709 lines
24 KiB
Plaintext
709 lines
24 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_HDR[1];
|
|
.global .align 1 .b8 image_HDR2[1];
|
|
.global .align 1 .b8 uvfacenormal[1];
|
|
.global .align 1 .b8 uvpos[1];
|
|
.global .align 1 .b8 uvsmoothpos[1];
|
|
.global .align 1 .b8 triangleMarks[1];
|
|
.global .align 1 .b8 image_Mask[1];
|
|
.global .align 4 .f32 fakeBias;
|
|
.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_typeinfo8fakeBiasE[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 8 .b8 _ZN21rti_internal_typename8fakeBiasE[6] = {102, 108, 111, 97, 116, 0};
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum7pixelIDE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum10resolutionE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum6normalE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum6camPosE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum4rootE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum12imageEnabledE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum8tileInfoE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum8additiveE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum8fakeBiasE = 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_semantic8fakeBiasE[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_annotation8fakeBiasE[1];
|
|
|
|
.visible .entry _Z6oxMainv(
|
|
|
|
)
|
|
{
|
|
.local .align 4 .b8 __local_depot0[28];
|
|
.reg .b64 %SP;
|
|
.reg .b64 %SPL;
|
|
.reg .pred %p<33>;
|
|
.reg .b16 %rs<12>;
|
|
.reg .f32 %f<388>;
|
|
.reg .b32 %r<129>;
|
|
.reg .b64 %rd<124>;
|
|
|
|
|
|
mov.u64 %rd123, __local_depot0;
|
|
cvta.local.u64 %SP, %rd123;
|
|
ld.global.v2.u32 {%r8, %r9}, [pixelID];
|
|
cvt.u64.u32 %rd8, %r8;
|
|
cvt.u64.u32 %rd9, %r9;
|
|
mov.u64 %rd12, uvfacenormal;
|
|
cvta.global.u64 %rd7, %rd12;
|
|
mov.u32 %r6, 2;
|
|
mov.u32 %r7, 4;
|
|
mov.u64 %rd11, 0;
|
|
// inline asm
|
|
call (%rd6), _rt_buffer_get_64, (%rd7, %r6, %r7, %rd8, %rd9, %rd11, %rd11);
|
|
// inline asm
|
|
ld.u32 %r1, [%rd6];
|
|
shr.u32 %r12, %r1, 16;
|
|
cvt.u16.u32 %rs1, %r12;
|
|
and.b16 %rs2, %rs1, 255;
|
|
cvt.u16.u32 %rs3, %r1;
|
|
or.b16 %rs4, %rs3, %rs2;
|
|
setp.eq.s16 %p1, %rs4, 0;
|
|
mov.f32 %f366, 0f00000000;
|
|
mov.f32 %f367, %f366;
|
|
mov.f32 %f368, %f366;
|
|
@%p1 bra BB0_2;
|
|
|
|
ld.u8 %rs5, [%rd6+1];
|
|
and.b16 %rs7, %rs3, 255;
|
|
cvt.rn.f32.u16 %f121, %rs7;
|
|
div.rn.f32 %f122, %f121, 0f437F0000;
|
|
fma.rn.f32 %f123, %f122, 0f40000000, 0fBF800000;
|
|
cvt.rn.f32.u16 %f124, %rs5;
|
|
div.rn.f32 %f125, %f124, 0f437F0000;
|
|
