2840 lines
74 KiB
Plaintext
2840 lines
74 KiB
Plaintext
//
|
|
// Generated by NVIDIA NVVM Compiler
|
|
//
|
|
// Compiler Build ID: CL-23083092
|
|
// Cuda compilation tools, release 9.1, V9.1.85
|
|
// Based on LLVM 3.4svn
|
|
//
|
|
|
|
.version 6.1
|
|
.target sm_30
|
|
.address_size 64
|
|
|
|
// .globl _Z6oxMainv
|
|
.global .align 8 .b8 pixelID[8];
|
|
.global .align 8 .b8 resolution[8];
|
|
.global .align 4 .b8 normal[12];
|
|
.global .align 4 .b8 camPos[12];
|
|
.global .align 4 .b8 root[4];
|
|
.global .align 4 .u32 imageEnabled;
|
|
.global .texref lightmap;
|
|
.global .align 16 .b8 tileInfo[16];
|
|
.global .align 4 .u32 additive;
|
|
.global .align 1 .b8 image[1];
|
|
.global .align 1 .b8 image_HDR[1];
|
|
.global .align 1 .b8 image_HDR2[1];
|
|
.global .align 1 .b8 image_Mask[1];
|
|
.global .align 1 .b8 image_RNM0[1];
|
|
.global .align 1 .b8 image_RNM1[1];
|
|
.global .align 1 .b8 image_RNM2[1];
|
|
.global .align 1 .b8 image_RNM3[1];
|
|
.global .align 1 .b8 image_RNM4[1];
|
|
.global .align 1 .b8 image_RNM5[1];
|
|
.global .align 1 .b8 image_RNM6[1];
|
|
.global .align 1 .b8 image_RNM7[1];
|
|
.global .align 1 .b8 image_RNM8[1];
|
|
.global .align 1 .b8 uvpos[1];
|
|
.global .align 1 .b8 uvnormal[1];
|
|
.global .align 1 .b8 rnd_seeds[1];
|
|
.global .align 4 .b8 directDir[12];
|
|
.global .align 4 .b8 directColor[12];
|
|
.global .align 4 .f32 shadowSpread;
|
|
.global .align 4 .u32 samples;
|
|
.global .align 4 .u32 ignoreNormal;
|
|
.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_typeinfo9directDirE[8] = {82, 97, 121, 0, 12, 0, 0, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo11directColorE[8] = {82, 97, 121, 0, 12, 0, 0, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo12shadowSpreadE[8] = {82, 97, 121, 0, 4, 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_typeinfo12ignoreNormalE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
|
|
.global .align 8 .u64 _ZN21rti_internal_register20reg_bitness_detectorE;
|
|
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail0E;
|
|
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail1E;
|
|
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail2E;
|
|
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail3E;
|
|
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail4E;
|
|
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail5E;
|
|
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail6E;
|
|
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail7E;
|
|
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail8E;
|
|
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail9E;
|
|
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail0E;
|
|
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail1E;
|
|
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail2E;
|
|
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail3E;
|
|
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail4E;
|
|
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail5E;
|
|
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail6E;
|
|
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail7E;
|
|
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail8E;
|
|
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail9E;
|
|
.global .align 4 .u32 _ZN21rti_internal_register14reg_rayIndex_xE;
|
|
.global .align 4 .u32 _ZN21rti_internal_register14reg_rayIndex_yE;
|
|
.global .align 4 .u32 _ZN21rti_internal_register14reg_rayIndex_zE;
|
|
.global .align 8 .b8 _ZN21rti_internal_typename7pixelIDE[6] = {117, 105, 110, 116, 50, 0};
|
|
.global .align 8 .b8 _ZN21rti_internal_typename10resolutionE[6] = {117, 105, 110, 116, 50, 0};
|
|
.global .align 8 .b8 _ZN21rti_internal_typename6normalE[7] = {102, 108, 111, 97, 116, 51, 0};
|
|
.global .align 8 .b8 _ZN21rti_internal_typename6camPosE[7] = {102, 108, 111, 97, 116, 51, 0};
|
|
.global .align 16 .b8 _ZN21rti_internal_typename4rootE[9] = {114, 116, 79, 98, 106, 101, 99, 116, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typename12imageEnabledE[4] = {105, 110, 116, 0};
|
|
.global .align 8 .b8 _ZN21rti_internal_typename8tileInfoE[6] = {117, 105, 110, 116, 52, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typename8additiveE[4] = {105, 110, 116, 0};
|
|
.global .align 8 .b8 _ZN21rti_internal_typename9directDirE[7] = {102, 108, 111, 97, 116, 51, 0};
|
|
.global .align 8 .b8 _ZN21rti_internal_typename11directColorE[7] = {102, 108, 111, 97, 116, 51, 0};
|
|
.global .align 8 .b8 _ZN21rti_internal_typename12shadowSpreadE[6] = {102, 108, 111, 97, 116, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typename7samplesE[4] = {105, 110, 116, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typename12ignoreNormalE[4] = {105, 110, 116, 0};
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum7pixelIDE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum10resolutionE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum6normalE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum6camPosE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum4rootE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum12imageEnabledE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum8tileInfoE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum8additiveE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum9directDirE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum11directColorE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum12shadowSpreadE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum7samplesE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum12ignoreNormalE = 4919;
|
|
.global .align 16 .b8 _ZN21rti_internal_semantic7pixelIDE[14] = {114, 116, 76, 97, 117, 110, 99, 104, 73, 110, 100, 101, 120, 0};
|
|
.global .align 16 .b8 _ZN21rti_internal_semantic10resolutionE[12] = {114, 116, 76, 97, 117, 110, 99, 104, 68, 105, 109, 0};
|
|
.global .align 16 .b8 _ZN21rti_internal_semantic6normalE[17] = {97, 116, 116, 114, 105, 98, 117, 116, 101, 32, 110, 111, 114, 109, 97, 108, 0};
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic6camPosE[1];
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic4rootE[1];
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic12imageEnabledE[1];
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic8tileInfoE[1];
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic8additiveE[1];
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic9directDirE[1];
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic11directColorE[1];
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic12shadowSpreadE[1];
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic7samplesE[1];
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic12ignoreNormalE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation7pixelIDE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation10resolutionE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation6normalE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation6camPosE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation4rootE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation12imageEnabledE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation8tileInfoE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation8additiveE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation9directDirE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation11directColorE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation12shadowSpreadE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation7samplesE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation12ignoreNormalE[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<134>;
|
|
.reg .b16 %rs<298>;
|
|
.reg .f32 %f<930>;
|
|
.reg .b32 %r<458>;
|
|
.reg .b64 %rd<486>;
|
|
|
|
|
|
mov.u64 %rd485, __local_depot0;
|
|
cvta.local.u64 %SP, %rd485;
|
|
ld.global.v2.u32 {%r101, %r102}, [pixelID];
|
|
cvt.u64.u32 %rd22, %r101;
|
|
cvt.u64.u32 %rd23, %r102;
|
|
mov.u64 %rd26, uvnormal;
|
|
cvta.global.u64 %rd21, %rd26;
|
|
mov.u32 %r99, 2;
|
|
mov.u32 %r100, 4;
|
|
mov.u64 %rd25, 0;
|
|
// inline asm
|
|
call (%rd20), _rt_buffer_get_64, (%rd21, %r99, %r100, %rd22, %rd23, %rd25, %rd25);
|
|
// inline asm
|
|
ld.u32 %r1, [%rd20];
|
|
shr.u32 %r105, %r1, 16;
|
|
cvt.u16.u32 %rs1, %r105;
|
|
and.b16 %rs12, %rs1, 255;
|
|
cvt.u16.u32 %rs13, %r1;
|
|
or.b16 %rs14, %rs13, %rs12;
|
|
setp.eq.s16 %p5, %rs14, 0;
|
|
mov.f32 %f898, 0f00000000;
|
|
mov.f32 %f899, %f898;
|
|
mov.f32 %f900, %f898;
|
|
@%p5 bra BB0_2;
|
|
|
|
ld.u8 %rs15, [%rd20+1];
|
|
and.b16 %rs17, %rs13, 255;
|
|
cvt.rn.f32.u16 %f161, %rs17;
|
|
div.rn.f32 %f162, %f161, 0f437F0000;
|
|
fma.rn.f32 %f163, %f162, 0f40000000, 0fBF800000;
|
|
cvt.rn.f32.u16 %f164, %rs15;
|
|
div.rn.f32 %f165, %f164, 0f437F0000;
|
|
fma.rn.f32 %f166, %f165, 0f40000000, 0fBF800000;
|
|
cvt.rn.f32.u16 %f167, %rs12;
|
|
div.rn.f32 %f168, %f167, 0f437F0000;
|
|
fma.rn.f32 %f169, %f168, 0f40000000, 0fBF800000;
|
|
mul.f32 %f170, %f166, %f166;
|
|
fma.rn.f32 %f171, %f163, %f163, %f170;
|
|
fma.rn.f32 %f172, %f169, %f169, %f171;
|
|
sqrt.rn.f32 %f173, %f172;
|
|
rcp.rn.f32 %f174, %f173;
|
|
mul.f32 %f898, %f163, %f174;
|
|
mul.f32 %f899, %f166, %f174;
|
|
mul.f32 %f900, %f169, %f174;
|
|
|
|
BB0_2:
|
|
ld.global.v2.u32 {%r106, %r107}, [pixelID];
|
|
ld.global.v2.u32 {%r109, %r110}, [tileInfo];
|
|
add.s32 %r2, %r106, %r109;
|
|
add.s32 %r3, %r107, %r110;
|
|
setp.eq.f32 %p6, %f899, 0f00000000;
|
|
setp.eq.f32 %p7, %f898, 0f00000000;
|
|
and.pred %p8, %p7, %p6;
|
|
setp.eq.f32 %p9, %f900, 0f00000000;
|
|
and.pred %p10, %p8, %p9;
|
|
@%p10 bra BB0_136;
|
|
bra.uni BB0_3;
|
|
|
|
BB0_136:
|
|
ld.global.u32 %r457, [imageEnabled];
|
|
and.b32 %r347, %r457, 1;
|
|
setp.eq.b32 %p121, %r347, 1;
|
|
@!%p121 bra BB0_138;
|
|
bra.uni BB0_137;
|
|
|
|
BB0_137:
|
|
cvt.u64.u32 %rd269, %r2;
|
|
cvt.u64.u32 %rd270, %r3;
|
|
mov.u64 %rd273, image;
|
|
cvta.global.u64 %rd268, %rd273;
|
|
// inline asm
|
|
call (%rd267), _rt_buffer_get_64, (%rd268, %r99, %r100, %rd269, %rd270, %rd25, %rd25);
|
|
// inline asm
|
|
mov.u16 %rs165, 0;
|
|
st.v4.u8 [%rd267], {%rs165, %rs165, %rs165, %rs165};
|
|
ld.global.u32 %r457, [imageEnabled];
|
|
|
|
BB0_138:
|
|
and.b32 %r350, %r457, 8;
|
|
