266 lines
11 KiB
Plaintext
266 lines
11 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 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 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<3>;
|
||
|
|
.reg .f32 %f<65>;
|
||
|
|
.reg .b32 %r<39>;
|
||
|
|
.reg .b64 %rd<76>;
|
||
|
|
|
||
|
|
|
||
|
|
ld.param.u32 %r5, [_Z11interpolatei_param_0];
|
||
|
|
// inline asm
|
||
|
|
call (%f6, %f7), _rt_get_triangle_barycentrics, ();
|
||
|
|
// inline asm
|
||
|
|
cvt.s64.s32 %rd3, %r5;
|
||
|
|
mov.u64 %rd7, index_buffer;
|
||
|
|
cvta.global.u64 %rd2, %rd7;
|
||
|
|
mov.u32 %r6, 1;
|
||
|
|
mov.u32 %r7, 12;
|
||
|
|
mov.u64 %rd6, 0;
|
||
|
|
// inline asm
|
||
|
|
call (%rd1), _rt_buffer_get_64, (%rd2, %r6, %r7, %rd3, %rd6, %rd6, %rd6);
|
||
|
|
// inline asm
|
||
|
|
ld.u32 %r1, [%rd1];
|
||
|
|
ld.u32 %r2, [%rd1+4];
|
||
|
|
ld.u32 %r3, [%rd1+8];
|
||
|
|
ld.global.u32 %r4, [firstAlphaTriangle];
|
||
|
|
setp.gt.s32 %p1, %r4, %r5;
|
||
|
|
mov.f32 %f64, 0f3F800000;
|
||
|
|
sub.f32 %f9, %f64, %f6;
|
||
|
|
sub.f32 %f3, %f9, %f7;
|
||
|
|
@%p1 bra BB0_2;
|
||
|
|
|
||
|
|
cvt.s64.s32 %rd28, %r3;
|
||
|
|
sub.s32 %r26, %r5, %r4;
|
||
|
|
cvt.s64.s32 %rd10, %r26;
|
||
|
|
mov.u64 %rd53, triangleAlphaIDs;
|
||
|
|
cvta.global.u64 %rd9, %rd53;
|
||
|
|
mov.u32 %r22, 4;
|
||
|
|
// inline asm
|
||
|
|
call (%rd8), _rt_buffer_get_64, (%rd9, %r6, %r22, %rd10, %rd6, %rd6, %rd6);
|
||
|
|
// inline asm
|
||
|
|
ld.u32 %rd54, [%rd8];
|
||
|
|
cvt.u32.u64 %r27, %rd54;
|
||
|
|
shr.u32 %r28, %r27, 16;
|
||
|
|
cvt.s64.s32 %rd16, %r1;
|
||
|
|
mov.u64 %rd55, vertex_buffer_uv;
|
||
|
|
cvta.global.u64 %rd15, %rd55;
|
||
|
|
mov.u32 %r15, 8;
|
||
|
|
// inline asm
|
||
|
|
call (%rd14), _rt_buffer_get_64, (%rd15, %r6, %r15, %rd16, %rd6, %rd6, %rd6);
|
||
|
|
// inline asm
|
||
|
|
ld.v2.f32 {%f10, %f11}, [%rd14];
|
||
|
|
cvt.s64.s32 %rd22, %r2;
|
||
|
|
// inline asm
|
||
|
|
call (%rd20), _rt_buffer_get_64, (%rd15, %r6, %r15, %rd22, %rd6, %rd6, %rd6);
|
||
|
|
// inline asm
|
||
|
|
ld.v2.f32 {%f14, %f15}, [%rd20];
|
||
|
|
// inline asm
|
||
|
|
call (%rd26), _rt_buffer_get_64, (%rd15, %r6, %r15, %rd28, %rd6, %rd6, %rd6);
|
||
|
|
// inline asm
|
||
|
|
mul.f32 %f18, %f6, %f14;
|
||
|
|
mul.f32 %f19, %f6, %f15;
|
||
|
|
fma.rn.f32 %f20, %f3, %f10, %f18;
|
||
|
|
fma.rn.f32 %f21, %f3, %f11, %f19;
|
||
|
|
ld.v2.f32 {%f22, %f23}, [%rd26];
|
||
|
|
fma.rn.f32 %f26, %f7, %f22, %f20;
|
||
|
|
fma.rn.f32 %f27, %f7, %f23, %f21;
|
||
|
|
abs.f32 %f28, %f26;
|
||
|
|
cvt.rmi.f32.f32 %f29, %f28;
|
||
|
|
sub.f32 %f30, %f28, %f29;
|
||
|
|
abs.f32 %f31, %f27;
|
||
|
|
cvt.rmi.f32.f32 %f32, %f31;
|
||
|
|
sub.f32 %f33, %f31, %f32;
|
||
|
|
and.b64 %rd34, %rd54, 65535;
|
||
|
|
mov.u64 %rd56, alphaTextures;
|
||
|
|
cvta.global.u64 %rd33, %rd56;
|
||
|
|
// inline asm
|
||
|
|
