1379 lines
40 KiB
Plaintext
1379 lines
40 KiB
Plaintext
|
//
|
||
|
|
// Generated by NVIDIA NVVM Compiler
|
||
|
|
//
|
||
|
|
// Compiler Build ID: CL-23083092
|
||
|
|
// Cuda compilation tools, release 9.1, V9.1.85
|
||
|
|
// Based on LLVM 3.4svn
|
||
|
|
//
|
||
|
|
|
||
|
|
.version 6.1
|
||
|
|
.target sm_30
|
||
|
|
.address_size 64
|
||
|
|
|
||
|
|
// .globl _Z6oxMainv
|
||
|
|
.global .align 8 .b8 pixelID[8];
|
||
|
|
.global .align 8 .b8 resolution[8];
|
||
|
|
.global .align 4 .b8 normal[12];
|
||
|
|
.global .align 4 .b8 camPos[12];
|
||
|
|
.global .align 4 .b8 root[4];
|
||
|
|
.global .align 4 .u32 imageEnabled;
|
||
|
|
.global .texref lightmap;
|
||
|
|
.global .align 16 .b8 tileInfo[16];
|
||
|
|
.global .align 4 .u32 additive;
|
||
|
|
.global .align 1 .b8 image[1];
|
||
|
|
.global .align 1 .b8 image_HDR[1];
|
||
|
|
.global .align 1 .b8 image_HDR2[1];
|
||
|
|
.global .align 1 .b8 uvpos[1];
|
||
|
|
.global .align 1 .b8 uvnormal[1];
|
||
|
|
.global .align 1 .b8 rnd_seeds[1];
|
||
|
|
.global .align 4 .u32 sky;
|
||
|
|
.global .align 4 .b8 skyColor[12];
|
||
|
|
.global .align 4 .u32 samples;
|
||
|
|
.global .align 4 .u32 hemispherical;
|
||
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo7pixelIDE[8] = {82, 97, 121, 0, 8, 0, 0, 0};
|
||
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo10resolutionE[8] = {82, 97, 121, 0, 8, 0, 0, 0};
|
||
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo6normalE[8] = {82, 97, 121, 0, 12, 0, 0, 0};
|
||
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo6camPosE[8] = {82, 97, 121, 0, 12, 0, 0, 0};
|
||
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo4rootE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
|
||
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo12imageEnabledE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
|
||
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo8tileInfoE[8] = {82, 97, 121, 0, 16, 0, 0, 0};
|
||
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo8additiveE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
|
||
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo3skyE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
|
||
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo8skyColorE[8] = {82, 97, 121, 0, 12, 0, 0, 0};
|
||
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo7samplesE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
|
||
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo13hemisphericalE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
|
||
|
|
.global .align 8 .u64 _ZN21rti_internal_register20reg_bitness_detectorE;
|
||
|
|
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail0E;
|
||
|
|
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail1E;
|
||
|
|
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail2E;
|
||
|
|
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail3E;
|
||
|
|
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail4E;
|
||
|
|
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail5E;
|
||
|
|
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail6E;
|
||
|
|
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail7E;
|
||
|
|
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail8E;
|
||
|
|
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail9E;
|
||
|
|
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail0E;
|
||
|
|
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail1E;
|
||
|
|
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail2E;
|
||
|
|
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail3E;
|
||
|
|
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail4E;
|
||
|
|
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail5E;
|
||
|
|
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail6E;
|
||
|
|
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail7E;
|
||
|
|
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail8E;
|
||
|
|
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail9E;
|
||
|
|
.global .align 4 .u32 _ZN21rti_internal_register14reg_rayIndex_xE;
|
||
|
|
.global .align 4 .u32 _ZN21rti_internal_register14reg_rayIndex_yE;
|
||
|
|
.global .align 4 .u32 _ZN21rti_internal_register14reg_rayIndex_zE;
|
||
|
|
.global .align 8 .b8 _ZN21rti_internal_typename7pixelIDE[6] = {117, 105, 110, 116, 50, 0};
|
||
|
|
.global .align 8 .b8 _ZN21rti_internal_typename10resolutionE[6] = {117, 105, 110, 116, 50, 0};
|
||
|
|
.global .align 8 .b8 _ZN21rti_internal_typename6normalE[7] = {102, 108, 111, 97, 116, 51, 0};
|
||
|
|
.global .align 8 .b8 _ZN21rti_internal_typename6camPosE[7] = {102, 108, 111, 97, 116, 51, 0};
|
||
|
|
.global .align 16 .b8 _ZN21rti_internal_typename4rootE[9] = {114, 116, 79, 98, 106, 101, 99, 116, 0};
|
||
|
|
.global .align 4 .b8 _ZN21rti_internal_typename12imageEnabledE[4] = {105, 110, 116, 0};
|
||
|
|
.global .align 8 .b8 _ZN21rti_internal_typename8tileInfoE[6] = {117, 105, 110, 116, 52, 0};
|
||
|
|
.global .align 4 .b8 _ZN21rti_internal_typename8additiveE[4] = {105, 110, 116, 0};
|
||
|
|
.global .align 4 .b8 _ZN21rti_internal_typename3skyE[4] = {105, 110, 116, 0};
|
||
|
|
.global .align 8 .b8 _ZN21rti_internal_typename8skyColorE[7] = {102, 108, 111, 97, 116, 51, 0};
|
||
|
|
.global .align 4 .b8 _ZN21rti_internal_typename7samplesE[4] = {105, 110, 116, 0};
|
||
|
|
.global .align 4 .b8 _ZN21rti_internal_typename13hemisphericalE[4] = {105, 110, 116, 0};
|
||
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum7pixelIDE = 4919;
|
||
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum10resolutionE = 4919;
|
||
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum6normalE = 4919;
|
||
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum6camPosE = 4919;
|
||
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum4rootE = 4919;
|
||
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum12imageEnabledE = 4919;
|
||
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum8tileInfoE = 4919;
|
||
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum8additiveE = 4919;
|
||
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum3skyE = 4919;
|
||
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum8skyColorE = 4919;
|
||
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum7samplesE = 4919;
|
||
