1986 lines
60 KiB
Plaintext
1986 lines
60 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 image_Mask[1];
|
|
.global .align 1 .b8 image_Dir[1];
|
|
.global .align 1 .b8 uvpos[1];
|
|
.global .align 1 .b8 uvnormal[1];
|
|
.global .align 4 .u32 ignoreNormal;
|
|
.global .align 1 .b8 localLights[1];
|
|
.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_typeinfo12ignoreNormalE[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_typename12ignoreNormalE[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_typeenum12ignoreNormalE = 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_semantic12ignoreNormalE[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_annotation12ignoreNormalE[1];
|
|
|
|
.visible .entry _Z6oxMainv(
|
|
|
|
)
|
|
{
|
|
.local .align 4 .b8 __local_depot0[12];
|
|
.reg .b64 %SP;
|
|
.reg .b64 %SPL;
|
|
.reg .pred %p<134>;
|
|
.reg .b16 %rs<63>;
|
|
.reg .f32 %f<1333>;
|
|
.reg .b32 %r<207>;
|
|
.reg .b64 %rd<126>;
|
|
|
|
|
|
mov.u64 %rd125, __local_depot0;
|
|
cvta.local.u64 %SP, %rd125;
|
|
ld.global.v2.u32 {%r30, %r31}, [pixelID];
|
|
cvt.u64.u32 %rd10, %r30;
|
|
cvt.u64.u32 %rd11, %r31;
|
|
mov.u64 %rd14, uvnormal;
|
|
cvta.global.u64 %rd9, %rd14;
|
|
mov.u32 %r28, 2;
|
|
mov.u32 %r29, 4;
|
|
mov.u64 %rd13, 0;
|
|
// inline asm
|
|
call (%rd8), _rt_buffer_get_64, (%rd9, %r28, %r29, %rd10, %rd11, %rd13, %rd13);
|
|
// inline asm
|
|
ld.u32 %r1, [%rd8];
|
|
shr.u32 %r34, %r1, 16;
|
|
cvt.u16.u32 %rs1, %r34;
|
|
and.b16 %rs5, %rs1, 255;
|
|
cvt.u16.u32 %rs6, %r1;
|
|
or.b16 %rs7, %rs6, %rs5;
|
|
setp.eq.s16 %p8, %rs7, 0;
|
|
mov.f32 %f1258, 0f00000000;
|
|
mov.f32 %f1259, %f1258;
|
|
mov.f32 %f1260, %f1258;
|
|
@%p8 bra BB0_2;
|
|
|
|
ld.u8 %rs8, [%rd8+1];
|
|
and.b16 %rs10, %rs6, 255;
|
|
cvt.rn.f32.u16 %f242, %rs10;
|
|
div.rn.f32 %f243, %f242, 0f437F0000;
|
|
fma.rn.f32 %f244, %f243, 0f40000000, 0fBF800000;
|
|
cvt.rn.f32.u16 %f245, %rs8;
|
|
div.rn.f32 %f246, %f245, 0f437F0000;
|
|
fma.rn.f32 %f247, %f246, 0f40000000, 0fBF800000;
|
|
cvt.rn.f32.u16 %f248, %rs5;
|
|
div.rn.f32 %f249, %f248, 0f437F0000;
|
|
fma.rn.f32 %f250, %f249, 0f40000000, 0fBF800000;
|
|
mul.f32 %f251, %f247, %f247;
|
|
fma.rn.f32 %f252, %f244, %f244, %f251;
|
|
fma.rn.f32 %f253, %f250, %f250, %f252;
|
|
sqrt.rn.f32 %f254, %f253;
|
|
rcp.rn.f32 %f255, %f254;
|
|
mul.f32 %f1258, %f244, %f255;
|
|
mul.f32 %f1259, %f247, %f255;
|
|
mul.f32 %f1260, %f250, %f255;
|
|
|
|
BB0_2:
|
|
ld.global.v2.u32 {%r35, %r36}, [pixelID];
|
|
ld.global.v2.u32 {%r38, %r39}, [tileInfo];
|
|
add.s32 %r2, %r35, %r38;
|
|
add.s32 %r3, %r36, %r39;
|
|
setp.eq.f32 %p9, %f1259, 0f00000000;
|
|
setp.eq.f32 %p10, %f1258, 0f00000000;
|
|
and.pred %p11, %p10, %p9;
|
|
setp.eq.f32 %p12, %f1260, 0f00000000;
|
|
and.pred %p13, %p11, %p12;
|
|
@%p13 bra BB0_103;
|
|
bra.uni BB0_3;
|
|
|
|
BB0_103:
|
|
ld.global.u32 %r206, [imageEnabled];
|
|
and.b32 %r179, %r206, 1;
|
|
setp.eq.b32 %p128, %r179, 1;
|
|
@!%p128 bra BB0_105;
|
|
bra.uni BB0_104;
|
|
|
|
BB0_104:
|
|
cvt.u64.u32 %rd86, %r2;
|
|
cvt.u64.u32 %rd87, %r3;
|
|
mov.u64 %rd90, image;
|
|
cvta.global.u64 %rd85, %rd90;
|
|
// inline asm
|
|
call (%rd84), _rt_buffer_get_64, (%rd85, %r28, %r29, %rd86, %rd87, %rd13, %rd13);
|
|
// inline asm
|
|
mov.u16 %rs43, 0;
|
|
st.v4.u8 [%rd84], {%rs43, %rs43, %rs43, %rs43};
|
|
ld.global.u32 %r206, [imageEnabled];
|
|
|
|
BB0_105:
|
|
and.b32 %r182, %r206, 8;
|
|
setp.eq.s32 %p129, %r182, 0;
|
|
@%p129 bra BB0_107;
|
|
|
|
cvt.u64.u32 %rd93, %r2;
|
|
cvt.u64.u32 %rd94, %r3;
|
|
mov.u64 %rd97, image_Mask;
|
|
cvta.global.u64 %rd92, %rd97;
|
|
// inline asm
|
|
call (%rd91), _rt_buffer_get_64, (%rd92, %r28, %r28, %rd93, %rd94, %rd13, %rd13);
|
|
// inline asm
|
|
mov.f32 %f1247, 0f00000000;
|
|
cvt.rzi.u32.f32 %r185, %f1247;
|
|
cvt.u16.u32 %rs44, %r185;
|
|
mov.u16 %rs45, 0;
|
|
st.v2.u8 [%rd91], {%rs44, %rs45};
|
|
ld.global.u32 %r206, [imageEnabled];
|
|
|
|
BB0_107:
|
|
and.b32 %r186, %r206, 4;
|
|
setp.eq.s32 %p130, %r186, 0;
|
|
@%p130 bra BB0_111;
|
|
|
|
ld.global.u32 %r187, [additive];
|
|
setp.eq.s32 %p131, %r187, 0;
|
|
cvt.u64.u32 %rd6, %r2;
|
|
cvt.u64.u32 %rd7, %r3;
|
|
@%p131 bra BB0_110;
|
|
|
|
mov.u64 %rd110, image_HDR;
|
|
cvta.global.u64 %rd99, %rd110;
|
|
mov.u32 %r191, 8;
|
|
// inline asm
|
|
call (%rd98), _rt_buffer_get_64, (%rd99, %r28, %r191, %rd6, %rd7, %rd13, %rd13);
|
|
// inline asm
|
|
ld.v4.u16 {%rs52, %rs53, %rs54, %rs55}, [%rd98];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1248, %rs52;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1249, %rs53;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1250, %rs54;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd104), _rt_buffer_get_64, (%rd99, %r28, %r191, %rd6, %rd7, %rd13, %rd13);
|
|
// inline asm
|
|
add.f32 %f1251, %f1248, 0f00000000;
|
|
add.f32 %f1252, %f1249, 0f00000000;
|
|
add.f32 %f1253, %f1250, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs51, %f1253;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs50, %f1252;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs49, %f1251;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs56, 0;
|
|
st.v4.u16 [%rd104], {%rs49, %rs50, %rs51, %rs56};
|
|
bra.uni BB0_111;
|
|
|
|
BB0_3:
|
|
ld.global.v2.u32 {%r47, %r48}, [pixelID];
|
|
cvt.u64.u32 %rd17, %r47;
|
|
cvt.u64.u32 %rd18, %r48;
|
|
mov.u64 %rd26, uvpos;
|
|
cvta.global.u64 %rd16, %rd26;
|
|
mov.u32 %r44, 12;
|
|
// inline asm
|
|
call (%rd15), _rt_buffer_get_64, (%rd16, %r28, %r44, %rd17, %rd18, %rd13, %rd13);
|
|
// inline asm
|
|
ld.f32 %f9, [%rd15+8];
|
|
ld.f32 %f8, [%rd15+4];
|
|
ld.f32 %f7, [%rd15];
|
|
mul.f32 %f263, %f7, 0f3456BF95;
|
|
mul.f32 %f264, %f8, 0f3456BF95;
|
|
mul.f32 %f265, %f9, 0f3456BF95;
|
|
abs.f32 %f266, %f1258;
|
|
div.rn.f32 %f267, %f263, %f266;
|
|
abs.f32 %f268, %f1259;
|
|
div.rn.f32 %f269, %f264, %f268;
|
|
abs.f32 %f270, %f1260;
|
|
div.rn.f32 %f271, %f265, %f270;
|
|
abs.f32 %f272, %f267;
|
|
abs.f32 %f273, %f269;
|
|
abs.f32 %f274, %f271;
|
|
mov.f32 %f275, 0f38D1B717;
|
|
max.f32 %f276, %f272, %f275;
|
|
max.f32 %f277, %f273, %f275;
|
|
max.f32 %f278, %f274, %f275;
|
|
fma.rn.f32 %f10, %f1258, %f276, %f7;
|
|
fma.rn.f32 %f11, %f1259, %f277, %f8;
|
|
fma.rn.f32 %f12, %f1260, %f278, %f9;
|
|
ld.global.v2.u32 {%r51, %r52}, [pixelID];
|
|
cvt.rn.f32.u32 %f13, %r51;
|
|
cvt.rn.f32.u32 %f14, %r52;
|
|
mov.u64 %rd27, localLights;
|
|
cvta.global.u64 %rd25, %rd27;
|
|
mov.u32 %r45, 1;
|
|
mov.u32 %r46, 96;
|
|
// inline asm
|
|
call (%rd21, %rd22, %rd23, %rd24), _rt_buffer_get_size_64, (%rd25, %r45, %r46);
|
|
// inline asm
|
|
cvt.u32.u64 %r4, %rd21;
|
|
setp.eq.s32 %p14, %r4, 0;
|
|
mov.f32 %f1261, 0f00000000;
|
|
mov.f32 %f22, %f1261;
|
|
mov.f32 %f23, %f1261;
|
|
mov.f32 %f24, %f1261;
|
|
mov.f32 %f1265, %f1261;
|
|
mov.f32 %f1266, %f1261;
|
|
mov.f32 %f1267, %f1261;
|
|
@%p14 bra BB0_49;
|
|
|
|
mov.f32 %f286, 0f40000000;
|
|
cvt.rzi.f32.f32 %f287, %f286;
|
|
add.f32 %f288, %f287, %f287;
|
|
mov.f32 %f289, 0f40800000;
|
|
sub.f32 %f290, %f289, %f288;
|
|
abs.f32 %f15, %f290;
|
|
mul.f32 %f16, %f10, 0f3456BF95;
|
|
mul.f32 %f17, %f11, 0f3456BF95;
|
|
mul.f32 %f18, %f12, 0f3456BF95;
|
|
mul.f32 %f19, %f13, 0f3DD32618;
|
|
mul.f32 %f20, %f14, 0f3DD2F1AA;
|
|
mov.f32 %f285, 0f00000000;
|
|
mov.u32 %r198, 0;
|
|
abs.f32 %f476, %f16;
|
|
abs.f32 %f477, %f17;
|
|
max.f32 %f478, %f476, %f477;
|
|
abs.f32 %f479, %f18;
|
|
max.f32 %f480, %f478, %f479;
|
|
mov.f32 %f1261, %f285;
|
|
mov.f32 %f22, %f285;
|
|
mov.f32 %f23, %f285;
|
|
mov.f32 %f24, %f285;
|
|
mov.f32 %f1265, %f285;
|
|
mov.f32 %f1266, %f285;
