1889 lines
54 KiB
Plaintext
1889 lines
54 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 image[1];
|
|
.global .align 1 .b8 image_HDR[1];
|
|
.global .align 1 .b8 image_HDR2[1];
|
|
.global .align 1 .b8 image_Mask[1];
|
|
.global .align 1 .b8 image_Dir[1];
|
|
.global .align 1 .b8 uvpos[1];
|
|
.global .align 1 .b8 uvnormal[1];
|
|
.global .align 1 .b8 lightMeshBuffer[1];
|
|
.global .align 4 .u32 lightMeshBufferSize;
|
|
.global .align 4 .f32 lightInvCutoff;
|
|
.global .align 4 .f32 lightPointSize;
|
|
.global .align 4 .b8 lightColor[12];
|
|
.global .align 1 .b8 rnd_seeds[1];
|
|
.global .align 4 .u32 samples;
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo7pixelIDE[8] = {82, 97, 121, 0, 8, 0, 0, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo10resolutionE[8] = {82, 97, 121, 0, 8, 0, 0, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo6normalE[8] = {82, 97, 121, 0, 12, 0, 0, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo6camPosE[8] = {82, 97, 121, 0, 12, 0, 0, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo4rootE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo12imageEnabledE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo8tileInfoE[8] = {82, 97, 121, 0, 16, 0, 0, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo8additiveE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo19lightMeshBufferSizeE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo14lightInvCutoffE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo14lightPointSizeE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo10lightColorE[8] = {82, 97, 121, 0, 12, 0, 0, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo7samplesE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
|
|
.global .align 8 .u64 _ZN21rti_internal_register20reg_bitness_detectorE;
|
|
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail0E;
|
|
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail1E;
|
|
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail2E;
|
|
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail3E;
|
|
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail4E;
|
|
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail5E;
|
|
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail6E;
|
|
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail7E;
|
|
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail8E;
|
|
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail9E;
|
|
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail0E;
|
|
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail1E;
|
|
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail2E;
|
|
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail3E;
|
|
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail4E;
|
|
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail5E;
|
|
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail6E;
|
|
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail7E;
|
|
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail8E;
|
|
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail9E;
|
|
.global .align 4 .u32 _ZN21rti_internal_register14reg_rayIndex_xE;
|
|
.global .align 4 .u32 _ZN21rti_internal_register14reg_rayIndex_yE;
|
|
.global .align 4 .u32 _ZN21rti_internal_register14reg_rayIndex_zE;
|
|
.global .align 8 .b8 _ZN21rti_internal_typename7pixelIDE[6] = {117, 105, 110, 116, 50, 0};
|
|
.global .align 8 .b8 _ZN21rti_internal_typename10resolutionE[6] = {117, 105, 110, 116, 50, 0};
|
|
.global .align 8 .b8 _ZN21rti_internal_typename6normalE[7] = {102, 108, 111, 97, 116, 51, 0};
|
|
.global .align 8 .b8 _ZN21rti_internal_typename6camPosE[7] = {102, 108, 111, 97, 116, 51, 0};
|
|
.global .align 16 .b8 _ZN21rti_internal_typename4rootE[9] = {114, 116, 79, 98, 106, 101, 99, 116, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typename12imageEnabledE[4] = {105, 110, 116, 0};
|
|
.global .align 8 .b8 _ZN21rti_internal_typename8tileInfoE[6] = {117, 105, 110, 116, 52, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typename8additiveE[4] = {105, 110, 116, 0};
|
|
.global .align 16 .b8 _ZN21rti_internal_typename19lightMeshBufferSizeE[13] = {117, 110, 115, 105, 103, 110, 101, 100, 32, 105, 110, 116, 0};
|
|
.global .align 8 .b8 _ZN21rti_internal_typename14lightInvCutoffE[6] = {102, 108, 111, 97, 116, 0};
|
|
.global .align 8 .b8 _ZN21rti_internal_typename14lightPointSizeE[6] = {102, 108, 111, 97, 116, 0};
|
|
.global .align 8 .b8 _ZN21rti_internal_typename10lightColorE[7] = {102, 108, 111, 97, 116, 51, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typename7samplesE[4] = {105, 110, 116, 0};
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum7pixelIDE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum10resolutionE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum6normalE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum6camPosE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum4rootE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum12imageEnabledE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum8tileInfoE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum8additiveE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum19lightMeshBufferSizeE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum14lightInvCutoffE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum14lightPointSizeE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum10lightColorE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum7samplesE = 4919;
|
|
.global .align 16 .b8 _ZN21rti_internal_semantic7pixelIDE[14] = {114, 116, 76, 97, 117, 110, 99, 104, 73, 110, 100, 101, 120, 0};
|
|
.global .align 16 .b8 _ZN21rti_internal_semantic10resolutionE[12] = {114, 116, 76, 97, 117, 110, 99, 104, 68, 105, 109, 0};
|
|
.global .align 16 .b8 _ZN21rti_internal_semantic6normalE[17] = {97, 116, 116, 114, 105, 98, 117, 116, 101, 32, 110, 111, 114, 109, 97, 108, 0};
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic6camPosE[1];
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic4rootE[1];
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic12imageEnabledE[1];
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic8tileInfoE[1];
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic8additiveE[1];
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic19lightMeshBufferSizeE[1];
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic14lightInvCutoffE[1];
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic14lightPointSizeE[1];
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic10lightColorE[1];
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic7samplesE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation7pixelIDE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation10resolutionE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation6normalE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation6camPosE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation4rootE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation12imageEnabledE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation8tileInfoE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation8additiveE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation19lightMeshBufferSizeE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation14lightInvCutoffE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation14lightPointSizeE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation10lightColorE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation7samplesE[1];
