2067 lines
56 KiB
Plaintext
2067 lines
56 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 uvpos[1];
|
||
|
|
.global .align 1 .b8 uvnormal[1];
|
||
|
|
.global .align 1 .b8 rnd_seeds[1];
|
||
|
|
.global .align 4 .u32 sky;
|
||
|
|
.global .align 4 .b8 skyColor[12];
|
||
|
|
.global .align 4 .u32 samples;
|
||
|
|
.global .align 4 .u32 hemispherical;
|
||
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo7pixelIDE[8] = {82, 97, 121, 0, 8, 0, 0, 0};
|
||
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo10resolutionE[8] = {82, 97, 121, 0, 8, 0, 0, 0};
|
||
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo6normalE[8] = {82, 97, 121, 0, 12, 0, 0, 0};
|
||
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo6camPosE[8] = {82, 97, 121, 0, 12, 0, 0, 0};
|
||
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo4rootE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
|
||
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo12imageEnabledE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
|
||
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo8tileInfoE[8] = {82, 97, 121, 0, 16, 0, 0, 0};
|
||
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo8additiveE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
|
||
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo3skyE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
|
||
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo8skyColorE[8] = {82, 97, 121, 0, 12, 0, 0, 0};
|
||
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo7samplesE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
|
||
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo13hemisphericalE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
|
||
|
|
.global .align 8 .u64 _ZN21rti_internal_register20reg_bitness_detectorE;
|
||
|
|
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail0E;
|
||
|
|
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail1E;
|
||
|
|
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail2E;
|
||
|
|
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail3E;
|
||
|
|
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail4E;
|
||
|
|
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail5E;
|
||
|
|
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail6E;
|
||
|
|
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail7E;
|
||
|
|
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail8E;
|
||
|
|
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail9E;
|
||
|
|
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail0E;
|
||
|
|
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail1E;
|
||
|
|
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail2E;
|
||
|
|
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail3E;
|
||
|
|
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail4E;
|
||
|
|
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail5E;
|
||
|
|
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail6E;
|
||
|
|
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail7E;
|
||
|
|
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail8E;
|
||
|
|
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail9E;
|
||
|
|
.global .align 4 .u32 _ZN21rti_internal_register14reg_rayIndex_xE;
|
||
|
|
.global .align 4 .u32 _ZN21rti_internal_register14reg_rayIndex_yE;
|
||
|
|
.global .align 4 .u32 _ZN21rti_internal_register14reg_rayIndex_zE;
|
||
|
|
.global .align 8 .b8 _ZN21rti_internal_typename7pixelIDE[6] = {117, 105, 110, 116, 50, 0};
|
||
|
|
.global .align 8 .b8 _ZN21rti_internal_typename10resolutionE[6] = {117, 105, 110, 116, 50, 0};
|
||
|
|
.global .align 8 .b8 _ZN21rti_internal_typename6normalE[7] = {102, 108, 111, 97, 116, 51, 0};
|
||
|
|
.global .align 8 .b8 _ZN21rti_internal_typename6camPosE[7] = {102, 108, 111, 97, 116, 51, 0};
|
||
|
|
.global .align 16 .b8 _ZN21rti_internal_typename4rootE[9] = {114, 116, 79, 98, 106, 101, 99, 116, 0};
|
||
|
|
.global .align 4 .b8 _ZN21rti_internal_typename12imageEnabledE[4] = {105, 110, 116, 0};
|
||
|
|
.global .align 8 .b8 _ZN21rti_internal_typename8tileInfoE[6] = {117, 105, 110, 116, 52, 0};
|
||
|
|
.global .align 4 .b8 _ZN21rti_internal_typename8additiveE[4] = {105, 110, 116, 0};
|
||
|
|
.global .align 4 .b8 _ZN21rti_internal_typename3skyE[4] = {105, 110, 116, 0};
|
||
|
|
.global .align 8 .b8 _ZN21rti_internal_typename8skyColorE[7] = {102, 108, 111, 97, 116, 51, 0};
|
||
|
|
.global .align 4 .b8 _ZN21rti_internal_typename7samplesE[4] = {105, 110, 116, 0};
|
||
|
|
.global .align 4 .b8 _ZN21rti_internal_typename13hemisphericalE[4] = {105, 110, 116, 0};
|
||
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum7pixelIDE = 4919;
|
||
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum10resolutionE = 4919;
|
||
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum6normalE = 4919;
|
||
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum6camPosE = 4919;
|
||
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum4rootE = 4919;
|
||
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum12imageEnabledE = 4919;
|
||
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum8tileInfoE = 4919;
|
||
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum8additiveE = 4919;
|
||
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum3skyE = 4919;
|
||
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum8skyColorE = 4919;
|
||
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum7samplesE = 4919;
|
||
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum13hemisphericalE = 4919;
|
||
|
|
.global .align 16 .b8 _ZN21rti_internal_semantic7pixelIDE[14] = {114, 116, 76, 97, 117, 110, 99, 104, 73, 110, 100, 101, 120, 0};
|
||
|
|
.global .align 16 .b8 _ZN21rti_internal_semantic10resolutionE[12] = {114, 116, 76, 97, 117, 110, 99, 104, 68, 105, 109, 0};
|
||
|
|
.global .align 16 .b8 _ZN21rti_internal_semantic6normalE[17] = {97, 116, 116, 114, 105, 98, 117, 116, 101, 32, 110, 111, 114, 109, 97, 108, 0};
|
||
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic6camPosE[1];
|
||
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic4rootE[1];
|
||
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic12imageEnabledE[1];
|
||
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic8tileInfoE[1];
|
||
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic8additiveE[1];
|
||
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic3skyE[1];
|
||
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic8skyColorE[1];
|
||
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic7samplesE[1];
|
||
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic13hemisphericalE[1];
|
||
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation7pixelIDE[1];
|
||
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation10resolutionE[1];
|
||
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation6normalE[1];
|
||
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation6camPosE[1];
|
||
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation4rootE[1];
|
||
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation12imageEnabledE[1];
|
||
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation8tileInfoE[1];
|
||
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation8additiveE[1];
|
||
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation3skyE[1];
|
||
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation8skyColorE[1];
|
||
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation7samplesE[1];
|
||
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation13hemisphericalE[1];
|
||
|
|
.const .align 4 .b8 __cudart_i2opi_f[24] = {65, 144, 67, 60, 153, 149, 98, 219, 192, 221, 52, 245, 209, 87, 39, 252, 41, 21, 68, 78, 110, 131, 249, 162};
|
||
|
|
|
||
|
|
.visible .entry _Z6oxMainv(
|
||
|
|
|
||
|
|
)
|
||
|
|
{
|
||
|
|
.local .align 4 .b8 __local_depot0[40];
|
||
|
|
.reg .b64 %SP;
|
||
|
|
.reg .b64 %SPL;
|
||
|
|
.reg .pred %p<106>;
|
||
|
|
.reg .b16 %rs<154>;
|
||
|
|
.reg .f32 %f<897>;
|
||
|
|
.reg .b32 %r<384>;
|
||
|
|
.reg .b64 %rd<288>;
|
||
|
|
|
||
|
|
|
||
|
|
mov.u64 %rd287, __local_depot0;
|
||
|
|
cvta.local.u64 %SP, %rd287;
|
||
|
|
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 %rs7, %rs1, 255;
|
||
|
|
cvt.u16.u32 %rs8, %r3;
|
||
|
|
or.b16 %rs9, %rs8, %rs7;
|
||
|
|
setp.eq.s16 %p4, %rs9, 0;
|
||
|
|
mov.f32 %f813, 0f00000000;
|
||
|
|
mov.f32 %f814, %f813;
|
||
|
|
mov.f32 %f815, %f813;
|
||
|
|
@%p4 bra BB0_2;
|
||
|
|
|
||
|
|
ld.u8 %rs10, [%rd20+1];
|
||
|
|
and.b16 %rs12, %rs8, 255;
|
||
|
|
cvt.rn.f32.u16 %f215, %rs12;
