2649 lines
76 KiB
Plaintext
2649 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<165>;
|
|
.reg .b16 %rs<162>;
|
|
.reg .f32 %f<1575>;
|
|
.reg .b32 %r<273>;
|
|
.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 %f1480, 0f00000000;
|
|
mov.f32 %f1481, %f1480;
|
|
mov.f32 %f1482, %f1480;
|
|
@%p9 bra BB0_2;
|
|
|
|
ld.u8 %rs12, [%rd8+1];
|
|
and.b16 %rs14, %rs10, 255;
|
|
cvt.rn.f32.u16 %f288, %rs14;
|
|
div.rn.f32 %f289, %f288, 0f437F0000;
|
|
fma.rn.f32 %f290, %f289, 0f40000000, 0fBF800000;
|
|
cvt.rn.f32.u16 %f291, %rs12;
|
|
div.rn.f32 %f292, %f291, 0f437F0000;
|
|
fma.rn.f32 %f293, %f292, 0f40000000, 0fBF800000;
|
|
cvt.rn.f32.u16 %f294, %rs9;
|
|
div.rn.f32 %f295, %f294, 0f437F0000;
|
|
fma.rn.f32 %f296, %f295, 0f40000000, 0fBF800000;
|
|
mul.f32 %f297, %f293, %f293;
|
|
fma.rn.f32 %f298, %f290, %f290, %f297;
|
|
fma.rn.f32 %f299, %f296, %f296, %f298;
|
|
sqrt.rn.f32 %f300, %f299;
|
|
rcp.rn.f32 %f301, %f300;
|
|
mul.f32 %f1480, %f290, %f301;
|
|
mul.f32 %f1481, %f293, %f301;
|
|
mul.f32 %f1482, %f296, %f301;
|
|
|
|
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, %f1481, 0f00000000;
|
|
setp.eq.f32 %p11, %f1480, 0f00000000;
|
|
and.pred %p12, %p11, %p10;
|
|
setp.eq.f32 %p13, %f1482, 0f00000000;
|
|
and.pred %p14, %p12, %p13;
|
|
@%p14 bra BB0_121;
|
|
bra.uni BB0_3;
|
|
|
|
BB0_121:
|
|
ld.global.u32 %r272, [imageEnabled];
|
|
and.b32 %r221, %r272, 1;
|
|
setp.eq.b32 %p156, %r221, 1;
|
|
@!%p156 bra BB0_123;
|
|
bra.uni BB0_122;
|
|
|
|
BB0_122:
|
|
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 %r272, [imageEnabled];
|
|
|
|
BB0_123:
|
|
and.b32 %r224, %r272, 8;
|
|
setp.eq.s32 %p157, %r224, 0;
|
|
@%p157 bra BB0_125;
|
|
|
|
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 %f1437, 0f00000000;
|
|
cvt.rzi.u32.f32 %r227, %f1437;
|
|
cvt.u16.u32 %rs94, %r227;
|
|
mov.u16 %rs95, 0;
|
|
st.v2.u8 [%rd164], {%rs94, %rs95};
|
|
ld.global.u32 %r272, [imageEnabled];
|
|
|
|
BB0_125:
|
|
cvt.u64.u32 %rd6, %r2;
|
|
cvt.u64.u32 %rd7, %r3;
|
|
and.b32 %r228, %r272, 4;
|
|
setp.eq.s32 %p158, %r228, 0;
|
|
@%p158 bra BB0_129;
|
|
|
|
ld.global.u32 %r229, [additive];
|
|
setp.eq.s32 %p159, %r229, 0;
|
|
@%p159 bra BB0_128;
|
|
|
|
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 %f1438, %rs102;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1439, %rs103;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1440, %rs104;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd177), _rt_buffer_get_64, (%rd172, %r28, %r233, %rd6, %rd7, %rd13, %rd13);
|
|
// inline asm
|
|
add.f32 %f1441, %f1438, 0f00000000;
|
|
add.f32 %f1442, %f1439, 0f00000000;
|
|
add.f32 %f1443, %f1440, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs101, %f1443;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs100, %f1442;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs99, %f1441;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs106, 0;
|
|
st.v4.u16 [%rd177], {%rs99, %rs100, %rs101, %rs106};
|
|
bra.uni BB0_129;
|
|
|
|
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 %f318, %f7, 0f3456BF95;
|
|
mul.f32 %f319, %f8, 0f3456BF95;
|
|
mul.f32 %f320, %f9, 0f3456BF95;
|
|
abs.f32 %f321, %f1480;
|
|
div.rn.f32 %f322, %f318, %f321;
|
|
abs.f32 %f323, %f1481;
|
|
div.rn.f32 %f324, %f319, %f323;
|
|
abs.f32 %f325, %f1482;
|
|
div.rn.f32 %f326, %f320, %f325;
|
|
abs.f32 %f327, %f322;
|
|
abs.f32 %f328, %f324;
|
|
abs.f32 %f329, %f326;
|
|
mov.f32 %f330, 0f38D1B717;
|
|
max.f32 %f331, %f327, %f330;
|
|
max.f32 %f332, %f328, %f330;
|
|
max.f32 %f333, %f329, %f330;
|
|
fma.rn.f32 %f10, %f1480, %f331, %f7;
|
|
fma.rn.f32 %f11, %f1481, %f332, %f8;
|
|
fma.rn.f32 %f12, %f1482, %f333, %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 %f1483, 0f00000000;
|
|
mov.f32 %f1484, %f1483;
|
|
mov.f32 %f1485, %f1483;
|
|
mov.f32 %f1486, %f1483;
|
|
mov.f32 %f1487, %f1483;
|
|
mov.f32 %f1488, %f1483;
|
|
mov.f32 %f1489, %f1483;
|
|
mov.f32 %f1490, %f1483;
|
|
mov.f32 %f1491, %f1483;
|
|
mov.f32 %f1492, %f1483;
|
|
mov.f32 %f1493, %f1483;
|
|
mov.f32 %f1494, %f1483;
|
|
mov.f32 %f1495, %f1483;
|
|
mov.f32 %f1496, %f1483;
|
|
mov.f32 %f1497, %f1483;
|
|
mov.f32 %f1498, %f1483;
|
|
@%p15 bra BB0_57;
|
|
|
|
mov.f32 %f350, 0f40000000;
|
|
cvt.rzi.f32.f32 %f351, %f350;
|
|
add.f32 %f352, %f351, %f351;
|
|
mov.f32 %f353, 0f40800000;
|
|
sub.f32 %f354, %f353, %f352;
|
|
abs.f32 %f15, %f354;
|
|
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 %f349, 0f00000000;
|
|
mov.u32 %r264, 0;
|
|
abs.f32 %f547, %f16;
|
|
abs.f32 %f548, %f17;
|
|
max.f32 %f549, %f547, %f548;
|
|
abs.f32 %f550, %f18;
|
|
max.f32 %f551, %f549, %f550;
|
|
mov.f32 %f1483, %f349;
|
|
mov.f32 %f1484, %f349;
|
|
mov.f32 %f1485, %f349;
|
|
mov.f32 %f1486, %f349;
|
|
mov.f32 %f1487, %f349;
|
|
mov.f32 %f1488, %f349;
|
|
mov.f32 %f1489, %f349;
|
|
mov.f32 %f1490, %f349;
|
|
mov.f32 %f1491, %f349;
|
|
mov.f32 %f1492, %f349;
|
|
mov.f32 %f1493, %f349;
|
|
mov.f32 %f1494, %f349;
|
|
mov.f32 %f1495, %f349;
|
|
mov.f32 %f1496, %f349;
|
|
mov.f32 %f1497, %f349;
|
|
mov.f32 %f1498, %f349;
|
|
|
|
BB0_5:
|
|
cvt.u64.u32 %rd30, %r264;
|
|
// inline asm
|
|
call (%rd28), _rt_buffer_get_64, (%rd25, %r45, %r46, %rd30, %rd13, %rd13, %rd13);
|
|
// inline asm
|
|
ld.v4.f32 {%f357, %f358, %f359, %f360}, [%rd28+80];
|
|
ld.v4.f32 {%f361, %f362, %f363, %f364}, [%rd28+64];
|
|
ld.v4.f32 {%f365, %f366, %f367, %f368}, [%rd28+48];
|
|
ld.v4.f32 {%f369, %f1503, %f1504, %f372}, [%rd28+32];
|
|
ld.v4.f32 {%f373, %f374, %f375, %f376}, [%rd28+16];
|
|
ld.v4.f32 {%f377, %f378, %f379, %f380}, [%rd28];
|
|
mov.b32 %r6, %f360;
|
|
sub.f32 %f382, %f378, %f7;
|
|
sub.f32 %f383, %f379, %f8;
|
|
sub.f32 %f384, %f380, %f9;
|
|
mul.f32 %f385, %f383, %f383;
|
|
fma.rn.f32 %f386, %f382, %f382, %f385;
|
|
fma.rn.f32 %f387, %f384, %f384, %f386;
|
|
sqrt.rn.f32 %f63, %f387;
|
|
rcp.rn.f32 %f388, %f63;
|
|
mul.f32 %f64, %f382, %f388;
|
|
mul.f32 %f65, %f383, %f388;
|
|
mul.f32 %f66, %f384, %f388;
|
|
mul.f32 %f67, %f63, %f376;
|
|
abs.f32 %f68, %f67;
|
|
setp.lt.f32 %p16, %f68, 0f00800000;
|
|
mul.f32 %f389, %f68, 0f4B800000;
|
|
selp.f32 %f390, 0fC3170000, 0fC2FE0000, %p16;
|
|
selp.f32 %f391, %f389, %f68, %p16;
|
|
mov.b32 %r58, %f391;
|
|
and.b32 %r59, %r58, 8388607;
|
|
or.b32 %r60, %r59, 1065353216;
|
|
mov.b32 %f392, %r60;
|
|
shr.u32 %r61, %r58, 23;
|
|
cvt.rn.f32.u32 %f393, %r61;
|
|
add.f32 %f394, %f390, %f393;
|
|
setp.gt.f32 %p17, %f392, 0f3FB504F3;
|
|
mul.f32 %f395, %f392, 0f3F000000;
|
|
add.f32 %f396, %f394, 0f3F800000;
|
|
selp.f32 %f397, %f395, %f392, %p17;
|
|
selp.f32 %f398, %f396, %f394, %p17;
|
|
add.f32 %f399, %f397, 0fBF800000;
|
|
add.f32 %f356, %f397, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f355,%f356;
|
|
// inline asm
|
|
add.f32 %f400, %f399, %f399;
|
|
mul.f32 %f401, %f355, %f400;
|
|
mul.f32 %f402, %f401, %f401;
|
|
mov.f32 %f403, 0f3C4CAF63;
|
|
mov.f32 %f404, 0f3B18F0FE;
|
|
fma.rn.f32 %f405, %f404, %f402, %f403;
|
|
mov.f32 %f406, 0f3DAAAABD;
|
|
fma.rn.f32 %f407, %f405, %f402, %f406;
|
|
mul.rn.f32 %f408, %f407, %f402;
|
|
mul.rn.f32 %f409, %f408, %f401;
|
|
sub.f32 %f410, %f399, %f401;
|
|
neg.f32 %f411, %f401;
|
|
add.f32 %f412, %f410, %f410;
|
|
fma.rn.f32 %f413, %f411, %f399, %f412;
|
|
mul.rn.f32 %f414, %f355, %f413;
|
|
add.f32 %f415, %f409, %f401;
|
|
sub.f32 %f416, %f401, %f415;
|
|
add.f32 %f417, %f409, %f416;
|
|
add.f32 %f418, %f414, %f417;
|
|
add.f32 %f419, %f415, %f418;
|
|
sub.f32 %f420, %f415, %f419;
|
|
add.f32 %f421, %f418, %f420;
|
|
mov.f32 %f422, 0f3F317200;
|
|
mul.rn.f32 %f423, %f398, %f422;
|
|
mov.f32 %f424, 0f35BFBE8E;
|
|
mul.rn.f32 %f425, %f398, %f424;
|
|
add.f32 %f426, %f423, %f419;
|
|
sub.f32 %f427, %f423, %f426;
|
|
add.f32 %f428, %f419, %f427;
|
|
add.f32 %f429, %f421, %f428;
|
|
add.f32 %f430, %f425, %f429;
|
|
add.f32 %f431, %f426, %f430;
|
|
sub.f32 %f432, %f426, %f431;
|
|
add.f32 %f433, %f430, %f432;
|
|
mul.rn.f32 %f69, %f353, %f431;
|
|
neg.f32 %f435, %f69;
|
|
fma.rn.f32 %f436, %f353, %f431, %f435;
|
|
fma.rn.f32 %f437, %f353, %f433, %f436;
|
|
fma.rn.f32 %f70, %f349, %f431, %f437;
|
|
add.rn.f32 %f71, %f69, %f70;
|
|
mov.b32 %r62, %f71;
|
|
setp.eq.s32 %p1, %r62, 1118925336;
|
|
add.s32 %r63, %r62, -1;
|
|
mov.b32 %f439, %r63;
|
|
selp.f32 %f440, %f439, %f71, %p1;
|
|
mul.f32 %f441, %f440, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f442, %f441;
|
|
mov.f32 %f443, 0fBF317200;
|
|
fma.rn.f32 %f444, %f442, %f443, %f440;
|
|
mov.f32 %f445, 0fB5BFBE8E;
|
|
fma.rn.f32 %f446, %f442, %f445, %f444;
|
|
mul.f32 %f447, %f446, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f448, %f447;
|
|
add.f32 %f449, %f442, 0f00000000;
|
|
ex2.approx.f32 %f450, %f449;
|
|
mul.f32 %f451, %f448, %f450;
|
|
setp.lt.f32 %p18, %f440, 0fC2D20000;
|
|
selp.f32 %f452, 0f00000000, %f451, %p18;
|
|
setp.gt.f32 %p19, %f440, 0f42D20000;
|
|
selp.f32 %f1499, 0f7F800000, %f452, %p19;
|
|
setp.eq.f32 %p20, %f1499, 0f7F800000;
|
|
@%p20 bra BB0_7;
|
|
|
