1941 lines
56 KiB
Plaintext
1941 lines
56 KiB
Plaintext
//
|
|
// Generated by NVIDIA NVVM Compiler
|
|
//
|
|
// Compiler Build ID: CL-23083092
|
|
// Cuda compilation tools, release 9.1, V9.1.85
|
|
// Based on LLVM 3.4svn
|
|
//
|
|
|
|
.version 6.1
|
|
.target sm_30
|
|
.address_size 64
|
|
|
|
// .globl _Z6oxMainv
|
|
.global .align 8 .b8 pixelID[8];
|
|
.global .align 8 .b8 resolution[8];
|
|
.global .align 4 .b8 normal[12];
|
|
.global .align 4 .b8 camPos[12];
|
|
.global .align 4 .b8 root[4];
|
|
.global .align 4 .u32 imageEnabled;
|
|
.global .texref lightmap;
|
|
.global .align 16 .b8 tileInfo[16];
|
|
.global .align 4 .u32 additive;
|
|
.global .align 1 .b8 image[1];
|
|
.global .align 1 .b8 image_HDR[1];
|
|
.global .align 1 .b8 image_HDR2[1];
|
|
.global .align 1 .b8 image_Mask[1];
|
|
.global .align 1 .b8 image_Dir[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<145>;
|
|
.reg .b16 %rs<56>;
|
|
.reg .f32 %f<992>;
|
|
.reg .b32 %r<373>;
|
|
.reg .b64 %rd<184>;
|
|
|
|
|
|
mov.u64 %rd183, __local_depot0;
|
|
cvta.local.u64 %SP, %rd183;
|
|
ld.global.u32 %r1, [lightMeshBufferSize];
|
|
abs.s32 %r2, %r1;
|
|
ld.global.u32 %r3, [samples];
|
|
ld.global.v2.u32 {%r105, %r106}, [pixelID];
|
|
cvt.u64.u32 %rd24, %r105;
|
|
cvt.u64.u32 %rd25, %r106;
|
|
mov.u64 %rd28, uvnormal;
|
|
cvta.global.u64 %rd23, %rd28;
|
|
mov.u32 %r103, 2;
|
|
mov.u32 %r104, 4;
|
|
mov.u64 %rd27, 0;
|
|
// inline asm
|
|
call (%rd22), _rt_buffer_get_64, (%rd23, %r103, %r104, %rd24, %rd25, %rd27, %rd27);
|
|
// inline asm
|
|
ld.u32 %r4, [%rd22];
|
|
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 %f909, 0f00000000;
|
|
mov.f32 %f910, %f909;
|
|
mov.f32 %f911, %f909;
|
|
@%p7 bra BB0_2;
|
|
|
|
ld.u8 %rs6, [%rd22+1];
|
|
and.b16 %rs8, %rs4, 255;
|
|
cvt.rn.f32.u16 %f224, %rs8;
|
|
div.rn.f32 %f225, %f224, 0f437F0000;
|
|
fma.rn.f32 %f226, %f225, 0f40000000, 0fBF800000;
|
|
cvt.rn.f32.u16 %f227, %rs6;
|
|
div.rn.f32 %f228, %f227, 0f437F0000;
|
|
fma.rn.f32 %f229, %f228, 0f40000000, 0fBF800000;
|
|
cvt.rn.f32.u16 %f230, %rs3;
|
|
div.rn.f32 %f231, %f230, 0f437F0000;
|
|
fma.rn.f32 %f232, %f231, 0f40000000, 0fBF800000;
|
|
mul.f32 %f233, %f229, %f229;
|
|
fma.rn.f32 %f234, %f226, %f226, %f233;
|
|
fma.rn.f32 %f235, %f232, %f232, %f234;
|
|
sqrt.rn.f32 %f236, %f235;
|
|
rcp.rn.f32 %f237, %f236;
|
|
mul.f32 %f909, %f226, %f237;
|
|
mul.f32 %f910, %f229, %f237;
|
|
mul.f32 %f911, %f232, %f237;
|
|
|
|
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, %f910, 0f00000000;
|
|
setp.eq.f32 %p9, %f909, 0f00000000;
|
|
and.pred %p10, %p9, %p8;
|
|
setp.eq.f32 %p11, %f911, 0f00000000;
|
|
and.pred %p12, %p10, %p11;
|
|
@%p12 bra BB0_132;
|
|
bra.uni BB0_3;
|
|
|
|
BB0_132:
|
|
ld.global.u32 %r372, [imageEnabled];
|
|
and.b32 %r320, %r372, 1;
|
|
setp.eq.b32 %p139, %r320, 1;
|
|
@!%p139 bra BB0_134;
|
|
bra.uni BB0_133;
|
|
|
|
BB0_133:
|
|
cvt.u64.u32 %rd133, %r5;
|
|
cvt.u64.u32 %rd134, %r6;
|
|
mov.u64 %rd137, image;
|
|
cvta.global.u64 %rd132, %rd137;
|
|
mov.u64 %rd136, 0;
|
|
// inline asm
|
|
call (%rd131), _rt_buffer_get_64, (%rd132, %r103, %r104, %rd133, %rd134, %rd136, %rd136);
|
|
// inline asm
|
|
mov.u16 %rs37, 0;
|
|
st.v4.u8 [%rd131], {%rs37, %rs37, %rs37, %rs37};
|
|
ld.global.u32 %r372, [imageEnabled];
|
|
|
|
BB0_134:
|
|
and.b32 %r323, %r372, 8;
|
|
setp.eq.s32 %p140, %r323, 0;
|
|
@%p140 bra BB0_136;
|
|
|
|
cvt.u64.u32 %rd140, %r5;
|
|
cvt.u64.u32 %rd141, %r6;
|
|
mov.u64 %rd144, image_Mask;
|
|
cvta.global.u64 %rd139, %rd144;
|
|
mov.u64 %rd143, 0;
|
|
// inline asm
|
|
call (%rd138), _rt_buffer_get_64, (%rd139, %r103, %r103, %rd140, %rd141, %rd143, %rd143);
|
|
// inline asm
|
|
mov.f32 %f892, 0f00000000;
|
|
cvt.rzi.u32.f32 %r326, %f892;
|
|
cvt.u16.u32 %rs38, %r326;
|
|
mov.u16 %rs39, 0;
|
|
st.v2.u8 [%rd138], {%rs38, %rs39};
|
|
ld.global.u32 %r372, [imageEnabled];
|
|
|
|
BB0_136:
|
|
and.b32 %r327, %r372, 4;
|
|
setp.eq.s32 %p141, %r327, 0;
|
|
@%p141 bra BB0_140;
|
|
|
|
ld.global.u32 %r328, [additive];
|
|
setp.eq.s32 %p142, %r328, 0;
|
|
cvt.u64.u32 %rd20, %r5;
|
|
cvt.u64.u32 %rd21, %r6;
|
|
@%p142 bra BB0_139;
|
|
|
|
mov.u64 %rd157, image_HDR;
|
|
cvta.global.u64 %rd146, %rd157;
|
|
mov.u32 %r332, 8;
|
|
mov.u64 %rd156, 0;
|
|
// inline asm
|
|
call (%rd145), _rt_buffer_get_64, (%rd146, %r103, %r332, %rd20, %rd21, %rd156, %rd156);
|
|
// inline asm
|
|
ld.v4.u16 {%rs46, %rs47, %rs48, %rs49}, [%rd145];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f893, %rs46;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f894, %rs47;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f895, %rs48;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd151), _rt_buffer_get_64, (%rd146, %r103, %r332, %rd20, %rd21, %rd156, %rd156);
|
|
// inline asm
|
|
add.f32 %f896, %f893, 0f00000000;
|
|
add.f32 %f897, %f894, 0f00000000;
|
|
add.f32 %f898, %f895, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs45, %f898;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs44, %f897;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs43, %f896;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs50, 0;
|
|
st.v4.u16 [%rd151], {%rs43, %rs44, %rs45, %rs50};
|
|
bra.uni BB0_140;
|
|
|
|
BB0_3:
|
|
ld.global.v2.u32 {%r123, %r124}, [pixelID];
|
|
cvt.u64.u32 %rd31, %r123;
|
|
cvt.u64.u32 %rd32, %r124;
|
|
mov.u64 %rd41, uvpos;
|
|
cvta.global.u64 %rd30, %rd41;
|
|
mov.u32 %r119, 12;
|
|
// inline asm
|
|
call (%rd29), _rt_buffer_get_64, (%rd30, %r103, %r119, %rd31, %rd32, %rd27, %rd27);
|
|
// inline asm
|
|
ld.f32 %f9, [%rd29+8];
|
|
ld.f32 %f8, [%rd29+4];
|
|
ld.f32 %f7, [%rd29];
|
|
mul.f32 %f245, %f7, 0f3456BF95;
|
|
mul.f32 %f246, %f8, 0f3456BF95;
|
|
mul.f32 %f247, %f9, 0f3456BF95;
|
|
abs.f32 %f10, %f909;
|
|
div.rn.f32 %f248, %f245, %f10;
|
|
abs.f32 %f249, %f910;
|
|
div.rn.f32 %f250, %f246, %f249;
|
|
abs.f32 %f11, %f911;
|
|
div.rn.f32 %f251, %f247, %f11;
|
|
abs.f32 %f252, %f248;
|
|
abs.f32 %f253, %f250;
|
|
abs.f32 %f254, %f251;
|
|
mov.f32 %f255, 0f38D1B717;
|
|
max.f32 %f256, %f252, %f255;
|
|
max.f32 %f257, %f253, %f255;
|
|
max.f32 %f258, %f254, %f255;
|
|
fma.rn.f32 %f12, %f909, %f256, %f7;
|
|
fma.rn.f32 %f13, %f910, %f257, %f8;
|
|
fma.rn.f32 %f14, %f911, %f258, %f9;
|
|
ld.global.v2.u32 {%r127, %r128}, [pixelID];
|
|
cvt.u64.u32 %rd37, %r127;
|
|
cvt.u64.u32 %rd38, %r128;
|
|
mov.u64 %rd42, rnd_seeds;
|
|
cvta.global.u64 %rd36, %rd42;
|
|
// inline asm
|
|
call (%rd35), _rt_buffer_get_64, (%rd36, %r103, %r104, %rd37, %rd38, %rd27, %rd27);
|
|
// inline asm
|
|
ld.u32 %r348, [%rd35];
|
|
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 %r341, 0;
|
|
