2890 lines
76 KiB
Plaintext
2890 lines
76 KiB
Plaintext
//
|
|
// Generated by NVIDIA NVVM Compiler
|
|
//
|
|
// Compiler Build ID: CL-23083092
|
|
// Cuda compilation tools, release 9.1, V9.1.85
|
|
// Based on LLVM 3.4svn
|
|
//
|
|
|
|
.version 6.1
|
|
.target sm_30
|
|
.address_size 64
|
|
|
|
// .globl _Z6oxMainv
|
|
.global .align 8 .b8 pixelID[8];
|
|
.global .align 8 .b8 resolution[8];
|
|
.global .align 4 .b8 normal[12];
|
|
.global .align 4 .b8 camPos[12];
|
|
.global .align 4 .b8 root[4];
|
|
.global .align 4 .u32 imageEnabled;
|
|
.global .texref lightmap;
|
|
.global .align 16 .b8 tileInfo[16];
|
|
.global .align 4 .u32 additive;
|
|
.global .align 1 .b8 image[1];
|
|
.global .align 1 .b8 image_HDR[1];
|
|
.global .align 1 .b8 image_HDR2[1];
|
|
.global .align 1 .b8 image_Mask[1];
|
|
.global .align 1 .b8 image_RNM0[1];
|
|
.global .align 1 .b8 image_RNM1[1];
|
|
.global .align 1 .b8 image_RNM2[1];
|
|
.global .align 1 .b8 image_RNM3[1];
|
|
.global .align 1 .b8 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[40];
|
|
.reg .b64 %SP;
|
|
.reg .b64 %SPL;
|
|
.reg .pred %p<134>;
|
|
.reg .b16 %rs<298>;
|
|
.reg .f32 %f<963>;
|
|
.reg .b32 %r<458>;
|
|
.reg .b64 %rd<497>;
|
|
|
|
|
|
mov.u64 %rd496, __local_depot0;
|
|
cvta.local.u64 %SP, %rd496;
|
|
ld.global.v2.u32 {%r101, %r102}, [pixelID];
|
|
cvt.u64.u32 %rd23, %r101;
|
|
cvt.u64.u32 %rd24, %r102;
|
|
mov.u64 %rd27, uvnormal;
|
|
cvta.global.u64 %rd22, %rd27;
|
|
mov.u32 %r99, 2;
|
|
mov.u32 %r100, 4;
|
|
mov.u64 %rd26, 0;
|
|
// inline asm
|
|
call (%rd21), _rt_buffer_get_64, (%rd22, %r99, %r100, %rd23, %rd24, %rd26, %rd26);
|
|
// inline asm
|
|
ld.u32 %r1, [%rd21];
|
|
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 %f921, 0f00000000;
|
|
mov.f32 %f922, %f921;
|
|
mov.f32 %f923, %f921;
|
|
@%p5 bra BB0_2;
|
|
|
|
ld.u8 %rs15, [%rd21+1];
|
|
and.b16 %rs17, %rs13, 255;
|
|
cvt.rn.f32.u16 %f175, %rs17;
|
|
div.rn.f32 %f176, %f175, 0f437F0000;
|
|
fma.rn.f32 %f177, %f176, 0f40000000, 0fBF800000;
|
|
cvt.rn.f32.u16 %f178, %rs15;
|
|
div.rn.f32 %f179, %f178, 0f437F0000;
|
|
fma.rn.f32 %f180, %f179, 0f40000000, 0fBF800000;
|
|
cvt.rn.f32.u16 %f181, %rs12;
|
|
div.rn.f32 %f182, %f181, 0f437F0000;
|
|
fma.rn.f32 %f183, %f182, 0f40000000, 0fBF800000;
|
|
mul.f32 %f184, %f180, %f180;
|
|
fma.rn.f32 %f185, %f177, %f177, %f184;
|
|
fma.rn.f32 %f186, %f183, %f183, %f185;
|
|
sqrt.rn.f32 %f187, %f186;
|
|
rcp.rn.f32 %f188, %f187;
|
|
mul.f32 %f921, %f177, %f188;
|
|
mul.f32 %f922, %f180, %f188;
|
|
mul.f32 %f923, %f183, %f188;
|
|
|
|
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, %f922, 0f00000000;
|
|
setp.eq.f32 %p7, %f921, 0f00000000;
|
|
and.pred %p8, %p7, %p6;
|
|
setp.eq.f32 %p9, %f923, 0f00000000;
|
|
and.pred %p10, %p8, %p9;
|
|
@%p10 bra BB0_137;
|
|
bra.uni BB0_3;
|
|
|
|
BB0_137:
|
|
ld.global.u32 %r457, [imageEnabled];
|
|
and.b32 %r347, %r457, 1;
|
|
setp.eq.b32 %p121, %r347, 1;
|
|
@!%p121 bra BB0_139;
|
|
bra.uni BB0_138;
|
|
|
|
BB0_138:
|
|
cvt.u64.u32 %rd280, %r2;
|
|
cvt.u64.u32 %rd281, %r3;
|
|
mov.u64 %rd284, image;
|
|
cvta.global.u64 %rd279, %rd284;
|
|
mov.u64 %rd283, 0;
|
|
// inline asm
|
|
call (%rd278), _rt_buffer_get_64, (%rd279, %r99, %r100, %rd280, %rd281, %rd283, %rd283);
|
|
// inline asm
|
|
mov.u16 %rs165, 0;
|
|
st.v4.u8 [%rd278], {%rs165, %rs165, %rs165, %rs165};
|
|
ld.global.u32 %r457, [imageEnabled];
|
|
|
|
BB0_139:
|
|
and.b32 %r350, %r457, 8;
|
|
setp.eq.s32 %p122, %r350, 0;
|
|
@%p122 bra BB0_141;
|
|
|
|
cvt.u64.u32 %rd288, %r3;
|
|
cvt.u64.u32 %rd287, %r2;
|
|
mov.u64 %rd291, image_Mask;
|
|
cvta.global.u64 %rd286, %rd291;
|
|
mov.u64 %rd290, 0;
|
|
// inline asm
|
|
call (%rd285), _rt_buffer_get_64, (%rd286, %r99, %r99, %rd287, %rd288, %rd290, %rd290);
|
|
// inline asm
|
|
mov.f32 %f808, 0f00000000;
|
|
cvt.rzi.u32.f32 %r353, %f808;
|
|
cvt.u16.u32 %rs166, %r353;
|
|
mov.u16 %rs167, 0;
|
|
st.v2.u8 [%rd285], {%rs166, %rs167};
|
|
ld.global.u32 %r457, [imageEnabled];
|
|
|
|
BB0_141:
|
|
cvt.u64.u32 %rd19, %r2;
|
|
cvt.u64.u32 %rd20, %r3;
|
|
and.b32 %r354, %r457, 4;
|
|
setp.eq.s32 %p123, %r354, 0;
|
|
@%p123 bra BB0_145;
|
|
|
|
ld.global.u32 %r355, [additive];
|
|
setp.eq.s32 %p124, %r355, 0;
|
|
@%p124 bra BB0_144;
|
|
|
|
mov.u64 %rd304, image_HDR;
|
|
cvta.global.u64 %rd293, %rd304;
|
|
mov.u32 %r359, 8;
|
|
mov.u64 %rd303, 0;
|
|
// inline asm
|
|
call (%rd292), _rt_buffer_get_64, (%rd293, %r99, %r359, %rd19, %rd20, %rd303, %rd303);
|
|
// inline asm
|
|
ld.v4.u16 {%rs174, %rs175, %rs176, %rs177}, [%rd292];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f809, %rs174;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f810, %rs175;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f811, %rs176;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd298), _rt_buffer_get_64, (%rd293, %r99, %r359, %rd19, %rd20, %rd303, %rd303);
|
|
// inline asm
|
|
add.f32 %f812, %f809, 0f00000000;
|
|
add.f32 %f813, %f810, 0f00000000;
|
|
add.f32 %f814, %f811, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs173, %f814;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs172, %f813;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs171, %f812;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs178, 0;
|
|
st.v4.u16 [%rd298], {%rs171, %rs172, %rs173, %rs178};
|
|
bra.uni BB0_145;
|
|
|
|
BB0_3:
|
|
ld.global.f32 %f9, [directDir+8];
|
|
ld.global.f32 %f8, [directDir+4];
|
|
ld.global.f32 %f7, [directDir];
|
|
mul.f32 %f192, %f921, %f7;
|
|
mul.f32 %f193, %f922, %f8;
|
|
neg.f32 %f194, %f193;
|
|
sub.f32 %f195, %f194, %f192;
|
|
mul.f32 %f196, %f923, %f9;
|
|
sub.f32 %f10, %f195, %f196;
|
|
ld.global.v2.u32 {%r118, %r119}, [pixelID];
|
|
cvt.u64.u32 %rd30, %r118;
|
|
cvt.u64.u32 %rd31, %r119;
|
|
mov.u64 %rd40, uvpos;
|
|
cvta.global.u64 %rd29, %rd40;
|
|
mov.u32 %r115, 12;
|
|
// inline asm
|
|
call (%rd28), _rt_buffer_get_64, (%rd29, %r99, %r115, %rd30, %rd31, %rd26, %rd26);
|
|
// inline asm
|
|
ld.f32 %f197, [%rd28+8];
|
|
ld.f32 %f198, [%rd28+4];
|
|
ld.f32 %f199, [%rd28];
|
|
mul.f32 %f200, %f199, 0f3456BF95;
|
|
mul.f32 %f201, %f198, 0f3456BF95;
|
|
mul.f32 %f202, %f197, 0f3456BF95;
|
|
abs.f32 %f203, %f921;
|
|
div.rn.f32 %f204, %f200, %f203;
|
|
abs.f32 %f205, %f922;
|
|
div.rn.f32 %f206, %f201, %f205;
|
|
abs.f32 %f207, %f923;
|
|
div.rn.f32 %f208, %f202, %f207;
|
|
abs.f32 %f209, %f204;
|
|
abs.f32 %f210, %f206;
|
|
abs.f32 %f211, %f208;
|
|
mov.f32 %f212, 0f38D1B717;
|
|
max.f32 %f213, %f209, %f212;
|
|
max.f32 %f214, %f210, %f212;
|
|
max.f32 %f215, %f211, %f212;
|
|
fma.rn.f32 %f11, %f921, %f213, %f199;
|
|
fma.rn.f32 %f12, %f922, %f214, %f198;
|
|
fma.rn.f32 %f13, %f923, %f215, %f197;
|
|
abs.f32 %f216, %f7;
|
|
abs.f32 %f217, %f9;
|
|
setp.gt.f32 %p11, %f216, %f217;
|
|
neg.f32 %f218, %f8;
|
|
neg.f32 %f219, %f9;
|
|
selp.f32 %f220, %f218, 0f00000000, %p11;
|
|
selp.f32 %f221, %f7, %f219, %p11;
|
|
selp.f32 %f222, 0f00000000, %f8, %p11;
|
|
mul.f32 %f223, %f221, %f221;
|
|
fma.rn.f32 %f224, %f220, %f220, %f223;
|
|
fma.rn.f32 %f225, %f222, %f222, %f224;
|
|
sqrt.rn.f32 %f226, %f225;
|
|
rcp.rn.f32 %f227, %f226;
|
|
mul.f32 %f14, %f220, %f227;
|
|
mul.f32 %f15, %f221, %f227;
|
|
mul.f32 %f16, %f222, %f227;
|
|
ld.global.v2.u32 {%r122, %r123}, [pixelID];
|
|
cvt.u64.u32 %rd36, %r122;
|
|
cvt.u64.u32 %rd37, %r123;
|
|
mov.u64 %rd41, rnd_seeds;
|
|
cvta.global.u64 %rd35, %rd41;
|
|
// inline asm
|
|
call (%rd34), _rt_buffer_get_64, (%rd35, %r99, %r100, %rd36, %rd37, %rd26, %rd26);
|
|
// inline asm
|
|
ld.global.u32 %r426, [samples];
|
|
mov.f32 %f942, 0f00000000;
|
|
setp.lt.s32 %p12, %r426, 1;
|
|
@%p12 bra BB0_4;
|
|
|
|
cvt.rn.f32.s32 %f231, %r426;
|
|
rcp.rn.f32 %f17, %f231;
|
|
ld.u32 %r452, [%rd34];
|
|
mul.f32 %f18, %f11, 0f3456BF95;
|
|
mul.f32 %f19, %f12, 0f3456BF95;
|
|
mul.f32 %f20, %f13, 0f3456BF95;
|
|
mul.f32 %f232, %f7, %f15;
|
|
mul.f32 %f233, %f8, %f14;
|
|
sub.f32 %f21, %f233, %f232;
|
|
mul.f32 %f234, %f9, %f14;
|
|
mul.f32 %f235, %f7, %f16;
|
|
sub.f32 %f22, %f235, %f234;
|
|
mul.f32 %f236, %f8, %f16;
|
|
mul.f32 %f237, %f9, %f15;
|
|
sub.f32 %f23, %f237, %f236;
|
|
mov.f32 %f942, 0f00000000;
|
|
mov.u32 %r427, 0;
|
|
abs.f32 %f238, %f19;
|
|
abs.f32 %f239, %f18;
|
|
max.f32 %f240, %f239, %f238;
|
|
abs.f32 %f241, %f20;
|
|
max.f32 %f242, %f240, %f241;
|
|
mov.f32 %f943, %f942;
|
|
mov.f32 %f944, %f942;
|
|
|
|
BB0_6:
|
|
setp.lt.s32 %p13, %r426, 1;
|
|
@%p13 bra BB0_55;
|
|
|
|
cvt.rn.f32.s32 %f27, %r427;
|
|
max.f32 %f28, %f242, %f212;
|
|
mov.u32 %r429, 0;
|
|
|
|
BB0_8:
|
|
mad.lo.s32 %r128, %r452, 1664525, 1013904223;
|
|
and.b32 %r129, %r128, 16777215;
|
|
cvt.rn.f32.u32 %f244, %r129;
|
|
fma.rn.f32 %f245, %f244, 0f33800000, %f27;
|
|
mul.f32 %f246, %f17, %f245;
|
|
mad.lo.s32 %r452, %r128, 1664525, 1013904223;
|
|
and.b32 %r130, %r452, 16777215;
|
|
cvt.rn.f32.u32 %f247, %r130;
|
|
cvt.rn.f32.s32 %f248, %r429;
|
|
fma.rn.f32 %f249, %f247, 0f33800000, %f248;
|
|
mul.f32 %f250, %f17, %f249;
|
|
sqrt.rn.f32 %f32, %f246;
|
|
mul.f32 %f936, %f250, 0f40C90FDB;
|
|
abs.f32 %f34, %f936;
|
|
setp.neu.f32 %p14, %f34, 0f7F800000;
|
|
mov.f32 %f930, %f936;
|
|
@%p14 bra BB0_10;
|
|
|
|
mov.f32 %f251, 0f00000000;
|
|
mul.rn.f32 %f930, %f936, %f251;
|
|
|
|
BB0_10:
|
|
mul.f32 %f252, %f930, 0f3F22F983;
|
|
cvt.rni.s32.f32 %r440, %f252;
