	.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 (.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 (.param .f32 __cudaretf__Z3MaxIfET_S0_S0_) _Z3MaxIfET_S0_S0_ (.param .f32 __cudaparmf1__Z3MaxIfET_S0_S0_, .param .f32 __cudaparmf2__Z3MaxIfET_S0_S0_)

	.visible .func (.param .f32 __cudaretf__Z11HueWrapNormf) _Z11HueWrapNormf (.param .f32 __cudaparmf1__Z11HueWrapNormf)

	.visible .func (.param .f32 __cudaretf__Z29CalculateThresholdRangeWeightffff) _Z29CalculateThresholdRangeWeightffff (.param .f32 __cudaparmf1__Z29CalculateThresholdRangeWeightffff, .param .f32 __cudaparmf2__Z29CalculateThresholdRangeWeightffff, .param .f32 __cudaparmf3__Z29CalculateThresholdRangeWeightffff, .param .f32 __cudaparmf4__Z29CalculateThresholdRangeWeightffff)

	.visible .func (.param .f32 __cudaretf__Z33CalculateStartSoftnessRangeWeightffffb) _Z33CalculateStartSoftnessRangeWeightffffb (.param .f32 __cudaparmf1__Z33CalculateStartSoftnessRangeWeightffffb, .param .f32 __cudaparmf2__Z33CalculateStartSoftnessRangeWeightffffb, .param .f32 __cudaparmf3__Z33CalculateStartSoftnessRangeWeightffffb, .param .f32 __cudaparmf4__Z33CalculateStartSoftnessRangeWeightffffb, .param .s32 __cudaparmf5__Z33CalculateStartSoftnessRangeWeightffffb)

	.visible .func (.param .f32 __cudaretf__Z31CalculateEndSoftnessRangeWeightffffb) _Z31CalculateEndSoftnessRangeWeightffffb (.param .f32 __cudaparmf1__Z31CalculateEndSoftnessRangeWeightffffb, .param .f32 __cudaparmf2__Z31CalculateEndSoftnessRangeWeightffffb, .param .f32 __cudaparmf3__Z31CalculateEndSoftnessRangeWeightffffb, .param .f32 __cudaparmf4__Z31CalculateEndSoftnessRangeWeightffffb, .param .s32 __cudaparmf5__Z31CalculateEndSoftnessRangeWeightffffb)

	.visible .func (.param .f32 __cudaretf__Z20CalculateRangeWeightffffffb) _Z20CalculateRangeWeightffffffb (.param .f32 __cudaparmf1__Z20CalculateRangeWeightffffffb, .param .f32 __cudaparmf2__Z20CalculateRangeWeightffffffb, .param .f32 __cudaparmf3__Z20CalculateRangeWeightffffffb, .param .f32 __cudaparmf4__Z20CalculateRangeWeightffffffb, .param .f32 __cudaparmf5__Z20CalculateRangeWeightffffffb, .param .f32 __cudaparmf6__Z20CalculateRangeWeightffffffb, .param .s32 __cudaparmf7__Z20CalculateRangeWeightffffffb)

	.visible .func (.param .f32 __cudaretf__Z18CalculateHueWeightfRK33SecondaryColorCorrectorParameters) _Z18CalculateHueWeightfRK33SecondaryColorCorrectorParameters (.param .f32 __cudaparmf1__Z18CalculateHueWeightfRK33SecondaryColorCorrectorParameters, .param .u64 __cudaparmf2__Z18CalculateHueWeightfRK33SecondaryColorCorrectorParameters)

	.visible .func (.param .f32 __cudaretf__Z25CalculateSaturationWeightfRK33SecondaryColorCorrectorParameters) _Z25CalculateSaturationWeightfRK33SecondaryColorCorrectorParameters (.param .f32 __cudaparmf1__Z25CalculateSaturationWeightfRK33SecondaryColorCorrectorParameters, .param .u64 __cudaparmf2__Z25CalculateSaturationWeightfRK33SecondaryColorCorrectorParameters)

	.visible .func (.param .f32 __cudaretf__Z19CalculateLumaWeightfRK33SecondaryColorCorrectorParameters) _Z19CalculateLumaWeightfRK33SecondaryColorCorrectorParameters (.param .f32 __cudaparmf1__Z19CalculateLumaWeightfRK33SecondaryColorCorrectorParameters, .param .u64 __cudaparmf2__Z19CalculateLumaWeightfRK33SecondaryColorCorrectorParameters)

	.visible .func (.param .f32 __cudaretf__Z15CalculateWeight8PixelYUVRK33SecondaryColorCorrectorParameters) _Z15CalculateWeight8PixelYUVRK33SecondaryColorCorrectorParameters (.param .align 16 .b8 __cudaparmf1__Z15CalculateWeight8PixelYUVRK33SecondaryColorCorrectorParameters[16], .param .u64 __cudaparmf2__Z15CalculateWeight8PixelYUVRK33SecondaryColorCorrectorParameters)

	.visible .func _Z7Write2DIfEvT_PS0_iii (.param .f32 __cudaparmf1__Z7Write2DIfEvT_PS0_iii, .param .u64 __cudaparmf2__Z7Write2DIfEvT_PS0_iii, .param .s32 __cudaparmf3__Z7Write2DIfEvT_PS0_iii, .param .s32 __cudaparmf4__Z7Write2DIfEvT_PS0_iii, .param .s32 __cudaparmf5__Z7Write2DIfEvT_PS0_iii)

	//-----------------------------------------------------------
	// Compiling C:/Users/dvaeng/AppData/Local/Temp/tmpxft_00003e8c_00000000-11_ColorCorrector_Secondary.cpp3.i (C:/Users/dvaeng/AppData/Local/Temp/ccBI#.a16296)
	//-----------------------------------------------------------

	//-----------------------------------------------------------
	// 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_00003e8c_00000000-10_ColorCorrector_Secondary.cudafe2.gpu"
	.file	2	"c:\Mulder64\shared\adobe\MediaCore\GPUFoundation\API\Inc\GPUFoundation/PixelFormat.h"
	.file	3	"c:\Mulder64\shared\adobe\MediaCore\GPUFoundation\API\Inc\GPUFoundation/KernelSupport/PixelRGB.h"
	.file	4	"c:\Mulder64\shared\adobe\MediaCore\GPUFoundation\API\Inc\GPUFoundation/KernelSupport/PixelYUV.h"
	.file	5	"c:\Mulder64\shared\adobe\MediaCore\GPUFoundation\API\Inc\GPUFoundation/KernelSupport/PixelHSL.h"
	.file	6	"c:\Mulder64\shared\adobe\MediaCore\Display\Inc\CUDA/Effects/ColorCorrector_Secondary.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/ColorCorrector_Secondary.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 (.param .align 16 .b8 __cudaretf__Z18UnpremultiplyPixel8PixelRGB[16]) _Z18UnpremultiplyPixel8PixelRGB (.param .align 16 .b8 __cudaparmf1__Z18UnpremultiplyPixel8PixelRGB[16])
	{
	.reg .f32 %f<23>;
	.reg .pred %p<3>;
	.loc	3	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	3	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_11_1282;
	.loc	3	213	0
	rcp.approx.ftz.f32 	%f14, %f9;
	mul.ftz.f32 	%f15, %f14, %f6;
	.loc	3	214	0
	mul.ftz.f32 	%f16, %f14, %f4;
	.loc	3	215	0
	mul.ftz.f32 	%f17, %f14, %f2;
	bra.uni 	$Lt_11_1026;
$Lt_11_1282:
	.loc	3	219	0
	mov.f32 	%f15, 0f00000000;    	// 0
	mov.f32 	%f16, 0f00000000;    	// 0
	mov.f32 	%f17, 0f00000000;    	// 0
	mov.f32 	%f10, 0f00000000;    	// 0
$Lt_11_1026:
	.loc	3	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	3	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_12_1026;
	.loc	3	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_12_1026:
	.loc	3	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	3	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_13_1026;
	.loc	3	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_13_1026:
	.loc	3	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	3	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	3	254	0
	cvt.ftz.sat.f32.f32 	%f9, %f8;
	.loc	3	255	0
	mov.f32 	%f10, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p1, %f2, %f10;
	@!%p1 bra 	$Lt_14_4098;
	.loc	3	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_191_5;
$Lt_14_4098:
	.loc	3	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_191_5:
	.loc	3	256	0
	mov.f32 	%f20, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p2, %f4, %f20;
	@!%p2 bra 	$Lt_14_4610;
	.loc	3	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_191_3;
$Lt_14_4610:
	.loc	3	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_191_3:
	.loc	3	257	0
	mov.f32 	%f30, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p3, %f6, %f30;
	@!%p3 bra 	$Lt_14_5122;
	.loc	3	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_191_1;
$Lt_14_5122:
	.loc	3	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_191_1:
	.loc	3	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	3	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	3	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_15_5122;
	.loc	3	213	0
	rcp.approx.ftz.f32 	%f14, %f9;
	mul.ftz.f32 	%f15, %f14, %f6;
	.loc	3	214	0
	mul.ftz.f32 	%f16, %f14, %f4;
	.loc	3	215	0
	mul.ftz.f32 	%f17, %f14, %f2;
	bra.uni 	$Lt_15_4866;
$Lt_15_5122:
	.loc	3	219	0
	mov.f32 	%f15, 0f00000000;    	// 0
	mov.f32 	%f16, 0f00000000;    	// 0
	mov.f32 	%f17, 0f00000000;    	// 0
	mov.f32 	%f10, 0f00000000;    	// 0
$Lt_15_4866:
	.loc	3	266	0
	mov.f32 	%f18, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p2, %f17, %f18;
	@!%p2 bra 	$Lt_15_5378;
	.loc	3	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_192_5;
$Lt_15_5378:
	.loc	3	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_192_5:
	.loc	3	267	0
	mov.f32 	%f28, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p3, %f16, %f28;
	@!%p3 bra 	$Lt_15_5890;
	.loc	3	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_192_3;
$Lt_15_5890:
	.loc	3	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_192_3:
	.loc	3	268	0
	mov.f32 	%f38, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p4, %f15, %f38;
	@!%p4 bra 	$Lt_15_6402;
	.loc	3	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_192_1;
$Lt_15_6402:
	.loc	3	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_192_1:
	.loc	3	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	3	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	3	254	0
	cvt.ftz.sat.f32.f32 	%f9, %f8;
	.loc	3	255	0
	mov.f32 	%f10, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p1, %f2, %f10;
	@!%p1 bra 	$Lt_16_4098;
	.loc	3	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	3	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	3	256	0
	mov.f32 	%f20, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p2, %f4, %f20;
	@!%p2 bra 	$Lt_16_4610;
	.loc	3	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	3	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	3	257	0
	mov.f32 	%f30, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p3, %f6, %f30;
	@!%p3 bra 	$Lt_16_5122;
	.loc	3	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	3	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	3	259	0
	mul.ftz.f32 	%f40, %f36, %f9;
	mul.ftz.f32 	%f41, %f26, %f9;
	.loc	3	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	3	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	3	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	3	213	0
	rcp.approx.ftz.f32 	%f14, %f9;
	mul.ftz.f32 	%f15, %f14, %f6;
	.loc	3	214	0
	mul.ftz.f32 	%f16, %f14, %f4;
	.loc	3	215	0
	mul.ftz.f32 	%f17, %f14, %f2;
	bra.uni 	$Lt_17_4866;
$Lt_17_5122:
	.loc	3	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	3	266	0
	mov.f32 	%f18, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p2, %f17, %f18;
	@!%p2 bra 	$Lt_17_5378;
	.loc	3	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	3	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	3	267	0
	mov.f32 	%f28, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p3, %f16, %f28;
	@!%p3 bra 	$Lt_17_5890;
	.loc	3	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	3	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	3	268	0
	mov.f32 	%f38, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p4, %f15, %f38;
	@!%p4 bra 	$Lt_17_6402;
	.loc	3	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	3	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	3	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	5	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	5	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_19_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_19_6914;
$Lt_19_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_19_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_19_7426;
$Lt_19_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_19_7426:
$Lt_19_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	5	75	0
	mov.f32 	%f54, 0f3e22f983;    	// 0.159155
	mul.ftz.f32 	%f55, %f53, %f54;
	st.f32 	[%rd2+0], %f55;
	.loc	5	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	5	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	5	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_20_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_20_9986;
$Lt_20_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_20_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_20_10498;
$Lt_20_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_20_10498:
$Lt_20_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	5	75	0
	mov.f32 	%f56, 0f3e22f983;    	// 0.159155
	mul.ftz.f32 	%f57, %f55, %f56;
	st.f32 	[%rd2+0], %f57;
	.loc	5	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	5	93	0
	st.f32 	[%rd2+8], %f6;
	.loc	5	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	5	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	5	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	5	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 (.param .f32 __cudaretf__Z11HueWrapNormf) _Z11HueWrapNormf (.param .f32 __cudaparmf1__Z11HueWrapNormf)
	{
	.reg .u32 %r<4>;
	.reg .f32 %f<12>;
	.reg .pred %p<4>;
	.loc	24	34	0
$LDWbegin__Z11HueWrapNormf:
	ld.param.f32 	%f1, [__cudaparmf1__Z11HueWrapNormf];
	mov.f32 	%f2, %f1;
	.loc	24	35	0
	mov.f32 	%f3, %f2;
	mov.f32 	%f4, 0f3f800000;     	// 1
	setp.gt.ftz.f32 	%p1, %f2, %f4;
	@!%p1 bra 	$Lt_22_2050;
	.loc	24	39	0
	cvt.rzi.ftz.s32.f32 	%r1, %f2;
	cvt.rn.f32.s32 	%f5, %r1;
	sub.ftz.f32 	%f3, %f2, %f5;
	bra.uni 	$Lt_22_1794;
$Lt_22_2050:
	mov.f32 	%f6, 0f00000000;     	// 0
	setp.lt.ftz.f32 	%p2, %f2, %f6;
	@!%p2 bra 	$Lt_22_2306;
	.loc	24	43	0
	cvt.rzi.ftz.s32.f32 	%r2, %f2;
	cvt.rn.f32.s32 	%f7, %r2;
	sub.ftz.f32 	%f8, %f2, %f7;
	mov.f32 	%f9, 0f3f800000;     	// 1
	add.ftz.f32 	%f3, %f8, %f9;
$Lt_22_2306:
$Lt_22_1794:
	.loc	24	46	0
	mov.f32 	%f10, %f3;
	st.param.f32 	[__cudaretf__Z11HueWrapNormf], %f10;
	ret;
$LDWend__Z11HueWrapNormf:
	} // _Z11HueWrapNormf

	.visible .func (.param .f32 __cudaretf__Z29CalculateThresholdRangeWeightffff) _Z29CalculateThresholdRangeWeightffff (.param .f32 __cudaparmf1__Z29CalculateThresholdRangeWeightffff, .param .f32 __cudaparmf2__Z29CalculateThresholdRangeWeightffff, .param .f32 __cudaparmf3__Z29CalculateThresholdRangeWeightffff, .param .f32 __cudaparmf4__Z29CalculateThresholdRangeWeightffff)
	{
	.reg .u32 %r<13>;
	.reg .f32 %f<24>;
	.reg .pred %p<9>;
	.loc	24	53	0
$LDWbegin__Z29CalculateThresholdRangeWeightffff:
	ld.param.f32 	%f1, [__cudaparmf1__Z29CalculateThresholdRangeWeightffff];
	mov.f32 	%f2, %f1;
	ld.param.f32 	%f3, [__cudaparmf2__Z29CalculateThresholdRangeWeightffff];
	mov.f32 	%f4, %f3;
	ld.param.f32 	%f5, [__cudaparmf3__Z29CalculateThresholdRangeWeightffff];
	mov.f32 	%f6, %f5;
	ld.param.f32 	%f7, [__cudaparmf4__Z29CalculateThresholdRangeWeightffff];
	mov.f32 	%f8, %f7;
	.loc	24	71	0
	sub.ftz.f32 	%f9, %f4, %f2;
	setp.le.ftz.f32 	%p1, %f2, %f4;
	@!%p1 bra 	$Lt_23_6402;
	mov.f32 	%f10, 0f358637bd;    	// 1e-006
	setp.lt.ftz.f32 	%p2, %f9, %f10;
	@!%p2 bra 	$L_23_4866;
	.loc	24	74	0
	mov.f32 	%f11, 0f00000000;    	// 0
	bra.uni 	$L_23_4610;
$Lt_23_6402:
$L_23_4866:
	.loc	24	77	0
	setp.gt.ftz.f32 	%p3, %f2, %f4;
	@!%p3 bra 	$Lt_23_6914;
	sub.ftz.f32 	%f12, %f2, %f4;
	mov.f32 	%f13, 0f358637bd;    	// 1e-006
	setp.lt.ftz.f32 	%p4, %f12, %f13;
	@!%p4 bra 	$L_23_5890;
	mov.s32 	%r1, 1;
	bra.uni 	$L_23_5634;
$Lt_23_6914:
$L_23_5890:
	mov.s32 	%r1, 0;
$L_23_5634:
	sub.ftz.f32 	%f14, %f8, %f2;
	mov.f32 	%f15, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p5, %f14, %f15;
	add.ftz.f32 	%f16, %f14, %f6;
	selp.f32 	%f17, %f16, %f14, %p5;
	add.ftz.f32 	%f18, %f9, %f6;
	mov.f32 	%f19, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p6, %f9, %f19;
	selp.f32 	%f20, %f18, %f9, %p6;
	set.ge.ftz.u32.f32 	%r2, %f20, %f17;
	neg.s32 	%r3, %r2;
	mov.f32 	%f21, 0f00000000;    	// 0
	set.ge.ftz.u32.f32 	%r4, %f17, %f21;
	neg.s32 	%r5, %r4;
	and.b32 	%r6, %r3, %r5;
	mov.u32 	%r7, 0;
	set.ne.u32.u32 	%r8, %r1, %r7;
	neg.s32 	%r9, %r8;
	or.b32 	%r10, %r9, %r6;
	mov.u32 	%r11, 0;
	setp.eq.s32 	%p7, %r10, %r11;
	@%p7 bra 	$Lt_23_7426;
	mov.f32 	%f11, 0f3f800000;    	// 1
	bra.uni 	$L_23_5378;
$Lt_23_7426:
	mov.f32 	%f11, 0f00000000;    	// 0
$L_23_5378:
$L_23_4610:
	.loc	24	87	0
	mov.f32 	%f22, %f11;
	st.param.f32 	[__cudaretf__Z29CalculateThresholdRangeWeightffff], %f22;
	ret;
$LDWend__Z29CalculateThresholdRangeWeightffff:
	} // _Z29CalculateThresholdRangeWeightffff

