2377 lines
66 KiB
Plaintext
2377 lines
66 KiB
Plaintext
//
|
|
// Generated by NVIDIA NVVM Compiler
|
|
//
|
|
// Compiler Build ID: CL-23083092
|
|
// Cuda compilation tools, release 9.1, V9.1.85
|
|
// Based on LLVM 3.4svn
|
|
//
|
|
|
|
.version 6.1
|
|
.target sm_30
|
|
.address_size 64
|
|
|
|
// .globl _Z6oxMainv
|
|
.global .align 8 .b8 pixelID[8];
|
|
.global .align 8 .b8 resolution[8];
|
|
.global .align 4 .b8 normal[12];
|
|
.global .align 4 .b8 camPos[12];
|
|
.global .align 4 .b8 root[4];
|
|
.global .align 4 .u32 imageEnabled;
|
|
.global .texref lightmap;
|
|
.global .align 16 .b8 tileInfo[16];
|
|
.global .align 4 .u32 additive;
|
|
.global .align 1 .b8 image[1];
|
|
.global .align 1 .b8 image_HDR[1];
|
|
.global .align 1 .b8 image_HDR2[1];
|
|
.global .align 1 .b8 image_Mask[1];
|
|
.global .align 1 .b8 image_RNM0[1];
|
|
.global .align 1 .b8 image_RNM1[1];
|
|
.global .align 1 .b8 image_RNM2[1];
|
|
.global .align 1 .b8 image_RNM3[1];
|
|
.global .align 1 .b8 uvpos[1];
|
|
.global .align 1 .b8 uvnormal[1];
|
|
.global .align 1 .b8 lightMeshBuffer[1];
|
|
.global .align 4 .u32 lightMeshBufferSize;
|
|
.global .align 4 .f32 lightInvCutoff;
|
|
.global .align 4 .f32 lightPointSize;
|
|
.global .align 4 .b8 lightColor[12];
|
|
.global .align 1 .b8 rnd_seeds[1];
|
|
.global .align 4 .u32 samples;
|
|
.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_typeinfo19lightMeshBufferSizeE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo14lightInvCutoffE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo14lightPointSizeE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo10lightColorE[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 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 16 .b8 _ZN21rti_internal_typename19lightMeshBufferSizeE[13] = {117, 110, 115, 105, 103, 110, 101, 100, 32, 105, 110, 116, 0};
|
|
.global .align 8 .b8 _ZN21rti_internal_typename14lightInvCutoffE[6] = {102, 108, 111, 97, 116, 0};
|
|
.global .align 8 .b8 _ZN21rti_internal_typename14lightPointSizeE[6] = {102, 108, 111, 97, 116, 0};
|
|
.global .align 8 .b8 _ZN21rti_internal_typename10lightColorE[7] = {102, 108, 111, 97, 116, 51, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typename7samplesE[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_typeenum19lightMeshBufferSizeE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum14lightInvCutoffE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum14lightPointSizeE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum10lightColorE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum7samplesE = 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_semantic19lightMeshBufferSizeE[1];
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic14lightInvCutoffE[1];
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic14lightPointSizeE[1];
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic10lightColorE[1];
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic7samplesE[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_annotation19lightMeshBufferSizeE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation14lightInvCutoffE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation14lightPointSizeE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation10lightColorE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation7samplesE[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[36];
|
|
.reg .b64 %SP;
|
|
.reg .b64 %SPL;
|
|
.reg .pred %p<151>;
|
|
.reg .b16 %rs<152>;
|
|
.reg .f32 %f<1034>;
|
|
.reg .b32 %r<415>;
|
|
.reg .b64 %rd<308>;
|
|
|
|
|
|
mov.u64 %rd307, __local_depot0;
|
|
cvta.local.u64 %SP, %rd307;
|
|
ld.global.u32 %r1, [samples];
|
|
ld.global.v2.u32 {%r103, %r104}, [pixelID];
|
|
cvt.u64.u32 %rd23, %r103;
|
|
cvt.u64.u32 %rd24, %r104;
|
|
mov.u64 %rd27, uvnormal;
|
|
cvta.global.u64 %rd22, %rd27;
|
|
mov.u32 %r101, 2;
|
|
mov.u32 %r102, 4;
|
|
mov.u64 %rd26, 0;
|
|
// inline asm
|
|
call (%rd21), _rt_buffer_get_64, (%rd22, %r101, %r102, %rd23, %rd24, %rd26, %rd26);
|
|
// inline asm
|
|
ld.u32 %r2, [%rd21];
|
|
shr.u32 %r107, %r2, 16;
|
|
cvt.u16.u32 %rs1, %r107;
|
|
and.b16 %rs7, %rs1, 255;
|
|
cvt.u16.u32 %rs8, %r2;
|
|
or.b16 %rs9, %rs8, %rs7;
|
|
setp.eq.s16 %p7, %rs9, 0;
|
|
mov.f32 %f956, 0f00000000;
|
|
mov.f32 %f957, %f956;
|
|
mov.f32 %f958, %f956;
|
|
@%p7 bra BB0_2;
|
|
|
|
ld.u8 %rs10, [%rd21+1];
|
|
and.b16 %rs12, %rs8, 255;
|
|
cvt.rn.f32.u16 %f231, %rs12;
|
|
div.rn.f32 %f232, %f231, 0f437F0000;
|
|
fma.rn.f32 %f233, %f232, 0f40000000, 0fBF800000;
|
|
cvt.rn.f32.u16 %f234, %rs10;
|
|
div.rn.f32 %f235, %f234, 0f437F0000;
|
|
fma.rn.f32 %f236, %f235, 0f40000000, 0fBF800000;
|
|
cvt.rn.f32.u16 %f237, %rs7;
|
|
div.rn.f32 %f238, %f237, 0f437F0000;
|
|
fma.rn.f32 %f239, %f238, 0f40000000, 0fBF800000;
|
|
mul.f32 %f240, %f236, %f236;
|
|
fma.rn.f32 %f241, %f233, %f233, %f240;
|
|
fma.rn.f32 %f242, %f239, %f239, %f241;
|
|
sqrt.rn.f32 %f243, %f242;
|
|
rcp.rn.f32 %f244, %f243;
|
|
mul.f32 %f956, %f233, %f244;
|
|
mul.f32 %f957, %f236, %f244;
|
|
mul.f32 %f958, %f239, %f244;
|
|
|
|
BB0_2:
|
|
ld.global.v2.u32 {%r108, %r109}, [pixelID];
|
|
ld.global.v2.u32 {%r111, %r112}, [tileInfo];
|
|
add.s32 %r3, %r108, %r111;
|
|
add.s32 %r4, %r109, %r112;
|
|
setp.eq.f32 %p8, %f957, 0f00000000;
|
|
setp.eq.f32 %p9, %f956, 0f00000000;
|
|
and.pred %p10, %p9, %p8;
|
|
setp.eq.f32 %p11, %f958, 0f00000000;
|
|
and.pred %p12, %p10, %p11;
|
|
@%p12 bra BB0_141;
|
|
bra.uni BB0_3;
|
|
|
|
BB0_141:
|
|
ld.global.u32 %r414, [imageEnabled];
|
|
and.b32 %r336, %r414, 1;
|
|
setp.eq.b32 %p142, %r336, 1;
|
|
@!%p142 bra BB0_143;
|
|
bra.uni BB0_142;
|
|
|
|
BB0_142:
|
|
cvt.u64.u32 %rd189, %r3;
|
|
cvt.u64.u32 %rd190, %r4;
|
|
mov.u64 %rd193, image;
|
|
cvta.global.u64 %rd188, %rd193;
|
|
mov.u64 %rd192, 0;
|
|
// inline asm
|
|
call (%rd187), _rt_buffer_get_64, (%rd188, %r101, %r102, %rd189, %rd190, %rd192, %rd192);
|
|
// inline asm
|
|
mov.u16 %rs84, 0;
|
|
st.v4.u8 [%rd187], {%rs84, %rs84, %rs84, %rs84};
|
|
ld.global.u32 %r414, [imageEnabled];
|
|
|
|
BB0_143:
|
|
and.b32 %r339, %r414, 8;
|
|
setp.eq.s32 %p143, %r339, 0;
|
|
@%p143 bra BB0_145;
|
|
|
|
cvt.u64.u32 %rd197, %r4;
|
|
cvt.u64.u32 %rd196, %r3;
|
|
mov.u64 %rd200, image_Mask;
|
|
cvta.global.u64 %rd195, %rd200;
|
|
mov.u64 %rd199, 0;
|
|
// inline asm
|
|
call (%rd194), _rt_buffer_get_64, (%rd195, %r101, %r101, %rd196, %rd197, %rd199, %rd199);
|
|
// inline asm
|
|
mov.f32 %f911, 0f00000000;
|
|
cvt.rzi.u32.f32 %r342, %f911;
|
|
cvt.u16.u32 %rs85, %r342;
|
|
mov.u16 %rs86, 0;
|
|
st.v2.u8 [%rd194], {%rs85, %rs86};
|
|
ld.global.u32 %r414, [imageEnabled];
|
|
|
|
BB0_145:
|
|
cvt.u64.u32 %rd19, %r3;
|
|
cvt.u64.u32 %rd20, %r4;
|
|
and.b32 %r343, %r414, 4;
|
|
setp.eq.s32 %p144, %r343, 0;
|
|
@%p144 bra BB0_149;
|
|
|
|
ld.global.u32 %r344, [additive];
|
|
setp.eq.s32 %p145, %r344, 0;
|
|
@%p145 bra BB0_148;
|
|
|
|
mov.u64 %rd213, image_HDR;
|
|
cvta.global.u64 %rd202, %rd213;
|
|
mov.u32 %r348, 8;
|
|
mov.u64 %rd212, 0;
|
|
// inline asm
|
|
call (%rd201), _rt_buffer_get_64, (%rd202, %r101, %r348, %rd19, %rd20, %rd212, %rd212);
|
|
// inline asm
|
|
ld.v4.u16 {%rs93, %rs94, %rs95, %rs96}, [%rd201];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f912, %rs93;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f913, %rs94;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f914, %rs95;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd207), _rt_buffer_get_64, (%rd202, %r101, %r348, %rd19, %rd20, %rd212, %rd212);
|
|
// inline asm
|
|
add.f32 %f915, %f912, 0f00000000;
|
|
add.f32 %f916, %f913, 0f00000000;
|
|
add.f32 %f917, %f914, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs92, %f917;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs91, %f916;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs90, %f915;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs97, 0;
|
|
st.v4.u16 [%rd207], {%rs90, %rs91, %rs92, %rs97};
|
|
bra.uni BB0_149;
|
|
|
|
BB0_3:
|
|
ld.global.v2.u32 {%r121, %r122}, [pixelID];
|
|
cvt.u64.u32 %rd30, %r121;
|
|
cvt.u64.u32 %rd31, %r122;
|
|
mov.u64 %rd40, uvpos;
|
|
cvta.global.u64 %rd29, %rd40;
|
|
mov.u32 %r117, 12;
|
|
// inline asm
|
|
