1853 lines
53 KiB
Plaintext
1853 lines
53 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 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 1 .b8 rnd_seeds2[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[44];
|
|
.reg .b64 %SP;
|
|
.reg .b64 %SPL;
|
|
.reg .pred %p<142>;
|
|
.reg .b16 %rs<47>;
|
|
.reg .f32 %f<906>;
|
|
.reg .b32 %r<366>;
|
|
.reg .b64 %rd<157>;
|
|
|
|
|
|
mov.u64 %rd156, __local_depot0;
|
|
cvta.local.u64 %SP, %rd156;
|
|
ld.global.u32 %r1, [lightMeshBufferSize];
|
|
abs.s32 %r2, %r1;
|
|
ld.global.u32 %r3, [samples];
|
|
ld.global.v2.u32 {%r105, %r106}, [pixelID];
|
|
cvt.u64.u32 %rd22, %r105;
|
|
cvt.u64.u32 %rd23, %r106;
|
|
mov.u64 %rd26, uvnormal;
|
|
cvta.global.u64 %rd21, %rd26;
|
|
mov.u32 %r103, 2;
|
|
mov.u32 %r104, 4;
|
|
mov.u64 %rd25, 0;
|
|
// inline asm
|
|
call (%rd20), _rt_buffer_get_64, (%rd21, %r103, %r104, %rd22, %rd23, %rd25, %rd25);
|
|
// inline asm
|
|
ld.u32 %r4, [%rd20];
|
|
shr.u32 %r109, %r4, 16;
|
|
cvt.u16.u32 %rs1, %r109;
|
|
and.b16 %rs3, %rs1, 255;
|
|
cvt.u16.u32 %rs4, %r4;
|
|
or.b16 %rs5, %rs4, %rs3;
|
|
setp.eq.s16 %p7, %rs5, 0;
|
|
mov.f32 %f847, 0f00000000;
|
|
mov.f32 %f848, %f847;
|
|
mov.f32 %f849, %f847;
|
|
@%p7 bra BB0_2;
|
|
|
|
ld.u8 %rs6, [%rd20+1];
|
|
and.b16 %rs8, %rs4, 255;
|
|
cvt.rn.f32.u16 %f193, %rs8;
|
|
div.rn.f32 %f194, %f193, 0f437F0000;
|
|
fma.rn.f32 %f195, %f194, 0f40000000, 0fBF800000;
|
|
cvt.rn.f32.u16 %f196, %rs6;
|
|
div.rn.f32 %f197, %f196, 0f437F0000;
|
|
fma.rn.f32 %f198, %f197, 0f40000000, 0fBF800000;
|
|
cvt.rn.f32.u16 %f199, %rs3;
|
|
div.rn.f32 %f200, %f199, 0f437F0000;
|
|
fma.rn.f32 %f201, %f200, 0f40000000, 0fBF800000;
|
|
mul.f32 %f202, %f198, %f198;
|
|
fma.rn.f32 %f203, %f195, %f195, %f202;
|
|
fma.rn.f32 %f204, %f201, %f201, %f203;
|
|
sqrt.rn.f32 %f205, %f204;
|
|
rcp.rn.f32 %f206, %f205;
|
|
mul.f32 %f847, %f195, %f206;
|
|
mul.f32 %f848, %f198, %f206;
|
|
mul.f32 %f849, %f201, %f206;
|
|
|
|
BB0_2:
|
|
ld.global.v2.u32 {%r110, %r111}, [pixelID];
|
|
ld.global.v2.u32 {%r113, %r114}, [tileInfo];
|
|
add.s32 %r5, %r110, %r113;
|
|
add.s32 %r6, %r111, %r114;
|
|
setp.eq.f32 %p8, %f848, 0f00000000;
|
|
setp.eq.f32 %p9, %f847, 0f00000000;
|
|
and.pred %p10, %p9, %p8;
|
|
setp.eq.f32 %p11, %f849, 0f00000000;
|
|
and.pred %p12, %p10, %p11;
|
|
@%p12 bra BB0_130;
|
|
bra.uni BB0_3;
|
|
|
|
BB0_130:
|
|
ld.global.u32 %r365, [imageEnabled];
|
|
and.b32 %r315, %r365, 1;
|
|
setp.eq.b32 %p137, %r315, 1;
|
|
@!%p137 bra BB0_132;
|
|
bra.uni BB0_131;
|
|
|
|
BB0_131:
|
|
cvt.u64.u32 %rd116, %r5;
|
|
cvt.u64.u32 %rd117, %r6;
|
|
mov.u64 %rd120, image;
|
|
cvta.global.u64 %rd115, %rd120;
|
|
// inline asm
|
|
call (%rd114), _rt_buffer_get_64, (%rd115, %r103, %r104, %rd116, %rd117, %rd25, %rd25);
|
|
// inline asm
|
|
mov.u16 %rs31, 0;
|
|
st.v4.u8 [%rd114], {%rs31, %rs31, %rs31, %rs31};
|
|
ld.global.u32 %r365, [imageEnabled];
|
|
|
|
BB0_132:
|
|
and.b32 %r318, %r365, 8;
|
|
setp.eq.s32 %p138, %r318, 0;
|
|
@%p138 bra BB0_134;
|
|
|
|
cvt.u64.u32 %rd123, %r5;
|
|
cvt.u64.u32 %rd124, %r6;
|
|
mov.u64 %rd127, image_Mask;
|
|
cvta.global.u64 %rd122, %rd127;
|
|
// inline asm
|
|
call (%rd121), _rt_buffer_get_64, (%rd122, %r103, %r103, %rd123, %rd124, %rd25, %rd25);
|
|
// inline asm
|
|
mov.f32 %f830, 0f00000000;
|
|
cvt.rzi.u32.f32 %r321, %f830;
|
|
cvt.u16.u32 %rs32, %r321;
|
|
mov.u16 %rs33, 0;
|
|
st.v2.u8 [%rd121], {%rs32, %rs33};
|
|
ld.global.u32 %r365, [imageEnabled];
|
|
|
|
BB0_134:
|
|
and.b32 %r322, %r365, 4;
|
|
setp.eq.s32 %p139, %r322, 0;
|
|
@%p139 bra BB0_138;
|
|
|
|
ld.global.u32 %r323, [additive];
|
|
setp.eq.s32 %p140, %r323, 0;
|
|
cvt.u64.u32 %rd18, %r5;
|
|
cvt.u64.u32 %rd19, %r6;
|
|
@%p140 bra BB0_137;
|
|
|
|
mov.u64 %rd140, image_HDR;
|
|
cvta.global.u64 %rd129, %rd140;
|
|
mov.u32 %r327, 8;
|
|
// inline asm
|
|
call (%rd128), _rt_buffer_get_64, (%rd129, %r103, %r327, %rd18, %rd19, %rd25, %rd25);
|
|
// inline asm
|
|
ld.v4.u16 {%rs40, %rs41, %rs42, %rs43}, [%rd128];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f831, %rs40;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f832, %rs41;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f833, %rs42;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd134), _rt_buffer_get_64, (%rd129, %r103, %r327, %rd18, %rd19, %rd25, %rd25);
|
|
// inline asm
|
|
add.f32 %f834, %f831, 0f00000000;
|
|
add.f32 %f835, %f832, 0f00000000;
|
|
add.f32 %f836, %f833, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs39, %f836;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs38, %f835;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs37, %f834;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs44, 0;
|
|
st.v4.u16 [%rd134], {%rs37, %rs38, %rs39, %rs44};
|
|
bra.uni BB0_138;
|
|
|
|
BB0_3:
|
|
ld.global.v2.u32 {%r123, %r124}, [pixelID];
|
|
cvt.u64.u32 %rd29, %r123;
|
|
cvt.u64.u32 %rd30, %r124;
|
|
mov.u64 %rd39, uvpos;
|
|
cvta.global.u64 %rd28, %rd39;
|
|
mov.u32 %r119, 12;
|
|
// inline asm
|
|
call (%rd27), _rt_buffer_get_64, (%rd28, %r103, %r119, %rd29, %rd30, %rd25, %rd25);
|
|
// inline asm
|
|
ld.f32 %f9, [%rd27+8];
|
|
ld.f32 %f8, [%rd27+4];
|
|
ld.f32 %f7, [%rd27];
|
|
mul.f32 %f211, %f7, 0f3456BF95;
|
|
mul.f32 %f212, %f8, 0f3456BF95;
|
|
mul.f32 %f213, %f9, 0f3456BF95;
|
|
abs.f32 %f10, %f847;
|
|
div.rn.f32 %f214, %f211, %f10;
|
|
abs.f32 %f215, %f848;
|
|
div.rn.f32 %f216, %f212, %f215;
|
|
abs.f32 %f11, %f849;
|
|
div.rn.f32 %f217, %f213, %f11;
|
|
abs.f32 %f218, %f214;
|
|
abs.f32 %f219, %f216;
|
|
abs.f32 %f220, %f217;
|
|
mov.f32 %f221, 0f38D1B717;
|
|
max.f32 %f222, %f218, %f221;
|
|
max.f32 %f223, %f219, %f221;
|
|
max.f32 %f224, %f220, %f221;
|
|
fma.rn.f32 %f12, %f847, %f222, %f7;
|
|
fma.rn.f32 %f13, %f848, %f223, %f8;
|
|
fma.rn.f32 %f14, %f849, %f224, %f9;
|
|
ld.global.v2.u32 {%r127, %r128}, [pixelID];
|
|
cvt.u64.u32 %rd35, %r127;
|
|
cvt.u64.u32 %rd36, %r128;
|
|
mov.u64 %rd40, rnd_seeds;
|
|
cvta.global.u64 %rd34, %rd40;
|
|
// inline asm
|
|
call (%rd33), _rt_buffer_get_64, (%rd34, %r103, %r104, %rd35, %rd36, %rd25, %rd25);
|
|
// inline asm
|
|
ld.u32 %r341, [%rd33];
|
|
mov.pred %p13, 0;
|
|
mov.f32 %f19, 0f00000000;
|
|
mov.u32 %r9, 0;
|
|
setp.lt.s32 %p14, %r2, 1;
|
|
@%p14 bra BB0_4;
|
|
|
|
ld.global.f32 %f15, [lightPointSize];
|
|
mul.f32 %f16, %f12, 0f3456BF95;
|
|
mul.f32 %f17, %f13, 0f3456BF95;
