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

	.visible .func (.param .f32 __cudaretf__Z22CalculateShadowsWeightfff) _Z22CalculateShadowsWeightfff (.param .f32 __cudaparmf1__Z22CalculateShadowsWeightfff, .param .f32 __cudaparmf2__Z22CalculateShadowsWeightfff, .param .f32 __cudaparmf3__Z22CalculateShadowsWeightfff)

	.visible .func (.param .f32 __cudaretf__Z25CalculateHighlightsWeightfff) _Z25CalculateHighlightsWeightfff (.param .f32 __cudaparmf1__Z25CalculateHighlightsWeightfff, .param .f32 __cudaparmf2__Z25CalculateHighlightsWeightfff, .param .f32 __cudaparmf3__Z25CalculateHighlightsWeightfff)

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

	.visible .func (.param .f32 __cudaretf__Z23ApplyBrightnessContrastffff) _Z23ApplyBrightnessContrastffff (.param .f32 __cudaparmf1__Z23ApplyBrightnessContrastffff, .param .f32 __cudaparmf2__Z23ApplyBrightnessContrastffff, .param .f32 __cudaparmf3__Z23ApplyBrightnessContrastffff, .param .f32 __cudaparmf4__Z23ApplyBrightnessContrastffff)

	.visible .func (.param .f32 __cudaretf__Z5ClampIfET_S0_S0_S0_) _Z5ClampIfET_S0_S0_S0_ (.param .f32 __cudaparmf1__Z5ClampIfET_S0_S0_S0_, .param .f32 __cudaparmf2__Z5ClampIfET_S0_S0_S0_, .param .f32 __cudaparmf3__Z5ClampIfET_S0_S0_S0_)

	.visible .func (.param .f32 __cudaretf__Z13CalculateLumaffRK34LumaCorrectorParameters_TonalRangeRK30LumaCorrectorParameters_Levels) _Z13CalculateLumaffRK34LumaCorrectorParameters_TonalRangeRK30LumaCorrectorParameters_Levels (.param .f32 __cudaparmf1__Z13CalculateLumaffRK34LumaCorrectorParameters_TonalRangeRK30LumaCorrectorParameters_Levels, .param .f32 __cudaparmf2__Z13CalculateLumaffRK34LumaCorrectorParameters_TonalRangeRK30LumaCorrectorParameters_Levels, .param .u64 __cudaparmf3__Z13CalculateLumaffRK34LumaCorrectorParameters_TonalRangeRK30LumaCorrectorParameters_Levels, .param .u64 __cudaparmf4__Z13CalculateLumaffRK34LumaCorrectorParameters_TonalRangeRK30LumaCorrectorParameters_Levels)

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

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

	//-----------------------------------------------------------
	// Compiling C:/Users/dvaeng/AppData/Local/Temp/tmpxft_0000368c_00000000-11_LumaCorrector.cpp3.i (C:/Users/dvaeng/AppData/Local/Temp/ccBI#.a16268)
	//-----------------------------------------------------------

	//-----------------------------------------------------------
	// Options:
	//-----------------------------------------------------------
	//  Target:ptx, ISA:sm_20, Endian:little, Pointer Size:64
	//  -O3	(Optimization level)
	//  -g0	(Debug level)
	//  -m2	(Report advisories)
	//-----------------------------------------------------------

	.file	1	"C:/Users/dvaeng/AppData/Local/Temp/tmpxft_0000368c_00000000-10_LumaCorrector.cudafe2.gpu"
	.file	2	"c:\Mulder64\shared\adobe\MediaCore\GPUFoundation\API\Inc\GPUFoundation/PixelFormat.h"
	.file	3	"c:\Mulder64\shared\adobe\MediaCore\Display\Inc\CUDA/Effects/LumaCorrector.h"
	.file	4	"c:\Mulder64\shared\adobe\MediaCore\GPUFoundation\API\Inc\GPUFoundation/KernelSupport/PixelRGB.h"
	.file	5	"c:\Mulder64\shared\adobe\MediaCore\GPUFoundation\API\Inc\GPUFoundation/KernelSupport/PixelYUV.h"
	.file	6	"C:\Program Files (x86)\Microsoft Visual Studio 9.0\VC\include\crtdefs.h"
	.file	7	"c:\Mulder64\shared\adobe\MediaCore\External\3rdParty\NVIDIA\CUDA\win\include\crt/device_runtime.h"
	.file	8	"c:\Mulder64\shared\adobe\MediaCore\External\3rdParty\NVIDIA\CUDA\win\include\host_defines.h"
	.file	9	"c:\Mulder64\shared\adobe\MediaCore\External\3rdParty\NVIDIA\CUDA\win\include\builtin_types.h"
	.file	10	"c:\mulder64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\include\device_types.h"
	.file	11	"c:\mulder64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\include\driver_types.h"
	.file	12	"c:\mulder64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\include\surface_types.h"
	.file	13	"c:\mulder64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\include\texture_types.h"
	.file	14	"c:\mulder64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\include\vector_types.h"
	.file	15	"c:\mulder64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\include\builtin_types.h"
	.file	16	"c:\mulder64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\include\host_defines.h"
	.file	17	"c:\Mulder64\shared\adobe\MediaCore\External\3rdParty\NVIDIA\CUDA\win\include\device_launch_parameters.h"
	.file	18	"c:\mulder64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\include\crt\storage_class.h"
	.file	19	"C:\Program Files (x86)\Microsoft Visual Studio 9.0\VC\include\time.h"
	.file	20	"c:\Mulder64\shared\adobe\MediaCore\Display\Inc\CUDA/Effects/ColorCorrector_Common.h"
	.file	21	"c:\Mulder64\shared\adobe\MediaCore\GPUFoundation\API\Inc\GPUFoundation/KernelSupport/Utils.h"
	.file	22	"c:\Mulder64\shared\adobe\MediaCore\GPUFoundation\API\Inc\GPUFoundation/KernelSupport/VectorUtils.h"
	.file	23	"c:/Mulder64/shared/adobe/MediaCore/Display/Src/CUDA/Effects/LumaCorrector.cu"
	.file	24	"c:\Mulder64\shared\adobe\MediaCore\GPUFoundation\API\Inc\GPUFoundation/Numeric.h"
	.file	25	"c:\Mulder64\shared\adobe\MediaCore\External\3rdParty\NVIDIA\CUDA\win\include\common_functions.h"
	.file	26	"c:\mulder64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\include\math_functions.h"
	.file	27	"c:\mulder64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\include\math_constants.h"
	.file	28	"c:\mulder64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\include\device_functions.h"
	.file	29	"c:\mulder64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\include\sm_11_atomic_functions.h"
	.file	30	"c:\mulder64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\include\sm_12_atomic_functions.h"
	.file	31	"c:\mulder64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\include\sm_13_double_functions.h"
	.file	32	"c:\mulder64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\include\sm_20_atomic_functions.h"
	.file	33	"c:\mulder64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\include\sm_20_intrinsics.h"
	.file	34	"c:\mulder64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\include\surface_functions.h"
	.file	35	"c:\mulder64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\include\texture_fetch_functions.h"
	.file	36	"c:\mulder64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\include\math_functions_dbl_ptx3.h"
	.file	37	"c:\Mulder64\shared\adobe\MediaCore\GPUFoundation\API\Inc\GPUFoundation/KernelSupport/ColorSpaceConvert.h"


	.visible .func (.param .f32 __cudaretf__Z22CalculateShadowsWeightfff) _Z22CalculateShadowsWeightfff (.param .f32 __cudaparmf1__Z22CalculateShadowsWeightfff, .param .f32 __cudaparmf2__Z22CalculateShadowsWeightfff, .param .f32 __cudaparmf3__Z22CalculateShadowsWeightfff)
	{
	.reg .f32 %f<12>;
	.reg .pred %p<4>;
	.loc	20	34	0
$LDWbegin__Z22CalculateShadowsWeightfff:
	ld.param.f32 	%f1, [__cudaparmf1__Z22CalculateShadowsWeightfff];
	mov.f32 	%f2, %f1;
	ld.param.f32 	%f3, [__cudaparmf2__Z22CalculateShadowsWeightfff];
	mov.f32 	%f4, %f3;
	ld.param.f32 	%f5, [__cudaparmf3__Z22CalculateShadowsWeightfff];
	mov.f32 	%f6, %f5;
	.loc	20	35	0
	setp.le.ftz.f32 	%p1, %f2, %f4;
	@!%p1 bra 	$Lt_0_1282;
	.loc	20	37	0
	mov.f32 	%f7, 0f3f800000;     	// 1
	bra.uni 	$LBB6__Z22CalculateShadowsWeightfff;
$Lt_0_1282:
	.loc	20	39	0
	add.ftz.f32 	%f8, %f4, %f6;
	setp.ge.ftz.f32 	%p2, %f2, %f8;
	@!%p2 bra 	$Lt_0_1538;
	.loc	20	41	0
	mov.f32 	%f7, 0f00000000;     	// 0
	bra.uni 	$LBB6__Z22CalculateShadowsWeightfff;
$Lt_0_1538:
	.loc	20	45	0
	sub.ftz.f32 	%f9, %f2, %f4;
	div.approx.ftz.f32 	%f7, %f9, %f6;
$LBB6__Z22CalculateShadowsWeightfff:
	mov.f32 	%f10, %f7;
	st.param.f32 	[__cudaretf__Z22CalculateShadowsWeightfff], %f10;
	ret;
$LDWend__Z22CalculateShadowsWeightfff:
	} // _Z22CalculateShadowsWeightfff

	.visible .func (.param .f32 __cudaretf__Z25CalculateHighlightsWeightfff) _Z25CalculateHighlightsWeightfff (.param .f32 __cudaparmf1__Z25CalculateHighlightsWeightfff, .param .f32 __cudaparmf2__Z25CalculateHighlightsWeightfff, .param .f32 __cudaparmf3__Z25CalculateHighlightsWeightfff)
	{
	.reg .f32 %f<12>;
	.reg .pred %p<4>;
	.loc	20	52	0
$LDWbegin__Z25CalculateHighlightsWeightfff:
	ld.param.f32 	%f1, [__cudaparmf1__Z25CalculateHighlightsWeightfff];
	mov.f32 	%f2, %f1;
	ld.param.f32 	%f3, [__cudaparmf2__Z25CalculateHighlightsWeightfff];
	mov.f32 	%f4, %f3;
	ld.param.f32 	%f5, [__cudaparmf3__Z25CalculateHighlightsWeightfff];
	mov.f32 	%f6, %f5;
	.loc	20	53	0
	sub.ftz.f32 	%f7, %f4, %f6;
	setp.gt.ftz.f32 	%p1, %f7, %f2;
	@!%p1 bra 	$Lt_1_1282;
	.loc	20	55	0
	mov.f32 	%f8, 0f00000000;     	// 0
	bra.uni 	$LBB6__Z25CalculateHighlightsWeightfff;
$Lt_1_1282:
	.loc	20	57	0
	setp.gt.ftz.f32 	%p2, %f2, %f4;
	@!%p2 bra 	$Lt_1_1538;
	.loc	20	59	0
	mov.f32 	%f8, 0f3f800000;     	// 1
	bra.uni 	$LBB6__Z25CalculateHighlightsWeightfff;
$Lt_1_1538:
	.loc	20	63	0
	sub.ftz.f32 	%f9, %f2, %f7;
	div.approx.ftz.f32 	%f8, %f9, %f6;
$LBB6__Z25CalculateHighlightsWeightfff:
	mov.f32 	%f10, %f8;
	st.param.f32 	[__cudaretf__Z25CalculateHighlightsWeightfff], %f10;
	ret;
$LDWend__Z25CalculateHighlightsWeightfff:
	} // _Z25CalculateHighlightsWeightfff

	.visible .func (.param .s32 __cudaretf__Z15IntegerMultiplyii) _Z15IntegerMultiplyii (.param .s32 __cudaparmf1__Z15IntegerMultiplyii, .param .s32 __cudaparmf2__Z15IntegerMultiplyii)
	{
	.reg .u32 %r<7>;
	.loc	21	60	0
$LDWbegin__Z15IntegerMultiplyii:
	ld.param.u32 	%r1, [__cudaparmf1__Z15IntegerMultiplyii];
	mov.s32 	%r2, %r1;
	ld.param.u32 	%r3, [__cudaparmf2__Z15IntegerMultiplyii];
	mov.s32 	%r4, %r3;
	.loc	21	64	0
	mul.lo.s32 	%r5, %r2, %r4;
	st.param.s32 	[__cudaretf__Z15IntegerMultiplyii], %r5;
	ret;
$LDWend__Z15IntegerMultiplyii:
	} // _Z15IntegerMultiplyii

	.visible .func (.param .s32 __cudaretf__Z17Standard2DKernelXv) _Z17Standard2DKernelXv ()
	{
	.reg .u32 %r<7>;
	.loc	21	73	0
$LDWbegin__Z17Standard2DKernelXv:
	.loc	21	74	0
	mov.u32 	%r1, %tid.x;
	cvt.s32.u32 	%r2, %ctaid.x;
	cvt.s32.u32 	%r3, %ntid.x;
	mul.lo.s32 	%r4, %r2, %r3;
	add.u32 	%r5, %r1, %r4;
	st.param.s32 	[__cudaretf__Z17Standard2DKernelXv], %r5;
	ret;
$LDWend__Z17Standard2DKernelXv:
	} // _Z17Standard2DKernelXv

	.visible .func (.param .s32 __cudaretf__Z17Standard2DKernelYv) _Z17Standard2DKernelYv ()
	{
	.reg .u32 %r<7>;
	.loc	21	77	0
$LDWbegin__Z17Standard2DKernelYv:
	.loc	21	78	0
	mov.u32 	%r1, %tid.y;
	cvt.s32.u32 	%r2, %ctaid.y;
	cvt.s32.u32 	%r3, %ntid.y;
	mul.lo.s32 	%r4, %r2, %r3;
	add.u32 	%r5, %r1, %r4;
	st.param.s32 	[__cudaretf__Z17Standard2DKernelYv], %r5;
	ret;
$LDWend__Z17Standard2DKernelYv:
	} // _Z17Standard2DKernelYv

	.visible .func (.param .align 16 .b8 __cudaretf__Z13Half4ToFloat47ushort4[16]) _Z13Half4ToFloat47ushort4 (.param .align 8 .b8 __cudaparmf1__Z13Half4ToFloat47ushort4[8])
	{
	.reg .u32 %r<14>;
	.reg .f32 %f<9>;
	.loc	21	86	0
$LDWbegin__Z13Half4ToFloat47ushort4:
	ld.param.u16 	%r1, [__cudaparmf1__Z13Half4ToFloat47ushort4+0];
	mov.s32 	%r2, %r1;
	ld.param.u16 	%r3, [__cudaparmf1__Z13Half4ToFloat47ushort4+2];
	mov.s32 	%r4, %r3;
	ld.param.u16 	%r5, [__cudaparmf1__Z13Half4ToFloat47ushort4+4];
	mov.s32 	%r6, %r5;
	ld.param.u16 	%r7, [__cudaparmf1__Z13Half4ToFloat47ushort4+6];
	mov.s32 	%r8, %r7;
	.loc	21	87	0
	cvt.u16.u32 	%r9, %r4;
	{ .reg .b32 %b1;
	mov.b32		%b1, %r9;
	cvt.ftz.f32.f16	%f1, %b1; }
	cvt.u16.u32 	%r10, %r6;
	{ .reg .b32 %b1;
	mov.b32		%b1, %r10;
	cvt.ftz.f32.f16	%f2, %b1; }
	cvt.u16.u32 	%r11, %r8;
	{ .reg .b32 %b1;
	mov.b32		%b1, %r11;
	cvt.ftz.f32.f16	%f3, %b1; }
	cvt.u16.u32 	%r12, %r2;
	{ .reg .b32 %b1;
	mov.b32		%b1, %r12;
	cvt.ftz.f32.f16	%f4, %b1; }
	st.param.f32 	[__cudaretf__Z13Half4ToFloat47ushort4+0], %f4;
	mov.f32 	%f5, %f1;
	st.param.f32 	[__cudaretf__Z13Half4ToFloat47ushort4+4], %f5;
	mov.f32 	%f6, %f2;
	st.param.f32 	[__cudaretf__Z13Half4ToFloat47ushort4+8], %f6;
	mov.f32 	%f7, %f3;
	st.param.f32 	[__cudaretf__Z13Half4ToFloat47ushort4+12], %f7;
	ret;
$LDWend__Z13Half4ToFloat47ushort4:
	} // _Z13Half4ToFloat47ushort4

	.visible .func (.param .align 8 .b8 __cudaretf__Z13Float4ToHalf46float4[8]) _Z13Float4ToHalf46float4 (.param .align 16 .b8 __cudaparmf1__Z13Float4ToHalf46float4[16])
	{
	.reg .u32 %r<13>;
	.reg .f32 %f<10>;
	.loc	21	95	0
$LDWbegin__Z13Float4ToHalf46float4:
	ld.param.f32 	%f1, [__cudaparmf1__Z13Float4ToHalf46float4+0];
	mov.f32 	%f2, %f1;
	ld.param.f32 	%f3, [__cudaparmf1__Z13Float4ToHalf46float4+4];
	mov.f32 	%f4, %f3;
	ld.param.f32 	%f5, [__cudaparmf1__Z13Float4ToHalf46float4+8];
	mov.f32 	%f6, %f5;
	ld.param.f32 	%f7, [__cudaparmf1__Z13Float4ToHalf46float4+12];
	mov.f32 	%f8, %f7;
	.loc	21	96	0
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f4;
	mov.b32		%r1, %b1; }
	cvt.u16.u32 	%r2, %r1;
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f6;
	mov.b32		%r3, %b1; }
	cvt.u16.u32 	%r4, %r3;
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f8;
	mov.b32		%r5, %b1; }
	cvt.u16.u32 	%r6, %r5;
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f2;
	mov.b32		%r7, %b1; }
	cvt.u16.u32 	%r8, %r7;
	st.param.u16 	[__cudaretf__Z13Float4ToHalf46float4+0], %r8;
	mov.s32 	%r9, %r2;
	st.param.u16 	[__cudaretf__Z13Float4ToHalf46float4+2], %r9;
	mov.s32 	%r10, %r4;
	st.param.u16 	[__cudaretf__Z13Float4ToHalf46float4+4], %r10;
	mov.s32 	%r11, %r6;
	st.param.u16 	[__cudaretf__Z13Float4ToHalf46float4+6], %r11;
	ret;
$LDWend__Z13Float4ToHalf46float4:
	} // _Z13Float4ToHalf46float4