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum13hemisphericalE = 4919;
|
||
|
|
.global .align 16 .b8 _ZN21rti_internal_semantic7pixelIDE[14] = {114, 116, 76, 97, 117, 110, 99, 104, 73, 110, 100, 101, 120, 0};
|
||
|
|
.global .align 16 .b8 _ZN21rti_internal_semantic10resolutionE[12] = {114, 116, 76, 97, 117, 110, 99, 104, 68, 105, 109, 0};
|
||
|
|
.global .align 16 .b8 _ZN21rti_internal_semantic6normalE[17] = {97, 116, 116, 114, 105, 98, 117, 116, 101, 32, 110, 111, 114, 109, 97, 108, 0};
|
||
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic6camPosE[1];
|
||
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic4rootE[1];
|
||
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic12imageEnabledE[1];
|
||
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic8tileInfoE[1];
|
||
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic8additiveE[1];
|
||
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic3skyE[1];
|
||
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic8skyColorE[1];
|
||
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic7samplesE[1];
|
||
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic13hemisphericalE[1];
|
||
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation7pixelIDE[1];
|
||
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation10resolutionE[1];
|
||
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation6normalE[1];
|
||
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation6camPosE[1];
|
||
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation4rootE[1];
|
||
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation12imageEnabledE[1];
|
||
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation8tileInfoE[1];
|
||
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation8additiveE[1];
|
||
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation3skyE[1];
|
||
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation8skyColorE[1];
|
||
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation7samplesE[1];
|
||
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation13hemisphericalE[1];
|
||
|
|
.const .align 4 .b8 __cudart_i2opi_f[24] = {65, 144, 67, 60, 153, 149, 98, 219, 192, 221, 52, 245, 209, 87, 39, 252, 41, 21, 68, 78, 110, 131, 249, 162};
|
||
|
|
|
||
|
|
.visible .entry _Z6oxMainv(
|
||
|
|
|
||
|
|
)
|
||
|
|
{
|
||
|
|
.local .align 4 .b8 __local_depot0[40];
|
||
|
|
.reg .b64 %SP;
|
||
|
|
.reg .b64 %SPL;
|
||
|
|
.reg .pred %p<97>;
|
||
|
|
.reg .b16 %rs<42>;
|
||
|
|
.reg .f32 %f<596>;
|
||
|
|
.reg .b32 %r<304>;
|
||
|
|
.reg .b64 %rd<114>;
|
||
|
|
|
||
|
|
|
||
|
|
mov.u64 %rd113, __local_depot0;
|
||
|
|
cvta.local.u64 %SP, %rd113;
|
||
|
|
ld.global.u32 %r1, [samples];
|
||
|
|
ld.global.v2.u32 {%r93, %r94}, [pixelID];
|
||
|
|
cvt.u64.u32 %rd21, %r93;
|
||
|
|
cvt.u64.u32 %rd22, %r94;
|
||
|
|
mov.u64 %rd25, uvnormal;
|
||
|
|
cvta.global.u64 %rd20, %rd25;
|
||
|
|
mov.u32 %r91, 2;
|
||
|
|
mov.u32 %r92, 4;
|
||
|
|
mov.u64 %rd24, 0;
|
||
|
|
// inline asm
|
||
|
|
call (%rd19), _rt_buffer_get_64, (%rd20, %r91, %r92, %rd21, %rd22, %rd24, %rd24);
|
||
|
|
// inline asm
|
||
|
|
ld.u32 %r2, [%rd19];
|
||
|
|
shr.u32 %r97, %r2, 16;
|
||
|
|
cvt.u16.u32 %rs1, %r97;
|
||
|
|
and.b16 %rs3, %rs1, 255;
|
||
|
|
cvt.u16.u32 %rs4, %r2;
|
||
|
|
or.b16 %rs5, %rs4, %rs3;
|
||
|
|
setp.eq.s16 %p4, %rs5, 0;
|
||
|
|
mov.f32 %f560, 0f00000000;
|
||
|
|
mov.f32 %f561, %f560;
|
||
|
|
mov.f32 %f562, %f560;
|
||
|
|
@%p4 bra BB0_2;
|
||
|
|
|
||
|
|
ld.u8 %rs6, [%rd19+1];
|
||
|
|
and.b16 %rs8, %rs4, 255;
|
||
|
|
cvt.rn.f32.u16 %f121, %rs8;
|
||
|
|
div.rn.f32 %f122, %f121, 0f437F0000;
|
||
|
|
fma.rn.f32 %f123, %f122, 0f40000000, 0fBF800000;
|
||
|
|
cvt.rn.f32.u16 %f124, %rs6;
|
||
|
|
div.rn.f32 %f125, %f124, 0f437F0000;
|
||
|
|
fma.rn.f32 %f126, %f125, 0f40000000, 0fBF800000;
|
||
|
|
cvt.rn.f32.u16 %f127, %rs3;
|
||
|
|
div.rn.f32 %f128, %f127, 0f437F0000;
|
||
|
|
fma.rn.f32 %f129, %f128, 0f40000000, 0fBF800000;
|
||
|
|
mul.f32 %f130, %f126, %f126;
|
||
|
|
fma.rn.f32 %f131, %f123, %f123, %f130;
|
||
|
|
fma.rn.f32 %f132, %f129, %f129, %f131;
|
||
|
|
sqrt.rn.f32 %f133, %f132;
|
||
|
|
rcp.rn.f32 %f134, %f133;
|
||
|
|
mul.f32 %f560, %f123, %f134;
|
||
|
|
mul.f32 %f561, %f126, %f134;
|
||
|
|
mul.f32 %f562, %f129, %f134;
|
||
|
|
|
||
|
|
BB0_2:
|
||
|
|
ld.global.v2.u32 {%r98, %r99}, [pixelID];
|
||
|
|
ld.global.v2.u32 {%r101, %r102}, [tileInfo];
|
||
|
|
add.s32 %r3, %r98, %r101;
|
||
|
|
add.s32 %r4, %r99, %r102;
|
||
|
|
setp.eq.f32 %p5, %f561, 0f00000000;
|
||
|
|
setp.eq.f32 %p6, %f560, 0f00000000;
|
||
|
|
and.pred %p7, %p6, %p5;
|
||
|
|
setp.eq.f32 %p8, %f562, 0f00000000;
|
||
|
|
and.pred %p9, %p7, %p8;
|
||
|
|
@%p9 bra BB0_95;
|
||
|
|
bra.uni BB0_3;
|
||
|
|
|
||
|
|
BB0_95:
|
||
|
|
ld.global.u32 %r303, [imageEnabled];
|
||
|
|
and.b32 %r263, %r303, 1;
|
||
|
|
setp.eq.b32 %p94, %r263, 1;
|
||
|
|
@!%p94 bra BB0_97;
|
||
|
|
bra.uni BB0_96;
|
||
|
|
|
||
|
|
BB0_96:
|
||
|
|
cvt.u64.u32 %rd81, %r3;
|
||
|
|
cvt.u64.u32 %rd82, %r4;
|
||
|
|
mov.u64 %rd85, image;
|
||
|
|
cvta.global.u64 %rd80, %rd85;
|
||
|
|
// inline asm
|
||
|
|
call (%rd79), _rt_buffer_get_64, (%rd80, %r91, %r92, %rd81, %rd82, %rd24, %rd24);
|
||
|
|
// inline asm
|
||
|
|
mov.u16 %rs28, 0;
|
||
|
|
st.v4.u8 [%rd79], {%rs28, %rs28, %rs28, %rs28};
|
||
|
|
ld.global.u32 %r303, [imageEnabled];
|
||
|
|
|
||
|
|
BB0_97:
|
||
|
|
and.b32 %r266, %r303, 4;
|
||
|
|
setp.eq.s32 %p95, %r266, 0;
|
||
|
|
@%p95 bra BB0_101;
|
||
|
|
|
||
|
|
ld.global.u32 %r267, [additive];
|
||
|
|
setp.eq.s32 %p96, %r267, 0;
|
||
|
|
cvt.u64.u32 %rd17, %r3;
|
||
|
|
cvt.u64.u32 %rd18, %r4;
|
||
|
|
@%p96 bra BB0_100;
|
||
|
|
|
||
|
|
mov.u64 %rd98, image_HDR;
|
||
|
|
cvta.global.u64 %rd87, %rd98;
|
||
|
|
mov.u32 %r271, 8;
|
||
|
|
// inline asm
|
||
|
|
call (%rd86), _rt_buffer_get_64, (%rd87, %r91, %r271, %rd17, %rd18, %rd24, %rd24);
|
||
|
|
// inline asm
|
||
|
|
ld.v4.u16 {%rs35, %rs36, %rs37, %rs38}, [%rd86];
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f532, %rs35;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f533, %rs36;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f534, %rs37;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
call (%rd92), _rt_buffer_get_64, (%rd87, %r91, %r271, %rd17, %rd18, %rd24, %rd24);
