline asm
|
|
mul.f32 %f1304, %f1302, %f1303;
|
|
mul.f32 %f1305, %f1301, %f1303;
|
|
mul.f32 %f1306, %f1300, %f1303;
|
|
mov.f32 %f1307, 0f3F800000;
|
|
st.v4.f32 [%rd271], {%f1306, %f1305, %f1304, %f1307};
|
|
ret;
|
|
}
|
|
|
|
|