1817 lines
52 KiB
Plaintext
1817 lines
52 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 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 1 .b8 rnd_seeds2[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<143>;
|
|
.reg .b16 %rs<47>;
|
|
.reg .f32 %f<853>;
|
|
.reg .b32 %r<362>;
|
|
.reg .b64 %rd<148>;
|
|
|
|
|
|
mov.u64 %rd147, __local_depot0;
|
|
cvta.local.u64 %SP, %rd147;
|
|
ld.global.u32 %r1, [lightMeshBufferSize];
|
|
abs.s32 %r2, %r1;
|
|
ld.global.u32 %r3, [samples];
|
|
ld.global.v2.u32 {%r105, %r106}, [pixelID];
|
|
cvt.u64.u32 %rd22, %r105;
|
|
cvt.u64.u32 %rd23, %r106;
|
|
mov.u64 %rd26, uvnormal;
|
|
cvta.global.u64 %rd21, %rd26;
|
|
mov.u32 %r103, 2;
|
|
mov.u32 %r104, 4;
|
|
mov.u64 %rd25, 0;
|
|
// inline asm
|
|
call (%rd20), _rt_buffer_get_64, (%rd21, %r103, %r104, %rd22, %rd23, %rd25, %rd25);
|
|
// inline asm
|
|
ld.u32 %r4, [%rd20];
|
|
shr.u32 %r109, %r4, 16;
|
|
cvt.u16.u32 %rs1, %r109;
|
|
and.b16 %rs3, %rs1, 255;
|
|
cvt.u16.u32 %rs4, %r4;
|
|
or.b16 %rs5, %rs4, %rs3;
|
|
setp.eq.s16 %p7, %rs5, 0;
|
|
mov.f32 %f807, 0f00000000;
|
|
mov.f32 %f808, %f807;
|
|
mov.f32 %f809, %f807;
|
|
@%p7 bra BB0_2;
|
|
|
|
ld.u8 %rs6, [%rd20+1];
|
|
and.b16 %rs8, %rs4, 255;
|
|
cvt.rn.f32.u16 %f173, %rs8;
|
|
div.rn.f32 %f174, %f173, 0f437F0000;
|
|
fma.rn.f32 %f175, %f174, 0f40000000, 0fBF800000;
|
|
cvt.rn.f32.u16 %f176, %rs6;
|
|
div.rn.f32 %f177, %f176, 0f437F0000;
|
|
fma.rn.f32 %f178, %f177, 0f40000000, 0fBF800000;
|
|
cvt.rn.f32.u16 %f179, %rs3;
|
|
div.rn.f32 %f180, %f179, 0f437F0000;
|
|
fma.rn.f32 %f181, %f180, 0f40000000, 0fBF800000;
|
|
mul.f32 %f182, %f178, %f178;
|
|
fma.rn.f32 %f183, %f175, %f175, %f182;
|
|
fma.rn.f32 %f184, %f181, %f181, %f183;
|
|
sqrt.rn.f32 %f185, %f184;
|
|
rcp.rn.f32 %f186, %f185;
|
|
mul.f32 %f807, %f175, %f186;
|
|
mul.f32 %f808, %f178, %f186;
|
|
mul.f32 %f809, %f181, %f186;
|
|
|
|
BB0_2:
|
|
ld.global.v2.u32 {%r110, %r111}, [pixelID];
|
|
ld.global.v2.u32 {%r113, %r114}, [tileInfo];
|
|
add.s32 %r5, %r110, %r113;
|
|
add.s32 %r6, %r111, %r114;
|
|
setp.eq.f32 %p8, %f808, 0f00000000;
|
|
setp.eq.f32 %p9, %f807, 0f00000000;
|
|
and.pred %p10, %p9, %p8;
|
|
setp.eq.f32 %p11, %f809, 0f00000000;
|
|
and.pred %p12, %p10, %p11;
|
|
@%p12 bra BB0_130;
|
|
bra.uni BB0_3;
|
|
|
|
BB0_130:
|
|
ld.global.u32 %r361, [imageEnabled];
|
|
and.b32 %r312, %r361, 1;
|
|
setp.eq.b32 %p138, %r312, 1;
|
|
@!%p138 bra BB0_132;
|
|
bra.uni BB0_131;
|
|
|
|
BB0_131:
|
|
cvt.u64.u32 %rd109, %r5;
|
|
cvt.u64.u32 %rd110, %r6;
|
|
mov.u64 %rd113, image;
|
|
cvta.global.u64 %rd108, %rd113;
|
|
// inline asm
|
|
call (%rd107), _rt_buffer_get_64, (%rd108, %r103, %r104, %rd109, %rd110, %rd25, %rd25);
|
|
// inline asm
|
|
mov.u16 %rs31, 0;
|
|
st.v4.u8 [%rd107], {%rs31, %rs31, %rs31, %rs31};
|
|
ld.global.u32 %r361, [imageEnabled];
|
|
|
|
BB0_132:
|
|
and.b32 %r315, %r361, 8;
|
|
setp.eq.s32 %p139, %r315, 0;
|
|
@%p139 bra BB0_134;
|
|
|
|
cvt.u64.u32 %rd116, %r5;
|
|
cvt.u64.u32 %rd117, %r6;
|
|
mov.u64 %rd120, image_Mask;
|
|
cvta.global.u64 %rd115, %rd120;
|
|
// inline asm
|
|
call (%rd114), _rt_buffer_get_64, (%rd115, %r103, %r103, %rd116, %rd117, %rd25, %rd25);
|
|
// inline asm
|
|
mov.f32 %f790, 0f00000000;
|
|
cvt.rzi.u32.f32 %r318, %f790;
|
|
cvt.u16.u32 %rs32, %r318;
|
|
mov.u16 %rs33, 0;
|
|
st.v2.u8 [%rd114], {%rs32, %rs33};
|
|
ld.global.u32 %r361, [imageEnabled];
|
|
|
|
BB0_134:
|
|
and.b32 %r319, %r361, 4;
|
|
setp.eq.s32 %p140, %r319, 0;
|
|
@%p140 bra BB0_138;
|
|
|
|
ld.global.u32 %r320, [additive];
|
|
setp.eq.s32 %p141, %r320, 0;
|
|
cvt.u64.u32 %rd18, %r5;
|
|
cvt.u64.u32 %rd19, %r6;
|
|
@%p141 bra BB0_137;
|
|
|
|
mov.u64 %rd133, image_HDR;
|
|
cvta.global.u64 %rd122, %rd133;
|
|
mov.u32 %r324, 8;
|
|
// inline asm
|
|
call (%rd121), _rt_buffer_get_64, (%rd122, %r103, %r324, %rd18, %rd19, %rd25, %rd25);
|
|
// inline asm
|
|
ld.v4.u16 {%rs40, %rs41, %rs42, %rs43}, [%rd121];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f791, %rs40;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f792, %rs41;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f793, %rs42;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd127), _rt_buffer_get_64, (%rd122, %r103, %r324, %rd18, %rd19, %rd25, %rd25);
|
|
// inline asm
|
|
add.f32 %f794, %f791, 0f00000000;
|
|
add.f32 %f795, %f792, 0f00000000;
|
|
add.f32 %f796, %f793, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs39, %f796;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs38, %f795;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs37, %f794;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs44, 0;
|
|
st.v4.u16 [%rd127], {%rs37, %rs38, %rs39, %rs44};
|
|
bra.uni BB0_138;
|
|
|
|
BB0_3:
|
|
ld.global.v2.u32 {%r123, %r124}, [pixelID];
|
|
cvt.u64.u32 %rd29, %r123;
|
|
cvt.u64.u32 %rd30, %r124;
|
|
mov.u64 %rd39, uvpos;
|
|
cvta.global.u64 %rd28, %rd39;
|
|
mov.u32 %r119, 12;
|
|
// inline asm
|
|
call (%rd27), _rt_buffer_get_64, (%rd28, %r103, %r119, %rd29, %rd30, %rd25, %rd25);
|
|
// inline asm
|
|
ld.f32 %f9, [%rd27+8];
|
|
ld.f32 %f8, [%rd27+4];
|
|
ld.f32 %f7, [%rd27];
|
|
mul.f32 %f189, %f7, 0f3456BF95;
|
|
mul.f32 %f190, %f8, 0f3456BF95;
|
|
mul.f32 %f191, %f9, 0f3456BF95;
|
|
abs.f32 %f10, %f807;
|
|
div.rn.f32 %f192, %f189, %f10;
|
|
abs.f32 %f193, %f808;
|
|
div.rn.f32 %f194, %f190, %f193;
|
|
abs.f32 %f11, %f809;
|
|
div.rn.f32 %f195, %f191, %f11;
|
|
abs.f32 %f196, %f192;
|
|
abs.f32 %f197, %f194;
|
|
abs.f32 %f198, %f195;
|
|
mov.f32 %f199, 0f38D1B717;
|
|
max.f32 %f200, %f196, %f199;
|
|
max.f32 %f201, %f197, %f199;
|
|
max.f32 %f202, %f198, %f199;
|
|
fma.rn.f32 %f12, %f807, %f200, %f7;
|
|
fma.rn.f32 %f13, %f808, %f201, %f8;
|
|
fma.rn.f32 %f14, %f809, %f202, %f9;
|
|
ld.global.v2.u32 {%r127, %r128}, [pixelID];
|
|
cvt.u64.u32 %rd35, %r127;
|
|
cvt.u64.u32 %rd36, %r128;
|
|
mov.u64 %rd40, rnd_seeds;
|
|
cvta.global.u64 %rd34, %rd40;
|
|
// inline asm
|
|
call (%rd33), _rt_buffer_get_64, (%rd34, %r103, %r104, %rd35, %rd36, %rd25, %rd25);
|
|
// inline asm
|
|
ld.u32 %r337, [%rd33];
|
|
mov.pred %p13, 0;
|
|
mov.f32 %f19, 0f00000000;
|
|
mov.u32 %r9, 0;
|
|
setp.lt.s32 %p14, %r2, 1;
|
|
@%p14 bra BB0_4;
|
|
|
|