|
||
|
|
div.rn.f32 %f216, %f215, 0f437F0000;
|
||
|
|
fma.rn.f32 %f217, %f216, 0f40000000, 0fBF800000;
|
||
|
|
cvt.rn.f32.u16 %f218, %rs10;
|
||
|
|
div.rn.f32 %f219, %f218, 0f437F0000;
|
||
|
|
fma.rn.f32 %f220, %f219, 0f40000000, 0fBF800000;
|
||
|
|
cvt.rn.f32.u16 %f221, %rs7;
|
||
|
|
div.rn.f32 %f222, %f221, 0f437F0000;
|
||
|
|
fma.rn.f32 %f223, %f222, 0f40000000, 0fBF800000;
|
||
|
|
mul.f32 %f224, %f220, %f220;
|
||
|
|
fma.rn.f32 %f225, %f217, %f217, %f224;
|
||
|
|
fma.rn.f32 %f226, %f223, %f223, %f225;
|
||
|
|
sqrt.rn.f32 %f227, %f226;
|
||
|
|
rcp.rn.f32 %f228, %f227;
|
||
|
|
mul.f32 %f813, %f217, %f228;
|
||
|
|
mul.f32 %f814, %f220, %f228;
|
||
|
|
mul.f32 %f815, %f223, %f228;
|
||
|
|
|
||
|
|
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, %f814, 0f00000000;
|
||
|
|
setp.eq.f32 %p6, %f813, 0f00000000;
|
||
|
|
and.pred %p7, %p6, %p5;
|
||
|
|
setp.eq.f32 %p8, %f815, 0f00000000;
|
||
|
|
and.pred %p9, %p7, %p8;
|
||
|
|
@%p9 bra BB0_107;
|
||
|
|
bra.uni BB0_3;
|
||
|
|
|
||
|
|
BB0_107:
|
||
|
|
ld.global.u32 %r383, [imageEnabled];
|
||
|
|
and.b32 %r292, %r383, 1;
|
||
|
|
setp.eq.b32 %p98, %r292, 1;
|
||
|
|
@!%p98 bra BB0_109;
|
||
|
|
bra.uni BB0_108;
|
||
|
|
|
||
|
|
BB0_108:
|
||
|
|
cvt.u64.u32 %rd173, %r5;
|
||
|
|
cvt.u64.u32 %rd172, %r4;
|
||
|
|
mov.u64 %rd176, image;
|
||
|
|
cvta.global.u64 %rd171, %rd176;
|
||
|
|
mov.u64 %rd175, 0;
|
||
|
|
// inline asm
|
||
|
|
call (%rd170), _rt_buffer_get_64, (%rd171, %r92, %r93, %rd172, %rd173, %rd175, %rd175);
|
||
|
|
// inline asm
|
||
|
|
mov.u16 %rs88, 0;
|
||
|
|
st.v4.u8 [%rd170], {%rs88, %rs88, %rs88, %rs88};
|
||
|
|
ld.global.u32 %r383, [imageEnabled];
|
||
|
|
|
||
|
|
BB0_109:
|
||
|
|
cvt.u64.u32 %rd18, %r4;
|
||
|
|
cvt.u64.u32 %rd19, %r5;
|
||
|
|
and.b32 %r295, %r383, 4;
|
||
|
|
setp.eq.s32 %p99, %r295, 0;
|
||
|
|
@%p99 bra BB0_113;
|
||
|
|
|
||
|
|
ld.global.u32 %r296, [additive];
|
||
|
|
setp.eq.s32 %p100, %r296, 0;
|
||
|
|
@%p100 bra BB0_112;
|
||
|
|
|
||
|
|
mov.u64 %rd189, image_HDR;
|
||
|
|
cvta.global.u64 %rd178, %rd189;
|
||
|
|
mov.u32 %r300, 8;
|
||
|
|
mov.u64 %rd188, 0;
|
||
|
|
// inline asm
|
||
|
|
call (%rd177), _rt_buffer_get_64, (%rd178, %r92, %r300, %rd18, %rd19, %rd188, %rd188);
|
||
|
|
// inline asm
|
||
|
|
ld.v4.u16 {%rs95, %rs96, %rs97, %rs98}, [%rd177];
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f722, %rs95;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f723, %rs96;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f724, %rs97;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
call (%rd183), _rt_buffer_get_64, (%rd178, %r92, %r300, %rd18, %rd19, %rd188, %rd188);
|
||
|
|
// inline asm
|
||
|
|
add.f32 %f725, %f722, 0f00000000;
|
||
|
|
add.f32 %f726, %f723, 0f00000000;
|
||
|
|
add.f32 %f727, %f724, 0f00000000;
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs94, %f727;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs93, %f726;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs92, %f725;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
mov.u16 %rs99, 0;
|
||
|
|
st.v4.u16 [%rd183], {%rs92, %rs93, %rs94, %rs99};
|
||
|
|
bra.uni BB0_113;
|
||
|
|
|
||
|
|
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 %f244, [%rd27+8];
|
||
|
|
ld.f32 %f245, [%rd27+4];
|
||
|
|
ld.f32 %f246, [%rd27];
|
||
|
|
mul.f32 %f247, %f246, 0f3456BF95;
|
||
|
|
mul.f32 %f248, %f245, 0f3456BF95;
|
||
|
|
mul.f32 %f249, %f244, 0f3456BF95;
|
||
|
|
abs.f32 %f250, %f813;
|
||
|
|
div.rn.f32 %f251, %f247, %f250;
|
||
|
|
abs.f32 %f252, %f814;
|
||
|
|
div.rn.f32 %f253, %f248, %f252;
|
||
|
|
abs.f32 %f254, %f815;
|
||
|
|
div.rn.f32 %f255, %f249, %f254;
|
||
|
|
abs.f32 %f256, %f251;
|
||
|
|
abs.f32 %f257, %f253;
|
||
|
|
abs.f32 %f258, %f255;
|
||
|
|
mov.f32 %f259, 0f38D1B717;
|
||
|
|
max.f32 %f260, %f256, %f259;
|
||
|
|
max.f32 %f261, %f257, %f259;
|
||
|
|
max.f32 %f262, %f258, %f259;
|
||
|
|
fma.rn.f32 %f7, %f813, %f260, %f246;
|
||
|
|
fma.rn.f32 %f8, %f814, %f261, %f245;
|
||
|
|
fma.rn.f32 %f9, %f815, %f262, %f244;
|
||
|
|
ld.global.u32 %r6, [hemispherical];
|
||
|
|
setp.gt.f32 %p10, %f250, %f254;
|
||
|
|
neg.f32 %f263, %f814;
|
||
|
|
selp.f32 %f264, %f263, 0f00000000, %p10;
|
||
|
|
neg.f32 %f265, %f815;
|
||
|
|
selp.f32 %f266, %f813, %f265, %p10;
|
||
|
|
selp.f32 %f267, 0f00000000, %f814, %p10;
|
||
|
|
mul.f32 %f268, %f266, %f266;
|
||
|
|
fma.rn.f32 %f269, %f264, %f264, %f268;
|
||
|
|
fma.rn.f32 %f270, %f267, %f267, %f269;
|
||
|
|
sqrt.rn.f32 %f271, %f270;
|
||
|
|
rcp.rn.f32 %f272, %f271;
|
||
|
|
mul.f32 %f10, %f264, %f272;
|
||
|
|
mul.f32 %f11, %f266, %f272;
|
||
|
|
mul.f32 %f12, %f267, %f272;
|
||
|
|
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 %f872, 0f00000000;
|
||
|
|
setp.lt.s32 %p11, %r1, 1;
|
||
|
|
mov.f32 %f871, %f872;
|
||
|
|
mov.f32 %f870, %f872;
|
||
|
|
mov.f32 %f869, %f872;
|
||
|
|
mov.f32 %f868, %f872;
|
||
|
|
mov.f32 %f867, %f872;
|
||
|
|
mov.f32 %f866, %f872;
|
||
|
|
mov.f32 %f865, %f872;
|
||
|
|
mov.f32 %f864, %f872;
|
||
|
|
mov.f32 %f863, %f872;
|
||
|
|
mov.f32 %f862, %f872;
|
||
|
|
mov.f32 %f861, %f872;
|
||
|
|
mov.f32 %f860, %f872;
|
||
|
|
mov.f32 %f859, %f872;
|
||
|
|
mov.f32 %f858, %f872;
|
||
|
|
@%p11 bra BB0_56;
|
||
|
|
|
||
|
|
cvt.rn.f32.s32 %f288, %r2;
|
||
|
|
rcp.rn.f32 %f13, %f288;
|
||
|
|
ld.u32 %r361, [%rd33];
|
||
|
|
mul.f32 %f14, %f7, 0f3456BF95;
|
||
|
|
mul.f32 %f15, %f8, 0f3456BF95;
|
||
|
|
mul.f32 %f16, %f9, 0f3456BF95;
|
||
|
|
mul.f32 %f289, %f813, %f11;
|
||
|
|
mul.f32 %f290, %f814, %f10;
|
||
|
|
sub.f32 %f17, %f290, %f289;
|
||
|
|
mul.f32 %f291, %f815, %f10;
|
||
|
|
mul.f32 %f292, %f813, %f12;
|
||
|
|
sub.f32 %f18, %f292, %f291;
|
||
|
|
mul.f32 %f293, %f814, %f12;
|
||
|
|
mul.f32 %f294, %f815, %f11;
|
||
|
|
sub.f32 %f19, %f294, %f293;
|
||
|
|
mov.f32 %f872, 0f00000000;
|
||
|
|
mov.u32 %r119, 0;
|
||
|
|
abs.f32 %f376, %f15;
|
||
|
|
abs.f32 %f377, %f14;
|
||
|
|
max.f32 %f378, %f377, %f376;
|
||
|
|
abs.f32 %f379, %f16;
|
||
|
|
max.f32 %f380, %f378, %f379;
|
||
|
|
mov.u32 %r358, %r119;
|
||
|
|
mov.f32 %f871, %f872;
|
||
|
|
mov.f32 %f870, %f872;
|
||
|
|
mov.f32 %f869, %f872;
|
||
|
|
mov.f32 %f868, %f872;
|
||
|
|
mov.f32 %f867, %f872;
|
||
|
|
mov.f32 %f866, %f872;
|
||
|
|
mov.f32 %f865, %f872;
|
||
|
|
mov.f32 %f864, %f872;
|
||
|
|
mov.f32 %f863, %f872;
|
||
|
|
mov.f32 %f862, %f872;
|
||
|
|
mov.f32 %f861, %f872;
|
||
|
|
mov.f32 %f860, %f872;
|
||
|
|
mov.f32 %f859, %f872;
|
||
|
|
mov.f32 %f858, %f872;
|
||
|
|
|
||
|
|
BB0_5:
|
||
|
|
mov.u32 %r360, %r119;
|
||
|
|
|
||
|
|
BB0_6:
|
||
|
|
mov.u32 %r11, %r361;
|
||
|
|
cvt.rn.f32.s32 %f763, %r358;
|
||
|
|
mad.lo.s32 %r121, %r11, 1664525, 1013904223;
|
||
|
|
and.b32 %r122, %r121, 16777215;
|
||
|
|
cvt.rn.f32.u32 %f295, %r122;
|
||
|
|
fma.rn.f32 %f296, %f295, 0f33800000, %f763;
|
||
|
|
mul.f32 %f297, %f13, %f296;
|
||
|
|
mad.lo.s32 %r12, %r121, 1664525, 1013904223;
|
||
|
|
and.b32 %r123, %r12, 16777215;
|
||
|
|
cvt.rn.f32.u32 %f298, %r123;
|
||
|
|
cvt.rn.f32.s32 %f299, %r360;
|
||
|
|
fma.rn.f32 %f300, %f298, 0f33800000, %f299;
|
||
|
|
mul.f32 %f301, %f13, %f300;
|
||
|
|
fma.rn.f32 %f51, %f297, 0fC0000000, 0f3F800000;
|
||
|
|
mul.f32 %f302, %f51, %f51;
|
||
|
|
mov.f32 %f303, 0f3F800000;
|
||
|
|
sub.f32 %f304, %f303, %f302;
|
||
|
|
mov.f32 %f305, 0f00000000;
|
||
|
|
max.f32 %f306, %f305, %f304;
|
||
|
|
sqrt.rn.f32 %f52, %f306;
|
||
|
|
mul.f32 %f852, %f301, 0f40C90FDB;
|
||
|
|
abs.f32 %f54, %f852;
|
||
|
|
setp.neu.f32 %p12, %f54, 0f7F800000;
|
||
|
|
mov.f32 %f846, %f852;
|
||
|
|
@%p12 bra BB0_8;
|
||
|
|
|
||
|
|
mul.rn.f32 %f846, %f852, %f305;
|
||
|
|
|
||
|
|
BB0_8:
|
||
|
|
mul.f32 %f308, %f846, 0f3F22F983;
|
||
|
|
cvt.rni.s32.f32 %r371, %f308;
|
||
|
|
cvt.rn.f32.s32 %f309, %r371;
|
||
|
|
neg.f32 %f310, %f309;
|
||
|
|
mov.f32 %f311, 0f3FC90FDA;
|
||
|
|
fma.rn.f32 %f312, %f310, %f311, %f846;
|
||
|
|
mov.f32 %f313, 0f33A22168;
|
||
|
|
fma.rn.f32 %f314, %f310, %f313, %f312;
|
||
|
|
mov.f32 %f315, 0f27C234C5;
|
||
|
|
fma.rn.f32 %f847, %f310, %f315, %f314;
|
||
|
|
abs.f32 %f316, %f846;
|
||
|
|
setp.leu.f32 %p13, %f316, 0f47CE4780;
|
||
|
|
@%p13 bra BB0_19;
|
||
|
|
|
||
|
|
add.u64 %rd42, %SP, 12;
|
||
|
|
cvta.to.local.u64 %rd283, %rd42;
|
||
|
|
mov.u32 %r362, 0;
|
||
|
|
mov.u64 %rd284, 0;
|
||
|
|
mov.u32 %r363, %r362;
|
||
|
|
|
||
|
|
BB0_10:
|
||
|
|
.pragma "nounroll";
|
||
|
|
mov.b32 %r333, %f846;
|
||
|
|
shl.b32 %r332, %r333, 8;
|
||
|
|
or.b32 %r331, %r332, -2147483648;
|
||
|
|
add.u64 %rd278, %SP, 12;
|
||
|
|
cvta.to.local.u64 %rd277, %rd278;
|
||
|
|