|
|
cvt.rn.f32.s32 %f253, %r440;
|
|
neg.f32 %f254, %f253;
|
|
mov.f32 %f255, 0f3FC90FDA;
|
|
fma.rn.f32 %f256, %f254, %f255, %f930;
|
|
mov.f32 %f257, 0f33A22168;
|
|
fma.rn.f32 %f258, %f254, %f257, %f256;
|
|
mov.f32 %f259, 0f27C234C5;
|
|
fma.rn.f32 %f931, %f254, %f259, %f258;
|
|
abs.f32 %f260, %f930;
|
|
setp.leu.f32 %p15, %f260, 0f47CE4780;
|
|
@%p15 bra BB0_21;
|
|
|
|
add.u64 %rd43, %SP, 12;
|
|
cvta.to.local.u64 %rd3, %rd43;
|
|
mov.b32 %r13, %f930;
|
|
shr.u32 %r14, %r13, 23;
|
|
shl.b32 %r133, %r13, 8;
|
|
or.b32 %r15, %r133, -2147483648;
|
|
mov.u32 %r431, 0;
|
|
mov.u64 %rd493, 0;
|
|
mov.u64 %rd492, %rd3;
|
|
mov.u32 %r432, %r431;
|
|
|
|
BB0_12:
|
|
.pragma "nounroll";
|
|
shl.b64 %rd44, %rd493, 2;
|
|
mov.u64 %rd45, __cudart_i2opi_f;
|
|
add.s64 %rd46, %rd45, %rd44;
|
|
ld.const.u32 %r136, [%rd46];
|
|
// inline asm
|
|
{
|
|
mad.lo.cc.u32 %r134, %r136, %r15, %r432;
|
|
madc.hi.u32 %r432, %r136, %r15, 0;
|
|
}
|
|
// inline asm
|
|
st.local.u32 [%rd492], %r134;
|
|
add.s32 %r431, %r431, 1;
|
|
cvt.s64.s32 %rd493, %r431;
|
|
mul.wide.s32 %rd49, %r431, 4;
|
|
add.s64 %rd492, %rd3, %rd49;
|
|
setp.ne.s32 %p16, %r431, 6;
|
|
@%p16 bra BB0_12;
|
|
|
|
and.b32 %r139, %r14, 255;
|
|
add.s32 %r140, %r139, -128;
|
|
shr.u32 %r141, %r140, 5;
|
|
and.b32 %r20, %r13, -2147483648;
|
|
cvta.to.local.u64 %rd51, %rd43;
|
|
st.local.u32 [%rd51+24], %r432;
|
|
mov.u32 %r142, 6;
|
|
sub.s32 %r143, %r142, %r141;
|
|
mul.wide.s32 %rd52, %r143, 4;
|
|
add.s64 %rd8, %rd51, %rd52;
|
|
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_15;
|
|
|
|
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_15:
|
|
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_16;
|
|
bra.uni BB0_17;
|
|
|
|
BB0_16:
|
|
mov.u32 %r436, %r20;
|
|
mov.u32 %r437, %r29;
|
|
bra.uni BB0_18;
|
|
|
|
BB0_17:
|
|
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_18:
|
|
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_20;
|
|
|
|
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_20:
|
|
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 %f931, %r175;
|
|
|
|
BB0_21:
|
|
mul.rn.f32 %f40, %f931, %f931;
|
|
add.s32 %r46, %r440, 1;
|
|
and.b32 %r47, %r46, 1;
|
|
setp.eq.s32 %p23, %r47, 0;
|
|
@%p23 bra BB0_23;
|
|
bra.uni BB0_22;
|
|
|
|
BB0_23:
|
|
mov.f32 %f263, 0f3C08839E;
|
|
mov.f32 %f264, 0fB94CA1F9;
|
|
fma.rn.f32 %f932, %f264, %f40, %f263;
|
|
bra.uni BB0_24;
|
|
|
|
BB0_22:
|
|
mov.f32 %f261, 0fBAB6061A;
|
|
mov.f32 %f262, 0f37CCF5CE;
|
|
fma.rn.f32 %f932, %f262, %f40, %f261;
|
|
|
|
BB0_24:
|
|
@%p23 bra BB0_26;
|
|
bra.uni BB0_25;
|
|
|
|
BB0_26:
|
|
mov.f32 %f268, 0fBE2AAAA3;
|
|
fma.rn.f32 %f269, %f932, %f40, %f268;
|
|
mov.f32 %f270, 0f00000000;
|
|
fma.rn.f32 %f933, %f269, %f40, %f270;
|
|
bra.uni BB0_27;
|
|
|
|
BB0_25:
|
|
mov.f32 %f265, 0f3D2AAAA5;
|
|
fma.rn.f32 %f266, %f932, %f40, %f265;
|
|
mov.f32 %f267, 0fBF000000;
|
|
fma.rn.f32 %f933, %f266, %f40, %f267;
|
|
|
|
BB0_27:
|
|
fma.rn.f32 %f934, %f933, %f931, %f931;
|
|
@%p23 bra BB0_29;
|
|
|
|
mov.f32 %f271, 0f3F800000;
|
|
fma.rn.f32 %f934, %f933, %f40, %f271;
|
|
|
|
BB0_29:
|
|
and.b32 %r176, %r46, 2;
|
|
setp.eq.s32 %p26, %r176, 0;
|
|
@%p26 bra BB0_31;
|
|
|
|
mov.f32 %f272, 0f00000000;
|
|
mov.f32 %f273, 0fBF800000;
|
|
fma.rn.f32 %f934, %f934, %f273, %f272;
|
|
|
|
BB0_31:
|
|
@%p14 bra BB0_33;
|
|
|
|
mov.f32 %f274, 0f00000000;
|
|
mul.rn.f32 %f936, %f936, %f274;
|
|
|
|
BB0_33:
|
|
mul.f32 %f275, %f936, 0f3F22F983;
|
|
cvt.rni.s32.f32 %r450, %f275;
|
|
cvt.rn.f32.s32 %f276, %r450;
|
|
neg.f32 %f277, %f276;
|
|
fma.rn.f32 %f279, %f277, %f255, %f936;
|
|
fma.rn.f32 %f281, %f277, %f257, %f279;
|
|
fma.rn.f32 %f937, %f277, %f259, %f281;
|
|
abs.f32 %f283, %f936;
|
|
setp.leu.f32 %p28, %f283, 0f47CE4780;
|
|
@%p28 bra BB0_44;
|
|
|
|
add.u64 %rd54, %SP, 12;
|
|
cvta.to.local.u64 %rd9, %rd54;
|
|
mov.b32 %r49, %f936;
|
|
shr.u32 %r50, %r49, 23;
|
|
shl.b32 %r179, %r49, 8;
|
|
or.b32 %r51, %r179, -2147483648;
|
|
mov.u32 %r441, 0;
|
|
mov.u64 %rd494, %rd9;
|
|
mov.u64 %rd495, %rd26;
|
|
mov.u32 %r442, %r441;
|
|
|
|
BB0_35:
|
|
.pragma "nounroll";
|
|
shl.b64 %rd55, %rd495, 2;
|
|
mov.u64 %rd56, __cudart_i2opi_f;
|
|
add.s64 %rd57, %rd56, %rd55;
|
|
ld.const.u32 %r182, [%rd57];
|
|
// inline asm
|
|
{
|
|
mad.lo.cc.u32 %r180, %r182, %r51, %r442;
|
|
madc.hi.u32 %r442, %r182, %r51, 0;
|
|
}
|
|
// inline asm
|
|
st.local.u32 [%rd494], %r180;
|
|
add.s32 %r441, %r441, 1;
|
|
cvt.s64.s32 %rd495, %r441;
|
|
mul.wide.s32 %rd58, %r441, 4;
|
|
add.s64 %rd494, %rd9, %rd58;
|
|
setp.ne.s32 %p29, %r441, 6;
|
|
@%p29 bra BB0_35;
|
|
|
|
and.b32 %r185, %r50, 255;
|
|
add.s32 %r186, %r185, -128;
|
|
shr.u32 %r187, %r186, 5;
|
|
and.b32 %r56, %r49, -2147483648;
|
|
cvta.to.local.u64 %rd60, %rd54;
|
|
st.local.u32 [%rd60+24], %r442;
|
|
mov.u32 %r188, 6;
|
|
sub.s32 %r189, %r188, %r187;
|
|
mul.wide.s32 %rd61, %r189, 4;
|
|
add.s64 %rd15, %rd60, %rd61;
|
|
ld.local.u32 %r443, [%rd15];
|
|
ld.local.u32 %r444, [%rd15+-4];
|
|
and.b32 %r59, %r50, 31;
|
|
setp.eq.s32 %p30, %r59, 0;
|
|
@%p30 bra BB0_38;
|
|
|
|
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, [%rd15+-8];
|
|
shr.u32 %r195, %r194, %r191;
|
|
shl.b32 %r196, %r444, %r59;
|
|
add.s32 %r444, %r195, %r196;
|
|
|
|
BB0_38:
|
|
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_39;
|
|
bra.uni BB0_40;
|
|
|
|
BB0_39:
|
|
mov.u32 %r446, %r56;
|
|
mov.u32 %r447, %r65;
|
|
bra.uni BB0_41;
|
|
|
|
BB0_40:
|
|
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_41:
|
|
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_43;
|
|
|
|
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_43:
|
|
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 %f937, %r221;
|
|
|
|
BB0_44:
|
|
mul.rn.f32 %f57, %f937, %f937;
|
|
and.b32 %r82, %r450, 1;
|
|
setp.eq.s32 %p36, %r82, 0;
|
|
@%p36 bra BB0_46;
|
|
bra.uni BB0_45;
|
|
|
|
BB0_46:
|
|
mov.f32 %f286, 0f3C08839E;
|
|
mov.f32 %f287, 0fB94CA1F9;
|
|
fma.rn.f32 %f938, %f287, %f57, %f286;
|
|
bra.uni BB0_47;
|
|
|
|
BB0_45:
|
|
mov.f32 %f284, 0fBAB6061A;
|
|
mov.f32 %f285, 0f37CCF5CE;
|
|
fma.rn.f32 %f938, %f285, %f57, %f284;
|
|
|
|
BB0_47:
|
|
@%p36 bra BB0_49;
|
|
bra.uni BB0_48;
|
|
|
|
BB0_49:
|
|
mov.f32 %f291, 0fBE2AAAA3;
|
|
fma.rn.f32 %f292, %f938, %f57, %f291;
|
|
mov.f32 %f293, 0f00000000;
|
|
fma.rn.f32 %f939, %f292, %f57, %f293;
|
|
bra.uni BB0_50;
|
|
|
|
BB0_48:
|
|
mov.f32 %f288, 0f3D2AAAA5;
|
|
fma.rn.f32 %f289, %f938, %f57, %f288;
|
|
mov.f32 %f290, 0fBF000000;
|
|
fma.rn.f32 %f939, %f289, %f57, %f290;
|
|
|
|
BB0_50:
|
|
fma.rn.f32 %f940, %f939, %f937, %f937;
|
|
@%p36 bra BB0_52;
|
|
|
|
mov.f32 %f294, 0f3F800000;
|
|
fma.rn.f32 %f940, %f939, %f57, %f294;
|
|
|
|
BB0_52:
|
|
and.b32 %r222, %r450, 2;
|
|
setp.eq.s32 %p39, %r222, 0;
|
|
@%p39 bra BB0_54;
|
|
|
|
mov.f32 %f295, 0f00000000;
|
|
mov.f32 %f296, 0fBF800000;
|
|
fma.rn.f32 %f940, %f940, %f296, %f295;
|
|
|
|
BB0_54:
|
|
mul.f32 %f305, %f32, %f934;
|
|
add.u64 %rd62, %SP, 0;
|
|
cvta.to.local.u64 %rd63, %rd62;
|
|
mul.f32 %f306, %f305, %f305;
|
|
mov.f32 %f307, 0f3F800000;
|
|
sub.f32 %f308, %f307, %f306;
|
|
mul.f32 %f309, %f32, %f940;
|
|
mul.f32 %f310, %f309, %f309;
|
|
sub.f32 %f311, %f308, %f310;
|
|
mov.f32 %f312, 0f00000000;
|
|
max.f32 %f313, %f312, %f311;
|
|
sqrt.rn.f32 %f314, %f313;
|
|
mul.f32 %f315, %f14, %f309;
|
|
mul.f32 %f316, %f15, %f309;
|
|
mul.f32 %f317, %f16, %f309;
|
|
fma.rn.f32 %f318, %f23, %f305, %f315;
|
|
fma.rn.f32 %f319, %f22, %f305, %f316;
|
|
fma.rn.f32 %f320, %f21, %f305, %f317;
|
|
fma.rn.f32 %f321, %f7, %f314, %f318;
|
|
fma.rn.f32 %f322, %f8, %f314, %f319;
|
|
fma.rn.f32 %f323, %f9, %f314, %f320;
|
|
add.f32 %f324, %f7, %f321;
|
|
add.f32 %f325, %f8, %f322;
|
|
add.f32 %f326, %f9, %f323;
|
|
ld.global.f32 %f327, [shadowSpread];
|
|
mul.f32 %f328, %f327, %f324;
|
|
mul.f32 %f329, %f327, %f325;
|
|
mul.f32 %f330, %f327, %f326;
|
|
sub.f32 %f331, %f328, %f7;
|
|
sub.f32 %f332, %f329, %f8;
|
|
sub.f32 %f333, %f330, %f9;
|
|
mul.f32 %f334, %f332, %f332;
|
|
fma.rn.f32 %f335, %f331, %f331, %f334;
|
|
fma.rn.f32 %f336, %f333, %f333, %f335;
|
|
sqrt.rn.f32 %f337, %f336;
|
|
rcp.rn.f32 %f338, %f337;
|
|
mul.f32 %f300, %f338, %f331;
|
|
mul.f32 %f301, %f338, %f332;
|
|
mul.f32 %f302, %f338, %f333;
|
|
ld.global.u32 %r226, [imageEnabled];
|
|
and.b32 %r227, %r226, 32;
|
|
setp.eq.s32 %p40, %r227, 0;
|
|
selp.f32 %f339, 0f3F800000, 0f41200000, %p40;
|
|
mul.f32 %f303, %f339, %f28;
|
|
mov.u32 %r228, 1065353216;
|
|
st.local.u32 [%rd63+8], %r228;
|
|
st.local.u32 [%rd63+4], %r228;
|
|
st.local.u32 [%rd63], %r228;
|
|
ld.global.u32 %r223, [root];
|
|
mov.u32 %r224, 1;
|
|
mov.f32 %f304, 0f6C4ECB8F;
|
|
// inline asm
|
|
call _rt_trace_64, (%r223, %f11, %f12, %f13, %f300, %f301, %f302, %r224, %f303, %f304, %rd62, %r115);
|
|
// inline asm
|
|
ld.local.f32 %f340, [%rd63];
|
|
add.f32 %f944, %f944, %f340;
|
|
ld.local.f32 %f341, [%rd63+4];
|
|
add.f32 %f943, %f943, %f341;
|
|
ld.local.f32 %f342, [%rd63+8];
|
|
add.f32 %f942, %f942, %f342;
|
|
ld.global.u32 %r426, [samples];
|
|
add.s32 %r429, %r429, 1;
|
|
setp.lt.s32 %p41, %r429, %r426;
|
|
@%p41 bra BB0_8;
|
|
|
|
BB0_55:
|
|
add.s32 %r427, %r427, 1;
|
|
setp.lt.s32 %p42, %r427, %r426;
|
|
@%p42 bra BB0_6;
|
|
bra.uni BB0_56;
|
|
|
|
BB0_4:
|
|
mov.f32 %f943, %f942;
|
|
