const char * gb_gpu_kernel_nbor = 
"	.version 1.4\n"
"	.target sm_13\n"
"	.entry kernel_gb_nbor (\n"
"		.param .u64 __cudaparm_kernel_gb_nbor_x_,\n"
"		.param .u64 __cudaparm_kernel_gb_nbor_cut_form,\n"
"		.param .s32 __cudaparm_kernel_gb_nbor_ntypes,\n"
"		.param .u64 __cudaparm_kernel_gb_nbor_dev_nbor,\n"
"		.param .s32 __cudaparm_kernel_gb_nbor_nbor_pitch,\n"
"		.param .s32 __cudaparm_kernel_gb_nbor_start,\n"
"		.param .s32 __cudaparm_kernel_gb_nbor_inum,\n"
"		.param .u64 __cudaparm_kernel_gb_nbor_dev_ij,\n"
"		.param .s32 __cudaparm_kernel_gb_nbor_form_low,\n"
"		.param .s32 __cudaparm_kernel_gb_nbor_form_high,\n"
"		.param .s32 __cudaparm_kernel_gb_nbor_nall)\n"
"	{\n"
"	.reg .u32 %r<25>;\n"
"	.reg .u64 %rd<33>;\n"
"	.reg .f32 %f<20>;\n"
"	.reg .pred %p<9>;\n"
"	.loc	14	58	0\n"
"$LBB1_kernel_gb_nbor:\n"
"	cvt.s32.u16 	%r1, %ctaid.x;\n"
"	cvt.s32.u16 	%r2, %ntid.x;\n"
"	mul24.lo.s32 	%r3, %r1, %r2;\n"
"	cvt.u32.u16 	%r4, %tid.x;\n"
"	add.u32 	%r5, %r3, %r4;\n"
"	ld.param.u32 	%r6, [__cudaparm_kernel_gb_nbor_start];\n"
"	add.u32 	%r7, %r6, %r5;\n"
"	ld.param.s32 	%r8, [__cudaparm_kernel_gb_nbor_inum];\n"
"	setp.le.s32 	%p1, %r8, %r7;\n"
"	@%p1 bra 	$Lt_0_5122;\n"
"	.loc	14	65	0\n"
"	cvt.u64.s32 	%rd1, %r7;\n"
"	ld.param.u64 	%rd2, [__cudaparm_kernel_gb_nbor_dev_ij];\n"
"	mul.lo.u64 	%rd3, %rd1, 4;\n"
"	add.u64 	%rd4, %rd2, %rd3;\n"
"	ld.global.s32 	%r9, [%rd4+0];\n"
"	.loc	14	67	0\n"
"	ld.param.s32 	%r10, [__cudaparm_kernel_gb_nbor_nbor_pitch];\n"
"	cvt.u64.s32 	%rd5, %r10;\n"
"	mul.lo.u64 	%rd6, %rd5, 4;\n"
"	add.u64 	%rd7, %rd4, %rd6;\n"
"	ld.global.s32 	%r11, [%rd7+0];\n"
"	.loc	14	68	0\n"
"	add.u64 	%rd8, %rd7, %rd6;\n"
"	mov.s64 	%rd9, %rd8;\n"
"	.loc	14	70	0\n"
"	ld.param.u64 	%rd10, [__cudaparm_kernel_gb_nbor_dev_nbor];\n"
"	add.u64 	%rd11, %rd5, %rd1;\n"
"	add.u64 	%rd12, %rd5, %rd11;\n"
"	mul.lo.u64 	%rd13, %rd12, 4;\n"
"	add.u64 	%rd14, %rd10, %rd13;\n"
"	.loc	14	72	0\n"
"	ld.param.u64 	%rd15, [__cudaparm_kernel_gb_nbor_x_];\n"
"	cvt.u64.s32 	%rd16, %r9;\n"
"	mul.lo.u64 	%rd17, %rd16, 16;\n"
"	add.u64 	%rd18, %rd15, %rd17;\n"
"	ld.global.v4.f32 	{%f1,%f2,%f3,%f4}, [%rd18+0];\n"
"	mul24.lo.s32 	%r12, %r11, %r10;\n"
"	cvt.s64.s32 	%rd19, %r12;\n"
"	mul.lo.u64 	%rd20, %rd19, 4;\n"
"	add.u64 	%rd21, %rd8, %rd20;\n"
"	setp.ge.u64 	%p2, %rd8, %rd21;\n"
"	mov.s32 	%r13, 0;\n"