|
|
mul.f32 %f18, %f14, 0f3456BF95;
|
|
mov.f32 %f19, 0f00000000;
|
|
mov.u32 %r334, 0;
|
|
abs.f32 %f352, %f17;
|
|
abs.f32 %f353, %f16;
|
|
max.f32 %f354, %f353, %f352;
|
|
abs.f32 %f355, %f18;
|
|
max.f32 %f356, %f354, %f355;
|
|
mov.u32 %r9, %r334;
|
|
mov.f32 %f20, %f19;
|
|
mov.f32 %f21, %f19;
|
|
mov.f32 %f22, %f19;
|
|
|
|
BB0_6:
|
|
mul.lo.s32 %r10, %r334, 3;
|
|
cvt.s64.s32 %rd43, %r10;
|
|
mov.u64 %rd47, lightMeshBuffer;
|
|
cvta.global.u64 %rd42, %rd47;
|
|
mov.u32 %r133, 1;
|
|
// inline asm
|
|
call (%rd41), _rt_buffer_get_64, (%rd42, %r133, %r119, %rd43, %rd25, %rd25, %rd25);
|
|
// inline asm
|
|
ld.f32 %f229, [%rd41];
|
|
sub.f32 %f230, %f229, %f7;
|
|
ld.f32 %f231, [%rd41+4];
|
|
sub.f32 %f232, %f231, %f8;
|
|
ld.f32 %f233, [%rd41+8];
|
|
sub.f32 %f234, %f233, %f9;
|
|
mul.f32 %f235, %f232, %f232;
|
|
fma.rn.f32 %f236, %f230, %f230, %f235;
|
|
fma.rn.f32 %f237, %f234, %f234, %f236;
|
|
sqrt.rn.f32 %f23, %f237;
|
|
rcp.rn.f32 %f238, %f23;
|
|
mul.f32 %f24, %f230, %f238;
|
|
mul.f32 %f25, %f232, %f238;
|
|
mul.f32 %f26, %f234, %f238;
|
|
mul.f32 %f239, %f848, %f25;
|
|
fma.rn.f32 %f240, %f847, %f24, %f239;
|
|
fma.rn.f32 %f27, %f849, %f26, %f240;
|
|
setp.leu.f32 %p15, %f27, 0f00000000;
|
|
@%p15 bra BB0_23;
|
|
|
|
setp.eq.s32 %p16, %r3, 0;
|
|
mul.f32 %f241, %f23, %f23;
|
|
div.rn.f32 %f28, %f15, %f241;
|
|
ld.global.v2.u32 {%r137, %r138}, [pixelID];
|
|
cvt.u64.u32 %rd54, %r138;
|
|
cvt.u64.u32 %rd55, %r137;
|
|
and.b64 %rd50, %rd55, 255;
|
|
and.b64 %rd51, %rd54, 255;
|
|
mov.u64 %rd56, rnd_seeds2;
|
|
cvta.global.u64 %rd49, %rd56;
|
|
mov.u32 %r136, 3;
|
|
// inline asm
|
|
call (%rd48), _rt_buffer_get_64, (%rd49, %r103, %r136, %rd50, %rd51, %rd25, %rd25);
|
|
// inline asm
|
|
@%p16 bra BB0_9;
|
|
|
|
ld.u8 %rs10, [%rd48];
|
|
cvt.rn.f32.u16 %f242, %rs10;
|
|
div.rn.f32 %f243, %f242, 0fC37F0000;
|
|
fma.rn.f32 %f244, %f243, 0f3F333333, 0f3F800000;
|
|
setp.gt.f32 %p18, %f28, %f244;
|
|
mov.pred %p141, -1;
|
|
@%p18 bra BB0_25;
|
|
|
|
BB0_9:
|
|
ld.global.f32 %f247, [lightInvCutoff];
|
|
mul.f32 %f29, %f23, %f247;
|
|
mov.f32 %f251, 0f40800000;
|
|
abs.f32 %f31, %f29;
|
|
setp.lt.f32 %p19, %f31, 0f00800000;
|
|
mul.f32 %f253, %f31, 0f4B800000;
|
|
selp.f32 %f254, 0fC3170000, 0fC2FE0000, %p19;
|
|
selp.f32 %f255, %f253, %f31, %p19;
|
|
mov.b32 %r141, %f255;
|
|
and.b32 %r142, %r141, 8388607;
|
|
or.b32 %r143, %r142, 1065353216;
|
|
mov.b32 %f256, %r143;
|
|
shr.u32 %r144, %r141, 23;
|
|
cvt.rn.f32.u32 %f257, %r144;
|
|
add.f32 %f258, %f254, %f257;
|
|
setp.gt.f32 %p20, %f256, 0f3FB504F3;
|
|
mul.f32 %f259, %f256, 0f3F000000;
|
|
add.f32 %f260, %f258, 0f3F800000;
|
|
selp.f32 %f261, %f259, %f256, %p20;
|
|
selp.f32 %f262, %f260, %f258, %p20;
|
|
add.f32 %f263, %f261, 0fBF800000;
|
|
add.f32 %f246, %f261, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f245,%f246;
|
|
// inline asm
|
|
add.f32 %f264, %f263, %f263;
|
|
mul.f32 %f265, %f245, %f264;
|
|
mul.f32 %f266, %f265, %f265;
|
|
mov.f32 %f267, 0f3C4CAF63;
|
|
mov.f32 %f268, 0f3B18F0FE;
|
|
fma.rn.f32 %f269, %f268, %f266, %f267;
|
|
mov.f32 %f270, 0f3DAAAABD;
|
|
fma.rn.f32 %f271, %f269, %f266, %f270;
|
|
mul.rn.f32 %f272, %f271, %f266;
|
|
mul.rn.f32 %f273, %f272, %f265;
|
|
sub.f32 %f274, %f263, %f265;
|
|
neg.f32 %f275, %f265;
|
|
add.f32 %f276, %f274, %f274;
|
|
fma.rn.f32 %f277, %f275, %f263, %f276;
|
|
mul.rn.f32 %f278, %f245, %f277;
|
|
add.f32 %f279, %f273, %f265;
|
|
sub.f32 %f280, %f265, %f279;
|
|
add.f32 %f281, %f273, %f280;
|
|
add.f32 %f282, %f278, %f281;
|
|
add.f32 %f283, %f279, %f282;
|
|
sub.f32 %f284, %f279, %f283;
|
|
add.f32 %f285, %f282, %f284;
|
|
mov.f32 %f286, 0f3F317200;
|
|
mul.rn.f32 %f287, %f262, %f286;
|
|
mov.f32 %f288, 0f35BFBE8E;
|
|
mul.rn.f32 %f289, %f262, %f288;
|
|
add.f32 %f290, %f287, %f283;
|
|
sub.f32 %f291, %f287, %f290;
|
|
add.f32 %f292, %f283, %f291;
|
|
add.f32 %f293, %f285, %f292;
|
|
add.f32 %f294, %f289, %f293;
|
|
add.f32 %f295, %f290, %f294;
|
|
sub.f32 %f296, %f290, %f295;
|
|
add.f32 %f297, %f294, %f296;
|
|
mul.rn.f32 %f298, %f251, %f295;
|
|
neg.f32 %f299, %f298;
|
|
fma.rn.f32 %f300, %f251, %f295, %f299;
|
|
fma.rn.f32 %f301, %f251, %f297, %f300;
|
|
mov.f32 %f302, 0f00000000;
|
|
fma.rn.f32 %f303, %f302, %f295, %f301;
|
|
add.rn.f32 %f304, %f298, %f303;
|
|
neg.f32 %f305, %f304;
|
|
add.rn.f32 %f306, %f298, %f305;
|
|
add.rn.f32 %f307, %f306, %f303;
|
|
mov.b32 %r145, %f304;
|
|
setp.eq.s32 %p21, %r145, 1118925336;
|
|
add.s32 %r146, %r145, -1;
|
|
mov.b32 %f308, %r146;
|
|
add.f32 %f309, %f307, 0f37000000;
|
|
selp.f32 %f310, %f308, %f304, %p21;
|
|
selp.f32 %f32, %f309, %f307, %p21;
|
|
mul.f32 %f311, %f310, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f312, %f311;
|
|
mov.f32 %f313, 0fBF317200;
|
|
fma.rn.f32 %f314, %f312, %f313, %f310;
|
|
mov.f32 %f315, 0fB5BFBE8E;
|
|
fma.rn.f32 %f316, %f312, %f315, %f314;
|
|
mul.f32 %f317, %f316, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f318, %f317;
|
|
add.f32 %f319, %f312, 0f00000000;
|
|
ex2.approx.f32 %f320, %f319;
|
|
mul.f32 %f321, %f318, %f320;
|
|
setp.lt.f32 %p22, %f310, 0fC2D20000;
|
|
selp.f32 %f322, 0f00000000, %f321, %p22;
|
|
setp.gt.f32 %p23, %f310, 0f42D20000;
|
|
selp.f32 %f854, 0f7F800000, %f322, %p23;
|
|
setp.eq.f32 %p24, %f854, 0f7F800000;
|
|
@%p24 bra BB0_11;
|
|
|
|
fma.rn.f32 %f854, %f854, %f32, %f854;
|
|
|
|
BB0_11:
|
|
mov.f32 %f843, 0f40000000;
|
|
cvt.rzi.f32.f32 %f842, %f843;
|
|
add.f32 %f841, %f842, %f842;
|
|
mov.f32 %f840, 0f40800000;
|
|
sub.f32 %f839, %f840, %f841;
|
|
abs.f32 %f838, %f839;
|
|
setp.lt.f32 %p25, %f29, 0f00000000;
|
|
setp.eq.f32 %p26, %f838, 0f3F800000;
|
|
and.pred %p1, %p25, %p26;
|
|
mov.b32 %r147, %f854;
|
|
xor.b32 %r148, %r147, -2147483648;
|
|
mov.b32 %f323, %r148;
|
|
selp.f32 %f856, %f323, %f854, %p1;
|
|
setp.eq.f32 %p27, %f29, 0f00000000;
|
|
@%p27 bra BB0_14;
|
|
bra.uni BB0_12;
|
|
|
|
BB0_14:
|
|
add.f32 %f326, %f29, %f29;
|
|
selp.f32 %f856, %f326, 0f00000000, %p26;
|
|
bra.uni BB0_15;
|
|
|
|
BB0_12:
|
|
setp.geu.f32 %p28, %f29, 0f00000000;
|
|
@%p28 bra BB0_15;
|
|
|
|
mov.f32 %f846, 0f40800000;
|
|
cvt.rzi.f32.f32 %f325, %f846;
|
|
setp.neu.f32 %p29, %f325, 0f40800000;
|
|
selp.f32 %f856, 0f7FFFFFFF, %f856, %p29;
|
|
|
|
BB0_15:
|
|
add.f32 %f327, %f31, 0f40800000;
|
|
mov.b32 %r149, %f327;
|
|
setp.lt.s32 %p31, %r149, 2139095040;
|
