302 lines
13 KiB
Plaintext
302 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 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<4>;
|
|
.reg .f32 %f<88>;
|
|
.reg .b32 %r<45>;
|
|
.reg .b64 %rd<95>;
|
|
|
|
|
|
ld.param.u32 %r5, [_Z11interpolatei_param_0];
|
|
// inline asm
|
|
call (%f15, %f16), _rt_get_triangle_barycentrics, ();
|
|
// inline asm
|
|
cvt.s64.s32 %rd4, %r5;
|
|
mov.u64 %rd8, index_buffer;
|
|
cvta.global.u64 %rd3, %rd8;
|
|
mov.u32 %r6, 1;
|
|
mov.u32 %r7, 12;
|
|
mov.u64 %rd7, 0;
|
|
// inline asm
|
|
call (%rd2), _rt_buffer_get_64, (%rd3, %r6, %r7, %rd4, %rd7, %rd7, %rd7);
|
|
// inline asm
|
|
ld.u32 %r1, [%rd2];
|
|
ld.u32 %r2, [%rd2+4];
|
|
ld.u32 %r3, [%rd2+8];
|
|
ld.global.u32 %r4, [firstAlphaTriangle];
|
|
setp.gt.s32 %p1, %r4, %r5;
|
|
mov.f32 %f85, 0f3F800000;
|
|
sub.f32 %f18, %f85, %f15;
|
|
sub.f32 %f3, %f18, %f16;
|
|
@%p1 bra BB0_2;
|
|
|
|
cvt.s64.s32 %rd29, %r3;
|
|
sub.s32 %r26, %r5, %r4;
|
|
cvt.s64.s32 %rd11, %r26;
|
|
mov.u64 %rd54, triangleAlphaIDs;
|
|
cvta.global.u64 %rd10, %rd54;
|
|
mov.u32 %r22, 4;
|
|
// inline asm
|
|
call (%rd9), _rt_buffer_get_64, (%rd10, %r6, %r22, %rd11, %rd7, %rd7, %rd7);
|
|
// inline asm
|
|
ld.u32 %rd55, [%rd9];
|
|
cvt.u32.u64 %r27, %rd55;
|
|
shr.u32 %r28, %r27, 16;
|
|
cvt.s64.s32 %rd17, %r1;
|
|
mov.u64 %rd56, vertex_buffer_uv;
|
|
cvta.global.u64 %rd16, %rd56;
|
|
mov.u32 %r15, 8;
|
|
// inline asm
|
|
call (%rd15), _rt_buffer_get_64, (%rd16, %r6, %r15, %rd17, %rd7, %rd7, %rd7);
|
|
// inline asm
|
|
ld.v2.f32 {%f19, %f20}, [%rd15];
|
|
cvt.s64.s32 %rd23, %r2;
|
|
// inline asm
|
|
call (%rd21), _rt_buffer_get_64, (%rd16, %r6, %r15, %rd23, %rd7, %rd7, %rd7);
|
|
// inline asm
|
|
ld.v2.f32 {%f23, %f24}, [%rd21];
|
|
// inline asm
|
|
call (%rd27), _rt_buffer_get_64, (%rd16, %r6, %r15, %rd29, %rd7, %rd7, %rd7);
|
|
// inline asm
|
|
mul.f32 %f27, %f15, %f23;
|
|
mul.f32 %f28, %f15, %f24;
|
|
fma.rn.f32 %f29, %f3, %f19, %f27;
|
|
fma.rn.f32 %f30, %f3, %f20, %f28;
|
|
ld.v2.f32 {%f31, %f32}, [%rd27];
|
|
fma.rn.f32 %f35, %f16, %f31, %f29;
|
|
fma.rn.f32 %f36, %f16, %f32, %f30;
|
|
abs.f32 %f37, %f35;
|
|
cvt.rmi.f32.f32 %f38, %f37;
|
|
sub.f32 %f39, %f37, %f38;
|
|
abs.f32 %f40, %f36;
|
|
cvt.rmi.f32.f32 %f41, %f40;
|
|
sub.f32 %f42, %f40, %f41;
|
|
and.b64 %rd35, %rd55, 65535;
|
|
mov.u64 %rd57, alphaTextures;
|
|
cvta.global.u64 %rd34, %rd57;
|
|
// inline asm
|
|
call (%rd33), _rt_buffer_get_64, (%rd34, %r6, %r22, %rd35, %rd7, %rd7, %rd7);
|
|
// inline asm
|
|
ld.u32 %r18, [%rd33];
|
|
mov.u32 %r24, 2;
|
|
// inline asm
|
|
call (%rd39, %rd40, %rd41, %rd42), _rt_buffer_get_id_size_64, (%r18, %r24, %r6);
|
|
// inline asm
|
|
cvt.rn.f32.u64 %f43, %rd39;
|
|
mul.f32 %f44, %f39, %f43;
|
|
cvt.rzi.u32.f32 %r29, %f44;
|
|
cvt.rn.f32.u64 %f45, %rd40;
|
|
mul.f32 %f46, %f42, %f45;
|
|
cvt.rzi.u32.f32 %r30, %f46;
|
|
// inline asm
|
|