|
|
mov.f32 %f1267, %f285;
|
|
|
|
BB0_5:
|
|
cvt.u64.u32 %rd30, %r198;
|
|
// inline asm
|
|
call (%rd28), _rt_buffer_get_64, (%rd25, %r45, %r46, %rd30, %rd13, %rd13, %rd13);
|
|
// inline asm
|
|
ld.v4.f32 {%f293, %f294, %f295, %f296}, [%rd28+80];
|
|
ld.v4.f32 {%f297, %f298, %f299, %f300}, [%rd28+64];
|
|
ld.v4.f32 {%f301, %f302, %f303, %f304}, [%rd28+48];
|
|
ld.v4.f32 {%f305, %f1276, %f1277, %f308}, [%rd28+32];
|
|
ld.v4.f32 {%f309, %f310, %f311, %f312}, [%rd28+16];
|
|
ld.v4.f32 {%f313, %f314, %f315, %f316}, [%rd28];
|
|
mov.b32 %r6, %f296;
|
|
sub.f32 %f52, %f314, %f7;
|
|
sub.f32 %f54, %f315, %f8;
|
|
sub.f32 %f56, %f316, %f9;
|
|
mul.f32 %f318, %f54, %f54;
|
|
fma.rn.f32 %f319, %f52, %f52, %f318;
|
|
fma.rn.f32 %f320, %f56, %f56, %f319;
|
|
sqrt.rn.f32 %f57, %f320;
|
|
rcp.rn.f32 %f321, %f57;
|
|
mul.f32 %f58, %f52, %f321;
|
|
mul.f32 %f59, %f54, %f321;
|
|
mul.f32 %f60, %f56, %f321;
|
|
mul.f32 %f61, %f57, %f312;
|
|
abs.f32 %f62, %f61;
|
|
setp.lt.f32 %p15, %f62, 0f00800000;
|
|
mul.f32 %f322, %f62, 0f4B800000;
|
|
selp.f32 %f323, 0fC3170000, 0fC2FE0000, %p15;
|
|
selp.f32 %f324, %f322, %f62, %p15;
|
|
mov.b32 %r58, %f324;
|
|
and.b32 %r59, %r58, 8388607;
|
|
or.b32 %r60, %r59, 1065353216;
|
|
mov.b32 %f325, %r60;
|
|
shr.u32 %r61, %r58, 23;
|
|
cvt.rn.f32.u32 %f326, %r61;
|
|
add.f32 %f327, %f323, %f326;
|
|
setp.gt.f32 %p16, %f325, 0f3FB504F3;
|
|
mul.f32 %f328, %f325, 0f3F000000;
|
|
add.f32 %f329, %f327, 0f3F800000;
|
|
selp.f32 %f330, %f328, %f325, %p16;
|
|
selp.f32 %f331, %f329, %f327, %p16;
|
|
add.f32 %f332, %f330, 0fBF800000;
|
|
add.f32 %f292, %f330, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f291,%f292;
|
|
// inline asm
|
|
add.f32 %f333, %f332, %f332;
|
|
mul.f32 %f334, %f291, %f333;
|
|
mul.f32 %f335, %f334, %f334;
|
|
mov.f32 %f336, 0f3C4CAF63;
|
|
mov.f32 %f337, 0f3B18F0FE;
|
|
fma.rn.f32 %f338, %f337, %f335, %f336;
|
|
mov.f32 %f339, 0f3DAAAABD;
|
|
fma.rn.f32 %f340, %f338, %f335, %f339;
|
|
mul.rn.f32 %f341, %f340, %f335;
|
|
mul.rn.f32 %f342, %f341, %f334;
|
|
sub.f32 %f343, %f332, %f334;
|
|
neg.f32 %f344, %f334;
|
|
add.f32 %f345, %f343, %f343;
|
|
fma.rn.f32 %f346, %f344, %f332, %f345;
|
|
mul.rn.f32 %f347, %f291, %f346;
|
|
add.f32 %f348, %f342, %f334;
|
|
sub.f32 %f349, %f334, %f348;
|
|
add.f32 %f350, %f342, %f349;
|
|
add.f32 %f351, %f347, %f350;
|
|
add.f32 %f352, %f348, %f351;
|
|
sub.f32 %f353, %f348, %f352;
|
|
add.f32 %f354, %f351, %f353;
|
|
mov.f32 %f355, 0f3F317200;
|
|
mul.rn.f32 %f356, %f331, %f355;
|
|
mov.f32 %f357, 0f35BFBE8E;
|
|
mul.rn.f32 %f358, %f331, %f357;
|
|
add.f32 %f359, %f356, %f352;
|
|
sub.f32 %f360, %f356, %f359;
|
|
add.f32 %f361, %f352, %f360;
|
|
add.f32 %f362, %f354, %f361;
|
|
add.f32 %f363, %f358, %f362;
|
|
add.f32 %f364, %f359, %f363;
|
|
sub.f32 %f365, %f359, %f364;
|
|
add.f32 %f366, %f363, %f365;
|
|
mul.rn.f32 %f63, %f289, %f364;
|
|
neg.f32 %f368, %f63;
|
|
fma.rn.f32 %f369, %f289, %f364, %f368;
|
|
fma.rn.f32 %f370, %f289, %f366, %f369;
|
|
fma.rn.f32 %f64, %f285, %f364, %f370;
|
|
add.rn.f32 %f65, %f63, %f64;
|
|
mov.b32 %r62, %f65;
|
|
setp.eq.s32 %p1, %r62, 1118925336;
|
|
add.s32 %r63, %r62, -1;
|
|
mov.b32 %f372, %r63;
|
|
selp.f32 %f373, %f372, %f65, %p1;
|
|
mul.f32 %f374, %f373, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f375, %f374;
|
|
mov.f32 %f376, 0fBF317200;
|
|
fma.rn.f32 %f377, %f375, %f376, %f373;
|
|
mov.f32 %f378, 0fB5BFBE8E;
|
|
fma.rn.f32 %f379, %f375, %f378, %f377;
|
|
mul.f32 %f380, %f379, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f381, %f380;
|
|
add.f32 %f382, %f375, 0f00000000;
|
|
ex2.approx.f32 %f383, %f382;
|
|
mul.f32 %f384, %f381, %f383;
|
|
setp.lt.f32 %p17, %f373, 0fC2D20000;
|
|
selp.f32 %f385, 0f00000000, %f384, %p17;
|
|
setp.gt.f32 %p18, %f373, 0f42D20000;
|
|
selp.f32 %f1268, 0f7F800000, %f385, %p18;
|
|
setp.eq.f32 %p19, %f1268, 0f7F800000;
|
|
@%p19 bra BB0_7;
|
|
|
|
neg.f32 %f386, %f65;
|
|
add.rn.f32 %f387, %f63, %f386;
|
|
add.rn.f32 %f388, %f387, %f64;
|
|
add.f32 %f389, %f388, 0f37000000;
|
|
selp.f32 %f390, %f389, %f388, %p1;
|
|
fma.rn.f32 %f1268, %f1268, %f390, %f1268;
|
|
|
|
BB0_7:
|
|
setp.lt.f32 %p20, %f61, 0f00000000;
|
|
setp.eq.f32 %p21, %f15, 0f3F800000;
|
|
and.pred %p2, %p20, %p21;
|
|
mov.b32 %r64, %f1268;
|
|
xor.b32 %r65, %r64, -2147483648;
|
|
mov.b32 %f391, %r65;
|
|
selp.f32 %f1270, %f391, %f1268, %p2;
|
|
setp.eq.f32 %p22, %f61, 0f00000000;
|
|
@%p22 bra BB0_10;
|
|
bra.uni BB0_8;
|
|
|
|
BB0_10:
|
|
add.f32 %f394, %f61, %f61;
|
|
selp.f32 %f1270, %f394, 0f00000000, %p21;
|
|
bra.uni BB0_11;
|
|
|
|
BB0_8:
|
|
setp.geu.f32 %p23, %f61, 0f00000000;
|
|
@%p23 bra BB0_11;
|
|
|
|
cvt.rzi.f32.f32 %f393, %f289;
|
|
setp.neu.f32 %p24, %f393, 0f40800000;
|
|
selp.f32 %f1270, 0f7FFFFFFF, %f1270, %p24;
|
|
|
|
BB0_11:
|
|
add.f32 %f395, %f62, 0f40800000;
|
|
mov.b32 %r66, %f395;
|
|
setp.lt.s32 %p26, %r66, 2139095040;
|
|
@%p26 bra BB0_16;
|
|
|
|
setp.gtu.f32 %p27, %f62, 0f7F800000;
|
|
@%p27 bra BB0_15;
|
|
bra.uni BB0_13;
|
|
|
|
BB0_15:
|
|
add.f32 %f1270, %f61, 0f40800000;
|
|
bra.uni BB0_16;
|
|
|
|
BB0_13:
|
|
setp.neu.f32 %p28, %f62, 0f7F800000;
|
|
@%p28 bra BB0_16;
|
|
|
|
selp.f32 %f1270, 0fFF800000, 0f7F800000, %p2;
|
|
|
|
BB0_16:
|
|
mul.f32 %f396, %f57, %f310;
|
|
mov.f32 %f1300, 0f3F800000;
|
|
sub.f32 %f398, %f1300, %f1270;
|
|
setp.eq.f32 %p29, %f61, 0f3F800000;
|
|
selp.f32 %f399, 0f00000000, %f398, %p29;
|
|
cvt.sat.f32.f32 %f400, %f399;
|
|
fma.rn.f32 %f401, %f396, %f396, %f311;
|
|
div.rn.f32 %f91, %f400, %f401;
|
|
mul.f32 %f402, %f1259, %f59;
|
|
fma.rn.f32 %f403, %f1258, %f58, %f402;
|
|
fma.rn.f32 %f404, %f1260, %f60, %f403;
|
|
ld.global.u32 %r67, [ignoreNormal];
|
|
setp.eq.s32 %p30, %r67, 0;
|
|
selp.f32 %f405, %f404, 0f3F800000, %p30;
|
|
cvt.sat.f32.f32 %f92, %f405;
|
|
setp.eq.f32 %p31, %f313, 0f3F800000;
|
|
@%p31 bra BB0_23;
|
|
bra.uni BB0_17;
|
|
|
|
BB0_23:
|
|
setp.leu.f32 %p35, %f308, 0f00000000;
|
|
@%p35 bra BB0_19;
|
|
|
|
setp.geu.f32 %p36, %f304, 0f00000000;
|
|
@%p36 bra BB0_25;
|
|
|
|
sub.f32 %f1257, %f316, %f9;
|
|
sub.f32 %f1256, %f314, %f7;
|
|
sub.f32 %f1255, %f315, %f8;
|
|
mul.f32 %f428, %f294, %f1255;
|
|
fma.rn.f32 %f429, %f293, %f1256, %f428;
|
|
fma.rn.f32 %f430, %f295, %f1257, %f429;
|
|
rcp.rn.f32 %f431, %f430;
|
|
mul.f32 %f1271, %f1256, %f431;
|
|
mul.f32 %f1272, %f1255, %f431;
|
|
mul.f32 %f1273, %f1257, %f431;
|
|
neg.f32 %f304, %f304;
|
|
bra.uni BB0_27;
|
|
|
|
BB0_17:
|
|
setp.eq.f32 %p32, %f313, 0f40000000;
|
|
@%p32 bra BB0_21;
|
|
bra.uni BB0_18;
|
|
|
|
BB0_21:
|
|
setp.leu.f32 %p34, %f308, 0f00000000;
|
|
@%p34 bra BB0_19;
|
|
|
|
mul.f32 %f422, %f302, %f59;
|
|
fma.rn.f32 %f423, %f301, %f58, %f422;
|
|
mul.f32 %f424, %f298, %f59;
|
|
fma.rn.f32 %f425, %f297, %f58, %f424;
|
|
mul.f32 %f426, %f294, %f59;
|
|
fma.rn.f32 %f427, %f293, %f58, %f426;
|
|
fma.rn.f32 %f419, %f303, %f60, %f423;
|
|
fma.rn.f32 %f420, %f299, %f60, %f425;
|
|
fma.rn.f32 %f421, %f295, %f60, %f427;
|
|
cvt.rzi.s32.f32 %r68, %f308;
|
|
mov.u32 %r69, 6;
|
|
mov.u32 %r70, 0;
|
|
// inline asm
|
|
call (%f415, %f416, %f417, %f418), _rt_texture_get_base_id, (%r68, %r69, %f419, %f420, %f421, %r70);
|
|
// inline asm
|
|
mul.f32 %f1275, %f305, %f415;
|
|
mul.f32 %f1276, %f1276, %f416;
|
|
mul.f32 %f1277, %f1277, %f417;
|
|
bra.uni BB0_28;
|
|
|
|
BB0_18:
|
|
setp.neu.f32 %p33, %f313, 0f40800000;
|
|
@%p33 bra BB0_19;
|
|
|
|
mul.f32 %f406, %f293, %f58;
|
|
mul.f32 %f407, %f294, %f59;
|
|
neg.f32 %f408, %f407;
|
|
sub.f32 %f409, %f408, %f406;
|
|
mul.f32 %f410, %f295, %f60;
|
|
sub.f32 %f411, %f409, %f410;
|
|
