	.version 1.4
	.target sm_11, map_f64_to_f32
	// compiled with ../../../External/3rdParty/NVIDIA/CUDA/win/64/bin64/../open64/lib//be.exe
	// nvopencc 2.3 built on 2009-07-14

	//-----------------------------------------------------------
	// Compiling C:\Users\dvaeng\AppData\Local\Temp/tmpxft_0001d4d0_00000000-9_ThreeWayColorCorrector.cpp3.i (C:/Users/dvaeng/AppData/Local/Temp/ccBI#.a21108)
	//-----------------------------------------------------------

	//-----------------------------------------------------------
	// Options:
	//-----------------------------------------------------------
	//  Target:ptx, ISA:sm_11, Endian:little, Pointer Size:64
	//  -O3	(Optimization level)
	//  -g0	(Debug level)
	//  -m2	(Report advisories)
	//-----------------------------------------------------------

	.file	1	"C:\Users\dvaeng\AppData\Local\Temp/tmpxft_0001d4d0_00000000-8_ThreeWayColorCorrector.cudafe2.gpu"
	.file	2	"c:\scully64\shared\adobe\MediaCore\Display\Inc\CUDA/Effects/ThreeWayColorCorrector.h"
	.file	3	"c:\scully64\shared\adobe\MediaCore\GPUFoundation\API\Inc\GPUFoundation/KernelSupport/PixelRGB.h"
	.file	4	"c:\scully64\shared\adobe\MediaCore\GPUFoundation\API\Inc\GPUFoundation/KernelSupport/PixelYUV.h"
	.file	5	"C:\Program Files (x86)\Microsoft Visual Studio 9.0\VC\include\crtdefs.h"
	.file	6	"c:\scully64\shared\adobe\MediaCore\External\3rdParty\NVIDIA\CUDA\win\64\include\crt/device_runtime.h"
	.file	7	"c:\scully64\shared\adobe\MediaCore\External\3rdParty\NVIDIA\CUDA\win\64\include\host_defines.h"
	.file	8	"c:\scully64\shared\adobe\MediaCore\External\3rdParty\NVIDIA\CUDA\win\64\include\builtin_types.h"
	.file	9	"c:\scully64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\64\include\device_types.h"
	.file	10	"c:\scully64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\64\include\driver_types.h"
	.file	11	"c:\scully64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\64\include\texture_types.h"
	.file	12	"c:\scully64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\64\include\vector_types.h"
	.file	13	"c:\scully64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\64\include\host_defines.h"
	.file	14	"c:\scully64\shared\adobe\MediaCore\External\3rdParty\NVIDIA\CUDA\win\64\include\device_launch_parameters.h"
	.file	15	"c:\scully64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\64\include\crt\storage_class.h"
	.file	16	"C:\Program Files (x86)\Microsoft Visual Studio 9.0\VC\include\time.h"
	.file	17	"c:\scully64\shared\adobe\MediaCore\GPUFoundation\API\Inc\GPUFoundation/KernelSupport/Utils.h"
	.file	18	"c:\scully64\shared\adobe\MediaCore\GPUFoundation\API\Inc\GPUFoundation/KernelSupport/VectorUtils.h"
	.file	19	"c:/scully64/shared/adobe/MediaCore/Display/Src/CUDA/Effects/ThreeWayColorCorrector.cu"
	.file	20	"c:\scully64\shared\adobe\MediaCore\GPUFoundation\API\Inc\GPUFoundation/Numeric.h"
	.file	21	"c:\scully64\shared\adobe\MediaCore\External\3rdParty\NVIDIA\CUDA\win\64\include\common_functions.h"
	.file	22	"c:\scully64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\64\include\crt/func_macro.h"
	.file	23	"c:\scully64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\64\include\math_functions.h"
	.file	24	"c:\scully64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\64\include\device_functions.h"
	.file	25	"c:\scully64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\64\include\math_constants.h"
	.file	26	"c:\scully64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\64\include\sm_11_atomic_functions.h"
	.file	27	"c:\scully64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\64\include\sm_12_atomic_functions.h"
	.file	28	"c:\scully64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\64\include\sm_13_double_functions.h"
	.file	29	"c:\scully64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\64\include\common_types.h"
	.file	30	"c:\scully64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\64\include\texture_fetch_functions.h"
	.file	31	"c:\scully64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\64\include\math_functions_dbl_ptx1.h"
	.file	32	"c:\scully64\shared\adobe\MediaCore\GPUFoundation\API\Inc\GPUFoundation/KernelSupport/ColorSpaceConvert.h"

	.const .align 4 .b8 kRGB32f_To_601YPbPr[36] = {135,22,153,62,162,69,22,63,213,120,233,61,33,201,44,190,111,155,169,190,0,0,0,63,0,0,0,63,70,94,214,190,232,134,166,189};
	.const .align 4 .b8 k601YPbPr_To_RGB32f[36] = {0,0,128,63,0,0,0,0,188,116,179,63,0,0,128,63,152,50,176,190,158,209,54,191,0,0,128,63,229,208,226,63,0,0,0,0};
	.const .align 4 .b8 kRGB32f_To_601YCbCr[36] = {70,246,130,66,145,141,0,67,94,186,199,65,33,48,23,194,240,103,148,194,0,0,224,66,0,0,224,66,111,146,187,194,70,182,145,193};
	.const .align 4 .b8 k601YCbCr_To_RGB32f[36] = {37,160,149,59,0,0,0,0,182,23,205,59,37,160,149,59,40,15,201,186,156,239,80,187,37,160,149,59,236,155,1,60,0,0,0,0};
	.const .align 4 .b8 kRGB8u_To_601YCbCr[36] = {219,121,131,62,152,14,1,63,18,131,200,61,174,199,23,190,238,252,148,190,197,224,224,62,197,224,224,62,217,78,188,190,174,71,146,189};
	.const .align 4 .b8 k601YCbCr_To_RGB8u[36] = {127,10,149,63,0,0,0,0,160,74,204,63,127,10,149,63,254,148,200,190,184,30,80,191,127,10,149,63,78,26,1,64,0,0,0,0};
	.const .align 4 .b8 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};

	.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_inWidth,
		.param .s32 __cudaparm_ThreeWayColorCorrector_MaskKernel_inHeight)
	{
	.reg .u32 %r<23>;
	.reg .u64 %rd<10>;
	.reg .f32 %f<11>;
	.reg .pred %p<4>;
	.loc	19	133	0
$LBB1_ThreeWayColorCorrector_MaskKernel:
	.loc	19	136	0
	cvt.s32.u16 	%r1, %ctaid.x;
	cvt.s32.u16 	%r2, %ntid.x;
	mul24.lo.s32 	%r3, %r1, %r2;
	cvt.s32.u16 	%r4, %ctaid.y;
	cvt.s32.u16 	%r5, %ntid.y;
	mul24.lo.s32 	%r6, %r4, %r5;
	cvt.u32.u16 	%r7, %tid.x;
	add.u32 	%r8, %r3, %r7;
	cvt.u32.u16 	%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_0_2050;
	.loc	17	89	0
	ld.param.s32 	%r19, [__cudaparm_ThreeWayColorCorrector_MaskKernel_inPitch];
	mul24.lo.s32 	%r20, %r10, %r19;
	add.s32 	%r21, %r20, %r8;
	cvt.s64.s32 	%rd1, %r21;