	.visible .func (.param .u32 __cudaretf__Z4Mix3RjS_S_) _Z4Mix3RjS_S_ (.param .u64 __cudaparmf1__Z4Mix3RjS_S_, .param .u64 __cudaparmf2__Z4Mix3RjS_S_, .param .u64 __cudaparmf3__Z4Mix3RjS_S_)
	{
	.reg .u32 %r<75>;
	.reg .u64 %rd<8>;
	.loc	21	138	0
$LDWbegin__Z4Mix3RjS_S_:
	ld.param.u64 	%rd1, [__cudaparmf1__Z4Mix3RjS_S_];
	mov.s64 	%rd2, %rd1;
	ld.param.u64 	%rd3, [__cudaparmf2__Z4Mix3RjS_S_];
	mov.s64 	%rd4, %rd3;
	ld.param.u64 	%rd5, [__cudaparmf3__Z4Mix3RjS_S_];
	mov.s64 	%rd6, %rd5;
	.loc	21	139	0
	ld.u32 	%r1, [%rd2+0];
	ld.u32 	%r2, [%rd4+0];
	sub.u32 	%r3, %r1, %r2;
	st.u32 	[%rd2+0], %r3;
	ld.u32 	%r4, [%rd6+0];
	sub.u32 	%r5, %r3, %r4;
	st.u32 	[%rd2+0], %r5;
	ld.u32 	%r6, [%rd6+0];
	shr.u32 	%r7, %r6, 13;
	xor.b32 	%r8, %r5, %r7;
	st.u32 	[%rd2+0], %r8;
	.loc	21	140	0
	ld.u32 	%r9, [%rd4+0];
	ld.u32 	%r10, [%rd6+0];
	sub.u32 	%r11, %r9, %r10;
	st.u32 	[%rd4+0], %r11;
	ld.u32 	%r12, [%rd2+0];
	sub.u32 	%r13, %r11, %r12;
	st.u32 	[%rd4+0], %r13;
	ld.u32 	%r14, [%rd2+0];
	shl.b32 	%r15, %r14, 8;
	xor.b32 	%r16, %r13, %r15;
	st.u32 	[%rd4+0], %r16;
	.loc	21	141	0
	ld.u32 	%r17, [%rd6+0];
	ld.u32 	%r18, [%rd2+0];
	sub.u32 	%r19, %r17, %r18;
	st.u32 	[%rd6+0], %r19;
	ld.u32 	%r20, [%rd4+0];
	sub.u32 	%r21, %r19, %r20;
	st.u32 	[%rd6+0], %r21;
	ld.u32 	%r22, [%rd4+0];
	shr.u32 	%r23, %r22, 13;
	xor.b32 	%r24, %r21, %r23;
	st.u32 	[%rd6+0], %r24;
	.loc	21	142	0
	ld.u32 	%r25, [%rd2+0];
	ld.u32 	%r26, [%rd4+0];
	sub.u32 	%r27, %r25, %r26;
	st.u32 	[%rd2+0], %r27;
	ld.u32 	%r28, [%rd6+0];
	sub.u32 	%r29, %r27, %r28;
	st.u32 	[%rd2+0], %r29;
	ld.u32 	%r30, [%rd6+0];
	shr.u32 	%r31, %r30, 12;
	xor.b32 	%r32, %r29, %r31;
	st.u32 	[%rd2+0], %r32;
	.loc	21	143	0
	ld.u32 	%r33, [%rd4+0];
	ld.u32 	%r34, [%rd6+0];
	sub.u32 	%r35, %r33, %r34;
	st.u32 	[%rd4+0], %r35;
	ld.u32 	%r36, [%rd2+0];
	sub.u32 	%r37, %r35, %r36;
	st.u32 	[%rd4+0], %r37;
	ld.u32 	%r38, [%rd2+0];
	shl.b32 	%r39, %r38, 16;
	xor.b32 	%r40, %r37, %r39;
	st.u32 	[%rd4+0], %r40;
	.loc	21	144	0
	ld.u32 	%r41, [%rd6+0];
	ld.u32 	%r42, [%rd2+0];
	sub.u32 	%r43, %r41, %r42;
	st.u32 	[%rd6+0], %r43;
	ld.u32 	%r44, [%rd4+0];
	sub.u32 	%r45, %r43, %r44;
	st.u32 	[%rd6+0], %r45;
	ld.u32 	%r46, [%rd4+0];
	shr.u32 	%r47, %r46, 5;
	xor.b32 	%r48, %r45, %r47;
	st.u32 	[%rd6+0], %r48;
	.loc	21	145	0
	ld.u32 	%r49, [%rd2+0];
	ld.u32 	%r50, [%rd4+0];
	sub.u32 	%r51, %r49, %r50;
	st.u32 	[%rd2+0], %r51;
	ld.u32 	%r52, [%rd6+0];
	sub.u32 	%r53, %r51, %r52;
	st.u32 	[%rd2+0], %r53;
	ld.u32 	%r54, [%rd6+0];
	shr.u32 	%r55, %r54, 3;
	xor.b32 	%r56, %r53, %r55;
	st.u32 	[%rd2+0], %r56;
	.loc	21	146	0
	ld.u32 	%r57, [%rd4+0];
	ld.u32 	%r58, [%rd6+0];
	sub.u32 	%r59, %r57, %r58;
	st.u32 	[%rd4+0], %r59;
	ld.u32 	%r60, [%rd2+0];
	sub.u32 	%r61, %r59, %r60;
	st.u32 	[%rd4+0], %r61;
	ld.u32 	%r62, [%rd2+0];
	shl.b32 	%r63, %r62, 10;
	xor.b32 	%r64, %r61, %r63;
	st.u32 	[%rd4+0], %r64;
	.loc	21	147	0
	ld.u32 	%r65, [%rd6+0];
	ld.u32 	%r66, [%rd2+0];
	sub.u32 	%r67, %r65, %r66;
	st.u32 	[%rd6+0], %r67;
	ld.u32 	%r68, [%rd4+0];
	sub.u32 	%r69, %r67, %r68;
	st.u32 	[%rd6+0], %r69;
	ld.u32 	%r70, [%rd4+0];
	shr.u32 	%r71, %r70, 15;
	xor.b32 	%r72, %r69, %r71;
	st.u32 	[%rd6+0], %r72;
	.loc	21	148	0
	mov.s32 	%r73, %r72;
	st.param.u32 	[__cudaretf__Z4Mix3RjS_S_], %r73;
	ret;
$LDWend__Z4Mix3RjS_S_:
	} // _Z4Mix3RjS_S_

	.visible .func (.param .s32 __cudaretf__Z4Randj) _Z4Randj (.param .u32 __cudaparmf1__Z4Randj)
	{
	.reg .u32 %r<14>;
	.loc	21	152	0
$LDWbegin__Z4Randj:
	ld.param.u32 	%r1, [__cudaparmf1__Z4Randj];
	mov.s32 	%r2, %r1;
	.loc	21	163	0
	mul.lo.u32 	%r3, %r2, 1103515245;
	add.u32 	%r4, %r3, 12345;
	shr.u32 	%r5, %r4, 16;
	and.b32 	%r6, %r5, 255;
	shl.b32 	%r7, %r6, 7;
	mul.lo.u32 	%r8, %r2, -1029531031;
	sub.u32 	%r9, %r8, 740551042;
	shr.u32 	%r10, %r9, 16;
	and.b32 	%r11, %r10, 255;
	xor.b32 	%r12, %r7, %r11;
	st.param.s32 	[__cudaretf__Z4Randj], %r12;
	ret;
$LDWend__Z4Randj:
	} // _Z4Randj

	.visible .func (.param .s32 __cudaretf__Z6Rand2Djjj) _Z6Rand2Djjj (.param .u32 __cudaparmf1__Z6Rand2Djjj, .param .u32 __cudaparmf2__Z6Rand2Djjj, .param .u32 __cudaparmf3__Z6Rand2Djjj)
	{
	.reg .u32 %r<54>;
	.loc	21	169	0
$LDWbegin__Z6Rand2Djjj:
	ld.param.u32 	%r1, [__cudaparmf1__Z6Rand2Djjj];
	mov.s32 	%r2, %r1;
	ld.param.u32 	%r3, [__cudaparmf2__Z6Rand2Djjj];
	mov.s32 	%r4, %r3;
	ld.param.u32 	%r5, [__cudaparmf3__Z6Rand2Djjj];
	mov.s32 	%r6, %r5;
	.loc	21	139	0
	sub.u32 	%r7, %r2, %r4;
	sub.u32 	%r8, %r7, %r6;
	shr.u32 	%r9, %r6, 13;
	xor.b32 	%r10, %r8, %r9;
	.loc	21	140	0
	sub.u32 	%r11, %r4, %r6;
	sub.u32 	%r12, %r11, %r10;
	shl.b32 	%r13, %r10, 8;
	xor.b32 	%r14, %r12, %r13;
	.loc	21	141	0
	sub.u32 	%r15, %r6, %r10;
	sub.u32 	%r16, %r15, %r14;
	shr.u32 	%r17, %r14, 13;
	xor.b32 	%r18, %r16, %r17;
	.loc	21	142	0
	sub.u32 	%r19, %r10, %r14;
	sub.u32 	%r20, %r19, %r18;
	shr.u32 	%r21, %r18, 12;
	xor.b32 	%r22, %r20, %r21;
	.loc	21	143	0
	sub.u32 	%r23, %r14, %r18;
	sub.u32 	%r24, %r23, %r22;
	shl.b32 	%r25, %r22, 16;
	xor.b32 	%r26, %r24, %r25;
	.loc	21	144	0
	sub.u32 	%r27, %r18, %r22;
	sub.u32 	%r28, %r27, %r26;
	shr.u32 	%r29, %r26, 5;
	xor.b32 	%r30, %r28, %r29;
	.loc	21	145	0
	sub.u32 	%r31, %r22, %r26;
	sub.u32 	%r32, %r31, %r30;
	shr.u32 	%r33, %r30, 3;
	xor.b32 	%r34, %r32, %r33;
	.loc	21	146	0
	sub.u32 	%r35, %r26, %r30;
	sub.u32 	%r36, %r35, %r34;
	shl.b32 	%r37, %r34, 10;
	xor.b32 	%r38, %r36, %r37;
	.loc	21	147	0
	sub.u32 	%r39, %r30, %r34;
	sub.u32 	%r40, %r39, %r38;
	shr.u32 	%r41, %r38, 15;
	xor.b32 	%r42, %r40, %r41;
	.loc	21	170	0
	mul.lo.u32 	%r43, %r42, 1103515245;
	add.u32 	%r44, %r43, 12345;
	shr.u32 	%r45, %r44, 16;
	and.b32 	%r46, %r45, 255;
	shl.b32 	%r47, %r46, 7;
	mul.lo.u32 	%r48, %r42, -1029531031;
	sub.u32 	%r49, %r48, 740551042;
	shr.u32 	%r50, %r49, 16;
	and.b32 	%r51, %r50, 255;
	xor.b32 	%r52, %r47, %r51;
	st.param.s32 	[__cudaretf__Z6Rand2Djjj], %r52;
	ret;
$LDWend__Z6Rand2Djjj:
	} // _Z6Rand2Djjj

	.visible .func (.param .s32 __cudaretf__Z6Rand2Dj) _Z6Rand2Dj (.param .u32 __cudaparmf1__Z6Rand2Dj)
	{
	.reg .u32 %r<60>;
	.loc	21	175	0
$LDWbegin__Z6Rand2Dj:
	ld.param.u32 	%r1, [__cudaparmf1__Z6Rand2Dj];
	mov.s32 	%r2, %r1;
	.loc	21	143	0
	cvt.s32.u32 	%r3, %ctaid.y;
	cvt.s32.u32 	%r4, %ntid.y;
	mul.lo.s32 	%r5, %r3, %r4;
	cvt.s32.u32 	%r6, %ctaid.x;
	cvt.s32.u32 	%r7, %ntid.x;
	mul.lo.s32 	%r8, %r6, %r7;
	mov.u32 	%r9, %tid.y;
	add.u32 	%r10, %r5, %r9;
	mov.u32 	%r11, %tid.x;
	add.u32 	%r12, %r8, %r11;
	shr.u32 	%r13, %r10, 13;
	sub.u32 	%r14, %r2, %r12;
	sub.u32 	%r15, %r12, %r10;
	sub.u32 	%r16, %r14, %r10;
	xor.b32 	%r17, %r13, %r16;
	shl.b32 	%r18, %r17, 8;
	sub.u32 	%r19, %r15, %r17;
	sub.u32 	%r20, %r10, %r17;
	xor.b32 	%r21, %r18, %r19;
	shr.u32 	%r22, %r21, 13;
	sub.u32 	%r23, %r20, %r21;
	sub.u32 	%r24, %r17, %r21;
	xor.b32 	%r25, %r22, %r23;
	shr.u32 	%r26, %r25, 12;
	sub.u32 	%r27, %r24, %r25;
	xor.b32 	%r28, %r26, %r27;
	sub.u32 	%r29, %r21, %r25;
	sub.u32 	%r30, %r29, %r28;
	shl.b32 	%r31, %r28, 16;
	xor.b32 	%r32, %r30, %r31;
	.loc	21	144	0
	sub.u32 	%r33, %r25, %r28;
	sub.u32 	%r34, %r33, %r32;
	shr.u32 	%r35, %r32, 5;
	xor.b32 	%r36, %r34, %r35;
	.loc	21	145	0
	sub.u32 	%r37, %r28, %r32;
	sub.u32 	%r38, %r37, %r36;
	shr.u32 	%r39, %r36, 3;
	xor.b32 	%r40, %r38, %r39;
	.loc	21	146	0
	sub.u32 	%r41, %r32, %r36;
	sub.u32 	%r42, %r41, %r40;
	shl.b32 	%r43, %r40, 10;
	xor.b32 	%r44, %r42, %r43;
	.loc	21	147	0
	sub.u32 	%r45, %r36, %r40;
	sub.u32 	%r46, %r45, %r44;
	shr.u32 	%r47, %r44, 15;
	xor.b32 	%r48, %r46, %r47;
	.loc	21	176	0
	mul.lo.u32 	%r49, %r48, 1103515245;
	add.u32 	%r50, %r49, 12345;
	shr.u32 	%r51, %r50, 16;
	and.b32 	%r52, %r51, 255;
	shl.b32 	%r53, %r52, 7;
	mul.lo.u32 	%r54, %r48, -1029531031;
	sub.u32 	%r55, %r54, 740551042;
	shr.u32 	%r56, %r55, 16;
	and.b32 	%r57, %r56, 255;
	xor.b32 	%r58, %r53, %r57;
	st.param.s32 	[__cudaretf__Z6Rand2Dj], %r58;
	ret;
$LDWend__Z6Rand2Dj:
	} // _Z6Rand2Dj

	.visible .func (.param .align 8 .b8 __cudaretf__Z6Read2DI7ushort4ET_PKS1_iii[8]) _Z6Read2DI7ushort4ET_PKS1_iii (.param .u64 __cudaparmf1__Z6Read2DI7ushort4ET_PKS1_iii, .param .s32 __cudaparmf2__Z6Read2DI7ushort4ET_PKS1_iii, .param .s32 __cudaparmf3__Z6Read2DI7ushort4ET_PKS1_iii, .param .s32 __cudaparmf4__Z6Read2DI7ushort4ET_PKS1_iii)
	{
	.reg .u32 %r<14>;
	.reg .u64 %rd<7>;
	.loc	21	114	0
$LDWbegin__Z6Read2DI7ushort4ET_PKS1_iii:
	ld.param.u64 	%rd1, [__cudaparmf1__Z6Read2DI7ushort4ET_PKS1_iii];
	mov.s64 	%rd2, %rd1;
	ld.param.u32 	%r1, [__cudaparmf2__Z6Read2DI7ushort4ET_PKS1_iii];
	mov.s32 	%r2, %r1;
	ld.param.u32 	%r3, [__cudaparmf3__Z6Read2DI7ushort4ET_PKS1_iii];
	mov.s32 	%r4, %r3;
	ld.param.u32 	%r5, [__cudaparmf4__Z6Read2DI7ushort4ET_PKS1_iii];
	mov.s32 	%r6, %r5;
	.loc	21	115	0
	mul.lo.s32 	%r7, %r2, %r6;
	add.s32 	%r8, %r4, %r7;
	cvt.s64.s32 	%rd3, %r8;
	mul.wide.s32 	%rd4, %r8, 8;
	add.u64 	%rd5, %rd2, %rd4;
	ld.v4.u16 	{%r9,%r10,%r11,%r12}, [%rd5+0];
	st.param.u16 	[__cudaretf__Z6Read2DI7ushort4ET_PKS1_iii+0], %r9;
	st.param.u16 	[__cudaretf__Z6Read2DI7ushort4ET_PKS1_iii+2], %r10;
	st.param.u16 	[__cudaretf__Z6Read2DI7ushort4ET_PKS1_iii+4], %r11;
	st.param.u16 	[__cudaretf__Z6Read2DI7ushort4ET_PKS1_iii+6], %r12;
	ret;
$LDWend__Z6Read2DI7ushort4ET_PKS1_iii:
	} // _Z6Read2DI7ushort4ET_PKS1_iii

	.visible .func (.param .align 16 .b8 __cudaretf__Z6Read2DI6float4ET_PKS1_iii[16]) _Z6Read2DI6float4ET_PKS1_iii (.param .u64 __cudaparmf1__Z6Read2DI6float4ET_PKS1_iii, .param .s32 __cudaparmf2__Z6Read2DI6float4ET_PKS1_iii, .param .s32 __cudaparmf3__Z6Read2DI6float4ET_PKS1_iii, .param .s32 __cudaparmf4__Z6Read2DI6float4ET_PKS1_iii)
	{
	.reg .u32 %r<10>;
	.reg .u64 %rd<7>;
	.reg .f32 %f<6>;
	.loc	21	114	0
$LDWbegin__Z6Read2DI6float4ET_PKS1_iii:
	ld.param.u64 	%rd1, [__cudaparmf1__Z6Read2DI6float4ET_PKS1_iii];
	mov.s64 	%rd2, %rd1;
	ld.param.u32 	%r1, [__cudaparmf2__Z6Read2DI6float4ET_PKS1_iii];
	mov.s32 	%r2, %r1;
	ld.param.u32 	%r3, [__cudaparmf3__Z6Read2DI6float4ET_PKS1_iii];
	mov.s32 	%r4, %r3;
	ld.param.u32 	%r5, [__cudaparmf4__Z6Read2DI6float4ET_PKS1_iii];
	mov.s32 	%r6, %r5;
	.loc	21	115	0
	mul.lo.s32 	%r7, %r2, %r6;
	add.s32 	%r8, %r4, %r7;
	cvt.s64.s32 	%rd3, %r8;
	mul.wide.s32 	%rd4, %r8, 16;
	add.u64 	%rd5, %rd2, %rd4;
	ld.v4.f32 	{%f1,%f2,%f3,%f4}, [%rd5+0];
	st.param.f32 	[__cudaretf__Z6Read2DI6float4ET_PKS1_iii+0], %f1;
	st.param.f32 	[__cudaretf__Z6Read2DI6float4ET_PKS1_iii+4], %f2;
	st.param.f32 	[__cudaretf__Z6Read2DI6float4ET_PKS1_iii+8], %f3;
	st.param.f32 	[__cudaretf__Z6Read2DI6float4ET_PKS1_iii+12], %f4;
	ret;
$LDWend__Z6Read2DI6float4ET_PKS1_iii:
	} // _Z6Read2DI6float4ET_PKS1_iii

	.visible .func _Z7Write2DI7ushort4EvT_PS1_iii (.param .align 8 .b8 __cudaparmf1__Z7Write2DI7ushort4EvT_PS1_iii[8], .param .u64 __cudaparmf2__Z7Write2DI7ushort4EvT_PS1_iii, .param .s32 __cudaparmf3__Z7Write2DI7ushort4EvT_PS1_iii, .param .s32 __cudaparmf4__Z7Write2DI7ushort4EvT_PS1_iii, .param .s32 __cudaparmf5__Z7Write2DI7ushort4EvT_PS1_iii)
	{
	.reg .u32 %r<18>;
	.reg .u64 %rd<7>;
	.loc	21	125	0
$LDWbegin__Z7Write2DI7ushort4EvT_PS1_iii:
	ld.param.u16 	%r1, [__cudaparmf1__Z7Write2DI7ushort4EvT_PS1_iii+0];
	mov.s32 	%r2, %r1;
	ld.param.u16 	%r3, [__cudaparmf1__Z7Write2DI7ushort4EvT_PS1_iii+2];
	mov.s32 	%r4, %r3;
	ld.param.u16 	%r5, [__cudaparmf1__Z7Write2DI7ushort4EvT_PS1_iii+4];
	mov.s32 	%r6, %r5;
	ld.param.u16 	%r7, [__cudaparmf1__Z7Write2DI7ushort4EvT_PS1_iii+6];
	mov.s32 	%r8, %r7;
	ld.param.u64 	%rd1, [__cudaparmf2__Z7Write2DI7ushort4EvT_PS1_iii];
	mov.s64 	%rd2, %rd1;
	ld.param.u32 	%r9, [__cudaparmf3__Z7Write2DI7ushort4EvT_PS1_iii];
	mov.s32 	%r10, %r9;
	ld.param.u32 	%r11, [__cudaparmf4__Z7Write2DI7ushort4EvT_PS1_iii];
	mov.s32 	%r12, %r11;
	ld.param.u32 	%r13, [__cudaparmf5__Z7Write2DI7ushort4EvT_PS1_iii];
	mov.s32 	%r14, %r13;
	.loc	21	126	0
	mul.lo.s32 	%r15, %r10, %r14;
	add.s32 	%r16, %r12, %r15;
	cvt.s64.s32 	%rd3, %r16;
	mul.wide.s32 	%rd4, %r16, 8;
	add.u64 	%rd5, %rd2, %rd4;
	st.v4.u16 	[%rd5+0], {%r2,%r4,%r6,%r8};
	.loc	21	127	0
	ret;
$LDWend__Z7Write2DI7ushort4EvT_PS1_iii:
	} // _Z7Write2DI7ushort4EvT_PS1_iii

