649 lines
22 KiB
Plaintext
649 lines
22 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 _Z9intersecti
|
|
.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 .b8 ray[36];
|
|
.global .align 4 .u32 heightmap;
|
|
.global .align 4 .f32 heightmapMaxMip;
|
|
.global .align 4 .b8 heightmapMin[12];
|
|
.global .align 4 .b8 heightmapMax[12];
|
|
.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_typeinfo3rayE[8] = {82, 97, 121, 0, 36, 0, 0, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo9heightmapE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo15heightmapMaxMipE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo12heightmapMinE[8] = {82, 97, 121, 0, 12, 0, 0, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo12heightmapMaxE[8] = {82, 97, 121, 0, 12, 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 4 .b8 _ZN21rti_internal_typename3rayE[4] = {82, 97, 121, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typename9heightmapE[4] = {105, 110, 116, 0};
|
|
.global .align 8 .b8 _ZN21rti_internal_typename15heightmapMaxMipE[6] = {102, 108, 111, 97, 116, 0};
|
|
.global .align 8 .b8 _ZN21rti_internal_typename12heightmapMinE[7] = {102, 108, 111, 97, 116, 51, 0};
|
|
.global .align 8 .b8 _ZN21rti_internal_typename12heightmapMaxE[7] = {102, 108, 111, 97, 116, 51, 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_typeenum3rayE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum9heightmapE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum15heightmapMaxMipE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum12heightmapMinE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum12heightmapMaxE = 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_semantic3rayE[13] = {114, 116, 67, 117, 114, 114, 101, 110, 116, 82, 97, 121, 0};
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic9heightmapE[1];
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic15heightmapMaxMipE[1];
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic12heightmapMinE[1];
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic12heightmapMaxE[1];
|
|
.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_annotation3rayE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation9heightmapE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation15heightmapMaxMipE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation12heightmapMinE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation12heightmapMaxE[1];
|
|
|
|
.visible .entry _Z9intersecti(
|
|
.param .u32 _Z9intersecti_param_0
|
|
)
|
|
{
|
|
.reg .pred %p<57>;
|
|
.reg .b16 %rs<9>;
|
|
.reg .f32 %f<439>;
|
