341 lines
14 KiB
Plaintext
341 lines
14 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 8 .b8 texCoords[8];
|
|
.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 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_typeinfo9texCoordsE[8] = {82, 97, 121, 0, 8, 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_typename9texCoordsE[7] = {102, 108, 111, 97, 116, 50, 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_typeenum9texCoordsE = 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_semantic9texCoordsE[20] = {97, 116, 116, 114, 105, 98, 117, 116, 101, 32, 116, 101, 120, 67, 111, 111, 114, 100, 115, 0};
|
|
.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_annotation9texCoordsE[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<85>;
|
|
.reg .b32 %r<58>;
|
|
.reg .b64 %rd<110>;
|
|
|
|
|
|
ld.param.u32 %r6, [_Z11interpolatei_param_0];
|
|
cvt.s64.s32 %rd5, %r6;
|
|
mov.u64 %rd9, lodPerTriangle;
|
|
cvta.global.u64 %rd4, %rd9;
|
|
mov.u32 %r7, 1;
|
|
mov.u32 %r8, 4;
|
|
mov.u64 %rd8, 0;
|
|
// inline asm
|
|
call (%rd3), _rt_buffer_get_64, (%rd4, %r7, %r8, %rd5, %rd8, %rd8, %rd8);
|
|
// inline asm
|
|
ld.u32 %r1, [%rd3];
|
|
setp.eq.s32 %p1, %r1, 0;
|
|
@%p1 bra BB0_3;
|
|
|
|
ld.global.v2.u32 {%r11, %r12}, [pixelID];
|
|
cvt.u64.u32 %rd12, %r11;
|
|
cvt.u64.u32 %rd13, %r12;
|
|
mov.u64 %rd16, uvlod;
|
|
cvta.global.u64 %rd11, %rd16;
|
|
mov.u32 %r9, 2;
|
|
// inline asm
|
|
call (%rd10), _rt_buffer_get_64, (%rd11, %r9, %r7, %rd12, %rd13, %rd8, %rd8);
|
|
// inline asm
|
|
ld.u8 %r15, [%rd10];
|
|
shl.b32 %r16, %r7, %r15;
|
|
and.b32 %r17, %r16, %r1;
|
|
setp.ne.s32 %p2, %r17, 0;
|
|
@%p2 bra BB0_3;
|
|
bra.uni BB0_2;
|
|
|
|
BB0_3:
|
|
// inline asm
|
|
call (%f11, %f12), _rt_get_triangle_barycentrics, ();
|
|
// inline asm
|
|
mov.u64 %rd23, index_buffer;
|
|
cvta.global.u64 %rd18, %rd23;
|
|
mov.u32 %r20, 12;
|
|
// inline asm
|
|
call (%rd17), _rt_buffer_get_64, (%rd18, %r7, %r20, %rd5, %rd8, %rd8, %rd8);
|
|
// inline asm
|
|
ld.u32 %r2, [%rd17];
|
|
ld.u32 %r3, [%rd17+4];
|
|
ld.u32 %r4, [%rd17+8];
|
|
ld.global.u32 %r5, [firstAlphaTriangle];
|
|
setp.gt.s32 %p3, %r5, %r6;
|
|
mov.f32 %f84, 0f3F800000;
|
|
sub.f32 %f14, %f84, %f11;
|
|
sub.f32 %f3, %f14, %f12;
|
|
@%p3 bra BB0_5;
|
|
|
|
cvt.s64.s32 %rd44, %r4;
|
|
sub.s32 %r39, %r6, %r5;
|
|
cvt.s64.s32 %rd26, %r39;
|
|
mov.u64 %rd69, triangleAlphaIDs;
|
|
cvta.global.u64 %rd25, %rd69;
|
|
// inline asm
|
|
call (%rd24), _rt_buffer_get_64, (%rd25, %r7, %r8, %rd26, %rd8, %rd8, %rd8);
