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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

	.visible .func (.param .f32 __cudaretf__Z21CalculateThreeWayLumaff50ThreeWayColorCorrectorParameters_BalanceTonalRange39ThreeWayColorCorrectorParameters_Levels) _Z21CalculateThreeWayLumaff50ThreeWayColorCorrectorParameters_BalanceTonalRange39ThreeWayColorCorrectorParameters_Levels (.param .f32 __cudaparmf1__Z21CalculateThreeWayLumaff50ThreeWayColorCorrectorParameters_BalanceTonalRange39ThreeWayColorCorrectorParameters_Levels, .param .f32 __cudaparmf2__Z21CalculateThreeWayLumaff50ThreeWayColorCorrectorParameters_BalanceTonalRange39ThreeWayColorCorrectorParameters_Levels, .param .align 4 .b8 __cudaparmf3__Z21CalculateThreeWayLumaff50ThreeWayColorCorrectorParameters_BalanceTonalRange39ThreeWayColorCorrectorParameters_Levels[92], .param .align 4 .b8 __cudaparmf4__Z21CalculateThreeWayLumaff50ThreeWayColorCorrectorParameters_BalanceTonalRange39ThreeWayColorCorrectorParameters_Levels[80])

	.visible .func (.param .f32 __cudaretf__Z4LERPIfET_S0_S0_S0_) _Z4LERPIfET_S0_S0_S0_ (.param .f32 __cudaparmf1__Z4LERPIfET_S0_S0_S0_, .param .f32 __cudaparmf2__Z4LERPIfET_S0_S0_S0_, .param .f32 __cudaparmf3__Z4LERPIfET_S0_S0_S0_)

	.visible .func (.param .f32 __cudaretf__Z6Read2DIfET_PKS0_iii) _Z6Read2DIfET_PKS0_iii (.param .u64 __cudaparmf1__Z6Read2DIfET_PKS0_iii, .param .s32 __cudaparmf2__Z6Read2DIfET_PKS0_iii, .param .s32 __cudaparmf3__Z6Read2DIfET_PKS0_iii, .param .s32 __cudaparmf4__Z6Read2DIfET_PKS0_iii)

	//-----------------------------------------------------------
	// Compiling C:/Users/dvaeng/AppData/Local/Temp/tmpxft_00003c10_00000000-11_ThreeWayColorCorrector.cpp3.i (C:/Users/dvaeng/AppData/Local/Temp/ccBI#.a15116)
	//-----------------------------------------------------------

	//-----------------------------------------------------------
	// 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_00003c10_00000000-10_ThreeWayColorCorrector.cudafe2.gpu"
	.file	2	"c:\Mulder64\shared\adobe\MediaCore\GPUFoundation\API\Inc\GPUFoundation/PixelFormat.h"
	.file	3	"c:\Mulder64\shared\adobe\MediaCore\Display\Inc\CUDA/Effects/ThreeWayColorCorrector.h"
	.file	4	"c:\Mulder64\shared\adobe\MediaCore\GPUFoundation\API\Inc\GPUFoundation/KernelSupport/PixelRGB.h"
	.file	5	"c:\Mulder64\shared\adobe\MediaCore\GPUFoundation\API\Inc\GPUFoundation/KernelSupport/PixelYUV.h"
	.file	6	"C:\Program Files (x86)\Microsoft Visual Studio 9.0\VC\include\crtdefs.h"
	.file	7	"c:\Mulder64\shared\adobe\MediaCore\External\3rdParty\NVIDIA\CUDA\win\include\crt/device_runtime.h"
	.file	8	"c:\Mulder64\shared\adobe\MediaCore\External\3rdParty\NVIDIA\CUDA\win\include\host_defines.h"
	.file	9	"c:\Mulder64\shared\adobe\MediaCore\External\3rdParty\NVIDIA\CUDA\win\include\builtin_types.h"
	.file	10	"c:\mulder64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\include\device_types.h"
	.file	11	"c:\mulder64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\include\driver_types.h"
	.file	12	"c:\mulder64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\include\surface_types.h"
	.file	13	"c:\mulder64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\include\texture_types.h"
	.file	14	"c:\mulder64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\include\vector_types.h"
	.file	15	"c:\mulder64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\include\builtin_types.h"
	.file	16	"c:\mulder64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\include\host_defines.h"
	.file	17	"c:\Mulder64\shared\adobe\MediaCore\External\3rdParty\NVIDIA\CUDA\win\include\device_launch_parameters.h"
	.file	18	"c:\mulder64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\include\crt\storage_class.h"
	.file	19	"C:\Program Files (x86)\Microsoft Visual Studio 9.0\VC\include\time.h"
	.file	20	"c:\Mulder64\shared\adobe\MediaCore\GPUFoundation\API\Inc\GPUFoundation/KernelSupport/Utils.h"
	.file	21	"c:\Mulder64\shared\adobe\MediaCore\GPUFoundation\API\Inc\GPUFoundation/KernelSupport/VectorUtils.h"
	.file	22	"c:/Mulder64/shared/adobe/MediaCore/Display/Src/CUDA/Effects/ThreeWayColorCorrector.cu"
	.file	23	"c:\Mulder64\shared\adobe\MediaCore\GPUFoundation\API\Inc\GPUFoundation/Numeric.h"
	.file	24	"c:\Mulder64\shared\adobe\MediaCore\External\3rdParty\NVIDIA\CUDA\win\include\common_functions.h"
	.file	25	"c:\mulder64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\include\math_functions.h"
	.file	26	"c:\mulder64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\include\math_constants.h"
	.file	27	"c:\mulder64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\include\device_functions.h"
	.file	28	"c:\mulder64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\include\sm_11_atomic_functions.h"
	.file	29	"c:\mulder64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\include\sm_12_atomic_functions.h"
	.file	30	"c:\mulder64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\include\sm_13_double_functions.h"
	.file	31	"c:\mulder64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\include\sm_20_atomic_functions.h"
	.file	32	"c:\mulder64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\include\sm_20_intrinsics.h"
	.file	33	"c:\mulder64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\include\surface_functions.h"
	.file	34	"c:\mulder64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\include\texture_fetch_functions.h"
	.file	35	"c:\mulder64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\include\math_functions_dbl_ptx3.h"
	.file	36	"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	20	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	20	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	20	73	0
$LDWbegin__Z17Standard2DKernelXv:
	.loc	20	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	20	77	0
$LDWbegin__Z17Standard2DKernelYv:
	.loc	20	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	20	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	20	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	20	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	20	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	20	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	20	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	20	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	20	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	20	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	20	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	20	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	20	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	20	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	20	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	20	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	20	152	0
$LDWbegin__Z4Randj:
	ld.param.u32 	%r1, [__cudaparmf1__Z4Randj];
	mov.s32 	%r2, %r1;
	.loc	20	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	20	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	20	139	0
	sub.u32 	%r7, %r2, %r4;
	sub.u32 	%r8, %r7, %r6;
	shr.u32 	%r9, %r6, 13;
	xor.b32 	%r10, %r8, %r9;
	.loc	20	140	0
	sub.u32 	%r11, %r4, %r6;
	sub.u32 	%r12, %r11, %r10;
	shl.b32 	%r13, %r10, 8;
	xor.b32 	%r14, %r12, %r13;
	.loc	20	141	0
	sub.u32 	%r15, %r6, %r10;
	sub.u32 	%r16, %r15, %r14;
	shr.u32 	%r17, %r14, 13;
	xor.b32 	%r18, %r16, %r17;
	.loc	20	142	0
	sub.u32 	%r19, %r10, %r14;
	sub.u32 	%r20, %r19, %r18;
	shr.u32 	%r21, %r18, 12;
	xor.b32 	%r22, %r20, %r21;
	.loc	20	143	0
	sub.u32 	%r23, %r14, %r18;
	sub.u32 	%r24, %r23, %r22;
	shl.b32 	%r25, %r22, 16;
	xor.b32 	%r26, %r24, %r25;
	.loc	20	144	0
	sub.u32 	%r27, %r18, %r22;
	sub.u32 	%r28, %r27, %r26;
	shr.u32 	%r29, %r26, 5;
	xor.b32 	%r30, %r28, %r29;
	.loc	20	145	0
	sub.u32 	%r31, %r22, %r26;
	sub.u32 	%r32, %r31, %r30;
	shr.u32 	%r33, %r30, 3;
	xor.b32 	%r34, %r32, %r33;
	.loc	20	146	0
	sub.u32 	%r35, %r26, %r30;
	sub.u32 	%r36, %r35, %r34;
	shl.b32 	%r37, %r34, 10;
	xor.b32 	%r38, %r36, %r37;
	.loc	20	147	0
	sub.u32 	%r39, %r30, %r34;
	sub.u32 	%r40, %r39, %r38;
	shr.u32 	%r41, %r38, 15;
	xor.b32 	%r42, %r40, %r41;
	.loc	20	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	20	175	0
$LDWbegin__Z6Rand2Dj:
	ld.param.u32 	%r1, [__cudaparmf1__Z6Rand2Dj];
	mov.s32 	%r2, %r1;
	.loc	20	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	20	144	0
	sub.u32 	%r33, %r25, %r28;
	sub.u32 	%r34, %r33, %r32;
	shr.u32 	%r35, %r32, 5;
	xor.b32 	%r36, %r34, %r35;
	.loc	20	145	0
	sub.u32 	%r37, %r28, %r32;
	sub.u32 	%r38, %r37, %r36;
	shr.u32 	%r39, %r36, 3;
	xor.b32 	%r40, %r38, %r39;
	.loc	20	146	0
	sub.u32 	%r41, %r32, %r36;
	sub.u32 	%r42, %r41, %r40;
	shl.b32 	%r43, %r40, 10;
	xor.b32 	%r44, %r42, %r43;
	.loc	20	147	0
	sub.u32 	%r45, %r36, %r40;
	sub.u32 	%r46, %r45, %r44;
	shr.u32 	%r47, %r44, 15;
	xor.b32 	%r48, %r46, %r47;
	.loc	20	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	20	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	20	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	20	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	20	115	0
	mul.lo.s32 	%r7, %r2, %r6;
	add.s32 	%r8, %r4, %r7;
	cvt.s64.s32 	%rd3, %r8;
	mul.wide.s32 	%rd4, %r8, 16;
	add.u64 	%rd5, %rd2, %rd4;
	ld.v4.f32 	{%f1,%f2,%f3,%f4}, [%rd5+0];
	st.param.f32 	[__cudaretf__Z6Read2DI6float4ET_PKS1_iii+0], %f1;
	st.param.f32 	[__cudaretf__Z6Read2DI6float4ET_PKS1_iii+4], %f2;
	st.param.f32 	[__cudaretf__Z6Read2DI6float4ET_PKS1_iii+8], %f3;
	st.param.f32 	[__cudaretf__Z6Read2DI6float4ET_PKS1_iii+12], %f4;
	ret;
$LDWend__Z6Read2DI6float4ET_PKS1_iii:
	} // _Z6Read2DI6float4ET_PKS1_iii

	.visible .func _Z7Write2DI7ushort4EvT_PS1_iii (.param .align 8 .b8 __cudaparmf1__Z7Write2DI7ushort4EvT_PS1_iii[8], .param .u64 __cudaparmf2__Z7Write2DI7ushort4EvT_PS1_iii, .param .s32 __cudaparmf3__Z7Write2DI7ushort4EvT_PS1_iii, .param .s32 __cudaparmf4__Z7Write2DI7ushort4EvT_PS1_iii, .param .s32 __cudaparmf5__Z7Write2DI7ushort4EvT_PS1_iii)
	{
	.reg .u32 %r<18>;
	.reg .u64 %rd<7>;
	.loc	20	125	0
$LDWbegin__Z7Write2DI7ushort4EvT_PS1_iii:
	ld.param.u16 	%r1, [__cudaparmf1__Z7Write2DI7ushort4EvT_PS1_iii+0];
	mov.s32 	%r2, %r1;
	ld.param.u16 	%r3, [__cudaparmf1__Z7Write2DI7ushort4EvT_PS1_iii+2];
	mov.s32 	%r4, %r3;
	ld.param.u16 	%r5, [__cudaparmf1__Z7Write2DI7ushort4EvT_PS1_iii+4];
	mov.s32 	%r6, %r5;
	ld.param.u16 	%r7, [__cudaparmf1__Z7Write2DI7ushort4EvT_PS1_iii+6];
	mov.s32 	%r8, %r7;
	ld.param.u64 	%rd1, [__cudaparmf2__Z7Write2DI7ushort4EvT_PS1_iii];
	mov.s64 	%rd2, %rd1;
	ld.param.u32 	%r9, [__cudaparmf3__Z7Write2DI7ushort4EvT_PS1_iii];
	mov.s32 	%r10, %r9;
	ld.param.u32 	%r11, [__cudaparmf4__Z7Write2DI7ushort4EvT_PS1_iii];
	mov.s32 	%r12, %r11;
	ld.param.u32 	%r13, [__cudaparmf5__Z7Write2DI7ushort4EvT_PS1_iii];
	mov.s32 	%r14, %r13;
	.loc	20	126	0
	mul.lo.s32 	%r15, %r10, %r14;
	add.s32 	%r16, %r12, %r15;
	cvt.s64.s32 	%rd3, %r16;
	mul.wide.s32 	%rd4, %r16, 8;
	add.u64 	%rd5, %rd2, %rd4;
	st.v4.u16 	[%rd5+0], {%r2,%r4,%r6,%r8};
	.loc	20	127	0
	ret;
$LDWend__Z7Write2DI7ushort4EvT_PS1_iii:
	} // _Z7Write2DI7ushort4EvT_PS1_iii

	.visible .func _Z7Write2DI6float4EvT_PS1_iii (.param .align 16 .b8 __cudaparmf1__Z7Write2DI6float4EvT_PS1_iii[16], .param .u64 __cudaparmf2__Z7Write2DI6float4EvT_PS1_iii, .param .s32 __cudaparmf3__Z7Write2DI6float4EvT_PS1_iii, .param .s32 __cudaparmf4__Z7Write2DI6float4EvT_PS1_iii, .param .s32 __cudaparmf5__Z7Write2DI6float4EvT_PS1_iii)
	{
	.reg .u32 %r<10>;
	.reg .u64 %rd<7>;
	.reg .f32 %f<10>;
	.loc	20	125	0
$LDWbegin__Z7Write2DI6float4EvT_PS1_iii:
	ld.param.f32 	%f1, [__cudaparmf1__Z7Write2DI6float4EvT_PS1_iii+0];
	mov.f32 	%f2, %f1;
	ld.param.f32 	%f3, [__cudaparmf1__Z7Write2DI6float4EvT_PS1_iii+4];
	mov.f32 	%f4, %f3;
	ld.param.f32 	%f5, [__cudaparmf1__Z7Write2DI6float4EvT_PS1_iii+8];
	mov.f32 	%f6, %f5;
	ld.param.f32 	%f7, [__cudaparmf1__Z7Write2DI6float4EvT_PS1_iii+12];
	mov.f32 	%f8, %f7;
	ld.param.u64 	%rd1, [__cudaparmf2__Z7Write2DI6float4EvT_PS1_iii];
	mov.s64 	%rd2, %rd1;
	ld.param.u32 	%r1, [__cudaparmf3__Z7Write2DI6float4EvT_PS1_iii];
	mov.s32 	%r2, %r1;
	ld.param.u32 	%r3, [__cudaparmf4__Z7Write2DI6float4EvT_PS1_iii];
	mov.s32 	%r4, %r3;
	ld.param.u32 	%r5, [__cudaparmf5__Z7Write2DI6float4EvT_PS1_iii];
	mov.s32 	%r6, %r5;
	.loc	20	126	0
	mul.lo.s32 	%r7, %r2, %r6;
	add.s32 	%r8, %r4, %r7;
	cvt.s64.s32 	%rd3, %r8;
	mul.wide.s32 	%rd4, %r8, 16;
	add.u64 	%rd5, %rd2, %rd4;
	st.v4.f32 	[%rd5+0], {%f2,%f4,%f6,%f8};
	.loc	20	127	0
	ret;
$LDWend__Z7Write2DI6float4EvT_PS1_iii:
	} // _Z7Write2DI6float4EvT_PS1_iii