|
||
|
|
// inline asm
|
||
|
|
add.f32 %f535, %f532, 0f00000000;
|
||
|
|
add.f32 %f536, %f533, 0f00000000;
|
||
|
|
add.f32 %f537, %f534, 0f00000000;
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs34, %f537;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs33, %f536;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs32, %f535;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
mov.u16 %rs39, 0;
|
||
|
|
st.v4.u16 [%rd92], {%rs32, %rs33, %rs34, %rs39};
|
||
|
|
bra.uni BB0_101;
|
||
|
|
|
||
|
|
BB0_3:
|
||
|
|
ld.global.v2.u32 {%r110, %r111}, [pixelID];
|
||
|
|
cvt.u64.u32 %rd28, %r110;
|
||
|
|
cvt.u64.u32 %rd29, %r111;
|
||
|
|
mov.u64 %rd38, uvpos;
|
||
|
|
cvta.global.u64 %rd27, %rd38;
|
||
|
|
mov.u32 %r107, 12;
|
||
|
|
// inline asm
|
||
|
|
call (%rd26), _rt_buffer_get_64, (%rd27, %r91, %r107, %rd28, %rd29, %rd24, %rd24);
|
||
|
|
// inline asm
|
||
|
|
ld.f32 %f138, [%rd26+8];
|
||
|
|
ld.f32 %f139, [%rd26+4];
|
||
|
|
ld.f32 %f140, [%rd26];
|
||
|
|
mul.f32 %f141, %f140, 0f3456BF95;
|
||
|
|
mul.f32 %f142, %f139, 0f3456BF95;
|
||
|
|
mul.f32 %f143, %f138, 0f3456BF95;
|
||
|
|
abs.f32 %f144, %f560;
|
||
|
|
div.rn.f32 %f145, %f141, %f144;
|
||
|
|
abs.f32 %f146, %f561;
|
||
|
|
div.rn.f32 %f147, %f142, %f146;
|
||
|
|
abs.f32 %f148, %f562;
|
||
|
|
div.rn.f32 %f149, %f143, %f148;
|
||
|
|
abs.f32 %f150, %f145;
|
||
|
|
abs.f32 %f151, %f147;
|
||
|
|
abs.f32 %f152, %f149;
|
||
|
|
mov.f32 %f153, 0f38D1B717;
|
||
|
|
max.f32 %f154, %f150, %f153;
|
||
|
|
max.f32 %f155, %f151, %f153;
|
||
|
|
max.f32 %f156, %f152, %f153;
|
||
|
|
fma.rn.f32 %f7, %f560, %f154, %f140;
|
||
|
|
fma.rn.f32 %f8, %f561, %f155, %f139;
|
||
|
|
fma.rn.f32 %f9, %f562, %f156, %f138;
|
||
|
|
ld.global.u32 %r5, [hemispherical];
|
||
|
|
setp.gt.f32 %p10, %f144, %f148;
|
||
|
|
neg.f32 %f157, %f561;
|
||
|
|
selp.f32 %f158, %f157, 0f00000000, %p10;
|
||
|
|
neg.f32 %f159, %f562;
|
||
|
|
selp.f32 %f160, %f560, %f159, %p10;
|
||
|
|
selp.f32 %f161, 0f00000000, %f561, %p10;
|
||
|
|
mul.f32 %f162, %f160, %f160;
|
||
|
|
fma.rn.f32 %f163, %f158, %f158, %f162;
|
||
|
|
fma.rn.f32 %f164, %f161, %f161, %f163;
|
||
|
|
sqrt.rn.f32 %f165, %f164;
|
||
|
|
rcp.rn.f32 %f166, %f165;
|
||
|
|
mul.f32 %f10, %f158, %f166;
|
||
|
|
mul.f32 %f11, %f160, %f166;
|
||
|
|
mul.f32 %f12, %f161, %f166;
|
||
|
|
ld.global.v2.u32 {%r114, %r115}, [pixelID];
|
||
|
|
cvt.u64.u32 %rd34, %r114;
|
||
|
|
cvt.u64.u32 %rd35, %r115;
|
||
|
|
mov.u64 %rd39, rnd_seeds;
|
||
|
|
cvta.global.u64 %rd33, %rd39;
|
||
|
|
// inline asm
|
||
|
|
call (%rd32), _rt_buffer_get_64, (%rd33, %r91, %r92, %rd34, %rd35, %rd24, %rd24);
|
||
|
|
// inline asm
|
||
|
|
mov.f32 %f583, 0f00000000;
|
||
|
|
setp.lt.s32 %p11, %r1, 1;
|
||
|
|
mov.f32 %f582, %f583;
|
||
|
|
mov.f32 %f581, %f583;
|
||
|
|
@%p11 bra BB0_56;
|
||
|
|
|
||
|
|
cvt.rn.f32.s32 %f170, %r1;
|
||
|
|
rcp.rn.f32 %f13, %f170;
|
||
|
|
ld.u32 %r281, [%rd32];
|
||
|
|
mul.f32 %f14, %f7, 0f3456BF95;
|
||
|
|
mul.f32 %f15, %f8, 0f3456BF95;
|
||
|
|
mul.f32 %f16, %f9, 0f3456BF95;
|
||
|
|
mul.f32 %f171, %f560, %f11;
|
||
|
|
mul.f32 %f172, %f561, %f10;
|
||
|
|
sub.f32 %f17, %f172, %f171;
|
||
|
|
mul.f32 %f173, %f562, %f10;
|
||
|
|
mul.f32 %f174, %f560, %f12;
|
||
|
|
sub.f32 %f18, %f174, %f173;
|
||
|
|
mul.f32 %f175, %f561, %f12;
|
||
|
|
mul.f32 %f176, %f562, %f11;
|
||
|
|
sub.f32 %f19, %f176, %f175;
|
||
|
|
mov.f32 %f583, 0f00000000;
|
||
|
|
mov.u32 %r118, 0;
|
||
|
|
abs.f32 %f261, %f15;
|
||
|
|
abs.f32 %f262, %f14;
|
||
|
|
max.f32 %f263, %f262, %f261;
|
||
|
|
abs.f32 %f264, %f16;
|
||
|
|
max.f32 %f265, %f263, %f264;
|
||
|
|
mov.u32 %r278, %r118;
|
||
|
|
mov.f32 %f582, %f583;
|
||
|
|
mov.f32 %f581, %f583;
|
||
|
|
|
||
|
|
BB0_5:
|
||
|
|
cvt.rn.f32.s32 %f23, %r278;
|
||
|
|
mov.u32 %r280, %r118;
|
||
|
|
|
||
|
|
BB0_6:
|
||
|
|
mad.lo.s32 %r120, %r281, 1664525, 1013904223;
|
||
|
|
and.b32 %r121, %r120, 16777215;
|
||
|
|
cvt.rn.f32.u32 %f177, %r121;
|
||
|
|
fma.rn.f32 %f178, %f177, 0f33800000, %f23;
|
||
|
|
mul.f32 %f179, %f13, %f178;
|
||
|
|
mad.lo.s32 %r281, %r120, 1664525, 1013904223;
|
||
|
|
and.b32 %r122, %r281, 16777215;
|
||
|
|
cvt.rn.f32.u32 %f180, %r122;
|
||
|
|
cvt.rn.f32.s32 %f181, %r280;
|
||
|
|
fma.rn.f32 %f182, %f180, 0f33800000, %f181;
|
||
|
|
mul.f32 %f183, %f13, %f182;
|
||
|
|
sqrt.rn.f32 %f27, %f179;
|
||
|
|
mul.f32 %f575, %f183, 0f40C90FDB;
|
||
|
|
abs.f32 %f29, %f575;
|
||
|
|
setp.neu.f32 %p12, %f29, 0f7F800000;
|
||
|
|
mov.f32 %f569, %f575;
|
||
|
|
@%p12 bra BB0_8;
|
||
|
|
|
||
|
|
mov.f32 %f184, 0f00000000;
|
||
|
|
mul.rn.f32 %f569, %f575, %f184;
|
||
|
|
|
||
|
|
BB0_8:
|
||
|
|
mul.f32 %f185, %f569, 0f3F22F983;
|
||
|
|
cvt.rni.s32.f32 %r291, %f185;
|
||
|
|
cvt.rn.f32.s32 %f186, %r291;
|
||
|
|
neg.f32 %f187, %f186;
|
||
|
|
mov.f32 %f188, 0f3FC90FDA;
|
||
|
|
fma.rn.f32 %f189, %f187, %f188, %f569;
|
||
|
|
mov.f32 %f190, 0f33A22168;
|
||
|
|
fma.rn.f32 %f191, %f187, %f190, %f189;
|
||
|
|
mov.f32 %f192, 0f27C234C5;
|
||
|
|
fma.rn.f32 %f570, %f187, %f192, %f191;
|
||
|
|
abs.f32 %f193, %f569;
|
||
|
|
setp.leu.f32 %p13, %f193, 0f47CE4780;
|
||
|
|
@%p13 bra BB0_19;
|
||
|
|
|
||
|
|
mov.b32 %r13, %f569;
|
||
|
|
shr.u32 %r14, %r13, 23;
|
||
|
|
shl.b32 %r125, %r13, 8;
|
||
|
|
or.b32 %r15, %r125, -2147483648;
|
||
|
|
add.u64 %rd41, %SP, 0;
|
||
|
|
cvta.to.local.u64 %rd110, %rd41;
|
||
|
|
mov.u32 %r283, 0;
|
||
|
|
mov.u64 %rd109, __cudart_i2opi_f;
|
||
|
|
mov.u32 %r282, -6;
|
||
|
|
|
||
|
|
BB0_10:
|
||
|
|
.pragma "nounroll";
|
||
|
|
ld.const.u32 %r128, [%rd109];
|
||
|
|
// inline asm
|
||
|
|
{
|
||
|
|
mad.lo.cc.u32 %r126, %r128, %r15, %r283;
|
||
|
|
madc.hi.u32 %r283, %r128, %r15, 0;
|
||
|
|
}
|
||
|
|
// inline asm
|
||
|
|
st.local.u32 [%rd110], %r126;
|
||
|
|
add.s64 %rd110, %rd110, 4;
|
||
|
|
add.s64 %rd109, %rd109, 4;
|
||
|
|
add.s32 %r282, %r282, 1;
|
||
|
|
setp.ne.s32 %p14, %r282, 0;
|
||
|
|
@%p14 bra BB0_10;
|
||
|
|
|
||
|
|
and.b32 %r131, %r14, 255;
|
||
|
|
add.s32 %r132, %r131, -128;
|
||
|
|
shr.u32 %r133, %r132, 5;
|
||
|
|
and.b32 %r20, %r13, -2147483648;
|
||
|
|
cvta.to.local.u64 %rd43, %rd41;
|
||
|
|
st.local.u32 [%rd43+24], %r283;
|
||
|
|
mov.u32 %r134, 6;
|
||
|
|
sub.s32 %r135, %r134, %r133;
|
||
|
|
mul.wide.s32 %rd44, %r135, 4;
|
||
|
|
add.s64 %rd8, %rd43, %rd44;
|
||
|
|
