1034 lines
30 KiB
Plaintext
1034 lines
30 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 1 .b8 image_Mask[1];
|
|
.global .align 1 .b8 image_HDR[1];
|
|
.global .align 1 .b8 image_HDR2[1];
|
|
.global .align 1 .b8 uvpos[1];
|
|
.global .align 1 .b8 uvnormal[1];
|
|
.global .align 1 .b8 rnd_seeds[1];
|
|
.global .align 4 .u32 samples;
|
|
.global .align 4 .f32 lightRadius;
|
|
.global .align 4 .f32 intensity;
|
|
.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_typeinfo7samplesE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo11lightRadiusE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo9intensityE[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 4 .b8 _ZN21rti_internal_typename7samplesE[4] = {105, 110, 116, 0};
|
|
.global .align 8 .b8 _ZN21rti_internal_typename11lightRadiusE[6] = {102, 108, 111, 97, 116, 0};
|
|
.global .align 8 .b8 _ZN21rti_internal_typename9intensityE[6] = {102, 108, 111, 97, 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_typeenum7samplesE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum11lightRadiusE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum9intensityE = 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 1 .b8 _ZN21rti_internal_semantic7samplesE[1];
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic11lightRadiusE[1];
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic9intensityE[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_annotation7samplesE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation11lightRadiusE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation9intensityE[1];
|
|
.const .align 4 .b8 __cudart_i2opi_f[24] = {65, 144, 67, 60, 153, 149, 98, 219, 192, 221, 52, 245, 209, 87, 39, 252, 41, 21, 68, 78, 110, 131, 249, 162};
|
|
|
|
.visible .entry _Z6oxMainv(
|
|
|
|
)
|
|
{
|
|
.local .align 4 .b8 __local_depot0[32];
|
|
.reg .b64 %SP;
|
|
.reg .b64 %SPL;
|
|
.reg .pred %p<61>;
|
|
.reg .b16 %rs<47>;
|
|
.reg .f32 %f<345>;
|
|
.reg .b32 %r<275>;
|
|
.reg .b64 %rd<112>;
|
|
|
|
|
|
mov.u64 %rd111, __local_depot0;
|
|
cvta.local.u64 %SP, %rd111;
|
|
ld.global.v2.u32 {%r91, %r92}, [pixelID];
|
|
cvt.u64.u32 %rd22, %r91;
|
|
cvt.u64.u32 %rd23, %r92;
|
|
mov.u64 %rd26, uvnormal;
|
|
cvta.global.u64 %rd21, %rd26;
|
|
mov.u32 %r89, 2;
|
|
mov.u32 %r90, 4;
|
|
mov.u64 %rd25, 0;
|
|
// inline asm
|
|
call (%rd20), _rt_buffer_get_64, (%rd21, %r89, %r90, %rd22, %rd23, %rd25, %rd25);
|
|
// inline asm
|
|
ld.u32 %r1, [%rd20];
|
|
shr.u32 %r95, %r1, 16;
|
|
cvt.u16.u32 %rs1, %r95;
|
|
and.b16 %rs3, %rs1, 255;
|
|
cvt.u16.u32 %rs4, %r1;
|
|
or.b16 %rs5, %rs4, %rs3;
|
|
setp.eq.s16 %p2, %rs5, 0;
|
|
mov.f32 %f323, 0f00000000;
|
|
mov.f32 %f324, %f323;
|
|
mov.f32 %f325, %f323;
|
|
@%p2 bra BB0_2;
|
|
|
|
ld.u8 %rs6, [%rd20+1];
|
|
and.b16 %rs8, %rs4, 255;
|
|
cvt.rn.f32.u16 %f81, %rs8;
|
|
div.rn.f32 %f82, %f81, 0f437F0000;
|
|
fma.rn.f32 %f83, %f82, 0f40000000, 0fBF800000;
|
|
cvt.rn.f32.u16 %f84, %rs6;
|
|
div.rn.f32 %f85, %f84, 0f437F0000;
|
|
fma.rn.f32 %f86, %f85, 0f40000000, 0fBF800000;
|
|
cvt.rn.f32.u16 %f87, %rs3;
|
|
div.rn.f32 %f88, %f87, 0f437F0000;
|
|
fma.rn.f32 %f89, %f88, 0f40000000, 0fBF800000;
|
|
mul.f32 %f90, %f86, %f86;
|
|
fma.rn.f32 %f91, %f83, %f83, %f90;
|
|
fma.rn.f32 %f92, %f89, %f89, %f91;
|
|
sqrt.rn.f32 %f93, %f92;
|
|
rcp.rn.f32 %f94, %f93;
|
|
mul.f32 %f323, %f83, %f94;
|
|
mul.f32 %f324, %f86, %f94;
|
|
mul.f32 %f325, %f89, %f94;