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

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

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

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

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

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

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

	.visible .func (.param .align 16 .b8 __cudaretf__Z18SwapComponentOrderI6float4ET_RKS1_[16]) _Z18SwapComponentOrderI6float4ET_RKS1_ (.param .u64 __cudaparmf1__Z18SwapComponentOrderI6float4ET_RKS1_)
	{
	.reg .u64 %rd<4>;
	.reg .f32 %f<6>;
	.loc	21	264	0
$LDWbegin__Z18SwapComponentOrderI6float4ET_RKS1_:
	ld.param.u64 	%rd1, [__cudaparmf1__Z18SwapComponentOrderI6float4ET_RKS1_];
	mov.s64 	%rd2, %rd1;
	.loc	21	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 (.param .f32 __cudaretf__Z21CalculateThreeWayLumaff50ThreeWayColorCorrectorParameters_BalanceTonalRange39ThreeWayColorCorrectorParameters_Levels) _Z21CalculateThreeWayLumaff50ThreeWayColorCorrectorParameters_BalanceTonalRange39ThreeWayColorCorrectorParameters_Levels (.param .f32 __cudaparmf1__Z21CalculateThreeWayLumaff50ThreeWayColorCorrectorParameters_BalanceTonalRange39ThreeWayColorCorrectorParameters_Levels, .param .f32 __cudaparmf2__Z21CalculateThreeWayLumaff50ThreeWayColorCorrectorParameters_BalanceTonalRange39ThreeWayColorCorrectorParameters_Levels, .param .align 4 .b8 __cudaparmf3__Z21CalculateThreeWayLumaff50ThreeWayColorCorrectorParameters_BalanceTonalRange39ThreeWayColorCorrectorParameters_Levels[92], .param .align 4 .b8 __cudaparmf4__Z21CalculateThreeWayLumaff50ThreeWayColorCorrectorParameters_BalanceTonalRange39ThreeWayColorCorrectorParameters_Levels[80])
	{
	.reg .u32 %r<89>;
	.reg .f32 %f<175>;
	.reg .pred %p<15>;
	.loc	22	37	0
$LDWbegin__Z21CalculateThreeWayLumaff50ThreeWayColorCorrectorParameters_BalanceTonalRange39ThreeWayColorCorrectorParameters_Levels:
	ld.param.f32 	%f1, [__cudaparmf1__Z21CalculateThreeWayLumaff50ThreeWayColorCorrectorParameters_BalanceTonalRange39ThreeWayColorCorrectorParameters_Levels];
	mov.f32 	%f2, %f1;
	ld.param.f32 	%f3, [__cudaparmf2__Z21CalculateThreeWayLumaff50ThreeWayColorCorrectorParameters_BalanceTonalRange39ThreeWayColorCorrectorParameters_Levels];
	mov.f32 	%f4, %f3;
	ld.param.f32 	%f5, [__cudaparmf3__Z21CalculateThreeWayLumaff50ThreeWayColorCorrectorParameters_BalanceTonalRange39ThreeWayColorCorrectorParameters_Levels+0];
	mov.f32 	%f6, %f5;
	ld.param.f32 	%f7, [__cudaparmf3__Z21CalculateThreeWayLumaff50ThreeWayColorCorrectorParameters_BalanceTonalRange39ThreeWayColorCorrectorParameters_Levels+4];
	mov.f32 	%f8, %f7;
	ld.param.f32 	%f9, [__cudaparmf3__Z21CalculateThreeWayLumaff50ThreeWayColorCorrectorParameters_BalanceTonalRange39ThreeWayColorCorrectorParameters_Levels+8];
	mov.f32 	%f10, %f9;
	ld.param.f32 	%f11, [__cudaparmf3__Z21CalculateThreeWayLumaff50ThreeWayColorCorrectorParameters_BalanceTonalRange39ThreeWayColorCorrectorParameters_Levels+12];
	mov.f32 	%f12, %f11;
	ld.param.f32 	%f13, [__cudaparmf4__Z21CalculateThreeWayLumaff50ThreeWayColorCorrectorParameters_BalanceTonalRange39ThreeWayColorCorrectorParameters_Levels+0];
	mov.f32 	%f14, %f13;
	ld.param.f32 	%f15, [__cudaparmf4__Z21CalculateThreeWayLumaff50ThreeWayColorCorrectorParameters_BalanceTonalRange39ThreeWayColorCorrectorParameters_Levels+4];
	mov.f32 	%f16, %f15;
	ld.param.f32 	%f17, [__cudaparmf4__Z21CalculateThreeWayLumaff50ThreeWayColorCorrectorParameters_BalanceTonalRange39ThreeWayColorCorrectorParameters_Levels+8];
	mov.f32 	%f18, %f17;
	ld.param.f32 	%f19, [__cudaparmf4__Z21CalculateThreeWayLumaff50ThreeWayColorCorrectorParameters_BalanceTonalRange39ThreeWayColorCorrectorParameters_Levels+12];
	mov.f32 	%f20, %f19;
	ld.param.f32 	%f21, [__cudaparmf4__Z21CalculateThreeWayLumaff50ThreeWayColorCorrectorParameters_BalanceTonalRange39ThreeWayColorCorrectorParameters_Levels+16];
	mov.f32 	%f22, %f21;
	ld.param.f32 	%f23, [__cudaparmf4__Z21CalculateThreeWayLumaff50ThreeWayColorCorrectorParameters_BalanceTonalRange39ThreeWayColorCorrectorParameters_Levels+20];
	mov.f32 	%f24, %f23;
	ld.param.f32 	%f25, [__cudaparmf4__Z21CalculateThreeWayLumaff50ThreeWayColorCorrectorParameters_BalanceTonalRange39ThreeWayColorCorrectorParameters_Levels+24];
	mov.f32 	%f26, %f25;
	ld.param.f32 	%f27, [__cudaparmf4__Z21CalculateThreeWayLumaff50ThreeWayColorCorrectorParameters_BalanceTonalRange39ThreeWayColorCorrectorParameters_Levels+28];
	mov.f32 	%f28, %f27;
	ld.param.f32 	%f29, [__cudaparmf4__Z21CalculateThreeWayLumaff50ThreeWayColorCorrectorParameters_BalanceTonalRange39ThreeWayColorCorrectorParameters_Levels+32];
	mov.f32 	%f30, %f29;
	ld.param.f32 	%f31, [__cudaparmf4__Z21CalculateThreeWayLumaff50ThreeWayColorCorrectorParameters_BalanceTonalRange39ThreeWayColorCorrectorParameters_Levels+36];
	mov.f32 	%f32, %f31;
	ld.param.f32 	%f33, [__cudaparmf4__Z21CalculateThreeWayLumaff50ThreeWayColorCorrectorParameters_BalanceTonalRange39ThreeWayColorCorrectorParameters_Levels+40];
	mov.f32 	%f34, %f33;
	ld.param.f32 	%f35, [__cudaparmf4__Z21CalculateThreeWayLumaff50ThreeWayColorCorrectorParameters_BalanceTonalRange39ThreeWayColorCorrectorParameters_Levels+44];
	mov.f32 	%f36, %f35;
	ld.param.f32 	%f37, [__cudaparmf4__Z21CalculateThreeWayLumaff50ThreeWayColorCorrectorParameters_BalanceTonalRange39ThreeWayColorCorrectorParameters_Levels+48];
	mov.f32 	%f38, %f37;
	ld.param.f32 	%f39, [__cudaparmf4__Z21CalculateThreeWayLumaff50ThreeWayColorCorrectorParameters_BalanceTonalRange39ThreeWayColorCorrectorParameters_Levels+52];
	mov.f32 	%f40, %f39;
	ld.param.f32 	%f41, [__cudaparmf4__Z21CalculateThreeWayLumaff50ThreeWayColorCorrectorParameters_BalanceTonalRange39ThreeWayColorCorrectorParameters_Levels+56];
	mov.f32 	%f42, %f41;
	ld.param.f32 	%f43, [__cudaparmf4__Z21CalculateThreeWayLumaff50ThreeWayColorCorrectorParameters_BalanceTonalRange39ThreeWayColorCorrectorParameters_Levels+60];
	mov.f32 	%f44, %f43;
	ld.param.f32 	%f45, [__cudaparmf4__Z21CalculateThreeWayLumaff50ThreeWayColorCorrectorParameters_BalanceTonalRange39ThreeWayColorCorrectorParameters_Levels+64];
	mov.f32 	%f46, %f45;
	ld.param.f32 	%f47, [__cudaparmf4__Z21CalculateThreeWayLumaff50ThreeWayColorCorrectorParameters_BalanceTonalRange39ThreeWayColorCorrectorParameters_Levels+68];
	mov.f32 	%f48, %f47;
	ld.param.f32 	%f49, [__cudaparmf4__Z21CalculateThreeWayLumaff50ThreeWayColorCorrectorParameters_BalanceTonalRange39ThreeWayColorCorrectorParameters_Levels+72];
	mov.f32 	%f50, %f49;
	ld.param.f32 	%f51, [__cudaparmf4__Z21CalculateThreeWayLumaff50ThreeWayColorCorrectorParameters_BalanceTonalRange39ThreeWayColorCorrectorParameters_Levels+76];
	mov.f32 	%f52, %f51;
	.loc	22	53	0
	mov.f32 	%f53, %f2;
	.loc	22	37	0
	sub.ftz.f32 	%f54, %f16, %f14;
	sub.ftz.f32 	%f55, %f22, %f20;
	mov.f32 	%f56, 0f3f800000;    	// 1
	setp.neu.ftz.f32 	%p1, %f54, %f56;
	@%p1 bra 	$Lt_21_14082;
	mov.f32 	%f57, 0f3f800000;    	// 1
	setp.neu.ftz.f32 	%p2, %f55, %f57;
	@%p2 bra 	$Lt_21_14082;
	sub.ftz.f32 	%f58, %f26, %f24;
	mov.f32 	%f59, 0f3f800000;    	// 1
	setp.neu.ftz.f32 	%p3, %f58, %f59;
	@%p3 bra 	$Lt_21_14082;
	sub.ftz.f32 	%f60, %f32, %f30;
	mov.f32 	%f61, 0f3f800000;    	// 1
	setp.neu.ftz.f32 	%p4, %f60, %f61;
	@%p4 bra 	$Lt_21_14082;
	sub.ftz.f32 	%f62, %f36, %f34;
	mov.f32 	%f63, 0f3f800000;    	// 1
	setp.neu.ftz.f32 	%p5, %f62, %f63;
	@%p5 bra 	$Lt_21_14082;
	sub.ftz.f32 	%f64, %f42, %f40;
	mov.f32 	%f65, 0f3f800000;    	// 1
	setp.neu.ftz.f32 	%p6, %f64, %f65;
	@%p6 bra 	$Lt_21_14082;
	sub.ftz.f32 	%f66, %f46, %f44;
	mov.f32 	%f67, 0f3f800000;    	// 1
	setp.neu.ftz.f32 	%p7, %f66, %f67;
	@%p7 bra 	$Lt_21_14082;
	sub.ftz.f32 	%f68, %f52, %f50;
	mov.f32 	%f69, 0f3f800000;    	// 1
	setp.neu.ftz.f32 	%p8, %f68, %f69;
	@!%p8 bra 	$L_21_8194;
$Lt_21_14082:
$L_21_8450:
	.loc	20	143	0
	cvt.s32.u32 	%r1, %ctaid.y;
	cvt.s32.u32 	%r2, %ntid.y;
	mul.lo.s32 	%r3, %r1, %r2;
	cvt.s32.u32 	%r4, %ctaid.x;
	cvt.s32.u32 	%r5, %ntid.x;
	mul.lo.s32 	%r6, %r4, %r5;
	mov.u32 	%r7, %tid.y;
	add.u32 	%r8, %r3, %r7;
	mov.u32 	%r9, %tid.x;
	add.u32 	%r10, %r6, %r9;
	shr.u32 	%r11, %r8, 13;
	mov.s32 	%r12, 1;
	sub.s32 	%r13, %r12, %r10;
	sub.u32 	%r14, %r10, %r8;
	sub.u32 	%r15, %r13, %r8;
	xor.b32 	%r16, %r11, %r15;
	shl.b32 	%r17, %r16, 8;
	sub.u32 	%r18, %r14, %r16;
	sub.u32 	%r19, %r8, %r16;
	xor.b32 	%r20, %r17, %r18;
	shr.u32 	%r21, %r20, 13;
	sub.u32 	%r22, %r19, %r20;
	sub.u32 	%r23, %r16, %r20;
	xor.b32 	%r24, %r21, %r22;
	shr.u32 	%r25, %r24, 12;
	sub.u32 	%r26, %r23, %r24;
	xor.b32 	%r27, %r25, %r26;
	sub.u32 	%r28, %r20, %r24;
	sub.u32 	%r29, %r28, %r27;
	shl.b32 	%r30, %r27, 16;
	xor.b32 	%r31, %r29, %r30;
	.loc	20	144	0
	sub.u32 	%r32, %r24, %r27;
	sub.u32 	%r33, %r32, %r31;
	shr.u32 	%r34, %r31, 5;
	xor.b32 	%r35, %r33, %r34;
	.loc	20	145	0
	sub.u32 	%r36, %r27, %r31;
	sub.u32 	%r37, %r36, %r35;
	shr.u32 	%r38, %r35, 3;
	xor.b32 	%r39, %r37, %r38;
	.loc	20	146	0
	sub.u32 	%r40, %r31, %r35;
	sub.u32 	%r41, %r40, %r39;
	shl.b32 	%r42, %r39, 10;
	xor.b32 	%r43, %r41, %r42;
	.loc	20	147	0
	sub.u32 	%r44, %r35, %r39;
	sub.u32 	%r45, %r44, %r43;
	shr.u32 	%r46, %r43, 15;
	xor.b32 	%r47, %r45, %r46;
	.loc	22	63	0
	mov.f32 	%f70, 0f3b270d73;    	// 0.00254902
	mul.lo.u32 	%r48, %r47, 1103515245;
	add.u32 	%r49, %r48, 12345;
	shr.u32 	%r50, %r49, 16;
	and.b32 	%r51, %r50, 255;
	shl.b32 	%r52, %r51, 7;
	mul.lo.u32 	%r53, %r47, -1029531031;
	sub.u32 	%r54, %r53, 740551042;
	shr.u32 	%r55, %r54, 16;
	and.b32 	%r56, %r55, 255;
	xor.b32 	%r57, %r52, %r56;
	cvt.rn.f32.s32 	%f71, %r57;
	mov.f32 	%f72, 0f467ffe00;    	// 16383.5
	div.approx.ftz.f32 	%f73, %f71, %f72;
	mov.f32 	%f74, 0fbf800000;    	// -1
	add.ftz.f32 	%f75, %f73, %f74;
	fma.rn.ftz.f32 	%f53, %f70, %f75, %f2;
$L_21_8194:
	.loc	22	67	0
	sub.ftz.f32 	%f76, %f53, %f14;
	div.approx.ftz.f32 	%f77, %f76, %f54;
	mov.f32 	%f78, 0f00000000;    	// 0
	max.ftz.f32 	%f79, %f77, %f78;
	lg2.approx.ftz.f32 	%f80, %f79;
	mul.ftz.f32 	%f81, %f18, %f80;
	ex2.approx.ftz.f32 	%f82, %f81;
	fma.rn.ftz.f32 	%f83, %f55, %f82, %f20;
	.loc	22	79	0
	add.ftz.f32 	%f84, %f6, %f8;
	mov.f32 	%f85, 0f3f800000;    	// 1
	min.ftz.f32 	%f86, %f84, %f85;
	set.ge.ftz.u32.f32 	%r58, %f6, %f53;
	neg.s32 	%r59, %r58;
	mov.f32 	%f87, 0f00000000;    	// 0
	set.eq.ftz.u32.f32 	%r60, %f8, %f87;
	neg.s32 	%r61, %r60;
	set.ge.ftz.u32.f32 	%r62, %f86, %f53;
	neg.s32 	%r63, %r62;
	and.b32 	%r64, %r61, %r63;
	or.b32 	%r65, %r59, %r64;
	mov.u32 	%r66, 0;
	setp.eq.s32 	%p9, %r65, %r66;
	@%p9 bra 	$Lt_21_10498;
	.loc	22	87	0
	sub.ftz.f32 	%f88, %f32, %f30;
	sub.ftz.f32 	%f89, %f53, %f24;
	sub.ftz.f32 	%f90, %f26, %f24;
	div.approx.ftz.f32 	%f91, %f89, %f90;
	mov.f32 	%f92, 0f00000000;    	// 0
	max.ftz.f32 	%f93, %f91, %f92;
	lg2.approx.ftz.f32 	%f94, %f93;
	mul.ftz.f32 	%f95, %f28, %f94;
	ex2.approx.ftz.f32 	%f96, %f95;
	fma.rn.ftz.f32 	%f53, %f88, %f96, %f30;
	bra.uni 	$Lt_21_12290;
$Lt_21_10498:
	sub.ftz.f32 	%f97, %f10, %f12;
	mov.f32 	%f98, 0f00000000;    	// 0
	max.ftz.f32 	%f99, %f97, %f98;
	set.ge.ftz.u32.f32 	%r67, %f53, %f10;
	neg.s32 	%r68, %r67;
	mov.f32 	%f100, 0f00000000;   	// 0
	set.eq.ftz.u32.f32 	%r69, %f12, %f100;
	neg.s32 	%r70, %r69;
	set.le.ftz.u32.f32 	%r71, %f99, %f53;
	neg.s32 	%r72, %r71;
	and.b32 	%r73, %r70, %r72;
	or.b32 	%r74, %r68, %r73;
	mov.u32 	%r75, 0;
	setp.eq.s32 	%p10, %r74, %r75;
	@%p10 bra 	$Lt_21_11010;
	.loc	22	92	0
	sub.ftz.f32 	%f101, %f52, %f50;
	sub.ftz.f32 	%f102, %f53, %f44;
	sub.ftz.f32 	%f103, %f46, %f44;
	div.approx.ftz.f32 	%f104, %f102, %f103;
	mov.f32 	%f105, 0f00000000;   	// 0
	max.ftz.f32 	%f106, %f104, %f105;
	lg2.approx.ftz.f32 	%f107, %f106;
	mul.ftz.f32 	%f108, %f48, %f107;
	ex2.approx.ftz.f32 	%f109, %f108;
	fma.rn.ftz.f32 	%f53, %f101, %f109, %f50;
	bra.uni 	$Lt_21_12290;
$Lt_21_11010:
	sub.ftz.f32 	%f110, %f42, %f40;
	sub.ftz.f32 	%f111, %f53, %f34;
	sub.ftz.f32 	%f112, %f36, %f34;
	set.gt.ftz.u32.f32 	%r76, %f99, %f53;
	neg.s32 	%r77, %r76;
	set.lt.ftz.u32.f32 	%r78, %f86, %f53;
	neg.s32 	%r79, %r78;
	and.b32 	%r80, %r77, %r79;
	mov.u32 	%r81, 0;
	setp.eq.s32 	%p11, %r80, %r81;
	@%p11 bra 	$Lt_21_11522;
	.loc	22	96	0
	div.approx.ftz.f32 	%f113, %f111, %f112;
	mov.f32 	%f114, 0f00000000;   	// 0
	max.ftz.f32 	%f115, %f113, %f114;
	lg2.approx.ftz.f32 	%f116, %f115;
	mul.ftz.f32 	%f117, %f38, %f116;
	ex2.approx.ftz.f32 	%f118, %f117;
	fma.rn.ftz.f32 	%f53, %f110, %f118, %f40;
	bra.uni 	$Lt_21_12290;
$Lt_21_11522:
	set.gt.ftz.u32.f32 	%r82, %f86, %f53;
	neg.s32 	%r83, %r82;
	set.lt.ftz.u32.f32 	%r84, %f6, %f53;
	neg.s32 	%r85, %r84;
	and.b32 	%r86, %r83, %r85;
	mov.u32 	%r87, 0;
	setp.eq.s32 	%p12, %r86, %r87;
	@%p12 bra 	$Lt_21_12034;
	.loc	23	79	0
	sub.ftz.f32 	%f119, %f32, %f30;
	sub.ftz.f32 	%f120, %f53, %f24;
	sub.ftz.f32 	%f121, %f26, %f24;
	div.approx.ftz.f32 	%f122, %f120, %f121;
	mov.f32 	%f123, 0f00000000;   	// 0
	max.ftz.f32 	%f124, %f122, %f123;
	lg2.approx.ftz.f32 	%f125, %f124;
	mul.ftz.f32 	%f126, %f125, %f28;
	ex2.approx.ftz.f32 	%f127, %f126;
	fma.rn.ftz.f32 	%f128, %f119, %f127, %f30;
	sub.ftz.f32 	%f129, %f53, %f6;
	sub.ftz.f32 	%f130, %f86, %f6;
	div.approx.ftz.f32 	%f131, %f129, %f130;
	div.approx.ftz.f32 	%f132, %f111, %f112;
	mov.f32 	%f133, 0f00000000;   	// 0
	max.ftz.f32 	%f134, %f132, %f133;
	lg2.approx.ftz.f32 	%f135, %f134;
	mul.ftz.f32 	%f136, %f38, %f135;
	ex2.approx.ftz.f32 	%f137, %f136;
	fma.rn.ftz.f32 	%f138, %f110, %f137, %f40;
	sub.ftz.f32 	%f139, %f138, %f128;
	fma.rn.ftz.f32 	%f140, %f131, %f139, %f128;
	.loc	22	102	0
	mov.f32 	%f53, %f140;
	bra.uni 	$Lt_21_12290;
$Lt_21_12034:
	setp.lt.ftz.f32 	%p13, %f99, %f53;
	@!%p13 bra 	$Lt_21_12546;
	.loc	23	79	0
	div.approx.ftz.f32 	%f141, %f111, %f112;
	mov.f32 	%f142, 0f00000000;   	// 0
	max.ftz.f32 	%f143, %f141, %f142;
	lg2.approx.ftz.f32 	%f144, %f143;
	mul.ftz.f32 	%f145, %f144, %f38;
	ex2.approx.ftz.f32 	%f146, %f145;
	fma.rn.ftz.f32 	%f147, %f110, %f146, %f40;
	sub.ftz.f32 	%f148, %f53, %f99;
	sub.ftz.f32 	%f149, %f10, %f99;
	div.approx.ftz.f32 	%f150, %f148, %f149;
	sub.ftz.f32 	%f151, %f52, %f50;
	sub.ftz.f32 	%f152, %f53, %f44;
	sub.ftz.f32 	%f153, %f46, %f44;
	div.approx.ftz.f32 	%f154, %f152, %f153;
	mov.f32 	%f155, 0f00000000;   	// 0
	max.ftz.f32 	%f156, %f154, %f155;
	lg2.approx.ftz.f32 	%f157, %f156;
	mul.ftz.f32 	%f158, %f48, %f157;
	ex2.approx.ftz.f32 	%f159, %f158;
	fma.rn.ftz.f32 	%f160, %f151, %f159, %f50;
	sub.ftz.f32 	%f161, %f160, %f147;
	fma.rn.ftz.f32 	%f162, %f150, %f161, %f147;
	.loc	22	108	0
	mov.f32 	%f53, %f162;
	bra.uni 	$Lt_21_12290;
$Lt_21_12546:
	.loc	22	112	0
	div.approx.ftz.f32 	%f163, %f111, %f112;
	mov.f32 	%f164, 0f00000000;   	// 0
	max.ftz.f32 	%f165, %f163, %f164;
	lg2.approx.ftz.f32 	%f166, %f165;
	mul.ftz.f32 	%f167, %f38, %f166;
	ex2.approx.ftz.f32 	%f168, %f167;
	fma.rn.ftz.f32 	%f53, %f110, %f168, %f40;
$Lt_21_12290:
$Lt_21_11778:
$Lt_21_11266:
$Lt_21_10754:
$Lt_21_10242:
	.loc	22	121	0
	sub.ftz.f32 	%f169, %f53, %f83;
	mov.f32 	%f170, 0f3f000000;   	// 0.5
	fma.rn.ftz.f32 	%f171, %f169, %f170, %f83;
	sub.ftz.f32 	%f172, %f171, %f2;
	fma.rn.ftz.f32 	%f173, %f4, %f172, %f2;
	st.param.f32 	[__cudaretf__Z21CalculateThreeWayLumaff50ThreeWayColorCorrectorParameters_BalanceTonalRange39ThreeWayColorCorrectorParameters_Levels], %f173;
	ret;
$LDWend__Z21CalculateThreeWayLumaff50ThreeWayColorCorrectorParameters_BalanceTonalRange39ThreeWayColorCorrectorParameters_Levels:
	} // _Z21CalculateThreeWayLumaff50ThreeWayColorCorrectorParameters_BalanceTonalRange39ThreeWayColorCorrectorParameters_Levels