shl.b64 %rd43, %rd284, 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, %r331, %r363;
|
||
|
|
madc.hi.u32 %r363, %r129, %r331, 0;
|
||
|
|
}
|
||
|
|
// inline asm
|
||
|
|
st.local.u32 [%rd283], %r127;
|
||
|
|
add.s32 %r362, %r362, 1;
|
||
|
|
cvt.s64.s32 %rd284, %r362;
|
||
|
|
mul.wide.s32 %rd48, %r362, 4;
|
||
|
|
add.s64 %rd283, %rd277, %rd48;
|
||
|
|
setp.ne.s32 %p14, %r362, 6;
|
||
|
|
@%p14 bra BB0_10;
|
||
|
|
|
||
|
|
mov.b32 %r335, %f846;
|
||
|
|
shr.u32 %r334, %r335, 23;
|
||
|
|
add.u64 %rd279, %SP, 12;
|
||
|
|
and.b32 %r132, %r334, 255;
|
||
|
|
add.s32 %r133, %r132, -128;
|
||
|
|
shr.u32 %r134, %r133, 5;
|
||
|
|
cvta.to.local.u64 %rd50, %rd279;
|
||
|
|
st.local.u32 [%rd50+24], %r363;
|
||
|
|
mov.u32 %r135, 6;
|
||
|
|
sub.s32 %r136, %r135, %r134;
|
||
|
|
mul.wide.s32 %rd51, %r136, 4;
|
||
|
|
add.s64 %rd8, %rd50, %rd51;
|
||
|
|
ld.local.u32 %r364, [%rd8];
|
||
|
|
ld.local.u32 %r365, [%rd8+-4];
|
||
|
|
and.b32 %r24, %r334, 31;
|
||
|
|
setp.eq.s32 %p15, %r24, 0;
|
||
|
|
@%p15 bra BB0_13;
|
||
|
|
|
||
|
|
mov.u32 %r137, 32;
|
||
|
|
sub.s32 %r138, %r137, %r24;
|
||
|
|
shr.u32 %r139, %r365, %r138;
|
||
|
|
shl.b32 %r140, %r364, %r24;
|
||
|
|
add.s32 %r364, %r139, %r140;
|
||
|
|
ld.local.u32 %r141, [%rd8+-8];
|
||
|
|
shr.u32 %r142, %r141, %r138;
|
||
|
|
shl.b32 %r143, %r365, %r24;
|
||
|
|
add.s32 %r365, %r142, %r143;
|
||
|
|
|
||
|
|
BB0_13:
|
||
|
|
mov.b32 %r343, %f846;
|
||
|
|
and.b32 %r367, %r343, -2147483648;
|
||
|
|
shr.u32 %r144, %r365, 30;
|
||
|
|
shl.b32 %r145, %r364, 2;
|
||
|
|
add.s32 %r366, %r144, %r145;
|
||
|
|
shl.b32 %r30, %r365, 2;
|
||
|
|
shr.u32 %r146, %r366, 31;
|
||
|
|
shr.u32 %r147, %r364, 30;
|
||
|
|
add.s32 %r31, %r146, %r147;
|
||
|
|
setp.eq.s32 %p16, %r146, 0;
|
||
|
|
@%p16 bra BB0_14;
|
||
|
|
bra.uni BB0_15;
|
||
|
|
|
||
|
|
BB0_14:
|
||
|
|
mov.u32 %r368, %r30;
|
||
|
|
bra.uni BB0_16;
|
||
|
|
|
||
|
|
BB0_15:
|
||
|
|
mov.b32 %r345, %f846;
|
||
|
|
and.b32 %r344, %r345, -2147483648;
|
||
|
|
not.b32 %r148, %r366;
|
||
|
|
neg.s32 %r368, %r30;
|
||
|
|
setp.eq.s32 %p17, %r30, 0;
|
||
|
|
selp.u32 %r149, 1, 0, %p17;
|
||
|
|
add.s32 %r366, %r149, %r148;
|
||
|
|
xor.b32 %r367, %r344, -2147483648;
|
||
|
|
|
||
|
|
BB0_16:
|
||
|
|
mov.b32 %r347, %f846;
|
||
|
|
and.b32 %r346, %r347, -2147483648;
|
||
|
|
clz.b32 %r370, %r366;
|
||
|
|
setp.eq.s32 %p18, %r370, 0;
|
||
|
|
shl.b32 %r150, %r366, %r370;
|
||
|
|
mov.u32 %r151, 32;
|
||
|
|
sub.s32 %r152, %r151, %r370;
|
||
|
|
shr.u32 %r153, %r368, %r152;
|
||
|
|
add.s32 %r154, %r153, %r150;
|
||
|
|
selp.b32 %r39, %r366, %r154, %p18;
|
||
|
|
mov.u32 %r155, -921707870;
|
||
|
|
mul.hi.u32 %r369, %r39, %r155;
|
||
|
|
setp.eq.s32 %p19, %r346, 0;
|
||
|
|
neg.s32 %r156, %r31;
|
||
|
|
selp.b32 %r371, %r31, %r156, %p19;
|
||
|
|
setp.lt.s32 %p20, %r369, 1;
|
||
|
|
@%p20 bra BB0_18;
|
||
|
|
|
||
|
|
mul.lo.s32 %r157, %r39, -921707870;
|
||
|
|
shr.u32 %r158, %r157, 31;
|
||
|
|
shl.b32 %r159, %r369, 1;
|
||
|
|
add.s32 %r369, %r158, %r159;
|
||
|
|
add.s32 %r370, %r370, 1;
|
||
|
|
|
||
|
|
BB0_18:
|
||
|
|
mov.u32 %r160, 126;
|
||
|
|
sub.s32 %r161, %r160, %r370;
|
||
|
|
shl.b32 %r162, %r161, 23;
|
||
|
|
add.s32 %r163, %r369, 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, %r367;
|
||
|
|
mov.b32 %f847, %r168;
|
||
|
|
|
||
|
|
BB0_19:
|
||
|
|
add.s32 %r47, %r371, 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 %f773, %f847, %f847;
|
||
|
|
mov.f32 %f319, 0f3C08839E;
|
||
|
|
mov.f32 %f320, 0fB94CA1F9;
|
||
|
|
fma.rn.f32 %f848, %f320, %f773, %f319;
|
||
|
|
bra.uni BB0_22;
|
||
|
|
|
||
|
|
BB0_20:
|
||
|
|
mul.rn.f32 %f769, %f847, %f847;
|
||
|
|
mov.f32 %f317, 0fBAB6061A;
|
||
|
|
mov.f32 %f318, 0f37CCF5CE;
|
||
|
|
fma.rn.f32 %f848, %f318, %f769, %f317;
|
||
|
|
|
||
|
|
BB0_22:
|
||
|
|
@%p21 bra BB0_24;
|
||
|
|
bra.uni BB0_23;
|
||
|
|
|
||
|
|
BB0_24:
|
||
|
|
mul.rn.f32 %f772, %f847, %f847;
|
||
|
|
mov.f32 %f768, 0f00000000;
|
||
|
|
mov.f32 %f324, 0fBE2AAAA3;
|
||
|
|
fma.rn.f32 %f325, %f848, %f772, %f324;
|
||
|
|
fma.rn.f32 %f849, %f325, %f772, %f768;
|
||
|
|
bra.uni BB0_25;
|
||
|
|
|
||
|
|
BB0_23:
|
||
|
|
mul.rn.f32 %f770, %f847, %f847;
|
||
|
|
mov.f32 %f321, 0f3D2AAAA5;
|
||
|
|
fma.rn.f32 %f322, %f848, %f770, %f321;
|
||
|
|
mov.f32 %f323, 0fBF000000;
|
||
|
|
fma.rn.f32 %f849, %f322, %f770, %f323;
|
||
|
|
|
||
|
|
BB0_25:
|
||
|
|
fma.rn.f32 %f850, %f849, %f847, %f847;
|
||
|
|
@%p21 bra BB0_27;
|
||
|
|
|
||
|
|
mul.rn.f32 %f771, %f847, %f847;
|
||
|
|
mov.f32 %f757, 0f3F800000;
|
||
|
|
fma.rn.f32 %f850, %f849, %f771, %f757;
|
||
|
|
|
||
|
|
BB0_27:
|
||
|
|
add.s32 %r348, %r371, 1;
|
||
|
|
and.b32 %r169, %r348, 2;
|
||
|
|
setp.eq.s32 %p24, %r169, 0;
|
||
|
|
@%p24 bra BB0_29;
|
||
|
|
|
||
|
|
mov.f32 %f764, 0f00000000;
|
||
|
|
mov.f32 %f329, 0fBF800000;
|
||
|
|
fma.rn.f32 %f850, %f850, %f329, %f764;
|
||
|
|
|
||
|
|
BB0_29:
|
||
|
|
abs.f32 %f758, %f852;
|
||
|
|
setp.neu.f32 %p105, %f758, 0f7F800000;
|
||
|
|
@%p105 bra BB0_31;
|
||
|
|
|
||
|
|
mov.f32 %f767, 0f00000000;
|
||
|
|
mul.rn.f32 %f852, %f852, %f767;
|
||
|
|
|
||
|
|
BB0_31:
|
||
|
|
mov.f32 %f761, 0f27C234C5;
|
||
|
|
mov.f32 %f760, 0f33A22168;
|
||
|
|
mov.f32 %f759, 0f3FC90FDA;
|
||
|
|
mul.f32 %f331, %f852, 0f3F22F983;
|
||
|
|
cvt.rni.s32.f32 %r381, %f331;
|
||
|
|
cvt.rn.f32.s32 %f332, %r381;
|
||
|
|
neg.f32 %f333, %f332;
|
||
|
|
fma.rn.f32 %f335, %f333, %f759, %f852;
|
||
|
|
fma.rn.f32 %f337, %f333, %f760, %f335;
|
||
|
|
fma.rn.f32 %f853, %f333, %f761, %f337;
|
||
|
|
abs.f32 %f339, %f852;
|
||
|
|
setp.leu.f32 %p26, %f339, 0f47CE4780;
|
||
|
|
@%p26 bra BB0_42;
|
||
|
|
|
||
|
|
add.u64 %rd53, %SP, 12;
|
||
|
|
cvta.to.local.u64 %rd285, %rd53;
|
||
|
|
mov.b32 %r50, %f852;
|
||
|
|
shl.b32 %r172, %r50, 8;
|
||
|
|
or.b32 %r52, %r172, -2147483648;
|
||
|
|
mov.u32 %r372, 0;
|
||
|
|
mov.u64 %rd286, %rd25;
|
||
|
|
mov.u32 %r373, %r372;
|
||
|
|
|
||
|
|
BB0_33:
|
||
|
|
.pragma "nounroll";
|
||
|
|
add.u64 %rd281, %SP, 12;
|
||
|
|
cvta.to.local.u64 %rd280, %rd281;
|
||
|
|
shl.b64 %rd54, %rd286, 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, %r373;
|
||
|
|
madc.hi.u32 %r373, %r175, %r52, 0;
|
||
|
|
}
|
||
|
|
// inline asm
|
||
|
|
st.local.u32 [%rd285], %r173;
|
||
|
|
add.s32 %r372, %r372, 1;
|
||
|
|
cvt.s64.s32 %rd286, %r372;
|
||
|
|
mul.wide.s32 %rd57, %r372, 4;
|
||
|
|
add.s64 %rd285, %rd280, %rd57;
|
||
|
|
setp.ne.s32 %p27, %r372, 6;
|
||
|
|
@%p27 bra BB0_33;
|
||
|
|
|
||
|
|
mov.b32 %r350, %f852;
|
||
|
|
shr.u32 %r349, %r350, 23;
|
||
|
|
add.u64 %rd282, %SP, 12;
|
||
|
|
and.b32 %r178, %r349, 255;
|
||
|
|
add.s32 %r179, %r178, -128;
|
||
|
|
shr.u32 %r180, %r179, 5;
|
||
|
|
cvta.to.local.u64 %rd59, %rd282;
|
||
|
|
st.local.u32 [%rd59+24], %r373;
|
||
|
|
mov.u32 %r181, 6;
|
||
|
|
sub.s32 %r182, %r181, %r180;
|
||
|
|
mul.wide.s32 %rd60, %r182, 4;
|
||
|
|
add.s64 %rd15, %rd59, %rd60;
|
||
|
|
ld.local.u32 %r374, [%rd15];
|
||
|
|
ld.local.u32 %r375, [%rd15+-4];
|
||
|
|
and.b32 %r60, %r349, 31;
|
||
|
|
setp.eq.s32 %p28, %r60, 0;
|
||
|
|
@%p28 bra BB0_36;
|
||
|
|
|
||
|
|
mov.u32 %r183, 32;
|
||
|
|
sub.s32 %r184, %r183, %r60;
|
||
|
|
shr.u32 %r185, %r375, %r184;
|
||
|
|
shl.b32 %r186, %r374, %r60;
|
||
|
|
add.s32 %r374, %r185, %r186;
|
||
|
|
ld.local.u32 %r187, [%rd15+-8];
|
||
|
|
shr.u32 %r188, %r187, %r184;
|
||
|
|
shl.b32 %r189, %r375, %r60;
|
||
|
|
add.s32 %r375, %r188, %r189;
|
||
|
|
|
||
|
|
BB0_36:
|
||
|
|
mov.b32 %r353, %f852;
|
||
|
|
and.b32 %r377, %r353, -2147483648;
|
||
|
|
shr.u32 %r190, %r375, 30;
|
||
|
|
shl.b32 %r191, %r374, 2;
|
||
|
|
add.s32 %r376, %r190, %r191;
|
||
|
|
shl.b32 %r66, %r375, 2;
|
||
|
|
shr.u32 %r192, %r376, 31;
|
||
|
|
shr.u32 %r193, %r374, 30;
|
||
|
|
add.s32 %r67, %r192, %r193;
|
||
|
|
setp.eq.s32 %p29, %r192, 0;
|
||
|
|
@%p29 bra BB0_37;
|
||
|
|
bra.uni BB0_38;
|
||
|
|
|
||
|
|
BB0_37:
|
||
|
|
mov.u32 %r378, %r66;
|
||
|
|
bra.uni BB0_39;
|
||
|
|
|
||
|
|
BB0_38:
|
||
|
|
mov.b32 %r355, %f852;
|
||
|
|
and.b32 %r354, %r355, -2147483648;
|
||
|
|
not.b32 %r194, %r376;
|
||
|
|
neg.s32 %r378, %r66;
|
||
|
|
setp.eq.s32 %p30, %r66, 0;
|
||
|
|
selp.u32 %r195, 1, 0, %p30;
|
||
|
|
add.s32 %r376, %r195, %r194;
|
||
|
|
xor.b32 %r377, %r354, -2147483648;
|
||
|
|
|
||
|
|
BB0_39:
|
||
|
|
mov.b32 %r357, %f852;
|
||
|
|
and.b32 %r356, %r357, -2147483648;
|
||
|
|
clz.b32 %r380, %r376;
|
||
|
|
setp.eq.s32 %p31, %r380, 0;
|
||
|
|
shl.b32 %r196, %r376, %r380;
|
||
|
|
mov.u32 %r197, 32;
|
||
|
|
sub.s32 %r198, %r197, %r380;
|
||
|
|
shr.u32 %r199, %r378, %r198;
|
||
|
|
add.s32 %r200, %r199, %r196;
