1360 lines
39 KiB
Plaintext
1360 lines
39 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 .texref 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_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 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_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_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_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[32];
|
||
|
|
.reg .b64 %SP;
|
||
|
|
.reg .b64 %SPL;
|
||
|
|
.reg .pred %p<97>;
|
||
|
|
.reg .b16 %rs<42>;
|
||
|
|
.reg .f32 %f<569>;
|
||
|
|
.reg .b32 %r<302>;
|
||
|
|
.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 %f541, 0f00000000;
|
||
|
|
mov.f32 %f542, %f541;
|
||
|
|
mov.f32 %f543, %f541;
|
||
|
|
@%p4 bra BB0_2;
|
||
|
|
|
||
|
|
ld.u8 %rs6, [%rd19+1];
|
||
|
|
and.b16 %rs8, %rs4, 255;
|
||
|
|
cvt.rn.f32.u16 %f111, %rs8;
|
||
|
|
div.rn.f32 %f112, %f111, 0f437F0000;
|
||
|
|
fma.rn.f32 %f113, %f112, 0f40000000, 0fBF800000;
|
||
|
|
cvt.rn.f32.u16 %f114, %rs6;
|
||
|
|
div.rn.f32 %f115, %f114, 0f437F0000;
|
||
|
|
fma.rn.f32 %f116, %f115, 0f40000000, 0fBF800000;
|
||
|
|
cvt.rn.f32.u16 %f117, %rs3;
|
||
|
|
div.rn.f32 %f118, %f117, 0f437F0000;
|
||
|
|
fma.rn.f32 %f119, %f118, 0f40000000, 0fBF800000;
|
||
|
|
mul.f32 %f120, %f116, %f116;
|
||
|
|
fma.rn.f32 %f121, %f113, %f113, %f120;
|
||
|
|
fma.rn.f32 %f122, %f119, %f119, %f121;
|
||
|
|
sqrt.rn.f32 %f123, %f122;
|
||
|
|
rcp.rn.f32 %f124, %f123;
|
||
|
|
mul.f32 %f541, %f113, %f124;
|
||
|
|
mul.f32 %f542, %f116, %f124;
|
||
|
|
mul.f32 %f543, %f119, %f124;
|
||
|
|
|
||
|
|
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, %f542, 0f00000000;
|
||
|
|
setp.eq.f32 %p6, %f541, 0f00000000;
|
||
|
|
and.pred %p7, %p6, %p5;
|
||
|
|
setp.eq.f32 %p8, %f543, 0f00000000;
|
||
|
|
and.pred %p9, %p7, %p8;
|
||
|
|
@%p9 bra BB0_95;
|
||
|
|
bra.uni BB0_3;
|
||
|
|
|
||
|
|
BB0_95:
|
||
|
|
ld.global.u32 %r301, [imageEnabled];
|
||
|
|
and.b32 %r261, %r301, 1;
|
||
|
|
setp.eq.b32 %p94, %r261, 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 %r301, [imageEnabled];
|
||
|
|
|
||
|
|
BB0_97:
|
||
|
|
and.b32 %r264, %r301, 4;
|
||
|
|
setp.eq.s32 %p95, %r264, 0;
|
||
|
|
@%p95 bra BB0_101;
|
||
|
|
|
||
|
|
ld.global.u32 %r265, [additive];
|
||
|
|
setp.eq.s32 %p96, %r265, 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 %r269, 8;
|
||
|
|
// inline asm
|
||
|
|
call (%rd86), _rt_buffer_get_64, (%rd87, %r91, %r269, %rd17, %rd18, %rd24, %rd24);
|
||
|
|
// inline asm
|
||
|
|
ld.v4.u16 {%rs35, %rs36, %rs37, %rs38}, [%rd86];
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f506, %rs35;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f507, %rs36;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f508, %rs37;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
call (%rd92), _rt_buffer_get_64, (%rd87, %r91, %r269, %rd17, %rd18, %rd24, %rd24);
|
||
|
|
// inline asm
|
||
|
|
add.f32 %f509, %f506, 0f00000000;
|
||
|
|
add.f32 %f510, %f507, 0f00000000;
|
||
|
|
add.f32 %f511, %f508, 0f00000000;
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs34, %f511;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs33, %f510;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs32, %f509;}
|
||
|
|
|
||
|
|
// 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 %f126, [%rd26+8];
|
||
|
|
ld.f32 %f127, [%rd26+4];
|
||
|
|
ld.f32 %f128, [%rd26];
|
||
|
|
mul.f32 %f129, %f128, 0f3456BF95;
|
||
|
|
mul.f32 %f130, %f127, 0f3456BF95;
|
||
|
|
mul.f32 %f131, %f126, 0f3456BF95;
|
||
|
|
abs.f32 %f132, %f541;
|
||
|
|
div.rn.f32 %f133, %f129, %f132;
|
||
|
|
abs.f32 %f134, %f542;
|
||
|
|
div.rn.f32 %f135, %f130, %f134;
|
||
|
|
abs.f32 %f136, %f543;
|
||
|
|
div.rn.f32 %f137, %f131, %f136;
|
||
|
|
abs.f32 %f138, %f133;
|
||
|
|
abs.f32 %f139, %f135;
|
||
|
|
abs.f32 %f140, %f137;
|
||
|
|
mov.f32 %f141, 0f38D1B717;
|
||
|
|
max.f32 %f142, %f138, %f141;
|
||
|
|
max.f32 %f143, %f139, %f141;
|
||
|
|
max.f32 %f144, %f140, %f141;
|
||
|
|
fma.rn.f32 %f7, %f541, %f142, %f128;
|
||
|
|
fma.rn.f32 %f8, %f542, %f143, %f127;
|
||
|
|
fma.rn.f32 %f9, %f543, %f144, %f126;
|
||
|
|
ld.global.u32 %r5, [hemispherical];
|
||
|
|
setp.gt.f32 %p10, %f132, %f136;
|
||
|
|
neg.f32 %f145, %f542;
|
||
|
|
selp.f32 %f146, %f145, 0f00000000, %p10;
|
||
|
|
neg.f32 %f147, %f543;
|
||
|
|
selp.f32 %f148, %f541, %f147, %p10;
|
||
|
|
selp.f32 %f149, 0f00000000, %f542, %p10;
|
||
|
|
mul.f32 %f150, %f148, %f148;
|
||
|
|
fma.rn.f32 %f151, %f146, %f146, %f150;
|
||
|
|
fma.rn.f32 %f152, %f149, %f149, %f151;
|
||
|
|
sqrt.rn.f32 %f153, %f152;
|
||
|
|
rcp.rn.f32 %f154, %f153;
|
||
|
|
mul.f32 %f10, %f146, %f154;
|
||
|
|
mul.f32 %f11, %f148, %f154;
|
||
|
|
mul.f32 %f12, %f149, %f154;
