2820 lines
74 KiB
Plaintext
2820 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_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 .texref sky;
|
|
.global .align 4 .b8 skyColor[12];
|
|
.global .align 4 .u32 samples;
|
|
.global .align 4 .u32 hemispherical;
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo7pixelIDE[8] = {82, 97, 121, 0, 8, 0, 0, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo10resolutionE[8] = {82, 97, 121, 0, 8, 0, 0, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo6normalE[8] = {82, 97, 121, 0, 12, 0, 0, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo6camPosE[8] = {82, 97, 121, 0, 12, 0, 0, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo4rootE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo12imageEnabledE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo8tileInfoE[8] = {82, 97, 121, 0, 16, 0, 0, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo8additiveE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo8skyColorE[8] = {82, 97, 121, 0, 12, 0, 0, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo7samplesE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo13hemisphericalE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
|
|
.global .align 8 .u64 _ZN21rti_internal_register20reg_bitness_detectorE;
|
|
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail0E;
|
|
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail1E;
|
|
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail2E;
|
|
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail3E;
|
|
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail4E;
|
|
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail5E;
|
|
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail6E;
|
|
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail7E;
|
|
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail8E;
|
|
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail9E;
|
|
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail0E;
|
|
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail1E;
|
|
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail2E;
|
|
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail3E;
|
|
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail4E;
|
|
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail5E;
|
|
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail6E;
|
|
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail7E;
|
|
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail8E;
|
|
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail9E;
|
|
.global .align 4 .u32 _ZN21rti_internal_register14reg_rayIndex_xE;
|
|
.global .align 4 .u32 _ZN21rti_internal_register14reg_rayIndex_yE;
|
|
.global .align 4 .u32 _ZN21rti_internal_register14reg_rayIndex_zE;
|
|
.global .align 8 .b8 _ZN21rti_internal_typename7pixelIDE[6] = {117, 105, 110, 116, 50, 0};
|
|
.global .align 8 .b8 _ZN21rti_internal_typename10resolutionE[6] = {117, 105, 110, 116, 50, 0};
|
|
.global .align 8 .b8 _ZN21rti_internal_typename6normalE[7] = {102, 108, 111, 97, 116, 51, 0};
|
|
.global .align 8 .b8 _ZN21rti_internal_typename6camPosE[7] = {102, 108, 111, 97, 116, 51, 0};
|
|
.global .align 16 .b8 _ZN21rti_internal_typename4rootE[9] = {114, 116, 79, 98, 106, 101, 99, 116, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typename12imageEnabledE[4] = {105, 110, 116, 0};
|
|
.global .align 8 .b8 _ZN21rti_internal_typename8tileInfoE[6] = {117, 105, 110, 116, 52, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typename8additiveE[4] = {105, 110, 116, 0};
|
|
.global .align 8 .b8 _ZN21rti_internal_typename8skyColorE[7] = {102, 108, 111, 97, 116, 51, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typename7samplesE[4] = {105, 110, 116, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typename13hemisphericalE[4] = {105, 110, 116, 0};
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum7pixelIDE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum10resolutionE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum6normalE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum6camPosE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum4rootE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum12imageEnabledE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum8tileInfoE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum8additiveE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum8skyColorE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum7samplesE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum13hemisphericalE = 4919;
|
|
.global .align 16 .b8 _ZN21rti_internal_semantic7pixelIDE[14] = {114, 116, 76, 97, 117, 110, 99, 104, 73, 110, 100, 101, 120, 0};
|
|
.global .align 16 .b8 _ZN21rti_internal_semantic10resolutionE[12] = {114, 116, 76, 97, 117, 110, 99, 104, 68, 105, 109, 0};
|
|
.global .align 16 .b8 _ZN21rti_internal_semantic6normalE[17] = {97, 116, 116, 114, 105, 98, 117, 116, 101, 32, 110, 111, 114, 109, 97, 108, 0};
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic6camPosE[1];
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic4rootE[1];
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic12imageEnabledE[1];
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic8tileInfoE[1];
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic8additiveE[1];
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic8skyColorE[1];
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic7samplesE[1];
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic13hemisphericalE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation7pixelIDE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation10resolutionE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation6normalE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation6camPosE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation4rootE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation12imageEnabledE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation8tileInfoE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation8additiveE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation8skyColorE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation7samplesE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation13hemisphericalE[1];
|
|
.const .align 4 .b8 __cudart_i2opi_f[24] = {65, 144, 67, 60, 153, 149, 98, 219, 192, 221, 52, 245, 209, 87, 39, 252, 41, 21, 68, 78, 110, 131, 249, 162};
|
|
|
|
.visible .entry _Z6oxMainv(
|
|
|
|
)
|
|
{
|
|
.local .align 4 .b8 __local_depot0[40];
|
|
.reg .b64 %SP;
|
|
.reg .b64 %SPL;
|
|
.reg .pred %p<116>;
|
|
.reg .b16 %rs<294>;
|
|
.reg .f32 %f<1187>;
|
|
.reg .b32 %r<452>;
|
|
.reg .b64 %rd<488>;
|
|
|
|
|
|
mov.u64 %rd487, __local_depot0;
|
|
cvta.local.u64 %SP, %rd487;
|
|
ld.global.u32 %r1, [samples];
|
|
shl.b32 %r2, %r1, 1;
|
|
ld.global.v2.u32 {%r94, %r95}, [pixelID];
|
|
cvt.u64.u32 %rd22, %r94;
|
|
cvt.u64.u32 %rd23, %r95;
|
|
mov.u64 %rd26, uvnormal;
|
|
cvta.global.u64 %rd21, %rd26;
|
|
mov.u32 %r92, 2;
|
|
mov.u32 %r93, 4;
|
|
mov.u64 %rd25, 0;
|
|
// inline asm
|
|
call (%rd20), _rt_buffer_get_64, (%rd21, %r92, %r93, %rd22, %rd23, %rd25, %rd25);
|
|
// inline asm
|
|
ld.u32 %r3, [%rd20];
|
|
shr.u32 %r98, %r3, 16;
|
|
cvt.u16.u32 %rs1, %r98;
|
|
and.b16 %rs12, %rs1, 255;
|
|
cvt.u16.u32 %rs13, %r3;
|
|
or.b16 %rs14, %rs13, %rs12;
|
|
setp.eq.s16 %p4, %rs14, 0;
|
|
mov.f32 %f1043, 0f00000000;
|
|
mov.f32 %f1044, %f1043;
|
|
mov.f32 %f1045, %f1043;
|
|
@%p4 bra BB0_2;
|
|
|
|
ld.u8 %rs15, [%rd20+1];
|
|
and.b16 %rs17, %rs13, 255;
|
|
cvt.rn.f32.u16 %f311, %rs17;
|
|
div.rn.f32 %f312, %f311, 0f437F0000;
|
|
fma.rn.f32 %f313, %f312, 0f40000000, 0fBF800000;
|
|
cvt.rn.f32.u16 %f314, %rs15;
|
|
div.rn.f32 %f315, %f314, 0f437F0000;
|
|
fma.rn.f32 %f316, %f315, 0f40000000, 0fBF800000;
|
|
cvt.rn.f32.u16 %f317, %rs12;
|
|
div.rn.f32 %f318, %f317, 0f437F0000;
|
|
fma.rn.f32 %f319, %f318, 0f40000000, 0fBF800000;
|
|
mul.f32 %f320, %f316, %f316;
|
|
fma.rn.f32 %f321, %f313, %f313, %f320;
|
|
fma.rn.f32 %f322, %f319, %f319, %f321;
|
|
sqrt.rn.f32 %f323, %f322;
|
|
rcp.rn.f32 %f324, %f323;
|
|
mul.f32 %f1043, %f313, %f324;
|
|
mul.f32 %f1044, %f316, %f324;
|
|
mul.f32 %f1045, %f319, %f324;
|
|
|
|
BB0_2:
|
|
ld.global.v2.u32 {%r99, %r100}, [pixelID];
|
|
ld.global.v2.u32 {%r102, %r103}, [tileInfo];
|
|
add.s32 %r4, %r99, %r102;
|
|
add.s32 %r5, %r100, %r103;
|
|
setp.eq.f32 %p5, %f1044, 0f00000000;
|
|
setp.eq.f32 %p6, %f1043, 0f00000000;
|
|
and.pred %p7, %p6, %p5;
|
|
setp.eq.f32 %p8, %f1045, 0f00000000;
|
|
and.pred %p9, %p7, %p8;
|
|
@%p9 bra BB0_122;
|
|
bra.uni BB0_3;
|
|
|
|
BB0_122:
|
|
ld.global.u32 %r451, [imageEnabled];
|
|
and.b32 %r325, %r451, 1;
|
|
setp.eq.b32 %p103, %r325, 1;
|
|
@!%p103 bra BB0_124;
|
|
bra.uni BB0_123;
|
|
|
|
BB0_123:
|
|
cvt.u64.u32 %rd273, %r5;
|
|
cvt.u64.u32 %rd272, %r4;
|
|
mov.u64 %rd276, image;
|
|
cvta.global.u64 %rd271, %rd276;
|
|
mov.u64 %rd275, 0;
|
|
// inline asm
|
|
call (%rd270), _rt_buffer_get_64, (%rd271, %r92, %r93, %rd272, %rd273, %rd275, %rd275);
|
|
// inline asm
|
|
mov.u16 %rs163, 0;
|
|
st.v4.u8 [%rd270], {%rs163, %rs163, %rs163, %rs163};
|
|
ld.global.u32 %r451, [imageEnabled];
|
|
|
|
BB0_124:
|
|
cvt.u64.u32 %rd18, %r4;
|
|
cvt.u64.u32 %rd19, %r5;
|
|
and.b32 %r328, %r451, 4;
|
|
setp.eq.s32 %p104, %r328, 0;
|
|
@%p104 bra BB0_128;
|
|
|
|
ld.global.u32 %r329, [additive];
|
|
setp.eq.s32 %p105, %r329, 0;
|
|
@%p105 bra BB0_127;
|
|
|
|
mov.u64 %rd289, image_HDR;
|
|
cvta.global.u64 %rd278, %rd289;
|
|
mov.u32 %r333, 8;
|
|
mov.u64 %rd288, 0;
|
|
// inline asm
|
|
call (%rd277), _rt_buffer_get_64, (%rd278, %r92, %r333, %rd18, %rd19, %rd288, %rd288);
|
|
// inline asm
|
|
ld.v4.u16 {%rs170, %rs171, %rs172, %rs173}, [%rd277];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f951, %rs170;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f952, %rs171;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f953, %rs172;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd283), _rt_buffer_get_64, (%rd278, %r92, %r333, %rd18, %rd19, %rd288, %rd288);
|
|
// inline asm
|
|
add.f32 %f954, %f951, 0f00000000;
|
|
add.f32 %f955, %f952, 0f00000000;
|
|