fma.rn.f32 %f126, %f125, 0f40000000, 0fBF800000;
|
|
cvt.rn.f32.u16 %f127, %rs2;
|
|
div.rn.f32 %f128, %f127, 0f437F0000;
|
|
fma.rn.f32 %f129, %f128, 0f40000000, 0fBF800000;
|
|
mul.f32 %f130, %f126, %f126;
|
|
fma.rn.f32 %f131, %f123, %f123, %f130;
|
|
fma.rn.f32 %f132, %f129, %f129, %f131;
|
|
sqrt.rn.f32 %f133, %f132;
|
|
rcp.rn.f32 %f134, %f133;
|
|
mul.f32 %f366, %f123, %f134;
|
|
mul.f32 %f367, %f126, %f134;
|
|
mul.f32 %f368, %f129, %f134;
|
|
|
|
BB0_2:
|
|
ld.global.v2.u32 {%r13, %r14}, [pixelID];
|
|
ld.global.v2.u32 {%r16, %r17}, [tileInfo];
|
|
add.s32 %r2, %r13, %r16;
|
|
add.s32 %r3, %r14, %r17;
|
|
setp.eq.f32 %p2, %f367, 0f00000000;
|
|
setp.eq.f32 %p3, %f366, 0f00000000;
|
|
and.pred %p4, %p3, %p2;
|
|
setp.eq.f32 %p5, %f368, 0f00000000;
|
|
and.pred %p6, %p4, %p5;
|
|
@%p6 bra BB0_27;
|
|
bra.uni BB0_3;
|
|
|
|
BB0_27:
|
|
cvt.u64.u32 %rd107, %r2;
|
|
cvt.u64.u32 %rd108, %r3;
|
|
mov.u64 %rd111, image_HDR;
|
|
cvta.global.u64 %rd106, %rd111;
|
|
mov.u32 %r109, 16;
|
|
// inline asm
|
|
call (%rd105), _rt_buffer_get_64, (%rd106, %r6, %r109, %rd107, %rd108, %rd11, %rd11);
|
|
// inline asm
|
|
mov.f32 %f354, 0f00000000;
|
|
st.v4.f32 [%rd105], {%f354, %f354, %f354, %f354};
|
|
bra.uni BB0_28;
|
|
|
|
BB0_3:
|
|
ld.global.v2.u32 {%r39, %r40}, [pixelID];
|
|
cvt.u64.u32 %rd15, %r39;
|
|
cvt.u64.u32 %rd16, %r40;
|
|
mov.u64 %rd67, uvpos;
|
|
cvta.global.u64 %rd14, %rd67;
|
|
mov.u32 %r36, 16;
|
|
// inline asm
|
|
call (%rd13), _rt_buffer_get_64, (%rd14, %r6, %r36, %rd15, %rd16, %rd11, %rd11);
|
|
// inline asm
|
|
ld.f32 %f135, [%rd13];
|
|
ld.global.v2.u32 {%r43, %r44}, [pixelID];
|
|
cvt.u64.u32 %rd21, %r43;
|
|
cvt.u64.u32 %rd22, %r44;
|
|
// inline asm
|
|
call (%rd19), _rt_buffer_get_64, (%rd14, %r6, %r36, %rd21, %rd22, %rd11, %rd11);
|
|
// inline asm
|
|
ld.f32 %f136, [%rd19+4];
|
|
ld.global.v2.u32 {%r47, %r48}, [pixelID];
|
|
cvt.u64.u32 %rd27, %r47;
|
|
cvt.u64.u32 %rd28, %r48;
|
|
// inline asm
|
|
call (%rd25), _rt_buffer_get_64, (%rd14, %r6, %r36, %rd27, %rd28, %rd11, %rd11);
|
|
// inline asm
|
|
abs.f32 %f137, %f135;
|
|
setp.gtu.f32 %p7, %f137, 0f7F800000;
|
|
selp.f32 %f7, 0f00000000, %f135, %p7;
|
|
abs.f32 %f138, %f136;
|
|
setp.gtu.f32 %p8, %f138, 0f7F800000;
|
|
selp.f32 %f8, 0f00000000, %f136, %p8;
|
|
ld.f32 %f139, [%rd25+8];
|
|
abs.f32 %f140, %f139;
|
|
setp.gtu.f32 %p9, %f140, 0f7F800000;
|
|
selp.f32 %f9, 0f00000000, %f139, %p9;
|
|
ld.global.v2.u32 {%r51, %r52}, [pixelID];
|
|
cvt.u64.u32 %rd33, %r51;
|
|
cvt.u64.u32 %rd34, %r52;
|
|
// inline asm
|
|
call (%rd31), _rt_buffer_get_64, (%rd14, %r6, %r36, %rd33, %rd34, %rd11, %rd11);