abs.f32 %f389, %f17;
|
|
abs.f32 %f390, %f16;
|
|
max.f32 %f391, %f390, %f389;
|
|
abs.f32 %f392, %f18;
|
|
max.f32 %f393, %f391, %f392;
|
|
mov.u32 %r9, %r341;
|
|
mov.f32 %f20, %f19;
|
|
mov.f32 %f21, %f19;
|
|
mov.f32 %f22, %f19;
|
|
mov.f32 %f940, %f19;
|
|
mov.f32 %f941, %f19;
|
|
mov.f32 %f942, %f19;
|
|
|
|
BB0_6:
|
|
mul.lo.s32 %r10, %r341, 3;
|
|
cvt.s64.s32 %rd45, %r10;
|
|
mov.u64 %rd49, lightMeshBuffer;
|
|
cvta.global.u64 %rd44, %rd49;
|
|
mov.u32 %r133, 1;
|
|
mov.u64 %rd48, 0;
|
|
// inline asm
|
|
call (%rd43), _rt_buffer_get_64, (%rd44, %r133, %r119, %rd45, %rd48, %rd48, %rd48);
|
|
// inline asm
|
|
ld.f32 %f266, [%rd43];
|
|
sub.f32 %f267, %f266, %f7;
|
|
ld.f32 %f268, [%rd43+4];
|
|
sub.f32 %f269, %f268, %f8;
|
|
ld.f32 %f270, [%rd43+8];
|
|
sub.f32 %f271, %f270, %f9;
|
|
mul.f32 %f272, %f269, %f269;
|
|
fma.rn.f32 %f273, %f267, %f267, %f272;
|
|
fma.rn.f32 %f274, %f271, %f271, %f273;
|
|
sqrt.rn.f32 %f26, %f274;
|
|
rcp.rn.f32 %f275, %f26;
|
|
mul.f32 %f27, %f267, %f275;
|
|
mul.f32 %f28, %f269, %f275;
|
|
mul.f32 %f29, %f271, %f275;
|
|
mul.f32 %f276, %f910, %f28;
|
|
fma.rn.f32 %f277, %f909, %f27, %f276;
|
|
fma.rn.f32 %f30, %f911, %f29, %f277;
|
|
setp.leu.f32 %p15, %f30, 0f00000000;
|
|
@%p15 bra BB0_23;
|
|
|
|
mov.u64 %rd175, 0;
|
|
setp.eq.s32 %p16, %r3, 0;
|
|
mul.f32 %f278, %f26, %f26;
|
|
div.rn.f32 %f31, %f15, %f278;
|
|
ld.global.v2.u32 {%r137, %r138}, [pixelID];
|
|
cvt.u64.u32 %rd56, %r138;
|
|
cvt.u64.u32 %rd57, %r137;
|
|
and.b64 %rd52, %rd57, 255;
|
|
and.b64 %rd53, %rd56, 255;
|
|
mov.u64 %rd58, rnd_seeds2;
|
|
cvta.global.u64 %rd51, %rd58;
|
|
mov.u32 %r136, 3;
|
|
// inline asm
|
|
call (%rd50), _rt_buffer_get_64, (%rd51, %r103, %r136, %rd52, %rd53, %rd175, %rd175);
|
|
// inline asm
|
|
@%p16 bra BB0_9;
|
|
|
|
ld.u8 %rs10, [%rd50];
|
|
cvt.rn.f32.u16 %f279, %rs10;
|
|
div.rn.f32 %f280, %f279, 0fC37F0000;
|
|
fma.rn.f32 %f281, %f280, 0f3F333333, 0f3F800000;
|
|
setp.gt.f32 %p18, %f31, %f281;
|
|
mov.pred %p144, -1;
|
|
@%p18 bra BB0_25;
|
|
|
|
BB0_9:
|
|
ld.global.f32 %f284, [lightInvCutoff];
|
|
mul.f32 %f32, %f26, %f284;
|
|
mov.f32 %f288, 0f40800000;
|
|
abs.f32 %f34, %f32;
|
|
setp.lt.f32 %p19, %f34, 0f00800000;
|
|
mul.f32 %f290, %f34, 0f4B800000;
|
|
selp.f32 %f291, 0fC3170000, 0fC2FE0000, %p19;
|
|
selp.f32 %f292, %f290, %f34, %p19;
|
|
mov.b32 %r141, %f292;
|
|
and.b32 %r142, %r141, 8388607;
|
|
or.b32 %r143, %r142, 1065353216;
|
|
mov.b32 %f293, %r143;
|
|
shr.u32 %r144, %r141, 23;
|
|
cvt.rn.f32.u32 %f294, %r144;
|
|
add.f32 %f295, %f291, %f294;
|
|
setp.gt.f32 %p20, %f293, 0f3FB504F3;
|
|
mul.f32 %f296, %f293, 0f3F000000;
|
|
add.f32 %f297, %f295, 0f3F800000;
|
|
selp.f32 %f298, %f296, %f293, %p20;
|
|
selp.f32 %f299, %f297, %f295, %p20;
|
|
add.f32 %f300, %f298, 0fBF800000;
|
|
add.f32 %f283, %f298, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f282,%f283;
|
|
// inline asm
|
|
add.f32 %f301, %f300, %f300;
|
|
mul.f32 %f302, %f282, %f301;
|
|
mul.f32 %f303, %f302, %f302;
|
|
mov.f32 %f304, 0f3C4CAF63;
|
|
mov.f32 %f305, 0f3B18F0FE;
|
|
fma.rn.f32 %f306, %f305, %f303, %f304;
|
|
mov.f32 %f307, 0f3DAAAABD;
|
|
fma.rn.f32 %f308, %f306, %f303, %f307;
|
|
mul.rn.f32 %f309, %f308, %f303;
|
|
mul.rn.f32 %f310, %f309, %f302;
|
|
sub.f32 %f311, %f300, %f302;
|
|
neg.f32 %f312, %f302;
|
|
add.f32 %f313, %f311, %f311;
|
|
fma.rn.f32 %f314, %f312, %f300, %f313;
|
|
mul.rn.f32 %f315, %f282, %f314;
|
|
add.f32 %f316, %f310, %f302;
|
|
sub.f32 %f317, %f302, %f316;
|
|
add.f32 %f318, %f310, %f317;
|
|
add.f32 %f319, %f315, %f318;
|
|
add.f32 %f320, %f316, %f319;
|
|
sub.f32 %f321, %f316, %f320;
|
|
add.f32 %f322, %f319, %f321;
|
|
mov.f32 %f323, 0f3F317200;
|
|
mul.rn.f32 %f324, %f299, %f323;
|
|
mov.f32 %f325, 0f35BFBE8E;
|
|
mul.rn.f32 %f326, %f299, %f325;
|
|
add.f32 %f327, %f324, %f320;
|
|
sub.f32 %f328, %f324, %f327;
|
|
add.f32 %f329, %f320, %f328;
|
|
add.f32 %f330, %f322, %f329;
|
|
add.f32 %f331, %f326, %f330;
|
|
add.f32 %f332, %f327, %f331;
|
|
sub.f32 %f333, %f327, %f332;
|
|
add.f32 %f334, %f331, %f333;
|
|
mul.rn.f32 %f335, %f288, %f332;
|
|
neg.f32 %f336, %f335;
|
|
fma.rn.f32 %f337, %f288, %f332, %f336;
|
|
fma.rn.f32 %f338, %f288, %f334, %f337;
|
|
mov.f32 %f339, 0f00000000;
|
|
fma.rn.f32 %f340, %f339, %f332, %f338;
|
|
add.rn.f32 %f341, %f335, %f340;
|
|
neg.f32 %f342, %f341;
|
|
add.rn.f32 %f343, %f335, %f342;
|
|
add.rn.f32 %f344, %f343, %f340;
|
|
mov.b32 %r145, %f341;
|
|
setp.eq.s32 %p21, %r145, 1118925336;
|
|
add.s32 %r146, %r145, -1;
|
|
mov.b32 %f345, %r146;
|
|
add.f32 %f346, %f344, 0f37000000;
|
|
selp.f32 %f347, %f345, %f341, %p21;
|
|
selp.f32 %f35, %f346, %f344, %p21;
|
|
mul.f32 %f348, %f347, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f349, %f348;
|
|
mov.f32 %f350, 0fBF317200;
|
|
fma.rn.f32 %f351, %f349, %f350, %f347;
|
|
mov.f32 %f352, 0fB5BFBE8E;
|
|
fma.rn.f32 %f353, %f349, %f352, %f351;
|
|
mul.f32 %f354, %f353, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f355, %f354;
|
|
add.f32 %f356, %f349, 0f00000000;
|
|
ex2.approx.f32 %f357, %f356;
|
|
mul.f32 %f358, %f355, %f357;
|
|
setp.lt.f32 %p22, %f347, 0fC2D20000;
|
|
selp.f32 %f359, 0f00000000, %f358, %p22;
|
|
setp.gt.f32 %p23, %f347, 0f42D20000;
|
|
selp.f32 %f919, 0f7F800000, %f359, %p23;
|
|
setp.eq.f32 %p24, %f919, 0f7F800000;
|
|
@%p24 bra BB0_11;
|
|
|
|
fma.rn.f32 %f919, %f919, %f35, %f919;
|
|
|
|
BB0_11:
|
|
mov.f32 %f905, 0f40000000;
|
|
cvt.rzi.f32.f32 %f904, %f905;
|
|
add.f32 %f903, %f904, %f904;
|
|
mov.f32 %f902, 0f40800000;
|
|
sub.f32 %f901, %f902, %f903;
|
|
abs.f32 %f900, %f901;
|
|
setp.lt.f32 %p25, %f32, 0f00000000;
|
|
setp.eq.f32 %p26, %f900, 0f3F800000;
|
|
and.pred %p1, %p25, %p26;
|
|
mov.b32 %r147, %f919;
|
|
xor.b32 %r148, %r147, -2147483648;
|
|
mov.b32 %f360, %r148;
|
|
selp.f32 %f921, %f360, %f919, %p1;
|
|
setp.eq.f32 %p27, %f32, 0f00000000;
|
|
@%p27 bra BB0_14;
|
|
bra.uni BB0_12;
|
|
|
|
BB0_14:
|
|
add.f32 %f363, %f32, %f32;
|
|
selp.f32 %f921, %f363, 0f00000000, %p26;
|
|
bra.uni BB0_15;
|
|
|
|
BB0_12:
|
|
setp.geu.f32 %p28, %f32, 0f00000000;
|
|
@%p28 bra BB0_15;
|
|
|
|
mov.f32 %f908, 0f40800000;
|
|
cvt.rzi.f32.f32 %f362, %f908;
|
|
setp.neu.f32 %p29, %f362, 0f40800000;
|
|
selp.f32 %f921, 0f7FFFFFFF, %f921, %p29;
|
|
|
|
BB0_15:
|
|
add.f32 %f364, %f34, 0f40800000;
|
|
mov.b32 %r149, %f364;
|
|
setp.lt.s32 %p31, %r149, 2139095040;
|
|
@%p31 bra BB0_20;
|
|
|
|
setp.gtu.f32 %p32, %f34, 0f7F800000;
|
|
@%p32 bra BB0_19;
|
|
bra.uni BB0_17;
|
|
|
|
BB0_19:
|
|
add.f32 %f921, %f32, 0f40800000;
|
|
bra.uni BB0_20;
|
|
|
|
BB0_17:
|
|
setp.neu.f32 %p33, %f34, 0f7F800000;
|
|
@%p33 bra BB0_20;
|
|