add.f32 %f956, %f953, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs169, %f956;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs168, %f955;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs167, %f954;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs174, 0;
|
|
st.v4.u16 [%rd283], {%rs167, %rs168, %rs169, %rs174};
|
|
bra.uni BB0_128;
|
|
|
|
BB0_3:
|
|
ld.global.v2.u32 {%r111, %r112}, [pixelID];
|
|
cvt.u64.u32 %rd29, %r111;
|
|
cvt.u64.u32 %rd30, %r112;
|
|
mov.u64 %rd39, uvpos;
|
|
cvta.global.u64 %rd28, %rd39;
|
|
mov.u32 %r108, 12;
|
|
// inline asm
|
|
call (%rd27), _rt_buffer_get_64, (%rd28, %r92, %r108, %rd29, %rd30, %rd25, %rd25);
|
|
// inline asm
|
|
ld.f32 %f355, [%rd27+8];
|
|
ld.f32 %f356, [%rd27+4];
|
|
ld.f32 %f357, [%rd27];
|
|
mul.f32 %f358, %f357, 0f3456BF95;
|
|
mul.f32 %f359, %f356, 0f3456BF95;
|
|
mul.f32 %f360, %f355, 0f3456BF95;
|
|
abs.f32 %f361, %f1043;
|
|
div.rn.f32 %f362, %f358, %f361;
|
|
abs.f32 %f363, %f1044;
|
|
div.rn.f32 %f364, %f359, %f363;
|
|
abs.f32 %f365, %f1045;
|
|
div.rn.f32 %f366, %f360, %f365;
|
|
abs.f32 %f367, %f362;
|
|
abs.f32 %f368, %f364;
|
|
abs.f32 %f369, %f366;
|
|
mov.f32 %f370, 0f38D1B717;
|
|
max.f32 %f371, %f367, %f370;
|
|
max.f32 %f372, %f368, %f370;
|
|
max.f32 %f373, %f369, %f370;
|
|
fma.rn.f32 %f7, %f1043, %f371, %f357;
|
|
fma.rn.f32 %f8, %f1044, %f372, %f356;
|
|
fma.rn.f32 %f9, %f1045, %f373, %f355;
|
|
ld.global.u32 %r6, [hemispherical];
|
|
setp.gt.f32 %p10, %f361, %f365;
|
|
neg.f32 %f374, %f1044;
|
|
selp.f32 %f375, %f374, 0f00000000, %p10;
|
|
neg.f32 %f376, %f1045;
|
|
selp.f32 %f377, %f1043, %f376, %p10;
|
|
selp.f32 %f378, 0f00000000, %f1044, %p10;
|
|
mul.f32 %f379, %f377, %f377;
|
|
fma.rn.f32 %f380, %f375, %f375, %f379;
|
|
fma.rn.f32 %f381, %f378, %f378, %f380;
|
|
sqrt.rn.f32 %f382, %f381;
|
|
rcp.rn.f32 %f383, %f382;
|
|
mul.f32 %f10, %f375, %f383;
|
|
mul.f32 %f11, %f377, %f383;
|
|
mul.f32 %f12, %f378, %f383;
|
|
ld.global.v2.u32 {%r115, %r116}, [pixelID];
|
|
cvt.u64.u32 %rd35, %r115;
|
|
cvt.u64.u32 %rd36, %r116;
|
|
mov.u64 %rd40, rnd_seeds;
|
|
cvta.global.u64 %rd34, %rd40;
|
|
// inline asm
|
|
call (%rd33), _rt_buffer_get_64, (%rd34, %r92, %r93, %rd35, %rd36, %rd25, %rd25);
|
|
// inline asm
|
|
mov.f32 %f1147, 0f00000000;
|
|
setp.lt.s32 %p11, %r1, 1;
|
|
mov.f32 %f1146, %f1147;
|
|
mov.f32 %f1145, %f1147;
|
|
mov.f32 %f1144, %f1147;
|
|
mov.f32 %f1143, %f1147;
|
|
mov.f32 %f1142, %f1147;
|
|
mov.f32 %f1141, %f1147;
|
|
mov.f32 %f1140, %f1147;
|
|
mov.f32 %f1139, %f1147;
|
|
mov.f32 %f1138, %f1147;
|
|
mov.f32 %f1137, %f1147;
|
|
mov.f32 %f1136, %f1147;
|
|
mov.f32 %f1135, %f1147;
|
|
mov.f32 %f1134, %f1147;
|
|
mov.f32 %f1133, %f1147;
|
|
mov.f32 %f1132, %f1147;
|
|
mov.f32 %f1131, %f1147;
|
|
mov.f32 %f1130, %f1147;
|
|
mov.f32 %f1129, %f1147;
|
|
mov.f32 %f1128, %f1147;
|
|
mov.f32 %f1127, %f1147;
|
|
mov.f32 %f1126, %f1147;
|
|
mov.f32 %f1125, %f1147;
|
|
mov.f32 %f1124, %f1147;
|
|
mov.f32 %f1123, %f1147;
|
|
mov.f32 %f1122, %f1147;
|
|
mov.f32 %f1121, %f1147;
|
|
mov.f32 %f1120, %f1147;
|
|
mov.f32 %f1119, %f1147;
|
|
mov.f32 %f1118, %f1147;
|
|
@%p11 bra BB0_56;
|
|
|
|
cvt.rn.f32.s32 %f414, %r2;
|
|
rcp.rn.f32 %f13, %f414;
|
|
ld.u32 %r429, [%rd33];
|
|
mul.f32 %f14, %f7, 0f3456BF95;
|
|
mul.f32 %f15, %f8, 0f3456BF95;
|
|
mul.f32 %f16, %f9, 0f3456BF95;
|
|
mul.f32 %f415, %f1043, %f11;
|
|
mul.f32 %f416, %f1044, %f10;
|
|
sub.f32 %f17, %f416, %f415;
|
|
mul.f32 %f417, %f1045, %f10;
|
|
mul.f32 %f418, %f1043, %f12;
|
|
sub.f32 %f18, %f418, %f417;
|
|
mul.f32 %f419, %f1044, %f12;
|
|
mul.f32 %f420, %f1045, %f11;
|
|
sub.f32 %f19, %f420, %f419;
|
|
mov.f32 %f1147, 0f00000000;
|
|
mov.u32 %r119, 0;
|
|
abs.f32 %f495, %f15;
|
|
abs.f32 %f496, %f14;
|
|
max.f32 %f497, %f496, %f495;
|
|
abs.f32 %f498, %f16;
|
|
max.f32 %f499, %f497, %f498;
|
|
mov.u32 %r426, %r119;
|
|
mov.f32 %f1146, %f1147;
|
|
mov.f32 %f1145, %f1147;
|
|
mov.f32 %f1144, %f1147;
|
|
mov.f32 %f1143, %f1147;
|
|
mov.f32 %f1142, %f1147;
|
|
mov.f32 %f1141, %f1147;
|
|
mov.f32 %f1140, %f1147;
|
|
mov.f32 %f1139, %f1147;
|
|
mov.f32 %f1138, %f1147;
|
|
mov.f32 %f1137, %f1147;
|
|
mov.f32 %f1136, %f1147;
|
|
mov.f32 %f1135, %f1147;
|
|
mov.f32 %f1134, %f1147;
|
|
mov.f32 %f1133, %f1147;
|
|
mov.f32 %f1132, %f1147;
|
|
mov.f32 %f1131, %f1147;
|
|
mov.f32 %f1130, %f1147;
|
|
mov.f32 %f1129, %f1147;
|
|
mov.f32 %f1128, %f1147;
|
|
mov.f32 %f1127, %f1147;
|
|
mov.f32 %f1126, %f1147;
|
|
mov.f32 %f1125, %f1147;
|
|
mov.f32 %f1124, %f1147;
|
|
mov.f32 %f1123, %f1147;
|
|
mov.f32 %f1122, %f1147;
|
|
mov.f32 %f1121, %f1147;
|
|
mov.f32 %f1120, %f1147;
|
|
mov.f32 %f1119, %f1147;
|
|
mov.f32 %f1118, %f1147;
|
|
|
|
BB0_5:
|
|
mov.u32 %r428, %r119;
|
|
|
|
BB0_6:
|
|
mov.u32 %r11, %r429;
|
|
cvt.rn.f32.s32 %f1027, %r426;
|
|
mad.lo.s32 %r121, %r11, 1664525, 1013904223;
|
|
and.b32 %r122, %r121, 16777215;
|
|
cvt.rn.f32.u32 %f421, %r122;
|
|
fma.rn.f32 %f422, %f421, 0f33800000, %f1027;
|
|
mul.f32 %f423, %f13, %f422;
|
|
mad.lo.s32 %r12, %r121, 1664525, 1013904223;
|
|
and.b32 %r123, %r12, 16777215;
|
|
cvt.rn.f32.u32 %f424, %r123;
|
|
cvt.rn.f32.s32 %f425, %r428;
|
|
fma.rn.f32 %f426, %f424, 0f33800000, %f425;
|
|
mul.f32 %f427, %f13, %f426;
|
|
fma.rn.f32 %f81, %f423, 0fC0000000, 0f3F800000;
|
|
mul.f32 %f428, %f81, %f81;
|
|
mov.f32 %f429, 0f3F800000;
|
|
sub.f32 %f430, %f429, %f428;
|
|
mov.f32 %f431, 0f00000000;
|
|
max.f32 %f432, %f431, %f430;
|
|
sqrt.rn.f32 %f82, %f432;
|
|
mul.f32 %f1112, %f427, 0f40C90FDB;
|
|
abs.f32 %f84, %f1112;
|
|
setp.neu.f32 %p12, %f84, 0f7F800000;
|
|
mov.f32 %f1106, %f1112;
|
|
@%p12 bra BB0_8;
|
|
|
|
mul.rn.f32 %f1106, %f1112, %f431;
|
|
|
|
BB0_8:
|
|
mul.f32 %f434, %f1106, 0f3F22F983;
|
|
cvt.rni.s32.f32 %r439, %f434;
|
|
cvt.rn.f32.s32 %f435, %r439;
|
|
neg.f32 %f436, %f435;
|
|
mov.f32 %f437, 0f3FC90FDA;
|
|
fma.rn.f32 %f438, %f436, %f437, %f1106;
|
|
mov.f32 %f439, 0f33A22168;
|
|
fma.rn.f32 %f440, %f436, %f439, %f438;
|
|
mov.f32 %f441, 0f27C234C5;
|
|
fma.rn.f32 %f1107, %f436, %f441, %f440;
|
|
abs.f32 %f442, %f1106;
|
|
setp.leu.f32 %p13, %f442, 0f47CE4780;
|
|
@%p13 bra BB0_19;
|
|
|
|
add.u64 %rd42, %SP, 12;
|
|
cvta.to.local.u64 %rd483, %rd42;
|
|
mov.u32 %r430, 0;
|
|
mov.u64 %rd484, 0;
|
|
mov.u32 %r431, %r430;
|
|
|
|
BB0_10:
|
|
.pragma "nounroll";
|
|
mov.b32 %r401, %f1106;
|
|
shl.b32 %r400, %r401, 8;
|
|
or.b32 %r399, %r400, -2147483648;
|
|
add.u64 %rd478, %SP, 12;
|
|
cvta.to.local.u64 %rd477, %rd478;
|
|
shl.b64 %rd43, %rd484, 2;
|
|
mov.u64 %rd44, __cudart_i2opi_f;
|
|
add.s64 %rd45, %rd44, %rd43;
|
|
ld.const.u32 %r129, [%rd45];
|
|
// inline asm
|
|
{
|
|
mad.lo.cc.u32 %r127, %r129, %r399, %r431;
|
|
madc.hi.u32 %r431, %r129, %r399, 0;
|
|
}
|
|
// inline asm
|
|
st.local.u32 [%rd483], %r127;
|
|
add.s32 %r430, %r430, 1;
|
|
cvt.s64.s32 %rd484, %r430;
|
|
mul.wide.s32 %rd48, %r430, 4;
|
|
add.s64 %rd483, %rd477, %rd48;
|
|
setp.ne.s32 %p14, %r430, 6;
|
|
@%p14 bra BB0_10;
|
|
|
|
mov.b32 %r403, %f1106;
|
|
shr.u32 %r402, %r403, 23;
|
|
add.u64 %rd479, %SP, 12;
|
|
and.b32 %r132, %r402, 255;
|
|
add.s32 %r133, %r132, -128;
|
|
shr.u32 %r134, %r133, 5;
|
|
cvta.to.local.u64 %rd50, %rd479;
|
|
st.local.u32 [%rd50+24], %r431;
|
|
mov.u32 %r135, 6;
|
|
sub.s32 %r136, %r135, %r134;
|
|
mul.wide.s32 %rd51, %r136, 4;
|
|
add.s64 %rd8, %rd50, %rd51;
|
|
ld.local.u32 %r432, [%rd8];
|
|
ld.local.u32 %r433, [%rd8+-4];
|
|
and.b32 %r24, %r402, 31;
|
|
setp.eq.s32 %p15, %r24, 0;
|
|
@%p15 bra BB0_13;
|
|
|
|
mov.u32 %r137, 32;
|
|
sub.s32 %r138, %r137, %r24;
|
|
shr.u32 %r139, %r433, %r138;
|
|
shl.b32 %r140, %r432, %r24;
|
|
add.s32 %r432, %r139, %r140;
|
|
ld.local.u32 %r141, [%rd8+-8];
|
|
shr.u32 %r142, %r141, %r138;
|
|
shl.b32 %r143, %r433, %r24;
|
|
add.s32 %r433, %r142, %r143;
|
|
|
|
BB0_13:
|
|
mov.b32 %r411, %f1106;
|
|
and.b32 %r435, %r411, -2147483648;
|
|
shr.u32 %r144, %r433, 30;
|
|
shl.b32 %r145, %r432, 2;
|
|
add.s32 %r434, %r144, %r145;
|
|
shl.b32 %r30, %r433, 2;
|
|
shr.u32 %r146, %r434, 31;
|
|
shr.u32 %r147, %r432, 30;
|
|
add.s32 %r31, %r146, %r147;
|
|
setp.eq.s32 %p16, %r146, 0;
|
|
@%p16 bra BB0_14;
|
|
bra.uni BB0_15;
|
|
|
|
BB0_14:
|
|
mov.u32 %r436, %r30;
|
|
bra.uni BB0_16;
|
|
|
|
BB0_15:
|
|
mov.b32 %r413, %f1106;
|
|
and.b32 %r412, %r413, -2147483648;
|
|
not.b32 %r148, %r434;
|
|
neg.s32 %r436, %r30;
|
|
setp.eq.s32 %p17, %r30, 0;
|
|
selp.u32 %r149, 1, 0, %p17;
|
|
add.s32 %r434, %r149, %r148;
|
|
xor.b32 %r435, %r412, -2147483648;
|
|
|
|
BB0_16:
|
|
mov.b32 %r415, %f1106;
|
|
and.b32 %r414, %r415, -2147483648;
|
|
clz.b32 %r438, %r434;
|
|
setp.eq.s32 %p18, %r438, 0;
|
|
shl.b32 %r150, %r434, %r438;
|
|
mov.u32 %r151, 32;
|
|
sub.s32 %r152, %r151, %r438;
|
|
shr.u32 %r153, %r436, %r152;
|
|
add.s32 %r154, %r153, %r150;
|
|
selp.b32 %r39, %r434, %r154, %p18;
|
|
mov.u32 %r155, -921707870;
|
|
mul.hi.u32 %r437, %r39, %r155;
|
|
setp.eq.s32 %p19, %r414, 0;
|
|
neg.s32 %r156, %r31;
|
|
selp.b32 %r439, %r31, %r156, %p19;
|
|
setp.lt.s32 %p20, %r437, 1;
|
|
@%p20 bra BB0_18;
|
|
|
|
mul.lo.s32 %r157, %r39, -921707870;
|
|
shr.u32 %r158, %r157, 31;
|
|
shl.b32 %r159, %r437, 1;
|
|
add.s32 %r437, %r158, %r159;
|
|
add.s32 %r438, %r438, 1;
|
|
|
|
BB0_18:
|
|
mov.u32 %r160, 126;
|
|
sub.s32 %r161, %r160, %r438;
|
|
shl.b32 %r162, %r161, 23;
|
|
add.s32 %r163, %r437, 1;
|
|
shr.u32 %r164, %r163, 7;
|
|
add.s32 %r165, %r164, 1;
|
|
shr.u32 %r166, %r165, 1;
|
|
add.s32 %r167, %r166, %r162;
|
|
or.b32 %r168, %r167, %r435;
|
|
mov.b32 %f1107, %r168;
|
|
|
|
BB0_19:
|
|
add.s32 %r47, %r439, 1;
|
|
and.b32 %r48, %r47, 1;
|
|
setp.eq.s32 %p21, %r48, 0;
|
|
@%p21 bra BB0_21;
|
|
bra.uni BB0_20;
|
|
|
|
BB0_21:
|
|
mul.rn.f32 %f1037, %f1107, %f1107;
|
|
mov.f32 %f445, 0f3C08839E;
|
|
mov.f32 %f446, 0fB94CA1F9;
|
|
fma.rn.f32 %f1108, %f446, %f1037, %f445;
|
|
bra.uni BB0_22;
|
|
|
|
BB0_20:
|
|
mul.rn.f32 %f1033, %f1107, %f1107;
|
|
mov.f32 %f443, 0fBAB6061A;
|
|
mov.f32 %f444, 0f37CCF5CE;
|
|
fma.rn.f32 %f1108, %f444, %f1033, %f443;
|
|
|
|
BB0_22:
|
|
@%p21 bra BB0_24;
|
|
bra.uni BB0_23;
|
|
|
|
BB0_24:
|
|
mul.rn.f32 %f1036, %f1107, %f1107;
|
|
mov.f32 %f1032, 0f00000000;
|
|
mov.f32 %f450, 0fBE2AAAA3;
|
|
fma.rn.f32 %f451, %f1108, %f1036, %f450;
|
|
fma.rn.f32 %f1109, %f451, %f1036, %f1032;
|
|
bra.uni BB0_25;
|
|
|
|
BB0_23:
|
|