|
||
|
|
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 %f558, 0f00000000;
|
||
|
|
setp.lt.s32 %p11, %r1, 1;
|
||
|
|
@%p11 bra BB0_56;
|
||
|
|
|
||
|
|
cvt.rn.f32.s32 %f156, %r1;
|
||
|
|
rcp.rn.f32 %f13, %f156;
|
||
|
|
ld.u32 %r279, [%rd32];
|
||
|
|
mul.f32 %f14, %f7, 0f3456BF95;
|
||
|
|
mul.f32 %f15, %f8, 0f3456BF95;
|
||
|
|
mul.f32 %f16, %f9, 0f3456BF95;
|
||
|
|
mul.f32 %f157, %f541, %f11;
|
||
|
|
mul.f32 %f158, %f542, %f10;
|
||
|
|
sub.f32 %f17, %f158, %f157;
|
||
|
|
mul.f32 %f159, %f543, %f10;
|
||
|
|
mul.f32 %f160, %f541, %f12;
|
||
|
|
sub.f32 %f18, %f160, %f159;
|
||
|
|
mul.f32 %f161, %f542, %f12;
|
||
|
|
mul.f32 %f162, %f543, %f11;
|
||
|
|
sub.f32 %f19, %f162, %f161;
|
||
|
|
mov.f32 %f558, 0f00000000;
|
||
|
|
mov.u32 %r118, 0;
|
||
|
|
abs.f32 %f240, %f15;
|
||
|
|
abs.f32 %f241, %f14;
|
||
|
|
max.f32 %f242, %f241, %f240;
|
||
|
|
abs.f32 %f243, %f16;
|
||
|
|
max.f32 %f244, %f242, %f243;
|
||
|
|
mov.u32 %r276, %r118;
|
||
|
|
|
||
|
|
BB0_5:
|
||
|
|
cvt.rn.f32.s32 %f21, %r276;
|
||
|
|
mov.u32 %r278, %r118;
|
||
|
|
|
||
|
|
BB0_6:
|
||
|
|
mad.lo.s32 %r120, %r279, 1664525, 1013904223;
|
||
|
|
and.b32 %r121, %r120, 16777215;
|
||
|
|
cvt.rn.f32.u32 %f163, %r121;
|
||
|
|
fma.rn.f32 %f164, %f163, 0f33800000, %f21;
|
||
|
|
mul.f32 %f165, %f13, %f164;
|
||
|
|
mad.lo.s32 %r279, %r120, 1664525, 1013904223;
|
||
|
|
and.b32 %r122, %r279, 16777215;
|
||
|
|
cvt.rn.f32.u32 %f166, %r122;
|
||
|
|
cvt.rn.f32.s32 %f167, %r278;
|
||
|
|
fma.rn.f32 %f168, %f166, 0f33800000, %f167;
|
||
|
|
mul.f32 %f169, %f13, %f168;
|
||
|
|
sqrt.rn.f32 %f23, %f165;
|
||
|
|
mul.f32 %f552, %f169, 0f40C90FDB;
|
||
|
|
abs.f32 %f25, %f552;
|
||
|
|
setp.neu.f32 %p12, %f25, 0f7F800000;
|
||
|
|
mov.f32 %f546, %f552;
|
||
|
|
@%p12 bra BB0_8;
|
||
|
|
|
||
|
|
mov.f32 %f170, 0f00000000;
|
||
|
|
mul.rn.f32 %f546, %f552, %f170;
|
||
|
|
|
||
|
|
BB0_8:
|
||
|
|
mul.f32 %f171, %f546, 0f3F22F983;
|
||
|
|
cvt.rni.s32.f32 %r289, %f171;
|
||
|
|
cvt.rn.f32.s32 %f172, %r289;
|
||
|
|
neg.f32 %f173, %f172;
|
||
|
|
mov.f32 %f174, 0f3FC90FDA;
|
||
|
|
fma.rn.f32 %f175, %f173, %f174, %f546;
|
||
|
|
mov.f32 %f176, 0f33A22168;
|
||
|
|
fma.rn.f32 %f177, %f173, %f176, %f175;
|
||
|
|
mov.f32 %f178, 0f27C234C5;
|
||
|
|
fma.rn.f32 %f547, %f173, %f178, %f177;
|
||
|
|
abs.f32 %f179, %f546;
|
||
|
|
setp.leu.f32 %p13, %f179, 0f47CE4780;
|
||
|
|
@%p13 bra BB0_19;
|
||
|
|
|
||
|
|
mov.b32 %r13, %f546;
|
||
|
|
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 %r281, 0;
|
||
|
|
mov.u64 %rd109, __cudart_i2opi_f;
|
||
|
|
mov.u32 %r280, -6;
|
||
|
|
|
||
|
|
BB0_10:
|
||
|
|
.pragma "nounroll";
|
||
|
|
ld.const.u32 %r128, [%rd109];
|
||
|
|
// inline asm
|
||
|
|
{
|
||
|
|
mad.lo.cc.u32 %r126, %r128, %r15, %r281;
|
||
|
|
madc.hi.u32 %r281, %r128, %r15, 0;
|
||
|
|
}
|
||
|
|
// inline asm
|
||
|
|
st.local.u32 [%rd110], %r126;
|
||
|
|
add.s64 %rd110, %rd110, 4;
|
||
|
|
add.s64 %rd109, %rd109, 4;
|
||
|
|
add.s32 %r280, %r280, 1;
|
||
|
|
setp.ne.s32 %p14, %r280, 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], %r281;
|
||
|
|
mov.u32 %r134, 6;
|
||
|
|
sub.s32 %r135, %r134, %r133;
|
||
|
|
mul.wide.s32 %rd44, %r135, 4;
|
||
|
|
add.s64 %rd8, %rd43, %rd44;
|
||
|
|
ld.local.u32 %r282, [%rd8];
|
||
|
|
ld.local.u32 %r283, [%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, %r283, %r137;
|
||
|
|
shl.b32 %r139, %r282, %r23;
|
||
|
|
add.s32 %r282, %r138, %r139;
|
||
|
|
ld.local.u32 %r140, [%rd8+-8];
|
||
|
|
shr.u32 %r141, %r140, %r137;
|
||
|
|
shl.b32 %r142, %r283, %r23;
|
||
|
|
add.s32 %r283, %r141, %r142;
|
||
|
|
|
||
|
|
BB0_13:
|
||
|
|
shr.u32 %r143, %r283, 30;
|
||
|
|
shl.b32 %r144, %r282, 2;
|
||
|
|
add.s32 %r284, %r143, %r144;
|
||
|
|
shl.b32 %r29, %r283, 2;
|
||
|
|
shr.u32 %r145, %r284, 31;
|
||
|
|
shr.u32 %r146, %r282, 30;
|
||
|
|
add.s32 %r30, %r145, %r146;
|
||
|
|
setp.eq.s32 %p16, %r145, 0;
|
||
|
|
@%p16 bra BB0_14;
|
||
|
|
bra.uni BB0_15;
|
||
|
|
|
||
|
|
BB0_14:
|
||
|
|
mov.u32 %r285, %r20;
|
||
|
|
mov.u32 %r286, %r29;
|
||
|
|
bra.uni BB0_16;
|
||
|
|
|
||
|
|
BB0_15:
|
||
|
|
not.b32 %r147, %r284;
|
||
|
|
neg.s32 %r286, %r29;
|
||
|
|
setp.eq.s32 %p17, %r29, 0;
|
||
|
|
selp.u32 %r148, 1, 0, %p17;
|
||
|
|
add.s32 %r284, %r148, %r147;
|
||
|
|
xor.b32 %r285, %r20, -2147483648;
|
||
|
|
|
||
|
|
BB0_16:
|
||
|
|
clz.b32 %r288, %r284;
|
||
|
|
setp.eq.s32 %p18, %r288, 0;
|
||
|
|
shl.b32 %r149, %r284, %r288;
|
||
|
|
mov.u32 %r150, 32;
|
||
|
|
sub.s32 %r151, %r150, %r288;
|
||
|
|
shr.u32 %r152, %r286, %r151;
|
||
|
|
add.s32 %r153, %r152, %r149;
|
||
|
|
selp.b32 %r38, %r284, %r153, %p18;
|
||
|
|
mov.u32 %r154, -921707870;
|
||
|
|
mul.hi.u32 %r287, %r38, %r154;
|
||
|
|
setp.eq.s32 %p19, %r20, 0;