|
|
.const .align 4 .b8 __cudart_i2opi_f[24] = {65, 144, 67, 60, 153, 149, 98, 219, 192, 221, 52, 245, 209, 87, 39, 252, 41, 21, 68, 78, 110, 131, 249, 162};
|
|
|
|
.visible .entry _Z6oxMainv(
|
|
|
|
)
|
|
{
|
|
.local .align 4 .b8 __local_depot0[36];
|
|
.reg .b64 %SP;
|
|
.reg .b64 %SPL;
|
|
.reg .pred %p<145>;
|
|
.reg .b16 %rs<55>;
|
|
.reg .f32 %f<935>;
|
|
.reg .b32 %r<371>;
|
|
.reg .b64 %rd<162>;
|
|
|
|
|
|
mov.u64 %rd161, __local_depot0;
|
|
cvta.local.u64 %SP, %rd161;
|
|
ld.global.u32 %r1, [samples];
|
|
ld.global.v2.u32 {%r103, %r104}, [pixelID];
|
|
cvt.u64.u32 %rd23, %r103;
|
|
cvt.u64.u32 %rd24, %r104;
|
|
mov.u64 %rd27, uvnormal;
|
|
cvta.global.u64 %rd22, %rd27;
|
|
mov.u32 %r101, 2;
|
|
mov.u32 %r102, 4;
|
|
mov.u64 %rd26, 0;
|
|
// inline asm
|
|
call (%rd21), _rt_buffer_get_64, (%rd22, %r101, %r102, %rd23, %rd24, %rd26, %rd26);
|
|
// inline asm
|
|
ld.u32 %r2, [%rd21];
|
|
shr.u32 %r107, %r2, 16;
|
|
cvt.u16.u32 %rs1, %r107;
|
|
and.b16 %rs3, %rs1, 255;
|
|
cvt.u16.u32 %rs4, %r2;
|
|
or.b16 %rs5, %rs4, %rs3;
|
|
setp.eq.s16 %p7, %rs5, 0;
|
|
mov.f32 %f865, 0f00000000;
|
|
mov.f32 %f866, %f865;
|
|
mov.f32 %f867, %f865;
|
|
@%p7 bra BB0_2;
|
|
|
|
ld.u8 %rs6, [%rd21+1];
|
|
and.b16 %rs8, %rs4, 255;
|
|
cvt.rn.f32.u16 %f204, %rs8;
|
|
div.rn.f32 %f205, %f204, 0f437F0000;
|
|
fma.rn.f32 %f206, %f205, 0f40000000, 0fBF800000;
|
|
cvt.rn.f32.u16 %f207, %rs6;
|
|
div.rn.f32 %f208, %f207, 0f437F0000;
|
|
fma.rn.f32 %f209, %f208, 0f40000000, 0fBF800000;
|
|
cvt.rn.f32.u16 %f210, %rs3;
|
|
div.rn.f32 %f211, %f210, 0f437F0000;
|
|
fma.rn.f32 %f212, %f211, 0f40000000, 0fBF800000;
|
|
mul.f32 %f213, %f209, %f209;
|
|
fma.rn.f32 %f214, %f206, %f206, %f213;
|
|
fma.rn.f32 %f215, %f212, %f212, %f214;
|
|
sqrt.rn.f32 %f216, %f215;
|
|
rcp.rn.f32 %f217, %f216;
|
|
mul.f32 %f865, %f206, %f217;
|
|
mul.f32 %f866, %f209, %f217;
|
|
mul.f32 %f867, %f212, %f217;
|
|
|
|
BB0_2:
|
|
ld.global.v2.u32 {%r108, %r109}, [pixelID];
|
|
ld.global.v2.u32 {%r111, %r112}, [tileInfo];
|
|
add.s32 %r3, %r108, %r111;
|
|
add.s32 %r4, %r109, %r112;
|
|
setp.eq.f32 %p8, %f866, 0f00000000;
|
|
setp.eq.f32 %p9, %f865, 0f00000000;
|
|
and.pred %p10, %p9, %p8;
|
|
setp.eq.f32 %p11, %f867, 0f00000000;
|
|
and.pred %p12, %p10, %p11;
|
|
@%p12 bra BB0_131;
|
|
bra.uni BB0_3;
|
|
|
|
BB0_131:
|
|
ld.global.u32 %r370, [imageEnabled];
|
|
and.b32 %r313, %r370, 1;
|
|
setp.eq.b32 %p139, %r313, 1;
|
|
@!%p139 bra BB0_133;
|
|
bra.uni BB0_132;
|
|
|
|
BB0_132:
|
|
cvt.u64.u32 %rd116, %r3;
|
|
cvt.u64.u32 %rd117, %r4;
|
|
mov.u64 %rd120, image;
|
|
cvta.global.u64 %rd115, %rd120;
|
|
mov.u64 %rd119, 0;
|
|
// inline asm
|
|
call (%rd114), _rt_buffer_get_64, (%rd115, %r101, %r102, %rd116, %rd117, %rd119, %rd119);
|
|
// inline asm
|
|
mov.u16 %rs36, 0;
|
|
st.v4.u8 [%rd114], {%rs36, %rs36, %rs36, %rs36};
|
|
ld.global.u32 %r370, [imageEnabled];
|
|
|
|
BB0_133:
|
|
and.b32 %r316, %r370, 8;
|
|
setp.eq.s32 %p140, %r316, 0;
|
|
@%p140 bra BB0_135;
|
|
|
|
cvt.u64.u32 %rd123, %r3;
|
|
cvt.u64.u32 %rd124, %r4;
|
|
mov.u64 %rd127, image_Mask;
|
|
cvta.global.u64 %rd122, %rd127;
|
|
mov.u64 %rd126, 0;
|
|
// inline asm
|
|
call (%rd121), _rt_buffer_get_64, (%rd122, %r101, %r101, %rd123, %rd124, %rd126, %rd126);
|
|
// inline asm
|
|
mov.f32 %f848, 0f00000000;
|
|
cvt.rzi.u32.f32 %r319, %f848;
|
|
cvt.u16.u32 %rs37, %r319;
|
|
mov.u16 %rs38, 0;
|
|
st.v2.u8 [%rd121], {%rs37, %rs38};
|
|
ld.global.u32 %r370, [imageEnabled];
|
|
|
|
BB0_135:
|
|
and.b32 %r320, %r370, 4;
|
|
setp.eq.s32 %p141, %r320, 0;
|
|
@%p141 bra BB0_139;
|
|
|
|
ld.global.u32 %r321, [additive];
|
|
setp.eq.s32 %p142, %r321, 0;
|
|
cvt.u64.u32 %rd19, %r3;
|
|
cvt.u64.u32 %rd20, %r4;
|
|
@%p142 bra BB0_138;
|
|
|
|
mov.u64 %rd140, image_HDR;
|
|
cvta.global.u64 %rd129, %rd140;
|
|
mov.u32 %r325, 8;
|
|
mov.u64 %rd139, 0;
|
|
// inline asm
|
|
call (%rd128), _rt_buffer_get_64, (%rd129, %r101, %r325, %rd19, %rd20, %rd139, %rd139);
|
|
// inline asm
|
|
ld.v4.u16 {%rs45, %rs46, %rs47, %rs48}, [%rd128];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f849, %rs45;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f850, %rs46;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f851, %rs47;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd134), _rt_buffer_get_64, (%rd129, %r101, %r325, %rd19, %rd20, %rd139, %rd139);
|
|
// inline asm
|
|
add.f32 %f852, %f849, 0f00000000;
|
|
add.f32 %f853, %f850, 0f00000000;
|
|
add.f32 %f854, %f851, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs44, %f854;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs43, %f853;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs42, %f852;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs49, 0;
|
|
st.v4.u16 [%rd134], {%rs42, %rs43, %rs44, %rs49};
|
|
bra.uni BB0_139;
|
|
|
|
BB0_3:
|
|
ld.global.v2.u32 {%r121, %r122}, [pixelID];
|
|
cvt.u64.u32 %rd30, %r121;
|
|
cvt.u64.u32 %rd31, %r122;
|
|
mov.u64 %rd40, uvpos;
|
|
cvta.global.u64 %rd29, %rd40;
|
|
mov.u32 %r117, 12;
|
|
// inline asm
|
|
call (%rd28), _rt_buffer_get_64, (%rd29, %r101, %r117, %rd30, %rd31, %rd26, %rd26);
|
|
// inline asm
|
|
ld.f32 %f9, [%rd28+8];
|
|
ld.f32 %f8, [%rd28+4];
|
|
ld.f32 %f7, [%rd28];
|
|
mul.f32 %f223, %f7, 0f3456BF95;
|
|
mul.f32 %f224, %f8, 0f3456BF95;
|
|
mul.f32 %f225, %f9, 0f3456BF95;
|
|
abs.f32 %f10, %f865;
|
|
div.rn.f32 %f226, %f223, %f10;
|
|
abs.f32 %f227, %f866;
|
|
div.rn.f32 %f228, %f224, %f227;
|
|
abs.f32 %f11, %f867;
|
|
div.rn.f32 %f229, %f225, %f11;
|
|
abs.f32 %f230, %f226;
|
|
abs.f32 %f231, %f228;
|
|
abs.f32 %f232, %f229;
|
|
mov.f32 %f233, 0f38D1B717;
|
|
max.f32 %f234, %f230, %f233;
|
|
max.f32 %f235, %f231, %f233;
|
|
max.f32 %f236, %f232, %f233;
|
|
fma.rn.f32 %f12, %f865, %f234, %f7;
|
|
fma.rn.f32 %f13, %f866, %f235, %f8;
|
|
fma.rn.f32 %f14, %f867, %f236, %f9;
|
|
ld.global.v2.u32 {%r125, %r126}, [pixelID];
|
|
cvt.u64.u32 %rd36, %r125;
|
|
cvt.u64.u32 %rd37, %r126;
|
|
mov.u64 %rd41, rnd_seeds;
|
|
cvta.global.u64 %rd35, %rd41;
|
|
// inline asm
|
|
call (%rd34), _rt_buffer_get_64, (%rd35, %r101, %r102, %rd36, %rd37, %rd26, %rd26);
|
|
// inline asm
|
|
ld.u32 %r129, [%rd34];
|
|
mad.lo.s32 %r5, %r129, 1664525, 1013904223;
|
|
ld.global.u32 %r130, [lightMeshBufferSize];
|
|
setp.eq.s32 %p14, %r130, 0;
|
|
mov.pred %p13, 0;
|
|
mov.f32 %f20, 0f00000000;
|
|
mov.u32 %r7, 0;
|
|
@%p14 bra BB0_4;
|
|
|
|
ld.global.f32 %f15, [lightPointSize];
|
|
mul.f32 %f16, %f12, 0f3456BF95;
|
|
mul.f32 %f17, %f13, 0f3456BF95;
|
|
mul.f32 %f18, %f14, 0f3456BF95;
|
|
and.b32 %r133, %r5, 16777215;
|
|
cvt.rn.f32.u32 %f242, %r133;
|
|
mul.f32 %f243, %f242, 0fB3800000;
|
|
fma.rn.f32 %f19, %f243, 0f3F333333, 0f3F800000;
|
|
mov.f32 %f20, 0f00000000;
|
|
mov.u32 %r339, 0;
|
|
abs.f32 %f366, %f17;
|
|
abs.f32 %f367, %f16;
|
|
max.f32 %f368, %f367, %f366;
|
|
abs.f32 %f369, %f18;
|
|
max.f32 %f370, %f368, %f369;
|
|
mov.u32 %r7, %r339;
|
|
mov.f32 %f21, %f20;
|
|
mov.f32 %f888, %f20;
|
|
mov.f32 %f889, %f20;
|
|
mov.f32 %f890, %f20;
|
|
|
|
BB0_6:
|
|
shl.b32 %r8, %r339, 1;
|
|
cvt.s64.s32 %rd44, %r8;
|
|
mov.u64 %rd48, lightMeshBuffer;
|
|
cvta.global.u64 %rd43, %rd48;
|
|
mov.u32 %r134, 1;
|
|
// inline asm
|
|
call (%rd42), _rt_buffer_get_64, (%rd43, %r134, %r117, %rd44, %rd26, %rd26, %rd26);
|
|
// inline asm
|
|
ld.f32 %f244, [%rd42];
|
|