	.visible .func (.param .f32 __cudaretf__Z4LERPIfET_S0_S0_S0_) _Z4LERPIfET_S0_S0_S0_ (.param .f32 __cudaparmf1__Z4LERPIfET_S0_S0_S0_, .param .f32 __cudaparmf2__Z4LERPIfET_S0_S0_S0_, .param .f32 __cudaparmf3__Z4LERPIfET_S0_S0_S0_)
	{
	.reg .f32 %f<10>;
	.loc	23	78	0
$LDWbegin__Z4LERPIfET_S0_S0_S0_:
	ld.param.f32 	%f1, [__cudaparmf1__Z4LERPIfET_S0_S0_S0_];
	mov.f32 	%f2, %f1;
	ld.param.f32 	%f3, [__cudaparmf2__Z4LERPIfET_S0_S0_S0_];
	mov.f32 	%f4, %f3;
	ld.param.f32 	%f5, [__cudaparmf3__Z4LERPIfET_S0_S0_S0_];
	mov.f32 	%f6, %f5;
	.loc	23	79	0
	sub.ftz.f32 	%f7, %f4, %f2;
	fma.rn.ftz.f32 	%f8, %f6, %f7, %f2;
	st.param.f32 	[__cudaretf__Z4LERPIfET_S0_S0_S0_], %f8;
	ret;
$LDWend__Z4LERPIfET_S0_S0_S0_:
	} // _Z4LERPIfET_S0_S0_S0_

	.visible .func (.param .f32 __cudaretf__Z6Read2DIfET_PKS0_iii) _Z6Read2DIfET_PKS0_iii (.param .u64 __cudaparmf1__Z6Read2DIfET_PKS0_iii, .param .s32 __cudaparmf2__Z6Read2DIfET_PKS0_iii, .param .s32 __cudaparmf3__Z6Read2DIfET_PKS0_iii, .param .s32 __cudaparmf4__Z6Read2DIfET_PKS0_iii)
	{
	.reg .u32 %r<10>;
	.reg .u64 %rd<7>;
	.reg .f32 %f<3>;
	.loc	20	114	0
$LDWbegin__Z6Read2DIfET_PKS0_iii:
	ld.param.u64 	%rd1, [__cudaparmf1__Z6Read2DIfET_PKS0_iii];
	mov.s64 	%rd2, %rd1;
	ld.param.u32 	%r1, [__cudaparmf2__Z6Read2DIfET_PKS0_iii];
	mov.s32 	%r2, %r1;
	ld.param.u32 	%r3, [__cudaparmf3__Z6Read2DIfET_PKS0_iii];
	mov.s32 	%r4, %r3;
	ld.param.u32 	%r5, [__cudaparmf4__Z6Read2DIfET_PKS0_iii];
	mov.s32 	%r6, %r5;
	.loc	20	115	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;
	ld.f32 	%f1, [%rd5+0];
	st.param.f32 	[__cudaretf__Z6Read2DIfET_PKS0_iii], %f1;
	ret;
$LDWend__Z6Read2DIfET_PKS0_iii:
	} // _Z6Read2DIfET_PKS0_iii
	.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};

