647 lines
22 KiB
Plaintext
647 lines
22 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[24];
|
||
|
|
.reg .b64 %SP;
|
||
|
|
.reg .b64 %SPL;
|
||
|
|
.reg .pred %p<28>;
|
||
|
|
.reg .b16 %rs<11>;
|
||
|
|
.reg .f32 %f<360>;
|
||
|
|
.reg .b32 %r<109>;
|
||
|
|
.reg .b64 %rd<101>;
|
||
|
|
|
||
|
|
|
||
|
|
mov.u64 %rd100, __local_depot0;
|
||
|
|
cvta.local.u64 %SP, %rd100;
|
||
|
|
ld.global.v2.u32 {%r6, %r7}, [pixelID];
|
||
|
|
cvt.u64.u32 %rd7, %r6;
|
||
|
|
cvt.u64.u32 %rd8, %r7;
|
||
|
|
mov.u64 %rd11, uvfacenormal;
|
||
|
|
cvta.global.u64 %rd6, %rd11;
|
||
|
|
mov.u32 %r4, 2;
|
||
|
|
mov.u32 %r5, 4;
|
||
|
|
mov.u64 %rd10, 0;
|
||
|
|
// inline asm
|
||
|
|
call (%rd5), _rt_buffer_get_64, (%rd6, %r4, %r5, %rd7, %rd8, %rd10, %rd10);
|
||
|
|
// inline asm
|
||
|
|
ld.u32 %r1, [%rd5];
|
||
|
|
shr.u32 %r10, %r1, 16;
|
||
|
|
cvt.u16.u32 %rs1, %r10;
|
||
|
|
and.b16 %rs2, %rs1, 255;
|
||
|
|
cvt.u16.u32 %rs3, %r1;
|
||
|
|
or.b16 %rs4, %rs3, %rs2;
|
||
|
|
setp.eq.s16 %p1, %rs4, 0;
|
||
|
|
mov.f32 %f338, 0f00000000;
|
||
|
|
mov.f32 %f339, %f338;
|
||
|
|
mov.f32 %f340, %f338;
|
||
|
|
@%p1 bra BB0_2;
|
||
|
|
|
||
|
|
ld.u8 %rs5, [%rd5+1];
|
||
|
|
and.b16 %rs7, %rs3, 255;
|
||
|
|
cvt.rn.f32.u16 %f103, %rs7;
|
||
|
|
div.rn.f32 %f104, %f103, 0f437F0000;
|
||
|
|
fma.rn.f32 %f105, %f104, 0f40000000, 0fBF800000;
|
||
|
|
cvt.rn.f32.u16 %f106, %rs5;
|
||
|
|
div.rn.f32 %f107, %f106, 0f437F0000;
|
||
|
|
fma.rn.f32 %f108, %f107, 0f40000000, 0fBF800000;
|
||
|
|
cvt.rn.f32.u16 %f109, %rs2;
|
||
|
|
div.rn.f32 %f110, %f109, 0f437F0000;
|
||
|
|
fma.rn.f32 %f111, %f110, 0f40000000, 0fBF800000;
|
||
|
|
mul.f32 %f112, %f108, %f108;
|
||
|
|
fma.rn.f32 %f113, %f105, %f105, %f112;
|
||
|
|
fma.rn.f32 %f114, %f111, %f111, %f113;
|
||
|
|
sqrt.rn.f32 %f115, %f114;
|
||
|
|
rcp.rn.f32 %f116, %f115;
|
||
|
|
mul.f32 %f338, %f105, %f116;
|
||
|
|
mul.f32 %f339, %f108, %f116;
|
||
|
|
mul.f32 %f340, %f111, %f116;
|
||
|
|
|
||
|
|
BB0_2:
|
||
|
|
ld.global.v2.u32 {%r11, %r12}, [pixelID];
|
||
|
|
ld.global.v2.u32 {%r14, %r15}, [tileInfo];
|
||
|
|
add.s32 %r2, %r11, %r14;
|
||
|
|
add.s32 %r3, %r12, %r15;
|
||
|
|
setp.eq.f32 %p2, %f339, 0f00000000;
|
||
|
|
setp.eq.f32 %p3, %f338, 0f00000000;
|
||
|
|
and.pred %p4, %p3, %p2;
|
||
|
|
setp.eq.f32 %p5, %f340, 0f00000000;
|
||
|
|
and.pred %p6, %p4, %p5;
|
||
|
|
@%p6 bra BB0_24;
|
||
|
|
bra.uni BB0_3;
|
||
|
|
|
||
|
|
BB0_24:
|
||
|
|
cvt.u64.u32 %rd84, %r2;
|
||
|
|
cvt.u64.u32 %rd85, %r3;
|
