2727 lines
79 KiB
Plaintext
2727 lines
79 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[12];
|
|
.reg .b64 %SP;
|
|
.reg .b64 %SPL;
|
|
.reg .pred %p<166>;
|
|
.reg .b16 %rs<162>;
|
|
.reg .f32 %f<1659>;
|
|
.reg .b32 %r<274>;
|
|
.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 %f1548, 0f00000000;
|
|
mov.f32 %f1549, %f1548;
|
|
mov.f32 %f1550, %f1548;
|
|
@%p9 bra BB0_2;
|
|
|
|
ld.u8 %rs12, [%rd8+1];
|
|
and.b16 %rs14, %rs10, 255;
|
|
cvt.rn.f32.u16 %f321, %rs14;
|
|
div.rn.f32 %f322, %f321, 0f437F0000;
|
|
fma.rn.f32 %f323, %f322, 0f40000000, 0fBF800000;
|
|
cvt.rn.f32.u16 %f324, %rs12;
|
|
div.rn.f32 %f325, %f324, 0f437F0000;
|
|
fma.rn.f32 %f326, %f325, 0f40000000, 0fBF800000;
|
|
cvt.rn.f32.u16 %f327, %rs9;
|
|
div.rn.f32 %f328, %f327, 0f437F0000;
|
|
fma.rn.f32 %f329, %f328, 0f40000000, 0fBF800000;
|
|
mul.f32 %f330, %f326, %f326;
|
|
fma.rn.f32 %f331, %f323, %f323, %f330;
|
|
fma.rn.f32 %f332, %f329, %f329, %f331;
|
|
sqrt.rn.f32 %f333, %f332;
|
|
rcp.rn.f32 %f334, %f333;
|
|
mul.f32 %f1548, %f323, %f334;
|
|
mul.f32 %f1549, %f326, %f334;
|
|
mul.f32 %f1550, %f329, %f334;
|
|
|
|
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, %f1549, 0f00000000;
|
|
setp.eq.f32 %p11, %f1548, 0f00000000;
|
|
and.pred %p12, %p11, %p10;
|
|
setp.eq.f32 %p13, %f1550, 0f00000000;
|
|
and.pred %p14, %p12, %p13;
|
|
@%p14 bra BB0_124;
|
|
bra.uni BB0_3;
|
|
|
|
BB0_124:
|
|
ld.global.u32 %r273, [imageEnabled];
|
|
and.b32 %r221, %r273, 1;
|
|
setp.eq.b32 %p157, %r221, 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 %r273, [imageEnabled];
|
|
|
|
BB0_126:
|
|
and.b32 %r224, %r273, 8;
|
|
setp.eq.s32 %p158, %r224, 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 %f1502, 0f00000000;
|
|
cvt.rzi.u32.f32 %r227, %f1502;
|
|
cvt.u16.u32 %rs94, %r227;
|
|
mov.u16 %rs95, 0;
|
|
st.v2.u8 [%rd164], {%rs94, %rs95};
|
|
ld.global.u32 %r273, [imageEnabled];
|
|
|
|
BB0_128:
|
|
cvt.u64.u32 %rd6, %r2;
|
|
cvt.u64.u32 %rd7, %r3;
|
|
and.b32 %r228, %r273, 4;
|
|
setp.eq.s32 %p159, %r228, 0;
|
|
@%p159 bra BB0_132;
|
|
|
|
ld.global.u32 %r229, [additive];
|
|
setp.eq.s32 %p160, %r229, 0;
|
|
@%p160 bra BB0_131;
|
|
|
|
mov.u64 %rd183, image_HDR;
|
|
cvta.global.u64 %rd172, %rd183;
|
|
mov.u32 %r233, 8;
|
|
// inline asm
|
|
call (%rd171), _rt_buffer_get_64, (%rd172, %r28, %r233, %rd6, %rd7, %rd13, %rd13);
|
|
// inline asm
|
|
ld.v4.u16 {%rs102, %rs103, %rs104, %rs105}, [%rd171];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1503, %rs102;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1504, %rs103;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1505, %rs104;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd177), _rt_buffer_get_64, (%rd172, %r28, %r233, %rd6, %rd7, %rd13, %rd13);
|
|
// inline asm
|
|
add.f32 %f1506, %f1503, 0f00000000;
|
|
add.f32 %f1507, %f1504, 0f00000000;
|
|
add.f32 %f1508, %f1505, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs101, %f1508;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs100, %f1507;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs99, %f1506;}
|
|
|
|
// 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 %f351, %f7, 0f3456BF95;
|
|
mul.f32 %f352, %f8, 0f3456BF95;
|
|
mul.f32 %f353, %f9, 0f3456BF95;
|
|
abs.f32 %f354, %f1548;
|
|
div.rn.f32 %f355, %f351, %f354;
|
|
abs.f32 %f356, %f1549;
|
|
div.rn.f32 %f357, %f352, %f356;
|
|
abs.f32 %f358, %f1550;
|
|
div.rn.f32 %f359, %f353, %f358;
|
|
abs.f32 %f360, %f355;
|
|
abs.f32 %f361, %f357;
|
|
abs.f32 %f362, %f359;
|
|
mov.f32 %f363, 0f38D1B717;
|
|
max.f32 %f364, %f360, %f363;
|
|
max.f32 %f365, %f361, %f363;
|
|
max.f32 %f366, %f362, %f363;
|
|
fma.rn.f32 %f10, %f1548, %f364, %f7;
|
|
fma.rn.f32 %f11, %f1549, %f365, %f8;
|
|
fma.rn.f32 %f12, %f1550, %f366, %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 %p15, %r4, 0;
|
|
mov.f32 %f1551, 0f00000000;
|
|
mov.f32 %f1552, %f1551;
|
|
mov.f32 %f1553, %f1551;
|
|
mov.f32 %f1554, %f1551;
|
|
mov.f32 %f1555, %f1551;
|
|
mov.f32 %f1556, %f1551;
|
|
mov.f32 %f1557, %f1551;
|
|
mov.f32 %f1558, %f1551;
|
|
mov.f32 %f1559, %f1551;
|
|
mov.f32 %f1560, %f1551;
|
|
mov.f32 %f1561, %f1551;
|
|
mov.f32 %f1562, %f1551;
|
|
mov.f32 %f1563, %f1551;
|
|
mov.f32 %f1564, %f1551;
|
|
mov.f32 %f1565, %f1551;
|
|
mov.f32 %f1566, %f1551;
|
|
@%p15 bra BB0_60;
|
|
|
|
mov.f32 %f383, 0f40000000;
|
|
cvt.rzi.f32.f32 %f384, %f383;
|
|
add.f32 %f385, %f384, %f384;
|
|
mov.f32 %f386, 0f40800000;
|
|
sub.f32 %f387, %f386, %f385;
|
|
abs.f32 %f15, %f387;
|
|
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 %f382, 0f00000000;
|
|
mov.u32 %r265, 0;
|
|
abs.f32 %f587, %f16;
|
|
abs.f32 %f588, %f17;
|
|
max.f32 %f589, %f587, %f588;
|
|
abs.f32 %f590, %f18;
|
|
max.f32 %f591, %f589, %f590;
|
|
mov.f32 %f1551, %f382;
|
|
mov.f32 %f1552, %f382;
|
|
mov.f32 %f1553, %f382;
|
|
mov.f32 %f1554, %f382;
|
|
mov.f32 %f1555, %f382;
|
|
mov.f32 %f1556, %f382;
|
|
mov.f32 %f1557, %f382;
|
|
mov.f32 %f1558, %f382;
|
|
mov.f32 %f1559, %f382;
|
|
mov.f32 %f1560, %f382;
|
|
mov.f32 %f1561, %f382;
|
|
mov.f32 %f1562, %f382;
|
|
mov.f32 %f1563, %f382;
|
|
mov.f32 %f1564, %f382;
|
|
mov.f32 %f1565, %f382;
|
|
mov.f32 %f1566, %f382;
|
|
|
|
BB0_5:
|
|
cvt.u64.u32 %rd30, %r265;
|
|
// inline asm
|
|
call (%rd28), _rt_buffer_get_64, (%rd25, %r45, %r46, %rd30, %rd13, %rd13, %rd13);
|
|
// inline asm
|
|
ld.v4.f32 {%f390, %f391, %f392, %f393}, [%rd28+80];
|
|
ld.v4.f32 {%f394, %f395, %f396, %f397}, [%rd28+64];
|
|
ld.v4.f32 {%f398, %f399, %f400, %f401}, [%rd28+48];
|
|
ld.v4.f32 {%f402, %f1575, %f1576, %f405}, [%rd28+32];
|
|
ld.v4.f32 {%f406, %f407, %f408, %f409}, [%rd28+16];
|
|
ld.v4.f32 {%f410, %f411, %f412, %f413}, [%rd28];
|
|
mov.b32 %r6, %f393;
|
|
sub.f32 %f61, %f411, %f7;
|
|
sub.f32 %f63, %f412, %f8;
|
|
sub.f32 %f65, %f413, %f9;
|
|
mul.f32 %f415, %f63, %f63;
|
|
fma.rn.f32 %f416, %f61, %f61, %f415;
|
|
fma.rn.f32 %f417, %f65, %f65, %f416;
|
|
sqrt.rn.f32 %f66, %f417;
|
|
rcp.rn.f32 %f418, %f66;
|
|
mul.f32 %f67, %f61, %f418;
|
|
mul.f32 %f68, %f63, %f418;
|
|
mul.f32 %f69, %f65, %f418;
|
|
mul.f32 %f70, %f66, %f409;
|
|
abs.f32 %f71, %f70;
|
|
setp.lt.f32 %p16, %f71, 0f00800000;
|
|
mul.f32 %f419, %f71, 0f4B800000;
|
|
selp.f32 %f420, 0fC3170000, 0fC2FE0000, %p16;
|
|
selp.f32 %f421, %f419, %f71, %p16;
|
|
mov.b32 %r58, %f421;
|
|
and.b32 %r59, %r58, 8388607;
|
|
or.b32 %r60, %r59, 1065353216;
|
|
mov.b32 %f422, %r60;
|
|
shr.u32 %r61, %r58, 23;
|
|
cvt.rn.f32.u32 %f423, %r61;
|
|
add.f32 %f424, %f420, %f423;
|
|
setp.gt.f32 %p17, %f422, 0f3FB504F3;
|
|
mul.f32 %f425, %f422, 0f3F000000;
|
|
add.f32 %f426, %f424, 0f3F800000;
|
|
selp.f32 %f427, %f425, %f422, %p17;
|
|
selp.f32 %f428, %f426, %f424, %p17;
|
|
add.f32 %f429, %f427, 0fBF800000;
|
|
add.f32 %f389, %f427, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f388,%f389;
|
|
// inline asm
|
|
add.f32 %f430, %f429, %f429;
|
|
mul.f32 %f431, %f388, %f430;
|
|
mul.f32 %f432, %f431, %f431;
|
|
mov.f32 %f433, 0f3C4CAF63;
|
|
mov.f32 %f434, 0f3B18F0FE;
|
|
fma.rn.f32 %f435, %f434, %f432, %f433;
|
|
mov.f32 %f436, 0f3DAAAABD;
|
|
fma.rn.f32 %f437, %f435, %f432, %f436;
|
|
mul.rn.f32 %f438, %f437, %f432;
|
|
mul.rn.f32 %f439, %f438, %f431;
|
|
sub.f32 %f440, %f429, %f431;
|
|
neg.f32 %f441, %f431;
|
|
add.f32 %f442, %f440, %f440;
|
|
fma.rn.f32 %f443, %f441, %f429, %f442;
|
|
mul.rn.f32 %f444, %f388, %f443;
|
|
add.f32 %f445, %f439, %f431;
|
|
sub.f32 %f446, %f431, %f445;
|
|
add.f32 %f447, %f439, %f446;
|
|
add.f32 %f448, %f444, %f447;
|
|
add.f32 %f449, %f445, %f448;
|
|
sub.f32 %f450, %f445, %f449;
|
|
add.f32 %f451, %f448, %f450;
|
|
mov.f32 %f452, 0f3F317200;
|
|
mul.rn.f32 %f453, %f428, %f452;
|
|
mov.f32 %f454, 0f35BFBE8E;
|
|
mul.rn.f32 %f455, %f428, %f454;
|
|
add.f32 %f456, %f453, %f449;
|
|
sub.f32 %f457, %f453, %f456;
|
|
add.f32 %f458, %f449, %f457;
|
|
add.f32 %f459, %f451, %f458;
|
|
add.f32 %f460, %f455, %f459;
|
|
add.f32 %f461, %f456, %f460;
|
|
sub.f32 %f462, %f456, %f461;
|
|
add.f32 %f463, %f460, %f462;
|
|
mul.rn.f32 %f72, %f386, %f461;
|
|
neg.f32 %f465, %f72;
|
|
fma.rn.f32 %f466, %f386, %f461, %f465;
|
|
fma.rn.f32 %f467, %f386, %f463, %f466;
|
|
fma.rn.f32 %f73, %f382, %f461, %f467;
|
|
add.rn.f32 %f74, %f72, %f73;
|
|
mov.b32 %r62, %f74;
|
|
setp.eq.s32 %p1, %r62, 1118925336;
|
|
add.s32 %r63, %r62, -1;
|
|
mov.b32 %f469, %r63;
|
|
selp.f32 %f470, %f469, %f74, %p1;
|
|
mul.f32 %f471, %f470, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f472, %f471;
|
|
mov.f32 %f473, 0fBF317200;
|
|
fma.rn.f32 %f474, %f472, %f473, %f470;
|
|
mov.f32 %f475, 0fB5BFBE8E;
|
|
fma.rn.f32 %f476, %f472, %f475, %f474;
|
|
mul.f32 %f477, %f476, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f478, %f477;
|
|
add.f32 %f479, %f472, 0f00000000;
|
|
ex2.approx.f32 %f480, %f479;
|
|
mul.f32 %f481, %f478, %f480;
|
|
setp.lt.f32 %p18, %f470, 0fC2D20000;
|
|
selp.f32 %f482, 0f00000000, %f481, %p18;
|
|
setp.gt.f32 %p19, %f470, 0f42D20000;
|
|
selp.f32 %f1567, 0f7F800000, %f482, %p19;
|
|
setp.eq.f32 %p20, %f1567, 0f7F800000;
|
|
@%p20 bra BB0_7;
|
|
|
|
neg.f32 %f483, %f74;
|
|
add.rn.f32 %f484, %f72, %f483;
|
|
add.rn.f32 %f485, %f484, %f73;
|
|
add.f32 %f486, %f485, 0f37000000;
|
|
selp.f32 %f487, %f486, %f485, %p1;
|
|
fma.rn.f32 %f1567, %f1567, %f487, %f1567;
|
|
|
|
BB0_7:
|
|
setp.lt.f32 %p21, %f70, 0f00000000;
|
|
setp.eq.f32 %p22, %f15, 0f3F800000;
|
|
and.pred %p2, %p21, %p22;
|
|
mov.b32 %r64, %f1567;
|
|