	.entry ThreeWayColorCorrector_MaskKernel (
		.param .u64 __cudaparm_ThreeWayColorCorrector_MaskKernel_inImage,
		.param .s32 __cudaparm_ThreeWayColorCorrector_MaskKernel_inPitch,
		.param .u64 __cudaparm_ThreeWayColorCorrector_MaskKernel_inSecondaryMask,
		.param .s32 __cudaparm_ThreeWayColorCorrector_MaskKernel_inSecondaryPitch,
		.param .u32 __cudaparm_ThreeWayColorCorrector_MaskKernel_inDeviceFormat,
		.param .s32 __cudaparm_ThreeWayColorCorrector_MaskKernel_inWidth,
		.param .s32 __cudaparm_ThreeWayColorCorrector_MaskKernel_inHeight)
	{
	.reg .u32 %r<33>;
	.reg .u64 %rd<17>;
	.reg .f32 %f<10>;
	.reg .pred %p<5>;
	.loc	22	135	0
$LDWbegin_ThreeWayColorCorrector_MaskKernel:
	.loc	22	138	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_ThreeWayColorCorrector_MaskKernel_inWidth];
	set.gt.u32.s32 	%r12, %r11, %r8;
	neg.s32 	%r13, %r12;
	ld.param.s32 	%r14, [__cudaparm_ThreeWayColorCorrector_MaskKernel_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_24_3586;
	ld.param.s32 	%r19, [__cudaparm_ThreeWayColorCorrector_MaskKernel_inDeviceFormat];
	mov.s32 	%r20, 0;
	setp.eq.s32 	%p2, %r19, %r20;
	ld.param.s32 	%r21, [__cudaparm_ThreeWayColorCorrector_MaskKernel_inPitch];
	mul.lo.s32 	%r22, %r21, %r10;
	add.s32 	%r23, %r8, %r22;
	cvt.s64.s32 	%rd1, %r23;
	ld.param.u64 	%rd2, [__cudaparm_ThreeWayColorCorrector_MaskKernel_inImage];
	@!%p2 bra 	$Lt_24_4354;
	.loc	20	115	0
	mul.lo.u64 	%rd3, %rd1, 8;
	add.u64 	%rd4, %rd2, %rd3;
	ld.global.u16 	%r24, [%rd4+6];
	.loc	22	141	0
	{ .reg .b32 %b1;
	mov.b32		%b1, %r24;
	cvt.ftz.f32.f16	%f1, %b1; }
	bra.uni 	$Lt_24_4098;
$Lt_24_4354:
	mul.lo.u64 	%rd5, %rd1, 16;
	add.u64 	%rd6, %rd2, %rd5;
	ld.global.f32 	%f1, [%rd6+12];
$Lt_24_4098:
	ld.param.u64 	%rd7, [__cudaparm_ThreeWayColorCorrector_MaskKernel_inSecondaryMask];
	mov.u64 	%rd8, 0;
	setp.eq.u64 	%p3, %rd7, %rd8;
	@%p3 bra 	$Lt_24_4866;
	.loc	22	143	0
	ld.param.s32 	%r25, [__cudaparm_ThreeWayColorCorrector_MaskKernel_inSecondaryPitch];
	mul.lo.s32 	%r26, %r25, %r10;
	add.s32 	%r27, %r8, %r26;
	cvt.s64.s32 	%rd9, %r27;
	mul.wide.s32 	%rd10, %r27, 4;
	add.u64 	%rd11, %rd7, %rd10;
	ld.global.f32 	%f2, [%rd11+0];
	bra.uni 	$Lt_24_4610;
$Lt_24_4866:
	mov.f32 	%f2, 0f3f800000;     	// 1
$Lt_24_4610:
	.loc	22	147	0
	ld.const.f32 	%f3, [k601YPbPr_To_RGB32f+24];
	mul.ftz.f32 	%f4, %f3, %f2;
	ld.const.f32 	%f5, [k601YPbPr_To_RGB32f+12];
	mul.ftz.f32 	%f6, %f5, %f2;
	ld.const.f32 	%f7, [k601YPbPr_To_RGB32f+0];
	mul.ftz.f32 	%f8, %f7, %f2;
	@!%p2 bra 	$Lt_24_5378;
	.loc	20	126	0
	mul.lo.u64 	%rd12, %rd1, 8;
	add.u64 	%rd13, %rd2, %rd12;
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f4;
	mov.b32		%r28, %b1; }
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f6;
	mov.b32		%r29, %b1; }
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f8;
	mov.b32		%r30, %b1; }
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f1;
	mov.b32		%r31, %b1; }
	st.global.v4.u16 	[%rd13+0], {%r28,%r29,%r30,%r31};
	.loc	22	147	0
	bra.uni 	$Lt_24_5122;
$Lt_24_5378:
	.loc	20	126	0
	mul.lo.u64 	%rd14, %rd1, 16;
	add.u64 	%rd15, %rd2, %rd14;
	st.global.v4.f32 	[%rd15+0], {%f4,%f6,%f8,%f1};
$Lt_24_5122:
$Lt_24_3586:
	.loc	22	149	0
	exit;
$LDWend_ThreeWayColorCorrector_MaskKernel:
	} // ThreeWayColorCorrector_MaskKernel
	.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 ThreeWayColorCorrector_TonalRangeKernel (
		.param .u64 __cudaparm_ThreeWayColorCorrector_TonalRangeKernel_inImage,
		.param .s32 __cudaparm_ThreeWayColorCorrector_TonalRangeKernel_inPitch,
		.param .u64 __cudaparm_ThreeWayColorCorrector_TonalRangeKernel_inSecondaryMask,
		.param .s32 __cudaparm_ThreeWayColorCorrector_TonalRangeKernel_inSecondaryPitch,
		.param .u32 __cudaparm_ThreeWayColorCorrector_TonalRangeKernel_inDeviceFormat,
		.param .s32 __cudaparm_ThreeWayColorCorrector_TonalRangeKernel_inWidth,
		.param .s32 __cudaparm_ThreeWayColorCorrector_TonalRangeKernel_inHeight,
		.param .align 4 .b8 __cudaparm_ThreeWayColorCorrector_TonalRangeKernel_inBalanceTonalRangeParameters[92])
	{
	.reg .u32 %r<63>;
	.reg .u64 %rd<12>;
	.reg .f32 %f<40>;
	.reg .pred %p<9>;
	.loc	22	162	0
$LDWbegin_ThreeWayColorCorrector_TonalRangeKernel:
	.loc	22	165	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_ThreeWayColorCorrector_TonalRangeKernel_inWidth];
	set.gt.u32.s32 	%r12, %r11, %r8;
	neg.s32 	%r13, %r12;
	ld.param.s32 	%r14, [__cudaparm_ThreeWayColorCorrector_TonalRangeKernel_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_25_8194;
	ld.param.s32 	%r19, [__cudaparm_ThreeWayColorCorrector_TonalRangeKernel_inDeviceFormat];
	mov.s32 	%r20, 0;
	setp.eq.s32 	%p2, %r19, %r20;
	ld.param.s32 	%r21, [__cudaparm_ThreeWayColorCorrector_TonalRangeKernel_inPitch];
	mul.lo.s32 	%r22, %r21, %r10;
	add.s32 	%r23, %r8, %r22;
	cvt.s64.s32 	%rd1, %r23;
	ld.param.u64 	%rd2, [__cudaparm_ThreeWayColorCorrector_TonalRangeKernel_inImage];
	@!%p2 bra 	$Lt_25_8962;
	.loc	20	115	0
	mul.lo.u64 	%rd3, %rd1, 8;
	add.u64 	%rd4, %rd2, %rd3;
	ld.global.v4.u16 	{%r24,%r25,%r26,%r27}, [%rd4+0];
	.loc	22	168	0
	{ .reg .b32 %b1;
	mov.b32		%b1, %r24;
	cvt.ftz.f32.f16	%f1, %b1; }
	{ .reg .b32 %b1;
	mov.b32		%b1, %r25;
	cvt.ftz.f32.f16	%f2, %b1; }
	{ .reg .b32 %b1;
	mov.b32		%b1, %r26;
	cvt.ftz.f32.f16	%f3, %b1; }
	{ .reg .b32 %b1;
	mov.b32		%b1, %r27;
	cvt.ftz.f32.f16	%f4, %b1; }
	bra.uni 	$Lt_25_8706;
$Lt_25_8962:
	mul.lo.u64 	%rd5, %rd1, 16;
	add.u64 	%rd6, %rd2, %rd5;
	ld.global.v4.f32 	{%f1,%f2,%f3,%f4}, [%rd6+0];
$Lt_25_8706:
	ld.const.f32 	%f5, [kRGB32f_To_601YPbPr+4];
	mul.ftz.f32 	%f6, %f5, %f2;
	ld.param.f32 	%f7, [__cudaparm_ThreeWayColorCorrector_TonalRangeKernel_inBalanceTonalRangeParameters+4];
	ld.param.f32 	%f8, [__cudaparm_ThreeWayColorCorrector_TonalRangeKernel_inBalanceTonalRangeParameters+0];
	add.ftz.f32 	%f9, %f7, %f8;
	ld.const.f32 	%f10, [kRGB32f_To_601YPbPr+0];
	fma.rn.ftz.f32 	%f11, %f10, %f3, %f6;
	mov.f32 	%f12, 0f3f800000;    	// 1
	min.ftz.f32 	%f13, %f9, %f12;
	ld.const.f32 	%f14, [kRGB32f_To_601YPbPr+8];
	fma.rn.ftz.f32 	%f15, %f14, %f1, %f11;
	set.le.ftz.u32.f32 	%r28, %f15, %f8;
	neg.s32 	%r29, %r28;
	mov.f32 	%f16, 0f00000000;    	// 0
	set.eq.ftz.u32.f32 	%r30, %f7, %f16;
	neg.s32 	%r31, %r30;
	set.ge.ftz.u32.f32 	%r32, %f13, %f15;
	neg.s32 	%r33, %r32;
	and.b32 	%r34, %r31, %r33;
	or.b32 	%r35, %r29, %r34;
	mov.u32 	%r36, 0;
	setp.eq.s32 	%p3, %r35, %r36;
	@%p3 bra 	$Lt_25_9474;
	mov.f32 	%f17, 0f00000000;    	// 0
	bra.uni 	$Lt_25_9218;
$Lt_25_9474:
	ld.param.f32 	%f18, [__cudaparm_ThreeWayColorCorrector_TonalRangeKernel_inBalanceTonalRangeParameters+12];
	ld.param.f32 	%f19, [__cudaparm_ThreeWayColorCorrector_TonalRangeKernel_inBalanceTonalRangeParameters+8];
	sub.ftz.f32 	%f20, %f19, %f18;
	mov.f32 	%f21, 0f00000000;    	// 0
	max.ftz.f32 	%f22, %f20, %f21;
	set.ge.ftz.u32.f32 	%r37, %f15, %f19;
	neg.s32 	%r38, %r37;
	mov.f32 	%f23, 0f00000000;    	// 0
	set.eq.ftz.u32.f32 	%r39, %f18, %f23;
	neg.s32 	%r40, %r39;
	set.le.ftz.u32.f32 	%r41, %f22, %f15;
	neg.s32 	%r42, %r41;
	and.b32 	%r43, %r40, %r42;
	or.b32 	%r44, %r38, %r43;
	mov.u32 	%r45, 0;
	setp.eq.s32 	%p4, %r44, %r45;
	@%p4 bra 	$Lt_25_9986;
	mov.f32 	%f17, 0f3f800000;    	// 1
	bra.uni 	$Lt_25_9730;
$Lt_25_9986:
	set.gt.ftz.u32.f32 	%r46, %f22, %f15;
	neg.s32 	%r47, %r46;
	set.lt.ftz.u32.f32 	%r48, %f13, %f15;
	neg.s32 	%r49, %r48;
	and.b32 	%r50, %r47, %r49;
	mov.u32 	%r51, 0;
	setp.eq.s32 	%p5, %r50, %r51;
	@%p5 bra 	$Lt_25_10498;
	mov.f32 	%f17, 0f3f000000;    	// 0.5
	bra.uni 	$Lt_25_10242;
$Lt_25_10498:
	set.gt.ftz.u32.f32 	%r52, %f13, %f15;
	neg.s32 	%r53, %r52;
	set.gt.ftz.u32.f32 	%r54, %f15, %f8;
	neg.s32 	%r55, %r54;
	and.b32 	%r56, %r53, %r55;
	mov.u32 	%r57, 0;
	setp.eq.s32 	%p6, %r56, %r57;
	@%p6 bra 	$Lt_25_11010;
	.loc	22	191	0
	sub.ftz.f32 	%f24, %f15, %f8;
	sub.ftz.f32 	%f25, %f13, %f8;
	div.approx.ftz.f32 	%f26, %f24, %f25;
	mov.f32 	%f27, 0f3f000000;    	// 0.5
	mul.ftz.f32 	%f17, %f26, %f27;
	bra.uni 	$Lt_25_10754;
$Lt_25_11010:
	setp.lt.ftz.f32 	%p7, %f22, %f15;
	@!%p7 bra 	$Lt_25_11522;
	.loc	22	197	0
	mov.f32 	%f28, 0f3f000000;    	// 0.5
	sub.ftz.f32 	%f29, %f15, %f22;
	sub.ftz.f32 	%f30, %f19, %f22;
	div.approx.ftz.f32 	%f31, %f29, %f30;
	mov.f32 	%f32, 0f3f000000;    	// 0.5
	fma.rn.ftz.f32 	%f17, %f31, %f32, %f28;
	bra.uni 	$Lt_25_11266;
$Lt_25_11522:
	.loc	22	201	0
	mov.f32 	%f17, 0f3f000000;    	// 0.5
$Lt_25_11266:
$Lt_25_10754:
$Lt_25_10242:
$Lt_25_9730:
$Lt_25_9218:
	.loc	22	206	0
	ld.const.f32 	%f33, [k601YPbPr_To_RGB32f+24];
	mul.ftz.f32 	%f34, %f33, %f17;
	ld.const.f32 	%f35, [k601YPbPr_To_RGB32f+12];
	mul.ftz.f32 	%f36, %f35, %f17;
	ld.const.f32 	%f37, [k601YPbPr_To_RGB32f+0];
	mul.ftz.f32 	%f38, %f37, %f17;
	@!%p2 bra 	$Lt_25_12034;
	.loc	20	126	0
	mul.lo.u64 	%rd7, %rd1, 8;
	add.u64 	%rd8, %rd2, %rd7;
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f34;
	mov.b32		%r58, %b1; }
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f36;
	mov.b32		%r59, %b1; }
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f38;
	mov.b32		%r60, %b1; }
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f4;
	mov.b32		%r61, %b1; }
	st.global.v4.u16 	[%rd8+0], {%r58,%r59,%r60,%r61};
	.loc	22	206	0
	bra.uni 	$Lt_25_11778;
$Lt_25_12034:
	.loc	20	126	0
	mul.lo.u64 	%rd9, %rd1, 16;
	add.u64 	%rd10, %rd2, %rd9;
	st.global.v4.f32 	[%rd10+0], {%f34,%f36,%f38,%f4};
$Lt_25_11778:
$Lt_25_8194:
	.loc	22	208	0
	exit;
$LDWend_ThreeWayColorCorrector_TonalRangeKernel:
	} // ThreeWayColorCorrector_TonalRangeKernel

