	.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_0001d9b8_00000000-9_DropShadow.cpp3.i (C:/Users/dvaeng/AppData/Local/Temp/ccBI#.a21492)
	//-----------------------------------------------------------

	//-----------------------------------------------------------
	// 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_0001d9b8_00000000-8_DropShadow.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/Display/Src/CUDA/Effects/DropShadow.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 CreateShadowMask (
		.param .u64 __cudaparm_CreateShadowMask_inSrc,
		.param .s32 __cudaparm_CreateShadowMask_inSrcPitch,
		.param .s32 __cudaparm_CreateShadowMask_inSrcWidth,
		.param .s32 __cudaparm_CreateShadowMask_inSrcHeight,
		.param .u64 __cudaparm_CreateShadowMask_inMask,
		.param .s32 __cudaparm_CreateShadowMask_inMaskPitch,
		.param .s32 __cudaparm_CreateShadowMask_inMaskWidth,
		.param .s32 __cudaparm_CreateShadowMask_inMaskHeight,
		.param .s32 __cudaparm_CreateShadowMask_inShadowOffsetX,
		.param .s32 __cudaparm_CreateShadowMask_inShadowOffsetY)
	{
	.reg .u32 %r<34>;
	.reg .u64 %rd<10>;
	.reg .f32 %f<4>;
	.reg .pred %p<7>;
	.loc	16	39	0
$LBB1_CreateShadowMask:
	.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_CreateShadowMask_inMaskWidth];
	set.gt.u32.s32 	%r12, %r11, %r8;
	neg.s32 	%r13, %r12;
	ld.param.s32 	%r14, [__cudaparm_CreateShadowMask_inMaskHeight];
	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_4098;
	.loc	16	45	0
	ld.param.s32 	%r19, [__cudaparm_CreateShadowMask_inShadowOffsetX];
	sub.s32 	%r20, %r8, %r19;
	mov.u32 	%r21, 0;
	setp.lt.s32 	%p2, %r20, %r21;
	@%p2 bra 	$Lt_0_5122;
	ld.param.s32 	%r22, [__cudaparm_CreateShadowMask_inSrcWidth];
	setp.le.s32 	%p3, %r22, %r20;
	@%p3 bra 	$Lt_0_5122;
	ld.param.s32 	%r23, [__cudaparm_CreateShadowMask_inShadowOffsetY];
	sub.s32 	%r24, %r10, %r23;
	mov.u32 	%r25, 0;
	setp.lt.s32 	%p4, %r24, %r25;
	@%p4 bra 	$Lt_0_5122;
	ld.param.s32 	%r26, [__cudaparm_CreateShadowMask_inSrcHeight];
	setp.le.s32 	%p5, %r26, %r24;
	@%p5 bra 	$Lt_0_5122;
	.loc	15	89	0
	ld.param.u64 	%rd1, [__cudaparm_CreateShadowMask_inSrc];
	ld.param.s32 	%r27, [__cudaparm_CreateShadowMask_inSrcPitch];
	mul24.lo.s32 	%r28, %r24, %r27;
	add.s32 	%r29, %r20, %r28;
	cvt.s64.s32 	%rd2, %r29;
	mul.lo.u64 	%rd3, %rd2, 16;
	add.u64 	%rd4, %rd1, %rd3;
	ld.global.f32 	%f1, [%rd4+12];
	.loc	16	52	0
	mov.f32 	%f2, %f1;
	bra.uni 	$L_0_2818;
$Lt_0_5122:
$L_0_3074:
	.loc	16	56	0
	mov.f32 	%f2, 0f00000000;     	// 0
$L_0_2818:
	.loc	15	100	0
	ld.param.u64 	%rd5, [__cudaparm_CreateShadowMask_inMask];
	ld.param.s32 	%r30, [__cudaparm_CreateShadowMask_inMaskPitch];
	mul24.lo.s32 	%r31, %r10, %r30;