|
||
|
|
neg.s32 %r155, %r30;
|
||
|
|
selp.b32 %r289, %r30, %r155, %p19;
|
||
|
|
setp.lt.s32 %p20, %r287, 1;
|
||
|
|
@%p20 bra BB0_18;
|
||
|
|
|
||
|
|
mul.lo.s32 %r156, %r38, -921707870;
|
||
|
|
shr.u32 %r157, %r156, 31;
|
||
|
|
shl.b32 %r158, %r287, 1;
|
||
|
|
add.s32 %r287, %r157, %r158;
|
||
|
|
add.s32 %r288, %r288, 1;
|
||
|
|
|
||
|
|
BB0_18:
|
||
|
|
mov.u32 %r159, 126;
|
||
|
|
sub.s32 %r160, %r159, %r288;
|
||
|
|
shl.b32 %r161, %r160, 23;
|
||
|
|
add.s32 %r162, %r287, 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, %r285;
|
||
|
|
mov.b32 %f547, %r167;
|
||
|
|
|
||
|
|
BB0_19:
|
||
|
|
mul.rn.f32 %f31, %f547, %f547;
|
||
|
|
add.s32 %r46, %r289, 1;
|
||
|
|
and.b32 %r47, %r46, 1;
|
||
|
|
setp.eq.s32 %p21, %r47, 0;
|
||
|
|
@%p21 bra BB0_21;
|
||
|
|
bra.uni BB0_20;
|
||
|
|
|
||
|
|
BB0_21:
|
||
|
|
mov.f32 %f182, 0f3C08839E;
|
||
|
|
mov.f32 %f183, 0fB94CA1F9;
|
||
|
|
fma.rn.f32 %f548, %f183, %f31, %f182;
|
||
|
|
bra.uni BB0_22;
|
||
|
|
|
||
|
|
BB0_20:
|
||
|
|
mov.f32 %f180, 0fBAB6061A;
|
||
|
|
mov.f32 %f181, 0f37CCF5CE;
|
||
|
|
fma.rn.f32 %f548, %f181, %f31, %f180;
|
||
|
|
|
||
|
|
BB0_22:
|
||
|
|
@%p21 bra BB0_24;
|
||
|
|
bra.uni BB0_23;
|
||
|
|
|
||
|
|
BB0_24:
|
||
|
|
mov.f32 %f187, 0fBE2AAAA3;
|
||
|
|
fma.rn.f32 %f188, %f548, %f31, %f187;
|
||
|
|
mov.f32 %f189, 0f00000000;
|
||
|
|
fma.rn.f32 %f549, %f188, %f31, %f189;
|
||
|
|
bra.uni BB0_25;
|
||
|
|
|
||
|
|
BB0_23:
|
||
|
|
mov.f32 %f184, 0f3D2AAAA5;
|
||
|
|
fma.rn.f32 %f185, %f548, %f31, %f184;
|
||
|
|
mov.f32 %f186, 0fBF000000;
|
||
|
|
fma.rn.f32 %f549, %f185, %f31, %f186;
|
||
|
|
|
||
|
|
BB0_25:
|
||
|
|
fma.rn.f32 %f550, %f549, %f547, %f547;
|
||
|
|
@%p21 bra BB0_27;
|
||
|
|
|
||
|
|
mov.f32 %f190, 0f3F800000;
|
||
|
|
fma.rn.f32 %f550, %f549, %f31, %f190;
|
||
|
|
|
||
|
|
BB0_27:
|
||
|
|
and.b32 %r168, %r46, 2;
|
||
|
|
setp.eq.s32 %p24, %r168, 0;
|
||
|
|
@%p24 bra BB0_29;
|
||
|
|
|
||
|
|
mov.f32 %f191, 0f00000000;
|
||
|
|
mov.f32 %f192, 0fBF800000;
|
||
|
|
fma.rn.f32 %f550, %f550, %f192, %f191;
|
||
|
|
|
||
|
|
BB0_29:
|
||
|
|
@%p12 bra BB0_31;
|
||
|
|
|
||
|
|
mov.f32 %f193, 0f00000000;
|
||
|
|
mul.rn.f32 %f552, %f552, %f193;
|
||
|
|
|
||
|
|
BB0_31:
|
||
|
|
mul.f32 %f194, %f552, 0f3F22F983;
|
||
|
|
cvt.rni.s32.f32 %r299, %f194;
|
||
|
|
cvt.rn.f32.s32 %f195, %r299;
|
||
|
|
neg.f32 %f196, %f195;
|
||
|
|
fma.rn.f32 %f198, %f196, %f174, %f552;
|
||
|
|
fma.rn.f32 %f200, %f196, %f176, %f198;
|
||
|
|
fma.rn.f32 %f553, %f196, %f178, %f200;
|
||
|
|
abs.f32 %f202, %f552;
|
||
|
|
setp.leu.f32 %p26, %f202, 0f47CE4780;
|
||
|
|
@%p26 bra BB0_42;
|
||
|
|
|
||
|
|
mov.b32 %r49, %f552;
|
||
|
|
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 %r291, 0;
|
||
|
|
mov.u64 %rd111, __cudart_i2opi_f;
|
||
|
|
mov.u32 %r290, -6;
|
||
|
|
|
||
|
|
BB0_33:
|
||
|
|
.pragma "nounroll";
|
||
|
|
ld.const.u32 %r174, [%rd111];
|
||
|
|
// inline asm
|
||
|
|
{
|
||
|
|
mad.lo.cc.u32 %r172, %r174, %r51, %r291;
|
||
|
|
madc.hi.u32 %r291, %r174, %r51, 0;
|
||
|
|
}
|
||
|
|
// inline asm
|
||
|
|
st.local.u32 [%rd112], %r172;
|
||
|
|
add.s64 %rd112, %rd112, 4;
|
||
|
|
add.s64 %rd111, %rd111, 4;
|
||
|
|
add.s32 %r290, %r290, 1;
|
||
|
|
setp.ne.s32 %p27, %r290, 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], %r291;
|
||
|
|
mov.u32 %r180, 6;
|
||
|
|
sub.s32 %r181, %r180, %r179;
|
||
|
|
mul.wide.s32 %rd49, %r181, 4;
|
||
|
|
add.s64 %rd14, %rd48, %rd49;
|
||
|
|
ld.local.u32 %r292, [%rd14];
|
||
|
|
ld.local.u32 %r293, [%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, %r293, %r183;
|
||
|
|
shl.b32 %r185, %r292, %r59;
|
||
|
|
add.s32 %r292, %r184, %r185;
|
||
|
|
ld.local.u32 %r186, [%rd14+-8];
|
||
|
|
shr.u32 %r187, %r186, %r183;
|
||
|
|
shl.b32 %r188, %r293, %r59;
|
||
|
|
add.s32 %r293, %r187, %r188;
|
||
|
|
|
||
|
|
BB0_36:
|
||
|
|
shr.u32 %r189, %r293, 30;
|
||
|
|
shl.b32 %r190, %r292, 2;
|
||
|
|
add.s32 %r294, %r189, %r190;
|
||
|
|
shl.b32 %r65, %r293, 2;
|
||
|
|
shr.u32 %r191, %r294, 31;
|
||
|
|
shr.u32 %r192, %r292, 30;
|
||
|
|
add.s32 %r66, %r191, %r192;
|
||
|
|
setp.eq.s32 %p29, %r191, 0;
|
||
|
|
@%p29 bra BB0_37;
|
||
|
|
bra.uni BB0_38;
|
||
|
|
|
||
|
|
BB0_37:
|
||
|
|
mov.u32 %r295, %r56;
|
||
|
|
mov.u32 %r296, %r65;
|
||
|
|
bra.uni BB0_39;
|
||
|
|
|
||
|
|
BB0_38:
|
||
|
|
not.b32 %r193, %r294;
|
||
|
|
neg.s32 %r296, %r65;
|
||
|
|
setp.eq.s32 %p30, %r65, 0;
|
||
|
|
selp.u32 %r194, 1, 0, %p30;
|
||
|
|
add.s32 %r294, %r194, %r193;
|
||
|
|
xor.b32 %r295, %r56, -2147483648;
|
||
|
|
|
||
|
|
BB0_39:
|
||
|
|
clz.b32 %r298, %r294;
|
||
|
|
setp.eq.s32 %p31, %r298, 0;
|
||
|
|
shl.b32 %r195, %r294, %r298;
|
||
|
|
mov.u32 %r196, 32;
|
||
|
|
sub.s32 %r197, %r196, %r298;
|
||
|
|
