343 lines
13 KiB
Plaintext
343 lines
13 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 _Z11interpolatei
|
|
.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 1 .b8 lodPerTriangle[1];
|
|
.global .align 1 .b8 uvlod[1];
|
|
.global .align 4 .u32 imageEnabled;
|
|
.global .texref lightmap;
|
|
.global .align 16 .b8 tileInfo[16];
|
|
.global .align 4 .u32 additive;
|
|
.global .align 4 .f32 triAlpha;
|
|
.global .align 1 .b8 vertex_buffer[1];
|
|
.global .align 1 .b8 index_buffer[1];
|
|
.global .align 4 .u32 firstAlphaTriangle;
|
|
.global .align 1 .b8 vertex_buffer_uv[1];
|
|
.global .align 1 .b8 triangleAlphaIDs[1];
|
|
.global .align 1 .b8 alphaTextures[1];
|
|
.global .align 1 .b8 translucentTexIDs[1];
|
|
.global .align 4 .b8 ray[36];
|
|
.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_typeinfo8triAlphaE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo18firstAlphaTriangleE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo3rayE[8] = {82, 97, 121, 0, 36, 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_typename8triAlphaE[6] = {102, 108, 111, 97, 116, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typename18firstAlphaTriangleE[4] = {105, 110, 116, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typename3rayE[4] = {82, 97, 121, 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_typeenum8triAlphaE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum18firstAlphaTriangleE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum3rayE = 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 16 .b8 _ZN21rti_internal_semantic8triAlphaE[19] = {97, 116, 116, 114, 105, 98, 117, 116, 101, 32, 116, 114, 105, 65, 108, 112, 104, 97, 0};
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic18firstAlphaTriangleE[1];
|
|
.global .align 16 .b8 _ZN21rti_internal_semantic3rayE[13] = {114, 116, 67, 117, 114, 114, 101, 110, 116, 82, 97, 121, 0};
|
|
.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_annotation8triAlphaE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation18firstAlphaTriangleE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation3rayE[1];
|
|
|
|
.visible .entry _Z11interpolatei(
|
|
.param .u32 _Z11interpolatei_param_0
|
|
)
|
|
{
|
|
.reg .pred %p<6>;
|
|
.reg .f32 %f<80>;
|
|
.reg .b32 %r<67>;
|
|
.reg .b64 %rd<97>;
|
|
|
|
|
|
ld.param.u32 %r7, [_Z11interpolatei_param_0];
|
|
cvt.s64.s32 %rd4, %r7;
|
|
mov.u64 %rd8, lodPerTriangle;
|
|
cvta.global.u64 %rd3, %rd8;
|
|
mov.u32 %r8, 1;
|
|
mov.u32 %r9, 4;
|
|
mov.u64 %rd7, 0;
|
|
// inline asm
|
|
call (%rd2), _rt_buffer_get_64, (%rd3, %r8, %r9, %rd4, %rd7, %rd7, %rd7);
|
|
// inline asm
|
|
ld.u32 %r1, [%rd2];
|
|
setp.eq.s32 %p1, %r1, 0;
|
|
@%p1 bra BB0_3;
|
|
|
|
ld.global.v2.u32 {%r12, %r13}, [pixelID];
|
|
cvt.u64.u32 %rd11, %r12;
|
|
cvt.u64.u32 %rd12, %r13;
|
|
mov.u64 %rd15, uvlod;
|
|
cvta.global.u64 %rd10, %rd15;
|
|
mov.u32 %r10, 2;
|
|
// inline asm
|
|
call (%rd9), _rt_buffer_get_64, (%rd10, %r10, %r8, %rd11, %rd12, %rd7, %rd7);