|
|
// inline asm
|
|
ld.u32 %rd70, [%rd24];
|
|
cvt.u32.u64 %r40, %rd70;
|
|
shr.u32 %r41, %r40, 16;
|
|
cvt.s64.s32 %rd32, %r2;
|
|
mov.u64 %rd71, vertex_buffer_uv;
|
|
cvta.global.u64 %rd31, %rd71;
|
|
mov.u32 %r28, 8;
|
|
// inline asm
|
|
call (%rd30), _rt_buffer_get_64, (%rd31, %r7, %r28, %rd32, %rd8, %rd8, %rd8);
|
|
// inline asm
|
|
ld.v2.f32 {%f15, %f16}, [%rd30];
|
|
cvt.s64.s32 %rd38, %r3;
|
|
// inline asm
|
|
call (%rd36), _rt_buffer_get_64, (%rd31, %r7, %r28, %rd38, %rd8, %rd8, %rd8);
|
|
// inline asm
|
|
ld.v2.f32 {%f19, %f20}, [%rd36];
|
|
// inline asm
|
|
call (%rd42), _rt_buffer_get_64, (%rd31, %r7, %r28, %rd44, %rd8, %rd8, %rd8);
|
|
// inline asm
|
|
mul.f32 %f23, %f11, %f19;
|
|
mul.f32 %f24, %f11, %f20;
|
|
fma.rn.f32 %f25, %f3, %f15, %f23;
|
|
fma.rn.f32 %f26, %f3, %f16, %f24;
|
|
ld.v2.f32 {%f27, %f28}, [%rd42];
|
|
fma.rn.f32 %f31, %f12, %f27, %f25;
|
|
fma.rn.f32 %f32, %f12, %f28, %f26;
|
|
abs.f32 %f33, %f31;
|
|
cvt.rmi.f32.f32 %f34, %f33;
|
|
sub.f32 %f35, %f33, %f34;
|
|
abs.f32 %f36, %f32;
|
|
cvt.rmi.f32.f32 %f37, %f36;
|
|
sub.f32 %f38, %f36, %f37;
|
|
and.b64 %rd50, %rd70, 65535;
|
|
mov.u64 %rd72, alphaTextures;
|
|
cvta.global.u64 %rd49, %rd72;
|
|
// inline asm
|
|
call (%rd48), _rt_buffer_get_64, (%rd49, %r7, %r8, %rd50, %rd8, %rd8, %rd8);
|
|
// inline asm
|
|
ld.u32 %r31, [%rd48];
|
|
mov.u32 %r37, 2;
|
|
// inline asm
|
|
call (%rd54, %rd55, %rd56, %rd57), _rt_buffer_get_id_size_64, (%r31, %r37, %r7);
|
|
// inline asm
|
|
cvt.rn.f32.u64 %f39, %rd54;
|
|
mul.f32 %f40, %f35, %f39;
|
|
cvt.rzi.u32.f32 %r42, %f40;
|
|
cvt.rn.f32.u64 %f41, %rd55;
|
|
mul.f32 %f42, %f38, %f41;
|
|
cvt.rzi.u32.f32 %r43, %f42;
|
|
// inline asm
|
|
call (%rd58), _rt_buffer_get_64, (%rd49, %r7, %r8, %rd50, %rd8, %rd8, %rd8);
|
|
// inline asm
|
|
ld.u32 %r36, [%rd58];
|
|
cvt.u64.u32 %rd65, %r42;
|
|
cvt.u64.u32 %rd66, %r43;
|
|
// inline asm
|
|
call (%rd64), _rt_buffer_get_id_64, (%r36, %r37, %r7, %rd65, %rd66, %rd8, %rd8);
|
|
// inline asm
|
|
ld.u8 %r44, [%rd64];
|
|
and.b32 %r45, %r44, %r41;
|
|
setp.eq.s32 %p4, %r45, 0;
|
|
selp.f32 %f84, 0f00000000, 0f3F800000, %p4;
|
|
|
|
BB0_5:
|
|
st.global.f32 [triAlpha], %f84;
|
|
cvt.s64.s32 %rd93, %r2;
|
|
mov.u64 %rd109, vertex_buffer;
|
|
cvta.global.u64 %rd74, %rd109;
|
|
mov.u32 %r57, 32;
|
|
// inline asm
|
|
call (%rd73), _rt_buffer_get_64, (%rd74, %r7, %r57, %rd93, %rd8, %rd8, %rd8);
|
|
// inline asm
|
|
ld.f32 %f43, [%rd73+20];