|
neg.f32 %f453, %f71;
|
|
add.rn.f32 %f454, %f69, %f453;
|
|
add.rn.f32 %f455, %f454, %f70;
|
|
add.f32 %f456, %f455, 0f37000000;
|
|
selp.f32 %f457, %f456, %f455, %p1;
|
|
fma.rn.f32 %f1499, %f1499, %f457, %f1499;
|
|
|
|
BB0_7:
|
|
setp.lt.f32 %p21, %f67, 0f00000000;
|
|
setp.eq.f32 %p22, %f15, 0f3F800000;
|
|
and.pred %p2, %p21, %p22;
|
|
mov.b32 %r64, %f1499;
|
|
xor.b32 %r65, %r64, -2147483648;
|
|
mov.b32 %f458, %r65;
|
|
selp.f32 %f1501, %f458, %f1499, %p2;
|
|
setp.eq.f32 %p23, %f67, 0f00000000;
|
|
@%p23 bra BB0_10;
|
|
bra.uni BB0_8;
|
|
|
|
BB0_10:
|
|
add.f32 %f461, %f67, %f67;
|
|
selp.f32 %f1501, %f461, 0f00000000, %p22;
|
|
bra.uni BB0_11;
|
|
|
|
BB0_8:
|
|
setp.geu.f32 %p24, %f67, 0f00000000;
|
|
@%p24 bra BB0_11;
|
|
|
|
cvt.rzi.f32.f32 %f460, %f353;
|
|
setp.neu.f32 %p25, %f460, 0f40800000;
|
|
selp.f32 %f1501, 0f7FFFFFFF, %f1501, %p25;
|
|
|
|
BB0_11:
|
|
add.f32 %f462, %f68, 0f40800000;
|
|
mov.b32 %r66, %f462;
|
|
setp.lt.s32 %p27, %r66, 2139095040;
|
|
@%p27 bra BB0_16;
|
|
|
|
setp.gtu.f32 %p28, %f68, 0f7F800000;
|
|
@%p28 bra BB0_15;
|
|
bra.uni BB0_13;
|
|
|
|
BB0_15:
|
|
add.f32 %f1501, %f67, 0f40800000;
|
|
bra.uni BB0_16;
|
|
|
|
BB0_13:
|
|
setp.neu.f32 %p29, %f68, 0f7F800000;
|
|
@%p29 bra BB0_16;
|
|
|
|
selp.f32 %f1501, 0fFF800000, 0f7F800000, %p2;
|
|
|
|
BB0_16:
|
|
mul.f32 %f463, %f63, %f374;
|
|
mov.f32 %f1526, 0f3F800000;
|
|
sub.f32 %f465, %f1526, %f1501;
|
|
setp.eq.f32 %p30, %f67, 0f3F800000;
|
|
selp.f32 %f466, 0f00000000, %f465, %p30;
|
|
cvt.sat.f32.f32 %f467, %f466;
|
|
fma.rn.f32 %f468, %f463, %f463, %f375;
|
|
div.rn.f32 %f97, %f467, %f468;
|
|
mul.f32 %f469, %f1481, %f65;
|
|
fma.rn.f32 %f470, %f1480, %f64, %f469;
|
|
fma.rn.f32 %f471, %f1482, %f66, %f470;
|
|
ld.global.u32 %r67, [ignoreNormal];
|
|
setp.eq.s32 %p31, %r67, 0;
|
|
selp.f32 %f98, %f471, 0f3F800000, %p31;
|
|
mul.f32 %f472, %f98, 0f40800000;
|
|
cvt.sat.f32.f32 %f99, %f472;
|
|
setp.eq.f32 %p32, %f377, 0f3F800000;
|
|
@%p32 bra BB0_23;
|
|
bra.uni BB0_17;
|
|
|
|
BB0_23:
|
|
setp.leu.f32 %p36, %f372, 0f00000000;
|
|
@%p36 bra BB0_19;
|
|
|
|
mul.f32 %f503, %f357, %f64;
|
|
mul.f32 %f504, %f358, %f65;
|
|
neg.f32 %f505, %f504;
|
|
sub.f32 %f506, %f505, %f503;
|
|
mul.f32 %f507, %f359, %f66;
|
|
sub.f32 %f508, %f506, %f507;
|
|
setp.gt.f32 %p37, %f508, 0f00000000;
|
|
selp.f32 %f509, 0f3F800000, 0f00000000, %p37;
|
|
mul.f32 %f510, %f366, %f65;
|
|
fma.rn.f32 %f511, %f365, %f64, %f510;
|
|
mul.f32 %f512, %f362, %f65;
|
|
fma.rn.f32 %f513, %f361, %f64, %f512;
|
|
fma.rn.f32 %f514, %f367, %f66, %f511;
|
|
fma.rn.f32 %f515, %f363, %f66, %f513;
|
|
fma.rn.f32 %f516, %f368, %f514, 0f3F000000;
|
|
mov.f32 %f517, 0f3F800000;
|
|
sub.f32 %f499, %f517, %f516;
|
|
fma.rn.f32 %f500, %f368, %f515, 0f3F000000;
|
|
cvt.rzi.s32.f32 %r71, %f372;
|
|
mov.f32 %f502, 0f00000000;
|
|
// inline asm
|
|
call (%f495, %f496, %f497, %f498), _rt_texture_get_f_id, (%r71, %r28, %f499, %f500, %f502, %f502);
|
|
// inline asm
|
|
mul.f32 %f518, %f509, %f495;
|
|
mul.f32 %f519, %f509, %f496;
|
|
mul.f32 %f520, %f509, %f497;
|
|
mul.f32 %f1502, %f369, %f518;
|
|
mul.f32 %f1503, %f1503, %f519;
|
|
mul.f32 %f1504, %f1504, %f520;
|
|
bra.uni BB0_25;
|
|
|
|
BB0_17:
|
|
setp.eq.f32 %p33, %f377, 0f40000000;
|
|
@%p33 bra BB0_21;
|
|
bra.uni BB0_18;
|
|
|
|
BB0_21:
|
|
setp.leu.f32 %p35, %f372, 0f00000000;
|
|
@%p35 bra BB0_19;
|
|
|
|
mul.f32 %f489, %f366, %f65;
|
|
fma.rn.f32 %f490, %f365, %f64, %f489;
|
|
mul.f32 %f491, %f362, %f65;
|
|
fma.rn.f32 %f492, %f361, %f64, %f491;
|
|
mul.f32 %f493, %f358, %f65;
|
|
fma.rn.f32 %f494, %f357, %f64, %f493;
|
|
fma.rn.f32 %f486, %f367, %f66, %f490;
|
|
fma.rn.f32 %f487, %f363, %f66, %f492;
|
|
fma.rn.f32 %f488, %f359, %f66, %f494;
|
|
cvt.rzi.s32.f32 %r68, %f372;
|
|
mov.u32 %r69, 6;
|
|
mov.u32 %r70, 0;
|
|
// inline asm
|
|
call (%f482, %f483, %f484, %f485), _rt_texture_get_base_id, (%r68, %r69, %f486, %f487, %f488, %r70);
|
|
// inline asm
|
|
mul.f32 %f1502, %f369, %f482;
|
|
mul.f32 %f1503, %f1503, %f483;
|
|
mul.f32 %f1504, %f1504, %f484;
|
|
bra.uni BB0_25;
|
|
|
|
BB0_18:
|
|
setp.neu.f32 %p34, %f377, 0f40800000;
|
|
@%p34 bra BB0_19;
|
|
|
|
mul.f32 %f473, %f357, %f64;
|
|
mul.f32 %f474, %f358, %f65;
|
|
neg.f32 %f475, %f474;
|
|
sub.f32 %f476, %f475, %f473;
|
|
mul.f32 %f477, %f359, %f66;
|
|
sub.f32 %f478, %f476, %f477;
|
|
fma.rn.f32 %f479, %f372, %f478, %f368;
|
|
cvt.sat.f32.f32 %f480, %f479;
|
|
mul.f32 %f481, %f480, %f480;
|
|
mul.f32 %f1505, %f97, %f481;
|
|
mov.f32 %f1502, %f369;
|
|
bra.uni BB0_26;
|
|
|
|
BB0_19:
|
|
mov.f32 %f1502, %f369;
|
|
|
|
BB0_25:
|
|
mov.f32 %f1505, %f97;
|
|
|
|
BB0_26:
|
|
max.f32 %f536, %f1502, %f1503;
|
|
max.f32 %f537, %f536, %f1504;
|
|
mul.f32 %f538, %f99, %f1505;
|
|
mul.f32 %f539, %f538, %f537;
|
|
setp.lt.f32 %p39, %f539, 0f3727C5AC;
|
|
mov.pred %p164, -1;
|
|
mov.f32 %f1506, 0f00000000;
|
|
mov.f32 %f1507, %f1506;
|
|
mov.f32 %f1508, %f1506;
|
|
mov.f32 %f129, %f1506;
|
|
mov.f32 %f1510, %f1506;
|
|
mov.f32 %f1511, %f1506;
|
|
mov.f32 %f132, %f1506;
|
|
mov.f32 %f1513, %f1506;
|
|
mov.f32 %f1514, %f1506;
|
|
mov.f32 %f135, %f1506;
|
|
mov.f32 %f1516, %f1506;
|
|
mov.f32 %f1517, %f1506;
|
|
mov.f32 %f138, %f1506;
|
|
mov.f32 %f1519, %f1506;
|
|
mov.f32 %f1520, %f1506;
|
|
@%p39 bra BB0_28;
|
|
|
|
cvt.sat.f32.f32 %f540, %f98;
|
|
mul.f32 %f541, %f1505, %f540;
|
|
mul.f32 %f1506, %f1502, %f541;
|
|
mul.f32 %f1507, %f1503, %f541;
|
|
mul.f32 %f1508, %f1504, %f541;
|
|
mul.f32 %f542, %f1505, 0f3E800000;
|
|
mul.f32 %f543, %f99, %f542;
|
|
mul.f32 %f129, %f1502, %f543;
|
|
mul.f32 %f1510, %f1503, %f543;
|
|
mul.f32 %f1511, %f1504, %f543;
|
|
mul.f32 %f132, %f64, %f129;
|
|
mul.f32 %f1513, %f64, %f1510;
|
|
mul.f32 %f1514, %f64, %f1511;
|
|
mul.f32 %f135, %f65, %f129;
|
|
mul.f32 %f1516, %f65, %f1510;
|
|
mul.f32 %f1517, %f65, %f1511;
|
|
mul.f32 %f138, %f66, %f129;
|
|
mul.f32 %f1519, %f66, %f1510;
|
|
mul.f32 %f1520, %f66, %f1511;
|
|
mov.pred %p164, 0;
|
|
|
|
BB0_28:
|
|
@%p164 bra BB0_56;
|
|
|
|
setp.eq.s32 %p41, %r6, 0;
|
|
mov.u16 %rs161, 0;
|
|
@%p41 bra BB0_40;
|
|
|
|
abs.s32 %r8, %r6;
|
|
mov.f32 %f1525, 0f00000000;
|
|
setp.lt.s32 %p42, %r8, 1;
|
|
@%p42 bra BB0_39;
|
|
|
|
max.f32 %f142, %f551, %f330;
|
|
and.b32 %r9, %r8, 3;
|
|
setp.eq.s32 %p43, %r9, 0;
|
|
add.u64 %rd35, %SP, 0;
|
|
cvta.to.local.u64 %rd2, %rd35;
|
|
mov.f32 %f1525, 0f00000000;
|
|
mov.u32 %r268, 0;
|
|
@%p43 bra BB0_37;
|
|
|
|
setp.eq.s32 %p44, %r9, 1;
|
|
mov.f32 %f1522, 0f00000000;
|
|
mov.u32 %r266, 0;
|
|
@%p44 bra BB0_36;
|
|
|
|
setp.eq.s32 %p45, %r9, 2;
|
|
mov.f32 %f1521, 0f00000000;
|
|
mov.u32 %r265, 0;
|
|
@%p45 bra BB0_35;
|
|
|
|
cvt.rmi.f32.f32 %f563, %f19;
|
|
sub.f32 %f564, %f19, %f563;
|
|
cvt.rmi.f32.f32 %f565, %f20;
|
|
sub.f32 %f566, %f20, %f565;
|
|
add.f32 %f567, %f566, 0f420551EC;
|
|
add.f32 %f568, %f564, 0f420551EC;
|
|
mul.f32 %f569, %f564, %f567;
|
|
fma.rn.f32 %f570, %f566, %f568, %f569;
|
|
add.f32 %f571, %f570, 0f00000000;
|
|
add.f32 %f572, %f564, %f571;
|
|
add.f32 %f573, %f566, %f571;
|
|
add.f32 %f574, %f572, %f573;
|
|
mul.f32 %f575, %f571, %f574;
|
|
cvt.rmi.f32.f32 %f576, %f575;
|
|
sub.f32 %f577, %f575, %f576;
|
|
add.f32 %f578, %f572, %f572;
|
|
mul.f32 %f579, %f573, %f578;
|
|
cvt.rmi.f32.f32 %f580, %f579;
|
|
sub.f32 %f581, %f579, %f580;
|
|
mul.f32 %f582, %f572, %f574;
|
|
cvt.rmi.f32.f32 %f583, %f582;
|
|
sub.f32 %f584, %f582, %f583;
|
|
fma.rn.f32 %f585, %f577, 0f40000000, 0fBF800000;
|
|
fma.rn.f32 %f586, %f581, 0f40000000, 0fBF800000;
|
|
fma.rn.f32 %f587, %f584, 0f40000000, 0fBF800000;
|
|
fma.rn.f32 %f588, %f373, %f585, %f378;
|
|
fma.rn.f32 %f589, %f373, %f586, %f379;
|
|
fma.rn.f32 %f590, %f373, %f587, %f380;
|
|
sub.f32 %f591, %f588, %f7;
|
|
sub.f32 %f592, %f589, %f8;
|
|
sub.f32 %f593, %f590, %f9;
|
|
mul.f32 %f594, %f592, %f592;
|
|
fma.rn.f32 %f595, %f591, %f591, %f594;
|
|
fma.rn.f32 %f596, %f593, %f593, %f595;
|
|
sqrt.rn.f32 %f562, %f596;
|
|
rcp.rn.f32 %f597, %f562;
|
|
mul.f32 %f558, %f597, %f591;
|
|
mul.f32 %f559, %f597, %f592;
|
|
mul.f32 %f560, %f597, %f593;
|
|
ld.global.u32 %r80, [imageEnabled];
|
|
and.b32 %r81, %r80, 32;
|
|
setp.eq.s32 %p46, %r81, 0;
|
|
selp.f32 %f598, 0f3F800000, 0f41200000, %p46;
|
|
mul.f32 %f561, %f598, %f142;
|
|
mov.u32 %r82, 1065353216;
|
|
st.local.u32 [%rd2], %r82;
|
|
ld.global.u32 %r76, [root];
|
|
// inline asm
|
|
call _rt_trace_64, (%r76, %f10, %f11, %f12, %f558, %f559, %f560, %r45, %f561, %f562, %rd35, %r29);
|
|
// inline asm
|
|
ld.local.f32 %f599, [%rd2];
|
|
add.f32 %f1521, %f599, 0f00000000;
|
|
mov.u32 %r265, %r45;
|
|
|
|
BB0_35:
|
|
cvt.rn.f32.s32 %f608, %r265;
|
|
add.f32 %f609, %f13, %f608;
|
|
sub.f32 %f610, %f14, %f608;
|
|
mul.f32 %f611, %f609, 0f3DD32618;
|
|
cvt.rmi.f32.f32 %f612, %f611;
|
|
sub.f32 %f613, %f611, %f612;
|
|
mul.f32 %f614, %f610, 0f3DD2F1AA;
|
|
cvt.rmi.f32.f32 %f615, %f614;
|
|
sub.f32 %f616, %f614, %f615;
|
|
mul.f32 %f617, %f608, 0f3DC74539;
|
|
