ArabDesert/Assets/Editor/x64/Bakery/addHalfSeparateIndirect.ptx

291 lines
11 KiB
Plaintext
Raw Normal View History

2024-05-25 09:10:35 +03:00
//
// 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 image2[1];
.global .align 1 .b8 imageIndirect[1];
.global .align 4 .f32 indirectIntensity;
.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_typeinfo17indirectIntensityE[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_typename17indirectIntensityE[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_typeenum17indirectIntensityE = 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_semantic17indirectIntensityE[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_annotation17indirectIntensityE[1];
.visible .entry _Z6oxMainv(
)
{
.reg .b16 %rs<19>;
.reg .f32 %f<20>;
.reg .b32 %r<79>;
.reg .b64 %rd<82>;
ld.global.v2.u32 {%r27, %r28}, [pixelID];
cvt.u64.u32 %rd3, %r27;
cvt.u64.u32 %rd4, %r28;
mov.u64 %rd79, image;
cvta.global.u64 %rd2, %rd79;
mov.u32 %r25, 2;
mov.u32 %r26, 8;
mov.u64 %rd78, 0;
// inline asm
call (%rd1), _rt_buffer_get_64, (%rd2, %r25, %r26, %rd3, %rd4, %rd78, %rd78);
// inline asm
ld.u16 %rs1, [%rd1+6];
// inline asm
{ cvt.f32.f16 %f1, %rs1;}
// inline asm
ld.global.v2.u32 {%r31, %r32}, [pixelID];
cvt.u64.u32 %rd9, %r31;
cvt.u64.u32 %rd10, %r32;
// inline asm
call (%rd7), _rt_buffer_get_64, (%rd2, %r25, %r26, %rd9, %rd10, %rd78, %rd78);
// inline asm
ld.u16 %rs2, [%rd7];
// inline asm
{ cvt.f32.f16 %f2, %rs2;}
// inline asm
ld.global.v2.u32 {%r35, %r36}, [pixelID];
cvt.u64.u32 %rd15, %r35;
cvt.u64.u32 %rd16, %r36;
// inline asm
call (%rd13), _rt_buffer_get_64, (%rd2, %r25, %r26, %rd15, %rd16, %rd78, %rd78);
// inline asm
ld.u16 %rs3, [%rd13+2];
// inline asm
{ cvt.f32.f16 %f3, %rs3;}
// inline asm
ld.global.v2.u32 {%r39, %r40}, [pixelID];
cvt.u64.u32 %rd21, %r39;
cvt.u64.u32 %rd22, %r40;
// inline asm
call (%rd19), _rt_buffer_get_64, (%rd2, %r25, %r26, %rd21, %rd22, %rd78, %rd78);
// inline asm
ld.u16 %rs4, [%rd19+4];
// inline asm
{ cvt.f32.f16 %f4, %rs4;}
// inline asm
ld.global.v2.u32 {%r43, %r44}, [pixelID];
cvt.u64.u32 %rd27, %r43;
cvt.u64.u32 %rd28, %r44;
mov.u64 %rd80, image2;
cvta.global.u64 %rd26, %rd80;
// inline asm
call (%rd25), _rt_buffer_get_64, (%rd26, %r25, %r26, %rd27, %rd28, %rd78, %rd78);
// inline asm
ld.u16 %rs5, [%rd25+6];
// inline asm
{ cvt.f32.f16 %f5, %rs5;}
// inline asm
ld.global.v2.u32 {%r47, %r48}, [pixelID];
cvt.u64.u32 %rd33, %r47;
cvt.u64.u32 %rd34, %r48;
// inline asm
call (%rd31), _rt_buffer_get_64, (%rd26, %r25, %r26, %rd33, %rd34, %rd78, %rd78);
// inline asm
ld.u16 %rs6, [%rd31];
// inline asm
{ cvt.f32.f16 %f6, %rs6;}
// inline asm