sub.f32 %f245, %f244, %f7;
|
|
ld.f32 %f246, [%rd42+4];
|
|
sub.f32 %f247, %f246, %f8;
|
|
ld.f32 %f248, [%rd42+8];
|
|
sub.f32 %f249, %f248, %f9;
|
|
mul.f32 %f250, %f247, %f247;
|
|
fma.rn.f32 %f251, %f245, %f245, %f250;
|
|
fma.rn.f32 %f252, %f249, %f249, %f251;
|
|
sqrt.rn.f32 %f25, %f252;
|
|
rcp.rn.f32 %f253, %f25;
|
|
mul.f32 %f26, %f245, %f253;
|
|
mul.f32 %f27, %f247, %f253;
|
|
mul.f32 %f28, %f249, %f253;
|
|
mul.f32 %f254, %f866, %f27;
|
|
fma.rn.f32 %f255, %f865, %f26, %f254;
|
|
fma.rn.f32 %f29, %f867, %f28, %f255;
|
|
setp.leu.f32 %p15, %f29, 0f00000000;
|
|
@%p15 bra BB0_22;
|
|
|
|
setp.ne.s32 %p17, %r1, 0;
|
|
mul.f32 %f256, %f25, %f25;
|
|
mul.f32 %f257, %f256, 0f40C90FDB;
|
|
div.rn.f32 %f258, %f15, %f257;
|
|
add.f32 %f30, %f258, %f258;
|
|
setp.gt.f32 %p18, %f30, %f19;
|
|
and.pred %p19, %p17, %p18;
|
|
mov.pred %p144, -1;
|
|
@%p19 bra BB0_24;
|
|
|
|
ld.global.f32 %f261, [lightInvCutoff];
|
|
mul.f32 %f31, %f25, %f261;
|
|
mov.f32 %f265, 0f40800000;
|
|
abs.f32 %f33, %f31;
|
|
setp.lt.f32 %p20, %f33, 0f00800000;
|
|
mul.f32 %f267, %f33, 0f4B800000;
|
|
selp.f32 %f268, 0fC3170000, 0fC2FE0000, %p20;
|
|
selp.f32 %f269, %f267, %f33, %p20;
|
|
mov.b32 %r136, %f269;
|
|
and.b32 %r137, %r136, 8388607;
|
|
or.b32 %r138, %r137, 1065353216;
|
|
mov.b32 %f270, %r138;
|
|
shr.u32 %r139, %r136, 23;
|
|
cvt.rn.f32.u32 %f271, %r139;
|
|
add.f32 %f272, %f268, %f271;
|
|
setp.gt.f32 %p21, %f270, 0f3FB504F3;
|
|
mul.f32 %f273, %f270, 0f3F000000;
|
|
add.f32 %f274, %f272, 0f3F800000;
|
|
selp.f32 %f275, %f273, %f270, %p21;
|
|
selp.f32 %f276, %f274, %f272, %p21;
|
|
add.f32 %f277, %f275, 0fBF800000;
|
|
add.f32 %f260, %f275, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f259,%f260;
|
|
// inline asm
|
|
add.f32 %f278, %f277, %f277;
|
|
mul.f32 %f279, %f259, %f278;
|
|
mul.f32 %f280, %f279, %f279;
|
|
mov.f32 %f281, 0f3C4CAF63;
|
|
mov.f32 %f282, 0f3B18F0FE;
|
|
fma.rn.f32 %f283, %f282, %f280, %f281;
|
|
mov.f32 %f284, 0f3DAAAABD;
|
|
fma.rn.f32 %f285, %f283, %f280, %f284;
|
|
mul.rn.f32 %f286, %f285, %f280;
|
|
mul.rn.f32 %f287, %f286, %f279;
|
|
sub.f32 %f288, %f277, %f279;
|
|
neg.f32 %f289, %f279;
|
|
add.f32 %f290, %f288, %f288;
|
|
fma.rn.f32 %f291, %f289, %f277, %f290;
|
|
mul.rn.f32 %f292, %f259, %f291;
|
|
add.f32 %f293, %f287, %f279;
|
|
sub.f32 %f294, %f279, %f293;
|
|
add.f32 %f295, %f287, %f294;
|
|
add.f32 %f296, %f292, %f295;
|
|
add.f32 %f297, %f293, %f296;
|
|
sub.f32 %f298, %f293, %f297;
|
|
add.f32 %f299, %f296, %f298;
|
|
mov.f32 %f300, 0f3F317200;
|
|
mul.rn.f32 %f301, %f276, %f300;
|
|
mov.f32 %f302, 0f35BFBE8E;
|
|
mul.rn.f32 %f303, %f276, %f302;
|
|
add.f32 %f304, %f301, %f297;
|
|
sub.f32 %f305, %f301, %f304;
|
|
add.f32 %f306, %f297, %f305;
|
|
add.f32 %f307, %f299, %f306;
|
|
add.f32 %f308, %f303, %f307;
|
|
add.f32 %f309, %f304, %f308;
|
|
sub.f32 %f310, %f304, %f309;
|
|
add.f32 %f311, %f308, %f310;
|
|
mul.rn.f32 %f312, %f265, %f309;
|
|
neg.f32 %f313, %f312;
|
|
fma.rn.f32 %f314, %f265, %f309, %f313;
|
|
fma.rn.f32 %f315, %f265, %f311, %f314;
|
|
mov.f32 %f316, 0f00000000;
|
|
fma.rn.f32 %f317, %f316, %f309, %f315;
|
|
add.rn.f32 %f318, %f312, %f317;
|
|
neg.f32 %f319, %f318;
|
|
add.rn.f32 %f320, %f312, %f319;
|
|
add.rn.f32 %f321, %f320, %f317;
|
|
mov.b32 %r140, %f318;
|
|
setp.eq.s32 %p22, %r140, 1118925336;
|
|
add.s32 %r141, %r140, -1;
|
|
mov.b32 %f322, %r141;
|
|
add.f32 %f323, %f321, 0f37000000;
|
|
selp.f32 %f324, %f322, %f318, %p22;
|
|
selp.f32 %f34, %f323, %f321, %p22;
|
|
mul.f32 %f325, %f324, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f326, %f325;
|
|
mov.f32 %f327, 0fBF317200;
|
|
fma.rn.f32 %f328, %f326, %f327, %f324;
|
|
mov.f32 %f329, 0fB5BFBE8E;
|
|
fma.rn.f32 %f330, %f326, %f329, %f328;
|
|
mul.f32 %f331, %f330, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f332, %f331;
|
|
add.f32 %f333, %f326, 0f00000000;
|
|
ex2.approx.f32 %f334, %f333;
|
|
mul.f32 %f335, %f332, %f334;
|
|
setp.lt.f32 %p23, %f324, 0fC2D20000;
|
|
selp.f32 %f336, 0f00000000, %f335, %p23;
|
|
setp.gt.f32 %p24, %f324, 0f42D20000;
|
|
selp.f32 %f873, 0f7F800000, %f336, %p24;
|
|
setp.eq.f32 %p25, %f873, 0f7F800000;
|
|
@%p25 bra BB0_10;
|
|
|
|
fma.rn.f32 %f873, %f873, %f34, %f873;
|
|
|
|
BB0_10:
|
|
mov.f32 %f861, 0f40000000;
|
|
cvt.rzi.f32.f32 %f860, %f861;
|
|
add.f32 %f859, %f860, %f860;
|
|
mov.f32 %f858, 0f40800000;
|
|
sub.f32 %f857, %f858, %f859;
|
|
abs.f32 %f856, %f857;
|
|
setp.lt.f32 %p26, %f31, 0f00000000;
|
|
setp.eq.f32 %p27, %f856, 0f3F800000;
|
|
and.pred %p1, %p26, %p27;
|
|
mov.b32 %r142, %f873;
|
|
xor.b32 %r143, %r142, -2147483648;
|
|
mov.b32 %f337, %r143;
|
|
selp.f32 %f875, %f337, %f873, %p1;
|
|
setp.eq.f32 %p28, %f31, 0f00000000;
|
|
@%p28 bra BB0_13;
|
|
bra.uni BB0_11;
|
|
|
|
BB0_13:
|
|
add.f32 %f340, %f31, %f31;
|
|
selp.f32 %f875, %f340, 0f00000000, %p27;
|
|
bra.uni BB0_14;
|
|
|
|
BB0_11:
|
|
setp.geu.f32 %p29, %f31, 0f00000000;
|
|
@%p29 bra BB0_14;
|
|
|
|
mov.f32 %f864, 0f40800000;
|
|
cvt.rzi.f32.f32 %f339, %f864;
|
|
setp.neu.f32 %p30, %f339, 0f40800000;
|
|
selp.f32 %f875, 0f7FFFFFFF, %f875, %p30;
|
|
|
|
BB0_14:
|
|
add.f32 %f341, %f33, 0f40800000;
|
|
mov.b32 %r144, %f341;
|
|
setp.lt.s32 %p32, %r144, 2139095040;
|
|
@%p32 bra BB0_19;
|
|
|
|
setp.gtu.f32 %p33, %f33, 0f7F800000;
|
|
@%p33 bra BB0_18;
|
|
bra.uni BB0_16;
|
|
|
|
BB0_18:
|
|
add.f32 %f875, %f31, 0f40800000;
|
|
bra.uni BB0_19;
|
|
|
|
BB0_16:
|
|
setp.neu.f32 %p34, %f33, 0f7F800000;
|
|
@%p34 bra BB0_19;
|
|
|
|
selp.f32 %f875, 0fFF800000, 0f7F800000, %p1;
|
|
|
|
BB0_19:
|
|
mov.u32 %r331, 1;
|
|
mov.u64 %rd156, lightMeshBuffer;
|
|
cvta.global.u64 %rd155, %rd156;
|
|
shl.b32 %r330, %r339, 1;
|
|
mov.f32 %f342, 0f3F800000;
|
|
sub.f32 %f343, %f342, %f875;
|
|
setp.eq.f32 %p35, %f31, 0f3F800000;
|
|
selp.f32 %f344, 0f00000000, %f343, %p35;
|
|
cvt.sat.f32.f32 %f345, %f344;
|
|
mul.f32 %f346, %f30, %f345;
|
|
add.s32 %r147, %r330, 1;
|
|
cvt.s64.s32 %rd51, %r147;
|
|
// inline asm
|
|
call (%rd49), _rt_buffer_get_64, (%rd155, %r331, %r117, %rd51, %rd26, %rd26, %rd26);
|
|
// inline asm
|
|
ld.f32 %f347, [%rd49];
|
|
mul.f32 %f348, %f26, %f347;
|
|
ld.f32 %f349, [%rd49+4];
|
|
mul.f32 %f350, %f27, %f349;
|
|
neg.f32 %f351, %f350;
|
|
sub.f32 %f352, %f351, %f348;
|
|
ld.f32 %f353, [%rd49+8];
|
|
mul.f32 %f354, %f28, %f353;
|
|
sub.f32 %f355, %f352, %f354;
|
|
cvt.sat.f32.f32 %f356, %f355;
|
|
mul.f32 %f45, %f346, %f356;
|
|
cvt.sat.f32.f32 %f46, %f29;
|
|
mul.f32 %f357, %f45, %f46;
|
|
setp.leu.f32 %p36, %f357, 0f3727C5AC;
|
|
@%p36 bra BB0_21;
|
|
|
|
mov.u32 %r338, 1;
|
|
add.u64 %rd56, %SP, 28;
|
|
cvta.to.local.u64 %rd57, %rd56;
|
|
max.f32 %f364, %f370, %f233;
|
|
sub.f32 %f365, %f25, %f364;
|
|
mov.u32 %r151, 1065353216;
|
|
st.local.u32 [%rd57], %r151;
|
|
ld.global.u32 %r148, [root];
|
|
// inline asm
|
|
call _rt_trace_64, (%r148, %f12, %f13, %f14, %f26, %f27, %f28, %r338, %f364, %f365, %rd56, %r102);
|
|
// inline asm
|
|
ld.local.f32 %f372, [%rd57];
|
|
mul.f32 %f373, %f45, %f372;
|
|
fma.rn.f32 %f21, %f46, %f373, %f21;
|
|
add.f32 %f20, %f20, %f372;
|
|
fma.rn.f32 %f890, %f26, %f372, %f890;
|
|
fma.rn.f32 %f889, %f27, %f372, %f889;
|
|
fma.rn.f32 %f888, %f28, %f372, %f888;
|
|
|
|
BB0_21:
|
|
add.s32 %r7, %r7, 1;
|
|
|
|
BB0_22:
|
|
ld.global.u32 %r152, [lightMeshBufferSize];
|
|
add.s32 %r339, %r339, 1;
|
|
setp.lt.u32 %p38, %r339, %r152;
|
|
@%p38 bra BB0_6;
|
|
bra.uni BB0_23;
|
|
|
|
BB0_4:
|
|
mov.f32 %f21, %f20;
|
|