||
|
|
mov.u64 %rd88, image_HDR;
|
||
|
|
cvta.global.u64 %rd83, %rd88;
|
||
|
|
mov.u32 %r91, 16;
|
||
|
|
// inline asm
|
||
|
|
call (%rd82), _rt_buffer_get_64, (%rd83, %r4, %r91, %rd84, %rd85, %rd10, %rd10);
|
||
|
|
// inline asm
|
||
|
|
mov.f32 %f326, 0f00000000;
|
||
|
|
st.v4.f32 [%rd82], {%f326, %f326, %f326, %f326};
|
||
|
|
bra.uni BB0_25;
|
||
|
|
|
||
|
|
BB0_3:
|
||
|
|
ld.global.v2.u32 {%r36, %r37}, [pixelID];
|
||
|
|
cvt.u64.u32 %rd14, %r36;
|
||
|
|
cvt.u64.u32 %rd15, %r37;
|
||
|
|
mov.u64 %rd55, uvpos;
|
||
|
|
cvta.global.u64 %rd13, %rd55;
|
||
|
|
mov.u32 %r32, 16;
|
||
|
|
// inline asm
|
||
|
|
call (%rd12), _rt_buffer_get_64, (%rd13, %r4, %r32, %rd14, %rd15, %rd10, %rd10);
|
||
|
|
// inline asm
|
||
|
|
ld.f32 %f126, [%rd12];
|
||
|
|
ld.global.v2.u32 {%r40, %r41}, [pixelID];
|
||
|
|
cvt.u64.u32 %rd20, %r40;
|
||
|
|
cvt.u64.u32 %rd21, %r41;
|
||
|
|
// inline asm
|
||
|
|
call (%rd18), _rt_buffer_get_64, (%rd13, %r4, %r32, %rd20, %rd21, %rd10, %rd10);
|
||
|
|
// inline asm
|
||
|
|
ld.f32 %f127, [%rd18+4];
|
||
|
|
ld.global.v2.u32 {%r44, %r45}, [pixelID];
|
||
|
|
cvt.u64.u32 %rd26, %r44;
|
||
|
|
cvt.u64.u32 %rd27, %r45;
|
||
|
|
// inline asm
|
||
|
|
call (%rd24), _rt_buffer_get_64, (%rd13, %r4, %r32, %rd26, %rd27, %rd10, %rd10);
|
||
|
|
// inline asm
|
||
|
|
abs.f32 %f128, %f126;
|
||
|
|
setp.gtu.f32 %p7, %f128, 0f7F800000;
|
||
|
|
abs.f32 %f129, %f127;
|
||
|
|
setp.gtu.f32 %p8, %f129, 0f7F800000;
|
||
|
|
ld.f32 %f130, [%rd24+8];
|
||
|
|
abs.f32 %f131, %f130;
|
||
|
|
setp.gtu.f32 %p9, %f131, 0f7F800000;
|
||
|
|
ld.global.v2.u32 {%r48, %r49}, [pixelID];
|
||
|
|
cvt.u64.u32 %rd32, %r48;
|
||
|
|
cvt.u64.u32 %rd33, %r49;
|
||
|
|
// inline asm
|
||
|
|
call (%rd30), _rt_buffer_get_64, (%rd13, %r4, %r32, %rd32, %rd33, %rd10, %rd10);
|
||
|
|
// inline asm
|
||
|
|
ld.f32 %f132, [%rd30+12];
|
||
|
|
add.f32 %f7, %f132, 0f38D1B717;
|
||
|
|
mul.f32 %f133, %f126, 0f3456BF95;
|
||
|
|
selp.f32 %f134, 0f00000000, %f133, %p7;
|
||
|
|
mul.f32 %f135, %f127, 0f3456BF95;
|
||
|
|
selp.f32 %f136, 0f00000000, %f135, %p8;
|
||
|
|
mul.f32 %f137, %f130, 0f3456BF95;
|
||
|
|
selp.f32 %f138, 0f00000000, %f137, %p9;
|
||
|
|
abs.f32 %f139, %f338;
|
||
|
|
div.rn.f32 %f140, %f134, %f139;
|
||
|
|
abs.f32 %f141, %f339;
|
||
|
|
div.rn.f32 %f142, %f136, %f141;
|
||
|
|
abs.f32 %f143, %f340;
|
||
|
|
div.rn.f32 %f144, %f138, %f143;
|
||
|
|
abs.f32 %f145, %f140;
|
||
|
|
abs.f32 %f146, %f142;
|
||
|
|
abs.f32 %f147, %f144;
|