xor.b32 %r65, %r64, -2147483648;
|
|
mov.b32 %f488, %r65;
|
|
selp.f32 %f1569, %f488, %f1567, %p2;
|
|
setp.eq.f32 %p23, %f70, 0f00000000;
|
|
@%p23 bra BB0_10;
|
|
bra.uni BB0_8;
|
|
|
|
BB0_10:
|
|
add.f32 %f491, %f70, %f70;
|
|
selp.f32 %f1569, %f491, 0f00000000, %p22;
|
|
bra.uni BB0_11;
|
|
|
|
BB0_8:
|
|
setp.geu.f32 %p24, %f70, 0f00000000;
|
|
@%p24 bra BB0_11;
|
|
|
|
cvt.rzi.f32.f32 %f490, %f386;
|
|
setp.neu.f32 %p25, %f490, 0f40800000;
|
|
selp.f32 %f1569, 0f7FFFFFFF, %f1569, %p25;
|
|
|
|
BB0_11:
|
|
add.f32 %f492, %f71, 0f40800000;
|
|
mov.b32 %r66, %f492;
|
|
setp.lt.s32 %p27, %r66, 2139095040;
|
|
@%p27 bra BB0_16;
|
|
|
|
setp.gtu.f32 %p28, %f71, 0f7F800000;
|
|
@%p28 bra BB0_15;
|
|
bra.uni BB0_13;
|
|
|
|
BB0_15:
|
|
add.f32 %f1569, %f70, 0f40800000;
|
|
bra.uni BB0_16;
|
|
|
|
BB0_13:
|
|
setp.neu.f32 %p29, %f71, 0f7F800000;
|
|
@%p29 bra BB0_16;
|
|
|
|
selp.f32 %f1569, 0fFF800000, 0f7F800000, %p2;
|
|
|
|
BB0_16:
|
|
mul.f32 %f493, %f66, %f407;
|
|
mov.f32 %f1608, 0f3F800000;
|
|
sub.f32 %f495, %f1608, %f1569;
|
|
setp.eq.f32 %p30, %f70, 0f3F800000;
|
|
selp.f32 %f496, 0f00000000, %f495, %p30;
|
|
cvt.sat.f32.f32 %f497, %f496;
|
|
fma.rn.f32 %f498, %f493, %f493, %f408;
|
|
div.rn.f32 %f100, %f497, %f498;
|
|
mul.f32 %f499, %f1549, %f68;
|
|
fma.rn.f32 %f500, %f1548, %f67, %f499;
|
|
fma.rn.f32 %f501, %f1550, %f69, %f500;
|
|
ld.global.u32 %r67, [ignoreNormal];
|
|
setp.eq.s32 %p31, %r67, 0;
|
|
selp.f32 %f101, %f501, 0f3F800000, %p31;
|
|
mul.f32 %f502, %f101, 0f40800000;
|
|
cvt.sat.f32.f32 %f102, %f502;
|
|
setp.eq.f32 %p32, %f410, 0f3F800000;
|
|
@%p32 bra BB0_23;
|
|
bra.uni BB0_17;
|
|
|
|
BB0_23:
|
|
setp.leu.f32 %p36, %f405, 0f00000000;
|
|
@%p36 bra BB0_19;
|
|
|
|
setp.geu.f32 %p37, %f401, 0f00000000;
|
|
@%p37 bra BB0_25;
|
|
|
|
sub.f32 %f1547, %f413, %f9;
|
|
sub.f32 %f1546, %f411, %f7;
|
|
sub.f32 %f1545, %f412, %f8;
|
|
mul.f32 %f525, %f391, %f1545;
|
|
fma.rn.f32 %f526, %f390, %f1546, %f525;
|
|
fma.rn.f32 %f527, %f392, %f1547, %f526;
|
|
rcp.rn.f32 %f528, %f527;
|
|
mul.f32 %f1570, %f1546, %f528;
|
|
mul.f32 %f1571, %f1545, %f528;
|
|
mul.f32 %f1572, %f1547, %f528;
|
|
neg.f32 %f401, %f401;
|
|
bra.uni BB0_27;
|
|
|
|
BB0_17:
|
|
setp.eq.f32 %p33, %f410, 0f40000000;
|
|
@%p33 bra BB0_21;
|
|
bra.uni BB0_18;
|
|
|
|
BB0_21:
|
|
setp.leu.f32 %p35, %f405, 0f00000000;
|
|
@%p35 bra BB0_19;
|
|
|
|
mul.f32 %f519, %f399, %f68;
|
|
fma.rn.f32 %f520, %f398, %f67, %f519;
|
|
mul.f32 %f521, %f395, %f68;
|
|
fma.rn.f32 %f522, %f394, %f67, %f521;
|
|
mul.f32 %f523, %f391, %f68;
|
|
fma.rn.f32 %f524, %f390, %f67, %f523;
|
|
fma.rn.f32 %f516, %f400, %f69, %f520;
|
|
fma.rn.f32 %f517, %f396, %f69, %f522;
|
|
fma.rn.f32 %f518, %f392, %f69, %f524;
|
|
cvt.rzi.s32.f32 %r68, %f405;
|
|
mov.u32 %r69, 6;
|
|
mov.u32 %r70, 0;
|
|
// inline asm
|
|
call (%f512, %f513, %f514, %f515), _rt_texture_get_base_id, (%r68, %r69, %f516, %f517, %f518, %r70);
|
|
// inline asm
|
|
mul.f32 %f1574, %f402, %f512;
|
|
mul.f32 %f1575, %f1575, %f513;
|
|
mul.f32 %f1576, %f1576, %f514;
|
|
bra.uni BB0_28;
|
|
|
|
BB0_18:
|
|
setp.neu.f32 %p34, %f410, 0f40800000;
|
|
@%p34 bra BB0_19;
|
|
|
|
mul.f32 %f503, %f390, %f67;
|
|
mul.f32 %f504, %f391, %f68;
|
|
neg.f32 %f505, %f504;
|
|
sub.f32 %f506, %f505, %f503;
|
|
mul.f32 %f507, %f392, %f69;
|
|
sub.f32 %f508, %f506, %f507;
|
|
fma.rn.f32 %f509, %f405, %f508, %f401;
|
|
cvt.sat.f32.f32 %f510, %f509;
|
|
mul.f32 %f511, %f510, %f510;
|
|
mul.f32 %f1577, %f100, %f511;
|
|
mov.f32 %f1574, %f402;
|
|
bra.uni BB0_29;
|
|
|
|
BB0_19:
|
|
mov.f32 %f1574, %f402;
|
|
bra.uni BB0_28;
|
|
|
|
BB0_25:
|
|
mov.f32 %f1570, %f67;
|
|
mov.f32 %f1571, %f68;
|
|
mov.f32 %f1572, %f69;
|
|
|
|
BB0_27:
|
|
mul.f32 %f537, %f392, %f69;
|
|
mul.f32 %f538, %f391, %f68;
|
|
neg.f32 %f539, %f538;
|
|
mul.f32 %f540, %f390, %f67;
|
|
sub.f32 %f541, %f539, %f540;
|
|
sub.f32 %f542, %f541, %f537;
|
|
setp.gt.f32 %p38, %f542, 0f00000000;
|
|
selp.f32 %f543, 0f3F800000, 0f00000000, %p38;
|
|
mul.f32 %f544, %f399, %f1571;
|
|
fma.rn.f32 %f545, %f398, %f1570, %f544;
|
|
mul.f32 %f546, %f395, %f1571;
|
|
fma.rn.f32 %f547, %f394, %f1570, %f546;
|
|
fma.rn.f32 %f548, %f400, %f1572, %f545;
|
|
fma.rn.f32 %f549, %f396, %f1572, %f547;
|
|
fma.rn.f32 %f550, %f401, %f548, 0f3F000000;
|
|
mov.f32 %f551, 0f3F800000;
|
|
sub.f32 %f533, %f551, %f550;
|
|
fma.rn.f32 %f534, %f401, %f549, 0f3F000000;
|
|
cvt.rzi.s32.f32 %r71, %f405;
|
|
mov.f32 %f536, 0f00000000;
|
|
// inline asm
|
|
call (%f529, %f530, %f531, %f532), _rt_texture_get_f_id, (%r71, %r28, %f533, %f534, %f536, %f536);
|
|
// inline asm
|
|
mul.f32 %f552, %f543, %f529;
|
|
mul.f32 %f553, %f543, %f530;
|
|
mul.f32 %f554, %f543, %f531;
|
|
mul.f32 %f1574, %f402, %f552;
|
|
mul.f32 %f1575, %f1575, %f553;
|
|
mul.f32 %f1576, %f1576, %f554;
|
|
|
|
BB0_28:
|
|
mov.f32 %f1577, %f100;
|
|
|
|
BB0_29:
|
|
max.f32 %f570, %f1574, %f1575;
|
|
max.f32 %f571, %f570, %f1576;
|
|
mul.f32 %f572, %f102, %f1577;
|
|
mul.f32 %f573, %f572, %f571;
|
|
setp.lt.f32 %p40, %f573, 0f3727C5AC;
|
|
mov.pred %p165, -1;
|
|
mov.f32 %f1578, 0f00000000;
|
|
mov.f32 %f1579, %f1578;
|
|
mov.f32 %f1580, %f1578;
|
|
mov.f32 %f140, %f1578;
|
|
mov.f32 %f1582, %f1578;
|
|
mov.f32 %f1583, %f1578;
|
|
mov.f32 %f143, %f1578;
|
|
mov.f32 %f1585, %f1578;
|
|
mov.f32 %f1586, %f1578;
|
|
mov.f32 %f146, %f1578;
|
|
mov.f32 %f1588, %f1578;
|
|
mov.f32 %f1589, %f1578;
|
|
mov.f32 %f149, %f1578;
|
|
mov.f32 %f1591, %f1578;
|
|
mov.f32 %f1592, %f1578;
|
|
@%p40 bra BB0_31;
|
|
|
|
cvt.sat.f32.f32 %f574, %f101;
|
|
mul.f32 %f575, %f1577, %f574;
|
|
mul.f32 %f1578, %f1574, %f575;
|
|
mul.f32 %f1579, %f1575, %f575;
|
|
mul.f32 %f1580, %f1576, %f575;
|
|
mul.f32 %f576, %f1577, 0f3E800000;
|
|
mul.f32 %f577, %f102, %f576;
|
|
mul.f32 %f140, %f1574, %f577;
|
|
mul.f32 %f1582, %f1575, %f577;
|
|
mul.f32 %f1583, %f1576, %f577;
|
|
mul.f32 %f143, %f67, %f140;
|
|
mul.f32 %f1585, %f67, %f1582;
|
|
mul.f32 %f1586, %f67, %f1583;
|
|
mul.f32 %f146, %f68, %f140;
|
|
mul.f32 %f1588, %f68, %f1582;
|
|
mul.f32 %f1589, %f68, %f1583;
|
|
mul.f32 %f149, %f69, %f140;
|
|
mul.f32 %f1591, %f69, %f1582;
|
|
mul.f32 %f1592, %f69, %f1583;
|
|
mov.pred %p165, 0;
|
|
|
|
BB0_31:
|
|
@%p165 bra BB0_59;
|
|
|
|
setp.eq.s32 %p42, %r6, 0;
|
|
mov.u16 %rs161, 0;
|
|
mov.f32 %f1609, %f1608;
|
|
mov.f32 %f1610, %f1608;
|
|
@%p42 bra BB0_43;
|
|
|
|
abs.s32 %r8, %r6;
|
|
mov.f32 %f1605, 0f00000000;
|
|
setp.lt.s32 %p43, %r8, 1;
|
|
mov.f32 %f1606, %f1605;
|
|
mov.f32 %f1607, %f1605;
|
|
@%p43 bra BB0_42;
|
|
|
|
max.f32 %f153, %f591, %f363;
|
|
and.b32 %r9, %r8, 3;
|
|
setp.eq.s32 %p44, %r9, 0;
|
|
add.u64 %rd35, %SP, 0;
|
|
cvta.to.local.u64 %rd2, %rd35;
|
|
mov.f32 %f1605, 0f00000000;
|
|
mov.u32 %r269, 0;
|
|
mov.f32 %f1606, %f1605;
|
|
mov.f32 %f1607, %f1605;
|
|
@%p44 bra BB0_40;
|
|
|
|
setp.eq.s32 %p45, %r9, 1;
|
|
mov.f32 %f1596, 0f00000000;
|
|
mov.u32 %r267, 0;
|
|
mov.f32 %f1597, %f1596;
|
|
mov.f32 %f1598, %f1596;
|
|
@%p45 bra BB0_39;
|
|
|
|
setp.eq.s32 %p46, %r9, 2;
|
|
mov.f32 %f1593, 0f00000000;
|
|
mov.u32 %r266, 0;
|
|
mov.f32 %f1594, %f1593;
|
|
mov.f32 %f1595, %f1593;
|
|
@%p46 bra BB0_38;
|
|
|
|
cvt.rmi.f32.f32 %f607, %f19;
|
|
sub.f32 %f608, %f19, %f607;
|
|
cvt.rmi.f32.f32 %f609, %f20;
|
|
sub.f32 %f610, %f20, %f609;
|
|
add.f32 %f611, %f610, 0f420551EC;
|
|
add.f32 %f612, %f608, 0f420551EC;
|
|
mul.f32 %f613, %f608, %f611;
|
|
fma.rn.f32 %f614, %f610, %f612, %f613;
|
|
add.f32 %f615, %f614, 0f00000000;
|
|
add.f32 %f616, %f608, %f615;
|
|
add.f32 %f617, %f610, %f615;
|
|
add.f32 %f618, %f616, %f617;
|
|
mul.f32 %f619, %f615, %f618;
|
|
cvt.rmi.f32.f32 %f620, %f619;
|
|
sub.f32 %f621, %f619, %f620;
|
|
add.f32 %f622, %f616, %f616;
|
|
mul.f32 %f623, %f617, %f622;
|
|
cvt.rmi.f32.f32 %f624, %f623;
|
|
sub.f32 %f625, %f623, %f624;
|
|
mul.f32 %f626, %f616, %f618;
|
|
cvt.rmi.f32.f32 %f627, %f626;
|
|
sub.f32 %f628, %f626, %f627;
|
|
fma.rn.f32 %f629, %f621, 0f40000000, 0fBF800000;
|
|
fma.rn.f32 %f630, %f625, 0f40000000, 0fBF800000;
|
|
fma.rn.f32 %f631, %f628, 0f40000000, 0fBF800000;
|
|
fma.rn.f32 %f632, %f406, %f629, %f411;
|
|
fma.rn.f32 %f633, %f406, %f630, %f412;
|
|
fma.rn.f32 %f634, %f406, %f631, %f413;
|
|
sub.f32 %f635, %f632, %f7;
|
|
sub.f32 %f636, %f633, %f8;
|
|
sub.f32 %f637, %f634, %f9;
|
|
mul.f32 %f638, %f636, %f636;
|
|
fma.rn.f32 %f639, %f635, %f635, %f638;
|
|
fma.rn.f32 %f640, %f637, %f637, %f639;
|
|
sqrt.rn.f32 %f606, %f640;
|
|
rcp.rn.f32 %f641, %f606;
|
|
mul.f32 %f602, %f641, %f635;
|
|
mul.f32 %f603, %f641, %f636;
|
|
mul.f32 %f604, %f641, %f637;
|
|
ld.global.u32 %r80, [imageEnabled];
|
|
and.b32 %r81, %r80, 32;
|
|
setp.eq.s32 %p47, %r81, 0;
|
|
selp.f32 %f642, 0f3F800000, 0f41200000, %p47;
|
|
mul.f32 %f605, %f642, %f153;
|
|
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, %f602, %f603, %f604, %r45, %f605, %f606, %rd35, %r44);
|
|
// inline asm
|
|
ld.local.f32 %f643, [%rd2];
|
|
add.f32 %f1595, %f643, 0f00000000;
|
|
ld.local.f32 %f644, [%rd2+4];
|
|
add.f32 %f1594, %f644, 0f00000000;
|
|
ld.local.f32 %f645, [%rd2+8];
|
|
add.f32 %f1593, %f645, 0f00000000;
|
|
mov.u32 %r266, %r45;
|
|
|
|
BB0_38:
|
|
cvt.rn.f32.s32 %f654, %r266;
|
|
add.f32 %f655, %f13, %f654;
|
|
sub.f32 %f656, %f14, %f654;
|
|
mul.f32 %f657, %f655, 0f3DD32618;
|
|
cvt.rmi.f32.f32 %f658, %f657;
|
|
sub.f32 %f659, %f657, %f658;
|
|
mul.f32 %f660, %f656, 0f3DD2F1AA;
|
|
cvt.rmi.f32.f32 %f661, %f660;
|
|