call (%rd28), _rt_buffer_get_64, (%rd29, %r101, %r117, %rd30, %rd31, %rd26, %rd26);
|
|
// inline asm
|
|
ld.f32 %f9, [%rd28+8];
|
|
ld.f32 %f8, [%rd28+4];
|
|
ld.f32 %f7, [%rd28];
|
|
mul.f32 %f251, %f7, 0f3456BF95;
|
|
mul.f32 %f252, %f8, 0f3456BF95;
|
|
mul.f32 %f253, %f9, 0f3456BF95;
|
|
abs.f32 %f10, %f956;
|
|
div.rn.f32 %f254, %f251, %f10;
|
|
abs.f32 %f255, %f957;
|
|
div.rn.f32 %f256, %f252, %f255;
|
|
abs.f32 %f11, %f958;
|
|
div.rn.f32 %f257, %f253, %f11;
|
|
abs.f32 %f258, %f254;
|
|
abs.f32 %f259, %f256;
|
|
abs.f32 %f260, %f257;
|
|
mov.f32 %f261, 0f38D1B717;
|
|
max.f32 %f262, %f258, %f261;
|
|
max.f32 %f263, %f259, %f261;
|
|
max.f32 %f264, %f260, %f261;
|
|
fma.rn.f32 %f12, %f956, %f262, %f7;
|
|
fma.rn.f32 %f13, %f957, %f263, %f8;
|
|
fma.rn.f32 %f14, %f958, %f264, %f9;
|
|
ld.global.v2.u32 {%r125, %r126}, [pixelID];
|
|
cvt.u64.u32 %rd36, %r125;
|
|
cvt.u64.u32 %rd37, %r126;
|
|
mov.u64 %rd41, rnd_seeds;
|
|
cvta.global.u64 %rd35, %rd41;
|
|
// inline asm
|
|
call (%rd34), _rt_buffer_get_64, (%rd35, %r101, %r102, %rd36, %rd37, %rd26, %rd26);
|
|
// inline asm
|
|
ld.u32 %r129, [%rd34];
|
|
mad.lo.s32 %r5, %r129, 1664525, 1013904223;
|
|
ld.global.u32 %r130, [lightMeshBufferSize];
|
|
setp.eq.s32 %p14, %r130, 0;
|
|
mov.pred %p13, 0;
|
|
mov.f32 %f20, 0f00000000;
|
|
mov.u32 %r7, 0;
|
|
@%p14 bra BB0_4;
|
|
|
|
ld.global.f32 %f15, [lightPointSize];
|
|
mul.f32 %f16, %f12, 0f3456BF95;
|
|
mul.f32 %f17, %f13, 0f3456BF95;
|
|
mul.f32 %f18, %f14, 0f3456BF95;
|
|
and.b32 %r133, %r5, 16777215;
|
|
cvt.rn.f32.u32 %f271, %r133;
|
|
mul.f32 %f272, %f271, 0fB3800000;
|
|
fma.rn.f32 %f19, %f272, 0f3F333333, 0f3F800000;
|
|
mov.f32 %f20, 0f00000000;
|
|
mov.u32 %r383, 0;
|
|
abs.f32 %f394, %f17;
|
|
abs.f32 %f395, %f16;
|
|
max.f32 %f396, %f395, %f394;
|
|
abs.f32 %f397, %f18;
|
|
max.f32 %f398, %f396, %f397;
|
|
mov.u32 %r7, %r383;
|
|
mov.f32 %f21, %f20;
|
|
mov.f32 %f22, %f20;
|
|
mov.f32 %f23, %f20;
|
|
mov.f32 %f24, %f20;
|
|
mov.f32 %f25, %f20;
|
|
|
|
BB0_6:
|
|
shl.b32 %r8, %r383, 1;
|
|
cvt.s64.s32 %rd44, %r8;
|
|
mov.u64 %rd48, lightMeshBuffer;
|
|
cvta.global.u64 %rd43, %rd48;
|
|
mov.u32 %r134, 1;
|
|
// inline asm
|
|
call (%rd42), _rt_buffer_get_64, (%rd43, %r134, %r117, %rd44, %rd26, %rd26, %rd26);
|
|
// inline asm
|
|
ld.f32 %f273, [%rd42];
|
|
sub.f32 %f274, %f273, %f7;
|
|
ld.f32 %f275, [%rd42+4];
|
|
sub.f32 %f276, %f275, %f8;
|
|
ld.f32 %f277, [%rd42+8];
|
|
sub.f32 %f278, %f277, %f9;
|
|
mul.f32 %f279, %f276, %f276;
|
|
fma.rn.f32 %f280, %f274, %f274, %f279;
|
|
fma.rn.f32 %f281, %f278, %f278, %f280;
|
|
sqrt.rn.f32 %f26, %f281;
|
|
rcp.rn.f32 %f282, %f26;
|
|
mul.f32 %f27, %f274, %f282;
|
|
mul.f32 %f28, %f276, %f282;
|
|
mul.f32 %f29, %f278, %f282;
|
|
mul.f32 %f283, %f957, %f28;
|
|
fma.rn.f32 %f284, %f956, %f27, %f283;
|
|
fma.rn.f32 %f30, %f958, %f29, %f284;
|
|
setp.leu.f32 %p15, %f30, 0f00000000;
|
|
@%p15 bra BB0_22;
|
|
|
|
setp.ne.s32 %p17, %r1, 0;
|
|
mul.f32 %f285, %f26, %f26;
|
|
mul.f32 %f286, %f285, 0f40C90FDB;
|
|
div.rn.f32 %f287, %f15, %f286;
|
|
add.f32 %f31, %f287, %f287;
|
|
setp.gt.f32 %p18, %f31, %f19;
|
|
and.pred %p19, %p17, %p18;
|
|
mov.pred %p150, -1;
|
|
@%p19 bra BB0_24;
|
|
|
|
ld.global.f32 %f290, [lightInvCutoff];
|
|
mul.f32 %f32, %f26, %f290;
|
|
mov.f32 %f294, 0f40800000;
|
|
abs.f32 %f34, %f32;
|
|
setp.lt.f32 %p20, %f34, 0f00800000;
|
|
mul.f32 %f296, %f34, 0f4B800000;
|
|
selp.f32 %f297, 0fC3170000, 0fC2FE0000, %p20;
|
|
selp.f32 %f298, %f296, %f34, %p20;
|
|
mov.b32 %r136, %f298;
|
|
and.b32 %r137, %r136, 8388607;
|
|
or.b32 %r138, %r137, 1065353216;
|
|
mov.b32 %f299, %r138;
|
|
shr.u32 %r139, %r136, 23;
|
|
cvt.rn.f32.u32 %f300, %r139;
|
|
add.f32 %f301, %f297, %f300;
|
|
setp.gt.f32 %p21, %f299, 0f3FB504F3;
|
|
mul.f32 %f302, %f299, 0f3F000000;
|
|
add.f32 %f303, %f301, 0f3F800000;
|
|
selp.f32 %f304, %f302, %f299, %p21;
|
|
selp.f32 %f305, %f303, %f301, %p21;
|
|
add.f32 %f306, %f304, 0fBF800000;
|
|
add.f32 %f289, %f304, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f288,%f289;
|
|
// inline asm
|
|
add.f32 %f307, %f306, %f306;
|
|
mul.f32 %f308, %f288, %f307;
|
|
mul.f32 %f309, %f308, %f308;
|
|
mov.f32 %f310, 0f3C4CAF63;
|
|
mov.f32 %f311, 0f3B18F0FE;
|
|
fma.rn.f32 %f312, %f311, %f309, %f310;
|
|
mov.f32 %f313, 0f3DAAAABD;
|
|
fma.rn.f32 %f314, %f312, %f309, %f313;
|
|
mul.rn.f32 %f315, %f314, %f309;
|
|
mul.rn.f32 %f316, %f315, %f308;
|
|
sub.f32 %f317, %f306, %f308;
|
|
neg.f32 %f318, %f308;
|
|
add.f32 %f319, %f317, %f317;
|
|
fma.rn.f32 %f320, %f318, %f306, %f319;
|
|
mul.rn.f32 %f321, %f288, %f320;
|
|
add.f32 %f322, %f316, %f308;
|
|
sub.f32 %f323, %f308, %f322;
|
|
add.f32 %f324, %f316, %f323;
|
|
add.f32 %f325, %f321, %f324;
|
|
add.f32 %f326, %f322, %f325;
|
|
sub.f32 %f327, %f322, %f326;
|
|
add.f32 %f328, %f325, %f327;
|
|
mov.f32 %f329, 0f3F317200;
|
|
mul.rn.f32 %f330, %f305, %f329;
|
|
mov.f32 %f331, 0f35BFBE8E;
|
|
mul.rn.f32 %f332, %f305, %f331;
|
|
add.f32 %f333, %f330, %f326;
|
|
sub.f32 %f334, %f330, %f333;
|
|
add.f32 %f335, %f326, %f334;
|
|
add.f32 %f336, %f328, %f335;
|
|
add.f32 %f337, %f332, %f336;
|
|
add.f32 %f338, %f333, %f337;
|
|
sub.f32 %f339, %f333, %f338;
|
|
add.f32 %f340, %f337, %f339;
|
|
mul.rn.f32 %f341, %f294, %f338;
|
|
neg.f32 %f342, %f341;
|
|
fma.rn.f32 %f343, %f294, %f338, %f342;
|
|
fma.rn.f32 %f344, %f294, %f340, %f343;
|
|
mov.f32 %f345, 0f00000000;
|
|
fma.rn.f32 %f346, %f345, %f338, %f344;
|
|
add.rn.f32 %f347, %f341, %f346;
|
|
neg.f32 %f348, %f347;
|
|
add.rn.f32 %f349, %f341, %f348;
|
|
add.rn.f32 %f350, %f349, %f346;
|
|
mov.b32 %r140, %f347;
|
|
setp.eq.s32 %p22, %r140, 1118925336;
|
|
add.s32 %r141, %r140, -1;
|
|
mov.b32 %f351, %r141;
|
|
add.f32 %f352, %f350, 0f37000000;
|
|
selp.f32 %f353, %f351, %f347, %p22;
|
|
selp.f32 %f35, %f352, %f350, %p22;
|
|
mul.f32 %f354, %f353, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f355, %f354;
|
|
mov.f32 %f356, 0fBF317200;
|
|
fma.rn.f32 %f357, %f355, %f356, %f353;
|
|
mov.f32 %f358, 0fB5BFBE8E;
|
|
fma.rn.f32 %f359, %f355, %f358, %f357;
|
|
mul.f32 %f360, %f359, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f361, %f360;
|
|
add.f32 %f362, %f355, 0f00000000;
|
|
ex2.approx.f32 %f363, %f362;
|
|
mul.f32 %f364, %f361, %f363;
|
|
setp.lt.f32 %p23, %f353, 0fC2D20000;
|
|
selp.f32 %f365, 0f00000000, %f364, %p23;
|
|
setp.gt.f32 %p24, %f353, 0f42D20000;
|
|
selp.f32 %f965, 0f7F800000, %f365, %p24;
|
|
setp.eq.f32 %p25, %f965, 0f7F800000;
|
|
@%p25 bra BB0_10;
|
|
|
|
fma.rn.f32 %f965, %f965, %f35, %f965;
|
|
|
|
BB0_10:
|
|
mov.f32 %f952, 0f40000000;
|
|
cvt.rzi.f32.f32 %f951, %f952;
|
|
add.f32 %f950, %f951, %f951;
|
|
mov.f32 %f949, 0f40800000;
|
|
sub.f32 %f948, %f949, %f950;
|
|
abs.f32 %f947, %f948;
|
|
setp.lt.f32 %p26, %f32, 0f00000000;
|
|
setp.eq.f32 %p27, %f947, 0f3F800000;
|
|
and.pred %p1, %p26, %p27;
|
|
mov.b32 %r142, %f965;
|
|
xor.b32 %r143, %r142, -2147483648;
|
|
mov.b32 %f366, %r143;
|
|
selp.f32 %f967, %f366, %f965, %p1;
|
|
setp.eq.f32 %p28, %f32, 0f00000000;
|
|
@%p28 bra BB0_13;
|
|
bra.uni BB0_11;
|
|
|
|
BB0_13:
|
|
add.f32 %f369, %f32, %f32;
|
|
selp.f32 %f967, %f369, 0f00000000, %p27;
|
|
bra.uni BB0_14;
|
|
|
|
BB0_11:
|
|
setp.geu.f32 %p29, %f32, 0f00000000;
|
|
@%p29 bra BB0_14;
|
|
|
|
mov.f32 %f955, 0f40800000;
|
|
cvt.rzi.f32.f32 %f368, %f955;
|
|
setp.neu.f32 %p30, %f368, 0f40800000;
|
|
selp.f32 %f967, 0f7FFFFFFF, %f967, %p30;
|
|
|
|
BB0_14:
|
|
add.f32 %f370, %f34, 0f40800000;
|
|
mov.b32 %r144, %f370;
|
|
setp.lt.s32 %p32, %r144, 2139095040;
|
|
@%p32 bra BB0_19;
|
|
|
|
setp.gtu.f32 %p33, %f34, 0f7F800000;
|
|
@%p33 bra BB0_18;
|
|
bra.uni BB0_16;
|
|
|
|
BB0_18:
|
|
add.f32 %f967, %f32, 0f40800000;
|
|
bra.uni BB0_19;
|
|
|
|
BB0_16:
|
|
setp.neu.f32 %p34, %f34, 0f7F800000;
|
|
@%p34 bra BB0_19;
|
|
|
|
selp.f32 %f967, 0fFF800000, 0f7F800000, %p1;
|
|
|
|
BB0_19:
|
|
mov.u32 %r380, 1;
|
|
mov.u64 %rd302, lightMeshBuffer;
|
|
cvta.global.u64 %rd301, %rd302;
|
|
shl.b32 %r379, %r383, 1;
|
|
mov.f32 %f371, 0f3F800000;
|
|
sub.f32 %f372, %f371, %f967;
|
|
setp.eq.f32 %p35, %f32, 0f3F800000;
|
|
selp.f32 %f373, 0f00000000, %f372, %p35;
|
|
cvt.sat.f32.f32 %f374, %f373;
|
|
mul.f32 %f375, %f31, %f374;
|
|
add.s32 %r147, %r379, 1;
|
|
cvt.s64.s32 %rd51, %r147;
|
|
// inline asm
|