mov.f32 %f888, %f20;
|
|
mov.f32 %f889, %f20;
|
|
mov.f32 %f890, %f20;
|
|
|
|
BB0_23:
|
|
mov.pred %p144, %p13;
|
|
|
|
BB0_24:
|
|
cvt.rn.f32.s32 %f374, %r7;
|
|
mov.f32 %f375, 0f3F800000;
|
|
max.f32 %f376, %f374, %f375;
|
|
div.rn.f32 %f921, %f21, %f376;
|
|
div.rn.f32 %f922, %f20, %f376;
|
|
@!%p144 bra BB0_77;
|
|
bra.uni BB0_25;
|
|
|
|
BB0_25:
|
|
abs.f32 %f863, %f867;
|
|
abs.f32 %f862, %f865;
|
|
setp.gt.f32 %p39, %f862, %f863;
|
|
neg.f32 %f382, %f866;
|
|
selp.f32 %f383, %f382, 0f00000000, %p39;
|
|
neg.f32 %f384, %f867;
|
|
selp.f32 %f385, %f865, %f384, %p39;
|
|
selp.f32 %f386, 0f00000000, %f866, %p39;
|
|
mul.f32 %f387, %f385, %f385;
|
|
fma.rn.f32 %f388, %f383, %f383, %f387;
|
|
fma.rn.f32 %f389, %f386, %f386, %f388;
|
|
sqrt.rn.f32 %f390, %f389;
|
|
rcp.rn.f32 %f391, %f390;
|
|
mul.f32 %f69, %f383, %f391;
|
|
mul.f32 %f70, %f385, %f391;
|
|
mul.f32 %f71, %f386, %f391;
|
|
mov.f32 %f896, 0f00000000;
|
|
setp.lt.s32 %p40, %r1, 1;
|
|
mov.f32 %f897, %f896;
|
|
mov.f32 %f888, %f896;
|
|
mov.f32 %f889, %f896;
|
|
mov.f32 %f890, %f896;
|
|
@%p40 bra BB0_76;
|
|
|
|
mad.lo.s32 %r346, %r129, 1664525, 1013904223;
|
|
cvt.rn.f32.s32 %f397, %r1;
|
|
rcp.rn.f32 %f72, %f397;
|
|
add.u64 %rd58, %SP, 0;
|
|
cvta.to.local.u64 %rd2, %rd58;
|
|
mul.f32 %f73, %f12, 0f3456BF95;
|
|
mul.f32 %f74, %f13, 0f3456BF95;
|
|
mul.f32 %f75, %f14, 0f3456BF95;
|
|
add.u64 %rd59, %SP, 32;
|
|
cvta.to.local.u64 %rd3, %rd59;
|
|
mul.f32 %f398, %f865, %f70;
|
|
mul.f32 %f399, %f866, %f69;
|
|
sub.f32 %f76, %f399, %f398;
|
|
mul.f32 %f400, %f867, %f69;
|
|
mul.f32 %f401, %f865, %f71;
|
|
sub.f32 %f77, %f401, %f400;
|
|
mul.f32 %f402, %f866, %f71;
|
|
mul.f32 %f403, %f867, %f70;
|
|
sub.f32 %f78, %f403, %f402;
|
|
mov.f32 %f896, 0f00000000;
|
|
mov.u32 %r153, 0;
|
|
abs.f32 %f404, %f74;
|
|
abs.f32 %f405, %f73;
|
|
max.f32 %f406, %f405, %f404;
|
|
abs.f32 %f407, %f75;
|
|
max.f32 %f408, %f406, %f407;
|
|
mov.u32 %r343, %r153;
|
|
mov.f32 %f897, %f896;
|
|
mov.f32 %f888, %f896;
|
|
mov.f32 %f889, %f896;
|
|
mov.f32 %f890, %f896;
|
|
|
|
BB0_27:
|
|
cvt.rn.f32.s32 %f84, %r343;
|
|
max.f32 %f85, %f408, %f233;
|
|
mov.u32 %r345, %r153;
|
|
|
|
BB0_28:
|
|
mad.lo.s32 %r155, %r346, 1664525, 1013904223;
|
|
and.b32 %r156, %r155, 16777215;
|
|
cvt.rn.f32.u32 %f410, %r156;
|
|
fma.rn.f32 %f411, %f410, 0f33800000, %f84;
|
|
mul.f32 %f91, %f72, %f411;
|
|
mad.lo.s32 %r346, %r155, 1664525, 1013904223;
|
|
and.b32 %r157, %r346, 16777215;
|
|
cvt.rn.f32.u32 %f412, %r157;
|
|
cvt.rn.f32.s32 %f413, %r345;
|
|
fma.rn.f32 %f414, %f412, 0f33800000, %f413;
|
|
mul.f32 %f415, %f72, %f414;
|
|
mul.f32 %f416, %f91, %f91;
|
|
sub.f32 %f418, %f375, %f416;
|
|
mov.f32 %f419, 0f00000000;
|
|
max.f32 %f420, %f419, %f418;
|
|
sqrt.rn.f32 %f92, %f420;
|
|
mul.f32 %f907, %f415, 0f40C90FDB;
|
|
abs.f32 %f94, %f907;
|
|
setp.neu.f32 %p41, %f94, 0f7F800000;
|
|
mov.f32 %f901, %f907;
|
|
@%p41 bra BB0_30;
|
|
|
|
mul.rn.f32 %f901, %f907, %f419;
|
|
|
|
BB0_30:
|
|
mul.f32 %f422, %f901, 0f3F22F983;
|
|
cvt.rni.s32.f32 %r356, %f422;
|
|
cvt.rn.f32.s32 %f423, %r356;
|
|
neg.f32 %f424, %f423;
|
|
mov.f32 %f425, 0f3FC90FDA;
|
|
fma.rn.f32 %f426, %f424, %f425, %f901;
|
|
mov.f32 %f427, 0f33A22168;
|
|
fma.rn.f32 %f428, %f424, %f427, %f426;
|
|
mov.f32 %f429, 0f27C234C5;
|
|
fma.rn.f32 %f902, %f424, %f429, %f428;
|
|
abs.f32 %f430, %f901;
|
|
setp.leu.f32 %p42, %f430, 0f47CE4780;
|
|
@%p42 bra BB0_41;
|
|
|
|
mov.b32 %r19, %f901;
|
|
shr.u32 %r20, %r19, 23;
|
|
shl.b32 %r160, %r19, 8;
|
|
or.b32 %r21, %r160, -2147483648;
|
|
mov.u32 %r347, 0;
|
|
mov.u64 %rd158, 0;
|
|
mov.u64 %rd157, %rd2;
|
|
mov.u32 %r348, %r347;
|
|
|
|
BB0_32:
|
|
.pragma "nounroll";
|
|
shl.b64 %rd61, %rd158, 2;
|
|
mov.u64 %rd62, __cudart_i2opi_f;
|
|
add.s64 %rd63, %rd62, %rd61;
|
|
ld.const.u32 %r163, [%rd63];
|
|
// inline asm
|
|
{
|
|
mad.lo.cc.u32 %r161, %r163, %r21, %r348;
|
|
madc.hi.u32 %r348, %r163, %r21, 0;
|
|
}
|
|
// inline asm
|
|
st.local.u32 [%rd157], %r161;
|
|
add.s32 %r347, %r347, 1;
|
|
cvt.s64.s32 %rd158, %r347;
|
|
mul.wide.s32 %rd64, %r347, 4;
|
|
add.s64 %rd157, %rd2, %rd64;
|
|
setp.ne.s32 %p43, %r347, 6;
|
|
@%p43 bra BB0_32;
|
|
|
|
and.b32 %r166, %r20, 255;
|
|
add.s32 %r167, %r166, -128;
|
|
shr.u32 %r168, %r167, 5;
|
|
and.b32 %r26, %r19, -2147483648;
|
|
st.local.u32 [%rd2+24], %r348;
|
|
mov.u32 %r169, 6;
|
|
sub.s32 %r170, %r169, %r168;
|
|
mul.wide.s32 %rd65, %r170, 4;
|
|
add.s64 %rd9, %rd2, %rd65;
|
|
ld.local.u32 %r349, [%rd9];
|
|
ld.local.u32 %r350, [%rd9+-4];
|
|
and.b32 %r29, %r20, 31;
|
|
setp.eq.s32 %p44, %r29, 0;
|
|
@%p44 bra BB0_35;
|
|
|
|
mov.u32 %r171, 32;
|
|
sub.s32 %r172, %r171, %r29;
|
|
shr.u32 %r173, %r350, %r172;
|
|
shl.b32 %r174, %r349, %r29;
|
|
add.s32 %r349, %r173, %r174;
|
|
ld.local.u32 %r175, [%rd9+-8];
|
|
shr.u32 %r176, %r175, %r172;
|
|
shl.b32 %r177, %r350, %r29;
|
|
add.s32 %r350, %r176, %r177;
|
|
|
|
BB0_35:
|
|
shr.u32 %r178, %r350, 30;
|
|
shl.b32 %r179, %r349, 2;
|
|
add.s32 %r351, %r178, %r179;
|
|
shl.b32 %r35, %r350, 2;
|
|
shr.u32 %r180, %r351, 31;
|
|
shr.u32 %r181, %r349, 30;
|
|
add.s32 %r36, %r180, %r181;
|
|
setp.eq.s32 %p45, %r180, 0;
|
|
@%p45 bra BB0_36;
|
|
bra.uni BB0_37;
|
|
|
|
BB0_36:
|
|
mov.u32 %r352, %r26;
|
|
mov.u32 %r353, %r35;
|
|
bra.uni BB0_38;
|
|
|
|
BB0_37:
|
|
not.b32 %r182, %r351;
|
|
neg.s32 %r353, %r35;
|
|
setp.eq.s32 %p46, %r35, 0;
|
|
selp.u32 %r183, 1, 0, %p46;
|
|
add.s32 %r351, %r183, %r182;
|
|
xor.b32 %r352, %r26, -2147483648;
|
|
|
|
BB0_38:
|
|
clz.b32 %r355, %r351;
|
|
setp.eq.s32 %p47, %r355, 0;
|
|
shl.b32 %r184, %r351, %r355;
|
|
mov.u32 %r185, 32;
|
|
sub.s32 %r186, %r185, %r355;
|
|
shr.u32 %r187, %r353, %r186;
|
|
add.s32 %r188, %r187, %r184;
|
|
selp.b32 %r44, %r351, %r188, %p47;
|
|
mov.u32 %r189, -921707870;
|
|
mul.hi.u32 %r354, %r44, %r189;
|
|
setp.eq.s32 %p48, %r26, 0;
|
|
neg.s32 %r190, %r36;
|
|
selp.b32 %r356, %r36, %r190, %p48;
|
|
setp.lt.s32 %p49, %r354, 1;
|
|
@%p49 bra BB0_40;
|
|
|
|
mul.lo.s32 %r191, %r44, -921707870;
|
|
shr.u32 %r192, %r191, 31;
|
|
shl.b32 %r193, %r354, 1;
|
|
add.s32 %r354, %r192, %r193;
|
|
add.s32 %r355, %r355, 1;
|
|
|
|
BB0_40:
|
|
mov.u32 %r194, 126;
|
|
sub.s32 %r195, %r194, %r355;
|
|
shl.b32 %r196, %r195, 23;
|
|
add.s32 %r197, %r354, 1;
|
|
shr.u32 %r198, %r197, 7;
|
|
add.s32 %r199, %r198, 1;
|
|
shr.u32 %r200, %r199, 1;
|
|
add.s32 %r201, %r200, %r196;
|
|
or.b32 %r202, %r201, %r352;
|
|
mov.b32 %f902, %r202;
|
|
|
|
BB0_41:
|
|
mul.rn.f32 %f100, %f902, %f902;
|
|
add.s32 %r52, %r356, 1;
|
|
and.b32 %r53, %r52, 1;
|
|
setp.eq.s32 %p50, %r53, 0;
|
|
@%p50 bra BB0_43;
|
|
bra.uni BB0_42;
|
|
|
|
BB0_43:
|
|
mov.f32 %f433, 0f3C08839E;
|
|
mov.f32 %f434, 0fB94CA1F9;
|
|
fma.rn.f32 %f903, %f434, %f100, %f433;
|
|
bra.uni BB0_44;
|
|
|
|
BB0_42:
|
|
mov.f32 %f431, 0fBAB6061A;
|
|
mov.f32 %f432, 0f37CCF5CE;
|
|
fma.rn.f32 %f903, %f432, %f100, %f431;
|
|
|
|
BB0_44:
|
|
@%p50 bra BB0_46;
|
|
bra.uni BB0_45;
|
|
|
|
BB0_46:
|
|
mov.f32 %f438, 0fBE2AAAA3;
|
|
fma.rn.f32 %f439, %f903, %f100, %f438;
|
|
fma.rn.f32 %f904, %f439, %f100, %f419;
|
|
bra.uni BB0_47;
|
|
|
|
BB0_45:
|
|
mov.f32 %f435, 0f3D2AAAA5;
|
|
fma.rn.f32 %f436, %f903, %f100, %f435;
|
|
mov.f32 %f437, 0fBF000000;
|
|
fma.rn.f32 %f904, %f436, %f100, %f437;
|
|
|
|
BB0_47:
|
|
fma.rn.f32 %f905, %f904, %f902, %f902;
|
|
@%p50 bra BB0_49;
|
|
|
|
fma.rn.f32 %f905, %f904, %f100, %f375;
|
|
|
|
BB0_49:
|
|
and.b32 %r203, %r52, 2;
|
|
setp.eq.s32 %p53, %r203, 0;
|