setp.eq.s32 %p122, %r350, 0;
|
|
@%p122 bra BB0_140;
|
|
|
|
cvt.u64.u32 %rd277, %r3;
|
|
cvt.u64.u32 %rd276, %r2;
|
|
mov.u64 %rd280, image_Mask;
|
|
cvta.global.u64 %rd275, %rd280;
|
|
// inline asm
|
|
call (%rd274), _rt_buffer_get_64, (%rd275, %r99, %r99, %rd276, %rd277, %rd25, %rd25);
|
|
// inline asm
|
|
mov.f32 %f785, 0f00000000;
|
|
cvt.rzi.u32.f32 %r353, %f785;
|
|
cvt.u16.u32 %rs166, %r353;
|
|
mov.u16 %rs167, 0;
|
|
st.v2.u8 [%rd274], {%rs166, %rs167};
|
|
ld.global.u32 %r457, [imageEnabled];
|
|
|
|
BB0_140:
|
|
cvt.u64.u32 %rd18, %r2;
|
|
cvt.u64.u32 %rd19, %r3;
|
|
and.b32 %r354, %r457, 4;
|
|
setp.eq.s32 %p123, %r354, 0;
|
|
@%p123 bra BB0_144;
|
|
|
|
ld.global.u32 %r355, [additive];
|
|
setp.eq.s32 %p124, %r355, 0;
|
|
@%p124 bra BB0_143;
|
|
|
|
mov.u64 %rd293, image_HDR;
|
|
cvta.global.u64 %rd282, %rd293;
|
|
mov.u32 %r359, 8;
|
|
// inline asm
|
|
call (%rd281), _rt_buffer_get_64, (%rd282, %r99, %r359, %rd18, %rd19, %rd25, %rd25);
|
|
// inline asm
|
|
ld.v4.u16 {%rs174, %rs175, %rs176, %rs177}, [%rd281];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f786, %rs174;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f787, %rs175;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f788, %rs176;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd287), _rt_buffer_get_64, (%rd282, %r99, %r359, %rd18, %rd19, %rd25, %rd25);
|
|
// inline asm
|
|
add.f32 %f789, %f786, 0f00000000;
|
|
add.f32 %f790, %f787, 0f00000000;
|
|
add.f32 %f791, %f788, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs173, %f791;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs172, %f790;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs171, %f789;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs178, 0;
|
|
st.v4.u16 [%rd287], {%rs171, %rs172, %rs173, %rs178};
|
|
bra.uni BB0_144;
|
|
|
|
BB0_3:
|
|
ld.global.f32 %f9, [directDir+8];
|
|
ld.global.f32 %f8, [directDir+4];
|
|
ld.global.f32 %f7, [directDir];
|
|
mul.f32 %f176, %f898, %f7;
|
|
mul.f32 %f177, %f899, %f8;
|
|
neg.f32 %f178, %f177;
|
|
sub.f32 %f179, %f178, %f176;
|
|
mul.f32 %f180, %f900, %f9;
|
|
sub.f32 %f10, %f179, %f180;
|
|
ld.global.v2.u32 {%r118, %r119}, [pixelID];
|
|
cvt.u64.u32 %rd29, %r118;
|
|
cvt.u64.u32 %rd30, %r119;
|
|
mov.u64 %rd39, uvpos;
|
|
cvta.global.u64 %rd28, %rd39;
|
|
mov.u32 %r115, 12;
|
|
// inline asm
|
|
call (%rd27), _rt_buffer_get_64, (%rd28, %r99, %r115, %rd29, %rd30, %rd25, %rd25);
|
|
// inline asm
|
|
ld.f32 %f181, [%rd27+8];
|
|
ld.f32 %f182, [%rd27+4];
|
|
ld.f32 %f183, [%rd27];
|
|
mul.f32 %f184, %f183, 0f3456BF95;
|
|
mul.f32 %f185, %f182, 0f3456BF95;
|
|
mul.f32 %f186, %f181, 0f3456BF95;
|
|
abs.f32 %f187, %f898;
|
|
div.rn.f32 %f188, %f184, %f187;
|
|
abs.f32 %f189, %f899;
|
|
div.rn.f32 %f190, %f185, %f189;
|
|
abs.f32 %f191, %f900;
|
|
div.rn.f32 %f192, %f186, %f191;
|
|
abs.f32 %f193, %f188;
|
|
abs.f32 %f194, %f190;
|
|
abs.f32 %f195, %f192;
|
|
mov.f32 %f196, 0f38D1B717;
|
|
max.f32 %f197, %f193, %f196;
|
|
max.f32 %f198, %f194, %f196;
|
|
max.f32 %f199, %f195, %f196;
|
|
fma.rn.f32 %f11, %f898, %f197, %f183;
|
|
fma.rn.f32 %f12, %f899, %f198, %f182;
|
|
fma.rn.f32 %f13, %f900, %f199, %f181;
|
|
abs.f32 %f200, %f7;
|
|
abs.f32 %f201, %f9;
|
|
setp.gt.f32 %p11, %f200, %f201;
|
|
neg.f32 %f202, %f8;
|
|
neg.f32 %f203, %f9;
|
|
selp.f32 %f204, %f202, 0f00000000, %p11;
|
|
selp.f32 %f205, %f7, %f203, %p11;
|
|
selp.f32 %f206, 0f00000000, %f8, %p11;
|
|
mul.f32 %f207, %f205, %f205;
|
|
fma.rn.f32 %f208, %f204, %f204, %f207;
|
|
fma.rn.f32 %f209, %f206, %f206, %f208;
|
|
sqrt.rn.f32 %f210, %f209;
|
|
rcp.rn.f32 %f211, %f210;
|
|
mul.f32 %f14, %f204, %f211;
|
|
mul.f32 %f15, %f205, %f211;
|
|
mul.f32 %f16, %f206, %f211;
|
|
ld.global.v2.u32 {%r122, %r123}, [pixelID];
|
|
cvt.u64.u32 %rd35, %r122;
|
|
cvt.u64.u32 %rd36, %r123;
|
|
mov.u64 %rd40, rnd_seeds;
|
|
cvta.global.u64 %rd34, %rd40;
|
|
// inline asm
|
|
call (%rd33), _rt_buffer_get_64, (%rd34, %r99, %r100, %rd35, %rd36, %rd25, %rd25);
|
|
// inline asm
|
|
ld.global.u32 %r426, [samples];
|
|
mov.f32 %f915, 0f00000000;
|
|
setp.lt.s32 %p12, %r426, 1;
|
|
@%p12 bra BB0_55;
|
|
|
|
cvt.rn.f32.s32 %f213, %r426;
|
|
rcp.rn.f32 %f17, %f213;
|
|
ld.u32 %r452, [%rd33];
|
|
mul.f32 %f18, %f11, 0f3456BF95;
|
|
mul.f32 %f19, %f12, 0f3456BF95;
|
|
mul.f32 %f20, %f13, 0f3456BF95;
|
|
mul.f32 %f214, %f7, %f15;
|
|
mul.f32 %f215, %f8, %f14;
|
|
sub.f32 %f21, %f215, %f214;
|
|
mul.f32 %f216, %f9, %f14;
|
|
mul.f32 %f217, %f7, %f16;
|
|
sub.f32 %f22, %f217, %f216;
|
|
mul.f32 %f218, %f8, %f16;
|
|
mul.f32 %f219, %f9, %f15;
|
|
sub.f32 %f23, %f219, %f218;
|
|
mov.f32 %f915, 0f00000000;
|
|
mov.u32 %r427, 0;
|
|
abs.f32 %f220, %f19;
|
|
abs.f32 %f221, %f18;
|
|
max.f32 %f222, %f221, %f220;
|
|
abs.f32 %f223, %f20;
|
|
max.f32 %f224, %f222, %f223;
|
|
|
|
BB0_5:
|
|
setp.lt.s32 %p13, %r426, 1;
|
|
@%p13 bra BB0_54;
|
|
|
|
cvt.rn.f32.s32 %f25, %r427;
|
|
max.f32 %f26, %f224, %f196;
|
|
mov.u32 %r429, 0;
|
|
|
|
BB0_7:
|
|
mad.lo.s32 %r128, %r452, 1664525, 1013904223;
|
|
and.b32 %r129, %r128, 16777215;
|
|
cvt.rn.f32.u32 %f226, %r129;
|
|
fma.rn.f32 %f227, %f226, 0f33800000, %f25;
|
|
mul.f32 %f228, %f17, %f227;
|
|
mad.lo.s32 %r452, %r128, 1664525, 1013904223;
|
|
and.b32 %r130, %r452, 16777215;
|
|
cvt.rn.f32.u32 %f229, %r130;
|
|
cvt.rn.f32.s32 %f230, %r429;
|
|
fma.rn.f32 %f231, %f229, 0f33800000, %f230;
|
|
mul.f32 %f232, %f17, %f231;
|
|
sqrt.rn.f32 %f28, %f228;
|
|
mul.f32 %f909, %f232, 0f40C90FDB;
|
|
abs.f32 %f30, %f909;
|
|
setp.neu.f32 %p14, %f30, 0f7F800000;
|
|
mov.f32 %f903, %f909;
|
|
@%p14 bra BB0_9;
|
|
|
|
mov.f32 %f233, 0f00000000;
|
|
mul.rn.f32 %f903, %f909, %f233;
|
|
|
|
BB0_9:
|
|
mul.f32 %f234, %f903, 0f3F22F983;
|
|
cvt.rni.s32.f32 %r440, %f234;
|
|
cvt.rn.f32.s32 %f235, %r440;
|
|
neg.f32 %f236, %f235;
|
|
mov.f32 %f237, 0f3FC90FDA;
|
|
fma.rn.f32 %f238, %f236, %f237, %f903;
|
|
mov.f32 %f239, 0f33A22168;
|
|
fma.rn.f32 %f240, %f236, %f239, %f238;
|
|
mov.f32 %f241, 0f27C234C5;
|
|
fma.rn.f32 %f904, %f236, %f241, %f240;
|
|
abs.f32 %f242, %f903;
|
|
setp.leu.f32 %p15, %f242, 0f47CE4780;
|
|
@%p15 bra BB0_20;
|
|
|
|
mov.b32 %r13, %f903;
|
|
shr.u32 %r14, %r13, 23;
|
|
shl.b32 %r133, %r13, 8;
|
|
or.b32 %r15, %r133, -2147483648;
|
|
add.u64 %rd42, %SP, 4;
|
|
cvta.to.local.u64 %rd482, %rd42;
|
|
mov.u32 %r432, 0;
|
|
mov.u64 %rd481, __cudart_i2opi_f;
|
|
mov.u32 %r431, -6;
|
|
|
|
BB0_11:
|
|
.pragma "nounroll";
|
|
ld.const.u32 %r136, [%rd481];
|
|
// inline asm
|
|
{
|
|
mad.lo.cc.u32 %r134, %r136, %r15, %r432;
|
|
madc.hi.u32 %r432, %r136, %r15, 0;
|
|
}
|
|
// inline asm
|
|
st.local.u32 [%rd482], %r134;
|
|
add.s64 %rd482, %rd482, 4;
|
|
add.s64 %rd481, %rd481, 4;
|
|
add.s32 %r431, %r431, 1;
|
|
setp.ne.s32 %p16, %r431, 0;
|
|
@%p16 bra BB0_11;
|
|
|
|
and.b32 %r139, %r14, 255;
|
|
add.s32 %r140, %r139, -128;
|
|
shr.u32 %r141, %r140, 5;
|
|
and.b32 %r20, %r13, -2147483648;
|
|
cvta.to.local.u64 %rd44, %rd42;
|
|
st.local.u32 [%rd44+24], %r432;
|
|
mov.u32 %r142, 6;
|
|
sub.s32 %r143, %r142, %r141;
|
|
mul.wide.s32 %rd45, %r143, 4;
|
|
add.s64 %rd8, %rd44, %rd45;
|
|
ld.local.u32 %r433, [%rd8];
|
|
ld.local.u32 %r434, [%rd8+-4];
|
|
and.b32 %r23, %r14, 31;
|
|
setp.eq.s32 %p17, %r23, 0;
|
|
@%p17 bra BB0_14;
|
|
|
|
mov.u32 %r144, 32;
|
|
sub.s32 %r145, %r144, %r23;
|
|
shr.u32 %r146, %r434, %r145;
|
|
shl.b32 %r147, %r433, %r23;
|
|
add.s32 %r433, %r146, %r147;
|
|
ld.local.u32 %r148, [%rd8+-8];
|
|
shr.u32 %r149, %r148, %r145;
|
|
shl.b32 %r150, %r434, %r23;
|
|
add.s32 %r434, %r149, %r150;
|
|
|
|
BB0_14:
|
|
shr.u32 %r151, %r434, 30;
|
|
shl.b32 %r152, %r433, 2;
|
|
add.s32 %r435, %r151, %r152;
|
|
shl.b32 %r29, %r434, 2;
|
|
shr.u32 %r153, %r435, 31;
|
|
shr.u32 %r154, %r433, 30;
|
|
add.s32 %r30, %r153, %r154;
|
|
setp.eq.s32 %p18, %r153, 0;
|
|
@%p18 bra BB0_15;
|
|
bra.uni BB0_16;
|
|
|
|
BB0_15:
|
|
mov.u32 %r436, %r20;
|
|
mov.u32 %r437, %r29;
|
|
bra.uni BB0_17;
|
|
|
|
BB0_16:
|
|
not.b32 %r155, %r435;
|
|
neg.s32 %r437, %r29;
|
|
setp.eq.s32 %p19, %r29, 0;
|
|
selp.u32 %r156, 1, 0, %p19;
|
|
add.s32 %r435, %r156, %r155;
|
|
xor.b32 %r436, %r20, -2147483648;
|
|
|
|
BB0_17:
|
|
clz.b32 %r439, %r435;
|
|
setp.eq.s32 %p20, %r439, 0;
|
|
shl.b32 %r157, %r435, %r439;
|
|
mov.u32 %r158, 32;
|
|
sub.s32 %r159, %r158, %r439;
|
|
shr.u32 %r160, %r437, %r159;
|
|
add.s32 %r161, %r160, %r157;
|
|
selp.b32 %r38, %r435, %r161, %p20;
|
|
mov.u32 %r162, -921707870;
|
|
mul.hi.u32 %r438, %r38, %r162;
|
|
setp.eq.s32 %p21, %r20, 0;
|
|
neg.s32 %r163, %r30;
|
|
selp.b32 %r440, %r30, %r163, %p21;
|
|
setp.lt.s32 %p22, %r438, 1;
|
|
@%p22 bra BB0_19;
|
|
|
|
mul.lo.s32 %r164, %r38, -921707870;
|
|
shr.u32 %r165, %r164, 31;
|
|
shl.b32 %r166, %r438, 1;
|
|
add.s32 %r438, %r165, %r166;
|
|
add.s32 %r439, %r439, 1;
|
|
|
|
BB0_19:
|
|
mov.u32 %r167, 126;
|
|
sub.s32 %r168, %r167, %r439;
|
|
shl.b32 %r169, %r168, 23;
|
|
add.s32 %r170, %r438, 1;
|
|
shr.u32 %r171, %r170, 7;
|
|
add.s32 %r172, %r171, 1;
|
|
shr.u32 %r173, %r172, 1;
|
|
add.s32 %r174, %r173, %r169;
|
|
or.b32 %r175, %r174, %r436;
|
|
mov.b32 %f904, %r175;
|
|
|
|
BB0_20:
|
|
mul.rn.f32 %f36, %f904, %f904;
|
|
add.s32 %r46, %r440, 1;
|
|
and.b32 %r47, %r46, 1;
|
|
setp.eq.s32 %p23, %r47, 0;
|
|
@%p23 bra BB0_22;
|
|
bra.uni BB0_21;
|
|
|
|
BB0_22:
|
|
mov.f32 %f245, 0f3C08839E;
|
|
mov.f32 %f246, 0fB94CA1F9;
|
|
fma.rn.f32 %f905, %f246, %f36, %f245;
|
|
bra.uni BB0_23;
|
|
|
|
BB0_21:
|
|
mov.f32 %f243, 0fBAB6061A;
|
|
mov.f32 %f244, 0f37CCF5CE;
|
|
fma.rn.f32 %f905, %f244, %f36, %f243;
|
|
|
|
BB0_23:
|
|
@%p23 bra BB0_25;
|
|
bra.uni BB0_24;
|
|
|
|
BB0_25:
|
|
mov.f32 %f250, 0fBE2AAAA3;
|
|
fma.rn.f32 %f251, %f905, %f36, %f250;
|
|
mov.f32 %f252, 0f00000000;
|
|
fma.rn.f32 %f906, %f251, %f36, %f252;
|
|
bra.uni BB0_26;
|
|
|
|
BB0_24:
|
|
mov.f32 %f247, 0f3D2AAAA5;
|
|
fma.rn.f32 %f248, %f905, %f36, %f247;
|
|
mov.f32 %f249, 0fBF000000;
|
|
fma.rn.f32 %f906, %f248, %f36, %f249;
|
|
|
|
BB0_26:
|
|
fma.rn.f32 %f907, %f906, %f904, %f904;
|
|
@%p23 bra BB0_28;
|
|
|
|
mov.f32 %f253, 0f3F800000;
|
|
fma.rn.f32 %f907, %f906, %f36, %f253;
|
|
|
|
BB0_28:
|
|
and.b32 %r176, %r46, 2;
|
|
setp.eq.s32 %p26, %r176, 0;
|
|
@%p26 bra BB0_30;
|
|
|
|
mov.f32 %f254, 0f00000000;
|
|
mov.f32 %f255, 0fBF800000;
|
|
fma.rn.f32 %f907, %f907, %f255, %f254;