|
|
// inline asm
|
|
ld.f32 %f141, [%rd31+12];
|
|
add.f32 %f10, %f141, 0f38D1B717;
|
|
mul.f32 %f142, %f7, 0f3456BF95;
|
|
mul.f32 %f143, %f8, 0f3456BF95;
|
|
mul.f32 %f144, %f9, 0f3456BF95;
|
|
abs.f32 %f11, %f366;
|
|
div.rn.f32 %f145, %f142, %f11;
|
|
abs.f32 %f146, %f367;
|
|
div.rn.f32 %f147, %f143, %f146;
|
|
abs.f32 %f12, %f368;
|
|
div.rn.f32 %f148, %f144, %f12;
|
|
abs.f32 %f149, %f145;
|
|
abs.f32 %f150, %f147;
|
|
abs.f32 %f151, %f148;
|
|
mov.f32 %f152, 0f38D1B717;
|
|
max.f32 %f153, %f149, %f152;
|
|
max.f32 %f154, %f150, %f152;
|
|
max.f32 %f155, %f151, %f152;
|
|
mul.f32 %f13, %f366, %f153;
|
|
mul.f32 %f14, %f367, %f154;
|
|
mul.f32 %f15, %f368, %f155;
|
|
add.f32 %f16, %f7, %f13;
|
|
add.f32 %f17, %f8, %f14;
|
|
add.f32 %f18, %f9, %f15;
|
|
ld.global.v2.u32 {%r55, %r56}, [pixelID];
|
|
cvt.u64.u32 %rd39, %r55;
|
|
cvt.u64.u32 %rd40, %r56;
|
|
mov.u64 %rd68, uvsmoothpos;
|
|
cvta.global.u64 %rd38, %rd68;
|
|
// inline asm
|
|
call (%rd37), _rt_buffer_get_64, (%rd38, %r6, %r36, %rd39, %rd40, %rd11, %rd11);
|
|
// inline asm
|
|
ld.f32 %f19, [%rd37];
|
|
ld.global.v2.u32 {%r59, %r60}, [pixelID];
|
|
cvt.u64.u32 %rd45, %r59;
|
|
cvt.u64.u32 %rd46, %r60;
|
|
// inline asm
|
|
call (%rd43), _rt_buffer_get_64, (%rd38, %r6, %r36, %rd45, %rd46, %rd11, %rd11);
|
|
// inline asm
|
|
ld.f32 %f20, [%rd43+4];
|
|
ld.global.v2.u32 {%r63, %r64}, [pixelID];
|
|
cvt.u64.u32 %rd51, %r63;
|
|
cvt.u64.u32 %rd52, %r64;
|
|
// inline asm
|
|
call (%rd49), _rt_buffer_get_64, (%rd38, %r6, %r36, %rd51, %rd52, %rd11, %rd11);
|
|
// inline asm
|
|
ld.f32 %f21, [%rd49+8];
|
|
ld.global.v2.u32 {%r67, %r68}, [pixelID];
|
|
cvt.u64.u32 %rd57, %r67;
|
|
cvt.u64.u32 %rd58, %r68;
|
|
// inline asm
|
|
call (%rd55), _rt_buffer_get_64, (%rd38, %r6, %r36, %rd57, %rd58, %rd11, %rd11);
|
|
// inline asm
|
|
ld.f32 %f156, [%rd55+12];
|
|
cvt.rzi.u32.f32 %r71, %f156;
|
|
cvt.u64.u32 %rd63, %r71;
|
|
mov.u64 %rd69, triangleMarks;
|
|
cvta.global.u64 %rd62, %rd69;
|
|
mov.u32 %r38, 1;
|
|
// inline asm
|
|
call (%rd61), _rt_buffer_get_64, (%rd62, %r38, %r38, %rd63, %rd11, %rd11, %rd11);
|
|
// inline asm
|
|
ld.u8 %r128, [%rd61];
|
|
sub.f32 %f22, %f19, %f7;
|
|
sub.f32 %f23, %f20, %f8;
|
|
sub.f32 %f24, %f21, %f9;
|
|
mul.f32 %f157, %f23, %f23;
|
|
fma.rn.f32 %f158, %f22, %f22, %f157;
|
|
fma.rn.f32 %f159, %f24, %f24, %f158;
|
|
sqrt.rn.f32 %f25, %f159;
|
|
setp.leu.f32 %p10, %f25, 0f3727C5AC;
|
|
setp.eq.s32 %p11, %r128, 255;
|
|
or.pred %p12, %p11, %p10;
|
|
@%p12 bra BB0_6;
|
|
|
|
rcp.rn.f32 %f168, %f25;
|
|