	ld.param.u64 	%rd2, [__cudaparm_ThreeWayColorCorrector_MaskKernel_inImage];
	mul.lo.u64 	%rd3, %rd1, 16;
	add.u64 	%rd4, %rd2, %rd3;
	ld.global.f32 	%f1, [%rd4+12];
	.loc	19	139	0
	ld.param.u64 	%rd5, [__cudaparm_ThreeWayColorCorrector_MaskKernel_inSecondaryMask];
	mov.u64 	%rd6, 0;
	setp.eq.u64 	%p2, %rd5, %rd6;
	@%p2 bra 	$Lt_0_2818;
	.loc	17	89	0
	mul.lo.u64 	%rd7, %rd1, 4;
	add.u64 	%rd8, %rd5, %rd7;
	ld.global.f32 	%f2, [%rd8+0];
	.loc	19	141	0
	mov.f32 	%f3, %f2;
	bra.uni 	$Lt_0_2562;
$Lt_0_2818:
	mov.f32 	%f3, 0f3f800000;     	// 1
$Lt_0_2562:
	.loc	17	100	0
	ld.const.f32 	%f4, [k601YPbPr_To_RGB32f+24];
	mul.f32 	%f5, %f4, %f3;
	ld.const.f32 	%f6, [k601YPbPr_To_RGB32f+12];
	mul.f32 	%f7, %f6, %f3;
	ld.const.f32 	%f8, [k601YPbPr_To_RGB32f+0];
	mul.f32 	%f9, %f8, %f3;
	st.global.v4.f32 	[%rd4+0], {%f5,%f7,%f9,%f1};
$Lt_0_2050:
	.loc	19	147	0
	exit;
$LDWend_ThreeWayColorCorrector_MaskKernel:
	} // ThreeWayColorCorrector_MaskKernel

	.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_inWidth,
		.param .s32 __cudaparm_ThreeWayColorCorrector_TonalRangeKernel_inHeight,
		.param .align 4 .b8 __cudaparm_ThreeWayColorCorrector_TonalRangeKernel_inBalanceTonalRangeParameters[92])
	{
	.reg .u32 %r<53>;
	.reg .u64 %rd<6>;
	.reg .f32 %f<40>;
	.reg .pred %p<8>;
	.loc	19	158	0
$LBB1_ThreeWayColorCorrector_TonalRangeKernel:
	.loc	19	161	0
	cvt.s32.u16 	%r1, %ctaid.x;
	cvt.s32.u16 	%r2, %ntid.x;
	mul24.lo.s32 	%r3, %r1, %r2;
	cvt.s32.u16 	%r4, %ctaid.y;
	cvt.s32.u16 	%r5, %ntid.y;
	mul24.lo.s32 	%r6, %r4, %r5;
	cvt.u32.u16 	%r7, %tid.x;
	add.u32 	%r8, %r3, %r7;
	cvt.u32.u16 	%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_1_6658;
	.loc	17	89	0
	ld.param.s32 	%r19, [__cudaparm_ThreeWayColorCorrector_TonalRangeKernel_inPitch];
	mul24.lo.s32 	%r20, %r10, %r19;
	ld.param.u64 	%rd1, [__cudaparm_ThreeWayColorCorrector_TonalRangeKernel_inImage];
	add.s32 	%r21, %r20, %r8;
	cvt.s64.s32 	%rd2, %r21;
	mul.lo.u64 	%rd3, %rd2, 16;
	add.u64 	%rd4, %rd1, %rd3;
	ld.global.v4.f32 	{%f1,%f2,%f3,%f4}, [%rd4+0];
	.loc	19	167	0
	ld.const.f32 	%f5, [kRGB32f_To_601YPbPr+4];
	mul.f32 	%f6, %f5, %f2;
	ld.param.f32 	%f7, [__cudaparm_ThreeWayColorCorrector_TonalRangeKernel_inBalanceTonalRangeParameters+4];
	ld.param.f32 	%f8, [__cudaparm_ThreeWayColorCorrector_TonalRangeKernel_inBalanceTonalRangeParameters+0];
	add.f32 	%f9, %f7, %f8;
	ld.const.f32 	%f10, [kRGB32f_To_601YPbPr+0];
	mad.f32 	%f11, %f10, %f3, %f6;
	mov.f32 	%f12, 0f3f800000;    	// 1
	min.f32 	%f13, %f9, %f12;
	ld.const.f32 	%f14, [kRGB32f_To_601YPbPr+8];
	mad.f32 	%f15, %f14, %f1, %f11;
	set.le.u32.f32 	%r22, %f15, %f8;
	neg.s32 	%r23, %r22;
	mov.f32 	%f16, 0f00000000;    	// 0
	set.eq.u32.f32 	%r24, %f7, %f16;
	neg.s32 	%r25, %r24;
	set.ge.u32.f32 	%r26, %f13, %f15;
	neg.s32 	%r27, %r26;
	and.b32 	%r28, %r25, %r27;
	or.b32 	%r29, %r23, %r28;
	mov.u32 	%r30, 0;
	setp.eq.s32 	%p2, %r29, %r30;
	@%p2 bra 	$Lt_1_7426;
	mov.f32 	%f17, 0f00000000;    	// 0
	bra.uni 	$Lt_1_7170;
$Lt_1_7426:
	ld.param.f32 	%f18, [__cudaparm_ThreeWayColorCorrector_TonalRangeKernel_inBalanceTonalRangeParameters+12];
	ld.param.f32 	%f19, [__cudaparm_ThreeWayColorCorrector_TonalRangeKernel_inBalanceTonalRangeParameters+8];
	sub.f32 	%f20, %f19, %f18;
	mov.f32 	%f21, 0f00000000;    	// 0
	max.f32 	%f22, %f20, %f21;
	set.ge.u32.f32 	%r31, %f15, %f19;
	neg.s32 	%r32, %r31;
	mov.f32 	%f23, 0f00000000;    	// 0
	set.eq.u32.f32 	%r33, %f18, %f23;
	neg.s32 	%r34, %r33;
	set.le.u32.f32 	%r35, %f22, %f15;
	neg.s32 	%r36, %r35;
	and.b32 	%r37, %r34, %r36;
	or.b32 	%r38, %r32, %r37;
	mov.u32 	%r39, 0;
	setp.eq.s32 	%p3, %r38, %r39;
	@%p3 bra 	$Lt_1_7938;
	mov.f32 	%f17, 0f3f800000;    	// 1
	bra.uni 	$Lt_1_7682;
$Lt_1_7938:
	set.gt.u32.f32 	%r40, %f22, %f15;
	neg.s32 	%r41, %r40;
	set.lt.u32.f32 	%r42, %f13, %f15;
	neg.s32 	%r43, %r42;
	and.b32 	%r44, %r41, %r43;
	mov.u32 	%r45, 0;
	setp.eq.s32 	%p4, %r44, %r45;
	@%p4 bra 	$Lt_1_8450;
	mov.f32 	%f17, 0f3f000000;    	// 0.5
	bra.uni 	$Lt_1_8194;
$Lt_1_8450:
	set.gt.u32.f32 	%r46, %f13, %f15;
	neg.s32 	%r47, %r46;
	set.gt.u32.f32 	%r48, %f15, %f8;
	neg.s32 	%r49, %r48;
	and.b32 	%r50, %r47, %r49;
	mov.u32 	%r51, 0;
	setp.eq.s32 	%p5, %r50, %r51;
	@%p5 bra 	$Lt_1_8962;
	.loc	19	187	0
	sub.f32 	%f24, %f15, %f8;
	sub.f32 	%f25, %f13, %f8;
	div.full.f32 	%f26, %f24, %f25;
	mov.f32 	%f27, 0f3f000000;    	// 0.5
	mul.f32 	%f17, %f26, %f27;
	bra.uni 	$Lt_1_8706;
$Lt_1_8962:
	setp.lt.f32 	%p6, %f22, %f15;
	@!%p6 bra 	$Lt_1_9474;
	.loc	19	193	0
	mov.f32 	%f28, 0f3f000000;    	// 0.5
	sub.f32 	%f29, %f15, %f22;
	sub.f32 	%f30, %f19, %f22;
	div.full.f32 	%f31, %f29, %f30;
	mov.f32 	%f32, 0f3f000000;    	// 0.5
	mad.f32 	%f17, %f31, %f32, %f28;
	bra.uni 	$Lt_1_9218;
$Lt_1_9474:
	.loc	19	197	0
	mov.f32 	%f17, 0f3f000000;    	// 0.5
$Lt_1_9218:
$Lt_1_8706:
$Lt_1_8194:
$Lt_1_7682:
$Lt_1_7170:
	.loc	17	100	0
	ld.const.f32 	%f33, [k601YPbPr_To_RGB32f+24];
	mul.f32 	%f34, %f33, %f17;
	ld.const.f32 	%f35, [k601YPbPr_To_RGB32f+12];
	mul.f32 	%f36, %f35, %f17;
	ld.const.f32 	%f37, [k601YPbPr_To_RGB32f+0];
	mul.f32 	%f38, %f37, %f17;
	st.global.v4.f32 	[%rd4+0], {%f34,%f36,%f38,%f4};
$Lt_1_6658:
	.loc	19	204	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_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<89>;
	.reg .u64 %rd<10>;
	.reg .f32 %f<176>;
	.reg .pred %p<17>;
	.loc	19	216	0
$LBB1_ThreeWayColorCorrector_LumaKernel:
	.loc	19	219	0
	cvt.s32.u16 	%r1, %ctaid.x;
	cvt.s32.u16 	%r2, %ntid.x;
	mul24.lo.s32 	%r3, %r1, %r2;
	cvt.s32.u16 	%r4, %ctaid.y;
	cvt.s32.u16 	%r5, %ntid.y;
	mul24.lo.s32 	%r6, %r4, %r5;
	cvt.u32.u16 	%r7, %tid.x;
	add.u32 	%r8, %r3, %r7;
	cvt.u32.u16 	%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_2_12034;
	.loc	17	89	0
	ld.param.s32 	%r19, [__cudaparm_ThreeWayColorCorrector_LumaKernel_inPitch];
	mul24.lo.s32 	%r20, %r10, %r19;
	add.s32 	%r21, %r20, %r8;
	cvt.s64.s32 	%rd1, %r21;
	ld.param.u64 	%rd2, [__cudaparm_ThreeWayColorCorrector_LumaKernel_inImage];
	mul.lo.u64 	%rd3, %rd1, 16;
	add.u64 	%rd4, %rd2, %rd3;
	ld.global.v4.f32 	{%f1,%f2,%f3,%f4}, [%rd4+0];