|
|
selp.f32 %f921, 0fFF800000, 0f7F800000, %p1;
|
|
|
|
BB0_20:
|
|
mov.u32 %r338, 1;
|
|
mov.u64 %rd174, lightMeshBuffer;
|
|
cvta.global.u64 %rd173, %rd174;
|
|
mul.lo.s32 %r337, %r341, 3;
|
|
mov.u64 %rd172, 0;
|
|
mov.f32 %f365, 0f3F800000;
|
|
sub.f32 %f366, %f365, %f921;
|
|
setp.eq.f32 %p34, %f32, 0f3F800000;
|
|
selp.f32 %f367, 0f00000000, %f366, %p34;
|
|
cvt.sat.f32.f32 %f368, %f367;
|
|
mul.f32 %f369, %f31, %f368;
|
|
add.s32 %r152, %r337, 1;
|
|
cvt.s64.s32 %rd61, %r152;
|
|
// inline asm
|
|
call (%rd59), _rt_buffer_get_64, (%rd173, %r338, %r119, %rd61, %rd172, %rd172, %rd172);
|
|
// inline asm
|
|
ld.f32 %f370, [%rd59];
|
|
mul.f32 %f371, %f27, %f370;
|
|
ld.f32 %f372, [%rd59+4];
|
|
mul.f32 %f373, %f28, %f372;
|
|
neg.f32 %f374, %f373;
|
|
sub.f32 %f375, %f374, %f371;
|
|
ld.f32 %f376, [%rd59+8];
|
|
mul.f32 %f377, %f29, %f376;
|
|
sub.f32 %f378, %f375, %f377;
|
|
cvt.sat.f32.f32 %f379, %f378;
|
|
mul.f32 %f46, %f369, %f379;
|
|
cvt.sat.f32.f32 %f47, %f30;
|
|
mul.f32 %f380, %f46, %f47;
|
|
setp.leu.f32 %p35, %f380, 0f3727C5AC;
|
|
@%p35 bra BB0_22;
|
|
|
|
mov.u32 %r340, 1;
|
|
mov.u64 %rd178, lightMeshBuffer;
|
|
cvta.global.u64 %rd177, %rd178;
|
|
mul.lo.s32 %r339, %r341, 3;
|
|
mov.u64 %rd176, 0;
|
|
add.u64 %rd66, %SP, 28;
|
|
cvta.to.local.u64 %rd73, %rd66;
|
|
max.f32 %f387, %f393, %f255;
|
|
sub.f32 %f388, %f26, %f387;
|
|
mov.u32 %r158, 1065353216;
|
|
st.local.u32 [%rd73], %r158;
|
|
ld.global.u32 %r153, [root];
|
|
// inline asm
|
|
call _rt_trace_64, (%r153, %f12, %f13, %f14, %f27, %f28, %f29, %r340, %f387, %f388, %rd66, %r104);
|
|
// inline asm
|
|
add.s32 %r159, %r339, 2;
|
|
cvt.s64.s32 %rd69, %r159;
|
|
// inline asm
|
|
call (%rd67), _rt_buffer_get_64, (%rd177, %r340, %r119, %rd69, %rd176, %rd176, %rd176);
|
|
// inline asm
|
|
ld.f32 %f395, [%rd67];
|
|
ld.local.f32 %f396, [%rd73];
|
|
mul.f32 %f397, %f396, %f395;
|
|
ld.f32 %f398, [%rd67+4];
|
|
mul.f32 %f399, %f396, %f398;
|
|
ld.f32 %f400, [%rd67+8];
|
|
mul.f32 %f401, %f396, %f400;
|
|
mul.f32 %f402, %f46, 0f3EA2F983;
|
|
mul.f32 %f403, %f402, %f397;
|
|
mul.f32 %f404, %f402, %f399;
|
|
mul.f32 %f405, %f402, %f401;
|
|
fma.rn.f32 %f22, %f47, %f403, %f22;
|
|
fma.rn.f32 %f21, %f47, %f404, %f21;
|
|
fma.rn.f32 %f20, %f47, %f405, %f20;
|
|
add.f32 %f19, %f19, %f396;
|
|
fma.rn.f32 %f942, %f27, %f396, %f942;
|
|
fma.rn.f32 %f941, %f28, %f396, %f941;
|
|
fma.rn.f32 %f940, %f29, %f396, %f940;
|
|
|
|
BB0_22:
|
|
add.s32 %r9, %r9, 1;
|
|
|
|
BB0_23:
|
|
add.s32 %r341, %r341, 1;
|
|
setp.lt.s32 %p37, %r341, %r2;
|
|
@%p37 bra BB0_6;
|
|
bra.uni BB0_24;
|
|
|
|
BB0_4:
|
|
mov.f32 %f20, %f19;
|
|
mov.f32 %f21, %f19;
|
|
mov.f32 %f22, %f19;
|
|
mov.f32 %f940, %f19;
|
|
mov.f32 %f941, %f19;
|
|
mov.f32 %f942, %f19;
|
|
|
|
BB0_24:
|
|
mov.pred %p144, %p13;
|
|
|
|
BB0_25:
|
|
cvt.rn.f32.s32 %f406, %r9;
|
|
mov.f32 %f407, 0f3F800000;
|
|
max.f32 %f408, %f406, %f407;
|
|
rcp.rn.f32 %f409, %f408;
|
|
mul.f32 %f976, %f22, %f409;
|
|
mul.f32 %f977, %f21, %f409;
|
|
mul.f32 %f978, %f20, %f409;
|
|
div.rn.f32 %f979, %f19, %f408;
|
|
@!%p144 bra BB0_78;
|
|
bra.uni BB0_26;
|
|
|
|
BB0_26:
|
|
abs.f32 %f907, %f911;
|
|
abs.f32 %f906, %f909;
|
|
setp.gt.f32 %p38, %f906, %f907;
|
|
neg.f32 %f416, %f910;
|
|
selp.f32 %f417, %f416, 0f00000000, %p38;
|
|
neg.f32 %f418, %f911;
|
|
selp.f32 %f419, %f909, %f418, %p38;
|
|
selp.f32 %f420, 0f00000000, %f910, %p38;
|
|
mul.f32 %f421, %f419, %f419;
|
|
fma.rn.f32 %f422, %f417, %f417, %f421;
|
|
fma.rn.f32 %f423, %f420, %f420, %f422;
|
|
sqrt.rn.f32 %f424, %f423;
|
|
rcp.rn.f32 %f425, %f424;
|
|
mul.f32 %f80, %f417, %f425;
|
|
mul.f32 %f81, %f419, %f425;
|
|
mul.f32 %f82, %f420, %f425;
|
|
mov.f32 %f415, 0f00000000;
|
|
setp.lt.s32 %p39, %r3, 1;
|
|
mov.f32 %f949, %f415;
|
|
mov.f32 %f950, %f415;
|
|
mov.f32 %f951, %f415;
|
|
mov.f32 %f940, %f415;
|
|
mov.f32 %f941, %f415;
|
|
mov.f32 %f942, %f415;
|
|
@%p39 bra BB0_77;
|
|
|
|
cvt.rn.f32.s32 %f432, %r3;
|
|
rcp.rn.f32 %f83, %f432;
|
|
add.u64 %rd75, %SP, 0;
|
|
cvta.to.local.u64 %rd3, %rd75;
|
|
mul.f32 %f84, %f12, 0f3456BF95;
|
|
mul.f32 %f85, %f13, 0f3456BF95;
|
|
mul.f32 %f86, %f14, 0f3456BF95;
|
|
add.u64 %rd76, %SP, 32;
|
|
cvta.to.local.u64 %rd4, %rd76;
|
|
mul.f32 %f433, %f909, %f81;
|
|
mul.f32 %f434, %f910, %f80;
|
|
sub.f32 %f87, %f434, %f433;
|
|
mul.f32 %f435, %f911, %f80;
|
|
mul.f32 %f436, %f909, %f82;
|
|
sub.f32 %f88, %f436, %f435;
|
|
mul.f32 %f437, %f910, %f82;
|
|
mul.f32 %f438, %f911, %f81;
|
|
sub.f32 %f89, %f438, %f437;
|
|
mov.f32 %f949, 0f00000000;
|
|
mov.u32 %r160, 0;
|
|
abs.f32 %f439, %f85;
|
|
abs.f32 %f440, %f84;
|
|
max.f32 %f441, %f440, %f439;
|
|
abs.f32 %f442, %f86;
|
|
max.f32 %f443, %f441, %f442;
|
|
mov.u32 %r345, %r160;
|
|
mov.f32 %f950, %f949;
|
|
mov.f32 %f951, %f949;
|
|
mov.f32 %f940, %f949;
|
|
mov.f32 %f941, %f949;
|
|
mov.f32 %f942, %f949;
|
|
|
|
BB0_28:
|
|
cvt.rn.f32.s32 %f96, %r345;
|
|
max.f32 %f97, %f443, %f255;
|
|
mov.u32 %r347, %r160;
|
|
|
|
BB0_29:
|
|
mad.lo.s32 %r162, %r348, 1664525, 1013904223;
|
|
and.b32 %r163, %r162, 16777215;
|
|
cvt.rn.f32.u32 %f445, %r163;
|
|
fma.rn.f32 %f446, %f445, 0f33800000, %f96;
|
|
mul.f32 %f104, %f83, %f446;
|
|
mad.lo.s32 %r348, %r162, 1664525, 1013904223;
|
|
and.b32 %r164, %r348, 16777215;
|
|
cvt.rn.f32.u32 %f447, %r164;
|
|
cvt.rn.f32.s32 %f448, %r347;
|
|
fma.rn.f32 %f449, %f447, 0f33800000, %f448;
|
|
mul.f32 %f450, %f83, %f449;
|
|
mul.f32 %f451, %f104, %f104;
|
|
sub.f32 %f453, %f407, %f451;
|
|
mov.f32 %f454, 0f00000000;
|
|
max.f32 %f455, %f454, %f453;
|
|
sqrt.rn.f32 %f105, %f455;
|
|
mul.f32 %f961, %f450, 0f40C90FDB;
|
|
abs.f32 %f107, %f961;
|
|
setp.neu.f32 %p40, %f107, 0f7F800000;
|
|
mov.f32 %f955, %f961;
|
|
@%p40 bra BB0_31;
|
|
|
|
mul.rn.f32 %f955, %f961, %f454;
|
|
|
|
BB0_31:
|
|
mul.f32 %f457, %f955, 0f3F22F983;
|
|
cvt.rni.s32.f32 %r358, %f457;
|
|
cvt.rn.f32.s32 %f458, %r358;
|
|
neg.f32 %f459, %f458;
|
|
mov.f32 %f460, 0f3FC90FDA;
|
|
fma.rn.f32 %f461, %f459, %f460, %f955;
|
|
mov.f32 %f462, 0f33A22168;
|
|
fma.rn.f32 %f463, %f459, %f462, %f461;
|
|
mov.f32 %f464, 0f27C234C5;
|
|
fma.rn.f32 %f956, %f459, %f464, %f463;
|
|
abs.f32 %f465, %f955;
|
|
setp.leu.f32 %p41, %f465, 0f47CE4780;
|
|
@%p41 bra BB0_42;
|
|
|
|
mov.b32 %r21, %f955;
|
|
shr.u32 %r22, %r21, 23;
|
|
shl.b32 %r167, %r21, 8;
|
|
or.b32 %r23, %r167, -2147483648;
|
|
mov.u32 %r349, 0;
|
|
mov.u64 %rd180, 0;
|
|
mov.u64 %rd179, %rd3;
|
|
mov.u32 %r350, %r349;
|
|
|
|
BB0_33:
|
|
.pragma "nounroll";
|
|
shl.b64 %rd78, %rd180, 2;
|
|
mov.u64 %rd79, __cudart_i2opi_f;
|
|
add.s64 %rd80, %rd79, %rd78;
|
|
ld.const.u32 %r170, [%rd80];
|
|
// inline asm