|
|
|
|
BB0_2:
|
|
ld.global.v2.u32 {%r96, %r97}, [pixelID];
|
|
ld.global.v2.u32 {%r99, %r100}, [tileInfo];
|
|
add.s32 %r2, %r96, %r99;
|
|
add.s32 %r3, %r97, %r100;
|
|
setp.eq.f32 %p3, %f324, 0f00000000;
|
|
setp.eq.f32 %p4, %f323, 0f00000000;
|
|
and.pred %p5, %p4, %p3;
|
|
setp.eq.f32 %p6, %f325, 0f00000000;
|
|
and.pred %p7, %p5, %p6;
|
|
@%p7 bra BB0_72;
|
|
bra.uni BB0_3;
|
|
|
|
BB0_72:
|
|
ld.global.u8 %rs28, [imageEnabled];
|
|
and.b16 %rs29, %rs28, 4;
|
|
setp.eq.s16 %p58, %rs29, 0;
|
|
@%p58 bra BB0_76;
|
|
|
|
ld.global.u32 %r237, [additive];
|
|
setp.eq.s32 %p59, %r237, 0;
|
|
cvt.u64.u32 %rd18, %r2;
|
|
cvt.u64.u32 %rd19, %r3;
|
|
@%p59 bra BB0_75;
|
|
|
|
mov.u64 %rd92, image_HDR;
|
|
cvta.global.u64 %rd81, %rd92;
|
|
mov.u32 %r241, 8;
|
|
// inline asm
|
|
call (%rd80), _rt_buffer_get_64, (%rd81, %r89, %r241, %rd18, %rd19, %rd25, %rd25);
|
|
// inline asm
|
|
ld.v4.u16 {%rs36, %rs37, %rs38, %rs39}, [%rd80];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f313, %rs36;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f314, %rs37;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f315, %rs38;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd86), _rt_buffer_get_64, (%rd81, %r89, %r241, %rd18, %rd19, %rd25, %rd25);
|
|
// inline asm
|
|
add.f32 %f316, %f313, 0f00000000;
|
|
add.f32 %f317, %f314, 0f00000000;
|
|
add.f32 %f318, %f315, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs35, %f318;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs34, %f317;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs33, %f316;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs40, 0;
|
|
st.v4.u16 [%rd86], {%rs33, %rs34, %rs35, %rs40};
|
|
bra.uni BB0_76;
|
|
|
|
BB0_3:
|
|
ld.global.v2.u32 {%r108, %r109}, [pixelID];
|
|
cvt.u64.u32 %rd29, %r108;
|
|
cvt.u64.u32 %rd30, %r109;
|
|
mov.u64 %rd39, uvpos;
|
|
cvta.global.u64 %rd28, %rd39;
|
|
mov.u32 %r105, 12;
|
|
// inline asm
|
|
call (%rd27), _rt_buffer_get_64, (%rd28, %r89, %r105, %rd29, %rd30, %rd25, %rd25);
|
|
// inline asm
|
|
ld.f32 %f96, [%rd27+8];
|
|
ld.f32 %f97, [%rd27+4];
|
|
ld.f32 %f98, [%rd27];
|
|
mul.f32 %f99, %f98, 0f3456BF95;
|
|
mul.f32 %f100, %f97, 0f3456BF95;
|
|
mul.f32 %f101, %f96, 0f3456BF95;
|
|
abs.f32 %f102, %f323;
|
|
div.rn.f32 %f103, %f99, %f102;
|
|
abs.f32 %f104, %f324;
|
|
div.rn.f32 %f105, %f100, %f104;
|
|
abs.f32 %f106, %f325;
|
|
div.rn.f32 %f107, %f101, %f106;
|
|
abs.f32 %f108, %f103;
|
|
abs.f32 %f109, %f105;
|
|
abs.f32 %f110, %f107;
|
|
mov.f32 %f111, 0f38D1B717;
|
|
max.f32 %f112, %f108, %f111;
|
|
max.f32 %f113, %f109, %f111;
|
|
max.f32 %f114, %f110, %f111;
|
|
fma.rn.f32 %f7, %f323, %f112, %f98;
|
|
fma.rn.f32 %f8, %f324, %f113, %f97;
|
|
fma.rn.f32 %f9, %f325, %f114, %f96;
|
|
setp.gt.f32 %p8, %f102, %f106;
|
|
neg.f32 %f115, %f324;
|
|
selp.f32 %f116, %f115, 0f00000000, %p8;
|
|
neg.f32 %f117, %f325;
|
|
selp.f32 %f118, %f323, %f117, %p8;
|
|
selp.f32 %f119, 0f00000000, %f324, %p8;
|
|
mul.f32 %f120, %f118, %f118;
|
|
fma.rn.f32 %f121, %f116, %f116, %f120;
|
|
fma.rn.f32 %f122, %f119, %f119, %f121;
|
|
sqrt.rn.f32 %f123, %f122;
|
|
rcp.rn.f32 %f124, %f123;
|
|
mul.f32 %f10, %f116, %f124;
|
|
mul.f32 %f11, %f118, %f124;
|
|
mul.f32 %f12, %f119, %f124;
|
|
ld.global.v2.u32 {%r112, %r113}, [pixelID];
|
|
cvt.u64.u32 %rd35, %r112;
|
|
cvt.u64.u32 %rd36, %r113;
|
|
mov.u64 %rd40, rnd_seeds;
|
|
cvta.global.u64 %rd34, %rd40;
|
|
// inline asm
|
