2635 lines
76 KiB
Plaintext
2635 lines
76 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_RNM0[1];
|
|
.global .align 1 .b8 image_RNM1[1];
|
|
.global .align 1 .b8 image_RNM2[1];
|
|
.global .align 1 .b8 image_RNM3[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[4];
|
|
.reg .b64 %SP;
|
|
.reg .b64 %SPL;
|
|
.reg .pred %p<166>;
|
|
.reg .b16 %rs<162>;
|
|
.reg .f32 %f<1559>;
|
|
.reg .b32 %r<269>;
|
|
.reg .b64 %rd<272>;
|
|
|
|
|
|
mov.u64 %rd271, __local_depot0;
|
|
cvta.local.u64 %SP, %rd271;
|
|
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 %rs9, %rs1, 255;
|
|
cvt.u16.u32 %rs10, %r1;
|
|
or.b16 %rs11, %rs10, %rs9;
|
|
setp.eq.s16 %p9, %rs11, 0;
|
|
mov.f32 %f1460, 0f00000000;
|
|
mov.f32 %f1461, %f1460;
|
|
mov.f32 %f1462, %f1460;
|
|
@%p9 bra BB0_2;
|
|
|
|
ld.u8 %rs12, [%rd8+1];
|
|
and.b16 %rs14, %rs10, 255;
|
|
cvt.rn.f32.u16 %f295, %rs14;
|
|
div.rn.f32 %f296, %f295, 0f437F0000;
|
|
fma.rn.f32 %f297, %f296, 0f40000000, 0fBF800000;
|
|
cvt.rn.f32.u16 %f298, %rs12;
|
|
div.rn.f32 %f299, %f298, 0f437F0000;
|
|
fma.rn.f32 %f300, %f299, 0f40000000, 0fBF800000;
|
|
cvt.rn.f32.u16 %f301, %rs9;
|
|
div.rn.f32 %f302, %f301, 0f437F0000;
|
|
fma.rn.f32 %f303, %f302, 0f40000000, 0fBF800000;
|
|
mul.f32 %f304, %f300, %f300;
|
|
fma.rn.f32 %f305, %f297, %f297, %f304;
|
|
fma.rn.f32 %f306, %f303, %f303, %f305;
|
|
sqrt.rn.f32 %f307, %f306;
|
|
rcp.rn.f32 %f308, %f307;
|
|
mul.f32 %f1460, %f297, %f308;
|
|
mul.f32 %f1461, %f300, %f308;
|
|
mul.f32 %f1462, %f303, %f308;
|
|
|
|
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 %p10, %f1461, 0f00000000;
|
|
setp.eq.f32 %p11, %f1460, 0f00000000;
|
|
and.pred %p12, %p11, %p10;
|
|
setp.eq.f32 %p13, %f1462, 0f00000000;
|
|
and.pred %p14, %p12, %p13;
|
|
@%p14 bra BB0_124;
|
|
bra.uni BB0_3;
|
|
|
|
BB0_124:
|
|
ld.global.u32 %r268, [imageEnabled];
|
|
and.b32 %r217, %r268, 1;
|
|
setp.eq.b32 %p157, %r217, 1;
|
|
@!%p157 bra BB0_126;
|
|
bra.uni BB0_125;
|
|
|
|
BB0_125:
|
|
cvt.u64.u32 %rd159, %r2;
|
|
cvt.u64.u32 %rd160, %r3;
|
|
mov.u64 %rd163, image;
|
|
cvta.global.u64 %rd158, %rd163;
|
|
// inline asm
|
|
call (%rd157), _rt_buffer_get_64, (%rd158, %r28, %r29, %rd159, %rd160, %rd13, %rd13);
|
|
// inline asm
|
|
mov.u16 %rs93, 0;
|
|
st.v4.u8 [%rd157], {%rs93, %rs93, %rs93, %rs93};
|
|
ld.global.u32 %r268, [imageEnabled];
|
|
|
|
BB0_126:
|
|
and.b32 %r220, %r268, 8;
|
|
setp.eq.s32 %p158, %r220, 0;
|
|
@%p158 bra BB0_128;
|
|
|
|
cvt.u64.u32 %rd167, %r3;
|
|
cvt.u64.u32 %rd166, %r2;
|
|
mov.u64 %rd170, image_Mask;
|
|
cvta.global.u64 %rd165, %rd170;
|
|
// inline asm
|
|
call (%rd164), _rt_buffer_get_64, (%rd165, %r28, %r28, %rd166, %rd167, %rd13, %rd13);
|
|
// inline asm
|
|
mov.f32 %f1414, 0f00000000;
|
|
cvt.rzi.u32.f32 %r223, %f1414;
|
|
cvt.u16.u32 %rs94, %r223;
|
|
mov.u16 %rs95, 0;
|
|
st.v2.u8 [%rd164], {%rs94, %rs95};
|
|
ld.global.u32 %r268, [imageEnabled];
|
|
|
|
BB0_128:
|
|
cvt.u64.u32 %rd6, %r2;
|
|
cvt.u64.u32 %rd7, %r3;
|
|
and.b32 %r224, %r268, 4;
|
|
setp.eq.s32 %p159, %r224, 0;
|
|
@%p159 bra BB0_132;
|
|
|
|
ld.global.u32 %r225, [additive];
|
|
setp.eq.s32 %p160, %r225, 0;
|
|
@%p160 bra BB0_131;
|
|
|
|
mov.u64 %rd183, image_HDR;
|
|
cvta.global.u64 %rd172, %rd183;
|
|
mov.u32 %r229, 8;
|
|
// inline asm
|
|
call (%rd171), _rt_buffer_get_64, (%rd172, %r28, %r229, %rd6, %rd7, %rd13, %rd13);
|
|
// inline asm
|
|
ld.v4.u16 {%rs102, %rs103, %rs104, %rs105}, [%rd171];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1415, %rs102;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1416, %rs103;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1417, %rs104;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd177), _rt_buffer_get_64, (%rd172, %r28, %r229, %rd6, %rd7, %rd13, %rd13);
|
|
// inline asm
|
|
add.f32 %f1418, %f1415, 0f00000000;
|
|
add.f32 %f1419, %f1416, 0f00000000;
|
|
add.f32 %f1420, %f1417, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs101, %f1420;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs100, %f1419;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs99, %f1418;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs106, 0;
|
|
st.v4.u16 [%rd177], {%rs99, %rs100, %rs101, %rs106};
|
|
bra.uni BB0_132;
|
|
|
|
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 %f325, %f7, 0f3456BF95;
|
|
mul.f32 %f326, %f8, 0f3456BF95;
|
|
mul.f32 %f327, %f9, 0f3456BF95;
|
|
abs.f32 %f328, %f1460;
|
|
div.rn.f32 %f329, %f325, %f328;
|
|
abs.f32 %f330, %f1461;
|
|
div.rn.f32 %f331, %f326, %f330;
|
|
abs.f32 %f332, %f1462;
|
|
div.rn.f32 %f333, %f327, %f332;
|
|
abs.f32 %f334, %f329;
|
|
abs.f32 %f335, %f331;
|
|
abs.f32 %f336, %f333;
|
|
mov.f32 %f337, 0f38D1B717;
|
|
max.f32 %f338, %f334, %f337;
|
|
max.f32 %f339, %f335, %f337;
|
|
max.f32 %f340, %f336, %f337;
|
|
fma.rn.f32 %f10, %f1460, %f338, %f7;
|
|
fma.rn.f32 %f11, %f1461, %f339, %f8;
|
|
fma.rn.f32 %f12, %f1462, %f340, %f9;
|
|
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 %p15, %r4, 0;
|
|
mov.f32 %f1463, 0f00000000;
|
|
mov.f32 %f1464, %f1463;
|
|
mov.f32 %f1465, %f1463;
|
|
mov.f32 %f1466, %f1463;
|
|
mov.f32 %f1467, %f1463;
|
|
mov.f32 %f1468, %f1463;
|
|
mov.f32 %f1469, %f1463;
|
|
mov.f32 %f1470, %f1463;
|
|
mov.f32 %f1471, %f1463;
|
|
mov.f32 %f1472, %f1463;
|
|
mov.f32 %f1473, %f1463;
|
|
mov.f32 %f1474, %f1463;
|
|
mov.f32 %f1475, %f1463;
|
|
mov.f32 %f1476, %f1463;
|
|
mov.f32 %f1477, %f1463;
|
|
mov.f32 %f1478, %f1463;
|
|
@%p15 bra BB0_60;
|
|
|
|
mov.f32 %f357, 0f40000000;
|
|
cvt.rzi.f32.f32 %f358, %f357;
|
|
add.f32 %f359, %f358, %f358;
|
|
mov.f32 %f360, 0f40800000;
|
|
sub.f32 %f361, %f360, %f359;
|
|
abs.f32 %f13, %f361;
|
|
mul.f32 %f14, %f10, 0f3456BF95;
|
|
mul.f32 %f15, %f11, 0f3456BF95;
|
|
mul.f32 %f16, %f12, 0f3456BF95;
|
|
mov.f32 %f356, 0f00000000;
|
|
mov.u32 %r260, 0;
|
|
abs.f32 %f555, %f14;
|
|
abs.f32 %f556, %f15;
|
|
max.f32 %f557, %f555, %f556;
|
|
abs.f32 %f558, %f16;
|
|
max.f32 %f559, %f557, %f558;
|
|
mov.f32 %f1463, %f356;
|
|
mov.f32 %f1464, %f356;
|
|
mov.f32 %f1465, %f356;
|
|
mov.f32 %f1466, %f356;
|
|
mov.f32 %f1467, %f356;
|
|
mov.f32 %f1468, %f356;
|
|
mov.f32 %f1469, %f356;
|
|
mov.f32 %f1470, %f356;
|
|
mov.f32 %f1471, %f356;
|
|
mov.f32 %f1472, %f356;
|
|
mov.f32 %f1473, %f356;
|
|
mov.f32 %f1474, %f356;
|
|
mov.f32 %f1475, %f356;
|
|
mov.f32 %f1476, %f356;
|
|
mov.f32 %f1477, %f356;
|
|
mov.f32 %f1478, %f356;
|
|
|
|
BB0_5:
|
|
cvt.u64.u32 %rd30, %r260;
|
|
// inline asm
|
|
call (%rd28), _rt_buffer_get_64, (%rd25, %r45, %r46, %rd30, %rd13, %rd13, %rd13);
|
|
// inline asm
|
|
ld.v4.f32 {%f364, %f365, %f366, %f367}, [%rd28+80];
|
|
ld.v4.f32 {%f368, %f369, %f370, %f371}, [%rd28+64];
|
|
ld.v4.f32 {%f372, %f373, %f374, %f375}, [%rd28+48];
|
|
ld.v4.f32 {%f376, %f1487, %f1488, %f379}, [%rd28+32];
|
|
ld.v4.f32 {%f380, %f381, %f382, %f383}, [%rd28+16];
|
|
ld.v4.f32 {%f384, %f385, %f386, %f387}, [%rd28];
|
|
mov.b32 %r6, %f367;
|
|
sub.f32 %f57, %f385, %f7;
|
|
sub.f32 %f59, %f386, %f8;
|
|
sub.f32 %f61, %f387, %f9;
|
|
mul.f32 %f389, %f59, %f59;
|
|
fma.rn.f32 %f390, %f57, %f57, %f389;
|
|
fma.rn.f32 %f391, %f61, %f61, %f390;
|
|
sqrt.rn.f32 %f62, %f391;
|
|
rcp.rn.f32 %f392, %f62;
|
|
mul.f32 %f63, %f57, %f392;
|
|
mul.f32 %f64, %f59, %f392;
|
|
mul.f32 %f65, %f61, %f392;
|
|
mul.f32 %f66, %f62, %f383;
|
|
abs.f32 %f67, %f66;
|
|
setp.lt.f32 %p16, %f67, 0f00800000;
|
|
mul.f32 %f393, %f67, 0f4B800000;
|
|
selp.f32 %f394, 0fC3170000, 0fC2FE0000, %p16;
|
|
selp.f32 %f395, %f393, %f67, %p16;
|
|
mov.b32 %r54, %f395;
|
|
and.b32 %r55, %r54, 8388607;
|
|
or.b32 %r56, %r55, 1065353216;
|
|
mov.b32 %f396, %r56;
|
|
shr.u32 %r57, %r54, 23;
|
|
cvt.rn.f32.u32 %f397, %r57;
|
|
add.f32 %f398, %f394, %f397;
|
|
setp.gt.f32 %p17, %f396, 0f3FB504F3;
|
|
mul.f32 %f399, %f396, 0f3F000000;
|
|
add.f32 %f400, %f398, 0f3F800000;
|
|
selp.f32 %f401, %f399, %f396, %p17;
|
|
selp.f32 %f402, %f400, %f398, %p17;
|
|
add.f32 %f403, %f401, 0fBF800000;
|
|
add.f32 %f363, %f401, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f362,%f363;
|
|
// inline asm
|
|
add.f32 %f404, %f403, %f403;
|
|
mul.f32 %f405, %f362, %f404;
|
|
mul.f32 %f406, %f405, %f405;
|
|
mov.f32 %f407, 0f3C4CAF63;
|
|
mov.f32 %f408, 0f3B18F0FE;
|
|
fma.rn.f32 %f409, %f408, %f406, %f407;
|
|
mov.f32 %f410, 0f3DAAAABD;
|
|
fma.rn.f32 %f411, %f409, %f406, %f410;
|
|
mul.rn.f32 %f412, %f411, %f406;
|
|
mul.rn.f32 %f413, %f412, %f405;
|
|
sub.f32 %f414, %f403, %f405;
|
|
neg.f32 %f415, %f405;
|
|
add.f32 %f416, %f414, %f414;
|
|
fma.rn.f32 %f417, %f415, %f403, %f416;
|
|
mul.rn.f32 %f418, %f362, %f417;
|
|
add.f32 %f419, %f413, %f405;
|
|
sub.f32 %f420, %f405, %f419;
|
|
add.f32 %f421, %f413, %f420;
|
|
add.f32 %f422, %f418, %f421;
|
|
add.f32 %f423, %f419, %f422;
|
|
sub.f32 %f424, %f419, %f423;
|
|
add.f32 %f425, %f422, %f424;
|
|
mov.f32 %f426, 0f3F317200;
|
|
mul.rn.f32 %f427, %f402, %f426;
|
|
mov.f32 %f428, 0f35BFBE8E;
|
|
mul.rn.f32 %f429, %f402, %f428;
|
|
add.f32 %f430, %f427, %f423;
|
|
sub.f32 %f431, %f427, %f430;
|
|
add.f32 %f432, %f423, %f431;
|
|
add.f32 %f433, %f425, %f432;
|
|
add.f32 %f434, %f429, %f433;
|
|
add.f32 %f435, %f430, %f434;
|
|
sub.f32 %f436, %f430, %f435;
|
|
add.f32 %f437, %f434, %f436;
|
|
mul.rn.f32 %f68, %f360, %f435;
|
|
neg.f32 %f439, %f68;
|
|
fma.rn.f32 %f440, %f360, %f435, %f439;
|
|
fma.rn.f32 %f441, %f360, %f437, %f440;
|
|
fma.rn.f32 %f69, %f356, %f435, %f441;
|
|
add.rn.f32 %f70, %f68, %f69;
|
|
mov.b32 %r58, %f70;
|
|
setp.eq.s32 %p1, %r58, 1118925336;
|
|
add.s32 %r59, %r58, -1;
|
|
mov.b32 %f443, %r59;
|
|
selp.f32 %f444, %f443, %f70, %p1;