mul.rn.f32 %f1034, %f1107, %f1107;
|
|
mov.f32 %f447, 0f3D2AAAA5;
|
|
fma.rn.f32 %f448, %f1108, %f1034, %f447;
|
|
mov.f32 %f449, 0fBF000000;
|
|
fma.rn.f32 %f1109, %f448, %f1034, %f449;
|
|
|
|
BB0_25:
|
|
fma.rn.f32 %f1110, %f1109, %f1107, %f1107;
|
|
@%p21 bra BB0_27;
|
|
|
|
mul.rn.f32 %f1035, %f1107, %f1107;
|
|
mov.f32 %f1021, 0f3F800000;
|
|
fma.rn.f32 %f1110, %f1109, %f1035, %f1021;
|
|
|
|
BB0_27:
|
|
add.s32 %r416, %r439, 1;
|
|
and.b32 %r169, %r416, 2;
|
|
setp.eq.s32 %p24, %r169, 0;
|
|
@%p24 bra BB0_29;
|
|
|
|
mov.f32 %f1028, 0f00000000;
|
|
mov.f32 %f455, 0fBF800000;
|
|
fma.rn.f32 %f1110, %f1110, %f455, %f1028;
|
|
|
|
BB0_29:
|
|
abs.f32 %f1022, %f1112;
|
|
setp.neu.f32 %p115, %f1022, 0f7F800000;
|
|
@%p115 bra BB0_31;
|
|
|
|
mov.f32 %f1031, 0f00000000;
|
|
mul.rn.f32 %f1112, %f1112, %f1031;
|
|
|
|
BB0_31:
|
|
mov.f32 %f1025, 0f27C234C5;
|
|
mov.f32 %f1024, 0f33A22168;
|
|
mov.f32 %f1023, 0f3FC90FDA;
|
|
mul.f32 %f457, %f1112, 0f3F22F983;
|
|
cvt.rni.s32.f32 %r449, %f457;
|
|
cvt.rn.f32.s32 %f458, %r449;
|
|
neg.f32 %f459, %f458;
|
|
fma.rn.f32 %f461, %f459, %f1023, %f1112;
|
|
fma.rn.f32 %f463, %f459, %f1024, %f461;
|
|
fma.rn.f32 %f1113, %f459, %f1025, %f463;
|
|
abs.f32 %f465, %f1112;
|
|
setp.leu.f32 %p26, %f465, 0f47CE4780;
|
|
@%p26 bra BB0_42;
|
|
|
|
add.u64 %rd53, %SP, 12;
|
|
cvta.to.local.u64 %rd485, %rd53;
|
|
mov.b32 %r50, %f1112;
|
|
shl.b32 %r172, %r50, 8;
|
|
or.b32 %r52, %r172, -2147483648;
|
|
mov.u32 %r440, 0;
|
|
mov.u64 %rd486, %rd25;
|
|
mov.u32 %r441, %r440;
|
|
|
|
BB0_33:
|
|
.pragma "nounroll";
|
|
add.u64 %rd481, %SP, 12;
|
|
cvta.to.local.u64 %rd480, %rd481;
|
|
shl.b64 %rd54, %rd486, 2;
|
|
mov.u64 %rd55, __cudart_i2opi_f;
|
|
add.s64 %rd56, %rd55, %rd54;
|
|
ld.const.u32 %r175, [%rd56];
|
|
// inline asm
|
|
{
|
|
mad.lo.cc.u32 %r173, %r175, %r52, %r441;
|
|
madc.hi.u32 %r441, %r175, %r52, 0;
|
|
}
|
|
// inline asm
|
|
st.local.u32 [%rd485], %r173;
|
|
add.s32 %r440, %r440, 1;
|
|
cvt.s64.s32 %rd486, %r440;
|
|
mul.wide.s32 %rd57, %r440, 4;
|
|
add.s64 %rd485, %rd480, %rd57;
|
|
setp.ne.s32 %p27, %r440, 6;
|
|
@%p27 bra BB0_33;
|
|
|
|
mov.b32 %r418, %f1112;
|
|
shr.u32 %r417, %r418, 23;
|
|
add.u64 %rd482, %SP, 12;
|
|
and.b32 %r178, %r417, 255;
|
|
add.s32 %r179, %r178, -128;
|
|
shr.u32 %r180, %r179, 5;
|
|
cvta.to.local.u64 %rd59, %rd482;
|
|
st.local.u32 [%rd59+24], %r441;
|
|
mov.u32 %r181, 6;
|
|
sub.s32 %r182, %r181, %r180;
|
|
mul.wide.s32 %rd60, %r182, 4;
|
|
add.s64 %rd15, %rd59, %rd60;
|
|
ld.local.u32 %r442, [%rd15];
|
|
ld.local.u32 %r443, [%rd15+-4];
|
|
and.b32 %r60, %r417, 31;
|
|
setp.eq.s32 %p28, %r60, 0;
|
|
@%p28 bra BB0_36;
|
|
|
|
mov.u32 %r183, 32;
|
|
sub.s32 %r184, %r183, %r60;
|
|
shr.u32 %r185, %r443, %r184;
|
|
shl.b32 %r186, %r442, %r60;
|
|
add.s32 %r442, %r185, %r186;
|
|
ld.local.u32 %r187, [%rd15+-8];
|
|
shr.u32 %r188, %r187, %r184;
|
|
shl.b32 %r189, %r443, %r60;
|
|
add.s32 %r443, %r188, %r189;
|
|
|
|
BB0_36:
|
|
mov.b32 %r421, %f1112;
|
|
and.b32 %r445, %r421, -2147483648;
|
|
shr.u32 %r190, %r443, 30;
|
|
shl.b32 %r191, %r442, 2;
|
|
add.s32 %r444, %r190, %r191;
|
|
shl.b32 %r66, %r443, 2;
|
|
shr.u32 %r192, %r444, 31;
|
|
shr.u32 %r193, %r442, 30;
|
|
add.s32 %r67, %r192, %r193;
|
|
setp.eq.s32 %p29, %r192, 0;
|
|
@%p29 bra BB0_37;
|
|
bra.uni BB0_38;
|
|
|
|
BB0_37:
|
|
mov.u32 %r446, %r66;
|
|
bra.uni BB0_39;
|
|
|
|
BB0_38:
|
|
mov.b32 %r423, %f1112;
|
|
and.b32 %r422, %r423, -2147483648;
|
|
not.b32 %r194, %r444;
|
|
neg.s32 %r446, %r66;
|
|
setp.eq.s32 %p30, %r66, 0;
|
|
selp.u32 %r195, 1, 0, %p30;
|
|
add.s32 %r444, %r195, %r194;
|
|
xor.b32 %r445, %r422, -2147483648;
|
|
|
|
BB0_39:
|
|
mov.b32 %r425, %f1112;
|
|
and.b32 %r424, %r425, -2147483648;
|
|
clz.b32 %r448, %r444;
|
|
setp.eq.s32 %p31, %r448, 0;
|
|
shl.b32 %r196, %r444, %r448;
|
|
mov.u32 %r197, 32;
|
|
sub.s32 %r198, %r197, %r448;
|
|
shr.u32 %r199, %r446, %r198;
|
|
add.s32 %r200, %r199, %r196;
|
|
selp.b32 %r75, %r444, %r200, %p31;
|
|
mov.u32 %r201, -921707870;
|
|
mul.hi.u32 %r447, %r75, %r201;
|
|
setp.eq.s32 %p32, %r424, 0;
|
|
neg.s32 %r202, %r67;
|
|
selp.b32 %r449, %r67, %r202, %p32;
|
|
setp.lt.s32 %p33, %r447, 1;
|
|
@%p33 bra BB0_41;
|
|
|
|
mul.lo.s32 %r203, %r75, -921707870;
|
|
shr.u32 %r204, %r203, 31;
|
|
shl.b32 %r205, %r447, 1;
|
|
add.s32 %r447, %r204, %r205;
|
|
add.s32 %r448, %r448, 1;
|
|
|
|
BB0_41:
|
|
mov.u32 %r206, 126;
|
|
sub.s32 %r207, %r206, %r448;
|
|
shl.b32 %r208, %r207, 23;
|
|
add.s32 %r209, %r447, 1;
|
|
shr.u32 %r210, %r209, 7;
|
|
add.s32 %r211, %r210, 1;
|
|
shr.u32 %r212, %r211, 1;
|
|
add.s32 %r213, %r212, %r208;
|
|
or.b32 %r214, %r213, %r445;
|
|
mov.b32 %f1113, %r214;
|
|
|
|
BB0_42:
|
|
and.b32 %r83, %r449, 1;
|
|
setp.eq.s32 %p34, %r83, 0;
|
|
@%p34 bra BB0_44;
|
|
bra.uni BB0_43;
|
|
|
|
BB0_44:
|
|
mul.rn.f32 %f1042, %f1113, %f1113;
|
|
mov.f32 %f468, 0f3C08839E;
|
|
mov.f32 %f469, 0fB94CA1F9;
|
|
fma.rn.f32 %f1114, %f469, %f1042, %f468;
|
|
bra.uni BB0_45;
|
|
|
|
BB0_43:
|
|
mul.rn.f32 %f1038, %f1113, %f1113;
|
|
mov.f32 %f466, 0fBAB6061A;
|
|
mov.f32 %f467, 0f37CCF5CE;
|
|
fma.rn.f32 %f1114, %f467, %f1038, %f466;
|
|
|
|
BB0_45:
|
|
@%p34 bra BB0_47;
|
|
bra.uni BB0_46;
|
|
|
|
BB0_47:
|
|
mul.rn.f32 %f1041, %f1113, %f1113;
|
|
mov.f32 %f1030, 0f00000000;
|
|
mov.f32 %f473, 0fBE2AAAA3;
|
|
fma.rn.f32 %f474, %f1114, %f1041, %f473;
|
|
fma.rn.f32 %f1115, %f474, %f1041, %f1030;
|
|
bra.uni BB0_48;
|
|
|
|
BB0_46:
|
|
mul.rn.f32 %f1039, %f1113, %f1113;
|
|
mov.f32 %f470, 0f3D2AAAA5;
|
|
fma.rn.f32 %f471, %f1114, %f1039, %f470;
|
|
mov.f32 %f472, 0fBF000000;
|
|
fma.rn.f32 %f1115, %f471, %f1039, %f472;
|
|
|
|
BB0_48:
|
|
fma.rn.f32 %f1116, %f1115, %f1113, %f1113;
|
|
@%p34 bra BB0_50;
|
|
|
|
mul.rn.f32 %f1040, %f1113, %f1113;
|
|
mov.f32 %f1026, 0f3F800000;
|
|
fma.rn.f32 %f1116, %f1115, %f1040, %f1026;
|
|
|
|
BB0_50:
|
|
and.b32 %r215, %r449, 2;
|
|
setp.eq.s32 %p37, %r215, 0;
|
|
@%p37 bra BB0_52;
|
|
|
|
mov.f32 %f1029, 0f00000000;
|
|
mov.f32 %f478, 0fBF800000;
|
|
fma.rn.f32 %f1116, %f1116, %f478, %f1029;
|
|
|
|
BB0_52:
|
|
mul.f32 %f479, %f82, %f1110;
|
|
mul.f32 %f480, %f82, %f1116;
|
|
mul.f32 %f481, %f10, %f480;
|
|
mul.f32 %f482, %f11, %f480;
|
|
mul.f32 %f483, %f12, %f480;
|
|
fma.rn.f32 %f484, %f19, %f479, %f481;
|
|
fma.rn.f32 %f485, %f18, %f479, %f482;
|
|
fma.rn.f32 %f486, %f17, %f479, %f483;
|
|
fma.rn.f32 %f119, %f1043, %f81, %f484;
|
|
fma.rn.f32 %f120, %f1044, %f81, %f485;
|
|
fma.rn.f32 %f121, %f1045, %f81, %f486;
|
|
setp.gt.f32 %p38, %f120, 0f00000000;
|
|
setp.eq.s32 %p39, %r6, 0;
|
|
or.pred %p40, %p39, %p38;
|
|
@!%p40 bra BB0_54;
|
|
bra.uni BB0_53;
|
|
|
|
BB0_53:
|
|
add.u64 %rd61, %SP, 0;
|
|
cvta.to.local.u64 %rd62, %rd61;
|
|
max.f32 %f493, %f499, %f370;
|
|
mov.u32 %r219, 1065353216;
|
|
st.local.u32 [%rd62+8], %r219;
|
|
st.local.u32 [%rd62+4], %r219;
|
|
st.local.u32 [%rd62], %r219;
|
|
ld.global.u32 %r216, [root];
|
|
mov.u32 %r217, 1;
|
|
mov.f32 %f494, 0f6C4ECB8F;
|
|
// inline asm
|
|
call _rt_trace_64, (%r216, %f7, %f8, %f9, %f119, %f120, %f121, %r217, %f493, %f494, %rd61, %r108);
|
|
// inline asm
|
|
mul.f32 %f501, %f1044, %f120;
|
|
fma.rn.f32 %f502, %f1043, %f119, %f501;
|
|
fma.rn.f32 %f503, %f1045, %f121, %f502;
|
|
ld.local.f32 %f504, [%rd62+8];
|
|
ld.local.f32 %f505, [%rd62+4];
|
|
ld.local.f32 %f506, [%rd62];
|
|
add.f32 %f1142, %f1142, %f506;
|
|
add.f32 %f1143, %f1143, %f505;
|
|
add.f32 %f1144, %f1144, %f504;
|
|
fma.rn.f32 %f1139, %f120, %f506, %f1139;
|
|
fma.rn.f32 %f1140, %f120, %f505, %f1140;
|
|
fma.rn.f32 %f1141, %f120, %f504, %f1141;
|
|
fma.rn.f32 %f1136, %f121, %f506, %f1136;
|
|
fma.rn.f32 %f1137, %f121, %f505, %f1137;
|
|
fma.rn.f32 %f1138, %f121, %f504, %f1138;
|
|
fma.rn.f32 %f1133, %f119, %f506, %f1133;
|
|
fma.rn.f32 %f1134, %f119, %f505, %f1134;
|
|
fma.rn.f32 %f1135, %f119, %f504, %f1135;
|
|
mul.f32 %f507, %f119, %f120;
|
|
fma.rn.f32 %f1130, %f507, %f506, %f1130;
|
|
fma.rn.f32 %f1131, %f507, %f505, %f1131;
|
|
fma.rn.f32 %f1132, %f507, %f504, %f1132;
|
|
mul.f32 %f508, %f120, %f121;
|
|
fma.rn.f32 %f1127, %f508, %f506, %f1127;
|
|
fma.rn.f32 %f1128, %f508, %f505, %f1128;
|
|
fma.rn.f32 %f1129, %f508, %f504, %f1129;
|
|
fma.rn.f32 %f509, %f121, %f121, 0fBEAAAAAB;
|
|
fma.rn.f32 %f1124, %f509, %f506, %f1124;
|
|
fma.rn.f32 %f1125, %f509, %f505, %f1125;
|
|
fma.rn.f32 %f1126, %f509, %f504, %f1126;
|
|
mul.f32 %f510, %f119, %f121;
|
|
fma.rn.f32 %f1121, %f510, %f506, %f1121;
|
|
fma.rn.f32 %f1122, %f510, %f505, %f1122;
|
|
fma.rn.f32 %f1123, %f510, %f504, %f1123;
|
|
mul.f32 %f511, %f120, %f120;
|
|
mul.f32 %f512, %f119, %f119;
|
|
sub.f32 %f513, %f512, %f511;
|
|
fma.rn.f32 %f1118, %f513, %f506, %f1118;
|
|
fma.rn.f32 %f1119, %f513, %f505, %f1119;
|
|
fma.rn.f32 %f1120, %f513, %f504, %f1120;
|
|
cvt.sat.f32.f32 %f514, %f503;
|
|
fma.rn.f32 %f1145, %f514, %f506, %f1145;
|
|
fma.rn.f32 %f1146, %f514, %f505, %f1146;
|
|
fma.rn.f32 %f1147, %f514, %f504, %f1147;
|
|
|
|
BB0_54:
|
|
mad.lo.s32 %r406, %r11, 1664525, 1013904223;
|
|
mad.lo.s32 %r429, %r406, 1664525, 1013904223;
|
|
add.s32 %r428, %r428, 1;
|
|
setp.lt.s32 %p41, %r428, %r2;
|
|
@%p41 bra BB0_6;
|
|
|
|
mad.lo.s32 %r408, %r11, 1664525, 1013904223;
|
|
mad.lo.s32 %r429, %r408, 1664525, 1013904223;
|
|
add.s32 %r426, %r426, 1;
|
|
setp.lt.s32 %p42, %r426, %r2;
|
|
@%p42 bra BB0_5;
|
|
|
|
BB0_56:
|
|
mul.lo.s32 %r220, %r2, %r2;
|
|
cvt.rn.f32.s32 %f515, %r220;
|
|
rcp.rn.f32 %f516, %f515;
|
|
mul.f32 %f517, %f1145, %f516;
|
|
mul.f32 %f518, %f1146, %f516;
|
|
mul.f32 %f519, %f1147, %f516;
|
|
mul.f32 %f212, %f1142, %f516;
|
|
mul.f32 %f213, %f1143, %f516;
|
|
mul.f32 %f214, %f1144, %f516;
|
|
mul.f32 %f215, %f1139, %f516;
|
|
mul.f32 %f216, %f1140, %f516;
|
|
mul.f32 %f217, %f1141, %f516;
|
|
mul.f32 %f218, %f1136, %f516;
|
|
mul.f32 %f219, %f1137, %f516;
|
|
mul.f32 %f220, %f1138, %f516;
|
|
mul.f32 %f221, %f1133, %f516;
|
|
mul.f32 %f222, %f1134, %f516;
|
|
mul.f32 %f223, %f1135, %f516;
|
|
mul.f32 %f224, %f1130, %f516;
|
|
mul.f32 %f225, %f1131, %f516;
|
|
mul.f32 %f226, %f1132, %f516;
|
|
mul.f32 %f227, %f1127, %f516;
|
|
mul.f32 %f228, %f1128, %f516;
|
|
mul.f32 %f229, %f1129, %f516;
|
|
mul.f32 %f230, %f1124, %f516;
|
|
mul.f32 %f231, %f1125, %f516;
|
|
mul.f32 %f232, %f1126, %f516;
|
|
