ArabDesert/Assets/Editor/x64/Bakery/lambert_arealight.ptx

330 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 _Z6oxMainv
.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 .b8 payload[4];
.global .align 4 .b8 ray[36];
.global .align 4 .f32 t;
.global .align 4 .f32 lightInvCutoff;
.global .align 4 .u32 lightLMID;
.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_typeinfo7payloadE[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_typeinfo1tE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
.global .align 4 .b8 _ZN21rti_internal_typeinfo14lightInvCutoffE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
.global .align 4 .b8 _ZN21rti_internal_typeinfo9lightLMIDE[8] = {82, 97, 121, 0, 4, 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_typename7payloadE[8] = {82, 97, 121, 68, 97, 116, 97, 0};
.global .align 4 .b8 _ZN21rti_internal_typename3rayE[4] = {82, 97, 121, 0};
.global .align 8 .b8 _ZN21rti_internal_typename1tE[6] = {102, 108, 111, 97, 116, 0};
.global .align 8 .b8 _ZN21rti_internal_typename14lightInvCutoffE[6] = {102, 108, 111, 97, 116, 0};
.global .align 4 .b8 _ZN21rti_internal_typename9lightLMIDE[4] = {105, 110, 116, 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_typeenum7payloadE = 4919;
.global .align 4 .u32 _ZN21rti_internal_typeenum3rayE = 4919;
.global .align 4 .u32 _ZN21rti_internal_typeenum1tE = 4919;
.global .align 4 .u32 _ZN21rti_internal_typeenum14lightInvCutoffE = 4919;
.global .align 4 .u32 _ZN21rti_internal_typeenum9lightLMIDE = 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_semantic7payloadE[10] = {114, 116, 80, 97, 121, 108, 111, 97, 100, 0};
.global .align 16 .b8 _ZN21rti_internal_semantic3rayE[13] = {114, 116, 67, 117, 114, 114, 101, 110, 116, 82, 97, 121, 0};
.global .align 16 .b8 _ZN21rti_internal_semantic1tE[23] = {114, 116, 73, 110, 116, 101, 114, 115, 101, 99, 116, 105, 111, 110, 68, 105, 115, 116, 97, 110, 99, 101, 0};
.global .align 1 .b8 _ZN21rti_internal_semantic14lightInvCutoffE[1];
.global .align 1 .b8 _ZN21rti_internal_semantic9lightLMIDE[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_annotation9texCoordsE[1];
.global .align 1 .b8 _ZN23rti_internal_annotation7payloadE[1];
.global .align 1 .b8 _ZN23rti_internal_annotation3rayE[1];
.global .align 1 .b8 _ZN23rti_internal_annotation1tE[1];
.global .align 1 .b8 _ZN23rti_internal_annotation14lightInvCutoffE[1];
.global .align 1 .b8 _ZN23rti_internal_annotation9lightLMIDE[1];
.visible .entry _Z6oxMainv(