shr.u32 %r198, %r296, %r197;
|
||
|
|
add.s32 %r199, %r198, %r195;
|
||
|
|
selp.b32 %r74, %r294, %r199, %p31;
|
||
|
|
mov.u32 %r200, -921707870;
|
||
|
|
mul.hi.u32 %r297, %r74, %r200;
|
||
|
|
setp.eq.s32 %p32, %r56, 0;
|
||
|
|
neg.s32 %r201, %r66;
|
||
|
|
selp.b32 %r299, %r66, %r201, %p32;
|
||
|
|
setp.lt.s32 %p33, %r297, 1;
|
||
|
|
@%p33 bra BB0_41;
|
||
|
|
|
||
|
|
mul.lo.s32 %r202, %r74, -921707870;
|
||
|
|
shr.u32 %r203, %r202, 31;
|
||
|
|
shl.b32 %r204, %r297, 1;
|
||
|
|
add.s32 %r297, %r203, %r204;
|
||
|
|
add.s32 %r298, %r298, 1;
|
||
|
|
|
||
|
|
BB0_41:
|
||
|
|
mov.u32 %r205, 126;
|
||
|
|
sub.s32 %r206, %r205, %r298;
|
||
|
|
shl.b32 %r207, %r206, 23;
|
||
|
|
add.s32 %r208, %r297, 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, %r295;
|
||
|
|
mov.b32 %f553, %r213;
|
||
|
|
|
||
|
|
BB0_42:
|
||
|
|
mul.rn.f32 %f48, %f553, %f553;
|
||
|
|
and.b32 %r82, %r299, 1;
|
||
|
|
setp.eq.s32 %p34, %r82, 0;
|
||
|
|
@%p34 bra BB0_44;
|
||
|
|
bra.uni BB0_43;
|
||
|
|
|
||
|
|
BB0_44:
|
||
|
|
mov.f32 %f205, 0f3C08839E;
|
||
|
|
mov.f32 %f206, 0fB94CA1F9;
|
||
|
|
fma.rn.f32 %f554, %f206, %f48, %f205;
|
||
|
|
bra.uni BB0_45;
|
||
|
|
|
||
|
|
BB0_43:
|
||
|
|
mov.f32 %f203, 0fBAB6061A;
|
||
|
|
mov.f32 %f204, 0f37CCF5CE;
|
||
|
|
fma.rn.f32 %f554, %f204, %f48, %f203;
|
||
|
|
|
||
|
|
BB0_45:
|
||
|
|
@%p34 bra BB0_47;
|
||
|
|
bra.uni BB0_46;
|
||
|
|
|
||
|
|
BB0_47:
|
||
|
|
mov.f32 %f210, 0fBE2AAAA3;
|
||
|
|
fma.rn.f32 %f211, %f554, %f48, %f210;
|
||
|
|
mov.f32 %f212, 0f00000000;
|
||
|
|
fma.rn.f32 %f555, %f211, %f48, %f212;
|
||
|
|
bra.uni BB0_48;
|
||
|
|
|
||
|
|
BB0_46:
|
||
|
|
mov.f32 %f207, 0f3D2AAAA5;
|
||
|
|
fma.rn.f32 %f208, %f554, %f48, %f207;
|
||
|
|
mov.f32 %f209, 0fBF000000;
|
||
|
|
fma.rn.f32 %f555, %f208, %f48, %f209;
|
||
|
|
|
||
|
|
BB0_48:
|
||
|
|
fma.rn.f32 %f556, %f555, %f553, %f553;
|
||
|
|
@%p34 bra BB0_50;
|
||
|
|
|
||
|
|
mov.f32 %f213, 0f3F800000;
|
||
|
|
fma.rn.f32 %f556, %f555, %f48, %f213;
|
||
|
|
|
||
|
|
BB0_50:
|
||
|
|
and.b32 %r214, %r299, 2;
|
||
|
|
setp.eq.s32 %p37, %r214, 0;
|
||
|
|
@%p37 bra BB0_52;
|
||
|
|
|
||
|
|
mov.f32 %f214, 0f00000000;
|
||
|
|
mov.f32 %f215, 0fBF800000;
|
||
|
|
fma.rn.f32 %f556, %f556, %f215, %f214;
|
||
|
|
|
||
|
|
BB0_52:
|
||
|
|
mul.f32 %f216, %f23, %f550;
|
||
|
|
mul.f32 %f217, %f216, %f216;
|
||
|
|
mov.f32 %f218, 0f3F800000;
|
||
|
|
sub.f32 %f219, %f218, %f217;
|
||
|
|
mul.f32 %f220, %f23, %f556;
|
||
|
|
mul.f32 %f221, %f220, %f220;
|
||
|
|
sub.f32 %f222, %f219, %f221;
|
||
|
|
mov.f32 %f223, 0f00000000;
|
||
|
|
max.f32 %f224, %f223, %f222;
|
||
|
|
sqrt.rn.f32 %f225, %f224;
|
||
|
|
mul.f32 %f226, %f10, %f220;
|
||
|
|
mul.f32 %f227, %f11, %f220;
|
||
|
|
mul.f32 %f228, %f12, %f220;
|
||
|
|
fma.rn.f32 %f229, %f19, %f216, %f226;
|
||
|
|
fma.rn.f32 %f230, %f18, %f216, %f227;
|
||
|
|
fma.rn.f32 %f231, %f17, %f216, %f228;
|
||
|
|
fma.rn.f32 %f60, %f541, %f225, %f229;
|
||
|
|
fma.rn.f32 %f61, %f542, %f225, %f230;
|
||
|
|
fma.rn.f32 %f62, %f543, %f225, %f231;
|
||
|
|
setp.gt.f32 %p38, %f61, 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 %f238, %f244, %f141;
|
||
|
|
mov.u32 %r218, 1065353216;
|
||
|
|
st.local.u32 [%rd51], %r218;
|
||
|
|
ld.global.u32 %r215, [root];
|
||
|
|
mov.u32 %r216, 1;
|
||
|
|
mov.f32 %f239, 0f6C4ECB8F;
|
||
|
|
// inline asm
|
||
|
|
call _rt_trace_64, (%r215, %f7, %f8, %f9, %f60, %f61, %f62, %r216, %f238, %f239, %rd50, %r92);
|
||
|
|
// inline asm
|
||
|
|
ld.local.f32 %f246, [%rd51];
|
||
|
|
add.f32 %f558, %f558, %f246;
|
||
|
|
|
||
|
|
BB0_54:
|
||
|
|
add.s32 %r278, %r278, 1;
|
||
|
|
setp.lt.s32 %p41, %r278, %r1;
|
||
|
|
@%p41 bra BB0_6;
|
||
|
|
|
||
|
|
add.s32 %r276, %r276, 1;
|
||
|
|
setp.lt.s32 %p42, %r276, %r1;
|
||
|
|
@%p42 bra BB0_5;
|
||
|
|
|
||
|
|
BB0_56:
|
||
|
|
mul.lo.s32 %r219, %r1, %r1;
|
||
|
|
cvt.rn.f32.s32 %f247, %r219;
|
||
|
|
div.rn.f32 %f248, %f558, %f247;
|
||
|
|
ld.global.f32 %f249, [skyColor];
|
||
|
|
mul.f32 %f66, %f248, %f249;
|
||
|
|
ld.global.f32 %f250, [skyColor+4];
|
||
|
|
mul.f32 %f67, %f248, %f250;
|
||
|
|
ld.global.f32 %f251, [skyColor+8];
|
||
|
|
mul.f32 %f68, %f248, %f251;
|
||
|
|
ld.global.u32 %r300, [imageEnabled];
|
||
|
|
and.b32 %r220, %r300, 1;
|
||
|
|
setp.eq.b32 %p43, %r220, 1;
|
||
|
|
@!%p43 bra BB0_91;
|
||
|
|
bra.uni BB0_57;
|
||
|
|
|
||
|
|
BB0_57:
|
||
|
|
mov.f32 %f254, 0f3E666666;
|
||
|
|
cvt.rzi.f32.f32 %f255, %f254;
|
||
|
|
fma.rn.f32 %f256, %f255, 0fC0000000, 0f3EE66666;
|
||
|
|
abs.f32 %f69, %f256;
|
||
|
|
abs.f32 %f70, %f66;
|
||
|
|
setp.lt.f32 %p44, %f70, 0f00800000;
|
||
|
|
mul.f32 %f257, %f70, 0f4B800000;
|
||
|
|
selp.f32 %f258, 0fC3170000, 0fC2FE0000, %p44;
|
||
|
|
selp.f32 %f259, %f257, %f70, %p44;
|
||
