	.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_0001d6a4_00000000-9_GarbageMatte.cpp3.i (C:/Users/dvaeng/AppData/Local/Temp/ccBI#.a21320)
	//-----------------------------------------------------------

	//-----------------------------------------------------------
	// 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_0001d6a4_00000000-8_GarbageMatte.cudafe2.gpu"
	.file	2	"C:\Program Files (x86)\Microsoft Visual Studio 9.0\VC\include\crtdefs.h"
	.file	3	"c:\scully64\shared\adobe\MediaCore\External\3rdParty\NVIDIA\CUDA\win\64\include\crt/device_runtime.h"
	.file	4	"c:\scully64\shared\adobe\MediaCore\External\3rdParty\NVIDIA\CUDA\win\64\include\host_defines.h"
	.file	5	"c:\scully64\shared\adobe\MediaCore\External\3rdParty\NVIDIA\CUDA\win\64\include\builtin_types.h"
	.file	6	"c:\scully64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\64\include\device_types.h"
	.file	7	"c:\scully64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\64\include\driver_types.h"
	.file	8	"c:\scully64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\64\include\texture_types.h"
	.file	9	"c:\scully64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\64\include\vector_types.h"
	.file	10	"c:\scully64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\64\include\host_defines.h"
	.file	11	"c:\scully64\shared\adobe\MediaCore\External\3rdParty\NVIDIA\CUDA\win\64\include\device_launch_parameters.h"
	.file	12	"c:\scully64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\64\include\crt\storage_class.h"
	.file	13	"C:\Program Files (x86)\Microsoft Visual Studio 9.0\VC\include\time.h"
	.file	14	"c:\scully64\shared\adobe\MediaCore\GPUFoundation\API\Inc\GPUFoundation/KernelSupport/Utils.h"
	.file	15	"c:/scully64/shared/adobe/MediaCore/Display/Src/CUDA/Effects/GarbageMatte.cu"
	.file	16	"c:\scully64\shared\adobe\MediaCore\External\3rdParty\NVIDIA\CUDA\win\64\include\common_functions.h"
	.file	17	"c:\scully64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\64\include\crt/func_macro.h"
	.file	18	"c:\scully64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\64\include\math_functions.h"
	.file	19	"c:\scully64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\64\include\device_functions.h"
	.file	20	"c:\scully64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\64\include\math_constants.h"
	.file	21	"c:\scully64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\64\include\sm_11_atomic_functions.h"
	.file	22	"c:\scully64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\64\include\sm_12_atomic_functions.h"
	.file	23	"c:\scully64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\64\include\sm_13_double_functions.h"
	.file	24	"c:\scully64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\64\include\common_types.h"
	.file	25	"c:\scully64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\64\include\texture_fetch_functions.h"
	.file	26	"c:\scully64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\64\include\math_functions_dbl_ptx1.h"

	.const .align 4 .b8 polyPoints[204];

	.entry cuda_kernel_bitmask (
		.param .u64 __cudaparm_cuda_kernel_bitmask_dstFrame,
		.param .s32 __cudaparm_cuda_kernel_bitmask_width,
		.param .s32 __cudaparm_cuda_kernel_bitmask_height,
		.param .s32 __cudaparm_cuda_kernel_bitmask_dstPitch,
		.param .s32 __cudaparm_cuda_kernel_bitmask_count)
	{
	.reg .u32 %r<101>;
	.reg .u64 %rd<17>;
	.reg .f32 %f<389>;
	.reg .pred %p<162>;
	.loc	15	85	0
$LBB1_cuda_kernel_bitmask:
	.loc	15	88	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;
	mov.s32 	%r11, 32;
	mul24.lo.s32 	%r12, %r8, %r11;
	ld.param.s32 	%r13, [__cudaparm_cuda_kernel_bitmask_width];
	set.ge.u32.s32 	%r14, %r12, %r13;
	neg.s32 	%r15, %r14;
	ld.param.s32 	%r16, [__cudaparm_cuda_kernel_bitmask_height];
	set.le.u32.s32 	%r17, %r16, %r10;
	neg.s32 	%r18, %r17;
	or.b32 	%r19, %r15, %r18;
	mov.u32 	%r20, 0;
	setp.eq.s32 	%p1, %r19, %r20;
	@%p1 bra 	$Lt_0_46594;
	bra.uni 	$LBB160_cuda_kernel_bitmask;
$Lt_0_46594:
	.loc	15	95	0
	sub.s32 	%r21, %r13, 32;
	setp.lt.s32 	%p2, %r21, %r12;
	selp.s32 	%r22, %r21, %r12, %p2;
	mov.s32 	%r23, 16;
	mul24.lo.s32 	%r24, %r22, %r23;
	ld.param.s32 	%r25, [__cudaparm_cuda_kernel_bitmask_dstPitch];
	mul24.lo.s32 	%r26, %r10, %r25;
	add.s32 	%r27, %r24, %r26;
	cvt.u64.s32 	%rd1, %r27;
	ld.param.u64 	%rd2, [__cudaparm_cuda_kernel_bitmask_dstFrame];
	add.u64 	%rd3, %rd1, %rd2;
	ld.param.s32 	%r28, [__cudaparm_cuda_kernel_bitmask_count];
	mov.u32 	%r29, 16;
	setp.ne.s32 	%p3, %r28, %r29;
	@%p3 bra 	$Lt_0_47362;
	.loc	15	97	0
	mov.s64 	%rd4, %rd3;
	.loc	15	50	0
	mov.f32 	%f1, 0fbf800000;     	// -1
	mov.f32 	%f2, %f1;
	mov.f32 	%f3, 0fbf800000;     	// -1
	mov.f32 	%f4, %f3;
	mov.f32 	%f5, 0fbf800000;     	// -1
	mov.f32 	%f6, %f5;