mul.f32 %f163, %f22, %f168;
|
|
mul.f32 %f164, %f23, %f168;
|
|
mul.f32 %f165, %f24, %f168;
|
|
add.u64 %rd70, %SP, 12;
|
|
cvta.to.local.u64 %rd71, %rd70;
|
|
mov.u32 %r75, 1065353216;
|
|
st.local.u32 [%rd71], %r75;
|
|
ld.global.u32 %r72, [root];
|
|
// inline asm
|
|
call _rt_trace_64, (%r72, %f16, %f17, %f18, %f163, %f164, %f165, %r38, %f152, %f25, %rd70, %r7);
|
|
// inline asm
|
|
ld.local.f32 %f169, [%rd71];
|
|
setp.neu.f32 %p13, %f169, 0f00000000;
|
|
@%p13 bra BB0_6;
|
|
|
|
// inline asm
|
|
call (%rd72), _rt_buffer_get_64, (%rd62, %r38, %r38, %rd63, %rd11, %rd11, %rd11);
|
|
// inline asm
|
|
mov.u16 %rs9, 255;
|
|
st.u8 [%rd72], %rs9;
|
|
mov.u32 %r128, 255;
|
|
|
|
BB0_6:
|
|
setp.eq.s32 %p14, %r128, 0;
|
|
selp.f32 %f373, %f19, %f7, %p14;
|
|
selp.f32 %f374, %f20, %f8, %p14;
|
|
selp.f32 %f375, %f21, %f9, %p14;
|
|
add.f32 %f170, %f13, %f373;
|
|
add.f32 %f171, %f14, %f374;
|
|
add.f32 %f172, %f15, %f375;
|
|
neg.f32 %f179, %f367;
|
|
setp.gt.f32 %p15, %f11, %f12;
|
|
selp.f32 %f180, %f179, 0f00000000, %p15;
|
|
neg.f32 %f181, %f368;
|
|
selp.f32 %f182, %f366, %f181, %p15;
|
|
selp.f32 %f183, 0f00000000, %f367, %p15;
|
|
mul.f32 %f184, %f182, %f182;
|
|
fma.rn.f32 %f185, %f180, %f180, %f184;
|
|
fma.rn.f32 %f186, %f183, %f183, %f185;
|
|
sqrt.rn.f32 %f187, %f186;
|
|
rcp.rn.f32 %f188, %f187;
|
|
mul.f32 %f32, %f180, %f188;
|
|
mul.f32 %f33, %f182, %f188;
|
|
mul.f32 %f34, %f183, %f188;
|
|
mul.f32 %f189, %f368, %f33;
|
|
mul.f32 %f190, %f367, %f34;
|
|
sub.f32 %f35, %f189, %f190;
|
|
mul.f32 %f191, %f366, %f34;
|
|
mul.f32 %f192, %f368, %f32;
|
|
sub.f32 %f36, %f191, %f192;
|
|
mul.f32 %f193, %f367, %f32;
|
|
mul.f32 %f194, %f366, %f33;
|
|
sub.f32 %f37, %f193, %f194;
|
|
mov.f32 %f195, 0f3F8147AE;
|
|
sqrt.rn.f32 %f196, %f195;
|
|
rcp.rn.f32 %f38, %f196;
|
|
neg.f32 %f39, %f38;
|
|
mul.f32 %f40, %f38, 0f00000000;
|
|
mul.f32 %f197, %f38, 0f3DCCCCCD;
|
|
mul.f32 %f198, %f170, 0f3456BF95;
|
|
mul.f32 %f199, %f171, 0f3456BF95;
|
|
mul.f32 %f200, %f172, 0f3456BF95;
|
|
abs.f32 %f201, %f198;
|
|
abs.f32 %f202, %f199;
|
|
abs.f32 %f203, %f200;
|
|
max.f32 %f204, %f201, %f202;
|
|
max.f32 %f205, %f204, %f203;
|
|
max.f32 %f41, %f205, %f152;
|
|
mul.f32 %f42, %f32, %f40;
|
|
mul.f32 %f43, %f33, %f40;
|
|
mul.f32 %f44, %f34, %f40;
|
|
fma.rn.f32 %f207, %f35, %f39, %f42;
|
|
fma.rn.f32 %f208, %f36, %f39, %f43;
|
|
fma.rn.f32 %f209, %f37, %f39, %f44;
|
|
mul.f32 %f45, %f366, %f197;
|
|
mul.f32 %f46, %f367, %f197;
|
|
mul.f32 %f47, %f368, %f197;
|
|
add.f32 %f173, %f207, %f45;
|
|
add.f32 %f174, %f208, %f46;
|
|
add.f32 %f175, %f209, %f47;
