	.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_0001d918_00000000-9_PointwiseFilterHost.cpp3.i (C:/Users/dvaeng/AppData/Local/Temp/ccBI#.a21776)
	//-----------------------------------------------------------

	//-----------------------------------------------------------
	// 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_0001d918_00000000-8_PointwiseFilterHost.cudafe2.gpu"
	.file	2	"c:\scully64\shared\adobe\MediaCore\Display\Inc\CUDA/Effects/PointwiseFilterHost.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\Display\Inc\CUDA/Effects/AlphaAdjust.h"
	.file	19	"c:\scully64\shared\adobe\MediaCore\GPUFoundation\API\Inc\GPUFoundation/KernelSupport/VectorUtils.h"
	.file	20	"c:\scully64\shared\adobe\MediaCore\Display\Inc\CUDA/Effects/BlackAndWhite.h"
	.file	21	"c:\scully64\shared\adobe\MediaCore\Display\Inc\CUDA/Effects/BrightnessAndContrast.h"
	.file	22	"c:\scully64\shared\adobe\MediaCore\Display\Inc\CUDA/Effects/ColorBalanceRGB.h"
	.file	23	"c:\scully64\shared\adobe\MediaCore\Display\Inc\CUDA/Effects/ColorPass.h"
	.file	24	"c:\scully64\shared\adobe\MediaCore\Display\Inc\CUDA/Effects/ColorReplace.h"
	.file	25	"c:\scully64\shared\adobe\MediaCore\Display\Inc\CUDA/Effects/EdgeFeather.h"
	.file	26	"c:\scully64\shared\adobe\MediaCore\GPUFoundation\API\Inc\GPUFoundation/Numeric.h"
	.file	27	"c:\scully64\shared\adobe\MediaCore\Display\Inc\CUDA/Effects/Extract.h"
	.file	28	"c:\scully64\shared\adobe\MediaCore\Display\Inc\CUDA/Effects/FastColorCorrector.h"
	.file	29	"c:\scully64\shared\adobe\MediaCore\Display\Inc\CUDA/Effects/GammaCorrection.h"
	.file	30	"c:\scully64\shared\adobe\MediaCore\Display\Inc\CUDA/Effects/HorizontalFlip.h"
	.file	31	"c:\scully64\shared\adobe\MediaCore\Display\Inc\CUDA/Effects/Noise.h"
	.file	32	"c:\scully64\shared\adobe\MediaCore\Display\Inc\CUDA/Effects/ProcAmp.h"
	.file	33	"c:\scully64\shared\adobe\MediaCore\Display\Inc\CUDA/Effects/Tint.h"
	.file	34	"c:\scully64\shared\adobe\MediaCore\Display\Inc\CUDA/Effects/VerticalFlip.h"
	.file	35	"c:/scully64/shared/adobe/MediaCore/Display/Src/CUDA/Effects/PointwiseFilterHost.cu"
	.file	36	"c:\scully64\shared\adobe\MediaCore\External\3rdParty\NVIDIA\CUDA\win\64\include\common_functions.h"
	.file	37	"c:\scully64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\64\include\crt/func_macro.h"
	.file	38	"c:\scully64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\64\include\math_functions.h"
	.file	39	"c:\scully64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\64\include\device_functions.h"
	.file	40	"c:\scully64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\64\include\math_constants.h"
	.file	41	"c:\scully64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\64\include\sm_11_atomic_functions.h"
	.file	42	"c:\scully64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\64\include\sm_12_atomic_functions.h"
	.file	43	"c:\scully64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\64\include\sm_13_double_functions.h"
	.file	44	"c:\scully64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\64\include\common_types.h"
	.file	45	"c:\scully64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\64\include\texture_fetch_functions.h"
	.file	46	"c:\scully64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\64\include\math_functions_dbl_ptx1.h"
	.file	47	"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 PointwiseFilterHostKernel1 (
		.param .u64 __cudaparm_PointwiseFilterHostKernel1_inSrcImage,
		.param .s32 __cudaparm_PointwiseFilterHostKernel1_inSrcPitch,
		.param .u64 __cudaparm_PointwiseFilterHostKernel1_inDestImage,
		.param .s32 __cudaparm_PointwiseFilterHostKernel1_inDestPitch,
		.param .s32 __cudaparm_PointwiseFilterHostKernel1_inWidth,
		.param .s32 __cudaparm_PointwiseFilterHostKernel1_inHeight,
		.param .align 4 .b8 __cudaparm_PointwiseFilterHostKernel1___val_paraminFilter0[48])
	{
	.reg .u32 %r<269>;
	.reg .u64 %rd<10>;
	.reg .f32 %f<454>;
	.reg .pred %p<60>;
	.loc	35	121	0
$LBB1_PointwiseFilterHostKernel1:
	.loc	35	123	0
	cvt.s32.u16 	%r1, %ctaid.x;
	cvt.s32.u16 	%r2, %ntid.x;
	mul24.lo.s32 	%r3, %r1, %r2;
	cvt.u32.u16 	%r4, %tid.x;
	add.u32 	%r5, %r3, %r4;
	mov.s32 	%r6, %r5;
	.loc	35	124	0
	cvt.s32.u16 	%r7, %ctaid.y;
	cvt.s32.u16 	%r8, %ntid.y;
	mul24.lo.s32 	%r9, %r7, %r8;
	cvt.u32.u16 	%r10, %tid.y;
	add.u32 	%r11, %r9, %r10;
	mov.s32 	%r12, %r11;
	ld.param.s32 	%r13, [__cudaparm_PointwiseFilterHostKernel1_inHeight];
	ld.param.s32 	%r14, [__cudaparm_PointwiseFilterHostKernel1_inWidth];
	set.gt.u32.s32 	%r15, %r13, %r11;
	neg.s32 	%r16, %r15;
	set.gt.u32.s32 	%r17, %r14, %r5;
	neg.s32 	%r18, %r17;
	and.b32 	%r19, %r16, %r18;
	mov.u32 	%r20, 0;
	setp.eq.s32 	%p1, %r19, %r20;
	@%p1 bra 	$Lt_0_52482;
	.loc	17	89	0
	ld.param.s32 	%r21, [__cudaparm_PointwiseFilterHostKernel1_inSrcPitch];
	mul24.lo.s32 	%r22, %r11, %r21;
	ld.param.u64 	%rd1, [__cudaparm_PointwiseFilterHostKernel1_inSrcImage];
	add.s32 	%r23, %r22, %r5;
	cvt.s64.s32 	%rd2, %r23;
	mul.lo.u64 	%rd3, %rd2, 16;
	add.u64 	%rd4, %rd1, %rd3;
	ld.global.v4.f32 	{%f1,%f2,%f3,%f4}, [%rd4+0];
	.loc	35	128	0
	mov.f32 	%f5, %f1;
	mov.f32 	%f6, %f2;
	mov.f32 	%f7, %f3;
	mov.f32 	%f8, %f4;
	.loc	35	54	0
	ld.param.u32 	%r24, [__cudaparm_PointwiseFilterHostKernel1___val_paraminFilter0+0];
	mov.u32 	%r25, 0;
	setp.eq.s32 	%p2, %r24, %r25;
	@%p2 bra 	$Lt_0_258;
	mov.u32 	%r26, 1;
	setp.eq.s32 	%p3, %r24, %r26;
	@%p3 bra 	$Lt_0_770;
	mov.u32 	%r27, 2;
	setp.eq.s32 	%p4, %r24, %r27;
	@%p4 bra 	$Lt_0_1026;
	mov.u32 	%r28, 3;
	setp.eq.s32 	%p5, %r24, %r28;
	@%p5 bra 	$Lt_0_1282;
	mov.u32 	%r29, 4;
	setp.eq.s32 	%p6, %r24, %r29;
	@%p6 bra 	$Lt_0_1538;
	mov.u32 	%r30, 5;
	setp.eq.s32 	%p7, %r24, %r30;
	@%p7 bra 	$Lt_0_1794;
	mov.u32 	%r31, 6;
	setp.eq.s32 	%p8, %r24, %r31;
	@%p8 bra 	$Lt_0_2050;
	mov.u32 	%r32, 7;
	setp.eq.s32 	%p9, %r24, %r32;
	@%p9 bra 	$Lt_0_2306;
	mov.u32 	%r33, 8;
	setp.eq.s32 	%p10, %r24, %r33;
	@%p10 bra 	$Lt_0_2562;
	mov.u32 	%r34, 9;
	setp.eq.s32 	%p11, %r24, %r34;
	@%p11 bra 	$Lt_0_2818;
	mov.u32 	%r35, 10;
	setp.eq.s32 	%p12, %r24, %r35;
	@%p12 bra 	$Lt_0_3074;
	mov.u32 	%r36, 11;
	setp.eq.s32 	%p13, %r24, %r36;
	@%p13 bra 	$Lt_0_3330;
	mov.u32 	%r37, 12;
	setp.eq.s32 	%p14, %r24, %r37;
	@%p14 bra 	$Lt_0_3586;
	mov.u32 	%r38, 13;
	setp.eq.s32 	%p15, %r24, %r38;
	@%p15 bra 	$Lt_0_3842;
	mov.u32 	%r39, 14;
	setp.eq.s32 	%p16, %r24, %r39;
	@%p16 bra 	$Lt_0_4098;
	mov.u32 	%r40, 15;
	setp.eq.s32 	%p17, %r24, %r40;
	@%p17 bra 	$Lt_0_4354;
	mov.u32 	%r41, 16;
	setp.eq.s32 	%p18, %r24, %r41;
	@%p18 bra 	$Lt_0_4610;
	bra.uni 	$Lt_0_514;
$Lt_0_258:
	.loc	35	57	0
	ld.param.f32 	%f9, [__cudaparm_PointwiseFilterHostKernel1___val_paraminFilter0+8];
	mov.f32 	%f10, 0f3f800000;    	// 1
	setp.eq.f32 	%p19, %f9, %f10;
	ld.param.f32 	%f11, [__cudaparm_PointwiseFilterHostKernel1___val_paraminFilter0+12];
	mov.f32 	%f12, 0f3f800000;    	// 1
	setp.eq.f32 	%p20, %f11, %f12;
	ld.param.f32 	%f13, [__cudaparm_PointwiseFilterHostKernel1___val_paraminFilter0+4];
	ld.param.f32 	%f14, [__cudaparm_PointwiseFilterHostKernel1___val_paraminFilter0+16];
	mov.f32 	%f15, 0f3f800000;    	// 1
	setp.eq.f32 	%p21, %f14, %f15;
	@!%p21 bra 	$Lt_0_53250;
	.loc	18	53	0
	cvt.sat.f32.f32 	%f16, %f4;
	mul.f32 	%f17, %f16, %f13;
	selp.f32 	%f18, %f13, %f17, %p19;
	mov.f32 	%f19, 0f3f800000;    	// 1
	sub.f32 	%f20, %f19, %f18;
	selp.f32 	%f21, %f20, %f18, %p20;
	mov.f32 	%f8, %f21;
	.loc	18	57	0
	mov.f32 	%f5, %f21;
	mov.f32 	%f6, %f21;
	mov.f32 	%f7, %f21;
	bra.uni 	$Lt_0_514;
$Lt_0_53250:
	@!%p19 bra 	$Lt_0_53762;
	.loc	18	61	0
	mov.f32 	%f22, 0f3f800000;    	// 1
	sub.f32 	%f23, %f22, %f13;
	selp.f32 	%f8, %f23, %f13, %p20;
	bra.uni 	$Lt_0_514;
$Lt_0_53762:
	.loc	18	73	0
	cvt.sat.f32.f32 	%f24, %f4;
	mov.f32 	%f25, 0f3f800000;    	// 1
	sub.f32 	%f26, %f25, %f24;
	selp.f32 	%f27, %f26, %f24, %p20;
	mul.f32 	%f8, %f27, %f13;
	bra.uni 	$Lt_0_514;
$Lt_0_770:
	.loc	19	267	0
	ld.const.f32 	%f28, [kRGB32f_To_601YPbPr+4];
	mul.f32 	%f29, %f28, %f2;
	ld.const.f32 	%f30, [kRGB32f_To_601YPbPr+0];
	mad.f32 	%f31, %f30, %f3, %f29;
	ld.const.f32 	%f32, [kRGB32f_To_601YPbPr+8];
	mad.f32 	%f33, %f32, %f1, %f31;
	ld.const.f32 	%f34, [k601YPbPr_To_RGB32f+12];
	mul.f32 	%f35, %f34, %f33;
	.loc	19	268	0
	ld.const.f32 	%f36, [k601YPbPr_To_RGB32f+0];
	mul.f32 	%f37, %f36, %f33;
	.loc	20	44	0
	ld.const.f32 	%f38, [k601YPbPr_To_RGB32f+24];
	mul.f32 	%f5, %f38, %f33;
	mov.f32 	%f6, %f35;
	mov.f32 	%f7, %f37;
	mov.f32 	%f8, %f4;
	.loc	35	61	0
	bra.uni 	$Lt_0_514;
$Lt_0_1026:
	.loc	35	63	0
	ld.param.f32 	%f39, [__cudaparm_PointwiseFilterHostKernel1___val_paraminFilter0+20];
	ld.param.f32 	%f40, [__cudaparm_PointwiseFilterHostKernel1___val_paraminFilter0+12];
	ld.param.f32 	%f41, [__cudaparm_PointwiseFilterHostKernel1___val_paraminFilter0+4];
	mov.f32 	%f42, 0f00000000;    	// 0
	setp.neu.f32 	%p22, %f41, %f42;
	@!%p22 bra 	$Lt_0_54274;
	.loc	21	44	0
	ld.param.f32 	%f43, [__cudaparm_PointwiseFilterHostKernel1___val_paraminFilter0+8];
	sub.f32 	%f44, %f3, %f40;
	mad.f32 	%f7, %f43, %f44, %f39;
	.loc	21	45	0
	sub.f32 	%f45, %f2, %f40;
	mad.f32 	%f6, %f43, %f45, %f39;
	.loc	21	46	0
	sub.f32 	%f46, %f1, %f40;
	mad.f32 	%f5, %f43, %f46, %f39;
	bra.uni 	$Lt_0_514;
$Lt_0_54274:
	.loc	21	50	0
	ld.param.f32 	%f47, [__cudaparm_PointwiseFilterHostKernel1___val_paraminFilter0+24];
	setp.gt.f32 	%p23, %f3, %f40;
	selp.f32 	%f7, %f47, %f39, %p23;
	.loc	21	51	0
	setp.gt.f32 	%p24, %f2, %f40;
	selp.f32 	%f6, %f47, %f39, %p24;
	.loc	21	52	0
	setp.gt.f32 	%p25, %f1, %f40;
	selp.f32 	%f5, %f47, %f39, %p25;
	bra.uni 	$Lt_0_514;
$Lt_0_1282:
	.loc	22	47	0
	ld.param.f32 	%f48, [__cudaparm_PointwiseFilterHostKernel1___val_paraminFilter0+4];
	mul.f32 	%f5, %f48, %f1;
	ld.param.f32 	%f49, [__cudaparm_PointwiseFilterHostKernel1___val_paraminFilter0+8];
	mul.f32 	%f6, %f49, %f2;
	ld.param.f32 	%f50, [__cudaparm_PointwiseFilterHostKernel1___val_paraminFilter0+12];
	mul.f32 	%f7, %f50, %f3;
	.loc	35	67	0
	bra.uni 	$Lt_0_514;
$Lt_0_1538:
	.loc	23	48	0
	ld.param.f32 	%f51, [__cudaparm_PointwiseFilterHostKernel1___val_paraminFilter0+4];
	mov.f32 	%f52, 0f00000000;    	// 0
	setp.eq.f32 	%p26, %f51, %f52;
	ld.param.f32 	%f53, [__cudaparm_PointwiseFilterHostKernel1___val_paraminFilter0+8];
	mov.f32 	%f54, 0f00000000;    	// 0
	max.f32 	%f55, %f1, %f54;
	mov.f32 	%f56, 0f3f800000;    	// 1
	min.f32 	%f57, %f55, %f56;
	ld.param.f32 	%f58, [__cudaparm_PointwiseFilterHostKernel1___val_paraminFilter0+12];
	sub.f32 	%f59, %f57, %f58;
	abs.f32 	%f60, %f59;
	mov.f32 	%f61, 0f00000000;    	// 0
	max.f32 	%f62, %f2, %f61;
	mov.f32 	%f63, 0f3f800000;    	// 1
	min.f32 	%f64, %f62, %f63;
	ld.param.f32 	%f65, [__cudaparm_PointwiseFilterHostKernel1___val_paraminFilter0+16];
	sub.f32 	%f66, %f64, %f65;
	abs.f32 	%f67, %f66;
	mov.f32 	%f68, 0f00000000;    	// 0
	max.f32 	%f69, %f4, %f68;
	mov.f32 	%f70, 0f3f800000;    	// 1
	min.f32 	%f71, %f69, %f70;
	sub.f32 	%f72, %f71, %f4;
	abs.f32 	%f73, %f72;
	mov.f32 	%f74, 0f00000000;    	// 0
	max.f32 	%f75, %f3, %f74;
	mov.f32 	%f76, 0f3f800000;    	// 1
	min.f32 	%f77, %f75, %f76;
	ld.param.f32 	%f78, [__cudaparm_PointwiseFilterHostKernel1___val_paraminFilter0+20];
	sub.f32 	%f79, %f77, %f78;
	abs.f32 	%f80, %f79;
	max.f32 	%f81, %f73, %f80;
	max.f32 	%f82, %f67, %f81;
	max.f32 	%f83, %f60, %f82;
	setp.ge.f32 	%p27, %f53, %f83;
	xor.pred 	%p28, %p26, %p27;
	@!%p28 bra 	$Lt_0_514;
	.loc	19	267	0
	ld.const.f32 	%f84, [kRGB32f_To_601YPbPr+4];
	mul.f32 	%f29, %f84, %f2;
	ld.const.f32 	%f85, [kRGB32f_To_601YPbPr+0];
	mad.f32 	%f31, %f85, %f3, %f29;
	ld.const.f32 	%f86, [kRGB32f_To_601YPbPr+8];
	mad.f32 	%f33, %f86, %f1, %f31;
	ld.const.f32 	%f87, [k601YPbPr_To_RGB32f+12];
	mul.f32 	%f88, %f87, %f33;
	.loc	19	268	0
	ld.const.f32 	%f89, [k601YPbPr_To_RGB32f+0];
	mul.f32 	%f90, %f89, %f33;
	.loc	20	44	0
	ld.const.f32 	%f91, [k601YPbPr_To_RGB32f+24];
	mul.f32 	%f5, %f91, %f33;
	mov.f32 	%f6, %f88;
	mov.f32 	%f7, %f90;
	bra.uni 	$Lt_0_514;
$Lt_0_1794:
	.loc	24	48	0
	ld.param.f32 	%f92, [__cudaparm_PointwiseFilterHostKernel1___val_paraminFilter0+8];
	mov.f32 	%f93, 0f00000000;    	// 0
	max.f32 	%f94, %f1, %f93;
	mov.f32 	%f95, 0f3f800000;    	// 1
	min.f32 	%f96, %f94, %f95;
	ld.param.f32 	%f97, [__cudaparm_PointwiseFilterHostKernel1___val_paraminFilter0+12];
	sub.f32 	%f98, %f96, %f97;
	abs.f32 	%f99, %f98;
	mov.f32 	%f100, 0f00000000;   	// 0
	max.f32 	%f101, %f2, %f100;
	mov.f32 	%f102, 0f3f800000;   	// 1
	min.f32 	%f103, %f101, %f102;
	ld.param.f32 	%f104, [__cudaparm_PointwiseFilterHostKernel1___val_paraminFilter0+16];
	sub.f32 	%f105, %f103, %f104;
	abs.f32 	%f106, %f105;
	mov.f32 	%f107, 0f00000000;   	// 0
	max.f32 	%f108, %f4, %f107;
	mov.f32 	%f109, 0f3f800000;   	// 1
	min.f32 	%f110, %f108, %f109;
	sub.f32 	%f111, %f110, %f4;
	abs.f32 	%f112, %f111;
	mov.f32 	%f113, 0f00000000;   	// 0
	max.f32 	%f114, %f3, %f113;
	mov.f32 	%f115, 0f3f800000;   	// 1
	min.f32 	%f116, %f114, %f115;
	ld.param.f32 	%f117, [__cudaparm_PointwiseFilterHostKernel1___val_paraminFilter0+20];
	sub.f32 	%f118, %f116, %f117;
	abs.f32 	%f119, %f118;
	max.f32 	%f120, %f112, %f119;
	max.f32 	%f121, %f106, %f120;
	max.f32 	%f122, %f99, %f121;
	setp.ge.f32 	%p29, %f92, %f122;
	@!%p29 bra 	$Lt_0_514;
	.loc	24	51	0
	ld.param.f32 	%f123, [__cudaparm_PointwiseFilterHostKernel1___val_paraminFilter0+24];
	mov.f32 	%f124, %f123;
	ld.param.f32 	%f125, [__cudaparm_PointwiseFilterHostKernel1___val_paraminFilter0+28];
	mov.f32 	%f126, %f125;
	ld.param.f32 	%f127, [__cudaparm_PointwiseFilterHostKernel1___val_paraminFilter0+32];
	mov.f32 	%f128, %f127;
	ld.param.f32 	%f129, [__cudaparm_PointwiseFilterHostKernel1___val_paraminFilter0+4];
	mov.f32 	%f130, 0f00000000;   	// 0
	setp.eq.f32 	%p30, %f129, %f130;
	@!%p30 bra 	$Lt_0_55554;
	.loc	24	60	0
	ld.const.f32 	%f131, [kRGB32f_To_601YPbPr+4];
	mul.f32 	%f29, %f131, %f2;
	ld.const.f32 	%f132, [kRGB32f_To_601YPbPr+0];
	mad.f32 	%f31, %f132, %f3, %f29;
	ld.const.f32 	%f133, [kRGB32f_To_601YPbPr+8];
	mad.f32 	%f33, %f133, %f1, %f31;
	mul.f32 	%f124, %f33, %f123;
	.loc	24	61	0
	mul.f32 	%f126, %f33, %f125;
	.loc	24	62	0
	mul.f32 	%f128, %f33, %f127;
$Lt_0_55554:
	.loc	24	65	0
	mov.f32 	%f5, %f124;
	mov.f32 	%f6, %f126;
	mov.f32 	%f7, %f128;
	bra.uni 	$Lt_0_514;
$Lt_0_2050:
	.loc	25	47	0
	sub.s32 	%r42, %r13, %r11;
	sub.s32 	%r43, %r14, %r5;
	cvt.rn.f32.s32 	%f134, %r5;
	cvt.rn.f32.s32 	%f135, %r11;
	cvt.rn.f32.s32 	%f136, %r42;
	cvt.rn.f32.s32 	%f137, %r43;
	ld.param.f32 	%f13, [__cudaparm_PointwiseFilterHostKernel1___val_paraminFilter0+4];
	sub.f32 	%f138, %f13, %f134;
	sub.f32 	%f139, %f13, %f135;
	sub.f32 	%f140, %f13, %f136;
	sub.f32 	%f141, %f13, %f137;
	cvt.rzi.s32.f32 	%r44, %f138;
	cvt.rzi.s32.f32 	%r45, %f139;
	cvt.rzi.s32.f32 	%r46, %f140;
	cvt.rzi.s32.f32 	%r47, %f141;
	max.s32 	%r48, %r44, %r45;
	max.s32 	%r49, %r47, %r48;
	max.s32 	%r50, %r46, %r49;
	mov.u32 	%r51, 0;
	setp.le.s32 	%p31, %r50, %r51;
	@%p31 bra 	$Lt_0_514;
	.loc	25	51	0
	cvt.rn.f32.s32 	%f142, %r50;
	sub.f32 	%f143, %f13, %f142;
	div.full.f32 	%f144, %f143, %f13;
	mul.f32 	%f8, %f4, %f144;
	bra.uni 	$Lt_0_514;
$Lt_0_2306:
	.loc	27	50	0
	ld.const.f32 	%f145, [kRGB32f_To_601YPbPr+4];
	mul.f32 	%f29, %f145, %f2;
	ld.param.f32 	%f146, [__cudaparm_PointwiseFilterHostKernel1___val_paraminFilter0+12];
	ld.param.f32 	%f43, [__cudaparm_PointwiseFilterHostKernel1___val_paraminFilter0+8];
	sub.f32 	%f147, %f146, %f43;
	ld.const.f32 	%f148, [kRGB32f_To_601YPbPr+0];
	mad.f32 	%f31, %f148, %f3, %f29;
	ld.param.f32 	%f149, [__cudaparm_PointwiseFilterHostKernel1___val_paraminFilter0+16];
	mul.f32 	%f150, %f149, %f147;
	ld.const.f32 	%f151, [kRGB32f_To_601YPbPr+8];
	mad.f32 	%f33, %f151, %f1, %f31;
	add.f32 	%f152, %f150, %f146;
	sub.f32 	%f153, %f43, %f150;
	mov.f32 	%f154, 0f00000000;   	// 0
	max.f32 	%f155, %f152, %f154;
	mov.f32 	%f156, 0f00000000;   	// 0
	max.f32 	%f157, %f153, %f156;
	mov.f32 	%f158, 0f3f800000;   	// 1
	min.f32 	%f159, %f155, %f158;
	mov.f32 	%f160, 0f3f800000;   	// 1
	min.f32 	%f161, %f157, %f160;
	set.gt.u32.f32 	%r52, %f161, %f33;
	neg.s32 	%r53, %r52;
	set.le.u32.f32 	%r54, %f159, %f33;
	neg.s32 	%r55, %r54;
	or.b32 	%r56, %r53, %r55;
	mov.u32 	%r57, 0;
	setp.eq.s32 	%p32, %r56, %r57;
	@%p32 bra 	$Lt_0_56834;
	mov.f32 	%f162, 0f00000000;   	// 0
	bra.uni 	$Lt_0_57602;
$Lt_0_56834:
	add.f32 	%f163, %f150, %f43;
	mov.f32 	%f164, 0f00000000;   	// 0
	max.f32 	%f165, %f163, %f164;
	mov.f32 	%f166, 0f3f800000;   	// 1
	min.f32 	%f167, %f165, %f166;
	set.le.u32.f32 	%r58, %f167, %f33;
	neg.s32 	%r59, %r58;
	sub.f32 	%f168, %f146, %f150;
	mov.f32 	%f169, 0f00000000;   	// 0
	max.f32 	%f170, %f168, %f169;
	mov.f32 	%f171, 0f3f800000;   	// 1
	min.f32 	%f172, %f170, %f171;
	set.lt.u32.f32 	%r60, %f33, %f172;
	neg.s32 	%r61, %r60;
	and.b32 	%r62, %r59, %r61;
	mov.u32 	%r63, 0;
	setp.eq.s32 	%p33, %r62, %r63;
	@%p33 bra 	$Lt_0_57346;
	mov.f32 	%f162, 0f3f800000;   	// 1
	bra.uni 	$Lt_0_57602;
$Lt_0_57346:
	add.f32 	%f173, %f150, %f150;
	setp.gt.f32 	%p34, %f167, %f33;
	@!%p34 bra 	$Lt_0_57858;
	.loc	27	62	0
	sub.f32 	%f174, %f33, %f161;
	div.full.f32 	%f162, %f174, %f173;
	bra.uni 	$Lt_0_57602;
$Lt_0_57858:
	.loc	27	66	0
	sub.f32 	%f175, %f159, %f33;
	div.full.f32 	%f162, %f175, %f173;
$Lt_0_57602:
$Lt_0_57090:
$Lt_0_56578:
	.loc	27	69	0
	mov.f32 	%f176, 0f3f800000;   	// 1
	sub.f32 	%f177, %f176, %f162;
	ld.param.f32 	%f178, [__cudaparm_PointwiseFilterHostKernel1___val_paraminFilter0+4];
	mov.f32 	%f179, 0f00000000;   	// 0
	setp.neu.f32 	%p35, %f178, %f179;
	selp.f32 	%f162, %f177, %f162, %p35;
	.loc	27	77	0
	ld.const.f32 	%f180, [k601YPbPr_To_RGB32f+24];
	mul.f32 	%f5, %f180, %f162;
	ld.const.f32 	%f181, [k601YPbPr_To_RGB32f+12];
	mul.f32 	%f6, %f181, %f162;
	ld.const.f32 	%f182, [k601YPbPr_To_RGB32f+0];
	mul.f32 	%f7, %f182, %f162;
	.loc	35	79	0
	bra.uni 	$Lt_0_514;
$Lt_0_2562:
	.loc	35	80	0
	ld.param.f32 	%f183, [__cudaparm_PointwiseFilterHostKernel1___val_paraminFilter0+44];
	mov.f32 	%f184, 0f00000000;   	// 0
	setp.ge.f32 	%p36, %f183, %f184;
	@!%p36 bra 	$L_0_50690;
	cvt.rn.f32.s32 	%f185, %r5;
	cvt.rn.f32.s32 	%f186, %r14;
	mul.f32 	%f187, %f186, %f183;
	setp.lt.f32 	%p37, %f185, %f187;
	@%p37 bra 	$L_0_50434;
$L_0_50690:
	mov.f32 	%f188, 0f00000000;   	// 0
	setp.lt.f32 	%p38, %f183, %f188;
	@!%p38 bra 	$Lt_0_514;
	cvt.rn.f32.s32 	%f189, %r11;
	cvt.rn.f32.s32 	%f190, %r13;
	mul.f32 	%f191, %f190, %f183;
	neg.f32 	%f192, %f191;
	setp.lt.f32 	%p39, %f189, %f192;
	@!%p39 bra 	$Lt_0_514;
$L_0_50434:
	.loc	28	47	0
	ld.const.f32 	%f193, [kRGB32f_To_601YPbPr+4];
	mul.f32 	%f194, %f193, %f2;
	ld.const.f32 	%f195, [kRGB32f_To_601YPbPr+0];
	mad.f32 	%f196, %f195, %f3, %f194;
	ld.const.f32 	%f197, [kRGB32f_To_601YPbPr+8];
	mad.f32 	%f33, %f197, %f1, %f196;
	mov.f32 	%f198, %f33;
	ld.param.f32 	%f43, [__cudaparm_PointwiseFilterHostKernel1___val_paraminFilter0+8];
	ld.param.f32 	%f199, [__cudaparm_PointwiseFilterHostKernel1___val_paraminFilter0+4];
	sub.f32 	%f200, %f199, %f43;
	ld.param.f32 	%f201, [__cudaparm_PointwiseFilterHostKernel1___val_paraminFilter0+16];
	ld.param.f32 	%f202, [__cudaparm_PointwiseFilterHostKernel1___val_paraminFilter0+12];
	sub.f32 	%f203, %f202, %f201;
	mov.f32 	%f204, 0f3f800000;   	// 1
	set.neu.u32.f32 	%r64, %f203, %f204;
	neg.s32 	%r65, %r64;
	mov.f32 	%f205, 0f3f800000;   	// 1
	set.neu.u32.f32 	%r66, %f200, %f205;
	neg.s32 	%r67, %r66;
	or.b32 	%r68, %r65, %r67;
	mov.u32 	%r69, 0;
	setp.eq.s32 	%p40, %r68, %r69;
	@%p40 bra 	$Lt_0_58114;
	.loc	17	116	0
	mov.u32 	%r70, 1;
	sub.u32 	%r71, %r70, %r5;
	shr.u32 	%r72, %r11, 13;
	sub.u32 	%r73, %r5, %r11;
	sub.u32 	%r74, %r71, %r11;
	xor.b32 	%r75, %r72, %r74;
	shl.b32 	%r76, %r75, 8;
	sub.u32 	%r77, %r73, %r75;
	sub.u32 	%r78, %r11, %r75;
	xor.b32 	%r79, %r76, %r77;
	shr.u32 	%r80, %r79, 13;
	sub.u32 	%r81, %r78, %r79;
	xor.b32 	%r82, %r80, %r81;
	shr.u32 	%r83, %r82, 12;
	sub.u32 	%r84, %r75, %r79;
	sub.u32 	%r85, %r84, %r82;
	xor.b32 	%r86, %r83, %r85;
	.loc	17	117	0
	sub.u32 	%r87, %r79, %r82;
	sub.u32 	%r88, %r87, %r86;
	shl.b32 	%r89, %r86, 16;
	xor.b32 	%r88, %r88, %r89;
	.loc	17	118	0
	sub.u32 	%r90, %r82, %r86;
	sub.u32 	%r91, %r90, %r88;
	shr.u32 	%r92, %r88, 5;
	xor.b32 	%r91, %r91, %r92;
	.loc	17	119	0
	sub.u32 	%r86, %r86, %r88;
	sub.u32 	%r86, %r86, %r91;
	shr.u32 	%r93, %r91, 3;
	xor.b32 	%r86, %r86, %r93;
	.loc	17	120	0
	sub.u32 	%r88, %r88, %r91;
	sub.u32 	%r88, %r88, %r86;
	shl.b32 	%r94, %r86, 10;
	xor.b32 	%r88, %r88, %r94;
	.loc	17	121	0
	sub.u32 	%r91, %r91, %r86;
	sub.u32 	%r91, %r91, %r88;
	shr.u32 	%r95, %r88, 15;
	xor.b32 	%r91, %r91, %r95;
	.loc	28	57	0
	mov.f32 	%f206, 0f3b270d73;   	// 0.00254902
	mul.lo.u32 	%r96, %r91, 1103515245;
	add.u32 	%r97, %r96, 12345;
	shr.u32 	%r98, %r97, 16;
	and.b32 	%r99, %r98, 255;
	shl.b32 	%r100, %r99, 7;
	mul.lo.u32 	%r101, %r91, -1029531031;
	sub.u32 	%r102, %r101, 740551042;
	shr.u32 	%r103, %r102, 16;
	and.b32 	%r104, %r103, 255;
	xor.b32 	%r105, %r100, %r104;
	cvt.rn.f32.s32 	%f207, %r105;
	mov.f32 	%f208, 0f467ffe00;   	// 16383.5
	div.full.f32 	%f209, %f207, %f208;
	mov.f32 	%f210, 0fbf800000;   	// -1
	add.f32 	%f211, %f209, %f210;
	mad.f32 	%f198, %f206, %f211, %f33;
$Lt_0_58114:
	sub.f32 	%f212, %f198, %f43;
	ld.param.f32 	%f213, [__cudaparm_PointwiseFilterHostKernel1___val_paraminFilter0+20];
	mov.f32 	%f214, 0f3f800000;   	// 1
	setp.neu.f32 	%p41, %f213, %f214;
	@!%p41 bra 	$Lt_0_58882;
	mov.f32 	%f215, 0f00000000;   	// 0
	setp.lt.f32 	%p42, %f212, %f215;
	@!%p42 bra 	$Lt_0_59394;
	.loc	28	66	0
	mov.f32 	%f198, %f201;
	bra.uni 	$Lt_0_58626;
$Lt_0_59394:
	.loc	28	70	0
	rcp.approx.f32 	%f216, %f200;
	mul.f32 	%f217, %f212, %f216;
	lg2.approx.f32 	%f218, %f217;
	mul.f32 	%f219, %f213, %f218;
	ex2.approx.f32 	%f220, %f219;
	mad.f32 	%f198, %f203, %f220, %f201;
	bra.uni 	$Lt_0_58626;
$Lt_0_58882:
	.loc	28	77	0
	rcp.approx.f32 	%f221, %f200;
	mul.f32 	%f222, %f212, %f221;
	mad.f32 	%f198, %f203, %f222, %f201;
$Lt_0_58626:
	.loc	19	267	0
	ld.param.f32 	%f223, [__cudaparm_PointwiseFilterHostKernel1___val_paraminFilter0+28];
	sin.approx.f32 	%f224, %f223;
	ld.const.f32 	%f225, [kRGB32f_To_601YPbPr+28];
	mul.f32 	%f226, %f225, %f2;
	ld.const.f32 	%f227, [kRGB32f_To_601YPbPr+16];
	mul.f32 	%f228, %f227, %f2;
	cos.approx.f32 	%f229, %f223;
	ld.const.f32 	%f230, [kRGB32f_To_601YPbPr+24];
	mad.f32 	%f231, %f230, %f3, %f226;
	ld.const.f32 	%f232, [kRGB32f_To_601YPbPr+12];
	mad.f32 	%f233, %f232, %f3, %f228;
	ld.const.f32 	%f234, [kRGB32f_To_601YPbPr+32];
	mad.f32 	%f235, %f234, %f1, %f231;
	ld.const.f32 	%f236, [kRGB32f_To_601YPbPr+20];
	mad.f32 	%f237, %f236, %f1, %f233;
	mul.f32 	%f238, %f224, %f235;
	mul.f32 	%f239, %f229, %f235;
	mul.f32 	%f240, %f237, %f229;
	sub.f32 	%f241, %f240, %f238;
	mad.f32 	%f242, %f237, %f224, %f239;
	ld.param.f32 	%f243, [__cudaparm_PointwiseFilterHostKernel1___val_paraminFilter0+32];
	sub.f32 	%f244, %f243, %f241;
	ld.param.f32 	%f245, [__cudaparm_PointwiseFilterHostKernel1___val_paraminFilter0+36];
	sub.f32 	%f246, %f245, %f242;
	ld.param.f32 	%f247, [__cudaparm_PointwiseFilterHostKernel1___val_paraminFilter0+40];
	mad.f32 	%f248, %f247, %f244, %f241;
	mad.f32 	%f249, %f247, %f246, %f242;
	ld.param.f32 	%f250, [__cudaparm_PointwiseFilterHostKernel1___val_paraminFilter0+24];
	mul.f32 	%f251, %f248, %f250;
	mul.f32 	%f252, %f249, %f250;
	ld.const.f32 	%f253, [k601YPbPr_To_RGB32f+16];
	mul.f32 	%f254, %f253, %f251;
	ld.const.f32 	%f255, [k601YPbPr_To_RGB32f+12];
	mad.f32 	%f256, %f255, %f198, %f254;
	ld.const.f32 	%f257, [k601YPbPr_To_RGB32f+20];
	mad.f32 	%f258, %f257, %f252, %f256;
	.loc	19	268	0
	ld.const.f32 	%f259, [k601YPbPr_To_RGB32f+4];
	mul.f32 	%f260, %f259, %f251;
	ld.const.f32 	%f261, [k601YPbPr_To_RGB32f+0];
	mad.f32 	%f262, %f261, %f198, %f260;
	ld.const.f32 	%f263, [k601YPbPr_To_RGB32f+8];
	mad.f32 	%f264, %f263, %f252, %f262;
	.loc	28	92	0
	ld.const.f32 	%f265, [k601YPbPr_To_RGB32f+28];
	mul.f32 	%f266, %f265, %f251;
	ld.const.f32 	%f267, [k601YPbPr_To_RGB32f+24];
	mad.f32 	%f268, %f267, %f198, %f266;
	ld.const.f32 	%f269, [k601YPbPr_To_RGB32f+32];
	mad.f32 	%f5, %f269, %f252, %f268;
	mov.f32 	%f6, %f258;
	mov.f32 	%f7, %f264;
	bra.uni 	$Lt_0_514;
$Lt_0_2818:
	.loc	35	83	0
	ld.param.f32 	%f270, [__cudaparm_PointwiseFilterHostKernel1___val_paraminFilter0+44];
	mov.f32 	%f271, 0f00000000;   	// 0
	setp.ge.f32 	%p43, %f270, %f271;
	@!%p43 bra 	$L_0_51458;
	cvt.rn.f32.s32 	%f272, %r5;
	cvt.rn.f32.s32 	%f273, %r14;
	mul.f32 	%f274, %f273, %f270;
	setp.lt.f32 	%p44, %f272, %f274;
	@%p44 bra 	$L_0_51202;
$L_0_51458:
	mov.f32 	%f275, 0f00000000;   	// 0
	setp.lt.f32 	%p45, %f270, %f275;
	@!%p45 bra 	$Lt_0_514;
	cvt.rn.f32.s32 	%f276, %r11;
	cvt.rn.f32.s32 	%f277, %r13;
	mul.f32 	%f278, %f277, %f270;
	neg.f32 	%f279, %f278;
	setp.lt.f32 	%p46, %f276, %f279;
	@!%p46 bra 	$Lt_0_514;
$L_0_51202:
	.loc	28	110	0
	ld.const.f32 	%f280, [kRGB32f_To_601YPbPr+4];
	mul.f32 	%f29, %f280, %f2;
	ld.param.f32 	%f43, [__cudaparm_PointwiseFilterHostKernel1___val_paraminFilter0+8];
	ld.param.f32 	%f281, [__cudaparm_PointwiseFilterHostKernel1___val_paraminFilter0+4];
	sub.f32 	%f282, %f281, %f43;
	ld.param.f32 	%f283, [__cudaparm_PointwiseFilterHostKernel1___val_paraminFilter0+16];
	ld.param.f32 	%f284, [__cudaparm_PointwiseFilterHostKernel1___val_paraminFilter0+12];
	sub.f32 	%f285, %f284, %f283;
	ld.const.f32 	%f286, [kRGB32f_To_601YPbPr+0];
	mad.f32 	%f31, %f286, %f3, %f29;
	ld.const.f32 	%f287, [kRGB32f_To_601YPbPr+8];
	mad.f32 	%f33, %f287, %f1, %f31;
	sub.f32 	%f288, %f33, %f43;
	ld.param.f32 	%f289, [__cudaparm_PointwiseFilterHostKernel1___val_paraminFilter0+20];
	mov.f32 	%f290, 0f3f800000;   	// 1
	setp.neu.f32 	%p47, %f289, %f290;
	@!%p47 bra 	$Lt_0_59906;
	.loc	28	120	0
	mov.f32 	%f291, 0f00000000;   	// 0
	max.f32 	%f292, %f288, %f291;
	div.full.f32 	%f293, %f292, %f282;
	lg2.approx.f32 	%f294, %f293;
	mul.f32 	%f295, %f289, %f294;
	ex2.approx.f32 	%f296, %f295;
	mad.f32 	%f297, %f285, %f296, %f283;
	bra.uni 	$Lt_0_59650;
$Lt_0_59906:
	.loc	28	129	0
	div.full.f32 	%f298, %f288, %f282;
	mad.f32 	%f297, %f285, %f298, %f283;
$Lt_0_59650:
	.loc	28	135	0
	ld.const.f32 	%f299, [k601YPbPr_To_RGB32f+24];
	mul.f32 	%f5, %f299, %f297;
	ld.const.f32 	%f300, [k601YPbPr_To_RGB32f+12];
	mul.f32 	%f6, %f300, %f297;
	ld.const.f32 	%f301, [k601YPbPr_To_RGB32f+0];
	mul.f32 	%f7, %f301, %f297;
	bra.uni 	$Lt_0_514;
$Lt_0_3074:
	.loc	35	86	0
	ld.param.f32 	%f302, [__cudaparm_PointwiseFilterHostKernel1___val_paraminFilter0+44];
	mov.f32 	%f303, 0f00000000;   	// 0
	setp.ge.f32 	%p48, %f302, %f303;
	@!%p48 bra 	$L_0_52226;
	cvt.rn.f32.s32 	%f304, %r5;
	cvt.rn.f32.s32 	%f305, %r14;
	mul.f32 	%f306, %f305, %f302;
	setp.lt.f32 	%p49, %f304, %f306;
	@%p49 bra 	$L_0_51970;
$L_0_52226:
	mov.f32 	%f307, 0f00000000;   	// 0
	setp.lt.f32 	%p50, %f302, %f307;
	@!%p50 bra 	$Lt_0_514;
	cvt.rn.f32.s32 	%f308, %r11;
	cvt.rn.f32.s32 	%f309, %r13;
	mul.f32 	%f310, %f309, %f302;
	neg.f32 	%f311, %f310;
	setp.lt.f32 	%p51, %f308, %f311;
	@!%p51 bra 	$Lt_0_514;
$L_0_51970:
	.loc	28	160	0
	ld.const.f32 	%f5, [k601YPbPr_To_RGB32f+24];
	ld.const.f32 	%f6, [k601YPbPr_To_RGB32f+12];
	ld.const.f32 	%f7, [k601YPbPr_To_RGB32f+0];
	bra.uni 	$Lt_0_514;
$Lt_0_3330:
	.loc	29	42	0
	ld.param.f32 	%f13, [__cudaparm_PointwiseFilterHostKernel1___val_paraminFilter0+4];
	mov.f32 	%f312, 0f00000000;   	// 0
	setp.ge.f32 	%p52, %f1, %f312;
	@!%p52 bra 	$Lt_0_60418;
	.loc	19	292	0
	lg2.approx.f32 	%f313, %f1;
	mul.f32 	%f314, %f13, %f313;
	ex2.approx.f32 	%f315, %f314;
	bra.uni 	$Lt_0_60162;
$Lt_0_60418:
	neg.f32 	%f316, %f1;
	lg2.approx.f32 	%f317, %f316;
	mul.f32 	%f318, %f13, %f317;
	ex2.approx.f32 	%f319, %f318;
	neg.f32 	%f315, %f319;
$Lt_0_60162:
	mov.f32 	%f320, 0f00000000;   	// 0
	setp.ge.f32 	%p53, %f2, %f320;
	@!%p53 bra 	$Lt_0_60930;
	.loc	19	293	0
	lg2.approx.f32 	%f321, %f2;
	mul.f32 	%f322, %f13, %f321;
	ex2.approx.f32 	%f323, %f322;
	bra.uni 	$Lt_0_60674;
$Lt_0_60930:
	neg.f32 	%f324, %f2;
	lg2.approx.f32 	%f325, %f324;
	mul.f32 	%f326, %f13, %f325;
	ex2.approx.f32 	%f327, %f326;
	neg.f32 	%f323, %f327;
$Lt_0_60674:
	mov.f32 	%f328, 0f00000000;   	// 0
	setp.ge.f32 	%p54, %f3, %f328;
	@!%p54 bra 	$Lt_0_61442;
	.loc	19	294	0
	lg2.approx.f32 	%f329, %f3;
	mul.f32 	%f330, %f13, %f329;
	ex2.approx.f32 	%f331, %f330;
	bra.uni 	$Lt_0_61186;
$Lt_0_61442:
	neg.f32 	%f332, %f3;
	lg2.approx.f32 	%f333, %f332;
	mul.f32 	%f334, %f13, %f333;
	ex2.approx.f32 	%f335, %f334;
	neg.f32 	%f331, %f335;
$Lt_0_61186:
	mov.f32 	%f336, 0f00000000;   	// 0
	setp.ge.f32 	%p55, %f4, %f336;
	@!%p55 bra 	$Lt_0_61954;
	.loc	19	295	0
	lg2.approx.f32 	%f337, %f4;
	mul.f32 	%f338, %f13, %f337;
	ex2.approx.f32 	%f339, %f338;
	bra.uni 	$Lt_0_61698;
$Lt_0_61954:
	neg.f32 	%f340, %f4;
	lg2.approx.f32 	%f341, %f340;
	mul.f32 	%f342, %f13, %f341;
	ex2.approx.f32 	%f343, %f342;
	neg.f32 	%f339, %f343;
$Lt_0_61698:
	.loc	29	42	0
	mov.f32 	%f5, %f315;
	mov.f32 	%f6, %f323;
	mov.f32 	%f7, %f331;
	mov.f32 	%f8, %f339;
	.loc	35	91	0
	bra.uni 	$Lt_0_514;
$Lt_0_3586:
	.loc	30	41	0
	sub.s32 	%r106, %r14, %r5;
	sub.s32 	%r6, %r106, 1;
	.loc	35	94	0
	bra.uni 	$Lt_0_514;
$Lt_0_3842:
	.loc	35	96	0
	sub.u32 	%r107, %r5, %r11;
	ld.param.f32 	%f13, [__cudaparm_PointwiseFilterHostKernel1___val_paraminFilter0+4];
	mov.f32 	%f344, 0f3e800000;   	// 0.25
	mul.f32 	%f345, %f13, %f344;
	ld.param.f32 	%f346, [__cudaparm_PointwiseFilterHostKernel1___val_paraminFilter0+16];
	ld.param.f32 	%f347, [__cudaparm_PointwiseFilterHostKernel1___val_paraminFilter0+8];
	mov.f32 	%f348, 0f3f800000;   	// 1
	setp.eq.f32 	%p56, %f347, %f348;
	@!%p56 bra 	$Lt_0_62466;
	.loc	17	117	0
	cvt.rzi.u32.f32 	%r108, %f346;
	shr.u32 	%r109, %r108, 13;
	sub.u32 	%r110, %r11, %r108;
	sub.u32 	%r111, %r107, %r108;
	xor.b32 	%r112, %r111, %r109;
	shl.b32 	%r113, %r112, 8;
	sub.u32 	%r114, %r110, %r112;
	sub.u32 	%r115, %r108, %r112;
	xor.b32 	%r116, %r114, %r113;
	shr.u32 	%r117, %r116, 13;
	sub.u32 	%r118, %r115, %r116;
	sub.u32 	%r119, %r112, %r116;
	xor.b32 	%r120, %r118, %r117;
	shr.u32 	%r121, %r120, 12;
	sub.u32 	%r122, %r119, %r120;
	xor.b32 	%r123, %r122, %r121;
	shl.b32 	%r124, %r123, 16;
	sub.u32 	%r125, %r116, %r120;
	sub.u32 	%r126, %r125, %r123;
	xor.b32 	%r127, %r124, %r126;
	.loc	17	118	0
	sub.u32 	%r128, %r120, %r123;
	sub.u32 	%r129, %r128, %r127;
	shr.u32 	%r130, %r127, 5;
	xor.b32 	%r131, %r129, %r130;
	.loc	17	119	0
	sub.u32 	%r132, %r123, %r127;
	sub.u32 	%r133, %r132, %r131;
	shr.u32 	%r134, %r131, 3;
	xor.b32 	%r135, %r133, %r134;
	.loc	17	120	0
	sub.u32 	%r136, %r127, %r131;
	sub.u32 	%r137, %r136, %r135;
	shl.b32 	%r138, %r135, 10;
	xor.b32 	%r139, %r137, %r138;
	.loc	17	121	0
	sub.u32 	%r140, %r131, %r135;
	sub.u32 	%r141, %r140, %r139;
	shr.u32 	%r142, %r139, 15;
	xor.b32 	%r143, %r141, %r142;
	.loc	31	48	0
	mul.lo.u32 	%r144, %r143, 1103515245;
	add.u32 	%r145, %r144, 12345;
	shr.u32 	%r146, %r145, 16;
	and.b32 	%r147, %r146, 255;
	shl.b32 	%r148, %r147, 7;
	mul.lo.u32 	%r149, %r143, -1029531031;
	sub.u32 	%r150, %r149, 740551042;
	shr.u32 	%r151, %r150, 16;
	and.b32 	%r152, %r151, 255;
	xor.b32 	%r153, %r148, %r152;
	cvt.rn.f32.s32 	%f349, %r153;
	mov.f32 	%f350, 0f46fffe00;   	// 32767
	div.full.f32 	%f351, %f349, %f350;
	mul.f32 	%f352, %f13, %f351;
	mov.f32 	%f353, 0f3f000000;   	// 0.5
	mul.f32 	%f354, %f352, %f353;
	sub.f32 	%f355, %f354, %f345;
	.loc	17	117	0
	add.u32 	%r154, %r111, 1;
	xor.b32 	%r155, %r154, %r109;
	shl.b32 	%r156, %r155, 8;
	sub.u32 	%r157, %r110, %r155;
	sub.u32 	%r158, %r108, %r155;
	xor.b32 	%r159, %r156, %r157;
	shr.u32 	%r160, %r159, 13;
	sub.u32 	%r161, %r158, %r159;
	sub.u32 	%r162, %r155, %r159;
	xor.b32 	%r163, %r160, %r161;
	shr.u32 	%r164, %r163, 12;
	sub.u32 	%r165, %r162, %r163;
	xor.b32 	%r166, %r164, %r165;
	sub.u32 	%r167, %r159, %r163;
	sub.u32 	%r168, %r167, %r166;
	shl.b32 	%r169, %r166, 16;
	xor.b32 	%r168, %r168, %r169;
	.loc	17	118	0
	sub.u32 	%r170, %r163, %r166;
	sub.u32 	%r170, %r170, %r168;
	shr.u32 	%r171, %r168, 5;
	xor.b32 	%r170, %r170, %r171;
	.loc	17	119	0
	sub.u32 	%r172, %r166, %r168;
	sub.u32 	%r172, %r172, %r170;
	shr.u32 	%r173, %r170, 3;
	xor.b32 	%r172, %r172, %r173;
	.loc	17	120	0
	sub.u32 	%r168, %r168, %r170;
	sub.u32 	%r168, %r168, %r172;
	shl.b32 	%r174, %r172, 10;
	xor.b32 	%r168, %r168, %r174;
	.loc	17	121	0
	sub.u32 	%r170, %r170, %r172;
	sub.u32 	%r170, %r170, %r168;
	shr.u32 	%r175, %r168, 15;
	xor.b32 	%r170, %r170, %r175;
	.loc	31	49	0
	mul.lo.u32 	%r176, %r170, 1103515245;
	add.u32 	%r177, %r176, 12345;
	shr.u32 	%r178, %r177, 16;
	and.b32 	%r179, %r178, 255;
	shl.b32 	%r180, %r179, 7;
	mul.lo.u32 	%r181, %r170, -1029531031;
	sub.u32 	%r182, %r181, 740551042;
	shr.u32 	%r183, %r182, 16;
	and.b32 	%r184, %r183, 255;
	xor.b32 	%r185, %r180, %r184;
	cvt.rn.f32.s32 	%f356, %r185;
	mov.f32 	%f357, 0f46fffe00;   	// 32767
	div.full.f32 	%f358, %f356, %f357;
	mul.f32 	%f359, %f13, %f358;
	mov.f32 	%f360, 0f3f000000;   	// 0.5
	mul.f32 	%f361, %f359, %f360;
	sub.f32 	%f362, %f361, %f345;
	.loc	17	117	0
	add.u32 	%r186, %r111, 2;
	xor.b32 	%r187, %r186, %r109;
	shl.b32 	%r188, %r187, 8;
	sub.u32 	%r189, %r110, %r187;
	sub.u32 	%r190, %r108, %r187;
	xor.b32 	%r191, %r188, %r189;
	shr.u32 	%r192, %r191, 13;
	sub.u32 	%r193, %r190, %r191;
	sub.u32 	%r194, %r187, %r191;
	xor.b32 	%r195, %r192, %r193;
	shr.u32 	%r196, %r195, 12;
	sub.u32 	%r197, %r194, %r195;
	xor.b32 	%r198, %r196, %r197;
	sub.u32 	%r199, %r191, %r195;
	sub.u32 	%r200, %r199, %r198;
	shl.b32 	%r201, %r198, 16;
	xor.b32 	%r200, %r200, %r201;
	.loc	17	118	0
	sub.u32 	%r202, %r195, %r198;
	sub.u32 	%r202, %r202, %r200;
	shr.u32 	%r203, %r200, 5;
	xor.b32 	%r202, %r202, %r203;
	.loc	17	119	0
	sub.u32 	%r204, %r198, %r200;
	sub.u32 	%r204, %r204, %r202;
	shr.u32 	%r205, %r202, 3;
	xor.b32 	%r204, %r204, %r205;
	.loc	17	120	0
	sub.u32 	%r200, %r200, %r202;
	sub.u32 	%r200, %r200, %r204;
	shl.b32 	%r206, %r204, 10;
	xor.b32 	%r200, %r200, %r206;
	.loc	17	121	0
	sub.u32 	%r202, %r202, %r204;
	sub.u32 	%r202, %r202, %r200;
	shr.u32 	%r207, %r200, 15;
	xor.b32 	%r202, %r202, %r207;
	.loc	31	50	0
	mul.lo.u32 	%r208, %r202, 1103515245;
	add.u32 	%r209, %r208, 12345;
	shr.u32 	%r210, %r209, 16;
	and.b32 	%r211, %r210, 255;
	shl.b32 	%r212, %r211, 7;
	mul.lo.u32 	%r213, %r202, -1029531031;
	sub.u32 	%r214, %r213, 740551042;
	shr.u32 	%r215, %r214, 16;
	and.b32 	%r216, %r215, 255;
	xor.b32 	%r217, %r212, %r216;
	cvt.rn.f32.s32 	%f363, %r217;
	mov.f32 	%f364, 0f46fffe00;   	// 32767
	div.full.f32 	%f365, %f363, %f364;
	mul.f32 	%f366, %f13, %f365;
	mov.f32 	%f367, 0f3f000000;   	// 0.5
	mul.f32 	%f368, %f366, %f367;
	sub.f32 	%f369, %f368, %f345;
	bra.uni 	$Lt_0_62210;
$Lt_0_62466:
	.loc	17	117	0
	cvt.rzi.u32.f32 	%r218, %f346;
	shr.u32 	%r219, %r218, 13;
	sub.u32 	%r220, %r11, %r218;
	sub.u32 	%r221, %r107, %r218;
	xor.b32 	%r222, %r221, %r219;
	shl.b32 	%r223, %r222, 8;
	sub.u32 	%r224, %r220, %r222;
	sub.u32 	%r225, %r218, %r222;
	xor.b32 	%r226, %r224, %r223;
	shr.u32 	%r227, %r226, 13;
	sub.u32 	%r228, %r225, %r226;
	sub.u32 	%r229, %r222, %r226;
	xor.b32 	%r230, %r228, %r227;
	shr.u32 	%r231, %r230, 12;
	sub.u32 	%r232, %r229, %r230;
	xor.b32 	%r233, %r232, %r231;
	shl.b32 	%r234, %r233, 16;
	sub.u32 	%r235, %r226, %r230;
	sub.u32 	%r236, %r235, %r233;
	xor.b32 	%r237, %r234, %r236;
	.loc	17	118	0
	sub.u32 	%r238, %r230, %r233;
	sub.u32 	%r239, %r238, %r237;
	shr.u32 	%r240, %r237, 5;
	xor.b32 	%r241, %r239, %r240;
	.loc	17	119	0
	sub.u32 	%r242, %r233, %r237;
	sub.u32 	%r243, %r242, %r241;
	shr.u32 	%r244, %r241, 3;
	xor.b32 	%r245, %r243, %r244;
	.loc	17	120	0
	sub.u32 	%r246, %r237, %r241;
	sub.u32 	%r247, %r246, %r245;
	shl.b32 	%r248, %r245, 10;
	xor.b32 	%r249, %r247, %r248;
	.loc	17	121	0
	sub.u32 	%r250, %r241, %r245;
	sub.u32 	%r251, %r250, %r249;
	shr.u32 	%r252, %r249, 15;
	xor.b32 	%r253, %r251, %r252;
	.loc	31	54	0
	mul.lo.u32 	%r254, %r253, 1103515245;
	mul.lo.u32 	%r255, %r253, -1029531031;
	add.u32 	%r256, %r254, 12345;
	sub.u32 	%r257, %r255, 740551042;
	shr.u32 	%r258, %r256, 16;
	shr.u32 	%r259, %r257, 16;
	and.b32 	%r260, %r258, 255;
	and.b32 	%r261, %r259, 255;
	shl.b32 	%r262, %r260, 7;
	xor.b32 	%r263, %r262, %r261;
	cvt.rn.f32.s32 	%f370, %r263;
	mov.f32 	%f371, 0f46fffe00;   	// 32767
	div.full.f32 	%f372, %f370, %f371;
	mul.f32 	%f373, %f372, %f13;
	mov.f32 	%f374, 0f3f000000;   	// 0.5
	mul.f32 	%f375, %f373, %f374;
	sub.f32 	%f355, %f375, %f345;
	mov.f32 	%f369, %f355;
	mov.f32 	%f362, %f355;
$Lt_0_62210:
	.loc	31	57	0
	add.f32 	%f7, %f355, %f3;
	.loc	31	58	0
	add.f32 	%f6, %f2, %f362;
	.loc	31	59	0
	add.f32 	%f5, %f1, %f369;
	ld.param.f32 	%f376, [__cudaparm_PointwiseFilterHostKernel1___val_paraminFilter0+12];
	mov.f32 	%f377, 0f3f800000;   	// 1
	setp.eq.f32 	%p57, %f376, %f377;
	@!%p57 bra 	$Lt_0_514;
	.loc	31	63	0
	mov.f32 	%f378, 0f00000000;   	// 0
	max.f32 	%f379, %f5, %f378;
	mov.f32 	%f380, 0f3f800000;   	// 1
	min.f32 	%f5, %f379, %f380;
	mov.f32 	%f381, 0f00000000;   	// 0
	max.f32 	%f382, %f6, %f381;
	mov.f32 	%f383, 0f3f800000;   	// 1
	min.f32 	%f6, %f382, %f383;
	mov.f32 	%f384, 0f00000000;   	// 0
	max.f32 	%f385, %f7, %f384;
	mov.f32 	%f386, 0f3f800000;   	// 1
	min.f32 	%f7, %f385, %f386;
	mov.f32 	%f387, 0f00000000;   	// 0
	max.f32 	%f388, %f4, %f387;
	mov.f32 	%f389, 0f3f800000;   	// 1
	min.f32 	%f8, %f388, %f389;
	bra.uni 	$Lt_0_514;
$Lt_0_4098:
	.loc	35	99	0
	cvt.rn.f32.s32 	%f390, %r5;
	cvt.rn.f32.s32 	%f391, %r14;
	ld.param.f32 	%f392, [__cudaparm_PointwiseFilterHostKernel1___val_paraminFilter0+20];
	mul.f32 	%f393, %f391, %f392;
	setp.lt.f32 	%p58, %f390, %f393;
	@!%p58 bra 	$Lt_0_514;
	.loc	19	267	0
	ld.const.f32 	%f394, [kRGB32f_To_601YPbPr+4];
	mul.f32 	%f29, %f394, %f2;
	ld.const.f32 	%f395, [kRGB32f_To_601YPbPr+28];
	mul.f32 	%f396, %f395, %f2;
	ld.const.f32 	%f397, [kRGB32f_To_601YPbPr+16];
	mul.f32 	%f398, %f397, %f2;
	ld.const.f32 	%f399, [kRGB32f_To_601YPbPr+0];
	mad.f32 	%f31, %f399, %f3, %f29;
	ld.const.f32 	%f400, [kRGB32f_To_601YPbPr+24];
	mad.f32 	%f401, %f400, %f3, %f396;
	ld.const.f32 	%f402, [kRGB32f_To_601YPbPr+12];
	mad.f32 	%f403, %f402, %f3, %f398;
	ld.const.f32 	%f404, [kRGB32f_To_601YPbPr+8];
	mad.f32 	%f33, %f404, %f1, %f31;
	ld.const.f32 	%f405, [kRGB32f_To_601YPbPr+32];
	mad.f32 	%f406, %f405, %f1, %f401;
	ld.const.f32 	%f407, [kRGB32f_To_601YPbPr+20];
	mad.f32 	%f408, %f407, %f1, %f403;
	ld.param.f32 	%f409, [__cudaparm_PointwiseFilterHostKernel1___val_paraminFilter0+4];
	ld.param.f32 	%f410, [__cudaparm_PointwiseFilterHostKernel1___val_paraminFilter0+8];
	mad.f32 	%f411, %f33, %f410, %f409;
	ld.param.f32 	%f412, [__cudaparm_PointwiseFilterHostKernel1___val_paraminFilter0+16];
	mul.f32 	%f413, %f406, %f412;
	ld.param.f32 	%f414, [__cudaparm_PointwiseFilterHostKernel1___val_paraminFilter0+12];
	mul.f32 	%f415, %f406, %f414;
	mul.f32 	%f416, %f408, %f414;
	sub.f32 	%f417, %f416, %f413;
	mad.f32 	%f418, %f408, %f412, %f415;
	ld.const.f32 	%f419, [k601YPbPr_To_RGB32f+16];
	mul.f32 	%f420, %f419, %f417;
	ld.const.f32 	%f421, [k601YPbPr_To_RGB32f+12];
	mad.f32 	%f422, %f421, %f411, %f420;
	ld.const.f32 	%f423, [k601YPbPr_To_RGB32f+20];
	mad.f32 	%f424, %f423, %f418, %f422;
	.loc	19	268	0
	ld.const.f32 	%f425, [k601YPbPr_To_RGB32f+4];
	mul.f32 	%f426, %f425, %f417;
	ld.const.f32 	%f427, [k601YPbPr_To_RGB32f+0];
	mad.f32 	%f428, %f427, %f411, %f426;
	ld.const.f32 	%f429, [k601YPbPr_To_RGB32f+8];
	mad.f32 	%f430, %f429, %f418, %f428;
	.loc	32	56	0
	ld.const.f32 	%f431, [k601YPbPr_To_RGB32f+28];
	mul.f32 	%f432, %f431, %f417;
	ld.const.f32 	%f433, [k601YPbPr_To_RGB32f+24];
	mad.f32 	%f434, %f433, %f411, %f432;
	ld.const.f32 	%f435, [k601YPbPr_To_RGB32f+32];
	mad.f32 	%f5, %f435, %f418, %f434;
	mov.f32 	%f6, %f424;
	mov.f32 	%f7, %f430;
	bra.uni 	$Lt_0_514;
$Lt_0_4354:
	.loc	33	46	0
	ld.const.f32 	%f436, [kRGB32f_To_601YPbPr+4];
	mul.f32 	%f29, %f436, %f2;
	ld.const.f32 	%f437, [kRGB32f_To_601YPbPr+0];
	mad.f32 	%f31, %f437, %f3, %f29;
	ld.const.f32 	%f438, [kRGB32f_To_601YPbPr+8];
	mad.f32 	%f33, %f438, %f1, %f31;
	ld.param.f32 	%f43, [__cudaparm_PointwiseFilterHostKernel1___val_paraminFilter0+8];
	ld.param.f32 	%f439, [__cudaparm_PointwiseFilterHostKernel1___val_paraminFilter0+20];
	sub.f32 	%f440, %f439, %f43;
	mad.f32 	%f441, %f33, %f440, %f43;
	.loc	33	47	0
	ld.param.f32 	%f442, [__cudaparm_PointwiseFilterHostKernel1___val_paraminFilter0+12];
	ld.param.f32 	%f443, [__cudaparm_PointwiseFilterHostKernel1___val_paraminFilter0+24];
	sub.f32 	%f444, %f443, %f442;
	mad.f32 	%f445, %f33, %f444, %f442;
	.loc	33	49	0
	ld.param.f32 	%f446, [__cudaparm_PointwiseFilterHostKernel1___val_paraminFilter0+28];
	ld.param.f32 	%f13, [__cudaparm_PointwiseFilterHostKernel1___val_paraminFilter0+4];
	ld.param.f32 	%f447, [__cudaparm_PointwiseFilterHostKernel1___val_paraminFilter0+16];
	sub.f32 	%f448, %f447, %f13;
	mad.f32 	%f449, %f33, %f448, %f13;
	sub.f32 	%f450, %f449, %f1;
	mad.f32 	%f5, %f446, %f450, %f1;
	.loc	33	50	0
	sub.f32 	%f451, %f441, %f2;
	mad.f32 	%f6, %f446, %f451, %f2;
	.loc	33	51	0
	sub.f32 	%f452, %f445, %f3;
	mad.f32 	%f7, %f446, %f452, %f3;
	.loc	35	103	0
	bra.uni 	$Lt_0_514;
$Lt_0_4610:
	.loc	34	41	0
	sub.s32 	%r264, %r13, %r11;
	sub.s32 	%r12, %r264, 1;
$Lt_0_514:
	.loc	17	100	0
	ld.param.s32 	%r265, [__cudaparm_PointwiseFilterHostKernel1_inDestPitch];
	mul24.lo.s32 	%r266, %r12, %r265;
	ld.param.u64 	%rd5, [__cudaparm_PointwiseFilterHostKernel1_inDestImage];
	add.s32 	%r267, %r266, %r6;
	cvt.s64.s32 	%rd6, %r267;
	mul.lo.u64 	%rd7, %rd6, 16;
	add.u64 	%rd8, %rd5, %rd7;
	st.global.v4.f32 	[%rd8+0], {%f5,%f6,%f7,%f8};
$Lt_0_52482:
	.loc	35	134	0
	exit;
$LDWend_PointwiseFilterHostKernel1:
	} // PointwiseFilterHostKernel1

	.entry PointwiseFilterHostKernel2 (
		.param .u64 __cudaparm_PointwiseFilterHostKernel2_inSrcImage,
		.param .s32 __cudaparm_PointwiseFilterHostKernel2_inSrcPitch,
		.param .u64 __cudaparm_PointwiseFilterHostKernel2_inDestImage,
		.param .s32 __cudaparm_PointwiseFilterHostKernel2_inDestPitch,
		.param .s32 __cudaparm_PointwiseFilterHostKernel2_inWidth,
		.param .s32 __cudaparm_PointwiseFilterHostKernel2_inHeight,
		.param .align 4 .b8 __cudaparm_PointwiseFilterHostKernel2___val_paraminFilter0[48],
		.param .align 4 .b8 __cudaparm_PointwiseFilterHostKernel2___val_paraminFilter1[48])
	{
	.reg .u32 %r<524>;
	.reg .u64 %rd<10>;
	.reg .f32 %f<885>;
	.reg .pred %p<117>;
	.loc	35	143	0
$LBB1_PointwiseFilterHostKernel2:
	.loc	35	145	0
	cvt.s32.u16 	%r1, %ctaid.x;
	cvt.s32.u16 	%r2, %ntid.x;
	mul24.lo.s32 	%r3, %r1, %r2;
	cvt.u32.u16 	%r4, %tid.x;
	add.u32 	%r5, %r3, %r4;
	mov.s32 	%r6, %r5;
	.loc	35	146	0
	cvt.s32.u16 	%r7, %ctaid.y;
	cvt.s32.u16 	%r8, %ntid.y;
	mul24.lo.s32 	%r9, %r7, %r8;
	cvt.u32.u16 	%r10, %tid.y;
	add.u32 	%r11, %r9, %r10;
	mov.s32 	%r12, %r11;
	ld.param.s32 	%r13, [__cudaparm_PointwiseFilterHostKernel2_inHeight];
	ld.param.s32 	%r14, [__cudaparm_PointwiseFilterHostKernel2_inWidth];
	set.gt.u32.s32 	%r15, %r13, %r11;
	neg.s32 	%r16, %r15;
	set.gt.u32.s32 	%r17, %r14, %r5;
	neg.s32 	%r18, %r17;
	and.b32 	%r19, %r16, %r18;
	mov.u32 	%r20, 0;
	setp.eq.s32 	%p1, %r19, %r20;
	@%p1 bra 	$Lt_1_103682;
	.loc	17	89	0
	ld.param.s32 	%r21, [__cudaparm_PointwiseFilterHostKernel2_inSrcPitch];
	mul24.lo.s32 	%r22, %r11, %r21;
	ld.param.u64 	%rd1, [__cudaparm_PointwiseFilterHostKernel2_inSrcImage];
	add.s32 	%r23, %r22, %r5;
	cvt.s64.s32 	%rd2, %r23;
	mul.lo.u64 	%rd3, %rd2, 16;
	add.u64 	%rd4, %rd1, %rd3;
	ld.global.v4.f32 	{%f1,%f2,%f3,%f4}, [%rd4+0];
	.loc	35	150	0
	mov.f32 	%f5, %f1;
	mov.f32 	%f6, %f2;
	mov.f32 	%f7, %f3;
	mov.f32 	%f8, %f4;
	.loc	35	54	0
	ld.param.u32 	%r24, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter0+0];
	mov.u32 	%r25, 0;
	setp.eq.s32 	%p2, %r24, %r25;
	@%p2 bra 	$Lt_1_5122;
	mov.u32 	%r26, 1;
	setp.eq.s32 	%p3, %r24, %r26;
	@%p3 bra 	$Lt_1_5634;
	mov.u32 	%r27, 2;
	setp.eq.s32 	%p4, %r24, %r27;
	@%p4 bra 	$Lt_1_5890;
	mov.u32 	%r28, 3;
	setp.eq.s32 	%p5, %r24, %r28;
	@%p5 bra 	$Lt_1_6146;
	mov.u32 	%r29, 4;
	setp.eq.s32 	%p6, %r24, %r29;
	@%p6 bra 	$Lt_1_6402;
	mov.u32 	%r30, 5;
	setp.eq.s32 	%p7, %r24, %r30;
	@%p7 bra 	$Lt_1_6658;
	mov.u32 	%r31, 6;
	setp.eq.s32 	%p8, %r24, %r31;
	@%p8 bra 	$Lt_1_6914;
	mov.u32 	%r32, 7;
	setp.eq.s32 	%p9, %r24, %r32;
	@%p9 bra 	$Lt_1_7170;
	mov.u32 	%r33, 8;
	setp.eq.s32 	%p10, %r24, %r33;
	@%p10 bra 	$Lt_1_7426;
	mov.u32 	%r34, 9;
	setp.eq.s32 	%p11, %r24, %r34;
	@%p11 bra 	$Lt_1_7682;
	mov.u32 	%r35, 10;
	setp.eq.s32 	%p12, %r24, %r35;
	@%p12 bra 	$Lt_1_7938;
	mov.u32 	%r36, 11;
	setp.eq.s32 	%p13, %r24, %r36;
	@%p13 bra 	$Lt_1_8194;
	mov.u32 	%r37, 12;
	setp.eq.s32 	%p14, %r24, %r37;
	@%p14 bra 	$Lt_1_8450;
	mov.u32 	%r38, 13;
	setp.eq.s32 	%p15, %r24, %r38;
	@%p15 bra 	$Lt_1_8706;
	mov.u32 	%r39, 14;
	setp.eq.s32 	%p16, %r24, %r39;
	@%p16 bra 	$Lt_1_8962;
	mov.u32 	%r40, 15;
	setp.eq.s32 	%p17, %r24, %r40;
	@%p17 bra 	$Lt_1_9218;
	mov.u32 	%r41, 16;
	setp.eq.s32 	%p18, %r24, %r41;
	@%p18 bra 	$Lt_1_9474;
	bra.uni 	$Lt_1_5378;
$Lt_1_5122:
	.loc	35	57	0
	ld.param.f32 	%f9, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter0+8];
	mov.f32 	%f10, 0f3f800000;    	// 1
	setp.eq.f32 	%p19, %f9, %f10;
	ld.param.f32 	%f11, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter0+12];
	mov.f32 	%f12, 0f3f800000;    	// 1
	setp.eq.f32 	%p20, %f11, %f12;
	ld.param.f32 	%f13, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter0+4];
	ld.param.f32 	%f14, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter0+16];
	mov.f32 	%f15, 0f3f800000;    	// 1
	setp.eq.f32 	%p21, %f14, %f15;
	@!%p21 bra 	$Lt_1_104450;
	.loc	18	53	0
	cvt.sat.f32.f32 	%f16, %f4;
	mul.f32 	%f17, %f16, %f13;
	selp.f32 	%f18, %f13, %f17, %p19;
	mov.f32 	%f19, 0f3f800000;    	// 1
	sub.f32 	%f20, %f19, %f18;
	selp.f32 	%f21, %f20, %f18, %p20;
	mov.f32 	%f8, %f21;
	.loc	18	57	0
	mov.f32 	%f5, %f21;
	mov.f32 	%f6, %f21;
	mov.f32 	%f7, %f21;
	bra.uni 	$Lt_1_5378;
$Lt_1_104450:
	@!%p19 bra 	$Lt_1_104962;
	.loc	18	61	0
	mov.f32 	%f22, 0f3f800000;    	// 1
	sub.f32 	%f23, %f22, %f13;
	selp.f32 	%f8, %f23, %f13, %p20;
	bra.uni 	$Lt_1_5378;
$Lt_1_104962:
	.loc	18	73	0
	cvt.sat.f32.f32 	%f24, %f4;
	mov.f32 	%f25, 0f3f800000;    	// 1
	sub.f32 	%f26, %f25, %f24;
	selp.f32 	%f27, %f26, %f24, %p20;
	mul.f32 	%f8, %f27, %f13;
	bra.uni 	$Lt_1_5378;
$Lt_1_5634:
	.loc	19	267	0
	ld.const.f32 	%f28, [kRGB32f_To_601YPbPr+4];
	mul.f32 	%f29, %f28, %f2;
	ld.const.f32 	%f30, [kRGB32f_To_601YPbPr+0];
	mad.f32 	%f31, %f30, %f3, %f29;
	ld.const.f32 	%f32, [kRGB32f_To_601YPbPr+8];
	mad.f32 	%f33, %f32, %f1, %f31;
	ld.const.f32 	%f34, [k601YPbPr_To_RGB32f+12];
	mul.f32 	%f35, %f34, %f33;
	.loc	19	268	0
	ld.const.f32 	%f36, [k601YPbPr_To_RGB32f+0];
	mul.f32 	%f37, %f36, %f33;
	.loc	20	44	0
	ld.const.f32 	%f38, [k601YPbPr_To_RGB32f+24];
	mul.f32 	%f5, %f38, %f33;
	mov.f32 	%f6, %f35;
	mov.f32 	%f7, %f37;
	mov.f32 	%f8, %f4;
	.loc	35	61	0
	bra.uni 	$Lt_1_5378;
$Lt_1_5890:
	.loc	35	63	0
	ld.param.f32 	%f39, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter0+20];
	ld.param.f32 	%f40, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter0+12];
	ld.param.f32 	%f41, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter0+4];
	mov.f32 	%f42, 0f00000000;    	// 0