|
|
|
|
BB0_30:
|
|
@%p14 bra BB0_32;
|
|
|
|
mov.f32 %f256, 0f00000000;
|
|
mul.rn.f32 %f909, %f909, %f256;
|
|
|
|
BB0_32:
|
|
mul.f32 %f257, %f909, 0f3F22F983;
|
|
cvt.rni.s32.f32 %r450, %f257;
|
|
cvt.rn.f32.s32 %f258, %r450;
|
|
neg.f32 %f259, %f258;
|
|
fma.rn.f32 %f261, %f259, %f237, %f909;
|
|
fma.rn.f32 %f263, %f259, %f239, %f261;
|
|
fma.rn.f32 %f910, %f259, %f241, %f263;
|
|
abs.f32 %f265, %f909;
|
|
setp.leu.f32 %p28, %f265, 0f47CE4780;
|
|
@%p28 bra BB0_43;
|
|
|
|
mov.b32 %r49, %f909;
|
|
shr.u32 %r50, %r49, 23;
|
|
shl.b32 %r179, %r49, 8;
|
|
or.b32 %r51, %r179, -2147483648;
|
|
add.u64 %rd47, %SP, 4;
|
|
cvta.to.local.u64 %rd484, %rd47;
|
|
mov.u32 %r442, 0;
|
|
mov.u64 %rd483, __cudart_i2opi_f;
|
|
mov.u32 %r441, -6;
|
|
|
|
BB0_34:
|
|
.pragma "nounroll";
|
|
ld.const.u32 %r182, [%rd483];
|
|
// inline asm
|
|
{
|
|
mad.lo.cc.u32 %r180, %r182, %r51, %r442;
|
|
madc.hi.u32 %r442, %r182, %r51, 0;
|
|
}
|
|
// inline asm
|
|
st.local.u32 [%rd484], %r180;
|
|
add.s64 %rd484, %rd484, 4;
|
|
add.s64 %rd483, %rd483, 4;
|
|
add.s32 %r441, %r441, 1;
|
|
setp.ne.s32 %p29, %r441, 0;
|
|
@%p29 bra BB0_34;
|
|
|
|
and.b32 %r185, %r50, 255;
|
|
add.s32 %r186, %r185, -128;
|
|
shr.u32 %r187, %r186, 5;
|
|
and.b32 %r56, %r49, -2147483648;
|
|
cvta.to.local.u64 %rd49, %rd47;
|
|
st.local.u32 [%rd49+24], %r442;
|
|
mov.u32 %r188, 6;
|
|
sub.s32 %r189, %r188, %r187;
|
|
mul.wide.s32 %rd50, %r189, 4;
|
|
add.s64 %rd14, %rd49, %rd50;
|
|
ld.local.u32 %r443, [%rd14];
|
|
ld.local.u32 %r444, [%rd14+-4];
|
|
and.b32 %r59, %r50, 31;
|
|
setp.eq.s32 %p30, %r59, 0;
|
|
@%p30 bra BB0_37;
|
|
|
|
mov.u32 %r190, 32;
|
|
sub.s32 %r191, %r190, %r59;
|
|
shr.u32 %r192, %r444, %r191;
|
|
shl.b32 %r193, %r443, %r59;
|
|
add.s32 %r443, %r192, %r193;
|
|
ld.local.u32 %r194, [%rd14+-8];
|
|
shr.u32 %r195, %r194, %r191;
|
|
shl.b32 %r196, %r444, %r59;
|
|
add.s32 %r444, %r195, %r196;
|
|
|
|
BB0_37:
|
|
shr.u32 %r197, %r444, 30;
|
|
shl.b32 %r198, %r443, 2;
|
|
add.s32 %r445, %r197, %r198;
|
|
shl.b32 %r65, %r444, 2;
|
|
shr.u32 %r199, %r445, 31;
|
|
shr.u32 %r200, %r443, 30;
|
|
add.s32 %r66, %r199, %r200;
|
|
setp.eq.s32 %p31, %r199, 0;
|
|
@%p31 bra BB0_38;
|
|
bra.uni BB0_39;
|
|
|
|
BB0_38:
|
|
mov.u32 %r446, %r56;
|
|
mov.u32 %r447, %r65;
|
|
bra.uni BB0_40;
|
|
|
|
BB0_39:
|
|
not.b32 %r201, %r445;
|
|
neg.s32 %r447, %r65;
|
|
setp.eq.s32 %p32, %r65, 0;
|
|
selp.u32 %r202, 1, 0, %p32;
|
|
add.s32 %r445, %r202, %r201;
|
|
xor.b32 %r446, %r56, -2147483648;
|
|
|
|
BB0_40:
|
|
clz.b32 %r449, %r445;
|
|
setp.eq.s32 %p33, %r449, 0;
|
|
shl.b32 %r203, %r445, %r449;
|
|
mov.u32 %r204, 32;
|
|
sub.s32 %r205, %r204, %r449;
|
|
shr.u32 %r206, %r447, %r205;
|
|
add.s32 %r207, %r206, %r203;
|
|
selp.b32 %r74, %r445, %r207, %p33;
|
|
mov.u32 %r208, -921707870;
|
|
mul.hi.u32 %r448, %r74, %r208;
|
|
setp.eq.s32 %p34, %r56, 0;
|
|
neg.s32 %r209, %r66;
|
|
selp.b32 %r450, %r66, %r209, %p34;
|
|
setp.lt.s32 %p35, %r448, 1;
|
|
@%p35 bra BB0_42;
|
|
|
|
mul.lo.s32 %r210, %r74, -921707870;
|
|
shr.u32 %r211, %r210, 31;
|
|
shl.b32 %r212, %r448, 1;
|
|
add.s32 %r448, %r211, %r212;
|
|
add.s32 %r449, %r449, 1;
|
|
|
|
BB0_42:
|
|
mov.u32 %r213, 126;
|
|
sub.s32 %r214, %r213, %r449;
|
|
shl.b32 %r215, %r214, 23;
|
|
add.s32 %r216, %r448, 1;
|
|
shr.u32 %r217, %r216, 7;
|
|
add.s32 %r218, %r217, 1;
|
|
shr.u32 %r219, %r218, 1;
|
|
add.s32 %r220, %r219, %r215;
|
|
or.b32 %r221, %r220, %r446;
|
|
mov.b32 %f910, %r221;
|
|
|
|
BB0_43:
|
|
mul.rn.f32 %f53, %f910, %f910;
|
|
and.b32 %r82, %r450, 1;
|
|
setp.eq.s32 %p36, %r82, 0;
|
|
@%p36 bra BB0_45;
|
|
bra.uni BB0_44;
|
|
|
|
BB0_45:
|
|
mov.f32 %f268, 0f3C08839E;
|
|
mov.f32 %f269, 0fB94CA1F9;
|
|
fma.rn.f32 %f911, %f269, %f53, %f268;
|
|
bra.uni BB0_46;
|
|
|
|
BB0_44:
|
|
mov.f32 %f266, 0fBAB6061A;
|
|
mov.f32 %f267, 0f37CCF5CE;
|
|
fma.rn.f32 %f911, %f267, %f53, %f266;
|
|
|
|
BB0_46:
|
|
@%p36 bra BB0_48;
|
|
bra.uni BB0_47;
|
|
|
|
BB0_48:
|
|
mov.f32 %f273, 0fBE2AAAA3;
|
|
fma.rn.f32 %f274, %f911, %f53, %f273;
|
|
mov.f32 %f275, 0f00000000;
|
|
fma.rn.f32 %f912, %f274, %f53, %f275;
|
|
bra.uni BB0_49;
|
|
|
|
BB0_47:
|
|
mov.f32 %f270, 0f3D2AAAA5;
|
|
fma.rn.f32 %f271, %f911, %f53, %f270;
|
|
mov.f32 %f272, 0fBF000000;
|
|
fma.rn.f32 %f912, %f271, %f53, %f272;
|
|
|
|
BB0_49:
|
|
fma.rn.f32 %f913, %f912, %f910, %f910;
|
|
@%p36 bra BB0_51;
|
|
|
|
mov.f32 %f276, 0f3F800000;
|
|
fma.rn.f32 %f913, %f912, %f53, %f276;
|
|
|
|
BB0_51:
|
|
and.b32 %r222, %r450, 2;
|
|
setp.eq.s32 %p39, %r222, 0;
|
|
@%p39 bra BB0_53;
|
|
|
|
mov.f32 %f277, 0f00000000;
|
|
mov.f32 %f278, 0fBF800000;
|
|
fma.rn.f32 %f913, %f913, %f278, %f277;
|
|
|
|
BB0_53:
|
|
mul.f32 %f287, %f28, %f907;
|
|
add.u64 %rd51, %SP, 0;
|
|
cvta.to.local.u64 %rd52, %rd51;
|
|
mul.f32 %f288, %f287, %f287;
|
|
mov.f32 %f289, 0f3F800000;
|
|
sub.f32 %f290, %f289, %f288;
|
|
mul.f32 %f291, %f28, %f913;
|
|
mul.f32 %f292, %f291, %f291;
|
|
sub.f32 %f293, %f290, %f292;
|
|
mov.f32 %f294, 0f00000000;
|
|
max.f32 %f295, %f294, %f293;
|
|
sqrt.rn.f32 %f296, %f295;
|
|
mul.f32 %f297, %f14, %f291;
|
|
mul.f32 %f298, %f15, %f291;
|
|
mul.f32 %f299, %f16, %f291;
|
|
fma.rn.f32 %f300, %f23, %f287, %f297;
|
|
fma.rn.f32 %f301, %f22, %f287, %f298;
|
|
fma.rn.f32 %f302, %f21, %f287, %f299;
|
|
fma.rn.f32 %f303, %f7, %f296, %f300;
|
|
fma.rn.f32 %f304, %f8, %f296, %f301;
|
|
fma.rn.f32 %f305, %f9, %f296, %f302;
|
|
add.f32 %f306, %f7, %f303;
|
|
add.f32 %f307, %f8, %f304;
|
|
add.f32 %f308, %f9, %f305;
|
|
ld.global.f32 %f309, [shadowSpread];
|
|
mul.f32 %f310, %f309, %f306;
|
|
mul.f32 %f311, %f309, %f307;
|
|
mul.f32 %f312, %f309, %f308;
|
|
sub.f32 %f313, %f310, %f7;
|
|
sub.f32 %f314, %f311, %f8;
|
|
sub.f32 %f315, %f312, %f9;
|
|
mul.f32 %f316, %f314, %f314;
|
|
fma.rn.f32 %f317, %f313, %f313, %f316;
|
|
fma.rn.f32 %f318, %f315, %f315, %f317;
|
|
sqrt.rn.f32 %f319, %f318;
|
|
rcp.rn.f32 %f320, %f319;
|
|
mul.f32 %f282, %f320, %f313;
|
|
mul.f32 %f283, %f320, %f314;
|
|
mul.f32 %f284, %f320, %f315;
|
|
ld.global.u32 %r226, [imageEnabled];
|
|
and.b32 %r227, %r226, 32;
|
|
setp.eq.s32 %p40, %r227, 0;
|
|
selp.f32 %f321, 0f3F800000, 0f41200000, %p40;
|
|
mul.f32 %f285, %f321, %f26;
|
|
mov.u32 %r228, 1065353216;
|
|
st.local.u32 [%rd52], %r228;
|
|
ld.global.u32 %r223, [root];
|
|
mov.u32 %r224, 1;
|
|
mov.f32 %f286, 0f6C4ECB8F;
|
|
// inline asm
|
|
call _rt_trace_64, (%r223, %f11, %f12, %f13, %f282, %f283, %f284, %r224, %f285, %f286, %rd51, %r100);
|
|
// inline asm
|
|
ld.local.f32 %f322, [%rd52];
|
|
add.f32 %f915, %f915, %f322;
|
|
ld.global.u32 %r426, [samples];
|
|
add.s32 %r429, %r429, 1;
|
|
setp.lt.s32 %p41, %r429, %r426;
|
|
@%p41 bra BB0_7;
|
|
|
|
BB0_54:
|
|
add.s32 %r427, %r427, 1;
|
|
setp.lt.s32 %p42, %r427, %r426;
|
|
@%p42 bra BB0_5;
|
|
|
|
BB0_55:
|
|
setp.eq.s32 %p43, %r426, 0;
|
|
mov.f32 %f917, 0f3F800000;
|
|
@%p43 bra BB0_57;
|
|
|
|
mul.lo.s32 %r229, %r426, %r426;
|
|
cvt.rn.f32.s32 %f324, %r229;
|
|
div.rn.f32 %f917, %f915, %f324;
|
|
|
|
BB0_57:
|
|
ld.global.f32 %f325, [directColor];
|
|
mul.f32 %f70, %f917, %f325;
|
|
ld.global.f32 %f326, [directColor+4];
|
|
mul.f32 %f71, %f917, %f326;
|
|
ld.global.f32 %f327, [directColor+8];
|
|
mul.f32 %f72, %f917, %f327;
|
|
cvt.sat.f32.f32 %f328, %f10;
|
|
mul.f32 %f73, %f70, %f328;
|
|
mul.f32 %f74, %f71, %f328;
|
|
mul.f32 %f75, %f72, %f328;
|
|
ld.global.u32 %r455, [imageEnabled];
|
|
and.b32 %r230, %r455, 8;
|
|
setp.eq.s32 %p44, %r230, 0;
|
|
@%p44 bra BB0_70;
|
|
|
|
cvt.u64.u32 %rd55, %r2;
|
|
cvt.u64.u32 %rd56, %r3;
|
|
mov.u64 %rd59, image_Mask;
|
|
cvta.global.u64 %rd54, %rd59;
|
|
// inline asm
|
|
call (%rd53), _rt_buffer_get_64, (%rd54, %r99, %r99, %rd55, %rd56, %rd25, %rd25);
|
|
// inline asm
|
|
abs.f32 %f77, %f917;
|
|
setp.lt.f32 %p45, %f77, 0f00800000;
|
|
mul.f32 %f334, %f77, 0f4B800000;
|
|
selp.f32 %f335, 0fC3170000, 0fC2FE0000, %p45;
|
|
selp.f32 %f336, %f334, %f77, %p45;
|
|
mov.b32 %r233, %f336;
|
|
and.b32 %r234, %r233, 8388607;
|
|
or.b32 %r235, %r234, 1065353216;
|
|
mov.b32 %f337, %r235;
|
|
shr.u32 %r236, %r233, 23;
|
|
cvt.rn.f32.u32 %f338, %r236;
|
|
add.f32 %f339, %f335, %f338;
|
|
setp.gt.f32 %p46, %f337, 0f3FB504F3;
|
|
mul.f32 %f340, %f337, 0f3F000000;
|
|
add.f32 %f341, %f339, 0f3F800000;
|
|
selp.f32 %f342, %f340, %f337, %p46;
|
|
selp.f32 %f343, %f341, %f339, %p46;
|
|
add.f32 %f344, %f342, 0fBF800000;
|
|
add.f32 %f330, %f342, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f329,%f330;
|
|
// inline asm
|
|
add.f32 %f345, %f344, %f344;
|
|
mul.f32 %f346, %f329, %f345;
|
|
mul.f32 %f347, %f346, %f346;
|
|
mov.f32 %f348, 0f3C4CAF63;
|
|
mov.f32 %f349, 0f3B18F0FE;
|
|
fma.rn.f32 %f350, %f349, %f347, %f348;
|
|
mov.f32 %f351, 0f3DAAAABD;
|
|
fma.rn.f32 %f352, %f350, %f347, %f351;
|
|
mul.rn.f32 %f353, %f352, %f347;
|
|
mul.rn.f32 %f354, %f353, %f346;
|
|
sub.f32 %f355, %f344, %f346;
|
|
neg.f32 %f356, %f346;
|
|
add.f32 %f357, %f355, %f355;
|
|
fma.rn.f32 %f358, %f356, %f344, %f357;
|
|
mul.rn.f32 %f359, %f329, %f358;
|
|
add.f32 %f360, %f354, %f346;
|
|
sub.f32 %f361, %f346, %f360;
|
|
add.f32 %f362, %f354, %f361;
|
|
add.f32 %f363, %f359, %f362;
|
|
add.f32 %f364, %f360, %f363;
|
|
sub.f32 %f365, %f360, %f364;
|
|
add.f32 %f366, %f363, %f365;
|
|
mov.f32 %f367, 0f3F317200;
|
|
mul.rn.f32 %f368, %f343, %f367;
|
|
mov.f32 %f369, 0f35BFBE8E;
|
|
mul.rn.f32 %f370, %f343, %f369;
|
|
add.f32 %f371, %f368, %f364;
|
|
sub.f32 %f372, %f368, %f371;
|
|
add.f32 %f373, %f364, %f372;
|
|
add.f32 %f374, %f366, %f373;
|
|
add.f32 %f375, %f370, %f374;
|
|
add.f32 %f376, %f371, %f375;
|
|
sub.f32 %f377, %f371, %f376;
|
|
add.f32 %f378, %f375, %f377;
|
|
mov.f32 %f379, 0f3EE8BA2E;
|
|
mul.rn.f32 %f380, %f379, %f376;
|
|
neg.f32 %f381, %f380;
|
|
fma.rn.f32 %f382, %f379, %f376, %f381;
|
|
fma.rn.f32 %f383, %f379, %f378, %f382;
|
|
mov.f32 %f384, 0f00000000;
|
|
fma.rn.f32 %f385, %f384, %f376, %f383;
|
|
add.rn.f32 %f386, %f380, %f385;
|
|
neg.f32 %f387, %f386;
|
|
add.rn.f32 %f388, %f380, %f387;
|
|
add.rn.f32 %f389, %f388, %f385;
|
|
mov.b32 %r237, %f386;
|
|
setp.eq.s32 %p47, %r237, 1118925336;
|
|
add.s32 %r238, %r237, -1;
|
|
mov.b32 %f390, %r238;
|
|
add.f32 %f391, %f389, 0f37000000;
|
|
selp.f32 %f392, %f390, %f386, %p47;
|
|
selp.f32 %f78, %f391, %f389, %p47;
|
|
mul.f32 %f393, %f392, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f394, %f393;