|
.reg .b32 %r<36>;
|
|
|
|
|
|
ld.global.f32 %f3, [heightmapMin+8];
|
|
ld.global.f32 %f2, [heightmapMin+4];
|
|
ld.global.f32 %f1, [heightmapMin];
|
|
ld.global.f32 %f139, [heightmapMax+8];
|
|
ld.global.f32 %f140, [heightmapMax+4];
|
|
ld.global.f32 %f141, [heightmapMax];
|
|
add.f32 %f142, %f1, %f141;
|
|
add.f32 %f143, %f2, %f140;
|
|
add.f32 %f144, %f3, %f139;
|
|
mul.f32 %f145, %f142, 0f3F000000;
|
|
mul.f32 %f146, %f143, 0f3F000000;
|
|
mul.f32 %f147, %f144, 0f3F000000;
|
|
sub.f32 %f4, %f141, %f1;
|
|
sub.f32 %f5, %f140, %f2;
|
|
sub.f32 %f6, %f139, %f3;
|
|
mul.f32 %f148, %f4, 0f3F000000;
|
|
mul.f32 %f149, %f5, 0f3F000000;
|
|
mul.f32 %f150, %f6, 0f3F000000;
|
|
ld.global.f32 %f7, [ray+12];
|
|
rcp.rn.f32 %f151, %f7;
|
|
ld.global.f32 %f8, [ray+16];
|
|
rcp.rn.f32 %f152, %f8;
|
|
ld.global.f32 %f9, [ray+20];
|
|
rcp.rn.f32 %f153, %f9;
|
|
ld.global.f32 %f10, [ray];
|
|
sub.f32 %f154, %f10, %f145;
|
|
ld.global.f32 %f11, [ray+4];
|
|
sub.f32 %f155, %f11, %f146;
|
|
ld.global.f32 %f12, [ray+8];
|
|
sub.f32 %f156, %f12, %f147;
|
|
mul.f32 %f157, %f151, %f154;
|
|
mul.f32 %f158, %f152, %f155;
|
|
mul.f32 %f159, %f153, %f156;
|
|
abs.f32 %f160, %f151;
|
|
abs.f32 %f161, %f152;
|
|
abs.f32 %f162, %f153;
|
|
mul.f32 %f163, %f148, %f160;
|
|
mul.f32 %f164, %f149, %f161;
|
|
mul.f32 %f165, %f150, %f162;
|
|
neg.f32 %f166, %f157;
|
|
neg.f32 %f167, %f158;
|
|
neg.f32 %f168, %f159;
|
|
sub.f32 %f169, %f166, %f163;
|
|
sub.f32 %f170, %f167, %f164;
|
|
sub.f32 %f171, %f168, %f165;
|
|
sub.f32 %f172, %f163, %f157;
|
|
sub.f32 %f173, %f164, %f158;
|
|
sub.f32 %f174, %f165, %f159;
|
|
max.f32 %f175, %f169, %f170;
|
|
max.f32 %f13, %f175, %f171;
|
|
min.f32 %f176, %f172, %f173;
|
|
min.f32 %f177, %f176, %f174;
|
|
setp.gt.f32 %p4, %f13, %f177;
|
|
setp.lt.f32 %p5, %f177, 0f00000000;
|
|
or.pred %p6, %p4, %p5;
|
|
@%p6 bra BB0_32;
|
|
|
|
add.f32 %f178, %f13, 0f3C23D70A;
|
|
setp.lt.f32 %p7, %f178, 0f00000000;
|
|
selp.f32 %f14, 0f00000000, %f178, %p7;
|
|
fma.rn.f32 %f179, %f14, %f7, %f10;
|
|
fma.rn.f32 %f180, %f14, %f8, %f11;
|
|
fma.rn.f32 %f181, %f14, %f9, %f12;
|
|
sub.f32 %f182, %f179, %f1;
|
|
sub.f32 %f15, %f180, %f2;
|
|
sub.f32 %f183, %f181, %f3;
|
|
div.rn.f32 %f419, %f182, %f4;
|
|
div.rn.f32 %f421, %f183, %f6;
|
|
abs.f32 %f184, %f15;
|
|
sub.f32 %f18, %f11, %f2;
|
|
setp.geu.f32 %p8, %f184, 0f38D1B717;
|
|
@%p8 bra BB0_3;
|
|
|
|
abs.f32 %f185, %f18;
|
|
setp.gt.f32 %p9, %f185, 0f38D1B717;
|
|
@%p9 bra BB0_30;
|
|
bra.uni BB0_3;
|
|
|
|
BB0_30:
|
|
// inline asm
|
|
call (%r29), _rt_potential_intersection, (%f14);
|
|
// inline asm
|
|
setp.eq.s32 %p55, %r29, 0;