|
@%p31 bra BB0_20;
|
|
|
|
setp.gtu.f32 %p32, %f31, 0f7F800000;
|
|
@%p32 bra BB0_19;
|
|
bra.uni BB0_17;
|
|
|
|
BB0_19:
|
|
add.f32 %f856, %f29, 0f40800000;
|
|
bra.uni BB0_20;
|
|
|
|
BB0_17:
|
|
setp.neu.f32 %p33, %f31, 0f7F800000;
|
|
@%p33 bra BB0_20;
|
|
|
|
selp.f32 %f856, 0fFF800000, 0f7F800000, %p1;
|
|
|
|
BB0_20:
|
|
mov.u32 %r331, 1;
|
|
mov.u64 %rd149, lightMeshBuffer;
|
|
cvta.global.u64 %rd148, %rd149;
|
|
mul.lo.s32 %r330, %r334, 3;
|
|
mov.f32 %f328, 0f3F800000;
|
|
sub.f32 %f329, %f328, %f856;
|
|
setp.eq.f32 %p34, %f29, 0f3F800000;
|
|
selp.f32 %f330, 0f00000000, %f329, %p34;
|
|
cvt.sat.f32.f32 %f331, %f330;
|
|
mul.f32 %f332, %f28, %f331;
|
|
add.s32 %r152, %r330, 1;
|
|
cvt.s64.s32 %rd59, %r152;
|
|
// inline asm
|
|
call (%rd57), _rt_buffer_get_64, (%rd148, %r331, %r119, %rd59, %rd25, %rd25, %rd25);
|
|
// inline asm
|
|
ld.f32 %f333, [%rd57];
|
|
mul.f32 %f334, %f24, %f333;
|
|
ld.f32 %f335, [%rd57+4];
|
|
mul.f32 %f336, %f25, %f335;
|
|
neg.f32 %f337, %f336;
|
|
sub.f32 %f338, %f337, %f334;
|
|
ld.f32 %f339, [%rd57+8];
|
|
mul.f32 %f340, %f26, %f339;
|
|
sub.f32 %f341, %f338, %f340;
|
|
cvt.sat.f32.f32 %f342, %f341;
|
|
mul.f32 %f43, %f332, %f342;
|
|
cvt.sat.f32.f32 %f44, %f27;
|
|
mul.f32 %f343, %f43, %f44;
|
|
setp.leu.f32 %p35, %f343, 0f3727C5AC;
|
|
@%p35 bra BB0_22;
|
|
|
|
mov.u32 %r333, 1;
|
|
mov.u64 %rd151, lightMeshBuffer;
|
|
cvta.global.u64 %rd150, %rd151;
|
|
mul.lo.s32 %r332, %r334, 3;
|
|
add.u64 %rd64, %SP, 28;
|
|
cvta.to.local.u64 %rd71, %rd64;
|
|
max.f32 %f350, %f356, %f221;
|
|
sub.f32 %f351, %f23, %f350;
|
|
mov.u32 %r158, 1065353216;
|
|
st.local.u32 [%rd71], %r158;
|
|
ld.global.u32 %r153, [root];
|
|
// inline asm
|
|
call _rt_trace_64, (%r153, %f12, %f13, %f14, %f24, %f25, %f26, %r333, %f350, %f351, %rd64, %r104);
|
|
// inline asm
|
|
add.s32 %r159, %r332, 2;
|
|
cvt.s64.s32 %rd67, %r159;
|
|
// inline asm
|
|
call (%rd65), _rt_buffer_get_64, (%rd150, %r333, %r119, %rd67, %rd25, %rd25, %rd25);
|
|
// inline asm
|
|
ld.f32 %f358, [%rd65];
|
|
ld.local.f32 %f359, [%rd71];
|
|
mul.f32 %f360, %f359, %f358;
|
|
ld.f32 %f361, [%rd65+4];
|
|
mul.f32 %f362, %f359, %f361;
|
|
ld.f32 %f363, [%rd65+8];
|
|
mul.f32 %f364, %f359, %f363;
|
|
mul.f32 %f365, %f43, 0f3EA2F983;
|
|
mul.f32 %f366, %f365, %f360;
|
|
mul.f32 %f367, %f365, %f362;
|
|
mul.f32 %f368, %f365, %f364;
|
|
fma.rn.f32 %f22, %f44, %f366, %f22;
|
|
fma.rn.f32 %f21, %f44, %f367, %f21;
|
|
fma.rn.f32 %f20, %f44, %f368, %f20;
|
|
add.f32 %f19, %f19, %f359;
|
|
|
|
BB0_22:
|
|
add.s32 %r9, %r9, 1;
|
|
|
|
BB0_23:
|
|
add.s32 %r334, %r334, 1;
|
|
setp.lt.s32 %p37, %r334, %r2;
|
|
@%p37 bra BB0_6;
|
|
bra.uni BB0_24;
|
|
|
|
BB0_4:
|
|
mov.f32 %f20, %f19;
|
|
mov.f32 %f21, %f19;
|
|
mov.f32 %f22, %f19;
|
|
|
|
BB0_24:
|
|
mov.pred %p141, %p13;
|
|
|
|
BB0_25:
|
|
cvt.rn.f32.s32 %f369, %r9;
|
|
mov.f32 %f370, 0f3F800000;
|
|
max.f32 %f371, %f369, %f370;
|
|
rcp.rn.f32 %f372, %f371;
|
|
mul.f32 %f890, %f22, %f372;
|
|
mul.f32 %f891, %f21, %f372;
|
|
mul.f32 %f892, %f20, %f372;
|
|
div.rn.f32 %f893, %f19, %f371;
|
|
@!%p141 bra BB0_78;
|
|
bra.uni BB0_26;
|
|
|
|
BB0_26:
|
|
abs.f32 %f845, %f849;
|
|
abs.f32 %f844, %f847;
|
|
setp.gt.f32 %p38, %f844, %f845;
|
|
neg.f32 %f376, %f848;
|
|
selp.f32 %f377, %f376, 0f00000000, %p38;
|
|
neg.f32 %f378, %f849;
|
|
selp.f32 %f379, %f847, %f378, %p38;
|
|
selp.f32 %f380, 0f00000000, %f848, %p38;
|
|
mul.f32 %f381, %f379, %f379;
|
|
fma.rn.f32 %f382, %f377, %f377, %f381;
|
|
fma.rn.f32 %f383, %f380, %f380, %f382;
|
|
sqrt.rn.f32 %f384, %f383;
|
|
rcp.rn.f32 %f385, %f384;
|
|
mul.f32 %f65, %f377, %f385;
|
|
mul.f32 %f66, %f379, %f385;
|
|
mul.f32 %f67, %f380, %f385;
|
|
mov.f32 %f375, 0f00000000;
|
|
setp.lt.s32 %p39, %r3, 1;
|
|
mov.f32 %f872, %f375;
|
|
mov.f32 %f873, %f375;
|
|
mov.f32 %f874, %f375;
|
|
@%p39 bra BB0_77;
|
|
|
|
cvt.rn.f32.s32 %f389, %r3;
|
|
rcp.rn.f32 %f68, %f389;
|
|
add.u64 %rd73, %SP, 0;
|
|
cvta.to.local.u64 %rd3, %rd73;
|
|
mul.f32 %f69, %f12, 0f3456BF95;
|
|
mul.f32 %f70, %f13, 0f3456BF95;
|
|
mul.f32 %f71, %f14, 0f3456BF95;
|
|
add.u64 %rd74, %SP, 32;
|
|
cvta.to.local.u64 %rd4, %rd74;
|
|
mul.f32 %f390, %f847, %f66;
|
|
mul.f32 %f391, %f848, %f65;
|
|
sub.f32 %f72, %f391, %f390;
|
|
mul.f32 %f392, %f849, %f65;
|
|
mul.f32 %f393, %f847, %f67;
|
|
sub.f32 %f73, %f393, %f392;
|
|
mul.f32 %f394, %f848, %f67;
|
|
mul.f32 %f395, %f849, %f66;
|
|
sub.f32 %f74, %f395, %f394;
|
|
mov.f32 %f872, 0f00000000;
|
|
mov.u32 %r160, 0;
|
|
abs.f32 %f396, %f70;
|
|
abs.f32 %f397, %f69;
|
|
max.f32 %f398, %f397, %f396;
|
|
abs.f32 %f399, %f71;
|
|
max.f32 %f400, %f398, %f399;
|
|
mov.u32 %r338, %r160;
|
|
mov.f32 %f873, %f872;
|
|
mov.f32 %f874, %f872;
|
|
|
|
BB0_28:
|
|
cvt.rn.f32.s32 %f78, %r338;
|
|
max.f32 %f79, %f400, %f221;
|
|
mov.u32 %r340, %r160;
|
|
|
|
BB0_29:
|
|
mad.lo.s32 %r162, %r341, 1664525, 1013904223;
|
|
and.b32 %r163, %r162, 16777215;
|
|
cvt.rn.f32.u32 %f402, %r163;
|
|
fma.rn.f32 %f403, %f402, 0f33800000, %f78;
|
|
mul.f32 %f404, %f68, %f403;
|
|
mad.lo.s32 %r341, %r162, 1664525, 1013904223;
|
|
and.b32 %r164, %r341, 16777215;
|
|
cvt.rn.f32.u32 %f405, %r164;
|
|
cvt.rn.f32.s32 %f406, %r340;
|
|
fma.rn.f32 %f407, %f405, 0f33800000, %f406;
|
|
mul.f32 %f408, %f68, %f407;
|
|
sqrt.rn.f32 %f83, %f404;
|
|
mul.f32 %f881, %f408, 0f40C90FDB;
|
|
abs.f32 %f85, %f881;
|
|
setp.neu.f32 %p40, %f85, 0f7F800000;
|
|
mov.f32 %f875, %f881;
|
|
@%p40 bra BB0_31;
|
|
|
|
mov.f32 %f409, 0f00000000;
|
|
mul.rn.f32 %f875, %f881, %f409;
|
|
|
|
BB0_31:
|
|
mul.f32 %f410, %f875, 0f3F22F983;
|
|
cvt.rni.s32.f32 %r351, %f410;
|
|
cvt.rn.f32.s32 %f411, %r351;
|
|
neg.f32 %f412, %f411;
|
|
mov.f32 %f413, 0f3FC90FDA;
|
|
fma.rn.f32 %f414, %f412, %f413, %f875;
|
|
mov.f32 %f415, 0f33A22168;
|
|
fma.rn.f32 %f416, %f412, %f415, %f414;
|
|
mov.f32 %f417, 0f27C234C5;
|
|
fma.rn.f32 %f876, %f412, %f417, %f416;
|
|
abs.f32 %f418, %f875;
|
|
setp.leu.f32 %p41, %f418, 0f47CE4780;
|
|
@%p41 bra BB0_42;
|
|
|
|
mov.b32 %r21, %f875;
|
|
shr.u32 %r22, %r21, 23;
|
|
shl.b32 %r167, %r21, 8;
|
|
or.b32 %r23, %r167, -2147483648;
|
|
mov.u32 %r343, 0;
|
|
mov.u64 %rd152, __cudart_i2opi_f;
|
|
mov.u32 %r342, -6;
|
|
mov.u64 %rd153, %rd3;
|
|
|
|
BB0_33:
|
|
.pragma "nounroll";
|
|
ld.const.u32 %r170, [%rd152];
|
|
// inline asm
|
|
{
|
|
mad.lo.cc.u32 %r168, %r170, %r23, %r343;
