	.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_0001dbfc_00000000-9_RGBCurves.cpp3.i (C:/Users/dvaeng/AppData/Local/Temp/ccBI#.a21168)
	//-----------------------------------------------------------

	//-----------------------------------------------------------
	// 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_0001dbfc_00000000-8_RGBCurves.cudafe2.gpu"
	.file	2	"c:\scully64\shared\adobe\MediaCore\GPUFoundation\API\Inc\GPUFoundation/KernelSupport/PixelRGB.h"
	.file	3	"c:\scully64\shared\adobe\MediaCore\GPUFoundation\API\Inc\GPUFoundation/KernelSupport/PixelYUV.h"
	.file	4	"C:\Program Files (x86)\Microsoft Visual Studio 9.0\VC\include\crtdefs.h"
	.file	5	"c:\scully64\shared\adobe\MediaCore\External\3rdParty\NVIDIA\CUDA\win\64\include\crt/device_runtime.h"
	.file	6	"c:\scully64\shared\adobe\MediaCore\External\3rdParty\NVIDIA\CUDA\win\64\include\host_defines.h"
	.file	7	"c:\scully64\shared\adobe\MediaCore\External\3rdParty\NVIDIA\CUDA\win\64\include\builtin_types.h"
	.file	8	"c:\scully64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\64\include\device_types.h"
	.file	9	"c:\scully64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\64\include\driver_types.h"
	.file	10	"c:\scully64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\64\include\texture_types.h"
	.file	11	"c:\scully64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\64\include\vector_types.h"
	.file	12	"c:\scully64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\64\include\host_defines.h"
	.file	13	"c:\scully64\shared\adobe\MediaCore\External\3rdParty\NVIDIA\CUDA\win\64\include\device_launch_parameters.h"
	.file	14	"c:\scully64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\64\include\crt\storage_class.h"
	.file	15	"C:\Program Files (x86)\Microsoft Visual Studio 9.0\VC\include\time.h"
	.file	16	"c:\scully64\shared\adobe\MediaCore\GPUFoundation\API\Inc\GPUFoundation/KernelSupport/Utils.h"
	.file	17	"c:\scully64\shared\adobe\MediaCore\GPUFoundation\API\Inc\GPUFoundation/KernelSupport/VectorUtils.h"
	.file	18	"c:\scully64\shared\adobe\MediaCore\Display\Inc\CUDA/Effects/ColorCorrector_Curves.h"
	.file	19	"c:/scully64/shared/adobe/MediaCore/Display/Src/CUDA/Effects/RGBCurves.cu"
	.file	20	"c:\scully64\shared\adobe\MediaCore\GPUFoundation\API\Inc\GPUFoundation/Numeric.h"
	.file	21	"c:\scully64\shared\adobe\MediaCore\External\3rdParty\NVIDIA\CUDA\win\64\include\common_functions.h"
	.file	22	"c:\scully64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\64\include\crt/func_macro.h"
	.file	23	"c:\scully64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\64\include\math_functions.h"
	.file	24	"c:\scully64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\64\include\device_functions.h"
	.file	25	"c:\scully64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\64\include\math_constants.h"
	.file	26	"c:\scully64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\64\include\sm_11_atomic_functions.h"
	.file	27	"c:\scully64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\64\include\sm_12_atomic_functions.h"
	.file	28	"c:\scully64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\64\include\sm_13_double_functions.h"
	.file	29	"c:\scully64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\64\include\common_types.h"
	.file	30	"c:\scully64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\64\include\texture_fetch_functions.h"
	.file	31	"c:\scully64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\64\include\math_functions_dbl_ptx1.h"
	.file	32	"c:\scully64\shared\adobe\MediaCore\GPUFoundation\API\Inc\GPUFoundation/KernelSupport/ColorSpaceConvert.h"

	.shared .align 4 .b8 u__0[64];
	.shared .align 4 .b8 sMasterPointsX__2[64];
	.shared .align 4 .b8 sMasterPointsY__3[64];
	.shared .align 4 .b8 sMasterPointsY2__4[64];
	.shared .align 4 .b8 sRedPointsX__6[64];
	.shared .align 4 .b8 sRedPointsY__7[64];
	.shared .align 4 .b8 sRedPointsY2__8[64];
	.shared .align 4 .b8 sGreenPointsX__10[64];
	.shared .align 4 .b8 sGreenPointsY__11[64];
	.shared .align 4 .b8 sGreenPointsY2__12[64];
	.shared .align 4 .b8 sBluePointsX__14[64];
	.shared .align 4 .b8 sBluePointsY__15[64];
	.shared .align 4 .b8 sBluePointsY2__16[64];
	.shared .s32 sMasterPointCount__1;
	.shared .s32 sRedPointCount__5;
	.shared .s32 sGreenPointCount__9;
	.shared .s32 sBluePointCount__13;
	.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 RGBCurves_MaskKernel (
		.param .u64 __cudaparm_RGBCurves_MaskKernel_inImage,
		.param .s32 __cudaparm_RGBCurves_MaskKernel_inPitch,
		.param .u64 __cudaparm_RGBCurves_MaskKernel_inSecondaryMask,
		.param .s32 __cudaparm_RGBCurves_MaskKernel_inWidth,
		.param .s32 __cudaparm_RGBCurves_MaskKernel_inHeight)
	{
	.reg .u32 %r<25>;
	.reg .u64 %rd<10>;
	.reg .f32 %f<5>;
	.reg .pred %p<4>;
	.loc	19	391	0
$LBB1_RGBCurves_MaskKernel:
	.loc	19	394	0
	cvt.s32.u16 	%r1, %ctaid.x;
	cvt.s32.u16 	%r2, %ntid.x;
	mul24.lo.s32 	%r3, %r1, %r2;
	cvt.s32.u16 	%r4, %ctaid.y;
	cvt.s32.u16 	%r5, %ntid.y;
	mul24.lo.s32 	%r6, %r4, %r5;
	cvt.u32.u16 	%r7, %tid.x;
	add.u32 	%r8, %r3, %r7;
	cvt.u32.u16 	%r9, %tid.y;
	add.u32 	%r10, %r6, %r9;
	ld.param.s32 	%r11, [__cudaparm_RGBCurves_MaskKernel_inWidth];
	set.gt.u32.s32 	%r12, %r11, %r8;
	neg.s32 	%r13, %r12;
	ld.param.s32 	%r14, [__cudaparm_RGBCurves_MaskKernel_inHeight];
	set.gt.u32.s32 	%r15, %r14, %r10;
	neg.s32 	%r16, %r15;
	and.b32 	%r17, %r13, %r16;
	mov.u32 	%r18, 0;
	setp.eq.s32 	%p1, %r17, %r18;
	@%p1 bra 	$Lt_0_2050;
	.loc	19	397	0
	ld.param.s32 	%r19, [__cudaparm_RGBCurves_MaskKernel_inPitch];
	ld.param.u64 	%rd1, [__cudaparm_RGBCurves_MaskKernel_inSecondaryMask];
	mov.u64 	%rd2, 0;
	setp.eq.u64 	%p2, %rd1, %rd2;
	@%p2 bra 	$Lt_0_2818;
	.loc	16	89	0
	mul24.lo.s32 	%r20, %r10, %r19;
	add.s32 	%r21, %r20, %r8;
	cvt.s64.s32 	%rd3, %r21;
	mul.lo.u64 	%rd4, %rd3, 4;
	add.u64 	%rd5, %rd1, %rd4;
	ld.global.f32 	%f1, [%rd5+0];
	.loc	19	399	0
	mov.f32 	%f2, %f1;
	bra.uni 	$Lt_0_2562;
$Lt_0_2818:
	mov.f32 	%f2, 0f3f800000;     	// 1
	mul24.lo.s32 	%r22, %r10, %r19;
	add.s32 	%r23, %r22, %r8;
	cvt.s64.s32 	%rd3, %r23;
$Lt_0_2562:
	.loc	16	100	0
	ld.param.u64 	%rd6, [__cudaparm_RGBCurves_MaskKernel_inImage];
	mul.lo.u64 	%rd7, %rd3, 16;
	add.u64 	%rd8, %rd6, %rd7;
	mov.f32 	%f3, 0f3f800000;     	// 1
	st.global.v4.f32 	[%rd8+0], {%f2,%f2,%f2,%f3};
$Lt_0_2050:
	.loc	19	405	0
	exit;
$LDWend_RGBCurves_MaskKernel:
	} // RGBCurves_MaskKernel

	.entry RGBCurves_LumaKernel (
		.param .u64 __cudaparm_RGBCurves_LumaKernel_inImage,
		.param .s32 __cudaparm_RGBCurves_LumaKernel_inPitch,
		.param .u64 __cudaparm_RGBCurves_LumaKernel_inSecondaryMask,
		.param .s32 __cudaparm_RGBCurves_LumaKernel_inWidth,
		.param .s32 __cudaparm_RGBCurves_LumaKernel_inHeight,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inMasterPointX0,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inMasterPointX1,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inMasterPointX2,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inMasterPointX3,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inMasterPointX4,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inMasterPointX5,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inMasterPointX6,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inMasterPointX7,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inMasterPointX8,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inMasterPointX9,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inMasterPointX10,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inMasterPointX11,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inMasterPointX12,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inMasterPointX13,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inMasterPointX14,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inMasterPointX15,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inMasterPointY0,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inMasterPointY1,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inMasterPointY2,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inMasterPointY3,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inMasterPointY4,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inMasterPointY5,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inMasterPointY6,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inMasterPointY7,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inMasterPointY8,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inMasterPointY9,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inMasterPointY10,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inMasterPointY11,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inMasterPointY12,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inMasterPointY13,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inMasterPointY14,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inMasterPointY15,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inRedPointX0,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inRedPointX1,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inRedPointX2,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inRedPointX3,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inRedPointX4,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inRedPointX5,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inRedPointX6,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inRedPointX7,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inRedPointX8,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inRedPointX9,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inRedPointX10,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inRedPointX11,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inRedPointX12,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inRedPointX13,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inRedPointX14,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inRedPointX15,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inRedPointY0,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inRedPointY1,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inRedPointY2,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inRedPointY3,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inRedPointY4,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inRedPointY5,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inRedPointY6,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inRedPointY7,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inRedPointY8,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inRedPointY9,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inRedPointY10,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inRedPointY11,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inRedPointY12,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inRedPointY13,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inRedPointY14,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inRedPointY15,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inGreenPointX0,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inGreenPointX1,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inGreenPointX2,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inGreenPointX3,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inGreenPointX4,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inGreenPointX5,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inGreenPointX6,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inGreenPointX7,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inGreenPointX8,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inGreenPointX9,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inGreenPointX10,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inGreenPointX11,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inGreenPointX12,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inGreenPointX13,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inGreenPointX14,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inGreenPointX15,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inGreenPointY0,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inGreenPointY1,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inGreenPointY2,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inGreenPointY3,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inGreenPointY4,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inGreenPointY5,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inGreenPointY6,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inGreenPointY7,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inGreenPointY8,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inGreenPointY9,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inGreenPointY10,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inGreenPointY11,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inGreenPointY12,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inGreenPointY13,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inGreenPointY14,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inGreenPointY15,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inBluePointX0,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inBluePointX1,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inBluePointX2,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inBluePointX3,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inBluePointX4,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inBluePointX5,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inBluePointX6,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inBluePointX7,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inBluePointX8,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inBluePointX9,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inBluePointX10,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inBluePointX11,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inBluePointX12,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inBluePointX13,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inBluePointX14,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inBluePointX15,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inBluePointY0,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inBluePointY1,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inBluePointY2,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inBluePointY3,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inBluePointY4,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inBluePointY5,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inBluePointY6,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inBluePointY7,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inBluePointY8,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inBluePointY9,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inBluePointY10,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inBluePointY11,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inBluePointY12,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inBluePointY13,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inBluePointY14,
		.param .u8 __cudaparm_RGBCurves_LumaKernel_inBluePointY15)
	{
	.reg .u32 %r<223>;
	.reg .u64 %rd<165>;
	.reg .f32 %f<720>;
	.reg .pred %p<82>;
	.loc	19	543	0
$LBB1_RGBCurves_LumaKernel:
	.loc	19	546	0
	cvt.s32.u16 	%r1, %ctaid.x;
	cvt.s32.u16 	%r2, %ntid.x;
	mul24.lo.s32 	%r3, %r1, %r2;
	cvt.s32.u16 	%r4, %ctaid.y;
	cvt.s32.u16 	%r5, %ntid.y;
	mul24.lo.s32 	%r6, %r4, %r5;
	cvt.u32.u16 	%r7, %tid.x;
	add.u32 	%r8, %r3, %r7;
	cvt.u32.u16 	%r9, %tid.y;
	add.u32 	%r10, %r6, %r9;
	ld.param.s32 	%r11, [__cudaparm_RGBCurves_LumaKernel_inWidth];
	set.gt.u32.s32 	%r12, %r11, %r8;
	neg.s32 	%r13, %r12;
	ld.param.s32 	%r14, [__cudaparm_RGBCurves_LumaKernel_inHeight];
	set.gt.u32.s32 	%r15, %r14, %r10;
	neg.s32 	%r16, %r15;
	and.b32 	%r17, %r13, %r16;
	mov.u32 	%r18, 0;
	setp.eq.s32 	%p1, %r17, %r18;
	@%p1 bra 	$Lt_1_56066;
	.loc	16	89	0
	ld.param.s32 	%r19, [__cudaparm_RGBCurves_LumaKernel_inPitch];
	mul24.lo.s32 	%r20, %r10, %r19;
	add.s32 	%r21, %r20, %r8;
	cvt.s64.s32 	%rd1, %r21;
	ld.param.u64 	%rd2, [__cudaparm_RGBCurves_LumaKernel_inImage];
	mul.lo.u64 	%rd3, %rd1, 16;
	add.u64 	%rd4, %rd2, %rd3;
	ld.global.v4.f32 	{%f1,%f2,%f3,%f4}, [%rd4+0];
	.loc	19	549	0
	ld.param.u64 	%rd5, [__cudaparm_RGBCurves_LumaKernel_inSecondaryMask];
	mov.u64 	%rd6, 0;
	setp.eq.u64 	%p2, %rd5, %rd6;
	@%p2 bra 	$Lt_1_56834;
	.loc	16	89	0
	mul.lo.u64 	%rd7, %rd1, 4;
	add.u64 	%rd8, %rd5, %rd7;
	ld.global.f32 	%f5, [%rd8+0];
	.loc	19	550	0
	mov.f32 	%f6, %f5;
	bra.uni 	$Lt_1_56578;
$Lt_1_56834:
	mov.f32 	%f6, 0f3f800000;     	// 1
$Lt_1_56578:
	.loc	19	161	0
	mov.f32 	%f7, %f1;
	mov.f32 	%f8, %f2;
	mov.f32 	%f9, %f3;
	mov.u32 	%r22, 0;
	set.eq.u32.u32 	%r23, %r9, %r22;
	neg.s32 	%r24, %r23;
	mov.u32 	%r25, 0;
	set.eq.u32.u32 	%r26, %r7, %r25;
	neg.s32 	%r27, %r26;
	and.b32 	%r28, %r24, %r27;
	mov.u32 	%r29, 0;
	setp.eq.s32 	%p3, %r28, %r29;
	@%p3 bra 	$Lt_1_70914;
	.loc	19	186	0
	mov.u64 	%rd9, sMasterPointsX__2;
	mov.s32 	%r30, 16;
	st.shared.s32 	[sMasterPointCount__1], %r30;
	ld.param.u8 	%r31, [__cudaparm_RGBCurves_LumaKernel_inMasterPointX0];
	cvt.rn.f32.u32 	%f10, %r31;
	mov.f32 	%f11, 0f3b808081;    	// 0.00392157
	mul.f32 	%f12, %f10, %f11;
	st.shared.f32 	[sMasterPointsX__2+0], %f12;
	ld.param.u8 	%r32, [__cudaparm_RGBCurves_LumaKernel_inMasterPointX1];
	cvt.rn.f32.u32 	%f13, %r32;
	mov.f32 	%f14, 0f3b808081;    	// 0.00392157
	mul.f32 	%f15, %f13, %f14;
	st.shared.f32 	[sMasterPointsX__2+4], %f15;
	ld.param.u8 	%r33, [__cudaparm_RGBCurves_LumaKernel_inMasterPointX2];
	cvt.rn.f32.u32 	%f16, %r33;
	mov.f32 	%f17, 0f3b808081;    	// 0.00392157
	mul.f32 	%f18, %f16, %f17;
	st.shared.f32 	[sMasterPointsX__2+8], %f18;
	ld.param.u8 	%r34, [__cudaparm_RGBCurves_LumaKernel_inMasterPointX3];
	cvt.rn.f32.u32 	%f19, %r34;
	mov.f32 	%f20, 0f3b808081;    	// 0.00392157
	mul.f32 	%f21, %f19, %f20;
	st.shared.f32 	[sMasterPointsX__2+12], %f21;
	ld.param.u8 	%r35, [__cudaparm_RGBCurves_LumaKernel_inMasterPointX4];
	cvt.rn.f32.u32 	%f22, %r35;
	mov.f32 	%f23, 0f3b808081;    	// 0.00392157
	mul.f32 	%f24, %f22, %f23;
	st.shared.f32 	[sMasterPointsX__2+16], %f24;
	ld.param.u8 	%r36, [__cudaparm_RGBCurves_LumaKernel_inMasterPointX5];
	cvt.rn.f32.u32 	%f25, %r36;
	mov.f32 	%f26, 0f3b808081;    	// 0.00392157
	mul.f32 	%f27, %f25, %f26;
	st.shared.f32 	[sMasterPointsX__2+20], %f27;
	ld.param.u8 	%r37, [__cudaparm_RGBCurves_LumaKernel_inMasterPointX6];
	cvt.rn.f32.u32 	%f28, %r37;
	mov.f32 	%f29, 0f3b808081;    	// 0.00392157
	mul.f32 	%f30, %f28, %f29;
	st.shared.f32 	[sMasterPointsX__2+24], %f30;
	ld.param.u8 	%r38, [__cudaparm_RGBCurves_LumaKernel_inMasterPointX7];
	cvt.rn.f32.u32 	%f31, %r38;
	mov.f32 	%f32, 0f3b808081;    	// 0.00392157
	mul.f32 	%f33, %f31, %f32;
	st.shared.f32 	[sMasterPointsX__2+28], %f33;
	ld.param.u8 	%r39, [__cudaparm_RGBCurves_LumaKernel_inMasterPointX8];
	cvt.rn.f32.u32 	%f34, %r39;
	mov.f32 	%f35, 0f3b808081;    	// 0.00392157
	mul.f32 	%f36, %f34, %f35;
	st.shared.f32 	[sMasterPointsX__2+32], %f36;
	ld.param.u8 	%r40, [__cudaparm_RGBCurves_LumaKernel_inMasterPointX9];
	cvt.rn.f32.u32 	%f37, %r40;
	mov.f32 	%f38, 0f3b808081;    	// 0.00392157
	mul.f32 	%f39, %f37, %f38;
	st.shared.f32 	[sMasterPointsX__2+36], %f39;
	ld.param.u8 	%r41, [__cudaparm_RGBCurves_LumaKernel_inMasterPointX10];
	cvt.rn.f32.u32 	%f40, %r41;
	mov.f32 	%f41, 0f3b808081;    	// 0.00392157
	mul.f32 	%f42, %f40, %f41;
	st.shared.f32 	[sMasterPointsX__2+40], %f42;
	ld.param.u8 	%r42, [__cudaparm_RGBCurves_LumaKernel_inMasterPointX11];
	cvt.rn.f32.u32 	%f43, %r42;
	mov.f32 	%f44, 0f3b808081;    	// 0.00392157
	mul.f32 	%f45, %f43, %f44;
	st.shared.f32 	[sMasterPointsX__2+44], %f45;
	ld.param.u8 	%r43, [__cudaparm_RGBCurves_LumaKernel_inMasterPointX12];
	cvt.rn.f32.u32 	%f46, %r43;
	mov.f32 	%f47, 0f3b808081;    	// 0.00392157
	mul.f32 	%f48, %f46, %f47;
	st.shared.f32 	[sMasterPointsX__2+48], %f48;
	ld.param.u8 	%r44, [__cudaparm_RGBCurves_LumaKernel_inMasterPointX13];
	cvt.rn.f32.u32 	%f49, %r44;
	mov.f32 	%f50, 0f3b808081;    	// 0.00392157