|
@%p53 bra BB0_51;
|
|
|
|
mov.f32 %f443, 0fBF800000;
|
|
fma.rn.f32 %f905, %f905, %f443, %f419;
|
|
|
|
BB0_51:
|
|
@%p41 bra BB0_53;
|
|
|
|
mul.rn.f32 %f907, %f907, %f419;
|
|
|
|
BB0_53:
|
|
mul.f32 %f445, %f907, 0f3F22F983;
|
|
cvt.rni.s32.f32 %r366, %f445;
|
|
cvt.rn.f32.s32 %f446, %r366;
|
|
neg.f32 %f447, %f446;
|
|
fma.rn.f32 %f449, %f447, %f425, %f907;
|
|
fma.rn.f32 %f451, %f447, %f427, %f449;
|
|
fma.rn.f32 %f908, %f447, %f429, %f451;
|
|
abs.f32 %f453, %f907;
|
|
setp.leu.f32 %p55, %f453, 0f47CE4780;
|
|
@%p55 bra BB0_64;
|
|
|
|
mov.b32 %r55, %f907;
|
|
shr.u32 %r56, %r55, 23;
|
|
shl.b32 %r206, %r55, 8;
|
|
or.b32 %r57, %r206, -2147483648;
|
|
mov.u32 %r357, 0;
|
|
mov.u64 %rd159, %rd2;
|
|
mov.u64 %rd160, %rd26;
|
|
mov.u32 %r358, %r357;
|
|
|
|
BB0_55:
|
|
.pragma "nounroll";
|
|
shl.b64 %rd67, %rd160, 2;
|
|
mov.u64 %rd68, __cudart_i2opi_f;
|
|
add.s64 %rd69, %rd68, %rd67;
|
|
ld.const.u32 %r209, [%rd69];
|
|
// inline asm
|
|
{
|
|
mad.lo.cc.u32 %r207, %r209, %r57, %r358;
|
|
madc.hi.u32 %r358, %r209, %r57, 0;
|
|
}
|
|
// inline asm
|
|
st.local.u32 [%rd159], %r207;
|
|
add.s32 %r357, %r357, 1;
|
|
cvt.s64.s32 %rd160, %r357;
|
|
mul.wide.s32 %rd70, %r357, 4;
|
|
add.s64 %rd159, %rd2, %rd70;
|
|
setp.ne.s32 %p56, %r357, 6;
|
|
@%p56 bra BB0_55;
|
|
|
|
and.b32 %r212, %r56, 255;
|
|
add.s32 %r213, %r212, -128;
|
|
shr.u32 %r214, %r213, 5;
|
|
and.b32 %r62, %r55, -2147483648;
|
|
st.local.u32 [%rd2+24], %r358;
|
|
mov.u32 %r215, 6;
|
|
sub.s32 %r216, %r215, %r214;
|
|
mul.wide.s32 %rd71, %r216, 4;
|
|
add.s64 %rd15, %rd2, %rd71;
|
|
ld.local.u32 %r359, [%rd15];
|
|
ld.local.u32 %r360, [%rd15+-4];
|
|
and.b32 %r65, %r56, 31;
|
|
setp.eq.s32 %p57, %r65, 0;
|
|
@%p57 bra BB0_58;
|
|
|
|
mov.u32 %r217, 32;
|
|
sub.s32 %r218, %r217, %r65;
|
|
shr.u32 %r219, %r360, %r218;
|
|
shl.b32 %r220, %r359, %r65;
|
|
add.s32 %r359, %r219, %r220;
|
|
ld.local.u32 %r221, [%rd15+-8];
|
|
shr.u32 %r222, %r221, %r218;
|
|
shl.b32 %r223, %r360, %r65;
|
|
add.s32 %r360, %r222, %r223;
|
|
|
|
BB0_58:
|
|
shr.u32 %r224, %r360, 30;
|
|
shl.b32 %r225, %r359, 2;
|
|
add.s32 %r361, %r224, %r225;
|
|
shl.b32 %r71, %r360, 2;
|
|
shr.u32 %r226, %r361, 31;
|
|
shr.u32 %r227, %r359, 30;
|
|
add.s32 %r72, %r226, %r227;
|
|
setp.eq.s32 %p58, %r226, 0;
|
|
@%p58 bra BB0_59;
|
|
bra.uni BB0_60;
|
|
|
|
BB0_59:
|
|
mov.u32 %r362, %r62;
|
|
mov.u32 %r363, %r71;
|
|
bra.uni BB0_61;
|
|
|
|
BB0_60:
|
|
not.b32 %r228, %r361;
|
|
neg.s32 %r363, %r71;
|
|
setp.eq.s32 %p59, %r71, 0;
|
|
selp.u32 %r229, 1, 0, %p59;
|
|
add.s32 %r361, %r229, %r228;
|
|
xor.b32 %r362, %r62, -2147483648;
|
|
|
|
BB0_61:
|
|
clz.b32 %r365, %r361;
|
|
setp.eq.s32 %p60, %r365, 0;
|
|
shl.b32 %r230, %r361, %r365;
|
|
mov.u32 %r231, 32;
|
|
sub.s32 %r232, %r231, %r365;
|
|
shr.u32 %r233, %r363, %r232;
|
|
add.s32 %r234, %r233, %r230;
|
|
selp.b32 %r80, %r361, %r234, %p60;
|
|
mov.u32 %r235, -921707870;
|
|
mul.hi.u32 %r364, %r80, %r235;
|
|
setp.eq.s32 %p61, %r62, 0;
|
|
neg.s32 %r236, %r72;
|
|
selp.b32 %r366, %r72, %r236, %p61;
|
|
setp.lt.s32 %p62, %r364, 1;
|
|
@%p62 bra BB0_63;
|
|
|
|
mul.lo.s32 %r237, %r80, -921707870;
|
|
shr.u32 %r238, %r237, 31;
|
|
shl.b32 %r239, %r364, 1;
|
|
add.s32 %r364, %r238, %r239;
|
|
add.s32 %r365, %r365, 1;
|
|
|
|
BB0_63:
|
|
mov.u32 %r240, 126;
|
|
sub.s32 %r241, %r240, %r365;
|
|
shl.b32 %r242, %r241, 23;
|
|
add.s32 %r243, %r364, 1;
|
|
shr.u32 %r244, %r243, 7;
|
|
add.s32 %r245, %r244, 1;
|
|
shr.u32 %r246, %r245, 1;
|
|
add.s32 %r247, %r246, %r242;
|
|
or.b32 %r248, %r247, %r362;
|
|
mov.b32 %f908, %r248;
|
|
|
|
BB0_64:
|
|
mul.rn.f32 %f117, %f908, %f908;
|
|
and.b32 %r88, %r366, 1;
|
|
setp.eq.s32 %p63, %r88, 0;
|
|
@%p63 bra BB0_66;
|
|
bra.uni BB0_65;
|
|
|
|
BB0_66:
|
|
mov.f32 %f456, 0f3C08839E;
|
|
mov.f32 %f457, 0fB94CA1F9;
|
|
fma.rn.f32 %f909, %f457, %f117, %f456;
|
|
bra.uni BB0_67;
|
|
|
|
BB0_65:
|
|
mov.f32 %f454, 0fBAB6061A;
|
|
mov.f32 %f455, 0f37CCF5CE;
|
|
fma.rn.f32 %f909, %f455, %f117, %f454;
|
|
|
|
BB0_67:
|
|
@%p63 bra BB0_69;
|
|
bra.uni BB0_68;
|
|
|
|
BB0_69:
|
|
mov.f32 %f461, 0fBE2AAAA3;
|
|
fma.rn.f32 %f462, %f909, %f117, %f461;
|
|
fma.rn.f32 %f910, %f462, %f117, %f419;
|
|
bra.uni BB0_70;
|
|
|
|
BB0_68:
|
|
mov.f32 %f458, 0f3D2AAAA5;
|
|
fma.rn.f32 %f459, %f909, %f117, %f458;
|
|
mov.f32 %f460, 0fBF000000;
|
|
fma.rn.f32 %f910, %f459, %f117, %f460;
|
|
|
|
BB0_70:
|
|
fma.rn.f32 %f911, %f910, %f908, %f908;
|
|
@%p63 bra BB0_72;
|
|
|
|
fma.rn.f32 %f911, %f910, %f117, %f375;
|
|
|
|
BB0_72:
|
|
and.b32 %r249, %r366, 2;
|
|
setp.eq.s32 %p66, %r249, 0;
|
|
@%p66 bra BB0_74;
|
|
|
|
mov.f32 %f466, 0fBF800000;
|
|
fma.rn.f32 %f911, %f911, %f466, %f419;
|
|
|
|
BB0_74:
|
|
mul.f32 %f475, %f92, %f905;
|
|
mul.f32 %f476, %f92, %f911;
|
|
mul.f32 %f477, %f69, %f476;
|
|
mul.f32 %f478, %f70, %f476;
|
|
mul.f32 %f479, %f71, %f476;
|
|
fma.rn.f32 %f480, %f78, %f475, %f477;
|
|
fma.rn.f32 %f481, %f77, %f475, %f478;
|
|
fma.rn.f32 %f482, %f76, %f475, %f479;
|
|
fma.rn.f32 %f470, %f865, %f91, %f480;
|
|
fma.rn.f32 %f471, %f866, %f91, %f481;
|
|
fma.rn.f32 %f472, %f867, %f91, %f482;
|
|
mov.u32 %r251, 0;
|
|
st.local.u32 [%rd3], %r251;
|
|
ld.global.u32 %r250, [root];
|
|
mov.f32 %f474, 0f6C4ECB8F;
|
|
// inline asm
|
|
call _rt_trace_64, (%r250, %f12, %f13, %f14, %f470, %f471, %f472, %r251, %f85, %f474, %rd59, %r102);
|
|
// inline asm
|
|
ld.local.f32 %f483, [%rd3];
|
|
setp.lt.f32 %p67, %f483, 0f00000000;
|
|
selp.f32 %f484, 0f00000000, %f483, %p67;
|
|
selp.f32 %f485, 0f00000000, 0f3F800000, %p67;
|
|
fma.rn.f32 %f890, %f470, %f485, %f890;
|
|
fma.rn.f32 %f889, %f471, %f485, %f889;
|
|
fma.rn.f32 %f888, %f472, %f485, %f888;
|
|
add.f32 %f896, %f896, %f485;
|
|
mul.f32 %f486, %f866, %f471;
|
|
fma.rn.f32 %f487, %f865, %f470, %f486;
|
|
fma.rn.f32 %f488, %f867, %f472, %f487;
|
|
cvt.sat.f32.f32 %f489, %f488;
|
|
fma.rn.f32 %f897, %f484, %f489, %f897;
|
|
add.s32 %r345, %r345, 1;
|
|
setp.lt.s32 %p68, %r345, %r1;
|
|
@%p68 bra BB0_28;
|
|
|
|
add.s32 %r343, %r343, 1;
|
|
setp.lt.s32 %p69, %r343, %r1;
|
|
@%p69 bra BB0_27;
|
|
|
|
BB0_76:
|
|
mul.lo.s32 %r253, %r1, %r1;
|
|
cvt.rn.f32.s32 %f490, %r253;
|
|
div.rn.f32 %f491, %f897, %f490;
|
|
div.rn.f32 %f922, %f896, %f490;
|
|
add.f32 %f921, %f491, %f491;
|
|
|
|
BB0_77:
|
|
ld.global.u32 %r368, [imageEnabled];
|
|
and.b32 %r254, %r368, 8;
|
|
setp.eq.s32 %p70, %r254, 0;
|
|
@%p70 bra BB0_90;
|
|
|
|
mov.u32 %r332, 2;
|
|
cvt.u64.u32 %rd75, %r3;
|
|
cvt.u64.u32 %rd76, %r4;
|
|
mov.u64 %rd79, image_Mask;
|
|
cvta.global.u64 %rd74, %rd79;
|
|
// inline asm
|
|
call (%rd73), _rt_buffer_get_64, (%rd74, %r332, %r332, %rd75, %rd76, %rd26, %rd26);
|
|
// inline asm
|
|
mov.f32 %f494, 0f3E68BA2E;
|
|
cvt.rzi.f32.f32 %f495, %f494;
|
|
fma.rn.f32 %f496, %f495, 0fC0000000, 0f3EE8BA2E;
|
|
abs.f32 %f146, %f496;
|
|
abs.f32 %f147, %f922;
|
|
setp.lt.f32 %p71, %f147, 0f00800000;
|
|
mul.f32 %f497, %f147, 0f4B800000;
|
|
selp.f32 %f498, 0fC3170000, 0fC2FE0000, %p71;
|
|
selp.f32 %f499, %f497, %f147, %p71;
|
|
mov.b32 %r257, %f499;
|
|
and.b32 %r258, %r257, 8388607;
|
|
or.b32 %r259, %r258, 1065353216;
|
|
mov.b32 %f500, %r259;
|
|
shr.u32 %r260, %r257, 23;
|
|
cvt.rn.f32.u32 %f501, %r260;
|
|
add.f32 %f502, %f498, %f501;
|
|
setp.gt.f32 %p72, %f500, 0f3FB504F3;
|
|
mul.f32 %f503, %f500, 0f3F000000;
|
|
add.f32 %f504, %f502, 0f3F800000;
|
|
selp.f32 %f505, %f503, %f500, %p72;
|
|
selp.f32 %f506, %f504, %f502, %p72;
|
|
add.f32 %f507, %f505, 0fBF800000;