	mov.f32 	%f7, 0fbf800000;     	// -1
	mov.f32 	%f8, %f7;
	mov.f32 	%f9, 0fbf800000;     	// -1
	mov.f32 	%f10, %f9;
	mov.f32 	%f11, 0fbf800000;    	// -1
	mov.f32 	%f12, %f11;
	mov.f32 	%f13, 0fbf800000;    	// -1
	mov.f32 	%f14, %f13;
	mov.f32 	%f15, 0fbf800000;    	// -1
	mov.f32 	%f16, %f15;
	mov.f32 	%f17, 0fbf800000;    	// -1
	mov.f32 	%f18, %f17;
	mov.f32 	%f19, 0fbf800000;    	// -1
	mov.f32 	%f20, %f19;
	mov.f32 	%f21, 0fbf800000;    	// -1
	mov.f32 	%f22, %f21;
	mov.f32 	%f23, 0fbf800000;    	// -1
	mov.f32 	%f24, %f23;
	mov.f32 	%f25, 0fbf800000;    	// -1
	mov.f32 	%f26, %f25;
	mov.f32 	%f27, 0fbf800000;    	// -1
	mov.f32 	%f28, %f27;
	mov.f32 	%f29, 0fbf800000;    	// -1
	mov.f32 	%f30, %f29;
	mov.f32 	%f31, 0fbf800000;    	// -1
	mov.f32 	%f32, %f31;
	.loc	15	24	0
	cvt.rn.f32.s32 	%f33, %r10;
	ld.const.f32 	%f34, [polyPoints+4];
	ld.const.f32 	%f35, [polyPoints+16];
	setp.ge.f32 	%p4, %f33, %f34;
	setp.lt.f32 	%p5, %f33, %f35;
	xor.pred 	%p6, %p4, %p5;
	@%p6 bra 	$Lt_0_47618;
	.loc	15	40	0
	cvt.rn.f32.s32 	%f36, %r22;
	ld.const.f32 	%f37, [polyPoints+0];
	sub.f32 	%f38, %f36, %f37;
	ld.const.f32 	%f39, [polyPoints+8];
	sub.f32 	%f40, %f33, %f34;
	mul.f32 	%f41, %f39, %f40;
	sub.f32 	%f42, %f41, %f38;
	mov.f32 	%f2, %f42;
$Lt_0_47618:
	ld.const.f32 	%f43, [polyPoints+28];
	setp.ge.f32 	%p7, %f33, %f35;
	setp.lt.f32 	%p8, %f33, %f43;
	xor.pred 	%p9, %p7, %p8;
	@%p9 bra 	$Lt_0_48130;
	cvt.rn.f32.s32 	%f44, %r22;
	ld.const.f32 	%f45, [polyPoints+12];
	sub.f32 	%f46, %f44, %f45;
	ld.const.f32 	%f47, [polyPoints+20];
	sub.f32 	%f48, %f33, %f35;
	mul.f32 	%f49, %f47, %f48;
	sub.f32 	%f50, %f49, %f46;
	mov.f32 	%f4, %f50;
$Lt_0_48130:
	ld.const.f32 	%f51, [polyPoints+40];
	setp.ge.f32 	%p10, %f33, %f43;
	setp.lt.f32 	%p11, %f33, %f51;
	xor.pred 	%p12, %p10, %p11;
	@%p12 bra 	$Lt_0_48642;
	cvt.rn.f32.s32 	%f52, %r22;
	ld.const.f32 	%f53, [polyPoints+24];
	sub.f32 	%f54, %f52, %f53;
	ld.const.f32 	%f55, [polyPoints+32];
	sub.f32 	%f56, %f33, %f43;
	mul.f32 	%f57, %f55, %f56;
	sub.f32 	%f58, %f57, %f54;
	mov.f32 	%f6, %f58;
$Lt_0_48642:
	ld.const.f32 	%f59, [polyPoints+52];
	setp.ge.f32 	%p13, %f33, %f51;
	setp.lt.f32 	%p14, %f33, %f59;
	xor.pred 	%p15, %p13, %p14;
	@%p15 bra 	$Lt_0_49154;
	cvt.rn.f32.s32 	%f60, %r22;
	ld.const.f32 	%f61, [polyPoints+36];
	sub.f32 	%f62, %f60, %f61;
	ld.const.f32 	%f63, [polyPoints+44];
	sub.f32 	%f64, %f33, %f51;
	mul.f32 	%f65, %f63, %f64;
	sub.f32 	%f66, %f65, %f62;
	mov.f32 	%f8, %f66;
$Lt_0_49154:
	ld.const.f32 	%f67, [polyPoints+64];
	setp.ge.f32 	%p16, %f33, %f59;
	setp.lt.f32 	%p17, %f33, %f67;
	xor.pred 	%p18, %p16, %p17;
	@%p18 bra 	$Lt_0_49666;
	cvt.rn.f32.s32 	%f68, %r22;
	ld.const.f32 	%f69, [polyPoints+48];
	sub.f32 	%f70, %f68, %f69;
	ld.const.f32 	%f71, [polyPoints+56];
	sub.f32 	%f72, %f33, %f59;
	mul.f32 	%f73, %f71, %f72;
	sub.f32 	%f74, %f73, %f70;
	mov.f32 	%f10, %f74;
$Lt_0_49666:
	ld.const.f32 	%f75, [polyPoints+76];
	setp.ge.f32 	%p19, %f33, %f67;
	setp.lt.f32 	%p20, %f33, %f75;
	xor.pred 	%p21, %p19, %p20;
	@%p21 bra 	$Lt_0_50178;
	cvt.rn.f32.s32 	%f76, %r22;
	ld.const.f32 	%f77, [polyPoints+60];
	sub.f32 	%f78, %f76, %f77;
	ld.const.f32 	%f79, [polyPoints+68];
	sub.f32 	%f80, %f33, %f67;
	mul.f32 	%f81, %f79, %f80;
	sub.f32 	%f82, %f81, %f78;
	mov.f32 	%f12, %f82;
$Lt_0_50178:
	ld.const.f32 	%f83, [polyPoints+88];
	setp.ge.f32 	%p22, %f33, %f75;
	setp.lt.f32 	%p23, %f33, %f83;
	xor.pred 	%p24, %p22, %p23;
	@%p24 bra 	$Lt_0_50690;
	cvt.rn.f32.s32 	%f84, %r22;
	ld.const.f32 	%f85, [polyPoints+72];
	sub.f32 	%f86, %f84, %f85;
	ld.const.f32 	%f87, [polyPoints+80];
	sub.f32 	%f88, %f33, %f75;
	mul.f32 	%f89, %f87, %f88;
	sub.f32 	%f90, %f89, %f86;
	mov.f32 	%f14, %f90;
$Lt_0_50690:
	ld.const.f32 	%f91, [polyPoints+100];
	setp.ge.f32 	%p25, %f33, %f83;
	setp.lt.f32 	%p26, %f33, %f91;
	xor.pred 	%p27, %p25, %p26;
	@%p27 bra 	$Lt_0_51202;
	cvt.rn.f32.s32 	%f92, %r22;
	ld.const.f32 	%f93, [polyPoints+84];
	sub.f32 	%f94, %f92, %f93;
	ld.const.f32 	%f95, [polyPoints+92];
	sub.f32 	%f96, %f33, %f83;
	mul.f32 	%f97, %f95, %f96;
	sub.f32 	%f98, %f97, %f94;
	mov.f32 	%f16, %f98;
$Lt_0_51202:
	ld.const.f32 	%f99, [polyPoints+112];
	setp.ge.f32 	%p28, %f33, %f91;
	setp.lt.f32 	%p29, %f33, %f99;
	xor.pred 	%p30, %p28, %p29;
	@%p30 bra 	$Lt_0_51714;
	cvt.rn.f32.s32 	%f100, %r22;
	ld.const.f32 	%f101, [polyPoints+96];
	sub.f32 	%f102, %f100, %f101;
	ld.const.f32 	%f103, [polyPoints+104];
	sub.f32 	%f104, %f33, %f91;
	mul.f32 	%f105, %f103, %f104;
	sub.f32 	%f106, %f105, %f102;
	mov.f32 	%f18, %f106;
$Lt_0_51714:
	ld.const.f32 	%f107, [polyPoints+124];
	setp.ge.f32 	%p31, %f33, %f99;
	setp.lt.f32 	%p32, %f33, %f107;
	xor.pred 	%p33, %p31, %p32;
	@%p33 bra 	$Lt_0_52226;
	cvt.rn.f32.s32 	%f108, %r22;
	ld.const.f32 	%f109, [polyPoints+108];