|
|
mul.f32 %f445, %f444, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f446, %f445;
|
|
mov.f32 %f447, 0fBF317200;
|
|
fma.rn.f32 %f448, %f446, %f447, %f444;
|
|
mov.f32 %f449, 0fB5BFBE8E;
|
|
fma.rn.f32 %f450, %f446, %f449, %f448;
|
|
mul.f32 %f451, %f450, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f452, %f451;
|
|
add.f32 %f453, %f446, 0f00000000;
|
|
ex2.approx.f32 %f454, %f453;
|
|
mul.f32 %f455, %f452, %f454;
|
|
setp.lt.f32 %p18, %f444, 0fC2D20000;
|
|
selp.f32 %f456, 0f00000000, %f455, %p18;
|
|
setp.gt.f32 %p19, %f444, 0f42D20000;
|
|
selp.f32 %f1479, 0f7F800000, %f456, %p19;
|
|
setp.eq.f32 %p20, %f1479, 0f7F800000;
|
|
@%p20 bra BB0_7;
|
|
|
|
neg.f32 %f457, %f70;
|
|
add.rn.f32 %f458, %f68, %f457;
|
|
add.rn.f32 %f459, %f458, %f69;
|
|
add.f32 %f460, %f459, 0f37000000;
|
|
selp.f32 %f461, %f460, %f459, %p1;
|
|
fma.rn.f32 %f1479, %f1479, %f461, %f1479;
|
|
|
|
BB0_7:
|
|
setp.lt.f32 %p21, %f66, 0f00000000;
|
|
setp.eq.f32 %p22, %f13, 0f3F800000;
|
|
and.pred %p2, %p21, %p22;
|
|
mov.b32 %r60, %f1479;
|
|
xor.b32 %r61, %r60, -2147483648;
|
|
mov.b32 %f462, %r61;
|
|
selp.f32 %f1481, %f462, %f1479, %p2;
|
|
setp.eq.f32 %p23, %f66, 0f00000000;
|
|
@%p23 bra BB0_10;
|
|
bra.uni BB0_8;
|
|
|
|
BB0_10:
|
|
add.f32 %f465, %f66, %f66;
|
|
selp.f32 %f1481, %f465, 0f00000000, %p22;
|
|
bra.uni BB0_11;
|
|
|
|
BB0_8:
|
|
setp.geu.f32 %p24, %f66, 0f00000000;
|
|
@%p24 bra BB0_11;
|
|
|
|
cvt.rzi.f32.f32 %f464, %f360;
|
|
setp.neu.f32 %p25, %f464, 0f40800000;
|
|
selp.f32 %f1481, 0f7FFFFFFF, %f1481, %p25;
|
|
|
|
BB0_11:
|
|
add.f32 %f466, %f67, 0f40800000;
|
|
mov.b32 %r62, %f466;
|
|
setp.lt.s32 %p27, %r62, 2139095040;
|
|
@%p27 bra BB0_16;
|
|
|
|
setp.gtu.f32 %p28, %f67, 0f7F800000;
|
|
@%p28 bra BB0_15;
|
|
bra.uni BB0_13;
|
|
|
|
BB0_15:
|
|
add.f32 %f1481, %f66, 0f40800000;
|
|
bra.uni BB0_16;
|
|
|
|
BB0_13:
|
|
setp.neu.f32 %p29, %f67, 0f7F800000;
|
|
@%p29 bra BB0_16;
|
|
|
|
selp.f32 %f1481, 0fFF800000, 0f7F800000, %p2;
|
|
|
|
BB0_16:
|
|
mul.f32 %f467, %f62, %f381;
|
|
mov.f32 %f1510, 0f3F800000;
|
|
sub.f32 %f469, %f1510, %f1481;
|
|
setp.eq.f32 %p30, %f66, 0f3F800000;
|
|
selp.f32 %f470, 0f00000000, %f469, %p30;
|
|
cvt.sat.f32.f32 %f471, %f470;
|
|
fma.rn.f32 %f472, %f467, %f467, %f382;
|
|
div.rn.f32 %f96, %f471, %f472;
|
|
mul.f32 %f473, %f1461, %f64;
|
|
fma.rn.f32 %f474, %f1460, %f63, %f473;
|
|
fma.rn.f32 %f475, %f1462, %f65, %f474;
|
|
ld.global.u32 %r63, [ignoreNormal];
|
|
setp.eq.s32 %p31, %r63, 0;
|
|
selp.f32 %f97, %f475, 0f3F800000, %p31;
|
|
mul.f32 %f476, %f97, 0f40800000;
|
|
cvt.sat.f32.f32 %f98, %f476;
|
|
setp.eq.f32 %p32, %f384, 0f3F800000;
|
|
@%p32 bra BB0_23;
|
|
bra.uni BB0_17;
|
|
|
|
BB0_23:
|
|
setp.leu.f32 %p36, %f379, 0f00000000;
|
|
@%p36 bra BB0_19;
|
|
|
|
setp.geu.f32 %p37, %f375, 0f00000000;
|
|
@%p37 bra BB0_25;
|
|
|
|
sub.f32 %f1459, %f387, %f9;
|
|
sub.f32 %f1458, %f385, %f7;
|
|
sub.f32 %f1457, %f386, %f8;
|
|
mul.f32 %f499, %f365, %f1457;
|
|
fma.rn.f32 %f500, %f364, %f1458, %f499;
|
|
fma.rn.f32 %f501, %f366, %f1459, %f500;
|
|
rcp.rn.f32 %f502, %f501;
|
|
mul.f32 %f1482, %f1458, %f502;
|
|
mul.f32 %f1483, %f1457, %f502;
|
|
mul.f32 %f1484, %f1459, %f502;
|
|
neg.f32 %f375, %f375;
|
|
bra.uni BB0_27;
|
|
|
|
BB0_17:
|
|
setp.eq.f32 %p33, %f384, 0f40000000;
|
|
@%p33 bra BB0_21;
|
|
bra.uni BB0_18;
|
|
|
|
BB0_21:
|
|
setp.leu.f32 %p35, %f379, 0f00000000;
|
|
@%p35 bra BB0_19;
|
|
|
|
mul.f32 %f493, %f373, %f64;
|
|
fma.rn.f32 %f494, %f372, %f63, %f493;
|
|
mul.f32 %f495, %f369, %f64;
|
|
fma.rn.f32 %f496, %f368, %f63, %f495;
|
|
mul.f32 %f497, %f365, %f64;
|
|
fma.rn.f32 %f498, %f364, %f63, %f497;
|
|
fma.rn.f32 %f490, %f374, %f65, %f494;
|
|
fma.rn.f32 %f491, %f370, %f65, %f496;
|
|
fma.rn.f32 %f492, %f366, %f65, %f498;
|
|
cvt.rzi.s32.f32 %r64, %f379;
|
|
mov.u32 %r65, 6;
|
|
mov.u32 %r66, 0;
|
|
// inline asm
|
|
call (%f486, %f487, %f488, %f489), _rt_texture_get_base_id, (%r64, %r65, %f490, %f491, %f492, %r66);
|
|
// inline asm
|
|
mul.f32 %f1486, %f376, %f486;
|
|
mul.f32 %f1487, %f1487, %f487;
|
|
mul.f32 %f1488, %f1488, %f488;
|
|
bra.uni BB0_28;
|
|
|
|
BB0_18:
|
|
setp.neu.f32 %p34, %f384, 0f40800000;
|
|
@%p34 bra BB0_19;
|
|
|
|
mul.f32 %f477, %f364, %f63;
|
|
mul.f32 %f478, %f365, %f64;
|
|
neg.f32 %f479, %f478;
|
|
sub.f32 %f480, %f479, %f477;
|
|
mul.f32 %f481, %f366, %f65;
|
|
sub.f32 %f482, %f480, %f481;
|
|
fma.rn.f32 %f483, %f379, %f482, %f375;
|
|
cvt.sat.f32.f32 %f484, %f483;
|
|
mul.f32 %f485, %f484, %f484;
|
|
mul.f32 %f1489, %f96, %f485;
|
|
mov.f32 %f1486, %f376;
|
|
bra.uni BB0_29;
|
|
|
|
BB0_19:
|
|
mov.f32 %f1486, %f376;
|
|
bra.uni BB0_28;
|
|
|
|
BB0_25:
|
|
mov.f32 %f1482, %f63;
|
|
mov.f32 %f1483, %f64;
|
|
mov.f32 %f1484, %f65;
|
|
|
|
BB0_27:
|
|
mul.f32 %f511, %f366, %f65;
|
|
mul.f32 %f512, %f365, %f64;
|
|
neg.f32 %f513, %f512;
|
|
mul.f32 %f514, %f364, %f63;
|
|
sub.f32 %f515, %f513, %f514;
|
|
sub.f32 %f516, %f515, %f511;
|
|
setp.gt.f32 %p38, %f516, 0f00000000;
|
|
selp.f32 %f517, 0f3F800000, 0f00000000, %p38;
|
|
mul.f32 %f518, %f373, %f1483;
|
|
fma.rn.f32 %f519, %f372, %f1482, %f518;
|
|
mul.f32 %f520, %f369, %f1483;
|
|
fma.rn.f32 %f521, %f368, %f1482, %f520;
|
|
fma.rn.f32 %f522, %f374, %f1484, %f519;
|
|
fma.rn.f32 %f523, %f370, %f1484, %f521;
|
|
fma.rn.f32 %f524, %f375, %f522, 0f3F000000;
|
|
mov.f32 %f525, 0f3F800000;
|
|
sub.f32 %f507, %f525, %f524;
|
|
fma.rn.f32 %f508, %f375, %f523, 0f3F000000;
|
|
cvt.rzi.s32.f32 %r67, %f379;
|
|
mov.f32 %f510, 0f00000000;
|
|
// inline asm
|
|
call (%f503, %f504, %f505, %f506), _rt_texture_get_f_id, (%r67, %r28, %f507, %f508, %f510, %f510);
|
|
// inline asm
|
|
mul.f32 %f526, %f517, %f503;
|
|
mul.f32 %f527, %f517, %f504;
|
|
mul.f32 %f528, %f517, %f505;
|
|
mul.f32 %f1486, %f376, %f526;
|
|
mul.f32 %f1487, %f1487, %f527;
|
|
mul.f32 %f1488, %f1488, %f528;
|
|
|
|
BB0_28:
|
|
mov.f32 %f1489, %f96;
|
|
|
|
BB0_29:
|
|
max.f32 %f544, %f1486, %f1487;
|
|
max.f32 %f545, %f544, %f1488;
|
|
mul.f32 %f546, %f98, %f1489;
|
|
mul.f32 %f547, %f546, %f545;
|
|
setp.lt.f32 %p40, %f547, 0f3727C5AC;
|
|
mov.pred %p165, -1;
|
|
mov.f32 %f1490, 0f00000000;
|
|
mov.f32 %f1491, %f1490;
|
|
mov.f32 %f1492, %f1490;
|
|
mov.f32 %f136, %f1490;
|
|
mov.f32 %f1494, %f1490;
|
|
mov.f32 %f1495, %f1490;
|
|
mov.f32 %f139, %f1490;
|
|
mov.f32 %f1497, %f1490;
|
|
mov.f32 %f1498, %f1490;
|
|
mov.f32 %f142, %f1490;
|
|
mov.f32 %f1500, %f1490;
|
|
mov.f32 %f1501, %f1490;
|
|
mov.f32 %f145, %f1490;
|
|
mov.f32 %f1503, %f1490;
|
|
mov.f32 %f1504, %f1490;
|
|
@%p40 bra BB0_31;
|
|
|
|
cvt.sat.f32.f32 %f548, %f97;
|
|
mul.f32 %f549, %f1489, %f548;
|
|
mul.f32 %f1490, %f1486, %f549;
|
|
mul.f32 %f1491, %f1487, %f549;
|
|
mul.f32 %f1492, %f1488, %f549;
|
|
mul.f32 %f550, %f1489, 0f3E800000;
|
|
mul.f32 %f551, %f98, %f550;
|
|
mul.f32 %f136, %f1486, %f551;
|
|
mul.f32 %f1494, %f1487, %f551;
|
|
mul.f32 %f1495, %f1488, %f551;
|
|
mul.f32 %f139, %f63, %f136;
|
|
mul.f32 %f1497, %f63, %f1494;
|
|
mul.f32 %f1498, %f63, %f1495;
|
|
mul.f32 %f142, %f64, %f136;
|
|
mul.f32 %f1500, %f64, %f1494;
|
|
mul.f32 %f1501, %f64, %f1495;
|
|
mul.f32 %f145, %f65, %f136;
|
|
mul.f32 %f1503, %f65, %f1494;
|
|
mul.f32 %f1504, %f65, %f1495;
|
|
mov.pred %p165, 0;
|
|
|
|
BB0_31:
|
|
@%p165 bra BB0_59;
|
|
|
|
setp.eq.s32 %p42, %r6, 0;
|
|
mov.u16 %rs161, 0;
|
|
@%p42 bra BB0_43;
|
|
|
|
abs.s32 %r8, %r6;
|
|
mov.f32 %f1509, 0f00000000;
|
|
setp.lt.s32 %p43, %r8, 1;
|
|
@%p43 bra BB0_42;
|
|
|
|
max.f32 %f149, %f559, %f337;
|
|
and.b32 %r9, %r8, 3;
|
|
setp.eq.s32 %p44, %r9, 0;
|
|
add.u64 %rd35, %SP, 0;
|
|
cvta.to.local.u64 %rd2, %rd35;
|
|
mov.f32 %f1509, 0f00000000;
|
|
mov.u32 %r264, 0;
|
|
@%p44 bra BB0_40;
|
|
|
|
setp.eq.s32 %p45, %r9, 1;
|
|
mov.f32 %f1506, 0f00000000;
|
|
mov.u32 %r262, 0;
|
|
@%p45 bra BB0_39;
|
|
|
|
setp.eq.s32 %p46, %r9, 2;
|
|
mov.f32 %f1505, 0f00000000;
|
|
mov.u32 %r261, 0;
|
|
@%p46 bra BB0_38;
|
|
|
|
sub.f32 %f571, %f385, %f380;
|
|
sub.f32 %f572, %f386, %f380;
|
|
sub.f32 %f573, %f387, %f380;
|
|
sub.f32 %f574, %f571, %f7;
|
|
sub.f32 %f575, %f572, %f8;
|
|
sub.f32 %f576, %f573, %f9;
|
|
mul.f32 %f577, %f575, %f575;
|
|
fma.rn.f32 %f578, %f574, %f574, %f577;
|
|
fma.rn.f32 %f579, %f576, %f576, %f578;
|
|
sqrt.rn.f32 %f570, %f579;
|
|
rcp.rn.f32 %f580, %f570;
|
|
mul.f32 %f566, %f580, %f574;
|
|
mul.f32 %f567, %f580, %f575;
|
|
mul.f32 %f568, %f580, %f576;
|
|
ld.global.u32 %r76, [imageEnabled];
|
|
and.b32 %r77, %r76, 32;
|
|
setp.eq.s32 %p47, %r77, 0;
|
|
selp.f32 %f581, 0f3F800000, 0f41200000, %p47;
|
|
mul.f32 %f569, %f581, %f149;
|
|
mov.u32 %r78, 1065353216;
|
|
st.local.u32 [%rd2], %r78;
|
|
ld.global.u32 %r72, [root];
|
|
// inline asm
|
|
call _rt_trace_64, (%r72, %f10, %f11, %f12, %f566, %f567, %f568, %r45, %f569, %f570, %rd35, %r29);
|
|
// inline asm
|
|
ld.local.f32 %f582, [%rd2];
|
|
add.f32 %f1505, %f582, 0f00000000;
|
|
mov.u32 %r261, %r45;
|
|
|
|
BB0_38:
|
|
cvt.rn.f32.s32 %f591, %r261;
|
|
mul.f32 %f592, %f591, 0f3DD32618;
|
|
cvt.rmi.f32.f32 %f593, %f592;
|
|
sub.f32 %f594, %f592, %f593;
|
|
mul.f32 %f595, %f591, 0f3DD2F1AA;
|
|
cvt.rmi.f32.f32 %f596, %f595;
|
|
sub.f32 %f597, %f595, %f596;
|
|
mul.f32 %f598, %f591, 0f3DC74539;
|
|
cvt.rmi.f32.f32 %f599, %f598;
|
|
sub.f32 %f600, %f598, %f599;
|
|
add.f32 %f601, %f597, 0f4199851F;
|
|
add.f32 %f602, %f600, 0f4199851F;
|
|
add.f32 %f603, %f594, 0f4199851F;
|
|
mul.f32 %f604, %f597, %f602;
|
|
fma.rn.f32 %f605, %f594, %f601, %f604;
|
|
fma.rn.f32 %f606, %f603, %f600, %f605;
|
|
add.f32 %f607, %f594, %f606;
|
|
add.f32 %f608, %f597, %f606;
|
|