fma.rn.f32 %f412, %f308, %f411, %f304;
|
|
cvt.sat.f32.f32 %f413, %f412;
|
|
mul.f32 %f414, %f413, %f413;
|
|
mul.f32 %f1278, %f91, %f414;
|
|
mov.f32 %f1275, %f305;
|
|
bra.uni BB0_29;
|
|
|
|
BB0_19:
|
|
mov.f32 %f1275, %f305;
|
|
bra.uni BB0_28;
|
|
|
|
BB0_25:
|
|
mov.f32 %f1271, %f58;
|
|
mov.f32 %f1272, %f59;
|
|
mov.f32 %f1273, %f60;
|
|
|
|
BB0_27:
|
|
mul.f32 %f440, %f295, %f60;
|
|
mul.f32 %f441, %f294, %f59;
|
|
neg.f32 %f442, %f441;
|
|
mul.f32 %f443, %f293, %f58;
|
|
sub.f32 %f444, %f442, %f443;
|
|
sub.f32 %f445, %f444, %f440;
|
|
setp.gt.f32 %p37, %f445, 0f00000000;
|
|
selp.f32 %f446, 0f3F800000, 0f00000000, %p37;
|
|
mul.f32 %f447, %f302, %f1272;
|
|
fma.rn.f32 %f448, %f301, %f1271, %f447;
|
|
mul.f32 %f449, %f298, %f1272;
|
|
fma.rn.f32 %f450, %f297, %f1271, %f449;
|
|
fma.rn.f32 %f451, %f303, %f1273, %f448;
|
|
fma.rn.f32 %f452, %f299, %f1273, %f450;
|
|
fma.rn.f32 %f453, %f304, %f451, 0f3F000000;
|
|
mov.f32 %f454, 0f3F800000;
|
|
sub.f32 %f436, %f454, %f453;
|
|
fma.rn.f32 %f437, %f304, %f452, 0f3F000000;
|
|
cvt.rzi.s32.f32 %r71, %f308;
|
|
mov.f32 %f439, 0f00000000;
|
|
// inline asm
|
|
call (%f432, %f433, %f434, %f435), _rt_texture_get_f_id, (%r71, %r28, %f436, %f437, %f439, %f439);
|
|
// inline asm
|
|
mul.f32 %f455, %f446, %f432;
|
|
mul.f32 %f456, %f446, %f433;
|
|
mul.f32 %f457, %f446, %f434;
|
|
mul.f32 %f1275, %f305, %f455;
|
|
mul.f32 %f1276, %f1276, %f456;
|
|
mul.f32 %f1277, %f1277, %f457;
|
|
|
|
BB0_28:
|
|
mov.f32 %f1278, %f91;
|
|
|
|
BB0_29:
|
|
max.f32 %f464, %f1275, %f1276;
|
|
max.f32 %f465, %f464, %f1277;
|
|
mul.f32 %f112, %f92, %f1278;
|
|
mul.f32 %f466, %f112, %f465;
|
|
setp.lt.f32 %p39, %f466, 0f3727C5AC;
|
|
mov.pred %p133, -1;
|
|
mov.f32 %f119, 0f00000000;
|
|
mov.f32 %f120, %f119;
|
|
mov.f32 %f121, %f119;
|
|
mov.f32 %f122, %f119;
|
|
mov.f32 %f123, %f119;
|
|
mov.f32 %f124, %f119;
|
|
@%p39 bra BB0_31;
|
|
|
|
mul.f32 %f119, %f1275, %f112;
|
|
mul.f32 %f120, %f1276, %f112;
|
|
mul.f32 %f121, %f1277, %f112;
|
|
ld.global.u8 %rs12, [imageEnabled];
|
|
and.b16 %rs13, %rs12, 64;
|
|
setp.eq.s16 %p41, %rs13, 0;
|
|
selp.f32 %f122, 0f00000000, %f58, %p41;
|
|
selp.f32 %f123, 0f00000000, %f59, %p41;
|
|
selp.f32 %f124, 0f00000000, %f60, %p41;
|
|
mov.pred %p133, 0;
|
|
|
|
BB0_31:
|
|
@%p133 bra BB0_48;
|
|
|
|
setp.eq.s32 %p42, %r6, 0;
|
|
mov.u16 %rs62, 0;
|
|
mov.f32 %f1301, %f1300;
|
|
mov.f32 %f1302, %f1300;
|
|
@%p42 bra BB0_43;
|
|
|
|
abs.s32 %r8, %r6;
|
|
mov.f32 %f1297, 0f00000000;
|
|
setp.lt.s32 %p43, %r8, 1;
|
|
mov.f32 %f1298, %f1297;
|
|
mov.f32 %f1299, %f1297;
|
|
@%p43 bra BB0_42;
|
|
|
|
max.f32 %f126, %f480, %f275;
|
|
and.b32 %r9, %r8, 3;
|
|
setp.eq.s32 %p44, %r9, 0;
|
|
add.u64 %rd35, %SP, 0;
|
|
cvta.to.local.u64 %rd2, %rd35;
|
|
mov.f32 %f1297, 0f00000000;
|
|
mov.u32 %r202, 0;
|
|
mov.f32 %f1298, %f1297;
|
|
mov.f32 %f1299, %f1297;
|
|
@%p44 bra BB0_40;
|
|
|
|
setp.eq.s32 %p45, %r9, 1;
|
|
mov.f32 %f1288, 0f00000000;
|
|
mov.u32 %r200, 0;
|
|
mov.f32 %f1289, %f1288;
|
|
mov.f32 %f1290, %f1288;
|
|
@%p45 bra BB0_39;
|
|
|
|
setp.eq.s32 %p46, %r9, 2;
|
|
mov.f32 %f1285, 0f00000000;
|
|
mov.u32 %r199, 0;
|
|
mov.f32 %f1286, %f1285;
|
|
mov.f32 %f1287, %f1285;
|
|
@%p46 bra BB0_38;
|
|
|
|
cvt.rmi.f32.f32 %f496, %f19;
|
|
sub.f32 %f497, %f19, %f496;
|
|
cvt.rmi.f32.f32 %f498, %f20;
|
|
sub.f32 %f499, %f20, %f498;
|
|
add.f32 %f500, %f499, 0f420551EC;
|
|
add.f32 %f501, %f497, 0f420551EC;
|
|
mul.f32 %f502, %f497, %f500;
|
|
fma.rn.f32 %f503, %f499, %f501, %f502;
|
|
add.f32 %f504, %f503, 0f00000000;
|
|
add.f32 %f505, %f497, %f504;
|
|
add.f32 %f506, %f499, %f504;
|
|
add.f32 %f507, %f505, %f506;
|
|
mul.f32 %f508, %f504, %f507;
|
|
cvt.rmi.f32.f32 %f509, %f508;
|
|
sub.f32 %f510, %f508, %f509;
|
|
add.f32 %f511, %f505, %f505;
|
|
mul.f32 %f512, %f506, %f511;
|
|
cvt.rmi.f32.f32 %f513, %f512;
|
|
sub.f32 %f514, %f512, %f513;
|
|
mul.f32 %f515, %f505, %f507;
|
|
cvt.rmi.f32.f32 %f516, %f515;
|
|
sub.f32 %f517, %f515, %f516;
|
|
fma.rn.f32 %f518, %f510, 0f40000000, 0fBF800000;
|
|
fma.rn.f32 %f519, %f514, 0f40000000, 0fBF800000;
|
|
fma.rn.f32 %f520, %f517, 0f40000000, 0fBF800000;
|
|
fma.rn.f32 %f521, %f309, %f518, %f314;
|
|
fma.rn.f32 %f522, %f309, %f519, %f315;
|
|
fma.rn.f32 %f523, %f309, %f520, %f316;
|
|
sub.f32 %f524, %f521, %f7;
|
|
sub.f32 %f525, %f522, %f8;
|
|
sub.f32 %f526, %f523, %f9;
|
|
mul.f32 %f527, %f525, %f525;
|
|
fma.rn.f32 %f528, %f524, %f524, %f527;
|
|
fma.rn.f32 %f529, %f526, %f526, %f528;
|
|
sqrt.rn.f32 %f495, %f529;
|
|
rcp.rn.f32 %f530, %f495;
|
|
mul.f32 %f491, %f530, %f524;
|
|
mul.f32 %f492, %f530, %f525;
|
|
mul.f32 %f493, %f530, %f526;
|
|
ld.global.u32 %r80, [imageEnabled];
|
|
and.b32 %r81, %r80, 32;
|
|
setp.eq.s32 %p47, %r81, 0;
|
|
selp.f32 %f531, 0f3F800000, 0f41200000, %p47;
|
|
mul.f32 %f494, %f531, %f126;
|
|
mov.u32 %r82, 1065353216;
|
|
st.local.u32 [%rd2+8], %r82;
|
|
st.local.u32 [%rd2+4], %r82;
|
|
st.local.u32 [%rd2], %r82;
|
|
ld.global.u32 %r76, [root];
|
|
// inline asm
|
|
call _rt_trace_64, (%r76, %f10, %f11, %f12, %f491, %f492, %f493, %r45, %f494, %f495, %rd35, %r44);
|
|
// inline asm
|
|
ld.local.f32 %f532, [%rd2];
|
|
add.f32 %f1287, %f532, 0f00000000;
|
|
ld.local.f32 %f533, [%rd2+4];
|
|
add.f32 %f1286, %f533, 0f00000000;
|
|
ld.local.f32 %f534, [%rd2+8];
|
|
add.f32 %f1285, %f534, 0f00000000;
|
|
mov.u32 %r199, %r45;
|
|
|
|
BB0_38:
|
|
cvt.rn.f32.s32 %f543, %r199;
|
|
add.f32 %f544, %f13, %f543;
|
|
sub.f32 %f545, %f14, %f543;
|
|
mul.f32 %f546, %f544, 0f3DD32618;
|
|
cvt.rmi.f32.f32 %f547, %f546;
|
|
sub.f32 %f548, %f546, %f547;
|
|
mul.f32 %f549, %f545, 0f3DD2F1AA;
|
|
cvt.rmi.f32.f32 %f550, %f549;
|
|
sub.f32 %f551, %f549, %f550;
|
|
mul.f32 %f552, %f543, 0f3DC74539;
|
|
cvt.rmi.f32.f32 %f553, %f552;
|
|
sub.f32 %f554, %f552, %f553;
|
|
add.f32 %f555, %f551, 0f420551EC;
|
|
add.f32 %f556, %f548, 0f420551EC;
|
|
add.f32 %f557, %f554, 0f420551EC;
|
|
mul.f32 %f558, %f548, %f555;
|
|
fma.rn.f32 %f559, %f551, %f556, %f558;
|
|
fma.rn.f32 %f560, %f554, %f557, %f559;
|
|
add.f32 %f561, %f548, %f560;
|
|
add.f32 %f562, %f551, %f560;
|
|
add.f32 %f563, %f554, %f560;
|
|
add.f32 %f564, %f561, %f562;
|
|
mul.f32 %f565, %f563, %f564;
|
|
cvt.rmi.f32.f32 %f566, %f565;
|
|
sub.f32 %f567, %f565, %f566;
|
|
add.f32 %f568, %f561, %f561;
|
|
mul.f32 %f569, %f562, %f568;
|
|
cvt.rmi.f32.f32 %f570, %f569;
|
|
sub.f32 %f571, %f569, %f570;
|
|
mul.f32 %f572, %f561, %f564;
|
|
cvt.rmi.f32.f32 %f573, %f572;
|
|
sub.f32 %f574, %f572, %f573;
|
|
fma.rn.f32 %f575, %f567, 0f40000000, 0fBF800000;
|
|
fma.rn.f32 %f576, %f571, 0f40000000, 0fBF800000;
|
|
fma.rn.f32 %f577, %f574, 0f40000000, 0fBF800000;
|
|
fma.rn.f32 %f578, %f309, %f575, %f314;
|
|
fma.rn.f32 %f579, %f309, %f576, %f315;
|
|
fma.rn.f32 %f580, %f309, %f577, %f316;
|
|
sub.f32 %f581, %f578, %f7;
|
|
sub.f32 %f582, %f579, %f8;
|
|
sub.f32 %f583, %f580, %f9;
|
|
mul.f32 %f584, %f582, %f582;
|
|
fma.rn.f32 %f585, %f581, %f581, %f584;
|
|
fma.rn.f32 %f586, %f583, %f583, %f585;
|
|
sqrt.rn.f32 %f542, %f586;
|
|
rcp.rn.f32 %f587, %f542;
|
|
mul.f32 %f538, %f587, %f581;
|
|
mul.f32 %f539, %f587, %f582;
|
|
mul.f32 %f540, %f587, %f583;
|
|
ld.global.u32 %r86, [imageEnabled];
|
|
and.b32 %r87, %r86, 32;
|
|
setp.eq.s32 %p48, %r87, 0;
|
|
selp.f32 %f588, 0f3F800000, 0f41200000, %p48;
|
|
mul.f32 %f541, %f588, %f126;
|
|
mov.u32 %r88, 1065353216;
|
|
st.local.u32 [%rd2+8], %r88;
|
|
st.local.u32 [%rd2+4], %r88;
|
|
st.local.u32 [%rd2], %r88;
|
|
ld.global.u32 %r83, [root];
|
|
// inline asm
|
|
call _rt_trace_64, (%r83, %f10, %f11, %f12, %f538, %f539, %f540, %r45, %f541, %f542, %rd35, %r44);