mul.f32 %f233, %f1121, %f516;
|
|
mul.f32 %f234, %f1122, %f516;
|
|
mul.f32 %f235, %f1123, %f516;
|
|
mul.f32 %f236, %f1118, %f516;
|
|
mul.f32 %f237, %f1119, %f516;
|
|
mul.f32 %f238, %f1120, %f516;
|
|
fma.rn.f32 %f520, %f1145, %f516, %f517;
|
|
fma.rn.f32 %f521, %f1146, %f516, %f518;
|
|
fma.rn.f32 %f522, %f1147, %f516, %f519;
|
|
ld.global.f32 %f523, [skyColor];
|
|
mul.f32 %f239, %f523, %f520;
|
|
ld.global.f32 %f524, [skyColor+4];
|
|
mul.f32 %f240, %f524, %f521;
|
|
ld.global.f32 %f525, [skyColor+8];
|
|
mul.f32 %f241, %f522, %f525;
|
|
ld.global.u32 %r450, [imageEnabled];
|
|
and.b32 %r221, %r450, 1;
|
|
setp.eq.b32 %p43, %r221, 1;
|
|
@!%p43 bra BB0_91;
|
|
bra.uni BB0_57;
|
|
|
|
BB0_57:
|
|
mov.f32 %f528, 0f3E666666;
|
|
cvt.rzi.f32.f32 %f529, %f528;
|
|
fma.rn.f32 %f530, %f529, 0fC0000000, 0f3EE66666;
|
|
abs.f32 %f242, %f530;
|
|
abs.f32 %f243, %f239;
|
|
setp.lt.f32 %p44, %f243, 0f00800000;
|
|
mul.f32 %f531, %f243, 0f4B800000;
|
|
selp.f32 %f532, 0fC3170000, 0fC2FE0000, %p44;
|
|
selp.f32 %f533, %f531, %f243, %p44;
|
|
mov.b32 %r222, %f533;
|
|
and.b32 %r223, %r222, 8388607;
|
|
or.b32 %r224, %r223, 1065353216;
|
|
mov.b32 %f534, %r224;
|
|
shr.u32 %r225, %r222, 23;
|
|
cvt.rn.f32.u32 %f535, %r225;
|
|
add.f32 %f536, %f532, %f535;
|
|
setp.gt.f32 %p45, %f534, 0f3FB504F3;
|
|
mul.f32 %f537, %f534, 0f3F000000;
|
|
add.f32 %f538, %f536, 0f3F800000;
|
|
selp.f32 %f539, %f537, %f534, %p45;
|
|
selp.f32 %f540, %f538, %f536, %p45;
|
|
add.f32 %f541, %f539, 0fBF800000;
|
|
add.f32 %f527, %f539, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f526,%f527;
|
|
// inline asm
|
|
add.f32 %f542, %f541, %f541;
|
|
mul.f32 %f543, %f526, %f542;
|
|
mul.f32 %f544, %f543, %f543;
|
|
mov.f32 %f545, 0f3C4CAF63;
|
|
mov.f32 %f546, 0f3B18F0FE;
|
|
fma.rn.f32 %f547, %f546, %f544, %f545;
|
|
mov.f32 %f548, 0f3DAAAABD;
|
|
fma.rn.f32 %f549, %f547, %f544, %f548;
|
|
mul.rn.f32 %f550, %f549, %f544;
|
|
mul.rn.f32 %f551, %f550, %f543;
|
|
sub.f32 %f552, %f541, %f543;
|
|
neg.f32 %f553, %f543;
|
|
add.f32 %f554, %f552, %f552;
|
|
fma.rn.f32 %f555, %f553, %f541, %f554;
|
|
mul.rn.f32 %f556, %f526, %f555;
|
|
add.f32 %f557, %f551, %f543;
|
|
sub.f32 %f558, %f543, %f557;
|
|
add.f32 %f559, %f551, %f558;
|
|
add.f32 %f560, %f556, %f559;
|
|
add.f32 %f561, %f557, %f560;
|
|
sub.f32 %f562, %f557, %f561;
|
|
add.f32 %f563, %f560, %f562;
|
|
mov.f32 %f564, 0f3F317200;
|
|
mul.rn.f32 %f565, %f540, %f564;
|
|
mov.f32 %f566, 0f35BFBE8E;
|
|
mul.rn.f32 %f567, %f540, %f566;
|
|
add.f32 %f568, %f565, %f561;
|
|
sub.f32 %f569, %f565, %f568;
|
|
add.f32 %f570, %f561, %f569;
|
|
add.f32 %f571, %f563, %f570;
|
|
add.f32 %f572, %f567, %f571;
|
|
add.f32 %f573, %f568, %f572;
|
|
sub.f32 %f574, %f568, %f573;
|
|
add.f32 %f575, %f572, %f574;
|
|
mov.f32 %f576, 0f3EE66666;
|
|
mul.rn.f32 %f577, %f576, %f573;
|
|
neg.f32 %f578, %f577;
|
|
fma.rn.f32 %f579, %f576, %f573, %f578;
|
|
fma.rn.f32 %f580, %f576, %f575, %f579;
|
|
mov.f32 %f581, 0f00000000;
|
|
fma.rn.f32 %f582, %f581, %f573, %f580;
|
|
add.rn.f32 %f583, %f577, %f582;
|
|
neg.f32 %f584, %f583;
|
|
add.rn.f32 %f585, %f577, %f584;
|
|
add.rn.f32 %f586, %f585, %f582;
|
|
mov.b32 %r226, %f583;
|
|
setp.eq.s32 %p46, %r226, 1118925336;
|
|
add.s32 %r227, %r226, -1;
|
|
mov.b32 %f587, %r227;
|
|
add.f32 %f588, %f586, 0f37000000;
|
|
selp.f32 %f589, %f587, %f583, %p46;
|
|
selp.f32 %f244, %f588, %f586, %p46;
|
|
mul.f32 %f590, %f589, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f591, %f590;
|
|
mov.f32 %f592, 0fBF317200;
|
|
fma.rn.f32 %f593, %f591, %f592, %f589;
|
|
mov.f32 %f594, 0fB5BFBE8E;
|
|
fma.rn.f32 %f595, %f591, %f594, %f593;
|
|
mul.f32 %f596, %f595, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f597, %f596;
|
|
add.f32 %f598, %f591, 0f00000000;
|
|
ex2.approx.f32 %f599, %f598;
|
|
mul.f32 %f600, %f597, %f599;
|
|
setp.lt.f32 %p47, %f589, 0fC2D20000;
|
|
selp.f32 %f601, 0f00000000, %f600, %p47;
|
|
setp.gt.f32 %p48, %f589, 0f42D20000;
|
|
selp.f32 %f1178, 0f7F800000, %f601, %p48;
|
|
setp.eq.f32 %p49, %f1178, 0f7F800000;
|
|
@%p49 bra BB0_59;
|
|
|
|
fma.rn.f32 %f1178, %f1178, %f244, %f1178;
|
|
|
|
BB0_59:
|
|
setp.lt.f32 %p50, %f239, 0f00000000;
|
|
setp.eq.f32 %p51, %f242, 0f3F800000;
|
|
and.pred %p1, %p50, %p51;
|
|
mov.b32 %r228, %f1178;
|
|
xor.b32 %r229, %r228, -2147483648;
|
|
mov.b32 %f602, %r229;
|
|
selp.f32 %f1180, %f602, %f1178, %p1;
|
|
setp.eq.f32 %p52, %f239, 0f00000000;
|
|
@%p52 bra BB0_62;
|
|
bra.uni BB0_60;
|
|
|
|
BB0_62:
|
|
add.f32 %f605, %f239, %f239;
|
|
selp.f32 %f1180, %f605, 0f00000000, %p51;
|
|
bra.uni BB0_63;
|
|
|
|
BB0_127:
|
|
mov.u64 %rd296, image_HDR;
|
|
cvta.global.u64 %rd291, %rd296;
|
|
mov.u32 %r335, 8;
|
|
mov.u64 %rd295, 0;
|
|
// inline asm
|
|
call (%rd290), _rt_buffer_get_64, (%rd291, %r92, %r335, %rd18, %rd19, %rd295, %rd295);
|
|
// inline asm
|
|
mov.f32 %f957, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs175, %f957;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs176, 0;
|
|
st.v4.u16 [%rd290], {%rs175, %rs175, %rs175, %rs176};
|
|
|
|
BB0_128:
|
|
ld.global.u32 %r336, [additive];
|
|
setp.eq.s32 %p106, %r336, 0;
|
|
@%p106 bra BB0_130;
|
|
|
|
mov.u64 %rd309, image_RNM0;
|
|
cvta.global.u64 %rd298, %rd309;
|
|
mov.u32 %r340, 8;
|
|
mov.u64 %rd308, 0;
|
|
// inline asm
|
|
call (%rd297), _rt_buffer_get_64, (%rd298, %r92, %r340, %rd18, %rd19, %rd308, %rd308);
|
|
// inline asm
|
|
ld.v4.u16 {%rs183, %rs184, %rs185, %rs186}, [%rd297];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f958, %rs183;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f959, %rs184;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f960, %rs185;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd303), _rt_buffer_get_64, (%rd298, %r92, %r340, %rd18, %rd19, %rd308, %rd308);
|
|
// inline asm
|
|
add.f32 %f961, %f958, 0f00000000;
|
|
add.f32 %f962, %f959, 0f00000000;
|
|
add.f32 %f963, %f960, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs182, %f963;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs181, %f962;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs180, %f961;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs187, 0;
|
|
st.v4.u16 [%rd303], {%rs180, %rs181, %rs182, %rs187};
|
|
bra.uni BB0_131;
|
|
|
|
BB0_130:
|
|
mov.u64 %rd316, image_RNM0;
|
|
cvta.global.u64 %rd311, %rd316;
|
|
mov.u32 %r342, 8;
|
|
mov.u64 %rd315, 0;
|
|
// inline asm
|
|
call (%rd310), _rt_buffer_get_64, (%rd311, %r92, %r342, %rd18, %rd19, %rd315, %rd315);
|
|
// inline asm
|
|
mov.f32 %f964, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs188, %f964;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs189, 0;
|
|
st.v4.u16 [%rd310], {%rs188, %rs188, %rs188, %rs189};
|
|
|
|
BB0_131:
|
|
ld.global.u32 %r343, [additive];
|
|
setp.eq.s32 %p107, %r343, 0;
|
|
@%p107 bra BB0_133;
|
|
|
|
mov.u64 %rd329, image_RNM1;
|
|
cvta.global.u64 %rd318, %rd329;
|
|
mov.u32 %r347, 8;
|
|
mov.u64 %rd328, 0;
|
|
// inline asm
|
|
call (%rd317), _rt_buffer_get_64, (%rd318, %r92, %r347, %rd18, %rd19, %rd328, %rd328);
|
|
// inline asm
|
|
ld.v4.u16 {%rs196, %rs197, %rs198, %rs199}, [%rd317];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f965, %rs196;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f966, %rs197;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f967, %rs198;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd323), _rt_buffer_get_64, (%rd318, %r92, %r347, %rd18, %rd19, %rd328, %rd328);
|
|
// inline asm
|
|
add.f32 %f968, %f965, 0f00000000;
|
|
add.f32 %f969, %f966, 0f00000000;
|
|
add.f32 %f970, %f967, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs195, %f970;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs194, %f969;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs193, %f968;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs200, 0;
|
|
st.v4.u16 [%rd323], {%rs193, %rs194, %rs195, %rs200};
|
|
bra.uni BB0_134;
|
|
|
|
BB0_133:
|
|
mov.u64 %rd336, image_RNM1;
|
|
cvta.global.u64 %rd331, %rd336;
|
|
mov.u32 %r349, 8;
|
|
mov.u64 %rd335, 0;
|
|
// inline asm
|
|
call (%rd330), _rt_buffer_get_64, (%rd331, %r92, %r349, %rd18, %rd19, %rd335, %rd335);
|
|
// inline asm
|
|
mov.f32 %f971, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs201, %f971;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs202, 0;
|
|
st.v4.u16 [%rd330], {%rs201, %rs201, %rs201, %rs202};
|
|
|
|
BB0_134:
|
|
ld.global.u32 %r350, [additive];
|
|
setp.eq.s32 %p108, %r350, 0;
|
|
@%p108 bra BB0_136;
|
|
|
|
mov.u64 %rd349, image_RNM2;
|
|
cvta.global.u64 %rd338, %rd349;
|
|
mov.u32 %r354, 8;
|
|
mov.u64 %rd348, 0;
|
|
// inline asm
|
|
call (%rd337), _rt_buffer_get_64, (%rd338, %r92, %r354, %rd18, %rd19, %rd348, %rd348);
|
|
// inline asm
|
|
ld.v4.u16 {%rs209, %rs210, %rs211, %rs212}, [%rd337];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f972, %rs209;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f973, %rs210;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f974, %rs211;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd343), _rt_buffer_get_64, (%rd338, %r92, %r354, %rd18, %rd19, %rd348, %rd348);
|
|
// inline asm
|
|
add.f32 %f975, %f972, 0f00000000;
|
|
add.f32 %f976, %f973, 0f00000000;
|
|
add.f32 %f977, %f974, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs208, %f977;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs207, %f976;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs206, %f975;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs213, 0;
|
|
st.v4.u16 [%rd343], {%rs206, %rs207, %rs208, %rs213};
|
|
bra.uni BB0_137;
|
|
|
|
BB0_136:
|
|
mov.u64 %rd356, image_RNM2;
|
|
cvta.global.u64 %rd351, %rd356;
|
|
mov.u32 %r356, 8;
|
|
mov.u64 %rd355, 0;
|
|
// inline asm
|
|
call (%rd350), _rt_buffer_get_64, (%rd351, %r92, %r356, %rd18, %rd19, %rd355, %rd355);
|
|
// inline asm
|
|
mov.f32 %f978, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs214, %f978;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs215, 0;
|
|
st.v4.u16 [%rd350], {%rs214, %rs214, %rs214, %rs215};
|
|
|
|
BB0_137:
|
|
ld.global.u32 %r357, [additive];
|
|
setp.eq.s32 %p109, %r357, 0;
|
|
@%p109 bra BB0_139;
|
|
|
|
mov.u64 %rd369, image_RNM3;
|
|
cvta.global.u64 %rd358, %rd369;
|
|
mov.u32 %r361, 8;
|
|
mov.u64 %rd368, 0;
|
|
// inline asm