||
|
|
mov.f32 %f148, 0f38D1B717;
|
||
|
|
max.f32 %f149, %f145, %f148;
|
||
|
|
max.f32 %f150, %f146, %f148;
|
||
|
|
max.f32 %f151, %f147, %f148;
|
||
|
|
ld.global.v2.u32 {%r52, %r53}, [pixelID];
|
||
|
|
cvt.u64.u32 %rd38, %r52;
|
||
|
|
cvt.u64.u32 %rd39, %r53;
|
||
|
|
mov.u64 %rd56, uvsmoothpos;
|
||
|
|
cvta.global.u64 %rd37, %rd56;
|
||
|
|
// inline asm
|
||
|
|
call (%rd36), _rt_buffer_get_64, (%rd37, %r4, %r32, %rd38, %rd39, %rd10, %rd10);
|
||
|
|
// inline asm
|
||
|
|
ld.f32 %f345, [%rd36];
|
||
|
|
ld.global.v2.u32 {%r56, %r57}, [pixelID];
|
||
|
|
cvt.u64.u32 %rd44, %r56;
|
||
|
|
cvt.u64.u32 %rd45, %r57;
|
||
|
|
// inline asm
|
||
|
|
call (%rd42), _rt_buffer_get_64, (%rd37, %r4, %r32, %rd44, %rd45, %rd10, %rd10);
|
||
|
|
// inline asm
|
||
|
|
ld.f32 %f346, [%rd42+4];
|
||
|
|
ld.global.v2.u32 {%r60, %r61}, [pixelID];
|
||
|
|
cvt.u64.u32 %rd50, %r60;
|
||
|
|
cvt.u64.u32 %rd51, %r61;
|
||
|
|
// inline asm
|
||
|
|
call (%rd48), _rt_buffer_get_64, (%rd37, %r4, %r32, %rd50, %rd51, %rd10, %rd10);
|
||
|
|
// inline asm
|
||
|
|
fma.rn.f32 %f10, %f338, %f149, %f345;
|
||
|
|
fma.rn.f32 %f11, %f339, %f150, %f346;
|
||
|
|
ld.f32 %f347, [%rd48+8];
|
||
|
|
fma.rn.f32 %f13, %f340, %f151, %f347;
|
||
|
|
setp.gt.f32 %p10, %f139, %f143;
|
||
|
|
neg.f32 %f152, %f339;
|
||
|
|
selp.f32 %f153, %f152, 0f00000000, %p10;
|
||
|
|
neg.f32 %f154, %f340;
|
||
|
|
selp.f32 %f155, %f338, %f154, %p10;
|
||
|
|
selp.f32 %f156, 0f00000000, %f339, %p10;
|
||
|
|
mul.f32 %f157, %f155, %f155;
|
||
|
|
fma.rn.f32 %f158, %f153, %f153, %f157;
|
||
|
|
fma.rn.f32 %f159, %f156, %f156, %f158;
|
||
|
|
sqrt.rn.f32 %f160, %f159;
|
||
|
|
rcp.rn.f32 %f161, %f160;
|
||
|
|
mul.f32 %f14, %f153, %f161;
|
||
|
|
mul.f32 %f15, %f155, %f161;
|
||
|
|
mul.f32 %f16, %f156, %f161;
|
||
|
|
mul.f32 %f162, %f340, %f15;
|
||
|
|
mul.f32 %f163, %f339, %f16;
|
||
|
|
sub.f32 %f17, %f162, %f163;
|
||
|
|
mul.f32 %f164, %f338, %f16;
|
||
|
|
mul.f32 %f165, %f340, %f14;
|
||
|
|
sub.f32 %f18, %f164, %f165;
|
||
|
|
mul.f32 %f166, %f339, %f14;
|
||
|
|
mul.f32 %f167, %f338, %f15;
|
||
|
|
sub.f32 %f19, %f166, %f167;
|
||
|
|
mov.f32 %f168, 0f3F8147AE;
|
||
|
|
sqrt.rn.f32 %f169, %f168;
|
||
|
|
rcp.rn.f32 %f20, %f169;
|
||
|
|
neg.f32 %f21, %f20;
|
||
|
|
mul.f32 %f22, %f20, 0f00000000;
|
||
|
|
mul.f32 %f170, %f20, 0f3DCCCCCD;
|
||
|
|
mul.f32 %f171, %f10, 0f3456BF95;
|
||
|
|
mul.f32 %f172, %f11, 0f3456BF95;
|
||
|
|
mul.f32 %f173, %f13, 0f3456BF95;
|
||
|
|
abs.f32 %f174, %f171;