ld.global.f32 %f15, [lightPointSize];
|
|
mul.f32 %f16, %f12, 0f3456BF95;
|
|
mul.f32 %f17, %f13, 0f3456BF95;
|
|
mul.f32 %f18, %f14, 0f3456BF95;
|
|
mov.f32 %f19, 0f00000000;
|
|
mov.u32 %r330, 0;
|
|
abs.f32 %f328, %f17;
|
|
abs.f32 %f329, %f16;
|
|
max.f32 %f330, %f329, %f328;
|
|
abs.f32 %f331, %f18;
|
|
max.f32 %f332, %f330, %f331;
|
|
mov.u32 %r9, %r330;
|
|
mov.f32 %f20, %f19;
|
|
|
|
BB0_6:
|
|
shl.b32 %r10, %r330, 1;
|
|
cvt.s64.s32 %rd43, %r10;
|
|
mov.u64 %rd47, lightMeshBuffer;
|
|
cvta.global.u64 %rd42, %rd47;
|
|
mov.u32 %r133, 1;
|
|
// inline asm
|
|
call (%rd41), _rt_buffer_get_64, (%rd42, %r133, %r119, %rd43, %rd25, %rd25, %rd25);
|
|
// inline asm
|
|
ld.f32 %f205, [%rd41];
|
|
sub.f32 %f206, %f205, %f7;
|
|
ld.f32 %f207, [%rd41+4];
|
|
sub.f32 %f208, %f207, %f8;
|
|
ld.f32 %f209, [%rd41+8];
|
|
sub.f32 %f210, %f209, %f9;
|
|
mul.f32 %f211, %f208, %f208;
|
|
fma.rn.f32 %f212, %f206, %f206, %f211;
|
|
fma.rn.f32 %f213, %f210, %f210, %f212;
|
|
sqrt.rn.f32 %f21, %f213;
|
|
rcp.rn.f32 %f214, %f21;
|
|
mul.f32 %f22, %f206, %f214;
|
|
mul.f32 %f23, %f208, %f214;
|
|
mul.f32 %f24, %f210, %f214;
|
|
mul.f32 %f215, %f808, %f23;
|
|
fma.rn.f32 %f216, %f807, %f22, %f215;
|
|
fma.rn.f32 %f25, %f809, %f24, %f216;
|
|
setp.leu.f32 %p15, %f25, 0f00000000;
|
|
@%p15 bra BB0_23;
|
|
|
|
setp.eq.s32 %p16, %r3, 0;
|
|
mul.f32 %f217, %f21, %f21;
|
|
div.rn.f32 %f26, %f15, %f217;
|
|
ld.global.v2.u32 {%r137, %r138}, [pixelID];
|
|
cvt.u64.u32 %rd54, %r138;
|
|
cvt.u64.u32 %rd55, %r137;
|
|
and.b64 %rd50, %rd55, 255;
|
|
and.b64 %rd51, %rd54, 255;
|
|
mov.u64 %rd56, rnd_seeds2;
|
|
cvta.global.u64 %rd49, %rd56;
|
|
mov.u32 %r136, 3;
|
|
// inline asm
|
|
call (%rd48), _rt_buffer_get_64, (%rd49, %r103, %r136, %rd50, %rd51, %rd25, %rd25);
|
|
// inline asm
|
|
@%p16 bra BB0_9;
|
|
|
|
ld.u8 %rs10, [%rd48];
|
|
cvt.rn.f32.u16 %f218, %rs10;
|
|
div.rn.f32 %f219, %f218, 0fC37F0000;
|
|
fma.rn.f32 %f220, %f219, 0f3F333333, 0f3F800000;
|
|
setp.gt.f32 %p18, %f26, %f220;
|
|
mov.pred %p142, -1;
|
|
@%p18 bra BB0_25;
|
|
|
|
BB0_9:
|
|
ld.global.f32 %f223, [lightInvCutoff];
|
|
mul.f32 %f27, %f21, %f223;
|
|
mov.f32 %f227, 0f40800000;
|
|
abs.f32 %f29, %f27;
|
|
setp.lt.f32 %p19, %f29, 0f00800000;
|
|
mul.f32 %f229, %f29, 0f4B800000;
|
|
selp.f32 %f230, 0fC3170000, 0fC2FE0000, %p19;
|
|
selp.f32 %f231, %f229, %f29, %p19;
|
|
mov.b32 %r141, %f231;
|
|
and.b32 %r142, %r141, 8388607;
|
|
or.b32 %r143, %r142, 1065353216;
|
|
mov.b32 %f232, %r143;
|
|
shr.u32 %r144, %r141, 23;
|
|
cvt.rn.f32.u32 %f233, %r144;
|
|
add.f32 %f234, %f230, %f233;
|
|
setp.gt.f32 %p20, %f232, 0f3FB504F3;
|
|
mul.f32 %f235, %f232, 0f3F000000;
|
|
add.f32 %f236, %f234, 0f3F800000;
|
|
selp.f32 %f237, %f235, %f232, %p20;
|
|
selp.f32 %f238, %f236, %f234, %p20;
|
|
add.f32 %f239, %f237, 0fBF800000;
|
|
add.f32 %f222, %f237, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f221,%f222;
|
|
// inline asm
|
|
add.f32 %f240, %f239, %f239;
|
|
mul.f32 %f241, %f221, %f240;
|
|
mul.f32 %f242, %f241, %f241;
|
|
mov.f32 %f243, 0f3C4CAF63;
|
|
mov.f32 %f244, 0f3B18F0FE;
|
|
fma.rn.f32 %f245, %f244, %f242, %f243;
|
|
mov.f32 %f246, 0f3DAAAABD;
|
|
fma.rn.f32 %f247, %f245, %f242, %f246;
|
|
mul.rn.f32 %f248, %f247, %f242;
|
|
mul.rn.f32 %f249, %f248, %f241;
|
|
sub.f32 %f250, %f239, %f241;
|
|
neg.f32 %f251, %f241;
|
|
add.f32 %f252, %f250, %f250;
|
|
fma.rn.f32 %f253, %f251, %f239, %f252;
|
|
mul.rn.f32 %f254, %f221, %f253;
|
|
add.f32 %f255, %f249, %f241;
|
|
sub.f32 %f256, %f241, %f255;
|
|
add.f32 %f257, %f249, %f256;
|
|
add.f32 %f258, %f254, %f257;
|
|
add.f32 %f259, %f255, %f258;
|
|
sub.f32 %f260, %f255, %f259;
|
|
add.f32 %f261, %f258, %f260;
|
|
mov.f32 %f262, 0f3F317200;
|
|
mul.rn.f32 %f263, %f238, %f262;
|
|
mov.f32 %f264, 0f35BFBE8E;
|
|
mul.rn.f32 %f265, %f238, %f264;
|
|
add.f32 %f266, %f263, %f259;
|
|
sub.f32 %f267, %f263, %f266;
|
|
add.f32 %f268, %f259, %f267;
|
|
add.f32 %f269, %f261, %f268;
|
|
add.f32 %f270, %f265, %f269;
|
|
add.f32 %f271, %f266, %f270;
|
|
sub.f32 %f272, %f266, %f271;
|
|
add.f32 %f273, %f270, %f272;
|
|
mul.rn.f32 %f274, %f227, %f271;
|
|
neg.f32 %f275, %f274;
|
|
fma.rn.f32 %f276, %f227, %f271, %f275;
|
|
fma.rn.f32 %f277, %f227, %f273, %f276;
|
|
mov.f32 %f278, 0f00000000;
|
|
fma.rn.f32 %f279, %f278, %f271, %f277;
|
|
add.rn.f32 %f280, %f274, %f279;
|
|
neg.f32 %f281, %f280;
|
|
add.rn.f32 %f282, %f274, %f281;
|
|
add.rn.f32 %f283, %f282, %f279;
|
|
mov.b32 %r145, %f280;
|
|
setp.eq.s32 %p21, %r145, 1118925336;
|
|
add.s32 %r146, %r145, -1;
|
|
mov.b32 %f284, %r146;
|
|
add.f32 %f285, %f283, 0f37000000;
|
|
selp.f32 %f286, %f284, %f280, %p21;
|
|
selp.f32 %f30, %f285, %f283, %p21;
|
|
mul.f32 %f287, %f286, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f288, %f287;
|
|
mov.f32 %f289, 0fBF317200;
|
|
fma.rn.f32 %f290, %f288, %f289, %f286;
|
|
mov.f32 %f291, 0fB5BFBE8E;
|
|
fma.rn.f32 %f292, %f288, %f291, %f290;
|
|
mul.f32 %f293, %f292, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f294, %f293;
|
|
add.f32 %f295, %f288, 0f00000000;
|
|
ex2.approx.f32 %f296, %f295;
|
|
mul.f32 %f297, %f294, %f296;
|
|
setp.lt.f32 %p22, %f286, 0fC2D20000;
|
|
selp.f32 %f298, 0f00000000, %f297, %p22;
|
|
setp.gt.f32 %p23, %f286, 0f42D20000;
|
|
selp.f32 %f812, 0f7F800000, %f298, %p23;
|
|
setp.eq.f32 %p24, %f812, 0f7F800000;
|
|
@%p24 bra BB0_11;
|
|
|
|
fma.rn.f32 %f812, %f812, %f30, %f812;
|
|
|
|
BB0_11:
|
|
mov.f32 %f803, 0f40000000;
|
|
cvt.rzi.f32.f32 %f802, %f803;
|
|
add.f32 %f801, %f802, %f802;
|
|
mov.f32 %f800, 0f40800000;
|
|
sub.f32 %f799, %f800, %f801;
|
|
abs.f32 %f798, %f799;
|
|
setp.lt.f32 %p25, %f27, 0f00000000;
|
|
setp.eq.f32 %p26, %f798, 0f3F800000;
|
|
and.pred %p1, %p25, %p26;
|
|
mov.b32 %r147, %f812;
|
|
xor.b32 %r148, %r147, -2147483648;
|
|
mov.b32 %f299, %r148;
|
|
selp.f32 %f814, %f299, %f812, %p1;
|
|
setp.eq.f32 %p27, %f27, 0f00000000;
|
|
@%p27 bra BB0_14;
|
|
bra.uni BB0_12;
|
|
|
|
BB0_14:
|
|
add.f32 %f302, %f27, %f27;
|
|
selp.f32 %f814, %f302, 0f00000000, %p26;
|
|
bra.uni BB0_15;
|
|
|
|
BB0_12:
|
|
setp.geu.f32 %p28, %f27, 0f00000000;
|
|
@%p28 bra BB0_15;
|
|
|
|
mov.f32 %f806, 0f40800000;
|
|
cvt.rzi.f32.f32 %f301, %f806;
|
|
setp.neu.f32 %p29, %f301, 0f40800000;