	.entry ThreeWayColorCorrector_LumaKernel (
		.param .u64 __cudaparm_ThreeWayColorCorrector_LumaKernel_inImage,
		.param .s32 __cudaparm_ThreeWayColorCorrector_LumaKernel_inPitch,
		.param .u64 __cudaparm_ThreeWayColorCorrector_LumaKernel_inSecondaryMask,
		.param .s32 __cudaparm_ThreeWayColorCorrector_LumaKernel_inSecondaryPitch,
		.param .u32 __cudaparm_ThreeWayColorCorrector_LumaKernel_inDeviceFormat,
		.param .s32 __cudaparm_ThreeWayColorCorrector_LumaKernel_inWidth,
		.param .s32 __cudaparm_ThreeWayColorCorrector_LumaKernel_inHeight,
		.param .align 4 .b8 __cudaparm_ThreeWayColorCorrector_LumaKernel_inBalanceTonalRangeParameters[92],
		.param .align 4 .b8 __cudaparm_ThreeWayColorCorrector_LumaKernel_inLevelsParameters[80])
	{
	.reg .u32 %r<113>;
	.reg .u64 %rd<17>;
	.reg .f32 %f<186>;
	.reg .pred %p<18>;
	.loc	22	222	0
$LDWbegin_ThreeWayColorCorrector_LumaKernel:
	.loc	22	225	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_ThreeWayColorCorrector_LumaKernel_inWidth];
	set.gt.u32.s32 	%r12, %r11, %r8;
	neg.s32 	%r13, %r12;
	ld.param.s32 	%r14, [__cudaparm_ThreeWayColorCorrector_LumaKernel_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_26_17666;
	ld.param.s32 	%r19, [__cudaparm_ThreeWayColorCorrector_LumaKernel_inDeviceFormat];
	mov.s32 	%r20, 0;
	setp.eq.s32 	%p2, %r19, %r20;
	ld.param.s32 	%r21, [__cudaparm_ThreeWayColorCorrector_LumaKernel_inPitch];
	mul.lo.s32 	%r22, %r21, %r10;
	add.s32 	%r23, %r8, %r22;
	cvt.s64.s32 	%rd1, %r23;
	ld.param.u64 	%rd2, [__cudaparm_ThreeWayColorCorrector_LumaKernel_inImage];
	@!%p2 bra 	$Lt_26_14338;
	.loc	20	115	0
	mul.lo.u64 	%rd3, %rd1, 8;
	add.u64 	%rd4, %rd2, %rd3;
	ld.global.v4.u16 	{%r24,%r25,%r26,%r27}, [%rd4+0];
	.loc	22	228	0
	{ .reg .b32 %b1;
	mov.b32		%b1, %r24;
	cvt.ftz.f32.f16	%f1, %b1; }
	{ .reg .b32 %b1;
	mov.b32		%b1, %r25;
	cvt.ftz.f32.f16	%f2, %b1; }
	{ .reg .b32 %b1;
	mov.b32		%b1, %r26;
	cvt.ftz.f32.f16	%f3, %b1; }
	{ .reg .b32 %b1;
	mov.b32		%b1, %r27;
	cvt.ftz.f32.f16	%f4, %b1; }
	bra.uni 	$Lt_26_14082;
$Lt_26_14338:
	mul.lo.u64 	%rd5, %rd1, 16;
	add.u64 	%rd6, %rd2, %rd5;
	ld.global.v4.f32 	{%f1,%f2,%f3,%f4}, [%rd6+0];
$Lt_26_14082:
	ld.param.u64 	%rd7, [__cudaparm_ThreeWayColorCorrector_LumaKernel_inSecondaryMask];
	mov.u64 	%rd8, 0;
	setp.eq.u64 	%p3, %rd7, %rd8;
	@%p3 bra 	$Lt_26_14850;
	.loc	22	229	0
	ld.param.s32 	%r28, [__cudaparm_ThreeWayColorCorrector_LumaKernel_inSecondaryPitch];
	mul.lo.s32 	%r29, %r28, %r10;
	add.s32 	%r30, %r8, %r29;
	cvt.s64.s32 	%rd9, %r30;
	mul.wide.s32 	%rd10, %r30, 4;
	add.u64 	%rd11, %rd7, %rd10;
	ld.global.f32 	%f5, [%rd11+0];
	bra.uni 	$Lt_26_14594;
$Lt_26_14850:
	mov.f32 	%f5, 0f3f800000;     	// 1
$Lt_26_14594:
	.loc	22	53	0
	ld.const.f32 	%f6, [kRGB32f_To_601YPbPr+4];
	mul.ftz.f32 	%f7, %f6, %f2;
	ld.const.f32 	%f8, [kRGB32f_To_601YPbPr+0];
	fma.rn.ftz.f32 	%f9, %f8, %f3, %f7;
	ld.const.f32 	%f10, [kRGB32f_To_601YPbPr+8];
	fma.rn.ftz.f32 	%f11, %f10, %f1, %f9;
	mov.f32 	%f12, %f11;
	.loc	22	229	0
	ld.param.f32 	%f13, [__cudaparm_ThreeWayColorCorrector_LumaKernel_inLevelsParameters+0];
	ld.param.f32 	%f14, [__cudaparm_ThreeWayColorCorrector_LumaKernel_inLevelsParameters+4];
	sub.ftz.f32 	%f15, %f14, %f13;
	ld.param.f32 	%f16, [__cudaparm_ThreeWayColorCorrector_LumaKernel_inLevelsParameters+12];
	ld.param.f32 	%f17, [__cudaparm_ThreeWayColorCorrector_LumaKernel_inLevelsParameters+16];
	sub.ftz.f32 	%f18, %f17, %f16;
	mov.f32 	%f19, 0f3f800000;    	// 1
	setp.neu.ftz.f32 	%p4, %f15, %f19;
	@%p4 bra 	$Lt_26_19458;
	mov.f32 	%f20, 0f3f800000;    	// 1
	setp.neu.ftz.f32 	%p5, %f18, %f20;
	@%p5 bra 	$Lt_26_19458;
	ld.param.f32 	%f21, [__cudaparm_ThreeWayColorCorrector_LumaKernel_inLevelsParameters+24];
	ld.param.f32 	%f22, [__cudaparm_ThreeWayColorCorrector_LumaKernel_inLevelsParameters+20];
	sub.ftz.f32 	%f23, %f21, %f22;
	mov.f32 	%f24, 0f3f800000;    	// 1
	setp.neu.ftz.f32 	%p6, %f23, %f24;
	@%p6 bra 	$Lt_26_19458;
	ld.param.f32 	%f25, [__cudaparm_ThreeWayColorCorrector_LumaKernel_inLevelsParameters+36];
	ld.param.f32 	%f26, [__cudaparm_ThreeWayColorCorrector_LumaKernel_inLevelsParameters+32];
	sub.ftz.f32 	%f27, %f25, %f26;
	mov.f32 	%f28, 0f3f800000;    	// 1
	setp.neu.ftz.f32 	%p7, %f27, %f28;
	@%p7 bra 	$Lt_26_19458;
	ld.param.f32 	%f29, [__cudaparm_ThreeWayColorCorrector_LumaKernel_inLevelsParameters+44];
	ld.param.f32 	%f30, [__cudaparm_ThreeWayColorCorrector_LumaKernel_inLevelsParameters+40];
	sub.ftz.f32 	%f31, %f29, %f30;
	mov.f32 	%f32, 0f3f800000;    	// 1
	setp.neu.ftz.f32 	%p8, %f31, %f32;
	@%p8 bra 	$Lt_26_19458;
	ld.param.f32 	%f33, [__cudaparm_ThreeWayColorCorrector_LumaKernel_inLevelsParameters+56];
	ld.param.f32 	%f34, [__cudaparm_ThreeWayColorCorrector_LumaKernel_inLevelsParameters+52];
	sub.ftz.f32 	%f35, %f33, %f34;
	mov.f32 	%f36, 0f3f800000;    	// 1
	setp.neu.ftz.f32 	%p9, %f35, %f36;
	@%p9 bra 	$Lt_26_19458;
	ld.param.f32 	%f37, [__cudaparm_ThreeWayColorCorrector_LumaKernel_inLevelsParameters+64];
	ld.param.f32 	%f38, [__cudaparm_ThreeWayColorCorrector_LumaKernel_inLevelsParameters+60];
	sub.ftz.f32 	%f39, %f37, %f38;
	mov.f32 	%f40, 0f3f800000;    	// 1
	setp.neu.ftz.f32 	%p10, %f39, %f40;
	@%p10 bra 	$Lt_26_19458;
	ld.param.f32 	%f41, [__cudaparm_ThreeWayColorCorrector_LumaKernel_inLevelsParameters+76];
	ld.param.f32 	%f42, [__cudaparm_ThreeWayColorCorrector_LumaKernel_inLevelsParameters+72];
	sub.ftz.f32 	%f43, %f41, %f42;
	mov.f32 	%f44, 0f3f800000;    	// 1
	setp.neu.ftz.f32 	%p11, %f43, %f44;
	@!%p11 bra 	$L_26_11522;
$Lt_26_19458:
$L_26_11778:
	.loc	20	143	0
	mov.s32 	%r31, 1;
	sub.s32 	%r32, %r31, %r8;
	shr.u32 	%r33, %r10, 13;
	sub.u32 	%r34, %r8, %r10;
	sub.u32 	%r35, %r32, %r10;
	xor.b32 	%r36, %r33, %r35;
	shl.b32 	%r37, %r36, 8;
	sub.u32 	%r38, %r34, %r36;
	sub.u32 	%r39, %r10, %r36;
	xor.b32 	%r40, %r37, %r38;
	shr.u32 	%r41, %r40, 13;
	sub.u32 	%r42, %r39, %r40;
	sub.u32 	%r43, %r36, %r40;
	xor.b32 	%r44, %r41, %r42;
	shr.u32 	%r45, %r44, 12;
	sub.u32 	%r46, %r43, %r44;
	xor.b32 	%r47, %r45, %r46;
	sub.u32 	%r48, %r40, %r44;
	sub.u32 	%r49, %r48, %r47;
	shl.b32 	%r50, %r47, 16;
	xor.b32 	%r51, %r49, %r50;
	.loc	20	144	0
	sub.u32 	%r52, %r44, %r47;
	sub.u32 	%r53, %r52, %r51;
	shr.u32 	%r54, %r51, 5;
	xor.b32 	%r55, %r53, %r54;
	.loc	20	145	0
	sub.u32 	%r56, %r47, %r51;
	sub.u32 	%r57, %r56, %r55;
	shr.u32 	%r58, %r55, 3;
	xor.b32 	%r59, %r57, %r58;
	.loc	20	146	0
	sub.u32 	%r60, %r51, %r55;
	sub.u32 	%r61, %r60, %r59;
	shl.b32 	%r62, %r59, 10;
	xor.b32 	%r63, %r61, %r62;
	.loc	20	147	0
	sub.u32 	%r64, %r55, %r59;
	sub.u32 	%r65, %r64, %r63;
	shr.u32 	%r66, %r63, 15;
	xor.b32 	%r67, %r65, %r66;
	.loc	22	63	0
	mov.f32 	%f45, 0f3b270d73;    	// 0.00254902
	mul.lo.u32 	%r68, %r67, 1103515245;
	add.u32 	%r69, %r68, 12345;
	shr.u32 	%r70, %r69, 16;
	and.b32 	%r71, %r70, 255;
	shl.b32 	%r72, %r71, 7;
	mul.lo.u32 	%r73, %r67, -1029531031;
	sub.u32 	%r74, %r73, 740551042;
	shr.u32 	%r75, %r74, 16;
	and.b32 	%r76, %r75, 255;
	xor.b32 	%r77, %r72, %r76;
	cvt.rn.f32.s32 	%f46, %r77;
	mov.f32 	%f47, 0f467ffe00;    	// 16383.5
	div.approx.ftz.f32 	%f48, %f46, %f47;
	mov.f32 	%f49, 0fbf800000;    	// -1
	add.ftz.f32 	%f50, %f48, %f49;
	fma.rn.ftz.f32 	%f12, %f45, %f50, %f11;
$L_26_11522:
	.loc	22	67	0
	ld.param.f32 	%f51, [__cudaparm_ThreeWayColorCorrector_LumaKernel_inLevelsParameters+8];
	sub.ftz.f32 	%f52, %f12, %f13;
	div.approx.ftz.f32 	%f53, %f52, %f15;
	mov.f32 	%f54, 0f00000000;    	// 0
	max.ftz.f32 	%f55, %f53, %f54;
	lg2.approx.ftz.f32 	%f56, %f55;
	mul.ftz.f32 	%f57, %f51, %f56;
	ex2.approx.ftz.f32 	%f58, %f57;
	fma.rn.ftz.f32 	%f59, %f18, %f58, %f16;
	.loc	22	79	0
	ld.param.f32 	%f60, [__cudaparm_ThreeWayColorCorrector_LumaKernel_inBalanceTonalRangeParameters+4];
	ld.param.f32 	%f61, [__cudaparm_ThreeWayColorCorrector_LumaKernel_inBalanceTonalRangeParameters+0];
	add.ftz.f32 	%f62, %f60, %f61;
	mov.f32 	%f63, 0f3f800000;    	// 1
	min.ftz.f32 	%f64, %f62, %f63;
	set.le.ftz.u32.f32 	%r78, %f12, %f61;
	neg.s32 	%r79, %r78;
	mov.f32 	%f65, 0f00000000;    	// 0
	set.eq.ftz.u32.f32 	%r80, %f60, %f65;
	neg.s32 	%r81, %r80;
	set.ge.ftz.u32.f32 	%r82, %f64, %f12;
	neg.s32 	%r83, %r82;
	and.b32 	%r84, %r81, %r83;
	or.b32 	%r85, %r79, %r84;
	mov.u32 	%r86, 0;