|
||
|
|
abs.f32 %f175, %f172;
|
||
|
|
abs.f32 %f176, %f173;
|
||
|
|
max.f32 %f177, %f174, %f175;
|
||
|
|
max.f32 %f178, %f177, %f176;
|
||
|
|
max.f32 %f23, %f178, %f148;
|
||
|
|
mul.f32 %f124, %f7, 0f3FB504F3;
|
||
|
|
mul.f32 %f25, %f14, %f22;
|
||
|
|
mul.f32 %f26, %f15, %f22;
|
||
|
|
mul.f32 %f27, %f16, %f22;
|
||
|
|
fma.rn.f32 %f179, %f17, %f21, %f25;
|
||
|
|
fma.rn.f32 %f180, %f18, %f21, %f26;
|
||
|
|
fma.rn.f32 %f181, %f19, %f21, %f27;
|
||
|
|
mul.f32 %f28, %f338, %f170;
|
||
|
|
mul.f32 %f29, %f339, %f170;
|
||
|
|
mul.f32 %f30, %f340, %f170;
|
||
|
|
add.f32 %f120, %f179, %f28;
|
||
|
|
add.f32 %f121, %f180, %f29;
|
||
|
|
add.f32 %f122, %f181, %f30;
|
||
|
|
add.u64 %rd54, %SP, 12;
|
||
|
|
cvta.to.local.u64 %rd57, %rd54;
|
||
|
|
mov.u32 %r34, 0;
|
||
|
|
st.local.u32 [%rd57+8], %r34;
|
||
|
|
st.local.u32 [%rd57+4], %r34;
|
||
|
|
st.local.u32 [%rd57], %r34;
|
||
|
|
ld.global.u32 %r33, [root];
|
||
|
|
mov.f32 %f123, 0f00000000;
|
||
|
|
mov.u32 %r35, 12;
|
||
|
|
// inline asm
|
||
|
|
call _rt_trace_64, (%r33, %f10, %f11, %f13, %f120, %f121, %f122, %r34, %f123, %f124, %rd54, %r35);
|
||
|
|
// inline asm
|
||
|
|
ld.local.f32 %f34, [%rd57+4];
|
||
|
|
ld.local.f32 %f35, [%rd57];
|
||
|
|
add.f32 %f182, %f35, %f34;
|
||
|
|
ld.local.f32 %f36, [%rd57+8];
|
||
|
|
add.f32 %f183, %f182, %f36;
|
||
|
|
mov.f32 %f348, 0f47C34F80;
|
||
|
|
setp.eq.f32 %p11, %f183, 0f00000000;
|
||
|
|
@%p11 bra BB0_6;
|
||
|
|
|
||
|
|
mul.f32 %f185, %f34, %f34;
|
||
|
|
fma.rn.f32 %f186, %f35, %f35, %f185;
|
||
|
|
fma.rn.f32 %f187, %f36, %f36, %f186;
|
||
|
|
sqrt.rn.f32 %f37, %f187;
|
||
|
|
rcp.rn.f32 %f188, %f37;
|
||
|
|
mul.f32 %f189, %f188, %f35;
|
||
|
|
mul.f32 %f190, %f188, %f34;
|
||
|
|
mul.f32 %f191, %f188, %f36;
|
||
|
|
fma.rn.f32 %f192, %f23, %f189, %f10;
|
||
|
|
fma.rn.f32 %f193, %f23, %f190, %f11;
|
||
|
|
fma.rn.f32 %f194, %f23, %f191, %f13;
|
||
|
|
fma.rn.f32 %f38, %f37, %f120, %f192;
|
||
|
|
fma.rn.f32 %f39, %f37, %f121, %f193;
|
||
|
|
fma.rn.f32 %f40, %f37, %f122, %f194;
|
||
|
|
setp.geu.f32 %p12, %f37, 0f47C34F80;
|
||
|
|
@%p12 bra BB0_6;
|
||
|
|
|
||
|
|
mov.f32 %f345, %f38;
|
||
|
|
mov.f32 %f346, %f39;
|
||
|
|
mov.f32 %f347, %f40;
|
||
|
|
mov.f32 %f348, %f37;
|
||
|
|
|
||
|
|
BB0_6:
|
||
|
|
mov.u32 %r93, 12;
|
||
|
|
mov.f32 %f327, 0f00000000;
|
||
|
|
add.u64 %rd90, %SP, 12;
|
||
|
|
cvta.to.local.u64 %rd89, %rd90;
|
||
|
|
mov.u32 %r92, 0;
|
||
|
|
fma.rn.f32 %f203, %f17, %f20, %f25;
|
||
|
|