|
|
// inline asm
|
|
ld.local.f32 %f589, [%rd2];
|
|
add.f32 %f1290, %f1287, %f589;
|
|
ld.local.f32 %f590, [%rd2+4];
|
|
add.f32 %f1289, %f1286, %f590;
|
|
ld.local.f32 %f591, [%rd2+8];
|
|
add.f32 %f1288, %f1285, %f591;
|
|
add.s32 %r200, %r199, 1;
|
|
|
|
BB0_39:
|
|
cvt.rn.f32.s32 %f600, %r200;
|
|
add.f32 %f601, %f13, %f600;
|
|
sub.f32 %f602, %f14, %f600;
|
|
mul.f32 %f603, %f601, 0f3DD32618;
|
|
cvt.rmi.f32.f32 %f604, %f603;
|
|
sub.f32 %f605, %f603, %f604;
|
|
mul.f32 %f606, %f602, 0f3DD2F1AA;
|
|
cvt.rmi.f32.f32 %f607, %f606;
|
|
sub.f32 %f608, %f606, %f607;
|
|
mul.f32 %f609, %f600, 0f3DC74539;
|
|
cvt.rmi.f32.f32 %f610, %f609;
|
|
sub.f32 %f611, %f609, %f610;
|
|
add.f32 %f612, %f608, 0f420551EC;
|
|
add.f32 %f613, %f605, 0f420551EC;
|
|
add.f32 %f614, %f611, 0f420551EC;
|
|
mul.f32 %f615, %f605, %f612;
|
|
fma.rn.f32 %f616, %f608, %f613, %f615;
|
|
fma.rn.f32 %f617, %f611, %f614, %f616;
|
|
add.f32 %f618, %f605, %f617;
|
|
add.f32 %f619, %f608, %f617;
|
|
add.f32 %f620, %f611, %f617;
|
|
add.f32 %f621, %f618, %f619;
|
|
mul.f32 %f622, %f620, %f621;
|
|
cvt.rmi.f32.f32 %f623, %f622;
|
|
sub.f32 %f624, %f622, %f623;
|
|
add.f32 %f625, %f618, %f618;
|
|
mul.f32 %f626, %f619, %f625;
|
|
cvt.rmi.f32.f32 %f627, %f626;
|
|
sub.f32 %f628, %f626, %f627;
|
|
mul.f32 %f629, %f618, %f621;
|
|
cvt.rmi.f32.f32 %f630, %f629;
|
|
sub.f32 %f631, %f629, %f630;
|
|
fma.rn.f32 %f632, %f624, 0f40000000, 0fBF800000;
|
|
fma.rn.f32 %f633, %f628, 0f40000000, 0fBF800000;
|
|
fma.rn.f32 %f634, %f631, 0f40000000, 0fBF800000;
|
|
fma.rn.f32 %f635, %f309, %f632, %f314;
|
|
fma.rn.f32 %f636, %f309, %f633, %f315;
|
|
fma.rn.f32 %f637, %f309, %f634, %f316;
|
|
sub.f32 %f638, %f635, %f7;
|
|
sub.f32 %f639, %f636, %f8;
|
|
sub.f32 %f640, %f637, %f9;
|
|
mul.f32 %f641, %f639, %f639;
|
|
fma.rn.f32 %f642, %f638, %f638, %f641;
|
|
fma.rn.f32 %f643, %f640, %f640, %f642;
|
|
sqrt.rn.f32 %f599, %f643;
|
|
rcp.rn.f32 %f644, %f599;
|
|
mul.f32 %f595, %f644, %f638;
|
|
mul.f32 %f596, %f644, %f639;
|
|
mul.f32 %f597, %f644, %f640;
|
|
ld.global.u32 %r92, [imageEnabled];
|
|
and.b32 %r93, %r92, 32;
|
|
setp.eq.s32 %p49, %r93, 0;
|
|
selp.f32 %f645, 0f3F800000, 0f41200000, %p49;
|
|
mul.f32 %f598, %f645, %f126;
|
|
mov.u32 %r94, 1065353216;
|
|
st.local.u32 [%rd2+8], %r94;
|
|
st.local.u32 [%rd2+4], %r94;
|
|
st.local.u32 [%rd2], %r94;
|
|
ld.global.u32 %r89, [root];
|
|
mov.u32 %r90, 1;
|
|
// inline asm
|
|
call _rt_trace_64, (%r89, %f10, %f11, %f12, %f595, %f596, %f597, %r90, %f598, %f599, %rd35, %r44);
|
|
// inline asm
|
|
ld.local.f32 %f646, [%rd2];
|
|
add.f32 %f1299, %f1290, %f646;
|
|
ld.local.f32 %f647, [%rd2+4];
|
|
add.f32 %f1298, %f1289, %f647;
|
|
ld.local.f32 %f648, [%rd2+8];
|
|
add.f32 %f1297, %f1288, %f648;
|
|
add.s32 %r202, %r200, 1;
|
|
|
|
BB0_40:
|
|
setp.lt.u32 %p50, %r8, 4;
|
|
@%p50 bra BB0_42;
|
|
|
|
BB0_41:
|
|
cvt.rn.f32.s32 %f681, %r202;
|
|
add.f32 %f682, %f13, %f681;
|
|
sub.f32 %f683, %f14, %f681;
|
|
mul.f32 %f684, %f682, 0f3DD32618;
|
|
cvt.rmi.f32.f32 %f685, %f684;
|
|
sub.f32 %f686, %f684, %f685;
|
|
mul.f32 %f687, %f683, 0f3DD2F1AA;
|
|
cvt.rmi.f32.f32 %f688, %f687;
|
|
sub.f32 %f689, %f687, %f688;
|
|
mul.f32 %f690, %f681, 0f3DC74539;
|
|
cvt.rmi.f32.f32 %f691, %f690;
|
|
sub.f32 %f692, %f690, %f691;
|
|
add.f32 %f693, %f689, 0f420551EC;
|
|
add.f32 %f694, %f686, 0f420551EC;
|
|
add.f32 %f695, %f692, 0f420551EC;
|
|
mul.f32 %f696, %f686, %f693;
|
|
fma.rn.f32 %f697, %f689, %f694, %f696;
|
|
fma.rn.f32 %f698, %f692, %f695, %f697;
|
|
add.f32 %f699, %f686, %f698;
|
|
add.f32 %f700, %f689, %f698;
|
|
add.f32 %f701, %f692, %f698;
|
|
add.f32 %f702, %f699, %f700;
|
|
mul.f32 %f703, %f701, %f702;
|
|
cvt.rmi.f32.f32 %f704, %f703;
|
|
sub.f32 %f705, %f703, %f704;
|
|
add.f32 %f706, %f699, %f699;
|
|
mul.f32 %f707, %f700, %f706;
|
|
cvt.rmi.f32.f32 %f708, %f707;
|
|
sub.f32 %f709, %f707, %f708;
|
|
mul.f32 %f710, %f699, %f702;
|
|
cvt.rmi.f32.f32 %f711, %f710;
|
|
sub.f32 %f712, %f710, %f711;
|
|
fma.rn.f32 %f713, %f705, 0f40000000, 0fBF800000;
|
|
fma.rn.f32 %f714, %f709, 0f40000000, 0fBF800000;
|
|
fma.rn.f32 %f715, %f712, 0f40000000, 0fBF800000;
|
|
fma.rn.f32 %f716, %f309, %f713, %f314;
|
|
fma.rn.f32 %f717, %f309, %f714, %f315;
|
|
fma.rn.f32 %f718, %f309, %f715, %f316;
|
|
sub.f32 %f719, %f716, %f7;
|
|
sub.f32 %f720, %f717, %f8;
|
|
sub.f32 %f721, %f718, %f9;
|
|
mul.f32 %f722, %f720, %f720;
|
|
fma.rn.f32 %f723, %f719, %f719, %f722;
|
|
fma.rn.f32 %f724, %f721, %f721, %f723;
|
|
sqrt.rn.f32 %f656, %f724;
|
|
rcp.rn.f32 %f725, %f656;
|
|
mul.f32 %f652, %f725, %f719;
|
|
mul.f32 %f653, %f725, %f720;
|
|
mul.f32 %f654, %f725, %f721;
|
|
ld.global.u32 %r107, [imageEnabled];
|
|
and.b32 %r108, %r107, 32;
|
|
setp.eq.s32 %p51, %r108, 0;
|
|
selp.f32 %f726, 0f3F800000, 0f41200000, %p51;
|
|
mul.f32 %f655, %f726, %f126;
|
|
mov.u32 %r109, 1065353216;
|
|
st.local.u32 [%rd2+8], %r109;
|
|
st.local.u32 [%rd2+4], %r109;
|
|
st.local.u32 [%rd2], %r109;
|
|
ld.global.u32 %r95, [root];
|
|
mov.u32 %r105, 1;
|
|
// inline asm
|
|
call _rt_trace_64, (%r95, %f10, %f11, %f12, %f652, %f653, %f654, %r105, %f655, %f656, %rd35, %r44);
|
|
// inline asm
|
|
ld.local.f32 %f727, [%rd2];
|
|
add.f32 %f728, %f1299, %f727;
|
|
ld.local.f32 %f729, [%rd2+4];
|
|
add.f32 %f730, %f1298, %f729;
|
|
ld.local.f32 %f731, [%rd2+8];
|
|
add.f32 %f732, %f1297, %f731;
|
|
add.s32 %r110, %r202, 1;
|
|
cvt.rn.f32.s32 %f733, %r110;
|
|
add.f32 %f734, %f13, %f733;
|
|
sub.f32 %f735, %f14, %f733;
|
|
mul.f32 %f736, %f734, 0f3DD32618;
|
|
cvt.rmi.f32.f32 %f737, %f736;
|
|
sub.f32 %f738, %f736, %f737;
|
|
mul.f32 %f739, %f735, 0f3DD2F1AA;
|
|
cvt.rmi.f32.f32 %f740, %f739;
|
|
sub.f32 %f741, %f739, %f740;
|
|
mul.f32 %f742, %f733, 0f3DC74539;
|
|
cvt.rmi.f32.f32 %f743, %f742;
|
|
sub.f32 %f744, %f742, %f743;
|
|
add.f32 %f745, %f741, 0f420551EC;
|
|
add.f32 %f746, %f738, 0f420551EC;
|
|
add.f32 %f747, %f744, 0f420551EC;
|
|
mul.f32 %f748, %f738, %f745;
|
|
fma.rn.f32 %f749, %f741, %f746, %f748;
|
|
fma.rn.f32 %f750, %f744, %f747, %f749;
|
|
add.f32 %f751, %f738, %f750;
|
|
add.f32 %f752, %f741, %f750;
|
|
add.f32 %f753, %f744, %f750;
|
|
add.f32 %f754, %f751, %f752;
|
|
mul.f32 %f755, %f753, %f754;
|
|
cvt.rmi.f32.f32 %f756, %f755;
|
|
sub.f32 %f757, %f755, %f756;
|
|
add.f32 %f758, %f751, %f751;
|
|
mul.f32 %f759, %f752, %f758;
|
|
cvt.rmi.f32.f32 %f760, %f759;
|
|
sub.f32 %f761, %f759, %f760;
|
|
mul.f32 %f762, %f751, %f754;
|
|
cvt.rmi.f32.f32 %f763, %f762;
|
|
sub.f32 %f764, %f762, %f763;
|
|
fma.rn.f32 %f765, %f757, 0f40000000, 0fBF800000;
|
|
fma.rn.f32 %f766, %f761, 0f40000000, 0fBF800000;
|
|
fma.rn.f32 %f767, %f764, 0f40000000, 0fBF800000;
|
|
fma.rn.f32 %f768, %f309, %f765, %f314;
|
|
fma.rn.f32 %f769, %f309, %f766, %f315;
|
|
fma.rn.f32 %f770, %f309, %f767, %f316;
|
|
sub.f32 %f771, %f768, %f7;
|
|
sub.f32 %f772, %f769, %f8;
|
|
sub.f32 %f773, %f770, %f9;
|
|
mul.f32 %f774, %f772, %f772;
|
|
fma.rn.f32 %f775, %f771, %f771, %f774;
|
|
fma.rn.f32 %f776, %f773, %f773, %f775;
|
|
sqrt.rn.f32 %f664, %f776;
|
|
rcp.rn.f32 %f777, %f664;
|
|
mul.f32 %f660, %f777, %f771;
|
|
mul.f32 %f661, %f777, %f772;
|
|
mul.f32 %f662, %f777, %f773;
|
|
ld.global.u32 %r111, [imageEnabled];
|
|
and.b32 %r112, %r111, 32;
|
|
setp.eq.s32 %p52, %r112, 0;
|
|
selp.f32 %f778, 0f3F800000, 0f41200000, %p52;
|
|
mul.f32 %f663, %f778, %f126;
|
|
st.local.u32 [%rd2+8], %r109;
|
|
st.local.u32 [%rd2+4], %r109;
|
|
st.local.u32 [%rd2], %r109;
|
|
ld.global.u32 %r98, [root];
|
|
// inline asm
|
|