mov.f32 %f944, %f942;
|
|
|
|
BB0_56:
|
|
setp.eq.s32 %p43, %r426, 0;
|
|
mov.f32 %f948, 0f3F800000;
|
|
mov.f32 %f949, %f948;
|
|
mov.f32 %f950, %f948;
|
|
@%p43 bra BB0_58;
|
|
|
|
mul.lo.s32 %r229, %r426, %r426;
|
|
cvt.rn.f32.s32 %f346, %r229;
|
|
rcp.rn.f32 %f347, %f346;
|
|
mul.f32 %f948, %f944, %f347;
|
|
mul.f32 %f949, %f943, %f347;
|
|
mul.f32 %f950, %f942, %f347;
|
|
|
|
BB0_58:
|
|
ld.global.f32 %f348, [directColor];
|
|
mul.f32 %f84, %f948, %f348;
|
|
ld.global.f32 %f349, [directColor+4];
|
|
mul.f32 %f85, %f949, %f349;
|
|
ld.global.f32 %f350, [directColor+8];
|
|
mul.f32 %f86, %f950, %f350;
|
|
cvt.sat.f32.f32 %f351, %f10;
|
|
mul.f32 %f87, %f84, %f351;
|
|
mul.f32 %f88, %f85, %f351;
|
|
mul.f32 %f89, %f86, %f351;
|
|
ld.global.u32 %r455, [imageEnabled];
|
|
and.b32 %r230, %r455, 8;
|
|
setp.eq.s32 %p44, %r230, 0;
|
|
@%p44 bra BB0_71;
|
|
|
|
cvt.u64.u32 %rd66, %r2;
|
|
cvt.u64.u32 %rd67, %r3;
|
|
mov.u64 %rd70, image_Mask;
|
|
cvta.global.u64 %rd65, %rd70;
|
|
// inline asm
|
|
call (%rd64), _rt_buffer_get_64, (%rd65, %r99, %r99, %rd66, %rd67, %rd26, %rd26);
|
|
// inline asm
|
|
abs.f32 %f91, %f948;
|
|
setp.lt.f32 %p45, %f91, 0f00800000;
|
|
mul.f32 %f357, %f91, 0f4B800000;
|
|
selp.f32 %f358, 0fC3170000, 0fC2FE0000, %p45;
|
|
selp.f32 %f359, %f357, %f91, %p45;
|
|
mov.b32 %r233, %f359;
|
|
and.b32 %r234, %r233, 8388607;
|
|
or.b32 %r235, %r234, 1065353216;
|
|
mov.b32 %f360, %r235;
|
|
shr.u32 %r236, %r233, 23;
|
|
cvt.rn.f32.u32 %f361, %r236;
|
|
add.f32 %f362, %f358, %f361;
|
|
setp.gt.f32 %p46, %f360, 0f3FB504F3;
|
|
mul.f32 %f363, %f360, 0f3F000000;
|
|
add.f32 %f364, %f362, 0f3F800000;
|
|
selp.f32 %f365, %f363, %f360, %p46;
|
|
selp.f32 %f366, %f364, %f362, %p46;
|
|
add.f32 %f367, %f365, 0fBF800000;
|
|
add.f32 %f353, %f365, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f352,%f353;
|
|
// inline asm
|
|
add.f32 %f368, %f367, %f367;
|
|
mul.f32 %f369, %f352, %f368;
|
|
mul.f32 %f370, %f369, %f369;
|
|
mov.f32 %f371, 0f3C4CAF63;
|
|
mov.f32 %f372, 0f3B18F0FE;
|
|
fma.rn.f32 %f373, %f372, %f370, %f371;
|
|
mov.f32 %f374, 0f3DAAAABD;
|
|
fma.rn.f32 %f375, %f373, %f370, %f374;
|
|
mul.rn.f32 %f376, %f375, %f370;
|
|
mul.rn.f32 %f377, %f376, %f369;
|
|
sub.f32 %f378, %f367, %f369;
|
|
neg.f32 %f379, %f369;
|
|
add.f32 %f380, %f378, %f378;
|
|
fma.rn.f32 %f381, %f379, %f367, %f380;
|
|
mul.rn.f32 %f382, %f352, %f381;
|
|
add.f32 %f383, %f377, %f369;
|
|
sub.f32 %f384, %f369, %f383;
|
|
add.f32 %f385, %f377, %f384;
|
|
add.f32 %f386, %f382, %f385;
|
|
add.f32 %f387, %f383, %f386;
|
|
sub.f32 %f388, %f383, %f387;
|
|
add.f32 %f389, %f386, %f388;
|
|
mov.f32 %f390, 0f3F317200;
|
|
mul.rn.f32 %f391, %f366, %f390;
|
|
mov.f32 %f392, 0f35BFBE8E;
|
|
mul.rn.f32 %f393, %f366, %f392;
|
|
add.f32 %f394, %f391, %f387;
|
|
sub.f32 %f395, %f391, %f394;
|
|
add.f32 %f396, %f387, %f395;
|
|
add.f32 %f397, %f389, %f396;
|
|
add.f32 %f398, %f393, %f397;
|
|
add.f32 %f399, %f394, %f398;
|
|
sub.f32 %f400, %f394, %f399;
|
|
add.f32 %f401, %f398, %f400;
|
|
mov.f32 %f402, 0f3EE8BA2E;
|
|
mul.rn.f32 %f403, %f402, %f399;
|
|
neg.f32 %f404, %f403;
|
|
fma.rn.f32 %f405, %f402, %f399, %f404;
|
|
fma.rn.f32 %f406, %f402, %f401, %f405;
|
|
mov.f32 %f407, 0f00000000;
|
|
fma.rn.f32 %f408, %f407, %f399, %f406;
|
|
add.rn.f32 %f409, %f403, %f408;
|
|
neg.f32 %f410, %f409;
|
|
add.rn.f32 %f411, %f403, %f410;
|
|
add.rn.f32 %f412, %f411, %f408;
|
|
mov.b32 %r237, %f409;
|
|
setp.eq.s32 %p47, %r237, 1118925336;
|
|
add.s32 %r238, %r237, -1;
|
|
mov.b32 %f413, %r238;
|
|
add.f32 %f414, %f412, 0f37000000;
|
|
selp.f32 %f415, %f413, %f409, %p47;
|
|
selp.f32 %f92, %f414, %f412, %p47;
|
|
mul.f32 %f416, %f415, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f417, %f416;
|
|
mov.f32 %f418, 0fBF317200;
|
|
fma.rn.f32 %f419, %f417, %f418, %f415;
|
|
mov.f32 %f420, 0fB5BFBE8E;
|
|
fma.rn.f32 %f421, %f417, %f420, %f419;
|
|
mul.f32 %f422, %f421, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f423, %f422;
|
|
add.f32 %f424, %f417, 0f00000000;
|
|
ex2.approx.f32 %f425, %f424;
|
|
mul.f32 %f426, %f423, %f425;
|
|
setp.lt.f32 %p48, %f415, 0fC2D20000;
|
|
selp.f32 %f427, 0f00000000, %f426, %p48;
|
|
setp.gt.f32 %p49, %f415, 0f42D20000;
|
|
selp.f32 %f951, 0f7F800000, %f427, %p49;
|
|
setp.eq.f32 %p50, %f951, 0f7F800000;
|
|
@%p50 bra BB0_61;
|
|
|
|
fma.rn.f32 %f951, %f951, %f92, %f951;
|
|
|
|
BB0_61:
|
|
mov.f32 %f913, 0f3E68BA2E;
|
|
cvt.rzi.f32.f32 %f912, %f913;
|
|
fma.rn.f32 %f911, %f912, 0fC0000000, 0f3EE8BA2E;
|
|
abs.f32 %f910, %f911;
|
|
setp.lt.f32 %p51, %f948, 0f00000000;
|
|
setp.eq.f32 %p52, %f910, 0f3F800000;
|
|
and.pred %p1, %p51, %p52;
|
|
mov.b32 %r239, %f951;
|
|
xor.b32 %r240, %r239, -2147483648;
|
|
mov.b32 %f428, %r240;
|
|
selp.f32 %f953, %f428, %f951, %p1;
|
|
setp.eq.f32 %p53, %f948, 0f00000000;
|
|
@%p53 bra BB0_64;
|
|
bra.uni BB0_62;
|
|
|
|
BB0_64:
|
|
add.f32 %f431, %f948, %f948;
|
|
selp.f32 %f953, %f431, 0f00000000, %p52;
|
|
bra.uni BB0_65;
|
|
|
|
BB0_144:
|
|
mov.u64 %rd311, image_HDR;
|
|
cvta.global.u64 %rd306, %rd311;
|
|
mov.u32 %r361, 8;
|
|
mov.u64 %rd310, 0;
|
|
// inline asm
|
|
call (%rd305), _rt_buffer_get_64, (%rd306, %r99, %r361, %rd19, %rd20, %rd310, %rd310);
|
|
// inline asm
|
|
mov.f32 %f815, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs179, %f815;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs180, 0;
|
|
st.v4.u16 [%rd305], {%rs179, %rs179, %rs179, %rs180};
|
|
|
|
BB0_145:
|
|
ld.global.u32 %r362, [additive];
|
|
setp.eq.s32 %p125, %r362, 0;
|
|
@%p125 bra BB0_147;
|
|
|
|
mov.u64 %rd324, image_RNM0;
|
|
cvta.global.u64 %rd313, %rd324;
|
|
mov.u32 %r366, 8;
|
|
mov.u64 %rd323, 0;
|
|
// inline asm
|
|
call (%rd312), _rt_buffer_get_64, (%rd313, %r99, %r366, %rd19, %rd20, %rd323, %rd323);
|
|
// inline asm
|
|
ld.v4.u16 {%rs187, %rs188, %rs189, %rs190}, [%rd312];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f816, %rs187;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f817, %rs188;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f818, %rs189;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd318), _rt_buffer_get_64, (%rd313, %r99, %r366, %rd19, %rd20, %rd323, %rd323);
|
|
// inline asm
|
|
add.f32 %f819, %f816, 0f00000000;
|
|
add.f32 %f820, %f817, 0f00000000;
|
|
add.f32 %f821, %f818, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs186, %f821;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs185, %f820;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs184, %f819;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs191, 0;
|
|
st.v4.u16 [%rd318], {%rs184, %rs185, %rs186, %rs191};
|
|
bra.uni BB0_148;
|
|
|
|
BB0_147:
|
|
mov.u64 %rd331, image_RNM0;
|
|
cvta.global.u64 %rd326, %rd331;
|
|
mov.u32 %r368, 8;
|
|
mov.u64 %rd330, 0;
|
|
// inline asm
|
|
call (%rd325), _rt_buffer_get_64, (%rd326, %r99, %r368, %rd19, %rd20, %rd330, %rd330);
|
|
// inline asm
|
|
mov.f32 %f822, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs192, %f822;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs193, 0;
|
|
st.v4.u16 [%rd325], {%rs192, %rs192, %rs192, %rs193};
|
|
|
|
BB0_148:
|
|
ld.global.u32 %r369, [additive];
|
|
setp.eq.s32 %p126, %r369, 0;
|
|
@%p126 bra BB0_150;
|
|
|
|
mov.u64 %rd344, image_RNM1;
|
|
cvta.global.u64 %rd333, %rd344;
|
|
mov.u32 %r373, 8;
|
|
mov.u64 %rd343, 0;
|
|
// inline asm
|
|
call (%rd332), _rt_buffer_get_64, (%rd333, %r99, %r373, %rd19, %rd20, %rd343, %rd343);
|
|
// inline asm
|
|
ld.v4.u16 {%rs200, %rs201, %rs202, %rs203}, [%rd332];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f823, %rs200;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f824, %rs201;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f825, %rs202;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd338), _rt_buffer_get_64, (%rd333, %r99, %r373, %rd19, %rd20, %rd343, %rd343);
|
|
// inline asm
|
|
add.f32 %f826, %f823, 0f00000000;
|
|
add.f32 %f827, %f824, 0f00000000;
|
|
add.f32 %f828, %f825, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs199, %f828;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs198, %f827;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs197, %f826;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs204, 0;
|
|
st.v4.u16 [%rd338], {%rs197, %rs198, %rs199, %rs204};
|
|
bra.uni BB0_151;
|
|
|
|
BB0_150:
|
|
mov.u64 %rd351, image_RNM1;
|
|
cvta.global.u64 %rd346, %rd351;
|
|
mov.u32 %r375, 8;
|
|
mov.u64 %rd350, 0;
|
|
// inline asm
|
|
call (%rd345), _rt_buffer_get_64, (%rd346, %r99, %r375, %rd19, %rd20, %rd350, %rd350);
|
|
// inline asm
|
|
mov.f32 %f829, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs205, %f829;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs206, 0;
|
|
st.v4.u16 [%rd345], {%rs205, %rs205, %rs205, %rs206};
|
|
|
|
BB0_151:
|
|
ld.global.u32 %r376, [additive];
|
|
setp.eq.s32 %p127, %r376, 0;
|
|
@%p127 bra BB0_153;
|
|
|
|
mov.u64 %rd364, image_RNM2;
|
|
cvta.global.u64 %rd353, %rd364;
|
|
mov.u32 %r380, 8;
|
|
mov.u64 %rd363, 0;
|
|
// inline asm
|
|
call (%rd352), _rt_buffer_get_64, (%rd353, %r99, %r380, %rd19, %rd20, %rd363, %rd363);
|
|
// inline asm
|
|