fma.rn.f32 %f204, %f18, %f20, %f26;
|
||
|
|
fma.rn.f32 %f205, %f19, %f20, %f27;
|
||
|
|
add.f32 %f198, %f203, %f28;
|
||
|
|
add.f32 %f199, %f204, %f29;
|
||
|
|
add.f32 %f200, %f205, %f30;
|
||
|
|
st.local.u32 [%rd89+8], %r92;
|
||
|
|
st.local.u32 [%rd89+4], %r92;
|
||
|
|
st.local.u32 [%rd89], %r92;
|
||
|
|
ld.global.u32 %r64, [root];
|
||
|
|
// inline asm
|
||
|
|
call _rt_trace_64, (%r64, %f10, %f11, %f13, %f198, %f199, %f200, %r92, %f327, %f124, %rd90, %r93);
|
||
|
|
// inline asm
|
||
|
|
ld.local.f32 %f48, [%rd89+4];
|
||
|
|
ld.local.f32 %f49, [%rd89];
|
||
|
|
add.f32 %f206, %f49, %f48;
|
||
|
|
ld.local.f32 %f50, [%rd89+8];
|
||
|
|
add.f32 %f207, %f206, %f50;
|
||
|
|
setp.eq.f32 %p13, %f207, 0f00000000;
|
||
|
|
@%p13 bra BB0_9;
|
||
|
|
|
||
|
|
mul.f32 %f208, %f48, %f48;
|
||
|
|
fma.rn.f32 %f209, %f49, %f49, %f208;
|
||
|
|
fma.rn.f32 %f210, %f50, %f50, %f209;
|
||
|
|
sqrt.rn.f32 %f51, %f210;
|
||
|
|
rcp.rn.f32 %f211, %f51;
|
||
|
|
mul.f32 %f212, %f211, %f49;
|
||
|
|
mul.f32 %f213, %f211, %f48;
|
||
|
|
mul.f32 %f214, %f211, %f50;
|
||
|
|
fma.rn.f32 %f215, %f23, %f212, %f10;
|
||
|
|
fma.rn.f32 %f216, %f23, %f213, %f11;
|
||
|
|
fma.rn.f32 %f217, %f23, %f214, %f13;
|
||
|
|
fma.rn.f32 %f52, %f51, %f198, %f215;
|
||
|
|
fma.rn.f32 %f53, %f51, %f199, %f216;
|
||
|
|
fma.rn.f32 %f54, %f51, %f200, %f217;
|
||
|
|
setp.geu.f32 %p14, %f51, %f348;
|
||
|
|
@%p14 bra BB0_9;
|
||
|
|
|
||
|
|
mov.f32 %f345, %f52;
|
||
|
|
mov.f32 %f346, %f53;
|
||
|
|
mov.f32 %f347, %f54;
|
||
|
|
mov.f32 %f348, %f51;
|
||
|
|
|
||
|
|
BB0_9:
|
||
|
|
neg.f32 %f330, %f20;
|
||
|
|
mul.f32 %f329, %f20, 0f00000000;
|
||
|
|
mov.u32 %r95, 12;
|
||
|
|
mov.f32 %f328, 0f00000000;
|
||
|
|
add.u64 %rd93, %SP, 12;
|
||
|
|
cvta.to.local.u64 %rd92, %rd93;
|
||
|
|
mov.u32 %r94, 0;
|
||
|
|
mul.f32 %f59, %f17, %f329;
|
||
|
|
fma.rn.f32 %f226, %f14, %f330, %f59;
|
||
|
|
mul.f32 %f60, %f18, %f329;
|
||
|
|
fma.rn.f32 %f227, %f15, %f330, %f60;
|
||
|
|
mul.f32 %f61, %f19, %f329;
|
||
|
|
fma.rn.f32 %f228, %f16, %f330, %f61;
|
||
|
|
add.f32 %f221, %f226, %f28;
|
||
|
|
add.f32 %f222, %f227, %f29;
|
||
|
|
add.f32 %f223, %f228, %f30;
|
||
|
|
st.local.u32 [%rd92+8], %r94;
|
||
|
|
st.local.u32 [%rd92+4], %r94;
|
||
|
|
st.local.u32 [%rd92], %r94;
|
||
|
|
ld.global.u32 %r67, [root];
|
||
|
|
// inline asm
|
||
|
|
call _rt_trace_64, (%r67, %f10, %f11, %f13, %f221, %f222, %f223, %r94, %f328, %f124, %rd93, %r95);
|
||
|
|
// inline asm
|
||
|
|
ld.local.f32 %f65, [%rd92+4];