)
{
.reg .pred %p<23>;
.reg .f32 %f<125>;
.reg .b32 %r<16>;
ld.global.f32 %f19, [texCoords];
div.rn.f32 %f1, %f19, 0f41200000;
abs.f32 %f20, %f1;
mov.b32 %r1, %f1;
and.b32 %r2, %r1, -2147483648;
or.b32 %r3, %r2, 1056964608;
mov.b32 %f21, %r3;
add.f32 %f22, %f1, %f21;
cvt.rzi.f32.f32 %f23, %f22;
setp.gt.f32 %p2, %f20, 0f4B000000;
selp.f32 %f121, %f1, %f23, %p2;
setp.geu.f32 %p3, %f20, 0f3F000000;
@%p3 bra BB0_2;
cvt.rzi.f32.f32 %f121, %f1;
BB0_2:
ld.global.f32 %f24, [ray+12];
ld.global.f32 %f25, [normal];
ld.global.f32 %f26, [ray+16];
ld.global.f32 %f27, [normal+4];
mul.f32 %f28, %f27, %f26;
fma.rn.f32 %f29, %f25, %f24, %f28;
ld.global.f32 %f30, [ray+20];
ld.global.f32 %f31, [normal+8];
fma.rn.f32 %f32, %f31, %f30, %f29;
mov.u32 %r4, -1082130432;
st.global.u32 [payload], %r4;
setp.leu.f32 %p4, %f32, 0f00000000;
ld.global.u32 %r5, [lightLMID];
cvt.rzi.s32.f32 %r6, %f121;
setp.eq.s32 %p5, %r6, %r5;
and.pred %p6, %p4, %p5;
@!%p6 bra BB0_15;
bra.uni BB0_3;
BB0_3:
ld.global.f32 %f35, [t];
ld.global.f32 %f36, [lightInvCutoff];
mul.f32 %f5, %f35, %f36;
mov.f32 %f37, 0f40000000;
cvt.rzi.f32.f32 %f38, %f37;
add.f32 %f39, %f38, %f38;
mov.f32 %f40, 0f40800000;
sub.f32 %f41, %f40, %f39;
abs.f32 %f6, %f41;
abs.f32 %f7, %f5;
setp.lt.f32 %p7, %f7, 0f00800000;
mul.f32 %f42, %f7, 0f4B800000;
selp.f32 %f43, 0fC3170000, 0fC2FE0000, %p7;
selp.f32 %f44, %f42, %f7, %p7;
mov.b32 %r7, %f44;
and.b32 %r8, %r7, 8388607;
or.b32 %r9, %r8, 1065353216;
mov.b32 %f45, %r9;
shr.u32 %r10, %r7, 23;
cvt.rn.f32.u32 %f46, %r10;
add.f32 %f47, %f43, %f46;
setp.gt.f32 %p8, %f45, 0f3FB504F3;
mul.f32 %f48, %f45, 0f3F000000;
add.f32 %f49, %f47, 0f3F800000;
selp.f32 %f50, %f48, %f45, %p8;
selp.f32 %f51, %f49, %f47, %p8;
add.f32 %f52, %f50, 0fBF800000;
add.f32 %f34, %f50, 0f3F800000;
// inline asm
rcp.approx.ftz.f32 %f33,%f34;
// inline asm
add.f32 %f53, %f52, %f52;
mul.f32 %f54, %f33, %f53;
mul.f32 %f55, %f54, %f54;
mov.f32 %f56, 0f3C4CAF63;
mov.f32 %f57, 0f3B18F0FE;
fma.rn.f32 %f58, %f57, %f55, %f56;
mov.f32 %f59, 0f3DAAAABD;
fma.rn.f32 %f60, %f58, %f55, %f59;
mul.rn.f32 %f61, %f60, %f55;
mul.rn.f32 %f62, %f61, %f54;
sub.f32 %f63, %f52, %f54;
neg.f32 %f64, %f54;
add.f32 %f65, %f63, %f63;
fma.rn.f32 %f66, %f64, %f52, %f65;
mul.rn.f32 %f67, %f33, %f66;
add.f32 %f68, %f62, %f54;
sub.f32 %f69, %f54, %f68;
add.f32 %f70, %f62, %f69;
add.f32 %f71, %f67, %f70;
add.f32 %f72, %f68, %f71;
sub.f32 %f73, %f68, %f72;
add.f32 %f74, %f71, %f73;
mov.f32 %f75, 0f3F317200;