|
|
add.f32 %f493, %f505, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f492,%f493;
|
|
// inline asm
|
|
add.f32 %f508, %f507, %f507;
|
|
mul.f32 %f509, %f492, %f508;
|
|
mul.f32 %f510, %f509, %f509;
|
|
mov.f32 %f511, 0f3C4CAF63;
|
|
mov.f32 %f512, 0f3B18F0FE;
|
|
fma.rn.f32 %f513, %f512, %f510, %f511;
|
|
mov.f32 %f514, 0f3DAAAABD;
|
|
fma.rn.f32 %f515, %f513, %f510, %f514;
|
|
mul.rn.f32 %f516, %f515, %f510;
|
|
mul.rn.f32 %f517, %f516, %f509;
|
|
sub.f32 %f518, %f507, %f509;
|
|
neg.f32 %f519, %f509;
|
|
add.f32 %f520, %f518, %f518;
|
|
fma.rn.f32 %f521, %f519, %f507, %f520;
|
|
mul.rn.f32 %f522, %f492, %f521;
|
|
add.f32 %f523, %f517, %f509;
|
|
sub.f32 %f524, %f509, %f523;
|
|
add.f32 %f525, %f517, %f524;
|
|
add.f32 %f526, %f522, %f525;
|
|
add.f32 %f527, %f523, %f526;
|
|
sub.f32 %f528, %f523, %f527;
|
|
add.f32 %f529, %f526, %f528;
|
|
mov.f32 %f530, 0f3F317200;
|
|
mul.rn.f32 %f531, %f506, %f530;
|
|
mov.f32 %f532, 0f35BFBE8E;
|
|
mul.rn.f32 %f533, %f506, %f532;
|
|
add.f32 %f534, %f531, %f527;
|
|
sub.f32 %f535, %f531, %f534;
|
|
add.f32 %f536, %f527, %f535;
|
|
add.f32 %f537, %f529, %f536;
|
|
add.f32 %f538, %f533, %f537;
|
|
add.f32 %f539, %f534, %f538;
|
|
sub.f32 %f540, %f534, %f539;
|
|
add.f32 %f541, %f538, %f540;
|
|
mov.f32 %f542, 0f3EE8BA2E;
|
|
mul.rn.f32 %f543, %f542, %f539;
|
|
neg.f32 %f544, %f543;
|
|
fma.rn.f32 %f545, %f542, %f539, %f544;
|
|
fma.rn.f32 %f546, %f542, %f541, %f545;
|
|
mov.f32 %f547, 0f00000000;
|
|
fma.rn.f32 %f548, %f547, %f539, %f546;
|
|
add.rn.f32 %f549, %f543, %f548;
|
|
neg.f32 %f550, %f549;
|
|
add.rn.f32 %f551, %f543, %f550;
|
|
add.rn.f32 %f552, %f551, %f548;
|
|
mov.b32 %r261, %f549;
|
|
setp.eq.s32 %p73, %r261, 1118925336;
|
|
add.s32 %r262, %r261, -1;
|
|
mov.b32 %f553, %r262;
|
|
add.f32 %f554, %f552, 0f37000000;
|
|
selp.f32 %f555, %f553, %f549, %p73;
|
|
selp.f32 %f148, %f554, %f552, %p73;
|
|
mul.f32 %f556, %f555, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f557, %f556;
|
|
mov.f32 %f558, 0fBF317200;
|
|
fma.rn.f32 %f559, %f557, %f558, %f555;
|
|
mov.f32 %f560, 0fB5BFBE8E;
|
|
fma.rn.f32 %f561, %f557, %f560, %f559;
|
|
mul.f32 %f562, %f561, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f563, %f562;
|
|
add.f32 %f564, %f557, 0f00000000;
|
|
ex2.approx.f32 %f565, %f564;
|
|
mul.f32 %f566, %f563, %f565;
|
|
setp.lt.f32 %p74, %f555, 0fC2D20000;
|
|
selp.f32 %f567, 0f00000000, %f566, %p74;
|
|
setp.gt.f32 %p75, %f555, 0f42D20000;
|
|
selp.f32 %f923, 0f7F800000, %f567, %p75;
|
|
setp.eq.f32 %p76, %f923, 0f7F800000;
|
|
@%p76 bra BB0_80;
|
|
|
|
fma.rn.f32 %f923, %f923, %f148, %f923;
|
|
|
|
BB0_80:
|
|
setp.lt.f32 %p77, %f922, 0f00000000;
|
|
setp.eq.f32 %p78, %f146, 0f3F800000;
|
|
and.pred %p3, %p77, %p78;
|
|
mov.b32 %r263, %f923;
|
|
xor.b32 %r264, %r263, -2147483648;
|
|
mov.b32 %f568, %r264;
|
|
selp.f32 %f925, %f568, %f923, %p3;
|
|
setp.eq.f32 %p79, %f922, 0f00000000;
|
|
@%p79 bra BB0_83;
|
|
bra.uni BB0_81;
|
|
|
|
BB0_83:
|
|
add.f32 %f571, %f922, %f922;
|
|
selp.f32 %f925, %f571, 0f00000000, %p78;
|
|
bra.uni BB0_84;
|
|
|
|
BB0_138:
|
|
mov.u64 %rd147, image_HDR;
|
|
cvta.global.u64 %rd142, %rd147;
|
|
mov.u32 %r327, 8;
|
|
mov.u64 %rd146, 0;
|
|
// inline asm
|
|
call (%rd141), _rt_buffer_get_64, (%rd142, %r101, %r327, %rd19, %rd20, %rd146, %rd146);
|
|
// inline asm
|
|
mov.f32 %f855, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs50, %f855;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs51, 0;
|
|
st.v4.u16 [%rd141], {%rs50, %rs50, %rs50, %rs51};
|
|
|
|
BB0_139:
|
|
ld.global.u8 %rs52, [imageEnabled];
|
|
and.b16 %rs53, %rs52, 64;
|
|
setp.eq.s16 %p143, %rs53, 0;
|
|
@%p143 bra BB0_141;
|
|
|
|
cvt.u64.u32 %rd150, %r3;
|
|
cvt.u64.u32 %rd151, %r4;
|
|
mov.u64 %rd154, image_Dir;
|
|
cvta.global.u64 %rd149, %rd154;
|
|
mov.u64 %rd153, 0;
|
|
// inline asm
|
|
call (%rd148), _rt_buffer_get_64, (%rd149, %r101, %r102, %rd150, %rd151, %rd153, %rd153);
|
|
// inline asm
|
|
mov.u16 %rs54, 0;
|
|
st.v4.u8 [%rd148], {%rs54, %rs54, %rs54, %rs54};
|
|
bra.uni BB0_141;
|
|
|
|
BB0_81:
|
|
setp.geu.f32 %p80, %f922, 0f00000000;
|
|
@%p80 bra BB0_84;
|
|
|
|
cvt.rzi.f32.f32 %f570, %f542;
|
|
setp.neu.f32 %p81, %f570, 0f3EE8BA2E;
|
|
selp.f32 %f925, 0f7FFFFFFF, %f925, %p81;
|
|
|
|
BB0_84:
|
|
add.f32 %f572, %f147, 0f3EE8BA2E;
|
|
mov.b32 %r265, %f572;
|
|
setp.lt.s32 %p83, %r265, 2139095040;
|
|
@%p83 bra BB0_89;
|
|
|
|
setp.gtu.f32 %p84, %f147, 0f7F800000;
|
|
@%p84 bra BB0_88;
|
|
bra.uni BB0_86;
|
|
|
|
BB0_88:
|
|
add.f32 %f925, %f922, 0f3EE8BA2E;
|
|
bra.uni BB0_89;
|
|
|
|
BB0_86:
|
|
setp.neu.f32 %p85, %f147, 0f7F800000;
|
|
@%p85 bra BB0_89;
|
|
|
|
selp.f32 %f925, 0fFF800000, 0f7F800000, %p3;
|
|
|
|
BB0_89:
|
|
mul.f32 %f573, %f925, 0f437F0000;
|
|
setp.eq.f32 %p86, %f922, 0f3F800000;
|
|
selp.f32 %f574, 0f437F0000, %f573, %p86;
|
|
cvt.rzi.u32.f32 %r266, %f574;
|
|
cvt.u16.u32 %rs10, %r266;
|
|
mov.u16 %rs11, 255;
|
|
st.v2.u8 [%rd73], {%rs10, %rs11};
|
|
ld.global.u32 %r368, [imageEnabled];
|
|
|
|
BB0_90:
|
|
ld.global.f32 %f575, [lightColor];
|
|
mul.f32 %f159, %f921, %f575;
|
|
ld.global.f32 %f576, [lightColor+4];
|
|
mul.f32 %f160, %f921, %f576;
|
|
ld.global.f32 %f577, [lightColor+8];
|
|
mul.f32 %f161, %f921, %f577;
|
|
and.b32 %r267, %r368, 1;
|
|
setp.eq.b32 %p87, %r267, 1;
|
|
@!%p87 bra BB0_125;
|
|
bra.uni BB0_91;
|
|
|
|
BB0_91:
|
|
mov.f32 %f580, 0f3E666666;
|
|
cvt.rzi.f32.f32 %f581, %f580;
|
|
fma.rn.f32 %f582, %f581, 0fC0000000, 0f3EE66666;
|
|
abs.f32 %f162, %f582;
|
|
abs.f32 %f163, %f159;
|
|
setp.lt.f32 %p88, %f163, 0f00800000;
|
|
mul.f32 %f583, %f163, 0f4B800000;
|
|
selp.f32 %f584, 0fC3170000, 0fC2FE0000, %p88;
|
|
selp.f32 %f585, %f583, %f163, %p88;
|
|
mov.b32 %r268, %f585;
|
|
and.b32 %r269, %r268, 8388607;
|
|
or.b32 %r270, %r269, 1065353216;
|
|
mov.b32 %f586, %r270;
|
|
shr.u32 %r271, %r268, 23;
|
|
cvt.rn.f32.u32 %f587, %r271;
|
|
add.f32 %f588, %f584, %f587;
|
|
setp.gt.f32 %p89, %f586, 0f3FB504F3;
|
|
mul.f32 %f589, %f586, 0f3F000000;
|
|
add.f32 %f590, %f588, 0f3F800000;
|
|
selp.f32 %f591, %f589, %f586, %p89;
|
|
selp.f32 %f592, %f590, %f588, %p89;
|
|
add.f32 %f593, %f591, 0fBF800000;
|
|
add.f32 %f579, %f591, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f578,%f579;
|
|
// inline asm
|
|
add.f32 %f594, %f593, %f593;
|
|
mul.f32 %f595, %f578, %f594;
|
|
mul.f32 %f596, %f595, %f595;
|
|
mov.f32 %f597, 0f3C4CAF63;
|
|
mov.f32 %f598, 0f3B18F0FE;
|
|
fma.rn.f32 %f599, %f598, %f596, %f597;
|
|
mov.f32 %f600, 0f3DAAAABD;
|
|
fma.rn.f32 %f601, %f599, %f596, %f600;
|
|
mul.rn.f32 %f602, %f601, %f596;
|
|
mul.rn.f32 %f603, %f602, %f595;
|
|
sub.f32 %f604, %f593, %f595;
|
|
neg.f32 %f605, %f595;
|
|
add.f32 %f606, %f604, %f604;
|
|
fma.rn.f32 %f607, %f605, %f593, %f606;
|
|
mul.rn.f32 %f608, %f578, %f607;
|
|
add.f32 %f609, %f603, %f595;
|
|
sub.f32 %f610, %f595, %f609;
|
|
add.f32 %f611, %f603, %f610;
|
|
add.f32 %f612, %f608, %f611;
|
|
add.f32 %f613, %f609, %f612;
|
|
sub.f32 %f614, %f609, %f613;
|
|
add.f32 %f615, %f612, %f614;
|
|
mov.f32 %f616, 0f3F317200;
|
|
mul.rn.f32 %f617, %f592, %f616;
|
|
mov.f32 %f618, 0f35BFBE8E;
|
|
mul.rn.f32 %f619, %f592, %f618;
|
|
add.f32 %f620, %f617, %f613;
|
|
sub.f32 %f621, %f617, %f620;
|
|
add.f32 %f622, %f613, %f621;
|
|
add.f32 %f623, %f615, %f622;
|
|
add.f32 %f624, %f619, %f623;
|
|
add.f32 %f625, %f620, %f624;
|
|
sub.f32 %f626, %f620, %f625;
|
|
add.f32 %f627, %f624, %f626;
|
|
mov.f32 %f628, 0f3EE66666;