|
|
add.u64 %rd79, %SP, 16;
|
|
cvta.to.local.u64 %rd80, %rd79;
|
|
mov.u32 %r80, 0;
|
|
st.local.u32 [%rd80+8], %r80;
|
|
st.local.u32 [%rd80+4], %r80;
|
|
st.local.u32 [%rd80], %r80;
|
|
ld.global.u32 %r79, [root];
|
|
mul.f32 %f177, %f10, 0f3FB504F3;
|
|
mov.f32 %f176, 0f00000000;
|
|
mov.u32 %r81, 12;
|
|
// inline asm
|
|
call _rt_trace_64, (%r79, %f170, %f171, %f172, %f173, %f174, %f175, %r80, %f176, %f177, %rd79, %r81);
|
|
// inline asm
|
|
ld.local.f32 %f52, [%rd80+4];
|
|
ld.local.f32 %f53, [%rd80];
|
|
add.f32 %f210, %f53, %f52;
|
|
ld.local.f32 %f54, [%rd80+8];
|
|
add.f32 %f211, %f210, %f54;
|
|
mov.f32 %f376, 0f47C34F80;
|
|
setp.eq.f32 %p16, %f211, 0f00000000;
|
|
@%p16 bra BB0_9;
|
|
|
|
mul.f32 %f213, %f52, %f52;
|
|
fma.rn.f32 %f214, %f53, %f53, %f213;
|
|
fma.rn.f32 %f215, %f54, %f54, %f214;
|
|
sqrt.rn.f32 %f55, %f215;
|
|
rcp.rn.f32 %f216, %f55;
|
|
mul.f32 %f217, %f216, %f53;
|
|
mul.f32 %f218, %f216, %f52;
|
|
mul.f32 %f219, %f216, %f54;
|
|
fma.rn.f32 %f220, %f41, %f217, %f170;
|
|
fma.rn.f32 %f221, %f41, %f218, %f171;
|
|
fma.rn.f32 %f222, %f41, %f219, %f172;
|
|
fma.rn.f32 %f56, %f55, %f173, %f220;
|
|
fma.rn.f32 %f57, %f55, %f174, %f221;
|
|
fma.rn.f32 %f58, %f55, %f175, %f222;
|
|
setp.geu.f32 %p17, %f55, 0f47C34F80;
|
|
@%p17 bra BB0_9;
|
|
|
|
mov.f32 %f373, %f56;
|
|
mov.f32 %f374, %f57;
|
|
mov.f32 %f375, %f58;
|
|
mov.f32 %f376, %f55;
|
|
|
|
BB0_9:
|
|
mov.u32 %r111, 12;
|
|
mov.f32 %f355, 0f00000000;
|
|
add.u64 %rd113, %SP, 16;
|
|
cvta.to.local.u64 %rd112, %rd113;
|
|
mov.u32 %r110, 0;
|
|
fma.rn.f32 %f231, %f35, %f38, %f42;
|
|
fma.rn.f32 %f232, %f36, %f38, %f43;
|
|
fma.rn.f32 %f233, %f37, %f38, %f44;
|
|
add.f32 %f226, %f231, %f45;
|
|
add.f32 %f227, %f232, %f46;
|
|
add.f32 %f228, %f233, %f47;
|
|
st.local.u32 [%rd112+8], %r110;
|
|
st.local.u32 [%rd112+4], %r110;
|
|
st.local.u32 [%rd112], %r110;
|
|
ld.global.u32 %r82, [root];
|
|
// inline asm
|
|
call _rt_trace_64, (%r82, %f170, %f171, %f172, %f226, %f227, %f228, %r110, %f355, %f177, %rd113, %r111);
|
|
// inline asm
|
|
ld.local.f32 %f66, [%rd112+4];
|
|
ld.local.f32 %f67, [%rd112];
|
|
add.f32 %f234, %f67, %f66;
|
|
ld.local.f32 %f68, [%rd112+8];
|
|
add.f32 %f235, %f234, %f68;
|
|
setp.eq.f32 %p18, %f235, 0f00000000;
|
|
@%p18 bra BB0_12;
|
|
|
|
mul.f32 %f236, %f66, %f66;
|
|
fma.rn.f32 %f237, %f67, %f67, %f236;
|
|
fma.rn.f32 %f238, %f68, %f68, %f237;
|
|
sqrt.rn.f32 %f69, %f238;
|
|
rcp.rn.f32 %f239, %f69;
|
|
mul.f32 %f240, %f239, %f67;
|
|
mul.f32 %f241, %f239, %f66;
|
|
mul.f32 %f242, %f239, %f68;
|
|