	setp.neu.f32 	%p22, %f41, %f42;
	@!%p22 bra 	$Lt_1_105474;
	.loc	21	44	0
	ld.param.f32 	%f43, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter0+8];
	sub.f32 	%f44, %f3, %f40;
	mad.f32 	%f7, %f43, %f44, %f39;
	.loc	21	45	0
	sub.f32 	%f45, %f2, %f40;
	mad.f32 	%f6, %f43, %f45, %f39;
	.loc	21	46	0
	sub.f32 	%f46, %f1, %f40;
	mad.f32 	%f5, %f43, %f46, %f39;
	bra.uni 	$Lt_1_5378;
$Lt_1_105474:
	.loc	21	50	0
	ld.param.f32 	%f47, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter0+24];
	setp.gt.f32 	%p23, %f3, %f40;
	selp.f32 	%f7, %f47, %f39, %p23;
	.loc	21	51	0
	setp.gt.f32 	%p24, %f2, %f40;
	selp.f32 	%f6, %f47, %f39, %p24;
	.loc	21	52	0
	setp.gt.f32 	%p25, %f1, %f40;
	selp.f32 	%f5, %f47, %f39, %p25;
	bra.uni 	$Lt_1_5378;
$Lt_1_6146:
	.loc	22	47	0
	ld.param.f32 	%f48, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter0+4];
	mul.f32 	%f5, %f48, %f1;
	ld.param.f32 	%f49, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter0+8];
	mul.f32 	%f6, %f49, %f2;
	ld.param.f32 	%f50, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter0+12];
	mul.f32 	%f7, %f50, %f3;
	.loc	35	67	0
	bra.uni 	$Lt_1_5378;
$Lt_1_6402:
	.loc	23	48	0
	ld.param.f32 	%f51, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter0+4];
	mov.f32 	%f52, 0f00000000;    	// 0
	setp.eq.f32 	%p26, %f51, %f52;
	ld.param.f32 	%f53, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter0+8];
	mov.f32 	%f54, 0f00000000;    	// 0
	max.f32 	%f55, %f1, %f54;
	mov.f32 	%f56, 0f3f800000;    	// 1
	min.f32 	%f57, %f55, %f56;
	ld.param.f32 	%f58, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter0+12];
	sub.f32 	%f59, %f57, %f58;
	abs.f32 	%f60, %f59;
	mov.f32 	%f61, 0f00000000;    	// 0
	max.f32 	%f62, %f2, %f61;
	mov.f32 	%f63, 0f3f800000;    	// 1
	min.f32 	%f64, %f62, %f63;
	ld.param.f32 	%f65, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter0+16];
	sub.f32 	%f66, %f64, %f65;
	abs.f32 	%f67, %f66;
	mov.f32 	%f68, 0f00000000;    	// 0
	max.f32 	%f69, %f4, %f68;
	mov.f32 	%f70, 0f3f800000;    	// 1
	min.f32 	%f71, %f69, %f70;
	sub.f32 	%f72, %f71, %f4;
	abs.f32 	%f73, %f72;
	mov.f32 	%f74, 0f00000000;    	// 0
	max.f32 	%f75, %f3, %f74;
	mov.f32 	%f76, 0f3f800000;    	// 1
	min.f32 	%f77, %f75, %f76;
	ld.param.f32 	%f78, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter0+20];
	sub.f32 	%f79, %f77, %f78;
	abs.f32 	%f80, %f79;
	max.f32 	%f81, %f73, %f80;
	max.f32 	%f82, %f67, %f81;
	max.f32 	%f83, %f60, %f82;
	setp.ge.f32 	%p27, %f53, %f83;
	xor.pred 	%p28, %p26, %p27;
	@!%p28 bra 	$Lt_1_5378;
	.loc	19	267	0
	ld.const.f32 	%f84, [kRGB32f_To_601YPbPr+4];
	mul.f32 	%f29, %f84, %f2;
	ld.const.f32 	%f85, [kRGB32f_To_601YPbPr+0];
	mad.f32 	%f31, %f85, %f3, %f29;
	ld.const.f32 	%f86, [kRGB32f_To_601YPbPr+8];
	mad.f32 	%f33, %f86, %f1, %f31;
	ld.const.f32 	%f87, [k601YPbPr_To_RGB32f+12];
	mul.f32 	%f88, %f87, %f33;
	.loc	19	268	0
	ld.const.f32 	%f89, [k601YPbPr_To_RGB32f+0];
	mul.f32 	%f90, %f89, %f33;
	.loc	20	44	0
	ld.const.f32 	%f91, [k601YPbPr_To_RGB32f+24];
	mul.f32 	%f5, %f91, %f33;
	mov.f32 	%f6, %f88;
	mov.f32 	%f7, %f90;
	bra.uni 	$Lt_1_5378;
$Lt_1_6658:
	.loc	24	48	0
	ld.param.f32 	%f92, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter0+8];
	mov.f32 	%f93, 0f00000000;    	// 0
	max.f32 	%f94, %f1, %f93;
	mov.f32 	%f95, 0f3f800000;    	// 1
	min.f32 	%f96, %f94, %f95;
	ld.param.f32 	%f97, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter0+12];
	sub.f32 	%f98, %f96, %f97;
	abs.f32 	%f99, %f98;
	mov.f32 	%f100, 0f00000000;   	// 0
	max.f32 	%f101, %f2, %f100;
	mov.f32 	%f102, 0f3f800000;   	// 1
	min.f32 	%f103, %f101, %f102;
	ld.param.f32 	%f104, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter0+16];
	sub.f32 	%f105, %f103, %f104;
	abs.f32 	%f106, %f105;
	mov.f32 	%f107, 0f00000000;   	// 0
	max.f32 	%f108, %f4, %f107;
	mov.f32 	%f109, 0f3f800000;   	// 1
	min.f32 	%f110, %f108, %f109;
	sub.f32 	%f111, %f110, %f4;
	abs.f32 	%f112, %f111;
	mov.f32 	%f113, 0f00000000;   	// 0
	max.f32 	%f114, %f3, %f113;
	mov.f32 	%f115, 0f3f800000;   	// 1
	min.f32 	%f116, %f114, %f115;
	ld.param.f32 	%f117, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter0+20];
	sub.f32 	%f118, %f116, %f117;
	abs.f32 	%f119, %f118;
	max.f32 	%f120, %f112, %f119;
	max.f32 	%f121, %f106, %f120;
	max.f32 	%f122, %f99, %f121;
	setp.ge.f32 	%p29, %f92, %f122;
	@!%p29 bra 	$Lt_1_5378;
	.loc	24	51	0
	ld.param.f32 	%f123, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter0+24];
	mov.f32 	%f124, %f123;
	ld.param.f32 	%f125, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter0+28];
	mov.f32 	%f126, %f125;
	ld.param.f32 	%f127, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter0+32];
	mov.f32 	%f128, %f127;
	ld.param.f32 	%f129, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter0+4];
	mov.f32 	%f130, 0f00000000;   	// 0
	setp.eq.f32 	%p30, %f129, %f130;
	@!%p30 bra 	$Lt_1_106754;
	.loc	24	60	0
	ld.const.f32 	%f131, [kRGB32f_To_601YPbPr+4];
	mul.f32 	%f29, %f131, %f2;
	ld.const.f32 	%f132, [kRGB32f_To_601YPbPr+0];
	mad.f32 	%f31, %f132, %f3, %f29;
	ld.const.f32 	%f133, [kRGB32f_To_601YPbPr+8];
	mad.f32 	%f33, %f133, %f1, %f31;
	mul.f32 	%f124, %f33, %f123;
	.loc	24	61	0
	mul.f32 	%f126, %f33, %f125;
	.loc	24	62	0
	mul.f32 	%f128, %f33, %f127;
$Lt_1_106754:
	.loc	24	65	0
	mov.f32 	%f5, %f124;
	mov.f32 	%f6, %f126;
	mov.f32 	%f7, %f128;
	bra.uni 	$Lt_1_5378;
$Lt_1_6914:
	.loc	25	47	0
	sub.s32 	%r42, %r13, %r11;
	sub.s32 	%r43, %r14, %r5;
	cvt.rn.f32.s32 	%f134, %r5;
	cvt.rn.f32.s32 	%f135, %r11;
	cvt.rn.f32.s32 	%f136, %r42;
	cvt.rn.f32.s32 	%f137, %r43;
	ld.param.f32 	%f13, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter0+4];
	sub.f32 	%f138, %f13, %f134;
	sub.f32 	%f139, %f13, %f135;
	sub.f32 	%f140, %f13, %f136;
	sub.f32 	%f141, %f13, %f137;
	cvt.rzi.s32.f32 	%r44, %f138;
	cvt.rzi.s32.f32 	%r45, %f139;
	cvt.rzi.s32.f32 	%r46, %f140;
	cvt.rzi.s32.f32 	%r47, %f141;
	max.s32 	%r48, %r44, %r45;
	max.s32 	%r49, %r47, %r48;
	max.s32 	%r50, %r46, %r49;
	mov.u32 	%r51, 0;
	setp.le.s32 	%p31, %r50, %r51;
	@%p31 bra 	$Lt_1_5378;
	.loc	25	51	0
	cvt.rn.f32.s32 	%f142, %r50;
	sub.f32 	%f143, %f13, %f142;
	div.full.f32 	%f144, %f143, %f13;
	mul.f32 	%f8, %f4, %f144;
	bra.uni 	$Lt_1_5378;
$Lt_1_7170:
	.loc	27	50	0
	ld.const.f32 	%f145, [kRGB32f_To_601YPbPr+4];
	mul.f32 	%f29, %f145, %f2;
	ld.param.f32 	%f146, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter0+12];
	ld.param.f32 	%f43, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter0+8];
	sub.f32 	%f147, %f146, %f43;
	ld.const.f32 	%f148, [kRGB32f_To_601YPbPr+0];
	mad.f32 	%f31, %f148, %f3, %f29;
	ld.param.f32 	%f149, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter0+16];
	mul.f32 	%f150, %f149, %f147;
	ld.const.f32 	%f151, [kRGB32f_To_601YPbPr+8];
	mad.f32 	%f33, %f151, %f1, %f31;
	add.f32 	%f152, %f150, %f146;
	sub.f32 	%f153, %f43, %f150;
	mov.f32 	%f154, 0f00000000;   	// 0
	max.f32 	%f155, %f152, %f154;
	mov.f32 	%f156, 0f00000000;   	// 0
	max.f32 	%f157, %f153, %f156;
	mov.f32 	%f158, 0f3f800000;   	// 1
	min.f32 	%f159, %f155, %f158;
	mov.f32 	%f160, 0f3f800000;   	// 1
	min.f32 	%f161, %f157, %f160;
	set.gt.u32.f32 	%r52, %f161, %f33;
	neg.s32 	%r53, %r52;
	set.le.u32.f32 	%r54, %f159, %f33;
	neg.s32 	%r55, %r54;
	or.b32 	%r56, %r53, %r55;
	mov.u32 	%r57, 0;
	setp.eq.s32 	%p32, %r56, %r57;
	@%p32 bra 	$Lt_1_108034;
	mov.f32 	%f162, 0f00000000;   	// 0
	bra.uni 	$Lt_1_108802;
$Lt_1_108034:
	add.f32 	%f163, %f150, %f43;
	mov.f32 	%f164, 0f00000000;   	// 0
	max.f32 	%f165, %f163, %f164;
	mov.f32 	%f166, 0f3f800000;   	// 1
	min.f32 	%f167, %f165, %f166;
	set.le.u32.f32 	%r58, %f167, %f33;
	neg.s32 	%r59, %r58;
	sub.f32 	%f168, %f146, %f150;
	mov.f32 	%f169, 0f00000000;   	// 0
	max.f32 	%f170, %f168, %f169;
	mov.f32 	%f171, 0f3f800000;   	// 1
	min.f32 	%f172, %f170, %f171;
	set.lt.u32.f32 	%r60, %f33, %f172;
	neg.s32 	%r61, %r60;
	and.b32 	%r62, %r59, %r61;
	mov.u32 	%r63, 0;
	setp.eq.s32 	%p33, %r62, %r63;
	@%p33 bra 	$Lt_1_108546;
	mov.f32 	%f162, 0f3f800000;   	// 1
	bra.uni 	$Lt_1_108802;
$Lt_1_108546:
	add.f32 	%f173, %f150, %f150;
	setp.gt.f32 	%p34, %f167, %f33;
	@!%p34 bra 	$Lt_1_109058;
	.loc	27	62	0
	sub.f32 	%f174, %f33, %f161;
	div.full.f32 	%f162, %f174, %f173;
	bra.uni 	$Lt_1_108802;
$Lt_1_109058:
	.loc	27	66	0
	sub.f32 	%f175, %f159, %f33;
	div.full.f32 	%f162, %f175, %f173;
$Lt_1_108802:
$Lt_1_108290:
$Lt_1_107778:
	.loc	27	69	0
	mov.f32 	%f176, 0f3f800000;   	// 1
	sub.f32 	%f177, %f176, %f162;
	ld.param.f32 	%f178, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter0+4];
	mov.f32 	%f179, 0f00000000;   	// 0
	setp.neu.f32 	%p35, %f178, %f179;
	selp.f32 	%f162, %f177, %f162, %p35;
	.loc	27	77	0
	ld.const.f32 	%f180, [k601YPbPr_To_RGB32f+24];
	mul.f32 	%f5, %f180, %f162;
	ld.const.f32 	%f181, [k601YPbPr_To_RGB32f+12];
	mul.f32 	%f6, %f181, %f162;
	ld.const.f32 	%f182, [k601YPbPr_To_RGB32f+0];
	mul.f32 	%f7, %f182, %f162;
	.loc	35	79	0
	bra.uni 	$Lt_1_5378;
$Lt_1_7426:
	.loc	35	80	0
	ld.param.f32 	%f183, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter0+44];
	mov.f32 	%f184, 0f00000000;   	// 0
	setp.ge.f32 	%p36, %f183, %f184;
	@!%p36 bra 	$L_1_99586;
	cvt.rn.f32.s32 	%f185, %r5;
	cvt.rn.f32.s32 	%f186, %r14;
	mul.f32 	%f187, %f186, %f183;
	setp.lt.f32 	%p37, %f185, %f187;
	@%p37 bra 	$L_1_99330;
$L_1_99586:
	mov.f32 	%f188, 0f00000000;   	// 0
	setp.lt.f32 	%p38, %f183, %f188;
	@!%p38 bra 	$Lt_1_5378;
	cvt.rn.f32.s32 	%f189, %r11;
	cvt.rn.f32.s32 	%f190, %r13;
	mul.f32 	%f191, %f190, %f183;
	neg.f32 	%f192, %f191;
	setp.lt.f32 	%p39, %f189, %f192;
	@!%p39 bra 	$Lt_1_5378;
$L_1_99330:
	.loc	28	47	0
	ld.const.f32 	%f193, [kRGB32f_To_601YPbPr+4];
	mul.f32 	%f194, %f193, %f2;
	ld.const.f32 	%f195, [kRGB32f_To_601YPbPr+0];
	mad.f32 	%f196, %f195, %f3, %f194;
	ld.const.f32 	%f197, [kRGB32f_To_601YPbPr+8];
	mad.f32 	%f33, %f197, %f1, %f196;
	mov.f32 	%f198, %f33;
	ld.param.f32 	%f43, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter0+8];
	ld.param.f32 	%f199, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter0+4];
	sub.f32 	%f200, %f199, %f43;
	ld.param.f32 	%f201, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter0+16];
	ld.param.f32 	%f202, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter0+12];
	sub.f32 	%f203, %f202, %f201;
	mov.f32 	%f204, 0f3f800000;   	// 1
	set.neu.u32.f32 	%r64, %f203, %f204;
	neg.s32 	%r65, %r64;
	mov.f32 	%f205, 0f3f800000;   	// 1
	set.neu.u32.f32 	%r66, %f200, %f205;
	neg.s32 	%r67, %r66;
	or.b32 	%r68, %r65, %r67;
	mov.u32 	%r69, 0;
	setp.eq.s32 	%p40, %r68, %r69;
	@%p40 bra 	$Lt_1_109314;
	.loc	17	116	0
	mov.u32 	%r70, 1;
	sub.u32 	%r71, %r70, %r5;
	shr.u32 	%r72, %r11, 13;
	sub.u32 	%r73, %r5, %r11;
	sub.u32 	%r74, %r71, %r11;
	xor.b32 	%r75, %r74, %r72;
	shl.b32 	%r76, %r75, 8;
	sub.u32 	%r77, %r73, %r75;
	sub.u32 	%r78, %r11, %r75;
	xor.b32 	%r79, %r77, %r76;
	shr.u32 	%r80, %r79, 13;
	sub.u32 	%r81, %r78, %r79;
	xor.b32 	%r82, %r81, %r80;
	shr.u32 	%r83, %r82, 12;
	sub.u32 	%r84, %r75, %r79;
	sub.u32 	%r85, %r84, %r82;
	xor.b32 	%r86, %r83, %r85;
	.loc	17	117	0
	sub.u32 	%r87, %r79, %r82;
	sub.u32 	%r88, %r87, %r86;
	shl.b32 	%r89, %r86, 16;
	xor.b32 	%r90, %r88, %r89;
	.loc	17	118	0
	sub.u32 	%r91, %r82, %r86;
	sub.u32 	%r92, %r91, %r90;
	shr.u32 	%r93, %r90, 5;
	xor.b32 	%r94, %r92, %r93;
	.loc	17	119	0
	sub.u32 	%r95, %r86, %r90;
	sub.u32 	%r96, %r95, %r94;
	shr.u32 	%r97, %r94, 3;
	xor.b32 	%r98, %r96, %r97;
	.loc	17	120	0
	sub.u32 	%r99, %r90, %r94;
	sub.u32 	%r100, %r99, %r98;
	shl.b32 	%r101, %r98, 10;
	xor.b32 	%r102, %r100, %r101;
	.loc	17	121	0
	sub.u32 	%r103, %r94, %r98;
	sub.u32 	%r104, %r103, %r102;
	shr.u32 	%r105, %r102, 15;
	xor.b32 	%r106, %r104, %r105;
	.loc	28	57	0
	mov.f32 	%f206, 0f3b270d73;   	// 0.00254902
	mul.lo.u32 	%r107, %r106, 1103515245;
	add.u32 	%r108, %r107, 12345;
	shr.u32 	%r109, %r108, 16;
	and.b32 	%r110, %r109, 255;
	shl.b32 	%r111, %r110, 7;
	mul.lo.u32 	%r112, %r106, -1029531031;
	sub.u32 	%r113, %r112, 740551042;
	shr.u32 	%r114, %r113, 16;
	and.b32 	%r115, %r114, 255;
	xor.b32 	%r116, %r111, %r115;
	cvt.rn.f32.s32 	%f207, %r116;
	mov.f32 	%f208, 0f467ffe00;   	// 16383.5
	div.full.f32 	%f209, %f207, %f208;
	mov.f32 	%f210, 0fbf800000;   	// -1
	add.f32 	%f211, %f209, %f210;
	mad.f32 	%f198, %f206, %f211, %f33;
$Lt_1_109314:
	sub.f32 	%f212, %f198, %f43;
	ld.param.f32 	%f213, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter0+20];
	mov.f32 	%f214, 0f3f800000;   	// 1
	setp.neu.f32 	%p41, %f213, %f214;
	@!%p41 bra 	$Lt_1_110082;
	mov.f32 	%f215, 0f00000000;   	// 0
	setp.lt.f32 	%p42, %f212, %f215;
	@!%p42 bra 	$Lt_1_110594;
	.loc	28	66	0
	mov.f32 	%f198, %f201;
	bra.uni 	$Lt_1_109826;
$Lt_1_110594:
	.loc	28	70	0
	rcp.approx.f32 	%f216, %f200;
	mul.f32 	%f217, %f212, %f216;
	lg2.approx.f32 	%f218, %f217;
	mul.f32 	%f219, %f213, %f218;
	ex2.approx.f32 	%f220, %f219;
	mad.f32 	%f198, %f203, %f220, %f201;
	bra.uni 	$Lt_1_109826;
$Lt_1_110082:
	.loc	28	77	0
	rcp.approx.f32 	%f221, %f200;
	mul.f32 	%f222, %f212, %f221;
	mad.f32 	%f198, %f203, %f222, %f201;
$Lt_1_109826:
	.loc	19	267	0
	ld.const.f32 	%f223, [kRGB32f_To_601YPbPr+28];
	mul.f32 	%f224, %f223, %f2;
	ld.const.f32 	%f225, [kRGB32f_To_601YPbPr+16];
	mul.f32 	%f226, %f225, %f2;
	ld.param.f32 	%f227, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter0+28];
	sin.approx.f32 	%f228, %f227;
	cos.approx.f32 	%f229, %f227;
	ld.const.f32 	%f230, [kRGB32f_To_601YPbPr+24];
	mad.f32 	%f231, %f230, %f3, %f224;
	ld.const.f32 	%f232, [kRGB32f_To_601YPbPr+12];
	mad.f32 	%f233, %f232, %f3, %f226;
	ld.const.f32 	%f234, [kRGB32f_To_601YPbPr+32];
	mad.f32 	%f235, %f234, %f1, %f231;
	ld.const.f32 	%f236, [kRGB32f_To_601YPbPr+20];
	mad.f32 	%f237, %f236, %f1, %f233;
	mul.f32 	%f238, %f228, %f235;
	mul.f32 	%f239, %f229, %f235;
	mul.f32 	%f240, %f237, %f229;
	sub.f32 	%f241, %f240, %f238;
	mad.f32 	%f242, %f237, %f228, %f239;
	ld.param.f32 	%f243, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter0+32];
	sub.f32 	%f244, %f243, %f241;
	ld.param.f32 	%f245, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter0+36];
	sub.f32 	%f246, %f245, %f242;
	ld.param.f32 	%f247, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter0+40];
	mad.f32 	%f248, %f247, %f244, %f241;
	mad.f32 	%f249, %f247, %f246, %f242;
	ld.param.f32 	%f250, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter0+24];
	mul.f32 	%f251, %f248, %f250;
	mul.f32 	%f252, %f249, %f250;
	ld.const.f32 	%f253, [k601YPbPr_To_RGB32f+16];
	mul.f32 	%f254, %f253, %f251;
	ld.const.f32 	%f255, [k601YPbPr_To_RGB32f+12];
	mad.f32 	%f256, %f255, %f198, %f254;
	ld.const.f32 	%f257, [k601YPbPr_To_RGB32f+20];
	mad.f32 	%f258, %f257, %f252, %f256;
	.loc	19	268	0
	ld.const.f32 	%f259, [k601YPbPr_To_RGB32f+4];
	mul.f32 	%f260, %f259, %f251;
	ld.const.f32 	%f261, [k601YPbPr_To_RGB32f+0];
	mad.f32 	%f262, %f261, %f198, %f260;
	ld.const.f32 	%f263, [k601YPbPr_To_RGB32f+8];
	mad.f32 	%f264, %f263, %f252, %f262;
	.loc	28	92	0
	ld.const.f32 	%f265, [k601YPbPr_To_RGB32f+28];
	mul.f32 	%f266, %f265, %f251;
	ld.const.f32 	%f267, [k601YPbPr_To_RGB32f+24];
	mad.f32 	%f268, %f267, %f198, %f266;
	ld.const.f32 	%f269, [k601YPbPr_To_RGB32f+32];
	mad.f32 	%f5, %f269, %f252, %f268;
	mov.f32 	%f6, %f258;
	mov.f32 	%f7, %f264;
	bra.uni 	$Lt_1_5378;
$Lt_1_7682:
	.loc	35	83	0
	ld.param.f32 	%f270, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter0+44];
	mov.f32 	%f271, 0f00000000;   	// 0
	setp.ge.f32 	%p43, %f270, %f271;
	@!%p43 bra 	$L_1_100354;
	cvt.rn.f32.s32 	%f272, %r5;
	cvt.rn.f32.s32 	%f273, %r14;
	mul.f32 	%f274, %f273, %f270;
	setp.lt.f32 	%p44, %f272, %f274;
	@%p44 bra 	$L_1_100098;
$L_1_100354:
	mov.f32 	%f275, 0f00000000;   	// 0
	setp.lt.f32 	%p45, %f270, %f275;
	@!%p45 bra 	$Lt_1_5378;
	cvt.rn.f32.s32 	%f276, %r11;
	cvt.rn.f32.s32 	%f277, %r13;
	mul.f32 	%f278, %f277, %f270;
	neg.f32 	%f279, %f278;
	setp.lt.f32 	%p46, %f276, %f279;
	@!%p46 bra 	$Lt_1_5378;
$L_1_100098:
	.loc	28	110	0
	ld.const.f32 	%f280, [kRGB32f_To_601YPbPr+4];
	mul.f32 	%f29, %f280, %f2;
	ld.param.f32 	%f43, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter0+8];
	ld.param.f32 	%f281, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter0+4];
	sub.f32 	%f282, %f281, %f43;
	ld.param.f32 	%f283, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter0+16];
	ld.param.f32 	%f284, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter0+12];
	sub.f32 	%f285, %f284, %f283;
	ld.const.f32 	%f286, [kRGB32f_To_601YPbPr+0];
	mad.f32 	%f31, %f286, %f3, %f29;
	ld.const.f32 	%f287, [kRGB32f_To_601YPbPr+8];
	mad.f32 	%f33, %f287, %f1, %f31;
	sub.f32 	%f288, %f33, %f43;
	ld.param.f32 	%f289, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter0+20];
	mov.f32 	%f290, 0f3f800000;   	// 1
	setp.neu.f32 	%p47, %f289, %f290;
	@!%p47 bra 	$Lt_1_111106;
	.loc	28	120	0
	mov.f32 	%f291, 0f00000000;   	// 0
	max.f32 	%f292, %f288, %f291;
	div.full.f32 	%f293, %f292, %f282;
	lg2.approx.f32 	%f294, %f293;
	mul.f32 	%f295, %f289, %f294;
	ex2.approx.f32 	%f296, %f295;
	mad.f32 	%f297, %f285, %f296, %f283;
	bra.uni 	$Lt_1_110850;
$Lt_1_111106:
	.loc	28	129	0
	div.full.f32 	%f298, %f288, %f282;
	mad.f32 	%f297, %f285, %f298, %f283;
$Lt_1_110850:
	.loc	28	135	0
	ld.const.f32 	%f299, [k601YPbPr_To_RGB32f+24];
	mul.f32 	%f5, %f299, %f297;
	ld.const.f32 	%f300, [k601YPbPr_To_RGB32f+12];
	mul.f32 	%f6, %f300, %f297;
	ld.const.f32 	%f301, [k601YPbPr_To_RGB32f+0];
	mul.f32 	%f7, %f301, %f297;
	bra.uni 	$Lt_1_5378;
$Lt_1_7938:
	.loc	35	86	0
	ld.param.f32 	%f302, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter0+44];
	mov.f32 	%f303, 0f00000000;   	// 0
	setp.ge.f32 	%p48, %f302, %f303;
	@!%p48 bra 	$L_1_101122;
	cvt.rn.f32.s32 	%f304, %r5;
	cvt.rn.f32.s32 	%f305, %r14;
	mul.f32 	%f306, %f305, %f302;
	setp.lt.f32 	%p49, %f304, %f306;
	@%p49 bra 	$L_1_100866;
$L_1_101122:
	mov.f32 	%f307, 0f00000000;   	// 0
	setp.lt.f32 	%p50, %f302, %f307;
	@!%p50 bra 	$Lt_1_5378;
	cvt.rn.f32.s32 	%f308, %r11;
	cvt.rn.f32.s32 	%f309, %r13;
	mul.f32 	%f310, %f309, %f302;
	neg.f32 	%f311, %f310;
	setp.lt.f32 	%p51, %f308, %f311;
	@!%p51 bra 	$Lt_1_5378;
$L_1_100866:
	.loc	28	160	0
	ld.const.f32 	%f5, [k601YPbPr_To_RGB32f+24];
	ld.const.f32 	%f6, [k601YPbPr_To_RGB32f+12];
	ld.const.f32 	%f7, [k601YPbPr_To_RGB32f+0];
	bra.uni 	$Lt_1_5378;
$Lt_1_8194:
	.loc	29	42	0
	ld.param.f32 	%f13, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter0+4];
	mov.f32 	%f312, 0f00000000;   	// 0
	setp.ge.f32 	%p52, %f1, %f312;
	@!%p52 bra 	$Lt_1_111618;
	.loc	19	292	0
	lg2.approx.f32 	%f313, %f1;
	mul.f32 	%f314, %f13, %f313;
	ex2.approx.f32 	%f315, %f314;
	bra.uni 	$Lt_1_111362;
$Lt_1_111618:
	neg.f32 	%f316, %f1;
	lg2.approx.f32 	%f317, %f316;
	mul.f32 	%f318, %f13, %f317;
	ex2.approx.f32 	%f319, %f318;
	neg.f32 	%f315, %f319;
$Lt_1_111362:
	mov.f32 	%f320, 0f00000000;   	// 0
	setp.ge.f32 	%p53, %f2, %f320;
	@!%p53 bra 	$Lt_1_112130;
	.loc	19	293	0
	lg2.approx.f32 	%f321, %f2;
	mul.f32 	%f322, %f13, %f321;
	ex2.approx.f32 	%f323, %f322;
	bra.uni 	$Lt_1_111874;
$Lt_1_112130:
	neg.f32 	%f324, %f2;
	lg2.approx.f32 	%f325, %f324;
	mul.f32 	%f326, %f13, %f325;
	ex2.approx.f32 	%f327, %f326;
	neg.f32 	%f323, %f327;
$Lt_1_111874:
	mov.f32 	%f328, 0f00000000;   	// 0
	setp.ge.f32 	%p54, %f3, %f328;
	@!%p54 bra 	$Lt_1_112642;
	.loc	19	294	0
	lg2.approx.f32 	%f329, %f3;
	mul.f32 	%f330, %f13, %f329;
	ex2.approx.f32 	%f331, %f330;
	bra.uni 	$Lt_1_112386;
$Lt_1_112642:
	neg.f32 	%f332, %f3;
	lg2.approx.f32 	%f333, %f332;
	mul.f32 	%f334, %f13, %f333;
	ex2.approx.f32 	%f335, %f334;
	neg.f32 	%f331, %f335;
$Lt_1_112386:
	mov.f32 	%f336, 0f00000000;   	// 0
	setp.ge.f32 	%p55, %f4, %f336;
	@!%p55 bra 	$Lt_1_113154;
	.loc	19	295	0
	lg2.approx.f32 	%f337, %f4;
	mul.f32 	%f338, %f13, %f337;
	ex2.approx.f32 	%f339, %f338;
	bra.uni 	$Lt_1_112898;
$Lt_1_113154:
	neg.f32 	%f340, %f4;
	lg2.approx.f32 	%f341, %f340;
	mul.f32 	%f342, %f13, %f341;
	ex2.approx.f32 	%f343, %f342;
	neg.f32 	%f339, %f343;
$Lt_1_112898:
	.loc	29	42	0
	mov.f32 	%f5, %f315;
	mov.f32 	%f6, %f323;
	mov.f32 	%f7, %f331;
	mov.f32 	%f8, %f339;
	.loc	35	91	0
	bra.uni 	$Lt_1_5378;
$Lt_1_8450:
	.loc	30	41	0
	sub.s32 	%r117, %r14, %r5;
	sub.s32 	%r6, %r117, 1;
	.loc	35	94	0
	bra.uni 	$Lt_1_5378;
$Lt_1_8706:
	.loc	35	96	0
	sub.u32 	%r73, %r5, %r11;
	ld.param.f32 	%f13, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter0+4];
	mov.f32 	%f344, 0f3e800000;   	// 0.25
	mul.f32 	%f345, %f13, %f344;
	ld.param.f32 	%f346, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter0+16];
	ld.param.f32 	%f347, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter0+8];
	mov.f32 	%f348, 0f3f800000;   	// 1
	setp.eq.f32 	%p56, %f347, %f348;
	@!%p56 bra 	$Lt_1_113666;
	.loc	17	117	0
	cvt.rzi.u32.f32 	%r118, %f346;
	shr.u32 	%r119, %r118, 13;
	sub.u32 	%r120, %r73, %r118;
	sub.u32 	%r121, %r11, %r118;
	xor.b32 	%r122, %r120, %r119;
	shl.b32 	%r123, %r122, 8;
	sub.u32 	%r124, %r121, %r122;
	sub.u32 	%r125, %r118, %r122;
	xor.b32 	%r126, %r124, %r123;
	shr.u32 	%r127, %r126, 13;
	sub.u32 	%r128, %r125, %r126;
	sub.u32 	%r129, %r122, %r126;
	xor.b32 	%r130, %r128, %r127;
	shr.u32 	%r131, %r130, 12;
	sub.u32 	%r132, %r129, %r130;
	xor.b32 	%r133, %r132, %r131;
	shl.b32 	%r134, %r133, 16;
	sub.u32 	%r135, %r126, %r130;
	sub.u32 	%r136, %r135, %r133;
	xor.b32 	%r137, %r134, %r136;
	.loc	17	118	0
	sub.u32 	%r138, %r130, %r133;
	sub.u32 	%r139, %r138, %r137;
	shr.u32 	%r140, %r137, 5;
	xor.b32 	%r141, %r139, %r140;
	.loc	17	119	0
	sub.u32 	%r142, %r133, %r137;
	sub.u32 	%r143, %r142, %r141;
	shr.u32 	%r144, %r141, 3;
	xor.b32 	%r145, %r143, %r144;
	.loc	17	120	0
	sub.u32 	%r146, %r137, %r141;
	sub.u32 	%r147, %r146, %r145;
	shl.b32 	%r148, %r145, 10;
	xor.b32 	%r149, %r147, %r148;
	.loc	17	121	0
	sub.u32 	%r150, %r141, %r145;
	sub.u32 	%r151, %r150, %r149;
	shr.u32 	%r152, %r149, 15;
	xor.b32 	%r153, %r151, %r152;
	.loc	31	48	0
	mul.lo.u32 	%r154, %r153, 1103515245;
	add.u32 	%r155, %r154, 12345;
	shr.u32 	%r156, %r155, 16;
	and.b32 	%r157, %r156, 255;
	shl.b32 	%r158, %r157, 7;
	mul.lo.u32 	%r159, %r153, -1029531031;
	sub.u32 	%r160, %r159, 740551042;
	shr.u32 	%r161, %r160, 16;
	and.b32 	%r162, %r161, 255;
	xor.b32 	%r163, %r158, %r162;
	cvt.rn.f32.s32 	%f349, %r163;
	mov.f32 	%f350, 0f46fffe00;   	// 32767
	div.full.f32 	%f351, %f349, %f350;
	mul.f32 	%f352, %f13, %f351;
	mov.f32 	%f353, 0f3f000000;   	// 0.5
	mul.f32 	%f354, %f352, %f353;
	sub.f32 	%f355, %f354, %f345;
	.loc	17	117	0
	add.u32 	%r164, %r120, 1;
	xor.b32 	%r165, %r164, %r119;
	shl.b32 	%r166, %r165, 8;
	sub.u32 	%r167, %r121, %r165;
	sub.u32 	%r168, %r118, %r165;
	xor.b32 	%r169, %r166, %r167;
	shr.u32 	%r170, %r169, 13;
	sub.u32 	%r171, %r168, %r169;
	sub.u32 	%r172, %r165, %r169;
	xor.b32 	%r173, %r170, %r171;
	shr.u32 	%r174, %r173, 12;
	sub.u32 	%r175, %r172, %r173;
	xor.b32 	%r176, %r174, %r175;
	sub.u32 	%r177, %r169, %r173;
	sub.u32 	%r178, %r177, %r176;
	shl.b32 	%r179, %r176, 16;
	xor.b32 	%r178, %r178, %r179;
	.loc	17	118	0
	sub.u32 	%r180, %r173, %r176;
	sub.u32 	%r180, %r180, %r178;
	shr.u32 	%r181, %r178, 5;
	xor.b32 	%r180, %r180, %r181;
	.loc	17	119	0
	sub.u32 	%r182, %r176, %r178;
	sub.u32 	%r182, %r182, %r180;
	shr.u32 	%r183, %r180, 3;
	xor.b32 	%r182, %r182, %r183;
	.loc	17	120	0
	sub.u32 	%r178, %r178, %r180;
	sub.u32 	%r178, %r178, %r182;
	shl.b32 	%r184, %r182, 10;
	xor.b32 	%r178, %r178, %r184;
	.loc	17	121	0
	sub.u32 	%r180, %r180, %r182;
	sub.u32 	%r180, %r180, %r178;
	shr.u32 	%r185, %r178, 15;
	xor.b32 	%r180, %r180, %r185;
	.loc	31	49	0
	mul.lo.u32 	%r186, %r180, 1103515245;
	add.u32 	%r187, %r186, 12345;
	shr.u32 	%r188, %r187, 16;
	and.b32 	%r189, %r188, 255;
	shl.b32 	%r190, %r189, 7;
	mul.lo.u32 	%r191, %r180, -1029531031;
	sub.u32 	%r192, %r191, 740551042;
	shr.u32 	%r193, %r192, 16;
	and.b32 	%r194, %r193, 255;
	xor.b32 	%r195, %r190, %r194;
	cvt.rn.f32.s32 	%f356, %r195;
	mov.f32 	%f357, 0f46fffe00;   	// 32767
	div.full.f32 	%f358, %f356, %f357;
	mul.f32 	%f359, %f13, %f358;
	mov.f32 	%f360, 0f3f000000;   	// 0.5
	mul.f32 	%f361, %f359, %f360;
	sub.f32 	%f362, %f361, %f345;
	.loc	17	117	0
	add.u32 	%r196, %r120, 2;
	xor.b32 	%r197, %r196, %r119;
	shl.b32 	%r198, %r197, 8;
	sub.u32 	%r199, %r121, %r197;
	sub.u32 	%r200, %r118, %r197;
	xor.b32 	%r201, %r198, %r199;
	shr.u32 	%r202, %r201, 13;
	sub.u32 	%r203, %r200, %r201;
	sub.u32 	%r204, %r197, %r201;
	xor.b32 	%r205, %r202, %r203;
	shr.u32 	%r206, %r205, 12;
	sub.u32 	%r207, %r204, %r205;
	xor.b32 	%r208, %r206, %r207;
	sub.u32 	%r209, %r201, %r205;
	sub.u32 	%r210, %r209, %r208;
	shl.b32 	%r211, %r208, 16;
	xor.b32 	%r210, %r210, %r211;
	.loc	17	118	0
	sub.u32 	%r212, %r205, %r208;
	sub.u32 	%r212, %r212, %r210;
	shr.u32 	%r213, %r210, 5;
	xor.b32 	%r212, %r212, %r213;
	.loc	17	119	0
	sub.u32 	%r214, %r208, %r210;
	sub.u32 	%r214, %r214, %r212;
	shr.u32 	%r215, %r212, 3;
	xor.b32 	%r214, %r214, %r215;
	.loc	17	120	0
	sub.u32 	%r210, %r210, %r212;
	sub.u32 	%r210, %r210, %r214;
	shl.b32 	%r216, %r214, 10;
	xor.b32 	%r210, %r210, %r216;
	.loc	17	121	0
	sub.u32 	%r212, %r212, %r214;
	sub.u32 	%r212, %r212, %r210;
	shr.u32 	%r217, %r210, 15;
	xor.b32 	%r212, %r212, %r217;
	.loc	31	50	0
	mul.lo.u32 	%r218, %r212, 1103515245;
	add.u32 	%r219, %r218, 12345;
	shr.u32 	%r220, %r219, 16;
	and.b32 	%r221, %r220, 255;
	shl.b32 	%r222, %r221, 7;
	mul.lo.u32 	%r223, %r212, -1029531031;
	sub.u32 	%r224, %r223, 740551042;
	shr.u32 	%r225, %r224, 16;
	and.b32 	%r226, %r225, 255;
	xor.b32 	%r227, %r222, %r226;
	cvt.rn.f32.s32 	%f363, %r227;
	mov.f32 	%f364, 0f46fffe00;   	// 32767
	div.full.f32 	%f365, %f363, %f364;
	mul.f32 	%f366, %f13, %f365;
	mov.f32 	%f367, 0f3f000000;   	// 0.5
	mul.f32 	%f368, %f366, %f367;
	sub.f32 	%f369, %f368, %f345;
	bra.uni 	$Lt_1_113410;
$Lt_1_113666:
	.loc	17	117	0
	cvt.rzi.u32.f32 	%r228, %f346;
	shr.u32 	%r229, %r228, 13;
	sub.u32 	%r230, %r73, %r228;
	sub.u32 	%r231, %r11, %r228;
	xor.b32 	%r232, %r230, %r229;
	shl.b32 	%r233, %r232, 8;
	sub.u32 	%r234, %r231, %r232;
	sub.u32 	%r235, %r228, %r232;
	xor.b32 	%r236, %r234, %r233;
	shr.u32 	%r237, %r236, 13;
	sub.u32 	%r238, %r235, %r236;
	sub.u32 	%r239, %r232, %r236;
	xor.b32 	%r240, %r238, %r237;
	shr.u32 	%r241, %r240, 12;
	sub.u32 	%r242, %r239, %r240;
	xor.b32 	%r243, %r242, %r241;
	shl.b32 	%r244, %r243, 16;
	sub.u32 	%r245, %r236, %r240;
	sub.u32 	%r246, %r245, %r243;
	xor.b32 	%r247, %r244, %r246;
	.loc	17	118	0
	sub.u32 	%r248, %r240, %r243;
	sub.u32 	%r249, %r248, %r247;
	shr.u32 	%r250, %r247, 5;
	xor.b32 	%r251, %r249, %r250;
	.loc	17	119	0
	sub.u32 	%r252, %r243, %r247;
	sub.u32 	%r253, %r252, %r251;
	shr.u32 	%r254, %r251, 3;
	xor.b32 	%r255, %r253, %r254;
	.loc	17	120	0
	sub.u32 	%r256, %r247, %r251;
	sub.u32 	%r257, %r256, %r255;
	shl.b32 	%r258, %r255, 10;
	xor.b32 	%r259, %r257, %r258;
	.loc	17	121	0
	sub.u32 	%r260, %r251, %r255;
	sub.u32 	%r261, %r260, %r259;
	shr.u32 	%r262, %r259, 15;
	xor.b32 	%r263, %r261, %r262;
	.loc	31	54	0
	mul.lo.u32 	%r264, %r263, 1103515245;
	mul.lo.u32 	%r265, %r263, -1029531031;
	add.u32 	%r266, %r264, 12345;
	sub.u32 	%r267, %r265, 740551042;
	shr.u32 	%r268, %r266, 16;
	shr.u32 	%r269, %r267, 16;
	and.b32 	%r270, %r268, 255;
	and.b32 	%r271, %r269, 255;
	shl.b32 	%r272, %r270, 7;
	xor.b32 	%r273, %r272, %r271;
	cvt.rn.f32.s32 	%f370, %r273;
	mov.f32 	%f371, 0f46fffe00;   	// 32767
	div.full.f32 	%f372, %f370, %f371;
	mul.f32 	%f373, %f372, %f13;
	mov.f32 	%f374, 0f3f000000;   	// 0.5
	mul.f32 	%f375, %f373, %f374;
	sub.f32 	%f355, %f375, %f345;
	mov.f32 	%f369, %f355;
	mov.f32 	%f362, %f355;
$Lt_1_113410:
	.loc	31	57	0
	add.f32 	%f7, %f355, %f3;
	.loc	31	58	0
	add.f32 	%f6, %f2, %f362;
	.loc	31	59	0
	add.f32 	%f5, %f1, %f369;
	ld.param.f32 	%f376, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter0+12];
	mov.f32 	%f377, 0f3f800000;   	// 1
	setp.eq.f32 	%p57, %f376, %f377;
	@!%p57 bra 	$Lt_1_5378;
	.loc	31	63	0
	mov.f32 	%f378, 0f00000000;   	// 0
	max.f32 	%f379, %f5, %f378;
	mov.f32 	%f380, 0f3f800000;   	// 1
	min.f32 	%f5, %f379, %f380;
	mov.f32 	%f381, 0f00000000;   	// 0
	max.f32 	%f382, %f6, %f381;
	mov.f32 	%f383, 0f3f800000;   	// 1
	min.f32 	%f6, %f382, %f383;
	mov.f32 	%f384, 0f00000000;   	// 0
	max.f32 	%f385, %f7, %f384;
	mov.f32 	%f386, 0f3f800000;   	// 1
	min.f32 	%f7, %f385, %f386;
	mov.f32 	%f387, 0f00000000;   	// 0
	max.f32 	%f388, %f4, %f387;
	mov.f32 	%f389, 0f3f800000;   	// 1
	min.f32 	%f8, %f388, %f389;
	bra.uni 	$Lt_1_5378;
$Lt_1_8962:
	.loc	35	99	0
	cvt.rn.f32.s32 	%f390, %r5;
	cvt.rn.f32.s32 	%f391, %r14;
	ld.param.f32 	%f392, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter0+20];
	mul.f32 	%f393, %f391, %f392;
	setp.lt.f32 	%p58, %f390, %f393;
	@!%p58 bra 	$Lt_1_5378;
	.loc	19	267	0
	ld.const.f32 	%f394, [kRGB32f_To_601YPbPr+4];
	mul.f32 	%f29, %f394, %f2;
	ld.const.f32 	%f395, [kRGB32f_To_601YPbPr+28];
	mul.f32 	%f396, %f395, %f2;
	ld.const.f32 	%f397, [kRGB32f_To_601YPbPr+16];
	mul.f32 	%f398, %f397, %f2;
	ld.const.f32 	%f399, [kRGB32f_To_601YPbPr+0];
	mad.f32 	%f31, %f399, %f3, %f29;
	ld.const.f32 	%f400, [kRGB32f_To_601YPbPr+24];
	mad.f32 	%f401, %f400, %f3, %f396;
	ld.const.f32 	%f402, [kRGB32f_To_601YPbPr+12];
	mad.f32 	%f403, %f402, %f3, %f398;
	ld.const.f32 	%f404, [kRGB32f_To_601YPbPr+8];
	mad.f32 	%f33, %f404, %f1, %f31;
	ld.const.f32 	%f405, [kRGB32f_To_601YPbPr+32];
	mad.f32 	%f406, %f405, %f1, %f401;
	ld.const.f32 	%f407, [kRGB32f_To_601YPbPr+20];
	mad.f32 	%f408, %f407, %f1, %f403;
	ld.param.f32 	%f409, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter0+4];
	ld.param.f32 	%f410, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter0+8];
	mad.f32 	%f411, %f33, %f410, %f409;
	ld.param.f32 	%f412, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter0+16];
	mul.f32 	%f413, %f406, %f412;
	ld.param.f32 	%f414, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter0+12];
	mul.f32 	%f415, %f406, %f414;
	mul.f32 	%f416, %f408, %f414;
	sub.f32 	%f417, %f416, %f413;
	mad.f32 	%f418, %f408, %f412, %f415;
	ld.const.f32 	%f419, [k601YPbPr_To_RGB32f+16];
	mul.f32 	%f420, %f419, %f417;
	ld.const.f32 	%f421, [k601YPbPr_To_RGB32f+12];
	mad.f32 	%f422, %f421, %f411, %f420;
	ld.const.f32 	%f423, [k601YPbPr_To_RGB32f+20];
	mad.f32 	%f424, %f423, %f418, %f422;
	.loc	19	268	0
	ld.const.f32 	%f425, [k601YPbPr_To_RGB32f+4];
	mul.f32 	%f426, %f425, %f417;
	ld.const.f32 	%f427, [k601YPbPr_To_RGB32f+0];
	mad.f32 	%f428, %f427, %f411, %f426;
	ld.const.f32 	%f429, [k601YPbPr_To_RGB32f+8];
	mad.f32 	%f430, %f429, %f418, %f428;
	.loc	32	56	0
	ld.const.f32 	%f431, [k601YPbPr_To_RGB32f+28];
	mul.f32 	%f432, %f431, %f417;
	ld.const.f32 	%f433, [k601YPbPr_To_RGB32f+24];
	mad.f32 	%f434, %f433, %f411, %f432;
	ld.const.f32 	%f435, [k601YPbPr_To_RGB32f+32];
	mad.f32 	%f5, %f435, %f418, %f434;
	mov.f32 	%f6, %f424;
	mov.f32 	%f7, %f430;
	bra.uni 	$Lt_1_5378;
$Lt_1_9218:
	.loc	33	46	0
	ld.const.f32 	%f436, [kRGB32f_To_601YPbPr+4];
	mul.f32 	%f29, %f436, %f2;
	ld.const.f32 	%f437, [kRGB32f_To_601YPbPr+0];
	mad.f32 	%f31, %f437, %f3, %f29;
	ld.const.f32 	%f438, [kRGB32f_To_601YPbPr+8];
	mad.f32 	%f33, %f438, %f1, %f31;
	ld.param.f32 	%f43, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter0+8];
	ld.param.f32 	%f439, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter0+20];
	sub.f32 	%f440, %f439, %f43;
	mad.f32 	%f441, %f33, %f440, %f43;
	.loc	33	47	0
	ld.param.f32 	%f442, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter0+12];
	ld.param.f32 	%f443, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter0+24];
	sub.f32 	%f444, %f443, %f442;
	mad.f32 	%f445, %f33, %f444, %f442;
	.loc	33	49	0
	ld.param.f32 	%f446, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter0+28];
	ld.param.f32 	%f13, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter0+4];
	ld.param.f32 	%f447, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter0+16];
	sub.f32 	%f448, %f447, %f13;
	mad.f32 	%f449, %f33, %f448, %f13;
	sub.f32 	%f450, %f449, %f1;
	mad.f32 	%f5, %f446, %f450, %f1;
	.loc	33	50	0
	sub.f32 	%f451, %f441, %f2;
	mad.f32 	%f6, %f446, %f451, %f2;
	.loc	33	51	0
	sub.f32 	%f452, %f445, %f3;
	mad.f32 	%f7, %f446, %f452, %f3;
	.loc	35	103	0
	bra.uni 	$Lt_1_5378;
$Lt_1_9474:
	.loc	34	41	0
	sub.s32 	%r274, %r13, %r11;
	sub.s32 	%r12, %r274, 1;
$Lt_1_5378:
	.loc	35	54	0
	ld.param.u32 	%r275, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter1+0];
	mov.u32 	%r276, 0;
	setp.eq.s32 	%p59, %r275, %r276;
	@%p59 bra 	$Lt_1_258;
	mov.u32 	%r277, 1;
	setp.eq.s32 	%p60, %r275, %r277;
	@%p60 bra 	$Lt_1_770;
	mov.u32 	%r278, 2;
	setp.eq.s32 	%p61, %r275, %r278;
	@%p61 bra 	$Lt_1_1026;
	mov.u32 	%r279, 3;
	setp.eq.s32 	%p62, %r275, %r279;
	@%p62 bra 	$Lt_1_1282;
	mov.u32 	%r280, 4;
	setp.eq.s32 	%p63, %r275, %r280;
	@%p63 bra 	$Lt_1_1538;
	mov.u32 	%r281, 5;
	setp.eq.s32 	%p64, %r275, %r281;
	@%p64 bra 	$Lt_1_1794;
	mov.u32 	%r282, 6;
	setp.eq.s32 	%p65, %r275, %r282;
	@%p65 bra 	$Lt_1_2050;
	mov.u32 	%r283, 7;
	setp.eq.s32 	%p66, %r275, %r283;
	@%p66 bra 	$Lt_1_2306;
	mov.u32 	%r284, 8;
	setp.eq.s32 	%p67, %r275, %r284;
	@%p67 bra 	$Lt_1_2562;
	mov.u32 	%r285, 9;
	setp.eq.s32 	%p68, %r275, %r285;
	@%p68 bra 	$Lt_1_2818;
	mov.u32 	%r286, 10;
	setp.eq.s32 	%p69, %r275, %r286;
	@%p69 bra 	$Lt_1_3074;
	mov.u32 	%r287, 11;
	setp.eq.s32 	%p70, %r275, %r287;
	@%p70 bra 	$Lt_1_3330;
	mov.u32 	%r288, 12;
	setp.eq.s32 	%p71, %r275, %r288;
	@%p71 bra 	$Lt_1_3586;
	mov.u32 	%r289, 13;
	setp.eq.s32 	%p72, %r275, %r289;
	@%p72 bra 	$Lt_1_3842;
	mov.u32 	%r290, 14;
	setp.eq.s32 	%p73, %r275, %r290;
	@%p73 bra 	$Lt_1_4098;
	mov.u32 	%r291, 15;
	setp.eq.s32 	%p74, %r275, %r291;
	@%p74 bra 	$Lt_1_4354;
	mov.u32 	%r292, 16;
	setp.eq.s32 	%p75, %r275, %r292;
	@%p75 bra 	$Lt_1_4610;
	bra.uni 	$Lt_1_514;
$Lt_1_258:
	.loc	18	42	0
	cvt.sat.f32.f32 	%f8, %f8;
	ld.param.f32 	%f453, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter1+8];
	mov.f32 	%f454, 0f3f800000;   	// 1
	setp.eq.f32 	%p76, %f453, %f454;
	ld.param.f32 	%f455, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter1+12];
	mov.f32 	%f456, 0f3f800000;   	// 1
	setp.eq.f32 	%p77, %f455, %f456;
	ld.param.f32 	%f457, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter1+4];
	ld.param.f32 	%f458, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter1+16];
	mov.f32 	%f459, 0f3f800000;   	// 1
	setp.eq.f32 	%p78, %f458, %f459;
	@!%p78 bra 	$Lt_1_115202;
	.loc	18	45	0
	mul.f32 	%f460, %f8, %f457;
	selp.f32 	%f8, %f457, %f460, %p76;
	.loc	18	53	0
	mov.f32 	%f461, 0f3f800000;   	// 1
	sub.f32 	%f462, %f461, %f8;
	selp.f32 	%f8, %f462, %f8, %p77;
	.loc	18	57	0
	mov.f32 	%f5, %f8;
	mov.f32 	%f6, %f8;
	mov.f32 	%f7, %f8;
	bra.uni 	$Lt_1_514;
$Lt_1_115202:
	@!%p76 bra 	$Lt_1_115714;
	.loc	18	61	0
	mov.f32 	%f463, 0f3f800000;   	// 1
	sub.f32 	%f464, %f463, %f457;
	selp.f32 	%f8, %f464, %f457, %p77;
	bra.uni 	$Lt_1_514;
$Lt_1_115714:
	.loc	18	69	0
	mov.f32 	%f465, 0f3f800000;   	// 1
	sub.f32 	%f466, %f465, %f8;
	selp.f32 	%f8, %f466, %f8, %p77;
	.loc	18	73	0
	mul.f32 	%f8, %f8, %f457;
	bra.uni 	$Lt_1_514;
$Lt_1_770:
	.loc	19	267	0
	ld.const.f32 	%f467, [kRGB32f_To_601YPbPr+4];
	mul.f32 	%f468, %f467, %f6;
	ld.const.f32 	%f469, [kRGB32f_To_601YPbPr+0];
	mad.f32 	%f470, %f469, %f7, %f468;
	ld.const.f32 	%f471, [kRGB32f_To_601YPbPr+8];
	mad.f32 	%f472, %f471, %f5, %f470;
	ld.const.f32 	%f473, [k601YPbPr_To_RGB32f+12];
	mul.f32 	%f474, %f473, %f472;
	.loc	19	268	0
	ld.const.f32 	%f475, [k601YPbPr_To_RGB32f+0];
	mul.f32 	%f476, %f475, %f472;
	.loc	20	44	0
	ld.const.f32 	%f477, [k601YPbPr_To_RGB32f+24];
	mul.f32 	%f5, %f477, %f472;
	mov.f32 	%f6, %f474;
	mov.f32 	%f7, %f476;
	.loc	35	61	0
	bra.uni 	$Lt_1_514;
$Lt_1_1026:
	.loc	35	63	0
	ld.param.f32 	%f478, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter1+20];
	ld.param.f32 	%f479, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter1+12];
	ld.param.f32 	%f480, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter1+4];
	mov.f32 	%f481, 0f00000000;   	// 0
	setp.neu.f32 	%p79, %f480, %f481;
	@!%p79 bra 	$Lt_1_116226;
	.loc	21	44	0
	ld.param.f32 	%f482, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter1+8];
	sub.f32 	%f483, %f7, %f479;
	mad.f32 	%f7, %f482, %f483, %f478;
	.loc	21	45	0
	sub.f32 	%f484, %f6, %f479;
	mad.f32 	%f6, %f482, %f484, %f478;
	.loc	21	46	0
	sub.f32 	%f485, %f5, %f479;
	mad.f32 	%f5, %f482, %f485, %f478;
	bra.uni 	$Lt_1_514;
$Lt_1_116226:
	.loc	21	50	0
	ld.param.f32 	%f486, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter1+24];
	setp.gt.f32 	%p80, %f7, %f479;
	selp.f32 	%f7, %f486, %f478, %p80;
	.loc	21	51	0
	setp.gt.f32 	%p81, %f6, %f479;
	selp.f32 	%f6, %f486, %f478, %p81;
	.loc	21	52	0
	setp.gt.f32 	%p82, %f5, %f479;
	selp.f32 	%f5, %f486, %f478, %p82;
	bra.uni 	$Lt_1_514;
$Lt_1_1282:
	.loc	22	47	0
	ld.param.f32 	%f487, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter1+4];
	mul.f32 	%f5, %f487, %f5;
	ld.param.f32 	%f488, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter1+8];
	mul.f32 	%f6, %f488, %f6;
	ld.param.f32 	%f489, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter1+12];
	mul.f32 	%f7, %f489, %f7;
	.loc	35	67	0
	bra.uni 	$Lt_1_514;
$Lt_1_1538:
	.loc	23	48	0
	ld.param.f32 	%f490, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter1+4];
	mov.f32 	%f491, 0f00000000;   	// 0
	setp.eq.f32 	%p83, %f490, %f491;
	ld.param.f32 	%f492, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter1+8];
	mov.f32 	%f493, 0f00000000;   	// 0
	max.f32 	%f494, %f5, %f493;
	mov.f32 	%f495, 0f3f800000;   	// 1
	min.f32 	%f496, %f494, %f495;
	ld.param.f32 	%f497, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter1+12];
	sub.f32 	%f498, %f496, %f497;
	abs.f32 	%f499, %f498;
	mov.f32 	%f500, 0f00000000;   	// 0
	max.f32 	%f501, %f6, %f500;
	mov.f32 	%f502, 0f3f800000;   	// 1
	min.f32 	%f503, %f501, %f502;
	ld.param.f32 	%f504, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter1+16];
	sub.f32 	%f505, %f503, %f504;
	abs.f32 	%f506, %f505;
	mov.f32 	%f507, 0f00000000;   	// 0
	max.f32 	%f508, %f7, %f507;
	mov.f32 	%f509, 0f3f800000;   	// 1
	min.f32 	%f510, %f508, %f509;
	ld.param.f32 	%f511, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter1+20];
	sub.f32 	%f512, %f510, %f511;
	abs.f32 	%f513, %f512;
	mov.f32 	%f514, 0f00000000;   	// 0
	max.f32 	%f515, %f8, %f514;
	mov.f32 	%f516, 0f3f800000;   	// 1
	min.f32 	%f517, %f515, %f516;
	sub.f32 	%f518, %f517, %f8;
	abs.f32 	%f519, %f518;
	max.f32 	%f520, %f513, %f519;
	max.f32 	%f521, %f506, %f520;
	max.f32 	%f522, %f499, %f521;
	setp.ge.f32 	%p84, %f492, %f522;
	xor.pred 	%p85, %p83, %p84;
	@!%p85 bra 	$Lt_1_514;
	.loc	19	267	0
	ld.const.f32 	%f523, [kRGB32f_To_601YPbPr+4];
	mul.f32 	%f468, %f523, %f6;
	ld.const.f32 	%f524, [kRGB32f_To_601YPbPr+0];
	mad.f32 	%f470, %f524, %f7, %f468;
	ld.const.f32 	%f525, [kRGB32f_To_601YPbPr+8];
	mad.f32 	%f472, %f525, %f5, %f470;
	ld.const.f32 	%f526, [k601YPbPr_To_RGB32f+12];
	mul.f32 	%f527, %f526, %f472;
	.loc	19	268	0
	ld.const.f32 	%f528, [k601YPbPr_To_RGB32f+0];
	mul.f32 	%f529, %f528, %f472;
	.loc	20	44	0
	ld.const.f32 	%f530, [k601YPbPr_To_RGB32f+24];
	mul.f32 	%f5, %f530, %f472;
	mov.f32 	%f6, %f527;
	mov.f32 	%f7, %f529;
	bra.uni 	$Lt_1_514;
$Lt_1_1794:
	.loc	24	48	0
	ld.param.f32 	%f531, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter1+8];
	mov.f32 	%f532, 0f00000000;   	// 0
	max.f32 	%f533, %f5, %f532;
	mov.f32 	%f534, 0f3f800000;   	// 1
	min.f32 	%f535, %f533, %f534;
	ld.param.f32 	%f536, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter1+12];
	sub.f32 	%f537, %f535, %f536;
	abs.f32 	%f538, %f537;
	mov.f32 	%f539, 0f00000000;   	// 0
	max.f32 	%f540, %f6, %f539;
	mov.f32 	%f541, 0f3f800000;   	// 1
	min.f32 	%f542, %f540, %f541;
	ld.param.f32 	%f543, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter1+16];
	sub.f32 	%f544, %f542, %f543;
	abs.f32 	%f545, %f544;
	mov.f32 	%f546, 0f00000000;   	// 0
	max.f32 	%f547, %f7, %f546;
	mov.f32 	%f548, 0f3f800000;   	// 1
	min.f32 	%f549, %f547, %f548;
	ld.param.f32 	%f550, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter1+20];
	sub.f32 	%f551, %f549, %f550;
	abs.f32 	%f552, %f551;
	mov.f32 	%f553, 0f00000000;   	// 0
	max.f32 	%f554, %f8, %f553;
	mov.f32 	%f555, 0f3f800000;   	// 1
	min.f32 	%f556, %f554, %f555;
	sub.f32 	%f557, %f556, %f8;
	abs.f32 	%f558, %f557;
	max.f32 	%f559, %f552, %f558;
	max.f32 	%f560, %f545, %f559;
	max.f32 	%f561, %f538, %f560;
	setp.ge.f32 	%p86, %f531, %f561;
	@!%p86 bra 	$Lt_1_514;
	.loc	24	51	0
	ld.param.f32 	%f562, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter1+24];
	mov.f32 	%f124, %f562;
	ld.param.f32 	%f563, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter1+28];
	mov.f32 	%f126, %f563;
	ld.param.f32 	%f564, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter1+32];
	mov.f32 	%f128, %f564;
	ld.param.f32 	%f565, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter1+4];
	mov.f32 	%f566, 0f00000000;   	// 0
	setp.eq.f32 	%p87, %f565, %f566;
	@!%p87 bra 	$Lt_1_117506;
	.loc	24	60	0
	ld.const.f32 	%f567, [kRGB32f_To_601YPbPr+4];
	mul.f32 	%f468, %f567, %f6;
	ld.const.f32 	%f568, [kRGB32f_To_601YPbPr+0];
	mad.f32 	%f470, %f568, %f7, %f468;
	ld.const.f32 	%f569, [kRGB32f_To_601YPbPr+8];
	mad.f32 	%f472, %f569, %f5, %f470;
	mul.f32 	%f124, %f472, %f562;
	.loc	24	61	0
	mul.f32 	%f126, %f472, %f563;
	.loc	24	62	0
	mul.f32 	%f128, %f472, %f564;
$Lt_1_117506:
	.loc	24	65	0
	mov.f32 	%f5, %f124;
	mov.f32 	%f6, %f126;
	mov.f32 	%f7, %f128;
	bra.uni 	$Lt_1_514;
$Lt_1_2050:
	.loc	25	47	0
	sub.s32 	%r293, %r13, %r12;
	sub.s32 	%r294, %r14, %r6;
	cvt.rn.f32.s32 	%f570, %r6;
	cvt.rn.f32.s32 	%f571, %r12;
	cvt.rn.f32.s32 	%f572, %r293;
	cvt.rn.f32.s32 	%f573, %r294;
	ld.param.f32 	%f457, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter1+4];
	sub.f32 	%f574, %f457, %f570;
	sub.f32 	%f575, %f457, %f571;
	sub.f32 	%f576, %f457, %f572;
	sub.f32 	%f577, %f457, %f573;
	cvt.rzi.s32.f32 	%r295, %f574;
	cvt.rzi.s32.f32 	%r296, %f575;
	cvt.rzi.s32.f32 	%r297, %f576;
	cvt.rzi.s32.f32 	%r298, %f577;
	max.s32 	%r299, %r295, %r296;
	max.s32 	%r300, %r298, %r299;
	max.s32 	%r301, %r297, %r300;
	mov.u32 	%r302, 0;
	setp.le.s32 	%p88, %r301, %r302;
	@%p88 bra 	$Lt_1_514;
	.loc	25	51	0
	cvt.rn.f32.s32 	%f578, %r301;
	sub.f32 	%f579, %f457, %f578;
	div.full.f32 	%f580, %f579, %f457;
	mul.f32 	%f8, %f8, %f580;
	bra.uni 	$Lt_1_514;
$Lt_1_2306:
	.loc	27	50	0
	ld.const.f32 	%f581, [kRGB32f_To_601YPbPr+4];
	mul.f32 	%f468, %f581, %f6;
	ld.param.f32 	%f582, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter1+12];
	ld.param.f32 	%f482, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter1+8];
	sub.f32 	%f583, %f582, %f482;
	ld.const.f32 	%f584, [kRGB32f_To_601YPbPr+0];
	mad.f32 	%f470, %f584, %f7, %f468;
	ld.param.f32 	%f585, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter1+16];
	mul.f32 	%f586, %f585, %f583;
	ld.const.f32 	%f587, [kRGB32f_To_601YPbPr+8];
	mad.f32 	%f472, %f587, %f5, %f470;
	add.f32 	%f588, %f586, %f582;
	sub.f32 	%f589, %f482, %f586;
	mov.f32 	%f590, 0f00000000;   	// 0
	max.f32 	%f591, %f588, %f590;
	mov.f32 	%f592, 0f00000000;   	// 0
	max.f32 	%f593, %f589, %f592;
	mov.f32 	%f594, 0f3f800000;   	// 1
	min.f32 	%f595, %f591, %f594;
	mov.f32 	%f596, 0f3f800000;   	// 1
	min.f32 	%f597, %f593, %f596;
	set.gt.u32.f32 	%r303, %f597, %f472;
	neg.s32 	%r304, %r303;
	set.le.u32.f32 	%r305, %f595, %f472;
	neg.s32 	%r306, %r305;
	or.b32 	%r307, %r304, %r306;
	mov.u32 	%r308, 0;
	setp.eq.s32 	%p89, %r307, %r308;
	@%p89 bra 	$Lt_1_118786;
	mov.f32 	%f162, 0f00000000;   	// 0
	bra.uni 	$Lt_1_119554;
$Lt_1_118786:
	add.f32 	%f598, %f586, %f482;
	mov.f32 	%f599, 0f00000000;   	// 0
	max.f32 	%f600, %f598, %f599;
	mov.f32 	%f601, 0f3f800000;   	// 1
	min.f32 	%f602, %f600, %f601;
	set.le.u32.f32 	%r309, %f602, %f472;
	neg.s32 	%r310, %r309;
	sub.f32 	%f603, %f582, %f586;
	mov.f32 	%f604, 0f00000000;   	// 0
	max.f32 	%f605, %f603, %f604;
	mov.f32 	%f606, 0f3f800000;   	// 1
	min.f32 	%f607, %f605, %f606;
	set.lt.u32.f32 	%r311, %f472, %f607;
	neg.s32 	%r312, %r311;
	and.b32 	%r313, %r310, %r312;
	mov.u32 	%r314, 0;
	setp.eq.s32 	%p90, %r313, %r314;
	@%p90 bra 	$Lt_1_119298;
	mov.f32 	%f162, 0f3f800000;   	// 1
	bra.uni 	$Lt_1_119554;
$Lt_1_119298:
	add.f32 	%f608, %f586, %f586;
	setp.gt.f32 	%p91, %f602, %f472;
	@!%p91 bra 	$Lt_1_119810;
	.loc	27	62	0
	sub.f32 	%f609, %f472, %f597;
	div.full.f32 	%f162, %f609, %f608;
	bra.uni 	$Lt_1_119554;
$Lt_1_119810:
	.loc	27	66	0
	sub.f32 	%f610, %f595, %f472;
	div.full.f32 	%f162, %f610, %f608;
$Lt_1_119554:
$Lt_1_119042:
$Lt_1_118530:
	.loc	27	69	0
	mov.f32 	%f611, 0f3f800000;   	// 1
	sub.f32 	%f612, %f611, %f162;
	ld.param.f32 	%f613, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter1+4];
	mov.f32 	%f614, 0f00000000;   	// 0
	setp.neu.f32 	%p92, %f613, %f614;
	selp.f32 	%f162, %f612, %f162, %p92;
	.loc	27	77	0
	ld.const.f32 	%f615, [k601YPbPr_To_RGB32f+24];
	mul.f32 	%f5, %f615, %f162;
	ld.const.f32 	%f616, [k601YPbPr_To_RGB32f+12];
	mul.f32 	%f6, %f616, %f162;
	ld.const.f32 	%f617, [k601YPbPr_To_RGB32f+0];
	mul.f32 	%f7, %f617, %f162;
	.loc	35	79	0
	bra.uni 	$Lt_1_514;
$Lt_1_2562:
	.loc	35	80	0
	ld.param.f32 	%f618, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter1+44];
	mov.f32 	%f619, 0f00000000;   	// 0
	setp.ge.f32 	%p93, %f618, %f619;
	@!%p93 bra 	$L_1_101890;
	cvt.rn.f32.s32 	%f620, %r6;
	cvt.rn.f32.s32 	%f621, %r14;
	mul.f32 	%f622, %f621, %f618;
	setp.lt.f32 	%p94, %f620, %f622;
	@%p94 bra 	$L_1_101634;
$L_1_101890:
	mov.f32 	%f623, 0f00000000;   	// 0
	setp.lt.f32 	%p95, %f618, %f623;
	@!%p95 bra 	$Lt_1_514;
	cvt.rn.f32.s32 	%f624, %r12;
	cvt.rn.f32.s32 	%f625, %r13;
	mul.f32 	%f626, %f625, %f618;
	neg.f32 	%f627, %f626;
	setp.lt.f32 	%p96, %f624, %f627;
	@!%p96 bra 	$Lt_1_514;
$L_1_101634:
	.loc	28	47	0
	ld.const.f32 	%f628, [kRGB32f_To_601YPbPr+4];
	mul.f32 	%f629, %f628, %f6;
	ld.const.f32 	%f630, [kRGB32f_To_601YPbPr+0];
	mad.f32 	%f631, %f630, %f7, %f629;
	ld.const.f32 	%f632, [kRGB32f_To_601YPbPr+8];
	mad.f32 	%f472, %f632, %f5, %f631;
	mov.f32 	%f198, %f472;
	ld.param.f32 	%f482, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter1+8];
	ld.param.f32 	%f633, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter1+4];
	sub.f32 	%f634, %f633, %f482;
	ld.param.f32 	%f635, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter1+16];
	ld.param.f32 	%f636, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter1+12];
	sub.f32 	%f637, %f636, %f635;
	mov.f32 	%f638, 0f3f800000;   	// 1
	set.neu.u32.f32 	%r315, %f637, %f638;
	neg.s32 	%r316, %r315;
	mov.f32 	%f639, 0f3f800000;   	// 1
	set.neu.u32.f32 	%r317, %f634, %f639;
	neg.s32 	%r318, %r317;
	or.b32 	%r319, %r316, %r318;
	mov.u32 	%r320, 0;
	setp.eq.s32 	%p97, %r319, %r320;
	@%p97 bra 	$Lt_1_120066;
	.loc	17	116	0
	mov.u32 	%r321, 1;
	sub.u32 	%r322, %r321, %r5;
	shr.u32 	%r323, %r11, 13;
	sub.u32 	%r73, %r5, %r11;
	sub.u32 	%r324, %r322, %r11;
	xor.b32 	%r325, %r324, %r323;
	shl.b32 	%r326, %r325, 8;
	sub.u32 	%r327, %r73, %r325;
	sub.u32 	%r328, %r11, %r325;
	xor.b32 	%r329, %r327, %r326;
	shr.u32 	%r330, %r329, 13;
	sub.u32 	%r331, %r328, %r329;
	xor.b32 	%r332, %r331, %r330;