add.f32 %f609, %f600, %f606;
|
|
add.f32 %f610, %f607, %f608;
|
|
mul.f32 %f611, %f609, %f610;
|
|
cvt.rmi.f32.f32 %f612, %f611;
|
|
sub.f32 %f613, %f611, %f612;
|
|
add.f32 %f614, %f607, %f609;
|
|
mul.f32 %f615, %f608, %f614;
|
|
cvt.rmi.f32.f32 %f616, %f615;
|
|
sub.f32 %f617, %f615, %f616;
|
|
add.f32 %f618, %f608, %f609;
|
|
mul.f32 %f619, %f607, %f618;
|
|
cvt.rmi.f32.f32 %f620, %f619;
|
|
sub.f32 %f621, %f619, %f620;
|
|
fma.rn.f32 %f622, %f613, 0f40000000, 0fBF800000;
|
|
fma.rn.f32 %f623, %f617, 0f40000000, 0fBF800000;
|
|
fma.rn.f32 %f624, %f621, 0f40000000, 0fBF800000;
|
|
fma.rn.f32 %f625, %f380, %f622, %f385;
|
|
fma.rn.f32 %f626, %f380, %f623, %f386;
|
|
fma.rn.f32 %f627, %f380, %f624, %f387;
|
|
sub.f32 %f628, %f625, %f7;
|
|
sub.f32 %f629, %f626, %f8;
|
|
sub.f32 %f630, %f627, %f9;
|
|
mul.f32 %f631, %f629, %f629;
|
|
fma.rn.f32 %f632, %f628, %f628, %f631;
|
|
fma.rn.f32 %f633, %f630, %f630, %f632;
|
|
sqrt.rn.f32 %f590, %f633;
|
|
rcp.rn.f32 %f634, %f590;
|
|
mul.f32 %f586, %f634, %f628;
|
|
mul.f32 %f587, %f634, %f629;
|
|
mul.f32 %f588, %f634, %f630;
|
|
ld.global.u32 %r82, [imageEnabled];
|
|
and.b32 %r83, %r82, 32;
|
|
setp.eq.s32 %p48, %r83, 0;
|
|
selp.f32 %f635, 0f3F800000, 0f41200000, %p48;
|
|
mul.f32 %f589, %f635, %f149;
|
|
mov.u32 %r84, 1065353216;
|
|
st.local.u32 [%rd2], %r84;
|
|
ld.global.u32 %r79, [root];
|
|
// inline asm
|
|
call _rt_trace_64, (%r79, %f10, %f11, %f12, %f586, %f587, %f588, %r45, %f589, %f590, %rd35, %r29);
|
|
// inline asm
|
|
ld.local.f32 %f636, [%rd2];
|
|
add.f32 %f1506, %f1505, %f636;
|
|
add.s32 %r262, %r261, 1;
|
|
|
|
BB0_39:
|
|
cvt.rn.f32.s32 %f645, %r262;
|
|
mul.f32 %f646, %f645, 0f3DD32618;
|
|
cvt.rmi.f32.f32 %f647, %f646;
|
|
sub.f32 %f648, %f646, %f647;
|
|
mul.f32 %f649, %f645, 0f3DD2F1AA;
|
|
cvt.rmi.f32.f32 %f650, %f649;
|
|
sub.f32 %f651, %f649, %f650;
|
|
mul.f32 %f652, %f645, 0f3DC74539;
|
|
cvt.rmi.f32.f32 %f653, %f652;
|
|
sub.f32 %f654, %f652, %f653;
|
|
add.f32 %f655, %f651, 0f4199851F;
|
|
add.f32 %f656, %f654, 0f4199851F;
|
|
add.f32 %f657, %f648, 0f4199851F;
|
|
mul.f32 %f658, %f651, %f656;
|
|
fma.rn.f32 %f659, %f648, %f655, %f658;
|
|
fma.rn.f32 %f660, %f657, %f654, %f659;
|
|
add.f32 %f661, %f648, %f660;
|
|
add.f32 %f662, %f651, %f660;
|
|
add.f32 %f663, %f654, %f660;
|
|
add.f32 %f664, %f661, %f662;
|
|
mul.f32 %f665, %f663, %f664;
|
|
cvt.rmi.f32.f32 %f666, %f665;
|
|
sub.f32 %f667, %f665, %f666;
|
|
add.f32 %f668, %f661, %f663;
|
|
mul.f32 %f669, %f662, %f668;
|
|
cvt.rmi.f32.f32 %f670, %f669;
|
|
sub.f32 %f671, %f669, %f670;
|
|
add.f32 %f672, %f662, %f663;
|
|
mul.f32 %f673, %f661, %f672;
|
|
cvt.rmi.f32.f32 %f674, %f673;
|
|
sub.f32 %f675, %f673, %f674;
|
|
fma.rn.f32 %f676, %f667, 0f40000000, 0fBF800000;
|
|
fma.rn.f32 %f677, %f671, 0f40000000, 0fBF800000;
|
|
fma.rn.f32 %f678, %f675, 0f40000000, 0fBF800000;
|
|
fma.rn.f32 %f679, %f380, %f676, %f385;
|
|
fma.rn.f32 %f680, %f380, %f677, %f386;
|
|
fma.rn.f32 %f681, %f380, %f678, %f387;
|
|
sub.f32 %f682, %f679, %f7;
|
|
sub.f32 %f683, %f680, %f8;
|
|
sub.f32 %f684, %f681, %f9;
|
|
mul.f32 %f685, %f683, %f683;
|
|
fma.rn.f32 %f686, %f682, %f682, %f685;
|
|
fma.rn.f32 %f687, %f684, %f684, %f686;
|
|
sqrt.rn.f32 %f644, %f687;
|
|
rcp.rn.f32 %f688, %f644;
|
|
mul.f32 %f640, %f688, %f682;
|
|
mul.f32 %f641, %f688, %f683;
|
|
mul.f32 %f642, %f688, %f684;
|
|
ld.global.u32 %r88, [imageEnabled];
|
|
and.b32 %r89, %r88, 32;
|
|
setp.eq.s32 %p49, %r89, 0;
|
|
selp.f32 %f689, 0f3F800000, 0f41200000, %p49;
|
|
mul.f32 %f643, %f689, %f149;
|
|
mov.u32 %r90, 1065353216;
|
|
st.local.u32 [%rd2], %r90;
|
|
ld.global.u32 %r85, [root];
|
|
mov.u32 %r86, 1;
|
|
// inline asm
|
|
call _rt_trace_64, (%r85, %f10, %f11, %f12, %f640, %f641, %f642, %r86, %f643, %f644, %rd35, %r29);
|
|
// inline asm
|
|
ld.local.f32 %f690, [%rd2];
|
|
add.f32 %f1509, %f1506, %f690;
|
|
add.s32 %r264, %r262, 1;
|
|
|
|
BB0_40:
|
|
setp.lt.u32 %p50, %r8, 4;
|
|
@%p50 bra BB0_42;
|
|
|
|
BB0_41:
|
|
cvt.rn.f32.s32 %f723, %r264;
|
|
mul.f32 %f724, %f723, 0f3DD32618;
|
|
cvt.rmi.f32.f32 %f725, %f724;
|
|
sub.f32 %f726, %f724, %f725;
|
|
mul.f32 %f727, %f723, 0f3DD2F1AA;
|
|
cvt.rmi.f32.f32 %f728, %f727;
|
|
sub.f32 %f729, %f727, %f728;
|
|
mul.f32 %f730, %f723, 0f3DC74539;
|
|
cvt.rmi.f32.f32 %f731, %f730;
|
|
sub.f32 %f732, %f730, %f731;
|
|
add.f32 %f733, %f729, 0f4199851F;
|
|
add.f32 %f734, %f732, 0f4199851F;
|
|
add.f32 %f735, %f726, 0f4199851F;
|
|
mul.f32 %f736, %f729, %f734;
|
|
fma.rn.f32 %f737, %f726, %f733, %f736;
|
|
fma.rn.f32 %f738, %f735, %f732, %f737;
|
|
add.f32 %f739, %f726, %f738;
|
|
add.f32 %f740, %f729, %f738;
|
|
add.f32 %f741, %f732, %f738;
|
|
add.f32 %f742, %f739, %f740;
|
|
mul.f32 %f743, %f741, %f742;
|
|
cvt.rmi.f32.f32 %f744, %f743;
|
|
sub.f32 %f745, %f743, %f744;
|
|
add.f32 %f746, %f739, %f741;
|
|
mul.f32 %f747, %f740, %f746;
|
|
cvt.rmi.f32.f32 %f748, %f747;
|
|
sub.f32 %f749, %f747, %f748;
|
|
add.f32 %f750, %f740, %f741;
|
|
mul.f32 %f751, %f739, %f750;
|
|
cvt.rmi.f32.f32 %f752, %f751;
|
|
sub.f32 %f753, %f751, %f752;
|
|
fma.rn.f32 %f754, %f745, 0f40000000, 0fBF800000;
|
|
fma.rn.f32 %f755, %f749, 0f40000000, 0fBF800000;
|
|
fma.rn.f32 %f756, %f753, 0f40000000, 0fBF800000;
|
|
fma.rn.f32 %f757, %f380, %f754, %f385;
|
|
fma.rn.f32 %f758, %f380, %f755, %f386;
|
|
fma.rn.f32 %f759, %f380, %f756, %f387;
|
|
sub.f32 %f760, %f757, %f7;
|
|
sub.f32 %f761, %f758, %f8;
|
|
sub.f32 %f762, %f759, %f9;
|
|
mul.f32 %f763, %f761, %f761;
|
|
fma.rn.f32 %f764, %f760, %f760, %f763;
|
|
fma.rn.f32 %f765, %f762, %f762, %f764;
|
|
sqrt.rn.f32 %f698, %f765;
|
|
rcp.rn.f32 %f766, %f698;
|
|
mul.f32 %f694, %f766, %f760;
|
|
mul.f32 %f695, %f766, %f761;
|
|
mul.f32 %f696, %f766, %f762;
|
|
ld.global.u32 %r103, [imageEnabled];
|
|
and.b32 %r104, %r103, 32;
|
|
setp.eq.s32 %p51, %r104, 0;
|
|
selp.f32 %f767, 0f3F800000, 0f41200000, %p51;
|
|
mul.f32 %f697, %f767, %f149;
|
|
mov.u32 %r105, 1065353216;
|
|
st.local.u32 [%rd2], %r105;
|
|
ld.global.u32 %r91, [root];
|
|
mov.u32 %r101, 1;
|
|
// inline asm
|
|
call _rt_trace_64, (%r91, %f10, %f11, %f12, %f694, %f695, %f696, %r101, %f697, %f698, %rd35, %r29);
|
|
// inline asm
|
|
ld.local.f32 %f768, [%rd2];
|
|
add.f32 %f769, %f1509, %f768;
|
|
add.s32 %r106, %r264, 1;
|
|
cvt.rn.f32.s32 %f770, %r106;
|
|
mul.f32 %f771, %f770, 0f3DD32618;
|
|
cvt.rmi.f32.f32 %f772, %f771;
|
|
sub.f32 %f773, %f771, %f772;
|
|
mul.f32 %f774, %f770, 0f3DD2F1AA;
|
|
cvt.rmi.f32.f32 %f775, %f774;
|
|
sub.f32 %f776, %f774, %f775;
|
|
mul.f32 %f777, %f770, 0f3DC74539;
|
|
cvt.rmi.f32.f32 %f778, %f777;
|
|
sub.f32 %f779, %f777, %f778;
|
|
add.f32 %f780, %f776, 0f4199851F;
|
|
add.f32 %f781, %f779, 0f4199851F;
|
|
add.f32 %f782, %f773, 0f4199851F;
|
|
mul.f32 %f783, %f776, %f781;
|
|
fma.rn.f32 %f784, %f773, %f780, %f783;
|
|
fma.rn.f32 %f785, %f782, %f779, %f784;
|
|
add.f32 %f786, %f773, %f785;
|
|
add.f32 %f787, %f776, %f785;
|
|
add.f32 %f788, %f779, %f785;
|
|
add.f32 %f789, %f786, %f787;
|
|
mul.f32 %f790, %f788, %f789;
|
|
cvt.rmi.f32.f32 %f791, %f790;
|
|
sub.f32 %f792, %f790, %f791;
|
|
add.f32 %f793, %f786, %f788;
|
|
mul.f32 %f794, %f787, %f793;
|
|
cvt.rmi.f32.f32 %f795, %f794;
|
|
sub.f32 %f796, %f794, %f795;
|
|
add.f32 %f797, %f787, %f788;
|
|
mul.f32 %f798, %f786, %f797;
|
|
cvt.rmi.f32.f32 %f799, %f798;
|
|
sub.f32 %f800, %f798, %f799;
|
|
fma.rn.f32 %f801, %f792, 0f40000000, 0fBF800000;
|
|
fma.rn.f32 %f802, %f796, 0f40000000, 0fBF800000;
|
|
fma.rn.f32 %f803, %f800, 0f40000000, 0fBF800000;
|
|
fma.rn.f32 %f804, %f380, %f801, %f385;
|
|
fma.rn.f32 %f805, %f380, %f802, %f386;
|
|
fma.rn.f32 %f806, %f380, %f803, %f387;
|
|
sub.f32 %f807, %f804, %f7;
|
|
sub.f32 %f808, %f805, %f8;
|
|
sub.f32 %f809, %f806, %f9;
|
|
mul.f32 %f810, %f808, %f808;
|
|
fma.rn.f32 %f811, %f807, %f807, %f810;
|
|
fma.rn.f32 %f812, %f809, %f809, %f811;
|
|
sqrt.rn.f32 %f706, %f812;
|
|
rcp.rn.f32 %f813, %f706;
|
|
mul.f32 %f702, %f813, %f807;
|
|
mul.f32 %f703, %f813, %f808;
|
|
mul.f32 %f704, %f813, %f809;
|
|
ld.global.u32 %r107, [imageEnabled];
|
|
and.b32 %r108, %r107, 32;
|
|
setp.eq.s32 %p52, %r108, 0;
|
|
selp.f32 %f814, 0f3F800000, 0f41200000, %p52;
|
|
mul.f32 %f705, %f814, %f149;
|
|
st.local.u32 [%rd2], %r105;
|
|
ld.global.u32 %r94, [root];
|
|
// inline asm
|
|
call _rt_trace_64, (%r94, %f10, %f11, %f12, %f702, %f703, %f704, %r101, %f705, %f706, %rd35, %r29);
|
|
// inline asm
|
|
ld.local.f32 %f815, [%rd2];
|
|
add.f32 %f816, %f769, %f815;
|
|
add.s32 %r109, %r264, 2;
|
|
cvt.rn.f32.s32 %f817, %r109;
|
|
mul.f32 %f818, %f817, 0f3DD32618;
|
|
cvt.rmi.f32.f32 %f819, %f818;
|
|
sub.f32 %f820, %f818, %f819;
|
|
mul.f32 %f821, %f817, 0f3DD2F1AA;
|
|
cvt.rmi.f32.f32 %f822, %f821;
|
|
sub.f32 %f823, %f821, %f822;
|
|
mul.f32 %f824, %f817, 0f3DC74539;
|
|
cvt.rmi.f32.f32 %f825, %f824;
|
|
sub.f32 %f826, %f824, %f825;
|
|
add.f32 %f827, %f823, 0f4199851F;
|
|
add.f32 %f828, %f826, 0f4199851F;
|
|
add.f32 %f829, %f820, 0f4199851F;
|
|
mul.f32 %f830, %f823, %f828;
|
|
fma.rn.f32 %f831, %f820, %f827, %f830;
|
|
fma.rn.f32 %f832, %f829, %f826, %f831;
|
|
add.f32 %f833, %f820, %f832;
|
|
add.f32 %f834, %f823, %f832;
|
|
add.f32 %f835, %f826, %f832;
|
|
add.f32 %f836, %f833, %f834;
|
|
mul.f32 %f837, %f835, %f836;
|
|
cvt.rmi.f32.f32 %f838, %f837;
|
|
sub.f32 %f839, %f837, %f838;
|
|
add.f32 %f840, %f833, %f835;
|
|
mul.f32 %f841, %f834, %f840;
|
|
cvt.rmi.f32.f32 %f842, %f841;
|
|
sub.f32 %f843, %f841, %f842;
|
|
add.f32 %f844, %f834, %f835;
|
|
mul.f32 %f845, %f833, %f844;
|
|
cvt.rmi.f32.f32 %f846, %f845;
|
|
sub.f32 %f847, %f845, %f846;