|
|
call (%rd357), _rt_buffer_get_64, (%rd358, %r92, %r361, %rd18, %rd19, %rd368, %rd368);
|
|
// inline asm
|
|
ld.v4.u16 {%rs222, %rs223, %rs224, %rs225}, [%rd357];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f979, %rs222;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f980, %rs223;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f981, %rs224;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd363), _rt_buffer_get_64, (%rd358, %r92, %r361, %rd18, %rd19, %rd368, %rd368);
|
|
// inline asm
|
|
add.f32 %f982, %f979, 0f00000000;
|
|
add.f32 %f983, %f980, 0f00000000;
|
|
add.f32 %f984, %f981, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs221, %f984;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs220, %f983;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs219, %f982;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs226, 0;
|
|
st.v4.u16 [%rd363], {%rs219, %rs220, %rs221, %rs226};
|
|
bra.uni BB0_140;
|
|
|
|
BB0_139:
|
|
mov.u64 %rd376, image_RNM3;
|
|
cvta.global.u64 %rd371, %rd376;
|
|
mov.u32 %r363, 8;
|
|
mov.u64 %rd375, 0;
|
|
// inline asm
|
|
call (%rd370), _rt_buffer_get_64, (%rd371, %r92, %r363, %rd18, %rd19, %rd375, %rd375);
|
|
// inline asm
|
|
mov.f32 %f985, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs227, %f985;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs228, 0;
|
|
st.v4.u16 [%rd370], {%rs227, %rs227, %rs227, %rs228};
|
|
|
|
BB0_140:
|
|
ld.global.u32 %r364, [additive];
|
|
setp.eq.s32 %p110, %r364, 0;
|
|
@%p110 bra BB0_142;
|
|
|
|
mov.u64 %rd389, image_RNM4;
|
|
cvta.global.u64 %rd378, %rd389;
|
|
mov.u32 %r368, 8;
|
|
mov.u64 %rd388, 0;
|
|
// inline asm
|
|
call (%rd377), _rt_buffer_get_64, (%rd378, %r92, %r368, %rd18, %rd19, %rd388, %rd388);
|
|
// inline asm
|
|
ld.v4.u16 {%rs235, %rs236, %rs237, %rs238}, [%rd377];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f986, %rs235;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f987, %rs236;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f988, %rs237;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd383), _rt_buffer_get_64, (%rd378, %r92, %r368, %rd18, %rd19, %rd388, %rd388);
|
|
// inline asm
|
|
add.f32 %f989, %f986, 0f00000000;
|
|
add.f32 %f990, %f987, 0f00000000;
|
|
add.f32 %f991, %f988, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs234, %f991;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs233, %f990;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs232, %f989;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs239, 0;
|
|
st.v4.u16 [%rd383], {%rs232, %rs233, %rs234, %rs239};
|
|
bra.uni BB0_143;
|
|
|
|
BB0_142:
|
|
mov.u64 %rd396, image_RNM4;
|
|
cvta.global.u64 %rd391, %rd396;
|
|
mov.u32 %r370, 8;
|
|
mov.u64 %rd395, 0;
|
|
// inline asm
|
|
call (%rd390), _rt_buffer_get_64, (%rd391, %r92, %r370, %rd18, %rd19, %rd395, %rd395);
|
|
// inline asm
|
|
mov.f32 %f992, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs240, %f992;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs241, 0;
|
|
st.v4.u16 [%rd390], {%rs240, %rs240, %rs240, %rs241};
|
|
|
|
BB0_143:
|
|
ld.global.u32 %r371, [additive];
|
|
setp.eq.s32 %p111, %r371, 0;
|
|
@%p111 bra BB0_145;
|
|
|
|
mov.u64 %rd409, image_RNM5;
|
|
cvta.global.u64 %rd398, %rd409;
|
|
mov.u32 %r375, 8;
|
|
mov.u64 %rd408, 0;
|
|
// inline asm
|
|
call (%rd397), _rt_buffer_get_64, (%rd398, %r92, %r375, %rd18, %rd19, %rd408, %rd408);
|
|
// inline asm
|
|
ld.v4.u16 {%rs248, %rs249, %rs250, %rs251}, [%rd397];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f993, %rs248;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f994, %rs249;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f995, %rs250;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd403), _rt_buffer_get_64, (%rd398, %r92, %r375, %rd18, %rd19, %rd408, %rd408);
|
|
// inline asm
|
|
add.f32 %f996, %f993, 0f00000000;
|
|
add.f32 %f997, %f994, 0f00000000;
|
|
add.f32 %f998, %f995, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs247, %f998;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs246, %f997;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs245, %f996;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs252, 0;
|
|
st.v4.u16 [%rd403], {%rs245, %rs246, %rs247, %rs252};
|
|
bra.uni BB0_146;
|
|
|
|
BB0_145:
|
|
mov.u64 %rd416, image_RNM5;
|
|
cvta.global.u64 %rd411, %rd416;
|
|
mov.u32 %r377, 8;
|
|
mov.u64 %rd415, 0;
|
|
// inline asm
|
|
call (%rd410), _rt_buffer_get_64, (%rd411, %r92, %r377, %rd18, %rd19, %rd415, %rd415);
|
|
// inline asm
|
|
mov.f32 %f999, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs253, %f999;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs254, 0;
|
|
st.v4.u16 [%rd410], {%rs253, %rs253, %rs253, %rs254};
|
|
|
|
BB0_146:
|
|
ld.global.u32 %r378, [additive];
|
|
setp.eq.s32 %p112, %r378, 0;
|
|
@%p112 bra BB0_148;
|
|
|
|
mov.u64 %rd429, image_RNM6;
|
|
cvta.global.u64 %rd418, %rd429;
|
|
mov.u32 %r382, 8;
|
|
mov.u64 %rd428, 0;
|
|
// inline asm
|
|
call (%rd417), _rt_buffer_get_64, (%rd418, %r92, %r382, %rd18, %rd19, %rd428, %rd428);
|
|
// inline asm
|
|
ld.v4.u16 {%rs261, %rs262, %rs263, %rs264}, [%rd417];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1000, %rs261;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1001, %rs262;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1002, %rs263;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd423), _rt_buffer_get_64, (%rd418, %r92, %r382, %rd18, %rd19, %rd428, %rd428);
|
|
// inline asm
|
|
add.f32 %f1003, %f1000, 0f00000000;
|
|
add.f32 %f1004, %f1001, 0f00000000;
|
|
add.f32 %f1005, %f1002, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs260, %f1005;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs259, %f1004;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs258, %f1003;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs265, 0;
|
|
st.v4.u16 [%rd423], {%rs258, %rs259, %rs260, %rs265};
|
|
bra.uni BB0_149;
|
|
|
|
BB0_148:
|
|
mov.u64 %rd436, image_RNM6;
|
|
cvta.global.u64 %rd431, %rd436;
|
|
mov.u32 %r384, 8;
|
|
mov.u64 %rd435, 0;
|
|
// inline asm
|
|
call (%rd430), _rt_buffer_get_64, (%rd431, %r92, %r384, %rd18, %rd19, %rd435, %rd435);
|
|
// inline asm
|
|
mov.f32 %f1006, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs266, %f1006;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs267, 0;
|
|
st.v4.u16 [%rd430], {%rs266, %rs266, %rs266, %rs267};
|
|
|
|
BB0_149:
|
|
ld.global.u32 %r385, [additive];
|
|
setp.eq.s32 %p113, %r385, 0;
|
|
@%p113 bra BB0_151;
|
|
|
|
mov.u64 %rd449, image_RNM7;
|
|
cvta.global.u64 %rd438, %rd449;
|
|
mov.u32 %r389, 8;
|
|
mov.u64 %rd448, 0;
|
|
// inline asm
|
|
call (%rd437), _rt_buffer_get_64, (%rd438, %r92, %r389, %rd18, %rd19, %rd448, %rd448);
|
|
// inline asm
|
|
ld.v4.u16 {%rs274, %rs275, %rs276, %rs277}, [%rd437];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1007, %rs274;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1008, %rs275;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1009, %rs276;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd443), _rt_buffer_get_64, (%rd438, %r92, %r389, %rd18, %rd19, %rd448, %rd448);
|
|
// inline asm
|
|
add.f32 %f1010, %f1007, 0f00000000;
|
|
add.f32 %f1011, %f1008, 0f00000000;
|
|
add.f32 %f1012, %f1009, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs273, %f1012;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs272, %f1011;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs271, %f1010;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs278, 0;
|
|
st.v4.u16 [%rd443], {%rs271, %rs272, %rs273, %rs278};
|
|
bra.uni BB0_152;
|
|
|
|
BB0_151:
|
|
mov.u64 %rd456, image_RNM7;
|
|
cvta.global.u64 %rd451, %rd456;
|
|
mov.u32 %r391, 8;
|
|
mov.u64 %rd455, 0;
|
|
// inline asm
|
|
call (%rd450), _rt_buffer_get_64, (%rd451, %r92, %r391, %rd18, %rd19, %rd455, %rd455);
|
|
// inline asm
|
|
mov.f32 %f1013, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs279, %f1013;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs280, 0;
|
|
st.v4.u16 [%rd450], {%rs279, %rs279, %rs279, %rs280};
|
|
|
|
BB0_152:
|
|
ld.global.u32 %r392, [additive];
|
|
setp.eq.s32 %p114, %r392, 0;
|
|
@%p114 bra BB0_154;
|
|
|
|
mov.u64 %rd469, image_RNM8;
|
|
cvta.global.u64 %rd458, %rd469;
|
|
mov.u32 %r396, 8;
|
|
mov.u64 %rd468, 0;
|
|
// inline asm
|
|
call (%rd457), _rt_buffer_get_64, (%rd458, %r92, %r396, %rd18, %rd19, %rd468, %rd468);
|
|
// inline asm
|
|
ld.v4.u16 {%rs287, %rs288, %rs289, %rs290}, [%rd457];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1014, %rs287;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1015, %rs288;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1016, %rs289;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd463), _rt_buffer_get_64, (%rd458, %r92, %r396, %rd18, %rd19, %rd468, %rd468);
|
|
// inline asm
|
|
add.f32 %f1017, %f1014, 0f00000000;
|
|
add.f32 %f1018, %f1015, 0f00000000;
|
|
add.f32 %f1019, %f1016, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs286, %f1019;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs285, %f1018;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs284, %f1017;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs291, 0;
|
|
st.v4.u16 [%rd463], {%rs284, %rs285, %rs286, %rs291};
|
|
bra.uni BB0_155;
|
|
|
|
BB0_154:
|
|
mov.u64 %rd476, image_RNM8;
|
|
cvta.global.u64 %rd471, %rd476;
|
|
mov.u32 %r398, 8;
|
|
mov.u64 %rd475, 0;
|
|
// inline asm
|
|
call (%rd470), _rt_buffer_get_64, (%rd471, %r92, %r398, %rd18, %rd19, %rd475, %rd475);
|
|
// inline asm
|
|
mov.f32 %f1020, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs292, %f1020;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs293, 0;
|
|
st.v4.u16 [%rd470], {%rs292, %rs292, %rs292, %rs293};
|
|
bra.uni BB0_155;
|
|
|
|
BB0_60:
|
|
setp.geu.f32 %p53, %f239, 0f00000000;
|
|
@%p53 bra BB0_63;
|
|
|
|
cvt.rzi.f32.f32 %f604, %f576;
|
|
setp.neu.f32 %p54, %f604, 0f3EE66666;
|
|
selp.f32 %f1180, 0f7FFFFFFF, %f1180, %p54;
|
|
|
|
BB0_63:
|
|
add.f32 %f606, %f243, 0f3EE66666;
|
|
mov.b32 %r230, %f606;
|
|
setp.lt.s32 %p56, %r230, 2139095040;
|
|
@%p56 bra BB0_68;
|
|
|
|
setp.gtu.f32 %p57, %f243, 0f7F800000;
|
|
@%p57 bra BB0_67;
|
|
bra.uni BB0_65;
|
|
|
|
BB0_67:
|
|
add.f32 %f1180, %f239, 0f3EE66666;