	shr.u32 	%r333, %r332, 12;
	sub.u32 	%r334, %r325, %r329;
	sub.u32 	%r335, %r334, %r332;
	xor.b32 	%r336, %r333, %r335;
	.loc	17	117	0
	sub.u32 	%r337, %r329, %r332;
	sub.u32 	%r338, %r337, %r336;
	shl.b32 	%r339, %r336, 16;
	xor.b32 	%r340, %r338, %r339;
	.loc	17	118	0
	sub.u32 	%r341, %r332, %r336;
	sub.u32 	%r342, %r341, %r340;
	shr.u32 	%r343, %r340, 5;
	xor.b32 	%r344, %r342, %r343;
	.loc	17	119	0
	sub.u32 	%r345, %r336, %r340;
	sub.u32 	%r346, %r345, %r344;
	shr.u32 	%r347, %r344, 3;
	xor.b32 	%r348, %r346, %r347;
	.loc	17	120	0
	sub.u32 	%r349, %r340, %r344;
	sub.u32 	%r350, %r349, %r348;
	shl.b32 	%r351, %r348, 10;
	xor.b32 	%r352, %r350, %r351;
	.loc	17	121	0
	sub.u32 	%r353, %r344, %r348;
	sub.u32 	%r354, %r353, %r352;
	shr.u32 	%r355, %r352, 15;
	xor.b32 	%r356, %r354, %r355;
	.loc	28	57	0
	mov.f32 	%f640, 0f3b270d73;   	// 0.00254902
	mul.lo.u32 	%r357, %r356, 1103515245;
	add.u32 	%r358, %r357, 12345;
	shr.u32 	%r359, %r358, 16;
	and.b32 	%r360, %r359, 255;
	shl.b32 	%r361, %r360, 7;
	mul.lo.u32 	%r362, %r356, -1029531031;
	sub.u32 	%r363, %r362, 740551042;
	shr.u32 	%r364, %r363, 16;
	and.b32 	%r365, %r364, 255;
	xor.b32 	%r366, %r361, %r365;
	cvt.rn.f32.s32 	%f641, %r366;
	mov.f32 	%f642, 0f467ffe00;   	// 16383.5
	div.full.f32 	%f643, %f641, %f642;
	mov.f32 	%f644, 0fbf800000;   	// -1
	add.f32 	%f645, %f643, %f644;
	mad.f32 	%f198, %f640, %f645, %f472;
$Lt_1_120066:
	sub.f32 	%f646, %f198, %f482;
	ld.param.f32 	%f647, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter1+20];
	mov.f32 	%f648, 0f3f800000;   	// 1
	setp.neu.f32 	%p98, %f647, %f648;
	@!%p98 bra 	$Lt_1_120834;
	mov.f32 	%f649, 0f00000000;   	// 0
	setp.lt.f32 	%p99, %f646, %f649;
	@!%p99 bra 	$Lt_1_121346;
	.loc	28	66	0
	mov.f32 	%f198, %f635;
	bra.uni 	$Lt_1_120578;
$Lt_1_121346:
	.loc	28	70	0
	rcp.approx.f32 	%f650, %f634;
	mul.f32 	%f651, %f646, %f650;
	lg2.approx.f32 	%f652, %f651;
	mul.f32 	%f653, %f647, %f652;
	ex2.approx.f32 	%f654, %f653;
	mad.f32 	%f198, %f637, %f654, %f635;
	bra.uni 	$Lt_1_120578;
$Lt_1_120834:
	.loc	28	77	0
	rcp.approx.f32 	%f655, %f634;
	mul.f32 	%f656, %f646, %f655;
	mad.f32 	%f198, %f637, %f656, %f635;
$Lt_1_120578:
	.loc	19	267	0
	ld.param.f32 	%f657, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter1+28];
	sin.approx.f32 	%f658, %f657;
	ld.const.f32 	%f659, [kRGB32f_To_601YPbPr+28];
	mul.f32 	%f660, %f659, %f6;
	ld.const.f32 	%f661, [kRGB32f_To_601YPbPr+16];
	mul.f32 	%f662, %f661, %f6;
	cos.approx.f32 	%f663, %f657;
	ld.const.f32 	%f664, [kRGB32f_To_601YPbPr+24];
	mad.f32 	%f665, %f664, %f7, %f660;
	ld.const.f32 	%f666, [kRGB32f_To_601YPbPr+12];
	mad.f32 	%f667, %f666, %f7, %f662;
	ld.const.f32 	%f668, [kRGB32f_To_601YPbPr+32];
	mad.f32 	%f669, %f668, %f5, %f665;
	ld.const.f32 	%f670, [kRGB32f_To_601YPbPr+20];
	mad.f32 	%f671, %f670, %f5, %f667;
	mul.f32 	%f672, %f658, %f669;
	mul.f32 	%f673, %f663, %f669;
	mul.f32 	%f674, %f671, %f663;
	sub.f32 	%f675, %f674, %f672;
	mad.f32 	%f676, %f671, %f658, %f673;
	ld.param.f32 	%f677, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter1+32];
	sub.f32 	%f678, %f677, %f675;
	ld.param.f32 	%f679, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter1+36];
	sub.f32 	%f680, %f679, %f676;
	ld.param.f32 	%f681, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter1+40];
	mad.f32 	%f682, %f681, %f678, %f675;
	mad.f32 	%f683, %f681, %f680, %f676;
	ld.param.f32 	%f684, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter1+24];
	mul.f32 	%f685, %f682, %f684;
	mul.f32 	%f686, %f683, %f684;
	ld.const.f32 	%f687, [k601YPbPr_To_RGB32f+16];
	mul.f32 	%f688, %f687, %f685;
	ld.const.f32 	%f689, [k601YPbPr_To_RGB32f+12];
	mad.f32 	%f690, %f689, %f198, %f688;
	ld.const.f32 	%f691, [k601YPbPr_To_RGB32f+20];
	mad.f32 	%f692, %f691, %f686, %f690;
	.loc	19	268	0
	ld.const.f32 	%f693, [k601YPbPr_To_RGB32f+4];
	mul.f32 	%f694, %f693, %f685;
	ld.const.f32 	%f695, [k601YPbPr_To_RGB32f+0];
	mad.f32 	%f696, %f695, %f198, %f694;
	ld.const.f32 	%f697, [k601YPbPr_To_RGB32f+8];
	mad.f32 	%f698, %f697, %f686, %f696;
	.loc	28	92	0
	ld.const.f32 	%f699, [k601YPbPr_To_RGB32f+28];
	mul.f32 	%f700, %f699, %f685;
	ld.const.f32 	%f701, [k601YPbPr_To_RGB32f+24];
	mad.f32 	%f702, %f701, %f198, %f700;
	ld.const.f32 	%f703, [k601YPbPr_To_RGB32f+32];
	mad.f32 	%f5, %f703, %f686, %f702;
	mov.f32 	%f6, %f692;
	mov.f32 	%f7, %f698;
	bra.uni 	$Lt_1_514;
$Lt_1_2818:
	.loc	35	83	0
	ld.param.f32 	%f704, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter1+44];
	mov.f32 	%f705, 0f00000000;   	// 0
	setp.ge.f32 	%p100, %f704, %f705;
	@!%p100 bra 	$L_1_102658;
	cvt.rn.f32.s32 	%f706, %r6;
	cvt.rn.f32.s32 	%f707, %r14;
	mul.f32 	%f708, %f707, %f704;
	setp.lt.f32 	%p101, %f706, %f708;
	@%p101 bra 	$L_1_102402;
$L_1_102658:
	mov.f32 	%f709, 0f00000000;   	// 0
	setp.lt.f32 	%p102, %f704, %f709;
	@!%p102 bra 	$Lt_1_514;
	cvt.rn.f32.s32 	%f710, %r12;
	cvt.rn.f32.s32 	%f711, %r13;
	mul.f32 	%f712, %f711, %f704;
	neg.f32 	%f713, %f712;
	setp.lt.f32 	%p103, %f710, %f713;
	@!%p103 bra 	$Lt_1_514;
$L_1_102402:
	.loc	28	110	0
	ld.const.f32 	%f714, [kRGB32f_To_601YPbPr+4];
	mul.f32 	%f468, %f714, %f6;
	ld.param.f32 	%f482, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter1+8];
	ld.param.f32 	%f715, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter1+4];
	sub.f32 	%f716, %f715, %f482;
	ld.param.f32 	%f717, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter1+16];
	ld.param.f32 	%f718, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter1+12];
	sub.f32 	%f719, %f718, %f717;
	ld.const.f32 	%f720, [kRGB32f_To_601YPbPr+0];
	mad.f32 	%f470, %f720, %f7, %f468;
	ld.const.f32 	%f721, [kRGB32f_To_601YPbPr+8];
	mad.f32 	%f472, %f721, %f5, %f470;
	sub.f32 	%f722, %f472, %f482;
	ld.param.f32 	%f723, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter1+20];
	mov.f32 	%f724, 0f3f800000;   	// 1
	setp.neu.f32 	%p104, %f723, %f724;
	@!%p104 bra 	$Lt_1_121858;
	.loc	28	120	0
	mov.f32 	%f725, 0f00000000;   	// 0
	max.f32 	%f726, %f722, %f725;
	div.full.f32 	%f727, %f726, %f716;
	lg2.approx.f32 	%f728, %f727;
	mul.f32 	%f729, %f723, %f728;
	ex2.approx.f32 	%f730, %f729;
	mad.f32 	%f297, %f719, %f730, %f717;
	bra.uni 	$Lt_1_121602;
$Lt_1_121858:
	.loc	28	129	0
	div.full.f32 	%f731, %f722, %f716;
	mad.f32 	%f297, %f719, %f731, %f717;
$Lt_1_121602:
	.loc	28	135	0
	ld.const.f32 	%f732, [k601YPbPr_To_RGB32f+24];
	mul.f32 	%f5, %f732, %f297;
	ld.const.f32 	%f733, [k601YPbPr_To_RGB32f+12];
	mul.f32 	%f6, %f733, %f297;
	ld.const.f32 	%f734, [k601YPbPr_To_RGB32f+0];
	mul.f32 	%f7, %f734, %f297;
	bra.uni 	$Lt_1_514;
$Lt_1_3074:
	.loc	35	86	0
	ld.param.f32 	%f735, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter1+44];
	mov.f32 	%f736, 0f00000000;   	// 0
	setp.ge.f32 	%p105, %f735, %f736;
	@!%p105 bra 	$L_1_103426;
	cvt.rn.f32.s32 	%f737, %r6;
	cvt.rn.f32.s32 	%f738, %r14;
	mul.f32 	%f739, %f738, %f735;
	setp.lt.f32 	%p106, %f737, %f739;
	@%p106 bra 	$L_1_103170;
$L_1_103426:
	mov.f32 	%f740, 0f00000000;   	// 0
	setp.lt.f32 	%p107, %f735, %f740;
	@!%p107 bra 	$Lt_1_514;
	cvt.rn.f32.s32 	%f741, %r12;
	cvt.rn.f32.s32 	%f742, %r13;
	mul.f32 	%f743, %f742, %f735;
	neg.f32 	%f744, %f743;
	setp.lt.f32 	%p108, %f741, %f744;
	@!%p108 bra 	$Lt_1_514;
$L_1_103170:
	.loc	28	160	0
	ld.const.f32 	%f5, [k601YPbPr_To_RGB32f+24];
	ld.const.f32 	%f6, [k601YPbPr_To_RGB32f+12];
	ld.const.f32 	%f7, [k601YPbPr_To_RGB32f+0];
	bra.uni 	$Lt_1_514;
$Lt_1_3330:
	.loc	29	42	0
	ld.param.f32 	%f457, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter1+4];
	mov.f32 	%f745, 0f00000000;   	// 0
	setp.ge.f32 	%p109, %f5, %f745;
	@!%p109 bra 	$Lt_1_122370;
	.loc	19	292	0
	lg2.approx.f32 	%f746, %f5;
	mul.f32 	%f747, %f457, %f746;
	ex2.approx.f32 	%f748, %f747;
	bra.uni 	$Lt_1_122114;
$Lt_1_122370:
	neg.f32 	%f749, %f5;
	lg2.approx.f32 	%f750, %f749;
	mul.f32 	%f751, %f457, %f750;
	ex2.approx.f32 	%f752, %f751;
	neg.f32 	%f748, %f752;
$Lt_1_122114:
	mov.f32 	%f753, 0f00000000;   	// 0
	setp.ge.f32 	%p110, %f6, %f753;
	@!%p110 bra 	$Lt_1_122882;
	.loc	19	293	0
	lg2.approx.f32 	%f754, %f6;
	mul.f32 	%f755, %f457, %f754;
	ex2.approx.f32 	%f756, %f755;
	bra.uni 	$Lt_1_122626;
$Lt_1_122882:
	neg.f32 	%f757, %f6;
	lg2.approx.f32 	%f758, %f757;
	mul.f32 	%f759, %f457, %f758;
	ex2.approx.f32 	%f760, %f759;
	neg.f32 	%f756, %f760;
$Lt_1_122626:
	mov.f32 	%f761, 0f00000000;   	// 0
	setp.ge.f32 	%p111, %f7, %f761;
	@!%p111 bra 	$Lt_1_123394;
	.loc	19	294	0
	lg2.approx.f32 	%f762, %f7;
	mul.f32 	%f763, %f457, %f762;
	ex2.approx.f32 	%f764, %f763;
	bra.uni 	$Lt_1_123138;
$Lt_1_123394:
	neg.f32 	%f765, %f7;
	lg2.approx.f32 	%f766, %f765;
	mul.f32 	%f767, %f457, %f766;
	ex2.approx.f32 	%f768, %f767;
	neg.f32 	%f764, %f768;
$Lt_1_123138:
	mov.f32 	%f769, 0f00000000;   	// 0
	setp.ge.f32 	%p112, %f8, %f769;
	@!%p112 bra 	$Lt_1_123906;
	.loc	19	295	0
	lg2.approx.f32 	%f770, %f8;
	mul.f32 	%f771, %f457, %f770;
	ex2.approx.f32 	%f772, %f771;
	bra.uni 	$Lt_1_123650;
$Lt_1_123906:
	neg.f32 	%f773, %f8;
	lg2.approx.f32 	%f774, %f773;
	mul.f32 	%f775, %f457, %f774;
	ex2.approx.f32 	%f776, %f775;
	neg.f32 	%f772, %f776;
$Lt_1_123650:
	.loc	29	42	0
	mov.f32 	%f5, %f748;
	mov.f32 	%f6, %f756;
	mov.f32 	%f7, %f764;
	mov.f32 	%f8, %f772;
	.loc	35	91	0
	bra.uni 	$Lt_1_514;
$Lt_1_3586:
	.loc	30	41	0
	sub.s32 	%r367, %r14, %r6;
	sub.s32 	%r6, %r367, 1;
	.loc	35	94	0
	bra.uni 	$Lt_1_514;
$Lt_1_3842:
	.loc	35	96	0
	sub.u32 	%r368, %r6, %r12;
	ld.param.f32 	%f457, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter1+4];
	mov.f32 	%f777, 0f3e800000;   	// 0.25
	mul.f32 	%f778, %f457, %f777;
	ld.param.f32 	%f779, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter1+16];
	ld.param.f32 	%f780, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter1+8];
	mov.f32 	%f781, 0f3f800000;   	// 1
	setp.eq.f32 	%p113, %f780, %f781;
	@!%p113 bra 	$Lt_1_124418;
	.loc	17	117	0
	cvt.rzi.u32.f32 	%r369, %f779;
	shr.u32 	%r370, %r369, 13;
	sub.u32 	%r371, %r12, %r369;
	sub.u32 	%r372, %r368, %r369;
	xor.b32 	%r373, %r372, %r370;
	shl.b32 	%r374, %r373, 8;
	sub.u32 	%r375, %r371, %r373;
	sub.u32 	%r376, %r369, %r373;
	xor.b32 	%r377, %r375, %r374;
	shr.u32 	%r378, %r377, 13;
	sub.u32 	%r379, %r376, %r377;
	sub.u32 	%r380, %r373, %r377;
	xor.b32 	%r381, %r379, %r378;
	shr.u32 	%r382, %r381, 12;
	sub.u32 	%r383, %r380, %r381;
	xor.b32 	%r384, %r383, %r382;
	shl.b32 	%r385, %r384, 16;
	sub.u32 	%r386, %r377, %r381;
	sub.u32 	%r387, %r386, %r384;
	xor.b32 	%r388, %r385, %r387;
	.loc	17	118	0
	sub.u32 	%r389, %r381, %r384;
	sub.u32 	%r390, %r389, %r388;
	shr.u32 	%r391, %r388, 5;
	xor.b32 	%r392, %r390, %r391;
	.loc	17	119	0
	sub.u32 	%r393, %r384, %r388;
	sub.u32 	%r394, %r393, %r392;
	shr.u32 	%r395, %r392, 3;
	xor.b32 	%r396, %r394, %r395;
	.loc	17	120	0
	sub.u32 	%r397, %r388, %r392;
	sub.u32 	%r398, %r397, %r396;
	shl.b32 	%r399, %r396, 10;
	xor.b32 	%r400, %r398, %r399;
	.loc	17	121	0
	sub.u32 	%r401, %r392, %r396;
	sub.u32 	%r402, %r401, %r400;
	shr.u32 	%r403, %r400, 15;
	xor.b32 	%r404, %r402, %r403;
	.loc	31	48	0
	mul.lo.u32 	%r405, %r404, 1103515245;
	add.u32 	%r406, %r405, 12345;
	shr.u32 	%r407, %r406, 16;
	and.b32 	%r408, %r407, 255;
	shl.b32 	%r409, %r408, 7;
	mul.lo.u32 	%r410, %r404, -1029531031;
	sub.u32 	%r411, %r410, 740551042;
	shr.u32 	%r412, %r411, 16;
	and.b32 	%r413, %r412, 255;
	xor.b32 	%r414, %r409, %r413;
	cvt.rn.f32.s32 	%f782, %r414;
	mov.f32 	%f783, 0f46fffe00;   	// 32767
	div.full.f32 	%f784, %f782, %f783;
	mul.f32 	%f785, %f457, %f784;
	mov.f32 	%f786, 0f3f000000;   	// 0.5
	mul.f32 	%f787, %f785, %f786;
	sub.f32 	%f788, %f787, %f778;
	.loc	17	117	0
	add.u32 	%r415, %r372, 1;
	xor.b32 	%r416, %r415, %r370;
	shl.b32 	%r417, %r416, 8;
	sub.u32 	%r418, %r371, %r416;
	sub.u32 	%r419, %r369, %r416;
	xor.b32 	%r420, %r417, %r418;
	shr.u32 	%r421, %r420, 13;
	sub.u32 	%r422, %r419, %r420;
	sub.u32 	%r423, %r416, %r420;
	xor.b32 	%r424, %r421, %r422;
	shr.u32 	%r425, %r424, 12;
	sub.u32 	%r426, %r423, %r424;
	xor.b32 	%r427, %r425, %r426;
	sub.u32 	%r428, %r420, %r424;
	sub.u32 	%r178, %r428, %r427;
	shl.b32 	%r429, %r427, 16;
	xor.b32 	%r178, %r178, %r429;
	.loc	17	118	0
	sub.u32 	%r180, %r424, %r427;
	sub.u32 	%r180, %r180, %r178;
	shr.u32 	%r430, %r178, 5;
	xor.b32 	%r180, %r180, %r430;
	.loc	17	119	0
	sub.u32 	%r182, %r427, %r178;
	sub.u32 	%r182, %r182, %r180;
	shr.u32 	%r431, %r180, 3;
	xor.b32 	%r182, %r182, %r431;
	.loc	17	120	0
	sub.u32 	%r178, %r178, %r180;
	sub.u32 	%r178, %r178, %r182;
	shl.b32 	%r432, %r182, 10;
	xor.b32 	%r178, %r178, %r432;
	.loc	17	121	0
	sub.u32 	%r180, %r180, %r182;
	sub.u32 	%r180, %r180, %r178;
	shr.u32 	%r433, %r178, 15;
	xor.b32 	%r180, %r180, %r433;
	.loc	31	49	0
	mul.lo.u32 	%r434, %r180, 1103515245;
	add.u32 	%r435, %r434, 12345;
	shr.u32 	%r436, %r435, 16;
	and.b32 	%r437, %r436, 255;
	shl.b32 	%r438, %r437, 7;
	mul.lo.u32 	%r439, %r180, -1029531031;
	sub.u32 	%r440, %r439, 740551042;
	shr.u32 	%r441, %r440, 16;
	and.b32 	%r442, %r441, 255;
	xor.b32 	%r443, %r438, %r442;
	cvt.rn.f32.s32 	%f789, %r443;
	mov.f32 	%f790, 0f46fffe00;   	// 32767
	div.full.f32 	%f791, %f789, %f790;
	mul.f32 	%f792, %f457, %f791;
	mov.f32 	%f793, 0f3f000000;   	// 0.5
	mul.f32 	%f794, %f792, %f793;
	sub.f32 	%f362, %f794, %f778;
	.loc	17	117	0
	add.u32 	%r444, %r372, 2;
	xor.b32 	%r445, %r444, %r370;
	shl.b32 	%r446, %r445, 8;
	sub.u32 	%r447, %r371, %r445;
	sub.u32 	%r448, %r369, %r445;
	xor.b32 	%r449, %r446, %r447;
	shr.u32 	%r450, %r449, 13;
	sub.u32 	%r451, %r448, %r449;
	sub.u32 	%r452, %r445, %r449;
	xor.b32 	%r453, %r450, %r451;
	shr.u32 	%r454, %r453, 12;
	sub.u32 	%r455, %r452, %r453;
	xor.b32 	%r456, %r454, %r455;
	sub.u32 	%r457, %r449, %r453;
	sub.u32 	%r210, %r457, %r456;
	shl.b32 	%r458, %r456, 16;
	xor.b32 	%r210, %r210, %r458;
	.loc	17	118	0
	sub.u32 	%r212, %r453, %r456;
	sub.u32 	%r212, %r212, %r210;
	shr.u32 	%r459, %r210, 5;
	xor.b32 	%r212, %r212, %r459;
	.loc	17	119	0
	sub.u32 	%r214, %r456, %r210;
	sub.u32 	%r214, %r214, %r212;
	shr.u32 	%r460, %r212, 3;
	xor.b32 	%r214, %r214, %r460;
	.loc	17	120	0
	sub.u32 	%r210, %r210, %r212;
	sub.u32 	%r210, %r210, %r214;
	shl.b32 	%r461, %r214, 10;
	xor.b32 	%r210, %r210, %r461;
	.loc	17	121	0
	sub.u32 	%r212, %r212, %r214;
	sub.u32 	%r212, %r212, %r210;
	shr.u32 	%r462, %r210, 15;
	xor.b32 	%r212, %r212, %r462;
	.loc	31	50	0
	mul.lo.u32 	%r463, %r212, 1103515245;
	add.u32 	%r464, %r463, 12345;
	shr.u32 	%r465, %r464, 16;
	and.b32 	%r466, %r465, 255;
	shl.b32 	%r467, %r466, 7;
	mul.lo.u32 	%r468, %r212, -1029531031;
	sub.u32 	%r469, %r468, 740551042;
	shr.u32 	%r470, %r469, 16;
	and.b32 	%r471, %r470, 255;
	xor.b32 	%r472, %r467, %r471;
	cvt.rn.f32.s32 	%f795, %r472;
	mov.f32 	%f796, 0f46fffe00;   	// 32767
	div.full.f32 	%f797, %f795, %f796;
	mul.f32 	%f798, %f457, %f797;
	mov.f32 	%f799, 0f3f000000;   	// 0.5
	mul.f32 	%f800, %f798, %f799;
	sub.f32 	%f369, %f800, %f778;
	bra.uni 	$Lt_1_124162;
$Lt_1_124418:
	.loc	17	117	0
	cvt.rzi.u32.f32 	%r473, %f779;
	shr.u32 	%r474, %r473, 13;
	sub.u32 	%r475, %r12, %r473;
	sub.u32 	%r476, %r368, %r473;
	xor.b32 	%r477, %r476, %r474;
	shl.b32 	%r478, %r477, 8;
	sub.u32 	%r479, %r475, %r477;
	sub.u32 	%r480, %r473, %r477;
	xor.b32 	%r481, %r479, %r478;
	shr.u32 	%r482, %r481, 13;
	sub.u32 	%r483, %r480, %r481;
	sub.u32 	%r484, %r477, %r481;
	xor.b32 	%r485, %r483, %r482;
	shr.u32 	%r486, %r485, 12;
	sub.u32 	%r487, %r484, %r485;
	xor.b32 	%r488, %r487, %r486;
	shl.b32 	%r489, %r488, 16;
	sub.u32 	%r490, %r481, %r485;
	sub.u32 	%r491, %r490, %r488;
	xor.b32 	%r492, %r489, %r491;
	.loc	17	118	0
	sub.u32 	%r493, %r485, %r488;
	sub.u32 	%r494, %r493, %r492;
	shr.u32 	%r495, %r492, 5;
	xor.b32 	%r496, %r494, %r495;
	.loc	17	119	0
	sub.u32 	%r497, %r488, %r492;
	sub.u32 	%r498, %r497, %r496;
	shr.u32 	%r499, %r496, 3;
	xor.b32 	%r500, %r498, %r499;
	.loc	17	120	0
	sub.u32 	%r501, %r492, %r496;
	sub.u32 	%r502, %r501, %r500;
	shl.b32 	%r503, %r500, 10;
	xor.b32 	%r504, %r502, %r503;
	.loc	17	121	0
	sub.u32 	%r505, %r496, %r500;
	sub.u32 	%r506, %r505, %r504;
	shr.u32 	%r507, %r504, 15;
	xor.b32 	%r508, %r506, %r507;
	.loc	31	54	0
	mul.lo.u32 	%r509, %r508, 1103515245;
	mul.lo.u32 	%r510, %r508, -1029531031;
	add.u32 	%r511, %r509, 12345;
	sub.u32 	%r512, %r510, 740551042;
	shr.u32 	%r513, %r511, 16;
	shr.u32 	%r514, %r512, 16;
	and.b32 	%r515, %r513, 255;
	and.b32 	%r516, %r514, 255;
	shl.b32 	%r517, %r515, 7;
	xor.b32 	%r518, %r517, %r516;
	cvt.rn.f32.s32 	%f801, %r518;
	mov.f32 	%f802, 0f46fffe00;   	// 32767
	div.full.f32 	%f803, %f801, %f802;
	mul.f32 	%f804, %f803, %f457;
	mov.f32 	%f805, 0f3f000000;   	// 0.5
	mul.f32 	%f806, %f804, %f805;
	sub.f32 	%f788, %f806, %f778;
	mov.f32 	%f369, %f788;
	mov.f32 	%f362, %f788;
$Lt_1_124162:
	.loc	31	57	0
	add.f32 	%f7, %f788, %f7;
	.loc	31	58	0
	add.f32 	%f6, %f362, %f6;
	.loc	31	59	0
	add.f32 	%f5, %f5, %f369;
	ld.param.f32 	%f807, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter1+12];
	mov.f32 	%f808, 0f3f800000;   	// 1
	setp.eq.f32 	%p114, %f807, %f808;
	@!%p114 bra 	$Lt_1_514;
	.loc	31	63	0
	mov.f32 	%f809, 0f00000000;   	// 0
	max.f32 	%f810, %f5, %f809;
	mov.f32 	%f811, 0f3f800000;   	// 1
	min.f32 	%f5, %f810, %f811;
	mov.f32 	%f812, 0f00000000;   	// 0
	max.f32 	%f813, %f6, %f812;
	mov.f32 	%f814, 0f3f800000;   	// 1
	min.f32 	%f6, %f813, %f814;
	mov.f32 	%f815, 0f00000000;   	// 0
	max.f32 	%f816, %f7, %f815;
	mov.f32 	%f817, 0f3f800000;   	// 1
	min.f32 	%f7, %f816, %f817;
	mov.f32 	%f818, 0f00000000;   	// 0
	max.f32 	%f819, %f8, %f818;
	mov.f32 	%f820, 0f3f800000;   	// 1
	min.f32 	%f8, %f819, %f820;
	bra.uni 	$Lt_1_514;
$Lt_1_4098:
	.loc	35	99	0
	cvt.rn.f32.s32 	%f821, %r6;
	cvt.rn.f32.s32 	%f822, %r14;
	ld.param.f32 	%f823, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter1+20];
	mul.f32 	%f824, %f822, %f823;
	setp.lt.f32 	%p115, %f821, %f824;
	@!%p115 bra 	$Lt_1_514;
	.loc	19	267	0
	ld.const.f32 	%f825, [kRGB32f_To_601YPbPr+4];
	mul.f32 	%f468, %f825, %f6;
	ld.const.f32 	%f826, [kRGB32f_To_601YPbPr+28];
	mul.f32 	%f827, %f826, %f6;
	ld.const.f32 	%f828, [kRGB32f_To_601YPbPr+16];
	mul.f32 	%f829, %f828, %f6;
	ld.const.f32 	%f830, [kRGB32f_To_601YPbPr+0];
	mad.f32 	%f470, %f830, %f7, %f468;
	ld.const.f32 	%f831, [kRGB32f_To_601YPbPr+24];
	mad.f32 	%f832, %f831, %f7, %f827;
	ld.const.f32 	%f833, [kRGB32f_To_601YPbPr+12];
	mad.f32 	%f834, %f833, %f7, %f829;
	ld.const.f32 	%f835, [kRGB32f_To_601YPbPr+8];
	mad.f32 	%f472, %f835, %f5, %f470;
	ld.const.f32 	%f836, [kRGB32f_To_601YPbPr+32];
	mad.f32 	%f837, %f836, %f5, %f832;
	ld.const.f32 	%f838, [kRGB32f_To_601YPbPr+20];
	mad.f32 	%f839, %f838, %f5, %f834;
	ld.param.f32 	%f840, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter1+4];
	ld.param.f32 	%f841, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter1+8];
	mad.f32 	%f842, %f472, %f841, %f840;
	ld.param.f32 	%f843, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter1+16];
	mul.f32 	%f844, %f837, %f843;
	ld.param.f32 	%f845, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter1+12];
	mul.f32 	%f846, %f837, %f845;
	mul.f32 	%f847, %f839, %f845;
	sub.f32 	%f848, %f847, %f844;
	mad.f32 	%f849, %f839, %f843, %f846;
	ld.const.f32 	%f850, [k601YPbPr_To_RGB32f+16];
	mul.f32 	%f851, %f850, %f848;
	ld.const.f32 	%f852, [k601YPbPr_To_RGB32f+12];
	mad.f32 	%f853, %f852, %f842, %f851;
	ld.const.f32 	%f854, [k601YPbPr_To_RGB32f+20];
	mad.f32 	%f855, %f854, %f849, %f853;
	.loc	19	268	0
	ld.const.f32 	%f856, [k601YPbPr_To_RGB32f+4];
	mul.f32 	%f857, %f856, %f848;
	ld.const.f32 	%f858, [k601YPbPr_To_RGB32f+0];
	mad.f32 	%f859, %f858, %f842, %f857;
	ld.const.f32 	%f860, [k601YPbPr_To_RGB32f+8];
	mad.f32 	%f861, %f860, %f849, %f859;
	.loc	32	56	0
	ld.const.f32 	%f862, [k601YPbPr_To_RGB32f+28];
	mul.f32 	%f863, %f862, %f848;
	ld.const.f32 	%f864, [k601YPbPr_To_RGB32f+24];
	mad.f32 	%f865, %f864, %f842, %f863;
	ld.const.f32 	%f866, [k601YPbPr_To_RGB32f+32];
	mad.f32 	%f5, %f866, %f849, %f865;
	mov.f32 	%f6, %f855;
	mov.f32 	%f7, %f861;
	bra.uni 	$Lt_1_514;
$Lt_1_4354:
	.loc	33	46	0
	ld.const.f32 	%f867, [kRGB32f_To_601YPbPr+4];
	mul.f32 	%f468, %f867, %f6;
	ld.const.f32 	%f868, [kRGB32f_To_601YPbPr+0];
	mad.f32 	%f470, %f868, %f7, %f468;
	ld.const.f32 	%f869, [kRGB32f_To_601YPbPr+8];
	mad.f32 	%f472, %f869, %f5, %f470;
	ld.param.f32 	%f482, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter1+8];
	ld.param.f32 	%f870, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter1+20];
	sub.f32 	%f871, %f870, %f482;
	mad.f32 	%f872, %f472, %f871, %f482;
	.loc	33	47	0
	ld.param.f32 	%f873, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter1+12];
	ld.param.f32 	%f874, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter1+24];
	sub.f32 	%f875, %f874, %f873;
	mad.f32 	%f876, %f472, %f875, %f873;
	.loc	33	49	0
	ld.param.f32 	%f877, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter1+28];
	ld.param.f32 	%f457, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter1+4];
	ld.param.f32 	%f878, [__cudaparm_PointwiseFilterHostKernel2___val_paraminFilter1+16];
	sub.f32 	%f879, %f878, %f457;
	mad.f32 	%f880, %f472, %f879, %f457;
	sub.f32 	%f881, %f880, %f5;
	mad.f32 	%f5, %f877, %f881, %f5;
	.loc	33	50	0
	sub.f32 	%f882, %f872, %f6;
	mad.f32 	%f6, %f877, %f882, %f6;
	.loc	33	51	0
	sub.f32 	%f883, %f876, %f7;
	mad.f32 	%f7, %f877, %f883, %f7;
	.loc	35	103	0
	bra.uni 	$Lt_1_514;
$Lt_1_4610:
	.loc	34	41	0
	sub.s32 	%r519, %r13, %r12;
	sub.s32 	%r12, %r519, 1;
$Lt_1_514:
	.loc	17	100	0
	ld.param.s32 	%r520, [__cudaparm_PointwiseFilterHostKernel2_inDestPitch];
	mul24.lo.s32 	%r521, %r12, %r520;
	ld.param.u64 	%rd5, [__cudaparm_PointwiseFilterHostKernel2_inDestImage];
	add.s32 	%r522, %r521, %r6;
	cvt.s64.s32 	%rd6, %r522;
	mul.lo.u64 	%rd7, %rd6, 16;
	add.u64 	%rd8, %rd5, %rd7;
	st.global.v4.f32 	[%rd8+0], {%f5,%f6,%f7,%f8};
$Lt_1_103682:
	.loc	35	157	0
	exit;
$LDWend_PointwiseFilterHostKernel2:
	} // PointwiseFilterHostKernel2

	.entry PointwiseFilterHostKernel3 (
		.param .u64 __cudaparm_PointwiseFilterHostKernel3_inSrcImage,
		.param .s32 __cudaparm_PointwiseFilterHostKernel3_inSrcPitch,
		.param .u64 __cudaparm_PointwiseFilterHostKernel3_inDestImage,
		.param .s32 __cudaparm_PointwiseFilterHostKernel3_inDestPitch,
		.param .s32 __cudaparm_PointwiseFilterHostKernel3_inWidth,
		.param .s32 __cudaparm_PointwiseFilterHostKernel3_inHeight,
		.param .align 4 .b8 __cudaparm_PointwiseFilterHostKernel3___val_paraminFilter0[48],
		.param .align 4 .b8 __cudaparm_PointwiseFilterHostKernel3___val_paraminFilter1[48],
		.param .align 4 .b8 __cudaparm_PointwiseFilterHostKernel3___val_paraminFilter2[48])
	{
	.reg .u32 %r<718>;
	.reg .u64 %rd<10>;
	.reg .f32 %f<1300>;
	.reg .pred %p<174>;
	.loc	35	167	0
$LBB1_PointwiseFilterHostKernel3:
	.loc	35	169	0
	cvt.s32.u16 	%r1, %ctaid.x;
	cvt.s32.u16 	%r2, %ntid.x;
	mul24.lo.s32 	%r3, %r1, %r2;
	cvt.u32.u16 	%r4, %tid.x;
	add.u32 	%r5, %r3, %r4;
	mov.s32 	%r6, %r5;
	.loc	35	170	0
	cvt.s32.u16 	%r7, %ctaid.y;
	cvt.s32.u16 	%r8, %ntid.y;
	mul24.lo.s32 	%r9, %r7, %r8;
	cvt.u32.u16 	%r10, %tid.y;
	add.u32 	%r11, %r9, %r10;
	mov.s32 	%r12, %r11;
	ld.param.s32 	%r13, [__cudaparm_PointwiseFilterHostKernel3_inHeight];
	ld.param.s32 	%r14, [__cudaparm_PointwiseFilterHostKernel3_inWidth];
	set.gt.u32.s32 	%r15, %r13, %r11;
	neg.s32 	%r16, %r15;
	set.gt.u32.s32 	%r17, %r14, %r5;
	neg.s32 	%r18, %r17;
	and.b32 	%r19, %r16, %r18;
	mov.u32 	%r20, 0;
	setp.eq.s32 	%p1, %r19, %r20;
	@%p1 bra 	$Lt_2_154882;
	.loc	17	89	0
	ld.param.s32 	%r21, [__cudaparm_PointwiseFilterHostKernel3_inSrcPitch];
	mul24.lo.s32 	%r22, %r11, %r21;
	ld.param.u64 	%rd1, [__cudaparm_PointwiseFilterHostKernel3_inSrcImage];
	add.s32 	%r23, %r22, %r5;
	cvt.s64.s32 	%rd2, %r23;
	mul.lo.u64 	%rd3, %rd2, 16;
	add.u64 	%rd4, %rd1, %rd3;
	ld.global.v4.f32 	{%f1,%f2,%f3,%f4}, [%rd4+0];
	.loc	35	174	0
	mov.f32 	%f5, %f1;
	mov.f32 	%f6, %f2;
	mov.f32 	%f7, %f3;
	mov.f32 	%f8, %f4;
	.loc	35	54	0
	ld.param.u32 	%r24, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter0+0];
	mov.u32 	%r25, 0;
	setp.eq.s32 	%p2, %r24, %r25;
	@%p2 bra 	$Lt_2_9986;
	mov.u32 	%r26, 1;
	setp.eq.s32 	%p3, %r24, %r26;
	@%p3 bra 	$Lt_2_10498;
	mov.u32 	%r27, 2;
	setp.eq.s32 	%p4, %r24, %r27;
	@%p4 bra 	$Lt_2_10754;
	mov.u32 	%r28, 3;
	setp.eq.s32 	%p5, %r24, %r28;
	@%p5 bra 	$Lt_2_11010;
	mov.u32 	%r29, 4;
	setp.eq.s32 	%p6, %r24, %r29;
	@%p6 bra 	$Lt_2_11266;
	mov.u32 	%r30, 5;
	setp.eq.s32 	%p7, %r24, %r30;
	@%p7 bra 	$Lt_2_11522;
	mov.u32 	%r31, 6;
	setp.eq.s32 	%p8, %r24, %r31;
	@%p8 bra 	$Lt_2_11778;
	mov.u32 	%r32, 7;
	setp.eq.s32 	%p9, %r24, %r32;
	@%p9 bra 	$Lt_2_12034;
	mov.u32 	%r33, 8;
	setp.eq.s32 	%p10, %r24, %r33;
	@%p10 bra 	$Lt_2_12290;
	mov.u32 	%r34, 9;
	setp.eq.s32 	%p11, %r24, %r34;
	@%p11 bra 	$Lt_2_12546;
	mov.u32 	%r35, 10;
	setp.eq.s32 	%p12, %r24, %r35;
	@%p12 bra 	$Lt_2_12802;
	mov.u32 	%r36, 11;
	setp.eq.s32 	%p13, %r24, %r36;
	@%p13 bra 	$Lt_2_13058;
	mov.u32 	%r37, 12;
	setp.eq.s32 	%p14, %r24, %r37;
	@%p14 bra 	$Lt_2_13314;
	mov.u32 	%r38, 13;
	setp.eq.s32 	%p15, %r24, %r38;
	@%p15 bra 	$Lt_2_13570;
	mov.u32 	%r39, 14;
	setp.eq.s32 	%p16, %r24, %r39;
	@%p16 bra 	$Lt_2_13826;
	mov.u32 	%r40, 15;
	setp.eq.s32 	%p17, %r24, %r40;
	@%p17 bra 	$Lt_2_14082;
	mov.u32 	%r41, 16;
	setp.eq.s32 	%p18, %r24, %r41;
	@%p18 bra 	$Lt_2_14338;
	bra.uni 	$Lt_2_10242;
$Lt_2_9986:
	.loc	35	57	0
	ld.param.f32 	%f9, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter0+8];
	mov.f32 	%f10, 0f3f800000;    	// 1
	setp.eq.f32 	%p19, %f9, %f10;
	ld.param.f32 	%f11, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter0+12];
	mov.f32 	%f12, 0f3f800000;    	// 1
	setp.eq.f32 	%p20, %f11, %f12;
	ld.param.f32 	%f13, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter0+4];
	ld.param.f32 	%f14, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter0+16];
	mov.f32 	%f15, 0f3f800000;    	// 1
	setp.eq.f32 	%p21, %f14, %f15;
	@!%p21 bra 	$Lt_2_155650;
	.loc	18	53	0
	cvt.sat.f32.f32 	%f16, %f4;
	mul.f32 	%f17, %f16, %f13;
	selp.f32 	%f18, %f13, %f17, %p19;
	mov.f32 	%f19, 0f3f800000;    	// 1
	sub.f32 	%f20, %f19, %f18;
	selp.f32 	%f21, %f20, %f18, %p20;
	mov.f32 	%f8, %f21;
	.loc	18	57	0
	mov.f32 	%f5, %f21;
	mov.f32 	%f6, %f21;
	mov.f32 	%f7, %f21;
	bra.uni 	$Lt_2_10242;
$Lt_2_155650:
	@!%p19 bra 	$Lt_2_156162;
	.loc	18	61	0
	mov.f32 	%f22, 0f3f800000;    	// 1
	sub.f32 	%f23, %f22, %f13;
	selp.f32 	%f8, %f23, %f13, %p20;
	bra.uni 	$Lt_2_10242;
$Lt_2_156162:
	.loc	18	73	0
	cvt.sat.f32.f32 	%f24, %f4;
	mov.f32 	%f25, 0f3f800000;    	// 1
	sub.f32 	%f26, %f25, %f24;
	selp.f32 	%f27, %f26, %f24, %p20;
	mul.f32 	%f8, %f27, %f13;
	bra.uni 	$Lt_2_10242;
$Lt_2_10498:
	.loc	19	267	0
	ld.const.f32 	%f28, [kRGB32f_To_601YPbPr+4];
	mul.f32 	%f29, %f28, %f2;
	ld.const.f32 	%f30, [kRGB32f_To_601YPbPr+0];
	mad.f32 	%f31, %f30, %f3, %f29;
	ld.const.f32 	%f32, [kRGB32f_To_601YPbPr+8];
	mad.f32 	%f33, %f32, %f1, %f31;
	ld.const.f32 	%f34, [k601YPbPr_To_RGB32f+12];
	mul.f32 	%f35, %f34, %f33;
	.loc	19	268	0
	ld.const.f32 	%f36, [k601YPbPr_To_RGB32f+0];
	mul.f32 	%f37, %f36, %f33;
	.loc	20	44	0
	ld.const.f32 	%f38, [k601YPbPr_To_RGB32f+24];
	mul.f32 	%f5, %f38, %f33;
	mov.f32 	%f6, %f35;
	mov.f32 	%f7, %f37;
	mov.f32 	%f8, %f4;
	.loc	35	61	0
	bra.uni 	$Lt_2_10242;
$Lt_2_10754:
	.loc	35	63	0
	ld.param.f32 	%f39, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter0+20];
	ld.param.f32 	%f40, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter0+12];
	ld.param.f32 	%f41, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter0+4];
	mov.f32 	%f42, 0f00000000;    	// 0
	setp.neu.f32 	%p22, %f41, %f42;
	@!%p22 bra 	$Lt_2_156674;
	.loc	21	44	0
	ld.param.f32 	%f43, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter0+8];
	sub.f32 	%f44, %f3, %f40;
	mad.f32 	%f7, %f43, %f44, %f39;
	.loc	21	45	0
	sub.f32 	%f45, %f2, %f40;
	mad.f32 	%f6, %f43, %f45, %f39;
	.loc	21	46	0
	sub.f32 	%f46, %f1, %f40;
	mad.f32 	%f5, %f43, %f46, %f39;
	bra.uni 	$Lt_2_10242;
$Lt_2_156674:
	.loc	21	50	0
	ld.param.f32 	%f47, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter0+24];
	setp.gt.f32 	%p23, %f3, %f40;
	selp.f32 	%f7, %f47, %f39, %p23;
	.loc	21	51	0
	setp.gt.f32 	%p24, %f2, %f40;
	selp.f32 	%f6, %f47, %f39, %p24;
	.loc	21	52	0
	setp.gt.f32 	%p25, %f1, %f40;
	selp.f32 	%f5, %f47, %f39, %p25;
	bra.uni 	$Lt_2_10242;
$Lt_2_11010:
	.loc	22	47	0
	ld.param.f32 	%f48, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter0+4];
	mul.f32 	%f5, %f48, %f1;
	ld.param.f32 	%f49, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter0+8];
	mul.f32 	%f6, %f49, %f2;
	ld.param.f32 	%f50, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter0+12];
	mul.f32 	%f7, %f50, %f3;
	.loc	35	67	0
	bra.uni 	$Lt_2_10242;
$Lt_2_11266:
	.loc	23	48	0
	ld.param.f32 	%f51, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter0+4];
	mov.f32 	%f52, 0f00000000;    	// 0
	setp.eq.f32 	%p26, %f51, %f52;
	ld.param.f32 	%f53, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter0+8];
	mov.f32 	%f54, 0f00000000;    	// 0
	max.f32 	%f55, %f1, %f54;
	mov.f32 	%f56, 0f3f800000;    	// 1
	min.f32 	%f57, %f55, %f56;
	ld.param.f32 	%f58, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter0+12];
	sub.f32 	%f59, %f57, %f58;
	abs.f32 	%f60, %f59;
	mov.f32 	%f61, 0f00000000;    	// 0
	max.f32 	%f62, %f2, %f61;
	mov.f32 	%f63, 0f3f800000;    	// 1
	min.f32 	%f64, %f62, %f63;
	ld.param.f32 	%f65, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter0+16];
	sub.f32 	%f66, %f64, %f65;
	abs.f32 	%f67, %f66;
	mov.f32 	%f68, 0f00000000;    	// 0
	max.f32 	%f69, %f4, %f68;
	mov.f32 	%f70, 0f3f800000;    	// 1
	min.f32 	%f71, %f69, %f70;
	sub.f32 	%f72, %f71, %f4;
	abs.f32 	%f73, %f72;
	mov.f32 	%f74, 0f00000000;    	// 0
	max.f32 	%f75, %f3, %f74;
	mov.f32 	%f76, 0f3f800000;    	// 1
	min.f32 	%f77, %f75, %f76;
	ld.param.f32 	%f78, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter0+20];
	sub.f32 	%f79, %f77, %f78;
	abs.f32 	%f80, %f79;
	max.f32 	%f81, %f73, %f80;
	max.f32 	%f82, %f67, %f81;
	max.f32 	%f83, %f60, %f82;
	setp.ge.f32 	%p27, %f53, %f83;
	xor.pred 	%p28, %p26, %p27;
	@!%p28 bra 	$Lt_2_10242;
	.loc	19	267	0
	ld.const.f32 	%f84, [kRGB32f_To_601YPbPr+4];
	mul.f32 	%f29, %f84, %f2;
	ld.const.f32 	%f85, [kRGB32f_To_601YPbPr+0];
	mad.f32 	%f31, %f85, %f3, %f29;
	ld.const.f32 	%f86, [kRGB32f_To_601YPbPr+8];
	mad.f32 	%f33, %f86, %f1, %f31;
	ld.const.f32 	%f87, [k601YPbPr_To_RGB32f+12];
	mul.f32 	%f88, %f87, %f33;
	.loc	19	268	0
	ld.const.f32 	%f89, [k601YPbPr_To_RGB32f+0];
	mul.f32 	%f90, %f89, %f33;
	.loc	20	44	0
	ld.const.f32 	%f91, [k601YPbPr_To_RGB32f+24];
	mul.f32 	%f5, %f91, %f33;
	mov.f32 	%f6, %f88;
	mov.f32 	%f7, %f90;
	bra.uni 	$Lt_2_10242;
$Lt_2_11522:
	.loc	24	48	0
	ld.param.f32 	%f92, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter0+8];
	mov.f32 	%f93, 0f00000000;    	// 0
	max.f32 	%f94, %f1, %f93;
	mov.f32 	%f95, 0f3f800000;    	// 1
	min.f32 	%f96, %f94, %f95;
	ld.param.f32 	%f97, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter0+12];
	sub.f32 	%f98, %f96, %f97;
	abs.f32 	%f99, %f98;
	mov.f32 	%f100, 0f00000000;   	// 0
	max.f32 	%f101, %f2, %f100;
	mov.f32 	%f102, 0f3f800000;   	// 1
	min.f32 	%f103, %f101, %f102;
	ld.param.f32 	%f104, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter0+16];
	sub.f32 	%f105, %f103, %f104;
	abs.f32 	%f106, %f105;
	mov.f32 	%f107, 0f00000000;   	// 0
	max.f32 	%f108, %f4, %f107;
	mov.f32 	%f109, 0f3f800000;   	// 1
	min.f32 	%f110, %f108, %f109;
	sub.f32 	%f111, %f110, %f4;
	abs.f32 	%f112, %f111;
	mov.f32 	%f113, 0f00000000;   	// 0
	max.f32 	%f114, %f3, %f113;
	mov.f32 	%f115, 0f3f800000;   	// 1
	min.f32 	%f116, %f114, %f115;
	ld.param.f32 	%f117, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter0+20];
	sub.f32 	%f118, %f116, %f117;
	abs.f32 	%f119, %f118;
	max.f32 	%f120, %f112, %f119;
	max.f32 	%f121, %f106, %f120;
	max.f32 	%f122, %f99, %f121;
	setp.ge.f32 	%p29, %f92, %f122;
	@!%p29 bra 	$Lt_2_10242;
	.loc	24	51	0
	ld.param.f32 	%f123, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter0+24];
	mov.f32 	%f124, %f123;
	ld.param.f32 	%f125, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter0+28];
	mov.f32 	%f126, %f125;
	ld.param.f32 	%f127, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter0+32];
	mov.f32 	%f128, %f127;
	ld.param.f32 	%f129, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter0+4];
	mov.f32 	%f130, 0f00000000;   	// 0
	setp.eq.f32 	%p30, %f129, %f130;
	@!%p30 bra 	$Lt_2_157954;
	.loc	24	60	0
	ld.const.f32 	%f131, [kRGB32f_To_601YPbPr+4];
	mul.f32 	%f29, %f131, %f2;
	ld.const.f32 	%f132, [kRGB32f_To_601YPbPr+0];
	mad.f32 	%f31, %f132, %f3, %f29;
	ld.const.f32 	%f133, [kRGB32f_To_601YPbPr+8];
	mad.f32 	%f33, %f133, %f1, %f31;
	mul.f32 	%f124, %f33, %f123;
	.loc	24	61	0
	mul.f32 	%f126, %f33, %f125;
	.loc	24	62	0
	mul.f32 	%f128, %f33, %f127;
$Lt_2_157954:
	.loc	24	65	0
	mov.f32 	%f5, %f124;
	mov.f32 	%f6, %f126;
	mov.f32 	%f7, %f128;
	bra.uni 	$Lt_2_10242;
$Lt_2_11778:
	.loc	25	47	0
	sub.s32 	%r42, %r13, %r11;
	sub.s32 	%r43, %r14, %r5;
	cvt.rn.f32.s32 	%f134, %r5;
	cvt.rn.f32.s32 	%f135, %r11;
	cvt.rn.f32.s32 	%f136, %r42;
	cvt.rn.f32.s32 	%f137, %r43;
	ld.param.f32 	%f13, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter0+4];
	sub.f32 	%f138, %f13, %f134;
	sub.f32 	%f139, %f13, %f135;
	sub.f32 	%f140, %f13, %f136;
	sub.f32 	%f141, %f13, %f137;
	cvt.rzi.s32.f32 	%r44, %f138;
	cvt.rzi.s32.f32 	%r45, %f139;
	cvt.rzi.s32.f32 	%r46, %f140;
	cvt.rzi.s32.f32 	%r47, %f141;
	max.s32 	%r48, %r44, %r45;
	max.s32 	%r49, %r47, %r48;
	max.s32 	%r50, %r46, %r49;
	mov.u32 	%r51, 0;
	setp.le.s32 	%p31, %r50, %r51;
	@%p31 bra 	$Lt_2_10242;
	.loc	25	51	0
	cvt.rn.f32.s32 	%f142, %r50;
	sub.f32 	%f143, %f13, %f142;
	div.full.f32 	%f144, %f143, %f13;
	mul.f32 	%f8, %f4, %f144;
	bra.uni 	$Lt_2_10242;
$Lt_2_12034:
	.loc	27	50	0
	ld.const.f32 	%f145, [kRGB32f_To_601YPbPr+4];
	mul.f32 	%f29, %f145, %f2;
	ld.param.f32 	%f146, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter0+12];
	ld.param.f32 	%f43, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter0+8];
	sub.f32 	%f147, %f146, %f43;
	ld.const.f32 	%f148, [kRGB32f_To_601YPbPr+0];
	mad.f32 	%f31, %f148, %f3, %f29;
	ld.param.f32 	%f149, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter0+16];
	mul.f32 	%f150, %f149, %f147;
	ld.const.f32 	%f151, [kRGB32f_To_601YPbPr+8];
	mad.f32 	%f33, %f151, %f1, %f31;
	add.f32 	%f152, %f150, %f146;
	sub.f32 	%f153, %f43, %f150;
	mov.f32 	%f154, 0f00000000;   	// 0
	max.f32 	%f155, %f152, %f154;
	mov.f32 	%f156, 0f00000000;   	// 0
	max.f32 	%f157, %f153, %f156;
	mov.f32 	%f158, 0f3f800000;   	// 1
	min.f32 	%f159, %f155, %f158;
	mov.f32 	%f160, 0f3f800000;   	// 1
	min.f32 	%f161, %f157, %f160;
	set.gt.u32.f32 	%r52, %f161, %f33;
	neg.s32 	%r53, %r52;
	set.le.u32.f32 	%r54, %f159, %f33;
	neg.s32 	%r55, %r54;
	or.b32 	%r56, %r53, %r55;
	mov.u32 	%r57, 0;
	setp.eq.s32 	%p32, %r56, %r57;
	@%p32 bra 	$Lt_2_159234;
	mov.f32 	%f162, 0f00000000;   	// 0
	bra.uni 	$Lt_2_160002;
$Lt_2_159234:
	add.f32 	%f163, %f150, %f43;
	mov.f32 	%f164, 0f00000000;   	// 0
	max.f32 	%f165, %f163, %f164;
	mov.f32 	%f166, 0f3f800000;   	// 1
	min.f32 	%f167, %f165, %f166;
	set.le.u32.f32 	%r58, %f167, %f33;
	neg.s32 	%r59, %r58;
	sub.f32 	%f168, %f146, %f150;
	mov.f32 	%f169, 0f00000000;   	// 0
	max.f32 	%f170, %f168, %f169;
	mov.f32 	%f171, 0f3f800000;   	// 1
	min.f32 	%f172, %f170, %f171;
	set.lt.u32.f32 	%r60, %f33, %f172;
	neg.s32 	%r61, %r60;
	and.b32 	%r62, %r59, %r61;
	mov.u32 	%r63, 0;
	setp.eq.s32 	%p33, %r62, %r63;
	@%p33 bra 	$Lt_2_159746;
	mov.f32 	%f162, 0f3f800000;   	// 1
	bra.uni 	$Lt_2_160002;
$Lt_2_159746:
	add.f32 	%f173, %f150, %f150;
	setp.gt.f32 	%p34, %f167, %f33;
	@!%p34 bra 	$Lt_2_160258;
	.loc	27	62	0
	sub.f32 	%f174, %f33, %f161;
	div.full.f32 	%f162, %f174, %f173;
	bra.uni 	$Lt_2_160002;
$Lt_2_160258:
	.loc	27	66	0
	sub.f32 	%f175, %f159, %f33;
	div.full.f32 	%f162, %f175, %f173;
$Lt_2_160002:
$Lt_2_159490:
$Lt_2_158978:
	.loc	27	69	0
	mov.f32 	%f176, 0f3f800000;   	// 1
	sub.f32 	%f177, %f176, %f162;
	ld.param.f32 	%f178, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter0+4];
	mov.f32 	%f179, 0f00000000;   	// 0
	setp.neu.f32 	%p35, %f178, %f179;
	selp.f32 	%f162, %f177, %f162, %p35;
	.loc	27	77	0
	ld.const.f32 	%f180, [k601YPbPr_To_RGB32f+24];
	mul.f32 	%f5, %f180, %f162;
	ld.const.f32 	%f181, [k601YPbPr_To_RGB32f+12];
	mul.f32 	%f6, %f181, %f162;
	ld.const.f32 	%f182, [k601YPbPr_To_RGB32f+0];
	mul.f32 	%f7, %f182, %f162;
	.loc	35	79	0
	bra.uni 	$Lt_2_10242;
$Lt_2_12290:
	.loc	35	80	0
	ld.param.f32 	%f183, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter0+44];
	mov.f32 	%f184, 0f00000000;   	// 0
	setp.ge.f32 	%p36, %f183, %f184;
	@!%p36 bra 	$L_2_148482;
	cvt.rn.f32.s32 	%f185, %r5;
	cvt.rn.f32.s32 	%f186, %r14;
	mul.f32 	%f187, %f186, %f183;
	setp.lt.f32 	%p37, %f185, %f187;
	@%p37 bra 	$L_2_148226;
$L_2_148482:
	mov.f32 	%f188, 0f00000000;   	// 0
	setp.lt.f32 	%p38, %f183, %f188;
	@!%p38 bra 	$Lt_2_10242;
	cvt.rn.f32.s32 	%f189, %r11;
	cvt.rn.f32.s32 	%f190, %r13;
	mul.f32 	%f191, %f190, %f183;
	neg.f32 	%f192, %f191;
	setp.lt.f32 	%p39, %f189, %f192;
	@!%p39 bra 	$Lt_2_10242;
$L_2_148226:
	.loc	28	47	0
	ld.const.f32 	%f193, [kRGB32f_To_601YPbPr+4];
	mul.f32 	%f194, %f193, %f2;
	ld.const.f32 	%f195, [kRGB32f_To_601YPbPr+0];
	mad.f32 	%f196, %f195, %f3, %f194;
	ld.const.f32 	%f197, [kRGB32f_To_601YPbPr+8];
	mad.f32 	%f33, %f197, %f1, %f196;
	mov.f32 	%f198, %f33;
	ld.param.f32 	%f43, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter0+8];
	ld.param.f32 	%f199, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter0+4];
	sub.f32 	%f200, %f199, %f43;
	ld.param.f32 	%f201, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter0+16];
	ld.param.f32 	%f202, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter0+12];
	sub.f32 	%f203, %f202, %f201;
	mov.f32 	%f204, 0f3f800000;   	// 1
	set.neu.u32.f32 	%r64, %f203, %f204;
	neg.s32 	%r65, %r64;
	mov.f32 	%f205, 0f3f800000;   	// 1
	set.neu.u32.f32 	%r66, %f200, %f205;
	neg.s32 	%r67, %r66;
	or.b32 	%r68, %r65, %r67;
	mov.u32 	%r69, 0;
	setp.eq.s32 	%p40, %r68, %r69;
	@%p40 bra 	$Lt_2_160514;
	.loc	17	116	0
	mov.u32 	%r70, 1;
	sub.u32 	%r71, %r70, %r5;
	shr.u32 	%r72, %r11, 13;
	sub.u32 	%r73, %r5, %r11;
	sub.u32 	%r74, %r71, %r11;
	xor.b32 	%r75, %r74, %r72;
	shl.b32 	%r76, %r75, 8;
	sub.u32 	%r77, %r73, %r75;
	sub.u32 	%r78, %r11, %r75;
	xor.b32 	%r79, %r77, %r76;
	shr.u32 	%r80, %r79, 13;
	sub.u32 	%r81, %r78, %r79;
	xor.b32 	%r82, %r81, %r80;
	shr.u32 	%r83, %r82, 12;
	sub.u32 	%r84, %r75, %r79;
	sub.u32 	%r85, %r84, %r82;
	xor.b32 	%r86, %r83, %r85;
	.loc	17	117	0
	sub.u32 	%r87, %r79, %r82;
	sub.u32 	%r88, %r87, %r86;
	shl.b32 	%r89, %r86, 16;
	xor.b32 	%r90, %r88, %r89;
	.loc	17	118	0
	sub.u32 	%r91, %r82, %r86;
	sub.u32 	%r92, %r91, %r90;
	shr.u32 	%r93, %r90, 5;
	xor.b32 	%r94, %r92, %r93;
	.loc	17	119	0
	sub.u32 	%r95, %r86, %r90;
	sub.u32 	%r96, %r95, %r94;
	shr.u32 	%r97, %r94, 3;
	xor.b32 	%r98, %r96, %r97;
	.loc	17	120	0
	sub.u32 	%r99, %r90, %r94;
	sub.u32 	%r100, %r99, %r98;
	shl.b32 	%r101, %r98, 10;
	xor.b32 	%r102, %r100, %r101;
	.loc	17	121	0
	sub.u32 	%r103, %r94, %r98;
	sub.u32 	%r104, %r103, %r102;
	shr.u32 	%r105, %r102, 15;
	xor.b32 	%r106, %r104, %r105;
	.loc	28	57	0
	mov.f32 	%f206, 0f3b270d73;   	// 0.00254902
	mul.lo.u32 	%r107, %r106, 1103515245;
	add.u32 	%r108, %r107, 12345;
	shr.u32 	%r109, %r108, 16;
	and.b32 	%r110, %r109, 255;
	shl.b32 	%r111, %r110, 7;
	mul.lo.u32 	%r112, %r106, -1029531031;
	sub.u32 	%r113, %r112, 740551042;
	shr.u32 	%r114, %r113, 16;
	and.b32 	%r115, %r114, 255;
	xor.b32 	%r116, %r111, %r115;
	cvt.rn.f32.s32 	%f207, %r116;
	mov.f32 	%f208, 0f467ffe00;   	// 16383.5
	div.full.f32 	%f209, %f207, %f208;
	mov.f32 	%f210, 0fbf800000;   	// -1
	add.f32 	%f211, %f209, %f210;
	mad.f32 	%f198, %f206, %f211, %f33;
$Lt_2_160514:
	sub.f32 	%f212, %f198, %f43;
	ld.param.f32 	%f213, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter0+20];
	mov.f32 	%f214, 0f3f800000;   	// 1
	setp.neu.f32 	%p41, %f213, %f214;
	@!%p41 bra 	$Lt_2_161282;
	mov.f32 	%f215, 0f00000000;   	// 0
	setp.lt.f32 	%p42, %f212, %f215;
	@!%p42 bra 	$Lt_2_161794;
	.loc	28	66	0
	mov.f32 	%f198, %f201;
	bra.uni 	$Lt_2_161026;
$Lt_2_161794:
	.loc	28	70	0
	rcp.approx.f32 	%f216, %f200;
	mul.f32 	%f217, %f212, %f216;
	lg2.approx.f32 	%f218, %f217;
	mul.f32 	%f219, %f213, %f218;
	ex2.approx.f32 	%f220, %f219;
	mad.f32 	%f198, %f203, %f220, %f201;
	bra.uni 	$Lt_2_161026;
$Lt_2_161282:
	.loc	28	77	0
	rcp.approx.f32 	%f221, %f200;
	mul.f32 	%f222, %f212, %f221;
	mad.f32 	%f198, %f203, %f222, %f201;
$Lt_2_161026:
	.loc	19	267	0
	ld.const.f32 	%f223, [kRGB32f_To_601YPbPr+28];
	mul.f32 	%f224, %f223, %f2;
	ld.const.f32 	%f225, [kRGB32f_To_601YPbPr+16];
	mul.f32 	%f226, %f225, %f2;
	ld.param.f32 	%f227, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter0+28];
	sin.approx.f32 	%f228, %f227;
	cos.approx.f32 	%f229, %f227;
	ld.const.f32 	%f230, [kRGB32f_To_601YPbPr+24];
	mad.f32 	%f231, %f230, %f3, %f224;
	ld.const.f32 	%f232, [kRGB32f_To_601YPbPr+12];
	mad.f32 	%f233, %f232, %f3, %f226;
	ld.const.f32 	%f234, [kRGB32f_To_601YPbPr+32];
	mad.f32 	%f235, %f234, %f1, %f231;
	ld.const.f32 	%f236, [kRGB32f_To_601YPbPr+20];
	mad.f32 	%f237, %f236, %f1, %f233;
	mul.f32 	%f238, %f228, %f235;
	mul.f32 	%f239, %f229, %f235;
	mul.f32 	%f240, %f237, %f229;
	sub.f32 	%f241, %f240, %f238;
	mad.f32 	%f242, %f237, %f228, %f239;
	ld.param.f32 	%f243, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter0+32];
	sub.f32 	%f244, %f243, %f241;
	ld.param.f32 	%f245, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter0+36];
	sub.f32 	%f246, %f245, %f242;
	ld.param.f32 	%f247, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter0+40];
	mad.f32 	%f248, %f247, %f244, %f241;
	mad.f32 	%f249, %f247, %f246, %f242;
	ld.param.f32 	%f250, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter0+24];
	mul.f32 	%f251, %f248, %f250;
	mul.f32 	%f252, %f249, %f250;
	ld.const.f32 	%f253, [k601YPbPr_To_RGB32f+16];
	mul.f32 	%f254, %f253, %f251;
	ld.const.f32 	%f255, [k601YPbPr_To_RGB32f+12];
	mad.f32 	%f256, %f255, %f198, %f254;
	ld.const.f32 	%f257, [k601YPbPr_To_RGB32f+20];
	mad.f32 	%f258, %f257, %f252, %f256;
	.loc	19	268	0
	ld.const.f32 	%f259, [k601YPbPr_To_RGB32f+4];
	mul.f32 	%f260, %f259, %f251;
	ld.const.f32 	%f261, [k601YPbPr_To_RGB32f+0];
	mad.f32 	%f262, %f261, %f198, %f260;
	ld.const.f32 	%f263, [k601YPbPr_To_RGB32f+8];
	mad.f32 	%f264, %f263, %f252, %f262;
	.loc	28	92	0
	ld.const.f32 	%f265, [k601YPbPr_To_RGB32f+28];
	mul.f32 	%f266, %f265, %f251;
	ld.const.f32 	%f267, [k601YPbPr_To_RGB32f+24];
	mad.f32 	%f268, %f267, %f198, %f266;
	ld.const.f32 	%f269, [k601YPbPr_To_RGB32f+32];
	mad.f32 	%f5, %f269, %f252, %f268;
	mov.f32 	%f6, %f258;
	mov.f32 	%f7, %f264;
	bra.uni 	$Lt_2_10242;
$Lt_2_12546:
	.loc	35	83	0
	ld.param.f32 	%f270, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter0+44];
	mov.f32 	%f271, 0f00000000;   	// 0
	setp.ge.f32 	%p43, %f270, %f271;
	@!%p43 bra 	$L_2_149250;
	cvt.rn.f32.s32 	%f272, %r5;
	cvt.rn.f32.s32 	%f273, %r14;
	mul.f32 	%f274, %f273, %f270;
	setp.lt.f32 	%p44, %f272, %f274;
	@%p44 bra 	$L_2_148994;
$L_2_149250:
	mov.f32 	%f275, 0f00000000;   	// 0
	setp.lt.f32 	%p45, %f270, %f275;
	@!%p45 bra 	$Lt_2_10242;
	cvt.rn.f32.s32 	%f276, %r11;
	cvt.rn.f32.s32 	%f277, %r13;
	mul.f32 	%f278, %f277, %f270;
	neg.f32 	%f279, %f278;
	setp.lt.f32 	%p46, %f276, %f279;
	@!%p46 bra 	$Lt_2_10242;
$L_2_148994:
	.loc	28	110	0
	ld.const.f32 	%f280, [kRGB32f_To_601YPbPr+4];
	mul.f32 	%f29, %f280, %f2;
	ld.param.f32 	%f43, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter0+8];
	ld.param.f32 	%f281, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter0+4];
	sub.f32 	%f282, %f281, %f43;
	ld.param.f32 	%f283, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter0+16];
	ld.param.f32 	%f284, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter0+12];
	sub.f32 	%f285, %f284, %f283;
	ld.const.f32 	%f286, [kRGB32f_To_601YPbPr+0];
	mad.f32 	%f31, %f286, %f3, %f29;
	ld.const.f32 	%f287, [kRGB32f_To_601YPbPr+8];
	mad.f32 	%f33, %f287, %f1, %f31;
	sub.f32 	%f288, %f33, %f43;
	ld.param.f32 	%f289, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter0+20];
	mov.f32 	%f290, 0f3f800000;   	// 1
	setp.neu.f32 	%p47, %f289, %f290;
	@!%p47 bra 	$Lt_2_162306;
	.loc	28	120	0
	mov.f32 	%f291, 0f00000000;   	// 0
	max.f32 	%f292, %f288, %f291;
	div.full.f32 	%f293, %f292, %f282;
	lg2.approx.f32 	%f294, %f293;
	mul.f32 	%f295, %f289, %f294;
	ex2.approx.f32 	%f296, %f295;
	mad.f32 	%f297, %f285, %f296, %f283;
	bra.uni 	$Lt_2_162050;
$Lt_2_162306:
	.loc	28	129	0
	div.full.f32 	%f298, %f288, %f282;
	mad.f32 	%f297, %f285, %f298, %f283;
$Lt_2_162050:
	.loc	28	135	0
	ld.const.f32 	%f299, [k601YPbPr_To_RGB32f+24];
	mul.f32 	%f5, %f299, %f297;
	ld.const.f32 	%f300, [k601YPbPr_To_RGB32f+12];
	mul.f32 	%f6, %f300, %f297;
	ld.const.f32 	%f301, [k601YPbPr_To_RGB32f+0];
	mul.f32 	%f7, %f301, %f297;
	bra.uni 	$Lt_2_10242;
$Lt_2_12802:
	.loc	35	86	0
	ld.param.f32 	%f302, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter0+44];
	mov.f32 	%f303, 0f00000000;   	// 0
	setp.ge.f32 	%p48, %f302, %f303;
	@!%p48 bra 	$L_2_150018;
	cvt.rn.f32.s32 	%f304, %r5;
	cvt.rn.f32.s32 	%f305, %r14;
	mul.f32 	%f306, %f305, %f302;
	setp.lt.f32 	%p49, %f304, %f306;
	@%p49 bra 	$L_2_149762;
$L_2_150018:
	mov.f32 	%f307, 0f00000000;   	// 0
	setp.lt.f32 	%p50, %f302, %f307;
	@!%p50 bra 	$Lt_2_10242;
	cvt.rn.f32.s32 	%f308, %r11;
	cvt.rn.f32.s32 	%f309, %r13;
	mul.f32 	%f310, %f309, %f302;
	neg.f32 	%f311, %f310;
	setp.lt.f32 	%p51, %f308, %f311;
	@!%p51 bra 	$Lt_2_10242;
$L_2_149762:
	.loc	28	160	0
	ld.const.f32 	%f5, [k601YPbPr_To_RGB32f+24];
	ld.const.f32 	%f6, [k601YPbPr_To_RGB32f+12];
	ld.const.f32 	%f7, [k601YPbPr_To_RGB32f+0];
	bra.uni 	$Lt_2_10242;
$Lt_2_13058:
	.loc	29	42	0
	ld.param.f32 	%f13, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter0+4];
	mov.f32 	%f312, 0f00000000;   	// 0
	setp.ge.f32 	%p52, %f1, %f312;
	@!%p52 bra 	$Lt_2_162818;
	.loc	19	292	0
	lg2.approx.f32 	%f313, %f1;
	mul.f32 	%f314, %f13, %f313;
	ex2.approx.f32 	%f315, %f314;
	bra.uni 	$Lt_2_162562;
$Lt_2_162818:
	neg.f32 	%f316, %f1;
	lg2.approx.f32 	%f317, %f316;
	mul.f32 	%f318, %f13, %f317;
	ex2.approx.f32 	%f319, %f318;
	neg.f32 	%f315, %f319;
$Lt_2_162562:
	mov.f32 	%f320, 0f00000000;   	// 0
	setp.ge.f32 	%p53, %f2, %f320;
	@!%p53 bra 	$Lt_2_163330;
	.loc	19	293	0
	lg2.approx.f32 	%f321, %f2;
	mul.f32 	%f322, %f13, %f321;
	ex2.approx.f32 	%f323, %f322;
	bra.uni 	$Lt_2_163074;
$Lt_2_163330:
	neg.f32 	%f324, %f2;
	lg2.approx.f32 	%f325, %f324;
	mul.f32 	%f326, %f13, %f325;
	ex2.approx.f32 	%f327, %f326;
	neg.f32 	%f323, %f327;
$Lt_2_163074:
	mov.f32 	%f328, 0f00000000;   	// 0
	setp.ge.f32 	%p54, %f3, %f328;
	@!%p54 bra 	$Lt_2_163842;
	.loc	19	294	0
	lg2.approx.f32 	%f329, %f3;
	mul.f32 	%f330, %f13, %f329;
	ex2.approx.f32 	%f331, %f330;
	bra.uni 	$Lt_2_163586;
$Lt_2_163842:
	neg.f32 	%f332, %f3;
	lg2.approx.f32 	%f333, %f332;
	mul.f32 	%f334, %f13, %f333;
	ex2.approx.f32 	%f335, %f334;