sub.f32 %f662, %f660, %f661;
|
|
mul.f32 %f663, %f654, 0f3DC74539;
|
|
cvt.rmi.f32.f32 %f664, %f663;
|
|
sub.f32 %f665, %f663, %f664;
|
|
add.f32 %f666, %f662, 0f420551EC;
|
|
add.f32 %f667, %f659, 0f420551EC;
|
|
add.f32 %f668, %f665, 0f420551EC;
|
|
mul.f32 %f669, %f659, %f666;
|
|
fma.rn.f32 %f670, %f662, %f667, %f669;
|
|
fma.rn.f32 %f671, %f665, %f668, %f670;
|
|
add.f32 %f672, %f659, %f671;
|
|
add.f32 %f673, %f662, %f671;
|
|
add.f32 %f674, %f665, %f671;
|
|
add.f32 %f675, %f672, %f673;
|
|
mul.f32 %f676, %f674, %f675;
|
|
cvt.rmi.f32.f32 %f677, %f676;
|
|
sub.f32 %f678, %f676, %f677;
|
|
add.f32 %f679, %f672, %f672;
|
|
mul.f32 %f680, %f673, %f679;
|
|
cvt.rmi.f32.f32 %f681, %f680;
|
|
sub.f32 %f682, %f680, %f681;
|
|
mul.f32 %f683, %f672, %f675;
|
|
cvt.rmi.f32.f32 %f684, %f683;
|
|
sub.f32 %f685, %f683, %f684;
|
|
fma.rn.f32 %f686, %f678, 0f40000000, 0fBF800000;
|
|
fma.rn.f32 %f687, %f682, 0f40000000, 0fBF800000;
|
|
fma.rn.f32 %f688, %f685, 0f40000000, 0fBF800000;
|
|
fma.rn.f32 %f689, %f406, %f686, %f411;
|
|
fma.rn.f32 %f690, %f406, %f687, %f412;
|
|
fma.rn.f32 %f691, %f406, %f688, %f413;
|
|
sub.f32 %f692, %f689, %f7;
|
|
sub.f32 %f693, %f690, %f8;
|
|
sub.f32 %f694, %f691, %f9;
|
|
mul.f32 %f695, %f693, %f693;
|
|
fma.rn.f32 %f696, %f692, %f692, %f695;
|
|
fma.rn.f32 %f697, %f694, %f694, %f696;
|
|
sqrt.rn.f32 %f653, %f697;
|
|
rcp.rn.f32 %f698, %f653;
|
|
mul.f32 %f649, %f698, %f692;
|
|
mul.f32 %f650, %f698, %f693;
|
|
mul.f32 %f651, %f698, %f694;
|
|
ld.global.u32 %r86, [imageEnabled];
|
|
and.b32 %r87, %r86, 32;
|
|
setp.eq.s32 %p48, %r87, 0;
|
|
selp.f32 %f699, 0f3F800000, 0f41200000, %p48;
|
|
mul.f32 %f652, %f699, %f153;
|
|
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, %f649, %f650, %f651, %r45, %f652, %f653, %rd35, %r44);
|
|
// inline asm
|
|
ld.local.f32 %f700, [%rd2];
|
|
add.f32 %f1598, %f1595, %f700;
|
|
ld.local.f32 %f701, [%rd2+4];
|
|
add.f32 %f1597, %f1594, %f701;
|
|
ld.local.f32 %f702, [%rd2+8];
|
|
add.f32 %f1596, %f1593, %f702;
|
|
add.s32 %r267, %r266, 1;
|
|
|
|
BB0_39:
|
|
cvt.rn.f32.s32 %f711, %r267;
|
|
add.f32 %f712, %f13, %f711;
|
|
sub.f32 %f713, %f14, %f711;
|
|
mul.f32 %f714, %f712, 0f3DD32618;
|
|
cvt.rmi.f32.f32 %f715, %f714;
|
|
sub.f32 %f716, %f714, %f715;
|
|
mul.f32 %f717, %f713, 0f3DD2F1AA;
|
|
cvt.rmi.f32.f32 %f718, %f717;
|
|
sub.f32 %f719, %f717, %f718;
|
|
mul.f32 %f720, %f711, 0f3DC74539;
|
|
cvt.rmi.f32.f32 %f721, %f720;
|
|
sub.f32 %f722, %f720, %f721;
|
|
add.f32 %f723, %f719, 0f420551EC;
|
|
add.f32 %f724, %f716, 0f420551EC;
|
|
add.f32 %f725, %f722, 0f420551EC;
|
|
mul.f32 %f726, %f716, %f723;
|
|
fma.rn.f32 %f727, %f719, %f724, %f726;
|
|
fma.rn.f32 %f728, %f722, %f725, %f727;
|
|
add.f32 %f729, %f716, %f728;
|
|
add.f32 %f730, %f719, %f728;
|
|
add.f32 %f731, %f722, %f728;
|
|
add.f32 %f732, %f729, %f730;
|
|
mul.f32 %f733, %f731, %f732;
|
|
cvt.rmi.f32.f32 %f734, %f733;
|
|
sub.f32 %f735, %f733, %f734;
|
|
add.f32 %f736, %f729, %f729;
|
|
mul.f32 %f737, %f730, %f736;
|
|
cvt.rmi.f32.f32 %f738, %f737;
|
|
sub.f32 %f739, %f737, %f738;
|
|
mul.f32 %f740, %f729, %f732;
|
|
cvt.rmi.f32.f32 %f741, %f740;
|
|
sub.f32 %f742, %f740, %f741;
|
|
fma.rn.f32 %f743, %f735, 0f40000000, 0fBF800000;
|
|
fma.rn.f32 %f744, %f739, 0f40000000, 0fBF800000;
|
|
fma.rn.f32 %f745, %f742, 0f40000000, 0fBF800000;
|
|
fma.rn.f32 %f746, %f406, %f743, %f411;
|
|
fma.rn.f32 %f747, %f406, %f744, %f412;
|
|
fma.rn.f32 %f748, %f406, %f745, %f413;
|
|
sub.f32 %f749, %f746, %f7;
|
|
sub.f32 %f750, %f747, %f8;
|
|
sub.f32 %f751, %f748, %f9;
|
|
mul.f32 %f752, %f750, %f750;
|
|
fma.rn.f32 %f753, %f749, %f749, %f752;
|
|
fma.rn.f32 %f754, %f751, %f751, %f753;
|
|
sqrt.rn.f32 %f710, %f754;
|
|
rcp.rn.f32 %f755, %f710;
|
|
mul.f32 %f706, %f755, %f749;
|
|
mul.f32 %f707, %f755, %f750;
|
|
mul.f32 %f708, %f755, %f751;
|
|
ld.global.u32 %r92, [imageEnabled];
|
|
and.b32 %r93, %r92, 32;
|
|
setp.eq.s32 %p49, %r93, 0;
|
|
selp.f32 %f756, 0f3F800000, 0f41200000, %p49;
|
|
mul.f32 %f709, %f756, %f153;
|
|
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, %f706, %f707, %f708, %r90, %f709, %f710, %rd35, %r44);
|
|
// inline asm
|
|
ld.local.f32 %f757, [%rd2];
|
|
add.f32 %f1607, %f1598, %f757;
|
|
ld.local.f32 %f758, [%rd2+4];
|
|
add.f32 %f1606, %f1597, %f758;
|
|
ld.local.f32 %f759, [%rd2+8];
|
|
add.f32 %f1605, %f1596, %f759;
|
|
add.s32 %r269, %r267, 1;
|
|
|
|
BB0_40:
|
|
setp.lt.u32 %p50, %r8, 4;
|
|
@%p50 bra BB0_42;
|
|
|
|
BB0_41:
|
|
cvt.rn.f32.s32 %f792, %r269;
|
|
add.f32 %f793, %f13, %f792;
|
|
sub.f32 %f794, %f14, %f792;
|
|
mul.f32 %f795, %f793, 0f3DD32618;
|
|
cvt.rmi.f32.f32 %f796, %f795;
|
|
sub.f32 %f797, %f795, %f796;
|
|
mul.f32 %f798, %f794, 0f3DD2F1AA;
|
|
cvt.rmi.f32.f32 %f799, %f798;
|
|
sub.f32 %f800, %f798, %f799;
|
|
mul.f32 %f801, %f792, 0f3DC74539;
|
|
cvt.rmi.f32.f32 %f802, %f801;
|
|
sub.f32 %f803, %f801, %f802;
|
|
add.f32 %f804, %f800, 0f420551EC;
|
|
add.f32 %f805, %f797, 0f420551EC;
|
|
add.f32 %f806, %f803, 0f420551EC;
|
|
mul.f32 %f807, %f797, %f804;
|
|
fma.rn.f32 %f808, %f800, %f805, %f807;
|
|
fma.rn.f32 %f809, %f803, %f806, %f808;
|
|
add.f32 %f810, %f797, %f809;
|
|
add.f32 %f811, %f800, %f809;
|
|
add.f32 %f812, %f803, %f809;
|
|
add.f32 %f813, %f810, %f811;
|
|
mul.f32 %f814, %f812, %f813;
|
|
cvt.rmi.f32.f32 %f815, %f814;
|
|
sub.f32 %f816, %f814, %f815;
|
|
add.f32 %f817, %f810, %f810;
|
|
mul.f32 %f818, %f811, %f817;
|
|
cvt.rmi.f32.f32 %f819, %f818;
|
|
sub.f32 %f820, %f818, %f819;
|
|
mul.f32 %f821, %f810, %f813;
|
|
cvt.rmi.f32.f32 %f822, %f821;
|
|
sub.f32 %f823, %f821, %f822;
|
|
fma.rn.f32 %f824, %f816, 0f40000000, 0fBF800000;
|
|
fma.rn.f32 %f825, %f820, 0f40000000, 0fBF800000;
|
|
fma.rn.f32 %f826, %f823, 0f40000000, 0fBF800000;
|
|
fma.rn.f32 %f827, %f406, %f824, %f411;
|
|
fma.rn.f32 %f828, %f406, %f825, %f412;
|
|
fma.rn.f32 %f829, %f406, %f826, %f413;
|
|
sub.f32 %f830, %f827, %f7;
|
|
sub.f32 %f831, %f828, %f8;
|
|
sub.f32 %f832, %f829, %f9;
|
|
mul.f32 %f833, %f831, %f831;
|
|
fma.rn.f32 %f834, %f830, %f830, %f833;
|
|
fma.rn.f32 %f835, %f832, %f832, %f834;
|
|
sqrt.rn.f32 %f767, %f835;
|
|
rcp.rn.f32 %f836, %f767;
|
|
mul.f32 %f763, %f836, %f830;
|
|
mul.f32 %f764, %f836, %f831;
|
|
mul.f32 %f765, %f836, %f832;
|
|
ld.global.u32 %r107, [imageEnabled];
|
|
and.b32 %r108, %r107, 32;
|
|
setp.eq.s32 %p51, %r108, 0;
|
|
selp.f32 %f837, 0f3F800000, 0f41200000, %p51;
|
|
mul.f32 %f766, %f837, %f153;
|
|
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, %f763, %f764, %f765, %r105, %f766, %f767, %rd35, %r44);
|
|
// inline asm
|
|
ld.local.f32 %f838, [%rd2];
|
|
add.f32 %f839, %f1607, %f838;
|
|
ld.local.f32 %f840, [%rd2+4];
|
|
add.f32 %f841, %f1606, %f840;
|
|
ld.local.f32 %f842, [%rd2+8];
|
|
add.f32 %f843, %f1605, %f842;
|
|
add.s32 %r110, %r269, 1;
|
|
cvt.rn.f32.s32 %f844, %r110;
|
|
add.f32 %f845, %f13, %f844;
|
|
sub.f32 %f846, %f14, %f844;
|
|
mul.f32 %f847, %f845, 0f3DD32618;
|
|
cvt.rmi.f32.f32 %f848, %f847;
|
|
sub.f32 %f849, %f847, %f848;
|
|
mul.f32 %f850, %f846, 0f3DD2F1AA;
|
|
cvt.rmi.f32.f32 %f851, %f850;
|
|
sub.f32 %f852, %f850, %f851;
|
|
mul.f32 %f853, %f844, 0f3DC74539;
|
|
cvt.rmi.f32.f32 %f854, %f853;
|
|
sub.f32 %f855, %f853, %f854;
|
|
add.f32 %f856, %f852, 0f420551EC;
|
|
add.f32 %f857, %f849, 0f420551EC;
|
|
add.f32 %f858, %f855, 0f420551EC;
|
|
mul.f32 %f859, %f849, %f856;
|
|
fma.rn.f32 %f860, %f852, %f857, %f859;
|
|
fma.rn.f32 %f861, %f855, %f858, %f860;
|
|
add.f32 %f862, %f849, %f861;
|
|
add.f32 %f863, %f852, %f861;
|
|
add.f32 %f864, %f855, %f861;
|
|
add.f32 %f865, %f862, %f863;
|
|
mul.f32 %f866, %f864, %f865;
|
|
cvt.rmi.f32.f32 %f867, %f866;
|
|
sub.f32 %f868, %f866, %f867;
|
|
add.f32 %f869, %f862, %f862;
|
|
mul.f32 %f870, %f863, %f869;
|
|
cvt.rmi.f32.f32 %f871, %f870;
|
|
sub.f32 %f872, %f870, %f871;
|
|
mul.f32 %f873, %f862, %f865;
|
|
cvt.rmi.f32.f32 %f874, %f873;
|
|
sub.f32 %f875, %f873, %f874;
|
|
fma.rn.f32 %f876, %f868, 0f40000000, 0fBF800000;
|
|
fma.rn.f32 %f877, %f872, 0f40000000, 0fBF800000;
|
|
fma.rn.f32 %f878, %f875, 0f40000000, 0fBF800000;
|
|
fma.rn.f32 %f879, %f406, %f876, %f411;
|
|
fma.rn.f32 %f880, %f406, %f877, %f412;
|
|
fma.rn.f32 %f881, %f406, %f878, %f413;
|
|
sub.f32 %f882, %f879, %f7;
|
|
sub.f32 %f883, %f880, %f8;
|
|
sub.f32 %f884, %f881, %f9;
|
|
mul.f32 %f885, %f883, %f883;
|
|
fma.rn.f32 %f886, %f882, %f882, %f885;
|
|
fma.rn.f32 %f887, %f884, %f884, %f886;
|
|
sqrt.rn.f32 %f775, %f887;
|
|
rcp.rn.f32 %f888, %f775;
|
|
mul.f32 %f771, %f888, %f882;
|
|
mul.f32 %f772, %f888, %f883;
|
|
mul.f32 %f773, %f888, %f884;
|
|
ld.global.u32 %r111, [imageEnabled];
|
|
and.b32 %r112, %r111, 32;
|
|
setp.eq.s32 %p52, %r112, 0;
|
|
selp.f32 %f889, 0f3F800000, 0f41200000, %p52;
|
|
mul.f32 %f774, %f889, %f153;
|
|
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, %f771, %f772, %f773, %r105, %f774, %f775, %rd35, %r44);
|
|
// inline asm
|
|
ld.local.f32 %f890, [%rd2];
|
|
add.f32 %f891, %f839, %f890;
|
|
ld.local.f32 %f892, [%rd2+4];
|
|
add.f32 %f893, %f841, %f892;
|
|
ld.local.f32 %f894, [%rd2+8];
|
|
add.f32 %f895, %f843, %f894;
|
|
add.s32 %r113, %r269, 2;
|
|
cvt.rn.f32.s32 %f896, %r113;
|
|
add.f32 %f897, %f13, %f896;
|
|
sub.f32 %f898, %f14, %f896;
|
|
mul.f32 %f899, %f897, 0f3DD32618;
|
|
cvt.rmi.f32.f32 %f900, %f899;
|
|
sub.f32 %f901, %f899, %f900;
|
|
mul.f32 %f902, %f898, 0f3DD2F1AA;
|
|
cvt.rmi.f32.f32 %f903, %f902;
|
|
sub.f32 %f904, %f902, %f903;
|
|