|
|
madc.hi.u32 %r343, %r170, %r23, 0;
|
|
}
|
|
// inline asm
|
|
st.local.u32 [%rd153], %r168;
|
|
add.s64 %rd153, %rd153, 4;
|
|
add.s64 %rd152, %rd152, 4;
|
|
add.s32 %r342, %r342, 1;
|
|
setp.ne.s32 %p42, %r342, 0;
|
|
@%p42 bra BB0_33;
|
|
|
|
and.b32 %r173, %r22, 255;
|
|
add.s32 %r174, %r173, -128;
|
|
shr.u32 %r175, %r174, 5;
|
|
and.b32 %r28, %r21, -2147483648;
|
|
st.local.u32 [%rd3+24], %r343;
|
|
mov.u32 %r176, 6;
|
|
sub.s32 %r177, %r176, %r175;
|
|
mul.wide.s32 %rd76, %r177, 4;
|
|
add.s64 %rd9, %rd3, %rd76;
|
|
ld.local.u32 %r344, [%rd9];
|
|
ld.local.u32 %r345, [%rd9+-4];
|
|
and.b32 %r31, %r22, 31;
|
|
setp.eq.s32 %p43, %r31, 0;
|
|
@%p43 bra BB0_36;
|
|
|
|
mov.u32 %r178, 32;
|
|
sub.s32 %r179, %r178, %r31;
|
|
shr.u32 %r180, %r345, %r179;
|
|
shl.b32 %r181, %r344, %r31;
|
|
add.s32 %r344, %r180, %r181;
|
|
ld.local.u32 %r182, [%rd9+-8];
|
|
shr.u32 %r183, %r182, %r179;
|
|
shl.b32 %r184, %r345, %r31;
|
|
add.s32 %r345, %r183, %r184;
|
|
|
|
BB0_36:
|
|
shr.u32 %r185, %r345, 30;
|
|
shl.b32 %r186, %r344, 2;
|
|
add.s32 %r346, %r185, %r186;
|
|
shl.b32 %r37, %r345, 2;
|
|
shr.u32 %r187, %r346, 31;
|
|
shr.u32 %r188, %r344, 30;
|
|
add.s32 %r38, %r187, %r188;
|
|
setp.eq.s32 %p44, %r187, 0;
|
|
@%p44 bra BB0_37;
|
|
bra.uni BB0_38;
|
|
|
|
BB0_37:
|
|
mov.u32 %r347, %r28;
|
|
mov.u32 %r348, %r37;
|
|
bra.uni BB0_39;
|
|
|
|
BB0_38:
|
|
not.b32 %r189, %r346;
|
|
neg.s32 %r348, %r37;
|
|
setp.eq.s32 %p45, %r37, 0;
|
|
selp.u32 %r190, 1, 0, %p45;
|
|
add.s32 %r346, %r190, %r189;
|
|
xor.b32 %r347, %r28, -2147483648;
|
|
|
|
BB0_39:
|
|
clz.b32 %r350, %r346;
|
|
setp.eq.s32 %p46, %r350, 0;
|
|
shl.b32 %r191, %r346, %r350;
|
|
mov.u32 %r192, 32;
|
|
sub.s32 %r193, %r192, %r350;
|
|
shr.u32 %r194, %r348, %r193;
|
|
add.s32 %r195, %r194, %r191;
|
|
selp.b32 %r46, %r346, %r195, %p46;
|
|
mov.u32 %r196, -921707870;
|
|
mul.hi.u32 %r349, %r46, %r196;
|
|
setp.eq.s32 %p47, %r28, 0;
|
|
neg.s32 %r197, %r38;
|
|
selp.b32 %r351, %r38, %r197, %p47;
|
|
setp.lt.s32 %p48, %r349, 1;
|
|
@%p48 bra BB0_41;
|
|
|
|
mul.lo.s32 %r198, %r46, -921707870;
|
|
shr.u32 %r199, %r198, 31;
|
|
shl.b32 %r200, %r349, 1;
|
|
add.s32 %r349, %r199, %r200;
|
|
add.s32 %r350, %r350, 1;
|
|
|
|
BB0_41:
|
|
mov.u32 %r201, 126;
|
|
sub.s32 %r202, %r201, %r350;
|
|
shl.b32 %r203, %r202, 23;
|
|
add.s32 %r204, %r349, 1;
|
|
shr.u32 %r205, %r204, 7;
|
|
add.s32 %r206, %r205, 1;
|
|
shr.u32 %r207, %r206, 1;
|
|
add.s32 %r208, %r207, %r203;
|
|
or.b32 %r209, %r208, %r347;
|
|
mov.b32 %f876, %r209;
|
|
|
|
BB0_42:
|
|
mul.rn.f32 %f91, %f876, %f876;
|
|
add.s32 %r54, %r351, 1;
|
|
and.b32 %r55, %r54, 1;
|
|
setp.eq.s32 %p49, %r55, 0;
|
|
@%p49 bra BB0_44;
|
|
bra.uni BB0_43;
|
|
|
|
BB0_44:
|
|
mov.f32 %f421, 0f3C08839E;
|
|
mov.f32 %f422, 0fB94CA1F9;
|
|
fma.rn.f32 %f877, %f422, %f91, %f421;
|
|
bra.uni BB0_45;
|
|
|
|
BB0_43:
|
|
mov.f32 %f419, 0fBAB6061A;
|
|
mov.f32 %f420, 0f37CCF5CE;
|
|
fma.rn.f32 %f877, %f420, %f91, %f419;
|
|
|
|
BB0_45:
|
|
@%p49 bra BB0_47;
|
|
bra.uni BB0_46;
|
|
|
|
BB0_47:
|
|
mov.f32 %f426, 0fBE2AAAA3;
|
|
fma.rn.f32 %f427, %f877, %f91, %f426;
|
|
mov.f32 %f428, 0f00000000;
|
|
fma.rn.f32 %f878, %f427, %f91, %f428;
|
|
bra.uni BB0_48;
|
|
|
|
BB0_46:
|
|
mov.f32 %f423, 0f3D2AAAA5;
|
|
fma.rn.f32 %f424, %f877, %f91, %f423;
|
|
mov.f32 %f425, 0fBF000000;
|
|
fma.rn.f32 %f878, %f424, %f91, %f425;
|
|
|
|
BB0_48:
|
|
fma.rn.f32 %f879, %f878, %f876, %f876;
|
|
@%p49 bra BB0_50;
|
|
|
|
fma.rn.f32 %f879, %f878, %f91, %f370;
|
|
|
|
BB0_50:
|
|
and.b32 %r210, %r54, 2;
|
|
setp.eq.s32 %p52, %r210, 0;
|
|
@%p52 bra BB0_52;
|
|
|
|
mov.f32 %f430, 0f00000000;
|
|
mov.f32 %f431, 0fBF800000;
|
|
fma.rn.f32 %f879, %f879, %f431, %f430;
|
|
|
|
BB0_52:
|
|
@%p40 bra BB0_54;
|
|
|
|
mov.f32 %f432, 0f00000000;
|
|
mul.rn.f32 %f881, %f881, %f432;
|
|
|
|
BB0_54:
|
|
mul.f32 %f433, %f881, 0f3F22F983;
|
|
cvt.rni.s32.f32 %r361, %f433;
|
|
cvt.rn.f32.s32 %f434, %r361;
|
|
neg.f32 %f435, %f434;
|
|
fma.rn.f32 %f437, %f435, %f413, %f881;
|
|
fma.rn.f32 %f439, %f435, %f415, %f437;
|
|
fma.rn.f32 %f882, %f435, %f417, %f439;
|
|
abs.f32 %f441, %f881;
|
|
setp.leu.f32 %p54, %f441, 0f47CE4780;
|
|
@%p54 bra BB0_65;
|
|
|
|
mov.b32 %r57, %f881;
|
|
shr.u32 %r58, %r57, 23;
|
|
shl.b32 %r213, %r57, 8;
|
|
or.b32 %r59, %r213, -2147483648;
|
|
mov.u32 %r353, 0;
|
|
mov.u64 %rd154, __cudart_i2opi_f;
|
|
mov.u32 %r352, -6;
|
|
mov.u64 %rd155, %rd3;
|
|
|
|
BB0_56:
|
|
.pragma "nounroll";
|
|
ld.const.u32 %r216, [%rd154];
|
|
// inline asm
|
|
{
|
|
mad.lo.cc.u32 %r214, %r216, %r59, %r353;
|
|
madc.hi.u32 %r353, %r216, %r59, 0;
|
|
}
|
|
// inline asm
|
|
st.local.u32 [%rd155], %r214;
|
|
add.s64 %rd155, %rd155, 4;
|
|
add.s64 %rd154, %rd154, 4;
|
|
add.s32 %r352, %r352, 1;
|
|
setp.ne.s32 %p55, %r352, 0;
|
|
@%p55 bra BB0_56;
|
|
|
|
and.b32 %r219, %r58, 255;
|
|
add.s32 %r220, %r219, -128;
|
|
shr.u32 %r221, %r220, 5;
|
|
and.b32 %r64, %r57, -2147483648;
|
|
st.local.u32 [%rd3+24], %r353;
|
|
mov.u32 %r222, 6;
|
|
sub.s32 %r223, %r222, %r221;
|
|
mul.wide.s32 %rd78, %r223, 4;
|
|
add.s64 %rd14, %rd3, %rd78;
|
|
ld.local.u32 %r354, [%rd14];
|
|
ld.local.u32 %r355, [%rd14+-4];
|
|
and.b32 %r67, %r58, 31;
|
|
setp.eq.s32 %p56, %r67, 0;
|
|
@%p56 bra BB0_59;
|
|
|
|
mov.u32 %r224, 32;
|
|
sub.s32 %r225, %r224, %r67;
|
|
shr.u32 %r226, %r355, %r225;
|
|
shl.b32 %r227, %r354, %r67;
|
|
add.s32 %r354, %r226, %r227;
|
|
ld.local.u32 %r228, [%rd14+-8];
|
|
shr.u32 %r229, %r228, %r225;
|
|
shl.b32 %r230, %r355, %r67;
|
|
add.s32 %r355, %r229, %r230;
|
|
|
|
BB0_59:
|
|
shr.u32 %r231, %r355, 30;
|
|
shl.b32 %r232, %r354, 2;
|
|
add.s32 %r356, %r231, %r232;
|
|
shl.b32 %r73, %r355, 2;
|
|
shr.u32 %r233, %r356, 31;
|
|
shr.u32 %r234, %r354, 30;
|
|
add.s32 %r74, %r233, %r234;
|
|
setp.eq.s32 %p57, %r233, 0;
|
|
@%p57 bra BB0_60;
|
|
bra.uni BB0_61;
|
|
|
|
BB0_60:
|
|
mov.u32 %r357, %r64;
|
|
mov.u32 %r358, %r73;
|
|
bra.uni BB0_62;
|
|
|
|
BB0_61:
|
|
not.b32 %r235, %r356;
|
|
neg.s32 %r358, %r73;
|
|
setp.eq.s32 %p58, %r73, 0;
|
|
selp.u32 %r236, 1, 0, %p58;
|
|
add.s32 %r356, %r236, %r235;
|
|
xor.b32 %r357, %r64, -2147483648;
|
|
|
|
BB0_62:
|
|
clz.b32 %r360, %r356;
|
|
setp.eq.s32 %p59, %r360, 0;
|
|
shl.b32 %r237, %r356, %r360;
|
|
mov.u32 %r238, 32;
|
|
sub.s32 %r239, %r238, %r360;
|
|
shr.u32 %r240, %r358, %r239;
|
|
add.s32 %r241, %r240, %r237;