	neg.f32 	%f331, %f335;
$Lt_2_163586:
	mov.f32 	%f336, 0f00000000;   	// 0
	setp.ge.f32 	%p55, %f4, %f336;
	@!%p55 bra 	$Lt_2_164354;
	.loc	19	295	0
	lg2.approx.f32 	%f337, %f4;
	mul.f32 	%f338, %f13, %f337;
	ex2.approx.f32 	%f339, %f338;
	bra.uni 	$Lt_2_164098;
$Lt_2_164354:
	neg.f32 	%f340, %f4;
	lg2.approx.f32 	%f341, %f340;
	mul.f32 	%f342, %f13, %f341;
	ex2.approx.f32 	%f343, %f342;
	neg.f32 	%f339, %f343;
$Lt_2_164098:
	.loc	29	42	0
	mov.f32 	%f5, %f315;
	mov.f32 	%f6, %f323;
	mov.f32 	%f7, %f331;
	mov.f32 	%f8, %f339;
	.loc	35	91	0
	bra.uni 	$Lt_2_10242;
$Lt_2_13314:
	.loc	30	41	0
	sub.s32 	%r117, %r14, %r5;
	sub.s32 	%r6, %r117, 1;
	.loc	35	94	0
	bra.uni 	$Lt_2_10242;
$Lt_2_13570:
	.loc	35	96	0
	sub.u32 	%r73, %r5, %r11;
	ld.param.f32 	%f13, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter0+4];
	mov.f32 	%f344, 0f3e800000;   	// 0.25
	mul.f32 	%f345, %f13, %f344;
	ld.param.f32 	%f346, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter0+16];
	ld.param.f32 	%f347, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter0+8];
	mov.f32 	%f348, 0f3f800000;   	// 1
	setp.eq.f32 	%p56, %f347, %f348;
	@!%p56 bra 	$Lt_2_164866;
	.loc	17	117	0
	cvt.rzi.u32.f32 	%r118, %f346;
	shr.u32 	%r119, %r118, 13;
	sub.u32 	%r120, %r73, %r118;
	sub.u32 	%r121, %r11, %r118;
	xor.b32 	%r122, %r120, %r119;
	shl.b32 	%r123, %r122, 8;
	sub.u32 	%r124, %r121, %r122;
	sub.u32 	%r125, %r118, %r122;
	xor.b32 	%r126, %r124, %r123;
	shr.u32 	%r127, %r126, 13;
	sub.u32 	%r128, %r125, %r126;
	sub.u32 	%r129, %r122, %r126;
	xor.b32 	%r130, %r128, %r127;
	shr.u32 	%r131, %r130, 12;
	sub.u32 	%r132, %r129, %r130;
	xor.b32 	%r133, %r132, %r131;
	shl.b32 	%r134, %r133, 16;
	sub.u32 	%r135, %r126, %r130;
	sub.u32 	%r136, %r135, %r133;
	xor.b32 	%r137, %r134, %r136;
	.loc	17	118	0
	sub.u32 	%r138, %r130, %r133;
	sub.u32 	%r139, %r138, %r137;
	shr.u32 	%r140, %r137, 5;
	xor.b32 	%r141, %r139, %r140;
	.loc	17	119	0
	sub.u32 	%r142, %r133, %r137;
	sub.u32 	%r143, %r142, %r141;
	shr.u32 	%r144, %r141, 3;
	xor.b32 	%r145, %r143, %r144;
	.loc	17	120	0
	sub.u32 	%r146, %r137, %r141;
	sub.u32 	%r147, %r146, %r145;
	shl.b32 	%r148, %r145, 10;
	xor.b32 	%r149, %r147, %r148;
	.loc	17	121	0
	sub.u32 	%r150, %r141, %r145;
	sub.u32 	%r151, %r150, %r149;
	shr.u32 	%r152, %r149, 15;
	xor.b32 	%r153, %r151, %r152;
	.loc	31	48	0
	mul.lo.u32 	%r154, %r153, 1103515245;
	add.u32 	%r155, %r154, 12345;
	shr.u32 	%r156, %r155, 16;
	and.b32 	%r157, %r156, 255;
	shl.b32 	%r158, %r157, 7;
	mul.lo.u32 	%r159, %r153, -1029531031;
	sub.u32 	%r160, %r159, 740551042;
	shr.u32 	%r161, %r160, 16;
	and.b32 	%r162, %r161, 255;
	xor.b32 	%r163, %r158, %r162;
	cvt.rn.f32.s32 	%f349, %r163;
	mov.f32 	%f350, 0f46fffe00;   	// 32767
	div.full.f32 	%f351, %f349, %f350;
	mul.f32 	%f352, %f13, %f351;
	mov.f32 	%f353, 0f3f000000;   	// 0.5
	mul.f32 	%f354, %f352, %f353;
	sub.f32 	%f355, %f354, %f345;
	.loc	17	117	0
	add.u32 	%r164, %r120, 1;
	xor.b32 	%r165, %r164, %r119;
	shl.b32 	%r166, %r165, 8;
	sub.u32 	%r167, %r121, %r165;
	sub.u32 	%r168, %r118, %r165;
	xor.b32 	%r169, %r166, %r167;
	shr.u32 	%r170, %r169, 13;
	sub.u32 	%r171, %r168, %r169;
	sub.u32 	%r172, %r165, %r169;
	xor.b32 	%r173, %r170, %r171;
	shr.u32 	%r174, %r173, 12;
	sub.u32 	%r175, %r172, %r173;
	xor.b32 	%r176, %r174, %r175;
	sub.u32 	%r177, %r169, %r173;
	sub.u32 	%r178, %r177, %r176;
	shl.b32 	%r179, %r176, 16;
	xor.b32 	%r178, %r178, %r179;
	.loc	17	118	0
	sub.u32 	%r180, %r173, %r176;
	sub.u32 	%r180, %r180, %r178;
	shr.u32 	%r181, %r178, 5;
	xor.b32 	%r180, %r180, %r181;
	.loc	17	119	0
	sub.u32 	%r182, %r176, %r178;
	sub.u32 	%r182, %r182, %r180;
	shr.u32 	%r183, %r180, 3;
	xor.b32 	%r182, %r182, %r183;
	.loc	17	120	0
	sub.u32 	%r178, %r178, %r180;
	sub.u32 	%r178, %r178, %r182;
	shl.b32 	%r184, %r182, 10;
	xor.b32 	%r178, %r178, %r184;
	.loc	17	121	0
	sub.u32 	%r180, %r180, %r182;
	sub.u32 	%r180, %r180, %r178;
	shr.u32 	%r185, %r178, 15;
	xor.b32 	%r180, %r180, %r185;
	.loc	31	49	0
	mul.lo.u32 	%r186, %r180, 1103515245;
	add.u32 	%r187, %r186, 12345;
	shr.u32 	%r188, %r187, 16;
	and.b32 	%r189, %r188, 255;
	shl.b32 	%r190, %r189, 7;
	mul.lo.u32 	%r191, %r180, -1029531031;
	sub.u32 	%r192, %r191, 740551042;
	shr.u32 	%r193, %r192, 16;
	and.b32 	%r194, %r193, 255;
	xor.b32 	%r195, %r190, %r194;
	cvt.rn.f32.s32 	%f356, %r195;
	mov.f32 	%f357, 0f46fffe00;   	// 32767
	div.full.f32 	%f358, %f356, %f357;
	mul.f32 	%f359, %f13, %f358;
	mov.f32 	%f360, 0f3f000000;   	// 0.5
	mul.f32 	%f361, %f359, %f360;
	sub.f32 	%f362, %f361, %f345;
	.loc	17	117	0
	add.u32 	%r196, %r120, 2;
	xor.b32 	%r197, %r196, %r119;
	shl.b32 	%r198, %r197, 8;
	sub.u32 	%r199, %r121, %r197;
	sub.u32 	%r200, %r118, %r197;
	xor.b32 	%r201, %r198, %r199;
	shr.u32 	%r202, %r201, 13;
	sub.u32 	%r203, %r200, %r201;
	sub.u32 	%r204, %r197, %r201;
	xor.b32 	%r205, %r202, %r203;
	shr.u32 	%r206, %r205, 12;
	sub.u32 	%r207, %r204, %r205;
	xor.b32 	%r208, %r206, %r207;
	sub.u32 	%r209, %r201, %r205;
	sub.u32 	%r210, %r209, %r208;
	shl.b32 	%r211, %r208, 16;
	xor.b32 	%r210, %r210, %r211;
	.loc	17	118	0
	sub.u32 	%r212, %r205, %r208;
	sub.u32 	%r212, %r212, %r210;
	shr.u32 	%r213, %r210, 5;
	xor.b32 	%r212, %r212, %r213;
	.loc	17	119	0
	sub.u32 	%r214, %r208, %r210;
	sub.u32 	%r214, %r214, %r212;
	shr.u32 	%r215, %r212, 3;
	xor.b32 	%r214, %r214, %r215;
	.loc	17	120	0
	sub.u32 	%r210, %r210, %r212;
	sub.u32 	%r210, %r210, %r214;
	shl.b32 	%r216, %r214, 10;
	xor.b32 	%r210, %r210, %r216;
	.loc	17	121	0
	sub.u32 	%r212, %r212, %r214;
	sub.u32 	%r212, %r212, %r210;
	shr.u32 	%r217, %r210, 15;
	xor.b32 	%r212, %r212, %r217;
	.loc	31	50	0
	mul.lo.u32 	%r218, %r212, 1103515245;
	add.u32 	%r219, %r218, 12345;
	shr.u32 	%r220, %r219, 16;
	and.b32 	%r221, %r220, 255;
	shl.b32 	%r222, %r221, 7;
	mul.lo.u32 	%r223, %r212, -1029531031;
	sub.u32 	%r224, %r223, 740551042;
	shr.u32 	%r225, %r224, 16;
	and.b32 	%r226, %r225, 255;
	xor.b32 	%r227, %r222, %r226;
	cvt.rn.f32.s32 	%f363, %r227;
	mov.f32 	%f364, 0f46fffe00;   	// 32767
	div.full.f32 	%f365, %f363, %f364;
	mul.f32 	%f366, %f13, %f365;
	mov.f32 	%f367, 0f3f000000;   	// 0.5
	mul.f32 	%f368, %f366, %f367;
	sub.f32 	%f369, %f368, %f345;
	bra.uni 	$Lt_2_164610;
$Lt_2_164866:
	.loc	17	117	0
	cvt.rzi.u32.f32 	%r228, %f346;
	shr.u32 	%r229, %r228, 13;
	sub.u32 	%r230, %r73, %r228;
	sub.u32 	%r231, %r11, %r228;
	xor.b32 	%r232, %r230, %r229;
	shl.b32 	%r233, %r232, 8;
	sub.u32 	%r234, %r231, %r232;
	sub.u32 	%r235, %r228, %r232;
	xor.b32 	%r236, %r234, %r233;
	shr.u32 	%r237, %r236, 13;
	sub.u32 	%r238, %r235, %r236;
	sub.u32 	%r239, %r232, %r236;
	xor.b32 	%r240, %r238, %r237;
	shr.u32 	%r241, %r240, 12;
	sub.u32 	%r242, %r239, %r240;
	xor.b32 	%r243, %r242, %r241;
	shl.b32 	%r244, %r243, 16;
	sub.u32 	%r245, %r236, %r240;
	sub.u32 	%r246, %r245, %r243;
	xor.b32 	%r247, %r244, %r246;
	.loc	17	118	0
	sub.u32 	%r248, %r240, %r243;
	sub.u32 	%r249, %r248, %r247;
	shr.u32 	%r250, %r247, 5;
	xor.b32 	%r251, %r249, %r250;
	.loc	17	119	0
	sub.u32 	%r252, %r243, %r247;
	sub.u32 	%r253, %r252, %r251;
	shr.u32 	%r254, %r251, 3;
	xor.b32 	%r255, %r253, %r254;
	.loc	17	120	0
	sub.u32 	%r256, %r247, %r251;
	sub.u32 	%r257, %r256, %r255;
	shl.b32 	%r258, %r255, 10;
	xor.b32 	%r259, %r257, %r258;
	.loc	17	121	0
	sub.u32 	%r260, %r251, %r255;
	sub.u32 	%r261, %r260, %r259;
	shr.u32 	%r262, %r259, 15;
	xor.b32 	%r263, %r261, %r262;
	.loc	31	54	0
	mul.lo.u32 	%r264, %r263, 1103515245;
	mul.lo.u32 	%r265, %r263, -1029531031;
	add.u32 	%r266, %r264, 12345;
	sub.u32 	%r267, %r265, 740551042;
	shr.u32 	%r268, %r266, 16;
	shr.u32 	%r269, %r267, 16;
	and.b32 	%r270, %r268, 255;
	and.b32 	%r271, %r269, 255;
	shl.b32 	%r272, %r270, 7;
	xor.b32 	%r273, %r272, %r271;
	cvt.rn.f32.s32 	%f370, %r273;
	mov.f32 	%f371, 0f46fffe00;   	// 32767
	div.full.f32 	%f372, %f370, %f371;
	mul.f32 	%f373, %f372, %f13;
	mov.f32 	%f374, 0f3f000000;   	// 0.5
	mul.f32 	%f375, %f373, %f374;
	sub.f32 	%f355, %f375, %f345;
	mov.f32 	%f369, %f355;
	mov.f32 	%f362, %f355;
$Lt_2_164610:
	.loc	31	57	0
	add.f32 	%f7, %f355, %f3;
	.loc	31	58	0
	add.f32 	%f6, %f2, %f362;
	.loc	31	59	0
	add.f32 	%f5, %f1, %f369;
	ld.param.f32 	%f376, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter0+12];
	mov.f32 	%f377, 0f3f800000;   	// 1
	setp.eq.f32 	%p57, %f376, %f377;
	@!%p57 bra 	$Lt_2_10242;
	.loc	31	63	0
	mov.f32 	%f378, 0f00000000;   	// 0
	max.f32 	%f379, %f5, %f378;
	mov.f32 	%f380, 0f3f800000;   	// 1
	min.f32 	%f5, %f379, %f380;
	mov.f32 	%f381, 0f00000000;   	// 0
	max.f32 	%f382, %f6, %f381;
	mov.f32 	%f383, 0f3f800000;   	// 1
	min.f32 	%f6, %f382, %f383;
	mov.f32 	%f384, 0f00000000;   	// 0
	max.f32 	%f385, %f7, %f384;
	mov.f32 	%f386, 0f3f800000;   	// 1
	min.f32 	%f7, %f385, %f386;
	mov.f32 	%f387, 0f00000000;   	// 0
	max.f32 	%f388, %f4, %f387;
	mov.f32 	%f389, 0f3f800000;   	// 1
	min.f32 	%f8, %f388, %f389;
	bra.uni 	$Lt_2_10242;
$Lt_2_13826:
	.loc	35	99	0
	cvt.rn.f32.s32 	%f390, %r5;
	cvt.rn.f32.s32 	%f391, %r14;
	ld.param.f32 	%f392, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter0+20];
	mul.f32 	%f393, %f391, %f392;
	setp.lt.f32 	%p58, %f390, %f393;
	@!%p58 bra 	$Lt_2_10242;
	.loc	19	267	0
	ld.const.f32 	%f394, [kRGB32f_To_601YPbPr+4];
	mul.f32 	%f29, %f394, %f2;
	ld.const.f32 	%f395, [kRGB32f_To_601YPbPr+28];
	mul.f32 	%f396, %f395, %f2;
	ld.const.f32 	%f397, [kRGB32f_To_601YPbPr+16];
	mul.f32 	%f398, %f397, %f2;
	ld.const.f32 	%f399, [kRGB32f_To_601YPbPr+0];
	mad.f32 	%f31, %f399, %f3, %f29;
	ld.const.f32 	%f400, [kRGB32f_To_601YPbPr+24];
	mad.f32 	%f401, %f400, %f3, %f396;
	ld.const.f32 	%f402, [kRGB32f_To_601YPbPr+12];
	mad.f32 	%f403, %f402, %f3, %f398;
	ld.const.f32 	%f404, [kRGB32f_To_601YPbPr+8];
	mad.f32 	%f33, %f404, %f1, %f31;
	ld.const.f32 	%f405, [kRGB32f_To_601YPbPr+32];
	mad.f32 	%f406, %f405, %f1, %f401;
	ld.const.f32 	%f407, [kRGB32f_To_601YPbPr+20];
	mad.f32 	%f408, %f407, %f1, %f403;
	ld.param.f32 	%f409, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter0+4];
	ld.param.f32 	%f410, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter0+8];
	mad.f32 	%f411, %f33, %f410, %f409;
	ld.param.f32 	%f412, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter0+16];
	mul.f32 	%f413, %f406, %f412;
	ld.param.f32 	%f414, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter0+12];
	mul.f32 	%f415, %f406, %f414;
	mul.f32 	%f416, %f408, %f414;
	sub.f32 	%f417, %f416, %f413;
	mad.f32 	%f418, %f408, %f412, %f415;
	ld.const.f32 	%f419, [k601YPbPr_To_RGB32f+16];
	mul.f32 	%f420, %f419, %f417;
	ld.const.f32 	%f421, [k601YPbPr_To_RGB32f+12];
	mad.f32 	%f422, %f421, %f411, %f420;
	ld.const.f32 	%f423, [k601YPbPr_To_RGB32f+20];
	mad.f32 	%f424, %f423, %f418, %f422;
	.loc	19	268	0
	ld.const.f32 	%f425, [k601YPbPr_To_RGB32f+4];
	mul.f32 	%f426, %f425, %f417;
	ld.const.f32 	%f427, [k601YPbPr_To_RGB32f+0];
	mad.f32 	%f428, %f427, %f411, %f426;
	ld.const.f32 	%f429, [k601YPbPr_To_RGB32f+8];
	mad.f32 	%f430, %f429, %f418, %f428;
	.loc	32	56	0
	ld.const.f32 	%f431, [k601YPbPr_To_RGB32f+28];
	mul.f32 	%f432, %f431, %f417;
	ld.const.f32 	%f433, [k601YPbPr_To_RGB32f+24];
	mad.f32 	%f434, %f433, %f411, %f432;
	ld.const.f32 	%f435, [k601YPbPr_To_RGB32f+32];
	mad.f32 	%f5, %f435, %f418, %f434;
	mov.f32 	%f6, %f424;
	mov.f32 	%f7, %f430;
	bra.uni 	$Lt_2_10242;
$Lt_2_14082:
	.loc	33	46	0
	ld.const.f32 	%f436, [kRGB32f_To_601YPbPr+4];
	mul.f32 	%f29, %f436, %f2;
	ld.const.f32 	%f437, [kRGB32f_To_601YPbPr+0];
	mad.f32 	%f31, %f437, %f3, %f29;
	ld.const.f32 	%f438, [kRGB32f_To_601YPbPr+8];
	mad.f32 	%f33, %f438, %f1, %f31;
	ld.param.f32 	%f43, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter0+8];
	ld.param.f32 	%f439, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter0+20];
	sub.f32 	%f440, %f439, %f43;
	mad.f32 	%f441, %f33, %f440, %f43;
	.loc	33	47	0
	ld.param.f32 	%f442, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter0+12];
	ld.param.f32 	%f443, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter0+24];
	sub.f32 	%f444, %f443, %f442;
	mad.f32 	%f445, %f33, %f444, %f442;
	.loc	33	49	0
	ld.param.f32 	%f446, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter0+28];
	ld.param.f32 	%f13, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter0+4];
	ld.param.f32 	%f447, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter0+16];
	sub.f32 	%f448, %f447, %f13;
	mad.f32 	%f449, %f33, %f448, %f13;
	sub.f32 	%f450, %f449, %f1;
	mad.f32 	%f5, %f446, %f450, %f1;
	.loc	33	50	0
	sub.f32 	%f451, %f441, %f2;
	mad.f32 	%f6, %f446, %f451, %f2;
	.loc	33	51	0
	sub.f32 	%f452, %f445, %f3;
	mad.f32 	%f7, %f446, %f452, %f3;
	.loc	35	103	0
	bra.uni 	$Lt_2_10242;
$Lt_2_14338:
	.loc	34	41	0
	sub.s32 	%r274, %r13, %r11;
	sub.s32 	%r12, %r274, 1;
$Lt_2_10242:
	.loc	35	54	0
	ld.param.u32 	%r275, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter1+0];
	mov.u32 	%r276, 0;
	setp.eq.s32 	%p59, %r275, %r276;
	@%p59 bra 	$Lt_2_5122;
	mov.u32 	%r277, 1;
	setp.eq.s32 	%p60, %r275, %r277;
	@%p60 bra 	$Lt_2_5634;
	mov.u32 	%r278, 2;
	setp.eq.s32 	%p61, %r275, %r278;
	@%p61 bra 	$Lt_2_5890;
	mov.u32 	%r279, 3;
	setp.eq.s32 	%p62, %r275, %r279;
	@%p62 bra 	$Lt_2_6146;
	mov.u32 	%r280, 4;
	setp.eq.s32 	%p63, %r275, %r280;
	@%p63 bra 	$Lt_2_6402;
	mov.u32 	%r281, 5;
	setp.eq.s32 	%p64, %r275, %r281;
	@%p64 bra 	$Lt_2_6658;
	mov.u32 	%r282, 6;
	setp.eq.s32 	%p65, %r275, %r282;
	@%p65 bra 	$Lt_2_6914;
	mov.u32 	%r283, 7;
	setp.eq.s32 	%p66, %r275, %r283;
	@%p66 bra 	$Lt_2_7170;
	mov.u32 	%r284, 8;
	setp.eq.s32 	%p67, %r275, %r284;
	@%p67 bra 	$Lt_2_7426;
	mov.u32 	%r285, 9;
	setp.eq.s32 	%p68, %r275, %r285;
	@%p68 bra 	$Lt_2_7682;
	mov.u32 	%r286, 10;
	setp.eq.s32 	%p69, %r275, %r286;
	@%p69 bra 	$Lt_2_7938;
	mov.u32 	%r287, 11;
	setp.eq.s32 	%p70, %r275, %r287;
	@%p70 bra 	$Lt_2_8194;
	mov.u32 	%r288, 12;
	setp.eq.s32 	%p71, %r275, %r288;
	@%p71 bra 	$Lt_2_8450;
	mov.u32 	%r289, 13;
	setp.eq.s32 	%p72, %r275, %r289;
	@%p72 bra 	$Lt_2_8706;
	mov.u32 	%r290, 14;
	setp.eq.s32 	%p73, %r275, %r290;
	@%p73 bra 	$Lt_2_8962;
	mov.u32 	%r291, 15;
	setp.eq.s32 	%p74, %r275, %r291;
	@%p74 bra 	$Lt_2_9218;
	mov.u32 	%r292, 16;
	setp.eq.s32 	%p75, %r275, %r292;
	@%p75 bra 	$Lt_2_9474;
	bra.uni 	$Lt_2_5378;
$Lt_2_5122:
	.loc	18	42	0
	cvt.sat.f32.f32 	%f8, %f8;
	ld.param.f32 	%f453, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter1+8];
	mov.f32 	%f454, 0f3f800000;   	// 1
	setp.eq.f32 	%p76, %f453, %f454;
	ld.param.f32 	%f455, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter1+12];
	mov.f32 	%f456, 0f3f800000;   	// 1
	setp.eq.f32 	%p77, %f455, %f456;
	ld.param.f32 	%f457, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter1+4];
	ld.param.f32 	%f458, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter1+16];
	mov.f32 	%f459, 0f3f800000;   	// 1
	setp.eq.f32 	%p78, %f458, %f459;
	@!%p78 bra 	$Lt_2_166402;
	.loc	18	45	0
	mul.f32 	%f460, %f8, %f457;
	selp.f32 	%f8, %f457, %f460, %p76;
	.loc	18	53	0
	mov.f32 	%f461, 0f3f800000;   	// 1
	sub.f32 	%f462, %f461, %f8;
	selp.f32 	%f8, %f462, %f8, %p77;
	.loc	18	57	0
	mov.f32 	%f5, %f8;
	mov.f32 	%f6, %f8;
	mov.f32 	%f7, %f8;
	bra.uni 	$Lt_2_5378;
$Lt_2_166402:
	@!%p76 bra 	$Lt_2_166914;
	.loc	18	61	0
	mov.f32 	%f463, 0f3f800000;   	// 1
	sub.f32 	%f464, %f463, %f457;
	selp.f32 	%f8, %f464, %f457, %p77;
	bra.uni 	$Lt_2_5378;
$Lt_2_166914:
	.loc	18	69	0
	mov.f32 	%f465, 0f3f800000;   	// 1
	sub.f32 	%f466, %f465, %f8;
	selp.f32 	%f8, %f466, %f8, %p77;
	.loc	18	73	0
	mul.f32 	%f8, %f8, %f457;
	bra.uni 	$Lt_2_5378;
$Lt_2_5634:
	.loc	19	267	0
	ld.const.f32 	%f467, [kRGB32f_To_601YPbPr+4];
	mul.f32 	%f468, %f467, %f6;
	ld.const.f32 	%f469, [kRGB32f_To_601YPbPr+0];
	mad.f32 	%f470, %f469, %f7, %f468;
	ld.const.f32 	%f471, [kRGB32f_To_601YPbPr+8];
	mad.f32 	%f472, %f471, %f5, %f470;
	ld.const.f32 	%f473, [k601YPbPr_To_RGB32f+12];
	mul.f32 	%f474, %f473, %f472;
	.loc	19	268	0
	ld.const.f32 	%f475, [k601YPbPr_To_RGB32f+0];
	mul.f32 	%f476, %f475, %f472;
	.loc	20	44	0
	ld.const.f32 	%f477, [k601YPbPr_To_RGB32f+24];
	mul.f32 	%f5, %f477, %f472;
	mov.f32 	%f6, %f474;
	mov.f32 	%f7, %f476;
	.loc	35	61	0
	bra.uni 	$Lt_2_5378;
$Lt_2_5890:
	.loc	35	63	0
	ld.param.f32 	%f478, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter1+20];
	ld.param.f32 	%f479, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter1+12];
	ld.param.f32 	%f480, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter1+4];
	mov.f32 	%f481, 0f00000000;   	// 0
	setp.neu.f32 	%p79, %f480, %f481;
	@!%p79 bra 	$Lt_2_167426;
	.loc	21	44	0
	ld.param.f32 	%f482, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter1+8];
	sub.f32 	%f483, %f7, %f479;
	mad.f32 	%f7, %f482, %f483, %f478;
	.loc	21	45	0
	sub.f32 	%f484, %f6, %f479;
	mad.f32 	%f6, %f482, %f484, %f478;
	.loc	21	46	0
	sub.f32 	%f485, %f5, %f479;
	mad.f32 	%f5, %f482, %f485, %f478;
	bra.uni 	$Lt_2_5378;
$Lt_2_167426:
	.loc	21	50	0
	ld.param.f32 	%f486, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter1+24];
	setp.gt.f32 	%p80, %f7, %f479;
	selp.f32 	%f7, %f486, %f478, %p80;
	.loc	21	51	0
	setp.gt.f32 	%p81, %f6, %f479;
	selp.f32 	%f6, %f486, %f478, %p81;
	.loc	21	52	0
	setp.gt.f32 	%p82, %f5, %f479;
	selp.f32 	%f5, %f486, %f478, %p82;
	bra.uni 	$Lt_2_5378;
$Lt_2_6146:
	.loc	22	47	0
	ld.param.f32 	%f487, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter1+4];
	mul.f32 	%f5, %f487, %f5;
	ld.param.f32 	%f488, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter1+8];
	mul.f32 	%f6, %f488, %f6;
	ld.param.f32 	%f489, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter1+12];
	mul.f32 	%f7, %f489, %f7;
	.loc	35	67	0
	bra.uni 	$Lt_2_5378;
$Lt_2_6402:
	.loc	23	48	0
	ld.param.f32 	%f490, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter1+4];
	mov.f32 	%f491, 0f00000000;   	// 0
	setp.eq.f32 	%p83, %f490, %f491;
	ld.param.f32 	%f492, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter1+8];
	mov.f32 	%f493, 0f00000000;   	// 0
	max.f32 	%f494, %f5, %f493;
	mov.f32 	%f495, 0f3f800000;   	// 1
	min.f32 	%f496, %f494, %f495;
	ld.param.f32 	%f497, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter1+12];
	sub.f32 	%f498, %f496, %f497;
	abs.f32 	%f499, %f498;
	mov.f32 	%f500, 0f00000000;   	// 0
	max.f32 	%f501, %f6, %f500;
	mov.f32 	%f502, 0f3f800000;   	// 1
	min.f32 	%f503, %f501, %f502;
	ld.param.f32 	%f504, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter1+16];
	sub.f32 	%f505, %f503, %f504;
	abs.f32 	%f506, %f505;
	mov.f32 	%f507, 0f00000000;   	// 0
	max.f32 	%f508, %f7, %f507;
	mov.f32 	%f509, 0f3f800000;   	// 1
	min.f32 	%f510, %f508, %f509;
	ld.param.f32 	%f511, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter1+20];
	sub.f32 	%f512, %f510, %f511;
	abs.f32 	%f513, %f512;
	mov.f32 	%f514, 0f00000000;   	// 0
	max.f32 	%f515, %f8, %f514;
	mov.f32 	%f516, 0f3f800000;   	// 1
	min.f32 	%f517, %f515, %f516;
	sub.f32 	%f518, %f517, %f8;
	abs.f32 	%f519, %f518;
	max.f32 	%f520, %f513, %f519;
	max.f32 	%f521, %f506, %f520;
	max.f32 	%f522, %f499, %f521;
	setp.ge.f32 	%p84, %f492, %f522;
	xor.pred 	%p85, %p83, %p84;
	@!%p85 bra 	$Lt_2_5378;
	.loc	19	267	0
	ld.const.f32 	%f523, [kRGB32f_To_601YPbPr+4];
	mul.f32 	%f468, %f523, %f6;
	ld.const.f32 	%f524, [kRGB32f_To_601YPbPr+0];
	mad.f32 	%f470, %f524, %f7, %f468;
	ld.const.f32 	%f525, [kRGB32f_To_601YPbPr+8];
	mad.f32 	%f472, %f525, %f5, %f470;
	ld.const.f32 	%f526, [k601YPbPr_To_RGB32f+12];
	mul.f32 	%f527, %f526, %f472;
	.loc	19	268	0
	ld.const.f32 	%f528, [k601YPbPr_To_RGB32f+0];
	mul.f32 	%f529, %f528, %f472;
	.loc	20	44	0
	ld.const.f32 	%f530, [k601YPbPr_To_RGB32f+24];
	mul.f32 	%f5, %f530, %f472;
	mov.f32 	%f6, %f527;
	mov.f32 	%f7, %f529;
	bra.uni 	$Lt_2_5378;
$Lt_2_6658:
	.loc	24	48	0
	ld.param.f32 	%f531, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter1+8];
	mov.f32 	%f532, 0f00000000;   	// 0
	max.f32 	%f533, %f5, %f532;
	mov.f32 	%f534, 0f3f800000;   	// 1
	min.f32 	%f535, %f533, %f534;
	ld.param.f32 	%f536, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter1+12];
	sub.f32 	%f537, %f535, %f536;
	abs.f32 	%f538, %f537;
	mov.f32 	%f539, 0f00000000;   	// 0
	max.f32 	%f540, %f6, %f539;
	mov.f32 	%f541, 0f3f800000;   	// 1
	min.f32 	%f542, %f540, %f541;
	ld.param.f32 	%f543, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter1+16];
	sub.f32 	%f544, %f542, %f543;
	abs.f32 	%f545, %f544;
	mov.f32 	%f546, 0f00000000;   	// 0
	max.f32 	%f547, %f7, %f546;
	mov.f32 	%f548, 0f3f800000;   	// 1
	min.f32 	%f549, %f547, %f548;
	ld.param.f32 	%f550, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter1+20];
	sub.f32 	%f551, %f549, %f550;
	abs.f32 	%f552, %f551;
	mov.f32 	%f553, 0f00000000;   	// 0
	max.f32 	%f554, %f8, %f553;
	mov.f32 	%f555, 0f3f800000;   	// 1
	min.f32 	%f556, %f554, %f555;
	sub.f32 	%f557, %f556, %f8;
	abs.f32 	%f558, %f557;
	max.f32 	%f559, %f552, %f558;
	max.f32 	%f560, %f545, %f559;
	max.f32 	%f561, %f538, %f560;
	setp.ge.f32 	%p86, %f531, %f561;
	@!%p86 bra 	$Lt_2_5378;
	.loc	24	51	0
	ld.param.f32 	%f562, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter1+24];
	mov.f32 	%f124, %f562;
	ld.param.f32 	%f563, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter1+28];
	mov.f32 	%f126, %f563;
	ld.param.f32 	%f564, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter1+32];
	mov.f32 	%f128, %f564;
	ld.param.f32 	%f565, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter1+4];
	mov.f32 	%f566, 0f00000000;   	// 0
	setp.eq.f32 	%p87, %f565, %f566;
	@!%p87 bra 	$Lt_2_168706;
	.loc	24	60	0
	ld.const.f32 	%f567, [kRGB32f_To_601YPbPr+4];
	mul.f32 	%f468, %f567, %f6;
	ld.const.f32 	%f568, [kRGB32f_To_601YPbPr+0];
	mad.f32 	%f470, %f568, %f7, %f468;
	ld.const.f32 	%f569, [kRGB32f_To_601YPbPr+8];
	mad.f32 	%f472, %f569, %f5, %f470;
	mul.f32 	%f124, %f472, %f562;
	.loc	24	61	0
	mul.f32 	%f126, %f472, %f563;
	.loc	24	62	0
	mul.f32 	%f128, %f472, %f564;
$Lt_2_168706:
	.loc	24	65	0
	mov.f32 	%f5, %f124;
	mov.f32 	%f6, %f126;
	mov.f32 	%f7, %f128;
	bra.uni 	$Lt_2_5378;
$Lt_2_6914:
	.loc	25	47	0
	sub.s32 	%r293, %r13, %r12;
	sub.s32 	%r294, %r14, %r6;
	cvt.rn.f32.s32 	%f570, %r6;
	cvt.rn.f32.s32 	%f571, %r12;
	cvt.rn.f32.s32 	%f136, %r293;
	cvt.rn.f32.s32 	%f137, %r294;
	ld.param.f32 	%f457, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter1+4];
	sub.f32 	%f572, %f457, %f570;
	sub.f32 	%f573, %f457, %f571;
	sub.f32 	%f574, %f457, %f136;
	sub.f32 	%f575, %f457, %f137;
	cvt.rzi.s32.f32 	%r295, %f572;
	cvt.rzi.s32.f32 	%r296, %f573;
	cvt.rzi.s32.f32 	%r297, %f574;
	cvt.rzi.s32.f32 	%r298, %f575;
	max.s32 	%r299, %r295, %r296;
	max.s32 	%r300, %r298, %r299;
	max.s32 	%r301, %r297, %r300;
	mov.u32 	%r302, 0;
	setp.le.s32 	%p88, %r301, %r302;
	@%p88 bra 	$Lt_2_5378;
	.loc	25	51	0
	cvt.rn.f32.s32 	%f576, %r301;
	sub.f32 	%f577, %f457, %f576;
	div.full.f32 	%f578, %f577, %f457;
	mul.f32 	%f8, %f8, %f578;
	bra.uni 	$Lt_2_5378;
$Lt_2_7170:
	.loc	27	50	0
	ld.const.f32 	%f579, [kRGB32f_To_601YPbPr+4];
	mul.f32 	%f468, %f579, %f6;
	ld.param.f32 	%f580, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter1+12];
	ld.param.f32 	%f482, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter1+8];
	sub.f32 	%f581, %f580, %f482;
	ld.const.f32 	%f582, [kRGB32f_To_601YPbPr+0];
	mad.f32 	%f470, %f582, %f7, %f468;
	ld.param.f32 	%f583, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter1+16];
	mul.f32 	%f584, %f583, %f581;
	ld.const.f32 	%f585, [kRGB32f_To_601YPbPr+8];
	mad.f32 	%f472, %f585, %f5, %f470;
	add.f32 	%f586, %f584, %f580;
	sub.f32 	%f587, %f482, %f584;
	mov.f32 	%f588, 0f00000000;   	// 0
	max.f32 	%f589, %f586, %f588;
	mov.f32 	%f590, 0f00000000;   	// 0
	max.f32 	%f591, %f587, %f590;
	mov.f32 	%f592, 0f3f800000;   	// 1
	min.f32 	%f593, %f589, %f592;
	mov.f32 	%f594, 0f3f800000;   	// 1
	min.f32 	%f595, %f591, %f594;
	set.gt.u32.f32 	%r303, %f595, %f472;
	neg.s32 	%r304, %r303;
	set.le.u32.f32 	%r305, %f593, %f472;
	neg.s32 	%r306, %r305;
	or.b32 	%r307, %r304, %r306;
	mov.u32 	%r308, 0;
	setp.eq.s32 	%p89, %r307, %r308;
	@%p89 bra 	$Lt_2_169986;
	mov.f32 	%f162, 0f00000000;   	// 0
	bra.uni 	$Lt_2_170754;
$Lt_2_169986:
	add.f32 	%f596, %f584, %f482;
	mov.f32 	%f597, 0f00000000;   	// 0
	max.f32 	%f598, %f596, %f597;
	mov.f32 	%f599, 0f3f800000;   	// 1
	min.f32 	%f600, %f598, %f599;
	set.le.u32.f32 	%r309, %f600, %f472;
	neg.s32 	%r310, %r309;
	sub.f32 	%f601, %f580, %f584;
	mov.f32 	%f602, 0f00000000;   	// 0
	max.f32 	%f603, %f601, %f602;
	mov.f32 	%f604, 0f3f800000;   	// 1
	min.f32 	%f605, %f603, %f604;
	set.lt.u32.f32 	%r311, %f472, %f605;
	neg.s32 	%r312, %r311;
	and.b32 	%r313, %r310, %r312;
	mov.u32 	%r314, 0;
	setp.eq.s32 	%p90, %r313, %r314;
	@%p90 bra 	$Lt_2_170498;
	mov.f32 	%f162, 0f3f800000;   	// 1
	bra.uni 	$Lt_2_170754;
$Lt_2_170498:
	add.f32 	%f606, %f584, %f584;
	setp.gt.f32 	%p91, %f600, %f472;
	@!%p91 bra 	$Lt_2_171010;
	.loc	27	62	0
	sub.f32 	%f607, %f472, %f595;
	div.full.f32 	%f162, %f607, %f606;
	bra.uni 	$Lt_2_170754;
$Lt_2_171010:
	.loc	27	66	0
	sub.f32 	%f608, %f593, %f472;
	div.full.f32 	%f162, %f608, %f606;
$Lt_2_170754:
$Lt_2_170242:
$Lt_2_169730:
	.loc	27	69	0
	mov.f32 	%f609, 0f3f800000;   	// 1
	sub.f32 	%f610, %f609, %f162;
	ld.param.f32 	%f611, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter1+4];
	mov.f32 	%f612, 0f00000000;   	// 0
	setp.neu.f32 	%p92, %f611, %f612;
	selp.f32 	%f162, %f610, %f162, %p92;
	.loc	27	77	0
	ld.const.f32 	%f613, [k601YPbPr_To_RGB32f+24];
	mul.f32 	%f5, %f613, %f162;
	ld.const.f32 	%f614, [k601YPbPr_To_RGB32f+12];
	mul.f32 	%f6, %f614, %f162;
	ld.const.f32 	%f615, [k601YPbPr_To_RGB32f+0];
	mul.f32 	%f7, %f615, %f162;
	.loc	35	79	0
	bra.uni 	$Lt_2_5378;
$Lt_2_7426:
	.loc	35	80	0
	ld.param.f32 	%f616, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter1+44];
	mov.f32 	%f617, 0f00000000;   	// 0
	setp.ge.f32 	%p93, %f616, %f617;
	@!%p93 bra 	$L_2_150786;
	cvt.rn.f32.s32 	%f618, %r6;
	cvt.rn.f32.s32 	%f619, %r14;
	mul.f32 	%f620, %f619, %f616;
	setp.lt.f32 	%p94, %f618, %f620;
	@%p94 bra 	$L_2_150530;
$L_2_150786:
	mov.f32 	%f621, 0f00000000;   	// 0
	setp.lt.f32 	%p95, %f616, %f621;
	@!%p95 bra 	$Lt_2_5378;
	cvt.rn.f32.s32 	%f622, %r12;
	cvt.rn.f32.s32 	%f623, %r13;
	mul.f32 	%f624, %f623, %f616;
	neg.f32 	%f625, %f624;
	setp.lt.f32 	%p96, %f622, %f625;
	@!%p96 bra 	$Lt_2_5378;
$L_2_150530:
	.loc	28	47	0
	ld.const.f32 	%f626, [kRGB32f_To_601YPbPr+4];
	mul.f32 	%f627, %f626, %f6;
	ld.const.f32 	%f628, [kRGB32f_To_601YPbPr+0];
	mad.f32 	%f629, %f628, %f7, %f627;
	ld.const.f32 	%f630, [kRGB32f_To_601YPbPr+8];
	mad.f32 	%f472, %f630, %f5, %f629;
	mov.f32 	%f198, %f472;
	ld.param.f32 	%f482, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter1+8];
	ld.param.f32 	%f631, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter1+4];
	sub.f32 	%f632, %f631, %f482;
	ld.param.f32 	%f633, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter1+16];
	ld.param.f32 	%f634, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter1+12];
	sub.f32 	%f635, %f634, %f633;
	mov.f32 	%f636, 0f3f800000;   	// 1
	set.neu.u32.f32 	%r315, %f635, %f636;
	neg.s32 	%r316, %r315;
	mov.f32 	%f637, 0f3f800000;   	// 1
	set.neu.u32.f32 	%r317, %f632, %f637;
	neg.s32 	%r318, %r317;
	or.b32 	%r319, %r316, %r318;
	mov.u32 	%r320, 0;
	setp.eq.s32 	%p97, %r319, %r320;
	@%p97 bra 	$Lt_2_171266;
	.loc	17	116	0
	mov.u32 	%r321, 1;
	sub.u32 	%r71, %r321, %r5;
	shr.u32 	%r72, %r11, 13;
	sub.u32 	%r73, %r5, %r11;
	sub.u32 	%r74, %r71, %r11;
	xor.b32 	%r75, %r74, %r72;
	shl.b32 	%r76, %r75, 8;
	sub.u32 	%r77, %r73, %r75;
	sub.u32 	%r78, %r11, %r75;
	xor.b32 	%r79, %r77, %r76;
	shr.u32 	%r80, %r79, 13;
	sub.u32 	%r81, %r78, %r79;
	xor.b32 	%r82, %r81, %r80;
	shr.u32 	%r322, %r82, 12;
	sub.u32 	%r323, %r75, %r79;
	sub.u32 	%r324, %r323, %r82;
	xor.b32 	%r86, %r322, %r324;
	.loc	17	117	0
	sub.u32 	%r325, %r79, %r82;
	sub.u32 	%r88, %r325, %r86;
	shl.b32 	%r326, %r86, 16;
	xor.b32 	%r90, %r88, %r326;
	.loc	17	118	0
	sub.u32 	%r327, %r82, %r86;
	sub.u32 	%r92, %r327, %r90;
	shr.u32 	%r328, %r90, 5;
	xor.b32 	%r94, %r92, %r328;
	.loc	17	119	0
	sub.u32 	%r95, %r86, %r90;
	sub.u32 	%r96, %r95, %r94;
	shr.u32 	%r329, %r94, 3;
	xor.b32 	%r98, %r96, %r329;
	.loc	17	120	0
	sub.u32 	%r99, %r90, %r94;
	sub.u32 	%r100, %r99, %r98;
	shl.b32 	%r330, %r98, 10;
	xor.b32 	%r102, %r100, %r330;
	.loc	17	121	0
	sub.u32 	%r103, %r94, %r98;
	sub.u32 	%r104, %r103, %r102;
	shr.u32 	%r331, %r102, 15;
	xor.b32 	%r106, %r104, %r331;
	.loc	28	57	0
	mov.f32 	%f638, 0f3b270d73;   	// 0.00254902
	mul.lo.u32 	%r332, %r106, 1103515245;
	add.u32 	%r333, %r332, 12345;
	shr.u32 	%r334, %r333, 16;
	and.b32 	%r335, %r334, 255;
	shl.b32 	%r336, %r335, 7;
	mul.lo.u32 	%r337, %r106, -1029531031;
	sub.u32 	%r338, %r337, 740551042;
	shr.u32 	%r339, %r338, 16;
	and.b32 	%r340, %r339, 255;
	xor.b32 	%r341, %r336, %r340;
	cvt.rn.f32.s32 	%f639, %r341;
	mov.f32 	%f640, 0f467ffe00;   	// 16383.5
	div.full.f32 	%f641, %f639, %f640;
	mov.f32 	%f642, 0fbf800000;   	// -1
	add.f32 	%f643, %f641, %f642;
	mad.f32 	%f198, %f638, %f643, %f472;
$Lt_2_171266:
	sub.f32 	%f644, %f198, %f482;
	ld.param.f32 	%f645, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter1+20];
	mov.f32 	%f646, 0f3f800000;   	// 1
	setp.neu.f32 	%p98, %f645, %f646;
	@!%p98 bra 	$Lt_2_172034;
	mov.f32 	%f647, 0f00000000;   	// 0
	setp.lt.f32 	%p99, %f644, %f647;
	@!%p99 bra 	$Lt_2_172546;
	.loc	28	66	0
	mov.f32 	%f198, %f633;
	bra.uni 	$Lt_2_171778;
$Lt_2_172546:
	.loc	28	70	0
	rcp.approx.f32 	%f648, %f632;
	mul.f32 	%f649, %f644, %f648;
	lg2.approx.f32 	%f650, %f649;
	mul.f32 	%f651, %f645, %f650;
	ex2.approx.f32 	%f652, %f651;
	mad.f32 	%f198, %f635, %f652, %f633;
	bra.uni 	$Lt_2_171778;
$Lt_2_172034:
	.loc	28	77	0
	rcp.approx.f32 	%f653, %f632;
	mul.f32 	%f654, %f644, %f653;
	mad.f32 	%f198, %f635, %f654, %f633;
$Lt_2_171778:
	.loc	19	267	0
	ld.const.f32 	%f655, [kRGB32f_To_601YPbPr+28];
	mul.f32 	%f656, %f655, %f6;
	ld.const.f32 	%f657, [kRGB32f_To_601YPbPr+16];
	mul.f32 	%f658, %f657, %f6;
	ld.param.f32 	%f659, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter1+28];
	sin.approx.f32 	%f660, %f659;
	cos.approx.f32 	%f661, %f659;
	ld.const.f32 	%f662, [kRGB32f_To_601YPbPr+24];
	mad.f32 	%f663, %f662, %f7, %f656;
	ld.const.f32 	%f664, [kRGB32f_To_601YPbPr+12];
	mad.f32 	%f665, %f664, %f7, %f658;
	ld.const.f32 	%f666, [kRGB32f_To_601YPbPr+32];
	mad.f32 	%f667, %f666, %f5, %f663;
	ld.const.f32 	%f668, [kRGB32f_To_601YPbPr+20];
	mad.f32 	%f669, %f668, %f5, %f665;
	mul.f32 	%f670, %f660, %f667;
	mul.f32 	%f671, %f661, %f667;
	mul.f32 	%f672, %f669, %f661;
	sub.f32 	%f673, %f672, %f670;
	mad.f32 	%f674, %f669, %f660, %f671;
	ld.param.f32 	%f675, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter1+32];
	sub.f32 	%f676, %f675, %f673;
	ld.param.f32 	%f677, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter1+36];
	sub.f32 	%f678, %f677, %f674;
	ld.param.f32 	%f679, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter1+40];
	mad.f32 	%f680, %f679, %f676, %f673;
	mad.f32 	%f681, %f679, %f678, %f674;
	ld.param.f32 	%f682, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter1+24];
	mul.f32 	%f683, %f680, %f682;
	mul.f32 	%f684, %f681, %f682;
	ld.const.f32 	%f685, [k601YPbPr_To_RGB32f+16];
	mul.f32 	%f686, %f685, %f683;
	ld.const.f32 	%f687, [k601YPbPr_To_RGB32f+12];
	mad.f32 	%f688, %f687, %f198, %f686;
	ld.const.f32 	%f689, [k601YPbPr_To_RGB32f+20];
	mad.f32 	%f258, %f689, %f684, %f688;
	.loc	19	268	0
	ld.const.f32 	%f690, [k601YPbPr_To_RGB32f+4];
	mul.f32 	%f691, %f690, %f683;
	ld.const.f32 	%f692, [k601YPbPr_To_RGB32f+0];
	mad.f32 	%f693, %f692, %f198, %f691;
	ld.const.f32 	%f694, [k601YPbPr_To_RGB32f+8];
	mad.f32 	%f264, %f694, %f684, %f693;
	.loc	28	92	0
	ld.const.f32 	%f695, [k601YPbPr_To_RGB32f+28];
	mul.f32 	%f696, %f695, %f683;
	ld.const.f32 	%f697, [k601YPbPr_To_RGB32f+24];
	mad.f32 	%f698, %f697, %f198, %f696;
	ld.const.f32 	%f699, [k601YPbPr_To_RGB32f+32];
	mad.f32 	%f5, %f699, %f684, %f698;
	mov.f32 	%f6, %f258;
	mov.f32 	%f7, %f264;
	bra.uni 	$Lt_2_5378;
$Lt_2_7682:
	.loc	35	83	0
	ld.param.f32 	%f700, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter1+44];
	mov.f32 	%f701, 0f00000000;   	// 0
	setp.ge.f32 	%p100, %f700, %f701;
	@!%p100 bra 	$L_2_151554;
	cvt.rn.f32.s32 	%f702, %r6;
	cvt.rn.f32.s32 	%f703, %r14;
	mul.f32 	%f704, %f703, %f700;
	setp.lt.f32 	%p101, %f702, %f704;
	@%p101 bra 	$L_2_151298;
$L_2_151554:
	mov.f32 	%f705, 0f00000000;   	// 0
	setp.lt.f32 	%p102, %f700, %f705;
	@!%p102 bra 	$Lt_2_5378;
	cvt.rn.f32.s32 	%f706, %r12;
	cvt.rn.f32.s32 	%f707, %r13;
	mul.f32 	%f708, %f707, %f700;
	neg.f32 	%f709, %f708;
	setp.lt.f32 	%p103, %f706, %f709;
	@!%p103 bra 	$Lt_2_5378;
$L_2_151298:
	.loc	28	110	0
	ld.const.f32 	%f710, [kRGB32f_To_601YPbPr+4];
	mul.f32 	%f468, %f710, %f6;
	ld.param.f32 	%f482, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter1+8];
	ld.param.f32 	%f711, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter1+4];
	sub.f32 	%f712, %f711, %f482;
	ld.param.f32 	%f713, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter1+16];
	ld.param.f32 	%f714, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter1+12];
	sub.f32 	%f715, %f714, %f713;
	ld.const.f32 	%f716, [kRGB32f_To_601YPbPr+0];
	mad.f32 	%f470, %f716, %f7, %f468;
	ld.const.f32 	%f717, [kRGB32f_To_601YPbPr+8];
	mad.f32 	%f472, %f717, %f5, %f470;
	sub.f32 	%f718, %f472, %f482;
	ld.param.f32 	%f719, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter1+20];
	mov.f32 	%f720, 0f3f800000;   	// 1
	setp.neu.f32 	%p104, %f719, %f720;
	@!%p104 bra 	$Lt_2_173058;
	.loc	28	120	0
	mov.f32 	%f721, 0f00000000;   	// 0
	max.f32 	%f722, %f718, %f721;
	div.full.f32 	%f723, %f722, %f712;
	lg2.approx.f32 	%f724, %f723;
	mul.f32 	%f725, %f719, %f724;
	ex2.approx.f32 	%f726, %f725;
	mad.f32 	%f297, %f715, %f726, %f713;
	bra.uni 	$Lt_2_172802;
$Lt_2_173058:
	.loc	28	129	0
	div.full.f32 	%f727, %f718, %f712;
	mad.f32 	%f297, %f715, %f727, %f713;
$Lt_2_172802:
	.loc	28	135	0
	ld.const.f32 	%f728, [k601YPbPr_To_RGB32f+24];
	mul.f32 	%f5, %f728, %f297;
	ld.const.f32 	%f729, [k601YPbPr_To_RGB32f+12];
	mul.f32 	%f6, %f729, %f297;
	ld.const.f32 	%f730, [k601YPbPr_To_RGB32f+0];
	mul.f32 	%f7, %f730, %f297;
	bra.uni 	$Lt_2_5378;
$Lt_2_7938:
	.loc	35	86	0
	ld.param.f32 	%f731, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter1+44];
	mov.f32 	%f732, 0f00000000;   	// 0
	setp.ge.f32 	%p105, %f731, %f732;
	@!%p105 bra 	$L_2_152322;
	cvt.rn.f32.s32 	%f733, %r6;
	cvt.rn.f32.s32 	%f734, %r14;
	mul.f32 	%f735, %f734, %f731;
	setp.lt.f32 	%p106, %f733, %f735;
	@%p106 bra 	$L_2_152066;
$L_2_152322:
	mov.f32 	%f736, 0f00000000;   	// 0
	setp.lt.f32 	%p107, %f731, %f736;
	@!%p107 bra 	$Lt_2_5378;
	cvt.rn.f32.s32 	%f737, %r12;
	cvt.rn.f32.s32 	%f738, %r13;
	mul.f32 	%f739, %f738, %f731;
	neg.f32 	%f740, %f739;
	setp.lt.f32 	%p108, %f737, %f740;
	@!%p108 bra 	$Lt_2_5378;
$L_2_152066:
	.loc	28	160	0
	ld.const.f32 	%f5, [k601YPbPr_To_RGB32f+24];
	ld.const.f32 	%f6, [k601YPbPr_To_RGB32f+12];
	ld.const.f32 	%f7, [k601YPbPr_To_RGB32f+0];
	bra.uni 	$Lt_2_5378;
$Lt_2_8194:
	.loc	29	42	0
	ld.param.f32 	%f457, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter1+4];
	mov.f32 	%f741, 0f00000000;   	// 0
	setp.ge.f32 	%p109, %f5, %f741;
	@!%p109 bra 	$Lt_2_173570;
	.loc	19	292	0
	lg2.approx.f32 	%f742, %f5;
	mul.f32 	%f743, %f457, %f742;
	ex2.approx.f32 	%f744, %f743;
	bra.uni 	$Lt_2_173314;
$Lt_2_173570:
	neg.f32 	%f745, %f5;
	lg2.approx.f32 	%f746, %f745;
	mul.f32 	%f747, %f457, %f746;
	ex2.approx.f32 	%f748, %f747;
	neg.f32 	%f744, %f748;
$Lt_2_173314:
	mov.f32 	%f749, 0f00000000;   	// 0
	setp.ge.f32 	%p110, %f6, %f749;
	@!%p110 bra 	$Lt_2_174082;
	.loc	19	293	0
	lg2.approx.f32 	%f750, %f6;
	mul.f32 	%f751, %f457, %f750;
	ex2.approx.f32 	%f752, %f751;
	bra.uni 	$Lt_2_173826;
$Lt_2_174082:
	neg.f32 	%f753, %f6;
	lg2.approx.f32 	%f754, %f753;
	mul.f32 	%f755, %f457, %f754;
	ex2.approx.f32 	%f756, %f755;
	neg.f32 	%f752, %f756;
$Lt_2_173826:
	mov.f32 	%f757, 0f00000000;   	// 0
	setp.ge.f32 	%p111, %f7, %f757;
	@!%p111 bra 	$Lt_2_174594;
	.loc	19	294	0
	lg2.approx.f32 	%f758, %f7;
	mul.f32 	%f759, %f457, %f758;
	ex2.approx.f32 	%f760, %f759;
	bra.uni 	$Lt_2_174338;
$Lt_2_174594:
	neg.f32 	%f761, %f7;
	lg2.approx.f32 	%f762, %f761;
	mul.f32 	%f763, %f457, %f762;
	ex2.approx.f32 	%f764, %f763;
	neg.f32 	%f760, %f764;
$Lt_2_174338:
	mov.f32 	%f765, 0f00000000;   	// 0
	setp.ge.f32 	%p112, %f8, %f765;
	@!%p112 bra 	$Lt_2_175106;
	.loc	19	295	0
	lg2.approx.f32 	%f766, %f8;
	mul.f32 	%f767, %f457, %f766;
	ex2.approx.f32 	%f768, %f767;
	bra.uni 	$Lt_2_174850;
$Lt_2_175106:
	neg.f32 	%f769, %f8;
	lg2.approx.f32 	%f770, %f769;
	mul.f32 	%f771, %f457, %f770;
	ex2.approx.f32 	%f772, %f771;
	neg.f32 	%f768, %f772;
$Lt_2_174850:
	.loc	29	42	0
	mov.f32 	%f5, %f744;
	mov.f32 	%f6, %f752;
	mov.f32 	%f7, %f760;
	mov.f32 	%f8, %f768;
	.loc	35	91	0
	bra.uni 	$Lt_2_5378;
$Lt_2_8450:
	.loc	30	41	0
	sub.s32 	%r342, %r14, %r6;
	sub.s32 	%r6, %r342, 1;
	.loc	35	94	0
	bra.uni 	$Lt_2_5378;
$Lt_2_8706:
	.loc	35	96	0
	sub.u32 	%r343, %r6, %r12;
	ld.param.f32 	%f457, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter1+4];
	mov.f32 	%f773, 0f3e800000;   	// 0.25
	mul.f32 	%f774, %f457, %f773;
	ld.param.f32 	%f775, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter1+16];
	ld.param.f32 	%f776, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter1+8];
	mov.f32 	%f777, 0f3f800000;   	// 1
	setp.eq.f32 	%p113, %f776, %f777;
	@!%p113 bra 	$Lt_2_175618;
	.loc	17	117	0
	cvt.rzi.u32.f32 	%r344, %f775;
	shr.u32 	%r345, %r344, 13;
	sub.u32 	%r346, %r343, %r344;
	sub.u32 	%r347, %r12, %r344;
	xor.b32 	%r348, %r346, %r345;
	shl.b32 	%r349, %r348, 8;
	sub.u32 	%r350, %r347, %r348;
	sub.u32 	%r351, %r344, %r348;
	xor.b32 	%r352, %r350, %r349;
	shr.u32 	%r353, %r352, 13;
	sub.u32 	%r354, %r351, %r352;
	sub.u32 	%r355, %r348, %r352;
	xor.b32 	%r356, %r354, %r353;
	shr.u32 	%r357, %r356, 12;
	sub.u32 	%r358, %r355, %r356;
	xor.b32 	%r359, %r358, %r357;
	shl.b32 	%r360, %r359, 16;
	sub.u32 	%r361, %r352, %r356;
	sub.u32 	%r362, %r361, %r359;
	xor.b32 	%r363, %r360, %r362;
	.loc	17	118	0
	sub.u32 	%r364, %r356, %r359;
	sub.u32 	%r365, %r364, %r363;
	shr.u32 	%r366, %r363, 5;
	xor.b32 	%r367, %r365, %r366;
	.loc	17	119	0
	sub.u32 	%r368, %r359, %r363;
	sub.u32 	%r369, %r368, %r367;
	shr.u32 	%r370, %r367, 3;
	xor.b32 	%r371, %r369, %r370;
	.loc	17	120	0
	sub.u32 	%r372, %r363, %r367;
	sub.u32 	%r373, %r372, %r371;
	shl.b32 	%r374, %r371, 10;
	xor.b32 	%r375, %r373, %r374;
	.loc	17	121	0
	sub.u32 	%r376, %r367, %r371;
	sub.u32 	%r377, %r376, %r375;
	shr.u32 	%r378, %r375, 15;
	xor.b32 	%r379, %r377, %r378;
	.loc	31	48	0
	mul.lo.u32 	%r380, %r379, 1103515245;
	add.u32 	%r381, %r380, 12345;
	shr.u32 	%r382, %r381, 16;
	and.b32 	%r383, %r382, 255;
	shl.b32 	%r384, %r383, 7;
	mul.lo.u32 	%r385, %r379, -1029531031;
	sub.u32 	%r386, %r385, 740551042;
	shr.u32 	%r387, %r386, 16;
	and.b32 	%r388, %r387, 255;
	xor.b32 	%r389, %r384, %r388;
	cvt.rn.f32.s32 	%f778, %r389;
	mov.f32 	%f779, 0f46fffe00;   	// 32767
	div.full.f32 	%f780, %f778, %f779;
	mul.f32 	%f781, %f457, %f780;
	mov.f32 	%f782, 0f3f000000;   	// 0.5
	mul.f32 	%f783, %f781, %f782;
	sub.f32 	%f784, %f783, %f774;
	.loc	17	117	0
	add.u32 	%r390, %r346, 1;
	xor.b32 	%r391, %r390, %r345;
	shl.b32 	%r392, %r391, 8;
	sub.u32 	%r393, %r347, %r391;
	sub.u32 	%r394, %r344, %r391;
	xor.b32 	%r395, %r392, %r393;
	shr.u32 	%r396, %r395, 13;
	sub.u32 	%r397, %r394, %r395;
	sub.u32 	%r398, %r391, %r395;
	xor.b32 	%r399, %r396, %r397;
	shr.u32 	%r400, %r399, 12;
	sub.u32 	%r401, %r398, %r399;
	xor.b32 	%r402, %r400, %r401;
	sub.u32 	%r403, %r395, %r399;
	sub.u32 	%r178, %r403, %r402;
	shl.b32 	%r404, %r402, 16;
	xor.b32 	%r178, %r178, %r404;
	.loc	17	118	0
	sub.u32 	%r180, %r399, %r402;
	sub.u32 	%r180, %r180, %r178;
	shr.u32 	%r405, %r178, 5;
	xor.b32 	%r180, %r180, %r405;
	.loc	17	119	0
	sub.u32 	%r182, %r402, %r178;
	sub.u32 	%r182, %r182, %r180;
	shr.u32 	%r406, %r180, 3;
	xor.b32 	%r182, %r182, %r406;
	.loc	17	120	0
	sub.u32 	%r178, %r178, %r180;
	sub.u32 	%r178, %r178, %r182;
	shl.b32 	%r407, %r182, 10;
	xor.b32 	%r178, %r178, %r407;
	.loc	17	121	0
	sub.u32 	%r180, %r180, %r182;
	sub.u32 	%r180, %r180, %r178;
	shr.u32 	%r408, %r178, 15;
	xor.b32 	%r180, %r180, %r408;
	.loc	31	49	0
	mul.lo.u32 	%r409, %r180, 1103515245;
	add.u32 	%r410, %r409, 12345;
	shr.u32 	%r411, %r410, 16;
	and.b32 	%r412, %r411, 255;
	shl.b32 	%r413, %r412, 7;
	mul.lo.u32 	%r414, %r180, -1029531031;
	sub.u32 	%r415, %r414, 740551042;
	shr.u32 	%r416, %r415, 16;
	and.b32 	%r417, %r416, 255;
	xor.b32 	%r418, %r413, %r417;
	cvt.rn.f32.s32 	%f785, %r418;
	mov.f32 	%f786, 0f46fffe00;   	// 32767
	div.full.f32 	%f787, %f785, %f786;
	mul.f32 	%f788, %f457, %f787;
	mov.f32 	%f789, 0f3f000000;   	// 0.5
	mul.f32 	%f790, %f788, %f789;
	sub.f32 	%f362, %f790, %f774;
	.loc	17	117	0
	add.u32 	%r419, %r346, 2;
	xor.b32 	%r420, %r419, %r345;
	shl.b32 	%r421, %r420, 8;
	sub.u32 	%r422, %r347, %r420;
	sub.u32 	%r423, %r344, %r420;
	xor.b32 	%r424, %r421, %r422;
	shr.u32 	%r425, %r424, 13;
	sub.u32 	%r426, %r423, %r424;
	sub.u32 	%r427, %r420, %r424;
	xor.b32 	%r428, %r425, %r426;
	shr.u32 	%r429, %r428, 12;
	sub.u32 	%r430, %r427, %r428;
	xor.b32 	%r431, %r429, %r430;
	sub.u32 	%r432, %r424, %r428;
	sub.u32 	%r210, %r432, %r431;
	shl.b32 	%r433, %r431, 16;
	xor.b32 	%r210, %r210, %r433;
	.loc	17	118	0
	sub.u32 	%r212, %r428, %r431;
	sub.u32 	%r212, %r212, %r210;
	shr.u32 	%r434, %r210, 5;
	xor.b32 	%r212, %r212, %r434;
	.loc	17	119	0
	sub.u32 	%r214, %r431, %r210;
	sub.u32 	%r214, %r214, %r212;
	shr.u32 	%r435, %r212, 3;
	xor.b32 	%r214, %r214, %r435;
	.loc	17	120	0
	sub.u32 	%r210, %r210, %r212;
	sub.u32 	%r210, %r210, %r214;
	shl.b32 	%r436, %r214, 10;
	xor.b32 	%r210, %r210, %r436;
	.loc	17	121	0
	sub.u32 	%r212, %r212, %r214;
	sub.u32 	%r212, %r212, %r210;
	shr.u32 	%r437, %r210, 15;
	xor.b32 	%r212, %r212, %r437;
	.loc	31	50	0
	mul.lo.u32 	%r438, %r212, 1103515245;
	add.u32 	%r439, %r438, 12345;
	shr.u32 	%r440, %r439, 16;
	and.b32 	%r441, %r440, 255;
	shl.b32 	%r442, %r441, 7;
	mul.lo.u32 	%r443, %r212, -1029531031;
	sub.u32 	%r444, %r443, 740551042;
	shr.u32 	%r445, %r444, 16;
	and.b32 	%r446, %r445, 255;
	xor.b32 	%r447, %r442, %r446;
	cvt.rn.f32.s32 	%f791, %r447;
	mov.f32 	%f792, 0f46fffe00;   	// 32767
	div.full.f32 	%f793, %f791, %f792;
	mul.f32 	%f794, %f457, %f793;
	mov.f32 	%f795, 0f3f000000;   	// 0.5
	mul.f32 	%f796, %f794, %f795;
	sub.f32 	%f369, %f796, %f774;
	bra.uni 	$Lt_2_175362;
$Lt_2_175618:
	.loc	17	117	0
	cvt.rzi.u32.f32 	%r448, %f775;
	shr.u32 	%r449, %r448, 13;
	sub.u32 	%r450, %r343, %r448;
	sub.u32 	%r451, %r12, %r448;
	xor.b32 	%r452, %r450, %r449;
	shl.b32 	%r453, %r452, 8;
	sub.u32 	%r454, %r451, %r452;
	sub.u32 	%r455, %r448, %r452;
	xor.b32 	%r456, %r454, %r453;
	shr.u32 	%r457, %r456, 13;
	sub.u32 	%r458, %r455, %r456;
	sub.u32 	%r459, %r452, %r456;
	xor.b32 	%r460, %r458, %r457;
	shr.u32 	%r461, %r460, 12;
	sub.u32 	%r462, %r459, %r460;
	xor.b32 	%r463, %r462, %r461;
	shl.b32 	%r464, %r463, 16;
	sub.u32 	%r465, %r456, %r460;
	sub.u32 	%r466, %r465, %r463;
	xor.b32 	%r467, %r464, %r466;
	.loc	17	118	0
	sub.u32 	%r468, %r460, %r463;
	sub.u32 	%r469, %r468, %r467;
	shr.u32 	%r470, %r467, 5;
	xor.b32 	%r471, %r469, %r470;
	.loc	17	119	0
	sub.u32 	%r472, %r463, %r467;
	sub.u32 	%r473, %r472, %r471;
	shr.u32 	%r474, %r471, 3;
	xor.b32 	%r475, %r473, %r474;
	.loc	17	120	0
	sub.u32 	%r476, %r467, %r471;
	sub.u32 	%r477, %r476, %r475;
	shl.b32 	%r478, %r475, 10;
	xor.b32 	%r479, %r477, %r478;
	.loc	17	121	0
	sub.u32 	%r480, %r471, %r475;
	sub.u32 	%r481, %r480, %r479;
	shr.u32 	%r482, %r479, 15;
	xor.b32 	%r483, %r481, %r482;
	.loc	31	54	0
	mul.lo.u32 	%r484, %r483, 1103515245;
	mul.lo.u32 	%r485, %r483, -1029531031;
	add.u32 	%r486, %r484, 12345;
	sub.u32 	%r487, %r485, 740551042;
	shr.u32 	%r488, %r486, 16;
	shr.u32 	%r489, %r487, 16;
	and.b32 	%r490, %r488, 255;
	and.b32 	%r491, %r489, 255;
	shl.b32 	%r492, %r490, 7;
	xor.b32 	%r493, %r492, %r491;
	cvt.rn.f32.s32 	%f797, %r493;
	mov.f32 	%f798, 0f46fffe00;   	// 32767
	div.full.f32 	%f799, %f797, %f798;
	mul.f32 	%f800, %f799, %f457;
	mov.f32 	%f801, 0f3f000000;   	// 0.5
	mul.f32 	%f802, %f800, %f801;
	sub.f32 	%f784, %f802, %f774;
	mov.f32 	%f369, %f784;
	mov.f32 	%f362, %f784;
$Lt_2_175362:
	.loc	31	57	0
	add.f32 	%f7, %f784, %f7;
	.loc	31	58	0
	add.f32 	%f6, %f362, %f6;
	.loc	31	59	0
	add.f32 	%f5, %f5, %f369;
	ld.param.f32 	%f803, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter1+12];
	mov.f32 	%f804, 0f3f800000;   	// 1
	setp.eq.f32 	%p114, %f803, %f804;
	@!%p114 bra 	$Lt_2_5378;
	.loc	31	63	0
	mov.f32 	%f805, 0f00000000;   	// 0
	max.f32 	%f806, %f5, %f805;
	mov.f32 	%f807, 0f3f800000;   	// 1
	min.f32 	%f5, %f806, %f807;
	mov.f32 	%f808, 0f00000000;   	// 0
	max.f32 	%f809, %f6, %f808;
	mov.f32 	%f810, 0f3f800000;   	// 1
	min.f32 	%f6, %f809, %f810;
	mov.f32 	%f811, 0f00000000;   	// 0
	max.f32 	%f812, %f7, %f811;
	mov.f32 	%f813, 0f3f800000;   	// 1
	min.f32 	%f7, %f812, %f813;
	mov.f32 	%f814, 0f00000000;   	// 0
	max.f32 	%f815, %f8, %f814;
	mov.f32 	%f816, 0f3f800000;   	// 1
	min.f32 	%f8, %f815, %f816;
	bra.uni 	$Lt_2_5378;
$Lt_2_8962:
	.loc	35	99	0
	cvt.rn.f32.s32 	%f817, %r6;
	cvt.rn.f32.s32 	%f818, %r14;
	ld.param.f32 	%f819, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter1+20];
	mul.f32 	%f820, %f818, %f819;
	setp.lt.f32 	%p115, %f817, %f820;
	@!%p115 bra 	$Lt_2_5378;
	.loc	19	267	0
	ld.const.f32 	%f821, [kRGB32f_To_601YPbPr+4];
	mul.f32 	%f468, %f821, %f6;
	ld.const.f32 	%f822, [kRGB32f_To_601YPbPr+28];
	mul.f32 	%f823, %f822, %f6;
	ld.const.f32 	%f824, [kRGB32f_To_601YPbPr+16];
	mul.f32 	%f825, %f824, %f6;
	ld.const.f32 	%f826, [kRGB32f_To_601YPbPr+0];
	mad.f32 	%f470, %f826, %f7, %f468;
	ld.const.f32 	%f827, [kRGB32f_To_601YPbPr+24];
	mad.f32 	%f828, %f827, %f7, %f823;
	ld.const.f32 	%f829, [kRGB32f_To_601YPbPr+12];
	mad.f32 	%f830, %f829, %f7, %f825;
	ld.const.f32 	%f831, [kRGB32f_To_601YPbPr+8];
	mad.f32 	%f472, %f831, %f5, %f470;
	ld.const.f32 	%f832, [kRGB32f_To_601YPbPr+32];
	mad.f32 	%f833, %f832, %f5, %f828;
	ld.const.f32 	%f834, [kRGB32f_To_601YPbPr+20];
	mad.f32 	%f835, %f834, %f5, %f830;
	ld.param.f32 	%f836, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter1+4];
	ld.param.f32 	%f837, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter1+8];
	mad.f32 	%f838, %f472, %f837, %f836;
	ld.param.f32 	%f839, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter1+16];
	mul.f32 	%f840, %f833, %f839;
	ld.param.f32 	%f841, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter1+12];
	mul.f32 	%f842, %f833, %f841;
	mul.f32 	%f843, %f835, %f841;
	sub.f32 	%f844, %f843, %f840;
	mad.f32 	%f845, %f835, %f839, %f842;
	ld.const.f32 	%f846, [k601YPbPr_To_RGB32f+16];
	mul.f32 	%f847, %f846, %f844;
	ld.const.f32 	%f848, [k601YPbPr_To_RGB32f+12];
	mad.f32 	%f849, %f848, %f838, %f847;
	ld.const.f32 	%f850, [k601YPbPr_To_RGB32f+20];
	mad.f32 	%f424, %f850, %f845, %f849;
	.loc	19	268	0
	ld.const.f32 	%f851, [k601YPbPr_To_RGB32f+4];
	mul.f32 	%f852, %f851, %f844;
	ld.const.f32 	%f853, [k601YPbPr_To_RGB32f+0];