"	@%p2 bra 	$Lt_0_7682;\n"
"	cvt.rzi.s32.f32 	%r14, %f4;\n"
"	ld.param.s32 	%r15, [__cudaparm_kernel_gb_nbor_form_low];\n"
"	cvt.rn.f32.s32 	%f5, %r15;\n"
"	ld.param.s32 	%r16, [__cudaparm_kernel_gb_nbor_ntypes];\n"
"	mul24.lo.s32 	%r17, %r14, %r16;\n"
"	ld.param.u64 	%rd22, [__cudaparm_kernel_gb_nbor_cut_form];\n"
"	ld.param.s32 	%r18, [__cudaparm_kernel_gb_nbor_nall];\n"
"$Lt_0_6146:\n"
"	.loc	14	77	0\n"
"	ld.global.s32 	%r19, [%rd9+0];\n"
"	setp.lt.s32 	%p3, %r19, %r18;\n"
"	@%p3 bra 	$Lt_0_6402;\n"
"	.loc	14	79	0\n"
"	rem.s32 	%r19, %r19, %r18;\n"
"$Lt_0_6402:\n"
"	.loc	14	80	0\n"
"	cvt.u64.s32 	%rd23, %r19;\n"
"	mul.lo.u64 	%rd24, %rd23, 16;\n"
"	add.u64 	%rd25, %rd15, %rd24;\n"
"	ld.global.v4.f32 	{%f6,%f7,%f8,%f9}, [%rd25+0];\n"
"	.loc	14	83	0\n"
"	cvt.rzi.s32.f32 	%r20, %f9;\n"
"	add.s32 	%r21, %r20, %r17;\n"
"	cvt.u64.s32 	%rd26, %r21;\n"
"	mul.lo.u64 	%rd27, %rd26, 8;\n"
"	add.u64 	%rd28, %rd22, %rd27;\n"
"	ld.global.v2.f32 	{%f10,%f11}, [%rd28+0];\n"
"	.loc	14	80	0\n"
"	setp.le.f32 	%p4, %f5, %f11;\n"
"	@!%p4 bra 	$Lt_0_7938;\n"
"	ld.param.s32 	%r22, [__cudaparm_kernel_gb_nbor_form_high];\n"
"	cvt.rn.f32.s32 	%f12, %r22;\n"
"	setp.ge.f32 	%p5, %f12, %f11;\n"
"	@!%p5 bra 	$Lt_0_7938;\n"
"	sub.f32 	%f13, %f6, %f1;\n"
"	sub.f32 	%f14, %f7, %f2;\n"
"	sub.f32 	%f15, %f8, %f3;\n"
"	mul.f32 	%f16, %f13, %f13;\n"
"	mad.f32 	%f17, %f14, %f14, %f16;\n"
"	mad.f32 	%f18, %f15, %f15, %f17;\n"
"	setp.gt.f32 	%p6, %f10, %f18;\n"
"	@!%p6 bra 	$Lt_0_7938;\n"
"	.loc	14	94	0\n"
"	st.global.s32 	[%rd14+0], %r19;\n"
"	.loc	14	95	0\n"
"	add.u64 	%rd14, %rd6, %rd14;\n"
"	.loc	14	96	0\n"
"	add.s32 	%r13, %r13, 1;\n"
"$Lt_0_7938:\n"
"$L_0_4610:\n"
"	.loc	14	76	0\n"
"	add.u64 	%rd9, %rd6, %rd9;\n"
"	setp.gt.u64 	%p7, %rd21, %rd9;\n"
"	@%p7 bra 	$Lt_0_6146;\n"
"	bra.uni 	$Lt_0_5634;\n"
"$Lt_0_7682:\n"
"$Lt_0_5634:\n"
"	.loc	14	100	0\n"
"	add.s32 	%r23, %r10, %r7;\n"
"	cvt.s64.s32 	%rd29, %r23;\n"
"	mul.lo.u64 	%rd30, %rd29, 4;\n"
"	add.u64 	%rd31, %rd10, %rd30;\n"
"	st.global.s32 	[%rd31+0], %r13;\n"
"$Lt_0_5122:\n"
"	.loc	14	102	0\n"
"	exit;\n"
"$LDWend_kernel_gb_nbor:\n"
"	}\n"
"	.entry kernel_gb_nbor_fast (\n"
"		.param .u64 __cudaparm_kernel_gb_nbor_fast_x_,\n"
"		.param .u64 __cudaparm_kernel_gb_nbor_fast_cut_form,\n"
"		.param .u64 __cudaparm_kernel_gb_nbor_fast_dev_nbor,\n"
"		.param .s32 __cudaparm_kernel_gb_nbor_fast_nbor_pitch,\n"