|
|
selp.b32 %r82, %r356, %r241, %p59;
|
|
mov.u32 %r242, -921707870;
|
|
mul.hi.u32 %r359, %r82, %r242;
|
|
setp.eq.s32 %p60, %r64, 0;
|
|
neg.s32 %r243, %r74;
|
|
selp.b32 %r361, %r74, %r243, %p60;
|
|
setp.lt.s32 %p61, %r359, 1;
|
|
@%p61 bra BB0_64;
|
|
|
|
mul.lo.s32 %r244, %r82, -921707870;
|
|
shr.u32 %r245, %r244, 31;
|
|
shl.b32 %r246, %r359, 1;
|
|
add.s32 %r359, %r245, %r246;
|
|
add.s32 %r360, %r360, 1;
|
|
|
|
BB0_64:
|
|
mov.u32 %r247, 126;
|
|
sub.s32 %r248, %r247, %r360;
|
|
shl.b32 %r249, %r248, 23;
|
|
add.s32 %r250, %r359, 1;
|
|
shr.u32 %r251, %r250, 7;
|
|
add.s32 %r252, %r251, 1;
|
|
shr.u32 %r253, %r252, 1;
|
|
add.s32 %r254, %r253, %r249;
|
|
or.b32 %r255, %r254, %r357;
|
|
mov.b32 %f882, %r255;
|
|
|
|
BB0_65:
|
|
mul.rn.f32 %f108, %f882, %f882;
|
|
and.b32 %r90, %r361, 1;
|
|
setp.eq.s32 %p62, %r90, 0;
|
|
@%p62 bra BB0_67;
|
|
bra.uni BB0_66;
|
|
|
|
BB0_67:
|
|
mov.f32 %f444, 0f3C08839E;
|
|
mov.f32 %f445, 0fB94CA1F9;
|
|
fma.rn.f32 %f883, %f445, %f108, %f444;
|
|
bra.uni BB0_68;
|
|
|
|
BB0_66:
|
|
mov.f32 %f442, 0fBAB6061A;
|
|
mov.f32 %f443, 0f37CCF5CE;
|
|
fma.rn.f32 %f883, %f443, %f108, %f442;
|
|
|
|
BB0_68:
|
|
@%p62 bra BB0_70;
|
|
bra.uni BB0_69;
|
|
|
|
BB0_70:
|
|
mov.f32 %f449, 0fBE2AAAA3;
|
|
fma.rn.f32 %f450, %f883, %f108, %f449;
|
|
mov.f32 %f451, 0f00000000;
|
|
fma.rn.f32 %f884, %f450, %f108, %f451;
|
|
bra.uni BB0_71;
|
|
|
|
BB0_69:
|
|
mov.f32 %f446, 0f3D2AAAA5;
|
|
fma.rn.f32 %f447, %f883, %f108, %f446;
|
|
mov.f32 %f448, 0fBF000000;
|
|
fma.rn.f32 %f884, %f447, %f108, %f448;
|
|
|
|
BB0_71:
|
|
fma.rn.f32 %f885, %f884, %f882, %f882;
|
|
@%p62 bra BB0_73;
|
|
|
|
fma.rn.f32 %f885, %f884, %f108, %f370;
|
|
|
|
BB0_73:
|
|
and.b32 %r256, %r361, 2;
|
|
setp.eq.s32 %p65, %r256, 0;
|
|
@%p65 bra BB0_75;
|
|
|
|
mov.f32 %f453, 0f00000000;
|
|
mov.f32 %f454, 0fBF800000;
|
|
fma.rn.f32 %f885, %f885, %f454, %f453;
|
|
|
|
BB0_75:
|
|
mul.f32 %f463, %f83, %f879;
|
|
mul.f32 %f464, %f463, %f463;
|
|
sub.f32 %f466, %f370, %f464;
|
|
mul.f32 %f467, %f83, %f885;
|
|
mul.f32 %f468, %f467, %f467;
|
|
sub.f32 %f469, %f466, %f468;
|
|
mov.f32 %f470, 0f00000000;
|
|
max.f32 %f471, %f470, %f469;
|
|
sqrt.rn.f32 %f472, %f471;
|
|
mul.f32 %f473, %f65, %f467;
|
|
mul.f32 %f474, %f66, %f467;
|
|
mul.f32 %f475, %f67, %f467;
|
|
fma.rn.f32 %f476, %f74, %f463, %f473;
|
|
fma.rn.f32 %f477, %f73, %f463, %f474;
|
|
fma.rn.f32 %f478, %f72, %f463, %f475;
|
|
fma.rn.f32 %f458, %f847, %f472, %f476;
|
|
fma.rn.f32 %f459, %f848, %f472, %f477;
|
|
fma.rn.f32 %f460, %f849, %f472, %f478;
|
|
mov.u32 %r258, 0;
|
|
st.local.u32 [%rd4+8], %r258;
|
|
st.local.u32 [%rd4+4], %r258;
|
|
st.local.u32 [%rd4], %r258;
|
|
ld.global.u32 %r257, [root];
|
|
mov.f32 %f462, 0f6C4ECB8F;
|
|
// inline asm
|
|
call _rt_trace_64, (%r257, %f12, %f13, %f14, %f458, %f459, %f460, %r258, %f79, %f462, %rd74, %r119);
|
|
// inline asm
|
|
ld.local.f32 %f479, [%rd4];
|
|
max.f32 %f480, %f479, %f470;
|
|
ld.local.f32 %f481, [%rd4+4];
|
|
max.f32 %f482, %f481, %f470;
|
|
ld.local.f32 %f483, [%rd4+8];
|
|
max.f32 %f484, %f483, %f470;
|
|
add.f32 %f874, %f874, %f480;
|
|
add.f32 %f873, %f873, %f482;
|
|
add.f32 %f872, %f872, %f484;
|
|
add.s32 %r340, %r340, 1;
|
|
setp.lt.s32 %p66, %r340, %r3;
|
|
@%p66 bra BB0_29;
|
|
|
|
add.s32 %r338, %r338, 1;
|
|
setp.lt.s32 %p67, %r338, %r3;
|
|
@%p67 bra BB0_28;
|
|
|
|
BB0_77:
|
|
mul.lo.s32 %r260, %r3, %r3;
|
|
cvt.rn.f32.s32 %f485, %r260;
|
|
rcp.rn.f32 %f486, %f485;
|
|
mul.f32 %f890, %f874, %f486;
|
|
mul.f32 %f891, %f873, %f486;
|
|
mul.f32 %f892, %f872, %f486;
|
|
div.rn.f32 %f893, %f375, %f485;
|
|
|
|
BB0_78:
|
|
setp.gt.s32 %p68, %r1, -1;
|
|
selp.f32 %f134, %f893, %f890, %p68;
|
|
ld.global.u32 %r363, [imageEnabled];
|
|
and.b32 %r261, %r363, 8;
|
|
setp.eq.s32 %p69, %r261, 0;
|
|
@%p69 bra BB0_91;
|
|
|
|
cvt.u64.u32 %rd82, %r5;
|
|
cvt.u64.u32 %rd83, %r6;
|
|
mov.u64 %rd86, image_Mask;
|
|
cvta.global.u64 %rd81, %rd86;
|
|
// inline asm
|
|
call (%rd80), _rt_buffer_get_64, (%rd81, %r103, %r103, %rd82, %rd83, %rd25, %rd25);
|
|
// inline asm
|
|
mov.f32 %f490, 0f3E68BA2E;
|
|
cvt.rzi.f32.f32 %f491, %f490;
|
|
fma.rn.f32 %f492, %f491, 0fC0000000, 0f3EE8BA2E;
|
|
abs.f32 %f135, %f492;
|
|
abs.f32 %f136, %f134;
|
|
setp.lt.f32 %p70, %f136, 0f00800000;
|
|
mul.f32 %f493, %f136, 0f4B800000;
|
|
selp.f32 %f494, 0fC3170000, 0fC2FE0000, %p70;
|
|
selp.f32 %f495, %f493, %f136, %p70;
|
|
mov.b32 %r264, %f495;
|
|
and.b32 %r265, %r264, 8388607;
|
|
or.b32 %r266, %r265, 1065353216;
|
|
mov.b32 %f496, %r266;
|
|
shr.u32 %r267, %r264, 23;
|
|
cvt.rn.f32.u32 %f497, %r267;
|
|
add.f32 %f498, %f494, %f497;
|
|
setp.gt.f32 %p71, %f496, 0f3FB504F3;
|
|
mul.f32 %f499, %f496, 0f3F000000;
|
|
add.f32 %f500, %f498, 0f3F800000;
|
|
selp.f32 %f501, %f499, %f496, %p71;
|
|
selp.f32 %f502, %f500, %f498, %p71;
|
|
add.f32 %f503, %f501, 0fBF800000;
|
|
add.f32 %f489, %f501, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f488,%f489;
|
|
// inline asm
|
|
add.f32 %f504, %f503, %f503;
|
|
mul.f32 %f505, %f488, %f504;
|
|
mul.f32 %f506, %f505, %f505;
|
|
mov.f32 %f507, 0f3C4CAF63;
|
|
mov.f32 %f508, 0f3B18F0FE;
|
|
fma.rn.f32 %f509, %f508, %f506, %f507;
|
|
mov.f32 %f510, 0f3DAAAABD;
|
|
fma.rn.f32 %f511, %f509, %f506, %f510;
|
|
mul.rn.f32 %f512, %f511, %f506;
|
|
mul.rn.f32 %f513, %f512, %f505;
|
|
sub.f32 %f514, %f503, %f505;
|
|
neg.f32 %f515, %f505;
|
|
add.f32 %f516, %f514, %f514;
|
|
fma.rn.f32 %f517, %f515, %f503, %f516;
|
|
mul.rn.f32 %f518, %f488, %f517;
|
|
add.f32 %f519, %f513, %f505;
|
|
sub.f32 %f520, %f505, %f519;
|
|
add.f32 %f521, %f513, %f520;
|
|
add.f32 %f522, %f518, %f521;
|
|
add.f32 %f523, %f519, %f522;
|
|
sub.f32 %f524, %f519, %f523;
|
|
add.f32 %f525, %f522, %f524;
|
|
mov.f32 %f526, 0f3F317200;
|
|
mul.rn.f32 %f527, %f502, %f526;
|
|
mov.f32 %f528, 0f35BFBE8E;
|
|
mul.rn.f32 %f529, %f502, %f528;
|
|
add.f32 %f530, %f527, %f523;
|
|
sub.f32 %f531, %f527, %f530;
|
|
add.f32 %f532, %f523, %f531;
|
|
add.f32 %f533, %f525, %f532;
|
|
add.f32 %f534, %f529, %f533;
|
|
add.f32 %f535, %f530, %f534;
|
|
sub.f32 %f536, %f530, %f535;
|
|
add.f32 %f537, %f534, %f536;
|
|
mov.f32 %f538, 0f3EE8BA2E;
|
|
mul.rn.f32 %f539, %f538, %f535;
|
|
neg.f32 %f540, %f539;
|
|
fma.rn.f32 %f541, %f538, %f535, %f540;
|
|
fma.rn.f32 %f542, %f538, %f537, %f541;
|
|
mov.f32 %f543, 0f00000000;
|
|