	.visible .func _Z7Write2DI6float4EvT_PS1_iii (.param .align 16 .b8 __cudaparmf1__Z7Write2DI6float4EvT_PS1_iii[16], .param .u64 __cudaparmf2__Z7Write2DI6float4EvT_PS1_iii, .param .s32 __cudaparmf3__Z7Write2DI6float4EvT_PS1_iii, .param .s32 __cudaparmf4__Z7Write2DI6float4EvT_PS1_iii, .param .s32 __cudaparmf5__Z7Write2DI6float4EvT_PS1_iii)
	{
	.reg .u32 %r<10>;
	.reg .u64 %rd<7>;
	.reg .f32 %f<10>;
	.loc	21	125	0
$LDWbegin__Z7Write2DI6float4EvT_PS1_iii:
	ld.param.f32 	%f1, [__cudaparmf1__Z7Write2DI6float4EvT_PS1_iii+0];
	mov.f32 	%f2, %f1;
	ld.param.f32 	%f3, [__cudaparmf1__Z7Write2DI6float4EvT_PS1_iii+4];
	mov.f32 	%f4, %f3;
	ld.param.f32 	%f5, [__cudaparmf1__Z7Write2DI6float4EvT_PS1_iii+8];
	mov.f32 	%f6, %f5;
	ld.param.f32 	%f7, [__cudaparmf1__Z7Write2DI6float4EvT_PS1_iii+12];
	mov.f32 	%f8, %f7;
	ld.param.u64 	%rd1, [__cudaparmf2__Z7Write2DI6float4EvT_PS1_iii];
	mov.s64 	%rd2, %rd1;
	ld.param.u32 	%r1, [__cudaparmf3__Z7Write2DI6float4EvT_PS1_iii];
	mov.s32 	%r2, %r1;
	ld.param.u32 	%r3, [__cudaparmf4__Z7Write2DI6float4EvT_PS1_iii];
	mov.s32 	%r4, %r3;
	ld.param.u32 	%r5, [__cudaparmf5__Z7Write2DI6float4EvT_PS1_iii];
	mov.s32 	%r6, %r5;
	.loc	21	126	0
	mul.lo.s32 	%r7, %r2, %r6;
	add.s32 	%r8, %r4, %r7;
	cvt.s64.s32 	%rd3, %r8;
	mul.wide.s32 	%rd4, %r8, 16;
	add.u64 	%rd5, %rd2, %rd4;
	st.v4.f32 	[%rd5+0], {%f2,%f4,%f6,%f8};
	.loc	21	127	0
	ret;
$LDWend__Z7Write2DI6float4EvT_PS1_iii:
	} // _Z7Write2DI6float4EvT_PS1_iii

	.visible .func (.param .align 16 .b8 __cudaretf__Z18UnpremultiplyPixel8PixelRGB[16]) _Z18UnpremultiplyPixel8PixelRGB (.param .align 16 .b8 __cudaparmf1__Z18UnpremultiplyPixel8PixelRGB[16])
	{
	.reg .f32 %f<23>;
	.reg .pred %p<3>;
	.loc	4	206	0
$LDWbegin__Z18UnpremultiplyPixel8PixelRGB:
	ld.param.f32 	%f1, [__cudaparmf1__Z18UnpremultiplyPixel8PixelRGB+0];
	mov.f32 	%f2, %f1;
	ld.param.f32 	%f3, [__cudaparmf1__Z18UnpremultiplyPixel8PixelRGB+4];
	mov.f32 	%f4, %f3;
	ld.param.f32 	%f5, [__cudaparmf1__Z18UnpremultiplyPixel8PixelRGB+8];
	mov.f32 	%f6, %f5;
	ld.param.f32 	%f7, [__cudaparmf1__Z18UnpremultiplyPixel8PixelRGB+12];
	mov.f32 	%f8, %f7;
	.loc	4	208	0
	cvt.ftz.sat.f32.f32 	%f9, %f8;
	mov.f32 	%f10, %f9;
	mov.f32 	%f11, 0fb70637bd;    	// -8e-006
	add.ftz.f32 	%f12, %f9, %f11;
	mov.f32 	%f13, 0f00000000;    	// 0
	setp.le.ftz.f32 	%p1, %f12, %f13;
	@%p1 bra 	$Lt_15_1282;
	.loc	4	213	0
	rcp.approx.ftz.f32 	%f14, %f9;
	mul.ftz.f32 	%f15, %f14, %f6;
	.loc	4	214	0
	mul.ftz.f32 	%f16, %f14, %f4;
	.loc	4	215	0
	mul.ftz.f32 	%f17, %f14, %f2;
	bra.uni 	$Lt_15_1026;
$Lt_15_1282:
	.loc	4	219	0
	mov.f32 	%f15, 0f00000000;    	// 0
	mov.f32 	%f16, 0f00000000;    	// 0
	mov.f32 	%f17, 0f00000000;    	// 0
	mov.f32 	%f10, 0f00000000;    	// 0
$Lt_15_1026:
	.loc	4	224	0
	mov.f32 	%f18, %f17;
	st.param.f32 	[__cudaretf__Z18UnpremultiplyPixel8PixelRGB+0], %f18;
	mov.f32 	%f19, %f16;
	st.param.f32 	[__cudaretf__Z18UnpremultiplyPixel8PixelRGB+4], %f19;
	mov.f32 	%f20, %f15;
	st.param.f32 	[__cudaretf__Z18UnpremultiplyPixel8PixelRGB+8], %f20;
	mov.f32 	%f21, %f10;
	st.param.f32 	[__cudaretf__Z18UnpremultiplyPixel8PixelRGB+12], %f21;
	ret;
$LDWend__Z18UnpremultiplyPixel8PixelRGB:
	} // _Z18UnpremultiplyPixel8PixelRGB

	.visible .func (.param .f32 __cudaretf__Z13ToLinearColorf) _Z13ToLinearColorf (.param .f32 __cudaparmf1__Z13ToLinearColorf)
	{
	.reg .f32 %f<15>;
	.reg .pred %p<3>;
	.loc	4	231	0
$LDWbegin__Z13ToLinearColorf:
	ld.param.f32 	%f1, [__cudaparmf1__Z13ToLinearColorf];
	mov.f32 	%f2, %f1;
	mov.f32 	%f3, 0f00000000;     	// 0
	setp.lt.ftz.f32 	%p1, %f2, %f3;
	@!%p1 bra 	$Lt_16_1026;
	.loc	4	234	0
	neg.ftz.f32 	%f4, %f2;
	lg2.approx.ftz.f32 	%f5, %f4;
	mov.f32 	%f6, 0f400ccccd;     	// 2.2
	mul.ftz.f32 	%f7, %f5, %f6;
	ex2.approx.ftz.f32 	%f8, %f7;
	neg.ftz.f32 	%f9, %f8;
	bra.uni 	$LBB4__Z13ToLinearColorf;
$Lt_16_1026:
	.loc	4	236	0
	lg2.approx.ftz.f32 	%f10, %f2;
	mov.f32 	%f11, 0f400ccccd;    	// 2.2
	mul.ftz.f32 	%f12, %f10, %f11;
	ex2.approx.ftz.f32 	%f9, %f12;
$LBB4__Z13ToLinearColorf:
	mov.f32 	%f13, %f9;
	st.param.f32 	[__cudaretf__Z13ToLinearColorf], %f13;
	ret;
$LDWend__Z13ToLinearColorf:
	} // _Z13ToLinearColorf

	.visible .func (.param .f32 __cudaretf__Z15FromLinearColorf) _Z15FromLinearColorf (.param .f32 __cudaparmf1__Z15FromLinearColorf)
	{
	.reg .f32 %f<15>;
	.reg .pred %p<3>;
	.loc	4	239	0
$LDWbegin__Z15FromLinearColorf:
	ld.param.f32 	%f1, [__cudaparmf1__Z15FromLinearColorf];
	mov.f32 	%f2, %f1;
	mov.f32 	%f3, 0f00000000;     	// 0
	setp.lt.ftz.f32 	%p1, %f2, %f3;
	@!%p1 bra 	$Lt_17_1026;
	.loc	4	242	0
	neg.ftz.f32 	%f4, %f2;
	lg2.approx.ftz.f32 	%f5, %f4;
	mov.f32 	%f6, 0f3ee8ba2e;     	// 0.454545
	mul.ftz.f32 	%f7, %f5, %f6;
	ex2.approx.ftz.f32 	%f8, %f7;
	neg.ftz.f32 	%f9, %f8;
	bra.uni 	$LBB4__Z15FromLinearColorf;
$Lt_17_1026:
	.loc	4	244	0
	lg2.approx.ftz.f32 	%f10, %f2;
	mov.f32 	%f11, 0f3ee8ba2e;    	// 0.454545
	mul.ftz.f32 	%f12, %f10, %f11;
	ex2.approx.ftz.f32 	%f9, %f12;
$LBB4__Z15FromLinearColorf:
	mov.f32 	%f13, %f9;
	st.param.f32 	[__cudaretf__Z15FromLinearColorf], %f13;
	ret;
$LDWend__Z15FromLinearColorf:
	} // _Z15FromLinearColorf

	.visible .func (.param .align 16 .b8 __cudaretf__Z25PremultiplyLinearizePixel8PixelRGB[16]) _Z25PremultiplyLinearizePixel8PixelRGB (.param .align 16 .b8 __cudaparmf1__Z25PremultiplyLinearizePixel8PixelRGB[16])
	{
	.reg .f32 %f<47>;
	.reg .pred %p<5>;
	.loc	4	252	0
$LDWbegin__Z25PremultiplyLinearizePixel8PixelRGB:
	ld.param.f32 	%f1, [__cudaparmf1__Z25PremultiplyLinearizePixel8PixelRGB+0];
	mov.f32 	%f2, %f1;
	ld.param.f32 	%f3, [__cudaparmf1__Z25PremultiplyLinearizePixel8PixelRGB+4];
	mov.f32 	%f4, %f3;
	ld.param.f32 	%f5, [__cudaparmf1__Z25PremultiplyLinearizePixel8PixelRGB+8];
	mov.f32 	%f6, %f5;
	ld.param.f32 	%f7, [__cudaparmf1__Z25PremultiplyLinearizePixel8PixelRGB+12];
	mov.f32 	%f8, %f7;
	.loc	4	254	0
	cvt.ftz.sat.f32.f32 	%f9, %f8;
	.loc	4	255	0
	mov.f32 	%f10, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p1, %f2, %f10;
	@!%p1 bra 	$Lt_18_4098;
	.loc	4	234	0
	neg.ftz.f32 	%f11, %f2;
	lg2.approx.ftz.f32 	%f12, %f11;
	mov.f32 	%f13, 0f400ccccd;    	// 2.2
	mul.ftz.f32 	%f14, %f12, %f13;
	ex2.approx.ftz.f32 	%f15, %f14;
	neg.ftz.f32 	%f16, %f15;
	bra.uni 	$LDWendi___log2f_195_5;
$Lt_18_4098:
	.loc	4	236	0
	lg2.approx.ftz.f32 	%f17, %f2;
	mov.f32 	%f18, 0f400ccccd;    	// 2.2
	mul.ftz.f32 	%f19, %f17, %f18;
	ex2.approx.ftz.f32 	%f16, %f19;
$LDWendi___log2f_195_5:
	.loc	4	256	0
	mov.f32 	%f20, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p2, %f4, %f20;
	@!%p2 bra 	$Lt_18_4610;
	.loc	4	234	0
	neg.ftz.f32 	%f21, %f4;
	lg2.approx.ftz.f32 	%f22, %f21;
	mov.f32 	%f23, 0f400ccccd;    	// 2.2
	mul.ftz.f32 	%f24, %f22, %f23;
	ex2.approx.ftz.f32 	%f25, %f24;
	neg.ftz.f32 	%f26, %f25;
	bra.uni 	$LDWendi___log2f_195_3;
$Lt_18_4610:
	.loc	4	236	0
	lg2.approx.ftz.f32 	%f27, %f4;
	mov.f32 	%f28, 0f400ccccd;    	// 2.2
	mul.ftz.f32 	%f29, %f27, %f28;
	ex2.approx.ftz.f32 	%f26, %f29;
$LDWendi___log2f_195_3:
	.loc	4	257	0
	mov.f32 	%f30, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p3, %f6, %f30;
	@!%p3 bra 	$Lt_18_5122;
	.loc	4	234	0
	neg.ftz.f32 	%f31, %f6;
	lg2.approx.ftz.f32 	%f32, %f31;
	mov.f32 	%f33, 0f400ccccd;    	// 2.2
	mul.ftz.f32 	%f34, %f32, %f33;
	ex2.approx.ftz.f32 	%f35, %f34;
	neg.ftz.f32 	%f36, %f35;
	bra.uni 	$LDWendi___log2f_195_1;
$Lt_18_5122:
	.loc	4	236	0
	lg2.approx.ftz.f32 	%f37, %f6;
	mov.f32 	%f38, 0f400ccccd;    	// 2.2
	mul.ftz.f32 	%f39, %f37, %f38;
	ex2.approx.ftz.f32 	%f36, %f39;
$LDWendi___log2f_195_1:
	.loc	4	259	0
	mul.ftz.f32 	%f40, %f36, %f9;
	mul.ftz.f32 	%f41, %f26, %f9;
	mul.ftz.f32 	%f42, %f16, %f9;
	st.param.f32 	[__cudaretf__Z25PremultiplyLinearizePixel8PixelRGB+0], %f42;
	mov.f32 	%f43, %f41;
	st.param.f32 	[__cudaretf__Z25PremultiplyLinearizePixel8PixelRGB+4], %f43;
	mov.f32 	%f44, %f40;
	st.param.f32 	[__cudaretf__Z25PremultiplyLinearizePixel8PixelRGB+8], %f44;
	mov.f32 	%f45, %f9;
	st.param.f32 	[__cudaretf__Z25PremultiplyLinearizePixel8PixelRGB+12], %f45;
	ret;
$LDWend__Z25PremultiplyLinearizePixel8PixelRGB:
	} // _Z25PremultiplyLinearizePixel8PixelRGB

	.visible .func (.param .align 16 .b8 __cudaretf__Z29UnpremultiplyUnlinearizePixel8PixelRGB[16]) _Z29UnpremultiplyUnlinearizePixel8PixelRGB (.param .align 16 .b8 __cudaparmf1__Z29UnpremultiplyUnlinearizePixel8PixelRGB[16])
	{
	.reg .f32 %f<53>;
	.reg .pred %p<6>;
	.loc	4	263	0
$LDWbegin__Z29UnpremultiplyUnlinearizePixel8PixelRGB:
	ld.param.f32 	%f1, [__cudaparmf1__Z29UnpremultiplyUnlinearizePixel8PixelRGB+0];
	mov.f32 	%f2, %f1;
	ld.param.f32 	%f3, [__cudaparmf1__Z29UnpremultiplyUnlinearizePixel8PixelRGB+4];
	mov.f32 	%f4, %f3;
	ld.param.f32 	%f5, [__cudaparmf1__Z29UnpremultiplyUnlinearizePixel8PixelRGB+8];
	mov.f32 	%f6, %f5;
	ld.param.f32 	%f7, [__cudaparmf1__Z29UnpremultiplyUnlinearizePixel8PixelRGB+12];
	mov.f32 	%f8, %f7;
	.loc	4	208	0
	cvt.ftz.sat.f32.f32 	%f9, %f8;
	mov.f32 	%f10, %f9;
	mov.f32 	%f11, 0fb70637bd;    	// -8e-006
	add.ftz.f32 	%f12, %f9, %f11;
	mov.f32 	%f13, 0f00000000;    	// 0
	setp.le.ftz.f32 	%p1, %f12, %f13;
	@%p1 bra 	$Lt_19_5122;
	.loc	4	213	0
	rcp.approx.ftz.f32 	%f14, %f9;
	mul.ftz.f32 	%f15, %f14, %f6;
	.loc	4	214	0
	mul.ftz.f32 	%f16, %f14, %f4;
	.loc	4	215	0
	mul.ftz.f32 	%f17, %f14, %f2;
	bra.uni 	$Lt_19_4866;
$Lt_19_5122:
	.loc	4	219	0
	mov.f32 	%f15, 0f00000000;    	// 0
	mov.f32 	%f16, 0f00000000;    	// 0
	mov.f32 	%f17, 0f00000000;    	// 0
	mov.f32 	%f10, 0f00000000;    	// 0
$Lt_19_4866:
	.loc	4	266	0
	mov.f32 	%f18, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p2, %f17, %f18;
	@!%p2 bra 	$Lt_19_5378;
	.loc	4	242	0
	neg.ftz.f32 	%f19, %f17;
	lg2.approx.ftz.f32 	%f20, %f19;
	mov.f32 	%f21, 0f3ee8ba2e;    	// 0.454545
	mul.ftz.f32 	%f22, %f20, %f21;
	ex2.approx.ftz.f32 	%f23, %f22;
	neg.ftz.f32 	%f24, %f23;
	bra.uni 	$LDWendi___log2f_196_5;
$Lt_19_5378:
	.loc	4	244	0
	lg2.approx.ftz.f32 	%f25, %f17;
	mov.f32 	%f26, 0f3ee8ba2e;    	// 0.454545
	mul.ftz.f32 	%f27, %f25, %f26;
	ex2.approx.ftz.f32 	%f24, %f27;
$LDWendi___log2f_196_5:
	.loc	4	267	0
	mov.f32 	%f28, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p3, %f16, %f28;
	@!%p3 bra 	$Lt_19_5890;
	.loc	4	242	0
	neg.ftz.f32 	%f29, %f16;
	lg2.approx.ftz.f32 	%f30, %f29;
	mov.f32 	%f31, 0f3ee8ba2e;    	// 0.454545
	mul.ftz.f32 	%f32, %f30, %f31;
	ex2.approx.ftz.f32 	%f33, %f32;
	neg.ftz.f32 	%f34, %f33;
	bra.uni 	$LDWendi___log2f_196_3;
$Lt_19_5890:
	.loc	4	244	0
	lg2.approx.ftz.f32 	%f35, %f16;
	mov.f32 	%f36, 0f3ee8ba2e;    	// 0.454545
	mul.ftz.f32 	%f37, %f35, %f36;
	ex2.approx.ftz.f32 	%f34, %f37;
$LDWendi___log2f_196_3:
	.loc	4	268	0
	mov.f32 	%f38, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p4, %f15, %f38;
	@!%p4 bra 	$Lt_19_6402;
	.loc	4	242	0
	neg.ftz.f32 	%f39, %f15;
	lg2.approx.ftz.f32 	%f40, %f39;
	mov.f32 	%f41, 0f3ee8ba2e;    	// 0.454545
	mul.ftz.f32 	%f42, %f40, %f41;
	ex2.approx.ftz.f32 	%f43, %f42;
	neg.ftz.f32 	%f44, %f43;
	bra.uni 	$LDWendi___log2f_196_1;
$Lt_19_6402:
	.loc	4	244	0
	lg2.approx.ftz.f32 	%f45, %f15;
	mov.f32 	%f46, 0f3ee8ba2e;    	// 0.454545
	mul.ftz.f32 	%f47, %f45, %f46;
	ex2.approx.ftz.f32 	%f44, %f47;
$LDWendi___log2f_196_1:
	.loc	4	269	0
	mov.f32 	%f48, %f24;
	st.param.f32 	[__cudaretf__Z29UnpremultiplyUnlinearizePixel8PixelRGB+0], %f48;
	mov.f32 	%f49, %f34;
	st.param.f32 	[__cudaretf__Z29UnpremultiplyUnlinearizePixel8PixelRGB+4], %f49;
	mov.f32 	%f50, %f44;
	st.param.f32 	[__cudaretf__Z29UnpremultiplyUnlinearizePixel8PixelRGB+8], %f50;
	mov.f32 	%f51, %f10;
	st.param.f32 	[__cudaretf__Z29UnpremultiplyUnlinearizePixel8PixelRGB+12], %f51;
	ret;
$LDWend__Z29UnpremultiplyUnlinearizePixel8PixelRGB:
	} // _Z29UnpremultiplyUnlinearizePixel8PixelRGB