cvt.rmi.f32.f32 %f618, %f617;
|
|
sub.f32 %f619, %f617, %f618;
|
|
add.f32 %f620, %f616, 0f420551EC;
|
|
add.f32 %f621, %f613, 0f420551EC;
|
|
add.f32 %f622, %f619, 0f420551EC;
|
|
mul.f32 %f623, %f613, %f620;
|
|
fma.rn.f32 %f624, %f616, %f621, %f623;
|
|
fma.rn.f32 %f625, %f619, %f622, %f624;
|
|
add.f32 %f626, %f613, %f625;
|
|
add.f32 %f627, %f616, %f625;
|
|
add.f32 %f628, %f619, %f625;
|
|
add.f32 %f629, %f626, %f627;
|
|
mul.f32 %f630, %f628, %f629;
|
|
cvt.rmi.f32.f32 %f631, %f630;
|
|
sub.f32 %f632, %f630, %f631;
|
|
add.f32 %f633, %f626, %f626;
|
|
mul.f32 %f634, %f627, %f633;
|
|
cvt.rmi.f32.f32 %f635, %f634;
|
|
sub.f32 %f636, %f634, %f635;
|
|
mul.f32 %f637, %f626, %f629;
|
|
cvt.rmi.f32.f32 %f638, %f637;
|
|
sub.f32 %f639, %f637, %f638;
|
|
fma.rn.f32 %f640, %f632, 0f40000000, 0fBF800000;
|
|
fma.rn.f32 %f641, %f636, 0f40000000, 0fBF800000;
|
|
fma.rn.f32 %f642, %f639, 0f40000000, 0fBF800000;
|
|
fma.rn.f32 %f643, %f373, %f640, %f378;
|
|
fma.rn.f32 %f644, %f373, %f641, %f379;
|
|
fma.rn.f32 %f645, %f373, %f642, %f380;
|
|
sub.f32 %f646, %f643, %f7;
|
|
sub.f32 %f647, %f644, %f8;
|
|
sub.f32 %f648, %f645, %f9;
|
|
mul.f32 %f649, %f647, %f647;
|
|
fma.rn.f32 %f650, %f646, %f646, %f649;
|
|
fma.rn.f32 %f651, %f648, %f648, %f650;
|
|
sqrt.rn.f32 %f607, %f651;
|
|
rcp.rn.f32 %f652, %f607;
|
|
mul.f32 %f603, %f652, %f646;
|
|
mul.f32 %f604, %f652, %f647;
|
|
mul.f32 %f605, %f652, %f648;
|
|
ld.global.u32 %r86, [imageEnabled];
|
|
and.b32 %r87, %r86, 32;
|
|
setp.eq.s32 %p47, %r87, 0;
|
|
selp.f32 %f653, 0f3F800000, 0f41200000, %p47;
|
|
mul.f32 %f606, %f653, %f142;
|
|
mov.u32 %r88, 1065353216;
|
|
st.local.u32 [%rd2], %r88;
|
|
ld.global.u32 %r83, [root];
|
|
// inline asm
|
|
call _rt_trace_64, (%r83, %f10, %f11, %f12, %f603, %f604, %f605, %r45, %f606, %f607, %rd35, %r29);
|
|
// inline asm
|
|
ld.local.f32 %f654, [%rd2];
|
|
add.f32 %f1522, %f1521, %f654;
|
|
add.s32 %r266, %r265, 1;
|
|
|
|
BB0_36:
|
|
cvt.rn.f32.s32 %f663, %r266;
|
|
add.f32 %f664, %f13, %f663;
|
|
sub.f32 %f665, %f14, %f663;
|
|
mul.f32 %f666, %f664, 0f3DD32618;
|
|
cvt.rmi.f32.f32 %f667, %f666;
|
|
sub.f32 %f668, %f666, %f667;
|
|
mul.f32 %f669, %f665, 0f3DD2F1AA;
|
|
cvt.rmi.f32.f32 %f670, %f669;
|
|
sub.f32 %f671, %f669, %f670;
|
|
mul.f32 %f672, %f663, 0f3DC74539;
|
|
cvt.rmi.f32.f32 %f673, %f672;
|
|
sub.f32 %f674, %f672, %f673;
|
|
add.f32 %f675, %f671, 0f420551EC;
|
|
add.f32 %f676, %f668, 0f420551EC;
|
|
add.f32 %f677, %f674, 0f420551EC;
|
|
mul.f32 %f678, %f668, %f675;
|
|
fma.rn.f32 %f679, %f671, %f676, %f678;
|
|
fma.rn.f32 %f680, %f674, %f677, %f679;
|
|
add.f32 %f681, %f668, %f680;
|
|
add.f32 %f682, %f671, %f680;
|
|
add.f32 %f683, %f674, %f680;
|
|
add.f32 %f684, %f681, %f682;
|
|
mul.f32 %f685, %f683, %f684;
|
|
cvt.rmi.f32.f32 %f686, %f685;
|
|
sub.f32 %f687, %f685, %f686;
|
|
add.f32 %f688, %f681, %f681;
|
|
mul.f32 %f689, %f682, %f688;
|
|
cvt.rmi.f32.f32 %f690, %f689;
|
|
sub.f32 %f691, %f689, %f690;
|
|
mul.f32 %f692, %f681, %f684;
|
|
cvt.rmi.f32.f32 %f693, %f692;
|
|
sub.f32 %f694, %f692, %f693;
|
|
fma.rn.f32 %f695, %f687, 0f40000000, 0fBF800000;
|
|
fma.rn.f32 %f696, %f691, 0f40000000, 0fBF800000;
|
|
fma.rn.f32 %f697, %f694, 0f40000000, 0fBF800000;
|
|
fma.rn.f32 %f698, %f373, %f695, %f378;
|
|
fma.rn.f32 %f699, %f373, %f696, %f379;
|
|
fma.rn.f32 %f700, %f373, %f697, %f380;
|
|
sub.f32 %f701, %f698, %f7;
|
|
sub.f32 %f702, %f699, %f8;
|
|
sub.f32 %f703, %f700, %f9;
|
|
mul.f32 %f704, %f702, %f702;
|
|
fma.rn.f32 %f705, %f701, %f701, %f704;
|
|
fma.rn.f32 %f706, %f703, %f703, %f705;
|
|
sqrt.rn.f32 %f662, %f706;
|
|
rcp.rn.f32 %f707, %f662;
|
|
mul.f32 %f658, %f707, %f701;
|
|
mul.f32 %f659, %f707, %f702;
|
|
mul.f32 %f660, %f707, %f703;
|
|
ld.global.u32 %r92, [imageEnabled];
|
|
and.b32 %r93, %r92, 32;
|
|
setp.eq.s32 %p48, %r93, 0;
|
|
selp.f32 %f708, 0f3F800000, 0f41200000, %p48;
|
|
mul.f32 %f661, %f708, %f142;
|
|
mov.u32 %r94, 1065353216;
|
|
st.local.u32 [%rd2], %r94;
|
|
ld.global.u32 %r89, [root];
|
|
mov.u32 %r90, 1;
|
|
// inline asm
|
|
call _rt_trace_64, (%r89, %f10, %f11, %f12, %f658, %f659, %f660, %r90, %f661, %f662, %rd35, %r29);
|
|
// inline asm
|
|
ld.local.f32 %f709, [%rd2];
|
|
add.f32 %f1525, %f1522, %f709;
|
|
add.s32 %r268, %r266, 1;
|
|
|
|
BB0_37:
|
|
setp.lt.u32 %p49, %r8, 4;
|
|
@%p49 bra BB0_39;
|
|
|
|
BB0_38:
|
|
cvt.rn.f32.s32 %f742, %r268;
|
|
add.f32 %f743, %f13, %f742;
|
|
sub.f32 %f744, %f14, %f742;
|
|
mul.f32 %f745, %f743, 0f3DD32618;
|
|
cvt.rmi.f32.f32 %f746, %f745;
|
|
sub.f32 %f747, %f745, %f746;
|
|
mul.f32 %f748, %f744, 0f3DD2F1AA;
|
|
cvt.rmi.f32.f32 %f749, %f748;
|
|
sub.f32 %f750, %f748, %f749;
|
|
mul.f32 %f751, %f742, 0f3DC74539;
|
|
cvt.rmi.f32.f32 %f752, %f751;
|
|
sub.f32 %f753, %f751, %f752;
|
|
add.f32 %f754, %f750, 0f420551EC;
|
|
add.f32 %f755, %f747, 0f420551EC;
|
|
add.f32 %f756, %f753, 0f420551EC;
|
|
mul.f32 %f757, %f747, %f754;
|
|
fma.rn.f32 %f758, %f750, %f755, %f757;
|
|
fma.rn.f32 %f759, %f753, %f756, %f758;
|
|
add.f32 %f760, %f747, %f759;
|
|
add.f32 %f761, %f750, %f759;
|
|
add.f32 %f762, %f753, %f759;
|
|
add.f32 %f763, %f760, %f761;
|
|
mul.f32 %f764, %f762, %f763;
|
|
cvt.rmi.f32.f32 %f765, %f764;
|
|
sub.f32 %f766, %f764, %f765;
|
|
add.f32 %f767, %f760, %f760;
|
|
mul.f32 %f768, %f761, %f767;
|
|
cvt.rmi.f32.f32 %f769, %f768;
|
|
sub.f32 %f770, %f768, %f769;
|
|
mul.f32 %f771, %f760, %f763;
|
|
cvt.rmi.f32.f32 %f772, %f771;
|
|
sub.f32 %f773, %f771, %f772;
|
|
fma.rn.f32 %f774, %f766, 0f40000000, 0fBF800000;
|
|
fma.rn.f32 %f775, %f770, 0f40000000, 0fBF800000;
|
|
fma.rn.f32 %f776, %f773, 0f40000000, 0fBF800000;
|
|
fma.rn.f32 %f777, %f373, %f774, %f378;
|
|
fma.rn.f32 %f778, %f373, %f775, %f379;
|
|
fma.rn.f32 %f779, %f373, %f776, %f380;
|
|
sub.f32 %f780, %f777, %f7;
|
|
sub.f32 %f781, %f778, %f8;
|
|
sub.f32 %f782, %f779, %f9;
|
|
mul.f32 %f783, %f781, %f781;
|
|
fma.rn.f32 %f784, %f780, %f780, %f783;
|
|
fma.rn.f32 %f785, %f782, %f782, %f784;
|
|
sqrt.rn.f32 %f717, %f785;
|
|
rcp.rn.f32 %f786, %f717;
|
|
mul.f32 %f713, %f786, %f780;
|
|
mul.f32 %f714, %f786, %f781;
|
|
mul.f32 %f715, %f786, %f782;
|
|
ld.global.u32 %r107, [imageEnabled];
|
|
and.b32 %r108, %r107, 32;
|
|
setp.eq.s32 %p50, %r108, 0;
|
|
selp.f32 %f787, 0f3F800000, 0f41200000, %p50;
|
|
mul.f32 %f716, %f787, %f142;
|
|
mov.u32 %r109, 1065353216;
|
|
st.local.u32 [%rd2], %r109;
|
|
ld.global.u32 %r95, [root];
|
|
mov.u32 %r105, 1;
|
|
// inline asm
|
|
call _rt_trace_64, (%r95, %f10, %f11, %f12, %f713, %f714, %f715, %r105, %f716, %f717, %rd35, %r29);
|
|
// inline asm
|
|
ld.local.f32 %f788, [%rd2];
|
|
add.f32 %f789, %f1525, %f788;
|
|
add.s32 %r110, %r268, 1;
|
|
cvt.rn.f32.s32 %f790, %r110;
|
|
add.f32 %f791, %f13, %f790;
|
|
sub.f32 %f792, %f14, %f790;
|
|
mul.f32 %f793, %f791, 0f3DD32618;
|
|
cvt.rmi.f32.f32 %f794, %f793;
|
|
sub.f32 %f795, %f793, %f794;
|
|
mul.f32 %f796, %f792, 0f3DD2F1AA;
|
|
cvt.rmi.f32.f32 %f797, %f796;
|
|
sub.f32 %f798, %f796, %f797;
|
|
mul.f32 %f799, %f790, 0f3DC74539;
|
|
cvt.rmi.f32.f32 %f800, %f799;
|
|
sub.f32 %f801, %f799, %f800;
|
|
add.f32 %f802, %f798, 0f420551EC;
|
|
add.f32 %f803, %f795, 0f420551EC;
|
|
add.f32 %f804, %f801, 0f420551EC;
|
|
mul.f32 %f805, %f795, %f802;
|
|
fma.rn.f32 %f806, %f798, %f803, %f805;
|
|
fma.rn.f32 %f807, %f801, %f804, %f806;
|
|
add.f32 %f808, %f795, %f807;
|
|
add.f32 %f809, %f798, %f807;
|
|
add.f32 %f810, %f801, %f807;
|
|
add.f32 %f811, %f808, %f809;
|
|
mul.f32 %f812, %f810, %f811;
|
|
cvt.rmi.f32.f32 %f813, %f812;
|
|
sub.f32 %f814, %f812, %f813;
|
|
add.f32 %f815, %f808, %f808;
|
|
mul.f32 %f816, %f809, %f815;
|
|
cvt.rmi.f32.f32 %f817, %f816;
|
|
sub.f32 %f818, %f816, %f817;
|
|
mul.f32 %f819, %f808, %f811;
|
|
cvt.rmi.f32.f32 %f820, %f819;
|
|
sub.f32 %f821, %f819, %f820;
|
|
fma.rn.f32 %f822, %f814, 0f40000000, 0fBF800000;
|
|
fma.rn.f32 %f823, %f818, 0f40000000, 0fBF800000;
|
|
fma.rn.f32 %f824, %f821, 0f40000000, 0fBF800000;
|
|
fma.rn.f32 %f825, %f373, %f822, %f378;
|
|
fma.rn.f32 %f826, %f373, %f823, %f379;
|
|
fma.rn.f32 %f827, %f373, %f824, %f380;
|
|
sub.f32 %f828, %f825, %f7;
|
|
sub.f32 %f829, %f826, %f8;
|
|
sub.f32 %f830, %f827, %f9;
|
|
mul.f32 %f831, %f829, %f829;
|
|
fma.rn.f32 %f832, %f828, %f828, %f831;
|
|
fma.rn.f32 %f833, %f830, %f830, %f832;
|
|
sqrt.rn.f32 %f725, %f833;
|
|
rcp.rn.f32 %f834, %f725;
|
|
mul.f32 %f721, %f834, %f828;
|
|
mul.f32 %f722, %f834, %f829;
|
|
mul.f32 %f723, %f834, %f830;
|
|
ld.global.u32 %r111, [imageEnabled];
|
|
and.b32 %r112, %r111, 32;
|
|
setp.eq.s32 %p51, %r112, 0;
|
|
selp.f32 %f835, 0f3F800000, 0f41200000, %p51;
|
|
mul.f32 %f724, %f835, %f142;
|
|
st.local.u32 [%rd2], %r109;
|
|
ld.global.u32 %r98, [root];
|
|
// inline asm
|
|
call _rt_trace_64, (%r98, %f10, %f11, %f12, %f721, %f722, %f723, %r105, %f724, %f725, %rd35, %r29);
|
|
// inline asm
|
|
ld.local.f32 %f836, [%rd2];
|
|
add.f32 %f837, %f789, %f836;
|
|
add.s32 %r113, %r268, 2;
|
|
cvt.rn.f32.s32 %f838, %r113;
