246 lines
9.7 KiB
Plaintext
246 lines
9.7 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 uvfacenormal[1];
|
||
|
|
.global .align 1 .b8 uvpos[1];
|
||
|
|
.global .align 1 .b8 uvsmoothpos[1];
|
||
|
|
.global .align 1 .b8 triangleMarks[1];
|
||
|
|
.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 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 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 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 _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];
|
||
|
|
|
||
|
|
.visible .entry _Z6oxMainv(
|
||
|
|
|
||
|
|
)
|
||
|
|
{
|
||
|
|
.local .align 4 .b8 __local_depot0[4];
|
||
|
|
.reg .b64 %SP;
|
||
|
|
.reg .b64 %SPL;
|
||
|
|
.reg .pred %p<8>;
|
||
|
|
.reg .b16 %rs<10>;
|
||
|
|
.reg .f32 %f<50>;
|
||
|
|
.reg .b32 %r<46>;
|
||
|
|
.reg .b64 %rd<51>;
|
||
|
|
|
||
|
|
|
||
|
|
mov.u64 %rd50, __local_depot0;
|
||
|
|
cvta.local.u64 %SP, %rd50;
|
||
|
|
ld.global.v2.u32 {%r4, %r5}, [pixelID];
|
||
|
|
cvt.u64.u32 %rd4, %r4;
|
||
|
|
cvt.u64.u32 %rd5, %r5;
|
||
|
|
mov.u64 %rd8, uvfacenormal;
|
||
|
|
cvta.global.u64 %rd3, %rd8;
|
||
|
|
mov.u32 %r2, 2;
|
||
|
|
mov.u32 %r3, 4;
|
||
|
|
mov.u64 %rd7, 0;
|
||
|
|
// inline asm
|
||
|
|
call (%rd2), _rt_buffer_get_64, (%rd3, %r2, %r3, %rd4, %rd5, %rd7, %rd7);
|
||
|
|
// inline asm
|
||
|
|
ld.u32 %r1, [%rd2];
|
||
|
|
shr.u32 %r8, %r1, 16;
|
||
|
|
cvt.u16.u32 %rs1, %r8;
|
||
|
|
and.b16 %rs2, %rs1, 255;
|
||
|
|
cvt.u16.u32 %rs3, %r1;
|
||
|
|
or.b16 %rs4, %rs3, %rs2;
|
||
|
|
setp.eq.s16 %p1, %rs4, 0;
|
||
|
|
mov.f32 %f47, 0f00000000;
|
||
|
|
mov.f32 %f48, %f47;
|
||
|
|
mov.f32 %f49, %f47;
|
||
|
|
@%p1 bra BB0_2;
|
||
|
|
|
||
|
|
ld.u8 %rs5, [%rd2+1];
|
||
|
|
and.b16 %rs7, %rs3, 255;
|
||
|
|
cvt.rn.f32.u16 %f11, %rs7;
|
||
|
|
div.rn.f32 %f12, %f11, 0f437F0000;
|
||
|
|
fma.rn.f32 %f13, %f12, 0f40000000, 0fBF800000;
|
||
|
|
cvt.rn.f32.u16 %f14, %rs5;
|
||
|
|
div.rn.f32 %f15, %f14, 0f437F0000;
|
||
|
|
fma.rn.f32 %f16, %f15, 0f40000000, 0fBF800000;
|
||
|
|
cvt.rn.f32.u16 %f17, %rs2;
|
||
|
|
div.rn.f32 %f18, %f17, 0f437F0000;
|
||
|
|
fma.rn.f32 %f19, %f18, 0f40000000, 0fBF800000;
|
||
|
|
mul.f32 %f20, %f16, %f16;
|
||
|
|
fma.rn.f32 %f21, %f13, %f13, %f20;
|
||
|
|
fma.rn.f32 %f22, %f19, %f19, %f21;
|
||
|
|
sqrt.rn.f32 %f23, %f22;
|
||
|
|
rcp.rn.f32 %f24, %f23;
|
||
|
|
mul.f32 %f47, %f13, %f24;
|
||
|
|
mul.f32 %f48, %f16, %f24;
|
||
|
|
mul.f32 %f49, %f19, %f24;
|
||
|
|
|
||
|
|
BB0_2:
|
||
|
|
setp.eq.f32 %p2, %f48, 0f00000000;
|
||
|
|
setp.eq.f32 %p3, %f47, 0f00000000;
|
||
|
|
and.pred %p4, %p3, %p2;
|
||
|
|
setp.eq.f32 %p5, %f49, 0f00000000;
|
||
|
|
and.pred %p6, %p4, %p5;
|
||
|
|
@%p6 bra BB0_5;
|
||
|
|
|
||
|
|
ld.global.v2.u32 {%r22, %r23}, [pixelID];
|
||
|
|
cvt.u64.u32 %rd11, %r22;
|
||
|
|
cvt.u64.u32 %rd12, %r23;
|
||
|
|
mov.u64 %rd40, uvpos;
|
||
|
|
cvta.global.u64 %rd10, %rd40;
|
||
|
|
mov.u32 %r10, 12;
|
||
|
|
// inline asm