ld.v4.u16 {%rs213, %rs214, %rs215, %rs216}, [%rd352];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f830, %rs213;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f831, %rs214;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f832, %rs215;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd358), _rt_buffer_get_64, (%rd353, %r99, %r380, %rd19, %rd20, %rd363, %rd363);
|
|
// inline asm
|
|
add.f32 %f833, %f830, 0f00000000;
|
|
add.f32 %f834, %f831, 0f00000000;
|
|
add.f32 %f835, %f832, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs212, %f835;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs211, %f834;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs210, %f833;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs217, 0;
|
|
st.v4.u16 [%rd358], {%rs210, %rs211, %rs212, %rs217};
|
|
bra.uni BB0_154;
|
|
|
|
BB0_153:
|
|
mov.u64 %rd371, image_RNM2;
|
|
cvta.global.u64 %rd366, %rd371;
|
|
mov.u32 %r382, 8;
|
|
mov.u64 %rd370, 0;
|
|
// inline asm
|
|
call (%rd365), _rt_buffer_get_64, (%rd366, %r99, %r382, %rd19, %rd20, %rd370, %rd370);
|
|
// inline asm
|
|
mov.f32 %f836, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs218, %f836;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs219, 0;
|
|
st.v4.u16 [%rd365], {%rs218, %rs218, %rs218, %rs219};
|
|
|
|
BB0_154:
|
|
ld.global.u32 %r383, [additive];
|
|
setp.eq.s32 %p128, %r383, 0;
|
|
@%p128 bra BB0_156;
|
|
|
|
mov.u64 %rd384, image_RNM3;
|
|
cvta.global.u64 %rd373, %rd384;
|
|
mov.u32 %r387, 8;
|
|
mov.u64 %rd383, 0;
|
|
// inline asm
|
|
call (%rd372), _rt_buffer_get_64, (%rd373, %r99, %r387, %rd19, %rd20, %rd383, %rd383);
|
|
// inline asm
|
|
ld.v4.u16 {%rs226, %rs227, %rs228, %rs229}, [%rd372];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f837, %rs226;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f838, %rs227;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f839, %rs228;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd378), _rt_buffer_get_64, (%rd373, %r99, %r387, %rd19, %rd20, %rd383, %rd383);
|
|
// inline asm
|
|
add.f32 %f840, %f837, 0f00000000;
|
|
add.f32 %f841, %f838, 0f00000000;
|
|
add.f32 %f842, %f839, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs225, %f842;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs224, %f841;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs223, %f840;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs230, 0;
|
|
st.v4.u16 [%rd378], {%rs223, %rs224, %rs225, %rs230};
|
|
bra.uni BB0_157;
|
|
|
|
BB0_156:
|
|
mov.u64 %rd391, image_RNM3;
|
|
cvta.global.u64 %rd386, %rd391;
|
|
mov.u32 %r389, 8;
|
|
mov.u64 %rd390, 0;
|
|
// inline asm
|
|
call (%rd385), _rt_buffer_get_64, (%rd386, %r99, %r389, %rd19, %rd20, %rd390, %rd390);
|
|
// inline asm
|
|
mov.f32 %f843, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs231, %f843;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs232, 0;
|
|
st.v4.u16 [%rd385], {%rs231, %rs231, %rs231, %rs232};
|
|
|
|
BB0_157:
|
|
ld.global.u32 %r390, [additive];
|
|
setp.eq.s32 %p129, %r390, 0;
|
|
@%p129 bra BB0_159;
|
|
|
|
mov.u64 %rd404, image_RNM4;
|
|
cvta.global.u64 %rd393, %rd404;
|
|
mov.u32 %r394, 8;
|
|
mov.u64 %rd403, 0;
|
|
// inline asm
|
|
call (%rd392), _rt_buffer_get_64, (%rd393, %r99, %r394, %rd19, %rd20, %rd403, %rd403);
|
|
// inline asm
|
|
ld.v4.u16 {%rs239, %rs240, %rs241, %rs242}, [%rd392];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f844, %rs239;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f845, %rs240;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f846, %rs241;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd398), _rt_buffer_get_64, (%rd393, %r99, %r394, %rd19, %rd20, %rd403, %rd403);
|
|
// inline asm
|
|
add.f32 %f847, %f844, 0f00000000;
|
|
add.f32 %f848, %f845, 0f00000000;
|
|
add.f32 %f849, %f846, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs238, %f849;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs237, %f848;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs236, %f847;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs243, 0;
|
|
st.v4.u16 [%rd398], {%rs236, %rs237, %rs238, %rs243};
|
|
bra.uni BB0_160;
|
|
|
|
BB0_159:
|
|
mov.u64 %rd411, image_RNM4;
|
|
cvta.global.u64 %rd406, %rd411;
|
|
mov.u32 %r396, 8;
|
|
mov.u64 %rd410, 0;
|
|
// inline asm
|
|
call (%rd405), _rt_buffer_get_64, (%rd406, %r99, %r396, %rd19, %rd20, %rd410, %rd410);
|
|
// inline asm
|
|
mov.f32 %f850, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs244, %f850;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs245, 0;
|
|
st.v4.u16 [%rd405], {%rs244, %rs244, %rs244, %rs245};
|
|
|
|
BB0_160:
|
|
ld.global.u32 %r397, [additive];
|
|
setp.eq.s32 %p130, %r397, 0;
|
|
@%p130 bra BB0_162;
|
|
|
|
mov.u64 %rd424, image_RNM5;
|
|
cvta.global.u64 %rd413, %rd424;
|
|
mov.u32 %r401, 8;
|
|
mov.u64 %rd423, 0;
|
|
// inline asm
|
|
call (%rd412), _rt_buffer_get_64, (%rd413, %r99, %r401, %rd19, %rd20, %rd423, %rd423);
|
|
// inline asm
|
|
ld.v4.u16 {%rs252, %rs253, %rs254, %rs255}, [%rd412];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f851, %rs252;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f852, %rs253;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f853, %rs254;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd418), _rt_buffer_get_64, (%rd413, %r99, %r401, %rd19, %rd20, %rd423, %rd423);
|
|
// inline asm
|
|
add.f32 %f854, %f851, 0f00000000;
|
|
add.f32 %f855, %f852, 0f00000000;
|
|
add.f32 %f856, %f853, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs251, %f856;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs250, %f855;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs249, %f854;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs256, 0;
|
|
st.v4.u16 [%rd418], {%rs249, %rs250, %rs251, %rs256};
|
|
bra.uni BB0_163;
|
|
|
|
BB0_162:
|
|
mov.u64 %rd431, image_RNM5;
|
|
cvta.global.u64 %rd426, %rd431;
|
|
mov.u32 %r403, 8;
|
|
mov.u64 %rd430, 0;
|
|
// inline asm
|
|
call (%rd425), _rt_buffer_get_64, (%rd426, %r99, %r403, %rd19, %rd20, %rd430, %rd430);
|
|
// inline asm
|
|
mov.f32 %f857, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs257, %f857;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs258, 0;
|
|
st.v4.u16 [%rd425], {%rs257, %rs257, %rs257, %rs258};
|
|
|
|
BB0_163:
|
|
ld.global.u32 %r404, [additive];
|
|
setp.eq.s32 %p131, %r404, 0;
|
|
@%p131 bra BB0_165;
|
|
|
|
mov.u64 %rd444, image_RNM6;
|
|
cvta.global.u64 %rd433, %rd444;
|
|
mov.u32 %r408, 8;
|
|
mov.u64 %rd443, 0;
|
|
// inline asm
|
|
call (%rd432), _rt_buffer_get_64, (%rd433, %r99, %r408, %rd19, %rd20, %rd443, %rd443);
|
|
// inline asm
|
|
ld.v4.u16 {%rs265, %rs266, %rs267, %rs268}, [%rd432];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f858, %rs265;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f859, %rs266;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f860, %rs267;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd438), _rt_buffer_get_64, (%rd433, %r99, %r408, %rd19, %rd20, %rd443, %rd443);
|
|
// inline asm
|
|
add.f32 %f861, %f858, 0f00000000;
|
|
add.f32 %f862, %f859, 0f00000000;
|
|
add.f32 %f863, %f860, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs264, %f863;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs263, %f862;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs262, %f861;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs269, 0;
|
|
st.v4.u16 [%rd438], {%rs262, %rs263, %rs264, %rs269};
|
|
bra.uni BB0_166;
|
|
|
|
BB0_165:
|
|
mov.u64 %rd451, image_RNM6;
|
|
cvta.global.u64 %rd446, %rd451;
|
|
mov.u32 %r410, 8;
|
|
mov.u64 %rd450, 0;
|
|
// inline asm
|
|
call (%rd445), _rt_buffer_get_64, (%rd446, %r99, %r410, %rd19, %rd20, %rd450, %rd450);
|
|
// inline asm
|
|
mov.f32 %f864, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs270, %f864;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs271, 0;
|
|
st.v4.u16 [%rd445], {%rs270, %rs270, %rs270, %rs271};
|
|
|
|
BB0_166:
|
|
ld.global.u32 %r411, [additive];
|
|
setp.eq.s32 %p132, %r411, 0;
|
|
@%p132 bra BB0_168;
|
|
|
|
mov.u64 %rd464, image_RNM7;
|
|
cvta.global.u64 %rd453, %rd464;
|
|
mov.u32 %r415, 8;
|
|
mov.u64 %rd463, 0;
|
|
// inline asm
|
|
call (%rd452), _rt_buffer_get_64, (%rd453, %r99, %r415, %rd19, %rd20, %rd463, %rd463);
|
|
// inline asm
|
|
ld.v4.u16 {%rs278, %rs279, %rs280, %rs281}, [%rd452];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f865, %rs278;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f866, %rs279;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f867, %rs280;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd458), _rt_buffer_get_64, (%rd453, %r99, %r415, %rd19, %rd20, %rd463, %rd463);
|
|
// inline asm
|
|
add.f32 %f868, %f865, 0f00000000;
|
|
add.f32 %f869, %f866, 0f00000000;
|
|
add.f32 %f870, %f867, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs277, %f870;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs276, %f869;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs275, %f868;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs282, 0;
|
|
st.v4.u16 [%rd458], {%rs275, %rs276, %rs277, %rs282};
|
|
bra.uni BB0_169;
|
|
|
|
BB0_168:
|
|
mov.u64 %rd471, image_RNM7;
|
|
cvta.global.u64 %rd466, %rd471;
|
|
mov.u32 %r417, 8;
|
|
mov.u64 %rd470, 0;
|
|
// inline asm
|
|
call (%rd465), _rt_buffer_get_64, (%rd466, %r99, %r417, %rd19, %rd20, %rd470, %rd470);
|
|
// inline asm
|
|