	.visible .func (.param .f32 __cudaretf__Z33CalculateStartSoftnessRangeWeightffffb) _Z33CalculateStartSoftnessRangeWeightffffb (.param .f32 __cudaparmf1__Z33CalculateStartSoftnessRangeWeightffffb, .param .f32 __cudaparmf2__Z33CalculateStartSoftnessRangeWeightffffb, .param .f32 __cudaparmf3__Z33CalculateStartSoftnessRangeWeightffffb, .param .f32 __cudaparmf4__Z33CalculateStartSoftnessRangeWeightffffb, .param .s32 __cudaparmf5__Z33CalculateStartSoftnessRangeWeightffffb)
	{
	.reg .u32 %r<16>;
	.reg .f32 %f<22>;
	.reg .pred %p<7>;
	.loc	24	95	0
$LDWbegin__Z33CalculateStartSoftnessRangeWeightffffb:
	ld.param.f32 	%f1, [__cudaparmf1__Z33CalculateStartSoftnessRangeWeightffffb];
	mov.f32 	%f2, %f1;
	ld.param.f32 	%f3, [__cudaparmf2__Z33CalculateStartSoftnessRangeWeightffffb];
	mov.f32 	%f4, %f3;
	ld.param.f32 	%f5, [__cudaparmf3__Z33CalculateStartSoftnessRangeWeightffffb];
	mov.f32 	%f6, %f5;
	ld.param.f32 	%f7, [__cudaparmf4__Z33CalculateStartSoftnessRangeWeightffffb];
	mov.f32 	%f8, %f7;
	ld.param.u32 	%r1, [__cudaparmf5__Z33CalculateStartSoftnessRangeWeightffffb];
	cvt.s8.s32 	%r2, %r1;
	.loc	24	108	0
	sub.ftz.f32 	%f9, %f8, %f4;
	mov.f32 	%f10, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p1, %f9, %f10;
	add.ftz.f32 	%f11, %f9, %f6;
	selp.f32 	%f12, %f11, %f9, %p1;
	mov.u32 	%r3, 0;
	setp.eq.s32 	%p2, %r2, %r3;
	@%p2 bra 	$Lt_24_4866;
	neg.ftz.f32 	%f13, %f4;
	setp.lt.ftz.f32 	%p3, %f12, %f13;
	@!%p3 bra 	$L_24_4354;
	mov.s32 	%r4, 1;
	bra.uni 	$L_24_4098;
$Lt_24_4866:
$L_24_4354:
	mov.s32 	%r4, 0;
$L_24_4098:
	sub.ftz.f32 	%f14, %f2, %f4;
	mov.f32 	%f15, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p4, %f14, %f15;
	add.ftz.f32 	%f16, %f14, %f6;
	selp.f32 	%f17, %f16, %f14, %p4;
	set.le.ftz.u32.f32 	%r5, %f17, %f12;
	neg.s32 	%r6, %r5;
	mov.f32 	%f18, 0f00000000;    	// 0
	set.lt.ftz.u32.f32 	%r7, %f12, %f18;
	neg.s32 	%r8, %r7;
	or.b32 	%r9, %r6, %r8;
	mov.u32 	%r10, 0;
	set.ne.u32.u32 	%r11, %r4, %r10;
	neg.s32 	%r12, %r11;
	or.b32 	%r13, %r12, %r9;
	mov.u32 	%r14, 0;
	setp.eq.s32 	%p5, %r13, %r14;
	@%p5 bra 	$L_24_3842;
	.loc	24	111	0
	mov.f32 	%f19, 0f00000000;    	// 0
	bra.uni 	$L_24_3586;
$L_24_3842:
	.loc	24	119	0
	div.approx.ftz.f32 	%f19, %f12, %f17;
$L_24_3586:
	.loc	24	122	0
	mov.f32 	%f20, %f19;
	st.param.f32 	[__cudaretf__Z33CalculateStartSoftnessRangeWeightffffb], %f20;
	ret;
$LDWend__Z33CalculateStartSoftnessRangeWeightffffb:
	} // _Z33CalculateStartSoftnessRangeWeightffffb

	.visible .func (.param .f32 __cudaretf__Z31CalculateEndSoftnessRangeWeightffffb) _Z31CalculateEndSoftnessRangeWeightffffb (.param .f32 __cudaparmf1__Z31CalculateEndSoftnessRangeWeightffffb, .param .f32 __cudaparmf2__Z31CalculateEndSoftnessRangeWeightffffb, .param .f32 __cudaparmf3__Z31CalculateEndSoftnessRangeWeightffffb, .param .f32 __cudaparmf4__Z31CalculateEndSoftnessRangeWeightffffb, .param .s32 __cudaparmf5__Z31CalculateEndSoftnessRangeWeightffffb)
	{
	.reg .u32 %r<16>;
	.reg .f32 %f<24>;
	.reg .pred %p<7>;
	.loc	24	130	0
$LDWbegin__Z31CalculateEndSoftnessRangeWeightffffb:
	ld.param.f32 	%f1, [__cudaparmf1__Z31CalculateEndSoftnessRangeWeightffffb];
	mov.f32 	%f2, %f1;
	ld.param.f32 	%f3, [__cudaparmf2__Z31CalculateEndSoftnessRangeWeightffffb];
	mov.f32 	%f4, %f3;
	ld.param.f32 	%f5, [__cudaparmf3__Z31CalculateEndSoftnessRangeWeightffffb];
	mov.f32 	%f6, %f5;
	ld.param.f32 	%f7, [__cudaparmf4__Z31CalculateEndSoftnessRangeWeightffffb];
	mov.f32 	%f8, %f7;
	ld.param.u32 	%r1, [__cudaparmf5__Z31CalculateEndSoftnessRangeWeightffffb];
	cvt.s8.s32 	%r2, %r1;
	.loc	24	143	0
	sub.ftz.f32 	%f9, %f8, %f2;
	mov.f32 	%f10, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p1, %f9, %f10;
	add.ftz.f32 	%f11, %f9, %f6;
	selp.f32 	%f12, %f11, %f9, %p1;
	mov.u32 	%r3, 0;
	setp.eq.s32 	%p2, %r2, %r3;
	@%p2 bra 	$Lt_25_4866;
	sub.ftz.f32 	%f13, %f6, %f2;
	setp.gt.ftz.f32 	%p3, %f12, %f13;
	@!%p3 bra 	$L_25_4354;
	mov.s32 	%r4, 1;
	bra.uni 	$L_25_4098;
$Lt_25_4866:
$L_25_4354:
	mov.s32 	%r4, 0;
$L_25_4098:
	sub.ftz.f32 	%f14, %f4, %f2;
	mov.f32 	%f15, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p4, %f14, %f15;
	add.ftz.f32 	%f16, %f14, %f6;
	selp.f32 	%f17, %f16, %f14, %p4;
	set.le.ftz.u32.f32 	%r5, %f17, %f12;
	neg.s32 	%r6, %r5;
	mov.f32 	%f18, 0f00000000;    	// 0
	set.lt.ftz.u32.f32 	%r7, %f12, %f18;
	neg.s32 	%r8, %r7;
	or.b32 	%r9, %r6, %r8;
	mov.u32 	%r10, 0;
	set.ne.u32.u32 	%r11, %r4, %r10;
	neg.s32 	%r12, %r11;
	or.b32 	%r13, %r12, %r9;
	mov.u32 	%r14, 0;
	setp.eq.s32 	%p5, %r13, %r14;
	@%p5 bra 	$L_25_3842;
	.loc	24	146	0
	mov.f32 	%f19, 0f00000000;    	// 0
	bra.uni 	$L_25_3586;
$L_25_3842:
	.loc	24	154	0
	mov.f32 	%f20, 0f3f800000;    	// 1
	div.approx.ftz.f32 	%f21, %f12, %f17;
	sub.ftz.f32 	%f19, %f20, %f21;
$L_25_3586:
	.loc	24	157	0
	mov.f32 	%f22, %f19;
	st.param.f32 	[__cudaretf__Z31CalculateEndSoftnessRangeWeightffffb], %f22;
	ret;
$LDWend__Z31CalculateEndSoftnessRangeWeightffffb:
	} // _Z31CalculateEndSoftnessRangeWeightffffb

	.visible .func (.param .f32 __cudaretf__Z20CalculateRangeWeightffffffb) _Z20CalculateRangeWeightffffffb (.param .f32 __cudaparmf1__Z20CalculateRangeWeightffffffb, .param .f32 __cudaparmf2__Z20CalculateRangeWeightffffffb, .param .f32 __cudaparmf3__Z20CalculateRangeWeightffffffb, .param .f32 __cudaparmf4__Z20CalculateRangeWeightffffffb, .param .f32 __cudaparmf5__Z20CalculateRangeWeightffffffb, .param .f32 __cudaparmf6__Z20CalculateRangeWeightffffffb, .param .s32 __cudaparmf7__Z20CalculateRangeWeightffffffb)
	{
	.reg .u32 %r<38>;
	.reg .f32 %f<57>;
	.reg .pred %p<20>;
	.loc	24	167	0
$LDWbegin__Z20CalculateRangeWeightffffffb:
	ld.param.f32 	%f1, [__cudaparmf1__Z20CalculateRangeWeightffffffb];
	mov.f32 	%f2, %f1;
	ld.param.f32 	%f3, [__cudaparmf2__Z20CalculateRangeWeightffffffb];
	mov.f32 	%f4, %f3;
	ld.param.f32 	%f5, [__cudaparmf3__Z20CalculateRangeWeightffffffb];
	mov.f32 	%f6, %f5;
	ld.param.f32 	%f7, [__cudaparmf4__Z20CalculateRangeWeightffffffb];
	mov.f32 	%f8, %f7;
	ld.param.f32 	%f9, [__cudaparmf5__Z20CalculateRangeWeightffffffb];
	mov.f32 	%f10, %f9;
	ld.param.f32 	%f11, [__cudaparmf6__Z20CalculateRangeWeightffffffb];
	mov.f32 	%f12, %f11;
	ld.param.u32 	%r1, [__cudaparmf7__Z20CalculateRangeWeightffffffb];
	cvt.s8.s32 	%r2, %r1;
	.loc	24	71	0
	sub.ftz.f32 	%f13, %f6, %f4;
	setp.le.ftz.f32 	%p1, %f4, %f6;
	@!%p1 bra 	$Lt_26_20226;
	mov.f32 	%f14, 0f358637bd;    	// 1e-006
	setp.lt.ftz.f32 	%p2, %f13, %f14;
	@!%p2 bra 	$Lt_26_20226;
	.loc	24	74	0
	mov.f32 	%f15, 0f00000000;    	// 0
	bra.uni 	$L_26_15618;
$Lt_26_20226:
$L_26_15106:
	.loc	24	77	0
	setp.gt.ftz.f32 	%p3, %f4, %f6;
	@!%p3 bra 	$Lt_26_20738;
	sub.ftz.f32 	%f16, %f4, %f6;
	mov.f32 	%f17, 0f358637bd;    	// 1e-006
	setp.lt.ftz.f32 	%p4, %f16, %f17;
	@!%p4 bra 	$Lt_26_20738;
	mov.s32 	%r3, 1;
	bra.uni 	$L_26_15874;
$Lt_26_20738:
$L_26_16130:
	mov.s32 	%r3, 0;
$L_26_15874:
	sub.ftz.f32 	%f18, %f12, %f4;
	mov.f32 	%f19, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p5, %f18, %f19;
	add.ftz.f32 	%f20, %f18, %f10;
	selp.f32 	%f21, %f20, %f18, %p5;
	add.ftz.f32 	%f22, %f13, %f10;
	mov.f32 	%f23, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p6, %f13, %f23;
	selp.f32 	%f24, %f22, %f13, %p6;
	set.ge.ftz.u32.f32 	%r4, %f24, %f21;
	neg.s32 	%r5, %r4;
	mov.f32 	%f25, 0f00000000;    	// 0
	set.ge.ftz.u32.f32 	%r6, %f21, %f25;
	neg.s32 	%r7, %r6;
	and.b32 	%r8, %r5, %r7;
	mov.u32 	%r9, 0;
	set.ne.u32.u32 	%r10, %r3, %r9;
	neg.s32 	%r11, %r10;
	or.b32 	%r12, %r11, %r8;
	mov.u32 	%r13, 0;
	setp.eq.s32 	%p7, %r12, %r13;
	@%p7 bra 	$Lt_26_21250;
	mov.f32 	%f15, 0f3f800000;    	// 1
	bra.uni 	$L_26_15618;
$Lt_26_21250:
	mov.f32 	%f15, 0f00000000;    	// 0
$L_26_15618:
$L_26_14850:
	.loc	24	173	0
	mov.f32 	%f26, 0f3f800000;    	// 1
	setp.lt.ftz.f32 	%p8, %f15, %f26;
	@!%p8 bra 	$Lt_26_19458;
	.loc	24	108	0
	mov.s32 	%r14, 0;
	setp.ne.s32 	%p9, %r2, %r14;
	sub.ftz.f32 	%f27, %f12, %f2;
	mov.f32 	%f28, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p10, %f27, %f28;
	add.ftz.f32 	%f29, %f27, %f10;
	selp.f32 	%f30, %f29, %f27, %p10;
	@!%p9 bra 	$Lt_26_21506;
	neg.ftz.f32 	%f31, %f2;
	setp.lt.ftz.f32 	%p11, %f30, %f31;
	@!%p11 bra 	$Lt_26_21506;
	mov.s32 	%r15, 1;
	bra.uni 	$L_26_17154;
$Lt_26_21506:
$L_26_17410:
	mov.s32 	%r15, 0;
$L_26_17154:
	sub.ftz.f32 	%f32, %f4, %f2;
	mov.f32 	%f33, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p12, %f32, %f33;
	add.ftz.f32 	%f34, %f32, %f10;
	selp.f32 	%f35, %f34, %f32, %p12;
	set.le.ftz.u32.f32 	%r16, %f35, %f30;
	neg.s32 	%r17, %r16;
	mov.f32 	%f36, 0f00000000;    	// 0
	set.lt.ftz.u32.f32 	%r18, %f30, %f36;
	neg.s32 	%r19, %r18;
	or.b32 	%r20, %r17, %r19;
	mov.u32 	%r21, 0;
	set.ne.u32.u32 	%r22, %r15, %r21;
	neg.s32 	%r23, %r22;
	or.b32 	%r24, %r23, %r20;
	mov.u32 	%r25, 0;
	setp.eq.s32 	%p13, %r24, %r25;
	@%p13 bra 	$L_26_16898;
	.loc	24	111	0
	mov.f32 	%f37, 0f00000000;    	// 0
	bra.uni 	$L_26_16642;
$L_26_16898:
	.loc	24	119	0
	div.approx.ftz.f32 	%f37, %f30, %f35;
$L_26_16642:
	.loc	24	180	0
	mov.f32 	%f38, %f37;
	mov.f32 	%f39, 0f3f800000;    	// 1
	setp.lt.ftz.f32 	%p14, %f37, %f39;
	@!%p14 bra 	$Lt_26_19970;
	.loc	24	143	0
	sub.ftz.f32 	%f40, %f12, %f6;
	mov.f32 	%f41, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p15, %f40, %f41;
	add.ftz.f32 	%f42, %f40, %f10;
	selp.f32 	%f43, %f42, %f40, %p15;
	@!%p9 bra 	$Lt_26_22018;
	sub.ftz.f32 	%f44, %f10, %f6;
	setp.gt.ftz.f32 	%p16, %f43, %f44;
	@!%p16 bra 	$Lt_26_22018;
	mov.s32 	%r26, 1;
	bra.uni 	$L_26_18434;
$Lt_26_22018:
$L_26_18690:
	mov.s32 	%r26, 0;
$L_26_18434:
	sub.ftz.f32 	%f45, %f8, %f6;
	mov.f32 	%f46, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p17, %f45, %f46;
	add.ftz.f32 	%f47, %f45, %f10;
	selp.f32 	%f48, %f47, %f45, %p17;
	set.le.ftz.u32.f32 	%r27, %f48, %f43;
	neg.s32 	%r28, %r27;
	mov.f32 	%f49, 0f00000000;    	// 0
	set.lt.ftz.u32.f32 	%r29, %f43, %f49;
	neg.s32 	%r30, %r29;
	or.b32 	%r31, %r28, %r30;
	mov.u32 	%r32, 0;
	set.ne.u32.u32 	%r33, %r26, %r32;
	neg.s32 	%r34, %r33;
	or.b32 	%r35, %r34, %r31;
	mov.u32 	%r36, 0;
	setp.eq.s32 	%p18, %r35, %r36;
	@%p18 bra 	$L_26_18178;
	.loc	24	146	0
	mov.f32 	%f50, 0f00000000;    	// 0
	bra.uni 	$L_26_17922;
$L_26_18178:
	.loc	24	154	0
	mov.f32 	%f51, 0f3f800000;    	// 1
	div.approx.ftz.f32 	%f52, %f43, %f48;
	sub.ftz.f32 	%f50, %f51, %f52;
$L_26_17922:
	.loc	24	188	0
	mov.f32 	%f53, %f50;
	bra.uni 	$Lt_26_19202;
$Lt_26_19970:
	mov.f32 	%f53, 0f00000000;    	// 0
	bra.uni 	$Lt_26_19202;
$Lt_26_19458:
	mov.f32 	%f53, 0f00000000;    	// 0
	mov.f32 	%f38, 0f00000000;    	// 0
$Lt_26_19202:
	.loc	24	202	0
	max.ftz.f32 	%f54, %f53, %f38;
	max.ftz.f32 	%f55, %f15, %f54;
	st.param.f32 	[__cudaretf__Z20CalculateRangeWeightffffffb], %f55;
	ret;
$LDWend__Z20CalculateRangeWeightffffffb:
	} // _Z20CalculateRangeWeightffffffb