	add.s32 	%r32, %r8, %r31;
	cvt.s64.s32 	%rd6, %r32;
	mul.lo.u64 	%rd7, %rd6, 4;
	add.u64 	%rd8, %rd5, %rd7;
	st.global.f32 	[%rd8+0], %f2;
$Lt_0_4098:
	.loc	16	61	0
	exit;
$LDWend_CreateShadowMask:
	} // CreateShadowMask

	.entry CompositeShadowMask (
		.param .u64 __cudaparm_CompositeShadowMask_inSrc,
		.param .s32 __cudaparm_CompositeShadowMask_inSrcPitch,
		.param .s32 __cudaparm_CompositeShadowMask_inSrcWidth,
		.param .s32 __cudaparm_CompositeShadowMask_inSrcHeight,
		.param .u64 __cudaparm_CompositeShadowMask_inMask,
		.param .s32 __cudaparm_CompositeShadowMask_inMaskPitch,
		.param .u64 __cudaparm_CompositeShadowMask_inDest,
		.param .s32 __cudaparm_CompositeShadowMask_inDestPitch,
		.param .s32 __cudaparm_CompositeShadowMask_inDestWidth,
		.param .s32 __cudaparm_CompositeShadowMask_inDestHeight,
		.param .s32 __cudaparm_CompositeShadowMask_inSrcOffsetX,
		.param .s32 __cudaparm_CompositeShadowMask_inSrcOffsetY,
		.param .f32 __cudaparm_CompositeShadowMask_inColorB,
		.param .f32 __cudaparm_CompositeShadowMask_inColorG,
		.param .f32 __cudaparm_CompositeShadowMask_inColorR,
		.param .f32 __cudaparm_CompositeShadowMask_inOpacity,
		.param .s8 __cudaparm_CompositeShadowMask_inShadowOnly)
	{
	.reg .u32 %r<39>;
	.reg .u64 %rd<14>;
	.reg .f32 %f<35>;
	.reg .pred %p<9>;
	.loc	16	83	0
$LBB1_CompositeShadowMask:
	.loc	16	86	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_CompositeShadowMask_inDestWidth];
	set.gt.u32.s32 	%r12, %r11, %r8;
	neg.s32 	%r13, %r12;
	ld.param.s32 	%r14, [__cudaparm_CompositeShadowMask_inDestHeight];
	set.gt.u32.s32 	%r15, %r14, %r10;
	neg.s32 	%r16, %r15;
	and.b32 	%r17, %r13, %r16;
	mov.u32 	%r18, 0;
	setp.eq.s32 	%p1, %r17, %r18;
	@%p1 bra 	$Lt_1_5634;
	.loc	15	89	0
	ld.param.u64 	%rd1, [__cudaparm_CompositeShadowMask_inMask];
	ld.param.s32 	%r19, [__cudaparm_CompositeShadowMask_inMaskPitch];
	mul24.lo.s32 	%r20, %r10, %r19;
	add.s32 	%r21, %r8, %r20;
	cvt.s64.s32 	%rd2, %r21;
	mul.lo.u64 	%rd3, %rd2, 4;
	add.u64 	%rd4, %rd1, %rd3;
	ld.global.f32 	%f1, [%rd4+0];
	.loc	16	90	0
	ld.param.f32 	%f2, [__cudaparm_CompositeShadowMask_inOpacity];
	mul.f32 	%f3, %f2, %f1;
	ld.param.s8 	%r22, [__cudaparm_CompositeShadowMask_inShadowOnly];
	mov.u32 	%r23, 0;
	setp.eq.s32 	%p2, %r22, %r23;
	@%p2 bra 	$Lt_1_6402;
	.loc	16	96	0
	ld.param.f32 	%f4, [__cudaparm_CompositeShadowMask_inColorB];
	ld.param.f32 	%f5, [__cudaparm_CompositeShadowMask_inColorG];
	ld.param.f32 	%f6, [__cudaparm_CompositeShadowMask_inColorR];
	mov.f32 	%f7, %f3;
	bra.uni 	$Lt_1_6658;
$Lt_1_6402:
	.loc	16	100	0
	ld.param.s32 	%r24, [__cudaparm_CompositeShadowMask_inSrcOffsetX];
	sub.s32 	%r25, %r8, %r24;
	mov.u32 	%r26, 0;