mul.f32 %f905, %f896, 0f3DC74539;
|
|
cvt.rmi.f32.f32 %f906, %f905;
|
|
sub.f32 %f907, %f905, %f906;
|
|
add.f32 %f908, %f904, 0f420551EC;
|
|
add.f32 %f909, %f901, 0f420551EC;
|
|
add.f32 %f910, %f907, 0f420551EC;
|
|
mul.f32 %f911, %f901, %f908;
|
|
fma.rn.f32 %f912, %f904, %f909, %f911;
|
|
fma.rn.f32 %f913, %f907, %f910, %f912;
|
|
add.f32 %f914, %f901, %f913;
|
|
add.f32 %f915, %f904, %f913;
|
|
add.f32 %f916, %f907, %f913;
|
|
add.f32 %f917, %f914, %f915;
|
|
mul.f32 %f918, %f916, %f917;
|
|
cvt.rmi.f32.f32 %f919, %f918;
|
|
sub.f32 %f920, %f918, %f919;
|
|
add.f32 %f921, %f914, %f914;
|
|
mul.f32 %f922, %f915, %f921;
|
|
cvt.rmi.f32.f32 %f923, %f922;
|
|
sub.f32 %f924, %f922, %f923;
|
|
mul.f32 %f925, %f914, %f917;
|
|
cvt.rmi.f32.f32 %f926, %f925;
|
|
sub.f32 %f927, %f925, %f926;
|
|
fma.rn.f32 %f928, %f920, 0f40000000, 0fBF800000;
|
|
fma.rn.f32 %f929, %f924, 0f40000000, 0fBF800000;
|
|
fma.rn.f32 %f930, %f927, 0f40000000, 0fBF800000;
|
|
fma.rn.f32 %f931, %f406, %f928, %f411;
|
|
fma.rn.f32 %f932, %f406, %f929, %f412;
|
|
fma.rn.f32 %f933, %f406, %f930, %f413;
|
|
sub.f32 %f934, %f931, %f7;
|
|
sub.f32 %f935, %f932, %f8;
|
|
sub.f32 %f936, %f933, %f9;
|
|
mul.f32 %f937, %f935, %f935;
|
|
fma.rn.f32 %f938, %f934, %f934, %f937;
|
|
fma.rn.f32 %f939, %f936, %f936, %f938;
|
|
sqrt.rn.f32 %f783, %f939;
|
|
rcp.rn.f32 %f940, %f783;
|
|
mul.f32 %f779, %f940, %f934;
|
|
mul.f32 %f780, %f940, %f935;
|
|
mul.f32 %f781, %f940, %f936;
|
|
ld.global.u32 %r114, [imageEnabled];
|
|
and.b32 %r115, %r114, 32;
|
|
setp.eq.s32 %p53, %r115, 0;
|
|
selp.f32 %f941, 0f3F800000, 0f41200000, %p53;
|
|
mul.f32 %f782, %f941, %f153;
|
|
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, %f779, %f780, %f781, %r105, %f782, %f783, %rd35, %r44);
|
|
// inline asm
|
|
ld.local.f32 %f942, [%rd2];
|
|
add.f32 %f943, %f891, %f942;
|
|
ld.local.f32 %f944, [%rd2+4];
|
|
add.f32 %f945, %f893, %f944;
|
|
ld.local.f32 %f946, [%rd2+8];
|
|
add.f32 %f947, %f895, %f946;
|
|
add.s32 %r116, %r269, 3;
|
|
cvt.rn.f32.s32 %f948, %r116;
|
|
add.f32 %f949, %f13, %f948;
|
|
sub.f32 %f950, %f14, %f948;
|
|
mul.f32 %f951, %f949, 0f3DD32618;
|
|
cvt.rmi.f32.f32 %f952, %f951;
|
|
sub.f32 %f953, %f951, %f952;
|
|
mul.f32 %f954, %f950, 0f3DD2F1AA;
|
|
cvt.rmi.f32.f32 %f955, %f954;
|
|
sub.f32 %f956, %f954, %f955;
|
|
mul.f32 %f957, %f948, 0f3DC74539;
|
|
cvt.rmi.f32.f32 %f958, %f957;
|
|
sub.f32 %f959, %f957, %f958;
|
|
add.f32 %f960, %f956, 0f420551EC;
|
|
add.f32 %f961, %f953, 0f420551EC;
|
|
add.f32 %f962, %f959, 0f420551EC;
|
|
mul.f32 %f963, %f953, %f960;
|
|
fma.rn.f32 %f964, %f956, %f961, %f963;
|
|
fma.rn.f32 %f965, %f959, %f962, %f964;
|
|
add.f32 %f966, %f953, %f965;
|
|
add.f32 %f967, %f956, %f965;
|
|
add.f32 %f968, %f959, %f965;
|
|
add.f32 %f969, %f966, %f967;
|
|
mul.f32 %f970, %f968, %f969;
|
|
cvt.rmi.f32.f32 %f971, %f970;
|
|
sub.f32 %f972, %f970, %f971;
|
|
add.f32 %f973, %f966, %f966;
|
|
mul.f32 %f974, %f967, %f973;
|
|
cvt.rmi.f32.f32 %f975, %f974;
|
|
sub.f32 %f976, %f974, %f975;
|
|
mul.f32 %f977, %f966, %f969;
|
|
cvt.rmi.f32.f32 %f978, %f977;
|
|
sub.f32 %f979, %f977, %f978;
|
|
fma.rn.f32 %f980, %f972, 0f40000000, 0fBF800000;
|
|
fma.rn.f32 %f981, %f976, 0f40000000, 0fBF800000;
|
|
fma.rn.f32 %f982, %f979, 0f40000000, 0fBF800000;
|
|
fma.rn.f32 %f983, %f406, %f980, %f411;
|
|
fma.rn.f32 %f984, %f406, %f981, %f412;
|
|
fma.rn.f32 %f985, %f406, %f982, %f413;
|
|
sub.f32 %f986, %f983, %f7;
|
|
sub.f32 %f987, %f984, %f8;
|
|
sub.f32 %f988, %f985, %f9;
|
|
mul.f32 %f989, %f987, %f987;
|
|
fma.rn.f32 %f990, %f986, %f986, %f989;
|
|
fma.rn.f32 %f991, %f988, %f988, %f990;
|
|
sqrt.rn.f32 %f791, %f991;
|
|
rcp.rn.f32 %f992, %f791;
|
|
mul.f32 %f787, %f992, %f986;
|
|
mul.f32 %f788, %f992, %f987;
|
|
mul.f32 %f789, %f992, %f988;
|
|
ld.global.u32 %r117, [imageEnabled];
|
|
and.b32 %r118, %r117, 32;
|
|
setp.eq.s32 %p54, %r118, 0;
|
|
selp.f32 %f993, 0f3F800000, 0f41200000, %p54;
|
|
mul.f32 %f790, %f993, %f153;
|
|
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, %f787, %f788, %f789, %r105, %f790, %f791, %rd35, %r44);
|
|
// inline asm
|
|
ld.local.f32 %f994, [%rd2];
|
|
add.f32 %f1607, %f943, %f994;
|
|
ld.local.f32 %f995, [%rd2+4];
|
|
add.f32 %f1606, %f945, %f995;
|
|
ld.local.f32 %f996, [%rd2+8];
|
|
add.f32 %f1605, %f947, %f996;
|
|
add.s32 %r269, %r269, 4;
|
|
setp.lt.s32 %p55, %r269, %r8;
|
|
@%p55 bra BB0_41;
|
|
|
|
BB0_42:
|
|
cvt.rn.f32.s32 %f997, %r8;
|
|
rcp.rn.f32 %f998, %f997;
|
|
mul.f32 %f1608, %f1607, %f998;
|
|
mul.f32 %f1609, %f1606, %f998;
|
|
mul.f32 %f1610, %f1605, %f998;
|
|
shr.u32 %r119, %r6, 31;
|
|
cvt.u16.u32 %rs161, %r119;
|
|
|
|
BB0_43:
|
|
fma.rn.f32 %f1566, %f1578, %f1608, %f1566;
|
|
fma.rn.f32 %f1565, %f1579, %f1609, %f1565;
|
|
fma.rn.f32 %f1564, %f1580, %f1610, %f1564;
|
|
fma.rn.f32 %f1563, %f140, %f1608, %f1563;
|
|
fma.rn.f32 %f1562, %f1582, %f1609, %f1562;
|
|
fma.rn.f32 %f1561, %f1583, %f1610, %f1561;
|
|
fma.rn.f32 %f1560, %f143, %f1608, %f1560;
|
|
fma.rn.f32 %f1559, %f1585, %f1609, %f1559;
|
|
fma.rn.f32 %f1558, %f1586, %f1610, %f1558;
|
|
fma.rn.f32 %f1557, %f146, %f1608, %f1557;
|
|
fma.rn.f32 %f1556, %f1588, %f1609, %f1556;
|
|
fma.rn.f32 %f1555, %f1589, %f1610, %f1555;
|
|
fma.rn.f32 %f1554, %f149, %f1608, %f1554;
|
|
fma.rn.f32 %f1553, %f1591, %f1609, %f1553;
|
|
fma.rn.f32 %f1552, %f1592, %f1610, %f1552;
|
|
setp.eq.s16 %p56, %rs161, 0;
|
|
@%p56 bra BB0_58;
|
|
|
|
mov.f32 %f1544, 0fB5BFBE8E;
|
|
mov.f32 %f1543, 0fBF317200;
|
|
mov.f32 %f1542, 0f35BFBE8E;
|
|
mov.f32 %f1541, 0f3F317200;
|
|
mov.f32 %f1540, 0f3DAAAABD;
|
|
mov.f32 %f1539, 0f3C4CAF63;
|
|
mov.f32 %f1538, 0f3B18F0FE;
|
|
mul.f32 %f1001, %f143, 0f3F000000;
|
|
mul.f32 %f1002, %f146, 0f3F000000;
|
|
mul.f32 %f1003, %f1002, %f1002;
|
|
fma.rn.f32 %f1004, %f1001, %f1001, %f1003;
|
|
mul.f32 %f1005, %f149, 0f3F000000;
|
|
fma.rn.f32 %f1006, %f1005, %f1005, %f1004;
|
|
sqrt.rn.f32 %f1007, %f1006;
|
|
rcp.rn.f32 %f1008, %f1007;
|
|
mul.f32 %f1009, %f1001, %f1008;
|
|
mul.f32 %f1010, %f1002, %f1008;
|
|
mul.f32 %f1011, %f1005, %f1008;
|
|
mul.f32 %f1012, %f1549, %f1010;
|
|
fma.rn.f32 %f1013, %f1548, %f1009, %f1012;
|
|
fma.rn.f32 %f1014, %f1550, %f1011, %f1013;
|
|
fma.rn.f32 %f202, %f1014, 0f3F000000, 0f3F000000;
|
|
add.f32 %f1015, %f1007, %f1007;
|
|
div.rn.f32 %f1016, %f1015, %f140;
|
|
add.f32 %f203, %f1016, 0f3F800000;
|
|
div.rn.f32 %f1017, %f1007, %f140;
|
|
mov.f32 %f1018, 0f3F800000;
|
|
sub.f32 %f1019, %f1018, %f1017;
|
|
add.f32 %f1020, %f1017, 0f3F800000;
|
|
div.rn.f32 %f204, %f1019, %f1020;
|
|
sub.f32 %f1021, %f1018, %f204;
|
|
add.f32 %f1022, %f203, 0f3F800000;
|
|
mul.f32 %f205, %f1022, %f1021;
|
|
mul.f32 %f1023, %f203, 0f3F000000;
|
|
cvt.rzi.f32.f32 %f1024, %f1023;
|
|
add.f32 %f1025, %f1024, %f1024;
|
|
sub.f32 %f1026, %f203, %f1025;
|
|
abs.f32 %f206, %f1026;
|
|
abs.f32 %f207, %f202;
|
|
setp.lt.f32 %p57, %f207, 0f00800000;
|
|
mul.f32 %f1027, %f207, 0f4B800000;
|
|
selp.f32 %f1028, 0fC3170000, 0fC2FE0000, %p57;
|
|
selp.f32 %f1029, %f1027, %f207, %p57;
|
|
mov.b32 %r120, %f1029;
|
|
and.b32 %r121, %r120, 8388607;
|
|
or.b32 %r122, %r121, 1065353216;
|
|
mov.b32 %f1030, %r122;
|
|
shr.u32 %r123, %r120, 23;
|
|
cvt.rn.f32.u32 %f1031, %r123;
|
|
add.f32 %f1032, %f1028, %f1031;
|
|
setp.gt.f32 %p58, %f1030, 0f3FB504F3;
|
|
mul.f32 %f1033, %f1030, 0f3F000000;
|
|
add.f32 %f1034, %f1032, 0f3F800000;
|
|
selp.f32 %f1035, %f1033, %f1030, %p58;
|
|
selp.f32 %f1036, %f1034, %f1032, %p58;
|
|
add.f32 %f1037, %f1035, 0fBF800000;
|
|
add.f32 %f1000, %f1035, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f999,%f1000;
|
|
// inline asm
|
|
add.f32 %f1038, %f1037, %f1037;
|
|
mul.f32 %f1039, %f999, %f1038;
|
|
mul.f32 %f1040, %f1039, %f1039;
|
|
fma.rn.f32 %f1043, %f1538, %f1040, %f1539;
|
|
fma.rn.f32 %f1045, %f1043, %f1040, %f1540;
|
|
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, %f999, %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;
|
|
mul.rn.f32 %f1061, %f1036, %f1541;
|
|
mul.rn.f32 %f1063, %f1036, %f1542;
|
|
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;
|
|
abs.f32 %f208, %f203;
|
|
setp.gt.f32 %p59, %f208, 0f77F684DF;
|
|
mul.f32 %f1072, %f203, 0f39000000;
|
|
selp.f32 %f1073, %f1072, %f203, %p59;
|
|
mul.rn.f32 %f1074, %f1073, %f1069;
|
|
neg.f32 %f1075, %f1074;
|
|
fma.rn.f32 %f1076, %f1073, %f1069, %f1075;
|
|
fma.rn.f32 %f1077, %f1073, %f1071, %f1076;
|
|
mov.f32 %f1078, 0f00000000;
|
|
fma.rn.f32 %f1079, %f1078, %f1069, %f1077;
|
|
add.rn.f32 %f1080, %f1074, %f1079;
|
|
neg.f32 %f1081, %f1080;
|
|
add.rn.f32 %f1082, %f1074, %f1081;
|
|
add.rn.f32 %f1083, %f1082, %f1079;
|
|
mov.b32 %r124, %f1080;
|
|
setp.eq.s32 %p60, %r124, 1118925336;
|
|
add.s32 %r125, %r124, -1;
|
|
mov.b32 %f1084, %r125;
|
|
add.f32 %f1085, %f1083, 0f37000000;
|
|
selp.f32 %f1086, %f1084, %f1080, %p60;
|
|
selp.f32 %f209, %f1085, %f1083, %p60;
|
|
mul.f32 %f1087, %f1086, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f1088, %f1087;
|
|
fma.rn.f32 %f1090, %f1088, %f1543, %f1086;
|
|
fma.rn.f32 %f1092, %f1088, %f1544, %f1090;
|
|
mul.f32 %f1093, %f1092, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f1094, %f1093;
|
|
add.f32 %f1095, %f1088, 0f00000000;
|
|
ex2.approx.f32 %f1096, %f1095;
|
|
mul.f32 %f1097, %f1094, %f1096;
|
|
setp.lt.f32 %p61, %f1086, 0fC2D20000;
|
|
selp.f32 %f1098, 0f00000000, %f1097, %p61;
|
|
setp.gt.f32 %p62, %f1086, 0f42D20000;
|
|
selp.f32 %f1611, 0f7F800000, %f1098, %p62;