|
|
mov.f32 %f395, 0fBF317200;
|
|
fma.rn.f32 %f396, %f394, %f395, %f392;
|
|
mov.f32 %f397, 0fB5BFBE8E;
|
|
fma.rn.f32 %f398, %f394, %f397, %f396;
|
|
mul.f32 %f399, %f398, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f400, %f399;
|
|
add.f32 %f401, %f394, 0f00000000;
|
|
ex2.approx.f32 %f402, %f401;
|
|
mul.f32 %f403, %f400, %f402;
|
|
setp.lt.f32 %p48, %f392, 0fC2D20000;
|
|
selp.f32 %f404, 0f00000000, %f403, %p48;
|
|
setp.gt.f32 %p49, %f392, 0f42D20000;
|
|
selp.f32 %f918, 0f7F800000, %f404, %p49;
|
|
setp.eq.f32 %p50, %f918, 0f7F800000;
|
|
@%p50 bra BB0_60;
|
|
|
|
fma.rn.f32 %f918, %f918, %f78, %f918;
|
|
|
|
BB0_60:
|
|
mov.f32 %f890, 0f3E68BA2E;
|
|
cvt.rzi.f32.f32 %f889, %f890;
|
|
fma.rn.f32 %f888, %f889, 0fC0000000, 0f3EE8BA2E;
|
|
abs.f32 %f887, %f888;
|
|
setp.lt.f32 %p51, %f917, 0f00000000;
|
|
setp.eq.f32 %p52, %f887, 0f3F800000;
|
|
and.pred %p1, %p51, %p52;
|
|
mov.b32 %r239, %f918;
|
|
xor.b32 %r240, %r239, -2147483648;
|
|
mov.b32 %f405, %r240;
|
|
selp.f32 %f920, %f405, %f918, %p1;
|
|
setp.eq.f32 %p53, %f917, 0f00000000;
|
|
@%p53 bra BB0_63;
|
|
bra.uni BB0_61;
|
|
|
|
BB0_63:
|
|
add.f32 %f408, %f917, %f917;
|
|
selp.f32 %f920, %f408, 0f00000000, %p52;
|
|
bra.uni BB0_64;
|
|
|
|
BB0_143:
|
|
mov.u64 %rd300, image_HDR;
|
|
cvta.global.u64 %rd295, %rd300;
|
|
mov.u32 %r361, 8;
|
|
// inline asm
|
|
call (%rd294), _rt_buffer_get_64, (%rd295, %r99, %r361, %rd18, %rd19, %rd25, %rd25);
|
|
// inline asm
|
|
mov.f32 %f792, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs179, %f792;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs180, 0;
|
|
st.v4.u16 [%rd294], {%rs179, %rs179, %rs179, %rs180};
|
|
|
|
BB0_144:
|
|
ld.global.u32 %r362, [additive];
|
|
setp.eq.s32 %p125, %r362, 0;
|
|
@%p125 bra BB0_146;
|
|
|
|
mov.u64 %rd313, image_RNM0;
|
|
cvta.global.u64 %rd302, %rd313;
|
|
mov.u32 %r366, 8;
|
|
// inline asm
|
|
call (%rd301), _rt_buffer_get_64, (%rd302, %r99, %r366, %rd18, %rd19, %rd25, %rd25);
|
|
// inline asm
|
|
ld.v4.u16 {%rs187, %rs188, %rs189, %rs190}, [%rd301];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f793, %rs187;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f794, %rs188;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f795, %rs189;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd307), _rt_buffer_get_64, (%rd302, %r99, %r366, %rd18, %rd19, %rd25, %rd25);
|
|
// inline asm
|
|
add.f32 %f796, %f793, 0f00000000;
|
|
add.f32 %f797, %f794, 0f00000000;
|
|
add.f32 %f798, %f795, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs186, %f798;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs185, %f797;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs184, %f796;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs191, 0;
|
|
st.v4.u16 [%rd307], {%rs184, %rs185, %rs186, %rs191};
|
|
bra.uni BB0_147;
|
|
|
|
BB0_146:
|
|
mov.u64 %rd320, image_RNM0;
|
|
cvta.global.u64 %rd315, %rd320;
|
|
mov.u32 %r368, 8;
|
|
// inline asm
|
|
call (%rd314), _rt_buffer_get_64, (%rd315, %r99, %r368, %rd18, %rd19, %rd25, %rd25);
|
|
// inline asm
|
|
mov.f32 %f799, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs192, %f799;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs193, 0;
|
|
st.v4.u16 [%rd314], {%rs192, %rs192, %rs192, %rs193};
|
|
|
|
BB0_147:
|
|
ld.global.u32 %r369, [additive];
|
|
setp.eq.s32 %p126, %r369, 0;
|
|
@%p126 bra BB0_149;
|
|
|
|
mov.u64 %rd333, image_RNM1;
|
|
cvta.global.u64 %rd322, %rd333;
|
|
mov.u32 %r373, 8;
|
|
// inline asm
|
|
call (%rd321), _rt_buffer_get_64, (%rd322, %r99, %r373, %rd18, %rd19, %rd25, %rd25);
|
|
// inline asm
|
|
ld.v4.u16 {%rs200, %rs201, %rs202, %rs203}, [%rd321];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f800, %rs200;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f801, %rs201;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f802, %rs202;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd327), _rt_buffer_get_64, (%rd322, %r99, %r373, %rd18, %rd19, %rd25, %rd25);
|
|
// inline asm
|
|
add.f32 %f803, %f800, 0f00000000;
|
|
add.f32 %f804, %f801, 0f00000000;
|
|
add.f32 %f805, %f802, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs199, %f805;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs198, %f804;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs197, %f803;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs204, 0;
|
|
st.v4.u16 [%rd327], {%rs197, %rs198, %rs199, %rs204};
|
|
bra.uni BB0_150;
|
|
|
|
BB0_149:
|
|
mov.u64 %rd340, image_RNM1;
|
|
cvta.global.u64 %rd335, %rd340;
|
|
mov.u32 %r375, 8;
|
|
// inline asm
|
|
call (%rd334), _rt_buffer_get_64, (%rd335, %r99, %r375, %rd18, %rd19, %rd25, %rd25);
|
|
// inline asm
|
|
mov.f32 %f806, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs205, %f806;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs206, 0;
|
|
st.v4.u16 [%rd334], {%rs205, %rs205, %rs205, %rs206};
|
|
|
|
BB0_150:
|
|
ld.global.u32 %r376, [additive];
|
|
setp.eq.s32 %p127, %r376, 0;
|
|
@%p127 bra BB0_152;
|
|
|
|
mov.u64 %rd353, image_RNM2;
|
|
cvta.global.u64 %rd342, %rd353;
|
|
mov.u32 %r380, 8;
|
|
// inline asm
|
|
call (%rd341), _rt_buffer_get_64, (%rd342, %r99, %r380, %rd18, %rd19, %rd25, %rd25);
|
|
// inline asm
|
|
ld.v4.u16 {%rs213, %rs214, %rs215, %rs216}, [%rd341];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f807, %rs213;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f808, %rs214;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f809, %rs215;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd347), _rt_buffer_get_64, (%rd342, %r99, %r380, %rd18, %rd19, %rd25, %rd25);
|
|
// inline asm
|
|
add.f32 %f810, %f807, 0f00000000;
|
|
add.f32 %f811, %f808, 0f00000000;
|
|
add.f32 %f812, %f809, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs212, %f812;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs211, %f811;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs210, %f810;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs217, 0;
|
|
st.v4.u16 [%rd347], {%rs210, %rs211, %rs212, %rs217};
|
|
bra.uni BB0_153;
|
|
|
|
BB0_152:
|
|
mov.u64 %rd360, image_RNM2;
|
|
cvta.global.u64 %rd355, %rd360;
|
|
mov.u32 %r382, 8;
|
|
// inline asm
|
|
call (%rd354), _rt_buffer_get_64, (%rd355, %r99, %r382, %rd18, %rd19, %rd25, %rd25);
|
|
// inline asm
|
|
mov.f32 %f813, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs218, %f813;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs219, 0;
|
|
st.v4.u16 [%rd354], {%rs218, %rs218, %rs218, %rs219};
|
|
|
|
BB0_153:
|
|
ld.global.u32 %r383, [additive];
|
|
setp.eq.s32 %p128, %r383, 0;
|
|
@%p128 bra BB0_155;
|
|
|
|
mov.u64 %rd373, image_RNM3;
|
|
cvta.global.u64 %rd362, %rd373;
|
|
mov.u32 %r387, 8;
|
|
// inline asm
|
|
call (%rd361), _rt_buffer_get_64, (%rd362, %r99, %r387, %rd18, %rd19, %rd25, %rd25);
|
|
// inline asm
|
|
ld.v4.u16 {%rs226, %rs227, %rs228, %rs229}, [%rd361];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f814, %rs226;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f815, %rs227;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f816, %rs228;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd367), _rt_buffer_get_64, (%rd362, %r99, %r387, %rd18, %rd19, %rd25, %rd25);
|
|
// inline asm
|
|
add.f32 %f817, %f814, 0f00000000;
|
|
add.f32 %f818, %f815, 0f00000000;
|
|
add.f32 %f819, %f816, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs225, %f819;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs224, %f818;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs223, %f817;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs230, 0;
|
|
st.v4.u16 [%rd367], {%rs223, %rs224, %rs225, %rs230};
|
|
bra.uni BB0_156;
|
|
|
|
BB0_155:
|
|
mov.u64 %rd380, image_RNM3;
|
|
cvta.global.u64 %rd375, %rd380;
|
|
mov.u32 %r389, 8;
|
|
// inline asm
|
|
call (%rd374), _rt_buffer_get_64, (%rd375, %r99, %r389, %rd18, %rd19, %rd25, %rd25);
|
|
// inline asm
|
|
mov.f32 %f820, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs231, %f820;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs232, 0;
|
|
st.v4.u16 [%rd374], {%rs231, %rs231, %rs231, %rs232};
|
|
|
|
BB0_156:
|
|
ld.global.u32 %r390, [additive];
|
|
setp.eq.s32 %p129, %r390, 0;
|
|
@%p129 bra BB0_158;
|
|
|
|
mov.u64 %rd393, image_RNM4;
|
|
cvta.global.u64 %rd382, %rd393;
|
|
mov.u32 %r394, 8;
|
|
// inline asm
|
|
call (%rd381), _rt_buffer_get_64, (%rd382, %r99, %r394, %rd18, %rd19, %rd25, %rd25);
|
|
// inline asm
|
|
ld.v4.u16 {%rs239, %rs240, %rs241, %rs242}, [%rd381];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f821, %rs239;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f822, %rs240;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f823, %rs241;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd387), _rt_buffer_get_64, (%rd382, %r99, %r394, %rd18, %rd19, %rd25, %rd25);
|
|
// inline asm
|
|
add.f32 %f824, %f821, 0f00000000;
|
|
add.f32 %f825, %f822, 0f00000000;
|
|
add.f32 %f826, %f823, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs238, %f826;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs237, %f825;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs236, %f824;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs243, 0;
|
|
st.v4.u16 [%rd387], {%rs236, %rs237, %rs238, %rs243};
|
|
bra.uni BB0_159;
|
|
|
|
BB0_158:
|
|
mov.u64 %rd400, image_RNM4;
|
|
cvta.global.u64 %rd395, %rd400;
|
|
mov.u32 %r396, 8;
|
|
// inline asm
|
|
call (%rd394), _rt_buffer_get_64, (%rd395, %r99, %r396, %rd18, %rd19, %rd25, %rd25);
|
|
// inline asm
|
|
mov.f32 %f827, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs244, %f827;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs245, 0;
|
|
st.v4.u16 [%rd394], {%rs244, %rs244, %rs244, %rs245};
|
|
|
|
BB0_159:
|
|
ld.global.u32 %r397, [additive];
|
|
setp.eq.s32 %p130, %r397, 0;
|
|
@%p130 bra BB0_161;
|
|
|
|
mov.u64 %rd413, image_RNM5;
|
|
cvta.global.u64 %rd402, %rd413;
|
|
mov.u32 %r401, 8;
|
|
// inline asm
|
|
call (%rd401), _rt_buffer_get_64, (%rd402, %r99, %r401, %rd18, %rd19, %rd25, %rd25);
|
|
// inline asm
|
|
ld.v4.u16 {%rs252, %rs253, %rs254, %rs255}, [%rd401];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f828, %rs252;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f829, %rs253;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f830, %rs254;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd407), _rt_buffer_get_64, (%rd402, %r99, %r401, %rd18, %rd19, %rd25, %rd25);