	.visible .func (.param .f32 __cudaretf__Z3MaxIfET_S0_S0_) _Z3MaxIfET_S0_S0_ (.param .f32 __cudaparmf1__Z3MaxIfET_S0_S0_, .param .f32 __cudaparmf2__Z3MaxIfET_S0_S0_)
	{
	.reg .f32 %f<7>;
	.loc	23	54	0
$LDWbegin__Z3MaxIfET_S0_S0_:
	ld.param.f32 	%f1, [__cudaparmf1__Z3MaxIfET_S0_S0_];
	mov.f32 	%f2, %f1;
	ld.param.f32 	%f3, [__cudaparmf2__Z3MaxIfET_S0_S0_];
	mov.f32 	%f4, %f3;
	.loc	23	55	0
	max.ftz.f32 	%f5, %f2, %f4;
	st.param.f32 	[__cudaretf__Z3MaxIfET_S0_S0_], %f5;
	ret;
$LDWend__Z3MaxIfET_S0_S0_:
	} // _Z3MaxIfET_S0_S0_

	.visible .func (.param .f32 __cudaretf__Z18CalculateHueWeightfRK33SecondaryColorCorrectorParameters) _Z18CalculateHueWeightfRK33SecondaryColorCorrectorParameters (.param .f32 __cudaparmf1__Z18CalculateHueWeightfRK33SecondaryColorCorrectorParameters, .param .u64 __cudaparmf2__Z18CalculateHueWeightfRK33SecondaryColorCorrectorParameters)
	{
	.reg .u32 %r<31>;
	.reg .u64 %rd<4>;
	.reg .f32 %f<78>;
	.reg .pred %p<22>;
	.loc	24	207	0
$LDWbegin__Z18CalculateHueWeightfRK33SecondaryColorCorrectorParameters:
	ld.param.f32 	%f1, [__cudaparmf1__Z18CalculateHueWeightfRK33SecondaryColorCorrectorParameters];
	mov.f32 	%f2, %f1;
	ld.param.u64 	%rd1, [__cudaparmf2__Z18CalculateHueWeightfRK33SecondaryColorCorrectorParameters];
	mov.s64 	%rd2, %rd1;
	.loc	24	208	0
	ld.f32 	%f3, [%rd2+4];
	ld.f32 	%f4, [%rd2+0];
	sub.ftz.f32 	%f5, %f3, %f4;
	.loc	24	35	0
	mov.f32 	%f6, %f5;
	mov.f32 	%f7, 0f3f800000;     	// 1
	setp.gt.ftz.f32 	%p1, %f5, %f7;
	@!%p1 bra 	$Lt_28_21762;
	.loc	24	39	0
	cvt.rzi.ftz.s32.f32 	%r1, %f5;
	cvt.rn.f32.s32 	%f8, %r1;
	sub.ftz.f32 	%f6, %f5, %f8;
	bra.uni 	$Lt_28_22018;
$Lt_28_21762:
	mov.f32 	%f9, 0f00000000;     	// 0
	setp.lt.ftz.f32 	%p2, %f5, %f9;
	@!%p2 bra 	$Lt_28_22018;
	.loc	24	43	0
	cvt.rzi.ftz.s32.f32 	%r2, %f5;
	cvt.rn.f32.s32 	%f10, %r2;
	sub.ftz.f32 	%f11, %f5, %f10;
	mov.f32 	%f12, 0f3f800000;    	// 1
	add.ftz.f32 	%f6, %f11, %f12;
$Lt_28_22018:
$Lt_28_21506:
	.loc	24	208	0
	mov.f32 	%f13, %f6;
	.loc	24	209	0
	ld.f32 	%f14, [%rd2+12];
	ld.f32 	%f15, [%rd2+8];
	add.ftz.f32 	%f16, %f15, %f14;
	.loc	24	35	0
	mov.f32 	%f6, %f16;
	mov.f32 	%f17, 0f3f800000;    	// 1
	setp.gt.ftz.f32 	%p3, %f16, %f17;
	@!%p3 bra 	$Lt_28_22786;
	.loc	24	39	0
	cvt.rzi.ftz.s32.f32 	%r3, %f16;
	cvt.rn.f32.s32 	%f18, %r3;
	sub.ftz.f32 	%f6, %f16, %f18;
	bra.uni 	$Lt_28_23042;
$Lt_28_22786:
	mov.f32 	%f19, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p4, %f16, %f19;
	@!%p4 bra 	$Lt_28_23042;
	.loc	24	43	0
	cvt.rzi.ftz.s32.f32 	%r4, %f16;
	cvt.rn.f32.s32 	%f20, %r4;
	sub.ftz.f32 	%f21, %f16, %f20;
	mov.f32 	%f22, 0f3f800000;    	// 1
	add.ftz.f32 	%f6, %f21, %f22;
$Lt_28_23042:
$Lt_28_22530:
	.loc	24	71	0
	sub.ftz.f32 	%f23, %f14, %f3;
	setp.ge.ftz.f32 	%p5, %f14, %f3;
	@!%p5 bra 	$Lt_28_25602;
	mov.f32 	%f24, 0f358637bd;    	// 1e-006
	setp.lt.ftz.f32 	%p6, %f23, %f24;
	@!%p6 bra 	$Lt_28_25602;
	.loc	24	74	0
	mov.f32 	%f25, 0f00000000;    	// 0
	bra.uni 	$L_28_20482;
$Lt_28_25602:
$L_28_19970:
	.loc	24	77	0
	setp.lt.ftz.f32 	%p7, %f14, %f3;
	@!%p7 bra 	$Lt_28_26114;
	sub.ftz.f32 	%f26, %f3, %f14;
	mov.f32 	%f27, 0f358637bd;    	// 1e-006
	setp.lt.ftz.f32 	%p8, %f26, %f27;
	@!%p8 bra 	$Lt_28_26114;
	mov.s32 	%r5, 1;
	bra.uni 	$L_28_20738;
$Lt_28_26114:
$L_28_20994:
	mov.s32 	%r5, 0;
$L_28_20738:
	sub.ftz.f32 	%f28, %f2, %f3;
	mov.f32 	%f29, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p9, %f28, %f29;
	mov.f32 	%f30, 0f3f800000;    	// 1
	add.ftz.f32 	%f31, %f28, %f30;
	selp.f32 	%f32, %f31, %f28, %p9;
	mov.f32 	%f33, 0f3f800000;    	// 1
	add.ftz.f32 	%f34, %f23, %f33;
	mov.f32 	%f35, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p10, %f23, %f35;
	selp.f32 	%f36, %f34, %f23, %p10;
	set.ge.ftz.u32.f32 	%r6, %f36, %f32;
	neg.s32 	%r7, %r6;
	mov.f32 	%f37, 0f00000000;    	// 0
	set.ge.ftz.u32.f32 	%r8, %f32, %f37;
	neg.s32 	%r9, %r8;
	and.b32 	%r10, %r7, %r9;
	mov.u32 	%r11, 0;
	set.ne.u32.u32 	%r12, %r5, %r11;
	neg.s32 	%r13, %r12;
	or.b32 	%r14, %r13, %r10;
	mov.u32 	%r15, 0;
	setp.eq.s32 	%p11, %r14, %r15;
	@%p11 bra 	$Lt_28_26626;
	mov.f32 	%f25, 0f3f800000;    	// 1
	bra.uni 	$L_28_20482;
$Lt_28_26626:
	mov.f32 	%f25, 0f00000000;    	// 0
$L_28_20482:
$L_28_19714:
	.loc	24	173	0
	mov.f32 	%f38, 0f3f800000;    	// 1
	setp.lt.ftz.f32 	%p12, %f25, %f38;
	@!%p12 bra 	$Lt_28_23810;
	.loc	24	180	0
	sub.ftz.f32 	%f39, %f2, %f13;
	sub.ftz.f32 	%f40, %f3, %f13;
	mov.f32 	%f41, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p13, %f39, %f41;
	mov.f32 	%f42, 0f3f800000;    	// 1
	add.ftz.f32 	%f43, %f39, %f42;
	mov.f32 	%f44, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p14, %f40, %f44;
	mov.f32 	%f45, 0f3f800000;    	// 1
	add.ftz.f32 	%f46, %f40, %f45;
	selp.f32 	%f47, %f43, %f39, %p13;
	selp.f32 	%f48, %f46, %f40, %p14;
	set.ge.ftz.u32.f32 	%r16, %f47, %f48;
	neg.s32 	%r17, %r16;
	mov.f32 	%f49, 0f00000000;    	// 0
	set.lt.ftz.u32.f32 	%r18, %f47, %f49;
	neg.s32 	%r19, %r18;
	or.b32 	%r20, %r17, %r19;
	mov.u32 	%r21, 0;
	setp.eq.s32 	%p15, %r20, %r21;
	@%p15 bra 	$Lt_28_24322;
	mov.f32 	%f50, 0f00000000;    	// 0
	bra.uni 	$Lt_28_24066;
$Lt_28_24322:
	.loc	24	119	0
	div.approx.ftz.f32 	%f50, %f47, %f48;
$Lt_28_24066:
	.loc	24	180	0
	mov.f32 	%f51, %f50;
	mov.f32 	%f52, 0f3f800000;    	// 1
	setp.lt.ftz.f32 	%p16, %f50, %f52;
	@!%p16 bra 	$Lt_28_24834;
	.loc	24	188	0
	sub.ftz.f32 	%f53, %f2, %f14;
	sub.ftz.f32 	%f54, %f6, %f14;
	mov.f32 	%f55, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p17, %f53, %f55;
	mov.f32 	%f56, 0f3f800000;    	// 1
	add.ftz.f32 	%f57, %f53, %f56;
	mov.f32 	%f58, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p18, %f54, %f58;
	mov.f32 	%f59, 0f3f800000;    	// 1
	add.ftz.f32 	%f60, %f54, %f59;
	selp.f32 	%f61, %f57, %f53, %p17;
	selp.f32 	%f62, %f60, %f54, %p18;
	set.ge.ftz.u32.f32 	%r22, %f61, %f62;
	neg.s32 	%r23, %r22;
	mov.f32 	%f63, 0f00000000;    	// 0
	set.lt.ftz.u32.f32 	%r24, %f61, %f63;
	neg.s32 	%r25, %r24;
	or.b32 	%r26, %r23, %r25;
	mov.u32 	%r27, 0;
	setp.eq.s32 	%p19, %r26, %r27;
	@%p19 bra 	$Lt_28_25346;
	mov.f32 	%f64, 0f00000000;    	// 0
	bra.uni 	$Lt_28_25090;
$Lt_28_25346:
	.loc	24	154	0
	mov.f32 	%f65, 0f3f800000;    	// 1
	div.approx.ftz.f32 	%f66, %f61, %f62;
	sub.ftz.f32 	%f64, %f65, %f66;
$Lt_28_25090:
	.loc	24	188	0
	mov.f32 	%f67, %f64;
	bra.uni 	$Lt_28_23554;
$Lt_28_24834:
	mov.f32 	%f67, 0f00000000;    	// 0
	bra.uni 	$Lt_28_23554;
$Lt_28_23810:
	mov.f32 	%f67, 0f00000000;    	// 0
	mov.f32 	%f51, 0f00000000;    	// 0
$Lt_28_23554:
	.loc	24	230	0
	max.ftz.f32 	%f68, %f67, %f51;
	max.ftz.f32 	%f69, %f68, %f25;
	mov.f32 	%f70, 0f3f800000;    	// 1
	sub.ftz.f32 	%f71, %f70, %f69;
	ld.s8 	%r28, [%rd2+60];
	mov.s32 	%r29, 0;
	setp.ne.s32 	%p20, %r28, %r29;
	selp.f32 	%f72, %f71, %f69, %p20;
	mov.f32 	%f73, 0f00000000;    	// 0
	max.ftz.f32 	%f74, %f72, %f73;
	mov.f32 	%f75, 0f3f800000;    	// 1
	min.ftz.f32 	%f76, %f74, %f75;
	st.param.f32 	[__cudaretf__Z18CalculateHueWeightfRK33SecondaryColorCorrectorParameters], %f76;
	ret;
$LDWend__Z18CalculateHueWeightfRK33SecondaryColorCorrectorParameters:
	} // _Z18CalculateHueWeightfRK33SecondaryColorCorrectorParameters

	.visible .func (.param .f32 __cudaretf__Z25CalculateSaturationWeightfRK33SecondaryColorCorrectorParameters) _Z25CalculateSaturationWeightfRK33SecondaryColorCorrectorParameters (.param .f32 __cudaparmf1__Z25CalculateSaturationWeightfRK33SecondaryColorCorrectorParameters, .param .u64 __cudaparmf2__Z25CalculateSaturationWeightfRK33SecondaryColorCorrectorParameters)
	{
	.reg .u32 %r<33>;
	.reg .u64 %rd<4>;
	.reg .f32 %f<67>;
	.reg .pred %p<18>;
	.loc	24	235	0
$LDWbegin__Z25CalculateSaturationWeightfRK33SecondaryColorCorrectorParameters:
	ld.param.f32 	%f1, [__cudaparmf1__Z25CalculateSaturationWeightfRK33SecondaryColorCorrectorParameters];
	mov.f32 	%f2, %f1;
	ld.param.u64 	%rd1, [__cudaparmf2__Z25CalculateSaturationWeightfRK33SecondaryColorCorrectorParameters];
	mov.s64 	%rd2, %rd1;
	.loc	24	236	0
	ld.f32 	%f3, [%rd2+20];
	ld.f32 	%f4, [%rd2+16];
	sub.ftz.f32 	%f5, %f3, %f4;
	ld.f32 	%f6, [%rd2+28];
	ld.f32 	%f7, [%rd2+24];
	add.ftz.f32 	%f8, %f7, %f6;
	.loc	24	71	0
	sub.ftz.f32 	%f9, %f6, %f3;
	setp.ge.ftz.f32 	%p1, %f6, %f3;
	@!%p1 bra 	$Lt_29_20482;
	mov.f32 	%f10, 0f358637bd;    	// 1e-006
	setp.lt.ftz.f32 	%p2, %f9, %f10;
	@!%p2 bra 	$Lt_29_20482;
	.loc	24	74	0
	mov.f32 	%f11, 0f00000000;    	// 0
	bra.uni 	$L_29_17410;
$Lt_29_20482:
$L_29_16898:
	.loc	24	77	0
	setp.lt.ftz.f32 	%p3, %f6, %f3;
	@!%p3 bra 	$Lt_29_20994;
	sub.ftz.f32 	%f12, %f3, %f6;
	mov.f32 	%f13, 0f358637bd;    	// 1e-006
	setp.lt.ftz.f32 	%p4, %f12, %f13;
	@!%p4 bra 	$Lt_29_20994;
	mov.s32 	%r1, 1;
	bra.uni 	$L_29_17666;
$Lt_29_20994:
$L_29_17922:
	mov.s32 	%r1, 0;
$L_29_17666:
	sub.ftz.f32 	%f14, %f2, %f3;
	mov.f32 	%f15, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p5, %f14, %f15;
	mov.f32 	%f16, 0f3f800000;    	// 1
	add.ftz.f32 	%f17, %f14, %f16;
	selp.f32 	%f18, %f17, %f14, %p5;
	mov.f32 	%f19, 0f3f800000;    	// 1
	add.ftz.f32 	%f20, %f9, %f19;
	mov.f32 	%f21, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p6, %f9, %f21;
	selp.f32 	%f22, %f20, %f9, %p6;
	set.ge.ftz.u32.f32 	%r2, %f22, %f18;
	neg.s32 	%r3, %r2;
	mov.f32 	%f23, 0f00000000;    	// 0
	set.ge.ftz.u32.f32 	%r4, %f18, %f23;
	neg.s32 	%r5, %r4;
	and.b32 	%r6, %r3, %r5;
	mov.u32 	%r7, 0;
	set.ne.u32.u32 	%r8, %r1, %r7;
	neg.s32 	%r9, %r8;
	or.b32 	%r10, %r9, %r6;
	mov.u32 	%r11, 0;
	setp.eq.s32 	%p7, %r10, %r11;
	@%p7 bra 	$Lt_29_21506;
	mov.f32 	%f11, 0f3f800000;    	// 1
	bra.uni 	$L_29_17410;
$Lt_29_21506:
	mov.f32 	%f11, 0f00000000;    	// 0
$L_29_17410:
$L_29_16642:
	.loc	24	173	0
	mov.f32 	%f24, 0f3f800000;    	// 1
	setp.lt.ftz.f32 	%p8, %f11, %f24;
	@!%p8 bra 	$Lt_29_19714;
	.loc	24	108	0
	sub.ftz.f32 	%f25, %f2, %f5;
	sub.ftz.f32 	%f26, %f3, %f5;
	mov.f32 	%f27, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p9, %f25, %f27;
	mov.f32 	%f28, 0f3f800000;    	// 1
	add.ftz.f32 	%f29, %f25, %f28;
	mov.f32 	%f30, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p10, %f26, %f30;
	mov.f32 	%f31, 0f3f800000;    	// 1
	add.ftz.f32 	%f32, %f26, %f31;
	selp.f32 	%f33, %f29, %f25, %p9;
	selp.f32 	%f34, %f32, %f26, %p10;
	neg.ftz.f32 	%f35, %f5;
	set.lt.ftz.u32.f32 	%r12, %f33, %f35;
	neg.s32 	%r13, %r12;
	set.ge.ftz.u32.f32 	%r14, %f33, %f34;
	neg.s32 	%r15, %r14;
	mov.f32 	%f36, 0f00000000;    	// 0
	set.lt.ftz.u32.f32 	%r16, %f33, %f36;
	neg.s32 	%r17, %r16;
	or.b32 	%r18, %r15, %r17;
	or.b32 	%r19, %r13, %r18;
	mov.u32 	%r20, 0;
	setp.eq.s32 	%p11, %r19, %r20;
	@%p11 bra 	$L_29_18690;
	.loc	24	111	0
	mov.f32 	%f37, 0f00000000;    	// 0
	bra.uni 	$L_29_18434;
$L_29_18690:
	.loc	24	119	0
	div.approx.ftz.f32 	%f37, %f33, %f34;
$L_29_18434:
	.loc	24	180	0
	mov.f32 	%f38, %f37;
	mov.f32 	%f39, 0f3f800000;    	// 1
	setp.lt.ftz.f32 	%p12, %f37, %f39;
	@!%p12 bra 	$Lt_29_20226;
	.loc	24	143	0
	sub.ftz.f32 	%f40, %f2, %f6;
	sub.ftz.f32 	%f41, %f8, %f6;
	mov.f32 	%f42, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p13, %f40, %f42;
	mov.f32 	%f43, 0f3f800000;    	// 1
	add.ftz.f32 	%f44, %f40, %f43;
	mov.f32 	%f45, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p14, %f41, %f45;
	mov.f32 	%f46, 0f3f800000;    	// 1
	add.ftz.f32 	%f47, %f41, %f46;
	selp.f32 	%f48, %f44, %f40, %p13;
	selp.f32 	%f49, %f47, %f41, %p14;
	mov.f32 	%f50, 0f3f800000;    	// 1
	sub.ftz.f32 	%f51, %f50, %f6;
	set.gt.ftz.u32.f32 	%r21, %f48, %f51;
	neg.s32 	%r22, %r21;
	set.ge.ftz.u32.f32 	%r23, %f48, %f49;
	neg.s32 	%r24, %r23;
	mov.f32 	%f52, 0f00000000;    	// 0
	set.lt.ftz.u32.f32 	%r25, %f48, %f52;
	neg.s32 	%r26, %r25;
	or.b32 	%r27, %r24, %r26;
	or.b32 	%r28, %r22, %r27;
	mov.u32 	%r29, 0;
	setp.eq.s32 	%p15, %r28, %r29;
	@%p15 bra 	$L_29_19202;
	.loc	24	146	0
	mov.f32 	%f53, 0f00000000;    	// 0
	bra.uni 	$L_29_18946;
$L_29_19202:
	.loc	24	154	0
	mov.f32 	%f54, 0f3f800000;    	// 1
	div.approx.ftz.f32 	%f55, %f48, %f49;
	sub.ftz.f32 	%f53, %f54, %f55;
$L_29_18946:
	.loc	24	188	0
	mov.f32 	%f56, %f53;
	bra.uni 	$Lt_29_19458;
$Lt_29_20226:
	mov.f32 	%f56, 0f00000000;    	// 0
	bra.uni 	$Lt_29_19458;
$Lt_29_19714:
	mov.f32 	%f56, 0f00000000;    	// 0
	mov.f32 	%f38, 0f00000000;    	// 0
$Lt_29_19458:
	.loc	24	255	0
	max.ftz.f32 	%f57, %f56, %f38;
	max.ftz.f32 	%f58, %f57, %f11;
	mov.f32 	%f59, 0f3f800000;    	// 1
	sub.ftz.f32 	%f60, %f59, %f58;
	ld.s8 	%r30, [%rd2+60];
	mov.s32 	%r31, 0;
	setp.ne.s32 	%p16, %r30, %r31;
	selp.f32 	%f61, %f60, %f58, %p16;
	mov.f32 	%f62, 0f00000000;    	// 0
	max.ftz.f32 	%f63, %f61, %f62;
	mov.f32 	%f64, 0f3f800000;    	// 1
	min.ftz.f32 	%f65, %f63, %f64;
	st.param.f32 	[__cudaretf__Z25CalculateSaturationWeightfRK33SecondaryColorCorrectorParameters], %f65;
	ret;
$LDWend__Z25CalculateSaturationWeightfRK33SecondaryColorCorrectorParameters:
	} // _Z25CalculateSaturationWeightfRK33SecondaryColorCorrectorParameters