	setp.eq.s32 	%p12, %r85, %r86;
	@%p12 bra 	$Lt_26_15362;
	.loc	22	87	0
	ld.param.f32 	%f66, [__cudaparm_ThreeWayColorCorrector_LumaKernel_inLevelsParameters+20];
	ld.param.f32 	%f67, [__cudaparm_ThreeWayColorCorrector_LumaKernel_inLevelsParameters+32];
	ld.param.f32 	%f68, [__cudaparm_ThreeWayColorCorrector_LumaKernel_inLevelsParameters+36];
	sub.ftz.f32 	%f69, %f68, %f67;
	ld.param.f32 	%f70, [__cudaparm_ThreeWayColorCorrector_LumaKernel_inLevelsParameters+28];
	sub.ftz.f32 	%f71, %f12, %f66;
	ld.param.f32 	%f72, [__cudaparm_ThreeWayColorCorrector_LumaKernel_inLevelsParameters+24];
	sub.ftz.f32 	%f73, %f72, %f66;
	div.approx.ftz.f32 	%f74, %f71, %f73;
	mov.f32 	%f75, 0f00000000;    	// 0
	max.ftz.f32 	%f76, %f74, %f75;
	lg2.approx.ftz.f32 	%f77, %f76;
	mul.ftz.f32 	%f78, %f70, %f77;
	ex2.approx.ftz.f32 	%f79, %f78;
	fma.rn.ftz.f32 	%f12, %f69, %f79, %f67;
	bra.uni 	$Lt_26_17154;
$Lt_26_15362:
	ld.param.f32 	%f80, [__cudaparm_ThreeWayColorCorrector_LumaKernel_inBalanceTonalRangeParameters+12];
	ld.param.f32 	%f81, [__cudaparm_ThreeWayColorCorrector_LumaKernel_inBalanceTonalRangeParameters+8];
	sub.ftz.f32 	%f82, %f81, %f80;
	mov.f32 	%f83, 0f00000000;    	// 0
	max.ftz.f32 	%f84, %f82, %f83;
	set.ge.ftz.u32.f32 	%r87, %f12, %f81;
	neg.s32 	%r88, %r87;
	mov.f32 	%f85, 0f00000000;    	// 0
	set.eq.ftz.u32.f32 	%r89, %f80, %f85;
	neg.s32 	%r90, %r89;
	set.le.ftz.u32.f32 	%r91, %f84, %f12;
	neg.s32 	%r92, %r91;
	and.b32 	%r93, %r90, %r92;
	or.b32 	%r94, %r88, %r93;
	mov.u32 	%r95, 0;
	setp.eq.s32 	%p13, %r94, %r95;
	@%p13 bra 	$Lt_26_15874;
	.loc	22	92	0
	ld.param.f32 	%f86, [__cudaparm_ThreeWayColorCorrector_LumaKernel_inLevelsParameters+60];
	ld.param.f32 	%f87, [__cudaparm_ThreeWayColorCorrector_LumaKernel_inLevelsParameters+72];
	ld.param.f32 	%f88, [__cudaparm_ThreeWayColorCorrector_LumaKernel_inLevelsParameters+76];
	sub.ftz.f32 	%f89, %f88, %f87;
	ld.param.f32 	%f90, [__cudaparm_ThreeWayColorCorrector_LumaKernel_inLevelsParameters+68];
	sub.ftz.f32 	%f91, %f12, %f86;
	ld.param.f32 	%f92, [__cudaparm_ThreeWayColorCorrector_LumaKernel_inLevelsParameters+64];
	sub.ftz.f32 	%f93, %f92, %f86;
	div.approx.ftz.f32 	%f94, %f91, %f93;
	mov.f32 	%f95, 0f00000000;    	// 0
	max.ftz.f32 	%f96, %f94, %f95;
	lg2.approx.ftz.f32 	%f97, %f96;
	mul.ftz.f32 	%f98, %f90, %f97;
	ex2.approx.ftz.f32 	%f99, %f98;
	fma.rn.ftz.f32 	%f12, %f89, %f99, %f87;
	bra.uni 	$Lt_26_17154;
$Lt_26_15874:
	ld.param.f32 	%f100, [__cudaparm_ThreeWayColorCorrector_LumaKernel_inLevelsParameters+52];
	ld.param.f32 	%f101, [__cudaparm_ThreeWayColorCorrector_LumaKernel_inLevelsParameters+56];
	sub.ftz.f32 	%f102, %f101, %f100;
	ld.param.f32 	%f103, [__cudaparm_ThreeWayColorCorrector_LumaKernel_inLevelsParameters+40];
	sub.ftz.f32 	%f104, %f12, %f103;
	ld.param.f32 	%f105, [__cudaparm_ThreeWayColorCorrector_LumaKernel_inLevelsParameters+44];
	sub.ftz.f32 	%f106, %f105, %f103;
	ld.param.f32 	%f107, [__cudaparm_ThreeWayColorCorrector_LumaKernel_inLevelsParameters+48];
	set.gt.ftz.u32.f32 	%r96, %f84, %f12;
	neg.s32 	%r97, %r96;
	set.lt.ftz.u32.f32 	%r98, %f64, %f12;
	neg.s32 	%r99, %r98;
	and.b32 	%r100, %r97, %r99;
	mov.u32 	%r101, 0;
	setp.eq.s32 	%p14, %r100, %r101;
	@%p14 bra 	$Lt_26_16386;
	.loc	22	96	0
	div.approx.ftz.f32 	%f108, %f104, %f106;
	mov.f32 	%f109, 0f00000000;   	// 0
	max.ftz.f32 	%f110, %f108, %f109;
	lg2.approx.ftz.f32 	%f111, %f110;
	mul.ftz.f32 	%f112, %f107, %f111;
	ex2.approx.ftz.f32 	%f113, %f112;
	fma.rn.ftz.f32 	%f12, %f102, %f113, %f100;
	bra.uni 	$Lt_26_17154;
$Lt_26_16386:
	set.gt.ftz.u32.f32 	%r102, %f64, %f12;
	neg.s32 	%r103, %r102;
	set.gt.ftz.u32.f32 	%r104, %f12, %f61;
	neg.s32 	%r105, %r104;
	and.b32 	%r106, %r103, %r105;
	mov.u32 	%r107, 0;
	setp.eq.s32 	%p15, %r106, %r107;
	@%p15 bra 	$Lt_26_16898;
	.loc	23	79	0
	ld.param.f32 	%f114, [__cudaparm_ThreeWayColorCorrector_LumaKernel_inLevelsParameters+32];
	ld.param.f32 	%f115, [__cudaparm_ThreeWayColorCorrector_LumaKernel_inLevelsParameters+36];
	sub.ftz.f32 	%f116, %f115, %f114;
	ld.param.f32 	%f117, [__cudaparm_ThreeWayColorCorrector_LumaKernel_inLevelsParameters+20];
	sub.ftz.f32 	%f118, %f12, %f117;
	ld.param.f32 	%f119, [__cudaparm_ThreeWayColorCorrector_LumaKernel_inLevelsParameters+24];
	sub.ftz.f32 	%f120, %f119, %f117;
	div.approx.ftz.f32 	%f121, %f118, %f120;
	mov.f32 	%f122, 0f00000000;   	// 0
	max.ftz.f32 	%f123, %f121, %f122;
	lg2.approx.ftz.f32 	%f124, %f123;
	ld.param.f32 	%f125, [__cudaparm_ThreeWayColorCorrector_LumaKernel_inLevelsParameters+28];
	mul.ftz.f32 	%f126, %f125, %f124;
	ex2.approx.ftz.f32 	%f127, %f126;
	fma.rn.ftz.f32 	%f128, %f116, %f127, %f114;
	sub.ftz.f32 	%f129, %f12, %f61;
	sub.ftz.f32 	%f130, %f64, %f61;
	div.approx.ftz.f32 	%f131, %f129, %f130;
	div.approx.ftz.f32 	%f132, %f104, %f106;
	mov.f32 	%f133, 0f00000000;   	// 0
	max.ftz.f32 	%f134, %f132, %f133;
	lg2.approx.ftz.f32 	%f135, %f134;
	mul.ftz.f32 	%f136, %f107, %f135;
	ex2.approx.ftz.f32 	%f137, %f136;
	fma.rn.ftz.f32 	%f138, %f102, %f137, %f100;
	sub.ftz.f32 	%f139, %f138, %f128;
	fma.rn.ftz.f32 	%f140, %f131, %f139, %f128;
	.loc	22	102	0
	mov.f32 	%f12, %f140;
	bra.uni 	$Lt_26_17154;
$Lt_26_16898:
	setp.lt.ftz.f32 	%p16, %f84, %f12;
	@!%p16 bra 	$Lt_26_17410;
	.loc	23	79	0
	div.approx.ftz.f32 	%f141, %f104, %f106;
	mov.f32 	%f142, 0f00000000;   	// 0
	max.ftz.f32 	%f143, %f141, %f142;
	lg2.approx.ftz.f32 	%f144, %f143;
	mul.ftz.f32 	%f145, %f144, %f107;
	ex2.approx.ftz.f32 	%f146, %f145;
	fma.rn.ftz.f32 	%f147, %f102, %f146, %f100;
	ld.param.f32 	%f148, [__cudaparm_ThreeWayColorCorrector_LumaKernel_inLevelsParameters+60];
	ld.param.f32 	%f149, [__cudaparm_ThreeWayColorCorrector_LumaKernel_inLevelsParameters+72];
	sub.ftz.f32 	%f150, %f12, %f84;
	sub.ftz.f32 	%f151, %f81, %f84;
	div.approx.ftz.f32 	%f152, %f150, %f151;
	ld.param.f32 	%f153, [__cudaparm_ThreeWayColorCorrector_LumaKernel_inLevelsParameters+76];
	sub.ftz.f32 	%f154, %f153, %f149;
	ld.param.f32 	%f155, [__cudaparm_ThreeWayColorCorrector_LumaKernel_inLevelsParameters+68];
	sub.ftz.f32 	%f156, %f12, %f148;
	ld.param.f32 	%f157, [__cudaparm_ThreeWayColorCorrector_LumaKernel_inLevelsParameters+64];
	sub.ftz.f32 	%f158, %f157, %f148;
	div.approx.ftz.f32 	%f159, %f156, %f158;
	mov.f32 	%f160, 0f00000000;   	// 0
	max.ftz.f32 	%f161, %f159, %f160;
	lg2.approx.ftz.f32 	%f162, %f161;
	mul.ftz.f32 	%f163, %f155, %f162;
	ex2.approx.ftz.f32 	%f164, %f163;
	fma.rn.ftz.f32 	%f165, %f154, %f164, %f149;
	sub.ftz.f32 	%f166, %f165, %f147;
	fma.rn.ftz.f32 	%f167, %f152, %f166, %f147;
	.loc	22	108	0
	mov.f32 	%f12, %f167;
	bra.uni 	$Lt_26_17154;
$Lt_26_17410:
	.loc	22	112	0
	div.approx.ftz.f32 	%f168, %f104, %f106;
	mov.f32 	%f169, 0f00000000;   	// 0
	max.ftz.f32 	%f170, %f168, %f169;
	lg2.approx.ftz.f32 	%f171, %f170;
	mul.ftz.f32 	%f172, %f107, %f171;
	ex2.approx.ftz.f32 	%f173, %f172;
	fma.rn.ftz.f32 	%f12, %f102, %f173, %f100;
$Lt_26_17154:
$Lt_26_16642:
$Lt_26_16130:
$Lt_26_15618:
$Lt_26_15106:
	.loc	22	238	0
	sub.ftz.f32 	%f174, %f12, %f59;
	mov.f32 	%f175, 0f3f000000;   	// 0.5
	fma.rn.ftz.f32 	%f176, %f174, %f175, %f59;
	sub.ftz.f32 	%f177, %f176, %f11;
	fma.rn.ftz.f32 	%f178, %f5, %f177, %f11;
	ld.const.f32 	%f179, [k601YPbPr_To_RGB32f+24];
	mul.ftz.f32 	%f180, %f179, %f178;
	ld.const.f32 	%f181, [k601YPbPr_To_RGB32f+12];
	mul.ftz.f32 	%f182, %f181, %f178;
	ld.const.f32 	%f183, [k601YPbPr_To_RGB32f+0];
	mul.ftz.f32 	%f184, %f183, %f178;
	@!%p2 bra 	$Lt_26_17922;
	.loc	20	126	0
	mul.lo.u64 	%rd12, %rd1, 8;
	add.u64 	%rd13, %rd2, %rd12;
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f180;
	mov.b32		%r108, %b1; }
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f182;
	mov.b32		%r109, %b1; }
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f184;
	mov.b32		%r110, %b1; }
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f4;
	mov.b32		%r111, %b1; }
	st.global.v4.u16 	[%rd13+0], {%r108,%r109,%r110,%r111};
	.loc	22	238	0
	bra.uni 	$Lt_26_17666;
$Lt_26_17922:
	.loc	20	126	0
	mul.lo.u64 	%rd14, %rd1, 16;
	add.u64 	%rd15, %rd2, %rd14;
	st.global.v4.f32 	[%rd15+0], {%f180,%f182,%f184,%f4};
$Lt_26_17666:
$Lt_26_13570:
	.loc	22	240	0
	exit;
$LDWend_ThreeWayColorCorrector_LumaKernel:
	} // ThreeWayColorCorrector_LumaKernel