|
|
@%p55 bra BB0_32;
|
|
|
|
mov.u32 %r31, 0;
|
|
st.global.u32 [normal+8], %r31;
|
|
mov.u32 %r32, 1065353216;
|
|
st.global.u32 [normal+4], %r32;
|
|
st.global.u32 [normal], %r31;
|
|
// inline asm
|
|
call (%r30), _rt_report_intersection, (%r31);
|
|
// inline asm
|
|
bra.uni BB0_32;
|
|
|
|
BB0_3:
|
|
sub.f32 %f187, %f10, %f1;
|
|
div.rn.f32 %f19, %f187, %f4;
|
|
div.rn.f32 %f20, %f18, %f5;
|
|
sub.f32 %f188, %f12, %f3;
|
|
div.rn.f32 %f21, %f188, %f6;
|
|
add.f32 %f189, %f419, %f419;
|
|
cvt.rmi.f32.f32 %f190, %f189;
|
|
mul.f32 %f422, %f190, 0f3F000000;
|
|
add.f32 %f191, %f421, %f421;
|
|
cvt.rmi.f32.f32 %f192, %f191;
|
|
mul.f32 %f423, %f192, 0f3F000000;
|
|
div.rn.f32 %f193, %f7, %f4;
|
|
div.rn.f32 %f194, %f8, %f5;
|
|
mul.f32 %f195, %f194, %f194;
|
|
fma.rn.f32 %f196, %f193, %f193, %f195;
|
|
div.rn.f32 %f197, %f9, %f6;
|
|
fma.rn.f32 %f198, %f197, %f197, %f196;
|
|
sqrt.rn.f32 %f199, %f198;
|
|
rcp.rn.f32 %f200, %f199;
|
|
mul.f32 %f24, %f193, %f200;
|
|
mul.f32 %f25, %f194, %f200;
|
|
mul.f32 %f26, %f197, %f200;
|
|
setp.gt.f32 %p10, %f24, 0f00000000;
|
|
mov.f32 %f437, 0f3F800000;
|
|
mov.f32 %f392, %f437;
|
|
@%p10 bra BB0_5;
|
|
|
|
setp.eq.f32 %p11, %f24, 0f00000000;
|
|
selp.f32 %f392, 0f00000000, 0fBF800000, %p11;
|
|
|
|
BB0_5:
|
|
setp.gt.f32 %p12, %f26, 0f00000000;
|
|
mov.f32 %f393, %f437;
|
|
@%p12 bra BB0_7;
|
|
|
|
setp.eq.f32 %p13, %f26, 0f00000000;
|
|
selp.f32 %f393, 0f00000000, 0fBF800000, %p13;
|
|
|
|
BB0_7:
|
|
ld.global.f32 %f425, [heightmapMaxMip];
|
|
abs.f32 %f207, %f26;
|
|
mov.f32 %f208, 0f38D1B717;
|
|
max.f32 %f32, %f207, %f208;
|
|
add.f32 %f209, %f422, 0f3F000000;
|
|
setp.lt.f32 %p14, %f392, 0f00000000;
|
|
selp.f32 %f210, %f422, %f209, %p14;
|
|
sub.f32 %f211, %f210, %f419;
|
|
abs.f32 %f212, %f211;
|
|
abs.f32 %f213, %f24;
|
|
max.f32 %f33, %f213, %f208;
|
|
div.rn.f32 %f427, %f212, %f33;
|
|
add.f32 %f214, %f423, 0f3F000000;
|
|
setp.lt.f32 %p15, %f393, 0f00000000;
|
|
selp.f32 %f215, %f423, %f214, %p15;
|
|
sub.f32 %f216, %f215, %f421;
|
|
abs.f32 %f217, %f216;
|
|
div.rn.f32 %f428, %f217, %f32;
|
|
rcp.rn.f32 %f218, %f24;
|
|
abs.f32 %f219, %f218;
|
|
mul.f32 %f429, %f219, 0f3F000000;
|
|
rcp.rn.f32 %f220, %f26;
|
|
abs.f32 %f221, %f220;
|
|
mul.f32 %f430, %f221, 0f3F000000;
|
|
mul.f32 %f431, %f392, 0f3F000000;
|
|
mul.f32 %f432, %f393, 0f3F000000;
|
|
rcp.rn.f32 %f40, %f25;
|
|
div.rn.f32 %f420, %f15, %f5;
|
|
mov.f32 %f438, 0f00000000;
|
|
mov.f32 %f424, 0f3F000000;
|
|
mov.f32 %f426, 0f40000000;
|
|
mov.u32 %r33, 0;
|
|
mov.u32 %r35, %r33;
|
|
mov.f32 %f436, %f438;
|
|
|
|
BB0_8:
|
|
ld.global.u32 %r3, [heightmap];