|
|
mov.b32 %r221, %f259;
|
||
|
|
and.b32 %r222, %r221, 8388607;
|
||
|
|
or.b32 %r223, %r222, 1065353216;
|
||
|
|
mov.b32 %f260, %r223;
|
||
|
|
shr.u32 %r224, %r221, 23;
|
||
|
|
cvt.rn.f32.u32 %f261, %r224;
|
||
|
|
add.f32 %f262, %f258, %f261;
|
||
|
|
setp.gt.f32 %p45, %f260, 0f3FB504F3;
|
||
|
|
mul.f32 %f263, %f260, 0f3F000000;
|
||
|
|
add.f32 %f264, %f262, 0f3F800000;
|
||
|
|
selp.f32 %f265, %f263, %f260, %p45;
|
||
|
|
selp.f32 %f266, %f264, %f262, %p45;
|
||
|
|
add.f32 %f267, %f265, 0fBF800000;
|
||
|
|
add.f32 %f253, %f265, 0f3F800000;
|
||
|
|
// inline asm
|
||
|
|
rcp.approx.ftz.f32 %f252,%f253;
|
||
|
|
// inline asm
|
||
|
|
add.f32 %f268, %f267, %f267;
|
||
|
|
mul.f32 %f269, %f252, %f268;
|
||
|
|
mul.f32 %f270, %f269, %f269;
|
||
|
|
mov.f32 %f271, 0f3C4CAF63;
|
||
|
|
mov.f32 %f272, 0f3B18F0FE;
|
||
|
|
fma.rn.f32 %f273, %f272, %f270, %f271;
|
||
|
|
mov.f32 %f274, 0f3DAAAABD;
|
||
|
|
fma.rn.f32 %f275, %f273, %f270, %f274;
|
||
|
|
mul.rn.f32 %f276, %f275, %f270;
|
||
|
|
mul.rn.f32 %f277, %f276, %f269;
|
||
|
|
sub.f32 %f278, %f267, %f269;
|
||
|
|
neg.f32 %f279, %f269;
|
||
|
|
add.f32 %f280, %f278, %f278;
|
||
|
|
fma.rn.f32 %f281, %f279, %f267, %f280;
|
||
|
|
mul.rn.f32 %f282, %f252, %f281;
|
||
|
|
add.f32 %f283, %f277, %f269;
|
||
|
|
sub.f32 %f284, %f269, %f283;
|
||
|
|
add.f32 %f285, %f277, %f284;
|
||
|
|
add.f32 %f286, %f282, %f285;
|
||
|
|
add.f32 %f287, %f283, %f286;
|
||
|
|
sub.f32 %f288, %f283, %f287;
|
||
|
|
add.f32 %f289, %f286, %f288;
|
||
|
|
mov.f32 %f290, 0f3F317200;
|
||
|
|
mul.rn.f32 %f291, %f266, %f290;
|
||
|
|
mov.f32 %f292, 0f35BFBE8E;
|
||
|
|
mul.rn.f32 %f293, %f266, %f292;
|
||
|
|
add.f32 %f294, %f291, %f287;
|
||
|
|
sub.f32 %f295, %f291, %f294;
|
||
|
|
add.f32 %f296, %f287, %f295;
|
||
|
|
add.f32 %f297, %f289, %f296;
|
||
|
|
add.f32 %f298, %f293, %f297;
|
||
|
|
add.f32 %f299, %f294, %f298;
|
||
|
|
sub.f32 %f300, %f294, %f299;
|
||
|
|
add.f32 %f301, %f298, %f300;
|
||
|
|
mov.f32 %f302, 0f3EE66666;
|
||
|
|
mul.rn.f32 %f303, %f302, %f299;
|
||
|
|
neg.f32 %f304, %f303;
|
||
|
|
fma.rn.f32 %f305, %f302, %f299, %f304;
|
||
|
|
fma.rn.f32 %f306, %f302, %f301, %f305;
|
||
|
|
mov.f32 %f307, 0f00000000;
|
||
|
|
fma.rn.f32 %f308, %f307, %f299, %f306;
|
||
|
|
add.rn.f32 %f309, %f303, %f308;
|
||
|
|
neg.f32 %f310, %f309;
|
||
|
|
add.rn.f32 %f311, %f303, %f310;
|
||
|
|
add.rn.f32 %f312, %f311, %f308;
|
||
|
|
mov.b32 %r225, %f309;
|
||
|
|
setp.eq.s32 %p46, %r225, 1118925336;
|
||
|
|
add.s32 %r226, %r225, -1;
|
||
|
|
mov.b32 %f313, %r226;
|
||
|
|
add.f32 %f314, %f312, 0f37000000;
|
||
|
|
selp.f32 %f315, %f313, %f309, %p46;
|
||
|
|
selp.f32 %f71, %f314, %f312, %p46;
|
||
|
|
mul.f32 %f316, %f315, 0f3FB8AA3B;
|
||
|
|
cvt.rzi.f32.f32 %f317, %f316;
|
||
|
|
mov.f32 %f318, 0fBF317200;
|
||
|
|
fma.rn.f32 %f319, %f317, %f318, %f315;
|
||
|
|
mov.f32 %f320, 0fB5BFBE8E;
|
||
|
|
fma.rn.f32 %f321, %f317, %f320, %f319;
|
||
|
|
mul.f32 %f322, %f321, 0f3FB8AA3B;
|
||
|
|
ex2.approx.ftz.f32 %f323, %f322;
|
||
|
|
add.f32 %f324, %f317, 0f00000000;
|
||
|
|
ex2.approx.f32 %f325, %f324;
|
||
|
|
mul.f32 %f326, %f323, %f325;
|
||
|
|
setp.lt.f32 %p47, %f315, 0fC2D20000;
|
||
|
|
selp.f32 %f327, 0f00000000, %f326, %p47;
|
||
|
|
setp.gt.f32 %p48, %f315, 0f42D20000;
|
||
|
|
selp.f32 %f560, 0f7F800000, %f327, %p48;
|
||
|
|
setp.eq.f32 %p49, %f560, 0f7F800000;
|
||
|
|
@%p49 bra BB0_59;
|
||
|
|
|
||
|
|
fma.rn.f32 %f560, %f560, %f71, %f560;
|
||
|
|
|
||
|
|
BB0_59:
|
||
|
|
setp.lt.f32 %p50, %f66, 0f00000000;
|
||
|
|
setp.eq.f32 %p51, %f69, 0f3F800000;
|
||
|
|
and.pred %p1, %p50, %p51;
|
||
|
|
mov.b32 %r227, %f560;
|
||
|
|
xor.b32 %r228, %r227, -2147483648;
|
||
|
|
mov.b32 %f328, %r228;
|
||
|
|
selp.f32 %f562, %f328, %f560, %p1;
|
||
|
|
setp.eq.f32 %p52, %f66, 0f00000000;
|
||
|
|
@%p52 bra BB0_62;
|
||
|
|
bra.uni BB0_60;
|
||
|
|
|
||
|
|
BB0_62:
|
||
|
|
add.f32 %f331, %f66, %f66;
|
||
|
|
selp.f32 %f562, %f331, 0f00000000, %p51;
|
||
|
|
bra.uni BB0_63;
|
||
|
|
|
||
|
|
BB0_100:
|
||
|
|
mov.u64 %rd105, image_HDR;
|
||
|
|
cvta.global.u64 %rd100, %rd105;
|
||
|
|
mov.u32 %r271, 8;
|
||
|
|
// inline asm
|
||
|
|
call (%rd99), _rt_buffer_get_64, (%rd100, %r91, %r271, %rd17, %rd18, %rd24, %rd24);
|
||
|
|
// inline asm
|
||
|
|
mov.f32 %f512, 0f00000000;
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs40, %f512;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
mov.u16 %rs41, 0;
|
||
|
|
st.v4.u16 [%rd99], {%rs40, %rs40, %rs40, %rs41};
|
||
|
|
bra.uni BB0_101;
|
||
|
|
|
||
|
|
BB0_60:
|
||
|
|
setp.geu.f32 %p53, %f66, 0f00000000;
|
||
|
|
@%p53 bra BB0_63;
|
||
|
|
|
||
|
|
cvt.rzi.f32.f32 %f330, %f302;
|
||
|
|
setp.neu.f32 %p54, %f330, 0f3EE66666;
|
||
|
|