|
|
add.f32 %f839, %f13, %f838;
|
|
sub.f32 %f840, %f14, %f838;
|
|
mul.f32 %f841, %f839, 0f3DD32618;
|
|
cvt.rmi.f32.f32 %f842, %f841;
|
|
sub.f32 %f843, %f841, %f842;
|
|
mul.f32 %f844, %f840, 0f3DD2F1AA;
|
|
cvt.rmi.f32.f32 %f845, %f844;
|
|
sub.f32 %f846, %f844, %f845;
|
|
mul.f32 %f847, %f838, 0f3DC74539;
|
|
cvt.rmi.f32.f32 %f848, %f847;
|
|
sub.f32 %f849, %f847, %f848;
|
|
add.f32 %f850, %f846, 0f420551EC;
|
|
add.f32 %f851, %f843, 0f420551EC;
|
|
add.f32 %f852, %f849, 0f420551EC;
|
|
mul.f32 %f853, %f843, %f850;
|
|
fma.rn.f32 %f854, %f846, %f851, %f853;
|
|
fma.rn.f32 %f855, %f849, %f852, %f854;
|
|
add.f32 %f856, %f843, %f855;
|
|
add.f32 %f857, %f846, %f855;
|
|
add.f32 %f858, %f849, %f855;
|
|
add.f32 %f859, %f856, %f857;
|
|
mul.f32 %f860, %f858, %f859;
|
|
cvt.rmi.f32.f32 %f861, %f860;
|
|
sub.f32 %f862, %f860, %f861;
|
|
add.f32 %f863, %f856, %f856;
|
|
mul.f32 %f864, %f857, %f863;
|
|
cvt.rmi.f32.f32 %f865, %f864;
|
|
sub.f32 %f866, %f864, %f865;
|
|
mul.f32 %f867, %f856, %f859;
|
|
cvt.rmi.f32.f32 %f868, %f867;
|
|
sub.f32 %f869, %f867, %f868;
|
|
fma.rn.f32 %f870, %f862, 0f40000000, 0fBF800000;
|
|
fma.rn.f32 %f871, %f866, 0f40000000, 0fBF800000;
|
|
fma.rn.f32 %f872, %f869, 0f40000000, 0fBF800000;
|
|
fma.rn.f32 %f873, %f373, %f870, %f378;
|
|
fma.rn.f32 %f874, %f373, %f871, %f379;
|
|
fma.rn.f32 %f875, %f373, %f872, %f380;
|
|
sub.f32 %f876, %f873, %f7;
|
|
sub.f32 %f877, %f874, %f8;
|
|
sub.f32 %f878, %f875, %f9;
|
|
mul.f32 %f879, %f877, %f877;
|
|
fma.rn.f32 %f880, %f876, %f876, %f879;
|
|
fma.rn.f32 %f881, %f878, %f878, %f880;
|
|
sqrt.rn.f32 %f733, %f881;
|
|
rcp.rn.f32 %f882, %f733;
|
|
mul.f32 %f729, %f882, %f876;
|
|
mul.f32 %f730, %f882, %f877;
|
|
mul.f32 %f731, %f882, %f878;
|
|
ld.global.u32 %r114, [imageEnabled];
|
|
and.b32 %r115, %r114, 32;
|
|
setp.eq.s32 %p52, %r115, 0;
|
|
selp.f32 %f883, 0f3F800000, 0f41200000, %p52;
|
|
mul.f32 %f732, %f883, %f142;
|
|
st.local.u32 [%rd2], %r109;
|
|
ld.global.u32 %r101, [root];
|
|
// inline asm
|
|
call _rt_trace_64, (%r101, %f10, %f11, %f12, %f729, %f730, %f731, %r105, %f732, %f733, %rd35, %r29);
|
|
// inline asm
|
|
ld.local.f32 %f884, [%rd2];
|
|
add.f32 %f885, %f837, %f884;
|
|
add.s32 %r116, %r268, 3;
|
|
cvt.rn.f32.s32 %f886, %r116;
|
|
add.f32 %f887, %f13, %f886;
|
|
sub.f32 %f888, %f14, %f886;
|
|
mul.f32 %f889, %f887, 0f3DD32618;
|
|
cvt.rmi.f32.f32 %f890, %f889;
|
|
sub.f32 %f891, %f889, %f890;
|
|
mul.f32 %f892, %f888, 0f3DD2F1AA;
|
|
cvt.rmi.f32.f32 %f893, %f892;
|
|
sub.f32 %f894, %f892, %f893;
|
|
mul.f32 %f895, %f886, 0f3DC74539;
|
|
cvt.rmi.f32.f32 %f896, %f895;
|
|
sub.f32 %f897, %f895, %f896;
|
|
add.f32 %f898, %f894, 0f420551EC;
|
|
add.f32 %f899, %f891, 0f420551EC;
|
|
add.f32 %f900, %f897, 0f420551EC;
|
|
mul.f32 %f901, %f891, %f898;
|
|
fma.rn.f32 %f902, %f894, %f899, %f901;
|
|
fma.rn.f32 %f903, %f897, %f900, %f902;
|
|
add.f32 %f904, %f891, %f903;
|
|
add.f32 %f905, %f894, %f903;
|
|
add.f32 %f906, %f897, %f903;
|
|
add.f32 %f907, %f904, %f905;
|
|
mul.f32 %f908, %f906, %f907;
|
|
cvt.rmi.f32.f32 %f909, %f908;
|
|
sub.f32 %f910, %f908, %f909;
|
|
add.f32 %f911, %f904, %f904;
|
|
mul.f32 %f912, %f905, %f911;
|
|
cvt.rmi.f32.f32 %f913, %f912;
|
|
sub.f32 %f914, %f912, %f913;
|
|
mul.f32 %f915, %f904, %f907;
|
|
cvt.rmi.f32.f32 %f916, %f915;
|
|
sub.f32 %f917, %f915, %f916;
|
|
fma.rn.f32 %f918, %f910, 0f40000000, 0fBF800000;
|
|
fma.rn.f32 %f919, %f914, 0f40000000, 0fBF800000;
|
|
fma.rn.f32 %f920, %f917, 0f40000000, 0fBF800000;
|
|
fma.rn.f32 %f921, %f373, %f918, %f378;
|
|
fma.rn.f32 %f922, %f373, %f919, %f379;
|
|
fma.rn.f32 %f923, %f373, %f920, %f380;
|
|
sub.f32 %f924, %f921, %f7;
|
|
sub.f32 %f925, %f922, %f8;
|
|
sub.f32 %f926, %f923, %f9;
|
|
mul.f32 %f927, %f925, %f925;
|
|
fma.rn.f32 %f928, %f924, %f924, %f927;
|
|
fma.rn.f32 %f929, %f926, %f926, %f928;
|
|
sqrt.rn.f32 %f741, %f929;
|
|
rcp.rn.f32 %f930, %f741;
|
|
mul.f32 %f737, %f930, %f924;
|
|
mul.f32 %f738, %f930, %f925;
|
|
mul.f32 %f739, %f930, %f926;
|
|
ld.global.u32 %r117, [imageEnabled];
|
|
and.b32 %r118, %r117, 32;
|
|
setp.eq.s32 %p53, %r118, 0;
|
|
selp.f32 %f931, 0f3F800000, 0f41200000, %p53;
|
|
mul.f32 %f740, %f931, %f142;
|
|
st.local.u32 [%rd2], %r109;
|
|
ld.global.u32 %r104, [root];
|
|
// inline asm
|
|
call _rt_trace_64, (%r104, %f10, %f11, %f12, %f737, %f738, %f739, %r105, %f740, %f741, %rd35, %r29);
|
|
// inline asm
|
|
ld.local.f32 %f932, [%rd2];
|
|
add.f32 %f1525, %f885, %f932;
|
|
add.s32 %r268, %r268, 4;
|
|
setp.lt.s32 %p54, %r268, %r8;
|
|
@%p54 bra BB0_38;
|
|
|
|
BB0_39:
|
|
cvt.rn.f32.s32 %f933, %r8;
|
|
div.rn.f32 %f1526, %f1525, %f933;
|
|
shr.u32 %r119, %r6, 31;
|
|
cvt.u16.u32 %rs161, %r119;
|
|
|
|
BB0_40:
|
|
fma.rn.f32 %f1498, %f1506, %f1526, %f1498;
|
|
fma.rn.f32 %f1497, %f1507, %f1526, %f1497;
|
|
fma.rn.f32 %f1496, %f1508, %f1526, %f1496;
|
|
fma.rn.f32 %f1495, %f129, %f1526, %f1495;
|
|
fma.rn.f32 %f1494, %f1510, %f1526, %f1494;
|
|
fma.rn.f32 %f1493, %f1511, %f1526, %f1493;
|
|
fma.rn.f32 %f1492, %f132, %f1526, %f1492;
|
|
fma.rn.f32 %f1491, %f1513, %f1526, %f1491;
|
|
fma.rn.f32 %f1490, %f1514, %f1526, %f1490;
|
|
fma.rn.f32 %f1489, %f135, %f1526, %f1489;
|
|
fma.rn.f32 %f1488, %f1516, %f1526, %f1488;
|
|
fma.rn.f32 %f1487, %f1517, %f1526, %f1487;
|
|
fma.rn.f32 %f1486, %f138, %f1526, %f1486;
|
|
fma.rn.f32 %f1485, %f1519, %f1526, %f1485;
|
|
fma.rn.f32 %f1484, %f1520, %f1526, %f1484;
|
|
setp.eq.s16 %p55, %rs161, 0;
|
|
@%p55 bra BB0_55;
|
|
|
|
mov.f32 %f1479, 0fB5BFBE8E;
|
|
mov.f32 %f1478, 0fBF317200;
|
|
mov.f32 %f1477, 0f35BFBE8E;
|
|
mov.f32 %f1476, 0f3F317200;
|
|
mov.f32 %f1475, 0f3DAAAABD;
|
|
mov.f32 %f1474, 0f3C4CAF63;
|
|
mov.f32 %f1473, 0f3B18F0FE;
|
|
mul.f32 %f936, %f132, 0f3F000000;
|
|
mul.f32 %f937, %f135, 0f3F000000;
|
|
mul.f32 %f938, %f937, %f937;
|
|
fma.rn.f32 %f939, %f936, %f936, %f938;
|
|
mul.f32 %f940, %f138, 0f3F000000;
|
|
fma.rn.f32 %f941, %f940, %f940, %f939;
|
|
sqrt.rn.f32 %f942, %f941;
|
|
rcp.rn.f32 %f943, %f942;
|
|
mul.f32 %f944, %f936, %f943;
|
|
mul.f32 %f945, %f937, %f943;
|
|
mul.f32 %f946, %f940, %f943;
|
|
mul.f32 %f947, %f1481, %f945;
|
|
fma.rn.f32 %f948, %f1480, %f944, %f947;
|
|
fma.rn.f32 %f949, %f1482, %f946, %f948;
|
|
fma.rn.f32 %f169, %f949, 0f3F000000, 0f3F000000;
|
|
add.f32 %f950, %f942, %f942;
|
|
div.rn.f32 %f951, %f950, %f129;
|
|
add.f32 %f170, %f951, 0f3F800000;
|
|
div.rn.f32 %f952, %f942, %f129;
|
|
mov.f32 %f953, 0f3F800000;
|
|
sub.f32 %f954, %f953, %f952;
|
|
add.f32 %f955, %f952, 0f3F800000;
|
|
div.rn.f32 %f171, %f954, %f955;
|
|
sub.f32 %f956, %f953, %f171;
|
|
add.f32 %f957, %f170, 0f3F800000;
|
|
mul.f32 %f172, %f957, %f956;
|
|
mul.f32 %f958, %f170, 0f3F000000;
|
|
cvt.rzi.f32.f32 %f959, %f958;
|
|
add.f32 %f960, %f959, %f959;
|
|
sub.f32 %f961, %f170, %f960;
|
|
abs.f32 %f173, %f961;
|
|
abs.f32 %f174, %f169;
|
|
setp.lt.f32 %p56, %f174, 0f00800000;
|
|
mul.f32 %f962, %f174, 0f4B800000;
|
|
selp.f32 %f963, 0fC3170000, 0fC2FE0000, %p56;
|
|
selp.f32 %f964, %f962, %f174, %p56;
|
|
mov.b32 %r120, %f964;
|
|
and.b32 %r121, %r120, 8388607;
|
|
or.b32 %r122, %r121, 1065353216;
|
|
mov.b32 %f965, %r122;
|
|
shr.u32 %r123, %r120, 23;
|
|
cvt.rn.f32.u32 %f966, %r123;
|
|
add.f32 %f967, %f963, %f966;
|
|
setp.gt.f32 %p57, %f965, 0f3FB504F3;
|
|
mul.f32 %f968, %f965, 0f3F000000;
|
|
add.f32 %f969, %f967, 0f3F800000;
|
|
selp.f32 %f970, %f968, %f965, %p57;
|
|
selp.f32 %f971, %f969, %f967, %p57;
|
|
add.f32 %f972, %f970, 0fBF800000;
|
|
add.f32 %f935, %f970, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f934,%f935;
|
|
// inline asm
|
|
add.f32 %f973, %f972, %f972;
|
|
mul.f32 %f974, %f934, %f973;
|
|
mul.f32 %f975, %f974, %f974;
|
|
fma.rn.f32 %f978, %f1473, %f975, %f1474;
|
|
fma.rn.f32 %f980, %f978, %f975, %f1475;
|
|
mul.rn.f32 %f981, %f980, %f975;
|
|
mul.rn.f32 %f982, %f981, %f974;
|
|
sub.f32 %f983, %f972, %f974;
|
|
neg.f32 %f984, %f974;
|
|
add.f32 %f985, %f983, %f983;
|
|
fma.rn.f32 %f986, %f984, %f972, %f985;
|
|
mul.rn.f32 %f987, %f934, %f986;
|
|
add.f32 %f988, %f982, %f974;
|
|
sub.f32 %f989, %f974, %f988;
|
|
add.f32 %f990, %f982, %f989;
|
|
add.f32 %f991, %f987, %f990;
|
|
add.f32 %f992, %f988, %f991;
|
|
sub.f32 %f993, %f988, %f992;
|
|
add.f32 %f994, %f991, %f993;
|
|
mul.rn.f32 %f996, %f971, %f1476;
|
|
mul.rn.f32 %f998, %f971, %f1477;
|
|
add.f32 %f999, %f996, %f992;
|
|
sub.f32 %f1000, %f996, %f999;
|
|
add.f32 %f1001, %f992, %f1000;
|
|
add.f32 %f1002, %f994, %f1001;
|
|
add.f32 %f1003, %f998, %f1002;
|
|
add.f32 %f1004, %f999, %f1003;
|
|
sub.f32 %f1005, %f999, %f1004;
|
|
add.f32 %f1006, %f1003, %f1005;
|
|
abs.f32 %f175, %f170;
|
|
setp.gt.f32 %p58, %f175, 0f77F684DF;
|
|
mul.f32 %f1007, %f170, 0f39000000;
|
|
selp.f32 %f1008, %f1007, %f170, %p58;
|
|
mul.rn.f32 %f1009, %f1008, %f1004;
|
|
neg.f32 %f1010, %f1009;
|
|
fma.rn.f32 %f1011, %f1008, %f1004, %f1010;
|
|
fma.rn.f32 %f1012, %f1008, %f1006, %f1011;
|
|