	.visible .func (.param .f32 __cudaretf__Z19CalculateLumaWeightfRK33SecondaryColorCorrectorParameters) _Z19CalculateLumaWeightfRK33SecondaryColorCorrectorParameters (.param .f32 __cudaparmf1__Z19CalculateLumaWeightfRK33SecondaryColorCorrectorParameters, .param .u64 __cudaparmf2__Z19CalculateLumaWeightfRK33SecondaryColorCorrectorParameters)
	{
	.reg .u32 %r<33>;
	.reg .u64 %rd<4>;
	.reg .f32 %f<67>;
	.reg .pred %p<18>;
	.loc	24	260	0
$LDWbegin__Z19CalculateLumaWeightfRK33SecondaryColorCorrectorParameters:
	ld.param.f32 	%f1, [__cudaparmf1__Z19CalculateLumaWeightfRK33SecondaryColorCorrectorParameters];
	mov.f32 	%f2, %f1;
	ld.param.u64 	%rd1, [__cudaparmf2__Z19CalculateLumaWeightfRK33SecondaryColorCorrectorParameters];
	mov.s64 	%rd2, %rd1;
	.loc	24	261	0
	ld.f32 	%f3, [%rd2+36];
	ld.f32 	%f4, [%rd2+32];
	sub.ftz.f32 	%f5, %f3, %f4;
	ld.f32 	%f6, [%rd2+44];
	ld.f32 	%f7, [%rd2+40];
	add.ftz.f32 	%f8, %f7, %f6;
	.loc	24	71	0
	sub.ftz.f32 	%f9, %f6, %f3;
	setp.ge.ftz.f32 	%p1, %f6, %f3;
	@!%p1 bra 	$Lt_30_20482;
	mov.f32 	%f10, 0f358637bd;    	// 1e-006
	setp.lt.ftz.f32 	%p2, %f9, %f10;
	@!%p2 bra 	$Lt_30_20482;
	.loc	24	74	0
	mov.f32 	%f11, 0f00000000;    	// 0
	bra.uni 	$L_30_17410;
$Lt_30_20482:
$L_30_16898:
	.loc	24	77	0
	setp.lt.ftz.f32 	%p3, %f6, %f3;
	@!%p3 bra 	$Lt_30_20994;
	sub.ftz.f32 	%f12, %f3, %f6;
	mov.f32 	%f13, 0f358637bd;    	// 1e-006
	setp.lt.ftz.f32 	%p4, %f12, %f13;
	@!%p4 bra 	$Lt_30_20994;
	mov.s32 	%r1, 1;
	bra.uni 	$L_30_17666;
$Lt_30_20994:
$L_30_17922:
	mov.s32 	%r1, 0;
$L_30_17666:
	sub.ftz.f32 	%f14, %f2, %f3;
	mov.f32 	%f15, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p5, %f14, %f15;
	mov.f32 	%f16, 0f3f800000;    	// 1
	add.ftz.f32 	%f17, %f14, %f16;
	selp.f32 	%f18, %f17, %f14, %p5;
	mov.f32 	%f19, 0f3f800000;    	// 1
	add.ftz.f32 	%f20, %f9, %f19;
	mov.f32 	%f21, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p6, %f9, %f21;
	selp.f32 	%f22, %f20, %f9, %p6;
	set.ge.ftz.u32.f32 	%r2, %f22, %f18;
	neg.s32 	%r3, %r2;
	mov.f32 	%f23, 0f00000000;    	// 0
	set.ge.ftz.u32.f32 	%r4, %f18, %f23;
	neg.s32 	%r5, %r4;
	and.b32 	%r6, %r3, %r5;
	mov.u32 	%r7, 0;
	set.ne.u32.u32 	%r8, %r1, %r7;
	neg.s32 	%r9, %r8;
	or.b32 	%r10, %r9, %r6;
	mov.u32 	%r11, 0;
	setp.eq.s32 	%p7, %r10, %r11;
	@%p7 bra 	$Lt_30_21506;
	mov.f32 	%f11, 0f3f800000;    	// 1
	bra.uni 	$L_30_17410;
$Lt_30_21506:
	mov.f32 	%f11, 0f00000000;    	// 0
$L_30_17410:
$L_30_16642:
	.loc	24	173	0
	mov.f32 	%f24, 0f3f800000;    	// 1
	setp.lt.ftz.f32 	%p8, %f11, %f24;
	@!%p8 bra 	$Lt_30_19714;
	.loc	24	108	0
	sub.ftz.f32 	%f25, %f2, %f5;
	sub.ftz.f32 	%f26, %f3, %f5;
	mov.f32 	%f27, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p9, %f25, %f27;
	mov.f32 	%f28, 0f3f800000;    	// 1
	add.ftz.f32 	%f29, %f25, %f28;
	mov.f32 	%f30, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p10, %f26, %f30;
	mov.f32 	%f31, 0f3f800000;    	// 1
	add.ftz.f32 	%f32, %f26, %f31;
	selp.f32 	%f33, %f29, %f25, %p9;
	selp.f32 	%f34, %f32, %f26, %p10;
	neg.ftz.f32 	%f35, %f5;
	set.lt.ftz.u32.f32 	%r12, %f33, %f35;
	neg.s32 	%r13, %r12;
	set.ge.ftz.u32.f32 	%r14, %f33, %f34;
	neg.s32 	%r15, %r14;
	mov.f32 	%f36, 0f00000000;    	// 0
	set.lt.ftz.u32.f32 	%r16, %f33, %f36;
	neg.s32 	%r17, %r16;
	or.b32 	%r18, %r15, %r17;
	or.b32 	%r19, %r13, %r18;
	mov.u32 	%r20, 0;
	setp.eq.s32 	%p11, %r19, %r20;
	@%p11 bra 	$L_30_18690;
	.loc	24	111	0
	mov.f32 	%f37, 0f00000000;    	// 0
	bra.uni 	$L_30_18434;
$L_30_18690:
	.loc	24	119	0
	div.approx.ftz.f32 	%f37, %f33, %f34;
$L_30_18434:
	.loc	24	180	0
	mov.f32 	%f38, %f37;
	mov.f32 	%f39, 0f3f800000;    	// 1
	setp.lt.ftz.f32 	%p12, %f37, %f39;
	@!%p12 bra 	$Lt_30_20226;
	.loc	24	143	0
	sub.ftz.f32 	%f40, %f2, %f6;
	sub.ftz.f32 	%f41, %f8, %f6;
	mov.f32 	%f42, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p13, %f40, %f42;
	mov.f32 	%f43, 0f3f800000;    	// 1
	add.ftz.f32 	%f44, %f40, %f43;
	mov.f32 	%f45, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p14, %f41, %f45;
	mov.f32 	%f46, 0f3f800000;    	// 1
	add.ftz.f32 	%f47, %f41, %f46;
	selp.f32 	%f48, %f44, %f40, %p13;
	selp.f32 	%f49, %f47, %f41, %p14;
	mov.f32 	%f50, 0f3f800000;    	// 1
	sub.ftz.f32 	%f51, %f50, %f6;
	set.gt.ftz.u32.f32 	%r21, %f48, %f51;
	neg.s32 	%r22, %r21;
	set.ge.ftz.u32.f32 	%r23, %f48, %f49;
	neg.s32 	%r24, %r23;
	mov.f32 	%f52, 0f00000000;    	// 0
	set.lt.ftz.u32.f32 	%r25, %f48, %f52;
	neg.s32 	%r26, %r25;
	or.b32 	%r27, %r24, %r26;
	or.b32 	%r28, %r22, %r27;
	mov.u32 	%r29, 0;
	setp.eq.s32 	%p15, %r28, %r29;
	@%p15 bra 	$L_30_19202;
	.loc	24	146	0
	mov.f32 	%f53, 0f00000000;    	// 0
	bra.uni 	$L_30_18946;
$L_30_19202:
	.loc	24	154	0
	mov.f32 	%f54, 0f3f800000;    	// 1
	div.approx.ftz.f32 	%f55, %f48, %f49;
	sub.ftz.f32 	%f53, %f54, %f55;
$L_30_18946:
	.loc	24	188	0
	mov.f32 	%f56, %f53;
	bra.uni 	$Lt_30_19458;
$Lt_30_20226:
	mov.f32 	%f56, 0f00000000;    	// 0
	bra.uni 	$Lt_30_19458;
$Lt_30_19714:
	mov.f32 	%f56, 0f00000000;    	// 0
	mov.f32 	%f38, 0f00000000;    	// 0
$Lt_30_19458:
	.loc	24	280	0
	max.ftz.f32 	%f57, %f56, %f38;
	max.ftz.f32 	%f58, %f57, %f11;
	mov.f32 	%f59, 0f3f800000;    	// 1
	sub.ftz.f32 	%f60, %f59, %f58;
	ld.s8 	%r30, [%rd2+60];
	mov.s32 	%r31, 0;
	setp.ne.s32 	%p16, %r30, %r31;
	selp.f32 	%f61, %f60, %f58, %p16;
	mov.f32 	%f62, 0f00000000;    	// 0
	max.ftz.f32 	%f63, %f61, %f62;
	mov.f32 	%f64, 0f3f800000;    	// 1
	min.ftz.f32 	%f65, %f63, %f64;
	st.param.f32 	[__cudaretf__Z19CalculateLumaWeightfRK33SecondaryColorCorrectorParameters], %f65;
	ret;
$LDWend__Z19CalculateLumaWeightfRK33SecondaryColorCorrectorParameters:
	} // _Z19CalculateLumaWeightfRK33SecondaryColorCorrectorParameters

