Add Bakery back
This commit is contained in:
Toomas Tamm
9
Assets/Editor/x64.meta
Normal file
9
Assets/Editor/x64.meta
Normal file
@@ -0,0 +1,9 @@
|
||||
fileFormatVersion: 2
|
||||
guid: 5d1c2c72ba259dc4d91764b043bc0132
|
||||
folderAsset: yes
|
||||
timeCreated: 1606240228
|
||||
licenseType: Store
|
||||
DefaultImporter:
|
||||
userData:
|
||||
assetBundleName:
|
||||
assetBundleVariant:
|
||||
9
Assets/Editor/x64/Bakery.meta
Normal file
9
Assets/Editor/x64/Bakery.meta
Normal file
@@ -0,0 +1,9 @@
|
||||
fileFormatVersion: 2
|
||||
guid: 77a27e27bd2771544be2b85bfe73d283
|
||||
folderAsset: yes
|
||||
timeCreated: 1606240228
|
||||
licenseType: Store
|
||||
DefaultImporter:
|
||||
userData:
|
||||
assetBundleName:
|
||||
assetBundleVariant:
|
||||
22
Assets/Editor/x64/Bakery/.denoiserlog.txt
Normal file
22
Assets/Editor/x64/Bakery/.denoiserlog.txt
Normal file
@@ -0,0 +1,22 @@
|
||||
2021-3-5
|
||||
Init...
|
||||
Tiles: 1
|
||||
Loading shader denoisePrepare.ptx::oxMain...
|
||||
Loading shader denoiseFinish.ptx::oxMain...
|
||||
2021-3-5
|
||||
Init...
|
||||
Tiles: 1
|
||||
Done
|
||||
Loading shader denoisePrepare.ptx::oxMain...
|
||||
Loading shader denoiseFinish.ptx::oxMain...
|
||||
Done
|
||||
2021-3-5
|
||||
Init...
|
||||
Error (7801): OPTIX_ERROR_UNSUPPORTED_ABI_VERSION
|
||||
|
||||
2021-3-5
|
||||
Init...
|
||||
Tiles: 1
|
||||
Loading shader denoisePrepareOIDN.ptx::oxMain...
|
||||
Loading shader denoiseFinishOIDN.ptx::oxMain...
|
||||
Done
|
||||
94
Assets/Editor/x64/Bakery/.ftracelog.txt
Normal file
94
Assets/Editor/x64/Bakery/.ftracelog.txt
Normal file
@@ -0,0 +1,94 @@
|
||||
Init...
|
||||
sun
|
||||
Loading geometry...
|
||||
Loading alphas...
|
||||
Tile size: 512
|
||||
Device 0: GeForce RTX 2080 Ti (ON)
|
||||
Used host memory: 0mb, available GPU memory: 10133mb
|
||||
MODE: sun
|
||||
USING HALF
|
||||
Load shaders...
|
||||
Loading shader lmSun.ptx::oxMain...
|
||||
Loading shader clip_rtx.ptx::oxMain...
|
||||
Loading shader shadow_rtx.ptx::oxMain...
|
||||
Loading shader miss.ptx::oxMain...
|
||||
Loading shader attrib_trimesh.ptx::interpolate...
|
||||
Used host memory: 0mb, available GPU memory: 10133mb
|
||||
Used host memory: 0mb, available GPU memory: 10133mb
|
||||
Create buffer...
|
||||
Num passes: 4
|
||||
Num tex: 4
|
||||
Skipping pass LDR
|
||||
Pass light_HDR...
|
||||
Skipping pass Mask
|
||||
Skipping pass Dir
|
||||
Used host memory: 0mb, available GPU memory: 10133mb
|
||||
Launch 1x1 tiles...
|
||||
Tile 0x0...
|
||||
Streaming...0
|
||||
npHD
|
||||
Launching...
|
||||
Used host memory: 0mb, available GPU memory: 10042mb
|
||||
-----
|
||||
Mem stats:
|
||||
Total: 90
|
||||
Alphas: 0
|
||||
GI albedos: 0
|
||||
GI lightmaps: 0
|
||||
Light data: 0
|
||||
Output: 0
|
||||
UV GBuffer: 0
|
||||
Scene: 8282
|
||||
-----
|
||||
Render time: 0.586 sec
|
||||
|
||||
Writing...
|
||||
light_HDR.lz4
|
||||
Done
|
||||
Init...
|
||||
sun
|
||||
Loading geometry...
|
||||
Loading alphas...
|
||||
Tile size: 512
|
||||
Device 0: GeForce RTX 2080 Ti (ON)
|
||||
Used host memory: 0mb, available GPU memory: 9742mb
|
||||
MODE: sun
|
||||
USING HALF
|
||||
Load shaders...
|
||||
Loading shader lmSun.ptx::oxMain...
|
||||
Loading shader shadow.ptx::oxMain...
|
||||
Loading shader miss.ptx::oxMain...
|
||||
Loading shader trimesh.ptx::intersect...
|
||||
Loading shader trimesh.ptx::bbox...
|
||||
Used host memory: 0mb, available GPU memory: 9742mb
|
||||
Used host memory: 0mb, available GPU memory: 9742mb
|
||||
Create buffer...
|
||||
Num passes: 4
|
||||
Num tex: 4
|
||||
Skipping pass LDR
|
||||
Pass light_HDR...
|
||||
Skipping pass Mask
|
||||
Skipping pass Dir
|
||||
Used host memory: 0mb, available GPU memory: 9742mb
|
||||
Launch 1x1 tiles...
|
||||
Tile 0x0...
|
||||
Streaming...0
|
||||
npHD
|
||||
Launching...
|
||||
Used host memory: 0mb, available GPU memory: 9740mb
|
||||
-----
|
||||
Mem stats:
|
||||
Total: 2
|
||||
Alphas: 0
|
||||
GI albedos: 0
|
||||
GI lightmaps: 0
|
||||
Light data: 0
|
||||
Output: 0
|
||||
UV GBuffer: 0
|
||||
Scene: 8193
|
||||
-----
|
||||
Render time: 0.52 sec
|
||||
|
||||
Writing...
|
||||
light_HDR.lz4
|
||||
Done
|
||||
BIN
Assets/Editor/x64/Bakery/NormalsFittingTexture_dds
Normal file
BIN
Assets/Editor/x64/Bakery/NormalsFittingTexture_dds
Normal file
Binary file not shown.
8
Assets/Editor/x64/Bakery/NormalsFittingTexture_dds.meta
Normal file
8
Assets/Editor/x64/Bakery/NormalsFittingTexture_dds.meta
Normal file
@@ -0,0 +1,8 @@
|
||||
fileFormatVersion: 2
|
||||
guid: fad0196ede3a41749af057ec68f43124
|
||||
timeCreated: 1531334118
|
||||
licenseType: Store
|
||||
DefaultImporter:
|
||||
userData:
|
||||
assetBundleName:
|
||||
assetBundleVariant:
|
||||
BIN
Assets/Editor/x64/Bakery/OpenImageDenoise.dll
LFS
Normal file
BIN
Assets/Editor/x64/Bakery/OpenImageDenoise.dll
LFS
Normal file
Binary file not shown.
28
Assets/Editor/x64/Bakery/OpenImageDenoise.dll.meta
Normal file
28
Assets/Editor/x64/Bakery/OpenImageDenoise.dll.meta
Normal file
@@ -0,0 +1,28 @@
|
||||
fileFormatVersion: 2
|
||||
guid: 4ccddc16f2ea5484dbb20164a60bf488
|
||||
timeCreated: 1589355275
|
||||
licenseType: Store
|
||||
PluginImporter:
|
||||
serializedVersion: 2
|
||||
iconMap: {}
|
||||
executionOrder: {}
|
||||
isPreloaded: 0
|
||||
isOverridable: 0
|
||||
platformData:
|
||||
data:
|
||||
first:
|
||||
Any:
|
||||
second:
|
||||
enabled: 0
|
||||
settings: {}
|
||||
data:
|
||||
first:
|
||||
Editor: Editor
|
||||
second:
|
||||
enabled: 1
|
||||
settings:
|
||||
CPU: x86_64
|
||||
DefaultValueInitialized: true
|
||||
userData:
|
||||
assetBundleName:
|
||||
assetBundleVariant:
|
||||
200
Assets/Editor/x64/Bakery/addAO.ptx
Normal file
200
Assets/Editor/x64/Bakery/addAO.ptx
Normal file
@@ -0,0 +1,200 @@
|
||||
//
|
||||
// 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 image2[1];
|
||||
.global .align 4 .f32 intensity;
|
||||
.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_typeinfo9intensityE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
|
||||
.global .align 8 .u64 _ZN21rti_internal_register20reg_bitness_detectorE;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail0E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail1E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail2E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail3E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail4E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail5E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail6E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail7E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail8E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail9E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail0E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail1E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail2E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail3E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail4E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail5E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail6E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail7E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail8E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail9E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register14reg_rayIndex_xE;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register14reg_rayIndex_yE;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register14reg_rayIndex_zE;
|
||||
.global .align 8 .b8 _ZN21rti_internal_typename7pixelIDE[6] = {117, 105, 110, 116, 50, 0};
|
||||
.global .align 8 .b8 _ZN21rti_internal_typename10resolutionE[6] = {117, 105, 110, 116, 50, 0};
|
||||
.global .align 8 .b8 _ZN21rti_internal_typename6normalE[7] = {102, 108, 111, 97, 116, 51, 0};
|
||||
.global .align 8 .b8 _ZN21rti_internal_typename6camPosE[7] = {102, 108, 111, 97, 116, 51, 0};
|
||||
.global .align 16 .b8 _ZN21rti_internal_typename4rootE[9] = {114, 116, 79, 98, 106, 101, 99, 116, 0};
|
||||
.global .align 4 .b8 _ZN21rti_internal_typename12imageEnabledE[4] = {105, 110, 116, 0};
|
||||
.global .align 8 .b8 _ZN21rti_internal_typename8tileInfoE[6] = {117, 105, 110, 116, 52, 0};
|
||||
.global .align 4 .b8 _ZN21rti_internal_typename8additiveE[4] = {105, 110, 116, 0};
|
||||
.global .align 8 .b8 _ZN21rti_internal_typename9intensityE[6] = {102, 108, 111, 97, 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_typeenum9intensityE = 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_semantic9intensityE[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_annotation9intensityE[1];
|
||||
|
||||
.visible .entry _Z6oxMainv(
|
||||
|
||||
)
|
||||
{
|
||||
.reg .b16 %rs<10>;
|
||||
.reg .f32 %f<15>;
|
||||
.reg .b32 %r<37>;
|
||||
.reg .b64 %rd<39>;
|
||||
|
||||
|
||||
ld.global.v2.u32 {%r13, %r14}, [pixelID];
|
||||
cvt.u64.u32 %rd3, %r13;
|
||||
cvt.u64.u32 %rd4, %r14;
|
||||
mov.u64 %rd37, image;
|
||||
cvta.global.u64 %rd2, %rd37;
|
||||
mov.u32 %r11, 2;
|
||||
mov.u32 %r12, 8;
|
||||
mov.u64 %rd36, 0;
|
||||
// inline asm
|
||||
call (%rd1), _rt_buffer_get_64, (%rd2, %r11, %r12, %rd3, %rd4, %rd36, %rd36);
|
||||
// inline asm
|
||||
ld.u16 %rs1, [%rd1+6];
|
||||
// inline asm
|
||||
{ cvt.f32.f16 %f1, %rs1;}
|
||||
|
||||
// inline asm
|
||||
ld.global.v2.u32 {%r17, %r18}, [pixelID];
|
||||
cvt.u64.u32 %rd9, %r17;
|
||||
cvt.u64.u32 %rd10, %r18;
|
||||
// inline asm
|
||||
call (%rd7), _rt_buffer_get_64, (%rd2, %r11, %r12, %rd9, %rd10, %rd36, %rd36);
|
||||
// inline asm
|
||||
ld.u16 %rs2, [%rd7];
|
||||
// inline asm
|
||||
{ cvt.f32.f16 %f2, %rs2;}
|
||||
|
||||
// inline asm
|
||||
ld.global.v2.u32 {%r21, %r22}, [pixelID];
|
||||
cvt.u64.u32 %rd15, %r21;
|
||||
cvt.u64.u32 %rd16, %r22;
|
||||
// inline asm
|
||||
call (%rd13), _rt_buffer_get_64, (%rd2, %r11, %r12, %rd15, %rd16, %rd36, %rd36);
|
||||
// inline asm
|
||||
ld.u16 %rs3, [%rd13+2];
|
||||
// inline asm
|
||||
{ cvt.f32.f16 %f3, %rs3;}
|
||||
|
||||
// inline asm
|
||||
ld.global.v2.u32 {%r25, %r26}, [pixelID];
|
||||
cvt.u64.u32 %rd21, %r25;
|
||||
cvt.u64.u32 %rd22, %r26;
|
||||
// inline asm
|
||||
call (%rd19), _rt_buffer_get_64, (%rd2, %r11, %r12, %rd21, %rd22, %rd36, %rd36);
|
||||
// inline asm
|
||||
ld.u16 %rs4, [%rd19+4];
|
||||
// inline asm
|
||||
{ cvt.f32.f16 %f4, %rs4;}
|
||||
|
||||
// inline asm
|
||||
ld.global.v2.u32 {%r29, %r30}, [pixelID];
|
||||
cvt.u64.u32 %rd27, %r29;
|
||||
cvt.u64.u32 %rd28, %r30;
|
||||
mov.u64 %rd38, image2;
|
||||
cvta.global.u64 %rd26, %rd38;
|
||||
mov.u32 %r10, 1;
|
||||
// inline asm
|
||||
call (%rd25), _rt_buffer_get_64, (%rd26, %r11, %r10, %rd27, %rd28, %rd36, %rd36);
|
||||
// inline asm
|
||||
ld.u8 %rs9, [%rd25];
|
||||
cvt.rn.f32.u16 %f9, %rs9;
|
||||
div.rn.f32 %f10, %f9, 0f437F0000;
|
||||
ld.global.f32 %f11, [intensity];
|
||||
add.f32 %f12, %f10, 0fBF800000;
|
||||
fma.rn.f32 %f13, %f11, %f12, 0f3F800000;
|
||||
cvt.sat.f32.f32 %f14, %f13;
|
||||
mul.f32 %f5, %f2, %f14;
|
||||
mul.f32 %f6, %f3, %f14;
|
||||
mul.f32 %f7, %f4, %f14;
|
||||
ld.global.v2.u32 {%r33, %r34}, [pixelID];
|
||||
cvt.u64.u32 %rd33, %r33;
|
||||
cvt.u64.u32 %rd34, %r34;
|
||||
// inline asm
|
||||
call (%rd31), _rt_buffer_get_64, (%rd2, %r11, %r12, %rd33, %rd34, %rd36, %rd36);
|
||||
// inline asm
|
||||
// inline asm
|
||||
{ cvt.rn.f16.f32 %rs7, %f7;}
|
||||
|
||||
// inline asm
|
||||
// inline asm
|
||||
{ cvt.rn.f16.f32 %rs6, %f6;}
|
||||
|
||||
// inline asm
|
||||
// inline asm
|
||||
{ cvt.rn.f16.f32 %rs5, %f5;}
|
||||
|
||||
// inline asm
|
||||
// inline asm
|
||||
{ cvt.rn.f16.f32 %rs8, %f1;}
|
||||
|
||||
// inline asm
|
||||
st.v4.u16 [%rd31], {%rs5, %rs6, %rs7, %rs8};
|
||||
ret;
|
||||
}
|
||||
|
||||
|
||||
8
Assets/Editor/x64/Bakery/addAO.ptx.meta
Normal file
8
Assets/Editor/x64/Bakery/addAO.ptx.meta
Normal file
@@ -0,0 +1,8 @@
|
||||
fileFormatVersion: 2
|
||||
guid: aa2c6f245990b69408c255eb48c206eb
|
||||
timeCreated: 1534430149
|
||||
licenseType: Store
|
||||
DefaultImporter:
|
||||
userData:
|
||||
assetBundleName:
|
||||
assetBundleVariant:
|
||||
327
Assets/Editor/x64/Bakery/addDir.ptx
Normal file
327
Assets/Editor/x64/Bakery/addDir.ptx
Normal file
@@ -0,0 +1,327 @@
|
||||
//
|
||||
// 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 uvnormal[1];
|
||||
.global .align 1 .b8 image[1];
|
||||
.global .align 1 .b8 image2[1];
|
||||
.global .align 1 .b8 imageHDR[1];
|
||||
.global .align 1 .b8 imageFinal[1];
|
||||
.global .align 4 .f32 DoNormalize;
|
||||
.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_typeinfo11DoNormalizeE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
|
||||
.global .align 8 .u64 _ZN21rti_internal_register20reg_bitness_detectorE;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail0E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail1E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail2E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail3E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail4E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail5E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail6E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail7E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail8E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail9E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail0E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail1E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail2E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail3E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail4E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail5E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail6E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail7E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail8E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail9E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register14reg_rayIndex_xE;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register14reg_rayIndex_yE;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register14reg_rayIndex_zE;
|
||||
.global .align 8 .b8 _ZN21rti_internal_typename7pixelIDE[6] = {117, 105, 110, 116, 50, 0};
|
||||
.global .align 8 .b8 _ZN21rti_internal_typename10resolutionE[6] = {117, 105, 110, 116, 50, 0};
|
||||
.global .align 8 .b8 _ZN21rti_internal_typename6normalE[7] = {102, 108, 111, 97, 116, 51, 0};
|
||||
.global .align 8 .b8 _ZN21rti_internal_typename6camPosE[7] = {102, 108, 111, 97, 116, 51, 0};
|
||||
.global .align 16 .b8 _ZN21rti_internal_typename4rootE[9] = {114, 116, 79, 98, 106, 101, 99, 116, 0};
|
||||
.global .align 4 .b8 _ZN21rti_internal_typename12imageEnabledE[4] = {105, 110, 116, 0};
|
||||
.global .align 8 .b8 _ZN21rti_internal_typename8tileInfoE[6] = {117, 105, 110, 116, 52, 0};
|
||||
.global .align 4 .b8 _ZN21rti_internal_typename8additiveE[4] = {105, 110, 116, 0};
|
||||
.global .align 8 .b8 _ZN21rti_internal_typename11DoNormalizeE[6] = {102, 108, 111, 97, 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_typeenum11DoNormalizeE = 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_semantic11DoNormalizeE[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_annotation11DoNormalizeE[1];
|
||||
|
||||
.visible .entry _Z6oxMainv(
|
||||
|
||||
)
|
||||
{
|
||||
.reg .pred %p<4>;
|
||||
.reg .b16 %rs<20>;
|
||||
.reg .f32 %f<88>;
|
||||
.reg .b32 %r<79>;
|
||||
.reg .b64 %rd<80>;
|
||||
|
||||
|
||||
ld.global.v2.u32 {%r20, %r21}, [pixelID];
|
||||
cvt.u64.u32 %rd4, %r20;
|
||||
cvt.u64.u32 %rd5, %r21;
|
||||
mov.u64 %rd56, imageHDR;
|
||||
cvta.global.u64 %rd3, %rd56;
|
||||
mov.u32 %r18, 2;
|
||||
mov.u32 %r7, 8;
|
||||
mov.u64 %rd55, 0;
|
||||
// inline asm
|
||||
call (%rd2), _rt_buffer_get_64, (%rd3, %r18, %r7, %rd4, %rd5, %rd55, %rd55);
|
||||
// inline asm
|
||||
ld.u16 %rs2, [%rd2];
|
||||
// inline asm
|
||||
{ cvt.f32.f16 %f17, %rs2;}
|
||||
|
||||
// inline asm
|
||||
ld.global.v2.u32 {%r24, %r25}, [pixelID];
|
||||
cvt.u64.u32 %rd10, %r24;
|
||||
cvt.u64.u32 %rd11, %r25;
|
||||
// inline asm
|
||||
call (%rd8), _rt_buffer_get_64, (%rd3, %r18, %r7, %rd10, %rd11, %rd55, %rd55);
|
||||
// inline asm
|
||||
ld.u16 %rs3, [%rd8+2];
|
||||
// inline asm
|
||||
{ cvt.f32.f16 %f18, %rs3;}
|
||||
|
||||
// inline asm
|
||||
ld.global.v2.u32 {%r28, %r29}, [pixelID];
|
||||
cvt.u64.u32 %rd16, %r28;
|
||||
cvt.u64.u32 %rd17, %r29;
|
||||
// inline asm
|
||||
call (%rd14), _rt_buffer_get_64, (%rd3, %r18, %r7, %rd16, %rd17, %rd55, %rd55);
|
||||
// inline asm
|
||||
ld.u16 %rs4, [%rd14+4];
|
||||
// inline asm
|
||||
{ cvt.f32.f16 %f19, %rs4;}
|
||||
|
||||
// inline asm
|
||||
mul.f32 %f20, %f18, 0f3F372474;
|
||||
fma.rn.f32 %f21, %f17, 0f3E59999A, %f20;
|
||||
fma.rn.f32 %f22, %f19, 0f3D93A92A, %f21;
|
||||
ld.global.v2.u32 {%r32, %r33}, [pixelID];
|
||||
cvt.u64.u32 %rd22, %r32;
|
||||
cvt.u64.u32 %rd23, %r33;
|
||||
mov.u64 %rd57, image2;
|
||||
cvta.global.u64 %rd21, %rd57;
|
||||
mov.u32 %r19, 16;
|
||||
// inline asm
|
||||
call (%rd20), _rt_buffer_get_64, (%rd21, %r18, %r19, %rd22, %rd23, %rd55, %rd55);
|
||||
// inline asm
|
||||
ld.v4.f32 {%f23, %f24, %f25, %f26}, [%rd20];
|
||||
ld.global.v2.u32 {%r36, %r37}, [pixelID];
|
||||
cvt.u64.u32 %rd28, %r36;
|
||||
cvt.u64.u32 %rd29, %r37;
|
||||
mov.u64 %rd58, image;
|
||||
cvta.global.u64 %rd27, %rd58;
|
||||
mov.u32 %r17, 4;
|
||||
// inline asm
|
||||
call (%rd26), _rt_buffer_get_64, (%rd27, %r18, %r17, %rd28, %rd29, %rd55, %rd55);
|
||||
// inline asm
|
||||
ld.u8 %rs5, [%rd26];
|
||||
cvt.rn.f32.u16 %f30, %rs5;
|
||||
div.rn.f32 %f31, %f30, 0f437F0000;
|
||||
fma.rn.f32 %f32, %f31, 0f40000000, 0fBF800000;
|
||||
ld.global.v2.u32 {%r40, %r41}, [pixelID];
|
||||
cvt.u64.u32 %rd34, %r40;
|
||||
cvt.u64.u32 %rd35, %r41;
|
||||
// inline asm
|
||||
call (%rd32), _rt_buffer_get_64, (%rd27, %r18, %r17, %rd34, %rd35, %rd55, %rd55);
|
||||
// inline asm
|
||||
ld.u8 %rs6, [%rd32+1];
|
||||
cvt.rn.f32.u16 %f33, %rs6;
|
||||
div.rn.f32 %f34, %f33, 0f437F0000;
|
||||
fma.rn.f32 %f35, %f34, 0f40000000, 0fBF800000;
|
||||
ld.global.v2.u32 {%r44, %r45}, [pixelID];
|
||||
cvt.u64.u32 %rd40, %r44;
|
||||
cvt.u64.u32 %rd41, %r45;
|
||||
// inline asm
|
||||
call (%rd38), _rt_buffer_get_64, (%rd27, %r18, %r17, %rd40, %rd41, %rd55, %rd55);
|
||||
// inline asm
|
||||
ld.u8 %rs7, [%rd38+2];
|
||||
cvt.rn.f32.u16 %f36, %rs7;
|
||||
div.rn.f32 %f37, %f36, 0f437F0000;
|
||||
fma.rn.f32 %f38, %f37, 0f40000000, 0fBF800000;
|
||||
mul.f32 %f39, %f35, %f35;
|
||||
fma.rn.f32 %f40, %f32, %f32, %f39;
|
||||
fma.rn.f32 %f41, %f38, %f38, %f40;
|
||||
sqrt.rn.f32 %f42, %f41;
|
||||
rcp.rn.f32 %f43, %f42;
|
||||
mul.f32 %f44, %f32, %f43;
|
||||
mul.f32 %f45, %f35, %f43;
|
||||
mul.f32 %f46, %f38, %f43;
|
||||
fma.rn.f32 %f1, %f22, %f44, %f23;
|
||||
fma.rn.f32 %f2, %f22, %f45, %f24;
|
||||
fma.rn.f32 %f3, %f22, %f46, %f25;
|
||||
ld.global.v2.u32 {%r48, %r49}, [pixelID];
|
||||
cvt.u64.u32 %rd46, %r48;
|
||||
cvt.u64.u32 %rd47, %r49;
|
||||
// inline asm
|
||||
call (%rd44), _rt_buffer_get_64, (%rd27, %r18, %r17, %rd46, %rd47, %rd55, %rd55);
|
||||
// inline asm
|
||||
ld.u8 %rs8, [%rd44+3];
|
||||
cvt.rn.f32.u16 %f47, %rs8;
|
||||
div.rn.f32 %f48, %f47, 0f437F0000;
|
||||
ld.global.v2.u32 {%r52, %r53}, [pixelID];
|
||||
cvt.u64.u32 %rd52, %r52;
|
||||
cvt.u64.u32 %rd53, %r53;
|
||||
// inline asm
|
||||
call (%rd50), _rt_buffer_get_64, (%rd21, %r18, %r19, %rd52, %rd53, %rd55, %rd55);
|
||||
// inline asm
|
||||
ld.f32 %f49, [%rd50+12];
|
||||
min.f32 %f87, %f48, %f49;
|
||||
ld.global.f32 %f50, [DoNormalize];
|
||||
setp.gt.f32 %p1, %f50, 0f3F000000;
|
||||
@%p1 bra BB0_2;
|
||||
bra.uni BB0_1;
|
||||
|
||||
BB0_2:
|
||||
mul.f32 %f54, %f2, %f2;
|
||||
fma.rn.f32 %f55, %f1, %f1, %f54;
|
||||
fma.rn.f32 %f56, %f3, %f3, %f55;
|
||||
sqrt.rn.f32 %f57, %f56;
|
||||
rcp.rn.f32 %f58, %f57;
|
||||
mul.f32 %f5, %f1, %f58;
|
||||
mul.f32 %f6, %f2, %f58;
|
||||
mul.f32 %f7, %f3, %f58;
|
||||
ld.global.v2.u32 {%r64, %r65}, [pixelID];
|
||||
cvt.u64.u32 %rd68, %r64;
|
||||
cvt.u64.u32 %rd69, %r65;
|
||||
mov.u64 %rd72, uvnormal;
|
||||
cvta.global.u64 %rd67, %rd72;
|
||||
// inline asm
|
||||
call (%rd66), _rt_buffer_get_64, (%rd67, %r18, %r17, %rd68, %rd69, %rd55, %rd55);
|
||||
// inline asm
|
||||
ld.u32 %r1, [%rd66];
|
||||
shr.u32 %r68, %r1, 16;
|
||||
cvt.u16.u32 %rs1, %r68;
|
||||
and.b16 %rs9, %rs1, 255;
|
||||
cvt.u16.u32 %rs10, %r1;
|
||||
or.b16 %rs11, %rs10, %rs9;
|
||||
setp.eq.s16 %p2, %rs11, 0;
|
||||
mov.f32 %f84, 0f00000000;
|
||||
mov.f32 %f85, %f84;
|
||||
mov.f32 %f86, %f84;
|
||||
@%p2 bra BB0_4;
|
||||
|
||||
ld.u8 %rs12, [%rd66+1];
|
||||
and.b16 %rs14, %rs10, 255;
|
||||
cvt.rn.f32.u16 %f59, %rs14;
|
||||
div.rn.f32 %f60, %f59, 0f437F0000;
|
||||
fma.rn.f32 %f61, %f60, 0f40000000, 0fBF800000;
|
||||
cvt.rn.f32.u16 %f62, %rs12;
|
||||
div.rn.f32 %f63, %f62, 0f437F0000;
|
||||
fma.rn.f32 %f64, %f63, 0f40000000, 0fBF800000;
|
||||
cvt.rn.f32.u16 %f65, %rs9;
|
||||
div.rn.f32 %f66, %f65, 0f437F0000;
|
||||
fma.rn.f32 %f67, %f66, 0f40000000, 0fBF800000;
|
||||
mul.f32 %f68, %f64, %f64;
|
||||
fma.rn.f32 %f69, %f61, %f61, %f68;
|
||||
fma.rn.f32 %f70, %f67, %f67, %f69;
|
||||
sqrt.rn.f32 %f71, %f70;
|
||||
rcp.rn.f32 %f72, %f71;
|
||||
mul.f32 %f84, %f61, %f72;
|
||||
mul.f32 %f85, %f64, %f72;
|
||||
mul.f32 %f86, %f67, %f72;
|
||||
|
||||
BB0_4:
|
||||
mul.f32 %f73, %f6, %f85;
|
||||
fma.rn.f32 %f74, %f5, %f84, %f73;
|
||||
fma.rn.f32 %f14, %f7, %f86, %f74;
|
||||
setp.leu.f32 %p3, %f87, 0f00000000;
|
||||
@%p3 bra BB0_6;
|
||||
|
||||
fma.rn.f32 %f75, %f14, 0f3F000000, 0f3F000000;
|
||||
mov.f32 %f76, 0f3B808081;
|
||||
max.f32 %f87, %f75, %f76;
|
||||
|
||||
BB0_6:
|
||||
ld.global.v2.u32 {%r71, %r72}, [pixelID];
|
||||
cvt.u64.u32 %rd75, %r71;
|
||||
cvt.u64.u32 %rd76, %r72;
|
||||
mov.u64 %rd79, imageFinal;
|
||||
cvta.global.u64 %rd74, %rd79;
|
||||
// inline asm
|
||||
call (%rd73), _rt_buffer_get_64, (%rd74, %r18, %r17, %rd75, %rd76, %rd55, %rd55);
|
||||
// inline asm
|
||||
fma.rn.f32 %f77, %f5, 0f3F000000, 0f3F000000;
|
||||
mul.f32 %f78, %f77, 0f437F0000;
|
||||
cvt.rzi.u32.f32 %r75, %f78;
|
||||
fma.rn.f32 %f79, %f6, 0f3F000000, 0f3F000000;
|
||||
mul.f32 %f80, %f79, 0f437F0000;
|
||||
cvt.rzi.u32.f32 %r76, %f80;
|
||||
fma.rn.f32 %f81, %f7, 0f3F000000, 0f3F000000;
|
||||
mul.f32 %f82, %f81, 0f437F0000;
|
||||
cvt.rzi.u32.f32 %r77, %f82;
|
||||
mul.f32 %f83, %f87, 0f437F0000;
|
||||
cvt.rzi.u32.f32 %r78, %f83;
|
||||
cvt.u16.u32 %rs16, %r78;
|
||||
cvt.u16.u32 %rs17, %r77;
|
||||
cvt.u16.u32 %rs18, %r76;
|
||||
cvt.u16.u32 %rs19, %r75;
|
||||
st.v4.u8 [%rd73], {%rs19, %rs18, %rs17, %rs16};
|
||||
bra.uni BB0_7;
|
||||
|
||||
BB0_1:
|
||||
ld.global.v2.u32 {%r58, %r59}, [pixelID];
|
||||
cvt.u64.u32 %rd61, %r58;
|
||||
cvt.u64.u32 %rd62, %r59;
|
||||
// inline asm
|
||||
call (%rd59), _rt_buffer_get_64, (%rd21, %r18, %r19, %rd61, %rd62, %rd55, %rd55);
|
||||
// inline asm
|
||||
st.v4.f32 [%rd59], {%f1, %f2, %f3, %f87};
|
||||
|
||||
BB0_7:
|
||||
ret;
|
||||
}
|
||||
|
||||
|
||||
8
Assets/Editor/x64/Bakery/addDir.ptx.meta
Normal file
8
Assets/Editor/x64/Bakery/addDir.ptx.meta
Normal file
@@ -0,0 +1,8 @@
|
||||
fileFormatVersion: 2
|
||||
guid: 2610b17ae27a5bb4493e1491e8e5dc4d
|
||||
timeCreated: 1537442170
|
||||
licenseType: Store
|
||||
DefaultImporter:
|
||||
userData:
|
||||
assetBundleName:
|
||||
assetBundleVariant:
|
||||
228
Assets/Editor/x64/Bakery/addHalf.ptx
Normal file
228
Assets/Editor/x64/Bakery/addHalf.ptx
Normal file
@@ -0,0 +1,228 @@
|
||||
//
|
||||
// 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 image2[1];
|
||||
.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 8 .u64 _ZN21rti_internal_register20reg_bitness_detectorE;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail0E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail1E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail2E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail3E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail4E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail5E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail6E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail7E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail8E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail9E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail0E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail1E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail2E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail3E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail4E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail5E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail6E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail7E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail8E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail9E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register14reg_rayIndex_xE;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register14reg_rayIndex_yE;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register14reg_rayIndex_zE;
|
||||
.global .align 8 .b8 _ZN21rti_internal_typename7pixelIDE[6] = {117, 105, 110, 116, 50, 0};
|
||||
.global .align 8 .b8 _ZN21rti_internal_typename10resolutionE[6] = {117, 105, 110, 116, 50, 0};
|
||||
.global .align 8 .b8 _ZN21rti_internal_typename6normalE[7] = {102, 108, 111, 97, 116, 51, 0};
|
||||
.global .align 8 .b8 _ZN21rti_internal_typename6camPosE[7] = {102, 108, 111, 97, 116, 51, 0};
|
||||
.global .align 16 .b8 _ZN21rti_internal_typename4rootE[9] = {114, 116, 79, 98, 106, 101, 99, 116, 0};
|
||||
.global .align 4 .b8 _ZN21rti_internal_typename12imageEnabledE[4] = {105, 110, 116, 0};
|
||||
.global .align 8 .b8 _ZN21rti_internal_typename8tileInfoE[6] = {117, 105, 110, 116, 52, 0};
|
||||
.global .align 4 .b8 _ZN21rti_internal_typename8additiveE[4] = {105, 110, 116, 0};
|
||||
.global .align 4 .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 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 _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];
|
||||
|
||||
.visible .entry _Z6oxMainv(
|
||||
|
||||
)
|
||||
{
|
||||
.reg .b16 %rs<13>;
|
||||
.reg .f32 %f<16>;
|
||||
.reg .b32 %r<55>;
|
||||
.reg .b64 %rd<57>;
|
||||
|
||||
|
||||
ld.global.v2.u32 {%r19, %r20}, [pixelID];
|
||||
cvt.u64.u32 %rd3, %r19;
|
||||
cvt.u64.u32 %rd4, %r20;
|
||||
mov.u64 %rd55, image;
|
||||
cvta.global.u64 %rd2, %rd55;
|
||||
mov.u32 %r17, 2;
|
||||
mov.u32 %r18, 8;
|
||||
mov.u64 %rd54, 0;
|
||||
// inline asm
|
||||
call (%rd1), _rt_buffer_get_64, (%rd2, %r17, %r18, %rd3, %rd4, %rd54, %rd54);
|
||||
// inline asm
|
||||
ld.u16 %rs1, [%rd1+6];
|
||||
// inline asm
|
||||
{ cvt.f32.f16 %f1, %rs1;}
|
||||
|
||||
// inline asm
|
||||
ld.global.v2.u32 {%r23, %r24}, [pixelID];
|
||||
cvt.u64.u32 %rd9, %r23;
|
||||
cvt.u64.u32 %rd10, %r24;
|
||||
// inline asm
|
||||
call (%rd7), _rt_buffer_get_64, (%rd2, %r17, %r18, %rd9, %rd10, %rd54, %rd54);
|
||||
// inline asm
|
||||
ld.u16 %rs2, [%rd7];
|
||||
// inline asm
|
||||
{ cvt.f32.f16 %f2, %rs2;}
|
||||
|
||||
// inline asm
|
||||
ld.global.v2.u32 {%r27, %r28}, [pixelID];
|
||||
cvt.u64.u32 %rd15, %r27;
|
||||
cvt.u64.u32 %rd16, %r28;
|
||||
// inline asm
|
||||
call (%rd13), _rt_buffer_get_64, (%rd2, %r17, %r18, %rd15, %rd16, %rd54, %rd54);
|
||||
// inline asm
|
||||
ld.u16 %rs3, [%rd13+2];
|
||||
// inline asm
|
||||
{ cvt.f32.f16 %f3, %rs3;}
|
||||
|
||||
// inline asm
|
||||
ld.global.v2.u32 {%r31, %r32}, [pixelID];
|
||||
cvt.u64.u32 %rd21, %r31;
|
||||
cvt.u64.u32 %rd22, %r32;
|
||||
// inline asm
|
||||
call (%rd19), _rt_buffer_get_64, (%rd2, %r17, %r18, %rd21, %rd22, %rd54, %rd54);
|
||||
// inline asm
|
||||
ld.u16 %rs4, [%rd19+4];
|
||||
// inline asm
|
||||
{ cvt.f32.f16 %f4, %rs4;}
|
||||
|
||||
// inline asm
|
||||
ld.global.v2.u32 {%r35, %r36}, [pixelID];
|
||||
cvt.u64.u32 %rd27, %r35;
|
||||
cvt.u64.u32 %rd28, %r36;
|
||||
mov.u64 %rd56, image2;
|
||||
cvta.global.u64 %rd26, %rd56;
|
||||
// inline asm
|
||||
call (%rd25), _rt_buffer_get_64, (%rd26, %r17, %r18, %rd27, %rd28, %rd54, %rd54);
|
||||
// inline asm
|
||||
ld.u16 %rs5, [%rd25+6];
|
||||
// inline asm
|
||||
{ cvt.f32.f16 %f5, %rs5;}
|
||||
|
||||
// inline asm
|
||||
ld.global.v2.u32 {%r39, %r40}, [pixelID];
|
||||
cvt.u64.u32 %rd33, %r39;
|
||||
cvt.u64.u32 %rd34, %r40;
|
||||
// inline asm
|
||||
call (%rd31), _rt_buffer_get_64, (%rd26, %r17, %r18, %rd33, %rd34, %rd54, %rd54);
|
||||
// inline asm
|
||||
ld.u16 %rs6, [%rd31];
|
||||
// inline asm
|
||||
{ cvt.f32.f16 %f6, %rs6;}
|
||||
|
||||
// inline asm
|
||||
ld.global.v2.u32 {%r43, %r44}, [pixelID];
|
||||
cvt.u64.u32 %rd39, %r43;
|
||||
cvt.u64.u32 %rd40, %r44;
|
||||
// inline asm
|
||||
call (%rd37), _rt_buffer_get_64, (%rd26, %r17, %r18, %rd39, %rd40, %rd54, %rd54);
|
||||
// inline asm
|
||||
ld.u16 %rs7, [%rd37+2];
|
||||
// inline asm
|
||||
{ cvt.f32.f16 %f7, %rs7;}
|
||||
|
||||
// inline asm
|
||||
ld.global.v2.u32 {%r47, %r48}, [pixelID];
|
||||
cvt.u64.u32 %rd45, %r47;
|
||||
cvt.u64.u32 %rd46, %r48;
|
||||
// inline asm
|
||||
call (%rd43), _rt_buffer_get_64, (%rd26, %r17, %r18, %rd45, %rd46, %rd54, %rd54);
|
||||
// inline asm
|
||||
ld.u16 %rs8, [%rd43+4];
|
||||
// inline asm
|
||||
{ cvt.f32.f16 %f8, %rs8;}
|
||||
|
||||
// inline asm
|
||||
add.f32 %f13, %f2, %f6;
|
||||
add.f32 %f14, %f3, %f7;
|
||||
add.f32 %f15, %f4, %f8;
|
||||
mul.f32 %f9, %f1, %f13;
|
||||
mul.f32 %f10, %f1, %f14;
|
||||
mul.f32 %f11, %f1, %f15;
|
||||
min.f32 %f12, %f1, %f5;
|
||||
ld.global.v2.u32 {%r51, %r52}, [pixelID];
|
||||
cvt.u64.u32 %rd51, %r51;
|
||||
cvt.u64.u32 %rd52, %r52;
|
||||
// inline asm
|
||||
call (%rd49), _rt_buffer_get_64, (%rd26, %r17, %r18, %rd51, %rd52, %rd54, %rd54);
|
||||
// inline asm
|
||||
// inline asm
|
||||
{ cvt.rn.f16.f32 %rs12, %f12;}
|
||||
|
||||
// inline asm
|
||||
// inline asm
|
||||
{ cvt.rn.f16.f32 %rs11, %f11;}
|
||||
|
||||
// inline asm
|
||||
// inline asm
|
||||
{ cvt.rn.f16.f32 %rs10, %f10;}
|
||||
|
||||
// inline asm
|
||||
// inline asm
|
||||
{ cvt.rn.f16.f32 %rs9, %f9;}
|
||||
|
||||
// inline asm
|
||||
st.v4.u16 [%rd49], {%rs9, %rs10, %rs11, %rs12};
|
||||
ret;
|
||||
}
|
||||
|
||||
|
||||
8
Assets/Editor/x64/Bakery/addHalf.ptx.meta
Normal file
8
Assets/Editor/x64/Bakery/addHalf.ptx.meta
Normal file
@@ -0,0 +1,8 @@
|
||||
fileFormatVersion: 2
|
||||
guid: e11fe20f1d7fb5c4e8a3e879f3aef436
|
||||
timeCreated: 1526161016
|
||||
licenseType: Store
|
||||
DefaultImporter:
|
||||
userData:
|
||||
assetBundleName:
|
||||
assetBundleVariant:
|
||||
290
Assets/Editor/x64/Bakery/addHalfSeparateIndirect.ptx
Normal file
290
Assets/Editor/x64/Bakery/addHalfSeparateIndirect.ptx
Normal file
@@ -0,0 +1,290 @@
|
||||
//
|
||||
// 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 image2[1];
|
||||
.global .align 1 .b8 imageIndirect[1];
|
||||
.global .align 4 .f32 indirectIntensity;
|
||||
.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_typeinfo17indirectIntensityE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
|
||||
.global .align 8 .u64 _ZN21rti_internal_register20reg_bitness_detectorE;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail0E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail1E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail2E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail3E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail4E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail5E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail6E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail7E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail8E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail9E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail0E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail1E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail2E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail3E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail4E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail5E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail6E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail7E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail8E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail9E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register14reg_rayIndex_xE;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register14reg_rayIndex_yE;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register14reg_rayIndex_zE;
|
||||
.global .align 8 .b8 _ZN21rti_internal_typename7pixelIDE[6] = {117, 105, 110, 116, 50, 0};
|
||||
.global .align 8 .b8 _ZN21rti_internal_typename10resolutionE[6] = {117, 105, 110, 116, 50, 0};
|
||||
.global .align 8 .b8 _ZN21rti_internal_typename6normalE[7] = {102, 108, 111, 97, 116, 51, 0};
|
||||
.global .align 8 .b8 _ZN21rti_internal_typename6camPosE[7] = {102, 108, 111, 97, 116, 51, 0};
|
||||
.global .align 16 .b8 _ZN21rti_internal_typename4rootE[9] = {114, 116, 79, 98, 106, 101, 99, 116, 0};
|
||||
.global .align 4 .b8 _ZN21rti_internal_typename12imageEnabledE[4] = {105, 110, 116, 0};
|
||||
.global .align 8 .b8 _ZN21rti_internal_typename8tileInfoE[6] = {117, 105, 110, 116, 52, 0};
|
||||
.global .align 4 .b8 _ZN21rti_internal_typename8additiveE[4] = {105, 110, 116, 0};
|
||||
.global .align 8 .b8 _ZN21rti_internal_typename17indirectIntensityE[6] = {102, 108, 111, 97, 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_typeenum17indirectIntensityE = 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_semantic17indirectIntensityE[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_annotation17indirectIntensityE[1];
|
||||
|
||||
.visible .entry _Z6oxMainv(
|
||||
|
||||
)
|
||||
{
|
||||
.reg .b16 %rs<19>;
|
||||
.reg .f32 %f<20>;
|
||||
.reg .b32 %r<79>;
|
||||
.reg .b64 %rd<82>;
|
||||
|
||||
|
||||
ld.global.v2.u32 {%r27, %r28}, [pixelID];
|
||||
cvt.u64.u32 %rd3, %r27;
|
||||
cvt.u64.u32 %rd4, %r28;
|
||||
mov.u64 %rd79, image;
|
||||
cvta.global.u64 %rd2, %rd79;
|
||||
mov.u32 %r25, 2;
|
||||
mov.u32 %r26, 8;
|
||||
mov.u64 %rd78, 0;
|
||||
// inline asm
|
||||
call (%rd1), _rt_buffer_get_64, (%rd2, %r25, %r26, %rd3, %rd4, %rd78, %rd78);
|
||||
// inline asm
|
||||
ld.u16 %rs1, [%rd1+6];
|
||||
// inline asm
|
||||
{ cvt.f32.f16 %f1, %rs1;}
|
||||
|
||||
// inline asm
|
||||
ld.global.v2.u32 {%r31, %r32}, [pixelID];
|
||||
cvt.u64.u32 %rd9, %r31;
|
||||
cvt.u64.u32 %rd10, %r32;
|
||||
// inline asm
|
||||
call (%rd7), _rt_buffer_get_64, (%rd2, %r25, %r26, %rd9, %rd10, %rd78, %rd78);
|
||||
// inline asm
|
||||
ld.u16 %rs2, [%rd7];
|
||||
// inline asm
|
||||
{ cvt.f32.f16 %f2, %rs2;}
|
||||
|
||||
// inline asm
|
||||
ld.global.v2.u32 {%r35, %r36}, [pixelID];
|
||||
cvt.u64.u32 %rd15, %r35;
|
||||
cvt.u64.u32 %rd16, %r36;
|
||||
// inline asm
|
||||
call (%rd13), _rt_buffer_get_64, (%rd2, %r25, %r26, %rd15, %rd16, %rd78, %rd78);
|
||||
// inline asm
|
||||
ld.u16 %rs3, [%rd13+2];
|
||||
// inline asm
|
||||
{ cvt.f32.f16 %f3, %rs3;}
|
||||
|
||||
// inline asm
|
||||
ld.global.v2.u32 {%r39, %r40}, [pixelID];
|
||||
cvt.u64.u32 %rd21, %r39;
|
||||
cvt.u64.u32 %rd22, %r40;
|
||||
// inline asm
|
||||
call (%rd19), _rt_buffer_get_64, (%rd2, %r25, %r26, %rd21, %rd22, %rd78, %rd78);
|
||||
// inline asm
|
||||
ld.u16 %rs4, [%rd19+4];
|
||||
// inline asm
|
||||
{ cvt.f32.f16 %f4, %rs4;}
|
||||
|
||||
// inline asm
|
||||
ld.global.v2.u32 {%r43, %r44}, [pixelID];
|
||||
cvt.u64.u32 %rd27, %r43;
|
||||
cvt.u64.u32 %rd28, %r44;
|
||||
mov.u64 %rd80, image2;
|
||||
cvta.global.u64 %rd26, %rd80;
|
||||
// inline asm
|
||||
call (%rd25), _rt_buffer_get_64, (%rd26, %r25, %r26, %rd27, %rd28, %rd78, %rd78);
|
||||
// inline asm
|
||||
ld.u16 %rs5, [%rd25+6];
|
||||
// inline asm
|
||||
{ cvt.f32.f16 %f5, %rs5;}
|
||||
|
||||
// inline asm
|
||||
ld.global.v2.u32 {%r47, %r48}, [pixelID];
|
||||
cvt.u64.u32 %rd33, %r47;
|
||||
cvt.u64.u32 %rd34, %r48;
|
||||
// inline asm
|
||||
call (%rd31), _rt_buffer_get_64, (%rd26, %r25, %r26, %rd33, %rd34, %rd78, %rd78);
|
||||
// inline asm
|
||||
ld.u16 %rs6, [%rd31];
|
||||
// inline asm
|
||||
{ cvt.f32.f16 %f6, %rs6;}
|
||||
|
||||
// inline asm
|
||||
ld.global.v2.u32 {%r51, %r52}, [pixelID];
|
||||
cvt.u64.u32 %rd39, %r51;
|
||||
cvt.u64.u32 %rd40, %r52;
|
||||
// inline asm
|
||||
call (%rd37), _rt_buffer_get_64, (%rd26, %r25, %r26, %rd39, %rd40, %rd78, %rd78);
|
||||
// inline asm
|
||||
ld.u16 %rs7, [%rd37+2];
|
||||
// inline asm
|
||||
{ cvt.f32.f16 %f7, %rs7;}
|
||||
|
||||
// inline asm
|
||||
ld.global.v2.u32 {%r55, %r56}, [pixelID];
|
||||
cvt.u64.u32 %rd45, %r55;
|
||||
cvt.u64.u32 %rd46, %r56;
|
||||
// inline asm
|
||||
call (%rd43), _rt_buffer_get_64, (%rd26, %r25, %r26, %rd45, %rd46, %rd78, %rd78);
|
||||
// inline asm
|
||||
ld.u16 %rs8, [%rd43+4];
|
||||
// inline asm
|
||||
{ cvt.f32.f16 %f8, %rs8;}
|
||||
|
||||
// inline asm
|
||||
add.f32 %f9, %f2, %f6;
|
||||
add.f32 %f10, %f3, %f7;
|
||||
add.f32 %f11, %f4, %f8;
|
||||
min.f32 %f12, %f1, %f5;
|
||||
ld.global.v2.u32 {%r59, %r60}, [pixelID];
|
||||
cvt.u64.u32 %rd51, %r59;
|
||||
cvt.u64.u32 %rd52, %r60;
|
||||
// inline asm
|
||||
call (%rd49), _rt_buffer_get_64, (%rd26, %r25, %r26, %rd51, %rd52, %rd78, %rd78);
|
||||
// inline asm
|
||||
// inline asm
|
||||
{ cvt.rn.f16.f32 %rs11, %f11;}
|
||||
|
||||
// inline asm
|
||||
// inline asm
|
||||
{ cvt.rn.f16.f32 %rs10, %f10;}
|
||||
|
||||
// inline asm
|
||||
// inline asm
|
||||
{ cvt.rn.f16.f32 %rs9, %f9;}
|
||||
|
||||
// inline asm
|
||||
// inline asm
|
||||
{ cvt.rn.f16.f32 %rs12, %f12;}
|
||||
|
||||
// inline asm
|
||||
st.v4.u16 [%rd49], {%rs9, %rs10, %rs11, %rs12};
|
||||
ld.global.v2.u32 {%r63, %r64}, [pixelID];
|
||||
cvt.u64.u32 %rd57, %r63;
|
||||
cvt.u64.u32 %rd58, %r64;
|
||||
mov.u64 %rd81, imageIndirect;
|
||||
cvta.global.u64 %rd56, %rd81;
|
||||
// inline asm
|
||||
call (%rd55), _rt_buffer_get_64, (%rd56, %r25, %r26, %rd57, %rd58, %rd78, %rd78);
|
||||
// inline asm
|
||||
ld.u16 %rs13, [%rd55];
|
||||
// inline asm
|
||||
{ cvt.f32.f16 %f13, %rs13;}
|
||||
|
||||
// inline asm
|
||||
ld.global.v2.u32 {%r67, %r68}, [pixelID];
|
||||
cvt.u64.u32 %rd63, %r67;
|
||||
cvt.u64.u32 %rd64, %r68;
|
||||
// inline asm
|
||||
call (%rd61), _rt_buffer_get_64, (%rd56, %r25, %r26, %rd63, %rd64, %rd78, %rd78);
|
||||
// inline asm
|
||||
ld.u16 %rs14, [%rd61+2];
|
||||
// inline asm
|
||||
{ cvt.f32.f16 %f14, %rs14;}
|
||||
|
||||
// inline asm
|
||||
ld.global.v2.u32 {%r71, %r72}, [pixelID];
|
||||
cvt.u64.u32 %rd69, %r71;
|
||||
cvt.u64.u32 %rd70, %r72;
|
||||
// inline asm
|
||||
call (%rd67), _rt_buffer_get_64, (%rd56, %r25, %r26, %rd69, %rd70, %rd78, %rd78);
|
||||
// inline asm
|
||||
ld.u16 %rs15, [%rd67+4];
|
||||
// inline asm
|
||||
{ cvt.f32.f16 %f15, %rs15;}
|
||||
|
||||
// inline asm
|
||||
ld.global.f32 %f19, [indirectIntensity];
|
||||
fma.rn.f32 %f16, %f2, %f19, %f13;
|
||||
fma.rn.f32 %f17, %f3, %f19, %f14;
|
||||
fma.rn.f32 %f18, %f4, %f19, %f15;
|
||||
ld.global.v2.u32 {%r75, %r76}, [pixelID];
|
||||
cvt.u64.u32 %rd75, %r75;
|
||||
cvt.u64.u32 %rd76, %r76;
|
||||
// inline asm
|
||||
call (%rd73), _rt_buffer_get_64, (%rd56, %r25, %r26, %rd75, %rd76, %rd78, %rd78);
|
||||
// inline asm
|
||||
// inline asm
|
||||
{ cvt.rn.f16.f32 %rs18, %f18;}
|
||||
|
||||
// inline asm
|
||||
// inline asm
|
||||
{ cvt.rn.f16.f32 %rs17, %f17;}
|
||||
|
||||
// inline asm
|
||||
// inline asm
|
||||
{ cvt.rn.f16.f32 %rs16, %f16;}
|
||||
|
||||
// inline asm
|
||||
st.v4.u16 [%rd73], {%rs16, %rs17, %rs18, %rs12};
|
||||
ret;
|
||||
}
|
||||
|
||||
|
||||
@@ -0,0 +1,8 @@
|
||||
fileFormatVersion: 2
|
||||
guid: 3eeb7fb131fab904397234e1258187d6
|
||||
timeCreated: 1536238249
|
||||
licenseType: Store
|
||||
DefaultImporter:
|
||||
userData:
|
||||
assetBundleName:
|
||||
assetBundleVariant:
|
||||
666
Assets/Editor/x64/Bakery/addSH.ptx
Normal file
666
Assets/Editor/x64/Bakery/addSH.ptx
Normal file
@@ -0,0 +1,666 @@
|
||||
//
|
||||
// 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 inputImageL0[1];
|
||||
.global .align 1 .b8 inputImageL1x[1];
|
||||
.global .align 1 .b8 inputImageL1y[1];
|
||||
.global .align 1 .b8 inputImageL1z[1];
|
||||
.global .align 1 .b8 outputImageL0[1];
|
||||
.global .align 1 .b8 outputImageL1x[1];
|
||||
.global .align 1 .b8 outputImageL1y[1];
|
||||
.global .align 1 .b8 outputImageL1z[1];
|
||||
.global .align 1 .b8 packedImageL1x[1];
|
||||
.global .align 1 .b8 packedImageL1y[1];
|
||||
.global .align 1 .b8 packedImageL1z[1];
|
||||
.global .align 4 .f32 DoPack;
|
||||
.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_typeinfo6DoPackE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
|
||||
.global .align 8 .u64 _ZN21rti_internal_register20reg_bitness_detectorE;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail0E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail1E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail2E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail3E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail4E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail5E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail6E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail7E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail8E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail9E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail0E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail1E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail2E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail3E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail4E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail5E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail6E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail7E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail8E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail9E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register14reg_rayIndex_xE;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register14reg_rayIndex_yE;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register14reg_rayIndex_zE;
|
||||
.global .align 8 .b8 _ZN21rti_internal_typename7pixelIDE[6] = {117, 105, 110, 116, 50, 0};
|
||||
.global .align 8 .b8 _ZN21rti_internal_typename10resolutionE[6] = {117, 105, 110, 116, 50, 0};
|
||||
.global .align 8 .b8 _ZN21rti_internal_typename6normalE[7] = {102, 108, 111, 97, 116, 51, 0};
|
||||
.global .align 8 .b8 _ZN21rti_internal_typename6camPosE[7] = {102, 108, 111, 97, 116, 51, 0};
|
||||
.global .align 16 .b8 _ZN21rti_internal_typename4rootE[9] = {114, 116, 79, 98, 106, 101, 99, 116, 0};
|
||||
.global .align 4 .b8 _ZN21rti_internal_typename12imageEnabledE[4] = {105, 110, 116, 0};
|
||||
.global .align 8 .b8 _ZN21rti_internal_typename8tileInfoE[6] = {117, 105, 110, 116, 52, 0};
|
||||
.global .align 4 .b8 _ZN21rti_internal_typename8additiveE[4] = {105, 110, 116, 0};
|
||||
.global .align 8 .b8 _ZN21rti_internal_typename6DoPackE[6] = {102, 108, 111, 97, 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_typeenum6DoPackE = 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_semantic6DoPackE[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_annotation6DoPackE[1];
|
||||
|
||||
.visible .entry _Z6oxMainv(
|
||||
|
||||
)
|
||||
{
|
||||
.reg .pred %p<2>;
|
||||
.reg .b16 %rs<51>;
|
||||
.reg .f32 %f<132>;
|
||||
.reg .b32 %r<209>;
|
||||
.reg .b64 %rd<213>;
|
||||
|
||||
|
||||
ld.global.v2.u32 {%r55, %r56}, [pixelID];
|
||||
cvt.u64.u32 %rd3, %r55;
|
||||
cvt.u64.u32 %rd4, %r56;
|
||||
mov.u64 %rd163, inputImageL0;
|
||||
cvta.global.u64 %rd2, %rd163;
|
||||
mov.u32 %r53, 2;
|
||||
mov.u32 %r54, 8;
|
||||
mov.u64 %rd162, 0;
|
||||
// inline asm
|
||||
call (%rd1), _rt_buffer_get_64, (%rd2, %r53, %r54, %rd3, %rd4, %rd162, %rd162);
|
||||
// inline asm
|
||||
ld.u16 %rs2, [%rd1];
|
||||
// inline asm
|
||||
{ cvt.f32.f16 %f11, %rs2;}
|
||||
|
||||
// inline asm
|
||||
ld.global.v2.u32 {%r59, %r60}, [pixelID];
|
||||
cvt.u64.u32 %rd9, %r59;
|
||||
cvt.u64.u32 %rd10, %r60;
|
||||
// inline asm
|
||||
call (%rd7), _rt_buffer_get_64, (%rd2, %r53, %r54, %rd9, %rd10, %rd162, %rd162);
|
||||
// inline asm
|
||||
ld.u16 %rs3, [%rd7+2];
|
||||
// inline asm
|
||||
{ cvt.f32.f16 %f12, %rs3;}
|
||||
|
||||
// inline asm
|
||||
ld.global.v2.u32 {%r63, %r64}, [pixelID];
|
||||
cvt.u64.u32 %rd15, %r63;
|
||||
cvt.u64.u32 %rd16, %r64;
|
||||
// inline asm
|
||||
call (%rd13), _rt_buffer_get_64, (%rd2, %r53, %r54, %rd15, %rd16, %rd162, %rd162);
|
||||
// inline asm
|
||||
ld.u16 %rs4, [%rd13+4];
|
||||
// inline asm
|
||||
{ cvt.f32.f16 %f13, %rs4;}
|
||||
|
||||
// inline asm
|
||||
ld.global.v2.u32 {%r67, %r68}, [pixelID];
|
||||
cvt.u64.u32 %rd21, %r67;
|
||||
cvt.u64.u32 %rd22, %r68;
|
||||
mov.u64 %rd164, inputImageL1x;
|
||||
cvta.global.u64 %rd20, %rd164;
|
||||
// inline asm
|
||||
call (%rd19), _rt_buffer_get_64, (%rd20, %r53, %r54, %rd21, %rd22, %rd162, %rd162);
|
||||
// inline asm
|
||||
ld.u16 %rs5, [%rd19];
|
||||
// inline asm
|
||||
{ cvt.f32.f16 %f14, %rs5;}
|
||||
|
||||
// inline asm
|
||||
ld.global.v2.u32 {%r71, %r72}, [pixelID];
|
||||
cvt.u64.u32 %rd27, %r71;
|
||||
cvt.u64.u32 %rd28, %r72;
|
||||
// inline asm
|
||||
call (%rd25), _rt_buffer_get_64, (%rd20, %r53, %r54, %rd27, %rd28, %rd162, %rd162);
|
||||
// inline asm
|
||||
ld.u16 %rs6, [%rd25+2];
|
||||
// inline asm
|
||||
{ cvt.f32.f16 %f15, %rs6;}
|
||||
|
||||
// inline asm
|
||||
ld.global.v2.u32 {%r75, %r76}, [pixelID];
|
||||
cvt.u64.u32 %rd33, %r75;
|
||||
cvt.u64.u32 %rd34, %r76;
|
||||
// inline asm
|
||||
call (%rd31), _rt_buffer_get_64, (%rd20, %r53, %r54, %rd33, %rd34, %rd162, %rd162);
|
||||
// inline asm
|
||||
ld.u16 %rs7, [%rd31+4];
|
||||
// inline asm
|
||||
{ cvt.f32.f16 %f16, %rs7;}
|
||||
|
||||
// inline asm
|
||||
fma.rn.f32 %f41, %f14, 0f40000000, 0fBF800000;
|
||||
fma.rn.f32 %f42, %f15, 0f40000000, 0fBF800000;
|
||||
fma.rn.f32 %f43, %f16, 0f40000000, 0fBF800000;
|
||||
ld.global.v2.u32 {%r79, %r80}, [pixelID];
|
||||
cvt.u64.u32 %rd39, %r79;
|
||||
cvt.u64.u32 %rd40, %r80;
|
||||
mov.u64 %rd165, inputImageL1y;
|
||||
cvta.global.u64 %rd38, %rd165;
|
||||
// inline asm
|
||||
call (%rd37), _rt_buffer_get_64, (%rd38, %r53, %r54, %rd39, %rd40, %rd162, %rd162);
|
||||
// inline asm
|
||||
ld.u16 %rs8, [%rd37];
|
||||
// inline asm
|
||||
{ cvt.f32.f16 %f17, %rs8;}
|
||||
|
||||
// inline asm
|
||||
ld.global.v2.u32 {%r83, %r84}, [pixelID];
|
||||
cvt.u64.u32 %rd45, %r83;
|
||||
cvt.u64.u32 %rd46, %r84;
|
||||
// inline asm
|
||||
call (%rd43), _rt_buffer_get_64, (%rd38, %r53, %r54, %rd45, %rd46, %rd162, %rd162);
|
||||
// inline asm
|
||||
ld.u16 %rs9, [%rd43+2];
|
||||
// inline asm
|
||||
{ cvt.f32.f16 %f18, %rs9;}
|
||||
|
||||
// inline asm
|
||||
ld.global.v2.u32 {%r87, %r88}, [pixelID];
|
||||
cvt.u64.u32 %rd51, %r87;
|
||||
cvt.u64.u32 %rd52, %r88;
|
||||
// inline asm
|
||||
call (%rd49), _rt_buffer_get_64, (%rd38, %r53, %r54, %rd51, %rd52, %rd162, %rd162);
|
||||
// inline asm
|
||||
ld.u16 %rs10, [%rd49+4];
|
||||
// inline asm
|
||||
{ cvt.f32.f16 %f19, %rs10;}
|
||||
|
||||
// inline asm
|
||||
fma.rn.f32 %f44, %f17, 0f40000000, 0fBF800000;
|
||||
fma.rn.f32 %f45, %f18, 0f40000000, 0fBF800000;
|
||||
fma.rn.f32 %f46, %f19, 0f40000000, 0fBF800000;
|
||||
ld.global.v2.u32 {%r91, %r92}, [pixelID];
|
||||
cvt.u64.u32 %rd57, %r91;
|
||||
cvt.u64.u32 %rd58, %r92;
|
||||
mov.u64 %rd166, inputImageL1z;
|
||||
cvta.global.u64 %rd56, %rd166;
|
||||
// inline asm
|
||||
call (%rd55), _rt_buffer_get_64, (%rd56, %r53, %r54, %rd57, %rd58, %rd162, %rd162);
|
||||
// inline asm
|
||||
ld.u16 %rs11, [%rd55];
|
||||
// inline asm
|
||||
{ cvt.f32.f16 %f20, %rs11;}
|
||||
|
||||
// inline asm
|
||||
ld.global.v2.u32 {%r95, %r96}, [pixelID];
|
||||
cvt.u64.u32 %rd63, %r95;
|
||||
cvt.u64.u32 %rd64, %r96;
|
||||
// inline asm
|
||||
call (%rd61), _rt_buffer_get_64, (%rd56, %r53, %r54, %rd63, %rd64, %rd162, %rd162);
|
||||
// inline asm
|
||||
ld.u16 %rs12, [%rd61+2];
|
||||
// inline asm
|
||||
{ cvt.f32.f16 %f21, %rs12;}
|
||||
|
||||
// inline asm
|
||||
ld.global.v2.u32 {%r99, %r100}, [pixelID];
|
||||
cvt.u64.u32 %rd69, %r99;
|
||||
cvt.u64.u32 %rd70, %r100;
|
||||
// inline asm
|
||||
call (%rd67), _rt_buffer_get_64, (%rd56, %r53, %r54, %rd69, %rd70, %rd162, %rd162);
|
||||
// inline asm
|
||||
ld.u16 %rs13, [%rd67+4];
|
||||
// inline asm
|
||||
{ cvt.f32.f16 %f22, %rs13;}
|
||||
|
||||
// inline asm
|
||||
fma.rn.f32 %f47, %f20, 0f40000000, 0fBF800000;
|
||||
fma.rn.f32 %f48, %f21, 0f40000000, 0fBF800000;
|
||||
fma.rn.f32 %f49, %f22, 0f40000000, 0fBF800000;
|
||||
ld.global.v2.u32 {%r103, %r104}, [pixelID];
|
||||
cvt.u64.u32 %rd75, %r103;
|
||||
cvt.u64.u32 %rd76, %r104;
|
||||
mov.u64 %rd167, outputImageL0;
|
||||
cvta.global.u64 %rd74, %rd167;
|
||||
// inline asm
|
||||
call (%rd73), _rt_buffer_get_64, (%rd74, %r53, %r54, %rd75, %rd76, %rd162, %rd162);
|
||||
// inline asm
|
||||
ld.u16 %rs14, [%rd73];
|
||||
// inline asm
|
||||
{ cvt.f32.f16 %f23, %rs14;}
|
||||
|
||||
// inline asm
|
||||
ld.global.v2.u32 {%r107, %r108}, [pixelID];
|
||||
cvt.u64.u32 %rd81, %r107;
|
||||
cvt.u64.u32 %rd82, %r108;
|
||||
// inline asm
|
||||
call (%rd79), _rt_buffer_get_64, (%rd74, %r53, %r54, %rd81, %rd82, %rd162, %rd162);
|
||||
// inline asm
|
||||
ld.u16 %rs15, [%rd79+2];
|
||||
// inline asm
|
||||
{ cvt.f32.f16 %f24, %rs15;}
|
||||
|
||||
// inline asm
|
||||
ld.global.v2.u32 {%r111, %r112}, [pixelID];
|
||||
cvt.u64.u32 %rd87, %r111;
|
||||
cvt.u64.u32 %rd88, %r112;
|
||||
// inline asm
|
||||
call (%rd85), _rt_buffer_get_64, (%rd74, %r53, %r54, %rd87, %rd88, %rd162, %rd162);
|
||||
// inline asm
|
||||
ld.u16 %rs16, [%rd85+4];
|
||||
// inline asm
|
||||
{ cvt.f32.f16 %f25, %rs16;}
|
||||
|
||||
// inline asm
|
||||
ld.global.v2.u32 {%r115, %r116}, [pixelID];
|
||||
cvt.u64.u32 %rd93, %r115;
|
||||
cvt.u64.u32 %rd94, %r116;
|
||||
mov.u64 %rd168, outputImageL1x;
|
||||
cvta.global.u64 %rd92, %rd168;
|
||||
// inline asm
|
||||
call (%rd91), _rt_buffer_get_64, (%rd92, %r53, %r54, %rd93, %rd94, %rd162, %rd162);
|
||||
// inline asm
|
||||
ld.u16 %rs17, [%rd91];
|
||||
// inline asm
|
||||
{ cvt.f32.f16 %f26, %rs17;}
|
||||
|
||||
// inline asm
|
||||
ld.global.v2.u32 {%r119, %r120}, [pixelID];
|
||||
cvt.u64.u32 %rd99, %r119;
|
||||
cvt.u64.u32 %rd100, %r120;
|
||||
// inline asm
|
||||
call (%rd97), _rt_buffer_get_64, (%rd92, %r53, %r54, %rd99, %rd100, %rd162, %rd162);
|
||||
// inline asm
|
||||
ld.u16 %rs18, [%rd97+2];
|
||||
// inline asm
|
||||
{ cvt.f32.f16 %f27, %rs18;}
|
||||
|
||||
// inline asm
|
||||
ld.global.v2.u32 {%r123, %r124}, [pixelID];
|
||||
cvt.u64.u32 %rd105, %r123;
|
||||
cvt.u64.u32 %rd106, %r124;
|
||||
// inline asm
|
||||
call (%rd103), _rt_buffer_get_64, (%rd92, %r53, %r54, %rd105, %rd106, %rd162, %rd162);
|
||||
// inline asm
|
||||
ld.u16 %rs19, [%rd103+4];
|
||||
// inline asm
|
||||
{ cvt.f32.f16 %f28, %rs19;}
|
||||
|
||||
// inline asm
|
||||
fma.rn.f32 %f50, %f26, 0f40000000, 0fBF800000;
|
||||
fma.rn.f32 %f51, %f27, 0f40000000, 0fBF800000;
|
||||
fma.rn.f32 %f52, %f28, 0f40000000, 0fBF800000;
|
||||
mul.f32 %f53, %f23, %f50;
|
||||
mul.f32 %f54, %f24, %f51;
|
||||
mul.f32 %f55, %f25, %f52;
|
||||
ld.global.v2.u32 {%r127, %r128}, [pixelID];
|
||||
cvt.u64.u32 %rd111, %r127;
|
||||
cvt.u64.u32 %rd112, %r128;
|
||||
mov.u64 %rd169, outputImageL1y;
|
||||
cvta.global.u64 %rd110, %rd169;
|
||||
// inline asm
|
||||
call (%rd109), _rt_buffer_get_64, (%rd110, %r53, %r54, %rd111, %rd112, %rd162, %rd162);
|
||||
// inline asm
|
||||
ld.u16 %rs20, [%rd109];
|
||||
// inline asm
|
||||
{ cvt.f32.f16 %f29, %rs20;}
|
||||
|
||||
// inline asm
|
||||
ld.global.v2.u32 {%r131, %r132}, [pixelID];
|
||||
cvt.u64.u32 %rd117, %r131;
|
||||
cvt.u64.u32 %rd118, %r132;
|
||||
// inline asm
|
||||
call (%rd115), _rt_buffer_get_64, (%rd110, %r53, %r54, %rd117, %rd118, %rd162, %rd162);
|
||||
// inline asm
|
||||
ld.u16 %rs21, [%rd115+2];
|
||||
// inline asm
|
||||
{ cvt.f32.f16 %f30, %rs21;}
|
||||
|
||||
// inline asm
|
||||
ld.global.v2.u32 {%r135, %r136}, [pixelID];
|
||||
cvt.u64.u32 %rd123, %r135;
|
||||
cvt.u64.u32 %rd124, %r136;
|
||||
// inline asm
|
||||
call (%rd121), _rt_buffer_get_64, (%rd110, %r53, %r54, %rd123, %rd124, %rd162, %rd162);
|
||||
// inline asm
|
||||
ld.u16 %rs22, [%rd121+4];
|
||||
// inline asm
|
||||
{ cvt.f32.f16 %f31, %rs22;}
|
||||
|
||||
// inline asm
|
||||
fma.rn.f32 %f56, %f29, 0f40000000, 0fBF800000;
|
||||
fma.rn.f32 %f57, %f30, 0f40000000, 0fBF800000;
|
||||
fma.rn.f32 %f58, %f31, 0f40000000, 0fBF800000;
|
||||
mul.f32 %f59, %f23, %f56;
|
||||
mul.f32 %f60, %f24, %f57;
|
||||
mul.f32 %f61, %f25, %f58;
|
||||
ld.global.v2.u32 {%r139, %r140}, [pixelID];
|
||||
cvt.u64.u32 %rd129, %r139;
|
||||
cvt.u64.u32 %rd130, %r140;
|
||||
mov.u64 %rd170, outputImageL1z;
|
||||
cvta.global.u64 %rd128, %rd170;
|
||||
// inline asm
|
||||
call (%rd127), _rt_buffer_get_64, (%rd128, %r53, %r54, %rd129, %rd130, %rd162, %rd162);
|
||||
// inline asm
|
||||
ld.u16 %rs23, [%rd127];
|
||||
// inline asm
|
||||
{ cvt.f32.f16 %f32, %rs23;}
|
||||
|
||||
// inline asm
|
||||
ld.global.v2.u32 {%r143, %r144}, [pixelID];
|
||||
cvt.u64.u32 %rd135, %r143;
|
||||
cvt.u64.u32 %rd136, %r144;
|
||||
// inline asm
|
||||
call (%rd133), _rt_buffer_get_64, (%rd128, %r53, %r54, %rd135, %rd136, %rd162, %rd162);
|
||||
// inline asm
|
||||
ld.u16 %rs24, [%rd133+2];
|
||||
// inline asm
|
||||
{ cvt.f32.f16 %f33, %rs24;}
|
||||
|
||||
// inline asm
|
||||
ld.global.v2.u32 {%r147, %r148}, [pixelID];
|
||||
cvt.u64.u32 %rd141, %r147;
|
||||
cvt.u64.u32 %rd142, %r148;
|
||||
// inline asm
|
||||
call (%rd139), _rt_buffer_get_64, (%rd128, %r53, %r54, %rd141, %rd142, %rd162, %rd162);
|
||||
// inline asm
|
||||
ld.u16 %rs25, [%rd139+4];
|
||||
// inline asm
|
||||
{ cvt.f32.f16 %f34, %rs25;}
|
||||
|
||||
// inline asm
|
||||
fma.rn.f32 %f62, %f32, 0f40000000, 0fBF800000;
|
||||
fma.rn.f32 %f63, %f33, 0f40000000, 0fBF800000;
|
||||
fma.rn.f32 %f64, %f34, 0f40000000, 0fBF800000;
|
||||
mul.f32 %f65, %f23, %f62;
|
||||
mul.f32 %f66, %f24, %f63;
|
||||
mul.f32 %f67, %f25, %f64;
|
||||
ld.global.v2.u32 {%r151, %r152}, [pixelID];
|
||||
cvt.u64.u32 %rd147, %r151;
|
||||
cvt.u64.u32 %rd148, %r152;
|
||||
// inline asm
|
||||
call (%rd145), _rt_buffer_get_64, (%rd2, %r53, %r54, %rd147, %rd148, %rd162, %rd162);
|
||||
// inline asm
|
||||
ld.u16 %rs26, [%rd145+6];
|
||||
// inline asm
|
||||
{ cvt.f32.f16 %f35, %rs26;}
|
||||
|
||||
// inline asm
|
||||
ld.global.v2.u32 {%r155, %r156}, [pixelID];
|
||||
cvt.u64.u32 %rd153, %r155;
|
||||
cvt.u64.u32 %rd154, %r156;
|
||||
// inline asm
|
||||
call (%rd151), _rt_buffer_get_64, (%rd74, %r53, %r54, %rd153, %rd154, %rd162, %rd162);
|
||||
// inline asm
|
||||
ld.u16 %rs27, [%rd151+6];
|
||||
// inline asm
|
||||
{ cvt.f32.f16 %f36, %rs27;}
|
||||
|
||||
// inline asm
|
||||
min.f32 %f40, %f35, %f36;
|
||||
add.f32 %f68, %f11, %f23;
|
||||
add.f32 %f69, %f12, %f24;
|
||||
add.f32 %f70, %f13, %f25;
|
||||
fma.rn.f32 %f71, %f11, %f41, %f53;
|
||||
fma.rn.f32 %f72, %f12, %f42, %f54;
|
||||
fma.rn.f32 %f73, %f13, %f43, %f55;
|
||||
fma.rn.f32 %f74, %f11, %f44, %f59;
|
||||
fma.rn.f32 %f75, %f12, %f45, %f60;
|
||||
fma.rn.f32 %f76, %f13, %f46, %f61;
|
||||
fma.rn.f32 %f77, %f11, %f47, %f65;
|
||||
fma.rn.f32 %f78, %f12, %f48, %f66;
|
||||
fma.rn.f32 %f79, %f13, %f49, %f67;
|
||||
mov.f32 %f80, 0f34000000;
|
||||
max.f32 %f81, %f68, %f80;
|
||||
max.f32 %f82, %f69, %f80;
|
||||
max.f32 %f83, %f70, %f80;
|
||||
div.rn.f32 %f84, %f71, %f81;
|
||||
div.rn.f32 %f85, %f72, %f82;
|
||||
div.rn.f32 %f86, %f73, %f83;
|
||||
fma.rn.f32 %f87, %f84, 0f3F000000, 0f3F000000;
|
||||
fma.rn.f32 %f88, %f85, 0f3F000000, 0f3F000000;
|
||||
fma.rn.f32 %f89, %f86, 0f3F000000, 0f3F000000;
|
||||
div.rn.f32 %f90, %f74, %f81;
|
||||
div.rn.f32 %f91, %f75, %f82;
|
||||
div.rn.f32 %f92, %f76, %f83;
|
||||
fma.rn.f32 %f93, %f90, 0f3F000000, 0f3F000000;
|
||||
fma.rn.f32 %f94, %f91, 0f3F000000, 0f3F000000;
|
||||
fma.rn.f32 %f95, %f92, 0f3F000000, 0f3F000000;
|
||||
div.rn.f32 %f96, %f77, %f81;
|
||||
div.rn.f32 %f97, %f78, %f82;
|
||||
div.rn.f32 %f98, %f79, %f83;
|
||||
fma.rn.f32 %f99, %f96, 0f3F000000, 0f3F000000;
|
||||
fma.rn.f32 %f100, %f97, 0f3F000000, 0f3F000000;
|
||||
fma.rn.f32 %f101, %f98, 0f3F000000, 0f3F000000;
|
||||
mul.f32 %f37, %f68, %f40;
|
||||
mul.f32 %f38, %f69, %f40;
|
||||
mul.f32 %f39, %f70, %f40;
|
||||
mul.f32 %f2, %f40, %f87;
|
||||
mul.f32 %f3, %f40, %f88;
|
||||
mul.f32 %f4, %f40, %f89;
|
||||
mul.f32 %f5, %f40, %f93;
|
||||
mul.f32 %f6, %f40, %f94;
|
||||
mul.f32 %f7, %f40, %f95;
|
||||
mul.f32 %f8, %f40, %f99;
|
||||
mul.f32 %f9, %f40, %f100;
|
||||
mul.f32 %f10, %f40, %f101;
|
||||
ld.global.v2.u32 {%r159, %r160}, [pixelID];
|
||||
cvt.u64.u32 %rd159, %r159;
|
||||
cvt.u64.u32 %rd160, %r160;
|
||||
// inline asm
|
||||
call (%rd157), _rt_buffer_get_64, (%rd74, %r53, %r54, %rd159, %rd160, %rd162, %rd162);
|
||||
// inline asm
|
||||
// inline asm
|
||||
{ cvt.rn.f16.f32 %rs31, %f40;}
|
||||
|
||||
// inline asm
|
||||
// inline asm
|
||||
{ cvt.rn.f16.f32 %rs30, %f39;}
|
||||
|
||||
// inline asm
|
||||
// inline asm
|
||||
{ cvt.rn.f16.f32 %rs29, %f38;}
|
||||
|
||||
// inline asm
|
||||
// inline asm
|
||||
{ cvt.rn.f16.f32 %rs28, %f37;}
|
||||
|
||||
// inline asm
|
||||
st.v4.u16 [%rd157], {%rs28, %rs29, %rs30, %rs31};
|
||||
ld.global.f32 %f102, [DoPack];
|
||||
setp.gt.f32 %p1, %f102, 0f3F000000;
|
||||
@%p1 bra BB0_2;
|
||||
bra.uni BB0_1;
|
||||
|
||||
BB0_2:
|
||||
mul.f32 %f112, %f2, 0f437F0000;
|
||||
mov.f32 %f113, 0f437F0000;
|
||||
min.f32 %f114, %f112, %f113;
|
||||
mul.f32 %f115, %f3, 0f437F0000;
|
||||
min.f32 %f116, %f115, %f113;
|
||||
mul.f32 %f117, %f4, 0f437F0000;
|
||||
min.f32 %f118, %f117, %f113;
|
||||
mul.f32 %f119, %f5, 0f437F0000;
|
||||
min.f32 %f120, %f119, %f113;
|
||||
mul.f32 %f121, %f6, 0f437F0000;
|
||||
min.f32 %f122, %f121, %f113;
|
||||
mul.f32 %f123, %f7, 0f437F0000;
|
||||
min.f32 %f124, %f123, %f113;
|
||||
mul.f32 %f125, %f8, 0f437F0000;
|
||||
min.f32 %f126, %f125, %f113;
|
||||
mul.f32 %f127, %f9, 0f437F0000;
|
||||
min.f32 %f128, %f127, %f113;
|
||||
mul.f32 %f129, %f10, 0f437F0000;
|
||||
min.f32 %f130, %f129, %f113;
|
||||
ld.global.v2.u32 {%r187, %r188}, [pixelID];
|
||||
cvt.u64.u32 %rd194, %r187;
|
||||
cvt.u64.u32 %rd195, %r188;
|
||||
mov.u64 %rd210, packedImageL1x;
|
||||
cvta.global.u64 %rd193, %rd210;
|
||||
mov.u32 %r186, 4;
|
||||
// inline asm
|
||||
call (%rd192), _rt_buffer_get_64, (%rd193, %r53, %r186, %rd194, %rd195, %rd162, %rd162);
|
||||
// inline asm
|
||||
cvt.rzi.u32.f32 %r191, %f114;
|
||||
cvt.rzi.u32.f32 %r192, %f116;
|
||||
cvt.rzi.u32.f32 %r193, %f118;
|
||||
mul.f32 %f131, %f40, 0f437F0000;
|
||||
cvt.rzi.u32.f32 %r194, %f131;
|
||||
cvt.u16.u32 %rs41, %r193;
|
||||
cvt.u16.u32 %rs42, %r192;
|
||||
cvt.u16.u32 %rs43, %r191;
|
||||
cvt.u16.u32 %rs44, %r194;
|
||||
st.v4.u8 [%rd192], {%rs43, %rs42, %rs41, %rs44};
|
||||
ld.global.v2.u32 {%r195, %r196}, [pixelID];
|
||||
cvt.u64.u32 %rd200, %r195;
|
||||
cvt.u64.u32 %rd201, %r196;
|
||||
mov.u64 %rd211, packedImageL1y;
|
||||
cvta.global.u64 %rd199, %rd211;
|
||||
// inline asm
|
||||
call (%rd198), _rt_buffer_get_64, (%rd199, %r53, %r186, %rd200, %rd201, %rd162, %rd162);
|
||||
// inline asm
|
||||
cvt.rzi.u32.f32 %r199, %f120;
|
||||
cvt.rzi.u32.f32 %r200, %f122;
|
||||
cvt.rzi.u32.f32 %r201, %f124;
|
||||
cvt.u16.u32 %rs45, %r201;
|
||||
cvt.u16.u32 %rs46, %r200;
|
||||
cvt.u16.u32 %rs47, %r199;
|
||||
st.v4.u8 [%rd198], {%rs47, %rs46, %rs45, %rs44};
|
||||
ld.global.v2.u32 {%r202, %r203}, [pixelID];
|
||||
cvt.u64.u32 %rd206, %r202;
|
||||
cvt.u64.u32 %rd207, %r203;
|
||||
mov.u64 %rd212, packedImageL1z;
|
||||
cvta.global.u64 %rd205, %rd212;
|
||||
// inline asm
|
||||
call (%rd204), _rt_buffer_get_64, (%rd205, %r53, %r186, %rd206, %rd207, %rd162, %rd162);
|
||||
// inline asm
|
||||
cvt.rzi.u32.f32 %r206, %f126;
|
||||
cvt.rzi.u32.f32 %r207, %f128;
|
||||
cvt.rzi.u32.f32 %r208, %f130;
|
||||
cvt.u16.u32 %rs48, %r208;
|
||||
cvt.u16.u32 %rs49, %r207;
|
||||
cvt.u16.u32 %rs50, %r206;
|
||||
st.v4.u8 [%rd204], {%rs50, %rs49, %rs48, %rs44};
|
||||
bra.uni BB0_3;
|
||||
|
||||
BB0_1:
|
||||
ld.global.v2.u32 {%r169, %r170}, [pixelID];
|
||||
cvt.u64.u32 %rd173, %r169;
|
||||
cvt.u64.u32 %rd174, %r170;
|
||||
// inline asm
|
||||
call (%rd171), _rt_buffer_get_64, (%rd92, %r53, %r54, %rd173, %rd174, %rd162, %rd162);
|
||||
// inline asm
|
||||
cvt.sat.f32.f32 %f103, %f2;
|
||||
cvt.sat.f32.f32 %f104, %f3;
|
||||
cvt.sat.f32.f32 %f105, %f4;
|
||||
// inline asm
|
||||
{ cvt.rn.f16.f32 %rs34, %f105;}
|
||||
|
||||
// inline asm
|
||||
// inline asm
|
||||
{ cvt.rn.f16.f32 %rs33, %f104;}
|
||||
|
||||
// inline asm
|
||||
// inline asm
|
||||
{ cvt.rn.f16.f32 %rs32, %f103;}
|
||||
|
||||
// inline asm
|
||||
st.v4.u16 [%rd171], {%rs32, %rs33, %rs34, %rs31};
|
||||
ld.global.v2.u32 {%r173, %r174}, [pixelID];
|
||||
cvt.u64.u32 %rd179, %r173;
|
||||
cvt.u64.u32 %rd180, %r174;
|
||||
// inline asm
|
||||
call (%rd177), _rt_buffer_get_64, (%rd110, %r53, %r54, %rd179, %rd180, %rd162, %rd162);
|
||||
// inline asm
|
||||
cvt.sat.f32.f32 %f106, %f5;
|
||||
cvt.sat.f32.f32 %f107, %f6;
|
||||
cvt.sat.f32.f32 %f108, %f7;
|
||||
// inline asm
|
||||
{ cvt.rn.f16.f32 %rs37, %f108;}
|
||||
|
||||
// inline asm
|
||||
// inline asm
|
||||
{ cvt.rn.f16.f32 %rs36, %f107;}
|
||||
|
||||
// inline asm
|
||||
// inline asm
|
||||
{ cvt.rn.f16.f32 %rs35, %f106;}
|
||||
|
||||
// inline asm
|
||||
st.v4.u16 [%rd177], {%rs35, %rs36, %rs37, %rs31};
|
||||
ld.global.v2.u32 {%r177, %r178}, [pixelID];
|
||||
cvt.u64.u32 %rd185, %r177;
|
||||
cvt.u64.u32 %rd186, %r178;
|
||||
// inline asm
|
||||
call (%rd183), _rt_buffer_get_64, (%rd128, %r53, %r54, %rd185, %rd186, %rd162, %rd162);
|
||||
// inline asm
|
||||
cvt.sat.f32.f32 %f109, %f8;
|
||||
cvt.sat.f32.f32 %f110, %f9;
|
||||
cvt.sat.f32.f32 %f111, %f10;
|
||||
// inline asm
|
||||
{ cvt.rn.f16.f32 %rs40, %f111;}
|
||||
|
||||
// inline asm
|
||||
// inline asm
|
||||
{ cvt.rn.f16.f32 %rs39, %f110;}
|
||||
|
||||
// inline asm
|
||||
// inline asm
|
||||
{ cvt.rn.f16.f32 %rs38, %f109;}
|
||||
|
||||
// inline asm
|
||||
st.v4.u16 [%rd183], {%rs38, %rs39, %rs40, %rs31};
|
||||
|
||||
BB0_3:
|
||||
ret;
|
||||
}
|
||||
|
||||
|
||||
8
Assets/Editor/x64/Bakery/addSH.ptx.meta
Normal file
8
Assets/Editor/x64/Bakery/addSH.ptx.meta
Normal file
@@ -0,0 +1,8 @@
|
||||
fileFormatVersion: 2
|
||||
guid: 8e93cdd837ccece47adea672aa4bcff2
|
||||
timeCreated: 1537695617
|
||||
licenseType: Store
|
||||
DefaultImporter:
|
||||
userData:
|
||||
assetBundleName:
|
||||
assetBundleVariant:
|
||||
BIN
Assets/Editor/x64/Bakery/alphabuffergen.dll
LFS
Normal file
BIN
Assets/Editor/x64/Bakery/alphabuffergen.dll
LFS
Normal file
Binary file not shown.