mov.f32 %f1013, 0f00000000;
|
|
fma.rn.f32 %f1014, %f1013, %f1004, %f1012;
|
|
add.rn.f32 %f1015, %f1009, %f1014;
|
|
neg.f32 %f1016, %f1015;
|
|
add.rn.f32 %f1017, %f1009, %f1016;
|
|
add.rn.f32 %f1018, %f1017, %f1014;
|
|
mov.b32 %r124, %f1015;
|
|
setp.eq.s32 %p59, %r124, 1118925336;
|
|
add.s32 %r125, %r124, -1;
|
|
mov.b32 %f1019, %r125;
|
|
add.f32 %f1020, %f1018, 0f37000000;
|
|
selp.f32 %f1021, %f1019, %f1015, %p59;
|
|
selp.f32 %f176, %f1020, %f1018, %p59;
|
|
mul.f32 %f1022, %f1021, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f1023, %f1022;
|
|
fma.rn.f32 %f1025, %f1023, %f1478, %f1021;
|
|
fma.rn.f32 %f1027, %f1023, %f1479, %f1025;
|
|
mul.f32 %f1028, %f1027, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f1029, %f1028;
|
|
add.f32 %f1030, %f1023, 0f00000000;
|
|
ex2.approx.f32 %f1031, %f1030;
|
|
mul.f32 %f1032, %f1029, %f1031;
|
|
setp.lt.f32 %p60, %f1021, 0fC2D20000;
|
|
selp.f32 %f1033, 0f00000000, %f1032, %p60;
|
|
setp.gt.f32 %p61, %f1021, 0f42D20000;
|
|
selp.f32 %f1527, 0f7F800000, %f1033, %p61;
|
|
setp.eq.f32 %p62, %f1527, 0f7F800000;
|
|
@%p62 bra BB0_43;
|
|
|
|
fma.rn.f32 %f1527, %f1527, %f176, %f1527;
|
|
|
|
BB0_43:
|
|
setp.lt.f32 %p63, %f169, 0f00000000;
|
|
setp.eq.f32 %p64, %f173, 0f3F800000;
|
|
and.pred %p4, %p63, %p64;
|
|
mov.b32 %r126, %f1527;
|
|
xor.b32 %r127, %r126, -2147483648;
|
|
mov.b32 %f1034, %r127;
|
|
selp.f32 %f1529, %f1034, %f1527, %p4;
|
|
setp.eq.f32 %p65, %f169, 0f00000000;
|
|
@%p65 bra BB0_46;
|
|
bra.uni BB0_44;
|
|
|
|
BB0_46:
|
|
add.f32 %f1036, %f169, %f169;
|
|
mov.b32 %r128, %f1036;
|
|
selp.b32 %r129, %r128, 0, %p64;
|
|
or.b32 %r130, %r129, 2139095040;
|
|
setp.lt.f32 %p69, %f170, 0f00000000;
|
|
selp.b32 %r131, %r130, %r129, %p69;
|
|
mov.b32 %f1529, %r131;
|
|
bra.uni BB0_47;
|
|
|
|
BB0_44:
|
|
setp.geu.f32 %p66, %f169, 0f00000000;
|
|
@%p66 bra BB0_47;
|
|
|
|
cvt.rzi.f32.f32 %f1035, %f170;
|
|
setp.neu.f32 %p67, %f1035, %f170;
|
|
selp.f32 %f1529, 0f7FFFFFFF, %f1529, %p67;
|
|
|
|
BB0_47:
|
|
add.f32 %f1037, %f174, %f175;
|
|
mov.b32 %r132, %f1037;
|
|
setp.lt.s32 %p70, %r132, 2139095040;
|
|
@%p70 bra BB0_54;
|
|
|
|
setp.gtu.f32 %p71, %f174, 0f7F800000;
|
|
setp.gtu.f32 %p72, %f175, 0f7F800000;
|
|
or.pred %p73, %p71, %p72;
|
|
@%p73 bra BB0_53;
|
|
bra.uni BB0_49;
|
|
|
|
BB0_53:
|
|
add.f32 %f1529, %f170, %f169;
|
|
bra.uni BB0_54;
|
|
|
|
BB0_49:
|
|
setp.eq.f32 %p74, %f175, 0f7F800000;
|
|
@%p74 bra BB0_52;
|
|
bra.uni BB0_50;
|
|
|
|
BB0_52:
|
|
setp.gt.f32 %p77, %f174, 0f3F800000;
|
|
selp.b32 %r136, 2139095040, 0, %p77;
|
|
xor.b32 %r137, %r136, 2139095040;
|
|
setp.lt.f32 %p78, %f170, 0f00000000;
|
|
selp.b32 %r138, %r137, %r136, %p78;
|
|
mov.b32 %f1038, %r138;
|
|
setp.eq.f32 %p79, %f169, 0fBF800000;
|
|
selp.f32 %f1529, 0f3F800000, %f1038, %p79;
|
|
bra.uni BB0_54;
|
|
|
|
BB0_50:
|
|
setp.neu.f32 %p75, %f174, 0f7F800000;
|
|
@%p75 bra BB0_54;
|
|
|
|
setp.ltu.f32 %p76, %f170, 0f00000000;
|
|
selp.b32 %r133, 0, 2139095040, %p76;
|
|
or.b32 %r134, %r133, -2147483648;
|
|
selp.b32 %r135, %r134, %r133, %p4;
|
|
mov.b32 %f1529, %r135;
|
|
|
|
BB0_54:
|
|
setp.eq.f32 %p80, %f170, 0f00000000;
|
|
setp.eq.f32 %p81, %f169, 0f3F800000;
|
|
or.pred %p82, %p81, %p80;
|
|
selp.f32 %f1039, 0f3F800000, %f1529, %p82;
|
|
fma.rn.f32 %f1040, %f172, %f1039, %f171;
|
|
mul.f32 %f1041, %f129, %f1040;
|
|
div.rn.f32 %f1042, %f1041, %f369;
|
|
div.rn.f32 %f1043, %f1042, %f97;
|
|
cvt.sat.f32.f32 %f1044, %f1043;
|
|
mul.f32 %f1526, %f1526, %f1044;
|
|
|
|
BB0_55:
|
|
add.f32 %f1483, %f1483, %f1526;
|
|
|
|
BB0_56:
|
|
add.s32 %r264, %r264, 1;
|
|
setp.lt.u32 %p83, %r264, %r4;
|
|
@%p83 bra BB0_5;
|
|
|
|
BB0_57:
|
|
ld.global.u32 %r270, [imageEnabled];
|
|
and.b32 %r139, %r270, 8;
|
|
setp.eq.s32 %p84, %r139, 0;
|
|
@%p84 bra BB0_70;
|
|
|
|
cvt.sat.f32.f32 %f223, %f1483;
|
|
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 %f1047, 0f3E68BA2E;
|
|
cvt.rzi.f32.f32 %f1048, %f1047;
|
|
fma.rn.f32 %f1049, %f1048, 0fC0000000, 0f3EE8BA2E;
|
|
abs.f32 %f224, %f1049;
|
|
abs.f32 %f225, %f223;
|
|
setp.lt.f32 %p85, %f225, 0f00800000;
|
|
mul.f32 %f1050, %f225, 0f4B800000;
|
|
selp.f32 %f1051, 0fC3170000, 0fC2FE0000, %p85;
|
|
selp.f32 %f1052, %f1050, %f225, %p85;
|
|
mov.b32 %r142, %f1052;
|
|
and.b32 %r143, %r142, 8388607;
|
|
or.b32 %r144, %r143, 1065353216;
|
|
mov.b32 %f1053, %r144;
|
|
shr.u32 %r145, %r142, 23;
|
|
cvt.rn.f32.u32 %f1054, %r145;
|
|
add.f32 %f1055, %f1051, %f1054;
|
|
setp.gt.f32 %p86, %f1053, 0f3FB504F3;
|
|
mul.f32 %f1056, %f1053, 0f3F000000;
|
|
add.f32 %f1057, %f1055, 0f3F800000;
|
|
selp.f32 %f1058, %f1056, %f1053, %p86;
|
|
selp.f32 %f1059, %f1057, %f1055, %p86;
|
|
add.f32 %f1060, %f1058, 0fBF800000;
|
|
add.f32 %f1046, %f1058, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f1045,%f1046;
|
|
// inline asm
|
|
add.f32 %f1061, %f1060, %f1060;
|
|
mul.f32 %f1062, %f1045, %f1061;
|
|
mul.f32 %f1063, %f1062, %f1062;
|
|
mov.f32 %f1064, 0f3C4CAF63;
|
|
mov.f32 %f1065, 0f3B18F0FE;
|
|
fma.rn.f32 %f1066, %f1065, %f1063, %f1064;
|
|
mov.f32 %f1067, 0f3DAAAABD;
|
|
fma.rn.f32 %f1068, %f1066, %f1063, %f1067;
|
|
mul.rn.f32 %f1069, %f1068, %f1063;
|
|
mul.rn.f32 %f1070, %f1069, %f1062;
|
|
sub.f32 %f1071, %f1060, %f1062;
|
|
neg.f32 %f1072, %f1062;
|
|
add.f32 %f1073, %f1071, %f1071;
|
|
fma.rn.f32 %f1074, %f1072, %f1060, %f1073;
|
|
mul.rn.f32 %f1075, %f1045, %f1074;
|
|
add.f32 %f1076, %f1070, %f1062;
|
|
sub.f32 %f1077, %f1062, %f1076;
|
|
add.f32 %f1078, %f1070, %f1077;
|
|
add.f32 %f1079, %f1075, %f1078;
|
|
add.f32 %f1080, %f1076, %f1079;
|
|
sub.f32 %f1081, %f1076, %f1080;
|
|
add.f32 %f1082, %f1079, %f1081;
|
|
mov.f32 %f1083, 0f3F317200;
|
|
mul.rn.f32 %f1084, %f1059, %f1083;
|
|
mov.f32 %f1085, 0f35BFBE8E;
|
|
mul.rn.f32 %f1086, %f1059, %f1085;
|
|
add.f32 %f1087, %f1084, %f1080;
|
|
sub.f32 %f1088, %f1084, %f1087;
|
|
add.f32 %f1089, %f1080, %f1088;
|
|
add.f32 %f1090, %f1082, %f1089;
|
|
add.f32 %f1091, %f1086, %f1090;
|
|
add.f32 %f1092, %f1087, %f1091;
|
|
sub.f32 %f1093, %f1087, %f1092;
|
|
add.f32 %f1094, %f1091, %f1093;
|
|
mov.f32 %f1095, 0f3EE8BA2E;
|
|
mul.rn.f32 %f1096, %f1095, %f1092;
|
|
neg.f32 %f1097, %f1096;
|
|
fma.rn.f32 %f1098, %f1095, %f1092, %f1097;
|
|
fma.rn.f32 %f1099, %f1095, %f1094, %f1098;
|
|
mov.f32 %f1100, 0f00000000;
|
|
fma.rn.f32 %f1101, %f1100, %f1092, %f1099;
|
|
add.rn.f32 %f1102, %f1096, %f1101;
|
|
neg.f32 %f1103, %f1102;
|
|
add.rn.f32 %f1104, %f1096, %f1103;
|
|
add.rn.f32 %f1105, %f1104, %f1101;
|
|
mov.b32 %r146, %f1102;
|
|
setp.eq.s32 %p87, %r146, 1118925336;
|
|
add.s32 %r147, %r146, -1;
|
|
mov.b32 %f1106, %r147;
|
|
add.f32 %f1107, %f1105, 0f37000000;
|
|
selp.f32 %f1108, %f1106, %f1102, %p87;
|
|
selp.f32 %f226, %f1107, %f1105, %p87;
|
|
mul.f32 %f1109, %f1108, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f1110, %f1109;
|
|
mov.f32 %f1111, 0fBF317200;
|
|
fma.rn.f32 %f1112, %f1110, %f1111, %f1108;
|
|
mov.f32 %f1113, 0fB5BFBE8E;
|
|
fma.rn.f32 %f1114, %f1110, %f1113, %f1112;
|
|
mul.f32 %f1115, %f1114, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f1116, %f1115;
|
|
add.f32 %f1117, %f1110, 0f00000000;
|
|
ex2.approx.f32 %f1118, %f1117;
|
|
mul.f32 %f1119, %f1116, %f1118;
|
|
setp.lt.f32 %p88, %f1108, 0fC2D20000;
|
|
selp.f32 %f1120, 0f00000000, %f1119, %p88;
|
|
setp.gt.f32 %p89, %f1108, 0f42D20000;
|
|
selp.f32 %f1563, 0f7F800000, %f1120, %p89;
|
|
setp.eq.f32 %p90, %f1563, 0f7F800000;
|
|
@%p90 bra BB0_60;
|
|
|
|
fma.rn.f32 %f1563, %f1563, %f226, %f1563;
|
|
|
|
BB0_60:
|
|
setp.lt.f32 %p91, %f223, 0f00000000;
|
|
setp.eq.f32 %p92, %f224, 0f3F800000;
|
|
and.pred %p5, %p91, %p92;
|
|
mov.b32 %r148, %f1563;
|
|
xor.b32 %r149, %r148, -2147483648;
|
|
mov.b32 %f1121, %r149;
|
|
selp.f32 %f1565, %f1121, %f1563, %p5;
|
|
setp.eq.f32 %p93, %f223, 0f00000000;
|
|
@%p93 bra BB0_63;
|
|
bra.uni BB0_61;
|
|
|
|
BB0_63:
|
|
add.f32 %f1124, %f223, %f223;
|
|
selp.f32 %f1565, %f1124, 0f00000000, %p92;
|
|
bra.uni BB0_64;
|
|
|
|
BB0_128:
|
|
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 %f1444, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs107, %f1444;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs108, 0;
|
|
st.v4.u16 [%rd184], {%rs107, %rs107, %rs107, %rs108};
|
|
|
|
BB0_129:
|
|
ld.global.u32 %r236, [additive];
|
|
setp.eq.s32 %p160, %r236, 0;
|
|
@%p160 bra BB0_131;
|
|
|
|
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 %f1445, %rs115;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1446, %rs116;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1447, %rs117;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd197), _rt_buffer_get_64, (%rd192, %r28, %r240, %rd6, %rd7, %rd13, %rd13);
|
|
// inline asm
|
|
add.f32 %f1448, %f1445, 0f00000000;
|
|
add.f32 %f1449, %f1446, 0f00000000;
|
|
add.f32 %f1450, %f1447, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs114, %f1450;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs113, %f1449;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs112, %f1448;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs119, 0;