	.visible .func (.param .f32 __cudaretf__Z15CalculateWeight8PixelYUVRK33SecondaryColorCorrectorParameters) _Z15CalculateWeight8PixelYUVRK33SecondaryColorCorrectorParameters (.param .align 16 .b8 __cudaparmf1__Z15CalculateWeight8PixelYUVRK33SecondaryColorCorrectorParameters[16], .param .u64 __cudaparmf2__Z15CalculateWeight8PixelYUVRK33SecondaryColorCorrectorParameters)
	{
	.reg .u32 %r<121>;
	.reg .u64 %rd<4>;
	.reg .f32 %f<319>;
	.reg .pred %p<67>;
	.loc	24	289	0
$LDWbegin__Z15CalculateWeight8PixelYUVRK33SecondaryColorCorrectorParameters:
	ld.param.f32 	%f1, [__cudaparmf1__Z15CalculateWeight8PixelYUVRK33SecondaryColorCorrectorParameters+4];
	mov.f32 	%f2, %f1;
	ld.param.f32 	%f3, [__cudaparmf1__Z15CalculateWeight8PixelYUVRK33SecondaryColorCorrectorParameters+8];
	mov.f32 	%f4, %f3;
	ld.param.f32 	%f5, [__cudaparmf1__Z15CalculateWeight8PixelYUVRK33SecondaryColorCorrectorParameters+12];
	mov.f32 	%f6, %f5;
	ld.param.u64 	%rd1, [__cudaparmf2__Z15CalculateWeight8PixelYUVRK33SecondaryColorCorrectorParameters];
	mov.s64 	%rd2, %rd1;
	.loc	5	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_31_74754;
	.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_31_75010;
$Lt_31_74754:
	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_31_75266;
	.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_31_75010;
$Lt_31_75266:
	.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_31_75010:
$Lt_31_74498:
	.loc	24	295	0
	ld.s8 	%r29, [%rd2+61];
	mov.u32 	%r30, 0;
	setp.eq.s32 	%p8, %r29, %r30;
	@%p8 bra 	$Lt_31_75778;
	mov.f32 	%f52, 0f3f800000;    	// 1
	bra.uni 	$Lt_31_75522;
$Lt_31_75778:
	.loc	24	208	0
	ld.f32 	%f53, [%rd2+4];
	ld.f32 	%f54, [%rd2+0];
	sub.ftz.f32 	%f55, %f53, %f54;
	.loc	24	35	0
	mov.f32 	%f56, %f55;
	mov.f32 	%f57, 0f3f800000;    	// 1
	setp.gt.ftz.f32 	%p9, %f55, %f57;
	@!%p9 bra 	$Lt_31_76290;
	.loc	24	39	0
	cvt.rzi.ftz.s32.f32 	%r31, %f55;
	cvt.rn.f32.s32 	%f58, %r31;
	sub.ftz.f32 	%f56, %f55, %f58;
	bra.uni 	$Lt_31_76546;
$Lt_31_76290:
	mov.f32 	%f59, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p10, %f55, %f59;
	@!%p10 bra 	$Lt_31_76546;
	.loc	24	43	0
	cvt.rzi.ftz.s32.f32 	%r32, %f55;
	cvt.rn.f32.s32 	%f60, %r32;
	sub.ftz.f32 	%f61, %f55, %f60;
	mov.f32 	%f62, 0f3f800000;    	// 1
	add.ftz.f32 	%f56, %f61, %f62;
$Lt_31_76546:
$Lt_31_76034:
	.loc	24	208	0
	mov.f32 	%f63, %f56;
	.loc	24	209	0
	ld.f32 	%f64, [%rd2+12];
	ld.f32 	%f65, [%rd2+8];
	add.ftz.f32 	%f66, %f65, %f64;
	.loc	24	35	0
	mov.f32 	%f56, %f66;
	mov.f32 	%f67, 0f3f800000;    	// 1
	setp.gt.ftz.f32 	%p11, %f66, %f67;
	@!%p11 bra 	$Lt_31_77314;
	.loc	24	39	0
	cvt.rzi.ftz.s32.f32 	%r33, %f66;
	cvt.rn.f32.s32 	%f68, %r33;
	sub.ftz.f32 	%f56, %f66, %f68;
	bra.uni 	$Lt_31_77570;
$Lt_31_77314:
	mov.f32 	%f69, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p12, %f66, %f69;
	@!%p12 bra 	$Lt_31_77570;
	.loc	24	43	0
	cvt.rzi.ftz.s32.f32 	%r34, %f66;
	cvt.rn.f32.s32 	%f70, %r34;
	sub.ftz.f32 	%f71, %f66, %f70;
	mov.f32 	%f72, 0f3f800000;    	// 1
	add.ftz.f32 	%f56, %f71, %f72;
$Lt_31_77570:
$Lt_31_77058:
	.loc	24	71	0
	sub.ftz.f32 	%f73, %f64, %f53;
	setp.ge.ftz.f32 	%p13, %f64, %f53;
	@!%p13 bra 	$Lt_31_83714;
	mov.f32 	%f74, 0f358637bd;    	// 1e-006
	setp.lt.ftz.f32 	%p14, %f73, %f74;
	@!%p14 bra 	$Lt_31_83714;
	.loc	24	74	0
	mov.f32 	%f75, 0f00000000;    	// 0
	bra.uni 	$L_31_67842;
$Lt_31_83714:
$L_31_67330:
	.loc	24	77	0
	setp.lt.ftz.f32 	%p15, %f64, %f53;
	@!%p15 bra 	$Lt_31_84226;
	sub.ftz.f32 	%f76, %f53, %f64;
	mov.f32 	%f77, 0f358637bd;    	// 1e-006
	setp.lt.ftz.f32 	%p16, %f76, %f77;
	@!%p16 bra 	$Lt_31_84226;
	mov.s32 	%r35, 1;
	bra.uni 	$L_31_68098;
$Lt_31_84226:
$L_31_68354:
	mov.s32 	%r35, 0;
$L_31_68098:
	mov.f32 	%f78, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p17, %f20, %f78;
	mov.f32 	%f79, 0f40c90fdb;    	// 6.28319
	add.ftz.f32 	%f80, %f20, %f79;
	selp.f32 	%f81, %f80, %f20, %p17;
	mov.f32 	%f82, 0f3e22f983;    	// 0.159155
	mul.ftz.f32 	%f83, %f81, %f82;
	mov.f32 	%f84, 0f00000000;    	// 0
	max.ftz.f32 	%f85, %f83, %f84;
	mov.f32 	%f86, 0f3f800000;    	// 1
	min.ftz.f32 	%f87, %f85, %f86;
	sub.ftz.f32 	%f88, %f87, %f53;
	mov.f32 	%f89, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p18, %f88, %f89;
	mov.f32 	%f90, 0f3f800000;    	// 1
	add.ftz.f32 	%f91, %f88, %f90;
	selp.f32 	%f92, %f91, %f88, %p18;
	mov.f32 	%f93, 0f3f800000;    	// 1
	add.ftz.f32 	%f94, %f73, %f93;
	mov.f32 	%f95, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p19, %f73, %f95;
	selp.f32 	%f96, %f94, %f73, %p19;
	set.ge.ftz.u32.f32 	%r36, %f96, %f92;
	neg.s32 	%r37, %r36;
	mov.f32 	%f97, 0f00000000;    	// 0
	set.ge.ftz.u32.f32 	%r38, %f92, %f97;
	neg.s32 	%r39, %r38;
	and.b32 	%r40, %r37, %r39;
	mov.u32 	%r41, 0;
	set.ne.u32.u32 	%r42, %r35, %r41;
	neg.s32 	%r43, %r42;
	or.b32 	%r44, %r43, %r40;
	mov.u32 	%r45, 0;
	setp.eq.s32 	%p20, %r44, %r45;
	@%p20 bra 	$Lt_31_84738;
	mov.f32 	%f75, 0f3f800000;    	// 1
	bra.uni 	$L_31_67842;
$Lt_31_84738:
	mov.f32 	%f75, 0f00000000;    	// 0
$L_31_67842:
$L_31_67074:
	.loc	24	173	0
	mov.f32 	%f98, 0f3f800000;    	// 1
	setp.lt.ftz.f32 	%p21, %f75, %f98;
	@!%p21 bra 	$Lt_31_78338;
	.loc	24	180	0
	mov.f32 	%f99, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p22, %f20, %f99;
	mov.f32 	%f100, 0f40c90fdb;   	// 6.28319
	add.ftz.f32 	%f101, %f20, %f100;
	sub.ftz.f32 	%f102, %f53, %f63;
	selp.f32 	%f103, %f101, %f20, %p22;
	mov.f32 	%f104, 0f00000000;   	// 0
	setp.lt.ftz.f32 	%p23, %f102, %f104;
	mov.f32 	%f105, 0f3f800000;   	// 1
	add.ftz.f32 	%f106, %f102, %f105;
	mov.f32 	%f107, 0f3e22f983;   	// 0.159155
	mul.ftz.f32 	%f108, %f103, %f107;
	selp.f32 	%f109, %f106, %f102, %p23;
	mov.f32 	%f110, 0f00000000;   	// 0
	max.ftz.f32 	%f111, %f108, %f110;
	mov.f32 	%f112, 0f3f800000;   	// 1
	min.ftz.f32 	%f87, %f111, %f112;
	sub.ftz.f32 	%f113, %f87, %f63;
	mov.f32 	%f114, 0f00000000;   	// 0
	setp.lt.ftz.f32 	%p24, %f113, %f114;
	mov.f32 	%f115, 0f3f800000;   	// 1
	add.ftz.f32 	%f116, %f113, %f115;
	selp.f32 	%f117, %f116, %f113, %p24;
	set.le.ftz.u32.f32 	%r46, %f109, %f117;
	neg.s32 	%r47, %r46;
	mov.f32 	%f118, 0f00000000;   	// 0
	set.lt.ftz.u32.f32 	%r48, %f117, %f118;
	neg.s32 	%r49, %r48;
	or.b32 	%r50, %r47, %r49;
	mov.u32 	%r51, 0;
	setp.eq.s32 	%p25, %r50, %r51;
	@%p25 bra 	$Lt_31_78850;
	mov.f32 	%f119, 0f00000000;   	// 0
	bra.uni 	$Lt_31_78594;
$Lt_31_78850:
	.loc	24	119	0
	div.approx.ftz.f32 	%f119, %f117, %f109;
$Lt_31_78594:
	.loc	24	180	0
	mov.f32 	%f120, %f119;
	mov.f32 	%f121, 0f3f800000;   	// 1
	setp.lt.ftz.f32 	%p26, %f119, %f121;
	@!%p26 bra 	$Lt_31_79362;
	.loc	24	188	0
	sub.ftz.f32 	%f122, %f56, %f64;
	mov.f32 	%f123, 0f00000000;   	// 0
	setp.lt.ftz.f32 	%p27, %f122, %f123;
	mov.f32 	%f124, 0f3f800000;   	// 1
	add.ftz.f32 	%f125, %f122, %f124;
	selp.f32 	%f126, %f125, %f122, %p27;
	sub.ftz.f32 	%f127, %f87, %f64;
	mov.f32 	%f128, 0f00000000;   	// 0
	setp.lt.ftz.f32 	%p28, %f127, %f128;
	mov.f32 	%f129, 0f3f800000;   	// 1
	add.ftz.f32 	%f130, %f127, %f129;
	selp.f32 	%f131, %f130, %f127, %p28;
	set.le.ftz.u32.f32 	%r52, %f126, %f131;
	neg.s32 	%r53, %r52;
	mov.f32 	%f132, 0f00000000;   	// 0
	set.lt.ftz.u32.f32 	%r54, %f131, %f132;
	neg.s32 	%r55, %r54;
	or.b32 	%r56, %r53, %r55;
	mov.u32 	%r57, 0;
	setp.eq.s32 	%p29, %r56, %r57;
	@%p29 bra 	$Lt_31_79874;
	mov.f32 	%f133, 0f00000000;   	// 0
	bra.uni 	$Lt_31_79618;
$Lt_31_79874:
	.loc	24	154	0
	mov.f32 	%f134, 0f3f800000;   	// 1
	div.approx.ftz.f32 	%f135, %f131, %f126;
	sub.ftz.f32 	%f133, %f134, %f135;
$Lt_31_79618:
	.loc	24	188	0
	mov.f32 	%f136, %f133;
	bra.uni 	$Lt_31_78082;
$Lt_31_79362:
	mov.f32 	%f136, 0f00000000;   	// 0
	bra.uni 	$Lt_31_78082;
$Lt_31_78338:
	mov.f32 	%f136, 0f00000000;   	// 0
	mov.f32 	%f120, 0f00000000;   	// 0
$Lt_31_78082:
	.loc	24	307	0
	max.ftz.f32 	%f137, %f136, %f120;
	max.ftz.f32 	%f138, %f137, %f75;
	mov.f32 	%f139, 0f3f800000;   	// 1
	sub.ftz.f32 	%f140, %f139, %f138;
	ld.s8 	%r58, [%rd2+60];
	mov.s32 	%r59, 0;
	setp.ne.s32 	%p30, %r58, %r59;
	selp.f32 	%f141, %f140, %f138, %p30;
	mov.f32 	%f142, 0f00000000;   	// 0
	max.ftz.f32 	%f143, %f141, %f142;
	mov.f32 	%f144, 0f3f800000;   	// 1
	min.ftz.f32 	%f52, %f143, %f144;
$Lt_31_75522:
	mov.f32 	%f145, 0f00000000;   	// 0
	setp.neu.ftz.f32 	%p31, %f52, %f145;
	@!%p31 bra 	$Lt_31_80386;
	.loc	24	236	0
	ld.f32 	%f146, [%rd2+20];
	ld.f32 	%f147, [%rd2+16];
	sub.ftz.f32 	%f148, %f146, %f147;
	ld.f32 	%f149, [%rd2+28];
	ld.f32 	%f150, [%rd2+24];
	add.ftz.f32 	%f151, %f150, %f149;
	.loc	24	71	0
	sub.ftz.f32 	%f152, %f149, %f146;
	setp.ge.ftz.f32 	%p32, %f149, %f146;
	@!%p32 bra 	$Lt_31_84994;
	mov.f32 	%f153, 0f358637bd;   	// 1e-006
	setp.lt.ftz.f32 	%p33, %f152, %f153;
	@!%p33 bra 	$Lt_31_84994;
	.loc	24	74	0
	mov.f32 	%f154, 0f00000000;   	// 0
	bra.uni 	$L_31_69634;
$Lt_31_84994:
$L_31_69122:
	.loc	24	77	0
	setp.lt.ftz.f32 	%p34, %f149, %f146;
	@!%p34 bra 	$Lt_31_85506;
	sub.ftz.f32 	%f155, %f146, %f149;
	mov.f32 	%f156, 0f358637bd;   	// 1e-006
	setp.lt.ftz.f32 	%p35, %f155, %f156;
	@!%p35 bra 	$Lt_31_85506;
	mov.s32 	%r60, 1;
	bra.uni 	$L_31_69890;
$Lt_31_85506:
$L_31_70146:
	mov.s32 	%r60, 0;
$L_31_69890:
	mul.ftz.f32 	%f157, %f4, %f4;
	fma.rn.ftz.f32 	%f158, %f2, %f2, %f157;
	sqrt.approx.ftz.f32 	%f159, %f158;
	mov.f32 	%f160, 0f3f3504f3;   	// 0.707107
	div.approx.ftz.f32 	%f161, %f159, %f160;
	mov.f32 	%f162, 0f00000000;   	// 0
	max.ftz.f32 	%f163, %f161, %f162;
	mov.f32 	%f164, 0f3f800000;   	// 1
	min.ftz.f32 	%f165, %f163, %f164;
	sub.ftz.f32 	%f166, %f165, %f146;
	mov.f32 	%f167, 0f00000000;   	// 0
	setp.lt.ftz.f32 	%p36, %f166, %f167;
	mov.f32 	%f168, 0f3f800000;   	// 1
	add.ftz.f32 	%f169, %f166, %f168;
	selp.f32 	%f170, %f169, %f166, %p36;
	mov.f32 	%f171, 0f3f800000;   	// 1
	add.ftz.f32 	%f172, %f152, %f171;
	mov.f32 	%f173, 0f00000000;   	// 0
	setp.lt.ftz.f32 	%p37, %f152, %f173;
	selp.f32 	%f174, %f172, %f152, %p37;
	set.ge.ftz.u32.f32 	%r61, %f174, %f170;
	neg.s32 	%r62, %r61;
	mov.f32 	%f175, 0f00000000;   	// 0
	set.ge.ftz.u32.f32 	%r63, %f170, %f175;
	neg.s32 	%r64, %r63;
	and.b32 	%r65, %r62, %r64;
	mov.u32 	%r66, 0;
	set.ne.u32.u32 	%r67, %r60, %r66;
	neg.s32 	%r68, %r67;
	or.b32 	%r69, %r68, %r65;
	mov.u32 	%r70, 0;
	setp.eq.s32 	%p38, %r69, %r70;
	@%p38 bra 	$Lt_31_86018;
	mov.f32 	%f154, 0f3f800000;   	// 1
	bra.uni 	$L_31_69634;
$Lt_31_86018:
	mov.f32 	%f154, 0f00000000;   	// 0
$L_31_69634:
$L_31_68866:
	.loc	24	173	0
	mov.f32 	%f176, 0f3f800000;   	// 1
	setp.lt.ftz.f32 	%p39, %f154, %f176;
	@!%p39 bra 	$Lt_31_80898;
	.loc	24	108	0
	mul.ftz.f32 	%f177, %f4, %f4;
	sub.ftz.f32 	%f178, %f146, %f148;
	fma.rn.ftz.f32 	%f179, %f2, %f2, %f177;
	mov.f32 	%f180, 0f00000000;   	// 0
	setp.lt.ftz.f32 	%p40, %f178, %f180;
	mov.f32 	%f181, 0f3f800000;   	// 1
	add.ftz.f32 	%f182, %f178, %f181;
	sqrt.approx.ftz.f32 	%f183, %f179;
	selp.f32 	%f184, %f182, %f178, %p40;
	mov.f32 	%f185, 0f3f3504f3;   	// 0.707107
	div.approx.ftz.f32 	%f186, %f183, %f185;
	mov.f32 	%f187, 0f00000000;   	// 0
	max.ftz.f32 	%f188, %f186, %f187;
	mov.f32 	%f189, 0f3f800000;   	// 1
	min.ftz.f32 	%f165, %f188, %f189;
	sub.ftz.f32 	%f190, %f165, %f148;
	mov.f32 	%f191, 0f00000000;   	// 0
	setp.lt.ftz.f32 	%p41, %f190, %f191;
	mov.f32 	%f192, 0f3f800000;   	// 1
	add.ftz.f32 	%f193, %f190, %f192;
	selp.f32 	%f194, %f193, %f190, %p41;
	neg.ftz.f32 	%f195, %f148;
	set.lt.ftz.u32.f32 	%r71, %f194, %f195;
	neg.s32 	%r72, %r71;
	set.le.ftz.u32.f32 	%r73, %f184, %f194;
	neg.s32 	%r74, %r73;
	mov.f32 	%f196, 0f00000000;   	// 0
	set.lt.ftz.u32.f32 	%r75, %f194, %f196;
	neg.s32 	%r76, %r75;
	or.b32 	%r77, %r74, %r76;
	or.b32 	%r78, %r72, %r77;
	mov.u32 	%r79, 0;
	setp.eq.s32 	%p42, %r78, %r79;
	@%p42 bra 	$L_31_70914;
	.loc	24	111	0
	mov.f32 	%f197, 0f00000000;   	// 0
	bra.uni 	$L_31_70658;
$L_31_70914:
	.loc	28	529	0
	div.approx.ftz.f32 	%f198, %f194, %f184;
	.loc	24	119	0
	mov.f32 	%f197, %f198;
$L_31_70658:
	.loc	24	180	0
	mov.f32 	%f199, %f197;
	mov.f32 	%f200, 0f3f800000;   	// 1
	setp.lt.ftz.f32 	%p43, %f197, %f200;
	@!%p43 bra 	$Lt_31_81410;
	.loc	24	143	0
	sub.ftz.f32 	%f201, %f151, %f149;
	mov.f32 	%f202, 0f00000000;   	// 0
	setp.lt.ftz.f32 	%p44, %f201, %f202;
	mov.f32 	%f203, 0f3f800000;   	// 1
	add.ftz.f32 	%f204, %f201, %f203;
	selp.f32 	%f205, %f204, %f201, %p44;
	sub.ftz.f32 	%f206, %f165, %f149;
	mov.f32 	%f207, 0f00000000;   	// 0
	setp.lt.ftz.f32 	%p45, %f206, %f207;
	mov.f32 	%f208, 0f3f800000;   	// 1
	add.ftz.f32 	%f209, %f206, %f208;
	selp.f32 	%f210, %f209, %f206, %p45;
	mov.f32 	%f211, 0f3f800000;   	// 1
	sub.ftz.f32 	%f212, %f211, %f149;
	set.gt.ftz.u32.f32 	%r80, %f210, %f212;
	neg.s32 	%r81, %r80;
	set.le.ftz.u32.f32 	%r82, %f205, %f210;
	neg.s32 	%r83, %r82;
	mov.f32 	%f213, 0f00000000;   	// 0
	set.lt.ftz.u32.f32 	%r84, %f210, %f213;
	neg.s32 	%r85, %r84;
	or.b32 	%r86, %r83, %r85;
	or.b32 	%r87, %r81, %r86;
	mov.u32 	%r88, 0;
	setp.eq.s32 	%p46, %r87, %r88;
	@%p46 bra 	$L_31_71426;
	.loc	24	146	0
	mov.f32 	%f214, 0f00000000;   	// 0
	bra.uni 	$L_31_71170;
$L_31_71426:
	.loc	28	529	0
	div.approx.ftz.f32 	%f215, %f210, %f205;
	.loc	24	154	0
	mov.f32 	%f216, 0f3f800000;   	// 1
	sub.ftz.f32 	%f214, %f216, %f215;
$L_31_71170:
	.loc	24	188	0
	mov.f32 	%f217, %f214;
	bra.uni 	$Lt_31_80642;
$Lt_31_81410:
	mov.f32 	%f217, 0f00000000;   	// 0
	bra.uni 	$Lt_31_80642;
$Lt_31_80898:
	mov.f32 	%f217, 0f00000000;   	// 0
	mov.f32 	%f199, 0f00000000;   	// 0
$Lt_31_80642:
	.loc	24	198	0
	ld.s8 	%r89, [%rd2+60];
	max.ftz.f32 	%f218, %f217, %f199;
	mov.s32 	%r90, 0;
	setp.ne.s32 	%p47, %r89, %r90;
	max.ftz.f32 	%f219, %f218, %f154;
	mov.f32 	%f220, 0f3f800000;   	// 1
	sub.ftz.f32 	%f221, %f220, %f219;
	selp.f32 	%f222, %f221, %f219, %p47;
	.loc	24	311	0
	mov.f32 	%f223, 0f00000000;   	// 0
	max.ftz.f32 	%f224, %f222, %f223;
	mov.f32 	%f225, 0f3f800000;   	// 1
	min.ftz.f32 	%f226, %f224, %f225;
	mov.f32 	%f227, 0f00000000;   	// 0
	setp.neu.ftz.f32 	%p48, %f226, %f227;
	@!%p48 bra 	$Lt_31_81922;
	.loc	24	261	0
	ld.f32 	%f228, [%rd2+36];
	ld.f32 	%f229, [%rd2+32];
	sub.ftz.f32 	%f230, %f228, %f229;
	ld.f32 	%f231, [%rd2+44];
	ld.f32 	%f232, [%rd2+40];
	add.ftz.f32 	%f233, %f232, %f231;
	.loc	24	71	0
	sub.ftz.f32 	%f234, %f231, %f228;
	setp.ge.ftz.f32 	%p49, %f231, %f228;
	@!%p49 bra 	$Lt_31_86274;
	mov.f32 	%f235, 0f358637bd;   	// 1e-006
	setp.lt.ftz.f32 	%p50, %f234, %f235;
	@!%p50 bra 	$Lt_31_86274;
	.loc	24	74	0
	mov.f32 	%f236, 0f00000000;   	// 0
	bra.uni 	$L_31_72450;
$Lt_31_86274:
$L_31_71938:
	.loc	24	77	0
	setp.lt.ftz.f32 	%p51, %f231, %f228;
	@!%p51 bra 	$Lt_31_86786;
	sub.ftz.f32 	%f237, %f228, %f231;
	mov.f32 	%f238, 0f358637bd;   	// 1e-006
	setp.lt.ftz.f32 	%p52, %f237, %f238;
	@!%p52 bra 	$Lt_31_86786;
	mov.s32 	%r91, 1;
	bra.uni 	$L_31_72706;
$Lt_31_86786:
$L_31_72962:
	mov.s32 	%r91, 0;
$L_31_72706:
	mov.f32 	%f239, 0f00000000;   	// 0
	max.ftz.f32 	%f240, %f6, %f239;
	mov.f32 	%f241, 0f3f800000;   	// 1
	min.ftz.f32 	%f242, %f240, %f241;
	sub.ftz.f32 	%f243, %f242, %f228;
	mov.f32 	%f244, 0f00000000;   	// 0
	setp.lt.ftz.f32 	%p53, %f243, %f244;
	mov.f32 	%f245, 0f3f800000;   	// 1
	add.ftz.f32 	%f246, %f243, %f245;
	selp.f32 	%f247, %f246, %f243, %p53;
	mov.f32 	%f248, 0f3f800000;   	// 1
	add.ftz.f32 	%f249, %f234, %f248;
	mov.f32 	%f250, 0f00000000;   	// 0
	setp.lt.ftz.f32 	%p54, %f234, %f250;
	selp.f32 	%f251, %f249, %f234, %p54;
	set.ge.ftz.u32.f32 	%r92, %f251, %f247;
	neg.s32 	%r93, %r92;
	mov.f32 	%f252, 0f00000000;   	// 0
	set.ge.ftz.u32.f32 	%r94, %f247, %f252;
	neg.s32 	%r95, %r94;
	and.b32 	%r96, %r93, %r95;
	mov.u32 	%r97, 0;
	set.ne.u32.u32 	%r98, %r91, %r97;
	neg.s32 	%r99, %r98;
	or.b32 	%r100, %r99, %r96;
	mov.u32 	%r101, 0;
	setp.eq.s32 	%p55, %r100, %r101;
	@%p55 bra 	$Lt_31_87298;
	mov.f32 	%f236, 0f3f800000;   	// 1
	bra.uni 	$L_31_72450;
$Lt_31_87298:
	mov.f32 	%f236, 0f00000000;   	// 0
$L_31_72450:
$L_31_71682:
	.loc	24	173	0
	mov.f32 	%f253, 0f3f800000;   	// 1
	setp.lt.ftz.f32 	%p56, %f236, %f253;
	@!%p56 bra 	$Lt_31_82434;
	.loc	24	108	0
	mov.f32 	%f254, 0f00000000;   	// 0
	max.ftz.f32 	%f255, %f6, %f254;
	sub.ftz.f32 	%f256, %f228, %f230;
	mov.f32 	%f257, 0f3f800000;   	// 1
	min.ftz.f32 	%f242, %f255, %f257;
	mov.f32 	%f258, 0f00000000;   	// 0
	setp.lt.ftz.f32 	%p57, %f256, %f258;
	mov.f32 	%f259, 0f3f800000;   	// 1
	add.ftz.f32 	%f260, %f256, %f259;
	sub.ftz.f32 	%f261, %f242, %f230;
	selp.f32 	%f262, %f260, %f256, %p57;
	mov.f32 	%f263, 0f00000000;   	// 0
	setp.lt.ftz.f32 	%p58, %f261, %f263;
	mov.f32 	%f264, 0f3f800000;   	// 1
	add.ftz.f32 	%f265, %f261, %f264;
	selp.f32 	%f266, %f265, %f261, %p58;
	neg.ftz.f32 	%f267, %f230;
	set.lt.ftz.u32.f32 	%r102, %f266, %f267;
	neg.s32 	%r103, %r102;
	set.le.ftz.u32.f32 	%r104, %f262, %f266;
	neg.s32 	%r105, %r104;
	mov.f32 	%f268, 0f00000000;   	// 0
	set.lt.ftz.u32.f32 	%r106, %f266, %f268;
	neg.s32 	%r107, %r106;
	or.b32 	%r108, %r105, %r107;
	or.b32 	%r109, %r103, %r108;
	mov.u32 	%r110, 0;
	setp.eq.s32 	%p59, %r109, %r110;
	@%p59 bra 	$L_31_73730;
	.loc	24	111	0
	mov.f32 	%f269, 0f00000000;   	// 0
	bra.uni 	$L_31_73474;
$L_31_73730:
	.loc	24	119	0
	div.approx.ftz.f32 	%f269, %f266, %f262;
$L_31_73474:
	.loc	24	180	0
	mov.f32 	%f270, %f269;
	mov.f32 	%f271, 0f3f800000;   	// 1
	setp.lt.ftz.f32 	%p60, %f269, %f271;
	@!%p60 bra 	$Lt_31_82946;
	.loc	24	143	0
	sub.ftz.f32 	%f272, %f233, %f231;
	mov.f32 	%f273, 0f00000000;   	// 0
	setp.lt.ftz.f32 	%p61, %f272, %f273;
	mov.f32 	%f274, 0f3f800000;   	// 1
	add.ftz.f32 	%f275, %f272, %f274;
	sub.ftz.f32 	%f276, %f242, %f231;
	selp.f32 	%f277, %f275, %f272, %p61;
	mov.f32 	%f278, 0f00000000;   	// 0
	setp.lt.ftz.f32 	%p62, %f276, %f278;
	mov.f32 	%f279, 0f3f800000;   	// 1
	add.ftz.f32 	%f280, %f276, %f279;
	selp.f32 	%f281, %f280, %f276, %p62;
	mov.f32 	%f282, 0f3f800000;   	// 1
	sub.ftz.f32 	%f283, %f282, %f231;
	set.gt.ftz.u32.f32 	%r111, %f281, %f283;
	neg.s32 	%r112, %r111;
	set.le.ftz.u32.f32 	%r113, %f277, %f281;
	neg.s32 	%r114, %r113;
	mov.f32 	%f284, 0f00000000;   	// 0
	set.lt.ftz.u32.f32 	%r115, %f281, %f284;
	neg.s32 	%r116, %r115;
	or.b32 	%r117, %r114, %r116;
	or.b32 	%r118, %r112, %r117;
	mov.u32 	%r119, 0;
	setp.eq.s32 	%p63, %r118, %r119;
	@%p63 bra 	$L_31_74242;
	.loc	24	146	0
	mov.f32 	%f285, 0f00000000;   	// 0
	bra.uni 	$L_31_73986;
$L_31_74242:
	.loc	24	154	0
	mov.f32 	%f286, 0f3f800000;   	// 1
	div.approx.ftz.f32 	%f287, %f281, %f277;
	sub.ftz.f32 	%f285, %f286, %f287;
$L_31_73986:
	.loc	24	188	0
	mov.f32 	%f288, %f285;
	bra.uni 	$Lt_31_82178;
$Lt_31_82946:
	mov.f32 	%f288, 0f00000000;   	// 0
	bra.uni 	$Lt_31_82178;
$Lt_31_82434:
	mov.f32 	%f288, 0f00000000;   	// 0
	mov.f32 	%f270, 0f00000000;   	// 0
$Lt_31_82178:
	.loc	24	198	0
	max.ftz.f32 	%f289, %f288, %f270;
	max.ftz.f32 	%f290, %f289, %f236;
	mov.f32 	%f291, 0f3f800000;   	// 1
	sub.ftz.f32 	%f292, %f291, %f290;
	selp.f32 	%f293, %f292, %f290, %p47;
	.loc	24	278	0
	mov.f32 	%f294, 0f00000000;   	// 0
	max.ftz.f32 	%f295, %f293, %f294;
	mov.f32 	%f296, 0f3f800000;   	// 1
	min.ftz.f32 	%f297, %f295, %f296;
	add.ftz.f32 	%f298, %f226, %f52;
	add.ftz.f32 	%f299, %f297, %f298;
	mov.f32 	%f300, 0f3eaaaaab;   	// 0.333333
	mul.ftz.f32 	%f301, %f299, %f300;
	mov.f32 	%f302, 0f00000000;   	// 0
	mov.f32 	%f303, 0f00000000;   	// 0
	setp.neu.ftz.f32 	%p64, %f297, %f303;
	selp.f32 	%f304, %f301, %f302, %p64;
	bra.uni 	$Lt_31_80130;
$Lt_31_81922:
	mov.f32 	%f304, 0f00000000;   	// 0
	bra.uni 	$Lt_31_80130;
$Lt_31_80386:
	mov.f32 	%f304, 0f00000000;   	// 0
$Lt_31_80130:
	ld.f32 	%f305, [%rd2+56];
	mov.f32 	%f306, 0f00000000;   	// 0
	setp.neu.ftz.f32 	%p65, %f305, %f306;
	@!%p65 bra 	$Lt_31_83202;
	.loc	24	327	0
	lg2.approx.ftz.f32 	%f307, %f304;
	mov.f32 	%f308, 0f41200000;   	// 10
	lg2.approx.ftz.f32 	%f309, %f308;
	mul.ftz.f32 	%f310, %f305, %f309;
	ex2.approx.ftz.f32 	%f311, %f310;
	rcp.approx.ftz.f32 	%f312, %f311;
	mul.ftz.f32 	%f313, %f307, %f312;
	ex2.approx.ftz.f32 	%f304, %f313;
$Lt_31_83202:
	.loc	24	332	0
	mov.f32 	%f314, 0f00000000;   	// 0
	max.ftz.f32 	%f315, %f304, %f314;
	mov.f32 	%f316, 0f3f800000;   	// 1
	min.ftz.f32 	%f317, %f315, %f316;
	st.param.f32 	[__cudaretf__Z15CalculateWeight8PixelYUVRK33SecondaryColorCorrectorParameters], %f317;
	ret;
$LDWend__Z15CalculateWeight8PixelYUVRK33SecondaryColorCorrectorParameters:
	} // _Z15CalculateWeight8PixelYUVRK33SecondaryColorCorrectorParameters