|
|
// inline asm
|
|
add.f32 %f831, %f828, 0f00000000;
|
|
add.f32 %f832, %f829, 0f00000000;
|
|
add.f32 %f833, %f830, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs251, %f833;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs250, %f832;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs249, %f831;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs256, 0;
|
|
st.v4.u16 [%rd407], {%rs249, %rs250, %rs251, %rs256};
|
|
bra.uni BB0_162;
|
|
|
|
BB0_161:
|
|
mov.u64 %rd420, image_RNM5;
|
|
cvta.global.u64 %rd415, %rd420;
|
|
mov.u32 %r403, 8;
|
|
// inline asm
|
|
call (%rd414), _rt_buffer_get_64, (%rd415, %r99, %r403, %rd18, %rd19, %rd25, %rd25);
|
|
// inline asm
|
|
mov.f32 %f834, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs257, %f834;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs258, 0;
|
|
st.v4.u16 [%rd414], {%rs257, %rs257, %rs257, %rs258};
|
|
|
|
BB0_162:
|
|
ld.global.u32 %r404, [additive];
|
|
setp.eq.s32 %p131, %r404, 0;
|
|
@%p131 bra BB0_164;
|
|
|
|
mov.u64 %rd433, image_RNM6;
|
|
cvta.global.u64 %rd422, %rd433;
|
|
mov.u32 %r408, 8;
|
|
// inline asm
|
|
call (%rd421), _rt_buffer_get_64, (%rd422, %r99, %r408, %rd18, %rd19, %rd25, %rd25);
|
|
// inline asm
|
|
ld.v4.u16 {%rs265, %rs266, %rs267, %rs268}, [%rd421];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f835, %rs265;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f836, %rs266;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f837, %rs267;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd427), _rt_buffer_get_64, (%rd422, %r99, %r408, %rd18, %rd19, %rd25, %rd25);
|
|
// inline asm
|
|
add.f32 %f838, %f835, 0f00000000;
|
|
add.f32 %f839, %f836, 0f00000000;
|
|
add.f32 %f840, %f837, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs264, %f840;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs263, %f839;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs262, %f838;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs269, 0;
|
|
st.v4.u16 [%rd427], {%rs262, %rs263, %rs264, %rs269};
|
|
bra.uni BB0_165;
|
|
|
|
BB0_164:
|
|
mov.u64 %rd440, image_RNM6;
|
|
cvta.global.u64 %rd435, %rd440;
|
|
mov.u32 %r410, 8;
|
|
// inline asm
|
|
call (%rd434), _rt_buffer_get_64, (%rd435, %r99, %r410, %rd18, %rd19, %rd25, %rd25);
|
|
// inline asm
|
|
mov.f32 %f841, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs270, %f841;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs271, 0;
|
|
st.v4.u16 [%rd434], {%rs270, %rs270, %rs270, %rs271};
|
|
|
|
BB0_165:
|
|
ld.global.u32 %r411, [additive];
|
|
setp.eq.s32 %p132, %r411, 0;
|
|
@%p132 bra BB0_167;
|
|
|
|
mov.u64 %rd453, image_RNM7;
|
|
cvta.global.u64 %rd442, %rd453;
|
|
mov.u32 %r415, 8;
|
|
// inline asm
|
|
call (%rd441), _rt_buffer_get_64, (%rd442, %r99, %r415, %rd18, %rd19, %rd25, %rd25);
|
|
// inline asm
|
|
ld.v4.u16 {%rs278, %rs279, %rs280, %rs281}, [%rd441];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f842, %rs278;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f843, %rs279;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f844, %rs280;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd447), _rt_buffer_get_64, (%rd442, %r99, %r415, %rd18, %rd19, %rd25, %rd25);
|
|
// inline asm
|
|
add.f32 %f845, %f842, 0f00000000;
|
|
add.f32 %f846, %f843, 0f00000000;
|
|
add.f32 %f847, %f844, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs277, %f847;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs276, %f846;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs275, %f845;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs282, 0;
|
|
st.v4.u16 [%rd447], {%rs275, %rs276, %rs277, %rs282};
|
|
bra.uni BB0_168;
|
|
|
|
BB0_167:
|
|
mov.u64 %rd460, image_RNM7;
|
|
cvta.global.u64 %rd455, %rd460;
|
|
mov.u32 %r417, 8;
|
|
// inline asm
|
|
call (%rd454), _rt_buffer_get_64, (%rd455, %r99, %r417, %rd18, %rd19, %rd25, %rd25);
|
|
// inline asm
|
|
mov.f32 %f848, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs283, %f848;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs284, 0;
|
|
st.v4.u16 [%rd454], {%rs283, %rs283, %rs283, %rs284};
|
|
|
|
BB0_168:
|
|
ld.global.u32 %r418, [additive];
|
|
setp.eq.s32 %p133, %r418, 0;
|
|
@%p133 bra BB0_170;
|
|
|
|
mov.u64 %rd473, image_RNM8;
|
|
cvta.global.u64 %rd462, %rd473;
|
|
mov.u32 %r422, 8;
|
|
// inline asm
|
|
call (%rd461), _rt_buffer_get_64, (%rd462, %r99, %r422, %rd18, %rd19, %rd25, %rd25);
|
|
// inline asm
|
|
ld.v4.u16 {%rs291, %rs292, %rs293, %rs294}, [%rd461];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f849, %rs291;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f850, %rs292;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f851, %rs293;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd467), _rt_buffer_get_64, (%rd462, %r99, %r422, %rd18, %rd19, %rd25, %rd25);
|
|
// inline asm
|
|
add.f32 %f852, %f849, 0f00000000;
|
|
add.f32 %f853, %f850, 0f00000000;
|
|
add.f32 %f854, %f851, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs290, %f854;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs289, %f853;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs288, %f852;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs295, 0;
|
|
st.v4.u16 [%rd467], {%rs288, %rs289, %rs290, %rs295};
|
|
bra.uni BB0_171;
|
|
|
|
BB0_170:
|
|
mov.u64 %rd480, image_RNM8;
|
|
cvta.global.u64 %rd475, %rd480;
|
|
mov.u32 %r424, 8;
|
|
// inline asm
|
|
call (%rd474), _rt_buffer_get_64, (%rd475, %r99, %r424, %rd18, %rd19, %rd25, %rd25);
|
|
// inline asm
|
|
mov.f32 %f855, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs296, %f855;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs297, 0;
|
|
st.v4.u16 [%rd474], {%rs296, %rs296, %rs296, %rs297};
|
|
bra.uni BB0_171;
|
|
|
|
BB0_61:
|
|
setp.geu.f32 %p54, %f917, 0f00000000;
|
|
@%p54 bra BB0_64;
|
|
|
|
mov.f32 %f894, 0f3EE8BA2E;
|
|
cvt.rzi.f32.f32 %f407, %f894;
|
|
setp.neu.f32 %p55, %f407, 0f3EE8BA2E;
|
|
selp.f32 %f920, 0f7FFFFFFF, %f920, %p55;
|
|
|
|
BB0_64:
|
|
abs.f32 %f891, %f917;
|
|
add.f32 %f409, %f891, 0f3EE8BA2E;
|
|
mov.b32 %r241, %f409;
|
|
setp.lt.s32 %p57, %r241, 2139095040;
|
|
@%p57 bra BB0_69;
|
|
|
|
abs.f32 %f892, %f917;
|
|
setp.gtu.f32 %p58, %f892, 0f7F800000;
|
|
@%p58 bra BB0_68;
|
|
bra.uni BB0_66;
|
|
|
|
BB0_68:
|
|
add.f32 %f920, %f917, 0f3EE8BA2E;
|
|
bra.uni BB0_69;
|
|
|
|
BB0_66:
|
|
abs.f32 %f893, %f917;
|
|
setp.neu.f32 %p59, %f893, 0f7F800000;
|
|
@%p59 bra BB0_69;
|
|
|
|
selp.f32 %f920, 0fFF800000, 0f7F800000, %p1;
|
|
|
|
BB0_69:
|
|
mul.f32 %f410, %f920, 0f437F0000;
|
|
setp.eq.f32 %p60, %f917, 0f3F800000;
|
|
selp.f32 %f411, 0f437F0000, %f410, %p60;
|
|
cvt.rzi.u32.f32 %r242, %f411;
|
|
cvt.u16.u32 %rs19, %r242;
|
|
mov.u16 %rs20, 255;
|
|
st.v2.u8 [%rd53], {%rs19, %rs20};
|
|
ld.global.u32 %r455, [imageEnabled];
|
|
|
|
BB0_70:
|
|
and.b32 %r243, %r455, 1;
|
|
setp.eq.b32 %p61, %r243, 1;
|
|
@!%p61 bra BB0_105;
|
|
bra.uni BB0_71;
|
|
|
|
BB0_71:
|
|
abs.f32 %f90, %f73;
|
|
setp.lt.f32 %p62, %f90, 0f00800000;
|
|
mul.f32 %f417, %f90, 0f4B800000;
|
|
selp.f32 %f418, 0fC3170000, 0fC2FE0000, %p62;
|
|
selp.f32 %f419, %f417, %f90, %p62;
|
|
mov.b32 %r244, %f419;
|
|
and.b32 %r245, %r244, 8388607;
|
|
or.b32 %r246, %r245, 1065353216;
|
|
mov.b32 %f420, %r246;
|
|
shr.u32 %r247, %r244, 23;
|
|
cvt.rn.f32.u32 %f421, %r247;
|
|
add.f32 %f422, %f418, %f421;
|
|
setp.gt.f32 %p63, %f420, 0f3FB504F3;
|
|
mul.f32 %f423, %f420, 0f3F000000;
|
|
add.f32 %f424, %f422, 0f3F800000;
|
|
selp.f32 %f425, %f423, %f420, %p63;
|
|
selp.f32 %f426, %f424, %f422, %p63;
|
|
add.f32 %f427, %f425, 0fBF800000;
|
|
add.f32 %f413, %f425, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f412,%f413;
|
|
// inline asm
|
|
add.f32 %f428, %f427, %f427;
|
|
mul.f32 %f429, %f412, %f428;
|
|
mul.f32 %f430, %f429, %f429;
|
|
mov.f32 %f431, 0f3C4CAF63;
|
|
mov.f32 %f432, 0f3B18F0FE;
|
|
fma.rn.f32 %f433, %f432, %f430, %f431;
|
|
mov.f32 %f434, 0f3DAAAABD;
|
|
fma.rn.f32 %f435, %f433, %f430, %f434;
|
|
mul.rn.f32 %f436, %f435, %f430;
|
|
mul.rn.f32 %f437, %f436, %f429;
|
|
sub.f32 %f438, %f427, %f429;
|
|
neg.f32 %f439, %f429;
|
|
add.f32 %f440, %f438, %f438;
|
|
fma.rn.f32 %f441, %f439, %f427, %f440;
|
|
mul.rn.f32 %f442, %f412, %f441;
|
|
add.f32 %f443, %f437, %f429;
|
|
sub.f32 %f444, %f429, %f443;
|
|
add.f32 %f445, %f437, %f444;
|
|
add.f32 %f446, %f442, %f445;
|
|
add.f32 %f447, %f443, %f446;
|
|
sub.f32 %f448, %f443, %f447;
|
|
add.f32 %f449, %f446, %f448;
|
|
mov.f32 %f450, 0f3F317200;
|
|
mul.rn.f32 %f451, %f426, %f450;
|
|
mov.f32 %f452, 0f35BFBE8E;
|
|
mul.rn.f32 %f453, %f426, %f452;
|
|
add.f32 %f454, %f451, %f447;
|
|
sub.f32 %f455, %f451, %f454;
|
|
add.f32 %f456, %f447, %f455;
|
|
add.f32 %f457, %f449, %f456;
|
|
add.f32 %f458, %f453, %f457;
|
|
add.f32 %f459, %f454, %f458;
|
|
sub.f32 %f460, %f454, %f459;
|
|
add.f32 %f461, %f458, %f460;
|
|
mov.f32 %f462, 0f3EE66666;
|
|
mul.rn.f32 %f463, %f462, %f459;
|
|
neg.f32 %f464, %f463;
|
|
fma.rn.f32 %f465, %f462, %f459, %f464;
|
|
fma.rn.f32 %f466, %f462, %f461, %f465;
|
|
mov.f32 %f467, 0f00000000;
|
|
fma.rn.f32 %f468, %f467, %f459, %f466;
|
|
add.rn.f32 %f469, %f463, %f468;
|
|
neg.f32 %f470, %f469;
|
|
add.rn.f32 %f471, %f463, %f470;
|
|
add.rn.f32 %f472, %f471, %f468;
|
|
mov.b32 %r248, %f469;
|
|
setp.eq.s32 %p64, %r248, 1118925336;
|
|
add.s32 %r249, %r248, -1;
|
|
mov.b32 %f473, %r249;
|
|
add.f32 %f474, %f472, 0f37000000;
|
|
selp.f32 %f475, %f473, %f469, %p64;
|
|
selp.f32 %f91, %f474, %f472, %p64;
|
|
mul.f32 %f476, %f475, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f477, %f476;
|
|
mov.f32 %f478, 0fBF317200;
|
|
fma.rn.f32 %f479, %f477, %f478, %f475;
|
|
mov.f32 %f480, 0fB5BFBE8E;
|
|
fma.rn.f32 %f481, %f477, %f480, %f479;
|
|
mul.f32 %f482, %f481, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f483, %f482;
|
|
add.f32 %f484, %f477, 0f00000000;
|
|
ex2.approx.f32 %f485, %f484;
|
|
mul.f32 %f486, %f483, %f485;
|
|
setp.lt.f32 %p65, %f475, 0fC2D20000;
|
|
selp.f32 %f487, 0f00000000, %f486, %p65;
|
|
setp.gt.f32 %p66, %f475, 0f42D20000;
|
|
selp.f32 %f921, 0f7F800000, %f487, %p66;
|
|
setp.eq.f32 %p67, %f921, 0f7F800000;