	mul.f32 	%f51, %f49, %f50;
	st.shared.f32 	[sMasterPointsX__2+52], %f51;
	ld.param.u8 	%r45, [__cudaparm_RGBCurves_LumaKernel_inMasterPointX14];
	cvt.rn.f32.u32 	%f52, %r45;
	mov.f32 	%f53, 0f3b808081;    	// 0.00392157
	mul.f32 	%f54, %f52, %f53;
	st.shared.f32 	[sMasterPointsX__2+56], %f54;
	ld.param.u8 	%r46, [__cudaparm_RGBCurves_LumaKernel_inMasterPointX15];
	cvt.rn.f32.u32 	%f55, %r46;
	mov.f32 	%f56, 0f3b808081;    	// 0.00392157
	mul.f32 	%f57, %f55, %f56;
	st.shared.f32 	[sMasterPointsX__2+60], %f57;
	ld.param.u8 	%r47, [__cudaparm_RGBCurves_LumaKernel_inMasterPointY0];
	cvt.rn.f32.u32 	%f58, %r47;
	mov.f32 	%f59, 0f3b808081;    	// 0.00392157
	mul.f32 	%f60, %f58, %f59;
	st.shared.f32 	[sMasterPointsY__3+0], %f60;
	ld.param.u8 	%r48, [__cudaparm_RGBCurves_LumaKernel_inMasterPointY1];
	cvt.rn.f32.u32 	%f61, %r48;
	mov.f32 	%f62, 0f3b808081;    	// 0.00392157
	mul.f32 	%f63, %f61, %f62;
	st.shared.f32 	[sMasterPointsY__3+4], %f63;
	ld.param.u8 	%r49, [__cudaparm_RGBCurves_LumaKernel_inMasterPointY2];
	cvt.rn.f32.u32 	%f64, %r49;
	mov.f32 	%f65, 0f3b808081;    	// 0.00392157
	mul.f32 	%f66, %f64, %f65;
	st.shared.f32 	[sMasterPointsY__3+8], %f66;
	ld.param.u8 	%r50, [__cudaparm_RGBCurves_LumaKernel_inMasterPointY3];
	cvt.rn.f32.u32 	%f67, %r50;
	mov.f32 	%f68, 0f3b808081;    	// 0.00392157
	mul.f32 	%f69, %f67, %f68;
	st.shared.f32 	[sMasterPointsY__3+12], %f69;
	ld.param.u8 	%r51, [__cudaparm_RGBCurves_LumaKernel_inMasterPointY4];
	cvt.rn.f32.u32 	%f70, %r51;
	mov.f32 	%f71, 0f3b808081;    	// 0.00392157
	mul.f32 	%f72, %f70, %f71;
	st.shared.f32 	[sMasterPointsY__3+16], %f72;
	ld.param.u8 	%r52, [__cudaparm_RGBCurves_LumaKernel_inMasterPointY5];
	cvt.rn.f32.u32 	%f73, %r52;
	mov.f32 	%f74, 0f3b808081;    	// 0.00392157
	mul.f32 	%f75, %f73, %f74;
	st.shared.f32 	[sMasterPointsY__3+20], %f75;
	ld.param.u8 	%r53, [__cudaparm_RGBCurves_LumaKernel_inMasterPointY6];
	cvt.rn.f32.u32 	%f76, %r53;
	mov.f32 	%f77, 0f3b808081;    	// 0.00392157
	mul.f32 	%f78, %f76, %f77;
	st.shared.f32 	[sMasterPointsY__3+24], %f78;
	ld.param.u8 	%r54, [__cudaparm_RGBCurves_LumaKernel_inMasterPointY7];
	cvt.rn.f32.u32 	%f79, %r54;
	mov.f32 	%f80, 0f3b808081;    	// 0.00392157
	mul.f32 	%f81, %f79, %f80;
	st.shared.f32 	[sMasterPointsY__3+28], %f81;
	ld.param.u8 	%r55, [__cudaparm_RGBCurves_LumaKernel_inMasterPointY8];
	cvt.rn.f32.u32 	%f82, %r55;
	mov.f32 	%f83, 0f3b808081;    	// 0.00392157
	mul.f32 	%f84, %f82, %f83;
	st.shared.f32 	[sMasterPointsY__3+32], %f84;
	ld.param.u8 	%r56, [__cudaparm_RGBCurves_LumaKernel_inMasterPointY9];
	cvt.rn.f32.u32 	%f85, %r56;
	mov.f32 	%f86, 0f3b808081;    	// 0.00392157
	mul.f32 	%f87, %f85, %f86;
	st.shared.f32 	[sMasterPointsY__3+36], %f87;
	ld.param.u8 	%r57, [__cudaparm_RGBCurves_LumaKernel_inMasterPointY10];
	cvt.rn.f32.u32 	%f88, %r57;
	mov.f32 	%f89, 0f3b808081;    	// 0.00392157
	mul.f32 	%f90, %f88, %f89;
	st.shared.f32 	[sMasterPointsY__3+40], %f90;
	ld.param.u8 	%r58, [__cudaparm_RGBCurves_LumaKernel_inMasterPointY11];
	cvt.rn.f32.u32 	%f91, %r58;
	mov.f32 	%f92, 0f3b808081;    	// 0.00392157
	mul.f32 	%f93, %f91, %f92;
	st.shared.f32 	[sMasterPointsY__3+44], %f93;
	ld.param.u8 	%r59, [__cudaparm_RGBCurves_LumaKernel_inMasterPointY12];
	cvt.rn.f32.u32 	%f94, %r59;
	mov.f32 	%f95, 0f3b808081;    	// 0.00392157
	mul.f32 	%f96, %f94, %f95;
	st.shared.f32 	[sMasterPointsY__3+48], %f96;
	ld.param.u8 	%r60, [__cudaparm_RGBCurves_LumaKernel_inMasterPointY13];
	cvt.rn.f32.u32 	%f97, %r60;
	mov.f32 	%f98, 0f3b808081;    	// 0.00392157
	mul.f32 	%f99, %f97, %f98;
	st.shared.f32 	[sMasterPointsY__3+52], %f99;
	ld.param.u8 	%r61, [__cudaparm_RGBCurves_LumaKernel_inMasterPointY14];
	cvt.rn.f32.u32 	%f100, %r61;
	mov.f32 	%f101, 0f3b808081;   	// 0.00392157
	mul.f32 	%f102, %f100, %f101;
	st.shared.f32 	[sMasterPointsY__3+56], %f102;
	ld.param.u8 	%r62, [__cudaparm_RGBCurves_LumaKernel_inMasterPointY15];
	cvt.rn.f32.u32 	%f103, %r62;
	mov.f32 	%f104, 0f3b808081;   	// 0.00392157
	mul.f32 	%f105, %f103, %f104;
	st.shared.f32 	[sMasterPointsY__3+60], %f105;
	mov.u64 	%rd10, 0;
	mov.s64 	%rd11, %rd9;
	mov.s32 	%r63, 0;
	mov.u64 	%rd12, sMasterPointsY__3;
$Lt_1_58114:
 //<loop> Loop body line 186, nesting depth: 1, iterations: 16
	.loc	19	221	0
	ld.shared.f32 	%f106, [%rd11+0];
	mov.f32 	%f107, 0f00000000;   	// 0
	setp.neu.f32 	%p4, %f106, %f107;
	@%p4 bra 	$L_1_52226;
 //<loop> Part of loop body line 186, head labeled $Lt_1_58114
	add.u64 	%rd13, %rd10, %rd12;
	ld.shared.f32 	%f108, [%rd13+0];
	mov.f32 	%f109, 0f00000000;   	// 0
	setp.neu.f32 	%p5, %f108, %f109;
	@!%p5 bra 	$L_1_51970;
$L_1_52226:
 //<loop> Part of loop body line 186, head labeled $Lt_1_58114
	.loc	19	223	0
	add.s32 	%r30, %r63, 1;
	st.shared.s32 	[sMasterPointCount__1], %r30;
$L_1_51970:
 //<loop> Part of loop body line 186, head labeled $Lt_1_58114
	add.s32 	%r64, %r63, 1;
	mov.s32 	%r63, %r64;
	add.u64 	%rd10, %rd10, 4;
	add.u64 	%rd11, %rd11, 4;
	mov.u32 	%r65, 16;
	setp.ne.s32 	%p6, %r64, %r65;
	@%p6 bra 	$Lt_1_58114;
	.loc	19	226	0
	mov.f32 	%f110, 0f00000000;   	// 0
	st.shared.f32 	[u__0+0], %f110;
	.loc	18	48	0
	mov.f32 	%f111, 0f00000000;   	// 0
	st.shared.f32 	[sMasterPointsY2__4+0], %f111;
	cvt.u64.s32 	%rd14, %r30;
	mov.s64 	%rd15, %rd14;
	cvt.u32.u64 	%r66, %rd15;
	sub.s32 	%r67, %r66, 1;
	sub.s32 	%r68, %r66, 2;
	mov.u32 	%r69, 1;
	setp.le.s32 	%p7, %r67, %r69;
	@%p7 bra 	$Lt_1_58626;
	mov.s32 	%r70, %r68;
	mov.u64 	%rd16, 4;
	mov.s32 	%r71, 1;
	mov.u64 	%rd17, sMasterPointsY2__4;
	mov.u64 	%rd18, u__0;
	mov.s32 	%r72, %r70;
$Lt_1_59138:
 //<loop> Loop body line 48, nesting depth: 1, estimated iterations: unknown
	.loc	18	53	0
	add.u64 	%rd19, %rd16, %rd18;
	mov.f32 	%f112, 0f00000000;   	// 0
	st.shared.f32 	[%rd19+0], %f112;
	add.u64 	%rd20, %rd16, %rd17;
	mov.f32 	%f113, 0f00000000;   	// 0
	st.shared.f32 	[%rd20+0], %f113;
	add.u64 	%rd21, %rd16, %rd9;
	ld.shared.f32 	%f114, [%rd21+-4];
	ld.shared.f32 	%f115, [%rd21+0];
	setp.lt.f32 	%p8, %f114, %f115;
	@!%p8 bra 	$Lt_1_91650;
 //<loop> Part of loop body line 48, head labeled $Lt_1_59138
	ld.shared.f32 	%f116, [%rd21+4];
	ld.shared.f32 	%f115, [%rd21+0];
	setp.lt.f32 	%p9, %f115, %f116;
	@!%p9 bra 	$Lt_1_91650;
 //<loop> Part of loop body line 48, head labeled $Lt_1_59138
	ld.shared.f32 	%f115, [%rd21+0];
	ld.shared.f32 	%f114, [%rd21+-4];
	.loc	18	56	0
	sub.f32 	%f117, %f115, %f114;
	.loc	18	53	0
	ld.shared.f32 	%f116, [%rd21+4];
	.loc	18	56	0
	sub.f32 	%f118, %f116, %f114;
	.loc	18	57	0
	div.full.f32 	%f119, %f117, %f118;
	mov.f32 	%f120, 0f40000000;   	// 2
	ld.shared.f32 	%f121, [%rd20+-4];
	mad.f32 	%f122, %f121, %f119, %f120;
	mov.f32 	%f123, 0f00000000;   	// 0
	setp.neu.f32 	%p10, %f122, %f123;
	@!%p10 bra 	$Lt_1_91650;
 //<loop> Part of loop body line 48, head labeled $Lt_1_59138
	.loc	18	60	0
	mov.f32 	%f124, 0fbf800000;   	// -1
	add.f32 	%f125, %f119, %f124;
	div.full.f32 	%f126, %f125, %f122;
	st.shared.f32 	[%rd20+0], %f126;
	.loc	18	61	0
	add.u64 	%rd22, %rd16, %rd12;
	ld.shared.f32 	%f127, [%rd22+0];
	ld.shared.f32 	%f128, [%rd22+4];
	sub.f32 	%f129, %f128, %f127;
	sub.f32 	%f130, %f116, %f115;
	ld.shared.f32 	%f131, [%rd22+-4];
	sub.f32 	%f132, %f127, %f131;
	div.full.f32 	%f133, %f129, %f130;
	div.full.f32 	%f134, %f132, %f117;
	sub.f32 	%f135, %f133, %f134;
	st.shared.f32 	[%rd19+0], %f135;
	.loc	18	62	0
	mov.f32 	%f136, 0f40c00000;   	// 6
	mul.f32 	%f137, %f135, %f136;
	div.full.f32 	%f138, %f137, %f118;
	ld.shared.f32 	%f139, [%rd19+-4];
	mul.f32 	%f140, %f139, %f119;
	sub.f32 	%f141, %f138, %f140;
	div.full.f32 	%f142, %f141, %f122;
	st.shared.f32 	[%rd19+0], %f142;
$Lt_1_91650:
$L_1_52482:
 //<loop> Part of loop body line 48, head labeled $Lt_1_59138
	.loc	18	56	0
	add.s32 	%r71, %r71, 1;
	add.u64 	%rd16, %rd16, 4;
	setp.ne.s32 	%p11, %r67, %r71;
	@%p11 bra 	$Lt_1_59138;
$Lt_1_58626:
	mov.u64 	%rd18, u__0;
	mov.u64 	%rd17, sMasterPointsY2__4;
	.loc	18	69	0
	mov.f32 	%f143, 0f00000000;   	// 0
	mul.lo.u64 	%rd23, %rd14, 4;
	add.u64 	%rd24, %rd17, %rd23;
	st.shared.f32 	[%rd24+-4], %f143;
	.loc	18	72	0
	mov.s32 	%r73, %r68;
	mov.u32 	%r74, 0;
	setp.lt.s32 	%p12, %r68, %r74;
	@%p12 bra 	$Lt_1_60162;
	mov.s32 	%r75, %r67;
	cvt.u64.s32 	%rd25, %r68;
	mul.lo.u64 	%rd26, %rd25, 4;
	add.u64 	%rd27, %rd26, %rd18;
	add.u64 	%rd28, %rd26, %rd17;
	mov.s32 	%r76, %r75;
$Lt_1_60674:
 //<loop> Loop body line 72, nesting depth: 1, estimated iterations: unknown
	.loc	18	74	0
	ld.shared.f32 	%f144, [%rd27+0];
	ld.shared.f32 	%f145, [%rd28+0];
	ld.shared.f32 	%f146, [%rd28+4];
	mad.f32 	%f147, %f145, %f146, %f144;
	st.shared.f32 	[%rd28+0], %f147;
	sub.s32 	%r73, %r73, 1;
	sub.u64 	%rd28, %rd28, 4;
	sub.u64 	%rd27, %rd27, 4;
	mov.u32 	%r77, -1;
	setp.ne.s32 	%p13, %r73, %r77;
	@%p13 bra 	$Lt_1_60674;
$Lt_1_60162:
	.loc	19	226	0
	mov.u64 	%rd29, sRedPointsX__6;
	.loc	19	229	0
	mov.s32 	%r78, 16;
	st.shared.s32 	[sRedPointCount__5], %r78;
	ld.param.u8 	%r79, [__cudaparm_RGBCurves_LumaKernel_inRedPointX0];
	cvt.rn.f32.u32 	%f148, %r79;
	mov.f32 	%f149, 0f3b808081;   	// 0.00392157
	mul.f32 	%f150, %f148, %f149;
	st.shared.f32 	[sRedPointsX__6+0], %f150;
	ld.param.u8 	%r80, [__cudaparm_RGBCurves_LumaKernel_inRedPointX1];
	cvt.rn.f32.u32 	%f151, %r80;
	mov.f32 	%f152, 0f3b808081;   	// 0.00392157
	mul.f32 	%f153, %f151, %f152;
	st.shared.f32 	[sRedPointsX__6+4], %f153;
	ld.param.u8 	%r81, [__cudaparm_RGBCurves_LumaKernel_inRedPointX2];
	cvt.rn.f32.u32 	%f154, %r81;
	mov.f32 	%f155, 0f3b808081;   	// 0.00392157
	mul.f32 	%f156, %f154, %f155;
	st.shared.f32 	[sRedPointsX__6+8], %f156;
	ld.param.u8 	%r82, [__cudaparm_RGBCurves_LumaKernel_inRedPointX3];
	cvt.rn.f32.u32 	%f157, %r82;
	mov.f32 	%f158, 0f3b808081;   	// 0.00392157
	mul.f32 	%f159, %f157, %f158;
	st.shared.f32 	[sRedPointsX__6+12], %f159;
	ld.param.u8 	%r83, [__cudaparm_RGBCurves_LumaKernel_inRedPointX4];
	cvt.rn.f32.u32 	%f160, %r83;
	mov.f32 	%f161, 0f3b808081;   	// 0.00392157
	mul.f32 	%f162, %f160, %f161;
	st.shared.f32 	[sRedPointsX__6+16], %f162;
	ld.param.u8 	%r84, [__cudaparm_RGBCurves_LumaKernel_inRedPointX5];
	cvt.rn.f32.u32 	%f163, %r84;
	mov.f32 	%f164, 0f3b808081;   	// 0.00392157
	mul.f32 	%f165, %f163, %f164;
	st.shared.f32 	[sRedPointsX__6+20], %f165;
	ld.param.u8 	%r85, [__cudaparm_RGBCurves_LumaKernel_inRedPointX6];
	cvt.rn.f32.u32 	%f166, %r85;
	mov.f32 	%f167, 0f3b808081;   	// 0.00392157
	mul.f32 	%f168, %f166, %f167;
	st.shared.f32 	[sRedPointsX__6+24], %f168;
	ld.param.u8 	%r86, [__cudaparm_RGBCurves_LumaKernel_inRedPointX7];
	cvt.rn.f32.u32 	%f169, %r86;
	mov.f32 	%f170, 0f3b808081;   	// 0.00392157
	mul.f32 	%f171, %f169, %f170;
	st.shared.f32 	[sRedPointsX__6+28], %f171;
	ld.param.u8 	%r87, [__cudaparm_RGBCurves_LumaKernel_inRedPointX8];
	cvt.rn.f32.u32 	%f172, %r87;
	mov.f32 	%f173, 0f3b808081;   	// 0.00392157
	mul.f32 	%f174, %f172, %f173;
	st.shared.f32 	[sRedPointsX__6+32], %f174;
	ld.param.u8 	%r88, [__cudaparm_RGBCurves_LumaKernel_inRedPointX9];
	cvt.rn.f32.u32 	%f175, %r88;
	mov.f32 	%f176, 0f3b808081;   	// 0.00392157
	mul.f32 	%f177, %f175, %f176;
	st.shared.f32 	[sRedPointsX__6+36], %f177;
	ld.param.u8 	%r89, [__cudaparm_RGBCurves_LumaKernel_inRedPointX10];
	cvt.rn.f32.u32 	%f178, %r89;
	mov.f32 	%f179, 0f3b808081;   	// 0.00392157
	mul.f32 	%f180, %f178, %f179;
	st.shared.f32 	[sRedPointsX__6+40], %f180;
	ld.param.u8 	%r90, [__cudaparm_RGBCurves_LumaKernel_inRedPointX11];
	cvt.rn.f32.u32 	%f181, %r90;
	mov.f32 	%f182, 0f3b808081;   	// 0.00392157
	mul.f32 	%f183, %f181, %f182;
	st.shared.f32 	[sRedPointsX__6+44], %f183;
	ld.param.u8 	%r91, [__cudaparm_RGBCurves_LumaKernel_inRedPointX12];
	cvt.rn.f32.u32 	%f184, %r91;
	mov.f32 	%f185, 0f3b808081;   	// 0.00392157
	mul.f32 	%f186, %f184, %f185;
	st.shared.f32 	[sRedPointsX__6+48], %f186;
	ld.param.u8 	%r92, [__cudaparm_RGBCurves_LumaKernel_inRedPointX13];
	cvt.rn.f32.u32 	%f187, %r92;
	mov.f32 	%f188, 0f3b808081;   	// 0.00392157
	mul.f32 	%f189, %f187, %f188;
	st.shared.f32 	[sRedPointsX__6+52], %f189;
	ld.param.u8 	%r93, [__cudaparm_RGBCurves_LumaKernel_inRedPointX14];
	cvt.rn.f32.u32 	%f190, %r93;
	mov.f32 	%f191, 0f3b808081;   	// 0.00392157
	mul.f32 	%f192, %f190, %f191;
	st.shared.f32 	[sRedPointsX__6+56], %f192;
	ld.param.u8 	%r94, [__cudaparm_RGBCurves_LumaKernel_inRedPointX15];
	cvt.rn.f32.u32 	%f193, %r94;
	mov.f32 	%f194, 0f3b808081;   	// 0.00392157
	mul.f32 	%f195, %f193, %f194;
	st.shared.f32 	[sRedPointsX__6+60], %f195;
	ld.param.u8 	%r95, [__cudaparm_RGBCurves_LumaKernel_inRedPointY0];
	cvt.rn.f32.u32 	%f196, %r95;
	mov.f32 	%f197, 0f3b808081;   	// 0.00392157
	mul.f32 	%f198, %f196, %f197;
	st.shared.f32 	[sRedPointsY__7+0], %f198;
	ld.param.u8 	%r96, [__cudaparm_RGBCurves_LumaKernel_inRedPointY1];
	cvt.rn.f32.u32 	%f199, %r96;
	mov.f32 	%f200, 0f3b808081;   	// 0.00392157
	mul.f32 	%f201, %f199, %f200;
	st.shared.f32 	[sRedPointsY__7+4], %f201;
	ld.param.u8 	%r97, [__cudaparm_RGBCurves_LumaKernel_inRedPointY2];
	cvt.rn.f32.u32 	%f202, %r97;
	mov.f32 	%f203, 0f3b808081;   	// 0.00392157
	mul.f32 	%f204, %f202, %f203;
	st.shared.f32 	[sRedPointsY__7+8], %f204;
	ld.param.u8 	%r98, [__cudaparm_RGBCurves_LumaKernel_inRedPointY3];
	cvt.rn.f32.u32 	%f205, %r98;
	mov.f32 	%f206, 0f3b808081;   	// 0.00392157
	mul.f32 	%f207, %f205, %f206;
	st.shared.f32 	[sRedPointsY__7+12], %f207;
	ld.param.u8 	%r99, [__cudaparm_RGBCurves_LumaKernel_inRedPointY4];
	cvt.rn.f32.u32 	%f208, %r99;
	mov.f32 	%f209, 0f3b808081;   	// 0.00392157
	mul.f32 	%f210, %f208, %f209;
	st.shared.f32 	[sRedPointsY__7+16], %f210;
	ld.param.u8 	%r100, [__cudaparm_RGBCurves_LumaKernel_inRedPointY5];
	cvt.rn.f32.u32 	%f211, %r100;
	mov.f32 	%f212, 0f3b808081;   	// 0.00392157
	mul.f32 	%f213, %f211, %f212;
	st.shared.f32 	[sRedPointsY__7+20], %f213;
	ld.param.u8 	%r101, [__cudaparm_RGBCurves_LumaKernel_inRedPointY6];
	cvt.rn.f32.u32 	%f214, %r101;
	mov.f32 	%f215, 0f3b808081;   	// 0.00392157
	mul.f32 	%f216, %f214, %f215;
	st.shared.f32 	[sRedPointsY__7+24], %f216;
	ld.param.u8 	%r102, [__cudaparm_RGBCurves_LumaKernel_inRedPointY7];
	cvt.rn.f32.u32 	%f217, %r102;
	mov.f32 	%f218, 0f3b808081;   	// 0.00392157
	mul.f32 	%f219, %f217, %f218;
	st.shared.f32 	[sRedPointsY__7+28], %f219;
	ld.param.u8 	%r103, [__cudaparm_RGBCurves_LumaKernel_inRedPointY8];
	cvt.rn.f32.u32 	%f220, %r103;
	mov.f32 	%f221, 0f3b808081;   	// 0.00392157
	mul.f32 	%f222, %f220, %f221;
	st.shared.f32 	[sRedPointsY__7+32], %f222;
	ld.param.u8 	%r104, [__cudaparm_RGBCurves_LumaKernel_inRedPointY9];
	cvt.rn.f32.u32 	%f223, %r104;
	mov.f32 	%f224, 0f3b808081;   	// 0.00392157
	mul.f32 	%f225, %f223, %f224;
	st.shared.f32 	[sRedPointsY__7+36], %f225;
	ld.param.u8 	%r105, [__cudaparm_RGBCurves_LumaKernel_inRedPointY10];
	cvt.rn.f32.u32 	%f226, %r105;
	mov.f32 	%f227, 0f3b808081;   	// 0.00392157
	mul.f32 	%f228, %f226, %f227;
	st.shared.f32 	[sRedPointsY__7+40], %f228;
	ld.param.u8 	%r106, [__cudaparm_RGBCurves_LumaKernel_inRedPointY11];
	cvt.rn.f32.u32 	%f229, %r106;
	mov.f32 	%f230, 0f3b808081;   	// 0.00392157
	mul.f32 	%f231, %f229, %f230;
	st.shared.f32 	[sRedPointsY__7+44], %f231;
	ld.param.u8 	%r107, [__cudaparm_RGBCurves_LumaKernel_inRedPointY12];
	cvt.rn.f32.u32 	%f232, %r107;
	mov.f32 	%f233, 0f3b808081;   	// 0.00392157
	mul.f32 	%f234, %f232, %f233;
	st.shared.f32 	[sRedPointsY__7+48], %f234;
	ld.param.u8 	%r108, [__cudaparm_RGBCurves_LumaKernel_inRedPointY13];
	cvt.rn.f32.u32 	%f235, %r108;
	mov.f32 	%f236, 0f3b808081;   	// 0.00392157
	mul.f32 	%f237, %f235, %f236;
	st.shared.f32 	[sRedPointsY__7+52], %f237;
	ld.param.u8 	%r109, [__cudaparm_RGBCurves_LumaKernel_inRedPointY14];
	cvt.rn.f32.u32 	%f238, %r109;
	mov.f32 	%f239, 0f3b808081;   	// 0.00392157
	mul.f32 	%f240, %f238, %f239;
	st.shared.f32 	[sRedPointsY__7+56], %f240;
	ld.param.u8 	%r110, [__cudaparm_RGBCurves_LumaKernel_inRedPointY15];
	cvt.rn.f32.u32 	%f241, %r110;
	mov.f32 	%f242, 0f3b808081;   	// 0.00392157
	mul.f32 	%f243, %f241, %f242;
	st.shared.f32 	[sRedPointsY__7+60], %f243;
	mov.u64 	%rd30, 0;
	mov.s64 	%rd31, %rd29;
	mov.s32 	%r111, 0;
	mov.u64 	%rd32, sRedPointsY__7;
$Lt_1_61698:
 //<loop> Loop body line 229, nesting depth: 1, iterations: 16
	.loc	19	264	0
	ld.shared.f32 	%f244, [%rd31+0];
	mov.f32 	%f245, 0f00000000;   	// 0
	setp.neu.f32 	%p14, %f244, %f245;
	@%p14 bra 	$L_1_53250;
 //<loop> Part of loop body line 229, head labeled $Lt_1_61698
	add.u64 	%rd33, %rd30, %rd32;
	ld.shared.f32 	%f246, [%rd33+0];
	mov.f32 	%f247, 0f00000000;   	// 0
	setp.neu.f32 	%p15, %f246, %f247;
	@!%p15 bra 	$L_1_52994;
$L_1_53250:
 //<loop> Part of loop body line 229, head labeled $Lt_1_61698
	.loc	19	266	0
	add.s32 	%r78, %r111, 1;
	st.shared.s32 	[sRedPointCount__5], %r78;
$L_1_52994:
 //<loop> Part of loop body line 229, head labeled $Lt_1_61698
	add.s32 	%r112, %r111, 1;
	mov.s32 	%r111, %r112;
	add.u64 	%rd30, %rd30, 4;
	add.u64 	%rd31, %rd31, 4;
	mov.u32 	%r113, 16;
	setp.ne.s32 	%p16, %r112, %r113;
	@%p16 bra 	$Lt_1_61698;
	.loc	19	269	0
	mov.f32 	%f248, 0f00000000;   	// 0
	st.shared.f32 	[u__0+0], %f248;
	.loc	18	48	0
	mov.f32 	%f249, 0f00000000;   	// 0
	st.shared.f32 	[sRedPointsY2__8+0], %f249;
	cvt.u64.s32 	%rd34, %r78;
	mov.s64 	%rd35, %rd34;
	cvt.u32.u64 	%r114, %rd35;
	sub.s32 	%r115, %r114, 1;
	sub.s32 	%r116, %r114, 2;
	mov.u32 	%r117, 1;
	setp.le.s32 	%p17, %r115, %r117;
	@%p17 bra 	$Lt_1_62210;
	mov.s32 	%r118, %r116;
	mov.u64 	%rd16, 4;
	mov.s32 	%r71, 1;
	mov.u64 	%rd36, sRedPointsY2__8;
	mov.s32 	%r119, %r118;
$Lt_1_62722:
 //<loop> Loop body line 48, nesting depth: 1, estimated iterations: unknown
	.loc	18	53	0
	add.u64 	%rd19, %rd16, %rd18;
	mov.f32 	%f250, 0f00000000;   	// 0
	st.shared.f32 	[%rd19+0], %f250;
	add.u64 	%rd37, %rd16, %rd36;
	mov.f32 	%f251, 0f00000000;   	// 0
	st.shared.f32 	[%rd37+0], %f251;
	add.u64 	%rd38, %rd16, %rd29;
	ld.shared.f32 	%f252, [%rd38+-4];
	ld.shared.f32 	%f253, [%rd38+0];
	setp.lt.f32 	%p18, %f252, %f253;
	@!%p18 bra 	$Lt_1_93186;
 //<loop> Part of loop body line 48, head labeled $Lt_1_62722
	ld.shared.f32 	%f254, [%rd38+4];
	ld.shared.f32 	%f253, [%rd38+0];
	setp.lt.f32 	%p19, %f253, %f254;
	@!%p19 bra 	$Lt_1_93186;
 //<loop> Part of loop body line 48, head labeled $Lt_1_62722
	ld.shared.f32 	%f253, [%rd38+0];
	ld.shared.f32 	%f252, [%rd38+-4];
	.loc	18	56	0
	sub.f32 	%f255, %f253, %f252;
	.loc	18	53	0
	ld.shared.f32 	%f254, [%rd38+4];
	.loc	18	56	0
	sub.f32 	%f256, %f254, %f252;
	.loc	18	57	0
	div.full.f32 	%f119, %f255, %f256;
	mov.f32 	%f257, 0f40000000;   	// 2
	ld.shared.f32 	%f258, [%rd37+-4];
	mad.f32 	%f122, %f258, %f119, %f257;
	mov.f32 	%f259, 0f00000000;   	// 0
	setp.neu.f32 	%p20, %f122, %f259;
	@!%p20 bra 	$Lt_1_93186;
 //<loop> Part of loop body line 48, head labeled $Lt_1_62722
	.loc	18	60	0
	mov.f32 	%f260, 0fbf800000;   	// -1
	add.f32 	%f261, %f119, %f260;
	div.full.f32 	%f262, %f261, %f122;
	st.shared.f32 	[%rd37+0], %f262;
	.loc	18	61	0
	add.u64 	%rd39, %rd16, %rd32;
	ld.shared.f32 	%f263, [%rd39+0];
	ld.shared.f32 	%f264, [%rd39+4];
	sub.f32 	%f129, %f264, %f263;
	sub.f32 	%f130, %f254, %f253;
	ld.shared.f32 	%f265, [%rd39+-4];
	sub.f32 	%f132, %f263, %f265;
	div.full.f32 	%f133, %f129, %f130;
	div.full.f32 	%f134, %f132, %f255;
	sub.f32 	%f135, %f133, %f134;
	st.shared.f32 	[%rd19+0], %f135;
	.loc	18	62	0
	mov.f32 	%f266, 0f40c00000;   	// 6
	mul.f32 	%f267, %f135, %f266;
	div.full.f32 	%f268, %f267, %f256;
	ld.shared.f32 	%f269, [%rd19+-4];
	mul.f32 	%f270, %f269, %f119;
	sub.f32 	%f271, %f268, %f270;
	div.full.f32 	%f272, %f271, %f122;
	st.shared.f32 	[%rd19+0], %f272;
$Lt_1_93186:
$L_1_53506:
 //<loop> Part of loop body line 48, head labeled $Lt_1_62722
	.loc	18	56	0
	add.s32 	%r71, %r71, 1;
	add.u64 	%rd16, %rd16, 4;
	setp.ne.s32 	%p21, %r115, %r71;
	@%p21 bra 	$Lt_1_62722;
$Lt_1_62210:
	mov.u64 	%rd36, sRedPointsY2__8;
	.loc	18	69	0
	mov.f32 	%f273, 0f00000000;   	// 0
	mul.lo.u64 	%rd40, %rd34, 4;
	add.u64 	%rd41, %rd36, %rd40;
	st.shared.f32 	[%rd41+-4], %f273;
	.loc	18	72	0
	mov.s32 	%r73, %r116;
	mov.u32 	%r120, 0;
	setp.lt.s32 	%p22, %r116, %r120;
	@%p22 bra 	$Lt_1_63746;
	mov.s32 	%r121, %r115;
	cvt.u64.s32 	%rd42, %r116;
	mul.lo.u64 	%rd26, %rd42, 4;
	add.u64 	%rd27, %rd26, %rd18;
	add.u64 	%rd43, %rd26, %rd36;
	mov.s32 	%r122, %r121;
$Lt_1_64258:
 //<loop> Loop body line 72, nesting depth: 1, estimated iterations: unknown
	.loc	18	74	0
	ld.shared.f32 	%f274, [%rd27+0];
	ld.shared.f32 	%f275, [%rd43+0];