selp.f32 %f562, 0f7FFFFFFF, %f562, %p54;
|
||
|
|
|
||
|
|
BB0_63:
|
||
|
|
abs.f32 %f513, %f66;
|
||
|
|
add.f32 %f332, %f513, 0f3EE66666;
|
||
|
|
mov.b32 %r229, %f332;
|
||
|
|
setp.lt.s32 %p56, %r229, 2139095040;
|
||
|
|
@%p56 bra BB0_68;
|
||
|
|
|
||
|
|
abs.f32 %f530, %f66;
|
||
|
|
setp.gtu.f32 %p57, %f530, 0f7F800000;
|
||
|
|
@%p57 bra BB0_67;
|
||
|
|
bra.uni BB0_65;
|
||
|
|
|
||
|
|
BB0_67:
|
||
|
|
add.f32 %f562, %f66, 0f3EE66666;
|
||
|
|
bra.uni BB0_68;
|
||
|
|
|
||
|
|
BB0_65:
|
||
|
|
abs.f32 %f531, %f66;
|
||
|
|
setp.neu.f32 %p58, %f531, 0f7F800000;
|
||
|
|
@%p58 bra BB0_68;
|
||
|
|
|
||
|
|
selp.f32 %f562, 0fFF800000, 0f7F800000, %p1;
|
||
|
|
|
||
|
|
BB0_68:
|
||
|
|
mov.f32 %f521, 0fB5BFBE8E;
|
||
|
|
mov.f32 %f520, 0fBF317200;
|
||
|
|
mov.f32 %f519, 0f00000000;
|
||
|
|
mov.f32 %f518, 0f35BFBE8E;
|
||
|
|
mov.f32 %f517, 0f3F317200;
|
||
|
|
mov.f32 %f516, 0f3DAAAABD;
|
||
|
|
mov.f32 %f515, 0f3C4CAF63;
|
||
|
|
mov.f32 %f514, 0f3B18F0FE;
|
||
|
|
setp.eq.f32 %p59, %f66, 0f3F800000;
|
||
|
|
selp.f32 %f82, 0f3F800000, %f562, %p59;
|
||
|
|
abs.f32 %f83, %f67;
|
||
|
|
setp.lt.f32 %p60, %f83, 0f00800000;
|
||
|
|
mul.f32 %f335, %f83, 0f4B800000;
|
||
|
|
selp.f32 %f336, 0fC3170000, 0fC2FE0000, %p60;
|
||
|
|
selp.f32 %f337, %f335, %f83, %p60;
|
||
|
|
mov.b32 %r230, %f337;
|
||
|
|
and.b32 %r231, %r230, 8388607;
|
||
|
|
or.b32 %r232, %r231, 1065353216;
|
||
|
|
mov.b32 %f338, %r232;
|
||
|
|
shr.u32 %r233, %r230, 23;
|
||
|
|
cvt.rn.f32.u32 %f339, %r233;
|
||
|
|
add.f32 %f340, %f336, %f339;
|
||
|
|
setp.gt.f32 %p61, %f338, 0f3FB504F3;
|
||
|
|
mul.f32 %f341, %f338, 0f3F000000;
|
||
|
|
add.f32 %f342, %f340, 0f3F800000;
|
||
|
|
selp.f32 %f343, %f341, %f338, %p61;
|
||
|
|
selp.f32 %f344, %f342, %f340, %p61;
|
||
|
|
add.f32 %f345, %f343, 0fBF800000;
|
||
|
|
add.f32 %f334, %f343, 0f3F800000;
|
||
|
|
// inline asm
|
||
|
|
rcp.approx.ftz.f32 %f333,%f334;
|
||
|
|
// inline asm
|
||
|
|
add.f32 %f346, %f345, %f345;
|
||
|
|
mul.f32 %f347, %f333, %f346;
|
||
|
|
mul.f32 %f348, %f347, %f347;
|
||
|
|
fma.rn.f32 %f351, %f514, %f348, %f515;
|
||
|
|
fma.rn.f32 %f353, %f351, %f348, %f516;
|
||
|
|
mul.rn.f32 %f354, %f353, %f348;
|
||
|
|
mul.rn.f32 %f355, %f354, %f347;
|
||
|
|
sub.f32 %f356, %f345, %f347;
|
||
|
|
neg.f32 %f357, %f347;
|
||
|
|
add.f32 %f358, %f356, %f356;
|
||
|
|
fma.rn.f32 %f359, %f357, %f345, %f358;
|
||
|
|
mul.rn.f32 %f360, %f333, %f359;
|
||
|
|
add.f32 %f361, %f355, %f347;
|
||
|
|
sub.f32 %f362, %f347, %f361;
|
||
|
|
add.f32 %f363, %f355, %f362;
|
||
|
|
add.f32 %f364, %f360, %f363;
|
||
|
|
add.f32 %f365, %f361, %f364;
|
||
|
|
sub.f32 %f366, %f361, %f365;
|
||
|
|
add.f32 %f367, %f364, %f366;
|
||
|
|
mul.rn.f32 %f369, %f344, %f517;
|
||
|
|
mul.rn.f32 %f371, %f344, %f518;
|
||
|
|
add.f32 %f372, %f369, %f365;
|
||
|
|
sub.f32 %f373, %f369, %f372;
|
||
|
|
add.f32 %f374, %f365, %f373;
|
||
|
|
add.f32 %f375, %f367, %f374;
|
||
|
|
add.f32 %f376, %f371, %f375;
|
||
|
|
add.f32 %f377, %f372, %f376;
|
||
|
|
sub.f32 %f378, %f372, %f377;
|
||
|
|
add.f32 %f379, %f376, %f378;
|
||
|
|
mul.rn.f32 %f381, %f302, %f377;
|
||
|
|
neg.f32 %f382, %f381;
|
||
|
|
fma.rn.f32 %f383, %f302, %f377, %f382;
|
||
|
|
fma.rn.f32 %f384, %f302, %f379, %f383;
|
||
|
|
fma.rn.f32 %f386, %f519, %f377, %f384;
|
||
|
|
add.rn.f32 %f387, %f381, %f386;
|
||
|
|
neg.f32 %f388, %f387;
|
||
|
|
add.rn.f32 %f389, %f381, %f388;
|
||
|
|
add.rn.f32 %f390, %f389, %f386;
|
||
|
|
mov.b32 %r234, %f387;
|
||
|
|
setp.eq.s32 %p62, %r234, 1118925336;
|
||
|
|
add.s32 %r235, %r234, -1;
|
||
|
|
mov.b32 %f391, %r235;
|
||
|
|
add.f32 %f392, %f390, 0f37000000;
|
||
|
|
selp.f32 %f393, %f391, %f387, %p62;
|
||
|
|
selp.f32 %f84, %f392, %f390, %p62;
|
||
|
|
mul.f32 %f394, %f393, 0f3FB8AA3B;
|
||
|
|
cvt.rzi.f32.f32 %f395, %f394;
|
||
|
|
fma.rn.f32 %f397, %f395, %f520, %f393;
|
||
|
|
fma.rn.f32 %f399, %f395, %f521, %f397;
|
||
|
|
mul.f32 %f400, %f399, 0f3FB8AA3B;
|
||
|
|
ex2.approx.ftz.f32 %f401, %f400;
|
||
|
|
add.f32 %f402, %f395, 0f00000000;
|
||
|
|
ex2.approx.f32 %f403, %f402;
|
||
|
|
mul.f32 %f404, %f401, %f403;
|
||
|
|
setp.lt.f32 %p63, %f393, 0fC2D20000;
|
||
|
|
selp.f32 %f405, 0f00000000, %f404, %p63;
|
||
|
|
setp.gt.f32 %p64, %f393, 0f42D20000;
|
||
|
|
selp.f32 %f563, 0f7F800000, %f405, %p64;
|
||
|
|
setp.eq.f32 %p65, %f563, 0f7F800000;
|
||
|
|
@%p65 bra BB0_70;
|
||
|
|
|
||
|
|
fma.rn.f32 %f563, %f563, %f84, %f563;
|
||
|
|
|
||
|
|
BB0_70:
|
||
|
|
setp.lt.f32 %p66, %f67, 0f00000000;
|
||
|
|
and.pred %p2, %p66, %p51;
|
||
|
|
mov.b32 %r236, %f563;
|
||
|
|
xor.b32 %r237, %r236, -2147483648;
|
||
|
|
mov.b32 %f406, %r237;
|
||
|
|
selp.f32 %f565, %f406, %f563, %p2;
|
||
|
|
setp.eq.f32 %p68, %f67, 0f00000000;
|
||
|
|