|
|
fma.rn.f32 %f848, %f839, 0f40000000, 0fBF800000;
|
|
fma.rn.f32 %f849, %f843, 0f40000000, 0fBF800000;
|
|
fma.rn.f32 %f850, %f847, 0f40000000, 0fBF800000;
|
|
fma.rn.f32 %f851, %f380, %f848, %f385;
|
|
fma.rn.f32 %f852, %f380, %f849, %f386;
|
|
fma.rn.f32 %f853, %f380, %f850, %f387;
|
|
sub.f32 %f854, %f851, %f7;
|
|
sub.f32 %f855, %f852, %f8;
|
|
sub.f32 %f856, %f853, %f9;
|
|
mul.f32 %f857, %f855, %f855;
|
|
fma.rn.f32 %f858, %f854, %f854, %f857;
|
|
fma.rn.f32 %f859, %f856, %f856, %f858;
|
|
sqrt.rn.f32 %f714, %f859;
|
|
rcp.rn.f32 %f860, %f714;
|
|
mul.f32 %f710, %f860, %f854;
|
|
mul.f32 %f711, %f860, %f855;
|
|
mul.f32 %f712, %f860, %f856;
|
|
ld.global.u32 %r110, [imageEnabled];
|
|
and.b32 %r111, %r110, 32;
|
|
setp.eq.s32 %p53, %r111, 0;
|
|
selp.f32 %f861, 0f3F800000, 0f41200000, %p53;
|
|
mul.f32 %f713, %f861, %f149;
|
|
st.local.u32 [%rd2], %r105;
|
|
ld.global.u32 %r97, [root];
|
|
// inline asm
|
|
call _rt_trace_64, (%r97, %f10, %f11, %f12, %f710, %f711, %f712, %r101, %f713, %f714, %rd35, %r29);
|
|
// inline asm
|
|
ld.local.f32 %f862, [%rd2];
|
|
add.f32 %f863, %f816, %f862;
|
|
add.s32 %r112, %r264, 3;
|
|
cvt.rn.f32.s32 %f864, %r112;
|
|
mul.f32 %f865, %f864, 0f3DD32618;
|
|
cvt.rmi.f32.f32 %f866, %f865;
|
|
sub.f32 %f867, %f865, %f866;
|
|
mul.f32 %f868, %f864, 0f3DD2F1AA;
|
|
cvt.rmi.f32.f32 %f869, %f868;
|
|
sub.f32 %f870, %f868, %f869;
|
|
mul.f32 %f871, %f864, 0f3DC74539;
|
|
cvt.rmi.f32.f32 %f872, %f871;
|
|
sub.f32 %f873, %f871, %f872;
|
|
add.f32 %f874, %f870, 0f4199851F;
|
|
add.f32 %f875, %f873, 0f4199851F;
|
|
add.f32 %f876, %f867, 0f4199851F;
|
|
mul.f32 %f877, %f870, %f875;
|
|
fma.rn.f32 %f878, %f867, %f874, %f877;
|
|
fma.rn.f32 %f879, %f876, %f873, %f878;
|
|
add.f32 %f880, %f867, %f879;
|
|
add.f32 %f881, %f870, %f879;
|
|
add.f32 %f882, %f873, %f879;
|
|
add.f32 %f883, %f880, %f881;
|
|
mul.f32 %f884, %f882, %f883;
|
|
cvt.rmi.f32.f32 %f885, %f884;
|
|
sub.f32 %f886, %f884, %f885;
|
|
add.f32 %f887, %f880, %f882;
|
|
mul.f32 %f888, %f881, %f887;
|
|
cvt.rmi.f32.f32 %f889, %f888;
|
|
sub.f32 %f890, %f888, %f889;
|
|
add.f32 %f891, %f881, %f882;
|
|
mul.f32 %f892, %f880, %f891;
|
|
cvt.rmi.f32.f32 %f893, %f892;
|
|
sub.f32 %f894, %f892, %f893;
|
|
fma.rn.f32 %f895, %f886, 0f40000000, 0fBF800000;
|
|
fma.rn.f32 %f896, %f890, 0f40000000, 0fBF800000;
|
|
fma.rn.f32 %f897, %f894, 0f40000000, 0fBF800000;
|
|
fma.rn.f32 %f898, %f380, %f895, %f385;
|
|
fma.rn.f32 %f899, %f380, %f896, %f386;
|
|
fma.rn.f32 %f900, %f380, %f897, %f387;
|
|
sub.f32 %f901, %f898, %f7;
|
|
sub.f32 %f902, %f899, %f8;
|
|
sub.f32 %f903, %f900, %f9;
|
|
mul.f32 %f904, %f902, %f902;
|
|
fma.rn.f32 %f905, %f901, %f901, %f904;
|
|
fma.rn.f32 %f906, %f903, %f903, %f905;
|
|
sqrt.rn.f32 %f722, %f906;
|
|
rcp.rn.f32 %f907, %f722;
|
|
mul.f32 %f718, %f907, %f901;
|
|
mul.f32 %f719, %f907, %f902;
|
|
mul.f32 %f720, %f907, %f903;
|
|
ld.global.u32 %r113, [imageEnabled];
|
|
and.b32 %r114, %r113, 32;
|
|
setp.eq.s32 %p54, %r114, 0;
|
|
selp.f32 %f908, 0f3F800000, 0f41200000, %p54;
|
|
mul.f32 %f721, %f908, %f149;
|
|
st.local.u32 [%rd2], %r105;
|
|
ld.global.u32 %r100, [root];
|
|
// inline asm
|
|
call _rt_trace_64, (%r100, %f10, %f11, %f12, %f718, %f719, %f720, %r101, %f721, %f722, %rd35, %r29);
|
|
// inline asm
|
|
ld.local.f32 %f909, [%rd2];
|
|
add.f32 %f1509, %f863, %f909;
|
|
add.s32 %r264, %r264, 4;
|
|
setp.lt.s32 %p55, %r264, %r8;
|
|
@%p55 bra BB0_41;
|
|
|
|
BB0_42:
|
|
cvt.rn.f32.s32 %f910, %r8;
|
|
div.rn.f32 %f1510, %f1509, %f910;
|
|
shr.u32 %r115, %r6, 31;
|
|
cvt.u16.u32 %rs161, %r115;
|
|
|
|
BB0_43:
|
|
fma.rn.f32 %f1478, %f1490, %f1510, %f1478;
|
|
fma.rn.f32 %f1477, %f1491, %f1510, %f1477;
|
|
fma.rn.f32 %f1476, %f1492, %f1510, %f1476;
|
|
fma.rn.f32 %f1475, %f136, %f1510, %f1475;
|
|
fma.rn.f32 %f1474, %f1494, %f1510, %f1474;
|
|
fma.rn.f32 %f1473, %f1495, %f1510, %f1473;
|
|
fma.rn.f32 %f1472, %f139, %f1510, %f1472;
|
|
fma.rn.f32 %f1471, %f1497, %f1510, %f1471;
|
|
fma.rn.f32 %f1470, %f1498, %f1510, %f1470;
|
|
fma.rn.f32 %f1469, %f142, %f1510, %f1469;
|
|
fma.rn.f32 %f1468, %f1500, %f1510, %f1468;
|
|
fma.rn.f32 %f1467, %f1501, %f1510, %f1467;
|
|
fma.rn.f32 %f1466, %f145, %f1510, %f1466;
|
|
fma.rn.f32 %f1465, %f1503, %f1510, %f1465;
|
|
fma.rn.f32 %f1464, %f1504, %f1510, %f1464;
|
|
setp.eq.s16 %p56, %rs161, 0;
|
|
@%p56 bra BB0_58;
|
|
|
|
mov.f32 %f1456, 0fB5BFBE8E;
|
|
mov.f32 %f1455, 0fBF317200;
|
|
mov.f32 %f1454, 0f35BFBE8E;
|
|
mov.f32 %f1453, 0f3F317200;
|
|
mov.f32 %f1452, 0f3DAAAABD;
|
|
mov.f32 %f1451, 0f3C4CAF63;
|
|
mov.f32 %f1450, 0f3B18F0FE;
|
|
mul.f32 %f913, %f139, 0f3F000000;
|
|
mul.f32 %f914, %f142, 0f3F000000;
|
|
mul.f32 %f915, %f914, %f914;
|
|
fma.rn.f32 %f916, %f913, %f913, %f915;
|
|
mul.f32 %f917, %f145, 0f3F000000;
|
|
fma.rn.f32 %f918, %f917, %f917, %f916;
|
|
sqrt.rn.f32 %f919, %f918;
|
|
rcp.rn.f32 %f920, %f919;
|
|
mul.f32 %f921, %f913, %f920;
|
|
mul.f32 %f922, %f914, %f920;
|
|
mul.f32 %f923, %f917, %f920;
|
|
mul.f32 %f924, %f1461, %f922;
|
|
fma.rn.f32 %f925, %f1460, %f921, %f924;
|
|
fma.rn.f32 %f926, %f1462, %f923, %f925;
|
|
fma.rn.f32 %f176, %f926, 0f3F000000, 0f3F000000;
|
|
add.f32 %f927, %f919, %f919;
|
|
div.rn.f32 %f928, %f927, %f136;
|
|
add.f32 %f177, %f928, 0f3F800000;
|
|
div.rn.f32 %f929, %f919, %f136;
|
|
mov.f32 %f930, 0f3F800000;
|
|
sub.f32 %f931, %f930, %f929;
|
|
add.f32 %f932, %f929, 0f3F800000;
|
|
div.rn.f32 %f178, %f931, %f932;
|
|
sub.f32 %f933, %f930, %f178;
|
|
add.f32 %f934, %f177, 0f3F800000;
|
|
mul.f32 %f179, %f934, %f933;
|
|
mul.f32 %f935, %f177, 0f3F000000;
|
|
cvt.rzi.f32.f32 %f936, %f935;
|
|
add.f32 %f937, %f936, %f936;
|
|
sub.f32 %f938, %f177, %f937;
|
|
abs.f32 %f180, %f938;
|
|
abs.f32 %f181, %f176;
|
|
setp.lt.f32 %p57, %f181, 0f00800000;
|
|
mul.f32 %f939, %f181, 0f4B800000;
|
|
selp.f32 %f940, 0fC3170000, 0fC2FE0000, %p57;
|
|
selp.f32 %f941, %f939, %f181, %p57;
|
|
mov.b32 %r116, %f941;
|
|
and.b32 %r117, %r116, 8388607;
|
|
or.b32 %r118, %r117, 1065353216;
|
|
mov.b32 %f942, %r118;
|
|
shr.u32 %r119, %r116, 23;
|
|
cvt.rn.f32.u32 %f943, %r119;
|
|
add.f32 %f944, %f940, %f943;
|
|
setp.gt.f32 %p58, %f942, 0f3FB504F3;
|
|
mul.f32 %f945, %f942, 0f3F000000;
|
|
add.f32 %f946, %f944, 0f3F800000;
|
|
selp.f32 %f947, %f945, %f942, %p58;
|
|
selp.f32 %f948, %f946, %f944, %p58;
|
|
add.f32 %f949, %f947, 0fBF800000;
|
|
add.f32 %f912, %f947, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f911,%f912;
|
|
// inline asm
|
|
add.f32 %f950, %f949, %f949;
|
|
mul.f32 %f951, %f911, %f950;
|
|
mul.f32 %f952, %f951, %f951;
|
|
fma.rn.f32 %f955, %f1450, %f952, %f1451;
|
|
fma.rn.f32 %f957, %f955, %f952, %f1452;
|
|
mul.rn.f32 %f958, %f957, %f952;
|
|
mul.rn.f32 %f959, %f958, %f951;
|
|
sub.f32 %f960, %f949, %f951;
|
|
neg.f32 %f961, %f951;
|
|
add.f32 %f962, %f960, %f960;
|
|
fma.rn.f32 %f963, %f961, %f949, %f962;
|
|
mul.rn.f32 %f964, %f911, %f963;
|
|
add.f32 %f965, %f959, %f951;
|
|
sub.f32 %f966, %f951, %f965;
|
|
add.f32 %f967, %f959, %f966;
|
|
add.f32 %f968, %f964, %f967;
|
|
add.f32 %f969, %f965, %f968;
|
|
sub.f32 %f970, %f965, %f969;
|
|
add.f32 %f971, %f968, %f970;
|
|
mul.rn.f32 %f973, %f948, %f1453;
|
|
mul.rn.f32 %f975, %f948, %f1454;
|
|
add.f32 %f976, %f973, %f969;
|
|
sub.f32 %f977, %f973, %f976;
|
|
add.f32 %f978, %f969, %f977;
|
|
add.f32 %f979, %f971, %f978;
|
|
add.f32 %f980, %f975, %f979;
|
|
add.f32 %f981, %f976, %f980;
|
|
sub.f32 %f982, %f976, %f981;
|
|
add.f32 %f983, %f980, %f982;
|
|
abs.f32 %f182, %f177;
|
|
setp.gt.f32 %p59, %f182, 0f77F684DF;
|
|
mul.f32 %f984, %f177, 0f39000000;
|
|
selp.f32 %f985, %f984, %f177, %p59;
|
|
mul.rn.f32 %f986, %f985, %f981;
|
|
neg.f32 %f987, %f986;
|
|
fma.rn.f32 %f988, %f985, %f981, %f987;
|
|
fma.rn.f32 %f989, %f985, %f983, %f988;
|
|
mov.f32 %f990, 0f00000000;
|
|
fma.rn.f32 %f991, %f990, %f981, %f989;
|
|
add.rn.f32 %f992, %f986, %f991;
|
|
neg.f32 %f993, %f992;
|
|
add.rn.f32 %f994, %f986, %f993;
|
|
add.rn.f32 %f995, %f994, %f991;
|
|
mov.b32 %r120, %f992;
|
|
setp.eq.s32 %p60, %r120, 1118925336;
|
|
add.s32 %r121, %r120, -1;
|
|
mov.b32 %f996, %r121;
|
|
add.f32 %f997, %f995, 0f37000000;
|
|
selp.f32 %f998, %f996, %f992, %p60;
|
|
selp.f32 %f183, %f997, %f995, %p60;
|
|
mul.f32 %f999, %f998, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f1000, %f999;
|
|
fma.rn.f32 %f1002, %f1000, %f1455, %f998;
|
|
fma.rn.f32 %f1004, %f1000, %f1456, %f1002;
|
|
mul.f32 %f1005, %f1004, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f1006, %f1005;
|
|
add.f32 %f1007, %f1000, 0f00000000;
|
|
ex2.approx.f32 %f1008, %f1007;
|
|
mul.f32 %f1009, %f1006, %f1008;
|
|
setp.lt.f32 %p61, %f998, 0fC2D20000;
|
|
selp.f32 %f1010, 0f00000000, %f1009, %p61;
|
|
setp.gt.f32 %p62, %f998, 0f42D20000;
|
|
selp.f32 %f1511, 0f7F800000, %f1010, %p62;
|
|
setp.eq.f32 %p63, %f1511, 0f7F800000;
|
|
@%p63 bra BB0_46;
|
|
|
|
fma.rn.f32 %f1511, %f1511, %f183, %f1511;
|
|
|
|
BB0_46:
|
|
setp.lt.f32 %p64, %f176, 0f00000000;
|
|
setp.eq.f32 %p65, %f180, 0f3F800000;
|
|
and.pred %p4, %p64, %p65;
|
|
mov.b32 %r122, %f1511;
|
|
xor.b32 %r123, %r122, -2147483648;
|
|
mov.b32 %f1011, %r123;
|
|
selp.f32 %f1513, %f1011, %f1511, %p4;
|
|
setp.eq.f32 %p66, %f176, 0f00000000;
|
|
@%p66 bra BB0_49;
|
|
bra.uni BB0_47;
|
|
|
|
BB0_49:
|
|
add.f32 %f1013, %f176, %f176;
|
|
mov.b32 %r124, %f1013;
|
|
selp.b32 %r125, %r124, 0, %p65;
|
|
or.b32 %r126, %r125, 2139095040;
|
|
setp.lt.f32 %p70, %f177, 0f00000000;
|
|
selp.b32 %r127, %r126, %r125, %p70;
|
|
mov.b32 %f1513, %r127;
|
|
bra.uni BB0_50;
|
|
|
|
BB0_47:
|
|
setp.geu.f32 %p67, %f176, 0f00000000;