|
|
@%p67 bra BB0_73;
|
|
|
|
fma.rn.f32 %f921, %f921, %f91, %f921;
|
|
|
|
BB0_73:
|
|
mov.f32 %f859, 0f3E666666;
|
|
cvt.rzi.f32.f32 %f858, %f859;
|
|
fma.rn.f32 %f857, %f858, 0fC0000000, 0f3EE66666;
|
|
abs.f32 %f856, %f857;
|
|
setp.lt.f32 %p68, %f73, 0f00000000;
|
|
setp.eq.f32 %p69, %f856, 0f3F800000;
|
|
and.pred %p2, %p68, %p69;
|
|
mov.b32 %r250, %f921;
|
|
xor.b32 %r251, %r250, -2147483648;
|
|
mov.b32 %f488, %r251;
|
|
selp.f32 %f923, %f488, %f921, %p2;
|
|
setp.eq.f32 %p70, %f73, 0f00000000;
|
|
@%p70 bra BB0_76;
|
|
bra.uni BB0_74;
|
|
|
|
BB0_76:
|
|
add.f32 %f491, %f73, %f73;
|
|
selp.f32 %f923, %f491, 0f00000000, %p69;
|
|
bra.uni BB0_77;
|
|
|
|
BB0_74:
|
|
setp.geu.f32 %p71, %f73, 0f00000000;
|
|
@%p71 bra BB0_77;
|
|
|
|
mov.f32 %f883, 0f3EE66666;
|
|
cvt.rzi.f32.f32 %f490, %f883;
|
|
setp.neu.f32 %p72, %f490, 0f3EE66666;
|
|
selp.f32 %f923, 0f7FFFFFFF, %f923, %p72;
|
|
|
|
BB0_77:
|
|
abs.f32 %f860, %f73;
|
|
add.f32 %f492, %f860, 0f3EE66666;
|
|
mov.b32 %r252, %f492;
|
|
setp.lt.s32 %p74, %r252, 2139095040;
|
|
@%p74 bra BB0_82;
|
|
|
|
abs.f32 %f881, %f73;
|
|
setp.gtu.f32 %p75, %f881, 0f7F800000;
|
|
@%p75 bra BB0_81;
|
|
bra.uni BB0_79;
|
|
|
|
BB0_81:
|
|
add.f32 %f923, %f73, 0f3EE66666;
|
|
bra.uni BB0_82;
|
|
|
|
BB0_79:
|
|
abs.f32 %f882, %f73;
|
|
setp.neu.f32 %p76, %f882, 0f7F800000;
|
|
@%p76 bra BB0_82;
|
|
|
|
selp.f32 %f923, 0fFF800000, 0f7F800000, %p2;
|
|
|
|
BB0_82:
|
|
mov.f32 %f869, 0fB5BFBE8E;
|
|
mov.f32 %f868, 0fBF317200;
|
|
mov.f32 %f867, 0f00000000;
|
|
mov.f32 %f866, 0f35BFBE8E;
|
|
mov.f32 %f865, 0f3F317200;
|
|
mov.f32 %f864, 0f3DAAAABD;
|
|
mov.f32 %f863, 0f3C4CAF63;
|
|
mov.f32 %f862, 0f3B18F0FE;
|
|
mov.f32 %f861, 0f3EE66666;
|
|
setp.eq.f32 %p77, %f73, 0f3F800000;
|
|
selp.f32 %f102, 0f3F800000, %f923, %p77;
|
|
abs.f32 %f103, %f74;
|
|
setp.lt.f32 %p78, %f103, 0f00800000;
|
|
mul.f32 %f495, %f103, 0f4B800000;
|
|
selp.f32 %f496, 0fC3170000, 0fC2FE0000, %p78;
|
|
selp.f32 %f497, %f495, %f103, %p78;
|
|
mov.b32 %r253, %f497;
|
|
and.b32 %r254, %r253, 8388607;
|
|
or.b32 %r255, %r254, 1065353216;
|
|
mov.b32 %f498, %r255;
|
|
shr.u32 %r256, %r253, 23;
|
|
cvt.rn.f32.u32 %f499, %r256;
|
|
add.f32 %f500, %f496, %f499;
|
|
setp.gt.f32 %p79, %f498, 0f3FB504F3;
|
|
mul.f32 %f501, %f498, 0f3F000000;
|
|
add.f32 %f502, %f500, 0f3F800000;
|
|
selp.f32 %f503, %f501, %f498, %p79;
|
|
selp.f32 %f504, %f502, %f500, %p79;
|
|
add.f32 %f505, %f503, 0fBF800000;
|
|
add.f32 %f494, %f503, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f493,%f494;
|
|
// inline asm
|
|
add.f32 %f506, %f505, %f505;
|
|
mul.f32 %f507, %f493, %f506;
|
|
mul.f32 %f508, %f507, %f507;
|
|
fma.rn.f32 %f511, %f862, %f508, %f863;
|
|
fma.rn.f32 %f513, %f511, %f508, %f864;
|
|
mul.rn.f32 %f514, %f513, %f508;
|
|
mul.rn.f32 %f515, %f514, %f507;
|
|
sub.f32 %f516, %f505, %f507;
|
|
neg.f32 %f517, %f507;
|
|
add.f32 %f518, %f516, %f516;
|
|
fma.rn.f32 %f519, %f517, %f505, %f518;
|
|
mul.rn.f32 %f520, %f493, %f519;
|
|
add.f32 %f521, %f515, %f507;
|
|
sub.f32 %f522, %f507, %f521;
|
|
add.f32 %f523, %f515, %f522;
|
|
add.f32 %f524, %f520, %f523;
|
|
add.f32 %f525, %f521, %f524;
|
|
sub.f32 %f526, %f521, %f525;
|
|
add.f32 %f527, %f524, %f526;
|
|
mul.rn.f32 %f529, %f504, %f865;
|
|
mul.rn.f32 %f531, %f504, %f866;
|
|
add.f32 %f532, %f529, %f525;
|
|
sub.f32 %f533, %f529, %f532;
|
|
add.f32 %f534, %f525, %f533;
|
|
add.f32 %f535, %f527, %f534;
|
|
add.f32 %f536, %f531, %f535;
|
|
add.f32 %f537, %f532, %f536;
|
|
sub.f32 %f538, %f532, %f537;
|
|
add.f32 %f539, %f536, %f538;
|
|
mul.rn.f32 %f541, %f861, %f537;
|
|
neg.f32 %f542, %f541;
|
|
fma.rn.f32 %f543, %f861, %f537, %f542;
|
|
fma.rn.f32 %f544, %f861, %f539, %f543;
|
|
fma.rn.f32 %f546, %f867, %f537, %f544;
|
|
add.rn.f32 %f547, %f541, %f546;
|
|
neg.f32 %f548, %f547;
|
|
add.rn.f32 %f549, %f541, %f548;
|
|
add.rn.f32 %f550, %f549, %f546;
|
|
mov.b32 %r257, %f547;
|
|
setp.eq.s32 %p80, %r257, 1118925336;
|
|
add.s32 %r258, %r257, -1;
|
|
mov.b32 %f551, %r258;
|
|
add.f32 %f552, %f550, 0f37000000;
|
|
selp.f32 %f553, %f551, %f547, %p80;
|
|
selp.f32 %f104, %f552, %f550, %p80;
|
|
mul.f32 %f554, %f553, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f555, %f554;
|
|
fma.rn.f32 %f557, %f555, %f868, %f553;
|
|
fma.rn.f32 %f559, %f555, %f869, %f557;
|
|
mul.f32 %f560, %f559, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f561, %f560;
|
|
add.f32 %f562, %f555, 0f00000000;
|
|
ex2.approx.f32 %f563, %f562;
|
|
mul.f32 %f564, %f561, %f563;
|
|
setp.lt.f32 %p81, %f553, 0fC2D20000;
|
|
selp.f32 %f565, 0f00000000, %f564, %p81;
|
|
setp.gt.f32 %p82, %f553, 0f42D20000;
|
|
selp.f32 %f924, 0f7F800000, %f565, %p82;
|
|
setp.eq.f32 %p83, %f924, 0f7F800000;
|
|
@%p83 bra BB0_84;
|
|
|
|
fma.rn.f32 %f924, %f924, %f104, %f924;
|
|
|
|
BB0_84:
|
|
setp.lt.f32 %p84, %f74, 0f00000000;
|
|
and.pred %p3, %p84, %p69;
|
|
mov.b32 %r259, %f924;
|
|
xor.b32 %r260, %r259, -2147483648;
|
|
mov.b32 %f566, %r260;
|
|
selp.f32 %f926, %f566, %f924, %p3;
|
|
setp.eq.f32 %p86, %f74, 0f00000000;
|
|
@%p86 bra BB0_87;
|
|
bra.uni BB0_85;
|
|
|
|
BB0_87:
|
|
add.f32 %f569, %f74, %f74;
|
|
selp.f32 %f926, %f569, 0f00000000, %p69;
|
|
bra.uni BB0_88;
|
|
|
|
BB0_85:
|
|
setp.geu.f32 %p87, %f74, 0f00000000;
|
|
@%p87 bra BB0_88;
|
|
|
|
mov.f32 %f880, 0f3EE66666;
|
|
cvt.rzi.f32.f32 %f568, %f880;
|
|
setp.neu.f32 %p88, %f568, 0f3EE66666;
|
|
selp.f32 %f926, 0f7FFFFFFF, %f926, %p88;
|
|
|
|
BB0_88:
|
|
abs.f32 %f884, %f74;
|
|
add.f32 %f570, %f884, 0f3EE66666;
|
|
mov.b32 %r261, %f570;
|
|
setp.lt.s32 %p90, %r261, 2139095040;
|
|
@%p90 bra BB0_93;
|
|
|
|
abs.f32 %f885, %f74;
|
|
setp.gtu.f32 %p91, %f885, 0f7F800000;
|
|
@%p91 bra BB0_92;
|
|
bra.uni BB0_90;
|
|
|
|
BB0_92:
|
|
add.f32 %f926, %f74, 0f3EE66666;
|
|
bra.uni BB0_93;
|
|
|
|
BB0_90:
|
|
abs.f32 %f886, %f74;
|
|
setp.neu.f32 %p92, %f886, 0f7F800000;
|
|
@%p92 bra BB0_93;
|
|
|
|
selp.f32 %f926, 0fFF800000, 0f7F800000, %p3;
|
|
|
|
BB0_93:
|
|
mov.f32 %f878, 0fB5BFBE8E;
|
|
mov.f32 %f877, 0fBF317200;
|
|
mov.f32 %f876, 0f00000000;
|
|
mov.f32 %f875, 0f35BFBE8E;
|
|
mov.f32 %f874, 0f3F317200;
|
|
mov.f32 %f873, 0f3DAAAABD;
|
|
mov.f32 %f872, 0f3C4CAF63;
|
|
mov.f32 %f871, 0f3B18F0FE;
|
|
mov.f32 %f870, 0f3EE66666;
|
|
setp.eq.f32 %p93, %f74, 0f3F800000;
|
|
selp.f32 %f115, 0f3F800000, %f926, %p93;
|
|
abs.f32 %f116, %f75;
|
|
setp.lt.f32 %p94, %f116, 0f00800000;
|
|
mul.f32 %f573, %f116, 0f4B800000;
|
|
selp.f32 %f574, 0fC3170000, 0fC2FE0000, %p94;
|
|
selp.f32 %f575, %f573, %f116, %p94;
|
|
mov.b32 %r262, %f575;
|
|
and.b32 %r263, %r262, 8388607;
|
|
or.b32 %r264, %r263, 1065353216;
|
|
mov.b32 %f576, %r264;
|
|
shr.u32 %r265, %r262, 23;
|
|
cvt.rn.f32.u32 %f577, %r265;
|
|
add.f32 %f578, %f574, %f577;
|
|
setp.gt.f32 %p95, %f576, 0f3FB504F3;
|
|
mul.f32 %f579, %f576, 0f3F000000;
|
|
add.f32 %f580, %f578, 0f3F800000;
|
|
selp.f32 %f581, %f579, %f576, %p95;
|
|
selp.f32 %f582, %f580, %f578, %p95;
|
|
add.f32 %f583, %f581, 0fBF800000;
|
|
add.f32 %f572, %f581, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f571,%f572;
|
|
// inline asm
|
|
add.f32 %f584, %f583, %f583;
|
|
mul.f32 %f585, %f571, %f584;
|
|
mul.f32 %f586, %f585, %f585;
|
|
fma.rn.f32 %f589, %f871, %f586, %f872;
|
|
fma.rn.f32 %f591, %f589, %f586, %f873;
|
|
mul.rn.f32 %f592, %f591, %f586;
|
|
mul.rn.f32 %f593, %f592, %f585;
|
|
sub.f32 %f594, %f583, %f585;
|
|
neg.f32 %f595, %f585;
|
|
add.f32 %f596, %f594, %f594;
|
|
fma.rn.f32 %f597, %f595, %f583, %f596;
|
|
mul.rn.f32 %f598, %f571, %f597;
|
|
add.f32 %f599, %f593, %f585;
|
|
sub.f32 %f600, %f585, %f599;
|
|
add.f32 %f601, %f593, %f600;
|
|
add.f32 %f602, %f598, %f601;
|
|
add.f32 %f603, %f599, %f602;
|
|
sub.f32 %f604, %f599, %f603;
|
|
add.f32 %f605, %f602, %f604;
|
|
mul.rn.f32 %f607, %f582, %f874;
|
|
mul.rn.f32 %f609, %f582, %f875;
|
|
add.f32 %f610, %f607, %f603;
|
|
sub.f32 %f611, %f607, %f610;
|
|
add.f32 %f612, %f603, %f611;
|
|
add.f32 %f613, %f605, %f612;
|
|
add.f32 %f614, %f609, %f613;
|
|
add.f32 %f615, %f610, %f614;
|
|
sub.f32 %f616, %f610, %f615;
|
|
add.f32 %f617, %f614, %f616;
|
|
mul.rn.f32 %f619, %f870, %f615;
|
|
neg.f32 %f620, %f619;
|
|
fma.rn.f32 %f621, %f870, %f615, %f620;
|
|
fma.rn.f32 %f622, %f870, %f617, %f621;
|
|
fma.rn.f32 %f624, %f876, %f615, %f622;
|
|
add.rn.f32 %f625, %f619, %f624;
|
|
neg.f32 %f626, %f625;
|
|
add.rn.f32 %f627, %f619, %f626;
|
|
add.rn.f32 %f628, %f627, %f624;
|
|
mov.b32 %r266, %f625;
|
|
setp.eq.s32 %p96, %r266, 1118925336;
|
|
add.s32 %r267, %r266, -1;
|
|
mov.b32 %f629, %r267;
|
|
add.f32 %f630, %f628, 0f37000000;
|
|
selp.f32 %f631, %f629, %f625, %p96;
|
|
selp.f32 %f117, %f630, %f628, %p96;
|
|
mul.f32 %f632, %f631, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f633, %f632;
|
|
fma.rn.f32 %f635, %f633, %f877, %f631;
|
|
fma.rn.f32 %f637, %f633, %f878, %f635;
|
|
mul.f32 %f638, %f637, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f639, %f638;
|
|
add.f32 %f640, %f633, 0f00000000;
|
|
ex2.approx.f32 %f641, %f640;
|
|
mul.f32 %f642, %f639, %f641;
|
|
setp.lt.f32 %p97, %f631, 0fC2D20000;
|
|
selp.f32 %f643, 0f00000000, %f642, %p97;
|
|
setp.gt.f32 %p98, %f631, 0f42D20000;
|
|
selp.f32 %f927, 0f7F800000, %f643, %p98;
|
|
setp.eq.f32 %p99, %f927, 0f7F800000;
|
|
@%p99 bra BB0_95;
|
|
|
|
fma.rn.f32 %f927, %f927, %f117, %f927;
|
|
|
|
BB0_95:
|
|
setp.lt.f32 %p100, %f75, 0f00000000;
|
|
and.pred %p4, %p100, %p69;
|
|
mov.b32 %r268, %f927;
|
|
xor.b32 %r269, %r268, -2147483648;
|
|
mov.b32 %f644, %r269;
|
|
selp.f32 %f929, %f644, %f927, %p4;
|
|
setp.eq.f32 %p102, %f75, 0f00000000;
|
|
@%p102 bra BB0_98;
|
|
bra.uni BB0_96;
|
|
|
|
BB0_98:
|
|
add.f32 %f647, %f75, %f75;
|
|
selp.f32 %f929, %f647, 0f00000000, %p69;
|
|
bra.uni BB0_99;
|
|
|
|
BB0_96:
|
|
setp.geu.f32 %p103, %f75, 0f00000000;
|
|
@%p103 bra BB0_99;
|
|
|
|
mov.f32 %f879, 0f3EE66666;
|
|
cvt.rzi.f32.f32 %f646, %f879;
|
|
setp.neu.f32 %p104, %f646, 0f3EE66666;
|
|
selp.f32 %f929, 0f7FFFFFFF, %f929, %p104;
|
|
|
|
BB0_99:
|
|
abs.f32 %f895, %f75;
|
|
add.f32 %f648, %f895, 0f3EE66666;
|
|
mov.b32 %r270, %f648;
|
|
setp.lt.s32 %p106, %r270, 2139095040;
|
|
@%p106 bra BB0_104;
|
|
|
|
abs.f32 %f896, %f75;
|
|
setp.gtu.f32 %p107, %f896, 0f7F800000;
|
|
@%p107 bra BB0_103;
|
|
bra.uni BB0_101;
|
|
|
|
BB0_103:
|
|