|
|
st.v4.u16 [%rd197], {%rs112, %rs113, %rs114, %rs119};
|
|
bra.uni BB0_132;
|
|
|
|
BB0_131:
|
|
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 %f1451, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs120, %f1451;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs121, 0;
|
|
st.v4.u16 [%rd204], {%rs120, %rs120, %rs120, %rs121};
|
|
|
|
BB0_132:
|
|
ld.global.u32 %r243, [additive];
|
|
setp.eq.s32 %p161, %r243, 0;
|
|
@%p161 bra BB0_134;
|
|
|
|
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 %f1452, %rs128;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1453, %rs129;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1454, %rs130;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd217), _rt_buffer_get_64, (%rd212, %r28, %r247, %rd6, %rd7, %rd13, %rd13);
|
|
// inline asm
|
|
add.f32 %f1455, %f1452, 0f00000000;
|
|
add.f32 %f1456, %f1453, 0f00000000;
|
|
add.f32 %f1457, %f1454, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs127, %f1457;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs126, %f1456;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs125, %f1455;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs132, 0;
|
|
st.v4.u16 [%rd217], {%rs125, %rs126, %rs127, %rs132};
|
|
bra.uni BB0_135;
|
|
|
|
BB0_134:
|
|
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 %f1458, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs133, %f1458;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs134, 0;
|
|
st.v4.u16 [%rd224], {%rs133, %rs133, %rs133, %rs134};
|
|
|
|
BB0_135:
|
|
ld.global.u32 %r250, [additive];
|
|
setp.eq.s32 %p162, %r250, 0;
|
|
@%p162 bra BB0_137;
|
|
|
|
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 %f1459, %rs141;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1460, %rs142;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1461, %rs143;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd237), _rt_buffer_get_64, (%rd232, %r28, %r254, %rd6, %rd7, %rd13, %rd13);
|
|
// inline asm
|
|
add.f32 %f1462, %f1459, 0f00000000;
|
|
add.f32 %f1463, %f1460, 0f00000000;
|
|
add.f32 %f1464, %f1461, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs140, %f1464;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs139, %f1463;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs138, %f1462;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs145, 0;
|
|
st.v4.u16 [%rd237], {%rs138, %rs139, %rs140, %rs145};
|
|
bra.uni BB0_138;
|
|
|
|
BB0_137:
|
|
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 %f1465, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs146, %f1465;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs147, 0;
|
|
st.v4.u16 [%rd244], {%rs146, %rs146, %rs146, %rs147};
|
|
|
|
BB0_138:
|
|
ld.global.u32 %r257, [additive];
|
|
setp.eq.s32 %p163, %r257, 0;
|
|
@%p163 bra BB0_140;
|
|
|
|
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 %f1466, %rs154;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1467, %rs155;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1468, %rs156;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd257), _rt_buffer_get_64, (%rd252, %r28, %r261, %rd6, %rd7, %rd13, %rd13);
|
|
// inline asm
|
|
add.f32 %f1469, %f1466, 0f00000000;
|
|
add.f32 %f1470, %f1467, 0f00000000;
|
|
add.f32 %f1471, %f1468, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs153, %f1471;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs152, %f1470;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs151, %f1469;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs158, 0;
|
|
st.v4.u16 [%rd257], {%rs151, %rs152, %rs153, %rs158};
|
|
bra.uni BB0_141;
|
|
|
|
BB0_140:
|
|
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 %f1472, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs159, %f1472;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs160, 0;
|
|
st.v4.u16 [%rd264], {%rs159, %rs159, %rs159, %rs160};
|
|
bra.uni BB0_141;
|
|
|
|
BB0_61:
|
|
setp.geu.f32 %p94, %f223, 0f00000000;
|
|
@%p94 bra BB0_64;
|
|
|
|
cvt.rzi.f32.f32 %f1123, %f1095;
|
|
setp.neu.f32 %p95, %f1123, 0f3EE8BA2E;
|
|
selp.f32 %f1565, 0f7FFFFFFF, %f1565, %p95;
|
|
|
|
BB0_64:
|
|
add.f32 %f1125, %f225, 0f3EE8BA2E;
|
|
mov.b32 %r150, %f1125;
|
|
setp.lt.s32 %p97, %r150, 2139095040;
|
|
@%p97 bra BB0_69;
|
|
|
|
setp.gtu.f32 %p98, %f225, 0f7F800000;
|
|
@%p98 bra BB0_68;
|
|
bra.uni BB0_66;
|
|
|
|
BB0_68:
|
|
add.f32 %f1565, %f223, 0f3EE8BA2E;
|
|
bra.uni BB0_69;
|
|
|
|
BB0_66:
|
|
setp.neu.f32 %p99, %f225, 0f7F800000;
|
|
@%p99 bra BB0_69;
|
|
|
|
selp.f32 %f1565, 0fFF800000, 0f7F800000, %p5;
|
|
|
|
BB0_69:
|
|
mul.f32 %f1126, %f1565, 0f437F0000;
|
|
setp.eq.f32 %p100, %f223, 0f3F800000;
|
|
selp.f32 %f1127, 0f437F0000, %f1126, %p100;
|
|
cvt.rzi.u32.f32 %r151, %f1127;
|
|
cvt.u16.u32 %rs17, %r151;
|
|
mov.u16 %rs18, 255;
|
|
st.v2.u8 [%rd43], {%rs17, %rs18};
|
|
ld.global.u32 %r270, [imageEnabled];
|
|
|
|
BB0_70:
|
|
and.b32 %r152, %r270, 1;
|
|
setp.eq.b32 %p101, %r152, 1;
|
|
@!%p101 bra BB0_105;
|
|
bra.uni BB0_71;
|
|
|
|
BB0_71:
|
|
mov.f32 %f1130, 0f3E666666;
|
|
cvt.rzi.f32.f32 %f1131, %f1130;
|
|
fma.rn.f32 %f1132, %f1131, 0fC0000000, 0f3EE66666;
|
|
abs.f32 %f237, %f1132;
|
|
abs.f32 %f238, %f1498;
|
|
setp.lt.f32 %p102, %f238, 0f00800000;
|
|
mul.f32 %f1133, %f238, 0f4B800000;
|
|
selp.f32 %f1134, 0fC3170000, 0fC2FE0000, %p102;
|
|
selp.f32 %f1135, %f1133, %f238, %p102;
|
|
mov.b32 %r153, %f1135;
|
|
and.b32 %r154, %r153, 8388607;
|
|
or.b32 %r155, %r154, 1065353216;
|
|
mov.b32 %f1136, %r155;
|
|
shr.u32 %r156, %r153, 23;
|
|
cvt.rn.f32.u32 %f1137, %r156;
|
|
add.f32 %f1138, %f1134, %f1137;
|
|
setp.gt.f32 %p103, %f1136, 0f3FB504F3;
|
|
mul.f32 %f1139, %f1136, 0f3F000000;
|
|
add.f32 %f1140, %f1138, 0f3F800000;
|
|
selp.f32 %f1141, %f1139, %f1136, %p103;
|
|
selp.f32 %f1142, %f1140, %f1138, %p103;
|
|
add.f32 %f1143, %f1141, 0fBF800000;
|
|
add.f32 %f1129, %f1141, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f1128,%f1129;
|
|
// inline asm
|
|
add.f32 %f1144, %f1143, %f1143;
|
|
mul.f32 %f1145, %f1128, %f1144;
|
|
mul.f32 %f1146, %f1145, %f1145;
|
|
mov.f32 %f1147, 0f3C4CAF63;
|
|
mov.f32 %f1148, 0f3B18F0FE;
|
|
fma.rn.f32 %f1149, %f1148, %f1146, %f1147;
|
|
mov.f32 %f1150, 0f3DAAAABD;
|
|
fma.rn.f32 %f1151, %f1149, %f1146, %f1150;
|
|
mul.rn.f32 %f1152, %f1151, %f1146;
|
|
mul.rn.f32 %f1153, %f1152, %f1145;
|
|
sub.f32 %f1154, %f1143, %f1145;
|
|
neg.f32 %f1155, %f1145;
|
|
add.f32 %f1156, %f1154, %f1154;
|
|
fma.rn.f32 %f1157, %f1155, %f1143, %f1156;
|
|
mul.rn.f32 %f1158, %f1128, %f1157;
|
|
add.f32 %f1159, %f1153, %f1145;
|
|
sub.f32 %f1160, %f1145, %f1159;
|
|
add.f32 %f1161, %f1153, %f1160;
|
|
add.f32 %f1162, %f1158, %f1161;
|
|
add.f32 %f1163, %f1159, %f1162;
|
|
sub.f32 %f1164, %f1159, %f1163;
|
|
add.f32 %f1165, %f1162, %f1164;
|
|
mov.f32 %f1166, 0f3F317200;
|
|
mul.rn.f32 %f1167, %f1142, %f1166;
|
|
mov.f32 %f1168, 0f35BFBE8E;
|
|
mul.rn.f32 %f1169, %f1142, %f1168;
|
|
add.f32 %f1170, %f1167, %f1163;
|
|
sub.f32 %f1171, %f1167, %f1170;
|
|
add.f32 %f1172, %f1163, %f1171;
|
|
add.f32 %f1173, %f1165, %f1172;
|
|
add.f32 %f1174, %f1169, %f1173;
|
|
add.f32 %f1175, %f1170, %f1174;
|
|
sub.f32 %f1176, %f1170, %f1175;
|
|
add.f32 %f1177, %f1174, %f1176;
|
|
mov.f32 %f1178, 0f3EE66666;
|
|
mul.rn.f32 %f1179, %f1178, %f1175;
|
|
neg.f32 %f1180, %f1179;
|
|
fma.rn.f32 %f1181, %f1178, %f1175, %f1180;
|
|
fma.rn.f32 %f1182, %f1178, %f1177, %f1181;
|
|
mov.f32 %f1183, 0f00000000;
|
|
fma.rn.f32 %f1184, %f1183, %f1175, %f1182;
|
|
add.rn.f32 %f1185, %f1179, %f1184;
|
|
neg.f32 %f1186, %f1185;
|
|
add.rn.f32 %f1187, %f1179, %f1186;
|
|
add.rn.f32 %f1188, %f1187, %f1184;
|
|
mov.b32 %r157, %f1185;
|
|
setp.eq.s32 %p104, %r157, 1118925336;
|
|
add.s32 %r158, %r157, -1;
|
|
mov.b32 %f1189, %r158;
|
|
add.f32 %f1190, %f1188, 0f37000000;
|
|
selp.f32 %f1191, %f1189, %f1185, %p104;
|
|
selp.f32 %f239, %f1190, %f1188, %p104;
|
|
mul.f32 %f1192, %f1191, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f1193, %f1192;
|
|
mov.f32 %f1194, 0fBF317200;
|
|
fma.rn.f32 %f1195, %f1193, %f1194, %f1191;
|
|
mov.f32 %f1196, 0fB5BFBE8E;
|
|
fma.rn.f32 %f1197, %f1193, %f1196, %f1195;
|
|
mul.f32 %f1198, %f1197, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f1199, %f1198;
|
|
add.f32 %f1200, %f1193, 0f00000000;
|
|
ex2.approx.f32 %f1201, %f1200;
|
|
mul.f32 %f1202, %f1199, %f1201;
|
|
setp.lt.f32 %p105, %f1191, 0fC2D20000;
|
|
selp.f32 %f1203, 0f00000000, %f1202, %p105;
|
|
setp.gt.f32 %p106, %f1191, 0f42D20000;
|
|
selp.f32 %f1566, 0f7F800000, %f1203, %p106;
|
|
setp.eq.f32 %p107, %f1566, 0f7F800000;
|
|
@%p107 bra BB0_73;
|
|
|
|
fma.rn.f32 %f1566, %f1566, %f239, %f1566;
|
|
|
|
BB0_73:
|
|
setp.lt.f32 %p108, %f1498, 0f00000000;
|
|
setp.eq.f32 %p109, %f237, 0f3F800000;
|
|
and.pred %p6, %p108, %p109;
|
|
mov.b32 %r159, %f1566;
|
|