|
||
|
|
ld.local.f32 %f66, [%rd92];
|
||
|
|
add.f32 %f229, %f66, %f65;
|
||
|
|
ld.local.f32 %f67, [%rd92+8];
|
||
|
|
add.f32 %f230, %f229, %f67;
|
||
|
|
setp.eq.f32 %p15, %f230, 0f00000000;
|
||
|
|
@%p15 bra BB0_12;
|
||
|
|
|
||
|
|
mul.f32 %f231, %f65, %f65;
|
||
|
|
fma.rn.f32 %f232, %f66, %f66, %f231;
|
||
|
|
fma.rn.f32 %f233, %f67, %f67, %f232;
|
||
|
|
sqrt.rn.f32 %f68, %f233;
|
||
|
|
rcp.rn.f32 %f234, %f68;
|
||
|
|
mul.f32 %f235, %f234, %f66;
|
||
|
|
mul.f32 %f236, %f234, %f65;
|
||
|
|
mul.f32 %f237, %f234, %f67;
|
||
|
|
fma.rn.f32 %f238, %f23, %f235, %f10;
|
||
|
|
fma.rn.f32 %f239, %f23, %f236, %f11;
|
||
|
|
fma.rn.f32 %f240, %f23, %f237, %f13;
|
||
|
|
fma.rn.f32 %f69, %f68, %f221, %f238;
|
||
|
|
fma.rn.f32 %f70, %f68, %f222, %f239;
|
||
|
|
fma.rn.f32 %f71, %f68, %f223, %f240;
|
||
|
|
setp.geu.f32 %p16, %f68, %f348;
|
||
|
|
@%p16 bra BB0_12;
|
||
|
|
|
||
|
|
mov.f32 %f345, %f69;
|
||
|
|
mov.f32 %f346, %f70;
|
||
|
|
mov.f32 %f347, %f71;
|
||
|
|
mov.f32 %f348, %f68;
|
||
|
|
|
||
|
|
BB0_12:
|
||
|
|
mov.u32 %r97, 12;
|
||
|
|
mov.f32 %f331, 0f00000000;
|
||
|
|
add.u64 %rd96, %SP, 12;
|
||
|
|
cvta.to.local.u64 %rd95, %rd96;
|
||
|
|
mov.u32 %r96, 0;
|
||
|
|
fma.rn.f32 %f249, %f14, %f20, %f59;
|
||
|
|
fma.rn.f32 %f250, %f15, %f20, %f60;
|
||
|
|
fma.rn.f32 %f251, %f16, %f20, %f61;
|
||
|
|
add.f32 %f244, %f249, %f28;
|
||
|
|
add.f32 %f245, %f250, %f29;
|
||
|
|
add.f32 %f246, %f251, %f30;
|
||
|
|
st.local.u32 [%rd95+8], %r96;
|
||
|
|
st.local.u32 [%rd95+4], %r96;
|
||
|
|
st.local.u32 [%rd95], %r96;
|
||
|
|
ld.global.u32 %r70, [root];
|
||
|
|
// inline asm
|
||
|
|
call _rt_trace_64, (%r70, %f10, %f11, %f13, %f244, %f245, %f246, %r96, %f331, %f124, %rd96, %r97);
|
||
|
|
// inline asm
|
||
|
|
ld.local.f32 %f79, [%rd95+4];
|
||
|
|
ld.local.f32 %f80, [%rd95];
|
||
|
|
add.f32 %f252, %f80, %f79;
|
||
|
|
ld.local.f32 %f81, [%rd95+8];
|
||
|
|
add.f32 %f253, %f252, %f81;
|
||
|
|
setp.eq.f32 %p17, %f253, 0f00000000;
|
||
|
|
@%p17 bra BB0_15;
|
||
|
|
|
||
|
|
mul.f32 %f254, %f79, %f79;
|
||
|
|
fma.rn.f32 %f255, %f80, %f80, %f254;
|
||
|
|
fma.rn.f32 %f256, %f81, %f81, %f255;
|
||
|
|
sqrt.rn.f32 %f257, %f256;
|
||
|
|
rcp.rn.f32 %f258, %f257;
|
||
|
|
mul.f32 %f259, %f258, %f80;
|
||
|
|
mul.f32 %f260, %f258, %f79;
|
||
|
|
mul.f32 %f261, %f258, %f81;
|
||
|
|
fma.rn.f32 %f262, %f23, %f259, %f10;
|
||
|
|
fma.rn.f32 %f263, %f23, %f260, %f11;
|
||
|
|
fma.rn.f32 %f264, %f23, %f261, %f13;
|
||
|
|