	setp.lt.s32 	%p3, %r25, %r26;
	@%p3 bra 	$Lt_1_7682;
	ld.param.s32 	%r27, [__cudaparm_CompositeShadowMask_inSrcWidth];
	setp.le.s32 	%p4, %r27, %r25;
	@%p4 bra 	$Lt_1_7682;
	ld.param.s32 	%r28, [__cudaparm_CompositeShadowMask_inSrcOffsetY];
	sub.s32 	%r29, %r10, %r28;
	mov.u32 	%r30, 0;
	setp.lt.s32 	%p5, %r29, %r30;
	@%p5 bra 	$Lt_1_7682;
	ld.param.s32 	%r31, [__cudaparm_CompositeShadowMask_inSrcHeight];
	setp.le.s32 	%p6, %r31, %r29;
	@%p6 bra 	$Lt_1_7682;
	.loc	15	89	0
	ld.param.s32 	%r32, [__cudaparm_CompositeShadowMask_inSrcPitch];
	mul24.lo.s32 	%r33, %r29, %r32;
	ld.param.u64 	%rd5, [__cudaparm_CompositeShadowMask_inSrc];
	add.s32 	%r34, %r25, %r33;
	cvt.s64.s32 	%rd6, %r34;
	mul.lo.u64 	%rd7, %rd6, 16;
	add.u64 	%rd8, %rd5, %rd7;
	ld.global.v4.f32 	{%f8,%f9,%f10,%f11}, [%rd8+0];
	.loc	16	107	0
	mov.f32 	%f12, %f8;
	mov.f32 	%f13, %f9;
	mov.f32 	%f14, %f10;
	mov.f32 	%f15, %f11;
	bra.uni 	$L_1_4354;
$Lt_1_7682:
$L_1_4610:
	.loc	16	111	0
	mov.f32 	%f15, 0f00000000;    	// 0
	mov.f32 	%f14, 0f00000000;    	// 0
	mov.f32 	%f13, 0f00000000;    	// 0
	mov.f32 	%f12, 0f00000000;    	// 0
$L_1_4354:
	.loc	16	114	0
	cvt.sat.f32.f32 	%f15, %f15;
	mov.f32 	%f16, 0f3f800000;    	// 1
	sub.f32 	%f17, %f16, %f15;
	cvt.sat.f32.f32 	%f18, %f3;
	mul.f32 	%f19, %f17, %f18;
	add.f32 	%f20, %f19, %f15;
	mov.f32 	%f21, 0fb70637bd;    	// -8e-006
	add.f32 	%f22, %f20, %f21;
	mov.f32 	%f23, 0f00000000;    	// 0
	setp.le.f32 	%p7, %f22, %f23;
	@!%p7 bra 	$Lt_1_6914;
	mov.f32 	%f6, 0f00000000;     	// 0
	mov.f32 	%f5, 0f00000000;     	// 0
	mov.f32 	%f4, 0f00000000;     	// 0
	mov.f32 	%f7, 0f00000000;     	// 0
	bra.uni 	$Lt_1_6658;
$Lt_1_6914:
	.loc	16	128	0
	mov.f32 	%f7, %f20;
	.loc	16	129	0
	rcp.approx.f32 	%f24, %f20;
	mul.f32 	%f25, %f15, %f12;
	ld.param.f32 	%f26, [__cudaparm_CompositeShadowMask_inColorB];
	mad.f32 	%f27, %f26, %f19, %f25;
	mul.f32 	%f4, %f24, %f27;
	.loc	16	130	0
	mul.f32 	%f28, %f15, %f13;
	ld.param.f32 	%f29, [__cudaparm_CompositeShadowMask_inColorG];
	mad.f32 	%f30, %f19, %f29, %f28;
	mul.f32 	%f5, %f24, %f30;
	.loc	16	131	0
	mul.f32 	%f31, %f15, %f14;
	ld.param.f32 	%f32, [__cudaparm_CompositeShadowMask_inColorR];
	mad.f32 	%f33, %f19, %f32, %f31;
	mul.f32 	%f6, %f24, %f33;
$Lt_1_6658:
$Lt_1_6146:
	.loc	15	100	0
	ld.param.s32 	%r35, [__cudaparm_CompositeShadowMask_inDestPitch];
	mul24.lo.s32 	%r36, %r10, %r35;
	ld.param.u64 	%rd9, [__cudaparm_CompositeShadowMask_inDest];
	add.s32 	%r37, %r36, %r8;
	cvt.s64.s32 	%rd10, %r37;
	mul.lo.u64 	%rd11, %rd10, 16;
	add.u64 	%rd12, %rd9, %rd11;
	st.global.v4.f32 	[%rd12+0], {%f4,%f5,%f6,%f7};
$Lt_1_5634:
	.loc	16	137	0
	exit;
$LDWend_CompositeShadowMask:
	} // CompositeShadowMask