|
||
|
|
selp.b32 %r75, %r376, %r200, %p31;
|
||
|
|
mov.u32 %r201, -921707870;
|
||
|
|
mul.hi.u32 %r379, %r75, %r201;
|
||
|
|
setp.eq.s32 %p32, %r356, 0;
|
||
|
|
neg.s32 %r202, %r67;
|
||
|
|
selp.b32 %r381, %r67, %r202, %p32;
|
||
|
|
setp.lt.s32 %p33, %r379, 1;
|
||
|
|
@%p33 bra BB0_41;
|
||
|
|
|
||
|
|
mul.lo.s32 %r203, %r75, -921707870;
|
||
|
|
shr.u32 %r204, %r203, 31;
|
||
|
|
shl.b32 %r205, %r379, 1;
|
||
|
|
add.s32 %r379, %r204, %r205;
|
||
|
|
add.s32 %r380, %r380, 1;
|
||
|
|
|
||
|
|
BB0_41:
|
||
|
|
mov.u32 %r206, 126;
|
||
|
|
sub.s32 %r207, %r206, %r380;
|
||
|
|
shl.b32 %r208, %r207, 23;
|
||
|
|
add.s32 %r209, %r379, 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, %r377;
|
||
|
|
mov.b32 %f853, %r214;
|
||
|
|
|
||
|
|
BB0_42:
|
||
|
|
and.b32 %r83, %r381, 1;
|
||
|
|
setp.eq.s32 %p34, %r83, 0;
|
||
|
|
@%p34 bra BB0_44;
|
||
|
|
bra.uni BB0_43;
|
||
|
|
|
||
|
|
BB0_44:
|
||
|
|
mul.rn.f32 %f778, %f853, %f853;
|
||
|
|
mov.f32 %f342, 0f3C08839E;
|
||
|
|
mov.f32 %f343, 0fB94CA1F9;
|
||
|
|
fma.rn.f32 %f854, %f343, %f778, %f342;
|
||
|
|
bra.uni BB0_45;
|
||
|
|
|
||
|
|
BB0_43:
|
||
|
|
mul.rn.f32 %f774, %f853, %f853;
|
||
|
|
mov.f32 %f340, 0fBAB6061A;
|
||
|
|
mov.f32 %f341, 0f37CCF5CE;
|
||
|
|
fma.rn.f32 %f854, %f341, %f774, %f340;
|
||
|
|
|
||
|
|
BB0_45:
|
||
|
|
@%p34 bra BB0_47;
|
||
|
|
bra.uni BB0_46;
|
||
|
|
|
||
|
|
BB0_47:
|
||
|
|
mul.rn.f32 %f777, %f853, %f853;
|
||
|
|
mov.f32 %f766, 0f00000000;
|
||
|
|
mov.f32 %f347, 0fBE2AAAA3;
|
||
|
|
fma.rn.f32 %f348, %f854, %f777, %f347;
|
||
|
|
fma.rn.f32 %f855, %f348, %f777, %f766;
|
||
|
|
bra.uni BB0_48;
|
||
|
|
|
||
|
|
BB0_46:
|
||
|
|
mul.rn.f32 %f775, %f853, %f853;
|
||
|
|
mov.f32 %f344, 0f3D2AAAA5;
|
||
|
|
fma.rn.f32 %f345, %f854, %f775, %f344;
|
||
|
|
mov.f32 %f346, 0fBF000000;
|
||
|
|
fma.rn.f32 %f855, %f345, %f775, %f346;
|
||
|
|
|
||
|
|
BB0_48:
|
||
|
|
fma.rn.f32 %f856, %f855, %f853, %f853;
|
||
|
|
@%p34 bra BB0_50;
|
||
|
|
|
||
|
|
mul.rn.f32 %f776, %f853, %f853;
|
||
|
|
mov.f32 %f762, 0f3F800000;
|
||
|
|
fma.rn.f32 %f856, %f855, %f776, %f762;
|
||
|
|
|
||
|
|
BB0_50:
|
||
|
|
and.b32 %r215, %r381, 2;
|
||
|
|
setp.eq.s32 %p37, %r215, 0;
|
||
|
|
@%p37 bra BB0_52;
|
||
|
|
|
||
|
|
mov.f32 %f765, 0f00000000;
|
||
|
|
mov.f32 %f352, 0fBF800000;
|
||
|
|
fma.rn.f32 %f856, %f856, %f352, %f765;
|
||
|
|
|
||
|
|
BB0_52:
|
||
|
|
mul.f32 %f353, %f52, %f850;
|
||
|
|
mul.f32 %f354, %f52, %f856;
|
||
|
|
mul.f32 %f355, %f10, %f354;
|
||
|
|
mul.f32 %f356, %f11, %f354;
|
||
|
|
mul.f32 %f357, %f12, %f354;
|
||
|
|
fma.rn.f32 %f358, %f19, %f353, %f355;
|
||
|
|
fma.rn.f32 %f359, %f18, %f353, %f356;
|
||
|
|
fma.rn.f32 %f360, %f17, %f353, %f357;
|
||
|
|
fma.rn.f32 %f89, %f813, %f51, %f358;
|
||
|
|
fma.rn.f32 %f90, %f814, %f51, %f359;
|
||
|
|
fma.rn.f32 %f91, %f815, %f51, %f360;
|
||
|
|
setp.gt.f32 %p38, %f90, 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 %f374, %f380, %f259;
|
||
|
|
ld.global.u32 %r216, [sky];
|
||
|
|
neg.f32 %f367, %f89;
|
||
|
|
neg.f32 %f366, %f90;
|
||
|
|
neg.f32 %f365, %f91;
|
||
|
|
mov.u32 %r217, 6;
|
||
|
|
mov.u32 %r218, 0;
|
||
|
|
// inline asm
|
||
|
|
call (%f361, %f362, %f363, %f364), _rt_texture_get_base_id, (%r216, %r217, %f365, %f366, %f367, %r218);
|
||
|
|
// inline asm
|
||
|
|
st.local.f32 [%rd62+8], %f363;
|
||
|
|
st.local.f32 [%rd62+4], %f362;
|
||
|
|
st.local.f32 [%rd62], %f361;
|
||
|
|
ld.global.u32 %r219, [root];
|
||
|
|
mov.u32 %r220, 1;
|
||
|
|
mov.f32 %f375, 0f6C4ECB8F;
|
||
|
|
// inline asm
|
||
|
|
call _rt_trace_64, (%r219, %f7, %f8, %f9, %f89, %f90, %f91, %r220, %f374, %f375, %rd61, %r108);
|
||
|
|
// inline asm
|
||
|
|
mul.f32 %f382, %f814, %f90;
|
||
|
|
fma.rn.f32 %f383, %f813, %f89, %f382;
|
||
|
|
fma.rn.f32 %f384, %f815, %f91, %f383;
|
||
|
|
ld.local.f32 %f385, [%rd62+8];
|
||
|
|
ld.local.f32 %f386, [%rd62+4];
|
||
|
|
ld.local.f32 %f387, [%rd62];
|
||
|
|
fma.rn.f32 %f864, %f89, %f387, %f864;
|
||
|
|
fma.rn.f32 %f865, %f89, %f386, %f865;
|
||
|
|
fma.rn.f32 %f866, %f89, %f385, %f866;
|
||
|
|
fma.rn.f32 %f861, %f90, %f387, %f861;
|
||
|
|
fma.rn.f32 %f862, %f90, %f386, %f862;
|
||
|
|
fma.rn.f32 %f863, %f90, %f385, %f863;
|
||
|
|
fma.rn.f32 %f858, %f91, %f387, %f858;
|
||
|
|
fma.rn.f32 %f859, %f91, %f386, %f859;
|
||
|
|
fma.rn.f32 %f860, %f91, %f385, %f860;
|
||
|
|
add.f32 %f867, %f867, %f387;
|
||
|
|
add.f32 %f868, %f868, %f386;
|
||
|
|
add.f32 %f869, %f869, %f385;
|
||
|
|
cvt.sat.f32.f32 %f388, %f384;
|
||
|
|
fma.rn.f32 %f870, %f388, %f387, %f870;
|
||
|
|
fma.rn.f32 %f871, %f388, %f386, %f871;
|
||
|
|
fma.rn.f32 %f872, %f388, %f385, %f872;
|
||
|
|
|
||
|
|
BB0_54:
|
||
|
|
mad.lo.s32 %r338, %r11, 1664525, 1013904223;
|
||
|
|
mad.lo.s32 %r361, %r338, 1664525, 1013904223;
|
||
|
|
add.s32 %r360, %r360, 1;
|
||
|
|
setp.lt.s32 %p41, %r360, %r2;
|
||
|
|
@%p41 bra BB0_6;
|
||
|
|
|
||
|
|
mad.lo.s32 %r340, %r11, 1664525, 1013904223;
|
||
|
|
mad.lo.s32 %r361, %r340, 1664525, 1013904223;
|
||
|
|
add.s32 %r358, %r358, 1;
|
||
|
|
setp.lt.s32 %p42, %r358, %r2;
|
||
|
|
@%p42 bra BB0_5;
|
||
|
|
|
||
|
|
BB0_56:
|
||
|
|
mul.lo.s32 %r222, %r2, %r2;
|
||
|
|
cvt.rn.f32.s32 %f389, %r222;
|
||
|
|
rcp.rn.f32 %f390, %f389;
|
||
|
|
mul.f32 %f391, %f870, %f390;
|
||
|
|
mul.f32 %f392, %f871, %f390;
|
||
|
|
mul.f32 %f393, %f872, %f390;
|
||
|
|
mul.f32 %f137, %f867, %f390;
|
||
|
|
mul.f32 %f138, %f868, %f390;
|
||
|
|
mul.f32 %f139, %f869, %f390;
|
||
|
|
mul.f32 %f140, %f864, %f390;
|
||
|
|
mul.f32 %f141, %f865, %f390;
|
||
|
|
mul.f32 %f142, %f866, %f390;
|
||
|
|
mul.f32 %f143, %f861, %f390;
|
||
|
|
mul.f32 %f144, %f862, %f390;
|
||
|
|
mul.f32 %f145, %f863, %f390;
|
||
|
|
mul.f32 %f146, %f858, %f390;
|
||
|
|
mul.f32 %f147, %f859, %f390;
|
||
|
|
mul.f32 %f148, %f860, %f390;
|
||
|
|
fma.rn.f32 %f394, %f870, %f390, %f391;
|
||
|
|
fma.rn.f32 %f395, %f871, %f390, %f392;
|
||
|
|
fma.rn.f32 %f396, %f872, %f390, %f393;
|
||
|
|
ld.global.f32 %f397, [skyColor];
|
||
|
|
mul.f32 %f149, %f397, %f394;
|
||
|
|
ld.global.f32 %f398, [skyColor+4];
|
||
|
|
mul.f32 %f150, %f395, %f398;
|
||
|
|
ld.global.f32 %f399, [skyColor+8];
|
||
|
|
mul.f32 %f151, %f396, %f399;
|
||
|
|
ld.global.u32 %r382, [imageEnabled];
|
||
|
|
and.b32 %r223, %r382, 1;
|
||
|
|
setp.eq.b32 %p43, %r223, 1;
|
||
|
|
@!%p43 bra BB0_91;
|
||
|
|
bra.uni BB0_57;
|
||
|
|
|
||
|
|
BB0_57:
|
||
|
|
abs.f32 %f153, %f149;
|
||
|
|
setp.lt.f32 %p44, %f153, 0f00800000;
|
||
|
|
mul.f32 %f405, %f153, 0f4B800000;
|
||
|
|
selp.f32 %f406, 0fC3170000, 0fC2FE0000, %p44;
|
||
|
|
selp.f32 %f407, %f405, %f153, %p44;
|
||
|
|
mov.b32 %r224, %f407;
|
||
|
|
and.b32 %r225, %r224, 8388607;
|
||
|
|
or.b32 %r226, %r225, 1065353216;
|
||
|
|
mov.b32 %f408, %r226;
|
||
|
|
shr.u32 %r227, %r224, 23;
|
||
|
|
cvt.rn.f32.u32 %f409, %r227;
|
||
|
|
add.f32 %f410, %f406, %f409;
|
||
|
|
setp.gt.f32 %p45, %f408, 0f3FB504F3;
|
||
|
|
mul.f32 %f411, %f408, 0f3F000000;
|
||
|
|
add.f32 %f412, %f410, 0f3F800000;
|
||
|
|
selp.f32 %f413, %f411, %f408, %p45;
|
||
|
|
selp.f32 %f414, %f412, %f410, %p45;
|
||
|
|
add.f32 %f415, %f413, 0fBF800000;
|
||
|
|
add.f32 %f401, %f413, 0f3F800000;
|
||
|
|
// inline asm
|
||
|
|
rcp.approx.ftz.f32 %f400,%f401;
|
||
|
|
// inline asm
|
||
|
|
add.f32 %f416, %f415, %f415;
|
||
|
|
mul.f32 %f417, %f400, %f416;
|
||
|
|
mul.f32 %f418, %f417, %f417;
|
||
|
|
mov.f32 %f419, 0f3C4CAF63;
|
||
|
|
mov.f32 %f420, 0f3B18F0FE;
|
||
|
|
fma.rn.f32 %f421, %f420, %f418, %f419;
|
||
|
|
mov.f32 %f422, 0f3DAAAABD;
|
||
|
|
fma.rn.f32 %f423, %f421, %f418, %f422;
|
||
|
|
mul.rn.f32 %f424, %f423, %f418;
|
||
|
|
mul.rn.f32 %f425, %f424, %f417;
|
||
|
|
sub.f32 %f426, %f415, %f417;
|
||
|
|
neg.f32 %f427, %f417;
|
||
|
|
add.f32 %f428, %f426, %f426;
|
||
|
|
fma.rn.f32 %f429, %f427, %f415, %f428;
|
||
|
|
mul.rn.f32 %f430, %f400, %f429;
|
||
|
|
add.f32 %f431, %f425, %f417;
|
||
|
|
sub.f32 %f432, %f417, %f431;
|
||
|
|
add.f32 %f433, %f425, %f432;
|
||
|
|
add.f32 %f434, %f430, %f433;
|
||
|
|
add.f32 %f435, %f431, %f434;
|
||
|
|
sub.f32 %f436, %f431, %f435;
|
||
|
|
add.f32 %f437, %f434, %f436;
|
||
|
|
mov.f32 %f438, 0f3F317200;
|
||
|
|
mul.rn.f32 %f439, %f414, %f438;
|
||
|
|
mov.f32 %f440, 0f35BFBE8E;
|
||
|
|
mul.rn.f32 %f441, %f414, %f440;
|
||
|
|
add.f32 %f442, %f439, %f435;
|
||
|
|
sub.f32 %f443, %f439, %f442;
|
||
|
|
add.f32 %f444, %f435, %f443;
|
||
|
|
add.f32 %f445, %f437, %f444;
|
||
|
|
add.f32 %f446, %f441, %f445;
|
||
|
|
add.f32 %f447, %f442, %f446;
|
||
|
|
sub.f32 %f448, %f442, %f447;
|
||
|
|
add.f32 %f449, %f446, %f448;
|
||
|
|