	.visible .func (.param .align 16 .b8 __cudaretf__Z20PremultiplyLinearize6float4[16]) _Z20PremultiplyLinearize6float4 (.param .align 16 .b8 __cudaparmf1__Z20PremultiplyLinearize6float4[16])
	{
	.reg .f32 %f<47>;
	.reg .pred %p<5>;
	.loc	4	277	0
$LDWbegin__Z20PremultiplyLinearize6float4:
	ld.param.f32 	%f1, [__cudaparmf1__Z20PremultiplyLinearize6float4+0];
	mov.f32 	%f2, %f1;
	ld.param.f32 	%f3, [__cudaparmf1__Z20PremultiplyLinearize6float4+4];
	mov.f32 	%f4, %f3;
	ld.param.f32 	%f5, [__cudaparmf1__Z20PremultiplyLinearize6float4+8];
	mov.f32 	%f6, %f5;
	ld.param.f32 	%f7, [__cudaparmf1__Z20PremultiplyLinearize6float4+12];
	mov.f32 	%f8, %f7;
	.loc	4	254	0
	cvt.ftz.sat.f32.f32 	%f9, %f8;
	.loc	4	255	0
	mov.f32 	%f10, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p1, %f2, %f10;
	@!%p1 bra 	$Lt_20_4098;
	.loc	4	234	0
	neg.ftz.f32 	%f11, %f2;
	lg2.approx.ftz.f32 	%f12, %f11;
	mov.f32 	%f13, 0f400ccccd;    	// 2.2
	mul.ftz.f32 	%f14, %f12, %f13;
	ex2.approx.ftz.f32 	%f15, %f14;
	neg.ftz.f32 	%f16, %f15;
	bra.uni 	$LDWendi___log2f_197_5;
$Lt_20_4098:
	.loc	4	236	0
	lg2.approx.ftz.f32 	%f17, %f2;
	mov.f32 	%f18, 0f400ccccd;    	// 2.2
	mul.ftz.f32 	%f19, %f17, %f18;
	ex2.approx.ftz.f32 	%f16, %f19;
$LDWendi___log2f_197_5:
	.loc	4	256	0
	mov.f32 	%f20, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p2, %f4, %f20;
	@!%p2 bra 	$Lt_20_4610;
	.loc	4	234	0
	neg.ftz.f32 	%f21, %f4;
	lg2.approx.ftz.f32 	%f22, %f21;
	mov.f32 	%f23, 0f400ccccd;    	// 2.2
	mul.ftz.f32 	%f24, %f22, %f23;
	ex2.approx.ftz.f32 	%f25, %f24;
	neg.ftz.f32 	%f26, %f25;
	bra.uni 	$LDWendi___log2f_197_3;
$Lt_20_4610:
	.loc	4	236	0
	lg2.approx.ftz.f32 	%f27, %f4;
	mov.f32 	%f28, 0f400ccccd;    	// 2.2
	mul.ftz.f32 	%f29, %f27, %f28;
	ex2.approx.ftz.f32 	%f26, %f29;
$LDWendi___log2f_197_3:
	.loc	4	257	0
	mov.f32 	%f30, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p3, %f6, %f30;
	@!%p3 bra 	$Lt_20_5122;
	.loc	4	234	0
	neg.ftz.f32 	%f31, %f6;
	lg2.approx.ftz.f32 	%f32, %f31;
	mov.f32 	%f33, 0f400ccccd;    	// 2.2
	mul.ftz.f32 	%f34, %f32, %f33;
	ex2.approx.ftz.f32 	%f35, %f34;
	neg.ftz.f32 	%f36, %f35;
	bra.uni 	$LDWendi___log2f_197_1;
$Lt_20_5122:
	.loc	4	236	0
	lg2.approx.ftz.f32 	%f37, %f6;
	mov.f32 	%f38, 0f400ccccd;    	// 2.2
	mul.ftz.f32 	%f39, %f37, %f38;
	ex2.approx.ftz.f32 	%f36, %f39;
$LDWendi___log2f_197_1:
	.loc	4	259	0
	mul.ftz.f32 	%f40, %f36, %f9;
	mul.ftz.f32 	%f41, %f26, %f9;
	.loc	4	278	0
	mul.ftz.f32 	%f42, %f16, %f9;
	st.param.f32 	[__cudaretf__Z20PremultiplyLinearize6float4+0], %f42;
	mov.f32 	%f43, %f41;
	st.param.f32 	[__cudaretf__Z20PremultiplyLinearize6float4+4], %f43;
	mov.f32 	%f44, %f40;
	st.param.f32 	[__cudaretf__Z20PremultiplyLinearize6float4+8], %f44;
	mov.f32 	%f45, %f9;
	st.param.f32 	[__cudaretf__Z20PremultiplyLinearize6float4+12], %f45;
	ret;
$LDWend__Z20PremultiplyLinearize6float4:
	} // _Z20PremultiplyLinearize6float4

	.visible .func (.param .align 16 .b8 __cudaretf__Z24UnpremultiplyUnlinearize6float4[16]) _Z24UnpremultiplyUnlinearize6float4 (.param .align 16 .b8 __cudaparmf1__Z24UnpremultiplyUnlinearize6float4[16])
	{
	.reg .f32 %f<53>;
	.reg .pred %p<6>;
	.loc	4	284	0
$LDWbegin__Z24UnpremultiplyUnlinearize6float4:
	ld.param.f32 	%f1, [__cudaparmf1__Z24UnpremultiplyUnlinearize6float4+0];
	mov.f32 	%f2, %f1;
	ld.param.f32 	%f3, [__cudaparmf1__Z24UnpremultiplyUnlinearize6float4+4];
	mov.f32 	%f4, %f3;
	ld.param.f32 	%f5, [__cudaparmf1__Z24UnpremultiplyUnlinearize6float4+8];
	mov.f32 	%f6, %f5;
	ld.param.f32 	%f7, [__cudaparmf1__Z24UnpremultiplyUnlinearize6float4+12];
	mov.f32 	%f8, %f7;
	.loc	4	208	0
	cvt.ftz.sat.f32.f32 	%f9, %f8;
	mov.f32 	%f10, %f9;
	mov.f32 	%f11, 0fb70637bd;    	// -8e-006
	add.ftz.f32 	%f12, %f9, %f11;
	mov.f32 	%f13, 0f00000000;    	// 0
	setp.le.ftz.f32 	%p1, %f12, %f13;
	@%p1 bra 	$Lt_21_5122;
	.loc	4	213	0
	rcp.approx.ftz.f32 	%f14, %f9;
	mul.ftz.f32 	%f15, %f14, %f6;
	.loc	4	214	0
	mul.ftz.f32 	%f16, %f14, %f4;
	.loc	4	215	0
	mul.ftz.f32 	%f17, %f14, %f2;
	bra.uni 	$Lt_21_4866;
$Lt_21_5122:
	.loc	4	219	0
	mov.f32 	%f15, 0f00000000;    	// 0
	mov.f32 	%f16, 0f00000000;    	// 0
	mov.f32 	%f17, 0f00000000;    	// 0
	mov.f32 	%f10, 0f00000000;    	// 0
$Lt_21_4866:
	.loc	4	266	0
	mov.f32 	%f18, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p2, %f17, %f18;
	@!%p2 bra 	$Lt_21_5378;
	.loc	4	242	0
	neg.ftz.f32 	%f19, %f17;
	lg2.approx.ftz.f32 	%f20, %f19;
	mov.f32 	%f21, 0f3ee8ba2e;    	// 0.454545
	mul.ftz.f32 	%f22, %f20, %f21;
	ex2.approx.ftz.f32 	%f23, %f22;
	neg.ftz.f32 	%f24, %f23;
	bra.uni 	$LDWendi___log2f_198_5;
$Lt_21_5378:
	.loc	4	244	0
	lg2.approx.ftz.f32 	%f25, %f17;
	mov.f32 	%f26, 0f3ee8ba2e;    	// 0.454545
	mul.ftz.f32 	%f27, %f25, %f26;
	ex2.approx.ftz.f32 	%f24, %f27;
$LDWendi___log2f_198_5:
	.loc	4	267	0
	mov.f32 	%f28, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p3, %f16, %f28;
	@!%p3 bra 	$Lt_21_5890;
	.loc	4	242	0
	neg.ftz.f32 	%f29, %f16;
	lg2.approx.ftz.f32 	%f30, %f29;
	mov.f32 	%f31, 0f3ee8ba2e;    	// 0.454545
	mul.ftz.f32 	%f32, %f30, %f31;
	ex2.approx.ftz.f32 	%f33, %f32;
	neg.ftz.f32 	%f34, %f33;
	bra.uni 	$LDWendi___log2f_198_3;
$Lt_21_5890:
	.loc	4	244	0
	lg2.approx.ftz.f32 	%f35, %f16;
	mov.f32 	%f36, 0f3ee8ba2e;    	// 0.454545
	mul.ftz.f32 	%f37, %f35, %f36;
	ex2.approx.ftz.f32 	%f34, %f37;
$LDWendi___log2f_198_3:
	.loc	4	268	0
	mov.f32 	%f38, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p4, %f15, %f38;
	@!%p4 bra 	$Lt_21_6402;
	.loc	4	242	0
	neg.ftz.f32 	%f39, %f15;
	lg2.approx.ftz.f32 	%f40, %f39;
	mov.f32 	%f41, 0f3ee8ba2e;    	// 0.454545
	mul.ftz.f32 	%f42, %f40, %f41;
	ex2.approx.ftz.f32 	%f43, %f42;
	neg.ftz.f32 	%f44, %f43;
	bra.uni 	$LDWendi___log2f_198_1;
$Lt_21_6402:
	.loc	4	244	0
	lg2.approx.ftz.f32 	%f45, %f15;
	mov.f32 	%f46, 0f3ee8ba2e;    	// 0.454545
	mul.ftz.f32 	%f47, %f45, %f46;
	ex2.approx.ftz.f32 	%f44, %f47;
$LDWendi___log2f_198_1:
	.loc	4	285	0
	mov.f32 	%f48, %f24;
	st.param.f32 	[__cudaretf__Z24UnpremultiplyUnlinearize6float4+0], %f48;
	mov.f32 	%f49, %f34;
	st.param.f32 	[__cudaretf__Z24UnpremultiplyUnlinearize6float4+4], %f49;
	mov.f32 	%f50, %f44;
	st.param.f32 	[__cudaretf__Z24UnpremultiplyUnlinearize6float4+8], %f50;
	mov.f32 	%f51, %f10;
	st.param.f32 	[__cudaretf__Z24UnpremultiplyUnlinearize6float4+12], %f51;
	ret;
$LDWend__Z24UnpremultiplyUnlinearize6float4:
	} // _Z24UnpremultiplyUnlinearize6float4

	.visible .func (.param .align 16 .b8 __cudaretf__Z18SwapComponentOrderI6float4ET_RKS1_[16]) _Z18SwapComponentOrderI6float4ET_RKS1_ (.param .u64 __cudaparmf1__Z18SwapComponentOrderI6float4ET_RKS1_)
	{
	.reg .u64 %rd<4>;
	.reg .f32 %f<6>;
	.loc	22	264	0
$LDWbegin__Z18SwapComponentOrderI6float4ET_RKS1_:
	ld.param.u64 	%rd1, [__cudaparmf1__Z18SwapComponentOrderI6float4ET_RKS1_];
	mov.s64 	%rd2, %rd1;
	.loc	22	270	0
	ld.f32 	%f1, [%rd2+12];
	st.param.f32 	[__cudaretf__Z18SwapComponentOrderI6float4ET_RKS1_+0], %f1;
	ld.f32 	%f2, [%rd2+8];
	st.param.f32 	[__cudaretf__Z18SwapComponentOrderI6float4ET_RKS1_+4], %f2;
	ld.f32 	%f3, [%rd2+4];
	st.param.f32 	[__cudaretf__Z18SwapComponentOrderI6float4ET_RKS1_+8], %f3;
	ld.f32 	%f4, [%rd2+0];
	st.param.f32 	[__cudaretf__Z18SwapComponentOrderI6float4ET_RKS1_+12], %f4;
	ret;
$LDWend__Z18SwapComponentOrderI6float4ET_RKS1_:
	} // _Z18SwapComponentOrderI6float4ET_RKS1_

	.visible .func (.param .f32 __cudaretf__Z23ApplyBrightnessContrastffff) _Z23ApplyBrightnessContrastffff (.param .f32 __cudaparmf1__Z23ApplyBrightnessContrastffff, .param .f32 __cudaparmf2__Z23ApplyBrightnessContrastffff, .param .f32 __cudaparmf3__Z23ApplyBrightnessContrastffff, .param .f32 __cudaparmf4__Z23ApplyBrightnessContrastffff)
	{
	.reg .f32 %f<32>;
	.reg .pred %p<6>;
	.loc	23	35	0
$LDWbegin__Z23ApplyBrightnessContrastffff:
	ld.param.f32 	%f1, [__cudaparmf1__Z23ApplyBrightnessContrastffff];
	mov.f32 	%f2, %f1;
	ld.param.f32 	%f3, [__cudaparmf2__Z23ApplyBrightnessContrastffff];
	mov.f32 	%f4, %f3;
	ld.param.f32 	%f5, [__cudaparmf3__Z23ApplyBrightnessContrastffff];
	mov.f32 	%f6, %f5;
	ld.param.f32 	%f7, [__cudaparmf4__Z23ApplyBrightnessContrastffff];
	mov.f32 	%f8, %f7;
	mov.f32 	%f9, 0f00000000;     	// 0
	setp.ge.ftz.f32 	%p1, %f6, %f9;
	@!%p1 bra 	$Lt_23_5378;
	mov.f32 	%f10, 0f3f800000;    	// 1
	sub.ftz.f32 	%f11, %f10, %f6;
	sub.ftz.f32 	%f12, %f8, %f4;
	setp.le.ftz.f32 	%p2, %f12, %f2;
	mov.f32 	%f13, 0f00000000;    	// 0
	setp.eq.ftz.f32 	%p3, %f11, %f13;
	@!%p3 bra 	$Lt_23_5890;
	.loc	23	46	0
	mov.f32 	%f14, 0f3f800000;    	// 1
	mov.f32 	%f15, 0f00000000;    	// 0
	selp.f32 	%f16, %f14, %f15, %p2;
	bra.uni 	$Lt_23_6658;
$Lt_23_5890:
	@!%p2 bra 	$Lt_23_6402;
	.loc	23	59	0
	sub.ftz.f32 	%f17, %f2, %f12;
	div.approx.ftz.f32 	%f18, %f17, %f11;
	add.ftz.f32 	%f16, %f8, %f18;
	bra.uni 	$Lt_23_6658;
$Lt_23_6402:
	.loc	23	63	0
	sub.ftz.f32 	%f19, %f12, %f2;
	div.approx.ftz.f32 	%f20, %f19, %f11;
	sub.ftz.f32 	%f16, %f8, %f20;
	bra.uni 	$Lt_23_6658;
$Lt_23_5378:
	add.ftz.f32 	%f21, %f4, %f8;
	mov.f32 	%f22, 0f3f800000;    	// 1
	add.ftz.f32 	%f23, %f6, %f22;
	setp.ge.ftz.f32 	%p4, %f2, %f8;
	@!%p4 bra 	$Lt_23_6914;
	.loc	23	76	0
	sub.ftz.f32 	%f24, %f2, %f8;
	fma.rn.ftz.f32 	%f16, %f23, %f24, %f21;
	bra.uni 	$Lt_23_6658;
$Lt_23_6914:
	.loc	23	80	0
	sub.ftz.f32 	%f25, %f8, %f2;
	mul.ftz.f32 	%f26, %f23, %f25;
	sub.ftz.f32 	%f16, %f21, %f26;
$Lt_23_6658:
$Lt_23_5122:
	.loc	23	84	0
	mov.f32 	%f27, 0f00000000;    	// 0
	max.ftz.f32 	%f28, %f16, %f27;
	mov.f32 	%f29, 0f3f800000;    	// 1
	min.ftz.f32 	%f30, %f28, %f29;
	st.param.f32 	[__cudaretf__Z23ApplyBrightnessContrastffff], %f30;
	ret;
$LDWend__Z23ApplyBrightnessContrastffff:
	} // _Z23ApplyBrightnessContrastffff

	.visible .func (.param .f32 __cudaretf__Z5ClampIfET_S0_S0_S0_) _Z5ClampIfET_S0_S0_S0_ (.param .f32 __cudaparmf1__Z5ClampIfET_S0_S0_S0_, .param .f32 __cudaparmf2__Z5ClampIfET_S0_S0_S0_, .param .f32 __cudaparmf3__Z5ClampIfET_S0_S0_S0_)
	{
	.reg .f32 %f<10>;
	.loc	24	72	0
$LDWbegin__Z5ClampIfET_S0_S0_S0_:
	ld.param.f32 	%f1, [__cudaparmf1__Z5ClampIfET_S0_S0_S0_];
	mov.f32 	%f2, %f1;
	ld.param.f32 	%f3, [__cudaparmf2__Z5ClampIfET_S0_S0_S0_];
	mov.f32 	%f4, %f3;
	ld.param.f32 	%f5, [__cudaparmf3__Z5ClampIfET_S0_S0_S0_];
	mov.f32 	%f6, %f5;
	.loc	24	73	0
	max.ftz.f32 	%f7, %f2, %f4;
	min.ftz.f32 	%f8, %f6, %f7;
	st.param.f32 	[__cudaretf__Z5ClampIfET_S0_S0_S0_], %f8;
	ret;
$LDWend__Z5ClampIfET_S0_S0_S0_:
	} // _Z5ClampIfET_S0_S0_S0_

