	.version 2.2
	.target sm_20
	// compiled with ../../../External/3rdParty/NVIDIA/CUDA/win/bin/../open64/lib//be.exe
	// nvopencc 3.2 built on 2010-11-04

	.visible .func (.param .s32 __cudaretf__Z15IntegerMultiplyii) _Z15IntegerMultiplyii (.param .s32 __cudaparmf1__Z15IntegerMultiplyii, .param .s32 __cudaparmf2__Z15IntegerMultiplyii)

	.visible .func (.param .s32 __cudaretf__Z17Standard2DKernelXv) _Z17Standard2DKernelXv ()

	.visible .func (.param .s32 __cudaretf__Z17Standard2DKernelYv) _Z17Standard2DKernelYv ()

	.visible .func (.param .align 16 .b8 __cudaretf__Z13Half4ToFloat47ushort4[16]) _Z13Half4ToFloat47ushort4 (.param .align 8 .b8 __cudaparmf1__Z13Half4ToFloat47ushort4[8])

	.visible .func (.param .align 8 .b8 __cudaretf__Z13Float4ToHalf46float4[8]) _Z13Float4ToHalf46float4 (.param .align 16 .b8 __cudaparmf1__Z13Float4ToHalf46float4[16])

	.visible .func (.param .u32 __cudaretf__Z4Mix3RjS_S_) _Z4Mix3RjS_S_ (.param .u64 __cudaparmf1__Z4Mix3RjS_S_, .param .u64 __cudaparmf2__Z4Mix3RjS_S_, .param .u64 __cudaparmf3__Z4Mix3RjS_S_)

	.visible .func (.param .s32 __cudaretf__Z4Randj) _Z4Randj (.param .u32 __cudaparmf1__Z4Randj)

	.visible .func (.param .s32 __cudaretf__Z6Rand2Djjj) _Z6Rand2Djjj (.param .u32 __cudaparmf1__Z6Rand2Djjj, .param .u32 __cudaparmf2__Z6Rand2Djjj, .param .u32 __cudaparmf3__Z6Rand2Djjj)

	.visible .func (.param .s32 __cudaretf__Z6Rand2Dj) _Z6Rand2Dj (.param .u32 __cudaparmf1__Z6Rand2Dj)

	.visible .func (.param .align 8 .b8 __cudaretf__Z6Read2DI7ushort4ET_PKS1_iii[8]) _Z6Read2DI7ushort4ET_PKS1_iii (.param .u64 __cudaparmf1__Z6Read2DI7ushort4ET_PKS1_iii, .param .s32 __cudaparmf2__Z6Read2DI7ushort4ET_PKS1_iii, .param .s32 __cudaparmf3__Z6Read2DI7ushort4ET_PKS1_iii, .param .s32 __cudaparmf4__Z6Read2DI7ushort4ET_PKS1_iii)

	.visible .func (.param .align 16 .b8 __cudaretf__Z6Read2DI6float4ET_PKS1_iii[16]) _Z6Read2DI6float4ET_PKS1_iii (.param .u64 __cudaparmf1__Z6Read2DI6float4ET_PKS1_iii, .param .s32 __cudaparmf2__Z6Read2DI6float4ET_PKS1_iii, .param .s32 __cudaparmf3__Z6Read2DI6float4ET_PKS1_iii, .param .s32 __cudaparmf4__Z6Read2DI6float4ET_PKS1_iii)

	.visible .func _Z7Write2DI7ushort4EvT_PS1_iii (.param .align 8 .b8 __cudaparmf1__Z7Write2DI7ushort4EvT_PS1_iii[8], .param .u64 __cudaparmf2__Z7Write2DI7ushort4EvT_PS1_iii, .param .s32 __cudaparmf3__Z7Write2DI7ushort4EvT_PS1_iii, .param .s32 __cudaparmf4__Z7Write2DI7ushort4EvT_PS1_iii, .param .s32 __cudaparmf5__Z7Write2DI7ushort4EvT_PS1_iii)

	.visible .func _Z7Write2DI6float4EvT_PS1_iii (.param .align 16 .b8 __cudaparmf1__Z7Write2DI6float4EvT_PS1_iii[16], .param .u64 __cudaparmf2__Z7Write2DI6float4EvT_PS1_iii, .param .s32 __cudaparmf3__Z7Write2DI6float4EvT_PS1_iii, .param .s32 __cudaparmf4__Z7Write2DI6float4EvT_PS1_iii, .param .s32 __cudaparmf5__Z7Write2DI6float4EvT_PS1_iii)

	.visible .func (.param .align 16 .b8 __cudaretf__Z18UnpremultiplyPixel8PixelRGB[16]) _Z18UnpremultiplyPixel8PixelRGB (.param .align 16 .b8 __cudaparmf1__Z18UnpremultiplyPixel8PixelRGB[16])

	.visible .func (.param .f32 __cudaretf__Z13ToLinearColorf) _Z13ToLinearColorf (.param .f32 __cudaparmf1__Z13ToLinearColorf)

	.visible .func (.param .f32 __cudaretf__Z15FromLinearColorf) _Z15FromLinearColorf (.param .f32 __cudaparmf1__Z15FromLinearColorf)

	.visible .func (.param .align 16 .b8 __cudaretf__Z25PremultiplyLinearizePixel8PixelRGB[16]) _Z25PremultiplyLinearizePixel8PixelRGB (.param .align 16 .b8 __cudaparmf1__Z25PremultiplyLinearizePixel8PixelRGB[16])

	.visible .func (.param .align 16 .b8 __cudaretf__Z29UnpremultiplyUnlinearizePixel8PixelRGB[16]) _Z29UnpremultiplyUnlinearizePixel8PixelRGB (.param .align 16 .b8 __cudaparmf1__Z29UnpremultiplyUnlinearizePixel8PixelRGB[16])

	.visible .func (.param .align 16 .b8 __cudaretf__Z20PremultiplyLinearize6float4[16]) _Z20PremultiplyLinearize6float4 (.param .align 16 .b8 __cudaparmf1__Z20PremultiplyLinearize6float4[16])

	.visible .func (.param .align 16 .b8 __cudaretf__Z24UnpremultiplyUnlinearize6float4[16]) _Z24UnpremultiplyUnlinearize6float4 (.param .align 16 .b8 __cudaparmf1__Z24UnpremultiplyUnlinearize6float4[16])

	.visible .func (.param .align 16 .b8 __cudaretf__Z18SwapComponentOrderI6float4ET_RKS1_[16]) _Z18SwapComponentOrderI6float4ET_RKS1_ (.param .u64 __cudaparmf1__Z18SwapComponentOrderI6float4ET_RKS1_)

	.visible .func _Z14ConvertUVToHueffRf (.param .f32 __cudaparmf1__Z14ConvertUVToHueffRf, .param .f32 __cudaparmf2__Z14ConvertUVToHueffRf, .param .u64 __cudaparmf3__Z14ConvertUVToHueffRf)

	.visible .func _Z15ConvertYUVToHSL8PixelYUVR8PixelHSL (.param .align 16 .b8 __cudaparmf1__Z15ConvertYUVToHSL8PixelYUVR8PixelHSL[16], .param .u64 __cudaparmf2__Z15ConvertYUVToHSL8PixelYUVR8PixelHSL)

	.visible .func (.param .f32 __cudaretf__Z5ClampIfET_S0_S0_S0_) _Z5ClampIfET_S0_S0_S0_ (.param .f32 __cudaparmf1__Z5ClampIfET_S0_S0_S0_, .param .f32 __cudaparmf2__Z5ClampIfET_S0_S0_S0_, .param .f32 __cudaparmf3__Z5ClampIfET_S0_S0_S0_)

	.visible .func _Z15ConvertRGBtoHSL8PixelRGBR8PixelHSL (.param .align 16 .b8 __cudaparmf1__Z15ConvertRGBtoHSL8PixelRGBR8PixelHSL[16], .param .u64 __cudaparmf2__Z15ConvertRGBtoHSL8PixelRGBR8PixelHSL)

	.visible .func _Z15ConvertHSLtoRGB8PixelHSLR8PixelRGB (.param .align 16 .b8 __cudaparmf1__Z15ConvertHSLtoRGB8PixelHSLR8PixelRGB[16], .param .u64 __cudaparmf2__Z15ConvertHSLtoRGB8PixelHSLR8PixelRGB)

	.visible .func (.param .align 16 .b8 __cudaretf__Z15ConvertRGBtoYIQ8PixelRGB[16]) _Z15ConvertRGBtoYIQ8PixelRGB (.param .align 16 .b8 __cudaparmf1__Z15ConvertRGBtoYIQ8PixelRGB[16])

	.visible .func (.param .align 16 .b8 __cudaretf__Z15ConvertYIQtoRGB6float4[16]) _Z15ConvertYIQtoRGB6float4 (.param .align 16 .b8 __cudaparmf1__Z15ConvertYIQtoRGB6float4[16])

	//-----------------------------------------------------------
	// Compiling C:/Users/dvaeng/AppData/Local/Temp/tmpxft_000039f0_00000000-11_Invert.cpp3.i (C:/Users/dvaeng/AppData/Local/Temp/ccBI#.a13648)
	//-----------------------------------------------------------

	//-----------------------------------------------------------
	// Options:
	//-----------------------------------------------------------
	//  Target:ptx, ISA:sm_20, Endian:little, Pointer Size:64
	//  -O3	(Optimization level)
	//  -g0	(Debug level)
	//  -m2	(Report advisories)
	//-----------------------------------------------------------

	.file	1	"C:/Users/dvaeng/AppData/Local/Temp/tmpxft_000039f0_00000000-10_Invert.cudafe2.gpu"
	.file	2	"c:\Mulder64\shared\adobe\MediaCore\GPUFoundation\API\Inc\GPUFoundation/PixelFormat.h"
	.file	3	"c:\Mulder64\shared\adobe\MediaCore\Display\Inc\CUDA/Effects/Invert.h"
	.file	4	"c:\Mulder64\shared\adobe\MediaCore\GPUFoundation\API\Inc\GPUFoundation/KernelSupport/PixelRGB.h"
	.file	5	"c:\Mulder64\shared\adobe\MediaCore\GPUFoundation\API\Inc\GPUFoundation/KernelSupport/PixelYUV.h"
	.file	6	"c:\Mulder64\shared\adobe\MediaCore\GPUFoundation\API\Inc\GPUFoundation/KernelSupport/PixelHSL.h"
	.file	7	"C:\Program Files (x86)\Microsoft Visual Studio 9.0\VC\include\crtdefs.h"
	.file	8	"c:\Mulder64\shared\adobe\MediaCore\External\3rdParty\NVIDIA\CUDA\win\include\crt/device_runtime.h"
	.file	9	"c:\Mulder64\shared\adobe\MediaCore\External\3rdParty\NVIDIA\CUDA\win\include\host_defines.h"
	.file	10	"c:\Mulder64\shared\adobe\MediaCore\External\3rdParty\NVIDIA\CUDA\win\include\builtin_types.h"
	.file	11	"c:\mulder64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\include\device_types.h"
	.file	12	"c:\mulder64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\include\driver_types.h"
	.file	13	"c:\mulder64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\include\surface_types.h"
	.file	14	"c:\mulder64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\include\texture_types.h"
	.file	15	"c:\mulder64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\include\vector_types.h"
	.file	16	"c:\mulder64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\include\builtin_types.h"
	.file	17	"c:\mulder64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\include\host_defines.h"
	.file	18	"c:\Mulder64\shared\adobe\MediaCore\External\3rdParty\NVIDIA\CUDA\win\include\device_launch_parameters.h"
	.file	19	"c:\mulder64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\include\crt\storage_class.h"
	.file	20	"C:\Program Files (x86)\Microsoft Visual Studio 9.0\VC\include\time.h"
	.file	21	"c:\Mulder64\shared\adobe\MediaCore\GPUFoundation\API\Inc\GPUFoundation/KernelSupport/Utils.h"
	.file	22	"c:\Mulder64\shared\adobe\MediaCore\GPUFoundation\API\Inc\GPUFoundation/KernelSupport/VectorUtils.h"
	.file	23	"c:\Mulder64\shared\adobe\MediaCore\GPUFoundation\API\Inc\GPUFoundation/Numeric.h"
	.file	24	"c:/Mulder64/shared/adobe/MediaCore/Display/Src/CUDA/Effects/Invert.cu"
	.file	25	"c:\Mulder64\shared\adobe\MediaCore\External\3rdParty\NVIDIA\CUDA\win\include\common_functions.h"
	.file	26	"c:\mulder64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\include\math_functions.h"
	.file	27	"c:\mulder64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\include\math_constants.h"
	.file	28	"c:\mulder64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\include\device_functions.h"
	.file	29	"c:\mulder64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\include\sm_11_atomic_functions.h"
	.file	30	"c:\mulder64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\include\sm_12_atomic_functions.h"
	.file	31	"c:\mulder64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\include\sm_13_double_functions.h"
	.file	32	"c:\mulder64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\include\sm_20_atomic_functions.h"
	.file	33	"c:\mulder64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\include\sm_20_intrinsics.h"
	.file	34	"c:\mulder64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\include\surface_functions.h"
	.file	35	"c:\mulder64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\include\texture_fetch_functions.h"
	.file	36	"c:\mulder64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\include\math_functions_dbl_ptx3.h"
	.file	37	"c:\Mulder64\shared\adobe\MediaCore\GPUFoundation\API\Inc\GPUFoundation/KernelSupport/ColorSpaceConvert.h"


	.visible .func (.param .s32 __cudaretf__Z15IntegerMultiplyii) _Z15IntegerMultiplyii (.param .s32 __cudaparmf1__Z15IntegerMultiplyii, .param .s32 __cudaparmf2__Z15IntegerMultiplyii)
	{
	.reg .u32 %r<7>;
	.loc	21	60	0
$LDWbegin__Z15IntegerMultiplyii:
	ld.param.u32 	%r1, [__cudaparmf1__Z15IntegerMultiplyii];
	mov.s32 	%r2, %r1;
	ld.param.u32 	%r3, [__cudaparmf2__Z15IntegerMultiplyii];
	mov.s32 	%r4, %r3;
	.loc	21	64	0
	mul.lo.s32 	%r5, %r2, %r4;
	st.param.s32 	[__cudaretf__Z15IntegerMultiplyii], %r5;
	ret;
$LDWend__Z15IntegerMultiplyii:
	} // _Z15IntegerMultiplyii

	.visible .func (.param .s32 __cudaretf__Z17Standard2DKernelXv) _Z17Standard2DKernelXv ()
	{
	.reg .u32 %r<7>;
	.loc	21	73	0
$LDWbegin__Z17Standard2DKernelXv:
	.loc	21	74	0
	mov.u32 	%r1, %tid.x;
	cvt.s32.u32 	%r2, %ctaid.x;
	cvt.s32.u32 	%r3, %ntid.x;
	mul.lo.s32 	%r4, %r2, %r3;
	add.u32 	%r5, %r1, %r4;
	st.param.s32 	[__cudaretf__Z17Standard2DKernelXv], %r5;
	ret;
$LDWend__Z17Standard2DKernelXv:
	} // _Z17Standard2DKernelXv

	.visible .func (.param .s32 __cudaretf__Z17Standard2DKernelYv) _Z17Standard2DKernelYv ()
	{
	.reg .u32 %r<7>;
	.loc	21	77	0
$LDWbegin__Z17Standard2DKernelYv:
	.loc	21	78	0
	mov.u32 	%r1, %tid.y;
	cvt.s32.u32 	%r2, %ctaid.y;
	cvt.s32.u32 	%r3, %ntid.y;
	mul.lo.s32 	%r4, %r2, %r3;
	add.u32 	%r5, %r1, %r4;
	st.param.s32 	[__cudaretf__Z17Standard2DKernelYv], %r5;
	ret;
$LDWend__Z17Standard2DKernelYv:
	} // _Z17Standard2DKernelYv

	.visible .func (.param .align 16 .b8 __cudaretf__Z13Half4ToFloat47ushort4[16]) _Z13Half4ToFloat47ushort4 (.param .align 8 .b8 __cudaparmf1__Z13Half4ToFloat47ushort4[8])
	{
	.reg .u32 %r<14>;
	.reg .f32 %f<9>;
	.loc	21	86	0
$LDWbegin__Z13Half4ToFloat47ushort4:
	ld.param.u16 	%r1, [__cudaparmf1__Z13Half4ToFloat47ushort4+0];
	mov.s32 	%r2, %r1;
	ld.param.u16 	%r3, [__cudaparmf1__Z13Half4ToFloat47ushort4+2];
	mov.s32 	%r4, %r3;
	ld.param.u16 	%r5, [__cudaparmf1__Z13Half4ToFloat47ushort4+4];
	mov.s32 	%r6, %r5;
	ld.param.u16 	%r7, [__cudaparmf1__Z13Half4ToFloat47ushort4+6];
	mov.s32 	%r8, %r7;
	.loc	21	87	0
	cvt.u16.u32 	%r9, %r4;
	{ .reg .b32 %b1;
	mov.b32		%b1, %r9;
	cvt.ftz.f32.f16	%f1, %b1; }
	cvt.u16.u32 	%r10, %r6;
	{ .reg .b32 %b1;
	mov.b32		%b1, %r10;
	cvt.ftz.f32.f16	%f2, %b1; }
	cvt.u16.u32 	%r11, %r8;
	{ .reg .b32 %b1;
	mov.b32		%b1, %r11;
	cvt.ftz.f32.f16	%f3, %b1; }
	cvt.u16.u32 	%r12, %r2;
	{ .reg .b32 %b1;
	mov.b32		%b1, %r12;
	cvt.ftz.f32.f16	%f4, %b1; }
	st.param.f32 	[__cudaretf__Z13Half4ToFloat47ushort4+0], %f4;
	mov.f32 	%f5, %f1;
	st.param.f32 	[__cudaretf__Z13Half4ToFloat47ushort4+4], %f5;
	mov.f32 	%f6, %f2;
	st.param.f32 	[__cudaretf__Z13Half4ToFloat47ushort4+8], %f6;
	mov.f32 	%f7, %f3;
	st.param.f32 	[__cudaretf__Z13Half4ToFloat47ushort4+12], %f7;
	ret;
$LDWend__Z13Half4ToFloat47ushort4:
	} // _Z13Half4ToFloat47ushort4