	.loc	19	222	0
	ld.param.u64 	%rd5, [__cudaparm_ThreeWayColorCorrector_LumaKernel_inSecondaryMask];
	mov.u64 	%rd6, 0;
	setp.eq.u64 	%p2, %rd5, %rd6;
	@%p2 bra 	$Lt_2_12802;
	.loc	17	89	0
	mul.lo.u64 	%rd7, %rd1, 4;
	add.u64 	%rd8, %rd5, %rd7;
	ld.global.f32 	%f5, [%rd8+0];
	.loc	19	223	0
	mov.f32 	%f6, %f5;
	bra.uni 	$Lt_2_12546;
$Lt_2_12802:
	mov.f32 	%f6, 0f3f800000;     	// 1
$Lt_2_12546:
	.loc	19	53	0
	ld.const.f32 	%f7, [kRGB32f_To_601YPbPr+4];
	mul.f32 	%f8, %f7, %f2;
	ld.const.f32 	%f9, [kRGB32f_To_601YPbPr+0];
	mad.f32 	%f10, %f9, %f3, %f8;
	ld.const.f32 	%f11, [kRGB32f_To_601YPbPr+8];
	mad.f32 	%f12, %f11, %f1, %f10;
	mov.f32 	%f13, %f12;
	.loc	19	223	0
	ld.param.f32 	%f14, [__cudaparm_ThreeWayColorCorrector_LumaKernel_inLevelsParameters+0];
	ld.param.f32 	%f15, [__cudaparm_ThreeWayColorCorrector_LumaKernel_inLevelsParameters+4];
	sub.f32 	%f16, %f15, %f14;
	ld.param.f32 	%f17, [__cudaparm_ThreeWayColorCorrector_LumaKernel_inLevelsParameters+12];
	ld.param.f32 	%f18, [__cudaparm_ThreeWayColorCorrector_LumaKernel_inLevelsParameters+16];
	sub.f32 	%f19, %f18, %f17;
	mov.f32 	%f20, 0f3f800000;    	// 1
	setp.neu.f32 	%p3, %f16, %f20;
	@%p3 bra 	$Lt_2_16898;
	mov.f32 	%f21, 0f3f800000;    	// 1
	setp.neu.f32 	%p4, %f19, %f21;
	@%p4 bra 	$Lt_2_16898;
	ld.param.f32 	%f22, [__cudaparm_ThreeWayColorCorrector_LumaKernel_inLevelsParameters+24];
	ld.param.f32 	%f23, [__cudaparm_ThreeWayColorCorrector_LumaKernel_inLevelsParameters+20];
	sub.f32 	%f24, %f22, %f23;
	mov.f32 	%f25, 0f3f800000;    	// 1
	setp.neu.f32 	%p5, %f24, %f25;
	@%p5 bra 	$Lt_2_16898;
	ld.param.f32 	%f26, [__cudaparm_ThreeWayColorCorrector_LumaKernel_inLevelsParameters+36];
	ld.param.f32 	%f27, [__cudaparm_ThreeWayColorCorrector_LumaKernel_inLevelsParameters+32];
	sub.f32 	%f28, %f26, %f27;
	mov.f32 	%f29, 0f3f800000;    	// 1
	setp.neu.f32 	%p6, %f28, %f29;
	@%p6 bra 	$Lt_2_16898;
	ld.param.f32 	%f30, [__cudaparm_ThreeWayColorCorrector_LumaKernel_inLevelsParameters+44];
	ld.param.f32 	%f31, [__cudaparm_ThreeWayColorCorrector_LumaKernel_inLevelsParameters+40];
	sub.f32 	%f32, %f30, %f31;
	mov.f32 	%f33, 0f3f800000;    	// 1
	setp.neu.f32 	%p7, %f32, %f33;
	@%p7 bra 	$Lt_2_16898;
	ld.param.f32 	%f34, [__cudaparm_ThreeWayColorCorrector_LumaKernel_inLevelsParameters+56];
	ld.param.f32 	%f35, [__cudaparm_ThreeWayColorCorrector_LumaKernel_inLevelsParameters+52];
	sub.f32 	%f36, %f34, %f35;
	mov.f32 	%f37, 0f3f800000;    	// 1
	setp.neu.f32 	%p8, %f36, %f37;
	@%p8 bra 	$Lt_2_16898;
	ld.param.f32 	%f38, [__cudaparm_ThreeWayColorCorrector_LumaKernel_inLevelsParameters+64];
	ld.param.f32 	%f39, [__cudaparm_ThreeWayColorCorrector_LumaKernel_inLevelsParameters+60];
	sub.f32 	%f40, %f38, %f39;
	mov.f32 	%f41, 0f3f800000;    	// 1
	setp.neu.f32 	%p9, %f40, %f41;
	@%p9 bra 	$Lt_2_16898;
	ld.param.f32 	%f42, [__cudaparm_ThreeWayColorCorrector_LumaKernel_inLevelsParameters+76];
	ld.param.f32 	%f43, [__cudaparm_ThreeWayColorCorrector_LumaKernel_inLevelsParameters+72];
	sub.f32 	%f44, %f42, %f43;
	mov.f32 	%f45, 0f3f800000;    	// 1
	setp.neu.f32 	%p10, %f44, %f45;
	@!%p10 bra 	$L_2_9986;
$Lt_2_16898:
$L_2_10242:
	.loc	17	116	0
	mov.u32 	%r22, 1;
	sub.u32 	%r23, %r22, %r8;
	shr.u32 	%r24, %r10, 13;
	sub.u32 	%r25, %r8, %r10;
	sub.u32 	%r26, %r23, %r10;
	xor.b32 	%r27, %r24, %r26;
	shl.b32 	%r28, %r27, 8;
	sub.u32 	%r29, %r25, %r27;
	sub.u32 	%r30, %r10, %r27;
	xor.b32 	%r31, %r28, %r29;
	shr.u32 	%r32, %r31, 13;
	sub.u32 	%r33, %r30, %r31;
	xor.b32 	%r34, %r32, %r33;
	shr.u32 	%r35, %r34, 12;
	sub.u32 	%r36, %r27, %r31;
	sub.u32 	%r37, %r36, %r34;
	xor.b32 	%r38, %r35, %r37;
	.loc	17	117	0
	sub.u32 	%r39, %r31, %r34;
	sub.u32 	%r40, %r39, %r38;
	shl.b32 	%r41, %r38, 16;
	xor.b32 	%r40, %r40, %r41;
	.loc	17	118	0
	sub.u32 	%r42, %r34, %r38;
	sub.u32 	%r43, %r42, %r40;
	shr.u32 	%r44, %r40, 5;
	xor.b32 	%r43, %r43, %r44;
	.loc	17	119	0
	sub.u32 	%r38, %r38, %r40;
	sub.u32 	%r38, %r38, %r43;
	shr.u32 	%r45, %r43, 3;
	xor.b32 	%r38, %r38, %r45;
	.loc	17	120	0
	sub.u32 	%r40, %r40, %r43;
	sub.u32 	%r40, %r40, %r38;
	shl.b32 	%r46, %r38, 10;
	xor.b32 	%r40, %r40, %r46;
	.loc	17	121	0
	sub.u32 	%r43, %r43, %r38;
	sub.u32 	%r43, %r43, %r40;
	shr.u32 	%r47, %r40, 15;
	xor.b32 	%r43, %r43, %r47;
	.loc	19	63	0
	mov.f32 	%f46, 0f3b270d73;    	// 0.00254902
	mul.lo.u32 	%r48, %r43, 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, %r43, -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 	%f47, %r57;
	mov.f32 	%f48, 0f467ffe00;    	// 16383.5
	div.full.f32 	%f49, %f47, %f48;
	mov.f32 	%f50, 0fbf800000;    	// -1
	add.f32 	%f51, %f49, %f50;
	mad.f32 	%f13, %f46, %f51, %f12;
$L_2_9986:
	.loc	19	67	0
	ld.param.f32 	%f52, [__cudaparm_ThreeWayColorCorrector_LumaKernel_inLevelsParameters+8];
	sub.f32 	%f53, %f13, %f14;
	mov.f32 	%f54, 0f00000000;    	// 0
	max.f32 	%f55, %f53, %f54;
	div.full.f32 	%f56, %f55, %f16;
	lg2.approx.f32 	%f57, %f56;
	mul.f32 	%f58, %f52, %f57;
	ex2.approx.f32 	%f59, %f58;
	mad.f32 	%f60, %f19, %f59, %f17;
	.loc	19	79	0
	ld.param.f32 	%f61, [__cudaparm_ThreeWayColorCorrector_LumaKernel_inBalanceTonalRangeParameters+4];
	ld.param.f32 	%f62, [__cudaparm_ThreeWayColorCorrector_LumaKernel_inBalanceTonalRangeParameters+0];
	add.f32 	%f63, %f61, %f62;
	mov.f32 	%f64, 0f3f800000;    	// 1
	min.f32 	%f65, %f63, %f64;
	set.le.u32.f32 	%r58, %f13, %f62;
	neg.s32 	%r59, %r58;
	mov.f32 	%f66, 0f00000000;    	// 0
	set.eq.u32.f32 	%r60, %f61, %f66;
	neg.s32 	%r61, %r60;
	set.ge.u32.f32 	%r62, %f65, %f13;
	neg.s32 	%r63, %r62;
	and.b32 	%r64, %r61, %r63;
	or.b32 	%r65, %r59, %r64;
	mov.u32 	%r66, 0;
	setp.eq.s32 	%p11, %r65, %r66;
	@%p11 bra 	$Lt_2_13314;
	.loc	19	87	0
	ld.param.f32 	%f67, [__cudaparm_ThreeWayColorCorrector_LumaKernel_inLevelsParameters+20];
	ld.param.f32 	%f68, [__cudaparm_ThreeWayColorCorrector_LumaKernel_inLevelsParameters+32];
	ld.param.f32 	%f69, [__cudaparm_ThreeWayColorCorrector_LumaKernel_inLevelsParameters+36];
	sub.f32 	%f70, %f69, %f68;
	ld.param.f32 	%f71, [__cudaparm_ThreeWayColorCorrector_LumaKernel_inLevelsParameters+28];