138
Assets/Editor/x64/Bakery/alphabuffergen.dll.meta
Normal file
138
Assets/Editor/x64/Bakery/alphabuffergen.dll.meta
Normal file
@@ -0,0 +1,138 @@
|
||||
fileFormatVersion: 2
|
||||
guid: 36d566e460135294e9e29bd27a12b9b5
|
||||
timeCreated: 1526150503
|
||||
licenseType: Store
|
||||
PluginImporter:
|
||||
serializedVersion: 2
|
||||
iconMap: {}
|
||||
executionOrder: {}
|
||||
isPreloaded: 0
|
||||
isOverridable: 0
|
||||
platformData:
|
||||
data:
|
||||
first:
|
||||
'': Any
|
||||
second:
|
||||
enabled: 0
|
||||
settings:
|
||||
Exclude Android: 1
|
||||
Exclude Editor: 0
|
||||
Exclude Linux: 1
|
||||
Exclude Linux64: 1
|
||||
Exclude LinuxUniversal: 1
|
||||
Exclude OSXIntel: 1
|
||||
Exclude OSXIntel64: 1
|
||||
Exclude OSXUniversal: 1
|
||||
Exclude WebGL: 1
|
||||
Exclude Win: 1
|
||||
Exclude Win64: 1
|
||||
Exclude iOS: 1
|
||||
data:
|
||||
first:
|
||||
'': Editor
|
||||
second:
|
||||
enabled: 0
|
||||
settings:
|
||||
CPU: AnyCPU
|
||||
OS: AnyOS
|
||||
data:
|
||||
first:
|
||||
Android: Android
|
||||
second:
|
||||
enabled: 0
|
||||
settings:
|
||||
CPU: ARMv7
|
||||
data:
|
||||
first:
|
||||
Any:
|
||||
second:
|
||||
enabled: 0
|
||||
settings: {}
|
||||
data:
|
||||
first:
|
||||
Editor: Editor
|
||||
second:
|
||||
enabled: 1
|
||||
settings:
|
||||
DefaultValueInitialized: true
|
||||
data:
|
||||
first:
|
||||
Facebook: Win
|
||||
second:
|
||||
enabled: 0
|
||||
settings:
|
||||
CPU: None
|
||||
data:
|
||||
first:
|
||||
Facebook: Win64
|
||||
second:
|
||||
enabled: 0
|
||||
settings:
|
||||
CPU: None
|
||||
data:
|
||||
first:
|
||||
Standalone: Linux
|
||||
second:
|
||||
enabled: 0
|
||||
settings:
|
||||
CPU: x86
|
||||
data:
|
||||
first:
|
||||
Standalone: Linux64
|
||||
second:
|
||||
enabled: 0
|
||||
settings:
|
||||
CPU: x86_64
|
||||
data:
|
||||
first:
|
||||
Standalone: LinuxUniversal
|
||||
second:
|
||||
enabled: 0
|
||||
settings:
|
||||
CPU: None
|
||||
data:
|
||||
first:
|
||||
Standalone: OSXIntel
|
||||
second:
|
||||
enabled: 0
|
||||
settings:
|
||||
CPU: AnyCPU
|
||||
data:
|
||||
first:
|
||||
Standalone: OSXIntel64
|
||||
second:
|
||||
enabled: 0
|
||||
settings:
|
||||
CPU: AnyCPU
|
||||
data:
|
||||
first:
|
||||
Standalone: OSXUniversal
|
||||
second:
|
||||
enabled: 0
|
||||
settings:
|
||||
CPU: None
|
||||
data:
|
||||
first:
|
||||
Standalone: Win
|
||||
second:
|
||||
enabled: 0
|
||||
settings:
|
||||
CPU: None
|
||||
data:
|
||||
first:
|
||||
Standalone: Win64
|
||||
second:
|
||||
enabled: 0
|
||||
settings:
|
||||
CPU: None
|
||||
data:
|
||||
first:
|
||||
iPhone: iOS
|
||||
second:
|
||||
enabled: 0
|
||||
settings:
|
||||
CompileFlags:
|
||||
FrameworkDependencies:
|
||||
userData:
|
||||
assetBundleName:
|
||||
assetBundleVariant:
|
||||
265
Assets/Editor/x64/Bakery/attrib_trimesh.ptx
Normal file
265
Assets/Editor/x64/Bakery/attrib_trimesh.ptx
Normal file
@@ -0,0 +1,265 @@
|
||||
//
|
||||
// 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 _Z11interpolatei
|
||||
.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 4 .f32 triAlpha;
|
||||
.global .align 1 .b8 vertex_buffer[1];
|
||||
.global .align 1 .b8 index_buffer[1];
|
||||
.global .align 4 .u32 firstAlphaTriangle;
|
||||
.global .align 1 .b8 vertex_buffer_uv[1];
|
||||
.global .align 1 .b8 triangleAlphaIDs[1];
|
||||
.global .align 1 .b8 alphaTextures[1];
|
||||
.global .align 4 .b8 ray[36];
|
||||
.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_typeinfo8triAlphaE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
|
||||
.global .align 4 .b8 _ZN21rti_internal_typeinfo18firstAlphaTriangleE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
|
||||
.global .align 4 .b8 _ZN21rti_internal_typeinfo3rayE[8] = {82, 97, 121, 0, 36, 0, 0, 0};
|
||||
.global .align 8 .u64 _ZN21rti_internal_register20reg_bitness_detectorE;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail0E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail1E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail2E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail3E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail4E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail5E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail6E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail7E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail8E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail9E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail0E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail1E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail2E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail3E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail4E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail5E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail6E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail7E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail8E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail9E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register14reg_rayIndex_xE;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register14reg_rayIndex_yE;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register14reg_rayIndex_zE;
|
||||
.global .align 8 .b8 _ZN21rti_internal_typename7pixelIDE[6] = {117, 105, 110, 116, 50, 0};
|
||||
.global .align 8 .b8 _ZN21rti_internal_typename10resolutionE[6] = {117, 105, 110, 116, 50, 0};
|
||||
.global .align 8 .b8 _ZN21rti_internal_typename6normalE[7] = {102, 108, 111, 97, 116, 51, 0};
|
||||
.global .align 8 .b8 _ZN21rti_internal_typename6camPosE[7] = {102, 108, 111, 97, 116, 51, 0};
|
||||
.global .align 16 .b8 _ZN21rti_internal_typename4rootE[9] = {114, 116, 79, 98, 106, 101, 99, 116, 0};
|
||||
.global .align 4 .b8 _ZN21rti_internal_typename12imageEnabledE[4] = {105, 110, 116, 0};
|
||||
.global .align 8 .b8 _ZN21rti_internal_typename8tileInfoE[6] = {117, 105, 110, 116, 52, 0};
|
||||
.global .align 4 .b8 _ZN21rti_internal_typename8additiveE[4] = {105, 110, 116, 0};
|
||||
.global .align 8 .b8 _ZN21rti_internal_typename8triAlphaE[6] = {102, 108, 111, 97, 116, 0};
|
||||
.global .align 4 .b8 _ZN21rti_internal_typename18firstAlphaTriangleE[4] = {105, 110, 116, 0};
|
||||
.global .align 4 .b8 _ZN21rti_internal_typename3rayE[4] = {82, 97, 121, 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_typeenum8triAlphaE = 4919;
|
||||
.global .align 4 .u32 _ZN21rti_internal_typeenum18firstAlphaTriangleE = 4919;
|
||||
.global .align 4 .u32 _ZN21rti_internal_typeenum3rayE = 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 16 .b8 _ZN21rti_internal_semantic8triAlphaE[19] = {97, 116, 116, 114, 105, 98, 117, 116, 101, 32, 116, 114, 105, 65, 108, 112, 104, 97, 0};
|
||||
.global .align 1 .b8 _ZN21rti_internal_semantic18firstAlphaTriangleE[1];
|
||||
.global .align 16 .b8 _ZN21rti_internal_semantic3rayE[13] = {114, 116, 67, 117, 114, 114, 101, 110, 116, 82, 97, 121, 0};
|
||||
.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_annotation8triAlphaE[1];
|
||||
.global .align 1 .b8 _ZN23rti_internal_annotation18firstAlphaTriangleE[1];
|
||||
.global .align 1 .b8 _ZN23rti_internal_annotation3rayE[1];
|
||||
|
||||
.visible .entry _Z11interpolatei(
|
||||
.param .u32 _Z11interpolatei_param_0
|
||||
)
|
||||
{
|
||||
.reg .pred %p<3>;
|
||||
.reg .f32 %f<65>;
|
||||
.reg .b32 %r<39>;
|
||||
.reg .b64 %rd<76>;
|
||||
|
||||
|
||||
ld.param.u32 %r5, [_Z11interpolatei_param_0];
|
||||
// inline asm
|
||||
call (%f6, %f7), _rt_get_triangle_barycentrics, ();
|
||||
// inline asm
|
||||
cvt.s64.s32 %rd3, %r5;
|
||||
mov.u64 %rd7, index_buffer;
|
||||
cvta.global.u64 %rd2, %rd7;
|
||||
mov.u32 %r6, 1;
|
||||
mov.u32 %r7, 12;
|
||||
mov.u64 %rd6, 0;
|
||||
// inline asm
|
||||
call (%rd1), _rt_buffer_get_64, (%rd2, %r6, %r7, %rd3, %rd6, %rd6, %rd6);
|
||||
// inline asm
|
||||
ld.u32 %r1, [%rd1];
|
||||
ld.u32 %r2, [%rd1+4];
|
||||
ld.u32 %r3, [%rd1+8];
|
||||
ld.global.u32 %r4, [firstAlphaTriangle];
|
||||
setp.gt.s32 %p1, %r4, %r5;
|
||||
mov.f32 %f64, 0f3F800000;
|
||||
sub.f32 %f9, %f64, %f6;
|
||||
sub.f32 %f3, %f9, %f7;
|
||||
@%p1 bra BB0_2;
|
||||
|
||||
cvt.s64.s32 %rd28, %r3;
|
||||
sub.s32 %r26, %r5, %r4;
|
||||
cvt.s64.s32 %rd10, %r26;
|
||||
mov.u64 %rd53, triangleAlphaIDs;
|
||||
cvta.global.u64 %rd9, %rd53;
|
||||
mov.u32 %r22, 4;
|
||||
// inline asm
|
||||
call (%rd8), _rt_buffer_get_64, (%rd9, %r6, %r22, %rd10, %rd6, %rd6, %rd6);
|
||||
// inline asm
|
||||
ld.u32 %rd54, [%rd8];
|
||||
cvt.u32.u64 %r27, %rd54;
|
||||
shr.u32 %r28, %r27, 16;
|
||||
cvt.s64.s32 %rd16, %r1;
|
||||
mov.u64 %rd55, vertex_buffer_uv;
|
||||
cvta.global.u64 %rd15, %rd55;
|
||||
mov.u32 %r15, 8;
|
||||
// inline asm
|
||||
call (%rd14), _rt_buffer_get_64, (%rd15, %r6, %r15, %rd16, %rd6, %rd6, %rd6);
|
||||
// inline asm
|
||||
ld.v2.f32 {%f10, %f11}, [%rd14];
|
||||
cvt.s64.s32 %rd22, %r2;
|
||||
// inline asm
|
||||
call (%rd20), _rt_buffer_get_64, (%rd15, %r6, %r15, %rd22, %rd6, %rd6, %rd6);
|
||||
// inline asm
|
||||
ld.v2.f32 {%f14, %f15}, [%rd20];
|
||||
// inline asm
|
||||
call (%rd26), _rt_buffer_get_64, (%rd15, %r6, %r15, %rd28, %rd6, %rd6, %rd6);
|
||||
// inline asm
|
||||
mul.f32 %f18, %f6, %f14;
|
||||
mul.f32 %f19, %f6, %f15;
|
||||
fma.rn.f32 %f20, %f3, %f10, %f18;
|
||||
fma.rn.f32 %f21, %f3, %f11, %f19;
|
||||
ld.v2.f32 {%f22, %f23}, [%rd26];
|
||||
fma.rn.f32 %f26, %f7, %f22, %f20;
|
||||
fma.rn.f32 %f27, %f7, %f23, %f21;
|
||||
abs.f32 %f28, %f26;
|
||||
cvt.rmi.f32.f32 %f29, %f28;
|
||||
sub.f32 %f30, %f28, %f29;
|
||||
abs.f32 %f31, %f27;
|
||||
cvt.rmi.f32.f32 %f32, %f31;
|
||||
sub.f32 %f33, %f31, %f32;
|
||||
and.b64 %rd34, %rd54, 65535;
|
||||
mov.u64 %rd56, alphaTextures;
|
||||
cvta.global.u64 %rd33, %rd56;
|
||||
// inline asm
|
||||
call (%rd32), _rt_buffer_get_64, (%rd33, %r6, %r22, %rd34, %rd6, %rd6, %rd6);
|
||||
// inline asm
|
||||
ld.u32 %r18, [%rd32];
|
||||
mov.u32 %r24, 2;
|
||||
// inline asm
|
||||
call (%rd38, %rd39, %rd40, %rd41), _rt_buffer_get_id_size_64, (%r18, %r24, %r6);
|
||||
// inline asm
|
||||
cvt.rn.f32.u64 %f34, %rd38;
|
||||
mul.f32 %f35, %f30, %f34;
|
||||
cvt.rzi.u32.f32 %r29, %f35;
|
||||
cvt.rn.f32.u64 %f36, %rd39;
|
||||
mul.f32 %f37, %f33, %f36;
|
||||
cvt.rzi.u32.f32 %r30, %f37;
|
||||
// inline asm
|
||||
call (%rd42), _rt_buffer_get_64, (%rd33, %r6, %r22, %rd34, %rd6, %rd6, %rd6);
|
||||
// inline asm
|
||||
ld.u32 %r23, [%rd42];
|
||||
cvt.u64.u32 %rd49, %r29;
|
||||
cvt.u64.u32 %rd50, %r30;
|
||||
// inline asm
|
||||
call (%rd48), _rt_buffer_get_id_64, (%r23, %r24, %r6, %rd49, %rd50, %rd6, %rd6);
|
||||
// inline asm
|
||||
ld.u8 %r31, [%rd48];
|
||||
and.b32 %r32, %r31, %r28;
|
||||
setp.eq.s32 %p2, %r32, 0;
|
||||
selp.f32 %f64, 0f00000000, 0f3F800000, %p2;
|
||||
|
||||
BB0_2:
|
||||
st.global.f32 [triAlpha], %f64;
|
||||
cvt.s64.s32 %rd59, %r1;
|
||||
mov.u64 %rd75, vertex_buffer;
|
||||
cvta.global.u64 %rd58, %rd75;
|
||||
mov.u32 %r38, 24;
|
||||
// inline asm
|
||||
call (%rd57), _rt_buffer_get_64, (%rd58, %r6, %r38, %rd59, %rd6, %rd6, %rd6);
|
||||
// inline asm
|
||||
ld.f32 %f38, [%rd57+20];
|
||||
ld.f32 %f39, [%rd57+16];
|
||||
ld.f32 %f40, [%rd57+12];
|
||||
cvt.s64.s32 %rd65, %r2;
|
||||
// inline asm
|
||||
call (%rd63), _rt_buffer_get_64, (%rd58, %r6, %r38, %rd65, %rd6, %rd6, %rd6);
|
||||
// inline asm
|
||||
ld.f32 %f41, [%rd63+20];
|
||||
ld.f32 %f42, [%rd63+16];
|
||||
ld.f32 %f43, [%rd63+12];
|
||||
cvt.s64.s32 %rd71, %r3;
|
||||
// inline asm
|
||||
call (%rd69), _rt_buffer_get_64, (%rd58, %r6, %r38, %rd71, %rd6, %rd6, %rd6);
|
||||
// inline asm
|
||||
mul.f32 %f44, %f6, %f43;
|
||||
mul.f32 %f45, %f6, %f42;
|
||||
mul.f32 %f46, %f6, %f41;
|
||||
fma.rn.f32 %f47, %f3, %f40, %f44;
|
||||
fma.rn.f32 %f48, %f3, %f39, %f45;
|
||||
fma.rn.f32 %f49, %f3, %f38, %f46;
|
||||
ld.f32 %f50, [%rd69+20];
|
||||
ld.f32 %f51, [%rd69+16];
|
||||
ld.f32 %f52, [%rd69+12];
|
||||
fma.rn.f32 %f53, %f7, %f52, %f47;
|
||||
fma.rn.f32 %f54, %f7, %f51, %f48;
|
||||
fma.rn.f32 %f55, %f7, %f50, %f49;
|
||||
mul.f32 %f56, %f54, %f54;
|
||||
fma.rn.f32 %f57, %f53, %f53, %f56;
|
||||
fma.rn.f32 %f58, %f55, %f55, %f57;
|
||||
sqrt.rn.f32 %f59, %f58;
|
||||
rcp.rn.f32 %f60, %f59;
|
||||
mul.f32 %f61, %f60, %f53;
|
||||
mul.f32 %f62, %f60, %f54;
|
||||
mul.f32 %f63, %f60, %f55;
|
||||
st.global.f32 [normal], %f61;
|
||||
st.global.f32 [normal+4], %f62;
|
||||
st.global.f32 [normal+8], %f63;
|
||||
ret;
|
||||
}
|
||||
|
||||
|
||||
8
Assets/Editor/x64/Bakery/attrib_trimesh.ptx.meta
Normal file
8
Assets/Editor/x64/Bakery/attrib_trimesh.ptx.meta
Normal file
@@ -0,0 +1,8 @@
|
||||
fileFormatVersion: 2
|
||||
guid: 84cdf34ffa295d046887fff61e29ed06
|
||||
timeCreated: 1549654014
|
||||
licenseType: Store
|
||||
DefaultImporter:
|
||||
userData:
|
||||
assetBundleName:
|
||||
assetBundleVariant:
|
||||
301
Assets/Editor/x64/Bakery/attrib_trimeshTex.ptx
Normal file
301
Assets/Editor/x64/Bakery/attrib_trimeshTex.ptx
Normal file
@@ -0,0 +1,301 @@
|
||||
//
|
||||
// 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 _Z11interpolatei
|
||||
.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 8 .b8 texCoords[8];
|
||||
.global .align 4 .f32 triAlpha;
|
||||
.global .align 1 .b8 vertex_buffer[1];
|
||||
.global .align 1 .b8 index_buffer[1];
|
||||
.global .align 4 .u32 firstAlphaTriangle;
|
||||
.global .align 1 .b8 vertex_buffer_uv[1];
|
||||
.global .align 1 .b8 triangleAlphaIDs[1];
|
||||
.global .align 1 .b8 alphaTextures[1];
|
||||
.global .align 4 .b8 ray[36];
|
||||
.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_typeinfo9texCoordsE[8] = {82, 97, 121, 0, 8, 0, 0, 0};
|
||||
.global .align 4 .b8 _ZN21rti_internal_typeinfo8triAlphaE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
|
||||
.global .align 4 .b8 _ZN21rti_internal_typeinfo18firstAlphaTriangleE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
|
||||
.global .align 4 .b8 _ZN21rti_internal_typeinfo3rayE[8] = {82, 97, 121, 0, 36, 0, 0, 0};
|
||||
.global .align 8 .u64 _ZN21rti_internal_register20reg_bitness_detectorE;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail0E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail1E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail2E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail3E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail4E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail5E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail6E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail7E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail8E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail9E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail0E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail1E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail2E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail3E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail4E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail5E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail6E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail7E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail8E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail9E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register14reg_rayIndex_xE;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register14reg_rayIndex_yE;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register14reg_rayIndex_zE;
|
||||
.global .align 8 .b8 _ZN21rti_internal_typename7pixelIDE[6] = {117, 105, 110, 116, 50, 0};
|
||||
.global .align 8 .b8 _ZN21rti_internal_typename10resolutionE[6] = {117, 105, 110, 116, 50, 0};
|
||||
.global .align 8 .b8 _ZN21rti_internal_typename6normalE[7] = {102, 108, 111, 97, 116, 51, 0};
|
||||
.global .align 8 .b8 _ZN21rti_internal_typename6camPosE[7] = {102, 108, 111, 97, 116, 51, 0};
|
||||
.global .align 16 .b8 _ZN21rti_internal_typename4rootE[9] = {114, 116, 79, 98, 106, 101, 99, 116, 0};
|
||||
.global .align 4 .b8 _ZN21rti_internal_typename12imageEnabledE[4] = {105, 110, 116, 0};
|
||||
.global .align 8 .b8 _ZN21rti_internal_typename8tileInfoE[6] = {117, 105, 110, 116, 52, 0};
|
||||
.global .align 4 .b8 _ZN21rti_internal_typename8additiveE[4] = {105, 110, 116, 0};
|
||||
.global .align 8 .b8 _ZN21rti_internal_typename9texCoordsE[7] = {102, 108, 111, 97, 116, 50, 0};
|
||||
.global .align 8 .b8 _ZN21rti_internal_typename8triAlphaE[6] = {102, 108, 111, 97, 116, 0};
|
||||
.global .align 4 .b8 _ZN21rti_internal_typename18firstAlphaTriangleE[4] = {105, 110, 116, 0};
|
||||
.global .align 4 .b8 _ZN21rti_internal_typename3rayE[4] = {82, 97, 121, 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_typeenum9texCoordsE = 4919;
|
||||
.global .align 4 .u32 _ZN21rti_internal_typeenum8triAlphaE = 4919;
|
||||
.global .align 4 .u32 _ZN21rti_internal_typeenum18firstAlphaTriangleE = 4919;
|
||||
.global .align 4 .u32 _ZN21rti_internal_typeenum3rayE = 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 16 .b8 _ZN21rti_internal_semantic9texCoordsE[20] = {97, 116, 116, 114, 105, 98, 117, 116, 101, 32, 116, 101, 120, 67, 111, 111, 114, 100, 115, 0};
|
||||
.global .align 16 .b8 _ZN21rti_internal_semantic8triAlphaE[19] = {97, 116, 116, 114, 105, 98, 117, 116, 101, 32, 116, 114, 105, 65, 108, 112, 104, 97, 0};
|
||||
.global .align 1 .b8 _ZN21rti_internal_semantic18firstAlphaTriangleE[1];
|
||||
.global .align 16 .b8 _ZN21rti_internal_semantic3rayE[13] = {114, 116, 67, 117, 114, 114, 101, 110, 116, 82, 97, 121, 0};
|
||||
.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_annotation9texCoordsE[1];
|
||||
.global .align 1 .b8 _ZN23rti_internal_annotation8triAlphaE[1];
|
||||
.global .align 1 .b8 _ZN23rti_internal_annotation18firstAlphaTriangleE[1];
|
||||
.global .align 1 .b8 _ZN23rti_internal_annotation3rayE[1];
|
||||
|
||||
.visible .entry _Z11interpolatei(
|
||||
.param .u32 _Z11interpolatei_param_0
|
||||
)
|
||||
{
|
||||
.reg .pred %p<4>;
|
||||
.reg .f32 %f<88>;
|
||||
.reg .b32 %r<45>;
|
||||
.reg .b64 %rd<95>;
|
||||
|
||||
|
||||
ld.param.u32 %r5, [_Z11interpolatei_param_0];
|
||||
// inline asm
|
||||
call (%f15, %f16), _rt_get_triangle_barycentrics, ();
|
||||
// inline asm
|
||||
cvt.s64.s32 %rd4, %r5;
|
||||
mov.u64 %rd8, index_buffer;
|
||||
cvta.global.u64 %rd3, %rd8;
|
||||
mov.u32 %r6, 1;
|
||||
mov.u32 %r7, 12;
|
||||
mov.u64 %rd7, 0;
|
||||
// inline asm
|
||||
call (%rd2), _rt_buffer_get_64, (%rd3, %r6, %r7, %rd4, %rd7, %rd7, %rd7);
|
||||
// inline asm
|
||||
ld.u32 %r1, [%rd2];
|
||||
ld.u32 %r2, [%rd2+4];
|
||||
ld.u32 %r3, [%rd2+8];
|
||||
ld.global.u32 %r4, [firstAlphaTriangle];
|
||||
setp.gt.s32 %p1, %r4, %r5;
|
||||
mov.f32 %f85, 0f3F800000;
|
||||
sub.f32 %f18, %f85, %f15;
|
||||
sub.f32 %f3, %f18, %f16;
|
||||
@%p1 bra BB0_2;
|
||||
|
||||
cvt.s64.s32 %rd29, %r3;
|
||||
sub.s32 %r26, %r5, %r4;
|
||||
cvt.s64.s32 %rd11, %r26;
|
||||
mov.u64 %rd54, triangleAlphaIDs;
|
||||
cvta.global.u64 %rd10, %rd54;
|
||||
mov.u32 %r22, 4;
|
||||
// inline asm
|
||||
call (%rd9), _rt_buffer_get_64, (%rd10, %r6, %r22, %rd11, %rd7, %rd7, %rd7);
|
||||
// inline asm
|
||||
ld.u32 %rd55, [%rd9];
|
||||
cvt.u32.u64 %r27, %rd55;
|
||||
shr.u32 %r28, %r27, 16;
|
||||
cvt.s64.s32 %rd17, %r1;
|
||||
mov.u64 %rd56, vertex_buffer_uv;
|
||||
cvta.global.u64 %rd16, %rd56;
|
||||
mov.u32 %r15, 8;
|
||||
// inline asm
|
||||
call (%rd15), _rt_buffer_get_64, (%rd16, %r6, %r15, %rd17, %rd7, %rd7, %rd7);
|
||||
// inline asm
|
||||
ld.v2.f32 {%f19, %f20}, [%rd15];
|
||||
cvt.s64.s32 %rd23, %r2;
|
||||
// inline asm
|
||||
call (%rd21), _rt_buffer_get_64, (%rd16, %r6, %r15, %rd23, %rd7, %rd7, %rd7);
|
||||
// inline asm
|
||||
ld.v2.f32 {%f23, %f24}, [%rd21];
|
||||
// inline asm
|
||||
call (%rd27), _rt_buffer_get_64, (%rd16, %r6, %r15, %rd29, %rd7, %rd7, %rd7);
|
||||
// inline asm
|
||||
mul.f32 %f27, %f15, %f23;
|
||||
mul.f32 %f28, %f15, %f24;
|
||||
fma.rn.f32 %f29, %f3, %f19, %f27;
|
||||
fma.rn.f32 %f30, %f3, %f20, %f28;
|
||||
ld.v2.f32 {%f31, %f32}, [%rd27];
|
||||
fma.rn.f32 %f35, %f16, %f31, %f29;
|
||||
fma.rn.f32 %f36, %f16, %f32, %f30;
|
||||
abs.f32 %f37, %f35;
|
||||
cvt.rmi.f32.f32 %f38, %f37;
|
||||
sub.f32 %f39, %f37, %f38;
|
||||
abs.f32 %f40, %f36;
|
||||
cvt.rmi.f32.f32 %f41, %f40;
|
||||
sub.f32 %f42, %f40, %f41;
|
||||
and.b64 %rd35, %rd55, 65535;
|
||||
mov.u64 %rd57, alphaTextures;
|
||||
cvta.global.u64 %rd34, %rd57;
|
||||
// inline asm
|
||||
call (%rd33), _rt_buffer_get_64, (%rd34, %r6, %r22, %rd35, %rd7, %rd7, %rd7);
|
||||
// inline asm
|
||||
ld.u32 %r18, [%rd33];
|
||||
mov.u32 %r24, 2;
|
||||
// inline asm
|
||||
call (%rd39, %rd40, %rd41, %rd42), _rt_buffer_get_id_size_64, (%r18, %r24, %r6);
|
||||
// inline asm
|
||||
cvt.rn.f32.u64 %f43, %rd39;
|
||||
mul.f32 %f44, %f39, %f43;
|
||||
cvt.rzi.u32.f32 %r29, %f44;
|
||||
cvt.rn.f32.u64 %f45, %rd40;
|
||||
mul.f32 %f46, %f42, %f45;
|
||||
cvt.rzi.u32.f32 %r30, %f46;
|
||||
// inline asm
|
||||
call (%rd43), _rt_buffer_get_64, (%rd34, %r6, %r22, %rd35, %rd7, %rd7, %rd7);
|
||||
// inline asm
|
||||
ld.u32 %r23, [%rd43];
|
||||
cvt.u64.u32 %rd50, %r29;
|
||||
cvt.u64.u32 %rd51, %r30;
|
||||
// inline asm
|
||||
call (%rd49), _rt_buffer_get_id_64, (%r23, %r24, %r6, %rd50, %rd51, %rd7, %rd7);
|
||||
// inline asm
|
||||
ld.u8 %r31, [%rd49];
|
||||
and.b32 %r32, %r31, %r28;
|
||||
setp.eq.s32 %p2, %r32, 0;
|
||||
selp.f32 %f85, 0f00000000, 0f3F800000, %p2;
|
||||
|
||||
BB0_2:
|
||||
st.global.f32 [triAlpha], %f85;
|
||||
cvt.s64.s32 %rd78, %r1;
|
||||
mov.u64 %rd94, vertex_buffer;
|
||||
cvta.global.u64 %rd59, %rd94;
|
||||
mov.u32 %r44, 32;
|
||||
// inline asm
|
||||
call (%rd58), _rt_buffer_get_64, (%rd59, %r6, %r44, %rd78, %rd7, %rd7, %rd7);
|
||||
// inline asm
|
||||
ld.f32 %f47, [%rd58+20];
|
||||
ld.f32 %f48, [%rd58+16];
|
||||
ld.f32 %f49, [%rd58+12];
|
||||
cvt.s64.s32 %rd84, %r2;
|
||||
// inline asm
|
||||
call (%rd64), _rt_buffer_get_64, (%rd59, %r6, %r44, %rd84, %rd7, %rd7, %rd7);
|
||||
// inline asm
|
||||
ld.f32 %f50, [%rd64+20];
|
||||
ld.f32 %f51, [%rd64+16];
|
||||
ld.f32 %f52, [%rd64+12];
|
||||
cvt.s64.s32 %rd90, %r3;
|
||||
// inline asm
|
||||
call (%rd70), _rt_buffer_get_64, (%rd59, %r6, %r44, %rd90, %rd7, %rd7, %rd7);
|
||||
// inline asm
|
||||
mul.f32 %f53, %f15, %f52;
|
||||
mul.f32 %f54, %f15, %f51;
|
||||
mul.f32 %f55, %f15, %f50;
|
||||
fma.rn.f32 %f56, %f3, %f49, %f53;
|
||||
fma.rn.f32 %f57, %f3, %f48, %f54;
|
||||
fma.rn.f32 %f58, %f3, %f47, %f55;
|
||||
ld.f32 %f59, [%rd70+20];
|
||||
ld.f32 %f60, [%rd70+16];
|
||||
ld.f32 %f61, [%rd70+12];
|
||||
fma.rn.f32 %f62, %f16, %f61, %f56;
|
||||
fma.rn.f32 %f63, %f16, %f60, %f57;
|
||||
fma.rn.f32 %f64, %f16, %f59, %f58;
|
||||
mul.f32 %f65, %f63, %f63;
|
||||
fma.rn.f32 %f66, %f62, %f62, %f65;
|
||||
fma.rn.f32 %f67, %f64, %f64, %f66;
|
||||
sqrt.rn.f32 %f68, %f67;
|
||||
rcp.rn.f32 %f69, %f68;
|
||||
mul.f32 %f70, %f69, %f62;
|
||||
mul.f32 %f71, %f69, %f63;
|
||||
mul.f32 %f72, %f69, %f64;
|
||||
st.global.f32 [normal], %f70;
|
||||
st.global.f32 [normal+4], %f71;
|
||||
st.global.f32 [normal+8], %f72;
|
||||
// inline asm
|
||||
call (%rd76), _rt_buffer_get_64, (%rd59, %r6, %r44, %rd78, %rd7, %rd7, %rd7);
|
||||
// inline asm
|
||||
ld.v2.f32 {%f86, %f74}, [%rd76+24];
|
||||
// inline asm
|
||||
call (%rd82), _rt_buffer_get_64, (%rd59, %r6, %r44, %rd84, %rd7, %rd7, %rd7);
|
||||
// inline asm
|
||||
ld.v2.f32 {%f75, %f76}, [%rd82+24];
|
||||
// inline asm
|
||||
call (%rd88), _rt_buffer_get_64, (%rd59, %r6, %r44, %rd90, %rd7, %rd7, %rd7);
|
||||
// inline asm
|
||||
setp.lt.f32 %p3, %f74, 0f00000000;
|
||||
@%p3 bra BB0_4;
|
||||
bra.uni BB0_3;
|
||||
|
||||
BB0_4:
|
||||
add.f32 %f87, %f74, 0f3F800000;
|
||||
bra.uni BB0_5;
|
||||
|
||||
BB0_3:
|
||||
ld.v2.f32 {%f77, %f78}, [%rd88+24];
|
||||
mul.f32 %f81, %f15, %f75;
|
||||
mul.f32 %f82, %f15, %f76;
|
||||
fma.rn.f32 %f83, %f3, %f86, %f81;
|
||||
fma.rn.f32 %f84, %f3, %f74, %f82;
|
||||
fma.rn.f32 %f86, %f16, %f77, %f83;
|
||||
fma.rn.f32 %f87, %f16, %f78, %f84;
|
||||
|
||||
BB0_5:
|
||||
st.global.v2.f32 [texCoords], {%f86, %f87};
|
||||
ret;
|
||||
}
|
||||
|
||||
|
||||
8
Assets/Editor/x64/Bakery/attrib_trimeshTex.ptx.meta
Normal file
8
Assets/Editor/x64/Bakery/attrib_trimeshTex.ptx.meta
Normal file
@@ -0,0 +1,8 @@
|
||||
fileFormatVersion: 2
|
||||
guid: 855bb282e20f49b4cb68df002aea39ce
|
||||
timeCreated: 1549654381
|
||||
licenseType: Store
|
||||
DefaultImporter:
|
||||
userData:
|
||||
assetBundleName:
|
||||
assetBundleVariant:
|
||||
341
Assets/Editor/x64/Bakery/attrib_trimeshTexLODselect.ptx
Normal file
341
Assets/Editor/x64/Bakery/attrib_trimeshTexLODselect.ptx
Normal file
@@ -0,0 +1,341 @@
|
||||
//
|
||||
// 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 _Z11interpolatei
|
||||
.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 8 .b8 texCoords[8];
|
||||
.global .align 4 .f32 triAlpha;
|
||||
.global .align 1 .b8 vertex_buffer[1];
|
||||
.global .align 1 .b8 index_buffer[1];
|
||||
.global .align 4 .u32 firstAlphaTriangle;
|
||||
.global .align 1 .b8 vertex_buffer_uv[1];
|
||||
.global .align 1 .b8 triangleAlphaIDs[1];
|
||||
.global .align 1 .b8 alphaTextures[1];
|
||||
.global .align 4 .b8 ray[36];
|
||||
.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_typeinfo9texCoordsE[8] = {82, 97, 121, 0, 8, 0, 0, 0};
|
||||
.global .align 4 .b8 _ZN21rti_internal_typeinfo8triAlphaE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
|
||||
.global .align 4 .b8 _ZN21rti_internal_typeinfo18firstAlphaTriangleE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
|
||||
.global .align 4 .b8 _ZN21rti_internal_typeinfo3rayE[8] = {82, 97, 121, 0, 36, 0, 0, 0};
|
||||
.global .align 8 .u64 _ZN21rti_internal_register20reg_bitness_detectorE;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail0E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail1E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail2E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail3E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail4E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail5E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail6E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail7E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail8E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail9E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail0E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail1E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail2E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail3E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail4E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail5E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail6E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail7E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail8E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail9E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register14reg_rayIndex_xE;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register14reg_rayIndex_yE;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register14reg_rayIndex_zE;
|
||||
.global .align 8 .b8 _ZN21rti_internal_typename7pixelIDE[6] = {117, 105, 110, 116, 50, 0};
|
||||
.global .align 8 .b8 _ZN21rti_internal_typename10resolutionE[6] = {117, 105, 110, 116, 50, 0};
|
||||
.global .align 8 .b8 _ZN21rti_internal_typename6normalE[7] = {102, 108, 111, 97, 116, 51, 0};
|
||||
.global .align 8 .b8 _ZN21rti_internal_typename6camPosE[7] = {102, 108, 111, 97, 116, 51, 0};
|
||||
.global .align 16 .b8 _ZN21rti_internal_typename4rootE[9] = {114, 116, 79, 98, 106, 101, 99, 116, 0};
|
||||
.global .align 4 .b8 _ZN21rti_internal_typename12imageEnabledE[4] = {105, 110, 116, 0};
|
||||
.global .align 8 .b8 _ZN21rti_internal_typename8tileInfoE[6] = {117, 105, 110, 116, 52, 0};
|
||||
.global .align 4 .b8 _ZN21rti_internal_typename8additiveE[4] = {105, 110, 116, 0};
|
||||
.global .align 8 .b8 _ZN21rti_internal_typename9texCoordsE[7] = {102, 108, 111, 97, 116, 50, 0};
|
||||
.global .align 8 .b8 _ZN21rti_internal_typename8triAlphaE[6] = {102, 108, 111, 97, 116, 0};
|
||||
.global .align 4 .b8 _ZN21rti_internal_typename18firstAlphaTriangleE[4] = {105, 110, 116, 0};
|
||||
.global .align 4 .b8 _ZN21rti_internal_typename3rayE[4] = {82, 97, 121, 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_typeenum9texCoordsE = 4919;
|
||||
.global .align 4 .u32 _ZN21rti_internal_typeenum8triAlphaE = 4919;
|
||||
.global .align 4 .u32 _ZN21rti_internal_typeenum18firstAlphaTriangleE = 4919;
|
||||
.global .align 4 .u32 _ZN21rti_internal_typeenum3rayE = 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 16 .b8 _ZN21rti_internal_semantic9texCoordsE[20] = {97, 116, 116, 114, 105, 98, 117, 116, 101, 32, 116, 101, 120, 67, 111, 111, 114, 100, 115, 0};
|
||||
.global .align 16 .b8 _ZN21rti_internal_semantic8triAlphaE[19] = {97, 116, 116, 114, 105, 98, 117, 116, 101, 32, 116, 114, 105, 65, 108, 112, 104, 97, 0};
|
||||
.global .align 1 .b8 _ZN21rti_internal_semantic18firstAlphaTriangleE[1];
|
||||
.global .align 16 .b8 _ZN21rti_internal_semantic3rayE[13] = {114, 116, 67, 117, 114, 114, 101, 110, 116, 82, 97, 121, 0};
|
||||
.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_annotation9texCoordsE[1];
|
||||
.global .align 1 .b8 _ZN23rti_internal_annotation8triAlphaE[1];
|
||||
.global .align 1 .b8 _ZN23rti_internal_annotation18firstAlphaTriangleE[1];
|
||||
.global .align 1 .b8 _ZN23rti_internal_annotation3rayE[1];
|
||||
|
||||
.visible .entry _Z11interpolatei(
|
||||
.param .u32 _Z11interpolatei_param_0
|
||||
)
|
||||
{
|
||||
.reg .pred %p<3>;
|
||||
.reg .f32 %f<134>;
|
||||
.reg .b32 %r<51>;
|
||||
.reg .b64 %rd<112>;
|
||||
|
||||
|
||||
ld.param.u32 %r5, [_Z11interpolatei_param_0];
|
||||
// inline asm
|
||||
call (%f6, %f7), _rt_get_triangle_barycentrics, ();
|
||||
// inline asm
|
||||
cvt.s64.s32 %rd3, %r5;
|
||||
mov.u64 %rd7, index_buffer;
|
||||
cvta.global.u64 %rd2, %rd7;
|
||||
mov.u32 %r6, 1;
|
||||
mov.u32 %r7, 12;
|
||||
mov.u64 %rd6, 0;
|
||||
// inline asm
|
||||
call (%rd1), _rt_buffer_get_64, (%rd2, %r6, %r7, %rd3, %rd6, %rd6, %rd6);
|
||||
// inline asm
|
||||
ld.u32 %r1, [%rd1];
|
||||
ld.u32 %r2, [%rd1+4];
|
||||
ld.u32 %r3, [%rd1+8];
|
||||
ld.global.u32 %r4, [firstAlphaTriangle];
|
||||
setp.gt.s32 %p1, %r4, %r5;
|
||||
mov.f32 %f133, 0f3F800000;
|
||||
sub.f32 %f9, %f133, %f6;
|
||||
sub.f32 %f3, %f9, %f7;
|
||||
@%p1 bra BB0_2;
|
||||
|
||||
cvt.s64.s32 %rd28, %r3;
|
||||
sub.s32 %r26, %r5, %r4;
|
||||
cvt.s64.s32 %rd10, %r26;
|
||||
mov.u64 %rd53, triangleAlphaIDs;
|
||||
cvta.global.u64 %rd9, %rd53;
|
||||
mov.u32 %r22, 4;
|
||||
// inline asm
|
||||
call (%rd8), _rt_buffer_get_64, (%rd9, %r6, %r22, %rd10, %rd6, %rd6, %rd6);
|
||||
// inline asm
|
||||
ld.u32 %rd54, [%rd8];
|
||||
cvt.u32.u64 %r27, %rd54;
|
||||
shr.u32 %r28, %r27, 16;
|
||||
cvt.s64.s32 %rd16, %r1;
|
||||
mov.u64 %rd55, vertex_buffer_uv;
|
||||
cvta.global.u64 %rd15, %rd55;
|
||||
mov.u32 %r15, 8;
|
||||
// inline asm
|
||||
call (%rd14), _rt_buffer_get_64, (%rd15, %r6, %r15, %rd16, %rd6, %rd6, %rd6);
|
||||
// inline asm
|
||||
ld.v2.f32 {%f10, %f11}, [%rd14];
|
||||
cvt.s64.s32 %rd22, %r2;
|
||||
// inline asm
|
||||
call (%rd20), _rt_buffer_get_64, (%rd15, %r6, %r15, %rd22, %rd6, %rd6, %rd6);
|
||||
// inline asm
|
||||
ld.v2.f32 {%f14, %f15}, [%rd20];
|
||||
// inline asm
|
||||
call (%rd26), _rt_buffer_get_64, (%rd15, %r6, %r15, %rd28, %rd6, %rd6, %rd6);
|
||||
// inline asm
|
||||
mul.f32 %f18, %f6, %f14;
|
||||
mul.f32 %f19, %f6, %f15;
|
||||
fma.rn.f32 %f20, %f3, %f10, %f18;
|
||||
fma.rn.f32 %f21, %f3, %f11, %f19;
|
||||
ld.v2.f32 {%f22, %f23}, [%rd26];
|
||||
fma.rn.f32 %f26, %f7, %f22, %f20;
|
||||
fma.rn.f32 %f27, %f7, %f23, %f21;
|
||||
abs.f32 %f28, %f26;
|
||||
cvt.rmi.f32.f32 %f29, %f28;
|
||||
sub.f32 %f30, %f28, %f29;
|
||||
abs.f32 %f31, %f27;
|
||||
cvt.rmi.f32.f32 %f32, %f31;
|
||||
sub.f32 %f33, %f31, %f32;
|
||||
and.b64 %rd34, %rd54, 65535;
|
||||
mov.u64 %rd56, alphaTextures;
|
||||
cvta.global.u64 %rd33, %rd56;
|
||||
// inline asm
|
||||
call (%rd32), _rt_buffer_get_64, (%rd33, %r6, %r22, %rd34, %rd6, %rd6, %rd6);
|
||||
// inline asm
|
||||
ld.u32 %r18, [%rd32];
|
||||
mov.u32 %r24, 2;
|
||||
// inline asm
|
||||
call (%rd38, %rd39, %rd40, %rd41), _rt_buffer_get_id_size_64, (%r18, %r24, %r6);
|
||||
// inline asm
|
||||
cvt.rn.f32.u64 %f34, %rd38;
|
||||
mul.f32 %f35, %f30, %f34;
|
||||
cvt.rzi.u32.f32 %r29, %f35;
|
||||
cvt.rn.f32.u64 %f36, %rd39;
|
||||
mul.f32 %f37, %f33, %f36;
|
||||
cvt.rzi.u32.f32 %r30, %f37;
|
||||
// inline asm
|
||||
call (%rd42), _rt_buffer_get_64, (%rd33, %r6, %r22, %rd34, %rd6, %rd6, %rd6);
|
||||
// inline asm
|
||||
ld.u32 %r23, [%rd42];
|
||||
cvt.u64.u32 %rd49, %r29;
|
||||
cvt.u64.u32 %rd50, %r30;
|
||||
// inline asm
|
||||
call (%rd48), _rt_buffer_get_id_64, (%r23, %r24, %r6, %rd49, %rd50, %rd6, %rd6);
|
||||
// inline asm
|
||||
ld.u8 %r31, [%rd48];
|
||||
and.b32 %r32, %r31, %r28;
|
||||
setp.eq.s32 %p2, %r32, 0;
|
||||
selp.f32 %f133, 0f00000000, 0f3F800000, %p2;
|
||||
|
||||
BB0_2:
|
||||
st.global.f32 [triAlpha], %f133;
|
||||
cvt.s64.s32 %rd95, %r1;
|
||||
mov.u64 %rd111, vertex_buffer;
|
||||
cvta.global.u64 %rd58, %rd111;
|
||||
mov.u32 %r50, 32;
|
||||
// inline asm
|
||||
call (%rd57), _rt_buffer_get_64, (%rd58, %r6, %r50, %rd95, %rd6, %rd6, %rd6);
|
||||
// inline asm
|
||||
ld.f32 %f38, [%rd57+8];
|
||||
ld.v2.f32 {%f39, %f40}, [%rd57];
|
||||
cvt.s64.s32 %rd101, %r2;
|
||||
// inline asm
|
||||
call (%rd63), _rt_buffer_get_64, (%rd58, %r6, %r50, %rd101, %rd6, %rd6, %rd6);
|
||||
// inline asm
|
||||
ld.f32 %f43, [%rd63+8];
|
||||
ld.v2.f32 {%f44, %f45}, [%rd63];
|
||||
cvt.s64.s32 %rd107, %r3;
|
||||
// inline asm
|
||||
call (%rd69), _rt_buffer_get_64, (%rd58, %r6, %r50, %rd107, %rd6, %rd6, %rd6);
|
||||
// inline asm
|
||||
ld.f32 %f48, [%rd69+8];
|
||||
ld.v2.f32 {%f49, %f50}, [%rd69];
|
||||
// inline asm
|
||||
call (%rd75), _rt_buffer_get_64, (%rd58, %r6, %r50, %rd95, %rd6, %rd6, %rd6);
|
||||
// inline asm
|
||||
ld.f32 %f53, [%rd75+20];
|
||||
ld.f32 %f54, [%rd75+16];
|
||||
ld.f32 %f55, [%rd75+12];
|
||||
// inline asm
|
||||
call (%rd81), _rt_buffer_get_64, (%rd58, %r6, %r50, %rd101, %rd6, %rd6, %rd6);
|
||||
// inline asm
|
||||
ld.f32 %f56, [%rd81+20];
|
||||
ld.f32 %f57, [%rd81+16];
|
||||
ld.f32 %f58, [%rd81+12];
|
||||
// inline asm
|
||||
call (%rd87), _rt_buffer_get_64, (%rd58, %r6, %r50, %rd107, %rd6, %rd6, %rd6);
|
||||
// inline asm
|
||||
mul.f32 %f59, %f6, %f58;
|
||||
mul.f32 %f60, %f6, %f57;
|
||||
mul.f32 %f61, %f6, %f56;
|
||||
fma.rn.f32 %f62, %f3, %f55, %f59;
|
||||
fma.rn.f32 %f63, %f3, %f54, %f60;
|
||||
fma.rn.f32 %f64, %f3, %f53, %f61;
|
||||
ld.f32 %f65, [%rd87+20];
|
||||
ld.f32 %f66, [%rd87+16];
|
||||
ld.f32 %f67, [%rd87+12];
|
||||
fma.rn.f32 %f68, %f7, %f67, %f62;
|
||||
fma.rn.f32 %f69, %f7, %f66, %f63;
|
||||
fma.rn.f32 %f70, %f7, %f65, %f64;
|
||||
mul.f32 %f71, %f69, %f69;
|
||||
fma.rn.f32 %f72, %f68, %f68, %f71;
|
||||
fma.rn.f32 %f73, %f70, %f70, %f72;
|
||||
sqrt.rn.f32 %f74, %f73;
|
||||
rcp.rn.f32 %f75, %f74;
|
||||
mul.f32 %f76, %f75, %f68;
|
||||
mul.f32 %f77, %f75, %f69;
|
||||
mul.f32 %f78, %f75, %f70;
|
||||
st.global.f32 [normal], %f76;
|
||||
st.global.f32 [normal+4], %f77;
|
||||
st.global.f32 [normal+8], %f78;
|
||||
// inline asm
|
||||
call (%rd93), _rt_buffer_get_64, (%rd58, %r6, %r50, %rd95, %rd6, %rd6, %rd6);
|
||||
// inline asm
|
||||
ld.v2.f32 {%f79, %f80}, [%rd93+24];
|
||||
// inline asm
|
||||
call (%rd99), _rt_buffer_get_64, (%rd58, %r6, %r50, %rd101, %rd6, %rd6, %rd6);
|
||||
// inline asm
|
||||
ld.v2.f32 {%f83, %f84}, [%rd99+24];
|
||||
// inline asm
|
||||
call (%rd105), _rt_buffer_get_64, (%rd58, %r6, %r50, %rd107, %rd6, %rd6, %rd6);
|
||||
// inline asm
|
||||
ld.v2.f32 {%f87, %f88}, [%rd105+24];
|
||||
mul.f32 %f91, %f6, %f83;
|
||||
fma.rn.f32 %f92, %f3, %f79, %f91;
|
||||
sub.f32 %f93, %f44, %f39;
|
||||
sub.f32 %f94, %f45, %f40;
|
||||
sub.f32 %f95, %f43, %f38;
|
||||
sub.f32 %f96, %f49, %f39;
|
||||
sub.f32 %f97, %f50, %f40;
|
||||
sub.f32 %f98, %f48, %f38;
|
||||
mul.f32 %f99, %f94, %f98;
|
||||
mul.f32 %f100, %f95, %f97;
|
||||
sub.f32 %f101, %f99, %f100;
|
||||
mul.f32 %f102, %f95, %f96;
|
||||
mul.f32 %f103, %f93, %f98;
|
||||
sub.f32 %f104, %f102, %f103;
|
||||
mul.f32 %f105, %f93, %f97;
|
||||
mul.f32 %f106, %f94, %f96;
|
||||
sub.f32 %f107, %f105, %f106;
|
||||
mul.f32 %f108, %f104, %f104;
|
||||
fma.rn.f32 %f109, %f101, %f101, %f108;
|
||||
fma.rn.f32 %f110, %f107, %f107, %f109;
|
||||
sqrt.rn.f32 %f111, %f110;
|
||||
mul.f32 %f112, %f111, 0f3F000000;
|
||||
sub.f32 %f113, %f83, %f79;
|
||||
sub.f32 %f114, %f84, %f80;
|
||||
sub.f32 %f115, %f87, %f79;
|
||||
sub.f32 %f116, %f88, %f80;
|
||||
mul.f32 %f117, %f114, 0f00000000;
|
||||
mul.f32 %f118, %f116, 0f00000000;
|
||||
sub.f32 %f119, %f117, %f118;
|
||||
mul.f32 %f120, %f115, 0f00000000;
|
||||
mul.f32 %f121, %f113, 0f00000000;
|
||||
sub.f32 %f122, %f120, %f121;
|
||||
mul.f32 %f123, %f113, %f116;
|
||||
mul.f32 %f124, %f114, %f115;
|
||||
sub.f32 %f125, %f123, %f124;
|
||||
mul.f32 %f126, %f122, %f122;
|
||||
fma.rn.f32 %f127, %f119, %f119, %f126;
|
||||
fma.rn.f32 %f128, %f125, %f125, %f127;
|
||||
sqrt.rn.f32 %f129, %f128;
|
||||
mul.f32 %f130, %f129, 0f3F000000;
|
||||
div.rn.f32 %f131, %f112, %f130;
|
||||
fma.rn.f32 %f132, %f7, %f87, %f92;
|
||||
st.global.v2.f32 [texCoords], {%f132, %f131};
|
||||
ret;
|
||||
}
|
||||
|
||||
|
||||
@@ -0,0 +1,8 @@
|
||||
fileFormatVersion: 2
|
||||
guid: aac069d2f1ef4f54b99e859563c0b380
|
||||
timeCreated: 1552566788
|
||||
licenseType: Store
|
||||
DefaultImporter:
|
||||
userData:
|
||||
assetBundleName:
|
||||
assetBundleVariant:
|
||||
114
Assets/Editor/x64/Bakery/clip_rtx.ptx
Normal file
114
Assets/Editor/x64/Bakery/clip_rtx.ptx
Normal file
@@ -0,0 +1,114 @@
|
||||
//
|
||||
// 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 4 .f32 triAlpha;
|
||||
.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_typeinfo8triAlphaE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
|
||||
.global .align 8 .u64 _ZN21rti_internal_register20reg_bitness_detectorE;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail0E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail1E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail2E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail3E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail4E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail5E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail6E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail7E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail8E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail9E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail0E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail1E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail2E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail3E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail4E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail5E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail6E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail7E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail8E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail9E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register14reg_rayIndex_xE;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register14reg_rayIndex_yE;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register14reg_rayIndex_zE;
|
||||
.global .align 8 .b8 _ZN21rti_internal_typename7pixelIDE[6] = {117, 105, 110, 116, 50, 0};
|
||||
.global .align 8 .b8 _ZN21rti_internal_typename10resolutionE[6] = {117, 105, 110, 116, 50, 0};
|
||||
.global .align 8 .b8 _ZN21rti_internal_typename6normalE[7] = {102, 108, 111, 97, 116, 51, 0};
|
||||
.global .align 8 .b8 _ZN21rti_internal_typename6camPosE[7] = {102, 108, 111, 97, 116, 51, 0};
|
||||
.global .align 16 .b8 _ZN21rti_internal_typename4rootE[9] = {114, 116, 79, 98, 106, 101, 99, 116, 0};
|
||||
.global .align 4 .b8 _ZN21rti_internal_typename12imageEnabledE[4] = {105, 110, 116, 0};
|
||||
.global .align 8 .b8 _ZN21rti_internal_typename8tileInfoE[6] = {117, 105, 110, 116, 52, 0};
|
||||
.global .align 4 .b8 _ZN21rti_internal_typename8additiveE[4] = {105, 110, 116, 0};
|
||||
.global .align 8 .b8 _ZN21rti_internal_typename8triAlphaE[6] = {102, 108, 111, 97, 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_typeenum8triAlphaE = 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 16 .b8 _ZN21rti_internal_semantic8triAlphaE[19] = {97, 116, 116, 114, 105, 98, 117, 116, 101, 32, 116, 114, 105, 65, 108, 112, 104, 97, 0};
|
||||
.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_annotation8triAlphaE[1];
|
||||
|
||||
.visible .entry _Z6oxMainv(
|
||||
|
||||
)
|
||||
{
|
||||
.reg .pred %p<2>;
|
||||
.reg .f32 %f<2>;
|
||||
|
||||
|
||||
ld.global.f32 %f1, [triAlpha];
|
||||
setp.neu.f32 %p1, %f1, 0f00000000;
|
||||
@%p1 bra BB0_2;
|
||||
|
||||
// inline asm
|
||||
call _rt_ignore_intersection, ();
|
||||
// inline asm
|
||||
|
||||
BB0_2:
|
||||
ret;
|
||||
}
|
||||
|
||||
|
||||
8
Assets/Editor/x64/Bakery/clip_rtx.ptx.meta
Normal file
8
Assets/Editor/x64/Bakery/clip_rtx.ptx.meta
Normal file
@@ -0,0 +1,8 @@
|
||||
fileFormatVersion: 2
|
||||
guid: 55fda1d6fc84d9e42838115c0ee88ec2
|
||||
timeCreated: 1549922863
|
||||
licenseType: Store
|
||||
DefaultImporter:
|
||||
userData:
|
||||
assetBundleName:
|
||||
assetBundleVariant:
|
||||
BIN
Assets/Editor/x64/Bakery/combineMasks.exe
Normal file
BIN
Assets/Editor/x64/Bakery/combineMasks.exe
Normal file
Binary file not shown.
8
Assets/Editor/x64/Bakery/combineMasks.exe.meta
Normal file
8
Assets/Editor/x64/Bakery/combineMasks.exe.meta
Normal file
@@ -0,0 +1,8 @@
|
||||
fileFormatVersion: 2
|
||||
guid: 5aaaa91ab09a47640a15dcc03d1d0fac
|
||||
timeCreated: 1534844129
|
||||
licenseType: Store
|
||||
DefaultImporter:
|
||||
userData:
|
||||
assetBundleName:
|
||||
assetBundleVariant:
|
||||
BIN
Assets/Editor/x64/Bakery/cookieOutput.pso
Normal file
BIN
Assets/Editor/x64/Bakery/cookieOutput.pso
Normal file
Binary file not shown.
8
Assets/Editor/x64/Bakery/cookieOutput.pso.meta
Normal file
8
Assets/Editor/x64/Bakery/cookieOutput.pso.meta
Normal file
@@ -0,0 +1,8 @@
|
||||
fileFormatVersion: 2
|
||||
guid: dd0648d1aa6ef1846b294f90270347e7
|
||||
timeCreated: 1526155640
|
||||
licenseType: Store
|
||||
DefaultImporter:
|
||||
userData:
|
||||
assetBundleName:
|
||||
assetBundleVariant:
|
||||
BIN
Assets/Editor/x64/Bakery/cudart32_91.dll
LFS
Normal file
BIN
Assets/Editor/x64/Bakery/cudart32_91.dll
LFS
Normal file
Binary file not shown.
115
Assets/Editor/x64/Bakery/cudart32_91.dll.meta
Normal file
115
Assets/Editor/x64/Bakery/cudart32_91.dll.meta
Normal file
@@ -0,0 +1,115 @@
|
||||
fileFormatVersion: 2
|
||||
guid: 4e60d954644e0a145b35eea29aeba48a
|
||||
timeCreated: 1526150503
|
||||
licenseType: Store
|
||||
PluginImporter:
|
||||
serializedVersion: 2
|
||||
iconMap: {}
|
||||
executionOrder: {}
|
||||
isPreloaded: 0
|
||||
isOverridable: 0
|
||||
platformData:
|
||||
data:
|
||||
first:
|
||||
'': Any
|
||||
second:
|
||||
enabled: 0
|
||||
settings:
|
||||
Exclude Android: 1
|
||||
Exclude Editor: 0
|
||||
Exclude Linux: 1
|
||||
Exclude Linux64: 1
|
||||
Exclude LinuxUniversal: 1
|
||||
Exclude OSXIntel: 1
|
||||
Exclude OSXIntel64: 1
|
||||
Exclude OSXUniversal: 1
|
||||
Exclude WebGL: 1
|
||||
Exclude Win: 1
|
||||
Exclude Win64: 1
|
||||
Exclude iOS: 1
|
||||
data:
|
||||
first:
|
||||
Any:
|
||||
second:
|
||||
enabled: 0
|
||||
settings: {}
|
||||
data:
|
||||
first:
|
||||
Editor: Editor
|
||||
second:
|
||||
enabled: 1
|
||||
settings:
|
||||
DefaultValueInitialized: true
|
||||
data:
|
||||
first:
|
||||
Facebook: Win
|
||||
second:
|
||||
enabled: 0
|
||||
settings:
|
||||
CPU: None
|
||||
data:
|
||||
first:
|
||||
Facebook: Win64
|
||||
second:
|
||||
enabled: 0
|
||||
settings:
|
||||
CPU: None
|
||||
data:
|
||||
first:
|
||||
Standalone: Linux
|
||||
second:
|
||||
enabled: 0
|
||||
settings:
|
||||
CPU: None
|
||||
data:
|
||||
first:
|
||||
Standalone: Linux64
|
||||
second:
|
||||
enabled: 0
|
||||
settings:
|
||||
CPU: None
|
||||
data:
|
||||
first:
|
||||
Standalone: LinuxUniversal
|
||||
second:
|
||||
enabled: 0
|
||||
settings:
|
||||
CPU: None
|
||||
data:
|
||||
first:
|
||||
Standalone: OSXIntel
|
||||
second:
|
||||
enabled: 0
|
||||
settings:
|
||||
CPU: None
|
||||
data:
|
||||
first:
|
||||
Standalone: OSXIntel64
|
||||
second:
|
||||
enabled: 0
|
||||
settings:
|
||||
CPU: None
|
||||
data:
|
||||
first:
|
||||
Standalone: OSXUniversal
|
||||
second:
|
||||
enabled: 0
|
||||
settings:
|
||||
CPU: None
|
||||
data:
|
||||
first:
|
||||
Standalone: Win
|
||||
second:
|
||||
enabled: 0
|
||||
settings:
|
||||
CPU: None
|
||||
data:
|
||||
first:
|
||||
Standalone: Win64
|
||||
second:
|
||||
enabled: 0
|
||||
settings:
|
||||
CPU: None
|
||||
userData:
|
||||
assetBundleName:
|
||||
assetBundleVariant:
|
||||
BIN
Assets/Editor/x64/Bakery/cudart64_90.dll
LFS
Normal file
BIN
Assets/Editor/x64/Bakery/cudart64_90.dll
LFS
Normal file
Binary file not shown.
115
Assets/Editor/x64/Bakery/cudart64_90.dll.meta
Normal file
115
Assets/Editor/x64/Bakery/cudart64_90.dll.meta
Normal file
@@ -0,0 +1,115 @@
|
||||
fileFormatVersion: 2
|
||||
guid: dce443ff26676e84a942d0b78415b969
|
||||
timeCreated: 1526150504
|
||||
licenseType: Store
|
||||
PluginImporter:
|
||||
serializedVersion: 2
|
||||
iconMap: {}
|
||||
executionOrder: {}
|
||||
isPreloaded: 0
|
||||
isOverridable: 0
|
||||
platformData:
|
||||
data:
|
||||
first:
|
||||
'': Any
|
||||
second:
|
||||
enabled: 0
|
||||
settings:
|
||||
Exclude Android: 1
|
||||
Exclude Editor: 0
|
||||
Exclude Linux: 1
|
||||
Exclude Linux64: 1
|
||||
Exclude LinuxUniversal: 1
|
||||
Exclude OSXIntel: 1
|
||||
Exclude OSXIntel64: 1
|
||||
Exclude OSXUniversal: 1
|
||||
Exclude WebGL: 1
|
||||
Exclude Win: 1
|
||||
Exclude Win64: 1
|
||||
Exclude iOS: 1
|
||||
data:
|
||||
first:
|
||||
Any:
|
||||
second:
|
||||
enabled: 0
|
||||
settings: {}
|
||||
data:
|
||||
first:
|
||||
Editor: Editor
|
||||
second:
|
||||
enabled: 1
|
||||
settings:
|
||||
DefaultValueInitialized: true
|
||||
data:
|
||||
first:
|
||||
Facebook: Win
|
||||
second:
|
||||
enabled: 0
|
||||
settings:
|
||||
CPU: None
|
||||
data:
|
||||
first:
|
||||
Facebook: Win64
|
||||
second:
|
||||
enabled: 0
|
||||
settings:
|
||||
CPU: None
|
||||
data:
|
||||
first:
|
||||
Standalone: Linux
|
||||
second:
|
||||
enabled: 0
|
||||
settings:
|
||||
CPU: None
|
||||
data:
|
||||
first:
|
||||
Standalone: Linux64
|
||||
second:
|
||||
enabled: 0
|
||||
settings:
|
||||
CPU: None
|
||||
data:
|
||||
first:
|
||||
Standalone: LinuxUniversal
|
||||
second:
|
||||
enabled: 0
|
||||
settings:
|
||||
CPU: None
|
||||
data:
|
||||
first:
|
||||
Standalone: OSXIntel
|
||||
second:
|
||||
enabled: 0
|
||||
settings:
|
||||
CPU: None
|
||||
data:
|
||||
first:
|
||||
Standalone: OSXIntel64
|
||||
second:
|
||||
enabled: 0
|
||||
settings:
|
||||
CPU: None
|
||||
data:
|
||||
first:
|
||||
Standalone: OSXUniversal
|
||||
second:
|
||||
enabled: 0
|
||||
settings:
|
||||
CPU: None
|
||||
data:
|
||||
first:
|
||||
Standalone: Win
|
||||
second:
|
||||
enabled: 0
|
||||
settings:
|
||||
CPU: None
|
||||
data:
|
||||
first:
|
||||
Standalone: Win64
|
||||
second:
|
||||
enabled: 0
|
||||
settings:
|
||||
CPU: None
|
||||
userData:
|
||||
assetBundleName:
|
||||
assetBundleVariant:
|
||||
BIN
Assets/Editor/x64/Bakery/cudart64_91.dll
LFS
Normal file
BIN
Assets/Editor/x64/Bakery/cudart64_91.dll
LFS
Normal file
Binary file not shown.
115
Assets/Editor/x64/Bakery/cudart64_91.dll.meta
Normal file
115
Assets/Editor/x64/Bakery/cudart64_91.dll.meta
Normal file
@@ -0,0 +1,115 @@
|
||||
fileFormatVersion: 2
|
||||
guid: 76bf0bd45bbd4f34598f7ce6ac52f5e9
|
||||
timeCreated: 1526150503
|
||||
licenseType: Store
|
||||
PluginImporter:
|
||||
serializedVersion: 2
|
||||
iconMap: {}
|
||||
executionOrder: {}
|
||||
isPreloaded: 0
|
||||
isOverridable: 0
|
||||
platformData:
|
||||
data:
|
||||
first:
|
||||
'': Any
|
||||
second:
|
||||
enabled: 0
|
||||
settings:
|
||||
Exclude Android: 1
|
||||
Exclude Editor: 0
|
||||
Exclude Linux: 1
|
||||
Exclude Linux64: 1
|
||||
Exclude LinuxUniversal: 1
|
||||
Exclude OSXIntel: 1
|
||||
Exclude OSXIntel64: 1
|
||||
Exclude OSXUniversal: 1
|
||||
Exclude WebGL: 1
|
||||
Exclude Win: 1
|
||||
Exclude Win64: 1
|
||||
Exclude iOS: 1
|
||||
data:
|
||||
first:
|
||||
Any:
|
||||
second:
|
||||
enabled: 0
|
||||
settings: {}
|
||||
data:
|
||||
first:
|
||||
Editor: Editor
|
||||
second:
|
||||
enabled: 1
|
||||
settings:
|
||||
DefaultValueInitialized: true
|
||||
data:
|
||||
first:
|
||||
Facebook: Win
|
||||
second:
|
||||
enabled: 0
|
||||
settings:
|
||||
CPU: None
|
||||
data:
|
||||
first:
|
||||
Facebook: Win64
|
||||
second:
|
||||
enabled: 0
|
||||
settings:
|
||||
CPU: None
|
||||
data:
|
||||
first:
|
||||
Standalone: Linux
|
||||
second:
|
||||
enabled: 0
|
||||
settings:
|
||||
CPU: None
|
||||
data:
|
||||
first:
|
||||
Standalone: Linux64
|
||||
second:
|
||||
enabled: 0
|
||||
settings:
|
||||
CPU: None
|
||||
data:
|
||||
first:
|
||||
Standalone: LinuxUniversal
|
||||
second:
|
||||
enabled: 0
|
||||
settings:
|
||||
CPU: None
|
||||
data:
|
||||
first:
|
||||
Standalone: OSXIntel
|
||||
second:
|
||||
enabled: 0
|
||||
settings:
|
||||
CPU: None
|
||||
data:
|
||||
first:
|
||||
Standalone: OSXIntel64
|
||||
second:
|
||||
enabled: 0
|
||||
settings:
|
||||
CPU: None
|
||||
data:
|
||||
first:
|
||||
Standalone: OSXUniversal
|
||||
second:
|
||||
enabled: 0
|
||||
settings:
|
||||
CPU: None
|
||||
data:
|
||||
first:
|
||||
Standalone: Win
|
||||
second:
|
||||
enabled: 0
|
||||
settings:
|
||||
CPU: None
|
||||
data:
|
||||
first:
|
||||
Standalone: Win64
|
||||
second:
|
||||
enabled: 0
|
||||
settings:
|
||||
CPU: None
|
||||
userData:
|
||||
assetBundleName:
|
||||
assetBundleVariant:
|
||||
BIN
Assets/Editor/x64/Bakery/cudnn64_7.dll
LFS
Normal file
BIN
Assets/Editor/x64/Bakery/cudnn64_7.dll
LFS
Normal file
Binary file not shown.
28
Assets/Editor/x64/Bakery/cudnn64_7.dll.meta
Normal file
28
Assets/Editor/x64/Bakery/cudnn64_7.dll.meta
Normal file
@@ -0,0 +1,28 @@
|
||||
fileFormatVersion: 2
|
||||
guid: 4442e59a46322ab45ae831f1bd869c5c
|
||||
timeCreated: 1526331884
|
||||
licenseType: Store
|
||||
PluginImporter:
|
||||
serializedVersion: 2
|
||||
iconMap: {}
|
||||
executionOrder: {}
|
||||
isPreloaded: 0
|
||||
isOverridable: 0
|
||||
platformData:
|
||||
data:
|
||||
first:
|
||||
Any:
|
||||
second:
|
||||
enabled: 0
|
||||
settings: {}
|
||||
data:
|
||||
first:
|
||||
Editor: Editor
|
||||
second:
|
||||
enabled: 1
|
||||
settings:
|
||||
CPU: x86_64
|
||||
DefaultValueInitialized: true
|
||||
userData:
|
||||
assetBundleName:
|
||||
assetBundleVariant:
|
||||
BIN
Assets/Editor/x64/Bakery/d3dcompiler_47.dll
LFS
Normal file
BIN
Assets/Editor/x64/Bakery/d3dcompiler_47.dll
LFS
Normal file
Binary file not shown.
138
Assets/Editor/x64/Bakery/d3dcompiler_47.dll.meta
Normal file
138
Assets/Editor/x64/Bakery/d3dcompiler_47.dll.meta
Normal file
@@ -0,0 +1,138 @@
|
||||
fileFormatVersion: 2
|
||||
guid: c17aba6c3878ffb4da8ecb9800b93d9e
|
||||
timeCreated: 1526150504
|
||||
licenseType: Store
|
||||
PluginImporter:
|
||||
serializedVersion: 2
|
||||
iconMap: {}
|
||||
executionOrder: {}
|
||||
isPreloaded: 0
|
||||
isOverridable: 0
|
||||
platformData:
|
||||
data:
|
||||
first:
|
||||
'': Any
|
||||
second:
|
||||
enabled: 0
|
||||
settings:
|
||||
Exclude Android: 1
|
||||
Exclude Editor: 0
|
||||
Exclude Linux: 1
|
||||
Exclude Linux64: 1
|
||||
Exclude LinuxUniversal: 1
|
||||
Exclude OSXIntel: 1
|
||||
Exclude OSXIntel64: 1
|
||||
Exclude OSXUniversal: 1
|
||||
Exclude WebGL: 1
|
||||
Exclude Win: 1
|
||||
Exclude Win64: 1
|
||||
Exclude iOS: 1
|
||||
data:
|
||||
first:
|
||||
'': Editor
|
||||
second:
|
||||
enabled: 0
|
||||
settings:
|
||||
CPU: AnyCPU
|
||||
OS: AnyOS
|
||||
data:
|
||||
first:
|
||||
Android: Android
|
||||
second:
|
||||
enabled: 0
|
||||
settings:
|
||||
CPU: ARMv7
|
||||
data:
|
||||
first:
|
||||
Any:
|
||||
second:
|
||||
enabled: 0
|
||||
settings: {}
|
||||
data:
|
||||
first:
|
||||
Editor: Editor
|
||||
second:
|
||||
enabled: 1
|
||||
settings:
|
||||
DefaultValueInitialized: true
|
||||
data:
|
||||
first:
|
||||
Facebook: Win
|
||||
second:
|
||||
enabled: 0
|
||||
settings:
|
||||
CPU: AnyCPU
|
||||
data:
|
||||
first:
|
||||
Facebook: Win64
|
||||
second:
|
||||
enabled: 0
|
||||
settings:
|
||||
CPU: AnyCPU
|
||||
data:
|
||||
first:
|
||||
Standalone: Linux
|
||||
second:
|
||||
enabled: 0
|
||||
settings:
|
||||
CPU: x86
|
||||
data:
|
||||
first:
|
||||
Standalone: Linux64
|
||||
second:
|
||||
enabled: 0
|
||||
settings:
|
||||
CPU: x86_64
|
||||
data:
|
||||
first:
|
||||
Standalone: LinuxUniversal
|
||||
second:
|
||||
enabled: 0
|
||||
settings:
|
||||
CPU: None
|
||||
data:
|
||||
first:
|
||||
Standalone: OSXIntel
|
||||
second:
|
||||
enabled: 0
|
||||
settings:
|
||||
CPU: AnyCPU
|
||||
data:
|
||||
first:
|
||||
Standalone: OSXIntel64
|
||||
second:
|
||||
enabled: 0
|
||||
settings:
|
||||
CPU: AnyCPU
|
||||
data:
|
||||
first:
|
||||
Standalone: OSXUniversal
|
||||
second:
|
||||
enabled: 0
|
||||
settings:
|
||||
CPU: None
|
||||
data:
|
||||
first:
|
||||
Standalone: Win
|
||||
second:
|
||||
enabled: 0
|
||||
settings:
|
||||
CPU: AnyCPU
|
||||
data:
|
||||
first:
|
||||
Standalone: Win64
|
||||
second:
|
||||
enabled: 0
|
||||
settings:
|
||||
CPU: AnyCPU
|
||||
data:
|
||||
first:
|
||||
iPhone: iOS
|
||||
second:
|
||||
enabled: 0
|
||||
settings:
|
||||
CompileFlags:
|
||||
FrameworkDependencies:
|
||||
userData:
|
||||
assetBundleName:
|
||||
assetBundleVariant:
|
||||
BIN
Assets/Editor/x64/Bakery/debugLines.pso
Normal file
BIN
Assets/Editor/x64/Bakery/debugLines.pso
Normal file
Binary file not shown.