|
|
ld.f32 %f44, [%rd73+16];
|
|
ld.f32 %f45, [%rd73+12];
|
|
cvt.s64.s32 %rd99, %r3;
|
|
// inline asm
|
|
call (%rd79), _rt_buffer_get_64, (%rd74, %r7, %r57, %rd99, %rd8, %rd8, %rd8);
|
|
// inline asm
|
|
ld.f32 %f46, [%rd79+20];
|
|
ld.f32 %f47, [%rd79+16];
|
|
ld.f32 %f48, [%rd79+12];
|
|
cvt.s64.s32 %rd105, %r4;
|
|
// inline asm
|
|
call (%rd85), _rt_buffer_get_64, (%rd74, %r7, %r57, %rd105, %rd8, %rd8, %rd8);
|
|
// inline asm
|
|
mul.f32 %f49, %f11, %f48;
|
|
mul.f32 %f50, %f11, %f47;
|
|
mul.f32 %f51, %f11, %f46;
|
|
fma.rn.f32 %f52, %f3, %f45, %f49;
|
|
fma.rn.f32 %f53, %f3, %f44, %f50;
|
|
fma.rn.f32 %f54, %f3, %f43, %f51;
|
|
ld.f32 %f55, [%rd85+20];
|
|
ld.f32 %f56, [%rd85+16];
|
|
ld.f32 %f57, [%rd85+12];
|
|
fma.rn.f32 %f58, %f12, %f57, %f52;
|
|
fma.rn.f32 %f59, %f12, %f56, %f53;
|
|
fma.rn.f32 %f60, %f12, %f55, %f54;
|
|
mul.f32 %f61, %f59, %f59;
|
|
fma.rn.f32 %f62, %f58, %f58, %f61;
|
|
fma.rn.f32 %f63, %f60, %f60, %f62;
|
|
sqrt.rn.f32 %f64, %f63;
|
|
rcp.rn.f32 %f65, %f64;
|
|
mul.f32 %f66, %f65, %f58;
|
|
mul.f32 %f67, %f65, %f59;
|
|
mul.f32 %f68, %f65, %f60;
|
|
st.global.f32 [normal], %f66;
|
|
st.global.f32 [normal+4], %f67;
|
|
st.global.f32 [normal+8], %f68;
|
|
// inline asm
|
|
call (%rd91), _rt_buffer_get_64, (%rd74, %r7, %r57, %rd93, %rd8, %rd8, %rd8);
|
|
// inline asm
|
|
ld.v2.f32 {%f69, %f70}, [%rd91+24];
|
|
// inline asm
|
|
call (%rd97), _rt_buffer_get_64, (%rd74, %r7, %r57, %rd99, %rd8, %rd8, %rd8);
|
|
// inline asm
|
|
ld.v2.f32 {%f71, %f72}, [%rd97+24];
|
|
// inline asm
|
|
call (%rd103), _rt_buffer_get_64, (%rd74, %r7, %r57, %rd105, %rd8, %rd8, %rd8);
|
|
// inline asm
|
|
setp.lt.f32 %p5, %f70, 0f00000000;
|
|
@%p5 bra BB0_7;
|
|
bra.uni BB0_6;
|
|
|
|
BB0_7:
|
|
add.f32 %f83, %f70, 0f3F800000;
|
|
st.global.v2.f32 [texCoords], {%f69, %f83};
|
|
bra.uni BB0_8;
|
|
|
|
BB0_6:
|
|
ld.v2.f32 {%f73, %f74}, [%rd103+24];
|
|
mul.f32 %f77, %f11, %f71;
|
|
mul.f32 %f78, %f11, %f72;
|
|
fma.rn.f32 %f79, %f3, %f69, %f77;
|
|
fma.rn.f32 %f80, %f3, %f70, %f78;
|
|
fma.rn.f32 %f81, %f12, %f73, %f79;
|
|
fma.rn.f32 %f82, %f12, %f74, %f80;
|
|
st.global.v2.f32 [texCoords], {%f81, %f82};
|
|
bra.uni BB0_8;
|
|
|
|
BB0_2:
|
|
mov.f32 %f10, 0f00000000;
|
|
st.global.v2.f32 [texCoords], {%f10, %f10};
|
|
mov.u32 %r18, 0;
|
|
st.global.u32 [normal+8], %r18;
|
|
st.global.u32 [normal+4], %r18;
|
|
st.global.u32 [normal], %r18;
|
|
st.global.u32 [triAlpha], %r18;
|
|
|
|
BB0_8:
|
|
ret;
|
|
}
|
|
|
|
|