mul.rn.f32 %f76, %f51, %f75;
mov.f32 %f77, 0f35BFBE8E;
mul.rn.f32 %f78, %f51, %f77;
add.f32 %f79, %f76, %f72;
sub.f32 %f80, %f76, %f79;
add.f32 %f81, %f72, %f80;
add.f32 %f82, %f74, %f81;
add.f32 %f83, %f78, %f82;
add.f32 %f84, %f79, %f83;
sub.f32 %f85, %f79, %f84;
add.f32 %f86, %f83, %f85;
mul.rn.f32 %f87, %f40, %f84;
neg.f32 %f88, %f87;
fma.rn.f32 %f89, %f40, %f84, %f88;
fma.rn.f32 %f90, %f40, %f86, %f89;
mov.f32 %f91, 0f00000000;
fma.rn.f32 %f92, %f91, %f84, %f90;
add.rn.f32 %f93, %f87, %f92;
neg.f32 %f94, %f93;
add.rn.f32 %f95, %f87, %f94;
add.rn.f32 %f96, %f95, %f92;
mov.b32 %r11, %f93;
setp.eq.s32 %p9, %r11, 1118925336;
add.s32 %r12, %r11, -1;
mov.b32 %f97, %r12;
add.f32 %f98, %f96, 0f37000000;
selp.f32 %f99, %f97, %f93, %p9;
selp.f32 %f8, %f98, %f96, %p9;
mul.f32 %f100, %f99, 0f3FB8AA3B;
cvt.rzi.f32.f32 %f101, %f100;
mov.f32 %f102, 0fBF317200;
fma.rn.f32 %f103, %f101, %f102, %f99;
mov.f32 %f104, 0fB5BFBE8E;
fma.rn.f32 %f105, %f101, %f104, %f103;
mul.f32 %f106, %f105, 0f3FB8AA3B;
ex2.approx.ftz.f32 %f107, %f106;
add.f32 %f108, %f101, 0f00000000;
ex2.approx.f32 %f109, %f108;
mul.f32 %f110, %f107, %f109;
setp.lt.f32 %p10, %f99, 0fC2D20000;
selp.f32 %f111, 0f00000000, %f110, %p10;
setp.gt.f32 %p11, %f99, 0f42D20000;
selp.f32 %f122, 0f7F800000, %f111, %p11;
setp.eq.f32 %p12, %f122, 0f7F800000;
@%p12 bra BB0_5;
fma.rn.f32 %f122, %f122, %f8, %f122;
BB0_5:
setp.lt.f32 %p13, %f5, 0f00000000;
setp.eq.f32 %p14, %f6, 0f3F800000;
and.pred %p1, %p13, %p14;
mov.b32 %r13, %f122;
xor.b32 %r14, %r13, -2147483648;
mov.b32 %f112, %r14;
selp.f32 %f124, %f112, %f122, %p1;
setp.eq.f32 %p15, %f5, 0f00000000;
@%p15 bra BB0_8;
bra.uni BB0_6;
BB0_8:
add.f32 %f115, %f5, %f5;
selp.f32 %f124, %f115, 0f00000000, %p14;
bra.uni BB0_9;
BB0_6:
setp.geu.f32 %p16, %f5, 0f00000000;
@%p16 bra BB0_9;
cvt.rzi.f32.f32 %f114, %f40;
setp.neu.f32 %p17, %f114, 0f40800000;
selp.f32 %f124, 0f7FFFFFFF, %f124, %p17;
BB0_9:
add.f32 %f116, %f7, 0f40800000;
mov.b32 %r15, %f116;
setp.lt.s32 %p19, %r15, 2139095040;
@%p19 bra BB0_14;
setp.gtu.f32 %p20, %f7, 0f7F800000;
@%p20 bra BB0_13;
bra.uni BB0_11;
BB0_13:
add.f32 %f124, %f5, 0f40800000;
bra.uni BB0_14;
BB0_11:
setp.neu.f32 %p21, %f7, 0f7F800000;
@%p21 bra BB0_14;
selp.f32 %f124, 0fFF800000, 0f7F800000, %p1;
BB0_14:
mov.f32 %f117, 0f3F800000;
sub.f32 %f118, %f117, %f124;
setp.eq.f32 %p22, %f5, 0f3F800000;
selp.f32 %f119, 0f00000000, %f118, %p22;
cvt.sat.f32.f32 %f120, %f119;
st.global.f32 [payload], %f120;
BB0_15:
ret;
}