call _rt_trace_64, (%r98, %f10, %f11, %f12, %f660, %f661, %f662, %r105, %f663, %f664, %rd35, %r44);
|
|
// inline asm
|
|
ld.local.f32 %f779, [%rd2];
|
|
add.f32 %f780, %f728, %f779;
|
|
ld.local.f32 %f781, [%rd2+4];
|
|
add.f32 %f782, %f730, %f781;
|
|
ld.local.f32 %f783, [%rd2+8];
|
|
add.f32 %f784, %f732, %f783;
|
|
add.s32 %r113, %r202, 2;
|
|
cvt.rn.f32.s32 %f785, %r113;
|
|
add.f32 %f786, %f13, %f785;
|
|
sub.f32 %f787, %f14, %f785;
|
|
mul.f32 %f788, %f786, 0f3DD32618;
|
|
cvt.rmi.f32.f32 %f789, %f788;
|
|
sub.f32 %f790, %f788, %f789;
|
|
mul.f32 %f791, %f787, 0f3DD2F1AA;
|
|
cvt.rmi.f32.f32 %f792, %f791;
|
|
sub.f32 %f793, %f791, %f792;
|
|
mul.f32 %f794, %f785, 0f3DC74539;
|
|
cvt.rmi.f32.f32 %f795, %f794;
|
|
sub.f32 %f796, %f794, %f795;
|
|
add.f32 %f797, %f793, 0f420551EC;
|
|
add.f32 %f798, %f790, 0f420551EC;
|
|
add.f32 %f799, %f796, 0f420551EC;
|
|
mul.f32 %f800, %f790, %f797;
|
|
fma.rn.f32 %f801, %f793, %f798, %f800;
|
|
fma.rn.f32 %f802, %f796, %f799, %f801;
|
|
add.f32 %f803, %f790, %f802;
|
|
add.f32 %f804, %f793, %f802;
|
|
add.f32 %f805, %f796, %f802;
|
|
add.f32 %f806, %f803, %f804;
|
|
mul.f32 %f807, %f805, %f806;
|
|
cvt.rmi.f32.f32 %f808, %f807;
|
|
sub.f32 %f809, %f807, %f808;
|
|
add.f32 %f810, %f803, %f803;
|
|
mul.f32 %f811, %f804, %f810;
|
|
cvt.rmi.f32.f32 %f812, %f811;
|
|
sub.f32 %f813, %f811, %f812;
|
|
mul.f32 %f814, %f803, %f806;
|
|
cvt.rmi.f32.f32 %f815, %f814;
|
|
sub.f32 %f816, %f814, %f815;
|
|
fma.rn.f32 %f817, %f809, 0f40000000, 0fBF800000;
|
|
fma.rn.f32 %f818, %f813, 0f40000000, 0fBF800000;
|
|
fma.rn.f32 %f819, %f816, 0f40000000, 0fBF800000;
|
|
fma.rn.f32 %f820, %f309, %f817, %f314;
|
|
fma.rn.f32 %f821, %f309, %f818, %f315;
|
|
fma.rn.f32 %f822, %f309, %f819, %f316;
|
|
sub.f32 %f823, %f820, %f7;
|
|
sub.f32 %f824, %f821, %f8;
|
|
sub.f32 %f825, %f822, %f9;
|
|
mul.f32 %f826, %f824, %f824;
|
|
fma.rn.f32 %f827, %f823, %f823, %f826;
|
|
fma.rn.f32 %f828, %f825, %f825, %f827;
|
|
sqrt.rn.f32 %f672, %f828;
|
|
rcp.rn.f32 %f829, %f672;
|
|
mul.f32 %f668, %f829, %f823;
|
|
mul.f32 %f669, %f829, %f824;
|
|
mul.f32 %f670, %f829, %f825;
|
|
ld.global.u32 %r114, [imageEnabled];
|
|
and.b32 %r115, %r114, 32;
|
|
setp.eq.s32 %p53, %r115, 0;
|
|
selp.f32 %f830, 0f3F800000, 0f41200000, %p53;
|
|
mul.f32 %f671, %f830, %f126;
|
|
st.local.u32 [%rd2+8], %r109;
|
|
st.local.u32 [%rd2+4], %r109;
|
|
st.local.u32 [%rd2], %r109;
|
|
ld.global.u32 %r101, [root];
|
|
// inline asm
|
|
call _rt_trace_64, (%r101, %f10, %f11, %f12, %f668, %f669, %f670, %r105, %f671, %f672, %rd35, %r44);
|
|
// inline asm
|
|
ld.local.f32 %f831, [%rd2];
|
|
add.f32 %f832, %f780, %f831;
|
|
ld.local.f32 %f833, [%rd2+4];
|
|
add.f32 %f834, %f782, %f833;
|
|
ld.local.f32 %f835, [%rd2+8];
|
|
add.f32 %f836, %f784, %f835;
|
|
add.s32 %r116, %r202, 3;
|
|
cvt.rn.f32.s32 %f837, %r116;
|
|
add.f32 %f838, %f13, %f837;
|
|
sub.f32 %f839, %f14, %f837;
|
|
mul.f32 %f840, %f838, 0f3DD32618;
|
|
cvt.rmi.f32.f32 %f841, %f840;
|
|
sub.f32 %f842, %f840, %f841;
|
|
mul.f32 %f843, %f839, 0f3DD2F1AA;
|
|
cvt.rmi.f32.f32 %f844, %f843;
|
|
sub.f32 %f845, %f843, %f844;
|
|
mul.f32 %f846, %f837, 0f3DC74539;
|
|
cvt.rmi.f32.f32 %f847, %f846;
|
|
sub.f32 %f848, %f846, %f847;
|
|
add.f32 %f849, %f845, 0f420551EC;
|
|
add.f32 %f850, %f842, 0f420551EC;
|
|
add.f32 %f851, %f848, 0f420551EC;
|
|
mul.f32 %f852, %f842, %f849;
|
|
fma.rn.f32 %f853, %f845, %f850, %f852;
|
|
fma.rn.f32 %f854, %f848, %f851, %f853;
|
|
add.f32 %f855, %f842, %f854;
|
|
add.f32 %f856, %f845, %f854;
|
|
add.f32 %f857, %f848, %f854;
|
|
add.f32 %f858, %f855, %f856;
|
|
mul.f32 %f859, %f857, %f858;
|
|
cvt.rmi.f32.f32 %f860, %f859;
|
|
sub.f32 %f861, %f859, %f860;
|
|
add.f32 %f862, %f855, %f855;
|
|
mul.f32 %f863, %f856, %f862;
|
|
cvt.rmi.f32.f32 %f864, %f863;
|
|
sub.f32 %f865, %f863, %f864;
|
|
mul.f32 %f866, %f855, %f858;
|
|
cvt.rmi.f32.f32 %f867, %f866;
|
|
sub.f32 %f868, %f866, %f867;
|
|
fma.rn.f32 %f869, %f861, 0f40000000, 0fBF800000;
|
|
fma.rn.f32 %f870, %f865, 0f40000000, 0fBF800000;
|
|
fma.rn.f32 %f871, %f868, 0f40000000, 0fBF800000;
|
|
fma.rn.f32 %f872, %f309, %f869, %f314;
|
|
fma.rn.f32 %f873, %f309, %f870, %f315;
|
|
fma.rn.f32 %f874, %f309, %f871, %f316;
|
|
sub.f32 %f875, %f872, %f7;
|
|
sub.f32 %f876, %f873, %f8;
|
|
sub.f32 %f877, %f874, %f9;
|
|
mul.f32 %f878, %f876, %f876;
|
|
fma.rn.f32 %f879, %f875, %f875, %f878;
|
|
fma.rn.f32 %f880, %f877, %f877, %f879;
|
|
sqrt.rn.f32 %f680, %f880;
|
|
rcp.rn.f32 %f881, %f680;
|
|
mul.f32 %f676, %f881, %f875;
|
|
mul.f32 %f677, %f881, %f876;
|
|
mul.f32 %f678, %f881, %f877;
|
|
ld.global.u32 %r117, [imageEnabled];
|
|
and.b32 %r118, %r117, 32;
|
|
setp.eq.s32 %p54, %r118, 0;
|
|
selp.f32 %f882, 0f3F800000, 0f41200000, %p54;
|
|
mul.f32 %f679, %f882, %f126;
|
|
st.local.u32 [%rd2+8], %r109;
|
|
st.local.u32 [%rd2+4], %r109;
|
|
st.local.u32 [%rd2], %r109;
|
|
ld.global.u32 %r104, [root];
|
|
// inline asm
|
|
call _rt_trace_64, (%r104, %f10, %f11, %f12, %f676, %f677, %f678, %r105, %f679, %f680, %rd35, %r44);
|
|
// inline asm
|
|
ld.local.f32 %f883, [%rd2];
|
|
add.f32 %f1299, %f832, %f883;
|
|
ld.local.f32 %f884, [%rd2+4];
|
|
add.f32 %f1298, %f834, %f884;
|
|
ld.local.f32 %f885, [%rd2+8];
|
|
add.f32 %f1297, %f836, %f885;
|
|
add.s32 %r202, %r202, 4;
|
|
setp.lt.s32 %p55, %r202, %r8;
|
|
@%p55 bra BB0_41;
|
|
|
|
BB0_42:
|
|
cvt.rn.f32.s32 %f886, %r8;
|
|
rcp.rn.f32 %f887, %f886;
|
|
mul.f32 %f1300, %f1299, %f887;
|
|
mul.f32 %f1301, %f1298, %f887;
|
|
mul.f32 %f1302, %f1297, %f887;
|
|
shr.u32 %r119, %r6, 31;
|
|
cvt.u16.u32 %rs62, %r119;
|
|
|
|
BB0_43:
|
|
fma.rn.f32 %f1267, %f119, %f1300, %f1267;
|
|
fma.rn.f32 %f1266, %f120, %f1301, %f1266;
|
|
fma.rn.f32 %f1265, %f121, %f1302, %f1265;
|
|
ld.global.u8 %rs15, [imageEnabled];
|
|
and.b16 %rs16, %rs15, 64;
|
|
setp.eq.s16 %p56, %rs16, 0;
|
|
@%p56 bra BB0_45;
|
|
|
|
mul.f32 %f888, %f120, 0f3F372474;
|
|
fma.rn.f32 %f889, %f119, 0f3E59999A, %f888;
|
|
fma.rn.f32 %f890, %f121, 0f3D93A92A, %f889;
|
|
fma.rn.f32 %f24, %f122, %f890, %f24;
|
|
fma.rn.f32 %f23, %f123, %f890, %f23;
|
|
fma.rn.f32 %f22, %f890, %f124, %f22;
|
|
|
|
BB0_45:
|
|
setp.eq.s16 %p57, %rs62, 0;
|
|
@%p57 bra BB0_47;
|
|
|
|
div.rn.f32 %f891, %f119, %f305;
|
|
div.rn.f32 %f892, %f891, %f91;
|
|
cvt.sat.f32.f32 %f893, %f892;
|
|
mul.f32 %f1300, %f1300, %f893;
|
|
|
|
BB0_47:
|
|
add.f32 %f1261, %f1261, %f1300;
|
|
|
|
BB0_48:
|
|
add.s32 %r198, %r198, 1;
|
|
setp.lt.u32 %p58, %r198, %r4;
|
|
@%p58 bra BB0_5;
|
|
|
|
BB0_49:
|
|
ld.global.u32 %r204, [imageEnabled];
|
|
and.b32 %r120, %r204, 8;
|
|
setp.eq.s32 %p59, %r120, 0;
|
|
@%p59 bra BB0_62;
|
|
|
|
cvt.sat.f32.f32 %f186, %f1261;
|
|
cvt.u64.u32 %rd46, %r3;
|
|
cvt.u64.u32 %rd45, %r2;
|
|
mov.u64 %rd49, image_Mask;
|
|
cvta.global.u64 %rd44, %rd49;
|
|
// inline asm
|
|
call (%rd43), _rt_buffer_get_64, (%rd44, %r28, %r28, %rd45, %rd46, %rd13, %rd13);
|
|
// inline asm
|
|
mov.f32 %f896, 0f3E68BA2E;
|
|
cvt.rzi.f32.f32 %f897, %f896;
|
|
fma.rn.f32 %f898, %f897, 0fC0000000, 0f3EE8BA2E;
|
|
abs.f32 %f187, %f898;
|
|
abs.f32 %f188, %f186;
|
|
setp.lt.f32 %p60, %f188, 0f00800000;
|
|
mul.f32 %f899, %f188, 0f4B800000;
|
|
selp.f32 %f900, 0fC3170000, 0fC2FE0000, %p60;
|
|
selp.f32 %f901, %f899, %f188, %p60;
|
|
mov.b32 %r123, %f901;
|
|
and.b32 %r124, %r123, 8388607;
|
|
or.b32 %r125, %r124, 1065353216;
|
|
mov.b32 %f902, %r125;
|
|
shr.u32 %r126, %r123, 23;
|
|
cvt.rn.f32.u32 %f903, %r126;
|
|
add.f32 %f904, %f900, %f903;
|
|
setp.gt.f32 %p61, %f902, 0f3FB504F3;
|
|
mul.f32 %f905, %f902, 0f3F000000;
|
|
add.f32 %f906, %f904, 0f3F800000;
|
|