call (%rd43), _rt_buffer_get_64, (%rd34, %r6, %r22, %rd35, %rd7, %rd7, %rd7);
|
|
// inline asm
|
|
ld.u32 %r23, [%rd43];
|
|
cvt.u64.u32 %rd50, %r29;
|
|
cvt.u64.u32 %rd51, %r30;
|
|
// inline asm
|
|
call (%rd49), _rt_buffer_get_id_64, (%r23, %r24, %r6, %rd50, %rd51, %rd7, %rd7);
|
|
// inline asm
|
|
ld.u8 %r31, [%rd49];
|
|
and.b32 %r32, %r31, %r28;
|
|
setp.eq.s32 %p2, %r32, 0;
|
|
selp.f32 %f85, 0f00000000, 0f3F800000, %p2;
|
|
|
|
BB0_2:
|
|
st.global.f32 [triAlpha], %f85;
|
|
cvt.s64.s32 %rd78, %r1;
|
|
mov.u64 %rd94, vertex_buffer;
|
|
cvta.global.u64 %rd59, %rd94;
|
|
mov.u32 %r44, 32;
|
|
// inline asm
|
|
call (%rd58), _rt_buffer_get_64, (%rd59, %r6, %r44, %rd78, %rd7, %rd7, %rd7);
|
|
// inline asm
|
|
ld.f32 %f47, [%rd58+20];
|
|
ld.f32 %f48, [%rd58+16];
|
|
ld.f32 %f49, [%rd58+12];
|
|
cvt.s64.s32 %rd84, %r2;
|
|
// inline asm
|
|
call (%rd64), _rt_buffer_get_64, (%rd59, %r6, %r44, %rd84, %rd7, %rd7, %rd7);
|
|
// inline asm
|
|
ld.f32 %f50, [%rd64+20];
|
|
ld.f32 %f51, [%rd64+16];
|
|
ld.f32 %f52, [%rd64+12];
|
|
cvt.s64.s32 %rd90, %r3;
|
|
// inline asm
|
|
call (%rd70), _rt_buffer_get_64, (%rd59, %r6, %r44, %rd90, %rd7, %rd7, %rd7);
|
|
// inline asm
|
|
mul.f32 %f53, %f15, %f52;
|
|
mul.f32 %f54, %f15, %f51;
|
|
mul.f32 %f55, %f15, %f50;
|
|
fma.rn.f32 %f56, %f3, %f49, %f53;
|
|
fma.rn.f32 %f57, %f3, %f48, %f54;
|
|
fma.rn.f32 %f58, %f3, %f47, %f55;
|
|
ld.f32 %f59, [%rd70+20];
|
|
ld.f32 %f60, [%rd70+16];
|
|
ld.f32 %f61, [%rd70+12];
|
|
fma.rn.f32 %f62, %f16, %f61, %f56;
|
|
fma.rn.f32 %f63, %f16, %f60, %f57;
|
|
fma.rn.f32 %f64, %f16, %f59, %f58;
|
|
mul.f32 %f65, %f63, %f63;
|
|
fma.rn.f32 %f66, %f62, %f62, %f65;
|
|
fma.rn.f32 %f67, %f64, %f64, %f66;
|
|
sqrt.rn.f32 %f68, %f67;
|
|
rcp.rn.f32 %f69, %f68;
|
|
mul.f32 %f70, %f69, %f62;
|
|
mul.f32 %f71, %f69, %f63;
|
|
mul.f32 %f72, %f69, %f64;
|
|
st.global.f32 [normal], %f70;
|
|
st.global.f32 [normal+4], %f71;
|
|
st.global.f32 [normal+8], %f72;
|
|
// inline asm
|
|
call (%rd76), _rt_buffer_get_64, (%rd59, %r6, %r44, %rd78, %rd7, %rd7, %rd7);
|
|
// inline asm
|
|
ld.v2.f32 {%f86, %f74}, [%rd76+24];
|
|
// inline asm
|
|
call (%rd82), _rt_buffer_get_64, (%rd59, %r6, %r44, %rd84, %rd7, %rd7, %rd7);
|
|
// inline asm
|
|
ld.v2.f32 {%f75, %f76}, [%rd82+24];
|
|
// inline asm
|
|
call (%rd88), _rt_buffer_get_64, (%rd59, %r6, %r44, %rd90, %rd7, %rd7, %rd7);
|
|
// inline asm
|
|
setp.lt.f32 %p3, %f74, 0f00000000;
|
|
@%p3 bra BB0_4;
|
|
bra.uni BB0_3;
|
|
|
|
BB0_4:
|
|
add.f32 %f87, %f74, 0f3F800000;
|
|
bra.uni BB0_5;
|
|
|
|
BB0_3:
|
|
ld.v2.f32 {%f77, %f78}, [%rd88+24];
|
|
mul.f32 %f81, %f15, %f75;
|
|
mul.f32 %f82, %f15, %f76;
|
|
fma.rn.f32 %f83, %f3, %f86, %f81;
|
|
fma.rn.f32 %f84, %f3, %f74, %f82;
|
|
fma.rn.f32 %f86, %f16, %f77, %f83;
|
|
fma.rn.f32 %f87, %f16, %f78, %f84;
|
|
|
|
BB0_5:
|
|
st.global.v2.f32 [texCoords], {%f86, %f87};
|
|
ret;
|
|
}
|
|
|
|
|