add.f32 %f929, %f75, 0f3EE66666;
|
|
bra.uni BB0_104;
|
|
|
|
BB0_101:
|
|
abs.f32 %f897, %f75;
|
|
setp.neu.f32 %p108, %f897, 0f7F800000;
|
|
@%p108 bra BB0_104;
|
|
|
|
selp.f32 %f929, 0fFF800000, 0f7F800000, %p4;
|
|
|
|
BB0_104:
|
|
mov.u32 %r425, 4;
|
|
setp.eq.f32 %p109, %f75, 0f3F800000;
|
|
selp.f32 %f649, 0f3F800000, %f929, %p109;
|
|
cvt.u64.u32 %rd63, %r3;
|
|
cvt.u64.u32 %rd62, %r2;
|
|
mov.u64 %rd66, image;
|
|
cvta.global.u64 %rd61, %rd66;
|
|
// inline asm
|
|
call (%rd60), _rt_buffer_get_64, (%rd61, %r99, %r425, %rd62, %rd63, %rd25, %rd25);
|
|
// inline asm
|
|
cvt.sat.f32.f32 %f650, %f649;
|
|
mul.f32 %f651, %f650, 0f437FFD71;
|
|
cvt.rzi.u32.f32 %r273, %f651;
|
|
cvt.sat.f32.f32 %f652, %f115;
|
|
mul.f32 %f653, %f652, 0f437FFD71;
|
|
cvt.rzi.u32.f32 %r274, %f653;
|
|
cvt.sat.f32.f32 %f654, %f102;
|
|
mul.f32 %f655, %f654, 0f437FFD71;
|
|
cvt.rzi.u32.f32 %r275, %f655;
|
|
cvt.u16.u32 %rs21, %r273;
|
|
cvt.u16.u32 %rs22, %r275;
|
|
cvt.u16.u32 %rs23, %r274;
|
|
mov.u16 %rs24, 255;
|
|
st.v4.u8 [%rd60], {%rs21, %rs23, %rs22, %rs24};
|
|
ld.global.u32 %r455, [imageEnabled];
|
|
|
|
BB0_105:
|
|
cvt.u64.u32 %rd16, %r2;
|
|
cvt.u64.u32 %rd17, %r3;
|
|
and.b32 %r276, %r455, 4;
|
|
setp.eq.s32 %p110, %r276, 0;
|
|
@%p110 bra BB0_109;
|
|
|
|
ld.global.u32 %r277, [additive];
|
|
setp.eq.s32 %p111, %r277, 0;
|
|
mov.f32 %f656, 0f3F800000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs25, %f656;}
|
|
|
|
// inline asm
|
|
@%p111 bra BB0_108;
|
|
|
|
mov.u64 %rd79, image_HDR;
|
|
cvta.global.u64 %rd68, %rd79;
|
|
mov.u32 %r281, 8;
|
|
// inline asm
|
|
call (%rd67), _rt_buffer_get_64, (%rd68, %r99, %r281, %rd16, %rd17, %rd25, %rd25);
|
|
// inline asm
|
|
ld.v4.u16 {%rs32, %rs33, %rs34, %rs35}, [%rd67];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f657, %rs32;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f658, %rs33;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f659, %rs34;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd73), _rt_buffer_get_64, (%rd68, %r99, %r281, %rd16, %rd17, %rd25, %rd25);
|
|
// inline asm
|
|
add.f32 %f660, %f73, %f657;
|
|
add.f32 %f661, %f74, %f658;
|
|
add.f32 %f662, %f75, %f659;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs31, %f662;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs30, %f661;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs29, %f660;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd73], {%rs29, %rs30, %rs31, %rs25};
|
|
bra.uni BB0_109;
|
|
|
|
BB0_108:
|
|
mov.u64 %rd86, image_HDR;
|
|
cvta.global.u64 %rd81, %rd86;
|
|
mov.u32 %r283, 8;
|
|
// inline asm
|
|
call (%rd80), _rt_buffer_get_64, (%rd81, %r99, %r283, %rd16, %rd17, %rd25, %rd25);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs38, %f75;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs37, %f74;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs36, %f73;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd80], {%rs36, %rs37, %rs38, %rs25};
|
|
|
|
BB0_109:
|
|
mul.f32 %f128, %f70, 0f403D3C19;
|
|
mul.f32 %f129, %f128, 0f3E906EBB;
|
|
mul.f32 %f130, %f71, 0f403D3C19;
|
|
mul.f32 %f131, %f130, 0f3E906EBB;
|
|
mul.f32 %f132, %f72, 0f403D3C19;
|
|
mul.f32 %f133, %f132, 0f3E906EBB;
|
|
ld.global.u32 %r284, [additive];
|
|
setp.eq.s32 %p112, %r284, 0;
|
|
mov.f32 %f666, 0f3F800000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs39, %f666;}
|
|
|
|
// inline asm
|
|
@%p112 bra BB0_111;
|
|
|
|
mov.u64 %rd99, image_RNM0;
|
|
cvta.global.u64 %rd88, %rd99;
|
|
mov.u32 %r288, 8;
|
|
// inline asm
|
|
call (%rd87), _rt_buffer_get_64, (%rd88, %r99, %r288, %rd16, %rd17, %rd25, %rd25);
|
|
// inline asm
|
|
ld.v4.u16 {%rs46, %rs47, %rs48, %rs49}, [%rd87];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f667, %rs46;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f668, %rs47;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f669, %rs48;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd93), _rt_buffer_get_64, (%rd88, %r99, %r288, %rd16, %rd17, %rd25, %rd25);
|
|
// inline asm
|
|
add.f32 %f670, %f129, %f667;
|
|
add.f32 %f671, %f131, %f668;
|
|
add.f32 %f672, %f133, %f669;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs45, %f672;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs44, %f671;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs43, %f670;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd93], {%rs43, %rs44, %rs45, %rs39};
|
|
bra.uni BB0_112;
|
|
|
|
BB0_111:
|
|
mov.u64 %rd106, image_RNM0;
|
|
cvta.global.u64 %rd101, %rd106;
|
|
mov.u32 %r290, 8;
|
|
// inline asm
|
|
call (%rd100), _rt_buffer_get_64, (%rd101, %r99, %r290, %rd16, %rd17, %rd25, %rd25);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs52, %f133;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs51, %f131;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs50, %f129;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd100], {%rs50, %rs51, %rs52, %rs39};
|
|
|
|
BB0_112:
|
|
ld.global.f32 %f677, [directDir+4];
|
|
mul.f32 %f678, %f677, 0fBEFA2A1C;
|
|
mul.f32 %f134, %f128, %f678;
|
|
mul.f32 %f135, %f130, %f678;
|
|
mul.f32 %f136, %f132, %f678;
|
|
ld.global.u32 %r291, [additive];
|
|
setp.eq.s32 %p113, %r291, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs53, %f666;}
|
|
|
|
// inline asm
|
|
@%p113 bra BB0_114;
|
|
|
|
mov.u64 %rd119, image_RNM1;
|
|
cvta.global.u64 %rd108, %rd119;
|
|
mov.u32 %r295, 8;
|
|
// inline asm
|
|
call (%rd107), _rt_buffer_get_64, (%rd108, %r99, %r295, %rd16, %rd17, %rd25, %rd25);
|
|
// inline asm
|
|
ld.v4.u16 {%rs60, %rs61, %rs62, %rs63}, [%rd107];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f679, %rs60;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f680, %rs61;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f681, %rs62;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd113), _rt_buffer_get_64, (%rd108, %r99, %r295, %rd16, %rd17, %rd25, %rd25);
|
|
// inline asm
|
|
add.f32 %f682, %f134, %f679;
|
|
add.f32 %f683, %f135, %f680;
|
|
add.f32 %f684, %f136, %f681;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs59, %f684;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs58, %f683;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs57, %f682;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd113], {%rs57, %rs58, %rs59, %rs53};
|
|
bra.uni BB0_115;
|
|
|
|
BB0_114:
|
|
mov.u64 %rd126, image_RNM1;
|
|
cvta.global.u64 %rd121, %rd126;
|
|
mov.u32 %r297, 8;
|
|
// inline asm
|
|
call (%rd120), _rt_buffer_get_64, (%rd121, %r99, %r297, %rd16, %rd17, %rd25, %rd25);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs66, %f136;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs65, %f135;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs64, %f134;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd120], {%rs64, %rs65, %rs66, %rs53};
|
|
|
|
BB0_115:
|
|
ld.global.f32 %f689, [directDir+8];
|
|
mul.f32 %f690, %f689, 0fBEFA2A1C;
|
|
mul.f32 %f137, %f128, %f690;
|
|
mul.f32 %f138, %f130, %f690;
|
|
mul.f32 %f139, %f132, %f690;
|
|
ld.global.u32 %r298, [additive];
|
|
setp.eq.s32 %p114, %r298, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs67, %f666;}
|
|
|
|
// inline asm
|
|
@%p114 bra BB0_117;
|
|
|
|
mov.u64 %rd139, image_RNM2;
|
|
cvta.global.u64 %rd128, %rd139;
|
|
mov.u32 %r302, 8;
|
|
// inline asm
|
|
call (%rd127), _rt_buffer_get_64, (%rd128, %r99, %r302, %rd16, %rd17, %rd25, %rd25);
|
|
// inline asm
|
|
ld.v4.u16 {%rs74, %rs75, %rs76, %rs77}, [%rd127];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f691, %rs74;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f692, %rs75;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f693, %rs76;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd133), _rt_buffer_get_64, (%rd128, %r99, %r302, %rd16, %rd17, %rd25, %rd25);
|
|
// inline asm
|
|
add.f32 %f694, %f137, %f691;
|
|
add.f32 %f695, %f138, %f692;
|
|
add.f32 %f696, %f139, %f693;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs73, %f696;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs72, %f695;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs71, %f694;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd133], {%rs71, %rs72, %rs73, %rs67};
|
|
bra.uni BB0_118;
|
|
|
|
BB0_117:
|
|
mov.u64 %rd146, image_RNM2;
|
|
cvta.global.u64 %rd141, %rd146;
|
|
mov.u32 %r304, 8;
|
|
// inline asm
|
|
call (%rd140), _rt_buffer_get_64, (%rd141, %r99, %r304, %rd16, %rd17, %rd25, %rd25);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs80, %f139;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs79, %f138;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs78, %f137;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd140], {%rs78, %rs79, %rs80, %rs67};
|
|
|
|
BB0_118:
|
|
ld.global.f32 %f701, [directDir];
|
|
mul.f32 %f702, %f701, 0fBEFA2A1C;
|
|
mul.f32 %f140, %f128, %f702;
|
|
mul.f32 %f141, %f130, %f702;
|
|
mul.f32 %f142, %f132, %f702;
|
|
ld.global.u32 %r305, [additive];
|
|
setp.eq.s32 %p115, %r305, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs81, %f666;}
|
|
|
|
// inline asm
|
|
@%p115 bra BB0_120;
|
|
|
|
mov.u64 %rd159, image_RNM3;
|
|
cvta.global.u64 %rd148, %rd159;
|
|
mov.u32 %r309, 8;
|
|
// inline asm
|
|
call (%rd147), _rt_buffer_get_64, (%rd148, %r99, %r309, %rd16, %rd17, %rd25, %rd25);
|
|
// inline asm
|
|
ld.v4.u16 {%rs88, %rs89, %rs90, %rs91}, [%rd147];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f703, %rs88;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f704, %rs89;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f705, %rs90;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd153), _rt_buffer_get_64, (%rd148, %r99, %r309, %rd16, %rd17, %rd25, %rd25);
|
|
// inline asm
|
|
add.f32 %f706, %f140, %f703;
|
|
add.f32 %f707, %f141, %f704;
|
|
add.f32 %f708, %f142, %f705;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs87, %f708;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs86, %f707;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs85, %f706;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd153], {%rs85, %rs86, %rs87, %rs81};
|
|
bra.uni BB0_121;
|
|
|
|
BB0_120:
|
|
mov.u64 %rd166, image_RNM3;
|
|
cvta.global.u64 %rd161, %rd166;
|
|
mov.u32 %r311, 8;
|
|
// inline asm
|
|