	sub.f32 	%f110, %f108, %f109;
	ld.const.f32 	%f111, [polyPoints+116];
	sub.f32 	%f112, %f33, %f99;
	mul.f32 	%f113, %f111, %f112;
	sub.f32 	%f114, %f113, %f110;
	mov.f32 	%f20, %f114;
$Lt_0_52226:
	ld.const.f32 	%f115, [polyPoints+136];
	setp.ge.f32 	%p34, %f33, %f107;
	setp.lt.f32 	%p35, %f33, %f115;
	xor.pred 	%p36, %p34, %p35;
	@%p36 bra 	$Lt_0_52738;
	cvt.rn.f32.s32 	%f116, %r22;
	ld.const.f32 	%f117, [polyPoints+120];
	sub.f32 	%f118, %f116, %f117;
	ld.const.f32 	%f119, [polyPoints+128];
	sub.f32 	%f120, %f33, %f107;
	mul.f32 	%f121, %f119, %f120;
	sub.f32 	%f122, %f121, %f118;
	mov.f32 	%f22, %f122;
$Lt_0_52738:
	ld.const.f32 	%f123, [polyPoints+148];
	setp.ge.f32 	%p37, %f33, %f115;
	setp.lt.f32 	%p38, %f33, %f123;
	xor.pred 	%p39, %p37, %p38;
	@%p39 bra 	$Lt_0_53250;
	cvt.rn.f32.s32 	%f124, %r22;
	ld.const.f32 	%f125, [polyPoints+132];
	sub.f32 	%f126, %f124, %f125;
	ld.const.f32 	%f127, [polyPoints+140];
	sub.f32 	%f128, %f33, %f115;
	mul.f32 	%f129, %f127, %f128;
	sub.f32 	%f130, %f129, %f126;
	mov.f32 	%f24, %f130;
$Lt_0_53250:
	ld.const.f32 	%f131, [polyPoints+160];
	setp.ge.f32 	%p40, %f33, %f123;
	setp.lt.f32 	%p41, %f33, %f131;
	xor.pred 	%p42, %p40, %p41;
	@%p42 bra 	$Lt_0_53762;
	cvt.rn.f32.s32 	%f132, %r22;
	ld.const.f32 	%f133, [polyPoints+144];
	sub.f32 	%f134, %f132, %f133;
	ld.const.f32 	%f135, [polyPoints+152];
	sub.f32 	%f136, %f33, %f123;
	mul.f32 	%f137, %f135, %f136;
	sub.f32 	%f138, %f137, %f134;
	mov.f32 	%f26, %f138;
$Lt_0_53762:
	ld.const.f32 	%f139, [polyPoints+172];
	setp.ge.f32 	%p43, %f33, %f131;
	setp.lt.f32 	%p44, %f33, %f139;
	xor.pred 	%p45, %p43, %p44;
	@%p45 bra 	$Lt_0_54274;
	cvt.rn.f32.s32 	%f140, %r22;
	ld.const.f32 	%f141, [polyPoints+156];
	sub.f32 	%f142, %f140, %f141;
	ld.const.f32 	%f143, [polyPoints+164];
	sub.f32 	%f144, %f33, %f131;
	mul.f32 	%f145, %f143, %f144;
	sub.f32 	%f146, %f145, %f142;
	mov.f32 	%f28, %f146;
$Lt_0_54274:
	ld.const.f32 	%f147, [polyPoints+184];
	setp.ge.f32 	%p46, %f33, %f139;
	setp.lt.f32 	%p47, %f33, %f147;
	xor.pred 	%p48, %p46, %p47;
	@%p48 bra 	$Lt_0_54786;
	cvt.rn.f32.s32 	%f148, %r22;
	ld.const.f32 	%f149, [polyPoints+168];
	sub.f32 	%f150, %f148, %f149;
	ld.const.f32 	%f151, [polyPoints+176];
	sub.f32 	%f152, %f33, %f139;
	mul.f32 	%f153, %f151, %f152;
	sub.f32 	%f154, %f153, %f150;
	mov.f32 	%f30, %f154;
$Lt_0_54786:
	setp.ge.f32 	%p49, %f33, %f147;
	ld.const.f32 	%f155, [polyPoints+196];
	setp.lt.f32 	%p50, %f33, %f155;
	xor.pred 	%p51, %p49, %p50;
	@%p51 bra 	$Lt_0_55298;
	cvt.rn.f32.s32 	%f156, %r22;
	ld.const.f32 	%f157, [polyPoints+180];
	sub.f32 	%f158, %f156, %f157;
	ld.const.f32 	%f159, [polyPoints+188];
	sub.f32 	%f160, %f33, %f147;
	mul.f32 	%f161, %f159, %f160;
	sub.f32 	%f162, %f161, %f158;
	mov.f32 	%f32, %f162;
$Lt_0_55298:
	.loc	15	51	0
	mov.s32 	%r30, 0;
$Lt_0_56322:
 //<loop> Loop body line 51, nesting depth: 1, iterations: 4
	cvt.rn.f32.s32 	%f163, %r30;
	mov.f32 	%f164, %f2;
	setp.lt.f32 	%p52, %f163, %f164;
	@!%p52 bra 	$Lt_0_56834;
 //<loop> Part of loop body line 51, head labeled $Lt_0_56322
	.loc	15	61	0
	mov.f32 	%f165, 0f47c35000;   	// 100000
	min.f32 	%f166, %f164, %f165;
	mov.s32 	%r31, 1;
	bra.uni 	$Lt_0_56578;
$Lt_0_56834:
 //<loop> Part of loop body line 51, head labeled $Lt_0_56322
	mov.s32 	%r31, 0;
	mov.f32 	%f166, 0f47c35000;   	// 100000
$Lt_0_56578:
 //<loop> Part of loop body line 51, head labeled $Lt_0_56322
	mov.f32 	%f167, %f4;
	setp.lt.f32 	%p53, %f163, %f167;
	@!%p53 bra 	$Lt_0_57090;
 //<loop> Part of loop body line 51, head labeled $Lt_0_56322
	.loc	15	60	0
	xor.b32 	%r31, %r31, 1;
	.loc	15	61	0
	min.f32 	%f166, %f166, %f167;
$Lt_0_57090:
 //<loop> Part of loop body line 51, head labeled $Lt_0_56322
	mov.f32 	%f168, %f6;
	setp.lt.f32 	%p54, %f163, %f168;
	@!%p54 bra 	$Lt_0_57602;
 //<loop> Part of loop body line 51, head labeled $Lt_0_56322
	.loc	15	60	0
	xor.b32 	%r31, %r31, 1;
	.loc	15	61	0
	min.f32 	%f166, %f166, %f168;
$Lt_0_57602:
 //<loop> Part of loop body line 51, head labeled $Lt_0_56322
	mov.f32 	%f169, %f8;
	setp.lt.f32 	%p55, %f163, %f169;
	@!%p55 bra 	$Lt_0_58114;
 //<loop> Part of loop body line 51, head labeled $Lt_0_56322
	.loc	15	60	0
	xor.b32 	%r31, %r31, 1;
	.loc	15	61	0
	min.f32 	%f166, %f166, %f169;
$Lt_0_58114:
 //<loop> Part of loop body line 51, head labeled $Lt_0_56322
	mov.f32 	%f170, %f10;
	setp.lt.f32 	%p56, %f163, %f170;
	@!%p56 bra 	$Lt_0_58626;
 //<loop> Part of loop body line 51, head labeled $Lt_0_56322
	.loc	15	60	0
	xor.b32 	%r31, %r31, 1;
	.loc	15	61	0
	min.f32 	%f166, %f166, %f170;
$Lt_0_58626:
 //<loop> Part of loop body line 51, head labeled $Lt_0_56322
	mov.f32 	%f171, %f12;
	setp.lt.f32 	%p57, %f163, %f171;
	@!%p57 bra 	$Lt_0_59138;
 //<loop> Part of loop body line 51, head labeled $Lt_0_56322