mov.f32 %f871, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs283, %f871;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs284, 0;
|
|
st.v4.u16 [%rd465], {%rs283, %rs283, %rs283, %rs284};
|
|
|
|
BB0_169:
|
|
ld.global.u32 %r418, [additive];
|
|
setp.eq.s32 %p133, %r418, 0;
|
|
@%p133 bra BB0_171;
|
|
|
|
mov.u64 %rd484, image_RNM8;
|
|
cvta.global.u64 %rd473, %rd484;
|
|
mov.u32 %r422, 8;
|
|
mov.u64 %rd483, 0;
|
|
// inline asm
|
|
call (%rd472), _rt_buffer_get_64, (%rd473, %r99, %r422, %rd19, %rd20, %rd483, %rd483);
|
|
// inline asm
|
|
ld.v4.u16 {%rs291, %rs292, %rs293, %rs294}, [%rd472];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f872, %rs291;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f873, %rs292;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f874, %rs293;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd478), _rt_buffer_get_64, (%rd473, %r99, %r422, %rd19, %rd20, %rd483, %rd483);
|
|
// inline asm
|
|
add.f32 %f875, %f872, 0f00000000;
|
|
add.f32 %f876, %f873, 0f00000000;
|
|
add.f32 %f877, %f874, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs290, %f877;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs289, %f876;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs288, %f875;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs295, 0;
|
|
st.v4.u16 [%rd478], {%rs288, %rs289, %rs290, %rs295};
|
|
bra.uni BB0_172;
|
|
|
|
BB0_171:
|
|
mov.u64 %rd491, image_RNM8;
|
|
cvta.global.u64 %rd486, %rd491;
|
|
mov.u32 %r424, 8;
|
|
mov.u64 %rd490, 0;
|
|
// inline asm
|
|
call (%rd485), _rt_buffer_get_64, (%rd486, %r99, %r424, %rd19, %rd20, %rd490, %rd490);
|
|
// inline asm
|
|
mov.f32 %f878, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs296, %f878;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs297, 0;
|
|
st.v4.u16 [%rd485], {%rs296, %rs296, %rs296, %rs297};
|
|
bra.uni BB0_172;
|
|
|
|
BB0_62:
|
|
setp.geu.f32 %p54, %f948, 0f00000000;
|
|
@%p54 bra BB0_65;
|
|
|
|
mov.f32 %f917, 0f3EE8BA2E;
|
|
cvt.rzi.f32.f32 %f430, %f917;
|
|
setp.neu.f32 %p55, %f430, 0f3EE8BA2E;
|
|
selp.f32 %f953, 0f7FFFFFFF, %f953, %p55;
|
|
|
|
BB0_65:
|
|
abs.f32 %f914, %f948;
|
|
add.f32 %f432, %f914, 0f3EE8BA2E;
|
|
mov.b32 %r241, %f432;
|
|
setp.lt.s32 %p57, %r241, 2139095040;
|
|
@%p57 bra BB0_70;
|
|
|
|
abs.f32 %f915, %f948;
|
|
setp.gtu.f32 %p58, %f915, 0f7F800000;
|
|
@%p58 bra BB0_69;
|
|
bra.uni BB0_67;
|
|
|
|
BB0_69:
|
|
add.f32 %f953, %f948, 0f3EE8BA2E;
|
|
bra.uni BB0_70;
|
|
|
|
BB0_67:
|
|
abs.f32 %f916, %f948;
|
|
setp.neu.f32 %p59, %f916, 0f7F800000;
|
|
@%p59 bra BB0_70;
|
|
|
|
selp.f32 %f953, 0fFF800000, 0f7F800000, %p1;
|
|
|
|
BB0_70:
|
|
mul.f32 %f433, %f953, 0f437F0000;
|
|
setp.eq.f32 %p60, %f948, 0f3F800000;
|
|
selp.f32 %f434, 0f437F0000, %f433, %p60;
|
|
cvt.rzi.u32.f32 %r242, %f434;
|
|
cvt.u16.u32 %rs19, %r242;
|
|
mov.u16 %rs20, 255;
|
|
st.v2.u8 [%rd64], {%rs19, %rs20};
|
|
ld.global.u32 %r455, [imageEnabled];
|
|
|
|
BB0_71:
|
|
and.b32 %r243, %r455, 1;
|
|
setp.eq.b32 %p61, %r243, 1;
|
|
@!%p61 bra BB0_106;
|
|
bra.uni BB0_72;
|
|
|
|
BB0_72:
|
|
abs.f32 %f104, %f87;
|
|
setp.lt.f32 %p62, %f104, 0f00800000;
|
|
mul.f32 %f440, %f104, 0f4B800000;
|
|
selp.f32 %f441, 0fC3170000, 0fC2FE0000, %p62;
|
|
selp.f32 %f442, %f440, %f104, %p62;
|
|
mov.b32 %r244, %f442;
|
|
and.b32 %r245, %r244, 8388607;
|
|
or.b32 %r246, %r245, 1065353216;
|
|
mov.b32 %f443, %r246;
|
|
shr.u32 %r247, %r244, 23;
|
|
cvt.rn.f32.u32 %f444, %r247;
|
|
add.f32 %f445, %f441, %f444;
|
|
setp.gt.f32 %p63, %f443, 0f3FB504F3;
|
|
mul.f32 %f446, %f443, 0f3F000000;
|
|
add.f32 %f447, %f445, 0f3F800000;
|
|
selp.f32 %f448, %f446, %f443, %p63;
|
|
selp.f32 %f449, %f447, %f445, %p63;
|
|
add.f32 %f450, %f448, 0fBF800000;
|
|
add.f32 %f436, %f448, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f435,%f436;
|
|
// inline asm
|
|
add.f32 %f451, %f450, %f450;
|
|
mul.f32 %f452, %f435, %f451;
|
|
mul.f32 %f453, %f452, %f452;
|
|
mov.f32 %f454, 0f3C4CAF63;
|
|
mov.f32 %f455, 0f3B18F0FE;
|
|
fma.rn.f32 %f456, %f455, %f453, %f454;
|
|
mov.f32 %f457, 0f3DAAAABD;
|
|
fma.rn.f32 %f458, %f456, %f453, %f457;
|
|
mul.rn.f32 %f459, %f458, %f453;
|
|
mul.rn.f32 %f460, %f459, %f452;
|
|
sub.f32 %f461, %f450, %f452;
|
|
neg.f32 %f462, %f452;
|
|
add.f32 %f463, %f461, %f461;
|
|
fma.rn.f32 %f464, %f462, %f450, %f463;
|
|
mul.rn.f32 %f465, %f435, %f464;
|
|
add.f32 %f466, %f460, %f452;
|
|
sub.f32 %f467, %f452, %f466;
|
|
add.f32 %f468, %f460, %f467;
|
|
add.f32 %f469, %f465, %f468;
|
|
add.f32 %f470, %f466, %f469;
|
|
sub.f32 %f471, %f466, %f470;
|
|
add.f32 %f472, %f469, %f471;
|
|
mov.f32 %f473, 0f3F317200;
|
|
mul.rn.f32 %f474, %f449, %f473;
|
|
mov.f32 %f475, 0f35BFBE8E;
|
|
mul.rn.f32 %f476, %f449, %f475;
|
|
add.f32 %f477, %f474, %f470;
|
|
sub.f32 %f478, %f474, %f477;
|
|
add.f32 %f479, %f470, %f478;
|
|
add.f32 %f480, %f472, %f479;
|
|
add.f32 %f481, %f476, %f480;
|
|
add.f32 %f482, %f477, %f481;
|
|
sub.f32 %f483, %f477, %f482;
|
|
add.f32 %f484, %f481, %f483;
|
|
mov.f32 %f485, 0f3EE66666;
|
|
mul.rn.f32 %f486, %f485, %f482;
|
|
neg.f32 %f487, %f486;
|
|
fma.rn.f32 %f488, %f485, %f482, %f487;
|
|
fma.rn.f32 %f489, %f485, %f484, %f488;
|
|
mov.f32 %f490, 0f00000000;
|
|
fma.rn.f32 %f491, %f490, %f482, %f489;
|
|
add.rn.f32 %f492, %f486, %f491;
|
|
neg.f32 %f493, %f492;
|
|
add.rn.f32 %f494, %f486, %f493;
|
|
add.rn.f32 %f495, %f494, %f491;
|
|
mov.b32 %r248, %f492;
|
|
setp.eq.s32 %p64, %r248, 1118925336;
|
|
add.s32 %r249, %r248, -1;
|
|
mov.b32 %f496, %r249;
|
|
add.f32 %f497, %f495, 0f37000000;
|
|
selp.f32 %f498, %f496, %f492, %p64;
|
|
selp.f32 %f105, %f497, %f495, %p64;
|
|
mul.f32 %f499, %f498, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f500, %f499;
|
|
mov.f32 %f501, 0fBF317200;
|
|
fma.rn.f32 %f502, %f500, %f501, %f498;
|
|
mov.f32 %f503, 0fB5BFBE8E;
|
|
fma.rn.f32 %f504, %f500, %f503, %f502;
|
|
mul.f32 %f505, %f504, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f506, %f505;
|
|
add.f32 %f507, %f500, 0f00000000;
|
|
ex2.approx.f32 %f508, %f507;
|
|
mul.f32 %f509, %f506, %f508;
|
|
setp.lt.f32 %p65, %f498, 0fC2D20000;
|
|
selp.f32 %f510, 0f00000000, %f509, %p65;
|
|
setp.gt.f32 %p66, %f498, 0f42D20000;
|
|
selp.f32 %f954, 0f7F800000, %f510, %p66;
|
|
setp.eq.f32 %p67, %f954, 0f7F800000;
|
|
@%p67 bra BB0_74;
|
|
|
|
fma.rn.f32 %f954, %f954, %f105, %f954;
|
|
|
|
BB0_74:
|
|
mov.f32 %f882, 0f3E666666;
|
|
cvt.rzi.f32.f32 %f881, %f882;
|
|
fma.rn.f32 %f880, %f881, 0fC0000000, 0f3EE66666;
|
|
abs.f32 %f879, %f880;
|
|
setp.lt.f32 %p68, %f87, 0f00000000;
|
|
setp.eq.f32 %p69, %f879, 0f3F800000;
|
|
and.pred %p2, %p68, %p69;
|
|
mov.b32 %r250, %f954;
|
|
xor.b32 %r251, %r250, -2147483648;
|
|
mov.b32 %f511, %r251;
|
|
selp.f32 %f956, %f511, %f954, %p2;
|
|
setp.eq.f32 %p70, %f87, 0f00000000;
|
|
@%p70 bra BB0_77;
|
|
bra.uni BB0_75;
|
|
|
|
BB0_77:
|
|
add.f32 %f514, %f87, %f87;
|
|
selp.f32 %f956, %f514, 0f00000000, %p69;
|
|
bra.uni BB0_78;
|
|
|
|
BB0_75:
|
|
setp.geu.f32 %p71, %f87, 0f00000000;
|
|
@%p71 bra BB0_78;
|
|
|
|
mov.f32 %f906, 0f3EE66666;
|
|
cvt.rzi.f32.f32 %f513, %f906;
|
|
setp.neu.f32 %p72, %f513, 0f3EE66666;
|
|
selp.f32 %f956, 0f7FFFFFFF, %f956, %p72;
|
|
|
|
BB0_78:
|
|
abs.f32 %f883, %f87;
|
|
add.f32 %f515, %f883, 0f3EE66666;
|
|
mov.b32 %r252, %f515;
|
|
setp.lt.s32 %p74, %r252, 2139095040;
|
|
@%p74 bra BB0_83;
|
|
|
|
abs.f32 %f904, %f87;
|
|
setp.gtu.f32 %p75, %f904, 0f7F800000;
|
|
@%p75 bra BB0_82;
|
|
bra.uni BB0_80;
|
|
|
|
BB0_82:
|
|
add.f32 %f956, %f87, 0f3EE66666;
|
|
bra.uni BB0_83;
|
|
|
|
BB0_80:
|
|
abs.f32 %f905, %f87;
|
|
setp.neu.f32 %p76, %f905, 0f7F800000;
|
|
@%p76 bra BB0_83;
|
|
|
|
selp.f32 %f956, 0fFF800000, 0f7F800000, %p2;
|
|
|
|
BB0_83:
|
|
mov.f32 %f892, 0fB5BFBE8E;
|
|
mov.f32 %f891, 0fBF317200;
|
|
mov.f32 %f890, 0f00000000;
|
|
mov.f32 %f889, 0f35BFBE8E;
|
|
mov.f32 %f888, 0f3F317200;
|
|
mov.f32 %f887, 0f3DAAAABD;
|
|
mov.f32 %f886, 0f3C4CAF63;
|
|
mov.f32 %f885, 0f3B18F0FE;
|
|
mov.f32 %f884, 0f3EE66666;
|
|
setp.eq.f32 %p77, %f87, 0f3F800000;
|
|
selp.f32 %f116, 0f3F800000, %f956, %p77;
|
|
abs.f32 %f117, %f88;
|
|
setp.lt.f32 %p78, %f117, 0f00800000;
|
|
mul.f32 %f518, %f117, 0f4B800000;
|
|
selp.f32 %f519, 0fC3170000, 0fC2FE0000, %p78;
|
|
selp.f32 %f520, %f518, %f117, %p78;
|
|
mov.b32 %r253, %f520;
|
|
and.b32 %r254, %r253, 8388607;
|
|
or.b32 %r255, %r254, 1065353216;
|
|
mov.b32 %f521, %r255;
|
|
shr.u32 %r256, %r253, 23;
|
|
cvt.rn.f32.u32 %f522, %r256;
|
|
add.f32 %f523, %f519, %f522;
|
|
setp.gt.f32 %p79, %f521, 0f3FB504F3;
|
|
mul.f32 %f524, %f521, 0f3F000000;
|
|
add.f32 %f525, %f523, 0f3F800000;
|
|
selp.f32 %f526, %f524, %f521, %p79;
|
|
selp.f32 %f527, %f525, %f523, %p79;
|
|
add.f32 %f528, %f526, 0fBF800000;
|
|
add.f32 %f517, %f526, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f516,%f517;
|
|
// inline asm
|
|
add.f32 %f529, %f528, %f528;
|
|
mul.f32 %f530, %f516, %f529;
|
|
mul.f32 %f531, %f530, %f530;
|
|
fma.rn.f32 %f534, %f885, %f531, %f886;
|
|
fma.rn.f32 %f536, %f534, %f531, %f887;
|
|
mul.rn.f32 %f537, %f536, %f531;
|
|
mul.rn.f32 %f538, %f537, %f530;
|
|
sub.f32 %f539, %f528, %f530;
|
|
neg.f32 %f540, %f530;
|
|
add.f32 %f541, %f539, %f539;
|
|
fma.rn.f32 %f542, %f540, %f528, %f541;
|
|
mul.rn.f32 %f543, %f516, %f542;
|
|
add.f32 %f544, %f538, %f530;
|
|