8
Assets/Editor/x64/Bakery/debugLines.pso.meta
Normal file
8
Assets/Editor/x64/Bakery/debugLines.pso.meta
Normal file
@@ -0,0 +1,8 @@
|
||||
fileFormatVersion: 2
|
||||
guid: 4c983cb8f06427e46be24df2114836d6
|
||||
timeCreated: 1580816320
|
||||
licenseType: Store
|
||||
DefaultImporter:
|
||||
userData:
|
||||
assetBundleName:
|
||||
assetBundleVariant:
|
||||
1050
Assets/Editor/x64/Bakery/denoiseFinish.ptx
Normal file
1050
Assets/Editor/x64/Bakery/denoiseFinish.ptx
Normal file
File diff suppressed because it is too large
Load Diff
8
Assets/Editor/x64/Bakery/denoiseFinish.ptx.meta
Normal file
8
Assets/Editor/x64/Bakery/denoiseFinish.ptx.meta
Normal file
@@ -0,0 +1,8 @@
|
||||
fileFormatVersion: 2
|
||||
guid: bbbbce75a9e818f4ea65df709d80bc4c
|
||||
timeCreated: 1526332011
|
||||
licenseType: Store
|
||||
DefaultImporter:
|
||||
userData:
|
||||
assetBundleName:
|
||||
assetBundleVariant:
|
||||
988
Assets/Editor/x64/Bakery/denoiseFinish72.ptx
Normal file
988
Assets/Editor/x64/Bakery/denoiseFinish72.ptx
Normal file
@@ -0,0 +1,988 @@
|
||||
//
|
||||
// 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 __raygen__oxMain
|
||||
.const .align 8 .b8 cs[32];
|
||||
|
||||
.visible .entry __raygen__oxMain(
|
||||
|
||||
)
|
||||
{
|
||||
.reg .pred %p<105>;
|
||||
.reg .b16 %rs<12>;
|
||||
.reg .f32 %f<674>;
|
||||
.reg .b32 %r<63>;
|
||||
.reg .b64 %rd<17>;
|
||||
|
||||
|
||||
// inline asm
|
||||
call (%r1), _optix_get_launch_index_x, ();
|
||||
// inline asm
|
||||
// inline asm
|
||||
call (%r2), _optix_get_launch_index_y, ();
|
||||
// inline asm
|
||||
ld.const.u64 %rd5, [cs+8];
|
||||
cvta.to.global.u64 %rd6, %rd5;
|
||||
ld.const.v2.u32 {%r4, %r5}, [cs+24];
|
||||
mad.lo.s32 %r7, %r4, %r2, %r1;
|
||||
cvt.u64.u32 %rd1, %r7;
|
||||
mul.wide.u32 %rd7, %r7, 16;
|
||||
add.s64 %rd8, %rd6, %rd7;
|
||||
ld.global.v4.f32 {%f96, %f97, %f98, %f99}, [%rd8];
|
||||
setp.eq.s32 %p7, %r5, 1;
|
||||
selp.f32 %f2, %f98, %f96, %p7;
|
||||
selp.f32 %f3, %f96, %f98, %p7;
|
||||
setp.eq.s32 %p8, %r5, 0;
|
||||
mov.f32 %f102, 0f3F8CCCCD;
|
||||
cvt.rzi.f32.f32 %f103, %f102;
|
||||
fma.rn.f32 %f104, %f103, 0fC0000000, 0f400CCCCD;
|
||||
abs.f32 %f4, %f104;
|
||||
@%p8 bra BB0_35;
|
||||
|
||||
ld.const.u64 %rd2, [cs];
|
||||
cvta.to.global.u64 %rd9, %rd2;
|
||||
shl.b64 %rd10, %rd1, 3;
|
||||
add.s64 %rd11, %rd9, %rd10;
|
||||
ld.global.v4.u16 {%rs4, %rs5, %rs6, %rs7}, [%rd11];
|
||||
// inline asm
|
||||
{ cvt.f32.f16 %f105, %rs4;}
|
||||
|
||||
// inline asm
|
||||
// inline asm
|
||||
{ cvt.f32.f16 %f106, %rs5;}
|
||||
|
||||
// inline asm
|
||||
// inline asm
|
||||
{ cvt.f32.f16 %f107, %rs6;}
|
||||
|
||||
// inline asm
|
||||
mov.f32 %f110, 0f3F800000;
|
||||
sub.f32 %f8, %f110, %f2;
|
||||
abs.f32 %f9, %f8;
|
||||
setp.lt.f32 %p9, %f9, 0f00800000;
|
||||
mul.f32 %f111, %f9, 0f4B800000;
|
||||
selp.f32 %f112, 0fC3170000, 0fC2FE0000, %p9;
|
||||
selp.f32 %f113, %f111, %f9, %p9;
|
||||
mov.b32 %r9, %f113;
|
||||
and.b32 %r10, %r9, 8388607;
|
||||
or.b32 %r11, %r10, 1065353216;
|
||||
mov.b32 %f114, %r11;
|
||||
shr.u32 %r12, %r9, 23;
|
||||
cvt.rn.f32.u32 %f115, %r12;
|
||||
add.f32 %f116, %f112, %f115;
|
||||
setp.gt.f32 %p10, %f114, 0f3FB504F3;
|
||||
mul.f32 %f117, %f114, 0f3F000000;
|
||||
add.f32 %f118, %f116, 0f3F800000;
|
||||
selp.f32 %f119, %f117, %f114, %p10;
|
||||
selp.f32 %f120, %f118, %f116, %p10;
|
||||
add.f32 %f121, %f119, 0fBF800000;
|
||||
add.f32 %f109, %f119, 0f3F800000;
|
||||
// inline asm
|
||||
rcp.approx.ftz.f32 %f108,%f109;
|
||||
// inline asm
|
||||
add.f32 %f122, %f121, %f121;
|
||||
mul.f32 %f123, %f108, %f122;
|
||||
mul.f32 %f124, %f123, %f123;
|
||||
mov.f32 %f125, 0f3C4CAF63;
|
||||
mov.f32 %f126, 0f3B18F0FE;
|
||||
fma.rn.f32 %f127, %f126, %f124, %f125;
|
||||
mov.f32 %f128, 0f3DAAAABD;
|
||||
fma.rn.f32 %f129, %f127, %f124, %f128;
|
||||
mul.rn.f32 %f130, %f129, %f124;
|
||||
mul.rn.f32 %f131, %f130, %f123;
|
||||
sub.f32 %f132, %f121, %f123;
|
||||
neg.f32 %f133, %f123;
|
||||
add.f32 %f134, %f132, %f132;
|
||||
fma.rn.f32 %f135, %f133, %f121, %f134;
|
||||
mul.rn.f32 %f136, %f108, %f135;
|
||||
add.f32 %f137, %f131, %f123;
|
||||
sub.f32 %f138, %f123, %f137;
|
||||
add.f32 %f139, %f131, %f138;
|
||||
add.f32 %f140, %f136, %f139;
|
||||
add.f32 %f141, %f137, %f140;
|
||||
sub.f32 %f142, %f137, %f141;
|
||||
add.f32 %f143, %f140, %f142;
|
||||
mov.f32 %f144, 0f3F317200;
|
||||
mul.rn.f32 %f145, %f120, %f144;
|
||||
mov.f32 %f146, 0f35BFBE8E;
|
||||
mul.rn.f32 %f147, %f120, %f146;
|
||||
add.f32 %f148, %f145, %f141;
|
||||
sub.f32 %f149, %f145, %f148;
|
||||
add.f32 %f150, %f141, %f149;
|
||||
add.f32 %f151, %f143, %f150;
|
||||
add.f32 %f152, %f147, %f151;
|
||||
add.f32 %f153, %f148, %f152;
|
||||
sub.f32 %f154, %f148, %f153;
|
||||
add.f32 %f155, %f152, %f154;
|
||||
mov.f32 %f156, 0f400CCCCD;
|
||||
mul.rn.f32 %f157, %f156, %f153;
|
||||
neg.f32 %f158, %f157;
|
||||
fma.rn.f32 %f159, %f156, %f153, %f158;
|
||||
fma.rn.f32 %f160, %f156, %f155, %f159;
|
||||
mov.f32 %f161, 0f00000000;
|
||||
fma.rn.f32 %f162, %f161, %f153, %f160;
|
||||
add.rn.f32 %f163, %f157, %f162;
|
||||
neg.f32 %f164, %f163;
|
||||
add.rn.f32 %f165, %f157, %f164;
|
||||
add.rn.f32 %f166, %f165, %f162;
|
||||
mov.b32 %r13, %f163;
|
||||
setp.eq.s32 %p11, %r13, 1118925336;
|
||||
add.s32 %r14, %r13, -1;
|
||||
mov.b32 %f167, %r14;
|
||||
add.f32 %f168, %f166, 0f37000000;
|
||||
selp.f32 %f169, %f167, %f163, %p11;
|
||||
selp.f32 %f10, %f168, %f166, %p11;
|
||||
mul.f32 %f170, %f169, 0f3FB8AA3B;
|
||||
cvt.rzi.f32.f32 %f171, %f170;
|
||||
mov.f32 %f172, 0fBF317200;
|
||||
fma.rn.f32 %f173, %f171, %f172, %f169;
|
||||
mov.f32 %f174, 0fB5BFBE8E;
|
||||
fma.rn.f32 %f175, %f171, %f174, %f173;
|
||||
mul.f32 %f176, %f175, 0f3FB8AA3B;
|
||||
ex2.approx.ftz.f32 %f177, %f176;
|
||||
add.f32 %f178, %f171, 0f00000000;
|
||||
ex2.approx.f32 %f179, %f178;
|
||||
mul.f32 %f180, %f177, %f179;
|
||||
setp.lt.f32 %p12, %f169, 0fC2D20000;
|
||||
selp.f32 %f181, 0f00000000, %f180, %p12;
|
||||
setp.gt.f32 %p13, %f169, 0f42D20000;
|
||||
selp.f32 %f653, 0f7F800000, %f181, %p13;
|
||||
setp.eq.f32 %p14, %f653, 0f7F800000;
|
||||
@%p14 bra BB0_3;
|
||||
|
||||
fma.rn.f32 %f653, %f653, %f10, %f653;
|
||||
|
||||
BB0_3:
|
||||
setp.lt.f32 %p15, %f8, 0f00000000;
|
||||
setp.eq.f32 %p16, %f4, 0f3F800000;
|
||||
and.pred %p1, %p15, %p16;
|
||||
mov.b32 %r15, %f653;
|
||||
xor.b32 %r16, %r15, -2147483648;
|
||||
mov.b32 %f182, %r16;
|
||||
selp.f32 %f655, %f182, %f653, %p1;
|
||||
setp.eq.f32 %p17, %f8, 0f00000000;
|
||||
@%p17 bra BB0_6;
|
||||
bra.uni BB0_4;
|
||||
|
||||
BB0_6:
|
||||
add.f32 %f185, %f8, %f8;
|
||||
selp.f32 %f655, %f185, 0f00000000, %p16;
|
||||
bra.uni BB0_7;
|
||||
|
||||
BB0_35:
|
||||
abs.f32 %f52, %f2;
|
||||
setp.lt.f32 %p57, %f52, 0f00800000;
|
||||
mul.f32 %f359, %f52, 0f4B800000;
|
||||
selp.f32 %f360, 0fC3170000, 0fC2FE0000, %p57;
|
||||
selp.f32 %f361, %f359, %f52, %p57;
|
||||
mov.b32 %r36, %f361;
|
||||
and.b32 %r37, %r36, 8388607;
|
||||
or.b32 %r38, %r37, 1065353216;
|
||||
mov.b32 %f362, %r38;
|
||||
shr.u32 %r39, %r36, 23;
|
||||
cvt.rn.f32.u32 %f363, %r39;
|
||||
add.f32 %f364, %f360, %f363;
|
||||
setp.gt.f32 %p58, %f362, 0f3FB504F3;
|
||||
mul.f32 %f365, %f362, 0f3F000000;
|
||||
add.f32 %f366, %f364, 0f3F800000;
|
||||
selp.f32 %f367, %f365, %f362, %p58;
|
||||
selp.f32 %f368, %f366, %f364, %p58;
|
||||
add.f32 %f369, %f367, 0fBF800000;
|
||||
add.f32 %f358, %f367, 0f3F800000;
|
||||
// inline asm
|
||||
rcp.approx.ftz.f32 %f357,%f358;
|
||||
// inline asm
|
||||
add.f32 %f370, %f369, %f369;
|
||||
mul.f32 %f371, %f357, %f370;
|
||||
mul.f32 %f372, %f371, %f371;
|
||||
mov.f32 %f373, 0f3C4CAF63;
|
||||
mov.f32 %f374, 0f3B18F0FE;
|
||||
fma.rn.f32 %f375, %f374, %f372, %f373;
|
||||
mov.f32 %f376, 0f3DAAAABD;
|
||||
fma.rn.f32 %f377, %f375, %f372, %f376;
|
||||
mul.rn.f32 %f378, %f377, %f372;
|
||||
mul.rn.f32 %f379, %f378, %f371;
|
||||
sub.f32 %f380, %f369, %f371;
|
||||
neg.f32 %f381, %f371;
|
||||
add.f32 %f382, %f380, %f380;
|
||||
fma.rn.f32 %f383, %f381, %f369, %f382;
|
||||
mul.rn.f32 %f384, %f357, %f383;
|
||||
add.f32 %f385, %f379, %f371;
|
||||
sub.f32 %f386, %f371, %f385;
|
||||
add.f32 %f387, %f379, %f386;
|
||||
add.f32 %f388, %f384, %f387;
|
||||
add.f32 %f389, %f385, %f388;
|
||||
sub.f32 %f390, %f385, %f389;
|
||||
add.f32 %f391, %f388, %f390;
|
||||
mov.f32 %f392, 0f3F317200;
|
||||
mul.rn.f32 %f393, %f368, %f392;
|
||||
mov.f32 %f394, 0f35BFBE8E;
|
||||
mul.rn.f32 %f395, %f368, %f394;
|
||||
add.f32 %f396, %f393, %f389;
|
||||
sub.f32 %f397, %f393, %f396;
|
||||
add.f32 %f398, %f389, %f397;
|
||||
add.f32 %f399, %f391, %f398;
|
||||
add.f32 %f400, %f395, %f399;
|
||||
add.f32 %f401, %f396, %f400;
|
||||
sub.f32 %f402, %f396, %f401;
|
||||
add.f32 %f403, %f400, %f402;
|
||||
mov.f32 %f404, 0f400CCCCD;
|
||||
mul.rn.f32 %f405, %f404, %f401;
|
||||
neg.f32 %f406, %f405;
|
||||
fma.rn.f32 %f407, %f404, %f401, %f406;
|
||||
fma.rn.f32 %f408, %f404, %f403, %f407;
|
||||
mov.f32 %f409, 0f00000000;
|
||||
fma.rn.f32 %f410, %f409, %f401, %f408;
|
||||
add.rn.f32 %f411, %f405, %f410;
|
||||
neg.f32 %f412, %f411;
|
||||
add.rn.f32 %f413, %f405, %f412;
|
||||
add.rn.f32 %f414, %f413, %f410;
|
||||
mov.b32 %r40, %f411;
|
||||
setp.eq.s32 %p59, %r40, 1118925336;
|
||||
add.s32 %r41, %r40, -1;
|
||||
mov.b32 %f415, %r41;
|
||||
add.f32 %f416, %f414, 0f37000000;
|
||||
selp.f32 %f417, %f415, %f411, %p59;
|
||||
selp.f32 %f53, %f416, %f414, %p59;
|
||||
mul.f32 %f418, %f417, 0f3FB8AA3B;
|
||||
cvt.rzi.f32.f32 %f419, %f418;
|
||||
mov.f32 %f420, 0fBF317200;
|
||||
fma.rn.f32 %f421, %f419, %f420, %f417;
|
||||
mov.f32 %f422, 0fB5BFBE8E;
|
||||
fma.rn.f32 %f423, %f419, %f422, %f421;
|
||||
mul.f32 %f424, %f423, 0f3FB8AA3B;
|
||||
ex2.approx.ftz.f32 %f425, %f424;
|
||||
add.f32 %f426, %f419, 0f00000000;
|
||||
ex2.approx.f32 %f427, %f426;
|
||||
mul.f32 %f428, %f425, %f427;
|
||||
setp.lt.f32 %p60, %f417, 0fC2D20000;
|
||||
selp.f32 %f429, 0f00000000, %f428, %p60;
|
||||
setp.gt.f32 %p61, %f417, 0f42D20000;
|
||||
selp.f32 %f662, 0f7F800000, %f429, %p61;
|
||||
setp.eq.f32 %p62, %f662, 0f7F800000;
|
||||
@%p62 bra BB0_37;
|
||||
|
||||
fma.rn.f32 %f662, %f662, %f53, %f662;
|
||||
|
||||
BB0_37:
|
||||
setp.lt.f32 %p63, %f2, 0f00000000;
|
||||
setp.eq.f32 %p64, %f4, 0f3F800000;
|
||||
and.pred %p4, %p63, %p64;
|
||||
mov.b32 %r42, %f662;
|
||||
xor.b32 %r43, %r42, -2147483648;
|
||||
mov.b32 %f430, %r43;
|
||||
selp.f32 %f664, %f430, %f662, %p4;
|
||||
setp.eq.f32 %p65, %f2, 0f00000000;
|
||||
@%p65 bra BB0_40;
|
||||
bra.uni BB0_38;
|
||||
|
||||
BB0_40:
|
||||
add.f32 %f433, %f2, %f2;
|
||||
selp.f32 %f664, %f433, 0f00000000, %p64;
|
||||
bra.uni BB0_41;
|
||||
|
||||
BB0_4:
|
||||
setp.geu.f32 %p18, %f8, 0f00000000;
|
||||
@%p18 bra BB0_7;
|
||||
|
||||
mov.f32 %f630, 0f400CCCCD;
|
||||
cvt.rzi.f32.f32 %f184, %f630;
|
||||
setp.neu.f32 %p19, %f184, 0f400CCCCD;
|
||||
selp.f32 %f655, 0f7FFFFFFF, %f655, %p19;
|
||||
|
||||
BB0_7:
|
||||
abs.f32 %f604, %f8;
|
||||
add.f32 %f186, %f604, 0f400CCCCD;
|
||||
mov.b32 %r17, %f186;
|
||||
setp.lt.s32 %p21, %r17, 2139095040;
|
||||
@%p21 bra BB0_12;
|
||||
|
||||
abs.f32 %f628, %f8;
|
||||
setp.gtu.f32 %p22, %f628, 0f7F800000;
|
||||
@%p22 bra BB0_11;
|
||||
bra.uni BB0_9;
|
||||
|
||||
BB0_11:
|
||||
add.f32 %f655, %f8, 0f400CCCCD;
|
||||
bra.uni BB0_12;
|
||||
|
||||
BB0_9:
|
||||
abs.f32 %f629, %f8;
|
||||
setp.neu.f32 %p23, %f629, 0f7F800000;
|
||||
@%p23 bra BB0_12;
|
||||
|
||||
selp.f32 %f655, 0fFF800000, 0f7F800000, %p1;
|
||||
|
||||
BB0_12:
|
||||
mov.f32 %f614, 0fB5BFBE8E;
|
||||
mov.f32 %f613, 0fBF317200;
|
||||
mov.f32 %f612, 0f00000000;
|
||||
mov.f32 %f611, 0f35BFBE8E;
|
||||
mov.f32 %f610, 0f3F317200;
|
||||
mov.f32 %f609, 0f3DAAAABD;
|
||||
mov.f32 %f608, 0f3C4CAF63;
|
||||
mov.f32 %f607, 0f3B18F0FE;
|
||||
mov.f32 %f606, 0f3F800000;
|
||||
mov.f32 %f605, 0f400CCCCD;
|
||||
setp.eq.f32 %p24, %f8, 0f3F800000;
|
||||
selp.f32 %f189, 0f3F800000, %f655, %p24;
|
||||
cvt.sat.f32.f32 %f21, %f189;
|
||||
sub.f32 %f22, %f606, %f97;
|
||||
abs.f32 %f23, %f22;
|
||||
setp.lt.f32 %p25, %f23, 0f00800000;
|
||||
mul.f32 %f191, %f23, 0f4B800000;
|
||||
selp.f32 %f192, 0fC3170000, 0fC2FE0000, %p25;
|
||||
selp.f32 %f193, %f191, %f23, %p25;
|
||||
mov.b32 %r18, %f193;
|
||||
and.b32 %r19, %r18, 8388607;
|
||||
or.b32 %r20, %r19, 1065353216;
|
||||
mov.b32 %f194, %r20;
|
||||
shr.u32 %r21, %r18, 23;
|
||||
cvt.rn.f32.u32 %f195, %r21;
|
||||
add.f32 %f196, %f192, %f195;
|
||||
setp.gt.f32 %p26, %f194, 0f3FB504F3;
|
||||
mul.f32 %f197, %f194, 0f3F000000;
|
||||
add.f32 %f198, %f196, 0f3F800000;
|
||||
selp.f32 %f199, %f197, %f194, %p26;
|
||||
selp.f32 %f200, %f198, %f196, %p26;
|
||||
add.f32 %f201, %f199, 0fBF800000;
|
||||
add.f32 %f188, %f199, 0f3F800000;
|
||||
// inline asm
|
||||
rcp.approx.ftz.f32 %f187,%f188;
|
||||
// inline asm
|
||||
add.f32 %f202, %f201, %f201;
|
||||
mul.f32 %f203, %f187, %f202;
|
||||
mul.f32 %f204, %f203, %f203;
|
||||
fma.rn.f32 %f207, %f607, %f204, %f608;
|
||||
fma.rn.f32 %f209, %f207, %f204, %f609;
|
||||
mul.rn.f32 %f210, %f209, %f204;
|
||||
mul.rn.f32 %f211, %f210, %f203;
|
||||
sub.f32 %f212, %f201, %f203;
|
||||
neg.f32 %f213, %f203;
|
||||
add.f32 %f214, %f212, %f212;
|
||||
fma.rn.f32 %f215, %f213, %f201, %f214;
|
||||
mul.rn.f32 %f216, %f187, %f215;
|
||||
add.f32 %f217, %f211, %f203;
|
||||
sub.f32 %f218, %f203, %f217;
|
||||
add.f32 %f219, %f211, %f218;
|
||||
add.f32 %f220, %f216, %f219;
|
||||
add.f32 %f221, %f217, %f220;
|
||||
sub.f32 %f222, %f217, %f221;
|
||||
add.f32 %f223, %f220, %f222;
|
||||
mul.rn.f32 %f225, %f200, %f610;
|
||||
mul.rn.f32 %f227, %f200, %f611;
|
||||
add.f32 %f228, %f225, %f221;
|
||||
sub.f32 %f229, %f225, %f228;
|
||||
add.f32 %f230, %f221, %f229;
|
||||
add.f32 %f231, %f223, %f230;
|
||||
add.f32 %f232, %f227, %f231;
|
||||
add.f32 %f233, %f228, %f232;
|
||||
sub.f32 %f234, %f228, %f233;
|
||||
add.f32 %f235, %f232, %f234;
|
||||
mul.rn.f32 %f237, %f605, %f233;
|
||||
neg.f32 %f238, %f237;
|
||||
fma.rn.f32 %f239, %f605, %f233, %f238;
|
||||
fma.rn.f32 %f240, %f605, %f235, %f239;
|
||||
fma.rn.f32 %f242, %f612, %f233, %f240;
|
||||
add.rn.f32 %f243, %f237, %f242;
|
||||
neg.f32 %f244, %f243;
|
||||
add.rn.f32 %f245, %f237, %f244;
|
||||
add.rn.f32 %f246, %f245, %f242;
|
||||
mov.b32 %r22, %f243;
|
||||
setp.eq.s32 %p27, %r22, 1118925336;
|
||||
add.s32 %r23, %r22, -1;
|
||||
mov.b32 %f247, %r23;
|
||||
add.f32 %f248, %f246, 0f37000000;
|
||||
selp.f32 %f249, %f247, %f243, %p27;
|
||||
selp.f32 %f24, %f248, %f246, %p27;
|
||||
mul.f32 %f250, %f249, 0f3FB8AA3B;
|
||||
cvt.rzi.f32.f32 %f251, %f250;
|
||||
fma.rn.f32 %f253, %f251, %f613, %f249;
|
||||
fma.rn.f32 %f255, %f251, %f614, %f253;
|
||||
mul.f32 %f256, %f255, 0f3FB8AA3B;
|
||||
ex2.approx.ftz.f32 %f257, %f256;
|
||||
add.f32 %f258, %f251, 0f00000000;
|
||||
ex2.approx.f32 %f259, %f258;
|
||||
mul.f32 %f260, %f257, %f259;
|
||||
setp.lt.f32 %p28, %f249, 0fC2D20000;
|
||||
selp.f32 %f261, 0f00000000, %f260, %p28;
|
||||
setp.gt.f32 %p29, %f249, 0f42D20000;
|
||||
selp.f32 %f656, 0f7F800000, %f261, %p29;
|
||||
setp.eq.f32 %p30, %f656, 0f7F800000;
|
||||
@%p30 bra BB0_14;
|
||||
|
||||
fma.rn.f32 %f656, %f656, %f24, %f656;
|
||||
|
||||
BB0_14:
|
||||
setp.lt.f32 %p31, %f22, 0f00000000;
|
||||
and.pred %p2, %p31, %p16;
|
||||
mov.b32 %r24, %f656;
|
||||
xor.b32 %r25, %r24, -2147483648;
|
||||
mov.b32 %f262, %r25;
|
||||
selp.f32 %f658, %f262, %f656, %p2;
|
||||
setp.eq.f32 %p33, %f22, 0f00000000;
|
||||
@%p33 bra BB0_17;
|
||||
bra.uni BB0_15;
|
||||
|
||||
BB0_17:
|
||||
add.f32 %f265, %f22, %f22;
|
||||
selp.f32 %f658, %f265, 0f00000000, %p16;
|
||||
bra.uni BB0_18;
|
||||
|
||||
BB0_15:
|
||||
setp.geu.f32 %p34, %f22, 0f00000000;
|
||||
@%p34 bra BB0_18;
|
||||
|
||||
mov.f32 %f627, 0f400CCCCD;
|
||||
cvt.rzi.f32.f32 %f264, %f627;
|
||||
setp.neu.f32 %p35, %f264, 0f400CCCCD;
|
||||
selp.f32 %f658, 0f7FFFFFFF, %f658, %p35;
|
||||
|
||||
BB0_18:
|
||||
add.f32 %f266, %f23, 0f400CCCCD;
|
||||
mov.b32 %r26, %f266;
|
||||
setp.lt.s32 %p37, %r26, 2139095040;
|
||||
@%p37 bra BB0_23;
|
||||
|
||||
setp.gtu.f32 %p38, %f23, 0f7F800000;
|
||||
@%p38 bra BB0_22;
|
||||
bra.uni BB0_20;
|
||||
|
||||
BB0_22:
|
||||
add.f32 %f658, %f22, 0f400CCCCD;
|
||||
bra.uni BB0_23;
|
||||
|
||||
BB0_20:
|
||||
setp.neu.f32 %p39, %f23, 0f7F800000;
|
||||
@%p39 bra BB0_23;
|
||||
|
||||
selp.f32 %f658, 0fFF800000, 0f7F800000, %p2;
|
||||
|
||||
BB0_23:
|
||||
mov.f32 %f624, 0fB5BFBE8E;
|
||||
mov.f32 %f623, 0fBF317200;
|
||||
mov.f32 %f622, 0f00000000;
|
||||
mov.f32 %f621, 0f35BFBE8E;
|
||||
mov.f32 %f620, 0f3F317200;
|
||||
mov.f32 %f619, 0f3DAAAABD;
|
||||
mov.f32 %f618, 0f3C4CAF63;
|
||||
mov.f32 %f617, 0f3B18F0FE;
|
||||
mov.f32 %f616, 0f3F800000;
|
||||
mov.f32 %f615, 0f400CCCCD;
|
||||
setp.eq.f32 %p40, %f22, 0f3F800000;
|
||||
selp.f32 %f269, 0f3F800000, %f658, %p40;
|
||||
cvt.sat.f32.f32 %f35, %f269;
|
||||
sub.f32 %f36, %f616, %f3;
|
||||
abs.f32 %f37, %f36;
|
||||
setp.lt.f32 %p41, %f37, 0f00800000;
|
||||
mul.f32 %f271, %f37, 0f4B800000;
|
||||
selp.f32 %f272, 0fC3170000, 0fC2FE0000, %p41;
|
||||
selp.f32 %f273, %f271, %f37, %p41;
|
||||
mov.b32 %r27, %f273;
|
||||
and.b32 %r28, %r27, 8388607;
|
||||
or.b32 %r29, %r28, 1065353216;
|
||||
mov.b32 %f274, %r29;
|
||||
shr.u32 %r30, %r27, 23;
|
||||
cvt.rn.f32.u32 %f275, %r30;
|
||||
add.f32 %f276, %f272, %f275;
|
||||
setp.gt.f32 %p42, %f274, 0f3FB504F3;
|
||||
mul.f32 %f277, %f274, 0f3F000000;
|
||||
add.f32 %f278, %f276, 0f3F800000;
|
||||
selp.f32 %f279, %f277, %f274, %p42;
|
||||
selp.f32 %f280, %f278, %f276, %p42;
|
||||
add.f32 %f281, %f279, 0fBF800000;
|
||||
add.f32 %f268, %f279, 0f3F800000;
|
||||
// inline asm
|
||||
rcp.approx.ftz.f32 %f267,%f268;
|
||||
// inline asm
|
||||
add.f32 %f282, %f281, %f281;
|
||||
mul.f32 %f283, %f267, %f282;
|
||||
mul.f32 %f284, %f283, %f283;
|
||||
fma.rn.f32 %f287, %f617, %f284, %f618;
|
||||
fma.rn.f32 %f289, %f287, %f284, %f619;
|
||||
mul.rn.f32 %f290, %f289, %f284;
|
||||
mul.rn.f32 %f291, %f290, %f283;
|
||||
sub.f32 %f292, %f281, %f283;
|
||||
neg.f32 %f293, %f283;
|
||||
add.f32 %f294, %f292, %f292;
|
||||
fma.rn.f32 %f295, %f293, %f281, %f294;
|
||||
mul.rn.f32 %f296, %f267, %f295;
|
||||
add.f32 %f297, %f291, %f283;
|
||||
sub.f32 %f298, %f283, %f297;
|
||||
add.f32 %f299, %f291, %f298;
|
||||
add.f32 %f300, %f296, %f299;
|
||||
add.f32 %f301, %f297, %f300;
|
||||
sub.f32 %f302, %f297, %f301;
|
||||
add.f32 %f303, %f300, %f302;
|
||||
mul.rn.f32 %f305, %f280, %f620;
|
||||
mul.rn.f32 %f307, %f280, %f621;
|
||||
add.f32 %f308, %f305, %f301;
|
||||
sub.f32 %f309, %f305, %f308;
|
||||
add.f32 %f310, %f301, %f309;
|
||||
add.f32 %f311, %f303, %f310;
|
||||
add.f32 %f312, %f307, %f311;
|
||||
add.f32 %f313, %f308, %f312;
|
||||
sub.f32 %f314, %f308, %f313;
|
||||
add.f32 %f315, %f312, %f314;
|
||||
mul.rn.f32 %f317, %f615, %f313;
|
||||
neg.f32 %f318, %f317;
|
||||
fma.rn.f32 %f319, %f615, %f313, %f318;
|
||||
fma.rn.f32 %f320, %f615, %f315, %f319;
|
||||
fma.rn.f32 %f322, %f622, %f313, %f320;
|
||||
add.rn.f32 %f323, %f317, %f322;
|
||||
neg.f32 %f324, %f323;
|
||||
add.rn.f32 %f325, %f317, %f324;
|
||||
add.rn.f32 %f326, %f325, %f322;
|
||||
mov.b32 %r31, %f323;
|
||||
setp.eq.s32 %p43, %r31, 1118925336;
|
||||
add.s32 %r32, %r31, -1;
|
||||
mov.b32 %f327, %r32;
|
||||
add.f32 %f328, %f326, 0f37000000;
|
||||
selp.f32 %f329, %f327, %f323, %p43;
|
||||
selp.f32 %f38, %f328, %f326, %p43;
|
||||
mul.f32 %f330, %f329, 0f3FB8AA3B;
|
||||
cvt.rzi.f32.f32 %f331, %f330;
|
||||
fma.rn.f32 %f333, %f331, %f623, %f329;
|
||||
fma.rn.f32 %f335, %f331, %f624, %f333;
|
||||
mul.f32 %f336, %f335, 0f3FB8AA3B;
|
||||
ex2.approx.ftz.f32 %f337, %f336;
|
||||
add.f32 %f338, %f331, 0f00000000;
|
||||
ex2.approx.f32 %f339, %f338;
|
||||
mul.f32 %f340, %f337, %f339;
|
||||
setp.lt.f32 %p44, %f329, 0fC2D20000;
|
||||
selp.f32 %f341, 0f00000000, %f340, %p44;
|
||||
setp.gt.f32 %p45, %f329, 0f42D20000;
|
||||
selp.f32 %f659, 0f7F800000, %f341, %p45;
|
||||
setp.eq.f32 %p46, %f659, 0f7F800000;
|
||||
@%p46 bra BB0_25;
|
||||
|
||||
fma.rn.f32 %f659, %f659, %f38, %f659;
|
||||
|
||||
BB0_25:
|
||||
setp.lt.f32 %p47, %f36, 0f00000000;
|
||||
and.pred %p3, %p47, %p16;
|
||||
mov.b32 %r33, %f659;
|
||||
xor.b32 %r34, %r33, -2147483648;
|
||||
mov.b32 %f342, %r34;
|
||||
selp.f32 %f661, %f342, %f659, %p3;
|
||||
setp.eq.f32 %p49, %f36, 0f00000000;
|
||||
@%p49 bra BB0_28;
|
||||
bra.uni BB0_26;
|
||||
|
||||
BB0_28:
|
||||
add.f32 %f345, %f36, %f36;
|
||||
selp.f32 %f661, %f345, 0f00000000, %p16;
|
||||
bra.uni BB0_29;
|
||||
|
||||
BB0_26:
|
||||
setp.geu.f32 %p50, %f36, 0f00000000;
|
||||
@%p50 bra BB0_29;
|
||||
|
||||
mov.f32 %f626, 0f400CCCCD;
|
||||
cvt.rzi.f32.f32 %f344, %f626;
|
||||
setp.neu.f32 %p51, %f344, 0f400CCCCD;
|
||||
selp.f32 %f661, 0f7FFFFFFF, %f661, %p51;
|
||||
|
||||
BB0_29:
|
||||
abs.f32 %f631, %f36;
|
||||
add.f32 %f346, %f631, 0f400CCCCD;
|
||||
mov.b32 %r35, %f346;
|
||||
setp.lt.s32 %p53, %r35, 2139095040;
|
||||
@%p53 bra BB0_34;
|
||||
|
||||
abs.f32 %f632, %f36;
|
||||
setp.gtu.f32 %p54, %f632, 0f7F800000;
|
||||
@%p54 bra BB0_33;
|
||||
bra.uni BB0_31;
|
||||
|
||||
BB0_33:
|
||||
add.f32 %f661, %f36, 0f400CCCCD;
|
||||
bra.uni BB0_34;
|
||||
|
||||
BB0_31:
|
||||
abs.f32 %f633, %f36;
|
||||
setp.neu.f32 %p55, %f633, 0f7F800000;
|
||||
@%p55 bra BB0_34;
|
||||
|
||||
selp.f32 %f661, 0fFF800000, 0f7F800000, %p3;
|
||||
|
||||
BB0_34:
|
||||
ld.const.u64 %rd16, [cs];
|
||||
mov.f32 %f625, 0f3F800000;
|
||||
setp.eq.f32 %p56, %f36, 0f3F800000;
|
||||
selp.f32 %f347, 0f3F800000, %f661, %p56;
|
||||
cvt.sat.f32.f32 %f348, %f347;
|
||||
max.f32 %f349, %f21, %f35;
|
||||
max.f32 %f350, %f349, %f348;
|
||||
sub.f32 %f352, %f625, %f350;
|
||||
rcp.rn.f32 %f353, %f352;
|
||||
mul.f32 %f354, %f21, %f353;
|
||||
mul.f32 %f355, %f35, %f353;
|
||||
mul.f32 %f356, %f348, %f353;
|
||||
min.f32 %f671, %f354, %f105;
|
||||
min.f32 %f672, %f355, %f106;
|
||||
min.f32 %f673, %f356, %f107;
|
||||
bra.uni BB0_69;
|
||||
|
||||
BB0_38:
|
||||
setp.geu.f32 %p66, %f2, 0f00000000;
|
||||
@%p66 bra BB0_41;
|
||||
|
||||
cvt.rzi.f32.f32 %f432, %f404;
|
||||
setp.neu.f32 %p67, %f432, 0f400CCCCD;
|
||||
selp.f32 %f664, 0f7FFFFFFF, %f664, %p67;
|
||||
|
||||
BB0_41:
|
||||
abs.f32 %f634, %f2;
|
||||
add.f32 %f434, %f634, 0f400CCCCD;
|
||||
mov.b32 %r44, %f434;
|
||||
setp.lt.s32 %p69, %r44, 2139095040;
|
||||
@%p69 bra BB0_46;
|
||||
|
||||
abs.f32 %f651, %f2;
|
||||
setp.gtu.f32 %p70, %f651, 0f7F800000;
|
||||
@%p70 bra BB0_45;
|
||||
bra.uni BB0_43;
|
||||
|
||||
BB0_45:
|
||||
add.f32 %f664, %f2, 0f400CCCCD;
|
||||
bra.uni BB0_46;
|
||||
|
||||
BB0_43:
|
||||
abs.f32 %f652, %f2;
|
||||
setp.neu.f32 %p71, %f652, 0f7F800000;
|
||||
@%p71 bra BB0_46;
|
||||
|
||||
selp.f32 %f664, 0fFF800000, 0f7F800000, %p4;
|
||||
|
||||
BB0_46:
|
||||
mov.f32 %f642, 0fB5BFBE8E;
|
||||
mov.f32 %f641, 0fBF317200;
|
||||
mov.f32 %f640, 0f00000000;
|
||||
mov.f32 %f639, 0f35BFBE8E;
|
||||
mov.f32 %f638, 0f3F317200;
|
||||
mov.f32 %f637, 0f3DAAAABD;
|
||||
mov.f32 %f636, 0f3C4CAF63;
|
||||
mov.f32 %f635, 0f3B18F0FE;
|
||||
setp.eq.f32 %p72, %f2, 0f3F800000;
|
||||
selp.f32 %f437, 0f3F800000, %f664, %p72;
|
||||
cvt.sat.f32.f32 %f64, %f437;
|
||||
abs.f32 %f65, %f97;
|
||||
setp.lt.f32 %p73, %f65, 0f00800000;
|
||||
mul.f32 %f438, %f65, 0f4B800000;
|
||||
selp.f32 %f439, 0fC3170000, 0fC2FE0000, %p73;
|
||||
selp.f32 %f440, %f438, %f65, %p73;
|
||||
mov.b32 %r45, %f440;
|
||||
and.b32 %r46, %r45, 8388607;
|
||||
or.b32 %r47, %r46, 1065353216;
|
||||
mov.b32 %f441, %r47;
|
||||
shr.u32 %r48, %r45, 23;
|
||||
cvt.rn.f32.u32 %f442, %r48;
|
||||
add.f32 %f443, %f439, %f442;
|
||||
setp.gt.f32 %p74, %f441, 0f3FB504F3;
|
||||
mul.f32 %f444, %f441, 0f3F000000;
|
||||
add.f32 %f445, %f443, 0f3F800000;
|
||||
selp.f32 %f446, %f444, %f441, %p74;
|
||||
selp.f32 %f447, %f445, %f443, %p74;
|
||||
add.f32 %f448, %f446, 0fBF800000;
|
||||
add.f32 %f436, %f446, 0f3F800000;
|
||||
// inline asm
|
||||
rcp.approx.ftz.f32 %f435,%f436;
|
||||
// inline asm
|
||||
add.f32 %f449, %f448, %f448;
|
||||
mul.f32 %f450, %f435, %f449;
|
||||
mul.f32 %f451, %f450, %f450;
|
||||
fma.rn.f32 %f454, %f635, %f451, %f636;
|
||||
fma.rn.f32 %f456, %f454, %f451, %f637;
|
||||
mul.rn.f32 %f457, %f456, %f451;
|
||||
mul.rn.f32 %f458, %f457, %f450;
|
||||
sub.f32 %f459, %f448, %f450;
|
||||
neg.f32 %f460, %f450;
|
||||
add.f32 %f461, %f459, %f459;
|
||||
fma.rn.f32 %f462, %f460, %f448, %f461;
|
||||
mul.rn.f32 %f463, %f435, %f462;
|
||||
add.f32 %f464, %f458, %f450;
|
||||
sub.f32 %f465, %f450, %f464;
|
||||
add.f32 %f466, %f458, %f465;
|
||||
add.f32 %f467, %f463, %f466;
|
||||
add.f32 %f468, %f464, %f467;
|
||||
sub.f32 %f469, %f464, %f468;
|
||||
add.f32 %f470, %f467, %f469;
|
||||
mul.rn.f32 %f472, %f447, %f638;
|
||||
mul.rn.f32 %f474, %f447, %f639;
|
||||
add.f32 %f475, %f472, %f468;
|
||||
sub.f32 %f476, %f472, %f475;
|
||||
add.f32 %f477, %f468, %f476;
|
||||
add.f32 %f478, %f470, %f477;
|
||||
add.f32 %f479, %f474, %f478;
|
||||
add.f32 %f480, %f475, %f479;
|
||||
sub.f32 %f481, %f475, %f480;
|
||||
add.f32 %f482, %f479, %f481;
|
||||
mul.rn.f32 %f484, %f404, %f480;
|
||||
neg.f32 %f485, %f484;
|
||||
fma.rn.f32 %f486, %f404, %f480, %f485;
|
||||
fma.rn.f32 %f487, %f404, %f482, %f486;
|
||||
fma.rn.f32 %f489, %f640, %f480, %f487;
|
||||
add.rn.f32 %f490, %f484, %f489;
|
||||
neg.f32 %f491, %f490;
|
||||
add.rn.f32 %f492, %f484, %f491;
|
||||
add.rn.f32 %f493, %f492, %f489;
|
||||
mov.b32 %r49, %f490;
|
||||
setp.eq.s32 %p75, %r49, 1118925336;
|
||||
add.s32 %r50, %r49, -1;
|
||||
mov.b32 %f494, %r50;
|
||||
add.f32 %f495, %f493, 0f37000000;
|
||||
selp.f32 %f496, %f494, %f490, %p75;
|
||||
selp.f32 %f66, %f495, %f493, %p75;
|
||||
mul.f32 %f497, %f496, 0f3FB8AA3B;
|
||||
cvt.rzi.f32.f32 %f498, %f497;
|
||||
fma.rn.f32 %f500, %f498, %f641, %f496;
|
||||
fma.rn.f32 %f502, %f498, %f642, %f500;
|
||||
mul.f32 %f503, %f502, 0f3FB8AA3B;
|
||||
ex2.approx.ftz.f32 %f504, %f503;
|
||||
add.f32 %f505, %f498, 0f00000000;
|
||||
ex2.approx.f32 %f506, %f505;
|
||||
mul.f32 %f507, %f504, %f506;
|
||||
setp.lt.f32 %p76, %f496, 0fC2D20000;
|
||||
selp.f32 %f508, 0f00000000, %f507, %p76;
|
||||
setp.gt.f32 %p77, %f496, 0f42D20000;
|
||||
selp.f32 %f665, 0f7F800000, %f508, %p77;
|
||||
setp.eq.f32 %p78, %f665, 0f7F800000;
|
||||
@%p78 bra BB0_48;
|
||||
|
||||
fma.rn.f32 %f665, %f665, %f66, %f665;
|
||||
|
||||
BB0_48:
|
||||
setp.lt.f32 %p79, %f97, 0f00000000;
|
||||
and.pred %p5, %p79, %p64;
|
||||
mov.b32 %r51, %f665;
|
||||
xor.b32 %r52, %r51, -2147483648;
|
||||
mov.b32 %f509, %r52;
|
||||
selp.f32 %f667, %f509, %f665, %p5;
|
||||
setp.eq.f32 %p81, %f97, 0f00000000;
|
||||
@%p81 bra BB0_51;
|
||||
bra.uni BB0_49;
|
||||
|
||||
BB0_51:
|
||||
add.f32 %f512, %f97, %f97;
|
||||
selp.f32 %f667, %f512, 0f00000000, %p64;
|
||||
bra.uni BB0_52;
|
||||
|
||||
BB0_49:
|
||||
setp.geu.f32 %p82, %f97, 0f00000000;
|
||||
@%p82 bra BB0_52;
|
||||
|
||||
cvt.rzi.f32.f32 %f511, %f404;
|
||||
setp.neu.f32 %p83, %f511, 0f400CCCCD;
|
||||
selp.f32 %f667, 0f7FFFFFFF, %f667, %p83;
|
||||
|
||||
BB0_52:
|
||||
add.f32 %f513, %f65, 0f400CCCCD;
|
||||
mov.b32 %r53, %f513;
|
||||
setp.lt.s32 %p85, %r53, 2139095040;
|
||||
@%p85 bra BB0_57;
|
||||
|
||||
setp.gtu.f32 %p86, %f65, 0f7F800000;
|
||||
@%p86 bra BB0_56;
|
||||
bra.uni BB0_54;
|
||||
|
||||
BB0_56:
|
||||
add.f32 %f667, %f97, 0f400CCCCD;
|
||||
bra.uni BB0_57;
|
||||
|
||||
BB0_54:
|
||||
setp.neu.f32 %p87, %f65, 0f7F800000;
|
||||
@%p87 bra BB0_57;
|
||||
|
||||
selp.f32 %f667, 0fFF800000, 0f7F800000, %p5;
|
||||
|
||||
BB0_57:
|
||||
mov.f32 %f650, 0fB5BFBE8E;
|
||||
mov.f32 %f649, 0fBF317200;
|
||||
mov.f32 %f648, 0f00000000;
|
||||
mov.f32 %f647, 0f35BFBE8E;
|
||||
mov.f32 %f646, 0f3F317200;
|
||||
mov.f32 %f645, 0f3DAAAABD;
|
||||
mov.f32 %f644, 0f3C4CAF63;
|
||||
mov.f32 %f643, 0f3B18F0FE;
|
||||
setp.eq.f32 %p88, %f97, 0f3F800000;
|
||||
selp.f32 %f516, 0f3F800000, %f667, %p88;
|
||||
cvt.sat.f32.f32 %f77, %f516;
|
||||
abs.f32 %f78, %f3;
|
||||
setp.lt.f32 %p89, %f78, 0f00800000;
|
||||
mul.f32 %f517, %f78, 0f4B800000;
|
||||
selp.f32 %f518, 0fC3170000, 0fC2FE0000, %p89;
|
||||
selp.f32 %f519, %f517, %f78, %p89;
|
||||
mov.b32 %r54, %f519;
|
||||
and.b32 %r55, %r54, 8388607;
|
||||
or.b32 %r56, %r55, 1065353216;
|
||||
mov.b32 %f520, %r56;
|
||||
shr.u32 %r57, %r54, 23;
|
||||
cvt.rn.f32.u32 %f521, %r57;
|
||||
add.f32 %f522, %f518, %f521;
|
||||
setp.gt.f32 %p90, %f520, 0f3FB504F3;
|
||||
mul.f32 %f523, %f520, 0f3F000000;
|
||||
add.f32 %f524, %f522, 0f3F800000;
|
||||
selp.f32 %f525, %f523, %f520, %p90;
|
||||
selp.f32 %f526, %f524, %f522, %p90;
|
||||
add.f32 %f527, %f525, 0fBF800000;
|
||||
add.f32 %f515, %f525, 0f3F800000;
|
||||
// inline asm
|
||||
rcp.approx.ftz.f32 %f514,%f515;
|
||||
// inline asm
|
||||
add.f32 %f528, %f527, %f527;
|
||||
mul.f32 %f529, %f514, %f528;
|
||||
mul.f32 %f530, %f529, %f529;
|
||||
fma.rn.f32 %f533, %f643, %f530, %f644;
|
||||
fma.rn.f32 %f535, %f533, %f530, %f645;
|
||||
mul.rn.f32 %f536, %f535, %f530;
|
||||
mul.rn.f32 %f537, %f536, %f529;
|
||||
sub.f32 %f538, %f527, %f529;
|
||||
neg.f32 %f539, %f529;
|
||||
add.f32 %f540, %f538, %f538;
|
||||
fma.rn.f32 %f541, %f539, %f527, %f540;
|
||||
mul.rn.f32 %f542, %f514, %f541;
|
||||
add.f32 %f543, %f537, %f529;
|
||||
sub.f32 %f544, %f529, %f543;
|
||||
add.f32 %f545, %f537, %f544;
|
||||
add.f32 %f546, %f542, %f545;
|
||||
add.f32 %f547, %f543, %f546;
|
||||
sub.f32 %f548, %f543, %f547;
|
||||
add.f32 %f549, %f546, %f548;
|
||||
mul.rn.f32 %f551, %f526, %f646;
|
||||
mul.rn.f32 %f553, %f526, %f647;
|
||||
add.f32 %f554, %f551, %f547;
|
||||
sub.f32 %f555, %f551, %f554;
|
||||
add.f32 %f556, %f547, %f555;
|
||||
add.f32 %f557, %f549, %f556;
|
||||
add.f32 %f558, %f553, %f557;
|
||||
add.f32 %f559, %f554, %f558;
|
||||
sub.f32 %f560, %f554, %f559;
|
||||
add.f32 %f561, %f558, %f560;
|
||||
mul.rn.f32 %f563, %f404, %f559;
|
||||
neg.f32 %f564, %f563;
|
||||
fma.rn.f32 %f565, %f404, %f559, %f564;
|
||||
fma.rn.f32 %f566, %f404, %f561, %f565;
|
||||
fma.rn.f32 %f568, %f648, %f559, %f566;
|
||||
add.rn.f32 %f569, %f563, %f568;
|
||||
neg.f32 %f570, %f569;
|
||||
add.rn.f32 %f571, %f563, %f570;
|
||||
add.rn.f32 %f572, %f571, %f568;
|
||||
mov.b32 %r58, %f569;
|
||||
setp.eq.s32 %p91, %r58, 1118925336;
|
||||
add.s32 %r59, %r58, -1;
|
||||
mov.b32 %f573, %r59;
|
||||
add.f32 %f574, %f572, 0f37000000;
|
||||
selp.f32 %f575, %f573, %f569, %p91;
|
||||
selp.f32 %f79, %f574, %f572, %p91;
|
||||
mul.f32 %f576, %f575, 0f3FB8AA3B;
|
||||
cvt.rzi.f32.f32 %f577, %f576;
|
||||
fma.rn.f32 %f579, %f577, %f649, %f575;
|
||||
fma.rn.f32 %f581, %f577, %f650, %f579;
|
||||
mul.f32 %f582, %f581, 0f3FB8AA3B;
|
||||
ex2.approx.ftz.f32 %f583, %f582;
|
||||
add.f32 %f584, %f577, 0f00000000;
|
||||
ex2.approx.f32 %f585, %f584;
|
||||
mul.f32 %f586, %f583, %f585;
|
||||
setp.lt.f32 %p92, %f575, 0fC2D20000;
|
||||
selp.f32 %f587, 0f00000000, %f586, %p92;
|
||||
setp.gt.f32 %p93, %f575, 0f42D20000;
|
||||
selp.f32 %f668, 0f7F800000, %f587, %p93;
|
||||
setp.eq.f32 %p94, %f668, 0f7F800000;
|
||||
@%p94 bra BB0_59;
|
||||
|
||||
fma.rn.f32 %f668, %f668, %f79, %f668;
|
||||
|
||||
BB0_59:
|
||||
setp.lt.f32 %p95, %f3, 0f00000000;
|
||||
and.pred %p6, %p95, %p64;
|
||||
mov.b32 %r60, %f668;
|
||||
xor.b32 %r61, %r60, -2147483648;
|
||||
mov.b32 %f588, %r61;
|
||||
selp.f32 %f670, %f588, %f668, %p6;
|
||||
setp.eq.f32 %p97, %f3, 0f00000000;
|
||||
@%p97 bra BB0_62;
|
||||
bra.uni BB0_60;
|
||||
|
||||
BB0_62:
|
||||
add.f32 %f591, %f3, %f3;
|
||||
selp.f32 %f670, %f591, 0f00000000, %p64;
|
||||
bra.uni BB0_63;
|
||||
|
||||
BB0_60:
|
||||
setp.geu.f32 %p98, %f3, 0f00000000;
|
||||
@%p98 bra BB0_63;
|
||||
|
||||
cvt.rzi.f32.f32 %f590, %f404;
|
||||
setp.neu.f32 %p99, %f590, 0f400CCCCD;
|
||||
selp.f32 %f670, 0f7FFFFFFF, %f670, %p99;
|
||||
|
||||
BB0_63:
|
||||
add.f32 %f592, %f78, 0f400CCCCD;
|
||||
mov.b32 %r62, %f592;
|
||||
setp.lt.s32 %p101, %r62, 2139095040;
|
||||
@%p101 bra BB0_68;
|
||||
|
||||
setp.gtu.f32 %p102, %f78, 0f7F800000;
|
||||
@%p102 bra BB0_67;
|
||||
bra.uni BB0_65;
|
||||
|
||||
BB0_67:
|
||||
add.f32 %f670, %f3, 0f400CCCCD;
|
||||
bra.uni BB0_68;
|
||||
|
||||
BB0_65:
|
||||
setp.neu.f32 %p103, %f78, 0f7F800000;
|
||||
@%p103 bra BB0_68;
|
||||
|
||||
selp.f32 %f670, 0fFF800000, 0f7F800000, %p6;
|
||||
|
||||
BB0_68:
|
||||
setp.eq.f32 %p104, %f3, 0f3F800000;
|
||||
selp.f32 %f593, 0f3F800000, %f670, %p104;
|
||||
cvt.sat.f32.f32 %f594, %f593;
|
||||
max.f32 %f595, %f64, %f77;
|
||||
max.f32 %f596, %f595, %f594;
|
||||
mov.f32 %f597, 0f3F800000;
|
||||
sub.f32 %f598, %f597, %f596;
|
||||
rcp.rn.f32 %f599, %f598;
|
||||
mul.f32 %f671, %f64, %f599;
|
||||
mul.f32 %f672, %f77, %f599;
|
||||
mul.f32 %f673, %f594, %f599;
|
||||
ld.const.u64 %rd16, [cs];
|
||||
|
||||
BB0_69:
|
||||
cvta.to.global.u64 %rd12, %rd16;
|
||||
shl.b64 %rd13, %rd1, 3;
|
||||
add.s64 %rd14, %rd12, %rd13;
|
||||
mov.f32 %f603, 0f3F800000;
|
||||
// inline asm
|
||||
{ cvt.rn.f16.f32 %rs11, %f603;}
|
||||
|
||||
// inline asm
|
||||
// inline asm
|
||||
{ cvt.rn.f16.f32 %rs10, %f673;}
|
||||
|
||||
// inline asm
|
||||
// inline asm
|
||||
{ cvt.rn.f16.f32 %rs9, %f672;}
|
||||
|
||||
// inline asm
|
||||
// inline asm
|
||||
{ cvt.rn.f16.f32 %rs8, %f671;}
|
||||
|
||||
// inline asm
|
||||
st.global.v4.u16 [%rd14], {%rs8, %rs9, %rs10, %rs11};
|
||||
ret;
|
||||
}
|
||||
|
||||
|
||||
8
Assets/Editor/x64/Bakery/denoiseFinish72.ptx.meta
Normal file
8
Assets/Editor/x64/Bakery/denoiseFinish72.ptx.meta
Normal file
@@ -0,0 +1,8 @@
|
||||
fileFormatVersion: 2
|
||||
guid: 09ae1bcf43095f94eadb21257b0ccff0
|
||||
timeCreated: 1605208943
|
||||
licenseType: Store
|
||||
DefaultImporter:
|
||||
userData:
|
||||
assetBundleName:
|
||||
assetBundleVariant:
|
||||
107
Assets/Editor/x64/Bakery/denoiseFinishOIDN.ptx
Normal file
107
Assets/Editor/x64/Bakery/denoiseFinishOIDN.ptx
Normal file
@@ -0,0 +1,107 @@
|
||||
//
|
||||
// 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 1 .b8 output_buffer[1];
|
||||
.global .align 1 .b8 image2[1];
|
||||
.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 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 4 .u32 _ZN21rti_internal_typeenum7pixelIDE = 4919;
|
||||
.global .align 4 .u32 _ZN21rti_internal_typeenum10resolutionE = 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 1 .b8 _ZN23rti_internal_annotation7pixelIDE[1];
|
||||
.global .align 1 .b8 _ZN23rti_internal_annotation10resolutionE[1];
|
||||
|
||||
.visible .entry _Z6oxMainv(
|
||||
|
||||
)
|
||||
{
|
||||
.reg .b16 %rs<5>;
|
||||
.reg .f32 %f<5>;
|
||||
.reg .b32 %r<13>;
|
||||
.reg .b64 %rd<15>;
|
||||
|
||||
|
||||
ld.global.v2.u32 {%r5, %r6}, [pixelID];
|
||||
cvt.u64.u32 %rd3, %r5;
|
||||
cvt.u64.u32 %rd4, %r6;
|
||||
mov.u64 %rd13, output_buffer;
|
||||
cvta.global.u64 %rd2, %rd13;
|
||||
mov.u32 %r3, 2;
|
||||
mov.u32 %r2, 12;
|
||||
mov.u64 %rd12, 0;
|
||||
// inline asm
|
||||
call (%rd1), _rt_buffer_get_64, (%rd2, %r3, %r2, %rd3, %rd4, %rd12, %rd12);
|
||||
// inline asm
|
||||
ld.f32 %f1, [%rd1];
|
||||
ld.f32 %f2, [%rd1+4];
|
||||
ld.f32 %f3, [%rd1+8];
|
||||
ld.global.v2.u32 {%r9, %r10}, [pixelID];
|
||||
cvt.u64.u32 %rd9, %r9;
|
||||
cvt.u64.u32 %rd10, %r10;
|
||||
mov.u64 %rd14, image2;
|
||||
cvta.global.u64 %rd8, %rd14;
|
||||
mov.u32 %r4, 8;
|
||||
// inline asm
|
||||
call (%rd7), _rt_buffer_get_64, (%rd8, %r3, %r4, %rd9, %rd10, %rd12, %rd12);
|
||||
// inline asm
|
||||
// inline asm
|
||||
{ cvt.rn.f16.f32 %rs3, %f3;}
|
||||
|
||||
// inline asm
|
||||
// inline asm
|
||||
{ cvt.rn.f16.f32 %rs2, %f2;}
|
||||
|
||||
// inline asm
|
||||
// inline asm
|
||||
{ cvt.rn.f16.f32 %rs1, %f1;}
|
||||
|
||||
// inline asm
|
||||
mov.f32 %f4, 0f3F800000;
|
||||
// inline asm
|
||||
{ cvt.rn.f16.f32 %rs4, %f4;}
|
||||
|
||||
// inline asm
|
||||
st.v4.u16 [%rd7], {%rs1, %rs2, %rs3, %rs4};
|
||||
ret;
|
||||
}
|
||||
|
||||
|
||||
8
Assets/Editor/x64/Bakery/denoiseFinishOIDN.ptx.meta
Normal file
8
Assets/Editor/x64/Bakery/denoiseFinishOIDN.ptx.meta
Normal file
@@ -0,0 +1,8 @@
|
||||
fileFormatVersion: 2
|
||||
guid: 37a0eda7ff3594543a94083548bef4fb
|
||||
timeCreated: 1589355275
|
||||
licenseType: Store
|
||||
DefaultImporter:
|
||||
userData:
|
||||
assetBundleName:
|
||||
assetBundleVariant:
|
||||
598
Assets/Editor/x64/Bakery/denoiseFinishSH.ptx
Normal file
598
Assets/Editor/x64/Bakery/denoiseFinishSH.ptx
Normal file
@@ -0,0 +1,598 @@
|
||||
//
|
||||
// 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 1 .b8 output_buffer[1];
|
||||
.global .align 1 .b8 image2[1];
|
||||
.global .align 1 .b8 image3[1];
|
||||
.global .align 4 .u32 mode;
|
||||
.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_typeinfo4modeE[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 4 .b8 _ZN21rti_internal_typename4modeE[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_typeenum4modeE = 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 1 .b8 _ZN21rti_internal_semantic4modeE[1];
|
||||
.global .align 1 .b8 _ZN23rti_internal_annotation7pixelIDE[1];
|
||||
.global .align 1 .b8 _ZN23rti_internal_annotation10resolutionE[1];
|
||||
.global .align 1 .b8 _ZN23rti_internal_annotation4modeE[1];
|
||||
|
||||
.visible .entry _Z6oxMainv(
|
||||
|
||||
)
|
||||
{
|
||||
.reg .pred %p<53>;
|
||||
.reg .b16 %rs<9>;
|
||||
.reg .f32 %f<338>;
|
||||
.reg .b32 %r<52>;
|
||||
.reg .b64 %rd<24>;
|
||||
|
||||
|
||||
ld.global.v2.u32 {%r3, %r4}, [pixelID];
|
||||
cvt.u64.u32 %rd3, %r3;
|
||||
cvt.u64.u32 %rd4, %r4;
|
||||
mov.u64 %rd7, output_buffer;
|
||||
cvta.global.u64 %rd2, %rd7;
|
||||
mov.u32 %r1, 2;
|
||||
mov.u32 %r2, 16;
|
||||
mov.u64 %rd6, 0;
|
||||
// inline asm
|
||||
call (%rd1), _rt_buffer_get_64, (%rd2, %r1, %r2, %rd3, %rd4, %rd6, %rd6);
|
||||
// inline asm
|
||||
ld.v4.f32 {%f46, %f47, %f48, %f49}, [%rd1];
|
||||
mov.f32 %f50, 0f3F8CCCCD;
|
||||
cvt.rzi.f32.f32 %f51, %f50;
|
||||
fma.rn.f32 %f52, %f51, 0fC0000000, 0f400CCCCD;
|
||||
abs.f32 %f3, %f52;
|
||||
abs.f32 %f5, %f46;
|
||||
setp.lt.f32 %p4, %f5, 0f00800000;
|
||||
mul.f32 %f53, %f5, 0f4B800000;
|
||||
selp.f32 %f54, 0fC3170000, 0fC2FE0000, %p4;
|
||||
selp.f32 %f55, %f53, %f5, %p4;
|
||||
mov.b32 %r7, %f55;
|
||||
and.b32 %r8, %r7, 8388607;
|
||||
or.b32 %r9, %r8, 1065353216;
|
||||
mov.b32 %f56, %r9;
|
||||
shr.u32 %r10, %r7, 23;
|
||||
cvt.rn.f32.u32 %f57, %r10;
|
||||
add.f32 %f58, %f54, %f57;
|
||||
setp.gt.f32 %p5, %f56, 0f3FB504F3;
|
||||
mul.f32 %f59, %f56, 0f3F000000;
|
||||
add.f32 %f60, %f58, 0f3F800000;
|
||||
selp.f32 %f61, %f59, %f56, %p5;
|
||||
selp.f32 %f62, %f60, %f58, %p5;
|
||||
add.f32 %f63, %f61, 0fBF800000;
|
||||
add.f32 %f45, %f61, 0f3F800000;
|
||||
// inline asm
|
||||
rcp.approx.ftz.f32 %f44,%f45;
|
||||
// inline asm
|
||||
add.f32 %f64, %f63, %f63;
|
||||
mul.f32 %f65, %f44, %f64;
|
||||
mul.f32 %f66, %f65, %f65;
|
||||
mov.f32 %f67, 0f3C4CAF63;
|
||||
mov.f32 %f68, 0f3B18F0FE;
|
||||
fma.rn.f32 %f69, %f68, %f66, %f67;
|
||||
mov.f32 %f70, 0f3DAAAABD;
|
||||
fma.rn.f32 %f71, %f69, %f66, %f70;
|
||||
mul.rn.f32 %f72, %f71, %f66;
|
||||
mul.rn.f32 %f73, %f72, %f65;
|
||||
sub.f32 %f74, %f63, %f65;
|
||||
neg.f32 %f75, %f65;
|
||||
add.f32 %f76, %f74, %f74;
|
||||
fma.rn.f32 %f77, %f75, %f63, %f76;
|
||||
mul.rn.f32 %f78, %f44, %f77;
|
||||
add.f32 %f79, %f73, %f65;
|
||||
sub.f32 %f80, %f65, %f79;
|
||||
add.f32 %f81, %f73, %f80;
|
||||
add.f32 %f82, %f78, %f81;
|
||||
add.f32 %f83, %f79, %f82;
|
||||
sub.f32 %f84, %f79, %f83;
|
||||
add.f32 %f85, %f82, %f84;
|
||||
mov.f32 %f86, 0f3F317200;
|
||||
mul.rn.f32 %f87, %f62, %f86;
|
||||
mov.f32 %f88, 0f35BFBE8E;
|
||||
mul.rn.f32 %f89, %f62, %f88;
|
||||
add.f32 %f90, %f87, %f83;
|
||||
sub.f32 %f91, %f87, %f90;
|
||||
add.f32 %f92, %f83, %f91;
|
||||
add.f32 %f93, %f85, %f92;
|
||||
add.f32 %f94, %f89, %f93;
|
||||
add.f32 %f95, %f90, %f94;
|
||||
sub.f32 %f96, %f90, %f95;
|
||||
add.f32 %f97, %f94, %f96;
|
||||
mov.f32 %f98, 0f400CCCCD;
|
||||
mul.rn.f32 %f99, %f98, %f95;
|
||||
neg.f32 %f100, %f99;
|
||||
fma.rn.f32 %f101, %f98, %f95, %f100;
|
||||
fma.rn.f32 %f102, %f98, %f97, %f101;
|
||||
mov.f32 %f103, 0f00000000;
|
||||
fma.rn.f32 %f104, %f103, %f95, %f102;
|
||||
add.rn.f32 %f105, %f99, %f104;
|
||||
neg.f32 %f106, %f105;
|
||||
add.rn.f32 %f107, %f99, %f106;
|
||||
add.rn.f32 %f108, %f107, %f104;
|
||||
mov.b32 %r11, %f105;
|
||||
setp.eq.s32 %p6, %r11, 1118925336;
|
||||
add.s32 %r12, %r11, -1;
|
||||
mov.b32 %f109, %r12;
|
||||
add.f32 %f110, %f108, 0f37000000;
|
||||
selp.f32 %f111, %f109, %f105, %p6;
|
||||
selp.f32 %f6, %f110, %f108, %p6;
|
||||
mul.f32 %f112, %f111, 0f3FB8AA3B;
|
||||
cvt.rzi.f32.f32 %f113, %f112;
|
||||
mov.f32 %f114, 0fBF317200;
|
||||
fma.rn.f32 %f115, %f113, %f114, %f111;
|
||||
mov.f32 %f116, 0fB5BFBE8E;
|
||||
fma.rn.f32 %f117, %f113, %f116, %f115;
|
||||
mul.f32 %f118, %f117, 0f3FB8AA3B;
|
||||
ex2.approx.ftz.f32 %f119, %f118;
|
||||
add.f32 %f120, %f113, 0f00000000;
|
||||
ex2.approx.f32 %f121, %f120;
|
||||
mul.f32 %f122, %f119, %f121;
|
||||
setp.lt.f32 %p7, %f111, 0fC2D20000;
|
||||
selp.f32 %f123, 0f00000000, %f122, %p7;
|
||||
setp.gt.f32 %p8, %f111, 0f42D20000;
|
||||
selp.f32 %f329, 0f7F800000, %f123, %p8;
|
||||
setp.eq.f32 %p9, %f329, 0f7F800000;
|
||||
@%p9 bra BB0_2;
|
||||
|
||||
fma.rn.f32 %f329, %f329, %f6, %f329;
|
||||
|
||||
BB0_2:
|
||||
setp.lt.f32 %p10, %f46, 0f00000000;
|
||||
setp.eq.f32 %p11, %f3, 0f3F800000;
|
||||
and.pred %p1, %p10, %p11;
|
||||
mov.b32 %r13, %f329;
|
||||
xor.b32 %r14, %r13, -2147483648;
|
||||
mov.b32 %f124, %r14;
|
||||
selp.f32 %f331, %f124, %f329, %p1;
|
||||
setp.eq.f32 %p12, %f46, 0f00000000;
|
||||
@%p12 bra BB0_5;
|
||||
bra.uni BB0_3;
|
||||
|
||||
BB0_5:
|
||||
add.f32 %f127, %f46, %f46;
|
||||
selp.f32 %f331, %f127, 0f00000000, %p11;
|
||||
bra.uni BB0_6;
|
||||
|
||||
BB0_3:
|
||||
setp.geu.f32 %p13, %f46, 0f00000000;
|
||||
@%p13 bra BB0_6;
|
||||
|
||||
mov.f32 %f328, 0f400CCCCD;
|
||||
cvt.rzi.f32.f32 %f126, %f328;
|
||||
setp.neu.f32 %p14, %f126, 0f400CCCCD;
|
||||
selp.f32 %f331, 0f7FFFFFFF, %f331, %p14;
|
||||
|
||||
BB0_6:
|
||||
abs.f32 %f305, %f46;
|
||||
add.f32 %f128, %f305, 0f400CCCCD;
|
||||
mov.b32 %r15, %f128;
|
||||
setp.lt.s32 %p16, %r15, 2139095040;
|
||||
@%p16 bra BB0_11;
|
||||
|
||||
abs.f32 %f326, %f46;
|
||||
setp.gtu.f32 %p17, %f326, 0f7F800000;
|
||||
@%p17 bra BB0_10;
|
||||
bra.uni BB0_8;
|
||||
|
||||
BB0_10:
|
||||
add.f32 %f331, %f46, 0f400CCCCD;
|
||||
bra.uni BB0_11;
|
||||
|
||||
BB0_8:
|
||||
abs.f32 %f327, %f46;
|
||||
setp.neu.f32 %p18, %f327, 0f7F800000;
|
||||
@%p18 bra BB0_11;
|
||||
|
||||
selp.f32 %f331, 0fFF800000, 0f7F800000, %p1;
|
||||
|
||||
BB0_11:
|
||||
mov.f32 %f314, 0fB5BFBE8E;
|
||||
mov.f32 %f313, 0fBF317200;
|
||||
mov.f32 %f312, 0f00000000;
|
||||
mov.f32 %f311, 0f35BFBE8E;
|
||||
mov.f32 %f310, 0f3F317200;
|
||||
mov.f32 %f309, 0f3DAAAABD;
|
||||
mov.f32 %f308, 0f3C4CAF63;
|
||||
mov.f32 %f307, 0f3B18F0FE;
|
||||
mov.f32 %f306, 0f400CCCCD;
|
||||
setp.eq.f32 %p19, %f46, 0f3F800000;
|
||||
selp.f32 %f131, 0f3F800000, %f331, %p19;
|
||||
cvt.sat.f32.f32 %f17, %f131;
|
||||
abs.f32 %f18, %f47;
|
||||
setp.lt.f32 %p20, %f18, 0f00800000;
|
||||
mul.f32 %f132, %f18, 0f4B800000;
|
||||
selp.f32 %f133, 0fC3170000, 0fC2FE0000, %p20;
|
||||
selp.f32 %f134, %f132, %f18, %p20;
|
||||
mov.b32 %r16, %f134;
|
||||
and.b32 %r17, %r16, 8388607;
|
||||
or.b32 %r18, %r17, 1065353216;
|
||||
mov.b32 %f135, %r18;
|
||||
shr.u32 %r19, %r16, 23;
|
||||
cvt.rn.f32.u32 %f136, %r19;
|
||||
add.f32 %f137, %f133, %f136;
|
||||
setp.gt.f32 %p21, %f135, 0f3FB504F3;
|
||||
mul.f32 %f138, %f135, 0f3F000000;
|
||||
add.f32 %f139, %f137, 0f3F800000;
|
||||
selp.f32 %f140, %f138, %f135, %p21;
|
||||
selp.f32 %f141, %f139, %f137, %p21;
|
||||
add.f32 %f142, %f140, 0fBF800000;
|
||||
add.f32 %f130, %f140, 0f3F800000;
|
||||
// inline asm
|
||||
rcp.approx.ftz.f32 %f129,%f130;
|
||||
// inline asm
|
||||
add.f32 %f143, %f142, %f142;
|
||||
mul.f32 %f144, %f129, %f143;
|
||||
mul.f32 %f145, %f144, %f144;
|
||||
fma.rn.f32 %f148, %f307, %f145, %f308;
|
||||
fma.rn.f32 %f150, %f148, %f145, %f309;
|
||||
mul.rn.f32 %f151, %f150, %f145;
|
||||
mul.rn.f32 %f152, %f151, %f144;
|
||||
sub.f32 %f153, %f142, %f144;
|
||||
neg.f32 %f154, %f144;
|
||||
add.f32 %f155, %f153, %f153;
|
||||
fma.rn.f32 %f156, %f154, %f142, %f155;
|
||||
mul.rn.f32 %f157, %f129, %f156;
|
||||
add.f32 %f158, %f152, %f144;
|
||||
sub.f32 %f159, %f144, %f158;
|
||||
add.f32 %f160, %f152, %f159;
|
||||
add.f32 %f161, %f157, %f160;
|
||||
add.f32 %f162, %f158, %f161;
|
||||
sub.f32 %f163, %f158, %f162;
|
||||
add.f32 %f164, %f161, %f163;
|
||||
mul.rn.f32 %f166, %f141, %f310;
|
||||
mul.rn.f32 %f168, %f141, %f311;
|
||||
add.f32 %f169, %f166, %f162;
|
||||
sub.f32 %f170, %f166, %f169;
|
||||
add.f32 %f171, %f162, %f170;
|
||||
add.f32 %f172, %f164, %f171;
|
||||
add.f32 %f173, %f168, %f172;
|
||||
add.f32 %f174, %f169, %f173;
|
||||
sub.f32 %f175, %f169, %f174;
|
||||
add.f32 %f176, %f173, %f175;
|
||||
mul.rn.f32 %f178, %f306, %f174;
|
||||
neg.f32 %f179, %f178;
|
||||
fma.rn.f32 %f180, %f306, %f174, %f179;
|
||||
fma.rn.f32 %f181, %f306, %f176, %f180;
|
||||
fma.rn.f32 %f183, %f312, %f174, %f181;
|
||||
add.rn.f32 %f184, %f178, %f183;
|
||||
neg.f32 %f185, %f184;
|
||||
add.rn.f32 %f186, %f178, %f185;
|
||||
add.rn.f32 %f187, %f186, %f183;
|
||||
mov.b32 %r20, %f184;
|
||||
setp.eq.s32 %p22, %r20, 1118925336;
|
||||
add.s32 %r21, %r20, -1;
|
||||
mov.b32 %f188, %r21;
|
||||
add.f32 %f189, %f187, 0f37000000;
|
||||
selp.f32 %f190, %f188, %f184, %p22;
|
||||
selp.f32 %f19, %f189, %f187, %p22;
|
||||
mul.f32 %f191, %f190, 0f3FB8AA3B;
|
||||
cvt.rzi.f32.f32 %f192, %f191;
|
||||
fma.rn.f32 %f194, %f192, %f313, %f190;
|
||||
fma.rn.f32 %f196, %f192, %f314, %f194;
|
||||
mul.f32 %f197, %f196, 0f3FB8AA3B;
|
||||
ex2.approx.ftz.f32 %f198, %f197;
|
||||
add.f32 %f199, %f192, 0f00000000;
|
||||
ex2.approx.f32 %f200, %f199;
|
||||
mul.f32 %f201, %f198, %f200;
|
||||
setp.lt.f32 %p23, %f190, 0fC2D20000;
|
||||
selp.f32 %f202, 0f00000000, %f201, %p23;
|
||||
setp.gt.f32 %p24, %f190, 0f42D20000;
|
||||
selp.f32 %f332, 0f7F800000, %f202, %p24;
|
||||
setp.eq.f32 %p25, %f332, 0f7F800000;
|
||||
@%p25 bra BB0_13;
|
||||
|
||||
fma.rn.f32 %f332, %f332, %f19, %f332;
|
||||
|
||||
BB0_13:
|
||||
setp.lt.f32 %p26, %f47, 0f00000000;
|
||||
and.pred %p2, %p26, %p11;
|
||||
mov.b32 %r22, %f332;
|
||||
xor.b32 %r23, %r22, -2147483648;
|
||||
mov.b32 %f203, %r23;
|
||||
selp.f32 %f334, %f203, %f332, %p2;
|
||||
setp.eq.f32 %p28, %f47, 0f00000000;
|
||||
@%p28 bra BB0_16;
|
||||
bra.uni BB0_14;
|
||||
|
||||
BB0_16:
|
||||
add.f32 %f206, %f47, %f47;
|
||||
selp.f32 %f334, %f206, 0f00000000, %p11;
|
||||
bra.uni BB0_17;
|
||||
|
||||
BB0_14:
|
||||
setp.geu.f32 %p29, %f47, 0f00000000;
|
||||
@%p29 bra BB0_17;
|
||||
|
||||
mov.f32 %f325, 0f400CCCCD;
|
||||
cvt.rzi.f32.f32 %f205, %f325;
|
||||
setp.neu.f32 %p30, %f205, 0f400CCCCD;
|
||||
selp.f32 %f334, 0f7FFFFFFF, %f334, %p30;
|
||||
|
||||
BB0_17:
|
||||
add.f32 %f207, %f18, 0f400CCCCD;
|
||||
mov.b32 %r24, %f207;
|
||||
setp.lt.s32 %p32, %r24, 2139095040;
|
||||
@%p32 bra BB0_22;
|
||||
|
||||
setp.gtu.f32 %p33, %f18, 0f7F800000;
|
||||
@%p33 bra BB0_21;
|
||||
bra.uni BB0_19;
|
||||
|
||||
BB0_21:
|
||||
add.f32 %f334, %f47, 0f400CCCCD;
|
||||
bra.uni BB0_22;
|
||||
|
||||
BB0_19:
|
||||
setp.neu.f32 %p34, %f18, 0f7F800000;
|
||||
@%p34 bra BB0_22;
|
||||
|
||||
selp.f32 %f334, 0fFF800000, 0f7F800000, %p2;
|
||||
|
||||
BB0_22:
|
||||
mov.f32 %f323, 0fB5BFBE8E;
|
||||
mov.f32 %f322, 0fBF317200;
|
||||
mov.f32 %f321, 0f00000000;
|
||||
mov.f32 %f320, 0f35BFBE8E;
|
||||
mov.f32 %f319, 0f3F317200;
|
||||
mov.f32 %f318, 0f3DAAAABD;
|
||||
mov.f32 %f317, 0f3C4CAF63;
|
||||
mov.f32 %f316, 0f3B18F0FE;
|
||||
mov.f32 %f315, 0f400CCCCD;
|
||||
setp.eq.f32 %p35, %f47, 0f3F800000;
|
||||
selp.f32 %f210, 0f3F800000, %f334, %p35;
|
||||
cvt.sat.f32.f32 %f30, %f210;
|
||||
abs.f32 %f31, %f48;
|
||||
setp.lt.f32 %p36, %f31, 0f00800000;
|
||||
mul.f32 %f211, %f31, 0f4B800000;
|
||||
selp.f32 %f212, 0fC3170000, 0fC2FE0000, %p36;
|
||||
selp.f32 %f213, %f211, %f31, %p36;
|
||||
mov.b32 %r25, %f213;
|
||||
and.b32 %r26, %r25, 8388607;
|
||||
or.b32 %r27, %r26, 1065353216;
|
||||
mov.b32 %f214, %r27;
|
||||
shr.u32 %r28, %r25, 23;
|
||||
cvt.rn.f32.u32 %f215, %r28;
|
||||
add.f32 %f216, %f212, %f215;
|
||||
setp.gt.f32 %p37, %f214, 0f3FB504F3;
|
||||
mul.f32 %f217, %f214, 0f3F000000;
|
||||
add.f32 %f218, %f216, 0f3F800000;
|
||||
selp.f32 %f219, %f217, %f214, %p37;
|
||||
selp.f32 %f220, %f218, %f216, %p37;
|
||||
add.f32 %f221, %f219, 0fBF800000;
|
||||
add.f32 %f209, %f219, 0f3F800000;
|
||||
// inline asm
|
||||
rcp.approx.ftz.f32 %f208,%f209;
|
||||
// inline asm
|
||||
add.f32 %f222, %f221, %f221;
|
||||
mul.f32 %f223, %f208, %f222;
|
||||
mul.f32 %f224, %f223, %f223;
|
||||
fma.rn.f32 %f227, %f316, %f224, %f317;
|
||||
fma.rn.f32 %f229, %f227, %f224, %f318;
|
||||
mul.rn.f32 %f230, %f229, %f224;
|
||||
mul.rn.f32 %f231, %f230, %f223;
|
||||
sub.f32 %f232, %f221, %f223;
|
||||
neg.f32 %f233, %f223;
|
||||
add.f32 %f234, %f232, %f232;
|
||||
fma.rn.f32 %f235, %f233, %f221, %f234;
|
||||
mul.rn.f32 %f236, %f208, %f235;
|
||||
add.f32 %f237, %f231, %f223;
|
||||
sub.f32 %f238, %f223, %f237;
|
||||
add.f32 %f239, %f231, %f238;
|
||||
add.f32 %f240, %f236, %f239;
|
||||
add.f32 %f241, %f237, %f240;
|
||||
sub.f32 %f242, %f237, %f241;
|
||||
add.f32 %f243, %f240, %f242;
|
||||
mul.rn.f32 %f245, %f220, %f319;
|
||||
mul.rn.f32 %f247, %f220, %f320;
|
||||
add.f32 %f248, %f245, %f241;
|
||||
sub.f32 %f249, %f245, %f248;
|
||||
add.f32 %f250, %f241, %f249;
|
||||
add.f32 %f251, %f243, %f250;
|
||||
add.f32 %f252, %f247, %f251;
|
||||
add.f32 %f253, %f248, %f252;
|
||||
sub.f32 %f254, %f248, %f253;
|
||||
add.f32 %f255, %f252, %f254;
|
||||
mul.rn.f32 %f257, %f315, %f253;
|
||||
neg.f32 %f258, %f257;
|
||||
fma.rn.f32 %f259, %f315, %f253, %f258;
|
||||
fma.rn.f32 %f260, %f315, %f255, %f259;
|
||||
fma.rn.f32 %f262, %f321, %f253, %f260;
|
||||
add.rn.f32 %f263, %f257, %f262;
|
||||
neg.f32 %f264, %f263;
|
||||
add.rn.f32 %f265, %f257, %f264;
|
||||
add.rn.f32 %f266, %f265, %f262;
|
||||
mov.b32 %r29, %f263;
|
||||
setp.eq.s32 %p38, %r29, 1118925336;
|
||||
add.s32 %r30, %r29, -1;
|
||||
mov.b32 %f267, %r30;
|
||||
add.f32 %f268, %f266, 0f37000000;
|
||||
selp.f32 %f269, %f267, %f263, %p38;
|
||||
selp.f32 %f32, %f268, %f266, %p38;
|
||||
mul.f32 %f270, %f269, 0f3FB8AA3B;
|
||||
cvt.rzi.f32.f32 %f271, %f270;
|
||||
fma.rn.f32 %f273, %f271, %f322, %f269;
|
||||
fma.rn.f32 %f275, %f271, %f323, %f273;
|
||||
mul.f32 %f276, %f275, 0f3FB8AA3B;
|
||||
ex2.approx.ftz.f32 %f277, %f276;
|
||||
add.f32 %f278, %f271, 0f00000000;
|
||||
ex2.approx.f32 %f279, %f278;
|
||||
mul.f32 %f280, %f277, %f279;
|
||||
setp.lt.f32 %p39, %f269, 0fC2D20000;
|
||||
selp.f32 %f281, 0f00000000, %f280, %p39;
|
||||
setp.gt.f32 %p40, %f269, 0f42D20000;
|
||||
selp.f32 %f335, 0f7F800000, %f281, %p40;
|
||||
setp.eq.f32 %p41, %f335, 0f7F800000;
|
||||
@%p41 bra BB0_24;
|
||||
|
||||
fma.rn.f32 %f335, %f335, %f32, %f335;
|
||||
|
||||
BB0_24:
|
||||
setp.lt.f32 %p42, %f48, 0f00000000;
|
||||
and.pred %p3, %p42, %p11;
|
||||
mov.b32 %r31, %f335;
|
||||
xor.b32 %r32, %r31, -2147483648;
|
||||
mov.b32 %f282, %r32;
|
||||
selp.f32 %f337, %f282, %f335, %p3;
|
||||
setp.eq.f32 %p44, %f48, 0f00000000;
|
||||
@%p44 bra BB0_27;
|
||||
bra.uni BB0_25;
|
||||
|
||||
BB0_27:
|
||||
add.f32 %f285, %f48, %f48;
|
||||
selp.f32 %f337, %f285, 0f00000000, %p11;
|
||||
bra.uni BB0_28;
|
||||
|
||||
BB0_25:
|
||||
setp.geu.f32 %p45, %f48, 0f00000000;
|
||||
@%p45 bra BB0_28;
|
||||
|
||||
mov.f32 %f324, 0f400CCCCD;
|
||||
cvt.rzi.f32.f32 %f284, %f324;
|
||||
setp.neu.f32 %p46, %f284, 0f400CCCCD;
|
||||
selp.f32 %f337, 0f7FFFFFFF, %f337, %p46;
|
||||
|
||||
BB0_28:
|
||||
add.f32 %f286, %f31, 0f400CCCCD;
|
||||
mov.b32 %r33, %f286;
|
||||
setp.lt.s32 %p48, %r33, 2139095040;
|