|
call (%rd33), _rt_buffer_get_64, (%rd34, %r89, %r90, %rd35, %rd36, %rd25, %rd25);
|
|
// inline asm
|
|
ld.global.u32 %r247, [samples];
|
|
mov.f32 %f340, 0f00000000;
|
|
setp.lt.s32 %p9, %r247, 1;
|
|
@%p9 bra BB0_55;
|
|
|
|
cvt.rn.f32.s32 %f126, %r247;
|
|
rcp.rn.f32 %f13, %f126;
|
|
ld.u32 %r273, [%rd33];
|
|
mul.f32 %f14, %f7, 0f3456BF95;
|
|
mul.f32 %f15, %f8, 0f3456BF95;
|
|
mul.f32 %f16, %f9, 0f3456BF95;
|
|
mul.f32 %f127, %f323, %f11;
|
|
mul.f32 %f128, %f324, %f10;
|
|
sub.f32 %f17, %f128, %f127;
|
|
mul.f32 %f129, %f325, %f10;
|
|
mul.f32 %f130, %f323, %f12;
|
|
sub.f32 %f18, %f130, %f129;
|
|
mul.f32 %f131, %f324, %f12;
|
|
mul.f32 %f132, %f325, %f11;
|
|
sub.f32 %f19, %f132, %f131;
|
|
mov.f32 %f340, 0f00000000;
|
|
mov.u32 %r248, 0;
|
|
abs.f32 %f133, %f15;
|
|
abs.f32 %f134, %f14;
|
|
max.f32 %f135, %f134, %f133;
|
|
abs.f32 %f136, %f16;
|
|
max.f32 %f137, %f135, %f136;
|
|
|
|
BB0_5:
|
|
setp.lt.s32 %p10, %r247, 1;
|
|
@%p10 bra BB0_54;
|
|
|
|
cvt.rn.f32.s32 %f21, %r248;
|
|
max.f32 %f22, %f137, %f111;
|
|
mov.u32 %r250, 0;
|
|
|
|
BB0_7:
|
|
mad.lo.s32 %r118, %r273, 1664525, 1013904223;
|
|
and.b32 %r119, %r118, 16777215;
|
|
cvt.rn.f32.u32 %f139, %r119;
|
|
fma.rn.f32 %f140, %f139, 0f33800000, %f21;
|
|
mul.f32 %f141, %f13, %f140;
|
|
mad.lo.s32 %r273, %r118, 1664525, 1013904223;
|
|
and.b32 %r120, %r273, 16777215;
|
|
cvt.rn.f32.u32 %f142, %r120;
|
|
cvt.rn.f32.s32 %f143, %r250;
|
|
fma.rn.f32 %f144, %f142, 0f33800000, %f143;
|
|
mul.f32 %f145, %f13, %f144;
|
|
sqrt.rn.f32 %f24, %f141;
|
|
mul.f32 %f334, %f145, 0f40C90FDB;
|
|
abs.f32 %f26, %f334;
|
|
setp.neu.f32 %p11, %f26, 0f7F800000;
|
|
mov.f32 %f328, %f334;
|
|
@%p11 bra BB0_9;
|
|
|
|
mov.f32 %f146, 0f00000000;
|
|
mul.rn.f32 %f328, %f334, %f146;
|
|
|
|
BB0_9:
|
|
mul.f32 %f147, %f328, 0f3F22F983;
|
|
cvt.rni.s32.f32 %r261, %f147;
|
|
cvt.rn.f32.s32 %f148, %r261;
|
|
neg.f32 %f149, %f148;
|
|
mov.f32 %f150, 0f3FC90FDA;
|
|
fma.rn.f32 %f151, %f149, %f150, %f328;
|
|
mov.f32 %f152, 0f33A22168;
|
|
fma.rn.f32 %f153, %f149, %f152, %f151;
|
|
mov.f32 %f154, 0f27C234C5;
|
|
fma.rn.f32 %f329, %f149, %f154, %f153;
|
|
abs.f32 %f155, %f328;
|
|
setp.leu.f32 %p12, %f155, 0f47CE4780;
|
|
@%p12 bra BB0_20;
|
|
|
|
mov.b32 %r13, %f328;
|
|
shr.u32 %r14, %r13, 23;
|
|
shl.b32 %r123, %r13, 8;
|
|
or.b32 %r15, %r123, -2147483648;
|
|
add.u64 %rd42, %SP, 4;
|
|
cvta.to.local.u64 %rd108, %rd42;
|
|
mov.u32 %r253, 0;
|
|
mov.u64 %rd107, __cudart_i2opi_f;
|
|
mov.u32 %r252, -6;
|
|
|
|
BB0_11:
|
|
.pragma "nounroll";
|
|
ld.const.u32 %r126, [%rd107];
|
|
// inline asm
|
|
{
|
|
mad.lo.cc.u32 %r124, %r126, %r15, %r253;
|
|
madc.hi.u32 %r253, %r126, %r15, 0;
|
|
}
|
|
// inline asm
|
|
st.local.u32 [%rd108], %r124;
|
|
add.s64 %rd108, %rd108, 4;
|
|
add.s64 %rd107, %rd107, 4;
|
|
add.s32 %r252, %r252, 1;
|
|
setp.ne.s32 %p13, %r252, 0;
|
|
@%p13 bra BB0_11;
|
|
|
|
and.b32 %r129, %r14, 255;
|
|
add.s32 %r130, %r129, -128;
|
|
shr.u32 %r131, %r130, 5;
|
|
and.b32 %r20, %r13, -2147483648;
|
|
cvta.to.local.u64 %rd44, %rd42;
|
|
st.local.u32 [%rd44+24], %r253;
|
|
mov.u32 %r132, 6;
|
|
sub.s32 %r133, %r132, %r131;
|
|
mul.wide.s32 %rd45, %r133, 4;
|
|
add.s64 %rd8, %rd44, %rd45;
|
|
ld.local.u32 %r254, [%rd8];
|
|
ld.local.u32 %r255, [%rd8+-4];
|
|
and.b32 %r23, %r14, 31;
|
|
setp.eq.s32 %p14, %r23, 0;
|
|
@%p14 bra BB0_14;
|
|
|
|
mov.u32 %r134, 32;
|
|
sub.s32 %r135, %r134, %r23;
|
|
shr.u32 %r136, %r255, %r135;