|
|
selp.f32 %f814, 0f7FFFFFFF, %f814, %p29;
|
|
|
|
BB0_15:
|
|
add.f32 %f303, %f29, 0f40800000;
|
|
mov.b32 %r149, %f303;
|
|
setp.lt.s32 %p31, %r149, 2139095040;
|
|
@%p31 bra BB0_20;
|
|
|
|
setp.gtu.f32 %p32, %f29, 0f7F800000;
|
|
@%p32 bra BB0_19;
|
|
bra.uni BB0_17;
|
|
|
|
BB0_19:
|
|
add.f32 %f814, %f27, 0f40800000;
|
|
bra.uni BB0_20;
|
|
|
|
BB0_17:
|
|
setp.neu.f32 %p33, %f29, 0f7F800000;
|
|
@%p33 bra BB0_20;
|
|
|
|
selp.f32 %f814, 0fFF800000, 0f7F800000, %p1;
|
|
|
|
BB0_20:
|
|
mov.u32 %r328, 1;
|
|
mov.u64 %rd142, lightMeshBuffer;
|
|
cvta.global.u64 %rd141, %rd142;
|
|
shl.b32 %r327, %r330, 1;
|
|
mov.f32 %f304, 0f3F800000;
|
|
sub.f32 %f305, %f304, %f814;
|
|
setp.eq.f32 %p34, %f27, 0f3F800000;
|
|
selp.f32 %f306, 0f00000000, %f305, %p34;
|
|
cvt.sat.f32.f32 %f307, %f306;
|
|
mul.f32 %f308, %f26, %f307;
|
|
add.s32 %r152, %r327, 1;
|
|
cvt.s64.s32 %rd59, %r152;
|
|
// inline asm
|
|
call (%rd57), _rt_buffer_get_64, (%rd141, %r328, %r119, %rd59, %rd25, %rd25, %rd25);
|
|
// inline asm
|
|
ld.f32 %f309, [%rd57];
|
|
mul.f32 %f310, %f22, %f309;
|
|
ld.f32 %f311, [%rd57+4];
|
|
mul.f32 %f312, %f23, %f311;
|
|
neg.f32 %f313, %f312;
|
|
sub.f32 %f314, %f313, %f310;
|
|
ld.f32 %f315, [%rd57+8];
|
|
mul.f32 %f316, %f24, %f315;
|
|
sub.f32 %f317, %f314, %f316;
|
|
cvt.sat.f32.f32 %f318, %f317;
|
|
mul.f32 %f41, %f308, %f318;
|
|
cvt.sat.f32.f32 %f42, %f25;
|
|
mul.f32 %f319, %f41, %f42;
|
|
setp.leu.f32 %p35, %f319, 0f3727C5AC;
|
|
@%p35 bra BB0_22;
|
|
|
|
mov.u32 %r329, 1;
|
|
add.u64 %rd64, %SP, 28;
|
|
cvta.to.local.u64 %rd65, %rd64;
|
|
max.f32 %f326, %f332, %f199;
|
|
sub.f32 %f327, %f21, %f326;
|
|
mov.u32 %r156, 1065353216;
|
|
st.local.u32 [%rd65], %r156;
|
|
ld.global.u32 %r153, [root];
|
|
// inline asm
|
|
call _rt_trace_64, (%r153, %f12, %f13, %f14, %f22, %f23, %f24, %r329, %f326, %f327, %rd64, %r104);
|
|
// inline asm
|
|
ld.local.f32 %f334, [%rd65];
|
|
mul.f32 %f335, %f41, 0f3EA2F983;
|
|
mul.f32 %f336, %f335, %f334;
|
|
fma.rn.f32 %f20, %f42, %f336, %f20;
|
|
add.f32 %f19, %f19, %f334;
|
|
|
|
BB0_22:
|
|
add.s32 %r9, %r9, 1;
|
|
|
|
BB0_23:
|
|
add.s32 %r330, %r330, 1;
|
|
setp.lt.s32 %p37, %r330, %r2;
|
|
@%p37 bra BB0_6;
|
|
bra.uni BB0_24;
|
|
|
|
BB0_4:
|
|
mov.f32 %f20, %f19;
|
|
|
|
BB0_24:
|
|
mov.pred %p142, %p13;
|
|
|
|
BB0_25:
|
|
cvt.rn.f32.s32 %f337, %r9;
|
|
mov.f32 %f338, 0f3F800000;
|
|
max.f32 %f339, %f337, %f338;
|
|
div.rn.f32 %f839, %f20, %f339;
|
|
div.rn.f32 %f840, %f19, %f339;
|
|
@!%p142 bra BB0_78;
|
|
bra.uni BB0_26;
|
|
|
|
BB0_26:
|
|
abs.f32 %f805, %f809;
|
|
abs.f32 %f804, %f807;
|
|
setp.gt.f32 %p38, %f804, %f805;
|
|
neg.f32 %f342, %f808;
|
|
selp.f32 %f343, %f342, 0f00000000, %p38;
|
|
neg.f32 %f344, %f809;
|
|
selp.f32 %f345, %f807, %f344, %p38;
|
|
selp.f32 %f346, 0f00000000, %f808, %p38;
|
|
mul.f32 %f347, %f345, %f345;
|
|
fma.rn.f32 %f348, %f343, %f343, %f347;
|
|
fma.rn.f32 %f349, %f346, %f346, %f348;
|
|
sqrt.rn.f32 %f350, %f349;
|
|
rcp.rn.f32 %f351, %f350;
|
|
mul.f32 %f53, %f343, %f351;
|
|
mul.f32 %f54, %f345, %f351;
|
|
mul.f32 %f55, %f346, %f351;
|
|
mov.f32 %f823, 0f00000000;
|
|
setp.lt.s32 %p39, %r3, 1;
|
|
mov.f32 %f824, %f823;
|
|
@%p39 bra BB0_77;
|
|
|
|
cvt.rn.f32.s32 %f354, %r3;
|
|
rcp.rn.f32 %f56, %f354;
|
|
add.u64 %rd66, %SP, 0;
|
|
cvta.to.local.u64 %rd3, %rd66;
|
|
mul.f32 %f57, %f12, 0f3456BF95;
|
|
mul.f32 %f58, %f13, 0f3456BF95;
|
|
mul.f32 %f59, %f14, 0f3456BF95;
|
|
add.u64 %rd67, %SP, 32;
|
|
cvta.to.local.u64 %rd4, %rd67;
|
|
mul.f32 %f355, %f807, %f54;
|
|
mul.f32 %f356, %f808, %f53;
|
|
sub.f32 %f60, %f356, %f355;
|
|
mul.f32 %f357, %f809, %f53;
|
|
mul.f32 %f358, %f807, %f55;
|
|
sub.f32 %f61, %f358, %f357;
|
|
mul.f32 %f359, %f808, %f55;
|
|
mul.f32 %f360, %f809, %f54;
|
|
sub.f32 %f62, %f360, %f359;
|
|
mov.f32 %f823, 0f00000000;
|
|
mov.u32 %r157, 0;
|
|
abs.f32 %f361, %f58;
|
|
abs.f32 %f362, %f57;
|
|
max.f32 %f363, %f362, %f361;
|
|
abs.f32 %f364, %f59;
|
|
max.f32 %f365, %f363, %f364;
|
|
mov.u32 %r334, %r157;
|
|
mov.f32 %f824, %f823;
|
|
|
|
BB0_28:
|
|
cvt.rn.f32.s32 %f65, %r334;
|
|
max.f32 %f66, %f365, %f199;
|
|
mov.u32 %r336, %r157;
|
|
|
|
BB0_29:
|
|
mad.lo.s32 %r159, %r337, 1664525, 1013904223;
|
|
and.b32 %r160, %r159, 16777215;
|
|
cvt.rn.f32.u32 %f367, %r160;
|
|
fma.rn.f32 %f368, %f367, 0f33800000, %f65;
|
|
mul.f32 %f369, %f56, %f368;
|
|
mad.lo.s32 %r337, %r159, 1664525, 1013904223;
|
|
and.b32 %r161, %r337, 16777215;
|
|
cvt.rn.f32.u32 %f370, %r161;
|
|
cvt.rn.f32.s32 %f371, %r336;
|
|
fma.rn.f32 %f372, %f370, 0f33800000, %f371;
|
|
mul.f32 %f373, %f56, %f372;
|
|
sqrt.rn.f32 %f69, %f369;
|
|
mul.f32 %f831, %f373, 0f40C90FDB;
|
|
abs.f32 %f71, %f831;
|
|
setp.neu.f32 %p40, %f71, 0f7F800000;
|
|
mov.f32 %f825, %f831;
|
|
@%p40 bra BB0_31;
|
|
|
|
mov.f32 %f374, 0f00000000;
|
|
mul.rn.f32 %f825, %f831, %f374;
|
|
|
|
BB0_31:
|
|
mul.f32 %f375, %f825, 0f3F22F983;
|
|
cvt.rni.s32.f32 %r347, %f375;
|
|
cvt.rn.f32.s32 %f376, %r347;
|
|
neg.f32 %f377, %f376;
|
|
mov.f32 %f378, 0f3FC90FDA;
|
|
fma.rn.f32 %f379, %f377, %f378, %f825;
|
|
mov.f32 %f380, 0f33A22168;
|
|
fma.rn.f32 %f381, %f377, %f380, %f379;
|
|
mov.f32 %f382, 0f27C234C5;
|
|
fma.rn.f32 %f826, %f377, %f382, %f381;
|
|
abs.f32 %f383, %f825;
|
|
setp.leu.f32 %p41, %f383, 0f47CE4780;
|
|
@%p41 bra BB0_42;
|
|
|
|
mov.b32 %r21, %f825;
|
|
shr.u32 %r22, %r21, 23;
|
|
shl.b32 %r164, %r21, 8;
|
|
or.b32 %r23, %r164, -2147483648;
|
|
mov.u32 %r339, 0;
|
|
mov.u64 %rd143, __cudart_i2opi_f;
|
|
mov.u32 %r338, -6;
|
|
mov.u64 %rd144, %rd3;
|
|
|
|
BB0_33:
|
|
.pragma "nounroll";
|
|
ld.const.u32 %r167, [%rd143];
|
|
// inline asm
|
|
{
|
|
mad.lo.cc.u32 %r165, %r167, %r23, %r339;
|
|
madc.hi.u32 %r339, %r167, %r23, 0;
|
|
}
|
|
// inline asm
|
|
st.local.u32 [%rd144], %r165;
|
|
add.s64 %rd144, %rd144, 4;
|
|
add.s64 %rd143, %rd143, 4;
|
|
add.s32 %r338, %r338, 1;
|
|
setp.ne.s32 %p42, %r338, 0;
|
|
@%p42 bra BB0_33;
|
|
|
|
and.b32 %r170, %r22, 255;
|
|
add.s32 %r171, %r170, -128;
|
|
shr.u32 %r172, %r171, 5;
|
|
and.b32 %r28, %r21, -2147483648;
|
|
st.local.u32 [%rd3+24], %r339;
|
|
mov.u32 %r173, 6;
|
|
sub.s32 %r174, %r173, %r172;
|
|
mul.wide.s32 %rd69, %r174, 4;
|
|