|
|
// inline asm
|
|
ld.u8 %r16, [%rd9];
|
|
shl.b32 %r17, %r8, %r16;
|
|
and.b32 %r18, %r17, %r1;
|
|
setp.ne.s32 %p2, %r18, 0;
|
|
@%p2 bra BB0_3;
|
|
bra.uni BB0_2;
|
|
|
|
BB0_3:
|
|
// inline asm
|
|
call (%f8, %f9), _rt_get_triangle_barycentrics, ();
|
|
// inline asm
|
|
mov.u64 %rd22, index_buffer;
|
|
cvta.global.u64 %rd17, %rd22;
|
|
mov.u32 %r21, 12;
|
|
// inline asm
|
|
call (%rd16), _rt_buffer_get_64, (%rd17, %r8, %r21, %rd4, %rd7, %rd7, %rd7);
|
|
// inline asm
|
|
ld.u32 %r2, [%rd16];
|
|
ld.u32 %r3, [%rd16+4];
|
|
ld.u32 %r4, [%rd16+8];
|
|
ld.global.u32 %r5, [firstAlphaTriangle];
|
|
setp.gt.s32 %p3, %r5, %r7;
|
|
mov.f32 %f79, 0f3F800000;
|
|
sub.f32 %f11, %f79, %f8;
|
|
sub.f32 %f3, %f11, %f9;
|
|
@%p3 bra BB0_6;
|
|
|
|
cvt.s64.s32 %rd43, %r4;
|
|
sub.s32 %r30, %r7, %r5;
|
|
cvt.s64.s32 %rd25, %r30;
|
|
mov.u64 %rd47, triangleAlphaIDs;
|
|
cvta.global.u64 %rd24, %rd47;
|
|
// inline asm
|
|
call (%rd23), _rt_buffer_get_64, (%rd24, %r8, %r9, %rd25, %rd7, %rd7, %rd7);
|
|
// inline asm
|
|
ld.u32 %r6, [%rd23];
|
|
cvt.s64.s32 %rd31, %r2;
|
|
mov.u64 %rd48, vertex_buffer_uv;
|
|
cvta.global.u64 %rd30, %rd48;
|
|
mov.u32 %r29, 8;
|
|
// inline asm
|
|
call (%rd29), _rt_buffer_get_64, (%rd30, %r8, %r29, %rd31, %rd7, %rd7, %rd7);
|
|
// inline asm
|
|
ld.v2.f32 {%f12, %f13}, [%rd29];
|
|
cvt.s64.s32 %rd37, %r3;
|
|
// inline asm
|
|
call (%rd35), _rt_buffer_get_64, (%rd30, %r8, %r29, %rd37, %rd7, %rd7, %rd7);
|
|
// inline asm
|
|
ld.v2.f32 {%f16, %f17}, [%rd35];
|
|
// inline asm
|
|
call (%rd41), _rt_buffer_get_64, (%rd30, %r8, %r29, %rd43, %rd7, %rd7, %rd7);
|
|
// inline asm
|
|
mul.f32 %f20, %f8, %f16;
|
|
mul.f32 %f21, %f8, %f17;
|
|
fma.rn.f32 %f22, %f3, %f12, %f20;
|
|
fma.rn.f32 %f23, %f3, %f13, %f21;
|
|
ld.v2.f32 {%f24, %f25}, [%rd41];
|
|
fma.rn.f32 %f4, %f9, %f24, %f22;
|
|
fma.rn.f32 %f5, %f9, %f25, %f23;
|
|
and.b32 %r31, %r6, -16384;
|
|
setp.eq.s32 %p4, %r31, 16384;
|
|
@%p4 bra BB0_8;
|
|
bra.uni BB0_5;
|
|
|
|
BB0_8:
|
|
add.s32 %r52, %r6, -16384;
|
|
cvt.u64.u32 %rd73, %r52;
|
|
mov.u64 %rd77, translucentTexIDs;
|
|
cvta.global.u64 %rd72, %rd77;
|
|
// inline asm
|
|
call (%rd71), _rt_buffer_get_64, (%rd72, %r8, %r9, %rd73, %rd7, %rd7, %rd7);
|
|
// inline asm
|
|
ld.u32 %r50, [%rd71];
|
|
mov.u32 %r51, 2;
|
|
mov.f32 %f45, 0f00000000;
|
|
// inline asm
|
|
call (%f38, %f39, %f40, %f41), _rt_texture_get_f_id, (%r50, %r51, %f4, %f5, %f45, %f45);
|
|
// inline asm
|
|
cvt.sat.f32.f32 %f46, %f38;
|
|
mul.f32 %f47, %f46, 0f437F0000;
|
|
cvt.rzi.s32.f32 %r53, %f47;
|
|
cvt.sat.f32.f32 %f48, %f39;
|
|
mul.f32 %f49, %f48, 0f437F0000;
|
|
cvt.rzi.s32.f32 %r54, %f49;
|
|
cvt.sat.f32.f32 %f50, %f40;
|
|
mul.f32 %f51, %f50, 0f437F0000;
|
|
cvt.rzi.s32.f32 %r55, %f51;
|
|
shl.b32 %r56, %r54, 8;
|
|
or.b32 %r57, %r56, %r53;
|
|
shl.b32 %r58, %r55, 16;
|
|
or.b32 %r59, %r57, %r58;
|
|
not.b32 %r60, %r59;
|
|
cvt.rn.f32.s32 %f52, %r60;
|
|
st.global.f32 [triAlpha], %f52;
|
|
bra.uni BB0_7;
|
|
|
|
BB0_2:
|
|