|
|
{
|
|
mad.lo.cc.u32 %r168, %r170, %r23, %r350;
|
|
madc.hi.u32 %r350, %r170, %r23, 0;
|
|
}
|
|
// inline asm
|
|
st.local.u32 [%rd179], %r168;
|
|
add.s32 %r349, %r349, 1;
|
|
cvt.s64.s32 %rd180, %r349;
|
|
mul.wide.s32 %rd81, %r349, 4;
|
|
add.s64 %rd179, %rd3, %rd81;
|
|
setp.ne.s32 %p42, %r349, 6;
|
|
@%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], %r350;
|
|
mov.u32 %r176, 6;
|
|
sub.s32 %r177, %r176, %r175;
|
|
mul.wide.s32 %rd82, %r177, 4;
|
|
add.s64 %rd10, %rd3, %rd82;
|
|
ld.local.u32 %r351, [%rd10];
|
|
ld.local.u32 %r352, [%rd10+-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, %r352, %r179;
|
|
shl.b32 %r181, %r351, %r31;
|
|
add.s32 %r351, %r180, %r181;
|
|
ld.local.u32 %r182, [%rd10+-8];
|
|
shr.u32 %r183, %r182, %r179;
|
|
shl.b32 %r184, %r352, %r31;
|
|
add.s32 %r352, %r183, %r184;
|
|
|
|
BB0_36:
|
|
shr.u32 %r185, %r352, 30;
|
|
shl.b32 %r186, %r351, 2;
|
|
add.s32 %r353, %r185, %r186;
|
|
shl.b32 %r37, %r352, 2;
|
|
shr.u32 %r187, %r353, 31;
|
|
shr.u32 %r188, %r351, 30;
|
|
add.s32 %r38, %r187, %r188;
|
|
setp.eq.s32 %p44, %r187, 0;
|
|
@%p44 bra BB0_37;
|
|
bra.uni BB0_38;
|
|
|
|
BB0_37:
|
|
mov.u32 %r354, %r28;
|
|
mov.u32 %r355, %r37;
|
|
bra.uni BB0_39;
|
|
|
|
BB0_38:
|
|
not.b32 %r189, %r353;
|
|
neg.s32 %r355, %r37;
|
|
setp.eq.s32 %p45, %r37, 0;
|
|
selp.u32 %r190, 1, 0, %p45;
|
|
add.s32 %r353, %r190, %r189;
|
|
xor.b32 %r354, %r28, -2147483648;
|
|
|
|
BB0_39:
|
|
clz.b32 %r357, %r353;
|
|
setp.eq.s32 %p46, %r357, 0;
|
|
shl.b32 %r191, %r353, %r357;
|
|
mov.u32 %r192, 32;
|
|
sub.s32 %r193, %r192, %r357;
|
|
shr.u32 %r194, %r355, %r193;
|
|
add.s32 %r195, %r194, %r191;
|
|
selp.b32 %r46, %r353, %r195, %p46;
|
|
mov.u32 %r196, -921707870;
|
|
mul.hi.u32 %r356, %r46, %r196;
|
|
setp.eq.s32 %p47, %r28, 0;
|
|
neg.s32 %r197, %r38;
|
|
selp.b32 %r358, %r38, %r197, %p47;
|
|
setp.lt.s32 %p48, %r356, 1;
|
|
@%p48 bra BB0_41;
|
|
|
|
mul.lo.s32 %r198, %r46, -921707870;
|
|
shr.u32 %r199, %r198, 31;
|
|
shl.b32 %r200, %r356, 1;
|
|
add.s32 %r356, %r199, %r200;
|
|
add.s32 %r357, %r357, 1;
|
|
|
|
BB0_41:
|
|
mov.u32 %r201, 126;
|
|
sub.s32 %r202, %r201, %r357;
|
|
shl.b32 %r203, %r202, 23;
|
|
add.s32 %r204, %r356, 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, %r354;
|
|
mov.b32 %f956, %r209;
|
|
|
|
BB0_42:
|
|
mul.rn.f32 %f113, %f956, %f956;
|
|
add.s32 %r54, %r358, 1;
|
|
and.b32 %r55, %r54, 1;
|
|
setp.eq.s32 %p49, %r55, 0;
|
|
@%p49 bra BB0_44;
|
|
bra.uni BB0_43;
|
|
|
|
BB0_44:
|
|
mov.f32 %f468, 0f3C08839E;
|
|
mov.f32 %f469, 0fB94CA1F9;
|
|
fma.rn.f32 %f957, %f469, %f113, %f468;
|
|
bra.uni BB0_45;
|
|
|
|
BB0_43:
|
|
mov.f32 %f466, 0fBAB6061A;
|
|
mov.f32 %f467, 0f37CCF5CE;
|
|
fma.rn.f32 %f957, %f467, %f113, %f466;
|
|
|
|
BB0_45:
|
|
@%p49 bra BB0_47;
|
|
bra.uni BB0_46;
|
|
|
|
BB0_47:
|
|
mov.f32 %f473, 0fBE2AAAA3;
|
|
fma.rn.f32 %f474, %f957, %f113, %f473;
|
|
fma.rn.f32 %f958, %f474, %f113, %f454;
|
|
bra.uni BB0_48;
|
|
|
|
BB0_46:
|
|
mov.f32 %f470, 0f3D2AAAA5;
|
|
fma.rn.f32 %f471, %f957, %f113, %f470;
|
|
mov.f32 %f472, 0fBF000000;
|
|
fma.rn.f32 %f958, %f471, %f113, %f472;
|
|
|
|
BB0_48:
|
|
fma.rn.f32 %f959, %f958, %f956, %f956;
|
|
@%p49 bra BB0_50;
|
|
|
|
fma.rn.f32 %f959, %f958, %f113, %f407;
|
|
|
|
BB0_50:
|
|
and.b32 %r210, %r54, 2;
|
|
setp.eq.s32 %p52, %r210, 0;
|
|
@%p52 bra BB0_52;
|
|
|
|
mov.f32 %f478, 0fBF800000;
|
|
fma.rn.f32 %f959, %f959, %f478, %f454;
|
|
|
|
BB0_52:
|
|
@%p40 bra BB0_54;
|
|
|
|
mul.rn.f32 %f961, %f961, %f454;
|
|
|
|
BB0_54:
|
|
mul.f32 %f480, %f961, 0f3F22F983;
|
|
cvt.rni.s32.f32 %r368, %f480;
|
|
cvt.rn.f32.s32 %f481, %r368;
|
|
neg.f32 %f482, %f481;
|
|
fma.rn.f32 %f484, %f482, %f460, %f961;
|
|
fma.rn.f32 %f486, %f482, %f462, %f484;
|
|
fma.rn.f32 %f962, %f482, %f464, %f486;
|
|
abs.f32 %f488, %f961;
|
|
setp.leu.f32 %p54, %f488, 0f47CE4780;
|
|
@%p54 bra BB0_65;
|
|
|
|
mov.b32 %r57, %f961;
|
|
shr.u32 %r58, %r57, 23;
|
|
shl.b32 %r213, %r57, 8;
|
|
or.b32 %r59, %r213, -2147483648;
|
|
mov.u32 %r359, 0;
|
|
mov.u64 %rd181, %rd3;
|
|
mov.u64 %rd182, %rd27;
|
|
mov.u32 %r360, %r359;
|
|
|
|
BB0_56:
|
|
.pragma "nounroll";
|
|
shl.b64 %rd84, %rd182, 2;
|
|
mov.u64 %rd85, __cudart_i2opi_f;
|
|
add.s64 %rd86, %rd85, %rd84;
|
|
ld.const.u32 %r216, [%rd86];
|
|
// inline asm
|
|
{
|
|
mad.lo.cc.u32 %r214, %r216, %r59, %r360;
|
|
madc.hi.u32 %r360, %r216, %r59, 0;
|
|
}
|
|
// inline asm
|
|
st.local.u32 [%rd181], %r214;
|
|
add.s32 %r359, %r359, 1;
|
|
cvt.s64.s32 %rd182, %r359;
|
|
mul.wide.s32 %rd87, %r359, 4;
|
|
add.s64 %rd181, %rd3, %rd87;
|
|
setp.ne.s32 %p55, %r359, 6;
|
|
@%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], %r360;
|
|
mov.u32 %r222, 6;
|
|
sub.s32 %r223, %r222, %r221;
|
|
mul.wide.s32 %rd88, %r223, 4;
|
|
add.s64 %rd16, %rd3, %rd88;
|
|
ld.local.u32 %r361, [%rd16];
|
|
ld.local.u32 %r362, [%rd16+-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, %r362, %r225;
|
|
shl.b32 %r227, %r361, %r67;
|
|
add.s32 %r361, %r226, %r227;
|
|
ld.local.u32 %r228, [%rd16+-8];
|
|
shr.u32 %r229, %r228, %r225;
|
|
shl.b32 %r230, %r362, %r67;
|
|
add.s32 %r362, %r229, %r230;
|
|
|
|
BB0_59:
|
|
shr.u32 %r231, %r362, 30;
|
|
shl.b32 %r232, %r361, 2;
|
|
add.s32 %r363, %r231, %r232;
|
|
shl.b32 %r73, %r362, 2;
|
|
shr.u32 %r233, %r363, 31;
|
|
shr.u32 %r234, %r361, 30;
|
|
add.s32 %r74, %r233, %r234;
|
|
setp.eq.s32 %p57, %r233, 0;
|
|
@%p57 bra BB0_60;
|
|
bra.uni BB0_61;
|
|
|
|
BB0_60:
|
|
mov.u32 %r364, %r64;
|
|
mov.u32 %r365, %r73;
|
|
bra.uni BB0_62;
|
|
|
|
BB0_61:
|
|
not.b32 %r235, %r363;
|
|
neg.s32 %r365, %r73;
|
|
setp.eq.s32 %p58, %r73, 0;
|
|
selp.u32 %r236, 1, 0, %p58;
|
|
add.s32 %r363, %r236, %r235;
|
|
xor.b32 %r364, %r64, -2147483648;
|
|
|
|
BB0_62:
|
|
clz.b32 %r367, %r363;
|
|
setp.eq.s32 %p59, %r367, 0;
|
|
shl.b32 %r237, %r363, %r367;
|
|
mov.u32 %r238, 32;
|
|
sub.s32 %r239, %r238, %r367;
|
|
shr.u32 %r240, %r365, %r239;
|
|
add.s32 %r241, %r240, %r237;
|
|
selp.b32 %r82, %r363, %r241, %p59;
|
|
mov.u32 %r242, -921707870;
|
|
mul.hi.u32 %r366, %r82, %r242;
|
|
setp.eq.s32 %p60, %r64, 0;
|
|
neg.s32 %r243, %r74;
|
|
selp.b32 %r368, %r74, %r243, %p60;
|
|
setp.lt.s32 %p61, %r366, 1;
|
|
@%p61 bra BB0_64;
|
|
|
|
mul.lo.s32 %r244, %r82, -921707870;
|
|
shr.u32 %r245, %r244, 31;
|
|
shl.b32 %r246, %r366, 1;
|
|
add.s32 %r366, %r245, %r246;
|
|
add.s32 %r367, %r367, 1;
|
|
|
|
BB0_64:
|
|
mov.u32 %r247, 126;
|
|
sub.s32 %r248, %r247, %r367;
|
|
shl.b32 %r249, %r248, 23;
|
|
add.s32 %r250, %r366, 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, %r364;
|
|
mov.b32 %f962, %r255;
|
|
|
|
BB0_65:
|
|
mul.rn.f32 %f130, %f962, %f962;