	.visible .func _Z7Write2DIfEvT_PS0_iii (.param .f32 __cudaparmf1__Z7Write2DIfEvT_PS0_iii, .param .u64 __cudaparmf2__Z7Write2DIfEvT_PS0_iii, .param .s32 __cudaparmf3__Z7Write2DIfEvT_PS0_iii, .param .s32 __cudaparmf4__Z7Write2DIfEvT_PS0_iii, .param .s32 __cudaparmf5__Z7Write2DIfEvT_PS0_iii)
	{
	.reg .u32 %r<10>;
	.reg .u64 %rd<7>;
	.reg .f32 %f<4>;
	.loc	21	125	0
$LDWbegin__Z7Write2DIfEvT_PS0_iii:
	ld.param.f32 	%f1, [__cudaparmf1__Z7Write2DIfEvT_PS0_iii];
	mov.f32 	%f2, %f1;
	ld.param.u64 	%rd1, [__cudaparmf2__Z7Write2DIfEvT_PS0_iii];
	mov.s64 	%rd2, %rd1;
	ld.param.u32 	%r1, [__cudaparmf3__Z7Write2DIfEvT_PS0_iii];
	mov.s32 	%r2, %r1;
	ld.param.u32 	%r3, [__cudaparmf4__Z7Write2DIfEvT_PS0_iii];
	mov.s32 	%r4, %r3;
	ld.param.u32 	%r5, [__cudaparmf5__Z7Write2DIfEvT_PS0_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, 4;
	add.u64 	%rd5, %rd2, %rd4;
	st.f32 	[%rd5+0], %f2;
	.loc	21	127	0
	ret;
$LDWend__Z7Write2DIfEvT_PS0_iii:
	} // _Z7Write2DIfEvT_PS0_iii
	.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};