sub.f32 %f545, %f530, %f544;
|
|
add.f32 %f546, %f538, %f545;
|
|
add.f32 %f547, %f543, %f546;
|
|
add.f32 %f548, %f544, %f547;
|
|
sub.f32 %f549, %f544, %f548;
|
|
add.f32 %f550, %f547, %f549;
|
|
mul.rn.f32 %f552, %f527, %f888;
|
|
mul.rn.f32 %f554, %f527, %f889;
|
|
add.f32 %f555, %f552, %f548;
|
|
sub.f32 %f556, %f552, %f555;
|
|
add.f32 %f557, %f548, %f556;
|
|
add.f32 %f558, %f550, %f557;
|
|
add.f32 %f559, %f554, %f558;
|
|
add.f32 %f560, %f555, %f559;
|
|
sub.f32 %f561, %f555, %f560;
|
|
add.f32 %f562, %f559, %f561;
|
|
mul.rn.f32 %f564, %f884, %f560;
|
|
neg.f32 %f565, %f564;
|
|
fma.rn.f32 %f566, %f884, %f560, %f565;
|
|
fma.rn.f32 %f567, %f884, %f562, %f566;
|
|
fma.rn.f32 %f569, %f890, %f560, %f567;
|
|
add.rn.f32 %f570, %f564, %f569;
|
|
neg.f32 %f571, %f570;
|
|
add.rn.f32 %f572, %f564, %f571;
|
|
add.rn.f32 %f573, %f572, %f569;
|
|
mov.b32 %r257, %f570;
|
|
setp.eq.s32 %p80, %r257, 1118925336;
|
|
add.s32 %r258, %r257, -1;
|
|
mov.b32 %f574, %r258;
|
|
add.f32 %f575, %f573, 0f37000000;
|
|
selp.f32 %f576, %f574, %f570, %p80;
|
|
selp.f32 %f118, %f575, %f573, %p80;
|
|
mul.f32 %f577, %f576, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f578, %f577;
|
|
fma.rn.f32 %f580, %f578, %f891, %f576;
|
|
fma.rn.f32 %f582, %f578, %f892, %f580;
|
|
mul.f32 %f583, %f582, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f584, %f583;
|
|
add.f32 %f585, %f578, 0f00000000;
|
|
ex2.approx.f32 %f586, %f585;
|
|
mul.f32 %f587, %f584, %f586;
|
|
setp.lt.f32 %p81, %f576, 0fC2D20000;
|
|
selp.f32 %f588, 0f00000000, %f587, %p81;
|
|
setp.gt.f32 %p82, %f576, 0f42D20000;
|
|
selp.f32 %f957, 0f7F800000, %f588, %p82;
|
|
setp.eq.f32 %p83, %f957, 0f7F800000;
|
|
@%p83 bra BB0_85;
|
|
|
|
fma.rn.f32 %f957, %f957, %f118, %f957;
|
|
|
|
BB0_85:
|
|
setp.lt.f32 %p84, %f88, 0f00000000;
|
|
and.pred %p3, %p84, %p69;
|
|
mov.b32 %r259, %f957;
|
|
xor.b32 %r260, %r259, -2147483648;
|
|
mov.b32 %f589, %r260;
|
|
selp.f32 %f959, %f589, %f957, %p3;
|
|
setp.eq.f32 %p86, %f88, 0f00000000;
|
|
@%p86 bra BB0_88;
|
|
bra.uni BB0_86;
|
|
|
|
BB0_88:
|
|
add.f32 %f592, %f88, %f88;
|
|
selp.f32 %f959, %f592, 0f00000000, %p69;
|
|
bra.uni BB0_89;
|
|
|
|
BB0_86:
|
|
setp.geu.f32 %p87, %f88, 0f00000000;
|
|
@%p87 bra BB0_89;
|
|
|
|
mov.f32 %f903, 0f3EE66666;
|
|
cvt.rzi.f32.f32 %f591, %f903;
|
|
setp.neu.f32 %p88, %f591, 0f3EE66666;
|
|
selp.f32 %f959, 0f7FFFFFFF, %f959, %p88;
|
|
|
|
BB0_89:
|
|
abs.f32 %f907, %f88;
|
|
add.f32 %f593, %f907, 0f3EE66666;
|
|
mov.b32 %r261, %f593;
|
|
setp.lt.s32 %p90, %r261, 2139095040;
|
|
@%p90 bra BB0_94;
|
|
|
|
abs.f32 %f908, %f88;
|
|
setp.gtu.f32 %p91, %f908, 0f7F800000;
|
|
@%p91 bra BB0_93;
|
|
bra.uni BB0_91;
|
|
|
|
BB0_93:
|
|
add.f32 %f959, %f88, 0f3EE66666;
|
|
bra.uni BB0_94;
|
|
|
|
BB0_91:
|
|
abs.f32 %f909, %f88;
|
|
setp.neu.f32 %p92, %f909, 0f7F800000;
|
|
@%p92 bra BB0_94;
|
|
|
|
selp.f32 %f959, 0fFF800000, 0f7F800000, %p3;
|
|
|
|
BB0_94:
|
|
mov.f32 %f901, 0fB5BFBE8E;
|
|
mov.f32 %f900, 0fBF317200;
|
|
mov.f32 %f899, 0f00000000;
|
|
mov.f32 %f898, 0f35BFBE8E;
|
|
mov.f32 %f897, 0f3F317200;
|
|
mov.f32 %f896, 0f3DAAAABD;
|
|
mov.f32 %f895, 0f3C4CAF63;
|
|
mov.f32 %f894, 0f3B18F0FE;
|
|
mov.f32 %f893, 0f3EE66666;
|
|
setp.eq.f32 %p93, %f88, 0f3F800000;
|
|
selp.f32 %f129, 0f3F800000, %f959, %p93;
|
|
abs.f32 %f130, %f89;
|
|
setp.lt.f32 %p94, %f130, 0f00800000;
|
|
mul.f32 %f596, %f130, 0f4B800000;
|
|
selp.f32 %f597, 0fC3170000, 0fC2FE0000, %p94;
|
|
selp.f32 %f598, %f596, %f130, %p94;
|
|
mov.b32 %r262, %f598;
|
|
and.b32 %r263, %r262, 8388607;
|
|
or.b32 %r264, %r263, 1065353216;
|
|
mov.b32 %f599, %r264;
|
|
shr.u32 %r265, %r262, 23;
|
|
cvt.rn.f32.u32 %f600, %r265;
|
|
add.f32 %f601, %f597, %f600;
|
|
setp.gt.f32 %p95, %f599, 0f3FB504F3;
|
|
mul.f32 %f602, %f599, 0f3F000000;
|
|
add.f32 %f603, %f601, 0f3F800000;
|
|
selp.f32 %f604, %f602, %f599, %p95;
|
|
selp.f32 %f605, %f603, %f601, %p95;
|
|
add.f32 %f606, %f604, 0fBF800000;
|
|
add.f32 %f595, %f604, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f594,%f595;
|
|
// inline asm
|
|
add.f32 %f607, %f606, %f606;
|
|
mul.f32 %f608, %f594, %f607;
|
|
mul.f32 %f609, %f608, %f608;
|
|
fma.rn.f32 %f612, %f894, %f609, %f895;
|
|
fma.rn.f32 %f614, %f612, %f609, %f896;
|
|
mul.rn.f32 %f615, %f614, %f609;
|
|
mul.rn.f32 %f616, %f615, %f608;
|
|
sub.f32 %f617, %f606, %f608;
|
|
neg.f32 %f618, %f608;
|
|
add.f32 %f619, %f617, %f617;
|
|
fma.rn.f32 %f620, %f618, %f606, %f619;
|
|
mul.rn.f32 %f621, %f594, %f620;
|
|
add.f32 %f622, %f616, %f608;
|
|
sub.f32 %f623, %f608, %f622;
|
|
add.f32 %f624, %f616, %f623;
|
|
add.f32 %f625, %f621, %f624;
|
|
add.f32 %f626, %f622, %f625;
|
|
sub.f32 %f627, %f622, %f626;
|
|
add.f32 %f628, %f625, %f627;
|
|
mul.rn.f32 %f630, %f605, %f897;
|
|
mul.rn.f32 %f632, %f605, %f898;
|
|
add.f32 %f633, %f630, %f626;
|
|
sub.f32 %f634, %f630, %f633;
|
|
add.f32 %f635, %f626, %f634;
|
|
add.f32 %f636, %f628, %f635;
|
|
add.f32 %f637, %f632, %f636;
|
|
add.f32 %f638, %f633, %f637;
|
|
sub.f32 %f639, %f633, %f638;
|
|
add.f32 %f640, %f637, %f639;
|
|
mul.rn.f32 %f642, %f893, %f638;
|
|
neg.f32 %f643, %f642;
|
|
fma.rn.f32 %f644, %f893, %f638, %f643;
|
|
fma.rn.f32 %f645, %f893, %f640, %f644;
|
|
fma.rn.f32 %f647, %f899, %f638, %f645;
|
|
add.rn.f32 %f648, %f642, %f647;
|
|
neg.f32 %f649, %f648;
|
|
add.rn.f32 %f650, %f642, %f649;
|
|
add.rn.f32 %f651, %f650, %f647;
|
|
mov.b32 %r266, %f648;
|
|
setp.eq.s32 %p96, %r266, 1118925336;
|
|
add.s32 %r267, %r266, -1;
|
|
mov.b32 %f652, %r267;
|
|
add.f32 %f653, %f651, 0f37000000;
|
|
selp.f32 %f654, %f652, %f648, %p96;
|
|
selp.f32 %f131, %f653, %f651, %p96;
|
|
mul.f32 %f655, %f654, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f656, %f655;
|
|
fma.rn.f32 %f658, %f656, %f900, %f654;
|
|
fma.rn.f32 %f660, %f656, %f901, %f658;
|
|
mul.f32 %f661, %f660, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f662, %f661;
|
|
add.f32 %f663, %f656, 0f00000000;
|
|
ex2.approx.f32 %f664, %f663;
|
|
mul.f32 %f665, %f662, %f664;
|
|
setp.lt.f32 %p97, %f654, 0fC2D20000;
|
|
selp.f32 %f666, 0f00000000, %f665, %p97;
|
|
setp.gt.f32 %p98, %f654, 0f42D20000;
|
|
selp.f32 %f960, 0f7F800000, %f666, %p98;
|
|
setp.eq.f32 %p99, %f960, 0f7F800000;
|
|
@%p99 bra BB0_96;
|
|
|
|
fma.rn.f32 %f960, %f960, %f131, %f960;
|
|
|
|
BB0_96:
|
|
setp.lt.f32 %p100, %f89, 0f00000000;
|
|
and.pred %p4, %p100, %p69;
|
|
mov.b32 %r268, %f960;
|
|
xor.b32 %r269, %r268, -2147483648;
|
|
mov.b32 %f667, %r269;
|
|
selp.f32 %f962, %f667, %f960, %p4;
|
|
setp.eq.f32 %p102, %f89, 0f00000000;
|
|
@%p102 bra BB0_99;
|
|
bra.uni BB0_97;
|
|
|
|
BB0_99:
|
|
add.f32 %f670, %f89, %f89;
|
|
selp.f32 %f962, %f670, 0f00000000, %p69;
|
|
bra.uni BB0_100;
|
|
|
|
BB0_97:
|
|
setp.geu.f32 %p103, %f89, 0f00000000;
|
|
@%p103 bra BB0_100;
|
|
|
|
mov.f32 %f902, 0f3EE66666;
|
|
cvt.rzi.f32.f32 %f669, %f902;
|
|
setp.neu.f32 %p104, %f669, 0f3EE66666;
|
|
selp.f32 %f962, 0f7FFFFFFF, %f962, %p104;
|
|
|
|
BB0_100:
|
|
abs.f32 %f918, %f89;
|
|
add.f32 %f671, %f918, 0f3EE66666;
|
|
mov.b32 %r270, %f671;
|
|
setp.lt.s32 %p106, %r270, 2139095040;
|
|
@%p106 bra BB0_105;
|
|
|
|
abs.f32 %f919, %f89;
|
|
setp.gtu.f32 %p107, %f919, 0f7F800000;
|
|
@%p107 bra BB0_104;
|
|
bra.uni BB0_102;
|
|
|
|
BB0_104:
|
|
add.f32 %f962, %f89, 0f3EE66666;
|
|
bra.uni BB0_105;
|
|
|
|
BB0_102:
|
|
abs.f32 %f920, %f89;
|
|
setp.neu.f32 %p108, %f920, 0f7F800000;
|
|
@%p108 bra BB0_105;
|
|
|
|
selp.f32 %f962, 0fFF800000, 0f7F800000, %p4;
|
|
|
|
BB0_105:
|
|
mov.u32 %r425, 4;
|
|
setp.eq.f32 %p109, %f89, 0f3F800000;
|
|
selp.f32 %f672, 0f3F800000, %f962, %p109;
|
|
cvt.u64.u32 %rd74, %r3;
|
|
cvt.u64.u32 %rd73, %r2;
|
|
mov.u64 %rd77, image;
|
|
cvta.global.u64 %rd72, %rd77;
|
|
// inline asm
|
|
call (%rd71), _rt_buffer_get_64, (%rd72, %r99, %r425, %rd73, %rd74, %rd26, %rd26);
|
|
// inline asm
|
|
cvt.sat.f32.f32 %f673, %f672;
|
|
mul.f32 %f674, %f673, 0f437FFD71;
|
|
cvt.rzi.u32.f32 %r273, %f674;
|
|
cvt.sat.f32.f32 %f675, %f129;
|
|
mul.f32 %f676, %f675, 0f437FFD71;
|
|
cvt.rzi.u32.f32 %r274, %f676;
|
|
cvt.sat.f32.f32 %f677, %f116;
|
|
mul.f32 %f678, %f677, 0f437FFD71;
|
|
cvt.rzi.u32.f32 %r275, %f678;
|
|
cvt.u16.u32 %rs21, %r273;
|
|
cvt.u16.u32 %rs22, %r275;
|
|
cvt.u16.u32 %rs23, %r274;
|
|
mov.u16 %rs24, 255;
|
|
st.v4.u8 [%rd71], {%rs21, %rs23, %rs22, %rs24};
|
|
ld.global.u32 %r455, [imageEnabled];
|
|
|
|
BB0_106:
|
|
cvt.u64.u32 %rd17, %r2;
|
|
cvt.u64.u32 %rd18, %r3;
|
|
and.b32 %r276, %r455, 4;
|
|
setp.eq.s32 %p110, %r276, 0;
|
|
@%p110 bra BB0_110;
|
|
|
|
ld.global.u32 %r277, [additive];
|
|
setp.eq.s32 %p111, %r277, 0;
|
|
mov.f32 %f679, 0f3F800000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs25, %f679;}
|
|
|
|
// inline asm
|
|
@%p111 bra BB0_109;
|
|
|
|
mov.u64 %rd90, image_HDR;
|
|
cvta.global.u64 %rd79, %rd90;
|
|
mov.u32 %r281, 8;
|
|
// inline asm
|
|
call (%rd78), _rt_buffer_get_64, (%rd79, %r99, %r281, %rd17, %rd18, %rd26, %rd26);
|
|
// inline asm
|
|
ld.v4.u16 {%rs32, %rs33, %rs34, %rs35}, [%rd78];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f680, %rs32;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f681, %rs33;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f682, %rs34;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd84), _rt_buffer_get_64, (%rd79, %r99, %r281, %rd17, %rd18, %rd26, %rd26);