|
|
and.b32 %r90, %r368, 1;
|
|
setp.eq.s32 %p62, %r90, 0;
|
|
@%p62 bra BB0_67;
|
|
bra.uni BB0_66;
|
|
|
|
BB0_67:
|
|
mov.f32 %f491, 0f3C08839E;
|
|
mov.f32 %f492, 0fB94CA1F9;
|
|
fma.rn.f32 %f963, %f492, %f130, %f491;
|
|
bra.uni BB0_68;
|
|
|
|
BB0_66:
|
|
mov.f32 %f489, 0fBAB6061A;
|
|
mov.f32 %f490, 0f37CCF5CE;
|
|
fma.rn.f32 %f963, %f490, %f130, %f489;
|
|
|
|
BB0_68:
|
|
@%p62 bra BB0_70;
|
|
bra.uni BB0_69;
|
|
|
|
BB0_70:
|
|
mov.f32 %f496, 0fBE2AAAA3;
|
|
fma.rn.f32 %f497, %f963, %f130, %f496;
|
|
fma.rn.f32 %f964, %f497, %f130, %f454;
|
|
bra.uni BB0_71;
|
|
|
|
BB0_69:
|
|
mov.f32 %f493, 0f3D2AAAA5;
|
|
fma.rn.f32 %f494, %f963, %f130, %f493;
|
|
mov.f32 %f495, 0fBF000000;
|
|
fma.rn.f32 %f964, %f494, %f130, %f495;
|
|
|
|
BB0_71:
|
|
fma.rn.f32 %f965, %f964, %f962, %f962;
|
|
@%p62 bra BB0_73;
|
|
|
|
fma.rn.f32 %f965, %f964, %f130, %f407;
|
|
|
|
BB0_73:
|
|
and.b32 %r256, %r368, 2;
|
|
setp.eq.s32 %p65, %r256, 0;
|
|
@%p65 bra BB0_75;
|
|
|
|
mov.f32 %f501, 0fBF800000;
|
|
fma.rn.f32 %f965, %f965, %f501, %f454;
|
|
|
|
BB0_75:
|
|
mul.f32 %f510, %f105, %f959;
|
|
mul.f32 %f511, %f105, %f965;
|
|
mul.f32 %f512, %f80, %f511;
|
|
mul.f32 %f513, %f81, %f511;
|
|
mul.f32 %f514, %f82, %f511;
|
|
fma.rn.f32 %f515, %f89, %f510, %f512;
|
|
fma.rn.f32 %f516, %f88, %f510, %f513;
|
|
fma.rn.f32 %f517, %f87, %f510, %f514;
|
|
fma.rn.f32 %f505, %f909, %f104, %f515;
|
|
fma.rn.f32 %f506, %f910, %f104, %f516;
|
|
fma.rn.f32 %f507, %f911, %f104, %f517;
|
|
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 %f509, 0f6C4ECB8F;
|
|
// inline asm
|
|
call _rt_trace_64, (%r257, %f12, %f13, %f14, %f505, %f506, %f507, %r258, %f97, %f509, %rd76, %r119);
|
|
// inline asm
|
|
ld.local.f32 %f518, [%rd4];
|
|
max.f32 %f520, %f518, %f454;
|
|
ld.local.f32 %f521, [%rd4+4];
|
|
max.f32 %f522, %f521, %f454;
|
|
ld.local.f32 %f523, [%rd4+8];
|
|
max.f32 %f524, %f523, %f454;
|
|
setp.lt.f32 %p66, %f518, 0f00000000;
|
|
selp.f32 %f525, 0f00000000, 0f3F800000, %p66;
|
|
fma.rn.f32 %f942, %f505, %f525, %f942;
|
|
fma.rn.f32 %f941, %f506, %f525, %f941;
|
|
fma.rn.f32 %f940, %f507, %f525, %f940;
|
|
mul.f32 %f526, %f910, %f506;
|
|
fma.rn.f32 %f527, %f909, %f505, %f526;
|
|
fma.rn.f32 %f528, %f911, %f507, %f527;
|
|
cvt.sat.f32.f32 %f529, %f528;
|
|
fma.rn.f32 %f951, %f520, %f529, %f951;
|
|
fma.rn.f32 %f950, %f522, %f529, %f950;
|
|
fma.rn.f32 %f949, %f524, %f529, %f949;
|
|
add.s32 %r347, %r347, 1;
|
|
setp.lt.s32 %p67, %r347, %r3;
|
|
@%p67 bra BB0_29;
|
|
|
|
add.s32 %r345, %r345, 1;
|
|
setp.lt.s32 %p68, %r345, %r3;
|
|
@%p68 bra BB0_28;
|
|
|
|
BB0_77:
|
|
mul.lo.s32 %r260, %r3, %r3;
|
|
cvt.rn.f32.s32 %f530, %r260;
|
|
rcp.rn.f32 %f531, %f530;
|
|
mul.f32 %f532, %f951, %f531;
|
|
mul.f32 %f533, %f950, %f531;
|
|
mul.f32 %f534, %f949, %f531;
|
|
div.rn.f32 %f979, %f415, %f530;
|
|
fma.rn.f32 %f976, %f951, %f531, %f532;
|
|
fma.rn.f32 %f977, %f950, %f531, %f533;
|
|
fma.rn.f32 %f978, %f949, %f531, %f534;
|
|
|
|
BB0_78:
|
|
setp.gt.s32 %p69, %r1, -1;
|
|
selp.f32 %f165, %f979, %f976, %p69;
|
|
ld.global.u32 %r370, [imageEnabled];
|
|
and.b32 %r261, %r370, 8;
|
|
setp.eq.s32 %p70, %r261, 0;
|
|
@%p70 bra BB0_91;
|
|
|
|
cvt.u64.u32 %rd92, %r5;
|
|
cvt.u64.u32 %rd93, %r6;
|
|
mov.u64 %rd96, image_Mask;
|
|
cvta.global.u64 %rd91, %rd96;
|
|
// inline asm
|
|
call (%rd90), _rt_buffer_get_64, (%rd91, %r103, %r103, %rd92, %rd93, %rd27, %rd27);
|
|
// inline asm
|
|
mov.f32 %f538, 0f3E68BA2E;
|
|
cvt.rzi.f32.f32 %f539, %f538;
|
|
fma.rn.f32 %f540, %f539, 0fC0000000, 0f3EE8BA2E;
|
|
abs.f32 %f166, %f540;
|
|
abs.f32 %f167, %f165;
|
|
setp.lt.f32 %p71, %f167, 0f00800000;
|
|
mul.f32 %f541, %f167, 0f4B800000;
|
|
selp.f32 %f542, 0fC3170000, 0fC2FE0000, %p71;
|
|
selp.f32 %f543, %f541, %f167, %p71;
|
|
mov.b32 %r264, %f543;
|
|
and.b32 %r265, %r264, 8388607;
|
|
or.b32 %r266, %r265, 1065353216;
|
|
mov.b32 %f544, %r266;
|
|
shr.u32 %r267, %r264, 23;
|
|
cvt.rn.f32.u32 %f545, %r267;
|
|
add.f32 %f546, %f542, %f545;
|
|
setp.gt.f32 %p72, %f544, 0f3FB504F3;
|
|
mul.f32 %f547, %f544, 0f3F000000;
|
|
add.f32 %f548, %f546, 0f3F800000;
|
|
selp.f32 %f549, %f547, %f544, %p72;
|
|
selp.f32 %f550, %f548, %f546, %p72;
|
|
add.f32 %f551, %f549, 0fBF800000;
|
|
add.f32 %f537, %f549, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f536,%f537;
|
|
// inline asm
|
|
add.f32 %f552, %f551, %f551;
|
|
mul.f32 %f553, %f536, %f552;
|
|
mul.f32 %f554, %f553, %f553;
|
|
mov.f32 %f555, 0f3C4CAF63;
|
|
mov.f32 %f556, 0f3B18F0FE;
|
|
fma.rn.f32 %f557, %f556, %f554, %f555;
|
|
mov.f32 %f558, 0f3DAAAABD;
|
|
fma.rn.f32 %f559, %f557, %f554, %f558;
|
|
mul.rn.f32 %f560, %f559, %f554;
|
|
mul.rn.f32 %f561, %f560, %f553;
|
|
sub.f32 %f562, %f551, %f553;
|
|
neg.f32 %f563, %f553;
|
|
add.f32 %f564, %f562, %f562;
|
|
fma.rn.f32 %f565, %f563, %f551, %f564;
|
|
mul.rn.f32 %f566, %f536, %f565;
|
|
add.f32 %f567, %f561, %f553;
|
|
sub.f32 %f568, %f553, %f567;
|
|
add.f32 %f569, %f561, %f568;
|
|
add.f32 %f570, %f566, %f569;
|
|
add.f32 %f571, %f567, %f570;
|
|
sub.f32 %f572, %f567, %f571;
|
|
add.f32 %f573, %f570, %f572;
|
|
mov.f32 %f574, 0f3F317200;
|
|
mul.rn.f32 %f575, %f550, %f574;
|
|
mov.f32 %f576, 0f35BFBE8E;
|
|
mul.rn.f32 %f577, %f550, %f576;
|
|
add.f32 %f578, %f575, %f571;
|
|
sub.f32 %f579, %f575, %f578;
|
|
add.f32 %f580, %f571, %f579;
|
|
add.f32 %f581, %f573, %f580;
|
|
add.f32 %f582, %f577, %f581;
|
|
add.f32 %f583, %f578, %f582;
|
|
sub.f32 %f584, %f578, %f583;
|
|
add.f32 %f585, %f582, %f584;
|
|
mov.f32 %f586, 0f3EE8BA2E;
|
|
mul.rn.f32 %f587, %f586, %f583;
|
|
neg.f32 %f588, %f587;
|
|
fma.rn.f32 %f589, %f586, %f583, %f588;
|
|
fma.rn.f32 %f590, %f586, %f585, %f589;
|
|
mov.f32 %f591, 0f00000000;
|
|
fma.rn.f32 %f592, %f591, %f583, %f590;
|
|
add.rn.f32 %f593, %f587, %f592;
|
|
neg.f32 %f594, %f593;
|
|
add.rn.f32 %f595, %f587, %f594;
|
|
add.rn.f32 %f596, %f595, %f592;
|
|
mov.b32 %r268, %f593;
|
|
setp.eq.s32 %p73, %r268, 1118925336;
|
|
add.s32 %r269, %r268, -1;
|
|
mov.b32 %f597, %r269;
|
|
add.f32 %f598, %f596, 0f37000000;
|
|
selp.f32 %f599, %f597, %f593, %p73;
|
|
selp.f32 %f168, %f598, %f596, %p73;
|
|
mul.f32 %f600, %f599, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f601, %f600;
|
|
mov.f32 %f602, 0fBF317200;
|
|
fma.rn.f32 %f603, %f601, %f602, %f599;
|
|
mov.f32 %f604, 0fB5BFBE8E;
|
|
fma.rn.f32 %f605, %f601, %f604, %f603;
|
|
mul.f32 %f606, %f605, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f607, %f606;
|
|
add.f32 %f608, %f601, 0f00000000;
|
|
ex2.approx.f32 %f609, %f608;
|
|
mul.f32 %f610, %f607, %f609;
|
|
setp.lt.f32 %p74, %f599, 0fC2D20000;
|
|
selp.f32 %f611, 0f00000000, %f610, %p74;