|
|
shl.b32 %r137, %r254, %r23;
|
|
add.s32 %r254, %r136, %r137;
|
|
ld.local.u32 %r138, [%rd8+-8];
|
|
shr.u32 %r139, %r138, %r135;
|
|
shl.b32 %r140, %r255, %r23;
|
|
add.s32 %r255, %r139, %r140;
|
|
|
|
BB0_14:
|
|
shr.u32 %r141, %r255, 30;
|
|
shl.b32 %r142, %r254, 2;
|
|
add.s32 %r256, %r141, %r142;
|
|
shl.b32 %r29, %r255, 2;
|
|
shr.u32 %r143, %r256, 31;
|
|
shr.u32 %r144, %r254, 30;
|
|
add.s32 %r30, %r143, %r144;
|
|
setp.eq.s32 %p15, %r143, 0;
|
|
@%p15 bra BB0_15;
|
|
bra.uni BB0_16;
|
|
|
|
BB0_15:
|
|
mov.u32 %r257, %r20;
|
|
mov.u32 %r258, %r29;
|
|
bra.uni BB0_17;
|
|
|
|
BB0_16:
|
|
not.b32 %r145, %r256;
|
|
neg.s32 %r258, %r29;
|
|
setp.eq.s32 %p16, %r29, 0;
|
|
selp.u32 %r146, 1, 0, %p16;
|
|
add.s32 %r256, %r146, %r145;
|
|
xor.b32 %r257, %r20, -2147483648;
|
|
|
|
BB0_17:
|
|
clz.b32 %r260, %r256;
|
|
setp.eq.s32 %p17, %r260, 0;
|
|
shl.b32 %r147, %r256, %r260;
|
|
mov.u32 %r148, 32;
|
|
sub.s32 %r149, %r148, %r260;
|
|
shr.u32 %r150, %r258, %r149;
|
|
add.s32 %r151, %r150, %r147;
|
|
selp.b32 %r38, %r256, %r151, %p17;
|
|
mov.u32 %r152, -921707870;
|
|
mul.hi.u32 %r259, %r38, %r152;
|
|
setp.eq.s32 %p18, %r20, 0;
|
|
neg.s32 %r153, %r30;
|
|
selp.b32 %r261, %r30, %r153, %p18;
|
|
setp.lt.s32 %p19, %r259, 1;
|
|
@%p19 bra BB0_19;
|
|
|
|
mul.lo.s32 %r154, %r38, -921707870;
|
|
shr.u32 %r155, %r154, 31;
|
|
shl.b32 %r156, %r259, 1;
|
|
add.s32 %r259, %r155, %r156;
|
|
add.s32 %r260, %r260, 1;
|
|
|
|
BB0_19:
|
|
mov.u32 %r157, 126;
|
|
sub.s32 %r158, %r157, %r260;
|
|
shl.b32 %r159, %r158, 23;
|
|
add.s32 %r160, %r259, 1;
|
|
shr.u32 %r161, %r160, 7;
|
|
add.s32 %r162, %r161, 1;
|
|
shr.u32 %r163, %r162, 1;
|
|
add.s32 %r164, %r163, %r159;
|
|
or.b32 %r165, %r164, %r257;
|
|
mov.b32 %f329, %r165;
|
|
|
|
BB0_20:
|
|
mul.rn.f32 %f32, %f329, %f329;
|
|
add.s32 %r46, %r261, 1;
|
|
and.b32 %r47, %r46, 1;
|
|
setp.eq.s32 %p20, %r47, 0;
|
|
@%p20 bra BB0_22;
|
|
bra.uni BB0_21;
|
|
|
|
BB0_22:
|
|
mov.f32 %f158, 0f3C08839E;
|
|
mov.f32 %f159, 0fB94CA1F9;
|
|
fma.rn.f32 %f330, %f159, %f32, %f158;
|
|
bra.uni BB0_23;
|
|
|
|
BB0_21:
|
|
mov.f32 %f156, 0fBAB6061A;
|
|
mov.f32 %f157, 0f37CCF5CE;
|
|
fma.rn.f32 %f330, %f157, %f32, %f156;
|
|
|
|
BB0_23:
|
|
@%p20 bra BB0_25;
|
|
bra.uni BB0_24;
|
|
|
|
BB0_25:
|
|
mov.f32 %f163, 0fBE2AAAA3;
|
|
fma.rn.f32 %f164, %f330, %f32, %f163;
|
|
mov.f32 %f165, 0f00000000;
|
|
fma.rn.f32 %f331, %f164, %f32, %f165;
|
|
bra.uni BB0_26;
|
|
|
|
BB0_24:
|
|
mov.f32 %f160, 0f3D2AAAA5;
|
|
fma.rn.f32 %f161, %f330, %f32, %f160;
|
|
mov.f32 %f162, 0fBF000000;
|
|
fma.rn.f32 %f331, %f161, %f32, %f162;
|
|
|
|
BB0_26:
|
|
fma.rn.f32 %f332, %f331, %f329, %f329;
|
|
@%p20 bra BB0_28;
|
|
|
|
mov.f32 %f166, 0f3F800000;
|
|
fma.rn.f32 %f332, %f331, %f32, %f166;
|
|
|
|
BB0_28:
|
|
and.b32 %r166, %r46, 2;
|
|
setp.eq.s32 %p23, %r166, 0;
|
|
@%p23 bra BB0_30;
|
|
|
|
mov.f32 %f167, 0f00000000;
|
|
mov.f32 %f168, 0fBF800000;
|
|
fma.rn.f32 %f332, %f332, %f168, %f167;
|
|
|
|
BB0_30:
|
|
@%p11 bra BB0_32;
|
|
|
|
mov.f32 %f169, 0f00000000;
|
|
mul.rn.f32 %f334, %f334, %f169;
|
|
|
|
BB0_32:
|
|
mul.f32 %f170, %f334, 0f3F22F983;
|
|
cvt.rni.s32.f32 %r271, %f170;
|
|
cvt.rn.f32.s32 %f171, %r271;
|
|
neg.f32 %f172, %f171;
|
|
fma.rn.f32 %f174, %f172, %f150, %f334;
|
|
fma.rn.f32 %f176, %f172, %f152, %f174;
|
|
fma.rn.f32 %f335, %f172, %f154, %f176;
|
|
abs.f32 %f178, %f334;
|
|
setp.leu.f32 %p25, %f178, 0f47CE4780;
|
|