	.visible .func (.param .align 8 .b8 __cudaretf__Z13Float4ToHalf46float4[8]) _Z13Float4ToHalf46float4 (.param .align 16 .b8 __cudaparmf1__Z13Float4ToHalf46float4[16])
	{
	.reg .u32 %r<13>;
	.reg .f32 %f<10>;
	.loc	21	95	0
$LDWbegin__Z13Float4ToHalf46float4:
	ld.param.f32 	%f1, [__cudaparmf1__Z13Float4ToHalf46float4+0];
	mov.f32 	%f2, %f1;
	ld.param.f32 	%f3, [__cudaparmf1__Z13Float4ToHalf46float4+4];
	mov.f32 	%f4, %f3;
	ld.param.f32 	%f5, [__cudaparmf1__Z13Float4ToHalf46float4+8];
	mov.f32 	%f6, %f5;
	ld.param.f32 	%f7, [__cudaparmf1__Z13Float4ToHalf46float4+12];
	mov.f32 	%f8, %f7;
	.loc	21	96	0
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f4;
	mov.b32		%r1, %b1; }
	cvt.u16.u32 	%r2, %r1;
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f6;
	mov.b32		%r3, %b1; }
	cvt.u16.u32 	%r4, %r3;
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f8;
	mov.b32		%r5, %b1; }
	cvt.u16.u32 	%r6, %r5;
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f2;
	mov.b32		%r7, %b1; }
	cvt.u16.u32 	%r8, %r7;
	st.param.u16 	[__cudaretf__Z13Float4ToHalf46float4+0], %r8;
	mov.s32 	%r9, %r2;
	st.param.u16 	[__cudaretf__Z13Float4ToHalf46float4+2], %r9;
	mov.s32 	%r10, %r4;
	st.param.u16 	[__cudaretf__Z13Float4ToHalf46float4+4], %r10;
	mov.s32 	%r11, %r6;
	st.param.u16 	[__cudaretf__Z13Float4ToHalf46float4+6], %r11;
	ret;
$LDWend__Z13Float4ToHalf46float4:
	} // _Z13Float4ToHalf46float4

	.visible .func (.param .u32 __cudaretf__Z4Mix3RjS_S_) _Z4Mix3RjS_S_ (.param .u64 __cudaparmf1__Z4Mix3RjS_S_, .param .u64 __cudaparmf2__Z4Mix3RjS_S_, .param .u64 __cudaparmf3__Z4Mix3RjS_S_)
	{
	.reg .u32 %r<75>;
	.reg .u64 %rd<8>;
	.loc	21	138	0
$LDWbegin__Z4Mix3RjS_S_:
	ld.param.u64 	%rd1, [__cudaparmf1__Z4Mix3RjS_S_];
	mov.s64 	%rd2, %rd1;
	ld.param.u64 	%rd3, [__cudaparmf2__Z4Mix3RjS_S_];
	mov.s64 	%rd4, %rd3;
	ld.param.u64 	%rd5, [__cudaparmf3__Z4Mix3RjS_S_];
	mov.s64 	%rd6, %rd5;
	.loc	21	139	0
	ld.u32 	%r1, [%rd2+0];
	ld.u32 	%r2, [%rd4+0];
	sub.u32 	%r3, %r1, %r2;
	st.u32 	[%rd2+0], %r3;
	ld.u32 	%r4, [%rd6+0];
	sub.u32 	%r5, %r3, %r4;
	st.u32 	[%rd2+0], %r5;
	ld.u32 	%r6, [%rd6+0];
	shr.u32 	%r7, %r6, 13;
	xor.b32 	%r8, %r5, %r7;
	st.u32 	[%rd2+0], %r8;
	.loc	21	140	0
	ld.u32 	%r9, [%rd4+0];
	ld.u32 	%r10, [%rd6+0];
	sub.u32 	%r11, %r9, %r10;
	st.u32 	[%rd4+0], %r11;
	ld.u32 	%r12, [%rd2+0];
	sub.u32 	%r13, %r11, %r12;
	st.u32 	[%rd4+0], %r13;
	ld.u32 	%r14, [%rd2+0];
	shl.b32 	%r15, %r14, 8;
	xor.b32 	%r16, %r13, %r15;
	st.u32 	[%rd4+0], %r16;
	.loc	21	141	0
	ld.u32 	%r17, [%rd6+0];
	ld.u32 	%r18, [%rd2+0];
	sub.u32 	%r19, %r17, %r18;
	st.u32 	[%rd6+0], %r19;
	ld.u32 	%r20, [%rd4+0];
	sub.u32 	%r21, %r19, %r20;
	st.u32 	[%rd6+0], %r21;
	ld.u32 	%r22, [%rd4+0];
	shr.u32 	%r23, %r22, 13;
	xor.b32 	%r24, %r21, %r23;
	st.u32 	[%rd6+0], %r24;
	.loc	21	142	0
	ld.u32 	%r25, [%rd2+0];
	ld.u32 	%r26, [%rd4+0];
	sub.u32 	%r27, %r25, %r26;
	st.u32 	[%rd2+0], %r27;
	ld.u32 	%r28, [%rd6+0];
	sub.u32 	%r29, %r27, %r28;
	st.u32 	[%rd2+0], %r29;
	ld.u32 	%r30, [%rd6+0];
	shr.u32 	%r31, %r30, 12;
	xor.b32 	%r32, %r29, %r31;
	st.u32 	[%rd2+0], %r32;
	.loc	21	143	0
	ld.u32 	%r33, [%rd4+0];
	ld.u32 	%r34, [%rd6+0];
	sub.u32 	%r35, %r33, %r34;
	st.u32 	[%rd4+0], %r35;
	ld.u32 	%r36, [%rd2+0];
	sub.u32 	%r37, %r35, %r36;
	st.u32 	[%rd4+0], %r37;
	ld.u32 	%r38, [%rd2+0];
	shl.b32 	%r39, %r38, 16;
	xor.b32 	%r40, %r37, %r39;
	st.u32 	[%rd4+0], %r40;
	.loc	21	144	0
	ld.u32 	%r41, [%rd6+0];
	ld.u32 	%r42, [%rd2+0];
	sub.u32 	%r43, %r41, %r42;
	st.u32 	[%rd6+0], %r43;
	ld.u32 	%r44, [%rd4+0];
	sub.u32 	%r45, %r43, %r44;
	st.u32 	[%rd6+0], %r45;
	ld.u32 	%r46, [%rd4+0];
	shr.u32 	%r47, %r46, 5;
	xor.b32 	%r48, %r45, %r47;
	st.u32 	[%rd6+0], %r48;
	.loc	21	145	0
	ld.u32 	%r49, [%rd2+0];
	ld.u32 	%r50, [%rd4+0];
	sub.u32 	%r51, %r49, %r50;
	st.u32 	[%rd2+0], %r51;
	ld.u32 	%r52, [%rd6+0];
	sub.u32 	%r53, %r51, %r52;
	st.u32 	[%rd2+0], %r53;
	ld.u32 	%r54, [%rd6+0];
	shr.u32 	%r55, %r54, 3;
	xor.b32 	%r56, %r53, %r55;
	st.u32 	[%rd2+0], %r56;
	.loc	21	146	0
	ld.u32 	%r57, [%rd4+0];
	ld.u32 	%r58, [%rd6+0];
	sub.u32 	%r59, %r57, %r58;
	st.u32 	[%rd4+0], %r59;
	ld.u32 	%r60, [%rd2+0];
	sub.u32 	%r61, %r59, %r60;
	st.u32 	[%rd4+0], %r61;
	ld.u32 	%r62, [%rd2+0];
	shl.b32 	%r63, %r62, 10;
	xor.b32 	%r64, %r61, %r63;
	st.u32 	[%rd4+0], %r64;
	.loc	21	147	0
	ld.u32 	%r65, [%rd6+0];
	ld.u32 	%r66, [%rd2+0];
	sub.u32 	%r67, %r65, %r66;
	st.u32 	[%rd6+0], %r67;
	ld.u32 	%r68, [%rd4+0];
	sub.u32 	%r69, %r67, %r68;
	st.u32 	[%rd6+0], %r69;
	ld.u32 	%r70, [%rd4+0];
	shr.u32 	%r71, %r70, 15;
	xor.b32 	%r72, %r69, %r71;
	st.u32 	[%rd6+0], %r72;
	.loc	21	148	0
	mov.s32 	%r73, %r72;
	st.param.u32 	[__cudaretf__Z4Mix3RjS_S_], %r73;
	ret;
$LDWend__Z4Mix3RjS_S_:
	} // _Z4Mix3RjS_S_

	.visible .func (.param .s32 __cudaretf__Z4Randj) _Z4Randj (.param .u32 __cudaparmf1__Z4Randj)
	{
	.reg .u32 %r<14>;
	.loc	21	152	0
$LDWbegin__Z4Randj:
	ld.param.u32 	%r1, [__cudaparmf1__Z4Randj];
	mov.s32 	%r2, %r1;
	.loc	21	163	0
	mul.lo.u32 	%r3, %r2, 1103515245;
	add.u32 	%r4, %r3, 12345;
	shr.u32 	%r5, %r4, 16;
	and.b32 	%r6, %r5, 255;
	shl.b32 	%r7, %r6, 7;
	mul.lo.u32 	%r8, %r2, -1029531031;
	sub.u32 	%r9, %r8, 740551042;
	shr.u32 	%r10, %r9, 16;
	and.b32 	%r11, %r10, 255;
	xor.b32 	%r12, %r7, %r11;
	st.param.s32 	[__cudaretf__Z4Randj], %r12;
	ret;
$LDWend__Z4Randj:
	} // _Z4Randj

	.visible .func (.param .s32 __cudaretf__Z6Rand2Djjj) _Z6Rand2Djjj (.param .u32 __cudaparmf1__Z6Rand2Djjj, .param .u32 __cudaparmf2__Z6Rand2Djjj, .param .u32 __cudaparmf3__Z6Rand2Djjj)
	{
	.reg .u32 %r<54>;
	.loc	21	169	0
$LDWbegin__Z6Rand2Djjj:
	ld.param.u32 	%r1, [__cudaparmf1__Z6Rand2Djjj];
	mov.s32 	%r2, %r1;
	ld.param.u32 	%r3, [__cudaparmf2__Z6Rand2Djjj];
	mov.s32 	%r4, %r3;
	ld.param.u32 	%r5, [__cudaparmf3__Z6Rand2Djjj];
	mov.s32 	%r6, %r5;
	.loc	21	139	0
	sub.u32 	%r7, %r2, %r4;
	sub.u32 	%r8, %r7, %r6;
	shr.u32 	%r9, %r6, 13;
	xor.b32 	%r10, %r8, %r9;
	.loc	21	140	0
	sub.u32 	%r11, %r4, %r6;
	sub.u32 	%r12, %r11, %r10;
	shl.b32 	%r13, %r10, 8;
	xor.b32 	%r14, %r12, %r13;
	.loc	21	141	0
	sub.u32 	%r15, %r6, %r10;
	sub.u32 	%r16, %r15, %r14;
	shr.u32 	%r17, %r14, 13;
	xor.b32 	%r18, %r16, %r17;
	.loc	21	142	0
	sub.u32 	%r19, %r10, %r14;
	sub.u32 	%r20, %r19, %r18;
	shr.u32 	%r21, %r18, 12;
	xor.b32 	%r22, %r20, %r21;
	.loc	21	143	0
	sub.u32 	%r23, %r14, %r18;
	sub.u32 	%r24, %r23, %r22;
	shl.b32 	%r25, %r22, 16;
	xor.b32 	%r26, %r24, %r25;
	.loc	21	144	0
	sub.u32 	%r27, %r18, %r22;
	sub.u32 	%r28, %r27, %r26;
	shr.u32 	%r29, %r26, 5;
	xor.b32 	%r30, %r28, %r29;
	.loc	21	145	0
	sub.u32 	%r31, %r22, %r26;
	sub.u32 	%r32, %r31, %r30;
	shr.u32 	%r33, %r30, 3;
	xor.b32 	%r34, %r32, %r33;
	.loc	21	146	0
	sub.u32 	%r35, %r26, %r30;
	sub.u32 	%r36, %r35, %r34;
	shl.b32 	%r37, %r34, 10;
	xor.b32 	%r38, %r36, %r37;
	.loc	21	147	0
	sub.u32 	%r39, %r30, %r34;
	sub.u32 	%r40, %r39, %r38;
	shr.u32 	%r41, %r38, 15;
	xor.b32 	%r42, %r40, %r41;
	.loc	21	170	0
	mul.lo.u32 	%r43, %r42, 1103515245;
	add.u32 	%r44, %r43, 12345;
	shr.u32 	%r45, %r44, 16;
	and.b32 	%r46, %r45, 255;
	shl.b32 	%r47, %r46, 7;
	mul.lo.u32 	%r48, %r42, -1029531031;
	sub.u32 	%r49, %r48, 740551042;
	shr.u32 	%r50, %r49, 16;
	and.b32 	%r51, %r50, 255;
	xor.b32 	%r52, %r47, %r51;
	st.param.s32 	[__cudaretf__Z6Rand2Djjj], %r52;
	ret;
$LDWend__Z6Rand2Djjj:
	} // _Z6Rand2Djjj

	.visible .func (.param .s32 __cudaretf__Z6Rand2Dj) _Z6Rand2Dj (.param .u32 __cudaparmf1__Z6Rand2Dj)
	{
	.reg .u32 %r<60>;
	.loc	21	175	0
$LDWbegin__Z6Rand2Dj:
	ld.param.u32 	%r1, [__cudaparmf1__Z6Rand2Dj];
	mov.s32 	%r2, %r1;
	.loc	21	143	0
	cvt.s32.u32 	%r3, %ctaid.y;
	cvt.s32.u32 	%r4, %ntid.y;
	mul.lo.s32 	%r5, %r3, %r4;
	cvt.s32.u32 	%r6, %ctaid.x;
	cvt.s32.u32 	%r7, %ntid.x;
	mul.lo.s32 	%r8, %r6, %r7;
	mov.u32 	%r9, %tid.y;
	add.u32 	%r10, %r5, %r9;
	mov.u32 	%r11, %tid.x;
	add.u32 	%r12, %r8, %r11;
	shr.u32 	%r13, %r10, 13;
	sub.u32 	%r14, %r2, %r12;
	sub.u32 	%r15, %r12, %r10;
	sub.u32 	%r16, %r14, %r10;
	xor.b32 	%r17, %r13, %r16;
	shl.b32 	%r18, %r17, 8;
	sub.u32 	%r19, %r15, %r17;
	sub.u32 	%r20, %r10, %r17;
	xor.b32 	%r21, %r18, %r19;
	shr.u32 	%r22, %r21, 13;
	sub.u32 	%r23, %r20, %r21;
	sub.u32 	%r24, %r17, %r21;
	xor.b32 	%r25, %r22, %r23;
	shr.u32 	%r26, %r25, 12;
	sub.u32 	%r27, %r24, %r25;
	xor.b32 	%r28, %r26, %r27;
	sub.u32 	%r29, %r21, %r25;
	sub.u32 	%r30, %r29, %r28;
	shl.b32 	%r31, %r28, 16;
	xor.b32 	%r32, %r30, %r31;
	.loc	21	144	0
	sub.u32 	%r33, %r25, %r28;
	sub.u32 	%r34, %r33, %r32;
	shr.u32 	%r35, %r32, 5;
	xor.b32 	%r36, %r34, %r35;
	.loc	21	145	0
	sub.u32 	%r37, %r28, %r32;
	sub.u32 	%r38, %r37, %r36;
	shr.u32 	%r39, %r36, 3;
	xor.b32 	%r40, %r38, %r39;
	.loc	21	146	0
	sub.u32 	%r41, %r32, %r36;
	sub.u32 	%r42, %r41, %r40;
	shl.b32 	%r43, %r40, 10;
	xor.b32 	%r44, %r42, %r43;
	.loc	21	147	0
	sub.u32 	%r45, %r36, %r40;
	sub.u32 	%r46, %r45, %r44;
	shr.u32 	%r47, %r44, 15;
	xor.b32 	%r48, %r46, %r47;
	.loc	21	176	0
	mul.lo.u32 	%r49, %r48, 1103515245;
	add.u32 	%r50, %r49, 12345;
	shr.u32 	%r51, %r50, 16;
	and.b32 	%r52, %r51, 255;
	shl.b32 	%r53, %r52, 7;
	mul.lo.u32 	%r54, %r48, -1029531031;
	sub.u32 	%r55, %r54, 740551042;
	shr.u32 	%r56, %r55, 16;
	and.b32 	%r57, %r56, 255;
	xor.b32 	%r58, %r53, %r57;
	st.param.s32 	[__cudaretf__Z6Rand2Dj], %r58;
	ret;
$LDWend__Z6Rand2Dj:
	} // _Z6Rand2Dj

	.visible .func (.param .align 8 .b8 __cudaretf__Z6Read2DI7ushort4ET_PKS1_iii[8]) _Z6Read2DI7ushort4ET_PKS1_iii (.param .u64 __cudaparmf1__Z6Read2DI7ushort4ET_PKS1_iii, .param .s32 __cudaparmf2__Z6Read2DI7ushort4ET_PKS1_iii, .param .s32 __cudaparmf3__Z6Read2DI7ushort4ET_PKS1_iii, .param .s32 __cudaparmf4__Z6Read2DI7ushort4ET_PKS1_iii)
	{
	.reg .u32 %r<14>;
	.reg .u64 %rd<7>;
	.loc	21	114	0
$LDWbegin__Z6Read2DI7ushort4ET_PKS1_iii:
	ld.param.u64 	%rd1, [__cudaparmf1__Z6Read2DI7ushort4ET_PKS1_iii];
	mov.s64 	%rd2, %rd1;
	ld.param.u32 	%r1, [__cudaparmf2__Z6Read2DI7ushort4ET_PKS1_iii];
	mov.s32 	%r2, %r1;
	ld.param.u32 	%r3, [__cudaparmf3__Z6Read2DI7ushort4ET_PKS1_iii];
	mov.s32 	%r4, %r3;
	ld.param.u32 	%r5, [__cudaparmf4__Z6Read2DI7ushort4ET_PKS1_iii];
	mov.s32 	%r6, %r5;
	.loc	21	115	0
	mul.lo.s32 	%r7, %r2, %r6;
	add.s32 	%r8, %r4, %r7;
	cvt.s64.s32 	%rd3, %r8;
	mul.wide.s32 	%rd4, %r8, 8;
	add.u64 	%rd5, %rd2, %rd4;
	ld.v4.u16 	{%r9,%r10,%r11,%r12}, [%rd5+0];
	st.param.u16 	[__cudaretf__Z6Read2DI7ushort4ET_PKS1_iii+0], %r9;
	st.param.u16 	[__cudaretf__Z6Read2DI7ushort4ET_PKS1_iii+2], %r10;
	st.param.u16 	[__cudaretf__Z6Read2DI7ushort4ET_PKS1_iii+4], %r11;
	st.param.u16 	[__cudaretf__Z6Read2DI7ushort4ET_PKS1_iii+6], %r12;
	ret;
$LDWend__Z6Read2DI7ushort4ET_PKS1_iii:
	} // _Z6Read2DI7ushort4ET_PKS1_iii