	ld.shared.f32 	%f276, [%rd43+4];
	mad.f32 	%f277, %f275, %f276, %f274;
	st.shared.f32 	[%rd43+0], %f277;
	sub.s32 	%r73, %r73, 1;
	sub.u64 	%rd43, %rd43, 4;
	sub.u64 	%rd27, %rd27, 4;
	mov.u32 	%r123, -1;
	setp.ne.s32 	%p23, %r73, %r123;
	@%p23 bra 	$Lt_1_64258;
$Lt_1_63746:
	.loc	19	269	0
	mov.u64 	%rd44, sGreenPointsX__10;
	.loc	19	272	0
	mov.s32 	%r124, 16;
	st.shared.s32 	[sGreenPointCount__9], %r124;
	ld.param.u8 	%r125, [__cudaparm_RGBCurves_LumaKernel_inGreenPointX0];
	cvt.rn.f32.u32 	%f278, %r125;
	mov.f32 	%f279, 0f3b808081;   	// 0.00392157
	mul.f32 	%f280, %f278, %f279;
	st.shared.f32 	[sGreenPointsX__10+0], %f280;
	ld.param.u8 	%r126, [__cudaparm_RGBCurves_LumaKernel_inGreenPointX1];
	cvt.rn.f32.u32 	%f281, %r126;
	mov.f32 	%f282, 0f3b808081;   	// 0.00392157
	mul.f32 	%f283, %f281, %f282;
	st.shared.f32 	[sGreenPointsX__10+4], %f283;
	ld.param.u8 	%r127, [__cudaparm_RGBCurves_LumaKernel_inGreenPointX2];
	cvt.rn.f32.u32 	%f284, %r127;
	mov.f32 	%f285, 0f3b808081;   	// 0.00392157
	mul.f32 	%f286, %f284, %f285;
	st.shared.f32 	[sGreenPointsX__10+8], %f286;
	ld.param.u8 	%r128, [__cudaparm_RGBCurves_LumaKernel_inGreenPointX3];
	cvt.rn.f32.u32 	%f287, %r128;
	mov.f32 	%f288, 0f3b808081;   	// 0.00392157
	mul.f32 	%f289, %f287, %f288;
	st.shared.f32 	[sGreenPointsX__10+12], %f289;
	ld.param.u8 	%r129, [__cudaparm_RGBCurves_LumaKernel_inGreenPointX4];
	cvt.rn.f32.u32 	%f290, %r129;
	mov.f32 	%f291, 0f3b808081;   	// 0.00392157
	mul.f32 	%f292, %f290, %f291;
	st.shared.f32 	[sGreenPointsX__10+16], %f292;
	ld.param.u8 	%r130, [__cudaparm_RGBCurves_LumaKernel_inGreenPointX5];
	cvt.rn.f32.u32 	%f293, %r130;
	mov.f32 	%f294, 0f3b808081;   	// 0.00392157
	mul.f32 	%f295, %f293, %f294;
	st.shared.f32 	[sGreenPointsX__10+20], %f295;
	ld.param.u8 	%r131, [__cudaparm_RGBCurves_LumaKernel_inGreenPointX6];
	cvt.rn.f32.u32 	%f296, %r131;
	mov.f32 	%f297, 0f3b808081;   	// 0.00392157
	mul.f32 	%f298, %f296, %f297;
	st.shared.f32 	[sGreenPointsX__10+24], %f298;
	ld.param.u8 	%r132, [__cudaparm_RGBCurves_LumaKernel_inGreenPointX7];
	cvt.rn.f32.u32 	%f299, %r132;
	mov.f32 	%f300, 0f3b808081;   	// 0.00392157
	mul.f32 	%f301, %f299, %f300;
	st.shared.f32 	[sGreenPointsX__10+28], %f301;
	ld.param.u8 	%r133, [__cudaparm_RGBCurves_LumaKernel_inGreenPointX8];
	cvt.rn.f32.u32 	%f302, %r133;
	mov.f32 	%f303, 0f3b808081;   	// 0.00392157
	mul.f32 	%f304, %f302, %f303;
	st.shared.f32 	[sGreenPointsX__10+32], %f304;
	ld.param.u8 	%r134, [__cudaparm_RGBCurves_LumaKernel_inGreenPointX9];
	cvt.rn.f32.u32 	%f305, %r134;
	mov.f32 	%f306, 0f3b808081;   	// 0.00392157
	mul.f32 	%f307, %f305, %f306;
	st.shared.f32 	[sGreenPointsX__10+36], %f307;
	ld.param.u8 	%r135, [__cudaparm_RGBCurves_LumaKernel_inGreenPointX10];
	cvt.rn.f32.u32 	%f308, %r135;
	mov.f32 	%f309, 0f3b808081;   	// 0.00392157
	mul.f32 	%f310, %f308, %f309;
	st.shared.f32 	[sGreenPointsX__10+40], %f310;
	ld.param.u8 	%r136, [__cudaparm_RGBCurves_LumaKernel_inGreenPointX11];
	cvt.rn.f32.u32 	%f311, %r136;
	mov.f32 	%f312, 0f3b808081;   	// 0.00392157
	mul.f32 	%f313, %f311, %f312;
	st.shared.f32 	[sGreenPointsX__10+44], %f313;
	ld.param.u8 	%r137, [__cudaparm_RGBCurves_LumaKernel_inGreenPointX12];
	cvt.rn.f32.u32 	%f314, %r137;
	mov.f32 	%f315, 0f3b808081;   	// 0.00392157
	mul.f32 	%f316, %f314, %f315;
	st.shared.f32 	[sGreenPointsX__10+48], %f316;
	ld.param.u8 	%r138, [__cudaparm_RGBCurves_LumaKernel_inGreenPointX13];
	cvt.rn.f32.u32 	%f317, %r138;
	mov.f32 	%f318, 0f3b808081;   	// 0.00392157
	mul.f32 	%f319, %f317, %f318;
	st.shared.f32 	[sGreenPointsX__10+52], %f319;
	ld.param.u8 	%r139, [__cudaparm_RGBCurves_LumaKernel_inGreenPointX14];
	cvt.rn.f32.u32 	%f320, %r139;
	mov.f32 	%f321, 0f3b808081;   	// 0.00392157
	mul.f32 	%f322, %f320, %f321;
	st.shared.f32 	[sGreenPointsX__10+56], %f322;
	ld.param.u8 	%r140, [__cudaparm_RGBCurves_LumaKernel_inGreenPointX15];
	cvt.rn.f32.u32 	%f323, %r140;
	mov.f32 	%f324, 0f3b808081;   	// 0.00392157
	mul.f32 	%f325, %f323, %f324;
	st.shared.f32 	[sGreenPointsX__10+60], %f325;
	ld.param.u8 	%r141, [__cudaparm_RGBCurves_LumaKernel_inGreenPointY0];
	cvt.rn.f32.u32 	%f326, %r141;
	mov.f32 	%f327, 0f3b808081;   	// 0.00392157
	mul.f32 	%f328, %f326, %f327;
	st.shared.f32 	[sGreenPointsY__11+0], %f328;
	ld.param.u8 	%r142, [__cudaparm_RGBCurves_LumaKernel_inGreenPointY1];
	cvt.rn.f32.u32 	%f329, %r142;
	mov.f32 	%f330, 0f3b808081;   	// 0.00392157
	mul.f32 	%f331, %f329, %f330;
	st.shared.f32 	[sGreenPointsY__11+4], %f331;
	ld.param.u8 	%r143, [__cudaparm_RGBCurves_LumaKernel_inGreenPointY2];
	cvt.rn.f32.u32 	%f332, %r143;
	mov.f32 	%f333, 0f3b808081;   	// 0.00392157
	mul.f32 	%f334, %f332, %f333;
	st.shared.f32 	[sGreenPointsY__11+8], %f334;
	ld.param.u8 	%r144, [__cudaparm_RGBCurves_LumaKernel_inGreenPointY3];
	cvt.rn.f32.u32 	%f335, %r144;
	mov.f32 	%f336, 0f3b808081;   	// 0.00392157
	mul.f32 	%f337, %f335, %f336;
	st.shared.f32 	[sGreenPointsY__11+12], %f337;
	ld.param.u8 	%r145, [__cudaparm_RGBCurves_LumaKernel_inGreenPointY4];
	cvt.rn.f32.u32 	%f338, %r145;
	mov.f32 	%f339, 0f3b808081;   	// 0.00392157
	mul.f32 	%f340, %f338, %f339;
	st.shared.f32 	[sGreenPointsY__11+16], %f340;
	ld.param.u8 	%r146, [__cudaparm_RGBCurves_LumaKernel_inGreenPointY5];
	cvt.rn.f32.u32 	%f341, %r146;
	mov.f32 	%f342, 0f3b808081;   	// 0.00392157
	mul.f32 	%f343, %f341, %f342;
	st.shared.f32 	[sGreenPointsY__11+20], %f343;
	ld.param.u8 	%r147, [__cudaparm_RGBCurves_LumaKernel_inGreenPointY6];
	cvt.rn.f32.u32 	%f344, %r147;
	mov.f32 	%f345, 0f3b808081;   	// 0.00392157
	mul.f32 	%f346, %f344, %f345;
	st.shared.f32 	[sGreenPointsY__11+24], %f346;
	ld.param.u8 	%r148, [__cudaparm_RGBCurves_LumaKernel_inGreenPointY7];
	cvt.rn.f32.u32 	%f347, %r148;
	mov.f32 	%f348, 0f3b808081;   	// 0.00392157
	mul.f32 	%f349, %f347, %f348;
	st.shared.f32 	[sGreenPointsY__11+28], %f349;
	ld.param.u8 	%r149, [__cudaparm_RGBCurves_LumaKernel_inGreenPointY8];
	cvt.rn.f32.u32 	%f350, %r149;
	mov.f32 	%f351, 0f3b808081;   	// 0.00392157
	mul.f32 	%f352, %f350, %f351;
	st.shared.f32 	[sGreenPointsY__11+32], %f352;
	ld.param.u8 	%r150, [__cudaparm_RGBCurves_LumaKernel_inGreenPointY9];
	cvt.rn.f32.u32 	%f353, %r150;
	mov.f32 	%f354, 0f3b808081;   	// 0.00392157
	mul.f32 	%f355, %f353, %f354;
	st.shared.f32 	[sGreenPointsY__11+36], %f355;
	ld.param.u8 	%r151, [__cudaparm_RGBCurves_LumaKernel_inGreenPointY10];
	cvt.rn.f32.u32 	%f356, %r151;
	mov.f32 	%f357, 0f3b808081;   	// 0.00392157
	mul.f32 	%f358, %f356, %f357;
	st.shared.f32 	[sGreenPointsY__11+40], %f358;
	ld.param.u8 	%r152, [__cudaparm_RGBCurves_LumaKernel_inGreenPointY11];
	cvt.rn.f32.u32 	%f359, %r152;
	mov.f32 	%f360, 0f3b808081;   	// 0.00392157
	mul.f32 	%f361, %f359, %f360;
	st.shared.f32 	[sGreenPointsY__11+44], %f361;
	ld.param.u8 	%r153, [__cudaparm_RGBCurves_LumaKernel_inGreenPointY12];
	cvt.rn.f32.u32 	%f362, %r153;
	mov.f32 	%f363, 0f3b808081;   	// 0.00392157
	mul.f32 	%f364, %f362, %f363;
	st.shared.f32 	[sGreenPointsY__11+48], %f364;
	ld.param.u8 	%r154, [__cudaparm_RGBCurves_LumaKernel_inGreenPointY13];
	cvt.rn.f32.u32 	%f365, %r154;
	mov.f32 	%f366, 0f3b808081;   	// 0.00392157
	mul.f32 	%f367, %f365, %f366;
	st.shared.f32 	[sGreenPointsY__11+52], %f367;
	ld.param.u8 	%r155, [__cudaparm_RGBCurves_LumaKernel_inGreenPointY14];
	cvt.rn.f32.u32 	%f368, %r155;
	mov.f32 	%f369, 0f3b808081;   	// 0.00392157
	mul.f32 	%f370, %f368, %f369;
	st.shared.f32 	[sGreenPointsY__11+56], %f370;
	ld.param.u8 	%r156, [__cudaparm_RGBCurves_LumaKernel_inGreenPointY15];
	cvt.rn.f32.u32 	%f371, %r156;
	mov.f32 	%f372, 0f3b808081;   	// 0.00392157
	mul.f32 	%f373, %f371, %f372;
	st.shared.f32 	[sGreenPointsY__11+60], %f373;
	mov.u64 	%rd45, 0;
	mov.s64 	%rd46, %rd44;
	mov.s32 	%r157, 0;
	mov.u64 	%rd47, sGreenPointsY__11;
$Lt_1_65282:
 //<loop> Loop body line 272, nesting depth: 1, iterations: 16
	.loc	19	307	0
	ld.shared.f32 	%f374, [%rd46+0];
	mov.f32 	%f375, 0f00000000;   	// 0
	setp.neu.f32 	%p24, %f374, %f375;
	@%p24 bra 	$L_1_54274;
 //<loop> Part of loop body line 272, head labeled $Lt_1_65282
	add.u64 	%rd48, %rd45, %rd47;
	ld.shared.f32 	%f376, [%rd48+0];
	mov.f32 	%f377, 0f00000000;   	// 0
	setp.neu.f32 	%p25, %f376, %f377;
	@!%p25 bra 	$L_1_54018;
$L_1_54274:
 //<loop> Part of loop body line 272, head labeled $Lt_1_65282
	.loc	19	309	0
	add.s32 	%r124, %r157, 1;
	st.shared.s32 	[sGreenPointCount__9], %r124;
$L_1_54018:
 //<loop> Part of loop body line 272, head labeled $Lt_1_65282
	add.s32 	%r158, %r157, 1;
	mov.s32 	%r157, %r158;
	add.u64 	%rd45, %rd45, 4;
	add.u64 	%rd46, %rd46, 4;
	mov.u32 	%r159, 16;
	setp.ne.s32 	%p26, %r158, %r159;
	@%p26 bra 	$Lt_1_65282;
	.loc	19	312	0
	mov.f32 	%f378, 0f00000000;   	// 0
	st.shared.f32 	[u__0+0], %f378;
	.loc	18	48	0
	mov.f32 	%f379, 0f00000000;   	// 0
	st.shared.f32 	[sGreenPointsY2__12+0], %f379;
	cvt.u64.s32 	%rd49, %r124;
	mov.s64 	%rd50, %rd49;
	cvt.u32.u64 	%r160, %rd50;
	sub.s32 	%r161, %r160, 1;
	sub.s32 	%r162, %r160, 2;
	mov.u32 	%r163, 1;
	setp.le.s32 	%p27, %r161, %r163;
	@%p27 bra 	$Lt_1_65794;
	mov.s32 	%r164, %r162;
	mov.u64 	%rd16, 4;
	mov.s32 	%r71, 1;
	mov.u64 	%rd51, sGreenPointsY2__12;
	mov.s32 	%r165, %r164;
$Lt_1_66306:
 //<loop> Loop body line 48, nesting depth: 1, estimated iterations: unknown
	.loc	18	53	0
	add.u64 	%rd19, %rd16, %rd18;
	mov.f32 	%f380, 0f00000000;   	// 0
	st.shared.f32 	[%rd19+0], %f380;
	add.u64 	%rd52, %rd16, %rd51;
	mov.f32 	%f381, 0f00000000;   	// 0
	st.shared.f32 	[%rd52+0], %f381;
	add.u64 	%rd53, %rd16, %rd44;
	ld.shared.f32 	%f382, [%rd53+-4];
	ld.shared.f32 	%f383, [%rd53+0];
	setp.lt.f32 	%p28, %f382, %f383;
	@!%p28 bra 	$Lt_1_94722;
 //<loop> Part of loop body line 48, head labeled $Lt_1_66306
	ld.shared.f32 	%f384, [%rd53+4];
	ld.shared.f32 	%f383, [%rd53+0];
	setp.lt.f32 	%p29, %f383, %f384;
	@!%p29 bra 	$Lt_1_94722;
 //<loop> Part of loop body line 48, head labeled $Lt_1_66306
	ld.shared.f32 	%f383, [%rd53+0];
	ld.shared.f32 	%f382, [%rd53+-4];
	.loc	18	56	0
	sub.f32 	%f385, %f383, %f382;
	.loc	18	53	0
	ld.shared.f32 	%f384, [%rd53+4];
	.loc	18	56	0
	sub.f32 	%f386, %f384, %f382;
	.loc	18	57	0
	div.full.f32 	%f119, %f385, %f386;
	mov.f32 	%f387, 0f40000000;   	// 2
	ld.shared.f32 	%f388, [%rd52+-4];
	mad.f32 	%f122, %f388, %f119, %f387;
	mov.f32 	%f389, 0f00000000;   	// 0
	setp.neu.f32 	%p30, %f122, %f389;
	@!%p30 bra 	$Lt_1_94722;
 //<loop> Part of loop body line 48, head labeled $Lt_1_66306
	.loc	18	60	0
	mov.f32 	%f390, 0fbf800000;   	// -1
	add.f32 	%f391, %f119, %f390;
	div.full.f32 	%f392, %f391, %f122;
	st.shared.f32 	[%rd52+0], %f392;
	.loc	18	61	0
	add.u64 	%rd54, %rd16, %rd47;
	ld.shared.f32 	%f393, [%rd54+0];
	ld.shared.f32 	%f394, [%rd54+4];
	sub.f32 	%f129, %f394, %f393;
	sub.f32 	%f130, %f384, %f383;
	ld.shared.f32 	%f395, [%rd54+-4];
	sub.f32 	%f132, %f393, %f395;
	div.full.f32 	%f133, %f129, %f130;
	div.full.f32 	%f134, %f132, %f385;
	sub.f32 	%f135, %f133, %f134;
	st.shared.f32 	[%rd19+0], %f135;
	.loc	18	62	0
	mov.f32 	%f396, 0f40c00000;   	// 6
	mul.f32 	%f397, %f135, %f396;
	div.full.f32 	%f398, %f397, %f386;
	ld.shared.f32 	%f399, [%rd19+-4];
	mul.f32 	%f400, %f399, %f119;
	sub.f32 	%f401, %f398, %f400;
	div.full.f32 	%f402, %f401, %f122;
	st.shared.f32 	[%rd19+0], %f402;
$Lt_1_94722:
$L_1_54530:
 //<loop> Part of loop body line 48, head labeled $Lt_1_66306
	.loc	18	56	0
	add.s32 	%r71, %r71, 1;
	add.u64 	%rd16, %rd16, 4;
	setp.ne.s32 	%p31, %r161, %r71;
	@%p31 bra 	$Lt_1_66306;
$Lt_1_65794:
	mov.u64 	%rd51, sGreenPointsY2__12;
	.loc	18	69	0
	mov.f32 	%f403, 0f00000000;   	// 0
	mul.lo.u64 	%rd55, %rd49, 4;
	add.u64 	%rd56, %rd51, %rd55;
	st.shared.f32 	[%rd56+-4], %f403;
	.loc	18	72	0
	mov.s32 	%r73, %r162;
	mov.u32 	%r166, 0;
	setp.lt.s32 	%p32, %r162, %r166;
	@%p32 bra 	$Lt_1_67330;
	mov.s32 	%r167, %r161;
	cvt.u64.s32 	%rd57, %r162;
	mul.lo.u64 	%rd26, %rd57, 4;
	add.u64 	%rd27, %rd26, %rd18;
	add.u64 	%rd58, %rd26, %rd51;
	mov.s32 	%r168, %r167;
$Lt_1_67842:
 //<loop> Loop body line 72, nesting depth: 1, estimated iterations: unknown
	.loc	18	74	0
	ld.shared.f32 	%f404, [%rd27+0];
	ld.shared.f32 	%f405, [%rd58+0];
	ld.shared.f32 	%f406, [%rd58+4];
	mad.f32 	%f407, %f405, %f406, %f404;
	st.shared.f32 	[%rd58+0], %f407;
	sub.s32 	%r73, %r73, 1;
	sub.u64 	%rd58, %rd58, 4;
	sub.u64 	%rd27, %rd27, 4;
	mov.u32 	%r169, -1;
	setp.ne.s32 	%p33, %r73, %r169;
	@%p33 bra 	$Lt_1_67842;
$Lt_1_67330:
	.loc	19	312	0
	mov.u64 	%rd59, sBluePointsX__14;
	.loc	19	315	0
	mov.s32 	%r170, 16;
	st.shared.s32 	[sBluePointCount__13], %r170;
	ld.param.u8 	%r171, [__cudaparm_RGBCurves_LumaKernel_inBluePointX0];
	cvt.rn.f32.u32 	%f408, %r171;
	mov.f32 	%f409, 0f3b808081;   	// 0.00392157
	mul.f32 	%f410, %f408, %f409;
	st.shared.f32 	[sBluePointsX__14+0], %f410;
	ld.param.u8 	%r172, [__cudaparm_RGBCurves_LumaKernel_inBluePointX1];
	cvt.rn.f32.u32 	%f411, %r172;
	mov.f32 	%f412, 0f3b808081;   	// 0.00392157
	mul.f32 	%f413, %f411, %f412;
	st.shared.f32 	[sBluePointsX__14+4], %f413;
	ld.param.u8 	%r173, [__cudaparm_RGBCurves_LumaKernel_inBluePointX2];
	cvt.rn.f32.u32 	%f414, %r173;
	mov.f32 	%f415, 0f3b808081;   	// 0.00392157
	mul.f32 	%f416, %f414, %f415;
	st.shared.f32 	[sBluePointsX__14+8], %f416;
	ld.param.u8 	%r174, [__cudaparm_RGBCurves_LumaKernel_inBluePointX3];
	cvt.rn.f32.u32 	%f417, %r174;
	mov.f32 	%f418, 0f3b808081;   	// 0.00392157
	mul.f32 	%f419, %f417, %f418;
	st.shared.f32 	[sBluePointsX__14+12], %f419;
	ld.param.u8 	%r175, [__cudaparm_RGBCurves_LumaKernel_inBluePointX4];
	cvt.rn.f32.u32 	%f420, %r175;
	mov.f32 	%f421, 0f3b808081;   	// 0.00392157
	mul.f32 	%f422, %f420, %f421;
	st.shared.f32 	[sBluePointsX__14+16], %f422;
	ld.param.u8 	%r176, [__cudaparm_RGBCurves_LumaKernel_inBluePointX5];
	cvt.rn.f32.u32 	%f423, %r176;
	mov.f32 	%f424, 0f3b808081;   	// 0.00392157
	mul.f32 	%f425, %f423, %f424;
	st.shared.f32 	[sBluePointsX__14+20], %f425;
	ld.param.u8 	%r177, [__cudaparm_RGBCurves_LumaKernel_inBluePointX6];
	cvt.rn.f32.u32 	%f426, %r177;
	mov.f32 	%f427, 0f3b808081;   	// 0.00392157
	mul.f32 	%f428, %f426, %f427;
	st.shared.f32 	[sBluePointsX__14+24], %f428;
	ld.param.u8 	%r178, [__cudaparm_RGBCurves_LumaKernel_inBluePointX7];
	cvt.rn.f32.u32 	%f429, %r178;
	mov.f32 	%f430, 0f3b808081;   	// 0.00392157
	mul.f32 	%f431, %f429, %f430;
	st.shared.f32 	[sBluePointsX__14+28], %f431;
	ld.param.u8 	%r179, [__cudaparm_RGBCurves_LumaKernel_inBluePointX8];
	cvt.rn.f32.u32 	%f432, %r179;
	mov.f32 	%f433, 0f3b808081;   	// 0.00392157
	mul.f32 	%f434, %f432, %f433;
	st.shared.f32 	[sBluePointsX__14+32], %f434;
	ld.param.u8 	%r180, [__cudaparm_RGBCurves_LumaKernel_inBluePointX9];
	cvt.rn.f32.u32 	%f435, %r180;
	mov.f32 	%f436, 0f3b808081;   	// 0.00392157
	mul.f32 	%f437, %f435, %f436;
	st.shared.f32 	[sBluePointsX__14+36], %f437;
	ld.param.u8 	%r181, [__cudaparm_RGBCurves_LumaKernel_inBluePointX10];
	cvt.rn.f32.u32 	%f438, %r181;
	mov.f32 	%f439, 0f3b808081;   	// 0.00392157
	mul.f32 	%f440, %f438, %f439;
	st.shared.f32 	[sBluePointsX__14+40], %f440;
	ld.param.u8 	%r182, [__cudaparm_RGBCurves_LumaKernel_inBluePointX11];
	cvt.rn.f32.u32 	%f441, %r182;
	mov.f32 	%f442, 0f3b808081;   	// 0.00392157
	mul.f32 	%f443, %f441, %f442;
	st.shared.f32 	[sBluePointsX__14+44], %f443;
	ld.param.u8 	%r183, [__cudaparm_RGBCurves_LumaKernel_inBluePointX12];
	cvt.rn.f32.u32 	%f444, %r183;
	mov.f32 	%f445, 0f3b808081;   	// 0.00392157
	mul.f32 	%f446, %f444, %f445;
	st.shared.f32 	[sBluePointsX__14+48], %f446;
	ld.param.u8 	%r184, [__cudaparm_RGBCurves_LumaKernel_inBluePointX13];
	cvt.rn.f32.u32 	%f447, %r184;
	mov.f32 	%f448, 0f3b808081;   	// 0.00392157
	mul.f32 	%f449, %f447, %f448;
	st.shared.f32 	[sBluePointsX__14+52], %f449;
	ld.param.u8 	%r185, [__cudaparm_RGBCurves_LumaKernel_inBluePointX14];
	cvt.rn.f32.u32 	%f450, %r185;
	mov.f32 	%f451, 0f3b808081;   	// 0.00392157
	mul.f32 	%f452, %f450, %f451;
	st.shared.f32 	[sBluePointsX__14+56], %f452;
	ld.param.u8 	%r186, [__cudaparm_RGBCurves_LumaKernel_inBluePointX15];
	cvt.rn.f32.u32 	%f453, %r186;
	mov.f32 	%f454, 0f3b808081;   	// 0.00392157
	mul.f32 	%f455, %f453, %f454;
	st.shared.f32 	[sBluePointsX__14+60], %f455;
	ld.param.u8 	%r187, [__cudaparm_RGBCurves_LumaKernel_inBluePointY0];
	cvt.rn.f32.u32 	%f456, %r187;
	mov.f32 	%f457, 0f3b808081;   	// 0.00392157
	mul.f32 	%f458, %f456, %f457;
	st.shared.f32 	[sBluePointsY__15+0], %f458;
	ld.param.u8 	%r188, [__cudaparm_RGBCurves_LumaKernel_inBluePointY1];
	cvt.rn.f32.u32 	%f459, %r188;
	mov.f32 	%f460, 0f3b808081;   	// 0.00392157
	mul.f32 	%f461, %f459, %f460;
	st.shared.f32 	[sBluePointsY__15+4], %f461;
	ld.param.u8 	%r189, [__cudaparm_RGBCurves_LumaKernel_inBluePointY2];
	cvt.rn.f32.u32 	%f462, %r189;
	mov.f32 	%f463, 0f3b808081;   	// 0.00392157
	mul.f32 	%f464, %f462, %f463;
	st.shared.f32 	[sBluePointsY__15+8], %f464;
	ld.param.u8 	%r190, [__cudaparm_RGBCurves_LumaKernel_inBluePointY3];
	cvt.rn.f32.u32 	%f465, %r190;
	mov.f32 	%f466, 0f3b808081;   	// 0.00392157
	mul.f32 	%f467, %f465, %f466;
	st.shared.f32 	[sBluePointsY__15+12], %f467;
	ld.param.u8 	%r191, [__cudaparm_RGBCurves_LumaKernel_inBluePointY4];
	cvt.rn.f32.u32 	%f468, %r191;
	mov.f32 	%f469, 0f3b808081;   	// 0.00392157
	mul.f32 	%f470, %f468, %f469;
	st.shared.f32 	[sBluePointsY__15+16], %f470;
	ld.param.u8 	%r192, [__cudaparm_RGBCurves_LumaKernel_inBluePointY5];
	cvt.rn.f32.u32 	%f471, %r192;
	mov.f32 	%f472, 0f3b808081;   	// 0.00392157
	mul.f32 	%f473, %f471, %f472;
	st.shared.f32 	[sBluePointsY__15+20], %f473;
	ld.param.u8 	%r193, [__cudaparm_RGBCurves_LumaKernel_inBluePointY6];
	cvt.rn.f32.u32 	%f474, %r193;
	mov.f32 	%f475, 0f3b808081;   	// 0.00392157
	mul.f32 	%f476, %f474, %f475;
	st.shared.f32 	[sBluePointsY__15+24], %f476;
	ld.param.u8 	%r194, [__cudaparm_RGBCurves_LumaKernel_inBluePointY7];
	cvt.rn.f32.u32 	%f477, %r194;
	mov.f32 	%f478, 0f3b808081;   	// 0.00392157
	mul.f32 	%f479, %f477, %f478;
	st.shared.f32 	[sBluePointsY__15+28], %f479;
	ld.param.u8 	%r195, [__cudaparm_RGBCurves_LumaKernel_inBluePointY8];
	cvt.rn.f32.u32 	%f480, %r195;
	mov.f32 	%f481, 0f3b808081;   	// 0.00392157
	mul.f32 	%f482, %f480, %f481;
	st.shared.f32 	[sBluePointsY__15+32], %f482;
	ld.param.u8 	%r196, [__cudaparm_RGBCurves_LumaKernel_inBluePointY9];
	cvt.rn.f32.u32 	%f483, %r196;
	mov.f32 	%f484, 0f3b808081;   	// 0.00392157
	mul.f32 	%f485, %f483, %f484;
	st.shared.f32 	[sBluePointsY__15+36], %f485;
	ld.param.u8 	%r197, [__cudaparm_RGBCurves_LumaKernel_inBluePointY10];
	cvt.rn.f32.u32 	%f486, %r197;
	mov.f32 	%f487, 0f3b808081;   	// 0.00392157
	mul.f32 	%f488, %f486, %f487;
	st.shared.f32 	[sBluePointsY__15+40], %f488;
	ld.param.u8 	%r198, [__cudaparm_RGBCurves_LumaKernel_inBluePointY11];
	cvt.rn.f32.u32 	%f489, %r198;
	mov.f32 	%f490, 0f3b808081;   	// 0.00392157
	mul.f32 	%f491, %f489, %f490;
	st.shared.f32 	[sBluePointsY__15+44], %f491;
	ld.param.u8 	%r199, [__cudaparm_RGBCurves_LumaKernel_inBluePointY12];
	cvt.rn.f32.u32 	%f492, %r199;
	mov.f32 	%f493, 0f3b808081;   	// 0.00392157
	mul.f32 	%f494, %f492, %f493;
	st.shared.f32 	[sBluePointsY__15+48], %f494;
	ld.param.u8 	%r200, [__cudaparm_RGBCurves_LumaKernel_inBluePointY13];
	cvt.rn.f32.u32 	%f495, %r200;
	mov.f32 	%f496, 0f3b808081;   	// 0.00392157
	mul.f32 	%f497, %f495, %f496;
	st.shared.f32 	[sBluePointsY__15+52], %f497;
	ld.param.u8 	%r201, [__cudaparm_RGBCurves_LumaKernel_inBluePointY14];
	cvt.rn.f32.u32 	%f498, %r201;
	mov.f32 	%f499, 0f3b808081;   	// 0.00392157
	mul.f32 	%f500, %f498, %f499;
	st.shared.f32 	[sBluePointsY__15+56], %f500;
	ld.param.u8 	%r202, [__cudaparm_RGBCurves_LumaKernel_inBluePointY15];
	cvt.rn.f32.u32 	%f501, %r202;
	mov.f32 	%f502, 0f3b808081;   	// 0.00392157
	mul.f32 	%f503, %f501, %f502;
	st.shared.f32 	[sBluePointsY__15+60], %f503;
	mov.u64 	%rd60, 0;
	mov.s64 	%rd61, %rd59;
	mov.s32 	%r203, 0;
	mov.u64 	%rd62, sBluePointsY__15;
$Lt_1_68866:
 //<loop> Loop body line 315, nesting depth: 1, iterations: 16
	.loc	19	350	0
	ld.shared.f32 	%f504, [%rd61+0];
	mov.f32 	%f505, 0f00000000;   	// 0
	setp.neu.f32 	%p34, %f504, %f505;
	@%p34 bra 	$L_1_55298;
 //<loop> Part of loop body line 315, head labeled $Lt_1_68866
	add.u64 	%rd63, %rd60, %rd62;
	ld.shared.f32 	%f506, [%rd63+0];
	mov.f32 	%f507, 0f00000000;   	// 0
	setp.neu.f32 	%p35, %f506, %f507;
	@!%p35 bra 	$L_1_55042;
$L_1_55298:
 //<loop> Part of loop body line 315, head labeled $Lt_1_68866
	.loc	19	352	0
	add.s32 	%r170, %r203, 1;