	mad.f32 	%f854, %f853, %f838, %f852;
	ld.const.f32 	%f855, [k601YPbPr_To_RGB32f+8];
	mad.f32 	%f430, %f855, %f845, %f854;
	.loc	32	56	0
	ld.const.f32 	%f856, [k601YPbPr_To_RGB32f+28];
	mul.f32 	%f857, %f856, %f844;
	ld.const.f32 	%f858, [k601YPbPr_To_RGB32f+24];
	mad.f32 	%f859, %f858, %f838, %f857;
	ld.const.f32 	%f860, [k601YPbPr_To_RGB32f+32];
	mad.f32 	%f5, %f860, %f845, %f859;
	mov.f32 	%f6, %f424;
	mov.f32 	%f7, %f430;
	bra.uni 	$Lt_2_5378;
$Lt_2_9218:
	.loc	33	46	0
	ld.const.f32 	%f861, [kRGB32f_To_601YPbPr+4];
	mul.f32 	%f468, %f861, %f6;
	ld.const.f32 	%f862, [kRGB32f_To_601YPbPr+0];
	mad.f32 	%f470, %f862, %f7, %f468;
	ld.const.f32 	%f863, [kRGB32f_To_601YPbPr+8];
	mad.f32 	%f472, %f863, %f5, %f470;
	ld.param.f32 	%f482, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter1+8];
	ld.param.f32 	%f864, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter1+20];
	sub.f32 	%f865, %f864, %f482;
	mad.f32 	%f441, %f472, %f865, %f482;
	.loc	33	47	0
	ld.param.f32 	%f866, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter1+12];
	ld.param.f32 	%f867, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter1+24];
	sub.f32 	%f868, %f867, %f866;
	mad.f32 	%f445, %f472, %f868, %f866;
	.loc	33	49	0
	ld.param.f32 	%f869, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter1+28];
	ld.param.f32 	%f457, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter1+4];
	ld.param.f32 	%f870, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter1+16];
	sub.f32 	%f871, %f870, %f457;
	mad.f32 	%f872, %f472, %f871, %f457;
	sub.f32 	%f873, %f872, %f5;
	mad.f32 	%f5, %f869, %f873, %f5;
	.loc	33	50	0
	sub.f32 	%f874, %f441, %f6;
	mad.f32 	%f6, %f869, %f874, %f6;
	.loc	33	51	0
	sub.f32 	%f875, %f445, %f7;
	mad.f32 	%f7, %f869, %f875, %f7;
	.loc	35	103	0
	bra.uni 	$Lt_2_5378;
$Lt_2_9474:
	.loc	34	41	0
	sub.s32 	%r494, %r13, %r12;
	sub.s32 	%r12, %r494, 1;
$Lt_2_5378:
	.loc	35	54	0
	ld.param.u32 	%r495, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter2+0];
	mov.u32 	%r496, 0;
	setp.eq.s32 	%p116, %r495, %r496;
	@%p116 bra 	$Lt_2_258;
	mov.u32 	%r497, 1;
	setp.eq.s32 	%p117, %r495, %r497;
	@%p117 bra 	$Lt_2_770;
	mov.u32 	%r498, 2;
	setp.eq.s32 	%p118, %r495, %r498;
	@%p118 bra 	$Lt_2_1026;
	mov.u32 	%r499, 3;
	setp.eq.s32 	%p119, %r495, %r499;
	@%p119 bra 	$Lt_2_1282;
	mov.u32 	%r500, 4;
	setp.eq.s32 	%p120, %r495, %r500;
	@%p120 bra 	$Lt_2_1538;
	mov.u32 	%r501, 5;
	setp.eq.s32 	%p121, %r495, %r501;
	@%p121 bra 	$Lt_2_1794;
	mov.u32 	%r502, 6;
	setp.eq.s32 	%p122, %r495, %r502;
	@%p122 bra 	$Lt_2_2050;
	mov.u32 	%r503, 7;
	setp.eq.s32 	%p123, %r495, %r503;
	@%p123 bra 	$Lt_2_2306;
	mov.u32 	%r504, 8;
	setp.eq.s32 	%p124, %r495, %r504;
	@%p124 bra 	$Lt_2_2562;
	mov.u32 	%r505, 9;
	setp.eq.s32 	%p125, %r495, %r505;
	@%p125 bra 	$Lt_2_2818;
	mov.u32 	%r506, 10;
	setp.eq.s32 	%p126, %r495, %r506;
	@%p126 bra 	$Lt_2_3074;
	mov.u32 	%r507, 11;
	setp.eq.s32 	%p127, %r495, %r507;
	@%p127 bra 	$Lt_2_3330;
	mov.u32 	%r508, 12;
	setp.eq.s32 	%p128, %r495, %r508;
	@%p128 bra 	$Lt_2_3586;
	mov.u32 	%r509, 13;
	setp.eq.s32 	%p129, %r495, %r509;
	@%p129 bra 	$Lt_2_3842;
	mov.u32 	%r510, 14;
	setp.eq.s32 	%p130, %r495, %r510;
	@%p130 bra 	$Lt_2_4098;
	mov.u32 	%r511, 15;
	setp.eq.s32 	%p131, %r495, %r511;
	@%p131 bra 	$Lt_2_4354;
	mov.u32 	%r512, 16;
	setp.eq.s32 	%p132, %r495, %r512;
	@%p132 bra 	$Lt_2_4610;
	bra.uni 	$Lt_2_514;
$Lt_2_258:
	.loc	18	42	0
	cvt.sat.f32.f32 	%f8, %f8;
	ld.param.f32 	%f876, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter2+8];
	mov.f32 	%f877, 0f3f800000;   	// 1
	setp.eq.f32 	%p133, %f876, %f877;
	ld.param.f32 	%f878, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter2+12];
	mov.f32 	%f879, 0f3f800000;   	// 1
	setp.eq.f32 	%p134, %f878, %f879;
	ld.param.f32 	%f880, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter2+4];
	ld.param.f32 	%f881, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter2+16];
	mov.f32 	%f882, 0f3f800000;   	// 1
	setp.eq.f32 	%p135, %f881, %f882;
	@!%p135 bra 	$Lt_2_177154;
	.loc	18	45	0
	mul.f32 	%f883, %f8, %f880;
	selp.f32 	%f8, %f880, %f883, %p133;
	.loc	18	53	0
	mov.f32 	%f884, 0f3f800000;   	// 1
	sub.f32 	%f885, %f884, %f8;
	selp.f32 	%f8, %f885, %f8, %p134;
	.loc	18	57	0
	mov.f32 	%f5, %f8;
	mov.f32 	%f6, %f8;
	mov.f32 	%f7, %f8;
	bra.uni 	$Lt_2_514;
$Lt_2_177154:
	@!%p133 bra 	$Lt_2_177666;
	.loc	18	61	0
	mov.f32 	%f886, 0f3f800000;   	// 1
	sub.f32 	%f887, %f886, %f880;
	selp.f32 	%f8, %f887, %f880, %p134;
	bra.uni 	$Lt_2_514;
$Lt_2_177666:
	.loc	18	69	0
	mov.f32 	%f888, 0f3f800000;   	// 1
	sub.f32 	%f889, %f888, %f8;
	selp.f32 	%f8, %f889, %f8, %p134;
	.loc	18	73	0
	mul.f32 	%f8, %f8, %f880;
	bra.uni 	$Lt_2_514;
$Lt_2_770:
	.loc	19	267	0
	ld.const.f32 	%f890, [kRGB32f_To_601YPbPr+4];
	mul.f32 	%f891, %f890, %f6;
	ld.const.f32 	%f892, [kRGB32f_To_601YPbPr+0];
	mad.f32 	%f893, %f892, %f7, %f891;
	ld.const.f32 	%f894, [kRGB32f_To_601YPbPr+8];
	mad.f32 	%f895, %f894, %f5, %f893;
	ld.const.f32 	%f896, [k601YPbPr_To_RGB32f+12];
	mul.f32 	%f897, %f896, %f895;
	.loc	19	268	0
	ld.const.f32 	%f898, [k601YPbPr_To_RGB32f+0];
	mul.f32 	%f899, %f898, %f895;
	.loc	20	44	0
	ld.const.f32 	%f900, [k601YPbPr_To_RGB32f+24];
	mul.f32 	%f5, %f900, %f895;
	mov.f32 	%f6, %f897;
	mov.f32 	%f7, %f899;
	.loc	35	61	0
	bra.uni 	$Lt_2_514;
$Lt_2_1026:
	.loc	35	63	0
	ld.param.f32 	%f901, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter2+20];
	ld.param.f32 	%f902, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter2+12];
	ld.param.f32 	%f903, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter2+4];
	mov.f32 	%f904, 0f00000000;   	// 0
	setp.neu.f32 	%p136, %f903, %f904;
	@!%p136 bra 	$Lt_2_178178;
	.loc	21	44	0
	ld.param.f32 	%f905, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter2+8];
	sub.f32 	%f906, %f7, %f902;
	mad.f32 	%f7, %f905, %f906, %f901;
	.loc	21	45	0
	sub.f32 	%f907, %f6, %f902;
	mad.f32 	%f6, %f905, %f907, %f901;
	.loc	21	46	0
	sub.f32 	%f908, %f5, %f902;
	mad.f32 	%f5, %f905, %f908, %f901;
	bra.uni 	$Lt_2_514;
$Lt_2_178178:
	.loc	21	50	0
	ld.param.f32 	%f909, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter2+24];
	setp.gt.f32 	%p137, %f7, %f902;
	selp.f32 	%f7, %f909, %f901, %p137;
	.loc	21	51	0
	setp.gt.f32 	%p138, %f6, %f902;
	selp.f32 	%f6, %f909, %f901, %p138;
	.loc	21	52	0
	setp.gt.f32 	%p139, %f5, %f902;
	selp.f32 	%f5, %f909, %f901, %p139;
	bra.uni 	$Lt_2_514;
$Lt_2_1282:
	.loc	22	47	0
	ld.param.f32 	%f910, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter2+4];
	mul.f32 	%f5, %f910, %f5;
	ld.param.f32 	%f911, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter2+8];
	mul.f32 	%f6, %f911, %f6;
	ld.param.f32 	%f912, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter2+12];
	mul.f32 	%f7, %f912, %f7;
	.loc	35	67	0
	bra.uni 	$Lt_2_514;
$Lt_2_1538:
	.loc	23	48	0
	ld.param.f32 	%f913, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter2+4];
	mov.f32 	%f914, 0f00000000;   	// 0
	setp.eq.f32 	%p140, %f913, %f914;
	ld.param.f32 	%f915, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter2+8];
	mov.f32 	%f916, 0f00000000;   	// 0
	max.f32 	%f917, %f5, %f916;
	mov.f32 	%f918, 0f3f800000;   	// 1
	min.f32 	%f919, %f917, %f918;
	ld.param.f32 	%f920, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter2+12];
	sub.f32 	%f921, %f919, %f920;
	abs.f32 	%f922, %f921;
	mov.f32 	%f923, 0f00000000;   	// 0
	max.f32 	%f924, %f6, %f923;
	mov.f32 	%f925, 0f3f800000;   	// 1
	min.f32 	%f926, %f924, %f925;
	ld.param.f32 	%f927, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter2+16];
	sub.f32 	%f928, %f926, %f927;
	abs.f32 	%f929, %f928;
	mov.f32 	%f930, 0f00000000;   	// 0
	max.f32 	%f931, %f7, %f930;
	mov.f32 	%f932, 0f3f800000;   	// 1
	min.f32 	%f933, %f931, %f932;
	ld.param.f32 	%f934, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter2+20];
	sub.f32 	%f935, %f933, %f934;
	abs.f32 	%f936, %f935;
	mov.f32 	%f937, 0f00000000;   	// 0
	max.f32 	%f938, %f8, %f937;
	mov.f32 	%f939, 0f3f800000;   	// 1
	min.f32 	%f940, %f938, %f939;
	sub.f32 	%f941, %f940, %f8;
	abs.f32 	%f942, %f941;
	max.f32 	%f943, %f936, %f942;
	max.f32 	%f944, %f929, %f943;
	max.f32 	%f945, %f922, %f944;
	setp.ge.f32 	%p141, %f915, %f945;
	xor.pred 	%p142, %p140, %p141;
	@!%p142 bra 	$Lt_2_514;
	.loc	19	267	0
	ld.const.f32 	%f946, [kRGB32f_To_601YPbPr+4];
	mul.f32 	%f891, %f946, %f6;
	ld.const.f32 	%f947, [kRGB32f_To_601YPbPr+0];
	mad.f32 	%f893, %f947, %f7, %f891;
	ld.const.f32 	%f948, [kRGB32f_To_601YPbPr+8];
	mad.f32 	%f895, %f948, %f5, %f893;
	ld.const.f32 	%f949, [k601YPbPr_To_RGB32f+12];
	mul.f32 	%f950, %f949, %f895;
	.loc	19	268	0
	ld.const.f32 	%f951, [k601YPbPr_To_RGB32f+0];
	mul.f32 	%f952, %f951, %f895;
	.loc	20	44	0
	ld.const.f32 	%f953, [k601YPbPr_To_RGB32f+24];
	mul.f32 	%f5, %f953, %f895;
	mov.f32 	%f6, %f950;
	mov.f32 	%f7, %f952;
	bra.uni 	$Lt_2_514;
$Lt_2_1794:
	.loc	24	48	0
	ld.param.f32 	%f954, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter2+8];
	mov.f32 	%f955, 0f00000000;   	// 0
	max.f32 	%f956, %f5, %f955;
	mov.f32 	%f957, 0f3f800000;   	// 1
	min.f32 	%f958, %f956, %f957;
	ld.param.f32 	%f959, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter2+12];
	sub.f32 	%f960, %f958, %f959;
	abs.f32 	%f961, %f960;
	mov.f32 	%f962, 0f00000000;   	// 0
	max.f32 	%f963, %f6, %f962;
	mov.f32 	%f964, 0f3f800000;   	// 1
	min.f32 	%f965, %f963, %f964;
	ld.param.f32 	%f966, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter2+16];
	sub.f32 	%f967, %f965, %f966;
	abs.f32 	%f968, %f967;
	mov.f32 	%f969, 0f00000000;   	// 0
	max.f32 	%f970, %f7, %f969;
	mov.f32 	%f971, 0f3f800000;   	// 1
	min.f32 	%f972, %f970, %f971;
	ld.param.f32 	%f973, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter2+20];
	sub.f32 	%f974, %f972, %f973;
	abs.f32 	%f975, %f974;
	mov.f32 	%f976, 0f00000000;   	// 0
	max.f32 	%f977, %f8, %f976;
	mov.f32 	%f978, 0f3f800000;   	// 1
	min.f32 	%f979, %f977, %f978;
	sub.f32 	%f980, %f979, %f8;
	abs.f32 	%f981, %f980;
	max.f32 	%f982, %f975, %f981;
	max.f32 	%f983, %f968, %f982;
	max.f32 	%f984, %f961, %f983;
	setp.ge.f32 	%p143, %f954, %f984;
	@!%p143 bra 	$Lt_2_514;
	.loc	24	51	0
	ld.param.f32 	%f985, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter2+24];
	mov.f32 	%f124, %f985;
	ld.param.f32 	%f986, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter2+28];
	mov.f32 	%f126, %f986;
	ld.param.f32 	%f987, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter2+32];
	mov.f32 	%f128, %f987;
	ld.param.f32 	%f988, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter2+4];
	mov.f32 	%f989, 0f00000000;   	// 0
	setp.eq.f32 	%p144, %f988, %f989;
	@!%p144 bra 	$Lt_2_179458;
	.loc	24	60	0
	ld.const.f32 	%f990, [kRGB32f_To_601YPbPr+4];
	mul.f32 	%f891, %f990, %f6;
	ld.const.f32 	%f991, [kRGB32f_To_601YPbPr+0];
	mad.f32 	%f893, %f991, %f7, %f891;
	ld.const.f32 	%f992, [kRGB32f_To_601YPbPr+8];
	mad.f32 	%f895, %f992, %f5, %f893;
	mul.f32 	%f124, %f895, %f985;
	.loc	24	61	0
	mul.f32 	%f126, %f895, %f986;
	.loc	24	62	0
	mul.f32 	%f128, %f895, %f987;
$Lt_2_179458:
	.loc	24	65	0
	mov.f32 	%f5, %f124;
	mov.f32 	%f6, %f126;
	mov.f32 	%f7, %f128;
	bra.uni 	$Lt_2_514;
$Lt_2_2050:
	.loc	25	47	0
	sub.s32 	%r513, %r13, %r12;
	sub.s32 	%r514, %r14, %r6;
	cvt.rn.f32.s32 	%f993, %r6;
	cvt.rn.f32.s32 	%f994, %r12;
	cvt.rn.f32.s32 	%f136, %r513;
	cvt.rn.f32.s32 	%f137, %r514;
	ld.param.f32 	%f880, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter2+4];
	sub.f32 	%f995, %f880, %f993;
	sub.f32 	%f996, %f880, %f994;
	sub.f32 	%f997, %f880, %f136;
	sub.f32 	%f998, %f880, %f137;
	cvt.rzi.s32.f32 	%r515, %f995;
	cvt.rzi.s32.f32 	%r516, %f996;
	cvt.rzi.s32.f32 	%r517, %f997;
	cvt.rzi.s32.f32 	%r518, %f998;
	max.s32 	%r519, %r515, %r516;
	max.s32 	%r520, %r518, %r519;
	max.s32 	%r521, %r517, %r520;
	mov.u32 	%r522, 0;
	setp.le.s32 	%p145, %r521, %r522;
	@%p145 bra 	$Lt_2_514;
	.loc	25	51	0
	cvt.rn.f32.s32 	%f999, %r521;
	sub.f32 	%f1000, %f880, %f999;
	div.full.f32 	%f1001, %f1000, %f880;
	mul.f32 	%f8, %f8, %f1001;
	bra.uni 	$Lt_2_514;
$Lt_2_2306:
	.loc	27	50	0
	ld.const.f32 	%f1002, [kRGB32f_To_601YPbPr+4];
	mul.f32 	%f891, %f1002, %f6;
	ld.param.f32 	%f1003, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter2+12];
	ld.param.f32 	%f905, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter2+8];
	sub.f32 	%f1004, %f1003, %f905;
	ld.const.f32 	%f1005, [kRGB32f_To_601YPbPr+0];
	mad.f32 	%f893, %f1005, %f7, %f891;
	ld.param.f32 	%f1006, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter2+16];
	mul.f32 	%f1007, %f1006, %f1004;
	ld.const.f32 	%f1008, [kRGB32f_To_601YPbPr+8];
	mad.f32 	%f895, %f1008, %f5, %f893;
	add.f32 	%f1009, %f1007, %f1003;
	sub.f32 	%f1010, %f905, %f1007;
	mov.f32 	%f1011, 0f00000000;  	// 0
	max.f32 	%f1012, %f1009, %f1011;
	mov.f32 	%f1013, 0f00000000;  	// 0
	max.f32 	%f1014, %f1010, %f1013;
	mov.f32 	%f1015, 0f3f800000;  	// 1
	min.f32 	%f1016, %f1012, %f1015;
	mov.f32 	%f1017, 0f3f800000;  	// 1
	min.f32 	%f1018, %f1014, %f1017;
	set.gt.u32.f32 	%r523, %f1018, %f895;
	neg.s32 	%r524, %r523;
	set.le.u32.f32 	%r525, %f1016, %f895;
	neg.s32 	%r526, %r525;
	or.b32 	%r527, %r524, %r526;
	mov.u32 	%r528, 0;
	setp.eq.s32 	%p146, %r527, %r528;
	@%p146 bra 	$Lt_2_180738;
	mov.f32 	%f162, 0f00000000;   	// 0
	bra.uni 	$Lt_2_181506;
$Lt_2_180738:
	add.f32 	%f1019, %f1007, %f905;
	mov.f32 	%f1020, 0f00000000;  	// 0
	max.f32 	%f1021, %f1019, %f1020;
	mov.f32 	%f1022, 0f3f800000;  	// 1
	min.f32 	%f1023, %f1021, %f1022;
	set.le.u32.f32 	%r529, %f1023, %f895;
	neg.s32 	%r530, %r529;
	sub.f32 	%f1024, %f1003, %f1007;
	mov.f32 	%f1025, 0f00000000;  	// 0
	max.f32 	%f1026, %f1024, %f1025;
	mov.f32 	%f1027, 0f3f800000;  	// 1
	min.f32 	%f1028, %f1026, %f1027;
	set.lt.u32.f32 	%r531, %f895, %f1028;
	neg.s32 	%r532, %r531;
	and.b32 	%r533, %r530, %r532;
	mov.u32 	%r534, 0;
	setp.eq.s32 	%p147, %r533, %r534;
	@%p147 bra 	$Lt_2_181250;
	mov.f32 	%f162, 0f3f800000;   	// 1
	bra.uni 	$Lt_2_181506;
$Lt_2_181250:
	add.f32 	%f1029, %f1007, %f1007;
	setp.gt.f32 	%p148, %f1023, %f895;
	@!%p148 bra 	$Lt_2_181762;
	.loc	27	62	0
	sub.f32 	%f1030, %f895, %f1018;
	div.full.f32 	%f162, %f1030, %f1029;
	bra.uni 	$Lt_2_181506;
$Lt_2_181762:
	.loc	27	66	0
	sub.f32 	%f1031, %f1016, %f895;
	div.full.f32 	%f162, %f1031, %f1029;
$Lt_2_181506:
$Lt_2_180994:
$Lt_2_180482:
	.loc	27	69	0
	mov.f32 	%f1032, 0f3f800000;  	// 1
	sub.f32 	%f1033, %f1032, %f162;
	ld.param.f32 	%f1034, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter2+4];
	mov.f32 	%f1035, 0f00000000;  	// 0
	setp.neu.f32 	%p149, %f1034, %f1035;
	selp.f32 	%f162, %f1033, %f162, %p149;
	.loc	27	77	0
	ld.const.f32 	%f1036, [k601YPbPr_To_RGB32f+24];
	mul.f32 	%f5, %f1036, %f162;
	ld.const.f32 	%f1037, [k601YPbPr_To_RGB32f+12];
	mul.f32 	%f6, %f1037, %f162;
	ld.const.f32 	%f1038, [k601YPbPr_To_RGB32f+0];
	mul.f32 	%f7, %f1038, %f162;
	.loc	35	79	0
	bra.uni 	$Lt_2_514;
$Lt_2_2562:
	.loc	35	80	0
	ld.param.f32 	%f1039, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter2+44];
	mov.f32 	%f1040, 0f00000000;  	// 0
	setp.ge.f32 	%p150, %f1039, %f1040;
	@!%p150 bra 	$L_2_153090;
	cvt.rn.f32.s32 	%f1041, %r6;
	cvt.rn.f32.s32 	%f1042, %r14;
	mul.f32 	%f1043, %f1042, %f1039;
	setp.lt.f32 	%p151, %f1041, %f1043;
	@%p151 bra 	$L_2_152834;
$L_2_153090:
	mov.f32 	%f1044, 0f00000000;  	// 0
	setp.lt.f32 	%p152, %f1039, %f1044;
	@!%p152 bra 	$Lt_2_514;
	cvt.rn.f32.s32 	%f1045, %r12;
	cvt.rn.f32.s32 	%f1046, %r13;
	mul.f32 	%f1047, %f1046, %f1039;
	neg.f32 	%f1048, %f1047;
	setp.lt.f32 	%p153, %f1045, %f1048;
	@!%p153 bra 	$Lt_2_514;
$L_2_152834:
	.loc	28	47	0
	ld.const.f32 	%f1049, [kRGB32f_To_601YPbPr+4];
	mul.f32 	%f1050, %f1049, %f6;
	ld.const.f32 	%f1051, [kRGB32f_To_601YPbPr+0];
	mad.f32 	%f1052, %f1051, %f7, %f1050;
	ld.const.f32 	%f1053, [kRGB32f_To_601YPbPr+8];
	mad.f32 	%f895, %f1053, %f5, %f1052;
	mov.f32 	%f198, %f895;
	ld.param.f32 	%f905, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter2+8];
	ld.param.f32 	%f1054, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter2+4];
	sub.f32 	%f1055, %f1054, %f905;
	ld.param.f32 	%f1056, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter2+16];
	ld.param.f32 	%f1057, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter2+12];
	sub.f32 	%f1058, %f1057, %f1056;
	mov.f32 	%f1059, 0f3f800000;  	// 1
	set.neu.u32.f32 	%r535, %f1058, %f1059;
	neg.s32 	%r536, %r535;
	mov.f32 	%f1060, 0f3f800000;  	// 1
	set.neu.u32.f32 	%r537, %f1055, %f1060;
	neg.s32 	%r538, %r537;
	or.b32 	%r539, %r536, %r538;
	mov.u32 	%r540, 0;
	setp.eq.s32 	%p154, %r539, %r540;
	@%p154 bra 	$Lt_2_182018;
	.loc	17	116	0
	mov.u32 	%r541, 1;
	sub.u32 	%r71, %r541, %r5;
	shr.u32 	%r72, %r11, 13;
	sub.u32 	%r73, %r5, %r11;
	sub.u32 	%r74, %r71, %r11;
	xor.b32 	%r75, %r74, %r72;
	shl.b32 	%r76, %r75, 8;
	sub.u32 	%r77, %r73, %r75;
	sub.u32 	%r78, %r11, %r75;
	xor.b32 	%r79, %r77, %r76;
	shr.u32 	%r80, %r79, 13;
	sub.u32 	%r81, %r78, %r79;
	xor.b32 	%r82, %r81, %r80;
	shr.u32 	%r542, %r82, 12;
	sub.u32 	%r543, %r75, %r79;
	sub.u32 	%r544, %r543, %r82;
	xor.b32 	%r86, %r542, %r544;
	.loc	17	117	0
	sub.u32 	%r545, %r79, %r82;
	sub.u32 	%r88, %r545, %r86;
	shl.b32 	%r546, %r86, 16;
	xor.b32 	%r90, %r88, %r546;
	.loc	17	118	0
	sub.u32 	%r547, %r82, %r86;
	sub.u32 	%r92, %r547, %r90;
	shr.u32 	%r548, %r90, 5;
	xor.b32 	%r94, %r92, %r548;
	.loc	17	119	0
	sub.u32 	%r95, %r86, %r90;
	sub.u32 	%r96, %r95, %r94;
	shr.u32 	%r549, %r94, 3;
	xor.b32 	%r98, %r96, %r549;
	.loc	17	120	0
	sub.u32 	%r99, %r90, %r94;
	sub.u32 	%r100, %r99, %r98;
	shl.b32 	%r550, %r98, 10;
	xor.b32 	%r102, %r100, %r550;
	.loc	17	121	0
	sub.u32 	%r103, %r94, %r98;
	sub.u32 	%r104, %r103, %r102;
	shr.u32 	%r551, %r102, 15;
	xor.b32 	%r106, %r104, %r551;
	.loc	28	57	0
	mov.f32 	%f1061, 0f3b270d73;  	// 0.00254902
	mul.lo.u32 	%r552, %r106, 1103515245;
	add.u32 	%r553, %r552, 12345;
	shr.u32 	%r554, %r553, 16;
	and.b32 	%r555, %r554, 255;
	shl.b32 	%r556, %r555, 7;
	mul.lo.u32 	%r557, %r106, -1029531031;
	sub.u32 	%r558, %r557, 740551042;
	shr.u32 	%r559, %r558, 16;
	and.b32 	%r560, %r559, 255;
	xor.b32 	%r561, %r556, %r560;
	cvt.rn.f32.s32 	%f1062, %r561;
	mov.f32 	%f1063, 0f467ffe00;  	// 16383.5
	div.full.f32 	%f1064, %f1062, %f1063;
	mov.f32 	%f1065, 0fbf800000;  	// -1
	add.f32 	%f1066, %f1064, %f1065;
	mad.f32 	%f198, %f1061, %f1066, %f895;
$Lt_2_182018:
	sub.f32 	%f1067, %f198, %f905;
	ld.param.f32 	%f1068, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter2+20];
	mov.f32 	%f1069, 0f3f800000;  	// 1
	setp.neu.f32 	%p155, %f1068, %f1069;
	@!%p155 bra 	$Lt_2_182786;
	mov.f32 	%f1070, 0f00000000;  	// 0
	setp.lt.f32 	%p156, %f1067, %f1070;
	@!%p156 bra 	$Lt_2_183298;
	.loc	28	66	0
	mov.f32 	%f198, %f1056;
	bra.uni 	$Lt_2_182530;
$Lt_2_183298:
	.loc	28	70	0
	rcp.approx.f32 	%f1071, %f1055;
	mul.f32 	%f1072, %f1067, %f1071;
	lg2.approx.f32 	%f1073, %f1072;
	mul.f32 	%f1074, %f1068, %f1073;
	ex2.approx.f32 	%f1075, %f1074;
	mad.f32 	%f198, %f1058, %f1075, %f1056;
	bra.uni 	$Lt_2_182530;
$Lt_2_182786:
	.loc	28	77	0
	rcp.approx.f32 	%f1076, %f1055;
	mul.f32 	%f1077, %f1067, %f1076;
	mad.f32 	%f198, %f1058, %f1077, %f1056;
$Lt_2_182530:
	.loc	19	267	0
	ld.param.f32 	%f1078, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter2+28];
	sin.approx.f32 	%f1079, %f1078;
	ld.const.f32 	%f1080, [kRGB32f_To_601YPbPr+28];
	mul.f32 	%f1081, %f1080, %f6;
	ld.const.f32 	%f1082, [kRGB32f_To_601YPbPr+16];
	mul.f32 	%f1083, %f1082, %f6;
	cos.approx.f32 	%f1084, %f1078;
	ld.const.f32 	%f1085, [kRGB32f_To_601YPbPr+24];
	mad.f32 	%f1086, %f1085, %f7, %f1081;
	ld.const.f32 	%f1087, [kRGB32f_To_601YPbPr+12];
	mad.f32 	%f1088, %f1087, %f7, %f1083;
	ld.const.f32 	%f1089, [kRGB32f_To_601YPbPr+32];
	mad.f32 	%f1090, %f1089, %f5, %f1086;
	ld.const.f32 	%f1091, [kRGB32f_To_601YPbPr+20];
	mad.f32 	%f1092, %f1091, %f5, %f1088;
	mul.f32 	%f1093, %f1079, %f1090;
	mul.f32 	%f1094, %f1084, %f1090;
	mul.f32 	%f1095, %f1092, %f1084;
	sub.f32 	%f1096, %f1095, %f1093;
	mad.f32 	%f1097, %f1092, %f1079, %f1094;
	ld.param.f32 	%f1098, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter2+32];
	sub.f32 	%f1099, %f1098, %f1096;
	ld.param.f32 	%f1100, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter2+36];
	sub.f32 	%f1101, %f1100, %f1097;
	ld.param.f32 	%f1102, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter2+40];
	mad.f32 	%f1103, %f1102, %f1099, %f1096;
	mad.f32 	%f1104, %f1102, %f1101, %f1097;
	ld.param.f32 	%f1105, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter2+24];
	mul.f32 	%f1106, %f1103, %f1105;
	mul.f32 	%f1107, %f1104, %f1105;
	ld.const.f32 	%f1108, [k601YPbPr_To_RGB32f+16];
	mul.f32 	%f1109, %f1108, %f1106;
	ld.const.f32 	%f1110, [k601YPbPr_To_RGB32f+12];
	mad.f32 	%f1111, %f1110, %f198, %f1109;
	ld.const.f32 	%f1112, [k601YPbPr_To_RGB32f+20];
	mad.f32 	%f258, %f1112, %f1107, %f1111;
	.loc	19	268	0
	ld.const.f32 	%f1113, [k601YPbPr_To_RGB32f+4];
	mul.f32 	%f1114, %f1113, %f1106;
	ld.const.f32 	%f1115, [k601YPbPr_To_RGB32f+0];
	mad.f32 	%f1116, %f1115, %f198, %f1114;
	ld.const.f32 	%f1117, [k601YPbPr_To_RGB32f+8];
	mad.f32 	%f264, %f1117, %f1107, %f1116;
	.loc	28	92	0
	ld.const.f32 	%f1118, [k601YPbPr_To_RGB32f+28];
	mul.f32 	%f1119, %f1118, %f1106;
	ld.const.f32 	%f1120, [k601YPbPr_To_RGB32f+24];
	mad.f32 	%f1121, %f1120, %f198, %f1119;
	ld.const.f32 	%f1122, [k601YPbPr_To_RGB32f+32];
	mad.f32 	%f5, %f1122, %f1107, %f1121;
	mov.f32 	%f6, %f258;
	mov.f32 	%f7, %f264;
	bra.uni 	$Lt_2_514;
$Lt_2_2818:
	.loc	35	83	0
	ld.param.f32 	%f1123, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter2+44];
	mov.f32 	%f1124, 0f00000000;  	// 0
	setp.ge.f32 	%p157, %f1123, %f1124;
	@!%p157 bra 	$L_2_153858;
	cvt.rn.f32.s32 	%f1125, %r6;
	cvt.rn.f32.s32 	%f1126, %r14;
	mul.f32 	%f1127, %f1126, %f1123;
	setp.lt.f32 	%p158, %f1125, %f1127;
	@%p158 bra 	$L_2_153602;
$L_2_153858:
	mov.f32 	%f1128, 0f00000000;  	// 0
	setp.lt.f32 	%p159, %f1123, %f1128;
	@!%p159 bra 	$Lt_2_514;
	cvt.rn.f32.s32 	%f1129, %r12;
	cvt.rn.f32.s32 	%f1130, %r13;
	mul.f32 	%f1131, %f1130, %f1123;
	neg.f32 	%f1132, %f1131;
	setp.lt.f32 	%p160, %f1129, %f1132;
	@!%p160 bra 	$Lt_2_514;
$L_2_153602:
	.loc	28	110	0
	ld.const.f32 	%f1133, [kRGB32f_To_601YPbPr+4];
	mul.f32 	%f891, %f1133, %f6;
	ld.param.f32 	%f905, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter2+8];
	ld.param.f32 	%f1134, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter2+4];
	sub.f32 	%f1135, %f1134, %f905;
	ld.param.f32 	%f1136, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter2+16];
	ld.param.f32 	%f1137, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter2+12];
	sub.f32 	%f1138, %f1137, %f1136;
	ld.const.f32 	%f1139, [kRGB32f_To_601YPbPr+0];
	mad.f32 	%f893, %f1139, %f7, %f891;
	ld.const.f32 	%f1140, [kRGB32f_To_601YPbPr+8];
	mad.f32 	%f895, %f1140, %f5, %f893;
	sub.f32 	%f1141, %f895, %f905;
	ld.param.f32 	%f1142, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter2+20];
	mov.f32 	%f1143, 0f3f800000;  	// 1
	setp.neu.f32 	%p161, %f1142, %f1143;
	@!%p161 bra 	$Lt_2_183810;
	.loc	28	120	0
	mov.f32 	%f1144, 0f00000000;  	// 0
	max.f32 	%f1145, %f1141, %f1144;
	div.full.f32 	%f1146, %f1145, %f1135;
	lg2.approx.f32 	%f1147, %f1146;
	mul.f32 	%f1148, %f1142, %f1147;
	ex2.approx.f32 	%f1149, %f1148;
	mad.f32 	%f297, %f1138, %f1149, %f1136;
	bra.uni 	$Lt_2_183554;
$Lt_2_183810:
	.loc	28	129	0
	div.full.f32 	%f1150, %f1141, %f1135;
	mad.f32 	%f297, %f1138, %f1150, %f1136;
$Lt_2_183554:
	.loc	28	135	0
	ld.const.f32 	%f1151, [k601YPbPr_To_RGB32f+24];
	mul.f32 	%f5, %f1151, %f297;
	ld.const.f32 	%f1152, [k601YPbPr_To_RGB32f+12];
	mul.f32 	%f6, %f1152, %f297;
	ld.const.f32 	%f1153, [k601YPbPr_To_RGB32f+0];
	mul.f32 	%f7, %f1153, %f297;
	bra.uni 	$Lt_2_514;
$Lt_2_3074:
	.loc	35	86	0
	ld.param.f32 	%f1154, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter2+44];
	mov.f32 	%f1155, 0f00000000;  	// 0
	setp.ge.f32 	%p162, %f1154, %f1155;
	@!%p162 bra 	$L_2_154626;
	cvt.rn.f32.s32 	%f1156, %r6;
	cvt.rn.f32.s32 	%f1157, %r14;
	mul.f32 	%f1158, %f1157, %f1154;
	setp.lt.f32 	%p163, %f1156, %f1158;
	@%p163 bra 	$L_2_154370;
$L_2_154626:
	mov.f32 	%f1159, 0f00000000;  	// 0
	setp.lt.f32 	%p164, %f1154, %f1159;
	@!%p164 bra 	$Lt_2_514;
	cvt.rn.f32.s32 	%f1160, %r12;
	cvt.rn.f32.s32 	%f1161, %r13;
	mul.f32 	%f1162, %f1161, %f1154;
	neg.f32 	%f1163, %f1162;
	setp.lt.f32 	%p165, %f1160, %f1163;
	@!%p165 bra 	$Lt_2_514;
$L_2_154370:
	.loc	28	160	0
	ld.const.f32 	%f5, [k601YPbPr_To_RGB32f+24];
	ld.const.f32 	%f6, [k601YPbPr_To_RGB32f+12];
	ld.const.f32 	%f7, [k601YPbPr_To_RGB32f+0];
	bra.uni 	$Lt_2_514;
$Lt_2_3330:
	.loc	29	42	0
	ld.param.f32 	%f880, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter2+4];
	mov.f32 	%f1164, 0f00000000;  	// 0
	setp.ge.f32 	%p166, %f5, %f1164;
	@!%p166 bra 	$Lt_2_184322;
	.loc	19	292	0
	lg2.approx.f32 	%f1165, %f5;
	mul.f32 	%f1166, %f880, %f1165;
	ex2.approx.f32 	%f1167, %f1166;
	bra.uni 	$Lt_2_184066;
$Lt_2_184322:
	neg.f32 	%f1168, %f5;
	lg2.approx.f32 	%f1169, %f1168;
	mul.f32 	%f1170, %f880, %f1169;
	ex2.approx.f32 	%f1171, %f1170;
	neg.f32 	%f1167, %f1171;
$Lt_2_184066:
	mov.f32 	%f1172, 0f00000000;  	// 0
	setp.ge.f32 	%p167, %f6, %f1172;
	@!%p167 bra 	$Lt_2_184834;
	.loc	19	293	0
	lg2.approx.f32 	%f1173, %f6;
	mul.f32 	%f1174, %f880, %f1173;
	ex2.approx.f32 	%f1175, %f1174;
	bra.uni 	$Lt_2_184578;
$Lt_2_184834:
	neg.f32 	%f1176, %f6;
	lg2.approx.f32 	%f1177, %f1176;
	mul.f32 	%f1178, %f880, %f1177;
	ex2.approx.f32 	%f1179, %f1178;
	neg.f32 	%f1175, %f1179;
$Lt_2_184578:
	mov.f32 	%f1180, 0f00000000;  	// 0
	setp.ge.f32 	%p168, %f7, %f1180;
	@!%p168 bra 	$Lt_2_185346;
	.loc	19	294	0
	lg2.approx.f32 	%f1181, %f7;
	mul.f32 	%f1182, %f880, %f1181;
	ex2.approx.f32 	%f1183, %f1182;
	bra.uni 	$Lt_2_185090;
$Lt_2_185346:
	neg.f32 	%f1184, %f7;
	lg2.approx.f32 	%f1185, %f1184;
	mul.f32 	%f1186, %f880, %f1185;
	ex2.approx.f32 	%f1187, %f1186;
	neg.f32 	%f1183, %f1187;
$Lt_2_185090:
	mov.f32 	%f1188, 0f00000000;  	// 0
	setp.ge.f32 	%p169, %f8, %f1188;
	@!%p169 bra 	$Lt_2_185858;
	.loc	19	295	0
	lg2.approx.f32 	%f1189, %f8;
	mul.f32 	%f1190, %f880, %f1189;
	ex2.approx.f32 	%f1191, %f1190;
	bra.uni 	$Lt_2_185602;
$Lt_2_185858:
	neg.f32 	%f1192, %f8;
	lg2.approx.f32 	%f1193, %f1192;
	mul.f32 	%f1194, %f880, %f1193;
	ex2.approx.f32 	%f1195, %f1194;
	neg.f32 	%f1191, %f1195;
$Lt_2_185602:
	.loc	29	42	0
	mov.f32 	%f5, %f1167;
	mov.f32 	%f6, %f1175;
	mov.f32 	%f7, %f1183;
	mov.f32 	%f8, %f1191;
	.loc	35	91	0
	bra.uni 	$Lt_2_514;
$Lt_2_3586:
	.loc	30	41	0
	sub.s32 	%r562, %r14, %r6;
	sub.s32 	%r6, %r562, 1;
	.loc	35	94	0
	bra.uni 	$Lt_2_514;
$Lt_2_3842:
	.loc	35	96	0
	sub.u32 	%r343, %r6, %r12;
	ld.param.f32 	%f880, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter2+4];
	mov.f32 	%f1196, 0f3e800000;  	// 0.25
	mul.f32 	%f1197, %f880, %f1196;
	ld.param.f32 	%f1198, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter2+16];
	ld.param.f32 	%f1199, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter2+8];
	mov.f32 	%f1200, 0f3f800000;  	// 1
	setp.eq.f32 	%p170, %f1199, %f1200;
	@!%p170 bra 	$Lt_2_186370;
	.loc	17	117	0
	cvt.rzi.u32.f32 	%r563, %f1198;
	shr.u32 	%r564, %r563, 13;
	sub.u32 	%r565, %r12, %r563;
	sub.u32 	%r566, %r343, %r563;
	xor.b32 	%r567, %r566, %r564;
	shl.b32 	%r568, %r567, 8;
	sub.u32 	%r569, %r565, %r567;
	sub.u32 	%r570, %r563, %r567;
	xor.b32 	%r571, %r569, %r568;
	shr.u32 	%r572, %r571, 13;
	sub.u32 	%r573, %r570, %r571;
	sub.u32 	%r574, %r567, %r571;
	xor.b32 	%r575, %r573, %r572;
	shr.u32 	%r576, %r575, 12;
	sub.u32 	%r577, %r574, %r575;
	xor.b32 	%r578, %r577, %r576;
	shl.b32 	%r579, %r578, 16;
	sub.u32 	%r580, %r571, %r575;
	sub.u32 	%r581, %r580, %r578;
	xor.b32 	%r582, %r579, %r581;
	.loc	17	118	0
	sub.u32 	%r583, %r575, %r578;
	sub.u32 	%r584, %r583, %r582;
	shr.u32 	%r585, %r582, 5;
	xor.b32 	%r586, %r584, %r585;
	.loc	17	119	0
	sub.u32 	%r587, %r578, %r582;
	sub.u32 	%r588, %r587, %r586;
	shr.u32 	%r589, %r586, 3;
	xor.b32 	%r590, %r588, %r589;
	.loc	17	120	0
	sub.u32 	%r591, %r582, %r586;
	sub.u32 	%r592, %r591, %r590;
	shl.b32 	%r593, %r590, 10;
	xor.b32 	%r594, %r592, %r593;
	.loc	17	121	0
	sub.u32 	%r595, %r586, %r590;
	sub.u32 	%r596, %r595, %r594;
	shr.u32 	%r597, %r594, 15;
	xor.b32 	%r598, %r596, %r597;
	.loc	31	48	0
	mul.lo.u32 	%r599, %r598, 1103515245;
	add.u32 	%r600, %r599, 12345;
	shr.u32 	%r601, %r600, 16;
	and.b32 	%r602, %r601, 255;
	shl.b32 	%r603, %r602, 7;
	mul.lo.u32 	%r604, %r598, -1029531031;
	sub.u32 	%r605, %r604, 740551042;
	shr.u32 	%r606, %r605, 16;
	and.b32 	%r607, %r606, 255;
	xor.b32 	%r608, %r603, %r607;
	cvt.rn.f32.s32 	%f1201, %r608;
	mov.f32 	%f1202, 0f46fffe00;  	// 32767
	div.full.f32 	%f1203, %f1201, %f1202;
	mul.f32 	%f1204, %f880, %f1203;
	mov.f32 	%f1205, 0f3f000000;  	// 0.5
	mul.f32 	%f1206, %f1204, %f1205;
	sub.f32 	%f1207, %f1206, %f1197;
	.loc	17	117	0
	add.u32 	%r609, %r566, 1;
	xor.b32 	%r610, %r609, %r564;
	shl.b32 	%r611, %r610, 8;
	sub.u32 	%r612, %r565, %r610;
	sub.u32 	%r613, %r563, %r610;
	xor.b32 	%r614, %r611, %r612;
	shr.u32 	%r615, %r614, 13;
	sub.u32 	%r616, %r613, %r614;
	sub.u32 	%r617, %r610, %r614;
	xor.b32 	%r618, %r615, %r616;
	shr.u32 	%r619, %r618, 12;
	sub.u32 	%r620, %r617, %r618;
	xor.b32 	%r621, %r619, %r620;
	sub.u32 	%r622, %r614, %r618;
	sub.u32 	%r178, %r622, %r621;
	shl.b32 	%r623, %r621, 16;
	xor.b32 	%r178, %r178, %r623;
	.loc	17	118	0
	sub.u32 	%r180, %r618, %r621;
	sub.u32 	%r180, %r180, %r178;
	shr.u32 	%r624, %r178, 5;
	xor.b32 	%r180, %r180, %r624;
	.loc	17	119	0
	sub.u32 	%r182, %r621, %r178;
	sub.u32 	%r182, %r182, %r180;
	shr.u32 	%r625, %r180, 3;
	xor.b32 	%r182, %r182, %r625;
	.loc	17	120	0
	sub.u32 	%r178, %r178, %r180;
	sub.u32 	%r178, %r178, %r182;
	shl.b32 	%r626, %r182, 10;
	xor.b32 	%r178, %r178, %r626;
	.loc	17	121	0
	sub.u32 	%r180, %r180, %r182;
	sub.u32 	%r180, %r180, %r178;
	shr.u32 	%r627, %r178, 15;
	xor.b32 	%r180, %r180, %r627;
	.loc	31	49	0
	mul.lo.u32 	%r628, %r180, 1103515245;
	add.u32 	%r629, %r628, 12345;
	shr.u32 	%r630, %r629, 16;
	and.b32 	%r631, %r630, 255;
	shl.b32 	%r632, %r631, 7;
	mul.lo.u32 	%r633, %r180, -1029531031;
	sub.u32 	%r634, %r633, 740551042;
	shr.u32 	%r635, %r634, 16;
	and.b32 	%r636, %r635, 255;
	xor.b32 	%r637, %r632, %r636;
	cvt.rn.f32.s32 	%f1208, %r637;
	mov.f32 	%f1209, 0f46fffe00;  	// 32767
	div.full.f32 	%f1210, %f1208, %f1209;
	mul.f32 	%f1211, %f880, %f1210;
	mov.f32 	%f1212, 0f3f000000;  	// 0.5
	mul.f32 	%f1213, %f1211, %f1212;
	sub.f32 	%f362, %f1213, %f1197;
	.loc	17	117	0
	add.u32 	%r638, %r566, 2;
	xor.b32 	%r639, %r638, %r564;
	shl.b32 	%r640, %r639, 8;
	sub.u32 	%r641, %r565, %r639;
	sub.u32 	%r642, %r563, %r639;
	xor.b32 	%r643, %r640, %r641;
	shr.u32 	%r644, %r643, 13;
	sub.u32 	%r645, %r642, %r643;
	sub.u32 	%r646, %r639, %r643;
	xor.b32 	%r647, %r644, %r645;
	shr.u32 	%r648, %r647, 12;
	sub.u32 	%r649, %r646, %r647;
	xor.b32 	%r650, %r648, %r649;
	sub.u32 	%r651, %r643, %r647;
	sub.u32 	%r210, %r651, %r650;
	shl.b32 	%r652, %r650, 16;
	xor.b32 	%r210, %r210, %r652;
	.loc	17	118	0
	sub.u32 	%r212, %r647, %r650;
	sub.u32 	%r212, %r212, %r210;
	shr.u32 	%r653, %r210, 5;
	xor.b32 	%r212, %r212, %r653;
	.loc	17	119	0
	sub.u32 	%r214, %r650, %r210;
	sub.u32 	%r214, %r214, %r212;
	shr.u32 	%r654, %r212, 3;
	xor.b32 	%r214, %r214, %r654;
	.loc	17	120	0
	sub.u32 	%r210, %r210, %r212;
	sub.u32 	%r210, %r210, %r214;
	shl.b32 	%r655, %r214, 10;
	xor.b32 	%r210, %r210, %r655;
	.loc	17	121	0
	sub.u32 	%r212, %r212, %r214;
	sub.u32 	%r212, %r212, %r210;
	shr.u32 	%r656, %r210, 15;
	xor.b32 	%r212, %r212, %r656;
	.loc	31	50	0
	mul.lo.u32 	%r657, %r212, 1103515245;
	add.u32 	%r658, %r657, 12345;
	shr.u32 	%r659, %r658, 16;
	and.b32 	%r660, %r659, 255;
	shl.b32 	%r661, %r660, 7;
	mul.lo.u32 	%r662, %r212, -1029531031;
	sub.u32 	%r663, %r662, 740551042;
	shr.u32 	%r664, %r663, 16;
	and.b32 	%r665, %r664, 255;
	xor.b32 	%r666, %r661, %r665;
	cvt.rn.f32.s32 	%f1214, %r666;
	mov.f32 	%f1215, 0f46fffe00;  	// 32767
	div.full.f32 	%f1216, %f1214, %f1215;
	mul.f32 	%f1217, %f880, %f1216;
	mov.f32 	%f1218, 0f3f000000;  	// 0.5
	mul.f32 	%f1219, %f1217, %f1218;
	sub.f32 	%f369, %f1219, %f1197;
	bra.uni 	$Lt_2_186114;
$Lt_2_186370:
	.loc	17	117	0
	cvt.rzi.u32.f32 	%r667, %f1198;
	shr.u32 	%r668, %r667, 13;
	sub.u32 	%r669, %r12, %r667;
	sub.u32 	%r670, %r343, %r667;
	xor.b32 	%r671, %r670, %r668;
	shl.b32 	%r672, %r671, 8;
	sub.u32 	%r673, %r669, %r671;
	sub.u32 	%r674, %r667, %r671;
	xor.b32 	%r675, %r673, %r672;
	shr.u32 	%r676, %r675, 13;
	sub.u32 	%r677, %r674, %r675;
	sub.u32 	%r678, %r671, %r675;
	xor.b32 	%r679, %r677, %r676;
	shr.u32 	%r680, %r679, 12;
	sub.u32 	%r681, %r678, %r679;
	xor.b32 	%r682, %r681, %r680;
	shl.b32 	%r683, %r682, 16;
	sub.u32 	%r684, %r675, %r679;
	sub.u32 	%r685, %r684, %r682;
	xor.b32 	%r686, %r683, %r685;
	.loc	17	118	0
	sub.u32 	%r687, %r679, %r682;
	sub.u32 	%r688, %r687, %r686;
	shr.u32 	%r689, %r686, 5;
	xor.b32 	%r690, %r688, %r689;
	.loc	17	119	0
	sub.u32 	%r691, %r682, %r686;
	sub.u32 	%r692, %r691, %r690;
	shr.u32 	%r693, %r690, 3;
	xor.b32 	%r694, %r692, %r693;
	.loc	17	120	0
	sub.u32 	%r695, %r686, %r690;
	sub.u32 	%r696, %r695, %r694;
	shl.b32 	%r697, %r694, 10;
	xor.b32 	%r698, %r696, %r697;
	.loc	17	121	0
	sub.u32 	%r699, %r690, %r694;
	sub.u32 	%r700, %r699, %r698;
	shr.u32 	%r701, %r698, 15;
	xor.b32 	%r702, %r700, %r701;
	.loc	31	54	0
	mul.lo.u32 	%r703, %r702, 1103515245;
	mul.lo.u32 	%r704, %r702, -1029531031;
	add.u32 	%r705, %r703, 12345;
	sub.u32 	%r706, %r704, 740551042;
	shr.u32 	%r707, %r705, 16;
	shr.u32 	%r708, %r706, 16;
	and.b32 	%r709, %r707, 255;
	and.b32 	%r710, %r708, 255;
	shl.b32 	%r711, %r709, 7;
	xor.b32 	%r712, %r711, %r710;
	cvt.rn.f32.s32 	%f1220, %r712;
	mov.f32 	%f1221, 0f46fffe00;  	// 32767
	div.full.f32 	%f1222, %f1220, %f1221;
	mul.f32 	%f1223, %f1222, %f880;
	mov.f32 	%f1224, 0f3f000000;  	// 0.5
	mul.f32 	%f1225, %f1223, %f1224;
	sub.f32 	%f1207, %f1225, %f1197;
	mov.f32 	%f369, %f1207;
	mov.f32 	%f362, %f1207;
$Lt_2_186114:
	.loc	31	57	0
	add.f32 	%f7, %f1207, %f7;
	.loc	31	58	0
	add.f32 	%f6, %f362, %f6;
	.loc	31	59	0
	add.f32 	%f5, %f5, %f369;
	ld.param.f32 	%f1226, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter2+12];
	mov.f32 	%f1227, 0f3f800000;  	// 1
	setp.eq.f32 	%p171, %f1226, %f1227;
	@!%p171 bra 	$Lt_2_514;
	.loc	31	63	0
	mov.f32 	%f1228, 0f00000000;  	// 0
	max.f32 	%f1229, %f5, %f1228;
	mov.f32 	%f1230, 0f3f800000;  	// 1
	min.f32 	%f5, %f1229, %f1230;
	mov.f32 	%f1231, 0f00000000;  	// 0
	max.f32 	%f1232, %f6, %f1231;
	mov.f32 	%f1233, 0f3f800000;  	// 1
	min.f32 	%f6, %f1232, %f1233;
	mov.f32 	%f1234, 0f00000000;  	// 0
	max.f32 	%f1235, %f7, %f1234;
	mov.f32 	%f1236, 0f3f800000;  	// 1
	min.f32 	%f7, %f1235, %f1236;
	mov.f32 	%f1237, 0f00000000;  	// 0
	max.f32 	%f1238, %f8, %f1237;
	mov.f32 	%f1239, 0f3f800000;  	// 1
	min.f32 	%f8, %f1238, %f1239;
	bra.uni 	$Lt_2_514;
$Lt_2_4098:
	.loc	35	99	0
	cvt.rn.f32.s32 	%f1240, %r6;
	cvt.rn.f32.s32 	%f1241, %r14;
	ld.param.f32 	%f1242, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter2+20];
	mul.f32 	%f1243, %f1241, %f1242;
	setp.lt.f32 	%p172, %f1240, %f1243;
	@!%p172 bra 	$Lt_2_514;
	.loc	19	267	0
	ld.const.f32 	%f1244, [kRGB32f_To_601YPbPr+4];
	mul.f32 	%f891, %f1244, %f6;
	ld.const.f32 	%f1245, [kRGB32f_To_601YPbPr+28];
	mul.f32 	%f1246, %f1245, %f6;
	ld.const.f32 	%f1247, [kRGB32f_To_601YPbPr+16];
	mul.f32 	%f1248, %f1247, %f6;
	ld.const.f32 	%f1249, [kRGB32f_To_601YPbPr+0];
	mad.f32 	%f893, %f1249, %f7, %f891;
	ld.const.f32 	%f1250, [kRGB32f_To_601YPbPr+24];
	mad.f32 	%f1251, %f1250, %f7, %f1246;
	ld.const.f32 	%f1252, [kRGB32f_To_601YPbPr+12];
	mad.f32 	%f1253, %f1252, %f7, %f1248;
	ld.const.f32 	%f1254, [kRGB32f_To_601YPbPr+8];
	mad.f32 	%f895, %f1254, %f5, %f893;
	ld.const.f32 	%f1255, [kRGB32f_To_601YPbPr+32];
	mad.f32 	%f1256, %f1255, %f5, %f1251;
	ld.const.f32 	%f1257, [kRGB32f_To_601YPbPr+20];
	mad.f32 	%f1258, %f1257, %f5, %f1253;
	ld.param.f32 	%f1259, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter2+4];
	ld.param.f32 	%f1260, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter2+8];
	mad.f32 	%f1261, %f895, %f1260, %f1259;
	ld.param.f32 	%f1262, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter2+16];
	mul.f32 	%f1263, %f1256, %f1262;
	ld.param.f32 	%f1264, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter2+12];
	mul.f32 	%f1265, %f1256, %f1264;
	mul.f32 	%f1266, %f1258, %f1264;
	sub.f32 	%f1267, %f1266, %f1263;
	mad.f32 	%f1268, %f1258, %f1262, %f1265;
	ld.const.f32 	%f1269, [k601YPbPr_To_RGB32f+16];
	mul.f32 	%f1270, %f1269, %f1267;
	ld.const.f32 	%f1271, [k601YPbPr_To_RGB32f+12];
	mad.f32 	%f1272, %f1271, %f1261, %f1270;
	ld.const.f32 	%f1273, [k601YPbPr_To_RGB32f+20];
	mad.f32 	%f424, %f1273, %f1268, %f1272;
	.loc	19	268	0
	ld.const.f32 	%f1274, [k601YPbPr_To_RGB32f+4];
	mul.f32 	%f1275, %f1274, %f1267;
	ld.const.f32 	%f1276, [k601YPbPr_To_RGB32f+0];
	mad.f32 	%f1277, %f1276, %f1261, %f1275;
	ld.const.f32 	%f1278, [k601YPbPr_To_RGB32f+8];
	mad.f32 	%f430, %f1278, %f1268, %f1277;
	.loc	32	56	0
	ld.const.f32 	%f1279, [k601YPbPr_To_RGB32f+28];
	mul.f32 	%f1280, %f1279, %f1267;
	ld.const.f32 	%f1281, [k601YPbPr_To_RGB32f+24];
	mad.f32 	%f1282, %f1281, %f1261, %f1280;
	ld.const.f32 	%f1283, [k601YPbPr_To_RGB32f+32];
	mad.f32 	%f5, %f1283, %f1268, %f1282;
	mov.f32 	%f6, %f424;
	mov.f32 	%f7, %f430;
	bra.uni 	$Lt_2_514;
$Lt_2_4354:
	.loc	33	46	0
	ld.const.f32 	%f1284, [kRGB32f_To_601YPbPr+4];
	mul.f32 	%f891, %f1284, %f6;
	ld.const.f32 	%f1285, [kRGB32f_To_601YPbPr+0];
	mad.f32 	%f893, %f1285, %f7, %f891;
	ld.const.f32 	%f1286, [kRGB32f_To_601YPbPr+8];
	mad.f32 	%f895, %f1286, %f5, %f893;
	ld.param.f32 	%f905, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter2+8];
	ld.param.f32 	%f1287, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter2+20];
	sub.f32 	%f1288, %f1287, %f905;
	mad.f32 	%f441, %f895, %f1288, %f905;
	.loc	33	47	0
	ld.param.f32 	%f1289, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter2+12];
	ld.param.f32 	%f1290, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter2+24];
	sub.f32 	%f1291, %f1290, %f1289;
	mad.f32 	%f445, %f895, %f1291, %f1289;
	.loc	33	49	0
	ld.param.f32 	%f1292, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter2+28];
	ld.param.f32 	%f880, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter2+4];
	ld.param.f32 	%f1293, [__cudaparm_PointwiseFilterHostKernel3___val_paraminFilter2+16];
	sub.f32 	%f1294, %f1293, %f880;
	mad.f32 	%f1295, %f895, %f1294, %f880;
	sub.f32 	%f1296, %f1295, %f5;
	mad.f32 	%f5, %f1292, %f1296, %f5;
	.loc	33	50	0
	sub.f32 	%f1297, %f441, %f6;
	mad.f32 	%f6, %f1292, %f1297, %f6;
	.loc	33	51	0
	sub.f32 	%f1298, %f445, %f7;
	mad.f32 	%f7, %f1292, %f1298, %f7;
	.loc	35	103	0
	bra.uni 	$Lt_2_514;
$Lt_2_4610:
	.loc	34	41	0
	sub.s32 	%r713, %r13, %r12;
	sub.s32 	%r12, %r713, 1;
$Lt_2_514:
	.loc	17	100	0
	ld.param.s32 	%r714, [__cudaparm_PointwiseFilterHostKernel3_inDestPitch];
	mul24.lo.s32 	%r715, %r12, %r714;
	ld.param.u64 	%rd5, [__cudaparm_PointwiseFilterHostKernel3_inDestImage];
	add.s32 	%r716, %r715, %r6;
	cvt.s64.s32 	%rd6, %r716;
	mul.lo.u64 	%rd7, %rd6, 16;
	add.u64 	%rd8, %rd5, %rd7;
	st.global.v4.f32 	[%rd8+0], {%f5,%f6,%f7,%f8};
$Lt_2_154882:
	.loc	35	182	0
	exit;
$LDWend_PointwiseFilterHostKernel3:
	} // PointwiseFilterHostKernel3

	.entry PointwiseFilterHostKernel4 (
		.param .u64 __cudaparm_PointwiseFilterHostKernel4_inSrcImage,
		.param .s32 __cudaparm_PointwiseFilterHostKernel4_inSrcPitch,
		.param .u64 __cudaparm_PointwiseFilterHostKernel4_inDestImage,
		.param .s32 __cudaparm_PointwiseFilterHostKernel4_inDestPitch,
		.param .s32 __cudaparm_PointwiseFilterHostKernel4_inWidth,
		.param .s32 __cudaparm_PointwiseFilterHostKernel4_inHeight,
		.param .align 4 .b8 __cudaparm_PointwiseFilterHostKernel4___val_paraminFilter0[48],
		.param .align 4 .b8 __cudaparm_PointwiseFilterHostKernel4___val_paraminFilter1[48],
		.param .align 4 .b8 __cudaparm_PointwiseFilterHostKernel4___val_paraminFilter2[48],
		.param .align 4 .b8 __cudaparm_PointwiseFilterHostKernel4___val_paraminFilter3[48])
	{
	.reg .u32 %r<937>;
	.reg .u64 %rd<10>;
	.reg .f32 %f<1723>;
	.reg .pred %p<231>;
	.loc	35	193	0
$LBB1_PointwiseFilterHostKernel4:
	.loc	35	195	0
	cvt.s32.u16 	%r1, %ctaid.x;
	cvt.s32.u16 	%r2, %ntid.x;
	mul24.lo.s32 	%r3, %r1, %r2;
	cvt.u32.u16 	%r4, %tid.x;
	add.u32 	%r5, %r3, %r4;
	mov.s32 	%r6, %r5;
	.loc	35	196	0
	cvt.s32.u16 	%r7, %ctaid.y;
	cvt.s32.u16 	%r8, %ntid.y;
	mul24.lo.s32 	%r9, %r7, %r8;
	cvt.u32.u16 	%r10, %tid.y;
	add.u32 	%r11, %r9, %r10;
	mov.s32 	%r12, %r11;
	ld.param.s32 	%r13, [__cudaparm_PointwiseFilterHostKernel4_inHeight];
	ld.param.s32 	%r14, [__cudaparm_PointwiseFilterHostKernel4_inWidth];
	set.gt.u32.s32 	%r15, %r13, %r11;
	neg.s32 	%r16, %r15;
	set.gt.u32.s32 	%r17, %r14, %r5;
	neg.s32 	%r18, %r17;
	and.b32 	%r19, %r16, %r18;
	mov.u32 	%r20, 0;
	setp.eq.s32 	%p1, %r19, %r20;
	@%p1 bra 	$Lt_3_206082;
	.loc	17	89	0
	ld.param.s32 	%r21, [__cudaparm_PointwiseFilterHostKernel4_inSrcPitch];
	mul24.lo.s32 	%r22, %r11, %r21;
	ld.param.u64 	%rd1, [__cudaparm_PointwiseFilterHostKernel4_inSrcImage];
	add.s32 	%r23, %r22, %r5;
	cvt.s64.s32 	%rd2, %r23;
	mul.lo.u64 	%rd3, %rd2, 16;
	add.u64 	%rd4, %rd1, %rd3;
	ld.global.v4.f32 	{%f1,%f2,%f3,%f4}, [%rd4+0];
	.loc	35	200	0
	mov.f32 	%f5, %f1;
	mov.f32 	%f6, %f2;
	mov.f32 	%f7, %f3;
	mov.f32 	%f8, %f4;
	.loc	35	54	0
	ld.param.u32 	%r24, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter0+0];
	mov.u32 	%r25, 0;
	setp.eq.s32 	%p2, %r24, %r25;
	@%p2 bra 	$Lt_3_14850;
	mov.u32 	%r26, 1;
	setp.eq.s32 	%p3, %r24, %r26;
	@%p3 bra 	$Lt_3_15362;
	mov.u32 	%r27, 2;
	setp.eq.s32 	%p4, %r24, %r27;
	@%p4 bra 	$Lt_3_15618;
	mov.u32 	%r28, 3;
	setp.eq.s32 	%p5, %r24, %r28;
	@%p5 bra 	$Lt_3_15874;
	mov.u32 	%r29, 4;
	setp.eq.s32 	%p6, %r24, %r29;
	@%p6 bra 	$Lt_3_16130;
	mov.u32 	%r30, 5;
	setp.eq.s32 	%p7, %r24, %r30;
	@%p7 bra 	$Lt_3_16386;
	mov.u32 	%r31, 6;
	setp.eq.s32 	%p8, %r24, %r31;
	@%p8 bra 	$Lt_3_16642;
	mov.u32 	%r32, 7;
	setp.eq.s32 	%p9, %r24, %r32;
	@%p9 bra 	$Lt_3_16898;
	mov.u32 	%r33, 8;
	setp.eq.s32 	%p10, %r24, %r33;
	@%p10 bra 	$Lt_3_17154;
	mov.u32 	%r34, 9;
	setp.eq.s32 	%p11, %r24, %r34;
	@%p11 bra 	$Lt_3_17410;
	mov.u32 	%r35, 10;
	setp.eq.s32 	%p12, %r24, %r35;
	@%p12 bra 	$Lt_3_17666;
	mov.u32 	%r36, 11;
	setp.eq.s32 	%p13, %r24, %r36;
	@%p13 bra 	$Lt_3_17922;
	mov.u32 	%r37, 12;
	setp.eq.s32 	%p14, %r24, %r37;
	@%p14 bra 	$Lt_3_18178;
	mov.u32 	%r38, 13;
	setp.eq.s32 	%p15, %r24, %r38;
	@%p15 bra 	$Lt_3_18434;
	mov.u32 	%r39, 14;
	setp.eq.s32 	%p16, %r24, %r39;
	@%p16 bra 	$Lt_3_18690;
	mov.u32 	%r40, 15;
	setp.eq.s32 	%p17, %r24, %r40;
	@%p17 bra 	$Lt_3_18946;
	mov.u32 	%r41, 16;
	setp.eq.s32 	%p18, %r24, %r41;
	@%p18 bra 	$Lt_3_19202;
	bra.uni 	$Lt_3_15106;
$Lt_3_14850:
	.loc	35	57	0
	ld.param.f32 	%f9, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter0+8];
	mov.f32 	%f10, 0f3f800000;    	// 1
	setp.eq.f32 	%p19, %f9, %f10;
	ld.param.f32 	%f11, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter0+12];
	mov.f32 	%f12, 0f3f800000;    	// 1
	setp.eq.f32 	%p20, %f11, %f12;
	ld.param.f32 	%f13, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter0+4];
	ld.param.f32 	%f14, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter0+16];
	mov.f32 	%f15, 0f3f800000;    	// 1
	setp.eq.f32 	%p21, %f14, %f15;
	@!%p21 bra 	$Lt_3_206850;
	.loc	18	53	0
	cvt.sat.f32.f32 	%f16, %f4;
	mul.f32 	%f17, %f16, %f13;
	selp.f32 	%f18, %f13, %f17, %p19;
	mov.f32 	%f19, 0f3f800000;    	// 1
	sub.f32 	%f20, %f19, %f18;
	selp.f32 	%f21, %f20, %f18, %p20;
	mov.f32 	%f8, %f21;
	.loc	18	57	0
	mov.f32 	%f5, %f21;
	mov.f32 	%f6, %f21;
	mov.f32 	%f7, %f21;
	bra.uni 	$Lt_3_15106;
$Lt_3_206850:
	@!%p19 bra 	$Lt_3_207362;
	.loc	18	61	0
	mov.f32 	%f22, 0f3f800000;    	// 1
	sub.f32 	%f23, %f22, %f13;
	selp.f32 	%f8, %f23, %f13, %p20;
	bra.uni 	$Lt_3_15106;
$Lt_3_207362:
	.loc	18	73	0
	cvt.sat.f32.f32 	%f24, %f4;
	mov.f32 	%f25, 0f3f800000;    	// 1
	sub.f32 	%f26, %f25, %f24;
	selp.f32 	%f27, %f26, %f24, %p20;
	mul.f32 	%f8, %f27, %f13;
	bra.uni 	$Lt_3_15106;
$Lt_3_15362:
	.loc	19	267	0
	ld.const.f32 	%f28, [kRGB32f_To_601YPbPr+4];
	mul.f32 	%f29, %f28, %f2;
	ld.const.f32 	%f30, [kRGB32f_To_601YPbPr+0];
	mad.f32 	%f31, %f30, %f3, %f29;
	ld.const.f32 	%f32, [kRGB32f_To_601YPbPr+8];
	mad.f32 	%f33, %f32, %f1, %f31;
	ld.const.f32 	%f34, [k601YPbPr_To_RGB32f+12];
	mul.f32 	%f35, %f34, %f33;
	.loc	19	268	0
	ld.const.f32 	%f36, [k601YPbPr_To_RGB32f+0];
	mul.f32 	%f37, %f36, %f33;
	.loc	20	44	0
	ld.const.f32 	%f38, [k601YPbPr_To_RGB32f+24];
	mul.f32 	%f5, %f38, %f33;
	mov.f32 	%f6, %f35;
	mov.f32 	%f7, %f37;
	mov.f32 	%f8, %f4;
	.loc	35	61	0
	bra.uni 	$Lt_3_15106;
$Lt_3_15618:
	.loc	35	63	0
	ld.param.f32 	%f39, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter0+20];
	ld.param.f32 	%f40, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter0+12];
	ld.param.f32 	%f41, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter0+4];
	mov.f32 	%f42, 0f00000000;    	// 0
	setp.neu.f32 	%p22, %f41, %f42;
	@!%p22 bra 	$Lt_3_207874;
	.loc	21	44	0
	ld.param.f32 	%f43, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter0+8];
	sub.f32 	%f44, %f3, %f40;
	mad.f32 	%f7, %f43, %f44, %f39;
	.loc	21	45	0
	sub.f32 	%f45, %f2, %f40;
	mad.f32 	%f6, %f43, %f45, %f39;
	.loc	21	46	0
	sub.f32 	%f46, %f1, %f40;
	mad.f32 	%f5, %f43, %f46, %f39;
	bra.uni 	$Lt_3_15106;
$Lt_3_207874:
	.loc	21	50	0
	ld.param.f32 	%f47, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter0+24];
	setp.gt.f32 	%p23, %f3, %f40;
	selp.f32 	%f7, %f47, %f39, %p23;
	.loc	21	51	0
	setp.gt.f32 	%p24, %f2, %f40;
	selp.f32 	%f6, %f47, %f39, %p24;
	.loc	21	52	0
	setp.gt.f32 	%p25, %f1, %f40;
	selp.f32 	%f5, %f47, %f39, %p25;
	bra.uni 	$Lt_3_15106;
$Lt_3_15874:
	.loc	22	47	0
	ld.param.f32 	%f48, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter0+4];
	mul.f32 	%f5, %f48, %f1;
	ld.param.f32 	%f49, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter0+8];
	mul.f32 	%f6, %f49, %f2;
	ld.param.f32 	%f50, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter0+12];
	mul.f32 	%f7, %f50, %f3;
	.loc	35	67	0
	bra.uni 	$Lt_3_15106;
$Lt_3_16130:
	.loc	23	48	0
	ld.param.f32 	%f51, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter0+4];
	mov.f32 	%f52, 0f00000000;    	// 0
	setp.eq.f32 	%p26, %f51, %f52;
	ld.param.f32 	%f53, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter0+8];
	mov.f32 	%f54, 0f00000000;    	// 0
	max.f32 	%f55, %f1, %f54;
	mov.f32 	%f56, 0f3f800000;    	// 1
	min.f32 	%f57, %f55, %f56;
	ld.param.f32 	%f58, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter0+12];
	sub.f32 	%f59, %f57, %f58;
	abs.f32 	%f60, %f59;
	mov.f32 	%f61, 0f00000000;    	// 0
	max.f32 	%f62, %f2, %f61;
	mov.f32 	%f63, 0f3f800000;    	// 1
	min.f32 	%f64, %f62, %f63;
	ld.param.f32 	%f65, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter0+16];
	sub.f32 	%f66, %f64, %f65;
	abs.f32 	%f67, %f66;
	mov.f32 	%f68, 0f00000000;    	// 0
	max.f32 	%f69, %f4, %f68;
	mov.f32 	%f70, 0f3f800000;    	// 1
	min.f32 	%f71, %f69, %f70;