	.visible .func (.param .align 16 .b8 __cudaretf__Z6Read2DI6float4ET_PKS1_iii[16]) _Z6Read2DI6float4ET_PKS1_iii (.param .u64 __cudaparmf1__Z6Read2DI6float4ET_PKS1_iii, .param .s32 __cudaparmf2__Z6Read2DI6float4ET_PKS1_iii, .param .s32 __cudaparmf3__Z6Read2DI6float4ET_PKS1_iii, .param .s32 __cudaparmf4__Z6Read2DI6float4ET_PKS1_iii)
	{
	.reg .u32 %r<10>;
	.reg .u64 %rd<7>;
	.reg .f32 %f<6>;
	.loc	21	114	0
$LDWbegin__Z6Read2DI6float4ET_PKS1_iii:
	ld.param.u64 	%rd1, [__cudaparmf1__Z6Read2DI6float4ET_PKS1_iii];
	mov.s64 	%rd2, %rd1;
	ld.param.u32 	%r1, [__cudaparmf2__Z6Read2DI6float4ET_PKS1_iii];
	mov.s32 	%r2, %r1;
	ld.param.u32 	%r3, [__cudaparmf3__Z6Read2DI6float4ET_PKS1_iii];
	mov.s32 	%r4, %r3;
	ld.param.u32 	%r5, [__cudaparmf4__Z6Read2DI6float4ET_PKS1_iii];
	mov.s32 	%r6, %r5;
	.loc	21	115	0
	mul.lo.s32 	%r7, %r2, %r6;
	add.s32 	%r8, %r4, %r7;
	cvt.s64.s32 	%rd3, %r8;
	mul.wide.s32 	%rd4, %r8, 16;
	add.u64 	%rd5, %rd2, %rd4;
	ld.v4.f32 	{%f1,%f2,%f3,%f4}, [%rd5+0];
	st.param.f32 	[__cudaretf__Z6Read2DI6float4ET_PKS1_iii+0], %f1;
	st.param.f32 	[__cudaretf__Z6Read2DI6float4ET_PKS1_iii+4], %f2;
	st.param.f32 	[__cudaretf__Z6Read2DI6float4ET_PKS1_iii+8], %f3;
	st.param.f32 	[__cudaretf__Z6Read2DI6float4ET_PKS1_iii+12], %f4;
	ret;
$LDWend__Z6Read2DI6float4ET_PKS1_iii:
	} // _Z6Read2DI6float4ET_PKS1_iii

	.visible .func _Z7Write2DI7ushort4EvT_PS1_iii (.param .align 8 .b8 __cudaparmf1__Z7Write2DI7ushort4EvT_PS1_iii[8], .param .u64 __cudaparmf2__Z7Write2DI7ushort4EvT_PS1_iii, .param .s32 __cudaparmf3__Z7Write2DI7ushort4EvT_PS1_iii, .param .s32 __cudaparmf4__Z7Write2DI7ushort4EvT_PS1_iii, .param .s32 __cudaparmf5__Z7Write2DI7ushort4EvT_PS1_iii)
	{
	.reg .u32 %r<18>;
	.reg .u64 %rd<7>;
	.loc	21	125	0
$LDWbegin__Z7Write2DI7ushort4EvT_PS1_iii:
	ld.param.u16 	%r1, [__cudaparmf1__Z7Write2DI7ushort4EvT_PS1_iii+0];
	mov.s32 	%r2, %r1;
	ld.param.u16 	%r3, [__cudaparmf1__Z7Write2DI7ushort4EvT_PS1_iii+2];
	mov.s32 	%r4, %r3;
	ld.param.u16 	%r5, [__cudaparmf1__Z7Write2DI7ushort4EvT_PS1_iii+4];
	mov.s32 	%r6, %r5;
	ld.param.u16 	%r7, [__cudaparmf1__Z7Write2DI7ushort4EvT_PS1_iii+6];
	mov.s32 	%r8, %r7;
	ld.param.u64 	%rd1, [__cudaparmf2__Z7Write2DI7ushort4EvT_PS1_iii];
	mov.s64 	%rd2, %rd1;
	ld.param.u32 	%r9, [__cudaparmf3__Z7Write2DI7ushort4EvT_PS1_iii];
	mov.s32 	%r10, %r9;
	ld.param.u32 	%r11, [__cudaparmf4__Z7Write2DI7ushort4EvT_PS1_iii];
	mov.s32 	%r12, %r11;
	ld.param.u32 	%r13, [__cudaparmf5__Z7Write2DI7ushort4EvT_PS1_iii];
	mov.s32 	%r14, %r13;
	.loc	21	126	0
	mul.lo.s32 	%r15, %r10, %r14;
	add.s32 	%r16, %r12, %r15;
	cvt.s64.s32 	%rd3, %r16;
	mul.wide.s32 	%rd4, %r16, 8;
	add.u64 	%rd5, %rd2, %rd4;
	st.v4.u16 	[%rd5+0], {%r2,%r4,%r6,%r8};
	.loc	21	127	0
	ret;
$LDWend__Z7Write2DI7ushort4EvT_PS1_iii:
	} // _Z7Write2DI7ushort4EvT_PS1_iii

	.visible .func _Z7Write2DI6float4EvT_PS1_iii (.param .align 16 .b8 __cudaparmf1__Z7Write2DI6float4EvT_PS1_iii[16], .param .u64 __cudaparmf2__Z7Write2DI6float4EvT_PS1_iii, .param .s32 __cudaparmf3__Z7Write2DI6float4EvT_PS1_iii, .param .s32 __cudaparmf4__Z7Write2DI6float4EvT_PS1_iii, .param .s32 __cudaparmf5__Z7Write2DI6float4EvT_PS1_iii)
	{
	.reg .u32 %r<10>;
	.reg .u64 %rd<7>;
	.reg .f32 %f<10>;
	.loc	21	125	0
$LDWbegin__Z7Write2DI6float4EvT_PS1_iii:
	ld.param.f32 	%f1, [__cudaparmf1__Z7Write2DI6float4EvT_PS1_iii+0];
	mov.f32 	%f2, %f1;
	ld.param.f32 	%f3, [__cudaparmf1__Z7Write2DI6float4EvT_PS1_iii+4];
	mov.f32 	%f4, %f3;
	ld.param.f32 	%f5, [__cudaparmf1__Z7Write2DI6float4EvT_PS1_iii+8];
	mov.f32 	%f6, %f5;
	ld.param.f32 	%f7, [__cudaparmf1__Z7Write2DI6float4EvT_PS1_iii+12];
	mov.f32 	%f8, %f7;
	ld.param.u64 	%rd1, [__cudaparmf2__Z7Write2DI6float4EvT_PS1_iii];
	mov.s64 	%rd2, %rd1;
	ld.param.u32 	%r1, [__cudaparmf3__Z7Write2DI6float4EvT_PS1_iii];
	mov.s32 	%r2, %r1;
	ld.param.u32 	%r3, [__cudaparmf4__Z7Write2DI6float4EvT_PS1_iii];
	mov.s32 	%r4, %r3;
	ld.param.u32 	%r5, [__cudaparmf5__Z7Write2DI6float4EvT_PS1_iii];
	mov.s32 	%r6, %r5;
	.loc	21	126	0
	mul.lo.s32 	%r7, %r2, %r6;
	add.s32 	%r8, %r4, %r7;
	cvt.s64.s32 	%rd3, %r8;
	mul.wide.s32 	%rd4, %r8, 16;
	add.u64 	%rd5, %rd2, %rd4;
	st.v4.f32 	[%rd5+0], {%f2,%f4,%f6,%f8};
	.loc	21	127	0
	ret;
$LDWend__Z7Write2DI6float4EvT_PS1_iii:
	} // _Z7Write2DI6float4EvT_PS1_iii

	.visible .func (.param .align 16 .b8 __cudaretf__Z18UnpremultiplyPixel8PixelRGB[16]) _Z18UnpremultiplyPixel8PixelRGB (.param .align 16 .b8 __cudaparmf1__Z18UnpremultiplyPixel8PixelRGB[16])
	{
	.reg .f32 %f<23>;
	.reg .pred %p<3>;
	.loc	4	206	0
$LDWbegin__Z18UnpremultiplyPixel8PixelRGB:
	ld.param.f32 	%f1, [__cudaparmf1__Z18UnpremultiplyPixel8PixelRGB+0];
	mov.f32 	%f2, %f1;
	ld.param.f32 	%f3, [__cudaparmf1__Z18UnpremultiplyPixel8PixelRGB+4];
	mov.f32 	%f4, %f3;
	ld.param.f32 	%f5, [__cudaparmf1__Z18UnpremultiplyPixel8PixelRGB+8];
	mov.f32 	%f6, %f5;
	ld.param.f32 	%f7, [__cudaparmf1__Z18UnpremultiplyPixel8PixelRGB+12];
	mov.f32 	%f8, %f7;
	.loc	4	208	0
	cvt.ftz.sat.f32.f32 	%f9, %f8;
	mov.f32 	%f10, %f9;
	mov.f32 	%f11, 0fb70637bd;    	// -8e-006
	add.ftz.f32 	%f12, %f9, %f11;
	mov.f32 	%f13, 0f00000000;    	// 0
	setp.le.ftz.f32 	%p1, %f12, %f13;
	@%p1 bra 	$Lt_13_1282;
	.loc	4	213	0
	rcp.approx.ftz.f32 	%f14, %f9;
	mul.ftz.f32 	%f15, %f14, %f6;
	.loc	4	214	0
	mul.ftz.f32 	%f16, %f14, %f4;
	.loc	4	215	0
	mul.ftz.f32 	%f17, %f14, %f2;
	bra.uni 	$Lt_13_1026;
$Lt_13_1282:
	.loc	4	219	0
	mov.f32 	%f15, 0f00000000;    	// 0
	mov.f32 	%f16, 0f00000000;    	// 0
	mov.f32 	%f17, 0f00000000;    	// 0
	mov.f32 	%f10, 0f00000000;    	// 0
$Lt_13_1026:
	.loc	4	224	0
	mov.f32 	%f18, %f17;
	st.param.f32 	[__cudaretf__Z18UnpremultiplyPixel8PixelRGB+0], %f18;
	mov.f32 	%f19, %f16;
	st.param.f32 	[__cudaretf__Z18UnpremultiplyPixel8PixelRGB+4], %f19;
	mov.f32 	%f20, %f15;
	st.param.f32 	[__cudaretf__Z18UnpremultiplyPixel8PixelRGB+8], %f20;
	mov.f32 	%f21, %f10;
	st.param.f32 	[__cudaretf__Z18UnpremultiplyPixel8PixelRGB+12], %f21;
	ret;
$LDWend__Z18UnpremultiplyPixel8PixelRGB:
	} // _Z18UnpremultiplyPixel8PixelRGB

	.visible .func (.param .f32 __cudaretf__Z13ToLinearColorf) _Z13ToLinearColorf (.param .f32 __cudaparmf1__Z13ToLinearColorf)
	{
	.reg .f32 %f<15>;
	.reg .pred %p<3>;
	.loc	4	231	0
$LDWbegin__Z13ToLinearColorf:
	ld.param.f32 	%f1, [__cudaparmf1__Z13ToLinearColorf];
	mov.f32 	%f2, %f1;
	mov.f32 	%f3, 0f00000000;     	// 0
	setp.lt.ftz.f32 	%p1, %f2, %f3;
	@!%p1 bra 	$Lt_14_1026;
	.loc	4	234	0
	neg.ftz.f32 	%f4, %f2;
	lg2.approx.ftz.f32 	%f5, %f4;
	mov.f32 	%f6, 0f400ccccd;     	// 2.2
	mul.ftz.f32 	%f7, %f5, %f6;
	ex2.approx.ftz.f32 	%f8, %f7;
	neg.ftz.f32 	%f9, %f8;
	bra.uni 	$LBB4__Z13ToLinearColorf;
$Lt_14_1026:
	.loc	4	236	0
	lg2.approx.ftz.f32 	%f10, %f2;
	mov.f32 	%f11, 0f400ccccd;    	// 2.2
	mul.ftz.f32 	%f12, %f10, %f11;
	ex2.approx.ftz.f32 	%f9, %f12;
$LBB4__Z13ToLinearColorf:
	mov.f32 	%f13, %f9;
	st.param.f32 	[__cudaretf__Z13ToLinearColorf], %f13;
	ret;
$LDWend__Z13ToLinearColorf:
	} // _Z13ToLinearColorf

	.visible .func (.param .f32 __cudaretf__Z15FromLinearColorf) _Z15FromLinearColorf (.param .f32 __cudaparmf1__Z15FromLinearColorf)
	{
	.reg .f32 %f<15>;
	.reg .pred %p<3>;
	.loc	4	239	0
$LDWbegin__Z15FromLinearColorf:
	ld.param.f32 	%f1, [__cudaparmf1__Z15FromLinearColorf];
	mov.f32 	%f2, %f1;
	mov.f32 	%f3, 0f00000000;     	// 0
	setp.lt.ftz.f32 	%p1, %f2, %f3;
	@!%p1 bra 	$Lt_15_1026;
	.loc	4	242	0
	neg.ftz.f32 	%f4, %f2;
	lg2.approx.ftz.f32 	%f5, %f4;
	mov.f32 	%f6, 0f3ee8ba2e;     	// 0.454545
	mul.ftz.f32 	%f7, %f5, %f6;
	ex2.approx.ftz.f32 	%f8, %f7;
	neg.ftz.f32 	%f9, %f8;
	bra.uni 	$LBB4__Z15FromLinearColorf;
$Lt_15_1026:
	.loc	4	244	0
	lg2.approx.ftz.f32 	%f10, %f2;
	mov.f32 	%f11, 0f3ee8ba2e;    	// 0.454545
	mul.ftz.f32 	%f12, %f10, %f11;
	ex2.approx.ftz.f32 	%f9, %f12;
$LBB4__Z15FromLinearColorf:
	mov.f32 	%f13, %f9;
	st.param.f32 	[__cudaretf__Z15FromLinearColorf], %f13;
	ret;
$LDWend__Z15FromLinearColorf:
	} // _Z15FromLinearColorf

	.visible .func (.param .align 16 .b8 __cudaretf__Z25PremultiplyLinearizePixel8PixelRGB[16]) _Z25PremultiplyLinearizePixel8PixelRGB (.param .align 16 .b8 __cudaparmf1__Z25PremultiplyLinearizePixel8PixelRGB[16])
	{
	.reg .f32 %f<47>;
	.reg .pred %p<5>;
	.loc	4	252	0
$LDWbegin__Z25PremultiplyLinearizePixel8PixelRGB:
	ld.param.f32 	%f1, [__cudaparmf1__Z25PremultiplyLinearizePixel8PixelRGB+0];
	mov.f32 	%f2, %f1;
	ld.param.f32 	%f3, [__cudaparmf1__Z25PremultiplyLinearizePixel8PixelRGB+4];
	mov.f32 	%f4, %f3;
	ld.param.f32 	%f5, [__cudaparmf1__Z25PremultiplyLinearizePixel8PixelRGB+8];
	mov.f32 	%f6, %f5;
	ld.param.f32 	%f7, [__cudaparmf1__Z25PremultiplyLinearizePixel8PixelRGB+12];
	mov.f32 	%f8, %f7;
	.loc	4	254	0
	cvt.ftz.sat.f32.f32 	%f9, %f8;
	.loc	4	255	0
	mov.f32 	%f10, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p1, %f2, %f10;
	@!%p1 bra 	$Lt_16_4098;
	.loc	4	234	0
	neg.ftz.f32 	%f11, %f2;
	lg2.approx.ftz.f32 	%f12, %f11;
	mov.f32 	%f13, 0f400ccccd;    	// 2.2
	mul.ftz.f32 	%f14, %f12, %f13;
	ex2.approx.ftz.f32 	%f15, %f14;
	neg.ftz.f32 	%f16, %f15;
	bra.uni 	$LDWendi___log2f_193_5;
$Lt_16_4098:
	.loc	4	236	0
	lg2.approx.ftz.f32 	%f17, %f2;
	mov.f32 	%f18, 0f400ccccd;    	// 2.2
	mul.ftz.f32 	%f19, %f17, %f18;
	ex2.approx.ftz.f32 	%f16, %f19;
$LDWendi___log2f_193_5:
	.loc	4	256	0
	mov.f32 	%f20, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p2, %f4, %f20;
	@!%p2 bra 	$Lt_16_4610;
	.loc	4	234	0
	neg.ftz.f32 	%f21, %f4;
	lg2.approx.ftz.f32 	%f22, %f21;
	mov.f32 	%f23, 0f400ccccd;    	// 2.2
	mul.ftz.f32 	%f24, %f22, %f23;
	ex2.approx.ftz.f32 	%f25, %f24;
	neg.ftz.f32 	%f26, %f25;
	bra.uni 	$LDWendi___log2f_193_3;
$Lt_16_4610:
	.loc	4	236	0
	lg2.approx.ftz.f32 	%f27, %f4;
	mov.f32 	%f28, 0f400ccccd;    	// 2.2
	mul.ftz.f32 	%f29, %f27, %f28;
	ex2.approx.ftz.f32 	%f26, %f29;
$LDWendi___log2f_193_3:
	.loc	4	257	0
	mov.f32 	%f30, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p3, %f6, %f30;
	@!%p3 bra 	$Lt_16_5122;
	.loc	4	234	0
	neg.ftz.f32 	%f31, %f6;
	lg2.approx.ftz.f32 	%f32, %f31;
	mov.f32 	%f33, 0f400ccccd;    	// 2.2
	mul.ftz.f32 	%f34, %f32, %f33;
	ex2.approx.ftz.f32 	%f35, %f34;
	neg.ftz.f32 	%f36, %f35;
	bra.uni 	$LDWendi___log2f_193_1;
$Lt_16_5122:
	.loc	4	236	0
	lg2.approx.ftz.f32 	%f37, %f6;
	mov.f32 	%f38, 0f400ccccd;    	// 2.2
	mul.ftz.f32 	%f39, %f37, %f38;
	ex2.approx.ftz.f32 	%f36, %f39;
$LDWendi___log2f_193_1:
	.loc	4	259	0
	mul.ftz.f32 	%f40, %f36, %f9;
	mul.ftz.f32 	%f41, %f26, %f9;
	mul.ftz.f32 	%f42, %f16, %f9;
	st.param.f32 	[__cudaretf__Z25PremultiplyLinearizePixel8PixelRGB+0], %f42;
	mov.f32 	%f43, %f41;
	st.param.f32 	[__cudaretf__Z25PremultiplyLinearizePixel8PixelRGB+4], %f43;
	mov.f32 	%f44, %f40;
	st.param.f32 	[__cudaretf__Z25PremultiplyLinearizePixel8PixelRGB+8], %f44;
	mov.f32 	%f45, %f9;
	st.param.f32 	[__cudaretf__Z25PremultiplyLinearizePixel8PixelRGB+12], %f45;
	ret;
$LDWend__Z25PremultiplyLinearizePixel8PixelRGB:
	} // _Z25PremultiplyLinearizePixel8PixelRGB