	st.shared.s32 	[sBluePointCount__13], %r170;
$L_1_55042:
 //<loop> Part of loop body line 315, head labeled $Lt_1_68866
	add.s32 	%r204, %r203, 1;
	mov.s32 	%r203, %r204;
	add.u64 	%rd60, %rd60, 4;
	add.u64 	%rd61, %rd61, 4;
	mov.u32 	%r205, 16;
	setp.ne.s32 	%p36, %r204, %r205;
	@%p36 bra 	$Lt_1_68866;
	.loc	19	355	0
	mov.f32 	%f508, 0f00000000;   	// 0
	st.shared.f32 	[u__0+0], %f508;
	.loc	18	48	0
	mov.f32 	%f509, 0f00000000;   	// 0
	st.shared.f32 	[sBluePointsY2__16+0], %f509;
	cvt.u64.s32 	%rd64, %r170;
	mov.s64 	%rd65, %rd64;
	cvt.u32.u64 	%r206, %rd65;
	sub.s32 	%r207, %r206, 1;
	sub.s32 	%r208, %r206, 2;
	mov.u32 	%r209, 1;
	setp.le.s32 	%p37, %r207, %r209;
	@%p37 bra 	$Lt_1_69378;
	mov.s32 	%r210, %r208;
	mov.u64 	%rd16, 4;
	mov.s32 	%r71, 1;
	mov.u64 	%rd66, sBluePointsY2__16;
	mov.s32 	%r211, %r210;
$Lt_1_69890:
 //<loop> Loop body line 48, nesting depth: 1, estimated iterations: unknown
	.loc	18	53	0
	add.u64 	%rd19, %rd16, %rd18;
	mov.f32 	%f510, 0f00000000;   	// 0
	st.shared.f32 	[%rd19+0], %f510;
	add.u64 	%rd67, %rd16, %rd66;
	mov.f32 	%f511, 0f00000000;   	// 0
	st.shared.f32 	[%rd67+0], %f511;
	add.u64 	%rd68, %rd16, %rd59;
	ld.shared.f32 	%f512, [%rd68+-4];
	ld.shared.f32 	%f513, [%rd68+0];
	setp.lt.f32 	%p38, %f512, %f513;
	@!%p38 bra 	$Lt_1_96258;
 //<loop> Part of loop body line 48, head labeled $Lt_1_69890
	ld.shared.f32 	%f514, [%rd68+4];
	ld.shared.f32 	%f513, [%rd68+0];
	setp.lt.f32 	%p39, %f513, %f514;
	@!%p39 bra 	$Lt_1_96258;
 //<loop> Part of loop body line 48, head labeled $Lt_1_69890
	ld.shared.f32 	%f513, [%rd68+0];
	ld.shared.f32 	%f512, [%rd68+-4];
	.loc	18	56	0
	sub.f32 	%f515, %f513, %f512;
	.loc	18	53	0
	ld.shared.f32 	%f514, [%rd68+4];
	.loc	18	56	0
	sub.f32 	%f516, %f514, %f512;
	.loc	18	57	0
	div.full.f32 	%f119, %f515, %f516;
	mov.f32 	%f517, 0f40000000;   	// 2
	ld.shared.f32 	%f518, [%rd67+-4];
	mad.f32 	%f122, %f518, %f119, %f517;
	mov.f32 	%f519, 0f00000000;   	// 0
	setp.neu.f32 	%p40, %f122, %f519;
	@!%p40 bra 	$Lt_1_96258;
 //<loop> Part of loop body line 48, head labeled $Lt_1_69890
	.loc	18	60	0
	mov.f32 	%f520, 0fbf800000;   	// -1
	add.f32 	%f521, %f119, %f520;
	div.full.f32 	%f522, %f521, %f122;
	st.shared.f32 	[%rd67+0], %f522;
	.loc	18	61	0
	add.u64 	%rd69, %rd16, %rd62;
	ld.shared.f32 	%f523, [%rd69+0];
	ld.shared.f32 	%f524, [%rd69+4];
	sub.f32 	%f129, %f524, %f523;
	sub.f32 	%f130, %f514, %f513;
	ld.shared.f32 	%f525, [%rd69+-4];
	sub.f32 	%f132, %f523, %f525;
	div.full.f32 	%f133, %f129, %f130;
	div.full.f32 	%f134, %f132, %f515;
	sub.f32 	%f135, %f133, %f134;
	st.shared.f32 	[%rd19+0], %f135;
	.loc	18	62	0
	mov.f32 	%f526, 0f40c00000;   	// 6
	mul.f32 	%f527, %f135, %f526;
	div.full.f32 	%f528, %f527, %f516;
	ld.shared.f32 	%f529, [%rd19+-4];
	mul.f32 	%f530, %f529, %f119;
	sub.f32 	%f531, %f528, %f530;
	div.full.f32 	%f532, %f531, %f122;
	st.shared.f32 	[%rd19+0], %f532;
$Lt_1_96258:
$L_1_55554:
 //<loop> Part of loop body line 48, head labeled $Lt_1_69890
	.loc	18	56	0
	add.s32 	%r71, %r71, 1;
	add.u64 	%rd16, %rd16, 4;
	setp.ne.s32 	%p41, %r207, %r71;
	@%p41 bra 	$Lt_1_69890;
$Lt_1_69378:
	mov.u64 	%rd66, sBluePointsY2__16;
	.loc	18	69	0
	mov.f32 	%f533, 0f00000000;   	// 0
	mul.lo.u64 	%rd70, %rd64, 4;
	add.u64 	%rd71, %rd66, %rd70;
	st.shared.f32 	[%rd71+-4], %f533;
	.loc	18	72	0
	mov.s32 	%r73, %r208;
	mov.u32 	%r212, 0;
	setp.lt.s32 	%p42, %r208, %r212;
	@%p42 bra 	$Lt_1_70914;
	mov.s32 	%r213, %r207;
	cvt.u64.s32 	%rd72, %r208;
	mul.lo.u64 	%rd26, %rd72, 4;
	add.u64 	%rd27, %rd26, %rd18;
	add.u64 	%rd73, %rd26, %rd66;
	mov.s32 	%r214, %r213;
$Lt_1_71426:
 //<loop> Loop body line 72, nesting depth: 1, estimated iterations: unknown
	.loc	18	74	0
	ld.shared.f32 	%f534, [%rd27+0];
	ld.shared.f32 	%f535, [%rd73+0];
	ld.shared.f32 	%f536, [%rd73+4];
	mad.f32 	%f537, %f535, %f536, %f534;
	st.shared.f32 	[%rd73+0], %f537;
	sub.s32 	%r73, %r73, 1;
	sub.u64 	%rd73, %rd73, 4;
	sub.u64 	%rd27, %rd27, 4;
	mov.u32 	%r215, -1;
	setp.ne.s32 	%p43, %r73, %r215;
	@%p43 bra 	$Lt_1_71426;
$Lt_1_70914:
$Lt_1_57090:
	mov.u64 	%rd9, sMasterPointsX__2;
	mov.u64 	%rd17, sMasterPointsY2__4;
	mov.u64 	%rd29, sRedPointsX__6;
	mov.u64 	%rd36, sRedPointsY2__8;
	mov.u64 	%rd44, sGreenPointsX__10;
	mov.u64 	%rd51, sGreenPointsY2__12;
	mov.u64 	%rd59, sBluePointsX__14;
	mov.u64 	%rd66, sBluePointsY2__16;
	mov.u64 	%rd12, sMasterPointsY__3;
	mov.u64 	%rd32, sRedPointsY__7;
	mov.u64 	%rd47, sGreenPointsY__11;
	mov.u64 	%rd62, sBluePointsY__15;
	.loc	19	358	0
	bar.sync 	0;
	.loc	19	360	0
	mov.f32 	%f538, 0f3a83126f;   	// 0.001
	setp.gt.f32 	%p44, %f6, %f538;
	@!%p44 bra 	$Lt_1_71938;
	.loc	19	365	0
	mov.f32 	%f539, 0f00000000;   	// 0
	setp.lt.f32 	%p45, %f3, %f539;
	@!%p45 bra 	$Lt_1_72706;
	.loc	18	134	0
	ld.shared.f32 	%f540, [sRedPointsY__7+0];
	add.f32 	%f541, %f540, %f3;
	bra.uni 	$Lt_1_72962;
$Lt_1_72706:
	ld.shared.s32 	%r216, [sRedPointCount__5];
	cvt.u64.s32 	%rd34, %r216;
	mov.f32 	%f542, 0f3f800000;   	// 1
	setp.gt.f32 	%p46, %f3, %f542;
	@!%p46 bra 	$Lt_1_73218;
	.loc	18	138	0
	mov.f32 	%f543, 0f3f800000;   	// 1
	mul.lo.u64 	%rd74, %rd34, 4;
	add.u64 	%rd75, %rd32, %rd74;
	ld.shared.f32 	%f544, [%rd75+-4];
	sub.f32 	%f545, %f543, %f544;
	sub.f32 	%f541, %f3, %f545;
	bra.uni 	$Lt_1_72962;
$Lt_1_73218:
	.loc	18	88	0
	mov.f32 	%f546, %f3;
	.loc	18	92	0
	sub.u64 	%rd76, %rd34, 1;
	mov.s64 	%rd77, %rd76;
	mov.u64 	%rd78, 1;
	setp.le.u64 	%p47, %rd76, %rd78;
	@%p47 bra 	$Lt_1_97026;
	mov.u64 	%rd79, 0;
$Lt_1_73986:
 //<loop> Loop body line 92
	add.u64 	%rd80, %rd77, %rd79;
	shr.u64 	%rd81, %rd80, 1;
	mul.lo.u64 	%rd82, %rd81, 4;
	add.u64 	%rd83, %rd29, %rd82;
	ld.shared.f32 	%f547, [%rd83+0];
	setp.gt.f32 	%p48, %f547, %f3;
	@!%p48 bra 	$Lt_1_74498;
 //<loop> Part of loop body line 92, head labeled $Lt_1_73986
	.loc	18	101	0
	mov.s64 	%rd77, %rd81;
	bra.uni 	$Lt_1_74242;
$Lt_1_74498:
 //<loop> Part of loop body line 92, head labeled $Lt_1_73986
	.loc	18	105	0
	mov.s64 	%rd79, %rd81;
$Lt_1_74242:
 //<loop> Part of loop body line 92, head labeled $Lt_1_73986
	sub.u64 	%rd84, %rd77, %rd79;
	mov.u64 	%rd85, 1;
	setp.gt.u64 	%p49, %rd84, %rd85;
	@%p49 bra 	$Lt_1_73986;
	bra.uni 	$Lt_1_73474;
$Lt_1_97026:
	mov.u64 	%rd79, 0;
$Lt_1_73474:
	.loc	18	109	0
	mul.lo.u64 	%rd86, %rd77, 4;
	mul.lo.u64 	%rd87, %rd79, 4;
	add.u64 	%rd88, %rd86, %rd29;
	add.u64 	%rd89, %rd87, %rd29;
	ld.shared.f32 	%f548, [%rd88+0];
	ld.shared.f32 	%f549, [%rd89+0];
	sub.f32 	%f550, %f548, %f549;
	mov.f32 	%f551, 0f00000000;   	// 0
	setp.neu.f32 	%p50, %f550, %f551;
	@!%p50 bra 	$Lt_1_75010;
	ld.shared.f32 	%f548, [%rd88+0];
	.loc	18	112	0
	sub.f32 	%f552, %f548, %f3;
	.loc	18	109	0
	ld.shared.f32 	%f549, [%rd89+0];
	.loc	18	113	0
	sub.f32 	%f553, %f3, %f549;
	.loc	18	114	0
	div.full.f32 	%f554, %f553, %f550;
	div.full.f32 	%f555, %f552, %f550;
	add.u64 	%rd90, %rd86, %rd36;
	ld.shared.f32 	%f556, [%rd90+0];
	mul.f32 	%f557, %f554, %f554;
	mul.f32 	%f558, %f554, %f557;
	sub.f32 	%f559, %f558, %f554;
	mul.f32 	%f560, %f556, %f559;
	add.u64 	%rd91, %rd87, %rd36;
	ld.shared.f32 	%f561, [%rd91+0];
	mul.f32 	%f562, %f555, %f555;
	mul.f32 	%f563, %f555, %f562;
	sub.f32 	%f564, %f563, %f555;
	mad.f32 	%f565, %f561, %f564, %f560;
	mul.f32 	%f566, %f550, %f550;
	mul.f32 	%f567, %f565, %f566;
	mov.f32 	%f568, 0f40c00000;   	// 6
	div.full.f32 	%f569, %f567, %f568;
	add.u64 	%rd92, %rd86, %rd32;
	ld.shared.f32 	%f570, [%rd92+0];
	mul.f32 	%f571, %f570, %f554;
	add.u64 	%rd93, %rd87, %rd32;
	ld.shared.f32 	%f572, [%rd93+0];
	mad.f32 	%f573, %f572, %f555, %f571;
	add.f32 	%f546, %f569, %f573;
$Lt_1_75010:
	.loc	18	142	0
	mov.f32 	%f541, %f546;
$Lt_1_72962:
$Lt_1_72450:
	.loc	19	366	0
	mov.f32 	%f574, 0f00000000;   	// 0
	setp.lt.f32 	%p51, %f541, %f574;
	@!%p51 bra 	$Lt_1_75778;
	.loc	18	134	0
	ld.shared.f32 	%f575, [sMasterPointsY__3+0];
	add.f32 	%f541, %f575, %f541;
	bra.uni 	$Lt_1_76034;
$Lt_1_75778:
	ld.shared.s32 	%r217, [sMasterPointCount__1];
	cvt.u64.s32 	%rd94, %r217;
	mov.f32 	%f576, 0f3f800000;   	// 1
	setp.gt.f32 	%p52, %f541, %f576;
	@!%p52 bra 	$Lt_1_76290;
	.loc	18	138	0
	mov.f32 	%f577, 0f3f800000;   	// 1
	mul.lo.u64 	%rd95, %rd94, 4;
	add.u64 	%rd96, %rd12, %rd95;
	ld.shared.f32 	%f578, [%rd96+-4];
	sub.f32 	%f579, %f577, %f578;
	sub.f32 	%f541, %f541, %f579;
	bra.uni 	$Lt_1_76034;
$Lt_1_76290:
	.loc	18	88	0
	mov.f32 	%f546, %f541;
	.loc	18	92	0
	sub.u64 	%rd97, %rd94, 1;
	mov.s64 	%rd77, %rd97;
	mov.u64 	%rd98, 1;
	setp.le.u64 	%p53, %rd97, %rd98;
	@%p53 bra 	$Lt_1_97282;
	mov.u64 	%rd79, 0;
$Lt_1_77058:
 //<loop> Loop body line 92
	add.u64 	%rd99, %rd77, %rd79;
	shr.u64 	%rd81, %rd99, 1;
	mul.lo.u64 	%rd100, %rd81, 4;
	add.u64 	%rd101, %rd9, %rd100;
	ld.shared.f32 	%f580, [%rd101+0];
	setp.gt.f32 	%p54, %f580, %f541;
	@!%p54 bra 	$Lt_1_77570;
 //<loop> Part of loop body line 92, head labeled $Lt_1_77058
	.loc	18	101	0
	mov.s64 	%rd77, %rd81;
	bra.uni 	$Lt_1_77314;
$Lt_1_77570:
 //<loop> Part of loop body line 92, head labeled $Lt_1_77058
	.loc	18	105	0
	mov.s64 	%rd79, %rd81;
$Lt_1_77314:
 //<loop> Part of loop body line 92, head labeled $Lt_1_77058
	sub.u64 	%rd102, %rd77, %rd79;
	mov.u64 	%rd103, 1;
	setp.gt.u64 	%p55, %rd102, %rd103;
	@%p55 bra 	$Lt_1_77058;
	bra.uni 	$Lt_1_76546;
$Lt_1_97282:
	mov.u64 	%rd79, 0;
$Lt_1_76546:
	.loc	18	109	0
	mul.lo.u64 	%rd86, %rd77, 4;
	mul.lo.u64 	%rd87, %rd79, 4;
	add.u64 	%rd104, %rd86, %rd9;
	add.u64 	%rd105, %rd87, %rd9;
	ld.shared.f32 	%f581, [%rd104+0];
	ld.shared.f32 	%f582, [%rd105+0];
	sub.f32 	%f550, %f581, %f582;
	mov.f32 	%f583, 0f00000000;   	// 0
	setp.neu.f32 	%p56, %f550, %f583;
	@!%p56 bra 	$Lt_1_78082;
	.loc	18	112	0
	sub.f32 	%f552, %f581, %f541;
	.loc	18	113	0
	sub.f32 	%f553, %f541, %f582;
	.loc	18	114	0
	div.full.f32 	%f554, %f553, %f550;
	div.full.f32 	%f555, %f552, %f550;
	add.u64 	%rd106, %rd86, %rd17;
	ld.shared.f32 	%f584, [%rd106+0];
	mul.f32 	%f585, %f554, %f554;
	mul.f32 	%f586, %f554, %f585;
	sub.f32 	%f587, %f586, %f554;
	mul.f32 	%f588, %f584, %f587;
	add.u64 	%rd107, %rd87, %rd17;
	ld.shared.f32 	%f589, [%rd107+0];
	mul.f32 	%f590, %f555, %f555;
	mul.f32 	%f591, %f555, %f590;
	sub.f32 	%f592, %f591, %f555;
	mad.f32 	%f593, %f589, %f592, %f588;
	mul.f32 	%f594, %f550, %f550;
	mul.f32 	%f595, %f593, %f594;
	mov.f32 	%f596, 0f40c00000;   	// 6
	div.full.f32 	%f597, %f595, %f596;
	add.u64 	%rd108, %rd86, %rd12;
	ld.shared.f32 	%f598, [%rd108+0];
	mul.f32 	%f599, %f598, %f554;
	add.u64 	%rd109, %rd87, %rd12;
	ld.shared.f32 	%f600, [%rd109+0];
	mad.f32 	%f601, %f600, %f555, %f599;
	add.f32 	%f546, %f597, %f601;
$Lt_1_78082:
	.loc	18	142	0
	mov.f32 	%f541, %f546;
$Lt_1_76034:
$Lt_1_75522:
	.loc	19	367	0
	sub.f32 	%f602, %f541, %f3;
	mad.f32 	%f9, %f6, %f602, %f3;
	.loc	19	370	0
	mov.f32 	%f603, 0f00000000;   	// 0
	setp.lt.f32 	%p57, %f2, %f603;
	@!%p57 bra 	$Lt_1_78850;
	.loc	18	134	0
	ld.shared.f32 	%f604, [sGreenPointsY__11+0];
	add.f32 	%f541, %f604, %f2;
	bra.uni 	$Lt_1_79106;
$Lt_1_78850:
	ld.shared.s32 	%r218, [sGreenPointCount__9];
	cvt.u64.s32 	%rd49, %r218;
	mov.f32 	%f605, 0f3f800000;   	// 1
	setp.gt.f32 	%p58, %f2, %f605;
	@!%p58 bra 	$Lt_1_79362;
	.loc	18	138	0
	mov.f32 	%f606, 0f3f800000;   	// 1
	mul.lo.u64 	%rd110, %rd49, 4;
	add.u64 	%rd111, %rd47, %rd110;
	ld.shared.f32 	%f607, [%rd111+-4];
	sub.f32 	%f608, %f606, %f607;
	sub.f32 	%f541, %f2, %f608;
	bra.uni 	$Lt_1_79106;
$Lt_1_79362:
	.loc	18	88	0
	mov.f32 	%f546, %f2;
	.loc	18	92	0
	sub.u64 	%rd112, %rd49, 1;
	mov.s64 	%rd77, %rd112;
	mov.u64 	%rd113, 1;
	setp.le.u64 	%p59, %rd112, %rd113;
	@%p59 bra 	$Lt_1_97538;
	mov.u64 	%rd79, 0;
$Lt_1_80130:
 //<loop> Loop body line 92
	add.u64 	%rd114, %rd77, %rd79;
	shr.u64 	%rd81, %rd114, 1;
	mul.lo.u64 	%rd115, %rd81, 4;
	add.u64 	%rd116, %rd44, %rd115;
	ld.shared.f32 	%f609, [%rd116+0];
	setp.gt.f32 	%p60, %f609, %f2;
	@!%p60 bra 	$Lt_1_80642;
 //<loop> Part of loop body line 92, head labeled $Lt_1_80130
	.loc	18	101	0
	mov.s64 	%rd77, %rd81;
	bra.uni 	$Lt_1_80386;
$Lt_1_80642:
 //<loop> Part of loop body line 92, head labeled $Lt_1_80130
	.loc	18	105	0
	mov.s64 	%rd79, %rd81;
$Lt_1_80386:
 //<loop> Part of loop body line 92, head labeled $Lt_1_80130
	sub.u64 	%rd117, %rd77, %rd79;
	mov.u64 	%rd118, 1;
	setp.gt.u64 	%p61, %rd117, %rd118;
	@%p61 bra 	$Lt_1_80130;
	bra.uni 	$Lt_1_79618;
$Lt_1_97538:
	mov.u64 	%rd79, 0;
$Lt_1_79618:
	.loc	18	109	0
	mul.lo.u64 	%rd86, %rd77, 4;
	mul.lo.u64 	%rd87, %rd79, 4;
	add.u64 	%rd119, %rd86, %rd44;
	add.u64 	%rd120, %rd87, %rd44;
	ld.shared.f32 	%f610, [%rd119+0];
	ld.shared.f32 	%f611, [%rd120+0];
	sub.f32 	%f550, %f610, %f611;
	mov.f32 	%f612, 0f00000000;   	// 0
	setp.neu.f32 	%p62, %f550, %f612;
	@!%p62 bra 	$Lt_1_81154;
	ld.shared.f32 	%f610, [%rd119+0];
	.loc	18	112	0
	sub.f32 	%f552, %f610, %f2;
	.loc	18	109	0
	ld.shared.f32 	%f611, [%rd120+0];
	.loc	18	113	0
	sub.f32 	%f553, %f2, %f611;
	.loc	18	114	0
	div.full.f32 	%f554, %f553, %f550;
	div.full.f32 	%f555, %f552, %f550;
	add.u64 	%rd121, %rd86, %rd51;
	ld.shared.f32 	%f613, [%rd121+0];
	mul.f32 	%f614, %f554, %f554;
	mul.f32 	%f615, %f554, %f614;
	sub.f32 	%f616, %f615, %f554;
	mul.f32 	%f617, %f613, %f616;
	add.u64 	%rd122, %rd87, %rd51;
	ld.shared.f32 	%f618, [%rd122+0];
	mul.f32 	%f619, %f555, %f555;
	mul.f32 	%f620, %f555, %f619;
	sub.f32 	%f621, %f620, %f555;
	mad.f32 	%f622, %f618, %f621, %f617;
	mul.f32 	%f623, %f550, %f550;
	mul.f32 	%f624, %f622, %f623;
	mov.f32 	%f625, 0f40c00000;   	// 6
	div.full.f32 	%f626, %f624, %f625;
	add.u64 	%rd123, %rd86, %rd47;
	ld.shared.f32 	%f627, [%rd123+0];
	mul.f32 	%f628, %f627, %f554;
	add.u64 	%rd124, %rd87, %rd47;
	ld.shared.f32 	%f629, [%rd124+0];
	mad.f32 	%f630, %f629, %f555, %f628;
	add.f32 	%f546, %f626, %f630;
$Lt_1_81154:
	.loc	18	142	0
	mov.f32 	%f541, %f546;
$Lt_1_79106:
$Lt_1_78594:
	.loc	19	371	0
	mov.f32 	%f631, 0f00000000;   	// 0
	setp.lt.f32 	%p63, %f541, %f631;
	@!%p63 bra 	$Lt_1_81922;
	.loc	18	134	0
	ld.shared.f32 	%f632, [sMasterPointsY__3+0];
	add.f32 	%f541, %f632, %f541;
	bra.uni 	$Lt_1_82178;
$Lt_1_81922:
	ld.shared.s32 	%r219, [sMasterPointCount__1];
	cvt.u64.s32 	%rd94, %r219;
	mov.f32 	%f633, 0f3f800000;   	// 1
	setp.gt.f32 	%p64, %f541, %f633;
	@!%p64 bra 	$Lt_1_82434;
	.loc	18	138	0
	mov.f32 	%f634, 0f3f800000;   	// 1
	mul.lo.u64 	%rd125, %rd94, 4;
	add.u64 	%rd126, %rd12, %rd125;
	ld.shared.f32 	%f635, [%rd126+-4];
	sub.f32 	%f636, %f634, %f635;
	sub.f32 	%f541, %f541, %f636;
	bra.uni 	$Lt_1_82178;
$Lt_1_82434:
	.loc	18	88	0
	mov.f32 	%f546, %f541;
	.loc	18	92	0
	sub.u64 	%rd97, %rd94, 1;
	mov.s64 	%rd77, %rd97;
	mov.u64 	%rd127, 1;
	setp.le.u64 	%p65, %rd97, %rd127;
	@%p65 bra 	$Lt_1_97794;
	mov.u64 	%rd79, 0;
$Lt_1_83202:
 //<loop> Loop body line 92
	add.u64 	%rd128, %rd77, %rd79;
	shr.u64 	%rd81, %rd128, 1;
	mul.lo.u64 	%rd129, %rd81, 4;
	add.u64 	%rd130, %rd9, %rd129;
	ld.shared.f32 	%f637, [%rd130+0];
	setp.gt.f32 	%p66, %f637, %f541;
	@!%p66 bra 	$Lt_1_83714;
 //<loop> Part of loop body line 92, head labeled $Lt_1_83202
	.loc	18	101	0
	mov.s64 	%rd77, %rd81;
	bra.uni 	$Lt_1_83458;
$Lt_1_83714:
 //<loop> Part of loop body line 92, head labeled $Lt_1_83202
	.loc	18	105	0
	mov.s64 	%rd79, %rd81;
$Lt_1_83458:
 //<loop> Part of loop body line 92, head labeled $Lt_1_83202
	sub.u64 	%rd131, %rd77, %rd79;
	mov.u64 	%rd132, 1;
	setp.gt.u64 	%p67, %rd131, %rd132;
	@%p67 bra 	$Lt_1_83202;
	bra.uni 	$Lt_1_82690;
$Lt_1_97794:
	mov.u64 	%rd79, 0;
$Lt_1_82690:
	.loc	18	109	0
	mul.lo.u64 	%rd86, %rd77, 4;
	mul.lo.u64 	%rd87, %rd79, 4;
	add.u64 	%rd104, %rd86, %rd9;
	add.u64 	%rd105, %rd87, %rd9;
	ld.shared.f32 	%f581, [%rd104+0];
	ld.shared.f32 	%f582, [%rd105+0];
	sub.f32 	%f550, %f581, %f582;
	mov.f32 	%f638, 0f00000000;   	// 0
	setp.neu.f32 	%p68, %f550, %f638;
	@!%p68 bra 	$Lt_1_84226;
	.loc	18	112	0
	sub.f32 	%f552, %f581, %f541;
	.loc	18	113	0
	sub.f32 	%f553, %f541, %f582;
	.loc	18	114	0
	div.full.f32 	%f554, %f553, %f550;
	div.full.f32 	%f555, %f552, %f550;
	add.u64 	%rd133, %rd86, %rd17;
	ld.shared.f32 	%f639, [%rd133+0];
	mul.f32 	%f640, %f554, %f554;
	mul.f32 	%f641, %f554, %f640;
	sub.f32 	%f642, %f641, %f554;
	mul.f32 	%f643, %f639, %f642;
	add.u64 	%rd134, %rd87, %rd17;
	ld.shared.f32 	%f644, [%rd134+0];
	mul.f32 	%f645, %f555, %f555;
	mul.f32 	%f646, %f555, %f645;
	sub.f32 	%f647, %f646, %f555;
	mad.f32 	%f648, %f644, %f647, %f643;
	mul.f32 	%f649, %f550, %f550;
	mul.f32 	%f650, %f648, %f649;
	mov.f32 	%f651, 0f40c00000;   	// 6
	div.full.f32 	%f652, %f650, %f651;
	add.u64 	%rd135, %rd86, %rd12;
	ld.shared.f32 	%f653, [%rd135+0];
	mul.f32 	%f654, %f653, %f554;
	add.u64 	%rd136, %rd87, %rd12;
	ld.shared.f32 	%f655, [%rd136+0];
	mad.f32 	%f656, %f655, %f555, %f654;
	add.f32 	%f546, %f652, %f656;
$Lt_1_84226:
	.loc	18	142	0
	mov.f32 	%f541, %f546;
$Lt_1_82178:
$Lt_1_81666:
	.loc	19	372	0
	sub.f32 	%f657, %f541, %f2;
	mad.f32 	%f8, %f6, %f657, %f2;
	.loc	19	375	0
	mov.f32 	%f658, 0f00000000;   	// 0
	setp.lt.f32 	%p69, %f1, %f658;
	@!%p69 bra 	$Lt_1_84994;
	.loc	18	134	0
	ld.shared.f32 	%f659, [sBluePointsY__15+0];
	add.f32 	%f541, %f659, %f1;
	bra.uni 	$Lt_1_85250;
$Lt_1_84994:
	ld.shared.s32 	%r220, [sBluePointCount__13];
	cvt.u64.s32 	%rd64, %r220;
	mov.f32 	%f660, 0f3f800000;   	// 1
	setp.gt.f32 	%p70, %f1, %f660;
	@!%p70 bra 	$Lt_1_85506;
	.loc	18	138	0
	mov.f32 	%f661, 0f3f800000;   	// 1
	mul.lo.u64 	%rd137, %rd64, 4;
	add.u64 	%rd138, %rd62, %rd137;
	ld.shared.f32 	%f662, [%rd138+-4];
	sub.f32 	%f663, %f661, %f662;
	sub.f32 	%f541, %f1, %f663;
	bra.uni 	$Lt_1_85250;
$Lt_1_85506:
	.loc	18	88	0
	mov.f32 	%f546, %f1;
	.loc	18	92	0
	sub.u64 	%rd139, %rd64, 1;
	mov.s64 	%rd77, %rd139;
	mov.u64 	%rd140, 1;
	setp.le.u64 	%p71, %rd139, %rd140;
	@%p71 bra 	$Lt_1_98050;
	mov.u64 	%rd79, 0;
$Lt_1_86274:
 //<loop> Loop body line 92
	add.u64 	%rd141, %rd77, %rd79;
	shr.u64 	%rd81, %rd141, 1;
	mul.lo.u64 	%rd142, %rd81, 4;
	add.u64 	%rd143, %rd59, %rd142;
	ld.shared.f32 	%f664, [%rd143+0];
	setp.gt.f32 	%p72, %f664, %f1;
	@!%p72 bra 	$Lt_1_86786;
 //<loop> Part of loop body line 92, head labeled $Lt_1_86274
	.loc	18	101	0
	mov.s64 	%rd77, %rd81;
	bra.uni 	$Lt_1_86530;
$Lt_1_86786:
 //<loop> Part of loop body line 92, head labeled $Lt_1_86274
	.loc	18	105	0
	mov.s64 	%rd79, %rd81;
$Lt_1_86530:
 //<loop> Part of loop body line 92, head labeled $Lt_1_86274
	sub.u64 	%rd144, %rd77, %rd79;
	mov.u64 	%rd145, 1;
	setp.gt.u64 	%p73, %rd144, %rd145;
	@%p73 bra 	$Lt_1_86274;
	bra.uni 	$Lt_1_85762;
$Lt_1_98050:
	mov.u64 	%rd79, 0;
$Lt_1_85762:
	.loc	18	109	0
	mul.lo.u64 	%rd86, %rd77, 4;
	mul.lo.u64 	%rd87, %rd79, 4;
	add.u64 	%rd146, %rd86, %rd59;
	add.u64 	%rd147, %rd87, %rd59;
	ld.shared.f32 	%f665, [%rd146+0];
	ld.shared.f32 	%f666, [%rd147+0];
	sub.f32 	%f550, %f665, %f666;
	mov.f32 	%f667, 0f00000000;   	// 0
	setp.neu.f32 	%p74, %f550, %f667;
	@!%p74 bra 	$Lt_1_87298;
	ld.shared.f32 	%f665, [%rd146+0];
	.loc	18	112	0
	sub.f32 	%f552, %f665, %f1;
	.loc	18	109	0
	ld.shared.f32 	%f666, [%rd147+0];
	.loc	18	113	0
	sub.f32 	%f553, %f1, %f666;
	.loc	18	114	0
	div.full.f32 	%f554, %f553, %f550;
	div.full.f32 	%f555, %f552, %f550;
	add.u64 	%rd148, %rd86, %rd66;
	ld.shared.f32 	%f668, [%rd148+0];
	mul.f32 	%f669, %f554, %f554;
	mul.f32 	%f670, %f554, %f669;
	sub.f32 	%f671, %f670, %f554;
	mul.f32 	%f672, %f668, %f671;
	add.u64 	%rd149, %rd87, %rd66;
	ld.shared.f32 	%f673, [%rd149+0];
	mul.f32 	%f674, %f555, %f555;
	mul.f32 	%f675, %f555, %f674;
	sub.f32 	%f676, %f675, %f555;
	mad.f32 	%f677, %f673, %f676, %f672;
	mul.f32 	%f678, %f550, %f550;
	mul.f32 	%f679, %f677, %f678;
	mov.f32 	%f680, 0f40c00000;   	// 6
	div.full.f32 	%f681, %f679, %f680;
	add.u64 	%rd150, %rd86, %rd62;
	ld.shared.f32 	%f682, [%rd150+0];
	mul.f32 	%f683, %f682, %f554;
	add.u64 	%rd151, %rd87, %rd62;
	ld.shared.f32 	%f684, [%rd151+0];
	mad.f32 	%f685, %f684, %f555, %f683;
	add.f32 	%f546, %f681, %f685;
$Lt_1_87298:
	.loc	18	142	0
	mov.f32 	%f541, %f546;
$Lt_1_85250:
$Lt_1_84738:
	.loc	19	376	0
	mov.f32 	%f686, 0f00000000;   	// 0
	setp.lt.f32 	%p75, %f541, %f686;
	@!%p75 bra 	$Lt_1_88066;
	.loc	18	134	0
	ld.shared.f32 	%f687, [sMasterPointsY__3+0];
	add.f32 	%f541, %f687, %f541;
	bra.uni 	$Lt_1_88322;
$Lt_1_88066:
	ld.shared.s32 	%r221, [sMasterPointCount__1];
	cvt.u64.s32 	%rd94, %r221;
	mov.f32 	%f688, 0f3f800000;   	// 1
	setp.gt.f32 	%p76, %f541, %f688;
	@!%p76 bra 	$Lt_1_88578;
	.loc	18	138	0
	mov.f32 	%f689, 0f3f800000;   	// 1
	mul.lo.u64 	%rd152, %rd94, 4;