|
||
|
|
call (%rd9), _rt_buffer_get_64, (%rd10, %r2, %r10, %rd11, %rd12, %rd7, %rd7);
|
||
|
|
// inline asm
|
||
|
|
ld.f32 %f33, [%rd9+8];
|
||
|
|
ld.f32 %f34, [%rd9+4];
|
||
|
|
ld.f32 %f35, [%rd9];
|
||
|
|
fma.rn.f32 %f25, %f47, 0f38D1B717, %f35;
|
||
|
|
fma.rn.f32 %f26, %f48, 0f38D1B717, %f34;
|
||
|
|
fma.rn.f32 %f27, %f49, 0f38D1B717, %f33;
|
||
|
|
ld.global.v2.u32 {%r26, %r27}, [pixelID];
|
||
|
|
cvt.u64.u32 %rd17, %r26;
|
||
|
|
cvt.u64.u32 %rd18, %r27;
|
||
|
|
mov.u64 %rd41, uvsmoothpos;
|
||
|
|
cvta.global.u64 %rd16, %rd41;
|
||
|
|
mov.u32 %r18, 16;
|
||
|
|
// inline asm
|
||
|
|
call (%rd15), _rt_buffer_get_64, (%rd16, %r2, %r18, %rd17, %rd18, %rd7, %rd7);
|
||
|
|
// inline asm
|
||
|
|
ld.f32 %f36, [%rd15];
|
||
|
|
ld.global.v2.u32 {%r30, %r31}, [pixelID];
|
||
|
|
cvt.u64.u32 %rd23, %r30;
|
||
|
|
cvt.u64.u32 %rd24, %r31;
|
||
|
|
// inline asm
|
||
|
|
call (%rd21), _rt_buffer_get_64, (%rd16, %r2, %r18, %rd23, %rd24, %rd7, %rd7);
|
||
|
|
// inline asm
|
||
|
|
ld.f32 %f37, [%rd21+4];
|
||
|
|
ld.global.v2.u32 {%r34, %r35}, [pixelID];
|
||
|
|
cvt.u64.u32 %rd29, %r34;
|
||
|
|
cvt.u64.u32 %rd30, %r35;
|
||
|
|
// inline asm
|
||
|
|
call (%rd27), _rt_buffer_get_64, (%rd16, %r2, %r18, %rd29, %rd30, %rd7, %rd7);
|
||
|
|
// inline asm
|
||
|
|
ld.f32 %f38, [%rd27+8];
|
||
|
|
ld.global.v2.u32 {%r38, %r39}, [pixelID];
|
||
|
|
cvt.u64.u32 %rd35, %r38;
|
||
|
|
cvt.u64.u32 %rd36, %r39;
|
||
|
|
// inline asm
|
||
|
|
call (%rd33), _rt_buffer_get_64, (%rd16, %r2, %r18, %rd35, %rd36, %rd7, %rd7);
|
||
|
|
// inline asm
|
||
|
|
ld.f32 %f7, [%rd33+12];
|
||
|
|
sub.f32 %f39, %f36, %f35;
|
||
|
|
sub.f32 %f40, %f37, %f34;
|
||
|
|
sub.f32 %f41, %f38, %f33;
|
||
|
|
mul.f32 %f42, %f40, %f40;
|
||
|
|
fma.rn.f32 %f43, %f39, %f39, %f42;
|
||
|
|
fma.rn.f32 %f44, %f41, %f41, %f43;
|
||
|
|
sqrt.rn.f32 %f32, %f44;
|
||
|
|
rcp.rn.f32 %f45, %f32;
|
||
|
|
mul.f32 %f28, %f39, %f45;
|
||
|
|
mul.f32 %f29, %f40, %f45;
|
||
|
|
mul.f32 %f30, %f41, %f45;
|
||
|
|
add.u64 %rd39, %SP, 0;
|
||
|
|
cvta.to.local.u64 %rd42, %rd39;
|
||
|
|
mov.u32 %r42, -1082130432;
|
||
|
|
st.local.u32 [%rd42], %r42;
|
||
|
|
ld.global.u32 %r19, [root];
|
||
|
|
mov.u32 %r20, 0;
|
||
|
|
mov.f32 %f31, 0f38D1B717;
|
||
|
|
// inline asm
|
||
|
|
call _rt_trace_64, (%r19, %f25, %f26, %f27, %f28, %f29, %f30, %r20, %f31, %f32, %rd39, %r3);
|
||
|
|
// inline asm
|
||
|
|
ld.local.f32 %f46, [%rd42];
|
||
|
|
setp.ltu.f32 %p7, %f46, 0f00000000;
|
||
|
|
@%p7 bra BB0_5;
|
||
|
|
|
||
|
|
cvt.rzi.u32.f32 %r45, %f7;
|
||
|
|
cvt.u64.u32 %rd45, %r45;
|
||
|
|
mov.u64 %rd49, triangleMarks;
|
||
|
|
cvta.global.u64 %rd44, %rd49;
|
||
|
|
mov.u32 %r44, 1;
|
||
|
|
// inline asm
|
||
|
|
call (%rd43), _rt_buffer_get_64, (%rd44, %r44, %r44, %rd45, %rd7, %rd7, %rd7);
|
||
|
|
// inline asm
|
||
|
|
mov.u16 %rs9, 255;
|
||
|
|
st.u8 [%rd43], %rs9;
|
||
|
|
|
||
|
|
BB0_5:
|
||
|
|
ret;
|
||
|
|
}
|
||
|
|
|
||
|
|
|