||||
@%p48 bra BB0_33;
|
||||
|
||||
setp.gtu.f32 %p49, %f31, 0f7F800000;
|
||||
@%p49 bra BB0_32;
|
||||
bra.uni BB0_30;
|
||||
|
||||
BB0_32:
|
||||
add.f32 %f337, %f48, 0f400CCCCD;
|
||||
bra.uni BB0_33;
|
||||
|
||||
BB0_30:
|
||||
setp.neu.f32 %p50, %f31, 0f7F800000;
|
||||
@%p50 bra BB0_33;
|
||||
|
||||
selp.f32 %f337, 0fFF800000, 0f7F800000, %p3;
|
||||
|
||||
BB0_33:
|
||||
setp.eq.f32 %p51, %f48, 0f3F800000;
|
||||
selp.f32 %f287, 0f3F800000, %f337, %p51;
|
||||
cvt.sat.f32.f32 %f43, %f287;
|
||||
ld.global.u32 %r34, [mode];
|
||||
setp.gt.s32 %p52, %r34, 0;
|
||||
@%p52 bra BB0_35;
|
||||
bra.uni BB0_34;
|
||||
|
||||
BB0_35:
|
||||
mov.u64 %rd23, 0;
|
||||
mov.u32 %r51, 2;
|
||||
ld.global.v2.u32 {%r43, %r44}, [pixelID];
|
||||
cvt.u64.u32 %rd17, %r43;
|
||||
cvt.u64.u32 %rd18, %r44;
|
||||
mov.u64 %rd21, image3;
|
||||
cvta.global.u64 %rd16, %rd21;
|
||||
mov.u32 %r42, 4;
|
||||
// inline asm
|
||||
call (%rd15), _rt_buffer_get_64, (%rd16, %r51, %r42, %rd17, %rd18, %rd23, %rd23);
|
||||
// inline asm
|
||||
cvt.sat.f32.f32 %f299, %f17;
|
||||
mul.f32 %f300, %f299, 0f437F0000;
|
||||
cvt.rzi.u32.f32 %r47, %f300;
|
||||
cvt.sat.f32.f32 %f301, %f30;
|
||||
mul.f32 %f302, %f301, 0f437F0000;
|
||||
cvt.rzi.u32.f32 %r48, %f302;
|
||||
cvt.sat.f32.f32 %f303, %f43;
|
||||
mul.f32 %f304, %f303, 0f437F0000;
|
||||
cvt.rzi.u32.f32 %r49, %f304;
|
||||
cvt.u16.u32 %rs5, %r49;
|
||||
cvt.u16.u32 %rs6, %r48;
|
||||
cvt.u16.u32 %rs7, %r47;
|
||||
mov.u16 %rs8, 255;
|
||||
st.v4.u8 [%rd15], {%rs7, %rs6, %rs5, %rs8};
|
||||
bra.uni BB0_36;
|
||||
|
||||
BB0_34:
|
||||
mov.u64 %rd22, 0;
|
||||
mov.u32 %r50, 2;
|
||||
max.f32 %f292, %f17, %f30;
|
||||
max.f32 %f293, %f292, %f43;
|
||||
mov.f32 %f291, 0f3F800000;
|
||||
sub.f32 %f294, %f291, %f293;
|
||||
rcp.rn.f32 %f295, %f294;
|
||||
mul.f32 %f296, %f17, %f295;
|
||||
mul.f32 %f297, %f30, %f295;
|
||||
mul.f32 %f298, %f43, %f295;
|
||||
mul.f32 %f288, %f296, 0f3E800000;
|
||||
mul.f32 %f289, %f297, 0f3E800000;
|
||||
mul.f32 %f290, %f298, 0f3E800000;
|
||||
ld.global.v2.u32 {%r37, %r38}, [pixelID];
|
||||
cvt.u64.u32 %rd10, %r37;
|
||||
cvt.u64.u32 %rd11, %r38;
|
||||
mov.u64 %rd14, image2;
|
||||
cvta.global.u64 %rd9, %rd14;
|
||||
mov.u32 %r36, 8;
|
||||
// inline asm
|
||||
call (%rd8), _rt_buffer_get_64, (%rd9, %r50, %r36, %rd10, %rd11, %rd22, %rd22);
|
||||
// inline asm
|
||||
// inline asm
|
||||
{ cvt.rn.f16.f32 %rs3, %f290;}
|
||||
|
||||
// inline asm
|
||||
// inline asm
|
||||
{ cvt.rn.f16.f32 %rs2, %f289;}
|
||||
|
||||
// inline asm
|
||||
// inline asm
|
||||
{ cvt.rn.f16.f32 %rs1, %f288;}
|
||||
|
||||
// inline asm
|
||||
// inline asm
|
||||
{ cvt.rn.f16.f32 %rs4, %f291;}
|
||||
|
||||
// inline asm
|
||||
st.v4.u16 [%rd8], {%rs1, %rs2, %rs3, %rs4};
|
||||
|
||||
BB0_36:
|
||||
ret;
|
||||
}
|
||||
|
||||
|
||||
8
Assets/Editor/x64/Bakery/denoiseFinishSH.ptx.meta
Normal file
8
Assets/Editor/x64/Bakery/denoiseFinishSH.ptx.meta
Normal file
@@ -0,0 +1,8 @@
|
||||
fileFormatVersion: 2
|
||||
guid: 7fd4e37367df86746b44d9f09203b06a
|
||||
timeCreated: 1538345689
|
||||
licenseType: Store
|
||||
DefaultImporter:
|
||||
userData:
|
||||
assetBundleName:
|
||||
assetBundleVariant:
|
||||
526
Assets/Editor/x64/Bakery/denoiseFinishSH72.ptx
Normal file
526
Assets/Editor/x64/Bakery/denoiseFinishSH72.ptx
Normal file
@@ -0,0 +1,526 @@
|
||||
//
|
||||
// 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 __raygen__oxMain
|
||||
.const .align 8 .b8 cs[32];
|
||||
|
||||
.visible .entry __raygen__oxMain(
|
||||
|
||||
)
|
||||
{
|
||||
.reg .pred %p<53>;
|
||||
.reg .b16 %rs<9>;
|
||||
.reg .f32 %f<326>;
|
||||
.reg .b32 %r<39>;
|
||||
.reg .b64 %rd<14>;
|
||||
|
||||
|
||||
// inline asm
|
||||
call (%r1), _optix_get_launch_index_x, ();
|
||||
// inline asm
|
||||
// inline asm
|
||||
call (%r2), _optix_get_launch_index_y, ();
|
||||
// inline asm
|
||||
ld.const.u64 %rd2, [cs+8];
|
||||
cvta.to.global.u64 %rd3, %rd2;
|
||||
ld.const.v2.u32 {%r4, %r5}, [cs+24];
|
||||
mad.lo.s32 %r7, %r4, %r2, %r1;
|
||||
cvt.u64.u32 %rd1, %r7;
|
||||
mul.wide.u32 %rd4, %r7, 16;
|
||||
add.s64 %rd5, %rd3, %rd4;
|
||||
ld.global.v4.f32 {%f43, %f44, %f45, %f46}, [%rd5];
|
||||
setp.gt.s32 %p4, %r5, 0;
|
||||
@%p4 bra BB0_2;
|
||||
bra.uni BB0_1;
|
||||
|
||||
BB0_2:
|
||||
mov.f32 %f53, 0f3F8CCCCD;
|
||||
cvt.rzi.f32.f32 %f54, %f53;
|
||||
fma.rn.f32 %f55, %f54, 0fC0000000, 0f400CCCCD;
|
||||
abs.f32 %f4, %f55;
|
||||
abs.f32 %f5, %f43;
|
||||
setp.lt.f32 %p5, %f5, 0f00800000;
|
||||
mul.f32 %f56, %f5, 0f4B800000;
|
||||
selp.f32 %f57, 0fC3170000, 0fC2FE0000, %p5;
|
||||
selp.f32 %f58, %f56, %f5, %p5;
|
||||
mov.b32 %r9, %f58;
|
||||
and.b32 %r10, %r9, 8388607;
|
||||
or.b32 %r11, %r10, 1065353216;
|
||||
mov.b32 %f59, %r11;
|
||||
shr.u32 %r12, %r9, 23;
|
||||
cvt.rn.f32.u32 %f60, %r12;
|
||||
add.f32 %f61, %f57, %f60;
|
||||
setp.gt.f32 %p6, %f59, 0f3FB504F3;
|
||||
mul.f32 %f62, %f59, 0f3F000000;
|
||||
add.f32 %f63, %f61, 0f3F800000;
|
||||
selp.f32 %f64, %f62, %f59, %p6;
|
||||
selp.f32 %f65, %f63, %f61, %p6;
|
||||
add.f32 %f66, %f64, 0fBF800000;
|
||||
add.f32 %f52, %f64, 0f3F800000;
|
||||
// inline asm
|
||||
rcp.approx.ftz.f32 %f51,%f52;
|
||||
// inline asm
|
||||
add.f32 %f67, %f66, %f66;
|
||||
mul.f32 %f68, %f51, %f67;
|
||||
mul.f32 %f69, %f68, %f68;
|
||||
mov.f32 %f70, 0f3C4CAF63;
|
||||
mov.f32 %f71, 0f3B18F0FE;
|
||||
fma.rn.f32 %f72, %f71, %f69, %f70;
|
||||
mov.f32 %f73, 0f3DAAAABD;
|
||||
fma.rn.f32 %f74, %f72, %f69, %f73;
|
||||
mul.rn.f32 %f75, %f74, %f69;
|
||||
mul.rn.f32 %f76, %f75, %f68;
|
||||
sub.f32 %f77, %f66, %f68;
|
||||
neg.f32 %f78, %f68;
|
||||
add.f32 %f79, %f77, %f77;
|
||||
fma.rn.f32 %f80, %f78, %f66, %f79;
|
||||
mul.rn.f32 %f81, %f51, %f80;
|
||||
add.f32 %f82, %f76, %f68;
|
||||
sub.f32 %f83, %f68, %f82;
|
||||
add.f32 %f84, %f76, %f83;
|
||||
add.f32 %f85, %f81, %f84;
|
||||
add.f32 %f86, %f82, %f85;
|
||||
sub.f32 %f87, %f82, %f86;
|
||||
add.f32 %f88, %f85, %f87;
|
||||
mov.f32 %f89, 0f3F317200;
|
||||
mul.rn.f32 %f90, %f65, %f89;
|
||||
mov.f32 %f91, 0f35BFBE8E;
|
||||
mul.rn.f32 %f92, %f65, %f91;
|
||||
add.f32 %f93, %f90, %f86;
|
||||
sub.f32 %f94, %f90, %f93;
|
||||
add.f32 %f95, %f86, %f94;
|
||||
add.f32 %f96, %f88, %f95;
|
||||
add.f32 %f97, %f92, %f96;
|
||||
add.f32 %f98, %f93, %f97;
|
||||
sub.f32 %f99, %f93, %f98;
|
||||
add.f32 %f100, %f97, %f99;
|
||||
mov.f32 %f101, 0f400CCCCD;
|
||||
mul.rn.f32 %f102, %f101, %f98;
|
||||
neg.f32 %f103, %f102;
|
||||
fma.rn.f32 %f104, %f101, %f98, %f103;
|
||||
fma.rn.f32 %f105, %f101, %f100, %f104;
|
||||
mov.f32 %f106, 0f00000000;
|
||||
fma.rn.f32 %f107, %f106, %f98, %f105;
|
||||
add.rn.f32 %f108, %f102, %f107;
|
||||
neg.f32 %f109, %f108;
|
||||
add.rn.f32 %f110, %f102, %f109;
|
||||
add.rn.f32 %f111, %f110, %f107;
|
||||
mov.b32 %r13, %f108;
|
||||
setp.eq.s32 %p7, %r13, 1118925336;
|
||||
add.s32 %r14, %r13, -1;
|
||||
mov.b32 %f112, %r14;
|
||||
add.f32 %f113, %f111, 0f37000000;
|
||||
selp.f32 %f114, %f112, %f108, %p7;
|
||||
selp.f32 %f6, %f113, %f111, %p7;
|
||||
mul.f32 %f115, %f114, 0f3FB8AA3B;
|
||||
cvt.rzi.f32.f32 %f116, %f115;
|
||||
mov.f32 %f117, 0fBF317200;
|
||||
fma.rn.f32 %f118, %f116, %f117, %f114;
|
||||
mov.f32 %f119, 0fB5BFBE8E;
|
||||
fma.rn.f32 %f120, %f116, %f119, %f118;
|
||||
mul.f32 %f121, %f120, 0f3FB8AA3B;
|
||||
ex2.approx.ftz.f32 %f122, %f121;
|
||||
add.f32 %f123, %f116, 0f00000000;
|
||||
ex2.approx.f32 %f124, %f123;
|
||||
mul.f32 %f125, %f122, %f124;
|
||||
setp.lt.f32 %p8, %f114, 0fC2D20000;
|
||||
selp.f32 %f126, 0f00000000, %f125, %p8;
|
||||
setp.gt.f32 %p9, %f114, 0f42D20000;
|
||||
selp.f32 %f317, 0f7F800000, %f126, %p9;
|
||||
setp.eq.f32 %p10, %f317, 0f7F800000;
|
||||
@%p10 bra BB0_4;
|
||||
|
||||
fma.rn.f32 %f317, %f317, %f6, %f317;
|
||||
|
||||
BB0_4:
|
||||
setp.lt.f32 %p11, %f43, 0f00000000;
|
||||
setp.eq.f32 %p12, %f4, 0f3F800000;
|
||||
and.pred %p1, %p11, %p12;
|
||||
mov.b32 %r15, %f317;
|
||||
xor.b32 %r16, %r15, -2147483648;
|
||||
mov.b32 %f127, %r16;
|
||||
selp.f32 %f319, %f127, %f317, %p1;
|
||||
setp.eq.f32 %p13, %f43, 0f00000000;
|
||||
@%p13 bra BB0_7;
|
||||
bra.uni BB0_5;
|
||||
|
||||
BB0_7:
|
||||
add.f32 %f130, %f43, %f43;
|
||||
selp.f32 %f319, %f130, 0f00000000, %p12;
|
||||
bra.uni BB0_8;
|
||||
|
||||
BB0_1:
|
||||
ld.const.u64 %rd6, [cs];
|
||||
cvta.to.global.u64 %rd7, %rd6;
|
||||
shl.b64 %rd8, %rd1, 3;
|
||||
add.s64 %rd9, %rd7, %rd8;
|
||||
mov.f32 %f50, 0f3F800000;
|
||||
// inline asm
|
||||
{ cvt.rn.f16.f32 %rs4, %f50;}
|
||||
|
||||
// inline asm
|
||||
// inline asm
|
||||
{ cvt.rn.f16.f32 %rs3, %f45;}
|
||||
|
||||
// inline asm
|
||||
// inline asm
|
||||
{ cvt.rn.f16.f32 %rs2, %f44;}
|
||||
|
||||
// inline asm
|
||||
// inline asm
|
||||
{ cvt.rn.f16.f32 %rs1, %f43;}
|
||||
|
||||
// inline asm
|
||||
st.global.v4.u16 [%rd9], {%rs1, %rs2, %rs3, %rs4};
|
||||
bra.uni BB0_36;
|
||||
|
||||
BB0_5:
|
||||
setp.geu.f32 %p14, %f43, 0f00000000;
|
||||
@%p14 bra BB0_8;
|
||||
|
||||
cvt.rzi.f32.f32 %f129, %f101;
|
||||
setp.neu.f32 %p15, %f129, 0f400CCCCD;
|
||||
selp.f32 %f319, 0f7FFFFFFF, %f319, %p15;
|
||||
|
||||
BB0_8:
|
||||
abs.f32 %f298, %f43;
|
||||
add.f32 %f131, %f298, 0f400CCCCD;
|
||||
mov.b32 %r17, %f131;
|
||||
setp.lt.s32 %p17, %r17, 2139095040;
|
||||
@%p17 bra BB0_13;
|
||||
|
||||
abs.f32 %f315, %f43;
|
||||
setp.gtu.f32 %p18, %f315, 0f7F800000;
|
||||
@%p18 bra BB0_12;
|
||||
bra.uni BB0_10;
|
||||
|
||||
BB0_12:
|
||||
add.f32 %f319, %f43, 0f400CCCCD;
|
||||
bra.uni BB0_13;
|
||||
|
||||
BB0_10:
|
||||
abs.f32 %f316, %f43;
|
||||
setp.neu.f32 %p19, %f316, 0f7F800000;
|
||||
@%p19 bra BB0_13;
|
||||
|
||||
selp.f32 %f319, 0fFF800000, 0f7F800000, %p1;
|
||||
|
||||
BB0_13:
|
||||
mov.f32 %f306, 0fB5BFBE8E;
|
||||
mov.f32 %f305, 0fBF317200;
|
||||
mov.f32 %f304, 0f00000000;
|
||||
mov.f32 %f303, 0f35BFBE8E;
|
||||
mov.f32 %f302, 0f3F317200;
|
||||
mov.f32 %f301, 0f3DAAAABD;
|
||||
mov.f32 %f300, 0f3C4CAF63;
|
||||
mov.f32 %f299, 0f3B18F0FE;
|
||||
setp.eq.f32 %p20, %f43, 0f3F800000;
|
||||
selp.f32 %f134, 0f3F800000, %f319, %p20;
|
||||
cvt.sat.f32.f32 %f17, %f134;
|
||||
abs.f32 %f18, %f44;
|
||||
setp.lt.f32 %p21, %f18, 0f00800000;
|
||||
mul.f32 %f135, %f18, 0f4B800000;
|
||||
selp.f32 %f136, 0fC3170000, 0fC2FE0000, %p21;
|
||||
selp.f32 %f137, %f135, %f18, %p21;
|
||||
mov.b32 %r18, %f137;
|
||||
and.b32 %r19, %r18, 8388607;
|
||||
or.b32 %r20, %r19, 1065353216;
|
||||
mov.b32 %f138, %r20;
|
||||
shr.u32 %r21, %r18, 23;
|
||||
cvt.rn.f32.u32 %f139, %r21;
|
||||
add.f32 %f140, %f136, %f139;
|
||||
setp.gt.f32 %p22, %f138, 0f3FB504F3;
|
||||
mul.f32 %f141, %f138, 0f3F000000;
|
||||
add.f32 %f142, %f140, 0f3F800000;
|
||||
selp.f32 %f143, %f141, %f138, %p22;
|
||||
selp.f32 %f144, %f142, %f140, %p22;
|
||||
add.f32 %f145, %f143, 0fBF800000;
|
||||
add.f32 %f133, %f143, 0f3F800000;
|
||||
// inline asm
|
||||
rcp.approx.ftz.f32 %f132,%f133;
|
||||
// inline asm
|
||||
add.f32 %f146, %f145, %f145;
|
||||
mul.f32 %f147, %f132, %f146;
|
||||
mul.f32 %f148, %f147, %f147;
|
||||
fma.rn.f32 %f151, %f299, %f148, %f300;
|
||||
fma.rn.f32 %f153, %f151, %f148, %f301;
|
||||
mul.rn.f32 %f154, %f153, %f148;
|
||||
mul.rn.f32 %f155, %f154, %f147;
|
||||
sub.f32 %f156, %f145, %f147;
|
||||
neg.f32 %f157, %f147;
|
||||
add.f32 %f158, %f156, %f156;
|
||||
fma.rn.f32 %f159, %f157, %f145, %f158;
|
||||
mul.rn.f32 %f160, %f132, %f159;
|
||||
add.f32 %f161, %f155, %f147;
|
||||
sub.f32 %f162, %f147, %f161;
|
||||
add.f32 %f163, %f155, %f162;
|
||||
add.f32 %f164, %f160, %f163;
|
||||
add.f32 %f165, %f161, %f164;
|
||||
sub.f32 %f166, %f161, %f165;
|
||||
add.f32 %f167, %f164, %f166;
|
||||
mul.rn.f32 %f169, %f144, %f302;
|
||||
mul.rn.f32 %f171, %f144, %f303;
|
||||
add.f32 %f172, %f169, %f165;
|
||||
sub.f32 %f173, %f169, %f172;
|
||||
add.f32 %f174, %f165, %f173;
|
||||
add.f32 %f175, %f167, %f174;
|
||||
add.f32 %f176, %f171, %f175;
|
||||
add.f32 %f177, %f172, %f176;
|
||||
sub.f32 %f178, %f172, %f177;
|
||||
add.f32 %f179, %f176, %f178;
|
||||
mul.rn.f32 %f181, %f101, %f177;
|
||||
neg.f32 %f182, %f181;
|
||||
fma.rn.f32 %f183, %f101, %f177, %f182;
|
||||
fma.rn.f32 %f184, %f101, %f179, %f183;
|
||||
fma.rn.f32 %f186, %f304, %f177, %f184;
|
||||
add.rn.f32 %f187, %f181, %f186;
|
||||
neg.f32 %f188, %f187;
|
||||
add.rn.f32 %f189, %f181, %f188;
|
||||
add.rn.f32 %f190, %f189, %f186;
|
||||
mov.b32 %r22, %f187;
|
||||
setp.eq.s32 %p23, %r22, 1118925336;
|
||||
add.s32 %r23, %r22, -1;
|
||||
mov.b32 %f191, %r23;
|
||||
add.f32 %f192, %f190, 0f37000000;
|
||||
selp.f32 %f193, %f191, %f187, %p23;
|
||||
selp.f32 %f19, %f192, %f190, %p23;
|
||||
mul.f32 %f194, %f193, 0f3FB8AA3B;
|
||||
cvt.rzi.f32.f32 %f195, %f194;
|
||||
fma.rn.f32 %f197, %f195, %f305, %f193;
|
||||
fma.rn.f32 %f199, %f195, %f306, %f197;
|
||||
mul.f32 %f200, %f199, 0f3FB8AA3B;
|
||||
ex2.approx.ftz.f32 %f201, %f200;
|
||||
add.f32 %f202, %f195, 0f00000000;
|
||||
ex2.approx.f32 %f203, %f202;
|
||||
mul.f32 %f204, %f201, %f203;
|
||||
setp.lt.f32 %p24, %f193, 0fC2D20000;
|
||||
selp.f32 %f205, 0f00000000, %f204, %p24;
|
||||
setp.gt.f32 %p25, %f193, 0f42D20000;
|
||||
selp.f32 %f320, 0f7F800000, %f205, %p25;
|
||||
setp.eq.f32 %p26, %f320, 0f7F800000;
|
||||
@%p26 bra BB0_15;
|
||||
|
||||
fma.rn.f32 %f320, %f320, %f19, %f320;
|
||||
|
||||
BB0_15:
|
||||
setp.lt.f32 %p27, %f44, 0f00000000;
|
||||
and.pred %p2, %p27, %p12;
|
||||
mov.b32 %r24, %f320;
|
||||
xor.b32 %r25, %r24, -2147483648;
|
||||
mov.b32 %f206, %r25;
|
||||
selp.f32 %f322, %f206, %f320, %p2;
|
||||
setp.eq.f32 %p29, %f44, 0f00000000;
|
||||
@%p29 bra BB0_18;
|
||||
bra.uni BB0_16;
|
||||
|
||||
BB0_18:
|
||||
add.f32 %f209, %f44, %f44;
|
||||
selp.f32 %f322, %f209, 0f00000000, %p12;
|
||||
bra.uni BB0_19;
|
||||
|
||||
BB0_16:
|
||||
setp.geu.f32 %p30, %f44, 0f00000000;
|
||||
@%p30 bra BB0_19;
|
||||
|
||||
cvt.rzi.f32.f32 %f208, %f101;
|
||||
setp.neu.f32 %p31, %f208, 0f400CCCCD;
|
||||
selp.f32 %f322, 0f7FFFFFFF, %f322, %p31;
|
||||
|
||||
BB0_19:
|
||||
add.f32 %f210, %f18, 0f400CCCCD;
|
||||
mov.b32 %r26, %f210;
|
||||
setp.lt.s32 %p33, %r26, 2139095040;
|
||||
@%p33 bra BB0_24;
|
||||
|
||||
setp.gtu.f32 %p34, %f18, 0f7F800000;
|
||||
@%p34 bra BB0_23;
|
||||
bra.uni BB0_21;
|
||||
|
||||
BB0_23:
|
||||
add.f32 %f322, %f44, 0f400CCCCD;
|
||||
bra.uni BB0_24;
|
||||
|
||||
BB0_21:
|
||||
setp.neu.f32 %p35, %f18, 0f7F800000;
|
||||
@%p35 bra BB0_24;
|
||||
|
||||
selp.f32 %f322, 0fFF800000, 0f7F800000, %p2;
|
||||
|
||||
BB0_24:
|
||||
mov.f32 %f314, 0fB5BFBE8E;
|
||||
mov.f32 %f313, 0fBF317200;
|
||||
mov.f32 %f312, 0f00000000;
|
||||
mov.f32 %f311, 0f35BFBE8E;
|
||||
mov.f32 %f310, 0f3F317200;
|
||||
mov.f32 %f309, 0f3DAAAABD;
|
||||
mov.f32 %f308, 0f3C4CAF63;
|
||||
mov.f32 %f307, 0f3B18F0FE;
|
||||
setp.eq.f32 %p36, %f44, 0f3F800000;
|
||||
selp.f32 %f213, 0f3F800000, %f322, %p36;
|
||||
cvt.sat.f32.f32 %f30, %f213;
|
||||
abs.f32 %f31, %f45;
|
||||
setp.lt.f32 %p37, %f31, 0f00800000;
|
||||
mul.f32 %f214, %f31, 0f4B800000;
|
||||
selp.f32 %f215, 0fC3170000, 0fC2FE0000, %p37;
|
||||
selp.f32 %f216, %f214, %f31, %p37;
|
||||
mov.b32 %r27, %f216;
|
||||
and.b32 %r28, %r27, 8388607;
|
||||
or.b32 %r29, %r28, 1065353216;
|
||||
mov.b32 %f217, %r29;
|
||||
shr.u32 %r30, %r27, 23;
|
||||
cvt.rn.f32.u32 %f218, %r30;
|
||||
add.f32 %f219, %f215, %f218;
|
||||
setp.gt.f32 %p38, %f217, 0f3FB504F3;
|
||||
mul.f32 %f220, %f217, 0f3F000000;
|
||||
add.f32 %f221, %f219, 0f3F800000;
|
||||
selp.f32 %f222, %f220, %f217, %p38;
|
||||
selp.f32 %f223, %f221, %f219, %p38;
|
||||
add.f32 %f224, %f222, 0fBF800000;
|
||||
add.f32 %f212, %f222, 0f3F800000;
|
||||
// inline asm
|
||||
rcp.approx.ftz.f32 %f211,%f212;
|
||||
// inline asm
|
||||
add.f32 %f225, %f224, %f224;
|
||||
mul.f32 %f226, %f211, %f225;
|
||||
mul.f32 %f227, %f226, %f226;
|
||||
fma.rn.f32 %f230, %f307, %f227, %f308;
|
||||
fma.rn.f32 %f232, %f230, %f227, %f309;
|
||||
mul.rn.f32 %f233, %f232, %f227;
|
||||
mul.rn.f32 %f234, %f233, %f226;
|
||||
sub.f32 %f235, %f224, %f226;
|
||||
neg.f32 %f236, %f226;
|
||||
add.f32 %f237, %f235, %f235;
|
||||
fma.rn.f32 %f238, %f236, %f224, %f237;
|
||||
mul.rn.f32 %f239, %f211, %f238;
|
||||
add.f32 %f240, %f234, %f226;
|
||||
sub.f32 %f241, %f226, %f240;
|
||||
add.f32 %f242, %f234, %f241;
|
||||
add.f32 %f243, %f239, %f242;
|
||||
add.f32 %f244, %f240, %f243;
|
||||
sub.f32 %f245, %f240, %f244;
|
||||
add.f32 %f246, %f243, %f245;
|
||||
mul.rn.f32 %f248, %f223, %f310;
|
||||
mul.rn.f32 %f250, %f223, %f311;
|
||||
add.f32 %f251, %f248, %f244;
|
||||
sub.f32 %f252, %f248, %f251;
|
||||
add.f32 %f253, %f244, %f252;
|
||||
add.f32 %f254, %f246, %f253;
|
||||
add.f32 %f255, %f250, %f254;
|
||||
add.f32 %f256, %f251, %f255;
|
||||
sub.f32 %f257, %f251, %f256;
|
||||
add.f32 %f258, %f255, %f257;
|
||||
mul.rn.f32 %f260, %f101, %f256;
|
||||
neg.f32 %f261, %f260;
|
||||
fma.rn.f32 %f262, %f101, %f256, %f261;
|
||||
fma.rn.f32 %f263, %f101, %f258, %f262;
|
||||
fma.rn.f32 %f265, %f312, %f256, %f263;
|
||||
add.rn.f32 %f266, %f260, %f265;
|
||||
neg.f32 %f267, %f266;
|
||||
add.rn.f32 %f268, %f260, %f267;
|
||||
add.rn.f32 %f269, %f268, %f265;
|
||||
mov.b32 %r31, %f266;
|
||||
setp.eq.s32 %p39, %r31, 1118925336;
|
||||
add.s32 %r32, %r31, -1;
|
||||
mov.b32 %f270, %r32;
|
||||
add.f32 %f271, %f269, 0f37000000;
|
||||
selp.f32 %f272, %f270, %f266, %p39;
|
||||
selp.f32 %f32, %f271, %f269, %p39;
|
||||
mul.f32 %f273, %f272, 0f3FB8AA3B;
|
||||
cvt.rzi.f32.f32 %f274, %f273;
|
||||
fma.rn.f32 %f276, %f274, %f313, %f272;
|
||||
fma.rn.f32 %f278, %f274, %f314, %f276;
|
||||
mul.f32 %f279, %f278, 0f3FB8AA3B;
|
||||
ex2.approx.ftz.f32 %f280, %f279;
|
||||
add.f32 %f281, %f274, 0f00000000;
|
||||
ex2.approx.f32 %f282, %f281;
|
||||
mul.f32 %f283, %f280, %f282;
|
||||
setp.lt.f32 %p40, %f272, 0fC2D20000;
|
||||
selp.f32 %f284, 0f00000000, %f283, %p40;
|
||||
setp.gt.f32 %p41, %f272, 0f42D20000;
|
||||
selp.f32 %f323, 0f7F800000, %f284, %p41;
|
||||
setp.eq.f32 %p42, %f323, 0f7F800000;
|
||||
@%p42 bra BB0_26;
|
||||
|
||||
fma.rn.f32 %f323, %f323, %f32, %f323;
|
||||
|
||||
BB0_26:
|
||||
setp.lt.f32 %p43, %f45, 0f00000000;
|
||||
and.pred %p3, %p43, %p12;
|
||||
mov.b32 %r33, %f323;
|
||||
xor.b32 %r34, %r33, -2147483648;
|
||||
mov.b32 %f285, %r34;
|
||||
selp.f32 %f325, %f285, %f323, %p3;
|
||||
setp.eq.f32 %p45, %f45, 0f00000000;
|
||||
@%p45 bra BB0_29;
|
||||
bra.uni BB0_27;
|
||||
|
||||
BB0_29:
|
||||
add.f32 %f288, %f45, %f45;
|
||||
selp.f32 %f325, %f288, 0f00000000, %p12;
|
||||
bra.uni BB0_30;
|
||||
|
||||
BB0_27:
|
||||
setp.geu.f32 %p46, %f45, 0f00000000;
|
||||
@%p46 bra BB0_30;
|
||||
|
||||
cvt.rzi.f32.f32 %f287, %f101;
|
||||
setp.neu.f32 %p47, %f287, 0f400CCCCD;
|
||||
selp.f32 %f325, 0f7FFFFFFF, %f325, %p47;
|
||||
|
||||
BB0_30:
|
||||
add.f32 %f289, %f31, 0f400CCCCD;
|
||||
mov.b32 %r35, %f289;
|
||||
setp.lt.s32 %p49, %r35, 2139095040;
|
||||
@%p49 bra BB0_35;
|
||||
|
||||
setp.gtu.f32 %p50, %f31, 0f7F800000;
|
||||
@%p50 bra BB0_34;
|
||||
bra.uni BB0_32;
|
||||
|
||||
BB0_34:
|
||||
add.f32 %f325, %f45, 0f400CCCCD;
|
||||
bra.uni BB0_35;
|
||||
|
||||
BB0_32:
|
||||
setp.neu.f32 %p51, %f31, 0f7F800000;
|
||||
@%p51 bra BB0_35;
|
||||
|
||||
selp.f32 %f325, 0fFF800000, 0f7F800000, %p3;
|
||||
|
||||
BB0_35:
|
||||
setp.eq.f32 %p52, %f45, 0f3F800000;
|
||||
selp.f32 %f290, 0f3F800000, %f325, %p52;
|
||||
cvt.sat.f32.f32 %f291, %f290;
|
||||
cvt.sat.f32.f32 %f292, %f17;
|
||||
mul.f32 %f293, %f292, 0f437F0000;
|
||||
cvt.rzi.u32.f32 %r36, %f293;
|
||||
cvt.sat.f32.f32 %f294, %f30;
|
||||
mul.f32 %f295, %f294, 0f437F0000;
|
||||
cvt.rzi.u32.f32 %r37, %f295;
|
||||
cvt.sat.f32.f32 %f296, %f291;
|
||||
mul.f32 %f297, %f296, 0f437F0000;
|
||||
cvt.rzi.u32.f32 %r38, %f297;
|
||||
ld.const.u64 %rd10, [cs+16];
|
||||
cvta.to.global.u64 %rd11, %rd10;
|
||||
shl.b64 %rd12, %rd1, 2;
|
||||
add.s64 %rd13, %rd11, %rd12;
|
||||
cvt.u16.u32 %rs5, %r38;
|
||||
cvt.u16.u32 %rs6, %r37;
|
||||
cvt.u16.u32 %rs7, %r36;
|
||||
mov.u16 %rs8, 255;
|
||||
st.global.v4.u8 [%rd13], {%rs7, %rs6, %rs5, %rs8};
|
||||
|
||||
BB0_36:
|
||||
ret;
|
||||
}
|
||||
|
||||
|
||||
8
Assets/Editor/x64/Bakery/denoiseFinishSH72.ptx.meta
Normal file
8
Assets/Editor/x64/Bakery/denoiseFinishSH72.ptx.meta
Normal file
@@ -0,0 +1,8 @@
|
||||
fileFormatVersion: 2
|
||||
guid: 941f8431b37da2447b84b09660fed6c7
|
||||
timeCreated: 1605379707
|
||||
licenseType: Store
|
||||
DefaultImporter:
|
||||
userData:
|
||||
assetBundleName:
|
||||
assetBundleVariant:
|
||||
167
Assets/Editor/x64/Bakery/denoiseFinishSH_OIDN.ptx
Normal file
167
Assets/Editor/x64/Bakery/denoiseFinishSH_OIDN.ptx
Normal file
@@ -0,0 +1,167 @@
|
||||
//
|
||||
// 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 1 .b8 output_buffer[1];
|
||||
.global .align 1 .b8 image2[1];
|
||||
.global .align 1 .b8 image3[1];
|
||||
.global .align 4 .u32 mode;
|
||||
.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_typeinfo4modeE[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 4 .b8 _ZN21rti_internal_typename4modeE[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_typeenum4modeE = 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 1 .b8 _ZN21rti_internal_semantic4modeE[1];
|
||||
.global .align 1 .b8 _ZN23rti_internal_annotation7pixelIDE[1];
|
||||
.global .align 1 .b8 _ZN23rti_internal_annotation10resolutionE[1];
|
||||
.global .align 1 .b8 _ZN23rti_internal_annotation4modeE[1];
|
||||
|
||||
.visible .entry _Z6oxMainv(
|
||||
|
||||
)
|
||||
{
|
||||
.reg .pred %p<2>;
|
||||
.reg .b16 %rs<12>;
|
||||
.reg .f32 %f<17>;
|
||||
.reg .b32 %r<37>;
|
||||
.reg .b64 %rd<43>;
|
||||
|
||||
|
||||
ld.global.v2.u32 {%r3, %r4}, [pixelID];
|
||||
cvt.u64.u32 %rd5, %r3;
|
||||
cvt.u64.u32 %rd6, %r4;
|
||||
mov.u64 %rd9, output_buffer;
|
||||
cvta.global.u64 %rd4, %rd9;
|
||||
mov.u32 %r1, 2;
|
||||
mov.u32 %r2, 12;
|
||||
mov.u64 %rd8, 0;
|
||||
// inline asm
|
||||
call (%rd3), _rt_buffer_get_64, (%rd4, %r1, %r2, %rd5, %rd6, %rd8, %rd8);
|
||||
// inline asm
|
||||
ld.f32 %f1, [%rd3];
|
||||
ld.f32 %f2, [%rd3+4];
|
||||
ld.f32 %f3, [%rd3+8];
|
||||
ld.global.u32 %r7, [mode];
|
||||
setp.gt.s32 %p1, %r7, 0;
|
||||
ld.global.v2.u32 {%r8, %r9}, [pixelID];
|
||||
cvt.u64.u32 %rd1, %r8;
|
||||
cvt.u64.u32 %rd2, %r9;
|
||||
@%p1 bra BB0_2;
|
||||
bra.uni BB0_1;
|
||||
|
||||
BB0_2:
|
||||
mov.u64 %rd41, image2;
|
||||
cvta.global.u64 %rd18, %rd41;
|
||||
mov.u32 %r19, 8;
|
||||
// inline asm
|
||||
call (%rd17), _rt_buffer_get_64, (%rd18, %r1, %r19, %rd1, %rd2, %rd8, %rd8);
|
||||
// inline asm
|
||||
ld.global.v2.u32 {%r22, %r23}, [pixelID];
|
||||
cvt.u64.u32 %rd25, %r22;
|
||||
cvt.u64.u32 %rd26, %r23;
|
||||
// inline asm
|
||||
call (%rd23), _rt_buffer_get_64, (%rd18, %r1, %r19, %rd25, %rd26, %rd8, %rd8);
|
||||
// inline asm
|
||||
ld.global.v2.u32 {%r26, %r27}, [pixelID];
|
||||
cvt.u64.u32 %rd31, %r26;
|
||||
cvt.u64.u32 %rd32, %r27;
|
||||
// inline asm
|
||||
call (%rd29), _rt_buffer_get_64, (%rd18, %r1, %r19, %rd31, %rd32, %rd8, %rd8);
|
||||
// inline asm
|
||||
ld.global.v2.u32 {%r30, %r31}, [pixelID];
|
||||
cvt.u64.u32 %rd37, %r30;
|
||||
cvt.u64.u32 %rd38, %r31;
|
||||
mov.u64 %rd42, image3;
|
||||
cvta.global.u64 %rd36, %rd42;
|
||||
mov.u32 %r21, 4;
|
||||
// inline asm
|
||||
call (%rd35), _rt_buffer_get_64, (%rd36, %r1, %r21, %rd37, %rd38, %rd8, %rd8);
|
||||
// inline asm
|
||||
cvt.sat.f32.f32 %f11, %f1;
|
||||
mul.f32 %f12, %f11, 0f437F0000;
|
||||
cvt.rzi.u32.f32 %r34, %f12;
|
||||
cvt.sat.f32.f32 %f13, %f2;
|
||||
mul.f32 %f14, %f13, 0f437F0000;
|
||||
cvt.rzi.u32.f32 %r35, %f14;
|
||||
cvt.sat.f32.f32 %f15, %f3;
|
||||
mul.f32 %f16, %f15, 0f437F0000;
|
||||
cvt.rzi.u32.f32 %r36, %f16;
|
||||
cvt.u16.u32 %rs8, %r36;
|
||||
cvt.u16.u32 %rs9, %r35;
|
||||
cvt.u16.u32 %rs10, %r34;
|
||||
mov.u16 %rs11, 255;
|
||||
st.v4.u8 [%rd35], {%rs10, %rs9, %rs8, %rs11};
|
||||
bra.uni BB0_3;
|
||||
|
||||
BB0_1:
|
||||
mov.u64 %rd16, image2;
|
||||
cvta.global.u64 %rd11, %rd16;
|
||||
mov.u32 %r13, 8;
|
||||
// inline asm
|
||||
call (%rd10), _rt_buffer_get_64, (%rd11, %r1, %r13, %rd1, %rd2, %rd8, %rd8);
|
||||
// inline asm
|
||||
mov.f32 %f7, 0f3F800000;
|
||||
// inline asm
|
||||
{ cvt.rn.f16.f32 %rs4, %f7;}
|
||||
|
||||
// inline asm
|
||||
// inline asm
|
||||
{ cvt.rn.f16.f32 %rs3, %f3;}
|
||||
|
||||
// inline asm
|
||||
// inline asm
|
||||
{ cvt.rn.f16.f32 %rs2, %f2;}
|
||||
|
||||
// inline asm
|
||||
// inline asm
|
||||
{ cvt.rn.f16.f32 %rs1, %f1;}
|
||||
|
||||
// inline asm
|
||||
st.v4.u16 [%rd10], {%rs1, %rs2, %rs3, %rs4};
|
||||
|
||||
BB0_3:
|
||||
ret;
|
||||
}
|
||||
|
||||
|
||||
8
Assets/Editor/x64/Bakery/denoiseFinishSH_OIDN.ptx.meta
Normal file
8
Assets/Editor/x64/Bakery/denoiseFinishSH_OIDN.ptx.meta
Normal file
@@ -0,0 +1,8 @@
|
||||
fileFormatVersion: 2
|
||||
guid: 0d06c9e9bf3986941abe247af43f8ae1
|
||||
timeCreated: 1599897795
|
||||
licenseType: Store
|
||||
DefaultImporter:
|
||||
userData:
|
||||
assetBundleName:
|
||||
assetBundleVariant:
|
||||
586
Assets/Editor/x64/Bakery/denoisePrepare.ptx
Normal file
586
Assets/Editor/x64/Bakery/denoisePrepare.ptx
Normal file
@@ -0,0 +1,586 @@
|
||||
//
|
||||
// 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 1 .b8 image[1];
|
||||
.global .align 1 .b8 input_buffer[1];
|
||||
.global .align 4 .u32 mode;
|
||||
.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_typeinfo4modeE[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 4 .b8 _ZN21rti_internal_typename4modeE[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_typeenum4modeE = 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 1 .b8 _ZN21rti_internal_semantic4modeE[1];
|
||||
.global .align 1 .b8 _ZN23rti_internal_annotation7pixelIDE[1];
|
||||
.global .align 1 .b8 _ZN23rti_internal_annotation10resolutionE[1];
|
||||
.global .align 1 .b8 _ZN23rti_internal_annotation4modeE[1];
|
||||
|
||||
.visible .entry _Z6oxMainv(
|
||||
|
||||
)
|
||||
{
|
||||
.reg .pred %p<54>;
|
||||
.reg .b16 %rs<4>;
|
||||
.reg .f32 %f<335>;
|
||||
.reg .b32 %r<55>;
|
||||
.reg .b64 %rd<29>;
|
||||
|
||||
|
||||
ld.global.v2.u32 {%r7, %r8}, [pixelID];
|
||||
cvt.u64.u32 %rd4, %r7;
|
||||
cvt.u64.u32 %rd5, %r8;
|
||||
mov.u64 %rd20, image;
|
||||
cvta.global.u64 %rd3, %rd20;
|
||||
mov.u32 %r5, 2;
|
||||
mov.u32 %r6, 8;
|
||||
mov.u64 %rd19, 0;
|
||||
// inline asm
|
||||
call (%rd2), _rt_buffer_get_64, (%rd3, %r5, %r6, %rd4, %rd5, %rd19, %rd19);
|
||||
// inline asm
|
||||
ld.u16 %rs1, [%rd2];
|
||||
// inline asm
|
||||
{ cvt.f32.f16 %f44, %rs1;}
|
||||
|
||||
// inline asm
|
||||
ld.global.v2.u32 {%r11, %r12}, [pixelID];
|
||||
cvt.u64.u32 %rd10, %r11;
|
||||
cvt.u64.u32 %rd11, %r12;
|
||||
// inline asm
|
||||
call (%rd8), _rt_buffer_get_64, (%rd3, %r5, %r6, %rd10, %rd11, %rd19, %rd19);
|
||||
// inline asm
|
||||
ld.u16 %rs2, [%rd8+2];
|
||||
// inline asm
|
||||
{ cvt.f32.f16 %f45, %rs2;}
|
||||
|
||||
// inline asm
|
||||
ld.global.v2.u32 {%r15, %r16}, [pixelID];
|
||||
cvt.u64.u32 %rd16, %r15;
|
||||
cvt.u64.u32 %rd17, %r16;
|
||||
// inline asm
|
||||
call (%rd14), _rt_buffer_get_64, (%rd3, %r5, %r6, %rd16, %rd17, %rd19, %rd19);
|
||||
// inline asm
|
||||
ld.u16 %rs3, [%rd14+4];
|
||||
// inline asm
|
||||
{ cvt.f32.f16 %f46, %rs3;}
|
||||
|
||||
// inline asm
|
||||
max.f32 %f49, %f44, %f45;
|
||||
max.f32 %f50, %f49, %f46;
|
||||
add.f32 %f51, %f50, 0f3F800000;
|
||||
rcp.rn.f32 %f52, %f51;
|
||||
mul.f32 %f53, %f44, %f52;
|
||||
mul.f32 %f1, %f45, %f52;
|
||||
mul.f32 %f54, %f46, %f52;
|
||||
ld.global.u32 %r19, [mode];
|
||||
setp.eq.s32 %p4, %r19, 1;
|
||||
selp.f32 %f2, %f54, %f53, %p4;
|
||||
selp.f32 %f3, %f53, %f54, %p4;
|
||||
abs.f32 %f5, %f2;
|
||||
setp.lt.f32 %p5, %f5, 0f00800000;
|
||||
mul.f32 %f58, %f5, 0f4B800000;
|
||||
selp.f32 %f59, 0fC3170000, 0fC2FE0000, %p5;
|
||||
selp.f32 %f60, %f58, %f5, %p5;
|
||||
mov.b32 %r20, %f60;
|
||||
and.b32 %r21, %r20, 8388607;
|
||||
or.b32 %r22, %r21, 1065353216;
|
||||
mov.b32 %f61, %r22;
|
||||
shr.u32 %r23, %r20, 23;
|
||||
cvt.rn.f32.u32 %f62, %r23;
|
||||
add.f32 %f63, %f59, %f62;
|
||||
setp.gt.f32 %p6, %f61, 0f3FB504F3;
|
||||
mul.f32 %f64, %f61, 0f3F000000;
|
||||
add.f32 %f65, %f63, 0f3F800000;
|
||||
selp.f32 %f66, %f64, %f61, %p6;
|
||||
selp.f32 %f67, %f65, %f63, %p6;
|
||||
add.f32 %f68, %f66, 0fBF800000;
|
||||
add.f32 %f48, %f66, 0f3F800000;
|
||||
// inline asm
|
||||
rcp.approx.ftz.f32 %f47,%f48;
|
||||
// inline asm
|
||||
add.f32 %f69, %f68, %f68;
|
||||
mul.f32 %f70, %f47, %f69;
|
||||
mul.f32 %f71, %f70, %f70;
|
||||
mov.f32 %f72, 0f3C4CAF63;
|
||||
mov.f32 %f73, 0f3B18F0FE;
|
||||
fma.rn.f32 %f74, %f73, %f71, %f72;
|
||||
mov.f32 %f75, 0f3DAAAABD;
|
||||
fma.rn.f32 %f76, %f74, %f71, %f75;
|
||||
mul.rn.f32 %f77, %f76, %f71;
|
||||
mul.rn.f32 %f78, %f77, %f70;
|
||||
sub.f32 %f79, %f68, %f70;
|
||||
neg.f32 %f80, %f70;
|
||||
add.f32 %f81, %f79, %f79;
|
||||
fma.rn.f32 %f82, %f80, %f68, %f81;
|
||||
mul.rn.f32 %f83, %f47, %f82;
|
||||
add.f32 %f84, %f78, %f70;
|
||||
sub.f32 %f85, %f70, %f84;
|
||||
add.f32 %f86, %f78, %f85;
|
||||
add.f32 %f87, %f83, %f86;
|
||||
add.f32 %f88, %f84, %f87;
|
||||
sub.f32 %f89, %f84, %f88;
|
||||
add.f32 %f90, %f87, %f89;
|
||||
mov.f32 %f91, 0f3F317200;
|
||||
mul.rn.f32 %f92, %f67, %f91;
|
||||
mov.f32 %f93, 0f35BFBE8E;
|
||||
mul.rn.f32 %f94, %f67, %f93;
|
||||
add.f32 %f95, %f92, %f88;
|
||||
sub.f32 %f96, %f92, %f95;
|
||||
add.f32 %f97, %f88, %f96;
|
||||
add.f32 %f98, %f90, %f97;
|
||||
add.f32 %f99, %f94, %f98;
|
||||
add.f32 %f100, %f95, %f99;
|
||||
sub.f32 %f101, %f95, %f100;
|
||||
add.f32 %f102, %f99, %f101;
|
||||
mov.f32 %f103, 0f3EE8BA2E;
|
||||
mul.rn.f32 %f104, %f103, %f100;
|
||||
neg.f32 %f105, %f104;
|
||||
fma.rn.f32 %f106, %f103, %f100, %f105;
|
||||
fma.rn.f32 %f107, %f103, %f102, %f106;
|
||||
mov.f32 %f108, 0f00000000;
|
||||
fma.rn.f32 %f109, %f108, %f100, %f107;
|
||||
add.rn.f32 %f110, %f104, %f109;
|
||||
neg.f32 %f111, %f110;
|
||||
add.rn.f32 %f112, %f104, %f111;
|
||||
add.rn.f32 %f113, %f112, %f109;
|
||||
mov.b32 %r24, %f110;
|
||||
setp.eq.s32 %p7, %r24, 1118925336;
|
||||
add.s32 %r25, %r24, -1;
|
||||
mov.b32 %f114, %r25;
|
||||
add.f32 %f115, %f113, 0f37000000;
|
||||
selp.f32 %f116, %f114, %f110, %p7;
|
||||
selp.f32 %f6, %f115, %f113, %p7;
|
||||
mul.f32 %f117, %f116, 0f3FB8AA3B;
|
||||
cvt.rzi.f32.f32 %f118, %f117;
|
||||
mov.f32 %f119, 0fBF317200;
|
||||
fma.rn.f32 %f120, %f118, %f119, %f116;
|
||||
mov.f32 %f121, 0fB5BFBE8E;
|
||||
fma.rn.f32 %f122, %f118, %f121, %f120;
|
||||
mul.f32 %f123, %f122, 0f3FB8AA3B;
|
||||
ex2.approx.ftz.f32 %f124, %f123;
|
||||
add.f32 %f125, %f118, 0f00000000;
|
||||
ex2.approx.f32 %f126, %f125;
|
||||
mul.f32 %f127, %f124, %f126;
|
||||
setp.lt.f32 %p8, %f116, 0fC2D20000;
|
||||
selp.f32 %f128, 0f00000000, %f127, %p8;
|
||||
setp.gt.f32 %p9, %f116, 0f42D20000;
|
||||
selp.f32 %f326, 0f7F800000, %f128, %p9;
|
||||
setp.eq.f32 %p10, %f326, 0f7F800000;
|
||||
@%p10 bra BB0_2;
|
||||
|
||||
fma.rn.f32 %f326, %f326, %f6, %f326;
|
||||
|
||||
BB0_2:
|
||||
mov.f32 %f325, 0f3E68BA2E;
|
||||
cvt.rzi.f32.f32 %f324, %f325;
|
||||
fma.rn.f32 %f323, %f324, 0fC0000000, 0f3EE8BA2E;
|
||||
abs.f32 %f322, %f323;
|
||||
setp.lt.f32 %p11, %f2, 0f00000000;
|
||||
setp.eq.f32 %p12, %f322, 0f3F800000;
|
||||
and.pred %p1, %p11, %p12;
|
||||
mov.b32 %r26, %f326;
|
||||
xor.b32 %r27, %r26, -2147483648;
|
||||
mov.b32 %f129, %r27;
|
||||
selp.f32 %f328, %f129, %f326, %p1;
|
||||
setp.eq.f32 %p13, %f2, 0f00000000;
|
||||
@%p13 bra BB0_5;
|
||||
bra.uni BB0_3;
|
||||
|
||||
BB0_5:
|
||||
add.f32 %f132, %f2, %f2;
|
||||
selp.f32 %f328, %f132, 0f00000000, %p12;
|
||||
bra.uni BB0_6;
|
||||
|
||||
BB0_3:
|
||||
setp.geu.f32 %p14, %f2, 0f00000000;
|
||||
@%p14 bra BB0_6;
|
||||
|
||||
mov.f32 %f321, 0f3EE8BA2E;
|
||||
cvt.rzi.f32.f32 %f131, %f321;
|
||||
setp.neu.f32 %p15, %f131, 0f3EE8BA2E;
|
||||
selp.f32 %f328, 0f7FFFFFFF, %f328, %p15;
|
||||
|
||||
BB0_6:
|
||||
abs.f32 %f298, %f2;
|
||||
add.f32 %f133, %f298, 0f3EE8BA2E;
|
||||
mov.b32 %r28, %f133;
|
||||
setp.lt.s32 %p17, %r28, 2139095040;
|
||||
@%p17 bra BB0_11;
|
||||
|
||||
abs.f32 %f319, %f2;
|
||||
setp.gtu.f32 %p18, %f319, 0f7F800000;
|
||||
@%p18 bra BB0_10;
|
||||
bra.uni BB0_8;
|
||||
|
||||
BB0_10:
|
||||
add.f32 %f328, %f2, 0f3EE8BA2E;
|
||||
bra.uni BB0_11;
|
||||
|
||||
BB0_8:
|
||||
abs.f32 %f320, %f2;
|
||||
setp.neu.f32 %p19, %f320, 0f7F800000;
|
||||
@%p19 bra BB0_11;
|
||||
|
||||
selp.f32 %f328, 0fFF800000, 0f7F800000, %p1;
|
||||
|
||||
BB0_11:
|
||||
mov.f32 %f307, 0fB5BFBE8E;
|
||||
mov.f32 %f306, 0fBF317200;
|
||||
mov.f32 %f305, 0f00000000;
|
||||
mov.f32 %f304, 0f35BFBE8E;
|
||||
mov.f32 %f303, 0f3F317200;
|
||||
mov.f32 %f302, 0f3DAAAABD;
|
||||
mov.f32 %f301, 0f3C4CAF63;
|
||||
mov.f32 %f300, 0f3B18F0FE;
|
||||
mov.f32 %f299, 0f3EE8BA2E;
|
||||
setp.eq.f32 %p20, %f2, 0f3F800000;
|
||||
selp.f32 %f136, 0f3F800000, %f328, %p20;
|
||||
cvt.sat.f32.f32 %f17, %f136;
|
||||
abs.f32 %f18, %f1;
|
||||
setp.lt.f32 %p21, %f18, 0f00800000;
|
||||
mul.f32 %f137, %f18, 0f4B800000;
|
||||
selp.f32 %f138, 0fC3170000, 0fC2FE0000, %p21;
|
||||
selp.f32 %f139, %f137, %f18, %p21;
|
||||
mov.b32 %r29, %f139;
|
||||
and.b32 %r30, %r29, 8388607;
|
||||
or.b32 %r31, %r30, 1065353216;
|
||||
mov.b32 %f140, %r31;
|
||||
shr.u32 %r32, %r29, 23;
|
||||
cvt.rn.f32.u32 %f141, %r32;
|
||||
add.f32 %f142, %f138, %f141;
|
||||
setp.gt.f32 %p22, %f140, 0f3FB504F3;
|
||||
mul.f32 %f143, %f140, 0f3F000000;
|
||||
add.f32 %f144, %f142, 0f3F800000;
|
||||
selp.f32 %f145, %f143, %f140, %p22;
|
||||
selp.f32 %f146, %f144, %f142, %p22;
|
||||
add.f32 %f147, %f145, 0fBF800000;
|
||||
add.f32 %f135, %f145, 0f3F800000;
|
||||
// inline asm
|
||||
rcp.approx.ftz.f32 %f134,%f135;
|
||||
// inline asm
|
||||
add.f32 %f148, %f147, %f147;
|
||||
mul.f32 %f149, %f134, %f148;
|
||||
mul.f32 %f150, %f149, %f149;
|
||||
fma.rn.f32 %f153, %f300, %f150, %f301;
|
||||
fma.rn.f32 %f155, %f153, %f150, %f302;
|
||||
mul.rn.f32 %f156, %f155, %f150;
|
||||
mul.rn.f32 %f157, %f156, %f149;
|
||||
sub.f32 %f158, %f147, %f149;
|
||||
neg.f32 %f159, %f149;
|
||||
add.f32 %f160, %f158, %f158;
|
||||
fma.rn.f32 %f161, %f159, %f147, %f160;
|
||||
mul.rn.f32 %f162, %f134, %f161;
|
||||
add.f32 %f163, %f157, %f149;
|
||||
sub.f32 %f164, %f149, %f163;
|
||||
add.f32 %f165, %f157, %f164;
|
||||
add.f32 %f166, %f162, %f165;
|
||||
add.f32 %f167, %f163, %f166;
|
||||
sub.f32 %f168, %f163, %f167;
|
||||
add.f32 %f169, %f166, %f168;
|
||||
mul.rn.f32 %f171, %f146, %f303;
|
||||
mul.rn.f32 %f173, %f146, %f304;
|
||||
add.f32 %f174, %f171, %f167;
|
||||
sub.f32 %f175, %f171, %f174;
|
||||
add.f32 %f176, %f167, %f175;
|
||||
add.f32 %f177, %f169, %f176;
|
||||
add.f32 %f178, %f173, %f177;
|
||||
add.f32 %f179, %f174, %f178;
|
||||
sub.f32 %f180, %f174, %f179;
|
||||
add.f32 %f181, %f178, %f180;
|
||||
mul.rn.f32 %f183, %f299, %f179;
|
||||
neg.f32 %f184, %f183;
|
||||
fma.rn.f32 %f185, %f299, %f179, %f184;
|
||||
fma.rn.f32 %f186, %f299, %f181, %f185;
|
||||
fma.rn.f32 %f188, %f305, %f179, %f186;
|
||||
add.rn.f32 %f189, %f183, %f188;
|
||||
neg.f32 %f190, %f189;
|
||||
add.rn.f32 %f191, %f183, %f190;
|
||||
add.rn.f32 %f192, %f191, %f188;
|
||||
mov.b32 %r33, %f189;
|
||||
setp.eq.s32 %p23, %r33, 1118925336;
|
||||
add.s32 %r34, %r33, -1;
|
||||
mov.b32 %f193, %r34;
|
||||
add.f32 %f194, %f192, 0f37000000;
|
||||
selp.f32 %f195, %f193, %f189, %p23;
|
||||
selp.f32 %f19, %f194, %f192, %p23;
|
||||
mul.f32 %f196, %f195, 0f3FB8AA3B;
|
||||
cvt.rzi.f32.f32 %f197, %f196;
|
||||
fma.rn.f32 %f199, %f197, %f306, %f195;
|
||||
fma.rn.f32 %f201, %f197, %f307, %f199;
|
||||
mul.f32 %f202, %f201, 0f3FB8AA3B;
|
||||
ex2.approx.ftz.f32 %f203, %f202;
|
||||
add.f32 %f204, %f197, 0f00000000;
|
||||
ex2.approx.f32 %f205, %f204;
|
||||
mul.f32 %f206, %f203, %f205;
|
||||
setp.lt.f32 %p24, %f195, 0fC2D20000;
|
||||
selp.f32 %f207, 0f00000000, %f206, %p24;
|
||||
setp.gt.f32 %p25, %f195, 0f42D20000;
|
||||
selp.f32 %f329, 0f7F800000, %f207, %p25;
|
||||
setp.eq.f32 %p26, %f329, 0f7F800000;
|
||||
@%p26 bra BB0_13;
|
||||
|
||||
fma.rn.f32 %f329, %f329, %f19, %f329;
|
||||
|
||||
BB0_13:
|
||||
setp.lt.f32 %p27, %f1, 0f00000000;
|
||||
and.pred %p2, %p27, %p12;
|
||||
mov.b32 %r35, %f329;
|
||||
xor.b32 %r36, %r35, -2147483648;
|
||||
mov.b32 %f208, %r36;
|
||||
selp.f32 %f331, %f208, %f329, %p2;
|
||||
setp.eq.f32 %p29, %f1, 0f00000000;
|
||||
@%p29 bra BB0_16;
|
||||
bra.uni BB0_14;
|
||||
|
||||
BB0_16:
|
||||
add.f32 %f211, %f1, %f1;
|
||||
selp.f32 %f331, %f211, 0f00000000, %p12;
|
||||
bra.uni BB0_17;
|
||||
|
||||
BB0_14:
|
||||
setp.geu.f32 %p30, %f1, 0f00000000;
|
||||
@%p30 bra BB0_17;
|
||||
|
||||
mov.f32 %f318, 0f3EE8BA2E;
|
||||
cvt.rzi.f32.f32 %f210, %f318;
|
||||
setp.neu.f32 %p31, %f210, 0f3EE8BA2E;
|
||||
selp.f32 %f331, 0f7FFFFFFF, %f331, %p31;
|
||||
|
||||
BB0_17:
|
||||
add.f32 %f212, %f18, 0f3EE8BA2E;
|
||||
mov.b32 %r37, %f212;
|
||||
setp.lt.s32 %p33, %r37, 2139095040;
|
||||
@%p33 bra BB0_22;
|
||||
|
||||
setp.gtu.f32 %p34, %f18, 0f7F800000;
|
||||
@%p34 bra BB0_21;
|
||||
bra.uni BB0_19;
|
||||
|
||||
BB0_21:
|
||||
add.f32 %f331, %f1, 0f3EE8BA2E;
|
||||
bra.uni BB0_22;
|
||||
|
||||
BB0_19:
|
||||
setp.neu.f32 %p35, %f18, 0f7F800000;
|
||||
@%p35 bra BB0_22;
|
||||
|
||||
selp.f32 %f331, 0fFF800000, 0f7F800000, %p2;
|
||||
|
||||
BB0_22:
|
||||
mov.f32 %f316, 0fB5BFBE8E;
|
||||
mov.f32 %f315, 0fBF317200;
|
||||
mov.f32 %f314, 0f00000000;
|
||||
mov.f32 %f313, 0f35BFBE8E;
|
||||
mov.f32 %f312, 0f3F317200;
|
||||
mov.f32 %f311, 0f3DAAAABD;
|
||||
mov.f32 %f310, 0f3C4CAF63;
|
||||
mov.f32 %f309, 0f3B18F0FE;
|
||||
mov.f32 %f308, 0f3EE8BA2E;
|
||||
setp.eq.f32 %p36, %f1, 0f3F800000;
|
||||
selp.f32 %f215, 0f3F800000, %f331, %p36;
|
||||
cvt.sat.f32.f32 %f30, %f215;
|
||||
abs.f32 %f31, %f3;
|
||||
setp.lt.f32 %p37, %f31, 0f00800000;
|
||||
mul.f32 %f216, %f31, 0f4B800000;
|
||||
selp.f32 %f217, 0fC3170000, 0fC2FE0000, %p37;
|
||||
selp.f32 %f218, %f216, %f31, %p37;
|
||||
mov.b32 %r38, %f218;
|
||||
and.b32 %r39, %r38, 8388607;
|
||||
or.b32 %r40, %r39, 1065353216;
|
||||
mov.b32 %f219, %r40;
|
||||
shr.u32 %r41, %r38, 23;
|
||||
cvt.rn.f32.u32 %f220, %r41;
|
||||
add.f32 %f221, %f217, %f220;
|
||||
setp.gt.f32 %p38, %f219, 0f3FB504F3;
|
||||
mul.f32 %f222, %f219, 0f3F000000;
|
||||
add.f32 %f223, %f221, 0f3F800000;
|
||||
selp.f32 %f224, %f222, %f219, %p38;
|
||||
selp.f32 %f225, %f223, %f221, %p38;
|
||||
add.f32 %f226, %f224, 0fBF800000;
|
||||
add.f32 %f214, %f224, 0f3F800000;
|
||||
// inline asm
|
||||
rcp.approx.ftz.f32 %f213,%f214;
|
||||
// inline asm
|
||||
add.f32 %f227, %f226, %f226;
|
||||
mul.f32 %f228, %f213, %f227;
|
||||
mul.f32 %f229, %f228, %f228;
|
||||
fma.rn.f32 %f232, %f309, %f229, %f310;
|
||||
fma.rn.f32 %f234, %f232, %f229, %f311;
|
||||
mul.rn.f32 %f235, %f234, %f229;
|
||||
mul.rn.f32 %f236, %f235, %f228;
|
||||
sub.f32 %f237, %f226, %f228;
|
||||
neg.f32 %f238, %f228;
|
||||
add.f32 %f239, %f237, %f237;
|
||||
fma.rn.f32 %f240, %f238, %f226, %f239;
|
||||
mul.rn.f32 %f241, %f213, %f240;
|
||||
add.f32 %f242, %f236, %f228;
|
||||
sub.f32 %f243, %f228, %f242;
|
||||
add.f32 %f244, %f236, %f243;
|
||||
add.f32 %f245, %f241, %f244;
|
||||
add.f32 %f246, %f242, %f245;
|
||||
sub.f32 %f247, %f242, %f246;
|
||||
add.f32 %f248, %f245, %f247;
|
||||
mul.rn.f32 %f250, %f225, %f312;
|
||||
mul.rn.f32 %f252, %f225, %f313;
|
||||
add.f32 %f253, %f250, %f246;
|
||||
sub.f32 %f254, %f250, %f253;
|
||||
add.f32 %f255, %f246, %f254;
|
||||
add.f32 %f256, %f248, %f255;
|
||||
add.f32 %f257, %f252, %f256;
|
||||
add.f32 %f258, %f253, %f257;
|
||||
sub.f32 %f259, %f253, %f258;
|
||||
add.f32 %f260, %f257, %f259;
|
||||
mul.rn.f32 %f262, %f308, %f258;
|
||||
neg.f32 %f263, %f262;
|
||||
fma.rn.f32 %f264, %f308, %f258, %f263;
|
||||
fma.rn.f32 %f265, %f308, %f260, %f264;
|
||||
fma.rn.f32 %f267, %f314, %f258, %f265;
|
||||
add.rn.f32 %f268, %f262, %f267;
|
||||
neg.f32 %f269, %f268;
|
||||
add.rn.f32 %f270, %f262, %f269;
|
||||
add.rn.f32 %f271, %f270, %f267;
|
||||
mov.b32 %r42, %f268;
|
||||
setp.eq.s32 %p39, %r42, 1118925336;
|
||||
add.s32 %r43, %r42, -1;
|
||||
mov.b32 %f272, %r43;
|
||||
add.f32 %f273, %f271, 0f37000000;
|
||||
selp.f32 %f274, %f272, %f268, %p39;
|
||||
selp.f32 %f32, %f273, %f271, %p39;
|
||||
mul.f32 %f275, %f274, 0f3FB8AA3B;
|
||||
cvt.rzi.f32.f32 %f276, %f275;
|
||||
fma.rn.f32 %f278, %f276, %f315, %f274;
|
||||
fma.rn.f32 %f280, %f276, %f316, %f278;
|
||||
mul.f32 %f281, %f280, 0f3FB8AA3B;
|
||||
ex2.approx.ftz.f32 %f282, %f281;
|
||||
add.f32 %f283, %f276, 0f00000000;
|
||||
ex2.approx.f32 %f284, %f283;
|
||||
mul.f32 %f285, %f282, %f284;
|
||||
setp.lt.f32 %p40, %f274, 0fC2D20000;
|
||||
selp.f32 %f286, 0f00000000, %f285, %p40;
|
||||
setp.gt.f32 %p41, %f274, 0f42D20000;
|
||||
selp.f32 %f332, 0f7F800000, %f286, %p41;
|
||||
setp.eq.f32 %p42, %f332, 0f7F800000;
|
||||
@%p42 bra BB0_24;
|
||||
|
||||
fma.rn.f32 %f332, %f332, %f32, %f332;
|
||||
|
||||
BB0_24:
|
||||
setp.lt.f32 %p43, %f3, 0f00000000;
|
||||
and.pred %p3, %p43, %p12;
|
||||
mov.b32 %r44, %f332;
|
||||
xor.b32 %r45, %r44, -2147483648;
|
||||
mov.b32 %f287, %r45;
|
||||
selp.f32 %f334, %f287, %f332, %p3;
|
||||
setp.eq.f32 %p45, %f3, 0f00000000;
|
||||
@%p45 bra BB0_27;
|
||||
bra.uni BB0_25;
|
||||
|
||||
BB0_27:
|
||||
add.f32 %f290, %f3, %f3;
|
||||
selp.f32 %f334, %f290, 0f00000000, %p12;
|
||||
bra.uni BB0_28;
|
||||
|
||||
BB0_25:
|
||||
setp.geu.f32 %p46, %f3, 0f00000000;
|
||||
@%p46 bra BB0_28;
|
||||
|
||||
mov.f32 %f317, 0f3EE8BA2E;
|
||||
cvt.rzi.f32.f32 %f289, %f317;
|
||||
setp.neu.f32 %p47, %f289, 0f3EE8BA2E;
|
||||
selp.f32 %f334, 0f7FFFFFFF, %f334, %p47;
|
||||
|
||||
BB0_28:
|
||||
add.f32 %f291, %f31, 0f3EE8BA2E;
|
||||
mov.b32 %r46, %f291;
|
||||
setp.lt.s32 %p49, %r46, 2139095040;
|
||||
@%p49 bra BB0_33;
|
||||
|
||||
setp.gtu.f32 %p50, %f31, 0f7F800000;
|
||||
@%p50 bra BB0_32;
|
||||
bra.uni BB0_30;
|
||||
|
||||
BB0_32:
|
||||
add.f32 %f334, %f3, 0f3EE8BA2E;
|
||||
bra.uni BB0_33;
|
||||
|
||||
BB0_30:
|
||||
setp.neu.f32 %p51, %f31, 0f7F800000;
|
||||
@%p51 bra BB0_33;
|
||||
|
||||
selp.f32 %f334, 0fFF800000, 0f7F800000, %p3;
|
||||
|
||||
BB0_33:
|
||||
mov.u64 %rd28, 0;
|
||||
mov.u32 %r54, 2;
|
||||
setp.eq.f32 %p52, %f3, 0f3F800000;
|
||||
selp.f32 %f292, 0f3F800000, %f334, %p52;
|
||||
cvt.sat.f32.f32 %f43, %f292;
|
||||
ld.global.v2.u32 {%r49, %r50}, [pixelID];
|
||||
cvt.u64.u32 %rd23, %r49;
|
||||
cvt.u64.u32 %rd24, %r50;
|
||||
mov.u64 %rd27, input_buffer;
|
||||
cvta.global.u64 %rd22, %rd27;
|
||||
mov.u32 %r48, 16;
|
||||
// inline asm
|
||||
call (%rd21), _rt_buffer_get_64, (%rd22, %r54, %r48, %rd23, %rd24, %rd28, %rd28);
|
||||
// inline asm
|
||||
ld.global.u32 %r53, [mode];
|
||||
setp.eq.s32 %p53, %r53, 0;
|
||||
@%p53 bra BB0_35;
|
||||
|
||||
mov.f32 %f293, 0f3F800000;
|
||||
sub.f32 %f294, %f293, %f43;
|
||||
sub.f32 %f295, %f293, %f30;
|
||||
sub.f32 %f296, %f293, %f17;
|
||||
st.v4.f32 [%rd21], {%f296, %f295, %f294, %f293};
|
||||
bra.uni BB0_36;
|
||||
|
||||
BB0_35:
|
||||
mov.f32 %f297, 0f3F800000;
|
||||
st.v4.f32 [%rd21], {%f17, %f30, %f43, %f297};
|
||||
|
||||
BB0_36:
|
||||
ret;
|
||||
}
|
||||
|
||||
|
||||
8
Assets/Editor/x64/Bakery/denoisePrepare.ptx.meta
Normal file
8
Assets/Editor/x64/Bakery/denoisePrepare.ptx.meta
Normal file
@@ -0,0 +1,8 @@
|
||||
fileFormatVersion: 2
|
||||
guid: 3c975724414595b499496e69340c5672
|
||||
timeCreated: 1526332011
|
||||
licenseType: Store
|
||||
DefaultImporter:
|
||||
userData:
|
||||
assetBundleName:
|
||||
assetBundleVariant:
|
||||
516
Assets/Editor/x64/Bakery/denoisePrepare72.ptx
Normal file
516
Assets/Editor/x64/Bakery/denoisePrepare72.ptx
Normal file
@@ -0,0 +1,516 @@
|
||||
//
|
||||
// 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 __raygen__oxMain
|
||||
.const .align 8 .b8 cs[32];
|
||||
|
||||
.visible .entry __raygen__oxMain(
|
||||
|
||||
)
|
||||
{
|
||||
.reg .pred %p<54>;
|
||||
.reg .b16 %rs<8>;
|
||||
.reg .f32 %f<337>;
|
||||
.reg .b32 %r<36>;
|
||||
.reg .b64 %rd<10>;
|
||||
|
||||
|
||||
// inline asm
|
||||
call (%r2), _optix_get_launch_index_x, ();
|
||||
// inline asm
|
||||
// inline asm
|
||||
call (%r3), _optix_get_launch_index_y, ();
|
||||
// inline asm
|
||||
ld.const.u64 %rd2, [cs];
|
||||
cvta.to.global.u64 %rd3, %rd2;
|
||||
ld.const.v2.u32 {%r5, %r6}, [cs+24];
|
||||
mad.lo.s32 %r8, %r5, %r3, %r2;
|
||||
cvt.u64.u32 %rd1, %r8;
|
||||
mul.wide.u32 %rd4, %r8, 8;
|
||||
add.s64 %rd5, %rd3, %rd4;
|
||||
ld.global.v4.u16 {%rs4, %rs5, %rs6, %rs7}, [%rd5];
|
||||
// inline asm
|
||||
{ cvt.f32.f16 %f43, %rs4;}
|
||||
|
||||
// inline asm
|
||||
// inline asm
|
||||
{ cvt.f32.f16 %f44, %rs5;}
|
||||
|
||||
// inline asm
|
||||
// inline asm
|
||||
{ cvt.f32.f16 %f45, %rs6;}
|
||||
|
||||
// inline asm
|
||||
max.f32 %f48, %f43, %f44;
|
||||
max.f32 %f49, %f48, %f45;
|
||||
add.f32 %f50, %f49, 0f3F800000;
|
||||
rcp.rn.f32 %f51, %f50;
|
||||
mul.f32 %f52, %f43, %f51;
|
||||
mul.f32 %f1, %f44, %f51;
|
||||
mul.f32 %f53, %f45, %f51;
|
||||
setp.eq.s32 %p4, %r6, 1;
|
||||
selp.f32 %f2, %f52, %f53, %p4;
|
||||
selp.f32 %f3, %f53, %f52, %p4;
|
||||
abs.f32 %f5, %f3;
|
||||
setp.lt.f32 %p5, %f5, 0f00800000;
|
||||
mul.f32 %f57, %f5, 0f4B800000;
|
||||
selp.f32 %f58, 0fC3170000, 0fC2FE0000, %p5;
|
||||
selp.f32 %f59, %f57, %f5, %p5;
|
||||
mov.b32 %r9, %f59;
|
||||
and.b32 %r10, %r9, 8388607;
|
||||
or.b32 %r11, %r10, 1065353216;
|
||||
mov.b32 %f60, %r11;
|
||||
shr.u32 %r12, %r9, 23;
|
||||
cvt.rn.f32.u32 %f61, %r12;
|
||||
add.f32 %f62, %f58, %f61;
|
||||
setp.gt.f32 %p6, %f60, 0f3FB504F3;
|
||||
mul.f32 %f63, %f60, 0f3F000000;
|
||||
add.f32 %f64, %f62, 0f3F800000;
|
||||
selp.f32 %f65, %f63, %f60, %p6;
|
||||
selp.f32 %f66, %f64, %f62, %p6;
|
||||
add.f32 %f67, %f65, 0fBF800000;
|
||||
add.f32 %f47, %f65, 0f3F800000;
|
||||
// inline asm
|
||||
rcp.approx.ftz.f32 %f46,%f47;
|
||||
// inline asm
|
||||
add.f32 %f68, %f67, %f67;
|
||||
mul.f32 %f69, %f46, %f68;
|
||||
mul.f32 %f70, %f69, %f69;
|
||||
mov.f32 %f71, 0f3C4CAF63;
|
||||
mov.f32 %f72, 0f3B18F0FE;
|
||||
fma.rn.f32 %f73, %f72, %f70, %f71;
|
||||
mov.f32 %f74, 0f3DAAAABD;
|
||||
fma.rn.f32 %f75, %f73, %f70, %f74;
|
||||
mul.rn.f32 %f76, %f75, %f70;
|
||||
mul.rn.f32 %f77, %f76, %f69;
|
||||
sub.f32 %f78, %f67, %f69;
|
||||
neg.f32 %f79, %f69;
|
||||
add.f32 %f80, %f78, %f78;
|
||||
fma.rn.f32 %f81, %f79, %f67, %f80;
|
||||
mul.rn.f32 %f82, %f46, %f81;
|
||||
add.f32 %f83, %f77, %f69;
|
||||
sub.f32 %f84, %f69, %f83;
|
||||
add.f32 %f85, %f77, %f84;
|
||||
add.f32 %f86, %f82, %f85;
|
||||
add.f32 %f87, %f83, %f86;
|
||||
sub.f32 %f88, %f83, %f87;
|
||||
add.f32 %f89, %f86, %f88;
|
||||
mov.f32 %f90, 0f3F317200;
|
||||
mul.rn.f32 %f91, %f66, %f90;
|
||||
mov.f32 %f92, 0f35BFBE8E;
|
||||
mul.rn.f32 %f93, %f66, %f92;
|
||||
add.f32 %f94, %f91, %f87;
|
||||
sub.f32 %f95, %f91, %f94;
|
||||
add.f32 %f96, %f87, %f95;
|
||||
add.f32 %f97, %f89, %f96;
|
||||
add.f32 %f98, %f93, %f97;
|
||||
add.f32 %f99, %f94, %f98;
|
||||
sub.f32 %f100, %f94, %f99;
|
||||
add.f32 %f101, %f98, %f100;
|
||||
mov.f32 %f102, 0f3EE8BA2E;
|
||||
mul.rn.f32 %f103, %f102, %f99;
|
||||
neg.f32 %f104, %f103;
|
||||
fma.rn.f32 %f105, %f102, %f99, %f104;
|
||||
fma.rn.f32 %f106, %f102, %f101, %f105;
|
||||
mov.f32 %f107, 0f00000000;
|
||||
fma.rn.f32 %f108, %f107, %f99, %f106;
|
||||
add.rn.f32 %f109, %f103, %f108;
|
||||
neg.f32 %f110, %f109;
|
||||
add.rn.f32 %f111, %f103, %f110;
|
||||
add.rn.f32 %f112, %f111, %f108;
|
||||
mov.b32 %r13, %f109;
|
||||
setp.eq.s32 %p7, %r13, 1118925336;
|
||||
add.s32 %r14, %r13, -1;
|
||||
mov.b32 %f113, %r14;
|
||||
add.f32 %f114, %f112, 0f37000000;
|
||||
selp.f32 %f115, %f113, %f109, %p7;
|
||||
selp.f32 %f6, %f114, %f112, %p7;
|
||||
mul.f32 %f116, %f115, 0f3FB8AA3B;
|
||||
cvt.rzi.f32.f32 %f117, %f116;
|
||||
mov.f32 %f118, 0fBF317200;
|
||||
fma.rn.f32 %f119, %f117, %f118, %f115;
|
||||
mov.f32 %f120, 0fB5BFBE8E;
|
||||
fma.rn.f32 %f121, %f117, %f120, %f119;
|
||||
mul.f32 %f122, %f121, 0f3FB8AA3B;
|
||||
ex2.approx.ftz.f32 %f123, %f122;
|
||||
add.f32 %f124, %f117, 0f00000000;
|
||||
ex2.approx.f32 %f125, %f124;
|
||||
mul.f32 %f126, %f123, %f125;
|
||||
setp.lt.f32 %p8, %f115, 0fC2D20000;
|
||||
selp.f32 %f127, 0f00000000, %f126, %p8;
|
||||
setp.gt.f32 %p9, %f115, 0f42D20000;
|
||||
selp.f32 %f328, 0f7F800000, %f127, %p9;
|
||||
setp.eq.f32 %p10, %f328, 0f7F800000;
|
||||
@%p10 bra BB0_2;
|
||||
|
||||
fma.rn.f32 %f328, %f328, %f6, %f328;
|
||||
|
||||
BB0_2:
|
||||
mov.f32 %f303, 0f3E68BA2E;
|
||||
cvt.rzi.f32.f32 %f302, %f303;
|
||||
fma.rn.f32 %f301, %f302, 0fC0000000, 0f3EE8BA2E;
|
||||
abs.f32 %f300, %f301;
|
||||
setp.lt.f32 %p11, %f3, 0f00000000;
|
||||
setp.eq.f32 %p12, %f300, 0f3F800000;
|
||||
and.pred %p1, %p11, %p12;
|
||||
mov.b32 %r15, %f328;
|
||||
xor.b32 %r16, %r15, -2147483648;
|
||||
mov.b32 %f128, %r16;
|
||||
selp.f32 %f330, %f128, %f328, %p1;
|
||||
setp.eq.f32 %p13, %f3, 0f00000000;
|
||||
@%p13 bra BB0_5;
|
||||
bra.uni BB0_3;
|
||||
|
||||
BB0_5:
|
||||
add.f32 %f131, %f3, %f3;
|
||||
selp.f32 %f330, %f131, 0f00000000, %p12;
|
||||
bra.uni BB0_6;
|
||||
|
||||
BB0_3:
|
||||
setp.geu.f32 %p14, %f3, 0f00000000;
|
||||
@%p14 bra BB0_6;
|
||||
|
||||
mov.f32 %f327, 0f3EE8BA2E;
|
||||
cvt.rzi.f32.f32 %f130, %f327;
|
||||
setp.neu.f32 %p15, %f130, 0f3EE8BA2E;
|
||||
selp.f32 %f330, 0f7FFFFFFF, %f330, %p15;
|
||||
|
||||
BB0_6:
|
||||
abs.f32 %f304, %f3;
|
||||
add.f32 %f132, %f304, 0f3EE8BA2E;
|
||||
mov.b32 %r17, %f132;
|
||||
setp.lt.s32 %p17, %r17, 2139095040;
|
||||
@%p17 bra BB0_11;
|
||||
|
||||
abs.f32 %f325, %f3;
|
||||
setp.gtu.f32 %p18, %f325, 0f7F800000;
|
||||
@%p18 bra BB0_10;
|
||||
bra.uni BB0_8;
|
||||
|
||||
BB0_10:
|
||||
add.f32 %f330, %f3, 0f3EE8BA2E;
|
||||
bra.uni BB0_11;
|
||||
|
||||
BB0_8:
|
||||
abs.f32 %f326, %f3;
|
||||
setp.neu.f32 %p19, %f326, 0f7F800000;
|
||||
@%p19 bra BB0_11;
|
||||
|
||||
selp.f32 %f330, 0fFF800000, 0f7F800000, %p1;
|
||||
|
||||
BB0_11:
|
||||
mov.f32 %f313, 0fB5BFBE8E;
|
||||
mov.f32 %f312, 0fBF317200;
|
||||
mov.f32 %f311, 0f00000000;
|
||||
mov.f32 %f310, 0f35BFBE8E;
|
||||
mov.f32 %f309, 0f3F317200;
|
||||
mov.f32 %f308, 0f3DAAAABD;
|
||||
mov.f32 %f307, 0f3C4CAF63;
|
||||
mov.f32 %f306, 0f3B18F0FE;
|
||||
mov.f32 %f305, 0f3EE8BA2E;
|
||||
setp.eq.f32 %p20, %f3, 0f3F800000;
|
||||
selp.f32 %f135, 0f3F800000, %f330, %p20;
|
||||
cvt.sat.f32.f32 %f17, %f135;
|
||||
abs.f32 %f18, %f1;
|
||||
setp.lt.f32 %p21, %f18, 0f00800000;
|
||||
mul.f32 %f136, %f18, 0f4B800000;
|
||||
selp.f32 %f137, 0fC3170000, 0fC2FE0000, %p21;
|
||||
selp.f32 %f138, %f136, %f18, %p21;
|
||||
mov.b32 %r18, %f138;
|
||||
and.b32 %r19, %r18, 8388607;
|
||||
or.b32 %r20, %r19, 1065353216;
|
||||
mov.b32 %f139, %r20;
|
||||
shr.u32 %r21, %r18, 23;
|
||||
cvt.rn.f32.u32 %f140, %r21;
|
||||
add.f32 %f141, %f137, %f140;
|
||||
setp.gt.f32 %p22, %f139, 0f3FB504F3;
|
||||
mul.f32 %f142, %f139, 0f3F000000;
|
||||
add.f32 %f143, %f141, 0f3F800000;
|
||||
selp.f32 %f144, %f142, %f139, %p22;
|
||||
selp.f32 %f145, %f143, %f141, %p22;
|
||||
add.f32 %f146, %f144, 0fBF800000;
|
||||
add.f32 %f134, %f144, 0f3F800000;
|
||||
// inline asm
|
||||
rcp.approx.ftz.f32 %f133,%f134;
|
||||
// inline asm
|
||||
add.f32 %f147, %f146, %f146;
|
||||
mul.f32 %f148, %f133, %f147;
|
||||
mul.f32 %f149, %f148, %f148;
|
||||
fma.rn.f32 %f152, %f306, %f149, %f307;
|
||||
fma.rn.f32 %f154, %f152, %f149, %f308;
|
||||
mul.rn.f32 %f155, %f154, %f149;
|
||||
mul.rn.f32 %f156, %f155, %f148;
|
||||
sub.f32 %f157, %f146, %f148;
|
||||
neg.f32 %f158, %f148;
|
||||
add.f32 %f159, %f157, %f157;
|
||||
fma.rn.f32 %f160, %f158, %f146, %f159;
|
||||
mul.rn.f32 %f161, %f133, %f160;
|
||||
add.f32 %f162, %f156, %f148;
|
||||
sub.f32 %f163, %f148, %f162;
|
||||
add.f32 %f164, %f156, %f163;
|
||||
add.f32 %f165, %f161, %f164;
|
||||
add.f32 %f166, %f162, %f165;
|
||||
sub.f32 %f167, %f162, %f166;
|
||||
add.f32 %f168, %f165, %f167;
|
||||
mul.rn.f32 %f170, %f145, %f309;
|
||||
mul.rn.f32 %f172, %f145, %f310;
|
||||
add.f32 %f173, %f170, %f166;
|
||||