ld.global.v2.u32 {%r51, %r52}, [pixelID];
cvt.u64.u32 %rd39, %r51;
cvt.u64.u32 %rd40, %r52;
// inline asm
call (%rd37), _rt_buffer_get_64, (%rd26, %r25, %r26, %rd39, %rd40, %rd78, %rd78);
// inline asm
ld.u16 %rs7, [%rd37+2];
// inline asm
{ cvt.f32.f16 %f7, %rs7;}
// inline asm
ld.global.v2.u32 {%r55, %r56}, [pixelID];
cvt.u64.u32 %rd45, %r55;
cvt.u64.u32 %rd46, %r56;
// inline asm
call (%rd43), _rt_buffer_get_64, (%rd26, %r25, %r26, %rd45, %rd46, %rd78, %rd78);
// inline asm
ld.u16 %rs8, [%rd43+4];
// inline asm
{ cvt.f32.f16 %f8, %rs8;}
// inline asm
add.f32 %f9, %f2, %f6;
add.f32 %f10, %f3, %f7;
add.f32 %f11, %f4, %f8;
min.f32 %f12, %f1, %f5;
ld.global.v2.u32 {%r59, %r60}, [pixelID];
cvt.u64.u32 %rd51, %r59;
cvt.u64.u32 %rd52, %r60;
// inline asm
call (%rd49), _rt_buffer_get_64, (%rd26, %r25, %r26, %rd51, %rd52, %rd78, %rd78);
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs11, %f11;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs10, %f10;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs9, %f9;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs12, %f12;}
// inline asm
st.v4.u16 [%rd49], {%rs9, %rs10, %rs11, %rs12};
ld.global.v2.u32 {%r63, %r64}, [pixelID];
cvt.u64.u32 %rd57, %r63;
cvt.u64.u32 %rd58, %r64;
mov.u64 %rd81, imageIndirect;
cvta.global.u64 %rd56, %rd81;
// inline asm
call (%rd55), _rt_buffer_get_64, (%rd56, %r25, %r26, %rd57, %rd58, %rd78, %rd78);
// inline asm
ld.u16 %rs13, [%rd55];
// inline asm
{ cvt.f32.f16 %f13, %rs13;}
// inline asm
ld.global.v2.u32 {%r67, %r68}, [pixelID];
cvt.u64.u32 %rd63, %r67;
cvt.u64.u32 %rd64, %r68;
// inline asm
call (%rd61), _rt_buffer_get_64, (%rd56, %r25, %r26, %rd63, %rd64, %rd78, %rd78);
// inline asm
ld.u16 %rs14, [%rd61+2];
// inline asm
{ cvt.f32.f16 %f14, %rs14;}
// inline asm
ld.global.v2.u32 {%r71, %r72}, [pixelID];
cvt.u64.u32 %rd69, %r71;
cvt.u64.u32 %rd70, %r72;
// inline asm
call (%rd67), _rt_buffer_get_64, (%rd56, %r25, %r26, %rd69, %rd70, %rd78, %rd78);
// inline asm
ld.u16 %rs15, [%rd67+4];
// inline asm
{ cvt.f32.f16 %f15, %rs15;}
// inline asm
ld.global.f32 %f19, [indirectIntensity];
fma.rn.f32 %f16, %f2, %f19, %f13;
fma.rn.f32 %f17, %f3, %f19, %f14;
fma.rn.f32 %f18, %f4, %f19, %f15;
ld.global.v2.u32 {%r75, %r76}, [pixelID];
cvt.u64.u32 %rd75, %r75;
cvt.u64.u32 %rd76, %r76;
// inline asm
call (%rd73), _rt_buffer_get_64, (%rd56, %r25, %r26, %rd75, %rd76, %rd78, %rd78);
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs18, %f18;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs17, %f17;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs16, %f16;}
// inline asm
st.v4.u16 [%rd73], {%rs16, %rs17, %rs18, %rs12};
ret;
}