|
|
setp.eq.f32 %p16, %f425, 0f00000000;
|
|
@%p16 bra BB0_12;
|
|
bra.uni BB0_9;
|
|
|
|
BB0_12:
|
|
mov.u32 %r22, 2;
|
|
mov.u32 %r23, 0;
|
|
mov.f32 %f264, 0f00000000;
|
|
// inline asm
|
|
call (%f233, %f234, %f235, %f236), _rt_texture_get_level_id, (%r3, %r22, %f422, %f423, %f264, %r23, %f264);
|
|
// inline asm
|
|
ld.global.u32 %r15, [heightmap];
|
|
add.f32 %f253, %f422, %f424;
|
|
// inline asm
|
|
call (%f241, %f242, %f243, %f244), _rt_texture_get_level_id, (%r15, %r22, %f253, %f423, %f264, %r23, %f264);
|
|
// inline asm
|
|
ld.global.u32 %r18, [heightmap];
|
|
add.f32 %f262, %f423, %f424;
|
|
// inline asm
|
|
call (%f249, %f250, %f251, %f252), _rt_texture_get_level_id, (%r18, %r22, %f253, %f262, %f264, %r23, %f264);
|
|
// inline asm
|
|
ld.global.u32 %r21, [heightmap];
|
|
// inline asm
|
|
call (%f257, %f258, %f259, %f260), _rt_texture_get_level_id, (%r21, %r22, %f422, %f262, %f264, %r23, %f264);
|
|
// inline asm
|
|
sub.f32 %f265, %f253, %f422;
|
|
sub.f32 %f266, %f249, %f233;
|
|
sub.f32 %f267, %f262, %f423;
|
|
sub.f32 %f268, %f233, %f257;
|
|
sub.f32 %f269, %f423, %f262;
|
|
mul.f32 %f270, %f267, %f268;
|
|
mul.f32 %f271, %f269, %f266;
|
|
sub.f32 %f436, %f270, %f271;
|
|
mul.f32 %f67, %f269, %f265;
|
|
sub.f32 %f272, %f422, %f422;
|
|
mul.f32 %f273, %f272, %f267;
|
|
sub.f32 %f437, %f67, %f273;
|
|
mul.f32 %f274, %f272, %f266;
|
|
mul.f32 %f275, %f265, %f268;
|
|
sub.f32 %f438, %f274, %f275;
|
|
mul.f32 %f276, %f24, %f436;
|
|
fma.rn.f32 %f277, %f25, %f437, %f276;
|
|
fma.rn.f32 %f278, %f26, %f438, %f277;
|
|
rcp.rn.f32 %f279, %f278;
|
|
sub.f32 %f70, %f233, %f20;
|
|
sub.f32 %f280, %f422, %f19;
|
|
mul.f32 %f281, %f280, %f279;
|
|
mul.f32 %f282, %f70, %f279;
|
|
sub.f32 %f283, %f423, %f21;
|
|
mul.f32 %f284, %f283, %f279;
|
|
mul.f32 %f285, %f25, %f284;
|
|
mul.f32 %f286, %f26, %f282;
|
|
sub.f32 %f287, %f285, %f286;
|
|
mul.f32 %f288, %f26, %f281;
|
|
mul.f32 %f289, %f24, %f284;
|
|
sub.f32 %f290, %f288, %f289;
|
|
mul.f32 %f291, %f24, %f282;
|
|
mul.f32 %f292, %f25, %f281;
|
|
sub.f32 %f293, %f291, %f292;
|
|
mul.f32 %f294, %f268, %f290;
|
|
fma.rn.f32 %f295, %f272, %f287, %f294;
|
|
fma.rn.f32 %f296, %f269, %f293, %f295;
|
|
mul.f32 %f297, %f266, %f290;
|
|
fma.rn.f32 %f298, %f265, %f287, %f297;
|
|
fma.rn.f32 %f299, %f267, %f293, %f298;
|
|
mul.f32 %f300, %f281, %f436;
|
|
fma.rn.f32 %f301, %f437, %f282, %f300;
|
|
fma.rn.f32 %f71, %f284, %f438, %f301;
|
|
setp.gt.f32 %p22, %f71, 0f3456BF95;
|
|
setp.ge.f32 %p23, %f296, 0f00000000;
|
|
and.pred %p24, %p22, %p23;
|
|
setp.ge.f32 %p25, %f299, 0f00000000;
|
|
and.pred %p26, %p24, %p25;
|
|
