1484 lines
42 KiB
Plaintext
1484 lines
42 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[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 .texref sky;
|
|
.global .align 4 .b8 skyColor[12];
|
|
.global .align 4 .u32 samples;
|
|
.global .align 4 .u32 hemispherical;
|
|
.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_typeinfo8skyColorE[8] = {82, 97, 121, 0, 12, 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_typeinfo13hemisphericalE[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_typename8skyColorE[7] = {102, 108, 111, 97, 116, 51, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typename7samplesE[4] = {105, 110, 116, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typename13hemisphericalE[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_typeenum8skyColorE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum7samplesE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum13hemisphericalE = 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_semantic8skyColorE[1];
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic7samplesE[1];
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic13hemisphericalE[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_annotation8skyColorE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation7samplesE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation13hemisphericalE[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[40];
|
|
.reg .b64 %SP;
|
|
.reg .b64 %SPL;
|
|
.reg .pred %p<109>;
|
|
.reg .b16 %rs<42>;
|
|
.reg .f32 %f<662>;
|
|
.reg .b32 %r<326>;
|
|
.reg .b64 %rd<126>;
|
|
|
|
|
|
mov.u64 %rd125, __local_depot0;
|
|
cvta.local.u64 %SP, %rd125;
|
|
ld.global.v2.u32 {%r102, %r103}, [pixelID];
|
|
cvt.u64.u32 %rd22, %r102;
|
|
cvt.u64.u32 %rd23, %r103;
|
|
mov.u64 %rd26, uvnormal;
|
|
cvta.global.u64 %rd21, %rd26;
|
|
mov.u32 %r100, 2;
|
|
mov.u32 %r101, 4;
|
|
mov.u64 %rd25, 0;
|
|
// inline asm
|
|
call (%rd20), _rt_buffer_get_64, (%rd21, %r100, %r101, %rd22, %rd23, %rd25, %rd25);
|
|
// inline asm
|
|
ld.u32 %r1, [%rd20];
|
|
shr.u32 %r106, %r1, 16;
|
|
cvt.u16.u32 %rs1, %r106;
|
|
and.b16 %rs3, %rs1, 255;
|
|
cvt.u16.u32 %rs4, %r1;
|
|
or.b16 %rs5, %rs4, %rs3;
|
|
setp.eq.s16 %p4, %rs5, 0;
|
|
mov.f32 %f622, 0f00000000;
|
|
mov.f32 %f623, %f622;
|
|
mov.f32 %f624, %f622;
|
|
@%p4 bra BB0_2;
|
|
|
|
ld.u8 %rs6, [%rd20+1];
|
|
and.b16 %rs8, %rs4, 255;
|
|
cvt.rn.f32.u16 %f131, %rs8;
|
|
div.rn.f32 %f132, %f131, 0f437F0000;
|
|
fma.rn.f32 %f133, %f132, 0f40000000, 0fBF800000;
|
|
cvt.rn.f32.u16 %f134, %rs6;
|
|
div.rn.f32 %f135, %f134, 0f437F0000;
|
|
fma.rn.f32 %f136, %f135, 0f40000000, 0fBF800000;
|
|
cvt.rn.f32.u16 %f137, %rs3;
|
|
div.rn.f32 %f138, %f137, 0f437F0000;
|
|
fma.rn.f32 %f139, %f138, 0f40000000, 0fBF800000;
|
|
mul.f32 %f140, %f136, %f136;
|
|
fma.rn.f32 %f141, %f133, %f133, %f140;
|
|
fma.rn.f32 %f142, %f139, %f139, %f141;
|
|
sqrt.rn.f32 %f143, %f142;
|
|
rcp.rn.f32 %f144, %f143;
|
|
mul.f32 %f622, %f133, %f144;
|
|
mul.f32 %f623, %f136, %f144;
|
|
mul.f32 %f624, %f139, %f144;
|
|
|
|
BB0_2:
|
|
ld.global.v2.u32 {%r107, %r108}, [pixelID];
|
|
ld.global.v2.u32 {%r110, %r111}, [tileInfo];
|
|
add.s32 %r2, %r107, %r110;
|
|
add.s32 %r3, %r108, %r111;
|
|
setp.eq.f32 %p5, %f623, 0f00000000;
|
|
setp.eq.f32 %p6, %f622, 0f00000000;
|
|
and.pred %p7, %p6, %p5;
|
|
setp.eq.f32 %p8, %f624, 0f00000000;
|
|
and.pred %p9, %p7, %p8;
|
|
@%p9 bra BB0_102;
|
|
bra.uni BB0_3;
|
|
|
|
BB0_102:
|
|
ld.global.u32 %r325, [imageEnabled];
|
|
and.b32 %r279, %r325, 1;
|
|
setp.eq.b32 %p106, %r279, 1;
|
|
@!%p106 bra BB0_104;
|
|
bra.uni BB0_103;
|
|
|
|
BB0_103:
|
|
cvt.u64.u32 %rd93, %r2;
|
|
cvt.u64.u32 %rd94, %r3;
|
|
mov.u64 %rd97, image;
|
|
cvta.global.u64 %rd92, %rd97;
|
|
mov.u64 %rd96, 0;
|
|
// inline asm
|
|
call (%rd91), _rt_buffer_get_64, (%rd92, %r100, %r101, %rd93, %rd94, %rd96, %rd96);
|
|
// inline asm
|
|
mov.u16 %rs28, 0;
|
|
st.v4.u8 [%rd91], {%rs28, %rs28, %rs28, %rs28};
|
|
ld.global.u32 %r325, [imageEnabled];
|
|
|
|
BB0_104:
|
|
and.b32 %r282, %r325, 4;
|
|
setp.eq.s32 %p107, %r282, 0;
|
|
@%p107 bra BB0_108;
|
|
|
|
ld.global.u32 %r283, [additive];
|
|
setp.eq.s32 %p108, %r283, 0;
|
|
cvt.u64.u32 %rd18, %r2;
|
|
cvt.u64.u32 %rd19, %r3;
|
|
@%p108 bra BB0_107;
|
|
|
|
mov.u64 %rd110, image_HDR;
|
|
cvta.global.u64 %rd99, %rd110;
|
|
mov.u32 %r287, 8;
|
|
mov.u64 %rd109, 0;
|
|
// inline asm
|
|
call (%rd98), _rt_buffer_get_64, (%rd99, %r100, %r287, %rd18, %rd19, %rd109, %rd109);
|
|
// inline asm
|
|
ld.v4.u16 {%rs35, %rs36, %rs37, %rs38}, [%rd98];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f594, %rs35;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f595, %rs36;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f596, %rs37;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd104), _rt_buffer_get_64, (%rd99, %r100, %r287, %rd18, %rd19, %rd109, %rd109);
|
|
// inline asm
|
|
add.f32 %f597, %f594, 0f00000000;
|
|
add.f32 %f598, %f595, 0f00000000;
|
|