|
|
bra.uni BB0_68;
|
|
|
|
BB0_65:
|
|
setp.neu.f32 %p58, %f243, 0f7F800000;
|
|
@%p58 bra BB0_68;
|
|
|
|
selp.f32 %f1180, 0fFF800000, 0f7F800000, %p1;
|
|
|
|
BB0_68:
|
|
setp.eq.f32 %p59, %f239, 0f3F800000;
|
|
selp.f32 %f255, 0f3F800000, %f1180, %p59;
|
|
abs.f32 %f256, %f240;
|
|
setp.lt.f32 %p60, %f256, 0f00800000;
|
|
mul.f32 %f609, %f256, 0f4B800000;
|
|
selp.f32 %f610, 0fC3170000, 0fC2FE0000, %p60;
|
|
selp.f32 %f611, %f609, %f256, %p60;
|
|
mov.b32 %r231, %f611;
|
|
and.b32 %r232, %r231, 8388607;
|
|
or.b32 %r233, %r232, 1065353216;
|
|
mov.b32 %f612, %r233;
|
|
shr.u32 %r234, %r231, 23;
|
|
cvt.rn.f32.u32 %f613, %r234;
|
|
add.f32 %f614, %f610, %f613;
|
|
setp.gt.f32 %p61, %f612, 0f3FB504F3;
|
|
mul.f32 %f615, %f612, 0f3F000000;
|
|
add.f32 %f616, %f614, 0f3F800000;
|
|
selp.f32 %f617, %f615, %f612, %p61;
|
|
selp.f32 %f618, %f616, %f614, %p61;
|
|
add.f32 %f619, %f617, 0fBF800000;
|
|
add.f32 %f608, %f617, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f607,%f608;
|
|
// inline asm
|
|
add.f32 %f620, %f619, %f619;
|
|
mul.f32 %f621, %f607, %f620;
|
|
mul.f32 %f622, %f621, %f621;
|
|
fma.rn.f32 %f625, %f546, %f622, %f545;
|
|
fma.rn.f32 %f627, %f625, %f622, %f548;
|
|
mul.rn.f32 %f628, %f627, %f622;
|
|
mul.rn.f32 %f629, %f628, %f621;
|
|
sub.f32 %f630, %f619, %f621;
|
|
neg.f32 %f631, %f621;
|
|
add.f32 %f632, %f630, %f630;
|
|
fma.rn.f32 %f633, %f631, %f619, %f632;
|
|
mul.rn.f32 %f634, %f607, %f633;
|
|
add.f32 %f635, %f629, %f621;
|
|
sub.f32 %f636, %f621, %f635;
|
|
add.f32 %f637, %f629, %f636;
|
|
add.f32 %f638, %f634, %f637;
|
|
add.f32 %f639, %f635, %f638;
|
|
sub.f32 %f640, %f635, %f639;
|
|
add.f32 %f641, %f638, %f640;
|
|
mul.rn.f32 %f643, %f618, %f564;
|
|
mul.rn.f32 %f645, %f618, %f566;
|
|
add.f32 %f646, %f643, %f639;
|
|
sub.f32 %f647, %f643, %f646;
|
|
add.f32 %f648, %f639, %f647;
|
|
add.f32 %f649, %f641, %f648;
|
|
add.f32 %f650, %f645, %f649;
|
|
add.f32 %f651, %f646, %f650;
|
|
sub.f32 %f652, %f646, %f651;
|
|
add.f32 %f653, %f650, %f652;
|
|
mul.rn.f32 %f655, %f576, %f651;
|
|
neg.f32 %f656, %f655;
|
|
fma.rn.f32 %f657, %f576, %f651, %f656;
|
|
fma.rn.f32 %f658, %f576, %f653, %f657;
|
|
fma.rn.f32 %f660, %f581, %f651, %f658;
|
|
add.rn.f32 %f661, %f655, %f660;
|
|
neg.f32 %f662, %f661;
|
|
add.rn.f32 %f663, %f655, %f662;
|
|
add.rn.f32 %f664, %f663, %f660;
|
|
mov.b32 %r235, %f661;
|
|
setp.eq.s32 %p62, %r235, 1118925336;
|
|
add.s32 %r236, %r235, -1;
|
|
mov.b32 %f665, %r236;
|
|
add.f32 %f666, %f664, 0f37000000;
|
|
selp.f32 %f667, %f665, %f661, %p62;
|
|
selp.f32 %f257, %f666, %f664, %p62;
|
|
mul.f32 %f668, %f667, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f669, %f668;
|
|
fma.rn.f32 %f671, %f669, %f592, %f667;
|
|
fma.rn.f32 %f673, %f669, %f594, %f671;
|
|
mul.f32 %f674, %f673, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f675, %f674;
|
|
add.f32 %f676, %f669, 0f00000000;
|
|
ex2.approx.f32 %f677, %f676;
|
|
mul.f32 %f678, %f675, %f677;
|
|
setp.lt.f32 %p63, %f667, 0fC2D20000;
|
|
selp.f32 %f679, 0f00000000, %f678, %p63;
|
|
setp.gt.f32 %p64, %f667, 0f42D20000;
|
|
selp.f32 %f1181, 0f7F800000, %f679, %p64;
|
|
setp.eq.f32 %p65, %f1181, 0f7F800000;
|
|
@%p65 bra BB0_70;
|
|
|
|
fma.rn.f32 %f1181, %f1181, %f257, %f1181;
|
|
|
|
BB0_70:
|
|
setp.lt.f32 %p66, %f240, 0f00000000;
|
|
and.pred %p2, %p66, %p51;
|
|
mov.b32 %r237, %f1181;
|
|
xor.b32 %r238, %r237, -2147483648;
|
|
mov.b32 %f680, %r238;
|
|
selp.f32 %f1183, %f680, %f1181, %p2;
|
|
setp.eq.f32 %p68, %f240, 0f00000000;
|
|
@%p68 bra BB0_73;
|
|
bra.uni BB0_71;
|
|
|
|
BB0_73:
|
|
add.f32 %f683, %f240, %f240;
|
|
selp.f32 %f1183, %f683, 0f00000000, %p51;
|
|
bra.uni BB0_74;
|
|
|
|
BB0_71:
|
|
setp.geu.f32 %p69, %f240, 0f00000000;
|
|
@%p69 bra BB0_74;
|
|
|
|
cvt.rzi.f32.f32 %f682, %f576;
|
|
setp.neu.f32 %p70, %f682, 0f3EE66666;
|
|
selp.f32 %f1183, 0f7FFFFFFF, %f1183, %p70;
|
|
|
|
BB0_74:
|
|
add.f32 %f684, %f256, 0f3EE66666;
|
|
mov.b32 %r239, %f684;
|
|
setp.lt.s32 %p72, %r239, 2139095040;
|
|
@%p72 bra BB0_79;
|
|
|
|
setp.gtu.f32 %p73, %f256, 0f7F800000;
|
|
@%p73 bra BB0_78;
|
|
bra.uni BB0_76;
|
|
|
|
BB0_78:
|
|
add.f32 %f1183, %f240, 0f3EE66666;
|
|
bra.uni BB0_79;
|
|
|
|
BB0_76:
|
|
setp.neu.f32 %p74, %f256, 0f7F800000;
|
|
@%p74 bra BB0_79;
|
|
|
|
selp.f32 %f1183, 0fFF800000, 0f7F800000, %p2;
|
|
|
|
BB0_79:
|
|
setp.eq.f32 %p75, %f240, 0f3F800000;
|
|
selp.f32 %f268, 0f3F800000, %f1183, %p75;
|
|
abs.f32 %f269, %f241;
|
|
setp.lt.f32 %p76, %f269, 0f00800000;
|
|
mul.f32 %f687, %f269, 0f4B800000;
|
|
selp.f32 %f688, 0fC3170000, 0fC2FE0000, %p76;
|
|
selp.f32 %f689, %f687, %f269, %p76;
|
|
mov.b32 %r240, %f689;
|
|
and.b32 %r241, %r240, 8388607;
|
|
or.b32 %r242, %r241, 1065353216;
|
|
mov.b32 %f690, %r242;
|
|
shr.u32 %r243, %r240, 23;
|
|
cvt.rn.f32.u32 %f691, %r243;
|
|
add.f32 %f692, %f688, %f691;
|
|
setp.gt.f32 %p77, %f690, 0f3FB504F3;
|
|
mul.f32 %f693, %f690, 0f3F000000;
|
|
add.f32 %f694, %f692, 0f3F800000;
|
|
selp.f32 %f695, %f693, %f690, %p77;
|
|
selp.f32 %f696, %f694, %f692, %p77;
|
|
add.f32 %f697, %f695, 0fBF800000;
|
|
add.f32 %f686, %f695, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f685,%f686;
|
|
// inline asm
|
|
add.f32 %f698, %f697, %f697;
|
|
mul.f32 %f699, %f685, %f698;
|
|
mul.f32 %f700, %f699, %f699;
|
|
fma.rn.f32 %f703, %f546, %f700, %f545;
|
|
fma.rn.f32 %f705, %f703, %f700, %f548;
|
|
mul.rn.f32 %f706, %f705, %f700;
|
|
mul.rn.f32 %f707, %f706, %f699;
|
|
sub.f32 %f708, %f697, %f699;
|
|
neg.f32 %f709, %f699;
|
|
add.f32 %f710, %f708, %f708;
|
|
fma.rn.f32 %f711, %f709, %f697, %f710;
|
|
mul.rn.f32 %f712, %f685, %f711;
|
|
add.f32 %f713, %f707, %f699;
|
|
sub.f32 %f714, %f699, %f713;
|
|
add.f32 %f715, %f707, %f714;
|
|
add.f32 %f716, %f712, %f715;
|
|
add.f32 %f717, %f713, %f716;
|
|
sub.f32 %f718, %f713, %f717;
|
|
add.f32 %f719, %f716, %f718;
|
|
mul.rn.f32 %f721, %f696, %f564;
|
|
mul.rn.f32 %f723, %f696, %f566;
|
|
add.f32 %f724, %f721, %f717;
|
|
sub.f32 %f725, %f721, %f724;
|
|
add.f32 %f726, %f717, %f725;
|
|
add.f32 %f727, %f719, %f726;
|
|
add.f32 %f728, %f723, %f727;
|
|
add.f32 %f729, %f724, %f728;
|
|
sub.f32 %f730, %f724, %f729;
|
|
add.f32 %f731, %f728, %f730;
|
|
mul.rn.f32 %f733, %f576, %f729;
|
|
neg.f32 %f734, %f733;
|
|
fma.rn.f32 %f735, %f576, %f729, %f734;
|
|
fma.rn.f32 %f736, %f576, %f731, %f735;
|
|
fma.rn.f32 %f738, %f581, %f729, %f736;
|
|
add.rn.f32 %f739, %f733, %f738;
|
|
neg.f32 %f740, %f739;
|
|
add.rn.f32 %f741, %f733, %f740;
|
|
add.rn.f32 %f742, %f741, %f738;
|
|
mov.b32 %r244, %f739;
|
|
setp.eq.s32 %p78, %r244, 1118925336;
|
|
add.s32 %r245, %r244, -1;
|
|
mov.b32 %f743, %r245;
|
|
add.f32 %f744, %f742, 0f37000000;
|
|
selp.f32 %f745, %f743, %f739, %p78;
|
|
selp.f32 %f270, %f744, %f742, %p78;
|
|
mul.f32 %f746, %f745, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f747, %f746;
|
|
fma.rn.f32 %f749, %f747, %f592, %f745;
|
|
fma.rn.f32 %f751, %f747, %f594, %f749;
|
|
mul.f32 %f752, %f751, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f753, %f752;
|
|
add.f32 %f754, %f747, 0f00000000;
|
|
ex2.approx.f32 %f755, %f754;
|
|
mul.f32 %f756, %f753, %f755;
|
|
setp.lt.f32 %p79, %f745, 0fC2D20000;
|
|
selp.f32 %f757, 0f00000000, %f756, %p79;
|
|
setp.gt.f32 %p80, %f745, 0f42D20000;
|
|
selp.f32 %f1184, 0f7F800000, %f757, %p80;
|
|
setp.eq.f32 %p81, %f1184, 0f7F800000;
|
|
@%p81 bra BB0_81;
|
|
|
|
fma.rn.f32 %f1184, %f1184, %f270, %f1184;
|
|
|
|
BB0_81:
|
|
setp.lt.f32 %p82, %f241, 0f00000000;
|
|
and.pred %p3, %p82, %p51;
|
|
mov.b32 %r246, %f1184;
|
|
xor.b32 %r247, %r246, -2147483648;
|
|
mov.b32 %f758, %r247;
|
|
selp.f32 %f1186, %f758, %f1184, %p3;
|
|
setp.eq.f32 %p84, %f241, 0f00000000;
|
|
@%p84 bra BB0_84;
|
|
bra.uni BB0_82;
|
|
|
|
BB0_84:
|
|
add.f32 %f761, %f241, %f241;
|
|
selp.f32 %f1186, %f761, 0f00000000, %p51;
|
|
bra.uni BB0_85;
|
|
|
|
BB0_82:
|
|
setp.geu.f32 %p85, %f241, 0f00000000;
|
|
@%p85 bra BB0_85;
|
|
|
|
cvt.rzi.f32.f32 %f760, %f576;
|
|
setp.neu.f32 %p86, %f760, 0f3EE66666;
|
|
selp.f32 %f1186, 0f7FFFFFFF, %f1186, %p86;
|
|
|
|
BB0_85:
|
|
add.f32 %f762, %f269, 0f3EE66666;
|
|
mov.b32 %r248, %f762;
|
|
setp.lt.s32 %p88, %r248, 2139095040;
|
|
@%p88 bra BB0_90;
|
|
|
|
setp.gtu.f32 %p89, %f269, 0f7F800000;
|
|
@%p89 bra BB0_89;
|
|
bra.uni BB0_87;
|
|
|
|
BB0_89:
|
|
add.f32 %f1186, %f241, 0f3EE66666;
|
|
bra.uni BB0_90;
|
|
|
|
BB0_87:
|
|
setp.neu.f32 %p90, %f269, 0f7F800000;
|
|
@%p90 bra BB0_90;
|
|
|
|
selp.f32 %f1186, 0fFF800000, 0f7F800000, %p3;
|
|
|
|
BB0_90:
|
|
mov.u32 %r409, 4;
|
|
setp.eq.f32 %p91, %f241, 0f3F800000;
|
|
selp.f32 %f763, 0f3F800000, %f1186, %p91;
|
|
cvt.u64.u32 %rd66, %r5;
|
|
cvt.u64.u32 %rd65, %r4;
|
|
mov.u64 %rd69, image;
|
|
cvta.global.u64 %rd64, %rd69;
|
|
// inline asm
|
|
call (%rd63), _rt_buffer_get_64, (%rd64, %r92, %r409, %rd65, %rd66, %rd25, %rd25);
|
|
// inline asm
|
|
cvt.sat.f32.f32 %f764, %f763;
|
|
mul.f32 %f765, %f764, 0f437FFD71;
|
|
cvt.rzi.u32.f32 %r251, %f765;
|
|
cvt.sat.f32.f32 %f766, %f268;
|
|
mul.f32 %f767, %f766, 0f437FFD71;
|
|
cvt.rzi.u32.f32 %r252, %f767;
|
|
cvt.sat.f32.f32 %f768, %f255;
|
|
mul.f32 %f769, %f768, 0f437FFD71;
|
|
cvt.rzi.u32.f32 %r253, %f769;
|
|
cvt.u16.u32 %rs19, %r251;
|
|
cvt.u16.u32 %rs20, %r253;
|
|
cvt.u16.u32 %rs21, %r252;
|
|
mov.u16 %rs22, 255;
|
|
st.v4.u8 [%rd63], {%rs19, %rs21, %rs20, %rs22};
|
|
ld.global.u32 %r450, [imageEnabled];
|
|
|
|
BB0_91:
|
|
cvt.u64.u32 %rd16, %r4;
|
|
cvt.u64.u32 %rd17, %r5;
|
|
and.b32 %r254, %r450, 4;
|
|
setp.eq.s32 %p92, %r254, 0;
|
|
@%p92 bra BB0_95;
|
|
|
|
ld.global.u32 %r255, [additive];
|
|
setp.eq.s32 %p93, %r255, 0;
|
|
mov.f32 %f770, 0f3F800000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs23, %f770;}
|
|
|
|
// inline asm
|
|
@%p93 bra BB0_94;
|
|
|
|
mov.u64 %rd82, image_HDR;
|
|
cvta.global.u64 %rd71, %rd82;
|
|
mov.u32 %r259, 8;
|
|
// inline asm
|
|
call (%rd70), _rt_buffer_get_64, (%rd71, %r92, %r259, %rd16, %rd17, %rd25, %rd25);
|
|
// inline asm
|
|
ld.v4.u16 {%rs30, %rs31, %rs32, %rs33}, [%rd70];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f771, %rs30;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f772, %rs31;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f773, %rs32;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd76), _rt_buffer_get_64, (%rd71, %r92, %r259, %rd16, %rd17, %rd25, %rd25);