	sub.f32 	%f72, %f71, %f4;
	abs.f32 	%f73, %f72;
	mov.f32 	%f74, 0f00000000;    	// 0
	max.f32 	%f75, %f3, %f74;
	mov.f32 	%f76, 0f3f800000;    	// 1
	min.f32 	%f77, %f75, %f76;
	ld.param.f32 	%f78, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter0+20];
	sub.f32 	%f79, %f77, %f78;
	abs.f32 	%f80, %f79;
	max.f32 	%f81, %f73, %f80;
	max.f32 	%f82, %f67, %f81;
	max.f32 	%f83, %f60, %f82;
	setp.ge.f32 	%p27, %f53, %f83;
	xor.pred 	%p28, %p26, %p27;
	@!%p28 bra 	$Lt_3_15106;
	.loc	19	267	0
	ld.const.f32 	%f84, [kRGB32f_To_601YPbPr+4];
	mul.f32 	%f29, %f84, %f2;
	ld.const.f32 	%f85, [kRGB32f_To_601YPbPr+0];
	mad.f32 	%f31, %f85, %f3, %f29;
	ld.const.f32 	%f86, [kRGB32f_To_601YPbPr+8];
	mad.f32 	%f33, %f86, %f1, %f31;
	ld.const.f32 	%f87, [k601YPbPr_To_RGB32f+12];
	mul.f32 	%f88, %f87, %f33;
	.loc	19	268	0
	ld.const.f32 	%f89, [k601YPbPr_To_RGB32f+0];
	mul.f32 	%f90, %f89, %f33;
	.loc	20	44	0
	ld.const.f32 	%f91, [k601YPbPr_To_RGB32f+24];
	mul.f32 	%f5, %f91, %f33;
	mov.f32 	%f6, %f88;
	mov.f32 	%f7, %f90;
	bra.uni 	$Lt_3_15106;
$Lt_3_16386:
	.loc	24	48	0
	ld.param.f32 	%f92, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter0+8];
	mov.f32 	%f93, 0f00000000;    	// 0
	max.f32 	%f94, %f1, %f93;
	mov.f32 	%f95, 0f3f800000;    	// 1
	min.f32 	%f96, %f94, %f95;
	ld.param.f32 	%f97, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter0+12];
	sub.f32 	%f98, %f96, %f97;
	abs.f32 	%f99, %f98;
	mov.f32 	%f100, 0f00000000;   	// 0
	max.f32 	%f101, %f2, %f100;
	mov.f32 	%f102, 0f3f800000;   	// 1
	min.f32 	%f103, %f101, %f102;
	ld.param.f32 	%f104, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter0+16];
	sub.f32 	%f105, %f103, %f104;
	abs.f32 	%f106, %f105;
	mov.f32 	%f107, 0f00000000;   	// 0
	max.f32 	%f108, %f4, %f107;
	mov.f32 	%f109, 0f3f800000;   	// 1
	min.f32 	%f110, %f108, %f109;
	sub.f32 	%f111, %f110, %f4;
	abs.f32 	%f112, %f111;
	mov.f32 	%f113, 0f00000000;   	// 0
	max.f32 	%f114, %f3, %f113;
	mov.f32 	%f115, 0f3f800000;   	// 1
	min.f32 	%f116, %f114, %f115;
	ld.param.f32 	%f117, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter0+20];
	sub.f32 	%f118, %f116, %f117;
	abs.f32 	%f119, %f118;
	max.f32 	%f120, %f112, %f119;
	max.f32 	%f121, %f106, %f120;
	max.f32 	%f122, %f99, %f121;
	setp.ge.f32 	%p29, %f92, %f122;
	@!%p29 bra 	$Lt_3_15106;
	.loc	24	51	0
	ld.param.f32 	%f123, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter0+24];
	mov.f32 	%f124, %f123;
	ld.param.f32 	%f125, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter0+28];
	mov.f32 	%f126, %f125;
	ld.param.f32 	%f127, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter0+32];
	mov.f32 	%f128, %f127;
	ld.param.f32 	%f129, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter0+4];
	mov.f32 	%f130, 0f00000000;   	// 0
	setp.eq.f32 	%p30, %f129, %f130;
	@!%p30 bra 	$Lt_3_209154;
	.loc	24	60	0
	ld.const.f32 	%f131, [kRGB32f_To_601YPbPr+4];
	mul.f32 	%f29, %f131, %f2;
	ld.const.f32 	%f132, [kRGB32f_To_601YPbPr+0];
	mad.f32 	%f31, %f132, %f3, %f29;
	ld.const.f32 	%f133, [kRGB32f_To_601YPbPr+8];
	mad.f32 	%f33, %f133, %f1, %f31;
	mul.f32 	%f124, %f33, %f123;
	.loc	24	61	0
	mul.f32 	%f126, %f33, %f125;
	.loc	24	62	0
	mul.f32 	%f128, %f33, %f127;
$Lt_3_209154:
	.loc	24	65	0
	mov.f32 	%f5, %f124;
	mov.f32 	%f6, %f126;
	mov.f32 	%f7, %f128;
	bra.uni 	$Lt_3_15106;
$Lt_3_16642:
	.loc	25	47	0
	sub.s32 	%r42, %r13, %r11;
	sub.s32 	%r43, %r14, %r5;
	cvt.rn.f32.s32 	%f134, %r5;
	cvt.rn.f32.s32 	%f135, %r11;
	cvt.rn.f32.s32 	%f136, %r42;
	cvt.rn.f32.s32 	%f137, %r43;
	ld.param.f32 	%f13, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter0+4];
	sub.f32 	%f138, %f13, %f134;
	sub.f32 	%f139, %f13, %f135;
	sub.f32 	%f140, %f13, %f136;
	sub.f32 	%f141, %f13, %f137;
	cvt.rzi.s32.f32 	%r44, %f138;
	cvt.rzi.s32.f32 	%r45, %f139;
	cvt.rzi.s32.f32 	%r46, %f140;
	cvt.rzi.s32.f32 	%r47, %f141;
	max.s32 	%r48, %r44, %r45;
	max.s32 	%r49, %r47, %r48;
	max.s32 	%r50, %r46, %r49;
	mov.u32 	%r51, 0;
	setp.le.s32 	%p31, %r50, %r51;
	@%p31 bra 	$Lt_3_15106;
	.loc	25	51	0
	cvt.rn.f32.s32 	%f142, %r50;
	sub.f32 	%f143, %f13, %f142;
	div.full.f32 	%f144, %f143, %f13;
	mul.f32 	%f8, %f4, %f144;
	bra.uni 	$Lt_3_15106;
$Lt_3_16898:
	.loc	27	50	0
	ld.const.f32 	%f145, [kRGB32f_To_601YPbPr+4];
	mul.f32 	%f29, %f145, %f2;
	ld.param.f32 	%f146, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter0+12];
	ld.param.f32 	%f43, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter0+8];
	sub.f32 	%f147, %f146, %f43;
	ld.const.f32 	%f148, [kRGB32f_To_601YPbPr+0];
	mad.f32 	%f31, %f148, %f3, %f29;
	ld.param.f32 	%f149, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter0+16];
	mul.f32 	%f150, %f149, %f147;
	ld.const.f32 	%f151, [kRGB32f_To_601YPbPr+8];
	mad.f32 	%f33, %f151, %f1, %f31;
	add.f32 	%f152, %f150, %f146;
	sub.f32 	%f153, %f43, %f150;
	mov.f32 	%f154, 0f00000000;   	// 0
	max.f32 	%f155, %f152, %f154;
	mov.f32 	%f156, 0f00000000;   	// 0
	max.f32 	%f157, %f153, %f156;
	mov.f32 	%f158, 0f3f800000;   	// 1
	min.f32 	%f159, %f155, %f158;
	mov.f32 	%f160, 0f3f800000;   	// 1
	min.f32 	%f161, %f157, %f160;
	set.gt.u32.f32 	%r52, %f161, %f33;
	neg.s32 	%r53, %r52;
	set.le.u32.f32 	%r54, %f159, %f33;
	neg.s32 	%r55, %r54;
	or.b32 	%r56, %r53, %r55;
	mov.u32 	%r57, 0;
	setp.eq.s32 	%p32, %r56, %r57;
	@%p32 bra 	$Lt_3_210434;
	mov.f32 	%f162, 0f00000000;   	// 0
	bra.uni 	$Lt_3_211202;
$Lt_3_210434:
	add.f32 	%f163, %f150, %f43;
	mov.f32 	%f164, 0f00000000;   	// 0
	max.f32 	%f165, %f163, %f164;
	mov.f32 	%f166, 0f3f800000;   	// 1
	min.f32 	%f167, %f165, %f166;
	set.le.u32.f32 	%r58, %f167, %f33;
	neg.s32 	%r59, %r58;
	sub.f32 	%f168, %f146, %f150;
	mov.f32 	%f169, 0f00000000;   	// 0
	max.f32 	%f170, %f168, %f169;
	mov.f32 	%f171, 0f3f800000;   	// 1
	min.f32 	%f172, %f170, %f171;
	set.lt.u32.f32 	%r60, %f33, %f172;
	neg.s32 	%r61, %r60;
	and.b32 	%r62, %r59, %r61;
	mov.u32 	%r63, 0;
	setp.eq.s32 	%p33, %r62, %r63;
	@%p33 bra 	$Lt_3_210946;
	mov.f32 	%f162, 0f3f800000;   	// 1
	bra.uni 	$Lt_3_211202;
$Lt_3_210946:
	add.f32 	%f173, %f150, %f150;
	setp.gt.f32 	%p34, %f167, %f33;
	@!%p34 bra 	$Lt_3_211458;
	.loc	27	62	0
	sub.f32 	%f174, %f33, %f161;
	div.full.f32 	%f162, %f174, %f173;
	bra.uni 	$Lt_3_211202;
$Lt_3_211458:
	.loc	27	66	0
	sub.f32 	%f175, %f159, %f33;
	div.full.f32 	%f162, %f175, %f173;
$Lt_3_211202:
$Lt_3_210690:
$Lt_3_210178:
	.loc	27	69	0
	mov.f32 	%f176, 0f3f800000;   	// 1
	sub.f32 	%f177, %f176, %f162;
	ld.param.f32 	%f178, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter0+4];
	mov.f32 	%f179, 0f00000000;   	// 0
	setp.neu.f32 	%p35, %f178, %f179;
	selp.f32 	%f162, %f177, %f162, %p35;
	.loc	27	77	0
	ld.const.f32 	%f180, [k601YPbPr_To_RGB32f+24];
	mul.f32 	%f5, %f180, %f162;
	ld.const.f32 	%f181, [k601YPbPr_To_RGB32f+12];
	mul.f32 	%f6, %f181, %f162;
	ld.const.f32 	%f182, [k601YPbPr_To_RGB32f+0];
	mul.f32 	%f7, %f182, %f162;
	.loc	35	79	0
	bra.uni 	$Lt_3_15106;
$Lt_3_17154:
	.loc	35	80	0
	ld.param.f32 	%f183, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter0+44];
	mov.f32 	%f184, 0f00000000;   	// 0
	setp.ge.f32 	%p36, %f183, %f184;
	@!%p36 bra 	$L_3_197378;
	cvt.rn.f32.s32 	%f185, %r5;
	cvt.rn.f32.s32 	%f186, %r14;
	mul.f32 	%f187, %f186, %f183;
	setp.lt.f32 	%p37, %f185, %f187;
	@%p37 bra 	$L_3_197122;
$L_3_197378:
	mov.f32 	%f188, 0f00000000;   	// 0
	setp.lt.f32 	%p38, %f183, %f188;
	@!%p38 bra 	$Lt_3_15106;
	cvt.rn.f32.s32 	%f189, %r11;
	cvt.rn.f32.s32 	%f190, %r13;
	mul.f32 	%f191, %f190, %f183;
	neg.f32 	%f192, %f191;
	setp.lt.f32 	%p39, %f189, %f192;
	@!%p39 bra 	$Lt_3_15106;
$L_3_197122:
	.loc	28	47	0
	ld.const.f32 	%f193, [kRGB32f_To_601YPbPr+4];
	mul.f32 	%f194, %f193, %f2;
	ld.const.f32 	%f195, [kRGB32f_To_601YPbPr+0];
	mad.f32 	%f196, %f195, %f3, %f194;
	ld.const.f32 	%f197, [kRGB32f_To_601YPbPr+8];
	mad.f32 	%f33, %f197, %f1, %f196;
	mov.f32 	%f198, %f33;
	ld.param.f32 	%f43, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter0+8];
	ld.param.f32 	%f199, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter0+4];
	sub.f32 	%f200, %f199, %f43;
	ld.param.f32 	%f201, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter0+16];
	ld.param.f32 	%f202, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter0+12];
	sub.f32 	%f203, %f202, %f201;
	mov.f32 	%f204, 0f3f800000;   	// 1
	set.neu.u32.f32 	%r64, %f203, %f204;
	neg.s32 	%r65, %r64;
	mov.f32 	%f205, 0f3f800000;   	// 1
	set.neu.u32.f32 	%r66, %f200, %f205;
	neg.s32 	%r67, %r66;
	or.b32 	%r68, %r65, %r67;
	mov.u32 	%r69, 0;
	setp.eq.s32 	%p40, %r68, %r69;
	@%p40 bra 	$Lt_3_211714;
	.loc	17	116	0
	mov.u32 	%r70, 1;
	sub.u32 	%r71, %r70, %r5;
	shr.u32 	%r72, %r11, 13;
	sub.u32 	%r73, %r5, %r11;
	sub.u32 	%r74, %r71, %r11;
	xor.b32 	%r75, %r74, %r72;
	shl.b32 	%r76, %r75, 8;
	sub.u32 	%r77, %r73, %r75;
	sub.u32 	%r78, %r11, %r75;
	xor.b32 	%r79, %r77, %r76;
	shr.u32 	%r80, %r79, 13;
	sub.u32 	%r81, %r78, %r79;
	xor.b32 	%r82, %r81, %r80;
	shr.u32 	%r83, %r82, 12;
	sub.u32 	%r84, %r75, %r79;
	sub.u32 	%r85, %r84, %r82;
	xor.b32 	%r86, %r83, %r85;
	.loc	17	117	0
	sub.u32 	%r87, %r79, %r82;
	sub.u32 	%r88, %r87, %r86;
	shl.b32 	%r89, %r86, 16;
	xor.b32 	%r90, %r88, %r89;
	.loc	17	118	0
	sub.u32 	%r91, %r82, %r86;
	sub.u32 	%r92, %r91, %r90;
	shr.u32 	%r93, %r90, 5;
	xor.b32 	%r94, %r92, %r93;
	.loc	17	119	0
	sub.u32 	%r95, %r86, %r90;
	sub.u32 	%r96, %r95, %r94;
	shr.u32 	%r97, %r94, 3;
	xor.b32 	%r98, %r96, %r97;
	.loc	17	120	0
	sub.u32 	%r99, %r90, %r94;
	sub.u32 	%r100, %r99, %r98;
	shl.b32 	%r101, %r98, 10;
	xor.b32 	%r102, %r100, %r101;
	.loc	17	121	0
	sub.u32 	%r103, %r94, %r98;
	sub.u32 	%r104, %r103, %r102;
	shr.u32 	%r105, %r102, 15;
	xor.b32 	%r106, %r104, %r105;
	.loc	28	57	0
	mov.f32 	%f206, 0f3b270d73;   	// 0.00254902
	mul.lo.u32 	%r107, %r106, 1103515245;
	add.u32 	%r108, %r107, 12345;
	shr.u32 	%r109, %r108, 16;
	and.b32 	%r110, %r109, 255;
	shl.b32 	%r111, %r110, 7;
	mul.lo.u32 	%r112, %r106, -1029531031;
	sub.u32 	%r113, %r112, 740551042;
	shr.u32 	%r114, %r113, 16;
	and.b32 	%r115, %r114, 255;
	xor.b32 	%r116, %r111, %r115;
	cvt.rn.f32.s32 	%f207, %r116;
	mov.f32 	%f208, 0f467ffe00;   	// 16383.5
	div.full.f32 	%f209, %f207, %f208;
	mov.f32 	%f210, 0fbf800000;   	// -1
	add.f32 	%f211, %f209, %f210;
	mad.f32 	%f198, %f206, %f211, %f33;
$Lt_3_211714:
	sub.f32 	%f212, %f198, %f43;
	ld.param.f32 	%f213, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter0+20];
	mov.f32 	%f214, 0f3f800000;   	// 1
	setp.neu.f32 	%p41, %f213, %f214;
	@!%p41 bra 	$Lt_3_212482;
	mov.f32 	%f215, 0f00000000;   	// 0
	setp.lt.f32 	%p42, %f212, %f215;
	@!%p42 bra 	$Lt_3_212994;
	.loc	28	66	0
	mov.f32 	%f198, %f201;
	bra.uni 	$Lt_3_212226;
$Lt_3_212994:
	.loc	28	70	0
	rcp.approx.f32 	%f216, %f200;
	mul.f32 	%f217, %f212, %f216;
	lg2.approx.f32 	%f218, %f217;
	mul.f32 	%f219, %f213, %f218;
	ex2.approx.f32 	%f220, %f219;
	mad.f32 	%f198, %f203, %f220, %f201;
	bra.uni 	$Lt_3_212226;
$Lt_3_212482:
	.loc	28	77	0
	rcp.approx.f32 	%f221, %f200;
	mul.f32 	%f222, %f212, %f221;
	mad.f32 	%f198, %f203, %f222, %f201;
$Lt_3_212226:
	.loc	19	267	0
	ld.const.f32 	%f223, [kRGB32f_To_601YPbPr+28];
	mul.f32 	%f224, %f223, %f2;
	ld.const.f32 	%f225, [kRGB32f_To_601YPbPr+16];
	mul.f32 	%f226, %f225, %f2;
	ld.param.f32 	%f227, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter0+28];
	sin.approx.f32 	%f228, %f227;
	cos.approx.f32 	%f229, %f227;
	ld.const.f32 	%f230, [kRGB32f_To_601YPbPr+24];
	mad.f32 	%f231, %f230, %f3, %f224;
	ld.const.f32 	%f232, [kRGB32f_To_601YPbPr+12];
	mad.f32 	%f233, %f232, %f3, %f226;
	ld.const.f32 	%f234, [kRGB32f_To_601YPbPr+32];
	mad.f32 	%f235, %f234, %f1, %f231;
	ld.const.f32 	%f236, [kRGB32f_To_601YPbPr+20];
	mad.f32 	%f237, %f236, %f1, %f233;
	mul.f32 	%f238, %f228, %f235;
	mul.f32 	%f239, %f229, %f235;
	mul.f32 	%f240, %f237, %f229;
	sub.f32 	%f241, %f240, %f238;
	mad.f32 	%f242, %f237, %f228, %f239;
	ld.param.f32 	%f243, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter0+32];
	sub.f32 	%f244, %f243, %f241;
	ld.param.f32 	%f245, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter0+36];
	sub.f32 	%f246, %f245, %f242;
	ld.param.f32 	%f247, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter0+40];
	mad.f32 	%f248, %f247, %f244, %f241;
	mad.f32 	%f249, %f247, %f246, %f242;
	ld.param.f32 	%f250, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter0+24];
	mul.f32 	%f251, %f248, %f250;
	mul.f32 	%f252, %f249, %f250;
	ld.const.f32 	%f253, [k601YPbPr_To_RGB32f+16];
	mul.f32 	%f254, %f253, %f251;
	ld.const.f32 	%f255, [k601YPbPr_To_RGB32f+12];
	mad.f32 	%f256, %f255, %f198, %f254;
	ld.const.f32 	%f257, [k601YPbPr_To_RGB32f+20];
	mad.f32 	%f258, %f257, %f252, %f256;
	.loc	19	268	0
	ld.const.f32 	%f259, [k601YPbPr_To_RGB32f+4];
	mul.f32 	%f260, %f259, %f251;
	ld.const.f32 	%f261, [k601YPbPr_To_RGB32f+0];
	mad.f32 	%f262, %f261, %f198, %f260;
	ld.const.f32 	%f263, [k601YPbPr_To_RGB32f+8];
	mad.f32 	%f264, %f263, %f252, %f262;
	.loc	28	92	0
	ld.const.f32 	%f265, [k601YPbPr_To_RGB32f+28];
	mul.f32 	%f266, %f265, %f251;
	ld.const.f32 	%f267, [k601YPbPr_To_RGB32f+24];
	mad.f32 	%f268, %f267, %f198, %f266;
	ld.const.f32 	%f269, [k601YPbPr_To_RGB32f+32];
	mad.f32 	%f5, %f269, %f252, %f268;
	mov.f32 	%f6, %f258;
	mov.f32 	%f7, %f264;
	bra.uni 	$Lt_3_15106;
$Lt_3_17410:
	.loc	35	83	0
	ld.param.f32 	%f270, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter0+44];
	mov.f32 	%f271, 0f00000000;   	// 0
	setp.ge.f32 	%p43, %f270, %f271;
	@!%p43 bra 	$L_3_198146;
	cvt.rn.f32.s32 	%f272, %r5;
	cvt.rn.f32.s32 	%f273, %r14;
	mul.f32 	%f274, %f273, %f270;
	setp.lt.f32 	%p44, %f272, %f274;
	@%p44 bra 	$L_3_197890;
$L_3_198146:
	mov.f32 	%f275, 0f00000000;   	// 0
	setp.lt.f32 	%p45, %f270, %f275;
	@!%p45 bra 	$Lt_3_15106;
	cvt.rn.f32.s32 	%f276, %r11;
	cvt.rn.f32.s32 	%f277, %r13;
	mul.f32 	%f278, %f277, %f270;
	neg.f32 	%f279, %f278;
	setp.lt.f32 	%p46, %f276, %f279;
	@!%p46 bra 	$Lt_3_15106;
$L_3_197890:
	.loc	28	110	0
	ld.const.f32 	%f280, [kRGB32f_To_601YPbPr+4];
	mul.f32 	%f29, %f280, %f2;
	ld.param.f32 	%f43, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter0+8];
	ld.param.f32 	%f281, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter0+4];
	sub.f32 	%f282, %f281, %f43;
	ld.param.f32 	%f283, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter0+16];
	ld.param.f32 	%f284, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter0+12];
	sub.f32 	%f285, %f284, %f283;
	ld.const.f32 	%f286, [kRGB32f_To_601YPbPr+0];
	mad.f32 	%f31, %f286, %f3, %f29;
	ld.const.f32 	%f287, [kRGB32f_To_601YPbPr+8];
	mad.f32 	%f33, %f287, %f1, %f31;
	sub.f32 	%f288, %f33, %f43;
	ld.param.f32 	%f289, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter0+20];
	mov.f32 	%f290, 0f3f800000;   	// 1
	setp.neu.f32 	%p47, %f289, %f290;
	@!%p47 bra 	$Lt_3_213506;
	.loc	28	120	0
	mov.f32 	%f291, 0f00000000;   	// 0
	max.f32 	%f292, %f288, %f291;
	div.full.f32 	%f293, %f292, %f282;
	lg2.approx.f32 	%f294, %f293;
	mul.f32 	%f295, %f289, %f294;
	ex2.approx.f32 	%f296, %f295;
	mad.f32 	%f297, %f285, %f296, %f283;
	bra.uni 	$Lt_3_213250;
$Lt_3_213506:
	.loc	28	129	0
	div.full.f32 	%f298, %f288, %f282;
	mad.f32 	%f297, %f285, %f298, %f283;
$Lt_3_213250:
	.loc	28	135	0
	ld.const.f32 	%f299, [k601YPbPr_To_RGB32f+24];
	mul.f32 	%f5, %f299, %f297;
	ld.const.f32 	%f300, [k601YPbPr_To_RGB32f+12];
	mul.f32 	%f6, %f300, %f297;
	ld.const.f32 	%f301, [k601YPbPr_To_RGB32f+0];
	mul.f32 	%f7, %f301, %f297;
	bra.uni 	$Lt_3_15106;
$Lt_3_17666:
	.loc	35	86	0
	ld.param.f32 	%f302, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter0+44];
	mov.f32 	%f303, 0f00000000;   	// 0
	setp.ge.f32 	%p48, %f302, %f303;
	@!%p48 bra 	$L_3_198914;
	cvt.rn.f32.s32 	%f304, %r5;
	cvt.rn.f32.s32 	%f305, %r14;
	mul.f32 	%f306, %f305, %f302;
	setp.lt.f32 	%p49, %f304, %f306;
	@%p49 bra 	$L_3_198658;
$L_3_198914:
	mov.f32 	%f307, 0f00000000;   	// 0
	setp.lt.f32 	%p50, %f302, %f307;
	@!%p50 bra 	$Lt_3_15106;
	cvt.rn.f32.s32 	%f308, %r11;
	cvt.rn.f32.s32 	%f309, %r13;
	mul.f32 	%f310, %f309, %f302;
	neg.f32 	%f311, %f310;
	setp.lt.f32 	%p51, %f308, %f311;
	@!%p51 bra 	$Lt_3_15106;
$L_3_198658:
	.loc	28	160	0
	ld.const.f32 	%f5, [k601YPbPr_To_RGB32f+24];
	ld.const.f32 	%f6, [k601YPbPr_To_RGB32f+12];
	ld.const.f32 	%f7, [k601YPbPr_To_RGB32f+0];
	bra.uni 	$Lt_3_15106;
$Lt_3_17922:
	.loc	29	42	0
	ld.param.f32 	%f13, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter0+4];
	mov.f32 	%f312, 0f00000000;   	// 0
	setp.ge.f32 	%p52, %f1, %f312;
	@!%p52 bra 	$Lt_3_214018;
	.loc	19	292	0
	lg2.approx.f32 	%f313, %f1;
	mul.f32 	%f314, %f13, %f313;
	ex2.approx.f32 	%f315, %f314;
	bra.uni 	$Lt_3_213762;
$Lt_3_214018:
	neg.f32 	%f316, %f1;
	lg2.approx.f32 	%f317, %f316;
	mul.f32 	%f318, %f13, %f317;
	ex2.approx.f32 	%f319, %f318;
	neg.f32 	%f315, %f319;
$Lt_3_213762:
	mov.f32 	%f320, 0f00000000;   	// 0
	setp.ge.f32 	%p53, %f2, %f320;
	@!%p53 bra 	$Lt_3_214530;
	.loc	19	293	0
	lg2.approx.f32 	%f321, %f2;
	mul.f32 	%f322, %f13, %f321;
	ex2.approx.f32 	%f323, %f322;
	bra.uni 	$Lt_3_214274;
$Lt_3_214530:
	neg.f32 	%f324, %f2;
	lg2.approx.f32 	%f325, %f324;
	mul.f32 	%f326, %f13, %f325;
	ex2.approx.f32 	%f327, %f326;
	neg.f32 	%f323, %f327;
$Lt_3_214274:
	mov.f32 	%f328, 0f00000000;   	// 0
	setp.ge.f32 	%p54, %f3, %f328;
	@!%p54 bra 	$Lt_3_215042;
	.loc	19	294	0
	lg2.approx.f32 	%f329, %f3;
	mul.f32 	%f330, %f13, %f329;
	ex2.approx.f32 	%f331, %f330;
	bra.uni 	$Lt_3_214786;
$Lt_3_215042:
	neg.f32 	%f332, %f3;
	lg2.approx.f32 	%f333, %f332;
	mul.f32 	%f334, %f13, %f333;
	ex2.approx.f32 	%f335, %f334;
	neg.f32 	%f331, %f335;
$Lt_3_214786:
	mov.f32 	%f336, 0f00000000;   	// 0
	setp.ge.f32 	%p55, %f4, %f336;
	@!%p55 bra 	$Lt_3_215554;
	.loc	19	295	0
	lg2.approx.f32 	%f337, %f4;
	mul.f32 	%f338, %f13, %f337;
	ex2.approx.f32 	%f339, %f338;
	bra.uni 	$Lt_3_215298;
$Lt_3_215554:
	neg.f32 	%f340, %f4;
	lg2.approx.f32 	%f341, %f340;
	mul.f32 	%f342, %f13, %f341;
	ex2.approx.f32 	%f343, %f342;
	neg.f32 	%f339, %f343;
$Lt_3_215298:
	.loc	29	42	0
	mov.f32 	%f5, %f315;
	mov.f32 	%f6, %f323;
	mov.f32 	%f7, %f331;
	mov.f32 	%f8, %f339;
	.loc	35	91	0
	bra.uni 	$Lt_3_15106;
$Lt_3_18178:
	.loc	30	41	0
	sub.s32 	%r117, %r14, %r5;
	sub.s32 	%r6, %r117, 1;
	.loc	35	94	0
	bra.uni 	$Lt_3_15106;
$Lt_3_18434:
	.loc	35	96	0
	sub.u32 	%r73, %r5, %r11;
	ld.param.f32 	%f13, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter0+4];
	mov.f32 	%f344, 0f3e800000;   	// 0.25
	mul.f32 	%f345, %f13, %f344;
	ld.param.f32 	%f346, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter0+16];
	ld.param.f32 	%f347, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter0+8];
	mov.f32 	%f348, 0f3f800000;   	// 1
	setp.eq.f32 	%p56, %f347, %f348;
	@!%p56 bra 	$Lt_3_216066;
	.loc	17	117	0
	cvt.rzi.u32.f32 	%r118, %f346;
	shr.u32 	%r119, %r118, 13;
	sub.u32 	%r120, %r73, %r118;
	sub.u32 	%r121, %r11, %r118;
	xor.b32 	%r122, %r120, %r119;
	shl.b32 	%r123, %r122, 8;
	sub.u32 	%r124, %r121, %r122;
	sub.u32 	%r125, %r118, %r122;
	xor.b32 	%r126, %r124, %r123;
	shr.u32 	%r127, %r126, 13;
	sub.u32 	%r128, %r125, %r126;
	sub.u32 	%r129, %r122, %r126;
	xor.b32 	%r130, %r128, %r127;
	shr.u32 	%r131, %r130, 12;
	sub.u32 	%r132, %r129, %r130;
	xor.b32 	%r133, %r132, %r131;
	shl.b32 	%r134, %r133, 16;
	sub.u32 	%r135, %r126, %r130;
	sub.u32 	%r136, %r135, %r133;
	xor.b32 	%r137, %r134, %r136;
	.loc	17	118	0
	sub.u32 	%r138, %r130, %r133;
	sub.u32 	%r139, %r138, %r137;
	shr.u32 	%r140, %r137, 5;
	xor.b32 	%r141, %r139, %r140;
	.loc	17	119	0
	sub.u32 	%r142, %r133, %r137;
	sub.u32 	%r143, %r142, %r141;
	shr.u32 	%r144, %r141, 3;
	xor.b32 	%r145, %r143, %r144;
	.loc	17	120	0
	sub.u32 	%r146, %r137, %r141;
	sub.u32 	%r147, %r146, %r145;
	shl.b32 	%r148, %r145, 10;
	xor.b32 	%r149, %r147, %r148;
	.loc	17	121	0
	sub.u32 	%r150, %r141, %r145;
	sub.u32 	%r151, %r150, %r149;
	shr.u32 	%r152, %r149, 15;
	xor.b32 	%r153, %r151, %r152;
	.loc	31	48	0
	mul.lo.u32 	%r154, %r153, 1103515245;
	add.u32 	%r155, %r154, 12345;
	shr.u32 	%r156, %r155, 16;
	and.b32 	%r157, %r156, 255;
	shl.b32 	%r158, %r157, 7;
	mul.lo.u32 	%r159, %r153, -1029531031;
	sub.u32 	%r160, %r159, 740551042;
	shr.u32 	%r161, %r160, 16;
	and.b32 	%r162, %r161, 255;
	xor.b32 	%r163, %r158, %r162;
	cvt.rn.f32.s32 	%f349, %r163;
	mov.f32 	%f350, 0f46fffe00;   	// 32767
	div.full.f32 	%f351, %f349, %f350;
	mul.f32 	%f352, %f13, %f351;
	mov.f32 	%f353, 0f3f000000;   	// 0.5
	mul.f32 	%f354, %f352, %f353;
	sub.f32 	%f355, %f354, %f345;
	.loc	17	117	0
	add.u32 	%r164, %r120, 1;
	xor.b32 	%r165, %r164, %r119;
	shl.b32 	%r166, %r165, 8;
	sub.u32 	%r167, %r121, %r165;
	sub.u32 	%r168, %r118, %r165;
	xor.b32 	%r169, %r166, %r167;
	shr.u32 	%r170, %r169, 13;
	sub.u32 	%r171, %r168, %r169;
	sub.u32 	%r172, %r165, %r169;
	xor.b32 	%r173, %r170, %r171;
	shr.u32 	%r174, %r173, 12;
	sub.u32 	%r175, %r172, %r173;
	xor.b32 	%r176, %r174, %r175;
	sub.u32 	%r177, %r169, %r173;
	sub.u32 	%r178, %r177, %r176;
	shl.b32 	%r179, %r176, 16;
	xor.b32 	%r178, %r178, %r179;
	.loc	17	118	0
	sub.u32 	%r180, %r173, %r176;
	sub.u32 	%r180, %r180, %r178;
	shr.u32 	%r181, %r178, 5;
	xor.b32 	%r180, %r180, %r181;
	.loc	17	119	0
	sub.u32 	%r182, %r176, %r178;
	sub.u32 	%r182, %r182, %r180;
	shr.u32 	%r183, %r180, 3;
	xor.b32 	%r182, %r182, %r183;
	.loc	17	120	0
	sub.u32 	%r178, %r178, %r180;
	sub.u32 	%r178, %r178, %r182;
	shl.b32 	%r184, %r182, 10;
	xor.b32 	%r178, %r178, %r184;
	.loc	17	121	0
	sub.u32 	%r180, %r180, %r182;
	sub.u32 	%r180, %r180, %r178;
	shr.u32 	%r185, %r178, 15;
	xor.b32 	%r180, %r180, %r185;
	.loc	31	49	0
	mul.lo.u32 	%r186, %r180, 1103515245;
	add.u32 	%r187, %r186, 12345;
	shr.u32 	%r188, %r187, 16;
	and.b32 	%r189, %r188, 255;
	shl.b32 	%r190, %r189, 7;
	mul.lo.u32 	%r191, %r180, -1029531031;
	sub.u32 	%r192, %r191, 740551042;
	shr.u32 	%r193, %r192, 16;
	and.b32 	%r194, %r193, 255;
	xor.b32 	%r195, %r190, %r194;
	cvt.rn.f32.s32 	%f356, %r195;
	mov.f32 	%f357, 0f46fffe00;   	// 32767
	div.full.f32 	%f358, %f356, %f357;
	mul.f32 	%f359, %f13, %f358;
	mov.f32 	%f360, 0f3f000000;   	// 0.5
	mul.f32 	%f361, %f359, %f360;
	sub.f32 	%f362, %f361, %f345;
	.loc	17	117	0
	add.u32 	%r196, %r120, 2;
	xor.b32 	%r197, %r196, %r119;
	shl.b32 	%r198, %r197, 8;
	sub.u32 	%r199, %r121, %r197;
	sub.u32 	%r200, %r118, %r197;
	xor.b32 	%r201, %r198, %r199;
	shr.u32 	%r202, %r201, 13;
	sub.u32 	%r203, %r200, %r201;
	sub.u32 	%r204, %r197, %r201;
	xor.b32 	%r205, %r202, %r203;
	shr.u32 	%r206, %r205, 12;
	sub.u32 	%r207, %r204, %r205;
	xor.b32 	%r208, %r206, %r207;
	sub.u32 	%r209, %r201, %r205;
	sub.u32 	%r210, %r209, %r208;
	shl.b32 	%r211, %r208, 16;
	xor.b32 	%r210, %r210, %r211;
	.loc	17	118	0
	sub.u32 	%r212, %r205, %r208;
	sub.u32 	%r212, %r212, %r210;
	shr.u32 	%r213, %r210, 5;
	xor.b32 	%r212, %r212, %r213;
	.loc	17	119	0
	sub.u32 	%r214, %r208, %r210;
	sub.u32 	%r214, %r214, %r212;
	shr.u32 	%r215, %r212, 3;
	xor.b32 	%r214, %r214, %r215;
	.loc	17	120	0
	sub.u32 	%r210, %r210, %r212;
	sub.u32 	%r210, %r210, %r214;
	shl.b32 	%r216, %r214, 10;
	xor.b32 	%r210, %r210, %r216;
	.loc	17	121	0
	sub.u32 	%r212, %r212, %r214;
	sub.u32 	%r212, %r212, %r210;
	shr.u32 	%r217, %r210, 15;
	xor.b32 	%r212, %r212, %r217;
	.loc	31	50	0
	mul.lo.u32 	%r218, %r212, 1103515245;
	add.u32 	%r219, %r218, 12345;
	shr.u32 	%r220, %r219, 16;
	and.b32 	%r221, %r220, 255;
	shl.b32 	%r222, %r221, 7;
	mul.lo.u32 	%r223, %r212, -1029531031;
	sub.u32 	%r224, %r223, 740551042;
	shr.u32 	%r225, %r224, 16;
	and.b32 	%r226, %r225, 255;
	xor.b32 	%r227, %r222, %r226;
	cvt.rn.f32.s32 	%f363, %r227;
	mov.f32 	%f364, 0f46fffe00;   	// 32767
	div.full.f32 	%f365, %f363, %f364;
	mul.f32 	%f366, %f13, %f365;
	mov.f32 	%f367, 0f3f000000;   	// 0.5
	mul.f32 	%f368, %f366, %f367;
	sub.f32 	%f369, %f368, %f345;
	bra.uni 	$Lt_3_215810;
$Lt_3_216066:
	.loc	17	117	0
	cvt.rzi.u32.f32 	%r228, %f346;
	shr.u32 	%r229, %r228, 13;
	sub.u32 	%r230, %r73, %r228;
	sub.u32 	%r231, %r11, %r228;
	xor.b32 	%r232, %r230, %r229;
	shl.b32 	%r233, %r232, 8;
	sub.u32 	%r234, %r231, %r232;
	sub.u32 	%r235, %r228, %r232;
	xor.b32 	%r236, %r234, %r233;
	shr.u32 	%r237, %r236, 13;
	sub.u32 	%r238, %r235, %r236;
	sub.u32 	%r239, %r232, %r236;
	xor.b32 	%r240, %r238, %r237;
	shr.u32 	%r241, %r240, 12;
	sub.u32 	%r242, %r239, %r240;
	xor.b32 	%r243, %r242, %r241;
	shl.b32 	%r244, %r243, 16;
	sub.u32 	%r245, %r236, %r240;
	sub.u32 	%r246, %r245, %r243;
	xor.b32 	%r247, %r244, %r246;
	.loc	17	118	0
	sub.u32 	%r248, %r240, %r243;
	sub.u32 	%r249, %r248, %r247;
	shr.u32 	%r250, %r247, 5;
	xor.b32 	%r251, %r249, %r250;
	.loc	17	119	0
	sub.u32 	%r252, %r243, %r247;
	sub.u32 	%r253, %r252, %r251;
	shr.u32 	%r254, %r251, 3;
	xor.b32 	%r255, %r253, %r254;
	.loc	17	120	0
	sub.u32 	%r256, %r247, %r251;
	sub.u32 	%r257, %r256, %r255;
	shl.b32 	%r258, %r255, 10;
	xor.b32 	%r259, %r257, %r258;
	.loc	17	121	0
	sub.u32 	%r260, %r251, %r255;
	sub.u32 	%r261, %r260, %r259;
	shr.u32 	%r262, %r259, 15;
	xor.b32 	%r263, %r261, %r262;
	.loc	31	54	0
	mul.lo.u32 	%r264, %r263, 1103515245;
	mul.lo.u32 	%r265, %r263, -1029531031;
	add.u32 	%r266, %r264, 12345;
	sub.u32 	%r267, %r265, 740551042;
	shr.u32 	%r268, %r266, 16;
	shr.u32 	%r269, %r267, 16;
	and.b32 	%r270, %r268, 255;
	and.b32 	%r271, %r269, 255;
	shl.b32 	%r272, %r270, 7;
	xor.b32 	%r273, %r272, %r271;
	cvt.rn.f32.s32 	%f370, %r273;
	mov.f32 	%f371, 0f46fffe00;   	// 32767
	div.full.f32 	%f372, %f370, %f371;
	mul.f32 	%f373, %f372, %f13;
	mov.f32 	%f374, 0f3f000000;   	// 0.5
	mul.f32 	%f375, %f373, %f374;
	sub.f32 	%f355, %f375, %f345;
	mov.f32 	%f369, %f355;
	mov.f32 	%f362, %f355;
$Lt_3_215810:
	.loc	31	57	0
	add.f32 	%f7, %f355, %f3;
	.loc	31	58	0
	add.f32 	%f6, %f2, %f362;
	.loc	31	59	0
	add.f32 	%f5, %f1, %f369;
	ld.param.f32 	%f376, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter0+12];
	mov.f32 	%f377, 0f3f800000;   	// 1
	setp.eq.f32 	%p57, %f376, %f377;
	@!%p57 bra 	$Lt_3_15106;
	.loc	31	63	0
	mov.f32 	%f378, 0f00000000;   	// 0
	max.f32 	%f379, %f5, %f378;
	mov.f32 	%f380, 0f3f800000;   	// 1
	min.f32 	%f5, %f379, %f380;
	mov.f32 	%f381, 0f00000000;   	// 0
	max.f32 	%f382, %f6, %f381;
	mov.f32 	%f383, 0f3f800000;   	// 1
	min.f32 	%f6, %f382, %f383;
	mov.f32 	%f384, 0f00000000;   	// 0
	max.f32 	%f385, %f7, %f384;
	mov.f32 	%f386, 0f3f800000;   	// 1
	min.f32 	%f7, %f385, %f386;
	mov.f32 	%f387, 0f00000000;   	// 0
	max.f32 	%f388, %f4, %f387;
	mov.f32 	%f389, 0f3f800000;   	// 1
	min.f32 	%f8, %f388, %f389;
	bra.uni 	$Lt_3_15106;
$Lt_3_18690:
	.loc	35	99	0
	cvt.rn.f32.s32 	%f390, %r5;
	cvt.rn.f32.s32 	%f391, %r14;
	ld.param.f32 	%f392, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter0+20];
	mul.f32 	%f393, %f391, %f392;
	setp.lt.f32 	%p58, %f390, %f393;
	@!%p58 bra 	$Lt_3_15106;
	.loc	19	267	0
	ld.const.f32 	%f394, [kRGB32f_To_601YPbPr+4];
	mul.f32 	%f29, %f394, %f2;
	ld.const.f32 	%f395, [kRGB32f_To_601YPbPr+28];
	mul.f32 	%f396, %f395, %f2;
	ld.const.f32 	%f397, [kRGB32f_To_601YPbPr+16];
	mul.f32 	%f398, %f397, %f2;
	ld.const.f32 	%f399, [kRGB32f_To_601YPbPr+0];
	mad.f32 	%f31, %f399, %f3, %f29;
	ld.const.f32 	%f400, [kRGB32f_To_601YPbPr+24];
	mad.f32 	%f401, %f400, %f3, %f396;
	ld.const.f32 	%f402, [kRGB32f_To_601YPbPr+12];
	mad.f32 	%f403, %f402, %f3, %f398;
	ld.const.f32 	%f404, [kRGB32f_To_601YPbPr+8];
	mad.f32 	%f33, %f404, %f1, %f31;
	ld.const.f32 	%f405, [kRGB32f_To_601YPbPr+32];
	mad.f32 	%f406, %f405, %f1, %f401;
	ld.const.f32 	%f407, [kRGB32f_To_601YPbPr+20];
	mad.f32 	%f408, %f407, %f1, %f403;
	ld.param.f32 	%f409, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter0+4];
	ld.param.f32 	%f410, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter0+8];
	mad.f32 	%f411, %f33, %f410, %f409;
	ld.param.f32 	%f412, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter0+16];
	mul.f32 	%f413, %f406, %f412;
	ld.param.f32 	%f414, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter0+12];
	mul.f32 	%f415, %f406, %f414;
	mul.f32 	%f416, %f408, %f414;
	sub.f32 	%f417, %f416, %f413;
	mad.f32 	%f418, %f408, %f412, %f415;
	ld.const.f32 	%f419, [k601YPbPr_To_RGB32f+16];
	mul.f32 	%f420, %f419, %f417;
	ld.const.f32 	%f421, [k601YPbPr_To_RGB32f+12];
	mad.f32 	%f422, %f421, %f411, %f420;
	ld.const.f32 	%f423, [k601YPbPr_To_RGB32f+20];
	mad.f32 	%f424, %f423, %f418, %f422;
	.loc	19	268	0
	ld.const.f32 	%f425, [k601YPbPr_To_RGB32f+4];
	mul.f32 	%f426, %f425, %f417;
	ld.const.f32 	%f427, [k601YPbPr_To_RGB32f+0];
	mad.f32 	%f428, %f427, %f411, %f426;
	ld.const.f32 	%f429, [k601YPbPr_To_RGB32f+8];
	mad.f32 	%f430, %f429, %f418, %f428;
	.loc	32	56	0
	ld.const.f32 	%f431, [k601YPbPr_To_RGB32f+28];
	mul.f32 	%f432, %f431, %f417;
	ld.const.f32 	%f433, [k601YPbPr_To_RGB32f+24];
	mad.f32 	%f434, %f433, %f411, %f432;
	ld.const.f32 	%f435, [k601YPbPr_To_RGB32f+32];
	mad.f32 	%f5, %f435, %f418, %f434;
	mov.f32 	%f6, %f424;
	mov.f32 	%f7, %f430;
	bra.uni 	$Lt_3_15106;
$Lt_3_18946:
	.loc	33	46	0
	ld.const.f32 	%f436, [kRGB32f_To_601YPbPr+4];
	mul.f32 	%f29, %f436, %f2;
	ld.const.f32 	%f437, [kRGB32f_To_601YPbPr+0];
	mad.f32 	%f31, %f437, %f3, %f29;
	ld.const.f32 	%f438, [kRGB32f_To_601YPbPr+8];
	mad.f32 	%f33, %f438, %f1, %f31;
	ld.param.f32 	%f43, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter0+8];
	ld.param.f32 	%f439, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter0+20];
	sub.f32 	%f440, %f439, %f43;
	mad.f32 	%f441, %f33, %f440, %f43;
	.loc	33	47	0
	ld.param.f32 	%f442, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter0+12];
	ld.param.f32 	%f443, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter0+24];
	sub.f32 	%f444, %f443, %f442;
	mad.f32 	%f445, %f33, %f444, %f442;
	.loc	33	49	0
	ld.param.f32 	%f446, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter0+28];
	ld.param.f32 	%f13, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter0+4];
	ld.param.f32 	%f447, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter0+16];
	sub.f32 	%f448, %f447, %f13;
	mad.f32 	%f449, %f33, %f448, %f13;
	sub.f32 	%f450, %f449, %f1;
	mad.f32 	%f5, %f446, %f450, %f1;
	.loc	33	50	0
	sub.f32 	%f451, %f441, %f2;
	mad.f32 	%f6, %f446, %f451, %f2;
	.loc	33	51	0
	sub.f32 	%f452, %f445, %f3;
	mad.f32 	%f7, %f446, %f452, %f3;
	.loc	35	103	0
	bra.uni 	$Lt_3_15106;
$Lt_3_19202:
	.loc	34	41	0
	sub.s32 	%r274, %r13, %r11;
	sub.s32 	%r12, %r274, 1;
$Lt_3_15106:
	.loc	35	54	0
	ld.param.u32 	%r275, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter1+0];
	mov.u32 	%r276, 0;
	setp.eq.s32 	%p59, %r275, %r276;
	@%p59 bra 	$Lt_3_9986;
	mov.u32 	%r277, 1;
	setp.eq.s32 	%p60, %r275, %r277;
	@%p60 bra 	$Lt_3_10498;
	mov.u32 	%r278, 2;
	setp.eq.s32 	%p61, %r275, %r278;
	@%p61 bra 	$Lt_3_10754;
	mov.u32 	%r279, 3;
	setp.eq.s32 	%p62, %r275, %r279;
	@%p62 bra 	$Lt_3_11010;
	mov.u32 	%r280, 4;
	setp.eq.s32 	%p63, %r275, %r280;
	@%p63 bra 	$Lt_3_11266;
	mov.u32 	%r281, 5;
	setp.eq.s32 	%p64, %r275, %r281;
	@%p64 bra 	$Lt_3_11522;
	mov.u32 	%r282, 6;
	setp.eq.s32 	%p65, %r275, %r282;
	@%p65 bra 	$Lt_3_11778;
	mov.u32 	%r283, 7;
	setp.eq.s32 	%p66, %r275, %r283;
	@%p66 bra 	$Lt_3_12034;
	mov.u32 	%r284, 8;
	setp.eq.s32 	%p67, %r275, %r284;
	@%p67 bra 	$Lt_3_12290;
	mov.u32 	%r285, 9;
	setp.eq.s32 	%p68, %r275, %r285;
	@%p68 bra 	$Lt_3_12546;
	mov.u32 	%r286, 10;
	setp.eq.s32 	%p69, %r275, %r286;
	@%p69 bra 	$Lt_3_12802;
	mov.u32 	%r287, 11;
	setp.eq.s32 	%p70, %r275, %r287;
	@%p70 bra 	$Lt_3_13058;
	mov.u32 	%r288, 12;
	setp.eq.s32 	%p71, %r275, %r288;
	@%p71 bra 	$Lt_3_13314;
	mov.u32 	%r289, 13;
	setp.eq.s32 	%p72, %r275, %r289;
	@%p72 bra 	$Lt_3_13570;
	mov.u32 	%r290, 14;
	setp.eq.s32 	%p73, %r275, %r290;
	@%p73 bra 	$Lt_3_13826;
	mov.u32 	%r291, 15;
	setp.eq.s32 	%p74, %r275, %r291;
	@%p74 bra 	$Lt_3_14082;
	mov.u32 	%r292, 16;
	setp.eq.s32 	%p75, %r275, %r292;
	@%p75 bra 	$Lt_3_14338;
	bra.uni 	$Lt_3_10242;
$Lt_3_9986:
	.loc	18	42	0
	cvt.sat.f32.f32 	%f8, %f8;
	ld.param.f32 	%f453, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter1+8];
	mov.f32 	%f454, 0f3f800000;   	// 1
	setp.eq.f32 	%p76, %f453, %f454;
	ld.param.f32 	%f455, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter1+12];
	mov.f32 	%f456, 0f3f800000;   	// 1
	setp.eq.f32 	%p77, %f455, %f456;
	ld.param.f32 	%f457, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter1+4];
	ld.param.f32 	%f458, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter1+16];
	mov.f32 	%f459, 0f3f800000;   	// 1
	setp.eq.f32 	%p78, %f458, %f459;
	@!%p78 bra 	$Lt_3_217602;
	.loc	18	45	0
	mul.f32 	%f460, %f8, %f457;
	selp.f32 	%f8, %f457, %f460, %p76;
	.loc	18	53	0
	mov.f32 	%f461, 0f3f800000;   	// 1
	sub.f32 	%f462, %f461, %f8;
	selp.f32 	%f8, %f462, %f8, %p77;
	.loc	18	57	0
	mov.f32 	%f5, %f8;
	mov.f32 	%f6, %f8;
	mov.f32 	%f7, %f8;
	bra.uni 	$Lt_3_10242;
$Lt_3_217602:
	@!%p76 bra 	$Lt_3_218114;
	.loc	18	61	0
	mov.f32 	%f463, 0f3f800000;   	// 1
	sub.f32 	%f464, %f463, %f457;
	selp.f32 	%f8, %f464, %f457, %p77;
	bra.uni 	$Lt_3_10242;
$Lt_3_218114:
	.loc	18	69	0
	mov.f32 	%f465, 0f3f800000;   	// 1
	sub.f32 	%f466, %f465, %f8;
	selp.f32 	%f8, %f466, %f8, %p77;
	.loc	18	73	0
	mul.f32 	%f8, %f8, %f457;
	bra.uni 	$Lt_3_10242;
$Lt_3_10498:
	.loc	19	267	0
	ld.const.f32 	%f467, [kRGB32f_To_601YPbPr+4];
	mul.f32 	%f468, %f467, %f6;
	ld.const.f32 	%f469, [kRGB32f_To_601YPbPr+0];
	mad.f32 	%f470, %f469, %f7, %f468;
	ld.const.f32 	%f471, [kRGB32f_To_601YPbPr+8];
	mad.f32 	%f472, %f471, %f5, %f470;
	ld.const.f32 	%f473, [k601YPbPr_To_RGB32f+12];
	mul.f32 	%f474, %f473, %f472;
	.loc	19	268	0
	ld.const.f32 	%f475, [k601YPbPr_To_RGB32f+0];
	mul.f32 	%f476, %f475, %f472;
	.loc	20	44	0
	ld.const.f32 	%f477, [k601YPbPr_To_RGB32f+24];
	mul.f32 	%f5, %f477, %f472;
	mov.f32 	%f6, %f474;
	mov.f32 	%f7, %f476;
	.loc	35	61	0
	bra.uni 	$Lt_3_10242;
$Lt_3_10754:
	.loc	35	63	0
	ld.param.f32 	%f478, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter1+20];
	ld.param.f32 	%f479, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter1+12];
	ld.param.f32 	%f480, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter1+4];
	mov.f32 	%f481, 0f00000000;   	// 0
	setp.neu.f32 	%p79, %f480, %f481;
	@!%p79 bra 	$Lt_3_218626;
	.loc	21	44	0
	ld.param.f32 	%f482, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter1+8];
	sub.f32 	%f483, %f7, %f479;
	mad.f32 	%f7, %f482, %f483, %f478;
	.loc	21	45	0
	sub.f32 	%f484, %f6, %f479;
	mad.f32 	%f6, %f482, %f484, %f478;
	.loc	21	46	0
	sub.f32 	%f485, %f5, %f479;
	mad.f32 	%f5, %f482, %f485, %f478;
	bra.uni 	$Lt_3_10242;
$Lt_3_218626:
	.loc	21	50	0
	ld.param.f32 	%f486, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter1+24];
	setp.gt.f32 	%p80, %f7, %f479;
	selp.f32 	%f7, %f486, %f478, %p80;
	.loc	21	51	0
	setp.gt.f32 	%p81, %f6, %f479;
	selp.f32 	%f6, %f486, %f478, %p81;
	.loc	21	52	0
	setp.gt.f32 	%p82, %f5, %f479;
	selp.f32 	%f5, %f486, %f478, %p82;
	bra.uni 	$Lt_3_10242;
$Lt_3_11010:
	.loc	22	47	0
	ld.param.f32 	%f487, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter1+4];
	mul.f32 	%f5, %f487, %f5;
	ld.param.f32 	%f488, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter1+8];
	mul.f32 	%f6, %f488, %f6;
	ld.param.f32 	%f489, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter1+12];
	mul.f32 	%f7, %f489, %f7;
	.loc	35	67	0
	bra.uni 	$Lt_3_10242;
$Lt_3_11266:
	.loc	23	48	0
	ld.param.f32 	%f490, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter1+4];
	mov.f32 	%f491, 0f00000000;   	// 0
	setp.eq.f32 	%p83, %f490, %f491;
	ld.param.f32 	%f492, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter1+8];
	mov.f32 	%f493, 0f00000000;   	// 0
	max.f32 	%f494, %f5, %f493;
	mov.f32 	%f495, 0f3f800000;   	// 1
	min.f32 	%f496, %f494, %f495;
	ld.param.f32 	%f497, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter1+12];
	sub.f32 	%f498, %f496, %f497;
	abs.f32 	%f499, %f498;
	mov.f32 	%f500, 0f00000000;   	// 0
	max.f32 	%f501, %f6, %f500;
	mov.f32 	%f502, 0f3f800000;   	// 1
	min.f32 	%f503, %f501, %f502;
	ld.param.f32 	%f504, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter1+16];
	sub.f32 	%f505, %f503, %f504;
	abs.f32 	%f506, %f505;
	mov.f32 	%f507, 0f00000000;   	// 0
	max.f32 	%f508, %f7, %f507;
	mov.f32 	%f509, 0f3f800000;   	// 1
	min.f32 	%f510, %f508, %f509;
	ld.param.f32 	%f511, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter1+20];
	sub.f32 	%f512, %f510, %f511;
	abs.f32 	%f513, %f512;
	mov.f32 	%f514, 0f00000000;   	// 0
	max.f32 	%f515, %f8, %f514;
	mov.f32 	%f516, 0f3f800000;   	// 1
	min.f32 	%f517, %f515, %f516;
	sub.f32 	%f518, %f517, %f8;
	abs.f32 	%f519, %f518;
	max.f32 	%f520, %f513, %f519;
	max.f32 	%f521, %f506, %f520;
	max.f32 	%f522, %f499, %f521;
	setp.ge.f32 	%p84, %f492, %f522;
	xor.pred 	%p85, %p83, %p84;
	@!%p85 bra 	$Lt_3_10242;
	.loc	19	267	0
	ld.const.f32 	%f523, [kRGB32f_To_601YPbPr+4];
	mul.f32 	%f468, %f523, %f6;
	ld.const.f32 	%f524, [kRGB32f_To_601YPbPr+0];
	mad.f32 	%f470, %f524, %f7, %f468;
	ld.const.f32 	%f525, [kRGB32f_To_601YPbPr+8];
	mad.f32 	%f472, %f525, %f5, %f470;
	ld.const.f32 	%f526, [k601YPbPr_To_RGB32f+12];
	mul.f32 	%f527, %f526, %f472;
	.loc	19	268	0
	ld.const.f32 	%f528, [k601YPbPr_To_RGB32f+0];
	mul.f32 	%f529, %f528, %f472;
	.loc	20	44	0
	ld.const.f32 	%f530, [k601YPbPr_To_RGB32f+24];
	mul.f32 	%f5, %f530, %f472;
	mov.f32 	%f6, %f527;
	mov.f32 	%f7, %f529;
	bra.uni 	$Lt_3_10242;
$Lt_3_11522:
	.loc	24	48	0
	ld.param.f32 	%f531, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter1+8];
	mov.f32 	%f532, 0f00000000;   	// 0
	max.f32 	%f533, %f5, %f532;
	mov.f32 	%f534, 0f3f800000;   	// 1
	min.f32 	%f535, %f533, %f534;
	ld.param.f32 	%f536, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter1+12];
	sub.f32 	%f537, %f535, %f536;
	abs.f32 	%f538, %f537;
	mov.f32 	%f539, 0f00000000;   	// 0
	max.f32 	%f540, %f6, %f539;
	mov.f32 	%f541, 0f3f800000;   	// 1
	min.f32 	%f542, %f540, %f541;
	ld.param.f32 	%f543, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter1+16];
	sub.f32 	%f544, %f542, %f543;
	abs.f32 	%f545, %f544;
	mov.f32 	%f546, 0f00000000;   	// 0
	max.f32 	%f547, %f7, %f546;
	mov.f32 	%f548, 0f3f800000;   	// 1
	min.f32 	%f549, %f547, %f548;
	ld.param.f32 	%f550, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter1+20];
	sub.f32 	%f551, %f549, %f550;
	abs.f32 	%f552, %f551;
	mov.f32 	%f553, 0f00000000;   	// 0
	max.f32 	%f554, %f8, %f553;
	mov.f32 	%f555, 0f3f800000;   	// 1
	min.f32 	%f556, %f554, %f555;
	sub.f32 	%f557, %f556, %f8;
	abs.f32 	%f558, %f557;
	max.f32 	%f559, %f552, %f558;
	max.f32 	%f560, %f545, %f559;
	max.f32 	%f561, %f538, %f560;
	setp.ge.f32 	%p86, %f531, %f561;
	@!%p86 bra 	$Lt_3_10242;
	.loc	24	51	0
	ld.param.f32 	%f562, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter1+24];
	mov.f32 	%f124, %f562;
	ld.param.f32 	%f563, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter1+28];
	mov.f32 	%f126, %f563;
	ld.param.f32 	%f564, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter1+32];
	mov.f32 	%f128, %f564;
	ld.param.f32 	%f565, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter1+4];
	mov.f32 	%f566, 0f00000000;   	// 0
	setp.eq.f32 	%p87, %f565, %f566;
	@!%p87 bra 	$Lt_3_219906;
	.loc	24	60	0
	ld.const.f32 	%f567, [kRGB32f_To_601YPbPr+4];
	mul.f32 	%f468, %f567, %f6;
	ld.const.f32 	%f568, [kRGB32f_To_601YPbPr+0];
	mad.f32 	%f470, %f568, %f7, %f468;
	ld.const.f32 	%f569, [kRGB32f_To_601YPbPr+8];
	mad.f32 	%f472, %f569, %f5, %f470;
	mul.f32 	%f124, %f472, %f562;
	.loc	24	61	0
	mul.f32 	%f126, %f472, %f563;
	.loc	24	62	0
	mul.f32 	%f128, %f472, %f564;
$Lt_3_219906:
	.loc	24	65	0
	mov.f32 	%f5, %f124;
	mov.f32 	%f6, %f126;
	mov.f32 	%f7, %f128;
	bra.uni 	$Lt_3_10242;
$Lt_3_11778:
	.loc	25	47	0
	sub.s32 	%r293, %r13, %r12;
	sub.s32 	%r294, %r14, %r6;
	cvt.rn.f32.s32 	%f570, %r6;
	cvt.rn.f32.s32 	%f571, %r12;
	cvt.rn.f32.s32 	%f136, %r293;
	cvt.rn.f32.s32 	%f137, %r294;
	ld.param.f32 	%f457, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter1+4];
	sub.f32 	%f572, %f457, %f570;
	sub.f32 	%f573, %f457, %f571;
	sub.f32 	%f574, %f457, %f136;
	sub.f32 	%f575, %f457, %f137;
	cvt.rzi.s32.f32 	%r295, %f572;
	cvt.rzi.s32.f32 	%r296, %f573;
	cvt.rzi.s32.f32 	%r297, %f574;
	cvt.rzi.s32.f32 	%r298, %f575;
	max.s32 	%r299, %r295, %r296;
	max.s32 	%r300, %r298, %r299;
	max.s32 	%r301, %r297, %r300;
	mov.u32 	%r302, 0;
	setp.le.s32 	%p88, %r301, %r302;
	@%p88 bra 	$Lt_3_10242;
	.loc	25	51	0
	cvt.rn.f32.s32 	%f576, %r301;
	sub.f32 	%f577, %f457, %f576;
	div.full.f32 	%f578, %f577, %f457;
	mul.f32 	%f8, %f8, %f578;
	bra.uni 	$Lt_3_10242;
$Lt_3_12034:
	.loc	27	50	0
	ld.const.f32 	%f579, [kRGB32f_To_601YPbPr+4];
	mul.f32 	%f468, %f579, %f6;
	ld.param.f32 	%f580, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter1+12];
	ld.param.f32 	%f482, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter1+8];
	sub.f32 	%f581, %f580, %f482;
	ld.const.f32 	%f582, [kRGB32f_To_601YPbPr+0];
	mad.f32 	%f470, %f582, %f7, %f468;
	ld.param.f32 	%f583, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter1+16];
	mul.f32 	%f584, %f583, %f581;
	ld.const.f32 	%f585, [kRGB32f_To_601YPbPr+8];
	mad.f32 	%f472, %f585, %f5, %f470;
	add.f32 	%f586, %f584, %f580;
	sub.f32 	%f587, %f482, %f584;
	mov.f32 	%f588, 0f00000000;   	// 0
	max.f32 	%f589, %f586, %f588;
	mov.f32 	%f590, 0f00000000;   	// 0
	max.f32 	%f591, %f587, %f590;
	mov.f32 	%f592, 0f3f800000;   	// 1
	min.f32 	%f593, %f589, %f592;
	mov.f32 	%f594, 0f3f800000;   	// 1
	min.f32 	%f595, %f591, %f594;
	set.gt.u32.f32 	%r303, %f595, %f472;
	neg.s32 	%r304, %r303;
	set.le.u32.f32 	%r305, %f593, %f472;
	neg.s32 	%r306, %r305;
	or.b32 	%r307, %r304, %r306;
	mov.u32 	%r308, 0;
	setp.eq.s32 	%p89, %r307, %r308;
	@%p89 bra 	$Lt_3_221186;
	mov.f32 	%f162, 0f00000000;   	// 0
	bra.uni 	$Lt_3_221954;
$Lt_3_221186:
	add.f32 	%f596, %f584, %f482;
	mov.f32 	%f597, 0f00000000;   	// 0
	max.f32 	%f598, %f596, %f597;
	mov.f32 	%f599, 0f3f800000;   	// 1
	min.f32 	%f600, %f598, %f599;
	set.le.u32.f32 	%r309, %f600, %f472;
	neg.s32 	%r310, %r309;
	sub.f32 	%f601, %f580, %f584;
	mov.f32 	%f602, 0f00000000;   	// 0
	max.f32 	%f603, %f601, %f602;
	mov.f32 	%f604, 0f3f800000;   	// 1
	min.f32 	%f605, %f603, %f604;
	set.lt.u32.f32 	%r311, %f472, %f605;
	neg.s32 	%r312, %r311;
	and.b32 	%r313, %r310, %r312;
	mov.u32 	%r314, 0;
	setp.eq.s32 	%p90, %r313, %r314;
	@%p90 bra 	$Lt_3_221698;
	mov.f32 	%f162, 0f3f800000;   	// 1
	bra.uni 	$Lt_3_221954;
$Lt_3_221698:
	add.f32 	%f606, %f584, %f584;
	setp.gt.f32 	%p91, %f600, %f472;
	@!%p91 bra 	$Lt_3_222210;
	.loc	27	62	0
	sub.f32 	%f607, %f472, %f595;
	div.full.f32 	%f162, %f607, %f606;
	bra.uni 	$Lt_3_221954;
$Lt_3_222210:
	.loc	27	66	0
	sub.f32 	%f608, %f593, %f472;
	div.full.f32 	%f162, %f608, %f606;
$Lt_3_221954:
$Lt_3_221442:
$Lt_3_220930:
	.loc	27	69	0
	mov.f32 	%f609, 0f3f800000;   	// 1
	sub.f32 	%f610, %f609, %f162;
	ld.param.f32 	%f611, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter1+4];
	mov.f32 	%f612, 0f00000000;   	// 0
	setp.neu.f32 	%p92, %f611, %f612;
	selp.f32 	%f162, %f610, %f162, %p92;
	.loc	27	77	0
	ld.const.f32 	%f613, [k601YPbPr_To_RGB32f+24];
	mul.f32 	%f5, %f613, %f162;
	ld.const.f32 	%f614, [k601YPbPr_To_RGB32f+12];
	mul.f32 	%f6, %f614, %f162;
	ld.const.f32 	%f615, [k601YPbPr_To_RGB32f+0];
	mul.f32 	%f7, %f615, %f162;
	.loc	35	79	0
	bra.uni 	$Lt_3_10242;
$Lt_3_12290:
	.loc	35	80	0
	ld.param.f32 	%f616, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter1+44];
	mov.f32 	%f617, 0f00000000;   	// 0
	setp.ge.f32 	%p93, %f616, %f617;
	@!%p93 bra 	$L_3_199682;
	cvt.rn.f32.s32 	%f618, %r6;
	cvt.rn.f32.s32 	%f619, %r14;
	mul.f32 	%f620, %f619, %f616;
	setp.lt.f32 	%p94, %f618, %f620;
	@%p94 bra 	$L_3_199426;
$L_3_199682:
	mov.f32 	%f621, 0f00000000;   	// 0
	setp.lt.f32 	%p95, %f616, %f621;
	@!%p95 bra 	$Lt_3_10242;
	cvt.rn.f32.s32 	%f622, %r12;
	cvt.rn.f32.s32 	%f623, %r13;
	mul.f32 	%f624, %f623, %f616;
	neg.f32 	%f625, %f624;
	setp.lt.f32 	%p96, %f622, %f625;
	@!%p96 bra 	$Lt_3_10242;
$L_3_199426:
	.loc	28	47	0
	ld.const.f32 	%f626, [kRGB32f_To_601YPbPr+4];
	mul.f32 	%f627, %f626, %f6;
	ld.const.f32 	%f628, [kRGB32f_To_601YPbPr+0];
	mad.f32 	%f629, %f628, %f7, %f627;
	ld.const.f32 	%f630, [kRGB32f_To_601YPbPr+8];
	mad.f32 	%f472, %f630, %f5, %f629;
	mov.f32 	%f198, %f472;
	ld.param.f32 	%f482, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter1+8];
	ld.param.f32 	%f631, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter1+4];
	sub.f32 	%f632, %f631, %f482;
	ld.param.f32 	%f633, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter1+16];
	ld.param.f32 	%f634, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter1+12];
	sub.f32 	%f635, %f634, %f633;
	mov.f32 	%f636, 0f3f800000;   	// 1
	set.neu.u32.f32 	%r315, %f635, %f636;
	neg.s32 	%r316, %r315;
	mov.f32 	%f637, 0f3f800000;   	// 1
	set.neu.u32.f32 	%r317, %f632, %f637;
	neg.s32 	%r318, %r317;
	or.b32 	%r319, %r316, %r318;
	mov.u32 	%r320, 0;
	setp.eq.s32 	%p97, %r319, %r320;
	@%p97 bra 	$Lt_3_222466;
	.loc	17	116	0
	mov.u32 	%r321, 1;
	sub.u32 	%r71, %r321, %r5;
	shr.u32 	%r72, %r11, 13;
	sub.u32 	%r73, %r5, %r11;
	sub.u32 	%r74, %r71, %r11;
	xor.b32 	%r75, %r74, %r72;
	shl.b32 	%r76, %r75, 8;
	sub.u32 	%r77, %r73, %r75;
	sub.u32 	%r78, %r11, %r75;
	xor.b32 	%r79, %r77, %r76;
	shr.u32 	%r80, %r79, 13;
	sub.u32 	%r81, %r78, %r79;
	xor.b32 	%r82, %r81, %r80;
	shr.u32 	%r322, %r82, 12;
	sub.u32 	%r323, %r75, %r79;
	sub.u32 	%r324, %r323, %r82;
	xor.b32 	%r86, %r322, %r324;
	.loc	17	117	0
	sub.u32 	%r325, %r79, %r82;
	sub.u32 	%r88, %r325, %r86;
	shl.b32 	%r326, %r86, 16;
	xor.b32 	%r90, %r88, %r326;
	.loc	17	118	0
	sub.u32 	%r327, %r82, %r86;
	sub.u32 	%r92, %r327, %r90;
	shr.u32 	%r328, %r90, 5;
	xor.b32 	%r94, %r92, %r328;
	.loc	17	119	0
	sub.u32 	%r95, %r86, %r90;
	sub.u32 	%r96, %r95, %r94;
	shr.u32 	%r329, %r94, 3;
	xor.b32 	%r98, %r96, %r329;
	.loc	17	120	0
	sub.u32 	%r99, %r90, %r94;
	sub.u32 	%r100, %r99, %r98;
	shl.b32 	%r330, %r98, 10;
	xor.b32 	%r102, %r100, %r330;
	.loc	17	121	0
	sub.u32 	%r103, %r94, %r98;
	sub.u32 	%r104, %r103, %r102;
	shr.u32 	%r331, %r102, 15;
	xor.b32 	%r106, %r104, %r331;
	.loc	28	57	0
	mov.f32 	%f638, 0f3b270d73;   	// 0.00254902
	mul.lo.u32 	%r332, %r106, 1103515245;
	add.u32 	%r333, %r332, 12345;
	shr.u32 	%r334, %r333, 16;
	and.b32 	%r335, %r334, 255;
	shl.b32 	%r336, %r335, 7;
	mul.lo.u32 	%r337, %r106, -1029531031;
	sub.u32 	%r338, %r337, 740551042;
	shr.u32 	%r339, %r338, 16;
	and.b32 	%r340, %r339, 255;
	xor.b32 	%r341, %r336, %r340;
	cvt.rn.f32.s32 	%f639, %r341;
	mov.f32 	%f640, 0f467ffe00;   	// 16383.5
	div.full.f32 	%f641, %f639, %f640;
	mov.f32 	%f642, 0fbf800000;   	// -1
	add.f32 	%f643, %f641, %f642;
	mad.f32 	%f198, %f638, %f643, %f472;
$Lt_3_222466:
	sub.f32 	%f644, %f198, %f482;
	ld.param.f32 	%f645, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter1+20];
	mov.f32 	%f646, 0f3f800000;   	// 1
	setp.neu.f32 	%p98, %f645, %f646;
	@!%p98 bra 	$Lt_3_223234;
	mov.f32 	%f647, 0f00000000;   	// 0
	setp.lt.f32 	%p99, %f644, %f647;
	@!%p99 bra 	$Lt_3_223746;
	.loc	28	66	0
	mov.f32 	%f198, %f633;
	bra.uni 	$Lt_3_222978;
$Lt_3_223746:
	.loc	28	70	0
	rcp.approx.f32 	%f648, %f632;
	mul.f32 	%f649, %f644, %f648;
	lg2.approx.f32 	%f650, %f649;
	mul.f32 	%f651, %f645, %f650;
	ex2.approx.f32 	%f652, %f651;
	mad.f32 	%f198, %f635, %f652, %f633;