|
|
@%p67 bra BB0_50;
|
|
|
|
cvt.rzi.f32.f32 %f1012, %f177;
|
|
setp.neu.f32 %p68, %f1012, %f177;
|
|
selp.f32 %f1513, 0f7FFFFFFF, %f1513, %p68;
|
|
|
|
BB0_50:
|
|
add.f32 %f1014, %f181, %f182;
|
|
mov.b32 %r128, %f1014;
|
|
setp.lt.s32 %p71, %r128, 2139095040;
|
|
@%p71 bra BB0_57;
|
|
|
|
setp.gtu.f32 %p72, %f181, 0f7F800000;
|
|
setp.gtu.f32 %p73, %f182, 0f7F800000;
|
|
or.pred %p74, %p72, %p73;
|
|
@%p74 bra BB0_56;
|
|
bra.uni BB0_52;
|
|
|
|
BB0_56:
|
|
add.f32 %f1513, %f177, %f176;
|
|
bra.uni BB0_57;
|
|
|
|
BB0_52:
|
|
setp.eq.f32 %p75, %f182, 0f7F800000;
|
|
@%p75 bra BB0_55;
|
|
bra.uni BB0_53;
|
|
|
|
BB0_55:
|
|
setp.gt.f32 %p78, %f181, 0f3F800000;
|
|
selp.b32 %r132, 2139095040, 0, %p78;
|
|
xor.b32 %r133, %r132, 2139095040;
|
|
setp.lt.f32 %p79, %f177, 0f00000000;
|
|
selp.b32 %r134, %r133, %r132, %p79;
|
|
mov.b32 %f1015, %r134;
|
|
setp.eq.f32 %p80, %f176, 0fBF800000;
|
|
selp.f32 %f1513, 0f3F800000, %f1015, %p80;
|
|
bra.uni BB0_57;
|
|
|
|
BB0_53:
|
|
setp.neu.f32 %p76, %f181, 0f7F800000;
|
|
@%p76 bra BB0_57;
|
|
|
|
setp.ltu.f32 %p77, %f177, 0f00000000;
|
|
selp.b32 %r129, 0, 2139095040, %p77;
|
|
or.b32 %r130, %r129, -2147483648;
|
|
selp.b32 %r131, %r130, %r129, %p4;
|
|
mov.b32 %f1513, %r131;
|
|
|
|
BB0_57:
|
|
setp.eq.f32 %p81, %f177, 0f00000000;
|
|
setp.eq.f32 %p82, %f176, 0f3F800000;
|
|
or.pred %p83, %p82, %p81;
|
|
selp.f32 %f1016, 0f3F800000, %f1513, %p83;
|
|
fma.rn.f32 %f1017, %f179, %f1016, %f178;
|
|
mul.f32 %f1018, %f136, %f1017;
|
|
div.rn.f32 %f1019, %f1018, %f376;
|
|
div.rn.f32 %f1020, %f1019, %f96;
|
|
cvt.sat.f32.f32 %f1021, %f1020;
|
|
mul.f32 %f1510, %f1510, %f1021;
|
|
|
|
BB0_58:
|
|
add.f32 %f1463, %f1463, %f1510;
|
|
|
|
BB0_59:
|
|
add.s32 %r260, %r260, 1;
|
|
setp.lt.u32 %p84, %r260, %r4;
|
|
@%p84 bra BB0_5;
|
|
|
|
BB0_60:
|
|
ld.global.u32 %r266, [imageEnabled];
|
|
and.b32 %r135, %r266, 8;
|
|
setp.eq.s32 %p85, %r135, 0;
|
|
@%p85 bra BB0_73;
|
|
|
|
cvt.sat.f32.f32 %f230, %f1463;
|
|
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 %f1024, 0f3E68BA2E;
|
|
cvt.rzi.f32.f32 %f1025, %f1024;
|
|
fma.rn.f32 %f1026, %f1025, 0fC0000000, 0f3EE8BA2E;
|
|
abs.f32 %f231, %f1026;
|
|
abs.f32 %f232, %f230;
|
|
setp.lt.f32 %p86, %f232, 0f00800000;
|
|
mul.f32 %f1027, %f232, 0f4B800000;
|
|
selp.f32 %f1028, 0fC3170000, 0fC2FE0000, %p86;
|
|
selp.f32 %f1029, %f1027, %f232, %p86;
|
|
mov.b32 %r138, %f1029;
|
|
and.b32 %r139, %r138, 8388607;
|
|
or.b32 %r140, %r139, 1065353216;
|
|
mov.b32 %f1030, %r140;
|
|
shr.u32 %r141, %r138, 23;
|
|
cvt.rn.f32.u32 %f1031, %r141;
|
|
add.f32 %f1032, %f1028, %f1031;
|
|
setp.gt.f32 %p87, %f1030, 0f3FB504F3;
|
|
mul.f32 %f1033, %f1030, 0f3F000000;
|
|
add.f32 %f1034, %f1032, 0f3F800000;
|
|
selp.f32 %f1035, %f1033, %f1030, %p87;
|
|
selp.f32 %f1036, %f1034, %f1032, %p87;
|
|
add.f32 %f1037, %f1035, 0fBF800000;
|
|
add.f32 %f1023, %f1035, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f1022,%f1023;
|
|
// inline asm
|
|
add.f32 %f1038, %f1037, %f1037;
|
|
mul.f32 %f1039, %f1022, %f1038;
|
|
mul.f32 %f1040, %f1039, %f1039;
|
|
mov.f32 %f1041, 0f3C4CAF63;
|
|
mov.f32 %f1042, 0f3B18F0FE;
|
|
fma.rn.f32 %f1043, %f1042, %f1040, %f1041;
|
|
mov.f32 %f1044, 0f3DAAAABD;
|
|
fma.rn.f32 %f1045, %f1043, %f1040, %f1044;
|
|
mul.rn.f32 %f1046, %f1045, %f1040;
|
|
mul.rn.f32 %f1047, %f1046, %f1039;
|
|
sub.f32 %f1048, %f1037, %f1039;
|
|
neg.f32 %f1049, %f1039;
|
|
add.f32 %f1050, %f1048, %f1048;
|
|
fma.rn.f32 %f1051, %f1049, %f1037, %f1050;
|
|
mul.rn.f32 %f1052, %f1022, %f1051;
|
|
add.f32 %f1053, %f1047, %f1039;
|
|
sub.f32 %f1054, %f1039, %f1053;
|
|
add.f32 %f1055, %f1047, %f1054;
|
|
add.f32 %f1056, %f1052, %f1055;
|
|
add.f32 %f1057, %f1053, %f1056;
|
|
sub.f32 %f1058, %f1053, %f1057;
|
|
add.f32 %f1059, %f1056, %f1058;
|
|
mov.f32 %f1060, 0f3F317200;
|
|
mul.rn.f32 %f1061, %f1036, %f1060;
|
|
mov.f32 %f1062, 0f35BFBE8E;
|
|
mul.rn.f32 %f1063, %f1036, %f1062;
|
|
add.f32 %f1064, %f1061, %f1057;
|
|
sub.f32 %f1065, %f1061, %f1064;
|
|
add.f32 %f1066, %f1057, %f1065;
|
|
add.f32 %f1067, %f1059, %f1066;
|
|
add.f32 %f1068, %f1063, %f1067;
|
|
add.f32 %f1069, %f1064, %f1068;
|
|
sub.f32 %f1070, %f1064, %f1069;
|
|
add.f32 %f1071, %f1068, %f1070;
|
|
mov.f32 %f1072, 0f3EE8BA2E;
|
|
mul.rn.f32 %f1073, %f1072, %f1069;
|
|
neg.f32 %f1074, %f1073;
|
|
fma.rn.f32 %f1075, %f1072, %f1069, %f1074;
|
|
fma.rn.f32 %f1076, %f1072, %f1071, %f1075;
|
|
mov.f32 %f1077, 0f00000000;
|
|
fma.rn.f32 %f1078, %f1077, %f1069, %f1076;
|
|
add.rn.f32 %f1079, %f1073, %f1078;
|
|
neg.f32 %f1080, %f1079;
|
|
add.rn.f32 %f1081, %f1073, %f1080;
|
|
add.rn.f32 %f1082, %f1081, %f1078;
|
|
mov.b32 %r142, %f1079;
|
|
setp.eq.s32 %p88, %r142, 1118925336;
|
|
add.s32 %r143, %r142, -1;
|
|
mov.b32 %f1083, %r143;
|
|
add.f32 %f1084, %f1082, 0f37000000;
|
|
selp.f32 %f1085, %f1083, %f1079, %p88;
|
|
selp.f32 %f233, %f1084, %f1082, %p88;
|
|
mul.f32 %f1086, %f1085, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f1087, %f1086;
|
|
mov.f32 %f1088, 0fBF317200;
|
|
fma.rn.f32 %f1089, %f1087, %f1088, %f1085;
|
|
mov.f32 %f1090, 0fB5BFBE8E;
|
|
fma.rn.f32 %f1091, %f1087, %f1090, %f1089;
|
|
mul.f32 %f1092, %f1091, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f1093, %f1092;
|
|
add.f32 %f1094, %f1087, 0f00000000;
|
|
ex2.approx.f32 %f1095, %f1094;
|
|
mul.f32 %f1096, %f1093, %f1095;
|
|
setp.lt.f32 %p89, %f1085, 0fC2D20000;
|
|
selp.f32 %f1097, 0f00000000, %f1096, %p89;
|
|
setp.gt.f32 %p90, %f1085, 0f42D20000;
|
|
selp.f32 %f1547, 0f7F800000, %f1097, %p90;
|
|
setp.eq.f32 %p91, %f1547, 0f7F800000;
|
|
@%p91 bra BB0_63;
|
|
|
|
fma.rn.f32 %f1547, %f1547, %f233, %f1547;
|
|
|
|
BB0_63:
|
|
setp.lt.f32 %p92, %f230, 0f00000000;
|
|
setp.eq.f32 %p93, %f231, 0f3F800000;
|
|
and.pred %p5, %p92, %p93;
|
|
mov.b32 %r144, %f1547;
|
|
xor.b32 %r145, %r144, -2147483648;
|
|
mov.b32 %f1098, %r145;
|
|
selp.f32 %f1549, %f1098, %f1547, %p5;
|
|
setp.eq.f32 %p94, %f230, 0f00000000;
|
|
@%p94 bra BB0_66;
|
|
bra.uni BB0_64;
|
|
|
|
BB0_66:
|
|
add.f32 %f1101, %f230, %f230;
|
|
selp.f32 %f1549, %f1101, 0f00000000, %p93;
|
|
bra.uni BB0_67;
|
|
|
|
BB0_131:
|
|
mov.u64 %rd190, image_HDR;
|
|
cvta.global.u64 %rd185, %rd190;
|
|
mov.u32 %r231, 8;
|
|
// inline asm
|
|
call (%rd184), _rt_buffer_get_64, (%rd185, %r28, %r231, %rd6, %rd7, %rd13, %rd13);
|
|
// inline asm
|
|
mov.f32 %f1421, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs107, %f1421;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs108, 0;
|
|
st.v4.u16 [%rd184], {%rs107, %rs107, %rs107, %rs108};
|
|
|
|
BB0_132:
|
|
ld.global.u32 %r232, [additive];
|
|
setp.eq.s32 %p161, %r232, 0;
|
|
@%p161 bra BB0_134;
|
|
|
|
mov.u64 %rd203, image_RNM0;
|
|
cvta.global.u64 %rd192, %rd203;
|
|
mov.u32 %r236, 8;
|
|
// inline asm
|
|
call (%rd191), _rt_buffer_get_64, (%rd192, %r28, %r236, %rd6, %rd7, %rd13, %rd13);
|
|
// inline asm
|
|
ld.v4.u16 {%rs115, %rs116, %rs117, %rs118}, [%rd191];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1422, %rs115;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1423, %rs116;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1424, %rs117;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd197), _rt_buffer_get_64, (%rd192, %r28, %r236, %rd6, %rd7, %rd13, %rd13);
|
|
// inline asm
|
|
add.f32 %f1425, %f1422, 0f00000000;
|
|
add.f32 %f1426, %f1423, 0f00000000;
|
|
add.f32 %f1427, %f1424, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs114, %f1427;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs113, %f1426;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs112, %f1425;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs119, 0;
|
|
st.v4.u16 [%rd197], {%rs112, %rs113, %rs114, %rs119};
|
|
bra.uni BB0_135;
|
|
|
|
BB0_134:
|
|
mov.u64 %rd210, image_RNM0;
|
|
cvta.global.u64 %rd205, %rd210;
|
|
mov.u32 %r238, 8;
|
|
// inline asm
|
|
call (%rd204), _rt_buffer_get_64, (%rd205, %r28, %r238, %rd6, %rd7, %rd13, %rd13);
|
|
// inline asm
|
|
mov.f32 %f1428, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs120, %f1428;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs121, 0;
|
|
st.v4.u16 [%rd204], {%rs120, %rs120, %rs120, %rs121};
|
|
|
|
BB0_135:
|
|
ld.global.u32 %r239, [additive];
|
|
setp.eq.s32 %p162, %r239, 0;
|
|
@%p162 bra BB0_137;
|
|
|
|
mov.u64 %rd223, image_RNM1;
|
|
cvta.global.u64 %rd212, %rd223;
|
|
mov.u32 %r243, 8;
|
|
// inline asm
|
|
call (%rd211), _rt_buffer_get_64, (%rd212, %r28, %r243, %rd6, %rd7, %rd13, %rd13);
|
|
// inline asm
|
|
ld.v4.u16 {%rs128, %rs129, %rs130, %rs131}, [%rd211];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1429, %rs128;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1430, %rs129;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1431, %rs130;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd217), _rt_buffer_get_64, (%rd212, %r28, %r243, %rd6, %rd7, %rd13, %rd13);
|
|
// inline asm
|
|
add.f32 %f1432, %f1429, 0f00000000;
|
|
add.f32 %f1433, %f1430, 0f00000000;
|
|
add.f32 %f1434, %f1431, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs127, %f1434;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs126, %f1433;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs125, %f1432;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs132, 0;
|
|
st.v4.u16 [%rd217], {%rs125, %rs126, %rs127, %rs132};
|
|
bra.uni BB0_138;
|
|
|
|
BB0_137:
|
|
mov.u64 %rd230, image_RNM1;
|
|
cvta.global.u64 %rd225, %rd230;
|
|
mov.u32 %r245, 8;
|
|
// inline asm
|
|
call (%rd224), _rt_buffer_get_64, (%rd225, %r28, %r245, %rd6, %rd7, %rd13, %rd13);
|
|