add.s64 %rd9, %rd3, %rd69;
|
|
ld.local.u32 %r340, [%rd9];
|
|
ld.local.u32 %r341, [%rd9+-4];
|
|
and.b32 %r31, %r22, 31;
|
|
setp.eq.s32 %p43, %r31, 0;
|
|
@%p43 bra BB0_36;
|
|
|
|
mov.u32 %r175, 32;
|
|
sub.s32 %r176, %r175, %r31;
|
|
shr.u32 %r177, %r341, %r176;
|
|
shl.b32 %r178, %r340, %r31;
|
|
add.s32 %r340, %r177, %r178;
|
|
ld.local.u32 %r179, [%rd9+-8];
|
|
shr.u32 %r180, %r179, %r176;
|
|
shl.b32 %r181, %r341, %r31;
|
|
add.s32 %r341, %r180, %r181;
|
|
|
|
BB0_36:
|
|
shr.u32 %r182, %r341, 30;
|
|
shl.b32 %r183, %r340, 2;
|
|
add.s32 %r342, %r182, %r183;
|
|
shl.b32 %r37, %r341, 2;
|
|
shr.u32 %r184, %r342, 31;
|
|
shr.u32 %r185, %r340, 30;
|
|
add.s32 %r38, %r184, %r185;
|
|
setp.eq.s32 %p44, %r184, 0;
|
|
@%p44 bra BB0_37;
|
|
bra.uni BB0_38;
|
|
|
|
BB0_37:
|
|
mov.u32 %r343, %r28;
|
|
mov.u32 %r344, %r37;
|
|
bra.uni BB0_39;
|
|
|
|
BB0_38:
|
|
not.b32 %r186, %r342;
|
|
neg.s32 %r344, %r37;
|
|
setp.eq.s32 %p45, %r37, 0;
|
|
selp.u32 %r187, 1, 0, %p45;
|
|
add.s32 %r342, %r187, %r186;
|
|
xor.b32 %r343, %r28, -2147483648;
|
|
|
|
BB0_39:
|
|
clz.b32 %r346, %r342;
|
|
setp.eq.s32 %p46, %r346, 0;
|
|
shl.b32 %r188, %r342, %r346;
|
|
mov.u32 %r189, 32;
|
|
sub.s32 %r190, %r189, %r346;
|
|
shr.u32 %r191, %r344, %r190;
|
|
add.s32 %r192, %r191, %r188;
|
|
selp.b32 %r46, %r342, %r192, %p46;
|
|
mov.u32 %r193, -921707870;
|
|
mul.hi.u32 %r345, %r46, %r193;
|
|
setp.eq.s32 %p47, %r28, 0;
|
|
neg.s32 %r194, %r38;
|
|
selp.b32 %r347, %r38, %r194, %p47;
|
|
setp.lt.s32 %p48, %r345, 1;
|
|
@%p48 bra BB0_41;
|
|
|
|
mul.lo.s32 %r195, %r46, -921707870;
|
|
shr.u32 %r196, %r195, 31;
|
|
shl.b32 %r197, %r345, 1;
|
|
add.s32 %r345, %r196, %r197;
|
|
add.s32 %r346, %r346, 1;
|
|
|
|
BB0_41:
|
|
mov.u32 %r198, 126;
|
|
sub.s32 %r199, %r198, %r346;
|
|
shl.b32 %r200, %r199, 23;
|
|
add.s32 %r201, %r345, 1;
|
|
shr.u32 %r202, %r201, 7;
|
|
add.s32 %r203, %r202, 1;
|
|
shr.u32 %r204, %r203, 1;
|
|
add.s32 %r205, %r204, %r200;
|
|
or.b32 %r206, %r205, %r343;
|
|
mov.b32 %f826, %r206;
|
|
|
|
BB0_42:
|
|
mul.rn.f32 %f77, %f826, %f826;
|
|
add.s32 %r54, %r347, 1;
|
|
and.b32 %r55, %r54, 1;
|
|
setp.eq.s32 %p49, %r55, 0;
|
|
@%p49 bra BB0_44;
|
|
bra.uni BB0_43;
|
|
|
|
BB0_44:
|
|
mov.f32 %f386, 0f3C08839E;
|
|
mov.f32 %f387, 0fB94CA1F9;
|
|
fma.rn.f32 %f827, %f387, %f77, %f386;
|
|
bra.uni BB0_45;
|
|
|
|
BB0_43:
|
|
mov.f32 %f384, 0fBAB6061A;
|
|
mov.f32 %f385, 0f37CCF5CE;
|
|
fma.rn.f32 %f827, %f385, %f77, %f384;
|
|
|
|
BB0_45:
|
|
@%p49 bra BB0_47;
|
|
bra.uni BB0_46;
|
|
|
|
BB0_47:
|
|
mov.f32 %f391, 0fBE2AAAA3;
|
|
fma.rn.f32 %f392, %f827, %f77, %f391;
|
|
mov.f32 %f393, 0f00000000;
|
|
fma.rn.f32 %f828, %f392, %f77, %f393;
|
|
bra.uni BB0_48;
|
|
|
|
BB0_46:
|
|
mov.f32 %f388, 0f3D2AAAA5;
|
|
fma.rn.f32 %f389, %f827, %f77, %f388;
|
|
mov.f32 %f390, 0fBF000000;
|
|
fma.rn.f32 %f828, %f389, %f77, %f390;
|
|
|
|
BB0_48:
|
|
fma.rn.f32 %f829, %f828, %f826, %f826;
|
|
@%p49 bra BB0_50;
|
|
|
|
fma.rn.f32 %f829, %f828, %f77, %f338;
|
|
|
|
BB0_50:
|
|
and.b32 %r207, %r54, 2;
|
|
setp.eq.s32 %p52, %r207, 0;
|
|
@%p52 bra BB0_52;
|
|
|
|
mov.f32 %f395, 0f00000000;
|
|
mov.f32 %f396, 0fBF800000;
|
|
fma.rn.f32 %f829, %f829, %f396, %f395;
|
|
|
|
BB0_52:
|
|
@%p40 bra BB0_54;
|
|
|
|
mov.f32 %f397, 0f00000000;
|
|
mul.rn.f32 %f831, %f831, %f397;
|
|
|
|
BB0_54:
|
|
mul.f32 %f398, %f831, 0f3F22F983;
|
|
cvt.rni.s32.f32 %r357, %f398;
|
|
cvt.rn.f32.s32 %f399, %r357;
|
|
neg.f32 %f400, %f399;
|
|
fma.rn.f32 %f402, %f400, %f378, %f831;
|
|
fma.rn.f32 %f404, %f400, %f380, %f402;
|
|
fma.rn.f32 %f832, %f400, %f382, %f404;
|
|
abs.f32 %f406, %f831;
|
|
setp.leu.f32 %p54, %f406, 0f47CE4780;
|
|
@%p54 bra BB0_65;
|
|
|
|
mov.b32 %r57, %f831;
|
|
shr.u32 %r58, %r57, 23;
|
|
shl.b32 %r210, %r57, 8;
|
|
or.b32 %r59, %r210, -2147483648;
|
|
mov.u32 %r349, 0;
|
|
mov.u64 %rd145, __cudart_i2opi_f;
|
|
mov.u32 %r348, -6;
|
|
mov.u64 %rd146, %rd3;
|
|
|
|
BB0_56:
|
|
.pragma "nounroll";
|
|
ld.const.u32 %r213, [%rd145];
|
|
// inline asm
|
|
{
|
|
mad.lo.cc.u32 %r211, %r213, %r59, %r349;
|
|
madc.hi.u32 %r349, %r213, %r59, 0;
|
|
}
|
|
// inline asm
|
|
st.local.u32 [%rd146], %r211;
|
|
add.s64 %rd146, %rd146, 4;
|
|
add.s64 %rd145, %rd145, 4;
|
|
add.s32 %r348, %r348, 1;
|
|
setp.ne.s32 %p55, %r348, 0;
|
|
@%p55 bra BB0_56;
|
|
|
|
and.b32 %r216, %r58, 255;
|
|
add.s32 %r217, %r216, -128;
|
|
shr.u32 %r218, %r217, 5;
|
|
and.b32 %r64, %r57, -2147483648;
|
|
st.local.u32 [%rd3+24], %r349;
|
|
mov.u32 %r219, 6;
|
|
sub.s32 %r220, %r219, %r218;
|
|
mul.wide.s32 %rd71, %r220, 4;
|
|
add.s64 %rd14, %rd3, %rd71;
|
|
ld.local.u32 %r350, [%rd14];
|
|
ld.local.u32 %r351, [%rd14+-4];
|
|
and.b32 %r67, %r58, 31;
|
|
setp.eq.s32 %p56, %r67, 0;
|
|
@%p56 bra BB0_59;
|
|
|
|
mov.u32 %r221, 32;
|
|
sub.s32 %r222, %r221, %r67;
|
|
shr.u32 %r223, %r351, %r222;
|
|
shl.b32 %r224, %r350, %r67;
|
|
add.s32 %r350, %r223, %r224;
|
|
ld.local.u32 %r225, [%rd14+-8];
|
|
shr.u32 %r226, %r225, %r222;
|
|
shl.b32 %r227, %r351, %r67;
|
|
add.s32 %r351, %r226, %r227;
|
|
|
|
BB0_59:
|
|
shr.u32 %r228, %r351, 30;
|
|
shl.b32 %r229, %r350, 2;
|
|
add.s32 %r352, %r228, %r229;
|
|
shl.b32 %r73, %r351, 2;
|
|
shr.u32 %r230, %r352, 31;
|
|
shr.u32 %r231, %r350, 30;
|
|
add.s32 %r74, %r230, %r231;
|
|
setp.eq.s32 %p57, %r230, 0;
|
|
@%p57 bra BB0_60;
|
|
bra.uni BB0_61;
|
|
|
|
BB0_60:
|
|
mov.u32 %r353, %r64;
|
|
mov.u32 %r354, %r73;
|
|
bra.uni BB0_62;
|
|
|
|
BB0_61:
|
|
not.b32 %r232, %r352;
|
|
neg.s32 %r354, %r73;
|
|
setp.eq.s32 %p58, %r73, 0;
|
|
selp.u32 %r233, 1, 0, %p58;
|
|
add.s32 %r352, %r233, %r232;
|
|
xor.b32 %r353, %r64, -2147483648;
|
|
|
|
BB0_62:
|
|
clz.b32 %r356, %r352;
|
|
setp.eq.s32 %p59, %r356, 0;
|
|
shl.b32 %r234, %r352, %r356;
|
|
mov.u32 %r235, 32;
|
|
sub.s32 %r236, %r235, %r356;
|
|
shr.u32 %r237, %r354, %r236;
|
|
add.s32 %r238, %r237, %r234;
|
|
selp.b32 %r82, %r352, %r238, %p59;
|
|
mov.u32 %r239, -921707870;
|
|
mul.hi.u32 %r355, %r82, %r239;
|
|
setp.eq.s32 %p60, %r64, 0;
|
|
neg.s32 %r240, %r74;
|
|
selp.b32 %r357, %r74, %r240, %p60;
|
|
setp.lt.s32 %p61, %r355, 1;
|
|
@%p61 bra BB0_64;
|
|
|
|
mul.lo.s32 %r241, %r82, -921707870;
|
|
shr.u32 %r242, %r241, 31;
|
|
shl.b32 %r243, %r355, 1;