@%p25 bra BB0_43;
|
|
|
|
mov.b32 %r49, %f334;
|
|
shr.u32 %r50, %r49, 23;
|
|
shl.b32 %r169, %r49, 8;
|
|
or.b32 %r51, %r169, -2147483648;
|
|
add.u64 %rd47, %SP, 4;
|
|
cvta.to.local.u64 %rd110, %rd47;
|
|
mov.u32 %r263, 0;
|
|
mov.u64 %rd109, __cudart_i2opi_f;
|
|
mov.u32 %r262, -6;
|
|
|
|
BB0_34:
|
|
.pragma "nounroll";
|
|
ld.const.u32 %r172, [%rd109];
|
|
// inline asm
|
|
{
|
|
mad.lo.cc.u32 %r170, %r172, %r51, %r263;
|
|
madc.hi.u32 %r263, %r172, %r51, 0;
|
|
}
|
|
// inline asm
|
|
st.local.u32 [%rd110], %r170;
|
|
add.s64 %rd110, %rd110, 4;
|
|
add.s64 %rd109, %rd109, 4;
|
|
add.s32 %r262, %r262, 1;
|
|
setp.ne.s32 %p26, %r262, 0;
|
|
@%p26 bra BB0_34;
|
|
|
|
and.b32 %r175, %r50, 255;
|
|
add.s32 %r176, %r175, -128;
|
|
shr.u32 %r177, %r176, 5;
|
|
and.b32 %r56, %r49, -2147483648;
|
|
cvta.to.local.u64 %rd49, %rd47;
|
|
st.local.u32 [%rd49+24], %r263;
|
|
mov.u32 %r178, 6;
|
|
sub.s32 %r179, %r178, %r177;
|
|
mul.wide.s32 %rd50, %r179, 4;
|
|
add.s64 %rd14, %rd49, %rd50;
|
|
ld.local.u32 %r264, [%rd14];
|
|
ld.local.u32 %r265, [%rd14+-4];
|
|
and.b32 %r59, %r50, 31;
|
|
setp.eq.s32 %p27, %r59, 0;
|
|
@%p27 bra BB0_37;
|
|
|
|
mov.u32 %r180, 32;
|
|
sub.s32 %r181, %r180, %r59;
|
|
shr.u32 %r182, %r265, %r181;
|
|
shl.b32 %r183, %r264, %r59;
|
|
add.s32 %r264, %r182, %r183;
|
|
ld.local.u32 %r184, [%rd14+-8];
|
|
shr.u32 %r185, %r184, %r181;
|
|
shl.b32 %r186, %r265, %r59;
|
|
add.s32 %r265, %r185, %r186;
|
|
|
|
BB0_37:
|
|
shr.u32 %r187, %r265, 30;
|
|
shl.b32 %r188, %r264, 2;
|
|
add.s32 %r266, %r187, %r188;
|
|
shl.b32 %r65, %r265, 2;
|
|
shr.u32 %r189, %r266, 31;
|
|
shr.u32 %r190, %r264, 30;
|
|
add.s32 %r66, %r189, %r190;
|
|
setp.eq.s32 %p28, %r189, 0;
|
|
@%p28 bra BB0_38;
|
|
bra.uni BB0_39;
|
|
|
|
BB0_38:
|
|
mov.u32 %r267, %r56;
|
|
mov.u32 %r268, %r65;
|
|
bra.uni BB0_40;
|
|
|
|
BB0_39:
|
|
not.b32 %r191, %r266;
|
|
neg.s32 %r268, %r65;
|
|
setp.eq.s32 %p29, %r65, 0;
|
|
selp.u32 %r192, 1, 0, %p29;
|
|
add.s32 %r266, %r192, %r191;
|
|
xor.b32 %r267, %r56, -2147483648;
|
|
|
|
BB0_40:
|
|
clz.b32 %r270, %r266;
|
|
setp.eq.s32 %p30, %r270, 0;
|
|
shl.b32 %r193, %r266, %r270;
|
|
mov.u32 %r194, 32;
|
|
sub.s32 %r195, %r194, %r270;
|
|
shr.u32 %r196, %r268, %r195;
|
|
add.s32 %r197, %r196, %r193;
|
|
selp.b32 %r74, %r266, %r197, %p30;
|
|
mov.u32 %r198, -921707870;
|
|
mul.hi.u32 %r269, %r74, %r198;
|
|
setp.eq.s32 %p31, %r56, 0;
|
|
neg.s32 %r199, %r66;
|
|
selp.b32 %r271, %r66, %r199, %p31;
|
|
setp.lt.s32 %p32, %r269, 1;
|
|
@%p32 bra BB0_42;
|
|
|
|
mul.lo.s32 %r200, %r74, -921707870;
|
|
shr.u32 %r201, %r200, 31;
|
|
shl.b32 %r202, %r269, 1;
|
|
add.s32 %r269, %r201, %r202;
|
|
add.s32 %r270, %r270, 1;
|
|
|
|
BB0_42:
|
|
mov.u32 %r203, 126;
|
|
sub.s32 %r204, %r203, %r270;
|
|
shl.b32 %r205, %r204, 23;
|
|
add.s32 %r206, %r269, 1;
|
|
shr.u32 %r207, %r206, 7;
|
|
add.s32 %r208, %r207, 1;
|
|
shr.u32 %r209, %r208, 1;
|
|
add.s32 %r210, %r209, %r205;
|
|
or.b32 %r211, %r210, %r267;
|
|
mov.b32 %f335, %r211;
|
|
|
|
BB0_43:
|
|
mul.rn.f32 %f49, %f335, %f335;
|
|
and.b32 %r82, %r271, 1;
|
|
setp.eq.s32 %p33, %r82, 0;
|
|
@%p33 bra BB0_45;
|
|
bra.uni BB0_44;
|
|
|
|
BB0_45:
|
|
mov.f32 %f181, 0f3C08839E;
|
|
mov.f32 %f182, 0fB94CA1F9;
|
|
fma.rn.f32 %f336, %f182, %f49, %f181;
|
|
bra.uni BB0_46;
|
|
|
|
BB0_44:
|
|
mov.f32 %f179, 0fBAB6061A;
|
|
mov.f32 %f180, 0f37CCF5CE;
|
|
fma.rn.f32 %f336, %f180, %f49, %f179;
|
|
|
|
BB0_46:
|
|