add.f32 %f599, %f596, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs34, %f599;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs33, %f598;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs32, %f597;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs39, 0;
|
|
st.v4.u16 [%rd104], {%rs32, %rs33, %rs34, %rs39};
|
|
bra.uni BB0_108;
|
|
|
|
BB0_3:
|
|
ld.global.v2.u32 {%r119, %r120}, [pixelID];
|
|
cvt.u64.u32 %rd29, %r119;
|
|
cvt.u64.u32 %rd30, %r120;
|
|
mov.u64 %rd39, uvpos;
|
|
cvta.global.u64 %rd28, %rd39;
|
|
mov.u32 %r116, 12;
|
|
// inline asm
|
|
call (%rd27), _rt_buffer_get_64, (%rd28, %r100, %r116, %rd29, %rd30, %rd25, %rd25);
|
|
// inline asm
|
|
ld.f32 %f148, [%rd27+8];
|
|
ld.f32 %f149, [%rd27+4];
|
|
ld.f32 %f150, [%rd27];
|
|
mul.f32 %f151, %f150, 0f3456BF95;
|
|
mul.f32 %f152, %f149, 0f3456BF95;
|
|
mul.f32 %f153, %f148, 0f3456BF95;
|
|
abs.f32 %f154, %f622;
|
|
div.rn.f32 %f155, %f151, %f154;
|
|
abs.f32 %f156, %f623;
|
|
div.rn.f32 %f157, %f152, %f156;
|
|
abs.f32 %f158, %f624;
|
|
div.rn.f32 %f159, %f153, %f158;
|
|
abs.f32 %f160, %f155;
|
|
abs.f32 %f161, %f157;
|
|
abs.f32 %f162, %f159;
|
|
mov.f32 %f163, 0f38D1B717;
|
|
max.f32 %f164, %f160, %f163;
|
|
max.f32 %f165, %f161, %f163;
|
|
max.f32 %f166, %f162, %f163;
|
|
fma.rn.f32 %f7, %f622, %f164, %f150;
|
|
fma.rn.f32 %f8, %f623, %f165, %f149;
|
|
fma.rn.f32 %f9, %f624, %f166, %f148;
|
|
ld.global.u32 %r4, [hemispherical];
|
|
setp.gt.f32 %p10, %f154, %f158;
|
|
neg.f32 %f167, %f623;
|
|
selp.f32 %f168, %f167, 0f00000000, %p10;
|
|
neg.f32 %f169, %f624;
|
|
selp.f32 %f170, %f622, %f169, %p10;
|
|
selp.f32 %f171, 0f00000000, %f623, %p10;
|
|
mul.f32 %f172, %f170, %f170;
|
|
fma.rn.f32 %f173, %f168, %f168, %f172;
|
|
fma.rn.f32 %f174, %f171, %f171, %f173;
|
|
sqrt.rn.f32 %f175, %f174;
|
|
rcp.rn.f32 %f176, %f175;
|
|
mul.f32 %f10, %f168, %f176;
|
|
mul.f32 %f11, %f170, %f176;
|
|
mul.f32 %f12, %f171, %f176;
|
|
ld.global.v2.u32 {%r123, %r124}, [pixelID];
|
|
cvt.u64.u32 %rd35, %r123;
|
|
cvt.u64.u32 %rd36, %r124;
|
|
mov.u64 %rd40, rnd_seeds;
|
|
cvta.global.u64 %rd34, %rd40;
|
|
// inline asm
|
|
call (%rd33), _rt_buffer_get_64, (%rd34, %r100, %r101, %rd35, %rd36, %rd25, %rd25);
|
|
// inline asm
|
|
ld.global.u32 %r320, [samples];
|
|
mov.f32 %f646, 0f00000000;
|
|
setp.lt.s32 %p11, %r320, 1;
|
|
@%p11 bra BB0_4;
|
|
|
|
cvt.rn.f32.s32 %f180, %r320;
|
|
rcp.rn.f32 %f13, %f180;
|
|
ld.u32 %r322, [%rd33];
|
|
mul.f32 %f14, %f7, 0f3456BF95;
|
|
mul.f32 %f15, %f8, 0f3456BF95;
|
|
mul.f32 %f16, %f9, 0f3456BF95;
|
|
mul.f32 %f181, %f622, %f11;
|
|
mul.f32 %f182, %f623, %f10;
|
|
sub.f32 %f17, %f182, %f181;
|
|
mul.f32 %f183, %f624, %f10;
|
|
mul.f32 %f184, %f622, %f12;
|
|
sub.f32 %f18, %f184, %f183;
|
|
mul.f32 %f185, %f623, %f12;
|
|
mul.f32 %f186, %f624, %f11;
|
|
sub.f32 %f19, %f186, %f185;
|
|
mov.f32 %f646, 0f00000000;
|
|
mov.u32 %r295, 0;
|
|
abs.f32 %f256, %f14;
|
|
abs.f32 %f257, %f15;
|
|
max.f32 %f258, %f256, %f257;
|
|
abs.f32 %f259, %f16;
|
|
max.f32 %f260, %f258, %f259;
|
|
mov.f32 %f645, %f646;
|
|
mov.f32 %f644, %f646;
|
|
|
|
BB0_6:
|
|
setp.lt.s32 %p12, %r320, 1;
|
|
@%p12 bra BB0_62;
|
|
|
|
cvt.rn.f32.s32 %f23, %r295;
|
|
mov.u32 %r298, 0;
|
|
|
|
BB0_8:
|
|
mad.lo.s32 %r129, %r322, 1664525, 1013904223;
|
|
and.b32 %r130, %r129, 16777215;
|
|
cvt.rn.f32.u32 %f187, %r130;
|
|
fma.rn.f32 %f188, %f187, 0f33800000, %f23;
|
|
mul.f32 %f189, %f13, %f188;
|
|
mad.lo.s32 %r322, %r129, 1664525, 1013904223;
|
|
and.b32 %r131, %r322, 16777215;
|
|
cvt.rn.f32.u32 %f190, %r131;
|
|
cvt.rn.f32.s32 %f191, %r298;
|
|
fma.rn.f32 %f192, %f190, 0f33800000, %f191;
|
|
mul.f32 %f193, %f13, %f192;
|
|
sqrt.rn.f32 %f27, %f189;
|
|
mul.f32 %f637, %f193, 0f40C90FDB;
|
|
abs.f32 %f29, %f637;
|
|
setp.neu.f32 %p13, %f29, 0f7F800000;
|
|
mov.f32 %f631, %f637;
|
|
@%p13 bra BB0_10;
|
|
|
|
mov.f32 %f194, 0f00000000;
|
|
mul.rn.f32 %f631, %f637, %f194;
|
|
|
|
BB0_10:
|
|
mul.f32 %f195, %f631, 0f3F22F983;
|
|
cvt.rni.s32.f32 %r309, %f195;
|
|
cvt.rn.f32.s32 %f196, %r309;
|
|
neg.f32 %f197, %f196;
|
|
mov.f32 %f198, 0f3FC90FDA;
|
|
fma.rn.f32 %f199, %f197, %f198, %f631;
|
|
mov.f32 %f200, 0f33A22168;
|
|
fma.rn.f32 %f201, %f197, %f200, %f199;
|
|
mov.f32 %f202, 0f27C234C5;
|
|
fma.rn.f32 %f632, %f197, %f202, %f201;
|
|
abs.f32 %f203, %f631;
|
|
setp.leu.f32 %p14, %f203, 0f47CE4780;
|
|
@%p14 bra BB0_21;
|
|
|
|
add.u64 %rd42, %SP, 0;
|
|
cvta.to.local.u64 %rd3, %rd42;
|
|
mov.b32 %r15, %f631;
|
|
shr.u32 %r16, %r15, 23;
|
|
shl.b32 %r134, %r15, 8;
|
|
or.b32 %r17, %r134, -2147483648;
|
|
mov.u32 %r300, 0;
|
|
mov.u64 %rd122, 0;
|
|
mov.u64 %rd121, %rd3;
|
|
mov.u32 %r301, %r300;
|
|
|
|
BB0_12:
|
|
.pragma "nounroll";
|
|
shl.b64 %rd43, %rd122, 2;
|
|
mov.u64 %rd44, __cudart_i2opi_f;
|
|
add.s64 %rd45, %rd44, %rd43;
|
|
ld.const.u32 %r137, [%rd45];
|
|
// inline asm
|
|
{
|
|
mad.lo.cc.u32 %r135, %r137, %r17, %r301;
|
|
madc.hi.u32 %r301, %r137, %r17, 0;
|
|
}
|
|
// inline asm
|
|
st.local.u32 [%rd121], %r135;
|
|
add.s32 %r300, %r300, 1;
|
|
cvt.s64.s32 %rd122, %r300;
|
|
mul.wide.s32 %rd48, %r300, 4;
|
|
add.s64 %rd121, %rd3, %rd48;
|
|
setp.ne.s32 %p15, %r300, 6;
|
|
@%p15 bra BB0_12;
|
|
|
|
and.b32 %r140, %r16, 255;
|
|