selp.f32 %f907, %f905, %f902, %p61;
|
|
selp.f32 %f908, %f906, %f904, %p61;
|
|
add.f32 %f909, %f907, 0fBF800000;
|
|
add.f32 %f895, %f907, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f894,%f895;
|
|
// inline asm
|
|
add.f32 %f910, %f909, %f909;
|
|
mul.f32 %f911, %f894, %f910;
|
|
mul.f32 %f912, %f911, %f911;
|
|
mov.f32 %f913, 0f3C4CAF63;
|
|
mov.f32 %f914, 0f3B18F0FE;
|
|
fma.rn.f32 %f915, %f914, %f912, %f913;
|
|
mov.f32 %f916, 0f3DAAAABD;
|
|
fma.rn.f32 %f917, %f915, %f912, %f916;
|
|
mul.rn.f32 %f918, %f917, %f912;
|
|
mul.rn.f32 %f919, %f918, %f911;
|
|
sub.f32 %f920, %f909, %f911;
|
|
neg.f32 %f921, %f911;
|
|
add.f32 %f922, %f920, %f920;
|
|
fma.rn.f32 %f923, %f921, %f909, %f922;
|
|
mul.rn.f32 %f924, %f894, %f923;
|
|
add.f32 %f925, %f919, %f911;
|
|
sub.f32 %f926, %f911, %f925;
|
|
add.f32 %f927, %f919, %f926;
|
|
add.f32 %f928, %f924, %f927;
|
|
add.f32 %f929, %f925, %f928;
|
|
sub.f32 %f930, %f925, %f929;
|
|
add.f32 %f931, %f928, %f930;
|
|
mov.f32 %f932, 0f3F317200;
|
|
mul.rn.f32 %f933, %f908, %f932;
|
|
mov.f32 %f934, 0f35BFBE8E;
|
|
mul.rn.f32 %f935, %f908, %f934;
|
|
add.f32 %f936, %f933, %f929;
|
|
sub.f32 %f937, %f933, %f936;
|
|
add.f32 %f938, %f929, %f937;
|
|
add.f32 %f939, %f931, %f938;
|
|
add.f32 %f940, %f935, %f939;
|
|
add.f32 %f941, %f936, %f940;
|
|
sub.f32 %f942, %f936, %f941;
|
|
add.f32 %f943, %f940, %f942;
|
|
mov.f32 %f944, 0f3EE8BA2E;
|
|
mul.rn.f32 %f945, %f944, %f941;
|
|
neg.f32 %f946, %f945;
|
|
fma.rn.f32 %f947, %f944, %f941, %f946;
|
|
fma.rn.f32 %f948, %f944, %f943, %f947;
|
|
mov.f32 %f949, 0f00000000;
|
|
fma.rn.f32 %f950, %f949, %f941, %f948;
|
|
add.rn.f32 %f951, %f945, %f950;
|
|
neg.f32 %f952, %f951;
|
|
add.rn.f32 %f953, %f945, %f952;
|
|
add.rn.f32 %f954, %f953, %f950;
|
|
mov.b32 %r127, %f951;
|
|
setp.eq.s32 %p62, %r127, 1118925336;
|
|
add.s32 %r128, %r127, -1;
|
|
mov.b32 %f955, %r128;
|
|
add.f32 %f956, %f954, 0f37000000;
|
|
selp.f32 %f957, %f955, %f951, %p62;
|
|
selp.f32 %f189, %f956, %f954, %p62;
|
|
mul.f32 %f958, %f957, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f959, %f958;
|
|
mov.f32 %f960, 0fBF317200;
|
|
fma.rn.f32 %f961, %f959, %f960, %f957;
|
|
mov.f32 %f962, 0fB5BFBE8E;
|
|
fma.rn.f32 %f963, %f959, %f962, %f961;
|
|
mul.f32 %f964, %f963, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f965, %f964;
|
|
add.f32 %f966, %f959, 0f00000000;
|
|
ex2.approx.f32 %f967, %f966;
|
|
mul.f32 %f968, %f965, %f967;
|
|
setp.lt.f32 %p63, %f957, 0fC2D20000;
|
|
selp.f32 %f969, 0f00000000, %f968, %p63;
|
|
setp.gt.f32 %p64, %f957, 0f42D20000;
|
|
selp.f32 %f1321, 0f7F800000, %f969, %p64;
|
|
setp.eq.f32 %p65, %f1321, 0f7F800000;
|
|
@%p65 bra BB0_52;
|
|
|
|
fma.rn.f32 %f1321, %f1321, %f189, %f1321;
|
|
|
|
BB0_52:
|
|
setp.lt.f32 %p66, %f186, 0f00000000;
|
|
setp.eq.f32 %p67, %f187, 0f3F800000;
|
|
and.pred %p4, %p66, %p67;
|
|
mov.b32 %r129, %f1321;
|
|
xor.b32 %r130, %r129, -2147483648;
|
|
mov.b32 %f970, %r130;
|
|
selp.f32 %f1323, %f970, %f1321, %p4;
|
|
setp.eq.f32 %p68, %f186, 0f00000000;
|
|
@%p68 bra BB0_55;
|
|
bra.uni BB0_53;
|
|
|
|
BB0_55:
|
|
add.f32 %f973, %f186, %f186;
|
|
selp.f32 %f1323, %f973, 0f00000000, %p67;
|
|
bra.uni BB0_56;
|
|
|
|
BB0_110:
|
|
mov.u64 %rd117, image_HDR;
|
|
cvta.global.u64 %rd112, %rd117;
|
|
mov.u32 %r193, 8;
|
|
// inline asm
|
|
call (%rd111), _rt_buffer_get_64, (%rd112, %r28, %r193, %rd6, %rd7, %rd13, %rd13);
|
|
// inline asm
|
|
mov.f32 %f1254, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs57, %f1254;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs58, 0;
|
|
st.v4.u16 [%rd111], {%rs57, %rs57, %rs57, %rs58};
|
|
|
|
BB0_111:
|
|
ld.global.u8 %rs59, [imageEnabled];
|
|
and.b16 %rs60, %rs59, 64;
|
|
setp.eq.s16 %p132, %rs60, 0;
|
|
@%p132 bra BB0_113;
|
|
|
|
cvt.u64.u32 %rd120, %r2;
|
|
cvt.u64.u32 %rd121, %r3;
|
|
mov.u64 %rd124, image_Dir;
|
|
cvta.global.u64 %rd119, %rd124;
|
|
// inline asm
|
|
call (%rd118), _rt_buffer_get_64, (%rd119, %r28, %r29, %rd120, %rd121, %rd13, %rd13);
|
|
// inline asm
|
|
mov.u16 %rs61, 0;
|
|
st.v4.u8 [%rd118], {%rs61, %rs61, %rs61, %rs61};
|
|
bra.uni BB0_113;
|
|
|
|
BB0_53:
|
|
setp.geu.f32 %p69, %f186, 0f00000000;
|
|
@%p69 bra BB0_56;
|
|
|
|
cvt.rzi.f32.f32 %f972, %f944;
|
|
setp.neu.f32 %p70, %f972, 0f3EE8BA2E;
|
|
selp.f32 %f1323, 0f7FFFFFFF, %f1323, %p70;
|
|
|
|
BB0_56:
|
|
add.f32 %f974, %f188, 0f3EE8BA2E;
|
|
mov.b32 %r131, %f974;
|
|
setp.lt.s32 %p72, %r131, 2139095040;
|
|
@%p72 bra BB0_61;
|
|
|
|
setp.gtu.f32 %p73, %f188, 0f7F800000;
|
|
@%p73 bra BB0_60;
|
|
bra.uni BB0_58;
|
|
|
|
BB0_60:
|
|
add.f32 %f1323, %f186, 0f3EE8BA2E;
|
|
bra.uni BB0_61;
|
|
|
|
BB0_58:
|
|
setp.neu.f32 %p74, %f188, 0f7F800000;
|
|
@%p74 bra BB0_61;
|
|
|
|
selp.f32 %f1323, 0fFF800000, 0f7F800000, %p4;
|
|
|
|
BB0_61:
|
|
mul.f32 %f975, %f1323, 0f437F0000;
|
|
setp.eq.f32 %p75, %f186, 0f3F800000;
|
|
selp.f32 %f976, 0f437F0000, %f975, %p75;
|
|
cvt.rzi.u32.f32 %r132, %f976;
|
|
cvt.u16.u32 %rs17, %r132;
|
|
mov.u16 %rs18, 255;
|
|
st.v2.u8 [%rd43], {%rs17, %rs18};
|
|
ld.global.u32 %r204, [imageEnabled];
|
|
|
|
BB0_62:
|
|
and.b32 %r133, %r204, 1;
|
|
setp.eq.b32 %p76, %r133, 1;
|
|
@!%p76 bra BB0_97;
|
|
bra.uni BB0_63;
|
|
|
|
BB0_63:
|
|
mov.f32 %f979, 0f3E666666;
|
|
cvt.rzi.f32.f32 %f980, %f979;
|
|
fma.rn.f32 %f981, %f980, 0fC0000000, 0f3EE66666;
|
|
abs.f32 %f200, %f981;
|
|
abs.f32 %f201, %f1267;
|
|
setp.lt.f32 %p77, %f201, 0f00800000;
|
|
mul.f32 %f982, %f201, 0f4B800000;
|
|
selp.f32 %f983, 0fC3170000, 0fC2FE0000, %p77;
|
|
selp.f32 %f984, %f982, %f201, %p77;
|
|
mov.b32 %r134, %f984;
|
|
and.b32 %r135, %r134, 8388607;
|
|
or.b32 %r136, %r135, 1065353216;
|
|
mov.b32 %f985, %r136;
|
|
shr.u32 %r137, %r134, 23;
|
|
cvt.rn.f32.u32 %f986, %r137;
|
|
add.f32 %f987, %f983, %f986;
|
|
setp.gt.f32 %p78, %f985, 0f3FB504F3;
|
|
mul.f32 %f988, %f985, 0f3F000000;
|
|
add.f32 %f989, %f987, 0f3F800000;
|
|
selp.f32 %f990, %f988, %f985, %p78;
|
|
selp.f32 %f991, %f989, %f987, %p78;
|
|
add.f32 %f992, %f990, 0fBF800000;
|
|
add.f32 %f978, %f990, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f977,%f978;
|
|
// inline asm
|
|
add.f32 %f993, %f992, %f992;
|
|
mul.f32 %f994, %f977, %f993;
|
|
mul.f32 %f995, %f994, %f994;
|
|
mov.f32 %f996, 0f3C4CAF63;
|
|
mov.f32 %f997, 0f3B18F0FE;
|
|
fma.rn.f32 %f998, %f997, %f995, %f996;
|
|
mov.f32 %f999, 0f3DAAAABD;
|
|
fma.rn.f32 %f1000, %f998, %f995, %f999;
|
|
mul.rn.f32 %f1001, %f1000, %f995;
|
|
mul.rn.f32 %f1002, %f1001, %f994;
|
|
sub.f32 %f1003, %f992, %f994;
|
|
neg.f32 %f1004, %f994;
|
|
add.f32 %f1005, %f1003, %f1003;
|
|
fma.rn.f32 %f1006, %f1004, %f992, %f1005;
|
|
mul.rn.f32 %f1007, %f977, %f1006;
|
|
add.f32 %f1008, %f1002, %f994;
|
|
sub.f32 %f1009, %f994, %f1008;
|
|
add.f32 %f1010, %f1002, %f1009;
|
|
add.f32 %f1011, %f1007, %f1010;
|
|
add.f32 %f1012, %f1008, %f1011;
|
|
sub.f32 %f1013, %f1008, %f1012;
|
|
add.f32 %f1014, %f1011, %f1013;
|
|
mov.f32 %f1015, 0f3F317200;
|
|
mul.rn.f32 %f1016, %f991, %f1015;
|
|
mov.f32 %f1017, 0f35BFBE8E;
|
|
mul.rn.f32 %f1018, %f991, %f1017;
|
|
add.f32 %f1019, %f1016, %f1012;
|
|
sub.f32 %f1020, %f1016, %f1019;
|
|
add.f32 %f1021, %f1012, %f1020;
|
|
add.f32 %f1022, %f1014, %f1021;
|
|
add.f32 %f1023, %f1018, %f1022;
|
|
add.f32 %f1024, %f1019, %f1023;
|
|
sub.f32 %f1025, %f1019, %f1024;
|
|
add.f32 %f1026, %f1023, %f1025;
|
|
mov.f32 %f1027, 0f3EE66666;
|
|
mul.rn.f32 %f1028, %f1027, %f1024;
|
|
neg.f32 %f1029, %f1028;
|
|
fma.rn.f32 %f1030, %f1027, %f1024, %f1029;
|
|
fma.rn.f32 %f1031, %f1027, %f1026, %f1030;
|
|
mov.f32 %f1032, 0f00000000;
|
|