add.f32 %f302, %f296, %f299;
|
|
setp.le.f32 %p27, %f302, 0f3F800000;
|
|
and.pred %p28, %p26, %p27;
|
|
@%p28 bra BB0_26;
|
|
|
|
sub.f32 %f303, %f241, %f233;
|
|
sub.f32 %f305, %f422, %f253;
|
|
sub.f32 %f306, %f233, %f249;
|
|
sub.f32 %f307, %f423, %f423;
|
|
mul.f32 %f308, %f307, %f306;
|
|
mul.f32 %f311, %f269, %f303;
|
|
sub.f32 %f436, %f308, %f311;
|
|
mul.f32 %f312, %f307, %f305;
|
|
sub.f32 %f437, %f67, %f312;
|
|
mul.f32 %f313, %f305, %f303;
|
|
mul.f32 %f315, %f265, %f306;
|
|
sub.f32 %f438, %f313, %f315;
|
|
mul.f32 %f316, %f24, %f436;
|
|
fma.rn.f32 %f317, %f25, %f437, %f316;
|
|
fma.rn.f32 %f318, %f26, %f438, %f317;
|
|
rcp.rn.f32 %f319, %f318;
|
|
mul.f32 %f321, %f280, %f319;
|
|
mul.f32 %f322, %f70, %f319;
|
|
mul.f32 %f324, %f283, %f319;
|
|
mul.f32 %f325, %f25, %f324;
|
|
mul.f32 %f326, %f26, %f322;
|
|
sub.f32 %f327, %f325, %f326;
|
|
mul.f32 %f328, %f26, %f321;
|
|
mul.f32 %f329, %f24, %f324;
|
|
sub.f32 %f330, %f328, %f329;
|
|
mul.f32 %f331, %f24, %f322;
|
|
mul.f32 %f332, %f25, %f321;
|
|
sub.f32 %f333, %f331, %f332;
|
|
mul.f32 %f334, %f306, %f330;
|
|
fma.rn.f32 %f335, %f305, %f327, %f334;
|
|
fma.rn.f32 %f336, %f269, %f333, %f335;
|
|
mul.f32 %f337, %f303, %f330;
|
|
fma.rn.f32 %f338, %f265, %f327, %f337;
|
|
fma.rn.f32 %f339, %f307, %f333, %f338;
|
|
mul.f32 %f340, %f437, %f322;
|
|
fma.rn.f32 %f341, %f436, %f321, %f340;
|
|
fma.rn.f32 %f75, %f438, %f324, %f341;
|
|
setp.gt.f32 %p29, %f75, 0f3456BF95;
|
|
setp.ge.f32 %p30, %f336, 0f00000000;
|
|
and.pred %p31, %p29, %p30;
|
|
setp.ge.f32 %p32, %f339, 0f00000000;
|
|
and.pred %p33, %p31, %p32;
|
|
add.f32 %f342, %f336, %f339;
|
|
setp.le.f32 %p34, %f342, 0f3F800000;
|
|
and.pred %p35, %p33, %p34;
|
|
selp.u16 %rs7, 1, 0, %p35;
|
|
@%p35 bra BB0_25;
|
|
bra.uni BB0_14;
|
|
|
|
BB0_9:
|
|
setp.lt.f32 %p17, %f25, 0f00000000;
|
|
mov.u32 %r10, 2;
|
|
mov.f32 %f228, 0f00000000;
|
|
mov.u32 %r11, 0;
|
|
// inline asm
|
|
call (%f222, %f223, %f224, %f225), _rt_texture_get_level_id, (%r3, %r10, %f422, %f423, %f228, %r11, %f425);
|
|
// inline asm
|
|
setp.gt.f32 %p56, %f222, %f420;
|
|
setp.leu.f32 %p18, %f222, %f420;
|
|
and.pred %p19, %p17, %p18;
|
|
@!%p19 bra BB0_11;
|
|
bra.uni BB0_10;
|
|
|
|
BB0_10:
|
|
sub.f32 %f230, %f420, %f222;
|
|
mul.f32 %f231, %f40, %f230;
|
|
neg.f32 %f232, %f231;
|
|
setp.gt.f32 %p20, %f427, %f232;
|
|
setp.gt.f32 %p21, %f428, %f232;
|
|
and.pred %p56, %p20, %p21;
|
|
|
|
BB0_11:
|
|
selp.u16 %rs7, 1, 0, %p56;
|
|
|
|
BB0_14:
|
|
setp.eq.s16 %p36, %rs7, 0;
|
|
@%p36 bra BB0_16;
|
|
bra.uni BB0_15;
|
|
|
|
BB0_16:
|
|
setp.eq.s32 %p39, %r35, 4;
|
|
@%p39 bra BB0_22;
|
|