|
|
add.s32 %r355, %r242, %r243;
|
|
add.s32 %r356, %r356, 1;
|
|
|
|
BB0_64:
|
|
mov.u32 %r244, 126;
|
|
sub.s32 %r245, %r244, %r356;
|
|
shl.b32 %r246, %r245, 23;
|
|
add.s32 %r247, %r355, 1;
|
|
shr.u32 %r248, %r247, 7;
|
|
add.s32 %r249, %r248, 1;
|
|
shr.u32 %r250, %r249, 1;
|
|
add.s32 %r251, %r250, %r246;
|
|
or.b32 %r252, %r251, %r353;
|
|
mov.b32 %f832, %r252;
|
|
|
|
BB0_65:
|
|
mul.rn.f32 %f94, %f832, %f832;
|
|
and.b32 %r90, %r357, 1;
|
|
setp.eq.s32 %p62, %r90, 0;
|
|
@%p62 bra BB0_67;
|
|
bra.uni BB0_66;
|
|
|
|
BB0_67:
|
|
mov.f32 %f409, 0f3C08839E;
|
|
mov.f32 %f410, 0fB94CA1F9;
|
|
fma.rn.f32 %f833, %f410, %f94, %f409;
|
|
bra.uni BB0_68;
|
|
|
|
BB0_66:
|
|
mov.f32 %f407, 0fBAB6061A;
|
|
mov.f32 %f408, 0f37CCF5CE;
|
|
fma.rn.f32 %f833, %f408, %f94, %f407;
|
|
|
|
BB0_68:
|
|
@%p62 bra BB0_70;
|
|
bra.uni BB0_69;
|
|
|
|
BB0_70:
|
|
mov.f32 %f414, 0fBE2AAAA3;
|
|
fma.rn.f32 %f415, %f833, %f94, %f414;
|
|
mov.f32 %f416, 0f00000000;
|
|
fma.rn.f32 %f834, %f415, %f94, %f416;
|
|
bra.uni BB0_71;
|
|
|
|
BB0_69:
|
|
mov.f32 %f411, 0f3D2AAAA5;
|
|
fma.rn.f32 %f412, %f833, %f94, %f411;
|
|
mov.f32 %f413, 0fBF000000;
|
|
fma.rn.f32 %f834, %f412, %f94, %f413;
|
|
|
|
BB0_71:
|
|
fma.rn.f32 %f835, %f834, %f832, %f832;
|
|
@%p62 bra BB0_73;
|
|
|
|
fma.rn.f32 %f835, %f834, %f94, %f338;
|
|
|
|
BB0_73:
|
|
and.b32 %r253, %r357, 2;
|
|
setp.eq.s32 %p65, %r253, 0;
|
|
@%p65 bra BB0_75;
|
|
|
|
mov.f32 %f418, 0f00000000;
|
|
mov.f32 %f419, 0fBF800000;
|
|
fma.rn.f32 %f835, %f835, %f419, %f418;
|
|
|
|
BB0_75:
|
|
mul.f32 %f428, %f69, %f829;
|
|
mul.f32 %f429, %f428, %f428;
|
|
sub.f32 %f431, %f338, %f429;
|
|
mul.f32 %f432, %f69, %f835;
|
|
mul.f32 %f433, %f432, %f432;
|
|
sub.f32 %f434, %f431, %f433;
|
|
mov.f32 %f435, 0f00000000;
|
|
max.f32 %f436, %f435, %f434;
|
|
sqrt.rn.f32 %f437, %f436;
|
|
mul.f32 %f438, %f53, %f432;
|
|
mul.f32 %f439, %f54, %f432;
|
|
mul.f32 %f440, %f55, %f432;
|
|
fma.rn.f32 %f441, %f62, %f428, %f438;
|
|
fma.rn.f32 %f442, %f61, %f428, %f439;
|
|
fma.rn.f32 %f443, %f60, %f428, %f440;
|
|
fma.rn.f32 %f423, %f807, %f437, %f441;
|
|
fma.rn.f32 %f424, %f808, %f437, %f442;
|
|
fma.rn.f32 %f425, %f809, %f437, %f443;
|
|
mov.u32 %r255, 0;
|
|
st.local.u32 [%rd4], %r255;
|
|
ld.global.u32 %r254, [root];
|
|
mov.f32 %f427, 0f6C4ECB8F;
|
|
// inline asm
|
|
call _rt_trace_64, (%r254, %f12, %f13, %f14, %f423, %f424, %f425, %r255, %f66, %f427, %rd67, %r104);
|
|
// inline asm
|
|
ld.local.f32 %f444, [%rd4];
|
|
setp.lt.f32 %p66, %f444, 0f00000000;
|
|
selp.f32 %f445, 0f00000000, %f444, %p66;
|
|
selp.f32 %f446, 0f00000000, 0f3F800000, %p66;
|
|
add.f32 %f823, %f823, %f446;
|
|
add.f32 %f824, %f824, %f445;
|
|
add.s32 %r336, %r336, 1;
|
|
setp.lt.s32 %p67, %r336, %r3;
|
|
@%p67 bra BB0_29;
|
|
|
|
add.s32 %r334, %r334, 1;
|
|
setp.lt.s32 %p68, %r334, %r3;
|
|
@%p68 bra BB0_28;
|
|
|
|
BB0_77:
|
|
mul.lo.s32 %r257, %r3, %r3;
|
|
cvt.rn.f32.s32 %f447, %r257;
|
|
div.rn.f32 %f839, %f824, %f447;
|
|
div.rn.f32 %f840, %f823, %f447;
|
|
|
|
BB0_78:
|
|
setp.gt.s32 %p69, %r1, -1;
|
|
selp.f32 %f114, %f840, %f839, %p69;
|
|
ld.global.u32 %r359, [imageEnabled];
|
|
and.b32 %r258, %r359, 8;
|
|
setp.eq.s32 %p70, %r258, 0;
|
|
@%p70 bra BB0_91;
|
|
|
|
cvt.u64.u32 %rd75, %r5;
|
|
cvt.u64.u32 %rd76, %r6;
|
|
mov.u64 %rd79, image_Mask;
|
|
cvta.global.u64 %rd74, %rd79;
|
|
// inline asm
|
|
call (%rd73), _rt_buffer_get_64, (%rd74, %r103, %r103, %rd75, %rd76, %rd25, %rd25);
|
|
// inline asm
|
|
mov.f32 %f450, 0f3E68BA2E;
|
|
cvt.rzi.f32.f32 %f451, %f450;
|
|
fma.rn.f32 %f452, %f451, 0fC0000000, 0f3EE8BA2E;
|
|
abs.f32 %f115, %f452;
|
|
abs.f32 %f116, %f114;
|
|
setp.lt.f32 %p71, %f116, 0f00800000;
|
|
mul.f32 %f453, %f116, 0f4B800000;
|
|
selp.f32 %f454, 0fC3170000, 0fC2FE0000, %p71;
|
|
selp.f32 %f455, %f453, %f116, %p71;
|
|
mov.b32 %r261, %f455;
|
|
and.b32 %r262, %r261, 8388607;
|
|
or.b32 %r263, %r262, 1065353216;
|
|
mov.b32 %f456, %r263;
|
|
shr.u32 %r264, %r261, 23;
|
|
cvt.rn.f32.u32 %f457, %r264;
|
|
add.f32 %f458, %f454, %f457;
|
|
setp.gt.f32 %p72, %f456, 0f3FB504F3;
|
|
mul.f32 %f459, %f456, 0f3F000000;
|
|
add.f32 %f460, %f458, 0f3F800000;
|
|
selp.f32 %f461, %f459, %f456, %p72;
|
|
selp.f32 %f462, %f460, %f458, %p72;
|
|
add.f32 %f463, %f461, 0fBF800000;
|
|
add.f32 %f449, %f461, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f448,%f449;
|
|
// inline asm
|
|
add.f32 %f464, %f463, %f463;
|
|
mul.f32 %f465, %f448, %f464;
|
|
mul.f32 %f466, %f465, %f465;
|
|
mov.f32 %f467, 0f3C4CAF63;
|
|
mov.f32 %f468, 0f3B18F0FE;
|
|
fma.rn.f32 %f469, %f468, %f466, %f467;
|
|
mov.f32 %f470, 0f3DAAAABD;
|
|
fma.rn.f32 %f471, %f469, %f466, %f470;
|
|
mul.rn.f32 %f472, %f471, %f466;
|
|
mul.rn.f32 %f473, %f472, %f465;
|
|
sub.f32 %f474, %f463, %f465;
|
|
neg.f32 %f475, %f465;
|
|
add.f32 %f476, %f474, %f474;
|
|
fma.rn.f32 %f477, %f475, %f463, %f476;
|
|
mul.rn.f32 %f478, %f448, %f477;
|
|
add.f32 %f479, %f473, %f465;
|
|
sub.f32 %f480, %f465, %f479;
|
|
add.f32 %f481, %f473, %f480;
|
|
add.f32 %f482, %f478, %f481;
|
|
add.f32 %f483, %f479, %f482;
|
|
sub.f32 %f484, %f479, %f483;
|
|
add.f32 %f485, %f482, %f484;
|
|
mov.f32 %f486, 0f3F317200;
|
|
mul.rn.f32 %f487, %f462, %f486;
|
|
mov.f32 %f488, 0f35BFBE8E;
|
|
mul.rn.f32 %f489, %f462, %f488;
|
|
add.f32 %f490, %f487, %f483;
|
|
sub.f32 %f491, %f487, %f490;
|
|
add.f32 %f492, %f483, %f491;
|
|
add.f32 %f493, %f485, %f492;
|
|
add.f32 %f494, %f489, %f493;
|
|
add.f32 %f495, %f490, %f494;
|
|
sub.f32 %f496, %f490, %f495;
|
|
add.f32 %f497, %f494, %f496;
|
|
mov.f32 %f498, 0f3EE8BA2E;
|
|
mul.rn.f32 %f499, %f498, %f495;
|
|
neg.f32 %f500, %f499;
|
|
fma.rn.f32 %f501, %f498, %f495, %f500;
|
|
fma.rn.f32 %f502, %f498, %f497, %f501;
|
|
mov.f32 %f503, 0f00000000;
|
|
fma.rn.f32 %f504, %f503, %f495, %f502;
|
|
add.rn.f32 %f505, %f499, %f504;
|
|
neg.f32 %f506, %f505;
|
|
add.rn.f32 %f507, %f499, %f506;
|
|
add.rn.f32 %f508, %f507, %f504;
|
|
mov.b32 %r265, %f505;
|
|
setp.eq.s32 %p73, %r265, 1118925336;
|
|
add.s32 %r266, %r265, -1;
|
|
mov.b32 %f509, %r266;
|
|
add.f32 %f510, %f508, 0f37000000;
|
|
selp.f32 %f511, %f509, %f505, %p73;
|
|
selp.f32 %f117, %f510, %f508, %p73;
|
|