"		.param .s32 __cudaparm_kernel_gb_nbor_fast_start,\n"
"		.param .s32 __cudaparm_kernel_gb_nbor_fast_inum,\n"
"		.param .u64 __cudaparm_kernel_gb_nbor_fast_dev_ij,\n"
"		.param .s32 __cudaparm_kernel_gb_nbor_fast_form_low,\n"
"		.param .s32 __cudaparm_kernel_gb_nbor_fast_form_high,\n"
"		.param .s32 __cudaparm_kernel_gb_nbor_fast_nall)\n"
"	{\n"
"	.reg .u32 %r<28>;\n"
"	.reg .u64 %rd<42>;\n"
"	.reg .f32 %f<19>;\n"
"	.reg .pred %p<10>;\n"
"	.shared .align 4 .b8 __cuda_cutsq132[256];\n"
"	.shared .align 4 .b8 __cuda_form388[256];\n"
"	.loc	14	116	0\n"
"$LBB1_kernel_gb_nbor_fast:\n"
"	cvt.s32.u16 	%r1, %tid.x;\n"
"	mov.u32 	%r2, 63;\n"
"	setp.gt.s32 	%p1, %r1, %r2;\n"
"	@%p1 bra 	$Lt_1_5890;\n"
"	.loc	14	122	0\n"
"	mov.u64 	%rd1, __cuda_cutsq132;\n"
"	mov.u64 	%rd2, __cuda_form388;\n"
"	cvt.u64.s32 	%rd3, %r1;\n"
"	mul.lo.u64 	%rd4, %rd3, 4;\n"
"	ld.param.u64 	%rd5, [__cudaparm_kernel_gb_nbor_fast_cut_form];\n"
"	mul.lo.u64 	%rd6, %rd3, 8;\n"
"	add.u64 	%rd7, %rd5, %rd6;\n"
"	ld.global.v2.f32 	{%f1,%f2}, [%rd7+0];\n"
"	add.u64 	%rd8, %rd4, %rd1;\n"
"	st.shared.f32 	[%rd8+0], %f1;\n"
"	.loc	14	123	0\n"
"	cvt.rzi.s32.f32 	%r3, %f2;\n"
"	add.u64 	%rd9, %rd4, %rd2;\n"
"	st.shared.s32 	[%rd9+0], %r3;\n"
"$Lt_1_5890:\n"
"	mov.u64 	%rd1, __cuda_cutsq132;\n"
"	mov.u64 	%rd2, __cuda_form388;\n"
"	.loc	14	126	0\n"
"	bar.sync 	0;\n"
"	cvt.s32.u16 	%r4, %ntid.x;\n"
"	cvt.s32.u16 	%r5, %ctaid.x;\n"
"	mul24.lo.s32 	%r6, %r4, %r5;\n"
"	ld.param.s32 	%r7, [__cudaparm_kernel_gb_nbor_fast_start];\n"
"	add.s32 	%r8, %r7, %r6;\n"
"	add.s32 	%r9, %r8, %r1;\n"
"	ld.param.s32 	%r10, [__cudaparm_kernel_gb_nbor_fast_inum];\n"
"	setp.le.s32 	%p2, %r10, %r9;\n"
"	@%p2 bra 	$Lt_1_6402;\n"
"	.loc	14	130	0\n"
"	cvt.u64.s32 	%rd10, %r9;\n"
"	ld.param.u64 	%rd11, [__cudaparm_kernel_gb_nbor_fast_dev_ij];\n"
"	mul.lo.u64 	%rd12, %rd10, 4;\n"
"	add.u64 	%rd13, %rd11, %rd12;\n"
"	ld.global.s32 	%r11, [%rd13+0];\n"
"	.loc	14	132	0\n"
"	ld.param.s32 	%r12, [__cudaparm_kernel_gb_nbor_fast_nbor_pitch];\n"
"	cvt.u64.s32 	%rd14, %r12;\n"
"	mul.lo.u64 	%rd15, %rd14, 4;\n"
"	add.u64 	%rd16, %rd13, %rd15;\n"
"	ld.global.s32 	%r13, [%rd16+0];\n"
"	.loc	14	133	0\n"
"	add.u64 	%rd17, %rd16, %rd15;\n"
"	mov.s64 	%rd18, %rd17;\n"
"	.loc	14	135	0\n"
"	ld.param.u64 	%rd19, [__cudaparm_kernel_gb_nbor_fast_dev_nbor];\n"