|
|
setp.gt.f32 %p75, %f599, 0f42D20000;
|
|
selp.f32 %f980, 0f7F800000, %f611, %p75;
|
|
setp.eq.f32 %p76, %f980, 0f7F800000;
|
|
@%p76 bra BB0_81;
|
|
|
|
fma.rn.f32 %f980, %f980, %f168, %f980;
|
|
|
|
BB0_81:
|
|
setp.lt.f32 %p77, %f165, 0f00000000;
|
|
setp.eq.f32 %p78, %f166, 0f3F800000;
|
|
and.pred %p3, %p77, %p78;
|
|
mov.b32 %r270, %f980;
|
|
xor.b32 %r271, %r270, -2147483648;
|
|
mov.b32 %f612, %r271;
|
|
selp.f32 %f982, %f612, %f980, %p3;
|
|
setp.eq.f32 %p79, %f165, 0f00000000;
|
|
@%p79 bra BB0_84;
|
|
bra.uni BB0_82;
|
|
|
|
BB0_84:
|
|
add.f32 %f615, %f165, %f165;
|
|
selp.f32 %f982, %f615, 0f00000000, %p78;
|
|
bra.uni BB0_85;
|
|
|
|
BB0_139:
|
|
mov.u64 %rd164, image_HDR;
|
|
cvta.global.u64 %rd159, %rd164;
|
|
mov.u32 %r334, 8;
|
|
mov.u64 %rd163, 0;
|
|
// inline asm
|
|
call (%rd158), _rt_buffer_get_64, (%rd159, %r103, %r334, %rd20, %rd21, %rd163, %rd163);
|
|
// inline asm
|
|
mov.f32 %f899, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs51, %f899;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs52, 0;
|
|
st.v4.u16 [%rd158], {%rs51, %rs51, %rs51, %rs52};
|
|
|
|
BB0_140:
|
|
ld.global.u8 %rs53, [imageEnabled];
|
|
and.b16 %rs54, %rs53, 64;
|
|
setp.eq.s16 %p143, %rs54, 0;
|
|
@%p143 bra BB0_142;
|
|
|
|
cvt.u64.u32 %rd167, %r5;
|
|
cvt.u64.u32 %rd168, %r6;
|
|
mov.u64 %rd171, image_Dir;
|
|
cvta.global.u64 %rd166, %rd171;
|
|
mov.u64 %rd170, 0;
|
|
// inline asm
|
|
call (%rd165), _rt_buffer_get_64, (%rd166, %r103, %r104, %rd167, %rd168, %rd170, %rd170);
|
|
// inline asm
|
|
mov.u16 %rs55, 0;
|
|
st.v4.u8 [%rd165], {%rs55, %rs55, %rs55, %rs55};
|
|
bra.uni BB0_142;
|
|
|
|
BB0_82:
|
|
setp.geu.f32 %p80, %f165, 0f00000000;
|
|
@%p80 bra BB0_85;
|
|
|
|
cvt.rzi.f32.f32 %f614, %f586;
|
|
setp.neu.f32 %p81, %f614, 0f3EE8BA2E;
|
|
selp.f32 %f982, 0f7FFFFFFF, %f982, %p81;
|
|
|
|
BB0_85:
|
|
add.f32 %f616, %f167, 0f3EE8BA2E;
|
|
mov.b32 %r272, %f616;
|
|
setp.lt.s32 %p83, %r272, 2139095040;
|
|
@%p83 bra BB0_90;
|
|
|
|
setp.gtu.f32 %p84, %f167, 0f7F800000;
|
|
@%p84 bra BB0_89;
|
|
bra.uni BB0_87;
|
|
|
|
BB0_89:
|
|
add.f32 %f982, %f165, 0f3EE8BA2E;
|
|
bra.uni BB0_90;
|
|
|
|
BB0_87:
|
|
setp.neu.f32 %p85, %f167, 0f7F800000;
|
|
@%p85 bra BB0_90;
|
|
|
|
selp.f32 %f982, 0fFF800000, 0f7F800000, %p3;
|
|
|
|
BB0_90:
|
|
mul.f32 %f617, %f982, 0f437F0000;
|
|
setp.eq.f32 %p86, %f165, 0f3F800000;
|
|
selp.f32 %f618, 0f437F0000, %f617, %p86;
|
|
cvt.rzi.u32.f32 %r273, %f618;
|
|
cvt.u16.u32 %rs11, %r273;
|
|
mov.u16 %rs12, 255;
|
|
st.v2.u8 [%rd90], {%rs11, %rs12};
|
|
ld.global.u32 %r370, [imageEnabled];
|
|
|
|
BB0_91:
|
|
ld.global.f32 %f619, [lightColor];
|
|
mul.f32 %f179, %f976, %f619;
|
|
ld.global.f32 %f620, [lightColor+4];
|
|
mul.f32 %f180, %f977, %f620;
|
|
ld.global.f32 %f621, [lightColor+8];
|
|
mul.f32 %f181, %f978, %f621;
|
|
and.b32 %r274, %r370, 1;
|
|
setp.eq.b32 %p87, %r274, 1;
|
|
@!%p87 bra BB0_126;
|
|
bra.uni BB0_92;
|
|
|
|
BB0_92:
|
|
mov.f32 %f624, 0f3E666666;
|
|
cvt.rzi.f32.f32 %f625, %f624;
|
|
fma.rn.f32 %f626, %f625, 0fC0000000, 0f3EE66666;
|
|
abs.f32 %f182, %f626;
|
|
abs.f32 %f183, %f179;
|
|
setp.lt.f32 %p88, %f183, 0f00800000;
|
|
mul.f32 %f627, %f183, 0f4B800000;
|
|
selp.f32 %f628, 0fC3170000, 0fC2FE0000, %p88;
|
|
selp.f32 %f629, %f627, %f183, %p88;
|
|
mov.b32 %r275, %f629;
|
|
and.b32 %r276, %r275, 8388607;
|
|
or.b32 %r277, %r276, 1065353216;
|
|
mov.b32 %f630, %r277;
|
|
shr.u32 %r278, %r275, 23;
|
|
cvt.rn.f32.u32 %f631, %r278;
|
|
add.f32 %f632, %f628, %f631;
|
|
setp.gt.f32 %p89, %f630, 0f3FB504F3;
|
|
mul.f32 %f633, %f630, 0f3F000000;
|
|
add.f32 %f634, %f632, 0f3F800000;
|
|
selp.f32 %f635, %f633, %f630, %p89;
|
|
selp.f32 %f636, %f634, %f632, %p89;
|
|
add.f32 %f637, %f635, 0fBF800000;
|
|
add.f32 %f623, %f635, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f622,%f623;
|
|
// inline asm
|
|
add.f32 %f638, %f637, %f637;
|
|
mul.f32 %f639, %f622, %f638;
|
|
mul.f32 %f640, %f639, %f639;
|
|
mov.f32 %f641, 0f3C4CAF63;
|
|
mov.f32 %f642, 0f3B18F0FE;
|
|
fma.rn.f32 %f643, %f642, %f640, %f641;
|
|
mov.f32 %f644, 0f3DAAAABD;
|
|
fma.rn.f32 %f645, %f643, %f640, %f644;
|
|
mul.rn.f32 %f646, %f645, %f640;
|
|
mul.rn.f32 %f647, %f646, %f639;
|
|
sub.f32 %f648, %f637, %f639;
|
|
neg.f32 %f649, %f639;
|
|
add.f32 %f650, %f648, %f648;
|
|
fma.rn.f32 %f651, %f649, %f637, %f650;
|
|
mul.rn.f32 %f652, %f622, %f651;
|
|
add.f32 %f653, %f647, %f639;
|
|
sub.f32 %f654, %f639, %f653;
|
|
add.f32 %f655, %f647, %f654;
|
|
add.f32 %f656, %f652, %f655;
|
|
add.f32 %f657, %f653, %f656;
|
|
sub.f32 %f658, %f653, %f657;
|
|
add.f32 %f659, %f656, %f658;
|
|
mov.f32 %f660, 0f3F317200;
|
|
mul.rn.f32 %f661, %f636, %f660;
|
|
mov.f32 %f662, 0f35BFBE8E;
|
|
mul.rn.f32 %f663, %f636, %f662;
|
|
add.f32 %f664, %f661, %f657;
|
|
sub.f32 %f665, %f661, %f664;
|
|
add.f32 %f666, %f657, %f665;
|
|
add.f32 %f667, %f659, %f666;
|
|
add.f32 %f668, %f663, %f667;
|
|
add.f32 %f669, %f664, %f668;
|
|
sub.f32 %f670, %f664, %f669;
|
|
add.f32 %f671, %f668, %f670;
|
|
mov.f32 %f672, 0f3EE66666;
|
|
mul.rn.f32 %f673, %f672, %f669;
|
|
neg.f32 %f674, %f673;
|
|
fma.rn.f32 %f675, %f672, %f669, %f674;
|
|
fma.rn.f32 %f676, %f672, %f671, %f675;
|
|
mov.f32 %f677, 0f00000000;
|
|
fma.rn.f32 %f678, %f677, %f669, %f676;
|
|
add.rn.f32 %f679, %f673, %f678;
|
|
neg.f32 %f680, %f679;
|
|
add.rn.f32 %f681, %f673, %f680;
|
|
add.rn.f32 %f682, %f681, %f678;
|
|
mov.b32 %r279, %f679;
|
|
setp.eq.s32 %p90, %r279, 1118925336;
|
|
add.s32 %r280, %r279, -1;
|
|
mov.b32 %f683, %r280;
|
|
add.f32 %f684, %f682, 0f37000000;
|
|
selp.f32 %f685, %f683, %f679, %p90;
|
|
selp.f32 %f184, %f684, %f682, %p90;
|
|
mul.f32 %f686, %f685, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f687, %f686;
|
|
mov.f32 %f688, 0fBF317200;
|
|
fma.rn.f32 %f689, %f687, %f688, %f685;
|
|
mov.f32 %f690, 0fB5BFBE8E;
|
|
fma.rn.f32 %f691, %f687, %f690, %f689;
|
|
mul.f32 %f692, %f691, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f693, %f692;
|
|
add.f32 %f694, %f687, 0f00000000;
|
|
ex2.approx.f32 %f695, %f694;
|
|
mul.f32 %f696, %f693, %f695;
|
|
setp.lt.f32 %p91, %f685, 0fC2D20000;
|
|
selp.f32 %f697, 0f00000000, %f696, %p91;
|
|
setp.gt.f32 %p92, %f685, 0f42D20000;
|
|
selp.f32 %f983, 0f7F800000, %f697, %p92;
|
|
setp.eq.f32 %p93, %f983, 0f7F800000;
|
|
@%p93 bra BB0_94;
|
|
|
|
fma.rn.f32 %f983, %f983, %f184, %f983;
|
|
|
|
BB0_94:
|
|
setp.lt.f32 %p94, %f179, 0f00000000;
|
|
setp.eq.f32 %p95, %f182, 0f3F800000;
|
|
and.pred %p4, %p94, %p95;
|
|
mov.b32 %r281, %f983;
|
|
xor.b32 %r282, %r281, -2147483648;
|
|
mov.b32 %f698, %r282;
|
|
selp.f32 %f985, %f698, %f983, %p4;
|
|