	.visible .func (.param .align 16 .b8 __cudaretf__Z29UnpremultiplyUnlinearizePixel8PixelRGB[16]) _Z29UnpremultiplyUnlinearizePixel8PixelRGB (.param .align 16 .b8 __cudaparmf1__Z29UnpremultiplyUnlinearizePixel8PixelRGB[16])
	{
	.reg .f32 %f<53>;
	.reg .pred %p<6>;
	.loc	4	263	0
$LDWbegin__Z29UnpremultiplyUnlinearizePixel8PixelRGB:
	ld.param.f32 	%f1, [__cudaparmf1__Z29UnpremultiplyUnlinearizePixel8PixelRGB+0];
	mov.f32 	%f2, %f1;
	ld.param.f32 	%f3, [__cudaparmf1__Z29UnpremultiplyUnlinearizePixel8PixelRGB+4];
	mov.f32 	%f4, %f3;
	ld.param.f32 	%f5, [__cudaparmf1__Z29UnpremultiplyUnlinearizePixel8PixelRGB+8];
	mov.f32 	%f6, %f5;
	ld.param.f32 	%f7, [__cudaparmf1__Z29UnpremultiplyUnlinearizePixel8PixelRGB+12];
	mov.f32 	%f8, %f7;
	.loc	4	208	0
	cvt.ftz.sat.f32.f32 	%f9, %f8;
	mov.f32 	%f10, %f9;
	mov.f32 	%f11, 0fb70637bd;    	// -8e-006
	add.ftz.f32 	%f12, %f9, %f11;
	mov.f32 	%f13, 0f00000000;    	// 0
	setp.le.ftz.f32 	%p1, %f12, %f13;
	@%p1 bra 	$Lt_17_5122;
	.loc	4	213	0
	rcp.approx.ftz.f32 	%f14, %f9;
	mul.ftz.f32 	%f15, %f14, %f6;
	.loc	4	214	0
	mul.ftz.f32 	%f16, %f14, %f4;
	.loc	4	215	0
	mul.ftz.f32 	%f17, %f14, %f2;
	bra.uni 	$Lt_17_4866;
$Lt_17_5122:
	.loc	4	219	0
	mov.f32 	%f15, 0f00000000;    	// 0
	mov.f32 	%f16, 0f00000000;    	// 0
	mov.f32 	%f17, 0f00000000;    	// 0
	mov.f32 	%f10, 0f00000000;    	// 0
$Lt_17_4866:
	.loc	4	266	0
	mov.f32 	%f18, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p2, %f17, %f18;
	@!%p2 bra 	$Lt_17_5378;
	.loc	4	242	0
	neg.ftz.f32 	%f19, %f17;
	lg2.approx.ftz.f32 	%f20, %f19;
	mov.f32 	%f21, 0f3ee8ba2e;    	// 0.454545
	mul.ftz.f32 	%f22, %f20, %f21;
	ex2.approx.ftz.f32 	%f23, %f22;
	neg.ftz.f32 	%f24, %f23;
	bra.uni 	$LDWendi___log2f_194_5;
$Lt_17_5378:
	.loc	4	244	0
	lg2.approx.ftz.f32 	%f25, %f17;
	mov.f32 	%f26, 0f3ee8ba2e;    	// 0.454545
	mul.ftz.f32 	%f27, %f25, %f26;
	ex2.approx.ftz.f32 	%f24, %f27;
$LDWendi___log2f_194_5:
	.loc	4	267	0
	mov.f32 	%f28, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p3, %f16, %f28;
	@!%p3 bra 	$Lt_17_5890;
	.loc	4	242	0
	neg.ftz.f32 	%f29, %f16;
	lg2.approx.ftz.f32 	%f30, %f29;
	mov.f32 	%f31, 0f3ee8ba2e;    	// 0.454545
	mul.ftz.f32 	%f32, %f30, %f31;
	ex2.approx.ftz.f32 	%f33, %f32;
	neg.ftz.f32 	%f34, %f33;
	bra.uni 	$LDWendi___log2f_194_3;
$Lt_17_5890:
	.loc	4	244	0
	lg2.approx.ftz.f32 	%f35, %f16;
	mov.f32 	%f36, 0f3ee8ba2e;    	// 0.454545
	mul.ftz.f32 	%f37, %f35, %f36;
	ex2.approx.ftz.f32 	%f34, %f37;
$LDWendi___log2f_194_3:
	.loc	4	268	0
	mov.f32 	%f38, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p4, %f15, %f38;
	@!%p4 bra 	$Lt_17_6402;
	.loc	4	242	0
	neg.ftz.f32 	%f39, %f15;
	lg2.approx.ftz.f32 	%f40, %f39;
	mov.f32 	%f41, 0f3ee8ba2e;    	// 0.454545
	mul.ftz.f32 	%f42, %f40, %f41;
	ex2.approx.ftz.f32 	%f43, %f42;
	neg.ftz.f32 	%f44, %f43;
	bra.uni 	$LDWendi___log2f_194_1;
$Lt_17_6402:
	.loc	4	244	0
	lg2.approx.ftz.f32 	%f45, %f15;
	mov.f32 	%f46, 0f3ee8ba2e;    	// 0.454545
	mul.ftz.f32 	%f47, %f45, %f46;
	ex2.approx.ftz.f32 	%f44, %f47;
$LDWendi___log2f_194_1:
	.loc	4	269	0
	mov.f32 	%f48, %f24;
	st.param.f32 	[__cudaretf__Z29UnpremultiplyUnlinearizePixel8PixelRGB+0], %f48;
	mov.f32 	%f49, %f34;
	st.param.f32 	[__cudaretf__Z29UnpremultiplyUnlinearizePixel8PixelRGB+4], %f49;
	mov.f32 	%f50, %f44;
	st.param.f32 	[__cudaretf__Z29UnpremultiplyUnlinearizePixel8PixelRGB+8], %f50;
	mov.f32 	%f51, %f10;
	st.param.f32 	[__cudaretf__Z29UnpremultiplyUnlinearizePixel8PixelRGB+12], %f51;
	ret;
$LDWend__Z29UnpremultiplyUnlinearizePixel8PixelRGB:
	} // _Z29UnpremultiplyUnlinearizePixel8PixelRGB

	.visible .func (.param .align 16 .b8 __cudaretf__Z20PremultiplyLinearize6float4[16]) _Z20PremultiplyLinearize6float4 (.param .align 16 .b8 __cudaparmf1__Z20PremultiplyLinearize6float4[16])
	{
	.reg .f32 %f<47>;
	.reg .pred %p<5>;
	.loc	4	277	0
$LDWbegin__Z20PremultiplyLinearize6float4:
	ld.param.f32 	%f1, [__cudaparmf1__Z20PremultiplyLinearize6float4+0];
	mov.f32 	%f2, %f1;
	ld.param.f32 	%f3, [__cudaparmf1__Z20PremultiplyLinearize6float4+4];
	mov.f32 	%f4, %f3;
	ld.param.f32 	%f5, [__cudaparmf1__Z20PremultiplyLinearize6float4+8];
	mov.f32 	%f6, %f5;
	ld.param.f32 	%f7, [__cudaparmf1__Z20PremultiplyLinearize6float4+12];
	mov.f32 	%f8, %f7;
	.loc	4	254	0
	cvt.ftz.sat.f32.f32 	%f9, %f8;
	.loc	4	255	0
	mov.f32 	%f10, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p1, %f2, %f10;
	@!%p1 bra 	$Lt_18_4098;
	.loc	4	234	0
	neg.ftz.f32 	%f11, %f2;
	lg2.approx.ftz.f32 	%f12, %f11;
	mov.f32 	%f13, 0f400ccccd;    	// 2.2
	mul.ftz.f32 	%f14, %f12, %f13;
	ex2.approx.ftz.f32 	%f15, %f14;
	neg.ftz.f32 	%f16, %f15;
	bra.uni 	$LDWendi___log2f_195_5;
$Lt_18_4098:
	.loc	4	236	0
	lg2.approx.ftz.f32 	%f17, %f2;
	mov.f32 	%f18, 0f400ccccd;    	// 2.2
	mul.ftz.f32 	%f19, %f17, %f18;
	ex2.approx.ftz.f32 	%f16, %f19;
$LDWendi___log2f_195_5:
	.loc	4	256	0
	mov.f32 	%f20, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p2, %f4, %f20;
	@!%p2 bra 	$Lt_18_4610;
	.loc	4	234	0
	neg.ftz.f32 	%f21, %f4;
	lg2.approx.ftz.f32 	%f22, %f21;
	mov.f32 	%f23, 0f400ccccd;    	// 2.2
	mul.ftz.f32 	%f24, %f22, %f23;
	ex2.approx.ftz.f32 	%f25, %f24;
	neg.ftz.f32 	%f26, %f25;
	bra.uni 	$LDWendi___log2f_195_3;
$Lt_18_4610:
	.loc	4	236	0
	lg2.approx.ftz.f32 	%f27, %f4;
	mov.f32 	%f28, 0f400ccccd;    	// 2.2
	mul.ftz.f32 	%f29, %f27, %f28;
	ex2.approx.ftz.f32 	%f26, %f29;
$LDWendi___log2f_195_3:
	.loc	4	257	0
	mov.f32 	%f30, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p3, %f6, %f30;
	@!%p3 bra 	$Lt_18_5122;
	.loc	4	234	0
	neg.ftz.f32 	%f31, %f6;
	lg2.approx.ftz.f32 	%f32, %f31;
	mov.f32 	%f33, 0f400ccccd;    	// 2.2
	mul.ftz.f32 	%f34, %f32, %f33;
	ex2.approx.ftz.f32 	%f35, %f34;
	neg.ftz.f32 	%f36, %f35;
	bra.uni 	$LDWendi___log2f_195_1;
$Lt_18_5122:
	.loc	4	236	0
	lg2.approx.ftz.f32 	%f37, %f6;
	mov.f32 	%f38, 0f400ccccd;    	// 2.2
	mul.ftz.f32 	%f39, %f37, %f38;
	ex2.approx.ftz.f32 	%f36, %f39;
$LDWendi___log2f_195_1:
	.loc	4	259	0
	mul.ftz.f32 	%f40, %f36, %f9;
	mul.ftz.f32 	%f41, %f26, %f9;
	.loc	4	278	0
	mul.ftz.f32 	%f42, %f16, %f9;
	st.param.f32 	[__cudaretf__Z20PremultiplyLinearize6float4+0], %f42;
	mov.f32 	%f43, %f41;
	st.param.f32 	[__cudaretf__Z20PremultiplyLinearize6float4+4], %f43;
	mov.f32 	%f44, %f40;
	st.param.f32 	[__cudaretf__Z20PremultiplyLinearize6float4+8], %f44;
	mov.f32 	%f45, %f9;
	st.param.f32 	[__cudaretf__Z20PremultiplyLinearize6float4+12], %f45;
	ret;
$LDWend__Z20PremultiplyLinearize6float4:
	} // _Z20PremultiplyLinearize6float4

	.visible .func (.param .align 16 .b8 __cudaretf__Z24UnpremultiplyUnlinearize6float4[16]) _Z24UnpremultiplyUnlinearize6float4 (.param .align 16 .b8 __cudaparmf1__Z24UnpremultiplyUnlinearize6float4[16])
	{
	.reg .f32 %f<53>;
	.reg .pred %p<6>;
	.loc	4	284	0
$LDWbegin__Z24UnpremultiplyUnlinearize6float4:
	ld.param.f32 	%f1, [__cudaparmf1__Z24UnpremultiplyUnlinearize6float4+0];
	mov.f32 	%f2, %f1;
	ld.param.f32 	%f3, [__cudaparmf1__Z24UnpremultiplyUnlinearize6float4+4];
	mov.f32 	%f4, %f3;
	ld.param.f32 	%f5, [__cudaparmf1__Z24UnpremultiplyUnlinearize6float4+8];
	mov.f32 	%f6, %f5;
	ld.param.f32 	%f7, [__cudaparmf1__Z24UnpremultiplyUnlinearize6float4+12];
	mov.f32 	%f8, %f7;
	.loc	4	208	0
	cvt.ftz.sat.f32.f32 	%f9, %f8;
	mov.f32 	%f10, %f9;
	mov.f32 	%f11, 0fb70637bd;    	// -8e-006
	add.ftz.f32 	%f12, %f9, %f11;
	mov.f32 	%f13, 0f00000000;    	// 0
	setp.le.ftz.f32 	%p1, %f12, %f13;
	@%p1 bra 	$Lt_19_5122;
	.loc	4	213	0
	rcp.approx.ftz.f32 	%f14, %f9;
	mul.ftz.f32 	%f15, %f14, %f6;
	.loc	4	214	0
	mul.ftz.f32 	%f16, %f14, %f4;
	.loc	4	215	0
	mul.ftz.f32 	%f17, %f14, %f2;
	bra.uni 	$Lt_19_4866;
$Lt_19_5122:
	.loc	4	219	0
	mov.f32 	%f15, 0f00000000;    	// 0
	mov.f32 	%f16, 0f00000000;    	// 0
	mov.f32 	%f17, 0f00000000;    	// 0
	mov.f32 	%f10, 0f00000000;    	// 0
$Lt_19_4866:
	.loc	4	266	0
	mov.f32 	%f18, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p2, %f17, %f18;
	@!%p2 bra 	$Lt_19_5378;
	.loc	4	242	0
	neg.ftz.f32 	%f19, %f17;
	lg2.approx.ftz.f32 	%f20, %f19;
	mov.f32 	%f21, 0f3ee8ba2e;    	// 0.454545
	mul.ftz.f32 	%f22, %f20, %f21;
	ex2.approx.ftz.f32 	%f23, %f22;
	neg.ftz.f32 	%f24, %f23;
	bra.uni 	$LDWendi___log2f_196_5;
$Lt_19_5378:
	.loc	4	244	0
	lg2.approx.ftz.f32 	%f25, %f17;
	mov.f32 	%f26, 0f3ee8ba2e;    	// 0.454545
	mul.ftz.f32 	%f27, %f25, %f26;
	ex2.approx.ftz.f32 	%f24, %f27;
$LDWendi___log2f_196_5:
	.loc	4	267	0
	mov.f32 	%f28, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p3, %f16, %f28;
	@!%p3 bra 	$Lt_19_5890;
	.loc	4	242	0
	neg.ftz.f32 	%f29, %f16;
	lg2.approx.ftz.f32 	%f30, %f29;
	mov.f32 	%f31, 0f3ee8ba2e;    	// 0.454545
	mul.ftz.f32 	%f32, %f30, %f31;
	ex2.approx.ftz.f32 	%f33, %f32;
	neg.ftz.f32 	%f34, %f33;
	bra.uni 	$LDWendi___log2f_196_3;
$Lt_19_5890:
	.loc	4	244	0
	lg2.approx.ftz.f32 	%f35, %f16;
	mov.f32 	%f36, 0f3ee8ba2e;    	// 0.454545
	mul.ftz.f32 	%f37, %f35, %f36;
	ex2.approx.ftz.f32 	%f34, %f37;
$LDWendi___log2f_196_3:
	.loc	4	268	0
	mov.f32 	%f38, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p4, %f15, %f38;
	@!%p4 bra 	$Lt_19_6402;
	.loc	4	242	0
	neg.ftz.f32 	%f39, %f15;
	lg2.approx.ftz.f32 	%f40, %f39;
	mov.f32 	%f41, 0f3ee8ba2e;    	// 0.454545
	mul.ftz.f32 	%f42, %f40, %f41;
	ex2.approx.ftz.f32 	%f43, %f42;
	neg.ftz.f32 	%f44, %f43;
	bra.uni 	$LDWendi___log2f_196_1;
$Lt_19_6402:
	.loc	4	244	0
	lg2.approx.ftz.f32 	%f45, %f15;
	mov.f32 	%f46, 0f3ee8ba2e;    	// 0.454545
	mul.ftz.f32 	%f47, %f45, %f46;
	ex2.approx.ftz.f32 	%f44, %f47;
$LDWendi___log2f_196_1:
	.loc	4	285	0
	mov.f32 	%f48, %f24;
	st.param.f32 	[__cudaretf__Z24UnpremultiplyUnlinearize6float4+0], %f48;
	mov.f32 	%f49, %f34;
	st.param.f32 	[__cudaretf__Z24UnpremultiplyUnlinearize6float4+4], %f49;
	mov.f32 	%f50, %f44;
	st.param.f32 	[__cudaretf__Z24UnpremultiplyUnlinearize6float4+8], %f50;
	mov.f32 	%f51, %f10;
	st.param.f32 	[__cudaretf__Z24UnpremultiplyUnlinearize6float4+12], %f51;
	ret;
$LDWend__Z24UnpremultiplyUnlinearize6float4:
	} // _Z24UnpremultiplyUnlinearize6float4

	.visible .func (.param .align 16 .b8 __cudaretf__Z18SwapComponentOrderI6float4ET_RKS1_[16]) _Z18SwapComponentOrderI6float4ET_RKS1_ (.param .u64 __cudaparmf1__Z18SwapComponentOrderI6float4ET_RKS1_)
	{
	.reg .u64 %rd<4>;
	.reg .f32 %f<6>;
	.loc	22	264	0
$LDWbegin__Z18SwapComponentOrderI6float4ET_RKS1_:
	ld.param.u64 	%rd1, [__cudaparmf1__Z18SwapComponentOrderI6float4ET_RKS1_];
	mov.s64 	%rd2, %rd1;
	.loc	22	270	0
	ld.f32 	%f1, [%rd2+12];
	st.param.f32 	[__cudaretf__Z18SwapComponentOrderI6float4ET_RKS1_+0], %f1;
	ld.f32 	%f2, [%rd2+8];
	st.param.f32 	[__cudaretf__Z18SwapComponentOrderI6float4ET_RKS1_+4], %f2;
	ld.f32 	%f3, [%rd2+4];
	st.param.f32 	[__cudaretf__Z18SwapComponentOrderI6float4ET_RKS1_+8], %f3;
	ld.f32 	%f4, [%rd2+0];
	st.param.f32 	[__cudaretf__Z18SwapComponentOrderI6float4ET_RKS1_+12], %f4;
	ret;
$LDWend__Z18SwapComponentOrderI6float4ET_RKS1_:
	} // _Z18SwapComponentOrderI6float4ET_RKS1_