	bra.uni 	$Lt_3_222978;
$Lt_3_223234:
	.loc	28	77	0
	rcp.approx.f32 	%f653, %f632;
	mul.f32 	%f654, %f644, %f653;
	mad.f32 	%f198, %f635, %f654, %f633;
$Lt_3_222978:
	.loc	19	267	0
	ld.const.f32 	%f655, [kRGB32f_To_601YPbPr+28];
	mul.f32 	%f656, %f655, %f6;
	ld.const.f32 	%f657, [kRGB32f_To_601YPbPr+16];
	mul.f32 	%f658, %f657, %f6;
	ld.param.f32 	%f659, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter1+28];
	sin.approx.f32 	%f660, %f659;
	cos.approx.f32 	%f661, %f659;
	ld.const.f32 	%f662, [kRGB32f_To_601YPbPr+24];
	mad.f32 	%f663, %f662, %f7, %f656;
	ld.const.f32 	%f664, [kRGB32f_To_601YPbPr+12];
	mad.f32 	%f665, %f664, %f7, %f658;
	ld.const.f32 	%f666, [kRGB32f_To_601YPbPr+32];
	mad.f32 	%f667, %f666, %f5, %f663;
	ld.const.f32 	%f668, [kRGB32f_To_601YPbPr+20];
	mad.f32 	%f669, %f668, %f5, %f665;
	mul.f32 	%f670, %f660, %f667;
	mul.f32 	%f671, %f661, %f667;
	mul.f32 	%f672, %f669, %f661;
	sub.f32 	%f673, %f672, %f670;
	mad.f32 	%f674, %f669, %f660, %f671;
	ld.param.f32 	%f675, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter1+32];
	sub.f32 	%f676, %f675, %f673;
	ld.param.f32 	%f677, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter1+36];
	sub.f32 	%f678, %f677, %f674;
	ld.param.f32 	%f679, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter1+40];
	mad.f32 	%f680, %f679, %f676, %f673;
	mad.f32 	%f681, %f679, %f678, %f674;
	ld.param.f32 	%f682, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter1+24];
	mul.f32 	%f683, %f680, %f682;
	mul.f32 	%f684, %f681, %f682;
	ld.const.f32 	%f685, [k601YPbPr_To_RGB32f+16];
	mul.f32 	%f686, %f685, %f683;
	ld.const.f32 	%f687, [k601YPbPr_To_RGB32f+12];
	mad.f32 	%f688, %f687, %f198, %f686;
	ld.const.f32 	%f689, [k601YPbPr_To_RGB32f+20];
	mad.f32 	%f258, %f689, %f684, %f688;
	.loc	19	268	0
	ld.const.f32 	%f690, [k601YPbPr_To_RGB32f+4];
	mul.f32 	%f691, %f690, %f683;
	ld.const.f32 	%f692, [k601YPbPr_To_RGB32f+0];
	mad.f32 	%f693, %f692, %f198, %f691;
	ld.const.f32 	%f694, [k601YPbPr_To_RGB32f+8];
	mad.f32 	%f264, %f694, %f684, %f693;
	.loc	28	92	0
	ld.const.f32 	%f695, [k601YPbPr_To_RGB32f+28];
	mul.f32 	%f696, %f695, %f683;
	ld.const.f32 	%f697, [k601YPbPr_To_RGB32f+24];
	mad.f32 	%f698, %f697, %f198, %f696;
	ld.const.f32 	%f699, [k601YPbPr_To_RGB32f+32];
	mad.f32 	%f5, %f699, %f684, %f698;
	mov.f32 	%f6, %f258;
	mov.f32 	%f7, %f264;
	bra.uni 	$Lt_3_10242;
$Lt_3_12546:
	.loc	35	83	0
	ld.param.f32 	%f700, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter1+44];
	mov.f32 	%f701, 0f00000000;   	// 0
	setp.ge.f32 	%p100, %f700, %f701;
	@!%p100 bra 	$L_3_200450;
	cvt.rn.f32.s32 	%f702, %r6;
	cvt.rn.f32.s32 	%f703, %r14;
	mul.f32 	%f704, %f703, %f700;
	setp.lt.f32 	%p101, %f702, %f704;
	@%p101 bra 	$L_3_200194;
$L_3_200450:
	mov.f32 	%f705, 0f00000000;   	// 0
	setp.lt.f32 	%p102, %f700, %f705;
	@!%p102 bra 	$Lt_3_10242;
	cvt.rn.f32.s32 	%f706, %r12;
	cvt.rn.f32.s32 	%f707, %r13;
	mul.f32 	%f708, %f707, %f700;
	neg.f32 	%f709, %f708;
	setp.lt.f32 	%p103, %f706, %f709;
	@!%p103 bra 	$Lt_3_10242;
$L_3_200194:
	.loc	28	110	0
	ld.const.f32 	%f710, [kRGB32f_To_601YPbPr+4];
	mul.f32 	%f468, %f710, %f6;
	ld.param.f32 	%f482, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter1+8];
	ld.param.f32 	%f711, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter1+4];
	sub.f32 	%f712, %f711, %f482;
	ld.param.f32 	%f713, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter1+16];
	ld.param.f32 	%f714, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter1+12];
	sub.f32 	%f715, %f714, %f713;
	ld.const.f32 	%f716, [kRGB32f_To_601YPbPr+0];
	mad.f32 	%f470, %f716, %f7, %f468;
	ld.const.f32 	%f717, [kRGB32f_To_601YPbPr+8];
	mad.f32 	%f472, %f717, %f5, %f470;
	sub.f32 	%f718, %f472, %f482;
	ld.param.f32 	%f719, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter1+20];
	mov.f32 	%f720, 0f3f800000;   	// 1
	setp.neu.f32 	%p104, %f719, %f720;
	@!%p104 bra 	$Lt_3_224258;
	.loc	28	120	0
	mov.f32 	%f721, 0f00000000;   	// 0
	max.f32 	%f722, %f718, %f721;
	div.full.f32 	%f723, %f722, %f712;
	lg2.approx.f32 	%f724, %f723;
	mul.f32 	%f725, %f719, %f724;
	ex2.approx.f32 	%f726, %f725;
	mad.f32 	%f297, %f715, %f726, %f713;
	bra.uni 	$Lt_3_224002;
$Lt_3_224258:
	.loc	28	129	0
	div.full.f32 	%f727, %f718, %f712;
	mad.f32 	%f297, %f715, %f727, %f713;
$Lt_3_224002:
	.loc	28	135	0
	ld.const.f32 	%f728, [k601YPbPr_To_RGB32f+24];
	mul.f32 	%f5, %f728, %f297;
	ld.const.f32 	%f729, [k601YPbPr_To_RGB32f+12];
	mul.f32 	%f6, %f729, %f297;
	ld.const.f32 	%f730, [k601YPbPr_To_RGB32f+0];
	mul.f32 	%f7, %f730, %f297;
	bra.uni 	$Lt_3_10242;
$Lt_3_12802:
	.loc	35	86	0
	ld.param.f32 	%f731, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter1+44];
	mov.f32 	%f732, 0f00000000;   	// 0
	setp.ge.f32 	%p105, %f731, %f732;
	@!%p105 bra 	$L_3_201218;
	cvt.rn.f32.s32 	%f733, %r6;
	cvt.rn.f32.s32 	%f734, %r14;
	mul.f32 	%f735, %f734, %f731;
	setp.lt.f32 	%p106, %f733, %f735;
	@%p106 bra 	$L_3_200962;
$L_3_201218:
	mov.f32 	%f736, 0f00000000;   	// 0
	setp.lt.f32 	%p107, %f731, %f736;
	@!%p107 bra 	$Lt_3_10242;
	cvt.rn.f32.s32 	%f737, %r12;
	cvt.rn.f32.s32 	%f738, %r13;
	mul.f32 	%f739, %f738, %f731;
	neg.f32 	%f740, %f739;
	setp.lt.f32 	%p108, %f737, %f740;
	@!%p108 bra 	$Lt_3_10242;
$L_3_200962:
	.loc	28	160	0
	ld.const.f32 	%f5, [k601YPbPr_To_RGB32f+24];
	ld.const.f32 	%f6, [k601YPbPr_To_RGB32f+12];
	ld.const.f32 	%f7, [k601YPbPr_To_RGB32f+0];
	bra.uni 	$Lt_3_10242;
$Lt_3_13058:
	.loc	29	42	0
	ld.param.f32 	%f457, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter1+4];
	mov.f32 	%f741, 0f00000000;   	// 0
	setp.ge.f32 	%p109, %f5, %f741;
	@!%p109 bra 	$Lt_3_224770;
	.loc	19	292	0
	lg2.approx.f32 	%f742, %f5;
	mul.f32 	%f743, %f457, %f742;
	ex2.approx.f32 	%f744, %f743;
	bra.uni 	$Lt_3_224514;
$Lt_3_224770:
	neg.f32 	%f745, %f5;
	lg2.approx.f32 	%f746, %f745;
	mul.f32 	%f747, %f457, %f746;
	ex2.approx.f32 	%f748, %f747;
	neg.f32 	%f744, %f748;
$Lt_3_224514:
	mov.f32 	%f749, 0f00000000;   	// 0
	setp.ge.f32 	%p110, %f6, %f749;
	@!%p110 bra 	$Lt_3_225282;
	.loc	19	293	0
	lg2.approx.f32 	%f750, %f6;
	mul.f32 	%f751, %f457, %f750;
	ex2.approx.f32 	%f752, %f751;
	bra.uni 	$Lt_3_225026;
$Lt_3_225282:
	neg.f32 	%f753, %f6;
	lg2.approx.f32 	%f754, %f753;
	mul.f32 	%f755, %f457, %f754;
	ex2.approx.f32 	%f756, %f755;
	neg.f32 	%f752, %f756;
$Lt_3_225026:
	mov.f32 	%f757, 0f00000000;   	// 0
	setp.ge.f32 	%p111, %f7, %f757;
	@!%p111 bra 	$Lt_3_225794;
	.loc	19	294	0
	lg2.approx.f32 	%f758, %f7;
	mul.f32 	%f759, %f457, %f758;
	ex2.approx.f32 	%f760, %f759;
	bra.uni 	$Lt_3_225538;
$Lt_3_225794:
	neg.f32 	%f761, %f7;
	lg2.approx.f32 	%f762, %f761;
	mul.f32 	%f763, %f457, %f762;
	ex2.approx.f32 	%f764, %f763;
	neg.f32 	%f760, %f764;
$Lt_3_225538:
	mov.f32 	%f765, 0f00000000;   	// 0
	setp.ge.f32 	%p112, %f8, %f765;
	@!%p112 bra 	$Lt_3_226306;
	.loc	19	295	0
	lg2.approx.f32 	%f766, %f8;
	mul.f32 	%f767, %f457, %f766;
	ex2.approx.f32 	%f768, %f767;
	bra.uni 	$Lt_3_226050;
$Lt_3_226306:
	neg.f32 	%f769, %f8;
	lg2.approx.f32 	%f770, %f769;
	mul.f32 	%f771, %f457, %f770;
	ex2.approx.f32 	%f772, %f771;
	neg.f32 	%f768, %f772;
$Lt_3_226050:
	.loc	29	42	0
	mov.f32 	%f5, %f744;
	mov.f32 	%f6, %f752;
	mov.f32 	%f7, %f760;
	mov.f32 	%f8, %f768;
	.loc	35	91	0
	bra.uni 	$Lt_3_10242;
$Lt_3_13314:
	.loc	30	41	0
	sub.s32 	%r342, %r14, %r6;
	sub.s32 	%r6, %r342, 1;
	.loc	35	94	0
	bra.uni 	$Lt_3_10242;
$Lt_3_13570:
	.loc	35	96	0
	sub.u32 	%r343, %r6, %r12;
	ld.param.f32 	%f457, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter1+4];
	mov.f32 	%f773, 0f3e800000;   	// 0.25
	mul.f32 	%f774, %f457, %f773;
	ld.param.f32 	%f775, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter1+16];
	ld.param.f32 	%f776, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter1+8];
	mov.f32 	%f777, 0f3f800000;   	// 1
	setp.eq.f32 	%p113, %f776, %f777;
	@!%p113 bra 	$Lt_3_226818;
	.loc	17	117	0
	cvt.rzi.u32.f32 	%r344, %f775;
	shr.u32 	%r345, %r344, 13;
	sub.u32 	%r346, %r343, %r344;
	sub.u32 	%r347, %r12, %r344;
	xor.b32 	%r348, %r346, %r345;
	shl.b32 	%r349, %r348, 8;
	sub.u32 	%r350, %r347, %r348;
	sub.u32 	%r351, %r344, %r348;
	xor.b32 	%r352, %r350, %r349;
	shr.u32 	%r353, %r352, 13;
	sub.u32 	%r354, %r351, %r352;
	sub.u32 	%r355, %r348, %r352;
	xor.b32 	%r356, %r354, %r353;
	shr.u32 	%r357, %r356, 12;
	sub.u32 	%r358, %r355, %r356;
	xor.b32 	%r359, %r358, %r357;
	shl.b32 	%r360, %r359, 16;
	sub.u32 	%r361, %r352, %r356;
	sub.u32 	%r362, %r361, %r359;
	xor.b32 	%r363, %r360, %r362;
	.loc	17	118	0
	sub.u32 	%r364, %r356, %r359;
	sub.u32 	%r365, %r364, %r363;
	shr.u32 	%r366, %r363, 5;
	xor.b32 	%r367, %r365, %r366;
	.loc	17	119	0
	sub.u32 	%r368, %r359, %r363;
	sub.u32 	%r369, %r368, %r367;
	shr.u32 	%r370, %r367, 3;
	xor.b32 	%r371, %r369, %r370;
	.loc	17	120	0
	sub.u32 	%r372, %r363, %r367;
	sub.u32 	%r373, %r372, %r371;
	shl.b32 	%r374, %r371, 10;
	xor.b32 	%r375, %r373, %r374;
	.loc	17	121	0
	sub.u32 	%r376, %r367, %r371;
	sub.u32 	%r377, %r376, %r375;
	shr.u32 	%r378, %r375, 15;
	xor.b32 	%r379, %r377, %r378;
	.loc	31	48	0
	mul.lo.u32 	%r380, %r379, 1103515245;
	add.u32 	%r381, %r380, 12345;
	shr.u32 	%r382, %r381, 16;
	and.b32 	%r383, %r382, 255;
	shl.b32 	%r384, %r383, 7;
	mul.lo.u32 	%r385, %r379, -1029531031;
	sub.u32 	%r386, %r385, 740551042;
	shr.u32 	%r387, %r386, 16;
	and.b32 	%r388, %r387, 255;
	xor.b32 	%r389, %r384, %r388;
	cvt.rn.f32.s32 	%f778, %r389;
	mov.f32 	%f779, 0f46fffe00;   	// 32767
	div.full.f32 	%f780, %f778, %f779;
	mul.f32 	%f781, %f457, %f780;
	mov.f32 	%f782, 0f3f000000;   	// 0.5
	mul.f32 	%f783, %f781, %f782;
	sub.f32 	%f784, %f783, %f774;
	.loc	17	117	0
	add.u32 	%r390, %r346, 1;
	xor.b32 	%r391, %r390, %r345;
	shl.b32 	%r392, %r391, 8;
	sub.u32 	%r393, %r347, %r391;
	sub.u32 	%r394, %r344, %r391;
	xor.b32 	%r395, %r392, %r393;
	shr.u32 	%r396, %r395, 13;
	sub.u32 	%r397, %r394, %r395;
	sub.u32 	%r398, %r391, %r395;
	xor.b32 	%r399, %r396, %r397;
	shr.u32 	%r400, %r399, 12;
	sub.u32 	%r401, %r398, %r399;
	xor.b32 	%r402, %r400, %r401;
	sub.u32 	%r403, %r395, %r399;
	sub.u32 	%r178, %r403, %r402;
	shl.b32 	%r404, %r402, 16;
	xor.b32 	%r178, %r178, %r404;
	.loc	17	118	0
	sub.u32 	%r180, %r399, %r402;
	sub.u32 	%r180, %r180, %r178;
	shr.u32 	%r405, %r178, 5;
	xor.b32 	%r180, %r180, %r405;
	.loc	17	119	0
	sub.u32 	%r182, %r402, %r178;
	sub.u32 	%r182, %r182, %r180;
	shr.u32 	%r406, %r180, 3;
	xor.b32 	%r182, %r182, %r406;
	.loc	17	120	0
	sub.u32 	%r178, %r178, %r180;
	sub.u32 	%r178, %r178, %r182;
	shl.b32 	%r407, %r182, 10;
	xor.b32 	%r178, %r178, %r407;
	.loc	17	121	0
	sub.u32 	%r180, %r180, %r182;
	sub.u32 	%r180, %r180, %r178;
	shr.u32 	%r408, %r178, 15;
	xor.b32 	%r180, %r180, %r408;
	.loc	31	49	0
	mul.lo.u32 	%r409, %r180, 1103515245;
	add.u32 	%r410, %r409, 12345;
	shr.u32 	%r411, %r410, 16;
	and.b32 	%r412, %r411, 255;
	shl.b32 	%r413, %r412, 7;
	mul.lo.u32 	%r414, %r180, -1029531031;
	sub.u32 	%r415, %r414, 740551042;
	shr.u32 	%r416, %r415, 16;
	and.b32 	%r417, %r416, 255;
	xor.b32 	%r418, %r413, %r417;
	cvt.rn.f32.s32 	%f785, %r418;
	mov.f32 	%f786, 0f46fffe00;   	// 32767
	div.full.f32 	%f787, %f785, %f786;
	mul.f32 	%f788, %f457, %f787;
	mov.f32 	%f789, 0f3f000000;   	// 0.5
	mul.f32 	%f790, %f788, %f789;
	sub.f32 	%f362, %f790, %f774;
	.loc	17	117	0
	add.u32 	%r419, %r346, 2;
	xor.b32 	%r420, %r419, %r345;
	shl.b32 	%r421, %r420, 8;
	sub.u32 	%r422, %r347, %r420;
	sub.u32 	%r423, %r344, %r420;
	xor.b32 	%r424, %r421, %r422;
	shr.u32 	%r425, %r424, 13;
	sub.u32 	%r426, %r423, %r424;
	sub.u32 	%r427, %r420, %r424;
	xor.b32 	%r428, %r425, %r426;
	shr.u32 	%r429, %r428, 12;
	sub.u32 	%r430, %r427, %r428;
	xor.b32 	%r431, %r429, %r430;
	sub.u32 	%r432, %r424, %r428;
	sub.u32 	%r210, %r432, %r431;
	shl.b32 	%r433, %r431, 16;
	xor.b32 	%r210, %r210, %r433;
	.loc	17	118	0
	sub.u32 	%r212, %r428, %r431;
	sub.u32 	%r212, %r212, %r210;
	shr.u32 	%r434, %r210, 5;
	xor.b32 	%r212, %r212, %r434;
	.loc	17	119	0
	sub.u32 	%r214, %r431, %r210;
	sub.u32 	%r214, %r214, %r212;
	shr.u32 	%r435, %r212, 3;
	xor.b32 	%r214, %r214, %r435;
	.loc	17	120	0
	sub.u32 	%r210, %r210, %r212;
	sub.u32 	%r210, %r210, %r214;
	shl.b32 	%r436, %r214, 10;
	xor.b32 	%r210, %r210, %r436;
	.loc	17	121	0
	sub.u32 	%r212, %r212, %r214;
	sub.u32 	%r212, %r212, %r210;
	shr.u32 	%r437, %r210, 15;
	xor.b32 	%r212, %r212, %r437;
	.loc	31	50	0
	mul.lo.u32 	%r438, %r212, 1103515245;
	add.u32 	%r439, %r438, 12345;
	shr.u32 	%r440, %r439, 16;
	and.b32 	%r441, %r440, 255;
	shl.b32 	%r442, %r441, 7;
	mul.lo.u32 	%r443, %r212, -1029531031;
	sub.u32 	%r444, %r443, 740551042;
	shr.u32 	%r445, %r444, 16;
	and.b32 	%r446, %r445, 255;
	xor.b32 	%r447, %r442, %r446;
	cvt.rn.f32.s32 	%f791, %r447;
	mov.f32 	%f792, 0f46fffe00;   	// 32767
	div.full.f32 	%f793, %f791, %f792;
	mul.f32 	%f794, %f457, %f793;
	mov.f32 	%f795, 0f3f000000;   	// 0.5
	mul.f32 	%f796, %f794, %f795;
	sub.f32 	%f369, %f796, %f774;
	bra.uni 	$Lt_3_226562;
$Lt_3_226818:
	.loc	17	117	0
	cvt.rzi.u32.f32 	%r448, %f775;
	shr.u32 	%r449, %r448, 13;
	sub.u32 	%r450, %r343, %r448;
	sub.u32 	%r451, %r12, %r448;
	xor.b32 	%r452, %r450, %r449;
	shl.b32 	%r453, %r452, 8;
	sub.u32 	%r454, %r451, %r452;
	sub.u32 	%r455, %r448, %r452;
	xor.b32 	%r456, %r454, %r453;
	shr.u32 	%r457, %r456, 13;
	sub.u32 	%r458, %r455, %r456;
	sub.u32 	%r459, %r452, %r456;
	xor.b32 	%r460, %r458, %r457;
	shr.u32 	%r461, %r460, 12;
	sub.u32 	%r462, %r459, %r460;
	xor.b32 	%r463, %r462, %r461;
	shl.b32 	%r464, %r463, 16;
	sub.u32 	%r465, %r456, %r460;
	sub.u32 	%r466, %r465, %r463;
	xor.b32 	%r467, %r464, %r466;
	.loc	17	118	0
	sub.u32 	%r468, %r460, %r463;
	sub.u32 	%r469, %r468, %r467;
	shr.u32 	%r470, %r467, 5;
	xor.b32 	%r471, %r469, %r470;
	.loc	17	119	0
	sub.u32 	%r472, %r463, %r467;
	sub.u32 	%r473, %r472, %r471;
	shr.u32 	%r474, %r471, 3;
	xor.b32 	%r475, %r473, %r474;
	.loc	17	120	0
	sub.u32 	%r476, %r467, %r471;
	sub.u32 	%r477, %r476, %r475;
	shl.b32 	%r478, %r475, 10;
	xor.b32 	%r479, %r477, %r478;
	.loc	17	121	0
	sub.u32 	%r480, %r471, %r475;
	sub.u32 	%r481, %r480, %r479;
	shr.u32 	%r482, %r479, 15;
	xor.b32 	%r483, %r481, %r482;
	.loc	31	54	0
	mul.lo.u32 	%r484, %r483, 1103515245;
	mul.lo.u32 	%r485, %r483, -1029531031;
	add.u32 	%r486, %r484, 12345;
	sub.u32 	%r487, %r485, 740551042;
	shr.u32 	%r488, %r486, 16;
	shr.u32 	%r489, %r487, 16;
	and.b32 	%r490, %r488, 255;
	and.b32 	%r491, %r489, 255;
	shl.b32 	%r492, %r490, 7;
	xor.b32 	%r493, %r492, %r491;
	cvt.rn.f32.s32 	%f797, %r493;
	mov.f32 	%f798, 0f46fffe00;   	// 32767
	div.full.f32 	%f799, %f797, %f798;
	mul.f32 	%f800, %f799, %f457;
	mov.f32 	%f801, 0f3f000000;   	// 0.5
	mul.f32 	%f802, %f800, %f801;
	sub.f32 	%f784, %f802, %f774;
	mov.f32 	%f369, %f784;
	mov.f32 	%f362, %f784;
$Lt_3_226562:
	.loc	31	57	0
	add.f32 	%f7, %f784, %f7;
	.loc	31	58	0
	add.f32 	%f6, %f362, %f6;
	.loc	31	59	0
	add.f32 	%f5, %f5, %f369;
	ld.param.f32 	%f803, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter1+12];
	mov.f32 	%f804, 0f3f800000;   	// 1
	setp.eq.f32 	%p114, %f803, %f804;
	@!%p114 bra 	$Lt_3_10242;
	.loc	31	63	0
	mov.f32 	%f805, 0f00000000;   	// 0
	max.f32 	%f806, %f5, %f805;
	mov.f32 	%f807, 0f3f800000;   	// 1
	min.f32 	%f5, %f806, %f807;
	mov.f32 	%f808, 0f00000000;   	// 0
	max.f32 	%f809, %f6, %f808;
	mov.f32 	%f810, 0f3f800000;   	// 1
	min.f32 	%f6, %f809, %f810;
	mov.f32 	%f811, 0f00000000;   	// 0
	max.f32 	%f812, %f7, %f811;
	mov.f32 	%f813, 0f3f800000;   	// 1
	min.f32 	%f7, %f812, %f813;
	mov.f32 	%f814, 0f00000000;   	// 0
	max.f32 	%f815, %f8, %f814;
	mov.f32 	%f816, 0f3f800000;   	// 1
	min.f32 	%f8, %f815, %f816;
	bra.uni 	$Lt_3_10242;
$Lt_3_13826:
	.loc	35	99	0
	cvt.rn.f32.s32 	%f817, %r6;
	cvt.rn.f32.s32 	%f818, %r14;
	ld.param.f32 	%f819, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter1+20];
	mul.f32 	%f820, %f818, %f819;
	setp.lt.f32 	%p115, %f817, %f820;
	@!%p115 bra 	$Lt_3_10242;
	.loc	19	267	0
	ld.const.f32 	%f821, [kRGB32f_To_601YPbPr+4];
	mul.f32 	%f468, %f821, %f6;
	ld.const.f32 	%f822, [kRGB32f_To_601YPbPr+28];
	mul.f32 	%f823, %f822, %f6;
	ld.const.f32 	%f824, [kRGB32f_To_601YPbPr+16];
	mul.f32 	%f825, %f824, %f6;
	ld.const.f32 	%f826, [kRGB32f_To_601YPbPr+0];
	mad.f32 	%f470, %f826, %f7, %f468;
	ld.const.f32 	%f827, [kRGB32f_To_601YPbPr+24];
	mad.f32 	%f828, %f827, %f7, %f823;
	ld.const.f32 	%f829, [kRGB32f_To_601YPbPr+12];
	mad.f32 	%f830, %f829, %f7, %f825;
	ld.const.f32 	%f831, [kRGB32f_To_601YPbPr+8];
	mad.f32 	%f472, %f831, %f5, %f470;
	ld.const.f32 	%f832, [kRGB32f_To_601YPbPr+32];
	mad.f32 	%f833, %f832, %f5, %f828;
	ld.const.f32 	%f834, [kRGB32f_To_601YPbPr+20];
	mad.f32 	%f835, %f834, %f5, %f830;
	ld.param.f32 	%f836, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter1+4];
	ld.param.f32 	%f837, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter1+8];
	mad.f32 	%f838, %f472, %f837, %f836;
	ld.param.f32 	%f839, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter1+16];
	mul.f32 	%f840, %f833, %f839;
	ld.param.f32 	%f841, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter1+12];
	mul.f32 	%f842, %f833, %f841;
	mul.f32 	%f843, %f835, %f841;
	sub.f32 	%f844, %f843, %f840;
	mad.f32 	%f845, %f835, %f839, %f842;
	ld.const.f32 	%f846, [k601YPbPr_To_RGB32f+16];
	mul.f32 	%f847, %f846, %f844;
	ld.const.f32 	%f848, [k601YPbPr_To_RGB32f+12];
	mad.f32 	%f849, %f848, %f838, %f847;
	ld.const.f32 	%f850, [k601YPbPr_To_RGB32f+20];
	mad.f32 	%f424, %f850, %f845, %f849;
	.loc	19	268	0
	ld.const.f32 	%f851, [k601YPbPr_To_RGB32f+4];
	mul.f32 	%f852, %f851, %f844;
	ld.const.f32 	%f853, [k601YPbPr_To_RGB32f+0];
	mad.f32 	%f854, %f853, %f838, %f852;
	ld.const.f32 	%f855, [k601YPbPr_To_RGB32f+8];
	mad.f32 	%f430, %f855, %f845, %f854;
	.loc	32	56	0
	ld.const.f32 	%f856, [k601YPbPr_To_RGB32f+28];
	mul.f32 	%f857, %f856, %f844;
	ld.const.f32 	%f858, [k601YPbPr_To_RGB32f+24];
	mad.f32 	%f859, %f858, %f838, %f857;
	ld.const.f32 	%f860, [k601YPbPr_To_RGB32f+32];
	mad.f32 	%f5, %f860, %f845, %f859;
	mov.f32 	%f6, %f424;
	mov.f32 	%f7, %f430;
	bra.uni 	$Lt_3_10242;
$Lt_3_14082:
	.loc	33	46	0
	ld.const.f32 	%f861, [kRGB32f_To_601YPbPr+4];
	mul.f32 	%f468, %f861, %f6;
	ld.const.f32 	%f862, [kRGB32f_To_601YPbPr+0];
	mad.f32 	%f470, %f862, %f7, %f468;
	ld.const.f32 	%f863, [kRGB32f_To_601YPbPr+8];
	mad.f32 	%f472, %f863, %f5, %f470;
	ld.param.f32 	%f482, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter1+8];
	ld.param.f32 	%f864, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter1+20];
	sub.f32 	%f865, %f864, %f482;
	mad.f32 	%f441, %f472, %f865, %f482;
	.loc	33	47	0
	ld.param.f32 	%f866, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter1+12];
	ld.param.f32 	%f867, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter1+24];
	sub.f32 	%f868, %f867, %f866;
	mad.f32 	%f445, %f472, %f868, %f866;
	.loc	33	49	0
	ld.param.f32 	%f869, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter1+28];
	ld.param.f32 	%f457, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter1+4];
	ld.param.f32 	%f870, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter1+16];
	sub.f32 	%f871, %f870, %f457;
	mad.f32 	%f872, %f472, %f871, %f457;
	sub.f32 	%f873, %f872, %f5;
	mad.f32 	%f5, %f869, %f873, %f5;
	.loc	33	50	0
	sub.f32 	%f874, %f441, %f6;
	mad.f32 	%f6, %f869, %f874, %f6;
	.loc	33	51	0
	sub.f32 	%f875, %f445, %f7;
	mad.f32 	%f7, %f869, %f875, %f7;
	.loc	35	103	0
	bra.uni 	$Lt_3_10242;
$Lt_3_14338:
	.loc	34	41	0
	sub.s32 	%r494, %r13, %r12;
	sub.s32 	%r12, %r494, 1;
$Lt_3_10242:
	.loc	35	54	0
	ld.param.u32 	%r495, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter2+0];
	mov.u32 	%r496, 0;
	setp.eq.s32 	%p116, %r495, %r496;
	@%p116 bra 	$Lt_3_5122;
	mov.u32 	%r497, 1;
	setp.eq.s32 	%p117, %r495, %r497;
	@%p117 bra 	$Lt_3_5634;
	mov.u32 	%r498, 2;
	setp.eq.s32 	%p118, %r495, %r498;
	@%p118 bra 	$Lt_3_5890;
	mov.u32 	%r499, 3;
	setp.eq.s32 	%p119, %r495, %r499;
	@%p119 bra 	$Lt_3_6146;
	mov.u32 	%r500, 4;
	setp.eq.s32 	%p120, %r495, %r500;
	@%p120 bra 	$Lt_3_6402;
	mov.u32 	%r501, 5;
	setp.eq.s32 	%p121, %r495, %r501;
	@%p121 bra 	$Lt_3_6658;
	mov.u32 	%r502, 6;
	setp.eq.s32 	%p122, %r495, %r502;
	@%p122 bra 	$Lt_3_6914;
	mov.u32 	%r503, 7;
	setp.eq.s32 	%p123, %r495, %r503;
	@%p123 bra 	$Lt_3_7170;
	mov.u32 	%r504, 8;
	setp.eq.s32 	%p124, %r495, %r504;
	@%p124 bra 	$Lt_3_7426;
	mov.u32 	%r505, 9;
	setp.eq.s32 	%p125, %r495, %r505;
	@%p125 bra 	$Lt_3_7682;
	mov.u32 	%r506, 10;
	setp.eq.s32 	%p126, %r495, %r506;
	@%p126 bra 	$Lt_3_7938;
	mov.u32 	%r507, 11;
	setp.eq.s32 	%p127, %r495, %r507;
	@%p127 bra 	$Lt_3_8194;
	mov.u32 	%r508, 12;
	setp.eq.s32 	%p128, %r495, %r508;
	@%p128 bra 	$Lt_3_8450;
	mov.u32 	%r509, 13;
	setp.eq.s32 	%p129, %r495, %r509;
	@%p129 bra 	$Lt_3_8706;
	mov.u32 	%r510, 14;
	setp.eq.s32 	%p130, %r495, %r510;
	@%p130 bra 	$Lt_3_8962;
	mov.u32 	%r511, 15;
	setp.eq.s32 	%p131, %r495, %r511;
	@%p131 bra 	$Lt_3_9218;
	mov.u32 	%r512, 16;
	setp.eq.s32 	%p132, %r495, %r512;
	@%p132 bra 	$Lt_3_9474;
	bra.uni 	$Lt_3_5378;
$Lt_3_5122:
	.loc	18	42	0
	cvt.sat.f32.f32 	%f8, %f8;
	ld.param.f32 	%f876, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter2+8];
	mov.f32 	%f877, 0f3f800000;   	// 1
	setp.eq.f32 	%p133, %f876, %f877;
	ld.param.f32 	%f878, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter2+12];
	mov.f32 	%f879, 0f3f800000;   	// 1
	setp.eq.f32 	%p134, %f878, %f879;
	ld.param.f32 	%f880, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter2+4];
	ld.param.f32 	%f881, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter2+16];
	mov.f32 	%f882, 0f3f800000;   	// 1
	setp.eq.f32 	%p135, %f881, %f882;
	@!%p135 bra 	$Lt_3_228354;
	.loc	18	45	0
	mul.f32 	%f883, %f8, %f880;
	selp.f32 	%f8, %f880, %f883, %p133;
	.loc	18	53	0
	mov.f32 	%f884, 0f3f800000;   	// 1
	sub.f32 	%f885, %f884, %f8;
	selp.f32 	%f8, %f885, %f8, %p134;
	.loc	18	57	0
	mov.f32 	%f5, %f8;
	mov.f32 	%f6, %f8;
	mov.f32 	%f7, %f8;
	bra.uni 	$Lt_3_5378;
$Lt_3_228354:
	@!%p133 bra 	$Lt_3_228866;
	.loc	18	61	0
	mov.f32 	%f886, 0f3f800000;   	// 1
	sub.f32 	%f887, %f886, %f880;
	selp.f32 	%f8, %f887, %f880, %p134;
	bra.uni 	$Lt_3_5378;
$Lt_3_228866:
	.loc	18	69	0
	mov.f32 	%f888, 0f3f800000;   	// 1
	sub.f32 	%f889, %f888, %f8;
	selp.f32 	%f8, %f889, %f8, %p134;
	.loc	18	73	0
	mul.f32 	%f8, %f8, %f880;
	bra.uni 	$Lt_3_5378;
$Lt_3_5634:
	.loc	19	267	0
	ld.const.f32 	%f890, [kRGB32f_To_601YPbPr+4];
	mul.f32 	%f891, %f890, %f6;
	ld.const.f32 	%f892, [kRGB32f_To_601YPbPr+0];
	mad.f32 	%f893, %f892, %f7, %f891;
	ld.const.f32 	%f894, [kRGB32f_To_601YPbPr+8];
	mad.f32 	%f895, %f894, %f5, %f893;
	ld.const.f32 	%f896, [k601YPbPr_To_RGB32f+12];
	mul.f32 	%f897, %f896, %f895;
	.loc	19	268	0
	ld.const.f32 	%f898, [k601YPbPr_To_RGB32f+0];
	mul.f32 	%f899, %f898, %f895;
	.loc	20	44	0
	ld.const.f32 	%f900, [k601YPbPr_To_RGB32f+24];
	mul.f32 	%f5, %f900, %f895;
	mov.f32 	%f6, %f897;
	mov.f32 	%f7, %f899;
	.loc	35	61	0
	bra.uni 	$Lt_3_5378;
$Lt_3_5890:
	.loc	35	63	0
	ld.param.f32 	%f901, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter2+20];
	ld.param.f32 	%f902, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter2+12];
	ld.param.f32 	%f903, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter2+4];
	mov.f32 	%f904, 0f00000000;   	// 0
	setp.neu.f32 	%p136, %f903, %f904;
	@!%p136 bra 	$Lt_3_229378;
	.loc	21	44	0
	ld.param.f32 	%f905, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter2+8];
	sub.f32 	%f906, %f7, %f902;
	mad.f32 	%f7, %f905, %f906, %f901;
	.loc	21	45	0
	sub.f32 	%f907, %f6, %f902;
	mad.f32 	%f6, %f905, %f907, %f901;
	.loc	21	46	0
	sub.f32 	%f908, %f5, %f902;
	mad.f32 	%f5, %f905, %f908, %f901;
	bra.uni 	$Lt_3_5378;
$Lt_3_229378:
	.loc	21	50	0
	ld.param.f32 	%f909, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter2+24];
	setp.gt.f32 	%p137, %f7, %f902;
	selp.f32 	%f7, %f909, %f901, %p137;
	.loc	21	51	0
	setp.gt.f32 	%p138, %f6, %f902;
	selp.f32 	%f6, %f909, %f901, %p138;
	.loc	21	52	0
	setp.gt.f32 	%p139, %f5, %f902;
	selp.f32 	%f5, %f909, %f901, %p139;
	bra.uni 	$Lt_3_5378;
$Lt_3_6146:
	.loc	22	47	0
	ld.param.f32 	%f910, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter2+4];
	mul.f32 	%f5, %f910, %f5;
	ld.param.f32 	%f911, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter2+8];
	mul.f32 	%f6, %f911, %f6;
	ld.param.f32 	%f912, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter2+12];
	mul.f32 	%f7, %f912, %f7;
	.loc	35	67	0
	bra.uni 	$Lt_3_5378;
$Lt_3_6402:
	.loc	23	48	0
	ld.param.f32 	%f913, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter2+4];
	mov.f32 	%f914, 0f00000000;   	// 0
	setp.eq.f32 	%p140, %f913, %f914;
	ld.param.f32 	%f915, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter2+8];
	mov.f32 	%f916, 0f00000000;   	// 0
	max.f32 	%f917, %f5, %f916;
	mov.f32 	%f918, 0f3f800000;   	// 1
	min.f32 	%f919, %f917, %f918;
	ld.param.f32 	%f920, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter2+12];
	sub.f32 	%f921, %f919, %f920;
	abs.f32 	%f922, %f921;
	mov.f32 	%f923, 0f00000000;   	// 0
	max.f32 	%f924, %f6, %f923;
	mov.f32 	%f925, 0f3f800000;   	// 1
	min.f32 	%f926, %f924, %f925;
	ld.param.f32 	%f927, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter2+16];
	sub.f32 	%f928, %f926, %f927;
	abs.f32 	%f929, %f928;
	mov.f32 	%f930, 0f00000000;   	// 0
	max.f32 	%f931, %f7, %f930;
	mov.f32 	%f932, 0f3f800000;   	// 1
	min.f32 	%f933, %f931, %f932;
	ld.param.f32 	%f934, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter2+20];
	sub.f32 	%f935, %f933, %f934;
	abs.f32 	%f936, %f935;
	mov.f32 	%f937, 0f00000000;   	// 0
	max.f32 	%f938, %f8, %f937;
	mov.f32 	%f939, 0f3f800000;   	// 1
	min.f32 	%f940, %f938, %f939;
	sub.f32 	%f941, %f940, %f8;
	abs.f32 	%f942, %f941;
	max.f32 	%f943, %f936, %f942;
	max.f32 	%f944, %f929, %f943;
	max.f32 	%f945, %f922, %f944;
	setp.ge.f32 	%p141, %f915, %f945;
	xor.pred 	%p142, %p140, %p141;
	@!%p142 bra 	$Lt_3_5378;
	.loc	19	267	0
	ld.const.f32 	%f946, [kRGB32f_To_601YPbPr+4];
	mul.f32 	%f891, %f946, %f6;
	ld.const.f32 	%f947, [kRGB32f_To_601YPbPr+0];
	mad.f32 	%f893, %f947, %f7, %f891;
	ld.const.f32 	%f948, [kRGB32f_To_601YPbPr+8];
	mad.f32 	%f895, %f948, %f5, %f893;
	ld.const.f32 	%f949, [k601YPbPr_To_RGB32f+12];
	mul.f32 	%f950, %f949, %f895;
	.loc	19	268	0
	ld.const.f32 	%f951, [k601YPbPr_To_RGB32f+0];
	mul.f32 	%f952, %f951, %f895;
	.loc	20	44	0
	ld.const.f32 	%f953, [k601YPbPr_To_RGB32f+24];
	mul.f32 	%f5, %f953, %f895;
	mov.f32 	%f6, %f950;
	mov.f32 	%f7, %f952;
	bra.uni 	$Lt_3_5378;
$Lt_3_6658:
	.loc	24	48	0
	ld.param.f32 	%f954, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter2+8];
	mov.f32 	%f955, 0f00000000;   	// 0
	max.f32 	%f956, %f5, %f955;
	mov.f32 	%f957, 0f3f800000;   	// 1
	min.f32 	%f958, %f956, %f957;
	ld.param.f32 	%f959, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter2+12];
	sub.f32 	%f960, %f958, %f959;
	abs.f32 	%f961, %f960;
	mov.f32 	%f962, 0f00000000;   	// 0
	max.f32 	%f963, %f6, %f962;
	mov.f32 	%f964, 0f3f800000;   	// 1
	min.f32 	%f965, %f963, %f964;
	ld.param.f32 	%f966, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter2+16];
	sub.f32 	%f967, %f965, %f966;
	abs.f32 	%f968, %f967;
	mov.f32 	%f969, 0f00000000;   	// 0
	max.f32 	%f970, %f7, %f969;
	mov.f32 	%f971, 0f3f800000;   	// 1
	min.f32 	%f972, %f970, %f971;
	ld.param.f32 	%f973, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter2+20];
	sub.f32 	%f974, %f972, %f973;
	abs.f32 	%f975, %f974;
	mov.f32 	%f976, 0f00000000;   	// 0
	max.f32 	%f977, %f8, %f976;
	mov.f32 	%f978, 0f3f800000;   	// 1
	min.f32 	%f979, %f977, %f978;
	sub.f32 	%f980, %f979, %f8;
	abs.f32 	%f981, %f980;
	max.f32 	%f982, %f975, %f981;
	max.f32 	%f983, %f968, %f982;
	max.f32 	%f984, %f961, %f983;
	setp.ge.f32 	%p143, %f954, %f984;
	@!%p143 bra 	$Lt_3_5378;
	.loc	24	51	0
	ld.param.f32 	%f985, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter2+24];
	mov.f32 	%f124, %f985;
	ld.param.f32 	%f986, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter2+28];
	mov.f32 	%f126, %f986;
	ld.param.f32 	%f987, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter2+32];
	mov.f32 	%f128, %f987;
	ld.param.f32 	%f988, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter2+4];
	mov.f32 	%f989, 0f00000000;   	// 0
	setp.eq.f32 	%p144, %f988, %f989;
	@!%p144 bra 	$Lt_3_230658;
	.loc	24	60	0
	ld.const.f32 	%f990, [kRGB32f_To_601YPbPr+4];
	mul.f32 	%f891, %f990, %f6;
	ld.const.f32 	%f991, [kRGB32f_To_601YPbPr+0];
	mad.f32 	%f893, %f991, %f7, %f891;
	ld.const.f32 	%f992, [kRGB32f_To_601YPbPr+8];
	mad.f32 	%f895, %f992, %f5, %f893;
	mul.f32 	%f124, %f895, %f985;
	.loc	24	61	0
	mul.f32 	%f126, %f895, %f986;
	.loc	24	62	0
	mul.f32 	%f128, %f895, %f987;
$Lt_3_230658:
	.loc	24	65	0
	mov.f32 	%f5, %f124;
	mov.f32 	%f6, %f126;
	mov.f32 	%f7, %f128;
	bra.uni 	$Lt_3_5378;
$Lt_3_6914:
	.loc	25	47	0
	sub.s32 	%r513, %r13, %r12;
	sub.s32 	%r514, %r14, %r6;
	cvt.rn.f32.s32 	%f993, %r6;
	cvt.rn.f32.s32 	%f994, %r12;
	cvt.rn.f32.s32 	%f136, %r513;
	cvt.rn.f32.s32 	%f137, %r514;
	ld.param.f32 	%f880, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter2+4];
	sub.f32 	%f995, %f880, %f993;
	sub.f32 	%f996, %f880, %f994;
	sub.f32 	%f997, %f880, %f136;
	sub.f32 	%f998, %f880, %f137;
	cvt.rzi.s32.f32 	%r515, %f995;
	cvt.rzi.s32.f32 	%r516, %f996;
	cvt.rzi.s32.f32 	%r517, %f997;
	cvt.rzi.s32.f32 	%r518, %f998;
	max.s32 	%r519, %r515, %r516;
	max.s32 	%r520, %r518, %r519;
	max.s32 	%r521, %r517, %r520;
	mov.u32 	%r522, 0;
	setp.le.s32 	%p145, %r521, %r522;
	@%p145 bra 	$Lt_3_5378;
	.loc	25	51	0
	cvt.rn.f32.s32 	%f999, %r521;
	sub.f32 	%f1000, %f880, %f999;
	div.full.f32 	%f1001, %f1000, %f880;
	mul.f32 	%f8, %f8, %f1001;
	bra.uni 	$Lt_3_5378;
$Lt_3_7170:
	.loc	27	50	0
	ld.const.f32 	%f1002, [kRGB32f_To_601YPbPr+4];
	mul.f32 	%f891, %f1002, %f6;
	ld.param.f32 	%f1003, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter2+12];
	ld.param.f32 	%f905, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter2+8];
	sub.f32 	%f1004, %f1003, %f905;
	ld.const.f32 	%f1005, [kRGB32f_To_601YPbPr+0];
	mad.f32 	%f893, %f1005, %f7, %f891;
	ld.param.f32 	%f1006, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter2+16];
	mul.f32 	%f1007, %f1006, %f1004;
	ld.const.f32 	%f1008, [kRGB32f_To_601YPbPr+8];
	mad.f32 	%f895, %f1008, %f5, %f893;
	add.f32 	%f1009, %f1007, %f1003;
	sub.f32 	%f1010, %f905, %f1007;
	mov.f32 	%f1011, 0f00000000;  	// 0
	max.f32 	%f1012, %f1009, %f1011;
	mov.f32 	%f1013, 0f00000000;  	// 0
	max.f32 	%f1014, %f1010, %f1013;
	mov.f32 	%f1015, 0f3f800000;  	// 1
	min.f32 	%f1016, %f1012, %f1015;
	mov.f32 	%f1017, 0f3f800000;  	// 1
	min.f32 	%f1018, %f1014, %f1017;
	set.gt.u32.f32 	%r523, %f1018, %f895;
	neg.s32 	%r524, %r523;
	set.le.u32.f32 	%r525, %f1016, %f895;
	neg.s32 	%r526, %r525;
	or.b32 	%r527, %r524, %r526;
	mov.u32 	%r528, 0;
	setp.eq.s32 	%p146, %r527, %r528;
	@%p146 bra 	$Lt_3_231938;
	mov.f32 	%f162, 0f00000000;   	// 0
	bra.uni 	$Lt_3_232706;
$Lt_3_231938:
	add.f32 	%f1019, %f1007, %f905;
	mov.f32 	%f1020, 0f00000000;  	// 0
	max.f32 	%f1021, %f1019, %f1020;
	mov.f32 	%f1022, 0f3f800000;  	// 1
	min.f32 	%f1023, %f1021, %f1022;
	set.le.u32.f32 	%r529, %f1023, %f895;
	neg.s32 	%r530, %r529;
	sub.f32 	%f1024, %f1003, %f1007;
	mov.f32 	%f1025, 0f00000000;  	// 0
	max.f32 	%f1026, %f1024, %f1025;
	mov.f32 	%f1027, 0f3f800000;  	// 1
	min.f32 	%f1028, %f1026, %f1027;
	set.lt.u32.f32 	%r531, %f895, %f1028;
	neg.s32 	%r532, %r531;
	and.b32 	%r533, %r530, %r532;
	mov.u32 	%r534, 0;
	setp.eq.s32 	%p147, %r533, %r534;
	@%p147 bra 	$Lt_3_232450;
	mov.f32 	%f162, 0f3f800000;   	// 1
	bra.uni 	$Lt_3_232706;
$Lt_3_232450:
	add.f32 	%f1029, %f1007, %f1007;
	setp.gt.f32 	%p148, %f1023, %f895;
	@!%p148 bra 	$Lt_3_232962;
	.loc	27	62	0
	sub.f32 	%f1030, %f895, %f1018;
	div.full.f32 	%f162, %f1030, %f1029;
	bra.uni 	$Lt_3_232706;
$Lt_3_232962:
	.loc	27	66	0
	sub.f32 	%f1031, %f1016, %f895;
	div.full.f32 	%f162, %f1031, %f1029;
$Lt_3_232706:
$Lt_3_232194:
$Lt_3_231682:
	.loc	27	69	0
	mov.f32 	%f1032, 0f3f800000;  	// 1
	sub.f32 	%f1033, %f1032, %f162;
	ld.param.f32 	%f1034, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter2+4];
	mov.f32 	%f1035, 0f00000000;  	// 0
	setp.neu.f32 	%p149, %f1034, %f1035;
	selp.f32 	%f162, %f1033, %f162, %p149;
	.loc	27	77	0
	ld.const.f32 	%f1036, [k601YPbPr_To_RGB32f+24];
	mul.f32 	%f5, %f1036, %f162;
	ld.const.f32 	%f1037, [k601YPbPr_To_RGB32f+12];
	mul.f32 	%f6, %f1037, %f162;
	ld.const.f32 	%f1038, [k601YPbPr_To_RGB32f+0];
	mul.f32 	%f7, %f1038, %f162;
	.loc	35	79	0
	bra.uni 	$Lt_3_5378;
$Lt_3_7426:
	.loc	35	80	0
	ld.param.f32 	%f1039, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter2+44];
	mov.f32 	%f1040, 0f00000000;  	// 0
	setp.ge.f32 	%p150, %f1039, %f1040;
	@!%p150 bra 	$L_3_201986;
	cvt.rn.f32.s32 	%f1041, %r6;
	cvt.rn.f32.s32 	%f1042, %r14;
	mul.f32 	%f1043, %f1042, %f1039;
	setp.lt.f32 	%p151, %f1041, %f1043;
	@%p151 bra 	$L_3_201730;
$L_3_201986:
	mov.f32 	%f1044, 0f00000000;  	// 0
	setp.lt.f32 	%p152, %f1039, %f1044;
	@!%p152 bra 	$Lt_3_5378;
	cvt.rn.f32.s32 	%f1045, %r12;
	cvt.rn.f32.s32 	%f1046, %r13;
	mul.f32 	%f1047, %f1046, %f1039;
	neg.f32 	%f1048, %f1047;
	setp.lt.f32 	%p153, %f1045, %f1048;
	@!%p153 bra 	$Lt_3_5378;
$L_3_201730:
	.loc	28	47	0
	ld.const.f32 	%f1049, [kRGB32f_To_601YPbPr+4];
	mul.f32 	%f1050, %f1049, %f6;
	ld.const.f32 	%f1051, [kRGB32f_To_601YPbPr+0];
	mad.f32 	%f1052, %f1051, %f7, %f1050;
	ld.const.f32 	%f1053, [kRGB32f_To_601YPbPr+8];
	mad.f32 	%f895, %f1053, %f5, %f1052;
	mov.f32 	%f198, %f895;
	ld.param.f32 	%f905, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter2+8];
	ld.param.f32 	%f1054, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter2+4];
	sub.f32 	%f1055, %f1054, %f905;
	ld.param.f32 	%f1056, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter2+16];
	ld.param.f32 	%f1057, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter2+12];
	sub.f32 	%f1058, %f1057, %f1056;
	mov.f32 	%f1059, 0f3f800000;  	// 1
	set.neu.u32.f32 	%r535, %f1058, %f1059;
	neg.s32 	%r536, %r535;
	mov.f32 	%f1060, 0f3f800000;  	// 1
	set.neu.u32.f32 	%r537, %f1055, %f1060;
	neg.s32 	%r538, %r537;
	or.b32 	%r539, %r536, %r538;
	mov.u32 	%r540, 0;
	setp.eq.s32 	%p154, %r539, %r540;
	@%p154 bra 	$Lt_3_233218;
	.loc	17	116	0
	mov.u32 	%r541, 1;
	sub.u32 	%r71, %r541, %r5;
	shr.u32 	%r72, %r11, 13;
	sub.u32 	%r73, %r5, %r11;
	sub.u32 	%r74, %r71, %r11;
	xor.b32 	%r75, %r74, %r72;
	shl.b32 	%r76, %r75, 8;
	sub.u32 	%r77, %r73, %r75;
	sub.u32 	%r78, %r11, %r75;
	xor.b32 	%r79, %r77, %r76;
	shr.u32 	%r80, %r79, 13;
	sub.u32 	%r81, %r78, %r79;
	xor.b32 	%r82, %r81, %r80;
	shr.u32 	%r542, %r82, 12;
	sub.u32 	%r543, %r75, %r79;
	sub.u32 	%r544, %r543, %r82;
	xor.b32 	%r86, %r542, %r544;
	.loc	17	117	0
	sub.u32 	%r545, %r79, %r82;
	sub.u32 	%r88, %r545, %r86;
	shl.b32 	%r546, %r86, 16;
	xor.b32 	%r90, %r88, %r546;
	.loc	17	118	0
	sub.u32 	%r547, %r82, %r86;
	sub.u32 	%r92, %r547, %r90;
	shr.u32 	%r548, %r90, 5;
	xor.b32 	%r94, %r92, %r548;
	.loc	17	119	0
	sub.u32 	%r95, %r86, %r90;
	sub.u32 	%r96, %r95, %r94;
	shr.u32 	%r549, %r94, 3;
	xor.b32 	%r98, %r96, %r549;
	.loc	17	120	0
	sub.u32 	%r99, %r90, %r94;
	sub.u32 	%r100, %r99, %r98;
	shl.b32 	%r550, %r98, 10;
	xor.b32 	%r102, %r100, %r550;
	.loc	17	121	0
	sub.u32 	%r103, %r94, %r98;
	sub.u32 	%r104, %r103, %r102;
	shr.u32 	%r551, %r102, 15;
	xor.b32 	%r106, %r104, %r551;
	.loc	28	57	0
	mov.f32 	%f1061, 0f3b270d73;  	// 0.00254902
	mul.lo.u32 	%r552, %r106, 1103515245;
	add.u32 	%r553, %r552, 12345;
	shr.u32 	%r554, %r553, 16;
	and.b32 	%r555, %r554, 255;
	shl.b32 	%r556, %r555, 7;
	mul.lo.u32 	%r557, %r106, -1029531031;
	sub.u32 	%r558, %r557, 740551042;
	shr.u32 	%r559, %r558, 16;
	and.b32 	%r560, %r559, 255;
	xor.b32 	%r561, %r556, %r560;
	cvt.rn.f32.s32 	%f1062, %r561;
	mov.f32 	%f1063, 0f467ffe00;  	// 16383.5
	div.full.f32 	%f1064, %f1062, %f1063;
	mov.f32 	%f1065, 0fbf800000;  	// -1
	add.f32 	%f1066, %f1064, %f1065;
	mad.f32 	%f198, %f1061, %f1066, %f895;
$Lt_3_233218:
	sub.f32 	%f1067, %f198, %f905;
	ld.param.f32 	%f1068, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter2+20];
	mov.f32 	%f1069, 0f3f800000;  	// 1
	setp.neu.f32 	%p155, %f1068, %f1069;
	@!%p155 bra 	$Lt_3_233986;
	mov.f32 	%f1070, 0f00000000;  	// 0
	setp.lt.f32 	%p156, %f1067, %f1070;
	@!%p156 bra 	$Lt_3_234498;
	.loc	28	66	0
	mov.f32 	%f198, %f1056;
	bra.uni 	$Lt_3_233730;
$Lt_3_234498:
	.loc	28	70	0
	rcp.approx.f32 	%f1071, %f1055;
	mul.f32 	%f1072, %f1067, %f1071;
	lg2.approx.f32 	%f1073, %f1072;
	mul.f32 	%f1074, %f1068, %f1073;
	ex2.approx.f32 	%f1075, %f1074;
	mad.f32 	%f198, %f1058, %f1075, %f1056;
	bra.uni 	$Lt_3_233730;
$Lt_3_233986:
	.loc	28	77	0
	rcp.approx.f32 	%f1076, %f1055;
	mul.f32 	%f1077, %f1067, %f1076;
	mad.f32 	%f198, %f1058, %f1077, %f1056;
$Lt_3_233730:
	.loc	19	267	0
	ld.const.f32 	%f1078, [kRGB32f_To_601YPbPr+28];
	mul.f32 	%f1079, %f1078, %f6;
	ld.const.f32 	%f1080, [kRGB32f_To_601YPbPr+16];
	mul.f32 	%f1081, %f1080, %f6;
	ld.param.f32 	%f1082, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter2+28];
	sin.approx.f32 	%f1083, %f1082;
	cos.approx.f32 	%f1084, %f1082;
	ld.const.f32 	%f1085, [kRGB32f_To_601YPbPr+24];
	mad.f32 	%f1086, %f1085, %f7, %f1079;
	ld.const.f32 	%f1087, [kRGB32f_To_601YPbPr+12];
	mad.f32 	%f1088, %f1087, %f7, %f1081;
	ld.const.f32 	%f1089, [kRGB32f_To_601YPbPr+32];
	mad.f32 	%f1090, %f1089, %f5, %f1086;
	ld.const.f32 	%f1091, [kRGB32f_To_601YPbPr+20];
	mad.f32 	%f1092, %f1091, %f5, %f1088;
	mul.f32 	%f1093, %f1083, %f1090;
	mul.f32 	%f1094, %f1084, %f1090;
	mul.f32 	%f1095, %f1092, %f1084;
	sub.f32 	%f1096, %f1095, %f1093;
	mad.f32 	%f1097, %f1092, %f1083, %f1094;
	ld.param.f32 	%f1098, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter2+32];
	sub.f32 	%f1099, %f1098, %f1096;
	ld.param.f32 	%f1100, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter2+36];
	sub.f32 	%f1101, %f1100, %f1097;
	ld.param.f32 	%f1102, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter2+40];
	mad.f32 	%f1103, %f1102, %f1099, %f1096;
	mad.f32 	%f1104, %f1102, %f1101, %f1097;
	ld.param.f32 	%f1105, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter2+24];
	mul.f32 	%f1106, %f1103, %f1105;
	mul.f32 	%f1107, %f1104, %f1105;
	ld.const.f32 	%f1108, [k601YPbPr_To_RGB32f+16];
	mul.f32 	%f1109, %f1108, %f1106;
	ld.const.f32 	%f1110, [k601YPbPr_To_RGB32f+12];
	mad.f32 	%f1111, %f1110, %f198, %f1109;
	ld.const.f32 	%f1112, [k601YPbPr_To_RGB32f+20];
	mad.f32 	%f258, %f1112, %f1107, %f1111;
	.loc	19	268	0
	ld.const.f32 	%f1113, [k601YPbPr_To_RGB32f+4];
	mul.f32 	%f1114, %f1113, %f1106;
	ld.const.f32 	%f1115, [k601YPbPr_To_RGB32f+0];
	mad.f32 	%f1116, %f1115, %f198, %f1114;
	ld.const.f32 	%f1117, [k601YPbPr_To_RGB32f+8];
	mad.f32 	%f264, %f1117, %f1107, %f1116;
	.loc	28	92	0
	ld.const.f32 	%f1118, [k601YPbPr_To_RGB32f+28];
	mul.f32 	%f1119, %f1118, %f1106;
	ld.const.f32 	%f1120, [k601YPbPr_To_RGB32f+24];
	mad.f32 	%f1121, %f1120, %f198, %f1119;
	ld.const.f32 	%f1122, [k601YPbPr_To_RGB32f+32];
	mad.f32 	%f5, %f1122, %f1107, %f1121;
	mov.f32 	%f6, %f258;
	mov.f32 	%f7, %f264;
	bra.uni 	$Lt_3_5378;
$Lt_3_7682:
	.loc	35	83	0
	ld.param.f32 	%f1123, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter2+44];
	mov.f32 	%f1124, 0f00000000;  	// 0
	setp.ge.f32 	%p157, %f1123, %f1124;
	@!%p157 bra 	$L_3_202754;
	cvt.rn.f32.s32 	%f1125, %r6;
	cvt.rn.f32.s32 	%f1126, %r14;
	mul.f32 	%f1127, %f1126, %f1123;
	setp.lt.f32 	%p158, %f1125, %f1127;
	@%p158 bra 	$L_3_202498;
$L_3_202754:
	mov.f32 	%f1128, 0f00000000;  	// 0
	setp.lt.f32 	%p159, %f1123, %f1128;
	@!%p159 bra 	$Lt_3_5378;
	cvt.rn.f32.s32 	%f1129, %r12;
	cvt.rn.f32.s32 	%f1130, %r13;
	mul.f32 	%f1131, %f1130, %f1123;
	neg.f32 	%f1132, %f1131;
	setp.lt.f32 	%p160, %f1129, %f1132;
	@!%p160 bra 	$Lt_3_5378;
$L_3_202498:
	.loc	28	110	0
	ld.const.f32 	%f1133, [kRGB32f_To_601YPbPr+4];
	mul.f32 	%f891, %f1133, %f6;
	ld.param.f32 	%f905, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter2+8];
	ld.param.f32 	%f1134, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter2+4];
	sub.f32 	%f1135, %f1134, %f905;
	ld.param.f32 	%f1136, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter2+16];
	ld.param.f32 	%f1137, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter2+12];
	sub.f32 	%f1138, %f1137, %f1136;
	ld.const.f32 	%f1139, [kRGB32f_To_601YPbPr+0];
	mad.f32 	%f893, %f1139, %f7, %f891;
	ld.const.f32 	%f1140, [kRGB32f_To_601YPbPr+8];
	mad.f32 	%f895, %f1140, %f5, %f893;
	sub.f32 	%f1141, %f895, %f905;
	ld.param.f32 	%f1142, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter2+20];
	mov.f32 	%f1143, 0f3f800000;  	// 1
	setp.neu.f32 	%p161, %f1142, %f1143;
	@!%p161 bra 	$Lt_3_235010;
	.loc	28	120	0
	mov.f32 	%f1144, 0f00000000;  	// 0
	max.f32 	%f1145, %f1141, %f1144;
	div.full.f32 	%f1146, %f1145, %f1135;
	lg2.approx.f32 	%f1147, %f1146;
	mul.f32 	%f1148, %f1142, %f1147;
	ex2.approx.f32 	%f1149, %f1148;
	mad.f32 	%f297, %f1138, %f1149, %f1136;
	bra.uni 	$Lt_3_234754;
$Lt_3_235010:
	.loc	28	129	0
	div.full.f32 	%f1150, %f1141, %f1135;
	mad.f32 	%f297, %f1138, %f1150, %f1136;
$Lt_3_234754:
	.loc	28	135	0
	ld.const.f32 	%f1151, [k601YPbPr_To_RGB32f+24];
	mul.f32 	%f5, %f1151, %f297;
	ld.const.f32 	%f1152, [k601YPbPr_To_RGB32f+12];
	mul.f32 	%f6, %f1152, %f297;
	ld.const.f32 	%f1153, [k601YPbPr_To_RGB32f+0];
	mul.f32 	%f7, %f1153, %f297;
	bra.uni 	$Lt_3_5378;
$Lt_3_7938:
	.loc	35	86	0
	ld.param.f32 	%f1154, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter2+44];
	mov.f32 	%f1155, 0f00000000;  	// 0
	setp.ge.f32 	%p162, %f1154, %f1155;
	@!%p162 bra 	$L_3_203522;
	cvt.rn.f32.s32 	%f1156, %r6;
	cvt.rn.f32.s32 	%f1157, %r14;
	mul.f32 	%f1158, %f1157, %f1154;
	setp.lt.f32 	%p163, %f1156, %f1158;
	@%p163 bra 	$L_3_203266;
$L_3_203522:
	mov.f32 	%f1159, 0f00000000;  	// 0
	setp.lt.f32 	%p164, %f1154, %f1159;
	@!%p164 bra 	$Lt_3_5378;
	cvt.rn.f32.s32 	%f1160, %r12;
	cvt.rn.f32.s32 	%f1161, %r13;
	mul.f32 	%f1162, %f1161, %f1154;
	neg.f32 	%f1163, %f1162;
	setp.lt.f32 	%p165, %f1160, %f1163;
	@!%p165 bra 	$Lt_3_5378;
$L_3_203266:
	.loc	28	160	0
	ld.const.f32 	%f5, [k601YPbPr_To_RGB32f+24];
	ld.const.f32 	%f6, [k601YPbPr_To_RGB32f+12];
	ld.const.f32 	%f7, [k601YPbPr_To_RGB32f+0];
	bra.uni 	$Lt_3_5378;
$Lt_3_8194:
	.loc	29	42	0
	ld.param.f32 	%f880, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter2+4];
	mov.f32 	%f1164, 0f00000000;  	// 0
	setp.ge.f32 	%p166, %f5, %f1164;
	@!%p166 bra 	$Lt_3_235522;
	.loc	19	292	0
	lg2.approx.f32 	%f1165, %f5;
	mul.f32 	%f1166, %f880, %f1165;
	ex2.approx.f32 	%f1167, %f1166;
	bra.uni 	$Lt_3_235266;
$Lt_3_235522:
	neg.f32 	%f1168, %f5;
	lg2.approx.f32 	%f1169, %f1168;
	mul.f32 	%f1170, %f880, %f1169;
	ex2.approx.f32 	%f1171, %f1170;
	neg.f32 	%f1167, %f1171;
$Lt_3_235266:
	mov.f32 	%f1172, 0f00000000;  	// 0
	setp.ge.f32 	%p167, %f6, %f1172;
	@!%p167 bra 	$Lt_3_236034;
	.loc	19	293	0
	lg2.approx.f32 	%f1173, %f6;
	mul.f32 	%f1174, %f880, %f1173;
	ex2.approx.f32 	%f1175, %f1174;
	bra.uni 	$Lt_3_235778;
$Lt_3_236034:
	neg.f32 	%f1176, %f6;
	lg2.approx.f32 	%f1177, %f1176;
	mul.f32 	%f1178, %f880, %f1177;
	ex2.approx.f32 	%f1179, %f1178;
	neg.f32 	%f1175, %f1179;
$Lt_3_235778:
	mov.f32 	%f1180, 0f00000000;  	// 0
	setp.ge.f32 	%p168, %f7, %f1180;
	@!%p168 bra 	$Lt_3_236546;
	.loc	19	294	0
	lg2.approx.f32 	%f1181, %f7;
	mul.f32 	%f1182, %f880, %f1181;
	ex2.approx.f32 	%f1183, %f1182;
	bra.uni 	$Lt_3_236290;
$Lt_3_236546:
	neg.f32 	%f1184, %f7;
	lg2.approx.f32 	%f1185, %f1184;
	mul.f32 	%f1186, %f880, %f1185;
	ex2.approx.f32 	%f1187, %f1186;
	neg.f32 	%f1183, %f1187;
$Lt_3_236290:
	mov.f32 	%f1188, 0f00000000;  	// 0
	setp.ge.f32 	%p169, %f8, %f1188;
	@!%p169 bra 	$Lt_3_237058;
	.loc	19	295	0
	lg2.approx.f32 	%f1189, %f8;
	mul.f32 	%f1190, %f880, %f1189;
	ex2.approx.f32 	%f1191, %f1190;
	bra.uni 	$Lt_3_236802;
$Lt_3_237058:
	neg.f32 	%f1192, %f8;
	lg2.approx.f32 	%f1193, %f1192;
	mul.f32 	%f1194, %f880, %f1193;
	ex2.approx.f32 	%f1195, %f1194;
	neg.f32 	%f1191, %f1195;
$Lt_3_236802:
	.loc	29	42	0
	mov.f32 	%f5, %f1167;
	mov.f32 	%f6, %f1175;
	mov.f32 	%f7, %f1183;
	mov.f32 	%f8, %f1191;
	.loc	35	91	0
	bra.uni 	$Lt_3_5378;
$Lt_3_8450:
	.loc	30	41	0
	sub.s32 	%r562, %r14, %r6;
	sub.s32 	%r6, %r562, 1;
	.loc	35	94	0
	bra.uni 	$Lt_3_5378;
$Lt_3_8706:
	.loc	35	96	0
	sub.u32 	%r343, %r6, %r12;
	ld.param.f32 	%f880, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter2+4];
	mov.f32 	%f1196, 0f3e800000;  	// 0.25
	mul.f32 	%f1197, %f880, %f1196;
	ld.param.f32 	%f1198, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter2+16];
	ld.param.f32 	%f1199, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter2+8];
	mov.f32 	%f1200, 0f3f800000;  	// 1
	setp.eq.f32 	%p170, %f1199, %f1200;
	@!%p170 bra 	$Lt_3_237570;
	.loc	17	117	0
	cvt.rzi.u32.f32 	%r563, %f1198;
	shr.u32 	%r564, %r563, 13;
	sub.u32 	%r565, %r343, %r563;
	sub.u32 	%r566, %r12, %r563;
	xor.b32 	%r567, %r565, %r564;
	shl.b32 	%r568, %r567, 8;
	sub.u32 	%r569, %r566, %r567;
	sub.u32 	%r570, %r563, %r567;
	xor.b32 	%r571, %r569, %r568;
	shr.u32 	%r572, %r571, 13;
	sub.u32 	%r573, %r570, %r571;
	sub.u32 	%r574, %r567, %r571;
	xor.b32 	%r575, %r573, %r572;
	shr.u32 	%r576, %r575, 12;
	sub.u32 	%r577, %r574, %r575;
	xor.b32 	%r578, %r577, %r576;
	shl.b32 	%r579, %r578, 16;
	sub.u32 	%r580, %r571, %r575;
	sub.u32 	%r581, %r580, %r578;
	xor.b32 	%r582, %r579, %r581;
	.loc	17	118	0
	sub.u32 	%r583, %r575, %r578;
	sub.u32 	%r584, %r583, %r582;
	shr.u32 	%r585, %r582, 5;
	xor.b32 	%r586, %r584, %r585;
	.loc	17	119	0
	sub.u32 	%r587, %r578, %r582;
	sub.u32 	%r588, %r587, %r586;
	shr.u32 	%r589, %r586, 3;
	xor.b32 	%r590, %r588, %r589;
	.loc	17	120	0
	sub.u32 	%r591, %r582, %r586;
	sub.u32 	%r592, %r591, %r590;
	shl.b32 	%r593, %r590, 10;
	xor.b32 	%r594, %r592, %r593;
	.loc	17	121	0
	sub.u32 	%r595, %r586, %r590;
	sub.u32 	%r596, %r595, %r594;
	shr.u32 	%r597, %r594, 15;
	xor.b32 	%r598, %r596, %r597;
	.loc	31	48	0
	mul.lo.u32 	%r599, %r598, 1103515245;
	add.u32 	%r600, %r599, 12345;
	shr.u32 	%r601, %r600, 16;
	and.b32 	%r602, %r601, 255;
	shl.b32 	%r603, %r602, 7;
	mul.lo.u32 	%r604, %r598, -1029531031;
	sub.u32 	%r605, %r604, 740551042;
	shr.u32 	%r606, %r605, 16;
	and.b32 	%r607, %r606, 255;
	xor.b32 	%r608, %r603, %r607;
	cvt.rn.f32.s32 	%f1201, %r608;
	mov.f32 	%f1202, 0f46fffe00;  	// 32767
	div.full.f32 	%f1203, %f1201, %f1202;
	mul.f32 	%f1204, %f880, %f1203;
	mov.f32 	%f1205, 0f3f000000;  	// 0.5
	mul.f32 	%f1206, %f1204, %f1205;
	sub.f32 	%f1207, %f1206, %f1197;
	.loc	17	117	0
	add.u32 	%r609, %r565, 1;
	xor.b32 	%r610, %r609, %r564;
	shl.b32 	%r611, %r610, 8;
	sub.u32 	%r612, %r566, %r610;
	sub.u32 	%r613, %r563, %r610;
	xor.b32 	%r614, %r611, %r612;
	shr.u32 	%r615, %r614, 13;
	sub.u32 	%r616, %r613, %r614;
	sub.u32 	%r617, %r610, %r614;
	xor.b32 	%r618, %r615, %r616;
	shr.u32 	%r619, %r618, 12;
	sub.u32 	%r620, %r617, %r618;
	xor.b32 	%r621, %r619, %r620;
	sub.u32 	%r622, %r614, %r618;
	sub.u32 	%r178, %r622, %r621;
	shl.b32 	%r623, %r621, 16;
	xor.b32 	%r178, %r178, %r623;
	.loc	17	118	0
	sub.u32 	%r180, %r618, %r621;
	sub.u32 	%r180, %r180, %r178;
	shr.u32 	%r624, %r178, 5;
	xor.b32 	%r180, %r180, %r624;
	.loc	17	119	0
	sub.u32 	%r182, %r621, %r178;