mov.f32 %f450, 0f3EE66666;
|
||
|
|
mul.rn.f32 %f451, %f450, %f447;
|
||
|
|
neg.f32 %f452, %f451;
|
||
|
|
fma.rn.f32 %f453, %f450, %f447, %f452;
|
||
|
|
fma.rn.f32 %f454, %f450, %f449, %f453;
|
||
|
|
mov.f32 %f455, 0f00000000;
|
||
|
|
fma.rn.f32 %f456, %f455, %f447, %f454;
|
||
|
|
add.rn.f32 %f457, %f451, %f456;
|
||
|
|
neg.f32 %f458, %f457;
|
||
|
|
add.rn.f32 %f459, %f451, %f458;
|
||
|
|
add.rn.f32 %f460, %f459, %f456;
|
||
|
|
mov.b32 %r228, %f457;
|
||
|
|
setp.eq.s32 %p46, %r228, 1118925336;
|
||
|
|
add.s32 %r229, %r228, -1;
|
||
|
|
mov.b32 %f461, %r229;
|
||
|
|
add.f32 %f462, %f460, 0f37000000;
|
||
|
|
selp.f32 %f463, %f461, %f457, %p46;
|
||
|
|
selp.f32 %f154, %f462, %f460, %p46;
|
||
|
|
mul.f32 %f464, %f463, 0f3FB8AA3B;
|
||
|
|
cvt.rzi.f32.f32 %f465, %f464;
|
||
|
|
mov.f32 %f466, 0fBF317200;
|
||
|
|
fma.rn.f32 %f467, %f465, %f466, %f463;
|
||
|
|
mov.f32 %f468, 0fB5BFBE8E;
|
||
|
|
fma.rn.f32 %f469, %f465, %f468, %f467;
|
||
|
|
mul.f32 %f470, %f469, 0f3FB8AA3B;
|
||
|
|
ex2.approx.ftz.f32 %f471, %f470;
|
||
|
|
add.f32 %f472, %f465, 0f00000000;
|
||
|
|
ex2.approx.f32 %f473, %f472;
|
||
|
|
mul.f32 %f474, %f471, %f473;
|
||
|
|
setp.lt.f32 %p47, %f463, 0fC2D20000;
|
||
|
|
selp.f32 %f475, 0f00000000, %f474, %p47;
|
||
|
|
setp.gt.f32 %p48, %f463, 0f42D20000;
|
||
|
|
selp.f32 %f888, 0f7F800000, %f475, %p48;
|
||
|
|
setp.eq.f32 %p49, %f888, 0f7F800000;
|
||
|
|
@%p49 bra BB0_59;
|
||
|
|
|
||
|
|
fma.rn.f32 %f888, %f888, %f154, %f888;
|
||
|
|
|
||
|
|
BB0_59:
|
||
|
|
mov.f32 %f782, 0f3E666666;
|
||
|
|
cvt.rzi.f32.f32 %f781, %f782;
|
||
|
|
fma.rn.f32 %f780, %f781, 0fC0000000, 0f3EE66666;
|
||
|
|
abs.f32 %f779, %f780;
|
||
|
|
setp.lt.f32 %p50, %f149, 0f00000000;
|
||
|
|
setp.eq.f32 %p51, %f779, 0f3F800000;
|
||
|
|
and.pred %p1, %p50, %p51;
|
||
|
|
mov.b32 %r230, %f888;
|
||
|
|
xor.b32 %r231, %r230, -2147483648;
|
||
|
|
mov.b32 %f476, %r231;
|
||
|
|
selp.f32 %f890, %f476, %f888, %p1;
|
||
|
|
setp.eq.f32 %p52, %f149, 0f00000000;
|
||
|
|
@%p52 bra BB0_62;
|
||
|
|
bra.uni BB0_60;
|
||
|
|
|
||
|
|
BB0_62:
|
||
|
|
add.f32 %f479, %f149, %f149;
|
||
|
|
selp.f32 %f890, %f479, 0f00000000, %p51;
|
||
|
|
bra.uni BB0_63;
|
||
|
|
|
||
|
|
BB0_112:
|
||
|
|
mov.u64 %rd196, image_HDR;
|
||
|
|
cvta.global.u64 %rd191, %rd196;
|
||
|
|
mov.u32 %r302, 8;
|
||
|
|
mov.u64 %rd195, 0;
|
||
|
|
// inline asm
|
||
|
|
call (%rd190), _rt_buffer_get_64, (%rd191, %r92, %r302, %rd18, %rd19, %rd195, %rd195);
|
||
|
|
// inline asm
|
||
|
|
mov.f32 %f728, 0f00000000;
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs100, %f728;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
mov.u16 %rs101, 0;
|
||
|
|
st.v4.u16 [%rd190], {%rs100, %rs100, %rs100, %rs101};
|
||
|
|
|
||
|
|
BB0_113:
|
||
|
|
ld.global.u32 %r303, [additive];
|
||
|
|
setp.eq.s32 %p101, %r303, 0;
|
||
|
|
@%p101 bra BB0_115;
|
||
|
|
|
||
|
|
mov.u64 %rd209, image_RNM0;
|
||
|
|
cvta.global.u64 %rd198, %rd209;
|
||
|
|
mov.u32 %r307, 8;
|
||
|
|
mov.u64 %rd208, 0;
|
||
|
|
// inline asm
|
||
|
|
call (%rd197), _rt_buffer_get_64, (%rd198, %r92, %r307, %rd18, %rd19, %rd208, %rd208);
|
||
|
|
// inline asm
|
||
|
|
ld.v4.u16 {%rs108, %rs109, %rs110, %rs111}, [%rd197];
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f729, %rs108;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f730, %rs109;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f731, %rs110;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
call (%rd203), _rt_buffer_get_64, (%rd198, %r92, %r307, %rd18, %rd19, %rd208, %rd208);
|
||
|
|
// inline asm
|
||
|
|
add.f32 %f732, %f729, 0f00000000;
|
||
|
|
add.f32 %f733, %f730, 0f00000000;
|
||
|
|
add.f32 %f734, %f731, 0f00000000;
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs107, %f734;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs106, %f733;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs105, %f732;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
mov.u16 %rs112, 0;
|
||
|
|
st.v4.u16 [%rd203], {%rs105, %rs106, %rs107, %rs112};
|
||
|
|
bra.uni BB0_116;
|
||
|
|
|
||
|
|
BB0_115:
|
||
|
|
mov.u64 %rd216, image_RNM0;
|
||
|
|
cvta.global.u64 %rd211, %rd216;
|
||
|
|
mov.u32 %r309, 8;
|
||
|
|
mov.u64 %rd215, 0;
|
||
|
|
// inline asm
|
||
|
|
call (%rd210), _rt_buffer_get_64, (%rd211, %r92, %r309, %rd18, %rd19, %rd215, %rd215);
|
||
|
|
// inline asm
|
||
|
|
mov.f32 %f735, 0f00000000;
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs113, %f735;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
mov.u16 %rs114, 0;
|
||
|
|
st.v4.u16 [%rd210], {%rs113, %rs113, %rs113, %rs114};
|
||
|
|
|
||
|
|
BB0_116:
|
||
|
|
ld.global.u32 %r310, [additive];
|
||
|
|
setp.eq.s32 %p102, %r310, 0;
|
||
|
|
@%p102 bra BB0_118;
|
||
|
|
|
||
|
|
mov.u64 %rd229, image_RNM1;
|
||
|
|
cvta.global.u64 %rd218, %rd229;
|
||
|
|
mov.u32 %r314, 8;
|
||
|
|
mov.u64 %rd228, 0;
|
||
|
|
// inline asm
|
||
|
|
call (%rd217), _rt_buffer_get_64, (%rd218, %r92, %r314, %rd18, %rd19, %rd228, %rd228);
|
||
|
|
// inline asm
|
||
|
|
ld.v4.u16 {%rs121, %rs122, %rs123, %rs124}, [%rd217];
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f736, %rs121;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f737, %rs122;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f738, %rs123;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
call (%rd223), _rt_buffer_get_64, (%rd218, %r92, %r314, %rd18, %rd19, %rd228, %rd228);
|
||
|
|
// inline asm
|
||
|
|
add.f32 %f739, %f736, 0f00000000;
|
||
|
|
add.f32 %f740, %f737, 0f00000000;
|
||
|
|
add.f32 %f741, %f738, 0f00000000;
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs120, %f741;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs119, %f740;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs118, %f739;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
mov.u16 %rs125, 0;
|
||
|
|
st.v4.u16 [%rd223], {%rs118, %rs119, %rs120, %rs125};
|
||
|
|
bra.uni BB0_119;
|
||
|
|
|
||
|
|
BB0_118:
|
||
|
|
mov.u64 %rd236, image_RNM1;
|
||
|
|
cvta.global.u64 %rd231, %rd236;
|
||
|
|
mov.u32 %r316, 8;
|
||
|
|
mov.u64 %rd235, 0;
|
||
|
|
// inline asm
|
||
|
|
call (%rd230), _rt_buffer_get_64, (%rd231, %r92, %r316, %rd18, %rd19, %rd235, %rd235);
|
||
|
|
// inline asm
|
||
|
|
mov.f32 %f742, 0f00000000;
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs126, %f742;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
mov.u16 %rs127, 0;
|
||
|
|
st.v4.u16 [%rd230], {%rs126, %rs126, %rs126, %rs127};
|
||
|
|
|
||
|
|
BB0_119:
|
||
|
|
ld.global.u32 %r317, [additive];
|
||
|
|
setp.eq.s32 %p103, %r317, 0;
|
||
|
|
@%p103 bra BB0_121;
|
||
|
|
|
||
|
|
mov.u64 %rd249, image_RNM2;
|
||
|
|
cvta.global.u64 %rd238, %rd249;
|
||
|
|
mov.u32 %r321, 8;
|
||
|
|
mov.u64 %rd248, 0;
|
||
|
|
// inline asm
|
||
|
|
call (%rd237), _rt_buffer_get_64, (%rd238, %r92, %r321, %rd18, %rd19, %rd248, %rd248);
|
||
|
|
// inline asm
|
||
|
|
ld.v4.u16 {%rs134, %rs135, %rs136, %rs137}, [%rd237];
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f743, %rs134;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f744, %rs135;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f745, %rs136;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
call (%rd243), _rt_buffer_get_64, (%rd238, %r92, %r321, %rd18, %rd19, %rd248, %rd248);
|
||
|
|
// inline asm
|
||
|
|
add.f32 %f746, %f743, 0f00000000;
|
||
|
|
add.f32 %f747, %f744, 0f00000000;
|
||
|
|
add.f32 %f748, %f745, 0f00000000;
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs133, %f748;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs132, %f747;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs131, %f746;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
mov.u16 %rs138, 0;
|
||
|
|
st.v4.u16 [%rd243], {%rs131, %rs132, %rs133, %rs138};
|
||
|
|
bra.uni BB0_122;
|
||
|
|
|
||
|
|
BB0_121:
|
||
|
|
mov.u64 %rd256, image_RNM2;
|
||
|
|
cvta.global.u64 %rd251, %rd256;
|
||
|
|
mov.u32 %r323, 8;
|
||
|
|
mov.u64 %rd255, 0;
|
||
|
|
// inline asm
|
||
|
|
call (%rd250), _rt_buffer_get_64, (%rd251, %r92, %r323, %rd18, %rd19, %rd255, %rd255);
|
||
|
|
// inline asm
|
||
|
|
mov.f32 %f749, 0f00000000;
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs139, %f749;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
mov.u16 %rs140, 0;
|
||
|
|
st.v4.u16 [%rd250], {%rs139, %rs139, %rs139, %rs140};
|
||
|
|
|
||
|
|
BB0_122:
|
||
|
|
ld.global.u32 %r324, [additive];
|
||
|
|
setp.eq.s32 %p104, %r324, 0;
|
||
|
|
@%p104 bra BB0_124;
|
||
|
|
|
||
|
|
mov.u64 %rd269, image_RNM3;
|
||
|
|
cvta.global.u64 %rd258, %rd269;
|
||
|
|
mov.u32 %r328, 8;