	.loc	15	60	0
	xor.b32 	%r31, %r31, 1;
	.loc	15	61	0
	min.f32 	%f166, %f166, %f171;
$Lt_0_59138:
 //<loop> Part of loop body line 51, head labeled $Lt_0_56322
	mov.f32 	%f172, %f14;
	setp.lt.f32 	%p58, %f163, %f172;
	@!%p58 bra 	$Lt_0_59650;
 //<loop> Part of loop body line 51, head labeled $Lt_0_56322
	.loc	15	60	0
	xor.b32 	%r31, %r31, 1;
	.loc	15	61	0
	min.f32 	%f166, %f166, %f172;
$Lt_0_59650:
 //<loop> Part of loop body line 51, head labeled $Lt_0_56322
	mov.f32 	%f173, %f16;
	setp.lt.f32 	%p59, %f163, %f173;
	@!%p59 bra 	$Lt_0_60162;
 //<loop> Part of loop body line 51, head labeled $Lt_0_56322
	.loc	15	60	0
	xor.b32 	%r31, %r31, 1;
	.loc	15	61	0
	min.f32 	%f166, %f166, %f173;
$Lt_0_60162:
 //<loop> Part of loop body line 51, head labeled $Lt_0_56322
	mov.f32 	%f174, %f18;
	setp.lt.f32 	%p60, %f163, %f174;
	@!%p60 bra 	$Lt_0_60674;
 //<loop> Part of loop body line 51, head labeled $Lt_0_56322
	.loc	15	60	0
	xor.b32 	%r31, %r31, 1;
	.loc	15	61	0
	min.f32 	%f166, %f166, %f174;
$Lt_0_60674:
 //<loop> Part of loop body line 51, head labeled $Lt_0_56322
	mov.f32 	%f175, %f20;
	setp.lt.f32 	%p61, %f163, %f175;
	@!%p61 bra 	$Lt_0_61186;
 //<loop> Part of loop body line 51, head labeled $Lt_0_56322
	.loc	15	60	0
	xor.b32 	%r31, %r31, 1;
	.loc	15	61	0
	min.f32 	%f166, %f166, %f175;
$Lt_0_61186:
 //<loop> Part of loop body line 51, head labeled $Lt_0_56322
	mov.f32 	%f176, %f22;
	setp.lt.f32 	%p62, %f163, %f176;
	@!%p62 bra 	$Lt_0_61698;
 //<loop> Part of loop body line 51, head labeled $Lt_0_56322
	.loc	15	60	0
	xor.b32 	%r31, %r31, 1;
	.loc	15	61	0
	min.f32 	%f166, %f166, %f176;
$Lt_0_61698:
 //<loop> Part of loop body line 51, head labeled $Lt_0_56322
	mov.f32 	%f177, %f24;
	setp.lt.f32 	%p63, %f163, %f177;
	@!%p63 bra 	$Lt_0_62210;
 //<loop> Part of loop body line 51, head labeled $Lt_0_56322
	.loc	15	60	0
	xor.b32 	%r31, %r31, 1;
	.loc	15	61	0
	min.f32 	%f166, %f166, %f177;
$Lt_0_62210:
 //<loop> Part of loop body line 51, head labeled $Lt_0_56322
	mov.f32 	%f178, %f26;
	setp.lt.f32 	%p64, %f163, %f178;
	@!%p64 bra 	$Lt_0_62722;
 //<loop> Part of loop body line 51, head labeled $Lt_0_56322
	.loc	15	60	0
	xor.b32 	%r31, %r31, 1;
	.loc	15	61	0
	min.f32 	%f166, %f166, %f178;
$Lt_0_62722:
 //<loop> Part of loop body line 51, head labeled $Lt_0_56322
	mov.f32 	%f179, %f28;
	setp.lt.f32 	%p65, %f163, %f179;
	@!%p65 bra 	$Lt_0_63234;
 //<loop> Part of loop body line 51, head labeled $Lt_0_56322
	.loc	15	60	0
	xor.b32 	%r31, %r31, 1;
	.loc	15	61	0
	min.f32 	%f166, %f166, %f179;
$Lt_0_63234:
 //<loop> Part of loop body line 51, head labeled $Lt_0_56322
	mov.f32 	%f180, %f30;
	setp.lt.f32 	%p66, %f163, %f180;
	@!%p66 bra 	$Lt_0_63746;
 //<loop> Part of loop body line 51, head labeled $Lt_0_56322
	.loc	15	60	0
	xor.b32 	%r31, %r31, 1;
	.loc	15	61	0
	min.f32 	%f166, %f166, %f180;
$Lt_0_63746:
 //<loop> Part of loop body line 51, head labeled $Lt_0_56322
	mov.f32 	%f181, %f32;
	setp.lt.f32 	%p67, %f163, %f181;
	@!%p67 bra 	$Lt_0_64258;
 //<loop> Part of loop body line 51, head labeled $Lt_0_56322
	.loc	15	60	0
	xor.b32 	%r31, %r31, 1;
	.loc	15	61	0
	min.f32 	%f166, %f166, %f181;
$Lt_0_64258:
 //<loop> Part of loop body line 51, head labeled $Lt_0_56322
	mov.s32 	%r32, 0;
	setp.eq.s32 	%p68, %r31, %r32;
	add.s32 	%r33, %r30, 8;
	cvt.rn.f32.s32 	%f182, %r33;
	setp.le.f32 	%p69, %f182, %f166;
	@!%p69 bra 	$Lt_0_65026;
 //<loop> Part of loop body line 51, head labeled $Lt_0_56322
	@!%p68 bra 	$Lt_0_64770;
 //<loop> Part of loop body line 51, head labeled $Lt_0_56322
	.loc	15	67	0
	mov.f32 	%f183, 0f00000000;   	// 0
	st.global.f32 	[%rd4+12], %f183;
	mov.f32 	%f184, 0f00000000;   	// 0
	st.global.f32 	[%rd4+28], %f184;
	mov.f32 	%f185, 0f00000000;   	// 0
	st.global.f32 	[%rd4+44], %f185;
	mov.f32 	%f186, 0f00000000;   	// 0
	st.global.f32 	[%rd4+60], %f186;
	mov.f32 	%f187, 0f00000000;   	// 0
	st.global.f32 	[%rd4+76], %f187;
	mov.f32 	%f188, 0f00000000;   	// 0
	st.global.f32 	[%rd4+92], %f188;
	mov.f32 	%f189, 0f00000000;   	// 0
	st.global.f32 	[%rd4+108], %f189;
	mov.f32 	%f190, 0f00000000;   	// 0
	st.global.f32 	[%rd4+124], %f190;
	bra.uni 	$Lt_0_64770;
$Lt_0_65026:
 //<loop> Part of loop body line 51, head labeled $Lt_0_56322
	@!%p68 bra 	$Lt_0_65794;
 //<loop> Part of loop body line 51, head labeled $Lt_0_56322
	.loc	15	71	0
	mov.f32 	%f191, 0f00000000;   	// 0
	st.global.f32 	[%rd4+12], %f191;
$Lt_0_65794:
 //<loop> Part of loop body line 51, head labeled $Lt_0_56322
	mov.s32 	%r34, 1;
$Lt_0_66818:
 //<loop> Loop body line 71, nesting depth: 2, iterations: 7
	.loc	15	74	0
	add.s32 	%r35, %r34, %r30;
	cvt.rn.f32.s32 	%f192, %r35;
	mov.f32 	%f193, %f10;
	setp.gt.f32 	%p70, %f193, %f192;
	mov.f32 	%f194, %f8;
	setp.gt.f32 	%p71, %f194, %f192;