	.entry SecondaryColorCorrection (
		.param .u64 __cudaparm_SecondaryColorCorrection_inImage,
		.param .s32 __cudaparm_SecondaryColorCorrection_inPitch,
		.param .u64 __cudaparm_SecondaryColorCorrection_inSecondaryMask,
		.param .s32 __cudaparm_SecondaryColorCorrection_inSecondaryPitch,
		.param .u32 __cudaparm_SecondaryColorCorrection_inDeviceFormat,
		.param .s32 __cudaparm_SecondaryColorCorrection_inWidth,
		.param .s32 __cudaparm_SecondaryColorCorrection_inHeight,
		.param .align 4 .b8 __cudaparm_SecondaryColorCorrection___val_paraminSecondaryParameters[64])
	{
	.reg .u32 %r<150>;
	.reg .u64 %rd<12>;
	.reg .f32 %f<331>;
	.reg .pred %p<69>;
	.loc	24	346	0
$LDWbegin_SecondaryColorCorrection:
	.loc	24	349	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_SecondaryColorCorrection_inWidth];
	set.gt.u32.s32 	%r12, %r11, %r8;
	neg.s32 	%r13, %r12;
	ld.param.s32 	%r14, [__cudaparm_SecondaryColorCorrection_inHeight];
	set.gt.u32.s32 	%r15, %r14, %r10;
	neg.s32 	%r16, %r15;
	and.b32 	%r17, %r13, %r16;
	mov.u32 	%r18, 0;
	setp.eq.s32 	%p1, %r17, %r18;
	@%p1 bra 	$Lt_33_76290;
	ld.param.s32 	%r19, [__cudaparm_SecondaryColorCorrection_inPitch];
	mul.lo.s32 	%r20, %r19, %r10;
	add.s32 	%r21, %r8, %r20;
	cvt.s64.s32 	%rd1, %r21;
	ld.param.u64 	%rd2, [__cudaparm_SecondaryColorCorrection_inImage];
	ld.param.s32 	%r22, [__cudaparm_SecondaryColorCorrection_inDeviceFormat];
	mov.u32 	%r23, 0;
	setp.ne.s32 	%p2, %r22, %r23;
	@%p2 bra 	$Lt_33_77058;
	.loc	21	115	0
	mul.lo.u64 	%rd3, %rd1, 8;
	add.u64 	%rd4, %rd2, %rd3;
	ld.global.v4.u16 	{%r24,%r25,%r26,_}, [%rd4+0];
	.loc	24	352	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; }
	bra.uni 	$Lt_33_76802;
$Lt_33_77058:
	mul.lo.u64 	%rd5, %rd1, 16;
	add.u64 	%rd6, %rd2, %rd5;
	ld.global.v4.f32 	{%f1,%f2,%f3,_}, [%rd6+0];
$Lt_33_76802:
	.loc	5	60	0
	ld.const.f32 	%f4, [kRGB32f_To_601YPbPr+28];
	mul.ftz.f32 	%f5, %f4, %f2;
	ld.const.f32 	%f6, [kRGB32f_To_601YPbPr+16];
	mul.ftz.f32 	%f7, %f6, %f2;
	ld.const.f32 	%f8, [kRGB32f_To_601YPbPr+24];
	fma.rn.ftz.f32 	%f9, %f8, %f3, %f5;
	ld.const.f32 	%f10, [kRGB32f_To_601YPbPr+12];
	fma.rn.ftz.f32 	%f11, %f10, %f3, %f7;
	ld.const.f32 	%f12, [kRGB32f_To_601YPbPr+32];
	fma.rn.ftz.f32 	%f13, %f12, %f1, %f9;
	ld.const.f32 	%f14, [kRGB32f_To_601YPbPr+20];
	fma.rn.ftz.f32 	%f15, %f14, %f1, %f11;
	mov.f32 	%f16, 0f3f728f61;    	// 0.9475
	mul.ftz.f32 	%f17, %f13, %f16;
	mov.f32 	%f18, 0f3ea3b6e9;    	// 0.319755
	mul.ftz.f32 	%f19, %f13, %f18;
	mov.f32 	%f20, 0fbea3b6e9;    	// -0.319755
	fma.rn.ftz.f32 	%f21, %f20, %f15, %f17;
	mov.f32 	%f22, 0fbf728f61;    	// -0.9475
	mul.ftz.f32 	%f23, %f22, %f15;
	sub.ftz.f32 	%f24, %f23, %f19;
	abs.ftz.f32 	%f25, %f21;
	abs.ftz.f32 	%f26, %f24;
	mov.b32 	%r27, %f24;
	and.b32 	%r28, %r27, -2147483648;
	mov.f32 	%f27, 0f00000000;    	// 0
	set.eq.ftz.u32.f32 	%r29, %f25, %f27;
	neg.s32 	%r30, %r29;
	mov.f32 	%f28, 0f00000000;    	// 0
	set.eq.ftz.u32.f32 	%r31, %f26, %f28;
	neg.s32 	%r32, %r31;
	and.b32 	%r33, %r30, %r32;
	mov.u32 	%r34, 0;
	setp.eq.s32 	%p3, %r33, %r34;
	@%p3 bra 	$Lt_33_77570;
	.loc	26	1842	0
	mov.s32 	%r35, 1078530011;
	mov.s32 	%r36, 0;
	mov.b32 	%r37, %f21;
	mov.s32 	%r38, 0;
	setp.lt.s32 	%p4, %r37, %r38;
	selp.s32 	%r39, %r35, %r36, %p4;
	or.b32 	%r40, %r39, %r28;
	mov.b32 	%f29, %r40;
	bra.uni 	$Lt_33_77826;
$Lt_33_77570:
	mov.f32 	%f30, 0f7f800000;    	// 1.#INF
	set.eq.ftz.u32.f32 	%r41, %f25, %f30;
	neg.s32 	%r42, %r41;
	mov.f32 	%f31, 0f7f800000;    	// 1.#INF
	set.eq.ftz.u32.f32 	%r43, %f26, %f31;
	neg.s32 	%r44, %r43;
	and.b32 	%r45, %r42, %r44;
	mov.u32 	%r46, 0;
	setp.eq.s32 	%p5, %r45, %r46;
	@%p5 bra 	$Lt_33_78082;
	.loc	26	1845	0
	mov.s32 	%r47, 1075235812;
	mov.s32 	%r48, 1061752795;
	mov.b32 	%r49, %f21;
	mov.s32 	%r50, 0;
	setp.lt.s32 	%p6, %r49, %r50;
	selp.s32 	%r51, %r47, %r48, %p6;
	or.b32 	%r52, %r51, %r28;
	mov.b32 	%f29, %r52;
	bra.uni 	$Lt_33_77826;
$Lt_33_78082:
	.loc	26	1173	0
	min.ftz.f32 	%f32, %f25, %f26;
	max.ftz.f32 	%f33, %f25, %f26;
	div.full.ftz.f32 	%f34, %f32, %f33;
	mul.ftz.f32 	%f35, %f34, %f34;
	mov.f32 	%f36, 0fbf52c7ea;    	// -0.823363
	mov.f32 	%f37, 0fc0b59883;    	// -5.67487
	fma.rn.ftz.f32 	%f38, %f35, %f36, %f37;
	mov.f32 	%f39, 0fc0d21907;    	// -6.56556
	fma.rn.ftz.f32 	%f40, %f38, %f35, %f39;
	mov.f32 	%f41, 0f41355dc0;    	// 11.3354
	add.ftz.f32 	%f42, %f35, %f41;
	mov.f32 	%f43, 0f41e6bd60;    	// 28.8425
	fma.rn.ftz.f32 	%f44, %f42, %f35, %f43;
	.loc	26	1211	0
	mov.f32 	%f45, 0f419d92c8;    	// 19.6967
	fma.rn.ftz.f32 	%f46, %f44, %f35, %f45;
	rcp.approx.ftz.f32 	%f47, %f46;
	.loc	26	1173	0
	mul.ftz.f32 	%f48, %f40, %f35;
	mul.ftz.f32 	%f49, %f34, %f48;
	fma.rn.ftz.f32 	%f50, %f49, %f47, %f34;
	.loc	26	1856	0
	mov.f32 	%f51, 0f3fc90fdb;    	// 1.5708
	sub.ftz.f32 	%f52, %f51, %f50;
	setp.lt.ftz.f32 	%p7, %f25, %f26;
	selp.f32 	%f53, %f52, %f50, %p7;
	mov.f32 	%f54, 0f40490fdb;    	// 3.14159
	sub.ftz.f32 	%f55, %f54, %f53;
	mov.f32 	%f56, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p8, %f21, %f56;
	selp.f32 	%f57, %f55, %f53, %p8;
	mov.b32 	%r53, %f57;
	or.b32 	%r54, %r28, %r53;
	mov.b32 	%f58, %r54;
	add.ftz.f32 	%f59, %f21, %f24;
	mov.f32 	%f60, 0f7f800000;    	// 1.#INF
	setp.le.ftz.f32 	%p9, %f59, %f60;
	selp.f32 	%f29, %f58, %f59, %p9;
$Lt_33_77826:
$Lt_33_77314:
	.loc	24	295	0
	ld.param.s8 	%r55, [__cudaparm_SecondaryColorCorrection___val_paraminSecondaryParameters+61];
	mov.u32 	%r56, 0;
	setp.eq.s32 	%p10, %r55, %r56;
	@%p10 bra 	$Lt_33_78594;
	mov.f32 	%f61, 0f3f800000;    	// 1
	bra.uni 	$Lt_33_78338;
$Lt_33_78594:
	.loc	24	35	0
	ld.param.f32 	%f62, [__cudaparm_SecondaryColorCorrection___val_paraminSecondaryParameters+4];
	ld.param.f32 	%f63, [__cudaparm_SecondaryColorCorrection___val_paraminSecondaryParameters+0];
	sub.ftz.f32 	%f64, %f62, %f63;
	mov.f32 	%f65, %f64;
	mov.f32 	%f66, 0f3f800000;    	// 1
	setp.gt.ftz.f32 	%p11, %f64, %f66;
	@!%p11 bra 	$Lt_33_79106;
	.loc	24	39	0
	cvt.rzi.ftz.s32.f32 	%r57, %f64;
	cvt.rn.f32.s32 	%f67, %r57;
	sub.ftz.f32 	%f65, %f64, %f67;
	bra.uni 	$Lt_33_79362;
$Lt_33_79106:
	mov.f32 	%f68, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p12, %f64, %f68;
	@!%p12 bra 	$Lt_33_79362;
	.loc	24	43	0
	cvt.rzi.ftz.s32.f32 	%r58, %f64;
	cvt.rn.f32.s32 	%f69, %r58;
	sub.ftz.f32 	%f70, %f64, %f69;
	mov.f32 	%f71, 0f3f800000;    	// 1
	add.ftz.f32 	%f65, %f70, %f71;
$Lt_33_79362:
$Lt_33_78850:
	.loc	24	208	0
	mov.f32 	%f72, %f65;
	.loc	24	35	0
	ld.param.f32 	%f73, [__cudaparm_SecondaryColorCorrection___val_paraminSecondaryParameters+12];
	ld.param.f32 	%f74, [__cudaparm_SecondaryColorCorrection___val_paraminSecondaryParameters+8];
	add.ftz.f32 	%f75, %f74, %f73;
	mov.f32 	%f65, %f75;
	mov.f32 	%f76, 0f3f800000;    	// 1
	setp.gt.ftz.f32 	%p13, %f75, %f76;
	@!%p13 bra 	$Lt_33_80130;
	.loc	24	39	0
	cvt.rzi.ftz.s32.f32 	%r59, %f75;
	cvt.rn.f32.s32 	%f77, %r59;
	sub.ftz.f32 	%f65, %f75, %f77;
	bra.uni 	$Lt_33_80386;
$Lt_33_80130:
	mov.f32 	%f78, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p14, %f75, %f78;
	@!%p14 bra 	$Lt_33_80386;
	.loc	24	43	0
	cvt.rzi.ftz.s32.f32 	%r60, %f75;
	cvt.rn.f32.s32 	%f79, %r60;
	sub.ftz.f32 	%f80, %f75, %f79;
	mov.f32 	%f81, 0f3f800000;    	// 1
	add.ftz.f32 	%f65, %f80, %f81;
$Lt_33_80386:
$Lt_33_79874:
	.loc	24	71	0
	sub.ftz.f32 	%f82, %f73, %f62;
	setp.ge.ftz.f32 	%p15, %f73, %f62;
	@!%p15 bra 	$Lt_33_86530;
	mov.f32 	%f83, 0f358637bd;    	// 1e-006
	setp.lt.ftz.f32 	%p16, %f82, %f83;
	@!%p16 bra 	$Lt_33_86530;
	.loc	24	74	0
	mov.f32 	%f84, 0f00000000;    	// 0
	bra.uni 	$L_33_69634;
$Lt_33_86530:
$L_33_69122:
	.loc	24	77	0
	setp.lt.ftz.f32 	%p17, %f73, %f62;
	@!%p17 bra 	$Lt_33_87042;
	sub.ftz.f32 	%f85, %f62, %f73;
	mov.f32 	%f86, 0f358637bd;    	// 1e-006
	setp.lt.ftz.f32 	%p18, %f85, %f86;
	@!%p18 bra 	$Lt_33_87042;
	mov.s32 	%r61, 1;
	bra.uni 	$L_33_69890;
$Lt_33_87042:
$L_33_70146:
	mov.s32 	%r61, 0;
$L_33_69890:
	mov.f32 	%f87, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p19, %f29, %f87;
	mov.f32 	%f88, 0f40c90fdb;    	// 6.28319
	add.ftz.f32 	%f89, %f29, %f88;
	selp.f32 	%f90, %f89, %f29, %p19;
	mov.f32 	%f91, 0f3e22f983;    	// 0.159155
	mul.ftz.f32 	%f92, %f90, %f91;
	mov.f32 	%f93, 0f00000000;    	// 0
	max.ftz.f32 	%f94, %f92, %f93;
	mov.f32 	%f95, 0f3f800000;    	// 1
	min.ftz.f32 	%f96, %f94, %f95;
	sub.ftz.f32 	%f97, %f96, %f62;
	mov.f32 	%f98, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p20, %f97, %f98;
	mov.f32 	%f99, 0f3f800000;    	// 1
	add.ftz.f32 	%f100, %f97, %f99;
	selp.f32 	%f101, %f100, %f97, %p20;
	mov.f32 	%f102, 0f3f800000;   	// 1
	add.ftz.f32 	%f103, %f82, %f102;
	mov.f32 	%f104, 0f00000000;   	// 0
	setp.lt.ftz.f32 	%p21, %f82, %f104;
	selp.f32 	%f105, %f103, %f82, %p21;
	set.ge.ftz.u32.f32 	%r62, %f105, %f101;
	neg.s32 	%r63, %r62;
	mov.f32 	%f106, 0f00000000;   	// 0
	set.ge.ftz.u32.f32 	%r64, %f101, %f106;
	neg.s32 	%r65, %r64;
	and.b32 	%r66, %r63, %r65;
	mov.u32 	%r67, 0;
	set.ne.u32.u32 	%r68, %r61, %r67;
	neg.s32 	%r69, %r68;
	or.b32 	%r70, %r69, %r66;
	mov.u32 	%r71, 0;
	setp.eq.s32 	%p22, %r70, %r71;
	@%p22 bra 	$Lt_33_87554;
	mov.f32 	%f84, 0f3f800000;    	// 1
	bra.uni 	$L_33_69634;
$Lt_33_87554:
	mov.f32 	%f84, 0f00000000;    	// 0
$L_33_69634:
$L_33_68866:
	.loc	24	173	0
	mov.f32 	%f107, 0f3f800000;   	// 1
	setp.lt.ftz.f32 	%p23, %f84, %f107;
	@!%p23 bra 	$Lt_33_81154;
	.loc	24	180	0
	mov.f32 	%f108, 0f00000000;   	// 0
	setp.lt.ftz.f32 	%p24, %f29, %f108;
	mov.f32 	%f109, 0f40c90fdb;   	// 6.28319
	add.ftz.f32 	%f110, %f29, %f109;
	sub.ftz.f32 	%f111, %f62, %f72;
	selp.f32 	%f112, %f110, %f29, %p24;
	mov.f32 	%f113, 0f00000000;   	// 0
	setp.lt.ftz.f32 	%p25, %f111, %f113;
	mov.f32 	%f114, 0f3f800000;   	// 1
	add.ftz.f32 	%f115, %f111, %f114;
	mov.f32 	%f116, 0f3e22f983;   	// 0.159155
	mul.ftz.f32 	%f117, %f112, %f116;
	selp.f32 	%f118, %f115, %f111, %p25;
	mov.f32 	%f119, 0f00000000;   	// 0
	max.ftz.f32 	%f120, %f117, %f119;
	mov.f32 	%f121, 0f3f800000;   	// 1
	min.ftz.f32 	%f96, %f120, %f121;
	sub.ftz.f32 	%f122, %f96, %f72;
	mov.f32 	%f123, 0f00000000;   	// 0
	setp.lt.ftz.f32 	%p26, %f122, %f123;
	mov.f32 	%f124, 0f3f800000;   	// 1
	add.ftz.f32 	%f125, %f122, %f124;
	selp.f32 	%f126, %f125, %f122, %p26;
	set.le.ftz.u32.f32 	%r72, %f118, %f126;
	neg.s32 	%r73, %r72;
	mov.f32 	%f127, 0f00000000;   	// 0
	set.lt.ftz.u32.f32 	%r74, %f126, %f127;
	neg.s32 	%r75, %r74;
	or.b32 	%r76, %r73, %r75;
	mov.u32 	%r77, 0;
	setp.eq.s32 	%p27, %r76, %r77;
	@%p27 bra 	$Lt_33_81666;
	mov.f32 	%f128, 0f00000000;   	// 0
	bra.uni 	$Lt_33_81410;
$Lt_33_81666:
	.loc	24	119	0
	div.approx.ftz.f32 	%f128, %f126, %f118;
$Lt_33_81410:
	.loc	24	180	0
	mov.f32 	%f129, %f128;
	mov.f32 	%f130, 0f3f800000;   	// 1
	setp.lt.ftz.f32 	%p28, %f128, %f130;
	@!%p28 bra 	$Lt_33_82178;
	.loc	24	188	0
	sub.ftz.f32 	%f131, %f65, %f73;
	mov.f32 	%f132, 0f00000000;   	// 0
	setp.lt.ftz.f32 	%p29, %f131, %f132;
	mov.f32 	%f133, 0f3f800000;   	// 1
	add.ftz.f32 	%f134, %f131, %f133;
	selp.f32 	%f135, %f134, %f131, %p29;
	sub.ftz.f32 	%f136, %f96, %f73;
	mov.f32 	%f137, 0f00000000;   	// 0
	setp.lt.ftz.f32 	%p30, %f136, %f137;
	mov.f32 	%f138, 0f3f800000;   	// 1
	add.ftz.f32 	%f139, %f136, %f138;
	selp.f32 	%f140, %f139, %f136, %p30;
	set.le.ftz.u32.f32 	%r78, %f135, %f140;
	neg.s32 	%r79, %r78;
	mov.f32 	%f141, 0f00000000;   	// 0
	set.lt.ftz.u32.f32 	%r80, %f140, %f141;
	neg.s32 	%r81, %r80;
	or.b32 	%r82, %r79, %r81;
	mov.u32 	%r83, 0;
	setp.eq.s32 	%p31, %r82, %r83;
	@%p31 bra 	$Lt_33_82690;
	mov.f32 	%f142, 0f00000000;   	// 0
	bra.uni 	$Lt_33_82434;
$Lt_33_82690:
	.loc	24	154	0
	mov.f32 	%f143, 0f3f800000;   	// 1
	div.approx.ftz.f32 	%f144, %f140, %f135;
	sub.ftz.f32 	%f142, %f143, %f144;
$Lt_33_82434:
	.loc	24	188	0
	mov.f32 	%f145, %f142;
	bra.uni 	$Lt_33_80898;
$Lt_33_82178:
	mov.f32 	%f145, 0f00000000;   	// 0
	bra.uni 	$Lt_33_80898;
$Lt_33_81154:
	mov.f32 	%f145, 0f00000000;   	// 0
	mov.f32 	%f129, 0f00000000;   	// 0
$Lt_33_80898:
	.loc	24	307	0
	max.ftz.f32 	%f146, %f145, %f129;
	max.ftz.f32 	%f147, %f146, %f84;
	mov.f32 	%f148, 0f3f800000;   	// 1
	sub.ftz.f32 	%f149, %f148, %f147;
	ld.param.s8 	%r84, [__cudaparm_SecondaryColorCorrection___val_paraminSecondaryParameters+60];
	mov.s32 	%r85, 0;
	setp.ne.s32 	%p32, %r84, %r85;
	selp.f32 	%f150, %f149, %f147, %p32;
	mov.f32 	%f151, 0f00000000;   	// 0
	max.ftz.f32 	%f152, %f150, %f151;
	mov.f32 	%f153, 0f3f800000;   	// 1
	min.ftz.f32 	%f61, %f152, %f153;
$Lt_33_78338:
	mov.f32 	%f154, 0f00000000;   	// 0
	setp.neu.ftz.f32 	%p33, %f61, %f154;
	@!%p33 bra 	$Lt_33_83202;
	.loc	24	311	0
	mul.ftz.f32 	%f155, %f13, %f13;
	fma.rn.ftz.f32 	%f156, %f15, %f15, %f155;
	sqrt.approx.ftz.f32 	%f157, %f156;
	mov.f32 	%f158, 0f3f3504f3;   	// 0.707107
	div.approx.ftz.f32 	%f159, %f157, %f158;
	mov.f32 	%f160, 0f00000000;   	// 0
	max.ftz.f32 	%f161, %f159, %f160;
	mov.f32 	%f162, 0f3f800000;   	// 1
	min.ftz.f32 	%f163, %f161, %f162;