	.visible .func _Z14ConvertUVToHueffRf (.param .f32 __cudaparmf1__Z14ConvertUVToHueffRf, .param .f32 __cudaparmf2__Z14ConvertUVToHueffRf, .param .u64 __cudaparmf3__Z14ConvertUVToHueffRf)
	{
	.reg .u32 %r<30>;
	.reg .u64 %rd<4>;
	.reg .f32 %f<57>;
	.reg .pred %p<10>;
	.loc	6	45	0
$LDWbegin__Z14ConvertUVToHueffRf:
	ld.param.f32 	%f1, [__cudaparmf1__Z14ConvertUVToHueffRf];
	mov.f32 	%f2, %f1;
	ld.param.f32 	%f3, [__cudaparmf2__Z14ConvertUVToHueffRf];
	mov.f32 	%f4, %f3;
	ld.param.u64 	%rd1, [__cudaparmf3__Z14ConvertUVToHueffRf];
	mov.s64 	%rd2, %rd1;
	.loc	6	60	0
	mov.f32 	%f5, 0f3f728f61;     	// 0.9475
	mul.ftz.f32 	%f6, %f4, %f5;
	mov.f32 	%f7, 0f3ea3b6e9;     	// 0.319755
	mul.ftz.f32 	%f8, %f4, %f7;
	mov.f32 	%f9, 0fbea3b6e9;     	// -0.319755
	fma.rn.ftz.f32 	%f10, %f9, %f2, %f6;
	mov.f32 	%f11, 0fbf728f61;    	// -0.9475
	mul.ftz.f32 	%f12, %f11, %f2;
	sub.ftz.f32 	%f13, %f12, %f8;
	abs.ftz.f32 	%f14, %f10;
	abs.ftz.f32 	%f15, %f13;
	mov.b32 	%r1, %f13;
	and.b32 	%r2, %r1, -2147483648;
	mov.f32 	%f16, 0f00000000;    	// 0
	set.eq.ftz.u32.f32 	%r3, %f14, %f16;
	neg.s32 	%r4, %r3;
	mov.f32 	%f17, 0f00000000;    	// 0
	set.eq.ftz.u32.f32 	%r5, %f15, %f17;
	neg.s32 	%r6, %r5;
	and.b32 	%r7, %r4, %r6;
	mov.u32 	%r8, 0;
	setp.eq.s32 	%p1, %r7, %r8;
	@%p1 bra 	$Lt_21_7170;
	.loc	26	1842	0
	mov.s32 	%r9, 1078530011;
	mov.s32 	%r10, 0;
	mov.b32 	%r11, %f10;
	mov.s32 	%r12, 0;
	setp.lt.s32 	%p2, %r11, %r12;
	selp.s32 	%r13, %r9, %r10, %p2;
	or.b32 	%r14, %r13, %r2;
	mov.b32 	%f18, %r14;
	bra.uni 	$Lt_21_6914;
$Lt_21_7170:
	mov.f32 	%f19, 0f7f800000;    	// 1.#INF
	set.eq.ftz.u32.f32 	%r15, %f14, %f19;
	neg.s32 	%r16, %r15;
	mov.f32 	%f20, 0f7f800000;    	// 1.#INF
	set.eq.ftz.u32.f32 	%r17, %f15, %f20;
	neg.s32 	%r18, %r17;
	and.b32 	%r19, %r16, %r18;
	mov.u32 	%r20, 0;
	setp.eq.s32 	%p3, %r19, %r20;
	@%p3 bra 	$Lt_21_7682;
	.loc	26	1845	0
	mov.s32 	%r21, 1075235812;
	mov.s32 	%r22, 1061752795;
	mov.b32 	%r23, %f10;
	mov.s32 	%r24, 0;
	setp.lt.s32 	%p4, %r23, %r24;
	selp.s32 	%r25, %r21, %r22, %p4;
	or.b32 	%r26, %r25, %r2;
	mov.b32 	%f18, %r26;
	bra.uni 	$Lt_21_7426;
$Lt_21_7682:
	.loc	26	1207	0
	min.ftz.f32 	%f21, %f14, %f15;
	max.ftz.f32 	%f22, %f14, %f15;
	div.full.ftz.f32 	%f23, %f21, %f22;
	mul.ftz.f32 	%f24, %f23, %f23;
	mov.f32 	%f25, 0fbf52c7ea;    	// -0.823363
	mov.f32 	%f26, 0fc0b59883;    	// -5.67487
	fma.rn.ftz.f32 	%f27, %f24, %f25, %f26;
	mov.f32 	%f28, 0fc0d21907;    	// -6.56556
	fma.rn.ftz.f32 	%f29, %f27, %f24, %f28;
	mul.ftz.f32 	%f30, %f24, %f29;
	mul.ftz.f32 	%f31, %f23, %f30;
	.loc	26	1211	0
	mov.f32 	%f32, 0f41355dc0;    	// 11.3354
	add.ftz.f32 	%f33, %f24, %f32;
	mov.f32 	%f34, 0f41e6bd60;    	// 28.8425
	fma.rn.ftz.f32 	%f35, %f33, %f24, %f34;
	mov.f32 	%f36, 0f419d92c8;    	// 19.6967
	fma.rn.ftz.f32 	%f37, %f35, %f24, %f36;
	rcp.approx.ftz.f32 	%f38, %f37;
	.loc	26	1856	0
	setp.lt.ftz.f32 	%p5, %f14, %f15;
	fma.rn.ftz.f32 	%f39, %f31, %f38, %f23;
	mov.f32 	%f40, 0f3fc90fdb;    	// 1.5708
	sub.ftz.f32 	%f41, %f40, %f39;
	selp.f32 	%f42, %f41, %f39, %p5;
	mov.f32 	%f43, 0f40490fdb;    	// 3.14159
	sub.ftz.f32 	%f44, %f43, %f42;
	mov.f32 	%f45, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p6, %f10, %f45;
	selp.f32 	%f46, %f44, %f42, %p6;
	mov.b32 	%r27, %f46;
	or.b32 	%r28, %r2, %r27;
	mov.b32 	%f47, %r28;
	add.ftz.f32 	%f48, %f10, %f13;
	mov.f32 	%f49, 0f7f800000;    	// 1.#INF
	setp.le.ftz.f32 	%p7, %f48, %f49;
	selp.f32 	%f18, %f47, %f48, %p7;
$Lt_21_7426:
$Lt_21_6914:
	.loc	26	1863	0
	mov.f32 	%f50, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p8, %f18, %f50;
	mov.f32 	%f51, 0f40c90fdb;    	// 6.28319
	add.ftz.f32 	%f52, %f18, %f51;
	selp.f32 	%f53, %f52, %f18, %p8;
	st.f32 	[%rd2+0], %f53;
	.loc	6	75	0
	mov.f32 	%f54, 0f3e22f983;    	// 0.159155
	mul.ftz.f32 	%f55, %f53, %f54;
	st.f32 	[%rd2+0], %f55;
	.loc	6	76	0
	ret;
$LDWend__Z14ConvertUVToHueffRf:
	} // _Z14ConvertUVToHueffRf

	.visible .func _Z15ConvertYUVToHSL8PixelYUVR8PixelHSL (.param .align 16 .b8 __cudaparmf1__Z15ConvertYUVToHSL8PixelYUVR8PixelHSL[16], .param .u64 __cudaparmf2__Z15ConvertYUVToHSL8PixelYUVR8PixelHSL)
	{
	.reg .u32 %r<30>;
	.reg .u64 %rd<4>;
	.reg .f32 %f<77>;
	.reg .pred %p<10>;
	.loc	6	81	0
$LDWbegin__Z15ConvertYUVToHSL8PixelYUVR8PixelHSL:
	ld.param.f32 	%f1, [__cudaparmf1__Z15ConvertYUVToHSL8PixelYUVR8PixelHSL+4];
	mov.f32 	%f2, %f1;
	ld.param.f32 	%f3, [__cudaparmf1__Z15ConvertYUVToHSL8PixelYUVR8PixelHSL+8];
	mov.f32 	%f4, %f3;
	ld.param.f32 	%f5, [__cudaparmf1__Z15ConvertYUVToHSL8PixelYUVR8PixelHSL+12];
	mov.f32 	%f6, %f5;
	ld.param.u64 	%rd1, [__cudaparmf2__Z15ConvertYUVToHSL8PixelYUVR8PixelHSL];
	mov.s64 	%rd2, %rd1;
	.loc	6	60	0
	mov.f32 	%f7, 0f3f728f61;     	// 0.9475
	mul.ftz.f32 	%f8, %f2, %f7;
	mov.f32 	%f9, 0f3ea3b6e9;     	// 0.319755
	mul.ftz.f32 	%f10, %f2, %f9;
	mov.f32 	%f11, 0fbea3b6e9;    	// -0.319755
	fma.rn.ftz.f32 	%f12, %f11, %f4, %f8;
	mov.f32 	%f13, 0fbf728f61;    	// -0.9475
	mul.ftz.f32 	%f14, %f13, %f4;
	sub.ftz.f32 	%f15, %f14, %f10;
	abs.ftz.f32 	%f16, %f12;
	abs.ftz.f32 	%f17, %f15;
	mov.b32 	%r1, %f15;
	and.b32 	%r2, %r1, -2147483648;
	mov.f32 	%f18, 0f00000000;    	// 0
	set.eq.ftz.u32.f32 	%r3, %f16, %f18;
	neg.s32 	%r4, %r3;
	mov.f32 	%f19, 0f00000000;    	// 0
	set.eq.ftz.u32.f32 	%r5, %f17, %f19;
	neg.s32 	%r6, %r5;
	and.b32 	%r7, %r4, %r6;
	mov.u32 	%r8, 0;
	setp.eq.s32 	%p1, %r7, %r8;
	@%p1 bra 	$Lt_22_10242;
	.loc	26	1842	0
	mov.s32 	%r9, 1078530011;
	mov.s32 	%r10, 0;
	mov.b32 	%r11, %f12;
	mov.s32 	%r12, 0;
	setp.lt.s32 	%p2, %r11, %r12;
	selp.s32 	%r13, %r9, %r10, %p2;
	or.b32 	%r14, %r13, %r2;
	mov.b32 	%f20, %r14;
	bra.uni 	$Lt_22_9986;
$Lt_22_10242:
	mov.f32 	%f21, 0f7f800000;    	// 1.#INF
	set.eq.ftz.u32.f32 	%r15, %f16, %f21;
	neg.s32 	%r16, %r15;
	mov.f32 	%f22, 0f7f800000;    	// 1.#INF
	set.eq.ftz.u32.f32 	%r17, %f17, %f22;
	neg.s32 	%r18, %r17;
	and.b32 	%r19, %r16, %r18;
	mov.u32 	%r20, 0;
	setp.eq.s32 	%p3, %r19, %r20;
	@%p3 bra 	$Lt_22_10754;
	.loc	26	1845	0
	mov.s32 	%r21, 1075235812;
	mov.s32 	%r22, 1061752795;
	mov.b32 	%r23, %f12;
	mov.s32 	%r24, 0;
	setp.lt.s32 	%p4, %r23, %r24;
	selp.s32 	%r25, %r21, %r22, %p4;
	or.b32 	%r26, %r25, %r2;
	mov.b32 	%f20, %r26;
	bra.uni 	$Lt_22_10498;
$Lt_22_10754:
	.loc	26	1207	0
	min.ftz.f32 	%f23, %f16, %f17;
	max.ftz.f32 	%f24, %f16, %f17;
	div.full.ftz.f32 	%f25, %f23, %f24;
	mul.ftz.f32 	%f26, %f25, %f25;
	mov.f32 	%f27, 0fbf52c7ea;    	// -0.823363
	mov.f32 	%f28, 0fc0b59883;    	// -5.67487
	fma.rn.ftz.f32 	%f29, %f26, %f27, %f28;
	mov.f32 	%f30, 0fc0d21907;    	// -6.56556
	fma.rn.ftz.f32 	%f31, %f29, %f26, %f30;
	mul.ftz.f32 	%f32, %f26, %f31;
	mul.ftz.f32 	%f33, %f25, %f32;
	.loc	26	1211	0
	mov.f32 	%f34, 0f41355dc0;    	// 11.3354
	add.ftz.f32 	%f35, %f26, %f34;
	mov.f32 	%f36, 0f41e6bd60;    	// 28.8425
	fma.rn.ftz.f32 	%f37, %f35, %f26, %f36;
	mov.f32 	%f38, 0f419d92c8;    	// 19.6967
	fma.rn.ftz.f32 	%f39, %f37, %f26, %f38;
	rcp.approx.ftz.f32 	%f40, %f39;
	.loc	26	1856	0
	setp.lt.ftz.f32 	%p5, %f16, %f17;
	fma.rn.ftz.f32 	%f41, %f33, %f40, %f25;
	mov.f32 	%f42, 0f3fc90fdb;    	// 1.5708
	sub.ftz.f32 	%f43, %f42, %f41;
	selp.f32 	%f44, %f43, %f41, %p5;
	mov.f32 	%f45, 0f40490fdb;    	// 3.14159
	sub.ftz.f32 	%f46, %f45, %f44;
	mov.f32 	%f47, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p6, %f12, %f47;
	selp.f32 	%f48, %f46, %f44, %p6;
	mov.b32 	%r27, %f48;
	or.b32 	%r28, %r2, %r27;
	mov.b32 	%f49, %r28;
	add.ftz.f32 	%f50, %f12, %f15;
	mov.f32 	%f51, 0f7f800000;    	// 1.#INF
	setp.le.ftz.f32 	%p7, %f50, %f51;
	selp.f32 	%f20, %f49, %f50, %p7;
$Lt_22_10498:
$Lt_22_9986:
	.loc	26	1863	0
	mov.f32 	%f52, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p8, %f20, %f52;
	mov.f32 	%f53, 0f40c90fdb;    	// 6.28319
	add.ftz.f32 	%f54, %f20, %f53;
	selp.f32 	%f55, %f54, %f20, %p8;
	st.f32 	[%rd2+0], %f55;
	.loc	6	75	0
	mov.f32 	%f56, 0f3e22f983;    	// 0.159155
	mul.ftz.f32 	%f57, %f55, %f56;
	st.f32 	[%rd2+0], %f57;
	.loc	6	89	0
	mul.ftz.f32 	%f58, %f4, %f4;
	fma.rn.ftz.f32 	%f59, %f2, %f2, %f58;
	sqrt.approx.ftz.f32 	%f60, %f59;
	mov.f32 	%f61, 0f3f3504f3;    	// 0.707107
	div.approx.ftz.f32 	%f62, %f60, %f61;
	.loc	6	93	0
	st.f32 	[%rd2+8], %f6;
	.loc	6	99	0
	ld.f32 	%f63, [%rd2+0];
	mov.f32 	%f64, 0f00000000;    	// 0
	max.ftz.f32 	%f65, %f63, %f64;
	mov.f32 	%f66, 0f3f800000;    	// 1
	min.ftz.f32 	%f67, %f65, %f66;
	st.v2.f32 	[%rd2+0], {%f67,%f62};
	.loc	6	100	0
	mov.f32 	%f68, 0f00000000;    	// 0
	max.ftz.f32 	%f69, %f62, %f68;
	mov.f32 	%f70, 0f3f800000;    	// 1
	min.ftz.f32 	%f71, %f69, %f70;
	st.f32 	[%rd2+4], %f71;
	.loc	6	101	0
	mov.f32 	%f72, 0f00000000;    	// 0
	max.ftz.f32 	%f73, %f6, %f72;
	mov.f32 	%f74, 0f3f800000;    	// 1
	min.ftz.f32 	%f75, %f73, %f74;
	st.f32 	[%rd2+8], %f75;
	.loc	6	102	0
	ret;
$LDWend__Z15ConvertYUVToHSL8PixelYUVR8PixelHSL:
	} // _Z15ConvertYUVToHSL8PixelYUVR8PixelHSL

	.visible .func (.param .f32 __cudaretf__Z5ClampIfET_S0_S0_S0_) _Z5ClampIfET_S0_S0_S0_ (.param .f32 __cudaparmf1__Z5ClampIfET_S0_S0_S0_, .param .f32 __cudaparmf2__Z5ClampIfET_S0_S0_S0_, .param .f32 __cudaparmf3__Z5ClampIfET_S0_S0_S0_)
	{
	.reg .f32 %f<10>;
	.loc	23	72	0
$LDWbegin__Z5ClampIfET_S0_S0_S0_:
	ld.param.f32 	%f1, [__cudaparmf1__Z5ClampIfET_S0_S0_S0_];
	mov.f32 	%f2, %f1;
	ld.param.f32 	%f3, [__cudaparmf2__Z5ClampIfET_S0_S0_S0_];
	mov.f32 	%f4, %f3;
	ld.param.f32 	%f5, [__cudaparmf3__Z5ClampIfET_S0_S0_S0_];
	mov.f32 	%f6, %f5;
	.loc	23	73	0
	max.ftz.f32 	%f7, %f2, %f4;
	min.ftz.f32 	%f8, %f6, %f7;
	st.param.f32 	[__cudaretf__Z5ClampIfET_S0_S0_S0_], %f8;
	ret;
$LDWend__Z5ClampIfET_S0_S0_S0_:
	} // _Z5ClampIfET_S0_S0_S0_