fma.rn.f32 %f82, %f257, %f244, %f262;
|
||
|
|
fma.rn.f32 %f83, %f257, %f245, %f263;
|
||
|
|
fma.rn.f32 %f84, %f257, %f246, %f264;
|
||
|
|
setp.geu.f32 %p18, %f257, %f348;
|
||
|
|
@%p18 bra BB0_15;
|
||
|
|
|
||
|
|
mov.f32 %f345, %f82;
|
||
|
|
mov.f32 %f346, %f83;
|
||
|
|
mov.f32 %f347, %f84;
|
||
|
|
|
||
|
|
BB0_15:
|
||
|
|
abs.f32 %f265, %f345;
|
||
|
|
setp.gtu.f32 %p19, %f265, 0f7F800000;
|
||
|
|
selp.f32 %f356, 0f00000000, %f345, %p19;
|
||
|
|
abs.f32 %f266, %f346;
|
||
|
|
setp.gtu.f32 %p20, %f266, 0f7F800000;
|
||
|
|
selp.f32 %f357, 0f00000000, %f346, %p20;
|
||
|
|
abs.f32 %f267, %f347;
|
||
|
|
setp.gtu.f32 %p21, %f267, 0f7F800000;
|
||
|
|
selp.f32 %f358, 0f00000000, %f347, %p21;
|
||
|
|
ld.global.f32 %f91, [fakeBias];
|
||
|
|
setp.eq.f32 %p22, %f91, 0f00000000;
|
||
|
|
@%p22 bra BB0_17;
|
||
|
|
|
||
|
|
fma.rn.f32 %f356, %f338, %f91, %f356;
|
||
|
|
fma.rn.f32 %f357, %f339, %f91, %f357;
|
||
|
|
fma.rn.f32 %f358, %f340, %f91, %f358;
|
||
|
|
|
||
|
|
BB0_17:
|
||
|
|
mov.u32 %r99, 16;
|
||
|
|
mov.u64 %rd98, 0;
|
||
|
|
mov.u32 %r98, 2;
|
||
|
|
cvt.u64.u32 %rd67, %r3;
|
||
|
|
cvt.u64.u32 %rd66, %r2;
|
||
|
|
mov.u64 %rd70, image_HDR;
|
||
|
|
cvta.global.u64 %rd65, %rd70;
|
||
|
|
// inline asm
|
||
|
|
call (%rd64), _rt_buffer_get_64, (%rd65, %r98, %r99, %rd66, %rd67, %rd98, %rd98);
|
||
|
|
// inline asm
|
||
|
|
mov.f32 %f268, 0f3F800000;
|
||
|
|
st.v4.f32 [%rd64], {%f356, %f357, %f358, %f268};
|
||
|
|
ld.global.u8 %rs9, [imageEnabled];
|
||
|
|
and.b16 %rs10, %rs9, 32;
|
||
|
|
setp.eq.s16 %p23, %rs10, 0;
|
||
|
|
@%p23 bra BB0_25;
|
||
|
|
|
||
|
|
add.f32 %f333, %f132, 0f38D1B717;
|
||
|
|
mov.f32 %f332, 0f38D1B717;
|
||
|
|
mov.u32 %r101, 12;
|
||
|
|
mov.u32 %r100, 0;
|
||
|
|
mul.f32 %f276, %f333, 0f41000000;
|
||
|
|
add.u64 %rd71, %SP, 0;
|
||
|
|
cvta.to.local.u64 %rd4, %rd71;
|
||
|
|
st.local.u32 [%rd4+8], %r100;
|
||
|
|
st.local.u32 [%rd4+4], %r100;
|
||
|
|
st.local.u32 [%rd4], %r100;
|
||
|
|
ld.global.u32 %r75, [root];
|
||
|
|
// inline asm
|
||
|
|
call _rt_trace_64, (%r75, %f10, %f11, %f13, %f120, %f121, %f122, %r100, %f332, %f276, %rd71, %r101);
|
||
|
|
// inline asm
|
||
|
|
ld.local.f32 %f278, [%rd4+4];
|
||
|
|
ld.local.f32 %f279, [%rd4];
|
||
|
|
add.f32 %f280, %f279, %f278;
|
||
|
|
ld.local.f32 %f281, [%rd4+8];
|
||
|
|
add.f32 %f282, %f280, %f281;
|
||
|
|
setp.neu.f32 %p24, %f282, 0f00000000;
|
||
|
|
mov.f32 %f359, 0f437F0000;
|
||
|
|
@%p24 bra BB0_23;
|
||
|
|
|
||
|
|