|
call (%rd49), _rt_buffer_get_64, (%rd301, %r380, %r117, %rd51, %rd26, %rd26, %rd26);
|
|
// inline asm
|
|
ld.f32 %f376, [%rd49];
|
|
mul.f32 %f377, %f27, %f376;
|
|
ld.f32 %f378, [%rd49+4];
|
|
mul.f32 %f379, %f28, %f378;
|
|
neg.f32 %f380, %f379;
|
|
sub.f32 %f381, %f380, %f377;
|
|
ld.f32 %f382, [%rd49+8];
|
|
mul.f32 %f383, %f29, %f382;
|
|
sub.f32 %f384, %f381, %f383;
|
|
cvt.sat.f32.f32 %f385, %f384;
|
|
mul.f32 %f46, %f375, %f385;
|
|
setp.leu.f32 %p36, %f46, 0f3727C5AC;
|
|
@%p36 bra BB0_21;
|
|
|
|
mov.u32 %r382, 1;
|
|
add.u64 %rd56, %SP, 4;
|
|
cvta.to.local.u64 %rd57, %rd56;
|
|
max.f32 %f392, %f398, %f261;
|
|
sub.f32 %f393, %f26, %f392;
|
|
mov.u32 %r151, 1065353216;
|
|
st.local.u32 [%rd57], %r151;
|
|
ld.global.u32 %r148, [root];
|
|
// inline asm
|
|
call _rt_trace_64, (%r148, %f12, %f13, %f14, %f27, %f28, %f29, %r382, %f392, %f393, %rd56, %r102);
|
|
// inline asm
|
|
ld.local.f32 %f400, [%rd57];
|
|
mul.f32 %f401, %f46, %f400;
|
|
cvt.sat.f32.f32 %f402, %f30;
|
|
fma.rn.f32 %f25, %f402, %f401, %f25;
|
|
mul.f32 %f403, %f30, 0f40800000;
|
|
cvt.sat.f32.f32 %f404, %f403;
|
|
mul.f32 %f405, %f401, %f404;
|
|
fma.rn.f32 %f23, %f27, %f405, %f23;
|
|
fma.rn.f32 %f22, %f28, %f405, %f22;
|
|
fma.rn.f32 %f21, %f29, %f405, %f21;
|
|
add.f32 %f24, %f24, %f405;
|
|
add.f32 %f20, %f20, %f400;
|
|
|
|
BB0_21:
|
|
add.s32 %r7, %r7, 1;
|
|
|
|
BB0_22:
|
|
ld.global.u32 %r152, [lightMeshBufferSize];
|
|
add.s32 %r383, %r383, 1;
|
|
setp.lt.u32 %p38, %r383, %r152;
|
|
@%p38 bra BB0_6;
|
|
bra.uni BB0_23;
|
|
|
|
BB0_4:
|
|
mov.f32 %f21, %f20;
|
|
mov.f32 %f22, %f20;
|
|
mov.f32 %f23, %f20;
|
|
mov.f32 %f24, %f20;
|
|
mov.f32 %f25, %f20;
|
|
|
|
BB0_23:
|
|
mov.pred %p150, %p13;
|
|
|
|
BB0_24:
|
|
cvt.rn.f32.s32 %f406, %r7;
|
|
mov.f32 %f407, 0f3F800000;
|
|
max.f32 %f408, %f406, %f407;
|
|
div.rn.f32 %f1016, %f25, %f408;
|
|
div.rn.f32 %f1021, %f20, %f408;
|
|
div.rn.f32 %f1017, %f24, %f408;
|
|
div.rn.f32 %f1018, %f23, %f408;
|
|
div.rn.f32 %f1019, %f22, %f408;
|
|
div.rn.f32 %f1020, %f21, %f408;
|
|
@!%p150 bra BB0_77;
|
|
bra.uni BB0_25;
|
|
|
|
BB0_25:
|
|
abs.f32 %f954, %f958;
|
|
abs.f32 %f953, %f956;
|
|
setp.gt.f32 %p39, %f953, %f954;
|
|
neg.f32 %f415, %f957;
|
|
selp.f32 %f416, %f415, 0f00000000, %p39;
|
|
neg.f32 %f417, %f958;
|
|
selp.f32 %f418, %f956, %f417, %p39;
|
|
selp.f32 %f419, 0f00000000, %f957, %p39;
|
|
mul.f32 %f420, %f418, %f418;
|
|
fma.rn.f32 %f421, %f416, %f416, %f420;
|
|
fma.rn.f32 %f422, %f419, %f419, %f421;
|
|
sqrt.rn.f32 %f423, %f422;
|
|
rcp.rn.f32 %f424, %f423;
|
|
mul.f32 %f77, %f416, %f424;
|
|
mul.f32 %f78, %f418, %f424;
|
|
mul.f32 %f79, %f419, %f424;
|
|
mov.f32 %f992, 0f00000000;
|
|
setp.lt.s32 %p40, %r1, 1;
|
|
mov.f32 %f993, %f992;
|
|
mov.f32 %f994, %f992;
|
|
mov.f32 %f995, %f992;
|
|
mov.f32 %f996, %f992;
|
|
mov.f32 %f997, %f992;
|
|
@%p40 bra BB0_76;
|
|
|
|
mad.lo.s32 %r390, %r129, 1664525, 1013904223;
|
|
cvt.rn.f32.s32 %f431, %r1;
|
|
rcp.rn.f32 %f80, %f431;
|
|
add.u64 %rd58, %SP, 8;
|
|
cvta.to.local.u64 %rd2, %rd58;
|
|
mul.f32 %f81, %f12, 0f3456BF95;
|
|
mul.f32 %f82, %f13, 0f3456BF95;
|
|
mul.f32 %f83, %f14, 0f3456BF95;
|
|
add.u64 %rd59, %SP, 0;
|
|
cvta.to.local.u64 %rd3, %rd59;
|
|
mul.f32 %f432, %f956, %f78;
|
|
mul.f32 %f433, %f957, %f77;
|
|
sub.f32 %f84, %f433, %f432;
|
|
mul.f32 %f434, %f958, %f77;
|
|
mul.f32 %f435, %f956, %f79;
|
|
sub.f32 %f85, %f435, %f434;
|
|
mul.f32 %f436, %f957, %f79;
|
|
mul.f32 %f437, %f958, %f78;
|
|
sub.f32 %f86, %f437, %f436;
|
|
mov.f32 %f992, 0f00000000;
|
|
mov.u32 %r153, 0;
|
|
abs.f32 %f438, %f82;
|
|
abs.f32 %f439, %f81;
|
|
max.f32 %f440, %f439, %f438;
|
|
abs.f32 %f441, %f83;
|
|
max.f32 %f442, %f440, %f441;
|
|
mov.u32 %r387, %r153;
|
|
mov.f32 %f993, %f992;
|
|
mov.f32 %f994, %f992;
|
|
mov.f32 %f995, %f992;
|
|
mov.f32 %f996, %f992;
|
|
mov.f32 %f997, %f992;
|
|
|
|
BB0_27:
|
|
cvt.rn.f32.s32 %f93, %r387;
|
|
max.f32 %f94, %f442, %f261;
|
|
mov.u32 %r389, %r153;
|
|
|
|
BB0_28:
|
|
mad.lo.s32 %r155, %r390, 1664525, 1013904223;
|
|
and.b32 %r156, %r155, 16777215;
|
|
cvt.rn.f32.u32 %f444, %r156;
|
|
fma.rn.f32 %f445, %f444, 0f33800000, %f93;
|
|
mul.f32 %f101, %f80, %f445;
|
|
mad.lo.s32 %r390, %r155, 1664525, 1013904223;
|
|
and.b32 %r157, %r390, 16777215;
|
|
cvt.rn.f32.u32 %f446, %r157;
|
|
cvt.rn.f32.s32 %f447, %r389;
|
|
fma.rn.f32 %f448, %f446, 0f33800000, %f447;
|
|
mul.f32 %f449, %f80, %f448;
|
|
mul.f32 %f450, %f101, %f101;
|
|
sub.f32 %f452, %f407, %f450;
|
|
mov.f32 %f453, 0f00000000;
|
|
max.f32 %f454, %f453, %f452;
|
|
sqrt.rn.f32 %f102, %f454;
|
|
mul.f32 %f1004, %f449, 0f40C90FDB;
|
|
abs.f32 %f104, %f1004;
|
|
setp.neu.f32 %p41, %f104, 0f7F800000;
|
|
mov.f32 %f998, %f1004;
|
|
@%p41 bra BB0_30;
|
|
|
|
mul.rn.f32 %f998, %f1004, %f453;
|
|
|
|
BB0_30:
|
|
mul.f32 %f456, %f998, 0f3F22F983;
|
|
cvt.rni.s32.f32 %r400, %f456;
|
|
cvt.rn.f32.s32 %f457, %r400;
|
|
neg.f32 %f458, %f457;
|
|
mov.f32 %f459, 0f3FC90FDA;
|
|
fma.rn.f32 %f460, %f458, %f459, %f998;
|
|
mov.f32 %f461, 0f33A22168;
|
|
fma.rn.f32 %f462, %f458, %f461, %f460;
|
|
mov.f32 %f463, 0f27C234C5;
|
|
fma.rn.f32 %f999, %f458, %f463, %f462;
|
|
abs.f32 %f464, %f998;
|
|
setp.leu.f32 %p42, %f464, 0f47CE4780;
|
|
@%p42 bra BB0_41;
|
|
|
|
mov.b32 %r19, %f998;
|
|
shr.u32 %r20, %r19, 23;
|
|
shl.b32 %r160, %r19, 8;
|
|
or.b32 %r21, %r160, -2147483648;
|
|
mov.u32 %r391, 0;
|
|
mov.u64 %rd304, 0;
|
|
mov.u64 %rd303, %rd2;
|
|
mov.u32 %r392, %r391;
|
|
|
|
BB0_32:
|
|
.pragma "nounroll";
|
|
shl.b64 %rd61, %rd304, 2;
|
|
mov.u64 %rd62, __cudart_i2opi_f;
|
|
add.s64 %rd63, %rd62, %rd61;
|
|
ld.const.u32 %r163, [%rd63];
|
|
// inline asm
|
|
{
|
|
mad.lo.cc.u32 %r161, %r163, %r21, %r392;
|
|
madc.hi.u32 %r392, %r163, %r21, 0;
|
|
}
|
|
// inline asm
|
|
st.local.u32 [%rd303], %r161;
|
|
add.s32 %r391, %r391, 1;
|
|
cvt.s64.s32 %rd304, %r391;
|
|
mul.wide.s32 %rd64, %r391, 4;
|
|
add.s64 %rd303, %rd2, %rd64;
|
|
setp.ne.s32 %p43, %r391, 6;
|
|
@%p43 bra BB0_32;
|
|
|
|
and.b32 %r166, %r20, 255;
|
|
add.s32 %r167, %r166, -128;
|
|
shr.u32 %r168, %r167, 5;
|
|
and.b32 %r26, %r19, -2147483648;
|
|
st.local.u32 [%rd2+24], %r392;
|
|
mov.u32 %r169, 6;
|
|
sub.s32 %r170, %r169, %r168;
|
|
mul.wide.s32 %rd65, %r170, 4;
|
|
add.s64 %rd9, %rd2, %rd65;
|
|
ld.local.u32 %r393, [%rd9];
|
|
ld.local.u32 %r394, [%rd9+-4];
|
|
and.b32 %r29, %r20, 31;
|
|
setp.eq.s32 %p44, %r29, 0;
|
|
@%p44 bra BB0_35;
|
|
|
|
mov.u32 %r171, 32;
|
|
sub.s32 %r172, %r171, %r29;
|
|
shr.u32 %r173, %r394, %r172;
|
|
shl.b32 %r174, %r393, %r29;
|
|
add.s32 %r393, %r173, %r174;
|
|
ld.local.u32 %r175, [%rd9+-8];
|
|
shr.u32 %r176, %r175, %r172;
|
|
shl.b32 %r177, %r394, %r29;
|
|
add.s32 %r394, %r176, %r177;
|
|
|
|
BB0_35:
|
|
shr.u32 %r178, %r394, 30;
|
|
shl.b32 %r179, %r393, 2;
|
|
add.s32 %r395, %r178, %r179;
|
|
shl.b32 %r35, %r394, 2;
|
|
shr.u32 %r180, %r395, 31;
|
|
shr.u32 %r181, %r393, 30;
|
|
add.s32 %r36, %r180, %r181;
|
|
setp.eq.s32 %p45, %r180, 0;
|
|
@%p45 bra BB0_36;
|
|
bra.uni BB0_37;
|
|
|
|
BB0_36:
|
|
mov.u32 %r396, %r26;
|
|
mov.u32 %r397, %r35;
|
|
bra.uni BB0_38;
|
|
|
|
BB0_37:
|
|
not.b32 %r182, %r395;
|
|
neg.s32 %r397, %r35;
|
|
setp.eq.s32 %p46, %r35, 0;
|
|
selp.u32 %r183, 1, 0, %p46;
|
|
add.s32 %r395, %r183, %r182;
|
|
xor.b32 %r396, %r26, -2147483648;
|
|
|
|
BB0_38:
|
|
clz.b32 %r399, %r395;
|
|
setp.eq.s32 %p47, %r399, 0;
|
|
shl.b32 %r184, %r395, %r399;
|
|
mov.u32 %r185, 32;
|
|
sub.s32 %r186, %r185, %r399;
|
|
shr.u32 %r187, %r397, %r186;
|
|
add.s32 %r188, %r187, %r184;
|
|
selp.b32 %r44, %r395, %r188, %p47;
|
|
mov.u32 %r189, -921707870;
|
|
mul.hi.u32 %r398, %r44, %r189;
|
|
setp.eq.s32 %p48, %r26, 0;
|
|
neg.s32 %r190, %r36;
|
|
selp.b32 %r400, %r36, %r190, %p48;
|
|
setp.lt.s32 %p49, %r398, 1;
|
|
@%p49 bra BB0_40;
|
|
|
|
mul.lo.s32 %r191, %r44, -921707870;
|
|
shr.u32 %r192, %r191, 31;
|
|
shl.b32 %r193, %r398, 1;
|
|
add.s32 %r398, %r192, %r193;
|
|
add.s32 %r399, %r399, 1;
|
|
|
|
BB0_40:
|
|
mov.u32 %r194, 126;
|
|
sub.s32 %r195, %r194, %r399;
|
|
shl.b32 %r196, %r195, 23;
|
|
add.s32 %r197, %r398, 1;
|
|
shr.u32 %r198, %r197, 7;
|
|
add.s32 %r199, %r198, 1;
|
|
shr.u32 %r200, %r199, 1;
|
|
add.s32 %r201, %r200, %r196;
|
|
or.b32 %r202, %r201, %r396;
|
|
mov.b32 %f999, %r202;