call (%rd160), _rt_buffer_get_64, (%rd161, %r99, %r311, %rd16, %rd17, %rd25, %rd25);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs94, %f142;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs93, %f141;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs92, %f140;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd160], {%rs92, %rs93, %rs94, %rs81};
|
|
|
|
BB0_121:
|
|
ld.global.f32 %f713, [directDir+4];
|
|
ld.global.f32 %f714, [directDir];
|
|
mul.f32 %f715, %f714, %f713;
|
|
mul.f32 %f716, %f715, 0f3F8BD8A1;
|
|
mul.f32 %f143, %f128, %f716;
|
|
mul.f32 %f144, %f130, %f716;
|
|
mul.f32 %f145, %f132, %f716;
|
|
ld.global.u32 %r312, [additive];
|
|
setp.eq.s32 %p116, %r312, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs95, %f666;}
|
|
|
|
// inline asm
|
|
@%p116 bra BB0_123;
|
|
|
|
mov.u64 %rd179, image_RNM4;
|
|
cvta.global.u64 %rd168, %rd179;
|
|
mov.u32 %r316, 8;
|
|
// inline asm
|
|
call (%rd167), _rt_buffer_get_64, (%rd168, %r99, %r316, %rd16, %rd17, %rd25, %rd25);
|
|
// inline asm
|
|
ld.v4.u16 {%rs102, %rs103, %rs104, %rs105}, [%rd167];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f717, %rs102;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f718, %rs103;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f719, %rs104;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd173), _rt_buffer_get_64, (%rd168, %r99, %r316, %rd16, %rd17, %rd25, %rd25);
|
|
// inline asm
|
|
add.f32 %f720, %f143, %f717;
|
|
add.f32 %f721, %f144, %f718;
|
|
add.f32 %f722, %f145, %f719;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs101, %f722;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs100, %f721;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs99, %f720;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd173], {%rs99, %rs100, %rs101, %rs95};
|
|
bra.uni BB0_124;
|
|
|
|
BB0_123:
|
|
mov.u64 %rd186, image_RNM4;
|
|
cvta.global.u64 %rd181, %rd186;
|
|
mov.u32 %r318, 8;
|
|
// inline asm
|
|
call (%rd180), _rt_buffer_get_64, (%rd181, %r99, %r318, %rd16, %rd17, %rd25, %rd25);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs108, %f145;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs107, %f144;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs106, %f143;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd180], {%rs106, %rs107, %rs108, %rs95};
|
|
|
|
BB0_124:
|
|
ld.global.f32 %f727, [directDir+4];
|
|
ld.global.f32 %f728, [directDir+8];
|
|
mul.f32 %f729, %f727, %f728;
|
|
mul.f32 %f730, %f729, 0f3F8BD8A1;
|
|
mul.f32 %f146, %f128, %f730;
|
|
mul.f32 %f147, %f130, %f730;
|
|
mul.f32 %f148, %f132, %f730;
|
|
ld.global.u32 %r319, [additive];
|
|
setp.eq.s32 %p117, %r319, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs109, %f666;}
|
|
|
|
// inline asm
|
|
@%p117 bra BB0_126;
|
|
|
|
mov.u64 %rd199, image_RNM5;
|
|
cvta.global.u64 %rd188, %rd199;
|
|
mov.u32 %r323, 8;
|
|
// inline asm
|
|
call (%rd187), _rt_buffer_get_64, (%rd188, %r99, %r323, %rd16, %rd17, %rd25, %rd25);
|
|
// inline asm
|
|
ld.v4.u16 {%rs116, %rs117, %rs118, %rs119}, [%rd187];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f731, %rs116;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f732, %rs117;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f733, %rs118;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd193), _rt_buffer_get_64, (%rd188, %r99, %r323, %rd16, %rd17, %rd25, %rd25);
|
|
// inline asm
|
|
add.f32 %f734, %f146, %f731;
|
|
add.f32 %f735, %f147, %f732;
|
|
add.f32 %f736, %f148, %f733;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs115, %f736;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs114, %f735;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs113, %f734;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd193], {%rs113, %rs114, %rs115, %rs109};
|
|
bra.uni BB0_127;
|
|
|
|
BB0_126:
|
|
mov.u64 %rd206, image_RNM5;
|
|
cvta.global.u64 %rd201, %rd206;
|
|
mov.u32 %r325, 8;
|
|
// inline asm
|
|
call (%rd200), _rt_buffer_get_64, (%rd201, %r99, %r325, %rd16, %rd17, %rd25, %rd25);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs122, %f148;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs121, %f147;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs120, %f146;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd200], {%rs120, %rs121, %rs122, %rs109};
|
|
|
|
BB0_127:
|
|
ld.global.f32 %f741, [directDir+8];
|
|
fma.rn.f32 %f742, %f741, %f741, 0fBEAAAAAB;
|
|
mul.f32 %f743, %f742, 0f3F723881;
|
|
mul.f32 %f149, %f128, %f743;
|
|
mul.f32 %f150, %f130, %f743;
|
|
mul.f32 %f151, %f132, %f743;
|
|
ld.global.u32 %r326, [additive];
|
|
setp.eq.s32 %p118, %r326, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs123, %f666;}
|
|
|
|
// inline asm
|
|
@%p118 bra BB0_129;
|
|
|
|
mov.u64 %rd219, image_RNM6;
|
|
cvta.global.u64 %rd208, %rd219;
|
|
mov.u32 %r330, 8;
|
|
// inline asm
|
|
call (%rd207), _rt_buffer_get_64, (%rd208, %r99, %r330, %rd16, %rd17, %rd25, %rd25);
|
|
// inline asm
|
|
ld.v4.u16 {%rs130, %rs131, %rs132, %rs133}, [%rd207];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f744, %rs130;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f745, %rs131;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f746, %rs132;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd213), _rt_buffer_get_64, (%rd208, %r99, %r330, %rd16, %rd17, %rd25, %rd25);
|
|
// inline asm
|
|
add.f32 %f747, %f149, %f744;
|
|
add.f32 %f748, %f150, %f745;
|
|
add.f32 %f749, %f151, %f746;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs129, %f749;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs128, %f748;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs127, %f747;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd213], {%rs127, %rs128, %rs129, %rs123};
|
|
bra.uni BB0_130;
|
|
|
|
BB0_129:
|
|
mov.u64 %rd226, image_RNM6;
|
|
cvta.global.u64 %rd221, %rd226;
|
|
mov.u32 %r332, 8;
|
|
// inline asm
|
|
call (%rd220), _rt_buffer_get_64, (%rd221, %r99, %r332, %rd16, %rd17, %rd25, %rd25);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs136, %f151;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs135, %f150;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs134, %f149;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd220], {%rs134, %rs135, %rs136, %rs123};
|
|
|
|
BB0_130:
|
|
ld.global.f32 %f754, [directDir+8];
|
|
ld.global.f32 %f755, [directDir];
|
|
mul.f32 %f756, %f755, %f754;
|
|
mul.f32 %f757, %f756, 0f3F8BD8A1;
|
|
mul.f32 %f152, %f128, %f757;
|
|
mul.f32 %f153, %f130, %f757;
|
|
mul.f32 %f154, %f132, %f757;
|
|
ld.global.u32 %r333, [additive];
|
|
setp.eq.s32 %p119, %r333, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs137, %f666;}
|
|
|
|
// inline asm
|
|
@%p119 bra BB0_132;
|
|
|
|
mov.u64 %rd239, image_RNM7;
|
|
cvta.global.u64 %rd228, %rd239;
|
|
mov.u32 %r337, 8;
|
|
// inline asm
|
|
call (%rd227), _rt_buffer_get_64, (%rd228, %r99, %r337, %rd16, %rd17, %rd25, %rd25);
|
|
// inline asm
|
|
ld.v4.u16 {%rs144, %rs145, %rs146, %rs147}, [%rd227];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f758, %rs144;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f759, %rs145;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f760, %rs146;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd233), _rt_buffer_get_64, (%rd228, %r99, %r337, %rd16, %rd17, %rd25, %rd25);
|
|
// inline asm
|
|
add.f32 %f761, %f152, %f758;
|
|
add.f32 %f762, %f153, %f759;
|
|
add.f32 %f763, %f154, %f760;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs143, %f763;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs142, %f762;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs141, %f761;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd233], {%rs141, %rs142, %rs143, %rs137};
|
|
bra.uni BB0_133;
|
|
|
|
BB0_132:
|
|
mov.u64 %rd246, image_RNM7;
|
|
cvta.global.u64 %rd241, %rd246;
|
|
mov.u32 %r339, 8;
|
|
// inline asm
|
|
call (%rd240), _rt_buffer_get_64, (%rd241, %r99, %r339, %rd16, %rd17, %rd25, %rd25);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs150, %f154;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs149, %f153;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs148, %f152;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd240], {%rs148, %rs149, %rs150, %rs137};
|
|
|
|
BB0_133:
|
|
ld.global.f32 %f768, [directDir];
|
|
mul.f32 %f769, %f768, %f768;
|
|
ld.global.f32 %f770, [directDir+4];
|
|
mul.f32 %f771, %f770, %f770;
|
|
sub.f32 %f772, %f769, %f771;
|
|
mul.f32 %f773, %f772, 0f3F0BD8A1;
|
|
mul.f32 %f155, %f128, %f773;
|
|
mul.f32 %f156, %f130, %f773;
|
|
mul.f32 %f157, %f132, %f773;
|
|
ld.global.u32 %r340, [additive];
|
|
setp.eq.s32 %p120, %r340, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs151, %f666;}
|
|
|
|
// inline asm
|
|
@%p120 bra BB0_135;
|
|
|
|
mov.u64 %rd259, image_RNM8;
|
|
cvta.global.u64 %rd248, %rd259;
|
|
mov.u32 %r344, 8;
|
|
// inline asm
|
|
call (%rd247), _rt_buffer_get_64, (%rd248, %r99, %r344, %rd16, %rd17, %rd25, %rd25);
|
|
// inline asm
|
|
ld.v4.u16 {%rs158, %rs159, %rs160, %rs161}, [%rd247];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f774, %rs158;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f775, %rs159;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f776, %rs160;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd253), _rt_buffer_get_64, (%rd248, %r99, %r344, %rd16, %rd17, %rd25, %rd25);
|
|
// inline asm
|
|
add.f32 %f777, %f155, %f774;
|
|
add.f32 %f778, %f156, %f775;
|
|
add.f32 %f779, %f157, %f776;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs157, %f779;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs156, %f778;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs155, %f777;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd253], {%rs155, %rs156, %rs157, %rs151};
|
|
bra.uni BB0_171;
|
|
|
|
BB0_135:
|
|
mov.u64 %rd266, image_RNM8;
|
|
cvta.global.u64 %rd261, %rd266;
|
|
mov.u32 %r346, 8;
|
|
// inline asm
|
|
call (%rd260), _rt_buffer_get_64, (%rd261, %r99, %r346, %rd16, %rd17, %rd25, %rd25);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs164, %f157;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs163, %f156;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs162, %f155;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd260], {%rs162, %rs163, %rs164, %rs151};
|
|
|
|
BB0_171:
|
|
ret;
|
|
}
|
|
|
|
|