	ld.param.f32 	%f164, [__cudaparm_SecondaryColorCorrection___val_paraminSecondaryParameters+20];
	sub.ftz.f32 	%f165, %f163, %f164;
	mov.f32 	%f166, 0f3f800000;   	// 1
	add.ftz.f32 	%f167, %f165, %f166;
	mov.f32 	%f168, 0f00000000;   	// 0
	setp.lt.ftz.f32 	%p34, %f165, %f168;
	selp.f32 	%f169, %f167, %f165, %p34;
	.loc	24	71	0
	ld.param.f32 	%f170, [__cudaparm_SecondaryColorCorrection___val_paraminSecondaryParameters+28];
	sub.ftz.f32 	%f171, %f170, %f164;
	setp.ge.ftz.f32 	%p35, %f170, %f164;
	@!%p35 bra 	$Lt_33_87810;
	mov.f32 	%f172, 0f358637bd;   	// 1e-006
	setp.lt.ftz.f32 	%p36, %f171, %f172;
	@!%p36 bra 	$Lt_33_87810;
	.loc	24	74	0
	mov.f32 	%f173, 0f00000000;   	// 0
	bra.uni 	$L_33_71426;
$Lt_33_87810:
$L_33_70914:
	.loc	24	77	0
	setp.lt.ftz.f32 	%p37, %f170, %f164;
	@!%p37 bra 	$Lt_33_88322;
	sub.ftz.f32 	%f174, %f164, %f170;
	mov.f32 	%f175, 0f358637bd;   	// 1e-006
	setp.lt.ftz.f32 	%p38, %f174, %f175;
	@!%p38 bra 	$Lt_33_88322;
	mov.s32 	%r86, 1;
	bra.uni 	$L_33_71682;
$Lt_33_88322:
$L_33_71938:
	mov.s32 	%r86, 0;
$L_33_71682:
	mov.f32 	%f176, 0f3f800000;   	// 1
	add.ftz.f32 	%f177, %f171, %f176;
	mov.f32 	%f178, 0f00000000;   	// 0
	setp.lt.ftz.f32 	%p39, %f171, %f178;
	selp.f32 	%f179, %f177, %f171, %p39;
	set.ge.ftz.u32.f32 	%r87, %f179, %f169;
	neg.s32 	%r88, %r87;
	mov.f32 	%f180, 0f00000000;   	// 0
	set.ge.ftz.u32.f32 	%r89, %f169, %f180;
	neg.s32 	%r90, %r89;
	and.b32 	%r91, %r88, %r90;
	mov.u32 	%r92, 0;
	set.ne.u32.u32 	%r93, %r86, %r92;
	neg.s32 	%r94, %r93;
	or.b32 	%r95, %r94, %r91;
	mov.u32 	%r96, 0;
	setp.eq.s32 	%p40, %r95, %r96;
	@%p40 bra 	$Lt_33_88834;
	mov.f32 	%f173, 0f3f800000;   	// 1
	bra.uni 	$L_33_71426;
$Lt_33_88834:
	mov.f32 	%f173, 0f00000000;   	// 0
$L_33_71426:
$L_33_70658:
	.loc	24	173	0
	mov.f32 	%f181, 0f3f800000;   	// 1
	setp.lt.ftz.f32 	%p41, %f173, %f181;
	@!%p41 bra 	$Lt_33_83714;
	.loc	24	180	0
	ld.param.f32 	%f182, [__cudaparm_SecondaryColorCorrection___val_paraminSecondaryParameters+16];
	sub.ftz.f32 	%f183, %f164, %f182;
	sub.ftz.f32 	%f184, %f163, %f183;
	mov.f32 	%f185, 0f3f800000;   	// 1
	add.ftz.f32 	%f186, %f184, %f185;
	mov.f32 	%f187, 0f00000000;   	// 0
	setp.lt.ftz.f32 	%p42, %f184, %f187;
	selp.f32 	%f188, %f186, %f184, %p42;
	.loc	24	108	0
	sub.ftz.f32 	%f189, %f164, %f183;
	mov.f32 	%f190, 0f00000000;   	// 0
	setp.lt.ftz.f32 	%p43, %f189, %f190;
	mov.f32 	%f191, 0f3f800000;   	// 1
	add.ftz.f32 	%f192, %f189, %f191;
	selp.f32 	%f193, %f192, %f189, %p43;
	sub.ftz.f32 	%f194, %f182, %f164;
	set.lt.ftz.u32.f32 	%r97, %f188, %f194;
	neg.s32 	%r98, %r97;
	set.ge.ftz.u32.f32 	%r99, %f188, %f193;
	neg.s32 	%r100, %r99;
	mov.f32 	%f195, 0f00000000;   	// 0
	set.lt.ftz.u32.f32 	%r101, %f188, %f195;
	neg.s32 	%r102, %r101;
	or.b32 	%r103, %r100, %r102;
	or.b32 	%r104, %r98, %r103;
	mov.u32 	%r105, 0;
	setp.eq.s32 	%p44, %r104, %r105;
	@%p44 bra 	$L_33_72706;
	.loc	24	111	0
	mov.f32 	%f196, 0f00000000;   	// 0
	bra.uni 	$L_33_72450;
$L_33_72706:
	.loc	24	119	0
	div.approx.ftz.f32 	%f196, %f188, %f193;
$L_33_72450:
	.loc	24	180	0
	mov.f32 	%f197, %f196;
	mov.f32 	%f198, 0f3f800000;   	// 1
	setp.lt.ftz.f32 	%p45, %f196, %f198;
	@!%p45 bra 	$Lt_33_84226;
	.loc	24	188	0
	sub.ftz.f32 	%f199, %f163, %f170;
	mov.f32 	%f200, 0f3f800000;   	// 1
	add.ftz.f32 	%f201, %f199, %f200;
	mov.f32 	%f202, 0f00000000;   	// 0
	setp.lt.ftz.f32 	%p46, %f199, %f202;
	selp.f32 	%f203, %f201, %f199, %p46;
	.loc	24	143	0
	ld.param.f32 	%f204, [__cudaparm_SecondaryColorCorrection___val_paraminSecondaryParameters+24];
	add.ftz.f32 	%f205, %f204, %f170;
	sub.ftz.f32 	%f206, %f205, %f170;
	mov.f32 	%f207, 0f00000000;   	// 0
	setp.lt.ftz.f32 	%p47, %f206, %f207;
	mov.f32 	%f208, 0f3f800000;   	// 1
	add.ftz.f32 	%f209, %f206, %f208;
	selp.f32 	%f210, %f209, %f206, %p47;
	mov.f32 	%f211, 0f3f800000;   	// 1
	sub.ftz.f32 	%f212, %f211, %f170;
	set.gt.ftz.u32.f32 	%r106, %f203, %f212;
	neg.s32 	%r107, %r106;
	set.ge.ftz.u32.f32 	%r108, %f203, %f210;
	neg.s32 	%r109, %r108;
	mov.f32 	%f213, 0f00000000;   	// 0
	set.lt.ftz.u32.f32 	%r110, %f203, %f213;
	neg.s32 	%r111, %r110;
	or.b32 	%r112, %r109, %r111;
	or.b32 	%r113, %r107, %r112;
	mov.u32 	%r114, 0;
	setp.eq.s32 	%p48, %r113, %r114;
	@%p48 bra 	$L_33_73218;
	.loc	24	146	0
	mov.f32 	%f214, 0f00000000;   	// 0
	bra.uni 	$L_33_72962;
$L_33_73218:
	.loc	24	154	0
	mov.f32 	%f215, 0f3f800000;   	// 1
	div.approx.ftz.f32 	%f216, %f203, %f210;
	sub.ftz.f32 	%f214, %f215, %f216;
$L_33_72962:
	.loc	24	188	0
	mov.f32 	%f217, %f214;
	bra.uni 	$Lt_33_83458;
$Lt_33_84226:
	mov.f32 	%f217, 0f00000000;   	// 0
	bra.uni 	$Lt_33_83458;
$Lt_33_83714:
	mov.f32 	%f217, 0f00000000;   	// 0
	mov.f32 	%f197, 0f00000000;   	// 0
$Lt_33_83458:
	.loc	24	311	0
	ld.param.s8 	%r115, [__cudaparm_SecondaryColorCorrection___val_paraminSecondaryParameters+60];
	mov.s32 	%r116, 0;
	setp.ne.s32 	%p49, %r115, %r116;
	max.ftz.f32 	%f218, %f217, %f197;
	max.ftz.f32 	%f219, %f218, %f173;
	mov.f32 	%f220, 0f3f800000;   	// 1
	sub.ftz.f32 	%f221, %f220, %f219;
	selp.f32 	%f222, %f221, %f219, %p49;
	mov.f32 	%f223, 0f00000000;   	// 0
	max.ftz.f32 	%f224, %f222, %f223;
	mov.f32 	%f225, 0f3f800000;   	// 1
	min.ftz.f32 	%f226, %f224, %f225;
	mov.f32 	%f227, 0f00000000;   	// 0
	setp.neu.ftz.f32 	%p50, %f226, %f227;
	@!%p50 bra 	$Lt_33_84738;
	.loc	24	71	0
	ld.param.f32 	%f228, [__cudaparm_SecondaryColorCorrection___val_paraminSecondaryParameters+36];
	ld.param.f32 	%f229, [__cudaparm_SecondaryColorCorrection___val_paraminSecondaryParameters+44];
	sub.ftz.f32 	%f230, %f229, %f228;
	setp.le.ftz.f32 	%p51, %f228, %f229;
	@!%p51 bra 	$Lt_33_89090;
	mov.f32 	%f231, 0f358637bd;   	// 1e-006
	setp.lt.ftz.f32 	%p52, %f230, %f231;
	@!%p52 bra 	$Lt_33_89090;
	.loc	24	74	0
	mov.f32 	%f232, 0f00000000;   	// 0
	bra.uni 	$L_33_74242;
$Lt_33_89090:
$L_33_73730:
	.loc	24	77	0
	setp.gt.ftz.f32 	%p53, %f228, %f229;
	@!%p53 bra 	$Lt_33_89602;
	sub.ftz.f32 	%f233, %f228, %f229;
	mov.f32 	%f234, 0f358637bd;   	// 1e-006
	setp.lt.ftz.f32 	%p54, %f233, %f234;
	@!%p54 bra 	$Lt_33_89602;
	mov.s32 	%r117, 1;
	bra.uni 	$L_33_74498;
$Lt_33_89602:
$L_33_74754:
	mov.s32 	%r117, 0;
$L_33_74498:
	ld.const.f32 	%f235, [kRGB32f_To_601YPbPr+4];
	mul.ftz.f32 	%f236, %f235, %f2;
	ld.const.f32 	%f237, [kRGB32f_To_601YPbPr+0];
	fma.rn.ftz.f32 	%f238, %f237, %f3, %f236;
	ld.const.f32 	%f239, [kRGB32f_To_601YPbPr+8];
	fma.rn.ftz.f32 	%f240, %f239, %f1, %f238;
	mov.f32 	%f241, 0f00000000;   	// 0
	max.ftz.f32 	%f242, %f240, %f241;
	mov.f32 	%f243, 0f3f800000;   	// 1
	min.ftz.f32 	%f244, %f242, %f243;
	sub.ftz.f32 	%f245, %f244, %f228;
	mov.f32 	%f246, 0f00000000;   	// 0
	setp.lt.ftz.f32 	%p55, %f245, %f246;
	mov.f32 	%f247, 0f3f800000;   	// 1
	add.ftz.f32 	%f248, %f245, %f247;
	selp.f32 	%f249, %f248, %f245, %p55;
	mov.f32 	%f250, 0f3f800000;   	// 1
	add.ftz.f32 	%f251, %f230, %f250;
	mov.f32 	%f252, 0f00000000;   	// 0
	setp.lt.ftz.f32 	%p56, %f230, %f252;
	selp.f32 	%f253, %f251, %f230, %p56;
	set.ge.ftz.u32.f32 	%r118, %f253, %f249;
	neg.s32 	%r119, %r118;
	mov.f32 	%f254, 0f00000000;   	// 0
	set.ge.ftz.u32.f32 	%r120, %f249, %f254;
	neg.s32 	%r121, %r120;
	and.b32 	%r122, %r119, %r121;
	mov.u32 	%r123, 0;
	set.ne.u32.u32 	%r124, %r117, %r123;
	neg.s32 	%r125, %r124;
	or.b32 	%r126, %r125, %r122;
	mov.u32 	%r127, 0;
	setp.eq.s32 	%p57, %r126, %r127;
	@%p57 bra 	$Lt_33_90114;
	mov.f32 	%f232, 0f3f800000;   	// 1
	bra.uni 	$L_33_74242;
$Lt_33_90114:
	mov.f32 	%f232, 0f00000000;   	// 0
$L_33_74242:
$L_33_73474:
	.loc	24	173	0
	mov.f32 	%f255, 0f3f800000;   	// 1
	setp.lt.ftz.f32 	%p58, %f232, %f255;
	@!%p58 bra 	$Lt_33_85250;
	.loc	24	108	0
	ld.const.f32 	%f256, [kRGB32f_To_601YPbPr+4];
	mul.ftz.f32 	%f257, %f256, %f2;
	ld.param.f32 	%f258, [__cudaparm_SecondaryColorCorrection___val_paraminSecondaryParameters+32];
	sub.ftz.f32 	%f259, %f228, %f258;
	ld.const.f32 	%f260, [kRGB32f_To_601YPbPr+0];
	fma.rn.ftz.f32 	%f261, %f260, %f3, %f257;
	sub.ftz.f32 	%f262, %f228, %f259;
	ld.const.f32 	%f263, [kRGB32f_To_601YPbPr+8];
	fma.rn.ftz.f32 	%f264, %f263, %f1, %f261;
	mov.f32 	%f265, 0f00000000;   	// 0
	setp.lt.ftz.f32 	%p59, %f262, %f265;
	mov.f32 	%f266, 0f3f800000;   	// 1
	add.ftz.f32 	%f267, %f262, %f266;
	mov.f32 	%f268, 0f00000000;   	// 0
	max.ftz.f32 	%f269, %f264, %f268;
	selp.f32 	%f270, %f267, %f262, %p59;
	mov.f32 	%f271, 0f3f800000;   	// 1
	min.ftz.f32 	%f244, %f269, %f271;
	sub.ftz.f32 	%f272, %f244, %f259;
	mov.f32 	%f273, 0f00000000;   	// 0
	setp.lt.ftz.f32 	%p60, %f272, %f273;
	mov.f32 	%f274, 0f3f800000;   	// 1
	add.ftz.f32 	%f275, %f272, %f274;
	selp.f32 	%f276, %f275, %f272, %p60;
	sub.ftz.f32 	%f277, %f258, %f228;
	set.lt.ftz.u32.f32 	%r128, %f276, %f277;
	neg.s32 	%r129, %r128;
	set.le.ftz.u32.f32 	%r130, %f270, %f276;
	neg.s32 	%r131, %r130;
	mov.f32 	%f278, 0f00000000;   	// 0
	set.lt.ftz.u32.f32 	%r132, %f276, %f278;
	neg.s32 	%r133, %r132;
	or.b32 	%r134, %r131, %r133;
	or.b32 	%r135, %r129, %r134;
	mov.u32 	%r136, 0;
	setp.eq.s32 	%p61, %r135, %r136;
	@%p61 bra 	$L_33_75522;
	.loc	24	111	0
	mov.f32 	%f279, 0f00000000;   	// 0
	bra.uni 	$L_33_75266;
$L_33_75522:
	.loc	24	119	0
	div.approx.ftz.f32 	%f279, %f276, %f270;
$L_33_75266:
	.loc	24	180	0
	mov.f32 	%f280, %f279;
	mov.f32 	%f281, 0f3f800000;   	// 1
	setp.lt.ftz.f32 	%p62, %f279, %f281;
	@!%p62 bra 	$Lt_33_85762;
	.loc	24	143	0
	ld.param.f32 	%f282, [__cudaparm_SecondaryColorCorrection___val_paraminSecondaryParameters+40];
	add.ftz.f32 	%f283, %f282, %f229;
	sub.ftz.f32 	%f284, %f283, %f229;
	mov.f32 	%f285, 0f00000000;   	// 0
	setp.lt.ftz.f32 	%p63, %f284, %f285;
	mov.f32 	%f286, 0f3f800000;   	// 1
	add.ftz.f32 	%f287, %f284, %f286;
	selp.f32 	%f288, %f287, %f284, %p63;
	sub.ftz.f32 	%f289, %f244, %f229;
	mov.f32 	%f290, 0f00000000;   	// 0
	setp.lt.ftz.f32 	%p64, %f289, %f290;
	mov.f32 	%f291, 0f3f800000;   	// 1
	add.ftz.f32 	%f292, %f289, %f291;
	selp.f32 	%f293, %f292, %f289, %p64;
	mov.f32 	%f294, 0f3f800000;   	// 1
	sub.ftz.f32 	%f295, %f294, %f229;
	set.gt.ftz.u32.f32 	%r137, %f293, %f295;
	neg.s32 	%r138, %r137;
	set.le.ftz.u32.f32 	%r139, %f288, %f293;
	neg.s32 	%r140, %r139;
	mov.f32 	%f296, 0f00000000;   	// 0
	set.lt.ftz.u32.f32 	%r141, %f293, %f296;
	neg.s32 	%r142, %r141;
	or.b32 	%r143, %r140, %r142;
	or.b32 	%r144, %r138, %r143;
	mov.u32 	%r145, 0;
	setp.eq.s32 	%p65, %r144, %r145;
	@%p65 bra 	$L_33_76034;
	.loc	24	146	0
	mov.f32 	%f297, 0f00000000;   	// 0
	bra.uni 	$L_33_75778;
$L_33_76034:
	.loc	24	154	0
	mov.f32 	%f298, 0f3f800000;   	// 1
	div.approx.ftz.f32 	%f299, %f293, %f288;
	sub.ftz.f32 	%f297, %f298, %f299;
$L_33_75778:
	.loc	24	188	0
	mov.f32 	%f300, %f297;
	bra.uni 	$Lt_33_84994;
$Lt_33_85762:
	mov.f32 	%f300, 0f00000000;   	// 0
	bra.uni 	$Lt_33_84994;
$Lt_33_85250:
	mov.f32 	%f300, 0f00000000;   	// 0
	mov.f32 	%f280, 0f00000000;   	// 0
$Lt_33_84994:
	.loc	24	278	0
	max.ftz.f32 	%f301, %f300, %f280;
	max.ftz.f32 	%f302, %f301, %f232;
	mov.f32 	%f303, 0f3f800000;   	// 1
	sub.ftz.f32 	%f304, %f303, %f302;
	selp.f32 	%f305, %f304, %f302, %p49;
	mov.f32 	%f306, 0f00000000;   	// 0
	max.ftz.f32 	%f307, %f305, %f306;
	mov.f32 	%f308, 0f3f800000;   	// 1
	min.ftz.f32 	%f309, %f307, %f308;
	add.ftz.f32 	%f310, %f226, %f61;
	add.ftz.f32 	%f311, %f309, %f310;
	mov.f32 	%f312, 0f3eaaaaab;   	// 0.333333
	mul.ftz.f32 	%f313, %f311, %f312;
	mov.f32 	%f314, 0f00000000;   	// 0
	mov.f32 	%f315, 0f00000000;   	// 0
	setp.neu.ftz.f32 	%p66, %f309, %f315;
	selp.f32 	%f316, %f313, %f314, %p66;
	bra.uni 	$Lt_33_82946;
$Lt_33_84738:
	mov.f32 	%f316, 0f00000000;   	// 0
	bra.uni 	$Lt_33_82946;
$Lt_33_83202:
	mov.f32 	%f316, 0f00000000;   	// 0
$Lt_33_82946:
	ld.param.f32 	%f317, [__cudaparm_SecondaryColorCorrection___val_paraminSecondaryParameters+56];
	mov.f32 	%f318, 0f00000000;   	// 0
	setp.neu.ftz.f32 	%p67, %f317, %f318;
	@!%p67 bra 	$Lt_33_86018;
	.loc	24	327	0
	lg2.approx.ftz.f32 	%f319, %f316;
	mov.f32 	%f320, 0f41200000;   	// 10
	lg2.approx.ftz.f32 	%f321, %f320;
	mul.ftz.f32 	%f322, %f317, %f321;
	ex2.approx.ftz.f32 	%f323, %f322;
	rcp.approx.ftz.f32 	%f324, %f323;
	mul.ftz.f32 	%f325, %f319, %f324;
	ex2.approx.ftz.f32 	%f316, %f325;
$Lt_33_86018:
	.loc	21	126	0
	mov.f32 	%f326, 0f00000000;   	// 0
	max.ftz.f32 	%f327, %f316, %f326;
	mov.f32 	%f328, 0f3f800000;   	// 1
	min.ftz.f32 	%f329, %f327, %f328;
	ld.param.u64 	%rd7, [__cudaparm_SecondaryColorCorrection_inSecondaryMask];
	ld.param.s32 	%r146, [__cudaparm_SecondaryColorCorrection_inSecondaryPitch];
	mul.lo.s32 	%r147, %r146, %r10;
	add.s32 	%r148, %r8, %r147;
	cvt.s64.s32 	%rd8, %r148;
	mul.wide.s32 	%rd9, %r148, 4;
	add.u64 	%rd10, %rd7, %rd9;
	st.global.f32 	[%rd10+0], %f329;
$Lt_33_76290:
	.loc	24	361	0
	exit;
$LDWend_SecondaryColorCorrection:
	} // SecondaryColorCorrection
	.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};
	.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};
	.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};

