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

	//-----------------------------------------------------------
	// Compiling C:\Users\dvaeng\AppData\Local\Temp/tmpxft_0001dac4_00000000-9_CrossDissolve.cpp3.i (C:/Users/dvaeng/AppData/Local/Temp/ccBI#.a21784)
	//-----------------------------------------------------------

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

	.file	1	"C:\Users\dvaeng\AppData\Local\Temp/tmpxft_0001dac4_00000000-8_CrossDissolve.cudafe2.gpu"
	.file	2	"c:\scully64\shared\adobe\MediaCore\GPUFoundation\API\Inc\GPUFoundation/KernelSupport/PixelRGB.h"
	.file	3	"C:\Program Files (x86)\Microsoft Visual Studio 9.0\VC\include\crtdefs.h"
	.file	4	"c:\scully64\shared\adobe\MediaCore\External\3rdParty\NVIDIA\CUDA\win\64\include\crt/device_runtime.h"
	.file	5	"c:\scully64\shared\adobe\MediaCore\External\3rdParty\NVIDIA\CUDA\win\64\include\host_defines.h"
	.file	6	"c:\scully64\shared\adobe\MediaCore\External\3rdParty\NVIDIA\CUDA\win\64\include\builtin_types.h"
	.file	7	"c:\scully64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\64\include\device_types.h"
	.file	8	"c:\scully64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\64\include\driver_types.h"
	.file	9	"c:\scully64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\64\include\texture_types.h"
	.file	10	"c:\scully64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\64\include\vector_types.h"
	.file	11	"c:\scully64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\64\include\host_defines.h"
	.file	12	"c:\scully64\shared\adobe\MediaCore\External\3rdParty\NVIDIA\CUDA\win\64\include\device_launch_parameters.h"
	.file	13	"c:\scully64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\64\include\crt\storage_class.h"
	.file	14	"C:\Program Files (x86)\Microsoft Visual Studio 9.0\VC\include\time.h"
	.file	15	"c:\scully64\shared\adobe\MediaCore\GPUFoundation\API\Inc\GPUFoundation/KernelSupport/Utils.h"
	.file	16	"c:/scully64/shared/adobe/MediaCore/GPUFoundation/Src/ImageProcessing/CrossDissolve.cu"
	.file	17	"c:\scully64\shared\adobe\MediaCore\External\3rdParty\NVIDIA\CUDA\win\64\include\common_functions.h"
	.file	18	"c:\scully64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\64\include\crt/func_macro.h"
	.file	19	"c:\scully64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\64\include\math_functions.h"
	.file	20	"c:\scully64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\64\include\device_functions.h"
	.file	21	"c:\scully64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\64\include\math_constants.h"
	.file	22	"c:\scully64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\64\include\sm_11_atomic_functions.h"
	.file	23	"c:\scully64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\64\include\sm_12_atomic_functions.h"
	.file	24	"c:\scully64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\64\include\sm_13_double_functions.h"
	.file	25	"c:\scully64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\64\include\common_types.h"
	.file	26	"c:\scully64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\64\include\texture_fetch_functions.h"
	.file	27	"c:\scully64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\64\include\math_functions_dbl_ptx1.h"