setp.eq.f32 %p96, %f179, 0f00000000;
|
|
@%p96 bra BB0_97;
|
|
bra.uni BB0_95;
|
|
|
|
BB0_97:
|
|
add.f32 %f701, %f179, %f179;
|
|
selp.f32 %f985, %f701, 0f00000000, %p95;
|
|
bra.uni BB0_98;
|
|
|
|
BB0_95:
|
|
setp.geu.f32 %p97, %f179, 0f00000000;
|
|
@%p97 bra BB0_98;
|
|
|
|
cvt.rzi.f32.f32 %f700, %f672;
|
|
setp.neu.f32 %p98, %f700, 0f3EE66666;
|
|
selp.f32 %f985, 0f7FFFFFFF, %f985, %p98;
|
|
|
|
BB0_98:
|
|
add.f32 %f702, %f183, 0f3EE66666;
|
|
mov.b32 %r283, %f702;
|
|
setp.lt.s32 %p100, %r283, 2139095040;
|
|
@%p100 bra BB0_103;
|
|
|
|
setp.gtu.f32 %p101, %f183, 0f7F800000;
|
|
@%p101 bra BB0_102;
|
|
bra.uni BB0_100;
|
|
|
|
BB0_102:
|
|
add.f32 %f985, %f179, 0f3EE66666;
|
|
bra.uni BB0_103;
|
|
|
|
BB0_100:
|
|
setp.neu.f32 %p102, %f183, 0f7F800000;
|
|
@%p102 bra BB0_103;
|
|
|
|
selp.f32 %f985, 0fFF800000, 0f7F800000, %p4;
|
|
|
|
BB0_103:
|
|
setp.eq.f32 %p103, %f179, 0f3F800000;
|
|
selp.f32 %f195, 0f3F800000, %f985, %p103;
|
|
abs.f32 %f196, %f180;
|
|
setp.lt.f32 %p104, %f196, 0f00800000;
|
|
mul.f32 %f705, %f196, 0f4B800000;
|
|
selp.f32 %f706, 0fC3170000, 0fC2FE0000, %p104;
|
|
selp.f32 %f707, %f705, %f196, %p104;
|
|
mov.b32 %r284, %f707;
|
|
and.b32 %r285, %r284, 8388607;
|
|
or.b32 %r286, %r285, 1065353216;
|
|
mov.b32 %f708, %r286;
|
|
shr.u32 %r287, %r284, 23;
|
|
cvt.rn.f32.u32 %f709, %r287;
|
|
add.f32 %f710, %f706, %f709;
|
|
setp.gt.f32 %p105, %f708, 0f3FB504F3;
|
|
mul.f32 %f711, %f708, 0f3F000000;
|
|
add.f32 %f712, %f710, 0f3F800000;
|
|
selp.f32 %f713, %f711, %f708, %p105;
|
|
selp.f32 %f714, %f712, %f710, %p105;
|
|
add.f32 %f715, %f713, 0fBF800000;
|
|
add.f32 %f704, %f713, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f703,%f704;
|
|
// inline asm
|
|
add.f32 %f716, %f715, %f715;
|
|
mul.f32 %f717, %f703, %f716;
|
|
mul.f32 %f718, %f717, %f717;
|
|
fma.rn.f32 %f721, %f642, %f718, %f641;
|
|
fma.rn.f32 %f723, %f721, %f718, %f644;
|
|
mul.rn.f32 %f724, %f723, %f718;
|
|
mul.rn.f32 %f725, %f724, %f717;
|
|
sub.f32 %f726, %f715, %f717;
|
|
neg.f32 %f727, %f717;
|
|
add.f32 %f728, %f726, %f726;
|
|
fma.rn.f32 %f729, %f727, %f715, %f728;
|
|
mul.rn.f32 %f730, %f703, %f729;
|
|
add.f32 %f731, %f725, %f717;
|
|
sub.f32 %f732, %f717, %f731;
|
|
add.f32 %f733, %f725, %f732;
|
|
add.f32 %f734, %f730, %f733;
|
|
add.f32 %f735, %f731, %f734;
|
|
sub.f32 %f736, %f731, %f735;
|
|
add.f32 %f737, %f734, %f736;
|
|
mul.rn.f32 %f739, %f714, %f660;
|
|
mul.rn.f32 %f741, %f714, %f662;
|
|
add.f32 %f742, %f739, %f735;
|
|
sub.f32 %f743, %f739, %f742;
|
|
add.f32 %f744, %f735, %f743;
|
|
add.f32 %f745, %f737, %f744;
|
|
add.f32 %f746, %f741, %f745;
|
|
add.f32 %f747, %f742, %f746;
|
|
sub.f32 %f748, %f742, %f747;
|
|
add.f32 %f749, %f746, %f748;
|
|
mul.rn.f32 %f751, %f672, %f747;
|
|
neg.f32 %f752, %f751;
|
|
fma.rn.f32 %f753, %f672, %f747, %f752;
|
|
fma.rn.f32 %f754, %f672, %f749, %f753;
|
|
fma.rn.f32 %f756, %f677, %f747, %f754;
|
|
add.rn.f32 %f757, %f751, %f756;
|
|
neg.f32 %f758, %f757;
|
|
add.rn.f32 %f759, %f751, %f758;
|
|
add.rn.f32 %f760, %f759, %f756;
|
|
mov.b32 %r288, %f757;
|
|
setp.eq.s32 %p106, %r288, 1118925336;
|
|
add.s32 %r289, %r288, -1;
|
|
mov.b32 %f761, %r289;
|
|
add.f32 %f762, %f760, 0f37000000;
|
|
selp.f32 %f763, %f761, %f757, %p106;
|
|
selp.f32 %f197, %f762, %f760, %p106;
|
|
mul.f32 %f764, %f763, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f765, %f764;
|
|
fma.rn.f32 %f767, %f765, %f688, %f763;
|
|
fma.rn.f32 %f769, %f765, %f690, %f767;
|
|
mul.f32 %f770, %f769, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f771, %f770;
|
|
add.f32 %f772, %f765, 0f00000000;
|
|
ex2.approx.f32 %f773, %f772;
|
|
mul.f32 %f774, %f771, %f773;
|
|
setp.lt.f32 %p107, %f763, 0fC2D20000;
|
|
selp.f32 %f775, 0f00000000, %f774, %p107;
|
|
setp.gt.f32 %p108, %f763, 0f42D20000;
|
|
selp.f32 %f986, 0f7F800000, %f775, %p108;
|
|
setp.eq.f32 %p109, %f986, 0f7F800000;
|
|
@%p109 bra BB0_105;
|
|
|
|
fma.rn.f32 %f986, %f986, %f197, %f986;
|
|
|
|
BB0_105:
|
|
setp.lt.f32 %p110, %f180, 0f00000000;
|
|
and.pred %p5, %p110, %p95;
|
|
mov.b32 %r290, %f986;
|
|
xor.b32 %r291, %r290, -2147483648;
|
|
mov.b32 %f776, %r291;
|
|
selp.f32 %f988, %f776, %f986, %p5;
|
|
setp.eq.f32 %p112, %f180, 0f00000000;
|
|
@%p112 bra BB0_108;
|
|
bra.uni BB0_106;
|
|
|
|
BB0_108:
|
|
add.f32 %f779, %f180, %f180;
|
|
selp.f32 %f988, %f779, 0f00000000, %p95;
|
|
bra.uni BB0_109;
|
|
|
|
BB0_106:
|
|
setp.geu.f32 %p113, %f180, 0f00000000;
|
|
@%p113 bra BB0_109;
|
|
|
|
cvt.rzi.f32.f32 %f778, %f672;
|
|
setp.neu.f32 %p114, %f778, 0f3EE66666;
|
|
selp.f32 %f988, 0f7FFFFFFF, %f988, %p114;
|
|
|
|
BB0_109:
|
|
add.f32 %f780, %f196, 0f3EE66666;
|
|
mov.b32 %r292, %f780;
|
|
setp.lt.s32 %p116, %r292, 2139095040;
|
|
@%p116 bra BB0_114;
|
|
|
|
setp.gtu.f32 %p117, %f196, 0f7F800000;
|
|
@%p117 bra BB0_113;
|
|
bra.uni BB0_111;
|
|
|
|
BB0_113:
|
|
add.f32 %f988, %f180, 0f3EE66666;
|
|
bra.uni BB0_114;
|
|
|
|
BB0_111:
|
|
setp.neu.f32 %p118, %f196, 0f7F800000;
|
|
@%p118 bra BB0_114;
|
|
|
|
selp.f32 %f988, 0fFF800000, 0f7F800000, %p5;
|
|
|
|
BB0_114:
|
|
setp.eq.f32 %p119, %f180, 0f3F800000;
|
|
selp.f32 %f208, 0f3F800000, %f988, %p119;
|
|
abs.f32 %f209, %f181;
|
|
setp.lt.f32 %p120, %f209, 0f00800000;
|
|
mul.f32 %f783, %f209, 0f4B800000;
|
|
selp.f32 %f784, 0fC3170000, 0fC2FE0000, %p120;
|
|
selp.f32 %f785, %f783, %f209, %p120;
|
|
mov.b32 %r293, %f785;
|
|
and.b32 %r294, %r293, 8388607;
|
|
or.b32 %r295, %r294, 1065353216;
|
|
mov.b32 %f786, %r295;
|
|
shr.u32 %r296, %r293, 23;
|
|
cvt.rn.f32.u32 %f787, %r296;
|
|
add.f32 %f788, %f784, %f787;
|
|
setp.gt.f32 %p121, %f786, 0f3FB504F3;
|
|
mul.f32 %f789, %f786, 0f3F000000;
|
|
add.f32 %f790, %f788, 0f3F800000;
|
|
selp.f32 %f791, %f789, %f786, %p121;
|
|
selp.f32 %f792, %f790, %f788, %p121;
|
|
add.f32 %f793, %f791, 0fBF800000;
|
|
add.f32 %f782, %f791, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f781,%f782;
|
|
// inline asm
|
|
add.f32 %f794, %f793, %f793;
|
|
mul.f32 %f795, %f781, %f794;
|
|
mul.f32 %f796, %f795, %f795;
|
|
fma.rn.f32 %f799, %f642, %f796, %f641;
|
|
fma.rn.f32 %f801, %f799, %f796, %f644;
|
|
mul.rn.f32 %f802, %f801, %f796;
|
|
mul.rn.f32 %f803, %f802, %f795;
|
|
sub.f32 %f804, %f793, %f795;
|
|
neg.f32 %f805, %f795;
|
|
add.f32 %f806, %f804, %f804;
|
|
fma.rn.f32 %f807, %f805, %f793, %f806;
|
|
mul.rn.f32 %f808, %f781, %f807;
|
|
add.f32 %f809, %f803, %f795;
|
|
sub.f32 %f810, %f795, %f809;
|
|
add.f32 %f811, %f803, %f810;
|
|
add.f32 %f812, %f808, %f811;
|
|
add.f32 %f813, %f809, %f812;
|
|
sub.f32 %f814, %f809, %f813;
|
|
add.f32 %f815, %f812, %f814;
|
|
mul.rn.f32 %f817, %f792, %f660;
|
|