	.visible .func _Z15ConvertRGBtoHSL8PixelRGBR8PixelHSL (.param .align 16 .b8 __cudaparmf1__Z15ConvertRGBtoHSL8PixelRGBR8PixelHSL[16], .param .u64 __cudaparmf2__Z15ConvertRGBtoHSL8PixelRGBR8PixelHSL)
	{
	.reg .u32 %r<20>;
	.reg .u64 %rd<4>;
	.reg .f32 %f<34>;
	.reg .pred %p<12>;
	.loc	24	193	0
$LDWbegin__Z15ConvertRGBtoHSL8PixelRGBR8PixelHSL:
	ld.param.f32 	%f1, [__cudaparmf1__Z15ConvertRGBtoHSL8PixelRGBR8PixelHSL+0];
	mov.f32 	%f2, %f1;
	ld.param.f32 	%f3, [__cudaparmf1__Z15ConvertRGBtoHSL8PixelRGBR8PixelHSL+4];
	mov.f32 	%f4, %f3;
	ld.param.f32 	%f5, [__cudaparmf1__Z15ConvertRGBtoHSL8PixelRGBR8PixelHSL+8];
	mov.f32 	%f6, %f5;
	ld.param.u64 	%rd1, [__cudaparmf2__Z15ConvertRGBtoHSL8PixelRGBR8PixelHSL];
	mov.s64 	%rd2, %rd1;
	setp.le.ftz.f32 	%p1, %f2, %f6;
	setp.le.ftz.f32 	%p2, %f4, %f6;
	selp.s32 	%r1, 1, 0, %p1;
	selp.s32 	%r2, 1, 0, %p2;
	and.b32 	%r3, %r1, %r2;
	mov.u32 	%r4, 0;
	setp.eq.s32 	%p3, %r3, %r4;
	@%p3 bra 	$Lt_24_8706;
	.loc	24	214	0
	sub.ftz.f32 	%f7, %f4, %f2;
	st.f32 	[%rd2+0], %f7;
$Lt_24_8706:
	mov.f32 	%f8, %f6;
	setp.le.ftz.f32 	%p4, %f2, %f4;
	selp.s32 	%r5, 1, 0, %p4;
	set.ge.ftz.u32.f32 	%r6, %f4, %f6;
	neg.s32 	%r7, %r6;
	and.b32 	%r8, %r5, %r7;
	mov.u32 	%r9, 0;
	setp.eq.s32 	%p5, %r8, %r9;
	@%p5 bra 	$Lt_24_9474;
	.loc	24	219	0
	mov.f32 	%f8, %f4;
	.loc	24	220	0
	sub.ftz.f32 	%f9, %f2, %f6;
	st.f32 	[%rd2+0], %f9;
	mov.f32 	%f10, 0f40000000;    	// 2
	bra.uni 	$Lt_24_9218;
$Lt_24_9474:
	mov.f32 	%f10, 0f00000000;    	// 0
$Lt_24_9218:
	setp.ge.ftz.f32 	%p6, %f2, %f4;
	selp.s32 	%r10, 1, 0, %p6;
	set.ge.ftz.u32.f32 	%r11, %f2, %f6;
	neg.s32 	%r12, %r11;
	and.b32 	%r13, %r10, %r12;
	mov.u32 	%r14, 0;
	setp.eq.s32 	%p7, %r13, %r14;
	@%p7 bra 	$Lt_24_9730;
	.loc	24	225	0
	mov.f32 	%f8, %f2;
	.loc	24	226	0
	sub.ftz.f32 	%f11, %f6, %f4;
	st.f32 	[%rd2+0], %f11;
	mov.f32 	%f10, 0f40800000;    	// 4
$Lt_24_9730:
	.loc	24	242	0
	and.b32 	%r15, %r1, %r5;
	and.b32 	%r16, %r2, %r10;
	neg.s32 	%r17, %r16;
	slct.f32.s32 	%f12, %f6, %f4, %r17;
	neg.s32 	%r18, %r15;
	slct.f32.s32 	%f13, %f12, %f2, %r18;
	add.ftz.f32 	%f14, %f13, %f8;
	mov.f32 	%f15, 0f40000000;    	// 2
	div.approx.ftz.f32 	%f16, %f14, %f15;
	st.f32 	[%rd2+8], %f16;
	.loc	24	243	0
	mov.f32 	%f17, 0f00000000;    	// 0
	st.f32 	[%rd2+4], %f17;
	mov.f32 	%f18, 0f38d1b717;    	// 0.0001
	setp.gt.ftz.f32 	%p8, %f16, %f18;
	@!%p8 bra 	$Lt_24_10498;
	.loc	24	247	0
	sub.ftz.f32 	%f19, %f8, %f13;
	st.f32 	[%rd2+4], %f19;
	mov.f32 	%f20, 0f38d1b717;    	// 0.0001
	setp.gt.ftz.f32 	%p9, %f19, %f20;
	@!%p9 bra 	$Lt_24_11010;
	.loc	24	250	0
	ld.f32 	%f21, [%rd2+0];
	.loc	24	251	0
	div.approx.ftz.f32 	%f22, %f21, %f19;
	add.ftz.f32 	%f23, %f22, %f10;
	mov.f32 	%f24, %f23;
	st.f32 	[%rd2+0], %f23;
	mov.f32 	%f25, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p10, %f23, %f25;
	@!%p10 bra 	$Lt_24_11266;
	.loc	24	256	0
	mov.f32 	%f26, 0f40c00000;    	// 6
	add.ftz.f32 	%f24, %f23, %f26;
	st.f32 	[%rd2+0], %f24;
$Lt_24_11266:
	.loc	24	258	0
	mov.f32 	%f27, %f24;
	mov.f32 	%f28, 0f40c00000;    	// 6
	div.approx.ftz.f32 	%f29, %f27, %f28;
	st.f32 	[%rd2+0], %f29;
	bra.uni 	$Lt_24_10242;
$Lt_24_11010:
	.loc	24	262	0
	mov.f32 	%f30, 0f00000000;    	// 0
	st.f32 	[%rd2+4], %f30;
	.loc	24	263	0
	mov.f32 	%f31, 0f00000000;    	// 0
	st.f32 	[%rd2+0], %f31;
	bra.uni 	$Lt_24_10242;
$Lt_24_10498:
	.loc	24	268	0
	mov.f32 	%f32, 0f00000000;    	// 0
	st.f32 	[%rd2+0], %f32;
$Lt_24_10242:
	.loc	24	271	0
	ret;
$LDWend__Z15ConvertRGBtoHSL8PixelRGBR8PixelHSL:
	} // _Z15ConvertRGBtoHSL8PixelRGBR8PixelHSL

	.visible .func _Z15ConvertHSLtoRGB8PixelHSLR8PixelRGB (.param .align 16 .b8 __cudaparmf1__Z15ConvertHSLtoRGB8PixelHSLR8PixelRGB[16], .param .u64 __cudaparmf2__Z15ConvertHSLtoRGB8PixelHSLR8PixelRGB)
	{
	.reg .u32 %r<9>;
	.reg .u64 %rd<4>;
	.reg .f32 %f<62>;
	.reg .pred %p<10>;
	.loc	24	277	0
$LDWbegin__Z15ConvertHSLtoRGB8PixelHSLR8PixelRGB:
	ld.param.f32 	%f1, [__cudaparmf1__Z15ConvertHSLtoRGB8PixelHSLR8PixelRGB+0];
	mov.f32 	%f2, %f1;
	ld.param.f32 	%f3, [__cudaparmf1__Z15ConvertHSLtoRGB8PixelHSLR8PixelRGB+4];
	mov.f32 	%f4, %f3;
	ld.param.f32 	%f5, [__cudaparmf1__Z15ConvertHSLtoRGB8PixelHSLR8PixelRGB+8];
	mov.f32 	%f6, %f5;
	ld.param.u64 	%rd1, [__cudaparmf2__Z15ConvertHSLtoRGB8PixelHSLR8PixelRGB];
	mov.s64 	%rd2, %rd1;
	mov.f32 	%f7, 0f3f000000;     	// 0.5
	setp.le.ftz.f32 	%p1, %f6, %f7;
	@!%p1 bra 	$Lt_25_4354;
	.loc	24	287	0
	mov.f32 	%f8, 0f3f800000;     	// 1
	add.ftz.f32 	%f9, %f4, %f8;
	mul.ftz.f32 	%f10, %f6, %f9;
	bra.uni 	$Lt_25_4098;
$Lt_25_4354:
	add.ftz.f32 	%f11, %f4, %f6;
	mul.ftz.f32 	%f12, %f4, %f6;
	sub.ftz.f32 	%f10, %f11, %f12;
$Lt_25_4098:
	.loc	24	289	0
	mov.f32 	%f13, 0f00000000;    	// 0
	st.f32 	[%rd2+0], %f13;
	mov.f32 	%f14, 0f00000000;    	// 0
	st.f32 	[%rd2+4], %f14;
	mov.f32 	%f15, 0f00000000;    	// 0
	st.f32 	[%rd2+8], %f15;
	mov.f32 	%f16, %f10;
	mov.f32 	%f17, 0f00000000;    	// 0
	setp.gt.ftz.f32 	%p2, %f16, %f17;
	@!%p2 bra 	$Lt_25_514;
	.loc	24	306	0
	mov.f32 	%f18, 0f40c00000;    	// 6
	mul.ftz.f32 	%f19, %f2, %f18;
	cvt.rzi.ftz.s32.f32 	%r1, %f19;
	mov.u32 	%r2, 0;
	setp.eq.s32 	%p3, %r1, %r2;
	@%p3 bra 	$Lt_25_258;
	mov.u32 	%r3, 1;
	setp.eq.s32 	%p4, %r1, %r3;
	@%p4 bra 	$Lt_25_770;
	mov.u32 	%r4, 2;
	setp.eq.s32 	%p5, %r1, %r4;
	@%p5 bra 	$Lt_25_1026;
	mov.u32 	%r5, 3;
	setp.eq.s32 	%p6, %r1, %r5;
	@%p6 bra 	$Lt_25_1282;
	mov.u32 	%r6, 4;
	setp.eq.s32 	%p7, %r1, %r6;
	@%p7 bra 	$Lt_25_1538;
	mov.u32 	%r7, 5;
	setp.eq.s32 	%p8, %r1, %r7;
	@%p8 bra 	$Lt_25_1794;
	bra.uni 	$Lt_25_514;
$Lt_25_258:
	.loc	24	309	0
	st.f32 	[%rd2+8], %f16;
	.loc	24	310	0
	add.ftz.f32 	%f20, %f6, %f6;
	sub.ftz.f32 	%f21, %f20, %f16;
	sub.ftz.f32 	%f22, %f16, %f21;
	div.approx.ftz.f32 	%f23, %f22, %f16;
	mul.ftz.f32 	%f24, %f16, %f23;
	cvt.rn.f32.s32 	%f25, %r1;
	sub.ftz.f32 	%f26, %f19, %f25;
	fma.rn.ftz.f32 	%f27, %f24, %f26, %f21;
	st.f32 	[%rd2+4], %f27;
	.loc	24	311	0
	st.f32 	[%rd2+0], %f21;
	.loc	24	312	0
	bra.uni 	$Lt_25_514;
$Lt_25_770:
	.loc	24	314	0
	add.ftz.f32 	%f20, %f6, %f6;
	sub.ftz.f32 	%f21, %f20, %f16;
	sub.ftz.f32 	%f28, %f16, %f21;
	div.approx.ftz.f32 	%f29, %f28, %f16;
	mul.ftz.f32 	%f30, %f16, %f29;
	cvt.rn.f32.s32 	%f31, %r1;
	sub.ftz.f32 	%f32, %f19, %f31;
	mul.ftz.f32 	%f33, %f30, %f32;
	sub.ftz.f32 	%f34, %f16, %f33;
	st.f32 	[%rd2+8], %f34;
	.loc	24	315	0
	st.f32 	[%rd2+4], %f16;
	.loc	24	316	0
	st.f32 	[%rd2+0], %f21;
	.loc	24	317	0
	bra.uni 	$Lt_25_514;
$Lt_25_1026:
	.loc	24	319	0
	add.ftz.f32 	%f20, %f6, %f6;
	sub.ftz.f32 	%f21, %f20, %f16;
	st.f32 	[%rd2+8], %f21;
	.loc	24	320	0
	st.f32 	[%rd2+4], %f16;
	.loc	24	321	0
	sub.ftz.f32 	%f35, %f16, %f21;
	div.approx.ftz.f32 	%f36, %f35, %f16;
	mul.ftz.f32 	%f37, %f16, %f36;
	cvt.rn.f32.s32 	%f38, %r1;
	sub.ftz.f32 	%f39, %f19, %f38;
	fma.rn.ftz.f32 	%f40, %f37, %f39, %f21;
	st.f32 	[%rd2+0], %f40;
	.loc	24	322	0
	bra.uni 	$Lt_25_514;
$Lt_25_1282:
	.loc	24	324	0
	add.ftz.f32 	%f20, %f6, %f6;
	sub.ftz.f32 	%f21, %f20, %f16;
	st.f32 	[%rd2+8], %f21;
	.loc	24	325	0
	sub.ftz.f32 	%f41, %f16, %f21;
	div.approx.ftz.f32 	%f42, %f41, %f16;
	mul.ftz.f32 	%f43, %f16, %f42;
	cvt.rn.f32.s32 	%f44, %r1;
	sub.ftz.f32 	%f45, %f19, %f44;
	mul.ftz.f32 	%f46, %f43, %f45;
	sub.ftz.f32 	%f47, %f16, %f46;
	st.f32 	[%rd2+4], %f47;
	.loc	24	326	0
	st.f32 	[%rd2+0], %f16;
	.loc	24	327	0
	bra.uni 	$Lt_25_514;
$Lt_25_1538:
	.loc	24	329	0
	add.ftz.f32 	%f20, %f6, %f6;
	sub.ftz.f32 	%f21, %f20, %f16;
	sub.ftz.f32 	%f48, %f16, %f21;
	div.approx.ftz.f32 	%f49, %f48, %f16;
	mul.ftz.f32 	%f50, %f16, %f49;
	cvt.rn.f32.s32 	%f51, %r1;
	sub.ftz.f32 	%f52, %f19, %f51;
	fma.rn.ftz.f32 	%f53, %f50, %f52, %f21;
	st.f32 	[%rd2+8], %f53;
	.loc	24	330	0
	st.f32 	[%rd2+4], %f21;
	.loc	24	331	0
	st.f32 	[%rd2+0], %f16;
	.loc	24	332	0
	bra.uni 	$Lt_25_514;
$Lt_25_1794:
	.loc	24	334	0
	st.f32 	[%rd2+8], %f16;
	.loc	24	335	0
	add.ftz.f32 	%f20, %f6, %f6;
	sub.ftz.f32 	%f21, %f20, %f16;
	st.f32 	[%rd2+4], %f21;
	.loc	24	336	0
	sub.ftz.f32 	%f54, %f16, %f21;
	div.approx.ftz.f32 	%f55, %f54, %f16;
	mul.ftz.f32 	%f56, %f16, %f55;
	cvt.rn.f32.s32 	%f57, %r1;
	sub.ftz.f32 	%f58, %f19, %f57;
	mul.ftz.f32 	%f59, %f56, %f58;
	sub.ftz.f32 	%f60, %f16, %f59;
	st.f32 	[%rd2+0], %f60;
$Lt_25_514:
$Lt_25_4610:
	.loc	24	340	0
	ret;
$LDWend__Z15ConvertHSLtoRGB8PixelHSLR8PixelRGB:
	} // _Z15ConvertHSLtoRGB8PixelHSLR8PixelRGB
	.const .align 4 .b8 kRGB32f_To_YIQ[36] = {135,22,153,62,162,69,22,63,213,120,233,61,216,128,24,63,27,133,140,190,149,124,164,190,236,135,88,62,134,200,5,191,22,77,159,62};

	.visible .func (.param .align 16 .b8 __cudaretf__Z15ConvertRGBtoYIQ8PixelRGB[16]) _Z15ConvertRGBtoYIQ8PixelRGB (.param .align 16 .b8 __cudaparmf1__Z15ConvertRGBtoYIQ8PixelRGB[16])
	{
	.reg .f32 %f<29>;
	.loc	24	345	0
$LDWbegin__Z15ConvertRGBtoYIQ8PixelRGB:
	ld.param.f32 	%f1, [__cudaparmf1__Z15ConvertRGBtoYIQ8PixelRGB+0];
	mov.f32 	%f2, %f1;
	ld.param.f32 	%f3, [__cudaparmf1__Z15ConvertRGBtoYIQ8PixelRGB+4];
	mov.f32 	%f4, %f3;
	ld.param.f32 	%f5, [__cudaparmf1__Z15ConvertRGBtoYIQ8PixelRGB+8];
	mov.f32 	%f6, %f5;
	ld.param.f32 	%f7, [__cudaparmf1__Z15ConvertRGBtoYIQ8PixelRGB+12];
	mov.f32 	%f8, %f7;
	.loc	24	346	0
	mov.f32 	%f9, %f8;
	st.param.f32 	[__cudaretf__Z15ConvertRGBtoYIQ8PixelRGB+0], %f9;
	ld.const.f32 	%f10, [kRGB32f_To_YIQ+4];
	mul.ftz.f32 	%f11, %f10, %f4;
	ld.const.f32 	%f12, [kRGB32f_To_YIQ+0];
	fma.rn.ftz.f32 	%f13, %f12, %f6, %f11;
	ld.const.f32 	%f14, [kRGB32f_To_YIQ+8];
	fma.rn.ftz.f32 	%f15, %f14, %f2, %f13;
	st.param.f32 	[__cudaretf__Z15ConvertRGBtoYIQ8PixelRGB+4], %f15;
	ld.const.f32 	%f16, [kRGB32f_To_YIQ+16];
	mul.ftz.f32 	%f17, %f16, %f4;
	ld.const.f32 	%f18, [kRGB32f_To_YIQ+12];
	fma.rn.ftz.f32 	%f19, %f18, %f6, %f17;
	ld.const.f32 	%f20, [kRGB32f_To_YIQ+20];
	fma.rn.ftz.f32 	%f21, %f20, %f2, %f19;
	st.param.f32 	[__cudaretf__Z15ConvertRGBtoYIQ8PixelRGB+8], %f21;
	ld.const.f32 	%f22, [kRGB32f_To_YIQ+28];
	mul.ftz.f32 	%f23, %f22, %f4;
	ld.const.f32 	%f24, [kRGB32f_To_YIQ+24];
	fma.rn.ftz.f32 	%f25, %f24, %f6, %f23;
	ld.const.f32 	%f26, [kRGB32f_To_YIQ+32];
	fma.rn.ftz.f32 	%f27, %f26, %f2, %f25;
	st.param.f32 	[__cudaretf__Z15ConvertRGBtoYIQ8PixelRGB+12], %f27;
	ret;
$LDWend__Z15ConvertRGBtoYIQ8PixelRGB:
	} // _Z15ConvertRGBtoYIQ8PixelRGB
	.const .align 4 .b8 kYIQ_To_RGB32f[36] = {0,0,128,63,20,208,116,63,219,249,30,63,0,0,128,63,177,80,139,190,2,188,37,191,0,0,128,63,45,178,141,191,85,48,218,63};