call (%rd32), _rt_buffer_get_64, (%rd33, %r6, %r22, %rd34, %rd6, %rd6, %rd6);
|
||
|
|
// inline asm
|
||
|
|
ld.u32 %r18, [%rd32];
|
||
|
|
mov.u32 %r24, 2;
|
||
|
|
// inline asm
|
||
|
|
call (%rd38, %rd39, %rd40, %rd41), _rt_buffer_get_id_size_64, (%r18, %r24, %r6);
|
||
|
|
// inline asm
|
||
|
|
cvt.rn.f32.u64 %f34, %rd38;
|
||
|
|
mul.f32 %f35, %f30, %f34;
|
||
|
|
cvt.rzi.u32.f32 %r29, %f35;
|
||
|
|
cvt.rn.f32.u64 %f36, %rd39;
|
||
|
|
mul.f32 %f37, %f33, %f36;
|
||
|
|
cvt.rzi.u32.f32 %r30, %f37;
|
||
|
|
// inline asm
|
||
|
|
call (%rd42), _rt_buffer_get_64, (%rd33, %r6, %r22, %rd34, %rd6, %rd6, %rd6);
|
||
|
|
// inline asm
|
||
|
|
ld.u32 %r23, [%rd42];
|
||
|
|
cvt.u64.u32 %rd49, %r29;
|
||
|
|
cvt.u64.u32 %rd50, %r30;
|
||
|
|
// inline asm
|
||
|
|
call (%rd48), _rt_buffer_get_id_64, (%r23, %r24, %r6, %rd49, %rd50, %rd6, %rd6);
|
||
|
|
// inline asm
|
||
|
|
ld.u8 %r31, [%rd48];
|
||
|
|
and.b32 %r32, %r31, %r28;
|
||
|
|
setp.eq.s32 %p2, %r32, 0;
|
||
|
|
selp.f32 %f64, 0f00000000, 0f3F800000, %p2;
|
||
|
|
|
||
|
|
BB0_2:
|
||
|
|
st.global.f32 [triAlpha], %f64;
|
||
|
|
cvt.s64.s32 %rd59, %r1;
|
||
|
|
mov.u64 %rd75, vertex_buffer;
|
||
|
|
cvta.global.u64 %rd58, %rd75;
|
||
|
|
mov.u32 %r38, 24;
|
||
|
|
// inline asm
|
||
|
|
call (%rd57), _rt_buffer_get_64, (%rd58, %r6, %r38, %rd59, %rd6, %rd6, %rd6);
|
||
|
|
// inline asm
|
||
|
|
ld.f32 %f38, [%rd57+20];
|
||
|
|
ld.f32 %f39, [%rd57+16];
|
||
|
|
ld.f32 %f40, [%rd57+12];
|
||
|
|
cvt.s64.s32 %rd65, %r2;
|
||
|
|
// inline asm
|
||
|
|
call (%rd63), _rt_buffer_get_64, (%rd58, %r6, %r38, %rd65, %rd6, %rd6, %rd6);
|
||
|
|
// inline asm
|
||
|
|
ld.f32 %f41, [%rd63+20];
|
||
|
|
ld.f32 %f42, [%rd63+16];
|
||
|
|
ld.f32 %f43, [%rd63+12];
|
||
|
|
cvt.s64.s32 %rd71, %r3;
|
||
|
|
// inline asm
|
||
|
|
call (%rd69), _rt_buffer_get_64, (%rd58, %r6, %r38, %rd71, %rd6, %rd6, %rd6);
|
||
|
|
// inline asm
|
||
|
|
mul.f32 %f44, %f6, %f43;
|
||
|
|
mul.f32 %f45, %f6, %f42;
|
||
|
|
mul.f32 %f46, %f6, %f41;
|
||
|
|
fma.rn.f32 %f47, %f3, %f40, %f44;
|
||
|
|
fma.rn.f32 %f48, %f3, %f39, %f45;
|
||
|
|
fma.rn.f32 %f49, %f3, %f38, %f46;
|
||
|
|
ld.f32 %f50, [%rd69+20];
|
||
|
|
ld.f32 %f51, [%rd69+16];
|
||
|
|
ld.f32 %f52, [%rd69+12];
|
||
|
|
fma.rn.f32 %f53, %f7, %f52, %f47;
|
||
|
|
fma.rn.f32 %f54, %f7, %f51, %f48;
|
||
|
|
fma.rn.f32 %f55, %f7, %f50, %f49;
|
||
|
|
mul.f32 %f56, %f54, %f54;
|
||
|
|
fma.rn.f32 %f57, %f53, %f53, %f56;
|
||
|
|
fma.rn.f32 %f58, %f55, %f55, %f57;
|
||
|
|
sqrt.rn.f32 %f59, %f58;
|
||
|
|
rcp.rn.f32 %f60, %f59;
|
||
|
|
mul.f32 %f61, %f60, %f53;
|
||
|
|
mul.f32 %f62, %f60, %f54;
|
||
|
|
mul.f32 %f63, %f60, %f55;
|
||
|
|
st.global.f32 [normal], %f61;
|
||
|
|
st.global.f32 [normal+4], %f62;
|
||
|
|
st.global.f32 [normal+8], %f63;
|
||
|
|
ret;
|
||
|
|
}
|
||
|
|
|
||
|
|
|