	.entry ThreeWayColorCorrector_CompositeKernel (
		.param .u64 __cudaparm_ThreeWayColorCorrector_CompositeKernel_inImage,
		.param .s32 __cudaparm_ThreeWayColorCorrector_CompositeKernel_inPitch,
		.param .u64 __cudaparm_ThreeWayColorCorrector_CompositeKernel_inSecondaryMask,
		.param .s32 __cudaparm_ThreeWayColorCorrector_CompositeKernel_inSecondaryPitch,
		.param .u32 __cudaparm_ThreeWayColorCorrector_CompositeKernel_inDeviceFormat,
		.param .s32 __cudaparm_ThreeWayColorCorrector_CompositeKernel_inWidth,
		.param .s32 __cudaparm_ThreeWayColorCorrector_CompositeKernel_inHeight,
		.param .align 4 .b8 __cudaparm_ThreeWayColorCorrector_CompositeKernel_inBalanceTonalRangeParameters[92],
		.param .align 4 .b8 __cudaparm_ThreeWayColorCorrector_CompositeKernel_inLevelsParameters[80])
	{
	.reg .u32 %r<140>;
	.reg .u64 %rd<17>;
	.reg .f32 %f<367>;
	.reg .pred %p<24>;
	.loc	22	254	0
$LDWbegin_ThreeWayColorCorrector_CompositeKernel:
	.loc	22	257	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_ThreeWayColorCorrector_CompositeKernel_inWidth];
	set.gt.u32.s32 	%r12, %r11, %r8;
	neg.s32 	%r13, %r12;
	ld.param.s32 	%r14, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_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_27_25602;
	ld.param.s32 	%r19, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inDeviceFormat];
	mov.s32 	%r20, 0;
	setp.eq.s32 	%p2, %r19, %r20;
	ld.param.s32 	%r21, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inPitch];
	mul.lo.s32 	%r22, %r21, %r10;
	add.s32 	%r23, %r8, %r22;
	cvt.s64.s32 	%rd1, %r23;
	ld.param.u64 	%rd2, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inImage];
	@!%p2 bra 	$Lt_27_19714;
	.loc	20	115	0
	mul.lo.u64 	%rd3, %rd1, 8;
	add.u64 	%rd4, %rd2, %rd3;
	ld.global.v4.u16 	{%r24,%r25,%r26,%r27}, [%rd4+0];
	.loc	22	260	0
	{ .reg .b32 %b1;
	mov.b32		%b1, %r24;
	cvt.ftz.f32.f16	%f1, %b1; }
	{ .reg .b32 %b1;
	mov.b32		%b1, %r25;
	cvt.ftz.f32.f16	%f2, %b1; }
	{ .reg .b32 %b1;
	mov.b32		%b1, %r26;
	cvt.ftz.f32.f16	%f3, %b1; }
	{ .reg .b32 %b1;
	mov.b32		%b1, %r27;
	cvt.ftz.f32.f16	%f4, %b1; }
	bra.uni 	$Lt_27_19458;
$Lt_27_19714:
	mul.lo.u64 	%rd5, %rd1, 16;
	add.u64 	%rd6, %rd2, %rd5;
	ld.global.v4.f32 	{%f1,%f2,%f3,%f4}, [%rd6+0];
$Lt_27_19458:
	ld.param.u64 	%rd7, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inSecondaryMask];
	mov.u64 	%rd8, 0;
	setp.eq.u64 	%p3, %rd7, %rd8;
	@%p3 bra 	$Lt_27_20226;
	.loc	22	261	0
	ld.param.s32 	%r28, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inSecondaryPitch];
	mul.lo.s32 	%r29, %r28, %r10;
	add.s32 	%r30, %r8, %r29;
	cvt.s64.s32 	%rd9, %r30;
	mul.wide.s32 	%rd10, %r30, 4;
	add.u64 	%rd11, %rd7, %rd10;
	ld.global.f32 	%f5, [%rd11+0];
	bra.uni 	$Lt_27_19970;
$Lt_27_20226:
	mov.f32 	%f5, 0f3f800000;     	// 1
$Lt_27_19970:
	.loc	22	299	0
	ld.const.f32 	%f6, [kRGB32f_To_601YPbPr+4];
	mul.ftz.f32 	%f7, %f6, %f2;
	ld.param.f32 	%f8, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inBalanceTonalRangeParameters+4];
	mov.f32 	%f9, 0f00000000;     	// 0
	setp.eq.ftz.f32 	%p4, %f8, %f9;
	ld.param.f32 	%f10, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inBalanceTonalRangeParameters+0];
	add.ftz.f32 	%f11, %f10, %f8;
	ld.const.f32 	%f12, [kRGB32f_To_601YPbPr+28];
	mul.ftz.f32 	%f13, %f12, %f2;
	ld.const.f32 	%f14, [kRGB32f_To_601YPbPr+16];
	mul.ftz.f32 	%f15, %f14, %f2;
	ld.const.f32 	%f16, [kRGB32f_To_601YPbPr+0];
	fma.rn.ftz.f32 	%f17, %f16, %f3, %f7;
	selp.s32 	%r31, 1, 0, %p4;
	mov.f32 	%f18, 0f3f800000;    	// 1
	min.ftz.f32 	%f19, %f11, %f18;
	ld.const.f32 	%f20, [kRGB32f_To_601YPbPr+24];
	fma.rn.ftz.f32 	%f21, %f20, %f3, %f13;
	ld.const.f32 	%f22, [kRGB32f_To_601YPbPr+12];
	fma.rn.ftz.f32 	%f23, %f22, %f3, %f15;
	ld.const.f32 	%f24, [kRGB32f_To_601YPbPr+8];
	fma.rn.ftz.f32 	%f25, %f24, %f1, %f17;
	ld.const.f32 	%f26, [kRGB32f_To_601YPbPr+32];
	fma.rn.ftz.f32 	%f27, %f26, %f1, %f21;
	ld.const.f32 	%f28, [kRGB32f_To_601YPbPr+20];
	fma.rn.ftz.f32 	%f29, %f28, %f1, %f23;
	set.le.ftz.u32.f32 	%r32, %f25, %f10;
	neg.s32 	%r33, %r32;
	set.ge.ftz.u32.f32 	%r34, %f19, %f25;
	neg.s32 	%r35, %r34;
	and.b32 	%r36, %r31, %r35;
	or.b32 	%r37, %r33, %r36;
	mov.u32 	%r38, 0;
	setp.eq.s32 	%p5, %r37, %r38;
	@%p5 bra 	$Lt_27_20738;
	.loc	22	317	0
	ld.param.f32 	%f30, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inBalanceTonalRangeParameters+32];
	sin.approx.ftz.f32 	%f31, %f30;
	cos.approx.ftz.f32 	%f32, %f30;
	mul.ftz.f32 	%f33, %f27, %f31;
	mul.ftz.f32 	%f34, %f29, %f32;
	sub.ftz.f32 	%f35, %f34, %f33;
	ld.param.f32 	%f36, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inBalanceTonalRangeParameters+20];
	ld.param.f32 	%f37, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inBalanceTonalRangeParameters+64];
	ld.param.f32 	%f38, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inBalanceTonalRangeParameters+56];
	sub.ftz.f32 	%f39, %f38, %f35;
	fma.rn.ftz.f32 	%f40, %f37, %f39, %f35;
	mul.ftz.f32 	%f41, %f36, %f40;
	.loc	22	318	0
	mul.ftz.f32 	%f42, %f27, %f32;
	fma.rn.ftz.f32 	%f43, %f29, %f31, %f42;
	ld.param.f32 	%f44, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inBalanceTonalRangeParameters+60];
	sub.ftz.f32 	%f45, %f44, %f43;
	fma.rn.ftz.f32 	%f46, %f37, %f45, %f43;
	mul.ftz.f32 	%f47, %f36, %f46;
	bra.uni 	$Lt_27_22530;
$Lt_27_20738:
	ld.param.f32 	%f48, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inBalanceTonalRangeParameters+12];
	ld.param.f32 	%f49, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inBalanceTonalRangeParameters+8];
	sub.ftz.f32 	%f50, %f49, %f48;
	mov.f32 	%f51, 0f00000000;    	// 0
	max.ftz.f32 	%f52, %f50, %f51;
	set.ge.ftz.u32.f32 	%r39, %f25, %f49;
	neg.s32 	%r40, %r39;
	mov.f32 	%f53, 0f00000000;    	// 0
	set.eq.ftz.u32.f32 	%r41, %f48, %f53;
	neg.s32 	%r42, %r41;
	set.ge.ftz.u32.f32 	%r43, %f25, %f52;
	neg.s32 	%r44, %r43;
	and.b32 	%r45, %r42, %r44;
	or.b32 	%r46, %r40, %r45;
	mov.u32 	%r47, 0;
	setp.eq.s32 	%p6, %r46, %r47;
	@%p6 bra 	$Lt_27_21250;
	.loc	22	323	0
	ld.param.f32 	%f54, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inBalanceTonalRangeParameters+40];
	sin.approx.ftz.f32 	%f55, %f54;
	cos.approx.ftz.f32 	%f56, %f54;
	mul.ftz.f32 	%f57, %f27, %f55;
	mul.ftz.f32 	%f58, %f29, %f56;
	sub.ftz.f32 	%f59, %f58, %f57;
	ld.param.f32 	%f60, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inBalanceTonalRangeParameters+88];
	ld.param.f32 	%f61, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inBalanceTonalRangeParameters+28];
	ld.param.f32 	%f62, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inBalanceTonalRangeParameters+80];
	sub.ftz.f32 	%f63, %f62, %f59;
	fma.rn.ftz.f32 	%f64, %f60, %f63, %f59;
	mul.ftz.f32 	%f41, %f61, %f64;
	.loc	22	324	0
	mul.ftz.f32 	%f65, %f27, %f56;
	fma.rn.ftz.f32 	%f66, %f29, %f55, %f65;
	ld.param.f32 	%f67, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inBalanceTonalRangeParameters+84];
	sub.ftz.f32 	%f68, %f67, %f66;
	fma.rn.ftz.f32 	%f69, %f60, %f68, %f66;
	mul.ftz.f32 	%f47, %f61, %f69;
	bra.uni 	$Lt_27_22530;
$Lt_27_21250:
	ld.param.f32 	%f70, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inBalanceTonalRangeParameters+72];
	ld.param.f32 	%f71, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inBalanceTonalRangeParameters+68];
	ld.param.f32 	%f72, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inBalanceTonalRangeParameters+76];
	ld.param.f32 	%f73, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inBalanceTonalRangeParameters+24];
	ld.param.f32 	%f74, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inBalanceTonalRangeParameters+36];
	set.lt.ftz.u32.f32 	%r48, %f19, %f25;
	neg.s32 	%r49, %r48;
	set.lt.ftz.u32.f32 	%r50, %f25, %f52;
	neg.s32 	%r51, %r50;
	and.b32 	%r52, %r49, %r51;
	mov.u32 	%r53, 0;
	setp.eq.s32 	%p7, %r52, %r53;
	@%p7 bra 	$Lt_27_21762;
	.loc	22	328	0
	sin.approx.ftz.f32 	%f75, %f74;
	cos.approx.ftz.f32 	%f76, %f74;
	mul.ftz.f32 	%f77, %f27, %f75;
	mul.ftz.f32 	%f78, %f29, %f76;
	sub.ftz.f32 	%f79, %f78, %f77;
	sub.ftz.f32 	%f80, %f71, %f79;
	fma.rn.ftz.f32 	%f81, %f72, %f80, %f79;
	mul.ftz.f32 	%f41, %f73, %f81;
	.loc	22	329	0
	mul.ftz.f32 	%f82, %f27, %f76;
	fma.rn.ftz.f32 	%f83, %f29, %f75, %f82;
	sub.ftz.f32 	%f84, %f70, %f83;
	fma.rn.ftz.f32 	%f85, %f72, %f84, %f83;
	mul.ftz.f32 	%f47, %f73, %f85;
	bra.uni 	$Lt_27_22530;
$Lt_27_21762:
	set.gt.ftz.u32.f32 	%r54, %f19, %f25;
	neg.s32 	%r55, %r54;
	set.gt.ftz.u32.f32 	%r56, %f25, %f10;
	neg.s32 	%r57, %r56;
	and.b32 	%r58, %r55, %r57;
	mov.u32 	%r59, 0;
	setp.eq.s32 	%p8, %r58, %r59;
	@%p8 bra 	$Lt_27_22274;
	.loc	22	335	0
	sin.approx.ftz.f32 	%f86, %f74;
	cos.approx.ftz.f32 	%f87, %f74;
	ld.param.f32 	%f88, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inBalanceTonalRangeParameters+32];
	sin.approx.ftz.f32 	%f89, %f88;
	cos.approx.ftz.f32 	%f90, %f88;
	sub.ftz.f32 	%f91, %f19, %f10;
	sub.ftz.f32 	%f92, %f25, %f10;
	mul.ftz.f32 	%f93, %f27, %f86;
	mul.ftz.f32 	%f94, %f27, %f89;
	div.approx.ftz.f32 	%f95, %f92, %f91;
	mul.ftz.f32 	%f96, %f29, %f87;
	sub.ftz.f32 	%f97, %f96, %f93;
	mul.ftz.f32 	%f98, %f29, %f90;
	sub.ftz.f32 	%f99, %f98, %f94;
	ld.param.f32 	%f100, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inBalanceTonalRangeParameters+56];
	sub.ftz.f32 	%f101, %f100, %f99;
	ld.param.f32 	%f102, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inBalanceTonalRangeParameters+64];
	fma.rn.ftz.f32 	%f103, %f102, %f101, %f99;
	ld.param.f32 	%f104, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inBalanceTonalRangeParameters+20];
	mul.ftz.f32 	%f105, %f103, %f104;
	sub.ftz.f32 	%f106, %f71, %f97;
	fma.rn.ftz.f32 	%f107, %f72, %f106, %f97;
	mul.ftz.f32 	%f108, %f73, %f107;
	sub.ftz.f32 	%f109, %f108, %f105;
	fma.rn.ftz.f32 	%f41, %f95, %f109, %f105;
	.loc	22	336	0
	mul.ftz.f32 	%f110, %f27, %f87;
	mul.ftz.f32 	%f111, %f27, %f90;
	fma.rn.ftz.f32 	%f112, %f29, %f86, %f110;
	fma.rn.ftz.f32 	%f113, %f29, %f89, %f111;
	ld.param.f32 	%f114, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inBalanceTonalRangeParameters+60];
	sub.ftz.f32 	%f115, %f114, %f113;
	fma.rn.ftz.f32 	%f116, %f102, %f115, %f113;
	mul.ftz.f32 	%f117, %f116, %f104;
	sub.ftz.f32 	%f118, %f70, %f112;
	fma.rn.ftz.f32 	%f119, %f72, %f118, %f112;
	mul.ftz.f32 	%f120, %f73, %f119;
	sub.ftz.f32 	%f121, %f120, %f117;
	fma.rn.ftz.f32 	%f47, %f95, %f121, %f117;
	bra.uni 	$Lt_27_22530;
$Lt_27_22274:
	setp.gt.ftz.f32 	%p9, %f25, %f52;
	@!%p9 bra 	$Lt_27_22786;
	.loc	22	342	0
	sin.approx.ftz.f32 	%f122, %f74;
	cos.approx.ftz.f32 	%f123, %f74;
	ld.param.f32 	%f124, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inBalanceTonalRangeParameters+40];
	sin.approx.ftz.f32 	%f125, %f124;
	cos.approx.ftz.f32 	%f126, %f124;
	sub.ftz.f32 	%f127, %f49, %f52;
	sub.ftz.f32 	%f128, %f25, %f52;
	mul.ftz.f32 	%f129, %f27, %f122;
	mul.ftz.f32 	%f130, %f27, %f125;
	div.approx.ftz.f32 	%f131, %f128, %f127;
	mul.ftz.f32 	%f132, %f29, %f123;
	sub.ftz.f32 	%f133, %f132, %f129;
	mul.ftz.f32 	%f134, %f29, %f126;
	sub.ftz.f32 	%f135, %f134, %f130;
	sub.ftz.f32 	%f136, %f71, %f133;
	fma.rn.ftz.f32 	%f137, %f72, %f136, %f133;
	mul.ftz.f32 	%f138, %f137, %f73;
	ld.param.f32 	%f139, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inBalanceTonalRangeParameters+88];
	ld.param.f32 	%f140, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inBalanceTonalRangeParameters+28];
	ld.param.f32 	%f141, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inBalanceTonalRangeParameters+80];
	sub.ftz.f32 	%f142, %f141, %f135;
	fma.rn.ftz.f32 	%f143, %f139, %f142, %f135;
	mul.ftz.f32 	%f144, %f140, %f143;
	sub.ftz.f32 	%f145, %f144, %f138;
	fma.rn.ftz.f32 	%f41, %f131, %f145, %f138;
	.loc	22	343	0
	mul.ftz.f32 	%f146, %f27, %f123;
	mul.ftz.f32 	%f147, %f27, %f126;
	fma.rn.ftz.f32 	%f148, %f29, %f122, %f146;
	fma.rn.ftz.f32 	%f149, %f29, %f125, %f147;
	sub.ftz.f32 	%f150, %f70, %f148;
	fma.rn.ftz.f32 	%f151, %f72, %f150, %f148;
	mul.ftz.f32 	%f152, %f151, %f73;
	ld.param.f32 	%f153, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inBalanceTonalRangeParameters+84];
	sub.ftz.f32 	%f154, %f153, %f149;
	fma.rn.ftz.f32 	%f155, %f139, %f154, %f149;
	mul.ftz.f32 	%f156, %f140, %f155;
	sub.ftz.f32 	%f157, %f156, %f152;
	fma.rn.ftz.f32 	%f47, %f131, %f157, %f152;
	bra.uni 	$Lt_27_22530;
$Lt_27_22786:
	.loc	22	347	0
	sin.approx.ftz.f32 	%f158, %f74;
	cos.approx.ftz.f32 	%f159, %f74;
	mul.ftz.f32 	%f160, %f27, %f158;
	mul.ftz.f32 	%f161, %f29, %f159;
	sub.ftz.f32 	%f162, %f161, %f160;
	sub.ftz.f32 	%f163, %f71, %f162;
	fma.rn.ftz.f32 	%f164, %f72, %f163, %f162;
	mul.ftz.f32 	%f41, %f73, %f164;
	.loc	22	348	0
	mul.ftz.f32 	%f165, %f27, %f159;
	fma.rn.ftz.f32 	%f166, %f29, %f158, %f165;
	sub.ftz.f32 	%f167, %f70, %f166;
	fma.rn.ftz.f32 	%f168, %f72, %f167, %f166;
	mul.ftz.f32 	%f47, %f73, %f168;
$Lt_27_22530:
$Lt_27_22018:
$Lt_27_21506:
$Lt_27_20994:
$Lt_27_20482:
	.loc	22	53	0
	mov.f32 	%f169, %f25;
	.loc	22	352	0
	ld.param.f32 	%f170, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inLevelsParameters+0];
	ld.param.f32 	%f171, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inLevelsParameters+4];
	sub.ftz.f32 	%f172, %f171, %f170;
	ld.param.f32 	%f173, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inLevelsParameters+12];
	ld.param.f32 	%f174, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inLevelsParameters+16];
	sub.ftz.f32 	%f175, %f174, %f173;
	mov.f32 	%f176, 0f3f800000;   	// 1
	setp.neu.ftz.f32 	%p10, %f172, %f176;
	@%p10 bra 	$Lt_27_27394;
	mov.f32 	%f177, 0f3f800000;   	// 1
	setp.neu.ftz.f32 	%p11, %f175, %f177;
	@%p11 bra 	$Lt_27_27394;
	ld.param.f32 	%f178, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inLevelsParameters+24];
	ld.param.f32 	%f179, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inLevelsParameters+20];
	sub.ftz.f32 	%f180, %f178, %f179;
	mov.f32 	%f181, 0f3f800000;   	// 1
	setp.neu.ftz.f32 	%p12, %f180, %f181;
	@%p12 bra 	$Lt_27_27394;
	ld.param.f32 	%f182, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inLevelsParameters+36];
	ld.param.f32 	%f183, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inLevelsParameters+32];