fma.rn.f32 %f544, %f543, %f535, %f542;
|
|
add.rn.f32 %f545, %f539, %f544;
|
|
neg.f32 %f546, %f545;
|
|
add.rn.f32 %f547, %f539, %f546;
|
|
add.rn.f32 %f548, %f547, %f544;
|
|
mov.b32 %r268, %f545;
|
|
setp.eq.s32 %p72, %r268, 1118925336;
|
|
add.s32 %r269, %r268, -1;
|
|
mov.b32 %f549, %r269;
|
|
add.f32 %f550, %f548, 0f37000000;
|
|
selp.f32 %f551, %f549, %f545, %p72;
|
|
selp.f32 %f137, %f550, %f548, %p72;
|
|
mul.f32 %f552, %f551, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f553, %f552;
|
|
mov.f32 %f554, 0fBF317200;
|
|
fma.rn.f32 %f555, %f553, %f554, %f551;
|
|
mov.f32 %f556, 0fB5BFBE8E;
|
|
fma.rn.f32 %f557, %f553, %f556, %f555;
|
|
mul.f32 %f558, %f557, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f559, %f558;
|
|
add.f32 %f560, %f553, 0f00000000;
|
|
ex2.approx.f32 %f561, %f560;
|
|
mul.f32 %f562, %f559, %f561;
|
|
setp.lt.f32 %p73, %f551, 0fC2D20000;
|
|
selp.f32 %f563, 0f00000000, %f562, %p73;
|
|
setp.gt.f32 %p74, %f551, 0f42D20000;
|
|
selp.f32 %f894, 0f7F800000, %f563, %p74;
|
|
setp.eq.f32 %p75, %f894, 0f7F800000;
|
|
@%p75 bra BB0_81;
|
|
|
|
fma.rn.f32 %f894, %f894, %f137, %f894;
|
|
|
|
BB0_81:
|
|
setp.lt.f32 %p76, %f134, 0f00000000;
|
|
setp.eq.f32 %p77, %f135, 0f3F800000;
|
|
and.pred %p3, %p76, %p77;
|
|
mov.b32 %r270, %f894;
|
|
xor.b32 %r271, %r270, -2147483648;
|
|
mov.b32 %f564, %r271;
|
|
selp.f32 %f896, %f564, %f894, %p3;
|
|
setp.eq.f32 %p78, %f134, 0f00000000;
|
|
@%p78 bra BB0_84;
|
|
bra.uni BB0_82;
|
|
|
|
BB0_84:
|
|
add.f32 %f567, %f134, %f134;
|
|
selp.f32 %f896, %f567, 0f00000000, %p77;
|
|
bra.uni BB0_85;
|
|
|
|
BB0_137:
|
|
mov.u64 %rd147, image_HDR;
|
|
cvta.global.u64 %rd142, %rd147;
|
|
mov.u32 %r329, 8;
|
|
// inline asm
|
|
call (%rd141), _rt_buffer_get_64, (%rd142, %r103, %r329, %rd18, %rd19, %rd25, %rd25);
|
|
// inline asm
|
|
mov.f32 %f837, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs45, %f837;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs46, 0;
|
|
st.v4.u16 [%rd141], {%rs45, %rs45, %rs45, %rs46};
|
|
bra.uni BB0_138;
|
|
|
|
BB0_82:
|
|
setp.geu.f32 %p79, %f134, 0f00000000;
|
|
@%p79 bra BB0_85;
|
|
|
|
cvt.rzi.f32.f32 %f566, %f538;
|
|
setp.neu.f32 %p80, %f566, 0f3EE8BA2E;
|
|
selp.f32 %f896, 0f7FFFFFFF, %f896, %p80;
|
|
|
|
BB0_85:
|
|
add.f32 %f568, %f136, 0f3EE8BA2E;
|
|
mov.b32 %r272, %f568;
|
|
setp.lt.s32 %p82, %r272, 2139095040;
|
|
@%p82 bra BB0_90;
|
|
|
|
setp.gtu.f32 %p83, %f136, 0f7F800000;
|
|
@%p83 bra BB0_89;
|
|
bra.uni BB0_87;
|
|
|
|
BB0_89:
|
|
add.f32 %f896, %f134, 0f3EE8BA2E;
|
|
bra.uni BB0_90;
|
|
|
|
BB0_87:
|
|
setp.neu.f32 %p84, %f136, 0f7F800000;
|
|
@%p84 bra BB0_90;
|
|
|
|
selp.f32 %f896, 0fFF800000, 0f7F800000, %p3;
|
|
|
|
BB0_90:
|
|
mul.f32 %f569, %f896, 0f437F0000;
|
|
setp.eq.f32 %p85, %f134, 0f3F800000;
|
|
selp.f32 %f570, 0f437F0000, %f569, %p85;
|
|
cvt.rzi.u32.f32 %r273, %f570;
|
|
cvt.u16.u32 %rs11, %r273;
|
|
mov.u16 %rs12, 255;
|
|
st.v2.u8 [%rd80], {%rs11, %rs12};
|
|
ld.global.u32 %r363, [imageEnabled];
|
|
|
|
BB0_91:
|
|
ld.global.f32 %f571, [lightColor];
|
|
mul.f32 %f148, %f890, %f571;
|
|
ld.global.f32 %f572, [lightColor+4];
|
|
mul.f32 %f149, %f891, %f572;
|
|
ld.global.f32 %f573, [lightColor+8];
|
|
mul.f32 %f150, %f892, %f573;
|
|
and.b32 %r274, %r363, 1;
|
|
setp.eq.b32 %p86, %r274, 1;
|
|
@!%p86 bra BB0_126;
|
|
bra.uni BB0_92;
|
|
|
|
BB0_92:
|
|
mov.f32 %f576, 0f3E666666;
|
|
cvt.rzi.f32.f32 %f577, %f576;
|
|
fma.rn.f32 %f578, %f577, 0fC0000000, 0f3EE66666;
|
|
abs.f32 %f151, %f578;
|
|
abs.f32 %f152, %f148;
|
|
setp.lt.f32 %p87, %f152, 0f00800000;
|
|
mul.f32 %f579, %f152, 0f4B800000;
|
|
selp.f32 %f580, 0fC3170000, 0fC2FE0000, %p87;
|
|
selp.f32 %f581, %f579, %f152, %p87;
|
|
mov.b32 %r275, %f581;
|
|
and.b32 %r276, %r275, 8388607;
|
|
or.b32 %r277, %r276, 1065353216;
|
|
mov.b32 %f582, %r277;
|
|
shr.u32 %r278, %r275, 23;
|
|
cvt.rn.f32.u32 %f583, %r278;
|
|
add.f32 %f584, %f580, %f583;
|
|
setp.gt.f32 %p88, %f582, 0f3FB504F3;
|
|
mul.f32 %f585, %f582, 0f3F000000;
|
|
add.f32 %f586, %f584, 0f3F800000;
|
|
selp.f32 %f587, %f585, %f582, %p88;
|
|
selp.f32 %f588, %f586, %f584, %p88;
|
|
add.f32 %f589, %f587, 0fBF800000;
|
|
add.f32 %f575, %f587, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f574,%f575;
|
|
// inline asm
|
|
add.f32 %f590, %f589, %f589;
|
|
mul.f32 %f591, %f574, %f590;
|
|
mul.f32 %f592, %f591, %f591;
|
|
mov.f32 %f593, 0f3C4CAF63;
|
|
mov.f32 %f594, 0f3B18F0FE;
|
|
fma.rn.f32 %f595, %f594, %f592, %f593;
|
|
mov.f32 %f596, 0f3DAAAABD;
|
|
fma.rn.f32 %f597, %f595, %f592, %f596;
|
|
mul.rn.f32 %f598, %f597, %f592;
|
|
mul.rn.f32 %f599, %f598, %f591;
|
|
sub.f32 %f600, %f589, %f591;
|
|
neg.f32 %f601, %f591;
|
|
add.f32 %f602, %f600, %f600;
|
|
fma.rn.f32 %f603, %f601, %f589, %f602;
|
|
mul.rn.f32 %f604, %f574, %f603;
|
|
add.f32 %f605, %f599, %f591;
|
|
sub.f32 %f606, %f591, %f605;
|
|
add.f32 %f607, %f599, %f606;
|
|
add.f32 %f608, %f604, %f607;
|
|
add.f32 %f609, %f605, %f608;
|
|
sub.f32 %f610, %f605, %f609;
|
|
add.f32 %f611, %f608, %f610;
|
|
mov.f32 %f612, 0f3F317200;
|
|
mul.rn.f32 %f613, %f588, %f612;
|
|
mov.f32 %f614, 0f35BFBE8E;
|
|
mul.rn.f32 %f615, %f588, %f614;
|
|
add.f32 %f616, %f613, %f609;
|
|
sub.f32 %f617, %f613, %f616;
|
|
add.f32 %f618, %f609, %f617;
|
|
add.f32 %f619, %f611, %f618;
|
|
add.f32 %f620, %f615, %f619;
|
|
add.f32 %f621, %f616, %f620;
|
|
sub.f32 %f622, %f616, %f621;
|
|
add.f32 %f623, %f620, %f622;
|
|
mov.f32 %f624, 0f3EE66666;
|
|
mul.rn.f32 %f625, %f624, %f621;
|
|
neg.f32 %f626, %f625;
|
|
fma.rn.f32 %f627, %f624, %f621, %f626;
|
|
fma.rn.f32 %f628, %f624, %f623, %f627;
|
|
mov.f32 %f629, 0f00000000;
|
|
fma.rn.f32 %f630, %f629, %f621, %f628;
|
|
add.rn.f32 %f631, %f625, %f630;
|
|
neg.f32 %f632, %f631;
|
|
add.rn.f32 %f633, %f625, %f632;
|
|
add.rn.f32 %f634, %f633, %f630;
|
|
mov.b32 %r279, %f631;
|
|
setp.eq.s32 %p89, %r279, 1118925336;
|
|
add.s32 %r280, %r279, -1;
|
|
mov.b32 %f635, %r280;
|
|
add.f32 %f636, %f634, 0f37000000;
|
|
selp.f32 %f637, %f635, %f631, %p89;
|
|
selp.f32 %f153, %f636, %f634, %p89;
|
|
mul.f32 %f638, %f637, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f639, %f638;
|
|
mov.f32 %f640, 0fBF317200;
|
|
fma.rn.f32 %f641, %f639, %f640, %f637;
|
|
mov.f32 %f642, 0fB5BFBE8E;
|
|
fma.rn.f32 %f643, %f639, %f642, %f641;
|
|
mul.f32 %f644, %f643, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f645, %f644;