fma.rn.f32 %f243, %f41, %f240, %f170;
|
|
fma.rn.f32 %f244, %f41, %f241, %f171;
|
|
fma.rn.f32 %f245, %f41, %f242, %f172;
|
|
fma.rn.f32 %f70, %f69, %f226, %f243;
|
|
fma.rn.f32 %f71, %f69, %f227, %f244;
|
|
fma.rn.f32 %f72, %f69, %f228, %f245;
|
|
setp.geu.f32 %p19, %f69, %f376;
|
|
@%p19 bra BB0_12;
|
|
|
|
mov.f32 %f373, %f70;
|
|
mov.f32 %f374, %f71;
|
|
mov.f32 %f375, %f72;
|
|
mov.f32 %f376, %f69;
|
|
|
|
BB0_12:
|
|
neg.f32 %f358, %f38;
|
|
mul.f32 %f357, %f38, 0f00000000;
|
|
mov.u32 %r113, 12;
|
|
mov.f32 %f356, 0f00000000;
|
|
add.u64 %rd116, %SP, 16;
|
|
cvta.to.local.u64 %rd115, %rd116;
|
|
mov.u32 %r112, 0;
|
|
mul.f32 %f77, %f35, %f357;
|
|
fma.rn.f32 %f254, %f32, %f358, %f77;
|
|
mul.f32 %f78, %f36, %f357;
|
|
fma.rn.f32 %f255, %f33, %f358, %f78;
|
|
mul.f32 %f79, %f37, %f357;
|
|
fma.rn.f32 %f256, %f34, %f358, %f79;
|
|
add.f32 %f249, %f254, %f45;
|
|
add.f32 %f250, %f255, %f46;
|
|
add.f32 %f251, %f256, %f47;
|
|
st.local.u32 [%rd115+8], %r112;
|
|
st.local.u32 [%rd115+4], %r112;
|
|
st.local.u32 [%rd115], %r112;
|
|
ld.global.u32 %r85, [root];
|
|
// inline asm
|
|
call _rt_trace_64, (%r85, %f170, %f171, %f172, %f249, %f250, %f251, %r112, %f356, %f177, %rd116, %r113);
|
|
// inline asm
|
|
ld.local.f32 %f83, [%rd115+4];
|
|
ld.local.f32 %f84, [%rd115];
|
|
add.f32 %f257, %f84, %f83;
|
|
ld.local.f32 %f85, [%rd115+8];
|
|
add.f32 %f258, %f257, %f85;
|
|
setp.eq.f32 %p20, %f258, 0f00000000;
|
|
@%p20 bra BB0_15;
|
|
|
|
mul.f32 %f259, %f83, %f83;
|
|
fma.rn.f32 %f260, %f84, %f84, %f259;
|
|
fma.rn.f32 %f261, %f85, %f85, %f260;
|
|
sqrt.rn.f32 %f86, %f261;
|
|
rcp.rn.f32 %f262, %f86;
|
|
mul.f32 %f263, %f262, %f84;
|
|
mul.f32 %f264, %f262, %f83;
|
|
mul.f32 %f265, %f262, %f85;
|
|
fma.rn.f32 %f266, %f41, %f263, %f170;
|
|
fma.rn.f32 %f267, %f41, %f264, %f171;
|
|
fma.rn.f32 %f268, %f41, %f265, %f172;
|
|
fma.rn.f32 %f87, %f86, %f249, %f266;
|
|
fma.rn.f32 %f88, %f86, %f250, %f267;
|
|
fma.rn.f32 %f89, %f86, %f251, %f268;
|
|
setp.geu.f32 %p21, %f86, %f376;
|
|
@%p21 bra BB0_15;
|
|
|
|
mov.f32 %f373, %f87;
|
|
mov.f32 %f374, %f88;
|
|
mov.f32 %f375, %f89;
|
|
mov.f32 %f376, %f86;
|
|
|
|
BB0_15:
|
|
mov.u32 %r115, 12;
|
|
mov.f32 %f359, 0f00000000;
|
|
add.u64 %rd119, %SP, 16;
|
|
cvta.to.local.u64 %rd118, %rd119;
|
|
mov.u32 %r114, 0;
|
|
fma.rn.f32 %f277, %f32, %f38, %f77;
|
|
fma.rn.f32 %f278, %f33, %f38, %f78;
|
|
fma.rn.f32 %f279, %f34, %f38, %f79;
|
|
add.f32 %f272, %f277, %f45;
|
|
add.f32 %f273, %f278, %f46;
|
|
add.f32 %f274, %f279, %f47;
|
|
st.local.u32 [%rd118+8], %r114;
|
|
st.local.u32 [%rd118+4], %r114;
|
|