	.visible .func (.param .align 16 .b8 __cudaretf__Z15ConvertYIQtoRGB6float4[16]) _Z15ConvertYIQtoRGB6float4 (.param .align 16 .b8 __cudaparmf1__Z15ConvertYIQtoRGB6float4[16])
	{
	.reg .f32 %f<29>;
	.loc	24	353	0
$LDWbegin__Z15ConvertYIQtoRGB6float4:
	ld.param.f32 	%f1, [__cudaparmf1__Z15ConvertYIQtoRGB6float4+0];
	mov.f32 	%f2, %f1;
	ld.param.f32 	%f3, [__cudaparmf1__Z15ConvertYIQtoRGB6float4+4];
	mov.f32 	%f4, %f3;
	ld.param.f32 	%f5, [__cudaparmf1__Z15ConvertYIQtoRGB6float4+8];
	mov.f32 	%f6, %f5;
	ld.param.f32 	%f7, [__cudaparmf1__Z15ConvertYIQtoRGB6float4+12];
	mov.f32 	%f8, %f7;
	.loc	24	354	0
	ld.const.f32 	%f9, [kYIQ_To_RGB32f+28];
	mul.ftz.f32 	%f10, %f9, %f6;
	ld.const.f32 	%f11, [kYIQ_To_RGB32f+24];
	fma.rn.ftz.f32 	%f12, %f11, %f4, %f10;
	ld.const.f32 	%f13, [kYIQ_To_RGB32f+32];
	fma.rn.ftz.f32 	%f14, %f13, %f8, %f12;
	st.param.f32 	[__cudaretf__Z15ConvertYIQtoRGB6float4+0], %f14;
	ld.const.f32 	%f15, [kYIQ_To_RGB32f+16];
	mul.ftz.f32 	%f16, %f15, %f6;
	ld.const.f32 	%f17, [kYIQ_To_RGB32f+12];
	fma.rn.ftz.f32 	%f18, %f17, %f4, %f16;
	ld.const.f32 	%f19, [kYIQ_To_RGB32f+20];
	fma.rn.ftz.f32 	%f20, %f19, %f8, %f18;
	st.param.f32 	[__cudaretf__Z15ConvertYIQtoRGB6float4+4], %f20;
	ld.const.f32 	%f21, [kYIQ_To_RGB32f+4];
	mul.ftz.f32 	%f22, %f21, %f6;
	ld.const.f32 	%f23, [kYIQ_To_RGB32f+0];
	fma.rn.ftz.f32 	%f24, %f23, %f4, %f22;
	ld.const.f32 	%f25, [kYIQ_To_RGB32f+8];
	fma.rn.ftz.f32 	%f26, %f25, %f8, %f24;
	st.param.f32 	[__cudaretf__Z15ConvertYIQtoRGB6float4+8], %f26;
	mov.f32 	%f27, %f2;
	st.param.f32 	[__cudaretf__Z15ConvertYIQtoRGB6float4+12], %f27;
	ret;
$LDWend__Z15ConvertYIQtoRGB6float4:
	} // _Z15ConvertYIQtoRGB6float4