|
|
add.f32 %f646, %f639, 0f00000000;
|
|
ex2.approx.f32 %f647, %f646;
|
|
mul.f32 %f648, %f645, %f647;
|
|
setp.lt.f32 %p90, %f637, 0fC2D20000;
|
|
selp.f32 %f649, 0f00000000, %f648, %p90;
|
|
setp.gt.f32 %p91, %f637, 0f42D20000;
|
|
selp.f32 %f897, 0f7F800000, %f649, %p91;
|
|
setp.eq.f32 %p92, %f897, 0f7F800000;
|
|
@%p92 bra BB0_94;
|
|
|
|
fma.rn.f32 %f897, %f897, %f153, %f897;
|
|
|
|
BB0_94:
|
|
setp.lt.f32 %p93, %f148, 0f00000000;
|
|
setp.eq.f32 %p94, %f151, 0f3F800000;
|
|
and.pred %p4, %p93, %p94;
|
|
mov.b32 %r281, %f897;
|
|
xor.b32 %r282, %r281, -2147483648;
|
|
mov.b32 %f650, %r282;
|
|
selp.f32 %f899, %f650, %f897, %p4;
|
|
setp.eq.f32 %p95, %f148, 0f00000000;
|
|
@%p95 bra BB0_97;
|
|
bra.uni BB0_95;
|
|
|
|
BB0_97:
|
|
add.f32 %f653, %f148, %f148;
|
|
selp.f32 %f899, %f653, 0f00000000, %p94;
|
|
bra.uni BB0_98;
|
|
|
|
BB0_95:
|
|
setp.geu.f32 %p96, %f148, 0f00000000;
|
|
@%p96 bra BB0_98;
|
|
|
|
cvt.rzi.f32.f32 %f652, %f624;
|
|
setp.neu.f32 %p97, %f652, 0f3EE66666;
|
|
selp.f32 %f899, 0f7FFFFFFF, %f899, %p97;
|
|
|
|
BB0_98:
|
|
add.f32 %f654, %f152, 0f3EE66666;
|
|
mov.b32 %r283, %f654;
|
|
setp.lt.s32 %p99, %r283, 2139095040;
|
|
@%p99 bra BB0_103;
|
|
|
|
setp.gtu.f32 %p100, %f152, 0f7F800000;
|
|
@%p100 bra BB0_102;
|
|
bra.uni BB0_100;
|
|
|
|
BB0_102:
|
|
add.f32 %f899, %f148, 0f3EE66666;
|
|
bra.uni BB0_103;
|
|
|
|
BB0_100:
|
|
setp.neu.f32 %p101, %f152, 0f7F800000;
|
|
@%p101 bra BB0_103;
|
|
|
|
selp.f32 %f899, 0fFF800000, 0f7F800000, %p4;
|
|
|
|
BB0_103:
|
|
setp.eq.f32 %p102, %f148, 0f3F800000;
|
|
selp.f32 %f164, 0f3F800000, %f899, %p102;
|
|
abs.f32 %f165, %f149;
|
|
setp.lt.f32 %p103, %f165, 0f00800000;
|
|
mul.f32 %f657, %f165, 0f4B800000;
|
|
selp.f32 %f658, 0fC3170000, 0fC2FE0000, %p103;
|
|
selp.f32 %f659, %f657, %f165, %p103;
|
|
mov.b32 %r284, %f659;
|
|
and.b32 %r285, %r284, 8388607;
|
|
or.b32 %r286, %r285, 1065353216;
|
|
mov.b32 %f660, %r286;
|
|
shr.u32 %r287, %r284, 23;
|
|
cvt.rn.f32.u32 %f661, %r287;
|
|
add.f32 %f662, %f658, %f661;
|
|
setp.gt.f32 %p104, %f660, 0f3FB504F3;
|
|
mul.f32 %f663, %f660, 0f3F000000;
|
|
add.f32 %f664, %f662, 0f3F800000;
|
|
selp.f32 %f665, %f663, %f660, %p104;
|
|
selp.f32 %f666, %f664, %f662, %p104;
|
|
add.f32 %f667, %f665, 0fBF800000;
|
|
add.f32 %f656, %f665, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f655,%f656;
|
|
// inline asm
|
|
add.f32 %f668, %f667, %f667;
|
|
mul.f32 %f669, %f655, %f668;
|
|
mul.f32 %f670, %f669, %f669;
|
|
fma.rn.f32 %f673, %f594, %f670, %f593;
|
|
fma.rn.f32 %f675, %f673, %f670, %f596;
|
|
mul.rn.f32 %f676, %f675, %f670;
|
|
mul.rn.f32 %f677, %f676, %f669;
|
|
sub.f32 %f678, %f667, %f669;
|
|
neg.f32 %f679, %f669;
|
|
add.f32 %f680, %f678, %f678;
|
|
fma.rn.f32 %f681, %f679, %f667, %f680;
|
|
mul.rn.f32 %f682, %f655, %f681;
|
|
add.f32 %f683, %f677, %f669;
|
|
sub.f32 %f684, %f669, %f683;
|
|
add.f32 %f685, %f677, %f684;
|
|
add.f32 %f686, %f682, %f685;
|
|
add.f32 %f687, %f683, %f686;
|
|
sub.f32 %f688, %f683, %f687;
|
|
add.f32 %f689, %f686, %f688;
|
|
mul.rn.f32 %f691, %f666, %f612;
|
|
mul.rn.f32 %f693, %f666, %f614;
|
|
add.f32 %f694, %f691, %f687;
|
|
sub.f32 %f695, %f691, %f694;
|
|
add.f32 %f696, %f687, %f695;
|
|
add.f32 %f697, %f689, %f696;
|
|
add.f32 %f698, %f693, %f697;
|
|
add.f32 %f699, %f694, %f698;
|
|
sub.f32 %f700, %f694, %f699;
|
|
add.f32 %f701, %f698, %f700;
|
|
mul.rn.f32 %f703, %f624, %f699;
|
|
neg.f32 %f704, %f703;
|
|
fma.rn.f32 %f705, %f624, %f699, %f704;
|
|
fma.rn.f32 %f706, %f624, %f701, %f705;
|
|
fma.rn.f32 %f708, %f629, %f699, %f706;
|
|
add.rn.f32 %f709, %f703, %f708;
|
|
neg.f32 %f710, %f709;
|
|
add.rn.f32 %f711, %f703, %f710;
|
|
add.rn.f32 %f712, %f711, %f708;
|
|
mov.b32 %r288, %f709;
|
|
setp.eq.s32 %p105, %r288, 1118925336;
|
|
add.s32 %r289, %r288, -1;
|
|
mov.b32 %f713, %r289;
|
|
add.f32 %f714, %f712, 0f37000000;
|
|
selp.f32 %f715, %f713, %f709, %p105;
|
|
selp.f32 %f166, %f714, %f712, %p105;
|
|
mul.f32 %f716, %f715, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f717, %f716;
|
|
fma.rn.f32 %f719, %f717, %f640, %f715;
|
|
fma.rn.f32 %f721, %f717, %f642, %f719;
|
|
mul.f32 %f722, %f721, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f723, %f722;
|
|
add.f32 %f724, %f717, 0f00000000;
|
|
ex2.approx.f32 %f725, %f724;
|
|
mul.f32 %f726, %f723, %f725;
|
|
setp.lt.f32 %p106, %f715, 0fC2D20000;
|
|
selp.f32 %f727, 0f00000000, %f726, %p106;
|
|
setp.gt.f32 %p107, %f715, 0f42D20000;
|
|
selp.f32 %f900, 0f7F800000, %f727, %p107;
|
|
setp.eq.f32 %p108, %f900, 0f7F800000;
|
|
@%p108 bra BB0_105;
|
|
|
|
fma.rn.f32 %f900, %f900, %f166, %f900;
|
|
|
|
BB0_105:
|
|
setp.lt.f32 %p109, %f149, 0f00000000;
|
|
and.pred %p5, %p109, %p94;
|
|
mov.b32 %r290, %f900;
|
|
xor.b32 %r291, %r290, -2147483648;
|
|
mov.b32 %f728, %r291;
|
|
selp.f32 %f902, %f728, %f900, %p5;
|
|
setp.eq.f32 %p111, %f149, 0f00000000;
|
|
@%p111 bra BB0_108;
|
|
bra.uni BB0_106;
|
|
|
|
BB0_108:
|
|
add.f32 %f731, %f149, %f149;
|
|
selp.f32 %f902, %f731, 0f00000000, %p94;
|
|
bra.uni BB0_109;
|
|
|
|
BB0_106:
|
|
setp.geu.f32 %p112, %f149, 0f00000000;
|
|
@%p112 bra BB0_109;
|
|
|
|
cvt.rzi.f32.f32 %f730, %f624;
|
|
setp.neu.f32 %p113, %f730, 0f3EE66666;
|
|
selp.f32 %f902, 0f7FFFFFFF, %f902, %p113;
|
|
|
|
BB0_109:
|
|
add.f32 %f732, %f165, 0f3EE66666;
|
|
mov.b32 %r292, %f732;
|
|
setp.lt.s32 %p115, %r292, 2139095040;
|
|
@%p115 bra BB0_114;
|
|
|
|
setp.gtu.f32 %p116, %f165, 0f7F800000;
|
|
@%p116 bra BB0_113;
|
|
bra.uni BB0_111;
|
|
|
|
BB0_113:
|
|
add.f32 %f902, %f149, 0f3EE66666;
|
|
bra.uni BB0_114;
|
|
|
|
BB0_111:
|
|
setp.neu.f32 %p117, %f165, 0f7F800000;
|
|
@%p117 bra BB0_114;
|
|
|
|
selp.f32 %f902, 0fFF800000, 0f7F800000, %p5;
|
|
|
|
BB0_114:
|
|
setp.eq.f32 %p118, %f149, 0f3F800000;
|
|
selp.f32 %f177, 0f3F800000, %f902, %p118;
|
|
abs.f32 %f178, %f150;
|
|
setp.lt.f32 %p119, %f178, 0f00800000;
|
|
mul.f32 %f735, %f178, 0f4B800000;
|
|
selp.f32 %f736, 0fC3170000, 0fC2FE0000, %p119;
|
|
selp.f32 %f737, %f735, %f178, %p119;
|
|
mov.b32 %r293, %f737;
|
|
and.b32 %r294, %r293, 8388607;
|
|
or.b32 %r295, %r294, 1065353216;
|
|
mov.b32 %f738, %r295;
|
|
shr.u32 %r296, %r293, 23;
|
|
cvt.rn.f32.u32 %f739, %r296;
|
|
add.f32 %f740, %f736, %f739;
|
|
setp.gt.f32 %p120, %f738, 0f3FB504F3;
|
|
mul.f32 %f741, %f738, 0f3F000000;
|
|