	sub.f32 	%f72, %f13, %f67;
	mov.f32 	%f73, 0f00000000;    	// 0
	max.f32 	%f74, %f72, %f73;
	ld.param.f32 	%f75, [__cudaparm_ThreeWayColorCorrector_LumaKernel_inLevelsParameters+24];
	sub.f32 	%f76, %f75, %f67;
	div.full.f32 	%f77, %f74, %f76;
	lg2.approx.f32 	%f78, %f77;
	mul.f32 	%f79, %f71, %f78;
	ex2.approx.f32 	%f80, %f79;
	mad.f32 	%f13, %f70, %f80, %f68;
	bra.uni 	$Lt_2_15106;
$Lt_2_13314:
	ld.param.f32 	%f81, [__cudaparm_ThreeWayColorCorrector_LumaKernel_inBalanceTonalRangeParameters+12];
	ld.param.f32 	%f82, [__cudaparm_ThreeWayColorCorrector_LumaKernel_inBalanceTonalRangeParameters+8];
	sub.f32 	%f83, %f82, %f81;
	mov.f32 	%f84, 0f00000000;    	// 0
	max.f32 	%f85, %f83, %f84;
	set.ge.u32.f32 	%r67, %f13, %f82;
	neg.s32 	%r68, %r67;
	mov.f32 	%f86, 0f00000000;    	// 0
	set.eq.u32.f32 	%r69, %f81, %f86;
	neg.s32 	%r70, %r69;
	set.le.u32.f32 	%r71, %f85, %f13;
	neg.s32 	%r72, %r71;
	and.b32 	%r73, %r70, %r72;
	or.b32 	%r74, %r68, %r73;
	mov.u32 	%r75, 0;
	setp.eq.s32 	%p12, %r74, %r75;
	@%p12 bra 	$Lt_2_13826;
	.loc	19	92	0
	ld.param.f32 	%f87, [__cudaparm_ThreeWayColorCorrector_LumaKernel_inLevelsParameters+60];
	ld.param.f32 	%f88, [__cudaparm_ThreeWayColorCorrector_LumaKernel_inLevelsParameters+72];
	ld.param.f32 	%f89, [__cudaparm_ThreeWayColorCorrector_LumaKernel_inLevelsParameters+76];
	sub.f32 	%f90, %f89, %f88;
	ld.param.f32 	%f91, [__cudaparm_ThreeWayColorCorrector_LumaKernel_inLevelsParameters+68];
	sub.f32 	%f92, %f13, %f87;
	mov.f32 	%f93, 0f00000000;    	// 0
	max.f32 	%f94, %f92, %f93;
	ld.param.f32 	%f95, [__cudaparm_ThreeWayColorCorrector_LumaKernel_inLevelsParameters+64];
	sub.f32 	%f96, %f95, %f87;
	div.full.f32 	%f97, %f94, %f96;
	lg2.approx.f32 	%f98, %f97;
	mul.f32 	%f99, %f91, %f98;
	ex2.approx.f32 	%f100, %f99;
	mad.f32 	%f13, %f90, %f100, %f88;
	bra.uni 	$Lt_2_15106;
$Lt_2_13826:
	ld.param.f32 	%f101, [__cudaparm_ThreeWayColorCorrector_LumaKernel_inLevelsParameters+52];
	ld.param.f32 	%f102, [__cudaparm_ThreeWayColorCorrector_LumaKernel_inLevelsParameters+56];
	sub.f32 	%f103, %f102, %f101;
	ld.param.f32 	%f104, [__cudaparm_ThreeWayColorCorrector_LumaKernel_inLevelsParameters+40];
	sub.f32 	%f105, %f13, %f104;
	ld.param.f32 	%f106, [__cudaparm_ThreeWayColorCorrector_LumaKernel_inLevelsParameters+44];
	sub.f32 	%f107, %f106, %f104;
	mov.f32 	%f108, 0f00000000;   	// 0
	max.f32 	%f109, %f105, %f108;
	div.full.f32 	%f110, %f109, %f107;
	ld.param.f32 	%f111, [__cudaparm_ThreeWayColorCorrector_LumaKernel_inLevelsParameters+48];
	set.gt.u32.f32 	%r76, %f85, %f13;
	neg.s32 	%r77, %r76;
	set.lt.u32.f32 	%r78, %f65, %f13;
	neg.s32 	%r79, %r78;
	and.b32 	%r80, %r77, %r79;
	mov.u32 	%r81, 0;
	setp.eq.s32 	%p13, %r80, %r81;
	@%p13 bra 	$Lt_2_14338;
	.loc	19	96	0
	lg2.approx.f32 	%f112, %f110;
	mul.f32 	%f113, %f111, %f112;
	ex2.approx.f32 	%f114, %f113;
	mad.f32 	%f13, %f103, %f114, %f101;
	bra.uni 	$Lt_2_15106;
$Lt_2_14338:
	set.gt.u32.f32 	%r82, %f65, %f13;
	neg.s32 	%r83, %r82;
	set.gt.u32.f32 	%r84, %f13, %f62;
	neg.s32 	%r85, %r84;
	and.b32 	%r86, %r83, %r85;
	mov.u32 	%r87, 0;
	setp.eq.s32 	%p14, %r86, %r87;
	@%p14 bra 	$Lt_2_14850;
	.loc	19	102	0
	ld.param.f32 	%f115, [__cudaparm_ThreeWayColorCorrector_LumaKernel_inLevelsParameters+32];
	ld.param.f32 	%f116, [__cudaparm_ThreeWayColorCorrector_LumaKernel_inLevelsParameters+36];
	sub.f32 	%f117, %f116, %f115;
	ld.param.f32 	%f118, [__cudaparm_ThreeWayColorCorrector_LumaKernel_inLevelsParameters+20];
	sub.f32 	%f119, %f13, %f118;
	ld.param.f32 	%f120, [__cudaparm_ThreeWayColorCorrector_LumaKernel_inLevelsParameters+24];
	sub.f32 	%f121, %f120, %f118;
	mov.f32 	%f122, 0f00000000;   	// 0
	max.f32 	%f123, %f119, %f122;
	div.full.f32 	%f124, %f123, %f121;
	lg2.approx.f32 	%f125, %f124;
	ld.param.f32 	%f126, [__cudaparm_ThreeWayColorCorrector_LumaKernel_inLevelsParameters+28];
	mul.f32 	%f127, %f126, %f125;
	ex2.approx.f32 	%f128, %f127;
	mad.f32 	%f129, %f117, %f128, %f115;
	sub.f32 	%f130, %f13, %f62;
	sub.f32 	%f131, %f65, %f62;
	div.full.f32 	%f132, %f130, %f131;
	lg2.approx.f32 	%f133, %f110;
	mul.f32 	%f134, %f111, %f133;
	ex2.approx.f32 	%f135, %f134;
	mad.f32 	%f136, %f103, %f135, %f101;
	sub.f32 	%f137, %f136, %f129;
	mad.f32 	%f13, %f132, %f137, %f129;
	bra.uni 	$Lt_2_15106;
$Lt_2_14850:
	setp.lt.f32 	%p15, %f85, %f13;
	@!%p15 bra 	$Lt_2_15362;
	.loc	19	108	0
	lg2.approx.f32 	%f138, %f110;
	mul.f32 	%f139, %f138, %f111;
	ex2.approx.f32 	%f140, %f139;
	mad.f32 	%f141, %f103, %f140, %f101;
	ld.param.f32 	%f142, [__cudaparm_ThreeWayColorCorrector_LumaKernel_inLevelsParameters+60];
	ld.param.f32 	%f143, [__cudaparm_ThreeWayColorCorrector_LumaKernel_inLevelsParameters+72];
	sub.f32 	%f144, %f13, %f85;
	sub.f32 	%f145, %f82, %f85;
	div.full.f32 	%f146, %f144, %f145;
	ld.param.f32 	%f147, [__cudaparm_ThreeWayColorCorrector_LumaKernel_inLevelsParameters+76];
	sub.f32 	%f148, %f147, %f143;
	ld.param.f32 	%f149, [__cudaparm_ThreeWayColorCorrector_LumaKernel_inLevelsParameters+68];
	sub.f32 	%f150, %f13, %f142;
	mov.f32 	%f151, 0f00000000;   	// 0
	max.f32 	%f152, %f150, %f151;
	ld.param.f32 	%f153, [__cudaparm_ThreeWayColorCorrector_LumaKernel_inLevelsParameters+64];
	sub.f32 	%f154, %f153, %f142;
	div.full.f32 	%f155, %f152, %f154;
	lg2.approx.f32 	%f156, %f155;
	mul.f32 	%f157, %f149, %f156;
	ex2.approx.f32 	%f158, %f157;
	mad.f32 	%f159, %f148, %f158, %f143;
	sub.f32 	%f160, %f159, %f141;
	mad.f32 	%f13, %f146, %f160, %f141;
	bra.uni 	$Lt_2_15106;
$Lt_2_15362:
	.loc	19	112	0
	lg2.approx.f32 	%f161, %f110;
	mul.f32 	%f162, %f111, %f161;
	ex2.approx.f32 	%f163, %f162;
	mad.f32 	%f13, %f103, %f163, %f101;
$Lt_2_15106:
$Lt_2_14594:
$Lt_2_14082:
$Lt_2_13570:
$Lt_2_13058:
	.loc	17	100	0
	sub.f32 	%f164, %f13, %f60;
	mov.f32 	%f165, 0f3f000000;   	// 0.5
	mad.f32 	%f166, %f164, %f165, %f60;
	sub.f32 	%f167, %f166, %f12;
	mad.f32 	%f168, %f6, %f167, %f12;
	ld.const.f32 	%f169, [k601YPbPr_To_RGB32f+24];
	mul.f32 	%f170, %f169, %f168;
	ld.const.f32 	%f171, [k601YPbPr_To_RGB32f+12];
	mul.f32 	%f172, %f171, %f168;
	ld.const.f32 	%f173, [k601YPbPr_To_RGB32f+0];