add.s32 %r141, %r140, -128;
|
|
shr.u32 %r142, %r141, 5;
|
|
and.b32 %r22, %r15, -2147483648;
|
|
cvta.to.local.u64 %rd50, %rd42;
|
|
st.local.u32 [%rd50+24], %r301;
|
|
mov.u32 %r143, 6;
|
|
sub.s32 %r144, %r143, %r142;
|
|
mul.wide.s32 %rd51, %r144, 4;
|
|
add.s64 %rd8, %rd50, %rd51;
|
|
ld.local.u32 %r302, [%rd8];
|
|
ld.local.u32 %r303, [%rd8+-4];
|
|
and.b32 %r25, %r16, 31;
|
|
setp.eq.s32 %p16, %r25, 0;
|
|
@%p16 bra BB0_15;
|
|
|
|
mov.u32 %r145, 32;
|
|
sub.s32 %r146, %r145, %r25;
|
|
shr.u32 %r147, %r303, %r146;
|
|
shl.b32 %r148, %r302, %r25;
|
|
add.s32 %r302, %r147, %r148;
|
|
ld.local.u32 %r149, [%rd8+-8];
|
|
shr.u32 %r150, %r149, %r146;
|
|
shl.b32 %r151, %r303, %r25;
|
|
add.s32 %r303, %r150, %r151;
|
|
|
|
BB0_15:
|
|
shr.u32 %r152, %r303, 30;
|
|
shl.b32 %r153, %r302, 2;
|
|
add.s32 %r304, %r152, %r153;
|
|
shl.b32 %r31, %r303, 2;
|
|
shr.u32 %r154, %r304, 31;
|
|
shr.u32 %r155, %r302, 30;
|
|
add.s32 %r32, %r154, %r155;
|
|
setp.eq.s32 %p17, %r154, 0;
|
|
@%p17 bra BB0_16;
|
|
bra.uni BB0_17;
|
|
|
|
BB0_16:
|
|
mov.u32 %r305, %r22;
|
|
mov.u32 %r306, %r31;
|
|
bra.uni BB0_18;
|
|
|
|
BB0_17:
|
|
not.b32 %r156, %r304;
|
|
neg.s32 %r306, %r31;
|
|
setp.eq.s32 %p18, %r31, 0;
|
|
selp.u32 %r157, 1, 0, %p18;
|
|
add.s32 %r304, %r157, %r156;
|
|
xor.b32 %r305, %r22, -2147483648;
|
|
|
|
BB0_18:
|
|
clz.b32 %r308, %r304;
|
|
setp.eq.s32 %p19, %r308, 0;
|
|
shl.b32 %r158, %r304, %r308;
|
|
mov.u32 %r159, 32;
|
|
sub.s32 %r160, %r159, %r308;
|
|
shr.u32 %r161, %r306, %r160;
|
|
add.s32 %r162, %r161, %r158;
|
|
selp.b32 %r40, %r304, %r162, %p19;
|
|
mov.u32 %r163, -921707870;
|
|
mul.hi.u32 %r307, %r40, %r163;
|
|
setp.eq.s32 %p20, %r22, 0;
|
|
neg.s32 %r164, %r32;
|
|
selp.b32 %r309, %r32, %r164, %p20;
|
|
setp.lt.s32 %p21, %r307, 1;
|
|
@%p21 bra BB0_20;
|
|
|
|
mul.lo.s32 %r165, %r40, -921707870;
|
|
shr.u32 %r166, %r165, 31;
|
|
shl.b32 %r167, %r307, 1;
|
|
add.s32 %r307, %r166, %r167;
|
|
add.s32 %r308, %r308, 1;
|
|
|
|
BB0_20:
|
|
mov.u32 %r168, 126;
|
|
sub.s32 %r169, %r168, %r308;
|
|
shl.b32 %r170, %r169, 23;
|
|
add.s32 %r171, %r307, 1;
|
|
shr.u32 %r172, %r171, 7;
|
|
add.s32 %r173, %r172, 1;
|
|
shr.u32 %r174, %r173, 1;
|
|
add.s32 %r175, %r174, %r170;
|
|
or.b32 %r176, %r175, %r305;
|
|
mov.b32 %f632, %r176;
|
|
|
|
BB0_21:
|
|
mul.rn.f32 %f35, %f632, %f632;
|
|
add.s32 %r48, %r309, 1;
|
|
and.b32 %r49, %r48, 1;
|
|
setp.eq.s32 %p22, %r49, 0;
|
|
@%p22 bra BB0_23;
|
|
bra.uni BB0_22;
|
|
|
|
BB0_23:
|
|
mov.f32 %f206, 0f3C08839E;
|
|
mov.f32 %f207, 0fB94CA1F9;
|
|
fma.rn.f32 %f633, %f207, %f35, %f206;
|
|
bra.uni BB0_24;
|
|
|
|
BB0_22:
|
|
mov.f32 %f204, 0fBAB6061A;
|
|
mov.f32 %f205, 0f37CCF5CE;
|
|
fma.rn.f32 %f633, %f205, %f35, %f204;
|
|
|
|
BB0_24:
|
|
@%p22 bra BB0_26;
|
|
bra.uni BB0_25;
|
|
|
|
BB0_26:
|
|
mov.f32 %f211, 0fBE2AAAA3;
|
|
fma.rn.f32 %f212, %f633, %f35, %f211;
|
|
mov.f32 %f213, 0f00000000;
|
|
fma.rn.f32 %f634, %f212, %f35, %f213;
|
|
bra.uni BB0_27;
|
|
|
|
BB0_25:
|
|
mov.f32 %f208, 0f3D2AAAA5;
|
|
fma.rn.f32 %f209, %f633, %f35, %f208;
|
|
mov.f32 %f210, 0fBF000000;
|
|
fma.rn.f32 %f634, %f209, %f35, %f210;
|
|
|
|
BB0_27:
|
|
fma.rn.f32 %f635, %f634, %f632, %f632;
|
|
@%p22 bra BB0_29;
|
|
|
|
mov.f32 %f214, 0f3F800000;
|
|
fma.rn.f32 %f635, %f634, %f35, %f214;
|
|
|
|
BB0_29:
|
|
and.b32 %r177, %r48, 2;
|
|
setp.eq.s32 %p25, %r177, 0;
|
|
@%p25 bra BB0_31;
|
|
|
|
mov.f32 %f215, 0f00000000;
|
|
mov.f32 %f216, 0fBF800000;
|
|
fma.rn.f32 %f635, %f635, %f216, %f215;
|
|
|
|
BB0_31:
|
|
@%p13 bra BB0_33;
|
|
|
|
mov.f32 %f217, 0f00000000;
|
|
mul.rn.f32 %f637, %f637, %f217;
|
|
|
|
BB0_33:
|
|
mul.f32 %f218, %f637, 0f3F22F983;
|
|
cvt.rni.s32.f32 %r319, %f218;
|
|
cvt.rn.f32.s32 %f219, %r319;
|
|
neg.f32 %f220, %f219;
|
|
fma.rn.f32 %f222, %f220, %f198, %f637;
|
|
fma.rn.f32 %f224, %f220, %f200, %f222;
|
|
fma.rn.f32 %f638, %f220, %f202, %f224;
|
|
abs.f32 %f226, %f637;
|
|
setp.leu.f32 %p27, %f226, 0f47CE4780;
|
|
@%p27 bra BB0_44;
|
|
|
|
add.u64 %rd53, %SP, 0;
|
|
cvta.to.local.u64 %rd9, %rd53;
|
|
mov.b32 %r51, %f637;
|
|
shr.u32 %r52, %r51, 23;
|
|
shl.b32 %r180, %r51, 8;
|
|
or.b32 %r53, %r180, -2147483648;
|
|
mov.u32 %r310, 0;
|
|
mov.u64 %rd123, %rd9;
|
|
mov.u64 %rd124, %rd25;
|
|
mov.u32 %r311, %r310;
|
|
|
|
BB0_35:
|
|
.pragma "nounroll";
|
|
shl.b64 %rd54, %rd124, 2;
|
|
mov.u64 %rd55, __cudart_i2opi_f;
|
|
add.s64 %rd56, %rd55, %rd54;
|
|
ld.const.u32 %r183, [%rd56];
|
|
// inline asm
|
|
{
|
|
mad.lo.cc.u32 %r181, %r183, %r53, %r311;
|
|
madc.hi.u32 %r311, %r183, %r53, 0;
|
|
}
|
|
// inline asm
|
|
st.local.u32 [%rd123], %r181;
|
|
add.s32 %r310, %r310, 1;
|
|
cvt.s64.s32 %rd124, %r310;
|
|
mul.wide.s32 %rd57, %r310, 4;
|
|
add.s64 %rd123, %rd9, %rd57;
|
|
setp.ne.s32 %p28, %r310, 6;
|
|
@%p28 bra BB0_35;
|
|
|
|
and.b32 %r186, %r52, 255;
|
|
add.s32 %r187, %r186, -128;
|
|
shr.u32 %r188, %r187, 5;
|
|
and.b32 %r58, %r51, -2147483648;
|
|
cvta.to.local.u64 %rd59, %rd53;
|
|
st.local.u32 [%rd59+24], %r311;
|
|
mov.u32 %r189, 6;
|
|
sub.s32 %r190, %r189, %r188;
|
|
mul.wide.s32 %rd60, %r190, 4;
|
|
add.s64 %rd15, %rd59, %rd60;
|
|