@%p68 bra BB0_73;
|
||
|
|
bra.uni BB0_71;
|
||
|
|
|
||
|
|
BB0_73:
|
||
|
|
add.f32 %f409, %f67, %f67;
|
||
|
|
selp.f32 %f565, %f409, 0f00000000, %p51;
|
||
|
|
bra.uni BB0_74;
|
||
|
|
|
||
|
|
BB0_71:
|
||
|
|
setp.geu.f32 %p69, %f67, 0f00000000;
|
||
|
|
@%p69 bra BB0_74;
|
||
|
|
|
||
|
|
mov.f32 %f537, 0f3EE66666;
|
||
|
|
cvt.rzi.f32.f32 %f408, %f537;
|
||
|
|
setp.neu.f32 %p70, %f408, 0f3EE66666;
|
||
|
|
selp.f32 %f565, 0f7FFFFFFF, %f565, %p70;
|
||
|
|
|
||
|
|
BB0_74:
|
||
|
|
abs.f32 %f532, %f67;
|
||
|
|
add.f32 %f410, %f532, 0f3EE66666;
|
||
|
|
mov.b32 %r238, %f410;
|
||
|
|
setp.lt.s32 %p72, %r238, 2139095040;
|
||
|
|
@%p72 bra BB0_79;
|
||
|
|
|
||
|
|
abs.f32 %f535, %f67;
|
||
|
|
setp.gtu.f32 %p73, %f535, 0f7F800000;
|
||
|
|
@%p73 bra BB0_78;
|
||
|
|
bra.uni BB0_76;
|
||
|
|
|
||
|
|
BB0_78:
|
||
|
|
add.f32 %f565, %f67, 0f3EE66666;
|
||
|
|
bra.uni BB0_79;
|
||
|
|
|
||
|
|
BB0_76:
|
||
|
|
abs.f32 %f536, %f67;
|
||
|
|
setp.neu.f32 %p74, %f536, 0f7F800000;
|
||
|
|
@%p74 bra BB0_79;
|
||
|
|
|
||
|
|
selp.f32 %f565, 0fFF800000, 0f7F800000, %p2;
|
||
|
|
|
||
|
|
BB0_79:
|
||
|
|
mov.f32 %f533, 0f3EE66666;
|
||
|
|
mov.f32 %f529, 0fB5BFBE8E;
|
||
|
|
mov.f32 %f528, 0fBF317200;
|
||
|
|
mov.f32 %f527, 0f00000000;
|
||
|
|
mov.f32 %f526, 0f35BFBE8E;
|
||
|
|
mov.f32 %f525, 0f3F317200;
|
||
|
|
mov.f32 %f524, 0f3DAAAABD;
|
||
|
|
mov.f32 %f523, 0f3C4CAF63;
|
||
|
|
mov.f32 %f522, 0f3B18F0FE;
|
||
|
|
setp.eq.f32 %p75, %f67, 0f3F800000;
|
||
|
|
selp.f32 %f95, 0f3F800000, %f565, %p75;
|
||
|
|
abs.f32 %f96, %f68;
|
||
|
|
setp.lt.f32 %p76, %f96, 0f00800000;
|
||
|
|
mul.f32 %f413, %f96, 0f4B800000;
|
||
|
|
selp.f32 %f414, 0fC3170000, 0fC2FE0000, %p76;
|
||
|
|
selp.f32 %f415, %f413, %f96, %p76;
|
||
|
|
mov.b32 %r239, %f415;
|
||
|
|
and.b32 %r240, %r239, 8388607;
|
||
|
|
or.b32 %r241, %r240, 1065353216;
|
||
|
|
mov.b32 %f416, %r241;
|
||
|
|
shr.u32 %r242, %r239, 23;
|
||
|
|
cvt.rn.f32.u32 %f417, %r242;
|
||
|
|
add.f32 %f418, %f414, %f417;
|
||
|
|
setp.gt.f32 %p77, %f416, 0f3FB504F3;
|
||
|
|
mul.f32 %f419, %f416, 0f3F000000;
|
||
|
|
add.f32 %f420, %f418, 0f3F800000;
|
||
|
|
selp.f32 %f421, %f419, %f416, %p77;
|
||
|
|
selp.f32 %f422, %f420, %f418, %p77;
|
||
|
|
add.f32 %f423, %f421, 0fBF800000;
|
||
|
|
add.f32 %f412, %f421, 0f3F800000;
|
||
|
|
// inline asm
|
||
|
|
rcp.approx.ftz.f32 %f411,%f412;
|
||
|
|
// inline asm
|
||
|
|
add.f32 %f424, %f423, %f423;
|
||
|
|
mul.f32 %f425, %f411, %f424;
|
||
|
|
mul.f32 %f426, %f425, %f425;
|
||
|
|
fma.rn.f32 %f429, %f522, %f426, %f523;
|
||
|
|
fma.rn.f32 %f431, %f429, %f426, %f524;
|
||
|
|
mul.rn.f32 %f432, %f431, %f426;
|
||
|
|
mul.rn.f32 %f433, %f432, %f425;
|
||
|
|
sub.f32 %f434, %f423, %f425;
|
||
|
|
neg.f32 %f435, %f425;
|
||
|
|
add.f32 %f436, %f434, %f434;
|
||
|
|
fma.rn.f32 %f437, %f435, %f423, %f436;
|
||
|
|
mul.rn.f32 %f438, %f411, %f437;
|
||
|
|
add.f32 %f439, %f433, %f425;
|
||
|
|
sub.f32 %f440, %f425, %f439;
|
||
|
|
add.f32 %f441, %f433, %f440;
|
||
|
|
add.f32 %f442, %f438, %f441;
|
||
|
|
add.f32 %f443, %f439, %f442;
|
||
|
|
sub.f32 %f444, %f439, %f443;
|
||
|
|
add.f32 %f445, %f442, %f444;
|
||
|
|
mul.rn.f32 %f447, %f422, %f525;
|
||
|
|
mul.rn.f32 %f449, %f422, %f526;
|
||
|
|
add.f32 %f450, %f447, %f443;
|
||
|
|
sub.f32 %f451, %f447, %f450;
|
||
|
|
add.f32 %f452, %f443, %f451;
|
||
|
|
add.f32 %f453, %f445, %f452;
|
||
|
|
add.f32 %f454, %f449, %f453;
|
||
|
|
add.f32 %f455, %f450, %f454;
|
||
|
|
sub.f32 %f456, %f450, %f455;
|
||
|
|
add.f32 %f457, %f454, %f456;
|
||
|
|
mul.rn.f32 %f459, %f533, %f455;
|
||
|
|
neg.f32 %f460, %f459;
|
||
|
|
fma.rn.f32 %f461, %f533, %f455, %f460;
|
||
|
|
fma.rn.f32 %f462, %f533, %f457, %f461;
|
||
|
|
fma.rn.f32 %f464, %f527, %f455, %f462;
|
||
|
|
add.rn.f32 %f465, %f459, %f464;
|
||
|
|
neg.f32 %f466, %f465;
|
||
|
|
add.rn.f32 %f467, %f459, %f466;
|
||
|
|
add.rn.f32 %f468, %f467, %f464;
|
||
|
|
mov.b32 %r243, %f465;
|
||
|
|
setp.eq.s32 %p78, %r243, 1118925336;
|
||
|
|
add.s32 %r244, %r243, -1;
|
||
|
|
mov.b32 %f469, %r244;
|
||
|
|
add.f32 %f470, %f468, 0f37000000;
|
||
|
|
selp.f32 %f471, %f469, %f465, %p78;
|
||
|
|
selp.f32 %f97, %f470, %f468, %p78;
|
||
|
|
mul.f32 %f472, %f471, 0f3FB8AA3B;
|
||
|
|
cvt.rzi.f32.f32 %f473, %f472;
|
||
|
|
fma.rn.f32 %f475, %f473, %f528, %f471;
|
||
|
|
fma.rn.f32 %f477, %f473, %f529, %f475;
|
||
|
|
mul.f32 %f478, %f477, 0f3FB8AA3B;
|
||
|
|
ex2.approx.ftz.f32 %f479, %f478;
|
||
|
|
add.f32 %f480, %f473, 0f00000000;
|
||
|
|
ex2.approx.f32 %f481, %f480;
|
||
|
|
mul.f32 %f482, %f479, %f481;
|
||
|
|
setp.lt.f32 %p79, %f471, 0fC2D20000;
|
||
|
|
selp.f32 %f483, 0f00000000, %f482, %p79;
|
||
|
|
setp.gt.f32 %p80, %f471, 0f42D20000;
|
||
|
|
selp.f32 %f566, 0f7F800000, %f483, %p80;