	.visible .func (.param .f32 __cudaretf__Z13CalculateLumaffRK34LumaCorrectorParameters_TonalRangeRK30LumaCorrectorParameters_Levels) _Z13CalculateLumaffRK34LumaCorrectorParameters_TonalRangeRK30LumaCorrectorParameters_Levels (.param .f32 __cudaparmf1__Z13CalculateLumaffRK34LumaCorrectorParameters_TonalRangeRK30LumaCorrectorParameters_Levels, .param .f32 __cudaparmf2__Z13CalculateLumaffRK34LumaCorrectorParameters_TonalRangeRK30LumaCorrectorParameters_Levels, .param .u64 __cudaparmf3__Z13CalculateLumaffRK34LumaCorrectorParameters_TonalRangeRK30LumaCorrectorParameters_Levels, .param .u64 __cudaparmf4__Z13CalculateLumaffRK34LumaCorrectorParameters_TonalRangeRK30LumaCorrectorParameters_Levels)
	{
	.reg .u64 %rd<6>;
	.reg .f32 %f<168>;
	.reg .pred %p<17>;
	.loc	23	92	0
$LDWbegin__Z13CalculateLumaffRK34LumaCorrectorParameters_TonalRangeRK30LumaCorrectorParameters_Levels:
	ld.param.f32 	%f1, [__cudaparmf1__Z13CalculateLumaffRK34LumaCorrectorParameters_TonalRangeRK30LumaCorrectorParameters_Levels];
	mov.f32 	%f2, %f1;
	ld.param.f32 	%f3, [__cudaparmf2__Z13CalculateLumaffRK34LumaCorrectorParameters_TonalRangeRK30LumaCorrectorParameters_Levels];
	mov.f32 	%f4, %f3;
	ld.param.u64 	%rd1, [__cudaparmf3__Z13CalculateLumaffRK34LumaCorrectorParameters_TonalRangeRK30LumaCorrectorParameters_Levels];
	mov.s64 	%rd2, %rd1;
	ld.param.u64 	%rd3, [__cudaparmf4__Z13CalculateLumaffRK34LumaCorrectorParameters_TonalRangeRK30LumaCorrectorParameters_Levels];
	mov.s64 	%rd4, %rd3;
	.loc	23	93	0
	mov.f32 	%f5, %f2;
	mov.f32 	%f6, 0f3a83126f;     	// 0.001
	setp.gt.ftz.f32 	%p1, %f4, %f6;
	@!%p1 bra 	$Lt_25_16642;
	.loc	23	99	0
	ld.f32 	%f7, [%rd2+4];
	ld.f32 	%f8, [%rd2+12];
	mov.f32 	%f9, 0f40000000;     	// 2
	mov.f32 	%f10, 0f3f800000;    	// 1
	setp.lt.ftz.f32 	%p2, %f8, %f7;
	selp.f32 	%f11, %f9, %f10, %p2;
	.loc	23	100	0
	ld.f32 	%f12, [%rd2+0];
	.loc	20	35	0
	setp.le.ftz.f32 	%p3, %f2, %f12;
	@!%p3 bra 	$Lt_25_12034;
	.loc	20	37	0
	mov.f32 	%f13, 0f3f800000;    	// 1
	bra.uni 	$LDWendi_fdividef_202_3;
$Lt_25_12034:
	.loc	20	39	0
	add.ftz.f32 	%f14, %f7, %f12;
	setp.ge.ftz.f32 	%p4, %f2, %f14;
	@!%p4 bra 	$Lt_25_12290;
	.loc	20	41	0
	mov.f32 	%f13, 0f00000000;    	// 0
	bra.uni 	$LDWendi_fdividef_202_3;
$Lt_25_12290:
	.loc	20	45	0
	sub.ftz.f32 	%f15, %f2, %f12;
	div.approx.ftz.f32 	%f13, %f15, %f7;
$LDWendi_fdividef_202_3:
	.loc	23	101	0
	ld.f32 	%f16, [%rd2+8];
	.loc	20	53	0
	sub.ftz.f32 	%f17, %f16, %f8;
	setp.gt.ftz.f32 	%p5, %f17, %f2;
	@!%p5 bra 	$Lt_25_12546;
	.loc	20	55	0
	mov.f32 	%f18, 0f00000000;    	// 0
	bra.uni 	$LDWendi_fdividef_202_1;
$Lt_25_12546:
	.loc	20	57	0
	setp.gt.ftz.f32 	%p6, %f2, %f16;
	@!%p6 bra 	$Lt_25_12802;
	.loc	20	59	0
	mov.f32 	%f18, 0f3f800000;    	// 1
	bra.uni 	$LDWendi_fdividef_202_1;
$Lt_25_12802:
	.loc	20	63	0
	sub.ftz.f32 	%f19, %f2, %f17;
	div.approx.ftz.f32 	%f18, %f19, %f8;
$LDWendi_fdividef_202_1:
	.loc	23	112	0
	div.approx.ftz.f32 	%f20, %f13, %f11;
	sub.ftz.f32 	%f21, %f11, %f13;
	div.approx.ftz.f32 	%f22, %f18, %f11;
	sub.ftz.f32 	%f23, %f21, %f18;
	div.approx.ftz.f32 	%f24, %f23, %f11;
	ld.f32 	%f25, [%rd4+72];
	mul.ftz.f32 	%f26, %f25, %f20;
	mul.ftz.f32 	%f27, %f4, %f26;
	ld.f32 	%f28, [%rd4+48];
	mul.ftz.f32 	%f29, %f28, %f24;
	fma.rn.ftz.f32 	%f30, %f4, %f29, %f27;
	ld.f32 	%f31, [%rd4+24];
	mul.ftz.f32 	%f32, %f31, %f22;
	fma.rn.ftz.f32 	%f33, %f4, %f32, %f30;
	.loc	23	117	0
	ld.f32 	%f34, [%rd4+76];
	mul.ftz.f32 	%f35, %f34, %f20;
	mul.ftz.f32 	%f36, %f4, %f35;
	ld.f32 	%f37, [%rd4+52];
	mul.ftz.f32 	%f38, %f37, %f24;
	fma.rn.ftz.f32 	%f39, %f4, %f38, %f36;
	ld.f32 	%f40, [%rd4+28];
	mul.ftz.f32 	%f41, %f40, %f22;
	fma.rn.ftz.f32 	%f42, %f4, %f41, %f39;
	.loc	23	122	0
	ld.f32 	%f43, [%rd4+80];
	mul.ftz.f32 	%f44, %f43, %f20;
	ld.f32 	%f45, [%rd4+56];
	fma.rn.ftz.f32 	%f46, %f45, %f24, %f44;
	ld.f32 	%f47, [%rd4+32];
	fma.rn.ftz.f32 	%f48, %f47, %f22, %f46;
	.loc	23	128	0
	ld.f32 	%f49, [%rd4+84];
	mul.ftz.f32 	%f50, %f49, %f20;
	ld.f32 	%f51, [%rd4+60];
	fma.rn.ftz.f32 	%f52, %f51, %f24, %f50;
	ld.f32 	%f53, [%rd4+36];
	fma.rn.ftz.f32 	%f54, %f53, %f22, %f52;
	.loc	23	134	0
	ld.f32 	%f55, [%rd4+88];
	mul.ftz.f32 	%f56, %f55, %f20;
	mul.ftz.f32 	%f57, %f4, %f56;
	ld.f32 	%f58, [%rd4+64];
	mul.ftz.f32 	%f59, %f58, %f24;
	fma.rn.ftz.f32 	%f60, %f4, %f59, %f57;
	ld.f32 	%f61, [%rd4+40];
	mul.ftz.f32 	%f62, %f61, %f22;
	fma.rn.ftz.f32 	%f63, %f4, %f62, %f60;
	.loc	23	139	0
	ld.f32 	%f64, [%rd4+92];
	mul.ftz.f32 	%f65, %f64, %f20;
	ld.f32 	%f66, [%rd4+68];
	fma.rn.ftz.f32 	%f67, %f66, %f24, %f65;
	ld.f32 	%f68, [%rd4+44];
	fma.rn.ftz.f32 	%f69, %f68, %f22, %f67;
	.loc	23	148	0
	mov.f32 	%f70, 0fbf000000;    	// -0.5
	add.ftz.f32 	%f71, %f48, %f70;
	mov.f32 	%f72, 0f3f000000;    	// 0.5
	fma.rn.ftz.f32 	%f73, %f4, %f71, %f72;
	mov.f32 	%f74, 0f00000000;    	// 0
	setp.ge.ftz.f32 	%p7, %f42, %f74;
	@!%p7 bra 	$Lt_25_17410;
	mov.f32 	%f75, 0f3f800000;    	// 1
	sub.ftz.f32 	%f76, %f75, %f42;
	sub.ftz.f32 	%f77, %f73, %f33;
	setp.le.ftz.f32 	%p8, %f77, %f2;
	mov.f32 	%f78, 0f00000000;    	// 0
	setp.eq.ftz.f32 	%p9, %f76, %f78;
	@!%p9 bra 	$Lt_25_17922;
	.loc	23	46	0
	mov.f32 	%f79, 0f3f800000;    	// 1
	mov.f32 	%f80, 0f00000000;    	// 0
	selp.f32 	%f81, %f79, %f80, %p8;
	bra.uni 	$Lt_25_18690;
$Lt_25_17922:
	@!%p8 bra 	$Lt_25_18434;
	.loc	23	59	0
	sub.ftz.f32 	%f82, %f2, %f77;
	div.approx.ftz.f32 	%f83, %f82, %f76;
	add.ftz.f32 	%f81, %f73, %f83;
	bra.uni 	$Lt_25_18690;
$Lt_25_18434:
	.loc	23	63	0
	sub.ftz.f32 	%f84, %f77, %f2;
	div.approx.ftz.f32 	%f85, %f84, %f76;
	sub.ftz.f32 	%f81, %f73, %f85;
	bra.uni 	$Lt_25_18690;
$Lt_25_17410:
	mov.f32 	%f86, 0f3f800000;    	// 1
	add.ftz.f32 	%f87, %f42, %f86;
	add.ftz.f32 	%f88, %f73, %f33;
	setp.le.ftz.f32 	%p10, %f73, %f2;
	@!%p10 bra 	$Lt_25_18946;
	.loc	23	76	0
	sub.ftz.f32 	%f89, %f2, %f73;
	fma.rn.ftz.f32 	%f81, %f87, %f89, %f88;
	bra.uni 	$Lt_25_18690;
$Lt_25_18946:
	.loc	23	80	0
	sub.ftz.f32 	%f90, %f73, %f2;
	mul.ftz.f32 	%f91, %f87, %f90;
	sub.ftz.f32 	%f81, %f88, %f91;
$Lt_25_18690:
$Lt_25_17154:
	.loc	23	166	0
	ld.f32 	%f92, [%rd4+8];
	.loc	23	167	0
	ld.f32 	%f93, [%rd4+0];
	mul.ftz.f32 	%f94, %f93, %f4;
	.loc	23	168	0
	ld.f32 	%f95, [%rd4+4];
	mul.ftz.f32 	%f96, %f95, %f4;
	.loc	23	169	0
	mov.f32 	%f97, 0fbf000000;    	// -0.5
	add.ftz.f32 	%f98, %f92, %f97;
	mov.f32 	%f99, 0f00000000;    	// 0
	max.ftz.f32 	%f100, %f81, %f99;
	mov.f32 	%f101, 0fbf800000;   	// -1
	add.ftz.f32 	%f102, %f54, %f101;
	mov.f32 	%f103, 0fbf800000;   	// -1
	add.ftz.f32 	%f104, %f69, %f103;
	mov.f32 	%f105, 0f3f000000;   	// 0.5
	fma.rn.ftz.f32 	%f106, %f4, %f98, %f105;
	mov.f32 	%f107, 0f3f800000;   	// 1
	min.ftz.f32 	%f108, %f100, %f107;
	mov.f32 	%f109, 0f3f800000;   	// 1
	fma.rn.ftz.f32 	%f110, %f4, %f102, %f109;
	mov.f32 	%f111, 0f3f800000;   	// 1
	fma.rn.ftz.f32 	%f112, %f4, %f104, %f111;
	rcp.approx.ftz.f32 	%f113, %f110;
	mov.f32 	%f114, 0f00000000;   	// 0
	setp.ge.ftz.f32 	%p11, %f96, %f114;
	@!%p11 bra 	$Lt_25_19458;
	mov.f32 	%f115, 0f3f800000;   	// 1
	sub.ftz.f32 	%f116, %f115, %f96;
	sub.ftz.f32 	%f117, %f106, %f94;
	mov.f32 	%f118, 0f00000000;   	// 0
	setp.eq.ftz.f32 	%p12, %f116, %f118;
	@!%p12 bra 	$Lt_25_19970;
	.loc	23	46	0
	mov.f32 	%f119, 0f3f800000;   	// 1
	mov.f32 	%f120, 0f00000000;   	// 0
	lg2.approx.ftz.f32 	%f121, %f108;
	mul.ftz.f32 	%f122, %f113, %f121;
	ex2.approx.ftz.f32 	%f123, %f122;
	add.ftz.f32 	%f124, %f63, %f123;
	mul.ftz.f32 	%f125, %f112, %f124;
	setp.le.ftz.f32 	%p13, %f117, %f125;
	selp.f32 	%f81, %f119, %f120, %p13;
	bra.uni 	$Lt_25_20738;
$Lt_25_19970:
	lg2.approx.ftz.f32 	%f126, %f108;
	mul.ftz.f32 	%f127, %f113, %f126;
	ex2.approx.ftz.f32 	%f128, %f127;
	add.ftz.f32 	%f129, %f128, %f63;
	mul.ftz.f32 	%f130, %f112, %f129;
	setp.le.ftz.f32 	%p14, %f117, %f130;
	@!%p14 bra 	$Lt_25_20482;
	.loc	28	529	0
	sub.ftz.f32 	%f131, %f130, %f117;
	div.approx.ftz.f32 	%f132, %f131, %f116;
	.loc	23	59	0
	add.ftz.f32 	%f81, %f132, %f106;
	bra.uni 	$Lt_25_20738;
$Lt_25_20482:
	.loc	28	529	0
	sub.ftz.f32 	%f133, %f117, %f130;
	div.approx.ftz.f32 	%f134, %f133, %f116;
	.loc	23	63	0
	sub.ftz.f32 	%f81, %f106, %f134;
	bra.uni 	$Lt_25_20738;
$Lt_25_19458:
	mov.f32 	%f135, 0f3f800000;   	// 1
	add.ftz.f32 	%f136, %f96, %f135;
	add.ftz.f32 	%f137, %f106, %f94;
	lg2.approx.ftz.f32 	%f138, %f108;
	mul.ftz.f32 	%f139, %f113, %f138;
	ex2.approx.ftz.f32 	%f140, %f139;
	add.ftz.f32 	%f141, %f140, %f63;
	mul.ftz.f32 	%f142, %f112, %f141;
	setp.le.ftz.f32 	%p15, %f106, %f142;
	@!%p15 bra 	$Lt_25_20994;
	.loc	23	76	0
	sub.ftz.f32 	%f143, %f142, %f106;
	fma.rn.ftz.f32 	%f81, %f136, %f143, %f137;
	bra.uni 	$Lt_25_20738;
$Lt_25_20994:
	.loc	23	80	0
	sub.ftz.f32 	%f144, %f106, %f142;
	mul.ftz.f32 	%f145, %f136, %f144;
	sub.ftz.f32 	%f81, %f137, %f145;
$Lt_25_20738:
$Lt_25_19202:
	.loc	23	176	0
	ld.f32 	%f146, [%rd4+12];
	.loc	23	184	0
	mov.f32 	%f147, 0f3f800000;   	// 1
	ld.f32 	%f148, [%rd4+20];
	mov.f32 	%f149, 0fbf800000;   	// -1
	add.ftz.f32 	%f150, %f148, %f149;
	fma.rn.ftz.f32 	%f151, %f4, %f150, %f147;
	mov.f32 	%f152, 0f00000000;   	// 0
	max.ftz.f32 	%f153, %f81, %f152;
	mov.f32 	%f154, 0f3f800000;   	// 1
	min.ftz.f32 	%f155, %f153, %f154;
	lg2.approx.ftz.f32 	%f156, %f155;
	mov.f32 	%f157, 0f3f800000;   	// 1
	mov.f32 	%f158, 0fbf800000;   	// -1
	add.ftz.f32 	%f159, %f146, %f158;
	fma.rn.ftz.f32 	%f160, %f4, %f159, %f157;
	rcp.approx.ftz.f32 	%f161, %f160;
	mul.ftz.f32 	%f162, %f156, %f161;
	ex2.approx.ftz.f32 	%f163, %f162;
	ld.f32 	%f164, [%rd4+16];
	fma.rn.ftz.f32 	%f165, %f164, %f4, %f163;
	mul.ftz.f32 	%f5, %f151, %f165;
$Lt_25_16642:
	.loc	23	187	0
	mov.f32 	%f166, %f5;
	st.param.f32 	[__cudaretf__Z13CalculateLumaffRK34LumaCorrectorParameters_TonalRangeRK30LumaCorrectorParameters_Levels], %f166;
	ret;
$LDWend__Z13CalculateLumaffRK34LumaCorrectorParameters_TonalRangeRK30LumaCorrectorParameters_Levels:
	} // _Z13CalculateLumaffRK34LumaCorrectorParameters_TonalRangeRK30LumaCorrectorParameters_Levels

	.visible .func (.param .f32 __cudaretf__Z4LERPIfET_S0_S0_S0_) _Z4LERPIfET_S0_S0_S0_ (.param .f32 __cudaparmf1__Z4LERPIfET_S0_S0_S0_, .param .f32 __cudaparmf2__Z4LERPIfET_S0_S0_S0_, .param .f32 __cudaparmf3__Z4LERPIfET_S0_S0_S0_)
	{
	.reg .f32 %f<10>;
	.loc	24	78	0
$LDWbegin__Z4LERPIfET_S0_S0_S0_:
	ld.param.f32 	%f1, [__cudaparmf1__Z4LERPIfET_S0_S0_S0_];
	mov.f32 	%f2, %f1;
	ld.param.f32 	%f3, [__cudaparmf2__Z4LERPIfET_S0_S0_S0_];
	mov.f32 	%f4, %f3;
	ld.param.f32 	%f5, [__cudaparmf3__Z4LERPIfET_S0_S0_S0_];
	mov.f32 	%f6, %f5;
	.loc	24	79	0
	sub.ftz.f32 	%f7, %f4, %f2;
	fma.rn.ftz.f32 	%f8, %f6, %f7, %f2;
	st.param.f32 	[__cudaretf__Z4LERPIfET_S0_S0_S0_], %f8;
	ret;
$LDWend__Z4LERPIfET_S0_S0_S0_:
	} // _Z4LERPIfET_S0_S0_S0_

	.visible .func (.param .f32 __cudaretf__Z6Read2DIfET_PKS0_iii) _Z6Read2DIfET_PKS0_iii (.param .u64 __cudaparmf1__Z6Read2DIfET_PKS0_iii, .param .s32 __cudaparmf2__Z6Read2DIfET_PKS0_iii, .param .s32 __cudaparmf3__Z6Read2DIfET_PKS0_iii, .param .s32 __cudaparmf4__Z6Read2DIfET_PKS0_iii)
	{
	.reg .u32 %r<10>;
	.reg .u64 %rd<7>;
	.reg .f32 %f<3>;
	.loc	21	114	0
$LDWbegin__Z6Read2DIfET_PKS0_iii:
	ld.param.u64 	%rd1, [__cudaparmf1__Z6Read2DIfET_PKS0_iii];
	mov.s64 	%rd2, %rd1;
	ld.param.u32 	%r1, [__cudaparmf2__Z6Read2DIfET_PKS0_iii];
	mov.s32 	%r2, %r1;
	ld.param.u32 	%r3, [__cudaparmf3__Z6Read2DIfET_PKS0_iii];
	mov.s32 	%r4, %r3;
	ld.param.u32 	%r5, [__cudaparmf4__Z6Read2DIfET_PKS0_iii];
	mov.s32 	%r6, %r5;
	.loc	21	115	0
	mul.lo.s32 	%r7, %r2, %r6;
	add.s32 	%r8, %r4, %r7;
	cvt.s64.s32 	%rd3, %r8;
	mul.wide.s32 	%rd4, %r8, 4;
	add.u64 	%rd5, %rd2, %rd4;
	ld.f32 	%f1, [%rd5+0];
	st.param.f32 	[__cudaretf__Z6Read2DIfET_PKS0_iii], %f1;
	ret;
$LDWend__Z6Read2DIfET_PKS0_iii:
	} // _Z6Read2DIfET_PKS0_iii
	.const .align 4 .b8 k601YPbPr_To_RGB32f[36] = {0,0,128,63,0,0,0,0,188,116,179,63,0,0,128,63,152,50,176,190,158,209,54,191,0,0,128,63,229,208,226,63,0,0,0,0};

	.entry LumaCorrector_MaskKernel (
		.param .u64 __cudaparm_LumaCorrector_MaskKernel_inImage,
		.param .s32 __cudaparm_LumaCorrector_MaskKernel_inPitch,
		.param .u64 __cudaparm_LumaCorrector_MaskKernel_inSecondaryMask,
		.param .s32 __cudaparm_LumaCorrector_MaskKernel_inSecondaryPitch,
		.param .u32 __cudaparm_LumaCorrector_MaskKernel_inDeviceFormat,
		.param .s32 __cudaparm_LumaCorrector_MaskKernel_inWidth,
		.param .s32 __cudaparm_LumaCorrector_MaskKernel_inHeight)
	{
	.reg .u32 %r<32>;
	.reg .u64 %rd<13>;
	.reg .f32 %f<11>;
	.reg .pred %p<5>;
	.loc	23	200	0
$LDWbegin_LumaCorrector_MaskKernel:
	.loc	23	203	0
	cvt.s32.u32 	%r1, %ctaid.x;
	cvt.s32.u32 	%r2, %ntid.x;
	mul.lo.s32 	%r3, %r1, %r2;
	cvt.s32.u32 	%r4, %ctaid.y;
	cvt.s32.u32 	%r5, %ntid.y;
	mul.lo.s32 	%r6, %r4, %r5;
	mov.u32 	%r7, %tid.x;
	add.u32 	%r8, %r3, %r7;
	mov.u32 	%r9, %tid.y;
	add.u32 	%r10, %r6, %r9;
	ld.param.s32 	%r11, [__cudaparm_LumaCorrector_MaskKernel_inWidth];
	set.gt.u32.s32 	%r12, %r11, %r8;
	neg.s32 	%r13, %r12;
	ld.param.s32 	%r14, [__cudaparm_LumaCorrector_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_28_3586;
	.loc	23	206	0
	ld.param.u64 	%rd1, [__cudaparm_LumaCorrector_MaskKernel_inSecondaryMask];
	mov.u64 	%rd2, 0;
	setp.eq.u64 	%p2, %rd1, %rd2;
	@%p2 bra 	$Lt_28_4866;
	.loc	23	208	0
	ld.param.s32 	%r19, [__cudaparm_LumaCorrector_MaskKernel_inSecondaryPitch];
	mul.lo.s32 	%r20, %r19, %r10;
	add.s32 	%r21, %r8, %r20;
	cvt.s64.s32 	%rd3, %r21;
	mul.wide.s32 	%rd4, %r21, 4;
	add.u64 	%rd5, %rd1, %rd4;
	ld.global.f32 	%f1, [%rd5+0];
	bra.uni 	$Lt_28_4610;
$Lt_28_4866:
	mov.f32 	%f1, 0f3f800000;     	// 1
$Lt_28_4610:
	.loc	23	213	0
	ld.const.f32 	%f2, [k601YPbPr_To_RGB32f+24];
	mul.ftz.f32 	%f3, %f2, %f1;
	ld.const.f32 	%f4, [k601YPbPr_To_RGB32f+12];
	mul.ftz.f32 	%f5, %f4, %f1;
	ld.const.f32 	%f6, [k601YPbPr_To_RGB32f+0];
	mul.ftz.f32 	%f7, %f6, %f1;
	ld.param.s32 	%r22, [__cudaparm_LumaCorrector_MaskKernel_inPitch];
	mul.lo.s32 	%r23, %r22, %r10;
	add.s32 	%r24, %r8, %r23;
	cvt.s64.s32 	%rd6, %r24;
	ld.param.u64 	%rd7, [__cudaparm_LumaCorrector_MaskKernel_inImage];
	ld.param.s32 	%r25, [__cudaparm_LumaCorrector_MaskKernel_inDeviceFormat];
	mov.u32 	%r26, 0;
	setp.ne.s32 	%p3, %r25, %r26;
	@%p3 bra 	$Lt_28_5378;
	.loc	21	126	0
	mul.lo.u64 	%rd8, %rd6, 8;
	add.u64 	%rd9, %rd7, %rd8;
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f3;
	mov.b32		%r27, %b1; }
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f5;
	mov.b32		%r28, %b1; }
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f7;
	mov.b32		%r29, %b1; }
	mov.f32 	%f8, 0f3f800000;     	// 1
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f8;
	mov.b32		%r30, %b1; }
	st.global.v4.u16 	[%rd9+0], {%r27,%r28,%r29,%r30};
	.loc	23	213	0
	bra.uni 	$Lt_28_5122;
$Lt_28_5378:
	.loc	21	126	0
	mul.lo.u64 	%rd10, %rd6, 16;
	add.u64 	%rd11, %rd7, %rd10;
	mov.f32 	%f9, 0f3f800000;     	// 1
	st.global.v4.f32 	[%rd11+0], {%f3,%f5,%f7,%f9};
$Lt_28_5122:
$Lt_28_3586:
	.loc	23	215	0
	exit;
$LDWend_LumaCorrector_MaskKernel:
	} // LumaCorrector_MaskKernel
	.const .align 4 .b8 kRGB32f_To_601YPbPr[36] = {135,22,153,62,162,69,22,63,213,120,233,61,33,201,44,190,111,155,169,190,0,0,0,63,0,0,0,63,70,94,214,190,232,134,166,189};