	mul.f32 	%f174, %f173, %f168;
	st.global.v4.f32 	[%rd4+0], {%f170,%f172,%f174,%f4};
$Lt_2_12034:
	.loc	19	234	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_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<116>;
	.reg .u64 %rd<10>;
	.reg .f32 %f<357>;
	.reg .pred %p<23>;
	.loc	19	246	0
$LBB1_ThreeWayColorCorrector_CompositeKernel:
	.loc	19	249	0
	cvt.s32.u16 	%r1, %ctaid.x;
	cvt.s32.u16 	%r2, %ntid.x;
	mul24.lo.s32 	%r3, %r1, %r2;
	cvt.s32.u16 	%r4, %ctaid.y;
	cvt.s32.u16 	%r5, %ntid.y;
	mul24.lo.s32 	%r6, %r4, %r5;
	cvt.u32.u16 	%r7, %tid.x;
	add.u32 	%r8, %r3, %r7;
	cvt.u32.u16 	%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_3_17410;
	.loc	17	89	0
	ld.param.s32 	%r19, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inPitch];
	mul24.lo.s32 	%r20, %r10, %r19;
	add.s32 	%r21, %r20, %r8;
	cvt.s64.s32 	%rd1, %r21;
	ld.param.u64 	%rd2, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inImage];
	mul.lo.u64 	%rd3, %rd1, 16;
	add.u64 	%rd4, %rd2, %rd3;
	ld.global.v4.f32 	{%f1,%f2,%f3,%f4}, [%rd4+0];
	.loc	19	252	0
	ld.param.u64 	%rd5, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inSecondaryMask];
	mov.u64 	%rd6, 0;
	setp.eq.u64 	%p2, %rd5, %rd6;
	@%p2 bra 	$Lt_3_18178;
	.loc	17	89	0
	mul.lo.u64 	%rd7, %rd1, 4;
	add.u64 	%rd8, %rd5, %rd7;
	ld.global.f32 	%f5, [%rd8+0];
	.loc	19	253	0
	mov.f32 	%f6, %f5;
	bra.uni 	$Lt_3_17922;
$Lt_3_18178:
	mov.f32 	%f6, 0f3f800000;     	// 1
$Lt_3_17922:
	.loc	19	291	0
	ld.const.f32 	%f7, [kRGB32f_To_601YPbPr+4];
	mul.f32 	%f8, %f7, %f2;
	ld.param.f32 	%f9, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inBalanceTonalRangeParameters+4];
	mov.f32 	%f10, 0f00000000;    	// 0
	setp.eq.f32 	%p3, %f9, %f10;
	ld.param.f32 	%f11, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inBalanceTonalRangeParameters+0];
	add.f32 	%f12, %f11, %f9;
	ld.const.f32 	%f13, [kRGB32f_To_601YPbPr+28];
	mul.f32 	%f14, %f13, %f2;
	ld.const.f32 	%f15, [kRGB32f_To_601YPbPr+16];
	mul.f32 	%f16, %f15, %f2;
	ld.const.f32 	%f17, [kRGB32f_To_601YPbPr+0];
	mad.f32 	%f18, %f17, %f3, %f8;
	selp.s32 	%r22, 1, 0, %p3;
	mov.f32 	%f19, 0f3f800000;    	// 1
	min.f32 	%f20, %f12, %f19;
	ld.const.f32 	%f21, [kRGB32f_To_601YPbPr+24];
	mad.f32 	%f22, %f21, %f3, %f14;
	ld.const.f32 	%f23, [kRGB32f_To_601YPbPr+12];
	mad.f32 	%f24, %f23, %f3, %f16;
	ld.const.f32 	%f25, [kRGB32f_To_601YPbPr+8];
	mad.f32 	%f26, %f25, %f1, %f18;
	ld.const.f32 	%f27, [kRGB32f_To_601YPbPr+32];
	mad.f32 	%f28, %f27, %f1, %f22;
	ld.const.f32 	%f29, [kRGB32f_To_601YPbPr+20];
	mad.f32 	%f30, %f29, %f1, %f24;
	set.le.u32.f32 	%r23, %f26, %f11;
	neg.s32 	%r24, %r23;
	set.ge.u32.f32 	%r25, %f20, %f26;
	neg.s32 	%r26, %r25;
	and.b32 	%r27, %r22, %r26;
	or.b32 	%r28, %r24, %r27;
	mov.u32 	%r29, 0;
	setp.eq.s32 	%p4, %r28, %r29;
	@%p4 bra 	$Lt_3_18690;
	.loc	19	309	0
	ld.param.f32 	%f31, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inBalanceTonalRangeParameters+32];
	sin.approx.f32 	%f32, %f31;
	cos.approx.f32 	%f33, %f31;
	mul.f32 	%f34, %f28, %f32;
	mul.f32 	%f35, %f30, %f33;
	sub.f32 	%f36, %f35, %f34;
	ld.param.f32 	%f37, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inBalanceTonalRangeParameters+20];
	ld.param.f32 	%f38, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inBalanceTonalRangeParameters+64];
	ld.param.f32 	%f39, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inBalanceTonalRangeParameters+56];
	sub.f32 	%f40, %f39, %f36;
	mad.f32 	%f41, %f38, %f40, %f36;
	mul.f32 	%f42, %f37, %f41;
	.loc	19	310	0
	mul.f32 	%f43, %f28, %f33;
	mad.f32 	%f44, %f30, %f32, %f43;
	ld.param.f32 	%f45, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inBalanceTonalRangeParameters+60];
	sub.f32 	%f46, %f45, %f44;
	mad.f32 	%f47, %f38, %f46, %f44;
	mul.f32 	%f48, %f37, %f47;
	bra.uni 	$Lt_3_20482;
$Lt_3_18690:
	ld.param.f32 	%f49, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inBalanceTonalRangeParameters+12];
	ld.param.f32 	%f50, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inBalanceTonalRangeParameters+8];
	sub.f32 	%f51, %f50, %f49;
	mov.f32 	%f52, 0f00000000;    	// 0
	max.f32 	%f53, %f51, %f52;
	set.ge.u32.f32 	%r30, %f26, %f50;
	neg.s32 	%r31, %r30;
	mov.f32 	%f54, 0f00000000;    	// 0
	set.eq.u32.f32 	%r32, %f49, %f54;
	neg.s32 	%r33, %r32;
	set.ge.u32.f32 	%r34, %f26, %f53;
	neg.s32 	%r35, %r34;
	and.b32 	%r36, %r33, %r35;
	or.b32 	%r37, %r31, %r36;
	mov.u32 	%r38, 0;
	setp.eq.s32 	%p5, %r37, %r38;
	@%p5 bra 	$Lt_3_19202;
	.loc	19	315	0
	ld.param.f32 	%f55, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inBalanceTonalRangeParameters+40];
	sin.approx.f32 	%f56, %f55;
	cos.approx.f32 	%f57, %f55;
	mul.f32 	%f58, %f28, %f56;
	mul.f32 	%f59, %f30, %f57;
	sub.f32 	%f60, %f59, %f58;
	ld.param.f32 	%f61, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inBalanceTonalRangeParameters+88];
	ld.param.f32 	%f62, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inBalanceTonalRangeParameters+28];
	ld.param.f32 	%f63, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inBalanceTonalRangeParameters+80];
	sub.f32 	%f64, %f63, %f60;
	mad.f32 	%f65, %f61, %f64, %f60;
	mul.f32 	%f42, %f62, %f65;
	.loc	19	316	0
	mul.f32 	%f66, %f28, %f57;
	mad.f32 	%f67, %f30, %f56, %f66;
	ld.param.f32 	%f68, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inBalanceTonalRangeParameters+84];
	sub.f32 	%f69, %f68, %f67;
	mad.f32 	%f70, %f61, %f69, %f67;
	mul.f32 	%f48, %f62, %f70;
	bra.uni 	$Lt_3_20482;
$Lt_3_19202:
	ld.param.f32 	%f71, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inBalanceTonalRangeParameters+72];
	ld.param.f32 	%f72, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inBalanceTonalRangeParameters+68];
	ld.param.f32 	%f73, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inBalanceTonalRangeParameters+76];
	ld.param.f32 	%f74, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inBalanceTonalRangeParameters+24];
	ld.param.f32 	%f75, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inBalanceTonalRangeParameters+36];
	set.lt.u32.f32 	%r39, %f20, %f26;
	neg.s32 	%r40, %r39;