sub.f32 %f174, %f170, %f173;
|
||||
add.f32 %f175, %f166, %f174;
|
||||
add.f32 %f176, %f168, %f175;
|
||||
add.f32 %f177, %f172, %f176;
|
||||
add.f32 %f178, %f173, %f177;
|
||||
sub.f32 %f179, %f173, %f178;
|
||||
add.f32 %f180, %f177, %f179;
|
||||
mul.rn.f32 %f182, %f305, %f178;
|
||||
neg.f32 %f183, %f182;
|
||||
fma.rn.f32 %f184, %f305, %f178, %f183;
|
||||
fma.rn.f32 %f185, %f305, %f180, %f184;
|
||||
fma.rn.f32 %f187, %f311, %f178, %f185;
|
||||
add.rn.f32 %f188, %f182, %f187;
|
||||
neg.f32 %f189, %f188;
|
||||
add.rn.f32 %f190, %f182, %f189;
|
||||
add.rn.f32 %f191, %f190, %f187;
|
||||
mov.b32 %r22, %f188;
|
||||
setp.eq.s32 %p23, %r22, 1118925336;
|
||||
add.s32 %r23, %r22, -1;
|
||||
mov.b32 %f192, %r23;
|
||||
add.f32 %f193, %f191, 0f37000000;
|
||||
selp.f32 %f194, %f192, %f188, %p23;
|
||||
selp.f32 %f19, %f193, %f191, %p23;
|
||||
mul.f32 %f195, %f194, 0f3FB8AA3B;
|
||||
cvt.rzi.f32.f32 %f196, %f195;
|
||||
fma.rn.f32 %f198, %f196, %f312, %f194;
|
||||
fma.rn.f32 %f200, %f196, %f313, %f198;
|
||||
mul.f32 %f201, %f200, 0f3FB8AA3B;
|
||||
ex2.approx.ftz.f32 %f202, %f201;
|
||||
add.f32 %f203, %f196, 0f00000000;
|
||||
ex2.approx.f32 %f204, %f203;
|
||||
mul.f32 %f205, %f202, %f204;
|
||||
setp.lt.f32 %p24, %f194, 0fC2D20000;
|
||||
selp.f32 %f206, 0f00000000, %f205, %p24;
|
||||
setp.gt.f32 %p25, %f194, 0f42D20000;
|
||||
selp.f32 %f331, 0f7F800000, %f206, %p25;
|
||||
setp.eq.f32 %p26, %f331, 0f7F800000;
|
||||
@%p26 bra BB0_13;
|
||||
|
||||
fma.rn.f32 %f331, %f331, %f19, %f331;
|
||||
|
||||
BB0_13:
|
||||
setp.lt.f32 %p27, %f1, 0f00000000;
|
||||
and.pred %p2, %p27, %p12;
|
||||
mov.b32 %r24, %f331;
|
||||
xor.b32 %r25, %r24, -2147483648;
|
||||
mov.b32 %f207, %r25;
|
||||
selp.f32 %f333, %f207, %f331, %p2;
|
||||
setp.eq.f32 %p29, %f1, 0f00000000;
|
||||
@%p29 bra BB0_16;
|
||||
bra.uni BB0_14;
|
||||
|
||||
BB0_16:
|
||||
add.f32 %f210, %f1, %f1;
|
||||
selp.f32 %f333, %f210, 0f00000000, %p12;
|
||||
bra.uni BB0_17;
|
||||
|
||||
BB0_14:
|
||||
setp.geu.f32 %p30, %f1, 0f00000000;
|
||||
@%p30 bra BB0_17;
|
||||
|
||||
mov.f32 %f324, 0f3EE8BA2E;
|
||||
cvt.rzi.f32.f32 %f209, %f324;
|
||||
setp.neu.f32 %p31, %f209, 0f3EE8BA2E;
|
||||
selp.f32 %f333, 0f7FFFFFFF, %f333, %p31;
|
||||
|
||||
BB0_17:
|
||||
add.f32 %f211, %f18, 0f3EE8BA2E;
|
||||
mov.b32 %r26, %f211;
|
||||
setp.lt.s32 %p33, %r26, 2139095040;
|
||||
@%p33 bra BB0_22;
|
||||
|
||||
setp.gtu.f32 %p34, %f18, 0f7F800000;
|
||||
@%p34 bra BB0_21;
|
||||
bra.uni BB0_19;
|
||||
|
||||
BB0_21:
|
||||
add.f32 %f333, %f1, 0f3EE8BA2E;
|
||||
bra.uni BB0_22;
|
||||
|
||||
BB0_19:
|
||||
setp.neu.f32 %p35, %f18, 0f7F800000;
|
||||
@%p35 bra BB0_22;
|
||||
|
||||
selp.f32 %f333, 0fFF800000, 0f7F800000, %p2;
|
||||
|
||||
BB0_22:
|
||||
mov.f32 %f322, 0fB5BFBE8E;
|
||||
mov.f32 %f321, 0fBF317200;
|
||||
mov.f32 %f320, 0f00000000;
|
||||
mov.f32 %f319, 0f35BFBE8E;
|
||||
mov.f32 %f318, 0f3F317200;
|
||||
mov.f32 %f317, 0f3DAAAABD;
|
||||
mov.f32 %f316, 0f3C4CAF63;
|
||||
mov.f32 %f315, 0f3B18F0FE;
|
||||
mov.f32 %f314, 0f3EE8BA2E;
|
||||
setp.eq.f32 %p36, %f1, 0f3F800000;
|
||||
selp.f32 %f214, 0f3F800000, %f333, %p36;
|
||||
cvt.sat.f32.f32 %f30, %f214;
|
||||
abs.f32 %f31, %f2;
|
||||
setp.lt.f32 %p37, %f31, 0f00800000;
|
||||
mul.f32 %f215, %f31, 0f4B800000;
|
||||
selp.f32 %f216, 0fC3170000, 0fC2FE0000, %p37;
|
||||
selp.f32 %f217, %f215, %f31, %p37;
|
||||
mov.b32 %r27, %f217;
|
||||
and.b32 %r28, %r27, 8388607;
|
||||
or.b32 %r29, %r28, 1065353216;
|
||||
mov.b32 %f218, %r29;
|
||||
shr.u32 %r30, %r27, 23;
|
||||
cvt.rn.f32.u32 %f219, %r30;
|
||||
add.f32 %f220, %f216, %f219;
|
||||
setp.gt.f32 %p38, %f218, 0f3FB504F3;
|
||||
mul.f32 %f221, %f218, 0f3F000000;
|
||||
add.f32 %f222, %f220, 0f3F800000;
|
||||
selp.f32 %f223, %f221, %f218, %p38;
|
||||
selp.f32 %f224, %f222, %f220, %p38;
|
||||
add.f32 %f225, %f223, 0fBF800000;
|
||||
add.f32 %f213, %f223, 0f3F800000;
|
||||
// inline asm
|
||||
rcp.approx.ftz.f32 %f212,%f213;
|
||||
// inline asm
|
||||
add.f32 %f226, %f225, %f225;
|
||||
mul.f32 %f227, %f212, %f226;
|
||||
mul.f32 %f228, %f227, %f227;
|
||||
fma.rn.f32 %f231, %f315, %f228, %f316;
|
||||
fma.rn.f32 %f233, %f231, %f228, %f317;
|
||||
mul.rn.f32 %f234, %f233, %f228;
|
||||
mul.rn.f32 %f235, %f234, %f227;
|
||||
sub.f32 %f236, %f225, %f227;
|
||||
neg.f32 %f237, %f227;
|
||||
add.f32 %f238, %f236, %f236;
|
||||
fma.rn.f32 %f239, %f237, %f225, %f238;
|
||||
mul.rn.f32 %f240, %f212, %f239;
|
||||
add.f32 %f241, %f235, %f227;
|
||||
sub.f32 %f242, %f227, %f241;
|
||||
add.f32 %f243, %f235, %f242;
|
||||
add.f32 %f244, %f240, %f243;
|
||||
add.f32 %f245, %f241, %f244;
|
||||
sub.f32 %f246, %f241, %f245;
|
||||
add.f32 %f247, %f244, %f246;
|
||||
mul.rn.f32 %f249, %f224, %f318;
|
||||
mul.rn.f32 %f251, %f224, %f319;
|
||||
add.f32 %f252, %f249, %f245;
|
||||
sub.f32 %f253, %f249, %f252;
|
||||
add.f32 %f254, %f245, %f253;
|
||||
add.f32 %f255, %f247, %f254;
|
||||
add.f32 %f256, %f251, %f255;
|
||||
add.f32 %f257, %f252, %f256;
|
||||
sub.f32 %f258, %f252, %f257;
|
||||
add.f32 %f259, %f256, %f258;
|
||||
mul.rn.f32 %f261, %f314, %f257;
|
||||
neg.f32 %f262, %f261;
|
||||
fma.rn.f32 %f263, %f314, %f257, %f262;
|
||||
fma.rn.f32 %f264, %f314, %f259, %f263;
|
||||
fma.rn.f32 %f266, %f320, %f257, %f264;
|
||||
add.rn.f32 %f267, %f261, %f266;
|
||||
neg.f32 %f268, %f267;
|
||||
add.rn.f32 %f269, %f261, %f268;
|
||||
add.rn.f32 %f270, %f269, %f266;
|
||||
mov.b32 %r31, %f267;
|
||||
setp.eq.s32 %p39, %r31, 1118925336;
|
||||
add.s32 %r32, %r31, -1;
|
||||
mov.b32 %f271, %r32;
|
||||
add.f32 %f272, %f270, 0f37000000;
|
||||
selp.f32 %f273, %f271, %f267, %p39;
|
||||
selp.f32 %f32, %f272, %f270, %p39;
|
||||
mul.f32 %f274, %f273, 0f3FB8AA3B;
|
||||
cvt.rzi.f32.f32 %f275, %f274;
|
||||
fma.rn.f32 %f277, %f275, %f321, %f273;
|
||||
fma.rn.f32 %f279, %f275, %f322, %f277;
|
||||
mul.f32 %f280, %f279, 0f3FB8AA3B;
|
||||
ex2.approx.ftz.f32 %f281, %f280;
|
||||
add.f32 %f282, %f275, 0f00000000;
|
||||
ex2.approx.f32 %f283, %f282;
|
||||
mul.f32 %f284, %f281, %f283;
|
||||
setp.lt.f32 %p40, %f273, 0fC2D20000;
|
||||
selp.f32 %f285, 0f00000000, %f284, %p40;
|
||||
setp.gt.f32 %p41, %f273, 0f42D20000;
|
||||
selp.f32 %f334, 0f7F800000, %f285, %p41;
|
||||
setp.eq.f32 %p42, %f334, 0f7F800000;
|
||||
@%p42 bra BB0_24;
|
||||
|
||||
fma.rn.f32 %f334, %f334, %f32, %f334;
|
||||
|
||||
BB0_24:
|
||||
setp.lt.f32 %p43, %f2, 0f00000000;
|
||||
and.pred %p3, %p43, %p12;
|
||||
mov.b32 %r33, %f334;
|
||||
xor.b32 %r34, %r33, -2147483648;
|
||||
mov.b32 %f286, %r34;
|
||||
selp.f32 %f336, %f286, %f334, %p3;
|
||||
setp.eq.f32 %p45, %f2, 0f00000000;
|
||||
@%p45 bra BB0_27;
|
||||
bra.uni BB0_25;
|
||||
|
||||
BB0_27:
|
||||
add.f32 %f289, %f2, %f2;
|
||||
selp.f32 %f336, %f289, 0f00000000, %p12;
|
||||
bra.uni BB0_28;
|
||||
|
||||
BB0_25:
|
||||
setp.geu.f32 %p46, %f2, 0f00000000;
|
||||
@%p46 bra BB0_28;
|
||||
|
||||
mov.f32 %f323, 0f3EE8BA2E;
|
||||
cvt.rzi.f32.f32 %f288, %f323;
|
||||
setp.neu.f32 %p47, %f288, 0f3EE8BA2E;
|
||||
selp.f32 %f336, 0f7FFFFFFF, %f336, %p47;
|
||||
|
||||
BB0_28:
|
||||
add.f32 %f290, %f31, 0f3EE8BA2E;
|
||||
mov.b32 %r35, %f290;
|
||||
setp.lt.s32 %p49, %r35, 2139095040;
|
||||
@%p49 bra BB0_33;
|
||||
|
||||
setp.gtu.f32 %p50, %f31, 0f7F800000;
|
||||
@%p50 bra BB0_32;
|
||||
bra.uni BB0_30;
|
||||
|
||||
BB0_32:
|
||||
add.f32 %f336, %f2, 0f3EE8BA2E;
|
||||
bra.uni BB0_33;
|
||||
|
||||
BB0_30:
|
||||
setp.neu.f32 %p51, %f31, 0f7F800000;
|
||||
@%p51 bra BB0_33;
|
||||
|
||||
selp.f32 %f336, 0fFF800000, 0f7F800000, %p3;
|
||||
|
||||
BB0_33:
|
||||
setp.eq.f32 %p52, %f2, 0f3F800000;
|
||||
selp.f32 %f291, 0f3F800000, %f336, %p52;
|
||||
cvt.sat.f32.f32 %f292, %f291;
|
||||
mov.f32 %f293, 0f3F800000;
|
||||
sub.f32 %f294, %f293, %f292;
|
||||
sub.f32 %f295, %f293, %f17;
|
||||
setp.eq.s32 %p53, %r6, 0;
|
||||
sub.f32 %f296, %f293, %f30;
|
||||
ld.const.u64 %rd6, [cs+8];
|
||||
cvta.to.global.u64 %rd7, %rd6;
|
||||
shl.b64 %rd8, %rd1, 4;
|
||||
add.s64 %rd9, %rd7, %rd8;
|
||||
selp.f32 %f297, %f292, %f294, %p53;
|
||||
selp.f32 %f298, %f30, %f296, %p53;
|
||||
selp.f32 %f299, %f17, %f295, %p53;
|
||||
st.global.v4.f32 [%rd9], {%f299, %f298, %f297, %f293};
|
||||
ret;
|
||||
}
|
||||
|
||||
|
||||
8
Assets/Editor/x64/Bakery/denoisePrepare72.ptx.meta
Normal file
8
Assets/Editor/x64/Bakery/denoisePrepare72.ptx.meta
Normal file
@@ -0,0 +1,8 @@
|
||||
fileFormatVersion: 2
|
||||
guid: 37759ebb3c14dd94c8fd9cb41ba67722
|
||||
timeCreated: 1605208943
|
||||
licenseType: Store
|
||||
DefaultImporter:
|
||||
userData:
|
||||
assetBundleName:
|
||||
assetBundleVariant:
|
||||
116
Assets/Editor/x64/Bakery/denoisePrepareOIDN.ptx
Normal file
116
Assets/Editor/x64/Bakery/denoisePrepareOIDN.ptx
Normal file
@@ -0,0 +1,116 @@
|
||||
//
|
||||
// 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 1 .b8 image[1];
|
||||
.global .align 1 .b8 input_buffer[1];
|
||||
.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 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 4 .u32 _ZN21rti_internal_typeenum7pixelIDE = 4919;
|
||||
.global .align 4 .u32 _ZN21rti_internal_typeenum10resolutionE = 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 1 .b8 _ZN23rti_internal_annotation7pixelIDE[1];
|
||||
.global .align 1 .b8 _ZN23rti_internal_annotation10resolutionE[1];
|
||||
|
||||
.visible .entry _Z6oxMainv(
|
||||
|
||||
)
|
||||
{
|
||||
.reg .b16 %rs<4>;
|
||||
.reg .f32 %f<4>;
|
||||
.reg .b32 %r<25>;
|
||||
.reg .b64 %rd<27>;
|
||||
|
||||
|
||||
ld.global.v2.u32 {%r9, %r10}, [pixelID];
|
||||
cvt.u64.u32 %rd3, %r9;
|
||||
cvt.u64.u32 %rd4, %r10;
|
||||
mov.u64 %rd25, image;
|
||||
cvta.global.u64 %rd2, %rd25;
|
||||
mov.u32 %r7, 2;
|
||||
mov.u32 %r6, 8;
|
||||
mov.u64 %rd24, 0;
|
||||
// inline asm
|
||||
call (%rd1), _rt_buffer_get_64, (%rd2, %r7, %r6, %rd3, %rd4, %rd24, %rd24);
|
||||
// inline asm
|
||||
ld.u16 %rs1, [%rd1];
|
||||
// inline asm
|
||||
{ cvt.f32.f16 %f1, %rs1;}
|
||||
|
||||
// inline asm
|
||||
ld.global.v2.u32 {%r13, %r14}, [pixelID];
|
||||
cvt.u64.u32 %rd9, %r13;
|
||||
cvt.u64.u32 %rd10, %r14;
|
||||
// inline asm
|
||||
call (%rd7), _rt_buffer_get_64, (%rd2, %r7, %r6, %rd9, %rd10, %rd24, %rd24);
|
||||
// inline asm
|
||||
ld.u16 %rs2, [%rd7+2];
|
||||
// inline asm
|
||||
{ cvt.f32.f16 %f2, %rs2;}
|
||||
|
||||
// inline asm
|
||||
ld.global.v2.u32 {%r17, %r18}, [pixelID];
|
||||
cvt.u64.u32 %rd15, %r17;
|
||||
cvt.u64.u32 %rd16, %r18;
|
||||
// inline asm
|
||||
call (%rd13), _rt_buffer_get_64, (%rd2, %r7, %r6, %rd15, %rd16, %rd24, %rd24);
|
||||
// inline asm
|
||||
ld.u16 %rs3, [%rd13+4];
|
||||
// inline asm
|
||||
{ cvt.f32.f16 %f3, %rs3;}
|
||||
|
||||
// inline asm
|
||||
ld.global.v2.u32 {%r21, %r22}, [pixelID];
|
||||
cvt.u64.u32 %rd21, %r21;
|
||||
cvt.u64.u32 %rd22, %r22;
|
||||
mov.u64 %rd26, input_buffer;
|
||||
cvta.global.u64 %rd20, %rd26;
|
||||
mov.u32 %r8, 12;
|
||||
// inline asm
|
||||
call (%rd19), _rt_buffer_get_64, (%rd20, %r7, %r8, %rd21, %rd22, %rd24, %rd24);
|
||||
// inline asm
|
||||
st.f32 [%rd19+8], %f3;
|
||||
st.f32 [%rd19+4], %f2;
|
||||
st.f32 [%rd19], %f1;
|
||||
ret;
|
||||
}
|
||||
|
||||
|
||||
8
Assets/Editor/x64/Bakery/denoisePrepareOIDN.ptx.meta
Normal file
8
Assets/Editor/x64/Bakery/denoisePrepareOIDN.ptx.meta
Normal file
@@ -0,0 +1,8 @@
|
||||
fileFormatVersion: 2
|
||||
guid: 239bdcce70626464e9dbf82737527c87
|
||||
timeCreated: 1589355275
|
||||
licenseType: Store
|
||||
DefaultImporter:
|
||||
userData:
|
||||
assetBundleName:
|
||||
assetBundleVariant:
|
||||
595
Assets/Editor/x64/Bakery/denoisePrepareSH.ptx
Normal file
595
Assets/Editor/x64/Bakery/denoisePrepareSH.ptx
Normal file
@@ -0,0 +1,595 @@
|
||||
//
|
||||
// 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 1 .b8 input_buffer[1];
|
||||
.global .align 1 .b8 image[1];
|
||||
.global .align 4 .u32 mode;
|
||||
.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_typeinfo4modeE[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 4 .b8 _ZN21rti_internal_typename4modeE[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_typeenum4modeE = 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 1 .b8 _ZN21rti_internal_semantic4modeE[1];
|
||||
.global .align 1 .b8 _ZN23rti_internal_annotation7pixelIDE[1];
|
||||
.global .align 1 .b8 _ZN23rti_internal_annotation10resolutionE[1];
|
||||
.global .align 1 .b8 _ZN23rti_internal_annotation4modeE[1];
|
||||
|
||||
.visible .entry _Z6oxMainv(
|
||||
|
||||
)
|
||||
{
|
||||
.reg .pred %p<53>;
|
||||
.reg .b16 %rs<4>;
|
||||
.reg .f32 %f<327>;
|
||||
.reg .b32 %r<67>;
|
||||
.reg .b64 %rd<48>;
|
||||
|
||||
|
||||
ld.global.u32 %r1, [mode];
|
||||
setp.gt.s32 %p4, %r1, 0;
|
||||
ld.global.v2.u32 {%r2, %r3}, [pixelID];
|
||||
cvt.u64.u32 %rd1, %r2;
|
||||
cvt.u64.u32 %rd2, %r3;
|
||||
@%p4 bra BB0_2;
|
||||
bra.uni BB0_1;
|
||||
|
||||
BB0_2:
|
||||
mov.u64 %rd40, input_buffer;
|
||||
cvta.global.u64 %rd23, %rd40;
|
||||
mov.u32 %r24, 2;
|
||||
mov.u32 %r25, 16;
|
||||
mov.u64 %rd39, 0;
|
||||
// inline asm
|
||||
call (%rd22), _rt_buffer_get_64, (%rd23, %r24, %r25, %rd1, %rd2, %rd39, %rd39);
|
||||
// inline asm
|
||||
ld.f32 %f315, [%rd22];
|
||||
ld.global.v2.u32 {%r26, %r27}, [pixelID];
|
||||
cvt.u64.u32 %rd30, %r26;
|
||||
cvt.u64.u32 %rd31, %r27;
|
||||
// inline asm
|
||||
call (%rd28), _rt_buffer_get_64, (%rd23, %r24, %r25, %rd30, %rd31, %rd39, %rd39);
|
||||
// inline asm
|
||||
ld.f32 %f316, [%rd28+4];
|
||||
ld.global.v2.u32 {%r30, %r31}, [pixelID];
|
||||
cvt.u64.u32 %rd36, %r30;
|
||||
cvt.u64.u32 %rd37, %r31;
|
||||
// inline asm
|
||||
call (%rd34), _rt_buffer_get_64, (%rd23, %r24, %r25, %rd36, %rd37, %rd39, %rd39);
|
||||
// inline asm
|
||||
ld.f32 %f317, [%rd34+8];
|
||||
bra.uni BB0_3;
|
||||
|
||||
BB0_1:
|
||||
mov.u64 %rd21, image;
|
||||
cvta.global.u64 %rd4, %rd21;
|
||||
mov.u32 %r10, 2;
|
||||
mov.u32 %r11, 8;
|
||||
mov.u64 %rd20, 0;
|
||||
// inline asm
|
||||
call (%rd3), _rt_buffer_get_64, (%rd4, %r10, %r11, %rd1, %rd2, %rd20, %rd20);
|
||||
// inline asm
|
||||
ld.u16 %rs1, [%rd3];
|
||||
// inline asm
|
||||
{ cvt.f32.f16 %f49, %rs1;}
|
||||
|
||||
// inline asm
|
||||
ld.global.v2.u32 {%r12, %r13}, [pixelID];
|
||||
cvt.u64.u32 %rd11, %r12;
|
||||
cvt.u64.u32 %rd12, %r13;
|
||||
// inline asm
|
||||
call (%rd9), _rt_buffer_get_64, (%rd4, %r10, %r11, %rd11, %rd12, %rd20, %rd20);
|
||||
// inline asm
|
||||
ld.u16 %rs2, [%rd9+2];
|
||||
// inline asm
|
||||
{ cvt.f32.f16 %f50, %rs2;}
|
||||
|
||||
// inline asm
|
||||
ld.global.v2.u32 {%r16, %r17}, [pixelID];
|
||||
cvt.u64.u32 %rd17, %r16;
|
||||
cvt.u64.u32 %rd18, %r17;
|
||||
// inline asm
|
||||
call (%rd15), _rt_buffer_get_64, (%rd4, %r10, %r11, %rd17, %rd18, %rd20, %rd20);
|
||||
// inline asm
|
||||
ld.u16 %rs3, [%rd15+4];
|
||||
// inline asm
|
||||
{ cvt.f32.f16 %f51, %rs3;}
|
||||
|
||||
// inline asm
|
||||
mul.f32 %f52, %f49, 0f40800000;
|
||||
mul.f32 %f53, %f50, 0f40800000;
|
||||
mul.f32 %f54, %f51, 0f40800000;
|
||||
max.f32 %f55, %f52, %f53;
|
||||
max.f32 %f56, %f55, %f54;
|
||||
add.f32 %f57, %f56, 0f3F800000;
|
||||
rcp.rn.f32 %f58, %f57;
|
||||
mul.f32 %f315, %f52, %f58;
|
||||
mul.f32 %f316, %f53, %f58;
|
||||
mul.f32 %f317, %f54, %f58;
|
||||
|
||||
BB0_3:
|
||||
mov.f32 %f61, 0f3E68BA2E;
|
||||
cvt.rzi.f32.f32 %f62, %f61;
|
||||
fma.rn.f32 %f63, %f62, 0fC0000000, 0f3EE8BA2E;
|
||||
abs.f32 %f10, %f63;
|
||||
abs.f32 %f11, %f315;
|
||||
setp.lt.f32 %p5, %f11, 0f00800000;
|
||||
mul.f32 %f64, %f11, 0f4B800000;
|
||||
selp.f32 %f65, 0fC3170000, 0fC2FE0000, %p5;
|
||||
selp.f32 %f66, %f64, %f11, %p5;
|
||||
mov.b32 %r34, %f66;
|
||||
and.b32 %r35, %r34, 8388607;
|
||||
or.b32 %r36, %r35, 1065353216;
|
||||
mov.b32 %f67, %r36;
|
||||
shr.u32 %r37, %r34, 23;
|
||||
cvt.rn.f32.u32 %f68, %r37;
|
||||
add.f32 %f69, %f65, %f68;
|
||||
setp.gt.f32 %p6, %f67, 0f3FB504F3;
|
||||
mul.f32 %f70, %f67, 0f3F000000;
|
||||
add.f32 %f71, %f69, 0f3F800000;
|
||||
selp.f32 %f72, %f70, %f67, %p6;
|
||||
selp.f32 %f73, %f71, %f69, %p6;
|
||||
add.f32 %f74, %f72, 0fBF800000;
|
||||
add.f32 %f60, %f72, 0f3F800000;
|
||||
// inline asm
|
||||
rcp.approx.ftz.f32 %f59,%f60;
|
||||
// inline asm
|
||||
add.f32 %f75, %f74, %f74;
|
||||
mul.f32 %f76, %f59, %f75;
|
||||
mul.f32 %f77, %f76, %f76;
|
||||
mov.f32 %f78, 0f3C4CAF63;
|
||||
mov.f32 %f79, 0f3B18F0FE;
|
||||
fma.rn.f32 %f80, %f79, %f77, %f78;
|
||||
mov.f32 %f81, 0f3DAAAABD;
|
||||
fma.rn.f32 %f82, %f80, %f77, %f81;
|
||||
mul.rn.f32 %f83, %f82, %f77;
|
||||
mul.rn.f32 %f84, %f83, %f76;
|
||||
sub.f32 %f85, %f74, %f76;
|
||||
neg.f32 %f86, %f76;
|
||||
add.f32 %f87, %f85, %f85;
|
||||
fma.rn.f32 %f88, %f86, %f74, %f87;
|
||||
mul.rn.f32 %f89, %f59, %f88;
|
||||
add.f32 %f90, %f84, %f76;
|
||||
sub.f32 %f91, %f76, %f90;
|
||||
add.f32 %f92, %f84, %f91;
|
||||
add.f32 %f93, %f89, %f92;
|
||||
add.f32 %f94, %f90, %f93;
|
||||
sub.f32 %f95, %f90, %f94;
|
||||
add.f32 %f96, %f93, %f95;
|
||||
mov.f32 %f97, 0f3F317200;
|
||||
mul.rn.f32 %f98, %f73, %f97;
|
||||
mov.f32 %f99, 0f35BFBE8E;
|
||||
mul.rn.f32 %f100, %f73, %f99;
|
||||
add.f32 %f101, %f98, %f94;
|
||||
sub.f32 %f102, %f98, %f101;
|
||||
add.f32 %f103, %f94, %f102;
|
||||
add.f32 %f104, %f96, %f103;
|
||||
add.f32 %f105, %f100, %f104;
|
||||
add.f32 %f106, %f101, %f105;
|
||||
sub.f32 %f107, %f101, %f106;
|
||||
add.f32 %f108, %f105, %f107;
|
||||
mov.f32 %f109, 0f3EE8BA2E;
|
||||
mul.rn.f32 %f110, %f109, %f106;
|
||||
neg.f32 %f111, %f110;
|
||||
fma.rn.f32 %f112, %f109, %f106, %f111;
|
||||
fma.rn.f32 %f113, %f109, %f108, %f112;
|
||||
mov.f32 %f114, 0f00000000;
|
||||
fma.rn.f32 %f115, %f114, %f106, %f113;
|
||||
add.rn.f32 %f116, %f110, %f115;
|
||||
neg.f32 %f117, %f116;
|
||||
add.rn.f32 %f118, %f110, %f117;
|
||||
add.rn.f32 %f119, %f118, %f115;
|
||||
mov.b32 %r38, %f116;
|
||||
setp.eq.s32 %p7, %r38, 1118925336;
|
||||
add.s32 %r39, %r38, -1;
|
||||
mov.b32 %f120, %r39;
|
||||
add.f32 %f121, %f119, 0f37000000;
|
||||
selp.f32 %f122, %f120, %f116, %p7;
|
||||
selp.f32 %f12, %f121, %f119, %p7;
|
||||
mul.f32 %f123, %f122, 0f3FB8AA3B;
|
||||
cvt.rzi.f32.f32 %f124, %f123;
|
||||
mov.f32 %f125, 0fBF317200;
|
||||
fma.rn.f32 %f126, %f124, %f125, %f122;
|
||||
mov.f32 %f127, 0fB5BFBE8E;
|
||||
fma.rn.f32 %f128, %f124, %f127, %f126;
|
||||
mul.f32 %f129, %f128, 0f3FB8AA3B;
|
||||
ex2.approx.ftz.f32 %f130, %f129;
|
||||
add.f32 %f131, %f124, 0f00000000;
|
||||
ex2.approx.f32 %f132, %f131;
|
||||
mul.f32 %f133, %f130, %f132;
|
||||
setp.lt.f32 %p8, %f122, 0fC2D20000;
|
||||
selp.f32 %f134, 0f00000000, %f133, %p8;
|
||||
setp.gt.f32 %p9, %f122, 0f42D20000;
|
||||
selp.f32 %f318, 0f7F800000, %f134, %p9;
|
||||
setp.eq.f32 %p10, %f318, 0f7F800000;
|
||||
@%p10 bra BB0_5;
|
||||
|
||||
fma.rn.f32 %f318, %f318, %f12, %f318;
|
||||
|
||||
BB0_5:
|
||||
setp.lt.f32 %p11, %f315, 0f00000000;
|
||||
setp.eq.f32 %p12, %f10, 0f3F800000;
|
||||
and.pred %p1, %p11, %p12;
|
||||
mov.b32 %r40, %f318;
|
||||
xor.b32 %r41, %r40, -2147483648;
|
||||
mov.b32 %f135, %r41;
|
||||
selp.f32 %f320, %f135, %f318, %p1;
|
||||
setp.eq.f32 %p13, %f315, 0f00000000;
|
||||
@%p13 bra BB0_8;
|
||||
bra.uni BB0_6;
|
||||
|
||||
BB0_8:
|
||||
add.f32 %f138, %f315, %f315;
|
||||
selp.f32 %f320, %f138, 0f00000000, %p12;
|
||||
bra.uni BB0_9;
|
||||
|
||||
BB0_6:
|
||||
setp.geu.f32 %p14, %f315, 0f00000000;
|
||||
@%p14 bra BB0_9;
|
||||
|
||||
cvt.rzi.f32.f32 %f137, %f109;
|
||||
setp.neu.f32 %p15, %f137, 0f3EE8BA2E;
|
||||
selp.f32 %f320, 0f7FFFFFFF, %f320, %p15;
|
||||
|
||||
BB0_9:
|
||||
add.f32 %f139, %f11, 0f3EE8BA2E;
|
||||
mov.b32 %r42, %f139;
|
||||
setp.lt.s32 %p17, %r42, 2139095040;
|
||||
@%p17 bra BB0_14;
|
||||
|
||||
setp.gtu.f32 %p18, %f11, 0f7F800000;
|
||||
@%p18 bra BB0_13;
|
||||
bra.uni BB0_11;
|
||||
|
||||
BB0_13:
|
||||
add.f32 %f320, %f315, 0f3EE8BA2E;
|
||||
bra.uni BB0_14;
|
||||
|
||||
BB0_11:
|
||||
setp.neu.f32 %p19, %f11, 0f7F800000;
|
||||
@%p19 bra BB0_14;
|
||||
|
||||
selp.f32 %f320, 0fFF800000, 0f7F800000, %p1;
|
||||
|
||||
BB0_14:
|
||||
mov.f32 %f307, 0fBF317200;
|
||||
mov.f32 %f306, 0f00000000;
|
||||
mov.f32 %f305, 0f35BFBE8E;
|
||||
mov.f32 %f304, 0f3F317200;
|
||||
mov.f32 %f303, 0f3DAAAABD;
|
||||
mov.f32 %f302, 0f3C4CAF63;
|
||||
mov.f32 %f301, 0f3B18F0FE;
|
||||
setp.eq.f32 %p20, %f315, 0f3F800000;
|
||||
selp.f32 %f142, 0f3F800000, %f320, %p20;
|
||||
cvt.sat.f32.f32 %f23, %f142;
|
||||
abs.f32 %f24, %f316;
|
||||
setp.lt.f32 %p21, %f24, 0f00800000;
|
||||
mul.f32 %f143, %f24, 0f4B800000;
|
||||
selp.f32 %f144, 0fC3170000, 0fC2FE0000, %p21;
|
||||
selp.f32 %f145, %f143, %f24, %p21;
|
||||
mov.b32 %r43, %f145;
|
||||
and.b32 %r44, %r43, 8388607;
|
||||
or.b32 %r45, %r44, 1065353216;
|
||||
mov.b32 %f146, %r45;
|
||||
shr.u32 %r46, %r43, 23;
|
||||
cvt.rn.f32.u32 %f147, %r46;
|
||||
add.f32 %f148, %f144, %f147;
|
||||
setp.gt.f32 %p22, %f146, 0f3FB504F3;
|
||||
mul.f32 %f149, %f146, 0f3F000000;
|
||||
add.f32 %f150, %f148, 0f3F800000;
|
||||
selp.f32 %f151, %f149, %f146, %p22;
|
||||
selp.f32 %f152, %f150, %f148, %p22;
|
||||
add.f32 %f153, %f151, 0fBF800000;
|
||||
add.f32 %f141, %f151, 0f3F800000;
|
||||
// inline asm
|
||||
rcp.approx.ftz.f32 %f140,%f141;
|
||||
// inline asm
|
||||
add.f32 %f154, %f153, %f153;
|
||||
mul.f32 %f155, %f140, %f154;
|
||||
mul.f32 %f156, %f155, %f155;
|
||||
fma.rn.f32 %f159, %f301, %f156, %f302;
|
||||
fma.rn.f32 %f161, %f159, %f156, %f303;
|
||||
mul.rn.f32 %f162, %f161, %f156;
|
||||
mul.rn.f32 %f163, %f162, %f155;
|
||||
sub.f32 %f164, %f153, %f155;
|
||||
neg.f32 %f165, %f155;
|
||||
add.f32 %f166, %f164, %f164;
|
||||
fma.rn.f32 %f167, %f165, %f153, %f166;
|
||||
mul.rn.f32 %f168, %f140, %f167;
|
||||
add.f32 %f169, %f163, %f155;
|
||||
sub.f32 %f170, %f155, %f169;
|
||||
add.f32 %f171, %f163, %f170;
|
||||
add.f32 %f172, %f168, %f171;
|
||||
add.f32 %f173, %f169, %f172;
|
||||
sub.f32 %f174, %f169, %f173;
|
||||
add.f32 %f175, %f172, %f174;
|
||||
mul.rn.f32 %f177, %f152, %f304;
|
||||
mul.rn.f32 %f179, %f152, %f305;
|
||||
add.f32 %f180, %f177, %f173;
|
||||
sub.f32 %f181, %f177, %f180;
|
||||
add.f32 %f182, %f173, %f181;
|
||||
add.f32 %f183, %f175, %f182;
|
||||
add.f32 %f184, %f179, %f183;
|
||||
add.f32 %f185, %f180, %f184;
|
||||
sub.f32 %f186, %f180, %f185;
|
||||
add.f32 %f187, %f184, %f186;
|
||||
mul.rn.f32 %f189, %f109, %f185;
|
||||
neg.f32 %f190, %f189;
|
||||
fma.rn.f32 %f191, %f109, %f185, %f190;
|
||||
fma.rn.f32 %f192, %f109, %f187, %f191;
|
||||
fma.rn.f32 %f194, %f306, %f185, %f192;
|
||||
add.rn.f32 %f195, %f189, %f194;
|
||||
neg.f32 %f196, %f195;
|
||||
add.rn.f32 %f197, %f189, %f196;
|
||||
add.rn.f32 %f198, %f197, %f194;
|
||||
mov.b32 %r47, %f195;
|
||||
setp.eq.s32 %p23, %r47, 1118925336;
|
||||
add.s32 %r48, %r47, -1;
|
||||
mov.b32 %f199, %r48;
|
||||
add.f32 %f200, %f198, 0f37000000;
|
||||
selp.f32 %f201, %f199, %f195, %p23;
|
||||
selp.f32 %f25, %f200, %f198, %p23;
|
||||
mul.f32 %f202, %f201, 0f3FB8AA3B;
|
||||
cvt.rzi.f32.f32 %f203, %f202;
|
||||
fma.rn.f32 %f205, %f203, %f307, %f201;
|
||||
fma.rn.f32 %f207, %f203, %f127, %f205;
|
||||
mul.f32 %f208, %f207, 0f3FB8AA3B;
|
||||
ex2.approx.ftz.f32 %f209, %f208;
|
||||
add.f32 %f210, %f203, 0f00000000;
|
||||
ex2.approx.f32 %f211, %f210;
|
||||
mul.f32 %f212, %f209, %f211;
|
||||
setp.lt.f32 %p24, %f201, 0fC2D20000;
|
||||
selp.f32 %f213, 0f00000000, %f212, %p24;
|
||||
setp.gt.f32 %p25, %f201, 0f42D20000;
|
||||
selp.f32 %f321, 0f7F800000, %f213, %p25;
|
||||
setp.eq.f32 %p26, %f321, 0f7F800000;
|
||||
@%p26 bra BB0_16;
|
||||
|
||||
fma.rn.f32 %f321, %f321, %f25, %f321;
|
||||
|
||||
BB0_16:
|
||||
setp.lt.f32 %p27, %f316, 0f00000000;
|
||||
and.pred %p2, %p27, %p12;
|
||||
mov.b32 %r49, %f321;
|
||||
xor.b32 %r50, %r49, -2147483648;
|
||||
mov.b32 %f214, %r50;
|
||||
selp.f32 %f323, %f214, %f321, %p2;
|
||||
setp.eq.f32 %p29, %f316, 0f00000000;
|
||||
@%p29 bra BB0_19;
|
||||
bra.uni BB0_17;
|
||||
|
||||
BB0_19:
|
||||
add.f32 %f217, %f316, %f316;
|
||||
selp.f32 %f323, %f217, 0f00000000, %p12;
|
||||
bra.uni BB0_20;
|
||||
|
||||
BB0_17:
|
||||
setp.geu.f32 %p30, %f316, 0f00000000;
|
||||
@%p30 bra BB0_20;
|
||||
|
||||
cvt.rzi.f32.f32 %f216, %f109;
|
||||
setp.neu.f32 %p31, %f216, 0f3EE8BA2E;
|
||||
selp.f32 %f323, 0f7FFFFFFF, %f323, %p31;
|
||||
|
||||
BB0_20:
|
||||
add.f32 %f218, %f24, 0f3EE8BA2E;
|
||||
mov.b32 %r51, %f218;
|
||||
setp.lt.s32 %p33, %r51, 2139095040;
|
||||
@%p33 bra BB0_25;
|
||||
|
||||
setp.gtu.f32 %p34, %f24, 0f7F800000;
|
||||
@%p34 bra BB0_24;
|
||||
bra.uni BB0_22;
|
||||
|
||||
BB0_24:
|
||||
add.f32 %f323, %f316, 0f3EE8BA2E;
|
||||
bra.uni BB0_25;
|
||||
|
||||
BB0_22:
|
||||
setp.neu.f32 %p35, %f24, 0f7F800000;
|
||||
@%p35 bra BB0_25;
|
||||
|
||||
selp.f32 %f323, 0fFF800000, 0f7F800000, %p2;
|
||||
|
||||
BB0_25:
|
||||
mov.f32 %f314, 0fBF317200;
|
||||
mov.f32 %f313, 0f00000000;
|
||||
mov.f32 %f312, 0f35BFBE8E;
|
||||
mov.f32 %f311, 0f3F317200;
|
||||
mov.f32 %f310, 0f3DAAAABD;
|
||||
mov.f32 %f309, 0f3C4CAF63;
|
||||
mov.f32 %f308, 0f3B18F0FE;
|
||||
setp.eq.f32 %p36, %f316, 0f3F800000;
|
||||
selp.f32 %f221, 0f3F800000, %f323, %p36;
|
||||
cvt.sat.f32.f32 %f36, %f221;
|
||||
abs.f32 %f37, %f317;
|
||||
setp.lt.f32 %p37, %f37, 0f00800000;
|
||||
mul.f32 %f222, %f37, 0f4B800000;
|
||||
selp.f32 %f223, 0fC3170000, 0fC2FE0000, %p37;
|
||||
selp.f32 %f224, %f222, %f37, %p37;
|
||||
mov.b32 %r52, %f224;
|
||||
and.b32 %r53, %r52, 8388607;
|
||||
or.b32 %r54, %r53, 1065353216;
|
||||
mov.b32 %f225, %r54;
|
||||
shr.u32 %r55, %r52, 23;
|
||||
cvt.rn.f32.u32 %f226, %r55;
|
||||
add.f32 %f227, %f223, %f226;
|
||||
setp.gt.f32 %p38, %f225, 0f3FB504F3;
|
||||
mul.f32 %f228, %f225, 0f3F000000;
|
||||
add.f32 %f229, %f227, 0f3F800000;
|
||||
selp.f32 %f230, %f228, %f225, %p38;
|
||||
selp.f32 %f231, %f229, %f227, %p38;
|
||||
add.f32 %f232, %f230, 0fBF800000;
|
||||
add.f32 %f220, %f230, 0f3F800000;
|
||||
// inline asm
|
||||
rcp.approx.ftz.f32 %f219,%f220;
|
||||
// inline asm
|
||||
add.f32 %f233, %f232, %f232;
|
||||
mul.f32 %f234, %f219, %f233;
|
||||
mul.f32 %f235, %f234, %f234;
|
||||
fma.rn.f32 %f238, %f308, %f235, %f309;
|
||||
fma.rn.f32 %f240, %f238, %f235, %f310;
|
||||
mul.rn.f32 %f241, %f240, %f235;
|
||||
mul.rn.f32 %f242, %f241, %f234;
|
||||
sub.f32 %f243, %f232, %f234;
|
||||
neg.f32 %f244, %f234;
|
||||
add.f32 %f245, %f243, %f243;
|
||||
fma.rn.f32 %f246, %f244, %f232, %f245;
|
||||
mul.rn.f32 %f247, %f219, %f246;
|
||||
add.f32 %f248, %f242, %f234;
|
||||
sub.f32 %f249, %f234, %f248;
|
||||
add.f32 %f250, %f242, %f249;
|
||||
add.f32 %f251, %f247, %f250;
|
||||
add.f32 %f252, %f248, %f251;
|
||||
sub.f32 %f253, %f248, %f252;
|
||||
add.f32 %f254, %f251, %f253;
|
||||
mul.rn.f32 %f256, %f231, %f311;
|
||||
mul.rn.f32 %f258, %f231, %f312;
|
||||
add.f32 %f259, %f256, %f252;
|
||||
sub.f32 %f260, %f256, %f259;
|
||||
add.f32 %f261, %f252, %f260;
|
||||
add.f32 %f262, %f254, %f261;
|
||||
add.f32 %f263, %f258, %f262;
|
||||
add.f32 %f264, %f259, %f263;
|
||||
sub.f32 %f265, %f259, %f264;
|
||||
add.f32 %f266, %f263, %f265;
|
||||
mul.rn.f32 %f268, %f109, %f264;
|
||||
neg.f32 %f269, %f268;
|
||||
fma.rn.f32 %f270, %f109, %f264, %f269;
|
||||
fma.rn.f32 %f271, %f109, %f266, %f270;
|
||||
fma.rn.f32 %f273, %f313, %f264, %f271;
|
||||
add.rn.f32 %f274, %f268, %f273;
|
||||
neg.f32 %f275, %f274;
|
||||
add.rn.f32 %f276, %f268, %f275;
|
||||
add.rn.f32 %f277, %f276, %f273;
|
||||
mov.b32 %r56, %f274;
|
||||
setp.eq.s32 %p39, %r56, 1118925336;
|
||||
add.s32 %r57, %r56, -1;
|
||||
mov.b32 %f278, %r57;
|
||||
add.f32 %f279, %f277, 0f37000000;
|
||||
selp.f32 %f280, %f278, %f274, %p39;
|
||||
selp.f32 %f38, %f279, %f277, %p39;
|
||||
mul.f32 %f281, %f280, 0f3FB8AA3B;
|
||||
cvt.rzi.f32.f32 %f282, %f281;
|
||||
fma.rn.f32 %f284, %f282, %f314, %f280;
|
||||
fma.rn.f32 %f286, %f282, %f127, %f284;
|
||||
mul.f32 %f287, %f286, 0f3FB8AA3B;
|
||||
ex2.approx.ftz.f32 %f288, %f287;
|
||||
add.f32 %f289, %f282, 0f00000000;
|
||||
ex2.approx.f32 %f290, %f289;
|
||||
mul.f32 %f291, %f288, %f290;
|
||||
setp.lt.f32 %p40, %f280, 0fC2D20000;
|
||||
selp.f32 %f292, 0f00000000, %f291, %p40;
|
||||
setp.gt.f32 %p41, %f280, 0f42D20000;
|
||||
selp.f32 %f324, 0f7F800000, %f292, %p41;
|
||||
setp.eq.f32 %p42, %f324, 0f7F800000;
|
||||
@%p42 bra BB0_27;
|
||||
|
||||
fma.rn.f32 %f324, %f324, %f38, %f324;
|
||||
|
||||
BB0_27:
|
||||
setp.lt.f32 %p43, %f317, 0f00000000;
|
||||
and.pred %p3, %p43, %p12;
|
||||
mov.b32 %r58, %f324;
|
||||
xor.b32 %r59, %r58, -2147483648;
|
||||
mov.b32 %f293, %r59;
|
||||
selp.f32 %f326, %f293, %f324, %p3;
|
||||
setp.eq.f32 %p45, %f317, 0f00000000;
|
||||
@%p45 bra BB0_30;
|
||||
bra.uni BB0_28;
|
||||
|
||||
BB0_30:
|
||||
add.f32 %f296, %f317, %f317;
|
||||
selp.f32 %f326, %f296, 0f00000000, %p12;
|
||||
bra.uni BB0_31;
|
||||
|
||||
BB0_28:
|
||||
setp.geu.f32 %p46, %f317, 0f00000000;
|
||||
@%p46 bra BB0_31;
|
||||
|
||||
cvt.rzi.f32.f32 %f295, %f109;
|
||||
setp.neu.f32 %p47, %f295, 0f3EE8BA2E;
|
||||
selp.f32 %f326, 0f7FFFFFFF, %f326, %p47;
|
||||
|
||||
BB0_31:
|
||||
add.f32 %f297, %f37, 0f3EE8BA2E;
|
||||
mov.b32 %r60, %f297;
|
||||
setp.lt.s32 %p49, %r60, 2139095040;
|
||||
@%p49 bra BB0_36;
|
||||
|
||||
setp.gtu.f32 %p50, %f37, 0f7F800000;
|
||||
@%p50 bra BB0_35;
|
||||
bra.uni BB0_33;
|
||||
|
||||
BB0_35:
|
||||
add.f32 %f326, %f317, 0f3EE8BA2E;
|
||||
bra.uni BB0_36;
|
||||
|
||||
BB0_33:
|
||||
setp.neu.f32 %p51, %f37, 0f7F800000;
|
||||
@%p51 bra BB0_36;
|
||||
|
||||
selp.f32 %f326, 0fFF800000, 0f7F800000, %p3;
|
||||
|
||||
BB0_36:
|
||||
setp.eq.f32 %p52, %f317, 0f3F800000;
|
||||
selp.f32 %f298, 0f3F800000, %f326, %p52;
|
||||
ld.global.v2.u32 {%r63, %r64}, [pixelID];
|
||||
cvt.u64.u32 %rd43, %r63;
|
||||
cvt.u64.u32 %rd44, %r64;
|
||||
mov.u64 %rd47, input_buffer;
|
||||
cvta.global.u64 %rd42, %rd47;
|
||||
mov.u32 %r61, 2;
|
||||
mov.u32 %r62, 16;
|
||||
mov.u64 %rd46, 0;
|
||||
// inline asm
|
||||
call (%rd41), _rt_buffer_get_64, (%rd42, %r61, %r62, %rd43, %rd44, %rd46, %rd46);
|
||||
// inline asm
|
||||
cvt.sat.f32.f32 %f299, %f298;
|
||||
mov.f32 %f300, 0f3F800000;
|
||||
st.v4.f32 [%rd41], {%f23, %f36, %f299, %f300};
|
||||
ret;
|
||||
}
|
||||
|
||||
|
||||
8
Assets/Editor/x64/Bakery/denoisePrepareSH.ptx.meta
Normal file
8
Assets/Editor/x64/Bakery/denoisePrepareSH.ptx.meta
Normal file
@@ -0,0 +1,8 @@
|
||||
fileFormatVersion: 2
|
||||
guid: 20b1479cc74212c4bb1e3ac7a32c7eb4
|
||||
timeCreated: 1538345689
|
||||
licenseType: Store
|
||||
DefaultImporter:
|
||||
userData:
|
||||
assetBundleName:
|
||||
assetBundleVariant:
|
||||
519
Assets/Editor/x64/Bakery/denoisePrepareSH72.ptx
Normal file
519
Assets/Editor/x64/Bakery/denoisePrepareSH72.ptx
Normal file
@@ -0,0 +1,519 @@
|
||||
//
|
||||
// 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 __raygen__oxMain
|
||||
.const .align 8 .b8 cs[32];
|
||||
|
||||
.visible .entry __raygen__oxMain(
|
||||
|
||||
)
|
||||
{
|
||||
.reg .pred %p<53>;
|
||||
.reg .b16 %rs<8>;
|
||||
.reg .f32 %f<334>;
|
||||
.reg .b32 %r<42>;
|
||||
.reg .b64 %rd<16>;
|
||||
|
||||
|
||||
// inline asm
|
||||
call (%r6), _optix_get_launch_index_x, ();
|
||||
// inline asm
|
||||
// inline asm
|
||||
call (%r7), _optix_get_launch_index_y, ();
|
||||
// inline asm
|
||||
ld.const.u32 %r9, [cs+28];
|
||||
setp.gt.s32 %p4, %r9, 0;
|
||||
@%p4 bra BB0_2;
|
||||
bra.uni BB0_1;
|
||||
|
||||
BB0_2:
|
||||
ld.const.u64 %rd2, [cs+8];
|
||||
cvta.to.global.u64 %rd8, %rd2;
|
||||
ld.const.u32 %r4, [cs+24];
|
||||
mad.lo.s32 %r11, %r4, %r7, %r6;
|
||||
mul.wide.u32 %rd9, %r11, 16;
|
||||
add.s64 %rd10, %rd8, %rd9;
|
||||
ld.global.v4.f32 {%f55, %f56, %f57, %f58}, [%rd10];
|
||||
mov.f32 %f59, 0f3E68BA2E;
|
||||
cvt.rzi.f32.f32 %f60, %f59;
|
||||
fma.rn.f32 %f61, %f60, 0fC0000000, 0f3EE8BA2E;
|
||||
abs.f32 %f6, %f61;
|
||||
abs.f32 %f8, %f55;
|
||||
setp.lt.f32 %p5, %f8, 0f00800000;
|
||||
mul.f32 %f62, %f8, 0f4B800000;
|
||||
selp.f32 %f63, 0fC3170000, 0fC2FE0000, %p5;
|
||||
selp.f32 %f64, %f62, %f8, %p5;
|
||||
mov.b32 %r12, %f64;
|
||||
and.b32 %r13, %r12, 8388607;
|
||||
or.b32 %r14, %r13, 1065353216;
|
||||
mov.b32 %f65, %r14;
|
||||
shr.u32 %r15, %r12, 23;
|
||||
cvt.rn.f32.u32 %f66, %r15;
|
||||
add.f32 %f67, %f63, %f66;
|
||||
setp.gt.f32 %p6, %f65, 0f3FB504F3;
|
||||
mul.f32 %f68, %f65, 0f3F000000;
|
||||
add.f32 %f69, %f67, 0f3F800000;
|
||||
selp.f32 %f70, %f68, %f65, %p6;
|
||||
selp.f32 %f71, %f69, %f67, %p6;
|
||||
add.f32 %f72, %f70, 0fBF800000;
|
||||
add.f32 %f54, %f70, 0f3F800000;
|
||||
// inline asm
|
||||
rcp.approx.ftz.f32 %f53,%f54;
|
||||
// inline asm
|
||||
add.f32 %f73, %f72, %f72;
|
||||
mul.f32 %f74, %f53, %f73;
|
||||
mul.f32 %f75, %f74, %f74;
|
||||
mov.f32 %f76, 0f3C4CAF63;
|
||||
mov.f32 %f77, 0f3B18F0FE;
|
||||
fma.rn.f32 %f78, %f77, %f75, %f76;
|
||||
mov.f32 %f79, 0f3DAAAABD;
|
||||
fma.rn.f32 %f80, %f78, %f75, %f79;
|
||||
mul.rn.f32 %f81, %f80, %f75;
|
||||
mul.rn.f32 %f82, %f81, %f74;
|
||||
sub.f32 %f83, %f72, %f74;
|
||||
neg.f32 %f84, %f74;
|
||||
add.f32 %f85, %f83, %f83;
|
||||
fma.rn.f32 %f86, %f84, %f72, %f85;
|
||||
mul.rn.f32 %f87, %f53, %f86;
|
||||
add.f32 %f88, %f82, %f74;
|
||||
sub.f32 %f89, %f74, %f88;
|
||||
add.f32 %f90, %f82, %f89;
|
||||
add.f32 %f91, %f87, %f90;
|
||||
add.f32 %f92, %f88, %f91;
|
||||
sub.f32 %f93, %f88, %f92;
|
||||
add.f32 %f94, %f91, %f93;
|
||||
mov.f32 %f95, 0f3F317200;
|
||||
mul.rn.f32 %f96, %f71, %f95;
|
||||
mov.f32 %f97, 0f35BFBE8E;
|
||||
mul.rn.f32 %f98, %f71, %f97;
|
||||
add.f32 %f99, %f96, %f92;
|
||||
sub.f32 %f100, %f96, %f99;
|
||||
add.f32 %f101, %f92, %f100;
|
||||
add.f32 %f102, %f94, %f101;
|
||||
add.f32 %f103, %f98, %f102;
|
||||
add.f32 %f104, %f99, %f103;
|
||||
sub.f32 %f105, %f99, %f104;
|
||||
add.f32 %f106, %f103, %f105;
|
||||
mov.f32 %f107, 0f3EE8BA2E;
|
||||
mul.rn.f32 %f108, %f107, %f104;
|
||||
neg.f32 %f109, %f108;
|
||||
fma.rn.f32 %f110, %f107, %f104, %f109;
|
||||
fma.rn.f32 %f111, %f107, %f106, %f110;
|
||||
mov.f32 %f112, 0f00000000;
|
||||
fma.rn.f32 %f113, %f112, %f104, %f111;
|
||||
add.rn.f32 %f114, %f108, %f113;
|
||||
neg.f32 %f115, %f114;
|
||||
add.rn.f32 %f116, %f108, %f115;
|
||||
add.rn.f32 %f117, %f116, %f113;
|
||||
mov.b32 %r16, %f114;
|
||||
setp.eq.s32 %p7, %r16, 1118925336;
|
||||
add.s32 %r17, %r16, -1;
|
||||
mov.b32 %f118, %r17;
|
||||
add.f32 %f119, %f117, 0f37000000;
|
||||
selp.f32 %f120, %f118, %f114, %p7;
|
||||
selp.f32 %f9, %f119, %f117, %p7;
|
||||
mul.f32 %f121, %f120, 0f3FB8AA3B;
|
||||
cvt.rzi.f32.f32 %f122, %f121;
|
||||
mov.f32 %f123, 0fBF317200;
|
||||
fma.rn.f32 %f124, %f122, %f123, %f120;
|
||||
mov.f32 %f125, 0fB5BFBE8E;
|
||||
fma.rn.f32 %f126, %f122, %f125, %f124;
|
||||
mul.f32 %f127, %f126, 0f3FB8AA3B;
|
||||
ex2.approx.ftz.f32 %f128, %f127;
|
||||
add.f32 %f129, %f122, 0f00000000;
|
||||
ex2.approx.f32 %f130, %f129;
|
||||
mul.f32 %f131, %f128, %f130;
|
||||
setp.lt.f32 %p8, %f120, 0fC2D20000;
|
||||
selp.f32 %f132, 0f00000000, %f131, %p8;
|
||||
setp.gt.f32 %p9, %f120, 0f42D20000;
|
||||
selp.f32 %f322, 0f7F800000, %f132, %p9;
|
||||
setp.eq.f32 %p10, %f322, 0f7F800000;
|
||||
@%p10 bra BB0_4;
|
||||
|
||||
fma.rn.f32 %f322, %f322, %f9, %f322;
|
||||
|
||||
BB0_4:
|
||||
setp.lt.f32 %p11, %f55, 0f00000000;
|
||||
setp.eq.f32 %p12, %f6, 0f3F800000;
|
||||
and.pred %p1, %p11, %p12;
|
||||
mov.b32 %r18, %f322;
|
||||
xor.b32 %r19, %r18, -2147483648;
|
||||
mov.b32 %f133, %r19;
|
||||
selp.f32 %f324, %f133, %f322, %p1;
|
||||
setp.eq.f32 %p13, %f55, 0f00000000;
|
||||
@%p13 bra BB0_7;
|
||||
bra.uni BB0_5;
|
||||
|
||||
BB0_7:
|
||||
add.f32 %f136, %f55, %f55;
|
||||
selp.f32 %f324, %f136, 0f00000000, %p12;
|
||||
bra.uni BB0_8;
|
||||
|
||||
BB0_1:
|
||||
ld.const.u64 %rd4, [cs];
|
||||
cvta.to.global.u64 %rd5, %rd4;
|
||||
ld.const.u32 %r41, [cs+24];
|
||||
mad.lo.s32 %r10, %r41, %r7, %r6;
|
||||
mul.wide.u32 %rd6, %r10, 8;
|
||||
add.s64 %rd7, %rd5, %rd6;
|
||||
ld.global.v4.u16 {%rs4, %rs5, %rs6, %rs7}, [%rd7];
|
||||
// inline asm
|
||||
{ cvt.f32.f16 %f331, %rs4;}
|
||||
|
||||
// inline asm
|
||||
// inline asm
|
||||
{ cvt.f32.f16 %f332, %rs5;}
|
||||
|
||||
// inline asm
|
||||
// inline asm
|
||||
{ cvt.f32.f16 %f333, %rs6;}
|
||||
|
||||
// inline asm
|
||||
ld.const.u64 %rd15, [cs+8];
|
||||
bra.uni BB0_36;
|
||||
|
||||
BB0_5:
|
||||
setp.geu.f32 %p14, %f55, 0f00000000;
|
||||
@%p14 bra BB0_8;
|
||||
|
||||
mov.f32 %f321, 0f3EE8BA2E;
|
||||
cvt.rzi.f32.f32 %f135, %f321;
|
||||
setp.neu.f32 %p15, %f135, 0f3EE8BA2E;
|
||||
selp.f32 %f324, 0f7FFFFFFF, %f324, %p15;
|
||||
|
||||
BB0_8:
|
||||
abs.f32 %f298, %f55;
|
||||
add.f32 %f137, %f298, 0f3EE8BA2E;
|
||||
mov.b32 %r20, %f137;
|
||||
setp.lt.s32 %p17, %r20, 2139095040;
|
||||
@%p17 bra BB0_13;
|
||||
|
||||
abs.f32 %f319, %f55;
|
||||
setp.gtu.f32 %p18, %f319, 0f7F800000;
|
||||
@%p18 bra BB0_12;
|
||||
bra.uni BB0_10;
|
||||
|
||||
BB0_12:
|
||||
add.f32 %f324, %f55, 0f3EE8BA2E;
|
||||
bra.uni BB0_13;
|
||||
|
||||
BB0_10:
|
||||
abs.f32 %f320, %f55;
|
||||
setp.neu.f32 %p19, %f320, 0f7F800000;
|
||||
@%p19 bra BB0_13;
|
||||
|
||||
selp.f32 %f324, 0fFF800000, 0f7F800000, %p1;
|
||||
|
||||
BB0_13:
|
||||
mov.f32 %f307, 0fB5BFBE8E;
|
||||
mov.f32 %f306, 0fBF317200;
|
||||
mov.f32 %f305, 0f00000000;
|
||||
mov.f32 %f304, 0f35BFBE8E;
|
||||
mov.f32 %f303, 0f3F317200;
|
||||
mov.f32 %f302, 0f3DAAAABD;
|
||||
mov.f32 %f301, 0f3C4CAF63;
|
||||
mov.f32 %f300, 0f3B18F0FE;
|
||||
mov.f32 %f299, 0f3EE8BA2E;
|
||||
setp.eq.f32 %p20, %f55, 0f3F800000;
|
||||
selp.f32 %f140, 0f3F800000, %f324, %p20;
|
||||
cvt.sat.f32.f32 %f331, %f140;
|
||||
abs.f32 %f21, %f56;
|
||||
setp.lt.f32 %p21, %f21, 0f00800000;
|
||||
mul.f32 %f141, %f21, 0f4B800000;
|
||||
selp.f32 %f142, 0fC3170000, 0fC2FE0000, %p21;
|
||||
selp.f32 %f143, %f141, %f21, %p21;
|
||||
mov.b32 %r21, %f143;
|
||||
and.b32 %r22, %r21, 8388607;
|
||||
or.b32 %r23, %r22, 1065353216;
|
||||
mov.b32 %f144, %r23;
|
||||
shr.u32 %r24, %r21, 23;
|
||||
cvt.rn.f32.u32 %f145, %r24;
|
||||
add.f32 %f146, %f142, %f145;
|
||||
setp.gt.f32 %p22, %f144, 0f3FB504F3;
|
||||
mul.f32 %f147, %f144, 0f3F000000;
|
||||
add.f32 %f148, %f146, 0f3F800000;
|
||||
selp.f32 %f149, %f147, %f144, %p22;
|
||||
selp.f32 %f150, %f148, %f146, %p22;
|
||||
add.f32 %f151, %f149, 0fBF800000;
|
||||
add.f32 %f139, %f149, 0f3F800000;
|
||||
// inline asm
|
||||
rcp.approx.ftz.f32 %f138,%f139;
|
||||
// inline asm
|
||||
add.f32 %f152, %f151, %f151;
|
||||
mul.f32 %f153, %f138, %f152;
|
||||
mul.f32 %f154, %f153, %f153;
|
||||
fma.rn.f32 %f157, %f300, %f154, %f301;
|
||||
fma.rn.f32 %f159, %f157, %f154, %f302;
|
||||
mul.rn.f32 %f160, %f159, %f154;
|
||||
mul.rn.f32 %f161, %f160, %f153;
|
||||
sub.f32 %f162, %f151, %f153;
|
||||
neg.f32 %f163, %f153;
|
||||
add.f32 %f164, %f162, %f162;
|
||||
fma.rn.f32 %f165, %f163, %f151, %f164;
|
||||
mul.rn.f32 %f166, %f138, %f165;
|
||||
add.f32 %f167, %f161, %f153;
|
||||
sub.f32 %f168, %f153, %f167;
|
||||
add.f32 %f169, %f161, %f168;
|
||||
add.f32 %f170, %f166, %f169;
|
||||
add.f32 %f171, %f167, %f170;
|
||||
sub.f32 %f172, %f167, %f171;
|
||||
add.f32 %f173, %f170, %f172;
|
||||
mul.rn.f32 %f175, %f150, %f303;
|
||||
mul.rn.f32 %f177, %f150, %f304;
|
||||
add.f32 %f178, %f175, %f171;
|
||||
sub.f32 %f179, %f175, %f178;
|
||||
add.f32 %f180, %f171, %f179;
|
||||
add.f32 %f181, %f173, %f180;
|
||||
add.f32 %f182, %f177, %f181;
|
||||
add.f32 %f183, %f178, %f182;
|
||||
sub.f32 %f184, %f178, %f183;
|
||||
add.f32 %f185, %f182, %f184;
|
||||
mul.rn.f32 %f187, %f299, %f183;
|
||||
neg.f32 %f188, %f187;
|
||||
fma.rn.f32 %f189, %f299, %f183, %f188;
|
||||
fma.rn.f32 %f190, %f299, %f185, %f189;
|
||||
fma.rn.f32 %f192, %f305, %f183, %f190;
|
||||
add.rn.f32 %f193, %f187, %f192;
|
||||
neg.f32 %f194, %f193;
|
||||
add.rn.f32 %f195, %f187, %f194;
|
||||
add.rn.f32 %f196, %f195, %f192;
|
||||
mov.b32 %r25, %f193;
|
||||
setp.eq.s32 %p23, %r25, 1118925336;
|
||||
add.s32 %r26, %r25, -1;
|
||||
mov.b32 %f197, %r26;
|
||||
add.f32 %f198, %f196, 0f37000000;
|
||||
selp.f32 %f199, %f197, %f193, %p23;
|
||||
selp.f32 %f22, %f198, %f196, %p23;
|
||||
mul.f32 %f200, %f199, 0f3FB8AA3B;
|
||||
cvt.rzi.f32.f32 %f201, %f200;
|
||||
fma.rn.f32 %f203, %f201, %f306, %f199;
|
||||
fma.rn.f32 %f205, %f201, %f307, %f203;
|
||||
mul.f32 %f206, %f205, 0f3FB8AA3B;
|
||||
ex2.approx.ftz.f32 %f207, %f206;
|
||||
add.f32 %f208, %f201, 0f00000000;
|
||||
ex2.approx.f32 %f209, %f208;
|
||||
mul.f32 %f210, %f207, %f209;
|
||||
setp.lt.f32 %p24, %f199, 0fC2D20000;
|
||||
selp.f32 %f211, 0f00000000, %f210, %p24;
|
||||
setp.gt.f32 %p25, %f199, 0f42D20000;
|
||||
selp.f32 %f325, 0f7F800000, %f211, %p25;
|
||||
setp.eq.f32 %p26, %f325, 0f7F800000;
|
||||
@%p26 bra BB0_15;
|
||||
|
||||
fma.rn.f32 %f325, %f325, %f22, %f325;
|
||||
|
||||
BB0_15:
|
||||
setp.lt.f32 %p27, %f56, 0f00000000;
|
||||
and.pred %p2, %p27, %p12;
|
||||
mov.b32 %r27, %f325;
|
||||
xor.b32 %r28, %r27, -2147483648;
|
||||
mov.b32 %f212, %r28;
|
||||
selp.f32 %f327, %f212, %f325, %p2;
|
||||
setp.eq.f32 %p29, %f56, 0f00000000;
|
||||
@%p29 bra BB0_18;
|
||||
bra.uni BB0_16;
|
||||
|
||||
BB0_18:
|
||||
add.f32 %f215, %f56, %f56;
|
||||
selp.f32 %f327, %f215, 0f00000000, %p12;
|
||||
bra.uni BB0_19;
|
||||
|
||||
BB0_16:
|
||||
setp.geu.f32 %p30, %f56, 0f00000000;
|
||||
@%p30 bra BB0_19;
|
||||
|
||||
mov.f32 %f318, 0f3EE8BA2E;
|
||||
cvt.rzi.f32.f32 %f214, %f318;
|
||||
setp.neu.f32 %p31, %f214, 0f3EE8BA2E;
|
||||
selp.f32 %f327, 0f7FFFFFFF, %f327, %p31;
|
||||
|
||||
BB0_19:
|
||||
add.f32 %f216, %f21, 0f3EE8BA2E;
|
||||
mov.b32 %r29, %f216;
|
||||
setp.lt.s32 %p33, %r29, 2139095040;
|
||||
@%p33 bra BB0_24;
|
||||
|
||||
setp.gtu.f32 %p34, %f21, 0f7F800000;
|
||||
@%p34 bra BB0_23;
|
||||
bra.uni BB0_21;
|
||||
|
||||
BB0_23:
|
||||
add.f32 %f327, %f56, 0f3EE8BA2E;
|
||||
bra.uni BB0_24;
|
||||
|
||||
BB0_21:
|
||||
setp.neu.f32 %p35, %f21, 0f7F800000;
|
||||
@%p35 bra BB0_24;
|
||||
|
||||
selp.f32 %f327, 0fFF800000, 0f7F800000, %p2;
|
||||
|
||||
BB0_24:
|
||||
mov.f32 %f316, 0fB5BFBE8E;
|
||||
mov.f32 %f315, 0fBF317200;
|
||||
mov.f32 %f314, 0f00000000;
|
||||
mov.f32 %f313, 0f35BFBE8E;
|
||||
mov.f32 %f312, 0f3F317200;
|
||||
mov.f32 %f311, 0f3DAAAABD;
|
||||
mov.f32 %f310, 0f3C4CAF63;
|
||||
mov.f32 %f309, 0f3B18F0FE;
|
||||
mov.f32 %f308, 0f3EE8BA2E;
|
||||
setp.eq.f32 %p36, %f56, 0f3F800000;
|
||||
selp.f32 %f219, 0f3F800000, %f327, %p36;
|
||||
cvt.sat.f32.f32 %f332, %f219;
|
||||
abs.f32 %f34, %f57;
|
||||
setp.lt.f32 %p37, %f34, 0f00800000;
|
||||
mul.f32 %f220, %f34, 0f4B800000;
|
||||
selp.f32 %f221, 0fC3170000, 0fC2FE0000, %p37;
|
||||
selp.f32 %f222, %f220, %f34, %p37;
|
||||
mov.b32 %r30, %f222;
|
||||
and.b32 %r31, %r30, 8388607;
|
||||
or.b32 %r32, %r31, 1065353216;
|
||||
mov.b32 %f223, %r32;
|
||||
shr.u32 %r33, %r30, 23;
|
||||
cvt.rn.f32.u32 %f224, %r33;
|
||||
add.f32 %f225, %f221, %f224;
|
||||
setp.gt.f32 %p38, %f223, 0f3FB504F3;
|
||||
mul.f32 %f226, %f223, 0f3F000000;
|
||||
add.f32 %f227, %f225, 0f3F800000;
|
||||
selp.f32 %f228, %f226, %f223, %p38;
|
||||
selp.f32 %f229, %f227, %f225, %p38;
|
||||
add.f32 %f230, %f228, 0fBF800000;
|
||||
add.f32 %f218, %f228, 0f3F800000;
|
||||
// inline asm
|
||||
rcp.approx.ftz.f32 %f217,%f218;
|
||||
// inline asm
|
||||
add.f32 %f231, %f230, %f230;
|
||||
mul.f32 %f232, %f217, %f231;
|
||||
mul.f32 %f233, %f232, %f232;
|
||||
fma.rn.f32 %f236, %f309, %f233, %f310;
|
||||
fma.rn.f32 %f238, %f236, %f233, %f311;
|
||||
mul.rn.f32 %f239, %f238, %f233;
|
||||
mul.rn.f32 %f240, %f239, %f232;
|
||||
sub.f32 %f241, %f230, %f232;
|
||||
neg.f32 %f242, %f232;
|
||||
add.f32 %f243, %f241, %f241;
|
||||
fma.rn.f32 %f244, %f242, %f230, %f243;
|
||||
mul.rn.f32 %f245, %f217, %f244;
|
||||
add.f32 %f246, %f240, %f232;
|
||||
sub.f32 %f247, %f232, %f246;
|
||||
add.f32 %f248, %f240, %f247;
|
||||
add.f32 %f249, %f245, %f248;
|
||||
add.f32 %f250, %f246, %f249;
|
||||
sub.f32 %f251, %f246, %f250;
|
||||
add.f32 %f252, %f249, %f251;
|
||||
mul.rn.f32 %f254, %f229, %f312;
|
||||
mul.rn.f32 %f256, %f229, %f313;
|
||||
add.f32 %f257, %f254, %f250;
|
||||
sub.f32 %f258, %f254, %f257;
|
||||
add.f32 %f259, %f250, %f258;
|
||||
add.f32 %f260, %f252, %f259;
|
||||
add.f32 %f261, %f256, %f260;
|
||||
add.f32 %f262, %f257, %f261;
|
||||
sub.f32 %f263, %f257, %f262;
|
||||
add.f32 %f264, %f261, %f263;
|
||||
mul.rn.f32 %f266, %f308, %f262;
|
||||
neg.f32 %f267, %f266;
|
||||
fma.rn.f32 %f268, %f308, %f262, %f267;
|
||||
fma.rn.f32 %f269, %f308, %f264, %f268;
|
||||
fma.rn.f32 %f271, %f314, %f262, %f269;
|
||||
add.rn.f32 %f272, %f266, %f271;
|
||||
neg.f32 %f273, %f272;
|
||||
add.rn.f32 %f274, %f266, %f273;
|
||||
add.rn.f32 %f275, %f274, %f271;
|
||||
mov.b32 %r34, %f272;
|
||||
setp.eq.s32 %p39, %r34, 1118925336;
|
||||
add.s32 %r35, %r34, -1;
|
||||
mov.b32 %f276, %r35;
|
||||
add.f32 %f277, %f275, 0f37000000;
|
||||
selp.f32 %f278, %f276, %f272, %p39;
|
||||
selp.f32 %f35, %f277, %f275, %p39;
|
||||
mul.f32 %f279, %f278, 0f3FB8AA3B;
|
||||
cvt.rzi.f32.f32 %f280, %f279;
|
||||
fma.rn.f32 %f282, %f280, %f315, %f278;
|
||||
fma.rn.f32 %f284, %f280, %f316, %f282;
|
||||
mul.f32 %f285, %f284, 0f3FB8AA3B;
|
||||
ex2.approx.ftz.f32 %f286, %f285;
|
||||
add.f32 %f287, %f280, 0f00000000;
|
||||
ex2.approx.f32 %f288, %f287;
|
||||
mul.f32 %f289, %f286, %f288;
|
||||
setp.lt.f32 %p40, %f278, 0fC2D20000;
|
||||
selp.f32 %f290, 0f00000000, %f289, %p40;
|
||||
setp.gt.f32 %p41, %f278, 0f42D20000;
|
||||
selp.f32 %f328, 0f7F800000, %f290, %p41;
|
||||
setp.eq.f32 %p42, %f328, 0f7F800000;
|
||||
@%p42 bra BB0_26;
|
||||
|
||||
fma.rn.f32 %f328, %f328, %f35, %f328;
|
||||
|
||||
BB0_26:
|
||||
setp.lt.f32 %p43, %f57, 0f00000000;
|
||||
and.pred %p3, %p43, %p12;
|
||||
mov.b32 %r36, %f328;
|
||||
xor.b32 %r37, %r36, -2147483648;
|
||||
mov.b32 %f291, %r37;
|
||||
selp.f32 %f330, %f291, %f328, %p3;
|
||||
setp.eq.f32 %p45, %f57, 0f00000000;
|
||||
@%p45 bra BB0_29;
|
||||
bra.uni BB0_27;
|
||||
|
||||
BB0_29:
|
||||
add.f32 %f294, %f57, %f57;
|
||||
selp.f32 %f330, %f294, 0f00000000, %p12;
|
||||
bra.uni BB0_30;
|
||||
|
||||
BB0_27:
|
||||
setp.geu.f32 %p46, %f57, 0f00000000;
|
||||
@%p46 bra BB0_30;
|
||||
|
||||
mov.f32 %f317, 0f3EE8BA2E;
|
||||
cvt.rzi.f32.f32 %f293, %f317;
|
||||
setp.neu.f32 %p47, %f293, 0f3EE8BA2E;
|
||||
selp.f32 %f330, 0f7FFFFFFF, %f330, %p47;
|
||||
|
||||
BB0_30:
|
||||
add.f32 %f295, %f34, 0f3EE8BA2E;
|
||||
mov.b32 %r38, %f295;
|
||||
setp.lt.s32 %p49, %r38, 2139095040;
|
||||
@%p49 bra BB0_35;
|
||||
|
||||
setp.gtu.f32 %p50, %f34, 0f7F800000;
|
||||
@%p50 bra BB0_34;
|
||||
bra.uni BB0_32;
|
||||
|
||||
BB0_34:
|
||||
add.f32 %f330, %f57, 0f3EE8BA2E;
|
||||
bra.uni BB0_35;
|
||||
|
||||
BB0_32:
|
||||
setp.neu.f32 %p51, %f34, 0f7F800000;
|
||||
@%p51 bra BB0_35;
|
||||
|
||||
selp.f32 %f330, 0fFF800000, 0f7F800000, %p3;
|
||||
|
||||
BB0_35:
|
||||
ld.const.u64 %rd15, [cs+8];
|
||||
ld.const.u32 %r41, [cs+24];
|
||||
setp.eq.f32 %p52, %f57, 0f3F800000;
|
||||
selp.f32 %f296, 0f3F800000, %f330, %p52;
|
||||
cvt.sat.f32.f32 %f333, %f296;
|
||||
|
||||
BB0_36:
|
||||
mad.lo.s32 %r39, %r41, %r7, %r6;
|
||||
cvta.to.global.u64 %rd11, %rd15;
|
||||
mul.wide.u32 %rd12, %r39, 16;
|
||||
add.s64 %rd13, %rd11, %rd12;
|
||||
mov.f32 %f297, 0f3F800000;
|
||||
st.global.v4.f32 [%rd13], {%f331, %f332, %f333, %f297};
|
||||
ret;
|
||||
}
|
||||
|
||||
|
||||
8
Assets/Editor/x64/Bakery/denoisePrepareSH72.ptx.meta
Normal file
8
Assets/Editor/x64/Bakery/denoisePrepareSH72.ptx.meta
Normal file
@@ -0,0 +1,8 @@
|
||||
fileFormatVersion: 2
|
||||
guid: cf96e815cf5f4354a9ccd4524316bf97
|
||||
timeCreated: 1605376410
|
||||
licenseType: Store
|
||||
DefaultImporter:
|
||||
userData:
|
||||
assetBundleName:
|
||||
assetBundleVariant:
|
||||
159
Assets/Editor/x64/Bakery/denoisePrepareSH_OIDN.ptx
Normal file
159
Assets/Editor/x64/Bakery/denoisePrepareSH_OIDN.ptx
Normal file
@@ -0,0 +1,159 @@
|
||||
//
|
||||
// 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 1 .b8 input_buffer[1];
|
||||
.global .align 1 .b8 image[1];
|
||||
.global .align 4 .u32 mode;
|
||||
.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_typeinfo4modeE[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 4 .b8 _ZN21rti_internal_typename4modeE[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_typeenum4modeE = 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 1 .b8 _ZN21rti_internal_semantic4modeE[1];
|
||||
.global .align 1 .b8 _ZN23rti_internal_annotation7pixelIDE[1];
|
||||
.global .align 1 .b8 _ZN23rti_internal_annotation10resolutionE[1];
|
||||
.global .align 1 .b8 _ZN23rti_internal_annotation4modeE[1];
|
||||
|
||||
.visible .entry _Z6oxMainv(
|
||||
|
||||
)
|
||||
{
|
||||
.reg .pred %p<2>;
|
||||
.reg .b16 %rs<4>;
|
||||
.reg .f32 %f<16>;
|
||||
.reg .b32 %r<40>;
|
||||
.reg .b64 %rd<48>;
|
||||
|
||||
|
||||
ld.global.u32 %r1, [mode];
|
||||
setp.gt.s32 %p1, %r1, 0;
|
||||
ld.global.v2.u32 {%r2, %r3}, [pixelID];
|
||||
cvt.u64.u32 %rd1, %r2;
|
||||
cvt.u64.u32 %rd2, %r3;
|
||||
@%p1 bra BB0_2;
|
||||
bra.uni BB0_1;
|
||||
|
||||
BB0_2:
|
||||
mov.u64 %rd40, input_buffer;
|
||||
cvta.global.u64 %rd23, %rd40;
|
||||
mov.u32 %r24, 2;
|
||||
mov.u32 %r25, 12;
|
||||
mov.u64 %rd39, 0;
|
||||
// inline asm
|
||||
call (%rd22), _rt_buffer_get_64, (%rd23, %r24, %r25, %rd1, %rd2, %rd39, %rd39);
|
||||
// inline asm
|
||||
ld.f32 %f13, [%rd22];
|
||||
ld.global.v2.u32 {%r26, %r27}, [pixelID];
|
||||
cvt.u64.u32 %rd30, %r26;
|
||||
cvt.u64.u32 %rd31, %r27;
|
||||
// inline asm
|
||||
call (%rd28), _rt_buffer_get_64, (%rd23, %r24, %r25, %rd30, %rd31, %rd39, %rd39);
|
||||
// inline asm
|
||||
ld.f32 %f14, [%rd28+4];
|
||||
ld.global.v2.u32 {%r30, %r31}, [pixelID];
|
||||
cvt.u64.u32 %rd36, %r30;
|
||||
cvt.u64.u32 %rd37, %r31;
|
||||
// inline asm
|
||||
call (%rd34), _rt_buffer_get_64, (%rd23, %r24, %r25, %rd36, %rd37, %rd39, %rd39);
|
||||
// inline asm
|
||||
ld.f32 %f15, [%rd34+8];
|
||||
bra.uni BB0_3;
|
||||
|
||||
BB0_1:
|
||||
mov.u64 %rd21, image;
|
||||
cvta.global.u64 %rd4, %rd21;
|
||||
mov.u32 %r10, 2;
|
||||
mov.u32 %r11, 8;
|
||||
mov.u64 %rd20, 0;
|
||||
// inline asm
|
||||
call (%rd3), _rt_buffer_get_64, (%rd4, %r10, %r11, %rd1, %rd2, %rd20, %rd20);
|
||||
// inline asm
|
||||
ld.u16 %rs1, [%rd3];
|
||||
// inline asm
|
||||
{ cvt.f32.f16 %f13, %rs1;}
|
||||
|
||||
// inline asm
|
||||
ld.global.v2.u32 {%r12, %r13}, [pixelID];
|
||||
cvt.u64.u32 %rd11, %r12;
|
||||
cvt.u64.u32 %rd12, %r13;
|
||||
// inline asm
|
||||
call (%rd9), _rt_buffer_get_64, (%rd4, %r10, %r11, %rd11, %rd12, %rd20, %rd20);
|
||||
// inline asm
|
||||
ld.u16 %rs2, [%rd9+2];
|
||||
// inline asm
|
||||
{ cvt.f32.f16 %f14, %rs2;}
|
||||
|
||||
// inline asm
|
||||
ld.global.v2.u32 {%r16, %r17}, [pixelID];
|
||||
cvt.u64.u32 %rd17, %r16;
|
||||
cvt.u64.u32 %rd18, %r17;
|
||||
// inline asm
|
||||
call (%rd15), _rt_buffer_get_64, (%rd4, %r10, %r11, %rd17, %rd18, %rd20, %rd20);
|
||||
// inline asm
|
||||
ld.u16 %rs3, [%rd15+4];
|
||||
// inline asm
|
||||
{ cvt.f32.f16 %f15, %rs3;}
|
||||
|
||||
// inline asm
|
||||
|
||||
BB0_3:
|
||||
ld.global.v2.u32 {%r36, %r37}, [pixelID];
|
||||
cvt.u64.u32 %rd43, %r36;
|
||||
cvt.u64.u32 %rd44, %r37;
|
||||
mov.u64 %rd47, input_buffer;
|
||||
cvta.global.u64 %rd42, %rd47;
|
||||
mov.u32 %r34, 2;
|
||||
mov.u32 %r35, 12;
|
||||
mov.u64 %rd46, 0;
|
||||
// inline asm
|
||||
call (%rd41), _rt_buffer_get_64, (%rd42, %r34, %r35, %rd43, %rd44, %rd46, %rd46);
|
||||
// inline asm
|
||||
st.f32 [%rd41+8], %f15;
|
||||
st.f32 [%rd41+4], %f14;
|
||||
st.f32 [%rd41], %f13;
|
||||
ret;
|
||||
}
|
||||
|
||||
|
||||
8
Assets/Editor/x64/Bakery/denoisePrepareSH_OIDN.ptx.meta
Normal file
8
Assets/Editor/x64/Bakery/denoisePrepareSH_OIDN.ptx.meta
Normal file
@@ -0,0 +1,8 @@
|
||||
fileFormatVersion: 2
|
||||
guid: f763370827bb0e94b89c753d02bf676d
|
||||
timeCreated: 1599897795
|
||||
licenseType: Store
|
||||
DefaultImporter:
|
||||
userData:
|
||||
assetBundleName:
|
||||
assetBundleVariant:
|
||||
BIN
Assets/Editor/x64/Bakery/denoiser.exe
Normal file
BIN
Assets/Editor/x64/Bakery/denoiser.exe
Normal file
Binary file not shown.