ld.local.u32 %r312, [%rd15];
|
|
ld.local.u32 %r313, [%rd15+-4];
|
|
and.b32 %r61, %r52, 31;
|
|
setp.eq.s32 %p29, %r61, 0;
|
|
@%p29 bra BB0_38;
|
|
|
|
mov.u32 %r191, 32;
|
|
sub.s32 %r192, %r191, %r61;
|
|
shr.u32 %r193, %r313, %r192;
|
|
shl.b32 %r194, %r312, %r61;
|
|
add.s32 %r312, %r193, %r194;
|
|
ld.local.u32 %r195, [%rd15+-8];
|
|
shr.u32 %r196, %r195, %r192;
|
|
shl.b32 %r197, %r313, %r61;
|
|
add.s32 %r313, %r196, %r197;
|
|
|
|
BB0_38:
|
|
shr.u32 %r198, %r313, 30;
|
|
shl.b32 %r199, %r312, 2;
|
|
add.s32 %r314, %r198, %r199;
|
|
shl.b32 %r67, %r313, 2;
|
|
shr.u32 %r200, %r314, 31;
|
|
shr.u32 %r201, %r312, 30;
|
|
add.s32 %r68, %r200, %r201;
|
|
setp.eq.s32 %p30, %r200, 0;
|
|
@%p30 bra BB0_39;
|
|
bra.uni BB0_40;
|
|
|
|
BB0_39:
|
|
mov.u32 %r315, %r58;
|
|
mov.u32 %r316, %r67;
|
|
bra.uni BB0_41;
|
|
|
|
BB0_40:
|
|
not.b32 %r202, %r314;
|
|
neg.s32 %r316, %r67;
|
|
setp.eq.s32 %p31, %r67, 0;
|
|
selp.u32 %r203, 1, 0, %p31;
|
|
add.s32 %r314, %r203, %r202;
|
|
xor.b32 %r315, %r58, -2147483648;
|
|
|
|
BB0_41:
|
|
clz.b32 %r318, %r314;
|
|
setp.eq.s32 %p32, %r318, 0;
|
|
shl.b32 %r204, %r314, %r318;
|
|
mov.u32 %r205, 32;
|
|
sub.s32 %r206, %r205, %r318;
|
|
shr.u32 %r207, %r316, %r206;
|
|
add.s32 %r208, %r207, %r204;
|
|
selp.b32 %r76, %r314, %r208, %p32;
|
|
mov.u32 %r209, -921707870;
|
|
mul.hi.u32 %r317, %r76, %r209;
|
|
setp.eq.s32 %p33, %r58, 0;
|
|
neg.s32 %r210, %r68;
|
|
selp.b32 %r319, %r68, %r210, %p33;
|
|
setp.lt.s32 %p34, %r317, 1;
|
|
@%p34 bra BB0_43;
|
|
|
|
mul.lo.s32 %r211, %r76, -921707870;
|
|
shr.u32 %r212, %r211, 31;
|
|
shl.b32 %r213, %r317, 1;
|
|
add.s32 %r317, %r212, %r213;
|
|
add.s32 %r318, %r318, 1;
|
|
|
|
BB0_43:
|
|
mov.u32 %r214, 126;
|
|
sub.s32 %r215, %r214, %r318;
|
|
shl.b32 %r216, %r215, 23;
|
|
add.s32 %r217, %r317, 1;
|
|
shr.u32 %r218, %r217, 7;
|
|
add.s32 %r219, %r218, 1;
|
|
shr.u32 %r220, %r219, 1;
|
|
add.s32 %r221, %r220, %r216;
|
|
or.b32 %r222, %r221, %r315;
|
|
mov.b32 %f638, %r222;
|
|
|
|
BB0_44:
|
|
mul.rn.f32 %f52, %f638, %f638;
|
|
and.b32 %r84, %r319, 1;
|
|
setp.eq.s32 %p35, %r84, 0;
|
|
@%p35 bra BB0_46;
|
|
bra.uni BB0_45;
|
|
|
|
BB0_46:
|
|
mov.f32 %f229, 0f3C08839E;
|
|
mov.f32 %f230, 0fB94CA1F9;
|
|
fma.rn.f32 %f639, %f230, %f52, %f229;
|
|
bra.uni BB0_47;
|
|
|
|
BB0_45:
|
|
mov.f32 %f227, 0fBAB6061A;
|
|
mov.f32 %f228, 0f37CCF5CE;
|
|
fma.rn.f32 %f639, %f228, %f52, %f227;
|
|
|
|
BB0_47:
|
|
@%p35 bra BB0_49;
|
|
bra.uni BB0_48;
|
|
|
|
BB0_49:
|
|
mov.f32 %f234, 0fBE2AAAA3;
|
|
fma.rn.f32 %f235, %f639, %f52, %f234;
|
|
mov.f32 %f236, 0f00000000;
|
|
fma.rn.f32 %f640, %f235, %f52, %f236;
|
|
bra.uni BB0_50;
|
|
|
|
BB0_48:
|
|
mov.f32 %f231, 0f3D2AAAA5;
|
|
fma.rn.f32 %f232, %f639, %f52, %f231;
|
|
mov.f32 %f233, 0fBF000000;
|
|
fma.rn.f32 %f640, %f232, %f52, %f233;
|
|
|
|
BB0_50:
|
|
fma.rn.f32 %f641, %f640, %f638, %f638;
|
|
@%p35 bra BB0_52;
|
|
|
|
mov.f32 %f237, 0f3F800000;
|
|
fma.rn.f32 %f641, %f640, %f52, %f237;
|
|
|
|
BB0_52:
|
|
and.b32 %r223, %r319, 2;
|
|
setp.eq.s32 %p38, %r223, 0;
|
|
@%p38 bra BB0_54;
|
|
|
|
mov.f32 %f238, 0f00000000;
|
|
mov.f32 %f239, 0fBF800000;
|
|
fma.rn.f32 %f641, %f641, %f239, %f238;
|
|
|
|
BB0_54:
|
|
mul.f32 %f240, %f27, %f635;
|
|
mul.f32 %f241, %f240, %f240;
|
|
mov.f32 %f242, 0f3F800000;
|
|
sub.f32 %f243, %f242, %f241;
|
|
mul.f32 %f244, %f27, %f641;
|
|
mul.f32 %f245, %f244, %f244;
|
|
sub.f32 %f246, %f243, %f245;
|
|
mov.f32 %f247, 0f00000000;
|
|
max.f32 %f248, %f247, %f246;
|
|
sqrt.rn.f32 %f249, %f248;
|
|
mul.f32 %f250, %f10, %f244;
|
|
mul.f32 %f251, %f11, %f244;
|
|
mul.f32 %f252, %f12, %f244;
|
|
fma.rn.f32 %f253, %f19, %f240, %f250;
|
|
fma.rn.f32 %f254, %f18, %f240, %f251;
|
|
fma.rn.f32 %f255, %f17, %f240, %f252;
|
|
fma.rn.f32 %f64, %f622, %f249, %f253;
|
|
fma.rn.f32 %f65, %f623, %f249, %f254;
|
|
fma.rn.f32 %f66, %f624, %f249, %f255;
|
|
setp.gt.f32 %p39, %f65, 0f00000000;
|
|
setp.eq.s32 %p40, %r4, 0;
|
|
or.pred %p41, %p40, %p39;
|
|
@!%p41 bra BB0_61;
|
|
bra.uni BB0_55;
|
|
|
|
BB0_55:
|
|
max.f32 %f67, %f260, %f163;
|
|
abs.f32 %f68, %f66;
|
|
setp.eq.f32 %p42, %f68, 0f00000000;
|
|
abs.f32 %f69, %f64;
|
|
setp.eq.f32 %p43, %f69, 0f00000000;
|
|
and.pred %p44, %p42, %p43;
|
|
mov.b32 %r85, %f66;
|
|
mov.b32 %r224, %f64;
|
|
and.b32 %r86, %r224, -2147483648;
|
|
@%p44 bra BB0_59;
|
|
bra.uni BB0_56;
|
|
|
|
BB0_59:
|
|
shr.s32 %r231, %r85, 31;
|
|
and.b32 %r232, %r231, 1078530011;
|
|
or.b32 %r233, %r232, %r86;
|
|
mov.b32 %f643, %r233;
|
|
bra.uni BB0_60;
|
|
|
|
BB0_56:
|
|
setp.eq.f32 %p45, %f68, 0f7F800000;
|
|
setp.eq.f32 %p46, %f69, 0f7F800000;
|
|
and.pred %p47, %p45, %p46;
|
|
@%p47 bra BB0_58;
|
|
bra.uni BB0_57;
|
|
|
|
BB0_58:
|
|
shr.s32 %r227, %r85, 31;
|
|
and.b32 %r228, %r227, 13483017;
|
|
add.s32 %r229, %r228, 1061752795;
|
|
or.b32 %r230, %r229, %r86;
|
|
mov.b32 %f643, %r230;
|
|
bra.uni BB0_60;
|
|
|
|
BB0_57:
|
|
max.f32 %f262, %f69, %f68;
|
|
min.f32 %f263, %f69, %f68;
|
|
div.rn.f32 %f264, %f263, %f262;
|
|
mul.rn.f32 %f265, %f264, %f264;
|
|
mov.f32 %f266, 0fC0B59883;
|
|
mov.f32 %f267, 0fBF52C7EA;
|
|
fma.rn.f32 %f268, %f265, %f267, %f266;
