339 lines
14 KiB
Plaintext
339 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 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 bounceFlag;
|
|
.global .align 1 .b8 areaLights[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_typeinfo10bounceFlagE[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_typename10bounceFlagE[4] = {105, 110, 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_typeenum10bounceFlagE = 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_semantic10bounceFlagE[1];
|
|
.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_annotation10bounceFlagE[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<7>;
|
|
.reg .f32 %f<97>;
|
|
.reg .b32 %r<51>;
|
|
.reg .b64 %rd<107>;
|
|
|
|
|
|
ld.param.u32 %r5, [_Z11interpolatei_param_0];
|
|
// inline asm
|
|
call (%f17, %f18), _rt_get_triangle_barycentrics, ();
|
|
// inline asm
|
|
cvt.s64.s32 %rd7, %r5;
|
|
mov.u64 %rd11, index_buffer;
|
|
cvta.global.u64 %rd6, %rd11;
|
|
mov.u32 %r6, 1;
|
|
mov.u32 %r7, 12;
|
|
mov.u64 %rd10, 0;
|
|
// inline asm
|
|
call (%rd5), _rt_buffer_get_64, (%rd6, %r6, %r7, %rd7, %rd10, %rd10, %rd10);
|
|
// inline asm
|
|
ld.u32 %r1, [%rd5];
|
|
ld.u32 %r2, [%rd5+4];
|
|
ld.u32 %r3, [%rd5+8];
|
|
ld.global.u32 %r4, [firstAlphaTriangle];
|
|
setp.gt.s32 %p1, %r4, %r5;
|
|
mov.f32 %f94, 0f3F800000;
|
|
@%p1 bra BB0_2;
|
|
|
|
cvt.s64.s32 %rd20, %r1;
|
|
sub.s32 %r26, %r5, %r4;
|
|
cvt.s64.s32 %rd14, %r26;
|
|
mov.u64 %rd57, triangleAlphaIDs;
|
|
cvta.global.u64 %rd13, %rd57;
|
|
mov.u32 %r22, 4;
|
|
// inline asm
|
|
call (%rd12), _rt_buffer_get_64, (%rd13, %r6, %r22, %rd14, %rd10, %rd10, %rd10);
|
|
// inline asm
|
|
ld.u32 %rd58, [%rd12];
|
|
cvt.u32.u64 %r27, %rd58;
|
|
shr.u32 %r28, %r27, 16;
|
|
mov.u64 %rd59, vertex_buffer_uv;
|
|
cvta.global.u64 %rd19, %rd59;
|
|
mov.u32 %r15, 8;
|
|
// inline asm
|
|
call (%rd18), _rt_buffer_get_64, (%rd19, %r6, %r15, %rd20, %rd10, %rd10, %rd10);
|
|
// inline asm
|
|
ld.v2.f32 {%f20, %f21}, [%rd18];
|
|
cvt.s64.s32 %rd26, %r2;
|
|
// inline asm
|
|
call (%rd24), _rt_buffer_get_64, (%rd19, %r6, %r15, %rd26, %rd10, %rd10, %rd10);
|
|
// inline asm
|
|
ld.v2.f32 {%f24, %f25}, [%rd24];
|
|
cvt.s64.s32 %rd32, %r3;
|
|
// inline asm
|
|
call (%rd30), _rt_buffer_get_64, (%rd19, %r6, %r15, %rd32, %rd10, %rd10, %rd10);
|
|
// inline asm
|
|
mov.f32 %f28, 0f3F800000;
|
|
sub.f32 %f29, %f28, %f17;