8
Assets/Editor/x64/Bakery/denoiser.exe.meta
Normal file
8
Assets/Editor/x64/Bakery/denoiser.exe.meta
Normal file
@@ -0,0 +1,8 @@
|
||||
fileFormatVersion: 2
|
||||
guid: 00324aa8cc137c44cb8bdb89b0b85076
|
||||
timeCreated: 1562931708
|
||||
licenseType: Store
|
||||
DefaultImporter:
|
||||
userData:
|
||||
assetBundleName:
|
||||
assetBundleVariant:
|
||||
BIN
Assets/Editor/x64/Bakery/denoiser72.exe
Normal file
BIN
Assets/Editor/x64/Bakery/denoiser72.exe
Normal file
Binary file not shown.
8
Assets/Editor/x64/Bakery/denoiser72.exe.meta
Normal file
8
Assets/Editor/x64/Bakery/denoiser72.exe.meta
Normal file
@@ -0,0 +1,8 @@
|
||||
fileFormatVersion: 2
|
||||
guid: 923624ea465edaf4e8fbde0e3317fa96
|
||||
timeCreated: 1605206208
|
||||
licenseType: Store
|
||||
DefaultImporter:
|
||||
userData:
|
||||
assetBundleName:
|
||||
assetBundleVariant:
|
||||
BIN
Assets/Editor/x64/Bakery/denoiserLegacy.exe
Normal file
BIN
Assets/Editor/x64/Bakery/denoiserLegacy.exe
Normal file
Binary file not shown.
8
Assets/Editor/x64/Bakery/denoiserLegacy.exe.meta
Normal file
8
Assets/Editor/x64/Bakery/denoiserLegacy.exe.meta
Normal file
@@ -0,0 +1,8 @@
|
||||
fileFormatVersion: 2
|
||||
guid: e835870912302c34abd30850e21d1f0b
|
||||
timeCreated: 1573289338
|
||||
licenseType: Store
|
||||
DefaultImporter:
|
||||
userData:
|
||||
assetBundleName:
|
||||
assetBundleVariant:
|
||||
BIN
Assets/Editor/x64/Bakery/denoiserOIDN.exe
Normal file
BIN
Assets/Editor/x64/Bakery/denoiserOIDN.exe
Normal file
Binary file not shown.
8
Assets/Editor/x64/Bakery/denoiserOIDN.exe.meta
Normal file
8
Assets/Editor/x64/Bakery/denoiserOIDN.exe.meta
Normal file
@@ -0,0 +1,8 @@
|
||||
fileFormatVersion: 2
|
||||
guid: 7c5044bbfa3bae54196b28dca3529e39
|
||||
timeCreated: 1605444920
|
||||
licenseType: Store
|
||||
DefaultImporter:
|
||||
userData:
|
||||
assetBundleName:
|
||||
assetBundleVariant:
|
||||
BIN
Assets/Editor/x64/Bakery/dilate.pso
Normal file
BIN
Assets/Editor/x64/Bakery/dilate.pso
Normal file
Binary file not shown.
8
Assets/Editor/x64/Bakery/dilate.pso.meta
Normal file
8
Assets/Editor/x64/Bakery/dilate.pso.meta
Normal file
@@ -0,0 +1,8 @@
|
||||
fileFormatVersion: 2
|
||||
guid: e5bdfd4e6f60ffb438e29f5ce906e4c0
|
||||
timeCreated: 1529751953
|
||||
licenseType: Store
|
||||
DefaultImporter:
|
||||
userData:
|
||||
assetBundleName:
|
||||
assetBundleVariant:
|
||||
473
Assets/Editor/x64/Bakery/dilate.ptx
Normal file
473
Assets/Editor/x64/Bakery/dilate.ptx
Normal file
@@ -0,0 +1,473 @@
|
||||
//
|
||||
// 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 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 8 .u64 _ZN21rti_internal_register20reg_bitness_detectorE;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail0E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail1E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail2E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail3E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail4E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail5E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail6E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail7E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail8E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail9E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail0E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail1E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail2E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail3E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail4E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail5E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail6E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail7E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail8E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail9E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register14reg_rayIndex_xE;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register14reg_rayIndex_yE;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register14reg_rayIndex_zE;
|
||||
.global .align 8 .b8 _ZN21rti_internal_typename7pixelIDE[6] = {117, 105, 110, 116, 50, 0};
|
||||
.global .align 8 .b8 _ZN21rti_internal_typename10resolutionE[6] = {117, 105, 110, 116, 50, 0};
|
||||
.global .align 8 .b8 _ZN21rti_internal_typename6normalE[7] = {102, 108, 111, 97, 116, 51, 0};
|
||||
.global .align 8 .b8 _ZN21rti_internal_typename6camPosE[7] = {102, 108, 111, 97, 116, 51, 0};
|
||||
.global .align 16 .b8 _ZN21rti_internal_typename4rootE[9] = {114, 116, 79, 98, 106, 101, 99, 116, 0};
|
||||
.global .align 4 .b8 _ZN21rti_internal_typename12imageEnabledE[4] = {105, 110, 116, 0};
|
||||
.global .align 8 .b8 _ZN21rti_internal_typename8tileInfoE[6] = {117, 105, 110, 116, 52, 0};
|
||||
.global .align 4 .b8 _ZN21rti_internal_typename8additiveE[4] = {105, 110, 116, 0};
|
||||
.global .align 4 .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 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 _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];
|
||||
|
||||
.visible .entry _Z6oxMainv(
|
||||
|
||||
)
|
||||
{
|
||||
.reg .pred %p<14>;
|
||||
.reg .b16 %rs<45>;
|
||||
.reg .f32 %f<159>;
|
||||
.reg .b32 %r<149>;
|
||||
.reg .b64 %rd<271>;
|
||||
|
||||
|
||||
ld.global.v2.u32 {%r18, %r19}, [pixelID];
|
||||
cvt.u64.u32 %rd9, %r18;
|
||||
cvt.u64.u32 %rd10, %r19;
|
||||
mov.u64 %rd13, image;
|
||||
cvta.global.u64 %rd8, %rd13;
|
||||
mov.u32 %r16, 2;
|
||||
mov.u32 %r17, 4;
|
||||
mov.u64 %rd12, 0;
|
||||
// inline asm
|
||||
call (%rd7), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd9, %rd10, %rd12, %rd12);
|
||||
// inline asm
|
||||
ld.u8 %rs1, [%rd7+3];
|
||||
setp.ne.s16 %p1, %rs1, 0;
|
||||
@%p1 bra BB0_17;
|
||||
|
||||
ld.global.v2.u32 {%r33, %r34}, [pixelID];
|
||||
cvt.u64.u32 %rd16, %r33;
|
||||
cvt.u64.u32 %rd17, %r34;
|
||||
// inline asm
|
||||
call (%rd14), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd16, %rd17, %rd12, %rd12);
|
||||
// inline asm
|
||||
ld.u8 %rs2, [%rd14+2];
|
||||
cvt.rn.f32.u16 %f61, %rs2;
|
||||
div.rn.f32 %f135, %f61, 0f437F0000;
|
||||
ld.global.v2.u32 {%r37, %r38}, [pixelID];
|
||||
cvt.u64.u32 %rd22, %r37;
|
||||
cvt.u64.u32 %rd23, %r38;
|
||||
// inline asm
|
||||
call (%rd20), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd22, %rd23, %rd12, %rd12);
|
||||
// inline asm
|
||||
ld.u8 %rs3, [%rd20+1];
|
||||
cvt.rn.f32.u16 %f62, %rs3;
|
||||
div.rn.f32 %f136, %f62, 0f437F0000;
|
||||
ld.global.v2.u32 {%r41, %r42}, [pixelID];
|
||||
cvt.u64.u32 %rd28, %r41;
|
||||
cvt.u64.u32 %rd29, %r42;
|
||||
// inline asm
|
||||
call (%rd26), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd28, %rd29, %rd12, %rd12);
|
||||
// inline asm
|
||||
ld.u8 %rs4, [%rd26];
|
||||
cvt.rn.f32.u16 %f63, %rs4;
|
||||
div.rn.f32 %f137, %f63, 0f437F0000;
|
||||
ld.global.v2.u32 {%r45, %r46}, [pixelID];
|
||||
cvt.u64.u32 %rd34, %r45;
|
||||
cvt.u64.u32 %rd35, %r46;
|
||||
// inline asm
|
||||
call (%rd32), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd34, %rd35, %rd12, %rd12);
|
||||
// inline asm
|
||||
ld.u8 %rs5, [%rd32+3];
|
||||
cvt.rn.f32.u16 %f64, %rs5;
|
||||
div.rn.f32 %f138, %f64, 0f437F0000;
|
||||
ld.global.v2.u32 {%r49, %r50}, [pixelID];
|
||||
setp.eq.s32 %p2, %r49, 0;
|
||||
add.s32 %r52, %r49, -1;
|
||||
cvt.u64.u32 %rd45, %r52;
|
||||
selp.b64 %rd40, 0, %rd45, %p2;
|
||||
setp.eq.s32 %p3, %r50, 0;
|
||||
add.s32 %r54, %r50, -1;
|
||||
cvt.u64.u32 %rd46, %r54;
|
||||
selp.b64 %rd41, 0, %rd46, %p3;
|
||||
ld.global.v2.u32 {%r55, %r56}, [resolution];
|
||||
add.s32 %r58, %r55, -1;
|
||||
setp.eq.s32 %p4, %r49, %r58;
|
||||
add.s32 %r59, %r49, 1;
|
||||
selp.b32 %r1, %r58, %r59, %p4;
|
||||
add.s32 %r61, %r56, -1;
|
||||
setp.eq.s32 %p5, %r50, %r61;
|
||||
add.s32 %r62, %r50, 1;
|
||||
selp.b32 %r2, %r61, %r62, %p5;
|
||||
// inline asm
|
||||
call (%rd38), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd40, %rd41, %rd12, %rd12);
|
||||
// inline asm
|
||||
ld.u8 %rs6, [%rd38+3];
|
||||
setp.eq.s16 %p6, %rs6, 0;
|
||||
mov.u32 %r143, 0;
|
||||
@%p6 bra BB0_3;
|
||||
|
||||
// inline asm
|
||||
call (%rd47), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd40, %rd41, %rd12, %rd12);
|
||||
// inline asm
|
||||
ld.u8 %rs7, [%rd47+2];
|
||||
cvt.rn.f32.u16 %f65, %rs7;
|
||||
div.rn.f32 %f66, %f65, 0f437F0000;
|
||||
// inline asm
|
||||
call (%rd53), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd40, %rd41, %rd12, %rd12);
|
||||
// inline asm
|
||||
ld.u8 %rs8, [%rd53+1];
|
||||
cvt.rn.f32.u16 %f67, %rs8;
|
||||
div.rn.f32 %f68, %f67, 0f437F0000;
|
||||
// inline asm
|
||||
call (%rd59), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd40, %rd41, %rd12, %rd12);
|
||||
// inline asm
|
||||
ld.u8 %rs9, [%rd59];
|
||||
cvt.rn.f32.u16 %f69, %rs9;
|
||||
div.rn.f32 %f70, %f69, 0f437F0000;
|
||||
// inline asm
|
||||
call (%rd65), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd40, %rd41, %rd12, %rd12);
|
||||
// inline asm
|
||||
ld.u8 %rs10, [%rd65+3];
|
||||
cvt.rn.f32.u16 %f71, %rs10;
|
||||
div.rn.f32 %f72, %f71, 0f437F0000;
|
||||
add.f32 %f135, %f135, %f66;
|
||||
add.f32 %f136, %f136, %f68;
|
||||
add.f32 %f137, %f137, %f70;
|
||||
add.f32 %f138, %f138, %f72;
|
||||
mov.u32 %r143, 1;
|
||||
|
||||
BB0_3:
|
||||
ld.global.u32 %rd74, [pixelID];
|
||||
// inline asm
|
||||
call (%rd72), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd74, %rd41, %rd12, %rd12);
|
||||
// inline asm
|
||||
ld.u8 %rs11, [%rd72+3];
|
||||
setp.eq.s16 %p7, %rs11, 0;
|
||||
@%p7 bra BB0_5;
|
||||
|
||||
// inline asm
|
||||
call (%rd79), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd74, %rd41, %rd12, %rd12);
|
||||
// inline asm
|
||||
ld.u8 %rs12, [%rd79+2];
|
||||
cvt.rn.f32.u16 %f73, %rs12;
|
||||
div.rn.f32 %f74, %f73, 0f437F0000;
|
||||
// inline asm
|
||||
call (%rd85), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd74, %rd41, %rd12, %rd12);
|
||||
// inline asm
|
||||
ld.u8 %rs13, [%rd85+1];
|
||||
cvt.rn.f32.u16 %f75, %rs13;
|
||||
div.rn.f32 %f76, %f75, 0f437F0000;
|
||||
// inline asm
|
||||
call (%rd91), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd74, %rd41, %rd12, %rd12);
|
||||
// inline asm
|
||||
ld.u8 %rs14, [%rd91];
|
||||
cvt.rn.f32.u16 %f77, %rs14;
|
||||
div.rn.f32 %f78, %f77, 0f437F0000;
|
||||
// inline asm
|
||||
call (%rd97), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd74, %rd41, %rd12, %rd12);
|
||||
// inline asm
|
||||
ld.u8 %rs15, [%rd97+3];
|
||||
cvt.rn.f32.u16 %f79, %rs15;
|
||||
div.rn.f32 %f80, %f79, 0f437F0000;
|
||||
add.f32 %f135, %f135, %f74;
|
||||
add.f32 %f136, %f136, %f76;
|
||||
add.f32 %f137, %f137, %f78;
|
||||
add.f32 %f138, %f138, %f80;
|
||||
add.s32 %r143, %r143, 1;
|
||||
|
||||
BB0_5:
|
||||
cvt.u64.u32 %rd106, %r1;
|
||||
// inline asm
|
||||
call (%rd104), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd106, %rd41, %rd12, %rd12);
|
||||
// inline asm
|
||||
ld.u8 %rs16, [%rd104+3];
|
||||
setp.eq.s16 %p8, %rs16, 0;
|
||||
@%p8 bra BB0_7;
|
||||
|
||||
// inline asm
|
||||
call (%rd111), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd106, %rd41, %rd12, %rd12);
|
||||
// inline asm
|
||||
ld.u8 %rs17, [%rd111+2];
|
||||
cvt.rn.f32.u16 %f81, %rs17;
|
||||
div.rn.f32 %f82, %f81, 0f437F0000;
|
||||
// inline asm
|
||||
call (%rd117), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd106, %rd41, %rd12, %rd12);
|
||||
// inline asm
|
||||
ld.u8 %rs18, [%rd117+1];
|
||||
cvt.rn.f32.u16 %f83, %rs18;
|
||||
div.rn.f32 %f84, %f83, 0f437F0000;
|
||||
// inline asm
|
||||
call (%rd123), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd106, %rd41, %rd12, %rd12);
|
||||
// inline asm
|
||||
ld.u8 %rs19, [%rd123];
|
||||
cvt.rn.f32.u16 %f85, %rs19;
|
||||
div.rn.f32 %f86, %f85, 0f437F0000;
|
||||
// inline asm
|
||||
call (%rd129), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd106, %rd41, %rd12, %rd12);
|
||||
// inline asm
|
||||
ld.u8 %rs20, [%rd129+3];
|
||||
cvt.rn.f32.u16 %f87, %rs20;
|
||||
div.rn.f32 %f88, %f87, 0f437F0000;
|
||||
add.f32 %f135, %f135, %f82;
|
||||
add.f32 %f136, %f136, %f84;
|
||||
add.f32 %f137, %f137, %f86;
|
||||
add.f32 %f138, %f138, %f88;
|
||||
add.s32 %r143, %r143, 1;
|
||||
|
||||
BB0_7:
|
||||
ld.global.u32 %rd139, [pixelID+4];
|
||||
// inline asm
|
||||
call (%rd136), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd40, %rd139, %rd12, %rd12);
|
||||
// inline asm
|
||||
ld.u8 %rs21, [%rd136+3];
|
||||
setp.eq.s16 %p9, %rs21, 0;
|
||||
@%p9 bra BB0_9;
|
||||
|
||||
// inline asm
|
||||
call (%rd143), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd40, %rd139, %rd12, %rd12);
|
||||
// inline asm
|
||||
ld.u8 %rs22, [%rd143+2];
|
||||
cvt.rn.f32.u16 %f89, %rs22;
|
||||
div.rn.f32 %f90, %f89, 0f437F0000;
|
||||
// inline asm
|
||||
call (%rd149), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd40, %rd139, %rd12, %rd12);
|
||||
// inline asm
|
||||
ld.u8 %rs23, [%rd149+1];
|
||||
cvt.rn.f32.u16 %f91, %rs23;
|
||||
div.rn.f32 %f92, %f91, 0f437F0000;
|
||||
// inline asm
|
||||
call (%rd155), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd40, %rd139, %rd12, %rd12);
|
||||
// inline asm
|
||||
ld.u8 %rs24, [%rd155];
|
||||
cvt.rn.f32.u16 %f93, %rs24;
|
||||
div.rn.f32 %f94, %f93, 0f437F0000;
|
||||
// inline asm
|
||||
call (%rd161), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd40, %rd139, %rd12, %rd12);
|
||||
// inline asm
|
||||
ld.u8 %rs25, [%rd161+3];
|
||||
cvt.rn.f32.u16 %f95, %rs25;
|
||||
div.rn.f32 %f96, %f95, 0f437F0000;
|
||||
add.f32 %f135, %f135, %f90;
|
||||
add.f32 %f136, %f136, %f92;
|
||||
add.f32 %f137, %f137, %f94;
|
||||
add.f32 %f138, %f138, %f96;
|
||||
add.s32 %r143, %r143, 1;
|
||||
|
||||
BB0_9:
|
||||
// inline asm
|
||||
call (%rd168), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd106, %rd139, %rd12, %rd12);
|
||||
// inline asm
|
||||
ld.u8 %rs26, [%rd168+3];
|
||||
setp.eq.s16 %p10, %rs26, 0;
|
||||
@%p10 bra BB0_11;
|
||||
|
||||
// inline asm
|
||||
call (%rd175), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd106, %rd139, %rd12, %rd12);
|
||||
// inline asm
|
||||
ld.u8 %rs27, [%rd175+2];
|
||||
cvt.rn.f32.u16 %f97, %rs27;
|
||||
div.rn.f32 %f98, %f97, 0f437F0000;
|
||||
// inline asm
|
||||
call (%rd181), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd106, %rd139, %rd12, %rd12);
|
||||
// inline asm
|
||||
ld.u8 %rs28, [%rd181+1];
|
||||
cvt.rn.f32.u16 %f99, %rs28;
|
||||
div.rn.f32 %f100, %f99, 0f437F0000;
|
||||
// inline asm
|
||||
call (%rd187), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd106, %rd139, %rd12, %rd12);
|
||||
// inline asm
|
||||
ld.u8 %rs29, [%rd187];
|
||||
cvt.rn.f32.u16 %f101, %rs29;
|
||||
div.rn.f32 %f102, %f101, 0f437F0000;
|
||||
// inline asm
|
||||
call (%rd193), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd106, %rd139, %rd12, %rd12);
|
||||
// inline asm
|
||||
ld.u8 %rs30, [%rd193+3];
|
||||
cvt.rn.f32.u16 %f103, %rs30;
|
||||
div.rn.f32 %f104, %f103, 0f437F0000;
|
||||
add.f32 %f135, %f135, %f98;
|
||||
add.f32 %f136, %f136, %f100;
|
||||
add.f32 %f137, %f137, %f102;
|
||||
add.f32 %f138, %f138, %f104;
|
||||
add.s32 %r143, %r143, 1;
|
||||
|
||||
BB0_11:
|
||||
cvt.u64.u32 %rd203, %r2;
|
||||
// inline asm
|
||||
call (%rd200), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd40, %rd203, %rd12, %rd12);
|
||||
// inline asm
|
||||
ld.u8 %rs31, [%rd200+3];
|
||||
setp.eq.s16 %p11, %rs31, 0;
|
||||
@%p11 bra BB0_13;
|
||||
|
||||
// inline asm
|
||||
call (%rd207), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd40, %rd203, %rd12, %rd12);
|
||||
// inline asm
|
||||
ld.u8 %rs32, [%rd207+2];
|
||||
cvt.rn.f32.u16 %f105, %rs32;
|
||||
div.rn.f32 %f106, %f105, 0f437F0000;
|
||||
// inline asm
|
||||
call (%rd213), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd40, %rd203, %rd12, %rd12);
|
||||
// inline asm
|
||||
ld.u8 %rs33, [%rd213+1];
|
||||
cvt.rn.f32.u16 %f107, %rs33;
|
||||
div.rn.f32 %f108, %f107, 0f437F0000;
|
||||
// inline asm
|
||||
call (%rd219), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd40, %rd203, %rd12, %rd12);
|
||||
// inline asm
|
||||
ld.u8 %rs34, [%rd219];
|
||||
cvt.rn.f32.u16 %f109, %rs34;
|
||||
div.rn.f32 %f110, %f109, 0f437F0000;
|
||||
// inline asm
|
||||
call (%rd225), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd40, %rd203, %rd12, %rd12);
|
||||
// inline asm
|
||||
ld.u8 %rs35, [%rd225+3];
|
||||
cvt.rn.f32.u16 %f111, %rs35;
|
||||
div.rn.f32 %f112, %f111, 0f437F0000;
|
||||
add.f32 %f135, %f135, %f106;
|
||||
add.f32 %f136, %f136, %f108;
|
||||
add.f32 %f137, %f137, %f110;
|
||||
add.f32 %f138, %f138, %f112;
|
||||
add.s32 %r143, %r143, 1;
|
||||
|
||||
BB0_13:
|
||||
// inline asm
|
||||
call (%rd232), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd106, %rd203, %rd12, %rd12);
|
||||
// inline asm
|
||||
ld.u8 %rs36, [%rd232+3];
|
||||
setp.eq.s16 %p12, %rs36, 0;
|
||||
@%p12 bra BB0_15;
|
||||
|
||||
// inline asm
|
||||
call (%rd239), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd106, %rd203, %rd12, %rd12);
|
||||
// inline asm
|
||||
ld.u8 %rs37, [%rd239+2];
|
||||
cvt.rn.f32.u16 %f113, %rs37;
|
||||
div.rn.f32 %f114, %f113, 0f437F0000;
|
||||
// inline asm
|
||||
call (%rd245), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd106, %rd203, %rd12, %rd12);
|
||||
// inline asm
|
||||
ld.u8 %rs38, [%rd245+1];
|
||||
cvt.rn.f32.u16 %f115, %rs38;
|
||||
div.rn.f32 %f116, %f115, 0f437F0000;
|
||||
// inline asm
|
||||
call (%rd251), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd106, %rd203, %rd12, %rd12);
|
||||
// inline asm
|
||||
ld.u8 %rs39, [%rd251];
|
||||
cvt.rn.f32.u16 %f117, %rs39;
|
||||
div.rn.f32 %f118, %f117, 0f437F0000;
|
||||
// inline asm
|
||||
call (%rd257), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd106, %rd203, %rd12, %rd12);
|
||||
// inline asm
|
||||
ld.u8 %rs40, [%rd257+3];
|
||||
cvt.rn.f32.u16 %f119, %rs40;
|
||||
div.rn.f32 %f120, %f119, 0f437F0000;
|
||||
add.f32 %f135, %f135, %f114;
|
||||
add.f32 %f136, %f136, %f116;
|
||||
add.f32 %f137, %f137, %f118;
|
||||
add.f32 %f138, %f138, %f120;
|
||||
add.s32 %r143, %r143, 1;
|
||||
|
||||
BB0_15:
|
||||
setp.eq.s32 %p13, %r143, 0;
|
||||
@%p13 bra BB0_17;
|
||||
|
||||
cvt.rn.f32.u32 %f121, %r143;
|
||||
rcp.rn.f32 %f122, %f121;
|
||||
mul.f32 %f123, %f135, %f122;
|
||||
mul.f32 %f124, %f136, %f122;
|
||||
mul.f32 %f125, %f137, %f122;
|
||||
mul.f32 %f126, %f138, %f122;
|
||||
mul.f32 %f127, %f123, 0f437F0000;
|
||||
mul.f32 %f128, %f124, 0f437F0000;
|
||||
mul.f32 %f129, %f125, 0f437F0000;
|
||||
mul.f32 %f130, %f126, 0f437F0000;
|
||||
ld.global.v2.u32 {%r134, %r135}, [pixelID];
|
||||
cvt.u64.u32 %rd266, %r134;
|
||||
cvt.u64.u32 %rd267, %r135;
|
||||
// inline asm
|
||||
call (%rd264), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd266, %rd267, %rd12, %rd12);
|
||||
// inline asm
|
||||
cvt.rzi.u32.f32 %r138, %f129;
|
||||
cvt.rzi.u32.f32 %r139, %f128;
|
||||
cvt.rzi.u32.f32 %r140, %f127;
|
||||
cvt.rzi.u32.f32 %r141, %f130;
|
||||
cvt.u16.u32 %rs41, %r141;
|
||||
cvt.u16.u32 %rs42, %r140;
|
||||
cvt.u16.u32 %rs43, %r139;
|
||||
cvt.u16.u32 %rs44, %r138;
|
||||
st.v4.u8 [%rd264], {%rs44, %rs43, %rs42, %rs41};
|
||||
|
||||
BB0_17:
|
||||
ret;
|
||||
}
|
||||
|
||||
|
||||
8
Assets/Editor/x64/Bakery/dilate.ptx.meta
Normal file
8
Assets/Editor/x64/Bakery/dilate.ptx.meta
Normal file
@@ -0,0 +1,8 @@
|
||||
fileFormatVersion: 2
|
||||
guid: 8714cded3f3d6d64681ce7f4c7b804a6
|
||||
timeCreated: 1526161016
|
||||
licenseType: Store
|
||||
DefaultImporter:
|
||||
userData:
|
||||
assetBundleName:
|
||||
assetBundleVariant:
|
||||
284
Assets/Editor/x64/Bakery/dilateFloat.ptx
Normal file
284
Assets/Editor/x64/Bakery/dilateFloat.ptx
Normal file
@@ -0,0 +1,284 @@
|
||||
//
|
||||
// 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 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 8 .u64 _ZN21rti_internal_register20reg_bitness_detectorE;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail0E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail1E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail2E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail3E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail4E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail5E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail6E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail7E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail8E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail9E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail0E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail1E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail2E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail3E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail4E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail5E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail6E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail7E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail8E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail9E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register14reg_rayIndex_xE;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register14reg_rayIndex_yE;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register14reg_rayIndex_zE;
|
||||
.global .align 8 .b8 _ZN21rti_internal_typename7pixelIDE[6] = {117, 105, 110, 116, 50, 0};
|
||||
.global .align 8 .b8 _ZN21rti_internal_typename10resolutionE[6] = {117, 105, 110, 116, 50, 0};
|
||||
.global .align 8 .b8 _ZN21rti_internal_typename6normalE[7] = {102, 108, 111, 97, 116, 51, 0};
|
||||
.global .align 8 .b8 _ZN21rti_internal_typename6camPosE[7] = {102, 108, 111, 97, 116, 51, 0};
|
||||
.global .align 16 .b8 _ZN21rti_internal_typename4rootE[9] = {114, 116, 79, 98, 106, 101, 99, 116, 0};
|
||||
.global .align 4 .b8 _ZN21rti_internal_typename12imageEnabledE[4] = {105, 110, 116, 0};
|
||||
.global .align 8 .b8 _ZN21rti_internal_typename8tileInfoE[6] = {117, 105, 110, 116, 52, 0};
|
||||
.global .align 4 .b8 _ZN21rti_internal_typename8additiveE[4] = {105, 110, 116, 0};
|
||||
.global .align 4 .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 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 _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];
|
||||
|
||||
.visible .entry _Z6oxMainv(
|
||||
|
||||
)
|
||||
{
|
||||
.reg .pred %p<14>;
|
||||
.reg .f32 %f<134>;
|
||||
.reg .b32 %r<85>;
|
||||
.reg .b64 %rd<127>;
|
||||
|
||||
|
||||
ld.global.v2.u32 {%r18, %r19}, [pixelID];
|
||||
cvt.u64.u32 %rd9, %r18;
|
||||
cvt.u64.u32 %rd10, %r19;
|
||||
mov.u64 %rd13, image;
|
||||
cvta.global.u64 %rd8, %rd13;
|
||||
mov.u32 %r16, 2;
|
||||
mov.u32 %r17, 16;
|
||||
mov.u64 %rd12, 0;
|
||||
// inline asm
|
||||
call (%rd7), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd9, %rd10, %rd12, %rd12);
|
||||
// inline asm
|
||||
ld.f32 %f46, [%rd7+12];
|
||||
setp.gt.f32 %p1, %f46, 0f00000000;
|
||||
@%p1 bra BB0_17;
|
||||
|
||||
ld.global.v2.u32 {%r27, %r28}, [pixelID];
|
||||
cvt.u64.u32 %rd16, %r27;
|
||||
cvt.u64.u32 %rd17, %r28;
|
||||
// inline asm
|
||||
call (%rd14), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd16, %rd17, %rd12, %rd12);
|
||||
// inline asm
|
||||
ld.v4.f32 {%f116, %f117, %f118, %f50}, [%rd14];
|
||||
ld.global.v2.u32 {%r31, %r32}, [pixelID];
|
||||
setp.eq.s32 %p2, %r31, 0;
|
||||
add.s32 %r34, %r31, -1;
|
||||
cvt.u64.u32 %rd27, %r34;
|
||||
selp.b64 %rd22, 0, %rd27, %p2;
|
||||
setp.eq.s32 %p3, %r32, 0;
|
||||
add.s32 %r36, %r32, -1;
|
||||
cvt.u64.u32 %rd28, %r36;
|
||||
selp.b64 %rd23, 0, %rd28, %p3;
|
||||
ld.global.v2.u32 {%r37, %r38}, [resolution];
|
||||
add.s32 %r40, %r37, -1;
|
||||
setp.eq.s32 %p4, %r31, %r40;
|
||||
add.s32 %r41, %r31, 1;
|
||||
selp.b32 %r1, %r40, %r41, %p4;
|
||||
add.s32 %r43, %r38, -1;
|
||||
setp.eq.s32 %p5, %r32, %r43;
|
||||
add.s32 %r44, %r32, 1;
|
||||
selp.b32 %r2, %r43, %r44, %p5;
|
||||
// inline asm
|
||||
call (%rd20), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd22, %rd23, %rd12, %rd12);
|
||||
// inline asm
|
||||
ld.f32 %f51, [%rd20+12];
|
||||
mov.u32 %r79, 0;
|
||||
setp.leu.f32 %p6, %f51, 0f00000000;
|
||||
@%p6 bra BB0_3;
|
||||
|
||||
// inline asm
|
||||
call (%rd29), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd22, %rd23, %rd12, %rd12);
|
||||
// inline asm
|
||||
ld.v4.f32 {%f52, %f53, %f54, %f55}, [%rd29];
|
||||
add.f32 %f116, %f116, %f52;
|
||||
add.f32 %f117, %f117, %f53;
|
||||
add.f32 %f118, %f118, %f54;
|
||||
mov.u32 %r79, 1;
|
||||
|
||||
BB0_3:
|
||||
ld.global.u32 %rd38, [pixelID];
|
||||
// inline asm
|
||||
call (%rd36), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd38, %rd23, %rd12, %rd12);
|
||||
// inline asm
|
||||
ld.f32 %f59, [%rd36+12];
|
||||
setp.leu.f32 %p7, %f59, 0f00000000;
|
||||
@%p7 bra BB0_5;
|
||||
|
||||
// inline asm
|
||||
call (%rd43), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd38, %rd23, %rd12, %rd12);
|
||||
// inline asm
|
||||
ld.v4.f32 {%f60, %f61, %f62, %f63}, [%rd43];
|
||||
add.f32 %f116, %f116, %f60;
|
||||
add.f32 %f117, %f117, %f61;
|
||||
add.f32 %f118, %f118, %f62;
|
||||
add.s32 %r79, %r79, 1;
|
||||
|
||||
BB0_5:
|
||||
cvt.u64.u32 %rd52, %r1;
|
||||
// inline asm
|
||||
call (%rd50), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd52, %rd23, %rd12, %rd12);
|
||||
// inline asm
|
||||
ld.f32 %f67, [%rd50+12];
|
||||
setp.leu.f32 %p8, %f67, 0f00000000;
|
||||
@%p8 bra BB0_7;
|
||||
|
||||
// inline asm
|
||||
call (%rd57), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd52, %rd23, %rd12, %rd12);
|
||||
// inline asm
|
||||
ld.v4.f32 {%f68, %f69, %f70, %f71}, [%rd57];
|
||||
add.f32 %f116, %f116, %f68;
|
||||
add.f32 %f117, %f117, %f69;
|
||||
add.f32 %f118, %f118, %f70;
|
||||
add.s32 %r79, %r79, 1;
|
||||
|
||||
BB0_7:
|
||||
ld.global.u32 %rd67, [pixelID+4];
|
||||
// inline asm
|
||||
call (%rd64), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd22, %rd67, %rd12, %rd12);
|
||||
// inline asm
|
||||
ld.f32 %f75, [%rd64+12];
|
||||
setp.leu.f32 %p9, %f75, 0f00000000;
|
||||
@%p9 bra BB0_9;
|
||||
|
||||
// inline asm
|
||||
call (%rd71), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd22, %rd67, %rd12, %rd12);
|
||||
// inline asm
|
||||
ld.v4.f32 {%f76, %f77, %f78, %f79}, [%rd71];
|
||||
add.f32 %f116, %f116, %f76;
|
||||
add.f32 %f117, %f117, %f77;
|
||||
add.f32 %f118, %f118, %f78;
|
||||
add.s32 %r79, %r79, 1;
|
||||
|
||||
BB0_9:
|
||||
// inline asm
|
||||
call (%rd78), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd52, %rd67, %rd12, %rd12);
|
||||
// inline asm
|
||||
ld.f32 %f83, [%rd78+12];
|
||||
setp.leu.f32 %p10, %f83, 0f00000000;
|
||||
@%p10 bra BB0_11;
|
||||
|
||||
// inline asm
|
||||
call (%rd85), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd52, %rd67, %rd12, %rd12);
|
||||
// inline asm
|
||||
ld.v4.f32 {%f84, %f85, %f86, %f87}, [%rd85];
|
||||
add.f32 %f116, %f116, %f84;
|
||||
add.f32 %f117, %f117, %f85;
|
||||
add.f32 %f118, %f118, %f86;
|
||||
add.s32 %r79, %r79, 1;
|
||||
|
||||
BB0_11:
|
||||
cvt.u64.u32 %rd95, %r2;
|
||||
// inline asm
|
||||
call (%rd92), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd22, %rd95, %rd12, %rd12);
|
||||
// inline asm
|
||||
ld.f32 %f91, [%rd92+12];
|
||||
setp.leu.f32 %p11, %f91, 0f00000000;
|
||||
@%p11 bra BB0_13;
|
||||
|
||||
// inline asm
|
||||
call (%rd99), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd22, %rd95, %rd12, %rd12);
|
||||
// inline asm
|
||||
ld.v4.f32 {%f92, %f93, %f94, %f95}, [%rd99];
|
||||
add.f32 %f116, %f116, %f92;
|
||||
add.f32 %f117, %f117, %f93;
|
||||
add.f32 %f118, %f118, %f94;
|
||||
add.s32 %r79, %r79, 1;
|
||||
|
||||
BB0_13:
|
||||
// inline asm
|
||||
call (%rd106), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd52, %rd95, %rd12, %rd12);
|
||||
// inline asm
|
||||
ld.f32 %f99, [%rd106+12];
|
||||
setp.leu.f32 %p12, %f99, 0f00000000;
|
||||
@%p12 bra BB0_15;
|
||||
|
||||
// inline asm
|
||||
call (%rd113), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd52, %rd95, %rd12, %rd12);
|
||||
// inline asm
|
||||
ld.v4.f32 {%f100, %f101, %f102, %f103}, [%rd113];
|
||||
add.f32 %f116, %f116, %f100;
|
||||
add.f32 %f117, %f117, %f101;
|
||||
add.f32 %f118, %f118, %f102;
|
||||
add.s32 %r79, %r79, 1;
|
||||
|
||||
BB0_15:
|
||||
setp.eq.s32 %p13, %r79, 0;
|
||||
@%p13 bra BB0_17;
|
||||
|
||||
cvt.rn.f32.u32 %f107, %r79;
|
||||
rcp.rn.f32 %f108, %f107;
|
||||
ld.global.v2.u32 {%r74, %r75}, [pixelID];
|
||||
cvt.u64.u32 %rd122, %r74;
|
||||
cvt.u64.u32 %rd123, %r75;
|
||||
// inline asm
|
||||
call (%rd120), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd122, %rd123, %rd12, %rd12);
|
||||
// inline asm
|
||||
mul.f32 %f109, %f118, %f108;
|
||||
mul.f32 %f110, %f117, %f108;
|
||||
mul.f32 %f111, %f116, %f108;
|
||||
mov.f32 %f112, 0f3F800000;
|
||||
st.v4.f32 [%rd120], {%f111, %f110, %f109, %f112};
|
||||
|
||||
BB0_17:
|
||||
ret;
|
||||
}
|
||||
|
||||
|
||||
8
Assets/Editor/x64/Bakery/dilateFloat.ptx.meta
Normal file
8
Assets/Editor/x64/Bakery/dilateFloat.ptx.meta
Normal file
@@ -0,0 +1,8 @@
|
||||
fileFormatVersion: 2
|
||||
guid: a3321abd087b2e44e8f1852158e8d948
|
||||
timeCreated: 1526161016
|
||||
licenseType: Store
|
||||
DefaultImporter:
|
||||
userData:
|
||||
assetBundleName:
|
||||
assetBundleVariant:
|
||||
472
Assets/Editor/x64/Bakery/dilateHalf.ptx
Normal file
472
Assets/Editor/x64/Bakery/dilateHalf.ptx
Normal file
@@ -0,0 +1,472 @@
|
||||
//
|
||||
// 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 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 8 .u64 _ZN21rti_internal_register20reg_bitness_detectorE;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail0E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail1E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail2E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail3E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail4E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail5E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail6E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail7E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail8E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail9E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail0E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail1E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail2E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail3E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail4E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail5E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail6E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail7E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail8E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail9E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register14reg_rayIndex_xE;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register14reg_rayIndex_yE;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register14reg_rayIndex_zE;
|
||||
.global .align 8 .b8 _ZN21rti_internal_typename7pixelIDE[6] = {117, 105, 110, 116, 50, 0};
|
||||
.global .align 8 .b8 _ZN21rti_internal_typename10resolutionE[6] = {117, 105, 110, 116, 50, 0};
|
||||
.global .align 8 .b8 _ZN21rti_internal_typename6normalE[7] = {102, 108, 111, 97, 116, 51, 0};
|
||||
.global .align 8 .b8 _ZN21rti_internal_typename6camPosE[7] = {102, 108, 111, 97, 116, 51, 0};
|
||||
.global .align 16 .b8 _ZN21rti_internal_typename4rootE[9] = {114, 116, 79, 98, 106, 101, 99, 116, 0};
|
||||
.global .align 4 .b8 _ZN21rti_internal_typename12imageEnabledE[4] = {105, 110, 116, 0};
|
||||
.global .align 8 .b8 _ZN21rti_internal_typename8tileInfoE[6] = {117, 105, 110, 116, 52, 0};
|
||||
.global .align 4 .b8 _ZN21rti_internal_typename8additiveE[4] = {105, 110, 116, 0};
|
||||
.global .align 4 .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 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 _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];
|
||||
|
||||
.visible .entry _Z6oxMainv(
|
||||
|
||||
)
|
||||
{
|
||||
.reg .pred %p<14>;
|
||||
.reg .b16 %rs<39>;
|
||||
.reg .f32 %f<98>;
|
||||
.reg .b32 %r<125>;
|
||||
.reg .b64 %rd<223>;
|
||||
|
||||
|
||||
ld.global.v2.u32 {%r18, %r19}, [pixelID];
|
||||
cvt.u64.u32 %rd9, %r18;
|
||||
cvt.u64.u32 %rd10, %r19;
|
||||
mov.u64 %rd13, image;
|
||||
cvta.global.u64 %rd8, %rd13;
|
||||
mov.u32 %r16, 2;
|
||||
mov.u32 %r17, 8;
|
||||
mov.u64 %rd12, 0;
|
||||
// inline asm
|
||||
call (%rd7), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd9, %rd10, %rd12, %rd12);
|
||||
// inline asm
|
||||
ld.u16 %rs3, [%rd7+6];
|
||||
mov.f32 %f46, 0f00000000;
|
||||
// inline asm
|
||||
{ cvt.rn.f16.f32 %rs2, %f46;}
|
||||
|
||||
// inline asm
|
||||
setp.gt.u16 %p1, %rs3, %rs2;
|
||||
@%p1 bra BB0_17;
|
||||
|
||||
ld.global.v2.u32 {%r31, %r32}, [pixelID];
|
||||
cvt.u64.u32 %rd16, %r31;
|
||||
cvt.u64.u32 %rd17, %r32;
|
||||
// inline asm
|
||||
call (%rd14), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd16, %rd17, %rd12, %rd12);
|
||||
// inline asm
|
||||
ld.u16 %rs4, [%rd14];
|
||||
// inline asm
|
||||
{ cvt.f32.f16 %f80, %rs4;}
|
||||
|
||||
// inline asm
|
||||
ld.global.v2.u32 {%r35, %r36}, [pixelID];
|
||||
cvt.u64.u32 %rd22, %r35;
|
||||
cvt.u64.u32 %rd23, %r36;
|
||||
// inline asm
|
||||
call (%rd20), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd22, %rd23, %rd12, %rd12);
|
||||
// inline asm
|
||||
ld.u16 %rs5, [%rd20+2];
|
||||
// inline asm
|
||||
{ cvt.f32.f16 %f81, %rs5;}
|
||||
|
||||
// inline asm
|
||||
ld.global.v2.u32 {%r39, %r40}, [pixelID];
|
||||
cvt.u64.u32 %rd28, %r39;
|
||||
cvt.u64.u32 %rd29, %r40;
|
||||
// inline asm
|
||||
call (%rd26), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd28, %rd29, %rd12, %rd12);
|
||||
// inline asm
|
||||
ld.u16 %rs6, [%rd26+4];
|
||||
// inline asm
|
||||
{ cvt.f32.f16 %f82, %rs6;}
|
||||
|
||||
// inline asm
|
||||
ld.global.v2.u32 {%r43, %r44}, [pixelID];
|
||||
setp.eq.s32 %p2, %r43, 0;
|
||||
add.s32 %r46, %r43, -1;
|
||||
cvt.u64.u32 %rd39, %r46;
|
||||
selp.b64 %rd34, 0, %rd39, %p2;
|
||||
setp.eq.s32 %p3, %r44, 0;
|
||||
add.s32 %r48, %r44, -1;
|
||||
cvt.u64.u32 %rd40, %r48;
|
||||
selp.b64 %rd35, 0, %rd40, %p3;
|
||||
ld.global.v2.u32 {%r49, %r50}, [resolution];
|
||||
add.s32 %r52, %r49, -1;
|
||||
setp.eq.s32 %p4, %r43, %r52;
|
||||
add.s32 %r53, %r43, 1;
|
||||
selp.b32 %r1, %r52, %r53, %p4;
|
||||
add.s32 %r55, %r50, -1;
|
||||
setp.eq.s32 %p5, %r44, %r55;
|
||||
add.s32 %r56, %r44, 1;
|
||||
selp.b32 %r2, %r55, %r56, %p5;
|
||||
// inline asm
|
||||
call (%rd32), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd34, %rd35, %rd12, %rd12);
|
||||
// inline asm
|
||||
ld.u16 %rs7, [%rd32+6];
|
||||
mov.u32 %r119, 0;
|
||||
setp.le.u16 %p6, %rs7, %rs2;
|
||||
@%p6 bra BB0_3;
|
||||
|
||||
// inline asm
|
||||
call (%rd41), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd34, %rd35, %rd12, %rd12);
|
||||
// inline asm
|
||||
ld.u16 %rs8, [%rd41];
|
||||
// inline asm
|
||||
{ cvt.f32.f16 %f50, %rs8;}
|
||||
|
||||
// inline asm
|
||||
// inline asm
|
||||
call (%rd47), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd34, %rd35, %rd12, %rd12);
|
||||
// inline asm
|
||||
ld.u16 %rs9, [%rd47+2];
|
||||
// inline asm
|
||||
{ cvt.f32.f16 %f51, %rs9;}
|
||||
|
||||
// inline asm
|
||||
// inline asm
|
||||
call (%rd53), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd34, %rd35, %rd12, %rd12);
|
||||
// inline asm
|
||||
ld.u16 %rs10, [%rd53+4];
|
||||
// inline asm
|
||||
{ cvt.f32.f16 %f52, %rs10;}
|
||||
|
||||
// inline asm
|
||||
add.f32 %f80, %f80, %f50;
|
||||
add.f32 %f81, %f81, %f51;
|
||||
add.f32 %f82, %f82, %f52;
|
||||
mov.u32 %r119, 1;
|
||||
|
||||
BB0_3:
|
||||
ld.global.u32 %rd62, [pixelID];
|
||||
// inline asm
|
||||
call (%rd60), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd62, %rd35, %rd12, %rd12);
|
||||
// inline asm
|
||||
ld.u16 %rs11, [%rd60+6];
|
||||
setp.le.u16 %p7, %rs11, %rs2;
|
||||
@%p7 bra BB0_5;
|
||||
|
||||
// inline asm
|
||||
call (%rd67), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd62, %rd35, %rd12, %rd12);
|
||||
// inline asm
|
||||
ld.u16 %rs12, [%rd67];
|
||||
// inline asm
|
||||
{ cvt.f32.f16 %f53, %rs12;}
|
||||
|
||||
// inline asm
|
||||
// inline asm
|
||||
call (%rd73), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd62, %rd35, %rd12, %rd12);
|
||||
// inline asm
|
||||
ld.u16 %rs13, [%rd73+2];
|
||||
// inline asm
|
||||
{ cvt.f32.f16 %f54, %rs13;}
|
||||
|
||||
// inline asm
|
||||
// inline asm
|
||||
call (%rd79), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd62, %rd35, %rd12, %rd12);
|
||||
// inline asm
|
||||
ld.u16 %rs14, [%rd79+4];
|
||||
// inline asm
|
||||
{ cvt.f32.f16 %f55, %rs14;}
|
||||
|
||||
// inline asm
|
||||
add.f32 %f80, %f80, %f53;
|
||||
add.f32 %f81, %f81, %f54;
|
||||
add.f32 %f82, %f82, %f55;
|
||||
add.s32 %r119, %r119, 1;
|
||||
|
||||
BB0_5:
|
||||
cvt.u64.u32 %rd88, %r1;
|
||||
// inline asm
|
||||
call (%rd86), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd88, %rd35, %rd12, %rd12);
|
||||
// inline asm
|
||||
ld.u16 %rs15, [%rd86+6];
|
||||
setp.le.u16 %p8, %rs15, %rs2;
|
||||
@%p8 bra BB0_7;
|
||||
|
||||
// inline asm
|
||||
call (%rd93), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd88, %rd35, %rd12, %rd12);
|
||||
// inline asm
|
||||
ld.u16 %rs16, [%rd93];
|
||||
// inline asm
|
||||
{ cvt.f32.f16 %f56, %rs16;}
|
||||
|
||||
// inline asm
|
||||
// inline asm
|
||||
call (%rd99), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd88, %rd35, %rd12, %rd12);
|
||||
// inline asm
|
||||
ld.u16 %rs17, [%rd99+2];
|
||||
// inline asm
|
||||
{ cvt.f32.f16 %f57, %rs17;}
|
||||
|
||||
// inline asm
|
||||
// inline asm
|
||||
call (%rd105), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd88, %rd35, %rd12, %rd12);
|
||||
// inline asm
|
||||
ld.u16 %rs18, [%rd105+4];
|
||||
// inline asm
|
||||
{ cvt.f32.f16 %f58, %rs18;}
|
||||
|
||||
// inline asm
|
||||
add.f32 %f80, %f80, %f56;
|
||||
add.f32 %f81, %f81, %f57;
|
||||
add.f32 %f82, %f82, %f58;
|
||||
add.s32 %r119, %r119, 1;
|
||||
|
||||
BB0_7:
|
||||
ld.global.u32 %rd115, [pixelID+4];
|
||||
// inline asm
|
||||
call (%rd112), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd34, %rd115, %rd12, %rd12);
|
||||
// inline asm
|
||||
ld.u16 %rs19, [%rd112+6];
|
||||
setp.le.u16 %p9, %rs19, %rs2;
|
||||
@%p9 bra BB0_9;
|
||||
|
||||
// inline asm
|
||||
call (%rd119), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd34, %rd115, %rd12, %rd12);
|
||||
// inline asm
|
||||
ld.u16 %rs20, [%rd119];
|
||||
// inline asm
|
||||
{ cvt.f32.f16 %f59, %rs20;}
|
||||
|
||||
// inline asm
|
||||
// inline asm
|
||||
call (%rd125), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd34, %rd115, %rd12, %rd12);
|
||||
// inline asm
|
||||
ld.u16 %rs21, [%rd125+2];
|
||||
// inline asm
|
||||
{ cvt.f32.f16 %f60, %rs21;}
|
||||
|
||||
// inline asm
|
||||
// inline asm
|
||||
call (%rd131), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd34, %rd115, %rd12, %rd12);
|
||||
// inline asm
|
||||
ld.u16 %rs22, [%rd131+4];
|
||||
// inline asm
|
||||
{ cvt.f32.f16 %f61, %rs22;}
|
||||
|
||||
// inline asm
|
||||
add.f32 %f80, %f80, %f59;
|
||||
add.f32 %f81, %f81, %f60;
|
||||
add.f32 %f82, %f82, %f61;
|
||||
add.s32 %r119, %r119, 1;
|
||||
|
||||
BB0_9:
|
||||
// inline asm
|
||||
call (%rd138), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd88, %rd115, %rd12, %rd12);
|
||||
// inline asm
|
||||
ld.u16 %rs23, [%rd138+6];
|
||||
setp.le.u16 %p10, %rs23, %rs2;
|
||||
@%p10 bra BB0_11;
|
||||
|
||||
// inline asm
|
||||
call (%rd145), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd88, %rd115, %rd12, %rd12);
|
||||
// inline asm
|
||||
ld.u16 %rs24, [%rd145];
|
||||
// inline asm
|
||||
{ cvt.f32.f16 %f62, %rs24;}
|
||||
|
||||
// inline asm
|
||||
// inline asm
|
||||
call (%rd151), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd88, %rd115, %rd12, %rd12);
|
||||
// inline asm
|
||||
ld.u16 %rs25, [%rd151+2];
|
||||
// inline asm
|
||||
{ cvt.f32.f16 %f63, %rs25;}
|
||||
|
||||
// inline asm
|
||||
// inline asm
|
||||
call (%rd157), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd88, %rd115, %rd12, %rd12);
|
||||
// inline asm
|
||||
ld.u16 %rs26, [%rd157+4];
|
||||
// inline asm
|
||||
{ cvt.f32.f16 %f64, %rs26;}
|
||||
|
||||
// inline asm
|
||||
add.f32 %f80, %f80, %f62;
|
||||
add.f32 %f81, %f81, %f63;
|
||||
add.f32 %f82, %f82, %f64;
|
||||
add.s32 %r119, %r119, 1;
|
||||
|
||||
BB0_11:
|
||||
cvt.u64.u32 %rd167, %r2;
|
||||
// inline asm
|
||||
call (%rd164), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd34, %rd167, %rd12, %rd12);
|
||||
// inline asm
|
||||
ld.u16 %rs27, [%rd164+6];
|
||||
setp.le.u16 %p11, %rs27, %rs2;
|
||||
@%p11 bra BB0_13;
|
||||
|
||||
// inline asm
|
||||
call (%rd171), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd34, %rd167, %rd12, %rd12);
|
||||
// inline asm
|
||||
ld.u16 %rs28, [%rd171];
|
||||
// inline asm
|
||||
{ cvt.f32.f16 %f65, %rs28;}
|
||||
|
||||
// inline asm
|
||||
// inline asm
|
||||
call (%rd177), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd34, %rd167, %rd12, %rd12);
|
||||
// inline asm
|
||||
ld.u16 %rs29, [%rd177+2];
|
||||
// inline asm
|
||||
{ cvt.f32.f16 %f66, %rs29;}
|
||||
|
||||
// inline asm
|
||||
// inline asm
|
||||
call (%rd183), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd34, %rd167, %rd12, %rd12);
|
||||
// inline asm
|
||||
ld.u16 %rs30, [%rd183+4];
|
||||
// inline asm
|
||||
{ cvt.f32.f16 %f67, %rs30;}
|
||||
|
||||
// inline asm
|
||||
add.f32 %f80, %f80, %f65;
|
||||
add.f32 %f81, %f81, %f66;
|
||||
add.f32 %f82, %f82, %f67;
|
||||
add.s32 %r119, %r119, 1;
|
||||
|
||||
BB0_13:
|
||||
// inline asm
|
||||
call (%rd190), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd88, %rd167, %rd12, %rd12);
|
||||
// inline asm
|
||||
ld.u16 %rs31, [%rd190+6];
|
||||
setp.le.u16 %p12, %rs31, %rs2;
|
||||
@%p12 bra BB0_15;
|
||||
|
||||
// inline asm
|
||||
call (%rd197), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd88, %rd167, %rd12, %rd12);
|
||||
// inline asm
|
||||
ld.u16 %rs32, [%rd197];
|
||||
// inline asm
|
||||
{ cvt.f32.f16 %f68, %rs32;}
|
||||
|
||||
// inline asm
|
||||
// inline asm
|
||||
call (%rd203), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd88, %rd167, %rd12, %rd12);
|
||||
// inline asm
|
||||
ld.u16 %rs33, [%rd203+2];
|
||||
// inline asm
|
||||
{ cvt.f32.f16 %f69, %rs33;}
|
||||
|
||||
// inline asm
|
||||
// inline asm
|
||||
call (%rd209), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd88, %rd167, %rd12, %rd12);
|
||||
// inline asm
|
||||
ld.u16 %rs34, [%rd209+4];
|
||||
// inline asm
|
||||
{ cvt.f32.f16 %f70, %rs34;}
|
||||
|
||||
// inline asm
|
||||
add.f32 %f80, %f80, %f68;
|
||||
add.f32 %f81, %f81, %f69;
|
||||
add.f32 %f82, %f82, %f70;
|
||||
add.s32 %r119, %r119, 1;
|
||||
|
||||
BB0_15:
|
||||
setp.eq.s32 %p13, %r119, 0;
|
||||
@%p13 bra BB0_17;
|
||||
|
||||
cvt.rn.f32.u32 %f75, %r119;
|
||||
rcp.rn.f32 %f76, %f75;
|
||||
mul.f32 %f71, %f80, %f76;
|
||||
mul.f32 %f72, %f81, %f76;
|
||||
mul.f32 %f73, %f82, %f76;
|
||||
ld.global.v2.u32 {%r114, %r115}, [pixelID];
|
||||
cvt.u64.u32 %rd218, %r114;
|
||||
cvt.u64.u32 %rd219, %r115;
|
||||
// inline asm
|
||||
call (%rd216), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd218, %rd219, %rd12, %rd12);
|
||||
// inline asm
|
||||
// inline asm
|
||||
{ cvt.rn.f16.f32 %rs37, %f73;}
|
||||
|
||||
// inline asm
|
||||
// inline asm
|
||||
{ cvt.rn.f16.f32 %rs36, %f72;}
|
||||
|
||||
// inline asm
|
||||
// inline asm
|
||||
{ cvt.rn.f16.f32 %rs35, %f71;}
|
||||
|
||||
// inline asm
|
||||
mov.f32 %f74, 0f3F800000;
|
||||
// inline asm
|
||||
{ cvt.rn.f16.f32 %rs38, %f74;}
|
||||
|
||||
// inline asm
|
||||
st.v4.u16 [%rd216], {%rs35, %rs36, %rs37, %rs38};
|
||||
|
||||
BB0_17:
|
||||
ret;
|
||||
}
|
||||
|
||||
|
||||
8
Assets/Editor/x64/Bakery/dilateHalf.ptx.meta
Normal file
8
Assets/Editor/x64/Bakery/dilateHalf.ptx.meta
Normal file
@@ -0,0 +1,8 @@
|
||||
fileFormatVersion: 2
|
||||
guid: 221f0729ddb8fc742a94d3f5ed9ae330
|
||||
timeCreated: 1526161016
|
||||
licenseType: Store
|
||||
DefaultImporter:
|
||||
userData:
|
||||
assetBundleName:
|
||||
assetBundleVariant:
|
||||
286
Assets/Editor/x64/Bakery/dilateMask.ptx
Normal file
286
Assets/Editor/x64/Bakery/dilateMask.ptx
Normal file
@@ -0,0 +1,286 @@
|
||||
//
|
||||
// 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 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 8 .u64 _ZN21rti_internal_register20reg_bitness_detectorE;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail0E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail1E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail2E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail3E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail4E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail5E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail6E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail7E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail8E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail9E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail0E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail1E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail2E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail3E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail4E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail5E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail6E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail7E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail8E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail9E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register14reg_rayIndex_xE;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register14reg_rayIndex_yE;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register14reg_rayIndex_zE;
|
||||
.global .align 8 .b8 _ZN21rti_internal_typename7pixelIDE[6] = {117, 105, 110, 116, 50, 0};
|
||||
.global .align 8 .b8 _ZN21rti_internal_typename10resolutionE[6] = {117, 105, 110, 116, 50, 0};
|
||||
.global .align 8 .b8 _ZN21rti_internal_typename6normalE[7] = {102, 108, 111, 97, 116, 51, 0};
|
||||
.global .align 8 .b8 _ZN21rti_internal_typename6camPosE[7] = {102, 108, 111, 97, 116, 51, 0};
|
||||
.global .align 16 .b8 _ZN21rti_internal_typename4rootE[9] = {114, 116, 79, 98, 106, 101, 99, 116, 0};
|
||||
.global .align 4 .b8 _ZN21rti_internal_typename12imageEnabledE[4] = {105, 110, 116, 0};
|
||||
.global .align 8 .b8 _ZN21rti_internal_typename8tileInfoE[6] = {117, 105, 110, 116, 52, 0};
|
||||
.global .align 4 .b8 _ZN21rti_internal_typename8additiveE[4] = {105, 110, 116, 0};
|
||||
.global .align 4 .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 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 _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];
|
||||
|
||||
.visible .entry _Z6oxMainv(
|
||||
|
||||
)
|
||||
{
|
||||
.reg .pred %p<14>;
|
||||
.reg .b16 %rs<19>;
|
||||
.reg .f32 %f<41>;
|
||||
.reg .b32 %r<86>;
|
||||
.reg .b64 %rd<127>;
|
||||
|
||||
|
||||
ld.global.v2.u32 {%r18, %r19}, [pixelID];
|
||||
cvt.u64.u32 %rd9, %r18;
|
||||
cvt.u64.u32 %rd10, %r19;
|
||||
mov.u64 %rd13, image;
|
||||
cvta.global.u64 %rd8, %rd13;
|
||||
mov.u32 %r17, 2;
|
||||
mov.u64 %rd12, 0;
|
||||
// inline asm
|
||||
call (%rd7), _rt_buffer_get_64, (%rd8, %r17, %r17, %rd9, %rd10, %rd12, %rd12);
|
||||
// inline asm
|
||||
ld.u8 %rs1, [%rd7+1];
|
||||
setp.ne.s16 %p1, %rs1, 0;
|
||||
@%p1 bra BB0_17;
|
||||
|
||||
ld.global.v2.u32 {%r27, %r28}, [pixelID];
|
||||
cvt.u64.u32 %rd16, %r27;
|
||||
cvt.u64.u32 %rd17, %r28;
|
||||
// inline asm
|
||||
call (%rd14), _rt_buffer_get_64, (%rd8, %r17, %r17, %rd16, %rd17, %rd12, %rd12);
|
||||
// inline asm
|
||||
ld.u8 %rs2, [%rd14];
|
||||
cvt.rn.f32.u16 %f16, %rs2;
|
||||
div.rn.f32 %f35, %f16, 0f437F0000;
|
||||
ld.global.v2.u32 {%r31, %r32}, [pixelID];
|
||||
setp.eq.s32 %p2, %r31, 0;
|
||||
add.s32 %r34, %r31, -1;
|
||||
cvt.u64.u32 %rd27, %r34;
|
||||
selp.b64 %rd22, 0, %rd27, %p2;
|
||||
setp.eq.s32 %p3, %r32, 0;
|
||||
add.s32 %r36, %r32, -1;
|
||||
cvt.u64.u32 %rd28, %r36;
|
||||
selp.b64 %rd23, 0, %rd28, %p3;
|
||||
ld.global.v2.u32 {%r37, %r38}, [resolution];
|
||||
add.s32 %r40, %r37, -1;
|
||||
setp.eq.s32 %p4, %r31, %r40;
|
||||
add.s32 %r41, %r31, 1;
|
||||
selp.b32 %r1, %r40, %r41, %p4;
|
||||
add.s32 %r43, %r38, -1;
|
||||
setp.eq.s32 %p5, %r32, %r43;
|
||||
add.s32 %r44, %r32, 1;
|
||||
selp.b32 %r2, %r43, %r44, %p5;
|
||||
// inline asm
|
||||
call (%rd20), _rt_buffer_get_64, (%rd8, %r17, %r17, %rd22, %rd23, %rd12, %rd12);
|
||||
// inline asm
|
||||
ld.u8 %rs3, [%rd20+1];
|
||||
setp.eq.s16 %p6, %rs3, 0;
|
||||
mov.u32 %r80, 0;
|
||||
@%p6 bra BB0_3;
|
||||
|
||||
// inline asm
|
||||
call (%rd29), _rt_buffer_get_64, (%rd8, %r17, %r17, %rd22, %rd23, %rd12, %rd12);
|
||||
// inline asm
|
||||
ld.u8 %rs4, [%rd29];
|
||||
cvt.rn.f32.u16 %f17, %rs4;
|
||||
div.rn.f32 %f18, %f17, 0f437F0000;
|
||||
add.f32 %f35, %f35, %f18;
|
||||
mov.u32 %r80, 1;
|
||||
|
||||
BB0_3:
|
||||
ld.global.u32 %rd38, [pixelID];
|
||||
// inline asm
|
||||
call (%rd36), _rt_buffer_get_64, (%rd8, %r17, %r17, %rd38, %rd23, %rd12, %rd12);
|
||||
// inline asm
|
||||
ld.u8 %rs5, [%rd36+1];
|
||||
setp.eq.s16 %p7, %rs5, 0;
|
||||
@%p7 bra BB0_5;
|
||||
|
||||
// inline asm
|
||||
call (%rd43), _rt_buffer_get_64, (%rd8, %r17, %r17, %rd38, %rd23, %rd12, %rd12);
|
||||
// inline asm
|
||||
ld.u8 %rs6, [%rd43];
|
||||
cvt.rn.f32.u16 %f19, %rs6;
|
||||
div.rn.f32 %f20, %f19, 0f437F0000;
|
||||
add.f32 %f35, %f35, %f20;
|
||||
add.s32 %r80, %r80, 1;
|
||||
|
||||
BB0_5:
|
||||
cvt.u64.u32 %rd52, %r1;
|
||||
// inline asm
|
||||
call (%rd50), _rt_buffer_get_64, (%rd8, %r17, %r17, %rd52, %rd23, %rd12, %rd12);
|
||||
// inline asm
|
||||
ld.u8 %rs7, [%rd50+1];
|
||||
setp.eq.s16 %p8, %rs7, 0;
|
||||
@%p8 bra BB0_7;
|
||||
|
||||
// inline asm
|
||||
call (%rd57), _rt_buffer_get_64, (%rd8, %r17, %r17, %rd52, %rd23, %rd12, %rd12);
|
||||
// inline asm
|
||||
ld.u8 %rs8, [%rd57];
|
||||
cvt.rn.f32.u16 %f21, %rs8;
|
||||
div.rn.f32 %f22, %f21, 0f437F0000;
|
||||
add.f32 %f35, %f35, %f22;
|
||||
add.s32 %r80, %r80, 1;
|
||||
|
||||
BB0_7:
|
||||
ld.global.u32 %rd67, [pixelID+4];
|
||||
// inline asm
|
||||
call (%rd64), _rt_buffer_get_64, (%rd8, %r17, %r17, %rd22, %rd67, %rd12, %rd12);
|
||||
// inline asm
|
||||
ld.u8 %rs9, [%rd64+1];
|
||||
setp.eq.s16 %p9, %rs9, 0;
|
||||
@%p9 bra BB0_9;
|
||||
|
||||
// inline asm
|
||||
call (%rd71), _rt_buffer_get_64, (%rd8, %r17, %r17, %rd22, %rd67, %rd12, %rd12);
|
||||
// inline asm
|
||||
ld.u8 %rs10, [%rd71];
|
||||
cvt.rn.f32.u16 %f23, %rs10;
|
||||
div.rn.f32 %f24, %f23, 0f437F0000;
|
||||
add.f32 %f35, %f35, %f24;
|
||||
add.s32 %r80, %r80, 1;
|
||||
|
||||
BB0_9:
|
||||
// inline asm
|
||||
call (%rd78), _rt_buffer_get_64, (%rd8, %r17, %r17, %rd52, %rd67, %rd12, %rd12);
|
||||
// inline asm
|
||||
ld.u8 %rs11, [%rd78+1];
|
||||
setp.eq.s16 %p10, %rs11, 0;
|
||||
@%p10 bra BB0_11;
|
||||
|
||||
// inline asm
|
||||
call (%rd85), _rt_buffer_get_64, (%rd8, %r17, %r17, %rd52, %rd67, %rd12, %rd12);
|
||||
// inline asm
|
||||
ld.u8 %rs12, [%rd85];
|
||||
cvt.rn.f32.u16 %f25, %rs12;
|
||||
div.rn.f32 %f26, %f25, 0f437F0000;
|
||||
add.f32 %f35, %f35, %f26;
|
||||
add.s32 %r80, %r80, 1;
|
||||
|
||||
BB0_11:
|
||||
cvt.u64.u32 %rd95, %r2;
|
||||
// inline asm
|
||||
call (%rd92), _rt_buffer_get_64, (%rd8, %r17, %r17, %rd22, %rd95, %rd12, %rd12);
|
||||
// inline asm
|
||||
ld.u8 %rs13, [%rd92+1];
|
||||
setp.eq.s16 %p11, %rs13, 0;
|
||||
@%p11 bra BB0_13;
|
||||
|
||||
// inline asm
|
||||
call (%rd99), _rt_buffer_get_64, (%rd8, %r17, %r17, %rd22, %rd95, %rd12, %rd12);
|
||||
// inline asm
|
||||
ld.u8 %rs14, [%rd99];
|
||||
cvt.rn.f32.u16 %f27, %rs14;
|
||||
div.rn.f32 %f28, %f27, 0f437F0000;
|
||||
add.f32 %f35, %f35, %f28;
|
||||
add.s32 %r80, %r80, 1;
|
||||
|
||||
BB0_13:
|
||||
// inline asm
|
||||
call (%rd106), _rt_buffer_get_64, (%rd8, %r17, %r17, %rd52, %rd95, %rd12, %rd12);
|
||||
// inline asm
|
||||
ld.u8 %rs15, [%rd106+1];
|
||||
setp.eq.s16 %p12, %rs15, 0;
|
||||
@%p12 bra BB0_15;
|
||||
|
||||
// inline asm
|
||||
call (%rd113), _rt_buffer_get_64, (%rd8, %r17, %r17, %rd52, %rd95, %rd12, %rd12);
|
||||
// inline asm
|
||||
ld.u8 %rs16, [%rd113];
|
||||
cvt.rn.f32.u16 %f29, %rs16;
|
||||
div.rn.f32 %f30, %f29, 0f437F0000;
|
||||
add.f32 %f35, %f35, %f30;
|
||||
add.s32 %r80, %r80, 1;
|
||||
|
||||
BB0_15:
|
||||
setp.eq.s32 %p13, %r80, 0;
|
||||
@%p13 bra BB0_17;
|
||||
|
||||
cvt.rn.f32.u32 %f31, %r80;
|
||||
div.rn.f32 %f32, %f35, %f31;
|
||||
ld.global.v2.u32 {%r74, %r75}, [pixelID];
|
||||
cvt.u64.u32 %rd122, %r74;
|
||||
cvt.u64.u32 %rd123, %r75;
|
||||
// inline asm
|
||||
call (%rd120), _rt_buffer_get_64, (%rd8, %r17, %r17, %rd122, %rd123, %rd12, %rd12);
|
||||
// inline asm
|
||||
mul.f32 %f33, %f32, 0f437F0000;
|
||||
cvt.rzi.u32.f32 %r78, %f33;
|
||||
cvt.u16.u32 %rs17, %r78;
|
||||
mov.u16 %rs18, 255;
|
||||
st.v2.u8 [%rd120], {%rs17, %rs18};
|
||||
|
||||
BB0_17:
|
||||
ret;
|
||||
}
|
||||
|
||||
|
||||
8
Assets/Editor/x64/Bakery/dilateMask.ptx.meta
Normal file
8
Assets/Editor/x64/Bakery/dilateMask.ptx.meta
Normal file
@@ -0,0 +1,8 @@
|
||||
fileFormatVersion: 2
|
||||
guid: 47011ee837ad730478e644f3ec8b3953
|
||||
timeCreated: 1534426493
|
||||
licenseType: Store
|
||||
DefaultImporter:
|
||||
userData:
|
||||
assetBundleName:
|
||||
assetBundleVariant:
|
||||
BIN
Assets/Editor/x64/Bakery/dilateSimple.pso
Normal file
BIN
Assets/Editor/x64/Bakery/dilateSimple.pso
Normal file
Binary file not shown.