|
|
mov.f32 %f269, 0fC0D21907;
|
|
fma.rn.f32 %f270, %f268, %f265, %f269;
|
|
mul.f32 %f271, %f265, %f270;
|
|
mul.f32 %f272, %f264, %f271;
|
|
add.f32 %f273, %f265, 0f41355DC0;
|
|
mov.f32 %f274, 0f41E6BD60;
|
|
fma.rn.f32 %f275, %f273, %f265, %f274;
|
|
mov.f32 %f276, 0f419D92C8;
|
|
fma.rn.f32 %f277, %f275, %f265, %f276;
|
|
rcp.rn.f32 %f278, %f277;
|
|
fma.rn.f32 %f279, %f272, %f278, %f264;
|
|
mov.f32 %f280, 0f3FC90FDB;
|
|
sub.f32 %f281, %f280, %f279;
|
|
setp.gt.f32 %p48, %f69, %f68;
|
|
selp.f32 %f282, %f281, %f279, %p48;
|
|
mov.f32 %f283, 0f40490FDB;
|
|
sub.f32 %f284, %f283, %f282;
|
|
setp.lt.s32 %p49, %r85, 0;
|
|
selp.f32 %f285, %f284, %f282, %p49;
|
|
mov.b32 %r225, %f285;
|
|
or.b32 %r226, %r225, %r86;
|
|
mov.b32 %f286, %r226;
|
|
add.f32 %f287, %f68, %f69;
|
|
setp.gtu.f32 %p50, %f287, 0f7F800000;
|
|
selp.f32 %f643, %f287, %f286, %p50;
|
|
|
|
BB0_60:
|
|
add.u64 %rd61, %SP, 28;
|
|
cvta.to.local.u64 %rd62, %rd61;
|
|
add.f32 %f296, %f643, 0f40490FDB;
|
|
mul.f32 %f297, %f296, 0f3E22F983;
|
|
neg.f32 %f298, %f65;
|
|
abs.f32 %f299, %f298;
|
|
sub.f32 %f301, %f242, %f299;
|
|
mul.f32 %f302, %f301, 0f3F000000;
|
|
sqrt.rn.f32 %f303, %f302;
|
|
setp.gt.f32 %p51, %f299, 0f3F11EB85;
|
|
selp.f32 %f304, %f303, %f299, %p51;
|
|
mul.f32 %f305, %f304, %f304;
|
|
mov.f32 %f306, 0f3C94D2E9;
|
|
mov.f32 %f307, 0f3D53F941;
|
|
fma.rn.f32 %f308, %f307, %f305, %f306;
|
|
mov.f32 %f309, 0f3D3F841F;
|
|
fma.rn.f32 %f310, %f308, %f305, %f309;
|
|
mov.f32 %f311, 0f3D994929;
|
|
fma.rn.f32 %f312, %f310, %f305, %f311;
|
|
mov.f32 %f313, 0f3E2AAB94;
|
|
fma.rn.f32 %f314, %f312, %f305, %f313;
|
|
mul.f32 %f315, %f305, %f314;
|
|
fma.rn.f32 %f316, %f315, %f304, %f304;
|
|
mov.f32 %f317, 0f3FC90FDB;
|
|
sub.f32 %f318, %f317, %f316;
|
|
add.f32 %f319, %f316, %f316;
|
|
selp.f32 %f320, %f319, %f318, %p51;
|
|
mov.f32 %f321, 0f40490FDB;
|
|
sub.f32 %f322, %f321, %f320;
|
|
setp.gt.f32 %p52, %f65, 0f80000000;
|
|
selp.f32 %f323, %f322, %f320, %p52;
|
|
fma.rn.f32 %f324, %f323, 0fBEA2F983, 0f3F800000;
|
|
tex.2d.v4.f32.f32 {%f325, %f326, %f327, %f328}, [sky, {%f297, %f324}];
|
|
st.local.f32 [%rd62], %f325;
|
|
st.local.f32 [%rd62+4], %f326;
|
|
st.local.f32 [%rd62+8], %f327;
|
|
ld.global.u32 %r234, [root];
|
|
mov.u32 %r235, 1;
|
|
mov.f32 %f295, 0f6C4ECB8F;
|
|
// inline asm
|
|
call _rt_trace_64, (%r234, %f7, %f8, %f9, %f64, %f65, %f66, %r235, %f67, %f295, %rd61, %r116);
|
|
// inline asm
|
|
ld.local.f32 %f329, [%rd62];
|
|
add.f32 %f644, %f644, %f329;
|
|
ld.local.f32 %f330, [%rd62+4];
|
|
add.f32 %f645, %f645, %f330;
|
|
ld.local.f32 %f331, [%rd62+8];
|
|
add.f32 %f646, %f646, %f331;
|
|
ld.global.u32 %r320, [samples];
|
|
|
|
BB0_61:
|
|
add.s32 %r298, %r298, 1;
|
|
setp.lt.s32 %p53, %r298, %r320;
|
|
@%p53 bra BB0_8;
|
|
|
|
BB0_62:
|
|
add.s32 %r295, %r295, 1;
|
|
setp.lt.s32 %p54, %r295, %r320;
|
|
@%p54 bra BB0_6;
|
|
bra.uni BB0_63;
|
|
|
|
BB0_4:
|
|
mov.f32 %f645, %f646;
|
|
mov.f32 %f644, %f646;
|
|
|
|
BB0_63:
|
|
mul.lo.s32 %r237, %r320, %r320;
|
|
cvt.rn.f32.s32 %f332, %r237;
|
|
rcp.rn.f32 %f333, %f332;
|
|
mul.f32 %f334, %f644, %f333;
|
|
mul.f32 %f335, %f645, %f333;
|
|
mul.f32 %f336, %f646, %f333;
|
|
ld.global.f32 %f337, [skyColor];
|
|
mul.f32 %f86, %f334, %f337;
|
|
ld.global.f32 %f338, [skyColor+4];
|
|
mul.f32 %f87, %f335, %f338;
|
|
ld.global.f32 %f339, [skyColor+8];
|
|
mul.f32 %f88, %f336, %f339;
|
|
ld.global.u32 %r324, [imageEnabled];
|
|
and.b32 %r238, %r324, 1;
|
|
setp.eq.b32 %p55, %r238, 1;
|
|
@!%p55 bra BB0_98;
|
|
bra.uni BB0_64;
|
|
|
|
BB0_64:
|
|
mov.f32 %f342, 0f3E666666;
|
|
cvt.rzi.f32.f32 %f343, %f342;
|
|
fma.rn.f32 %f344, %f343, 0fC0000000, 0f3EE66666;
|
|
abs.f32 %f89, %f344;
|
|
abs.f32 %f90, %f86;
|
|
setp.lt.f32 %p56, %f90, 0f00800000;
|
|
mul.f32 %f345, %f90, 0f4B800000;
|
|
selp.f32 %f346, 0fC3170000, 0fC2FE0000, %p56;
|
|
selp.f32 %f347, %f345, %f90, %p56;
|
|
mov.b32 %r239, %f347;
|
|
and.b32 %r240, %r239, 8388607;
|
|
or.b32 %r241, %r240, 1065353216;
|
|
mov.b32 %f348, %r241;
|
|
shr.u32 %r242, %r239, 23;
|
|
cvt.rn.f32.u32 %f349, %r242;
|
|
add.f32 %f350, %f346, %f349;
|
|
setp.gt.f32 %p57, %f348, 0f3FB504F3;
|
|
mul.f32 %f351, %f348, 0f3F000000;
|
|
add.f32 %f352, %f350, 0f3F800000;
|
|
selp.f32 %f353, %f351, %f348, %p57;
|
|
selp.f32 %f354, %f352, %f350, %p57;
|
|
add.f32 %f355, %f353, 0fBF800000;
|
|
add.f32 %f341, %f353, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f340,%f341;
|
|
// inline asm
|
|
add.f32 %f356, %f355, %f355;
|
|
mul.f32 %f357, %f340, %f356;
|
|
mul.f32 %f358, %f357, %f357;
|
|
mov.f32 %f359, 0f3C4CAF63;
|
|
mov.f32 %f360, 0f3B18F0FE;
|
|
fma.rn.f32 %f361, %f360, %f358, %f359;
|
|
mov.f32 %f362, 0f3DAAAABD;
|
|
fma.rn.f32 %f363, %f361, %f358, %f362;
|
|
mul.rn.f32 %f364, %f363, %f358;
|
|
mul.rn.f32 %f365, %f364, %f357;
|
|
sub.f32 %f366, %f355, %f357;
|
|
neg.f32 %f367, %f357;
|
|
add.f32 %f368, %f366, %f366;
|
|
fma.rn.f32 %f369, %f367, %f355, %f368;
|
|
mul.rn.f32 %f370, %f340, %f369;
|
|
add.f32 %f371, %f365, %f357;
|
|
sub.f32 %f372, %f357, %f371;
|
|
add.f32 %f373, %f365, %f372;
|
|