|
|
|
|
BB0_41:
|
|
mul.rn.f32 %f110, %f999, %f999;
|
|
add.s32 %r52, %r400, 1;
|
|
and.b32 %r53, %r52, 1;
|
|
setp.eq.s32 %p50, %r53, 0;
|
|
@%p50 bra BB0_43;
|
|
bra.uni BB0_42;
|
|
|
|
BB0_43:
|
|
mov.f32 %f467, 0f3C08839E;
|
|
mov.f32 %f468, 0fB94CA1F9;
|
|
fma.rn.f32 %f1000, %f468, %f110, %f467;
|
|
bra.uni BB0_44;
|
|
|
|
BB0_42:
|
|
mov.f32 %f465, 0fBAB6061A;
|
|
mov.f32 %f466, 0f37CCF5CE;
|
|
fma.rn.f32 %f1000, %f466, %f110, %f465;
|
|
|
|
BB0_44:
|
|
@%p50 bra BB0_46;
|
|
bra.uni BB0_45;
|
|
|
|
BB0_46:
|
|
mov.f32 %f472, 0fBE2AAAA3;
|
|
fma.rn.f32 %f473, %f1000, %f110, %f472;
|
|
fma.rn.f32 %f1001, %f473, %f110, %f453;
|
|
bra.uni BB0_47;
|
|
|
|
BB0_45:
|
|
mov.f32 %f469, 0f3D2AAAA5;
|
|
fma.rn.f32 %f470, %f1000, %f110, %f469;
|
|
mov.f32 %f471, 0fBF000000;
|
|
fma.rn.f32 %f1001, %f470, %f110, %f471;
|
|
|
|
BB0_47:
|
|
fma.rn.f32 %f1002, %f1001, %f999, %f999;
|
|
@%p50 bra BB0_49;
|
|
|
|
fma.rn.f32 %f1002, %f1001, %f110, %f407;
|
|
|
|
BB0_49:
|
|
and.b32 %r203, %r52, 2;
|
|
setp.eq.s32 %p53, %r203, 0;
|
|
@%p53 bra BB0_51;
|
|
|
|
mov.f32 %f477, 0fBF800000;
|
|
fma.rn.f32 %f1002, %f1002, %f477, %f453;
|
|
|
|
BB0_51:
|
|
@%p41 bra BB0_53;
|
|
|
|
mul.rn.f32 %f1004, %f1004, %f453;
|
|
|
|
BB0_53:
|
|
mul.f32 %f479, %f1004, 0f3F22F983;
|
|
cvt.rni.s32.f32 %r410, %f479;
|
|
cvt.rn.f32.s32 %f480, %r410;
|
|
neg.f32 %f481, %f480;
|
|
fma.rn.f32 %f483, %f481, %f459, %f1004;
|
|
fma.rn.f32 %f485, %f481, %f461, %f483;
|
|
fma.rn.f32 %f1005, %f481, %f463, %f485;
|
|
abs.f32 %f487, %f1004;
|
|
setp.leu.f32 %p55, %f487, 0f47CE4780;
|
|
@%p55 bra BB0_64;
|
|
|
|
mov.b32 %r55, %f1004;
|
|
shr.u32 %r56, %r55, 23;
|
|
shl.b32 %r206, %r55, 8;
|
|
or.b32 %r57, %r206, -2147483648;
|
|
mov.u32 %r401, 0;
|
|
mov.u64 %rd305, %rd2;
|
|
mov.u64 %rd306, %rd26;
|
|
mov.u32 %r402, %r401;
|
|
|
|
BB0_55:
|
|
.pragma "nounroll";
|
|
shl.b64 %rd67, %rd306, 2;
|
|
mov.u64 %rd68, __cudart_i2opi_f;
|
|
add.s64 %rd69, %rd68, %rd67;
|
|
ld.const.u32 %r209, [%rd69];
|
|
// inline asm
|
|
{
|
|
mad.lo.cc.u32 %r207, %r209, %r57, %r402;
|
|
madc.hi.u32 %r402, %r209, %r57, 0;
|
|
}
|
|
// inline asm
|
|
st.local.u32 [%rd305], %r207;
|
|
add.s32 %r401, %r401, 1;
|
|
cvt.s64.s32 %rd306, %r401;
|
|
mul.wide.s32 %rd70, %r401, 4;
|
|
add.s64 %rd305, %rd2, %rd70;
|
|
setp.ne.s32 %p56, %r401, 6;
|
|
@%p56 bra BB0_55;
|
|
|
|
and.b32 %r212, %r56, 255;
|
|
add.s32 %r213, %r212, -128;
|
|
shr.u32 %r214, %r213, 5;
|
|
and.b32 %r62, %r55, -2147483648;
|
|
st.local.u32 [%rd2+24], %r402;
|
|
mov.u32 %r215, 6;
|
|
sub.s32 %r216, %r215, %r214;
|
|
mul.wide.s32 %rd71, %r216, 4;
|
|
add.s64 %rd15, %rd2, %rd71;
|
|
ld.local.u32 %r403, [%rd15];
|
|
ld.local.u32 %r404, [%rd15+-4];
|
|
and.b32 %r65, %r56, 31;
|
|
setp.eq.s32 %p57, %r65, 0;
|
|
@%p57 bra BB0_58;
|
|
|
|
mov.u32 %r217, 32;
|
|
sub.s32 %r218, %r217, %r65;
|
|
shr.u32 %r219, %r404, %r218;
|
|
shl.b32 %r220, %r403, %r65;
|
|
add.s32 %r403, %r219, %r220;
|
|
ld.local.u32 %r221, [%rd15+-8];
|
|
shr.u32 %r222, %r221, %r218;
|
|
shl.b32 %r223, %r404, %r65;
|
|
add.s32 %r404, %r222, %r223;
|
|
|
|
BB0_58:
|
|
shr.u32 %r224, %r404, 30;
|
|
shl.b32 %r225, %r403, 2;
|
|
add.s32 %r405, %r224, %r225;
|
|
shl.b32 %r71, %r404, 2;
|
|
shr.u32 %r226, %r405, 31;
|
|
shr.u32 %r227, %r403, 30;
|
|
add.s32 %r72, %r226, %r227;
|
|
setp.eq.s32 %p58, %r226, 0;
|
|
@%p58 bra BB0_59;
|
|
bra.uni BB0_60;
|
|
|
|
BB0_59:
|
|
mov.u32 %r406, %r62;
|
|
mov.u32 %r407, %r71;
|
|
bra.uni BB0_61;
|
|
|
|
BB0_60:
|
|
not.b32 %r228, %r405;
|
|
neg.s32 %r407, %r71;
|
|
setp.eq.s32 %p59, %r71, 0;
|
|
selp.u32 %r229, 1, 0, %p59;
|
|
add.s32 %r405, %r229, %r228;
|
|
xor.b32 %r406, %r62, -2147483648;
|
|
|
|
BB0_61:
|
|
clz.b32 %r409, %r405;
|
|
setp.eq.s32 %p60, %r409, 0;
|
|
shl.b32 %r230, %r405, %r409;
|
|
mov.u32 %r231, 32;
|
|
sub.s32 %r232, %r231, %r409;
|
|
shr.u32 %r233, %r407, %r232;
|
|
add.s32 %r234, %r233, %r230;
|
|
selp.b32 %r80, %r405, %r234, %p60;
|
|
mov.u32 %r235, -921707870;
|
|
mul.hi.u32 %r408, %r80, %r235;
|
|
setp.eq.s32 %p61, %r62, 0;
|
|
neg.s32 %r236, %r72;
|
|
selp.b32 %r410, %r72, %r236, %p61;
|
|
setp.lt.s32 %p62, %r408, 1;
|
|
@%p62 bra BB0_63;
|
|
|
|
mul.lo.s32 %r237, %r80, -921707870;
|
|
shr.u32 %r238, %r237, 31;
|
|
shl.b32 %r239, %r408, 1;
|
|
add.s32 %r408, %r238, %r239;
|
|
add.s32 %r409, %r409, 1;
|
|
|
|
BB0_63:
|
|
mov.u32 %r240, 126;
|
|
sub.s32 %r241, %r240, %r409;
|
|
shl.b32 %r242, %r241, 23;
|
|
add.s32 %r243, %r408, 1;
|
|
shr.u32 %r244, %r243, 7;
|
|
add.s32 %r245, %r244, 1;
|
|
shr.u32 %r246, %r245, 1;
|
|
add.s32 %r247, %r246, %r242;
|
|
or.b32 %r248, %r247, %r406;
|
|
mov.b32 %f1005, %r248;
|
|
|
|
BB0_64:
|
|
mul.rn.f32 %f127, %f1005, %f1005;
|
|
and.b32 %r88, %r410, 1;
|
|
setp.eq.s32 %p63, %r88, 0;
|
|
@%p63 bra BB0_66;
|
|
bra.uni BB0_65;
|
|
|
|
BB0_66:
|
|
mov.f32 %f490, 0f3C08839E;
|
|
mov.f32 %f491, 0fB94CA1F9;
|
|
fma.rn.f32 %f1006, %f491, %f127, %f490;
|
|
bra.uni BB0_67;
|
|
|
|
BB0_65:
|
|
mov.f32 %f488, 0fBAB6061A;
|
|
mov.f32 %f489, 0f37CCF5CE;
|
|
fma.rn.f32 %f1006, %f489, %f127, %f488;
|
|
|
|
BB0_67:
|
|
@%p63 bra BB0_69;
|
|
bra.uni BB0_68;
|
|
|
|
BB0_69:
|
|
mov.f32 %f495, 0fBE2AAAA3;
|
|
fma.rn.f32 %f496, %f1006, %f127, %f495;
|
|
fma.rn.f32 %f1007, %f496, %f127, %f453;
|
|
bra.uni BB0_70;
|
|
|
|
BB0_68:
|
|
mov.f32 %f492, 0f3D2AAAA5;
|
|
fma.rn.f32 %f493, %f1006, %f127, %f492;
|
|
mov.f32 %f494, 0fBF000000;
|
|
fma.rn.f32 %f1007, %f493, %f127, %f494;
|
|
|
|
BB0_70:
|
|
fma.rn.f32 %f1008, %f1007, %f1005, %f1005;
|
|
@%p63 bra BB0_72;
|
|
|
|
fma.rn.f32 %f1008, %f1007, %f127, %f407;
|
|
|
|
BB0_72:
|
|
and.b32 %r249, %r410, 2;
|
|
setp.eq.s32 %p66, %r249, 0;
|
|
@%p66 bra BB0_74;
|
|
|
|
mov.f32 %f500, 0fBF800000;
|
|
fma.rn.f32 %f1008, %f1008, %f500, %f453;
|
|
|
|
BB0_74:
|
|
mul.f32 %f509, %f102, %f1002;
|
|
mul.f32 %f510, %f102, %f1008;
|
|
mul.f32 %f511, %f77, %f510;
|
|
mul.f32 %f512, %f78, %f510;
|
|
mul.f32 %f513, %f79, %f510;
|
|
fma.rn.f32 %f514, %f86, %f509, %f511;
|
|
fma.rn.f32 %f515, %f85, %f509, %f512;
|
|
fma.rn.f32 %f516, %f84, %f509, %f513;
|
|
fma.rn.f32 %f504, %f956, %f101, %f514;
|
|
fma.rn.f32 %f505, %f957, %f101, %f515;
|
|
fma.rn.f32 %f506, %f958, %f101, %f516;
|
|
mov.u32 %r251, 0;
|
|
st.local.u32 [%rd3], %r251;
|
|
ld.global.u32 %r250, [root];
|
|
mov.f32 %f508, 0f6C4ECB8F;
|
|
// inline asm
|
|
call _rt_trace_64, (%r250, %f12, %f13, %f14, %f504, %f505, %f506, %r251, %f94, %f508, %rd59, %r102);
|
|
// inline asm
|
|
ld.local.f32 %f517, [%rd3];
|
|
setp.lt.f32 %p67, %f517, 0f00000000;
|
|
selp.f32 %f518, 0f00000000, %f517, %p67;
|
|
selp.f32 %f519, 0f00000000, 0f3F800000, %p67;
|
|
add.f32 %f992, %f992, %f519;
|
|
fma.rn.f32 %f995, %f504, %f518, %f995;
|
|
fma.rn.f32 %f994, %f505, %f518, %f994;
|
|
fma.rn.f32 %f993, %f506, %f518, %f993;
|
|
add.f32 %f996, %f996, %f518;
|
|
mul.f32 %f520, %f957, %f505;
|
|
fma.rn.f32 %f521, %f956, %f504, %f520;
|
|
fma.rn.f32 %f522, %f958, %f506, %f521;
|
|
cvt.sat.f32.f32 %f523, %f522;
|
|
fma.rn.f32 %f997, %f518, %f523, %f997;
|
|
add.s32 %r389, %r389, 1;
|
|
setp.lt.s32 %p68, %r389, %r1;
|
|
@%p68 bra BB0_28;
|
|
|
|
add.s32 %r387, %r387, 1;
|
|
setp.lt.s32 %p69, %r387, %r1;
|
|
@%p69 bra BB0_27;
|
|
|
|
BB0_76:
|
|
mul.lo.s32 %r253, %r1, %r1;
|
|
cvt.rn.f32.s32 %f524, %r253;
|
|
div.rn.f32 %f525, %f997, %f524;
|
|
div.rn.f32 %f1021, %f992, %f524;
|
|
div.rn.f32 %f1017, %f996, %f524;
|
|
div.rn.f32 %f1018, %f995, %f524;
|
|
div.rn.f32 %f1019, %f994, %f524;
|
|
div.rn.f32 %f1020, %f993, %f524;
|
|
add.f32 %f1016, %f525, %f525;
|
|
|
|
BB0_77:
|
|
ld.global.u32 %r412, [imageEnabled];
|
|
and.b32 %r254, %r412, 8;
|
|
setp.eq.s32 %p70, %r254, 0;
|
|
@%p70 bra BB0_90;
|
|
|
|
cvt.u64.u32 %rd75, %r3;
|
|
cvt.u64.u32 %rd76, %r4;
|
|
mov.u64 %rd79, image_Mask;
|
|
cvta.global.u64 %rd74, %rd79;
|
|
// inline asm
|
|
call (%rd73), _rt_buffer_get_64, (%rd74, %r101, %r101, %rd75, %rd76, %rd26, %rd26);
|
|
// inline asm
|
|
mov.f32 %f528, 0f3E68BA2E;
|
|
cvt.rzi.f32.f32 %f529, %f528;
|
|
fma.rn.f32 %f530, %f529, 0fC0000000, 0f3EE8BA2E;
|
|
abs.f32 %f163, %f530;
|
|
abs.f32 %f164, %f1021;
|
|
setp.lt.f32 %p71, %f164, 0f00800000;
|
|
mul.f32 %f531, %f164, 0f4B800000;
|
|
selp.f32 %f532, 0fC3170000, 0fC2FE0000, %p71;
|
|
selp.f32 %f533, %f531, %f164, %p71;
|