|
||
|
|
mov.u64 %rd268, 0;
|
||
|
|
// inline asm
|
||
|
|
call (%rd257), _rt_buffer_get_64, (%rd258, %r92, %r328, %rd18, %rd19, %rd268, %rd268);
|
||
|
|
// inline asm
|
||
|
|
ld.v4.u16 {%rs147, %rs148, %rs149, %rs150}, [%rd257];
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f750, %rs147;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f751, %rs148;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f752, %rs149;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
call (%rd263), _rt_buffer_get_64, (%rd258, %r92, %r328, %rd18, %rd19, %rd268, %rd268);
|
||
|
|
// inline asm
|
||
|
|
add.f32 %f753, %f750, 0f00000000;
|
||
|
|
add.f32 %f754, %f751, 0f00000000;
|
||
|
|
add.f32 %f755, %f752, 0f00000000;
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs146, %f755;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs145, %f754;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs144, %f753;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
mov.u16 %rs151, 0;
|
||
|
|
st.v4.u16 [%rd263], {%rs144, %rs145, %rs146, %rs151};
|
||
|
|
bra.uni BB0_125;
|
||
|
|
|
||
|
|
BB0_124:
|
||
|
|
mov.u64 %rd276, image_RNM3;
|
||
|
|
cvta.global.u64 %rd271, %rd276;
|
||
|
|
mov.u32 %r330, 8;
|
||
|
|
mov.u64 %rd275, 0;
|
||
|
|
// inline asm
|
||
|
|
call (%rd270), _rt_buffer_get_64, (%rd271, %r92, %r330, %rd18, %rd19, %rd275, %rd275);
|
||
|
|
// inline asm
|
||
|
|
mov.f32 %f756, 0f00000000;
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs152, %f756;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
mov.u16 %rs153, 0;
|
||
|
|
st.v4.u16 [%rd270], {%rs152, %rs152, %rs152, %rs153};
|
||
|
|
bra.uni BB0_125;
|
||
|
|
|
||
|
|
BB0_60:
|
||
|
|
setp.geu.f32 %p53, %f149, 0f00000000;
|
||
|
|
@%p53 bra BB0_63;
|
||
|
|
|
||
|
|
mov.f32 %f806, 0f3EE66666;
|
||
|
|
cvt.rzi.f32.f32 %f478, %f806;
|
||
|
|
setp.neu.f32 %p54, %f478, 0f3EE66666;
|
||
|
|
selp.f32 %f890, 0f7FFFFFFF, %f890, %p54;
|
||
|
|
|
||
|
|
BB0_63:
|
||
|
|
abs.f32 %f783, %f149;
|
||
|
|
add.f32 %f480, %f783, 0f3EE66666;
|
||
|
|
mov.b32 %r232, %f480;
|
||
|
|
setp.lt.s32 %p56, %r232, 2139095040;
|
||
|
|
@%p56 bra BB0_68;
|
||
|
|
|
||
|
|
abs.f32 %f804, %f149;
|
||
|
|
setp.gtu.f32 %p57, %f804, 0f7F800000;
|
||
|
|
@%p57 bra BB0_67;
|
||
|
|
bra.uni BB0_65;
|
||
|
|
|
||
|
|
BB0_67:
|
||
|
|
add.f32 %f890, %f149, 0f3EE66666;
|
||
|
|
bra.uni BB0_68;
|
||
|
|
|
||
|
|
BB0_65:
|
||
|
|
abs.f32 %f805, %f149;
|
||
|
|
setp.neu.f32 %p58, %f805, 0f7F800000;
|
||
|
|
@%p58 bra BB0_68;
|
||
|
|
|
||
|
|
selp.f32 %f890, 0fFF800000, 0f7F800000, %p1;
|
||
|
|
|
||
|
|
BB0_68:
|
||
|
|
mov.f32 %f792, 0fB5BFBE8E;
|
||
|
|
mov.f32 %f791, 0fBF317200;
|
||
|
|
mov.f32 %f790, 0f00000000;
|
||
|
|
mov.f32 %f789, 0f35BFBE8E;
|
||
|
|
mov.f32 %f788, 0f3F317200;
|
||
|
|
mov.f32 %f787, 0f3DAAAABD;
|
||
|
|
mov.f32 %f786, 0f3C4CAF63;
|
||
|
|
mov.f32 %f785, 0f3B18F0FE;
|
||
|
|
mov.f32 %f784, 0f3EE66666;
|
||
|
|
setp.eq.f32 %p59, %f149, 0f3F800000;
|
||
|
|
selp.f32 %f165, 0f3F800000, %f890, %p59;
|
||
|
|
abs.f32 %f166, %f150;
|
||
|
|
setp.lt.f32 %p60, %f166, 0f00800000;
|
||
|
|
mul.f32 %f483, %f166, 0f4B800000;
|
||
|
|
selp.f32 %f484, 0fC3170000, 0fC2FE0000, %p60;
|
||
|
|
selp.f32 %f485, %f483, %f166, %p60;
|
||
|
|
mov.b32 %r233, %f485;
|
||
|
|
and.b32 %r234, %r233, 8388607;
|
||
|
|
or.b32 %r235, %r234, 1065353216;
|
||
|
|
mov.b32 %f486, %r235;
|
||
|
|
shr.u32 %r236, %r233, 23;
|
||
|
|
cvt.rn.f32.u32 %f487, %r236;
|
||
|
|
add.f32 %f488, %f484, %f487;
|
||
|
|
setp.gt.f32 %p61, %f486, 0f3FB504F3;
|
||
|
|
mul.f32 %f489, %f486, 0f3F000000;
|
||
|
|
add.f32 %f490, %f488, 0f3F800000;
|
||
|
|
selp.f32 %f491, %f489, %f486, %p61;
|
||
|
|
selp.f32 %f492, %f490, %f488, %p61;
|
||
|
|
add.f32 %f493, %f491, 0fBF800000;
|
||
|
|
add.f32 %f482, %f491, 0f3F800000;
|
||
|
|
// inline asm
|
||
|
|
rcp.approx.ftz.f32 %f481,%f482;
|
||
|
|
// inline asm
|
||
|
|
add.f32 %f494, %f493, %f493;
|
||
|
|
mul.f32 %f495, %f481, %f494;
|
||
|
|
mul.f32 %f496, %f495, %f495;
|
||
|
|
fma.rn.f32 %f499, %f785, %f496, %f786;
|
||
|
|
fma.rn.f32 %f501, %f499, %f496, %f787;
|
||
|
|
mul.rn.f32 %f502, %f501, %f496;
|
||
|
|
mul.rn.f32 %f503, %f502, %f495;
|
||
|
|
sub.f32 %f504, %f493, %f495;
|
||
|
|
neg.f32 %f505, %f495;
|
||
|
|
add.f32 %f506, %f504, %f504;
|
||
|
|
fma.rn.f32 %f507, %f505, %f493, %f506;
|
||
|
|
mul.rn.f32 %f508, %f481, %f507;
|
||
|
|
add.f32 %f509, %f503, %f495;
|
||
|
|
sub.f32 %f510, %f495, %f509;
|
||
|
|
add.f32 %f511, %f503, %f510;
|
||
|
|
add.f32 %f512, %f508, %f511;
|
||
|
|
add.f32 %f513, %f509, %f512;
|
||
|
|
sub.f32 %f514, %f509, %f513;
|
||
|
|
add.f32 %f515, %f512, %f514;
|
||
|
|
mul.rn.f32 %f517, %f492, %f788;
|
||
|
|
mul.rn.f32 %f519, %f492, %f789;
|
||
|
|
add.f32 %f520, %f517, %f513;
|
||
|
|
sub.f32 %f521, %f517, %f520;
|
||
|
|
add.f32 %f522, %f513, %f521;
|
||
|
|
add.f32 %f523, %f515, %f522;
|
||
|
|
add.f32 %f524, %f519, %f523;
|
||
|
|
add.f32 %f525, %f520, %f524;
|
||
|
|
sub.f32 %f526, %f520, %f525;
|
||
|
|
add.f32 %f527, %f524, %f526;
|
||
|
|
mul.rn.f32 %f529, %f784, %f525;
|
||
|
|
neg.f32 %f530, %f529;
|
||
|
|
fma.rn.f32 %f531, %f784, %f525, %f530;
|
||
|
|
fma.rn.f32 %f532, %f784, %f527, %f531;
|
||
|
|
fma.rn.f32 %f534, %f790, %f525, %f532;
|
||
|
|
add.rn.f32 %f535, %f529, %f534;
|
||
|
|
neg.f32 %f536, %f535;
|
||
|
|
add.rn.f32 %f537, %f529, %f536;
|
||
|
|
add.rn.f32 %f538, %f537, %f534;
|
||
|
|
mov.b32 %r237, %f535;
|
||
|
|
setp.eq.s32 %p62, %r237, 1118925336;
|
||
|
|
add.s32 %r238, %r237, -1;
|
||
|
|
mov.b32 %f539, %r238;
|
||
|
|
add.f32 %f540, %f538, 0f37000000;
|
||
|
|
selp.f32 %f541, %f539, %f535, %p62;
|
||
|
|
selp.f32 %f167, %f540, %f538, %p62;
|
||
|
|
mul.f32 %f542, %f541, 0f3FB8AA3B;
|
||
|
|
cvt.rzi.f32.f32 %f543, %f542;
|
||
|
|
fma.rn.f32 %f545, %f543, %f791, %f541;
|
||
|
|
fma.rn.f32 %f547, %f543, %f792, %f545;
|
||
|
|
mul.f32 %f548, %f547, 0f3FB8AA3B;
|
||
|
|
ex2.approx.ftz.f32 %f549, %f548;
|
||
|
|
add.f32 %f550, %f543, 0f00000000;
|
||
|
|
ex2.approx.f32 %f551, %f550;
|
||
|
|
mul.f32 %f552, %f549, %f551;
|
||
|
|
setp.lt.f32 %p63, %f541, 0fC2D20000;
|
||
|
|
selp.f32 %f553, 0f00000000, %f552, %p63;
|
||
|
|
setp.gt.f32 %p64, %f541, 0f42D20000;
|
||
|
|
selp.f32 %f891, 0f7F800000, %f553, %p64;
|
||
|
|
setp.eq.f32 %p65, %f891, 0f7F800000;
|
||
|
|
@%p65 bra BB0_70;
|
||
|
|
|
||
|
|
fma.rn.f32 %f891, %f891, %f167, %f891;
|
||
|
|
|
||
|
|
BB0_70:
|
||
|
|
setp.lt.f32 %p66, %f150, 0f00000000;
|
||
|
|
and.pred %p2, %p66, %p51;
|
||
|
|
mov.b32 %r239, %f891;
|
||
|
|
xor.b32 %r240, %r239, -2147483648;
|
||
|
|
mov.b32 %f554, %r240;
|
||
|
|
selp.f32 %f893, %f554, %f891, %p2;
|
||
|
|
setp.eq.f32 %p68, %f150, 0f00000000;
|
||
|
|
@%p68 bra BB0_73;
|
||
|
|
bra.uni BB0_71;
|
||
|
|
|
||
|
|
BB0_73:
|
||
|
|
add.f32 %f557, %f150, %f150;
|
||
|
|
selp.f32 %f893, %f557, 0f00000000, %p51;
|
||
|
|
bra.uni BB0_74;
|
||
|
|
|
||
|
|
BB0_71:
|
||
|
|
setp.geu.f32 %p69, %f150, 0f00000000;
|
||
|
|
@%p69 bra BB0_74;
|
||
|
|
|
||
|
|
mov.f32 %f803, 0f3EE66666;
|
||
|
|
cvt.rzi.f32.f32 %f556, %f803;
|
||
|
|
setp.neu.f32 %p70, %f556, 0f3EE66666;
|
||
|
|
selp.f32 %f893, 0f7FFFFFFF, %f893, %p70;
|
||
|
|
|
||
|
|
BB0_74:
|
||
|
|
abs.f32 %f807, %f150;
|
||
|
|
add.f32 %f558, %f807, 0f3EE66666;
|
||
|
|
mov.b32 %r241, %f558;
|
||
|
|
setp.lt.s32 %p72, %r241, 2139095040;
|
||
|
|
@%p72 bra BB0_79;
|
||
|
|
|
||
|
|
abs.f32 %f808, %f150;
|
||
|
|
setp.gtu.f32 %p73, %f808, 0f7F800000;
|
||
|
|
@%p73 bra BB0_78;
|
||
|
|
bra.uni BB0_76;
|
||
|
|
|
||
|
|
BB0_78:
|
||
|
|
add.f32 %f893, %f150, 0f3EE66666;
|
||
|
|
bra.uni BB0_79;
|
||
|
|
|
||
|
|
BB0_76:
|
||
|
|
abs.f32 %f809, %f150;
|
||
|
|
setp.neu.f32 %p74, %f809, 0f7F800000;
|
||
|
|
@%p74 bra BB0_79;
|
||
|
|
|
||
|
|
selp.f32 %f893, 0fFF800000, 0f7F800000, %p2;
|
||
|
|
|
||
|
|
BB0_79:
|
||
|
|
mov.f32 %f801, 0fB5BFBE8E;
|
||
|
|
mov.f32 %f800, 0fBF317200;
|
||
|
|
mov.f32 %f799, 0f00000000;
|
||
|
|
mov.f32 %f798, 0f35BFBE8E;
|
||
|
|
mov.f32 %f797, 0f3F317200;
|
||
|
|
mov.f32 %f796, 0f3DAAAABD;
|
||
|
|
mov.f32 %f795, 0f3C4CAF63;
|
||
|
|
mov.f32 %f794, 0f3B18F0FE;
|
||
|
|
mov.f32 %f793, 0f3EE66666;
|
||
|
|
setp.eq.f32 %p75, %f150, 0f3F800000;
|
||
|
|
selp.f32 %f178, 0f3F800000, %f893, %p75;
|
||
|
|
abs.f32 %f179, %f151;
|
||
|
|
setp.lt.f32 %p76, %f179, 0f00800000;
|
||
|
|
mul.f32 %f561, %f179, 0f4B800000;
|
||
|
|
selp.f32 %f562, 0fC3170000, 0fC2FE0000, %p76;
|
||
|
|
selp.f32 %f563, %f561, %f179, %p76;
|
||
|
|
mov.b32 %r242, %f563;
|
||
|
|
and.b32 %r243, %r242, 8388607;
|
||
|
|
or.b32 %r244, %r243, 1065353216;
|
||
|
|
mov.b32 %f564, %r244;