|
||
|
|
setp.eq.f32 %p81, %f566, 0f7F800000;
|
||
|
|
@%p81 bra BB0_81;
|
||
|
|
|
||
|
|
fma.rn.f32 %f566, %f566, %f97, %f566;
|
||
|
|
|
||
|
|
BB0_81:
|
||
|
|
setp.lt.f32 %p82, %f68, 0f00000000;
|
||
|
|
and.pred %p3, %p82, %p51;
|
||
|
|
mov.b32 %r245, %f566;
|
||
|
|
xor.b32 %r246, %r245, -2147483648;
|
||
|
|
mov.b32 %f484, %r246;
|
||
|
|
selp.f32 %f568, %f484, %f566, %p3;
|
||
|
|
setp.eq.f32 %p84, %f68, 0f00000000;
|
||
|
|
@%p84 bra BB0_84;
|
||
|
|
bra.uni BB0_82;
|
||
|
|
|
||
|
|
BB0_84:
|
||
|
|
add.f32 %f487, %f68, %f68;
|
||
|
|
selp.f32 %f568, %f487, 0f00000000, %p51;
|
||
|
|
bra.uni BB0_85;
|
||
|
|
|
||
|
|
BB0_82:
|
||
|
|
setp.geu.f32 %p85, %f68, 0f00000000;
|
||
|
|
@%p85 bra BB0_85;
|
||
|
|
|
||
|
|
mov.f32 %f534, 0f3EE66666;
|
||
|
|
cvt.rzi.f32.f32 %f486, %f534;
|
||
|
|
setp.neu.f32 %p86, %f486, 0f3EE66666;
|
||
|
|
selp.f32 %f568, 0f7FFFFFFF, %f568, %p86;
|
||
|
|
|
||
|
|
BB0_85:
|
||
|
|
abs.f32 %f538, %f68;
|
||
|
|
add.f32 %f488, %f538, 0f3EE66666;
|
||
|
|
mov.b32 %r247, %f488;
|
||
|
|
setp.lt.s32 %p88, %r247, 2139095040;
|
||
|
|
@%p88 bra BB0_90;
|
||
|
|
|
||
|
|
abs.f32 %f539, %f68;
|
||
|
|
setp.gtu.f32 %p89, %f539, 0f7F800000;
|
||
|
|
@%p89 bra BB0_89;
|
||
|
|
bra.uni BB0_87;
|
||
|
|
|
||
|
|
BB0_89:
|
||
|
|
add.f32 %f568, %f68, 0f3EE66666;
|
||
|
|
bra.uni BB0_90;
|
||
|
|
|
||
|
|
BB0_87:
|
||
|
|
abs.f32 %f540, %f68;
|
||
|
|
setp.neu.f32 %p90, %f540, 0f7F800000;
|
||
|
|
@%p90 bra BB0_90;
|
||
|
|
|
||
|
|
selp.f32 %f568, 0fFF800000, 0f7F800000, %p3;
|
||
|
|
|
||
|
|
BB0_90:
|
||
|
|
mov.u32 %r273, 4;
|
||
|
|
mov.u64 %rd106, 0;
|
||
|
|
mov.u32 %r272, 2;
|
||
|
|
setp.eq.f32 %p91, %f68, 0f3F800000;
|
||
|
|
selp.f32 %f489, 0f3F800000, %f568, %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, %r272, %r273, %rd54, %rd55, %rd106, %rd106);
|
||
|
|
// inline asm
|
||
|
|
cvt.sat.f32.f32 %f490, %f489;
|
||
|
|
mul.f32 %f491, %f490, 0f437FFD71;
|
||
|
|
cvt.rzi.u32.f32 %r250, %f491;
|
||
|
|
cvt.sat.f32.f32 %f492, %f95;
|
||
|
|
mul.f32 %f493, %f492, 0f437FFD71;
|
||
|
|
cvt.rzi.u32.f32 %r251, %f493;
|
||
|
|
cvt.sat.f32.f32 %f494, %f82;
|
||
|
|
mul.f32 %f495, %f494, 0f437FFD71;
|
||
|
|
cvt.rzi.u32.f32 %r252, %f495;
|
||
|
|
cvt.u16.u32 %rs10, %r250;
|
||
|
|
cvt.u16.u32 %rs11, %r252;
|
||
|
|
cvt.u16.u32 %rs12, %r251;
|
||
|
|
mov.u16 %rs13, 255;
|
||
|
|
st.v4.u8 [%rd52], {%rs10, %rs12, %rs11, %rs13};
|
||
|
|
ld.global.u32 %r300, [imageEnabled];
|
||
|
|
|
||
|
|
BB0_91:
|
||
|
|
and.b32 %r253, %r300, 4;
|
||
|
|
setp.eq.s32 %p92, %r253, 0;
|
||
|
|
@%p92 bra BB0_101;
|
||
|
|
|
||
|
|
ld.global.u32 %r254, [additive];
|
||
|
|
setp.eq.s32 %p93, %r254, 0;
|
||
|
|
cvt.u64.u32 %rd15, %r3;
|
||
|
|
cvt.u64.u32 %rd16, %r4;
|
||
|
|
mov.f32 %f496, 0f3F800000;
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs14, %f496;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
@%p93 bra BB0_94;
|
||
|
|
|
||
|
|
mov.u64 %rd107, 0;
|
||
|
|
mov.u32 %r274, 2;
|
||
|
|
mov.u64 %rd71, image_HDR;
|
||
|
|
cvta.global.u64 %rd60, %rd71;
|
||
|
|
mov.u32 %r258, 8;
|
||
|
|
// inline asm
|
||
|
|
call (%rd59), _rt_buffer_get_64, (%rd60, %r274, %r258, %rd15, %rd16, %rd107, %rd107);
|
||
|
|
// inline asm
|
||
|
|
ld.v4.u16 {%rs21, %rs22, %rs23, %rs24}, [%rd59];
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f497, %rs21;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f498, %rs22;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f499, %rs23;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
call (%rd65), _rt_buffer_get_64, (%rd60, %r274, %r258, %rd15, %rd16, %rd107, %rd107);
|
||
|
|
// inline asm
|
||
|
|
add.f32 %f500, %f66, %f497;
|
||
|
|
add.f32 %f501, %f67, %f498;
|
||
|
|
add.f32 %f502, %f68, %f499;
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs20, %f502;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs19, %f501;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs18, %f500;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
st.v4.u16 [%rd65], {%rs18, %rs19, %rs20, %rs14};
|
||
|
|
bra.uni BB0_101;
|
||
|
|
|
||
|
|
BB0_94:
|
||
|
|
mov.u64 %rd108, 0;
|
||
|
|
mov.u32 %r275, 2;
|
||
|
|
mov.u64 %rd78, image_HDR;
|
||
|
|
cvta.global.u64 %rd73, %rd78;
|
||
|
|
mov.u32 %r260, 8;
|
||
|
|
// inline asm
|
||
|
|
call (%rd72), _rt_buffer_get_64, (%rd73, %r275, %r260, %rd15, %rd16, %rd108, %rd108);
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs27, %f68;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs26, %f67;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs25, %f66;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
st.v4.u16 [%rd72], {%rs25, %rs26, %rs27, %rs14};
|
||
|
|
|
||
|
|
BB0_101:
|
||
|
|
ret;
|
||
|
|
}
|
||
|
|
|
||
|
|
|