|
mov.b32 %r257, %f533;
|
|
and.b32 %r258, %r257, 8388607;
|
|
or.b32 %r259, %r258, 1065353216;
|
|
mov.b32 %f534, %r259;
|
|
shr.u32 %r260, %r257, 23;
|
|
cvt.rn.f32.u32 %f535, %r260;
|
|
add.f32 %f536, %f532, %f535;
|
|
setp.gt.f32 %p72, %f534, 0f3FB504F3;
|
|
mul.f32 %f537, %f534, 0f3F000000;
|
|
add.f32 %f538, %f536, 0f3F800000;
|
|
selp.f32 %f539, %f537, %f534, %p72;
|
|
selp.f32 %f540, %f538, %f536, %p72;
|
|
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, 0f3EE8BA2E;
|
|
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 %r261, %f583;
|
|
setp.eq.s32 %p73, %r261, 1118925336;
|
|
add.s32 %r262, %r261, -1;
|
|
mov.b32 %f587, %r262;
|
|
add.f32 %f588, %f586, 0f37000000;
|
|
selp.f32 %f589, %f587, %f583, %p73;
|
|
selp.f32 %f165, %f588, %f586, %p73;
|
|
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 %p74, %f589, 0fC2D20000;
|
|
selp.f32 %f601, 0f00000000, %f600, %p74;
|
|
setp.gt.f32 %p75, %f589, 0f42D20000;
|
|
selp.f32 %f1022, 0f7F800000, %f601, %p75;
|
|
setp.eq.f32 %p76, %f1022, 0f7F800000;
|
|
@%p76 bra BB0_80;
|
|
|
|
fma.rn.f32 %f1022, %f1022, %f165, %f1022;
|
|
|
|
BB0_80:
|
|
setp.lt.f32 %p77, %f1021, 0f00000000;
|
|
setp.eq.f32 %p78, %f163, 0f3F800000;
|
|
and.pred %p3, %p77, %p78;
|
|
mov.b32 %r263, %f1022;
|
|
xor.b32 %r264, %r263, -2147483648;
|
|
mov.b32 %f602, %r264;
|
|
selp.f32 %f1024, %f602, %f1022, %p3;
|
|
setp.eq.f32 %p79, %f1021, 0f00000000;
|
|
@%p79 bra BB0_83;
|
|
bra.uni BB0_81;
|
|
|
|
BB0_83:
|
|
add.f32 %f605, %f1021, %f1021;
|
|
selp.f32 %f1024, %f605, 0f00000000, %p78;
|
|
bra.uni BB0_84;
|
|
|
|
BB0_148:
|
|
mov.u64 %rd220, image_HDR;
|
|
cvta.global.u64 %rd215, %rd220;
|
|
mov.u32 %r350, 8;
|
|
mov.u64 %rd219, 0;
|
|
// inline asm
|
|
call (%rd214), _rt_buffer_get_64, (%rd215, %r101, %r350, %rd19, %rd20, %rd219, %rd219);
|
|
// inline asm
|
|
mov.f32 %f918, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs98, %f918;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs99, 0;
|
|
st.v4.u16 [%rd214], {%rs98, %rs98, %rs98, %rs99};
|
|
|
|
BB0_149:
|
|
ld.global.u32 %r351, [additive];
|
|
setp.eq.s32 %p146, %r351, 0;
|
|
@%p146 bra BB0_151;
|
|
|
|
mov.u64 %rd233, image_RNM0;
|
|
cvta.global.u64 %rd222, %rd233;
|
|
mov.u32 %r355, 8;
|
|
mov.u64 %rd232, 0;
|
|
// inline asm
|
|
call (%rd221), _rt_buffer_get_64, (%rd222, %r101, %r355, %rd19, %rd20, %rd232, %rd232);
|
|
// inline asm
|
|
ld.v4.u16 {%rs106, %rs107, %rs108, %rs109}, [%rd221];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f919, %rs106;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f920, %rs107;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f921, %rs108;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd227), _rt_buffer_get_64, (%rd222, %r101, %r355, %rd19, %rd20, %rd232, %rd232);
|
|
// inline asm
|
|
add.f32 %f922, %f919, 0f00000000;
|
|
add.f32 %f923, %f920, 0f00000000;
|
|
add.f32 %f924, %f921, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs105, %f924;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs104, %f923;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs103, %f922;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs110, 0;
|
|
st.v4.u16 [%rd227], {%rs103, %rs104, %rs105, %rs110};
|
|
bra.uni BB0_152;
|
|
|
|
BB0_151:
|
|
mov.u64 %rd240, image_RNM0;
|
|
cvta.global.u64 %rd235, %rd240;
|
|
mov.u32 %r357, 8;
|
|
mov.u64 %rd239, 0;
|
|
// inline asm
|
|
call (%rd234), _rt_buffer_get_64, (%rd235, %r101, %r357, %rd19, %rd20, %rd239, %rd239);
|
|
// inline asm
|
|
mov.f32 %f925, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs111, %f925;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs112, 0;
|
|
st.v4.u16 [%rd234], {%rs111, %rs111, %rs111, %rs112};
|
|
|
|
BB0_152:
|
|
ld.global.u32 %r358, [additive];
|
|
setp.eq.s32 %p147, %r358, 0;
|
|
@%p147 bra BB0_154;
|
|
|
|
mov.u64 %rd253, image_RNM1;
|
|
cvta.global.u64 %rd242, %rd253;
|
|
mov.u32 %r362, 8;
|
|
mov.u64 %rd252, 0;
|
|
// inline asm
|
|
call (%rd241), _rt_buffer_get_64, (%rd242, %r101, %r362, %rd19, %rd20, %rd252, %rd252);
|
|
// inline asm
|
|
ld.v4.u16 {%rs119, %rs120, %rs121, %rs122}, [%rd241];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f926, %rs119;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f927, %rs120;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f928, %rs121;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd247), _rt_buffer_get_64, (%rd242, %r101, %r362, %rd19, %rd20, %rd252, %rd252);
|
|
// inline asm
|
|
add.f32 %f929, %f926, 0f00000000;
|
|
add.f32 %f930, %f927, 0f00000000;
|
|
add.f32 %f931, %f928, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs118, %f931;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs117, %f930;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs116, %f929;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs123, 0;
|
|
st.v4.u16 [%rd247], {%rs116, %rs117, %rs118, %rs123};
|
|
bra.uni BB0_155;
|
|
|
|
BB0_154:
|
|
mov.u64 %rd260, image_RNM1;
|
|
cvta.global.u64 %rd255, %rd260;
|
|
mov.u32 %r364, 8;
|
|
mov.u64 %rd259, 0;
|
|
// inline asm
|
|
call (%rd254), _rt_buffer_get_64, (%rd255, %r101, %r364, %rd19, %rd20, %rd259, %rd259);
|
|
// inline asm
|
|
mov.f32 %f932, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs124, %f932;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs125, 0;
|
|
st.v4.u16 [%rd254], {%rs124, %rs124, %rs124, %rs125};
|
|
|
|
BB0_155:
|
|
ld.global.u32 %r365, [additive];
|
|
setp.eq.s32 %p148, %r365, 0;
|
|
@%p148 bra BB0_157;
|
|
|
|
mov.u64 %rd273, image_RNM2;
|
|
cvta.global.u64 %rd262, %rd273;
|
|
mov.u32 %r369, 8;
|
|
mov.u64 %rd272, 0;
|
|
// inline asm
|
|
call (%rd261), _rt_buffer_get_64, (%rd262, %r101, %r369, %rd19, %rd20, %rd272, %rd272);
|
|
// inline asm
|
|
ld.v4.u16 {%rs132, %rs133, %rs134, %rs135}, [%rd261];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f933, %rs132;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f934, %rs133;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f935, %rs134;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd267), _rt_buffer_get_64, (%rd262, %r101, %r369, %rd19, %rd20, %rd272, %rd272);
|
|
// inline asm
|
|
add.f32 %f936, %f933, 0f00000000;
|
|
add.f32 %f937, %f934, 0f00000000;
|
|
add.f32 %f938, %f935, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs131, %f938;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs130, %f937;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs129, %f936;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs136, 0;
|
|
st.v4.u16 [%rd267], {%rs129, %rs130, %rs131, %rs136};
|
|
bra.uni BB0_158;
|
|
|
|
BB0_157:
|
|
mov.u64 %rd280, image_RNM2;
|
|
cvta.global.u64 %rd275, %rd280;
|
|
mov.u32 %r371, 8;
|
|
mov.u64 %rd279, 0;
|
|
// inline asm
|
|
call (%rd274), _rt_buffer_get_64, (%rd275, %r101, %r371, %rd19, %rd20, %rd279, %rd279);
|
|
// inline asm
|
|
mov.f32 %f939, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs137, %f939;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs138, 0;
|
|
st.v4.u16 [%rd274], {%rs137, %rs137, %rs137, %rs138};
|
|
|
|
BB0_158:
|
|
ld.global.u32 %r372, [additive];
|
|
setp.eq.s32 %p149, %r372, 0;
|
|
@%p149 bra BB0_160;
|
|
|
|
mov.u64 %rd293, image_RNM3;
|
|
cvta.global.u64 %rd282, %rd293;
|
|
mov.u32 %r376, 8;
|
|
mov.u64 %rd292, 0;
|
|
// inline asm
|
|
call (%rd281), _rt_buffer_get_64, (%rd282, %r101, %r376, %rd19, %rd20, %rd292, %rd292);
|
|
// inline asm
|
|
ld.v4.u16 {%rs145, %rs146, %rs147, %rs148}, [%rd281];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f940, %rs145;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f941, %rs146;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f942, %rs147;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd287), _rt_buffer_get_64, (%rd282, %r101, %r376, %rd19, %rd20, %rd292, %rd292);
|
|
// inline asm
|
|
add.f32 %f943, %f940, 0f00000000;
|
|
add.f32 %f944, %f941, 0f00000000;
|
|
add.f32 %f945, %f942, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs144, %f945;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs143, %f944;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs142, %f943;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs149, 0;
|
|
st.v4.u16 [%rd287], {%rs142, %rs143, %rs144, %rs149};
|
|
bra.uni BB0_161;
|
|
|
|
BB0_160:
|
|
mov.u64 %rd300, image_RNM3;
|
|
cvta.global.u64 %rd295, %rd300;
|
|
mov.u32 %r378, 8;
|
|
mov.u64 %rd299, 0;
|
|
// inline asm
|
|
call (%rd294), _rt_buffer_get_64, (%rd295, %r101, %r378, %rd19, %rd20, %rd299, %rd299);
|
|
// inline asm
|
|
mov.f32 %f946, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs150, %f946;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs151, 0;
|
|
st.v4.u16 [%rd294], {%rs150, %rs150, %rs150, %rs151};
|
|
bra.uni BB0_161;
|
|
|
|
BB0_81:
|
|
setp.geu.f32 %p80, %f1021, 0f00000000;
|
|
@%p80 bra BB0_84;
|
|
|
|
cvt.rzi.f32.f32 %f604, %f576;
|
|
setp.neu.f32 %p81, %f604, 0f3EE8BA2E;
|
|
selp.f32 %f1024, 0f7FFFFFFF, %f1024, %p81;
|
|
|
|
BB0_84:
|
|
add.f32 %f606, %f164, 0f3EE8BA2E;
|
|
mov.b32 %r265, %f606;
|
|
setp.lt.s32 %p83, %r265, 2139095040;
|
|
@%p83 bra BB0_89;
|
|
|
|
setp.gtu.f32 %p84, %f164, 0f7F800000;
|
|
@%p84 bra BB0_88;
|
|
bra.uni BB0_86;
|
|
|
|
BB0_88:
|
|
add.f32 %f1024, %f1021, 0f3EE8BA2E;
|
|
bra.uni BB0_89;
|
|
|
|
BB0_86:
|
|
setp.neu.f32 %p85, %f164, 0f7F800000;
|
|
@%p85 bra BB0_89;
|
|
|
|
selp.f32 %f1024, 0fFF800000, 0f7F800000, %p3;
|
|
|
|
BB0_89:
|
|
mul.f32 %f607, %f1024, 0f437F0000;
|
|
setp.eq.f32 %p86, %f1021, 0f3F800000;
|
|
selp.f32 %f608, 0f437F0000, %f607, %p86;
|
|
cvt.rzi.u32.f32 %r266, %f608;
|
|
cvt.u16.u32 %rs14, %r266;
|
|
mov.u16 %rs15, 255;
|
|
st.v2.u8 [%rd73], {%rs14, %rs15};
|
|
ld.global.u32 %r412, [imageEnabled];
|
|
|
|
BB0_90:
|
|
ld.global.f32 %f609, [lightColor];
|
|
mul.f32 %f176, %f1016, %f609;
|
|
ld.global.f32 %f610, [lightColor+4];
|
|
mul.f32 %f177, %f1016, %f610;
|
|
ld.global.f32 %f611, [lightColor+8];
|
|
mul.f32 %f178, %f1016, %f611;
|
|
and.b32 %r267, %r412, 1;
|
|
setp.eq.b32 %p87, %r267, 1;
|
|
@!%p87 bra BB0_125;
|
|
bra.uni BB0_91;
|
|
|
|
BB0_91:
|
|
mov.f32 %f614, 0f3E666666;
|
|
cvt.rzi.f32.f32 %f615, %f614;
|
|
fma.rn.f32 %f616, %f615, 0fC0000000, 0f3EE66666;
|
|
abs.f32 %f179, %f616;
|
|
abs.f32 %f180, %f176;
|
|
setp.lt.f32 %p88, %f180, 0f00800000;
|
|
mul.f32 %f617, %f180, 0f4B800000;
|
|
selp.f32 %f618, 0fC3170000, 0fC2FE0000, %p88;
|
|
selp.f32 %f619, %f617, %f180, %p88;
|
|
mov.b32 %r268, %f619;
|
|
and.b32 %r269, %r268, 8388607;
|
|
or.b32 %r270, %r269, 1065353216;
|
|
mov.b32 %f620, %r270;
|
|
shr.u32 %r271, %r268, 23;
|
|
cvt.rn.f32.u32 %f621, %r271;
|
|
add.f32 %f622, %f618, %f621;
|
|
setp.gt.f32 %p89, %f620, 0f3FB504F3;
|
|
mul.f32 %f623, %f620, 0f3F000000;
|
|
add.f32 %f624, %f622, 0f3F800000;
|
|
selp.f32 %f625, %f623, %f620, %p89;
|
|
selp.f32 %f626, %f624, %f622, %p89;
|
|
add.f32 %f627, %f625, 0fBF800000;
|
|
add.f32 %f613, %f625, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f612,%f613;
|
|
// inline asm
|
|
add.f32 %f628, %f627, %f627;
|
|
mul.f32 %f629, %f612, %f628;
|
|
mul.f32 %f630, %f629, %f629;
|
|
mov.f32 %f631, 0f3C4CAF63;
|
|
mov.f32 %f632, 0f3B18F0FE;
|
|
fma.rn.f32 %f633, %f632, %f630, %f631;
|
|
mov.f32 %f634, 0f3DAAAABD;
|
|
fma.rn.f32 %f635, %f633, %f630, %f634;
|
|
mul.rn.f32 %f636, %f635, %f630;
|
|
mul.rn.f32 %f637, %f636, %f629;
|
|
sub.f32 %f638, %f627, %f629;
|
|
neg.f32 %f639, %f629;
|
|
add.f32 %f640, %f638, %f638;
|
|
fma.rn.f32 %f641, %f639, %f627, %f640;
|
|
mul.rn.f32 %f642, %f612, %f641;
|
|
add.f32 %f643, %f637, %f629;
|
|
sub.f32 %f644, %f629, %f643;
|
|
add.f32 %f645, %f637, %f644;
|
|
add.f32 %f646, %f642, %f645;
|
|
add.f32 %f647, %f643, %f646;
|
|
sub.f32 %f648, %f643, %f647;
|
|
add.f32 %f649, %f646, %f648;
|
|
mov.f32 %f650, 0f3F317200;
|
|
mul.rn.f32 %f651, %f626, %f650;
|
|
mov.f32 %f652, 0f35BFBE8E;
|
|
mul.rn.f32 %f653, %f626, %f652;
|
|
add.f32 %f654, %f651, %f647;
|
|
sub.f32 %f655, %f651, %f654;
|
|
add.f32 %f656, %f647, %f655;
|
|
add.f32 %f657, %f649, %f656;
|
|
add.f32 %f658, %f653, %f657;
|
|
add.f32 %f659, %f654, %f658;
|
|
sub.f32 %f660, %f654, %f659;
|
|
add.f32 %f661, %f658, %f660;
|
|
mov.f32 %f662, 0f3EE66666;
|
|
mul.rn.f32 %f663, %f662, %f659;
|
|
neg.f32 %f664, %f663;
|
|
fma.rn.f32 %f665, %f662, %f659, %f664;
|
|
fma.rn.f32 %f666, %f662, %f661, %f665;
|
|
mov.f32 %f667, 0f00000000;
|
|
fma.rn.f32 %f668, %f667, %f659, %f666;
|
|
add.rn.f32 %f669, %f663, %f668;
|
|
neg.f32 %f670, %f669;
|
|
add.rn.f32 %f671, %f663, %f670;
|
|
add.rn.f32 %f672, %f671, %f668;
|
|
mov.b32 %r272, %f669;
|
|
setp.eq.s32 %p90, %r272, 1118925336;
|
|
add.s32 %r273, %r272, -1;
|
|
mov.b32 %f673, %r273;
|
|
add.f32 %f674, %f672, 0f37000000;
|
|
selp.f32 %f675, %f673, %f669, %p90;
|
|
selp.f32 %f181, %f674, %f672, %p90;
|
|
mul.f32 %f676, %f675, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f677, %f676;
|
|
mov.f32 %f678, 0fBF317200;
|
|
fma.rn.f32 %f679, %f677, %f678, %f675;
|
|
mov.f32 %f680, 0fB5BFBE8E;
|
|
fma.rn.f32 %f681, %f677, %f680, %f679;
|
|
mul.f32 %f682, %f681, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f683, %f682;
|
|
add.f32 %f684, %f677, 0f00000000;
|
|
ex2.approx.f32 %f685, %f684;
|
|
mul.f32 %f686, %f683, %f685;
|
|
setp.lt.f32 %p91, %f675, 0fC2D20000;
|
|
selp.f32 %f687, 0f00000000, %f686, %p91;
|
|
setp.gt.f32 %p92, %f675, 0f42D20000;
|
|
selp.f32 %f1025, 0f7F800000, %f687, %p92;
|
|
setp.eq.f32 %p93, %f1025, 0f7F800000;
|
|
@%p93 bra BB0_93;
|
|
|
|
fma.rn.f32 %f1025, %f1025, %f181, %f1025;
|
|
|
|
BB0_93:
|
|
setp.lt.f32 %p94, %f176, 0f00000000;
|
|
setp.eq.f32 %p95, %f179, 0f3F800000;
|
|
and.pred %p4, %p94, %p95;
|
|
mov.b32 %r274, %f1025;
|
|
xor.b32 %r275, %r274, -2147483648;
|
|
mov.b32 %f688, %r275;
|
|
selp.f32 %f1027, %f688, %f1025, %p4;
|
|
setp.eq.f32 %p96, %f176, 0f00000000;
|
|
@%p96 bra BB0_96;
|
|
bra.uni BB0_94;
|
|
|
|
BB0_96:
|
|
add.f32 %f691, %f176, %f176;
|
|
selp.f32 %f1027, %f691, 0f00000000, %p95;
|
|
bra.uni BB0_97;
|
|
|
|
BB0_94:
|
|
setp.geu.f32 %p97, %f176, 0f00000000;
|
|
@%p97 bra BB0_97;
|
|
|
|
cvt.rzi.f32.f32 %f690, %f662;
|
|
setp.neu.f32 %p98, %f690, 0f3EE66666;
|
|
selp.f32 %f1027, 0f7FFFFFFF, %f1027, %p98;
|
|
|
|
BB0_97:
|
|
add.f32 %f692, %f180, 0f3EE66666;
|
|
mov.b32 %r276, %f692;
|
|
setp.lt.s32 %p100, %r276, 2139095040;
|
|
@%p100 bra BB0_102;
|
|
|
|
setp.gtu.f32 %p101, %f180, 0f7F800000;
|
|
@%p101 bra BB0_101;
|
|
bra.uni BB0_99;
|
|
|
|
BB0_101:
|
|
add.f32 %f1027, %f176, 0f3EE66666;
|
|
bra.uni BB0_102;
|
|
|
|
BB0_99:
|
|
setp.neu.f32 %p102, %f180, 0f7F800000;
|
|
@%p102 bra BB0_102;
|
|
|
|
selp.f32 %f1027, 0fFF800000, 0f7F800000, %p4;
|
|
|
|
BB0_102:
|
|
setp.eq.f32 %p103, %f176, 0f3F800000;
|
|
selp.f32 %f192, 0f3F800000, %f1027, %p103;
|
|
abs.f32 %f193, %f177;
|
|
setp.lt.f32 %p104, %f193, 0f00800000;
|
|
mul.f32 %f695, %f193, 0f4B800000;
|
|
selp.f32 %f696, 0fC3170000, 0fC2FE0000, %p104;
|
|
selp.f32 %f697, %f695, %f193, %p104;
|
|
mov.b32 %r277, %f697;
|
|
and.b32 %r278, %r277, 8388607;
|
|
or.b32 %r279, %r278, 1065353216;
|
|
mov.b32 %f698, %r279;
|
|
shr.u32 %r280, %r277, 23;
|
|
cvt.rn.f32.u32 %f699, %r280;
|
|
add.f32 %f700, %f696, %f699;
|
|
setp.gt.f32 %p105, %f698, 0f3FB504F3;
|
|
mul.f32 %f701, %f698, 0f3F000000;
|
|
add.f32 %f702, %f700, 0f3F800000;
|
|
selp.f32 %f703, %f701, %f698, %p105;
|
|
selp.f32 %f704, %f702, %f700, %p105;
|
|
add.f32 %f705, %f703, 0fBF800000;
|
|
add.f32 %f694, %f703, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f693,%f694;
|
|
// inline asm
|
|
add.f32 %f706, %f705, %f705;
|
|
mul.f32 %f707, %f693, %f706;
|
|
mul.f32 %f708, %f707, %f707;
|
|
fma.rn.f32 %f711, %f632, %f708, %f631;
|
|
fma.rn.f32 %f713, %f711, %f708, %f634;
|
|
mul.rn.f32 %f714, %f713, %f708;
|
|
mul.rn.f32 %f715, %f714, %f707;
|
|
sub.f32 %f716, %f705, %f707;
|
|
neg.f32 %f717, %f707;
|
|
add.f32 %f718, %f716, %f716;
|
|
fma.rn.f32 %f719, %f717, %f705, %f718;
|
|
mul.rn.f32 %f720, %f693, %f719;
|
|
add.f32 %f721, %f715, %f707;
|
|
sub.f32 %f722, %f707, %f721;
|
|
add.f32 %f723, %f715, %f722;
|
|
add.f32 %f724, %f720, %f723;
|
|
add.f32 %f725, %f721, %f724;
|
|
sub.f32 %f726, %f721, %f725;
|
|
add.f32 %f727, %f724, %f726;
|
|
mul.rn.f32 %f729, %f704, %f650;
|
|
mul.rn.f32 %f731, %f704, %f652;
|
|
add.f32 %f732, %f729, %f725;
|
|
sub.f32 %f733, %f729, %f732;
|
|
add.f32 %f734, %f725, %f733;
|
|
add.f32 %f735, %f727, %f734;
|
|
add.f32 %f736, %f731, %f735;
|
|
add.f32 %f737, %f732, %f736;
|
|
sub.f32 %f738, %f732, %f737;
|
|
add.f32 %f739, %f736, %f738;
|
|
mul.rn.f32 %f741, %f662, %f737;
|
|