|
|
sub.f32 %f30, %f29, %f18;
|
|
mul.f32 %f31, %f17, %f24;
|
|
mul.f32 %f32, %f17, %f25;
|
|
fma.rn.f32 %f33, %f30, %f20, %f31;
|
|
fma.rn.f32 %f34, %f30, %f21, %f32;
|
|
ld.v2.f32 {%f35, %f36}, [%rd30];
|
|
fma.rn.f32 %f39, %f18, %f35, %f33;
|
|
fma.rn.f32 %f40, %f18, %f36, %f34;
|
|
abs.f32 %f41, %f39;
|
|
cvt.rmi.f32.f32 %f42, %f41;
|
|
sub.f32 %f43, %f41, %f42;
|
|
abs.f32 %f44, %f40;
|
|
cvt.rmi.f32.f32 %f45, %f44;
|
|
sub.f32 %f46, %f44, %f45;
|
|
and.b64 %rd38, %rd58, 65535;
|
|
mov.u64 %rd60, alphaTextures;
|
|
cvta.global.u64 %rd37, %rd60;
|
|
// inline asm
|
|
call (%rd36), _rt_buffer_get_64, (%rd37, %r6, %r22, %rd38, %rd10, %rd10, %rd10);
|
|
// inline asm
|
|
ld.u32 %r18, [%rd36];
|
|
mov.u32 %r24, 2;
|
|
// inline asm
|
|
call (%rd42, %rd43, %rd44, %rd45), _rt_buffer_get_id_size_64, (%r18, %r24, %r6);
|
|
// inline asm
|
|
cvt.rn.f32.u64 %f47, %rd42;
|
|
mul.f32 %f48, %f43, %f47;
|
|
cvt.rzi.u32.f32 %r29, %f48;
|
|
cvt.rn.f32.u64 %f49, %rd43;
|
|
mul.f32 %f50, %f46, %f49;
|
|
cvt.rzi.u32.f32 %r30, %f50;
|
|
// inline asm
|
|
call (%rd46), _rt_buffer_get_64, (%rd37, %r6, %r22, %rd38, %rd10, %rd10, %rd10);
|
|
// inline asm
|
|
ld.u32 %r23, [%rd46];
|
|
cvt.u64.u32 %rd53, %r29;
|
|
cvt.u64.u32 %rd54, %r30;
|
|
// inline asm
|
|
call (%rd52), _rt_buffer_get_id_64, (%r23, %r24, %r6, %rd53, %rd54, %rd10, %rd10);
|
|
// inline asm
|
|
ld.u8 %r31, [%rd52];
|
|
and.b32 %r32, %r31, %r28;
|
|
setp.eq.s32 %p2, %r32, 0;
|
|
selp.f32 %f94, 0f00000000, 0f3F800000, %p2;
|
|
|
|
BB0_2:
|
|
st.global.f32 [triAlpha], %f94;
|
|
cvt.s64.s32 %rd63, %r1;
|
|
mov.u64 %rd67, vertex_buffer;
|
|
cvta.global.u64 %rd62, %rd67;
|
|
mov.u32 %r34, 32;
|
|
// inline asm
|
|
call (%rd61), _rt_buffer_get_64, (%rd62, %r6, %r34, %rd63, %rd10, %rd10, %rd10);
|
|
// inline asm
|
|
ld.v2.f32 {%f95, %f52}, [%rd61+24];
|
|
setp.geu.f32 %p3, %f95, 0f00000000;
|
|
@%p3 bra BB0_6;
|
|
|
|
ld.global.u32 %r35, [bounceFlag];
|
|
cvt.rn.f32.s32 %f53, %r35;
|
|
setp.neu.f32 %p4, %f53, 0f00000000;
|
|
@%p4 bra BB0_6;
|
|
|
|
div.rn.f32 %f54, %f95, 0f41200000;
|
|
cvt.rzi.s32.f32 %r38, %f54;
|
|
neg.s32 %r39, %r38;
|
|
cvt.s64.s32 %rd70, %r39;
|
|
mov.u64 %rd74, areaLights;
|
|
cvta.global.u64 %rd69, %rd74;
|
|
mov.u32 %r37, 16;
|
|
// inline asm
|
|
call (%rd68), _rt_buffer_get_64, (%rd69, %r6, %r37, %rd70, %rd10, %rd10, %rd10);
|
|
// inline asm
|
|
ld.f32 %f55, [%rd68+8];