	.entry InvertKernel (
		.param .u32 __cudaparm_InvertKernel_inDeviceFormat,
		.param .u64 __cudaparm_InvertKernel_inOutput,
		.param .s32 __cudaparm_InvertKernel_inWidth,
		.param .s32 __cudaparm_InvertKernel_inHeight,
		.param .s32 __cudaparm_InvertKernel_inOutputPitch,
		.param .u64 __cudaparm_InvertKernel_inInput,
		.param .s32 __cudaparm_InvertKernel_inInputPitch,
		.param .align 4 .b8 __cudaparm_InvertKernel_inParameters[8])
	{
	.reg .u32 %r<82>;
	.reg .u64 %rd<14>;
	.reg .f32 %f<176>;
	.reg .pred %p<44>;
	// __cuda_local_var_92229_11_non_const_inv = 16
	// __cuda_local_var_92265_12_non_const_hsl = 32
	.loc	24	60	0
$LDWbegin_InvertKernel:
	.loc	24	63	0
	cvt.s32.u32 	%r1, %ctaid.x;
	cvt.s32.u32 	%r2, %ntid.x;
	mul.lo.s32 	%r3, %r1, %r2;
	cvt.s32.u32 	%r4, %ctaid.y;
	cvt.s32.u32 	%r5, %ntid.y;
	mul.lo.s32 	%r6, %r4, %r5;
	mov.u32 	%r7, %tid.x;
	add.u32 	%r8, %r3, %r7;
	mov.u32 	%r9, %tid.y;
	add.u32 	%r10, %r6, %r9;
	ld.param.s32 	%r11, [__cudaparm_InvertKernel_inWidth];
	set.le.u32.s32 	%r12, %r11, %r8;
	neg.s32 	%r13, %r12;
	ld.param.s32 	%r14, [__cudaparm_InvertKernel_inHeight];
	set.le.u32.s32 	%r15, %r14, %r10;
	neg.s32 	%r16, %r15;
	or.b32 	%r17, %r13, %r16;
	mov.u32 	%r18, 0;
	setp.eq.s32 	%p1, %r17, %r18;
	@%p1 bra 	$Lt_28_29698;
	bra.uni 	$LBB83_InvertKernel;
$Lt_28_29698:
	.loc	24	67	0
	ld.param.s32 	%r19, [__cudaparm_InvertKernel_inDeviceFormat];
	mov.s32 	%r20, 0;
	setp.eq.s32 	%p2, %r19, %r20;
	ld.param.s32 	%r21, [__cudaparm_InvertKernel_inInputPitch];
	mul.lo.s32 	%r22, %r21, %r10;
	add.s32 	%r23, %r8, %r22;
	cvt.s64.s32 	%rd1, %r23;
	ld.param.u64 	%rd2, [__cudaparm_InvertKernel_inInput];
	@!%p2 bra 	$Lt_28_30466;
	.loc	21	115	0
	mul.lo.u64 	%rd3, %rd1, 8;
	add.u64 	%rd4, %rd2, %rd3;
	ld.global.v4.u16 	{%r24,%r25,%r26,%r27}, [%rd4+0];
	.loc	24	70	0
	{ .reg .b32 %b1;
	mov.b32		%b1, %r24;
	cvt.ftz.f32.f16	%f1, %b1; }
	{ .reg .b32 %b1;
	mov.b32		%b1, %r25;
	cvt.ftz.f32.f16	%f2, %b1; }
	{ .reg .b32 %b1;
	mov.b32		%b1, %r26;
	cvt.ftz.f32.f16	%f3, %b1; }
	{ .reg .b32 %b1;
	mov.b32		%b1, %r27;
	cvt.ftz.f32.f16	%f4, %b1; }
	bra.uni 	$Lt_28_30210;
$Lt_28_30466:
	mul.lo.u64 	%rd5, %rd1, 16;
	add.u64 	%rd6, %rd2, %rd5;
	ld.global.v4.f32 	{%f1,%f2,%f3,%f4}, [%rd6+0];
$Lt_28_30210:
	.loc	24	72	0
	mov.f32 	%f5, 0f00000000;     	// 0
	max.ftz.f32 	%f6, %f3, %f5;
	mov.f32 	%f7, 0f3f800000;     	// 1
	min.ftz.f32 	%f8, %f6, %f7;
	.loc	24	73	0
	mov.f32 	%f9, 0f00000000;     	// 0
	max.ftz.f32 	%f10, %f2, %f9;
	mov.f32 	%f11, 0f3f800000;    	// 1
	min.ftz.f32 	%f12, %f10, %f11;
	.loc	24	74	0
	mov.f32 	%f13, 0f00000000;    	// 0
	max.ftz.f32 	%f14, %f1, %f13;
	mov.f32 	%f15, 0f3f800000;    	// 1
	min.ftz.f32 	%f16, %f14, %f15;
	.loc	24	75	0
	mov.f32 	%f17, 0f00000000;    	// 0
	max.ftz.f32 	%f18, %f4, %f17;
	mov.f32 	%f19, 0f3f800000;    	// 1
	min.ftz.f32 	%f20, %f18, %f19;
	mov.f32 	%f21, %f20;
	.loc	24	79	0
	ld.param.s32 	%r28, [__cudaparm_InvertKernel_inParameters+0];
	mov.u32 	%r29, 15;
	setp.eq.s32 	%p3, %r28, %r29;
	@%p3 bra 	$Lt_28_258;
	mov.u32 	%r30, 0;
	setp.eq.s32 	%p4, %r28, %r30;
	@%p4 bra 	$Lt_28_770;
	mov.u32 	%r31, 1;
	setp.eq.s32 	%p5, %r28, %r31;
	@%p5 bra 	$Lt_28_1026;
	mov.u32 	%r32, 2;
	setp.eq.s32 	%p6, %r28, %r32;
	@%p6 bra 	$Lt_28_1282;
	mov.u32 	%r33, 3;
	setp.eq.s32 	%p7, %r28, %r33;
	@%p7 bra 	$Lt_28_1538;
	mov.s32 	%r34, 5;
	setp.eq.s32 	%p8, %r28, %r34;
	@%p8 bra 	$Lt_28_1794;
	mov.u32 	%r35, 6;
	setp.eq.s32 	%p9, %r28, %r35;
	@%p9 bra 	$Lt_28_1794;
	mov.u32 	%r36, 7;
	setp.eq.s32 	%p10, %r28, %r36;
	@%p10 bra 	$Lt_28_1794;
	mov.u32 	%r37, 8;
	setp.eq.s32 	%p11, %r28, %r37;
	@%p11 bra 	$Lt_28_1794;
	mov.s32 	%r38, 10;
	setp.eq.s32 	%p12, %r28, %r38;
	@%p12 bra 	$Lt_28_4866;
	mov.u32 	%r39, 11;
	setp.eq.s32 	%p13, %r28, %r39;
	@%p13 bra 	$Lt_28_4866;
	mov.u32 	%r40, 12;
	setp.eq.s32 	%p14, %r28, %r40;
	@%p14 bra 	$Lt_28_4866;
	mov.u32 	%r41, 13;
	setp.eq.s32 	%p15, %r28, %r41;
	@%p15 bra 	$Lt_28_4866;
	bra.uni 	$Lt_28_514;
$Lt_28_258:
	.loc	24	82	0
	ld.param.f32 	%f22, [__cudaparm_InvertKernel_inParameters+4];
	mov.f32 	%f23, 0f3f800000;    	// 1
	sub.ftz.f32 	%f24, %f23, %f20;
	mov.f32 	%f25, 0f3f800000;    	// 1
	sub.ftz.f32 	%f26, %f25, %f22;
	mul.ftz.f32 	%f27, %f24, %f26;
	fma.rn.ftz.f32 	%f21, %f22, %f20, %f27;
	.loc	24	83	0
	bra.uni 	$Lt_28_514;
$Lt_28_770:
	.loc	24	86	0
	mov.f32 	%f28, 0f3f800000;    	// 1
	sub.ftz.f32 	%f29, %f28, %f8;
	mov.f32 	%f30, %f29;
	.loc	24	87	0
	mov.f32 	%f31, 0f3f800000;    	// 1
	sub.ftz.f32 	%f32, %f31, %f12;
	mov.f32 	%f33, %f32;
	.loc	24	88	0
	mov.f32 	%f34, 0f3f800000;    	// 1
	sub.ftz.f32 	%f35, %f34, %f16;
	mov.f32 	%f36, %f35;
	.loc	24	89	0
	bra.uni 	$Lt_28_514;
$Lt_28_1026:
	.loc	24	92	0
	mov.f32 	%f37, 0f3f800000;    	// 1
	sub.ftz.f32 	%f38, %f37, %f8;
	mov.f32 	%f30, %f38;
	.loc	24	93	0
	mov.f32 	%f33, %f12;
	.loc	24	94	0
	mov.f32 	%f36, %f16;
	.loc	24	95	0
	bra.uni 	$Lt_28_514;
$Lt_28_1282:
	.loc	24	98	0
	mov.f32 	%f30, %f8;
	.loc	24	99	0
	mov.f32 	%f39, 0f3f800000;    	// 1
	sub.ftz.f32 	%f40, %f39, %f12;
	mov.f32 	%f33, %f40;
	.loc	24	100	0
	mov.f32 	%f36, %f16;
	.loc	24	101	0
	bra.uni 	$Lt_28_514;
$Lt_28_1538:
	.loc	24	104	0
	mov.f32 	%f30, %f8;
	.loc	24	105	0
	mov.f32 	%f33, %f12;
	.loc	24	106	0
	mov.f32 	%f41, 0f3f800000;    	// 1
	sub.ftz.f32 	%f42, %f41, %f16;
	mov.f32 	%f36, %f42;
	.loc	24	107	0
	bra.uni 	$Lt_28_514;
$Lt_28_1794:
$Lt_28_2050:
$Lt_28_2306:
$Lt_28_2562:
	.loc	24	114	0
	setp.ge.ftz.f32 	%p16, %f8, %f16;
	setp.ge.ftz.f32 	%p17, %f8, %f12;
	selp.s32 	%r42, 1, 0, %p16;
	selp.s32 	%r43, 1, 0, %p17;
	and.b32 	%r44, %r42, %r43;
	mov.u32 	%r45, 0;
	setp.eq.s32 	%p18, %r44, %r45;
	@%p18 bra 	$Lt_28_30722;
	.loc	24	214	0
	sub.ftz.f32 	%f43, %f12, %f16;
	mov.f32 	%f44, %f43;
$Lt_28_30722:
	mov.f32 	%f45, %f8;
	setp.ge.ftz.f32 	%p19, %f12, %f16;
	selp.s32 	%r46, 1, 0, %p19;
	set.le.ftz.u32.f32 	%r47, %f8, %f12;
	neg.s32 	%r48, %r47;
	and.b32 	%r49, %r46, %r48;
	mov.u32 	%r50, 0;
	setp.eq.s32 	%p20, %r49, %r50;
	@%p20 bra 	$Lt_28_31490;
	.loc	24	219	0
	mov.f32 	%f45, %f12;
	.loc	24	220	0
	sub.ftz.f32 	%f46, %f16, %f8;
	mov.f32 	%f44, %f46;
	mov.f32 	%f47, 0f40000000;    	// 2
	bra.uni 	$Lt_28_31234;
$Lt_28_31490:
	mov.f32 	%f47, 0f00000000;    	// 0
$Lt_28_31234:
	setp.le.ftz.f32 	%p21, %f12, %f16;
	selp.s32 	%r51, 1, 0, %p21;
	set.le.ftz.u32.f32 	%r52, %f8, %f16;
	neg.s32 	%r53, %r52;
	and.b32 	%r54, %r51, %r53;
	mov.u32 	%r55, 0;
	setp.eq.s32 	%p22, %r54, %r55;
	@%p22 bra 	$Lt_28_31746;
	.loc	24	225	0
	mov.f32 	%f45, %f16;
	.loc	24	226	0
	sub.ftz.f32 	%f48, %f8, %f12;
	mov.f32 	%f44, %f48;
	mov.f32 	%f47, 0f40800000;    	// 4
$Lt_28_31746:
	.loc	24	242	0
	and.b32 	%r56, %r42, %r46;
	and.b32 	%r57, %r43, %r51;
	neg.s32 	%r58, %r57;
	slct.f32.s32 	%f49, %f8, %f12, %r58;
	neg.s32 	%r59, %r56;
	slct.f32.s32 	%f50, %f49, %f16, %r59;
	add.ftz.f32 	%f51, %f50, %f45;
	mov.f32 	%f52, 0f40000000;    	// 2
	div.approx.ftz.f32 	%f53, %f51, %f52;
	mov.f32 	%f54, %f53;
	.loc	24	243	0
	mov.f32 	%f55, 0f00000000;    	// 0
	mov.f32 	%f56, 0f38d1b717;    	// 0.0001
	setp.gt.ftz.f32 	%p23, %f53, %f56;
	@!%p23 bra 	$Lt_28_32514;
	.loc	24	247	0
	sub.ftz.f32 	%f57, %f45, %f50;
	mov.f32 	%f55, %f57;
	mov.f32 	%f58, 0f38d1b717;    	// 0.0001
	setp.gt.ftz.f32 	%p24, %f57, %f58;
	@!%p24 bra 	$Lt_28_33026;
	.loc	24	251	0
	mov.f32 	%f59, %f44;
	div.approx.ftz.f32 	%f60, %f59, %f57;
	add.ftz.f32 	%f61, %f47, %f60;
	.loc	24	258	0
	mov.f32 	%f62, 0f40c00000;    	// 6
	add.ftz.f32 	%f63, %f61, %f62;
	mov.f32 	%f64, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p25, %f61, %f64;
	selp.f32 	%f65, %f63, %f61, %p25;
	mov.f32 	%f66, 0f40c00000;    	// 6
	div.approx.ftz.f32 	%f67, %f65, %f66;
	bra.uni 	$Lt_28_32258;
$Lt_28_33026:
	.loc	24	262	0
	mov.f32 	%f55, 0f00000000;    	// 0
	.loc	24	263	0
	mov.f32 	%f67, 0f00000000;    	// 0
	bra.uni 	$Lt_28_32258;
$Lt_28_32514:
	.loc	24	268	0
	mov.f32 	%f67, 0f00000000;    	// 0
$Lt_28_32258:
	.loc	24	115	0
	@%p8 bra 	$Lt_28_2818;
	mov.u32 	%r60, 6;
	setp.eq.s32 	%p26, %r28, %r60;
	@%p26 bra 	$Lt_28_3330;
	mov.u32 	%r61, 7;
	setp.eq.s32 	%p27, %r28, %r61;
	@%p27 bra 	$Lt_28_3586;
	mov.u32 	%r62, 8;
	setp.eq.s32 	%p28, %r28, %r62;
	@%p28 bra 	$Lt_28_3842;
	bra.uni 	$Lt_28_3074;
$Lt_28_2818:
	mov.f32 	%f68, 0f3f000000;    	// 0.5
	sub.ftz.f32 	%f69, %f68, %f67;
	mov.f32 	%f70, 0f3f800000;    	// 1
	add.ftz.f32 	%f71, %f69, %f70;
	mov.f32 	%f72, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p29, %f69, %f72;
	selp.f32 	%f67, %f71, %f69, %p29;
	.loc	24	123	0
	mov.f32 	%f73, 0f3f800000;    	// 1
	sub.ftz.f32 	%f54, %f73, %f53;
	.loc	24	124	0
	mov.f32 	%f74, 0f3f800000;    	// 1
	sub.ftz.f32 	%f55, %f74, %f55;
	.loc	24	125	0
	bra.uni 	$Lt_28_3074;
$Lt_28_3330:
	.loc	24	126	0
	mov.f32 	%f75, 0f3f000000;    	// 0.5
	sub.ftz.f32 	%f76, %f75, %f67;
	mov.f32 	%f77, 0f3f800000;    	// 1
	add.ftz.f32 	%f78, %f76, %f77;
	mov.f32 	%f79, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p30, %f76, %f79;
	selp.f32 	%f67, %f78, %f76, %p30;
	.loc	24	132	0
	bra.uni 	$Lt_28_3074;
$Lt_28_3586:
	.loc	24	134	0
	mov.f32 	%f80, 0f3f800000;    	// 1
	sub.ftz.f32 	%f54, %f80, %f53;
	.loc	24	135	0
	bra.uni 	$Lt_28_3074;
$Lt_28_3842:
	.loc	24	138	0
	mov.f32 	%f81, 0f3f800000;    	// 1
	sub.ftz.f32 	%f55, %f81, %f55;
$Lt_28_3074:
	.loc	24	141	0
	mov.f32 	%f82, 0f3f000000;    	// 0.5
	setp.le.ftz.f32 	%p31, %f54, %f82;
	@!%p31 bra 	$Lt_28_33538;
	.loc	24	287	0
	mov.f32 	%f83, 0f3f800000;    	// 1
	add.ftz.f32 	%f84, %f55, %f83;
	mul.ftz.f32 	%f85, %f54, %f84;
	bra.uni 	$Lt_28_33282;
$Lt_28_33538:
	add.ftz.f32 	%f86, %f55, %f54;
	mul.ftz.f32 	%f87, %f55, %f54;
	sub.ftz.f32 	%f85, %f86, %f87;
$Lt_28_33282:
	.loc	24	289	0
	mov.f32 	%f88, 0f00000000;    	// 0
	mov.f32 	%f89, 0f00000000;    	// 0
	mov.f32 	%f90, 0f00000000;    	// 0
	mov.f32 	%f91, %f85;
	mov.f32 	%f92, 0f00000000;    	// 0
	setp.gt.ftz.f32 	%p32, %f91, %f92;
	@!%p32 bra 	$Lt_28_34050;
	.loc	24	306	0
	mov.f32 	%f93, 0f40c00000;    	// 6
	mul.ftz.f32 	%f94, %f67, %f93;
	cvt.rzi.ftz.s32.f32 	%r63, %f94;
	mov.u32 	%r64, 0;
	setp.eq.s32 	%p33, %r63, %r64;
	@%p33 bra 	$Lt_28_8450;
	mov.u32 	%r65, 1;
	setp.eq.s32 	%p34, %r63, %r65;
	@%p34 bra 	$Lt_28_8962;
	mov.u32 	%r66, 2;
	setp.eq.s32 	%p35, %r63, %r66;
	@%p35 bra 	$Lt_28_9218;
	mov.u32 	%r67, 3;
	setp.eq.s32 	%p36, %r63, %r67;
	@%p36 bra 	$Lt_28_9474;
	mov.u32 	%r68, 4;
	setp.eq.s32 	%p37, %r63, %r68;
	@%p37 bra 	$Lt_28_9730;
	mov.u32 	%r69, 5;
	setp.eq.s32 	%p38, %r63, %r69;
	@%p38 bra 	$Lt_28_9986;
	mov.f32 	%f36, %f88;
	mov.f32 	%f33, %f89;
	mov.f32 	%f30, %f90;
	bra.uni 	$Lt_28_514;
$Lt_28_8450:
	.loc	24	309	0
	mov.f32 	%f30, %f91;
	.loc	24	310	0
	add.ftz.f32 	%f95, %f54, %f54;
	sub.ftz.f32 	%f96, %f95, %f91;
	sub.ftz.f32 	%f97, %f91, %f96;
	div.approx.ftz.f32 	%f98, %f97, %f91;
	mul.ftz.f32 	%f99, %f91, %f98;
	cvt.rn.f32.s32 	%f100, %r63;
	sub.ftz.f32 	%f101, %f94, %f100;
	fma.rn.ftz.f32 	%f89, %f99, %f101, %f96;
	mov.f32 	%f33, %f89;
	.loc	24	311	0
	mov.f32 	%f36, %f96;
	.loc	24	312	0
	bra.uni 	$Lt_28_514;
$Lt_28_8962:
	.loc	24	314	0
	add.ftz.f32 	%f95, %f54, %f54;
	sub.ftz.f32 	%f96, %f95, %f91;
	sub.ftz.f32 	%f102, %f91, %f96;
	div.approx.ftz.f32 	%f103, %f102, %f91;
	mul.ftz.f32 	%f104, %f91, %f103;
	cvt.rn.f32.s32 	%f105, %r63;
	sub.ftz.f32 	%f106, %f94, %f105;
	mul.ftz.f32 	%f107, %f104, %f106;
	sub.ftz.f32 	%f90, %f91, %f107;
	mov.f32 	%f30, %f90;
	.loc	24	315	0
	mov.f32 	%f33, %f91;
	.loc	24	316	0
	mov.f32 	%f36, %f96;
	.loc	24	317	0
	bra.uni 	$Lt_28_514;
$Lt_28_9218:
	.loc	24	319	0
	add.ftz.f32 	%f95, %f54, %f54;
	sub.ftz.f32 	%f96, %f95, %f91;
	mov.f32 	%f30, %f96;
	.loc	24	320	0
	mov.f32 	%f33, %f91;
	.loc	24	321	0
	sub.ftz.f32 	%f108, %f91, %f96;
	div.approx.ftz.f32 	%f109, %f108, %f91;
	mul.ftz.f32 	%f110, %f91, %f109;
	cvt.rn.f32.s32 	%f111, %r63;
	sub.ftz.f32 	%f112, %f94, %f111;
	fma.rn.ftz.f32 	%f88, %f110, %f112, %f96;
	mov.f32 	%f36, %f88;
	.loc	24	322	0
	bra.uni 	$Lt_28_514;
$Lt_28_9474:
	.loc	24	324	0
	add.ftz.f32 	%f95, %f54, %f54;
	sub.ftz.f32 	%f96, %f95, %f91;
	mov.f32 	%f30, %f96;
	.loc	24	325	0
	sub.ftz.f32 	%f113, %f91, %f96;
	div.approx.ftz.f32 	%f114, %f113, %f91;
	mul.ftz.f32 	%f115, %f91, %f114;
	cvt.rn.f32.s32 	%f116, %r63;
	sub.ftz.f32 	%f117, %f94, %f116;
	mul.ftz.f32 	%f118, %f115, %f117;
	sub.ftz.f32 	%f89, %f91, %f118;
	mov.f32 	%f33, %f89;
	.loc	24	326	0
	mov.f32 	%f36, %f91;
	.loc	24	327	0
	bra.uni 	$Lt_28_514;
$Lt_28_9730:
	.loc	24	329	0
	add.ftz.f32 	%f95, %f54, %f54;
	sub.ftz.f32 	%f96, %f95, %f91;
	sub.ftz.f32 	%f119, %f91, %f96;
	div.approx.ftz.f32 	%f120, %f119, %f91;
	mul.ftz.f32 	%f121, %f91, %f120;
	cvt.rn.f32.s32 	%f122, %r63;
	sub.ftz.f32 	%f123, %f94, %f122;
	fma.rn.ftz.f32 	%f90, %f121, %f123, %f96;
	mov.f32 	%f30, %f90;
	.loc	24	330	0
	mov.f32 	%f33, %f96;
	.loc	24	331	0
	mov.f32 	%f36, %f91;
	.loc	24	332	0
	bra.uni 	$Lt_28_514;
$Lt_28_9986:
	.loc	24	334	0
	mov.f32 	%f30, %f91;
	.loc	24	335	0
	add.ftz.f32 	%f95, %f54, %f54;
	sub.ftz.f32 	%f96, %f95, %f91;
	mov.f32 	%f33, %f96;
	.loc	24	336	0
	sub.ftz.f32 	%f124, %f91, %f96;
	div.approx.ftz.f32 	%f125, %f124, %f91;
	mul.ftz.f32 	%f126, %f91, %f125;
	cvt.rn.f32.s32 	%f127, %r63;
	sub.ftz.f32 	%f128, %f94, %f127;
	mul.ftz.f32 	%f129, %f126, %f128;
	sub.ftz.f32 	%f88, %f91, %f129;
	mov.f32 	%f36, %f88;
	bra.uni 	$Lt_28_514;
$Lt_28_34050:
	.loc	24	338	0
	mov.f32 	%f36, %f88;
	mov.f32 	%f33, %f89;
	mov.f32 	%f30, %f90;
	bra.uni 	$Lt_28_514;
$Lt_28_4866:
$Lt_28_5122:
$Lt_28_5378:
$Lt_28_5634:
	.loc	24	149	0
	ld.const.f32 	%f130, [kRGB32f_To_YIQ+4];
	mul.ftz.f32 	%f131, %f130, %f12;
	ld.const.f32 	%f132, [kRGB32f_To_YIQ+0];
	fma.rn.ftz.f32 	%f133, %f132, %f8, %f131;
	ld.const.f32 	%f134, [kRGB32f_To_YIQ+8];
	fma.rn.ftz.f32 	%f135, %f134, %f16, %f133;
	mov.f32 	%f136, %f135;
	ld.const.f32 	%f137, [kRGB32f_To_YIQ+16];
	mul.ftz.f32 	%f138, %f137, %f12;
	ld.const.f32 	%f139, [kRGB32f_To_YIQ+12];
	fma.rn.ftz.f32 	%f140, %f139, %f8, %f138;
	ld.const.f32 	%f141, [kRGB32f_To_YIQ+20];
	fma.rn.ftz.f32 	%f142, %f141, %f16, %f140;
	mov.f32 	%f143, %f142;
	ld.const.f32 	%f144, [kRGB32f_To_YIQ+28];
	mul.ftz.f32 	%f145, %f144, %f12;
	ld.const.f32 	%f146, [kRGB32f_To_YIQ+24];
	fma.rn.ftz.f32 	%f147, %f146, %f8, %f145;
	ld.const.f32 	%f148, [kRGB32f_To_YIQ+32];
	fma.rn.ftz.f32 	%f149, %f148, %f16, %f147;
	mov.f32 	%f150, %f149;
	.loc	24	150	0
	@%p12 bra 	$Lt_28_5890;
	mov.u32 	%r70, 11;
	setp.eq.s32 	%p39, %r28, %r70;
	@%p39 bra 	$Lt_28_6402;
	mov.u32 	%r71, 12;
	setp.eq.s32 	%p40, %r28, %r71;
	@%p40 bra 	$Lt_28_6658;
	mov.u32 	%r72, 13;
	setp.eq.s32 	%p41, %r28, %r72;
	@%p41 bra 	$Lt_28_6914;
	bra.uni 	$Lt_28_6146;
$Lt_28_5890:
	.loc	24	153	0
	mov.f32 	%f151, 0f3f800000;   	// 1
	sub.ftz.f32 	%f136, %f151, %f135;
	.loc	24	154	0
	neg.ftz.f32 	%f143, %f142;
	.loc	24	155	0
	neg.ftz.f32 	%f150, %f149;
	.loc	24	156	0
	bra.uni 	$Lt_28_6146;
$Lt_28_6402:
	.loc	24	158	0
	mov.f32 	%f152, 0f3f800000;   	// 1
	sub.ftz.f32 	%f136, %f152, %f135;
	.loc	24	159	0
	bra.uni 	$Lt_28_6146;
$Lt_28_6658:
	.loc	24	161	0
	neg.ftz.f32 	%f143, %f142;
	.loc	24	162	0
	bra.uni 	$Lt_28_6146;
$Lt_28_6914:
	.loc	24	164	0
	neg.ftz.f32 	%f150, %f149;
$Lt_28_6146:
	.loc	24	168	0
	ld.const.f32 	%f153, [kYIQ_To_RGB32f+28];
	mul.ftz.f32 	%f154, %f153, %f143;
	ld.const.f32 	%f155, [kYIQ_To_RGB32f+24];
	fma.rn.ftz.f32 	%f156, %f155, %f136, %f154;
	ld.const.f32 	%f157, [kYIQ_To_RGB32f+32];
	fma.rn.ftz.f32 	%f158, %f157, %f150, %f156;
	mov.f32 	%f36, %f158;
	ld.const.f32 	%f159, [kYIQ_To_RGB32f+16];
	mul.ftz.f32 	%f160, %f159, %f143;
	ld.const.f32 	%f161, [kYIQ_To_RGB32f+12];
	fma.rn.ftz.f32 	%f162, %f161, %f136, %f160;
	ld.const.f32 	%f163, [kYIQ_To_RGB32f+20];
	fma.rn.ftz.f32 	%f164, %f163, %f150, %f162;
	mov.f32 	%f33, %f164;
	ld.const.f32 	%f165, [kYIQ_To_RGB32f+4];
	mul.ftz.f32 	%f166, %f165, %f143;
	ld.const.f32 	%f167, [kYIQ_To_RGB32f+0];
	fma.rn.ftz.f32 	%f168, %f167, %f136, %f166;
	ld.const.f32 	%f169, [kYIQ_To_RGB32f+8];
	fma.rn.ftz.f32 	%f170, %f169, %f150, %f168;
	mov.f32 	%f30, %f170;
$Lt_28_514:
	.loc	24	170	0
	mov.u32 	%r73, 15;
	setp.eq.s32 	%p42, %r28, %r73;
	@%p42 bra 	$Lt_28_34306;
	.loc	24	174	0
	mov.f32 	%f90, %f30;
	ld.param.f32 	%f171, [__cudaparm_InvertKernel_inParameters+4];
	sub.ftz.f32 	%f172, %f8, %f90;
	fma.rn.ftz.f32 	%f8, %f171, %f172, %f90;
	.loc	24	175	0
	mov.f32 	%f89, %f33;
	sub.ftz.f32 	%f173, %f12, %f89;
	fma.rn.ftz.f32 	%f12, %f171, %f173, %f89;
	.loc	24	176	0
	mov.f32 	%f88, %f36;
	sub.ftz.f32 	%f174, %f16, %f88;
	fma.rn.ftz.f32 	%f16, %f171, %f174, %f88;
$Lt_28_34306:
	ld.param.s32 	%r74, [__cudaparm_InvertKernel_inOutputPitch];
	mul.lo.s32 	%r75, %r74, %r10;
	add.s32 	%r76, %r8, %r75;
	cvt.s64.s32 	%rd7, %r76;
	ld.param.u64 	%rd8, [__cudaparm_InvertKernel_inOutput];
	@!%p2 bra 	$Lt_28_35074;
	.loc	21	126	0
	mul.lo.u64 	%rd9, %rd7, 8;
	add.u64 	%rd10, %rd8, %rd9;
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f16;
	mov.b32		%r77, %b1; }
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f12;
	mov.b32		%r78, %b1; }
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f8;
	mov.b32		%r79, %b1; }
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f21;
	mov.b32		%r80, %b1; }
	st.global.v4.u16 	[%rd10+0], {%r77,%r78,%r79,%r80};
	.loc	24	179	0
	bra.uni 	$LBB83_InvertKernel;
$Lt_28_35074:
	.loc	21	126	0
	mul.lo.u64 	%rd11, %rd7, 16;
	add.u64 	%rd12, %rd8, %rd11;
	st.global.v4.f32 	[%rd12+0], {%f16,%f12,%f8,%f21};
$LBB83_InvertKernel:
	.loc	24	181	0
	exit;
$LDWend_InvertKernel:
	} // InvertKernel
	.const .align 4 .b8 kRGB32f_To_601YPbPr[36] = {135,22,153,62,162,69,22,63,213,120,233,61,33,201,44,190,111,155,169,190,0,0,0,63,0,0,0,63,70,94,214,190,232,134,166,189};
	.const .align 4 .b8 k601YPbPr_To_RGB32f[36] = {0,0,128,63,0,0,0,0,188,116,179,63,0,0,128,63,152,50,176,190,158,209,54,191,0,0,128,63,229,208,226,63,0,0,0,0};
	.const .align 4 .b8 kRGB32f_To_601YCbCr[36] = {70,246,130,66,145,141,0,67,94,186,199,65,33,48,23,194,240,103,148,194,0,0,224,66,0,0,224,66,111,146,187,194,70,182,145,193};
	.const .align 4 .b8 k601YCbCr_To_RGB32f[36] = {37,160,149,59,0,0,0,0,182,23,205,59,37,160,149,59,40,15,201,186,156,239,80,187,37,160,149,59,236,155,1,60,0,0,0,0};
	.const .align 4 .b8 kRGB8u_To_601YCbCr[36] = {219,121,131,62,152,14,1,63,18,131,200,61,174,199,23,190,238,252,148,190,197,224,224,62,197,224,224,62,217,78,188,190,174,71,146,189};
	.const .align 4 .b8 k601YCbCr_To_RGB8u[36] = {127,10,149,63,0,0,0,0,160,74,204,63,127,10,149,63,254,148,200,190,184,30,80,191,127,10,149,63,78,26,1,64,0,0,0,0};
	.const .align 4 .b8 kRGB8u_To_601YCbCrFullRange[36] = {135,22,153,62,162,69,22,63,213,120,233,61,166,27,44,190,39,241,168,190,250,254,254,62,250,254,254,62,43,135,213,190,59,223,165,189};
	.const .align 4 .b8 k601YCbCrFullRange_To_RGB8u[36] = {0,0,128,63,0,0,0,0,72,193,178,63,0,0,128,63,143,130,175,190,225,26,54,191,0,0,128,63,20,238,225,63,0,0,0,0};
	.const .align 4 .b8 kRGB32f_To_601YCbCrFullRange[36] = {113,125,152,66,92,175,21,67,92,143,232,65,158,111,43,194,49,72,168,194,0,0,254,66,0,0,254,66,170,177,212,194,88,57,165,193};
	.const .align 4 .b8 k601YCbCrFullRange_To_RGB32f[36] = {129,128,128,59,0,0,0,0,188,116,179,59,129,128,128,59,194,50,176,186,179,209,54,187,129,128,128,59,229,208,226,59,0,0,0,0};
	.const .align 4 .b8 kRGB32f_To_709YPbPr[36] = {208,179,89,62,89,23,55,63,152,221,147,61,186,164,234,189,210,86,197,190,0,0,0,63,0,0,0,63,190,134,232,190,16,202,59,189};
	.const .align 4 .b8 k709YPbPr_To_RGB32f[36] = {0,0,128,63,0,0,0,0,12,147,201,63,0,0,128,63,221,209,63,190,243,173,239,190,0,0,128,63,77,132,237,63,0,0,0,0};
	.const .align 4 .b8 kRGB32f_To_709YCbCr[36] = {106,60,58,66,6,161,28,67,244,253,124,65,223,79,205,193,8,172,172,194,0,0,224,66,0,0,224,66,195,117,203,194,236,81,36,193};
	.const .align 4 .b8 k709YCbCr_To_RGB32f[36] = {37,160,149,59,0,0,0,0,239,94,230,59,37,160,149,59,33,57,91,186,178,245,8,187,37,160,149,59,82,185,7,60,0,0,0,0};
	.const .align 4 .b8 kRGB8u_To_709YCbCr[36] = {207,247,58,62,53,62,29,63,231,251,125,61,147,24,206,61,23,89,173,190,197,224,224,62,197,224,224,62,12,66,204,190,195,245,36,189};
	.const .align 4 .b8 k709YCbCr_To_RGB8u[36] = {127,10,149,63,0,0,0,0,147,120,229,63,127,10,149,63,53,94,90,190,205,108,8,191,127,10,149,63,154,49,7,64,0,0,0,0};
	.const .align 4 .b8 k709YCbCr_To_601YCbCr[36] = {0,0,128,63,23,100,203,61,1,77,68,62,0,0,0,0,18,103,125,63,10,158,226,189,0,0,0,0,61,98,148,189,249,191,123,63};
	.const .align 4 .b8 k601YCbCr_To_709YCbCr[36] = {0,0,128,63,122,165,236,189,179,237,84,190,0,0,0,0,204,98,130,63,216,188,234,61,0,0,0,0,74,179,153,61,234,61,131,63};
	.const .align 4 .b8 kYCbCrOffset[12] = {0,0,128,65,0,0,0,67,0,0,0,67};
	.const .align 4 .b8 kYCbCrFullRangeOffset[12] = {0,0,0,0,0,0,0,67,0,0,0,67};