|
||
|
|
shr.u32 %r245, %r242, 23;
|
||
|
|
cvt.rn.f32.u32 %f565, %r245;
|
||
|
|
add.f32 %f566, %f562, %f565;
|
||
|
|
setp.gt.f32 %p77, %f564, 0f3FB504F3;
|
||
|
|
mul.f32 %f567, %f564, 0f3F000000;
|
||
|
|
add.f32 %f568, %f566, 0f3F800000;
|
||
|
|
selp.f32 %f569, %f567, %f564, %p77;
|
||
|
|
selp.f32 %f570, %f568, %f566, %p77;
|
||
|
|
add.f32 %f571, %f569, 0fBF800000;
|
||
|
|
add.f32 %f560, %f569, 0f3F800000;
|
||
|
|
// inline asm
|
||
|
|
rcp.approx.ftz.f32 %f559,%f560;
|
||
|
|
// inline asm
|
||
|
|
add.f32 %f572, %f571, %f571;
|
||
|
|
mul.f32 %f573, %f559, %f572;
|
||
|
|
mul.f32 %f574, %f573, %f573;
|
||
|
|
fma.rn.f32 %f577, %f794, %f574, %f795;
|
||
|
|
fma.rn.f32 %f579, %f577, %f574, %f796;
|
||
|
|
mul.rn.f32 %f580, %f579, %f574;
|
||
|
|
mul.rn.f32 %f581, %f580, %f573;
|
||
|
|
sub.f32 %f582, %f571, %f573;
|
||
|
|
neg.f32 %f583, %f573;
|
||
|
|
add.f32 %f584, %f582, %f582;
|
||
|
|
fma.rn.f32 %f585, %f583, %f571, %f584;
|
||
|
|
mul.rn.f32 %f586, %f559, %f585;
|
||
|
|
add.f32 %f587, %f581, %f573;
|
||
|
|
sub.f32 %f588, %f573, %f587;
|
||
|
|
add.f32 %f589, %f581, %f588;
|
||
|
|
add.f32 %f590, %f586, %f589;
|
||
|
|
add.f32 %f591, %f587, %f590;
|
||
|
|
sub.f32 %f592, %f587, %f591;
|
||
|
|
add.f32 %f593, %f590, %f592;
|
||
|
|
mul.rn.f32 %f595, %f570, %f797;
|
||
|
|
mul.rn.f32 %f597, %f570, %f798;
|
||
|
|
add.f32 %f598, %f595, %f591;
|
||
|
|
sub.f32 %f599, %f595, %f598;
|
||
|
|
add.f32 %f600, %f591, %f599;
|
||
|
|
add.f32 %f601, %f593, %f600;
|
||
|
|
add.f32 %f602, %f597, %f601;
|
||
|
|
add.f32 %f603, %f598, %f602;
|
||
|
|
sub.f32 %f604, %f598, %f603;
|
||
|
|
add.f32 %f605, %f602, %f604;
|
||
|
|
mul.rn.f32 %f607, %f793, %f603;
|
||
|
|
neg.f32 %f608, %f607;
|
||
|
|
fma.rn.f32 %f609, %f793, %f603, %f608;
|
||
|
|
fma.rn.f32 %f610, %f793, %f605, %f609;
|
||
|
|
fma.rn.f32 %f612, %f799, %f603, %f610;
|
||
|
|
add.rn.f32 %f613, %f607, %f612;
|
||
|
|
neg.f32 %f614, %f613;
|
||
|
|
add.rn.f32 %f615, %f607, %f614;
|
||
|
|
add.rn.f32 %f616, %f615, %f612;
|
||
|
|
mov.b32 %r246, %f613;
|
||
|
|
setp.eq.s32 %p78, %r246, 1118925336;
|
||
|
|
add.s32 %r247, %r246, -1;
|
||
|
|
mov.b32 %f617, %r247;
|
||
|
|
add.f32 %f618, %f616, 0f37000000;
|
||
|
|
selp.f32 %f619, %f617, %f613, %p78;
|
||
|
|
selp.f32 %f180, %f618, %f616, %p78;
|
||
|
|
mul.f32 %f620, %f619, 0f3FB8AA3B;
|
||
|
|
cvt.rzi.f32.f32 %f621, %f620;
|
||
|
|
fma.rn.f32 %f623, %f621, %f800, %f619;
|
||
|
|
fma.rn.f32 %f625, %f621, %f801, %f623;
|
||
|
|
mul.f32 %f626, %f625, 0f3FB8AA3B;
|
||
|
|
ex2.approx.ftz.f32 %f627, %f626;
|
||
|
|
add.f32 %f628, %f621, 0f00000000;
|
||
|
|
ex2.approx.f32 %f629, %f628;
|
||
|
|
mul.f32 %f630, %f627, %f629;
|
||
|
|
setp.lt.f32 %p79, %f619, 0fC2D20000;
|
||
|
|
selp.f32 %f631, 0f00000000, %f630, %p79;
|
||
|
|
setp.gt.f32 %p80, %f619, 0f42D20000;
|
||
|
|
selp.f32 %f894, 0f7F800000, %f631, %p80;
|
||
|
|
setp.eq.f32 %p81, %f894, 0f7F800000;
|
||
|
|
@%p81 bra BB0_81;
|
||
|
|
|
||
|
|
fma.rn.f32 %f894, %f894, %f180, %f894;
|
||
|
|
|
||
|
|
BB0_81:
|
||
|
|
setp.lt.f32 %p82, %f151, 0f00000000;
|
||
|
|
and.pred %p3, %p82, %p51;
|
||
|
|
mov.b32 %r248, %f894;
|
||
|
|
xor.b32 %r249, %r248, -2147483648;
|
||
|
|
mov.b32 %f632, %r249;
|
||
|
|
selp.f32 %f896, %f632, %f894, %p3;
|
||
|
|
setp.eq.f32 %p84, %f151, 0f00000000;
|
||
|
|
@%p84 bra BB0_84;
|
||
|
|
bra.uni BB0_82;
|
||
|
|
|
||
|
|
BB0_84:
|
||
|
|
add.f32 %f635, %f151, %f151;
|
||
|
|
selp.f32 %f896, %f635, 0f00000000, %p51;
|
||
|
|
bra.uni BB0_85;
|
||
|
|
|
||
|
|
BB0_82:
|
||
|
|
setp.geu.f32 %p85, %f151, 0f00000000;
|
||
|
|
@%p85 bra BB0_85;
|
||
|
|
|
||
|
|
mov.f32 %f802, 0f3EE66666;
|
||
|
|
cvt.rzi.f32.f32 %f634, %f802;
|
||
|
|
setp.neu.f32 %p86, %f634, 0f3EE66666;
|
||
|
|
selp.f32 %f896, 0f7FFFFFFF, %f896, %p86;
|
||
|
|
|
||
|
|
BB0_85:
|
||
|
|
abs.f32 %f810, %f151;
|
||
|
|
add.f32 %f636, %f810, 0f3EE66666;
|
||
|
|
mov.b32 %r250, %f636;
|
||
|
|
setp.lt.s32 %p88, %r250, 2139095040;
|
||
|
|
@%p88 bra BB0_90;
|
||
|
|
|
||
|
|
abs.f32 %f811, %f151;
|
||
|
|
setp.gtu.f32 %p89, %f811, 0f7F800000;
|
||
|
|
@%p89 bra BB0_89;
|
||
|
|
bra.uni BB0_87;
|
||
|
|
|
||
|
|
BB0_89:
|
||
|
|
add.f32 %f896, %f151, 0f3EE66666;
|
||
|
|
bra.uni BB0_90;
|
||
|
|
|
||
|
|
BB0_87:
|
||
|
|
abs.f32 %f812, %f151;
|
||
|
|
setp.neu.f32 %p90, %f812, 0f7F800000;
|
||
|
|
@%p90 bra BB0_90;
|
||
|
|
|
||
|
|
selp.f32 %f896, 0fFF800000, 0f7F800000, %p3;
|
||
|
|
|
||
|
|
BB0_90:
|
||
|
|
mov.u32 %r341, 4;
|
||
|
|
setp.eq.f32 %p91, %f151, 0f3F800000;
|
||
|
|
selp.f32 %f637, 0f3F800000, %f896, %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, %r341, %rd65, %rd66, %rd25, %rd25);
|
||
|
|
// inline asm
|
||
|
|
cvt.sat.f32.f32 %f638, %f637;
|
||
|
|
mul.f32 %f639, %f638, 0f437FFD71;
|
||
|
|
cvt.rzi.u32.f32 %r253, %f639;
|
||
|
|
cvt.sat.f32.f32 %f640, %f178;
|
||
|
|
mul.f32 %f641, %f640, 0f437FFD71;
|
||
|
|
cvt.rzi.u32.f32 %r254, %f641;
|
||
|
|
cvt.sat.f32.f32 %f642, %f165;
|
||
|
|
mul.f32 %f643, %f642, 0f437FFD71;
|
||
|
|
cvt.rzi.u32.f32 %r255, %f643;
|
||
|
|
cvt.u16.u32 %rs14, %r253;
|
||
|
|
cvt.u16.u32 %rs15, %r255;
|
||
|
|
cvt.u16.u32 %rs16, %r254;
|
||
|
|
mov.u16 %rs17, 255;
|
||
|
|
st.v4.u8 [%rd63], {%rs14, %rs16, %rs15, %rs17};
|
||
|
|
ld.global.u32 %r382, [imageEnabled];
|
||
|
|
|
||
|
|
BB0_91:
|
||
|
|
cvt.u64.u32 %rd16, %r4;
|
||
|
|
cvt.u64.u32 %rd17, %r5;
|
||
|
|
and.b32 %r256, %r382, 4;
|
||
|
|
setp.eq.s32 %p92, %r256, 0;
|
||
|
|
@%p92 bra BB0_95;
|
||
|
|
|
||
|
|
ld.global.u32 %r257, [additive];
|
||
|
|
setp.eq.s32 %p93, %r257, 0;
|
||
|
|
mov.f32 %f644, 0f3F800000;
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs18, %f644;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
@%p93 bra BB0_94;
|
||
|
|
|
||
|
|
mov.u64 %rd82, image_HDR;
|
||
|
|
cvta.global.u64 %rd71, %rd82;
|
||
|
|
mov.u32 %r261, 8;
|
||
|
|
// inline asm
|
||
|
|
call (%rd70), _rt_buffer_get_64, (%rd71, %r92, %r261, %rd16, %rd17, %rd25, %rd25);
|
||
|
|
// inline asm
|
||
|
|
ld.v4.u16 {%rs25, %rs26, %rs27, %rs28}, [%rd70];
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f645, %rs25;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f646, %rs26;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f647, %rs27;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
call (%rd76), _rt_buffer_get_64, (%rd71, %r92, %r261, %rd16, %rd17, %rd25, %rd25);
|
||
|
|
// inline asm
|
||
|
|
add.f32 %f648, %f149, %f645;
|
||
|
|
add.f32 %f649, %f150, %f646;
|
||
|
|
add.f32 %f650, %f151, %f647;
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs24, %f650;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs23, %f649;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs22, %f648;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
st.v4.u16 [%rd76], {%rs22, %rs23, %rs24, %rs18};
|
||
|
|
bra.uni BB0_95;
|
||
|
|
|
||
|
|
BB0_94:
|
||
|
|
mov.u64 %rd89, image_HDR;
|
||
|
|
cvta.global.u64 %rd84, %rd89;
|
||
|
|
mov.u32 %r263, 8;
|
||
|
|
// inline asm
|
||
|
|
call (%rd83), _rt_buffer_get_64, (%rd84, %r92, %r263, %rd16, %rd17, %rd25, %rd25);
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs31, %f151;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs30, %f150;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs29, %f149;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
st.v4.u16 [%rd83], {%rs29, %rs30, %rs31, %rs18};
|
||
|
|
|
||
|
|
BB0_95:
|
||
|
|
ld.global.f32 %f655, [skyColor];
|
||
|
|
mul.f32 %f656, %f137, %f655;
|
||
|
|
ld.global.f32 %f657, [skyColor+4];
|
||
|
|
mul.f32 %f658, %f138, %f657;
|
||
|
|
ld.global.f32 %f659, [skyColor+8];
|
||
|
|
mul.f32 %f660, %f139, %f659;
|
||
|
|
mul.f32 %f191, %f140, %f655;
|
||
|
|
mul.f32 %f192, %f141, %f657;
|
||
|
|
mul.f32 %f193, %f142, %f659;
|
||
|
|
mul.f32 %f194, %f143, %f655;
|
||
|
|
mul.f32 %f195, %f144, %f657;
|
||
|
|
mul.f32 %f196, %f145, %f659;
|
||
|
|
mul.f32 %f197, %f146, %f655;
|
||
|
|
mul.f32 %f198, %f147, %f657;
|
||
|
|
mul.f32 %f199, %f148, %f659;
|
||
|
|
mul.f32 %f200, %f656, 0f3F000000;
|
||
|
|
mul.f32 %f201, %f658, 0f3F000000;
|
||
|
|
mul.f32 %f202, %f660, 0f3F000000;
|
||
|
|
ld.global.u32 %r264, [additive];
|
||
|
|
setp.eq.s32 %p94, %r264, 0;
|
||
|
|
mov.f32 %f654, 0f3F800000;
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs32, %f654;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
@%p94 bra BB0_97;
|
||
|
|
|
||
|
|
mov.u64 %rd102, image_RNM0;
|
||
|
|
cvta.global.u64 %rd91, %rd102;
|
||
|
|
mov.u32 %r268, 8;
|
||
|
|
// inline asm
|
||
|
|
call (%rd90), _rt_buffer_get_64, (%rd91, %r92, %r268, %rd16, %rd17, %rd25, %rd25);
|
||
|
|