fma.rn.f32 %f1033, %f1032, %f1024, %f1031;
|
|
add.rn.f32 %f1034, %f1028, %f1033;
|
|
neg.f32 %f1035, %f1034;
|
|
add.rn.f32 %f1036, %f1028, %f1035;
|
|
add.rn.f32 %f1037, %f1036, %f1033;
|
|
mov.b32 %r138, %f1034;
|
|
setp.eq.s32 %p79, %r138, 1118925336;
|
|
add.s32 %r139, %r138, -1;
|
|
mov.b32 %f1038, %r139;
|
|
add.f32 %f1039, %f1037, 0f37000000;
|
|
selp.f32 %f1040, %f1038, %f1034, %p79;
|
|
selp.f32 %f202, %f1039, %f1037, %p79;
|
|
mul.f32 %f1041, %f1040, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f1042, %f1041;
|
|
mov.f32 %f1043, 0fBF317200;
|
|
fma.rn.f32 %f1044, %f1042, %f1043, %f1040;
|
|
mov.f32 %f1045, 0fB5BFBE8E;
|
|
fma.rn.f32 %f1046, %f1042, %f1045, %f1044;
|
|
mul.f32 %f1047, %f1046, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f1048, %f1047;
|
|
add.f32 %f1049, %f1042, 0f00000000;
|
|
ex2.approx.f32 %f1050, %f1049;
|
|
mul.f32 %f1051, %f1048, %f1050;
|
|
setp.lt.f32 %p80, %f1040, 0fC2D20000;
|
|
selp.f32 %f1052, 0f00000000, %f1051, %p80;
|
|
setp.gt.f32 %p81, %f1040, 0f42D20000;
|
|
selp.f32 %f1324, 0f7F800000, %f1052, %p81;
|
|
setp.eq.f32 %p82, %f1324, 0f7F800000;
|
|
@%p82 bra BB0_65;
|
|
|
|
fma.rn.f32 %f1324, %f1324, %f202, %f1324;
|
|
|
|
BB0_65:
|
|
setp.lt.f32 %p83, %f1267, 0f00000000;
|
|
setp.eq.f32 %p84, %f200, 0f3F800000;
|
|
and.pred %p5, %p83, %p84;
|
|
mov.b32 %r140, %f1324;
|
|
xor.b32 %r141, %r140, -2147483648;
|
|
mov.b32 %f1053, %r141;
|
|
selp.f32 %f1326, %f1053, %f1324, %p5;
|
|
setp.eq.f32 %p85, %f1267, 0f00000000;
|
|
@%p85 bra BB0_68;
|
|
bra.uni BB0_66;
|
|
|
|
BB0_68:
|
|
add.f32 %f1056, %f1267, %f1267;
|
|
selp.f32 %f1326, %f1056, 0f00000000, %p84;
|
|
bra.uni BB0_69;
|
|
|
|
BB0_66:
|
|
setp.geu.f32 %p86, %f1267, 0f00000000;
|
|
@%p86 bra BB0_69;
|
|
|
|
cvt.rzi.f32.f32 %f1055, %f1027;
|
|
setp.neu.f32 %p87, %f1055, 0f3EE66666;
|
|
selp.f32 %f1326, 0f7FFFFFFF, %f1326, %p87;
|
|
|
|
BB0_69:
|
|
add.f32 %f1057, %f201, 0f3EE66666;
|
|
mov.b32 %r142, %f1057;
|
|
setp.lt.s32 %p89, %r142, 2139095040;
|
|
@%p89 bra BB0_74;
|
|
|
|
setp.gtu.f32 %p90, %f201, 0f7F800000;
|
|
@%p90 bra BB0_73;
|
|
bra.uni BB0_71;
|
|
|
|
BB0_73:
|
|
add.f32 %f1326, %f1267, 0f3EE66666;
|
|
bra.uni BB0_74;
|
|
|
|
BB0_71:
|
|
setp.neu.f32 %p91, %f201, 0f7F800000;
|
|
@%p91 bra BB0_74;
|
|
|
|
selp.f32 %f1326, 0fFF800000, 0f7F800000, %p5;
|
|
|
|
BB0_74:
|
|
setp.eq.f32 %p92, %f1267, 0f3F800000;
|
|
selp.f32 %f213, 0f3F800000, %f1326, %p92;
|
|
abs.f32 %f214, %f1266;
|
|
setp.lt.f32 %p93, %f214, 0f00800000;
|
|
mul.f32 %f1060, %f214, 0f4B800000;
|
|
selp.f32 %f1061, 0fC3170000, 0fC2FE0000, %p93;
|
|
selp.f32 %f1062, %f1060, %f214, %p93;
|
|
mov.b32 %r143, %f1062;
|
|
and.b32 %r144, %r143, 8388607;
|
|
or.b32 %r145, %r144, 1065353216;
|
|
mov.b32 %f1063, %r145;
|
|
shr.u32 %r146, %r143, 23;
|
|
cvt.rn.f32.u32 %f1064, %r146;
|
|
add.f32 %f1065, %f1061, %f1064;
|
|
setp.gt.f32 %p94, %f1063, 0f3FB504F3;
|
|
mul.f32 %f1066, %f1063, 0f3F000000;
|
|
add.f32 %f1067, %f1065, 0f3F800000;
|
|
selp.f32 %f1068, %f1066, %f1063, %p94;
|
|
selp.f32 %f1069, %f1067, %f1065, %p94;
|
|
add.f32 %f1070, %f1068, 0fBF800000;
|
|
add.f32 %f1059, %f1068, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f1058,%f1059;
|
|
// inline asm
|
|
add.f32 %f1071, %f1070, %f1070;
|
|
mul.f32 %f1072, %f1058, %f1071;
|
|
mul.f32 %f1073, %f1072, %f1072;
|
|
fma.rn.f32 %f1076, %f997, %f1073, %f996;
|
|
fma.rn.f32 %f1078, %f1076, %f1073, %f999;
|
|
mul.rn.f32 %f1079, %f1078, %f1073;
|
|
mul.rn.f32 %f1080, %f1079, %f1072;
|
|
sub.f32 %f1081, %f1070, %f1072;
|
|
neg.f32 %f1082, %f1072;
|
|
add.f32 %f1083, %f1081, %f1081;
|
|
fma.rn.f32 %f1084, %f1082, %f1070, %f1083;
|
|
mul.rn.f32 %f1085, %f1058, %f1084;
|
|
add.f32 %f1086, %f1080, %f1072;
|
|
sub.f32 %f1087, %f1072, %f1086;
|
|
add.f32 %f1088, %f1080, %f1087;
|
|
add.f32 %f1089, %f1085, %f1088;
|
|
add.f32 %f1090, %f1086, %f1089;
|
|
sub.f32 %f1091, %f1086, %f1090;
|
|
add.f32 %f1092, %f1089, %f1091;
|
|
mul.rn.f32 %f1094, %f1069, %f1015;
|
|
mul.rn.f32 %f1096, %f1069, %f1017;
|
|
add.f32 %f1097, %f1094, %f1090;
|
|
sub.f32 %f1098, %f1094, %f1097;
|
|
add.f32 %f1099, %f1090, %f1098;
|
|
add.f32 %f1100, %f1092, %f1099;
|
|
add.f32 %f1101, %f1096, %f1100;
|
|
add.f32 %f1102, %f1097, %f1101;
|
|
sub.f32 %f1103, %f1097, %f1102;
|
|
add.f32 %f1104, %f1101, %f1103;
|
|
mul.rn.f32 %f1106, %f1027, %f1102;
|
|
neg.f32 %f1107, %f1106;
|
|
fma.rn.f32 %f1108, %f1027, %f1102, %f1107;
|
|
fma.rn.f32 %f1109, %f1027, %f1104, %f1108;
|
|
fma.rn.f32 %f1111, %f1032, %f1102, %f1109;
|
|
add.rn.f32 %f1112, %f1106, %f1111;
|
|
neg.f32 %f1113, %f1112;
|
|
add.rn.f32 %f1114, %f1106, %f1113;
|
|
add.rn.f32 %f1115, %f1114, %f1111;
|
|
mov.b32 %r147, %f1112;
|
|
setp.eq.s32 %p95, %r147, 1118925336;
|
|
add.s32 %r148, %r147, -1;
|
|
mov.b32 %f1116, %r148;
|
|
add.f32 %f1117, %f1115, 0f37000000;
|
|
selp.f32 %f1118, %f1116, %f1112, %p95;
|
|
selp.f32 %f215, %f1117, %f1115, %p95;
|
|
mul.f32 %f1119, %f1118, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f1120, %f1119;
|
|
fma.rn.f32 %f1122, %f1120, %f1043, %f1118;
|
|
fma.rn.f32 %f1124, %f1120, %f1045, %f1122;
|
|
mul.f32 %f1125, %f1124, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f1126, %f1125;
|
|
add.f32 %f1127, %f1120, 0f00000000;
|
|
ex2.approx.f32 %f1128, %f1127;
|
|
mul.f32 %f1129, %f1126, %f1128;
|
|
setp.lt.f32 %p96, %f1118, 0fC2D20000;
|
|
selp.f32 %f1130, 0f00000000, %f1129, %p96;
|
|
setp.gt.f32 %p97, %f1118, 0f42D20000;
|
|
selp.f32 %f1327, 0f7F800000, %f1130, %p97;
|
|
setp.eq.f32 %p98, %f1327, 0f7F800000;
|
|
@%p98 bra BB0_76;
|
|
|
|
fma.rn.f32 %f1327, %f1327, %f215, %f1327;
|
|
|
|
BB0_76:
|
|
setp.lt.f32 %p99, %f1266, 0f00000000;
|
|
and.pred %p6, %p99, %p84;
|
|
mov.b32 %r149, %f1327;
|
|
xor.b32 %r150, %r149, -2147483648;
|
|
mov.b32 %f1131, %r150;
|
|
selp.f32 %f1329, %f1131, %f1327, %p6;
|
|
setp.eq.f32 %p101, %f1266, 0f00000000;
|
|
@%p101 bra BB0_79;
|
|
bra.uni BB0_77;
|
|
|
|
BB0_79:
|
|
add.f32 %f1134, %f1266, %f1266;
|
|
selp.f32 %f1329, %f1134, 0f00000000, %p84;
|
|
bra.uni BB0_80;
|
|
|
|
BB0_77:
|
|
setp.geu.f32 %p102, %f1266, 0f00000000;
|
|
@%p102 bra BB0_80;
|
|
|
|
cvt.rzi.f32.f32 %f1133, %f1027;
|
|
setp.neu.f32 %p103, %f1133, 0f3EE66666;
|
|
selp.f32 %f1329, 0f7FFFFFFF, %f1329, %p103;
|
|
|
|
BB0_80:
|
|
add.f32 %f1135, %f214, 0f3EE66666;
|
|
mov.b32 %r151, %f1135;
|
|
setp.lt.s32 %p105, %r151, 2139095040;
|
|
@%p105 bra BB0_85;
|
|
|
|
setp.gtu.f32 %p106, %f214, 0f7F800000;
|
|
@%p106 bra BB0_84;
|
|
bra.uni BB0_82;
|
|
|
|
BB0_84:
|
|
add.f32 %f1329, %f1266, 0f3EE66666;
|
|
bra.uni BB0_85;
|
|
|
|
BB0_82:
|
|
setp.neu.f32 %p107, %f214, 0f7F800000;
|
|
@%p107 bra BB0_85;
|
|
|
|
selp.f32 %f1329, 0fFF800000, 0f7F800000, %p6;
|
|
|
|
BB0_85:
|
|
setp.eq.f32 %p108, %f1266, 0f3F800000;
|
|
selp.f32 %f226, 0f3F800000, %f1329, %p108;
|
|
abs.f32 %f227, %f1265;
|
|
setp.lt.f32 %p109, %f227, 0f00800000;
|
|
mul.f32 %f1138, %f227, 0f4B800000;
|
|
selp.f32 %f1139, 0fC3170000, 0fC2FE0000, %p109;
|
|
selp.f32 %f1140, %f1138, %f227, %p109;
|
|
mov.b32 %r152, %f1140;
|
|
and.b32 %r153, %r152, 8388607;
|
|
or.b32 %r154, %r153, 1065353216;
|
|
mov.b32 %f1141, %r154;
|
|
shr.u32 %r155, %r152, 23;
|
|
cvt.rn.f32.u32 %f1142, %r155;
|
|
add.f32 %f1143, %f1139, %f1142;
|
|
setp.gt.f32 %p110, %f1141, 0f3FB504F3;
|
|
mul.f32 %f1144, %f1141, 0f3F000000;
|
|
add.f32 %f1145, %f1143, 0f3F800000;
|
|
selp.f32 %f1146, %f1144, %f1141, %p110;
|
|
selp.f32 %f1147, %f1145, %f1143, %p110;
|
|
add.f32 %f1148, %f1146, 0fBF800000;
|
|
add.f32 %f1137, %f1146, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f1136,%f1137;
|
|
// inline asm
|
|