	sub.ftz.f32 	%f184, %f182, %f183;
	mov.f32 	%f185, 0f3f800000;   	// 1
	setp.neu.ftz.f32 	%p13, %f184, %f185;
	@%p13 bra 	$Lt_27_27394;
	ld.param.f32 	%f186, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inLevelsParameters+44];
	ld.param.f32 	%f187, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inLevelsParameters+40];
	sub.ftz.f32 	%f188, %f186, %f187;
	mov.f32 	%f189, 0f3f800000;   	// 1
	setp.neu.ftz.f32 	%p14, %f188, %f189;
	@%p14 bra 	$Lt_27_27394;
	ld.param.f32 	%f190, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inLevelsParameters+56];
	ld.param.f32 	%f191, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inLevelsParameters+52];
	sub.ftz.f32 	%f192, %f190, %f191;
	mov.f32 	%f193, 0f3f800000;   	// 1
	setp.neu.ftz.f32 	%p15, %f192, %f193;
	@%p15 bra 	$Lt_27_27394;
	ld.param.f32 	%f194, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inLevelsParameters+64];
	ld.param.f32 	%f195, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inLevelsParameters+60];
	sub.ftz.f32 	%f196, %f194, %f195;
	mov.f32 	%f197, 0f3f800000;   	// 1
	setp.neu.ftz.f32 	%p16, %f196, %f197;
	@%p16 bra 	$Lt_27_27394;
	ld.param.f32 	%f198, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inLevelsParameters+76];
	ld.param.f32 	%f199, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inLevelsParameters+72];
	sub.ftz.f32 	%f200, %f198, %f199;
	mov.f32 	%f201, 0f3f800000;   	// 1
	setp.neu.ftz.f32 	%p17, %f200, %f201;
	@!%p17 bra 	$L_27_16898;
$Lt_27_27394:
$L_27_17154:
	.loc	20	143	0
	mov.s32 	%r60, 1;
	sub.s32 	%r61, %r60, %r8;
	shr.u32 	%r62, %r10, 13;
	sub.u32 	%r63, %r8, %r10;
	sub.u32 	%r64, %r61, %r10;
	xor.b32 	%r65, %r62, %r64;
	shl.b32 	%r66, %r65, 8;
	sub.u32 	%r67, %r63, %r65;
	sub.u32 	%r68, %r10, %r65;
	xor.b32 	%r69, %r66, %r67;
	shr.u32 	%r70, %r69, 13;
	sub.u32 	%r71, %r68, %r69;
	sub.u32 	%r72, %r65, %r69;
	xor.b32 	%r73, %r70, %r71;
	shr.u32 	%r74, %r73, 12;
	sub.u32 	%r75, %r72, %r73;
	xor.b32 	%r76, %r74, %r75;
	sub.u32 	%r77, %r69, %r73;
	sub.u32 	%r78, %r77, %r76;
	shl.b32 	%r79, %r76, 16;
	xor.b32 	%r80, %r78, %r79;
	.loc	20	144	0
	sub.u32 	%r81, %r73, %r76;
	sub.u32 	%r82, %r81, %r80;
	shr.u32 	%r83, %r80, 5;
	xor.b32 	%r84, %r82, %r83;
	.loc	20	145	0
	sub.u32 	%r85, %r76, %r80;
	sub.u32 	%r86, %r85, %r84;
	shr.u32 	%r87, %r84, 3;
	xor.b32 	%r88, %r86, %r87;
	.loc	20	146	0
	sub.u32 	%r89, %r80, %r84;
	sub.u32 	%r90, %r89, %r88;
	shl.b32 	%r91, %r88, 10;
	xor.b32 	%r92, %r90, %r91;
	.loc	20	147	0
	sub.u32 	%r93, %r84, %r88;
	sub.u32 	%r94, %r93, %r92;
	shr.u32 	%r95, %r92, 15;
	xor.b32 	%r96, %r94, %r95;
	.loc	22	63	0
	mov.f32 	%f202, 0f3b270d73;   	// 0.00254902
	mul.lo.u32 	%r97, %r96, 1103515245;
	add.u32 	%r98, %r97, 12345;
	shr.u32 	%r99, %r98, 16;
	and.b32 	%r100, %r99, 255;
	shl.b32 	%r101, %r100, 7;
	mul.lo.u32 	%r102, %r96, -1029531031;
	sub.u32 	%r103, %r102, 740551042;
	shr.u32 	%r104, %r103, 16;
	and.b32 	%r105, %r104, 255;
	xor.b32 	%r106, %r101, %r105;
	cvt.rn.f32.s32 	%f203, %r106;
	mov.f32 	%f204, 0f467ffe00;   	// 16383.5
	div.approx.ftz.f32 	%f205, %f203, %f204;
	mov.f32 	%f206, 0fbf800000;   	// -1
	add.ftz.f32 	%f207, %f205, %f206;
	fma.rn.ftz.f32 	%f169, %f202, %f207, %f25;
$L_27_16898:
	.loc	22	67	0
	ld.param.f32 	%f208, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inLevelsParameters+8];
	sub.ftz.f32 	%f209, %f169, %f170;
	div.approx.ftz.f32 	%f210, %f209, %f172;
	mov.f32 	%f211, 0f00000000;   	// 0
	max.ftz.f32 	%f212, %f210, %f211;
	lg2.approx.ftz.f32 	%f213, %f212;
	mul.ftz.f32 	%f214, %f208, %f213;
	ex2.approx.ftz.f32 	%f215, %f214;
	fma.rn.ftz.f32 	%f216, %f175, %f215, %f173;
	.loc	22	79	0
	set.le.ftz.u32.f32 	%r107, %f169, %f10;
	neg.s32 	%r108, %r107;
	set.ge.ftz.u32.f32 	%r109, %f19, %f169;
	neg.s32 	%r110, %r109;
	and.b32 	%r111, %r31, %r110;
	or.b32 	%r112, %r108, %r111;
	mov.u32 	%r113, 0;
	setp.eq.s32 	%p18, %r112, %r113;
	@%p18 bra 	$Lt_27_23298;
	.loc	22	87	0
	ld.param.f32 	%f217, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inLevelsParameters+20];
	ld.param.f32 	%f218, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inLevelsParameters+32];
	ld.param.f32 	%f219, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inLevelsParameters+36];
	sub.ftz.f32 	%f220, %f219, %f218;
	ld.param.f32 	%f221, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inLevelsParameters+28];
	sub.ftz.f32 	%f222, %f169, %f217;
	ld.param.f32 	%f223, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inLevelsParameters+24];
	sub.ftz.f32 	%f224, %f223, %f217;
	div.approx.ftz.f32 	%f225, %f222, %f224;
	mov.f32 	%f226, 0f00000000;   	// 0
	max.ftz.f32 	%f227, %f225, %f226;
	lg2.approx.ftz.f32 	%f228, %f227;
	mul.ftz.f32 	%f229, %f221, %f228;
	ex2.approx.ftz.f32 	%f230, %f229;
	fma.rn.ftz.f32 	%f169, %f220, %f230, %f218;
	bra.uni 	$Lt_27_25090;
$Lt_27_23298:
	ld.param.f32 	%f231, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inBalanceTonalRangeParameters+12];
	ld.param.f32 	%f49, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inBalanceTonalRangeParameters+8];
	sub.ftz.f32 	%f232, %f49, %f231;
	mov.f32 	%f233, 0f00000000;   	// 0
	max.ftz.f32 	%f52, %f232, %f233;
	set.ge.ftz.u32.f32 	%r114, %f169, %f49;
	neg.s32 	%r115, %r114;
	mov.f32 	%f234, 0f00000000;   	// 0
	set.eq.ftz.u32.f32 	%r116, %f231, %f234;
	neg.s32 	%r117, %r116;
	set.le.ftz.u32.f32 	%r118, %f52, %f169;
	neg.s32 	%r119, %r118;
	and.b32 	%r120, %r117, %r119;
	or.b32 	%r121, %r115, %r120;
	mov.u32 	%r122, 0;
	setp.eq.s32 	%p19, %r121, %r122;
	@%p19 bra 	$Lt_27_23810;
	.loc	22	92	0
	ld.param.f32 	%f235, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inLevelsParameters+60];
	ld.param.f32 	%f236, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inLevelsParameters+72];
	ld.param.f32 	%f237, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inLevelsParameters+76];
	sub.ftz.f32 	%f238, %f237, %f236;
	ld.param.f32 	%f239, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inLevelsParameters+68];
	sub.ftz.f32 	%f240, %f169, %f235;
	ld.param.f32 	%f241, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inLevelsParameters+64];
	sub.ftz.f32 	%f242, %f241, %f235;
	div.approx.ftz.f32 	%f243, %f240, %f242;
	mov.f32 	%f244, 0f00000000;   	// 0
	max.ftz.f32 	%f245, %f243, %f244;
	lg2.approx.ftz.f32 	%f246, %f245;
	mul.ftz.f32 	%f247, %f239, %f246;
	ex2.approx.ftz.f32 	%f248, %f247;
	fma.rn.ftz.f32 	%f169, %f238, %f248, %f236;
	bra.uni 	$Lt_27_25090;
$Lt_27_23810:
	ld.param.f32 	%f249, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inLevelsParameters+52];
	ld.param.f32 	%f250, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inLevelsParameters+56];
	sub.ftz.f32 	%f251, %f250, %f249;
	ld.param.f32 	%f252, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inLevelsParameters+40];
	sub.ftz.f32 	%f253, %f169, %f252;
	ld.param.f32 	%f254, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inLevelsParameters+44];
	sub.ftz.f32 	%f255, %f254, %f252;
	ld.param.f32 	%f256, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inLevelsParameters+48];
	set.gt.ftz.u32.f32 	%r123, %f52, %f169;
	neg.s32 	%r124, %r123;
	set.lt.ftz.u32.f32 	%r125, %f19, %f169;
	neg.s32 	%r126, %r125;
	and.b32 	%r127, %r124, %r126;
	mov.u32 	%r128, 0;
	setp.eq.s32 	%p20, %r127, %r128;
	@%p20 bra 	$Lt_27_24322;
	.loc	22	96	0
	div.approx.ftz.f32 	%f257, %f253, %f255;
	mov.f32 	%f258, 0f00000000;   	// 0
	max.ftz.f32 	%f259, %f257, %f258;
	lg2.approx.ftz.f32 	%f260, %f259;
	mul.ftz.f32 	%f261, %f256, %f260;
	ex2.approx.ftz.f32 	%f262, %f261;
	fma.rn.ftz.f32 	%f169, %f251, %f262, %f249;
	bra.uni 	$Lt_27_25090;
$Lt_27_24322:
	set.gt.ftz.u32.f32 	%r129, %f19, %f169;
	neg.s32 	%r130, %r129;
	set.gt.ftz.u32.f32 	%r131, %f169, %f10;
	neg.s32 	%r132, %r131;
	and.b32 	%r133, %r130, %r132;
	mov.u32 	%r134, 0;
	setp.eq.s32 	%p21, %r133, %r134;
	@%p21 bra 	$Lt_27_24834;
	.loc	23	79	0
	ld.param.f32 	%f263, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inLevelsParameters+32];
	ld.param.f32 	%f264, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inLevelsParameters+36];
	sub.ftz.f32 	%f265, %f264, %f263;
	ld.param.f32 	%f266, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inLevelsParameters+20];
	sub.ftz.f32 	%f267, %f169, %f266;
	ld.param.f32 	%f268, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inLevelsParameters+24];
	sub.ftz.f32 	%f269, %f268, %f266;
	div.approx.ftz.f32 	%f270, %f267, %f269;
	mov.f32 	%f271, 0f00000000;   	// 0
	max.ftz.f32 	%f272, %f270, %f271;
	lg2.approx.ftz.f32 	%f273, %f272;
	ld.param.f32 	%f274, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inLevelsParameters+28];
	mul.ftz.f32 	%f275, %f274, %f273;
	ex2.approx.ftz.f32 	%f276, %f275;
	fma.rn.ftz.f32 	%f277, %f265, %f276, %f263;
	sub.ftz.f32 	%f278, %f169, %f10;
	sub.ftz.f32 	%f279, %f19, %f10;
	div.approx.ftz.f32 	%f280, %f278, %f279;
	div.approx.ftz.f32 	%f281, %f253, %f255;
	mov.f32 	%f282, 0f00000000;   	// 0
	max.ftz.f32 	%f283, %f281, %f282;
	lg2.approx.ftz.f32 	%f284, %f283;
	mul.ftz.f32 	%f285, %f256, %f284;
	ex2.approx.ftz.f32 	%f286, %f285;
	fma.rn.ftz.f32 	%f287, %f251, %f286, %f249;
	sub.ftz.f32 	%f288, %f287, %f277;
	fma.rn.ftz.f32 	%f289, %f280, %f288, %f277;
	.loc	22	102	0
	mov.f32 	%f169, %f289;
	bra.uni 	$Lt_27_25090;
$Lt_27_24834:
	setp.lt.ftz.f32 	%p22, %f52, %f169;
	@!%p22 bra 	$Lt_27_25346;
	.loc	23	79	0
	div.approx.ftz.f32 	%f290, %f253, %f255;
	mov.f32 	%f291, 0f00000000;   	// 0
	max.ftz.f32 	%f292, %f290, %f291;
	lg2.approx.ftz.f32 	%f293, %f292;
	mul.ftz.f32 	%f294, %f293, %f256;
	ex2.approx.ftz.f32 	%f295, %f294;
	fma.rn.ftz.f32 	%f296, %f251, %f295, %f249;
	ld.param.f32 	%f297, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inLevelsParameters+60];
	ld.param.f32 	%f298, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inLevelsParameters+72];
	sub.ftz.f32 	%f299, %f169, %f52;
	sub.ftz.f32 	%f300, %f49, %f52;
	div.approx.ftz.f32 	%f301, %f299, %f300;
	ld.param.f32 	%f302, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inLevelsParameters+76];
	sub.ftz.f32 	%f303, %f302, %f298;
	ld.param.f32 	%f304, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inLevelsParameters+68];
	sub.ftz.f32 	%f305, %f169, %f297;
	ld.param.f32 	%f306, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inLevelsParameters+64];
	sub.ftz.f32 	%f307, %f306, %f297;
	div.approx.ftz.f32 	%f308, %f305, %f307;
	mov.f32 	%f309, 0f00000000;   	// 0
	max.ftz.f32 	%f310, %f308, %f309;
	lg2.approx.ftz.f32 	%f311, %f310;
	mul.ftz.f32 	%f312, %f304, %f311;
	ex2.approx.ftz.f32 	%f313, %f312;
	fma.rn.ftz.f32 	%f314, %f303, %f313, %f298;
	sub.ftz.f32 	%f315, %f314, %f296;
	fma.rn.ftz.f32 	%f316, %f301, %f315, %f296;
	.loc	22	108	0
	mov.f32 	%f169, %f316;
	bra.uni 	$Lt_27_25090;
$Lt_27_25346:
	.loc	22	112	0
	div.approx.ftz.f32 	%f317, %f253, %f255;
	mov.f32 	%f318, 0f00000000;   	// 0
	max.ftz.f32 	%f319, %f317, %f318;
	lg2.approx.ftz.f32 	%f320, %f319;
	mul.ftz.f32 	%f321, %f256, %f320;
	ex2.approx.ftz.f32 	%f322, %f321;
	fma.rn.ftz.f32 	%f169, %f251, %f322, %f249;
$Lt_27_25090:
$Lt_27_24578:
$Lt_27_24066:
$Lt_27_23554:
$Lt_27_23042:
	.loc	22	366	0
	sub.ftz.f32 	%f323, %f169, %f216;
	mov.f32 	%f324, 0f3f000000;   	// 0.5
	fma.rn.ftz.f32 	%f325, %f323, %f324, %f216;
	sub.ftz.f32 	%f326, %f325, %f25;
	ld.param.f32 	%f327, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inBalanceTonalRangeParameters+48];
	sub.ftz.f32 	%f328, %f327, %f27;
	ld.param.f32 	%f329, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inBalanceTonalRangeParameters+44];
	sub.ftz.f32 	%f330, %f329, %f29;
	fma.rn.ftz.f32 	%f331, %f5, %f326, %f25;
	ld.param.f32 	%f332, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inBalanceTonalRangeParameters+52];
	fma.rn.ftz.f32 	%f333, %f332, %f328, %f27;
	fma.rn.ftz.f32 	%f334, %f332, %f330, %f29;
	ld.param.f32 	%f335, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inBalanceTonalRangeParameters+16];
	mul.ftz.f32 	%f336, %f333, %f335;
	mul.ftz.f32 	%f337, %f334, %f335;
	sub.ftz.f32 	%f338, %f47, %f336;
	sub.ftz.f32 	%f339, %f41, %f337;
	mov.f32 	%f340, 0f3f000000;   	// 0.5
	fma.rn.ftz.f32 	%f341, %f338, %f340, %f336;
	mov.f32 	%f342, 0f3f000000;   	// 0.5
	fma.rn.ftz.f32 	%f343, %f339, %f342, %f337;
	sub.ftz.f32 	%f344, %f341, %f27;
	sub.ftz.f32 	%f345, %f343, %f29;
	fma.rn.ftz.f32 	%f346, %f5, %f344, %f27;
	fma.rn.ftz.f32 	%f347, %f5, %f345, %f29;
	ld.const.f32 	%f348, [k601YPbPr_To_RGB32f+4];
	mul.ftz.f32 	%f349, %f348, %f347;
	ld.const.f32 	%f350, [k601YPbPr_To_RGB32f+0];
	fma.rn.ftz.f32 	%f351, %f350, %f331, %f349;
	ld.const.f32 	%f352, [k601YPbPr_To_RGB32f+8];
	fma.rn.ftz.f32 	%f353, %f352, %f346, %f351;
	ld.const.f32 	%f354, [k601YPbPr_To_RGB32f+16];
	mul.ftz.f32 	%f355, %f354, %f347;
	ld.const.f32 	%f356, [k601YPbPr_To_RGB32f+12];
	fma.rn.ftz.f32 	%f357, %f356, %f331, %f355;
	ld.const.f32 	%f358, [k601YPbPr_To_RGB32f+20];
	fma.rn.ftz.f32 	%f359, %f358, %f346, %f357;
	ld.const.f32 	%f360, [k601YPbPr_To_RGB32f+28];
	mul.ftz.f32 	%f361, %f360, %f347;
	ld.const.f32 	%f362, [k601YPbPr_To_RGB32f+24];
	fma.rn.ftz.f32 	%f363, %f362, %f331, %f361;
	ld.const.f32 	%f364, [k601YPbPr_To_RGB32f+32];
	fma.rn.ftz.f32 	%f365, %f364, %f346, %f363;
	@!%p2 bra 	$Lt_27_25858;
	.loc	20	126	0
	mul.lo.u64 	%rd12, %rd1, 8;
	add.u64 	%rd13, %rd2, %rd12;
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f365;
	mov.b32		%r135, %b1; }
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f359;
	mov.b32		%r136, %b1; }
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f353;
	mov.b32		%r137, %b1; }
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f4;
	mov.b32		%r138, %b1; }
	st.global.v4.u16 	[%rd13+0], {%r135,%r136,%r137,%r138};
	.loc	22	366	0
	bra.uni 	$Lt_27_25602;
$Lt_27_25858:
	.loc	20	126	0
	mul.lo.u64 	%rd14, %rd1, 16;
	add.u64 	%rd15, %rd2, %rd14;
	st.global.v4.f32 	[%rd15+0], {%f365,%f359,%f353,%f4};
$Lt_27_25602:
$Lt_27_18946:
	.loc	22	368	0
	exit;
$LDWend_ThreeWayColorCorrector_CompositeKernel:
	} // ThreeWayColorCorrector_CompositeKernel
	.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};