|
|
setp.eq.f32 %p63, %f1611, 0f7F800000;
|
|
@%p63 bra BB0_46;
|
|
|
|
fma.rn.f32 %f1611, %f1611, %f209, %f1611;
|
|
|
|
BB0_46:
|
|
setp.lt.f32 %p64, %f202, 0f00000000;
|
|
setp.eq.f32 %p65, %f206, 0f3F800000;
|
|
and.pred %p4, %p64, %p65;
|
|
mov.b32 %r126, %f1611;
|
|
xor.b32 %r127, %r126, -2147483648;
|
|
mov.b32 %f1099, %r127;
|
|
selp.f32 %f1613, %f1099, %f1611, %p4;
|
|
setp.eq.f32 %p66, %f202, 0f00000000;
|
|
@%p66 bra BB0_49;
|
|
bra.uni BB0_47;
|
|
|
|
BB0_49:
|
|
add.f32 %f1101, %f202, %f202;
|
|
mov.b32 %r128, %f1101;
|
|
selp.b32 %r129, %r128, 0, %p65;
|
|
or.b32 %r130, %r129, 2139095040;
|
|
setp.lt.f32 %p70, %f203, 0f00000000;
|
|
selp.b32 %r131, %r130, %r129, %p70;
|
|
mov.b32 %f1613, %r131;
|
|
bra.uni BB0_50;
|
|
|
|
BB0_47:
|
|
setp.geu.f32 %p67, %f202, 0f00000000;
|
|
@%p67 bra BB0_50;
|
|
|
|
cvt.rzi.f32.f32 %f1100, %f203;
|
|
setp.neu.f32 %p68, %f1100, %f203;
|
|
selp.f32 %f1613, 0f7FFFFFFF, %f1613, %p68;
|
|
|
|
BB0_50:
|
|
add.f32 %f1102, %f207, %f208;
|
|
mov.b32 %r132, %f1102;
|
|
setp.lt.s32 %p71, %r132, 2139095040;
|
|
@%p71 bra BB0_57;
|
|
|
|
setp.gtu.f32 %p72, %f207, 0f7F800000;
|
|
setp.gtu.f32 %p73, %f208, 0f7F800000;
|
|
or.pred %p74, %p72, %p73;
|
|
@%p74 bra BB0_56;
|
|
bra.uni BB0_52;
|
|
|
|
BB0_56:
|
|
add.f32 %f1613, %f203, %f202;
|
|
bra.uni BB0_57;
|
|
|
|
BB0_52:
|
|
setp.eq.f32 %p75, %f208, 0f7F800000;
|
|
@%p75 bra BB0_55;
|
|
bra.uni BB0_53;
|
|
|
|
BB0_55:
|
|
setp.gt.f32 %p78, %f207, 0f3F800000;
|
|
selp.b32 %r136, 2139095040, 0, %p78;
|
|
xor.b32 %r137, %r136, 2139095040;
|
|
setp.lt.f32 %p79, %f203, 0f00000000;
|
|
selp.b32 %r138, %r137, %r136, %p79;
|
|
mov.b32 %f1103, %r138;
|
|
setp.eq.f32 %p80, %f202, 0fBF800000;
|
|
selp.f32 %f1613, 0f3F800000, %f1103, %p80;
|
|
bra.uni BB0_57;
|
|
|
|
BB0_53:
|
|
setp.neu.f32 %p76, %f207, 0f7F800000;
|
|
@%p76 bra BB0_57;
|
|
|
|
setp.ltu.f32 %p77, %f203, 0f00000000;
|
|
selp.b32 %r133, 0, 2139095040, %p77;
|
|
or.b32 %r134, %r133, -2147483648;
|
|
selp.b32 %r135, %r134, %r133, %p4;
|
|
mov.b32 %f1613, %r135;
|
|
|
|
BB0_57:
|
|
setp.eq.f32 %p81, %f203, 0f00000000;
|
|
setp.eq.f32 %p82, %f202, 0f3F800000;
|
|
or.pred %p83, %p82, %p81;
|
|
selp.f32 %f1104, 0f3F800000, %f1613, %p83;
|
|
fma.rn.f32 %f1105, %f205, %f1104, %f204;
|
|
mul.f32 %f1106, %f140, %f1105;
|
|
div.rn.f32 %f1107, %f1106, %f402;
|
|
div.rn.f32 %f1108, %f1107, %f100;
|
|
cvt.sat.f32.f32 %f1109, %f1108;
|
|
mul.f32 %f1608, %f1608, %f1109;
|
|
|
|
BB0_58:
|
|
add.f32 %f1551, %f1551, %f1608;
|
|
|
|
BB0_59:
|
|
add.s32 %r265, %r265, 1;
|
|
setp.lt.u32 %p84, %r265, %r4;
|
|
@%p84 bra BB0_5;
|
|
|
|
BB0_60:
|
|
ld.global.u32 %r271, [imageEnabled];
|
|
and.b32 %r139, %r271, 8;
|
|
setp.eq.s32 %p85, %r139, 0;
|
|
@%p85 bra BB0_73;
|
|
|
|
cvt.sat.f32.f32 %f256, %f1551;
|
|
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 %f1112, 0f3E68BA2E;
|
|
cvt.rzi.f32.f32 %f1113, %f1112;
|
|
fma.rn.f32 %f1114, %f1113, 0fC0000000, 0f3EE8BA2E;
|
|
abs.f32 %f257, %f1114;
|
|
abs.f32 %f258, %f256;
|
|
setp.lt.f32 %p86, %f258, 0f00800000;
|
|
mul.f32 %f1115, %f258, 0f4B800000;
|
|
selp.f32 %f1116, 0fC3170000, 0fC2FE0000, %p86;
|
|
selp.f32 %f1117, %f1115, %f258, %p86;
|
|
mov.b32 %r142, %f1117;
|
|
and.b32 %r143, %r142, 8388607;
|
|
or.b32 %r144, %r143, 1065353216;
|
|
mov.b32 %f1118, %r144;
|
|
shr.u32 %r145, %r142, 23;
|
|
cvt.rn.f32.u32 %f1119, %r145;
|
|
add.f32 %f1120, %f1116, %f1119;
|
|
setp.gt.f32 %p87, %f1118, 0f3FB504F3;
|
|
mul.f32 %f1121, %f1118, 0f3F000000;
|
|
add.f32 %f1122, %f1120, 0f3F800000;
|
|
selp.f32 %f1123, %f1121, %f1118, %p87;
|
|
selp.f32 %f1124, %f1122, %f1120, %p87;
|
|
add.f32 %f1125, %f1123, 0fBF800000;
|
|
add.f32 %f1111, %f1123, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f1110,%f1111;
|
|
// inline asm
|
|
add.f32 %f1126, %f1125, %f1125;
|
|
mul.f32 %f1127, %f1110, %f1126;
|
|
mul.f32 %f1128, %f1127, %f1127;
|
|
mov.f32 %f1129, 0f3C4CAF63;
|
|
mov.f32 %f1130, 0f3B18F0FE;
|
|
fma.rn.f32 %f1131, %f1130, %f1128, %f1129;
|
|
mov.f32 %f1132, 0f3DAAAABD;
|
|
fma.rn.f32 %f1133, %f1131, %f1128, %f1132;
|
|
mul.rn.f32 %f1134, %f1133, %f1128;
|
|
mul.rn.f32 %f1135, %f1134, %f1127;
|
|
sub.f32 %f1136, %f1125, %f1127;
|
|
neg.f32 %f1137, %f1127;
|
|
add.f32 %f1138, %f1136, %f1136;
|
|
fma.rn.f32 %f1139, %f1137, %f1125, %f1138;
|
|
mul.rn.f32 %f1140, %f1110, %f1139;
|
|
add.f32 %f1141, %f1135, %f1127;
|
|
sub.f32 %f1142, %f1127, %f1141;
|
|
add.f32 %f1143, %f1135, %f1142;
|
|
add.f32 %f1144, %f1140, %f1143;
|
|
add.f32 %f1145, %f1141, %f1144;
|
|
sub.f32 %f1146, %f1141, %f1145;
|
|
add.f32 %f1147, %f1144, %f1146;
|
|
mov.f32 %f1148, 0f3F317200;
|
|
mul.rn.f32 %f1149, %f1124, %f1148;
|
|
mov.f32 %f1150, 0f35BFBE8E;
|
|
mul.rn.f32 %f1151, %f1124, %f1150;
|
|
add.f32 %f1152, %f1149, %f1145;
|
|
sub.f32 %f1153, %f1149, %f1152;
|
|
add.f32 %f1154, %f1145, %f1153;
|
|
add.f32 %f1155, %f1147, %f1154;
|
|
add.f32 %f1156, %f1151, %f1155;
|
|
add.f32 %f1157, %f1152, %f1156;
|
|
sub.f32 %f1158, %f1152, %f1157;
|
|
add.f32 %f1159, %f1156, %f1158;
|
|
mov.f32 %f1160, 0f3EE8BA2E;
|
|
mul.rn.f32 %f1161, %f1160, %f1157;
|
|
neg.f32 %f1162, %f1161;
|
|
fma.rn.f32 %f1163, %f1160, %f1157, %f1162;
|
|
fma.rn.f32 %f1164, %f1160, %f1159, %f1163;
|
|
mov.f32 %f1165, 0f00000000;
|
|
fma.rn.f32 %f1166, %f1165, %f1157, %f1164;
|
|
add.rn.f32 %f1167, %f1161, %f1166;
|
|
neg.f32 %f1168, %f1167;
|
|
add.rn.f32 %f1169, %f1161, %f1168;
|
|
add.rn.f32 %f1170, %f1169, %f1166;
|
|
mov.b32 %r146, %f1167;
|
|
setp.eq.s32 %p88, %r146, 1118925336;
|
|
add.s32 %r147, %r146, -1;
|
|
mov.b32 %f1171, %r147;
|
|
add.f32 %f1172, %f1170, 0f37000000;
|
|
selp.f32 %f1173, %f1171, %f1167, %p88;
|
|
selp.f32 %f259, %f1172, %f1170, %p88;
|
|
mul.f32 %f1174, %f1173, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f1175, %f1174;
|
|
mov.f32 %f1176, 0fBF317200;
|
|
fma.rn.f32 %f1177, %f1175, %f1176, %f1173;
|
|
mov.f32 %f1178, 0fB5BFBE8E;
|
|
fma.rn.f32 %f1179, %f1175, %f1178, %f1177;
|
|
mul.f32 %f1180, %f1179, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f1181, %f1180;
|
|
add.f32 %f1182, %f1175, 0f00000000;
|
|
ex2.approx.f32 %f1183, %f1182;
|
|
mul.f32 %f1184, %f1181, %f1183;
|
|
setp.lt.f32 %p89, %f1173, 0fC2D20000;
|
|
selp.f32 %f1185, 0f00000000, %f1184, %p89;
|
|
setp.gt.f32 %p90, %f1173, 0f42D20000;
|
|
selp.f32 %f1647, 0f7F800000, %f1185, %p90;
|
|
setp.eq.f32 %p91, %f1647, 0f7F800000;
|
|
@%p91 bra BB0_63;
|
|
|
|
fma.rn.f32 %f1647, %f1647, %f259, %f1647;
|
|
|
|
BB0_63:
|
|
setp.lt.f32 %p92, %f256, 0f00000000;
|
|
setp.eq.f32 %p93, %f257, 0f3F800000;
|
|
and.pred %p5, %p92, %p93;
|
|
mov.b32 %r148, %f1647;
|
|
xor.b32 %r149, %r148, -2147483648;
|
|
mov.b32 %f1186, %r149;
|
|
selp.f32 %f1649, %f1186, %f1647, %p5;
|
|
setp.eq.f32 %p94, %f256, 0f00000000;
|
|
@%p94 bra BB0_66;
|
|
bra.uni BB0_64;
|
|
|
|
BB0_66:
|
|
add.f32 %f1189, %f256, %f256;
|
|
selp.f32 %f1649, %f1189, 0f00000000, %p93;
|
|
bra.uni BB0_67;
|
|
|
|
BB0_131:
|
|
mov.u64 %rd190, image_HDR;
|
|
cvta.global.u64 %rd185, %rd190;
|
|
mov.u32 %r235, 8;
|
|
// inline asm
|
|
call (%rd184), _rt_buffer_get_64, (%rd185, %r28, %r235, %rd6, %rd7, %rd13, %rd13);
|
|
// inline asm
|
|
mov.f32 %f1509, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs107, %f1509;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs108, 0;
|
|
st.v4.u16 [%rd184], {%rs107, %rs107, %rs107, %rs108};
|
|
|
|
BB0_132:
|
|
ld.global.u32 %r236, [additive];
|
|
setp.eq.s32 %p161, %r236, 0;
|
|
@%p161 bra BB0_134;
|
|
|
|
mov.u64 %rd203, image_RNM0;
|
|
cvta.global.u64 %rd192, %rd203;
|
|
mov.u32 %r240, 8;
|
|
// inline asm
|
|
call (%rd191), _rt_buffer_get_64, (%rd192, %r28, %r240, %rd6, %rd7, %rd13, %rd13);
|
|
// inline asm
|
|
ld.v4.u16 {%rs115, %rs116, %rs117, %rs118}, [%rd191];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1510, %rs115;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1511, %rs116;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1512, %rs117;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd197), _rt_buffer_get_64, (%rd192, %r28, %r240, %rd6, %rd7, %rd13, %rd13);
|
|
// inline asm
|
|
add.f32 %f1513, %f1510, 0f00000000;
|
|
add.f32 %f1514, %f1511, 0f00000000;
|
|
add.f32 %f1515, %f1512, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs114, %f1515;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs113, %f1514;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs112, %f1513;}
|
|
|
|
// 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 %r242, 8;
|
|
// inline asm
|
|
call (%rd204), _rt_buffer_get_64, (%rd205, %r28, %r242, %rd6, %rd7, %rd13, %rd13);
|
|
// inline asm
|
|
mov.f32 %f1516, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs120, %f1516;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs121, 0;
|
|
st.v4.u16 [%rd204], {%rs120, %rs120, %rs120, %rs121};
|
|
|
|
BB0_135:
|
|
ld.global.u32 %r243, [additive];
|
|
setp.eq.s32 %p162, %r243, 0;
|
|
@%p162 bra BB0_137;
|
|
|
|
mov.u64 %rd223, image_RNM1;
|
|
cvta.global.u64 %rd212, %rd223;
|
|
mov.u32 %r247, 8;
|
|
// inline asm
|
|
call (%rd211), _rt_buffer_get_64, (%rd212, %r28, %r247, %rd6, %rd7, %rd13, %rd13);
|
|
// inline asm
|
|
ld.v4.u16 {%rs128, %rs129, %rs130, %rs131}, [%rd211];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1517, %rs128;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1518, %rs129;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1519, %rs130;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd217), _rt_buffer_get_64, (%rd212, %r28, %r247, %rd6, %rd7, %rd13, %rd13);