add.f32 %f374, %f370, %f373;
|
|
add.f32 %f375, %f371, %f374;
|
|
sub.f32 %f376, %f371, %f375;
|
|
add.f32 %f377, %f374, %f376;
|
|
mov.f32 %f378, 0f3F317200;
|
|
mul.rn.f32 %f379, %f354, %f378;
|
|
mov.f32 %f380, 0f35BFBE8E;
|
|
mul.rn.f32 %f381, %f354, %f380;
|
|
add.f32 %f382, %f379, %f375;
|
|
sub.f32 %f383, %f379, %f382;
|
|
add.f32 %f384, %f375, %f383;
|
|
add.f32 %f385, %f377, %f384;
|
|
add.f32 %f386, %f381, %f385;
|
|
add.f32 %f387, %f382, %f386;
|
|
sub.f32 %f388, %f382, %f387;
|
|
add.f32 %f389, %f386, %f388;
|
|
mov.f32 %f390, 0f3EE66666;
|
|
mul.rn.f32 %f391, %f390, %f387;
|
|
neg.f32 %f392, %f391;
|
|
fma.rn.f32 %f393, %f390, %f387, %f392;
|
|
fma.rn.f32 %f394, %f390, %f389, %f393;
|
|
mov.f32 %f395, 0f00000000;
|
|
fma.rn.f32 %f396, %f395, %f387, %f394;
|
|
add.rn.f32 %f397, %f391, %f396;
|
|
neg.f32 %f398, %f397;
|
|
add.rn.f32 %f399, %f391, %f398;
|
|
add.rn.f32 %f400, %f399, %f396;
|
|
mov.b32 %r243, %f397;
|
|
setp.eq.s32 %p58, %r243, 1118925336;
|
|
add.s32 %r244, %r243, -1;
|
|
mov.b32 %f401, %r244;
|
|
add.f32 %f402, %f400, 0f37000000;
|
|
selp.f32 %f403, %f401, %f397, %p58;
|
|
selp.f32 %f91, %f402, %f400, %p58;
|
|
mul.f32 %f404, %f403, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f405, %f404;
|
|
mov.f32 %f406, 0fBF317200;
|
|
fma.rn.f32 %f407, %f405, %f406, %f403;
|
|
mov.f32 %f408, 0fB5BFBE8E;
|
|
fma.rn.f32 %f409, %f405, %f408, %f407;
|
|
mul.f32 %f410, %f409, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f411, %f410;
|
|
add.f32 %f412, %f405, 0f00000000;
|
|
ex2.approx.f32 %f413, %f412;
|
|
mul.f32 %f414, %f411, %f413;
|
|
setp.lt.f32 %p59, %f403, 0fC2D20000;
|
|
selp.f32 %f415, 0f00000000, %f414, %p59;
|
|
setp.gt.f32 %p60, %f403, 0f42D20000;
|
|
selp.f32 %f653, 0f7F800000, %f415, %p60;
|
|
setp.eq.f32 %p61, %f653, 0f7F800000;
|
|
@%p61 bra BB0_66;
|
|
|
|
fma.rn.f32 %f653, %f653, %f91, %f653;
|
|
|
|
BB0_66:
|
|
setp.lt.f32 %p62, %f86, 0f00000000;
|
|
setp.eq.f32 %p63, %f89, 0f3F800000;
|
|
and.pred %p1, %p62, %p63;
|
|
mov.b32 %r245, %f653;
|
|
xor.b32 %r246, %r245, -2147483648;
|
|
mov.b32 %f416, %r246;
|
|
selp.f32 %f655, %f416, %f653, %p1;
|
|
setp.eq.f32 %p64, %f86, 0f00000000;
|
|
@%p64 bra BB0_69;
|
|
bra.uni BB0_67;
|
|
|
|
BB0_69:
|
|
add.f32 %f419, %f86, %f86;
|
|
selp.f32 %f655, %f419, 0f00000000, %p63;
|
|
bra.uni BB0_70;
|
|
|
|
BB0_107:
|
|
mov.u64 %rd117, image_HDR;
|
|
cvta.global.u64 %rd112, %rd117;
|
|
mov.u32 %r289, 8;
|
|
mov.u64 %rd116, 0;
|
|
// inline asm
|
|
call (%rd111), _rt_buffer_get_64, (%rd112, %r100, %r289, %rd18, %rd19, %rd116, %rd116);
|
|
// inline asm
|
|
mov.f32 %f600, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs40, %f600;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs41, 0;
|
|
st.v4.u16 [%rd111], {%rs40, %rs40, %rs40, %rs41};
|
|
bra.uni BB0_108;
|
|
|
|
BB0_67:
|
|
setp.geu.f32 %p65, %f86, 0f00000000;
|
|
@%p65 bra BB0_70;
|
|
|
|
mov.f32 %f621, 0f3EE66666;
|
|
cvt.rzi.f32.f32 %f418, %f621;
|
|
setp.neu.f32 %p66, %f418, 0f3EE66666;
|
|
selp.f32 %f655, 0f7FFFFFFF, %f655, %p66;
|
|
|
|
BB0_70:
|
|
add.f32 %f420, %f90, 0f3EE66666;
|
|
mov.b32 %r247, %f420;
|
|
setp.lt.s32 %p68, %r247, 2139095040;
|
|
@%p68 bra BB0_75;
|
|
|
|
setp.gtu.f32 %p69, %f90, 0f7F800000;
|
|
@%p69 bra BB0_74;
|
|
bra.uni BB0_72;
|
|
|
|
BB0_74:
|
|
add.f32 %f655, %f86, 0f3EE66666;
|
|
bra.uni BB0_75;
|
|
|
|
BB0_72:
|
|
setp.neu.f32 %p70, %f90, 0f7F800000;
|
|
@%p70 bra BB0_75;
|
|
|
|
selp.f32 %f655, 0fFF800000, 0f7F800000, %p1;
|
|
|
|
BB0_75:
|
|
mov.f32 %f609, 0fB5BFBE8E;
|
|
mov.f32 %f608, 0fBF317200;
|
|
mov.f32 %f607, 0f00000000;
|
|
mov.f32 %f606, 0f35BFBE8E;
|
|
mov.f32 %f605, 0f3F317200;
|
|
mov.f32 %f604, 0f3DAAAABD;
|
|
mov.f32 %f603, 0f3C4CAF63;
|
|
mov.f32 %f602, 0f3B18F0FE;
|
|
mov.f32 %f601, 0f3EE66666;
|
|
setp.eq.f32 %p71, %f86, 0f3F800000;
|
|
selp.f32 %f102, 0f3F800000, %f655, %p71;
|
|
abs.f32 %f103, %f87;
|
|
setp.lt.f32 %p72, %f103, 0f00800000;
|
|
mul.f32 %f423, %f103, 0f4B800000;
|
|
selp.f32 %f424, 0fC3170000, 0fC2FE0000, %p72;
|
|
selp.f32 %f425, %f423, %f103, %p72;
|
|
mov.b32 %r248, %f425;
|
|
and.b32 %r249, %r248, 8388607;
|
|
or.b32 %r250, %r249, 1065353216;
|
|
mov.b32 %f426, %r250;
|
|
shr.u32 %r251, %r248, 23;
|
|
cvt.rn.f32.u32 %f427, %r251;
|
|
add.f32 %f428, %f424, %f427;
|
|
setp.gt.f32 %p73, %f426, 0f3FB504F3;
|
|
mul.f32 %f429, %f426, 0f3F000000;
|
|
add.f32 %f430, %f428, 0f3F800000;
|
|
selp.f32 %f431, %f429, %f426, %p73;
|
|
selp.f32 %f432, %f430, %f428, %p73;
|
|
add.f32 %f433, %f431, 0fBF800000;
|
|
add.f32 %f422, %f431, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f421,%f422;
|
|
// inline asm
|
|
add.f32 %f434, %f433, %f433;
|
|
mul.f32 %f435, %f421, %f434;
|
|
mul.f32 %f436, %f435, %f435;
|
|
fma.rn.f32 %f439, %f602, %f436, %f603;
|
|
fma.rn.f32 %f441, %f439, %f436, %f604;
|
|
mul.rn.f32 %f442, %f441, %f436;
|
|
mul.rn.f32 %f443, %f442, %f435;
|
|
sub.f32 %f444, %f433, %f435;
|
|
neg.f32 %f445, %f435;
|
|
add.f32 %f446, %f444, %f444;
|
|
fma.rn.f32 %f447, %f445, %f433, %f446;
|
|
mul.rn.f32 %f448, %f421, %f447;
|
|
add.f32 %f449, %f443, %f435;
|
|
sub.f32 %f450, %f435, %f449;
|
|