|
|
// inline asm
|
|
add.f32 %f774, %f239, %f771;
|
|
add.f32 %f775, %f240, %f772;
|
|
add.f32 %f776, %f241, %f773;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs29, %f776;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs28, %f775;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs27, %f774;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd76], {%rs27, %rs28, %rs29, %rs23};
|
|
bra.uni BB0_95;
|
|
|
|
BB0_94:
|
|
mov.u64 %rd89, image_HDR;
|
|
cvta.global.u64 %rd84, %rd89;
|
|
mov.u32 %r261, 8;
|
|
// inline asm
|
|
call (%rd83), _rt_buffer_get_64, (%rd84, %r92, %r261, %rd16, %rd17, %rd25, %rd25);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs36, %f241;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs35, %f240;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs34, %f239;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd83], {%rs34, %rs35, %rs36, %rs23};
|
|
|
|
BB0_95:
|
|
ld.global.f32 %f781, [skyColor];
|
|
mul.f32 %f782, %f781, 0f40800000;
|
|
ld.global.f32 %f783, [skyColor+4];
|
|
mul.f32 %f784, %f783, 0f40800000;
|
|
ld.global.f32 %f785, [skyColor+8];
|
|
mul.f32 %f786, %f785, 0f40800000;
|
|
mul.f32 %f787, %f212, %f782;
|
|
mul.f32 %f788, %f213, %f784;
|
|
mul.f32 %f789, %f214, %f786;
|
|
mul.f32 %f281, %f787, 0f3F558750;
|
|
mul.f32 %f282, %f788, 0f3F558750;
|
|
mul.f32 %f283, %f789, 0f3F558750;
|
|
ld.global.u32 %r262, [additive];
|
|
setp.eq.s32 %p94, %r262, 0;
|
|
mov.f32 %f780, 0f3F800000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs37, %f780;}
|
|
|
|
// inline asm
|
|
@%p94 bra BB0_97;
|
|
|
|
mov.u64 %rd102, image_RNM0;
|
|
cvta.global.u64 %rd91, %rd102;
|
|
mov.u32 %r266, 8;
|
|
// inline asm
|
|
call (%rd90), _rt_buffer_get_64, (%rd91, %r92, %r266, %rd16, %rd17, %rd25, %rd25);
|
|
// inline asm
|
|
ld.v4.u16 {%rs44, %rs45, %rs46, %rs47}, [%rd90];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f790, %rs44;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f791, %rs45;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f792, %rs46;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd96), _rt_buffer_get_64, (%rd91, %r92, %r266, %rd16, %rd17, %rd25, %rd25);
|
|
// inline asm
|
|
add.f32 %f793, %f281, %f790;
|
|
add.f32 %f794, %f282, %f791;
|
|
add.f32 %f795, %f283, %f792;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs43, %f795;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs42, %f794;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs41, %f793;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd96], {%rs41, %rs42, %rs43, %rs37};
|
|
bra.uni BB0_98;
|
|
|
|
BB0_97:
|
|
mov.u64 %rd109, image_RNM0;
|
|
cvta.global.u64 %rd104, %rd109;
|
|
mov.u32 %r268, 8;
|
|
// inline asm
|
|
call (%rd103), _rt_buffer_get_64, (%rd104, %r92, %r268, %rd16, %rd17, %rd25, %rd25);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs50, %f283;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs49, %f282;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs48, %f281;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd103], {%rs48, %rs49, %rs50, %rs37};
|
|
|
|
BB0_98:
|
|
ld.global.f32 %f800, [skyColor];
|
|
mul.f32 %f801, %f800, 0f40800000;
|
|
ld.global.f32 %f802, [skyColor+4];
|
|
mul.f32 %f803, %f802, 0f40800000;
|
|
ld.global.f32 %f804, [skyColor+8];
|
|
mul.f32 %f805, %f804, 0f40800000;
|
|
mul.f32 %f806, %f215, %f801;
|
|
mul.f32 %f807, %f216, %f803;
|
|
mul.f32 %f808, %f217, %f805;
|
|
mul.f32 %f284, %f806, 0f3FB8EBD1;
|
|
mul.f32 %f285, %f807, 0f3FB8EBD1;
|
|
mul.f32 %f286, %f808, 0f3FB8EBD1;
|
|
ld.global.u32 %r269, [additive];
|
|
setp.eq.s32 %p95, %r269, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs51, %f780;}
|
|
|
|
// inline asm
|
|
@%p95 bra BB0_100;
|
|
|
|
mov.u64 %rd122, image_RNM1;
|
|
cvta.global.u64 %rd111, %rd122;
|
|
mov.u32 %r273, 8;
|
|
// inline asm
|
|
call (%rd110), _rt_buffer_get_64, (%rd111, %r92, %r273, %rd16, %rd17, %rd25, %rd25);
|
|
// inline asm
|
|
ld.v4.u16 {%rs58, %rs59, %rs60, %rs61}, [%rd110];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f809, %rs58;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f810, %rs59;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f811, %rs60;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd116), _rt_buffer_get_64, (%rd111, %r92, %r273, %rd16, %rd17, %rd25, %rd25);
|
|
// inline asm
|
|
add.f32 %f812, %f284, %f809;
|
|
add.f32 %f813, %f285, %f810;
|
|
add.f32 %f814, %f286, %f811;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs57, %f814;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs56, %f813;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs55, %f812;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd116], {%rs55, %rs56, %rs57, %rs51};
|
|
bra.uni BB0_101;
|
|
|
|
BB0_100:
|
|
mov.u64 %rd129, image_RNM1;
|
|
cvta.global.u64 %rd124, %rd129;
|
|
mov.u32 %r275, 8;
|
|
// inline asm
|
|
call (%rd123), _rt_buffer_get_64, (%rd124, %r92, %r275, %rd16, %rd17, %rd25, %rd25);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs64, %f286;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs63, %f285;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs62, %f284;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd123], {%rs62, %rs63, %rs64, %rs51};
|
|
|
|
BB0_101:
|
|
ld.global.f32 %f819, [skyColor];
|
|
mul.f32 %f820, %f819, 0f40800000;
|
|
ld.global.f32 %f821, [skyColor+4];
|
|
mul.f32 %f822, %f821, 0f40800000;
|
|
ld.global.f32 %f823, [skyColor+8];
|
|
mul.f32 %f824, %f823, 0f40800000;
|
|
mul.f32 %f825, %f218, %f820;
|
|
mul.f32 %f826, %f219, %f822;
|
|
mul.f32 %f827, %f220, %f824;
|
|
mul.f32 %f287, %f825, 0f3FB8EBD1;
|
|
mul.f32 %f288, %f826, 0f3FB8EBD1;
|
|
mul.f32 %f289, %f827, 0f3FB8EBD1;
|
|
ld.global.u32 %r276, [additive];
|
|
setp.eq.s32 %p96, %r276, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs65, %f780;}
|
|
|
|
// inline asm
|
|
@%p96 bra BB0_103;
|
|
|
|
mov.u64 %rd142, image_RNM2;
|
|
cvta.global.u64 %rd131, %rd142;
|
|
mov.u32 %r280, 8;
|
|
// inline asm
|
|
call (%rd130), _rt_buffer_get_64, (%rd131, %r92, %r280, %rd16, %rd17, %rd25, %rd25);
|
|
// inline asm
|
|
ld.v4.u16 {%rs72, %rs73, %rs74, %rs75}, [%rd130];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f828, %rs72;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f829, %rs73;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f830, %rs74;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd136), _rt_buffer_get_64, (%rd131, %r92, %r280, %rd16, %rd17, %rd25, %rd25);
|
|
// inline asm
|
|
add.f32 %f831, %f287, %f828;
|
|
add.f32 %f832, %f288, %f829;
|
|
add.f32 %f833, %f289, %f830;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs71, %f833;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs70, %f832;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs69, %f831;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd136], {%rs69, %rs70, %rs71, %rs65};
|
|
bra.uni BB0_104;
|
|
|
|
BB0_103:
|
|
mov.u64 %rd149, image_RNM2;
|
|
cvta.global.u64 %rd144, %rd149;
|
|
mov.u32 %r282, 8;
|
|
// inline asm
|
|
call (%rd143), _rt_buffer_get_64, (%rd144, %r92, %r282, %rd16, %rd17, %rd25, %rd25);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs78, %f289;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs77, %f288;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs76, %f287;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd143], {%rs76, %rs77, %rs78, %rs65};
|
|
|
|
BB0_104:
|
|
ld.global.f32 %f838, [skyColor];
|
|
mul.f32 %f839, %f838, 0f40800000;
|
|
ld.global.f32 %f840, [skyColor+4];
|
|
mul.f32 %f841, %f840, 0f40800000;
|
|
ld.global.f32 %f842, [skyColor+8];
|
|
mul.f32 %f843, %f842, 0f40800000;
|
|
mul.f32 %f844, %f221, %f839;
|
|
mul.f32 %f845, %f222, %f841;
|
|
mul.f32 %f846, %f223, %f843;
|
|
mul.f32 %f290, %f844, 0f3FB8EBD1;
|
|
mul.f32 %f291, %f845, 0f3FB8EBD1;
|
|
mul.f32 %f292, %f846, 0f3FB8EBD1;
|
|
ld.global.u32 %r283, [additive];
|
|
setp.eq.s32 %p97, %r283, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs79, %f780;}
|
|
|
|
// inline asm
|
|
@%p97 bra BB0_106;
|
|
|
|
mov.u64 %rd162, image_RNM3;
|
|
cvta.global.u64 %rd151, %rd162;
|
|
mov.u32 %r287, 8;
|
|
// inline asm
|
|
call (%rd150), _rt_buffer_get_64, (%rd151, %r92, %r287, %rd16, %rd17, %rd25, %rd25);
|
|
// inline asm
|
|
ld.v4.u16 {%rs86, %rs87, %rs88, %rs89}, [%rd150];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f847, %rs86;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f848, %rs87;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f849, %rs88;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd156), _rt_buffer_get_64, (%rd151, %r92, %r287, %rd16, %rd17, %rd25, %rd25);
|
|
// inline asm
|
|
add.f32 %f850, %f290, %f847;
|
|
add.f32 %f851, %f291, %f848;
|
|
add.f32 %f852, %f292, %f849;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs85, %f852;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs84, %f851;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs83, %f850;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd156], {%rs83, %rs84, %rs85, %rs79};
|
|
bra.uni BB0_107;
|
|
|
|
BB0_106:
|
|
mov.u64 %rd169, image_RNM3;
|
|
cvta.global.u64 %rd164, %rd169;
|
|
mov.u32 %r289, 8;
|
|
// inline asm
|
|
call (%rd163), _rt_buffer_get_64, (%rd164, %r92, %r289, %rd16, %rd17, %rd25, %rd25);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs92, %f292;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs91, %f291;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs90, %f290;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd163], {%rs90, %rs91, %rs92, %rs79};
|
|
|
|
BB0_107:
|
|
ld.global.f32 %f857, [skyColor];
|
|
mul.f32 %f858, %f857, 0f40800000;
|
|
ld.global.f32 %f859, [skyColor+4];
|
|
mul.f32 %f860, %f859, 0f40800000;
|
|
ld.global.f32 %f861, [skyColor+8];
|
|
mul.f32 %f862, %f861, 0f40800000;
|
|
mul.f32 %f863, %f224, %f858;
|
|
mul.f32 %f864, %f225, %f860;
|
|
mul.f32 %f865, %f226, %f862;
|
|
mul.f32 %f293, %f863, 0f404EBF87;
|
|
mul.f32 %f294, %f864, 0f404EBF87;
|
|
mul.f32 %f295, %f865, 0f404EBF87;
|
|
ld.global.u32 %r290, [additive];
|
|
setp.eq.s32 %p98, %r290, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs93, %f780;}
|
|
|
|
// inline asm
|
|
@%p98 bra BB0_109;
|
|
|
|