|
|
// inline asm
|
|
add.f32 %f1520, %f1517, 0f00000000;
|
|
add.f32 %f1521, %f1518, 0f00000000;
|
|
add.f32 %f1522, %f1519, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs127, %f1522;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs126, %f1521;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs125, %f1520;}
|
|
|
|
// 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 %r249, 8;
|
|
// inline asm
|
|
call (%rd224), _rt_buffer_get_64, (%rd225, %r28, %r249, %rd6, %rd7, %rd13, %rd13);
|
|
// inline asm
|
|
mov.f32 %f1523, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs133, %f1523;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs134, 0;
|
|
st.v4.u16 [%rd224], {%rs133, %rs133, %rs133, %rs134};
|
|
|
|
BB0_138:
|
|
ld.global.u32 %r250, [additive];
|
|
setp.eq.s32 %p163, %r250, 0;
|
|
@%p163 bra BB0_140;
|
|
|
|
mov.u64 %rd243, image_RNM2;
|
|
cvta.global.u64 %rd232, %rd243;
|
|
mov.u32 %r254, 8;
|
|
// inline asm
|
|
call (%rd231), _rt_buffer_get_64, (%rd232, %r28, %r254, %rd6, %rd7, %rd13, %rd13);
|
|
// inline asm
|
|
ld.v4.u16 {%rs141, %rs142, %rs143, %rs144}, [%rd231];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1524, %rs141;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1525, %rs142;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1526, %rs143;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd237), _rt_buffer_get_64, (%rd232, %r28, %r254, %rd6, %rd7, %rd13, %rd13);
|
|
// inline asm
|
|
add.f32 %f1527, %f1524, 0f00000000;
|
|
add.f32 %f1528, %f1525, 0f00000000;
|
|
add.f32 %f1529, %f1526, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs140, %f1529;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs139, %f1528;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs138, %f1527;}
|
|
|
|
// 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 %r256, 8;
|
|
// inline asm
|
|
call (%rd244), _rt_buffer_get_64, (%rd245, %r28, %r256, %rd6, %rd7, %rd13, %rd13);
|
|
// inline asm
|
|
mov.f32 %f1530, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs146, %f1530;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs147, 0;
|
|
st.v4.u16 [%rd244], {%rs146, %rs146, %rs146, %rs147};
|
|
|
|
BB0_141:
|
|
ld.global.u32 %r257, [additive];
|
|
setp.eq.s32 %p164, %r257, 0;
|
|
@%p164 bra BB0_143;
|
|
|
|
mov.u64 %rd263, image_RNM3;
|
|
cvta.global.u64 %rd252, %rd263;
|
|
mov.u32 %r261, 8;
|
|
// inline asm
|
|
call (%rd251), _rt_buffer_get_64, (%rd252, %r28, %r261, %rd6, %rd7, %rd13, %rd13);
|
|
// inline asm
|
|
ld.v4.u16 {%rs154, %rs155, %rs156, %rs157}, [%rd251];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1531, %rs154;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1532, %rs155;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1533, %rs156;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd257), _rt_buffer_get_64, (%rd252, %r28, %r261, %rd6, %rd7, %rd13, %rd13);
|
|
// inline asm
|
|
add.f32 %f1534, %f1531, 0f00000000;
|
|
add.f32 %f1535, %f1532, 0f00000000;
|
|
add.f32 %f1536, %f1533, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs153, %f1536;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs152, %f1535;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs151, %f1534;}
|
|
|
|
// 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 %r263, 8;
|
|
// inline asm
|
|
call (%rd264), _rt_buffer_get_64, (%rd265, %r28, %r263, %rd6, %rd7, %rd13, %rd13);
|
|
// inline asm
|
|
mov.f32 %f1537, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs159, %f1537;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs160, 0;
|
|
st.v4.u16 [%rd264], {%rs159, %rs159, %rs159, %rs160};
|
|
bra.uni BB0_144;
|
|
|
|
BB0_64:
|
|
setp.geu.f32 %p95, %f256, 0f00000000;
|
|
@%p95 bra BB0_67;
|
|
|
|
cvt.rzi.f32.f32 %f1188, %f1160;
|
|
setp.neu.f32 %p96, %f1188, 0f3EE8BA2E;
|
|
selp.f32 %f1649, 0f7FFFFFFF, %f1649, %p96;
|
|
|
|
BB0_67:
|
|
add.f32 %f1190, %f258, 0f3EE8BA2E;
|
|
mov.b32 %r150, %f1190;
|
|
setp.lt.s32 %p98, %r150, 2139095040;
|
|
@%p98 bra BB0_72;
|
|
|
|
setp.gtu.f32 %p99, %f258, 0f7F800000;
|
|
@%p99 bra BB0_71;
|
|
bra.uni BB0_69;
|
|
|
|
BB0_71:
|
|
add.f32 %f1649, %f256, 0f3EE8BA2E;
|
|
bra.uni BB0_72;
|
|
|
|
BB0_69:
|
|
setp.neu.f32 %p100, %f258, 0f7F800000;
|
|
@%p100 bra BB0_72;
|
|
|
|
selp.f32 %f1649, 0fFF800000, 0f7F800000, %p5;
|
|
|
|
BB0_72:
|
|
mul.f32 %f1191, %f1649, 0f437F0000;
|
|
setp.eq.f32 %p101, %f256, 0f3F800000;
|
|
selp.f32 %f1192, 0f437F0000, %f1191, %p101;
|
|
cvt.rzi.u32.f32 %r151, %f1192;
|
|
cvt.u16.u32 %rs17, %r151;
|
|
mov.u16 %rs18, 255;
|
|
st.v2.u8 [%rd43], {%rs17, %rs18};
|
|
ld.global.u32 %r271, [imageEnabled];
|
|
|
|
BB0_73:
|
|
and.b32 %r152, %r271, 1;
|
|
setp.eq.b32 %p102, %r152, 1;
|
|
@!%p102 bra BB0_108;
|
|
bra.uni BB0_74;
|
|
|
|
BB0_74:
|
|
mov.f32 %f1195, 0f3E666666;
|
|
cvt.rzi.f32.f32 %f1196, %f1195;
|
|
fma.rn.f32 %f1197, %f1196, 0fC0000000, 0f3EE66666;
|
|
abs.f32 %f270, %f1197;
|
|
abs.f32 %f271, %f1566;
|
|
setp.lt.f32 %p103, %f271, 0f00800000;
|
|
mul.f32 %f1198, %f271, 0f4B800000;
|
|
selp.f32 %f1199, 0fC3170000, 0fC2FE0000, %p103;
|
|
selp.f32 %f1200, %f1198, %f271, %p103;
|
|
mov.b32 %r153, %f1200;
|
|
and.b32 %r154, %r153, 8388607;
|
|
or.b32 %r155, %r154, 1065353216;
|
|
mov.b32 %f1201, %r155;
|
|
shr.u32 %r156, %r153, 23;
|
|
cvt.rn.f32.u32 %f1202, %r156;
|
|
add.f32 %f1203, %f1199, %f1202;
|
|
setp.gt.f32 %p104, %f1201, 0f3FB504F3;
|
|
mul.f32 %f1204, %f1201, 0f3F000000;
|
|
add.f32 %f1205, %f1203, 0f3F800000;
|
|
selp.f32 %f1206, %f1204, %f1201, %p104;
|
|
selp.f32 %f1207, %f1205, %f1203, %p104;
|
|
add.f32 %f1208, %f1206, 0fBF800000;
|
|
add.f32 %f1194, %f1206, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f1193,%f1194;
|
|
// inline asm
|
|
add.f32 %f1209, %f1208, %f1208;
|
|
mul.f32 %f1210, %f1193, %f1209;
|
|
mul.f32 %f1211, %f1210, %f1210;
|
|
mov.f32 %f1212, 0f3C4CAF63;
|
|
mov.f32 %f1213, 0f3B18F0FE;
|
|
fma.rn.f32 %f1214, %f1213, %f1211, %f1212;
|
|
mov.f32 %f1215, 0f3DAAAABD;
|
|
fma.rn.f32 %f1216, %f1214, %f1211, %f1215;
|
|
mul.rn.f32 %f1217, %f1216, %f1211;
|
|
mul.rn.f32 %f1218, %f1217, %f1210;
|
|
sub.f32 %f1219, %f1208, %f1210;
|
|
neg.f32 %f1220, %f1210;
|
|
add.f32 %f1221, %f1219, %f1219;
|
|
fma.rn.f32 %f1222, %f1220, %f1208, %f1221;
|
|
mul.rn.f32 %f1223, %f1193, %f1222;
|
|
add.f32 %f1224, %f1218, %f1210;
|
|
sub.f32 %f1225, %f1210, %f1224;
|
|
add.f32 %f1226, %f1218, %f1225;
|
|
add.f32 %f1227, %f1223, %f1226;
|
|
add.f32 %f1228, %f1224, %f1227;
|
|
sub.f32 %f1229, %f1224, %f1228;
|
|
add.f32 %f1230, %f1227, %f1229;
|
|
mov.f32 %f1231, 0f3F317200;
|
|
mul.rn.f32 %f1232, %f1207, %f1231;
|
|
mov.f32 %f1233, 0f35BFBE8E;
|
|
mul.rn.f32 %f1234, %f1207, %f1233;
|
|
add.f32 %f1235, %f1232, %f1228;
|
|
sub.f32 %f1236, %f1232, %f1235;
|
|
add.f32 %f1237, %f1228, %f1236;
|
|
add.f32 %f1238, %f1230, %f1237;
|
|
add.f32 %f1239, %f1234, %f1238;
|
|
add.f32 %f1240, %f1235, %f1239;
|
|
sub.f32 %f1241, %f1235, %f1240;
|
|
add.f32 %f1242, %f1239, %f1241;
|
|
mov.f32 %f1243, 0f3EE66666;
|
|
mul.rn.f32 %f1244, %f1243, %f1240;
|
|
neg.f32 %f1245, %f1244;
|
|
fma.rn.f32 %f1246, %f1243, %f1240, %f1245;
|
|
fma.rn.f32 %f1247, %f1243, %f1242, %f1246;
|
|
mov.f32 %f1248, 0f00000000;
|
|
fma.rn.f32 %f1249, %f1248, %f1240, %f1247;
|
|
add.rn.f32 %f1250, %f1244, %f1249;
|
|
neg.f32 %f1251, %f1250;
|
|
add.rn.f32 %f1252, %f1244, %f1251;
|
|
add.rn.f32 %f1253, %f1252, %f1249;
|
|
mov.b32 %r157, %f1250;
|
|
setp.eq.s32 %p105, %r157, 1118925336;
|
|
add.s32 %r158, %r157, -1;
|
|
mov.b32 %f1254, %r158;
|
|
add.f32 %f1255, %f1253, 0f37000000;
|
|
selp.f32 %f1256, %f1254, %f1250, %p105;
|
|
selp.f32 %f272, %f1255, %f1253, %p105;
|
|
mul.f32 %f1257, %f1256, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f1258, %f1257;
|
|
mov.f32 %f1259, 0fBF317200;
|
|
fma.rn.f32 %f1260, %f1258, %f1259, %f1256;
|
|
mov.f32 %f1261, 0fB5BFBE8E;
|
|
fma.rn.f32 %f1262, %f1258, %f1261, %f1260;
|
|
mul.f32 %f1263, %f1262, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f1264, %f1263;
|
|
add.f32 %f1265, %f1258, 0f00000000;
|
|
ex2.approx.f32 %f1266, %f1265;
|
|
mul.f32 %f1267, %f1264, %f1266;
|
|
setp.lt.f32 %p106, %f1256, 0fC2D20000;
|
|
selp.f32 %f1268, 0f00000000, %f1267, %p106;
|
|
setp.gt.f32 %p107, %f1256, 0f42D20000;
|
|
selp.f32 %f1650, 0f7F800000, %f1268, %p107;
|
|
setp.eq.f32 %p108, %f1650, 0f7F800000;
|
|
@%p108 bra BB0_76;
|
|
|
|
fma.rn.f32 %f1650, %f1650, %f272, %f1650;
|
|
|
|
BB0_76:
|
|
setp.lt.f32 %p109, %f1566, 0f00000000;
|
|
setp.eq.f32 %p110, %f270, 0f3F800000;
|
|
and.pred %p6, %p109, %p110;
|
|
mov.b32 %r159, %f1650;
|
|
xor.b32 %r160, %r159, -2147483648;
|
|
mov.b32 %f1269, %r160;
|
|
selp.f32 %f1652, %f1269, %f1650, %p6;
|
|
setp.eq.f32 %p111, %f1566, 0f00000000;
|
|
@%p111 bra BB0_79;
|
|
bra.uni BB0_77;
|
|
|
|
BB0_79:
|
|
add.f32 %f1272, %f1566, %f1566;
|
|
selp.f32 %f1652, %f1272, 0f00000000, %p110;
|
|
bra.uni BB0_80;
|
|
|
|
BB0_77:
|
|
setp.geu.f32 %p112, %f1566, 0f00000000;
|
|
@%p112 bra BB0_80;
|
|
|
|
cvt.rzi.f32.f32 %f1271, %f1243;
|
|
setp.neu.f32 %p113, %f1271, 0f3EE66666;
|
|
selp.f32 %f1652, 0f7FFFFFFF, %f1652, %p113;
|
|
|
|
BB0_80:
|
|
add.f32 %f1273, %f271, 0f3EE66666;
|
|
mov.b32 %r161, %f1273;
|
|
setp.lt.s32 %p115, %r161, 2139095040;
|
|
@%p115 bra BB0_85;
|
|
|
|
setp.gtu.f32 %p116, %f271, 0f7F800000;
|
|
@%p116 bra BB0_84;
|
|
bra.uni BB0_82;
|
|
|
|
BB0_84:
|
|