bra.uni BB0_17;
|
|
|
|
BB0_22:
|
|
add.f32 %f425, %f425, 0f3F800000;
|
|
mul.f32 %f426, %f426, 0f3F000000;
|
|
mul.f32 %f355, %f419, %f426;
|
|
cvt.rmi.f32.f32 %f356, %f355;
|
|
div.rn.f32 %f422, %f356, %f426;
|
|
mul.f32 %f357, %f421, %f426;
|
|
cvt.rmi.f32.f32 %f358, %f357;
|
|
div.rn.f32 %f423, %f358, %f426;
|
|
add.f32 %f424, %f424, %f424;
|
|
add.f32 %f359, %f424, %f422;
|
|
selp.f32 %f360, %f422, %f359, %p14;
|
|
sub.f32 %f361, %f360, %f419;
|
|
abs.f32 %f362, %f361;
|
|
div.rn.f32 %f427, %f362, %f33;
|
|
add.f32 %f363, %f424, %f423;
|
|
selp.f32 %f364, %f423, %f363, %p15;
|
|
sub.f32 %f365, %f364, %f421;
|
|
abs.f32 %f366, %f365;
|
|
div.rn.f32 %f428, %f366, %f32;
|
|
add.f32 %f429, %f429, %f429;
|
|
add.f32 %f430, %f430, %f430;
|
|
add.f32 %f431, %f431, %f431;
|
|
add.f32 %f432, %f432, %f432;
|
|
bra.uni BB0_23;
|
|
|
|
BB0_15:
|
|
add.f32 %f425, %f425, 0fBF800000;
|
|
add.f32 %f426, %f426, %f426;
|
|
mul.f32 %f343, %f419, %f426;
|
|
cvt.rmi.f32.f32 %f344, %f343;
|
|
div.rn.f32 %f422, %f344, %f426;
|
|
mul.f32 %f345, %f421, %f426;
|
|
cvt.rmi.f32.f32 %f346, %f345;
|
|
div.rn.f32 %f423, %f346, %f426;
|
|
mul.f32 %f424, %f424, 0f3F000000;
|
|
add.f32 %f347, %f424, %f422;
|
|
selp.f32 %f348, %f422, %f347, %p14;
|
|
sub.f32 %f349, %f348, %f419;
|
|
abs.f32 %f350, %f349;
|
|
div.rn.f32 %f427, %f350, %f33;
|
|
add.f32 %f351, %f424, %f423;
|
|
selp.f32 %f352, %f423, %f351, %p15;
|
|
sub.f32 %f353, %f352, %f421;
|
|
abs.f32 %f354, %f353;
|
|
div.rn.f32 %f428, %f354, %f32;
|
|
mul.f32 %f429, %f429, 0f3F000000;
|
|
mul.f32 %f430, %f430, 0f3F000000;
|
|
mul.f32 %f431, %f431, 0f3F000000;
|
|
mul.f32 %f432, %f432, 0f3F000000;
|
|
|
|
BB0_23:
|
|
mov.u32 %r35, 0;
|
|
bra.uni BB0_24;
|
|
|
|
BB0_17:
|
|
setp.lt.f32 %p40, %f427, %f428;
|
|
@%p40 bra BB0_19;
|
|
bra.uni BB0_18;
|
|
|
|
BB0_19:
|
|
sub.f32 %f418, %f428, %f427;
|
|
add.f32 %f422, %f422, %f431;
|
|
mov.f32 %f428, %f427;
|
|
mov.f32 %f427, %f429;
|
|
bra.uni BB0_20;
|
|
|
|
BB0_18:
|
|
sub.f32 %f427, %f427, %f428;
|
|
add.f32 %f423, %f423, %f432;
|
|
mov.f32 %f418, %f430;
|
|
|
|
BB0_20:
|
|
fma.rn.f32 %f421, %f26, %f428, %f421;
|
|
fma.rn.f32 %f420, %f25, %f428, %f420;
|
|
fma.rn.f32 %f419, %f24, %f428, %f419;
|
|
setp.ge.f32 %p41, %f419, 0f00000000;
|
|
setp.ge.f32 %p42, %f421, 0f00000000;
|
|
and.pred %p43, %p41, %p42;
|
|
setp.lt.f32 %p44, %f419, 0f3F800000;
|
|
and.pred %p45, %p43, %p44;
|
|
setp.lt.f32 %p46, %f421, 0f3F800000;
|
|
and.pred %p47, %p45, %p46;
|
|
setp.lt.f32 %p48, %f420, 0f3F800000;
|
|
and.pred %p49, %p47, %p48;
|
|
mov.u16 %rs8, 0;
|
|
@!%p49 bra BB0_27;
|
|
bra.uni BB0_21;
|
|
|
|
BB0_21:
|
|