	set.lt.u32.f32 	%r41, %f26, %f53;
	neg.s32 	%r42, %r41;
	and.b32 	%r43, %r40, %r42;
	mov.u32 	%r44, 0;
	setp.eq.s32 	%p6, %r43, %r44;
	@%p6 bra 	$Lt_3_19714;
	.loc	19	320	0
	sin.approx.f32 	%f76, %f75;
	cos.approx.f32 	%f77, %f75;
	mul.f32 	%f78, %f28, %f76;
	mul.f32 	%f79, %f30, %f77;
	sub.f32 	%f80, %f79, %f78;
	sub.f32 	%f81, %f72, %f80;
	mad.f32 	%f82, %f73, %f81, %f80;
	mul.f32 	%f42, %f74, %f82;
	.loc	19	321	0
	mul.f32 	%f83, %f28, %f77;
	mad.f32 	%f84, %f30, %f76, %f83;
	sub.f32 	%f85, %f71, %f84;
	mad.f32 	%f86, %f73, %f85, %f84;
	mul.f32 	%f48, %f74, %f86;
	bra.uni 	$Lt_3_20482;
$Lt_3_19714:
	set.gt.u32.f32 	%r45, %f20, %f26;
	neg.s32 	%r46, %r45;
	set.gt.u32.f32 	%r47, %f26, %f11;
	neg.s32 	%r48, %r47;
	and.b32 	%r49, %r46, %r48;
	mov.u32 	%r50, 0;
	setp.eq.s32 	%p7, %r49, %r50;
	@%p7 bra 	$Lt_3_20226;
	.loc	19	327	0
	sin.approx.f32 	%f87, %f75;
	cos.approx.f32 	%f88, %f75;
	ld.param.f32 	%f89, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inBalanceTonalRangeParameters+32];
	sin.approx.f32 	%f90, %f89;
	cos.approx.f32 	%f91, %f89;
	sub.f32 	%f92, %f20, %f11;
	sub.f32 	%f93, %f26, %f11;
	mul.f32 	%f94, %f28, %f87;
	mul.f32 	%f95, %f28, %f90;
	div.full.f32 	%f96, %f93, %f92;
	mul.f32 	%f97, %f30, %f88;
	sub.f32 	%f98, %f97, %f94;
	mul.f32 	%f99, %f30, %f91;
	sub.f32 	%f100, %f99, %f95;
	ld.param.f32 	%f101, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inBalanceTonalRangeParameters+56];
	sub.f32 	%f102, %f101, %f100;
	ld.param.f32 	%f103, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inBalanceTonalRangeParameters+64];
	mad.f32 	%f104, %f103, %f102, %f100;
	ld.param.f32 	%f105, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inBalanceTonalRangeParameters+20];
	mul.f32 	%f106, %f104, %f105;
	sub.f32 	%f107, %f72, %f98;
	mad.f32 	%f108, %f73, %f107, %f98;
	mul.f32 	%f109, %f74, %f108;
	sub.f32 	%f110, %f109, %f106;
	mad.f32 	%f42, %f96, %f110, %f106;
	.loc	19	328	0
	mul.f32 	%f111, %f28, %f88;
	mul.f32 	%f112, %f28, %f91;
	mad.f32 	%f113, %f30, %f87, %f111;
	mad.f32 	%f114, %f30, %f90, %f112;
	ld.param.f32 	%f115, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inBalanceTonalRangeParameters+60];
	sub.f32 	%f116, %f115, %f114;
	mad.f32 	%f117, %f103, %f116, %f114;
	mul.f32 	%f118, %f117, %f105;
	sub.f32 	%f119, %f71, %f113;
	mad.f32 	%f120, %f73, %f119, %f113;
	mul.f32 	%f121, %f74, %f120;
	sub.f32 	%f122, %f121, %f118;
	mad.f32 	%f48, %f96, %f122, %f118;
	bra.uni 	$Lt_3_20482;
$Lt_3_20226:
	setp.gt.f32 	%p8, %f26, %f53;
	@!%p8 bra 	$Lt_3_20738;
	.loc	19	334	0
	sin.approx.f32 	%f123, %f75;
	cos.approx.f32 	%f124, %f75;
	ld.param.f32 	%f125, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inBalanceTonalRangeParameters+40];
	sin.approx.f32 	%f126, %f125;
	cos.approx.f32 	%f127, %f125;
	sub.f32 	%f128, %f50, %f53;
	sub.f32 	%f129, %f26, %f53;
	mul.f32 	%f130, %f28, %f123;
	mul.f32 	%f131, %f28, %f126;
	div.full.f32 	%f132, %f129, %f128;
	mul.f32 	%f133, %f30, %f124;
	sub.f32 	%f134, %f133, %f130;
	mul.f32 	%f135, %f30, %f127;
	sub.f32 	%f136, %f135, %f131;
	sub.f32 	%f137, %f72, %f134;
	mad.f32 	%f138, %f73, %f137, %f134;
	mul.f32 	%f139, %f138, %f74;
	ld.param.f32 	%f140, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inBalanceTonalRangeParameters+88];
	ld.param.f32 	%f141, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inBalanceTonalRangeParameters+28];
	ld.param.f32 	%f142, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inBalanceTonalRangeParameters+80];
	sub.f32 	%f143, %f142, %f136;
	mad.f32 	%f144, %f140, %f143, %f136;
	mul.f32 	%f145, %f141, %f144;
	sub.f32 	%f146, %f145, %f139;
	mad.f32 	%f42, %f132, %f146, %f139;
	.loc	19	335	0
	mul.f32 	%f147, %f28, %f124;
	mul.f32 	%f148, %f28, %f127;
	mad.f32 	%f149, %f30, %f123, %f147;
	mad.f32 	%f150, %f30, %f126, %f148;
	sub.f32 	%f151, %f71, %f149;
	mad.f32 	%f152, %f73, %f151, %f149;
	mul.f32 	%f153, %f152, %f74;
	ld.param.f32 	%f154, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inBalanceTonalRangeParameters+84];
	sub.f32 	%f155, %f154, %f150;
	mad.f32 	%f156, %f140, %f155, %f150;
	mul.f32 	%f157, %f141, %f156;
	sub.f32 	%f158, %f157, %f153;
	mad.f32 	%f48, %f132, %f158, %f153;
	bra.uni 	$Lt_3_20482;
$Lt_3_20738:
	.loc	19	339	0
	sin.approx.f32 	%f159, %f75;
	cos.approx.f32 	%f160, %f75;
	mul.f32 	%f161, %f28, %f159;
	mul.f32 	%f162, %f30, %f160;
	sub.f32 	%f163, %f162, %f161;
	sub.f32 	%f164, %f72, %f163;
	mad.f32 	%f165, %f73, %f164, %f163;
	mul.f32 	%f42, %f74, %f165;
	.loc	19	340	0
	mul.f32 	%f166, %f28, %f160;
	mad.f32 	%f167, %f30, %f159, %f166;
	sub.f32 	%f168, %f71, %f167;
	mad.f32 	%f169, %f73, %f168, %f167;
	mul.f32 	%f48, %f74, %f169;
$Lt_3_20482:
$Lt_3_19970:
$Lt_3_19458:
$Lt_3_18946:
$Lt_3_18434:
	.loc	19	53	0
	mov.f32 	%f170, %f26;
	.loc	19	344	0
	ld.param.f32 	%f171, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inLevelsParameters+0];
	ld.param.f32 	%f172, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inLevelsParameters+4];
	sub.f32 	%f173, %f172, %f171;
	ld.param.f32 	%f174, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inLevelsParameters+12];
	ld.param.f32 	%f175, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inLevelsParameters+16];
	sub.f32 	%f176, %f175, %f174;
	mov.f32 	%f177, 0f3f800000;   	// 1
	setp.neu.f32 	%p9, %f173, %f177;
	@%p9 bra 	$Lt_3_24834;
	mov.f32 	%f178, 0f3f800000;   	// 1
	setp.neu.f32 	%p10, %f176, %f178;
	@%p10 bra 	$Lt_3_24834;
	ld.param.f32 	%f179, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inLevelsParameters+24];
	ld.param.f32 	%f180, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inLevelsParameters+20];
	sub.f32 	%f181, %f179, %f180;
	mov.f32 	%f182, 0f3f800000;   	// 1
	setp.neu.f32 	%p11, %f181, %f182;
	@%p11 bra 	$Lt_3_24834;
	ld.param.f32 	%f183, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inLevelsParameters+36];
	ld.param.f32 	%f184, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inLevelsParameters+32];
	sub.f32 	%f185, %f183, %f184;
	mov.f32 	%f186, 0f3f800000;   	// 1
	setp.neu.f32 	%p12, %f185, %f186;
	@%p12 bra 	$Lt_3_24834;