mov.u32 %r19, 0;
|
|
st.global.u32 [normal+8], %r19;
|
|
st.global.u32 [normal+4], %r19;
|
|
st.global.u32 [normal], %r19;
|
|
st.global.u32 [triAlpha], %r19;
|
|
bra.uni BB0_7;
|
|
|
|
BB0_5:
|
|
and.b32 %r42, %r6, 65535;
|
|
abs.f32 %f28, %f4;
|
|
cvt.rmi.f32.f32 %f29, %f28;
|
|
sub.f32 %f30, %f28, %f29;
|
|
abs.f32 %f31, %f5;
|
|
cvt.rmi.f32.f32 %f32, %f31;
|
|
sub.f32 %f33, %f31, %f32;
|
|
cvt.u64.u32 %rd61, %r42;
|
|
mov.u64 %rd70, alphaTextures;
|
|
cvta.global.u64 %rd50, %rd70;
|
|
// inline asm
|
|
call (%rd49), _rt_buffer_get_64, (%rd50, %r8, %r9, %rd61, %rd7, %rd7, %rd7);
|
|
// inline asm
|
|
ld.u32 %r34, [%rd49];
|
|
mov.u32 %r40, 2;
|
|
// inline asm
|
|
call (%rd55, %rd56, %rd57, %rd58), _rt_buffer_get_id_size_64, (%r34, %r40, %r8);
|
|
// inline asm
|
|
cvt.rn.f32.u64 %f34, %rd55;
|
|
mul.f32 %f35, %f30, %f34;
|
|
cvt.rzi.u32.f32 %r43, %f35;
|
|
cvt.rn.f32.u64 %f36, %rd56;
|
|
mul.f32 %f37, %f33, %f36;
|
|
cvt.rzi.u32.f32 %r44, %f37;
|
|
// inline asm
|
|
call (%rd59), _rt_buffer_get_64, (%rd50, %r8, %r9, %rd61, %rd7, %rd7, %rd7);
|
|
// inline asm
|
|
ld.u32 %r39, [%rd59];
|
|
cvt.u64.u32 %rd66, %r43;
|
|
cvt.u64.u32 %rd67, %r44;
|
|
// inline asm
|
|
call (%rd65), _rt_buffer_get_id_64, (%r39, %r40, %r8, %rd66, %rd67, %rd7, %rd7);
|
|
// inline asm
|
|
ld.u8 %r45, [%rd65];
|
|
shr.u32 %r46, %r6, 16;
|
|
and.b32 %r47, %r45, %r46;
|
|
setp.eq.s32 %p5, %r47, 0;
|
|
selp.f32 %f79, 0f00000000, 0f3F800000, %p5;
|
|
|
|
BB0_6:
|
|
st.global.f32 [triAlpha], %f79;
|
|
cvt.s64.s32 %rd80, %r2;
|
|
mov.u64 %rd96, vertex_buffer;
|
|
cvta.global.u64 %rd79, %rd96;
|
|
mov.u32 %r66, 24;
|
|
// inline asm
|
|
call (%rd78), _rt_buffer_get_64, (%rd79, %r8, %r66, %rd80, %rd7, %rd7, %rd7);
|
|
// inline asm
|
|
ld.f32 %f53, [%rd78+20];
|
|
ld.f32 %f54, [%rd78+16];
|
|
ld.f32 %f55, [%rd78+12];
|
|
cvt.s64.s32 %rd86, %r3;
|
|
// inline asm
|
|
call (%rd84), _rt_buffer_get_64, (%rd79, %r8, %r66, %rd86, %rd7, %rd7, %rd7);
|
|
// inline asm
|
|
ld.f32 %f56, [%rd84+20];
|
|
ld.f32 %f57, [%rd84+16];
|
|
ld.f32 %f58, [%rd84+12];
|
|
cvt.s64.s32 %rd92, %r4;
|
|
// inline asm
|
|
call (%rd90), _rt_buffer_get_64, (%rd79, %r8, %r66, %rd92, %rd7, %rd7, %rd7);
|
|
// inline asm
|
|
mul.f32 %f59, %f8, %f58;
|
|
mul.f32 %f60, %f8, %f57;
|
|
mul.f32 %f61, %f8, %f56;
|
|
fma.rn.f32 %f62, %f3, %f55, %f59;
|
|
fma.rn.f32 %f63, %f3, %f54, %f60;
|
|
fma.rn.f32 %f64, %f3, %f53, %f61;
|
|
ld.f32 %f65, [%rd90+20];
|
|
ld.f32 %f66, [%rd90+16];
|
|
ld.f32 %f67, [%rd90+12];
|
|
fma.rn.f32 %f68, %f9, %f67, %f62;
|
|
fma.rn.f32 %f69, %f9, %f66, %f63;
|
|
fma.rn.f32 %f70, %f9, %f65, %f64;
|
|
mul.f32 %f71, %f69, %f69;
|
|
fma.rn.f32 %f72, %f68, %f68, %f71;
|
|
fma.rn.f32 %f73, %f70, %f70, %f72;
|
|
sqrt.rn.f32 %f74, %f73;
|
|
rcp.rn.f32 %f75, %f74;
|
|
mul.f32 %f76, %f75, %f68;
|
|
mul.f32 %f77, %f75, %f69;
|
|
mul.f32 %f78, %f75, %f70;
|
|
st.global.f32 [normal], %f76;
|
|
st.global.f32 [normal+4], %f77;
|
|
st.global.f32 [normal+8], %f78;
|
|
|
|
BB0_7:
|
|
ret;
|
|
}
|
|
|
|
|