mul.f32 %f512, %f511, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f513, %f512;
|
|
mov.f32 %f514, 0fBF317200;
|
|
fma.rn.f32 %f515, %f513, %f514, %f511;
|
|
mov.f32 %f516, 0fB5BFBE8E;
|
|
fma.rn.f32 %f517, %f513, %f516, %f515;
|
|
mul.f32 %f518, %f517, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f519, %f518;
|
|
add.f32 %f520, %f513, 0f00000000;
|
|
ex2.approx.f32 %f521, %f520;
|
|
mul.f32 %f522, %f519, %f521;
|
|
setp.lt.f32 %p74, %f511, 0fC2D20000;
|
|
selp.f32 %f523, 0f00000000, %f522, %p74;
|
|
setp.gt.f32 %p75, %f511, 0f42D20000;
|
|
selp.f32 %f841, 0f7F800000, %f523, %p75;
|
|
setp.eq.f32 %p76, %f841, 0f7F800000;
|
|
@%p76 bra BB0_81;
|
|
|
|
fma.rn.f32 %f841, %f841, %f117, %f841;
|
|
|
|
BB0_81:
|
|
setp.lt.f32 %p77, %f114, 0f00000000;
|
|
setp.eq.f32 %p78, %f115, 0f3F800000;
|
|
and.pred %p3, %p77, %p78;
|
|
mov.b32 %r267, %f841;
|
|
xor.b32 %r268, %r267, -2147483648;
|
|
mov.b32 %f524, %r268;
|
|
selp.f32 %f843, %f524, %f841, %p3;
|
|
setp.eq.f32 %p79, %f114, 0f00000000;
|
|
@%p79 bra BB0_84;
|
|
bra.uni BB0_82;
|
|
|
|
BB0_84:
|
|
add.f32 %f527, %f114, %f114;
|
|
selp.f32 %f843, %f527, 0f00000000, %p78;
|
|
bra.uni BB0_85;
|
|
|
|
BB0_137:
|
|
mov.u64 %rd140, image_HDR;
|
|
cvta.global.u64 %rd135, %rd140;
|
|
mov.u32 %r326, 8;
|
|
// inline asm
|
|
call (%rd134), _rt_buffer_get_64, (%rd135, %r103, %r326, %rd18, %rd19, %rd25, %rd25);
|
|
// inline asm
|
|
mov.f32 %f797, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs45, %f797;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs46, 0;
|
|
st.v4.u16 [%rd134], {%rs45, %rs45, %rs45, %rs46};
|
|
bra.uni BB0_138;
|
|
|
|
BB0_82:
|
|
setp.geu.f32 %p80, %f114, 0f00000000;
|
|
@%p80 bra BB0_85;
|
|
|
|
cvt.rzi.f32.f32 %f526, %f498;
|
|
setp.neu.f32 %p81, %f526, 0f3EE8BA2E;
|
|
selp.f32 %f843, 0f7FFFFFFF, %f843, %p81;
|
|
|
|
BB0_85:
|
|
add.f32 %f528, %f116, 0f3EE8BA2E;
|
|
mov.b32 %r269, %f528;
|
|
setp.lt.s32 %p83, %r269, 2139095040;
|
|
@%p83 bra BB0_90;
|
|
|
|
setp.gtu.f32 %p84, %f116, 0f7F800000;
|
|
@%p84 bra BB0_89;
|
|
bra.uni BB0_87;
|
|
|
|
BB0_89:
|
|
add.f32 %f843, %f114, 0f3EE8BA2E;
|
|
bra.uni BB0_90;
|
|
|
|
BB0_87:
|
|
setp.neu.f32 %p85, %f116, 0f7F800000;
|
|
@%p85 bra BB0_90;
|
|
|
|
selp.f32 %f843, 0fFF800000, 0f7F800000, %p3;
|
|
|
|
BB0_90:
|
|
mul.f32 %f529, %f843, 0f437F0000;
|
|
setp.eq.f32 %p86, %f114, 0f3F800000;
|
|
selp.f32 %f530, 0f437F0000, %f529, %p86;
|
|
cvt.rzi.u32.f32 %r270, %f530;
|
|
cvt.u16.u32 %rs11, %r270;
|
|
mov.u16 %rs12, 255;
|
|
st.v2.u8 [%rd73], {%rs11, %rs12};
|
|
ld.global.u32 %r359, [imageEnabled];
|
|
|
|
BB0_91:
|
|
ld.global.f32 %f531, [lightColor];
|
|
mul.f32 %f128, %f839, %f531;
|
|
ld.global.f32 %f532, [lightColor+4];
|
|
mul.f32 %f129, %f839, %f532;
|
|
ld.global.f32 %f533, [lightColor+8];
|
|
mul.f32 %f130, %f839, %f533;
|
|
and.b32 %r271, %r359, 1;
|
|
setp.eq.b32 %p87, %r271, 1;
|
|
@!%p87 bra BB0_126;
|
|
bra.uni BB0_92;
|
|
|
|
BB0_92:
|
|
mov.f32 %f536, 0f3E666666;
|
|
cvt.rzi.f32.f32 %f537, %f536;
|
|
fma.rn.f32 %f538, %f537, 0fC0000000, 0f3EE66666;
|
|
abs.f32 %f131, %f538;
|
|
abs.f32 %f132, %f128;
|
|
setp.lt.f32 %p88, %f132, 0f00800000;
|
|
mul.f32 %f539, %f132, 0f4B800000;
|
|
selp.f32 %f540, 0fC3170000, 0fC2FE0000, %p88;
|
|
selp.f32 %f541, %f539, %f132, %p88;
|
|
mov.b32 %r272, %f541;
|
|
and.b32 %r273, %r272, 8388607;
|
|
or.b32 %r274, %r273, 1065353216;
|
|
mov.b32 %f542, %r274;
|
|
shr.u32 %r275, %r272, 23;
|
|
cvt.rn.f32.u32 %f543, %r275;
|
|
add.f32 %f544, %f540, %f543;
|
|
setp.gt.f32 %p89, %f542, 0f3FB504F3;
|
|
mul.f32 %f545, %f542, 0f3F000000;
|
|
add.f32 %f546, %f544, 0f3F800000;
|
|
selp.f32 %f547, %f545, %f542, %p89;
|
|
selp.f32 %f548, %f546, %f544, %p89;
|
|
add.f32 %f549, %f547, 0fBF800000;
|
|
add.f32 %f535, %f547, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f534,%f535;
|
|
// inline asm
|
|
add.f32 %f550, %f549, %f549;
|
|
mul.f32 %f551, %f534, %f550;
|
|
mul.f32 %f552, %f551, %f551;
|
|
mov.f32 %f553, 0f3C4CAF63;
|
|
mov.f32 %f554, 0f3B18F0FE;
|
|
fma.rn.f32 %f555, %f554, %f552, %f553;
|
|
mov.f32 %f556, 0f3DAAAABD;
|
|
fma.rn.f32 %f557, %f555, %f552, %f556;
|
|
mul.rn.f32 %f558, %f557, %f552;
|
|
mul.rn.f32 %f559, %f558, %f551;
|
|
sub.f32 %f560, %f549, %f551;
|
|
neg.f32 %f561, %f551;
|
|
add.f32 %f562, %f560, %f560;
|
|
fma.rn.f32 %f563, %f561, %f549, %f562;
|
|
mul.rn.f32 %f564, %f534, %f563;
|
|
add.f32 %f565, %f559, %f551;
|
|
sub.f32 %f566, %f551, %f565;
|
|
add.f32 %f567, %f559, %f566;
|
|
add.f32 %f568, %f564, %f567;
|
|
add.f32 %f569, %f565, %f568;
|
|
sub.f32 %f570, %f565, %f569;
|
|
add.f32 %f571, %f568, %f570;
|
|
mov.f32 %f572, 0f3F317200;
|
|
mul.rn.f32 %f573, %f548, %f572;
|
|
mov.f32 %f574, 0f35BFBE8E;
|
|
mul.rn.f32 %f575, %f548, %f574;
|
|
add.f32 %f576, %f573, %f569;
|
|
sub.f32 %f577, %f573, %f576;
|
|
add.f32 %f578, %f569, %f577;
|
|
add.f32 %f579, %f571, %f578;
|
|
add.f32 %f580, %f575, %f579;
|
|
add.f32 %f581, %f576, %f580;
|
|
sub.f32 %f582, %f576, %f581;
|
|
add.f32 %f583, %f580, %f582;
|
|
mov.f32 %f584, 0f3EE66666;
|
|
mul.rn.f32 %f585, %f584, %f581;
|
|
neg.f32 %f586, %f585;
|
|
fma.rn.f32 %f587, %f584, %f581, %f586;
|
|
fma.rn.f32 %f588, %f584, %f583, %f587;
|
|
mov.f32 %f589, 0f00000000;
|
|
fma.rn.f32 %f590, %f589, %f581, %f588;
|
|
add.rn.f32 %f591, %f585, %f590;
|
|
neg.f32 %f592, %f591;
|
|
add.rn.f32 %f593, %f585, %f592;
|
|
add.rn.f32 %f594, %f593, %f590;
|
|
mov.b32 %r276, %f591;
|
|
setp.eq.s32 %p90, %r276, 1118925336;
|
|
add.s32 %r277, %r276, -1;
|
|
mov.b32 %f595, %r277;
|
|
add.f32 %f596, %f594, 0f37000000;
|
|
selp.f32 %f597, %f595, %f591, %p90;
|
|
selp.f32 %f133, %f596, %f594, %p90;
|
|
mul.f32 %f598, %f597, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f599, %f598;
|
|
mov.f32 %f600, 0fBF317200;
|
|
fma.rn.f32 %f601, %f599, %f600, %f597;
|
|
mov.f32 %f602, 0fB5BFBE8E;
|
|
fma.rn.f32 %f603, %f599, %f602, %f601;
|
|
mul.f32 %f604, %f603, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f605, %f604;
|
|
add.f32 %f606, %f599, 0f00000000;
|
|
ex2.approx.f32 %f607, %f606;
|
|
mul.f32 %f608, %f605, %f607;
|
|
setp.lt.f32 %p91, %f597, 0fC2D20000;
|
|
selp.f32 %f609, 0f00000000, %f608, %p91;
|
|
setp.gt.f32 %p92, %f597, 0f42D20000;
|
|
selp.f32 %f844, 0f7F800000, %f609, %p92;