	.entry LumaCorrector_TonalRangeKernel (
		.param .u64 __cudaparm_LumaCorrector_TonalRangeKernel_inImage,
		.param .s32 __cudaparm_LumaCorrector_TonalRangeKernel_inPitch,
		.param .u64 __cudaparm_LumaCorrector_TonalRangeKernel_inSecondaryMask,
		.param .s32 __cudaparm_LumaCorrector_TonalRangeKernel_inSecondaryPitch,
		.param .u32 __cudaparm_LumaCorrector_TonalRangeKernel_inDeviceFormat,
		.param .s32 __cudaparm_LumaCorrector_TonalRangeKernel_inWidth,
		.param .s32 __cudaparm_LumaCorrector_TonalRangeKernel_inHeight,
		.param .align 4 .b8 __cudaparm_LumaCorrector_TonalRangeKernel_inTonalRangeParameters[16])
	{
	.reg .u32 %r<33>;
	.reg .u64 %rd<12>;
	.reg .f32 %f<41>;
	.reg .pred %p<12>;
	.loc	23	228	0
$LDWbegin_LumaCorrector_TonalRangeKernel:
	.loc	23	231	0
	cvt.s32.u32 	%r1, %ctaid.x;
	cvt.s32.u32 	%r2, %ntid.x;
	mul.lo.s32 	%r3, %r1, %r2;
	cvt.s32.u32 	%r4, %ctaid.y;
	cvt.s32.u32 	%r5, %ntid.y;
	mul.lo.s32 	%r6, %r4, %r5;
	mov.u32 	%r7, %tid.x;
	add.u32 	%r8, %r3, %r7;
	mov.u32 	%r9, %tid.y;
	add.u32 	%r10, %r6, %r9;
	ld.param.s32 	%r11, [__cudaparm_LumaCorrector_TonalRangeKernel_inWidth];
	set.gt.u32.s32 	%r12, %r11, %r8;
	neg.s32 	%r13, %r12;
	ld.param.s32 	%r14, [__cudaparm_LumaCorrector_TonalRangeKernel_inHeight];
	set.gt.u32.s32 	%r15, %r14, %r10;
	neg.s32 	%r16, %r15;
	and.b32 	%r17, %r13, %r16;
	mov.u32 	%r18, 0;
	setp.eq.s32 	%p1, %r17, %r18;
	@%p1 bra 	$Lt_29_11010;
	ld.param.s32 	%r19, [__cudaparm_LumaCorrector_TonalRangeKernel_inDeviceFormat];
	mov.s32 	%r20, 0;
	setp.eq.s32 	%p2, %r19, %r20;
	ld.param.s32 	%r21, [__cudaparm_LumaCorrector_TonalRangeKernel_inPitch];
	mul.lo.s32 	%r22, %r21, %r10;
	add.s32 	%r23, %r8, %r22;
	cvt.s64.s32 	%rd1, %r23;
	ld.param.u64 	%rd2, [__cudaparm_LumaCorrector_TonalRangeKernel_inImage];
	@!%p2 bra 	$Lt_29_11778;
	.loc	21	115	0
	mul.lo.u64 	%rd3, %rd1, 8;
	add.u64 	%rd4, %rd2, %rd3;
	ld.global.v4.u16 	{%r24,%r25,%r26,%r27}, [%rd4+0];
	.loc	23	234	0
	{ .reg .b32 %b1;
	mov.b32		%b1, %r24;
	cvt.ftz.f32.f16	%f1, %b1; }
	{ .reg .b32 %b1;
	mov.b32		%b1, %r25;
	cvt.ftz.f32.f16	%f2, %b1; }
	{ .reg .b32 %b1;
	mov.b32		%b1, %r26;
	cvt.ftz.f32.f16	%f3, %b1; }
	{ .reg .b32 %b1;
	mov.b32		%b1, %r27;
	cvt.ftz.f32.f16	%f4, %b1; }
	bra.uni 	$Lt_29_11522;
$Lt_29_11778:
	mul.lo.u64 	%rd5, %rd1, 16;
	add.u64 	%rd6, %rd2, %rd5;
	ld.global.v4.f32 	{%f1,%f2,%f3,%f4}, [%rd6+0];
$Lt_29_11522:
	.loc	20	35	0
	ld.const.f32 	%f5, [kRGB32f_To_601YPbPr+4];
	mul.ftz.f32 	%f6, %f5, %f2;
	ld.const.f32 	%f7, [kRGB32f_To_601YPbPr+0];
	fma.rn.ftz.f32 	%f8, %f7, %f3, %f6;
	ld.const.f32 	%f9, [kRGB32f_To_601YPbPr+8];
	fma.rn.ftz.f32 	%f10, %f9, %f1, %f8;
	ld.param.f32 	%f11, [__cudaparm_LumaCorrector_TonalRangeKernel_inTonalRangeParameters+4];
	ld.param.f32 	%f12, [__cudaparm_LumaCorrector_TonalRangeKernel_inTonalRangeParameters+0];
	setp.le.ftz.f32 	%p3, %f10, %f12;
	@!%p3 bra 	$Lt_29_7938;
	.loc	20	37	0
	mov.f32 	%f13, 0f3f800000;    	// 1
	bra.uni 	$LDWendi_fdividef_206_3;
$Lt_29_7938:
	.loc	20	39	0
	add.ftz.f32 	%f14, %f11, %f12;
	setp.ge.ftz.f32 	%p4, %f10, %f14;
	@!%p4 bra 	$Lt_29_8194;
	.loc	20	41	0
	mov.f32 	%f13, 0f00000000;    	// 0
	bra.uni 	$LDWendi_fdividef_206_3;
$Lt_29_8194:
	.loc	20	45	0
	sub.ftz.f32 	%f15, %f10, %f12;
	div.approx.ftz.f32 	%f13, %f15, %f11;
$LDWendi_fdividef_206_3:
	.loc	20	53	0
	ld.param.f32 	%f16, [__cudaparm_LumaCorrector_TonalRangeKernel_inTonalRangeParameters+12];
	ld.param.f32 	%f17, [__cudaparm_LumaCorrector_TonalRangeKernel_inTonalRangeParameters+8];
	sub.ftz.f32 	%f18, %f17, %f16;
	setp.gt.ftz.f32 	%p5, %f18, %f10;
	@!%p5 bra 	$Lt_29_8450;
	.loc	20	55	0
	mov.f32 	%f19, 0f00000000;    	// 0
	bra.uni 	$LDWendi_fdividef_206_1;
$Lt_29_8450:
	.loc	20	57	0
	setp.gt.ftz.f32 	%p6, %f10, %f17;
	@!%p6 bra 	$Lt_29_8706;
	.loc	20	59	0
	mov.f32 	%f19, 0f3f800000;    	// 1
	bra.uni 	$LDWendi_fdividef_206_1;
$Lt_29_8706:
	.loc	20	63	0
	sub.ftz.f32 	%f20, %f10, %f18;
	div.approx.ftz.f32 	%f19, %f20, %f16;
$LDWendi_fdividef_206_1:
	.loc	23	243	0
	setp.lt.ftz.f32 	%p7, %f16, %f11;
	mov.f32 	%f21, 0f40000000;    	// 2
	mov.f32 	%f22, 0f3f800000;    	// 1
	selp.f32 	%f23, %f21, %f22, %p7;
	div.approx.ftz.f32 	%f24, %f13, %f23;
	mov.f32 	%f25, 0f3f800000;    	// 1
	setp.eq.ftz.f32 	%p8, %f24, %f25;
	@!%p8 bra 	$Lt_29_12290;
	mov.f32 	%f26, 0f00000000;    	// 0
	bra.uni 	$Lt_29_12034;
$Lt_29_12290:
	sub.ftz.f32 	%f27, %f23, %f13;
	sub.ftz.f32 	%f28, %f27, %f19;
	div.approx.ftz.f32 	%f29, %f28, %f23;
	mov.f32 	%f30, 0f3f800000;    	// 1
	setp.eq.ftz.f32 	%p9, %f29, %f30;
	@!%p9 bra 	$Lt_29_12802;
	mov.f32 	%f26, 0f3f000000;    	// 0.5
	bra.uni 	$Lt_29_12546;
$Lt_29_12802:
	div.approx.ftz.f32 	%f31, %f19, %f23;
	mov.f32 	%f32, 0f3f800000;    	// 1
	setp.eq.ftz.f32 	%p10, %f31, %f32;
	@!%p10 bra 	$Lt_29_13314;
	mov.f32 	%f26, 0f3f800000;    	// 1
	bra.uni 	$Lt_29_13058;
$Lt_29_13314:
	.loc	23	259	0
	mov.f32 	%f33, 0f3f000000;    	// 0.5
	fma.rn.ftz.f32 	%f26, %f29, %f33, %f31;
$Lt_29_13058:
$Lt_29_12546:
$Lt_29_12034:
	.loc	23	264	0
	ld.const.f32 	%f34, [k601YPbPr_To_RGB32f+24];
	mul.ftz.f32 	%f35, %f34, %f26;
	ld.const.f32 	%f36, [k601YPbPr_To_RGB32f+12];
	mul.ftz.f32 	%f37, %f36, %f26;
	ld.const.f32 	%f38, [k601YPbPr_To_RGB32f+0];
	mul.ftz.f32 	%f39, %f38, %f26;
	@!%p2 bra 	$Lt_29_13826;
	.loc	21	126	0
	mul.lo.u64 	%rd7, %rd1, 8;
	add.u64 	%rd8, %rd2, %rd7;
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f35;
	mov.b32		%r28, %b1; }
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f37;
	mov.b32		%r29, %b1; }
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f39;
	mov.b32		%r30, %b1; }
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f4;
	mov.b32		%r31, %b1; }
	st.global.v4.u16 	[%rd8+0], {%r28,%r29,%r30,%r31};
	.loc	23	264	0
	bra.uni 	$Lt_29_13570;
$Lt_29_13826:
	.loc	21	126	0
	mul.lo.u64 	%rd9, %rd1, 16;
	add.u64 	%rd10, %rd2, %rd9;
	st.global.v4.f32 	[%rd10+0], {%f35,%f37,%f39,%f4};
$Lt_29_13570:
$Lt_29_11010:
	.loc	23	266	0
	exit;
$LDWend_LumaCorrector_TonalRangeKernel:
	} // LumaCorrector_TonalRangeKernel