|
|
// inline asm
|
|
add.f32 %f683, %f87, %f680;
|
|
add.f32 %f684, %f88, %f681;
|
|
add.f32 %f685, %f89, %f682;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs31, %f685;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs30, %f684;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs29, %f683;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd84], {%rs29, %rs30, %rs31, %rs25};
|
|
bra.uni BB0_110;
|
|
|
|
BB0_109:
|
|
mov.u64 %rd97, image_HDR;
|
|
cvta.global.u64 %rd92, %rd97;
|
|
mov.u32 %r283, 8;
|
|
// inline asm
|
|
call (%rd91), _rt_buffer_get_64, (%rd92, %r99, %r283, %rd17, %rd18, %rd26, %rd26);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs38, %f89;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs37, %f88;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs36, %f87;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd91], {%rs36, %rs37, %rs38, %rs25};
|
|
|
|
BB0_110:
|
|
mul.f32 %f142, %f84, 0f403D3C19;
|
|
mul.f32 %f143, %f142, 0f3E906EBB;
|
|
mul.f32 %f144, %f85, 0f403D3C19;
|
|
mul.f32 %f145, %f144, 0f3E906EBB;
|
|
mul.f32 %f146, %f86, 0f403D3C19;
|
|
mul.f32 %f147, %f146, 0f3E906EBB;
|
|
ld.global.u32 %r284, [additive];
|
|
setp.eq.s32 %p112, %r284, 0;
|
|
mov.f32 %f689, 0f3F800000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs39, %f689;}
|
|
|
|
// inline asm
|
|
@%p112 bra BB0_112;
|
|
|
|
mov.u64 %rd110, image_RNM0;
|
|
cvta.global.u64 %rd99, %rd110;
|
|
mov.u32 %r288, 8;
|
|
// inline asm
|
|
call (%rd98), _rt_buffer_get_64, (%rd99, %r99, %r288, %rd17, %rd18, %rd26, %rd26);
|
|
// inline asm
|
|
ld.v4.u16 {%rs46, %rs47, %rs48, %rs49}, [%rd98];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f690, %rs46;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f691, %rs47;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f692, %rs48;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd104), _rt_buffer_get_64, (%rd99, %r99, %r288, %rd17, %rd18, %rd26, %rd26);
|
|
// inline asm
|
|
add.f32 %f693, %f143, %f690;
|
|
add.f32 %f694, %f145, %f691;
|
|
add.f32 %f695, %f147, %f692;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs45, %f695;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs44, %f694;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs43, %f693;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd104], {%rs43, %rs44, %rs45, %rs39};
|
|
bra.uni BB0_113;
|
|
|
|
BB0_112:
|
|
mov.u64 %rd117, image_RNM0;
|
|
cvta.global.u64 %rd112, %rd117;
|
|
mov.u32 %r290, 8;
|
|
// inline asm
|
|
call (%rd111), _rt_buffer_get_64, (%rd112, %r99, %r290, %rd17, %rd18, %rd26, %rd26);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs52, %f147;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs51, %f145;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs50, %f143;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd111], {%rs50, %rs51, %rs52, %rs39};
|
|
|
|
BB0_113:
|
|
ld.global.f32 %f700, [directDir+4];
|
|
mul.f32 %f701, %f700, 0fBEFA2A1C;
|
|
mul.f32 %f148, %f142, %f701;
|
|
mul.f32 %f149, %f144, %f701;
|
|
mul.f32 %f150, %f146, %f701;
|
|
ld.global.u32 %r291, [additive];
|
|
setp.eq.s32 %p113, %r291, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs53, %f689;}
|
|
|
|
// inline asm
|
|
@%p113 bra BB0_115;
|
|
|
|
mov.u64 %rd130, image_RNM1;
|
|
cvta.global.u64 %rd119, %rd130;
|
|
mov.u32 %r295, 8;
|
|
// inline asm
|
|
call (%rd118), _rt_buffer_get_64, (%rd119, %r99, %r295, %rd17, %rd18, %rd26, %rd26);
|
|
// inline asm
|
|
ld.v4.u16 {%rs60, %rs61, %rs62, %rs63}, [%rd118];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f702, %rs60;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f703, %rs61;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f704, %rs62;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd124), _rt_buffer_get_64, (%rd119, %r99, %r295, %rd17, %rd18, %rd26, %rd26);
|
|
// inline asm
|
|
add.f32 %f705, %f148, %f702;
|
|
add.f32 %f706, %f149, %f703;
|
|
add.f32 %f707, %f150, %f704;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs59, %f707;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs58, %f706;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs57, %f705;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd124], {%rs57, %rs58, %rs59, %rs53};
|
|
bra.uni BB0_116;
|
|
|
|
BB0_115:
|
|
mov.u64 %rd137, image_RNM1;
|
|
cvta.global.u64 %rd132, %rd137;
|
|
mov.u32 %r297, 8;
|
|
// inline asm
|
|
call (%rd131), _rt_buffer_get_64, (%rd132, %r99, %r297, %rd17, %rd18, %rd26, %rd26);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs66, %f150;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs65, %f149;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs64, %f148;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd131], {%rs64, %rs65, %rs66, %rs53};
|
|
|
|
BB0_116:
|
|
ld.global.f32 %f712, [directDir+8];
|
|
mul.f32 %f713, %f712, 0fBEFA2A1C;
|
|
mul.f32 %f151, %f142, %f713;
|
|
mul.f32 %f152, %f144, %f713;
|
|
mul.f32 %f153, %f146, %f713;
|
|
ld.global.u32 %r298, [additive];
|
|
setp.eq.s32 %p114, %r298, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs67, %f689;}
|
|
|
|
// inline asm
|
|
@%p114 bra BB0_118;
|
|
|
|
mov.u64 %rd150, image_RNM2;
|
|
cvta.global.u64 %rd139, %rd150;
|
|
mov.u32 %r302, 8;
|
|
// inline asm
|
|
call (%rd138), _rt_buffer_get_64, (%rd139, %r99, %r302, %rd17, %rd18, %rd26, %rd26);
|
|
// inline asm
|
|
ld.v4.u16 {%rs74, %rs75, %rs76, %rs77}, [%rd138];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f714, %rs74;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f715, %rs75;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f716, %rs76;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd144), _rt_buffer_get_64, (%rd139, %r99, %r302, %rd17, %rd18, %rd26, %rd26);
|
|
// inline asm
|
|
add.f32 %f717, %f151, %f714;
|
|
add.f32 %f718, %f152, %f715;
|
|
add.f32 %f719, %f153, %f716;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs73, %f719;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs72, %f718;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs71, %f717;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd144], {%rs71, %rs72, %rs73, %rs67};
|
|
bra.uni BB0_119;
|
|
|
|
BB0_118:
|
|
mov.u64 %rd157, image_RNM2;
|
|
cvta.global.u64 %rd152, %rd157;
|
|
mov.u32 %r304, 8;
|
|
// inline asm
|
|
call (%rd151), _rt_buffer_get_64, (%rd152, %r99, %r304, %rd17, %rd18, %rd26, %rd26);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs80, %f153;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs79, %f152;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs78, %f151;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd151], {%rs78, %rs79, %rs80, %rs67};
|
|
|
|
BB0_119:
|
|
ld.global.f32 %f724, [directDir];
|
|
mul.f32 %f725, %f724, 0fBEFA2A1C;
|
|
mul.f32 %f154, %f142, %f725;
|
|
mul.f32 %f155, %f144, %f725;
|
|
mul.f32 %f156, %f146, %f725;
|
|
ld.global.u32 %r305, [additive];
|
|
setp.eq.s32 %p115, %r305, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs81, %f689;}
|
|
|
|
// inline asm
|
|
@%p115 bra BB0_121;
|
|
|
|
mov.u64 %rd170, image_RNM3;
|
|
cvta.global.u64 %rd159, %rd170;
|
|
mov.u32 %r309, 8;
|
|
// inline asm
|
|
call (%rd158), _rt_buffer_get_64, (%rd159, %r99, %r309, %rd17, %rd18, %rd26, %rd26);
|
|
// inline asm
|
|
ld.v4.u16 {%rs88, %rs89, %rs90, %rs91}, [%rd158];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f726, %rs88;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f727, %rs89;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f728, %rs90;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd164), _rt_buffer_get_64, (%rd159, %r99, %r309, %rd17, %rd18, %rd26, %rd26);
|
|
// inline asm
|
|
add.f32 %f729, %f154, %f726;
|
|
add.f32 %f730, %f155, %f727;
|
|
add.f32 %f731, %f156, %f728;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs87, %f731;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs86, %f730;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs85, %f729;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd164], {%rs85, %rs86, %rs87, %rs81};
|
|
bra.uni BB0_122;
|
|
|
|
BB0_121:
|
|
mov.u64 %rd177, image_RNM3;
|
|
cvta.global.u64 %rd172, %rd177;
|
|
mov.u32 %r311, 8;
|
|
// inline asm
|
|
call (%rd171), _rt_buffer_get_64, (%rd172, %r99, %r311, %rd17, %rd18, %rd26, %rd26);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs94, %f156;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs93, %f155;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs92, %f154;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd171], {%rs92, %rs93, %rs94, %rs81};
|
|
|
|
BB0_122:
|
|
ld.global.f32 %f736, [directDir+4];
|
|
ld.global.f32 %f737, [directDir];
|
|
mul.f32 %f738, %f737, %f736;
|
|
mul.f32 %f739, %f738, 0f3F8BD8A1;
|
|
mul.f32 %f157, %f142, %f739;
|
|
mul.f32 %f158, %f144, %f739;
|
|
mul.f32 %f159, %f146, %f739;
|
|
ld.global.u32 %r312, [additive];
|
|
setp.eq.s32 %p116, %r312, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs95, %f689;}
|
|
|
|
// inline asm
|
|
@%p116 bra BB0_124;
|
|
|
|
mov.u64 %rd190, image_RNM4;
|
|
cvta.global.u64 %rd179, %rd190;
|
|
mov.u32 %r316, 8;
|
|
// inline asm
|
|
call (%rd178), _rt_buffer_get_64, (%rd179, %r99, %r316, %rd17, %rd18, %rd26, %rd26);
|
|
// inline asm
|
|