ld.local.u32 %r284, [%rd8];
|
||
|
|
ld.local.u32 %r285, [%rd8+-4];
|
||
|
|
and.b32 %r23, %r14, 31;
|
||
|
|
setp.eq.s32 %p15, %r23, 0;
|
||
|
|
@%p15 bra BB0_13;
|
||
|
|
|
||
|
|
mov.u32 %r136, 32;
|
||
|
|
sub.s32 %r137, %r136, %r23;
|
||
|
|
shr.u32 %r138, %r285, %r137;
|
||
|
|
shl.b32 %r139, %r284, %r23;
|
||
|
|
add.s32 %r284, %r138, %r139;
|
||
|
|
ld.local.u32 %r140, [%rd8+-8];
|
||
|
|
shr.u32 %r141, %r140, %r137;
|
||
|
|
shl.b32 %r142, %r285, %r23;
|
||
|
|
add.s32 %r285, %r141, %r142;
|
||
|
|
|
||
|
|
BB0_13:
|
||
|
|
shr.u32 %r143, %r285, 30;
|
||
|
|
shl.b32 %r144, %r284, 2;
|
||
|
|
add.s32 %r286, %r143, %r144;
|
||
|
|
shl.b32 %r29, %r285, 2;
|
||
|
|
shr.u32 %r145, %r286, 31;
|
||
|
|
shr.u32 %r146, %r284, 30;
|
||
|
|
add.s32 %r30, %r145, %r146;
|
||
|
|
setp.eq.s32 %p16, %r145, 0;
|
||
|
|
@%p16 bra BB0_14;
|
||
|
|
bra.uni BB0_15;
|
||
|
|
|
||
|
|
BB0_14:
|
||
|
|
mov.u32 %r287, %r20;
|
||
|
|
mov.u32 %r288, %r29;
|
||
|
|
bra.uni BB0_16;
|
||
|
|
|
||
|
|
BB0_15:
|
||
|
|
not.b32 %r147, %r286;
|
||
|
|
neg.s32 %r288, %r29;
|
||
|
|
setp.eq.s32 %p17, %r29, 0;
|
||
|
|
selp.u32 %r148, 1, 0, %p17;
|
||
|
|
add.s32 %r286, %r148, %r147;
|
||
|
|
xor.b32 %r287, %r20, -2147483648;
|
||
|
|
|
||
|
|
BB0_16:
|
||
|
|
clz.b32 %r290, %r286;
|
||
|
|
setp.eq.s32 %p18, %r290, 0;
|
||
|
|
shl.b32 %r149, %r286, %r290;
|
||
|
|
mov.u32 %r150, 32;
|
||
|
|
sub.s32 %r151, %r150, %r290;
|
||
|
|
shr.u32 %r152, %r288, %r151;
|
||
|
|
add.s32 %r153, %r152, %r149;
|
||
|
|
selp.b32 %r38, %r286, %r153, %p18;
|
||
|
|
mov.u32 %r154, -921707870;
|
||
|
|
mul.hi.u32 %r289, %r38, %r154;
|
||
|
|
setp.eq.s32 %p19, %r20, 0;
|
||
|
|
neg.s32 %r155, %r30;
|
||
|
|
selp.b32 %r291, %r30, %r155, %p19;
|
||
|
|
setp.lt.s32 %p20, %r289, 1;
|
||
|
|
@%p20 bra BB0_18;
|
||
|
|
|
||
|
|
mul.lo.s32 %r156, %r38, -921707870;
|
||
|
|
shr.u32 %r157, %r156, 31;
|
||
|
|
shl.b32 %r158, %r289, 1;
|
||
|
|
add.s32 %r289, %r157, %r158;
|
||
|
|
add.s32 %r290, %r290, 1;
|
||
|
|
|
||
|
|
BB0_18:
|
||
|
|
mov.u32 %r159, 126;
|
||
|
|
sub.s32 %r160, %r159, %r290;
|
||
|
|
shl.b32 %r161, %r160, 23;
|
||
|
|
add.s32 %r162, %r289, 1;
|
||
|
|
shr.u32 %r163, %r162, 7;
|
||
|
|
add.s32 %r164, %r163, 1;
|
||
|
|
shr.u32 %r165, %r164, 1;
|
||
|
|
add.s32 %r166, %r165, %r161;
|
||
|
|
or.b32 %r167, %r166, %r287;
|
||
|
|
mov.b32 %f570, %r167;
|
||
|
|
|
||
|
|
BB0_19:
|
||
|
|
mul.rn.f32 %f35, %f570, %f570;
|
||
|
|
add.s32 %r46, %r291, 1;
|
||
|
|
and.b32 %r47, %r46, 1;
|
||
|
|
setp.eq.s32 %p21, %r47, 0;
|
||
|
|
@%p21 bra BB0_21;
|
||
|
|
bra.uni BB0_20;
|
||
|
|
|
||
|
|
BB0_21:
|
||
|
|
mov.f32 %f196, 0f3C08839E;
|
||
|
|
mov.f32 %f197, 0fB94CA1F9;
|
||
|
|
fma.rn.f32 %f571, %f197, %f35, %f196;
|
||
|
|
bra.uni BB0_22;
|
||
|
|
|
||
|
|
BB0_20:
|
||
|
|
mov.f32 %f194, 0fBAB6061A;
|
||
|
|
mov.f32 %f195, 0f37CCF5CE;
|
||
|
|
fma.rn.f32 %f571, %f195, %f35, %f194;
|
||
|
|
|
||
|
|
BB0_22:
|
||
|
|
@%p21 bra BB0_24;
|
||
|
|
bra.uni BB0_23;
|
||
|
|
|
||
|
|
BB0_24:
|
||
|
|
mov.f32 %f201, 0fBE2AAAA3;
|
||
|
|
fma.rn.f32 %f202, %f571, %f35, %f201;
|
||
|
|
mov.f32 %f203, 0f00000000;
|
||
|
|
fma.rn.f32 %f572, %f202, %f35, %f203;
|
||
|
|
bra.uni BB0_25;
|
||
|
|
|
||
|
|
BB0_23:
|
||
|
|
mov.f32 %f198, 0f3D2AAAA5;
|
||
|
|
fma.rn.f32 %f199, %f571, %f35, %f198;
|
||
|
|
mov.f32 %f200, 0fBF000000;
|
||
|
|
fma.rn.f32 %f572, %f199, %f35, %f200;
|
||
|
|
|
||
|
|
BB0_25:
|
||
|
|
fma.rn.f32 %f573, %f572, %f570, %f570;
|
||
|
|
@%p21 bra BB0_27;
|
||
|
|
|
||
|
|
mov.f32 %f204, 0f3F800000;
|
||
|
|
fma.rn.f32 %f573, %f572, %f35, %f204;
|
||
|
|
|
||
|
|
BB0_27:
|
||
|
|
and.b32 %r168, %r46, 2;
|
||
|
|
setp.eq.s32 %p24, %r168, 0;
|
||
|
|
@%p24 bra BB0_29;
|
||
|
|
|
||
|
|
mov.f32 %f205, 0f00000000;
|
||
|
|
mov.f32 %f206, 0fBF800000;
|
||
|
|
fma.rn.f32 %f573, %f573, %f206, %f205;
|
||
|
|
|
||
|
|
BB0_29:
|
||
|
|
@%p12 bra BB0_31;
|
||
|
|
|
||
|
|
mov.f32 %f207, 0f00000000;
|
||
|
|
mul.rn.f32 %f575, %f575, %f207;
|
||
|
|
|
||
|
|
BB0_31:
|
||
|
|
mul.f32 %f208, %f575, 0f3F22F983;
|
||
|
|
cvt.rni.s32.f32 %r301, %f208;
|
||
|
|
cvt.rn.f32.s32 %f209, %r301;
|
||
|
|
neg.f32 %f210, %f209;
|
||
|
|
fma.rn.f32 %f212, %f210, %f188, %f575;
|
||
|
|
fma.rn.f32 %f214, %f210, %f190, %f212;
|
||
|
|
fma.rn.f32 %f576, %f210, %f192, %f214;
|
||
|
|
abs.f32 %f216, %f575;
|
||
|
|
setp.leu.f32 %p26, %f216, 0f47CE4780;
|
||
|
|
@%p26 bra BB0_42;
|
||
|
|
|
||
|
|
mov.b32 %r49, %f575;
|
||
|
|
shr.u32 %r50, %r49, 23;
|
||
|
|
shl.b32 %r171, %r49, 8;
|
||
|
|
or.b32 %r51, %r171, -2147483648;
|
||
|
|
add.u64 %rd46, %SP, 0;
|
||
|
|
cvta.to.local.u64 %rd112, %rd46;
|
||
|
|
mov.u32 %r293, 0;
|
||
|
|
mov.u64 %rd111, __cudart_i2opi_f;
|
||
|
|
mov.u32 %r292, -6;
|
||
|
|
|
||
|
|
BB0_33:
|
||
|
|
.pragma "nounroll";
|
||
|
|
ld.const.u32 %r174, [%rd111];
|
||
|
|
// inline asm
|
||
|
|
{
|
||
|
|
mad.lo.cc.u32 %r172, %r174, %r51, %r293;
|
||
|
|
madc.hi.u32 %r293, %r174, %r51, 0;
|
||
|
|
}
|
||
|
|
// inline asm
|
||
|
|
st.local.u32 [%rd112], %r172;
|
||
|
|
add.s64 %rd112, %rd112, 4;
|
||
|
|
add.s64 %rd111, %rd111, 4;
|
||
|
|
add.s32 %r292, %r292, 1;
|
||
|
|
setp.ne.s32 %p27, %r292, 0;
|
||
|
|
@%p27 bra BB0_33;
|
||
|
|
|
||
|
|
and.b32 %r177, %r50, 255;
|
||
|
|
add.s32 %r178, %r177, -128;
|
||
|
|
shr.u32 %r179, %r178, 5;
|
||
|
|
and.b32 %r56, %r49, -2147483648;
|
||
|
|
cvta.to.local.u64 %rd48, %rd46;
|
||
|
|
st.local.u32 [%rd48+24], %r293;
|
||
|
|
mov.u32 %r180, 6;
|
||
|
|
sub.s32 %r181, %r180, %r179;
|
||
|
|
mul.wide.s32 %rd49, %r181, 4;
|
||
|
|
add.s64 %rd14, %rd48, %rd49;
|
||
|
|
ld.local.u32 %r294, [%rd14];
|
||
|
|
ld.local.u32 %r295, [%rd14+-4];
|
||
|
|
and.b32 %r59, %r50, 31;
|
||
|
|
setp.eq.s32 %p28, %r59, 0;
|
||
|
|
@%p28 bra BB0_36;
|
||
|
|
|
||
|
|
mov.u32 %r182, 32;
|
||
|
|
sub.s32 %r183, %r182, %r59;
|
||
|
|
shr.u32 %r184, %r295, %r183;
|
||
|
|
shl.b32 %r185, %r294, %r59;
|
||
|
|
add.s32 %r294, %r184, %r185;
|
||
|
|
ld.local.u32 %r186, [%rd14+-8];
|
||
|
|
shr.u32 %r187, %r186, %r183;
|
||
|
|
shl.b32 %r188, %r295, %r59;
|
||
|
|
add.s32 %r295, %r187, %r188;
|
||
|
|
|
||
|
|
BB0_36:
|
||
|
|
shr.u32 %r189, %r295, 30;
|
||
|
|
shl.b32 %r190, %r294, 2;
|
||
|
|
add.s32 %r296, %r189, %r190;
|
||
|
|
shl.b32 %r65, %r295, 2;
|
||
|
|
shr.u32 %r191, %r296, 31;
|
||
|
|
shr.u32 %r192, %r294, 30;
|
||
|
|
add.s32 %r66, %r191, %r192;