|
|
setp.eq.f32 %p93, %f844, 0f7F800000;
|
|
@%p93 bra BB0_94;
|
|
|
|
fma.rn.f32 %f844, %f844, %f133, %f844;
|
|
|
|
BB0_94:
|
|
setp.lt.f32 %p94, %f128, 0f00000000;
|
|
setp.eq.f32 %p95, %f131, 0f3F800000;
|
|
and.pred %p4, %p94, %p95;
|
|
mov.b32 %r278, %f844;
|
|
xor.b32 %r279, %r278, -2147483648;
|
|
mov.b32 %f610, %r279;
|
|
selp.f32 %f846, %f610, %f844, %p4;
|
|
setp.eq.f32 %p96, %f128, 0f00000000;
|
|
@%p96 bra BB0_97;
|
|
bra.uni BB0_95;
|
|
|
|
BB0_97:
|
|
add.f32 %f613, %f128, %f128;
|
|
selp.f32 %f846, %f613, 0f00000000, %p95;
|
|
bra.uni BB0_98;
|
|
|
|
BB0_95:
|
|
setp.geu.f32 %p97, %f128, 0f00000000;
|
|
@%p97 bra BB0_98;
|
|
|
|
cvt.rzi.f32.f32 %f612, %f584;
|
|
setp.neu.f32 %p98, %f612, 0f3EE66666;
|
|
selp.f32 %f846, 0f7FFFFFFF, %f846, %p98;
|
|
|
|
BB0_98:
|
|
add.f32 %f614, %f132, 0f3EE66666;
|
|
mov.b32 %r280, %f614;
|
|
setp.lt.s32 %p100, %r280, 2139095040;
|
|
@%p100 bra BB0_103;
|
|
|
|
setp.gtu.f32 %p101, %f132, 0f7F800000;
|
|
@%p101 bra BB0_102;
|
|
bra.uni BB0_100;
|
|
|
|
BB0_102:
|
|
add.f32 %f846, %f128, 0f3EE66666;
|
|
bra.uni BB0_103;
|
|
|
|
BB0_100:
|
|
setp.neu.f32 %p102, %f132, 0f7F800000;
|
|
@%p102 bra BB0_103;
|
|
|
|
selp.f32 %f846, 0fFF800000, 0f7F800000, %p4;
|
|
|
|
BB0_103:
|
|
setp.eq.f32 %p103, %f128, 0f3F800000;
|
|
selp.f32 %f144, 0f3F800000, %f846, %p103;
|
|
abs.f32 %f145, %f129;
|
|
setp.lt.f32 %p104, %f145, 0f00800000;
|
|
mul.f32 %f617, %f145, 0f4B800000;
|
|
selp.f32 %f618, 0fC3170000, 0fC2FE0000, %p104;
|
|
selp.f32 %f619, %f617, %f145, %p104;
|
|
mov.b32 %r281, %f619;
|
|
and.b32 %r282, %r281, 8388607;
|
|
or.b32 %r283, %r282, 1065353216;
|
|
mov.b32 %f620, %r283;
|
|
shr.u32 %r284, %r281, 23;
|
|
cvt.rn.f32.u32 %f621, %r284;
|
|
add.f32 %f622, %f618, %f621;
|
|
setp.gt.f32 %p105, %f620, 0f3FB504F3;
|
|
mul.f32 %f623, %f620, 0f3F000000;
|
|
add.f32 %f624, %f622, 0f3F800000;
|
|
selp.f32 %f625, %f623, %f620, %p105;
|
|
selp.f32 %f626, %f624, %f622, %p105;
|
|
add.f32 %f627, %f625, 0fBF800000;
|
|
add.f32 %f616, %f625, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f615,%f616;
|
|
// inline asm
|
|
add.f32 %f628, %f627, %f627;
|
|
mul.f32 %f629, %f615, %f628;
|
|
mul.f32 %f630, %f629, %f629;
|
|
fma.rn.f32 %f633, %f554, %f630, %f553;
|
|
fma.rn.f32 %f635, %f633, %f630, %f556;
|
|
mul.rn.f32 %f636, %f635, %f630;
|
|
mul.rn.f32 %f637, %f636, %f629;
|
|
sub.f32 %f638, %f627, %f629;
|
|
neg.f32 %f639, %f629;
|
|
add.f32 %f640, %f638, %f638;
|
|
fma.rn.f32 %f641, %f639, %f627, %f640;
|
|
mul.rn.f32 %f642, %f615, %f641;
|
|
add.f32 %f643, %f637, %f629;
|
|
sub.f32 %f644, %f629, %f643;
|
|
add.f32 %f645, %f637, %f644;
|
|
add.f32 %f646, %f642, %f645;
|
|
add.f32 %f647, %f643, %f646;
|
|
sub.f32 %f648, %f643, %f647;
|
|
add.f32 %f649, %f646, %f648;
|
|
mul.rn.f32 %f651, %f626, %f572;
|
|
mul.rn.f32 %f653, %f626, %f574;
|
|
add.f32 %f654, %f651, %f647;
|
|
sub.f32 %f655, %f651, %f654;
|
|
add.f32 %f656, %f647, %f655;
|
|
add.f32 %f657, %f649, %f656;
|
|
add.f32 %f658, %f653, %f657;
|
|
add.f32 %f659, %f654, %f658;
|
|
sub.f32 %f660, %f654, %f659;
|
|
add.f32 %f661, %f658, %f660;
|
|
mul.rn.f32 %f663, %f584, %f659;
|
|
neg.f32 %f664, %f663;
|
|
fma.rn.f32 %f665, %f584, %f659, %f664;
|
|
fma.rn.f32 %f666, %f584, %f661, %f665;
|
|
fma.rn.f32 %f668, %f589, %f659, %f666;
|
|
add.rn.f32 %f669, %f663, %f668;
|
|
neg.f32 %f670, %f669;
|
|
add.rn.f32 %f671, %f663, %f670;
|
|
add.rn.f32 %f672, %f671, %f668;
|
|
mov.b32 %r285, %f669;
|
|
setp.eq.s32 %p106, %r285, 1118925336;
|
|
add.s32 %r286, %r285, -1;
|
|
mov.b32 %f673, %r286;
|
|
add.f32 %f674, %f672, 0f37000000;
|
|
selp.f32 %f675, %f673, %f669, %p106;
|
|
selp.f32 %f146, %f674, %f672, %p106;
|
|
mul.f32 %f676, %f675, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f677, %f676;
|
|
fma.rn.f32 %f679, %f677, %f600, %f675;
|
|
fma.rn.f32 %f681, %f677, %f602, %f679;
|
|
mul.f32 %f682, %f681, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f683, %f682;
|
|
add.f32 %f684, %f677, 0f00000000;
|
|
ex2.approx.f32 %f685, %f684;
|
|
mul.f32 %f686, %f683, %f685;
|
|
setp.lt.f32 %p107, %f675, 0fC2D20000;
|
|
selp.f32 %f687, 0f00000000, %f686, %p107;
|
|
setp.gt.f32 %p108, %f675, 0f42D20000;
|
|
selp.f32 %f847, 0f7F800000, %f687, %p108;
|
|
setp.eq.f32 %p109, %f847, 0f7F800000;
|
|
@%p109 bra BB0_105;
|
|
|
|
fma.rn.f32 %f847, %f847, %f146, %f847;
|
|
|
|
BB0_105:
|
|
setp.lt.f32 %p110, %f129, 0f00000000;
|
|
and.pred %p5, %p110, %p95;
|
|
mov.b32 %r287, %f847;
|
|
xor.b32 %r288, %r287, -2147483648;
|
|
mov.b32 %f688, %r288;
|
|
selp.f32 %f849, %f688, %f847, %p5;
|
|
setp.eq.f32 %p112, %f129, 0f00000000;
|
|
@%p112 bra BB0_108;
|
|
bra.uni BB0_106;
|
|
|
|
BB0_108:
|
|
add.f32 %f691, %f129, %f129;
|
|
selp.f32 %f849, %f691, 0f00000000, %p95;
|
|
bra.uni BB0_109;
|
|
|
|
BB0_106:
|
|
setp.geu.f32 %p113, %f129, 0f00000000;
|
|
@%p113 bra BB0_109;
|
|
|
|
cvt.rzi.f32.f32 %f690, %f584;
|
|
setp.neu.f32 %p114, %f690, 0f3EE66666;
|
|
selp.f32 %f849, 0f7FFFFFFF, %f849, %p114;
|
|
|
|
BB0_109:
|
|
add.f32 %f692, %f145, 0f3EE66666;
|
|
mov.b32 %r289, %f692;
|
|
setp.lt.s32 %p116, %r289, 2139095040;
|
|
@%p116 bra BB0_114;
|
|
|
|
setp.gtu.f32 %p117, %f145, 0f7F800000;
|
|
@%p117 bra BB0_113;
|
|
bra.uni BB0_111;
|
|
|
|
BB0_113:
|
|
add.f32 %f849, %f129, 0f3EE66666;
|
|
bra.uni BB0_114;
|
|
|
|
BB0_111:
|
|
setp.neu.f32 %p118, %f145, 0f7F800000;
|
|
@%p118 bra BB0_114;
|
|
|
|
selp.f32 %f849, 0fFF800000, 0f7F800000, %p5;
|
|
|
|
BB0_114:
|
|
setp.eq.f32 %p119, %f129, 0f3F800000;
|
|
selp.f32 %f157, 0f3F800000, %f849, %p119;
|
|
abs.f32 %f158, %f130;
|
|
setp.lt.f32 %p120, %f158, 0f00800000;
|
|
mul.f32 %f695, %f158, 0f4B800000;
|
|
selp.f32 %f696, 0fC3170000, 0fC2FE0000, %p120;
|
|
selp.f32 %f697, %f695, %f158, %p120;
|
|
mov.b32 %r290, %f697;
|
|
and.b32 %r291, %r290, 8388607;
|
|
or.b32 %r292, %r291, 1065353216;
|
|
mov.b32 %f698, %r292;
|
|
shr.u32 %r293, %r290, 23;
|
|
cvt.rn.f32.u32 %f699, %r293;
|
|
add.f32 %f700, %f696, %f699;
|
|
setp.gt.f32 %p121, %f698, 0f3FB504F3;
|
|
mul.f32 %f701, %f698, 0f3F000000;
|
|
add.f32 %f702, %f700, 0f3F800000;
|
|
selp.f32 %f703, %f701, %f698, %p121;
|
|
selp.f32 %f704, %f702, %f700, %p121;
|
|
add.f32 %f705, %f703, 0fBF800000;
|