	sub.u32 	%r182, %r182, %r180;
	shr.u32 	%r625, %r180, 3;
	xor.b32 	%r182, %r182, %r625;
	.loc	17	120	0
	sub.u32 	%r178, %r178, %r180;
	sub.u32 	%r178, %r178, %r182;
	shl.b32 	%r626, %r182, 10;
	xor.b32 	%r178, %r178, %r626;
	.loc	17	121	0
	sub.u32 	%r180, %r180, %r182;
	sub.u32 	%r180, %r180, %r178;
	shr.u32 	%r627, %r178, 15;
	xor.b32 	%r180, %r180, %r627;
	.loc	31	49	0
	mul.lo.u32 	%r628, %r180, 1103515245;
	add.u32 	%r629, %r628, 12345;
	shr.u32 	%r630, %r629, 16;
	and.b32 	%r631, %r630, 255;
	shl.b32 	%r632, %r631, 7;
	mul.lo.u32 	%r633, %r180, -1029531031;
	sub.u32 	%r634, %r633, 740551042;
	shr.u32 	%r635, %r634, 16;
	and.b32 	%r636, %r635, 255;
	xor.b32 	%r637, %r632, %r636;
	cvt.rn.f32.s32 	%f1208, %r637;
	mov.f32 	%f1209, 0f46fffe00;  	// 32767
	div.full.f32 	%f1210, %f1208, %f1209;
	mul.f32 	%f1211, %f880, %f1210;
	mov.f32 	%f1212, 0f3f000000;  	// 0.5
	mul.f32 	%f1213, %f1211, %f1212;
	sub.f32 	%f362, %f1213, %f1197;
	.loc	17	117	0
	add.u32 	%r638, %r565, 2;
	xor.b32 	%r639, %r638, %r564;
	shl.b32 	%r640, %r639, 8;
	sub.u32 	%r641, %r566, %r639;
	sub.u32 	%r642, %r563, %r639;
	xor.b32 	%r643, %r640, %r641;
	shr.u32 	%r644, %r643, 13;
	sub.u32 	%r645, %r642, %r643;
	sub.u32 	%r646, %r639, %r643;
	xor.b32 	%r647, %r644, %r645;
	shr.u32 	%r648, %r647, 12;
	sub.u32 	%r649, %r646, %r647;
	xor.b32 	%r650, %r648, %r649;
	sub.u32 	%r651, %r643, %r647;
	sub.u32 	%r210, %r651, %r650;
	shl.b32 	%r652, %r650, 16;
	xor.b32 	%r210, %r210, %r652;
	.loc	17	118	0
	sub.u32 	%r212, %r647, %r650;
	sub.u32 	%r212, %r212, %r210;
	shr.u32 	%r653, %r210, 5;
	xor.b32 	%r212, %r212, %r653;
	.loc	17	119	0
	sub.u32 	%r214, %r650, %r210;
	sub.u32 	%r214, %r214, %r212;
	shr.u32 	%r654, %r212, 3;
	xor.b32 	%r214, %r214, %r654;
	.loc	17	120	0
	sub.u32 	%r210, %r210, %r212;
	sub.u32 	%r210, %r210, %r214;
	shl.b32 	%r655, %r214, 10;
	xor.b32 	%r210, %r210, %r655;
	.loc	17	121	0
	sub.u32 	%r212, %r212, %r214;
	sub.u32 	%r212, %r212, %r210;
	shr.u32 	%r656, %r210, 15;
	xor.b32 	%r212, %r212, %r656;
	.loc	31	50	0
	mul.lo.u32 	%r657, %r212, 1103515245;
	add.u32 	%r658, %r657, 12345;
	shr.u32 	%r659, %r658, 16;
	and.b32 	%r660, %r659, 255;
	shl.b32 	%r661, %r660, 7;
	mul.lo.u32 	%r662, %r212, -1029531031;
	sub.u32 	%r663, %r662, 740551042;
	shr.u32 	%r664, %r663, 16;
	and.b32 	%r665, %r664, 255;
	xor.b32 	%r666, %r661, %r665;
	cvt.rn.f32.s32 	%f1214, %r666;
	mov.f32 	%f1215, 0f46fffe00;  	// 32767
	div.full.f32 	%f1216, %f1214, %f1215;
	mul.f32 	%f1217, %f880, %f1216;
	mov.f32 	%f1218, 0f3f000000;  	// 0.5
	mul.f32 	%f1219, %f1217, %f1218;
	sub.f32 	%f369, %f1219, %f1197;
	bra.uni 	$Lt_3_237314;
$Lt_3_237570:
	.loc	17	117	0
	cvt.rzi.u32.f32 	%r667, %f1198;
	shr.u32 	%r668, %r667, 13;
	sub.u32 	%r669, %r343, %r667;
	sub.u32 	%r670, %r12, %r667;
	xor.b32 	%r671, %r669, %r668;
	shl.b32 	%r672, %r671, 8;
	sub.u32 	%r673, %r670, %r671;
	sub.u32 	%r674, %r667, %r671;
	xor.b32 	%r675, %r673, %r672;
	shr.u32 	%r676, %r675, 13;
	sub.u32 	%r677, %r674, %r675;
	sub.u32 	%r678, %r671, %r675;
	xor.b32 	%r679, %r677, %r676;
	shr.u32 	%r680, %r679, 12;
	sub.u32 	%r681, %r678, %r679;
	xor.b32 	%r682, %r681, %r680;
	shl.b32 	%r683, %r682, 16;
	sub.u32 	%r684, %r675, %r679;
	sub.u32 	%r685, %r684, %r682;
	xor.b32 	%r686, %r683, %r685;
	.loc	17	118	0
	sub.u32 	%r687, %r679, %r682;
	sub.u32 	%r688, %r687, %r686;
	shr.u32 	%r689, %r686, 5;
	xor.b32 	%r690, %r688, %r689;
	.loc	17	119	0
	sub.u32 	%r691, %r682, %r686;
	sub.u32 	%r692, %r691, %r690;
	shr.u32 	%r693, %r690, 3;
	xor.b32 	%r694, %r692, %r693;
	.loc	17	120	0
	sub.u32 	%r695, %r686, %r690;
	sub.u32 	%r696, %r695, %r694;
	shl.b32 	%r697, %r694, 10;
	xor.b32 	%r698, %r696, %r697;
	.loc	17	121	0
	sub.u32 	%r699, %r690, %r694;
	sub.u32 	%r700, %r699, %r698;
	shr.u32 	%r701, %r698, 15;
	xor.b32 	%r702, %r700, %r701;
	.loc	31	54	0
	mul.lo.u32 	%r703, %r702, 1103515245;
	mul.lo.u32 	%r704, %r702, -1029531031;
	add.u32 	%r705, %r703, 12345;
	sub.u32 	%r706, %r704, 740551042;
	shr.u32 	%r707, %r705, 16;
	shr.u32 	%r708, %r706, 16;
	and.b32 	%r709, %r707, 255;
	and.b32 	%r710, %r708, 255;
	shl.b32 	%r711, %r709, 7;
	xor.b32 	%r712, %r711, %r710;
	cvt.rn.f32.s32 	%f1220, %r712;
	mov.f32 	%f1221, 0f46fffe00;  	// 32767
	div.full.f32 	%f1222, %f1220, %f1221;
	mul.f32 	%f1223, %f1222, %f880;
	mov.f32 	%f1224, 0f3f000000;  	// 0.5
	mul.f32 	%f1225, %f1223, %f1224;
	sub.f32 	%f1207, %f1225, %f1197;
	mov.f32 	%f369, %f1207;
	mov.f32 	%f362, %f1207;
$Lt_3_237314:
	.loc	31	57	0
	add.f32 	%f7, %f1207, %f7;
	.loc	31	58	0
	add.f32 	%f6, %f362, %f6;
	.loc	31	59	0
	add.f32 	%f5, %f5, %f369;
	ld.param.f32 	%f1226, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter2+12];
	mov.f32 	%f1227, 0f3f800000;  	// 1
	setp.eq.f32 	%p171, %f1226, %f1227;
	@!%p171 bra 	$Lt_3_5378;
	.loc	31	63	0
	mov.f32 	%f1228, 0f00000000;  	// 0
	max.f32 	%f1229, %f5, %f1228;
	mov.f32 	%f1230, 0f3f800000;  	// 1
	min.f32 	%f5, %f1229, %f1230;
	mov.f32 	%f1231, 0f00000000;  	// 0
	max.f32 	%f1232, %f6, %f1231;
	mov.f32 	%f1233, 0f3f800000;  	// 1
	min.f32 	%f6, %f1232, %f1233;
	mov.f32 	%f1234, 0f00000000;  	// 0
	max.f32 	%f1235, %f7, %f1234;
	mov.f32 	%f1236, 0f3f800000;  	// 1
	min.f32 	%f7, %f1235, %f1236;
	mov.f32 	%f1237, 0f00000000;  	// 0
	max.f32 	%f1238, %f8, %f1237;
	mov.f32 	%f1239, 0f3f800000;  	// 1
	min.f32 	%f8, %f1238, %f1239;
	bra.uni 	$Lt_3_5378;
$Lt_3_8962:
	.loc	35	99	0
	cvt.rn.f32.s32 	%f1240, %r6;
	cvt.rn.f32.s32 	%f1241, %r14;
	ld.param.f32 	%f1242, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter2+20];
	mul.f32 	%f1243, %f1241, %f1242;
	setp.lt.f32 	%p172, %f1240, %f1243;
	@!%p172 bra 	$Lt_3_5378;
	.loc	19	267	0
	ld.const.f32 	%f1244, [kRGB32f_To_601YPbPr+4];
	mul.f32 	%f891, %f1244, %f6;
	ld.const.f32 	%f1245, [kRGB32f_To_601YPbPr+28];
	mul.f32 	%f1246, %f1245, %f6;
	ld.const.f32 	%f1247, [kRGB32f_To_601YPbPr+16];
	mul.f32 	%f1248, %f1247, %f6;
	ld.const.f32 	%f1249, [kRGB32f_To_601YPbPr+0];
	mad.f32 	%f893, %f1249, %f7, %f891;
	ld.const.f32 	%f1250, [kRGB32f_To_601YPbPr+24];
	mad.f32 	%f1251, %f1250, %f7, %f1246;
	ld.const.f32 	%f1252, [kRGB32f_To_601YPbPr+12];
	mad.f32 	%f1253, %f1252, %f7, %f1248;
	ld.const.f32 	%f1254, [kRGB32f_To_601YPbPr+8];
	mad.f32 	%f895, %f1254, %f5, %f893;
	ld.const.f32 	%f1255, [kRGB32f_To_601YPbPr+32];
	mad.f32 	%f1256, %f1255, %f5, %f1251;
	ld.const.f32 	%f1257, [kRGB32f_To_601YPbPr+20];
	mad.f32 	%f1258, %f1257, %f5, %f1253;
	ld.param.f32 	%f1259, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter2+4];
	ld.param.f32 	%f1260, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter2+8];
	mad.f32 	%f1261, %f895, %f1260, %f1259;
	ld.param.f32 	%f1262, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter2+16];
	mul.f32 	%f1263, %f1256, %f1262;
	ld.param.f32 	%f1264, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter2+12];
	mul.f32 	%f1265, %f1256, %f1264;
	mul.f32 	%f1266, %f1258, %f1264;
	sub.f32 	%f1267, %f1266, %f1263;
	mad.f32 	%f1268, %f1258, %f1262, %f1265;
	ld.const.f32 	%f1269, [k601YPbPr_To_RGB32f+16];
	mul.f32 	%f1270, %f1269, %f1267;
	ld.const.f32 	%f1271, [k601YPbPr_To_RGB32f+12];
	mad.f32 	%f1272, %f1271, %f1261, %f1270;
	ld.const.f32 	%f1273, [k601YPbPr_To_RGB32f+20];
	mad.f32 	%f424, %f1273, %f1268, %f1272;
	.loc	19	268	0
	ld.const.f32 	%f1274, [k601YPbPr_To_RGB32f+4];
	mul.f32 	%f1275, %f1274, %f1267;
	ld.const.f32 	%f1276, [k601YPbPr_To_RGB32f+0];
	mad.f32 	%f1277, %f1276, %f1261, %f1275;
	ld.const.f32 	%f1278, [k601YPbPr_To_RGB32f+8];
	mad.f32 	%f430, %f1278, %f1268, %f1277;
	.loc	32	56	0
	ld.const.f32 	%f1279, [k601YPbPr_To_RGB32f+28];
	mul.f32 	%f1280, %f1279, %f1267;
	ld.const.f32 	%f1281, [k601YPbPr_To_RGB32f+24];
	mad.f32 	%f1282, %f1281, %f1261, %f1280;
	ld.const.f32 	%f1283, [k601YPbPr_To_RGB32f+32];
	mad.f32 	%f5, %f1283, %f1268, %f1282;
	mov.f32 	%f6, %f424;
	mov.f32 	%f7, %f430;
	bra.uni 	$Lt_3_5378;
$Lt_3_9218:
	.loc	33	46	0
	ld.const.f32 	%f1284, [kRGB32f_To_601YPbPr+4];
	mul.f32 	%f891, %f1284, %f6;
	ld.const.f32 	%f1285, [kRGB32f_To_601YPbPr+0];
	mad.f32 	%f893, %f1285, %f7, %f891;
	ld.const.f32 	%f1286, [kRGB32f_To_601YPbPr+8];
	mad.f32 	%f895, %f1286, %f5, %f893;
	ld.param.f32 	%f905, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter2+8];
	ld.param.f32 	%f1287, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter2+20];
	sub.f32 	%f1288, %f1287, %f905;
	mad.f32 	%f441, %f895, %f1288, %f905;
	.loc	33	47	0
	ld.param.f32 	%f1289, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter2+12];
	ld.param.f32 	%f1290, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter2+24];
	sub.f32 	%f1291, %f1290, %f1289;
	mad.f32 	%f445, %f895, %f1291, %f1289;
	.loc	33	49	0
	ld.param.f32 	%f1292, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter2+28];
	ld.param.f32 	%f880, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter2+4];
	ld.param.f32 	%f1293, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter2+16];
	sub.f32 	%f1294, %f1293, %f880;
	mad.f32 	%f1295, %f895, %f1294, %f880;
	sub.f32 	%f1296, %f1295, %f5;
	mad.f32 	%f5, %f1292, %f1296, %f5;
	.loc	33	50	0
	sub.f32 	%f1297, %f441, %f6;
	mad.f32 	%f6, %f1292, %f1297, %f6;
	.loc	33	51	0
	sub.f32 	%f1298, %f445, %f7;
	mad.f32 	%f7, %f1292, %f1298, %f7;
	.loc	35	103	0
	bra.uni 	$Lt_3_5378;
$Lt_3_9474:
	.loc	34	41	0
	sub.s32 	%r713, %r13, %r12;
	sub.s32 	%r12, %r713, 1;
$Lt_3_5378:
	.loc	35	54	0
	ld.param.u32 	%r714, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter3+0];
	mov.u32 	%r715, 0;
	setp.eq.s32 	%p173, %r714, %r715;
	@%p173 bra 	$Lt_3_258;
	mov.u32 	%r716, 1;
	setp.eq.s32 	%p174, %r714, %r716;
	@%p174 bra 	$Lt_3_770;
	mov.u32 	%r717, 2;
	setp.eq.s32 	%p175, %r714, %r717;
	@%p175 bra 	$Lt_3_1026;
	mov.u32 	%r718, 3;
	setp.eq.s32 	%p176, %r714, %r718;
	@%p176 bra 	$Lt_3_1282;
	mov.u32 	%r719, 4;
	setp.eq.s32 	%p177, %r714, %r719;
	@%p177 bra 	$Lt_3_1538;
	mov.u32 	%r720, 5;
	setp.eq.s32 	%p178, %r714, %r720;
	@%p178 bra 	$Lt_3_1794;
	mov.u32 	%r721, 6;
	setp.eq.s32 	%p179, %r714, %r721;
	@%p179 bra 	$Lt_3_2050;
	mov.u32 	%r722, 7;
	setp.eq.s32 	%p180, %r714, %r722;
	@%p180 bra 	$Lt_3_2306;
	mov.u32 	%r723, 8;
	setp.eq.s32 	%p181, %r714, %r723;
	@%p181 bra 	$Lt_3_2562;
	mov.u32 	%r724, 9;
	setp.eq.s32 	%p182, %r714, %r724;
	@%p182 bra 	$Lt_3_2818;
	mov.u32 	%r725, 10;
	setp.eq.s32 	%p183, %r714, %r725;
	@%p183 bra 	$Lt_3_3074;
	mov.u32 	%r726, 11;
	setp.eq.s32 	%p184, %r714, %r726;
	@%p184 bra 	$Lt_3_3330;
	mov.u32 	%r727, 12;
	setp.eq.s32 	%p185, %r714, %r727;
	@%p185 bra 	$Lt_3_3586;
	mov.u32 	%r728, 13;
	setp.eq.s32 	%p186, %r714, %r728;
	@%p186 bra 	$Lt_3_3842;
	mov.u32 	%r729, 14;
	setp.eq.s32 	%p187, %r714, %r729;
	@%p187 bra 	$Lt_3_4098;
	mov.u32 	%r730, 15;
	setp.eq.s32 	%p188, %r714, %r730;
	@%p188 bra 	$Lt_3_4354;
	mov.u32 	%r731, 16;
	setp.eq.s32 	%p189, %r714, %r731;
	@%p189 bra 	$Lt_3_4610;
	bra.uni 	$Lt_3_514;
$Lt_3_258:
	.loc	18	42	0
	cvt.sat.f32.f32 	%f8, %f8;
	ld.param.f32 	%f1299, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter3+8];
	mov.f32 	%f1300, 0f3f800000;  	// 1
	setp.eq.f32 	%p190, %f1299, %f1300;
	ld.param.f32 	%f1301, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter3+12];
	mov.f32 	%f1302, 0f3f800000;  	// 1
	setp.eq.f32 	%p191, %f1301, %f1302;
	ld.param.f32 	%f1303, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter3+4];
	ld.param.f32 	%f1304, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter3+16];
	mov.f32 	%f1305, 0f3f800000;  	// 1
	setp.eq.f32 	%p192, %f1304, %f1305;
	@!%p192 bra 	$Lt_3_239106;
	.loc	18	45	0
	mul.f32 	%f1306, %f8, %f1303;
	selp.f32 	%f8, %f1303, %f1306, %p190;
	.loc	18	53	0
	mov.f32 	%f1307, 0f3f800000;  	// 1
	sub.f32 	%f1308, %f1307, %f8;
	selp.f32 	%f8, %f1308, %f8, %p191;
	.loc	18	57	0
	mov.f32 	%f5, %f8;
	mov.f32 	%f6, %f8;
	mov.f32 	%f7, %f8;
	bra.uni 	$Lt_3_514;
$Lt_3_239106:
	@!%p190 bra 	$Lt_3_239618;
	.loc	18	61	0
	mov.f32 	%f1309, 0f3f800000;  	// 1
	sub.f32 	%f1310, %f1309, %f1303;
	selp.f32 	%f8, %f1310, %f1303, %p191;
	bra.uni 	$Lt_3_514;
$Lt_3_239618:
	.loc	18	69	0
	mov.f32 	%f1311, 0f3f800000;  	// 1
	sub.f32 	%f1312, %f1311, %f8;
	selp.f32 	%f8, %f1312, %f8, %p191;
	.loc	18	73	0
	mul.f32 	%f8, %f8, %f1303;
	bra.uni 	$Lt_3_514;
$Lt_3_770:
	.loc	19	267	0
	ld.const.f32 	%f1313, [kRGB32f_To_601YPbPr+4];
	mul.f32 	%f1314, %f1313, %f6;
	ld.const.f32 	%f1315, [kRGB32f_To_601YPbPr+0];
	mad.f32 	%f1316, %f1315, %f7, %f1314;
	ld.const.f32 	%f1317, [kRGB32f_To_601YPbPr+8];
	mad.f32 	%f1318, %f1317, %f5, %f1316;
	ld.const.f32 	%f1319, [k601YPbPr_To_RGB32f+12];
	mul.f32 	%f1320, %f1319, %f1318;
	.loc	19	268	0
	ld.const.f32 	%f1321, [k601YPbPr_To_RGB32f+0];
	mul.f32 	%f1322, %f1321, %f1318;
	.loc	20	44	0
	ld.const.f32 	%f1323, [k601YPbPr_To_RGB32f+24];
	mul.f32 	%f5, %f1323, %f1318;
	mov.f32 	%f6, %f1320;
	mov.f32 	%f7, %f1322;
	.loc	35	61	0
	bra.uni 	$Lt_3_514;
$Lt_3_1026:
	.loc	35	63	0
	ld.param.f32 	%f1324, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter3+20];
	ld.param.f32 	%f1325, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter3+12];
	ld.param.f32 	%f1326, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter3+4];
	mov.f32 	%f1327, 0f00000000;  	// 0
	setp.neu.f32 	%p193, %f1326, %f1327;
	@!%p193 bra 	$Lt_3_240130;
	.loc	21	44	0
	ld.param.f32 	%f1328, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter3+8];
	sub.f32 	%f1329, %f7, %f1325;
	mad.f32 	%f7, %f1328, %f1329, %f1324;
	.loc	21	45	0
	sub.f32 	%f1330, %f6, %f1325;
	mad.f32 	%f6, %f1328, %f1330, %f1324;
	.loc	21	46	0
	sub.f32 	%f1331, %f5, %f1325;
	mad.f32 	%f5, %f1328, %f1331, %f1324;
	bra.uni 	$Lt_3_514;
$Lt_3_240130:
	.loc	21	50	0
	ld.param.f32 	%f1332, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter3+24];
	setp.gt.f32 	%p194, %f7, %f1325;
	selp.f32 	%f7, %f1332, %f1324, %p194;
	.loc	21	51	0
	setp.gt.f32 	%p195, %f6, %f1325;
	selp.f32 	%f6, %f1332, %f1324, %p195;
	.loc	21	52	0
	setp.gt.f32 	%p196, %f5, %f1325;
	selp.f32 	%f5, %f1332, %f1324, %p196;
	bra.uni 	$Lt_3_514;
$Lt_3_1282:
	.loc	22	47	0
	ld.param.f32 	%f1333, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter3+4];
	mul.f32 	%f5, %f1333, %f5;
	ld.param.f32 	%f1334, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter3+8];
	mul.f32 	%f6, %f1334, %f6;
	ld.param.f32 	%f1335, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter3+12];
	mul.f32 	%f7, %f1335, %f7;
	.loc	35	67	0
	bra.uni 	$Lt_3_514;
$Lt_3_1538:
	.loc	23	48	0
	ld.param.f32 	%f1336, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter3+4];
	mov.f32 	%f1337, 0f00000000;  	// 0
	setp.eq.f32 	%p197, %f1336, %f1337;
	ld.param.f32 	%f1338, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter3+8];
	mov.f32 	%f1339, 0f00000000;  	// 0
	max.f32 	%f1340, %f5, %f1339;
	mov.f32 	%f1341, 0f3f800000;  	// 1
	min.f32 	%f1342, %f1340, %f1341;
	ld.param.f32 	%f1343, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter3+12];
	sub.f32 	%f1344, %f1342, %f1343;
	abs.f32 	%f1345, %f1344;
	mov.f32 	%f1346, 0f00000000;  	// 0
	max.f32 	%f1347, %f6, %f1346;
	mov.f32 	%f1348, 0f3f800000;  	// 1
	min.f32 	%f1349, %f1347, %f1348;
	ld.param.f32 	%f1350, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter3+16];
	sub.f32 	%f1351, %f1349, %f1350;
	abs.f32 	%f1352, %f1351;
	mov.f32 	%f1353, 0f00000000;  	// 0
	max.f32 	%f1354, %f7, %f1353;
	mov.f32 	%f1355, 0f3f800000;  	// 1
	min.f32 	%f1356, %f1354, %f1355;
	ld.param.f32 	%f1357, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter3+20];
	sub.f32 	%f1358, %f1356, %f1357;
	abs.f32 	%f1359, %f1358;
	mov.f32 	%f1360, 0f00000000;  	// 0
	max.f32 	%f1361, %f8, %f1360;
	mov.f32 	%f1362, 0f3f800000;  	// 1
	min.f32 	%f1363, %f1361, %f1362;
	sub.f32 	%f1364, %f1363, %f8;
	abs.f32 	%f1365, %f1364;
	max.f32 	%f1366, %f1359, %f1365;
	max.f32 	%f1367, %f1352, %f1366;
	max.f32 	%f1368, %f1345, %f1367;
	setp.ge.f32 	%p198, %f1338, %f1368;
	xor.pred 	%p199, %p197, %p198;
	@!%p199 bra 	$Lt_3_514;
	.loc	19	267	0
	ld.const.f32 	%f1369, [kRGB32f_To_601YPbPr+4];
	mul.f32 	%f1314, %f1369, %f6;
	ld.const.f32 	%f1370, [kRGB32f_To_601YPbPr+0];
	mad.f32 	%f1316, %f1370, %f7, %f1314;
	ld.const.f32 	%f1371, [kRGB32f_To_601YPbPr+8];
	mad.f32 	%f1318, %f1371, %f5, %f1316;
	ld.const.f32 	%f1372, [k601YPbPr_To_RGB32f+12];
	mul.f32 	%f1373, %f1372, %f1318;
	.loc	19	268	0
	ld.const.f32 	%f1374, [k601YPbPr_To_RGB32f+0];
	mul.f32 	%f1375, %f1374, %f1318;
	.loc	20	44	0
	ld.const.f32 	%f1376, [k601YPbPr_To_RGB32f+24];
	mul.f32 	%f5, %f1376, %f1318;
	mov.f32 	%f6, %f1373;
	mov.f32 	%f7, %f1375;
	bra.uni 	$Lt_3_514;
$Lt_3_1794:
	.loc	24	48	0
	ld.param.f32 	%f1377, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter3+8];
	mov.f32 	%f1378, 0f00000000;  	// 0
	max.f32 	%f1379, %f5, %f1378;
	mov.f32 	%f1380, 0f3f800000;  	// 1
	min.f32 	%f1381, %f1379, %f1380;
	ld.param.f32 	%f1382, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter3+12];
	sub.f32 	%f1383, %f1381, %f1382;
	abs.f32 	%f1384, %f1383;
	mov.f32 	%f1385, 0f00000000;  	// 0
	max.f32 	%f1386, %f6, %f1385;
	mov.f32 	%f1387, 0f3f800000;  	// 1
	min.f32 	%f1388, %f1386, %f1387;
	ld.param.f32 	%f1389, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter3+16];
	sub.f32 	%f1390, %f1388, %f1389;
	abs.f32 	%f1391, %f1390;
	mov.f32 	%f1392, 0f00000000;  	// 0
	max.f32 	%f1393, %f7, %f1392;
	mov.f32 	%f1394, 0f3f800000;  	// 1
	min.f32 	%f1395, %f1393, %f1394;
	ld.param.f32 	%f1396, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter3+20];
	sub.f32 	%f1397, %f1395, %f1396;
	abs.f32 	%f1398, %f1397;
	mov.f32 	%f1399, 0f00000000;  	// 0
	max.f32 	%f1400, %f8, %f1399;
	mov.f32 	%f1401, 0f3f800000;  	// 1
	min.f32 	%f1402, %f1400, %f1401;
	sub.f32 	%f1403, %f1402, %f8;
	abs.f32 	%f1404, %f1403;
	max.f32 	%f1405, %f1398, %f1404;
	max.f32 	%f1406, %f1391, %f1405;
	max.f32 	%f1407, %f1384, %f1406;
	setp.ge.f32 	%p200, %f1377, %f1407;
	@!%p200 bra 	$Lt_3_514;
	.loc	24	51	0
	ld.param.f32 	%f1408, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter3+24];
	mov.f32 	%f124, %f1408;
	ld.param.f32 	%f1409, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter3+28];
	mov.f32 	%f126, %f1409;
	ld.param.f32 	%f1410, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter3+32];
	mov.f32 	%f128, %f1410;
	ld.param.f32 	%f1411, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter3+4];
	mov.f32 	%f1412, 0f00000000;  	// 0
	setp.eq.f32 	%p201, %f1411, %f1412;
	@!%p201 bra 	$Lt_3_241410;
	.loc	24	60	0
	ld.const.f32 	%f1413, [kRGB32f_To_601YPbPr+4];
	mul.f32 	%f1314, %f1413, %f6;
	ld.const.f32 	%f1414, [kRGB32f_To_601YPbPr+0];
	mad.f32 	%f1316, %f1414, %f7, %f1314;
	ld.const.f32 	%f1415, [kRGB32f_To_601YPbPr+8];
	mad.f32 	%f1318, %f1415, %f5, %f1316;
	mul.f32 	%f124, %f1318, %f1408;
	.loc	24	61	0
	mul.f32 	%f126, %f1318, %f1409;
	.loc	24	62	0
	mul.f32 	%f128, %f1318, %f1410;
$Lt_3_241410:
	.loc	24	65	0
	mov.f32 	%f5, %f124;
	mov.f32 	%f6, %f126;
	mov.f32 	%f7, %f128;
	bra.uni 	$Lt_3_514;
$Lt_3_2050:
	.loc	25	47	0
	sub.s32 	%r732, %r13, %r12;
	sub.s32 	%r733, %r14, %r6;
	cvt.rn.f32.s32 	%f1416, %r6;
	cvt.rn.f32.s32 	%f1417, %r12;
	cvt.rn.f32.s32 	%f136, %r732;
	cvt.rn.f32.s32 	%f137, %r733;
	ld.param.f32 	%f1303, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter3+4];
	sub.f32 	%f1418, %f1303, %f1416;
	sub.f32 	%f1419, %f1303, %f1417;
	sub.f32 	%f1420, %f1303, %f136;
	sub.f32 	%f1421, %f1303, %f137;
	cvt.rzi.s32.f32 	%r734, %f1418;
	cvt.rzi.s32.f32 	%r735, %f1419;
	cvt.rzi.s32.f32 	%r736, %f1420;
	cvt.rzi.s32.f32 	%r737, %f1421;
	max.s32 	%r738, %r734, %r735;
	max.s32 	%r739, %r737, %r738;
	max.s32 	%r740, %r736, %r739;
	mov.u32 	%r741, 0;
	setp.le.s32 	%p202, %r740, %r741;
	@%p202 bra 	$Lt_3_514;
	.loc	25	51	0
	cvt.rn.f32.s32 	%f1422, %r740;
	sub.f32 	%f1423, %f1303, %f1422;
	div.full.f32 	%f1424, %f1423, %f1303;
	mul.f32 	%f8, %f8, %f1424;
	bra.uni 	$Lt_3_514;
$Lt_3_2306:
	.loc	27	50	0
	ld.const.f32 	%f1425, [kRGB32f_To_601YPbPr+4];
	mul.f32 	%f1314, %f1425, %f6;
	ld.param.f32 	%f1426, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter3+12];
	ld.param.f32 	%f1328, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter3+8];
	sub.f32 	%f1427, %f1426, %f1328;
	ld.const.f32 	%f1428, [kRGB32f_To_601YPbPr+0];
	mad.f32 	%f1316, %f1428, %f7, %f1314;
	ld.param.f32 	%f1429, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter3+16];
	mul.f32 	%f1430, %f1429, %f1427;
	ld.const.f32 	%f1431, [kRGB32f_To_601YPbPr+8];
	mad.f32 	%f1318, %f1431, %f5, %f1316;
	add.f32 	%f1432, %f1430, %f1426;
	sub.f32 	%f1433, %f1328, %f1430;
	mov.f32 	%f1434, 0f00000000;  	// 0
	max.f32 	%f1435, %f1432, %f1434;
	mov.f32 	%f1436, 0f00000000;  	// 0
	max.f32 	%f1437, %f1433, %f1436;
	mov.f32 	%f1438, 0f3f800000;  	// 1
	min.f32 	%f1439, %f1435, %f1438;
	mov.f32 	%f1440, 0f3f800000;  	// 1
	min.f32 	%f1441, %f1437, %f1440;
	set.gt.u32.f32 	%r742, %f1441, %f1318;
	neg.s32 	%r743, %r742;
	set.le.u32.f32 	%r744, %f1439, %f1318;
	neg.s32 	%r745, %r744;
	or.b32 	%r746, %r743, %r745;
	mov.u32 	%r747, 0;
	setp.eq.s32 	%p203, %r746, %r747;
	@%p203 bra 	$Lt_3_242690;
	mov.f32 	%f162, 0f00000000;   	// 0
	bra.uni 	$Lt_3_243458;
$Lt_3_242690:
	add.f32 	%f1442, %f1430, %f1328;
	mov.f32 	%f1443, 0f00000000;  	// 0
	max.f32 	%f1444, %f1442, %f1443;
	mov.f32 	%f1445, 0f3f800000;  	// 1
	min.f32 	%f1446, %f1444, %f1445;
	set.le.u32.f32 	%r748, %f1446, %f1318;
	neg.s32 	%r749, %r748;
	sub.f32 	%f1447, %f1426, %f1430;
	mov.f32 	%f1448, 0f00000000;  	// 0
	max.f32 	%f1449, %f1447, %f1448;
	mov.f32 	%f1450, 0f3f800000;  	// 1
	min.f32 	%f1451, %f1449, %f1450;
	set.lt.u32.f32 	%r750, %f1318, %f1451;
	neg.s32 	%r751, %r750;
	and.b32 	%r752, %r749, %r751;
	mov.u32 	%r753, 0;
	setp.eq.s32 	%p204, %r752, %r753;
	@%p204 bra 	$Lt_3_243202;
	mov.f32 	%f162, 0f3f800000;   	// 1
	bra.uni 	$Lt_3_243458;
$Lt_3_243202:
	add.f32 	%f1452, %f1430, %f1430;
	setp.gt.f32 	%p205, %f1446, %f1318;
	@!%p205 bra 	$Lt_3_243714;
	.loc	27	62	0
	sub.f32 	%f1453, %f1318, %f1441;
	div.full.f32 	%f162, %f1453, %f1452;
	bra.uni 	$Lt_3_243458;
$Lt_3_243714:
	.loc	27	66	0
	sub.f32 	%f1454, %f1439, %f1318;
	div.full.f32 	%f162, %f1454, %f1452;
$Lt_3_243458:
$Lt_3_242946:
$Lt_3_242434:
	.loc	27	69	0
	mov.f32 	%f1455, 0f3f800000;  	// 1
	sub.f32 	%f1456, %f1455, %f162;
	ld.param.f32 	%f1457, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter3+4];
	mov.f32 	%f1458, 0f00000000;  	// 0
	setp.neu.f32 	%p206, %f1457, %f1458;
	selp.f32 	%f162, %f1456, %f162, %p206;
	.loc	27	77	0
	ld.const.f32 	%f1459, [k601YPbPr_To_RGB32f+24];
	mul.f32 	%f5, %f1459, %f162;
	ld.const.f32 	%f1460, [k601YPbPr_To_RGB32f+12];
	mul.f32 	%f6, %f1460, %f162;
	ld.const.f32 	%f1461, [k601YPbPr_To_RGB32f+0];
	mul.f32 	%f7, %f1461, %f162;
	.loc	35	79	0
	bra.uni 	$Lt_3_514;
$Lt_3_2562:
	.loc	35	80	0
	ld.param.f32 	%f1462, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter3+44];
	mov.f32 	%f1463, 0f00000000;  	// 0
	setp.ge.f32 	%p207, %f1462, %f1463;
	@!%p207 bra 	$L_3_204290;
	cvt.rn.f32.s32 	%f1464, %r6;
	cvt.rn.f32.s32 	%f1465, %r14;
	mul.f32 	%f1466, %f1465, %f1462;
	setp.lt.f32 	%p208, %f1464, %f1466;
	@%p208 bra 	$L_3_204034;
$L_3_204290:
	mov.f32 	%f1467, 0f00000000;  	// 0
	setp.lt.f32 	%p209, %f1462, %f1467;
	@!%p209 bra 	$Lt_3_514;
	cvt.rn.f32.s32 	%f1468, %r12;
	cvt.rn.f32.s32 	%f1469, %r13;
	mul.f32 	%f1470, %f1469, %f1462;
	neg.f32 	%f1471, %f1470;
	setp.lt.f32 	%p210, %f1468, %f1471;
	@!%p210 bra 	$Lt_3_514;
$L_3_204034:
	.loc	28	47	0
	ld.const.f32 	%f1472, [kRGB32f_To_601YPbPr+4];
	mul.f32 	%f1473, %f1472, %f6;
	ld.const.f32 	%f1474, [kRGB32f_To_601YPbPr+0];
	mad.f32 	%f1475, %f1474, %f7, %f1473;
	ld.const.f32 	%f1476, [kRGB32f_To_601YPbPr+8];
	mad.f32 	%f1318, %f1476, %f5, %f1475;
	mov.f32 	%f198, %f1318;
	ld.param.f32 	%f1328, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter3+8];
	ld.param.f32 	%f1477, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter3+4];
	sub.f32 	%f1478, %f1477, %f1328;
	ld.param.f32 	%f1479, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter3+16];
	ld.param.f32 	%f1480, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter3+12];
	sub.f32 	%f1481, %f1480, %f1479;
	mov.f32 	%f1482, 0f3f800000;  	// 1
	set.neu.u32.f32 	%r754, %f1481, %f1482;
	neg.s32 	%r755, %r754;
	mov.f32 	%f1483, 0f3f800000;  	// 1
	set.neu.u32.f32 	%r756, %f1478, %f1483;
	neg.s32 	%r757, %r756;
	or.b32 	%r758, %r755, %r757;
	mov.u32 	%r759, 0;
	setp.eq.s32 	%p211, %r758, %r759;
	@%p211 bra 	$Lt_3_243970;
	.loc	17	116	0
	mov.u32 	%r760, 1;
	sub.u32 	%r71, %r760, %r5;
	shr.u32 	%r72, %r11, 13;
	sub.u32 	%r73, %r5, %r11;
	sub.u32 	%r74, %r71, %r11;
	xor.b32 	%r75, %r74, %r72;
	shl.b32 	%r76, %r75, 8;
	sub.u32 	%r77, %r73, %r75;
	sub.u32 	%r78, %r11, %r75;
	xor.b32 	%r79, %r77, %r76;
	shr.u32 	%r80, %r79, 13;
	sub.u32 	%r81, %r78, %r79;
	xor.b32 	%r82, %r81, %r80;
	shr.u32 	%r761, %r82, 12;
	sub.u32 	%r762, %r75, %r79;
	sub.u32 	%r763, %r762, %r82;
	xor.b32 	%r86, %r761, %r763;
	.loc	17	117	0
	sub.u32 	%r764, %r79, %r82;
	sub.u32 	%r88, %r764, %r86;
	shl.b32 	%r765, %r86, 16;
	xor.b32 	%r90, %r88, %r765;
	.loc	17	118	0
	sub.u32 	%r766, %r82, %r86;
	sub.u32 	%r92, %r766, %r90;
	shr.u32 	%r767, %r90, 5;
	xor.b32 	%r94, %r92, %r767;
	.loc	17	119	0
	sub.u32 	%r95, %r86, %r90;
	sub.u32 	%r96, %r95, %r94;
	shr.u32 	%r768, %r94, 3;
	xor.b32 	%r98, %r96, %r768;
	.loc	17	120	0
	sub.u32 	%r99, %r90, %r94;
	sub.u32 	%r100, %r99, %r98;
	shl.b32 	%r769, %r98, 10;
	xor.b32 	%r102, %r100, %r769;
	.loc	17	121	0
	sub.u32 	%r103, %r94, %r98;
	sub.u32 	%r104, %r103, %r102;
	shr.u32 	%r770, %r102, 15;
	xor.b32 	%r106, %r104, %r770;
	.loc	28	57	0
	mov.f32 	%f1484, 0f3b270d73;  	// 0.00254902
	mul.lo.u32 	%r771, %r106, 1103515245;
	add.u32 	%r772, %r771, 12345;
	shr.u32 	%r773, %r772, 16;
	and.b32 	%r774, %r773, 255;
	shl.b32 	%r775, %r774, 7;
	mul.lo.u32 	%r776, %r106, -1029531031;
	sub.u32 	%r777, %r776, 740551042;
	shr.u32 	%r778, %r777, 16;
	and.b32 	%r779, %r778, 255;
	xor.b32 	%r780, %r775, %r779;
	cvt.rn.f32.s32 	%f1485, %r780;
	mov.f32 	%f1486, 0f467ffe00;  	// 16383.5
	div.full.f32 	%f1487, %f1485, %f1486;
	mov.f32 	%f1488, 0fbf800000;  	// -1
	add.f32 	%f1489, %f1487, %f1488;
	mad.f32 	%f198, %f1484, %f1489, %f1318;
$Lt_3_243970:
	sub.f32 	%f1490, %f198, %f1328;
	ld.param.f32 	%f1491, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter3+20];
	mov.f32 	%f1492, 0f3f800000;  	// 1
	setp.neu.f32 	%p212, %f1491, %f1492;
	@!%p212 bra 	$Lt_3_244738;
	mov.f32 	%f1493, 0f00000000;  	// 0
	setp.lt.f32 	%p213, %f1490, %f1493;
	@!%p213 bra 	$Lt_3_245250;
	.loc	28	66	0
	mov.f32 	%f198, %f1479;
	bra.uni 	$Lt_3_244482;
$Lt_3_245250:
	.loc	28	70	0
	rcp.approx.f32 	%f1494, %f1478;
	mul.f32 	%f1495, %f1490, %f1494;
	lg2.approx.f32 	%f1496, %f1495;
	mul.f32 	%f1497, %f1491, %f1496;
	ex2.approx.f32 	%f1498, %f1497;
	mad.f32 	%f198, %f1481, %f1498, %f1479;
	bra.uni 	$Lt_3_244482;
$Lt_3_244738:
	.loc	28	77	0
	rcp.approx.f32 	%f1499, %f1478;
	mul.f32 	%f1500, %f1490, %f1499;
	mad.f32 	%f198, %f1481, %f1500, %f1479;
$Lt_3_244482:
	.loc	19	267	0
	ld.param.f32 	%f1501, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter3+28];
	sin.approx.f32 	%f1502, %f1501;
	ld.const.f32 	%f1503, [kRGB32f_To_601YPbPr+28];
	mul.f32 	%f1504, %f1503, %f6;
	ld.const.f32 	%f1505, [kRGB32f_To_601YPbPr+16];
	mul.f32 	%f1506, %f1505, %f6;
	cos.approx.f32 	%f1507, %f1501;
	ld.const.f32 	%f1508, [kRGB32f_To_601YPbPr+24];
	mad.f32 	%f1509, %f1508, %f7, %f1504;
	ld.const.f32 	%f1510, [kRGB32f_To_601YPbPr+12];
	mad.f32 	%f1511, %f1510, %f7, %f1506;
	ld.const.f32 	%f1512, [kRGB32f_To_601YPbPr+32];
	mad.f32 	%f1513, %f1512, %f5, %f1509;
	ld.const.f32 	%f1514, [kRGB32f_To_601YPbPr+20];
	mad.f32 	%f1515, %f1514, %f5, %f1511;
	mul.f32 	%f1516, %f1502, %f1513;
	mul.f32 	%f1517, %f1507, %f1513;
	mul.f32 	%f1518, %f1515, %f1507;
	sub.f32 	%f1519, %f1518, %f1516;
	mad.f32 	%f1520, %f1515, %f1502, %f1517;
	ld.param.f32 	%f1521, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter3+32];
	sub.f32 	%f1522, %f1521, %f1519;
	ld.param.f32 	%f1523, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter3+36];
	sub.f32 	%f1524, %f1523, %f1520;
	ld.param.f32 	%f1525, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter3+40];
	mad.f32 	%f1526, %f1525, %f1522, %f1519;
	mad.f32 	%f1527, %f1525, %f1524, %f1520;
	ld.param.f32 	%f1528, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter3+24];
	mul.f32 	%f1529, %f1526, %f1528;
	mul.f32 	%f1530, %f1527, %f1528;
	ld.const.f32 	%f1531, [k601YPbPr_To_RGB32f+16];
	mul.f32 	%f1532, %f1531, %f1529;
	ld.const.f32 	%f1533, [k601YPbPr_To_RGB32f+12];
	mad.f32 	%f1534, %f1533, %f198, %f1532;
	ld.const.f32 	%f1535, [k601YPbPr_To_RGB32f+20];
	mad.f32 	%f258, %f1535, %f1530, %f1534;
	.loc	19	268	0
	ld.const.f32 	%f1536, [k601YPbPr_To_RGB32f+4];
	mul.f32 	%f1537, %f1536, %f1529;
	ld.const.f32 	%f1538, [k601YPbPr_To_RGB32f+0];
	mad.f32 	%f1539, %f1538, %f198, %f1537;
	ld.const.f32 	%f1540, [k601YPbPr_To_RGB32f+8];
	mad.f32 	%f264, %f1540, %f1530, %f1539;
	.loc	28	92	0
	ld.const.f32 	%f1541, [k601YPbPr_To_RGB32f+28];
	mul.f32 	%f1542, %f1541, %f1529;
	ld.const.f32 	%f1543, [k601YPbPr_To_RGB32f+24];
	mad.f32 	%f1544, %f1543, %f198, %f1542;
	ld.const.f32 	%f1545, [k601YPbPr_To_RGB32f+32];
	mad.f32 	%f5, %f1545, %f1530, %f1544;
	mov.f32 	%f6, %f258;
	mov.f32 	%f7, %f264;
	bra.uni 	$Lt_3_514;
$Lt_3_2818:
	.loc	35	83	0
	ld.param.f32 	%f1546, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter3+44];
	mov.f32 	%f1547, 0f00000000;  	// 0
	setp.ge.f32 	%p214, %f1546, %f1547;
	@!%p214 bra 	$L_3_205058;
	cvt.rn.f32.s32 	%f1548, %r6;
	cvt.rn.f32.s32 	%f1549, %r14;
	mul.f32 	%f1550, %f1549, %f1546;
	setp.lt.f32 	%p215, %f1548, %f1550;
	@%p215 bra 	$L_3_204802;
$L_3_205058:
	mov.f32 	%f1551, 0f00000000;  	// 0
	setp.lt.f32 	%p216, %f1546, %f1551;
	@!%p216 bra 	$Lt_3_514;
	cvt.rn.f32.s32 	%f1552, %r12;
	cvt.rn.f32.s32 	%f1553, %r13;
	mul.f32 	%f1554, %f1553, %f1546;
	neg.f32 	%f1555, %f1554;
	setp.lt.f32 	%p217, %f1552, %f1555;
	@!%p217 bra 	$Lt_3_514;
$L_3_204802:
	.loc	28	110	0
	ld.const.f32 	%f1556, [kRGB32f_To_601YPbPr+4];
	mul.f32 	%f1314, %f1556, %f6;
	ld.param.f32 	%f1328, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter3+8];
	ld.param.f32 	%f1557, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter3+4];
	sub.f32 	%f1558, %f1557, %f1328;
	ld.param.f32 	%f1559, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter3+16];
	ld.param.f32 	%f1560, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter3+12];
	sub.f32 	%f1561, %f1560, %f1559;
	ld.const.f32 	%f1562, [kRGB32f_To_601YPbPr+0];
	mad.f32 	%f1316, %f1562, %f7, %f1314;
	ld.const.f32 	%f1563, [kRGB32f_To_601YPbPr+8];
	mad.f32 	%f1318, %f1563, %f5, %f1316;
	sub.f32 	%f1564, %f1318, %f1328;
	ld.param.f32 	%f1565, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter3+20];
	mov.f32 	%f1566, 0f3f800000;  	// 1
	setp.neu.f32 	%p218, %f1565, %f1566;
	@!%p218 bra 	$Lt_3_245762;
	.loc	28	120	0
	mov.f32 	%f1567, 0f00000000;  	// 0
	max.f32 	%f1568, %f1564, %f1567;
	div.full.f32 	%f1569, %f1568, %f1558;
	lg2.approx.f32 	%f1570, %f1569;
	mul.f32 	%f1571, %f1565, %f1570;
	ex2.approx.f32 	%f1572, %f1571;
	mad.f32 	%f297, %f1561, %f1572, %f1559;
	bra.uni 	$Lt_3_245506;
$Lt_3_245762:
	.loc	28	129	0
	div.full.f32 	%f1573, %f1564, %f1558;
	mad.f32 	%f297, %f1561, %f1573, %f1559;
$Lt_3_245506:
	.loc	28	135	0
	ld.const.f32 	%f1574, [k601YPbPr_To_RGB32f+24];
	mul.f32 	%f5, %f1574, %f297;
	ld.const.f32 	%f1575, [k601YPbPr_To_RGB32f+12];
	mul.f32 	%f6, %f1575, %f297;
	ld.const.f32 	%f1576, [k601YPbPr_To_RGB32f+0];
	mul.f32 	%f7, %f1576, %f297;
	bra.uni 	$Lt_3_514;
$Lt_3_3074:
	.loc	35	86	0
	ld.param.f32 	%f1577, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter3+44];
	mov.f32 	%f1578, 0f00000000;  	// 0
	setp.ge.f32 	%p219, %f1577, %f1578;
	@!%p219 bra 	$L_3_205826;
	cvt.rn.f32.s32 	%f1579, %r6;
	cvt.rn.f32.s32 	%f1580, %r14;
	mul.f32 	%f1581, %f1580, %f1577;
	setp.lt.f32 	%p220, %f1579, %f1581;
	@%p220 bra 	$L_3_205570;
$L_3_205826:
	mov.f32 	%f1582, 0f00000000;  	// 0
	setp.lt.f32 	%p221, %f1577, %f1582;
	@!%p221 bra 	$Lt_3_514;
	cvt.rn.f32.s32 	%f1583, %r12;
	cvt.rn.f32.s32 	%f1584, %r13;
	mul.f32 	%f1585, %f1584, %f1577;
	neg.f32 	%f1586, %f1585;
	setp.lt.f32 	%p222, %f1583, %f1586;
	@!%p222 bra 	$Lt_3_514;
$L_3_205570:
	.loc	28	160	0
	ld.const.f32 	%f5, [k601YPbPr_To_RGB32f+24];
	ld.const.f32 	%f6, [k601YPbPr_To_RGB32f+12];
	ld.const.f32 	%f7, [k601YPbPr_To_RGB32f+0];
	bra.uni 	$Lt_3_514;
$Lt_3_3330:
	.loc	29	42	0
	ld.param.f32 	%f1303, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter3+4];
	mov.f32 	%f1587, 0f00000000;  	// 0
	setp.ge.f32 	%p223, %f5, %f1587;
	@!%p223 bra 	$Lt_3_246274;
	.loc	19	292	0
	lg2.approx.f32 	%f1588, %f5;
	mul.f32 	%f1589, %f1303, %f1588;
	ex2.approx.f32 	%f1590, %f1589;
	bra.uni 	$Lt_3_246018;
$Lt_3_246274:
	neg.f32 	%f1591, %f5;
	lg2.approx.f32 	%f1592, %f1591;
	mul.f32 	%f1593, %f1303, %f1592;
	ex2.approx.f32 	%f1594, %f1593;
	neg.f32 	%f1590, %f1594;
$Lt_3_246018:
	mov.f32 	%f1595, 0f00000000;  	// 0
	setp.ge.f32 	%p224, %f6, %f1595;
	@!%p224 bra 	$Lt_3_246786;
	.loc	19	293	0
	lg2.approx.f32 	%f1596, %f6;
	mul.f32 	%f1597, %f1303, %f1596;
	ex2.approx.f32 	%f1598, %f1597;
	bra.uni 	$Lt_3_246530;
$Lt_3_246786:
	neg.f32 	%f1599, %f6;
	lg2.approx.f32 	%f1600, %f1599;
	mul.f32 	%f1601, %f1303, %f1600;
	ex2.approx.f32 	%f1602, %f1601;
	neg.f32 	%f1598, %f1602;
$Lt_3_246530:
	mov.f32 	%f1603, 0f00000000;  	// 0
	setp.ge.f32 	%p225, %f7, %f1603;
	@!%p225 bra 	$Lt_3_247298;
	.loc	19	294	0
	lg2.approx.f32 	%f1604, %f7;
	mul.f32 	%f1605, %f1303, %f1604;
	ex2.approx.f32 	%f1606, %f1605;
	bra.uni 	$Lt_3_247042;
$Lt_3_247298:
	neg.f32 	%f1607, %f7;
	lg2.approx.f32 	%f1608, %f1607;
	mul.f32 	%f1609, %f1303, %f1608;
	ex2.approx.f32 	%f1610, %f1609;
	neg.f32 	%f1606, %f1610;
$Lt_3_247042:
	mov.f32 	%f1611, 0f00000000;  	// 0
	setp.ge.f32 	%p226, %f8, %f1611;
	@!%p226 bra 	$Lt_3_247810;
	.loc	19	295	0
	lg2.approx.f32 	%f1612, %f8;
	mul.f32 	%f1613, %f1303, %f1612;
	ex2.approx.f32 	%f1614, %f1613;
	bra.uni 	$Lt_3_247554;
$Lt_3_247810:
	neg.f32 	%f1615, %f8;
	lg2.approx.f32 	%f1616, %f1615;
	mul.f32 	%f1617, %f1303, %f1616;
	ex2.approx.f32 	%f1618, %f1617;
	neg.f32 	%f1614, %f1618;
$Lt_3_247554:
	.loc	29	42	0
	mov.f32 	%f5, %f1590;
	mov.f32 	%f6, %f1598;
	mov.f32 	%f7, %f1606;
	mov.f32 	%f8, %f1614;
	.loc	35	91	0
	bra.uni 	$Lt_3_514;
$Lt_3_3586:
	.loc	30	41	0
	sub.s32 	%r781, %r14, %r6;
	sub.s32 	%r6, %r781, 1;
	.loc	35	94	0
	bra.uni 	$Lt_3_514;
$Lt_3_3842:
	.loc	35	96	0
	sub.u32 	%r343, %r6, %r12;
	ld.param.f32 	%f1303, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter3+4];
	mov.f32 	%f1619, 0f3e800000;  	// 0.25
	mul.f32 	%f1620, %f1303, %f1619;
	ld.param.f32 	%f1621, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter3+16];
	ld.param.f32 	%f1622, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter3+8];
	mov.f32 	%f1623, 0f3f800000;  	// 1
	setp.eq.f32 	%p227, %f1622, %f1623;
	@!%p227 bra 	$Lt_3_248322;
	.loc	17	117	0
	cvt.rzi.u32.f32 	%r782, %f1621;
	shr.u32 	%r783, %r782, 13;
	sub.u32 	%r784, %r12, %r782;
	sub.u32 	%r785, %r343, %r782;
	xor.b32 	%r786, %r785, %r783;
	shl.b32 	%r787, %r786, 8;
	sub.u32 	%r788, %r784, %r786;
	sub.u32 	%r789, %r782, %r786;
	xor.b32 	%r790, %r788, %r787;
	shr.u32 	%r791, %r790, 13;
	sub.u32 	%r792, %r789, %r790;
	sub.u32 	%r793, %r786, %r790;
	xor.b32 	%r794, %r792, %r791;
	shr.u32 	%r795, %r794, 12;
	sub.u32 	%r796, %r793, %r794;
	xor.b32 	%r797, %r796, %r795;
	shl.b32 	%r798, %r797, 16;
	sub.u32 	%r799, %r790, %r794;
	sub.u32 	%r800, %r799, %r797;
	xor.b32 	%r801, %r798, %r800;
	.loc	17	118	0
	sub.u32 	%r802, %r794, %r797;
	sub.u32 	%r803, %r802, %r801;
	shr.u32 	%r804, %r801, 5;
	xor.b32 	%r805, %r803, %r804;
	.loc	17	119	0
	sub.u32 	%r806, %r797, %r801;
	sub.u32 	%r807, %r806, %r805;
	shr.u32 	%r808, %r805, 3;
	xor.b32 	%r809, %r807, %r808;
	.loc	17	120	0
	sub.u32 	%r810, %r801, %r805;
	sub.u32 	%r811, %r810, %r809;
	shl.b32 	%r812, %r809, 10;
	xor.b32 	%r813, %r811, %r812;
	.loc	17	121	0
	sub.u32 	%r814, %r805, %r809;
	sub.u32 	%r815, %r814, %r813;
	shr.u32 	%r816, %r813, 15;
	xor.b32 	%r817, %r815, %r816;
	.loc	31	48	0
	mul.lo.u32 	%r818, %r817, 1103515245;
	add.u32 	%r819, %r818, 12345;
	shr.u32 	%r820, %r819, 16;
	and.b32 	%r821, %r820, 255;
	shl.b32 	%r822, %r821, 7;
	mul.lo.u32 	%r823, %r817, -1029531031;
	sub.u32 	%r824, %r823, 740551042;
	shr.u32 	%r825, %r824, 16;
	and.b32 	%r826, %r825, 255;
	xor.b32 	%r827, %r822, %r826;
	cvt.rn.f32.s32 	%f1624, %r827;
	mov.f32 	%f1625, 0f46fffe00;  	// 32767
	div.full.f32 	%f1626, %f1624, %f1625;
	mul.f32 	%f1627, %f1303, %f1626;
	mov.f32 	%f1628, 0f3f000000;  	// 0.5
	mul.f32 	%f1629, %f1627, %f1628;
	sub.f32 	%f1630, %f1629, %f1620;
	.loc	17	117	0
	add.u32 	%r828, %r785, 1;
	xor.b32 	%r829, %r828, %r783;
	shl.b32 	%r830, %r829, 8;
	sub.u32 	%r831, %r784, %r829;
	sub.u32 	%r832, %r782, %r829;
	xor.b32 	%r833, %r830, %r831;
	shr.u32 	%r834, %r833, 13;
	sub.u32 	%r835, %r832, %r833;
	sub.u32 	%r836, %r829, %r833;
	xor.b32 	%r837, %r834, %r835;
	shr.u32 	%r838, %r837, 12;
	sub.u32 	%r839, %r836, %r837;
	xor.b32 	%r840, %r838, %r839;
	sub.u32 	%r841, %r833, %r837;
	sub.u32 	%r178, %r841, %r840;
	shl.b32 	%r842, %r840, 16;
	xor.b32 	%r178, %r178, %r842;
	.loc	17	118	0
	sub.u32 	%r180, %r837, %r840;
	sub.u32 	%r180, %r180, %r178;
	shr.u32 	%r843, %r178, 5;
	xor.b32 	%r180, %r180, %r843;
	.loc	17	119	0
	sub.u32 	%r182, %r840, %r178;
	sub.u32 	%r182, %r182, %r180;
	shr.u32 	%r844, %r180, 3;
	xor.b32 	%r182, %r182, %r844;
	.loc	17	120	0
	sub.u32 	%r178, %r178, %r180;
	sub.u32 	%r178, %r178, %r182;
	shl.b32 	%r845, %r182, 10;
	xor.b32 	%r178, %r178, %r845;
	.loc	17	121	0
	sub.u32 	%r180, %r180, %r182;
	sub.u32 	%r180, %r180, %r178;
	shr.u32 	%r846, %r178, 15;
	xor.b32 	%r180, %r180, %r846;
	.loc	31	49	0
	mul.lo.u32 	%r847, %r180, 1103515245;
	add.u32 	%r848, %r847, 12345;
	shr.u32 	%r849, %r848, 16;
	and.b32 	%r850, %r849, 255;
	shl.b32 	%r851, %r850, 7;
	mul.lo.u32 	%r852, %r180, -1029531031;
	sub.u32 	%r853, %r852, 740551042;
	shr.u32 	%r854, %r853, 16;
	and.b32 	%r855, %r854, 255;
	xor.b32 	%r856, %r851, %r855;
	cvt.rn.f32.s32 	%f1631, %r856;
	mov.f32 	%f1632, 0f46fffe00;  	// 32767
	div.full.f32 	%f1633, %f1631, %f1632;
	mul.f32 	%f1634, %f1303, %f1633;
	mov.f32 	%f1635, 0f3f000000;  	// 0.5
	mul.f32 	%f1636, %f1634, %f1635;
	sub.f32 	%f362, %f1636, %f1620;
	.loc	17	117	0
	add.u32 	%r857, %r785, 2;
	xor.b32 	%r858, %r857, %r783;
	shl.b32 	%r859, %r858, 8;
	sub.u32 	%r860, %r784, %r858;
	sub.u32 	%r861, %r782, %r858;
	xor.b32 	%r862, %r859, %r860;
	shr.u32 	%r863, %r862, 13;
	sub.u32 	%r864, %r861, %r862;
	sub.u32 	%r865, %r858, %r862;
	xor.b32 	%r866, %r863, %r864;
	shr.u32 	%r867, %r866, 12;
	sub.u32 	%r868, %r865, %r866;
	xor.b32 	%r869, %r867, %r868;
	sub.u32 	%r870, %r862, %r866;
	sub.u32 	%r210, %r870, %r869;
	shl.b32 	%r871, %r869, 16;
	xor.b32 	%r210, %r210, %r871;
	.loc	17	118	0
	sub.u32 	%r212, %r866, %r869;
	sub.u32 	%r212, %r212, %r210;
	shr.u32 	%r872, %r210, 5;
	xor.b32 	%r212, %r212, %r872;
	.loc	17	119	0
	sub.u32 	%r214, %r869, %r210;
	sub.u32 	%r214, %r214, %r212;
	shr.u32 	%r873, %r212, 3;
	xor.b32 	%r214, %r214, %r873;
	.loc	17	120	0
	sub.u32 	%r210, %r210, %r212;
	sub.u32 	%r210, %r210, %r214;
	shl.b32 	%r874, %r214, 10;
	xor.b32 	%r210, %r210, %r874;
	.loc	17	121	0
	sub.u32 	%r212, %r212, %r214;
	sub.u32 	%r212, %r212, %r210;
	shr.u32 	%r875, %r210, 15;
	xor.b32 	%r212, %r212, %r875;
	.loc	31	50	0
	mul.lo.u32 	%r876, %r212, 1103515245;
	add.u32 	%r877, %r876, 12345;
	shr.u32 	%r878, %r877, 16;
	and.b32 	%r879, %r878, 255;
	shl.b32 	%r880, %r879, 7;
	mul.lo.u32 	%r881, %r212, -1029531031;
	sub.u32 	%r882, %r881, 740551042;
	shr.u32 	%r883, %r882, 16;
	and.b32 	%r884, %r883, 255;
	xor.b32 	%r885, %r880, %r884;
	cvt.rn.f32.s32 	%f1637, %r885;
	mov.f32 	%f1638, 0f46fffe00;  	// 32767
	div.full.f32 	%f1639, %f1637, %f1638;
	mul.f32 	%f1640, %f1303, %f1639;
	mov.f32 	%f1641, 0f3f000000;  	// 0.5
	mul.f32 	%f1642, %f1640, %f1641;
	sub.f32 	%f369, %f1642, %f1620;
	bra.uni 	$Lt_3_248066;
$Lt_3_248322:
	.loc	17	117	0
	cvt.rzi.u32.f32 	%r886, %f1621;
	shr.u32 	%r887, %r886, 13;
	sub.u32 	%r888, %r12, %r886;
	sub.u32 	%r889, %r343, %r886;
	xor.b32 	%r890, %r889, %r887;
	shl.b32 	%r891, %r890, 8;
	sub.u32 	%r892, %r888, %r890;
	sub.u32 	%r893, %r886, %r890;
	xor.b32 	%r894, %r892, %r891;
	shr.u32 	%r895, %r894, 13;
	sub.u32 	%r896, %r893, %r894;
	sub.u32 	%r897, %r890, %r894;
	xor.b32 	%r898, %r896, %r895;
	shr.u32 	%r899, %r898, 12;
	sub.u32 	%r900, %r897, %r898;
	xor.b32 	%r901, %r900, %r899;
	shl.b32 	%r902, %r901, 16;
	sub.u32 	%r903, %r894, %r898;
	sub.u32 	%r904, %r903, %r901;
	xor.b32 	%r905, %r902, %r904;
	.loc	17	118	0
	sub.u32 	%r906, %r898, %r901;
	sub.u32 	%r907, %r906, %r905;
	shr.u32 	%r908, %r905, 5;
	xor.b32 	%r909, %r907, %r908;
	.loc	17	119	0
	sub.u32 	%r910, %r901, %r905;
	sub.u32 	%r911, %r910, %r909;
	shr.u32 	%r912, %r909, 3;
	xor.b32 	%r913, %r911, %r912;
	.loc	17	120	0
	sub.u32 	%r914, %r905, %r909;
	sub.u32 	%r915, %r914, %r913;
	shl.b32 	%r916, %r913, 10;
	xor.b32 	%r917, %r915, %r916;
	.loc	17	121	0
	sub.u32 	%r918, %r909, %r913;
	sub.u32 	%r919, %r918, %r917;
	shr.u32 	%r920, %r917, 15;
	xor.b32 	%r921, %r919, %r920;
	.loc	31	54	0
	mul.lo.u32 	%r922, %r921, 1103515245;
	mul.lo.u32 	%r923, %r921, -1029531031;
	add.u32 	%r924, %r922, 12345;
	sub.u32 	%r925, %r923, 740551042;
	shr.u32 	%r926, %r924, 16;
	shr.u32 	%r927, %r925, 16;
	and.b32 	%r928, %r926, 255;
	and.b32 	%r929, %r927, 255;
	shl.b32 	%r930, %r928, 7;
	xor.b32 	%r931, %r930, %r929;
	cvt.rn.f32.s32 	%f1643, %r931;
	mov.f32 	%f1644, 0f46fffe00;  	// 32767
	div.full.f32 	%f1645, %f1643, %f1644;
	mul.f32 	%f1646, %f1645, %f1303;
	mov.f32 	%f1647, 0f3f000000;  	// 0.5
	mul.f32 	%f1648, %f1646, %f1647;
	sub.f32 	%f1630, %f1648, %f1620;
	mov.f32 	%f369, %f1630;
	mov.f32 	%f362, %f1630;
$Lt_3_248066:
	.loc	31	57	0
	add.f32 	%f7, %f1630, %f7;
	.loc	31	58	0
	add.f32 	%f6, %f362, %f6;
	.loc	31	59	0
	add.f32 	%f5, %f5, %f369;
	ld.param.f32 	%f1649, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter3+12];
	mov.f32 	%f1650, 0f3f800000;  	// 1
	setp.eq.f32 	%p228, %f1649, %f1650;
	@!%p228 bra 	$Lt_3_514;
	.loc	31	63	0
	mov.f32 	%f1651, 0f00000000;  	// 0
	max.f32 	%f1652, %f5, %f1651;
	mov.f32 	%f1653, 0f3f800000;  	// 1
	min.f32 	%f5, %f1652, %f1653;
	mov.f32 	%f1654, 0f00000000;  	// 0
	max.f32 	%f1655, %f6, %f1654;
	mov.f32 	%f1656, 0f3f800000;  	// 1
	min.f32 	%f6, %f1655, %f1656;
	mov.f32 	%f1657, 0f00000000;  	// 0
	max.f32 	%f1658, %f7, %f1657;
	mov.f32 	%f1659, 0f3f800000;  	// 1
	min.f32 	%f7, %f1658, %f1659;
	mov.f32 	%f1660, 0f00000000;  	// 0
	max.f32 	%f1661, %f8, %f1660;
	mov.f32 	%f1662, 0f3f800000;  	// 1
	min.f32 	%f8, %f1661, %f1662;
	bra.uni 	$Lt_3_514;
$Lt_3_4098:
	.loc	35	99	0
	cvt.rn.f32.s32 	%f1663, %r6;
	cvt.rn.f32.s32 	%f1664, %r14;
	ld.param.f32 	%f1665, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter3+20];
	mul.f32 	%f1666, %f1664, %f1665;
	setp.lt.f32 	%p229, %f1663, %f1666;
	@!%p229 bra 	$Lt_3_514;
	.loc	19	267	0
	ld.const.f32 	%f1667, [kRGB32f_To_601YPbPr+4];
	mul.f32 	%f1314, %f1667, %f6;
	ld.const.f32 	%f1668, [kRGB32f_To_601YPbPr+28];
	mul.f32 	%f1669, %f1668, %f6;
	ld.const.f32 	%f1670, [kRGB32f_To_601YPbPr+16];
	mul.f32 	%f1671, %f1670, %f6;
	ld.const.f32 	%f1672, [kRGB32f_To_601YPbPr+0];
	mad.f32 	%f1316, %f1672, %f7, %f1314;
	ld.const.f32 	%f1673, [kRGB32f_To_601YPbPr+24];
	mad.f32 	%f1674, %f1673, %f7, %f1669;
	ld.const.f32 	%f1675, [kRGB32f_To_601YPbPr+12];
	mad.f32 	%f1676, %f1675, %f7, %f1671;
	ld.const.f32 	%f1677, [kRGB32f_To_601YPbPr+8];
	mad.f32 	%f1318, %f1677, %f5, %f1316;
	ld.const.f32 	%f1678, [kRGB32f_To_601YPbPr+32];
	mad.f32 	%f1679, %f1678, %f5, %f1674;
	ld.const.f32 	%f1680, [kRGB32f_To_601YPbPr+20];
	mad.f32 	%f1681, %f1680, %f5, %f1676;
	ld.param.f32 	%f1682, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter3+4];
	ld.param.f32 	%f1683, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter3+8];
	mad.f32 	%f1684, %f1318, %f1683, %f1682;
	ld.param.f32 	%f1685, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter3+16];
	mul.f32 	%f1686, %f1679, %f1685;
	ld.param.f32 	%f1687, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter3+12];
	mul.f32 	%f1688, %f1679, %f1687;
	mul.f32 	%f1689, %f1681, %f1687;
	sub.f32 	%f1690, %f1689, %f1686;
	mad.f32 	%f1691, %f1681, %f1685, %f1688;
	ld.const.f32 	%f1692, [k601YPbPr_To_RGB32f+16];
	mul.f32 	%f1693, %f1692, %f1690;
	ld.const.f32 	%f1694, [k601YPbPr_To_RGB32f+12];
	mad.f32 	%f1695, %f1694, %f1684, %f1693;
	ld.const.f32 	%f1696, [k601YPbPr_To_RGB32f+20];
	mad.f32 	%f424, %f1696, %f1691, %f1695;
	.loc	19	268	0
	ld.const.f32 	%f1697, [k601YPbPr_To_RGB32f+4];
	mul.f32 	%f1698, %f1697, %f1690;
	ld.const.f32 	%f1699, [k601YPbPr_To_RGB32f+0];
	mad.f32 	%f1700, %f1699, %f1684, %f1698;
	ld.const.f32 	%f1701, [k601YPbPr_To_RGB32f+8];
	mad.f32 	%f430, %f1701, %f1691, %f1700;
	.loc	32	56	0
	ld.const.f32 	%f1702, [k601YPbPr_To_RGB32f+28];
	mul.f32 	%f1703, %f1702, %f1690;
	ld.const.f32 	%f1704, [k601YPbPr_To_RGB32f+24];
	mad.f32 	%f1705, %f1704, %f1684, %f1703;
	ld.const.f32 	%f1706, [k601YPbPr_To_RGB32f+32];
	mad.f32 	%f5, %f1706, %f1691, %f1705;
	mov.f32 	%f6, %f424;
	mov.f32 	%f7, %f430;
	bra.uni 	$Lt_3_514;
$Lt_3_4354:
	.loc	33	46	0
	ld.const.f32 	%f1707, [kRGB32f_To_601YPbPr+4];
	mul.f32 	%f1314, %f1707, %f6;
	ld.const.f32 	%f1708, [kRGB32f_To_601YPbPr+0];
	mad.f32 	%f1316, %f1708, %f7, %f1314;
	ld.const.f32 	%f1709, [kRGB32f_To_601YPbPr+8];
	mad.f32 	%f1318, %f1709, %f5, %f1316;
	ld.param.f32 	%f1328, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter3+8];
	ld.param.f32 	%f1710, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter3+20];
	sub.f32 	%f1711, %f1710, %f1328;
	mad.f32 	%f441, %f1318, %f1711, %f1328;
	.loc	33	47	0
	ld.param.f32 	%f1712, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter3+12];
	ld.param.f32 	%f1713, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter3+24];
	sub.f32 	%f1714, %f1713, %f1712;
	mad.f32 	%f445, %f1318, %f1714, %f1712;
	.loc	33	49	0
	ld.param.f32 	%f1715, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter3+28];
	ld.param.f32 	%f1303, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter3+4];
	ld.param.f32 	%f1716, [__cudaparm_PointwiseFilterHostKernel4___val_paraminFilter3+16];
	sub.f32 	%f1717, %f1716, %f1303;
	mad.f32 	%f1718, %f1318, %f1717, %f1303;
	sub.f32 	%f1719, %f1718, %f5;
	mad.f32 	%f5, %f1715, %f1719, %f5;
	.loc	33	50	0
	sub.f32 	%f1720, %f441, %f6;
	mad.f32 	%f6, %f1715, %f1720, %f6;
	.loc	33	51	0
	sub.f32 	%f1721, %f445, %f7;
	mad.f32 	%f7, %f1715, %f1721, %f7;
	.loc	35	103	0
	bra.uni 	$Lt_3_514;
$Lt_3_4610:
	.loc	34	41	0
	sub.s32 	%r932, %r13, %r12;
	sub.s32 	%r12, %r932, 1;
$Lt_3_514:
	.loc	17	100	0
	ld.param.s32 	%r933, [__cudaparm_PointwiseFilterHostKernel4_inDestPitch];
	mul24.lo.s32 	%r934, %r12, %r933;
	ld.param.u64 	%rd5, [__cudaparm_PointwiseFilterHostKernel4_inDestImage];
	add.s32 	%r935, %r934, %r6;
	cvt.s64.s32 	%rd6, %r935;
	mul.lo.u64 	%rd7, %rd6, 16;
	add.u64 	%rd8, %rd5, %rd7;
	st.global.v4.f32 	[%rd8+0], {%f5,%f6,%f7,%f8};
$Lt_3_206082:
	.loc	35	209	0
	exit;
$LDWend_PointwiseFilterHostKernel4:
	} // PointwiseFilterHostKernel4