"	add.u64 	%rd20, %rd14, %rd10;\n"
"	add.u64 	%rd21, %rd14, %rd20;\n"
"	mul.lo.u64 	%rd22, %rd21, 4;\n"
"	add.u64 	%rd23, %rd19, %rd22;\n"
"	.loc	14	137	0\n"
"	ld.param.u64 	%rd24, [__cudaparm_kernel_gb_nbor_fast_x_];\n"
"	cvt.u64.s32 	%rd25, %r11;\n"
"	mul.lo.u64 	%rd26, %rd25, 16;\n"
"	add.u64 	%rd27, %rd24, %rd26;\n"
"	ld.global.v4.f32 	{%f3,%f4,%f5,%f6}, [%rd27+0];\n"
"	mul24.lo.s32 	%r14, %r13, %r12;\n"
"	cvt.s64.s32 	%rd28, %r14;\n"
"	mul.lo.u64 	%rd29, %rd28, 4;\n"
"	add.u64 	%rd30, %rd17, %rd29;\n"
"	setp.ge.u64 	%p3, %rd17, %rd30;\n"
"	mov.s32 	%r15, 0;\n"
"	@%p3 bra 	$Lt_1_8962;\n"
"	cvt.rzi.s32.f32 	%r16, %f6;\n"
"	mov.s32 	%r17, 8;\n"
"	mul24.lo.s32 	%r18, %r17, %r16;\n"
"	ld.param.s32 	%r19, [__cudaparm_kernel_gb_nbor_fast_form_low];\n"
"	ld.param.s32 	%r20, [__cudaparm_kernel_gb_nbor_fast_nall];\n"
"$Lt_1_7426:\n"
"	.loc	14	143	0\n"
"	ld.global.s32 	%r21, [%rd18+0];\n"
"	setp.lt.s32 	%p4, %r21, %r20;\n"
"	@%p4 bra 	$Lt_1_7682;\n"
"	.loc	14	145	0\n"
"	rem.s32 	%r21, %r21, %r20;\n"
"$Lt_1_7682:\n"
"	.loc	14	146	0\n"
"	cvt.u64.s32 	%rd31, %r21;\n"
"	mul.lo.u64 	%rd32, %rd31, 16;\n"
"	add.u64 	%rd33, %rd24, %rd32;\n"
"	ld.global.v4.f32 	{%f7,%f8,%f9,%f10}, [%rd33+0];\n"
"	cvt.rzi.s32.f32 	%r22, %f10;\n"
"	add.s32 	%r23, %r22, %r18;\n"
"	cvt.u64.s32 	%rd34, %r23;\n"
"	mul.lo.u64 	%rd35, %rd34, 4;\n"
"	add.u64 	%rd36, %rd35, %rd2;\n"
"	ld.shared.s32 	%r24, [%rd36+0];\n"
"	setp.lt.s32 	%p5, %r24, %r19;\n"
"	@%p5 bra 	$Lt_1_9218;\n"
"	ld.param.s32 	%r25, [__cudaparm_kernel_gb_nbor_fast_form_high];\n"
"	ld.shared.s32 	%r24, [%rd36+0];\n"
"	setp.lt.s32 	%p6, %r25, %r24;\n"
"	@%p6 bra 	$Lt_1_9218;\n"
"	sub.f32 	%f11, %f7, %f3;\n"
"	sub.f32 	%f12, %f8, %f4;\n"
"	sub.f32 	%f13, %f9, %f5;\n"
"	add.u64 	%rd37, %rd35, %rd1;\n"
"	ld.shared.f32 	%f14, [%rd37+0];\n"
"	mul.f32 	%f15, %f11, %f11;\n"
"	mad.f32 	%f16, %f12, %f12, %f15;\n"
"	mad.f32 	%f17, %f13, %f13, %f16;\n"
"	setp.gt.f32 	%p7, %f14, %f17;\n"
"	@!%p7 bra 	$Lt_1_9218;\n"
"	.loc	14	160	0\n"
"	st.global.s32 	[%rd23+0], %r21;\n"
"	.loc	14	161	0\n"
"	add.u64 	%rd23, %rd15, %rd23;\n"
"	.loc	14	162	0\n"
"	add.s32 	%r15, %r15, 1;\n"
"$Lt_1_9218:\n"
"$L_1_5378:\n"
"	.loc	14	142	0\n"
"	add.u64 	%rd18, %rd15, %rd18;\n"
"	setp.gt.u64 	%p8, %rd30, %rd18;\n"
"	@%p8 bra 	$Lt_1_7426;\n"
"	bra.uni 	$Lt_1_6914;\n"
"$Lt_1_8962:\n"
"$Lt_1_6914:\n"
"	.loc	14	166	0\n"