|
||
|
|
setp.eq.s32 %p29, %r191, 0;
|
||
|
|
@%p29 bra BB0_37;
|
||
|
|
bra.uni BB0_38;
|
||
|
|
|
||
|
|
BB0_37:
|
||
|
|
mov.u32 %r297, %r56;
|
||
|
|
mov.u32 %r298, %r65;
|
||
|
|
bra.uni BB0_39;
|
||
|
|
|
||
|
|
BB0_38:
|
||
|
|
not.b32 %r193, %r296;
|
||
|
|
neg.s32 %r298, %r65;
|
||
|
|
setp.eq.s32 %p30, %r65, 0;
|
||
|
|
selp.u32 %r194, 1, 0, %p30;
|
||
|
|
add.s32 %r296, %r194, %r193;
|
||
|
|
xor.b32 %r297, %r56, -2147483648;
|
||
|
|
|
||
|
|
BB0_39:
|
||
|
|
clz.b32 %r300, %r296;
|
||
|
|
setp.eq.s32 %p31, %r300, 0;
|
||
|
|
shl.b32 %r195, %r296, %r300;
|
||
|
|
mov.u32 %r196, 32;
|
||
|
|
sub.s32 %r197, %r196, %r300;
|
||
|
|
shr.u32 %r198, %r298, %r197;
|
||
|
|
add.s32 %r199, %r198, %r195;
|
||
|
|
selp.b32 %r74, %r296, %r199, %p31;
|
||
|
|
mov.u32 %r200, -921707870;
|
||
|
|
mul.hi.u32 %r299, %r74, %r200;
|
||
|
|
setp.eq.s32 %p32, %r56, 0;
|
||
|
|
neg.s32 %r201, %r66;
|
||
|
|
selp.b32 %r301, %r66, %r201, %p32;
|
||
|
|
setp.lt.s32 %p33, %r299, 1;
|
||
|
|
@%p33 bra BB0_41;
|
||
|
|
|
||
|
|
mul.lo.s32 %r202, %r74, -921707870;
|
||
|
|
shr.u32 %r203, %r202, 31;
|
||
|
|
shl.b32 %r204, %r299, 1;
|
||
|
|
add.s32 %r299, %r203, %r204;
|
||
|
|
add.s32 %r300, %r300, 1;
|
||
|
|
|
||
|
|
BB0_41:
|
||
|
|
mov.u32 %r205, 126;
|
||
|
|
sub.s32 %r206, %r205, %r300;
|
||
|
|
shl.b32 %r207, %r206, 23;
|
||
|
|
add.s32 %r208, %r299, 1;
|
||
|
|
shr.u32 %r209, %r208, 7;
|
||
|
|
add.s32 %r210, %r209, 1;
|
||
|
|
shr.u32 %r211, %r210, 1;
|
||
|
|
add.s32 %r212, %r211, %r207;
|
||
|
|
or.b32 %r213, %r212, %r297;
|
||
|
|
mov.b32 %f576, %r213;
|
||
|
|
|
||
|
|
BB0_42:
|
||
|
|
mul.rn.f32 %f52, %f576, %f576;
|
||
|
|
and.b32 %r82, %r301, 1;
|
||
|
|
setp.eq.s32 %p34, %r82, 0;
|
||
|
|
@%p34 bra BB0_44;
|
||
|
|
bra.uni BB0_43;
|
||
|
|
|
||
|
|
BB0_44:
|
||
|
|
mov.f32 %f219, 0f3C08839E;
|
||
|
|
mov.f32 %f220, 0fB94CA1F9;
|
||
|
|
fma.rn.f32 %f577, %f220, %f52, %f219;
|
||
|
|
bra.uni BB0_45;
|
||
|
|
|
||
|
|
BB0_43:
|
||
|
|
mov.f32 %f217, 0fBAB6061A;
|
||
|
|
mov.f32 %f218, 0f37CCF5CE;
|
||
|
|
fma.rn.f32 %f577, %f218, %f52, %f217;
|
||
|
|
|
||
|
|
BB0_45:
|
||
|
|
@%p34 bra BB0_47;
|
||
|
|
bra.uni BB0_46;
|
||
|
|
|
||
|
|
BB0_47:
|
||
|
|
mov.f32 %f224, 0fBE2AAAA3;
|
||
|
|
fma.rn.f32 %f225, %f577, %f52, %f224;
|
||
|
|
mov.f32 %f226, 0f00000000;
|
||
|
|
fma.rn.f32 %f578, %f225, %f52, %f226;
|
||
|
|
bra.uni BB0_48;
|
||
|
|
|
||
|
|
BB0_46:
|
||
|
|
mov.f32 %f221, 0f3D2AAAA5;
|
||
|
|
fma.rn.f32 %f222, %f577, %f52, %f221;
|
||
|
|
mov.f32 %f223, 0fBF000000;
|
||
|
|
fma.rn.f32 %f578, %f222, %f52, %f223;
|
||
|
|
|
||
|
|
BB0_48:
|
||
|
|
fma.rn.f32 %f579, %f578, %f576, %f576;
|
||
|
|
@%p34 bra BB0_50;
|
||
|
|
|
||
|
|
mov.f32 %f227, 0f3F800000;
|
||
|
|
fma.rn.f32 %f579, %f578, %f52, %f227;
|
||
|
|
|
||
|
|
BB0_50:
|
||
|
|
and.b32 %r214, %r301, 2;
|
||
|
|
setp.eq.s32 %p37, %r214, 0;
|
||
|
|
@%p37 bra BB0_52;
|
||
|
|
|
||
|
|
mov.f32 %f228, 0f00000000;
|
||
|
|
mov.f32 %f229, 0fBF800000;
|
||
|
|
fma.rn.f32 %f579, %f579, %f229, %f228;
|
||
|
|
|
||
|
|
BB0_52:
|
||
|
|
mul.f32 %f230, %f27, %f573;
|
||
|
|
mul.f32 %f231, %f230, %f230;
|
||
|
|
mov.f32 %f232, 0f3F800000;
|
||
|
|
sub.f32 %f233, %f232, %f231;
|
||
|
|
mul.f32 %f234, %f27, %f579;
|
||
|
|
mul.f32 %f235, %f234, %f234;
|
||
|
|
sub.f32 %f236, %f233, %f235;
|
||
|
|
mov.f32 %f237, 0f00000000;
|
||
|
|
max.f32 %f238, %f237, %f236;
|
||
|
|
sqrt.rn.f32 %f239, %f238;
|
||
|
|
mul.f32 %f240, %f10, %f234;
|
||
|
|
mul.f32 %f241, %f11, %f234;
|
||
|
|
mul.f32 %f242, %f12, %f234;
|
||
|
|
fma.rn.f32 %f243, %f19, %f230, %f240;
|
||
|
|
fma.rn.f32 %f244, %f18, %f230, %f241;
|
||
|
|
fma.rn.f32 %f245, %f17, %f230, %f242;
|
||
|
|
fma.rn.f32 %f64, %f560, %f239, %f243;
|
||
|
|
fma.rn.f32 %f65, %f561, %f239, %f244;
|
||
|
|
fma.rn.f32 %f66, %f562, %f239, %f245;
|
||
|
|
setp.gt.f32 %p38, %f65, 0f00000000;
|
||
|
|
setp.eq.s32 %p39, %r5, 0;
|
||
|
|
or.pred %p40, %p39, %p38;
|
||
|
|
@!%p40 bra BB0_54;
|
||
|
|
bra.uni BB0_53;
|
||
|
|
|
||
|
|
BB0_53:
|
||
|
|
add.u64 %rd50, %SP, 28;
|
||
|
|
cvta.to.local.u64 %rd51, %rd50;
|
||
|
|
max.f32 %f259, %f265, %f153;
|
||
|
|
ld.global.u32 %r215, [sky];
|
||
|
|
neg.f32 %f252, %f64;
|
||
|
|
neg.f32 %f251, %f65;
|
||
|
|
neg.f32 %f250, %f66;
|
||
|
|
mov.u32 %r216, 6;
|
||
|
|
mov.u32 %r217, 0;
|
||
|
|
// inline asm
|
||
|
|
call (%f246, %f247, %f248, %f249), _rt_texture_get_base_id, (%r215, %r216, %f250, %f251, %f252, %r217);
|
||
|
|
// inline asm
|
||
|
|
st.local.f32 [%rd51], %f246;
|
||
|
|
st.local.f32 [%rd51+4], %f247;
|
||
|
|
st.local.f32 [%rd51+8], %f248;
|
||
|
|
ld.global.u32 %r218, [root];
|
||
|
|
mov.u32 %r219, 1;
|
||
|
|
mov.f32 %f260, 0f6C4ECB8F;
|
||
|
|
// inline asm
|
||
|
|
call _rt_trace_64, (%r218, %f7, %f8, %f9, %f64, %f65, %f66, %r219, %f259, %f260, %rd50, %r107);
|
||
|
|
// inline asm
|
||
|
|
ld.local.f32 %f267, [%rd51];
|
||
|
|
add.f32 %f581, %f581, %f267;
|
||
|
|
ld.local.f32 %f268, [%rd51+4];
|
||
|
|
add.f32 %f582, %f582, %f268;
|
||
|
|
ld.local.f32 %f269, [%rd51+8];
|
||
|
|
add.f32 %f583, %f583, %f269;
|
||
|
|
|
||
|
|
BB0_54:
|
||
|
|
add.s32 %r280, %r280, 1;
|
||
|
|
setp.lt.s32 %p41, %r280, %r1;
|
||
|
|
@%p41 bra BB0_6;
|
||
|
|
|
||
|
|
add.s32 %r278, %r278, 1;
|
||
|
|
setp.lt.s32 %p42, %r278, %r1;
|
||
|
|
@%p42 bra BB0_5;
|
||
|
|
|
||
|
|
BB0_56:
|
||
|
|
mul.lo.s32 %r221, %r1, %r1;
|
||
|
|
cvt.rn.f32.s32 %f270, %r221;
|
||
|
|
rcp.rn.f32 %f271, %f270;
|
||
|
|
mul.f32 %f272, %f581, %f271;
|
||
|
|
mul.f32 %f273, %f582, %f271;
|
||
|
|
mul.f32 %f274, %f583, %f271;
|
||
|
|
ld.global.f32 %f275, [skyColor];
|
||
|
|
mul.f32 %f76, %f272, %f275;
|
||
|
|
ld.global.f32 %f276, [skyColor+4];
|
||
|
|
mul.f32 %f77, %f273, %f276;
|
||
|
|
ld.global.f32 %f277, [skyColor+8];
|
||
|
|
mul.f32 %f78, %f274, %f277;
|
||
|
|
ld.global.u32 %r302, [imageEnabled];
|
||
|
|
and.b32 %r222, %r302, 1;
|
||
|
|
setp.eq.b32 %p43, %r222, 1;
|
||
|
|
@!%p43 bra BB0_91;
|
||
|
|
bra.uni BB0_57;
|
||
|
|
|
||
|
|
BB0_57:
|
||
|
|
mov.f32 %f280, 0f3E666666;
|
||
|
|
cvt.rzi.f32.f32 %f281, %f280;
|
||
|
|
fma.rn.f32 %f282, %f281, 0fC0000000, 0f3EE66666;
|
||
|
|
abs.f32 %f79, %f282;
|
||
|
|
abs.f32 %f80, %f76;
|
||
|
|
setp.lt.f32 %p44, %f80, 0f00800000;
|
||
|
|
mul.f32 %f283, %f80, 0f4B800000;
|
||
|
|
selp.f32 %f284, 0fC3170000, 0fC2FE0000, %p44;
|