	mov.f32 	%f195, %f6;
	setp.gt.f32 	%p72, %f195, %f192;
	mov.f32 	%f196, %f4;
	setp.gt.f32 	%p73, %f196, %f192;
	mov.f32 	%f197, %f2;
	setp.gt.f32 	%p74, %f197, %f192;
	selp.s32 	%r36, 1, 0, %p74;
	xor.b32 	%r37, %r36, 1;
	selp.s32 	%r38, %r37, %r36, %p73;
	xor.b32 	%r39, %r38, 1;
	selp.s32 	%r40, %r39, %r38, %p72;
	xor.b32 	%r41, %r40, 1;
	selp.s32 	%r42, %r41, %r40, %p71;
	xor.b32 	%r43, %r42, 1;
	selp.s32 	%r44, %r43, %r42, %p70;
	xor.b32 	%r45, %r44, 1;
	mov.f32 	%f198, %f12;
	setp.gt.f32 	%p75, %f198, %f192;
	selp.s32 	%r31, %r45, %r44, %p75;
	xor.b32 	%r46, %r31, 1;
	mov.f32 	%f199, %f14;
	setp.gt.f32 	%p76, %f199, %f192;
	selp.s32 	%r31, %r46, %r31, %p76;
	xor.b32 	%r47, %r31, 1;
	mov.f32 	%f200, %f16;
	setp.gt.f32 	%p77, %f200, %f192;
	selp.s32 	%r31, %r47, %r31, %p77;
	xor.b32 	%r48, %r31, 1;
	mov.f32 	%f201, %f18;
	setp.gt.f32 	%p78, %f201, %f192;
	selp.s32 	%r31, %r48, %r31, %p78;
	xor.b32 	%r49, %r31, 1;
	mov.f32 	%f202, %f20;
	setp.gt.f32 	%p79, %f202, %f192;
	selp.s32 	%r31, %r49, %r31, %p79;
	xor.b32 	%r50, %r31, 1;
	mov.f32 	%f203, %f22;
	setp.gt.f32 	%p80, %f203, %f192;
	selp.s32 	%r31, %r50, %r31, %p80;
	xor.b32 	%r51, %r31, 1;
	mov.f32 	%f204, %f24;
	setp.gt.f32 	%p81, %f204, %f192;
	selp.s32 	%r31, %r51, %r31, %p81;
	xor.b32 	%r52, %r31, 1;
	mov.f32 	%f205, %f26;
	setp.gt.f32 	%p82, %f205, %f192;
	selp.s32 	%r31, %r52, %r31, %p82;
	xor.b32 	%r53, %r31, 1;
	mov.f32 	%f206, %f28;
	setp.gt.f32 	%p83, %f206, %f192;
	selp.s32 	%r31, %r53, %r31, %p83;
	xor.b32 	%r54, %r31, 1;
	mov.f32 	%f207, %f30;
	setp.gt.f32 	%p84, %f207, %f192;
	selp.s32 	%r31, %r54, %r31, %p84;
	xor.b32 	%r55, %r31, 1;
	mov.f32 	%f208, %f32;
	setp.gt.f32 	%p85, %f208, %f192;
	selp.s32 	%r56, %r55, %r31, %p85;
	mov.u32 	%r57, 0;
	setp.ne.s32 	%p86, %r56, %r57;
	@%p86 bra 	$Lt_0_67074;
 //<loop> Part of loop body line 71, head labeled $Lt_0_66818
	.loc	15	75	0
	mov.f32 	%f209, 0f00000000;   	// 0
	cvt.u64.s32 	%rd5, %r34;
	mul.lo.u64 	%rd6, %rd5, 16;
	add.u64 	%rd7, %rd4, %rd6;
	st.global.f32 	[%rd7+12], %f209;
$Lt_0_67074:
 //<loop> Part of loop body line 71, head labeled $Lt_0_66818
	add.s32 	%r34, %r34, 1;
	mov.u32 	%r58, 8;
	setp.ne.s32 	%p87, %r34, %r58;
	@%p87 bra 	$Lt_0_66818;
$Lt_0_64770:
 //<loop> Part of loop body line 51, head labeled $Lt_0_56322
	.loc	15	79	0
	add.u64 	%rd4, %rd4, 128;
	mov.s32 	%r30, %r33;
	mov.u32 	%r59, 31;
	setp.le.s32 	%p88, %r33, %r59;
	@%p88 bra 	$Lt_0_56322;
	bra.uni 	$LBB160_cuda_kernel_bitmask;
$Lt_0_47362:
	mov.u32 	%r60, 8;
	setp.ne.s32 	%p89, %r28, %r60;
	@%p89 bra 	$Lt_0_68354;
	.loc	15	98	0
	mov.s64 	%rd8, %rd3;
	.loc	15	50	0
	mov.f32 	%f210, 0fbf800000;   	// -1
	mov.f32 	%f211, %f210;
	mov.f32 	%f212, 0fbf800000;   	// -1
	mov.f32 	%f213, %f212;
	mov.f32 	%f214, 0fbf800000;   	// -1
	mov.f32 	%f215, %f214;
	mov.f32 	%f216, 0fbf800000;   	// -1
	mov.f32 	%f217, %f216;
	mov.f32 	%f218, 0fbf800000;   	// -1
	mov.f32 	%f219, %f218;
	mov.f32 	%f220, 0fbf800000;   	// -1
	mov.f32 	%f221, %f220;
	mov.f32 	%f222, 0fbf800000;   	// -1
	mov.f32 	%f223, %f222;
	mov.f32 	%f224, 0fbf800000;   	// -1
	mov.f32 	%f225, %f224;
	.loc	15	24	0
	cvt.rn.f32.s32 	%f226, %r10;
	ld.const.f32 	%f227, [polyPoints+4];
	ld.const.f32 	%f228, [polyPoints+16];
	setp.ge.f32 	%p90, %f226, %f227;
	setp.lt.f32 	%p91, %f226, %f228;
	xor.pred 	%p92, %p90, %p91;
	@%p92 bra 	$Lt_0_68610;
	.loc	15	40	0
	cvt.rn.f32.s32 	%f229, %r22;
	ld.const.f32 	%f230, [polyPoints+0];
	sub.f32 	%f231, %f229, %f230;
	ld.const.f32 	%f232, [polyPoints+8];
	sub.f32 	%f233, %f226, %f227;
	mul.f32 	%f234, %f232, %f233;
	sub.f32 	%f235, %f234, %f231;
	mov.f32 	%f211, %f235;
$Lt_0_68610:
	ld.const.f32 	%f236, [polyPoints+28];
	setp.ge.f32 	%p93, %f226, %f228;
	setp.lt.f32 	%p94, %f226, %f236;
	xor.pred 	%p95, %p93, %p94;
	@%p95 bra 	$Lt_0_69122;
	cvt.rn.f32.s32 	%f237, %r22;
	ld.const.f32 	%f238, [polyPoints+12];
	sub.f32 	%f239, %f237, %f238;
	ld.const.f32 	%f240, [polyPoints+20];
	sub.f32 	%f241, %f226, %f228;
	mul.f32 	%f242, %f240, %f241;
	sub.f32 	%f243, %f242, %f239;
	mov.f32 	%f213, %f243;
$Lt_0_69122:
	ld.const.f32 	%f244, [polyPoints+40];
	setp.ge.f32 	%p96, %f226, %f236;
	setp.lt.f32 	%p97, %f226, %f244;
	xor.pred 	%p98, %p96, %p97;
	@%p98 bra 	$Lt_0_69634;
	cvt.rn.f32.s32 	%f245, %r22;
	ld.const.f32 	%f246, [polyPoints+24];
	sub.f32 	%f247, %f245, %f246;
	ld.const.f32 	%f248, [polyPoints+32];
	sub.f32 	%f249, %f226, %f236;
	mul.f32 	%f250, %f248, %f249;
	sub.f32 	%f251, %f250, %f247;
	mov.f32 	%f215, %f251;
$Lt_0_69634:
	ld.const.f32 	%f252, [polyPoints+52];
	setp.ge.f32 	%p99, %f226, %f244;
	setp.lt.f32 	%p100, %f226, %f252;
	xor.pred 	%p101, %p99, %p100;
	@%p101 bra 	$Lt_0_70146;
	cvt.rn.f32.s32 	%f253, %r22;
	ld.const.f32 	%f254, [polyPoints+36];
	sub.f32 	%f255, %f253, %f254;