8
Assets/Editor/x64/Bakery/dilateSimple.pso.meta
Normal file
8
Assets/Editor/x64/Bakery/dilateSimple.pso.meta
Normal file
@@ -0,0 +1,8 @@
|
||||
fileFormatVersion: 2
|
||||
guid: fb7aef4959b95904399631eb75d7f0e5
|
||||
timeCreated: 1530103296
|
||||
licenseType: Store
|
||||
DefaultImporter:
|
||||
userData:
|
||||
assetBundleName:
|
||||
assetBundleVariant:
|
||||
934
Assets/Editor/x64/Bakery/exception.ptx
Normal file
934
Assets/Editor/x64/Bakery/exception.ptx
Normal file
@@ -0,0 +1,934 @@
|
||||
//
|
||||
// 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
|
||||
.extern .func (.param .b32 func_retval0) vprintf
|
||||
(
|
||||
.param .b64 vprintf_param_0,
|
||||
.param .b64 vprintf_param_1
|
||||
)
|
||||
;
|
||||
.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 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 8 .u64 _ZN21rti_internal_register20reg_bitness_detectorE;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail0E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail1E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail2E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail3E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail4E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail5E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail6E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail7E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail8E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail9E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail0E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail1E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail2E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail3E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail4E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail5E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail6E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail7E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail8E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail9E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register14reg_rayIndex_xE;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register14reg_rayIndex_yE;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register14reg_rayIndex_zE;
|
||||
.global .align 8 .b8 _ZN21rti_internal_typename7pixelIDE[6] = {117, 105, 110, 116, 50, 0};
|
||||
.global .align 8 .b8 _ZN21rti_internal_typename10resolutionE[6] = {117, 105, 110, 116, 50, 0};
|
||||
.global .align 8 .b8 _ZN21rti_internal_typename6normalE[7] = {102, 108, 111, 97, 116, 51, 0};
|
||||
.global .align 8 .b8 _ZN21rti_internal_typename6camPosE[7] = {102, 108, 111, 97, 116, 51, 0};
|
||||
.global .align 16 .b8 _ZN21rti_internal_typename4rootE[9] = {114, 116, 79, 98, 106, 101, 99, 116, 0};
|
||||
.global .align 4 .b8 _ZN21rti_internal_typename12imageEnabledE[4] = {105, 110, 116, 0};
|
||||
.global .align 8 .b8 _ZN21rti_internal_typename8tileInfoE[6] = {117, 105, 110, 116, 52, 0};
|
||||
.global .align 4 .b8 _ZN21rti_internal_typename8additiveE[4] = {105, 110, 116, 0};
|
||||
.global .align 4 .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 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 _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 16 .b8 $str[64] = {67, 97, 117, 103, 104, 116, 32, 82, 84, 95, 69, 88, 67, 69, 80, 84, 73, 79, 78, 95, 83, 84, 65, 67, 75, 95, 79, 86, 69, 82, 70, 76, 79, 87, 10, 32, 32, 108, 97, 117, 110, 99, 104, 32, 105, 110, 100, 101, 120, 32, 58, 32, 37, 100, 44, 32, 37, 100, 44, 32, 37, 100, 10, 0};
|
||||
.global .align 16 .b8 $str1[218] = {67, 97, 117, 103, 104, 116, 32, 82, 84, 95, 69, 88, 67, 69, 80, 84, 73, 79, 78, 95, 66, 85, 70, 70, 69, 82, 95, 73, 78, 68, 69, 88, 95, 79, 85, 84, 95, 79, 70, 95, 66, 79, 85, 78, 68, 83, 10, 32, 32, 108, 97, 117, 110, 99, 104, 32, 105, 110, 100, 101, 120, 32, 32, 32, 58, 32, 37, 100, 44, 32, 37, 100, 44, 32, 37, 100, 10, 32, 32, 98, 117, 102, 102, 101, 114, 32, 97, 100, 100, 114, 101, 115, 115, 32, 58, 32, 48, 120, 37, 108, 108, 88, 10, 32, 32, 100, 105, 109, 101, 110, 115, 105, 111, 110, 97, 108, 105, 116, 121, 32, 58, 32, 37, 100, 10, 32, 32, 115, 105, 122, 101, 32, 32, 32, 32, 32, 32, 32, 32, 32, 32, 32, 58, 32, 37, 108, 108, 100, 120, 37, 108, 108, 100, 120, 37, 108, 108, 100, 10, 32, 32, 101, 108, 101, 109, 101, 110, 116, 32, 115, 105, 122, 101, 32, 32, 32, 58, 32, 37, 100, 10, 32, 32, 97, 99, 99, 101, 115, 115, 101, 100, 32, 105, 110, 100, 101, 120, 32, 58, 32, 37, 108, 108, 100, 44, 32, 37, 108, 108, 100, 44, 32, 37, 108, 108, 100, 10, 0};
|
||||
.global .align 16 .b8 $str2[40] = {67, 97, 117, 103, 104, 116, 32, 82, 84, 95, 69, 88, 67, 69, 80, 84, 73, 79, 78, 95, 80, 82, 79, 71, 82, 65, 77, 95, 73, 68, 95, 73, 78, 86, 65, 76, 73, 68, 10, 0};
|
||||
.global .align 16 .b8 $str3[46] = {9, 112, 114, 111, 103, 114, 97, 109, 32, 73, 68, 32, 101, 113, 117, 97, 108, 32, 116, 111, 32, 82, 84, 95, 80, 82, 79, 71, 82, 65, 77, 95, 73, 68, 95, 78, 85, 76, 76, 32, 117, 115, 101, 100, 10, 0};
|
||||
.global .align 16 .b8 $str4[56] = {9, 112, 114, 111, 103, 114, 97, 109, 32, 73, 68, 32, 40, 37, 100, 41, 32, 105, 115, 32, 110, 111, 116, 32, 105, 110, 32, 116, 104, 101, 32, 118, 97, 108, 105, 100, 32, 114, 97, 110, 103, 101, 32, 111, 102, 32, 91, 49, 44, 115, 105, 122, 101, 41, 10, 0};
|
||||
.global .align 16 .b8 $str5[39] = {9, 112, 114, 111, 103, 114, 97, 109, 32, 73, 68, 32, 111, 102, 32, 97, 32, 100, 101, 108, 101, 116, 101, 100, 32, 112, 114, 111, 103, 114, 97, 109, 32, 117, 115, 101, 100, 10, 0};
|
||||
.global .align 16 .b8 $str6[40] = {67, 97, 117, 103, 104, 116, 32, 82, 84, 95, 69, 88, 67, 69, 80, 84, 73, 79, 78, 95, 84, 69, 88, 84, 85, 82, 69, 95, 73, 68, 95, 73, 78, 86, 65, 76, 73, 68, 10, 0};
|
||||
.global .align 16 .b8 $str7[33] = {9, 116, 101, 120, 116, 117, 114, 101, 32, 73, 68, 32, 40, 37, 100, 41, 32, 105, 115, 32, 105, 110, 118, 97, 108, 105, 100, 32, 40, 48, 41, 10, 0};
|
||||
.global .align 16 .b8 $str8[56] = {9, 116, 101, 120, 116, 117, 114, 101, 32, 73, 68, 32, 40, 37, 100, 41, 32, 105, 115, 32, 110, 111, 116, 32, 105, 110, 32, 116, 104, 101, 32, 118, 97, 108, 105, 100, 32, 114, 97, 110, 103, 101, 32, 111, 102, 32, 91, 49, 44, 115, 105, 122, 101, 41, 10, 0};
|
||||
.global .align 16 .b8 $str9[34] = {9, 116, 101, 120, 116, 117, 114, 101, 32, 73, 68, 32, 40, 37, 100, 41, 32, 105, 115, 32, 105, 110, 118, 97, 108, 105, 100, 32, 40, 45, 49, 41, 10, 0};
|
||||
.global .align 16 .b8 $str10[39] = {67, 97, 117, 103, 104, 116, 32, 82, 84, 95, 69, 88, 67, 69, 80, 84, 73, 79, 78, 95, 66, 85, 70, 70, 69, 82, 95, 73, 68, 95, 73, 78, 86, 65, 76, 73, 68, 10, 0};
|
||||
.global .align 16 .b8 $str11[44] = {9, 98, 117, 102, 102, 101, 114, 32, 73, 68, 32, 101, 113, 117, 97, 108, 32, 116, 111, 32, 82, 84, 95, 66, 85, 70, 70, 69, 82, 95, 73, 68, 95, 78, 85, 76, 76, 32, 117, 115, 101, 100, 10, 0};
|
||||
.global .align 16 .b8 $str12[55] = {9, 98, 117, 102, 102, 101, 114, 32, 73, 68, 32, 40, 37, 100, 41, 32, 105, 115, 32, 110, 111, 116, 32, 105, 110, 32, 116, 104, 101, 32, 118, 97, 108, 105, 100, 32, 114, 97, 110, 103, 101, 32, 111, 102, 32, 91, 49, 44, 115, 105, 122, 101, 41, 10, 0};
|
||||
.global .align 16 .b8 $str13[37] = {9, 66, 117, 102, 102, 101, 114, 32, 73, 68, 32, 111, 102, 32, 97, 32, 100, 101, 108, 101, 116, 101, 100, 32, 98, 117, 102, 102, 101, 114, 32, 117, 115, 101, 100, 10, 0};
|
||||
.global .align 16 .b8 $str14[145] = {67, 97, 117, 103, 104, 116, 32, 82, 84, 95, 69, 88, 67, 69, 80, 84, 73, 79, 78, 95, 73, 78, 68, 69, 88, 95, 79, 85, 84, 95, 79, 70, 95, 66, 79, 85, 78, 68, 83, 10, 32, 32, 108, 97, 117, 110, 99, 104, 32, 105, 110, 100, 101, 120, 32, 32, 32, 58, 32, 37, 100, 44, 32, 37, 100, 44, 32, 37, 100, 10, 32, 32, 98, 117, 102, 102, 101, 114, 32, 97, 100, 100, 114, 101, 115, 115, 32, 58, 32, 48, 120, 37, 108, 108, 88, 10, 32, 32, 115, 105, 122, 101, 32, 32, 32, 32, 32, 32, 32, 32, 32, 32, 32, 58, 32, 37, 108, 108, 100, 10, 32, 32, 97, 99, 99, 101, 115, 115, 101, 100, 32, 105, 110, 100, 101, 120, 32, 58, 32, 37, 108, 108, 100, 10, 0};
|
||||
.global .align 16 .b8 $str15[179] = {67, 97, 117, 103, 104, 116, 32, 82, 84, 95, 69, 88, 67, 69, 80, 84, 73, 79, 78, 95, 73, 78, 86, 65, 76, 73, 68, 95, 82, 65, 89, 10, 32, 32, 108, 97, 117, 110, 99, 104, 32, 105, 110, 100, 101, 120, 32, 32, 58, 32, 37, 100, 44, 32, 37, 100, 44, 32, 37, 100, 10, 32, 32, 114, 97, 121, 32, 111, 114, 105, 103, 105, 110, 32, 32, 32, 32, 58, 32, 37, 102, 32, 37, 102, 32, 37, 102, 10, 32, 32, 114, 97, 121, 32, 100, 105, 114, 101, 99, 116, 105, 111, 110, 32, 58, 32, 37, 102, 32, 37, 102, 32, 37, 102, 10, 32, 32, 114, 97, 121, 32, 116, 121, 112, 101, 32, 32, 32, 32, 32, 32, 58, 32, 37, 100, 10, 32, 32, 114, 97, 121, 32, 116, 109, 105, 110, 32, 32, 32, 32, 32, 32, 58, 32, 37, 102, 10, 32, 32, 114, 97, 121, 32, 116, 109, 97, 120, 32, 32, 32, 32, 32, 32, 58, 32, 37, 102, 10, 0};
|
||||
.global .align 16 .b8 $str16[84] = {67, 97, 117, 103, 104, 116, 32, 82, 84, 95, 69, 88, 67, 69, 80, 84, 73, 79, 78, 95, 73, 78, 84, 69, 82, 78, 65, 76, 95, 69, 82, 82, 79, 82, 10, 32, 32, 108, 97, 117, 110, 99, 104, 32, 105, 110, 100, 101, 120, 32, 58, 32, 37, 100, 44, 32, 37, 100, 44, 32, 37, 100, 10, 32, 32, 101, 114, 114, 111, 114, 32, 105, 100, 32, 32, 32, 32, 32, 58, 32, 37, 100, 10, 0};
|
||||
.global .align 16 .b8 $str17[57] = {67, 97, 117, 103, 104, 116, 32, 82, 84, 95, 69, 88, 67, 69, 80, 84, 73, 79, 78, 95, 85, 83, 69, 82, 43, 37, 100, 10, 32, 32, 108, 97, 117, 110, 99, 104, 32, 105, 110, 100, 101, 120, 32, 58, 32, 37, 100, 44, 32, 37, 100, 44, 32, 37, 100, 10, 0};
|
||||
.global .align 16 .b8 $str18[54] = {67, 97, 117, 103, 104, 116, 32, 117, 110, 107, 110, 111, 119, 110, 32, 101, 120, 99, 101, 112, 116, 105, 111, 110, 10, 32, 32, 108, 97, 117, 110, 99, 104, 32, 105, 110, 100, 101, 120, 32, 58, 32, 37, 100, 44, 32, 37, 100, 44, 32, 37, 100, 10, 0};
|
||||
|
||||
.visible .entry _Z6oxMainv(
|
||||
|
||||
)
|
||||
{
|
||||
.local .align 16 .b8 __local_depot0[208];
|
||||
.reg .b64 %SP;
|
||||
.reg .b64 %SPL;
|
||||
.reg .pred %p<40>;
|
||||
.reg .f32 %f<9>;
|
||||
.reg .b32 %r<84>;
|
||||
.reg .f64 %fd<9>;
|
||||
.reg .b64 %rd<90>;
|
||||
|
||||
|
||||
mov.u64 %rd89, __local_depot0;
|
||||
cvta.local.u64 %SP, %rd89;
|
||||
// inline asm
|
||||
call (%r39), _rt_get_exception_code, ();
|
||||
// inline asm
|
||||
// inline asm
|
||||
call (%r40), _rt_get_exception_code, ();
|
||||
// inline asm
|
||||
setp.eq.s32 %p1, %r40, 1020;
|
||||
@%p1 bra BB0_58;
|
||||
bra.uni BB0_1;
|
||||
|
||||
BB0_58:
|
||||
ld.volatile.global.u32 %r36, [_ZN21rti_internal_register14reg_rayIndex_xE];
|
||||
ld.volatile.global.u32 %r37, [_ZN21rti_internal_register14reg_rayIndex_yE];
|
||||
ld.volatile.global.u32 %r38, [_ZN21rti_internal_register14reg_rayIndex_zE];
|
||||
// inline asm
|
||||
call (%r81), _rt_print_active, ();
|
||||
// inline asm
|
||||
setp.eq.s32 %p39, %r81, 0;
|
||||
@%p39 bra BB0_60;
|
||||
|
||||
add.u64 %rd83, %SP, 184;
|
||||
cvta.to.local.u64 %rd84, %rd83;
|
||||
st.local.v2.u32 [%rd84], {%r36, %r37};
|
||||
st.local.u32 [%rd84+8], %r38;
|
||||
mov.u64 %rd85, $str;
|
||||
cvta.global.u64 %rd86, %rd85;
|
||||
// Callseq Start 18
|
||||
{
|
||||
.reg .b32 temp_param_reg;
|
||||
// <end>}
|
||||
.param .b64 param0;
|
||||
st.param.b64 [param0+0], %rd86;
|
||||
.param .b64 param1;
|
||||
st.param.b64 [param1+0], %rd83;
|
||||
.param .b32 retval0;
|
||||
call.uni (retval0),
|
||||
vprintf,
|
||||
(
|
||||
param0,
|
||||
param1
|
||||
);
|
||||
ld.param.b32 %r82, [retval0+0];
|
||||
|
||||
//{
|
||||
}// Callseq End 18
|
||||
bra.uni BB0_60;
|
||||
|
||||
BB0_1:
|
||||
setp.eq.s32 %p2, %r40, 1021;
|
||||
@%p2 bra BB0_52;
|
||||
bra.uni BB0_2;
|
||||
|
||||
BB0_52:
|
||||
ld.volatile.global.u32 %r30, [_ZN21rti_internal_register21reg_exception_detail0E];
|
||||
ld.volatile.global.u32 %r31, [_ZN21rti_internal_register14reg_rayIndex_xE];
|
||||
ld.volatile.global.u32 %r32, [_ZN21rti_internal_register14reg_rayIndex_yE];
|
||||
ld.volatile.global.u32 %r33, [_ZN21rti_internal_register14reg_rayIndex_zE];
|
||||
ld.volatile.global.u64 %rd5, [_ZN21rti_internal_register24reg_exception_64_detail0E];
|
||||
ld.volatile.global.u32 %r34, [_ZN21rti_internal_register21reg_exception_detail0E];
|
||||
ld.volatile.global.u64 %rd6, [_ZN21rti_internal_register24reg_exception_64_detail1E];
|
||||
mov.u64 %rd88, 1;
|
||||
setp.lt.u32 %p36, %r30, 2;
|
||||
mov.u64 %rd87, %rd88;
|
||||
@%p36 bra BB0_54;
|
||||
|
||||
ld.volatile.global.u64 %rd87, [_ZN21rti_internal_register24reg_exception_64_detail2E];
|
||||
|
||||
BB0_54:
|
||||
setp.lt.u32 %p37, %r30, 3;
|
||||
@%p37 bra BB0_56;
|
||||
|
||||
ld.volatile.global.u64 %rd88, [_ZN21rti_internal_register24reg_exception_64_detail3E];
|
||||
|
||||
BB0_56:
|
||||
ld.volatile.global.u32 %r35, [_ZN21rti_internal_register21reg_exception_detail1E];
|
||||
ld.volatile.global.u64 %rd11, [_ZN21rti_internal_register24reg_exception_64_detail4E];
|
||||
ld.volatile.global.u64 %rd12, [_ZN21rti_internal_register24reg_exception_64_detail5E];
|
||||
ld.volatile.global.u64 %rd13, [_ZN21rti_internal_register24reg_exception_64_detail6E];
|
||||
// inline asm
|
||||
call (%r79), _rt_print_active, ();
|
||||
// inline asm
|
||||
setp.eq.s32 %p38, %r79, 0;
|
||||
@%p38 bra BB0_60;
|
||||
|
||||
add.u64 %rd79, %SP, 96;
|
||||
cvta.to.local.u64 %rd80, %rd79;
|
||||
st.local.v2.u32 [%rd80], {%r31, %r32};
|
||||
st.local.u32 [%rd80+8], %r33;
|
||||
st.local.u32 [%rd80+24], %r34;
|
||||
st.local.u32 [%rd80+56], %r35;
|
||||
st.local.u64 [%rd80+16], %rd5;
|
||||
st.local.u64 [%rd80+32], %rd6;
|
||||
st.local.u64 [%rd80+40], %rd87;
|
||||
st.local.u64 [%rd80+48], %rd88;
|
||||
st.local.u64 [%rd80+64], %rd11;
|
||||
st.local.u64 [%rd80+72], %rd12;
|
||||
st.local.u64 [%rd80+80], %rd13;
|
||||
mov.u64 %rd81, $str1;
|
||||
cvta.global.u64 %rd82, %rd81;
|
||||
// Callseq Start 17
|
||||
{
|
||||
.reg .b32 temp_param_reg;
|
||||
// <end>}
|
||||
.param .b64 param0;
|
||||
st.param.b64 [param0+0], %rd82;
|
||||
.param .b64 param1;
|
||||
st.param.b64 [param1+0], %rd79;
|
||||
.param .b32 retval0;
|
||||
call.uni (retval0),
|
||||
vprintf,
|
||||
(
|
||||
param0,
|
||||
param1
|
||||
);
|
||||
ld.param.b32 %r80, [retval0+0];
|
||||
|
||||
//{
|
||||
}// Callseq End 17
|
||||
|
||||
BB0_60:
|
||||
ret;
|
||||
|
||||
BB0_2:
|
||||
setp.eq.s32 %p3, %r40, 1006;
|
||||
@%p3 bra BB0_41;
|
||||
bra.uni BB0_3;
|
||||
|
||||
BB0_41:
|
||||
// inline asm
|
||||
call (%r70), _rt_print_active, ();
|
||||
// inline asm
|
||||
setp.eq.s32 %p29, %r70, 0;
|
||||
@%p29 bra BB0_43;
|
||||
|
||||
mov.u64 %rd64, $str2;
|
||||
cvta.global.u64 %rd65, %rd64;
|
||||
mov.u64 %rd66, 0;
|
||||
// Callseq Start 13
|
||||
{
|
||||
.reg .b32 temp_param_reg;
|
||||
// <end>}
|
||||
.param .b64 param0;
|
||||
st.param.b64 [param0+0], %rd65;
|
||||
.param .b64 param1;
|
||||
st.param.b64 [param1+0], %rd66;
|
||||
.param .b32 retval0;
|
||||
call.uni (retval0),
|
||||
vprintf,
|
||||
(
|
||||
param0,
|
||||
param1
|
||||
);
|
||||
ld.param.b32 %r71, [retval0+0];
|
||||
|
||||
//{
|
||||
}// Callseq End 13
|
||||
|
||||
BB0_43:
|
||||
ld.volatile.global.u32 %r72, [_ZN21rti_internal_register21reg_exception_detail1E];
|
||||
setp.eq.s32 %p30, %r72, 0;
|
||||
@%p30 bra BB0_50;
|
||||
|
||||
setp.eq.s32 %p31, %r72, 1;
|
||||
@%p31 bra BB0_48;
|
||||
bra.uni BB0_45;
|
||||
|
||||
BB0_48:
|
||||
ld.volatile.global.u32 %r29, [_ZN21rti_internal_register21reg_exception_detail0E];
|
||||
// inline asm
|
||||
call (%r75), _rt_print_active, ();
|
||||
// inline asm
|
||||
setp.eq.s32 %p34, %r75, 0;
|
||||
@%p34 bra BB0_60;
|
||||
|
||||
add.u64 %rd70, %SP, 88;
|
||||
cvta.to.local.u64 %rd71, %rd70;
|
||||
st.local.u32 [%rd71], %r29;
|
||||
mov.u64 %rd72, $str4;
|
||||
cvta.global.u64 %rd73, %rd72;
|
||||
// Callseq Start 15
|
||||
{
|
||||
.reg .b32 temp_param_reg;
|
||||
// <end>}
|
||||
.param .b64 param0;
|
||||
st.param.b64 [param0+0], %rd73;
|
||||
.param .b64 param1;
|
||||
st.param.b64 [param1+0], %rd70;
|
||||
.param .b32 retval0;
|
||||
call.uni (retval0),
|
||||
vprintf,
|
||||
(
|
||||
param0,
|
||||
param1
|
||||
);
|
||||
ld.param.b32 %r76, [retval0+0];
|
||||
|
||||
//{
|
||||
}// Callseq End 15
|
||||
bra.uni BB0_60;
|
||||
|
||||
BB0_3:
|
||||
setp.eq.s32 %p4, %r40, 1007;
|
||||
@%p4 bra BB0_30;
|
||||
bra.uni BB0_4;
|
||||
|
||||
BB0_30:
|
||||
// inline asm
|
||||
call (%r61), _rt_print_active, ();
|
||||
// inline asm
|
||||
setp.eq.s32 %p22, %r61, 0;
|
||||
@%p22 bra BB0_32;
|
||||
|
||||
mov.u64 %rd49, $str6;
|
||||
cvta.global.u64 %rd50, %rd49;
|
||||
mov.u64 %rd51, 0;
|
||||
// Callseq Start 9
|
||||
{
|
||||
.reg .b32 temp_param_reg;
|
||||
// <end>}
|
||||
.param .b64 param0;
|
||||
st.param.b64 [param0+0], %rd50;
|
||||
.param .b64 param1;
|
||||
st.param.b64 [param1+0], %rd51;
|
||||
.param .b32 retval0;
|
||||
call.uni (retval0),
|
||||
vprintf,
|
||||
(
|
||||
param0,
|
||||
param1
|
||||
);
|
||||
ld.param.b32 %r62, [retval0+0];
|
||||
|
||||
//{
|
||||
}// Callseq End 9
|
||||
|
||||
BB0_32:
|
||||
ld.volatile.global.u32 %r63, [_ZN21rti_internal_register21reg_exception_detail1E];
|
||||
setp.eq.s32 %p23, %r63, 0;
|
||||
@%p23 bra BB0_39;
|
||||
|
||||
setp.eq.s32 %p24, %r63, 1;
|
||||
@%p24 bra BB0_37;
|
||||
bra.uni BB0_34;
|
||||
|
||||
BB0_37:
|
||||
ld.volatile.global.u32 %r27, [_ZN21rti_internal_register21reg_exception_detail0E];
|
||||
// inline asm
|
||||
call (%r66), _rt_print_active, ();
|
||||
// inline asm
|
||||
setp.eq.s32 %p27, %r66, 0;
|
||||
@%p27 bra BB0_60;
|
||||
|
||||
add.u64 %rd56, %SP, 72;
|
||||
cvta.to.local.u64 %rd57, %rd56;
|
||||
st.local.u32 [%rd57], %r27;
|
||||
mov.u64 %rd58, $str8;
|
||||
cvta.global.u64 %rd59, %rd58;
|
||||
// Callseq Start 11
|
||||
{
|
||||
.reg .b32 temp_param_reg;
|
||||
// <end>}
|
||||
.param .b64 param0;
|
||||
st.param.b64 [param0+0], %rd59;
|
||||
.param .b64 param1;
|
||||
st.param.b64 [param1+0], %rd56;
|
||||
.param .b32 retval0;
|
||||
call.uni (retval0),
|
||||
vprintf,
|
||||
(
|
||||
param0,
|
||||
param1
|
||||
);
|
||||
ld.param.b32 %r67, [retval0+0];
|
||||
|
||||
//{
|
||||
}// Callseq End 11
|
||||
bra.uni BB0_60;
|
||||
|
||||
BB0_4:
|
||||
setp.eq.s32 %p5, %r40, 1018;
|
||||
@%p5 bra BB0_19;
|
||||
bra.uni BB0_5;
|
||||
|
||||
BB0_19:
|
||||
// inline asm
|
||||
call (%r52), _rt_print_active, ();
|
||||
// inline asm
|
||||
setp.eq.s32 %p15, %r52, 0;
|
||||
@%p15 bra BB0_21;
|
||||
|
||||
mov.u64 %rd36, $str10;
|
||||
cvta.global.u64 %rd37, %rd36;
|
||||
mov.u64 %rd38, 0;
|
||||
// Callseq Start 5
|
||||
{
|
||||
.reg .b32 temp_param_reg;
|
||||
// <end>}
|
||||
.param .b64 param0;
|
||||
st.param.b64 [param0+0], %rd37;
|
||||
.param .b64 param1;
|
||||
st.param.b64 [param1+0], %rd38;
|
||||
.param .b32 retval0;
|
||||
call.uni (retval0),
|
||||
vprintf,
|
||||
(
|
||||
param0,
|
||||
param1
|
||||
);
|
||||
ld.param.b32 %r53, [retval0+0];
|
||||
|
||||
//{
|
||||
}// Callseq End 5
|
||||
|
||||
BB0_21:
|
||||
ld.volatile.global.u32 %r54, [_ZN21rti_internal_register21reg_exception_detail1E];
|
||||
setp.eq.s32 %p16, %r54, 0;
|
||||
@%p16 bra BB0_28;
|
||||
|
||||
setp.eq.s32 %p17, %r54, 1;
|
||||
@%p17 bra BB0_26;
|
||||
bra.uni BB0_23;
|
||||
|
||||
BB0_26:
|
||||
ld.volatile.global.u32 %r25, [_ZN21rti_internal_register21reg_exception_detail0E];
|
||||
// inline asm
|
||||
call (%r57), _rt_print_active, ();
|
||||
// inline asm
|
||||
setp.eq.s32 %p20, %r57, 0;
|
||||
@%p20 bra BB0_60;
|
||||
|
||||
add.u64 %rd42, %SP, 56;
|
||||
cvta.to.local.u64 %rd43, %rd42;
|
||||
st.local.u32 [%rd43], %r25;
|
||||
mov.u64 %rd44, $str12;
|
||||
cvta.global.u64 %rd45, %rd44;
|
||||
// Callseq Start 7
|
||||
{
|
||||
.reg .b32 temp_param_reg;
|
||||
// <end>}
|
||||
.param .b64 param0;
|
||||
st.param.b64 [param0+0], %rd45;
|
||||
.param .b64 param1;
|
||||
st.param.b64 [param1+0], %rd42;
|
||||
.param .b32 retval0;
|
||||
call.uni (retval0),
|
||||
vprintf,
|
||||
(
|
||||
param0,
|
||||
param1
|
||||
);
|
||||
ld.param.b32 %r58, [retval0+0];
|
||||
|
||||
//{
|
||||
}// Callseq End 7
|
||||
bra.uni BB0_60;
|
||||
|
||||
BB0_50:
|
||||
// inline asm
|
||||
call (%r77), _rt_print_active, ();
|
||||
// inline asm
|
||||
setp.eq.s32 %p35, %r77, 0;
|
||||
@%p35 bra BB0_60;
|
||||
|
||||
mov.u64 %rd74, $str3;
|
||||
cvta.global.u64 %rd75, %rd74;
|
||||
mov.u64 %rd76, 0;
|
||||
// Callseq Start 16
|
||||
{
|
||||
.reg .b32 temp_param_reg;
|
||||
// <end>}
|
||||
.param .b64 param0;
|
||||
st.param.b64 [param0+0], %rd75;
|
||||
.param .b64 param1;
|
||||
st.param.b64 [param1+0], %rd76;
|
||||
.param .b32 retval0;
|
||||
call.uni (retval0),
|
||||
vprintf,
|
||||
(
|
||||
param0,
|
||||
param1
|
||||
);
|
||||
ld.param.b32 %r78, [retval0+0];
|
||||
|
||||
//{
|
||||
}// Callseq End 16
|
||||
bra.uni BB0_60;
|
||||
|
||||
BB0_45:
|
||||
setp.ne.s32 %p32, %r72, 2;
|
||||
@%p32 bra BB0_60;
|
||||
|
||||
// inline asm
|
||||
call (%r73), _rt_print_active, ();
|
||||
// inline asm
|
||||
setp.eq.s32 %p33, %r73, 0;
|
||||
@%p33 bra BB0_60;
|
||||
|
||||
mov.u64 %rd67, $str5;
|
||||
cvta.global.u64 %rd68, %rd67;
|
||||
mov.u64 %rd69, 0;
|
||||
// Callseq Start 14
|
||||
{
|
||||
.reg .b32 temp_param_reg;
|
||||
// <end>}
|
||||
.param .b64 param0;
|
||||
st.param.b64 [param0+0], %rd68;
|
||||
.param .b64 param1;
|
||||
st.param.b64 [param1+0], %rd69;
|
||||
.param .b32 retval0;
|
||||
call.uni (retval0),
|
||||
vprintf,
|
||||
(
|
||||
param0,
|
||||
param1
|
||||
);
|
||||
ld.param.b32 %r74, [retval0+0];
|
||||
|
||||
//{
|
||||
}// Callseq End 14
|
||||
bra.uni BB0_60;
|
||||
|
||||
BB0_5:
|
||||
setp.eq.s32 %p6, %r40, 1019;
|
||||
@%p6 bra BB0_17;
|
||||
bra.uni BB0_6;
|
||||
|
||||
BB0_17:
|
||||
ld.volatile.global.u32 %r50, [_ZN21rti_internal_register21reg_exception_detail0E];
|
||||
ld.volatile.global.u32 %r22, [_ZN21rti_internal_register14reg_rayIndex_xE];
|
||||
ld.volatile.global.u32 %r23, [_ZN21rti_internal_register14reg_rayIndex_yE];
|
||||
ld.volatile.global.u32 %r24, [_ZN21rti_internal_register14reg_rayIndex_zE];
|
||||
ld.volatile.global.u64 %rd2, [_ZN21rti_internal_register24reg_exception_64_detail0E];
|
||||
ld.volatile.global.u64 %rd3, [_ZN21rti_internal_register24reg_exception_64_detail1E];
|
||||
ld.volatile.global.u64 %rd4, [_ZN21rti_internal_register24reg_exception_64_detail2E];
|
||||
// inline asm
|
||||
call (%r49), _rt_print_active, ();
|
||||
// inline asm
|
||||
setp.eq.s32 %p14, %r49, 0;
|
||||
@%p14 bra BB0_60;
|
||||
|
||||
add.u64 %rd32, %SP, 16;
|
||||
cvta.to.local.u64 %rd33, %rd32;
|
||||
st.local.v2.u32 [%rd33], {%r22, %r23};
|
||||
st.local.u32 [%rd33+8], %r24;
|
||||
st.local.u64 [%rd33+16], %rd2;
|
||||
st.local.u64 [%rd33+24], %rd3;
|
||||
st.local.u64 [%rd33+32], %rd4;
|
||||
mov.u64 %rd34, $str14;
|
||||
cvta.global.u64 %rd35, %rd34;
|
||||
// Callseq Start 4
|
||||
{
|
||||
.reg .b32 temp_param_reg;
|
||||
// <end>}
|
||||
.param .b64 param0;
|
||||
st.param.b64 [param0+0], %rd35;
|
||||
.param .b64 param1;
|
||||
st.param.b64 [param1+0], %rd32;
|
||||
.param .b32 retval0;
|
||||
call.uni (retval0),
|
||||
vprintf,
|
||||
(
|
||||
param0,
|
||||
param1
|
||||
);
|
||||
ld.param.b32 %r51, [retval0+0];
|
||||
|
||||
//{
|
||||
}// Callseq End 4
|
||||
bra.uni BB0_60;
|
||||
|
||||
BB0_39:
|
||||
ld.volatile.global.u32 %r28, [_ZN21rti_internal_register21reg_exception_detail0E];
|
||||
// inline asm
|
||||
call (%r68), _rt_print_active, ();
|
||||
// inline asm
|
||||
setp.eq.s32 %p28, %r68, 0;
|
||||
@%p28 bra BB0_60;
|
||||
|
||||
add.u64 %rd60, %SP, 80;
|
||||
cvta.to.local.u64 %rd61, %rd60;
|
||||
st.local.u32 [%rd61], %r28;
|
||||
mov.u64 %rd62, $str7;
|
||||
cvta.global.u64 %rd63, %rd62;
|
||||
// Callseq Start 12
|
||||
{
|
||||
.reg .b32 temp_param_reg;
|
||||
// <end>}
|
||||
.param .b64 param0;
|
||||
st.param.b64 [param0+0], %rd63;
|
||||
.param .b64 param1;
|
||||
st.param.b64 [param1+0], %rd60;
|
||||
.param .b32 retval0;
|
||||
call.uni (retval0),
|
||||
vprintf,
|
||||
(
|
||||
param0,
|
||||
param1
|
||||
);
|
||||
ld.param.b32 %r69, [retval0+0];
|
||||
|
||||
//{
|
||||
}// Callseq End 12
|
||||
bra.uni BB0_60;
|
||||
|
||||
BB0_34:
|
||||
setp.ne.s32 %p25, %r63, 2;
|
||||
@%p25 bra BB0_60;
|
||||
|
||||
ld.volatile.global.u32 %r26, [_ZN21rti_internal_register21reg_exception_detail0E];
|
||||
// inline asm
|
||||
call (%r64), _rt_print_active, ();
|
||||
// inline asm
|
||||
setp.eq.s32 %p26, %r64, 0;
|
||||
@%p26 bra BB0_60;
|
||||
|
||||
add.u64 %rd52, %SP, 64;
|
||||
cvta.to.local.u64 %rd53, %rd52;
|
||||
st.local.u32 [%rd53], %r26;
|
||||
mov.u64 %rd54, $str9;
|
||||
cvta.global.u64 %rd55, %rd54;
|
||||
// Callseq Start 10
|
||||
{
|
||||
.reg .b32 temp_param_reg;
|
||||
// <end>}
|
||||
.param .b64 param0;
|
||||
st.param.b64 [param0+0], %rd55;
|
||||
.param .b64 param1;
|
||||
st.param.b64 [param1+0], %rd52;
|
||||
.param .b32 retval0;
|
||||
call.uni (retval0),
|
||||
vprintf,
|
||||
(
|
||||
param0,
|
||||
param1
|
||||
);
|
||||
ld.param.b32 %r65, [retval0+0];
|
||||
|
||||
//{
|
||||
}// Callseq End 10
|
||||
bra.uni BB0_60;
|
||||
|
||||
BB0_6:
|
||||
setp.eq.s32 %p7, %r40, 1022;
|
||||
@%p7 bra BB0_15;
|
||||
bra.uni BB0_7;
|
||||
|
||||
BB0_15:
|
||||
ld.volatile.global.u32 %r10, [_ZN21rti_internal_register14reg_rayIndex_xE];
|
||||
ld.volatile.global.u32 %r11, [_ZN21rti_internal_register14reg_rayIndex_yE];
|
||||
ld.volatile.global.u32 %r12, [_ZN21rti_internal_register14reg_rayIndex_zE];
|
||||
ld.volatile.global.u32 %r13, [_ZN21rti_internal_register21reg_exception_detail0E];
|
||||
ld.volatile.global.u32 %r14, [_ZN21rti_internal_register21reg_exception_detail1E];
|
||||
ld.volatile.global.u32 %r15, [_ZN21rti_internal_register21reg_exception_detail2E];
|
||||
ld.volatile.global.u32 %r16, [_ZN21rti_internal_register21reg_exception_detail3E];
|
||||
ld.volatile.global.u32 %r17, [_ZN21rti_internal_register21reg_exception_detail4E];
|
||||
ld.volatile.global.u32 %r18, [_ZN21rti_internal_register21reg_exception_detail5E];
|
||||
ld.volatile.global.u32 %r19, [_ZN21rti_internal_register21reg_exception_detail6E];
|
||||
ld.volatile.global.u32 %r20, [_ZN21rti_internal_register21reg_exception_detail7E];
|
||||
ld.volatile.global.u32 %r21, [_ZN21rti_internal_register21reg_exception_detail8E];
|
||||
// inline asm
|
||||
call (%r47), _rt_print_active, ();
|
||||
// inline asm
|
||||
setp.eq.s32 %p13, %r47, 0;
|
||||
@%p13 bra BB0_60;
|
||||
|
||||
mov.b32 %f1, %r13;
|
||||
cvt.f64.f32 %fd1, %f1;
|
||||
mov.b32 %f2, %r14;
|
||||
cvt.f64.f32 %fd2, %f2;
|
||||
mov.b32 %f3, %r15;
|
||||
cvt.f64.f32 %fd3, %f3;
|
||||
mov.b32 %f4, %r16;
|
||||
cvt.f64.f32 %fd4, %f4;
|
||||
mov.b32 %f5, %r17;
|
||||
cvt.f64.f32 %fd5, %f5;
|
||||
mov.b32 %f6, %r18;
|
||||
cvt.f64.f32 %fd6, %f6;
|
||||
mov.b32 %f7, %r20;
|
||||
cvt.f64.f32 %fd7, %f7;
|
||||
mov.b32 %f8, %r21;
|
||||
cvt.f64.f32 %fd8, %f8;
|
||||
add.u64 %rd28, %SP, 96;
|
||||
cvta.to.local.u64 %rd29, %rd28;
|
||||
st.local.v2.u32 [%rd29], {%r10, %r11};
|
||||
st.local.u32 [%rd29+8], %r12;
|
||||
st.local.u32 [%rd29+64], %r19;
|
||||
st.local.f64 [%rd29+16], %fd1;
|
||||
st.local.f64 [%rd29+24], %fd2;
|
||||
st.local.f64 [%rd29+32], %fd3;
|
||||
st.local.f64 [%rd29+40], %fd4;
|
||||
st.local.f64 [%rd29+48], %fd5;
|
||||
st.local.f64 [%rd29+56], %fd6;
|
||||
st.local.f64 [%rd29+72], %fd7;
|
||||
st.local.f64 [%rd29+80], %fd8;
|
||||
mov.u64 %rd30, $str15;
|
||||
cvta.global.u64 %rd31, %rd30;
|
||||
// Callseq Start 3
|
||||
{
|
||||
.reg .b32 temp_param_reg;
|
||||
// <end>}
|
||||
.param .b64 param0;
|
||||
st.param.b64 [param0+0], %rd31;
|
||||
.param .b64 param1;
|
||||
st.param.b64 [param1+0], %rd28;
|
||||
.param .b32 retval0;
|
||||
call.uni (retval0),
|
||||
vprintf,
|
||||
(
|
||||
param0,
|
||||
param1
|
||||
);
|
||||
ld.param.b32 %r48, [retval0+0];
|
||||
|
||||
//{
|
||||
}// Callseq End 3
|
||||
bra.uni BB0_60;
|
||||
|
||||
BB0_28:
|
||||
// inline asm
|
||||
call (%r59), _rt_print_active, ();
|
||||
// inline asm
|
||||
setp.eq.s32 %p21, %r59, 0;
|
||||
@%p21 bra BB0_60;
|
||||
|
||||
mov.u64 %rd46, $str11;
|
||||
cvta.global.u64 %rd47, %rd46;
|
||||
mov.u64 %rd48, 0;
|
||||
// Callseq Start 8
|
||||
{
|
||||
.reg .b32 temp_param_reg;
|
||||
// <end>}
|
||||
.param .b64 param0;
|
||||
st.param.b64 [param0+0], %rd47;
|
||||
.param .b64 param1;
|
||||
st.param.b64 [param1+0], %rd48;
|
||||
.param .b32 retval0;
|
||||
call.uni (retval0),
|
||||
vprintf,
|
||||
(
|
||||
param0,
|
||||
param1
|
||||
);
|
||||
ld.param.b32 %r60, [retval0+0];
|
||||
|
||||
//{
|
||||
}// Callseq End 8
|
||||
bra.uni BB0_60;
|
||||
|
||||
BB0_23:
|
||||
setp.ne.s32 %p18, %r54, 2;
|
||||
@%p18 bra BB0_60;
|
||||
|
||||
// inline asm
|
||||
call (%r55), _rt_print_active, ();
|
||||
// inline asm
|
||||
setp.eq.s32 %p19, %r55, 0;
|
||||
@%p19 bra BB0_60;
|
||||
|
||||
mov.u64 %rd39, $str13;
|
||||
cvta.global.u64 %rd40, %rd39;
|
||||
mov.u64 %rd41, 0;
|
||||
// Callseq Start 6
|
||||
{
|
||||
.reg .b32 temp_param_reg;
|
||||
// <end>}
|
||||
.param .b64 param0;
|
||||
st.param.b64 [param0+0], %rd40;
|
||||
.param .b64 param1;
|
||||
st.param.b64 [param1+0], %rd41;
|
||||
.param .b32 retval0;
|
||||
call.uni (retval0),
|
||||
vprintf,
|
||||
(
|
||||
param0,
|
||||
param1
|
||||
);
|
||||
ld.param.b32 %r56, [retval0+0];
|
||||
|
||||
//{
|
||||
}// Callseq End 6
|
||||
bra.uni BB0_60;
|
||||
|
||||
BB0_7:
|
||||
setp.eq.s32 %p8, %r40, 1023;
|
||||
add.u64 %rd14, %SP, 0;
|
||||
cvta.to.local.u64 %rd15, %rd14;
|
||||
add.s64 %rd1, %rd15, 4;
|
||||
@%p8 bra BB0_13;
|
||||
bra.uni BB0_8;
|
||||
|
||||
BB0_13:
|
||||
ld.volatile.global.u32 %r6, [_ZN21rti_internal_register14reg_rayIndex_xE];
|
||||
ld.volatile.global.u32 %r7, [_ZN21rti_internal_register14reg_rayIndex_yE];
|
||||
ld.volatile.global.u32 %r8, [_ZN21rti_internal_register14reg_rayIndex_zE];
|
||||
ld.volatile.global.u32 %r9, [_ZN21rti_internal_register21reg_exception_detail0E];
|
||||
// inline asm
|
||||
call (%r45), _rt_print_active, ();
|
||||
// inline asm
|
||||
setp.eq.s32 %p12, %r45, 0;
|
||||
@%p12 bra BB0_60;
|
||||
|
||||
st.local.u32 [%rd15], %r6;
|
||||
st.local.u32 [%rd1], %r7;
|
||||
st.local.v2.u32 [%rd1+4], {%r8, %r9};
|
||||
mov.u64 %rd26, $str16;
|
||||
cvta.global.u64 %rd27, %rd26;
|
||||
// Callseq Start 2
|
||||
{
|
||||
.reg .b32 temp_param_reg;
|
||||
// <end>}
|
||||
.param .b64 param0;
|
||||
st.param.b64 [param0+0], %rd27;
|
||||
.param .b64 param1;
|
||||
st.param.b64 [param1+0], %rd14;
|
||||
.param .b32 retval0;
|
||||
call.uni (retval0),
|
||||
vprintf,
|
||||
(
|
||||
param0,
|
||||
param1
|
||||
);
|
||||
ld.param.b32 %r46, [retval0+0];
|
||||
|
||||
//{
|
||||
}// Callseq End 2
|
||||
bra.uni BB0_60;
|
||||
|
||||
BB0_8:
|
||||
add.s32 %r2, %r40, -1024;
|
||||
setp.lt.u32 %p9, %r2, 64512;
|
||||
ld.volatile.global.u32 %r3, [_ZN21rti_internal_register14reg_rayIndex_xE];
|
||||
ld.volatile.global.u32 %r4, [_ZN21rti_internal_register14reg_rayIndex_yE];
|
||||
ld.volatile.global.u32 %r5, [_ZN21rti_internal_register14reg_rayIndex_zE];
|
||||
@%p9 bra BB0_11;
|
||||
bra.uni BB0_9;
|
||||
|
||||
BB0_11:
|
||||
// inline asm
|
||||
call (%r43), _rt_print_active, ();
|
||||
// inline asm
|
||||
setp.eq.s32 %p11, %r43, 0;
|
||||
@%p11 bra BB0_60;
|
||||
|
||||
add.s32 %r83, %r40, -1024;
|
||||
st.local.u32 [%rd15], %r83;
|
||||
st.local.u32 [%rd1], %r3;
|
||||
st.local.v2.u32 [%rd1+4], {%r4, %r5};
|
||||
mov.u64 %rd22, $str17;
|
||||
cvta.global.u64 %rd23, %rd22;
|
||||
// Callseq Start 1
|
||||
{
|
||||
.reg .b32 temp_param_reg;
|
||||
// <end>}
|
||||
.param .b64 param0;
|
||||
st.param.b64 [param0+0], %rd23;
|
||||
.param .b64 param1;
|
||||
st.param.b64 [param1+0], %rd14;
|
||||
.param .b32 retval0;
|
||||
call.uni (retval0),
|
||||
vprintf,
|
||||
(
|
||||
param0,
|
||||
param1
|
||||
);
|
||||
ld.param.b32 %r44, [retval0+0];
|
||||
|
||||
//{
|
||||
}// Callseq End 1
|
||||
bra.uni BB0_60;
|
||||
|
||||
BB0_9:
|
||||
// inline asm
|
||||
call (%r41), _rt_print_active, ();
|
||||
// inline asm
|
||||
setp.eq.s32 %p10, %r41, 0;
|
||||
@%p10 bra BB0_60;
|
||||
|
||||
add.u64 %rd16, %SP, 184;
|
||||
cvta.to.local.u64 %rd17, %rd16;
|
||||
st.local.v2.u32 [%rd17], {%r3, %r4};
|
||||
st.local.u32 [%rd17+8], %r5;
|
||||
mov.u64 %rd18, $str18;
|
||||
cvta.global.u64 %rd19, %rd18;
|
||||
// Callseq Start 0
|
||||
{
|
||||
.reg .b32 temp_param_reg;
|
||||
// <end>}
|
||||
.param .b64 param0;
|
||||
st.param.b64 [param0+0], %rd19;
|
||||
.param .b64 param1;
|
||||
st.param.b64 [param1+0], %rd16;
|
||||
.param .b32 retval0;
|
||||
call.uni (retval0),
|
||||
vprintf,
|
||||
(
|
||||
param0,
|
||||
param1
|
||||
);
|
||||
ld.param.b32 %r42, [retval0+0];
|
||||
|
||||
//{
|
||||
}// Callseq End 0
|
||||
bra.uni BB0_60;
|
||||
}
|
||||
|
||||
|
||||
8
Assets/Editor/x64/Bakery/exception.ptx.meta
Normal file
8
Assets/Editor/x64/Bakery/exception.ptx.meta
Normal file
@@ -0,0 +1,8 @@
|
||||
fileFormatVersion: 2
|
||||
guid: 9d9bc0dee203fca4e82a2ec45087ad9e
|
||||
timeCreated: 1527356601
|
||||
licenseType: Store
|
||||
DefaultImporter:
|
||||
userData:
|
||||
assetBundleName:
|
||||
assetBundleVariant:
|
||||
337
Assets/Editor/x64/Bakery/fixPos.ptx
Normal file
337
Assets/Editor/x64/Bakery/fixPos.ptx
Normal file
@@ -0,0 +1,337 @@
|
||||
//
|
||||
// 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_HDR[1];
|
||||
.global .align 1 .b8 image_HDR2[1];
|
||||
.global .align 1 .b8 uvfacenormal[1];
|
||||
.global .align 1 .b8 uvpos[1];
|
||||
.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 8 .u64 _ZN21rti_internal_register20reg_bitness_detectorE;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail0E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail1E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail2E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail3E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail4E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail5E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail6E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail7E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail8E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail9E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail0E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail1E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail2E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail3E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail4E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail5E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail6E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail7E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail8E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail9E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register14reg_rayIndex_xE;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register14reg_rayIndex_yE;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register14reg_rayIndex_zE;
|
||||
.global .align 8 .b8 _ZN21rti_internal_typename7pixelIDE[6] = {117, 105, 110, 116, 50, 0};
|
||||
.global .align 8 .b8 _ZN21rti_internal_typename10resolutionE[6] = {117, 105, 110, 116, 50, 0};
|
||||
.global .align 8 .b8 _ZN21rti_internal_typename6normalE[7] = {102, 108, 111, 97, 116, 51, 0};
|
||||
.global .align 8 .b8 _ZN21rti_internal_typename6camPosE[7] = {102, 108, 111, 97, 116, 51, 0};
|
||||
.global .align 16 .b8 _ZN21rti_internal_typename4rootE[9] = {114, 116, 79, 98, 106, 101, 99, 116, 0};
|
||||
.global .align 4 .b8 _ZN21rti_internal_typename12imageEnabledE[4] = {105, 110, 116, 0};
|
||||
.global .align 8 .b8 _ZN21rti_internal_typename8tileInfoE[6] = {117, 105, 110, 116, 52, 0};
|
||||
.global .align 4 .b8 _ZN21rti_internal_typename8additiveE[4] = {105, 110, 116, 0};
|
||||
.global .align 4 .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 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 _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];
|
||||
|
||||
.visible .entry _Z6oxMainv(
|
||||
|
||||
)
|
||||
{
|
||||
.local .align 4 .b8 __local_depot0[4];
|
||||
.reg .b64 %SP;
|
||||
.reg .b64 %SPL;
|
||||
.reg .pred %p<12>;
|
||||
.reg .b16 %rs<9>;
|
||||
.reg .f32 %f<153>;
|
||||
.reg .b32 %r<59>;
|
||||
.reg .b64 %rd<57>;
|
||||
|
||||
|
||||
mov.u64 %rd56, __local_depot0;
|
||||
cvta.local.u64 %SP, %rd56;
|
||||
ld.global.v2.u32 {%r6, %r7}, [pixelID];
|
||||
cvt.u64.u32 %rd4, %r6;
|
||||
cvt.u64.u32 %rd5, %r7;
|
||||
mov.u64 %rd8, uvfacenormal;
|
||||
cvta.global.u64 %rd3, %rd8;
|
||||
mov.u32 %r4, 2;
|
||||
mov.u32 %r5, 4;
|
||||
mov.u64 %rd7, 0;
|
||||
// inline asm
|
||||
call (%rd2), _rt_buffer_get_64, (%rd3, %r4, %r5, %rd4, %rd5, %rd7, %rd7);
|
||||
// inline asm
|
||||
ld.u32 %r1, [%rd2];
|
||||
shr.u32 %r10, %r1, 16;
|
||||
cvt.u16.u32 %rs1, %r10;
|
||||
and.b16 %rs2, %rs1, 255;
|
||||
cvt.u16.u32 %rs3, %r1;
|
||||
or.b16 %rs4, %rs3, %rs2;
|
||||
setp.eq.s16 %p1, %rs4, 0;
|
||||
mov.f32 %f143, 0f00000000;
|
||||
mov.f32 %f144, %f143;
|
||||
mov.f32 %f145, %f143;
|
||||
@%p1 bra BB0_2;
|
||||
|
||||
ld.u8 %rs5, [%rd2+1];
|
||||
and.b16 %rs7, %rs3, 255;
|
||||
cvt.rn.f32.u16 %f61, %rs7;
|
||||
div.rn.f32 %f62, %f61, 0f437F0000;
|
||||
fma.rn.f32 %f63, %f62, 0f40000000, 0fBF800000;
|
||||
cvt.rn.f32.u16 %f64, %rs5;
|
||||
div.rn.f32 %f65, %f64, 0f437F0000;
|
||||
fma.rn.f32 %f66, %f65, 0f40000000, 0fBF800000;
|
||||
cvt.rn.f32.u16 %f67, %rs2;
|
||||
div.rn.f32 %f68, %f67, 0f437F0000;
|
||||
fma.rn.f32 %f69, %f68, 0f40000000, 0fBF800000;
|
||||
mul.f32 %f70, %f66, %f66;
|
||||
fma.rn.f32 %f71, %f63, %f63, %f70;
|
||||
fma.rn.f32 %f72, %f69, %f69, %f71;
|
||||
sqrt.rn.f32 %f73, %f72;
|
||||
rcp.rn.f32 %f74, %f73;
|
||||
mul.f32 %f143, %f63, %f74;
|
||||
mul.f32 %f144, %f66, %f74;
|
||||
mul.f32 %f145, %f69, %f74;
|
||||
|
||||
BB0_2:
|
||||
ld.global.v2.u32 {%r11, %r12}, [pixelID];
|
||||
ld.global.v2.u32 {%r14, %r15}, [tileInfo];
|
||||
add.s32 %r2, %r11, %r14;
|
||||
add.s32 %r3, %r12, %r15;
|
||||
setp.eq.f32 %p2, %f144, 0f00000000;
|
||||
setp.eq.f32 %p3, %f143, 0f00000000;
|
||||
and.pred %p4, %p3, %p2;
|
||||
setp.eq.f32 %p5, %f145, 0f00000000;
|
||||
and.pred %p6, %p4, %p5;
|
||||
@%p6 bra BB0_9;
|
||||
bra.uni BB0_3;
|
||||
|
||||
BB0_9:
|
||||
cvt.u64.u32 %rd51, %r2;
|
||||
cvt.u64.u32 %rd52, %r3;
|
||||
mov.u64 %rd55, image_HDR;
|
||||
cvta.global.u64 %rd50, %rd55;
|
||||
mov.u32 %r58, 16;
|
||||
// inline asm
|
||||
call (%rd49), _rt_buffer_get_64, (%rd50, %r4, %r58, %rd51, %rd52, %rd7, %rd7);
|
||||
// inline asm
|
||||
mov.f32 %f142, 0f00000000;
|
||||
st.v4.f32 [%rd49], {%f142, %f142, %f142, %f142};
|
||||
bra.uni BB0_10;
|
||||
|
||||
BB0_3:
|
||||
ld.global.v2.u32 {%r30, %r31}, [pixelID];
|
||||
cvt.u64.u32 %rd11, %r30;
|
||||
cvt.u64.u32 %rd12, %r31;
|
||||
mov.u64 %rd34, uvpos;
|
||||
cvta.global.u64 %rd10, %rd34;
|
||||
mov.u32 %r26, 16;
|
||||
// inline asm
|
||||
call (%rd9), _rt_buffer_get_64, (%rd10, %r4, %r26, %rd11, %rd12, %rd7, %rd7);
|
||||
// inline asm
|
||||
ld.f32 %f150, [%rd9];
|
||||
ld.global.v2.u32 {%r34, %r35}, [pixelID];
|
||||
cvt.u64.u32 %rd17, %r34;
|
||||
cvt.u64.u32 %rd18, %r35;
|
||||
// inline asm
|
||||
call (%rd15), _rt_buffer_get_64, (%rd10, %r4, %r26, %rd17, %rd18, %rd7, %rd7);
|
||||
// inline asm
|
||||
ld.f32 %f151, [%rd15+4];
|
||||
ld.global.v2.u32 {%r38, %r39}, [pixelID];
|
||||
cvt.u64.u32 %rd23, %r38;
|
||||
cvt.u64.u32 %rd24, %r39;
|
||||
// inline asm
|
||||
call (%rd21), _rt_buffer_get_64, (%rd10, %r4, %r26, %rd23, %rd24, %rd7, %rd7);
|
||||
// inline asm
|
||||
ld.f32 %f152, [%rd21+8];
|
||||
ld.global.v2.u32 {%r42, %r43}, [pixelID];
|
||||
cvt.u64.u32 %rd29, %r42;
|
||||
cvt.u64.u32 %rd30, %r43;
|
||||
// inline asm
|
||||
call (%rd27), _rt_buffer_get_64, (%rd10, %r4, %r26, %rd29, %rd30, %rd7, %rd7);
|
||||
// inline asm
|
||||
ld.f32 %f83, [%rd27+12];
|
||||
add.f32 %f82, %f83, 0f38D1B717;
|
||||
fma.rn.f32 %f11, %f143, 0f38D1B717, %f150;
|
||||
fma.rn.f32 %f12, %f144, 0f38D1B717, %f151;
|
||||
fma.rn.f32 %f13, %f145, 0f38D1B717, %f152;
|
||||
abs.f32 %f84, %f145;
|
||||
abs.f32 %f85, %f143;
|
||||
setp.gt.f32 %p7, %f85, %f84;
|
||||
neg.f32 %f86, %f144;
|
||||
selp.f32 %f87, %f86, 0f00000000, %p7;
|
||||
neg.f32 %f88, %f145;
|
||||
selp.f32 %f89, %f143, %f88, %p7;
|
||||
selp.f32 %f90, 0f00000000, %f144, %p7;
|
||||
mul.f32 %f91, %f89, %f89;
|
||||
fma.rn.f32 %f92, %f87, %f87, %f91;
|
||||
fma.rn.f32 %f93, %f90, %f90, %f92;
|
||||
sqrt.rn.f32 %f94, %f93;
|
||||
rcp.rn.f32 %f95, %f94;
|
||||
mul.f32 %f14, %f87, %f95;
|
||||
mul.f32 %f15, %f89, %f95;
|
||||
mul.f32 %f16, %f90, %f95;
|
||||
mul.f32 %f96, %f145, %f15;
|
||||
mul.f32 %f97, %f144, %f16;
|
||||
sub.f32 %f17, %f96, %f97;
|
||||
mul.f32 %f98, %f143, %f16;
|
||||
mul.f32 %f99, %f145, %f14;
|
||||
sub.f32 %f18, %f98, %f99;
|
||||
mul.f32 %f100, %f144, %f14;
|
||||
mul.f32 %f101, %f143, %f15;
|
||||
sub.f32 %f19, %f100, %f101;
|
||||
mov.f32 %f102, 0f3F8147AE;
|
||||
sqrt.rn.f32 %f103, %f102;
|
||||
rcp.rn.f32 %f20, %f103;
|
||||
neg.f32 %f21, %f20;
|
||||
mul.f32 %f22, %f20, 0f00000000;
|
||||
mul.f32 %f104, %f20, 0f3DCCCCCD;
|
||||
mul.f32 %f23, %f14, %f22;
|
||||
mul.f32 %f24, %f15, %f22;
|
||||
mul.f32 %f25, %f16, %f22;
|
||||
fma.rn.f32 %f105, %f17, %f21, %f23;
|
||||
fma.rn.f32 %f106, %f18, %f21, %f24;
|
||||
fma.rn.f32 %f107, %f19, %f21, %f25;
|
||||
mul.f32 %f26, %f143, %f104;
|
||||
mul.f32 %f27, %f144, %f104;
|
||||
mul.f32 %f28, %f145, %f104;
|
||||
add.f32 %f149, %f105, %f26;
|
||||
add.f32 %f148, %f106, %f27;
|
||||
add.f32 %f147, %f107, %f28;
|
||||
add.u64 %rd33, %SP, 0;
|
||||
cvta.to.local.u64 %rd35, %rd33;
|
||||
mov.u32 %r28, 0;
|
||||
st.local.u32 [%rd35], %r28;
|
||||
ld.global.u32 %r27, [root];
|
||||
mov.f32 %f81, 0f38D1B717;
|
||||
// inline asm
|
||||
call _rt_trace_64, (%r27, %f11, %f12, %f13, %f149, %f148, %f147, %r28, %f81, %f82, %rd33, %r5);
|
||||
// inline asm
|
||||
ld.local.f32 %f146, [%rd35];
|
||||
setp.gt.f32 %p8, %f146, 0f00000000;
|
||||
@%p8 bra BB0_7;
|
||||
|
||||
fma.rn.f32 %f116, %f17, %f20, %f23;
|
||||
fma.rn.f32 %f117, %f18, %f20, %f24;
|
||||
fma.rn.f32 %f118, %f19, %f20, %f25;
|
||||
add.f32 %f149, %f116, %f26;
|
||||
add.f32 %f148, %f117, %f27;
|
||||
add.f32 %f147, %f118, %f28;
|
||||
st.local.u32 [%rd35], %r28;
|
||||
ld.global.u32 %r46, [root];
|
||||
// inline asm
|
||||
call _rt_trace_64, (%r46, %f11, %f12, %f13, %f149, %f148, %f147, %r28, %f81, %f82, %rd33, %r5);
|
||||
// inline asm
|
||||
ld.local.f32 %f146, [%rd35];
|
||||
setp.gt.f32 %p9, %f146, 0f00000000;
|
||||
@%p9 bra BB0_7;
|
||||
|
||||
mul.f32 %f37, %f17, %f22;
|
||||
fma.rn.f32 %f127, %f14, %f21, %f37;
|
||||
mul.f32 %f38, %f18, %f22;
|
||||
fma.rn.f32 %f128, %f15, %f21, %f38;
|
||||
mul.f32 %f39, %f19, %f22;
|
||||
fma.rn.f32 %f129, %f16, %f21, %f39;
|
||||
add.f32 %f149, %f127, %f26;
|
||||
add.f32 %f148, %f128, %f27;
|
||||
add.f32 %f147, %f129, %f28;
|
||||
st.local.u32 [%rd35], %r28;
|
||||
ld.global.u32 %r49, [root];
|
||||
// inline asm
|
||||
call _rt_trace_64, (%r49, %f11, %f12, %f13, %f149, %f148, %f147, %r28, %f81, %f82, %rd33, %r5);
|
||||
// inline asm
|
||||
ld.local.f32 %f146, [%rd35];
|
||||
setp.gt.f32 %p10, %f146, 0f00000000;
|
||||
@%p10 bra BB0_7;
|
||||
|
||||
fma.rn.f32 %f138, %f14, %f20, %f37;
|
||||
fma.rn.f32 %f139, %f15, %f20, %f38;
|
||||
fma.rn.f32 %f140, %f16, %f20, %f39;
|
||||
add.f32 %f149, %f138, %f26;
|
||||
add.f32 %f148, %f139, %f27;
|
||||
add.f32 %f147, %f140, %f28;
|
||||
st.local.u32 [%rd35], %r28;
|
||||
ld.global.u32 %r52, [root];
|
||||
// inline asm
|
||||
call _rt_trace_64, (%r52, %f11, %f12, %f13, %f149, %f148, %f147, %r28, %f81, %f82, %rd33, %r5);
|
||||
// inline asm
|
||||
ld.local.f32 %f146, [%rd35];
|
||||
setp.leu.f32 %p11, %f146, 0f00000000;
|
||||
@%p11 bra BB0_8;
|
||||
|
||||
BB0_7:
|
||||
fma.rn.f32 %f150, %f146, %f149, %f11;
|
||||
fma.rn.f32 %f151, %f146, %f148, %f12;
|
||||
fma.rn.f32 %f152, %f146, %f147, %f13;
|
||||
|
||||
BB0_8:
|
||||
cvt.u64.u32 %rd45, %r3;
|
||||
cvt.u64.u32 %rd44, %r2;
|
||||
mov.u64 %rd48, image_HDR;
|
||||
cvta.global.u64 %rd43, %rd48;
|
||||
// inline asm
|
||||
call (%rd42), _rt_buffer_get_64, (%rd43, %r4, %r26, %rd44, %rd45, %rd7, %rd7);
|
||||
// inline asm
|
||||
mov.f32 %f141, 0f3F800000;
|
||||
st.v4.f32 [%rd42], {%f150, %f151, %f152, %f141};
|
||||
|
||||
BB0_10:
|
||||
ret;
|
||||
}
|
||||
|
||||
|
||||
8
Assets/Editor/x64/Bakery/fixPos.ptx.meta
Normal file
8
Assets/Editor/x64/Bakery/fixPos.ptx.meta
Normal file
@@ -0,0 +1,8 @@
|
||||
fileFormatVersion: 2
|
||||
guid: 04bc0e20327a72741902db186ca7e332
|
||||
timeCreated: 1526161016
|
||||
licenseType: Store
|
||||
DefaultImporter:
|
||||
userData:
|
||||
assetBundleName:
|
||||
assetBundleVariant:
|
||||
245
Assets/Editor/x64/Bakery/fixPos1.ptx
Normal file
245
Assets/Editor/x64/Bakery/fixPos1.ptx
Normal file
@@ -0,0 +1,245 @@
|
||||
//
|
||||
// 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 uvfacenormal[1];
|
||||
.global .align 1 .b8 uvpos[1];
|
||||
.global .align 1 .b8 uvsmoothpos[1];
|
||||
.global .align 1 .b8 triangleMarks[1];
|
||||
.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 8 .u64 _ZN21rti_internal_register20reg_bitness_detectorE;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail0E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail1E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail2E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail3E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail4E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail5E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail6E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail7E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail8E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail9E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail0E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail1E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail2E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail3E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail4E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail5E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail6E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail7E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail8E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail9E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register14reg_rayIndex_xE;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register14reg_rayIndex_yE;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register14reg_rayIndex_zE;
|
||||
.global .align 8 .b8 _ZN21rti_internal_typename7pixelIDE[6] = {117, 105, 110, 116, 50, 0};
|
||||
.global .align 8 .b8 _ZN21rti_internal_typename10resolutionE[6] = {117, 105, 110, 116, 50, 0};
|
||||
.global .align 8 .b8 _ZN21rti_internal_typename6normalE[7] = {102, 108, 111, 97, 116, 51, 0};
|
||||
.global .align 8 .b8 _ZN21rti_internal_typename6camPosE[7] = {102, 108, 111, 97, 116, 51, 0};
|
||||
.global .align 16 .b8 _ZN21rti_internal_typename4rootE[9] = {114, 116, 79, 98, 106, 101, 99, 116, 0};
|
||||
.global .align 4 .b8 _ZN21rti_internal_typename12imageEnabledE[4] = {105, 110, 116, 0};
|
||||
.global .align 8 .b8 _ZN21rti_internal_typename8tileInfoE[6] = {117, 105, 110, 116, 52, 0};
|
||||
.global .align 4 .b8 _ZN21rti_internal_typename8additiveE[4] = {105, 110, 116, 0};
|
||||
.global .align 4 .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 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 _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];
|
||||
|
||||
.visible .entry _Z6oxMainv(
|
||||
|
||||
)
|
||||
{
|
||||
.local .align 4 .b8 __local_depot0[4];
|
||||
.reg .b64 %SP;
|
||||
.reg .b64 %SPL;
|
||||
.reg .pred %p<8>;
|
||||
.reg .b16 %rs<10>;
|
||||
.reg .f32 %f<50>;
|
||||
.reg .b32 %r<46>;
|
||||
.reg .b64 %rd<51>;
|
||||
|
||||
|
||||
mov.u64 %rd50, __local_depot0;
|
||||
cvta.local.u64 %SP, %rd50;
|
||||
ld.global.v2.u32 {%r4, %r5}, [pixelID];
|
||||
cvt.u64.u32 %rd4, %r4;
|
||||
cvt.u64.u32 %rd5, %r5;
|
||||
mov.u64 %rd8, uvfacenormal;
|
||||
cvta.global.u64 %rd3, %rd8;
|
||||
mov.u32 %r2, 2;
|
||||
mov.u32 %r3, 4;
|
||||
mov.u64 %rd7, 0;
|
||||
// inline asm
|
||||
call (%rd2), _rt_buffer_get_64, (%rd3, %r2, %r3, %rd4, %rd5, %rd7, %rd7);
|
||||
// inline asm
|
||||
ld.u32 %r1, [%rd2];
|
||||
shr.u32 %r8, %r1, 16;
|
||||
cvt.u16.u32 %rs1, %r8;
|
||||
and.b16 %rs2, %rs1, 255;
|
||||
cvt.u16.u32 %rs3, %r1;
|
||||
or.b16 %rs4, %rs3, %rs2;
|
||||
setp.eq.s16 %p1, %rs4, 0;
|
||||
mov.f32 %f47, 0f00000000;
|
||||
mov.f32 %f48, %f47;
|
||||
mov.f32 %f49, %f47;
|
||||
@%p1 bra BB0_2;
|
||||
|
||||
ld.u8 %rs5, [%rd2+1];
|
||||
and.b16 %rs7, %rs3, 255;
|
||||
cvt.rn.f32.u16 %f11, %rs7;
|
||||
div.rn.f32 %f12, %f11, 0f437F0000;
|
||||
fma.rn.f32 %f13, %f12, 0f40000000, 0fBF800000;
|
||||
cvt.rn.f32.u16 %f14, %rs5;
|
||||
div.rn.f32 %f15, %f14, 0f437F0000;
|
||||
fma.rn.f32 %f16, %f15, 0f40000000, 0fBF800000;
|
||||
cvt.rn.f32.u16 %f17, %rs2;
|
||||
div.rn.f32 %f18, %f17, 0f437F0000;
|
||||
fma.rn.f32 %f19, %f18, 0f40000000, 0fBF800000;
|
||||
mul.f32 %f20, %f16, %f16;
|
||||
fma.rn.f32 %f21, %f13, %f13, %f20;
|
||||
fma.rn.f32 %f22, %f19, %f19, %f21;
|
||||
sqrt.rn.f32 %f23, %f22;
|
||||
rcp.rn.f32 %f24, %f23;
|
||||
mul.f32 %f47, %f13, %f24;
|
||||
mul.f32 %f48, %f16, %f24;
|
||||
mul.f32 %f49, %f19, %f24;
|
||||
|
||||
BB0_2:
|
||||
setp.eq.f32 %p2, %f48, 0f00000000;
|
||||
setp.eq.f32 %p3, %f47, 0f00000000;
|
||||
and.pred %p4, %p3, %p2;
|
||||
setp.eq.f32 %p5, %f49, 0f00000000;
|
||||
and.pred %p6, %p4, %p5;
|
||||
@%p6 bra BB0_5;
|
||||
|
||||
ld.global.v2.u32 {%r22, %r23}, [pixelID];
|
||||
cvt.u64.u32 %rd11, %r22;
|
||||
cvt.u64.u32 %rd12, %r23;
|
||||
mov.u64 %rd40, uvpos;
|
||||
cvta.global.u64 %rd10, %rd40;
|
||||
mov.u32 %r10, 12;
|
||||
// inline asm
|
||||
call (%rd9), _rt_buffer_get_64, (%rd10, %r2, %r10, %rd11, %rd12, %rd7, %rd7);
|
||||
// inline asm
|
||||
ld.f32 %f33, [%rd9+8];
|
||||
ld.f32 %f34, [%rd9+4];
|
||||
ld.f32 %f35, [%rd9];
|
||||
fma.rn.f32 %f25, %f47, 0f38D1B717, %f35;
|
||||
fma.rn.f32 %f26, %f48, 0f38D1B717, %f34;
|
||||
fma.rn.f32 %f27, %f49, 0f38D1B717, %f33;
|
||||
ld.global.v2.u32 {%r26, %r27}, [pixelID];
|
||||
cvt.u64.u32 %rd17, %r26;
|
||||
cvt.u64.u32 %rd18, %r27;
|
||||
mov.u64 %rd41, uvsmoothpos;
|
||||
cvta.global.u64 %rd16, %rd41;
|
||||
mov.u32 %r18, 16;
|
||||
// inline asm
|
||||
call (%rd15), _rt_buffer_get_64, (%rd16, %r2, %r18, %rd17, %rd18, %rd7, %rd7);
|
||||
// inline asm
|
||||
ld.f32 %f36, [%rd15];
|
||||
ld.global.v2.u32 {%r30, %r31}, [pixelID];
|
||||
cvt.u64.u32 %rd23, %r30;
|
||||
cvt.u64.u32 %rd24, %r31;
|
||||
// inline asm
|
||||
call (%rd21), _rt_buffer_get_64, (%rd16, %r2, %r18, %rd23, %rd24, %rd7, %rd7);
|
||||
// inline asm
|
||||
ld.f32 %f37, [%rd21+4];
|
||||
ld.global.v2.u32 {%r34, %r35}, [pixelID];
|
||||
cvt.u64.u32 %rd29, %r34;
|
||||
cvt.u64.u32 %rd30, %r35;
|
||||
// inline asm
|
||||
call (%rd27), _rt_buffer_get_64, (%rd16, %r2, %r18, %rd29, %rd30, %rd7, %rd7);
|
||||
// inline asm
|
||||
ld.f32 %f38, [%rd27+8];
|
||||
ld.global.v2.u32 {%r38, %r39}, [pixelID];
|
||||
cvt.u64.u32 %rd35, %r38;
|
||||
cvt.u64.u32 %rd36, %r39;
|
||||
// inline asm
|
||||
call (%rd33), _rt_buffer_get_64, (%rd16, %r2, %r18, %rd35, %rd36, %rd7, %rd7);
|
||||
// inline asm
|
||||
ld.f32 %f7, [%rd33+12];
|
||||
sub.f32 %f39, %f36, %f35;
|
||||
sub.f32 %f40, %f37, %f34;
|
||||
sub.f32 %f41, %f38, %f33;
|
||||
mul.f32 %f42, %f40, %f40;
|
||||
fma.rn.f32 %f43, %f39, %f39, %f42;
|
||||
fma.rn.f32 %f44, %f41, %f41, %f43;
|
||||
sqrt.rn.f32 %f32, %f44;
|
||||
rcp.rn.f32 %f45, %f32;
|
||||
mul.f32 %f28, %f39, %f45;
|
||||
mul.f32 %f29, %f40, %f45;
|
||||
mul.f32 %f30, %f41, %f45;
|
||||
add.u64 %rd39, %SP, 0;
|
||||
cvta.to.local.u64 %rd42, %rd39;
|
||||
mov.u32 %r42, -1082130432;
|
||||
st.local.u32 [%rd42], %r42;
|
||||
ld.global.u32 %r19, [root];
|
||||
mov.u32 %r20, 0;
|
||||
mov.f32 %f31, 0f38D1B717;
|
||||
// inline asm
|
||||
call _rt_trace_64, (%r19, %f25, %f26, %f27, %f28, %f29, %f30, %r20, %f31, %f32, %rd39, %r3);
|
||||
// inline asm
|
||||
ld.local.f32 %f46, [%rd42];
|
||||
setp.ltu.f32 %p7, %f46, 0f00000000;
|
||||
@%p7 bra BB0_5;
|
||||
|
||||
cvt.rzi.u32.f32 %r45, %f7;
|
||||
cvt.u64.u32 %rd45, %r45;
|
||||
mov.u64 %rd49, triangleMarks;
|
||||
cvta.global.u64 %rd44, %rd49;
|
||||
mov.u32 %r44, 1;
|
||||
// inline asm
|
||||
call (%rd43), _rt_buffer_get_64, (%rd44, %r44, %r44, %rd45, %rd7, %rd7, %rd7);
|
||||
// inline asm
|
||||
mov.u16 %rs9, 255;
|
||||
st.u8 [%rd43], %rs9;
|
||||
|
||||
BB0_5:
|
||||
ret;
|
||||
}
|
||||
|
||||
|
||||
8
Assets/Editor/x64/Bakery/fixPos1.ptx.meta
Normal file
8
Assets/Editor/x64/Bakery/fixPos1.ptx.meta
Normal file
@@ -0,0 +1,8 @@
|
||||
fileFormatVersion: 2
|
||||
guid: fdf30ebf12520be429e3ad31eb009930
|
||||
timeCreated: 1526161016
|
||||
licenseType: Store
|
||||
DefaultImporter:
|
||||
userData:
|
||||
assetBundleName:
|
||||
assetBundleVariant:
|
||||
684
Assets/Editor/x64/Bakery/fixPos12.ptx
Normal file
684
Assets/Editor/x64/Bakery/fixPos12.ptx
Normal file
@@ -0,0 +1,684 @@
|
||||
//
|
||||
// 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_HDR[1];
|
||||
.global .align 1 .b8 image_HDR2[1];
|
||||
.global .align 1 .b8 uvfacenormal[1];
|
||||
.global .align 1 .b8 uvpos[1];
|
||||
.global .align 1 .b8 uvsmoothpos[1];
|
||||
.global .align 1 .b8 triangleMarks[1];
|
||||
.global .align 1 .b8 image_Mask[1];
|
||||
.global .align 4 .f32 fakeBias;
|
||||
.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_typeinfo8fakeBiasE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
|
||||