	add.u64 	%rd153, %rd12, %rd152;
	ld.shared.f32 	%f690, [%rd153+-4];
	sub.f32 	%f691, %f689, %f690;
	sub.f32 	%f541, %f541, %f691;
	bra.uni 	$Lt_1_88322;
$Lt_1_88578:
	.loc	18	88	0
	mov.f32 	%f546, %f541;
	.loc	18	92	0
	sub.u64 	%rd97, %rd94, 1;
	mov.s64 	%rd77, %rd97;
	mov.u64 	%rd154, 1;
	setp.le.u64 	%p77, %rd97, %rd154;
	@%p77 bra 	$Lt_1_98306;
	mov.u64 	%rd79, 0;
$Lt_1_89346:
 //<loop> Loop body line 92
	add.u64 	%rd155, %rd77, %rd79;
	shr.u64 	%rd81, %rd155, 1;
	mul.lo.u64 	%rd156, %rd81, 4;
	add.u64 	%rd157, %rd9, %rd156;
	ld.shared.f32 	%f692, [%rd157+0];
	setp.gt.f32 	%p78, %f692, %f541;
	@!%p78 bra 	$Lt_1_89858;
 //<loop> Part of loop body line 92, head labeled $Lt_1_89346
	.loc	18	101	0
	mov.s64 	%rd77, %rd81;
	bra.uni 	$Lt_1_89602;
$Lt_1_89858:
 //<loop> Part of loop body line 92, head labeled $Lt_1_89346
	.loc	18	105	0
	mov.s64 	%rd79, %rd81;
$Lt_1_89602:
 //<loop> Part of loop body line 92, head labeled $Lt_1_89346
	sub.u64 	%rd158, %rd77, %rd79;
	mov.u64 	%rd159, 1;
	setp.gt.u64 	%p79, %rd158, %rd159;
	@%p79 bra 	$Lt_1_89346;
	bra.uni 	$Lt_1_88834;
$Lt_1_98306:
	mov.u64 	%rd79, 0;
$Lt_1_88834:
	.loc	18	109	0
	mul.lo.u64 	%rd86, %rd77, 4;
	mul.lo.u64 	%rd87, %rd79, 4;
	add.u64 	%rd104, %rd86, %rd9;
	add.u64 	%rd105, %rd87, %rd9;
	ld.shared.f32 	%f581, [%rd104+0];
	ld.shared.f32 	%f582, [%rd105+0];
	sub.f32 	%f550, %f581, %f582;
	mov.f32 	%f693, 0f00000000;   	// 0
	setp.neu.f32 	%p80, %f550, %f693;
	@!%p80 bra 	$Lt_1_90370;
	.loc	18	112	0
	sub.f32 	%f552, %f581, %f541;
	.loc	18	113	0
	sub.f32 	%f553, %f541, %f582;
	.loc	18	114	0
	div.full.f32 	%f554, %f553, %f550;
	div.full.f32 	%f555, %f552, %f550;
	add.u64 	%rd160, %rd86, %rd17;
	ld.shared.f32 	%f694, [%rd160+0];
	mul.f32 	%f695, %f554, %f554;
	mul.f32 	%f696, %f554, %f695;
	sub.f32 	%f697, %f696, %f554;
	mul.f32 	%f698, %f694, %f697;
	add.u64 	%rd161, %rd87, %rd17;
	ld.shared.f32 	%f699, [%rd161+0];
	mul.f32 	%f700, %f555, %f555;
	mul.f32 	%f701, %f555, %f700;
	sub.f32 	%f702, %f701, %f555;
	mad.f32 	%f703, %f699, %f702, %f698;
	mul.f32 	%f704, %f550, %f550;
	mul.f32 	%f705, %f703, %f704;
	mov.f32 	%f706, 0f40c00000;   	// 6
	div.full.f32 	%f707, %f705, %f706;
	add.u64 	%rd162, %rd86, %rd12;
	ld.shared.f32 	%f708, [%rd162+0];
	mul.f32 	%f709, %f708, %f554;
	add.u64 	%rd163, %rd87, %rd12;
	ld.shared.f32 	%f710, [%rd163+0];
	mad.f32 	%f711, %f710, %f555, %f709;
	add.f32 	%f546, %f707, %f711;
$Lt_1_90370:
	.loc	18	142	0
	mov.f32 	%f541, %f546;
$Lt_1_88322:
$Lt_1_87810:
	.loc	19	377	0
	sub.f32 	%f712, %f541, %f1;
	mad.f32 	%f7, %f6, %f712, %f1;
$Lt_1_71938:
	.loc	16	100	0
	ld.const.f32 	%f713, [kRGB32f_To_601YPbPr+4];
	mul.f32 	%f714, %f713, %f8;
	ld.const.f32 	%f715, [kRGB32f_To_601YPbPr+0];
	mad.f32 	%f716, %f715, %f9, %f714;
	ld.const.f32 	%f717, [kRGB32f_To_601YPbPr+8];
	mad.f32 	%f718, %f717, %f7, %f716;
	st.global.v4.f32 	[%rd4+0], {%f718,%f718,%f718,%f4};
$Lt_1_56066:
	.loc	19	690	0
	exit;
$LDWend_RGBCurves_LumaKernel:
	} // RGBCurves_LumaKernel

	.entry RGBCurves_CompositeKernel (
		.param .u64 __cudaparm_RGBCurves_CompositeKernel_inImage,
		.param .s32 __cudaparm_RGBCurves_CompositeKernel_inPitch,
		.param .u64 __cudaparm_RGBCurves_CompositeKernel_inSecondaryMask,
		.param .s32 __cudaparm_RGBCurves_CompositeKernel_inWidth,
		.param .s32 __cudaparm_RGBCurves_CompositeKernel_inHeight,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inMasterPointX0,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inMasterPointX1,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inMasterPointX2,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inMasterPointX3,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inMasterPointX4,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inMasterPointX5,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inMasterPointX6,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inMasterPointX7,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inMasterPointX8,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inMasterPointX9,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inMasterPointX10,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inMasterPointX11,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inMasterPointX12,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inMasterPointX13,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inMasterPointX14,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inMasterPointX15,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inMasterPointY0,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inMasterPointY1,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inMasterPointY2,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inMasterPointY3,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inMasterPointY4,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inMasterPointY5,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inMasterPointY6,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inMasterPointY7,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inMasterPointY8,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inMasterPointY9,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inMasterPointY10,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inMasterPointY11,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inMasterPointY12,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inMasterPointY13,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inMasterPointY14,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inMasterPointY15,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inRedPointX0,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inRedPointX1,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inRedPointX2,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inRedPointX3,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inRedPointX4,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inRedPointX5,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inRedPointX6,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inRedPointX7,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inRedPointX8,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inRedPointX9,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inRedPointX10,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inRedPointX11,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inRedPointX12,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inRedPointX13,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inRedPointX14,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inRedPointX15,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inRedPointY0,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inRedPointY1,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inRedPointY2,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inRedPointY3,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inRedPointY4,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inRedPointY5,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inRedPointY6,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inRedPointY7,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inRedPointY8,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inRedPointY9,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inRedPointY10,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inRedPointY11,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inRedPointY12,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inRedPointY13,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inRedPointY14,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inRedPointY15,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inGreenPointX0,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inGreenPointX1,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inGreenPointX2,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inGreenPointX3,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inGreenPointX4,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inGreenPointX5,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inGreenPointX6,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inGreenPointX7,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inGreenPointX8,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inGreenPointX9,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inGreenPointX10,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inGreenPointX11,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inGreenPointX12,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inGreenPointX13,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inGreenPointX14,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inGreenPointX15,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inGreenPointY0,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inGreenPointY1,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inGreenPointY2,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inGreenPointY3,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inGreenPointY4,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inGreenPointY5,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inGreenPointY6,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inGreenPointY7,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inGreenPointY8,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inGreenPointY9,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inGreenPointY10,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inGreenPointY11,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inGreenPointY12,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inGreenPointY13,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inGreenPointY14,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inGreenPointY15,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inBluePointX0,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inBluePointX1,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inBluePointX2,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inBluePointX3,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inBluePointX4,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inBluePointX5,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inBluePointX6,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inBluePointX7,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inBluePointX8,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inBluePointX9,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inBluePointX10,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inBluePointX11,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inBluePointX12,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inBluePointX13,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inBluePointX14,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inBluePointX15,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inBluePointY0,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inBluePointY1,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inBluePointY2,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inBluePointY3,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inBluePointY4,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inBluePointY5,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inBluePointY6,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inBluePointY7,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inBluePointY8,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inBluePointY9,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inBluePointY10,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inBluePointY11,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inBluePointY12,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inBluePointY13,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inBluePointY14,
		.param .u8 __cudaparm_RGBCurves_CompositeKernel_inBluePointY15)
	{
	.reg .u32 %r<223>;
	.reg .u64 %rd<165>;
	.reg .f32 %f<714>;
	.reg .pred %p<82>;
	.loc	19	828	0
$LBB1_RGBCurves_CompositeKernel:
	.loc	19	831	0
	cvt.s32.u16 	%r1, %ctaid.x;
	cvt.s32.u16 	%r2, %ntid.x;
	mul24.lo.s32 	%r3, %r1, %r2;
	cvt.s32.u16 	%r4, %ctaid.y;
	cvt.s32.u16 	%r5, %ntid.y;
	mul24.lo.s32 	%r6, %r4, %r5;
	cvt.u32.u16 	%r7, %tid.x;
	add.u32 	%r8, %r3, %r7;
	cvt.u32.u16 	%r9, %tid.y;
	add.u32 	%r10, %r6, %r9;
	ld.param.s32 	%r11, [__cudaparm_RGBCurves_CompositeKernel_inWidth];
	set.gt.u32.s32 	%r12, %r11, %r8;
	neg.s32 	%r13, %r12;
	ld.param.s32 	%r14, [__cudaparm_RGBCurves_CompositeKernel_inHeight];
	set.gt.u32.s32 	%r15, %r14, %r10;
	neg.s32 	%r16, %r15;
	and.b32 	%r17, %r13, %r16;
	mov.u32 	%r18, 0;
	setp.eq.s32 	%p1, %r17, %r18;
	@%p1 bra 	$Lt_2_56066;
	.loc	16	89	0
	ld.param.s32 	%r19, [__cudaparm_RGBCurves_CompositeKernel_inPitch];
	mul24.lo.s32 	%r20, %r10, %r19;
	add.s32 	%r21, %r20, %r8;
	cvt.s64.s32 	%rd1, %r21;
	ld.param.u64 	%rd2, [__cudaparm_RGBCurves_CompositeKernel_inImage];
	mul.lo.u64 	%rd3, %rd1, 16;
	add.u64 	%rd4, %rd2, %rd3;
	ld.global.v4.f32 	{%f1,%f2,%f3,%f4}, [%rd4+0];
	.loc	19	835	0
	ld.param.u64 	%rd5, [__cudaparm_RGBCurves_CompositeKernel_inSecondaryMask];
	mov.u64 	%rd6, 0;
	setp.eq.u64 	%p2, %rd5, %rd6;
	@%p2 bra 	$Lt_2_56834;
	.loc	16	89	0
	mul.lo.u64 	%rd7, %rd1, 4;
	add.u64 	%rd8, %rd5, %rd7;
	ld.global.f32 	%f5, [%rd8+0];
	.loc	19	836	0
	mov.f32 	%f6, %f5;
	bra.uni 	$Lt_2_56578;
$Lt_2_56834:
	mov.f32 	%f6, 0f3f800000;     	// 1
$Lt_2_56578:
	.loc	19	161	0
	mov.f32 	%f7, %f1;
	mov.f32 	%f8, %f2;
	mov.f32 	%f9, %f3;
	mov.u32 	%r22, 0;
	set.eq.u32.u32 	%r23, %r9, %r22;
	neg.s32 	%r24, %r23;
	mov.u32 	%r25, 0;
	set.eq.u32.u32 	%r26, %r7, %r25;
	neg.s32 	%r27, %r26;
	and.b32 	%r28, %r24, %r27;
	mov.u32 	%r29, 0;
	setp.eq.s32 	%p3, %r28, %r29;
	@%p3 bra 	$Lt_2_70914;
	.loc	19	186	0
	mov.u64 	%rd9, sMasterPointsX__2;
	mov.s32 	%r30, 16;
	st.shared.s32 	[sMasterPointCount__1], %r30;
	ld.param.u8 	%r31, [__cudaparm_RGBCurves_CompositeKernel_inMasterPointX0];
	cvt.rn.f32.u32 	%f10, %r31;
	mov.f32 	%f11, 0f3b808081;    	// 0.00392157
	mul.f32 	%f12, %f10, %f11;
	st.shared.f32 	[sMasterPointsX__2+0], %f12;
	ld.param.u8 	%r32, [__cudaparm_RGBCurves_CompositeKernel_inMasterPointX1];
	cvt.rn.f32.u32 	%f13, %r32;
	mov.f32 	%f14, 0f3b808081;    	// 0.00392157
	mul.f32 	%f15, %f13, %f14;
	st.shared.f32 	[sMasterPointsX__2+4], %f15;
	ld.param.u8 	%r33, [__cudaparm_RGBCurves_CompositeKernel_inMasterPointX2];
	cvt.rn.f32.u32 	%f16, %r33;
	mov.f32 	%f17, 0f3b808081;    	// 0.00392157
	mul.f32 	%f18, %f16, %f17;
	st.shared.f32 	[sMasterPointsX__2+8], %f18;
	ld.param.u8 	%r34, [__cudaparm_RGBCurves_CompositeKernel_inMasterPointX3];
	cvt.rn.f32.u32 	%f19, %r34;
	mov.f32 	%f20, 0f3b808081;    	// 0.00392157
	mul.f32 	%f21, %f19, %f20;
	st.shared.f32 	[sMasterPointsX__2+12], %f21;
	ld.param.u8 	%r35, [__cudaparm_RGBCurves_CompositeKernel_inMasterPointX4];
	cvt.rn.f32.u32 	%f22, %r35;
	mov.f32 	%f23, 0f3b808081;    	// 0.00392157
	mul.f32 	%f24, %f22, %f23;
	st.shared.f32 	[sMasterPointsX__2+16], %f24;
	ld.param.u8 	%r36, [__cudaparm_RGBCurves_CompositeKernel_inMasterPointX5];
	cvt.rn.f32.u32 	%f25, %r36;
	mov.f32 	%f26, 0f3b808081;    	// 0.00392157
	mul.f32 	%f27, %f25, %f26;
	st.shared.f32 	[sMasterPointsX__2+20], %f27;
	ld.param.u8 	%r37, [__cudaparm_RGBCurves_CompositeKernel_inMasterPointX6];
	cvt.rn.f32.u32 	%f28, %r37;
	mov.f32 	%f29, 0f3b808081;    	// 0.00392157
	mul.f32 	%f30, %f28, %f29;
	st.shared.f32 	[sMasterPointsX__2+24], %f30;
	ld.param.u8 	%r38, [__cudaparm_RGBCurves_CompositeKernel_inMasterPointX7];
	cvt.rn.f32.u32 	%f31, %r38;
	mov.f32 	%f32, 0f3b808081;    	// 0.00392157
	mul.f32 	%f33, %f31, %f32;
	st.shared.f32 	[sMasterPointsX__2+28], %f33;
	ld.param.u8 	%r39, [__cudaparm_RGBCurves_CompositeKernel_inMasterPointX8];
	cvt.rn.f32.u32 	%f34, %r39;
	mov.f32 	%f35, 0f3b808081;    	// 0.00392157
	mul.f32 	%f36, %f34, %f35;
	st.shared.f32 	[sMasterPointsX__2+32], %f36;
	ld.param.u8 	%r40, [__cudaparm_RGBCurves_CompositeKernel_inMasterPointX9];
	cvt.rn.f32.u32 	%f37, %r40;
	mov.f32 	%f38, 0f3b808081;    	// 0.00392157
	mul.f32 	%f39, %f37, %f38;
	st.shared.f32 	[sMasterPointsX__2+36], %f39;
	ld.param.u8 	%r41, [__cudaparm_RGBCurves_CompositeKernel_inMasterPointX10];
	cvt.rn.f32.u32 	%f40, %r41;
	mov.f32 	%f41, 0f3b808081;    	// 0.00392157
	mul.f32 	%f42, %f40, %f41;
	st.shared.f32 	[sMasterPointsX__2+40], %f42;
	ld.param.u8 	%r42, [__cudaparm_RGBCurves_CompositeKernel_inMasterPointX11];
	cvt.rn.f32.u32 	%f43, %r42;
	mov.f32 	%f44, 0f3b808081;    	// 0.00392157
	mul.f32 	%f45, %f43, %f44;
	st.shared.f32 	[sMasterPointsX__2+44], %f45;
	ld.param.u8 	%r43, [__cudaparm_RGBCurves_CompositeKernel_inMasterPointX12];
	cvt.rn.f32.u32 	%f46, %r43;
	mov.f32 	%f47, 0f3b808081;    	// 0.00392157
	mul.f32 	%f48, %f46, %f47;
	st.shared.f32 	[sMasterPointsX__2+48], %f48;
	ld.param.u8 	%r44, [__cudaparm_RGBCurves_CompositeKernel_inMasterPointX13];
	cvt.rn.f32.u32 	%f49, %r44;
	mov.f32 	%f50, 0f3b808081;    	// 0.00392157
	mul.f32 	%f51, %f49, %f50;
	st.shared.f32 	[sMasterPointsX__2+52], %f51;
	ld.param.u8 	%r45, [__cudaparm_RGBCurves_CompositeKernel_inMasterPointX14];
	cvt.rn.f32.u32 	%f52, %r45;
	mov.f32 	%f53, 0f3b808081;    	// 0.00392157
	mul.f32 	%f54, %f52, %f53;
	st.shared.f32 	[sMasterPointsX__2+56], %f54;
	ld.param.u8 	%r46, [__cudaparm_RGBCurves_CompositeKernel_inMasterPointX15];
	cvt.rn.f32.u32 	%f55, %r46;
	mov.f32 	%f56, 0f3b808081;    	// 0.00392157
	mul.f32 	%f57, %f55, %f56;
	st.shared.f32 	[sMasterPointsX__2+60], %f57;
	ld.param.u8 	%r47, [__cudaparm_RGBCurves_CompositeKernel_inMasterPointY0];
	cvt.rn.f32.u32 	%f58, %r47;
	mov.f32 	%f59, 0f3b808081;    	// 0.00392157
	mul.f32 	%f60, %f58, %f59;
	st.shared.f32 	[sMasterPointsY__3+0], %f60;
	ld.param.u8 	%r48, [__cudaparm_RGBCurves_CompositeKernel_inMasterPointY1];
	cvt.rn.f32.u32 	%f61, %r48;
	mov.f32 	%f62, 0f3b808081;    	// 0.00392157
	mul.f32 	%f63, %f61, %f62;
	st.shared.f32 	[sMasterPointsY__3+4], %f63;
	ld.param.u8 	%r49, [__cudaparm_RGBCurves_CompositeKernel_inMasterPointY2];
	cvt.rn.f32.u32 	%f64, %r49;
	mov.f32 	%f65, 0f3b808081;    	// 0.00392157
	mul.f32 	%f66, %f64, %f65;
	st.shared.f32 	[sMasterPointsY__3+8], %f66;
	ld.param.u8 	%r50, [__cudaparm_RGBCurves_CompositeKernel_inMasterPointY3];
	cvt.rn.f32.u32 	%f67, %r50;
	mov.f32 	%f68, 0f3b808081;    	// 0.00392157
	mul.f32 	%f69, %f67, %f68;
	st.shared.f32 	[sMasterPointsY__3+12], %f69;
	ld.param.u8 	%r51, [__cudaparm_RGBCurves_CompositeKernel_inMasterPointY4];
	cvt.rn.f32.u32 	%f70, %r51;
	mov.f32 	%f71, 0f3b808081;    	// 0.00392157
	mul.f32 	%f72, %f70, %f71;
	st.shared.f32 	[sMasterPointsY__3+16], %f72;
	ld.param.u8 	%r52, [__cudaparm_RGBCurves_CompositeKernel_inMasterPointY5];
	cvt.rn.f32.u32 	%f73, %r52;
	mov.f32 	%f74, 0f3b808081;    	// 0.00392157
	mul.f32 	%f75, %f73, %f74;
	st.shared.f32 	[sMasterPointsY__3+20], %f75;
	ld.param.u8 	%r53, [__cudaparm_RGBCurves_CompositeKernel_inMasterPointY6];
	cvt.rn.f32.u32 	%f76, %r53;
	mov.f32 	%f77, 0f3b808081;    	// 0.00392157
	mul.f32 	%f78, %f76, %f77;
	st.shared.f32 	[sMasterPointsY__3+24], %f78;
	ld.param.u8 	%r54, [__cudaparm_RGBCurves_CompositeKernel_inMasterPointY7];
	cvt.rn.f32.u32 	%f79, %r54;
	mov.f32 	%f80, 0f3b808081;    	// 0.00392157
	mul.f32 	%f81, %f79, %f80;
	st.shared.f32 	[sMasterPointsY__3+28], %f81;
	ld.param.u8 	%r55, [__cudaparm_RGBCurves_CompositeKernel_inMasterPointY8];
	cvt.rn.f32.u32 	%f82, %r55;
	mov.f32 	%f83, 0f3b808081;    	// 0.00392157
	mul.f32 	%f84, %f82, %f83;
	st.shared.f32 	[sMasterPointsY__3+32], %f84;
	ld.param.u8 	%r56, [__cudaparm_RGBCurves_CompositeKernel_inMasterPointY9];
	cvt.rn.f32.u32 	%f85, %r56;
	mov.f32 	%f86, 0f3b808081;    	// 0.00392157
	mul.f32 	%f87, %f85, %f86;
	st.shared.f32 	[sMasterPointsY__3+36], %f87;
	ld.param.u8 	%r57, [__cudaparm_RGBCurves_CompositeKernel_inMasterPointY10];
	cvt.rn.f32.u32 	%f88, %r57;
	mov.f32 	%f89, 0f3b808081;    	// 0.00392157
	mul.f32 	%f90, %f88, %f89;
	st.shared.f32 	[sMasterPointsY__3+40], %f90;
	ld.param.u8 	%r58, [__cudaparm_RGBCurves_CompositeKernel_inMasterPointY11];
	cvt.rn.f32.u32 	%f91, %r58;
	mov.f32 	%f92, 0f3b808081;    	// 0.00392157
	mul.f32 	%f93, %f91, %f92;
	st.shared.f32 	[sMasterPointsY__3+44], %f93;
	ld.param.u8 	%r59, [__cudaparm_RGBCurves_CompositeKernel_inMasterPointY12];