add.f32 %f742, %f740, 0f3F800000;
|
|
selp.f32 %f743, %f741, %f738, %p120;
|
|
selp.f32 %f744, %f742, %f740, %p120;
|
|
add.f32 %f745, %f743, 0fBF800000;
|
|
add.f32 %f734, %f743, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f733,%f734;
|
|
// inline asm
|
|
add.f32 %f746, %f745, %f745;
|
|
mul.f32 %f747, %f733, %f746;
|
|
mul.f32 %f748, %f747, %f747;
|
|
fma.rn.f32 %f751, %f594, %f748, %f593;
|
|
fma.rn.f32 %f753, %f751, %f748, %f596;
|
|
mul.rn.f32 %f754, %f753, %f748;
|
|
mul.rn.f32 %f755, %f754, %f747;
|
|
sub.f32 %f756, %f745, %f747;
|
|
neg.f32 %f757, %f747;
|
|
add.f32 %f758, %f756, %f756;
|
|
fma.rn.f32 %f759, %f757, %f745, %f758;
|
|
mul.rn.f32 %f760, %f733, %f759;
|
|
add.f32 %f761, %f755, %f747;
|
|
sub.f32 %f762, %f747, %f761;
|
|
add.f32 %f763, %f755, %f762;
|
|
add.f32 %f764, %f760, %f763;
|
|
add.f32 %f765, %f761, %f764;
|
|
sub.f32 %f766, %f761, %f765;
|
|
add.f32 %f767, %f764, %f766;
|
|
mul.rn.f32 %f769, %f744, %f612;
|
|
mul.rn.f32 %f771, %f744, %f614;
|
|
add.f32 %f772, %f769, %f765;
|
|
sub.f32 %f773, %f769, %f772;
|
|
add.f32 %f774, %f765, %f773;
|
|
add.f32 %f775, %f767, %f774;
|
|
add.f32 %f776, %f771, %f775;
|
|
add.f32 %f777, %f772, %f776;
|
|
sub.f32 %f778, %f772, %f777;
|
|
add.f32 %f779, %f776, %f778;
|
|
mul.rn.f32 %f781, %f624, %f777;
|
|
neg.f32 %f782, %f781;
|
|
fma.rn.f32 %f783, %f624, %f777, %f782;
|
|
fma.rn.f32 %f784, %f624, %f779, %f783;
|
|
fma.rn.f32 %f786, %f629, %f777, %f784;
|
|
add.rn.f32 %f787, %f781, %f786;
|
|
neg.f32 %f788, %f787;
|
|
add.rn.f32 %f789, %f781, %f788;
|
|
add.rn.f32 %f790, %f789, %f786;
|
|
mov.b32 %r297, %f787;
|
|
setp.eq.s32 %p121, %r297, 1118925336;
|
|
add.s32 %r298, %r297, -1;
|
|
mov.b32 %f791, %r298;
|
|
add.f32 %f792, %f790, 0f37000000;
|
|
selp.f32 %f793, %f791, %f787, %p121;
|
|
selp.f32 %f179, %f792, %f790, %p121;
|
|
mul.f32 %f794, %f793, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f795, %f794;
|
|
fma.rn.f32 %f797, %f795, %f640, %f793;
|
|
fma.rn.f32 %f799, %f795, %f642, %f797;
|
|
mul.f32 %f800, %f799, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f801, %f800;
|
|
add.f32 %f802, %f795, 0f00000000;
|
|
ex2.approx.f32 %f803, %f802;
|
|
mul.f32 %f804, %f801, %f803;
|
|
setp.lt.f32 %p122, %f793, 0fC2D20000;
|
|
selp.f32 %f805, 0f00000000, %f804, %p122;
|
|
setp.gt.f32 %p123, %f793, 0f42D20000;
|
|
selp.f32 %f903, 0f7F800000, %f805, %p123;
|
|
setp.eq.f32 %p124, %f903, 0f7F800000;
|
|
@%p124 bra BB0_116;
|
|
|
|
fma.rn.f32 %f903, %f903, %f179, %f903;
|
|
|
|
BB0_116:
|
|
setp.lt.f32 %p125, %f150, 0f00000000;
|
|
and.pred %p6, %p125, %p94;
|
|
mov.b32 %r299, %f903;
|
|
xor.b32 %r300, %r299, -2147483648;
|
|
mov.b32 %f806, %r300;
|
|
selp.f32 %f905, %f806, %f903, %p6;
|
|
setp.eq.f32 %p127, %f150, 0f00000000;
|
|
@%p127 bra BB0_119;
|
|
bra.uni BB0_117;
|
|
|
|
BB0_119:
|
|
add.f32 %f809, %f150, %f150;
|
|
selp.f32 %f905, %f809, 0f00000000, %p94;
|
|
bra.uni BB0_120;
|
|
|
|
BB0_117:
|
|
setp.geu.f32 %p128, %f150, 0f00000000;
|
|
@%p128 bra BB0_120;
|
|
|
|
cvt.rzi.f32.f32 %f808, %f624;
|
|
setp.neu.f32 %p129, %f808, 0f3EE66666;
|
|
selp.f32 %f905, 0f7FFFFFFF, %f905, %p129;
|
|
|
|
BB0_120:
|
|
add.f32 %f810, %f178, 0f3EE66666;
|
|
mov.b32 %r301, %f810;
|
|
setp.lt.s32 %p131, %r301, 2139095040;
|
|
@%p131 bra BB0_125;
|
|
|
|
setp.gtu.f32 %p132, %f178, 0f7F800000;
|
|
@%p132 bra BB0_124;
|
|
bra.uni BB0_122;
|
|
|
|
BB0_124:
|
|
add.f32 %f905, %f150, 0f3EE66666;
|
|
bra.uni BB0_125;
|
|
|
|
BB0_122:
|
|
setp.neu.f32 %p133, %f178, 0f7F800000;
|
|
@%p133 bra BB0_125;
|
|
|
|
selp.f32 %f905, 0fFF800000, 0f7F800000, %p6;
|
|
|
|
BB0_125:
|
|
setp.eq.f32 %p134, %f150, 0f3F800000;
|
|
selp.f32 %f811, 0f3F800000, %f905, %p134;
|
|
cvt.u64.u32 %rd90, %r6;
|
|
cvt.u64.u32 %rd89, %r5;
|
|
mov.u64 %rd93, image;
|
|
cvta.global.u64 %rd88, %rd93;
|
|
// inline asm
|
|
call (%rd87), _rt_buffer_get_64, (%rd88, %r103, %r104, %rd89, %rd90, %rd25, %rd25);
|
|
// inline asm
|
|
cvt.sat.f32.f32 %f812, %f811;
|
|
mul.f32 %f813, %f812, 0f437FFD71;
|
|
cvt.rzi.u32.f32 %r304, %f813;
|
|
cvt.sat.f32.f32 %f814, %f177;
|
|
mul.f32 %f815, %f814, 0f437FFD71;
|
|
cvt.rzi.u32.f32 %r305, %f815;
|
|
cvt.sat.f32.f32 %f816, %f164;
|
|
mul.f32 %f817, %f816, 0f437FFD71;
|
|
cvt.rzi.u32.f32 %r306, %f817;
|
|
cvt.u16.u32 %rs13, %r304;
|
|
cvt.u16.u32 %rs14, %r306;
|
|
cvt.u16.u32 %rs15, %r305;
|
|
mov.u16 %rs16, 255;
|
|
st.v4.u8 [%rd87], {%rs13, %rs15, %rs14, %rs16};
|
|
ld.global.u32 %r363, [imageEnabled];
|
|
|
|
BB0_126:
|
|
and.b32 %r307, %r363, 4;
|
|
setp.eq.s32 %p135, %r307, 0;
|
|
@%p135 bra BB0_138;
|
|
|
|
ld.global.u32 %r308, [additive];
|
|
setp.eq.s32 %p136, %r308, 0;
|
|
cvt.u64.u32 %rd16, %r5;
|
|
cvt.u64.u32 %rd17, %r6;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs17, %f370;}
|
|
|
|
// inline asm
|
|
@%p136 bra BB0_129;
|
|
|
|
mov.u64 %rd106, image_HDR;
|
|
cvta.global.u64 %rd95, %rd106;
|
|
mov.u32 %r312, 8;
|
|
// inline asm
|
|
call (%rd94), _rt_buffer_get_64, (%rd95, %r103, %r312, %rd16, %rd17, %rd25, %rd25);
|
|
// inline asm
|
|
ld.v4.u16 {%rs24, %rs25, %rs26, %rs27}, [%rd94];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f819, %rs24;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f820, %rs25;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f821, %rs26;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd100), _rt_buffer_get_64, (%rd95, %r103, %r312, %rd16, %rd17, %rd25, %rd25);
|
|
// inline asm
|
|
add.f32 %f822, %f148, %f819;
|
|
add.f32 %f823, %f149, %f820;
|
|
add.f32 %f824, %f150, %f821;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs23, %f824;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs22, %f823;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs21, %f822;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd100], {%rs21, %rs22, %rs23, %rs17};
|
|
bra.uni BB0_138;
|
|
|
|
BB0_129:
|
|
mov.u64 %rd113, image_HDR;
|
|
cvta.global.u64 %rd108, %rd113;
|
|
mov.u32 %r314, 8;
|
|
// inline asm
|
|
call (%rd107), _rt_buffer_get_64, (%rd108, %r103, %r314, %rd16, %rd17, %rd25, %rd25);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs30, %f150;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs29, %f149;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs28, %f148;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd107], {%rs28, %rs29, %rs30, %rs17};
|
|
|
|
BB0_138:
|
|
ret;
|
|
}
|
|
|
|
|