|
|
setp.geu.f32 %p5, %f55, 0f00000000;
|
|
@%p5 bra BB0_6;
|
|
|
|
mov.u32 %r40, 0;
|
|
st.global.u32 [triAlpha], %r40;
|
|
|
|
BB0_6:
|
|
cvt.s64.s32 %rd77, %r2;
|
|
// inline asm
|
|
call (%rd75), _rt_buffer_get_64, (%rd62, %r6, %r34, %rd77, %rd10, %rd10, %rd10);
|
|
// inline asm
|
|
ld.v2.f32 {%f56, %f57}, [%rd75+24];
|
|
cvt.s64.s32 %rd83, %r3;
|
|
// inline asm
|
|
call (%rd81), _rt_buffer_get_64, (%rd62, %r6, %r34, %rd83, %rd10, %rd10, %rd10);
|
|
// inline asm
|
|
mov.f32 %f58, 0f3F800000;
|
|
sub.f32 %f59, %f58, %f17;
|
|
sub.f32 %f11, %f59, %f18;
|
|
setp.lt.f32 %p6, %f52, 0f00000000;
|
|
@%p6 bra BB0_8;
|
|
bra.uni BB0_7;
|
|
|
|
BB0_8:
|
|
add.f32 %f96, %f52, 0f3F800000;
|
|
bra.uni BB0_9;
|
|
|
|
BB0_7:
|
|
ld.v2.f32 {%f60, %f61}, [%rd81+24];
|
|
mul.f32 %f64, %f17, %f56;
|
|
mul.f32 %f65, %f17, %f57;
|
|
fma.rn.f32 %f66, %f11, %f95, %f64;
|
|
fma.rn.f32 %f67, %f11, %f52, %f65;
|
|
fma.rn.f32 %f95, %f18, %f60, %f66;
|
|
fma.rn.f32 %f96, %f18, %f61, %f67;
|
|
|
|
BB0_9:
|
|
st.global.v2.f32 [texCoords], {%f95, %f96};
|
|
// inline asm
|
|
call (%rd88), _rt_buffer_get_64, (%rd62, %r6, %r34, %rd63, %rd10, %rd10, %rd10);
|
|
// inline asm
|
|
ld.f32 %f68, [%rd88+20];
|
|
ld.f32 %f69, [%rd88+16];
|
|
ld.f32 %f70, [%rd88+12];
|
|
// inline asm
|
|
call (%rd94), _rt_buffer_get_64, (%rd62, %r6, %r34, %rd77, %rd10, %rd10, %rd10);
|
|
// inline asm
|
|
ld.f32 %f71, [%rd94+20];
|
|
ld.f32 %f72, [%rd94+16];
|
|
ld.f32 %f73, [%rd94+12];
|
|
// inline asm
|
|
call (%rd100), _rt_buffer_get_64, (%rd62, %r6, %r34, %rd83, %rd10, %rd10, %rd10);
|
|
// inline asm
|
|
mul.f32 %f74, %f17, %f73;
|
|
mul.f32 %f75, %f17, %f72;
|
|
mul.f32 %f76, %f17, %f71;
|
|
fma.rn.f32 %f77, %f11, %f70, %f74;
|
|
fma.rn.f32 %f78, %f11, %f69, %f75;
|
|
fma.rn.f32 %f79, %f11, %f68, %f76;
|
|
ld.f32 %f80, [%rd100+20];
|
|
ld.f32 %f81, [%rd100+16];
|
|
ld.f32 %f82, [%rd100+12];
|
|
fma.rn.f32 %f83, %f18, %f82, %f77;
|
|
fma.rn.f32 %f84, %f18, %f81, %f78;
|
|
fma.rn.f32 %f85, %f18, %f80, %f79;
|
|
mul.f32 %f86, %f84, %f84;
|
|
fma.rn.f32 %f87, %f83, %f83, %f86;
|
|
fma.rn.f32 %f88, %f85, %f85, %f87;
|
|
sqrt.rn.f32 %f89, %f88;
|
|
rcp.rn.f32 %f90, %f89;
|
|
mul.f32 %f91, %f90, %f83;
|
|
mul.f32 %f92, %f90, %f84;
|
|
mul.f32 %f93, %f90, %f85;
|
|
st.global.f32 [normal], %f91;
|
|
st.global.f32 [normal+4], %f92;
|
|
st.global.f32 [normal+8], %f93;
|
|
ret;
|
|
}
|
|
|
|
|