mov.u64 %rd182, image_RNM4;
|
|
cvta.global.u64 %rd171, %rd182;
|
|
mov.u32 %r294, 8;
|
|
// inline asm
|
|
call (%rd170), _rt_buffer_get_64, (%rd171, %r92, %r294, %rd16, %rd17, %rd25, %rd25);
|
|
// inline asm
|
|
ld.v4.u16 {%rs100, %rs101, %rs102, %rs103}, [%rd170];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f866, %rs100;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f867, %rs101;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f868, %rs102;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd176), _rt_buffer_get_64, (%rd171, %r92, %r294, %rd16, %rd17, %rd25, %rd25);
|
|
// inline asm
|
|
add.f32 %f869, %f293, %f866;
|
|
add.f32 %f870, %f294, %f867;
|
|
add.f32 %f871, %f295, %f868;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs99, %f871;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs98, %f870;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs97, %f869;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd176], {%rs97, %rs98, %rs99, %rs93};
|
|
bra.uni BB0_110;
|
|
|
|
BB0_109:
|
|
mov.u64 %rd189, image_RNM4;
|
|
cvta.global.u64 %rd184, %rd189;
|
|
mov.u32 %r296, 8;
|
|
// inline asm
|
|
call (%rd183), _rt_buffer_get_64, (%rd184, %r92, %r296, %rd16, %rd17, %rd25, %rd25);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs106, %f295;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs105, %f294;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs104, %f293;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd183], {%rs104, %rs105, %rs106, %rs93};
|
|
|
|
BB0_110:
|
|
ld.global.f32 %f876, [skyColor];
|
|
mul.f32 %f877, %f876, 0f40800000;
|
|
ld.global.f32 %f878, [skyColor+4];
|
|
mul.f32 %f879, %f878, 0f40800000;
|
|
ld.global.f32 %f880, [skyColor+8];
|
|
mul.f32 %f881, %f880, 0f40800000;
|
|
mul.f32 %f882, %f227, %f877;
|
|
mul.f32 %f883, %f228, %f879;
|
|
mul.f32 %f884, %f229, %f881;
|
|
mul.f32 %f296, %f882, 0f404EBF87;
|
|
mul.f32 %f297, %f883, 0f404EBF87;
|
|
mul.f32 %f298, %f884, 0f404EBF87;
|
|
ld.global.u32 %r297, [additive];
|
|
setp.eq.s32 %p99, %r297, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs107, %f780;}
|
|
|
|
// inline asm
|
|
@%p99 bra BB0_112;
|
|
|
|
mov.u64 %rd202, image_RNM5;
|
|
cvta.global.u64 %rd191, %rd202;
|
|
mov.u32 %r301, 8;
|
|
// inline asm
|
|
call (%rd190), _rt_buffer_get_64, (%rd191, %r92, %r301, %rd16, %rd17, %rd25, %rd25);
|
|
// inline asm
|
|
ld.v4.u16 {%rs114, %rs115, %rs116, %rs117}, [%rd190];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f885, %rs114;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f886, %rs115;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f887, %rs116;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd196), _rt_buffer_get_64, (%rd191, %r92, %r301, %rd16, %rd17, %rd25, %rd25);
|
|
// inline asm
|
|
add.f32 %f888, %f296, %f885;
|
|
add.f32 %f889, %f297, %f886;
|
|
add.f32 %f890, %f298, %f887;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs113, %f890;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs112, %f889;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs111, %f888;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd196], {%rs111, %rs112, %rs113, %rs107};
|
|
bra.uni BB0_113;
|
|
|
|
BB0_112:
|
|
mov.u64 %rd209, image_RNM5;
|
|
cvta.global.u64 %rd204, %rd209;
|
|
mov.u32 %r303, 8;
|
|
// inline asm
|
|
call (%rd203), _rt_buffer_get_64, (%rd204, %r92, %r303, %rd16, %rd17, %rd25, %rd25);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs120, %f298;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs119, %f297;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs118, %f296;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd203], {%rs118, %rs119, %rs120, %rs107};
|
|
|
|
BB0_113:
|
|
ld.global.f32 %f895, [skyColor];
|
|
mul.f32 %f896, %f895, 0f40800000;
|
|
ld.global.f32 %f897, [skyColor+4];
|
|
mul.f32 %f898, %f897, 0f40800000;
|
|
ld.global.f32 %f899, [skyColor+8];
|
|
mul.f32 %f900, %f899, 0f40800000;
|
|
mul.f32 %f901, %f230, %f896;
|
|
mul.f32 %f902, %f231, %f898;
|
|
mul.f32 %f903, %f232, %f900;
|
|
mul.f32 %f299, %f901, 0f40330C94;
|
|
mul.f32 %f300, %f902, 0f40330C94;
|
|
mul.f32 %f301, %f903, 0f40330C94;
|
|
ld.global.u32 %r304, [additive];
|
|
setp.eq.s32 %p100, %r304, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs121, %f780;}
|
|
|
|
// inline asm
|
|
@%p100 bra BB0_115;
|
|
|
|
mov.u64 %rd222, image_RNM6;
|
|
cvta.global.u64 %rd211, %rd222;
|
|
mov.u32 %r308, 8;
|
|
// inline asm
|
|
call (%rd210), _rt_buffer_get_64, (%rd211, %r92, %r308, %rd16, %rd17, %rd25, %rd25);
|
|
// inline asm
|
|
ld.v4.u16 {%rs128, %rs129, %rs130, %rs131}, [%rd210];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f904, %rs128;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f905, %rs129;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f906, %rs130;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd216), _rt_buffer_get_64, (%rd211, %r92, %r308, %rd16, %rd17, %rd25, %rd25);
|
|
// inline asm
|
|
add.f32 %f907, %f299, %f904;
|
|
add.f32 %f908, %f300, %f905;
|
|
add.f32 %f909, %f301, %f906;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs127, %f909;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs126, %f908;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs125, %f907;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd216], {%rs125, %rs126, %rs127, %rs121};
|
|
bra.uni BB0_116;
|
|
|
|
BB0_115:
|
|
mov.u64 %rd229, image_RNM6;
|
|
cvta.global.u64 %rd224, %rd229;
|
|
mov.u32 %r310, 8;
|
|
// inline asm
|
|
call (%rd223), _rt_buffer_get_64, (%rd224, %r92, %r310, %rd16, %rd17, %rd25, %rd25);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs134, %f301;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs133, %f300;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs132, %f299;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd223], {%rs132, %rs133, %rs134, %rs121};
|
|
|
|
BB0_116:
|
|
ld.global.f32 %f914, [skyColor];
|
|
mul.f32 %f915, %f914, 0f40800000;
|
|
ld.global.f32 %f916, [skyColor+4];
|
|
mul.f32 %f917, %f916, 0f40800000;
|
|
ld.global.f32 %f918, [skyColor+8];
|
|
mul.f32 %f919, %f918, 0f40800000;
|
|
mul.f32 %f920, %f233, %f915;
|
|
mul.f32 %f921, %f234, %f917;
|
|
mul.f32 %f922, %f235, %f919;
|
|
mul.f32 %f302, %f920, 0f404EBF87;
|
|
mul.f32 %f303, %f921, 0f404EBF87;
|
|
mul.f32 %f304, %f922, 0f404EBF87;
|
|
ld.global.u32 %r311, [additive];
|
|
setp.eq.s32 %p101, %r311, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs135, %f780;}
|
|
|
|
// inline asm
|
|
@%p101 bra BB0_118;
|
|
|
|
mov.u64 %rd242, image_RNM7;
|
|
cvta.global.u64 %rd231, %rd242;
|
|
mov.u32 %r315, 8;
|
|
// inline asm
|
|
call (%rd230), _rt_buffer_get_64, (%rd231, %r92, %r315, %rd16, %rd17, %rd25, %rd25);
|
|
// inline asm
|
|
ld.v4.u16 {%rs142, %rs143, %rs144, %rs145}, [%rd230];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f923, %rs142;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f924, %rs143;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f925, %rs144;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd236), _rt_buffer_get_64, (%rd231, %r92, %r315, %rd16, %rd17, %rd25, %rd25);
|
|
// inline asm
|
|
add.f32 %f926, %f302, %f923;
|
|
add.f32 %f927, %f303, %f924;
|
|
add.f32 %f928, %f304, %f925;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs141, %f928;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs140, %f927;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs139, %f926;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd236], {%rs139, %rs140, %rs141, %rs135};
|
|
bra.uni BB0_119;
|
|
|
|
BB0_118:
|
|
mov.u64 %rd249, image_RNM7;
|
|
cvta.global.u64 %rd244, %rd249;
|
|
mov.u32 %r317, 8;
|
|
// inline asm
|
|
call (%rd243), _rt_buffer_get_64, (%rd244, %r92, %r317, %rd16, %rd17, %rd25, %rd25);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs148, %f304;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs147, %f303;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs146, %f302;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd243], {%rs146, %rs147, %rs148, %rs135};
|
|
|
|
BB0_119:
|
|
ld.global.f32 %f933, [skyColor];
|
|
mul.f32 %f934, %f933, 0f40800000;
|
|
ld.global.f32 %f935, [skyColor+4];
|
|
mul.f32 %f936, %f935, 0f40800000;
|
|
ld.global.f32 %f937, [skyColor+8];
|
|
mul.f32 %f938, %f937, 0f40800000;
|
|
mul.f32 %f939, %f236, %f934;
|
|
mul.f32 %f940, %f237, %f936;
|
|
mul.f32 %f941, %f238, %f938;
|
|
mul.f32 %f305, %f939, 0f3FCEBF87;
|
|
mul.f32 %f306, %f940, 0f3FCEBF87;
|
|
mul.f32 %f307, %f941, 0f3FCEBF87;
|
|
ld.global.u32 %r318, [additive];
|
|
setp.eq.s32 %p102, %r318, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs149, %f780;}
|
|
|
|
// inline asm
|
|
@%p102 bra BB0_121;
|
|
|
|
mov.u64 %rd262, image_RNM8;
|
|
cvta.global.u64 %rd251, %rd262;
|
|
mov.u32 %r322, 8;
|
|
// inline asm
|
|
call (%rd250), _rt_buffer_get_64, (%rd251, %r92, %r322, %rd16, %rd17, %rd25, %rd25);
|
|
// inline asm
|
|
ld.v4.u16 {%rs156, %rs157, %rs158, %rs159}, [%rd250];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f942, %rs156;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f943, %rs157;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f944, %rs158;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd256), _rt_buffer_get_64, (%rd251, %r92, %r322, %rd16, %rd17, %rd25, %rd25);
|
|
// inline asm
|
|
add.f32 %f945, %f305, %f942;
|
|
add.f32 %f946, %f306, %f943;
|
|
add.f32 %f947, %f307, %f944;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs155, %f947;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs154, %f946;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs153, %f945;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd256], {%rs153, %rs154, %rs155, %rs149};
|
|
bra.uni BB0_155;
|
|
|
|
BB0_121:
|
|
mov.u64 %rd269, image_RNM8;
|
|
cvta.global.u64 %rd264, %rd269;
|
|
mov.u32 %r324, 8;
|
|
// inline asm
|
|
call (%rd263), _rt_buffer_get_64, (%rd264, %r92, %r324, %rd16, %rd17, %rd25, %rd25);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs162, %f307;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs161, %f306;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs160, %f305;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd263], {%rs160, %rs161, %rs162, %rs149};
|
|
|
|
BB0_155:
|
|
ret;
|
|
}
|
|
|
|
|