add.f32 %f1149, %f1148, %f1148;
|
|
mul.f32 %f1150, %f1136, %f1149;
|
|
mul.f32 %f1151, %f1150, %f1150;
|
|
fma.rn.f32 %f1154, %f997, %f1151, %f996;
|
|
fma.rn.f32 %f1156, %f1154, %f1151, %f999;
|
|
mul.rn.f32 %f1157, %f1156, %f1151;
|
|
mul.rn.f32 %f1158, %f1157, %f1150;
|
|
sub.f32 %f1159, %f1148, %f1150;
|
|
neg.f32 %f1160, %f1150;
|
|
add.f32 %f1161, %f1159, %f1159;
|
|
fma.rn.f32 %f1162, %f1160, %f1148, %f1161;
|
|
mul.rn.f32 %f1163, %f1136, %f1162;
|
|
add.f32 %f1164, %f1158, %f1150;
|
|
sub.f32 %f1165, %f1150, %f1164;
|
|
add.f32 %f1166, %f1158, %f1165;
|
|
add.f32 %f1167, %f1163, %f1166;
|
|
add.f32 %f1168, %f1164, %f1167;
|
|
sub.f32 %f1169, %f1164, %f1168;
|
|
add.f32 %f1170, %f1167, %f1169;
|
|
mul.rn.f32 %f1172, %f1147, %f1015;
|
|
mul.rn.f32 %f1174, %f1147, %f1017;
|
|
add.f32 %f1175, %f1172, %f1168;
|
|
sub.f32 %f1176, %f1172, %f1175;
|
|
add.f32 %f1177, %f1168, %f1176;
|
|
add.f32 %f1178, %f1170, %f1177;
|
|
add.f32 %f1179, %f1174, %f1178;
|
|
add.f32 %f1180, %f1175, %f1179;
|
|
sub.f32 %f1181, %f1175, %f1180;
|
|
add.f32 %f1182, %f1179, %f1181;
|
|
mul.rn.f32 %f1184, %f1027, %f1180;
|
|
neg.f32 %f1185, %f1184;
|
|
fma.rn.f32 %f1186, %f1027, %f1180, %f1185;
|
|
fma.rn.f32 %f1187, %f1027, %f1182, %f1186;
|
|
fma.rn.f32 %f1189, %f1032, %f1180, %f1187;
|
|
add.rn.f32 %f1190, %f1184, %f1189;
|
|
neg.f32 %f1191, %f1190;
|
|
add.rn.f32 %f1192, %f1184, %f1191;
|
|
add.rn.f32 %f1193, %f1192, %f1189;
|
|
mov.b32 %r156, %f1190;
|
|
setp.eq.s32 %p111, %r156, 1118925336;
|
|
add.s32 %r157, %r156, -1;
|
|
mov.b32 %f1194, %r157;
|
|
add.f32 %f1195, %f1193, 0f37000000;
|
|
selp.f32 %f1196, %f1194, %f1190, %p111;
|
|
selp.f32 %f228, %f1195, %f1193, %p111;
|
|
mul.f32 %f1197, %f1196, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f1198, %f1197;
|
|
fma.rn.f32 %f1200, %f1198, %f1043, %f1196;
|
|
fma.rn.f32 %f1202, %f1198, %f1045, %f1200;
|
|
mul.f32 %f1203, %f1202, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f1204, %f1203;
|
|
add.f32 %f1205, %f1198, 0f00000000;
|
|
ex2.approx.f32 %f1206, %f1205;
|
|
mul.f32 %f1207, %f1204, %f1206;
|
|
setp.lt.f32 %p112, %f1196, 0fC2D20000;
|
|
selp.f32 %f1208, 0f00000000, %f1207, %p112;
|
|
setp.gt.f32 %p113, %f1196, 0f42D20000;
|
|
selp.f32 %f1330, 0f7F800000, %f1208, %p113;
|
|
setp.eq.f32 %p114, %f1330, 0f7F800000;
|
|
@%p114 bra BB0_87;
|
|
|
|
fma.rn.f32 %f1330, %f1330, %f228, %f1330;
|
|
|
|
BB0_87:
|
|
setp.lt.f32 %p115, %f1265, 0f00000000;
|
|
and.pred %p7, %p115, %p84;
|
|
mov.b32 %r158, %f1330;
|
|
xor.b32 %r159, %r158, -2147483648;
|
|
mov.b32 %f1209, %r159;
|
|
selp.f32 %f1332, %f1209, %f1330, %p7;
|
|
setp.eq.f32 %p117, %f1265, 0f00000000;
|
|
@%p117 bra BB0_90;
|
|
bra.uni BB0_88;
|
|
|
|
BB0_90:
|
|
add.f32 %f1212, %f1265, %f1265;
|
|
selp.f32 %f1332, %f1212, 0f00000000, %p84;
|
|
bra.uni BB0_91;
|
|
|
|
BB0_88:
|
|
setp.geu.f32 %p118, %f1265, 0f00000000;
|
|
@%p118 bra BB0_91;
|
|
|
|
cvt.rzi.f32.f32 %f1211, %f1027;
|
|
setp.neu.f32 %p119, %f1211, 0f3EE66666;
|
|
selp.f32 %f1332, 0f7FFFFFFF, %f1332, %p119;
|
|
|
|
BB0_91:
|
|
add.f32 %f1213, %f227, 0f3EE66666;
|
|
mov.b32 %r160, %f1213;
|
|
setp.lt.s32 %p121, %r160, 2139095040;
|
|
@%p121 bra BB0_96;
|
|
|
|
setp.gtu.f32 %p122, %f227, 0f7F800000;
|
|
@%p122 bra BB0_95;
|
|
bra.uni BB0_93;
|
|
|
|
BB0_95:
|
|
add.f32 %f1332, %f1265, 0f3EE66666;
|
|
bra.uni BB0_96;
|
|
|
|
BB0_93:
|
|
setp.neu.f32 %p123, %f227, 0f7F800000;
|
|
@%p123 bra BB0_96;
|
|
|
|
selp.f32 %f1332, 0fFF800000, 0f7F800000, %p7;
|
|
|
|
BB0_96:
|
|
mov.u32 %r196, 4;
|
|
setp.eq.f32 %p124, %f1265, 0f3F800000;
|
|
selp.f32 %f1214, 0f3F800000, %f1332, %p124;
|
|
cvt.u64.u32 %rd53, %r3;
|
|
cvt.u64.u32 %rd52, %r2;
|
|
mov.u64 %rd56, image;
|
|
cvta.global.u64 %rd51, %rd56;
|
|
// inline asm
|
|
call (%rd50), _rt_buffer_get_64, (%rd51, %r28, %r196, %rd52, %rd53, %rd13, %rd13);
|
|
// inline asm
|
|
cvt.sat.f32.f32 %f1215, %f1214;
|
|
mul.f32 %f1216, %f1215, 0f437FFD71;
|
|
cvt.rzi.u32.f32 %r163, %f1216;
|
|
cvt.sat.f32.f32 %f1217, %f226;
|
|
mul.f32 %f1218, %f1217, 0f437FFD71;
|
|
cvt.rzi.u32.f32 %r164, %f1218;
|
|
cvt.sat.f32.f32 %f1219, %f213;
|
|
mul.f32 %f1220, %f1219, 0f437FFD71;
|
|
cvt.rzi.u32.f32 %r165, %f1220;
|
|
cvt.u16.u32 %rs19, %r163;
|
|
cvt.u16.u32 %rs20, %r165;
|
|
cvt.u16.u32 %rs21, %r164;
|
|
mov.u16 %rs22, 255;
|
|
st.v4.u8 [%rd50], {%rs19, %rs21, %rs20, %rs22};
|
|
ld.global.u32 %r204, [imageEnabled];
|
|
|
|
BB0_97:
|
|
and.b32 %r166, %r204, 4;
|
|
setp.eq.s32 %p125, %r166, 0;
|
|
@%p125 bra BB0_101;
|
|
|
|
ld.global.u32 %r167, [additive];
|
|
setp.eq.s32 %p126, %r167, 0;
|
|
cvt.u64.u32 %rd4, %r2;
|
|
cvt.u64.u32 %rd5, %r3;
|
|
mov.f32 %f1221, 0f3F800000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs23, %f1221;}
|
|
|
|
// inline asm
|
|
@%p126 bra BB0_100;
|
|
|
|
mov.u64 %rd69, image_HDR;
|
|
cvta.global.u64 %rd58, %rd69;
|
|
mov.u32 %r171, 8;
|
|
// inline asm
|
|
call (%rd57), _rt_buffer_get_64, (%rd58, %r28, %r171, %rd4, %rd5, %rd13, %rd13);
|
|
// inline asm
|
|
ld.v4.u16 {%rs30, %rs31, %rs32, %rs33}, [%rd57];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1222, %rs30;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1223, %rs31;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1224, %rs32;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd63), _rt_buffer_get_64, (%rd58, %r28, %r171, %rd4, %rd5, %rd13, %rd13);
|
|
// inline asm
|
|
add.f32 %f1225, %f1267, %f1222;
|
|
add.f32 %f1226, %f1266, %f1223;
|
|
add.f32 %f1227, %f1265, %f1224;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs29, %f1227;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs28, %f1226;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs27, %f1225;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd63], {%rs27, %rs28, %rs29, %rs23};
|
|
bra.uni BB0_101;
|
|
|
|
BB0_100:
|
|
mov.u64 %rd76, image_HDR;
|
|
cvta.global.u64 %rd71, %rd76;
|
|
mov.u32 %r173, 8;
|
|
// inline asm
|
|
call (%rd70), _rt_buffer_get_64, (%rd71, %r28, %r173, %rd4, %rd5, %rd13, %rd13);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs36, %f1265;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs35, %f1266;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs34, %f1267;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd70], {%rs34, %rs35, %rs36, %rs23};
|
|
|
|
BB0_101:
|
|
ld.global.u8 %rs37, [imageEnabled];
|
|
and.b16 %rs38, %rs37, 64;
|
|
setp.eq.s16 %p127, %rs38, 0;
|
|
@%p127 bra BB0_113;
|
|
|
|
mov.u32 %r197, 4;
|
|
mul.f32 %f1231, %f23, %f23;
|
|
fma.rn.f32 %f1232, %f24, %f24, %f1231;
|
|
fma.rn.f32 %f1233, %f22, %f22, %f1232;
|
|
sqrt.rn.f32 %f1234, %f1233;
|
|
rcp.rn.f32 %f1235, %f1234;
|
|
mul.f32 %f1236, %f24, %f1235;
|
|
mul.f32 %f1237, %f23, %f1235;
|
|
mul.f32 %f1238, %f22, %f1235;
|
|
cvt.u64.u32 %rd80, %r3;
|
|
cvt.u64.u32 %rd79, %r2;
|
|
mov.u64 %rd83, image_Dir;
|
|
cvta.global.u64 %rd78, %rd83;
|
|
// inline asm
|
|
call (%rd77), _rt_buffer_get_64, (%rd78, %r28, %r197, %rd79, %rd80, %rd13, %rd13);
|
|
// inline asm
|
|
fma.rn.f32 %f1239, %f1236, 0f3F000000, 0f3F000000;
|
|
mul.f32 %f1240, %f1239, 0f437F0000;
|
|
cvt.rzi.u32.f32 %r176, %f1240;
|
|
fma.rn.f32 %f1241, %f1237, 0f3F000000, 0f3F000000;
|
|
mul.f32 %f1242, %f1241, 0f437F0000;
|
|
cvt.rzi.u32.f32 %r177, %f1242;
|
|
fma.rn.f32 %f1243, %f1238, 0f3F000000, 0f3F000000;
|
|
mul.f32 %f1244, %f1243, 0f437F0000;
|
|
cvt.rzi.u32.f32 %r178, %f1244;
|
|
cvt.u16.u32 %rs39, %r178;
|
|
cvt.u16.u32 %rs40, %r177;
|
|
cvt.u16.u32 %rs41, %r176;
|
|
mov.u16 %rs42, 255;
|
|
st.v4.u8 [%rd77], {%rs41, %rs40, %rs39, %rs42};
|
|
|
|
BB0_113:
|
|
ret;
|
|
}
|
|
|
|
|