@%p33 bra BB0_48;
|
|
bra.uni BB0_47;
|
|
|
|
BB0_48:
|
|
mov.f32 %f186, 0fBE2AAAA3;
|
|
fma.rn.f32 %f187, %f336, %f49, %f186;
|
|
mov.f32 %f188, 0f00000000;
|
|
fma.rn.f32 %f337, %f187, %f49, %f188;
|
|
bra.uni BB0_49;
|
|
|
|
BB0_47:
|
|
mov.f32 %f183, 0f3D2AAAA5;
|
|
fma.rn.f32 %f184, %f336, %f49, %f183;
|
|
mov.f32 %f185, 0fBF000000;
|
|
fma.rn.f32 %f337, %f184, %f49, %f185;
|
|
|
|
BB0_49:
|
|
fma.rn.f32 %f338, %f337, %f335, %f335;
|
|
@%p33 bra BB0_51;
|
|
|
|
mov.f32 %f189, 0f3F800000;
|
|
fma.rn.f32 %f338, %f337, %f49, %f189;
|
|
|
|
BB0_51:
|
|
and.b32 %r212, %r271, 2;
|
|
setp.eq.s32 %p36, %r212, 0;
|
|
@%p36 bra BB0_53;
|
|
|
|
mov.f32 %f190, 0f00000000;
|
|
mov.f32 %f191, 0fBF800000;
|
|
fma.rn.f32 %f338, %f338, %f191, %f190;
|
|
|
|
BB0_53:
|
|
mul.f32 %f200, %f24, %f332;
|
|
add.u64 %rd51, %SP, 0;
|
|
cvta.to.local.u64 %rd52, %rd51;
|
|
mul.f32 %f201, %f200, %f200;
|
|
mov.f32 %f202, 0f3F800000;
|
|
sub.f32 %f203, %f202, %f201;
|
|
mul.f32 %f204, %f24, %f338;
|
|
mul.f32 %f205, %f204, %f204;
|
|
sub.f32 %f206, %f203, %f205;
|
|
mov.f32 %f207, 0f00000000;
|
|
max.f32 %f208, %f207, %f206;
|
|
sqrt.rn.f32 %f209, %f208;
|
|
mul.f32 %f210, %f10, %f204;
|
|
mul.f32 %f211, %f11, %f204;
|
|
mul.f32 %f212, %f12, %f204;
|
|
fma.rn.f32 %f213, %f19, %f200, %f210;
|
|
fma.rn.f32 %f214, %f18, %f200, %f211;
|
|
fma.rn.f32 %f215, %f17, %f200, %f212;
|
|
fma.rn.f32 %f195, %f323, %f209, %f213;
|
|
fma.rn.f32 %f196, %f324, %f209, %f214;
|
|
fma.rn.f32 %f197, %f325, %f209, %f215;
|
|
ld.global.f32 %f199, [lightRadius];
|
|
mov.u32 %r216, 1065353216;
|
|
st.local.u32 [%rd52], %r216;
|
|
ld.global.u32 %r213, [root];
|
|
mov.u32 %r214, 1;
|
|
// inline asm
|
|
call _rt_trace_64, (%r213, %f7, %f8, %f9, %f195, %f196, %f197, %r214, %f22, %f199, %rd51, %r90);
|
|
// inline asm
|
|
ld.local.f32 %f216, [%rd52];
|
|
add.f32 %f340, %f340, %f216;
|
|
ld.global.u32 %r247, [samples];
|
|
add.s32 %r250, %r250, 1;
|
|
setp.lt.s32 %p37, %r250, %r247;
|
|
@%p37 bra BB0_7;
|
|
|
|
BB0_54:
|
|
add.s32 %r248, %r248, 1;
|
|
setp.lt.s32 %p38, %r248, %r247;
|
|
@%p38 bra BB0_5;
|
|
|
|
BB0_55:
|
|
mul.lo.s32 %r217, %r247, %r247;
|
|
cvt.rn.f32.s32 %f217, %r217;
|
|
div.rn.f32 %f218, %f340, %f217;
|
|
add.f32 %f219, %f218, 0fBF800000;
|
|
ld.global.f32 %f220, [intensity];
|
|
fma.rn.f32 %f221, %f220, %f219, 0f3F800000;
|
|
cvt.sat.f32.f32 %f64, %f221;
|
|
ld.global.u8 %rs10, [imageEnabled];
|
|
and.b16 %rs11, %rs10, 4;
|
|
setp.eq.s16 %p39, %rs11, 0;
|
|
@%p39 bra BB0_59;
|
|
|
|
ld.global.u32 %r218, [additive];
|
|
setp.eq.s32 %p40, %r218, 0;
|
|
cvt.u64.u32 %rd15, %r2;
|
|
cvt.u64.u32 %rd16, %r3;
|
|
mov.f32 %f222, 0f3F800000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs12, %f222;}
|
|
|
|
// inline asm
|
|
@%p40 bra BB0_58;
|
|
|
|
mov.u64 %rd65, image_HDR;
|
|
cvta.global.u64 %rd54, %rd65;
|
|
mov.u32 %r222, 8;
|
|
// inline asm
|
|
call (%rd53), _rt_buffer_get_64, (%rd54, %r89, %r222, %rd15, %rd16, %rd25, %rd25);
|
|
// inline asm
|
|
ld.v4.u16 {%rs19, %rs20, %rs21, %rs22}, [%rd53];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f223, %rs19;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f224, %rs20;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f225, %rs21;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd59), _rt_buffer_get_64, (%rd54, %r89, %r222, %rd15, %rd16, %rd25, %rd25);
|
|
// inline asm
|
|
add.f32 %f226, %f64, %f223;
|
|
add.f32 %f227, %f64, %f224;
|
|