	.entry LumaCorrector_LumaKernel (
		.param .u64 __cudaparm_LumaCorrector_LumaKernel_inImage,
		.param .s32 __cudaparm_LumaCorrector_LumaKernel_inPitch,
		.param .u64 __cudaparm_LumaCorrector_LumaKernel_inSecondaryMask,
		.param .s32 __cudaparm_LumaCorrector_LumaKernel_inSecondaryPitch,
		.param .u32 __cudaparm_LumaCorrector_LumaKernel_inDeviceFormat,
		.param .s32 __cudaparm_LumaCorrector_LumaKernel_inWidth,
		.param .s32 __cudaparm_LumaCorrector_LumaKernel_inHeight,
		.param .align 4 .b8 __cudaparm_LumaCorrector_LumaKernel___val_paraminTonalRangeParameters[16],
		.param .align 4 .b8 __cudaparm_LumaCorrector_LumaKernel___val_paraminLevelsParameters[96])
	{
	.reg .u32 %r<35>;
	.reg .u64 %rd<17>;
	.reg .f32 %f<171>;
	.reg .pred %p<19>;
	.loc	23	280	0
$LDWbegin_LumaCorrector_LumaKernel:
	.loc	23	283	0
	cvt.s32.u32 	%r1, %ctaid.x;
	cvt.s32.u32 	%r2, %ntid.x;
	mul.lo.s32 	%r3, %r1, %r2;
	cvt.s32.u32 	%r4, %ctaid.y;
	cvt.s32.u32 	%r5, %ntid.y;
	mul.lo.s32 	%r6, %r4, %r5;
	mov.u32 	%r7, %tid.x;
	add.u32 	%r8, %r3, %r7;
	mov.u32 	%r9, %tid.y;
	add.u32 	%r10, %r6, %r9;
	ld.param.s32 	%r11, [__cudaparm_LumaCorrector_LumaKernel_inWidth];
	set.gt.u32.s32 	%r12, %r11, %r8;
	neg.s32 	%r13, %r12;
	ld.param.s32 	%r14, [__cudaparm_LumaCorrector_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_30_26114;
	ld.param.s32 	%r19, [__cudaparm_LumaCorrector_LumaKernel_inDeviceFormat];
	mov.s32 	%r20, 0;
	setp.eq.s32 	%p2, %r19, %r20;
	ld.param.s32 	%r21, [__cudaparm_LumaCorrector_LumaKernel_inPitch];
	mul.lo.s32 	%r22, %r21, %r10;
	add.s32 	%r23, %r8, %r22;
	cvt.s64.s32 	%rd1, %r23;
	ld.param.u64 	%rd2, [__cudaparm_LumaCorrector_LumaKernel_inImage];
	@!%p2 bra 	$Lt_30_20738;
	.loc	21	115	0
	mul.lo.u64 	%rd3, %rd1, 8;
	add.u64 	%rd4, %rd2, %rd3;
	ld.global.v4.u16 	{%r24,%r25,%r26,_}, [%rd4+0];
	.loc	23	286	0
	{ .reg .b32 %b1;
	mov.b32		%b1, %r24;
	cvt.ftz.f32.f16	%f1, %b1; }
	{ .reg .b32 %b1;
	mov.b32		%b1, %r25;
	cvt.ftz.f32.f16	%f2, %b1; }
	{ .reg .b32 %b1;
	mov.b32		%b1, %r26;
	cvt.ftz.f32.f16	%f3, %b1; }
	bra.uni 	$Lt_30_20482;
$Lt_30_20738:
	mul.lo.u64 	%rd5, %rd1, 16;
	add.u64 	%rd6, %rd2, %rd5;
	ld.global.v4.f32 	{%f1,%f2,%f3,_}, [%rd6+0];
$Lt_30_20482:
	ld.param.u64 	%rd7, [__cudaparm_LumaCorrector_LumaKernel_inSecondaryMask];
	mov.u64 	%rd8, 0;
	setp.eq.u64 	%p3, %rd7, %rd8;
	@%p3 bra 	$Lt_30_21250;
	.loc	23	287	0
	ld.param.s32 	%r27, [__cudaparm_LumaCorrector_LumaKernel_inSecondaryPitch];
	mul.lo.s32 	%r28, %r27, %r10;
	add.s32 	%r29, %r8, %r28;
	cvt.s64.s32 	%rd9, %r29;
	mul.wide.s32 	%rd10, %r29, 4;
	add.u64 	%rd11, %rd7, %rd10;
	ld.global.f32 	%f4, [%rd11+0];
	bra.uni 	$Lt_30_20994;
$Lt_30_21250:
	mov.f32 	%f4, 0f3f800000;     	// 1
$Lt_30_20994:
	.loc	23	93	0
	ld.const.f32 	%f5, [kRGB32f_To_601YPbPr+4];
	mul.ftz.f32 	%f6, %f5, %f2;
	ld.const.f32 	%f7, [kRGB32f_To_601YPbPr+0];
	fma.rn.ftz.f32 	%f8, %f7, %f3, %f6;
	ld.const.f32 	%f9, [kRGB32f_To_601YPbPr+8];
	fma.rn.ftz.f32 	%f10, %f9, %f1, %f8;
	mov.f32 	%f11, %f10;
	mov.f32 	%f12, 0f3a83126f;    	// 0.001
	setp.gt.ftz.f32 	%p4, %f4, %f12;
	@!%p4 bra 	$Lt_30_21506;
	.loc	20	35	0
	ld.param.f32 	%f13, [__cudaparm_LumaCorrector_LumaKernel___val_paraminTonalRangeParameters+4];
	ld.param.f32 	%f14, [__cudaparm_LumaCorrector_LumaKernel___val_paraminTonalRangeParameters+0];
	setp.le.ftz.f32 	%p5, %f10, %f14;
	@!%p5 bra 	$Lt_30_15106;
	.loc	20	37	0
	mov.f32 	%f15, 0f3f800000;    	// 1
	bra.uni 	$LDWendi_fdividef_207_3;
$Lt_30_15106:
	.loc	20	39	0
	add.ftz.f32 	%f16, %f13, %f14;
	setp.ge.ftz.f32 	%p6, %f10, %f16;
	@!%p6 bra 	$Lt_30_15362;
	.loc	20	41	0
	mov.f32 	%f15, 0f00000000;    	// 0
	bra.uni 	$LDWendi_fdividef_207_3;
$Lt_30_15362:
	.loc	20	45	0
	sub.ftz.f32 	%f17, %f10, %f14;
	div.approx.ftz.f32 	%f15, %f17, %f13;
$LDWendi_fdividef_207_3:
	.loc	20	53	0
	ld.param.f32 	%f18, [__cudaparm_LumaCorrector_LumaKernel___val_paraminTonalRangeParameters+12];
	ld.param.f32 	%f19, [__cudaparm_LumaCorrector_LumaKernel___val_paraminTonalRangeParameters+8];
	sub.ftz.f32 	%f20, %f19, %f18;
	setp.gt.ftz.f32 	%p7, %f20, %f10;
	@!%p7 bra 	$Lt_30_15618;
	.loc	20	55	0
	mov.f32 	%f21, 0f00000000;    	// 0
	bra.uni 	$LDWendi_fdividef_207_1;
$Lt_30_15618:
	.loc	20	57	0
	setp.gt.ftz.f32 	%p8, %f10, %f19;
	@!%p8 bra 	$Lt_30_15874;
	.loc	20	59	0
	mov.f32 	%f21, 0f3f800000;    	// 1
	bra.uni 	$LDWendi_fdividef_207_1;
$Lt_30_15874:
	.loc	20	63	0
	sub.ftz.f32 	%f22, %f10, %f20;
	div.approx.ftz.f32 	%f21, %f22, %f18;
$LDWendi_fdividef_207_1:
	.loc	23	148	0
	setp.lt.ftz.f32 	%p9, %f18, %f13;
	mov.f32 	%f23, 0f40000000;    	// 2
	mov.f32 	%f24, 0f3f800000;    	// 1
	selp.f32 	%f25, %f23, %f24, %p9;
	div.approx.ftz.f32 	%f26, %f15, %f25;
	sub.ftz.f32 	%f27, %f25, %f15;
	div.approx.ftz.f32 	%f28, %f21, %f25;
	ld.param.f32 	%f29, [__cudaparm_LumaCorrector_LumaKernel___val_paraminLevelsParameters+76];
	mul.ftz.f32 	%f30, %f29, %f26;
	ld.param.f32 	%f31, [__cudaparm_LumaCorrector_LumaKernel___val_paraminLevelsParameters+80];
	mul.ftz.f32 	%f32, %f31, %f26;
	ld.param.f32 	%f33, [__cudaparm_LumaCorrector_LumaKernel___val_paraminLevelsParameters+72];
	mul.ftz.f32 	%f34, %f33, %f26;
	sub.ftz.f32 	%f35, %f27, %f21;
	ld.param.f32 	%f36, [__cudaparm_LumaCorrector_LumaKernel___val_paraminLevelsParameters+28];
	mul.ftz.f32 	%f37, %f36, %f28;
	ld.param.f32 	%f38, [__cudaparm_LumaCorrector_LumaKernel___val_paraminLevelsParameters+24];
	mul.ftz.f32 	%f39, %f38, %f28;
	mul.ftz.f32 	%f40, %f4, %f30;
	mul.ftz.f32 	%f41, %f4, %f34;
	div.approx.ftz.f32 	%f42, %f35, %f25;
	ld.param.f32 	%f43, [__cudaparm_LumaCorrector_LumaKernel___val_paraminLevelsParameters+52];
	mul.ftz.f32 	%f44, %f43, %f42;
	ld.param.f32 	%f45, [__cudaparm_LumaCorrector_LumaKernel___val_paraminLevelsParameters+56];
	fma.rn.ftz.f32 	%f46, %f45, %f42, %f32;
	ld.param.f32 	%f47, [__cudaparm_LumaCorrector_LumaKernel___val_paraminLevelsParameters+48];
	mul.ftz.f32 	%f48, %f47, %f42;
	fma.rn.ftz.f32 	%f49, %f4, %f44, %f40;
	ld.param.f32 	%f50, [__cudaparm_LumaCorrector_LumaKernel___val_paraminLevelsParameters+32];
	fma.rn.ftz.f32 	%f51, %f50, %f28, %f46;
	fma.rn.ftz.f32 	%f52, %f4, %f48, %f41;
	fma.rn.ftz.f32 	%f53, %f4, %f37, %f49;
	mov.f32 	%f54, 0fbf000000;    	// -0.5
	add.ftz.f32 	%f55, %f51, %f54;
	fma.rn.ftz.f32 	%f56, %f4, %f39, %f52;
	mov.f32 	%f57, 0f3f000000;    	// 0.5
	fma.rn.ftz.f32 	%f58, %f4, %f55, %f57;
	mov.f32 	%f59, 0f00000000;    	// 0
	setp.ge.ftz.f32 	%p10, %f53, %f59;
	@!%p10 bra 	$Lt_30_22274;
	mov.f32 	%f60, 0f3f800000;    	// 1
	sub.ftz.f32 	%f61, %f60, %f53;
	sub.ftz.f32 	%f62, %f58, %f56;
	setp.le.ftz.f32 	%p11, %f62, %f10;
	mov.f32 	%f63, 0f00000000;    	// 0
	setp.eq.ftz.f32 	%p12, %f61, %f63;
	@!%p12 bra 	$Lt_30_22786;
	.loc	23	46	0
	mov.f32 	%f64, 0f3f800000;    	// 1
	mov.f32 	%f65, 0f00000000;    	// 0
	selp.f32 	%f66, %f64, %f65, %p11;
	bra.uni 	$Lt_30_23554;
$Lt_30_22786:
	@!%p11 bra 	$Lt_30_23298;
	.loc	23	59	0
	sub.ftz.f32 	%f67, %f10, %f62;
	.loc	28	529	0
	div.approx.ftz.f32 	%f68, %f67, %f61;
	.loc	23	59	0
	add.ftz.f32 	%f66, %f68, %f58;
	bra.uni 	$Lt_30_23554;
$Lt_30_23298:
	.loc	23	63	0
	sub.ftz.f32 	%f69, %f62, %f10;
	.loc	28	529	0
	div.approx.ftz.f32 	%f70, %f69, %f61;
	.loc	23	63	0
	sub.ftz.f32 	%f66, %f58, %f70;
	bra.uni 	$Lt_30_23554;
$Lt_30_22274:
	mov.f32 	%f71, 0f3f800000;    	// 1
	add.ftz.f32 	%f72, %f53, %f71;
	add.ftz.f32 	%f73, %f56, %f58;
	setp.le.ftz.f32 	%p13, %f58, %f10;
	@!%p13 bra 	$Lt_30_23810;
	.loc	23	76	0
	sub.ftz.f32 	%f74, %f10, %f58;
	fma.rn.ftz.f32 	%f66, %f72, %f74, %f73;
	bra.uni 	$Lt_30_23554;
$Lt_30_23810:
	.loc	23	80	0
	sub.ftz.f32 	%f75, %f58, %f10;
	mul.ftz.f32 	%f76, %f72, %f75;
	sub.ftz.f32 	%f66, %f73, %f76;
$Lt_30_23554:
$Lt_30_22018:
	.loc	28	523	0
	mov.f32 	%f77, 0f00000000;    	// 0
	max.ftz.f32 	%f78, %f66, %f77;
	mov.f32 	%f79, 0f3f800000;    	// 1
	min.ftz.f32 	%f80, %f78, %f79;
	lg2.approx.ftz.f32 	%f81, %f80;
	mov.f32 	%f82, 0f3f800000;    	// 1
	ld.param.f32 	%f83, [__cudaparm_LumaCorrector_LumaKernel___val_paraminLevelsParameters+84];
	mul.ftz.f32 	%f84, %f83, %f26;
	ld.param.f32 	%f85, [__cudaparm_LumaCorrector_LumaKernel___val_paraminLevelsParameters+60];
	fma.rn.ftz.f32 	%f86, %f85, %f42, %f84;
	ld.param.f32 	%f87, [__cudaparm_LumaCorrector_LumaKernel___val_paraminLevelsParameters+36];
	fma.rn.ftz.f32 	%f88, %f87, %f28, %f86;
	mov.f32 	%f89, 0fbf800000;    	// -1
	add.ftz.f32 	%f90, %f88, %f89;
	fma.rn.ftz.f32 	%f91, %f4, %f90, %f82;
	rcp.approx.ftz.f32 	%f92, %f91;
	mul.ftz.f32 	%f93, %f81, %f92;
	ex2.approx.ftz.f32 	%f94, %f93;
	.loc	23	169	0
	ld.param.f32 	%f95, [__cudaparm_LumaCorrector_LumaKernel___val_paraminLevelsParameters+4];
	mul.ftz.f32 	%f96, %f95, %f4;
	ld.param.f32 	%f97, [__cudaparm_LumaCorrector_LumaKernel___val_paraminLevelsParameters+8];
	mov.f32 	%f98, 0fbf000000;    	// -0.5
	add.ftz.f32 	%f99, %f97, %f98;
	ld.param.f32 	%f100, [__cudaparm_LumaCorrector_LumaKernel___val_paraminLevelsParameters+0];
	mul.ftz.f32 	%f101, %f100, %f4;
	mov.f32 	%f102, 0f3f000000;   	// 0.5
	fma.rn.ftz.f32 	%f103, %f4, %f99, %f102;
	ld.param.f32 	%f104, [__cudaparm_LumaCorrector_LumaKernel___val_paraminLevelsParameters+88];
	mul.ftz.f32 	%f105, %f104, %f26;
	ld.param.f32 	%f106, [__cudaparm_LumaCorrector_LumaKernel___val_paraminLevelsParameters+92];
	mul.ftz.f32 	%f107, %f106, %f26;
	ld.param.f32 	%f108, [__cudaparm_LumaCorrector_LumaKernel___val_paraminLevelsParameters+40];
	mul.ftz.f32 	%f109, %f108, %f28;
	mul.ftz.f32 	%f110, %f4, %f105;
	ld.param.f32 	%f111, [__cudaparm_LumaCorrector_LumaKernel___val_paraminLevelsParameters+64];
	mul.ftz.f32 	%f112, %f111, %f42;
	ld.param.f32 	%f113, [__cudaparm_LumaCorrector_LumaKernel___val_paraminLevelsParameters+68];
	fma.rn.ftz.f32 	%f114, %f113, %f42, %f107;
	fma.rn.ftz.f32 	%f115, %f4, %f112, %f110;
	ld.param.f32 	%f116, [__cudaparm_LumaCorrector_LumaKernel___val_paraminLevelsParameters+44];
	fma.rn.ftz.f32 	%f117, %f116, %f28, %f114;
	fma.rn.ftz.f32 	%f118, %f4, %f109, %f115;
	mov.f32 	%f119, 0fbf800000;   	// -1
	add.ftz.f32 	%f120, %f117, %f119;
	add.ftz.f32 	%f121, %f94, %f118;
	mov.f32 	%f122, 0f3f800000;   	// 1
	fma.rn.ftz.f32 	%f123, %f4, %f120, %f122;
	mul.ftz.f32 	%f124, %f121, %f123;
	mov.f32 	%f125, 0f00000000;   	// 0
	setp.ge.ftz.f32 	%p14, %f96, %f125;
	@!%p14 bra 	$Lt_30_24322;
	mov.f32 	%f126, 0f3f800000;   	// 1
	sub.ftz.f32 	%f127, %f126, %f96;
	sub.ftz.f32 	%f128, %f103, %f101;
	setp.le.ftz.f32 	%p15, %f128, %f124;
	mov.f32 	%f129, 0f00000000;   	// 0
	setp.eq.ftz.f32 	%p16, %f127, %f129;
	@!%p16 bra 	$Lt_30_24834;
	.loc	23	46	0
	mov.f32 	%f130, 0f3f800000;   	// 1
	mov.f32 	%f131, 0f00000000;   	// 0
	selp.f32 	%f66, %f130, %f131, %p15;
	bra.uni 	$Lt_30_25602;
$Lt_30_24834:
	@!%p15 bra 	$Lt_30_25346;
	.loc	23	59	0
	sub.ftz.f32 	%f132, %f124, %f128;
	div.approx.ftz.f32 	%f133, %f132, %f127;
	add.ftz.f32 	%f66, %f103, %f133;
	bra.uni 	$Lt_30_25602;
$Lt_30_25346:
	.loc	23	63	0
	sub.ftz.f32 	%f134, %f128, %f124;
	div.approx.ftz.f32 	%f135, %f134, %f127;
	sub.ftz.f32 	%f66, %f103, %f135;
	bra.uni 	$Lt_30_25602;
$Lt_30_24322:
	mov.f32 	%f136, 0f3f800000;   	// 1
	add.ftz.f32 	%f137, %f96, %f136;
	add.ftz.f32 	%f138, %f101, %f103;
	setp.le.ftz.f32 	%p17, %f103, %f124;
	@!%p17 bra 	$Lt_30_25858;
	.loc	23	76	0
	sub.ftz.f32 	%f139, %f124, %f103;
	fma.rn.ftz.f32 	%f66, %f137, %f139, %f138;
	bra.uni 	$Lt_30_25602;
$Lt_30_25858:
	.loc	23	80	0
	sub.ftz.f32 	%f140, %f103, %f124;
	mul.ftz.f32 	%f141, %f137, %f140;
	sub.ftz.f32 	%f66, %f138, %f141;
$Lt_30_25602:
$Lt_30_24066:
	.loc	23	184	0
	mov.f32 	%f142, 0f3f800000;   	// 1
	ld.param.f32 	%f143, [__cudaparm_LumaCorrector_LumaKernel___val_paraminLevelsParameters+20];
	mov.f32 	%f144, 0fbf800000;   	// -1
	add.ftz.f32 	%f145, %f143, %f144;
	fma.rn.ftz.f32 	%f146, %f4, %f145, %f142;
	mov.f32 	%f147, 0f00000000;   	// 0
	max.ftz.f32 	%f148, %f66, %f147;
	mov.f32 	%f149, 0f3f800000;   	// 1
	min.ftz.f32 	%f150, %f148, %f149;
	lg2.approx.ftz.f32 	%f151, %f150;
	mov.f32 	%f152, 0f3f800000;   	// 1
	ld.param.f32 	%f153, [__cudaparm_LumaCorrector_LumaKernel___val_paraminLevelsParameters+12];
	mov.f32 	%f154, 0fbf800000;   	// -1
	add.ftz.f32 	%f155, %f153, %f154;
	fma.rn.ftz.f32 	%f156, %f4, %f155, %f152;
	rcp.approx.ftz.f32 	%f157, %f156;
	mul.ftz.f32 	%f158, %f151, %f157;
	ex2.approx.ftz.f32 	%f159, %f158;
	ld.param.f32 	%f160, [__cudaparm_LumaCorrector_LumaKernel___val_paraminLevelsParameters+16];
	fma.rn.ftz.f32 	%f161, %f160, %f4, %f159;
	mul.ftz.f32 	%f11, %f146, %f161;
$Lt_30_21506:
	.loc	23	297	0
	ld.const.f32 	%f162, [k601YPbPr_To_RGB32f+24];
	mul.ftz.f32 	%f163, %f162, %f11;
	ld.const.f32 	%f164, [k601YPbPr_To_RGB32f+12];
	mul.ftz.f32 	%f165, %f164, %f11;
	ld.const.f32 	%f166, [k601YPbPr_To_RGB32f+0];
	mul.ftz.f32 	%f167, %f166, %f11;
	@!%p2 bra 	$Lt_30_26370;
	.loc	21	126	0
	mul.lo.u64 	%rd12, %rd1, 8;
	add.u64 	%rd13, %rd2, %rd12;
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f163;
	mov.b32		%r30, %b1; }
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f165;
	mov.b32		%r31, %b1; }
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f167;
	mov.b32		%r32, %b1; }
	mov.f32 	%f168, 0f3f800000;   	// 1
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f168;
	mov.b32		%r33, %b1; }
	st.global.v4.u16 	[%rd13+0], {%r30,%r31,%r32,%r33};
	.loc	23	297	0
	bra.uni 	$Lt_30_26114;
$Lt_30_26370:
	.loc	21	126	0
	mul.lo.u64 	%rd14, %rd1, 16;
	add.u64 	%rd15, %rd2, %rd14;
	mov.f32 	%f169, 0f3f800000;   	// 1
	st.global.v4.f32 	[%rd15+0], {%f163,%f165,%f167,%f169};
$Lt_30_26114:
$Lt_30_19970:
	.loc	23	299	0
	exit;
$LDWend_LumaCorrector_LumaKernel:
	} // LumaCorrector_LumaKernel