.global .align 8 .u64 _ZN21rti_internal_register20reg_bitness_detectorE;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail0E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail1E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail2E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail3E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail4E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail5E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail6E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail7E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail8E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail9E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail0E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail1E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail2E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail3E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail4E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail5E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail6E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail7E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail8E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail9E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register14reg_rayIndex_xE;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register14reg_rayIndex_yE;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register14reg_rayIndex_zE;
|
||||
.global .align 8 .b8 _ZN21rti_internal_typename7pixelIDE[6] = {117, 105, 110, 116, 50, 0};
|
||||
.global .align 8 .b8 _ZN21rti_internal_typename10resolutionE[6] = {117, 105, 110, 116, 50, 0};
|
||||
.global .align 8 .b8 _ZN21rti_internal_typename6normalE[7] = {102, 108, 111, 97, 116, 51, 0};
|
||||
.global .align 8 .b8 _ZN21rti_internal_typename6camPosE[7] = {102, 108, 111, 97, 116, 51, 0};
|
||||
.global .align 16 .b8 _ZN21rti_internal_typename4rootE[9] = {114, 116, 79, 98, 106, 101, 99, 116, 0};
|
||||
.global .align 4 .b8 _ZN21rti_internal_typename12imageEnabledE[4] = {105, 110, 116, 0};
|
||||
.global .align 8 .b8 _ZN21rti_internal_typename8tileInfoE[6] = {117, 105, 110, 116, 52, 0};
|
||||
.global .align 4 .b8 _ZN21rti_internal_typename8additiveE[4] = {105, 110, 116, 0};
|
||||
.global .align 8 .b8 _ZN21rti_internal_typename8fakeBiasE[6] = {102, 108, 111, 97, 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_typeenum8fakeBiasE = 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_semantic8fakeBiasE[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_annotation8fakeBiasE[1];
|
||||
|
||||
.visible .entry _Z6oxMainv(
|
||||
|
||||
)
|
||||
{
|
||||
.local .align 4 .b8 __local_depot0[28];
|
||||
.reg .b64 %SP;
|
||||
.reg .b64 %SPL;
|
||||
.reg .pred %p<27>;
|
||||
.reg .b16 %rs<12>;
|
||||
.reg .f32 %f<370>;
|
||||
.reg .b32 %r<129>;
|
||||
.reg .b64 %rd<124>;
|
||||
|
||||
|
||||
mov.u64 %rd123, __local_depot0;
|
||||
cvta.local.u64 %SP, %rd123;
|
||||
ld.global.v2.u32 {%r8, %r9}, [pixelID];
|
||||
cvt.u64.u32 %rd8, %r8;
|
||||
cvt.u64.u32 %rd9, %r9;
|
||||
mov.u64 %rd12, uvfacenormal;
|
||||
cvta.global.u64 %rd7, %rd12;
|
||||
mov.u32 %r6, 2;
|
||||
mov.u32 %r7, 4;
|
||||
mov.u64 %rd11, 0;
|
||||
// inline asm
|
||||
call (%rd6), _rt_buffer_get_64, (%rd7, %r6, %r7, %rd8, %rd9, %rd11, %rd11);
|
||||
// inline asm
|
||||
ld.u32 %r1, [%rd6];
|
||||
shr.u32 %r12, %r1, 16;
|
||||
cvt.u16.u32 %rs1, %r12;
|
||||
and.b16 %rs2, %rs1, 255;
|
||||
cvt.u16.u32 %rs3, %r1;
|
||||
or.b16 %rs4, %rs3, %rs2;
|
||||
setp.eq.s16 %p1, %rs4, 0;
|
||||
mov.f32 %f348, 0f00000000;
|
||||
mov.f32 %f349, %f348;
|
||||
mov.f32 %f350, %f348;
|
||||
@%p1 bra BB0_2;
|
||||
|
||||
ld.u8 %rs5, [%rd6+1];
|
||||
and.b16 %rs7, %rs3, 255;
|
||||
cvt.rn.f32.u16 %f112, %rs7;
|
||||
div.rn.f32 %f113, %f112, 0f437F0000;
|
||||
fma.rn.f32 %f114, %f113, 0f40000000, 0fBF800000;
|
||||
cvt.rn.f32.u16 %f115, %rs5;
|
||||
div.rn.f32 %f116, %f115, 0f437F0000;
|
||||
fma.rn.f32 %f117, %f116, 0f40000000, 0fBF800000;
|
||||
cvt.rn.f32.u16 %f118, %rs2;
|
||||
div.rn.f32 %f119, %f118, 0f437F0000;
|
||||
fma.rn.f32 %f120, %f119, 0f40000000, 0fBF800000;
|
||||
mul.f32 %f121, %f117, %f117;
|
||||
fma.rn.f32 %f122, %f114, %f114, %f121;
|
||||
fma.rn.f32 %f123, %f120, %f120, %f122;
|
||||
sqrt.rn.f32 %f124, %f123;
|
||||
rcp.rn.f32 %f125, %f124;
|
||||
mul.f32 %f348, %f114, %f125;
|
||||
mul.f32 %f349, %f117, %f125;
|
||||
mul.f32 %f350, %f120, %f125;
|
||||
|
||||
BB0_2:
|
||||
ld.global.v2.u32 {%r13, %r14}, [pixelID];
|
||||
ld.global.v2.u32 {%r16, %r17}, [tileInfo];
|
||||
add.s32 %r2, %r13, %r16;
|
||||
add.s32 %r3, %r14, %r17;
|
||||
setp.eq.f32 %p2, %f349, 0f00000000;
|
||||
setp.eq.f32 %p3, %f348, 0f00000000;
|
||||
and.pred %p4, %p3, %p2;
|
||||
setp.eq.f32 %p5, %f350, 0f00000000;
|
||||
and.pred %p6, %p4, %p5;
|
||||
@%p6 bra BB0_27;
|
||||
bra.uni BB0_3;
|
||||
|
||||
BB0_27:
|
||||
cvt.u64.u32 %rd107, %r2;
|
||||
cvt.u64.u32 %rd108, %r3;
|
||||
mov.u64 %rd111, image_HDR;
|
||||
cvta.global.u64 %rd106, %rd111;
|
||||
mov.u32 %r109, 16;
|
||||
// inline asm
|
||||
call (%rd105), _rt_buffer_get_64, (%rd106, %r6, %r109, %rd107, %rd108, %rd11, %rd11);
|
||||
// inline asm
|
||||
mov.f32 %f336, 0f00000000;
|
||||
st.v4.f32 [%rd105], {%f336, %f336, %f336, %f336};
|
||||
bra.uni BB0_28;
|
||||
|
||||
BB0_3:
|
||||
ld.global.v2.u32 {%r39, %r40}, [pixelID];
|
||||
cvt.u64.u32 %rd15, %r39;
|
||||
cvt.u64.u32 %rd16, %r40;
|
||||
mov.u64 %rd67, uvpos;
|
||||
cvta.global.u64 %rd14, %rd67;
|
||||
mov.u32 %r36, 16;
|
||||
// inline asm
|
||||
call (%rd13), _rt_buffer_get_64, (%rd14, %r6, %r36, %rd15, %rd16, %rd11, %rd11);
|
||||
// inline asm
|
||||
ld.f32 %f7, [%rd13];
|
||||
ld.global.v2.u32 {%r43, %r44}, [pixelID];
|
||||
cvt.u64.u32 %rd21, %r43;
|
||||
cvt.u64.u32 %rd22, %r44;
|
||||
// inline asm
|
||||
call (%rd19), _rt_buffer_get_64, (%rd14, %r6, %r36, %rd21, %rd22, %rd11, %rd11);
|
||||
// inline asm
|
||||
ld.f32 %f8, [%rd19+4];
|
||||
ld.global.v2.u32 {%r47, %r48}, [pixelID];
|
||||
cvt.u64.u32 %rd27, %r47;
|
||||
cvt.u64.u32 %rd28, %r48;
|
||||
// inline asm
|
||||
call (%rd25), _rt_buffer_get_64, (%rd14, %r6, %r36, %rd27, %rd28, %rd11, %rd11);
|
||||
// inline asm
|
||||
ld.f32 %f9, [%rd25+8];
|
||||
ld.global.v2.u32 {%r51, %r52}, [pixelID];
|
||||
cvt.u64.u32 %rd33, %r51;
|
||||
cvt.u64.u32 %rd34, %r52;
|
||||
// inline asm
|
||||
call (%rd31), _rt_buffer_get_64, (%rd14, %r6, %r36, %rd33, %rd34, %rd11, %rd11);
|
||||
// inline asm
|
||||
ld.f32 %f126, [%rd31+12];
|
||||
add.f32 %f10, %f126, 0f38D1B717;
|
||||
mul.f32 %f127, %f7, 0f3456BF95;
|
||||
mul.f32 %f128, %f8, 0f3456BF95;
|
||||
mul.f32 %f129, %f9, 0f3456BF95;
|
||||
abs.f32 %f11, %f348;
|
||||
div.rn.f32 %f130, %f127, %f11;
|
||||
abs.f32 %f131, %f349;
|
||||
div.rn.f32 %f132, %f128, %f131;
|
||||
abs.f32 %f12, %f350;
|
||||
div.rn.f32 %f133, %f129, %f12;
|
||||
abs.f32 %f134, %f130;
|
||||
abs.f32 %f135, %f132;
|
||||
abs.f32 %f136, %f133;
|
||||
mov.f32 %f137, 0f38D1B717;
|
||||
max.f32 %f138, %f134, %f137;
|
||||
max.f32 %f139, %f135, %f137;
|
||||
max.f32 %f140, %f136, %f137;
|
||||
fma.rn.f32 %f13, %f348, %f138, %f7;
|
||||
fma.rn.f32 %f14, %f349, %f139, %f8;
|
||||
fma.rn.f32 %f15, %f350, %f140, %f9;
|
||||
ld.global.v2.u32 {%r55, %r56}, [pixelID];
|
||||
cvt.u64.u32 %rd39, %r55;
|
||||
cvt.u64.u32 %rd40, %r56;
|
||||
mov.u64 %rd68, uvsmoothpos;
|
||||
cvta.global.u64 %rd38, %rd68;
|
||||
// inline asm
|
||||
call (%rd37), _rt_buffer_get_64, (%rd38, %r6, %r36, %rd39, %rd40, %rd11, %rd11);
|
||||
// inline asm
|
||||
ld.f32 %f16, [%rd37];
|
||||
ld.global.v2.u32 {%r59, %r60}, [pixelID];
|
||||
cvt.u64.u32 %rd45, %r59;
|
||||
cvt.u64.u32 %rd46, %r60;
|
||||
// inline asm
|
||||
call (%rd43), _rt_buffer_get_64, (%rd38, %r6, %r36, %rd45, %rd46, %rd11, %rd11);
|
||||
// inline asm
|
||||
ld.f32 %f17, [%rd43+4];
|
||||
ld.global.v2.u32 {%r63, %r64}, [pixelID];
|
||||
cvt.u64.u32 %rd51, %r63;
|
||||
cvt.u64.u32 %rd52, %r64;
|
||||
// inline asm
|
||||
call (%rd49), _rt_buffer_get_64, (%rd38, %r6, %r36, %rd51, %rd52, %rd11, %rd11);
|
||||
// inline asm
|
||||
ld.f32 %f18, [%rd49+8];
|
||||
ld.global.v2.u32 {%r67, %r68}, [pixelID];
|
||||
cvt.u64.u32 %rd57, %r67;
|
||||
cvt.u64.u32 %rd58, %r68;
|
||||
// inline asm
|
||||
call (%rd55), _rt_buffer_get_64, (%rd38, %r6, %r36, %rd57, %rd58, %rd11, %rd11);
|
||||
// inline asm
|
||||
ld.f32 %f141, [%rd55+12];
|
||||
cvt.rzi.u32.f32 %r71, %f141;
|
||||
cvt.u64.u32 %rd63, %r71;
|
||||
mov.u64 %rd69, triangleMarks;
|
||||
cvta.global.u64 %rd62, %rd69;
|
||||
mov.u32 %r38, 1;
|
||||
// inline asm
|
||||
call (%rd61), _rt_buffer_get_64, (%rd62, %r38, %r38, %rd63, %rd11, %rd11, %rd11);
|
||||
// inline asm
|
||||
ld.u8 %r128, [%rd61];
|
||||
sub.f32 %f19, %f16, %f7;
|
||||
sub.f32 %f20, %f17, %f8;
|
||||
sub.f32 %f21, %f18, %f9;
|
||||
mul.f32 %f142, %f20, %f20;
|
||||
fma.rn.f32 %f143, %f19, %f19, %f142;
|
||||
fma.rn.f32 %f144, %f21, %f21, %f143;
|
||||
sqrt.rn.f32 %f22, %f144;
|
||||
setp.leu.f32 %p7, %f22, 0f3727C5AC;
|
||||
setp.eq.s32 %p8, %r128, 255;
|
||||
or.pred %p9, %p8, %p7;
|
||||
@%p9 bra BB0_6;
|
||||
|
||||
rcp.rn.f32 %f153, %f22;
|
||||
mul.f32 %f148, %f19, %f153;
|
||||
mul.f32 %f149, %f20, %f153;
|
||||
mul.f32 %f150, %f21, %f153;
|
||||
add.u64 %rd70, %SP, 12;
|
||||
cvta.to.local.u64 %rd71, %rd70;
|
||||
mov.u32 %r75, 1065353216;
|
||||
st.local.u32 [%rd71], %r75;
|
||||
ld.global.u32 %r72, [root];
|
||||
// inline asm
|
||||
call _rt_trace_64, (%r72, %f13, %f14, %f15, %f148, %f149, %f150, %r38, %f137, %f22, %rd70, %r7);
|
||||
// inline asm
|
||||
ld.local.f32 %f154, [%rd71];
|
||||
setp.neu.f32 %p10, %f154, 0f00000000;
|
||||
@%p10 bra BB0_6;
|
||||
|
||||
// inline asm
|
||||
call (%rd72), _rt_buffer_get_64, (%rd62, %r38, %r38, %rd63, %rd11, %rd11, %rd11);
|
||||
// inline asm
|
||||
mov.u16 %rs9, 255;
|
||||
st.u8 [%rd72], %rs9;
|
||||
mov.u32 %r128, 255;
|
||||
|
||||
BB0_6:
|
||||
setp.eq.s32 %p11, %r128, 0;
|
||||
selp.f32 %f355, %f16, %f7, %p11;
|
||||
selp.f32 %f356, %f17, %f8, %p11;
|
||||
selp.f32 %f357, %f18, %f9, %p11;
|
||||
neg.f32 %f164, %f349;
|
||||
setp.gt.f32 %p12, %f11, %f12;
|
||||
selp.f32 %f165, %f164, 0f00000000, %p12;
|
||||
neg.f32 %f166, %f350;
|
||||
selp.f32 %f167, %f348, %f166, %p12;
|
||||
selp.f32 %f168, 0f00000000, %f349, %p12;
|
||||
mul.f32 %f169, %f167, %f167;
|
||||
fma.rn.f32 %f170, %f165, %f165, %f169;
|
||||
fma.rn.f32 %f171, %f168, %f168, %f170;
|
||||
sqrt.rn.f32 %f172, %f171;
|
||||
rcp.rn.f32 %f173, %f172;
|
||||
mul.f32 %f26, %f165, %f173;
|
||||
mul.f32 %f27, %f167, %f173;
|
||||
mul.f32 %f28, %f168, %f173;
|
||||
mul.f32 %f174, %f350, %f27;
|
||||
mul.f32 %f175, %f349, %f28;
|
||||
sub.f32 %f29, %f174, %f175;
|
||||
mul.f32 %f176, %f348, %f28;
|
||||
mul.f32 %f177, %f350, %f26;
|
||||
sub.f32 %f30, %f176, %f177;
|
||||
mul.f32 %f178, %f349, %f26;
|
||||
mul.f32 %f179, %f348, %f27;
|
||||
sub.f32 %f31, %f178, %f179;
|
||||
mov.f32 %f180, 0f3F8147AE;
|
||||
sqrt.rn.f32 %f181, %f180;
|
||||
rcp.rn.f32 %f32, %f181;
|
||||
neg.f32 %f33, %f32;
|
||||
mul.f32 %f34, %f32, 0f00000000;
|
||||
mul.f32 %f182, %f32, 0f3DCCCCCD;
|
||||
mul.f32 %f183, %f13, 0f3456BF95;
|
||||
abs.f32 %f184, %f183;
|
||||
mul.f32 %f185, %f14, 0f3456BF95;
|
||||
abs.f32 %f186, %f185;
|
||||
mul.f32 %f187, %f15, 0f3456BF95;
|
||||
abs.f32 %f188, %f187;
|
||||
max.f32 %f189, %f184, %f186;
|
||||
max.f32 %f190, %f189, %f188;
|
||||
max.f32 %f35, %f190, %f137;
|
||||
mul.f32 %f36, %f26, %f34;
|
||||
mul.f32 %f37, %f27, %f34;
|
||||
mul.f32 %f38, %f28, %f34;
|
||||
fma.rn.f32 %f192, %f29, %f33, %f36;
|
||||
fma.rn.f32 %f193, %f30, %f33, %f37;
|
||||
fma.rn.f32 %f194, %f31, %f33, %f38;
|
||||
mul.f32 %f39, %f348, %f182;
|
||||
mul.f32 %f40, %f349, %f182;
|
||||
mul.f32 %f41, %f350, %f182;
|
||||
add.f32 %f158, %f192, %f39;
|
||||
add.f32 %f159, %f193, %f40;
|
||||
add.f32 %f160, %f194, %f41;
|
||||
add.u64 %rd79, %SP, 16;
|
||||
cvta.to.local.u64 %rd80, %rd79;
|
||||
mov.u32 %r80, 0;
|
||||
st.local.u32 [%rd80+8], %r80;
|
||||
st.local.u32 [%rd80+4], %r80;
|
||||
st.local.u32 [%rd80], %r80;
|
||||
ld.global.u32 %r79, [root];
|
||||
mul.f32 %f162, %f10, 0f3FB504F3;
|
||||
mov.f32 %f161, 0f00000000;
|
||||
mov.u32 %r81, 12;
|
||||
// inline asm
|
||||
call _rt_trace_64, (%r79, %f13, %f14, %f15, %f158, %f159, %f160, %r80, %f161, %f162, %rd79, %r81);
|
||||
// inline asm
|
||||
ld.local.f32 %f46, [%rd80+4];
|
||||
ld.local.f32 %f47, [%rd80];
|
||||
add.f32 %f195, %f47, %f46;
|
||||
ld.local.f32 %f48, [%rd80+8];
|
||||
add.f32 %f196, %f195, %f48;
|
||||
mov.f32 %f358, 0f47C34F80;
|
||||
setp.eq.f32 %p13, %f196, 0f00000000;
|
||||
@%p13 bra BB0_9;
|
||||
|
||||
mul.f32 %f198, %f46, %f46;
|
||||
fma.rn.f32 %f199, %f47, %f47, %f198;
|
||||
fma.rn.f32 %f200, %f48, %f48, %f199;
|
||||
sqrt.rn.f32 %f49, %f200;
|
||||
rcp.rn.f32 %f201, %f49;
|
||||
mul.f32 %f202, %f201, %f47;
|
||||
mul.f32 %f203, %f201, %f46;
|
||||
mul.f32 %f204, %f201, %f48;
|
||||
fma.rn.f32 %f205, %f35, %f202, %f13;
|
||||
fma.rn.f32 %f206, %f35, %f203, %f14;
|
||||
fma.rn.f32 %f207, %f35, %f204, %f15;
|
||||
fma.rn.f32 %f50, %f49, %f158, %f205;
|
||||
fma.rn.f32 %f51, %f49, %f159, %f206;
|
||||
fma.rn.f32 %f52, %f49, %f160, %f207;
|
||||
setp.geu.f32 %p14, %f49, 0f47C34F80;
|
||||
@%p14 bra BB0_9;
|
||||
|
||||
mov.f32 %f355, %f50;
|
||||
mov.f32 %f356, %f51;
|
||||
mov.f32 %f357, %f52;
|
||||
mov.f32 %f358, %f49;
|
||||
|
||||
BB0_9:
|
||||
mov.u32 %r111, 12;
|
||||
mov.f32 %f337, 0f00000000;
|
||||
add.u64 %rd113, %SP, 16;
|
||||
cvta.to.local.u64 %rd112, %rd113;
|
||||
mov.u32 %r110, 0;
|
||||
fma.rn.f32 %f216, %f29, %f32, %f36;
|
||||
fma.rn.f32 %f217, %f30, %f32, %f37;
|
||||
fma.rn.f32 %f218, %f31, %f32, %f38;
|
||||
add.f32 %f211, %f216, %f39;
|
||||
add.f32 %f212, %f217, %f40;
|
||||
add.f32 %f213, %f218, %f41;
|
||||
st.local.u32 [%rd112+8], %r110;
|
||||
st.local.u32 [%rd112+4], %r110;
|
||||
st.local.u32 [%rd112], %r110;
|
||||
ld.global.u32 %r82, [root];
|
||||
// inline asm
|
||||
call _rt_trace_64, (%r82, %f13, %f14, %f15, %f211, %f212, %f213, %r110, %f337, %f162, %rd113, %r111);
|
||||
// inline asm
|
||||
ld.local.f32 %f60, [%rd112+4];
|
||||
ld.local.f32 %f61, [%rd112];
|
||||
add.f32 %f219, %f61, %f60;
|
||||
ld.local.f32 %f62, [%rd112+8];
|
||||
add.f32 %f220, %f219, %f62;
|
||||
setp.eq.f32 %p15, %f220, 0f00000000;
|
||||
@%p15 bra BB0_12;
|
||||
|
||||
mul.f32 %f221, %f60, %f60;
|
||||
fma.rn.f32 %f222, %f61, %f61, %f221;
|
||||
fma.rn.f32 %f223, %f62, %f62, %f222;
|
||||
sqrt.rn.f32 %f63, %f223;
|
||||
rcp.rn.f32 %f224, %f63;
|
||||
mul.f32 %f225, %f224, %f61;
|
||||
mul.f32 %f226, %f224, %f60;
|
||||
mul.f32 %f227, %f224, %f62;
|
||||
fma.rn.f32 %f228, %f35, %f225, %f13;
|
||||
fma.rn.f32 %f229, %f35, %f226, %f14;
|
||||
fma.rn.f32 %f230, %f35, %f227, %f15;
|
||||
fma.rn.f32 %f64, %f63, %f211, %f228;
|
||||
fma.rn.f32 %f65, %f63, %f212, %f229;
|
||||
fma.rn.f32 %f66, %f63, %f213, %f230;
|
||||
setp.geu.f32 %p16, %f63, %f358;
|
||||
@%p16 bra BB0_12;
|
||||
|
||||
mov.f32 %f355, %f64;
|
||||
mov.f32 %f356, %f65;
|
||||
mov.f32 %f357, %f66;
|
||||
mov.f32 %f358, %f63;
|
||||
|
||||
BB0_12:
|
||||
neg.f32 %f340, %f32;
|
||||
mul.f32 %f339, %f32, 0f00000000;
|
||||
mov.u32 %r113, 12;
|
||||
mov.f32 %f338, 0f00000000;
|
||||
add.u64 %rd116, %SP, 16;
|
||||
cvta.to.local.u64 %rd115, %rd116;
|
||||
mov.u32 %r112, 0;
|
||||
mul.f32 %f71, %f29, %f339;
|
||||
fma.rn.f32 %f239, %f26, %f340, %f71;
|
||||
mul.f32 %f72, %f30, %f339;
|
||||
fma.rn.f32 %f240, %f27, %f340, %f72;
|
||||
mul.f32 %f73, %f31, %f339;
|
||||
fma.rn.f32 %f241, %f28, %f340, %f73;
|
||||
add.f32 %f234, %f239, %f39;
|
||||
add.f32 %f235, %f240, %f40;
|
||||
add.f32 %f236, %f241, %f41;
|
||||
st.local.u32 [%rd115+8], %r112;
|
||||
st.local.u32 [%rd115+4], %r112;
|
||||
st.local.u32 [%rd115], %r112;
|
||||
ld.global.u32 %r85, [root];
|
||||
// inline asm
|
||||
call _rt_trace_64, (%r85, %f13, %f14, %f15, %f234, %f235, %f236, %r112, %f338, %f162, %rd116, %r113);
|
||||
// inline asm
|
||||
ld.local.f32 %f77, [%rd115+4];
|
||||
ld.local.f32 %f78, [%rd115];
|
||||
add.f32 %f242, %f78, %f77;
|
||||
ld.local.f32 %f79, [%rd115+8];
|
||||
add.f32 %f243, %f242, %f79;
|
||||
setp.eq.f32 %p17, %f243, 0f00000000;
|
||||
@%p17 bra BB0_15;
|
||||
|
||||
mul.f32 %f244, %f77, %f77;
|
||||
fma.rn.f32 %f245, %f78, %f78, %f244;
|
||||
fma.rn.f32 %f246, %f79, %f79, %f245;
|
||||
sqrt.rn.f32 %f80, %f246;
|
||||
rcp.rn.f32 %f247, %f80;
|
||||
mul.f32 %f248, %f247, %f78;
|
||||
mul.f32 %f249, %f247, %f77;
|
||||
mul.f32 %f250, %f247, %f79;
|
||||
fma.rn.f32 %f251, %f35, %f248, %f13;
|
||||
fma.rn.f32 %f252, %f35, %f249, %f14;
|
||||
fma.rn.f32 %f253, %f35, %f250, %f15;
|
||||
fma.rn.f32 %f81, %f80, %f234, %f251;
|
||||
fma.rn.f32 %f82, %f80, %f235, %f252;
|
||||
fma.rn.f32 %f83, %f80, %f236, %f253;
|
||||
setp.geu.f32 %p18, %f80, %f358;
|
||||
@%p18 bra BB0_15;
|
||||
|
||||
mov.f32 %f355, %f81;
|
||||
mov.f32 %f356, %f82;
|
||||
mov.f32 %f357, %f83;
|
||||
mov.f32 %f358, %f80;
|
||||
|
||||
BB0_15:
|
||||
mov.u32 %r115, 12;
|
||||
mov.f32 %f341, 0f00000000;
|
||||
add.u64 %rd119, %SP, 16;
|
||||
cvta.to.local.u64 %rd118, %rd119;
|
||||
mov.u32 %r114, 0;
|
||||
fma.rn.f32 %f262, %f26, %f32, %f71;
|
||||
fma.rn.f32 %f263, %f27, %f32, %f72;
|
||||
fma.rn.f32 %f264, %f28, %f32, %f73;
|
||||
add.f32 %f257, %f262, %f39;
|
||||
add.f32 %f258, %f263, %f40;
|
||||
add.f32 %f259, %f264, %f41;
|
||||
st.local.u32 [%rd118+8], %r114;
|
||||
st.local.u32 [%rd118+4], %r114;
|
||||
st.local.u32 [%rd118], %r114;
|
||||
ld.global.u32 %r88, [root];
|
||||
// inline asm
|
||||
call _rt_trace_64, (%r88, %f13, %f14, %f15, %f257, %f258, %f259, %r114, %f341, %f162, %rd119, %r115);
|
||||
// inline asm
|
||||
ld.local.f32 %f91, [%rd118+4];
|
||||
ld.local.f32 %f92, [%rd118];
|
||||
add.f32 %f265, %f92, %f91;
|
||||
ld.local.f32 %f93, [%rd118+8];
|
||||
add.f32 %f266, %f265, %f93;
|
||||
setp.eq.f32 %p19, %f266, 0f00000000;
|
||||
@%p19 bra BB0_18;
|
||||
|
||||
mul.f32 %f267, %f91, %f91;
|
||||
fma.rn.f32 %f268, %f92, %f92, %f267;
|
||||
fma.rn.f32 %f269, %f93, %f93, %f268;
|
||||
sqrt.rn.f32 %f270, %f269;
|
||||
rcp.rn.f32 %f271, %f270;
|
||||
mul.f32 %f272, %f271, %f92;
|
||||
mul.f32 %f273, %f271, %f91;
|
||||
mul.f32 %f274, %f271, %f93;
|
||||
fma.rn.f32 %f275, %f35, %f272, %f13;
|
||||
fma.rn.f32 %f276, %f35, %f273, %f14;
|
||||
fma.rn.f32 %f277, %f35, %f274, %f15;
|
||||
fma.rn.f32 %f94, %f270, %f257, %f275;
|
||||
fma.rn.f32 %f95, %f270, %f258, %f276;
|
||||
fma.rn.f32 %f96, %f270, %f259, %f277;
|
||||
setp.geu.f32 %p20, %f270, %f358;
|
||||
@%p20 bra BB0_18;
|
||||
|
||||
mov.f32 %f355, %f94;
|
||||
mov.f32 %f356, %f95;
|
||||
mov.f32 %f357, %f96;
|
||||
|
||||
BB0_18:
|
||||
ld.global.f32 %f100, [fakeBias];
|
||||
setp.eq.f32 %p21, %f100, 0f00000000;
|
||||
@%p21 bra BB0_20;
|
||||
|
||||
fma.rn.f32 %f355, %f348, %f100, %f355;
|
||||
fma.rn.f32 %f356, %f349, %f100, %f356;
|
||||
fma.rn.f32 %f357, %f350, %f100, %f357;
|
||||
|
||||
BB0_20:
|
||||
mov.u32 %r117, 16;
|
||||
mov.u64 %rd121, 0;
|
||||
mov.u32 %r116, 2;
|
||||
cvt.u64.u32 %rd90, %r3;
|
||||
cvt.u64.u32 %rd89, %r2;
|
||||
mov.u64 %rd93, image_HDR;
|
||||
cvta.global.u64 %rd88, %rd93;
|
||||
// inline asm
|
||||
call (%rd87), _rt_buffer_get_64, (%rd88, %r116, %r117, %rd89, %rd90, %rd121, %rd121);
|
||||
// inline asm
|
||||
mov.f32 %f278, 0f3F800000;
|
||||
st.v4.f32 [%rd87], {%f355, %f356, %f357, %f278};
|
||||
ld.global.u8 %rs10, [imageEnabled];
|
||||
and.b16 %rs11, %rs10, 32;
|
||||
setp.eq.s16 %p22, %rs11, 0;
|
||||
@%p22 bra BB0_28;
|
||||
|
||||
add.f32 %f343, %f126, 0f38D1B717;
|
||||
mov.u32 %r119, 12;
|
||||
mov.u32 %r118, 0;
|
||||
mov.f32 %f342, 0f38D1B717;
|
||||
mul.f32 %f286, %f343, 0f41000000;
|
||||
add.u64 %rd94, %SP, 0;
|
||||
cvta.to.local.u64 %rd5, %rd94;
|
||||
st.local.u32 [%rd5+8], %r118;
|
||||
st.local.u32 [%rd5+4], %r118;
|
||||
st.local.u32 [%rd5], %r118;
|
||||
ld.global.u32 %r93, [root];
|
||||
// inline asm
|
||||
call _rt_trace_64, (%r93, %f13, %f14, %f15, %f158, %f159, %f160, %r118, %f342, %f286, %rd94, %r119);
|
||||
// inline asm
|
||||
ld.local.f32 %f288, [%rd5+4];
|
||||
ld.local.f32 %f289, [%rd5];
|
||||
add.f32 %f290, %f289, %f288;
|
||||
ld.local.f32 %f291, [%rd5+8];
|
||||
add.f32 %f292, %f290, %f291;
|
||||
setp.neu.f32 %p23, %f292, 0f00000000;
|
||||
mov.f32 %f369, 0f437F0000;
|
||||
@%p23 bra BB0_26;
|
||||
|
||||
mov.u32 %r121, 12;
|
||||
mov.u32 %r120, 0;
|
||||
mov.f32 %f344, 0f38D1B717;
|
||||
st.local.u32 [%rd5+8], %r120;
|
||||
st.local.u32 [%rd5+4], %r120;
|
||||
st.local.u32 [%rd5], %r120;
|
||||
ld.global.u32 %r96, [root];
|
||||
// inline asm
|
||||
call _rt_trace_64, (%r96, %f13, %f14, %f15, %f211, %f212, %f213, %r120, %f344, %f286, %rd94, %r121);
|
||||
// inline asm
|
||||
ld.local.f32 %f302, [%rd5+4];
|
||||
ld.local.f32 %f303, [%rd5];
|
||||
add.f32 %f304, %f303, %f302;
|
||||
ld.local.f32 %f305, [%rd5+8];
|
||||
add.f32 %f306, %f304, %f305;
|
||||
setp.neu.f32 %p24, %f306, 0f00000000;
|
||||
@%p24 bra BB0_26;
|
||||
|
||||
mov.u32 %r123, 12;
|
||||
mov.u32 %r122, 0;
|
||||
mov.f32 %f345, 0f38D1B717;
|
||||
st.local.u32 [%rd5+8], %r122;
|
||||
st.local.u32 [%rd5+4], %r122;
|
||||
st.local.u32 [%rd5], %r122;
|
||||
ld.global.u32 %r99, [root];
|
||||
// inline asm
|
||||
call _rt_trace_64, (%r99, %f13, %f14, %f15, %f234, %f235, %f236, %r122, %f345, %f286, %rd94, %r123);
|
||||
// inline asm
|
||||
ld.local.f32 %f316, [%rd5+4];
|
||||
ld.local.f32 %f317, [%rd5];
|
||||
add.f32 %f318, %f317, %f316;
|
||||
ld.local.f32 %f319, [%rd5+8];
|
||||
add.f32 %f320, %f318, %f319;
|
||||
setp.neu.f32 %p25, %f320, 0f00000000;
|
||||
@%p25 bra BB0_26;
|
||||
|
||||
mov.u32 %r125, 12;
|
||||
mov.u32 %r124, 0;
|
||||
mov.f32 %f346, 0f38D1B717;
|
||||
st.local.u32 [%rd5+8], %r124;
|
||||
st.local.u32 [%rd5+4], %r124;
|
||||
st.local.u32 [%rd5], %r124;
|
||||
ld.global.u32 %r102, [root];
|
||||
// inline asm
|
||||
call _rt_trace_64, (%r102, %f13, %f14, %f15, %f257, %f258, %f259, %r124, %f346, %f286, %rd94, %r125);
|
||||
// inline asm
|
||||
ld.local.f32 %f330, [%rd5+4];
|
||||
ld.local.f32 %f331, [%rd5];
|
||||
add.f32 %f332, %f331, %f330;
|
||||
ld.local.f32 %f333, [%rd5+8];
|
||||
add.f32 %f334, %f332, %f333;
|
||||
setp.neu.f32 %p26, %f334, 0f00000000;
|
||||
@%p26 bra BB0_26;
|
||||
|
||||
mov.f32 %f369, 0f00000000;
|
||||
|
||||
BB0_26:
|
||||
mov.u32 %r127, 1;
|
||||
mov.u64 %rd122, 0;
|
||||
mov.u32 %r126, 2;
|
||||
mov.u64 %rd104, image_Mask;
|
||||
cvta.global.u64 %rd99, %rd104;
|
||||
// inline asm
|
||||
call (%rd98), _rt_buffer_get_64, (%rd99, %r126, %r127, %rd89, %rd90, %rd122, %rd122);
|
||||
// inline asm
|
||||
cvt.rzi.u32.f32 %r107, %f369;
|
||||
st.u8 [%rd98], %r107;
|
||||
|
||||
BB0_28:
|
||||
ret;
|
||||
}
|
||||
|
||||
|
||||
8
Assets/Editor/x64/Bakery/fixPos12.ptx.meta
Normal file
8
Assets/Editor/x64/Bakery/fixPos12.ptx.meta
Normal file
@@ -0,0 +1,8 @@
|
||||
fileFormatVersion: 2
|
||||
guid: 8ac7ed1b9a8960c4980b929d79748557
|
||||
timeCreated: 1526320804
|
||||
licenseType: Store
|
||||
DefaultImporter:
|
||||
userData:
|
||||
assetBundleName:
|
||||
assetBundleVariant:
|
||||
382
Assets/Editor/x64/Bakery/fixPos2.ptx
Normal file
382
Assets/Editor/x64/Bakery/fixPos2.ptx
Normal file
@@ -0,0 +1,382 @@
|
||||
//
|
||||
// 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_HDR[1];
|
||||
.global .align 1 .b8 image_HDR2[1];
|
||||
.global .align 1 .b8 uvfacenormal[1];
|
||||
.global .align 1 .b8 uvpos[1];
|
||||
.global .align 1 .b8 uvsmoothpos[1];
|
||||
.global .align 1 .b8 triangleMarks[1];
|
||||
.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 8 .u64 _ZN21rti_internal_register20reg_bitness_detectorE;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail0E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail1E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail2E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail3E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail4E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail5E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail6E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail7E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail8E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail9E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail0E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail1E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail2E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail3E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail4E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail5E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail6E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail7E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail8E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail9E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register14reg_rayIndex_xE;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register14reg_rayIndex_yE;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register14reg_rayIndex_zE;
|
||||
.global .align 8 .b8 _ZN21rti_internal_typename7pixelIDE[6] = {117, 105, 110, 116, 50, 0};
|
||||
.global .align 8 .b8 _ZN21rti_internal_typename10resolutionE[6] = {117, 105, 110, 116, 50, 0};
|
||||
.global .align 8 .b8 _ZN21rti_internal_typename6normalE[7] = {102, 108, 111, 97, 116, 51, 0};
|
||||
.global .align 8 .b8 _ZN21rti_internal_typename6camPosE[7] = {102, 108, 111, 97, 116, 51, 0};
|
||||
.global .align 16 .b8 _ZN21rti_internal_typename4rootE[9] = {114, 116, 79, 98, 106, 101, 99, 116, 0};
|
||||
.global .align 4 .b8 _ZN21rti_internal_typename12imageEnabledE[4] = {105, 110, 116, 0};
|
||||
.global .align 8 .b8 _ZN21rti_internal_typename8tileInfoE[6] = {117, 105, 110, 116, 52, 0};
|
||||
.global .align 4 .b8 _ZN21rti_internal_typename8additiveE[4] = {105, 110, 116, 0};
|
||||
.global .align 4 .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 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 _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];
|
||||
|
||||
.visible .entry _Z6oxMainv(
|
||||
|
||||
)
|
||||
{
|
||||
.local .align 4 .b8 __local_depot0[4];
|
||||
.reg .b64 %SP;
|
||||
.reg .b64 %SPL;
|
||||
.reg .pred %p<14>;
|
||||
.reg .b16 %rs<10>;
|
||||
.reg .f32 %f<160>;
|
||||
.reg .b32 %r<86>;
|
||||
.reg .b64 %rd<89>;
|
||||
|
||||
|
||||
mov.u64 %rd88, __local_depot0;
|
||||
cvta.local.u64 %SP, %rd88;
|
||||
ld.global.v2.u32 {%r6, %r7}, [pixelID];
|
||||
cvt.u64.u32 %rd4, %r6;
|
||||
cvt.u64.u32 %rd5, %r7;
|
||||
mov.u64 %rd8, uvfacenormal;
|
||||
cvta.global.u64 %rd3, %rd8;
|
||||
mov.u32 %r4, 2;
|
||||
mov.u32 %r5, 4;
|
||||
mov.u64 %rd7, 0;
|
||||
// inline asm
|
||||
call (%rd2), _rt_buffer_get_64, (%rd3, %r4, %r5, %rd4, %rd5, %rd7, %rd7);
|
||||
// inline asm
|
||||
ld.u32 %r1, [%rd2];
|
||||
shr.u32 %r10, %r1, 16;
|
||||
cvt.u16.u32 %rs1, %r10;
|
||||
and.b16 %rs3, %rs1, 255;
|
||||
cvt.u16.u32 %rs4, %r1;
|
||||
or.b16 %rs5, %rs4, %rs3;
|
||||
setp.eq.s16 %p1, %rs5, 0;
|
||||
mov.f32 %f150, 0f00000000;
|
||||
mov.f32 %f151, %f150;
|
||||
mov.f32 %f152, %f150;
|
||||
@%p1 bra BB0_2;
|
||||
|
||||
ld.u8 %rs6, [%rd2+1];
|
||||
and.b16 %rs8, %rs4, 255;
|
||||
cvt.rn.f32.u16 %f61, %rs8;
|
||||
div.rn.f32 %f62, %f61, 0f437F0000;
|
||||
fma.rn.f32 %f63, %f62, 0f40000000, 0fBF800000;
|
||||
cvt.rn.f32.u16 %f64, %rs6;
|
||||
div.rn.f32 %f65, %f64, 0f437F0000;
|
||||
fma.rn.f32 %f66, %f65, 0f40000000, 0fBF800000;
|
||||
cvt.rn.f32.u16 %f67, %rs3;
|
||||
div.rn.f32 %f68, %f67, 0f437F0000;
|
||||
fma.rn.f32 %f69, %f68, 0f40000000, 0fBF800000;
|
||||
mul.f32 %f70, %f66, %f66;
|
||||
fma.rn.f32 %f71, %f63, %f63, %f70;
|
||||
fma.rn.f32 %f72, %f69, %f69, %f71;
|
||||
sqrt.rn.f32 %f73, %f72;
|
||||
rcp.rn.f32 %f74, %f73;
|
||||
mul.f32 %f150, %f63, %f74;
|
||||
mul.f32 %f151, %f66, %f74;
|
||||
mul.f32 %f152, %f69, %f74;
|
||||
|
||||
BB0_2:
|
||||
ld.global.v2.u32 {%r11, %r12}, [pixelID];
|
||||
ld.global.v2.u32 {%r14, %r15}, [tileInfo];
|
||||
add.s32 %r2, %r11, %r14;
|
||||
add.s32 %r3, %r12, %r15;
|
||||
setp.eq.f32 %p2, %f151, 0f00000000;
|
||||
setp.eq.f32 %p3, %f150, 0f00000000;
|
||||
and.pred %p4, %p3, %p2;
|
||||
setp.eq.f32 %p5, %f152, 0f00000000;
|
||||
and.pred %p6, %p4, %p5;
|
||||
@%p6 bra BB0_9;
|
||||
bra.uni BB0_3;
|
||||
|
||||
BB0_9:
|
||||
cvt.u64.u32 %rd83, %r2;
|
||||
cvt.u64.u32 %rd84, %r3;
|
||||
mov.u64 %rd87, image_HDR;
|
||||
cvta.global.u64 %rd82, %rd87;
|
||||
mov.u32 %r85, 16;
|
||||
// inline asm
|
||||
call (%rd81), _rt_buffer_get_64, (%rd82, %r4, %r85, %rd83, %rd84, %rd7, %rd7);
|
||||
// inline asm
|
||||
mov.f32 %f149, 0f00000000;
|
||||
st.v4.f32 [%rd81], {%f149, %f149, %f149, %f149};
|
||||
bra.uni BB0_10;
|
||||
|
||||
BB0_3:
|
||||
ld.global.v2.u32 {%r40, %r41}, [pixelID];
|
||||
cvt.u64.u32 %rd11, %r40;
|
||||
cvt.u64.u32 %rd12, %r41;
|
||||
mov.u64 %rd64, uvpos;
|
||||
cvta.global.u64 %rd10, %rd64;
|
||||
mov.u32 %r34, 16;
|
||||
// inline asm
|
||||
call (%rd9), _rt_buffer_get_64, (%rd10, %r4, %r34, %rd11, %rd12, %rd7, %rd7);
|
||||
// inline asm
|
||||
ld.f32 %f83, [%rd9];
|
||||
ld.global.v2.u32 {%r44, %r45}, [pixelID];
|
||||
cvt.u64.u32 %rd17, %r44;
|
||||
cvt.u64.u32 %rd18, %r45;
|
||||
// inline asm
|
||||
call (%rd15), _rt_buffer_get_64, (%rd10, %r4, %r34, %rd17, %rd18, %rd7, %rd7);
|
||||
// inline asm
|
||||
ld.f32 %f84, [%rd15+4];
|
||||
ld.global.v2.u32 {%r48, %r49}, [pixelID];
|
||||
cvt.u64.u32 %rd23, %r48;
|
||||
cvt.u64.u32 %rd24, %r49;
|
||||
// inline asm
|
||||
call (%rd21), _rt_buffer_get_64, (%rd10, %r4, %r34, %rd23, %rd24, %rd7, %rd7);
|
||||
// inline asm
|
||||
ld.f32 %f85, [%rd21+8];
|
||||
ld.global.v2.u32 {%r52, %r53}, [pixelID];
|
||||
cvt.u64.u32 %rd29, %r52;
|
||||
cvt.u64.u32 %rd30, %r53;
|
||||
// inline asm
|
||||
call (%rd27), _rt_buffer_get_64, (%rd10, %r4, %r34, %rd29, %rd30, %rd7, %rd7);
|
||||
// inline asm
|
||||
ld.f32 %f86, [%rd27+12];
|
||||
add.f32 %f82, %f86, 0f38D1B717;
|
||||
ld.global.v2.u32 {%r56, %r57}, [pixelID];
|
||||
cvt.u64.u32 %rd35, %r56;
|
||||
cvt.u64.u32 %rd36, %r57;
|
||||
mov.u64 %rd65, uvsmoothpos;
|
||||
cvta.global.u64 %rd34, %rd65;
|
||||
// inline asm
|
||||
call (%rd33), _rt_buffer_get_64, (%rd34, %r4, %r34, %rd35, %rd36, %rd7, %rd7);
|
||||
// inline asm
|
||||
ld.f32 %f87, [%rd33];
|
||||
ld.global.v2.u32 {%r60, %r61}, [pixelID];
|
||||
cvt.u64.u32 %rd41, %r60;
|
||||
cvt.u64.u32 %rd42, %r61;
|
||||
// inline asm
|
||||
call (%rd39), _rt_buffer_get_64, (%rd34, %r4, %r34, %rd41, %rd42, %rd7, %rd7);
|
||||
// inline asm
|
||||
ld.f32 %f88, [%rd39+4];
|
||||
ld.global.v2.u32 {%r64, %r65}, [pixelID];
|
||||
cvt.u64.u32 %rd47, %r64;
|
||||
cvt.u64.u32 %rd48, %r65;
|
||||
// inline asm
|
||||
call (%rd45), _rt_buffer_get_64, (%rd34, %r4, %r34, %rd47, %rd48, %rd7, %rd7);
|
||||
// inline asm
|
||||
ld.f32 %f89, [%rd45+8];
|
||||
ld.global.v2.u32 {%r68, %r69}, [pixelID];
|
||||
cvt.u64.u32 %rd53, %r68;
|
||||
cvt.u64.u32 %rd54, %r69;
|
||||
// inline asm
|
||||
call (%rd51), _rt_buffer_get_64, (%rd34, %r4, %r34, %rd53, %rd54, %rd7, %rd7);
|
||||
// inline asm
|
||||
ld.f32 %f90, [%rd51+12];
|
||||
cvt.rzi.u32.f32 %r72, %f90;
|
||||
cvt.u64.u32 %rd59, %r72;
|
||||
mov.u64 %rd66, triangleMarks;
|
||||
cvta.global.u64 %rd58, %rd66;
|
||||
mov.u32 %r36, 1;
|
||||
// inline asm
|
||||
call (%rd57), _rt_buffer_get_64, (%rd58, %r36, %r36, %rd59, %rd7, %rd7, %rd7);
|
||||
// inline asm
|
||||
ld.u8 %rs2, [%rd57];
|
||||
setp.eq.s16 %p7, %rs2, 0;
|
||||
selp.f32 %f157, %f87, %f83, %p7;
|
||||
selp.f32 %f158, %f88, %f84, %p7;
|
||||
selp.f32 %f159, %f89, %f85, %p7;
|
||||
fma.rn.f32 %f11, %f150, 0f38D1B717, %f157;
|
||||
fma.rn.f32 %f12, %f151, 0f38D1B717, %f158;
|
||||
fma.rn.f32 %f13, %f152, 0f38D1B717, %f159;
|
||||
abs.f32 %f91, %f152;
|
||||
abs.f32 %f92, %f150;
|
||||
setp.gt.f32 %p8, %f92, %f91;
|
||||
neg.f32 %f93, %f151;
|
||||
selp.f32 %f94, %f93, 0f00000000, %p8;
|
||||
neg.f32 %f95, %f152;
|
||||
selp.f32 %f96, %f150, %f95, %p8;
|
||||
selp.f32 %f97, 0f00000000, %f151, %p8;
|
||||
mul.f32 %f98, %f96, %f96;
|
||||
fma.rn.f32 %f99, %f94, %f94, %f98;
|
||||
fma.rn.f32 %f100, %f97, %f97, %f99;
|
||||
sqrt.rn.f32 %f101, %f100;
|
||||
rcp.rn.f32 %f102, %f101;
|
||||
mul.f32 %f14, %f94, %f102;
|
||||
mul.f32 %f15, %f96, %f102;
|
||||
mul.f32 %f16, %f97, %f102;
|
||||
mul.f32 %f103, %f152, %f15;
|
||||
mul.f32 %f104, %f151, %f16;
|
||||
sub.f32 %f17, %f103, %f104;
|
||||
mul.f32 %f105, %f150, %f16;
|
||||
mul.f32 %f106, %f152, %f14;
|
||||
sub.f32 %f18, %f105, %f106;
|
||||
mul.f32 %f107, %f151, %f14;
|
||||
mul.f32 %f108, %f150, %f15;
|
||||
sub.f32 %f19, %f107, %f108;
|
||||
mov.f32 %f109, 0f3F8147AE;
|
||||
sqrt.rn.f32 %f110, %f109;
|
||||
rcp.rn.f32 %f20, %f110;
|
||||
neg.f32 %f21, %f20;
|
||||
mul.f32 %f22, %f20, 0f00000000;
|
||||
mul.f32 %f111, %f20, 0f3DCCCCCD;
|
||||
mul.f32 %f23, %f14, %f22;
|
||||
mul.f32 %f24, %f15, %f22;
|
||||
mul.f32 %f25, %f16, %f22;
|
||||
fma.rn.f32 %f112, %f17, %f21, %f23;
|
||||
fma.rn.f32 %f113, %f18, %f21, %f24;
|
||||
fma.rn.f32 %f114, %f19, %f21, %f25;
|
||||
mul.f32 %f26, %f150, %f111;
|
||||
mul.f32 %f27, %f151, %f111;
|
||||
mul.f32 %f28, %f152, %f111;
|
||||
add.f32 %f156, %f112, %f26;
|
||||
add.f32 %f155, %f113, %f27;
|
||||
add.f32 %f154, %f114, %f28;
|
||||
add.u64 %rd63, %SP, 0;
|
||||
cvta.to.local.u64 %rd67, %rd63;
|
||||
mov.u32 %r38, 0;
|
||||
st.local.u32 [%rd67], %r38;
|
||||
ld.global.u32 %r37, [root];
|
||||
mov.f32 %f81, 0f38D1B717;
|
||||
// inline asm
|
||||
call _rt_trace_64, (%r37, %f11, %f12, %f13, %f156, %f155, %f154, %r38, %f81, %f82, %rd63, %r5);
|
||||
// inline asm
|
||||
ld.local.f32 %f153, [%rd67];
|
||||
setp.gt.f32 %p9, %f153, 0f00000000;
|
||||
@%p9 bra BB0_7;
|
||||
|
||||
fma.rn.f32 %f123, %f17, %f20, %f23;
|
||||
fma.rn.f32 %f124, %f18, %f20, %f24;
|
||||
fma.rn.f32 %f125, %f19, %f20, %f25;
|
||||
add.f32 %f156, %f123, %f26;
|
||||
add.f32 %f155, %f124, %f27;
|
||||
add.f32 %f154, %f125, %f28;
|
||||
st.local.u32 [%rd67], %r38;
|
||||
ld.global.u32 %r73, [root];
|
||||
// inline asm
|
||||
call _rt_trace_64, (%r73, %f11, %f12, %f13, %f156, %f155, %f154, %r38, %f81, %f82, %rd63, %r5);
|
||||
// inline asm
|
||||
ld.local.f32 %f153, [%rd67];
|
||||
setp.gt.f32 %p10, %f153, 0f00000000;
|
||||
@%p10 bra BB0_7;
|
||||
|
||||
mul.f32 %f37, %f17, %f22;
|
||||
fma.rn.f32 %f134, %f14, %f21, %f37;
|
||||
mul.f32 %f38, %f18, %f22;
|
||||
fma.rn.f32 %f135, %f15, %f21, %f38;
|
||||
mul.f32 %f39, %f19, %f22;
|
||||
fma.rn.f32 %f136, %f16, %f21, %f39;
|
||||
add.f32 %f156, %f134, %f26;
|
||||
add.f32 %f155, %f135, %f27;
|
||||
add.f32 %f154, %f136, %f28;
|
||||
st.local.u32 [%rd67], %r38;
|
||||
ld.global.u32 %r76, [root];
|
||||
// inline asm
|
||||
call _rt_trace_64, (%r76, %f11, %f12, %f13, %f156, %f155, %f154, %r38, %f81, %f82, %rd63, %r5);
|
||||
// inline asm
|
||||
ld.local.f32 %f153, [%rd67];
|
||||
setp.gt.f32 %p11, %f153, 0f00000000;
|
||||
@%p11 bra BB0_7;
|
||||
|
||||
fma.rn.f32 %f145, %f14, %f20, %f37;
|
||||
fma.rn.f32 %f146, %f15, %f20, %f38;
|
||||
fma.rn.f32 %f147, %f16, %f20, %f39;
|
||||
add.f32 %f156, %f145, %f26;
|
||||
add.f32 %f155, %f146, %f27;
|
||||
add.f32 %f154, %f147, %f28;
|
||||
st.local.u32 [%rd67], %r38;
|
||||
ld.global.u32 %r79, [root];
|
||||
// inline asm
|
||||
call _rt_trace_64, (%r79, %f11, %f12, %f13, %f156, %f155, %f154, %r38, %f81, %f82, %rd63, %r5);
|
||||
// inline asm
|
||||
ld.local.f32 %f153, [%rd67];
|
||||
setp.leu.f32 %p12, %f153, 0f00000000;
|
||||
@%p12 bra BB0_8;
|
||||
|
||||
BB0_7:
|
||||
fma.rn.f32 %f157, %f153, %f156, %f11;
|
||||
fma.rn.f32 %f158, %f153, %f155, %f12;
|
||||
fma.rn.f32 %f159, %f153, %f154, %f13;
|
||||
|
||||
BB0_8:
|
||||
cvt.u64.u32 %rd77, %r3;
|
||||
cvt.u64.u32 %rd76, %r2;
|
||||
mov.u64 %rd80, image_HDR;
|
||||
cvta.global.u64 %rd75, %rd80;
|
||||
// inline asm
|
||||
call (%rd74), _rt_buffer_get_64, (%rd75, %r4, %r34, %rd76, %rd77, %rd7, %rd7);
|
||||
// inline asm
|
||||
selp.f32 %f148, 0f3F000000, 0f3F800000, %p7;
|
||||
st.v4.f32 [%rd74], {%f157, %f158, %f159, %f148};
|
||||
|
||||
BB0_10:
|
||||
ret;
|
||||
}
|
||||
|
||||
|
||||
8
Assets/Editor/x64/Bakery/fixPos2.ptx.meta
Normal file
8
Assets/Editor/x64/Bakery/fixPos2.ptx.meta
Normal file
@@ -0,0 +1,8 @@
|
||||
fileFormatVersion: 2
|
||||
guid: 81c36c1f006bd8848b64a84a623af92d
|
||||
timeCreated: 1526161016
|
||||
licenseType: Store
|
||||
DefaultImporter:
|
||||
userData:
|
||||
assetBundleName:
|
||||
assetBundleVariant:
|
||||
513
Assets/Editor/x64/Bakery/fixPos3D.ptx
Normal file
513
Assets/Editor/x64/Bakery/fixPos3D.ptx
Normal file
@@ -0,0 +1,513 @@
|
||||
//
|
||||
// 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_HDR[1];
|
||||
.global .align 1 .b8 image_HDR2[1];
|
||||
.global .align 1 .b8 uvnormal[1];
|
||||
.global .align 1 .b8 uvpos[1];
|
||||
.global .align 4 .b8 voxelSize[12];
|
||||
.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_typeinfo9voxelSizeE[8] = {82, 97, 121, 0, 12, 0, 0, 0};
|
||||
.global .align 8 .u64 _ZN21rti_internal_register20reg_bitness_detectorE;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail0E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail1E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail2E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail3E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail4E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail5E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail6E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail7E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail8E;
|
||||
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail9E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail0E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail1E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail2E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail3E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail4E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail5E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail6E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail7E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail8E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail9E;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register14reg_rayIndex_xE;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register14reg_rayIndex_yE;
|
||||
.global .align 4 .u32 _ZN21rti_internal_register14reg_rayIndex_zE;
|
||||
.global .align 8 .b8 _ZN21rti_internal_typename7pixelIDE[6] = {117, 105, 110, 116, 50, 0};
|
||||
.global .align 8 .b8 _ZN21rti_internal_typename10resolutionE[6] = {117, 105, 110, 116, 50, 0};
|
||||
.global .align 8 .b8 _ZN21rti_internal_typename6normalE[7] = {102, 108, 111, 97, 116, 51, 0};
|
||||
.global .align 8 .b8 _ZN21rti_internal_typename6camPosE[7] = {102, 108, 111, 97, 116, 51, 0};
|
||||
.global .align 16 .b8 _ZN21rti_internal_typename4rootE[9] = {114, 116, 79, 98, 106, 101, 99, 116, 0};
|
||||
.global .align 4 .b8 _ZN21rti_internal_typename12imageEnabledE[4] = {105, 110, 116, 0};
|
||||
.global .align 8 .b8 _ZN21rti_internal_typename8tileInfoE[6] = {117, 105, 110, 116, 52, 0};
|
||||
.global .align 4 .b8 _ZN21rti_internal_typename8additiveE[4] = {105, 110, 116, 0};
|
||||
.global .align 8 .b8 _ZN21rti_internal_typename9voxelSizeE[7] = {102, 108, 111, 97, 116, 51, 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_typeenum9voxelSizeE = 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_semantic9voxelSizeE[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_annotation9voxelSizeE[1];
|
||||
|
||||
.visible .entry _Z6oxMainv(
|
||||
|
||||
)
|
||||
{
|
||||
.local .align 4 .b8 __local_depot0[12];
|
||||
.reg .b64 %SP;
|
||||
.reg .b64 %SPL;
|
||||
.reg .pred %p<28>;
|
||||
.reg .b16 %rs<9>;
|
||||
.reg .f32 %f<277>;
|
||||
.reg .b32 %r<59>;
|
||||
.reg .b64 %rd<55>;
|
||||
|
||||
|
||||
mov.u64 %rd54, __local_depot0;
|
||||
cvta.local.u64 %SP, %rd54;
|
||||
ld.global.v2.u32 {%r6, %r7}, [pixelID];
|
||||
cvt.u64.u32 %rd4, %r6;
|
||||
cvt.u64.u32 %rd5, %r7;
|
||||
mov.u64 %rd8, uvnormal;
|
||||
cvta.global.u64 %rd3, %rd8;
|
||||
mov.u32 %r4, 2;
|
||||
mov.u32 %r5, 4;
|
||||
mov.u64 %rd7, 0;
|
||||
// inline asm
|
||||
call (%rd2), _rt_buffer_get_64, (%rd3, %r4, %r5, %rd4, %rd5, %rd7, %rd7);
|
||||
// inline asm
|
||||
ld.u32 %r1, [%rd2];
|
||||
shr.u32 %r10, %r1, 16;
|
||||
cvt.u16.u32 %rs1, %r10;
|
||||
and.b16 %rs2, %rs1, 255;
|
||||
cvt.u16.u32 %rs3, %r1;
|
||||
or.b16 %rs4, %rs3, %rs2;
|
||||
setp.eq.s16 %p10, %rs4, 0;
|
||||
mov.pred %p25, -1;
|
||||
mov.pred %p26, %p25;
|
||||
mov.pred %p27, %p25;
|
||||
@%p10 bra BB0_2;
|
||||
|
||||
ld.u8 %rs5, [%rd2+1];
|
||||
and.b16 %rs7, %rs3, 255;
|
||||
cvt.rn.f32.u16 %f71, %rs7;
|
||||
div.rn.f32 %f72, %f71, 0f437F0000;
|
||||
fma.rn.f32 %f73, %f72, 0f40000000, 0fBF800000;
|
||||
cvt.rn.f32.u16 %f74, %rs5;
|
||||
div.rn.f32 %f75, %f74, 0f437F0000;
|
||||
fma.rn.f32 %f76, %f75, 0f40000000, 0fBF800000;
|
||||
cvt.rn.f32.u16 %f77, %rs2;
|
||||
div.rn.f32 %f78, %f77, 0f437F0000;
|
||||
fma.rn.f32 %f79, %f78, 0f40000000, 0fBF800000;
|
||||
mul.f32 %f80, %f76, %f76;
|
||||
fma.rn.f32 %f81, %f73, %f73, %f80;
|
||||
fma.rn.f32 %f82, %f79, %f79, %f81;
|
||||
sqrt.rn.f32 %f83, %f82;
|
||||
rcp.rn.f32 %f84, %f83;
|
||||
mul.f32 %f85, %f73, %f84;
|
||||
mul.f32 %f86, %f76, %f84;
|
||||
mul.f32 %f87, %f79, %f84;
|
||||
setp.eq.f32 %p25, %f85, 0f00000000;
|
||||
setp.eq.f32 %p26, %f86, 0f00000000;
|
||||
setp.eq.f32 %p27, %f87, 0f00000000;
|
||||
|
||||
BB0_2:
|
||||
ld.global.v2.u32 {%r11, %r12}, [pixelID];
|
||||
ld.global.v2.u32 {%r14, %r15}, [tileInfo];
|
||||
add.s32 %r2, %r11, %r14;
|
||||
add.s32 %r3, %r12, %r15;
|
||||
and.pred %p11, %p25, %p26;
|
||||
and.pred %p12, %p11, %p27;
|
||||
@%p12 bra BB0_23;
|
||||
bra.uni BB0_3;
|
||||
|
||||
BB0_23:
|
||||
cvt.u64.u32 %rd49, %r2;
|
||||
cvt.u64.u32 %rd50, %r3;
|
||||
mov.u64 %rd53, image_HDR;
|
||||
cvta.global.u64 %rd48, %rd53;
|
||||
mov.u32 %r58, 16;
|
||||
// inline asm
|
||||
call (%rd47), _rt_buffer_get_64, (%rd48, %r4, %r58, %rd49, %rd50, %rd7, %rd7);
|
||||
// inline asm
|
||||
mov.f32 %f253, 0f00000000;
|
||||
st.v4.f32 [%rd47], {%f253, %f253, %f253, %f253};
|
||||
bra.uni BB0_24;
|
||||
|
||||
BB0_3:
|
||||
ld.global.v2.u32 {%r28, %r29}, [pixelID];
|
||||
cvt.u64.u32 %rd11, %r28;
|
||||
cvt.u64.u32 %rd12, %r29;
|
||||
mov.u64 %rd28, uvpos;
|
||||
cvta.global.u64 %rd10, %rd28;
|
||||
mov.u32 %r27, 12;
|
||||
// inline asm
|
||||
call (%rd9), _rt_buffer_get_64, (%rd10, %r4, %r27, %rd11, %rd12, %rd7, %rd7);
|
||||
// inline asm
|
||||
ld.f32 %f1, [%rd9];
|
||||
ld.global.v2.u32 {%r32, %r33}, [pixelID];
|
||||
cvt.u64.u32 %rd17, %r32;
|
||||
cvt.u64.u32 %rd18, %r33;
|
||||
// inline asm
|
||||
call (%rd15), _rt_buffer_get_64, (%rd10, %r4, %r27, %rd17, %rd18, %rd7, %rd7);
|
||||
// inline asm
|
||||
ld.f32 %f2, [%rd15+4];
|
||||
ld.global.v2.u32 {%r36, %r37}, [pixelID];
|
||||
cvt.u64.u32 %rd23, %r36;
|
||||
cvt.u64.u32 %rd24, %r37;
|
||||
// inline asm
|
||||
call (%rd21), _rt_buffer_get_64, (%rd10, %r4, %r27, %rd23, %rd24, %rd7, %rd7);
|
||||
// inline asm
|
||||
mul.f32 %f97, %f1, 0f3456BF95;
|
||||
mul.f32 %f98, %f2, 0f3456BF95;
|
||||
ld.f32 %f3, [%rd21+8];
|
||||
mul.f32 %f99, %f3, 0f3456BF95;
|
||||
abs.f32 %f100, %f97;
|
||||
abs.f32 %f101, %f98;
|
||||
abs.f32 %f102, %f99;
|
||||
max.f32 %f103, %f100, %f101;
|
||||
max.f32 %f104, %f103, %f102;
|
||||
mov.f32 %f105, 0f38D1B717;
|
||||
max.f32 %f4, %f104, %f105;
|
||||
ld.global.f32 %f106, [voxelSize];
|
||||
ld.global.f32 %f107, [voxelSize+4];
|
||||
mul.f32 %f108, %f107, 0f00000000;
|
||||
sub.f32 %f109, %f108, %f106;
|
||||
ld.global.f32 %f110, [voxelSize+8];
|
||||
fma.rn.f32 %f111, %f110, 0f00000000, %f109;
|
||||
abs.f32 %f95, %f111;
|
||||
add.u64 %rd27, %SP, 0;
|
||||
cvta.to.local.u64 %rd29, %rd27;
|
||||
mov.u32 %r26, 0;
|
||||
st.local.u32 [%rd29+8], %r26;
|
||||
st.local.u32 [%rd29+4], %r26;
|
||||
st.local.u32 [%rd29], %r26;
|
||||
ld.global.u32 %r25, [root];
|
||||
mov.f32 %f91, 0fBF800000;
|
||||
mov.f32 %f94, 0f00000000;
|
||||
// inline asm
|
||||
call _rt_trace_64, (%r25, %f1, %f2, %f3, %f91, %f94, %f94, %r26, %f94, %f95, %rd27, %r27);
|
||||
// inline asm
|
||||
ld.local.f32 %f5, [%rd29+4];
|
||||
ld.local.f32 %f6, [%rd29];
|
||||
add.f32 %f112, %f6, %f5;
|
||||
ld.local.f32 %f7, [%rd29+8];
|
||||
add.f32 %f113, %f112, %f7;
|
||||
mov.f32 %f261, 0f47C34F80;
|
||||
setp.eq.f32 %p13, %f113, 0f00000000;
|
||||
@%p13 bra BB0_4;
|
||||
|
||||
mul.f32 %f115, %f5, %f5;
|
||||
fma.rn.f32 %f116, %f6, %f6, %f115;
|
||||
fma.rn.f32 %f117, %f7, %f7, %f116;
|
||||
sqrt.rn.f32 %f8, %f117;
|
||||
rcp.rn.f32 %f118, %f8;
|
||||
mul.f32 %f119, %f118, %f6;
|
||||
mul.f32 %f120, %f118, %f5;
|
||||
mul.f32 %f121, %f118, %f7;
|
||||
fma.rn.f32 %f122, %f4, %f119, %f1;
|
||||
fma.rn.f32 %f123, %f4, %f120, %f2;
|
||||
fma.rn.f32 %f124, %f4, %f121, %f3;
|
||||
sub.f32 %f9, %f122, %f8;
|
||||
fma.rn.f32 %f10, %f8, 0f00000000, %f123;
|
||||
fma.rn.f32 %f11, %f8, 0f00000000, %f124;
|
||||
setp.geu.f32 %p14, %f8, 0f47C34F80;
|
||||
mov.f32 %f258, %f1;
|
||||
mov.f32 %f259, %f2;
|
||||
mov.f32 %f260, %f3;
|
||||
@%p14 bra BB0_7;
|
||||
|
||||
mov.f32 %f258, %f9;
|
||||
mov.f32 %f259, %f10;
|
||||
mov.f32 %f260, %f11;
|
||||
mov.f32 %f261, %f8;
|
||||
bra.uni BB0_7;
|
||||
|
||||
BB0_4:
|
||||
mov.f32 %f258, %f1;
|
||||
mov.f32 %f259, %f2;
|
||||
mov.f32 %f260, %f3;
|
||||
|
||||
BB0_7:
|
||||
ld.global.f32 %f133, [voxelSize+4];
|
||||
ld.global.f32 %f134, [voxelSize];
|
||||
fma.rn.f32 %f135, %f133, 0f00000000, %f134;
|
||||
ld.global.f32 %f136, [voxelSize+8];
|
||||
fma.rn.f32 %f137, %f136, 0f00000000, %f135;
|
||||
abs.f32 %f132, %f137;
|
||||
st.local.u32 [%rd29+8], %r26;
|
||||
st.local.u32 [%rd29+4], %r26;
|
||||
st.local.u32 [%rd29], %r26;
|
||||
ld.global.u32 %r40, [root];
|
||||
mov.f32 %f128, 0f3F800000;
|
||||
// inline asm
|
||||
call _rt_trace_64, (%r40, %f1, %f2, %f3, %f128, %f94, %f94, %r26, %f94, %f132, %rd27, %r27);
|
||||
// inline asm
|
||||
ld.local.f32 %f16, [%rd29+4];
|
||||
ld.local.f32 %f17, [%rd29];
|
||||
add.f32 %f138, %f17, %f16;
|
||||
ld.local.f32 %f18, [%rd29+8];
|
||||
add.f32 %f139, %f138, %f18;
|
||||
setp.eq.f32 %p15, %f139, 0f00000000;
|
||||
@%p15 bra BB0_10;
|
||||
|
||||
mul.f32 %f140, %f16, %f16;
|
||||
fma.rn.f32 %f141, %f17, %f17, %f140;
|
||||
fma.rn.f32 %f142, %f18, %f18, %f141;
|
||||
sqrt.rn.f32 %f19, %f142;
|
||||
rcp.rn.f32 %f143, %f19;
|
||||
mul.f32 %f144, %f143, %f17;
|
||||
mul.f32 %f145, %f143, %f16;
|
||||
mul.f32 %f146, %f143, %f18;
|
||||
fma.rn.f32 %f20, %f4, %f144, %f1;
|
||||
fma.rn.f32 %f147, %f4, %f145, %f2;
|
||||
fma.rn.f32 %f148, %f4, %f146, %f3;
|
||||
fma.rn.f32 %f21, %f19, 0f00000000, %f147;
|
||||
fma.rn.f32 %f22, %f19, 0f00000000, %f148;
|
||||
setp.geu.f32 %p16, %f19, %f261;
|
||||
@%p16 bra BB0_10;
|
||||
|
||||
add.f32 %f258, %f20, %f19;
|
||||
mov.f32 %f259, %f21;
|
||||
mov.f32 %f260, %f22;
|
||||
mov.f32 %f261, %f19;
|
||||
|
||||
BB0_10:
|
||||
ld.global.f32 %f157, [voxelSize];
|
||||
ld.global.f32 %f158, [voxelSize+4];
|
||||
neg.f32 %f159, %f158;
|
||||
fma.rn.f32 %f160, %f157, 0f00000000, %f159;
|
||||
ld.global.f32 %f161, [voxelSize+8];
|
||||
fma.rn.f32 %f162, %f161, 0f00000000, %f160;
|
||||
abs.f32 %f156, %f162;
|
||||
st.local.u32 [%rd29+8], %r26;
|
||||
st.local.u32 [%rd29+4], %r26;
|
||||
st.local.u32 [%rd29], %r26;
|
||||
ld.global.u32 %r43, [root];
|
||||
// inline asm
|
||||
call _rt_trace_64, (%r43, %f1, %f2, %f3, %f94, %f91, %f94, %r26, %f94, %f156, %rd27, %r27);
|
||||
// inline asm
|
||||
ld.local.f32 %f28, [%rd29+4];
|
||||
ld.local.f32 %f29, [%rd29];
|
||||
add.f32 %f163, %f29, %f28;
|
||||
ld.local.f32 %f30, [%rd29+8];
|
||||
add.f32 %f164, %f163, %f30;
|
||||
setp.eq.f32 %p17, %f164, 0f00000000;
|
||||
@%p17 bra BB0_13;
|
||||
|
||||
mul.f32 %f165, %f28, %f28;
|
||||
fma.rn.f32 %f166, %f29, %f29, %f165;
|
||||
fma.rn.f32 %f167, %f30, %f30, %f166;
|
||||
sqrt.rn.f32 %f31, %f167;
|
||||
rcp.rn.f32 %f168, %f31;
|
||||
mul.f32 %f169, %f168, %f29;
|
||||
mul.f32 %f170, %f168, %f28;
|
||||
mul.f32 %f171, %f168, %f30;
|
||||
fma.rn.f32 %f172, %f4, %f169, %f1;
|
||||
fma.rn.f32 %f173, %f4, %f170, %f2;
|
||||
fma.rn.f32 %f174, %f4, %f171, %f3;
|
||||
fma.rn.f32 %f32, %f31, 0f00000000, %f172;
|
||||
sub.f32 %f33, %f173, %f31;
|
||||
fma.rn.f32 %f34, %f31, 0f00000000, %f174;
|
||||
setp.geu.f32 %p18, %f31, %f261;
|
||||
@%p18 bra BB0_13;
|
||||
|
||||
mov.f32 %f258, %f32;
|
||||
mov.f32 %f259, %f33;
|
||||
mov.f32 %f260, %f34;
|
||||
mov.f32 %f261, %f31;
|
||||
|
||||
BB0_13:
|
||||
ld.global.f32 %f183, [voxelSize];
|
||||
ld.global.f32 %f184, [voxelSize+4];
|
||||
fma.rn.f32 %f185, %f183, 0f00000000, %f184;
|
||||
ld.global.f32 %f186, [voxelSize+8];
|
||||
fma.rn.f32 %f187, %f186, 0f00000000, %f185;
|
||||
abs.f32 %f182, %f187;
|
||||
st.local.u32 [%rd29+8], %r26;
|
||||
st.local.u32 [%rd29+4], %r26;
|
||||
st.local.u32 [%rd29], %r26;
|
||||
ld.global.u32 %r46, [root];
|
||||
// inline asm
|
||||
call _rt_trace_64, (%r46, %f1, %f2, %f3, %f94, %f128, %f94, %r26, %f94, %f182, %rd27, %r27);
|
||||
// inline asm
|
||||
ld.local.f32 %f39, [%rd29+4];
|
||||
ld.local.f32 %f40, [%rd29];
|
||||
add.f32 %f188, %f40, %f39;
|
||||
ld.local.f32 %f41, [%rd29+8];
|
||||
add.f32 %f189, %f188, %f41;
|
||||
setp.eq.f32 %p19, %f189, 0f00000000;
|
||||
@%p19 bra BB0_16;
|
||||
|
||||
mul.f32 %f190, %f39, %f39;
|
||||
fma.rn.f32 %f191, %f40, %f40, %f190;
|
||||
fma.rn.f32 %f192, %f41, %f41, %f191;
|
||||
sqrt.rn.f32 %f42, %f192;
|
||||
rcp.rn.f32 %f193, %f42;
|
||||
mul.f32 %f194, %f193, %f40;
|
||||
mul.f32 %f195, %f193, %f39;
|
||||
mul.f32 %f196, %f193, %f41;
|
||||
fma.rn.f32 %f197, %f4, %f194, %f1;
|
||||
fma.rn.f32 %f43, %f4, %f195, %f2;
|
||||
fma.rn.f32 %f198, %f4, %f196, %f3;
|
||||
fma.rn.f32 %f44, %f42, 0f00000000, %f197;
|
||||
fma.rn.f32 %f45, %f42, 0f00000000, %f198;
|
||||
setp.geu.f32 %p20, %f42, %f261;
|
||||
@%p20 bra BB0_16;
|
||||
|
||||
add.f32 %f259, %f43, %f42;
|
||||
mov.f32 %f258, %f44;
|
||||
mov.f32 %f260, %f45;
|
||||
mov.f32 %f261, %f42;
|
||||
|
||||
BB0_16:
|
||||
ld.global.f32 %f207, [voxelSize];
|
||||
ld.global.f32 %f208, [voxelSize+4];
|
||||
mul.f32 %f209, %f208, 0f00000000;
|
||||
fma.rn.f32 %f210, %f207, 0f00000000, %f209;
|
||||
ld.global.f32 %f211, [voxelSize+8];
|
||||
sub.f32 %f212, %f210, %f211;
|
||||
abs.f32 %f206, %f212;
|
||||
st.local.u32 [%rd29+8], %r26;
|
||||
st.local.u32 [%rd29+4], %r26;
|
||||
st.local.u32 [%rd29], %r26;
|
||||
ld.global.u32 %r49, [root];
|
||||
// inline asm
|
||||
call _rt_trace_64, (%r49, %f1, %f2, %f3, %f94, %f94, %f91, %r26, %f94, %f206, %rd27, %r27);
|
||||
// inline asm
|
||||
ld.local.f32 %f51, [%rd29+4];
|
||||
ld.local.f32 %f52, [%rd29];
|
||||
add.f32 %f213, %f52, %f51;
|
||||
ld.local.f32 %f53, [%rd29+8];
|
||||
add.f32 %f214, %f213, %f53;
|
||||
setp.eq.f32 %p21, %f214, 0f00000000;
|
||||
@%p21 bra BB0_19;
|
||||
|
||||
mul.f32 %f215, %f51, %f51;
|
||||
fma.rn.f32 %f216, %f52, %f52, %f215;
|
||||
fma.rn.f32 %f217, %f53, %f53, %f216;
|
||||
sqrt.rn.f32 %f54, %f217;
|
||||
rcp.rn.f32 %f218, %f54;
|
||||
mul.f32 %f219, %f218, %f52;
|
||||
mul.f32 %f220, %f218, %f51;
|
||||
mul.f32 %f221, %f218, %f53;
|
||||
fma.rn.f32 %f222, %f4, %f219, %f1;
|
||||
fma.rn.f32 %f223, %f4, %f220, %f2;
|
||||
fma.rn.f32 %f224, %f4, %f221, %f3;
|
||||
fma.rn.f32 %f55, %f54, 0f00000000, %f222;
|
||||
fma.rn.f32 %f56, %f54, 0f00000000, %f223;
|
||||
sub.f32 %f57, %f224, %f54;
|
||||
setp.geu.f32 %p22, %f54, %f261;
|
||||
@%p22 bra BB0_19;
|
||||
|
||||
mov.f32 %f258, %f55;
|
||||
mov.f32 %f259, %f56;
|
||||
mov.f32 %f260, %f57;
|
||||
mov.f32 %f261, %f54;
|
||||
|
||||
BB0_19:
|
||||
ld.global.f32 %f233, [voxelSize];
|
||||
ld.global.f32 %f234, [voxelSize+4];
|
||||
mul.f32 %f235, %f234, 0f00000000;
|
||||
fma.rn.f32 %f236, %f233, 0f00000000, %f235;
|
||||
ld.global.f32 %f237, [voxelSize+8];
|
||||
add.f32 %f238, %f236, %f237;
|
||||
abs.f32 %f232, %f238;
|
||||
st.local.u32 [%rd29+8], %r26;
|
||||
st.local.u32 [%rd29+4], %r26;
|
||||
st.local.u32 [%rd29], %r26;
|
||||
ld.global.u32 %r52, [root];
|
||||
// inline asm
|
||||
call _rt_trace_64, (%r52, %f1, %f2, %f3, %f94, %f94, %f128, %r26, %f94, %f232, %rd27, %r27);
|
||||
// inline asm
|
||||
ld.local.f32 %f62, [%rd29+4];
|
||||
ld.local.f32 %f63, [%rd29];
|
||||
add.f32 %f239, %f63, %f62;
|
||||
ld.local.f32 %f64, [%rd29+8];
|
||||
add.f32 %f240, %f239, %f64;
|
||||
setp.eq.f32 %p23, %f240, 0f00000000;
|
||||
@%p23 bra BB0_22;
|
||||
|
||||
mul.f32 %f241, %f62, %f62;
|
||||
fma.rn.f32 %f242, %f63, %f63, %f241;
|
||||
fma.rn.f32 %f243, %f64, %f64, %f242;
|
||||
sqrt.rn.f32 %f244, %f243;
|
||||
rcp.rn.f32 %f245, %f244;
|
||||
mul.f32 %f246, %f245, %f63;
|
||||
mul.f32 %f247, %f245, %f62;
|
||||
mul.f32 %f248, %f245, %f64;
|
||||
fma.rn.f32 %f249, %f4, %f246, %f1;
|
||||
fma.rn.f32 %f250, %f4, %f247, %f2;
|
||||
fma.rn.f32 %f251, %f4, %f248, %f3;
|
||||
fma.rn.f32 %f65, %f244, 0f00000000, %f249;
|
||||
fma.rn.f32 %f66, %f244, 0f00000000, %f250;
|
||||
add.f32 %f67, %f251, %f244;
|
||||
setp.geu.f32 %p24, %f244, %f261;
|
||||
@%p24 bra BB0_22;
|
||||
|
||||
mov.f32 %f258, %f65;
|
||||
mov.f32 %f259, %f66;
|
||||
mov.f32 %f260, %f67;
|
||||
|
||||
BB0_22:
|
||||
cvt.u64.u32 %rd43, %r3;
|
||||
cvt.u64.u32 %rd42, %r2;
|
||||
mov.u64 %rd46, image_HDR;
|
||||
cvta.global.u64 %rd41, %rd46;
|
||||
mov.u32 %r56, 16;
|
||||
// inline asm
|
||||
call (%rd40), _rt_buffer_get_64, (%rd41, %r4, %r56, %rd42, %rd43, %rd7, %rd7);
|
||||
// inline asm
|
||||
st.v4.f32 [%rd40], {%f258, %f259, %f260, %f128};
|
||||
|
||||
BB0_24:
|
||||
ret;
|
||||
}
|
||||
|
||||
|
||||
8
Assets/Editor/x64/Bakery/fixPos3D.ptx.meta
Normal file
8
Assets/Editor/x64/Bakery/fixPos3D.ptx.meta
Normal file
@@ -0,0 +1,8 @@
|
||||
fileFormatVersion: 2
|
||||
guid: 6d3c23da2dc42294fa8ca290df4f84c2
|
||||
timeCreated: 1589639893
|
||||
licenseType: Store
|
||||
DefaultImporter:
|
||||
userData:
|
||||
assetBundleName:
|
||||
assetBundleVariant:
|
||||
Some files were not shown because too many files have changed in this diff Show More
Reference in New Issue
Block a user