	ld.param.f32 	%f187, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inLevelsParameters+44];
	ld.param.f32 	%f188, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inLevelsParameters+40];
	sub.f32 	%f189, %f187, %f188;
	mov.f32 	%f190, 0f3f800000;   	// 1
	setp.neu.f32 	%p13, %f189, %f190;
	@%p13 bra 	$Lt_3_24834;
	ld.param.f32 	%f191, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inLevelsParameters+56];
	ld.param.f32 	%f192, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inLevelsParameters+52];
	sub.f32 	%f193, %f191, %f192;
	mov.f32 	%f194, 0f3f800000;   	// 1
	setp.neu.f32 	%p14, %f193, %f194;
	@%p14 bra 	$Lt_3_24834;
	ld.param.f32 	%f195, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inLevelsParameters+64];
	ld.param.f32 	%f196, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inLevelsParameters+60];
	sub.f32 	%f197, %f195, %f196;
	mov.f32 	%f198, 0f3f800000;   	// 1
	setp.neu.f32 	%p15, %f197, %f198;
	@%p15 bra 	$Lt_3_24834;
	ld.param.f32 	%f199, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inLevelsParameters+76];
	ld.param.f32 	%f200, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inLevelsParameters+72];
	sub.f32 	%f201, %f199, %f200;
	mov.f32 	%f202, 0f3f800000;   	// 1
	setp.neu.f32 	%p16, %f201, %f202;
	@!%p16 bra 	$L_3_15362;
$Lt_3_24834:
$L_3_15618:
	.loc	17	116	0
	mov.u32 	%r51, 1;
	sub.u32 	%r52, %r51, %r8;
	shr.u32 	%r53, %r10, 13;
	sub.u32 	%r54, %r8, %r10;
	sub.u32 	%r55, %r52, %r10;
	xor.b32 	%r56, %r53, %r55;
	shl.b32 	%r57, %r56, 8;
	sub.u32 	%r58, %r54, %r56;
	sub.u32 	%r59, %r10, %r56;
	xor.b32 	%r60, %r57, %r58;
	shr.u32 	%r61, %r60, 13;
	sub.u32 	%r62, %r59, %r60;
	xor.b32 	%r63, %r61, %r62;
	shr.u32 	%r64, %r63, 12;
	sub.u32 	%r65, %r56, %r60;
	sub.u32 	%r66, %r65, %r63;
	xor.b32 	%r67, %r64, %r66;
	.loc	17	117	0
	sub.u32 	%r68, %r60, %r63;
	sub.u32 	%r69, %r68, %r67;
	shl.b32 	%r70, %r67, 16;
	xor.b32 	%r69, %r69, %r70;
	.loc	17	118	0
	sub.u32 	%r71, %r63, %r67;
	sub.u32 	%r72, %r71, %r69;
	shr.u32 	%r73, %r69, 5;
	xor.b32 	%r72, %r72, %r73;
	.loc	17	119	0
	sub.u32 	%r67, %r67, %r69;
	sub.u32 	%r67, %r67, %r72;
	shr.u32 	%r74, %r72, 3;
	xor.b32 	%r67, %r67, %r74;
	.loc	17	120	0
	sub.u32 	%r69, %r69, %r72;
	sub.u32 	%r69, %r69, %r67;
	shl.b32 	%r75, %r67, 10;
	xor.b32 	%r69, %r69, %r75;
	.loc	17	121	0
	sub.u32 	%r72, %r72, %r67;
	sub.u32 	%r72, %r72, %r69;
	shr.u32 	%r76, %r69, 15;
	xor.b32 	%r72, %r72, %r76;
	.loc	19	63	0
	mov.f32 	%f203, 0f3b270d73;   	// 0.00254902
	mul.lo.u32 	%r77, %r72, 1103515245;
	add.u32 	%r78, %r77, 12345;
	shr.u32 	%r79, %r78, 16;
	and.b32 	%r80, %r79, 255;
	shl.b32 	%r81, %r80, 7;
	mul.lo.u32 	%r82, %r72, -1029531031;
	sub.u32 	%r83, %r82, 740551042;
	shr.u32 	%r84, %r83, 16;
	and.b32 	%r85, %r84, 255;
	xor.b32 	%r86, %r81, %r85;
	cvt.rn.f32.s32 	%f204, %r86;
	mov.f32 	%f205, 0f467ffe00;   	// 16383.5
	div.full.f32 	%f206, %f204, %f205;
	mov.f32 	%f207, 0fbf800000;   	// -1
	add.f32 	%f208, %f206, %f207;
	mad.f32 	%f170, %f203, %f208, %f26;
$L_3_15362:
	.loc	19	67	0
	ld.param.f32 	%f209, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inLevelsParameters+8];
	sub.f32 	%f210, %f170, %f171;
	mov.f32 	%f211, 0f00000000;   	// 0
	max.f32 	%f212, %f210, %f211;
	div.full.f32 	%f213, %f212, %f173;
	lg2.approx.f32 	%f214, %f213;
	mul.f32 	%f215, %f209, %f214;
	ex2.approx.f32 	%f216, %f215;
	mad.f32 	%f217, %f176, %f216, %f174;
	.loc	19	79	0
	set.le.u32.f32 	%r87, %f170, %f11;
	neg.s32 	%r88, %r87;
	set.ge.u32.f32 	%r89, %f20, %f170;
	neg.s32 	%r90, %r89;
	and.b32 	%r91, %r22, %r90;
	or.b32 	%r92, %r88, %r91;
	mov.u32 	%r93, 0;
	setp.eq.s32 	%p17, %r92, %r93;
	@%p17 bra 	$Lt_3_21250;
	.loc	19	87	0
	ld.param.f32 	%f218, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inLevelsParameters+20];
	ld.param.f32 	%f219, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inLevelsParameters+32];
	ld.param.f32 	%f220, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inLevelsParameters+36];
	sub.f32 	%f221, %f220, %f219;
	ld.param.f32 	%f222, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inLevelsParameters+28];
	sub.f32 	%f223, %f170, %f218;
	mov.f32 	%f224, 0f00000000;   	// 0
	max.f32 	%f225, %f223, %f224;
	ld.param.f32 	%f226, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inLevelsParameters+24];
	sub.f32 	%f227, %f226, %f218;
	div.full.f32 	%f228, %f225, %f227;
	lg2.approx.f32 	%f229, %f228;
	mul.f32 	%f230, %f222, %f229;
	ex2.approx.f32 	%f231, %f230;
	mad.f32 	%f170, %f221, %f231, %f219;
	bra.uni 	$Lt_3_23042;
$Lt_3_21250:
	ld.param.f32 	%f232, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inBalanceTonalRangeParameters+12];
	ld.param.f32 	%f50, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inBalanceTonalRangeParameters+8];
	sub.f32 	%f233, %f50, %f232;
	mov.f32 	%f234, 0f00000000;   	// 0
	max.f32 	%f53, %f233, %f234;
	set.ge.u32.f32 	%r94, %f170, %f50;
	neg.s32 	%r95, %r94;
	mov.f32 	%f235, 0f00000000;   	// 0
	set.eq.u32.f32 	%r96, %f232, %f235;
	neg.s32 	%r97, %r96;
	set.le.u32.f32 	%r98, %f53, %f170;
	neg.s32 	%r99, %r98;
	and.b32 	%r100, %r97, %r99;
	or.b32 	%r101, %r95, %r100;
	mov.u32 	%r102, 0;
	setp.eq.s32 	%p18, %r101, %r102;
	@%p18 bra 	$Lt_3_21762;
	.loc	19	92	0
	ld.param.f32 	%f236, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inLevelsParameters+60];
	ld.param.f32 	%f237, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inLevelsParameters+72];
	ld.param.f32 	%f238, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inLevelsParameters+76];
	sub.f32 	%f239, %f238, %f237;
	ld.param.f32 	%f240, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inLevelsParameters+68];
	sub.f32 	%f241, %f170, %f236;
	mov.f32 	%f242, 0f00000000;   	// 0
	max.f32 	%f243, %f241, %f242;
	ld.param.f32 	%f244, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inLevelsParameters+64];
	sub.f32 	%f245, %f244, %f236;
	div.full.f32 	%f246, %f243, %f245;
	lg2.approx.f32 	%f247, %f246;
	mul.f32 	%f248, %f240, %f247;
	ex2.approx.f32 	%f249, %f248;
	mad.f32 	%f170, %f239, %f249, %f237;
	bra.uni 	$Lt_3_23042;
$Lt_3_21762:
	ld.param.f32 	%f250, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inLevelsParameters+52];
	ld.param.f32 	%f251, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inLevelsParameters+56];
	sub.f32 	%f252, %f251, %f250;
	ld.param.f32 	%f253, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inLevelsParameters+40];
	sub.f32 	%f254, %f170, %f253;
	ld.param.f32 	%f255, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inLevelsParameters+44];
	sub.f32 	%f256, %f255, %f253;
	mov.f32 	%f257, 0f00000000;   	// 0
	max.f32 	%f258, %f254, %f257;
	div.full.f32 	%f259, %f258, %f256;
	ld.param.f32 	%f260, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inLevelsParameters+48];
	set.gt.u32.f32 	%r103, %f53, %f170;
	neg.s32 	%r104, %r103;
	set.lt.u32.f32 	%r105, %f20, %f170;
	neg.s32 	%r106, %r105;
	and.b32 	%r107, %r104, %r106;
	mov.u32 	%r108, 0;
	setp.eq.s32 	%p19, %r107, %r108;
	@%p19 bra 	$Lt_3_22274;
	.loc	19	96	0
	lg2.approx.f32 	%f261, %f259;
	mul.f32 	%f262, %f260, %f261;
	ex2.approx.f32 	%f263, %f262;
	mad.f32 	%f170, %f252, %f263, %f250;
	bra.uni 	$Lt_3_23042;
$Lt_3_22274:
	set.gt.u32.f32 	%r109, %f20, %f170;
	neg.s32 	%r110, %r109;
	set.gt.u32.f32 	%r111, %f170, %f11;
	neg.s32 	%r112, %r111;
	and.b32 	%r113, %r110, %r112;
	mov.u32 	%r114, 0;
	setp.eq.s32 	%p20, %r113, %r114;
	@%p20 bra 	$Lt_3_22786;
	.loc	19	102	0
	ld.param.f32 	%f264, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inLevelsParameters+32];
	ld.param.f32 	%f265, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inLevelsParameters+36];
	sub.f32 	%f266, %f265, %f264;
	ld.param.f32 	%f267, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inLevelsParameters+20];
	sub.f32 	%f268, %f170, %f267;
	ld.param.f32 	%f269, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inLevelsParameters+24];
	sub.f32 	%f270, %f269, %f267;
	mov.f32 	%f271, 0f00000000;   	// 0
	max.f32 	%f272, %f268, %f271;
	div.full.f32 	%f273, %f272, %f270;
	lg2.approx.f32 	%f274, %f273;
	ld.param.f32 	%f275, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inLevelsParameters+28];
	mul.f32 	%f276, %f275, %f274;
	ex2.approx.f32 	%f277, %f276;
	mad.f32 	%f278, %f266, %f277, %f264;
	sub.f32 	%f279, %f170, %f11;
	sub.f32 	%f280, %f20, %f11;
	div.full.f32 	%f281, %f279, %f280;
	lg2.approx.f32 	%f282, %f259;
	mul.f32 	%f283, %f260, %f282;
	ex2.approx.f32 	%f284, %f283;
	mad.f32 	%f285, %f252, %f284, %f250;
	sub.f32 	%f286, %f285, %f278;
	mad.f32 	%f170, %f281, %f286, %f278;
	bra.uni 	$Lt_3_23042;
$Lt_3_22786:
	setp.lt.f32 	%p21, %f53, %f170;
	@!%p21 bra 	$Lt_3_23298;
	.loc	19	108	0
	lg2.approx.f32 	%f287, %f259;
	mul.f32 	%f288, %f287, %f260;
	ex2.approx.f32 	%f289, %f288;
	mad.f32 	%f290, %f252, %f289, %f250;
	ld.param.f32 	%f291, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inLevelsParameters+60];
	ld.param.f32 	%f292, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inLevelsParameters+72];
	sub.f32 	%f293, %f170, %f53;
	sub.f32 	%f294, %f50, %f53;
	div.full.f32 	%f295, %f293, %f294;
	ld.param.f32 	%f296, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inLevelsParameters+76];
	sub.f32 	%f297, %f296, %f292;
	ld.param.f32 	%f298, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inLevelsParameters+68];
	sub.f32 	%f299, %f170, %f291;
	mov.f32 	%f300, 0f00000000;   	// 0
	max.f32 	%f301, %f299, %f300;
	ld.param.f32 	%f302, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inLevelsParameters+64];
	sub.f32 	%f303, %f302, %f291;
	div.full.f32 	%f304, %f301, %f303;
	lg2.approx.f32 	%f305, %f304;
	mul.f32 	%f306, %f298, %f305;
	ex2.approx.f32 	%f307, %f306;
	mad.f32 	%f308, %f297, %f307, %f292;
	sub.f32 	%f309, %f308, %f290;
	mad.f32 	%f170, %f295, %f309, %f290;
	bra.uni 	$Lt_3_23042;
$Lt_3_23298:
	.loc	19	112	0
	lg2.approx.f32 	%f310, %f259;
	mul.f32 	%f311, %f260, %f310;
	ex2.approx.f32 	%f312, %f311;
	mad.f32 	%f170, %f252, %f312, %f250;
$Lt_3_23042:
$Lt_3_22530:
$Lt_3_22018:
$Lt_3_21506:
$Lt_3_20994:
	.loc	19	358	0
	sub.f32 	%f313, %f170, %f217;
	mov.f32 	%f314, 0f3f000000;   	// 0.5
	mad.f32 	%f315, %f313, %f314, %f217;
	sub.f32 	%f316, %f315, %f26;
	ld.param.f32 	%f317, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inBalanceTonalRangeParameters+48];
	sub.f32 	%f318, %f317, %f28;
	ld.param.f32 	%f319, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inBalanceTonalRangeParameters+44];
	sub.f32 	%f320, %f319, %f30;
	mad.f32 	%f321, %f6, %f316, %f26;
	ld.param.f32 	%f322, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inBalanceTonalRangeParameters+52];
	mad.f32 	%f323, %f322, %f318, %f28;
	mad.f32 	%f324, %f322, %f320, %f30;
	ld.param.f32 	%f325, [__cudaparm_ThreeWayColorCorrector_CompositeKernel_inBalanceTonalRangeParameters+16];
	mul.f32 	%f326, %f323, %f325;
	mul.f32 	%f327, %f324, %f325;
	sub.f32 	%f328, %f48, %f326;
	sub.f32 	%f329, %f42, %f327;
	mov.f32 	%f330, 0f3f000000;   	// 0.5
	mad.f32 	%f331, %f328, %f330, %f326;
	mov.f32 	%f332, 0f3f000000;   	// 0.5
	mad.f32 	%f333, %f329, %f332, %f327;
	sub.f32 	%f334, %f331, %f28;
	sub.f32 	%f335, %f333, %f30;
	mad.f32 	%f336, %f6, %f334, %f28;
	mad.f32 	%f337, %f6, %f335, %f30;
	ld.const.f32 	%f338, [k601YPbPr_To_RGB32f+4];
	mul.f32 	%f339, %f338, %f337;
	ld.const.f32 	%f340, [k601YPbPr_To_RGB32f+0];
	mad.f32 	%f341, %f340, %f321, %f339;
	ld.const.f32 	%f342, [k601YPbPr_To_RGB32f+8];
	mad.f32 	%f343, %f342, %f336, %f341;
	ld.const.f32 	%f344, [k601YPbPr_To_RGB32f+16];
	mul.f32 	%f345, %f344, %f337;
	ld.const.f32 	%f346, [k601YPbPr_To_RGB32f+12];
	mad.f32 	%f347, %f346, %f321, %f345;
	ld.const.f32 	%f348, [k601YPbPr_To_RGB32f+20];
	mad.f32 	%f349, %f348, %f336, %f347;
	ld.const.f32 	%f350, [k601YPbPr_To_RGB32f+28];
	mul.f32 	%f351, %f350, %f337;
	ld.const.f32 	%f352, [k601YPbPr_To_RGB32f+24];
	mad.f32 	%f353, %f352, %f321, %f351;
	ld.const.f32 	%f354, [k601YPbPr_To_RGB32f+32];
	mad.f32 	%f355, %f354, %f336, %f353;
	st.global.v4.f32 	[%rd4+0], {%f355,%f349,%f343,%f4};
$Lt_3_17410:
	.loc	19	360	0
	exit;
$LDWend_ThreeWayColorCorrector_CompositeKernel:
	} // ThreeWayColorCorrector_CompositeKernel