||
|
|
selp.f32 %f285, %f283, %f80, %p44;
|
||
|
|
mov.b32 %r223, %f285;
|
||
|
|
and.b32 %r224, %r223, 8388607;
|
||
|
|
or.b32 %r225, %r224, 1065353216;
|
||
|
|
mov.b32 %f286, %r225;
|
||
|
|
shr.u32 %r226, %r223, 23;
|
||
|
|
cvt.rn.f32.u32 %f287, %r226;
|
||
|
|
add.f32 %f288, %f284, %f287;
|
||
|
|
setp.gt.f32 %p45, %f286, 0f3FB504F3;
|
||
|
|
mul.f32 %f289, %f286, 0f3F000000;
|
||
|
|
add.f32 %f290, %f288, 0f3F800000;
|
||
|
|
selp.f32 %f291, %f289, %f286, %p45;
|
||
|
|
selp.f32 %f292, %f290, %f288, %p45;
|
||
|
|
add.f32 %f293, %f291, 0fBF800000;
|
||
|
|
add.f32 %f279, %f291, 0f3F800000;
|
||
|
|
// inline asm
|
||
|
|
rcp.approx.ftz.f32 %f278,%f279;
|
||
|
|
// inline asm
|
||
|
|
add.f32 %f294, %f293, %f293;
|
||
|
|
mul.f32 %f295, %f278, %f294;
|
||
|
|
mul.f32 %f296, %f295, %f295;
|
||
|
|
mov.f32 %f297, 0f3C4CAF63;
|
||
|
|
mov.f32 %f298, 0f3B18F0FE;
|
||
|
|
fma.rn.f32 %f299, %f298, %f296, %f297;
|
||
|
|
mov.f32 %f300, 0f3DAAAABD;
|
||
|
|
fma.rn.f32 %f301, %f299, %f296, %f300;
|
||
|
|
mul.rn.f32 %f302, %f301, %f296;
|
||
|
|
mul.rn.f32 %f303, %f302, %f295;
|
||
|
|
sub.f32 %f304, %f293, %f295;
|
||
|
|
neg.f32 %f305, %f295;
|
||
|
|
add.f32 %f306, %f304, %f304;
|
||
|
|
fma.rn.f32 %f307, %f305, %f293, %f306;
|
||
|
|
mul.rn.f32 %f308, %f278, %f307;
|
||
|
|
add.f32 %f309, %f303, %f295;
|
||
|
|
sub.f32 %f310, %f295, %f309;
|
||
|
|
add.f32 %f311, %f303, %f310;
|
||
|
|
add.f32 %f312, %f308, %f311;
|
||
|
|
add.f32 %f313, %f309, %f312;
|
||
|
|
sub.f32 %f314, %f309, %f313;
|
||
|
|
add.f32 %f315, %f312, %f314;
|
||
|
|
mov.f32 %f316, 0f3F317200;
|
||
|
|
mul.rn.f32 %f317, %f292, %f316;
|
||
|
|
mov.f32 %f318, 0f35BFBE8E;
|
||
|
|
mul.rn.f32 %f319, %f292, %f318;
|
||
|
|
add.f32 %f320, %f317, %f313;
|
||
|
|
sub.f32 %f321, %f317, %f320;
|
||
|
|
add.f32 %f322, %f313, %f321;
|
||
|
|
add.f32 %f323, %f315, %f322;
|
||
|
|
add.f32 %f324, %f319, %f323;
|
||
|
|
add.f32 %f325, %f320, %f324;
|
||
|
|
sub.f32 %f326, %f320, %f325;
|
||
|
|
add.f32 %f327, %f324, %f326;
|
||
|
|
mov.f32 %f328, 0f3EE66666;
|
||
|
|
mul.rn.f32 %f329, %f328, %f325;
|
||
|
|
neg.f32 %f330, %f329;
|
||
|
|
fma.rn.f32 %f331, %f328, %f325, %f330;
|
||
|
|
fma.rn.f32 %f332, %f328, %f327, %f331;
|
||
|
|
mov.f32 %f333, 0f00000000;
|
||
|
|
fma.rn.f32 %f334, %f333, %f325, %f332;
|
||
|
|
add.rn.f32 %f335, %f329, %f334;
|
||
|
|
neg.f32 %f336, %f335;
|
||
|
|
add.rn.f32 %f337, %f329, %f336;
|
||
|
|
add.rn.f32 %f338, %f337, %f334;
|
||
|
|
mov.b32 %r227, %f335;
|
||
|
|
setp.eq.s32 %p46, %r227, 1118925336;
|
||
|
|
add.s32 %r228, %r227, -1;
|
||
|
|
mov.b32 %f339, %r228;
|
||
|
|
add.f32 %f340, %f338, 0f37000000;
|
||
|
|
selp.f32 %f341, %f339, %f335, %p46;
|
||
|
|
selp.f32 %f81, %f340, %f338, %p46;
|
||
|
|
mul.f32 %f342, %f341, 0f3FB8AA3B;
|
||
|
|
cvt.rzi.f32.f32 %f343, %f342;
|
||
|
|
mov.f32 %f344, 0fBF317200;
|
||
|
|
fma.rn.f32 %f345, %f343, %f344, %f341;
|
||
|
|
mov.f32 %f346, 0fB5BFBE8E;
|
||
|
|
fma.rn.f32 %f347, %f343, %f346, %f345;
|
||
|
|
mul.f32 %f348, %f347, 0f3FB8AA3B;
|
||
|
|
ex2.approx.ftz.f32 %f349, %f348;
|
||
|
|
add.f32 %f350, %f343, 0f00000000;
|
||
|
|
ex2.approx.f32 %f351, %f350;
|
||
|
|
mul.f32 %f352, %f349, %f351;
|
||
|
|
setp.lt.f32 %p47, %f341, 0fC2D20000;
|
||
|
|
selp.f32 %f353, 0f00000000, %f352, %p47;
|
||
|
|
setp.gt.f32 %p48, %f341, 0f42D20000;
|
||
|
|
selp.f32 %f587, 0f7F800000, %f353, %p48;
|
||
|
|
setp.eq.f32 %p49, %f587, 0f7F800000;
|
||
|
|
@%p49 bra BB0_59;
|
||
|
|
|
||
|
|
fma.rn.f32 %f587, %f587, %f81, %f587;
|
||
|
|
|
||
|
|
BB0_59:
|
||
|
|
setp.lt.f32 %p50, %f76, 0f00000000;
|
||
|
|
setp.eq.f32 %p51, %f79, 0f3F800000;
|
||
|
|
and.pred %p1, %p50, %p51;
|
||
|
|
mov.b32 %r229, %f587;
|
||
|
|
xor.b32 %r230, %r229, -2147483648;
|
||
|
|
mov.b32 %f354, %r230;
|
||
|
|
selp.f32 %f589, %f354, %f587, %p1;
|
||
|
|
setp.eq.f32 %p52, %f76, 0f00000000;
|
||
|
|
@%p52 bra BB0_62;
|
||
|
|
bra.uni BB0_60;
|
||
|
|
|
||
|
|
BB0_62:
|
||
|
|
add.f32 %f357, %f76, %f76;
|
||
|
|
selp.f32 %f589, %f357, 0f00000000, %p51;
|
||
|
|
bra.uni BB0_63;
|
||
|
|
|
||
|
|
BB0_100:
|
||
|
|
mov.u64 %rd105, image_HDR;
|
||
|
|
cvta.global.u64 %rd100, %rd105;
|
||
|
|
mov.u32 %r273, 8;
|
||
|
|
// inline asm
|
||
|
|
call (%rd99), _rt_buffer_get_64, (%rd100, %r91, %r273, %rd17, %rd18, %rd24, %rd24);
|
||
|
|
// inline asm
|
||
|
|
mov.f32 %f538, 0f00000000;
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs40, %f538;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
mov.u16 %rs41, 0;
|
||
|
|
st.v4.u16 [%rd99], {%rs40, %rs40, %rs40, %rs41};
|
||
|
|
bra.uni BB0_101;
|
||
|
|
|
||
|
|
BB0_60:
|
||
|
|
setp.geu.f32 %p53, %f76, 0f00000000;
|
||
|
|
@%p53 bra BB0_63;
|
||
|
|
|
||
|
|
mov.f32 %f559, 0f3EE66666;
|
||
|
|
cvt.rzi.f32.f32 %f356, %f559;
|
||
|
|
setp.neu.f32 %p54, %f356, 0f3EE66666;
|
||
|
|
selp.f32 %f589, 0f7FFFFFFF, %f589, %p54;
|
||
|
|
|
||
|
|
BB0_63:
|
||
|
|
add.f32 %f358, %f80, 0f3EE66666;
|
||
|
|
mov.b32 %r231, %f358;
|
||
|
|
setp.lt.s32 %p56, %r231, 2139095040;
|
||
|
|
@%p56 bra BB0_68;
|
||
|
|
|
||
|
|
setp.gtu.f32 %p57, %f80, 0f7F800000;
|
||
|
|
@%p57 bra BB0_67;
|
||
|
|
bra.uni BB0_65;
|
||
|
|
|
||
|
|
BB0_67:
|
||
|
|
add.f32 %f589, %f76, 0f3EE66666;
|
||
|
|
bra.uni BB0_68;
|
||
|
|
|
||
|
|
BB0_65:
|
||
|
|
setp.neu.f32 %p58, %f80, 0f7F800000;
|
||
|
|
@%p58 bra BB0_68;
|
||
|
|
|
||
|
|
selp.f32 %f589, 0fFF800000, 0f7F800000, %p1;
|
||
|
|
|
||
|
|
BB0_68:
|
||
|
|
mov.f32 %f547, 0fB5BFBE8E;
|
||
|
|
mov.f32 %f546, 0fBF317200;
|
||
|
|
mov.f32 %f545, 0f00000000;
|
||
|
|
mov.f32 %f544, 0f35BFBE8E;
|
||
|
|
mov.f32 %f543, 0f3F317200;
|
||
|
|
mov.f32 %f542, 0f3DAAAABD;
|
||
|
|
mov.f32 %f541, 0f3C4CAF63;
|
||
|
|
mov.f32 %f540, 0f3B18F0FE;
|
||
|
|
mov.f32 %f539, 0f3EE66666;
|
||
|
|
setp.eq.f32 %p59, %f76, 0f3F800000;
|
||
|
|
selp.f32 %f92, 0f3F800000, %f589, %p59;
|
||
|
|
abs.f32 %f93, %f77;
|
||
|
|
setp.lt.f32 %p60, %f93, 0f00800000;
|
||
|
|
mul.f32 %f361, %f93, 0f4B800000;
|
||
|
|
selp.f32 %f362, 0fC3170000, 0fC2FE0000, %p60;
|
||
|
|
selp.f32 %f363, %f361, %f93, %p60;
|
||
|
|
mov.b32 %r232, %f363;
|
||
|
|
and.b32 %r233, %r232, 8388607;
|
||
|
|
or.b32 %r234, %r233, 1065353216;
|
||
|
|
mov.b32 %f364, %r234;
|
||
|
|
shr.u32 %r235, %r232, 23;
|
||
|
|
cvt.rn.f32.u32 %f365, %r235;
|
||
|
|
add.f32 %f366, %f362, %f365;
|
||
|
|