neg.f32 %f742, %f741;
|
|
fma.rn.f32 %f743, %f662, %f737, %f742;
|
|
fma.rn.f32 %f744, %f662, %f739, %f743;
|
|
fma.rn.f32 %f746, %f667, %f737, %f744;
|
|
add.rn.f32 %f747, %f741, %f746;
|
|
neg.f32 %f748, %f747;
|
|
add.rn.f32 %f749, %f741, %f748;
|
|
add.rn.f32 %f750, %f749, %f746;
|
|
mov.b32 %r281, %f747;
|
|
setp.eq.s32 %p106, %r281, 1118925336;
|
|
add.s32 %r282, %r281, -1;
|
|
mov.b32 %f751, %r282;
|
|
add.f32 %f752, %f750, 0f37000000;
|
|
selp.f32 %f753, %f751, %f747, %p106;
|
|
selp.f32 %f194, %f752, %f750, %p106;
|
|
mul.f32 %f754, %f753, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f755, %f754;
|
|
fma.rn.f32 %f757, %f755, %f678, %f753;
|
|
fma.rn.f32 %f759, %f755, %f680, %f757;
|
|
mul.f32 %f760, %f759, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f761, %f760;
|
|
add.f32 %f762, %f755, 0f00000000;
|
|
ex2.approx.f32 %f763, %f762;
|
|
mul.f32 %f764, %f761, %f763;
|
|
setp.lt.f32 %p107, %f753, 0fC2D20000;
|
|
selp.f32 %f765, 0f00000000, %f764, %p107;
|
|
setp.gt.f32 %p108, %f753, 0f42D20000;
|
|
selp.f32 %f1028, 0f7F800000, %f765, %p108;
|
|
setp.eq.f32 %p109, %f1028, 0f7F800000;
|
|
@%p109 bra BB0_104;
|
|
|
|
fma.rn.f32 %f1028, %f1028, %f194, %f1028;
|
|
|
|
BB0_104:
|
|
setp.lt.f32 %p110, %f177, 0f00000000;
|
|
and.pred %p5, %p110, %p95;
|
|
mov.b32 %r283, %f1028;
|
|
xor.b32 %r284, %r283, -2147483648;
|
|
mov.b32 %f766, %r284;
|
|
selp.f32 %f1030, %f766, %f1028, %p5;
|
|
setp.eq.f32 %p112, %f177, 0f00000000;
|
|
@%p112 bra BB0_107;
|
|
bra.uni BB0_105;
|
|
|
|
BB0_107:
|
|
add.f32 %f769, %f177, %f177;
|
|
selp.f32 %f1030, %f769, 0f00000000, %p95;
|
|
bra.uni BB0_108;
|
|
|
|
BB0_105:
|
|
setp.geu.f32 %p113, %f177, 0f00000000;
|
|
@%p113 bra BB0_108;
|
|
|
|
cvt.rzi.f32.f32 %f768, %f662;
|
|
setp.neu.f32 %p114, %f768, 0f3EE66666;
|
|
selp.f32 %f1030, 0f7FFFFFFF, %f1030, %p114;
|
|
|
|
BB0_108:
|
|
add.f32 %f770, %f193, 0f3EE66666;
|
|
mov.b32 %r285, %f770;
|
|
setp.lt.s32 %p116, %r285, 2139095040;
|
|
@%p116 bra BB0_113;
|
|
|
|
setp.gtu.f32 %p117, %f193, 0f7F800000;
|
|
@%p117 bra BB0_112;
|
|
bra.uni BB0_110;
|
|
|
|
BB0_112:
|
|
add.f32 %f1030, %f177, 0f3EE66666;
|
|
bra.uni BB0_113;
|
|
|
|
BB0_110:
|
|
setp.neu.f32 %p118, %f193, 0f7F800000;
|
|
@%p118 bra BB0_113;
|
|
|
|
selp.f32 %f1030, 0fFF800000, 0f7F800000, %p5;
|
|
|
|
BB0_113:
|
|
setp.eq.f32 %p119, %f177, 0f3F800000;
|
|
selp.f32 %f205, 0f3F800000, %f1030, %p119;
|
|
abs.f32 %f206, %f178;
|
|
setp.lt.f32 %p120, %f206, 0f00800000;
|
|
mul.f32 %f773, %f206, 0f4B800000;
|
|
selp.f32 %f774, 0fC3170000, 0fC2FE0000, %p120;
|
|
selp.f32 %f775, %f773, %f206, %p120;
|
|
mov.b32 %r286, %f775;
|
|
and.b32 %r287, %r286, 8388607;
|
|
or.b32 %r288, %r287, 1065353216;
|
|
mov.b32 %f776, %r288;
|
|
shr.u32 %r289, %r286, 23;
|
|
cvt.rn.f32.u32 %f777, %r289;
|
|
add.f32 %f778, %f774, %f777;
|
|
setp.gt.f32 %p121, %f776, 0f3FB504F3;
|
|
mul.f32 %f779, %f776, 0f3F000000;
|
|
add.f32 %f780, %f778, 0f3F800000;
|
|
selp.f32 %f781, %f779, %f776, %p121;
|
|
selp.f32 %f782, %f780, %f778, %p121;
|
|
add.f32 %f783, %f781, 0fBF800000;
|
|
add.f32 %f772, %f781, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f771,%f772;
|
|
// inline asm
|
|
add.f32 %f784, %f783, %f783;
|
|
mul.f32 %f785, %f771, %f784;
|
|
mul.f32 %f786, %f785, %f785;
|
|
fma.rn.f32 %f789, %f632, %f786, %f631;
|
|
fma.rn.f32 %f791, %f789, %f786, %f634;
|
|
mul.rn.f32 %f792, %f791, %f786;
|
|
mul.rn.f32 %f793, %f792, %f785;
|
|
sub.f32 %f794, %f783, %f785;
|
|
neg.f32 %f795, %f785;
|
|
add.f32 %f796, %f794, %f794;
|
|
fma.rn.f32 %f797, %f795, %f783, %f796;
|
|
mul.rn.f32 %f798, %f771, %f797;
|
|
add.f32 %f799, %f793, %f785;
|
|
sub.f32 %f800, %f785, %f799;
|
|
add.f32 %f801, %f793, %f800;
|
|
add.f32 %f802, %f798, %f801;
|
|
add.f32 %f803, %f799, %f802;
|
|
sub.f32 %f804, %f799, %f803;
|
|
add.f32 %f805, %f802, %f804;
|
|
mul.rn.f32 %f807, %f782, %f650;
|
|
mul.rn.f32 %f809, %f782, %f652;
|
|
add.f32 %f810, %f807, %f803;
|
|
sub.f32 %f811, %f807, %f810;
|
|
add.f32 %f812, %f803, %f811;
|
|
add.f32 %f813, %f805, %f812;
|
|
add.f32 %f814, %f809, %f813;
|
|
add.f32 %f815, %f810, %f814;
|
|
sub.f32 %f816, %f810, %f815;
|
|
add.f32 %f817, %f814, %f816;
|
|
mul.rn.f32 %f819, %f662, %f815;
|
|
neg.f32 %f820, %f819;
|
|
fma.rn.f32 %f821, %f662, %f815, %f820;
|
|
fma.rn.f32 %f822, %f662, %f817, %f821;
|
|
fma.rn.f32 %f824, %f667, %f815, %f822;
|
|
add.rn.f32 %f825, %f819, %f824;
|
|
neg.f32 %f826, %f825;
|
|
add.rn.f32 %f827, %f819, %f826;
|
|
add.rn.f32 %f828, %f827, %f824;
|
|
mov.b32 %r290, %f825;
|
|
setp.eq.s32 %p122, %r290, 1118925336;
|
|
add.s32 %r291, %r290, -1;
|
|
mov.b32 %f829, %r291;
|
|
add.f32 %f830, %f828, 0f37000000;
|
|
selp.f32 %f831, %f829, %f825, %p122;
|
|
selp.f32 %f207, %f830, %f828, %p122;
|
|
mul.f32 %f832, %f831, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f833, %f832;
|
|
fma.rn.f32 %f835, %f833, %f678, %f831;
|
|
fma.rn.f32 %f837, %f833, %f680, %f835;
|
|
mul.f32 %f838, %f837, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f839, %f838;
|
|
add.f32 %f840, %f833, 0f00000000;
|
|
ex2.approx.f32 %f841, %f840;
|
|
mul.f32 %f842, %f839, %f841;
|
|
setp.lt.f32 %p123, %f831, 0fC2D20000;
|
|
selp.f32 %f843, 0f00000000, %f842, %p123;
|
|
setp.gt.f32 %p124, %f831, 0f42D20000;
|
|
selp.f32 %f1031, 0f7F800000, %f843, %p124;
|
|
setp.eq.f32 %p125, %f1031, 0f7F800000;
|
|
@%p125 bra BB0_115;
|
|
|
|
fma.rn.f32 %f1031, %f1031, %f207, %f1031;
|
|
|
|
BB0_115:
|
|
setp.lt.f32 %p126, %f178, 0f00000000;
|
|
and.pred %p6, %p126, %p95;
|
|
mov.b32 %r292, %f1031;
|
|
xor.b32 %r293, %r292, -2147483648;
|
|
mov.b32 %f844, %r293;
|
|
selp.f32 %f1033, %f844, %f1031, %p6;
|
|
setp.eq.f32 %p128, %f178, 0f00000000;
|
|
@%p128 bra BB0_118;
|
|
bra.uni BB0_116;
|
|
|
|
BB0_118:
|
|
add.f32 %f847, %f178, %f178;
|
|
selp.f32 %f1033, %f847, 0f00000000, %p95;
|
|
bra.uni BB0_119;
|
|
|
|
BB0_116:
|
|
setp.geu.f32 %p129, %f178, 0f00000000;
|
|
@%p129 bra BB0_119;
|
|
|
|
cvt.rzi.f32.f32 %f846, %f662;
|
|
setp.neu.f32 %p130, %f846, 0f3EE66666;
|
|
selp.f32 %f1033, 0f7FFFFFFF, %f1033, %p130;
|
|
|
|
BB0_119:
|
|
add.f32 %f848, %f206, 0f3EE66666;
|
|
mov.b32 %r294, %f848;
|
|
setp.lt.s32 %p132, %r294, 2139095040;
|
|
@%p132 bra BB0_124;
|
|
|
|
setp.gtu.f32 %p133, %f206, 0f7F800000;
|
|
@%p133 bra BB0_123;
|
|
bra.uni BB0_121;
|
|
|
|
BB0_123:
|
|
add.f32 %f1033, %f178, 0f3EE66666;
|
|
bra.uni BB0_124;
|
|
|
|
BB0_121:
|
|
setp.neu.f32 %p134, %f206, 0f7F800000;
|
|
@%p134 bra BB0_124;
|
|
|
|
selp.f32 %f1033, 0fFF800000, 0f7F800000, %p6;
|
|
|
|
BB0_124:
|
|
setp.eq.f32 %p135, %f178, 0f3F800000;
|
|
selp.f32 %f849, 0f3F800000, %f1033, %p135;
|
|
cvt.u64.u32 %rd83, %r4;
|
|
cvt.u64.u32 %rd82, %r3;
|
|
mov.u64 %rd86, image;
|
|
cvta.global.u64 %rd81, %rd86;
|
|
// inline asm
|
|
call (%rd80), _rt_buffer_get_64, (%rd81, %r101, %r102, %rd82, %rd83, %rd26, %rd26);
|
|
// inline asm
|
|
cvt.sat.f32.f32 %f850, %f849;
|
|
mul.f32 %f851, %f850, 0f437FFD71;
|
|
cvt.rzi.u32.f32 %r297, %f851;
|
|
cvt.sat.f32.f32 %f852, %f205;
|
|
mul.f32 %f853, %f852, 0f437FFD71;
|
|
cvt.rzi.u32.f32 %r298, %f853;
|
|
cvt.sat.f32.f32 %f854, %f192;
|
|
mul.f32 %f855, %f854, 0f437FFD71;
|
|
cvt.rzi.u32.f32 %r299, %f855;
|
|
cvt.u16.u32 %rs16, %r297;
|
|
cvt.u16.u32 %rs17, %r299;
|
|
cvt.u16.u32 %rs18, %r298;
|
|
mov.u16 %rs19, 255;
|
|
st.v4.u8 [%rd80], {%rs16, %rs18, %rs17, %rs19};
|
|
ld.global.u32 %r412, [imageEnabled];
|
|
|
|
BB0_125:
|
|
cvt.u64.u32 %rd17, %r3;
|
|
cvt.u64.u32 %rd18, %r4;
|
|
and.b32 %r300, %r412, 4;
|
|
setp.eq.s32 %p136, %r300, 0;
|
|
@%p136 bra BB0_129;
|
|
|
|
ld.global.u32 %r301, [additive];
|
|
setp.eq.s32 %p137, %r301, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs20, %f407;}
|
|
|
|
// inline asm
|
|
@%p137 bra BB0_128;
|
|
|
|
mov.u64 %rd99, image_HDR;
|
|
cvta.global.u64 %rd88, %rd99;
|
|
mov.u32 %r305, 8;
|
|
// inline asm
|
|
call (%rd87), _rt_buffer_get_64, (%rd88, %r101, %r305, %rd17, %rd18, %rd26, %rd26);
|
|
// inline asm
|
|
ld.v4.u16 {%rs27, %rs28, %rs29, %rs30}, [%rd87];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f857, %rs27;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f858, %rs28;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f859, %rs29;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd93), _rt_buffer_get_64, (%rd88, %r101, %r305, %rd17, %rd18, %rd26, %rd26);