add.f32 %f451, %f443, %f450;
|
|
add.f32 %f452, %f448, %f451;
|
|
add.f32 %f453, %f449, %f452;
|
|
sub.f32 %f454, %f449, %f453;
|
|
add.f32 %f455, %f452, %f454;
|
|
mul.rn.f32 %f457, %f432, %f605;
|
|
mul.rn.f32 %f459, %f432, %f606;
|
|
add.f32 %f460, %f457, %f453;
|
|
sub.f32 %f461, %f457, %f460;
|
|
add.f32 %f462, %f453, %f461;
|
|
add.f32 %f463, %f455, %f462;
|
|
add.f32 %f464, %f459, %f463;
|
|
add.f32 %f465, %f460, %f464;
|
|
sub.f32 %f466, %f460, %f465;
|
|
add.f32 %f467, %f464, %f466;
|
|
mul.rn.f32 %f469, %f601, %f465;
|
|
neg.f32 %f470, %f469;
|
|
fma.rn.f32 %f471, %f601, %f465, %f470;
|
|
fma.rn.f32 %f472, %f601, %f467, %f471;
|
|
fma.rn.f32 %f474, %f607, %f465, %f472;
|
|
add.rn.f32 %f475, %f469, %f474;
|
|
neg.f32 %f476, %f475;
|
|
add.rn.f32 %f477, %f469, %f476;
|
|
add.rn.f32 %f478, %f477, %f474;
|
|
mov.b32 %r252, %f475;
|
|
setp.eq.s32 %p74, %r252, 1118925336;
|
|
add.s32 %r253, %r252, -1;
|
|
mov.b32 %f479, %r253;
|
|
add.f32 %f480, %f478, 0f37000000;
|
|
selp.f32 %f481, %f479, %f475, %p74;
|
|
selp.f32 %f104, %f480, %f478, %p74;
|
|
mul.f32 %f482, %f481, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f483, %f482;
|
|
fma.rn.f32 %f485, %f483, %f608, %f481;
|
|
fma.rn.f32 %f487, %f483, %f609, %f485;
|
|
mul.f32 %f488, %f487, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f489, %f488;
|
|
add.f32 %f490, %f483, 0f00000000;
|
|
ex2.approx.f32 %f491, %f490;
|
|
mul.f32 %f492, %f489, %f491;
|
|
setp.lt.f32 %p75, %f481, 0fC2D20000;
|
|
selp.f32 %f493, 0f00000000, %f492, %p75;
|
|
setp.gt.f32 %p76, %f481, 0f42D20000;
|
|
selp.f32 %f656, 0f7F800000, %f493, %p76;
|
|
setp.eq.f32 %p77, %f656, 0f7F800000;
|
|
@%p77 bra BB0_77;
|
|
|
|
fma.rn.f32 %f656, %f656, %f104, %f656;
|
|
|
|
BB0_77:
|
|
setp.lt.f32 %p78, %f87, 0f00000000;
|
|
and.pred %p2, %p78, %p63;
|
|
mov.b32 %r254, %f656;
|
|
xor.b32 %r255, %r254, -2147483648;
|
|
mov.b32 %f494, %r255;
|
|
selp.f32 %f658, %f494, %f656, %p2;
|
|
setp.eq.f32 %p80, %f87, 0f00000000;
|
|
@%p80 bra BB0_80;
|
|
bra.uni BB0_78;
|
|
|
|
BB0_80:
|
|
add.f32 %f497, %f87, %f87;
|
|
selp.f32 %f658, %f497, 0f00000000, %p63;
|
|
bra.uni BB0_81;
|
|
|
|
BB0_78:
|
|
setp.geu.f32 %p81, %f87, 0f00000000;
|
|
@%p81 bra BB0_81;
|
|
|
|
mov.f32 %f620, 0f3EE66666;
|
|
cvt.rzi.f32.f32 %f496, %f620;
|
|
setp.neu.f32 %p82, %f496, 0f3EE66666;
|
|
selp.f32 %f658, 0f7FFFFFFF, %f658, %p82;
|
|
|
|
BB0_81:
|
|
add.f32 %f498, %f103, 0f3EE66666;
|
|
mov.b32 %r256, %f498;
|
|
setp.lt.s32 %p84, %r256, 2139095040;
|
|
@%p84 bra BB0_86;
|
|
|
|
setp.gtu.f32 %p85, %f103, 0f7F800000;
|
|
@%p85 bra BB0_85;
|
|
bra.uni BB0_83;
|
|
|
|
BB0_85:
|
|
add.f32 %f658, %f87, 0f3EE66666;
|
|
bra.uni BB0_86;
|
|
|
|
BB0_83:
|
|
setp.neu.f32 %p86, %f103, 0f7F800000;
|
|
@%p86 bra BB0_86;
|
|
|
|
selp.f32 %f658, 0fFF800000, 0f7F800000, %p2;
|
|
|
|
BB0_86:
|
|
mov.f32 %f618, 0fB5BFBE8E;
|
|
mov.f32 %f617, 0fBF317200;
|
|
mov.f32 %f616, 0f00000000;
|
|
mov.f32 %f615, 0f35BFBE8E;
|
|
mov.f32 %f614, 0f3F317200;
|
|
mov.f32 %f613, 0f3DAAAABD;
|
|
mov.f32 %f612, 0f3C4CAF63;
|
|
mov.f32 %f611, 0f3B18F0FE;
|
|
mov.f32 %f610, 0f3EE66666;
|
|
setp.eq.f32 %p87, %f87, 0f3F800000;
|
|
selp.f32 %f115, 0f3F800000, %f658, %p87;
|
|
abs.f32 %f116, %f88;
|
|
setp.lt.f32 %p88, %f116, 0f00800000;
|
|
mul.f32 %f501, %f116, 0f4B800000;
|
|
selp.f32 %f502, 0fC3170000, 0fC2FE0000, %p88;
|
|
selp.f32 %f503, %f501, %f116, %p88;
|
|
mov.b32 %r257, %f503;
|
|
and.b32 %r258, %r257, 8388607;
|
|
or.b32 %r259, %r258, 1065353216;
|
|
mov.b32 %f504, %r259;
|
|
shr.u32 %r260, %r257, 23;
|
|
cvt.rn.f32.u32 %f505, %r260;
|
|
add.f32 %f506, %f502, %f505;
|
|
setp.gt.f32 %p89, %f504, 0f3FB504F3;
|
|
mul.f32 %f507, %f504, 0f3F000000;
|
|
add.f32 %f508, %f506, 0f3F800000;
|
|
selp.f32 %f509, %f507, %f504, %p89;
|
|
selp.f32 %f510, %f508, %f506, %p89;
|
|
add.f32 %f511, %f509, 0fBF800000;
|
|
add.f32 %f500, %f509, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f499,%f500;
|
|
// inline asm
|
|
add.f32 %f512, %f511, %f511;
|
|
mul.f32 %f513, %f499, %f512;
|
|
mul.f32 %f514, %f513, %f513;
|
|
fma.rn.f32 %f517, %f611, %f514, %f612;
|
|
fma.rn.f32 %f519, %f517, %f514, %f613;
|
|
mul.rn.f32 %f520, %f519, %f514;
|
|
mul.rn.f32 %f521, %f520, %f513;
|
|
sub.f32 %f522, %f511, %f513;
|
|
neg.f32 %f523, %f513;
|
|
add.f32 %f524, %f522, %f522;
|
|
fma.rn.f32 %f525, %f523, %f511, %f524;
|
|
mul.rn.f32 %f526, %f499, %f525;
|
|
add.f32 %f527, %f521, %f513;
|
|
sub.f32 %f528, %f513, %f527;
|
|
add.f32 %f529, %f521, %f528;
|
|
add.f32 %f530, %f526, %f529;
|
|
add.f32 %f531, %f527, %f530;
|
|
sub.f32 %f532, %f527, %f531;
|
|
add.f32 %f533, %f530, %f532;
|
|
mul.rn.f32 %f535, %f510, %f614;
|
|
mul.rn.f32 %f537, %f510, %f615;
|
|
add.f32 %f538, %f535, %f531;
|
|
sub.f32 %f539, %f535, %f538;
|
|
add.f32 %f540, %f531, %f539;
|
|
add.f32 %f541, %f533, %f540;
|
|
add.f32 %f542, %f537, %f541;
|
|
add.f32 %f543, %f538, %f542;
|
|
sub.f32 %f544, %f538, %f543;
|
|
add.f32 %f545, %f542, %f544;
|
|
mul.rn.f32 %f547, %f610, %f543;
|
|
neg.f32 %f548, %f547;
|
|
fma.rn.f32 %f549, %f610, %f543, %f548;
|
|
fma.rn.f32 %f550, %f610, %f545, %f549;
|
|