	.entry CrossDissolveKernel (
		.param .u64 __cudaparm_CrossDissolveKernel_inSrc0,
		.param .s32 __cudaparm_CrossDissolveKernel_inSrcPitch0,
		.param .u64 __cudaparm_CrossDissolveKernel_inSrc1,
		.param .s32 __cudaparm_CrossDissolveKernel_inSrcPitch1,
		.param .u64 __cudaparm_CrossDissolveKernel_inDest,
		.param .s32 __cudaparm_CrossDissolveKernel_inDestPitch,
		.param .s32 __cudaparm_CrossDissolveKernel_inWidth,
		.param .s32 __cudaparm_CrossDissolveKernel_inHeight,
		.param .f32 __cudaparm_CrossDissolveKernel_inDissolve)
	{
	.reg .u32 %r<29>;
	.reg .u64 %rd<14>;
	.reg .f32 %f<41>;
	.reg .pred %p<6>;
	.loc	16	39	0
$LBB1_CrossDissolveKernel:
	.loc	16	42	0
	cvt.s32.u16 	%r1, %ctaid.x;
	cvt.s32.u16 	%r2, %ntid.x;
	mul24.lo.s32 	%r3, %r1, %r2;
	cvt.s32.u16 	%r4, %ctaid.y;
	cvt.s32.u16 	%r5, %ntid.y;
	mul24.lo.s32 	%r6, %r4, %r5;
	cvt.u32.u16 	%r7, %tid.x;
	add.u32 	%r8, %r3, %r7;
	cvt.u32.u16 	%r9, %tid.y;
	add.u32 	%r10, %r6, %r9;
	ld.param.s32 	%r11, [__cudaparm_CrossDissolveKernel_inWidth];
	set.gt.u32.s32 	%r12, %r11, %r8;
	neg.s32 	%r13, %r12;
	ld.param.s32 	%r14, [__cudaparm_CrossDissolveKernel_inHeight];
	set.gt.u32.s32 	%r15, %r14, %r10;
	neg.s32 	%r16, %r15;
	and.b32 	%r17, %r13, %r16;
	mov.u32 	%r18, 0;
	setp.eq.s32 	%p1, %r17, %r18;
	@%p1 bra 	$Lt_0_4354;
	.loc	15	89	0
	ld.param.s32 	%r19, [__cudaparm_CrossDissolveKernel_inSrcPitch0];
	mul24.lo.s32 	%r20, %r10, %r19;
	ld.param.u64 	%rd1, [__cudaparm_CrossDissolveKernel_inSrc0];
	add.s32 	%r21, %r20, %r8;
	cvt.s64.s32 	%rd2, %r21;
	mul.lo.u64 	%rd3, %rd2, 16;
	add.u64 	%rd4, %rd1, %rd3;
	ld.global.v4.f32 	{%f1,%f2,%f3,%f4}, [%rd4+0];
	ld.param.s32 	%r22, [__cudaparm_CrossDissolveKernel_inSrcPitch1];
	mul24.lo.s32 	%r23, %r10, %r22;
	ld.param.u64 	%rd5, [__cudaparm_CrossDissolveKernel_inSrc1];
	add.s32 	%r24, %r23, %r8;
	cvt.s64.s32 	%rd6, %r24;
	mul.lo.u64 	%rd7, %rd6, 16;
	add.u64 	%rd8, %rd5, %rd7;
	ld.global.v4.f32 	{%f5,%f6,%f7,%f8}, [%rd8+0];
	.loc	16	47	0
	ld.param.f32 	%f9, [__cudaparm_CrossDissolveKernel_inDissolve];
	mov.f32 	%f10, 0fb70637bd;    	// -8e-006
	add.f32 	%f11, %f4, %f10;
	mov.f32 	%f12, 0f00000000;    	// 0
	setp.le.f32 	%p2, %f11, %f12;
	@!%p2 bra 	$Lt_0_5122;
	.loc	16	54	0
	mov.f32 	%f13, %f5;
	.loc	16	55	0
	mov.f32 	%f14, %f6;
	.loc	16	56	0
	mov.f32 	%f15, %f7;
	.loc	16	57	0
	mov.f32 	%f16, 0f3f800000;    	// 1
	sub.f32 	%f17, %f16, %f9;
	mul.f32 	%f18, %f8, %f17;
	bra.uni 	$Lt_0_4866;
$Lt_0_5122:
	mov.f32 	%f19, 0fb70637bd;    	// -8e-006
	add.f32 	%f20, %f8, %f19;
	mov.f32 	%f21, 0f00000000;    	// 0
	setp.le.f32 	%p3, %f20, %f21;
	@!%p3 bra 	$Lt_0_5634;
	.loc	16	62	0
	mov.f32 	%f13, %f1;
	.loc	16	63	0
	mov.f32 	%f14, %f2;
	.loc	16	64	0
	mov.f32 	%f15, %f3;
	.loc	16	65	0
	mul.f32 	%f18, %f4, %f9;
	bra.uni 	$Lt_0_5378;
$Lt_0_5634:
	.loc	16	67	0
	mov.f32 	%f22, 0f3f800000;    	// 1
	sub.f32 	%f23, %f22, %f9;
	sub.f32 	%f24, %f4, %f8;
	abs.f32 	%f25, %f24;
	mov.f32 	%f26, 0f370637bd;    	// 8e-006
	setp.lt.f32 	%p4, %f25, %f26;
	@!%p4 bra 	$Lt_0_6146;
	.loc	16	70	0
	mul.f32 	%f27, %f23, %f5;
	mad.f32 	%f13, %f9, %f1, %f27;
	.loc	16	71	0
	mul.f32 	%f28, %f23, %f6;
	mad.f32 	%f14, %f9, %f2, %f28;
	.loc	16	72	0
	mul.f32 	%f29, %f23, %f7;
	mad.f32 	%f15, %f9, %f3, %f29;
	.loc	16	73	0
	mov.f32 	%f18, %f4;
	bra.uni 	$Lt_0_5890;
$Lt_0_6146:
	.loc	16	81	0
	mul.f32 	%f30, %f4, %f9;
	mul.f32 	%f31, %f23, %f8;
	add.f32 	%f32, %f30, %f31;
	rcp.approx.f32 	%f33, %f32;
	mul.f32 	%f34, %f31, %f5;
	mad.f32 	%f35, %f30, %f1, %f34;
	mul.f32 	%f13, %f33, %f35;
	.loc	16	82	0
	mul.f32 	%f36, %f31, %f6;
	mad.f32 	%f37, %f30, %f2, %f36;
	mul.f32 	%f14, %f33, %f37;
	.loc	16	83	0
	mul.f32 	%f38, %f31, %f7;
	mad.f32 	%f39, %f30, %f3, %f38;
	mul.f32 	%f15, %f33, %f39;
	.loc	16	84	0
	mov.f32 	%f18, %f32;
$Lt_0_5890:
$Lt_0_5378:
$Lt_0_4866:
	.loc	15	100	0
	ld.param.s32 	%r25, [__cudaparm_CrossDissolveKernel_inDestPitch];
	mul24.lo.s32 	%r26, %r10, %r25;
	ld.param.u64 	%rd9, [__cudaparm_CrossDissolveKernel_inDest];
	add.s32 	%r27, %r26, %r8;
	cvt.s64.s32 	%rd10, %r27;
	mul.lo.u64 	%rd11, %rd10, 16;
	add.u64 	%rd12, %rd9, %rd11;
	st.global.v4.f32 	[%rd12+0], {%f13,%f14,%f15,%f18};
$Lt_0_4354:
	.loc	16	89	0
	exit;
$LDWend_CrossDissolveKernel:
	} // CrossDissolveKernel