add.f32 %f228, %f64, %f225;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs18, %f228;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs17, %f227;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs16, %f226;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd59], {%rs16, %rs17, %rs18, %rs12};
|
|
bra.uni BB0_59;
|
|
|
|
BB0_75:
|
|
mov.u64 %rd99, image_HDR;
|
|
cvta.global.u64 %rd94, %rd99;
|
|
mov.u32 %r243, 8;
|
|
// inline asm
|
|
call (%rd93), _rt_buffer_get_64, (%rd94, %r89, %r243, %rd18, %rd19, %rd25, %rd25);
|
|
// inline asm
|
|
mov.f32 %f319, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs41, %f319;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs42, 0;
|
|
st.v4.u16 [%rd93], {%rs41, %rs41, %rs41, %rs42};
|
|
|
|
BB0_76:
|
|
ld.global.u8 %rs43, [imageEnabled];
|
|
and.b16 %rs44, %rs43, 8;
|
|
setp.eq.s16 %p60, %rs44, 0;
|
|
@%p60 bra BB0_78;
|
|
|
|
cvt.u64.u32 %rd102, %r2;
|
|
cvt.u64.u32 %rd103, %r3;
|
|
mov.u64 %rd106, image_Mask;
|
|
cvta.global.u64 %rd101, %rd106;
|
|
// inline asm
|
|
call (%rd100), _rt_buffer_get_64, (%rd101, %r89, %r89, %rd102, %rd103, %rd25, %rd25);
|
|
// inline asm
|
|
mov.f32 %f322, 0f00000000;
|
|
cvt.rzi.u32.f32 %r246, %f322;
|
|
cvt.u16.u32 %rs45, %r246;
|
|
mov.u16 %rs46, 0;
|
|
st.v2.u8 [%rd100], {%rs45, %rs46};
|
|
bra.uni BB0_78;
|
|
|
|
BB0_58:
|
|
mov.u64 %rd72, image_HDR;
|
|
cvta.global.u64 %rd67, %rd72;
|
|
mov.u32 %r224, 8;
|
|
// inline asm
|
|
call (%rd66), _rt_buffer_get_64, (%rd67, %r89, %r224, %rd15, %rd16, %rd25, %rd25);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs23, %f64;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd66], {%rs23, %rs23, %rs23, %rs12};
|
|
|
|
BB0_59:
|
|
ld.global.u8 %rs24, [imageEnabled];
|
|
and.b16 %rs25, %rs24, 8;
|
|
setp.eq.s16 %p41, %rs25, 0;
|
|
@%p41 bra BB0_78;
|
|
|
|
cvt.u64.u32 %rd75, %r2;
|
|
cvt.u64.u32 %rd76, %r3;
|
|
mov.u64 %rd79, image_Mask;
|
|
cvta.global.u64 %rd74, %rd79;
|
|
// inline asm
|
|
call (%rd73), _rt_buffer_get_64, (%rd74, %r89, %r89, %rd75, %rd76, %rd25, %rd25);
|
|
// inline asm
|
|
mov.f32 %f232, 0f3E68BA2E;
|
|
cvt.rzi.f32.f32 %f233, %f232;
|
|
fma.rn.f32 %f234, %f233, 0fC0000000, 0f3EE8BA2E;
|
|
abs.f32 %f65, %f234;
|
|
abs.f32 %f66, %f64;
|
|
setp.lt.f32 %p42, %f66, 0f00800000;
|
|
mul.f32 %f235, %f66, 0f4B800000;
|
|
selp.f32 %f236, 0fC3170000, 0fC2FE0000, %p42;
|
|
selp.f32 %f237, %f235, %f66, %p42;
|
|
mov.b32 %r227, %f237;
|
|
and.b32 %r228, %r227, 8388607;
|
|
or.b32 %r229, %r228, 1065353216;
|
|
mov.b32 %f238, %r229;
|
|
shr.u32 %r230, %r227, 23;
|
|
cvt.rn.f32.u32 %f239, %r230;
|
|
add.f32 %f240, %f236, %f239;
|
|
setp.gt.f32 %p43, %f238, 0f3FB504F3;
|
|
mul.f32 %f241, %f238, 0f3F000000;
|
|
add.f32 %f242, %f240, 0f3F800000;
|
|
selp.f32 %f243, %f241, %f238, %p43;
|
|
selp.f32 %f244, %f242, %f240, %p43;
|
|
add.f32 %f245, %f243, 0fBF800000;
|
|
add.f32 %f231, %f243, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f230,%f231;
|
|
// inline asm
|
|
add.f32 %f246, %f245, %f245;
|
|
mul.f32 %f247, %f230, %f246;
|
|
mul.f32 %f248, %f247, %f247;
|
|
mov.f32 %f249, 0f3C4CAF63;
|
|
mov.f32 %f250, 0f3B18F0FE;
|
|
fma.rn.f32 %f251, %f250, %f248, %f249;
|
|
mov.f32 %f252, 0f3DAAAABD;
|
|
fma.rn.f32 %f253, %f251, %f248, %f252;
|
|
mul.rn.f32 %f254, %f253, %f248;
|
|
mul.rn.f32 %f255, %f254, %f247;
|
|
sub.f32 %f256, %f245, %f247;
|
|
neg.f32 %f257, %f247;
|
|
add.f32 %f258, %f256, %f256;
|
|