	ld.const.f32 	%f256, [polyPoints+44];
	sub.f32 	%f257, %f226, %f244;
	mul.f32 	%f258, %f256, %f257;
	sub.f32 	%f259, %f258, %f255;
	mov.f32 	%f217, %f259;
$Lt_0_70146:
	ld.const.f32 	%f260, [polyPoints+64];
	setp.ge.f32 	%p102, %f226, %f252;
	setp.lt.f32 	%p103, %f226, %f260;
	xor.pred 	%p104, %p102, %p103;
	@%p104 bra 	$Lt_0_70658;
	cvt.rn.f32.s32 	%f261, %r22;
	ld.const.f32 	%f262, [polyPoints+48];
	sub.f32 	%f263, %f261, %f262;
	ld.const.f32 	%f264, [polyPoints+56];
	sub.f32 	%f265, %f226, %f252;
	mul.f32 	%f266, %f264, %f265;
	sub.f32 	%f267, %f266, %f263;
	mov.f32 	%f219, %f267;
$Lt_0_70658:
	ld.const.f32 	%f268, [polyPoints+76];
	setp.ge.f32 	%p105, %f226, %f260;
	setp.lt.f32 	%p106, %f226, %f268;
	xor.pred 	%p107, %p105, %p106;
	@%p107 bra 	$Lt_0_71170;
	cvt.rn.f32.s32 	%f269, %r22;
	ld.const.f32 	%f270, [polyPoints+60];
	sub.f32 	%f271, %f269, %f270;
	ld.const.f32 	%f272, [polyPoints+68];
	sub.f32 	%f273, %f226, %f260;
	mul.f32 	%f274, %f272, %f273;
	sub.f32 	%f275, %f274, %f271;
	mov.f32 	%f221, %f275;
$Lt_0_71170:
	ld.const.f32 	%f276, [polyPoints+88];
	setp.ge.f32 	%p108, %f226, %f268;
	setp.lt.f32 	%p109, %f226, %f276;
	xor.pred 	%p110, %p108, %p109;
	@%p110 bra 	$Lt_0_71682;
	cvt.rn.f32.s32 	%f277, %r22;
	ld.const.f32 	%f278, [polyPoints+72];
	sub.f32 	%f279, %f277, %f278;
	ld.const.f32 	%f280, [polyPoints+80];
	sub.f32 	%f281, %f226, %f268;
	mul.f32 	%f282, %f280, %f281;
	sub.f32 	%f283, %f282, %f279;
	mov.f32 	%f223, %f283;
$Lt_0_71682:
	setp.ge.f32 	%p111, %f226, %f276;
	ld.const.f32 	%f284, [polyPoints+100];
	setp.gt.f32 	%p112, %f284, %f226;
	xor.pred 	%p113, %p111, %p112;
	@%p113 bra 	$Lt_0_72194;
	cvt.rn.f32.s32 	%f285, %r22;
	ld.const.f32 	%f286, [polyPoints+84];
	sub.f32 	%f287, %f285, %f286;
	ld.const.f32 	%f288, [polyPoints+92];
	sub.f32 	%f289, %f226, %f276;
	mul.f32 	%f290, %f288, %f289;
	sub.f32 	%f291, %f290, %f287;
	mov.f32 	%f225, %f291;
$Lt_0_72194:
	.loc	15	51	0
	mov.s32 	%r61, 0;
$Lt_0_73218:
 //<loop> Loop body line 51, nesting depth: 1, iterations: 4
	cvt.rn.f32.s32 	%f292, %r61;
	mov.f32 	%f293, %f211;
	setp.lt.f32 	%p114, %f292, %f293;
	@!%p114 bra 	$Lt_0_73730;
 //<loop> Part of loop body line 51, head labeled $Lt_0_73218
	.loc	15	61	0
	mov.f32 	%f294, 0f47c35000;   	// 100000
	min.f32 	%f295, %f293, %f294;
	mov.s32 	%r62, 1;
	bra.uni 	$Lt_0_73474;
$Lt_0_73730:
 //<loop> Part of loop body line 51, head labeled $Lt_0_73218
	mov.s32 	%r62, 0;
	mov.f32 	%f295, 0f47c35000;   	// 100000
$Lt_0_73474:
 //<loop> Part of loop body line 51, head labeled $Lt_0_73218
	mov.f32 	%f296, %f213;
	setp.lt.f32 	%p115, %f292, %f296;
	@!%p115 bra 	$Lt_0_73986;
 //<loop> Part of loop body line 51, head labeled $Lt_0_73218
	.loc	15	60	0
	xor.b32 	%r62, %r62, 1;
	.loc	15	61	0
	min.f32 	%f295, %f295, %f296;
$Lt_0_73986:
 //<loop> Part of loop body line 51, head labeled $Lt_0_73218
	mov.f32 	%f297, %f215;
	setp.lt.f32 	%p116, %f292, %f297;
	@!%p116 bra 	$Lt_0_74498;
 //<loop> Part of loop body line 51, head labeled $Lt_0_73218
	.loc	15	60	0
	xor.b32 	%r62, %r62, 1;
	.loc	15	61	0
	min.f32 	%f295, %f295, %f297;
$Lt_0_74498:
 //<loop> Part of loop body line 51, head labeled $Lt_0_73218
	mov.f32 	%f298, %f217;
	setp.lt.f32 	%p117, %f292, %f298;
	@!%p117 bra 	$Lt_0_75010;
 //<loop> Part of loop body line 51, head labeled $Lt_0_73218
	.loc	15	60	0
	xor.b32 	%r62, %r62, 1;
	.loc	15	61	0
	min.f32 	%f295, %f295, %f298;
$Lt_0_75010:
 //<loop> Part of loop body line 51, head labeled $Lt_0_73218
	mov.f32 	%f299, %f219;
	setp.lt.f32 	%p118, %f292, %f299;
	@!%p118 bra 	$Lt_0_75522;
 //<loop> Part of loop body line 51, head labeled $Lt_0_73218
	.loc	15	60	0
	xor.b32 	%r62, %r62, 1;
	.loc	15	61	0
	min.f32 	%f295, %f295, %f299;
$Lt_0_75522:
 //<loop> Part of loop body line 51, head labeled $Lt_0_73218
	mov.f32 	%f300, %f221;
	setp.lt.f32 	%p119, %f292, %f300;
	@!%p119 bra 	$Lt_0_76034;
 //<loop> Part of loop body line 51, head labeled $Lt_0_73218
	.loc	15	60	0
	xor.b32 	%r62, %r62, 1;
	.loc	15	61	0
	min.f32 	%f295, %f295, %f300;
$Lt_0_76034:
 //<loop> Part of loop body line 51, head labeled $Lt_0_73218
	mov.f32 	%f301, %f223;
	setp.lt.f32 	%p120, %f292, %f301;
	@!%p120 bra 	$Lt_0_76546;
 //<loop> Part of loop body line 51, head labeled $Lt_0_73218
	.loc	15	60	0
	xor.b32 	%r62, %r62, 1;
	.loc	15	61	0
	min.f32 	%f295, %f295, %f301;
$Lt_0_76546:
 //<loop> Part of loop body line 51, head labeled $Lt_0_73218
	mov.f32 	%f302, %f225;
	setp.lt.f32 	%p121, %f292, %f302;
	@!%p121 bra 	$Lt_0_77058;
 //<loop> Part of loop body line 51, head labeled $Lt_0_73218
	.loc	15	60	0
	xor.b32 	%r62, %r62, 1;
	.loc	15	61	0
	min.f32 	%f295, %f295, %f302;
$Lt_0_77058:
 //<loop> Part of loop body line 51, head labeled $Lt_0_73218
	mov.s32 	%r63, 0;
	setp.eq.s32 	%p122, %r62, %r63;
	add.s32 	%r64, %r61, 8;