xor.b32 %r160, %r159, -2147483648;
|
|
mov.b32 %f1204, %r160;
|
|
selp.f32 %f1568, %f1204, %f1566, %p6;
|
|
setp.eq.f32 %p110, %f1498, 0f00000000;
|
|
@%p110 bra BB0_76;
|
|
bra.uni BB0_74;
|
|
|
|
BB0_76:
|
|
add.f32 %f1207, %f1498, %f1498;
|
|
selp.f32 %f1568, %f1207, 0f00000000, %p109;
|
|
bra.uni BB0_77;
|
|
|
|
BB0_74:
|
|
setp.geu.f32 %p111, %f1498, 0f00000000;
|
|
@%p111 bra BB0_77;
|
|
|
|
cvt.rzi.f32.f32 %f1206, %f1178;
|
|
setp.neu.f32 %p112, %f1206, 0f3EE66666;
|
|
selp.f32 %f1568, 0f7FFFFFFF, %f1568, %p112;
|
|
|
|
BB0_77:
|
|
add.f32 %f1208, %f238, 0f3EE66666;
|
|
mov.b32 %r161, %f1208;
|
|
setp.lt.s32 %p114, %r161, 2139095040;
|
|
@%p114 bra BB0_82;
|
|
|
|
setp.gtu.f32 %p115, %f238, 0f7F800000;
|
|
@%p115 bra BB0_81;
|
|
bra.uni BB0_79;
|
|
|
|
BB0_81:
|
|
add.f32 %f1568, %f1498, 0f3EE66666;
|
|
bra.uni BB0_82;
|
|
|
|
BB0_79:
|
|
setp.neu.f32 %p116, %f238, 0f7F800000;
|
|
@%p116 bra BB0_82;
|
|
|
|
selp.f32 %f1568, 0fFF800000, 0f7F800000, %p6;
|
|
|
|
BB0_82:
|
|
setp.eq.f32 %p117, %f1498, 0f3F800000;
|
|
selp.f32 %f250, 0f3F800000, %f1568, %p117;
|
|
abs.f32 %f251, %f1497;
|
|
setp.lt.f32 %p118, %f251, 0f00800000;
|
|
mul.f32 %f1211, %f251, 0f4B800000;
|
|
selp.f32 %f1212, 0fC3170000, 0fC2FE0000, %p118;
|
|
selp.f32 %f1213, %f1211, %f251, %p118;
|
|
mov.b32 %r162, %f1213;
|
|
and.b32 %r163, %r162, 8388607;
|
|
or.b32 %r164, %r163, 1065353216;
|
|
mov.b32 %f1214, %r164;
|
|
shr.u32 %r165, %r162, 23;
|
|
cvt.rn.f32.u32 %f1215, %r165;
|
|
add.f32 %f1216, %f1212, %f1215;
|
|
setp.gt.f32 %p119, %f1214, 0f3FB504F3;
|
|
mul.f32 %f1217, %f1214, 0f3F000000;
|
|
add.f32 %f1218, %f1216, 0f3F800000;
|
|
selp.f32 %f1219, %f1217, %f1214, %p119;
|
|
selp.f32 %f1220, %f1218, %f1216, %p119;
|
|
add.f32 %f1221, %f1219, 0fBF800000;
|
|
add.f32 %f1210, %f1219, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f1209,%f1210;
|
|
// inline asm
|
|
add.f32 %f1222, %f1221, %f1221;
|
|
mul.f32 %f1223, %f1209, %f1222;
|
|
mul.f32 %f1224, %f1223, %f1223;
|
|
fma.rn.f32 %f1227, %f1148, %f1224, %f1147;
|
|
fma.rn.f32 %f1229, %f1227, %f1224, %f1150;
|
|
mul.rn.f32 %f1230, %f1229, %f1224;
|
|
mul.rn.f32 %f1231, %f1230, %f1223;
|
|
sub.f32 %f1232, %f1221, %f1223;
|
|
neg.f32 %f1233, %f1223;
|
|
add.f32 %f1234, %f1232, %f1232;
|
|
fma.rn.f32 %f1235, %f1233, %f1221, %f1234;
|
|
mul.rn.f32 %f1236, %f1209, %f1235;
|
|
add.f32 %f1237, %f1231, %f1223;
|
|
sub.f32 %f1238, %f1223, %f1237;
|
|
add.f32 %f1239, %f1231, %f1238;
|
|
add.f32 %f1240, %f1236, %f1239;
|
|
add.f32 %f1241, %f1237, %f1240;
|
|
sub.f32 %f1242, %f1237, %f1241;
|
|
add.f32 %f1243, %f1240, %f1242;
|
|
mul.rn.f32 %f1245, %f1220, %f1166;
|
|
mul.rn.f32 %f1247, %f1220, %f1168;
|
|
add.f32 %f1248, %f1245, %f1241;
|
|
sub.f32 %f1249, %f1245, %f1248;
|
|
add.f32 %f1250, %f1241, %f1249;
|
|
add.f32 %f1251, %f1243, %f1250;
|
|
add.f32 %f1252, %f1247, %f1251;
|
|
add.f32 %f1253, %f1248, %f1252;
|
|
sub.f32 %f1254, %f1248, %f1253;
|
|
add.f32 %f1255, %f1252, %f1254;
|
|
mul.rn.f32 %f1257, %f1178, %f1253;
|
|
neg.f32 %f1258, %f1257;
|
|
fma.rn.f32 %f1259, %f1178, %f1253, %f1258;
|
|
fma.rn.f32 %f1260, %f1178, %f1255, %f1259;
|
|
fma.rn.f32 %f1262, %f1183, %f1253, %f1260;
|
|
add.rn.f32 %f1263, %f1257, %f1262;
|
|
neg.f32 %f1264, %f1263;
|
|
add.rn.f32 %f1265, %f1257, %f1264;
|
|
add.rn.f32 %f1266, %f1265, %f1262;
|
|
mov.b32 %r166, %f1263;
|
|
setp.eq.s32 %p120, %r166, 1118925336;
|
|
add.s32 %r167, %r166, -1;
|
|
mov.b32 %f1267, %r167;
|
|
add.f32 %f1268, %f1266, 0f37000000;
|
|
selp.f32 %f1269, %f1267, %f1263, %p120;
|
|
selp.f32 %f252, %f1268, %f1266, %p120;
|
|
mul.f32 %f1270, %f1269, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f1271, %f1270;
|
|
fma.rn.f32 %f1273, %f1271, %f1194, %f1269;
|
|
fma.rn.f32 %f1275, %f1271, %f1196, %f1273;
|
|
mul.f32 %f1276, %f1275, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f1277, %f1276;
|
|
add.f32 %f1278, %f1271, 0f00000000;
|
|
ex2.approx.f32 %f1279, %f1278;
|
|
mul.f32 %f1280, %f1277, %f1279;
|
|
setp.lt.f32 %p121, %f1269, 0fC2D20000;
|
|
selp.f32 %f1281, 0f00000000, %f1280, %p121;
|
|
setp.gt.f32 %p122, %f1269, 0f42D20000;
|
|
selp.f32 %f1569, 0f7F800000, %f1281, %p122;
|
|
setp.eq.f32 %p123, %f1569, 0f7F800000;
|
|
@%p123 bra BB0_84;
|
|
|
|
fma.rn.f32 %f1569, %f1569, %f252, %f1569;
|
|
|
|
BB0_84:
|
|
setp.lt.f32 %p124, %f1497, 0f00000000;
|
|
and.pred %p7, %p124, %p109;
|
|
mov.b32 %r168, %f1569;
|
|
xor.b32 %r169, %r168, -2147483648;
|
|
mov.b32 %f1282, %r169;
|
|
selp.f32 %f1571, %f1282, %f1569, %p7;
|
|
setp.eq.f32 %p126, %f1497, 0f00000000;
|
|
@%p126 bra BB0_87;
|
|
bra.uni BB0_85;
|
|
|
|
BB0_87:
|
|
add.f32 %f1285, %f1497, %f1497;
|
|
selp.f32 %f1571, %f1285, 0f00000000, %p109;
|
|
bra.uni BB0_88;
|
|
|
|
BB0_85:
|
|
setp.geu.f32 %p127, %f1497, 0f00000000;
|
|
@%p127 bra BB0_88;
|
|
|
|
cvt.rzi.f32.f32 %f1284, %f1178;
|
|
setp.neu.f32 %p128, %f1284, 0f3EE66666;
|
|
selp.f32 %f1571, 0f7FFFFFFF, %f1571, %p128;
|
|
|
|
BB0_88:
|
|
add.f32 %f1286, %f251, 0f3EE66666;
|
|
mov.b32 %r170, %f1286;
|
|
setp.lt.s32 %p130, %r170, 2139095040;
|
|
@%p130 bra BB0_93;
|
|
|
|
setp.gtu.f32 %p131, %f251, 0f7F800000;
|
|
@%p131 bra BB0_92;
|
|
bra.uni BB0_90;
|
|
|
|
BB0_92:
|
|
add.f32 %f1571, %f1497, 0f3EE66666;
|
|
bra.uni BB0_93;
|
|
|
|
BB0_90:
|
|
setp.neu.f32 %p132, %f251, 0f7F800000;
|
|
@%p132 bra BB0_93;
|
|
|
|
selp.f32 %f1571, 0fFF800000, 0f7F800000, %p7;
|
|
|
|
BB0_93:
|
|
setp.eq.f32 %p133, %f1497, 0f3F800000;
|
|
selp.f32 %f263, 0f3F800000, %f1571, %p133;
|
|
abs.f32 %f264, %f1496;
|
|
setp.lt.f32 %p134, %f264, 0f00800000;
|
|
mul.f32 %f1289, %f264, 0f4B800000;
|
|
selp.f32 %f1290, 0fC3170000, 0fC2FE0000, %p134;
|
|
selp.f32 %f1291, %f1289, %f264, %p134;
|
|
mov.b32 %r171, %f1291;
|
|
and.b32 %r172, %r171, 8388607;
|
|
or.b32 %r173, %r172, 1065353216;
|
|
mov.b32 %f1292, %r173;
|
|
shr.u32 %r174, %r171, 23;
|
|
cvt.rn.f32.u32 %f1293, %r174;
|
|
add.f32 %f1294, %f1290, %f1293;
|
|
setp.gt.f32 %p135, %f1292, 0f3FB504F3;
|
|
mul.f32 %f1295, %f1292, 0f3F000000;
|
|
add.f32 %f1296, %f1294, 0f3F800000;
|
|
selp.f32 %f1297, %f1295, %f1292, %p135;
|
|
selp.f32 %f1298, %f1296, %f1294, %p135;
|
|
add.f32 %f1299, %f1297, 0fBF800000;
|
|
add.f32 %f1288, %f1297, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f1287,%f1288;
|
|
// inline asm
|
|
add.f32 %f1300, %f1299, %f1299;
|
|
mul.f32 %f1301, %f1287, %f1300;
|
|
mul.f32 %f1302, %f1301, %f1301;
|
|
fma.rn.f32 %f1305, %f1148, %f1302, %f1147;
|
|
fma.rn.f32 %f1307, %f1305, %f1302, %f1150;
|
|
mul.rn.f32 %f1308, %f1307, %f1302;
|
|
mul.rn.f32 %f1309, %f1308, %f1301;
|
|
sub.f32 %f1310, %f1299, %f1301;
|
|
neg.f32 %f1311, %f1301;
|
|
add.f32 %f1312, %f1310, %f1310;
|
|
fma.rn.f32 %f1313, %f1311, %f1299, %f1312;
|
|
mul.rn.f32 %f1314, %f1287, %f1313;
|
|
add.f32 %f1315, %f1309, %f1301;
|
|
sub.f32 %f1316, %f1301, %f1315;
|
|
add.f32 %f1317, %f1309, %f1316;
|
|
add.f32 %f1318, %f1314, %f1317;
|
|
add.f32 %f1319, %f1315, %f1318;
|
|
sub.f32 %f1320, %f1315, %f1319;
|
|
add.f32 %f1321, %f1318, %f1320;
|
|
mul.rn.f32 %f1323, %f1298, %f1166;
|
|
mul.rn.f32 %f1325, %f1298, %f1168;
|
|
add.f32 %f1326, %f1323, %f1319;
|
|
sub.f32 %f1327, %f1323, %f1326;
|
|
add.f32 %f1328, %f1319, %f1327;
|
|
add.f32 %f1329, %f1321, %f1328;
|
|
add.f32 %f1330, %f1325, %f1329;
|
|
add.f32 %f1331, %f1326, %f1330;
|
|
sub.f32 %f1332, %f1326, %f1331;
|
|
add.f32 %f1333, %f1330, %f1332;
|
|
mul.rn.f32 %f1335, %f1178, %f1331;
|
|
neg.f32 %f1336, %f1335;
|
|
fma.rn.f32 %f1337, %f1178, %f1331, %f1336;
|
|
fma.rn.f32 %f1338, %f1178, %f1333, %f1337;
|
|
fma.rn.f32 %f1340, %f1183, %f1331, %f1338;
|
|
add.rn.f32 %f1341, %f1335, %f1340;
|
|
neg.f32 %f1342, %f1341;
|
|
add.rn.f32 %f1343, %f1335, %f1342;
|
|
add.rn.f32 %f1344, %f1343, %f1340;
|
|
mov.b32 %r175, %f1341;
|
|
setp.eq.s32 %p136, %r175, 1118925336;
|
|
add.s32 %r176, %r175, -1;
|
|
mov.b32 %f1345, %r176;
|
|
add.f32 %f1346, %f1344, 0f37000000;
|
|
selp.f32 %f1347, %f1345, %f1341, %p136;
|
|
selp.f32 %f265, %f1346, %f1344, %p136;
|
|
mul.f32 %f1348, %f1347, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f1349, %f1348;
|
|
fma.rn.f32 %f1351, %f1349, %f1194, %f1347;
|
|
fma.rn.f32 %f1353, %f1349, %f1196, %f1351;
|
|
mul.f32 %f1354, %f1353, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f1355, %f1354;
|
|
add.f32 %f1356, %f1349, 0f00000000;
|
|
ex2.approx.f32 %f1357, %f1356;
|
|
mul.f32 %f1358, %f1355, %f1357;
|
|
setp.lt.f32 %p137, %f1347, 0fC2D20000;
|
|
selp.f32 %f1359, 0f00000000, %f1358, %p137;
|
|
setp.gt.f32 %p138, %f1347, 0f42D20000;
|
|
selp.f32 %f1572, 0f7F800000, %f1359, %p138;
|
|
setp.eq.f32 %p139, %f1572, 0f7F800000;
|
|
@%p139 bra BB0_95;
|
|
|
|
fma.rn.f32 %f1572, %f1572, %f265, %f1572;
|
|
|
|
BB0_95:
|
|
setp.lt.f32 %p140, %f1496, 0f00000000;
|
|
and.pred %p8, %p140, %p109;
|
|
mov.b32 %r177, %f1572;
|
|
xor.b32 %r178, %r177, -2147483648;
|
|
mov.b32 %f1360, %r178;
|
|
selp.f32 %f1574, %f1360, %f1572, %p8;
|
|
setp.eq.f32 %p142, %f1496, 0f00000000;
|
|