st.local.u32 [%rd118], %r114;
|
|
ld.global.u32 %r88, [root];
|
|
// inline asm
|
|
call _rt_trace_64, (%r88, %f170, %f171, %f172, %f272, %f273, %f274, %r114, %f359, %f177, %rd119, %r115);
|
|
// inline asm
|
|
ld.local.f32 %f97, [%rd118+4];
|
|
ld.local.f32 %f98, [%rd118];
|
|
add.f32 %f280, %f98, %f97;
|
|
ld.local.f32 %f99, [%rd118+8];
|
|
add.f32 %f281, %f280, %f99;
|
|
setp.eq.f32 %p22, %f281, 0f00000000;
|
|
@%p22 bra BB0_18;
|
|
|
|
mul.f32 %f282, %f97, %f97;
|
|
fma.rn.f32 %f283, %f98, %f98, %f282;
|
|
fma.rn.f32 %f284, %f99, %f99, %f283;
|
|
sqrt.rn.f32 %f285, %f284;
|
|
rcp.rn.f32 %f286, %f285;
|
|
mul.f32 %f287, %f286, %f98;
|
|
mul.f32 %f288, %f286, %f97;
|
|
mul.f32 %f289, %f286, %f99;
|
|
fma.rn.f32 %f290, %f41, %f287, %f170;
|
|
fma.rn.f32 %f291, %f41, %f288, %f171;
|
|
fma.rn.f32 %f292, %f41, %f289, %f172;
|
|
fma.rn.f32 %f100, %f285, %f272, %f290;
|
|
fma.rn.f32 %f101, %f285, %f273, %f291;
|
|
fma.rn.f32 %f102, %f285, %f274, %f292;
|
|
setp.geu.f32 %p23, %f285, %f376;
|
|
@%p23 bra BB0_18;
|
|
|
|
mov.f32 %f373, %f100;
|
|
mov.f32 %f374, %f101;
|
|
mov.f32 %f375, %f102;
|
|
|
|
BB0_18:
|
|
abs.f32 %f293, %f373;
|
|
setp.gtu.f32 %p24, %f293, 0f7F800000;
|
|
selp.f32 %f384, 0f00000000, %f373, %p24;
|
|
abs.f32 %f294, %f374;
|
|
setp.gtu.f32 %p25, %f294, 0f7F800000;
|
|
selp.f32 %f385, 0f00000000, %f374, %p25;
|
|
abs.f32 %f295, %f375;
|
|
setp.gtu.f32 %p26, %f295, 0f7F800000;
|
|
selp.f32 %f386, 0f00000000, %f375, %p26;
|
|
ld.global.f32 %f109, [fakeBias];
|
|
setp.eq.f32 %p27, %f109, 0f00000000;
|
|
@%p27 bra BB0_20;
|
|
|
|
fma.rn.f32 %f384, %f366, %f109, %f384;
|
|
fma.rn.f32 %f385, %f367, %f109, %f385;
|
|
fma.rn.f32 %f386, %f368, %f109, %f386;
|
|
|
|
BB0_20:
|
|
mov.u32 %r117, 16;
|
|
mov.u64 %rd121, 0;
|
|
mov.u32 %r116, 2;
|
|
cvt.u64.u32 %rd90, %r3;
|
|
cvt.u64.u32 %rd89, %r2;
|
|
mov.u64 %rd93, image_HDR;
|
|
cvta.global.u64 %rd88, %rd93;
|
|
// inline asm
|
|
call (%rd87), _rt_buffer_get_64, (%rd88, %r116, %r117, %rd89, %rd90, %rd121, %rd121);
|
|
// inline asm
|
|
mov.f32 %f296, 0f3F800000;
|
|
st.v4.f32 [%rd87], {%f384, %f385, %f386, %f296};
|
|
ld.global.u8 %rs10, [imageEnabled];
|
|
and.b16 %rs11, %rs10, 32;
|
|
setp.eq.s16 %p28, %rs11, 0;
|
|
@%p28 bra BB0_28;
|
|
|
|
add.f32 %f361, %f141, 0f38D1B717;
|
|
mov.u32 %r119, 12;
|
|
mov.u32 %r118, 0;
|
|
mov.f32 %f360, 0f38D1B717;
|
|
mul.f32 %f304, %f361, 0f41000000;
|
|
add.u64 %rd94, %SP, 0;
|
|
cvta.to.local.u64 %rd5, %rd94;
|
|
st.local.u32 [%rd5+8], %r118;
|
|
st.local.u32 [%rd5+4], %r118;
|
|
st.local.u32 [%rd5], %r118;
|
|
ld.global.u32 %r93, [root];
|