add.s32 %r35, %r35, 1;
|
|
mov.f32 %f428, %f418;
|
|
|
|
BB0_24:
|
|
add.s32 %r33, %r33, 1;
|
|
setp.lt.s32 %p52, %r33, 512;
|
|
mov.u16 %rs8, %rs7;
|
|
@%p52 bra BB0_8;
|
|
bra.uni BB0_27;
|
|
|
|
BB0_26:
|
|
fma.rn.f32 %f419, %f24, %f71, %f19;
|
|
fma.rn.f32 %f420, %f25, %f71, %f20;
|
|
fma.rn.f32 %f421, %f26, %f71, %f21;
|
|
mov.u16 %rs8, 1;
|
|
bra.uni BB0_27;
|
|
|
|
BB0_25:
|
|
fma.rn.f32 %f419, %f24, %f75, %f19;
|
|
fma.rn.f32 %f420, %f25, %f75, %f20;
|
|
fma.rn.f32 %f421, %f26, %f75, %f21;
|
|
mov.u16 %rs8, %rs7;
|
|
|
|
BB0_27:
|
|
setp.eq.s16 %p53, %rs8, 0;
|
|
@%p53 bra BB0_32;
|
|
|
|
fma.rn.f32 %f368, %f4, %f419, %f1;
|
|
fma.rn.f32 %f369, %f5, %f420, %f2;
|
|
fma.rn.f32 %f370, %f6, %f421, %f3;
|
|
ld.global.f32 %f371, [ray];
|
|
sub.f32 %f372, %f368, %f371;
|
|
ld.global.f32 %f373, [ray+4];
|
|
sub.f32 %f374, %f369, %f373;
|
|
ld.global.f32 %f375, [ray+8];
|
|
sub.f32 %f376, %f370, %f375;
|
|
mul.f32 %f377, %f374, %f374;
|
|
fma.rn.f32 %f378, %f372, %f372, %f377;
|
|
fma.rn.f32 %f379, %f376, %f376, %f378;
|
|
sqrt.rn.f32 %f367, %f379;
|
|
// inline asm
|
|
call (%r26), _rt_potential_intersection, (%f367);
|
|
// inline asm
|
|
setp.eq.s32 %p54, %r26, 0;
|
|
@%p54 bra BB0_32;
|
|
|
|
mul.f32 %f380, %f437, %f437;
|
|
fma.rn.f32 %f381, %f436, %f436, %f380;
|
|
fma.rn.f32 %f382, %f438, %f438, %f381;
|
|
sqrt.rn.f32 %f383, %f382;
|
|
rcp.rn.f32 %f384, %f383;
|
|
mul.f32 %f385, %f436, %f384;
|
|
neg.f32 %f386, %f385;
|
|
mul.f32 %f387, %f437, %f384;
|
|
neg.f32 %f388, %f387;
|
|
mul.f32 %f389, %f438, %f384;
|
|
neg.f32 %f390, %f389;
|
|
st.global.f32 [normal], %f386;
|
|
st.global.f32 [normal+4], %f388;
|
|
st.global.f32 [normal+8], %f390;
|
|
mov.u32 %r28, 0;
|
|
// inline asm
|
|
call (%r27), _rt_report_intersection, (%r28);
|
|
// inline asm
|
|
|
|
BB0_32:
|
|
ret;
|
|
}
|
|
|
|
// .globl _Z4bboxiPN5optix4AabbE
|
|
.visible .entry _Z4bboxiPN5optix4AabbE(
|
|
.param .u32 _Z4bboxiPN5optix4AabbE_param_0,
|
|
.param .u64 _Z4bboxiPN5optix4AabbE_param_1
|
|
)
|
|
{
|
|
.reg .f32 %f<7>;
|
|
.reg .b64 %rd<3>;
|
|
|
|
|
|
ld.param.u64 %rd1, [_Z4bboxiPN5optix4AabbE_param_1];
|
|
cvta.to.global.u64 %rd2, %rd1;
|
|
ld.global.f32 %f1, [heightmapMin];
|
|
ld.global.f32 %f2, [heightmapMin+4];
|
|
ld.global.f32 %f3, [heightmapMin+8];
|
|
st.global.f32 [%rd2+8], %f3;
|
|
st.global.f32 [%rd2+4], %f2;
|
|
st.global.f32 [%rd2], %f1;
|
|
ld.global.f32 %f4, [heightmapMax];
|
|
ld.global.f32 %f5, [heightmapMax+4];
|
|
ld.global.f32 %f6, [heightmapMax+8];
|
|
st.global.f32 [%rd2+20], %f6;
|
|
st.global.f32 [%rd2+16], %f5;
|
|
st.global.f32 [%rd2+12], %f4;
|
|
ret;
|
|
}
|
|
|
|
|