|
|
mul.rn.f32 %f629, %f628, %f625;
|
|
neg.f32 %f630, %f629;
|
|
fma.rn.f32 %f631, %f628, %f625, %f630;
|
|
fma.rn.f32 %f632, %f628, %f627, %f631;
|
|
mov.f32 %f633, 0f00000000;
|
|
fma.rn.f32 %f634, %f633, %f625, %f632;
|
|
add.rn.f32 %f635, %f629, %f634;
|
|
neg.f32 %f636, %f635;
|
|
add.rn.f32 %f637, %f629, %f636;
|
|
add.rn.f32 %f638, %f637, %f634;
|
|
mov.b32 %r272, %f635;
|
|
setp.eq.s32 %p90, %r272, 1118925336;
|
|
add.s32 %r273, %r272, -1;
|
|
mov.b32 %f639, %r273;
|
|
add.f32 %f640, %f638, 0f37000000;
|
|
selp.f32 %f641, %f639, %f635, %p90;
|
|
selp.f32 %f164, %f640, %f638, %p90;
|
|
mul.f32 %f642, %f641, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f643, %f642;
|
|
mov.f32 %f644, 0fBF317200;
|
|
fma.rn.f32 %f645, %f643, %f644, %f641;
|
|
mov.f32 %f646, 0fB5BFBE8E;
|
|
fma.rn.f32 %f647, %f643, %f646, %f645;
|
|
mul.f32 %f648, %f647, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f649, %f648;
|
|
add.f32 %f650, %f643, 0f00000000;
|
|
ex2.approx.f32 %f651, %f650;
|
|
mul.f32 %f652, %f649, %f651;
|
|
setp.lt.f32 %p91, %f641, 0fC2D20000;
|
|
selp.f32 %f653, 0f00000000, %f652, %p91;
|
|
setp.gt.f32 %p92, %f641, 0f42D20000;
|
|
selp.f32 %f926, 0f7F800000, %f653, %p92;
|
|
setp.eq.f32 %p93, %f926, 0f7F800000;
|
|
@%p93 bra BB0_93;
|
|
|
|
fma.rn.f32 %f926, %f926, %f164, %f926;
|
|
|
|
BB0_93:
|
|
setp.lt.f32 %p94, %f159, 0f00000000;
|
|
setp.eq.f32 %p95, %f162, 0f3F800000;
|
|
and.pred %p4, %p94, %p95;
|
|
mov.b32 %r274, %f926;
|
|
xor.b32 %r275, %r274, -2147483648;
|
|
mov.b32 %f654, %r275;
|
|
selp.f32 %f928, %f654, %f926, %p4;
|
|
setp.eq.f32 %p96, %f159, 0f00000000;
|
|
@%p96 bra BB0_96;
|
|
bra.uni BB0_94;
|
|
|
|
BB0_96:
|
|
add.f32 %f657, %f159, %f159;
|
|
selp.f32 %f928, %f657, 0f00000000, %p95;
|
|
bra.uni BB0_97;
|
|
|
|
BB0_94:
|
|
setp.geu.f32 %p97, %f159, 0f00000000;
|
|
@%p97 bra BB0_97;
|
|
|
|
cvt.rzi.f32.f32 %f656, %f628;
|
|
setp.neu.f32 %p98, %f656, 0f3EE66666;
|
|
selp.f32 %f928, 0f7FFFFFFF, %f928, %p98;
|
|
|
|
BB0_97:
|
|
add.f32 %f658, %f163, 0f3EE66666;
|
|
mov.b32 %r276, %f658;
|
|
setp.lt.s32 %p100, %r276, 2139095040;
|
|
@%p100 bra BB0_102;
|
|
|
|
setp.gtu.f32 %p101, %f163, 0f7F800000;
|
|
@%p101 bra BB0_101;
|
|
bra.uni BB0_99;
|
|
|
|
BB0_101:
|
|
add.f32 %f928, %f159, 0f3EE66666;
|
|
bra.uni BB0_102;
|
|
|
|
BB0_99:
|
|
setp.neu.f32 %p102, %f163, 0f7F800000;
|
|
@%p102 bra BB0_102;
|
|
|
|
selp.f32 %f928, 0fFF800000, 0f7F800000, %p4;
|
|
|
|
BB0_102:
|
|
setp.eq.f32 %p103, %f159, 0f3F800000;
|
|
selp.f32 %f175, 0f3F800000, %f928, %p103;
|
|
abs.f32 %f176, %f160;
|
|
setp.lt.f32 %p104, %f176, 0f00800000;
|
|
mul.f32 %f661, %f176, 0f4B800000;
|
|
selp.f32 %f662, 0fC3170000, 0fC2FE0000, %p104;
|
|
selp.f32 %f663, %f661, %f176, %p104;
|
|
mov.b32 %r277, %f663;
|
|
and.b32 %r278, %r277, 8388607;
|
|
or.b32 %r279, %r278, 1065353216;
|
|
mov.b32 %f664, %r279;
|
|
shr.u32 %r280, %r277, 23;
|
|
cvt.rn.f32.u32 %f665, %r280;
|
|
add.f32 %f666, %f662, %f665;
|
|
setp.gt.f32 %p105, %f664, 0f3FB504F3;
|
|
mul.f32 %f667, %f664, 0f3F000000;
|
|
add.f32 %f668, %f666, 0f3F800000;
|
|
selp.f32 %f669, %f667, %f664, %p105;
|
|
selp.f32 %f670, %f668, %f666, %p105;
|
|
add.f32 %f671, %f669, 0fBF800000;
|
|
add.f32 %f660, %f669, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f659,%f660;
|
|
// inline asm
|
|
add.f32 %f672, %f671, %f671;
|
|
mul.f32 %f673, %f659, %f672;
|
|
mul.f32 %f674, %f673, %f673;
|
|
fma.rn.f32 %f677, %f598, %f674, %f597;
|
|
fma.rn.f32 %f679, %f677, %f674, %f600;
|
|
mul.rn.f32 %f680, %f679, %f674;
|
|
mul.rn.f32 %f681, %f680, %f673;
|
|
sub.f32 %f682, %f671, %f673;
|
|
neg.f32 %f683, %f673;
|
|
add.f32 %f684, %f682, %f682;
|
|
fma.rn.f32 %f685, %f683, %f671, %f684;
|
|
mul.rn.f32 %f686, %f659, %f685;
|
|
add.f32 %f687, %f681, %f673;
|
|
sub.f32 %f688, %f673, %f687;
|
|
add.f32 %f689, %f681, %f688;
|
|
add.f32 %f690, %f686, %f689;
|
|
add.f32 %f691, %f687, %f690;
|
|
sub.f32 %f692, %f687, %f691;
|
|
add.f32 %f693, %f690, %f692;
|
|
mul.rn.f32 %f695, %f670, %f616;
|
|
mul.rn.f32 %f697, %f670, %f618;
|
|
add.f32 %f698, %f695, %f691;
|
|
sub.f32 %f699, %f695, %f698;
|
|
add.f32 %f700, %f691, %f699;
|
|
add.f32 %f701, %f693, %f700;
|
|
add.f32 %f702, %f697, %f701;
|
|
add.f32 %f703, %f698, %f702;
|
|
sub.f32 %f704, %f698, %f703;
|
|
add.f32 %f705, %f702, %f704;
|
|
mul.rn.f32 %f707, %f628, %f703;
|
|
neg.f32 %f708, %f707;
|
|
fma.rn.f32 %f709, %f628, %f703, %f708;
|
|
fma.rn.f32 %f710, %f628, %f705, %f709;
|
|
fma.rn.f32 %f712, %f633, %f703, %f710;
|
|
add.rn.f32 %f713, %f707, %f712;
|
|
neg.f32 %f714, %f713;
|
|
add.rn.f32 %f715, %f707, %f714;
|
|
add.rn.f32 %f716, %f715, %f712;
|
|
mov.b32 %r281, %f713;
|
|
setp.eq.s32 %p106, %r281, 1118925336;
|
|
add.s32 %r282, %r281, -1;
|
|
mov.b32 %f717, %r282;
|
|
add.f32 %f718, %f716, 0f37000000;
|
|
selp.f32 %f719, %f717, %f713, %p106;
|
|
selp.f32 %f177, %f718, %f716, %p106;
|
|
mul.f32 %f720, %f719, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f721, %f720;
|
|
fma.rn.f32 %f723, %f721, %f644, %f719;
|
|
fma.rn.f32 %f725, %f721, %f646, %f723;
|
|
mul.f32 %f726, %f725, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f727, %f726;
|
|
add.f32 %f728, %f721, 0f00000000;
|
|
ex2.approx.f32 %f729, %f728;
|
|
mul.f32 %f730, %f727, %f729;
|
|
setp.lt.f32 %p107, %f719, 0fC2D20000;
|
|
selp.f32 %f731, 0f00000000, %f730, %p107;
|
|
setp.gt.f32 %p108, %f719, 0f42D20000;
|
|
selp.f32 %f929, 0f7F800000, %f731, %p108;
|
|
setp.eq.f32 %p109, %f929, 0f7F800000;
|
|
@%p109 bra BB0_104;
|
|
|
|
fma.rn.f32 %f929, %f929, %f177, %f929;
|
|
|
|
BB0_104:
|
|
setp.lt.f32 %p110, %f160, 0f00000000;
|
|
and.pred %p5, %p110, %p95;
|
|
mov.b32 %r283, %f929;
|
|
xor.b32 %r284, %r283, -2147483648;
|
|
mov.b32 %f732, %r284;
|
|
selp.f32 %f931, %f732, %f929, %p5;
|
|
setp.eq.f32 %p112, %f160, 0f00000000;
|
|
@%p112 bra BB0_107;
|
|
bra.uni BB0_105;
|
|
|
|
BB0_107:
|
|
add.f32 %f735, %f160, %f160;
|
|
selp.f32 %f931, %f735, 0f00000000, %p95;
|
|
bra.uni BB0_108;
|
|
|
|
BB0_105:
|
|
setp.geu.f32 %p113, %f160, 0f00000000;
|
|
@%p113 bra BB0_108;
|
|
|
|
cvt.rzi.f32.f32 %f734, %f628;
|
|
setp.neu.f32 %p114, %f734, 0f3EE66666;
|
|
selp.f32 %f931, 0f7FFFFFFF, %f931, %p114;
|
|
|
|
BB0_108:
|
|
add.f32 %f736, %f176, 0f3EE66666;
|
|
mov.b32 %r285, %f736;
|
|
setp.lt.s32 %p116, %r285, 2139095040;
|
|
@%p116 bra BB0_113;
|
|
|
|
setp.gtu.f32 %p117, %f176, 0f7F800000;
|
|
@%p117 bra BB0_112;
|
|
bra.uni BB0_110;
|
|
|
|
BB0_112:
|
|
add.f32 %f931, %f160, 0f3EE66666;
|
|
bra.uni BB0_113;
|
|
|
|
BB0_110:
|
|
setp.neu.f32 %p118, %f176, 0f7F800000;
|
|
@%p118 bra BB0_113;
|
|
|
|
selp.f32 %f931, 0fFF800000, 0f7F800000, %p5;
|
|
|
|
BB0_113:
|
|
setp.eq.f32 %p119, %f160, 0f3F800000;
|
|
selp.f32 %f188, 0f3F800000, %f931, %p119;
|
|
abs.f32 %f189, %f161;
|
|
setp.lt.f32 %p120, %f189, 0f00800000;
|
|
mul.f32 %f739, %f189, 0f4B800000;
|
|
selp.f32 %f740, 0fC3170000, 0fC2FE0000, %p120;
|
|
selp.f32 %f741, %f739, %f189, %p120;
|
|
mov.b32 %r286, %f741;
|
|
and.b32 %r287, %r286, 8388607;
|
|
or.b32 %r288, %r287, 1065353216;
|
|
mov.b32 %f742, %r288;
|
|
shr.u32 %r289, %r286, 23;
|
|
cvt.rn.f32.u32 %f743, %r289;
|
|
add.f32 %f744, %f740, %f743;
|
|
setp.gt.f32 %p121, %f742, 0f3FB504F3;
|
|
mul.f32 %f745, %f742, 0f3F000000;
|
|
add.f32 %f746, %f744, 0f3F800000;
|
|
selp.f32 %f747, %f745, %f742, %p121;
|
|
selp.f32 %f748, %f746, %f744, %p121;
|
|
add.f32 %f749, %f747, 0fBF800000;
|
|
add.f32 %f738, %f747, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f737,%f738;