	cvt.rn.f32.u32 	%f94, %r59;
	mov.f32 	%f95, 0f3b808081;    	// 0.00392157
	mul.f32 	%f96, %f94, %f95;
	st.shared.f32 	[sMasterPointsY__3+48], %f96;
	ld.param.u8 	%r60, [__cudaparm_RGBCurves_CompositeKernel_inMasterPointY13];
	cvt.rn.f32.u32 	%f97, %r60;
	mov.f32 	%f98, 0f3b808081;    	// 0.00392157
	mul.f32 	%f99, %f97, %f98;
	st.shared.f32 	[sMasterPointsY__3+52], %f99;
	ld.param.u8 	%r61, [__cudaparm_RGBCurves_CompositeKernel_inMasterPointY14];
	cvt.rn.f32.u32 	%f100, %r61;
	mov.f32 	%f101, 0f3b808081;   	// 0.00392157
	mul.f32 	%f102, %f100, %f101;
	st.shared.f32 	[sMasterPointsY__3+56], %f102;
	ld.param.u8 	%r62, [__cudaparm_RGBCurves_CompositeKernel_inMasterPointY15];
	cvt.rn.f32.u32 	%f103, %r62;
	mov.f32 	%f104, 0f3b808081;   	// 0.00392157
	mul.f32 	%f105, %f103, %f104;
	st.shared.f32 	[sMasterPointsY__3+60], %f105;
	mov.u64 	%rd10, 0;
	mov.s64 	%rd11, %rd9;
	mov.s32 	%r63, 0;
	mov.u64 	%rd12, sMasterPointsY__3;
$Lt_2_58114:
 //<loop> Loop body line 186, nesting depth: 1, iterations: 16
	.loc	19	221	0
	ld.shared.f32 	%f106, [%rd11+0];
	mov.f32 	%f107, 0f00000000;   	// 0
	setp.neu.f32 	%p4, %f106, %f107;
	@%p4 bra 	$L_2_52226;
 //<loop> Part of loop body line 186, head labeled $Lt_2_58114
	add.u64 	%rd13, %rd10, %rd12;
	ld.shared.f32 	%f108, [%rd13+0];
	mov.f32 	%f109, 0f00000000;   	// 0
	setp.neu.f32 	%p5, %f108, %f109;
	@!%p5 bra 	$L_2_51970;
$L_2_52226:
 //<loop> Part of loop body line 186, head labeled $Lt_2_58114
	.loc	19	223	0
	add.s32 	%r30, %r63, 1;
	st.shared.s32 	[sMasterPointCount__1], %r30;
$L_2_51970:
 //<loop> Part of loop body line 186, head labeled $Lt_2_58114
	add.s32 	%r64, %r63, 1;
	mov.s32 	%r63, %r64;
	add.u64 	%rd10, %rd10, 4;
	add.u64 	%rd11, %rd11, 4;
	mov.u32 	%r65, 16;
	setp.ne.s32 	%p6, %r64, %r65;
	@%p6 bra 	$Lt_2_58114;
	.loc	19	226	0
	mov.f32 	%f110, 0f00000000;   	// 0
	st.shared.f32 	[u__0+0], %f110;
	.loc	18	48	0
	mov.f32 	%f111, 0f00000000;   	// 0
	st.shared.f32 	[sMasterPointsY2__4+0], %f111;
	cvt.u64.s32 	%rd14, %r30;
	mov.s64 	%rd15, %rd14;
	cvt.u32.u64 	%r66, %rd15;
	sub.s32 	%r67, %r66, 1;
	sub.s32 	%r68, %r66, 2;
	mov.u32 	%r69, 1;
	setp.le.s32 	%p7, %r67, %r69;
	@%p7 bra 	$Lt_2_58626;
	mov.s32 	%r70, %r68;
	mov.u64 	%rd16, 4;
	mov.s32 	%r71, 1;
	mov.u64 	%rd17, sMasterPointsY2__4;
	mov.u64 	%rd18, u__0;
	mov.s32 	%r72, %r70;
$Lt_2_59138:
 //<loop> Loop body line 48, nesting depth: 1, estimated iterations: unknown
	.loc	18	53	0
	add.u64 	%rd19, %rd16, %rd18;
	mov.f32 	%f112, 0f00000000;   	// 0
	st.shared.f32 	[%rd19+0], %f112;
	add.u64 	%rd20, %rd16, %rd17;
	mov.f32 	%f113, 0f00000000;   	// 0
	st.shared.f32 	[%rd20+0], %f113;
	add.u64 	%rd21, %rd16, %rd9;
	ld.shared.f32 	%f114, [%rd21+-4];
	ld.shared.f32 	%f115, [%rd21+0];
	setp.lt.f32 	%p8, %f114, %f115;
	@!%p8 bra 	$Lt_2_91650;
 //<loop> Part of loop body line 48, head labeled $Lt_2_59138
	ld.shared.f32 	%f116, [%rd21+4];
	ld.shared.f32 	%f115, [%rd21+0];
	setp.lt.f32 	%p9, %f115, %f116;
	@!%p9 bra 	$Lt_2_91650;
 //<loop> Part of loop body line 48, head labeled $Lt_2_59138
	ld.shared.f32 	%f115, [%rd21+0];
	ld.shared.f32 	%f114, [%rd21+-4];
	.loc	18	56	0
	sub.f32 	%f117, %f115, %f114;
	.loc	18	53	0
	ld.shared.f32 	%f116, [%rd21+4];
	.loc	18	56	0
	sub.f32 	%f118, %f116, %f114;
	.loc	18	57	0
	div.full.f32 	%f119, %f117, %f118;
	mov.f32 	%f120, 0f40000000;   	// 2
	ld.shared.f32 	%f121, [%rd20+-4];
	mad.f32 	%f122, %f121, %f119, %f120;
	mov.f32 	%f123, 0f00000000;   	// 0
	setp.neu.f32 	%p10, %f122, %f123;
	@!%p10 bra 	$Lt_2_91650;
 //<loop> Part of loop body line 48, head labeled $Lt_2_59138
	.loc	18	60	0
	mov.f32 	%f124, 0fbf800000;   	// -1
	add.f32 	%f125, %f119, %f124;
	div.full.f32 	%f126, %f125, %f122;
	st.shared.f32 	[%rd20+0], %f126;
	.loc	18	61	0
	add.u64 	%rd22, %rd16, %rd12;
	ld.shared.f32 	%f127, [%rd22+0];
	ld.shared.f32 	%f128, [%rd22+4];
	sub.f32 	%f129, %f128, %f127;
	sub.f32 	%f130, %f116, %f115;
	ld.shared.f32 	%f131, [%rd22+-4];
	sub.f32 	%f132, %f127, %f131;
	div.full.f32 	%f133, %f129, %f130;
	div.full.f32 	%f134, %f132, %f117;
	sub.f32 	%f135, %f133, %f134;
	st.shared.f32 	[%rd19+0], %f135;
	.loc	18	62	0
	mov.f32 	%f136, 0f40c00000;   	// 6
	mul.f32 	%f137, %f135, %f136;
	div.full.f32 	%f138, %f137, %f118;
	ld.shared.f32 	%f139, [%rd19+-4];
	mul.f32 	%f140, %f139, %f119;
	sub.f32 	%f141, %f138, %f140;
	div.full.f32 	%f142, %f141, %f122;
	st.shared.f32 	[%rd19+0], %f142;
$Lt_2_91650:
$L_2_52482:
 //<loop> Part of loop body line 48, head labeled $Lt_2_59138
	.loc	18	56	0
	add.s32 	%r71, %r71, 1;
	add.u64 	%rd16, %rd16, 4;
	setp.ne.s32 	%p11, %r67, %r71;
	@%p11 bra 	$Lt_2_59138;
$Lt_2_58626:
	mov.u64 	%rd18, u__0;
	mov.u64 	%rd17, sMasterPointsY2__4;
	.loc	18	69	0
	mov.f32 	%f143, 0f00000000;   	// 0
	mul.lo.u64 	%rd23, %rd14, 4;
	add.u64 	%rd24, %rd17, %rd23;
	st.shared.f32 	[%rd24+-4], %f143;
	.loc	18	72	0
	mov.s32 	%r73, %r68;
	mov.u32 	%r74, 0;
	setp.lt.s32 	%p12, %r68, %r74;
	@%p12 bra 	$Lt_2_60162;
	mov.s32 	%r75, %r67;
	cvt.u64.s32 	%rd25, %r68;
	mul.lo.u64 	%rd26, %rd25, 4;
	add.u64 	%rd27, %rd26, %rd18;
	add.u64 	%rd28, %rd26, %rd17;
	mov.s32 	%r76, %r75;
$Lt_2_60674:
 //<loop> Loop body line 72, nesting depth: 1, estimated iterations: unknown
	.loc	18	74	0
	ld.shared.f32 	%f144, [%rd27+0];
	ld.shared.f32 	%f145, [%rd28+0];
	ld.shared.f32 	%f146, [%rd28+4];
	mad.f32 	%f147, %f145, %f146, %f144;
	st.shared.f32 	[%rd28+0], %f147;
	sub.s32 	%r73, %r73, 1;
	sub.u64 	%rd28, %rd28, 4;
	sub.u64 	%rd27, %rd27, 4;
	mov.u32 	%r77, -1;
	setp.ne.s32 	%p13, %r73, %r77;
	@%p13 bra 	$Lt_2_60674;
$Lt_2_60162:
	.loc	19	226	0
	mov.u64 	%rd29, sRedPointsX__6;
	.loc	19	229	0
	mov.s32 	%r78, 16;
	st.shared.s32 	[sRedPointCount__5], %r78;
	ld.param.u8 	%r79, [__cudaparm_RGBCurves_CompositeKernel_inRedPointX0];
	cvt.rn.f32.u32 	%f148, %r79;
	mov.f32 	%f149, 0f3b808081;   	// 0.00392157
	mul.f32 	%f150, %f148, %f149;
	st.shared.f32 	[sRedPointsX__6+0], %f150;
	ld.param.u8 	%r80, [__cudaparm_RGBCurves_CompositeKernel_inRedPointX1];
	cvt.rn.f32.u32 	%f151, %r80;
	mov.f32 	%f152, 0f3b808081;   	// 0.00392157
	mul.f32 	%f153, %f151, %f152;
	st.shared.f32 	[sRedPointsX__6+4], %f153;
	ld.param.u8 	%r81, [__cudaparm_RGBCurves_CompositeKernel_inRedPointX2];
	cvt.rn.f32.u32 	%f154, %r81;
	mov.f32 	%f155, 0f3b808081;   	// 0.00392157
	mul.f32 	%f156, %f154, %f155;
	st.shared.f32 	[sRedPointsX__6+8], %f156;
	ld.param.u8 	%r82, [__cudaparm_RGBCurves_CompositeKernel_inRedPointX3];
	cvt.rn.f32.u32 	%f157, %r82;
	mov.f32 	%f158, 0f3b808081;   	// 0.00392157
	mul.f32 	%f159, %f157, %f158;
	st.shared.f32 	[sRedPointsX__6+12], %f159;
	ld.param.u8 	%r83, [__cudaparm_RGBCurves_CompositeKernel_inRedPointX4];
	cvt.rn.f32.u32 	%f160, %r83;
	mov.f32 	%f161, 0f3b808081;   	// 0.00392157
	mul.f32 	%f162, %f160, %f161;
	st.shared.f32 	[sRedPointsX__6+16], %f162;
	ld.param.u8 	%r84, [__cudaparm_RGBCurves_CompositeKernel_inRedPointX5];
	cvt.rn.f32.u32 	%f163, %r84;
	mov.f32 	%f164, 0f3b808081;   	// 0.00392157
	mul.f32 	%f165, %f163, %f164;
	st.shared.f32 	[sRedPointsX__6+20], %f165;
	ld.param.u8 	%r85, [__cudaparm_RGBCurves_CompositeKernel_inRedPointX6];
	cvt.rn.f32.u32 	%f166, %r85;
	mov.f32 	%f167, 0f3b808081;   	// 0.00392157
	mul.f32 	%f168, %f166, %f167;
	st.shared.f32 	[sRedPointsX__6+24], %f168;
	ld.param.u8 	%r86, [__cudaparm_RGBCurves_CompositeKernel_inRedPointX7];
	cvt.rn.f32.u32 	%f169, %r86;
	mov.f32 	%f170, 0f3b808081;   	// 0.00392157
	mul.f32 	%f171, %f169, %f170;
	st.shared.f32 	[sRedPointsX__6+28], %f171;
	ld.param.u8 	%r87, [__cudaparm_RGBCurves_CompositeKernel_inRedPointX8];
	cvt.rn.f32.u32 	%f172, %r87;
	mov.f32 	%f173, 0f3b808081;   	// 0.00392157
	mul.f32 	%f174, %f172, %f173;
	st.shared.f32 	[sRedPointsX__6+32], %f174;
	ld.param.u8 	%r88, [__cudaparm_RGBCurves_CompositeKernel_inRedPointX9];
	cvt.rn.f32.u32 	%f175, %r88;
	mov.f32 	%f176, 0f3b808081;   	// 0.00392157
	mul.f32 	%f177, %f175, %f176;
	st.shared.f32 	[sRedPointsX__6+36], %f177;
	ld.param.u8 	%r89, [__cudaparm_RGBCurves_CompositeKernel_inRedPointX10];
	cvt.rn.f32.u32 	%f178, %r89;
	mov.f32 	%f179, 0f3b808081;   	// 0.00392157
	mul.f32 	%f180, %f178, %f179;
	st.shared.f32 	[sRedPointsX__6+40], %f180;
	ld.param.u8 	%r90, [__cudaparm_RGBCurves_CompositeKernel_inRedPointX11];
	cvt.rn.f32.u32 	%f181, %r90;
	mov.f32 	%f182, 0f3b808081;   	// 0.00392157
	mul.f32 	%f183, %f181, %f182;
	st.shared.f32 	[sRedPointsX__6+44], %f183;
	ld.param.u8 	%r91, [__cudaparm_RGBCurves_CompositeKernel_inRedPointX12];
	cvt.rn.f32.u32 	%f184, %r91;
	mov.f32 	%f185, 0f3b808081;   	// 0.00392157
	mul.f32 	%f186, %f184, %f185;
	st.shared.f32 	[sRedPointsX__6+48], %f186;
	ld.param.u8 	%r92, [__cudaparm_RGBCurves_CompositeKernel_inRedPointX13];
	cvt.rn.f32.u32 	%f187, %r92;
	mov.f32 	%f188, 0f3b808081;   	// 0.00392157
	mul.f32 	%f189, %f187, %f188;
	st.shared.f32 	[sRedPointsX__6+52], %f189;
	ld.param.u8 	%r93, [__cudaparm_RGBCurves_CompositeKernel_inRedPointX14];
	cvt.rn.f32.u32 	%f190, %r93;
	mov.f32 	%f191, 0f3b808081;   	// 0.00392157
	mul.f32 	%f192, %f190, %f191;
	st.shared.f32 	[sRedPointsX__6+56], %f192;
	ld.param.u8 	%r94, [__cudaparm_RGBCurves_CompositeKernel_inRedPointX15];
	cvt.rn.f32.u32 	%f193, %r94;
	mov.f32 	%f194, 0f3b808081;   	// 0.00392157
	mul.f32 	%f195, %f193, %f194;
	st.shared.f32 	[sRedPointsX__6+60], %f195;
	ld.param.u8 	%r95, [__cudaparm_RGBCurves_CompositeKernel_inRedPointY0];
	cvt.rn.f32.u32 	%f196, %r95;
	mov.f32 	%f197, 0f3b808081;   	// 0.00392157
	mul.f32 	%f198, %f196, %f197;
	st.shared.f32 	[sRedPointsY__7+0], %f198;
	ld.param.u8 	%r96, [__cudaparm_RGBCurves_CompositeKernel_inRedPointY1];
	cvt.rn.f32.u32 	%f199, %r96;
	mov.f32 	%f200, 0f3b808081;   	// 0.00392157
	mul.f32 	%f201, %f199, %f200;
	st.shared.f32 	[sRedPointsY__7+4], %f201;
	ld.param.u8 	%r97, [__cudaparm_RGBCurves_CompositeKernel_inRedPointY2];
	cvt.rn.f32.u32 	%f202, %r97;
	mov.f32 	%f203, 0f3b808081;   	// 0.00392157
	mul.f32 	%f204, %f202, %f203;
	st.shared.f32 	[sRedPointsY__7+8], %f204;
	ld.param.u8 	%r98, [__cudaparm_RGBCurves_CompositeKernel_inRedPointY3];
	cvt.rn.f32.u32 	%f205, %r98;
	mov.f32 	%f206, 0f3b808081;   	// 0.00392157
	mul.f32 	%f207, %f205, %f206;
	st.shared.f32 	[sRedPointsY__7+12], %f207;
	ld.param.u8 	%r99, [__cudaparm_RGBCurves_CompositeKernel_inRedPointY4];
	cvt.rn.f32.u32 	%f208, %r99;
	mov.f32 	%f209, 0f3b808081;   	// 0.00392157
	mul.f32 	%f210, %f208, %f209;
	st.shared.f32 	[sRedPointsY__7+16], %f210;
	ld.param.u8 	%r100, [__cudaparm_RGBCurves_CompositeKernel_inRedPointY5];
	cvt.rn.f32.u32 	%f211, %r100;
	mov.f32 	%f212, 0f3b808081;   	// 0.00392157
	mul.f32 	%f213, %f211, %f212;
	st.shared.f32 	[sRedPointsY__7+20], %f213;
	ld.param.u8 	%r101, [__cudaparm_RGBCurves_CompositeKernel_inRedPointY6];
	cvt.rn.f32.u32 	%f214, %r101;
	mov.f32 	%f215, 0f3b808081;   	// 0.00392157
	mul.f32 	%f216, %f214, %f215;
	st.shared.f32 	[sRedPointsY__7+24], %f216;
	ld.param.u8 	%r102, [__cudaparm_RGBCurves_CompositeKernel_inRedPointY7];
	cvt.rn.f32.u32 	%f217, %r102;
	mov.f32 	%f218, 0f3b808081;   	// 0.00392157
	mul.f32 	%f219, %f217, %f218;
	st.shared.f32 	[sRedPointsY__7+28], %f219;
	ld.param.u8 	%r103, [__cudaparm_RGBCurves_CompositeKernel_inRedPointY8];
	cvt.rn.f32.u32 	%f220, %r103;
	mov.f32 	%f221, 0f3b808081;   	// 0.00392157
	mul.f32 	%f222, %f220, %f221;
	st.shared.f32 	[sRedPointsY__7+32], %f222;
	ld.param.u8 	%r104, [__cudaparm_RGBCurves_CompositeKernel_inRedPointY9];
	cvt.rn.f32.u32 	%f223, %r104;
	mov.f32 	%f224, 0f3b808081;   	// 0.00392157
	mul.f32 	%f225, %f223, %f224;
	st.shared.f32 	[sRedPointsY__7+36], %f225;
	ld.param.u8 	%r105, [__cudaparm_RGBCurves_CompositeKernel_inRedPointY10];
	cvt.rn.f32.u32 	%f226, %r105;
	mov.f32 	%f227, 0f3b808081;   	// 0.00392157
	mul.f32 	%f228, %f226, %f227;
	st.shared.f32 	[sRedPointsY__7+40], %f228;
	ld.param.u8 	%r106, [__cudaparm_RGBCurves_CompositeKernel_inRedPointY11];
	cvt.rn.f32.u32 	%f229, %r106;
	mov.f32 	%f230, 0f3b808081;   	// 0.00392157
	mul.f32 	%f231, %f229, %f230;
	st.shared.f32 	[sRedPointsY__7+44], %f231;
	ld.param.u8 	%r107, [__cudaparm_RGBCurves_CompositeKernel_inRedPointY12];
	cvt.rn.f32.u32 	%f232, %r107;
	mov.f32 	%f233, 0f3b808081;   	// 0.00392157
	mul.f32 	%f234, %f232, %f233;
	st.shared.f32 	[sRedPointsY__7+48], %f234;
	ld.param.u8 	%r108, [__cudaparm_RGBCurves_CompositeKernel_inRedPointY13];
	cvt.rn.f32.u32 	%f235, %r108;
	mov.f32 	%f236, 0f3b808081;   	// 0.00392157
	mul.f32 	%f237, %f235, %f236;
	st.shared.f32 	[sRedPointsY__7+52], %f237;
	ld.param.u8 	%r109, [__cudaparm_RGBCurves_CompositeKernel_inRedPointY14];
	cvt.rn.f32.u32 	%f238, %r109;
	mov.f32 	%f239, 0f3b808081;   	// 0.00392157
	mul.f32 	%f240, %f238, %f239;
	st.shared.f32 	[sRedPointsY__7+56], %f240;
	ld.param.u8 	%r110, [__cudaparm_RGBCurves_CompositeKernel_inRedPointY15];
	cvt.rn.f32.u32 	%f241, %r110;
	mov.f32 	%f242, 0f3b808081;   	// 0.00392157
	mul.f32 	%f243, %f241, %f242;
	st.shared.f32 	[sRedPointsY__7+60], %f243;
	mov.u64 	%rd30, 0;
	mov.s64 	%rd31, %rd29;
	mov.s32 	%r111, 0;
	mov.u64 	%rd32, sRedPointsY__7;
$Lt_2_61698:
 //<loop> Loop body line 229, nesting depth: 1, iterations: 16
	.loc	19	264	0
	ld.shared.f32 	%f244, [%rd31+0];
	mov.f32 	%f245, 0f00000000;   	// 0
	setp.neu.f32 	%p14, %f244, %f245;
	@%p14 bra 	$L_2_53250;
 //<loop> Part of loop body line 229, head labeled $Lt_2_61698
	add.u64 	%rd33, %rd30, %rd32;
	ld.shared.f32 	%f246, [%rd33+0];
	mov.f32 	%f247, 0f00000000;   	// 0
	setp.neu.f32 	%p15, %f246, %f247;
	@!%p15 bra 	$L_2_52994;
$L_2_53250:
 //<loop> Part of loop body line 229, head labeled $Lt_2_61698
	.loc	19	266	0
	add.s32 	%r78, %r111, 1;
	st.shared.s32 	[sRedPointCount__5], %r78;
$L_2_52994:
 //<loop> Part of loop body line 229, head labeled $Lt_2_61698
	add.s32 	%r112, %r111, 1;
	mov.s32 	%r111, %r112;
	add.u64 	%rd30, %rd30, 4;
	add.u64 	%rd31, %rd31, 4;
	mov.u32 	%r113, 16;
	setp.ne.s32 	%p16, %r112, %r113;
	@%p16 bra 	$Lt_2_61698;
	.loc	19	269	0
	mov.f32 	%f248, 0f00000000;   	// 0
	st.shared.f32 	[u__0+0], %f248;
	.loc	18	48	0
	mov.f32 	%f249, 0f00000000;   	// 0
	st.shared.f32 	[sRedPointsY2__8+0], %f249;
	cvt.u64.s32 	%rd34, %r78;
	mov.s64 	%rd35, %rd34;
	cvt.u32.u64 	%r114, %rd35;
	sub.s32 	%r115, %r114, 1;
	sub.s32 	%r116, %r114, 2;
	mov.u32 	%r117, 1;
	setp.le.s32 	%p17, %r115, %r117;
	@%p17 bra 	$Lt_2_62210;
	mov.s32 	%r118, %r116;
	mov.u64 	%rd16, 4;
	mov.s32 	%r71, 1;
	mov.u64 	%rd36, sRedPointsY2__8;
	mov.s32 	%r119, %r118;
$Lt_2_62722:
 //<loop> Loop body line 48, nesting depth: 1, estimated iterations: unknown
	.loc	18	53	0
	add.u64 	%rd19, %rd16, %rd18;
	mov.f32 	%f250, 0f00000000;   	// 0
	st.shared.f32 	[%rd19+0], %f250;
	add.u64 	%rd37, %rd16, %rd36;
	mov.f32 	%f251, 0f00000000;   	// 0
	st.shared.f32 	[%rd37+0], %f251;
	add.u64 	%rd38, %rd16, %rd29;
	ld.shared.f32 	%f252, [%rd38+-4];
	ld.shared.f32 	%f253, [%rd38+0];
	setp.lt.f32 	%p18, %f252, %f253;
	@!%p18 bra 	$Lt_2_93186;
 //<loop> Part of loop body line 48, head labeled $Lt_2_62722
	ld.shared.f32 	%f254, [%rd38+4];
	ld.shared.f32 	%f253, [%rd38+0];
	setp.lt.f32 	%p19, %f253, %f254;
	@!%p19 bra 	$Lt_2_93186;
 //<loop> Part of loop body line 48, head labeled $Lt_2_62722
	ld.shared.f32 	%f253, [%rd38+0];
	ld.shared.f32 	%f252, [%rd38+-4];
	.loc	18	56	0
	sub.f32 	%f255, %f253, %f252;
	.loc	18	53	0
	ld.shared.f32 	%f254, [%rd38+4];
	.loc	18	56	0
	sub.f32 	%f256, %f254, %f252;
	.loc	18	57	0
	div.full.f32 	%f119, %f255, %f256;
	mov.f32 	%f257, 0f40000000;   	// 2
	ld.shared.f32 	%f258, [%rd37+-4];
	mad.f32 	%f122, %f258, %f119, %f257;
	mov.f32 	%f259, 0f00000000;   	// 0
	setp.neu.f32 	%p20, %f122, %f259;
	@!%p20 bra 	$Lt_2_93186;
 //<loop> Part of loop body line 48, head labeled $Lt_2_62722
	.loc	18	60	0
	mov.f32 	%f260, 0fbf800000;   	// -1
	add.f32 	%f261, %f119, %f260;
	div.full.f32 	%f262, %f261, %f122;
	st.shared.f32 	[%rd37+0], %f262;
	.loc	18	61	0
	add.u64 	%rd39, %rd16, %rd32;
	ld.shared.f32 	%f263, [%rd39+0];
	ld.shared.f32 	%f264, [%rd39+4];
	sub.f32 	%f129, %f264, %f263;
	sub.f32 	%f130, %f254, %f253;
	ld.shared.f32 	%f265, [%rd39+-4];
	sub.f32 	%f132, %f263, %f265;
	div.full.f32 	%f133, %f129, %f130;
	div.full.f32 	%f134, %f132, %f255;
	sub.f32 	%f135, %f133, %f134;
	st.shared.f32 	[%rd19+0], %f135;
	.loc	18	62	0
	mov.f32 	%f266, 0f40c00000;   	// 6
	mul.f32 	%f267, %f135, %f266;
	div.full.f32 	%f268, %f267, %f256;
	ld.shared.f32 	%f269, [%rd19+-4];
	mul.f32 	%f270, %f269, %f119;
	sub.f32 	%f271, %f268, %f270;
	div.full.f32 	%f272, %f271, %f122;
	st.shared.f32 	[%rd19+0], %f272;
$Lt_2_93186:
$L_2_53506:
 //<loop> Part of loop body line 48, head labeled $Lt_2_62722
	.loc	18	56	0
	add.s32 	%r71, %r71, 1;
	add.u64 	%rd16, %rd16, 4;
	setp.ne.s32 	%p21, %r115, %r71;
	@%p21 bra 	$Lt_2_62722;
$Lt_2_62210:
	mov.u64 	%rd36, sRedPointsY2__8;
	.loc	18	69	0
	mov.f32 	%f273, 0f00000000;   	// 0
	mul.lo.u64 	%rd40, %rd34, 4;
	add.u64 	%rd41, %rd36, %rd40;
	st.shared.f32 	[%rd41+-4], %f273;
	.loc	18	72	0
	mov.s32 	%r73, %r116;
	mov.u32 	%r120, 0;
	setp.lt.s32 	%p22, %r116, %r120;
	@%p22 bra 	$Lt_2_63746;
	mov.s32 	%r121, %r115;
	cvt.u64.s32 	%rd42, %r116;
	mul.lo.u64 	%rd26, %rd42, 4;
	add.u64 	%rd27, %rd26, %rd18;
	add.u64 	%rd43, %rd26, %rd36;
	mov.s32 	%r122, %r121;
$Lt_2_64258:
 //<loop> Loop body line 72, nesting depth: 1, estimated iterations: unknown
	.loc	18	74	0
	ld.shared.f32 	%f274, [%rd27+0];
	ld.shared.f32 	%f275, [%rd43+0];
	ld.shared.f32 	%f276, [%rd43+4];
	mad.f32 	%f277, %f275, %f276, %f274;
	st.shared.f32 	[%rd43+0], %f277;
	sub.s32 	%r73, %r73, 1;
	sub.u64 	%rd43, %rd43, 4;
	sub.u64 	%rd27, %rd27, 4;
	mov.u32 	%r123, -1;
	setp.ne.s32 	%p23, %r73, %r123;
	@%p23 bra 	$Lt_2_64258;
$Lt_2_63746:
	.loc	19	269	0
	mov.u64 	%rd44, sGreenPointsX__10;
	.loc	19	272	0
	mov.s32 	%r124, 16;
	st.shared.s32 	[sGreenPointCount__9], %r124;
	ld.param.u8 	%r125, [__cudaparm_RGBCurves_CompositeKernel_inGreenPointX0];
	cvt.rn.f32.u32 	%f278, %r125;
	mov.f32 	%f279, 0f3b808081;   	// 0.00392157
	mul.f32 	%f280, %f278, %f279;
	st.shared.f32 	[sGreenPointsX__10+0], %f280;
	ld.param.u8 	%r126, [__cudaparm_RGBCurves_CompositeKernel_inGreenPointX1];
	cvt.rn.f32.u32 	%f281, %r126;
	mov.f32 	%f282, 0f3b808081;   	// 0.00392157
	mul.f32 	%f283, %f281, %f282;
	st.shared.f32 	[sGreenPointsX__10+4], %f283;
	ld.param.u8 	%r127, [__cudaparm_RGBCurves_CompositeKernel_inGreenPointX2];
	cvt.rn.f32.u32 	%f284, %r127;
	mov.f32 	%f285, 0f3b808081;   	// 0.00392157
	mul.f32 	%f286, %f284, %f285;
	st.shared.f32 	[sGreenPointsX__10+8], %f286;
	ld.param.u8 	%r128, [__cudaparm_RGBCurves_CompositeKernel_inGreenPointX3];
	cvt.rn.f32.u32 	%f287, %r128;
	mov.f32 	%f288, 0f3b808081;   	// 0.00392157
	mul.f32 	%f289, %f287, %f288;
	st.shared.f32 	[sGreenPointsX__10+12], %f289;
	ld.param.u8 	%r129, [__cudaparm_RGBCurves_CompositeKernel_inGreenPointX4];
	cvt.rn.f32.u32 	%f290, %r129;
	mov.f32 	%f291, 0f3b808081;   	// 0.00392157
	mul.f32 	%f292, %f290, %f291;
	st.shared.f32 	[sGreenPointsX__10+16], %f292;
	ld.param.u8 	%r130, [__cudaparm_RGBCurves_CompositeKernel_inGreenPointX5];
	cvt.rn.f32.u32 	%f293, %r130;
	mov.f32 	%f294, 0f3b808081;   	// 0.00392157
	mul.f32 	%f295, %f293, %f294;
	st.shared.f32 	[sGreenPointsX__10+20], %f295;
	ld.param.u8 	%r131, [__cudaparm_RGBCurves_CompositeKernel_inGreenPointX6];
	cvt.rn.f32.u32 	%f296, %r131;
	mov.f32 	%f297, 0f3b808081;   	// 0.00392157
	mul.f32 	%f298, %f296, %f297;
	st.shared.f32 	[sGreenPointsX__10+24], %f298;
	ld.param.u8 	%r132, [__cudaparm_RGBCurves_CompositeKernel_inGreenPointX7];
	cvt.rn.f32.u32 	%f299, %r132;
	mov.f32 	%f300, 0f3b808081;   	// 0.00392157
	mul.f32 	%f301, %f299, %f300;
	st.shared.f32 	[sGreenPointsX__10+28], %f301;
	ld.param.u8 	%r133, [__cudaparm_RGBCurves_CompositeKernel_inGreenPointX8];
	cvt.rn.f32.u32 	%f302, %r133;
	mov.f32 	%f303, 0f3b808081;   	// 0.00392157
	mul.f32 	%f304, %f302, %f303;
	st.shared.f32 	[sGreenPointsX__10+32], %f304;
	ld.param.u8 	%r134, [__cudaparm_RGBCurves_CompositeKernel_inGreenPointX9];
	cvt.rn.f32.u32 	%f305, %r134;
	mov.f32 	%f306, 0f3b808081;   	// 0.00392157
	mul.f32 	%f307, %f305, %f306;
	st.shared.f32 	[sGreenPointsX__10+36], %f307;
	ld.param.u8 	%r135, [__cudaparm_RGBCurves_CompositeKernel_inGreenPointX10];
	cvt.rn.f32.u32 	%f308, %r135;
	mov.f32 	%f309, 0f3b808081;   	// 0.00392157
	mul.f32 	%f310, %f308, %f309;
	st.shared.f32 	[sGreenPointsX__10+40], %f310;
	ld.param.u8 	%r136, [__cudaparm_RGBCurves_CompositeKernel_inGreenPointX11];