|
// inline asm
|
|
call _rt_trace_64, (%r93, %f170, %f171, %f172, %f173, %f174, %f175, %r118, %f360, %f304, %rd94, %r119);
|
|
// inline asm
|
|
ld.local.f32 %f306, [%rd5+4];
|
|
ld.local.f32 %f307, [%rd5];
|
|
add.f32 %f308, %f307, %f306;
|
|
ld.local.f32 %f309, [%rd5+8];
|
|
add.f32 %f310, %f308, %f309;
|
|
setp.neu.f32 %p29, %f310, 0f00000000;
|
|
mov.f32 %f387, 0f437F0000;
|
|
@%p29 bra BB0_26;
|
|
|
|
mov.u32 %r121, 12;
|
|
mov.u32 %r120, 0;
|
|
mov.f32 %f362, 0f38D1B717;
|
|
st.local.u32 [%rd5+8], %r120;
|
|
st.local.u32 [%rd5+4], %r120;
|
|
st.local.u32 [%rd5], %r120;
|
|
ld.global.u32 %r96, [root];
|
|
// inline asm
|
|
call _rt_trace_64, (%r96, %f170, %f171, %f172, %f226, %f227, %f228, %r120, %f362, %f304, %rd94, %r121);
|
|
// inline asm
|
|
ld.local.f32 %f320, [%rd5+4];
|
|
ld.local.f32 %f321, [%rd5];
|
|
add.f32 %f322, %f321, %f320;
|
|
ld.local.f32 %f323, [%rd5+8];
|
|
add.f32 %f324, %f322, %f323;
|
|
setp.neu.f32 %p30, %f324, 0f00000000;
|
|
@%p30 bra BB0_26;
|
|
|
|
mov.u32 %r123, 12;
|
|
mov.u32 %r122, 0;
|
|
mov.f32 %f363, 0f38D1B717;
|
|
st.local.u32 [%rd5+8], %r122;
|
|
st.local.u32 [%rd5+4], %r122;
|
|
st.local.u32 [%rd5], %r122;
|
|
ld.global.u32 %r99, [root];
|
|
// inline asm
|
|
call _rt_trace_64, (%r99, %f170, %f171, %f172, %f249, %f250, %f251, %r122, %f363, %f304, %rd94, %r123);
|
|
// inline asm
|
|
ld.local.f32 %f334, [%rd5+4];
|
|
ld.local.f32 %f335, [%rd5];
|
|
add.f32 %f336, %f335, %f334;
|
|
ld.local.f32 %f337, [%rd5+8];
|
|
add.f32 %f338, %f336, %f337;
|
|
setp.neu.f32 %p31, %f338, 0f00000000;
|
|
@%p31 bra BB0_26;
|
|
|
|
mov.u32 %r125, 12;
|
|
mov.u32 %r124, 0;
|
|
mov.f32 %f364, 0f38D1B717;
|
|
st.local.u32 [%rd5+8], %r124;
|
|
st.local.u32 [%rd5+4], %r124;
|
|
st.local.u32 [%rd5], %r124;
|
|
ld.global.u32 %r102, [root];
|
|
// inline asm
|
|
call _rt_trace_64, (%r102, %f170, %f171, %f172, %f272, %f273, %f274, %r124, %f364, %f304, %rd94, %r125);
|
|
// inline asm
|
|
ld.local.f32 %f348, [%rd5+4];
|
|
ld.local.f32 %f349, [%rd5];
|
|
add.f32 %f350, %f349, %f348;
|
|
ld.local.f32 %f351, [%rd5+8];
|
|
add.f32 %f352, %f350, %f351;
|
|
setp.neu.f32 %p32, %f352, 0f00000000;
|
|
@%p32 bra BB0_26;
|
|
|
|
mov.f32 %f387, 0f00000000;
|
|
|
|
BB0_26:
|
|
mov.u32 %r127, 1;
|
|
mov.u64 %rd122, 0;
|
|
mov.u32 %r126, 2;
|
|
mov.u64 %rd104, image_Mask;
|
|
cvta.global.u64 %rd99, %rd104;
|
|
// inline asm
|
|
call (%rd98), _rt_buffer_get_64, (%rd99, %r126, %r127, %rd89, %rd90, %rd122, %rd122);
|
|
// inline asm
|
|
cvt.rzi.u32.f32 %r107, %f387;
|
|
st.u8 [%rd98], %r107;
|
|
|
|
BB0_28:
|
|
ret;
|
|
}
|
|
|
|
|