|
|
// inline asm
|
|
add.f32 %f750, %f749, %f749;
|
|
mul.f32 %f751, %f737, %f750;
|
|
mul.f32 %f752, %f751, %f751;
|
|
fma.rn.f32 %f755, %f598, %f752, %f597;
|
|
fma.rn.f32 %f757, %f755, %f752, %f600;
|
|
mul.rn.f32 %f758, %f757, %f752;
|
|
mul.rn.f32 %f759, %f758, %f751;
|
|
sub.f32 %f760, %f749, %f751;
|
|
neg.f32 %f761, %f751;
|
|
add.f32 %f762, %f760, %f760;
|
|
fma.rn.f32 %f763, %f761, %f749, %f762;
|
|
mul.rn.f32 %f764, %f737, %f763;
|
|
add.f32 %f765, %f759, %f751;
|
|
sub.f32 %f766, %f751, %f765;
|
|
add.f32 %f767, %f759, %f766;
|
|
add.f32 %f768, %f764, %f767;
|
|
add.f32 %f769, %f765, %f768;
|
|
sub.f32 %f770, %f765, %f769;
|
|
add.f32 %f771, %f768, %f770;
|
|
mul.rn.f32 %f773, %f748, %f616;
|
|
mul.rn.f32 %f775, %f748, %f618;
|
|
add.f32 %f776, %f773, %f769;
|
|
sub.f32 %f777, %f773, %f776;
|
|
add.f32 %f778, %f769, %f777;
|
|
add.f32 %f779, %f771, %f778;
|
|
add.f32 %f780, %f775, %f779;
|
|
add.f32 %f781, %f776, %f780;
|
|
sub.f32 %f782, %f776, %f781;
|
|
add.f32 %f783, %f780, %f782;
|
|
mul.rn.f32 %f785, %f628, %f781;
|
|
neg.f32 %f786, %f785;
|
|
fma.rn.f32 %f787, %f628, %f781, %f786;
|
|
fma.rn.f32 %f788, %f628, %f783, %f787;
|
|
fma.rn.f32 %f790, %f633, %f781, %f788;
|
|
add.rn.f32 %f791, %f785, %f790;
|
|
neg.f32 %f792, %f791;
|
|
add.rn.f32 %f793, %f785, %f792;
|
|
add.rn.f32 %f794, %f793, %f790;
|
|
mov.b32 %r290, %f791;
|
|
setp.eq.s32 %p122, %r290, 1118925336;
|
|
add.s32 %r291, %r290, -1;
|
|
mov.b32 %f795, %r291;
|
|
add.f32 %f796, %f794, 0f37000000;
|
|
selp.f32 %f797, %f795, %f791, %p122;
|
|
selp.f32 %f190, %f796, %f794, %p122;
|
|
mul.f32 %f798, %f797, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f799, %f798;
|
|
fma.rn.f32 %f801, %f799, %f644, %f797;
|
|
fma.rn.f32 %f803, %f799, %f646, %f801;
|
|
mul.f32 %f804, %f803, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f805, %f804;
|
|
add.f32 %f806, %f799, 0f00000000;
|
|
ex2.approx.f32 %f807, %f806;
|
|
mul.f32 %f808, %f805, %f807;
|
|
setp.lt.f32 %p123, %f797, 0fC2D20000;
|
|
selp.f32 %f809, 0f00000000, %f808, %p123;
|
|
setp.gt.f32 %p124, %f797, 0f42D20000;
|
|
selp.f32 %f932, 0f7F800000, %f809, %p124;
|
|
setp.eq.f32 %p125, %f932, 0f7F800000;
|
|
@%p125 bra BB0_115;
|
|
|
|
fma.rn.f32 %f932, %f932, %f190, %f932;
|
|
|
|
BB0_115:
|
|
setp.lt.f32 %p126, %f161, 0f00000000;
|
|
and.pred %p6, %p126, %p95;
|
|
mov.b32 %r292, %f932;
|
|
xor.b32 %r293, %r292, -2147483648;
|
|
mov.b32 %f810, %r293;
|
|
selp.f32 %f934, %f810, %f932, %p6;
|
|
setp.eq.f32 %p128, %f161, 0f00000000;
|
|
@%p128 bra BB0_118;
|
|
bra.uni BB0_116;
|
|
|
|
BB0_118:
|
|
add.f32 %f813, %f161, %f161;
|
|
selp.f32 %f934, %f813, 0f00000000, %p95;
|
|
bra.uni BB0_119;
|
|
|
|
BB0_116:
|
|
setp.geu.f32 %p129, %f161, 0f00000000;
|
|
@%p129 bra BB0_119;
|
|
|
|
cvt.rzi.f32.f32 %f812, %f628;
|
|
setp.neu.f32 %p130, %f812, 0f3EE66666;
|
|
selp.f32 %f934, 0f7FFFFFFF, %f934, %p130;
|
|
|
|
BB0_119:
|
|
add.f32 %f814, %f189, 0f3EE66666;
|
|
mov.b32 %r294, %f814;
|
|
setp.lt.s32 %p132, %r294, 2139095040;
|
|
@%p132 bra BB0_124;
|
|
|
|
setp.gtu.f32 %p133, %f189, 0f7F800000;
|
|
@%p133 bra BB0_123;
|
|
bra.uni BB0_121;
|
|
|
|
BB0_123:
|
|
add.f32 %f934, %f161, 0f3EE66666;
|
|
bra.uni BB0_124;
|
|
|
|
BB0_121:
|
|
setp.neu.f32 %p134, %f189, 0f7F800000;
|
|
@%p134 bra BB0_124;
|
|
|
|
selp.f32 %f934, 0fFF800000, 0f7F800000, %p6;
|
|
|
|
BB0_124:
|
|
mov.u32 %r333, 2;
|
|
setp.eq.f32 %p135, %f161, 0f3F800000;
|
|
selp.f32 %f815, 0f3F800000, %f934, %p135;
|
|
cvt.u64.u32 %rd83, %r4;
|
|
cvt.u64.u32 %rd82, %r3;
|
|
mov.u64 %rd86, image;
|
|
cvta.global.u64 %rd81, %rd86;
|
|
// inline asm
|
|
call (%rd80), _rt_buffer_get_64, (%rd81, %r333, %r102, %rd82, %rd83, %rd26, %rd26);
|
|
// inline asm
|
|
cvt.sat.f32.f32 %f816, %f815;
|
|
mul.f32 %f817, %f816, 0f437FFD71;
|
|
cvt.rzi.u32.f32 %r297, %f817;
|
|
cvt.sat.f32.f32 %f818, %f188;
|
|
mul.f32 %f819, %f818, 0f437FFD71;
|
|
cvt.rzi.u32.f32 %r298, %f819;
|
|
cvt.sat.f32.f32 %f820, %f175;
|
|
mul.f32 %f821, %f820, 0f437FFD71;
|
|
cvt.rzi.u32.f32 %r299, %f821;
|
|
cvt.u16.u32 %rs12, %r297;
|
|
cvt.u16.u32 %rs13, %r299;
|
|
cvt.u16.u32 %rs14, %r298;
|
|
mov.u16 %rs15, 255;
|
|
st.v4.u8 [%rd80], {%rs12, %rs14, %rs13, %rs15};
|
|
ld.global.u32 %r368, [imageEnabled];
|
|
|
|
BB0_125:
|
|
and.b32 %r300, %r368, 4;
|
|
setp.eq.s32 %p136, %r300, 0;
|
|
@%p136 bra BB0_129;
|
|
|
|
ld.global.u32 %r301, [additive];
|
|
setp.eq.s32 %p137, %r301, 0;
|
|
cvt.u64.u32 %rd17, %r3;
|
|
cvt.u64.u32 %rd18, %r4;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs16, %f375;}
|
|
|
|
// inline asm
|
|
@%p137 bra BB0_128;
|
|
|
|
mov.u32 %r334, 2;
|
|
mov.u64 %rd99, image_HDR;
|
|
cvta.global.u64 %rd88, %rd99;
|
|
mov.u32 %r305, 8;
|
|
// inline asm
|
|
call (%rd87), _rt_buffer_get_64, (%rd88, %r334, %r305, %rd17, %rd18, %rd26, %rd26);
|
|
// inline asm
|
|
ld.v4.u16 {%rs23, %rs24, %rs25, %rs26}, [%rd87];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f823, %rs23;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f824, %rs24;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f825, %rs25;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd93), _rt_buffer_get_64, (%rd88, %r334, %r305, %rd17, %rd18, %rd26, %rd26);
|
|
// inline asm
|
|
add.f32 %f826, %f159, %f823;
|
|
add.f32 %f827, %f160, %f824;
|
|
add.f32 %f828, %f161, %f825;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs22, %f828;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs21, %f827;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs20, %f826;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd93], {%rs20, %rs21, %rs22, %rs16};
|
|
bra.uni BB0_129;
|
|
|
|
BB0_128:
|
|
mov.u32 %r335, 2;
|
|
mov.u64 %rd106, image_HDR;
|
|
cvta.global.u64 %rd101, %rd106;
|
|
mov.u32 %r307, 8;
|
|
// inline asm
|
|
call (%rd100), _rt_buffer_get_64, (%rd101, %r335, %r307, %rd17, %rd18, %rd26, %rd26);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs29, %f161;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs28, %f160;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs27, %f159;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd100], {%rs27, %rs28, %rs29, %rs16};
|
|
|
|
BB0_129:
|
|
ld.global.u8 %rs30, [imageEnabled];
|
|
and.b16 %rs31, %rs30, 64;
|
|
setp.eq.s16 %p138, %rs31, 0;
|
|
@%p138 bra BB0_141;
|
|
|
|
mov.u32 %r336, 2;
|
|
mul.f32 %f832, %f889, %f889;
|
|
fma.rn.f32 %f833, %f890, %f890, %f832;
|
|
fma.rn.f32 %f834, %f888, %f888, %f833;
|
|
sqrt.rn.f32 %f835, %f834;
|
|
rcp.rn.f32 %f836, %f835;
|
|
mul.f32 %f837, %f890, %f836;
|
|
mul.f32 %f838, %f889, %f836;
|
|
mul.f32 %f839, %f888, %f836;
|
|
cvt.u64.u32 %rd110, %r4;
|
|
cvt.u64.u32 %rd109, %r3;
|
|
mov.u64 %rd113, image_Dir;
|
|
cvta.global.u64 %rd108, %rd113;
|
|
// inline asm
|
|
call (%rd107), _rt_buffer_get_64, (%rd108, %r336, %r102, %rd109, %rd110, %rd26, %rd26);
|
|
// inline asm
|
|
fma.rn.f32 %f840, %f837, 0f3F000000, 0f3F000000;
|
|
mul.f32 %f841, %f840, 0f437F0000;
|
|
cvt.rzi.u32.f32 %r310, %f841;
|
|
fma.rn.f32 %f842, %f838, 0f3F000000, 0f3F000000;
|
|
mul.f32 %f843, %f842, 0f437F0000;
|
|
cvt.rzi.u32.f32 %r311, %f843;
|
|
fma.rn.f32 %f844, %f839, 0f3F000000, 0f3F000000;
|
|
mul.f32 %f845, %f844, 0f437F0000;
|
|
cvt.rzi.u32.f32 %r312, %f845;
|
|
cvt.u16.u32 %rs32, %r312;
|
|
cvt.u16.u32 %rs33, %r311;
|
|
cvt.u16.u32 %rs34, %r310;
|
|
mov.u16 %rs35, 255;
|
|
st.v4.u8 [%rd107], {%rs34, %rs33, %rs32, %rs35};
|
|
|
|
BB0_141:
|
|
ret;
|
|
}
|
|
|
|
|