@%p142 bra BB0_98;
|
|
bra.uni BB0_96;
|
|
|
|
BB0_98:
|
|
add.f32 %f1363, %f1496, %f1496;
|
|
selp.f32 %f1574, %f1363, 0f00000000, %p109;
|
|
bra.uni BB0_99;
|
|
|
|
BB0_96:
|
|
setp.geu.f32 %p143, %f1496, 0f00000000;
|
|
@%p143 bra BB0_99;
|
|
|
|
cvt.rzi.f32.f32 %f1362, %f1178;
|
|
setp.neu.f32 %p144, %f1362, 0f3EE66666;
|
|
selp.f32 %f1574, 0f7FFFFFFF, %f1574, %p144;
|
|
|
|
BB0_99:
|
|
add.f32 %f1364, %f264, 0f3EE66666;
|
|
mov.b32 %r179, %f1364;
|
|
setp.lt.s32 %p146, %r179, 2139095040;
|
|
@%p146 bra BB0_104;
|
|
|
|
setp.gtu.f32 %p147, %f264, 0f7F800000;
|
|
@%p147 bra BB0_103;
|
|
bra.uni BB0_101;
|
|
|
|
BB0_103:
|
|
add.f32 %f1574, %f1496, 0f3EE66666;
|
|
bra.uni BB0_104;
|
|
|
|
BB0_101:
|
|
setp.neu.f32 %p148, %f264, 0f7F800000;
|
|
@%p148 bra BB0_104;
|
|
|
|
selp.f32 %f1574, 0fFF800000, 0f7F800000, %p8;
|
|
|
|
BB0_104:
|
|
setp.eq.f32 %p149, %f1496, 0f3F800000;
|
|
selp.f32 %f1365, 0f3F800000, %f1574, %p149;
|
|
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 %f1366, %f1365;
|
|
mul.f32 %f1367, %f1366, 0f437FFD71;
|
|
cvt.rzi.u32.f32 %r182, %f1367;
|
|
cvt.sat.f32.f32 %f1368, %f263;
|
|
mul.f32 %f1369, %f1368, 0f437FFD71;
|
|
cvt.rzi.u32.f32 %r183, %f1369;
|
|
cvt.sat.f32.f32 %f1370, %f250;
|
|
mul.f32 %f1371, %f1370, 0f437FFD71;
|
|
cvt.rzi.u32.f32 %r184, %f1371;
|
|
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 %r270, [imageEnabled];
|
|
|
|
BB0_105:
|
|
cvt.u64.u32 %rd4, %r2;
|
|
cvt.u64.u32 %rd5, %r3;
|
|
and.b32 %r185, %r270, 4;
|
|
setp.eq.s32 %p150, %r185, 0;
|
|
@%p150 bra BB0_109;
|
|
|
|
ld.global.u32 %r186, [additive];
|
|
setp.eq.s32 %p151, %r186, 0;
|
|
mov.f32 %f1372, 0f3F800000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs23, %f1372;}
|
|
|
|
// inline asm
|
|
@%p151 bra BB0_108;
|
|
|
|
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 %f1373, %rs30;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1374, %rs31;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1375, %rs32;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd63), _rt_buffer_get_64, (%rd58, %r28, %r190, %rd4, %rd5, %rd13, %rd13);
|
|
// inline asm
|
|
add.f32 %f1376, %f1498, %f1373;
|
|
add.f32 %f1377, %f1497, %f1374;
|
|
add.f32 %f1378, %f1496, %f1375;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs29, %f1378;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs28, %f1377;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs27, %f1376;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd63], {%rs27, %rs28, %rs29, %rs23};
|
|
bra.uni BB0_109;
|
|
|
|
BB0_108:
|
|
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, %f1496;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs35, %f1497;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs34, %f1498;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd70], {%rs34, %rs35, %rs36, %rs23};
|
|
|
|
BB0_109:
|
|
mov.f32 %f1383, 0f34000000;
|
|
max.f32 %f1384, %f1495, %f1383;
|
|
div.rn.f32 %f1385, %f1492, %f1384;
|
|
max.f32 %f1386, %f1494, %f1383;
|
|
div.rn.f32 %f1387, %f1491, %f1386;
|
|
max.f32 %f1388, %f1493, %f1383;
|
|
div.rn.f32 %f1389, %f1490, %f1388;
|
|
fma.rn.f32 %f276, %f1385, 0f3F000000, 0f3F000000;
|
|
fma.rn.f32 %f277, %f1387, 0f3F000000, 0f3F000000;
|
|
fma.rn.f32 %f278, %f1389, 0f3F000000, 0f3F000000;
|
|
div.rn.f32 %f1390, %f1489, %f1384;
|
|
div.rn.f32 %f1391, %f1488, %f1386;
|
|
div.rn.f32 %f1392, %f1487, %f1388;
|
|
fma.rn.f32 %f279, %f1390, 0f3F000000, 0f3F000000;
|
|
fma.rn.f32 %f280, %f1391, 0f3F000000, 0f3F000000;
|
|
fma.rn.f32 %f281, %f1392, 0f3F000000, 0f3F000000;
|
|
div.rn.f32 %f1393, %f1486, %f1384;
|
|
div.rn.f32 %f1394, %f1485, %f1386;
|
|
div.rn.f32 %f1395, %f1484, %f1388;
|
|
fma.rn.f32 %f282, %f1393, 0f3F000000, 0f3F000000;
|
|
fma.rn.f32 %f283, %f1394, 0f3F000000, 0f3F000000;
|
|
fma.rn.f32 %f284, %f1395, 0f3F000000, 0f3F000000;
|
|
ld.global.u32 %r193, [additive];
|
|
setp.eq.s32 %p152, %r193, 0;
|
|
mov.f32 %f1382, 0f3F800000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs37, %f1382;}
|
|
|
|
// inline asm
|
|
@%p152 bra BB0_111;
|
|
|
|
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 %f1396, %rs44;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1397, %rs45;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1398, %rs46;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd83), _rt_buffer_get_64, (%rd78, %r28, %r197, %rd4, %rd5, %rd13, %rd13);
|
|
// inline asm
|
|
add.f32 %f1399, %f1495, %f1396;
|
|
add.f32 %f1400, %f1494, %f1397;
|
|
add.f32 %f1401, %f1493, %f1398;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs43, %f1401;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs42, %f1400;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs41, %f1399;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd83], {%rs41, %rs42, %rs43, %rs37};
|
|
bra.uni BB0_112;
|
|
|
|
BB0_111:
|
|
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, %f1493;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs49, %f1494;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs48, %f1495;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd90], {%rs48, %rs49, %rs50, %rs37};
|
|
|
|
BB0_112:
|
|
ld.global.u32 %r200, [additive];
|
|
setp.eq.s32 %p153, %r200, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs51, %f1382;}
|
|
|
|
// inline asm
|
|
@%p153 bra BB0_114;
|
|
|
|
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 %f1406, %rs58;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1407, %rs59;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1408, %rs60;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd103), _rt_buffer_get_64, (%rd98, %r28, %r204, %rd4, %rd5, %rd13, %rd13);
|
|
// inline asm
|
|
add.f32 %f1409, %f276, %f1406;
|
|
add.f32 %f1410, %f277, %f1407;
|
|
add.f32 %f1411, %f278, %f1408;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs57, %f1411;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs56, %f1410;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs55, %f1409;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd103], {%rs55, %rs56, %rs57, %rs51};
|
|
bra.uni BB0_115;
|
|
|
|
BB0_114:
|
|
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, %f278;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs63, %f277;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs62, %f276;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd110], {%rs62, %rs63, %rs64, %rs51};
|
|
|
|
BB0_115:
|
|
ld.global.u32 %r207, [additive];
|
|
setp.eq.s32 %p154, %r207, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs65, %f1382;}
|
|
|
|
// inline asm
|
|
@%p154 bra BB0_117;
|
|
|
|
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 %f1416, %rs72;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1417, %rs73;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1418, %rs74;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd123), _rt_buffer_get_64, (%rd118, %r28, %r211, %rd4, %rd5, %rd13, %rd13);
|
|
// inline asm
|
|
add.f32 %f1419, %f279, %f1416;
|
|
add.f32 %f1420, %f280, %f1417;
|
|
add.f32 %f1421, %f281, %f1418;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs71, %f1421;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs70, %f1420;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs69, %f1419;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd123], {%rs69, %rs70, %rs71, %rs65};
|
|
bra.uni BB0_118;
|
|
|
|
BB0_117:
|
|
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, %f281;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs77, %f280;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs76, %f279;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd130], {%rs76, %rs77, %rs78, %rs65};
|
|
|
|
BB0_118:
|
|
ld.global.u32 %r214, [additive];
|
|
setp.eq.s32 %p155, %r214, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs79, %f1382;}
|
|
|
|
// inline asm
|
|
@%p155 bra BB0_120;
|
|
|
|
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 %f1426, %rs86;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1427, %rs87;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1428, %rs88;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd143), _rt_buffer_get_64, (%rd138, %r28, %r218, %rd4, %rd5, %rd13, %rd13);
|
|
// inline asm
|
|
add.f32 %f1429, %f282, %f1426;
|
|
add.f32 %f1430, %f283, %f1427;
|
|
add.f32 %f1431, %f284, %f1428;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs85, %f1431;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs84, %f1430;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs83, %f1429;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd143], {%rs83, %rs84, %rs85, %rs79};
|
|
bra.uni BB0_141;
|
|
|
|
BB0_120:
|
|
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, %f284;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs91, %f283;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs90, %f282;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd150], {%rs90, %rs91, %rs92, %rs79};
|
|
|
|
BB0_141:
|
|
ret;
|
|
}
|
|
|
|
|