// inline asm
|
|
mov.f32 %f1435, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs133, %f1435;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs134, 0;
|
|
st.v4.u16 [%rd224], {%rs133, %rs133, %rs133, %rs134};
|
|
|
|
BB0_138:
|
|
ld.global.u32 %r246, [additive];
|
|
setp.eq.s32 %p163, %r246, 0;
|
|
@%p163 bra BB0_140;
|
|
|
|
mov.u64 %rd243, image_RNM2;
|
|
cvta.global.u64 %rd232, %rd243;
|
|
mov.u32 %r250, 8;
|
|
// inline asm
|
|
call (%rd231), _rt_buffer_get_64, (%rd232, %r28, %r250, %rd6, %rd7, %rd13, %rd13);
|
|
// inline asm
|
|
ld.v4.u16 {%rs141, %rs142, %rs143, %rs144}, [%rd231];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1436, %rs141;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1437, %rs142;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1438, %rs143;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd237), _rt_buffer_get_64, (%rd232, %r28, %r250, %rd6, %rd7, %rd13, %rd13);
|
|
// inline asm
|
|
add.f32 %f1439, %f1436, 0f00000000;
|
|
add.f32 %f1440, %f1437, 0f00000000;
|
|
add.f32 %f1441, %f1438, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs140, %f1441;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs139, %f1440;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs138, %f1439;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs145, 0;
|
|
st.v4.u16 [%rd237], {%rs138, %rs139, %rs140, %rs145};
|
|
bra.uni BB0_141;
|
|
|
|
BB0_140:
|
|
mov.u64 %rd250, image_RNM2;
|
|
cvta.global.u64 %rd245, %rd250;
|
|
mov.u32 %r252, 8;
|
|
// inline asm
|
|
call (%rd244), _rt_buffer_get_64, (%rd245, %r28, %r252, %rd6, %rd7, %rd13, %rd13);
|
|
// inline asm
|
|
mov.f32 %f1442, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs146, %f1442;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs147, 0;
|
|
st.v4.u16 [%rd244], {%rs146, %rs146, %rs146, %rs147};
|
|
|
|
BB0_141:
|
|
ld.global.u32 %r253, [additive];
|
|
setp.eq.s32 %p164, %r253, 0;
|
|
@%p164 bra BB0_143;
|
|
|
|
mov.u64 %rd263, image_RNM3;
|
|
cvta.global.u64 %rd252, %rd263;
|
|
mov.u32 %r257, 8;
|
|
// inline asm
|
|
call (%rd251), _rt_buffer_get_64, (%rd252, %r28, %r257, %rd6, %rd7, %rd13, %rd13);
|
|
// inline asm
|
|
ld.v4.u16 {%rs154, %rs155, %rs156, %rs157}, [%rd251];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1443, %rs154;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1444, %rs155;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1445, %rs156;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd257), _rt_buffer_get_64, (%rd252, %r28, %r257, %rd6, %rd7, %rd13, %rd13);
|
|
// inline asm
|
|
add.f32 %f1446, %f1443, 0f00000000;
|
|
add.f32 %f1447, %f1444, 0f00000000;
|
|
add.f32 %f1448, %f1445, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs153, %f1448;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs152, %f1447;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs151, %f1446;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs158, 0;
|
|
st.v4.u16 [%rd257], {%rs151, %rs152, %rs153, %rs158};
|
|
bra.uni BB0_144;
|
|
|
|
BB0_143:
|
|
mov.u64 %rd270, image_RNM3;
|
|
cvta.global.u64 %rd265, %rd270;
|
|
mov.u32 %r259, 8;
|
|
// inline asm
|
|
call (%rd264), _rt_buffer_get_64, (%rd265, %r28, %r259, %rd6, %rd7, %rd13, %rd13);
|
|
// inline asm
|
|
mov.f32 %f1449, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs159, %f1449;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs160, 0;
|
|
st.v4.u16 [%rd264], {%rs159, %rs159, %rs159, %rs160};
|
|
bra.uni BB0_144;
|
|
|
|
BB0_64:
|
|
setp.geu.f32 %p95, %f230, 0f00000000;
|
|
@%p95 bra BB0_67;
|
|
|
|
cvt.rzi.f32.f32 %f1100, %f1072;
|
|
setp.neu.f32 %p96, %f1100, 0f3EE8BA2E;
|
|
selp.f32 %f1549, 0f7FFFFFFF, %f1549, %p96;
|
|
|
|
BB0_67:
|
|
add.f32 %f1102, %f232, 0f3EE8BA2E;
|
|
mov.b32 %r146, %f1102;
|
|
setp.lt.s32 %p98, %r146, 2139095040;
|
|
@%p98 bra BB0_72;
|
|
|
|
setp.gtu.f32 %p99, %f232, 0f7F800000;
|
|
@%p99 bra BB0_71;
|
|
bra.uni BB0_69;
|
|
|
|
BB0_71:
|
|
add.f32 %f1549, %f230, 0f3EE8BA2E;
|
|
bra.uni BB0_72;
|
|
|
|
BB0_69:
|
|
setp.neu.f32 %p100, %f232, 0f7F800000;
|
|
@%p100 bra BB0_72;
|
|
|
|
selp.f32 %f1549, 0fFF800000, 0f7F800000, %p5;
|
|
|
|
BB0_72:
|
|
mul.f32 %f1103, %f1549, 0f437F0000;
|
|
setp.eq.f32 %p101, %f230, 0f3F800000;
|
|
selp.f32 %f1104, 0f437F0000, %f1103, %p101;
|
|
cvt.rzi.u32.f32 %r147, %f1104;
|
|
cvt.u16.u32 %rs17, %r147;
|
|
mov.u16 %rs18, 255;
|
|
st.v2.u8 [%rd43], {%rs17, %rs18};
|
|
ld.global.u32 %r266, [imageEnabled];
|
|
|
|
BB0_73:
|
|
and.b32 %r148, %r266, 1;
|
|
setp.eq.b32 %p102, %r148, 1;
|
|
@!%p102 bra BB0_108;
|
|
bra.uni BB0_74;
|
|
|
|
BB0_74:
|
|
mov.f32 %f1107, 0f3E666666;
|
|
cvt.rzi.f32.f32 %f1108, %f1107;
|
|
fma.rn.f32 %f1109, %f1108, 0fC0000000, 0f3EE66666;
|
|
abs.f32 %f244, %f1109;
|
|
abs.f32 %f245, %f1478;
|
|
setp.lt.f32 %p103, %f245, 0f00800000;
|
|
mul.f32 %f1110, %f245, 0f4B800000;
|
|
selp.f32 %f1111, 0fC3170000, 0fC2FE0000, %p103;
|
|
selp.f32 %f1112, %f1110, %f245, %p103;
|
|
mov.b32 %r149, %f1112;
|
|
and.b32 %r150, %r149, 8388607;
|
|
or.b32 %r151, %r150, 1065353216;
|
|
mov.b32 %f1113, %r151;
|
|
shr.u32 %r152, %r149, 23;
|
|
cvt.rn.f32.u32 %f1114, %r152;
|
|
add.f32 %f1115, %f1111, %f1114;
|
|
setp.gt.f32 %p104, %f1113, 0f3FB504F3;
|
|
mul.f32 %f1116, %f1113, 0f3F000000;
|
|
add.f32 %f1117, %f1115, 0f3F800000;
|
|
selp.f32 %f1118, %f1116, %f1113, %p104;
|
|
selp.f32 %f1119, %f1117, %f1115, %p104;
|
|
add.f32 %f1120, %f1118, 0fBF800000;
|
|
add.f32 %f1106, %f1118, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f1105,%f1106;
|
|
// inline asm
|
|
add.f32 %f1121, %f1120, %f1120;
|
|
mul.f32 %f1122, %f1105, %f1121;
|
|
mul.f32 %f1123, %f1122, %f1122;
|
|
mov.f32 %f1124, 0f3C4CAF63;
|
|
mov.f32 %f1125, 0f3B18F0FE;
|
|
fma.rn.f32 %f1126, %f1125, %f1123, %f1124;
|
|
mov.f32 %f1127, 0f3DAAAABD;
|
|
fma.rn.f32 %f1128, %f1126, %f1123, %f1127;
|
|
mul.rn.f32 %f1129, %f1128, %f1123;
|
|
mul.rn.f32 %f1130, %f1129, %f1122;
|
|
sub.f32 %f1131, %f1120, %f1122;
|
|
neg.f32 %f1132, %f1122;
|
|
add.f32 %f1133, %f1131, %f1131;
|
|
fma.rn.f32 %f1134, %f1132, %f1120, %f1133;
|
|
mul.rn.f32 %f1135, %f1105, %f1134;
|
|
add.f32 %f1136, %f1130, %f1122;
|
|
sub.f32 %f1137, %f1122, %f1136;
|
|
add.f32 %f1138, %f1130, %f1137;
|
|
add.f32 %f1139, %f1135, %f1138;
|
|
add.f32 %f1140, %f1136, %f1139;
|
|
sub.f32 %f1141, %f1136, %f1140;
|
|
add.f32 %f1142, %f1139, %f1141;
|
|
mov.f32 %f1143, 0f3F317200;
|
|
mul.rn.f32 %f1144, %f1119, %f1143;
|
|
mov.f32 %f1145, 0f35BFBE8E;
|
|
mul.rn.f32 %f1146, %f1119, %f1145;
|
|
add.f32 %f1147, %f1144, %f1140;
|
|
sub.f32 %f1148, %f1144, %f1147;
|
|
add.f32 %f1149, %f1140, %f1148;
|
|
add.f32 %f1150, %f1142, %f1149;
|
|
add.f32 %f1151, %f1146, %f1150;
|
|
add.f32 %f1152, %f1147, %f1151;
|
|
sub.f32 %f1153, %f1147, %f1152;
|
|
add.f32 %f1154, %f1151, %f1153;
|
|
mov.f32 %f1155, 0f3EE66666;
|
|
mul.rn.f32 %f1156, %f1155, %f1152;
|
|
neg.f32 %f1157, %f1156;
|
|
fma.rn.f32 %f1158, %f1155, %f1152, %f1157;
|
|
fma.rn.f32 %f1159, %f1155, %f1154, %f1158;
|
|
mov.f32 %f1160, 0f00000000;
|
|
fma.rn.f32 %f1161, %f1160, %f1152, %f1159;
|
|
add.rn.f32 %f1162, %f1156, %f1161;
|
|
neg.f32 %f1163, %f1162;
|
|
add.rn.f32 %f1164, %f1156, %f1163;
|
|
add.rn.f32 %f1165, %f1164, %f1161;
|
|
mov.b32 %r153, %f1162;
|
|
setp.eq.s32 %p105, %r153, 1118925336;
|
|
add.s32 %r154, %r153, -1;
|
|
mov.b32 %f1166, %r154;
|
|
add.f32 %f1167, %f1165, 0f37000000;
|
|
selp.f32 %f1168, %f1166, %f1162, %p105;
|
|
selp.f32 %f246, %f1167, %f1165, %p105;
|
|
mul.f32 %f1169, %f1168, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f1170, %f1169;
|
|
mov.f32 %f1171, 0fBF317200;
|
|
fma.rn.f32 %f1172, %f1170, %f1171, %f1168;
|
|
mov.f32 %f1173, 0fB5BFBE8E;
|
|
fma.rn.f32 %f1174, %f1170, %f1173, %f1172;
|
|
mul.f32 %f1175, %f1174, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f1176, %f1175;
|
|
add.f32 %f1177, %f1170, 0f00000000;
|
|
ex2.approx.f32 %f1178, %f1177;
|
|
mul.f32 %f1179, %f1176, %f1178;
|
|
setp.lt.f32 %p106, %f1168, 0fC2D20000;
|
|
selp.f32 %f1180, 0f00000000, %f1179, %p106;
|
|
setp.gt.f32 %p107, %f1168, 0f42D20000;
|
|
selp.f32 %f1550, 0f7F800000, %f1180, %p107;
|
|
setp.eq.f32 %p108, %f1550, 0f7F800000;
|
|
@%p108 bra BB0_76;
|
|
|
|
fma.rn.f32 %f1550, %f1550, %f246, %f1550;
|
|
|
|
BB0_76:
|
|
setp.lt.f32 %p109, %f1478, 0f00000000;
|
|
setp.eq.f32 %p110, %f244, 0f3F800000;
|
|
and.pred %p6, %p109, %p110;
|
|
mov.b32 %r155, %f1550;
|
|
xor.b32 %r156, %r155, -2147483648;
|
|
mov.b32 %f1181, %r156;
|
|
selp.f32 %f1552, %f1181, %f1550, %p6;
|
|
setp.eq.f32 %p111, %f1478, 0f00000000;
|
|
@%p111 bra BB0_79;
|
|
bra.uni BB0_77;
|
|
|
|
BB0_79:
|
|
add.f32 %f1184, %f1478, %f1478;
|
|
selp.f32 %f1552, %f1184, 0f00000000, %p110;
|
|
bra.uni BB0_80;
|
|
|
|
BB0_77:
|
|
setp.geu.f32 %p112, %f1478, 0f00000000;
|
|
@%p112 bra BB0_80;
|
|
|
|
cvt.rzi.f32.f32 %f1183, %f1155;
|
|
setp.neu.f32 %p113, %f1183, 0f3EE66666;
|
|
selp.f32 %f1552, 0f7FFFFFFF, %f1552, %p113;
|
|
|
|
BB0_80:
|
|
add.f32 %f1185, %f245, 0f3EE66666;
|
|
mov.b32 %r157, %f1185;
|
|
setp.lt.s32 %p115, %r157, 2139095040;
|
|
@%p115 bra BB0_85;
|
|
|
|
setp.gtu.f32 %p116, %f245, 0f7F800000;
|
|
@%p116 bra BB0_84;
|
|
bra.uni BB0_82;
|
|
|
|
BB0_84:
|
|
add.f32 %f1552, %f1478, 0f3EE66666;
|
|
bra.uni BB0_85;
|
|
|
|
BB0_82:
|
|
setp.neu.f32 %p117, %f245, 0f7F800000;
|
|
@%p117 bra BB0_85;
|
|
|
|
selp.f32 %f1552, 0fFF800000, 0f7F800000, %p6;
|
|
|
|
BB0_85:
|
|
setp.eq.f32 %p118, %f1478, 0f3F800000;
|
|
selp.f32 %f257, 0f3F800000, %f1552, %p118;
|
|
abs.f32 %f258, %f1477;
|
|
setp.lt.f32 %p119, %f258, 0f00800000;
|
|
mul.f32 %f1188, %f258, 0f4B800000;
|