	cvt.rn.f32.u32 	%f311, %r136;
	mov.f32 	%f312, 0f3b808081;   	// 0.00392157
	mul.f32 	%f313, %f311, %f312;
	st.shared.f32 	[sGreenPointsX__10+44], %f313;
	ld.param.u8 	%r137, [__cudaparm_RGBCurves_CompositeKernel_inGreenPointX12];
	cvt.rn.f32.u32 	%f314, %r137;
	mov.f32 	%f315, 0f3b808081;   	// 0.00392157
	mul.f32 	%f316, %f314, %f315;
	st.shared.f32 	[sGreenPointsX__10+48], %f316;
	ld.param.u8 	%r138, [__cudaparm_RGBCurves_CompositeKernel_inGreenPointX13];
	cvt.rn.f32.u32 	%f317, %r138;
	mov.f32 	%f318, 0f3b808081;   	// 0.00392157
	mul.f32 	%f319, %f317, %f318;
	st.shared.f32 	[sGreenPointsX__10+52], %f319;
	ld.param.u8 	%r139, [__cudaparm_RGBCurves_CompositeKernel_inGreenPointX14];
	cvt.rn.f32.u32 	%f320, %r139;
	mov.f32 	%f321, 0f3b808081;   	// 0.00392157
	mul.f32 	%f322, %f320, %f321;
	st.shared.f32 	[sGreenPointsX__10+56], %f322;
	ld.param.u8 	%r140, [__cudaparm_RGBCurves_CompositeKernel_inGreenPointX15];
	cvt.rn.f32.u32 	%f323, %r140;
	mov.f32 	%f324, 0f3b808081;   	// 0.00392157
	mul.f32 	%f325, %f323, %f324;
	st.shared.f32 	[sGreenPointsX__10+60], %f325;
	ld.param.u8 	%r141, [__cudaparm_RGBCurves_CompositeKernel_inGreenPointY0];
	cvt.rn.f32.u32 	%f326, %r141;
	mov.f32 	%f327, 0f3b808081;   	// 0.00392157
	mul.f32 	%f328, %f326, %f327;
	st.shared.f32 	[sGreenPointsY__11+0], %f328;
	ld.param.u8 	%r142, [__cudaparm_RGBCurves_CompositeKernel_inGreenPointY1];
	cvt.rn.f32.u32 	%f329, %r142;
	mov.f32 	%f330, 0f3b808081;   	// 0.00392157
	mul.f32 	%f331, %f329, %f330;
	st.shared.f32 	[sGreenPointsY__11+4], %f331;
	ld.param.u8 	%r143, [__cudaparm_RGBCurves_CompositeKernel_inGreenPointY2];
	cvt.rn.f32.u32 	%f332, %r143;
	mov.f32 	%f333, 0f3b808081;   	// 0.00392157
	mul.f32 	%f334, %f332, %f333;
	st.shared.f32 	[sGreenPointsY__11+8], %f334;
	ld.param.u8 	%r144, [__cudaparm_RGBCurves_CompositeKernel_inGreenPointY3];
	cvt.rn.f32.u32 	%f335, %r144;
	mov.f32 	%f336, 0f3b808081;   	// 0.00392157
	mul.f32 	%f337, %f335, %f336;
	st.shared.f32 	[sGreenPointsY__11+12], %f337;
	ld.param.u8 	%r145, [__cudaparm_RGBCurves_CompositeKernel_inGreenPointY4];
	cvt.rn.f32.u32 	%f338, %r145;
	mov.f32 	%f339, 0f3b808081;   	// 0.00392157
	mul.f32 	%f340, %f338, %f339;
	st.shared.f32 	[sGreenPointsY__11+16], %f340;
	ld.param.u8 	%r146, [__cudaparm_RGBCurves_CompositeKernel_inGreenPointY5];
	cvt.rn.f32.u32 	%f341, %r146;
	mov.f32 	%f342, 0f3b808081;   	// 0.00392157
	mul.f32 	%f343, %f341, %f342;
	st.shared.f32 	[sGreenPointsY__11+20], %f343;
	ld.param.u8 	%r147, [__cudaparm_RGBCurves_CompositeKernel_inGreenPointY6];
	cvt.rn.f32.u32 	%f344, %r147;
	mov.f32 	%f345, 0f3b808081;   	// 0.00392157
	mul.f32 	%f346, %f344, %f345;
	st.shared.f32 	[sGreenPointsY__11+24], %f346;
	ld.param.u8 	%r148, [__cudaparm_RGBCurves_CompositeKernel_inGreenPointY7];
	cvt.rn.f32.u32 	%f347, %r148;
	mov.f32 	%f348, 0f3b808081;   	// 0.00392157
	mul.f32 	%f349, %f347, %f348;
	st.shared.f32 	[sGreenPointsY__11+28], %f349;
	ld.param.u8 	%r149, [__cudaparm_RGBCurves_CompositeKernel_inGreenPointY8];
	cvt.rn.f32.u32 	%f350, %r149;
	mov.f32 	%f351, 0f3b808081;   	// 0.00392157
	mul.f32 	%f352, %f350, %f351;
	st.shared.f32 	[sGreenPointsY__11+32], %f352;
	ld.param.u8 	%r150, [__cudaparm_RGBCurves_CompositeKernel_inGreenPointY9];
	cvt.rn.f32.u32 	%f353, %r150;
	mov.f32 	%f354, 0f3b808081;   	// 0.00392157
	mul.f32 	%f355, %f353, %f354;
	st.shared.f32 	[sGreenPointsY__11+36], %f355;
	ld.param.u8 	%r151, [__cudaparm_RGBCurves_CompositeKernel_inGreenPointY10];
	cvt.rn.f32.u32 	%f356, %r151;
	mov.f32 	%f357, 0f3b808081;   	// 0.00392157
	mul.f32 	%f358, %f356, %f357;
	st.shared.f32 	[sGreenPointsY__11+40], %f358;
	ld.param.u8 	%r152, [__cudaparm_RGBCurves_CompositeKernel_inGreenPointY11];
	cvt.rn.f32.u32 	%f359, %r152;
	mov.f32 	%f360, 0f3b808081;   	// 0.00392157
	mul.f32 	%f361, %f359, %f360;
	st.shared.f32 	[sGreenPointsY__11+44], %f361;
	ld.param.u8 	%r153, [__cudaparm_RGBCurves_CompositeKernel_inGreenPointY12];
	cvt.rn.f32.u32 	%f362, %r153;
	mov.f32 	%f363, 0f3b808081;   	// 0.00392157
	mul.f32 	%f364, %f362, %f363;
	st.shared.f32 	[sGreenPointsY__11+48], %f364;
	ld.param.u8 	%r154, [__cudaparm_RGBCurves_CompositeKernel_inGreenPointY13];
	cvt.rn.f32.u32 	%f365, %r154;
	mov.f32 	%f366, 0f3b808081;   	// 0.00392157
	mul.f32 	%f367, %f365, %f366;
	st.shared.f32 	[sGreenPointsY__11+52], %f367;
	ld.param.u8 	%r155, [__cudaparm_RGBCurves_CompositeKernel_inGreenPointY14];
	cvt.rn.f32.u32 	%f368, %r155;
	mov.f32 	%f369, 0f3b808081;   	// 0.00392157
	mul.f32 	%f370, %f368, %f369;
	st.shared.f32 	[sGreenPointsY__11+56], %f370;
	ld.param.u8 	%r156, [__cudaparm_RGBCurves_CompositeKernel_inGreenPointY15];
	cvt.rn.f32.u32 	%f371, %r156;
	mov.f32 	%f372, 0f3b808081;   	// 0.00392157
	mul.f32 	%f373, %f371, %f372;
	st.shared.f32 	[sGreenPointsY__11+60], %f373;
	mov.u64 	%rd45, 0;
	mov.s64 	%rd46, %rd44;
	mov.s32 	%r157, 0;
	mov.u64 	%rd47, sGreenPointsY__11;
$Lt_2_65282:
 //<loop> Loop body line 272, nesting depth: 1, iterations: 16
	.loc	19	307	0
	ld.shared.f32 	%f374, [%rd46+0];
	mov.f32 	%f375, 0f00000000;   	// 0
	setp.neu.f32 	%p24, %f374, %f375;
	@%p24 bra 	$L_2_54274;
 //<loop> Part of loop body line 272, head labeled $Lt_2_65282
	add.u64 	%rd48, %rd45, %rd47;
	ld.shared.f32 	%f376, [%rd48+0];
	mov.f32 	%f377, 0f00000000;   	// 0
	setp.neu.f32 	%p25, %f376, %f377;
	@!%p25 bra 	$L_2_54018;
$L_2_54274:
 //<loop> Part of loop body line 272, head labeled $Lt_2_65282
	.loc	19	309	0
	add.s32 	%r124, %r157, 1;
	st.shared.s32 	[sGreenPointCount__9], %r124;
$L_2_54018:
 //<loop> Part of loop body line 272, head labeled $Lt_2_65282
	add.s32 	%r158, %r157, 1;
	mov.s32 	%r157, %r158;
	add.u64 	%rd45, %rd45, 4;
	add.u64 	%rd46, %rd46, 4;
	mov.u32 	%r159, 16;
	setp.ne.s32 	%p26, %r158, %r159;
	@%p26 bra 	$Lt_2_65282;
	.loc	19	312	0
	mov.f32 	%f378, 0f00000000;   	// 0
	st.shared.f32 	[u__0+0], %f378;
	.loc	18	48	0
	mov.f32 	%f379, 0f00000000;   	// 0
	st.shared.f32 	[sGreenPointsY2__12+0], %f379;
	cvt.u64.s32 	%rd49, %r124;
	mov.s64 	%rd50, %rd49;
	cvt.u32.u64 	%r160, %rd50;
	sub.s32 	%r161, %r160, 1;
	sub.s32 	%r162, %r160, 2;
	mov.u32 	%r163, 1;
	setp.le.s32 	%p27, %r161, %r163;
	@%p27 bra 	$Lt_2_65794;
	mov.s32 	%r164, %r162;
	mov.u64 	%rd16, 4;
	mov.s32 	%r71, 1;
	mov.u64 	%rd51, sGreenPointsY2__12;
	mov.s32 	%r165, %r164;
$Lt_2_66306:
 //<loop> Loop body line 48, nesting depth: 1, estimated iterations: unknown
	.loc	18	53	0
	add.u64 	%rd19, %rd16, %rd18;
	mov.f32 	%f380, 0f00000000;   	// 0
	st.shared.f32 	[%rd19+0], %f380;
	add.u64 	%rd52, %rd16, %rd51;
	mov.f32 	%f381, 0f00000000;   	// 0
	st.shared.f32 	[%rd52+0], %f381;
	add.u64 	%rd53, %rd16, %rd44;
	ld.shared.f32 	%f382, [%rd53+-4];
	ld.shared.f32 	%f383, [%rd53+0];
	setp.lt.f32 	%p28, %f382, %f383;
	@!%p28 bra 	$Lt_2_94722;
 //<loop> Part of loop body line 48, head labeled $Lt_2_66306
	ld.shared.f32 	%f384, [%rd53+4];
	ld.shared.f32 	%f383, [%rd53+0];
	setp.lt.f32 	%p29, %f383, %f384;
	@!%p29 bra 	$Lt_2_94722;
 //<loop> Part of loop body line 48, head labeled $Lt_2_66306
	ld.shared.f32 	%f383, [%rd53+0];
	ld.shared.f32 	%f382, [%rd53+-4];
	.loc	18	56	0
	sub.f32 	%f385, %f383, %f382;
	.loc	18	53	0
	ld.shared.f32 	%f384, [%rd53+4];
	.loc	18	56	0
	sub.f32 	%f386, %f384, %f382;
	.loc	18	57	0
	div.full.f32 	%f119, %f385, %f386;
	mov.f32 	%f387, 0f40000000;   	// 2
	ld.shared.f32 	%f388, [%rd52+-4];
	mad.f32 	%f122, %f388, %f119, %f387;
	mov.f32 	%f389, 0f00000000;   	// 0
	setp.neu.f32 	%p30, %f122, %f389;
	@!%p30 bra 	$Lt_2_94722;
 //<loop> Part of loop body line 48, head labeled $Lt_2_66306
	.loc	18	60	0
	mov.f32 	%f390, 0fbf800000;   	// -1
	add.f32 	%f391, %f119, %f390;
	div.full.f32 	%f392, %f391, %f122;
	st.shared.f32 	[%rd52+0], %f392;
	.loc	18	61	0
	add.u64 	%rd54, %rd16, %rd47;
	ld.shared.f32 	%f393, [%rd54+0];
	ld.shared.f32 	%f394, [%rd54+4];
	sub.f32 	%f129, %f394, %f393;
	sub.f32 	%f130, %f384, %f383;
	ld.shared.f32 	%f395, [%rd54+-4];
	sub.f32 	%f132, %f393, %f395;
	div.full.f32 	%f133, %f129, %f130;
	div.full.f32 	%f134, %f132, %f385;
	sub.f32 	%f135, %f133, %f134;
	st.shared.f32 	[%rd19+0], %f135;
	.loc	18	62	0
	mov.f32 	%f396, 0f40c00000;   	// 6
	mul.f32 	%f397, %f135, %f396;
	div.full.f32 	%f398, %f397, %f386;
	ld.shared.f32 	%f399, [%rd19+-4];
	mul.f32 	%f400, %f399, %f119;
	sub.f32 	%f401, %f398, %f400;
	div.full.f32 	%f402, %f401, %f122;
	st.shared.f32 	[%rd19+0], %f402;
$Lt_2_94722:
$L_2_54530:
 //<loop> Part of loop body line 48, head labeled $Lt_2_66306
	.loc	18	56	0
	add.s32 	%r71, %r71, 1;
	add.u64 	%rd16, %rd16, 4;
	setp.ne.s32 	%p31, %r161, %r71;
	@%p31 bra 	$Lt_2_66306;
$Lt_2_65794:
	mov.u64 	%rd51, sGreenPointsY2__12;
	.loc	18	69	0
	mov.f32 	%f403, 0f00000000;   	// 0
	mul.lo.u64 	%rd55, %rd49, 4;
	add.u64 	%rd56, %rd51, %rd55;
	st.shared.f32 	[%rd56+-4], %f403;
	.loc	18	72	0
	mov.s32 	%r73, %r162;
	mov.u32 	%r166, 0;
	setp.lt.s32 	%p32, %r162, %r166;
	@%p32 bra 	$Lt_2_67330;
	mov.s32 	%r167, %r161;
	cvt.u64.s32 	%rd57, %r162;
	mul.lo.u64 	%rd26, %rd57, 4;
	add.u64 	%rd27, %rd26, %rd18;
	add.u64 	%rd58, %rd26, %rd51;
	mov.s32 	%r168, %r167;
$Lt_2_67842:
 //<loop> Loop body line 72, nesting depth: 1, estimated iterations: unknown
	.loc	18	74	0
	ld.shared.f32 	%f404, [%rd27+0];
	ld.shared.f32 	%f405, [%rd58+0];
	ld.shared.f32 	%f406, [%rd58+4];
	mad.f32 	%f407, %f405, %f406, %f404;
	st.shared.f32 	[%rd58+0], %f407;
	sub.s32 	%r73, %r73, 1;
	sub.u64 	%rd58, %rd58, 4;
	sub.u64 	%rd27, %rd27, 4;
	mov.u32 	%r169, -1;
	setp.ne.s32 	%p33, %r73, %r169;
	@%p33 bra 	$Lt_2_67842;
$Lt_2_67330:
	.loc	19	312	0
	mov.u64 	%rd59, sBluePointsX__14;
	.loc	19	315	0
	mov.s32 	%r170, 16;
	st.shared.s32 	[sBluePointCount__13], %r170;
	ld.param.u8 	%r171, [__cudaparm_RGBCurves_CompositeKernel_inBluePointX0];
	cvt.rn.f32.u32 	%f408, %r171;
	mov.f32 	%f409, 0f3b808081;   	// 0.00392157
	mul.f32 	%f410, %f408, %f409;
	st.shared.f32 	[sBluePointsX__14+0], %f410;
	ld.param.u8 	%r172, [__cudaparm_RGBCurves_CompositeKernel_inBluePointX1];
	cvt.rn.f32.u32 	%f411, %r172;
	mov.f32 	%f412, 0f3b808081;   	// 0.00392157
	mul.f32 	%f413, %f411, %f412;
	st.shared.f32 	[sBluePointsX__14+4], %f413;
	ld.param.u8 	%r173, [__cudaparm_RGBCurves_CompositeKernel_inBluePointX2];
	cvt.rn.f32.u32 	%f414, %r173;
	mov.f32 	%f415, 0f3b808081;   	// 0.00392157
	mul.f32 	%f416, %f414, %f415;
	st.shared.f32 	[sBluePointsX__14+8], %f416;
	ld.param.u8 	%r174, [__cudaparm_RGBCurves_CompositeKernel_inBluePointX3];
	cvt.rn.f32.u32 	%f417, %r174;
	mov.f32 	%f418, 0f3b808081;   	// 0.00392157
	mul.f32 	%f419, %f417, %f418;
	st.shared.f32 	[sBluePointsX__14+12], %f419;
	ld.param.u8 	%r175, [__cudaparm_RGBCurves_CompositeKernel_inBluePointX4];
	cvt.rn.f32.u32 	%f420, %r175;
	mov.f32 	%f421, 0f3b808081;   	// 0.00392157
	mul.f32 	%f422, %f420, %f421;
	st.shared.f32 	[sBluePointsX__14+16], %f422;
	ld.param.u8 	%r176, [__cudaparm_RGBCurves_CompositeKernel_inBluePointX5];
	cvt.rn.f32.u32 	%f423, %r176;
	mov.f32 	%f424, 0f3b808081;   	// 0.00392157
	mul.f32 	%f425, %f423, %f424;
	st.shared.f32 	[sBluePointsX__14+20], %f425;
	ld.param.u8 	%r177, [__cudaparm_RGBCurves_CompositeKernel_inBluePointX6];
	cvt.rn.f32.u32 	%f426, %r177;
	mov.f32 	%f427, 0f3b808081;   	// 0.00392157
	mul.f32 	%f428, %f426, %f427;
	st.shared.f32 	[sBluePointsX__14+24], %f428;
	ld.param.u8 	%r178, [__cudaparm_RGBCurves_CompositeKernel_inBluePointX7];
	cvt.rn.f32.u32 	%f429, %r178;
	mov.f32 	%f430, 0f3b808081;   	// 0.00392157
	mul.f32 	%f431, %f429, %f430;
	st.shared.f32 	[sBluePointsX__14+28], %f431;
	ld.param.u8 	%r179, [__cudaparm_RGBCurves_CompositeKernel_inBluePointX8];
	cvt.rn.f32.u32 	%f432, %r179;
	mov.f32 	%f433, 0f3b808081;   	// 0.00392157
	mul.f32 	%f434, %f432, %f433;
	st.shared.f32 	[sBluePointsX__14+32], %f434;
	ld.param.u8 	%r180, [__cudaparm_RGBCurves_CompositeKernel_inBluePointX9];
	cvt.rn.f32.u32 	%f435, %r180;
	mov.f32 	%f436, 0f3b808081;   	// 0.00392157
	mul.f32 	%f437, %f435, %f436;
	st.shared.f32 	[sBluePointsX__14+36], %f437;
	ld.param.u8 	%r181, [__cudaparm_RGBCurves_CompositeKernel_inBluePointX10];
	cvt.rn.f32.u32 	%f438, %r181;
	mov.f32 	%f439, 0f3b808081;   	// 0.00392157
	mul.f32 	%f440, %f438, %f439;
	st.shared.f32 	[sBluePointsX__14+40], %f440;
	ld.param.u8 	%r182, [__cudaparm_RGBCurves_CompositeKernel_inBluePointX11];
	cvt.rn.f32.u32 	%f441, %r182;
	mov.f32 	%f442, 0f3b808081;   	// 0.00392157
	mul.f32 	%f443, %f441, %f442;
	st.shared.f32 	[sBluePointsX__14+44], %f443;
	ld.param.u8 	%r183, [__cudaparm_RGBCurves_CompositeKernel_inBluePointX12];
	cvt.rn.f32.u32 	%f444, %r183;
	mov.f32 	%f445, 0f3b808081;   	// 0.00392157
	mul.f32 	%f446, %f444, %f445;
	st.shared.f32 	[sBluePointsX__14+48], %f446;
	ld.param.u8 	%r184, [__cudaparm_RGBCurves_CompositeKernel_inBluePointX13];
	cvt.rn.f32.u32 	%f447, %r184;
	mov.f32 	%f448, 0f3b808081;   	// 0.00392157
	mul.f32 	%f449, %f447, %f448;
	st.shared.f32 	[sBluePointsX__14+52], %f449;
	ld.param.u8 	%r185, [__cudaparm_RGBCurves_CompositeKernel_inBluePointX14];
	cvt.rn.f32.u32 	%f450, %r185;
	mov.f32 	%f451, 0f3b808081;   	// 0.00392157
	mul.f32 	%f452, %f450, %f451;
	st.shared.f32 	[sBluePointsX__14+56], %f452;
	ld.param.u8 	%r186, [__cudaparm_RGBCurves_CompositeKernel_inBluePointX15];
	cvt.rn.f32.u32 	%f453, %r186;
	mov.f32 	%f454, 0f3b808081;   	// 0.00392157
	mul.f32 	%f455, %f453, %f454;
	st.shared.f32 	[sBluePointsX__14+60], %f455;
	ld.param.u8 	%r187, [__cudaparm_RGBCurves_CompositeKernel_inBluePointY0];
	cvt.rn.f32.u32 	%f456, %r187;
	mov.f32 	%f457, 0f3b808081;   	// 0.00392157
	mul.f32 	%f458, %f456, %f457;
	st.shared.f32 	[sBluePointsY__15+0], %f458;
	ld.param.u8 	%r188, [__cudaparm_RGBCurves_CompositeKernel_inBluePointY1];
	cvt.rn.f32.u32 	%f459, %r188;
	mov.f32 	%f460, 0f3b808081;   	// 0.00392157
	mul.f32 	%f461, %f459, %f460;
	st.shared.f32 	[sBluePointsY__15+4], %f461;
	ld.param.u8 	%r189, [__cudaparm_RGBCurves_CompositeKernel_inBluePointY2];
	cvt.rn.f32.u32 	%f462, %r189;
	mov.f32 	%f463, 0f3b808081;   	// 0.00392157
	mul.f32 	%f464, %f462, %f463;
	st.shared.f32 	[sBluePointsY__15+8], %f464;
	ld.param.u8 	%r190, [__cudaparm_RGBCurves_CompositeKernel_inBluePointY3];
	cvt.rn.f32.u32 	%f465, %r190;
	mov.f32 	%f466, 0f3b808081;   	// 0.00392157
	mul.f32 	%f467, %f465, %f466;
	st.shared.f32 	[sBluePointsY__15+12], %f467;
	ld.param.u8 	%r191, [__cudaparm_RGBCurves_CompositeKernel_inBluePointY4];
	cvt.rn.f32.u32 	%f468, %r191;
	mov.f32 	%f469, 0f3b808081;   	// 0.00392157
	mul.f32 	%f470, %f468, %f469;
	st.shared.f32 	[sBluePointsY__15+16], %f470;
	ld.param.u8 	%r192, [__cudaparm_RGBCurves_CompositeKernel_inBluePointY5];
	cvt.rn.f32.u32 	%f471, %r192;
	mov.f32 	%f472, 0f3b808081;   	// 0.00392157
	mul.f32 	%f473, %f471, %f472;
	st.shared.f32 	[sBluePointsY__15+20], %f473;
	ld.param.u8 	%r193, [__cudaparm_RGBCurves_CompositeKernel_inBluePointY6];
	cvt.rn.f32.u32 	%f474, %r193;
	mov.f32 	%f475, 0f3b808081;   	// 0.00392157
	mul.f32 	%f476, %f474, %f475;
	st.shared.f32 	[sBluePointsY__15+24], %f476;
	ld.param.u8 	%r194, [__cudaparm_RGBCurves_CompositeKernel_inBluePointY7];
	cvt.rn.f32.u32 	%f477, %r194;
	mov.f32 	%f478, 0f3b808081;   	// 0.00392157
	mul.f32 	%f479, %f477, %f478;
	st.shared.f32 	[sBluePointsY__15+28], %f479;
	ld.param.u8 	%r195, [__cudaparm_RGBCurves_CompositeKernel_inBluePointY8];
	cvt.rn.f32.u32 	%f480, %r195;
	mov.f32 	%f481, 0f3b808081;   	// 0.00392157
	mul.f32 	%f482, %f480, %f481;
	st.shared.f32 	[sBluePointsY__15+32], %f482;
	ld.param.u8 	%r196, [__cudaparm_RGBCurves_CompositeKernel_inBluePointY9];
	cvt.rn.f32.u32 	%f483, %r196;
	mov.f32 	%f484, 0f3b808081;   	// 0.00392157
	mul.f32 	%f485, %f483, %f484;
	st.shared.f32 	[sBluePointsY__15+36], %f485;
	ld.param.u8 	%r197, [__cudaparm_RGBCurves_CompositeKernel_inBluePointY10];
	cvt.rn.f32.u32 	%f486, %r197;
	mov.f32 	%f487, 0f3b808081;   	// 0.00392157
	mul.f32 	%f488, %f486, %f487;
	st.shared.f32 	[sBluePointsY__15+40], %f488;
	ld.param.u8 	%r198, [__cudaparm_RGBCurves_CompositeKernel_inBluePointY11];
	cvt.rn.f32.u32 	%f489, %r198;
	mov.f32 	%f490, 0f3b808081;   	// 0.00392157
	mul.f32 	%f491, %f489, %f490;
	st.shared.f32 	[sBluePointsY__15+44], %f491;
	ld.param.u8 	%r199, [__cudaparm_RGBCurves_CompositeKernel_inBluePointY12];
	cvt.rn.f32.u32 	%f492, %r199;
	mov.f32 	%f493, 0f3b808081;   	// 0.00392157
	mul.f32 	%f494, %f492, %f493;
	st.shared.f32 	[sBluePointsY__15+48], %f494;
	ld.param.u8 	%r200, [__cudaparm_RGBCurves_CompositeKernel_inBluePointY13];
	cvt.rn.f32.u32 	%f495, %r200;
	mov.f32 	%f496, 0f3b808081;   	// 0.00392157
	mul.f32 	%f497, %f495, %f496;
	st.shared.f32 	[sBluePointsY__15+52], %f497;
	ld.param.u8 	%r201, [__cudaparm_RGBCurves_CompositeKernel_inBluePointY14];
	cvt.rn.f32.u32 	%f498, %r201;
	mov.f32 	%f499, 0f3b808081;   	// 0.00392157
	mul.f32 	%f500, %f498, %f499;
	st.shared.f32 	[sBluePointsY__15+56], %f500;
	ld.param.u8 	%r202, [__cudaparm_RGBCurves_CompositeKernel_inBluePointY15];
	cvt.rn.f32.u32 	%f501, %r202;
	mov.f32 	%f502, 0f3b808081;   	// 0.00392157
	mul.f32 	%f503, %f501, %f502;
	st.shared.f32 	[sBluePointsY__15+60], %f503;
	mov.u64 	%rd60, 0;
	mov.s64 	%rd61, %rd59;
	mov.s32 	%r203, 0;
	mov.u64 	%rd62, sBluePointsY__15;
$Lt_2_68866:
 //<loop> Loop body line 315, nesting depth: 1, iterations: 16
	.loc	19	350	0
	ld.shared.f32 	%f504, [%rd61+0];
	mov.f32 	%f505, 0f00000000;   	// 0
	setp.neu.f32 	%p34, %f504, %f505;
	@%p34 bra 	$L_2_55298;
 //<loop> Part of loop body line 315, head labeled $Lt_2_68866
	add.u64 	%rd63, %rd60, %rd62;
	ld.shared.f32 	%f506, [%rd63+0];
	mov.f32 	%f507, 0f00000000;   	// 0
	setp.neu.f32 	%p35, %f506, %f507;
	@!%p35 bra 	$L_2_55042;
$L_2_55298:
 //<loop> Part of loop body line 315, head labeled $Lt_2_68866
	.loc	19	352	0
	add.s32 	%r170, %r203, 1;
	st.shared.s32 	[sBluePointCount__13], %r170;
$L_2_55042:
 //<loop> Part of loop body line 315, head labeled $Lt_2_68866
	add.s32 	%r204, %r203, 1;
	mov.s32 	%r203, %r204;
	add.u64 	%rd60, %rd60, 4;
	add.u64 	%rd61, %rd61, 4;
	mov.u32 	%r205, 16;
	setp.ne.s32 	%p36, %r204, %r205;
	@%p36 bra 	$Lt_2_68866;
	.loc	19	355	0
	mov.f32 	%f508, 0f00000000;   	// 0
	st.shared.f32 	[u__0+0], %f508;
	.loc	18	48	0
	mov.f32 	%f509, 0f00000000;   	// 0
	st.shared.f32 	[sBluePointsY2__16+0], %f509;
	cvt.u64.s32 	%rd64, %r170;
	mov.s64 	%rd65, %rd64;
	cvt.u32.u64 	%r206, %rd65;
	sub.s32 	%r207, %r206, 1;
	sub.s32 	%r208, %r206, 2;
	mov.u32 	%r209, 1;
	setp.le.s32 	%p37, %r207, %r209;
	@%p37 bra 	$Lt_2_69378;
	mov.s32 	%r210, %r208;
	mov.u64 	%rd16, 4;
	mov.s32 	%r71, 1;
	mov.u64 	%rd66, sBluePointsY2__16;
	mov.s32 	%r211, %r210;
$Lt_2_69890:
 //<loop> Loop body line 48, nesting depth: 1, estimated iterations: unknown
	.loc	18	53	0
	add.u64 	%rd19, %rd16, %rd18;
	mov.f32 	%f510, 0f00000000;   	// 0
	st.shared.f32 	[%rd19+0], %f510;
	add.u64 	%rd67, %rd16, %rd66;
	mov.f32 	%f511, 0f00000000;   	// 0
	st.shared.f32 	[%rd67+0], %f511;
	add.u64 	%rd68, %rd16, %rd59;
	ld.shared.f32 	%f512, [%rd68+-4];
	ld.shared.f32 	%f513, [%rd68+0];
	setp.lt.f32 	%p38, %f512, %f513;
	@!%p38 bra 	$Lt_2_96258;
 //<loop> Part of loop body line 48, head labeled $Lt_2_69890
	ld.shared.f32 	%f514, [%rd68+4];
	ld.shared.f32 	%f513, [%rd68+0];
	setp.lt.f32 	%p39, %f513, %f514;
	@!%p39 bra 	$Lt_2_96258;
 //<loop> Part of loop body line 48, head labeled $Lt_2_69890
	ld.shared.f32 	%f513, [%rd68+0];
	ld.shared.f32 	%f512, [%rd68+-4];
	.loc	18	56	0
	sub.f32 	%f515, %f513, %f512;
	.loc	18	53	0
	ld.shared.f32 	%f514, [%rd68+4];
	.loc	18	56	0
	sub.f32 	%f516, %f514, %f512;
	.loc	18	57	0
	div.full.f32 	%f119, %f515, %f516;
	mov.f32 	%f517, 0f40000000;   	// 2
	ld.shared.f32 	%f518, [%rd67+-4];
	mad.f32 	%f122, %f518, %f119, %f517;
	mov.f32 	%f519, 0f00000000;   	// 0
	setp.neu.f32 	%p40, %f122, %f519;
	@!%p40 bra 	$Lt_2_96258;
 //<loop> Part of loop body line 48, head labeled $Lt_2_69890
	.loc	18	60	0
	mov.f32 	%f520, 0fbf800000;   	// -1
	add.f32 	%f521, %f119, %f520;
	div.full.f32 	%f522, %f521, %f122;
	st.shared.f32 	[%rd67+0], %f522;
	.loc	18	61	0
	add.u64 	%rd69, %rd16, %rd62;
	ld.shared.f32 	%f523, [%rd69+0];
	ld.shared.f32 	%f524, [%rd69+4];
	sub.f32 	%f129, %f524, %f523;
	sub.f32 	%f130, %f514, %f513;
	ld.shared.f32 	%f525, [%rd69+-4];
	sub.f32 	%f132, %f523, %f525;
	div.full.f32 	%f133, %f129, %f130;
	div.full.f32 	%f134, %f132, %f515;
	sub.f32 	%f135, %f133, %f134;
	st.shared.f32 	[%rd19+0], %f135;
	.loc	18	62	0
	mov.f32 	%f526, 0f40c00000;   	// 6
	mul.f32 	%f527, %f135, %f526;
	div.full.f32 	%f528, %f527, %f516;