setp.gt.f32 %p61, %f364, 0f3FB504F3;
|
||
|
|
mul.f32 %f367, %f364, 0f3F000000;
|
||
|
|
add.f32 %f368, %f366, 0f3F800000;
|
||
|
|
selp.f32 %f369, %f367, %f364, %p61;
|
||
|
|
selp.f32 %f370, %f368, %f366, %p61;
|
||
|
|
add.f32 %f371, %f369, 0fBF800000;
|
||
|
|
add.f32 %f360, %f369, 0f3F800000;
|
||
|
|
// inline asm
|
||
|
|
rcp.approx.ftz.f32 %f359,%f360;
|
||
|
|
// inline asm
|
||
|
|
add.f32 %f372, %f371, %f371;
|
||
|
|
mul.f32 %f373, %f359, %f372;
|
||
|
|
mul.f32 %f374, %f373, %f373;
|
||
|
|
fma.rn.f32 %f377, %f540, %f374, %f541;
|
||
|
|
fma.rn.f32 %f379, %f377, %f374, %f542;
|
||
|
|
mul.rn.f32 %f380, %f379, %f374;
|
||
|
|
mul.rn.f32 %f381, %f380, %f373;
|
||
|
|
sub.f32 %f382, %f371, %f373;
|
||
|
|
neg.f32 %f383, %f373;
|
||
|
|
add.f32 %f384, %f382, %f382;
|
||
|
|
fma.rn.f32 %f385, %f383, %f371, %f384;
|
||
|
|
mul.rn.f32 %f386, %f359, %f385;
|
||
|
|
add.f32 %f387, %f381, %f373;
|
||
|
|
sub.f32 %f388, %f373, %f387;
|
||
|
|
add.f32 %f389, %f381, %f388;
|
||
|
|
add.f32 %f390, %f386, %f389;
|
||
|
|
add.f32 %f391, %f387, %f390;
|
||
|
|
sub.f32 %f392, %f387, %f391;
|
||
|
|
add.f32 %f393, %f390, %f392;
|
||
|
|
mul.rn.f32 %f395, %f370, %f543;
|
||
|
|
mul.rn.f32 %f397, %f370, %f544;
|
||
|
|
add.f32 %f398, %f395, %f391;
|
||
|
|
sub.f32 %f399, %f395, %f398;
|
||
|
|
add.f32 %f400, %f391, %f399;
|
||
|
|
add.f32 %f401, %f393, %f400;
|
||
|
|
add.f32 %f402, %f397, %f401;
|
||
|
|
add.f32 %f403, %f398, %f402;
|
||
|
|
sub.f32 %f404, %f398, %f403;
|
||
|
|
add.f32 %f405, %f402, %f404;
|
||
|
|
mul.rn.f32 %f407, %f539, %f403;
|
||
|
|
neg.f32 %f408, %f407;
|
||
|
|
fma.rn.f32 %f409, %f539, %f403, %f408;
|
||
|
|
fma.rn.f32 %f410, %f539, %f405, %f409;
|
||
|
|
fma.rn.f32 %f412, %f545, %f403, %f410;
|
||
|
|
add.rn.f32 %f413, %f407, %f412;
|
||
|
|
neg.f32 %f414, %f413;
|
||
|
|
add.rn.f32 %f415, %f407, %f414;
|
||
|
|
add.rn.f32 %f416, %f415, %f412;
|
||
|
|
mov.b32 %r236, %f413;
|
||
|
|
setp.eq.s32 %p62, %r236, 1118925336;
|
||
|
|
add.s32 %r237, %r236, -1;
|
||
|
|
mov.b32 %f417, %r237;
|
||
|
|
add.f32 %f418, %f416, 0f37000000;
|
||
|
|
selp.f32 %f419, %f417, %f413, %p62;
|
||
|
|
selp.f32 %f94, %f418, %f416, %p62;
|
||
|
|
mul.f32 %f420, %f419, 0f3FB8AA3B;
|
||
|
|
cvt.rzi.f32.f32 %f421, %f420;
|
||
|
|
fma.rn.f32 %f423, %f421, %f546, %f419;
|
||
|
|
fma.rn.f32 %f425, %f421, %f547, %f423;
|
||
|
|
mul.f32 %f426, %f425, 0f3FB8AA3B;
|
||
|
|
ex2.approx.ftz.f32 %f427, %f426;
|
||
|
|
add.f32 %f428, %f421, 0f00000000;
|
||
|
|
ex2.approx.f32 %f429, %f428;
|
||
|
|
mul.f32 %f430, %f427, %f429;
|
||
|
|
setp.lt.f32 %p63, %f419, 0fC2D20000;
|
||
|
|
selp.f32 %f431, 0f00000000, %f430, %p63;
|
||
|
|
setp.gt.f32 %p64, %f419, 0f42D20000;
|
||
|
|
selp.f32 %f590, 0f7F800000, %f431, %p64;
|
||
|
|
setp.eq.f32 %p65, %f590, 0f7F800000;
|
||
|
|
@%p65 bra BB0_70;
|
||
|
|
|
||
|
|
fma.rn.f32 %f590, %f590, %f94, %f590;
|
||
|
|
|
||
|
|
BB0_70:
|
||
|
|
setp.lt.f32 %p66, %f77, 0f00000000;
|
||
|
|
and.pred %p2, %p66, %p51;
|
||
|
|
mov.b32 %r238, %f590;
|
||
|
|
xor.b32 %r239, %r238, -2147483648;
|
||
|
|
mov.b32 %f432, %r239;
|
||
|
|
selp.f32 %f592, %f432, %f590, %p2;
|
||
|
|
setp.eq.f32 %p68, %f77, 0f00000000;
|
||
|
|
@%p68 bra BB0_73;
|
||
|
|
bra.uni BB0_71;
|
||
|
|
|
||
|
|
BB0_73:
|
||
|
|
add.f32 %f435, %f77, %f77;
|
||
|
|
selp.f32 %f592, %f435, 0f00000000, %p51;
|
||
|
|
bra.uni BB0_74;
|
||
|
|
|
||
|
|
BB0_71:
|
||
|
|
setp.geu.f32 %p69, %f77, 0f00000000;
|
||
|
|
@%p69 bra BB0_74;
|
||
|
|
|
||
|
|
mov.f32 %f558, 0f3EE66666;
|
||
|
|
cvt.rzi.f32.f32 %f434, %f558;
|
||
|
|
setp.neu.f32 %p70, %f434, 0f3EE66666;
|
||
|
|
selp.f32 %f592, 0f7FFFFFFF, %f592, %p70;
|
||
|
|
|
||
|
|
BB0_74:
|
||
|
|
add.f32 %f436, %f93, 0f3EE66666;
|
||
|
|
mov.b32 %r240, %f436;
|
||
|
|
setp.lt.s32 %p72, %r240, 2139095040;
|
||
|
|
@%p72 bra BB0_79;
|
||
|
|
|
||
|
|
setp.gtu.f32 %p73, %f93, 0f7F800000;
|
||
|
|
@%p73 bra BB0_78;
|
||
|
|
bra.uni BB0_76;
|
||
|
|
|
||
|
|
BB0_78:
|
||
|
|
add.f32 %f592, %f77, 0f3EE66666;
|
||
|
|
bra.uni BB0_79;
|
||
|
|
|
||
|
|
BB0_76:
|
||
|
|
setp.neu.f32 %p74, %f93, 0f7F800000;
|
||
|
|
@%p74 bra BB0_79;
|
||
|
|
|
||
|
|
selp.f32 %f592, 0fFF800000, 0f7F800000, %p2;
|
||
|
|
|
||
|
|
BB0_79:
|
||
|
|
mov.f32 %f556, 0fB5BFBE8E;
|
||
|
|
mov.f32 %f555, 0fBF317200;
|
||
|
|
mov.f32 %f554, 0f00000000;
|
||
|
|
mov.f32 %f553, 0f35BFBE8E;
|
||
|
|
mov.f32 %f552, 0f3F317200;
|
||
|
|
mov.f32 %f551, 0f3DAAAABD;
|
||
|
|
mov.f32 %f550, 0f3C4CAF63;
|
||
|
|
mov.f32 %f549, 0f3B18F0FE;
|
||
|
|
mov.f32 %f548, 0f3EE66666;
|
||
|
|
setp.eq.f32 %p75, %f77, 0f3F800000;
|
||
|
|
selp.f32 %f105, 0f3F800000, %f592, %p75;
|
||
|
|
abs.f32 %f106, %f78;
|
||
|
|
setp.lt.f32 %p76, %f106, 0f00800000;
|
||
|
|
mul.f32 %f439, %f106, 0f4B800000;
|
||
|
|
selp.f32 %f440, 0fC3170000, 0fC2FE0000, %p76;
|
||
|
|
selp.f32 %f441, %f439, %f106, %p76;
|
||
|
|
mov.b32 %r241, %f441;
|
||
|
|
and.b32 %r242, %r241, 8388607;
|
||
|
|
or.b32 %r243, %r242, 1065353216;
|
||
|
|
mov.b32 %f442, %r243;
|
||
|
|
shr.u32 %r244, %r241, 23;
|
||
|
|
cvt.rn.f32.u32 %f443, %r244;
|
||
|
|
add.f32 %f444, %f440, %f443;
|
||
|
|
setp.gt.f32 %p77, %f442, 0f3FB504F3;
|
||
|
|
mul.f32 %f445, %f442, 0f3F000000;
|
||
|
|
add.f32 %f446, %f444, 0f3F800000;
|
||
|
|
selp.f32 %f447, %f445, %f442, %p77;
|
||
|
|
selp.f32 %f448, %f446, %f444, %p77;
|
||
|
|
add.f32 %f449, %f447, 0fBF800000;
|
||
|
|
add.f32 %f438, %f447, 0f3F800000;
|
||
|
|
// inline asm
|
||
|
|
rcp.approx.ftz.f32 %f437,%f438;
|
||
|
|
// inline asm
|
||
|
|
add.f32 %f450, %f449, %f449;
|
||
|
|
mul.f32 %f451, %f437, %f450;
|
||
|
|
mul.f32 %f452, %f451, %f451;
|
||
|
|
fma.rn.f32 %f455, %f549, %f452, %f550;
|
||
|
|
fma.rn.f32 %f457, %f455, %f452, %f551;
|
||
|
|
mul.rn.f32 %f458, %f457, %f452;
|
||
|
|
mul.rn.f32 %f459, %f458, %f451;
|
||
|
|
sub.f32 %f460, %f449, %f451;
|
||
|
|
neg.f32 %f461, %f451;
|
||
|
|
add.f32 %f462, %f460, %f460;
|
||
|
|
fma.rn.f32 %f463, %f461, %f449, %f462;
|
||
|
|
mul.rn.f32 %f464, %f437, %f463;
|
||
|
|
add.f32 %f465, %f459, %f451;
|
||
|
|
sub.f32 %f466, %f451, %f465;
|
||
|
|
add.f32 %f467, %f459, %f466;
|
||
|
|
add.f32 %f468, %f464, %f467;
|
||
|
|
add.f32 %f469, %f465, %f468;
|
||
|
|
sub.f32 %f470, %f465, %f469;
|
||
|
|
add.f32 %f471, %f468, %f470;
|
||
|
|
mul.rn.f32 %f473, %f448, %f552;
|
||
|
|
mul.rn.f32 %f475, %f448, %f553;