fma.rn.f32 %f259, %f257, %f245, %f258;
|
|
mul.rn.f32 %f260, %f230, %f259;
|
|
add.f32 %f261, %f255, %f247;
|
|
sub.f32 %f262, %f247, %f261;
|
|
add.f32 %f263, %f255, %f262;
|
|
add.f32 %f264, %f260, %f263;
|
|
add.f32 %f265, %f261, %f264;
|
|
sub.f32 %f266, %f261, %f265;
|
|
add.f32 %f267, %f264, %f266;
|
|
mov.f32 %f268, 0f3F317200;
|
|
mul.rn.f32 %f269, %f244, %f268;
|
|
mov.f32 %f270, 0f35BFBE8E;
|
|
mul.rn.f32 %f271, %f244, %f270;
|
|
add.f32 %f272, %f269, %f265;
|
|
sub.f32 %f273, %f269, %f272;
|
|
add.f32 %f274, %f265, %f273;
|
|
add.f32 %f275, %f267, %f274;
|
|
add.f32 %f276, %f271, %f275;
|
|
add.f32 %f277, %f272, %f276;
|
|
sub.f32 %f278, %f272, %f277;
|
|
add.f32 %f279, %f276, %f278;
|
|
mov.f32 %f280, 0f3EE8BA2E;
|
|
mul.rn.f32 %f281, %f280, %f277;
|
|
neg.f32 %f282, %f281;
|
|
fma.rn.f32 %f283, %f280, %f277, %f282;
|
|
fma.rn.f32 %f284, %f280, %f279, %f283;
|
|
mov.f32 %f285, 0f00000000;
|
|
fma.rn.f32 %f286, %f285, %f277, %f284;
|
|
add.rn.f32 %f287, %f281, %f286;
|
|
neg.f32 %f288, %f287;
|
|
add.rn.f32 %f289, %f281, %f288;
|
|
add.rn.f32 %f290, %f289, %f286;
|
|
mov.b32 %r231, %f287;
|
|
setp.eq.s32 %p44, %r231, 1118925336;
|
|
add.s32 %r232, %r231, -1;
|
|
mov.b32 %f291, %r232;
|
|
add.f32 %f292, %f290, 0f37000000;
|
|
selp.f32 %f293, %f291, %f287, %p44;
|
|
selp.f32 %f67, %f292, %f290, %p44;
|
|
mul.f32 %f294, %f293, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f295, %f294;
|
|
mov.f32 %f296, 0fBF317200;
|
|
fma.rn.f32 %f297, %f295, %f296, %f293;
|
|
mov.f32 %f298, 0fB5BFBE8E;
|
|
fma.rn.f32 %f299, %f295, %f298, %f297;
|
|
mul.f32 %f300, %f299, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f301, %f300;
|
|
add.f32 %f302, %f295, 0f00000000;
|
|
ex2.approx.f32 %f303, %f302;
|
|
mul.f32 %f304, %f301, %f303;
|
|
setp.lt.f32 %p45, %f293, 0fC2D20000;
|
|
selp.f32 %f305, 0f00000000, %f304, %p45;
|
|
setp.gt.f32 %p46, %f293, 0f42D20000;
|
|
selp.f32 %f342, 0f7F800000, %f305, %p46;
|
|
setp.eq.f32 %p47, %f342, 0f7F800000;
|
|
@%p47 bra BB0_62;
|
|
|
|
fma.rn.f32 %f342, %f342, %f67, %f342;
|
|
|
|
BB0_62:
|
|
setp.lt.f32 %p48, %f64, 0f00000000;
|
|
setp.eq.f32 %p49, %f65, 0f3F800000;
|
|
and.pred %p1, %p48, %p49;
|
|
mov.b32 %r233, %f342;
|
|
xor.b32 %r234, %r233, -2147483648;
|
|
mov.b32 %f306, %r234;
|
|
selp.f32 %f344, %f306, %f342, %p1;
|
|
setp.eq.f32 %p50, %f64, 0f00000000;
|
|
@%p50 bra BB0_65;
|
|
bra.uni BB0_63;
|
|
|
|
BB0_65:
|
|
add.f32 %f309, %f64, %f64;
|
|
selp.f32 %f344, %f309, 0f00000000, %p49;
|
|
bra.uni BB0_66;
|
|
|
|
BB0_63:
|
|
setp.geu.f32 %p51, %f64, 0f00000000;
|
|
@%p51 bra BB0_66;
|
|
|
|
cvt.rzi.f32.f32 %f308, %f280;
|
|
setp.neu.f32 %p52, %f308, 0f3EE8BA2E;
|
|
selp.f32 %f344, 0f7FFFFFFF, %f344, %p52;
|
|
|
|
BB0_66:
|
|
add.f32 %f310, %f66, 0f3EE8BA2E;
|
|
mov.b32 %r235, %f310;
|
|
setp.lt.s32 %p54, %r235, 2139095040;
|
|
@%p54 bra BB0_71;
|
|
|
|
setp.gtu.f32 %p55, %f66, 0f7F800000;
|
|
@%p55 bra BB0_70;
|
|
bra.uni BB0_68;
|
|
|
|
BB0_70:
|
|
add.f32 %f344, %f64, 0f3EE8BA2E;
|
|
bra.uni BB0_71;
|
|
|
|
BB0_68:
|
|
setp.neu.f32 %p56, %f66, 0f7F800000;
|
|
@%p56 bra BB0_71;
|
|
|
|
selp.f32 %f344, 0fFF800000, 0f7F800000, %p1;
|
|
|
|
BB0_71:
|
|
mul.f32 %f311, %f344, 0f437F0000;
|
|
setp.eq.f32 %p57, %f64, 0f3F800000;
|
|
selp.f32 %f312, 0f437F0000, %f311, %p57;
|
|
cvt.rzi.u32.f32 %r236, %f312;
|
|
cvt.u16.u32 %rs26, %r236;
|
|
mov.u16 %rs27, 255;
|
|
st.v2.u8 [%rd73], {%rs26, %rs27};
|
|
|
|
BB0_78:
|
|
ret;
|
|
}
|
|
|
|
|