add.f32 %f1652, %f1566, 0f3EE66666;
|
|
bra.uni BB0_85;
|
|
|
|
BB0_82:
|
|
setp.neu.f32 %p117, %f271, 0f7F800000;
|
|
@%p117 bra BB0_85;
|
|
|
|
selp.f32 %f1652, 0fFF800000, 0f7F800000, %p6;
|
|
|
|
BB0_85:
|
|
setp.eq.f32 %p118, %f1566, 0f3F800000;
|
|
selp.f32 %f283, 0f3F800000, %f1652, %p118;
|
|
abs.f32 %f284, %f1565;
|
|
setp.lt.f32 %p119, %f284, 0f00800000;
|
|
mul.f32 %f1276, %f284, 0f4B800000;
|
|
selp.f32 %f1277, 0fC3170000, 0fC2FE0000, %p119;
|
|
selp.f32 %f1278, %f1276, %f284, %p119;
|
|
mov.b32 %r162, %f1278;
|
|
and.b32 %r163, %r162, 8388607;
|
|
or.b32 %r164, %r163, 1065353216;
|
|
mov.b32 %f1279, %r164;
|
|
shr.u32 %r165, %r162, 23;
|
|
cvt.rn.f32.u32 %f1280, %r165;
|
|
add.f32 %f1281, %f1277, %f1280;
|
|
setp.gt.f32 %p120, %f1279, 0f3FB504F3;
|
|
mul.f32 %f1282, %f1279, 0f3F000000;
|
|
add.f32 %f1283, %f1281, 0f3F800000;
|
|
selp.f32 %f1284, %f1282, %f1279, %p120;
|
|
selp.f32 %f1285, %f1283, %f1281, %p120;
|
|
add.f32 %f1286, %f1284, 0fBF800000;
|
|
add.f32 %f1275, %f1284, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f1274,%f1275;
|
|
// inline asm
|
|
add.f32 %f1287, %f1286, %f1286;
|
|
mul.f32 %f1288, %f1274, %f1287;
|
|
mul.f32 %f1289, %f1288, %f1288;
|
|
fma.rn.f32 %f1292, %f1213, %f1289, %f1212;
|
|
fma.rn.f32 %f1294, %f1292, %f1289, %f1215;
|
|
mul.rn.f32 %f1295, %f1294, %f1289;
|
|
mul.rn.f32 %f1296, %f1295, %f1288;
|
|
sub.f32 %f1297, %f1286, %f1288;
|
|
neg.f32 %f1298, %f1288;
|
|
add.f32 %f1299, %f1297, %f1297;
|
|
fma.rn.f32 %f1300, %f1298, %f1286, %f1299;
|
|
mul.rn.f32 %f1301, %f1274, %f1300;
|
|
add.f32 %f1302, %f1296, %f1288;
|
|
sub.f32 %f1303, %f1288, %f1302;
|
|
add.f32 %f1304, %f1296, %f1303;
|
|
add.f32 %f1305, %f1301, %f1304;
|
|
add.f32 %f1306, %f1302, %f1305;
|
|
sub.f32 %f1307, %f1302, %f1306;
|
|
add.f32 %f1308, %f1305, %f1307;
|
|
mul.rn.f32 %f1310, %f1285, %f1231;
|
|
mul.rn.f32 %f1312, %f1285, %f1233;
|
|
add.f32 %f1313, %f1310, %f1306;
|
|
sub.f32 %f1314, %f1310, %f1313;
|
|
add.f32 %f1315, %f1306, %f1314;
|
|
add.f32 %f1316, %f1308, %f1315;
|
|
add.f32 %f1317, %f1312, %f1316;
|
|
add.f32 %f1318, %f1313, %f1317;
|
|
sub.f32 %f1319, %f1313, %f1318;
|
|
add.f32 %f1320, %f1317, %f1319;
|
|
mul.rn.f32 %f1322, %f1243, %f1318;
|
|
neg.f32 %f1323, %f1322;
|
|
fma.rn.f32 %f1324, %f1243, %f1318, %f1323;
|
|
fma.rn.f32 %f1325, %f1243, %f1320, %f1324;
|
|
fma.rn.f32 %f1327, %f1248, %f1318, %f1325;
|
|
add.rn.f32 %f1328, %f1322, %f1327;
|
|
neg.f32 %f1329, %f1328;
|
|
add.rn.f32 %f1330, %f1322, %f1329;
|
|
add.rn.f32 %f1331, %f1330, %f1327;
|
|
mov.b32 %r166, %f1328;
|
|
setp.eq.s32 %p121, %r166, 1118925336;
|
|
add.s32 %r167, %r166, -1;
|
|
mov.b32 %f1332, %r167;
|
|
add.f32 %f1333, %f1331, 0f37000000;
|
|
selp.f32 %f1334, %f1332, %f1328, %p121;
|
|
selp.f32 %f285, %f1333, %f1331, %p121;
|
|
mul.f32 %f1335, %f1334, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f1336, %f1335;
|
|
fma.rn.f32 %f1338, %f1336, %f1259, %f1334;
|
|
fma.rn.f32 %f1340, %f1336, %f1261, %f1338;
|
|
mul.f32 %f1341, %f1340, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f1342, %f1341;
|
|
add.f32 %f1343, %f1336, 0f00000000;
|
|
ex2.approx.f32 %f1344, %f1343;
|
|
mul.f32 %f1345, %f1342, %f1344;
|
|
setp.lt.f32 %p122, %f1334, 0fC2D20000;
|
|
selp.f32 %f1346, 0f00000000, %f1345, %p122;
|
|
setp.gt.f32 %p123, %f1334, 0f42D20000;
|
|
selp.f32 %f1653, 0f7F800000, %f1346, %p123;
|
|
setp.eq.f32 %p124, %f1653, 0f7F800000;
|
|
@%p124 bra BB0_87;
|
|
|
|
fma.rn.f32 %f1653, %f1653, %f285, %f1653;
|
|
|
|
BB0_87:
|
|
setp.lt.f32 %p125, %f1565, 0f00000000;
|
|
and.pred %p7, %p125, %p110;
|
|
mov.b32 %r168, %f1653;
|
|
xor.b32 %r169, %r168, -2147483648;
|
|
mov.b32 %f1347, %r169;
|
|
selp.f32 %f1655, %f1347, %f1653, %p7;
|
|
setp.eq.f32 %p127, %f1565, 0f00000000;
|
|
@%p127 bra BB0_90;
|
|
bra.uni BB0_88;
|
|
|
|
BB0_90:
|
|
add.f32 %f1350, %f1565, %f1565;
|
|
selp.f32 %f1655, %f1350, 0f00000000, %p110;
|
|
bra.uni BB0_91;
|
|
|
|
BB0_88:
|
|
setp.geu.f32 %p128, %f1565, 0f00000000;
|
|
@%p128 bra BB0_91;
|
|
|
|
cvt.rzi.f32.f32 %f1349, %f1243;
|
|
setp.neu.f32 %p129, %f1349, 0f3EE66666;
|
|
selp.f32 %f1655, 0f7FFFFFFF, %f1655, %p129;
|
|
|
|
BB0_91:
|
|
add.f32 %f1351, %f284, 0f3EE66666;
|
|
mov.b32 %r170, %f1351;
|
|
setp.lt.s32 %p131, %r170, 2139095040;
|
|
@%p131 bra BB0_96;
|
|
|
|
setp.gtu.f32 %p132, %f284, 0f7F800000;
|
|
@%p132 bra BB0_95;
|
|
bra.uni BB0_93;
|
|
|
|
BB0_95:
|
|
add.f32 %f1655, %f1565, 0f3EE66666;
|
|
bra.uni BB0_96;
|
|
|
|
BB0_93:
|
|
setp.neu.f32 %p133, %f284, 0f7F800000;
|
|
@%p133 bra BB0_96;
|
|
|
|
selp.f32 %f1655, 0fFF800000, 0f7F800000, %p7;
|
|
|
|
BB0_96:
|
|
setp.eq.f32 %p134, %f1565, 0f3F800000;
|
|
selp.f32 %f296, 0f3F800000, %f1655, %p134;
|
|
abs.f32 %f297, %f1564;
|
|
setp.lt.f32 %p135, %f297, 0f00800000;
|
|
mul.f32 %f1354, %f297, 0f4B800000;
|
|
selp.f32 %f1355, 0fC3170000, 0fC2FE0000, %p135;
|
|
selp.f32 %f1356, %f1354, %f297, %p135;
|
|
mov.b32 %r171, %f1356;
|
|
and.b32 %r172, %r171, 8388607;
|
|
or.b32 %r173, %r172, 1065353216;
|
|
mov.b32 %f1357, %r173;
|
|
shr.u32 %r174, %r171, 23;
|
|
cvt.rn.f32.u32 %f1358, %r174;
|
|
add.f32 %f1359, %f1355, %f1358;
|
|
setp.gt.f32 %p136, %f1357, 0f3FB504F3;
|
|
mul.f32 %f1360, %f1357, 0f3F000000;
|
|
add.f32 %f1361, %f1359, 0f3F800000;
|
|
selp.f32 %f1362, %f1360, %f1357, %p136;
|
|
selp.f32 %f1363, %f1361, %f1359, %p136;
|
|
add.f32 %f1364, %f1362, 0fBF800000;
|
|
add.f32 %f1353, %f1362, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f1352,%f1353;
|
|
// inline asm
|
|
add.f32 %f1365, %f1364, %f1364;
|
|
mul.f32 %f1366, %f1352, %f1365;
|
|
mul.f32 %f1367, %f1366, %f1366;
|
|
fma.rn.f32 %f1370, %f1213, %f1367, %f1212;
|
|
fma.rn.f32 %f1372, %f1370, %f1367, %f1215;
|
|
mul.rn.f32 %f1373, %f1372, %f1367;
|
|
mul.rn.f32 %f1374, %f1373, %f1366;
|
|
sub.f32 %f1375, %f1364, %f1366;
|
|
neg.f32 %f1376, %f1366;
|
|
add.f32 %f1377, %f1375, %f1375;
|
|
fma.rn.f32 %f1378, %f1376, %f1364, %f1377;
|
|
mul.rn.f32 %f1379, %f1352, %f1378;
|
|
add.f32 %f1380, %f1374, %f1366;
|
|
sub.f32 %f1381, %f1366, %f1380;
|
|
add.f32 %f1382, %f1374, %f1381;
|
|
add.f32 %f1383, %f1379, %f1382;
|
|
add.f32 %f1384, %f1380, %f1383;
|
|
sub.f32 %f1385, %f1380, %f1384;
|
|
add.f32 %f1386, %f1383, %f1385;
|
|
mul.rn.f32 %f1388, %f1363, %f1231;
|
|
mul.rn.f32 %f1390, %f1363, %f1233;
|
|
add.f32 %f1391, %f1388, %f1384;
|
|
sub.f32 %f1392, %f1388, %f1391;
|
|
add.f32 %f1393, %f1384, %f1392;
|
|
add.f32 %f1394, %f1386, %f1393;
|
|
add.f32 %f1395, %f1390, %f1394;
|
|
add.f32 %f1396, %f1391, %f1395;
|
|
sub.f32 %f1397, %f1391, %f1396;
|
|
add.f32 %f1398, %f1395, %f1397;
|
|
mul.rn.f32 %f1400, %f1243, %f1396;
|
|
neg.f32 %f1401, %f1400;
|
|
fma.rn.f32 %f1402, %f1243, %f1396, %f1401;
|
|
fma.rn.f32 %f1403, %f1243, %f1398, %f1402;
|
|
fma.rn.f32 %f1405, %f1248, %f1396, %f1403;
|
|
add.rn.f32 %f1406, %f1400, %f1405;
|
|
neg.f32 %f1407, %f1406;
|
|
add.rn.f32 %f1408, %f1400, %f1407;
|
|
add.rn.f32 %f1409, %f1408, %f1405;
|
|
mov.b32 %r175, %f1406;
|
|
setp.eq.s32 %p137, %r175, 1118925336;
|
|
add.s32 %r176, %r175, -1;
|
|
mov.b32 %f1410, %r176;
|
|
add.f32 %f1411, %f1409, 0f37000000;
|
|
selp.f32 %f1412, %f1410, %f1406, %p137;
|
|
selp.f32 %f298, %f1411, %f1409, %p137;
|
|
mul.f32 %f1413, %f1412, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f1414, %f1413;
|
|
fma.rn.f32 %f1416, %f1414, %f1259, %f1412;
|
|
fma.rn.f32 %f1418, %f1414, %f1261, %f1416;
|
|
mul.f32 %f1419, %f1418, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f1420, %f1419;
|
|
add.f32 %f1421, %f1414, 0f00000000;
|
|
ex2.approx.f32 %f1422, %f1421;
|
|
mul.f32 %f1423, %f1420, %f1422;
|
|
setp.lt.f32 %p138, %f1412, 0fC2D20000;
|
|
selp.f32 %f1424, 0f00000000, %f1423, %p138;
|
|
setp.gt.f32 %p139, %f1412, 0f42D20000;
|
|
selp.f32 %f1656, 0f7F800000, %f1424, %p139;
|
|
setp.eq.f32 %p140, %f1656, 0f7F800000;
|
|
@%p140 bra BB0_98;
|
|
|
|
fma.rn.f32 %f1656, %f1656, %f298, %f1656;
|
|
|
|
BB0_98:
|
|
setp.lt.f32 %p141, %f1564, 0f00000000;
|
|
and.pred %p8, %p141, %p110;
|
|
mov.b32 %r177, %f1656;
|
|
xor.b32 %r178, %r177, -2147483648;
|
|
mov.b32 %f1425, %r178;
|
|
selp.f32 %f1658, %f1425, %f1656, %p8;
|
|
setp.eq.f32 %p143, %f1564, 0f00000000;
|
|
@%p143 bra BB0_101;
|
|
bra.uni BB0_99;
|
|
|
|
BB0_101:
|
|
add.f32 %f1428, %f1564, %f1564;
|
|
selp.f32 %f1658, %f1428, 0f00000000, %p110;
|
|
bra.uni BB0_102;
|
|
|
|
BB0_99:
|
|
setp.geu.f32 %p144, %f1564, 0f00000000;
|
|
@%p144 bra BB0_102;
|
|
|
|
cvt.rzi.f32.f32 %f1427, %f1243;
|
|
setp.neu.f32 %p145, %f1427, 0f3EE66666;
|
|
selp.f32 %f1658, 0f7FFFFFFF, %f1658, %p145;
|
|
|
|
BB0_102:
|
|
add.f32 %f1429, %f297, 0f3EE66666;
|
|
mov.b32 %r179, %f1429;
|
|
setp.lt.s32 %p147, %r179, 2139095040;
|
|
@%p147 bra BB0_107;
|
|
|
|
setp.gtu.f32 %p148, %f297, 0f7F800000;
|
|
@%p148 bra BB0_106;
|
|
bra.uni BB0_104;
|
|
|
|
BB0_106:
|
|
add.f32 %f1658, %f1564, 0f3EE66666;
|
|
bra.uni BB0_107;
|
|
|
|
BB0_104:
|
|
setp.neu.f32 %p149, %f297, 0f7F800000;
|
|
@%p149 bra BB0_107;
|
|
|
|
selp.f32 %f1658, 0fFF800000, 0f7F800000, %p8;
|
|
|
|
BB0_107:
|
|
mov.u32 %r264, 4;
|
|
setp.eq.f32 %p150, %f1564, 0f3F800000;
|
|
selp.f32 %f1430, 0f3F800000, %f1658, %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, %r264, %rd52, %rd53, %rd13, %rd13);
|
|
// inline asm
|
|
cvt.sat.f32.f32 %f1431, %f1430;
|
|
mul.f32 %f1432, %f1431, 0f437FFD71;
|
|
cvt.rzi.u32.f32 %r182, %f1432;
|
|
cvt.sat.f32.f32 %f1433, %f296;
|
|
mul.f32 %f1434, %f1433, 0f437FFD71;
|
|
cvt.rzi.u32.f32 %r183, %f1434;