|
add.f32 %f694, %f703, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f693,%f694;
|
|
// inline asm
|
|
add.f32 %f706, %f705, %f705;
|
|
mul.f32 %f707, %f693, %f706;
|
|
mul.f32 %f708, %f707, %f707;
|
|
fma.rn.f32 %f711, %f554, %f708, %f553;
|
|
fma.rn.f32 %f713, %f711, %f708, %f556;
|
|
mul.rn.f32 %f714, %f713, %f708;
|
|
mul.rn.f32 %f715, %f714, %f707;
|
|
sub.f32 %f716, %f705, %f707;
|
|
neg.f32 %f717, %f707;
|
|
add.f32 %f718, %f716, %f716;
|
|
fma.rn.f32 %f719, %f717, %f705, %f718;
|
|
mul.rn.f32 %f720, %f693, %f719;
|
|
add.f32 %f721, %f715, %f707;
|
|
sub.f32 %f722, %f707, %f721;
|
|
add.f32 %f723, %f715, %f722;
|
|
add.f32 %f724, %f720, %f723;
|
|
add.f32 %f725, %f721, %f724;
|
|
sub.f32 %f726, %f721, %f725;
|
|
add.f32 %f727, %f724, %f726;
|
|
mul.rn.f32 %f729, %f704, %f572;
|
|
mul.rn.f32 %f731, %f704, %f574;
|
|
add.f32 %f732, %f729, %f725;
|
|
sub.f32 %f733, %f729, %f732;
|
|
add.f32 %f734, %f725, %f733;
|
|
add.f32 %f735, %f727, %f734;
|
|
add.f32 %f736, %f731, %f735;
|
|
add.f32 %f737, %f732, %f736;
|
|
sub.f32 %f738, %f732, %f737;
|
|
add.f32 %f739, %f736, %f738;
|
|
mul.rn.f32 %f741, %f584, %f737;
|
|
neg.f32 %f742, %f741;
|
|
fma.rn.f32 %f743, %f584, %f737, %f742;
|
|
fma.rn.f32 %f744, %f584, %f739, %f743;
|
|
fma.rn.f32 %f746, %f589, %f737, %f744;
|
|
add.rn.f32 %f747, %f741, %f746;
|
|
neg.f32 %f748, %f747;
|
|
add.rn.f32 %f749, %f741, %f748;
|
|
add.rn.f32 %f750, %f749, %f746;
|
|
mov.b32 %r294, %f747;
|
|
setp.eq.s32 %p122, %r294, 1118925336;
|
|
add.s32 %r295, %r294, -1;
|
|
mov.b32 %f751, %r295;
|
|
add.f32 %f752, %f750, 0f37000000;
|
|
selp.f32 %f753, %f751, %f747, %p122;
|
|
selp.f32 %f159, %f752, %f750, %p122;
|
|
mul.f32 %f754, %f753, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f755, %f754;
|
|
fma.rn.f32 %f757, %f755, %f600, %f753;
|
|
fma.rn.f32 %f759, %f755, %f602, %f757;
|
|
mul.f32 %f760, %f759, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f761, %f760;
|
|
add.f32 %f762, %f755, 0f00000000;
|
|
ex2.approx.f32 %f763, %f762;
|
|
mul.f32 %f764, %f761, %f763;
|
|
setp.lt.f32 %p123, %f753, 0fC2D20000;
|
|
selp.f32 %f765, 0f00000000, %f764, %p123;
|
|
setp.gt.f32 %p124, %f753, 0f42D20000;
|
|
selp.f32 %f850, 0f7F800000, %f765, %p124;
|
|
setp.eq.f32 %p125, %f850, 0f7F800000;
|
|
@%p125 bra BB0_116;
|
|
|
|
fma.rn.f32 %f850, %f850, %f159, %f850;
|
|
|
|
BB0_116:
|
|
setp.lt.f32 %p126, %f130, 0f00000000;
|
|
and.pred %p6, %p126, %p95;
|
|
mov.b32 %r296, %f850;
|
|
xor.b32 %r297, %r296, -2147483648;
|
|
mov.b32 %f766, %r297;
|
|
selp.f32 %f852, %f766, %f850, %p6;
|
|
setp.eq.f32 %p128, %f130, 0f00000000;
|
|
@%p128 bra BB0_119;
|
|
bra.uni BB0_117;
|
|
|
|
BB0_119:
|
|
add.f32 %f769, %f130, %f130;
|
|
selp.f32 %f852, %f769, 0f00000000, %p95;
|
|
bra.uni BB0_120;
|
|
|
|
BB0_117:
|
|
setp.geu.f32 %p129, %f130, 0f00000000;
|
|
@%p129 bra BB0_120;
|
|
|
|
cvt.rzi.f32.f32 %f768, %f584;
|
|
setp.neu.f32 %p130, %f768, 0f3EE66666;
|
|
selp.f32 %f852, 0f7FFFFFFF, %f852, %p130;
|
|
|
|
BB0_120:
|
|
add.f32 %f770, %f158, 0f3EE66666;
|
|
mov.b32 %r298, %f770;
|
|
setp.lt.s32 %p132, %r298, 2139095040;
|
|
@%p132 bra BB0_125;
|
|
|
|
setp.gtu.f32 %p133, %f158, 0f7F800000;
|
|
@%p133 bra BB0_124;
|
|
bra.uni BB0_122;
|
|
|
|
BB0_124:
|
|
add.f32 %f852, %f130, 0f3EE66666;
|
|
bra.uni BB0_125;
|
|
|
|
BB0_122:
|
|
setp.neu.f32 %p134, %f158, 0f7F800000;
|
|
@%p134 bra BB0_125;
|
|
|
|
selp.f32 %f852, 0fFF800000, 0f7F800000, %p6;
|
|
|
|
BB0_125:
|
|
setp.eq.f32 %p135, %f130, 0f3F800000;
|
|
selp.f32 %f771, 0f3F800000, %f852, %p135;
|
|
cvt.u64.u32 %rd83, %r6;
|
|
cvt.u64.u32 %rd82, %r5;
|
|
mov.u64 %rd86, image;
|
|
cvta.global.u64 %rd81, %rd86;
|
|
// inline asm
|
|
call (%rd80), _rt_buffer_get_64, (%rd81, %r103, %r104, %rd82, %rd83, %rd25, %rd25);
|
|
// inline asm
|
|
cvt.sat.f32.f32 %f772, %f771;
|
|
mul.f32 %f773, %f772, 0f437FFD71;
|
|
cvt.rzi.u32.f32 %r301, %f773;
|
|
cvt.sat.f32.f32 %f774, %f157;
|
|
mul.f32 %f775, %f774, 0f437FFD71;
|
|
cvt.rzi.u32.f32 %r302, %f775;
|
|
cvt.sat.f32.f32 %f776, %f144;
|
|
mul.f32 %f777, %f776, 0f437FFD71;
|
|
cvt.rzi.u32.f32 %r303, %f777;
|
|
cvt.u16.u32 %rs13, %r301;
|
|
cvt.u16.u32 %rs14, %r303;
|
|
cvt.u16.u32 %rs15, %r302;
|
|
mov.u16 %rs16, 255;
|
|
st.v4.u8 [%rd80], {%rs13, %rs15, %rs14, %rs16};
|
|
ld.global.u32 %r359, [imageEnabled];
|
|
|
|
BB0_126:
|
|
and.b32 %r304, %r359, 4;
|
|
setp.eq.s32 %p136, %r304, 0;
|
|
@%p136 bra BB0_138;
|
|
|
|
ld.global.u32 %r305, [additive];
|
|
setp.eq.s32 %p137, %r305, 0;
|
|
cvt.u64.u32 %rd16, %r5;
|
|
cvt.u64.u32 %rd17, %r6;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs17, %f338;}
|
|
|
|
// inline asm
|
|
@%p137 bra BB0_129;
|
|
|
|
mov.u64 %rd99, image_HDR;
|
|
cvta.global.u64 %rd88, %rd99;
|
|
mov.u32 %r309, 8;
|
|
// inline asm
|
|
call (%rd87), _rt_buffer_get_64, (%rd88, %r103, %r309, %rd16, %rd17, %rd25, %rd25);
|
|
// inline asm
|
|
ld.v4.u16 {%rs24, %rs25, %rs26, %rs27}, [%rd87];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f779, %rs24;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f780, %rs25;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f781, %rs26;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd93), _rt_buffer_get_64, (%rd88, %r103, %r309, %rd16, %rd17, %rd25, %rd25);
|
|
// inline asm
|
|
add.f32 %f782, %f128, %f779;
|
|
add.f32 %f783, %f129, %f780;
|
|
add.f32 %f784, %f130, %f781;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs23, %f784;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs22, %f783;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs21, %f782;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd93], {%rs21, %rs22, %rs23, %rs17};
|
|
bra.uni BB0_138;
|
|
|
|
BB0_129:
|
|
mov.u64 %rd106, image_HDR;
|
|
cvta.global.u64 %rd101, %rd106;
|
|
mov.u32 %r311, 8;
|
|
// inline asm
|
|
call (%rd100), _rt_buffer_get_64, (%rd101, %r103, %r311, %rd16, %rd17, %rd25, %rd25);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs30, %f130;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs29, %f129;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs28, %f128;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd100], {%rs28, %rs29, %rs30, %rs17};
|
|
|
|
BB0_138:
|
|
ret;
|
|
}
|
|
|
|
|