	.entry LumaCorrector_CompositeKernel (
		.param .u64 __cudaparm_LumaCorrector_CompositeKernel_inImage,
		.param .s32 __cudaparm_LumaCorrector_CompositeKernel_inPitch,
		.param .u64 __cudaparm_LumaCorrector_CompositeKernel_inSecondaryMask,
		.param .s32 __cudaparm_LumaCorrector_CompositeKernel_inSecondaryPitch,
		.param .u32 __cudaparm_LumaCorrector_CompositeKernel_inDeviceFormat,
		.param .s32 __cudaparm_LumaCorrector_CompositeKernel_inWidth,
		.param .s32 __cudaparm_LumaCorrector_CompositeKernel_inHeight,
		.param .align 4 .b8 __cudaparm_LumaCorrector_CompositeKernel___val_paraminTonalRangeParameters[16],
		.param .align 4 .b8 __cudaparm_LumaCorrector_CompositeKernel___val_paraminLevelsParameters[96])
	{
	.reg .u32 %r<36>;
	.reg .u64 %rd<17>;
	.reg .f32 %f<194>;
	.reg .pred %p<19>;
	.loc	23	313	0
$LDWbegin_LumaCorrector_CompositeKernel:
	.loc	23	316	0
	cvt.s32.u32 	%r1, %ctaid.x;
	cvt.s32.u32 	%r2, %ntid.x;
	mul.lo.s32 	%r3, %r1, %r2;
	cvt.s32.u32 	%r4, %ctaid.y;
	cvt.s32.u32 	%r5, %ntid.y;
	mul.lo.s32 	%r6, %r4, %r5;
	mov.u32 	%r7, %tid.x;
	add.u32 	%r8, %r3, %r7;
	mov.u32 	%r9, %tid.y;
	add.u32 	%r10, %r6, %r9;
	ld.param.s32 	%r11, [__cudaparm_LumaCorrector_CompositeKernel_inWidth];
	set.gt.u32.s32 	%r12, %r11, %r8;
	neg.s32 	%r13, %r12;
	ld.param.s32 	%r14, [__cudaparm_LumaCorrector_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_31_26114;
	ld.param.s32 	%r19, [__cudaparm_LumaCorrector_CompositeKernel_inDeviceFormat];
	mov.s32 	%r20, 0;
	setp.eq.s32 	%p2, %r19, %r20;
	ld.param.s32 	%r21, [__cudaparm_LumaCorrector_CompositeKernel_inPitch];
	mul.lo.s32 	%r22, %r21, %r10;
	add.s32 	%r23, %r8, %r22;
	cvt.s64.s32 	%rd1, %r23;
	ld.param.u64 	%rd2, [__cudaparm_LumaCorrector_CompositeKernel_inImage];
	@!%p2 bra 	$Lt_31_20738;
	.loc	21	115	0
	mul.lo.u64 	%rd3, %rd1, 8;
	add.u64 	%rd4, %rd2, %rd3;
	ld.global.v4.u16 	{%r24,%r25,%r26,%r27}, [%rd4+0];
	.loc	23	319	0
	{ .reg .b32 %b1;
	mov.b32		%b1, %r24;
	cvt.ftz.f32.f16	%f1, %b1; }
	{ .reg .b32 %b1;
	mov.b32		%b1, %r25;
	cvt.ftz.f32.f16	%f2, %b1; }
	{ .reg .b32 %b1;
	mov.b32		%b1, %r26;
	cvt.ftz.f32.f16	%f3, %b1; }
	{ .reg .b32 %b1;
	mov.b32		%b1, %r27;
	cvt.ftz.f32.f16	%f4, %b1; }
	bra.uni 	$Lt_31_20482;
$Lt_31_20738:
	mul.lo.u64 	%rd5, %rd1, 16;
	add.u64 	%rd6, %rd2, %rd5;
	ld.global.v4.f32 	{%f1,%f2,%f3,%f4}, [%rd6+0];
$Lt_31_20482:
	ld.param.u64 	%rd7, [__cudaparm_LumaCorrector_CompositeKernel_inSecondaryMask];
	mov.u64 	%rd8, 0;
	setp.eq.u64 	%p3, %rd7, %rd8;
	@%p3 bra 	$Lt_31_21250;
	.loc	23	320	0
	ld.param.s32 	%r28, [__cudaparm_LumaCorrector_CompositeKernel_inSecondaryPitch];
	mul.lo.s32 	%r29, %r28, %r10;
	add.s32 	%r30, %r8, %r29;
	cvt.s64.s32 	%rd9, %r30;
	mul.wide.s32 	%rd10, %r30, 4;
	add.u64 	%rd11, %rd7, %rd10;
	ld.global.f32 	%f5, [%rd11+0];
	bra.uni 	$Lt_31_20994;
$Lt_31_21250:
	mov.f32 	%f5, 0f3f800000;     	// 1
$Lt_31_20994:
	.loc	23	93	0
	ld.const.f32 	%f6, [kRGB32f_To_601YPbPr+4];
	mul.ftz.f32 	%f7, %f6, %f2;
	ld.const.f32 	%f8, [kRGB32f_To_601YPbPr+0];
	fma.rn.ftz.f32 	%f9, %f8, %f3, %f7;
	ld.const.f32 	%f10, [kRGB32f_To_601YPbPr+8];
	fma.rn.ftz.f32 	%f11, %f10, %f1, %f9;
	mov.f32 	%f12, %f11;
	mov.f32 	%f13, 0f3a83126f;    	// 0.001
	setp.gt.ftz.f32 	%p4, %f5, %f13;
	@!%p4 bra 	$Lt_31_21506;
	.loc	20	35	0
	ld.param.f32 	%f14, [__cudaparm_LumaCorrector_CompositeKernel___val_paraminTonalRangeParameters+4];
	ld.param.f32 	%f15, [__cudaparm_LumaCorrector_CompositeKernel___val_paraminTonalRangeParameters+0];
	setp.le.ftz.f32 	%p5, %f11, %f15;
	@!%p5 bra 	$Lt_31_15106;
	.loc	20	37	0
	mov.f32 	%f16, 0f3f800000;    	// 1
	bra.uni 	$LDWendi_fdividef_208_3;
$Lt_31_15106:
	.loc	20	39	0
	add.ftz.f32 	%f17, %f14, %f15;
	setp.ge.ftz.f32 	%p6, %f11, %f17;
	@!%p6 bra 	$Lt_31_15362;
	.loc	20	41	0
	mov.f32 	%f16, 0f00000000;    	// 0
	bra.uni 	$LDWendi_fdividef_208_3;
$Lt_31_15362:
	.loc	20	45	0
	sub.ftz.f32 	%f18, %f11, %f15;
	div.approx.ftz.f32 	%f16, %f18, %f14;
$LDWendi_fdividef_208_3:
	.loc	20	53	0
	ld.param.f32 	%f19, [__cudaparm_LumaCorrector_CompositeKernel___val_paraminTonalRangeParameters+12];
	ld.param.f32 	%f20, [__cudaparm_LumaCorrector_CompositeKernel___val_paraminTonalRangeParameters+8];
	sub.ftz.f32 	%f21, %f20, %f19;
	setp.gt.ftz.f32 	%p7, %f21, %f11;
	@!%p7 bra 	$Lt_31_15618;
	.loc	20	55	0
	mov.f32 	%f22, 0f00000000;    	// 0
	bra.uni 	$LDWendi_fdividef_208_1;
$Lt_31_15618:
	.loc	20	57	0
	setp.gt.ftz.f32 	%p8, %f11, %f20;
	@!%p8 bra 	$Lt_31_15874;
	.loc	20	59	0
	mov.f32 	%f22, 0f3f800000;    	// 1
	bra.uni 	$LDWendi_fdividef_208_1;
$Lt_31_15874:
	.loc	20	63	0
	sub.ftz.f32 	%f23, %f11, %f21;
	div.approx.ftz.f32 	%f22, %f23, %f19;
$LDWendi_fdividef_208_1:
	.loc	23	148	0
	setp.lt.ftz.f32 	%p9, %f19, %f14;
	mov.f32 	%f24, 0f40000000;    	// 2
	mov.f32 	%f25, 0f3f800000;    	// 1
	selp.f32 	%f26, %f24, %f25, %p9;
	div.approx.ftz.f32 	%f27, %f16, %f26;
	sub.ftz.f32 	%f28, %f26, %f16;
	div.approx.ftz.f32 	%f29, %f22, %f26;
	ld.param.f32 	%f30, [__cudaparm_LumaCorrector_CompositeKernel___val_paraminLevelsParameters+76];
	mul.ftz.f32 	%f31, %f30, %f27;
	ld.param.f32 	%f32, [__cudaparm_LumaCorrector_CompositeKernel___val_paraminLevelsParameters+80];
	mul.ftz.f32 	%f33, %f32, %f27;
	ld.param.f32 	%f34, [__cudaparm_LumaCorrector_CompositeKernel___val_paraminLevelsParameters+72];
	mul.ftz.f32 	%f35, %f34, %f27;
	sub.ftz.f32 	%f36, %f28, %f22;
	ld.param.f32 	%f37, [__cudaparm_LumaCorrector_CompositeKernel___val_paraminLevelsParameters+28];
	mul.ftz.f32 	%f38, %f37, %f29;
	ld.param.f32 	%f39, [__cudaparm_LumaCorrector_CompositeKernel___val_paraminLevelsParameters+24];
	mul.ftz.f32 	%f40, %f39, %f29;
	mul.ftz.f32 	%f41, %f5, %f31;
	mul.ftz.f32 	%f42, %f5, %f35;
	div.approx.ftz.f32 	%f43, %f36, %f26;
	ld.param.f32 	%f44, [__cudaparm_LumaCorrector_CompositeKernel___val_paraminLevelsParameters+52];
	mul.ftz.f32 	%f45, %f44, %f43;
	ld.param.f32 	%f46, [__cudaparm_LumaCorrector_CompositeKernel___val_paraminLevelsParameters+56];
	fma.rn.ftz.f32 	%f47, %f46, %f43, %f33;
	ld.param.f32 	%f48, [__cudaparm_LumaCorrector_CompositeKernel___val_paraminLevelsParameters+48];
	mul.ftz.f32 	%f49, %f48, %f43;
	fma.rn.ftz.f32 	%f50, %f5, %f45, %f41;
	ld.param.f32 	%f51, [__cudaparm_LumaCorrector_CompositeKernel___val_paraminLevelsParameters+32];
	fma.rn.ftz.f32 	%f52, %f51, %f29, %f47;
	fma.rn.ftz.f32 	%f53, %f5, %f49, %f42;
	fma.rn.ftz.f32 	%f54, %f5, %f38, %f50;
	mov.f32 	%f55, 0fbf000000;    	// -0.5
	add.ftz.f32 	%f56, %f52, %f55;
	fma.rn.ftz.f32 	%f57, %f5, %f40, %f53;
	mov.f32 	%f58, 0f3f000000;    	// 0.5
	fma.rn.ftz.f32 	%f59, %f5, %f56, %f58;
	mov.f32 	%f60, 0f00000000;    	// 0
	setp.ge.ftz.f32 	%p10, %f54, %f60;
	@!%p10 bra 	$Lt_31_22274;
	mov.f32 	%f61, 0f3f800000;    	// 1
	sub.ftz.f32 	%f62, %f61, %f54;
	sub.ftz.f32 	%f63, %f59, %f57;
	setp.le.ftz.f32 	%p11, %f63, %f11;
	mov.f32 	%f64, 0f00000000;    	// 0
	setp.eq.ftz.f32 	%p12, %f62, %f64;
	@!%p12 bra 	$Lt_31_22786;
	.loc	23	46	0
	mov.f32 	%f65, 0f3f800000;    	// 1
	mov.f32 	%f66, 0f00000000;    	// 0
	selp.f32 	%f67, %f65, %f66, %p11;
	bra.uni 	$Lt_31_23554;
$Lt_31_22786:
	@!%p11 bra 	$Lt_31_23298;
	.loc	23	59	0
	sub.ftz.f32 	%f68, %f11, %f63;
	.loc	28	529	0
	div.approx.ftz.f32 	%f69, %f68, %f62;
	.loc	23	59	0
	add.ftz.f32 	%f67, %f69, %f59;
	bra.uni 	$Lt_31_23554;
$Lt_31_23298:
	.loc	23	63	0
	sub.ftz.f32 	%f70, %f63, %f11;
	.loc	28	529	0
	div.approx.ftz.f32 	%f71, %f70, %f62;
	.loc	23	63	0
	sub.ftz.f32 	%f67, %f59, %f71;
	bra.uni 	$Lt_31_23554;
$Lt_31_22274:
	mov.f32 	%f72, 0f3f800000;    	// 1
	add.ftz.f32 	%f73, %f54, %f72;
	add.ftz.f32 	%f74, %f57, %f59;
	setp.le.ftz.f32 	%p13, %f59, %f11;
	@!%p13 bra 	$Lt_31_23810;
	.loc	23	76	0
	sub.ftz.f32 	%f75, %f11, %f59;
	fma.rn.ftz.f32 	%f67, %f73, %f75, %f74;
	bra.uni 	$Lt_31_23554;
$Lt_31_23810:
	.loc	23	80	0
	sub.ftz.f32 	%f76, %f59, %f11;
	mul.ftz.f32 	%f77, %f73, %f76;
	sub.ftz.f32 	%f67, %f74, %f77;
$Lt_31_23554:
$Lt_31_22018:
	.loc	28	523	0
	mov.f32 	%f78, 0f00000000;    	// 0
	max.ftz.f32 	%f79, %f67, %f78;
	mov.f32 	%f80, 0f3f800000;    	// 1
	min.ftz.f32 	%f81, %f79, %f80;
	lg2.approx.ftz.f32 	%f82, %f81;
	mov.f32 	%f83, 0f3f800000;    	// 1
	ld.param.f32 	%f84, [__cudaparm_LumaCorrector_CompositeKernel___val_paraminLevelsParameters+84];
	mul.ftz.f32 	%f85, %f84, %f27;
	ld.param.f32 	%f86, [__cudaparm_LumaCorrector_CompositeKernel___val_paraminLevelsParameters+60];
	fma.rn.ftz.f32 	%f87, %f86, %f43, %f85;
	ld.param.f32 	%f88, [__cudaparm_LumaCorrector_CompositeKernel___val_paraminLevelsParameters+36];
	fma.rn.ftz.f32 	%f89, %f88, %f29, %f87;
	mov.f32 	%f90, 0fbf800000;    	// -1
	add.ftz.f32 	%f91, %f89, %f90;
	fma.rn.ftz.f32 	%f92, %f5, %f91, %f83;
	rcp.approx.ftz.f32 	%f93, %f92;
	mul.ftz.f32 	%f94, %f82, %f93;
	ex2.approx.ftz.f32 	%f95, %f94;
	.loc	23	169	0
	ld.param.f32 	%f96, [__cudaparm_LumaCorrector_CompositeKernel___val_paraminLevelsParameters+4];
	mul.ftz.f32 	%f97, %f96, %f5;
	ld.param.f32 	%f98, [__cudaparm_LumaCorrector_CompositeKernel___val_paraminLevelsParameters+8];
	mov.f32 	%f99, 0fbf000000;    	// -0.5
	add.ftz.f32 	%f100, %f98, %f99;
	ld.param.f32 	%f101, [__cudaparm_LumaCorrector_CompositeKernel___val_paraminLevelsParameters+0];
	mul.ftz.f32 	%f102, %f101, %f5;
	mov.f32 	%f103, 0f3f000000;   	// 0.5
	fma.rn.ftz.f32 	%f104, %f5, %f100, %f103;
	ld.param.f32 	%f105, [__cudaparm_LumaCorrector_CompositeKernel___val_paraminLevelsParameters+88];
	mul.ftz.f32 	%f106, %f105, %f27;
	ld.param.f32 	%f107, [__cudaparm_LumaCorrector_CompositeKernel___val_paraminLevelsParameters+92];
	mul.ftz.f32 	%f108, %f107, %f27;
	ld.param.f32 	%f109, [__cudaparm_LumaCorrector_CompositeKernel___val_paraminLevelsParameters+40];
	mul.ftz.f32 	%f110, %f109, %f29;
	mul.ftz.f32 	%f111, %f5, %f106;
	ld.param.f32 	%f112, [__cudaparm_LumaCorrector_CompositeKernel___val_paraminLevelsParameters+64];
	mul.ftz.f32 	%f113, %f112, %f43;
	ld.param.f32 	%f114, [__cudaparm_LumaCorrector_CompositeKernel___val_paraminLevelsParameters+68];
	fma.rn.ftz.f32 	%f115, %f114, %f43, %f108;
	fma.rn.ftz.f32 	%f116, %f5, %f113, %f111;
	ld.param.f32 	%f117, [__cudaparm_LumaCorrector_CompositeKernel___val_paraminLevelsParameters+44];
	fma.rn.ftz.f32 	%f118, %f117, %f29, %f115;
	fma.rn.ftz.f32 	%f119, %f5, %f110, %f116;
	mov.f32 	%f120, 0fbf800000;   	// -1
	add.ftz.f32 	%f121, %f118, %f120;
	add.ftz.f32 	%f122, %f95, %f119;
	mov.f32 	%f123, 0f3f800000;   	// 1
	fma.rn.ftz.f32 	%f124, %f5, %f121, %f123;
	mul.ftz.f32 	%f125, %f122, %f124;
	mov.f32 	%f126, 0f00000000;   	// 0
	setp.ge.ftz.f32 	%p14, %f97, %f126;
	@!%p14 bra 	$Lt_31_24322;
	mov.f32 	%f127, 0f3f800000;   	// 1
	sub.ftz.f32 	%f128, %f127, %f97;
	sub.ftz.f32 	%f129, %f104, %f102;
	setp.le.ftz.f32 	%p15, %f129, %f125;
	mov.f32 	%f130, 0f00000000;   	// 0
	setp.eq.ftz.f32 	%p16, %f128, %f130;
	@!%p16 bra 	$Lt_31_24834;
	.loc	23	46	0
	mov.f32 	%f131, 0f3f800000;   	// 1
	mov.f32 	%f132, 0f00000000;   	// 0
	selp.f32 	%f67, %f131, %f132, %p15;
	bra.uni 	$Lt_31_25602;
$Lt_31_24834:
	@!%p15 bra 	$Lt_31_25346;
	.loc	23	59	0
	sub.ftz.f32 	%f133, %f125, %f129;
	div.approx.ftz.f32 	%f134, %f133, %f128;
	add.ftz.f32 	%f67, %f104, %f134;
	bra.uni 	$Lt_31_25602;
$Lt_31_25346:
	.loc	23	63	0
	sub.ftz.f32 	%f135, %f129, %f125;
	div.approx.ftz.f32 	%f136, %f135, %f128;
	sub.ftz.f32 	%f67, %f104, %f136;
	bra.uni 	$Lt_31_25602;
$Lt_31_24322:
	mov.f32 	%f137, 0f3f800000;   	// 1
	add.ftz.f32 	%f138, %f97, %f137;
	add.ftz.f32 	%f139, %f102, %f104;
	setp.le.ftz.f32 	%p17, %f104, %f125;
	@!%p17 bra 	$Lt_31_25858;
	.loc	23	76	0
	sub.ftz.f32 	%f140, %f125, %f104;
	fma.rn.ftz.f32 	%f67, %f138, %f140, %f139;
	bra.uni 	$Lt_31_25602;
$Lt_31_25858:
	.loc	23	80	0
	sub.ftz.f32 	%f141, %f104, %f125;
	mul.ftz.f32 	%f142, %f138, %f141;
	sub.ftz.f32 	%f67, %f139, %f142;
$Lt_31_25602:
$Lt_31_24066:
	.loc	23	184	0
	mov.f32 	%f143, 0f3f800000;   	// 1
	ld.param.f32 	%f144, [__cudaparm_LumaCorrector_CompositeKernel___val_paraminLevelsParameters+20];
	mov.f32 	%f145, 0fbf800000;   	// -1
	add.ftz.f32 	%f146, %f144, %f145;
	fma.rn.ftz.f32 	%f147, %f5, %f146, %f143;
	mov.f32 	%f148, 0f00000000;   	// 0
	max.ftz.f32 	%f149, %f67, %f148;
	mov.f32 	%f150, 0f3f800000;   	// 1
	min.ftz.f32 	%f151, %f149, %f150;
	lg2.approx.ftz.f32 	%f152, %f151;
	mov.f32 	%f153, 0f3f800000;   	// 1
	ld.param.f32 	%f154, [__cudaparm_LumaCorrector_CompositeKernel___val_paraminLevelsParameters+12];
	mov.f32 	%f155, 0fbf800000;   	// -1
	add.ftz.f32 	%f156, %f154, %f155;
	fma.rn.ftz.f32 	%f157, %f5, %f156, %f153;
	rcp.approx.ftz.f32 	%f158, %f157;
	mul.ftz.f32 	%f159, %f152, %f158;
	ex2.approx.ftz.f32 	%f160, %f159;
	ld.param.f32 	%f161, [__cudaparm_LumaCorrector_CompositeKernel___val_paraminLevelsParameters+16];
	fma.rn.ftz.f32 	%f162, %f161, %f5, %f160;
	mul.ftz.f32 	%f12, %f147, %f162;
$Lt_31_21506:
	.loc	23	328	0
	ld.const.f32 	%f163, [kRGB32f_To_601YPbPr+16];
	mul.ftz.f32 	%f164, %f163, %f2;
	ld.const.f32 	%f165, [kRGB32f_To_601YPbPr+28];
	mul.ftz.f32 	%f166, %f165, %f2;
	ld.const.f32 	%f167, [kRGB32f_To_601YPbPr+12];
	fma.rn.ftz.f32 	%f168, %f167, %f3, %f164;
	ld.const.f32 	%f169, [kRGB32f_To_601YPbPr+24];
	fma.rn.ftz.f32 	%f170, %f169, %f3, %f166;
	ld.const.f32 	%f171, [kRGB32f_To_601YPbPr+20];
	fma.rn.ftz.f32 	%f172, %f171, %f1, %f168;
	ld.const.f32 	%f173, [kRGB32f_To_601YPbPr+32];
	fma.rn.ftz.f32 	%f174, %f173, %f1, %f170;
	ld.const.f32 	%f175, [k601YPbPr_To_RGB32f+28];
	mul.ftz.f32 	%f176, %f175, %f172;
	ld.const.f32 	%f177, [k601YPbPr_To_RGB32f+16];
	mul.ftz.f32 	%f178, %f177, %f172;
	ld.const.f32 	%f179, [k601YPbPr_To_RGB32f+4];
	mul.ftz.f32 	%f180, %f179, %f172;
	ld.const.f32 	%f181, [k601YPbPr_To_RGB32f+24];
	fma.rn.ftz.f32 	%f182, %f181, %f12, %f176;
	ld.const.f32 	%f183, [k601YPbPr_To_RGB32f+12];
	fma.rn.ftz.f32 	%f184, %f183, %f12, %f178;
	ld.const.f32 	%f185, [k601YPbPr_To_RGB32f+0];
	fma.rn.ftz.f32 	%f186, %f185, %f12, %f180;
	ld.const.f32 	%f187, [k601YPbPr_To_RGB32f+32];
	fma.rn.ftz.f32 	%f188, %f187, %f174, %f182;
	ld.const.f32 	%f189, [k601YPbPr_To_RGB32f+20];
	fma.rn.ftz.f32 	%f190, %f189, %f174, %f184;
	ld.const.f32 	%f191, [k601YPbPr_To_RGB32f+8];
	fma.rn.ftz.f32 	%f192, %f191, %f174, %f186;
	@!%p2 bra 	$Lt_31_26370;
	.loc	21	126	0
	mul.lo.u64 	%rd12, %rd1, 8;
	add.u64 	%rd13, %rd2, %rd12;
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f188;
	mov.b32		%r31, %b1; }
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f190;
	mov.b32		%r32, %b1; }
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f192;
	mov.b32		%r33, %b1; }
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f4;
	mov.b32		%r34, %b1; }
	st.global.v4.u16 	[%rd13+0], {%r31,%r32,%r33,%r34};
	.loc	23	328	0
	bra.uni 	$Lt_31_26114;
$Lt_31_26370:
	.loc	21	126	0
	mul.lo.u64 	%rd14, %rd1, 16;
	add.u64 	%rd15, %rd2, %rd14;
	st.global.v4.f32 	[%rd15+0], {%f188,%f190,%f192,%f4};
$Lt_31_26114:
$Lt_31_19970:
	.loc	23	330	0
	exit;
$LDWend_LumaCorrector_CompositeKernel:
	} // LumaCorrector_CompositeKernel
	.const .align 4 .b8 kRGB32f_To_601YCbCr[36] = {70,246,130,66,145,141,0,67,94,186,199,65,33,48,23,194,240,103,148,194,0,0,224,66,0,0,224,66,111,146,187,194,70,182,145,193};
	.const .align 4 .b8 k601YCbCr_To_RGB32f[36] = {37,160,149,59,0,0,0,0,182,23,205,59,37,160,149,59,40,15,201,186,156,239,80,187,37,160,149,59,236,155,1,60,0,0,0,0};
	.const .align 4 .b8 kRGB8u_To_601YCbCr[36] = {219,121,131,62,152,14,1,63,18,131,200,61,174,199,23,190,238,252,148,190,197,224,224,62,197,224,224,62,217,78,188,190,174,71,146,189};
	.const .align 4 .b8 k601YCbCr_To_RGB8u[36] = {127,10,149,63,0,0,0,0,160,74,204,63,127,10,149,63,254,148,200,190,184,30,80,191,127,10,149,63,78,26,1,64,0,0,0,0};
	.const .align 4 .b8 kRGB8u_To_601YCbCrFullRange[36] = {135,22,153,62,162,69,22,63,213,120,233,61,166,27,44,190,39,241,168,190,250,254,254,62,250,254,254,62,43,135,213,190,59,223,165,189};
	.const .align 4 .b8 k601YCbCrFullRange_To_RGB8u[36] = {0,0,128,63,0,0,0,0,72,193,178,63,0,0,128,63,143,130,175,190,225,26,54,191,0,0,128,63,20,238,225,63,0,0,0,0};
	.const .align 4 .b8 kRGB32f_To_601YCbCrFullRange[36] = {113,125,152,66,92,175,21,67,92,143,232,65,158,111,43,194,49,72,168,194,0,0,254,66,0,0,254,66,170,177,212,194,88,57,165,193};
	.const .align 4 .b8 k601YCbCrFullRange_To_RGB32f[36] = {129,128,128,59,0,0,0,0,188,116,179,59,129,128,128,59,194,50,176,186,179,209,54,187,129,128,128,59,229,208,226,59,0,0,0,0};
	.const .align 4 .b8 kRGB32f_To_709YPbPr[36] = {208,179,89,62,89,23,55,63,152,221,147,61,186,164,234,189,210,86,197,190,0,0,0,63,0,0,0,63,190,134,232,190,16,202,59,189};
	.const .align 4 .b8 k709YPbPr_To_RGB32f[36] = {0,0,128,63,0,0,0,0,12,147,201,63,0,0,128,63,221,209,63,190,243,173,239,190,0,0,128,63,77,132,237,63,0,0,0,0};
	.const .align 4 .b8 kRGB32f_To_709YCbCr[36] = {106,60,58,66,6,161,28,67,244,253,124,65,223,79,205,193,8,172,172,194,0,0,224,66,0,0,224,66,195,117,203,194,236,81,36,193};
	.const .align 4 .b8 k709YCbCr_To_RGB32f[36] = {37,160,149,59,0,0,0,0,239,94,230,59,37,160,149,59,33,57,91,186,178,245,8,187,37,160,149,59,82,185,7,60,0,0,0,0};
	.const .align 4 .b8 kRGB8u_To_709YCbCr[36] = {207,247,58,62,53,62,29,63,231,251,125,61,147,24,206,61,23,89,173,190,197,224,224,62,197,224,224,62,12,66,204,190,195,245,36,189};
	.const .align 4 .b8 k709YCbCr_To_RGB8u[36] = {127,10,149,63,0,0,0,0,147,120,229,63,127,10,149,63,53,94,90,190,205,108,8,191,127,10,149,63,154,49,7,64,0,0,0,0};
	.const .align 4 .b8 k709YCbCr_To_601YCbCr[36] = {0,0,128,63,23,100,203,61,1,77,68,62,0,0,0,0,18,103,125,63,10,158,226,189,0,0,0,0,61,98,148,189,249,191,123,63};
	.const .align 4 .b8 k601YCbCr_To_709YCbCr[36] = {0,0,128,63,122,165,236,189,179,237,84,190,0,0,0,0,204,98,130,63,216,188,234,61,0,0,0,0,74,179,153,61,234,61,131,63};
	.const .align 4 .b8 kYCbCrOffset[12] = {0,0,128,65,0,0,0,67,0,0,0,67};
	.const .align 4 .b8 kYCbCrFullRangeOffset[12] = {0,0,0,0,0,0,0,67,0,0,0,67};
	.const .align 4 .b8 kRGB32f_To_YIQ[36] = {135,22,153,62,162,69,22,63,213,120,233,61,216,128,24,63,27,133,140,190,149,124,164,190,236,135,88,62,134,200,5,191,22,77,159,62};
	.const .align 4 .b8 kYIQ_To_RGB32f[36] = {0,0,128,63,20,208,116,63,219,249,30,63,0,0,128,63,177,80,139,190,2,188,37,191,0,0,128,63,45,178,141,191,85,48,218,63};