ld.v4.u16 {%rs102, %rs103, %rs104, %rs105}, [%rd178];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f740, %rs102;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f741, %rs103;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f742, %rs104;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd184), _rt_buffer_get_64, (%rd179, %r99, %r316, %rd17, %rd18, %rd26, %rd26);
|
|
// inline asm
|
|
add.f32 %f743, %f157, %f740;
|
|
add.f32 %f744, %f158, %f741;
|
|
add.f32 %f745, %f159, %f742;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs101, %f745;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs100, %f744;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs99, %f743;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd184], {%rs99, %rs100, %rs101, %rs95};
|
|
bra.uni BB0_125;
|
|
|
|
BB0_124:
|
|
mov.u64 %rd197, image_RNM4;
|
|
cvta.global.u64 %rd192, %rd197;
|
|
mov.u32 %r318, 8;
|
|
// inline asm
|
|
call (%rd191), _rt_buffer_get_64, (%rd192, %r99, %r318, %rd17, %rd18, %rd26, %rd26);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs108, %f159;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs107, %f158;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs106, %f157;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd191], {%rs106, %rs107, %rs108, %rs95};
|
|
|
|
BB0_125:
|
|
ld.global.f32 %f750, [directDir+4];
|
|
ld.global.f32 %f751, [directDir+8];
|
|
mul.f32 %f752, %f750, %f751;
|
|
mul.f32 %f753, %f752, 0f3F8BD8A1;
|
|
mul.f32 %f160, %f142, %f753;
|
|
mul.f32 %f161, %f144, %f753;
|
|
mul.f32 %f162, %f146, %f753;
|
|
ld.global.u32 %r319, [additive];
|
|
setp.eq.s32 %p117, %r319, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs109, %f689;}
|
|
|
|
// inline asm
|
|
@%p117 bra BB0_127;
|
|
|
|
mov.u64 %rd210, image_RNM5;
|
|
cvta.global.u64 %rd199, %rd210;
|
|
mov.u32 %r323, 8;
|
|
// inline asm
|
|
call (%rd198), _rt_buffer_get_64, (%rd199, %r99, %r323, %rd17, %rd18, %rd26, %rd26);
|
|
// inline asm
|
|
ld.v4.u16 {%rs116, %rs117, %rs118, %rs119}, [%rd198];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f754, %rs116;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f755, %rs117;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f756, %rs118;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd204), _rt_buffer_get_64, (%rd199, %r99, %r323, %rd17, %rd18, %rd26, %rd26);
|
|
// inline asm
|
|
add.f32 %f757, %f160, %f754;
|
|
add.f32 %f758, %f161, %f755;
|
|
add.f32 %f759, %f162, %f756;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs115, %f759;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs114, %f758;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs113, %f757;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd204], {%rs113, %rs114, %rs115, %rs109};
|
|
bra.uni BB0_128;
|
|
|
|
BB0_127:
|
|
mov.u64 %rd217, image_RNM5;
|
|
cvta.global.u64 %rd212, %rd217;
|
|
mov.u32 %r325, 8;
|
|
// inline asm
|
|
call (%rd211), _rt_buffer_get_64, (%rd212, %r99, %r325, %rd17, %rd18, %rd26, %rd26);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs122, %f162;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs121, %f161;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs120, %f160;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd211], {%rs120, %rs121, %rs122, %rs109};
|
|
|
|
BB0_128:
|
|
ld.global.f32 %f764, [directDir+8];
|
|
fma.rn.f32 %f765, %f764, %f764, 0fBEAAAAAB;
|
|
mul.f32 %f766, %f765, 0f3F723881;
|
|
mul.f32 %f163, %f142, %f766;
|
|
mul.f32 %f164, %f144, %f766;
|
|
mul.f32 %f165, %f146, %f766;
|
|
ld.global.u32 %r326, [additive];
|
|
setp.eq.s32 %p118, %r326, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs123, %f689;}
|
|
|
|
// inline asm
|
|
@%p118 bra BB0_130;
|
|
|
|
mov.u64 %rd230, image_RNM6;
|
|
cvta.global.u64 %rd219, %rd230;
|
|
mov.u32 %r330, 8;
|
|
// inline asm
|
|
call (%rd218), _rt_buffer_get_64, (%rd219, %r99, %r330, %rd17, %rd18, %rd26, %rd26);
|
|
// inline asm
|
|
ld.v4.u16 {%rs130, %rs131, %rs132, %rs133}, [%rd218];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f767, %rs130;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f768, %rs131;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f769, %rs132;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd224), _rt_buffer_get_64, (%rd219, %r99, %r330, %rd17, %rd18, %rd26, %rd26);
|
|
// inline asm
|
|
add.f32 %f770, %f163, %f767;
|
|
add.f32 %f771, %f164, %f768;
|
|
add.f32 %f772, %f165, %f769;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs129, %f772;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs128, %f771;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs127, %f770;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd224], {%rs127, %rs128, %rs129, %rs123};
|
|
bra.uni BB0_131;
|
|
|
|
BB0_130:
|
|
mov.u64 %rd237, image_RNM6;
|
|
cvta.global.u64 %rd232, %rd237;
|
|
mov.u32 %r332, 8;
|
|
// inline asm
|
|
call (%rd231), _rt_buffer_get_64, (%rd232, %r99, %r332, %rd17, %rd18, %rd26, %rd26);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs136, %f165;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs135, %f164;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs134, %f163;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd231], {%rs134, %rs135, %rs136, %rs123};
|
|
|
|
BB0_131:
|
|
ld.global.f32 %f777, [directDir+8];
|
|
ld.global.f32 %f778, [directDir];
|
|
mul.f32 %f779, %f778, %f777;
|
|
mul.f32 %f780, %f779, 0f3F8BD8A1;
|
|
mul.f32 %f166, %f142, %f780;
|
|
mul.f32 %f167, %f144, %f780;
|
|
mul.f32 %f168, %f146, %f780;
|
|
ld.global.u32 %r333, [additive];
|
|
setp.eq.s32 %p119, %r333, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs137, %f689;}
|
|
|
|
// inline asm
|
|
@%p119 bra BB0_133;
|
|
|
|
mov.u64 %rd250, image_RNM7;
|
|
cvta.global.u64 %rd239, %rd250;
|
|
mov.u32 %r337, 8;
|
|
// inline asm
|
|
call (%rd238), _rt_buffer_get_64, (%rd239, %r99, %r337, %rd17, %rd18, %rd26, %rd26);
|
|
// inline asm
|
|
ld.v4.u16 {%rs144, %rs145, %rs146, %rs147}, [%rd238];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f781, %rs144;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f782, %rs145;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f783, %rs146;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd244), _rt_buffer_get_64, (%rd239, %r99, %r337, %rd17, %rd18, %rd26, %rd26);
|
|
// inline asm
|
|
add.f32 %f784, %f166, %f781;
|
|
add.f32 %f785, %f167, %f782;
|
|
add.f32 %f786, %f168, %f783;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs143, %f786;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs142, %f785;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs141, %f784;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd244], {%rs141, %rs142, %rs143, %rs137};
|
|
bra.uni BB0_134;
|
|
|
|
BB0_133:
|
|
mov.u64 %rd257, image_RNM7;
|
|
cvta.global.u64 %rd252, %rd257;
|
|
mov.u32 %r339, 8;
|
|
// inline asm
|
|
call (%rd251), _rt_buffer_get_64, (%rd252, %r99, %r339, %rd17, %rd18, %rd26, %rd26);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs150, %f168;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs149, %f167;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs148, %f166;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd251], {%rs148, %rs149, %rs150, %rs137};
|
|
|
|
BB0_134:
|
|
ld.global.f32 %f791, [directDir];
|
|
mul.f32 %f792, %f791, %f791;
|
|
ld.global.f32 %f793, [directDir+4];
|
|
mul.f32 %f794, %f793, %f793;
|
|
sub.f32 %f795, %f792, %f794;
|
|
mul.f32 %f796, %f795, 0f3F0BD8A1;
|
|
mul.f32 %f169, %f142, %f796;
|
|
mul.f32 %f170, %f144, %f796;
|
|
mul.f32 %f171, %f146, %f796;
|
|
ld.global.u32 %r340, [additive];
|
|
setp.eq.s32 %p120, %r340, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs151, %f689;}
|
|
|
|
// inline asm
|
|
@%p120 bra BB0_136;
|
|
|
|
mov.u64 %rd270, image_RNM8;
|
|
cvta.global.u64 %rd259, %rd270;
|
|
mov.u32 %r344, 8;
|
|
// inline asm
|
|
call (%rd258), _rt_buffer_get_64, (%rd259, %r99, %r344, %rd17, %rd18, %rd26, %rd26);
|
|
// inline asm
|
|
ld.v4.u16 {%rs158, %rs159, %rs160, %rs161}, [%rd258];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f797, %rs158;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f798, %rs159;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f799, %rs160;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd264), _rt_buffer_get_64, (%rd259, %r99, %r344, %rd17, %rd18, %rd26, %rd26);
|
|
// inline asm
|
|
add.f32 %f800, %f169, %f797;
|
|
add.f32 %f801, %f170, %f798;
|
|
add.f32 %f802, %f171, %f799;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs157, %f802;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs156, %f801;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs155, %f800;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd264], {%rs155, %rs156, %rs157, %rs151};
|
|
bra.uni BB0_172;
|
|
|
|
BB0_136:
|
|
mov.u64 %rd277, image_RNM8;
|
|
cvta.global.u64 %rd272, %rd277;
|
|
mov.u32 %r346, 8;
|
|
// inline asm
|
|
call (%rd271), _rt_buffer_get_64, (%rd272, %r99, %r346, %rd17, %rd18, %rd26, %rd26);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs164, %f171;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs163, %f170;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs162, %f169;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd271], {%rs162, %rs163, %rs164, %rs151};
|
|
|
|
BB0_172:
|
|
ret;
|
|
}
|
|
|
|
|