	ld.shared.f32 	%f529, [%rd19+-4];
	mul.f32 	%f530, %f529, %f119;
	sub.f32 	%f531, %f528, %f530;
	div.full.f32 	%f532, %f531, %f122;
	st.shared.f32 	[%rd19+0], %f532;
$Lt_2_96258:
$L_2_55554:
 //<loop> Part of loop body line 48, head labeled $Lt_2_69890
	.loc	18	56	0
	add.s32 	%r71, %r71, 1;
	add.u64 	%rd16, %rd16, 4;
	setp.ne.s32 	%p41, %r207, %r71;
	@%p41 bra 	$Lt_2_69890;
$Lt_2_69378:
	mov.u64 	%rd66, sBluePointsY2__16;
	.loc	18	69	0
	mov.f32 	%f533, 0f00000000;   	// 0
	mul.lo.u64 	%rd70, %rd64, 4;
	add.u64 	%rd71, %rd66, %rd70;
	st.shared.f32 	[%rd71+-4], %f533;
	.loc	18	72	0
	mov.s32 	%r73, %r208;
	mov.u32 	%r212, 0;
	setp.lt.s32 	%p42, %r208, %r212;
	@%p42 bra 	$Lt_2_70914;
	mov.s32 	%r213, %r207;
	cvt.u64.s32 	%rd72, %r208;
	mul.lo.u64 	%rd26, %rd72, 4;
	add.u64 	%rd27, %rd26, %rd18;
	add.u64 	%rd73, %rd26, %rd66;
	mov.s32 	%r214, %r213;
$Lt_2_71426:
 //<loop> Loop body line 72, nesting depth: 1, estimated iterations: unknown
	.loc	18	74	0
	ld.shared.f32 	%f534, [%rd27+0];
	ld.shared.f32 	%f535, [%rd73+0];
	ld.shared.f32 	%f536, [%rd73+4];
	mad.f32 	%f537, %f535, %f536, %f534;
	st.shared.f32 	[%rd73+0], %f537;
	sub.s32 	%r73, %r73, 1;
	sub.u64 	%rd73, %rd73, 4;
	sub.u64 	%rd27, %rd27, 4;
	mov.u32 	%r215, -1;
	setp.ne.s32 	%p43, %r73, %r215;
	@%p43 bra 	$Lt_2_71426;
$Lt_2_70914:
$Lt_2_57090:
	mov.u64 	%rd9, sMasterPointsX__2;
	mov.u64 	%rd17, sMasterPointsY2__4;
	mov.u64 	%rd29, sRedPointsX__6;
	mov.u64 	%rd36, sRedPointsY2__8;
	mov.u64 	%rd44, sGreenPointsX__10;
	mov.u64 	%rd51, sGreenPointsY2__12;
	mov.u64 	%rd59, sBluePointsX__14;
	mov.u64 	%rd66, sBluePointsY2__16;
	mov.u64 	%rd12, sMasterPointsY__3;
	mov.u64 	%rd32, sRedPointsY__7;
	mov.u64 	%rd47, sGreenPointsY__11;
	mov.u64 	%rd62, sBluePointsY__15;
	.loc	19	358	0
	bar.sync 	0;
	.loc	19	360	0
	mov.f32 	%f538, 0f3a83126f;   	// 0.001
	setp.gt.f32 	%p44, %f6, %f538;
	@!%p44 bra 	$Lt_2_71938;
	.loc	19	365	0
	mov.f32 	%f539, 0f00000000;   	// 0
	setp.lt.f32 	%p45, %f3, %f539;
	@!%p45 bra 	$Lt_2_72706;
	.loc	18	134	0
	ld.shared.f32 	%f540, [sRedPointsY__7+0];
	add.f32 	%f541, %f540, %f3;
	bra.uni 	$Lt_2_72962;
$Lt_2_72706:
	ld.shared.s32 	%r216, [sRedPointCount__5];
	cvt.u64.s32 	%rd34, %r216;
	mov.f32 	%f542, 0f3f800000;   	// 1
	setp.gt.f32 	%p46, %f3, %f542;
	@!%p46 bra 	$Lt_2_73218;
	.loc	18	138	0
	mov.f32 	%f543, 0f3f800000;   	// 1
	mul.lo.u64 	%rd74, %rd34, 4;
	add.u64 	%rd75, %rd32, %rd74;
	ld.shared.f32 	%f544, [%rd75+-4];
	sub.f32 	%f545, %f543, %f544;
	sub.f32 	%f541, %f3, %f545;
	bra.uni 	$Lt_2_72962;
$Lt_2_73218:
	.loc	18	88	0
	mov.f32 	%f546, %f3;
	.loc	18	92	0
	sub.u64 	%rd76, %rd34, 1;
	mov.s64 	%rd77, %rd76;
	mov.u64 	%rd78, 1;
	setp.le.u64 	%p47, %rd76, %rd78;
	@%p47 bra 	$Lt_2_97026;
	mov.u64 	%rd79, 0;
$Lt_2_73986:
 //<loop> Loop body line 92
	add.u64 	%rd80, %rd77, %rd79;
	shr.u64 	%rd81, %rd80, 1;
	mul.lo.u64 	%rd82, %rd81, 4;
	add.u64 	%rd83, %rd29, %rd82;
	ld.shared.f32 	%f547, [%rd83+0];
	setp.gt.f32 	%p48, %f547, %f3;
	@!%p48 bra 	$Lt_2_74498;
 //<loop> Part of loop body line 92, head labeled $Lt_2_73986
	.loc	18	101	0
	mov.s64 	%rd77, %rd81;
	bra.uni 	$Lt_2_74242;
$Lt_2_74498:
 //<loop> Part of loop body line 92, head labeled $Lt_2_73986
	.loc	18	105	0
	mov.s64 	%rd79, %rd81;
$Lt_2_74242:
 //<loop> Part of loop body line 92, head labeled $Lt_2_73986
	sub.u64 	%rd84, %rd77, %rd79;
	mov.u64 	%rd85, 1;
	setp.gt.u64 	%p49, %rd84, %rd85;
	@%p49 bra 	$Lt_2_73986;
	bra.uni 	$Lt_2_73474;
$Lt_2_97026:
	mov.u64 	%rd79, 0;
$Lt_2_73474:
	.loc	18	109	0
	mul.lo.u64 	%rd86, %rd77, 4;
	mul.lo.u64 	%rd87, %rd79, 4;
	add.u64 	%rd88, %rd86, %rd29;
	add.u64 	%rd89, %rd87, %rd29;
	ld.shared.f32 	%f548, [%rd88+0];
	ld.shared.f32 	%f549, [%rd89+0];
	sub.f32 	%f550, %f548, %f549;
	mov.f32 	%f551, 0f00000000;   	// 0
	setp.neu.f32 	%p50, %f550, %f551;
	@!%p50 bra 	$Lt_2_75010;
	ld.shared.f32 	%f548, [%rd88+0];
	.loc	18	112	0
	sub.f32 	%f552, %f548, %f3;
	.loc	18	109	0
	ld.shared.f32 	%f549, [%rd89+0];
	.loc	18	113	0
	sub.f32 	%f553, %f3, %f549;
	.loc	18	114	0
	div.full.f32 	%f554, %f553, %f550;
	div.full.f32 	%f555, %f552, %f550;
	add.u64 	%rd90, %rd86, %rd36;
	ld.shared.f32 	%f556, [%rd90+0];
	mul.f32 	%f557, %f554, %f554;
	mul.f32 	%f558, %f554, %f557;
	sub.f32 	%f559, %f558, %f554;
	mul.f32 	%f560, %f556, %f559;
	add.u64 	%rd91, %rd87, %rd36;
	ld.shared.f32 	%f561, [%rd91+0];
	mul.f32 	%f562, %f555, %f555;
	mul.f32 	%f563, %f555, %f562;
	sub.f32 	%f564, %f563, %f555;
	mad.f32 	%f565, %f561, %f564, %f560;
	mul.f32 	%f566, %f550, %f550;
	mul.f32 	%f567, %f565, %f566;
	mov.f32 	%f568, 0f40c00000;   	// 6
	div.full.f32 	%f569, %f567, %f568;
	add.u64 	%rd92, %rd86, %rd32;
	ld.shared.f32 	%f570, [%rd92+0];
	mul.f32 	%f571, %f570, %f554;
	add.u64 	%rd93, %rd87, %rd32;
	ld.shared.f32 	%f572, [%rd93+0];
	mad.f32 	%f573, %f572, %f555, %f571;
	add.f32 	%f546, %f569, %f573;
$Lt_2_75010:
	.loc	18	142	0
	mov.f32 	%f541, %f546;
$Lt_2_72962:
$Lt_2_72450:
	.loc	19	366	0
	mov.f32 	%f574, 0f00000000;   	// 0
	setp.lt.f32 	%p51, %f541, %f574;
	@!%p51 bra 	$Lt_2_75778;
	.loc	18	134	0
	ld.shared.f32 	%f575, [sMasterPointsY__3+0];
	add.f32 	%f541, %f575, %f541;
	bra.uni 	$Lt_2_76034;
$Lt_2_75778:
	ld.shared.s32 	%r217, [sMasterPointCount__1];
	cvt.u64.s32 	%rd94, %r217;
	mov.f32 	%f576, 0f3f800000;   	// 1
	setp.gt.f32 	%p52, %f541, %f576;
	@!%p52 bra 	$Lt_2_76290;
	.loc	18	138	0
	mov.f32 	%f577, 0f3f800000;   	// 1
	mul.lo.u64 	%rd95, %rd94, 4;
	add.u64 	%rd96, %rd12, %rd95;
	ld.shared.f32 	%f578, [%rd96+-4];
	sub.f32 	%f579, %f577, %f578;
	sub.f32 	%f541, %f541, %f579;
	bra.uni 	$Lt_2_76034;
$Lt_2_76290:
	.loc	18	88	0
	mov.f32 	%f546, %f541;
	.loc	18	92	0
	sub.u64 	%rd97, %rd94, 1;
	mov.s64 	%rd77, %rd97;
	mov.u64 	%rd98, 1;
	setp.le.u64 	%p53, %rd97, %rd98;
	@%p53 bra 	$Lt_2_97282;
	mov.u64 	%rd79, 0;
$Lt_2_77058:
 //<loop> Loop body line 92
	add.u64 	%rd99, %rd77, %rd79;
	shr.u64 	%rd81, %rd99, 1;
	mul.lo.u64 	%rd100, %rd81, 4;
	add.u64 	%rd101, %rd9, %rd100;
	ld.shared.f32 	%f580, [%rd101+0];
	setp.gt.f32 	%p54, %f580, %f541;
	@!%p54 bra 	$Lt_2_77570;
 //<loop> Part of loop body line 92, head labeled $Lt_2_77058
	.loc	18	101	0
	mov.s64 	%rd77, %rd81;
	bra.uni 	$Lt_2_77314;
$Lt_2_77570:
 //<loop> Part of loop body line 92, head labeled $Lt_2_77058
	.loc	18	105	0
	mov.s64 	%rd79, %rd81;
$Lt_2_77314:
 //<loop> Part of loop body line 92, head labeled $Lt_2_77058
	sub.u64 	%rd102, %rd77, %rd79;
	mov.u64 	%rd103, 1;
	setp.gt.u64 	%p55, %rd102, %rd103;
	@%p55 bra 	$Lt_2_77058;
	bra.uni 	$Lt_2_76546;
$Lt_2_97282:
	mov.u64 	%rd79, 0;
$Lt_2_76546:
	.loc	18	109	0
	mul.lo.u64 	%rd86, %rd77, 4;
	mul.lo.u64 	%rd87, %rd79, 4;
	add.u64 	%rd104, %rd86, %rd9;
	add.u64 	%rd105, %rd87, %rd9;
	ld.shared.f32 	%f581, [%rd104+0];
	ld.shared.f32 	%f582, [%rd105+0];
	sub.f32 	%f550, %f581, %f582;
	mov.f32 	%f583, 0f00000000;   	// 0
	setp.neu.f32 	%p56, %f550, %f583;
	@!%p56 bra 	$Lt_2_78082;
	.loc	18	112	0
	sub.f32 	%f552, %f581, %f541;
	.loc	18	113	0
	sub.f32 	%f553, %f541, %f582;
	.loc	18	114	0
	div.full.f32 	%f554, %f553, %f550;
	div.full.f32 	%f555, %f552, %f550;
	add.u64 	%rd106, %rd86, %rd17;
	ld.shared.f32 	%f584, [%rd106+0];
	mul.f32 	%f585, %f554, %f554;
	mul.f32 	%f586, %f554, %f585;
	sub.f32 	%f587, %f586, %f554;
	mul.f32 	%f588, %f584, %f587;
	add.u64 	%rd107, %rd87, %rd17;
	ld.shared.f32 	%f589, [%rd107+0];
	mul.f32 	%f590, %f555, %f555;
	mul.f32 	%f591, %f555, %f590;
	sub.f32 	%f592, %f591, %f555;
	mad.f32 	%f593, %f589, %f592, %f588;
	mul.f32 	%f594, %f550, %f550;
	mul.f32 	%f595, %f593, %f594;
	mov.f32 	%f596, 0f40c00000;   	// 6
	div.full.f32 	%f597, %f595, %f596;
	add.u64 	%rd108, %rd86, %rd12;
	ld.shared.f32 	%f598, [%rd108+0];
	mul.f32 	%f599, %f598, %f554;
	add.u64 	%rd109, %rd87, %rd12;
	ld.shared.f32 	%f600, [%rd109+0];
	mad.f32 	%f601, %f600, %f555, %f599;
	add.f32 	%f546, %f597, %f601;
$Lt_2_78082:
	.loc	18	142	0
	mov.f32 	%f541, %f546;
$Lt_2_76034:
$Lt_2_75522:
	.loc	19	367	0
	sub.f32 	%f602, %f541, %f3;
	mad.f32 	%f9, %f6, %f602, %f3;
	.loc	19	370	0
	mov.f32 	%f603, 0f00000000;   	// 0
	setp.lt.f32 	%p57, %f2, %f603;
	@!%p57 bra 	$Lt_2_78850;
	.loc	18	134	0
	ld.shared.f32 	%f604, [sGreenPointsY__11+0];
	add.f32 	%f541, %f604, %f2;
	bra.uni 	$Lt_2_79106;
$Lt_2_78850:
	ld.shared.s32 	%r218, [sGreenPointCount__9];
	cvt.u64.s32 	%rd49, %r218;
	mov.f32 	%f605, 0f3f800000;   	// 1
	setp.gt.f32 	%p58, %f2, %f605;
	@!%p58 bra 	$Lt_2_79362;
	.loc	18	138	0
	mov.f32 	%f606, 0f3f800000;   	// 1
	mul.lo.u64 	%rd110, %rd49, 4;
	add.u64 	%rd111, %rd47, %rd110;
	ld.shared.f32 	%f607, [%rd111+-4];
	sub.f32 	%f608, %f606, %f607;
	sub.f32 	%f541, %f2, %f608;
	bra.uni 	$Lt_2_79106;
$Lt_2_79362:
	.loc	18	88	0
	mov.f32 	%f546, %f2;
	.loc	18	92	0
	sub.u64 	%rd112, %rd49, 1;
	mov.s64 	%rd77, %rd112;
	mov.u64 	%rd113, 1;
	setp.le.u64 	%p59, %rd112, %rd113;
	@%p59 bra 	$Lt_2_97538;
	mov.u64 	%rd79, 0;
$Lt_2_80130:
 //<loop> Loop body line 92
	add.u64 	%rd114, %rd77, %rd79;
	shr.u64 	%rd81, %rd114, 1;
	mul.lo.u64 	%rd115, %rd81, 4;
	add.u64 	%rd116, %rd44, %rd115;
	ld.shared.f32 	%f609, [%rd116+0];
	setp.gt.f32 	%p60, %f609, %f2;
	@!%p60 bra 	$Lt_2_80642;
 //<loop> Part of loop body line 92, head labeled $Lt_2_80130
	.loc	18	101	0
	mov.s64 	%rd77, %rd81;
	bra.uni 	$Lt_2_80386;
$Lt_2_80642:
 //<loop> Part of loop body line 92, head labeled $Lt_2_80130
	.loc	18	105	0
	mov.s64 	%rd79, %rd81;
$Lt_2_80386:
 //<loop> Part of loop body line 92, head labeled $Lt_2_80130
	sub.u64 	%rd117, %rd77, %rd79;
	mov.u64 	%rd118, 1;
	setp.gt.u64 	%p61, %rd117, %rd118;
	@%p61 bra 	$Lt_2_80130;
	bra.uni 	$Lt_2_79618;
$Lt_2_97538:
	mov.u64 	%rd79, 0;
$Lt_2_79618:
	.loc	18	109	0
	mul.lo.u64 	%rd86, %rd77, 4;
	mul.lo.u64 	%rd87, %rd79, 4;
	add.u64 	%rd119, %rd86, %rd44;
	add.u64 	%rd120, %rd87, %rd44;
	ld.shared.f32 	%f610, [%rd119+0];
	ld.shared.f32 	%f611, [%rd120+0];
	sub.f32 	%f550, %f610, %f611;
	mov.f32 	%f612, 0f00000000;   	// 0
	setp.neu.f32 	%p62, %f550, %f612;
	@!%p62 bra 	$Lt_2_81154;
	ld.shared.f32 	%f610, [%rd119+0];
	.loc	18	112	0
	sub.f32 	%f552, %f610, %f2;
	.loc	18	109	0
	ld.shared.f32 	%f611, [%rd120+0];
	.loc	18	113	0
	sub.f32 	%f553, %f2, %f611;
	.loc	18	114	0
	div.full.f32 	%f554, %f553, %f550;
	div.full.f32 	%f555, %f552, %f550;
	add.u64 	%rd121, %rd86, %rd51;
	ld.shared.f32 	%f613, [%rd121+0];
	mul.f32 	%f614, %f554, %f554;
	mul.f32 	%f615, %f554, %f614;
	sub.f32 	%f616, %f615, %f554;
	mul.f32 	%f617, %f613, %f616;
	add.u64 	%rd122, %rd87, %rd51;
	ld.shared.f32 	%f618, [%rd122+0];
	mul.f32 	%f619, %f555, %f555;
	mul.f32 	%f620, %f555, %f619;
	sub.f32 	%f621, %f620, %f555;
	mad.f32 	%f622, %f618, %f621, %f617;
	mul.f32 	%f623, %f550, %f550;
	mul.f32 	%f624, %f622, %f623;
	mov.f32 	%f625, 0f40c00000;   	// 6
	div.full.f32 	%f626, %f624, %f625;
	add.u64 	%rd123, %rd86, %rd47;
	ld.shared.f32 	%f627, [%rd123+0];
	mul.f32 	%f628, %f627, %f554;
	add.u64 	%rd124, %rd87, %rd47;
	ld.shared.f32 	%f629, [%rd124+0];
	mad.f32 	%f630, %f629, %f555, %f628;
	add.f32 	%f546, %f626, %f630;
$Lt_2_81154:
	.loc	18	142	0
	mov.f32 	%f541, %f546;
$Lt_2_79106:
$Lt_2_78594:
	.loc	19	371	0
	mov.f32 	%f631, 0f00000000;   	// 0
	setp.lt.f32 	%p63, %f541, %f631;
	@!%p63 bra 	$Lt_2_81922;
	.loc	18	134	0
	ld.shared.f32 	%f632, [sMasterPointsY__3+0];
	add.f32 	%f541, %f632, %f541;
	bra.uni 	$Lt_2_82178;
$Lt_2_81922:
	ld.shared.s32 	%r219, [sMasterPointCount__1];
	cvt.u64.s32 	%rd94, %r219;
	mov.f32 	%f633, 0f3f800000;   	// 1
	setp.gt.f32 	%p64, %f541, %f633;
	@!%p64 bra 	$Lt_2_82434;
	.loc	18	138	0
	mov.f32 	%f634, 0f3f800000;   	// 1
	mul.lo.u64 	%rd125, %rd94, 4;
	add.u64 	%rd126, %rd12, %rd125;
	ld.shared.f32 	%f635, [%rd126+-4];
	sub.f32 	%f636, %f634, %f635;
	sub.f32 	%f541, %f541, %f636;
	bra.uni 	$Lt_2_82178;
$Lt_2_82434:
	.loc	18	88	0
	mov.f32 	%f546, %f541;
	.loc	18	92	0
	sub.u64 	%rd97, %rd94, 1;
	mov.s64 	%rd77, %rd97;
	mov.u64 	%rd127, 1;
	setp.le.u64 	%p65, %rd97, %rd127;
	@%p65 bra 	$Lt_2_97794;
	mov.u64 	%rd79, 0;
$Lt_2_83202:
 //<loop> Loop body line 92
	add.u64 	%rd128, %rd77, %rd79;
	shr.u64 	%rd81, %rd128, 1;
	mul.lo.u64 	%rd129, %rd81, 4;
	add.u64 	%rd130, %rd9, %rd129;
	ld.shared.f32 	%f637, [%rd130+0];
	setp.gt.f32 	%p66, %f637, %f541;
	@!%p66 bra 	$Lt_2_83714;
 //<loop> Part of loop body line 92, head labeled $Lt_2_83202
	.loc	18	101	0
	mov.s64 	%rd77, %rd81;
	bra.uni 	$Lt_2_83458;
$Lt_2_83714:
 //<loop> Part of loop body line 92, head labeled $Lt_2_83202
	.loc	18	105	0
	mov.s64 	%rd79, %rd81;
$Lt_2_83458:
 //<loop> Part of loop body line 92, head labeled $Lt_2_83202
	sub.u64 	%rd131, %rd77, %rd79;
	mov.u64 	%rd132, 1;
	setp.gt.u64 	%p67, %rd131, %rd132;
	@%p67 bra 	$Lt_2_83202;
	bra.uni 	$Lt_2_82690;
$Lt_2_97794:
	mov.u64 	%rd79, 0;
$Lt_2_82690:
	.loc	18	109	0
	mul.lo.u64 	%rd86, %rd77, 4;
	mul.lo.u64 	%rd87, %rd79, 4;
	add.u64 	%rd104, %rd86, %rd9;
	add.u64 	%rd105, %rd87, %rd9;
	ld.shared.f32 	%f581, [%rd104+0];
	ld.shared.f32 	%f582, [%rd105+0];
	sub.f32 	%f550, %f581, %f582;
	mov.f32 	%f638, 0f00000000;   	// 0
	setp.neu.f32 	%p68, %f550, %f638;
	@!%p68 bra 	$Lt_2_84226;
	.loc	18	112	0
	sub.f32 	%f552, %f581, %f541;
	.loc	18	113	0
	sub.f32 	%f553, %f541, %f582;
	.loc	18	114	0
	div.full.f32 	%f554, %f553, %f550;
	div.full.f32 	%f555, %f552, %f550;
	add.u64 	%rd133, %rd86, %rd17;
	ld.shared.f32 	%f639, [%rd133+0];
	mul.f32 	%f640, %f554, %f554;
	mul.f32 	%f641, %f554, %f640;
	sub.f32 	%f642, %f641, %f554;
	mul.f32 	%f643, %f639, %f642;
	add.u64 	%rd134, %rd87, %rd17;
	ld.shared.f32 	%f644, [%rd134+0];
	mul.f32 	%f645, %f555, %f555;
	mul.f32 	%f646, %f555, %f645;
	sub.f32 	%f647, %f646, %f555;
	mad.f32 	%f648, %f644, %f647, %f643;
	mul.f32 	%f649, %f550, %f550;
	mul.f32 	%f650, %f648, %f649;
	mov.f32 	%f651, 0f40c00000;   	// 6
	div.full.f32 	%f652, %f650, %f651;
	add.u64 	%rd135, %rd86, %rd12;
	ld.shared.f32 	%f653, [%rd135+0];
	mul.f32 	%f654, %f653, %f554;
	add.u64 	%rd136, %rd87, %rd12;
	ld.shared.f32 	%f655, [%rd136+0];
	mad.f32 	%f656, %f655, %f555, %f654;
	add.f32 	%f546, %f652, %f656;
$Lt_2_84226:
	.loc	18	142	0
	mov.f32 	%f541, %f546;
$Lt_2_82178:
$Lt_2_81666:
	.loc	19	372	0
	sub.f32 	%f657, %f541, %f2;
	mad.f32 	%f8, %f6, %f657, %f2;
	.loc	19	375	0
	mov.f32 	%f658, 0f00000000;   	// 0
	setp.lt.f32 	%p69, %f1, %f658;
	@!%p69 bra 	$Lt_2_84994;
	.loc	18	134	0
	ld.shared.f32 	%f659, [sBluePointsY__15+0];
	add.f32 	%f541, %f659, %f1;
	bra.uni 	$Lt_2_85250;
$Lt_2_84994:
	ld.shared.s32 	%r220, [sBluePointCount__13];
	cvt.u64.s32 	%rd64, %r220;
	mov.f32 	%f660, 0f3f800000;   	// 1
	setp.gt.f32 	%p70, %f1, %f660;
	@!%p70 bra 	$Lt_2_85506;
	.loc	18	138	0
	mov.f32 	%f661, 0f3f800000;   	// 1
	mul.lo.u64 	%rd137, %rd64, 4;
	add.u64 	%rd138, %rd62, %rd137;
	ld.shared.f32 	%f662, [%rd138+-4];
	sub.f32 	%f663, %f661, %f662;
	sub.f32 	%f541, %f1, %f663;
	bra.uni 	$Lt_2_85250;
$Lt_2_85506:
	.loc	18	88	0
	mov.f32 	%f546, %f1;
	.loc	18	92	0
	sub.u64 	%rd139, %rd64, 1;
	mov.s64 	%rd77, %rd139;
	mov.u64 	%rd140, 1;
	setp.le.u64 	%p71, %rd139, %rd140;
	@%p71 bra 	$Lt_2_98050;
	mov.u64 	%rd79, 0;
$Lt_2_86274:
 //<loop> Loop body line 92
	add.u64 	%rd141, %rd77, %rd79;
	shr.u64 	%rd81, %rd141, 1;
	mul.lo.u64 	%rd142, %rd81, 4;
	add.u64 	%rd143, %rd59, %rd142;
	ld.shared.f32 	%f664, [%rd143+0];
	setp.gt.f32 	%p72, %f664, %f1;
	@!%p72 bra 	$Lt_2_86786;
 //<loop> Part of loop body line 92, head labeled $Lt_2_86274
	.loc	18	101	0
	mov.s64 	%rd77, %rd81;
	bra.uni 	$Lt_2_86530;
$Lt_2_86786:
 //<loop> Part of loop body line 92, head labeled $Lt_2_86274
	.loc	18	105	0
	mov.s64 	%rd79, %rd81;
$Lt_2_86530:
 //<loop> Part of loop body line 92, head labeled $Lt_2_86274
	sub.u64 	%rd144, %rd77, %rd79;
	mov.u64 	%rd145, 1;
	setp.gt.u64 	%p73, %rd144, %rd145;
	@%p73 bra 	$Lt_2_86274;
	bra.uni 	$Lt_2_85762;
$Lt_2_98050:
	mov.u64 	%rd79, 0;
$Lt_2_85762:
	.loc	18	109	0
	mul.lo.u64 	%rd86, %rd77, 4;
	mul.lo.u64 	%rd87, %rd79, 4;
	add.u64 	%rd146, %rd86, %rd59;
	add.u64 	%rd147, %rd87, %rd59;
	ld.shared.f32 	%f665, [%rd146+0];
	ld.shared.f32 	%f666, [%rd147+0];
	sub.f32 	%f550, %f665, %f666;
	mov.f32 	%f667, 0f00000000;   	// 0
	setp.neu.f32 	%p74, %f550, %f667;
	@!%p74 bra 	$Lt_2_87298;
	ld.shared.f32 	%f665, [%rd146+0];
	.loc	18	112	0
	sub.f32 	%f552, %f665, %f1;
	.loc	18	109	0
	ld.shared.f32 	%f666, [%rd147+0];
	.loc	18	113	0
	sub.f32 	%f553, %f1, %f666;
	.loc	18	114	0
	div.full.f32 	%f554, %f553, %f550;
	div.full.f32 	%f555, %f552, %f550;
	add.u64 	%rd148, %rd86, %rd66;
	ld.shared.f32 	%f668, [%rd148+0];
	mul.f32 	%f669, %f554, %f554;
	mul.f32 	%f670, %f554, %f669;
	sub.f32 	%f671, %f670, %f554;
	mul.f32 	%f672, %f668, %f671;
	add.u64 	%rd149, %rd87, %rd66;
	ld.shared.f32 	%f673, [%rd149+0];
	mul.f32 	%f674, %f555, %f555;
	mul.f32 	%f675, %f555, %f674;
	sub.f32 	%f676, %f675, %f555;
	mad.f32 	%f677, %f673, %f676, %f672;
	mul.f32 	%f678, %f550, %f550;
	mul.f32 	%f679, %f677, %f678;
	mov.f32 	%f680, 0f40c00000;   	// 6
	div.full.f32 	%f681, %f679, %f680;
	add.u64 	%rd150, %rd86, %rd62;
	ld.shared.f32 	%f682, [%rd150+0];
	mul.f32 	%f683, %f682, %f554;
	add.u64 	%rd151, %rd87, %rd62;
	ld.shared.f32 	%f684, [%rd151+0];
	mad.f32 	%f685, %f684, %f555, %f683;
	add.f32 	%f546, %f681, %f685;
$Lt_2_87298:
	.loc	18	142	0
	mov.f32 	%f541, %f546;
$Lt_2_85250:
$Lt_2_84738:
	.loc	19	376	0
	mov.f32 	%f686, 0f00000000;   	// 0
	setp.lt.f32 	%p75, %f541, %f686;
	@!%p75 bra 	$Lt_2_88066;
	.loc	18	134	0
	ld.shared.f32 	%f687, [sMasterPointsY__3+0];
	add.f32 	%f541, %f687, %f541;
	bra.uni 	$Lt_2_88322;
$Lt_2_88066:
	ld.shared.s32 	%r221, [sMasterPointCount__1];
	cvt.u64.s32 	%rd94, %r221;
	mov.f32 	%f688, 0f3f800000;   	// 1
	setp.gt.f32 	%p76, %f541, %f688;
	@!%p76 bra 	$Lt_2_88578;
	.loc	18	138	0
	mov.f32 	%f689, 0f3f800000;   	// 1
	mul.lo.u64 	%rd152, %rd94, 4;
	add.u64 	%rd153, %rd12, %rd152;
	ld.shared.f32 	%f690, [%rd153+-4];
	sub.f32 	%f691, %f689, %f690;
	sub.f32 	%f541, %f541, %f691;
	bra.uni 	$Lt_2_88322;
$Lt_2_88578:
	.loc	18	88	0
	mov.f32 	%f546, %f541;
	.loc	18	92	0
	sub.u64 	%rd97, %rd94, 1;
	mov.s64 	%rd77, %rd97;
	mov.u64 	%rd154, 1;
	setp.le.u64 	%p77, %rd97, %rd154;
	@%p77 bra 	$Lt_2_98306;
	mov.u64 	%rd79, 0;
$Lt_2_89346:
 //<loop> Loop body line 92
	add.u64 	%rd155, %rd77, %rd79;
	shr.u64 	%rd81, %rd155, 1;
	mul.lo.u64 	%rd156, %rd81, 4;
	add.u64 	%rd157, %rd9, %rd156;
	ld.shared.f32 	%f692, [%rd157+0];
	setp.gt.f32 	%p78, %f692, %f541;
	@!%p78 bra 	$Lt_2_89858;
 //<loop> Part of loop body line 92, head labeled $Lt_2_89346
	.loc	18	101	0
	mov.s64 	%rd77, %rd81;
	bra.uni 	$Lt_2_89602;
$Lt_2_89858:
 //<loop> Part of loop body line 92, head labeled $Lt_2_89346
	.loc	18	105	0
	mov.s64 	%rd79, %rd81;
$Lt_2_89602:
 //<loop> Part of loop body line 92, head labeled $Lt_2_89346
	sub.u64 	%rd158, %rd77, %rd79;
	mov.u64 	%rd159, 1;
	setp.gt.u64 	%p79, %rd158, %rd159;
	@%p79 bra 	$Lt_2_89346;
	bra.uni 	$Lt_2_88834;
$Lt_2_98306:
	mov.u64 	%rd79, 0;
$Lt_2_88834:
	.loc	18	109	0
	mul.lo.u64 	%rd86, %rd77, 4;
	mul.lo.u64 	%rd87, %rd79, 4;
	add.u64 	%rd104, %rd86, %rd9;
	add.u64 	%rd105, %rd87, %rd9;
	ld.shared.f32 	%f581, [%rd104+0];
	ld.shared.f32 	%f582, [%rd105+0];
	sub.f32 	%f550, %f581, %f582;
	mov.f32 	%f693, 0f00000000;   	// 0
	setp.neu.f32 	%p80, %f550, %f693;
	@!%p80 bra 	$Lt_2_90370;
	.loc	18	112	0
	sub.f32 	%f552, %f581, %f541;
	.loc	18	113	0
	sub.f32 	%f553, %f541, %f582;
	.loc	18	114	0
	div.full.f32 	%f554, %f553, %f550;
	div.full.f32 	%f555, %f552, %f550;
	add.u64 	%rd160, %rd86, %rd17;
	ld.shared.f32 	%f694, [%rd160+0];
	mul.f32 	%f695, %f554, %f554;
	mul.f32 	%f696, %f554, %f695;
	sub.f32 	%f697, %f696, %f554;
	mul.f32 	%f698, %f694, %f697;
	add.u64 	%rd161, %rd87, %rd17;
	ld.shared.f32 	%f699, [%rd161+0];
	mul.f32 	%f700, %f555, %f555;
	mul.f32 	%f701, %f555, %f700;
	sub.f32 	%f702, %f701, %f555;
	mad.f32 	%f703, %f699, %f702, %f698;
	mul.f32 	%f704, %f550, %f550;
	mul.f32 	%f705, %f703, %f704;
	mov.f32 	%f706, 0f40c00000;   	// 6
	div.full.f32 	%f707, %f705, %f706;
	add.u64 	%rd162, %rd86, %rd12;
	ld.shared.f32 	%f708, [%rd162+0];
	mul.f32 	%f709, %f708, %f554;
	add.u64 	%rd163, %rd87, %rd12;
	ld.shared.f32 	%f710, [%rd163+0];
	mad.f32 	%f711, %f710, %f555, %f709;
	add.f32 	%f546, %f707, %f711;
$Lt_2_90370:
	.loc	18	142	0
	mov.f32 	%f541, %f546;
$Lt_2_88322:
$Lt_2_87810:
	.loc	19	377	0
	sub.f32 	%f712, %f541, %f1;
	mad.f32 	%f7, %f6, %f712, %f1;
$Lt_2_71938:
	st.global.v4.f32 	[%rd4+0], {%f7,%f8,%f9,%f4};
$Lt_2_56066:
	.loc	19	975	0
	exit;
$LDWend_RGBCurves_CompositeKernel:
	} // RGBCurves_CompositeKernel