|
||
|
|
add.f32 %f476, %f473, %f469;
|
||
|
|
sub.f32 %f477, %f473, %f476;
|
||
|
|
add.f32 %f478, %f469, %f477;
|
||
|
|
add.f32 %f479, %f471, %f478;
|
||
|
|
add.f32 %f480, %f475, %f479;
|
||
|
|
add.f32 %f481, %f476, %f480;
|
||
|
|
sub.f32 %f482, %f476, %f481;
|
||
|
|
add.f32 %f483, %f480, %f482;
|
||
|
|
mul.rn.f32 %f485, %f548, %f481;
|
||
|
|
neg.f32 %f486, %f485;
|
||
|
|
fma.rn.f32 %f487, %f548, %f481, %f486;
|
||
|
|
fma.rn.f32 %f488, %f548, %f483, %f487;
|
||
|
|
fma.rn.f32 %f490, %f554, %f481, %f488;
|
||
|
|
add.rn.f32 %f491, %f485, %f490;
|
||
|
|
neg.f32 %f492, %f491;
|
||
|
|
add.rn.f32 %f493, %f485, %f492;
|
||
|
|
add.rn.f32 %f494, %f493, %f490;
|
||
|
|
mov.b32 %r245, %f491;
|
||
|
|
setp.eq.s32 %p78, %r245, 1118925336;
|
||
|
|
add.s32 %r246, %r245, -1;
|
||
|
|
mov.b32 %f495, %r246;
|
||
|
|
add.f32 %f496, %f494, 0f37000000;
|
||
|
|
selp.f32 %f497, %f495, %f491, %p78;
|
||
|
|
selp.f32 %f107, %f496, %f494, %p78;
|
||
|
|
mul.f32 %f498, %f497, 0f3FB8AA3B;
|
||
|
|
cvt.rzi.f32.f32 %f499, %f498;
|
||
|
|
fma.rn.f32 %f501, %f499, %f555, %f497;
|
||
|
|
fma.rn.f32 %f503, %f499, %f556, %f501;
|
||
|
|
mul.f32 %f504, %f503, 0f3FB8AA3B;
|
||
|
|
ex2.approx.ftz.f32 %f505, %f504;
|
||
|
|
add.f32 %f506, %f499, 0f00000000;
|
||
|
|
ex2.approx.f32 %f507, %f506;
|
||
|
|
mul.f32 %f508, %f505, %f507;
|
||
|
|
setp.lt.f32 %p79, %f497, 0fC2D20000;
|
||
|
|
selp.f32 %f509, 0f00000000, %f508, %p79;
|
||
|
|
setp.gt.f32 %p80, %f497, 0f42D20000;
|
||
|
|
selp.f32 %f593, 0f7F800000, %f509, %p80;
|
||
|
|
setp.eq.f32 %p81, %f593, 0f7F800000;
|
||
|
|
@%p81 bra BB0_81;
|
||
|
|
|
||
|
|
fma.rn.f32 %f593, %f593, %f107, %f593;
|
||
|
|
|
||
|
|
BB0_81:
|
||
|
|
setp.lt.f32 %p82, %f78, 0f00000000;
|
||
|
|
and.pred %p3, %p82, %p51;
|
||
|
|
mov.b32 %r247, %f593;
|
||
|
|
xor.b32 %r248, %r247, -2147483648;
|
||
|
|
mov.b32 %f510, %r248;
|
||
|
|
selp.f32 %f595, %f510, %f593, %p3;
|
||
|
|
setp.eq.f32 %p84, %f78, 0f00000000;
|
||
|
|
@%p84 bra BB0_84;
|
||
|
|
bra.uni BB0_82;
|
||
|
|
|
||
|
|
BB0_84:
|
||
|
|
add.f32 %f513, %f78, %f78;
|
||
|
|
selp.f32 %f595, %f513, 0f00000000, %p51;
|
||
|
|
bra.uni BB0_85;
|
||
|
|
|
||
|
|
BB0_82:
|
||
|
|
setp.geu.f32 %p85, %f78, 0f00000000;
|
||
|
|
@%p85 bra BB0_85;
|
||
|
|
|
||
|
|
mov.f32 %f557, 0f3EE66666;
|
||
|
|
cvt.rzi.f32.f32 %f512, %f557;
|
||
|
|
setp.neu.f32 %p86, %f512, 0f3EE66666;
|
||
|
|
selp.f32 %f595, 0f7FFFFFFF, %f595, %p86;
|
||
|
|
|
||
|
|
BB0_85:
|
||
|
|
add.f32 %f514, %f106, 0f3EE66666;
|
||
|
|
mov.b32 %r249, %f514;
|
||
|
|
setp.lt.s32 %p88, %r249, 2139095040;
|
||
|
|
@%p88 bra BB0_90;
|
||
|
|
|
||
|
|
setp.gtu.f32 %p89, %f106, 0f7F800000;
|
||
|
|
@%p89 bra BB0_89;
|
||
|
|
bra.uni BB0_87;
|
||
|
|
|
||
|
|
BB0_89:
|
||
|
|
add.f32 %f595, %f78, 0f3EE66666;
|
||
|
|
bra.uni BB0_90;
|
||
|
|
|
||
|
|
BB0_87:
|
||
|
|
setp.neu.f32 %p90, %f106, 0f7F800000;
|
||
|
|
@%p90 bra BB0_90;
|
||
|
|
|
||
|
|
selp.f32 %f595, 0fFF800000, 0f7F800000, %p3;
|
||
|
|
|
||
|
|
BB0_90:
|
||
|
|
mov.u32 %r275, 4;
|
||
|
|
mov.u64 %rd106, 0;
|
||
|
|
mov.u32 %r274, 2;
|
||
|
|
setp.eq.f32 %p91, %f78, 0f3F800000;
|
||
|
|
selp.f32 %f515, 0f3F800000, %f595, %p91;
|
||
|
|
cvt.u64.u32 %rd55, %r4;
|
||
|
|
cvt.u64.u32 %rd54, %r3;
|
||
|
|
mov.u64 %rd58, image;
|
||
|
|
cvta.global.u64 %rd53, %rd58;
|
||
|
|
// inline asm
|
||
|
|
call (%rd52), _rt_buffer_get_64, (%rd53, %r274, %r275, %rd54, %rd55, %rd106, %rd106);
|
||
|
|
// inline asm
|
||
|
|
cvt.sat.f32.f32 %f516, %f515;
|
||
|
|
mul.f32 %f517, %f516, 0f437FFD71;
|
||
|
|
cvt.rzi.u32.f32 %r252, %f517;
|
||
|
|
cvt.sat.f32.f32 %f518, %f105;
|
||
|
|
mul.f32 %f519, %f518, 0f437FFD71;
|
||
|
|
cvt.rzi.u32.f32 %r253, %f519;
|
||
|
|
cvt.sat.f32.f32 %f520, %f92;
|
||
|
|
mul.f32 %f521, %f520, 0f437FFD71;
|
||
|
|
cvt.rzi.u32.f32 %r254, %f521;
|
||
|
|
cvt.u16.u32 %rs10, %r252;
|
||
|
|
cvt.u16.u32 %rs11, %r254;
|
||
|
|
cvt.u16.u32 %rs12, %r253;
|
||
|
|
mov.u16 %rs13, 255;
|
||
|
|
st.v4.u8 [%rd52], {%rs10, %rs12, %rs11, %rs13};
|
||
|
|
ld.global.u32 %r302, [imageEnabled];
|
||
|
|
|
||
|
|
BB0_91:
|
||
|
|
and.b32 %r255, %r302, 4;
|
||
|
|
setp.eq.s32 %p92, %r255, 0;
|
||
|
|
@%p92 bra BB0_101;
|
||
|
|
|
||
|
|
ld.global.u32 %r256, [additive];
|
||
|
|
setp.eq.s32 %p93, %r256, 0;
|
||
|
|
cvt.u64.u32 %rd15, %r3;
|
||
|
|
cvt.u64.u32 %rd16, %r4;
|
||
|
|
mov.f32 %f522, 0f3F800000;
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs14, %f522;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
@%p93 bra BB0_94;
|
||
|
|
|
||
|
|
mov.u64 %rd107, 0;
|
||
|
|
mov.u32 %r276, 2;
|
||
|
|
mov.u64 %rd71, image_HDR;
|
||
|
|
cvta.global.u64 %rd60, %rd71;
|
||
|
|
mov.u32 %r260, 8;
|
||
|
|
// inline asm
|
||
|
|
call (%rd59), _rt_buffer_get_64, (%rd60, %r276, %r260, %rd15, %rd16, %rd107, %rd107);
|
||
|
|
// inline asm
|
||
|
|
ld.v4.u16 {%rs21, %rs22, %rs23, %rs24}, [%rd59];
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f523, %rs21;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f524, %rs22;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f525, %rs23;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
call (%rd65), _rt_buffer_get_64, (%rd60, %r276, %r260, %rd15, %rd16, %rd107, %rd107);
|
||
|
|
// inline asm
|
||
|
|
add.f32 %f526, %f76, %f523;
|
||
|
|
add.f32 %f527, %f77, %f524;
|
||
|
|
add.f32 %f528, %f78, %f525;
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs20, %f528;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs19, %f527;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs18, %f526;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
st.v4.u16 [%rd65], {%rs18, %rs19, %rs20, %rs14};
|
||
|
|
bra.uni BB0_101;
|
||
|
|
|
||
|
|
BB0_94:
|
||
|
|
mov.u64 %rd108, 0;
|
||
|
|
mov.u32 %r277, 2;
|
||
|
|
mov.u64 %rd78, image_HDR;
|
||
|
|
cvta.global.u64 %rd73, %rd78;
|
||
|
|
mov.u32 %r262, 8;
|
||
|
|
// inline asm
|
||
|
|
call (%rd72), _rt_buffer_get_64, (%rd73, %r277, %r262, %rd15, %rd16, %rd108, %rd108);
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs27, %f78;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs26, %f77;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs25, %f76;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
st.v4.u16 [%rd72], {%rs25, %rs26, %rs27, %rs14};
|
||
|
|
|
||
|
|
BB0_101:
|
||
|
|
ret;
|
||
|
|
}
|
||
|
|
|
||
|
|
|