|
selp.f32 %f1189, 0fC3170000, 0fC2FE0000, %p119;
|
|
selp.f32 %f1190, %f1188, %f258, %p119;
|
|
mov.b32 %r158, %f1190;
|
|
and.b32 %r159, %r158, 8388607;
|
|
or.b32 %r160, %r159, 1065353216;
|
|
mov.b32 %f1191, %r160;
|
|
shr.u32 %r161, %r158, 23;
|
|
cvt.rn.f32.u32 %f1192, %r161;
|
|
add.f32 %f1193, %f1189, %f1192;
|
|
setp.gt.f32 %p120, %f1191, 0f3FB504F3;
|
|
mul.f32 %f1194, %f1191, 0f3F000000;
|
|
add.f32 %f1195, %f1193, 0f3F800000;
|
|
selp.f32 %f1196, %f1194, %f1191, %p120;
|
|
selp.f32 %f1197, %f1195, %f1193, %p120;
|
|
add.f32 %f1198, %f1196, 0fBF800000;
|
|
add.f32 %f1187, %f1196, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f1186,%f1187;
|
|
// inline asm
|
|
add.f32 %f1199, %f1198, %f1198;
|
|
mul.f32 %f1200, %f1186, %f1199;
|
|
mul.f32 %f1201, %f1200, %f1200;
|
|
fma.rn.f32 %f1204, %f1125, %f1201, %f1124;
|
|
fma.rn.f32 %f1206, %f1204, %f1201, %f1127;
|
|
mul.rn.f32 %f1207, %f1206, %f1201;
|
|
mul.rn.f32 %f1208, %f1207, %f1200;
|
|
sub.f32 %f1209, %f1198, %f1200;
|
|
neg.f32 %f1210, %f1200;
|
|
add.f32 %f1211, %f1209, %f1209;
|
|
fma.rn.f32 %f1212, %f1210, %f1198, %f1211;
|
|
mul.rn.f32 %f1213, %f1186, %f1212;
|
|
add.f32 %f1214, %f1208, %f1200;
|
|
sub.f32 %f1215, %f1200, %f1214;
|
|
add.f32 %f1216, %f1208, %f1215;
|
|
add.f32 %f1217, %f1213, %f1216;
|
|
add.f32 %f1218, %f1214, %f1217;
|
|
sub.f32 %f1219, %f1214, %f1218;
|
|
add.f32 %f1220, %f1217, %f1219;
|
|
mul.rn.f32 %f1222, %f1197, %f1143;
|
|
mul.rn.f32 %f1224, %f1197, %f1145;
|
|
add.f32 %f1225, %f1222, %f1218;
|
|
sub.f32 %f1226, %f1222, %f1225;
|
|
add.f32 %f1227, %f1218, %f1226;
|
|
add.f32 %f1228, %f1220, %f1227;
|
|
add.f32 %f1229, %f1224, %f1228;
|
|
add.f32 %f1230, %f1225, %f1229;
|
|
sub.f32 %f1231, %f1225, %f1230;
|
|
add.f32 %f1232, %f1229, %f1231;
|
|
mul.rn.f32 %f1234, %f1155, %f1230;
|
|
neg.f32 %f1235, %f1234;
|
|
fma.rn.f32 %f1236, %f1155, %f1230, %f1235;
|
|
fma.rn.f32 %f1237, %f1155, %f1232, %f1236;
|
|
fma.rn.f32 %f1239, %f1160, %f1230, %f1237;
|
|
add.rn.f32 %f1240, %f1234, %f1239;
|
|
neg.f32 %f1241, %f1240;
|
|
add.rn.f32 %f1242, %f1234, %f1241;
|
|
add.rn.f32 %f1243, %f1242, %f1239;
|
|
mov.b32 %r162, %f1240;
|
|
setp.eq.s32 %p121, %r162, 1118925336;
|
|
add.s32 %r163, %r162, -1;
|
|
mov.b32 %f1244, %r163;
|
|
add.f32 %f1245, %f1243, 0f37000000;
|
|
selp.f32 %f1246, %f1244, %f1240, %p121;
|
|
selp.f32 %f259, %f1245, %f1243, %p121;
|
|
mul.f32 %f1247, %f1246, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f1248, %f1247;
|
|
fma.rn.f32 %f1250, %f1248, %f1171, %f1246;
|
|
fma.rn.f32 %f1252, %f1248, %f1173, %f1250;
|
|
mul.f32 %f1253, %f1252, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f1254, %f1253;
|
|
add.f32 %f1255, %f1248, 0f00000000;
|
|
ex2.approx.f32 %f1256, %f1255;
|
|
mul.f32 %f1257, %f1254, %f1256;
|
|
setp.lt.f32 %p122, %f1246, 0fC2D20000;
|
|
selp.f32 %f1258, 0f00000000, %f1257, %p122;
|
|
setp.gt.f32 %p123, %f1246, 0f42D20000;
|
|
selp.f32 %f1553, 0f7F800000, %f1258, %p123;
|
|
setp.eq.f32 %p124, %f1553, 0f7F800000;
|
|
@%p124 bra BB0_87;
|
|
|
|
fma.rn.f32 %f1553, %f1553, %f259, %f1553;
|
|
|
|
BB0_87:
|
|
setp.lt.f32 %p125, %f1477, 0f00000000;
|
|
and.pred %p7, %p125, %p110;
|
|
mov.b32 %r164, %f1553;
|
|
xor.b32 %r165, %r164, -2147483648;
|
|
mov.b32 %f1259, %r165;
|
|
selp.f32 %f1555, %f1259, %f1553, %p7;
|
|
setp.eq.f32 %p127, %f1477, 0f00000000;
|
|
@%p127 bra BB0_90;
|
|
bra.uni BB0_88;
|
|
|
|
BB0_90:
|
|
add.f32 %f1262, %f1477, %f1477;
|
|
selp.f32 %f1555, %f1262, 0f00000000, %p110;
|
|
bra.uni BB0_91;
|
|
|
|
BB0_88:
|
|
setp.geu.f32 %p128, %f1477, 0f00000000;
|
|
@%p128 bra BB0_91;
|
|
|
|
cvt.rzi.f32.f32 %f1261, %f1155;
|
|
setp.neu.f32 %p129, %f1261, 0f3EE66666;
|
|
selp.f32 %f1555, 0f7FFFFFFF, %f1555, %p129;
|
|
|
|
BB0_91:
|
|
add.f32 %f1263, %f258, 0f3EE66666;
|
|
mov.b32 %r166, %f1263;
|
|
setp.lt.s32 %p131, %r166, 2139095040;
|
|
@%p131 bra BB0_96;
|
|
|
|
setp.gtu.f32 %p132, %f258, 0f7F800000;
|
|
@%p132 bra BB0_95;
|
|
bra.uni BB0_93;
|
|
|
|
BB0_95:
|
|
add.f32 %f1555, %f1477, 0f3EE66666;
|
|
bra.uni BB0_96;
|
|
|
|
BB0_93:
|
|
setp.neu.f32 %p133, %f258, 0f7F800000;
|
|
@%p133 bra BB0_96;
|
|
|
|
selp.f32 %f1555, 0fFF800000, 0f7F800000, %p7;
|
|
|
|
BB0_96:
|
|
setp.eq.f32 %p134, %f1477, 0f3F800000;
|
|
selp.f32 %f270, 0f3F800000, %f1555, %p134;
|
|
abs.f32 %f271, %f1476;
|
|
setp.lt.f32 %p135, %f271, 0f00800000;
|
|
mul.f32 %f1266, %f271, 0f4B800000;
|
|
selp.f32 %f1267, 0fC3170000, 0fC2FE0000, %p135;
|
|
selp.f32 %f1268, %f1266, %f271, %p135;
|
|
mov.b32 %r167, %f1268;
|
|
and.b32 %r168, %r167, 8388607;
|
|
or.b32 %r169, %r168, 1065353216;
|
|
mov.b32 %f1269, %r169;
|
|
shr.u32 %r170, %r167, 23;
|
|
cvt.rn.f32.u32 %f1270, %r170;
|
|
add.f32 %f1271, %f1267, %f1270;
|
|
setp.gt.f32 %p136, %f1269, 0f3FB504F3;
|
|
mul.f32 %f1272, %f1269, 0f3F000000;
|
|
add.f32 %f1273, %f1271, 0f3F800000;
|
|
selp.f32 %f1274, %f1272, %f1269, %p136;
|
|
selp.f32 %f1275, %f1273, %f1271, %p136;
|
|
add.f32 %f1276, %f1274, 0fBF800000;
|
|
add.f32 %f1265, %f1274, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f1264,%f1265;
|
|
// inline asm
|
|
add.f32 %f1277, %f1276, %f1276;
|
|
mul.f32 %f1278, %f1264, %f1277;
|
|
mul.f32 %f1279, %f1278, %f1278;
|
|
fma.rn.f32 %f1282, %f1125, %f1279, %f1124;
|
|
fma.rn.f32 %f1284, %f1282, %f1279, %f1127;
|
|
mul.rn.f32 %f1285, %f1284, %f1279;
|
|
mul.rn.f32 %f1286, %f1285, %f1278;
|
|
sub.f32 %f1287, %f1276, %f1278;
|
|
neg.f32 %f1288, %f1278;
|
|
add.f32 %f1289, %f1287, %f1287;
|
|
fma.rn.f32 %f1290, %f1288, %f1276, %f1289;
|
|
mul.rn.f32 %f1291, %f1264, %f1290;
|
|
add.f32 %f1292, %f1286, %f1278;
|
|
sub.f32 %f1293, %f1278, %f1292;
|
|
add.f32 %f1294, %f1286, %f1293;
|
|
add.f32 %f1295, %f1291, %f1294;
|
|
add.f32 %f1296, %f1292, %f1295;
|
|
sub.f32 %f1297, %f1292, %f1296;
|
|
add.f32 %f1298, %f1295, %f1297;
|
|
mul.rn.f32 %f1300, %f1275, %f1143;
|
|
mul.rn.f32 %f1302, %f1275, %f1145;
|
|
add.f32 %f1303, %f1300, %f1296;
|
|
sub.f32 %f1304, %f1300, %f1303;
|
|
add.f32 %f1305, %f1296, %f1304;
|
|
add.f32 %f1306, %f1298, %f1305;
|
|
add.f32 %f1307, %f1302, %f1306;
|
|
add.f32 %f1308, %f1303, %f1307;
|
|
sub.f32 %f1309, %f1303, %f1308;
|
|
add.f32 %f1310, %f1307, %f1309;
|
|
mul.rn.f32 %f1312, %f1155, %f1308;
|
|
neg.f32 %f1313, %f1312;
|
|
fma.rn.f32 %f1314, %f1155, %f1308, %f1313;
|
|
fma.rn.f32 %f1315, %f1155, %f1310, %f1314;
|
|
fma.rn.f32 %f1317, %f1160, %f1308, %f1315;
|
|
add.rn.f32 %f1318, %f1312, %f1317;
|
|
neg.f32 %f1319, %f1318;
|
|
add.rn.f32 %f1320, %f1312, %f1319;
|
|
add.rn.f32 %f1321, %f1320, %f1317;
|
|
mov.b32 %r171, %f1318;
|
|
setp.eq.s32 %p137, %r171, 1118925336;
|
|
add.s32 %r172, %r171, -1;
|
|
mov.b32 %f1322, %r172;
|
|
add.f32 %f1323, %f1321, 0f37000000;
|
|
selp.f32 %f1324, %f1322, %f1318, %p137;
|
|
selp.f32 %f272, %f1323, %f1321, %p137;
|
|
mul.f32 %f1325, %f1324, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f1326, %f1325;
|
|
fma.rn.f32 %f1328, %f1326, %f1171, %f1324;
|
|
fma.rn.f32 %f1330, %f1326, %f1173, %f1328;
|
|
mul.f32 %f1331, %f1330, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f1332, %f1331;
|
|
add.f32 %f1333, %f1326, 0f00000000;
|
|
ex2.approx.f32 %f1334, %f1333;
|
|
mul.f32 %f1335, %f1332, %f1334;
|
|
setp.lt.f32 %p138, %f1324, 0fC2D20000;
|
|
selp.f32 %f1336, 0f00000000, %f1335, %p138;
|
|
setp.gt.f32 %p139, %f1324, 0f42D20000;
|
|
selp.f32 %f1556, 0f7F800000, %f1336, %p139;
|
|
setp.eq.f32 %p140, %f1556, 0f7F800000;
|
|
@%p140 bra BB0_98;
|
|
|
|
fma.rn.f32 %f1556, %f1556, %f272, %f1556;
|
|
|
|
BB0_98:
|
|
setp.lt.f32 %p141, %f1476, 0f00000000;
|
|
and.pred %p8, %p141, %p110;
|
|
mov.b32 %r173, %f1556;
|
|
xor.b32 %r174, %r173, -2147483648;
|
|
mov.b32 %f1337, %r174;
|
|
selp.f32 %f1558, %f1337, %f1556, %p8;
|
|
setp.eq.f32 %p143, %f1476, 0f00000000;
|
|
@%p143 bra BB0_101;
|
|
bra.uni BB0_99;
|
|
|
|
BB0_101:
|
|
add.f32 %f1340, %f1476, %f1476;
|
|
selp.f32 %f1558, %f1340, 0f00000000, %p110;
|
|
bra.uni BB0_102;
|
|
|
|
BB0_99:
|
|
setp.geu.f32 %p144, %f1476, 0f00000000;
|
|
@%p144 bra BB0_102;
|
|
|
|
cvt.rzi.f32.f32 %f1339, %f1155;
|
|
setp.neu.f32 %p145, %f1339, 0f3EE66666;
|
|
selp.f32 %f1558, 0f7FFFFFFF, %f1558, %p145;
|
|
|
|
BB0_102:
|
|
add.f32 %f1341, %f271, 0f3EE66666;
|
|
mov.b32 %r175, %f1341;
|
|
setp.lt.s32 %p147, %r175, 2139095040;
|
|
@%p147 bra BB0_107;
|
|
|
|
setp.gtu.f32 %p148, %f271, 0f7F800000;
|
|
@%p148 bra BB0_106;
|
|
bra.uni BB0_104;
|
|
|
|
BB0_106:
|
|
add.f32 %f1558, %f1476, 0f3EE66666;
|
|
bra.uni BB0_107;
|
|
|
|
BB0_104:
|
|
setp.neu.f32 %p149, %f271, 0f7F800000;
|
|
@%p149 bra BB0_107;
|
|
|
|
selp.f32 %f1558, 0fFF800000, 0f7F800000, %p8;
|
|
|
|
BB0_107:
|
|
setp.eq.f32 %p150, %f1476, 0f3F800000;
|
|
selp.f32 %f1342, 0f3F800000, %f1558, %p150;
|
|
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, %r29, %rd52, %rd53, %rd13, %rd13);
|
|
// inline asm
|
|
cvt.sat.f32.f32 %f1343, %f1342;
|
|
mul.f32 %f1344, %f1343, 0f437FFD71;
|
|
cvt.rzi.u32.f32 %r178, %f1344;
|
|
cvt.sat.f32.f32 %f1345, %f270;
|
|
mul.f32 %f1346, %f1345, 0f437FFD71;
|
|
cvt.rzi.u32.f32 %r179, %f1346;
|
|
cvt.sat.f32.f32 %f1347, %f257;
|
|
mul.f32 %f1348, %f1347, 0f437FFD71;
|
|
cvt.rzi.u32.f32 %r180, %f1348;
|
|
cvt.u16.u32 %rs19, %r178;
|
|
cvt.u16.u32 %rs20, %r180;
|
|
cvt.u16.u32 %rs21, %r179;
|
|
mov.u16 %rs22, 255;
|
|
st.v4.u8 [%rd50], {%rs19, %rs21, %rs20, %rs22};
|
|
ld.global.u32 %r266, [imageEnabled];
|
|
|
|
BB0_108:
|
|
cvt.u64.u32 %rd4, %r2;
|
|
cvt.u64.u32 %rd5, %r3;
|
|
and.b32 %r181, %r266, 4;
|
|
setp.eq.s32 %p151, %r181, 0;
|
|
@%p151 bra BB0_112;
|
|
|
|
ld.global.u32 %r182, [additive];
|
|
setp.eq.s32 %p152, %r182, 0;