// inline asm
|
||
|
|
ld.v4.u16 {%rs39, %rs40, %rs41, %rs42}, [%rd90];
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f661, %rs39;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f662, %rs40;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f663, %rs41;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
call (%rd96), _rt_buffer_get_64, (%rd91, %r92, %r268, %rd16, %rd17, %rd25, %rd25);
|
||
|
|
// inline asm
|
||
|
|
add.f32 %f664, %f200, %f661;
|
||
|
|
add.f32 %f665, %f201, %f662;
|
||
|
|
add.f32 %f666, %f202, %f663;
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs38, %f666;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs37, %f665;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs36, %f664;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
st.v4.u16 [%rd96], {%rs36, %rs37, %rs38, %rs32};
|
||
|
|
bra.uni BB0_98;
|
||
|
|
|
||
|
|
BB0_97:
|
||
|
|
mov.u64 %rd109, image_RNM0;
|
||
|
|
cvta.global.u64 %rd104, %rd109;
|
||
|
|
mov.u32 %r270, 8;
|
||
|
|
// inline asm
|
||
|
|
call (%rd103), _rt_buffer_get_64, (%rd104, %r92, %r270, %rd16, %rd17, %rd25, %rd25);
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs45, %f202;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs44, %f201;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs43, %f200;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
st.v4.u16 [%rd103], {%rs43, %rs44, %rs45, %rs32};
|
||
|
|
|
||
|
|
BB0_98:
|
||
|
|
mov.f32 %f671, 0f34000000;
|
||
|
|
max.f32 %f672, %f200, %f671;
|
||
|
|
mul.f32 %f673, %f191, 0f3F000000;
|
||
|
|
div.rn.f32 %f674, %f673, %f672;
|
||
|
|
max.f32 %f675, %f201, %f671;
|
||
|
|
mul.f32 %f676, %f192, 0f3F000000;
|
||
|
|
div.rn.f32 %f677, %f676, %f675;
|
||
|
|
max.f32 %f678, %f202, %f671;
|
||
|
|
mul.f32 %f679, %f193, 0f3F000000;
|
||
|
|
div.rn.f32 %f680, %f679, %f678;
|
||
|
|
fma.rn.f32 %f203, %f674, 0f3F000000, 0f3F000000;
|
||
|
|
fma.rn.f32 %f204, %f677, 0f3F000000, 0f3F000000;
|
||
|
|
fma.rn.f32 %f205, %f680, 0f3F000000, 0f3F000000;
|
||
|
|
mul.f32 %f681, %f194, 0f3F000000;
|
||
|
|
div.rn.f32 %f682, %f681, %f672;
|
||
|
|
mul.f32 %f683, %f195, 0f3F000000;
|
||
|
|
div.rn.f32 %f684, %f683, %f675;
|
||
|
|
mul.f32 %f685, %f196, 0f3F000000;
|
||
|
|
div.rn.f32 %f686, %f685, %f678;
|
||
|
|
fma.rn.f32 %f206, %f682, 0f3F000000, 0f3F000000;
|
||
|
|
fma.rn.f32 %f207, %f684, 0f3F000000, 0f3F000000;
|
||
|
|
fma.rn.f32 %f208, %f686, 0f3F000000, 0f3F000000;
|
||
|
|
mul.f32 %f687, %f197, 0f3F000000;
|
||
|
|
div.rn.f32 %f688, %f687, %f672;
|
||
|
|
mul.f32 %f689, %f198, 0f3F000000;
|
||
|
|
div.rn.f32 %f690, %f689, %f675;
|
||
|
|
mul.f32 %f691, %f199, 0f3F000000;
|
||
|
|
div.rn.f32 %f692, %f691, %f678;
|
||
|
|
fma.rn.f32 %f209, %f688, 0f3F000000, 0f3F000000;
|
||
|
|
fma.rn.f32 %f210, %f690, 0f3F000000, 0f3F000000;
|
||
|
|
fma.rn.f32 %f211, %f692, 0f3F000000, 0f3F000000;
|
||
|
|
ld.global.u32 %r271, [additive];
|
||
|
|
setp.eq.s32 %p95, %r271, 0;
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs46, %f654;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
@%p95 bra BB0_100;
|
||
|
|
|
||
|
|
mov.u64 %rd122, image_RNM1;
|
||
|
|
cvta.global.u64 %rd111, %rd122;
|
||
|
|
mov.u32 %r275, 8;
|
||
|
|
// inline asm
|
||
|
|
call (%rd110), _rt_buffer_get_64, (%rd111, %r92, %r275, %rd16, %rd17, %rd25, %rd25);
|
||
|
|
// inline asm
|
||
|
|
ld.v4.u16 {%rs53, %rs54, %rs55, %rs56}, [%rd110];
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f693, %rs53;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f694, %rs54;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f695, %rs55;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
call (%rd116), _rt_buffer_get_64, (%rd111, %r92, %r275, %rd16, %rd17, %rd25, %rd25);
|
||
|
|
// inline asm
|
||
|
|
add.f32 %f696, %f203, %f693;
|
||
|
|
add.f32 %f697, %f204, %f694;
|
||
|
|
add.f32 %f698, %f205, %f695;
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs52, %f698;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs51, %f697;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs50, %f696;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
st.v4.u16 [%rd116], {%rs50, %rs51, %rs52, %rs46};
|
||
|
|
bra.uni BB0_101;
|
||
|
|
|
||
|
|
BB0_100:
|
||
|
|
mov.u64 %rd129, image_RNM1;
|
||
|
|
cvta.global.u64 %rd124, %rd129;
|
||
|
|
mov.u32 %r277, 8;
|
||
|
|
// inline asm
|
||
|
|
call (%rd123), _rt_buffer_get_64, (%rd124, %r92, %r277, %rd16, %rd17, %rd25, %rd25);
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs59, %f205;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs58, %f204;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs57, %f203;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
st.v4.u16 [%rd123], {%rs57, %rs58, %rs59, %rs46};
|
||
|
|
|
||
|
|
BB0_101:
|
||
|
|
ld.global.u32 %r278, [additive];
|
||
|
|
setp.eq.s32 %p96, %r278, 0;
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs60, %f654;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
@%p96 bra BB0_103;
|
||
|
|
|
||
|
|
mov.u64 %rd142, image_RNM2;
|
||
|
|
cvta.global.u64 %rd131, %rd142;
|
||
|
|
mov.u32 %r282, 8;
|
||
|
|
// inline asm
|
||
|
|
call (%rd130), _rt_buffer_get_64, (%rd131, %r92, %r282, %rd16, %rd17, %rd25, %rd25);
|
||
|
|
// inline asm
|
||
|
|
ld.v4.u16 {%rs67, %rs68, %rs69, %rs70}, [%rd130];
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f703, %rs67;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f704, %rs68;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f705, %rs69;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
call (%rd136), _rt_buffer_get_64, (%rd131, %r92, %r282, %rd16, %rd17, %rd25, %rd25);
|
||
|
|
// inline asm
|
||
|
|
add.f32 %f706, %f206, %f703;
|
||
|
|
add.f32 %f707, %f207, %f704;
|
||
|
|
add.f32 %f708, %f208, %f705;
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs66, %f708;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs65, %f707;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs64, %f706;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
st.v4.u16 [%rd136], {%rs64, %rs65, %rs66, %rs60};
|
||
|
|
bra.uni BB0_104;
|
||
|
|
|
||
|
|
BB0_103:
|
||
|
|
mov.u64 %rd149, image_RNM2;
|
||
|
|
cvta.global.u64 %rd144, %rd149;
|
||
|
|
mov.u32 %r284, 8;
|
||
|
|
// inline asm
|
||
|
|
call (%rd143), _rt_buffer_get_64, (%rd144, %r92, %r284, %rd16, %rd17, %rd25, %rd25);
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs73, %f208;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs72, %f207;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs71, %f206;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
st.v4.u16 [%rd143], {%rs71, %rs72, %rs73, %rs60};
|
||
|
|
|
||
|
|
BB0_104:
|
||
|
|
ld.global.u32 %r285, [additive];
|
||
|
|
setp.eq.s32 %p97, %r285, 0;
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs74, %f654;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
@%p97 bra BB0_106;
|
||
|
|
|
||
|
|
mov.u64 %rd162, image_RNM3;
|
||
|
|
cvta.global.u64 %rd151, %rd162;
|
||
|
|
mov.u32 %r289, 8;
|
||
|
|
// inline asm
|
||
|
|
call (%rd150), _rt_buffer_get_64, (%rd151, %r92, %r289, %rd16, %rd17, %rd25, %rd25);
|
||
|
|
// inline asm
|
||
|
|
ld.v4.u16 {%rs81, %rs82, %rs83, %rs84}, [%rd150];
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f713, %rs81;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f714, %rs82;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f715, %rs83;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
call (%rd156), _rt_buffer_get_64, (%rd151, %r92, %r289, %rd16, %rd17, %rd25, %rd25);
|
||
|
|
// inline asm
|
||
|
|
add.f32 %f716, %f209, %f713;
|
||
|
|
add.f32 %f717, %f210, %f714;
|
||
|
|
add.f32 %f718, %f211, %f715;
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs80, %f718;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs79, %f717;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs78, %f716;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
st.v4.u16 [%rd156], {%rs78, %rs79, %rs80, %rs74};
|
||
|
|
bra.uni BB0_125;
|
||
|
|
|
||
|
|
BB0_106:
|
||
|
|
mov.u64 %rd169, image_RNM3;
|
||
|
|
cvta.global.u64 %rd164, %rd169;
|
||
|
|
mov.u32 %r291, 8;
|
||
|
|
// inline asm
|
||
|
|
call (%rd163), _rt_buffer_get_64, (%rd164, %r92, %r291, %rd16, %rd17, %rd25, %rd25);
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs87, %f211;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs86, %f210;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs85, %f209;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
st.v4.u16 [%rd163], {%rs85, %rs86, %rs87, %rs74};
|
||
|
|
|
||
|
|
BB0_125:
|
||
|
|
ret;
|
||
|
|
}
|
||
|
|
|
||
|
|
|