	cvt.rn.f32.s32 	%f303, %r64;
	setp.le.f32 	%p123, %f303, %f295;
	@!%p123 bra 	$Lt_0_77826;
 //<loop> Part of loop body line 51, head labeled $Lt_0_73218
	@!%p122 bra 	$Lt_0_77570;
 //<loop> Part of loop body line 51, head labeled $Lt_0_73218
	.loc	15	67	0
	mov.f32 	%f304, 0f00000000;   	// 0
	st.global.f32 	[%rd8+12], %f304;
	mov.f32 	%f305, 0f00000000;   	// 0
	st.global.f32 	[%rd8+28], %f305;
	mov.f32 	%f306, 0f00000000;   	// 0
	st.global.f32 	[%rd8+44], %f306;
	mov.f32 	%f307, 0f00000000;   	// 0
	st.global.f32 	[%rd8+60], %f307;
	mov.f32 	%f308, 0f00000000;   	// 0
	st.global.f32 	[%rd8+76], %f308;
	mov.f32 	%f309, 0f00000000;   	// 0
	st.global.f32 	[%rd8+92], %f309;
	mov.f32 	%f310, 0f00000000;   	// 0
	st.global.f32 	[%rd8+108], %f310;
	mov.f32 	%f311, 0f00000000;   	// 0
	st.global.f32 	[%rd8+124], %f311;
	bra.uni 	$Lt_0_77570;
$Lt_0_77826:
 //<loop> Part of loop body line 51, head labeled $Lt_0_73218
	@!%p122 bra 	$Lt_0_78594;
 //<loop> Part of loop body line 51, head labeled $Lt_0_73218
	.loc	15	71	0
	mov.f32 	%f312, 0f00000000;   	// 0
	st.global.f32 	[%rd8+12], %f312;
$Lt_0_78594:
 //<loop> Part of loop body line 51, head labeled $Lt_0_73218
	mov.s32 	%r65, 1;
$Lt_0_79618:
 //<loop> Loop body line 71, nesting depth: 2, iterations: 7
	.loc	15	74	0
	add.s32 	%r66, %r65, %r61;
	cvt.rn.f32.s32 	%f313, %r66;
	mov.f32 	%f314, %f219;
	setp.gt.f32 	%p124, %f314, %f313;
	mov.f32 	%f315, %f217;
	setp.gt.f32 	%p125, %f315, %f313;
	mov.f32 	%f316, %f215;
	setp.gt.f32 	%p126, %f316, %f313;
	mov.f32 	%f317, %f213;
	setp.gt.f32 	%p127, %f317, %f313;
	mov.f32 	%f318, %f211;
	setp.gt.f32 	%p128, %f318, %f313;
	selp.s32 	%r67, 1, 0, %p128;
	xor.b32 	%r68, %r67, 1;
	selp.s32 	%r69, %r68, %r67, %p127;
	xor.b32 	%r70, %r69, 1;
	selp.s32 	%r71, %r70, %r69, %p126;
	xor.b32 	%r72, %r71, 1;
	selp.s32 	%r73, %r72, %r71, %p125;
	xor.b32 	%r74, %r73, 1;
	selp.s32 	%r75, %r74, %r73, %p124;
	xor.b32 	%r76, %r75, 1;
	mov.f32 	%f319, %f221;
	setp.gt.f32 	%p129, %f319, %f313;
	selp.s32 	%r62, %r76, %r75, %p129;
	xor.b32 	%r77, %r62, 1;
	mov.f32 	%f320, %f223;
	setp.gt.f32 	%p130, %f320, %f313;
	selp.s32 	%r62, %r77, %r62, %p130;
	xor.b32 	%r78, %r62, 1;
	mov.f32 	%f321, %f225;
	setp.gt.f32 	%p131, %f321, %f313;
	selp.s32 	%r79, %r78, %r62, %p131;
	mov.u32 	%r80, 0;
	setp.ne.s32 	%p132, %r79, %r80;
	@%p132 bra 	$Lt_0_79874;
 //<loop> Part of loop body line 71, head labeled $Lt_0_79618
	.loc	15	75	0
	mov.f32 	%f322, 0f00000000;   	// 0
	cvt.u64.s32 	%rd9, %r65;
	mul.lo.u64 	%rd10, %rd9, 16;
	add.u64 	%rd11, %rd8, %rd10;
	st.global.f32 	[%rd11+12], %f322;
$Lt_0_79874:
 //<loop> Part of loop body line 71, head labeled $Lt_0_79618
	add.s32 	%r65, %r65, 1;
	mov.u32 	%r81, 8;
	setp.ne.s32 	%p133, %r65, %r81;
	@%p133 bra 	$Lt_0_79618;
$Lt_0_77570:
 //<loop> Part of loop body line 51, head labeled $Lt_0_73218
	.loc	15	79	0
	add.u64 	%rd8, %rd8, 128;
	mov.s32 	%r61, %r64;
	mov.u32 	%r82, 31;
	setp.le.s32 	%p134, %r64, %r82;
	@%p134 bra 	$Lt_0_73218;
	bra.uni 	$LBB160_cuda_kernel_bitmask;
$Lt_0_68354:
	mov.u32 	%r83, 4;
	setp.ne.s32 	%p135, %r28, %r83;
	@%p135 bra 	$LBB160_cuda_kernel_bitmask;
	.loc	15	99	0
	mov.s64 	%rd12, %rd3;
	.loc	15	50	0
	mov.f32 	%f323, 0fbf800000;   	// -1
	mov.f32 	%f324, %f323;
	mov.f32 	%f325, 0fbf800000;   	// -1
	mov.f32 	%f326, %f325;
	mov.f32 	%f327, 0fbf800000;   	// -1
	mov.f32 	%f328, %f327;
	mov.f32 	%f329, 0fbf800000;   	// -1
	mov.f32 	%f330, %f329;
	.loc	15	24	0
	cvt.rn.f32.s32 	%f331, %r10;
	ld.const.f32 	%f332, [polyPoints+4];
	ld.const.f32 	%f333, [polyPoints+16];
	setp.ge.f32 	%p136, %f331, %f332;
	setp.lt.f32 	%p137, %f331, %f333;
	xor.pred 	%p138, %p136, %p137;
	@%p138 bra 	$Lt_0_81410;
	.loc	15	40	0
	cvt.rn.f32.s32 	%f334, %r22;
	ld.const.f32 	%f335, [polyPoints+0];
	sub.f32 	%f336, %f334, %f335;
	ld.const.f32 	%f337, [polyPoints+8];
	sub.f32 	%f338, %f331, %f332;
	mul.f32 	%f339, %f337, %f338;
	sub.f32 	%f340, %f339, %f336;
	mov.f32 	%f324, %f340;
$Lt_0_81410:
	ld.const.f32 	%f341, [polyPoints+28];
	setp.ge.f32 	%p139, %f331, %f333;
	setp.lt.f32 	%p140, %f331, %f341;
	xor.pred 	%p141, %p139, %p140;
	@%p141 bra 	$Lt_0_81922;
	cvt.rn.f32.s32 	%f342, %r22;
	ld.const.f32 	%f343, [polyPoints+12];
	sub.f32 	%f344, %f342, %f343;
	ld.const.f32 	%f345, [polyPoints+20];
	sub.f32 	%f346, %f331, %f333;
	mul.f32 	%f347, %f345, %f346;
	sub.f32 	%f348, %f347, %f344;
	mov.f32 	%f326, %f348;
$Lt_0_81922:
	ld.const.f32 	%f349, [polyPoints+40];
	setp.ge.f32 	%p142, %f331, %f341;
	setp.lt.f32 	%p143, %f331, %f349;
	xor.pred 	%p144, %p142, %p143;
	@%p144 bra 	$Lt_0_82434;
	cvt.rn.f32.s32 	%f350, %r22;
	ld.const.f32 	%f351, [polyPoints+24];
	sub.f32 	%f352, %f350, %f351;
	ld.const.f32 	%f353, [polyPoints+32];
	sub.f32 	%f354, %f331, %f341;