|
|
mov.f32 %f1349, 0f3F800000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs23, %f1349;}
|
|
|
|
// inline asm
|
|
@%p152 bra BB0_111;
|
|
|
|
mov.u64 %rd69, image_HDR;
|
|
cvta.global.u64 %rd58, %rd69;
|
|
mov.u32 %r186, 8;
|
|
// inline asm
|
|
call (%rd57), _rt_buffer_get_64, (%rd58, %r28, %r186, %rd4, %rd5, %rd13, %rd13);
|
|
// inline asm
|
|
ld.v4.u16 {%rs30, %rs31, %rs32, %rs33}, [%rd57];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1350, %rs30;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1351, %rs31;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1352, %rs32;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd63), _rt_buffer_get_64, (%rd58, %r28, %r186, %rd4, %rd5, %rd13, %rd13);
|
|
// inline asm
|
|
add.f32 %f1353, %f1478, %f1350;
|
|
add.f32 %f1354, %f1477, %f1351;
|
|
add.f32 %f1355, %f1476, %f1352;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs29, %f1355;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs28, %f1354;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs27, %f1353;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd63], {%rs27, %rs28, %rs29, %rs23};
|
|
bra.uni BB0_112;
|
|
|
|
BB0_111:
|
|
mov.u64 %rd76, image_HDR;
|
|
cvta.global.u64 %rd71, %rd76;
|
|
mov.u32 %r188, 8;
|
|
// inline asm
|
|
call (%rd70), _rt_buffer_get_64, (%rd71, %r28, %r188, %rd4, %rd5, %rd13, %rd13);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs36, %f1476;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs35, %f1477;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs34, %f1478;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd70], {%rs34, %rs35, %rs36, %rs23};
|
|
|
|
BB0_112:
|
|
mov.f32 %f1360, 0f34000000;
|
|
max.f32 %f1361, %f1475, %f1360;
|
|
div.rn.f32 %f1362, %f1472, %f1361;
|
|
max.f32 %f1363, %f1474, %f1360;
|
|
div.rn.f32 %f1364, %f1471, %f1363;
|
|
max.f32 %f1365, %f1473, %f1360;
|
|
div.rn.f32 %f1366, %f1470, %f1365;
|
|
fma.rn.f32 %f283, %f1362, 0f3F000000, 0f3F000000;
|
|
fma.rn.f32 %f284, %f1364, 0f3F000000, 0f3F000000;
|
|
fma.rn.f32 %f285, %f1366, 0f3F000000, 0f3F000000;
|
|
div.rn.f32 %f1367, %f1469, %f1361;
|
|
div.rn.f32 %f1368, %f1468, %f1363;
|
|
div.rn.f32 %f1369, %f1467, %f1365;
|
|
fma.rn.f32 %f286, %f1367, 0f3F000000, 0f3F000000;
|
|
fma.rn.f32 %f287, %f1368, 0f3F000000, 0f3F000000;
|
|
fma.rn.f32 %f288, %f1369, 0f3F000000, 0f3F000000;
|
|
div.rn.f32 %f1370, %f1466, %f1361;
|
|
div.rn.f32 %f1371, %f1465, %f1363;
|
|
div.rn.f32 %f1372, %f1464, %f1365;
|
|
fma.rn.f32 %f289, %f1370, 0f3F000000, 0f3F000000;
|
|
fma.rn.f32 %f290, %f1371, 0f3F000000, 0f3F000000;
|
|
fma.rn.f32 %f291, %f1372, 0f3F000000, 0f3F000000;
|
|
ld.global.u32 %r189, [additive];
|
|
setp.eq.s32 %p153, %r189, 0;
|
|
mov.f32 %f1359, 0f3F800000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs37, %f1359;}
|
|
|
|
// inline asm
|
|
@%p153 bra BB0_114;
|
|
|
|
mov.u64 %rd89, image_RNM0;
|
|
cvta.global.u64 %rd78, %rd89;
|
|
mov.u32 %r193, 8;
|
|
// inline asm
|
|
call (%rd77), _rt_buffer_get_64, (%rd78, %r28, %r193, %rd4, %rd5, %rd13, %rd13);
|
|
// inline asm
|
|
ld.v4.u16 {%rs44, %rs45, %rs46, %rs47}, [%rd77];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1373, %rs44;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1374, %rs45;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1375, %rs46;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd83), _rt_buffer_get_64, (%rd78, %r28, %r193, %rd4, %rd5, %rd13, %rd13);
|
|
// inline asm
|
|
add.f32 %f1376, %f1475, %f1373;
|
|
add.f32 %f1377, %f1474, %f1374;
|
|
add.f32 %f1378, %f1473, %f1375;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs43, %f1378;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs42, %f1377;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs41, %f1376;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd83], {%rs41, %rs42, %rs43, %rs37};
|
|
bra.uni BB0_115;
|
|
|
|
BB0_114:
|
|
mov.u64 %rd96, image_RNM0;
|
|
cvta.global.u64 %rd91, %rd96;
|
|
mov.u32 %r195, 8;
|
|
// inline asm
|
|
call (%rd90), _rt_buffer_get_64, (%rd91, %r28, %r195, %rd4, %rd5, %rd13, %rd13);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs50, %f1473;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs49, %f1474;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs48, %f1475;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd90], {%rs48, %rs49, %rs50, %rs37};
|
|
|
|
BB0_115:
|
|
ld.global.u32 %r196, [additive];
|
|
setp.eq.s32 %p154, %r196, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs51, %f1359;}
|
|
|
|
// inline asm
|
|
@%p154 bra BB0_117;
|
|
|
|
mov.u64 %rd109, image_RNM1;
|
|
cvta.global.u64 %rd98, %rd109;
|
|
mov.u32 %r200, 8;
|
|
// inline asm
|
|
call (%rd97), _rt_buffer_get_64, (%rd98, %r28, %r200, %rd4, %rd5, %rd13, %rd13);
|
|
// inline asm
|
|
ld.v4.u16 {%rs58, %rs59, %rs60, %rs61}, [%rd97];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1383, %rs58;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1384, %rs59;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1385, %rs60;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd103), _rt_buffer_get_64, (%rd98, %r28, %r200, %rd4, %rd5, %rd13, %rd13);
|
|
// inline asm
|
|
add.f32 %f1386, %f283, %f1383;
|
|
add.f32 %f1387, %f284, %f1384;
|
|
add.f32 %f1388, %f285, %f1385;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs57, %f1388;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs56, %f1387;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs55, %f1386;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd103], {%rs55, %rs56, %rs57, %rs51};
|
|
bra.uni BB0_118;
|
|
|
|
BB0_117:
|
|
mov.u64 %rd116, image_RNM1;
|
|
cvta.global.u64 %rd111, %rd116;
|
|
mov.u32 %r202, 8;
|
|
// inline asm
|
|
call (%rd110), _rt_buffer_get_64, (%rd111, %r28, %r202, %rd4, %rd5, %rd13, %rd13);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs64, %f285;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs63, %f284;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs62, %f283;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd110], {%rs62, %rs63, %rs64, %rs51};
|
|
|
|
BB0_118:
|
|
ld.global.u32 %r203, [additive];
|
|
setp.eq.s32 %p155, %r203, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs65, %f1359;}
|
|
|
|
// inline asm
|
|
@%p155 bra BB0_120;
|
|
|
|
mov.u64 %rd129, image_RNM2;
|
|
cvta.global.u64 %rd118, %rd129;
|
|
mov.u32 %r207, 8;
|
|
// inline asm
|
|
call (%rd117), _rt_buffer_get_64, (%rd118, %r28, %r207, %rd4, %rd5, %rd13, %rd13);
|
|
// inline asm
|
|
ld.v4.u16 {%rs72, %rs73, %rs74, %rs75}, [%rd117];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1393, %rs72;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1394, %rs73;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1395, %rs74;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd123), _rt_buffer_get_64, (%rd118, %r28, %r207, %rd4, %rd5, %rd13, %rd13);
|
|
// inline asm
|
|
add.f32 %f1396, %f286, %f1393;
|
|
add.f32 %f1397, %f287, %f1394;
|
|
add.f32 %f1398, %f288, %f1395;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs71, %f1398;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs70, %f1397;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs69, %f1396;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd123], {%rs69, %rs70, %rs71, %rs65};
|
|
bra.uni BB0_121;
|
|
|
|
BB0_120:
|
|
mov.u64 %rd136, image_RNM2;
|
|
cvta.global.u64 %rd131, %rd136;
|
|
mov.u32 %r209, 8;
|
|
// inline asm
|
|
call (%rd130), _rt_buffer_get_64, (%rd131, %r28, %r209, %rd4, %rd5, %rd13, %rd13);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs78, %f288;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs77, %f287;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs76, %f286;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd130], {%rs76, %rs77, %rs78, %rs65};
|
|
|
|
BB0_121:
|
|
ld.global.u32 %r210, [additive];
|
|
setp.eq.s32 %p156, %r210, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs79, %f1359;}
|
|
|
|
// inline asm
|
|
@%p156 bra BB0_123;
|
|
|
|
mov.u64 %rd149, image_RNM3;
|
|
cvta.global.u64 %rd138, %rd149;
|
|
mov.u32 %r214, 8;
|
|
// inline asm
|
|
call (%rd137), _rt_buffer_get_64, (%rd138, %r28, %r214, %rd4, %rd5, %rd13, %rd13);
|
|
// inline asm
|
|
ld.v4.u16 {%rs86, %rs87, %rs88, %rs89}, [%rd137];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1403, %rs86;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1404, %rs87;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1405, %rs88;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd143), _rt_buffer_get_64, (%rd138, %r28, %r214, %rd4, %rd5, %rd13, %rd13);
|
|
// inline asm
|
|
add.f32 %f1406, %f289, %f1403;
|
|
add.f32 %f1407, %f290, %f1404;
|
|
add.f32 %f1408, %f291, %f1405;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs85, %f1408;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs84, %f1407;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs83, %f1406;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd143], {%rs83, %rs84, %rs85, %rs79};
|
|
bra.uni BB0_144;
|
|
|
|
BB0_123:
|
|
mov.u64 %rd156, image_RNM3;
|
|
cvta.global.u64 %rd151, %rd156;
|
|
mov.u32 %r216, 8;
|
|
// inline asm
|
|
call (%rd150), _rt_buffer_get_64, (%rd151, %r28, %r216, %rd4, %rd5, %rd13, %rd13);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs92, %f291;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs91, %f290;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs90, %f289;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd150], {%rs90, %rs91, %rs92, %rs79};
|
|
|
|
BB0_144:
|
|
ret;
|
|
}
|
|
|
|
|