|
|
cvt.sat.f32.f32 %f1435, %f283;
|
|
mul.f32 %f1436, %f1435, 0f437FFD71;
|
|
cvt.rzi.u32.f32 %r184, %f1436;
|
|
cvt.u16.u32 %rs19, %r182;
|
|
cvt.u16.u32 %rs20, %r184;
|
|
cvt.u16.u32 %rs21, %r183;
|
|
mov.u16 %rs22, 255;
|
|
st.v4.u8 [%rd50], {%rs19, %rs21, %rs20, %rs22};
|
|
ld.global.u32 %r271, [imageEnabled];
|
|
|
|
BB0_108:
|
|
cvt.u64.u32 %rd4, %r2;
|
|
cvt.u64.u32 %rd5, %r3;
|
|
and.b32 %r185, %r271, 4;
|
|
setp.eq.s32 %p151, %r185, 0;
|
|
@%p151 bra BB0_112;
|
|
|
|
ld.global.u32 %r186, [additive];
|
|
setp.eq.s32 %p152, %r186, 0;
|
|
mov.f32 %f1437, 0f3F800000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs23, %f1437;}
|
|
|
|
// inline asm
|
|
@%p152 bra BB0_111;
|
|
|
|
mov.u64 %rd69, image_HDR;
|
|
cvta.global.u64 %rd58, %rd69;
|
|
mov.u32 %r190, 8;
|
|
// inline asm
|
|
call (%rd57), _rt_buffer_get_64, (%rd58, %r28, %r190, %rd4, %rd5, %rd13, %rd13);
|
|
// inline asm
|
|
ld.v4.u16 {%rs30, %rs31, %rs32, %rs33}, [%rd57];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1438, %rs30;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1439, %rs31;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1440, %rs32;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd63), _rt_buffer_get_64, (%rd58, %r28, %r190, %rd4, %rd5, %rd13, %rd13);
|
|
// inline asm
|
|
add.f32 %f1441, %f1566, %f1438;
|
|
add.f32 %f1442, %f1565, %f1439;
|
|
add.f32 %f1443, %f1564, %f1440;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs29, %f1443;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs28, %f1442;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs27, %f1441;}
|
|
|
|
// 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 %r192, 8;
|
|
// inline asm
|
|
call (%rd70), _rt_buffer_get_64, (%rd71, %r28, %r192, %rd4, %rd5, %rd13, %rd13);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs36, %f1564;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs35, %f1565;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs34, %f1566;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd70], {%rs34, %rs35, %rs36, %rs23};
|
|
|
|
BB0_112:
|
|
mov.f32 %f1448, 0f34000000;
|
|
max.f32 %f1449, %f1563, %f1448;
|
|
div.rn.f32 %f1450, %f1560, %f1449;
|
|
max.f32 %f1451, %f1562, %f1448;
|
|
div.rn.f32 %f1452, %f1559, %f1451;
|
|
max.f32 %f1453, %f1561, %f1448;
|
|
div.rn.f32 %f1454, %f1558, %f1453;
|
|
fma.rn.f32 %f309, %f1450, 0f3F000000, 0f3F000000;
|
|
fma.rn.f32 %f310, %f1452, 0f3F000000, 0f3F000000;
|
|
fma.rn.f32 %f311, %f1454, 0f3F000000, 0f3F000000;
|
|
div.rn.f32 %f1455, %f1557, %f1449;
|
|
div.rn.f32 %f1456, %f1556, %f1451;
|
|
div.rn.f32 %f1457, %f1555, %f1453;
|
|
fma.rn.f32 %f312, %f1455, 0f3F000000, 0f3F000000;
|
|
fma.rn.f32 %f313, %f1456, 0f3F000000, 0f3F000000;
|
|
fma.rn.f32 %f314, %f1457, 0f3F000000, 0f3F000000;
|
|
div.rn.f32 %f1458, %f1554, %f1449;
|
|
div.rn.f32 %f1459, %f1553, %f1451;
|
|
div.rn.f32 %f1460, %f1552, %f1453;
|
|
fma.rn.f32 %f315, %f1458, 0f3F000000, 0f3F000000;
|
|
fma.rn.f32 %f316, %f1459, 0f3F000000, 0f3F000000;
|
|
fma.rn.f32 %f317, %f1460, 0f3F000000, 0f3F000000;
|
|
ld.global.u32 %r193, [additive];
|
|
setp.eq.s32 %p153, %r193, 0;
|
|
mov.f32 %f1447, 0f3F800000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs37, %f1447;}
|
|
|
|
// inline asm
|
|
@%p153 bra BB0_114;
|
|
|
|
mov.u64 %rd89, image_RNM0;
|
|
cvta.global.u64 %rd78, %rd89;
|
|
mov.u32 %r197, 8;
|
|
// inline asm
|
|
call (%rd77), _rt_buffer_get_64, (%rd78, %r28, %r197, %rd4, %rd5, %rd13, %rd13);
|
|
// inline asm
|
|
ld.v4.u16 {%rs44, %rs45, %rs46, %rs47}, [%rd77];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1461, %rs44;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1462, %rs45;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1463, %rs46;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd83), _rt_buffer_get_64, (%rd78, %r28, %r197, %rd4, %rd5, %rd13, %rd13);
|
|
// inline asm
|
|
add.f32 %f1464, %f1563, %f1461;
|
|
add.f32 %f1465, %f1562, %f1462;
|
|
add.f32 %f1466, %f1561, %f1463;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs43, %f1466;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs42, %f1465;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs41, %f1464;}
|
|
|
|
// 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 %r199, 8;
|
|
// inline asm
|
|
call (%rd90), _rt_buffer_get_64, (%rd91, %r28, %r199, %rd4, %rd5, %rd13, %rd13);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs50, %f1561;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs49, %f1562;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs48, %f1563;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd90], {%rs48, %rs49, %rs50, %rs37};
|
|
|
|
BB0_115:
|
|
ld.global.u32 %r200, [additive];
|
|
setp.eq.s32 %p154, %r200, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs51, %f1447;}
|
|
|
|
// inline asm
|
|
@%p154 bra BB0_117;
|
|
|
|
mov.u64 %rd109, image_RNM1;
|
|
cvta.global.u64 %rd98, %rd109;
|
|
mov.u32 %r204, 8;
|
|
// inline asm
|
|
call (%rd97), _rt_buffer_get_64, (%rd98, %r28, %r204, %rd4, %rd5, %rd13, %rd13);
|
|
// inline asm
|
|
ld.v4.u16 {%rs58, %rs59, %rs60, %rs61}, [%rd97];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1471, %rs58;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1472, %rs59;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1473, %rs60;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd103), _rt_buffer_get_64, (%rd98, %r28, %r204, %rd4, %rd5, %rd13, %rd13);
|
|
// inline asm
|
|
add.f32 %f1474, %f309, %f1471;
|
|
add.f32 %f1475, %f310, %f1472;
|
|
add.f32 %f1476, %f311, %f1473;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs57, %f1476;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs56, %f1475;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs55, %f1474;}
|
|
|
|
// 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 %r206, 8;
|
|
// inline asm
|
|
call (%rd110), _rt_buffer_get_64, (%rd111, %r28, %r206, %rd4, %rd5, %rd13, %rd13);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs64, %f311;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs63, %f310;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs62, %f309;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd110], {%rs62, %rs63, %rs64, %rs51};
|
|
|
|
BB0_118:
|
|
ld.global.u32 %r207, [additive];
|
|
setp.eq.s32 %p155, %r207, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs65, %f1447;}
|
|
|
|
// inline asm
|
|
@%p155 bra BB0_120;
|
|
|
|
mov.u64 %rd129, image_RNM2;
|
|
cvta.global.u64 %rd118, %rd129;
|
|
mov.u32 %r211, 8;
|
|
// inline asm
|
|
call (%rd117), _rt_buffer_get_64, (%rd118, %r28, %r211, %rd4, %rd5, %rd13, %rd13);
|
|
// inline asm
|
|
ld.v4.u16 {%rs72, %rs73, %rs74, %rs75}, [%rd117];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1481, %rs72;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1482, %rs73;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1483, %rs74;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd123), _rt_buffer_get_64, (%rd118, %r28, %r211, %rd4, %rd5, %rd13, %rd13);
|
|
// inline asm
|
|
add.f32 %f1484, %f312, %f1481;
|
|
add.f32 %f1485, %f313, %f1482;
|
|
add.f32 %f1486, %f314, %f1483;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs71, %f1486;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs70, %f1485;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs69, %f1484;}
|
|
|
|
// 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 %r213, 8;
|
|
// inline asm
|
|
call (%rd130), _rt_buffer_get_64, (%rd131, %r28, %r213, %rd4, %rd5, %rd13, %rd13);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs78, %f314;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs77, %f313;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs76, %f312;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd130], {%rs76, %rs77, %rs78, %rs65};
|
|
|
|
BB0_121:
|
|
ld.global.u32 %r214, [additive];
|
|
setp.eq.s32 %p156, %r214, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs79, %f1447;}
|
|
|
|
// inline asm
|
|
@%p156 bra BB0_123;
|
|
|
|
mov.u64 %rd149, image_RNM3;
|
|
cvta.global.u64 %rd138, %rd149;
|
|
mov.u32 %r218, 8;
|
|
// inline asm
|
|
call (%rd137), _rt_buffer_get_64, (%rd138, %r28, %r218, %rd4, %rd5, %rd13, %rd13);
|
|
// inline asm
|
|
ld.v4.u16 {%rs86, %rs87, %rs88, %rs89}, [%rd137];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1491, %rs86;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1492, %rs87;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1493, %rs88;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd143), _rt_buffer_get_64, (%rd138, %r28, %r218, %rd4, %rd5, %rd13, %rd13);
|
|
// inline asm
|
|
add.f32 %f1494, %f315, %f1491;
|
|
add.f32 %f1495, %f316, %f1492;
|
|
add.f32 %f1496, %f317, %f1493;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs85, %f1496;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs84, %f1495;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs83, %f1494;}
|
|
|
|
// 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 %r220, 8;
|
|
// inline asm
|
|
call (%rd150), _rt_buffer_get_64, (%rd151, %r28, %r220, %rd4, %rd5, %rd13, %rd13);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs92, %f317;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs91, %f316;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs90, %f315;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd150], {%rs90, %rs91, %rs92, %rs79};
|
|
|
|
BB0_144:
|
|
ret;
|
|
}
|
|
|
|
|