	mul.f32 	%f355, %f353, %f354;
	sub.f32 	%f356, %f355, %f352;
	mov.f32 	%f328, %f356;
$Lt_0_82434:
	setp.ge.f32 	%p145, %f331, %f349;
	ld.const.f32 	%f357, [polyPoints+52];
	setp.gt.f32 	%p146, %f357, %f331;
	xor.pred 	%p147, %p145, %p146;
	@%p147 bra 	$Lt_0_82946;
	cvt.rn.f32.s32 	%f358, %r22;
	ld.const.f32 	%f359, [polyPoints+36];
	sub.f32 	%f360, %f358, %f359;
	ld.const.f32 	%f361, [polyPoints+44];
	sub.f32 	%f362, %f331, %f349;
	mul.f32 	%f363, %f361, %f362;
	sub.f32 	%f364, %f363, %f360;
	mov.f32 	%f330, %f364;
$Lt_0_82946:
	.loc	15	51	0
	mov.s32 	%r84, 0;
$Lt_0_83970:
 //<loop> Loop body line 51, nesting depth: 1, iterations: 4
	cvt.rn.f32.s32 	%f365, %r84;
	mov.f32 	%f366, %f324;
	setp.lt.f32 	%p148, %f365, %f366;
	@!%p148 bra 	$Lt_0_84482;
 //<loop> Part of loop body line 51, head labeled $Lt_0_83970
	.loc	15	61	0
	mov.f32 	%f367, 0f47c35000;   	// 100000
	min.f32 	%f368, %f366, %f367;
	mov.s32 	%r85, 1;
	bra.uni 	$Lt_0_84226;
$Lt_0_84482:
 //<loop> Part of loop body line 51, head labeled $Lt_0_83970
	mov.s32 	%r85, 0;
	mov.f32 	%f368, 0f47c35000;   	// 100000
$Lt_0_84226:
 //<loop> Part of loop body line 51, head labeled $Lt_0_83970
	mov.f32 	%f369, %f326;
	setp.lt.f32 	%p149, %f365, %f369;
	@!%p149 bra 	$Lt_0_84738;
 //<loop> Part of loop body line 51, head labeled $Lt_0_83970
	.loc	15	60	0
	xor.b32 	%r85, %r85, 1;
	.loc	15	61	0
	min.f32 	%f368, %f368, %f369;
$Lt_0_84738:
 //<loop> Part of loop body line 51, head labeled $Lt_0_83970
	mov.f32 	%f370, %f328;
	setp.lt.f32 	%p150, %f365, %f370;
	@!%p150 bra 	$Lt_0_85250;
 //<loop> Part of loop body line 51, head labeled $Lt_0_83970
	.loc	15	60	0
	xor.b32 	%r85, %r85, 1;
	.loc	15	61	0
	min.f32 	%f368, %f368, %f370;
$Lt_0_85250:
 //<loop> Part of loop body line 51, head labeled $Lt_0_83970
	mov.f32 	%f371, %f330;
	setp.lt.f32 	%p151, %f365, %f371;
	@!%p151 bra 	$Lt_0_85762;
 //<loop> Part of loop body line 51, head labeled $Lt_0_83970
	.loc	15	60	0
	xor.b32 	%r85, %r85, 1;
	.loc	15	61	0
	min.f32 	%f368, %f368, %f371;
$Lt_0_85762:
 //<loop> Part of loop body line 51, head labeled $Lt_0_83970
	mov.s32 	%r86, 0;
	setp.eq.s32 	%p152, %r85, %r86;
	add.s32 	%r87, %r84, 8;
	cvt.rn.f32.s32 	%f372, %r87;
	setp.le.f32 	%p153, %f372, %f368;
	@!%p153 bra 	$Lt_0_86530;
 //<loop> Part of loop body line 51, head labeled $Lt_0_83970
	@!%p152 bra 	$Lt_0_86274;
 //<loop> Part of loop body line 51, head labeled $Lt_0_83970
	.loc	15	67	0
	mov.f32 	%f373, 0f00000000;   	// 0
	st.global.f32 	[%rd12+12], %f373;
	mov.f32 	%f374, 0f00000000;   	// 0
	st.global.f32 	[%rd12+28], %f374;
	mov.f32 	%f375, 0f00000000;   	// 0
	st.global.f32 	[%rd12+44], %f375;
	mov.f32 	%f376, 0f00000000;   	// 0
	st.global.f32 	[%rd12+60], %f376;
	mov.f32 	%f377, 0f00000000;   	// 0
	st.global.f32 	[%rd12+76], %f377;
	mov.f32 	%f378, 0f00000000;   	// 0
	st.global.f32 	[%rd12+92], %f378;
	mov.f32 	%f379, 0f00000000;   	// 0
	st.global.f32 	[%rd12+108], %f379;
	mov.f32 	%f380, 0f00000000;   	// 0
	st.global.f32 	[%rd12+124], %f380;
	bra.uni 	$Lt_0_86274;
$Lt_0_86530:
 //<loop> Part of loop body line 51, head labeled $Lt_0_83970
	@!%p152 bra 	$Lt_0_87298;
 //<loop> Part of loop body line 51, head labeled $Lt_0_83970
	.loc	15	71	0
	mov.f32 	%f381, 0f00000000;   	// 0
	st.global.f32 	[%rd12+12], %f381;
$Lt_0_87298:
 //<loop> Part of loop body line 51, head labeled $Lt_0_83970
	mov.s32 	%r88, 1;
$Lt_0_88322:
 //<loop> Loop body line 71, nesting depth: 2, iterations: 7
	add.s32 	%r89, %r88, %r84;
	cvt.rn.f32.s32 	%f382, %r89;
	mov.f32 	%f383, %f328;
	setp.gt.f32 	%p154, %f383, %f382;
	mov.f32 	%f384, %f326;
	setp.gt.f32 	%p155, %f384, %f382;
	mov.f32 	%f385, %f324;
	setp.gt.f32 	%p156, %f385, %f382;
	selp.s32 	%r90, 1, 0, %p156;
	xor.b32 	%r91, %r90, 1;
	selp.s32 	%r92, %r91, %r90, %p155;
	xor.b32 	%r93, %r92, 1;
	selp.s32 	%r94, %r93, %r92, %p154;
	xor.b32 	%r95, %r94, 1;
	mov.f32 	%f386, %f330;
	setp.gt.f32 	%p157, %f386, %f382;
	selp.s32 	%r96, %r95, %r94, %p157;
	mov.u32 	%r97, 0;
	setp.ne.s32 	%p158, %r96, %r97;
	@%p158 bra 	$Lt_0_88578;
 //<loop> Part of loop body line 71, head labeled $Lt_0_88322
	.loc	15	75	0
	mov.f32 	%f387, 0f00000000;   	// 0
	cvt.u64.s32 	%rd13, %r88;
	mul.lo.u64 	%rd14, %rd13, 16;
	add.u64 	%rd15, %rd12, %rd14;
	st.global.f32 	[%rd15+12], %f387;
$Lt_0_88578:
 //<loop> Part of loop body line 71, head labeled $Lt_0_88322
	add.s32 	%r88, %r88, 1;
	mov.u32 	%r98, 8;
	setp.ne.s32 	%p159, %r88, %r98;
	@%p159 bra 	$Lt_0_88322;
$Lt_0_86274:
 //<loop> Part of loop body line 51, head labeled $Lt_0_83970
	.loc	15	79	0
	add.u64 	%rd12, %rd12, 128;
	mov.s32 	%r84, %r87;
	mov.u32 	%r99, 31;
	setp.le.s32 	%p160, %r87, %r99;
	@%p160 bra 	$Lt_0_83970;
$LBB160_cuda_kernel_bitmask:
	.loc	15	101	0
	exit;
$LDWend_cuda_kernel_bitmask:
	} // cuda_kernel_bitmask