fma.rn.f32 %f552, %f616, %f543, %f550;
|
|
add.rn.f32 %f553, %f547, %f552;
|
|
neg.f32 %f554, %f553;
|
|
add.rn.f32 %f555, %f547, %f554;
|
|
add.rn.f32 %f556, %f555, %f552;
|
|
mov.b32 %r261, %f553;
|
|
setp.eq.s32 %p90, %r261, 1118925336;
|
|
add.s32 %r262, %r261, -1;
|
|
mov.b32 %f557, %r262;
|
|
add.f32 %f558, %f556, 0f37000000;
|
|
selp.f32 %f559, %f557, %f553, %p90;
|
|
selp.f32 %f117, %f558, %f556, %p90;
|
|
mul.f32 %f560, %f559, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f561, %f560;
|
|
fma.rn.f32 %f563, %f561, %f617, %f559;
|
|
fma.rn.f32 %f565, %f561, %f618, %f563;
|
|
mul.f32 %f566, %f565, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f567, %f566;
|
|
add.f32 %f568, %f561, 0f00000000;
|
|
ex2.approx.f32 %f569, %f568;
|
|
mul.f32 %f570, %f567, %f569;
|
|
setp.lt.f32 %p91, %f559, 0fC2D20000;
|
|
selp.f32 %f571, 0f00000000, %f570, %p91;
|
|
setp.gt.f32 %p92, %f559, 0f42D20000;
|
|
selp.f32 %f659, 0f7F800000, %f571, %p92;
|
|
setp.eq.f32 %p93, %f659, 0f7F800000;
|
|
@%p93 bra BB0_88;
|
|
|
|
fma.rn.f32 %f659, %f659, %f117, %f659;
|
|
|
|
BB0_88:
|
|
setp.lt.f32 %p94, %f88, 0f00000000;
|
|
and.pred %p3, %p94, %p63;
|
|
mov.b32 %r263, %f659;
|
|
xor.b32 %r264, %r263, -2147483648;
|
|
mov.b32 %f572, %r264;
|
|
selp.f32 %f661, %f572, %f659, %p3;
|
|
setp.eq.f32 %p96, %f88, 0f00000000;
|
|
@%p96 bra BB0_91;
|
|
bra.uni BB0_89;
|
|
|
|
BB0_91:
|
|
add.f32 %f575, %f88, %f88;
|
|
selp.f32 %f661, %f575, 0f00000000, %p63;
|
|
bra.uni BB0_92;
|
|
|
|
BB0_89:
|
|
setp.geu.f32 %p97, %f88, 0f00000000;
|
|
@%p97 bra BB0_92;
|
|
|
|
mov.f32 %f619, 0f3EE66666;
|
|
cvt.rzi.f32.f32 %f574, %f619;
|
|
setp.neu.f32 %p98, %f574, 0f3EE66666;
|
|
selp.f32 %f661, 0f7FFFFFFF, %f661, %p98;
|
|
|
|
BB0_92:
|
|
add.f32 %f576, %f116, 0f3EE66666;
|
|
mov.b32 %r265, %f576;
|
|
setp.lt.s32 %p100, %r265, 2139095040;
|
|
@%p100 bra BB0_97;
|
|
|
|
setp.gtu.f32 %p101, %f116, 0f7F800000;
|
|
@%p101 bra BB0_96;
|
|
bra.uni BB0_94;
|
|
|
|
BB0_96:
|
|
add.f32 %f661, %f88, 0f3EE66666;
|
|
bra.uni BB0_97;
|
|
|
|
BB0_94:
|
|
setp.neu.f32 %p102, %f116, 0f7F800000;
|
|
@%p102 bra BB0_97;
|
|
|
|
selp.f32 %f661, 0fFF800000, 0f7F800000, %p3;
|
|
|
|
BB0_97:
|
|
mov.u32 %r291, 4;
|
|
mov.u64 %rd118, 0;
|
|
mov.u32 %r290, 2;
|
|
setp.eq.f32 %p103, %f88, 0f3F800000;
|
|
selp.f32 %f577, 0f3F800000, %f661, %p103;
|
|
cvt.u64.u32 %rd67, %r3;
|
|
cvt.u64.u32 %rd66, %r2;
|
|
mov.u64 %rd70, image;
|
|
cvta.global.u64 %rd65, %rd70;
|
|
// inline asm
|
|
call (%rd64), _rt_buffer_get_64, (%rd65, %r290, %r291, %rd66, %rd67, %rd118, %rd118);
|
|
// inline asm
|
|
cvt.sat.f32.f32 %f578, %f577;
|
|
mul.f32 %f579, %f578, 0f437FFD71;
|
|
cvt.rzi.u32.f32 %r268, %f579;
|
|
cvt.sat.f32.f32 %f580, %f115;
|
|
mul.f32 %f581, %f580, 0f437FFD71;
|
|
cvt.rzi.u32.f32 %r269, %f581;
|
|
cvt.sat.f32.f32 %f582, %f102;
|
|
mul.f32 %f583, %f582, 0f437FFD71;
|
|
cvt.rzi.u32.f32 %r270, %f583;
|
|
cvt.u16.u32 %rs10, %r268;
|
|
cvt.u16.u32 %rs11, %r270;
|
|
cvt.u16.u32 %rs12, %r269;
|
|
mov.u16 %rs13, 255;
|
|
st.v4.u8 [%rd64], {%rs10, %rs12, %rs11, %rs13};
|
|
ld.global.u32 %r324, [imageEnabled];
|
|
|
|
BB0_98:
|
|
and.b32 %r271, %r324, 4;
|
|
setp.eq.s32 %p104, %r271, 0;
|
|
@%p104 bra BB0_108;
|
|
|
|
ld.global.u32 %r272, [additive];
|
|
setp.eq.s32 %p105, %r272, 0;
|
|
cvt.u64.u32 %rd16, %r2;
|
|
cvt.u64.u32 %rd17, %r3;
|
|
mov.f32 %f584, 0f3F800000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs14, %f584;}
|
|
|
|
// inline asm
|
|
@%p105 bra BB0_101;
|
|
|
|
mov.u64 %rd119, 0;
|
|
mov.u32 %r292, 2;
|
|
mov.u64 %rd83, image_HDR;
|
|
cvta.global.u64 %rd72, %rd83;
|
|
mov.u32 %r276, 8;
|
|
// inline asm
|
|
call (%rd71), _rt_buffer_get_64, (%rd72, %r292, %r276, %rd16, %rd17, %rd119, %rd119);
|
|
// inline asm
|
|
ld.v4.u16 {%rs21, %rs22, %rs23, %rs24}, [%rd71];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f585, %rs21;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f586, %rs22;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f587, %rs23;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd77), _rt_buffer_get_64, (%rd72, %r292, %r276, %rd16, %rd17, %rd119, %rd119);
|
|
// inline asm
|
|
add.f32 %f588, %f86, %f585;
|
|
add.f32 %f589, %f87, %f586;
|
|
add.f32 %f590, %f88, %f587;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs20, %f590;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs19, %f589;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs18, %f588;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd77], {%rs18, %rs19, %rs20, %rs14};
|
|
bra.uni BB0_108;
|
|
|
|
BB0_101:
|
|
mov.u64 %rd120, 0;
|
|
mov.u32 %r293, 2;
|
|
mov.u64 %rd90, image_HDR;
|
|
cvta.global.u64 %rd85, %rd90;
|
|
mov.u32 %r278, 8;
|
|
// inline asm
|
|
call (%rd84), _rt_buffer_get_64, (%rd85, %r293, %r278, %rd16, %rd17, %rd120, %rd120);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs27, %f88;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs26, %f87;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs25, %f86;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd84], {%rs25, %rs26, %rs27, %rs14};
|
|
|
|
BB0_108:
|
|
ret;
|
|
}
|
|
|
|
|