mul.rn.f32 %f819, %f792, %f662;
|
|
add.f32 %f820, %f817, %f813;
|
|
sub.f32 %f821, %f817, %f820;
|
|
add.f32 %f822, %f813, %f821;
|
|
add.f32 %f823, %f815, %f822;
|
|
add.f32 %f824, %f819, %f823;
|
|
add.f32 %f825, %f820, %f824;
|
|
sub.f32 %f826, %f820, %f825;
|
|
add.f32 %f827, %f824, %f826;
|
|
mul.rn.f32 %f829, %f672, %f825;
|
|
neg.f32 %f830, %f829;
|
|
fma.rn.f32 %f831, %f672, %f825, %f830;
|
|
fma.rn.f32 %f832, %f672, %f827, %f831;
|
|
fma.rn.f32 %f834, %f677, %f825, %f832;
|
|
add.rn.f32 %f835, %f829, %f834;
|
|
neg.f32 %f836, %f835;
|
|
add.rn.f32 %f837, %f829, %f836;
|
|
add.rn.f32 %f838, %f837, %f834;
|
|
mov.b32 %r297, %f835;
|
|
setp.eq.s32 %p122, %r297, 1118925336;
|
|
add.s32 %r298, %r297, -1;
|
|
mov.b32 %f839, %r298;
|
|
add.f32 %f840, %f838, 0f37000000;
|
|
selp.f32 %f841, %f839, %f835, %p122;
|
|
selp.f32 %f210, %f840, %f838, %p122;
|
|
mul.f32 %f842, %f841, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f843, %f842;
|
|
fma.rn.f32 %f845, %f843, %f688, %f841;
|
|
fma.rn.f32 %f847, %f843, %f690, %f845;
|
|
mul.f32 %f848, %f847, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f849, %f848;
|
|
add.f32 %f850, %f843, 0f00000000;
|
|
ex2.approx.f32 %f851, %f850;
|
|
mul.f32 %f852, %f849, %f851;
|
|
setp.lt.f32 %p123, %f841, 0fC2D20000;
|
|
selp.f32 %f853, 0f00000000, %f852, %p123;
|
|
setp.gt.f32 %p124, %f841, 0f42D20000;
|
|
selp.f32 %f989, 0f7F800000, %f853, %p124;
|
|
setp.eq.f32 %p125, %f989, 0f7F800000;
|
|
@%p125 bra BB0_116;
|
|
|
|
fma.rn.f32 %f989, %f989, %f210, %f989;
|
|
|
|
BB0_116:
|
|
setp.lt.f32 %p126, %f181, 0f00000000;
|
|
and.pred %p6, %p126, %p95;
|
|
mov.b32 %r299, %f989;
|
|
xor.b32 %r300, %r299, -2147483648;
|
|
mov.b32 %f854, %r300;
|
|
selp.f32 %f991, %f854, %f989, %p6;
|
|
setp.eq.f32 %p128, %f181, 0f00000000;
|
|
@%p128 bra BB0_119;
|
|
bra.uni BB0_117;
|
|
|
|
BB0_119:
|
|
add.f32 %f857, %f181, %f181;
|
|
selp.f32 %f991, %f857, 0f00000000, %p95;
|
|
bra.uni BB0_120;
|
|
|
|
BB0_117:
|
|
setp.geu.f32 %p129, %f181, 0f00000000;
|
|
@%p129 bra BB0_120;
|
|
|
|
cvt.rzi.f32.f32 %f856, %f672;
|
|
setp.neu.f32 %p130, %f856, 0f3EE66666;
|
|
selp.f32 %f991, 0f7FFFFFFF, %f991, %p130;
|
|
|
|
BB0_120:
|
|
add.f32 %f858, %f209, 0f3EE66666;
|
|
mov.b32 %r301, %f858;
|
|
setp.lt.s32 %p132, %r301, 2139095040;
|
|
@%p132 bra BB0_125;
|
|
|
|
setp.gtu.f32 %p133, %f209, 0f7F800000;
|
|
@%p133 bra BB0_124;
|
|
bra.uni BB0_122;
|
|
|
|
BB0_124:
|
|
add.f32 %f991, %f181, 0f3EE66666;
|
|
bra.uni BB0_125;
|
|
|
|
BB0_122:
|
|
setp.neu.f32 %p134, %f209, 0f7F800000;
|
|
@%p134 bra BB0_125;
|
|
|
|
selp.f32 %f991, 0fFF800000, 0f7F800000, %p6;
|
|
|
|
BB0_125:
|
|
setp.eq.f32 %p135, %f181, 0f3F800000;
|
|
selp.f32 %f859, 0f3F800000, %f991, %p135;
|
|
cvt.u64.u32 %rd100, %r6;
|
|
cvt.u64.u32 %rd99, %r5;
|
|
mov.u64 %rd103, image;
|
|
cvta.global.u64 %rd98, %rd103;
|
|
// inline asm
|
|
call (%rd97), _rt_buffer_get_64, (%rd98, %r103, %r104, %rd99, %rd100, %rd27, %rd27);
|
|
// inline asm
|
|
cvt.sat.f32.f32 %f860, %f859;
|
|
mul.f32 %f861, %f860, 0f437FFD71;
|
|
cvt.rzi.u32.f32 %r304, %f861;
|
|
cvt.sat.f32.f32 %f862, %f208;
|
|
mul.f32 %f863, %f862, 0f437FFD71;
|
|
cvt.rzi.u32.f32 %r305, %f863;
|
|
cvt.sat.f32.f32 %f864, %f195;
|
|
mul.f32 %f865, %f864, 0f437FFD71;
|
|
cvt.rzi.u32.f32 %r306, %f865;
|
|
cvt.u16.u32 %rs13, %r304;
|
|
cvt.u16.u32 %rs14, %r306;
|
|
cvt.u16.u32 %rs15, %r305;
|
|
mov.u16 %rs16, 255;
|
|
st.v4.u8 [%rd97], {%rs13, %rs15, %rs14, %rs16};
|
|
ld.global.u32 %r370, [imageEnabled];
|
|
|
|
BB0_126:
|
|
and.b32 %r307, %r370, 4;
|
|
setp.eq.s32 %p136, %r307, 0;
|
|
@%p136 bra BB0_130;
|
|
|
|
ld.global.u32 %r308, [additive];
|
|
setp.eq.s32 %p137, %r308, 0;
|
|
cvt.u64.u32 %rd18, %r5;
|
|
cvt.u64.u32 %rd19, %r6;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs17, %f407;}
|
|
|
|
// inline asm
|
|
@%p137 bra BB0_129;
|
|
|
|
mov.u64 %rd116, image_HDR;
|
|
cvta.global.u64 %rd105, %rd116;
|
|
mov.u32 %r312, 8;
|
|
// inline asm
|
|
call (%rd104), _rt_buffer_get_64, (%rd105, %r103, %r312, %rd18, %rd19, %rd27, %rd27);
|
|
// inline asm
|
|
ld.v4.u16 {%rs24, %rs25, %rs26, %rs27}, [%rd104];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f867, %rs24;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f868, %rs25;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f869, %rs26;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd110), _rt_buffer_get_64, (%rd105, %r103, %r312, %rd18, %rd19, %rd27, %rd27);
|
|
// inline asm
|
|
add.f32 %f870, %f179, %f867;
|
|
add.f32 %f871, %f180, %f868;
|
|
add.f32 %f872, %f181, %f869;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs23, %f872;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs22, %f871;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs21, %f870;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd110], {%rs21, %rs22, %rs23, %rs17};
|
|
bra.uni BB0_130;
|
|
|
|
BB0_129:
|
|
mov.u64 %rd123, image_HDR;
|
|
cvta.global.u64 %rd118, %rd123;
|
|
mov.u32 %r314, 8;
|
|
// inline asm
|
|
call (%rd117), _rt_buffer_get_64, (%rd118, %r103, %r314, %rd18, %rd19, %rd27, %rd27);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs30, %f181;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs29, %f180;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs28, %f179;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd117], {%rs28, %rs29, %rs30, %rs17};
|
|
|
|
BB0_130:
|
|
ld.global.u8 %rs31, [imageEnabled];
|
|
and.b16 %rs32, %rs31, 64;
|
|
setp.eq.s16 %p138, %rs32, 0;
|
|
@%p138 bra BB0_142;
|
|
|
|
mul.f32 %f876, %f941, %f941;
|
|
fma.rn.f32 %f877, %f942, %f942, %f876;
|
|
fma.rn.f32 %f878, %f940, %f940, %f877;
|
|
sqrt.rn.f32 %f879, %f878;
|
|
rcp.rn.f32 %f880, %f879;
|
|
mul.f32 %f881, %f942, %f880;
|
|
mul.f32 %f882, %f941, %f880;
|
|
mul.f32 %f883, %f940, %f880;
|
|
cvt.u64.u32 %rd127, %r6;
|
|
cvt.u64.u32 %rd126, %r5;
|
|
mov.u64 %rd130, image_Dir;
|
|
cvta.global.u64 %rd125, %rd130;
|
|
// inline asm
|
|
call (%rd124), _rt_buffer_get_64, (%rd125, %r103, %r104, %rd126, %rd127, %rd27, %rd27);
|
|
// inline asm
|
|
fma.rn.f32 %f884, %f881, 0f3F000000, 0f3F000000;
|
|
mul.f32 %f885, %f884, 0f437F0000;
|
|
cvt.rzi.u32.f32 %r317, %f885;
|
|
fma.rn.f32 %f886, %f882, 0f3F000000, 0f3F000000;
|
|
mul.f32 %f887, %f886, 0f437F0000;
|
|
cvt.rzi.u32.f32 %r318, %f887;
|
|
fma.rn.f32 %f888, %f883, 0f3F000000, 0f3F000000;
|
|
mul.f32 %f889, %f888, 0f437F0000;
|
|
cvt.rzi.u32.f32 %r319, %f889;
|
|
cvt.u16.u32 %rs33, %r319;
|
|
cvt.u16.u32 %rs34, %r318;
|
|
cvt.u16.u32 %rs35, %r317;
|
|
mov.u16 %rs36, 255;
|
|
st.v4.u8 [%rd124], {%rs35, %rs34, %rs33, %rs36};
|
|
|
|
BB0_142:
|
|
ret;
|
|
}
|
|
|
|
|