mov.f32 %f334, 0f38D1B717;
|
||
|
|
mov.u32 %r103, 12;
|
||
|
|
mov.u32 %r102, 0;
|
||
|
|
st.local.u32 [%rd4+8], %r102;
|
||
|
|
st.local.u32 [%rd4+4], %r102;
|
||
|
|
st.local.u32 [%rd4], %r102;
|
||
|
|
ld.global.u32 %r78, [root];
|
||
|
|
// inline asm
|
||
|
|
call _rt_trace_64, (%r78, %f10, %f11, %f13, %f198, %f199, %f200, %r102, %f334, %f276, %rd71, %r103);
|
||
|
|
// inline asm
|
||
|
|
ld.local.f32 %f292, [%rd4+4];
|
||
|
|
ld.local.f32 %f293, [%rd4];
|
||
|
|
add.f32 %f294, %f293, %f292;
|
||
|
|
ld.local.f32 %f295, [%rd4+8];
|
||
|
|
add.f32 %f296, %f294, %f295;
|
||
|
|
setp.neu.f32 %p25, %f296, 0f00000000;
|
||
|
|
@%p25 bra BB0_23;
|
||
|
|
|
||
|
|
mov.f32 %f335, 0f38D1B717;
|
||
|
|
mov.u32 %r105, 12;
|
||
|
|
mov.u32 %r104, 0;
|
||
|
|
st.local.u32 [%rd4+8], %r104;
|
||
|
|
st.local.u32 [%rd4+4], %r104;
|
||
|
|
st.local.u32 [%rd4], %r104;
|
||
|
|
ld.global.u32 %r81, [root];
|
||
|
|
// inline asm
|
||
|
|
call _rt_trace_64, (%r81, %f10, %f11, %f13, %f221, %f222, %f223, %r104, %f335, %f276, %rd71, %r105);
|
||
|
|
// inline asm
|
||
|
|
ld.local.f32 %f306, [%rd4+4];
|
||
|
|
ld.local.f32 %f307, [%rd4];
|
||
|
|
add.f32 %f308, %f307, %f306;
|
||
|
|
ld.local.f32 %f309, [%rd4+8];
|
||
|
|
add.f32 %f310, %f308, %f309;
|
||
|
|
setp.neu.f32 %p26, %f310, 0f00000000;
|
||
|
|
@%p26 bra BB0_23;
|
||
|
|
|
||
|
|
mov.f32 %f336, 0f38D1B717;
|
||
|
|
mov.u32 %r107, 12;
|
||
|
|
mov.u32 %r106, 0;
|
||
|
|
st.local.u32 [%rd4+8], %r106;
|
||
|
|
st.local.u32 [%rd4+4], %r106;
|
||
|
|
st.local.u32 [%rd4], %r106;
|
||
|
|
ld.global.u32 %r84, [root];
|
||
|
|
// inline asm
|
||
|
|
call _rt_trace_64, (%r84, %f10, %f11, %f13, %f244, %f245, %f246, %r106, %f336, %f276, %rd71, %r107);
|
||
|
|
// inline asm
|
||
|
|
ld.local.f32 %f320, [%rd4+4];
|
||
|
|
ld.local.f32 %f321, [%rd4];
|
||
|
|
add.f32 %f322, %f321, %f320;
|
||
|
|
ld.local.f32 %f323, [%rd4+8];
|
||
|
|
add.f32 %f324, %f322, %f323;
|
||
|
|
setp.neu.f32 %p27, %f324, 0f00000000;
|
||
|
|
@%p27 bra BB0_23;
|
||
|
|
|
||
|
|
mov.f32 %f359, 0f00000000;
|
||
|
|
|
||
|
|
BB0_23:
|
||
|
|
mov.u64 %rd99, 0;
|
||
|
|
mov.u32 %r108, 2;
|
||
|
|
mov.u64 %rd81, image_Mask;
|
||
|
|
cvta.global.u64 %rd76, %rd81;
|
||
|
|
mov.u32 %r88, 1;
|
||
|
|
// inline asm
|
||
|
|
call (%rd75), _rt_buffer_get_64, (%rd76, %r108, %r88, %rd66, %rd67, %rd99, %rd99);
|
||
|
|
// inline asm
|
||
|
|
cvt.rzi.u32.f32 %r89, %f359;
|
||
|
|
st.u8 [%rd75], %r89;
|
||
|
|
|
||
|
|
BB0_25:
|
||
|
|
ret;
|
||
|
|
}
|
||
|
|
|
||
|
|
|