|
|
// inline asm
|
|
add.f32 %f860, %f176, %f857;
|
|
add.f32 %f861, %f177, %f858;
|
|
add.f32 %f862, %f178, %f859;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs26, %f862;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs25, %f861;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs24, %f860;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd93], {%rs24, %rs25, %rs26, %rs20};
|
|
bra.uni BB0_129;
|
|
|
|
BB0_128:
|
|
mov.u64 %rd106, image_HDR;
|
|
cvta.global.u64 %rd101, %rd106;
|
|
mov.u32 %r307, 8;
|
|
// inline asm
|
|
call (%rd100), _rt_buffer_get_64, (%rd101, %r101, %r307, %rd17, %rd18, %rd26, %rd26);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs33, %f178;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs32, %f177;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs31, %f176;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd100], {%rs31, %rs32, %rs33, %rs20};
|
|
|
|
BB0_129:
|
|
selp.f32 %f867, 0f3F000000, 0f3E800000, %p150;
|
|
mul.f32 %f218, %f867, %f1017;
|
|
mul.f32 %f219, %f867, %f1018;
|
|
mul.f32 %f220, %f867, %f1019;
|
|
mul.f32 %f221, %f867, %f1020;
|
|
ld.global.f32 %f868, [lightColor];
|
|
mul.f32 %f222, %f218, %f868;
|
|
ld.global.f32 %f869, [lightColor+4];
|
|
mul.f32 %f223, %f218, %f869;
|
|
ld.global.f32 %f870, [lightColor+8];
|
|
mul.f32 %f224, %f218, %f870;
|
|
ld.global.u32 %r308, [additive];
|
|
setp.eq.s32 %p138, %r308, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs34, %f407;}
|
|
|
|
// inline asm
|
|
@%p138 bra BB0_131;
|
|
|
|
mov.u64 %rd119, image_RNM0;
|
|
cvta.global.u64 %rd108, %rd119;
|
|
mov.u32 %r312, 8;
|
|
// inline asm
|
|
call (%rd107), _rt_buffer_get_64, (%rd108, %r101, %r312, %rd17, %rd18, %rd26, %rd26);
|
|
// inline asm
|
|
ld.v4.u16 {%rs41, %rs42, %rs43, %rs44}, [%rd107];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f871, %rs41;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f872, %rs42;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f873, %rs43;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd113), _rt_buffer_get_64, (%rd108, %r101, %r312, %rd17, %rd18, %rd26, %rd26);
|
|
// inline asm
|
|
add.f32 %f874, %f222, %f871;
|
|
add.f32 %f875, %f223, %f872;
|
|
add.f32 %f876, %f224, %f873;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs40, %f876;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs39, %f875;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs38, %f874;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd113], {%rs38, %rs39, %rs40, %rs34};
|
|
bra.uni BB0_132;
|
|
|
|
BB0_131:
|
|
mov.u64 %rd126, image_RNM0;
|
|
cvta.global.u64 %rd121, %rd126;
|
|
mov.u32 %r314, 8;
|
|
// inline asm
|
|
call (%rd120), _rt_buffer_get_64, (%rd121, %r101, %r314, %rd17, %rd18, %rd26, %rd26);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs47, %f224;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs46, %f223;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs45, %f222;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd120], {%rs45, %rs46, %rs47, %rs34};
|
|
|
|
BB0_132:
|
|
mov.f32 %f881, 0f34000000;
|
|
max.f32 %f882, %f218, %f881;
|
|
div.rn.f32 %f883, %f219, %f882;
|
|
fma.rn.f32 %f225, %f883, 0f3F000000, 0f3F000000;
|
|
div.rn.f32 %f884, %f220, %f882;
|
|
fma.rn.f32 %f226, %f884, 0f3F000000, 0f3F000000;
|
|
div.rn.f32 %f885, %f221, %f882;
|
|
fma.rn.f32 %f227, %f885, 0f3F000000, 0f3F000000;
|
|
ld.global.u32 %r315, [additive];
|
|
setp.eq.s32 %p139, %r315, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs48, %f407;}
|
|
|
|
// inline asm
|
|
@%p139 bra BB0_134;
|
|
|
|
mov.u64 %rd139, image_RNM1;
|
|
cvta.global.u64 %rd128, %rd139;
|
|
mov.u32 %r319, 8;
|
|
// inline asm
|
|
call (%rd127), _rt_buffer_get_64, (%rd128, %r101, %r319, %rd17, %rd18, %rd26, %rd26);
|
|
// inline asm
|
|
ld.v4.u16 {%rs55, %rs56, %rs57, %rs58}, [%rd127];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f886, %rs55;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f887, %rs56;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f888, %rs57;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd133), _rt_buffer_get_64, (%rd128, %r101, %r319, %rd17, %rd18, %rd26, %rd26);
|
|
// inline asm
|
|
add.f32 %f889, %f225, %f886;
|
|
add.f32 %f890, %f225, %f887;
|
|
add.f32 %f891, %f225, %f888;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs54, %f891;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs53, %f890;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs52, %f889;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd133], {%rs52, %rs53, %rs54, %rs48};
|
|
bra.uni BB0_135;
|
|
|
|
BB0_134:
|
|
mov.u64 %rd146, image_RNM1;
|
|
cvta.global.u64 %rd141, %rd146;
|
|
mov.u32 %r321, 8;
|
|
// inline asm
|
|
call (%rd140), _rt_buffer_get_64, (%rd141, %r101, %r321, %rd17, %rd18, %rd26, %rd26);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs59, %f225;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd140], {%rs59, %rs59, %rs59, %rs48};
|
|
|
|
BB0_135:
|
|
ld.global.u32 %r322, [additive];
|
|
setp.eq.s32 %p140, %r322, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs60, %f407;}
|
|
|
|
// inline asm
|
|
@%p140 bra BB0_137;
|
|
|
|
mov.u64 %rd159, image_RNM2;
|
|
cvta.global.u64 %rd148, %rd159;
|
|
mov.u32 %r326, 8;
|
|
// inline asm
|
|
call (%rd147), _rt_buffer_get_64, (%rd148, %r101, %r326, %rd17, %rd18, %rd26, %rd26);
|
|
// inline asm
|
|
ld.v4.u16 {%rs67, %rs68, %rs69, %rs70}, [%rd147];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f894, %rs67;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f895, %rs68;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f896, %rs69;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd153), _rt_buffer_get_64, (%rd148, %r101, %r326, %rd17, %rd18, %rd26, %rd26);
|
|
// inline asm
|
|
add.f32 %f897, %f226, %f894;
|
|
add.f32 %f898, %f226, %f895;
|
|
add.f32 %f899, %f226, %f896;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs66, %f899;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs65, %f898;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs64, %f897;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd153], {%rs64, %rs65, %rs66, %rs60};
|
|
bra.uni BB0_138;
|
|
|
|
BB0_137:
|
|
mov.u64 %rd166, image_RNM2;
|
|
cvta.global.u64 %rd161, %rd166;
|
|
mov.u32 %r328, 8;
|
|
// inline asm
|
|
call (%rd160), _rt_buffer_get_64, (%rd161, %r101, %r328, %rd17, %rd18, %rd26, %rd26);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs71, %f226;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd160], {%rs71, %rs71, %rs71, %rs60};
|
|
|
|
BB0_138:
|
|
ld.global.u32 %r329, [additive];
|
|
setp.eq.s32 %p141, %r329, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs72, %f407;}
|
|
|
|
// inline asm
|
|
@%p141 bra BB0_140;
|
|
|
|
mov.u64 %rd179, image_RNM3;
|
|
cvta.global.u64 %rd168, %rd179;
|
|
mov.u32 %r333, 8;
|
|
// inline asm
|
|
call (%rd167), _rt_buffer_get_64, (%rd168, %r101, %r333, %rd17, %rd18, %rd26, %rd26);
|
|
// inline asm
|
|
ld.v4.u16 {%rs79, %rs80, %rs81, %rs82}, [%rd167];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f902, %rs79;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f903, %rs80;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f904, %rs81;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd173), _rt_buffer_get_64, (%rd168, %r101, %r333, %rd17, %rd18, %rd26, %rd26);
|
|
// inline asm
|
|
add.f32 %f905, %f227, %f902;
|
|
add.f32 %f906, %f227, %f903;
|
|
add.f32 %f907, %f227, %f904;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs78, %f907;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs77, %f906;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs76, %f905;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd173], {%rs76, %rs77, %rs78, %rs72};
|
|
bra.uni BB0_161;
|
|
|
|
BB0_140:
|
|
mov.u64 %rd186, image_RNM3;
|
|
cvta.global.u64 %rd181, %rd186;
|
|
mov.u32 %r335, 8;
|
|
// inline asm
|
|
call (%rd180), _rt_buffer_get_64, (%rd181, %r101, %r335, %rd17, %rd18, %rd26, %rd26);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs83, %f227;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd180], {%rs83, %rs83, %rs83, %rs72};
|
|
|
|
BB0_161:
|
|
ret;
|
|
}
|
|
|
|
|