"	add.s32 	%r26, %r9, %r12;\n"
"	cvt.s64.s32 	%rd38, %r26;\n"
"	mul.lo.u64 	%rd39, %rd38, 4;\n"
"	add.u64 	%rd40, %rd19, %rd39;\n"
"	st.global.s32 	[%rd40+0], %r15;\n"
"$Lt_1_6402:\n"
"	.loc	14	168	0\n"
"	exit;\n"
"$LDWend_kernel_gb_nbor_fast:\n"
"	}\n"
;
const char * gb_gpu_kernel = 
"	.version 1.4\n"
"	.target sm_13\n"
"	.entry kernel_gayberne (\n"
"		.param .u64 __cudaparm_kernel_gayberne_x_,\n"
"		.param .u64 __cudaparm_kernel_gayberne_q,\n"
"		.param .u64 __cudaparm_kernel_gayberne_shape,\n"
"		.param .u64 __cudaparm_kernel_gayberne_well,\n"
"		.param .u64 __cudaparm_kernel_gayberne_gum,\n"
"		.param .u64 __cudaparm_kernel_gayberne_sig_eps,\n"
"		.param .s32 __cudaparm_kernel_gayberne_ntypes,\n"
"		.param .u64 __cudaparm_kernel_gayberne_lshape,\n"
"		.param .u64 __cudaparm_kernel_gayberne_dev_nbor,\n"
"		.param .s32 __cudaparm_kernel_gayberne_stride,\n"
"		.param .u64 __cudaparm_kernel_gayberne_ans,\n"
"		.param .s32 __cudaparm_kernel_gayberne_astride,\n"
"		.param .u64 __cudaparm_kernel_gayberne_engv,\n"
"		.param .u64 __cudaparm_kernel_gayberne_err_flag,\n"
"		.param .s32 __cudaparm_kernel_gayberne_eflag,\n"
"		.param .s32 __cudaparm_kernel_gayberne_vflag,\n"
"		.param .s32 __cudaparm_kernel_gayberne_inum,\n"
"		.param .s32 __cudaparm_kernel_gayberne_nall)\n"
"	{\n"
"	.reg .u32 %r<33>;\n"
"	.reg .u64 %rd<71>;\n"
"	.reg .f32 %f<862>;\n"
"	.reg .pred %p<26>;\n"
"	.shared .align 4 .b8 __cuda_sp_lj128[16];\n"
"	.loc	15	100	0\n"
"$LBB1_kernel_gayberne:\n"
"	cvt.s32.u16 	%r1, %tid.x;\n"
"	mov.u32 	%r2, 3;\n"
"	setp.gt.s32 	%p1, %r1, %r2;\n"
"	@%p1 bra 	$Lt_0_30466;\n"
"	.loc	15	106	0\n"
"	mov.u64 	%rd1, __cuda_sp_lj128;\n"
"	cvt.u64.s32 	%rd2, %r1;\n"
"	mul.lo.u64 	%rd3, %rd2, 4;\n"
"	ld.param.u64 	%rd4, [__cudaparm_kernel_gayberne_gum];\n"
"	add.u64 	%rd5, %rd4, %rd3;\n"
"	ld.global.f32 	%f1, [%rd5+12];\n"
"	add.u64 	%rd6, %rd3, %rd1;\n"
"	st.shared.f32 	[%rd6+0], %f1;\n"
"$Lt_0_30466:\n"
"	mov.u64 	%rd1, __cuda_sp_lj128;\n"
"	.loc	15	108	0\n"
"	bar.sync 	0;\n"
"	cvt.s32.u16 	%r3, %ctaid.x;\n"
"	cvt.s32.u16 	%r4, %ntid.x;\n"
"	mul24.lo.s32 	%r5, %r3, %r4;\n"
"	add.s32 	%r6, %r5, %r1;\n"
"	ld.param.s32 	%r7, [__cudaparm_kernel_gayberne_inum];\n"
"	setp.le.s32 	%p2, %r7, %r6;\n"
"	@%p2 bra 	$Lt_0_30978;\n"
"	.loc	15	123	0\n"
"	mov.f32 	%f2, 0f00000000;     	\n"
"	mov.f32 	%f3, %f2;\n"
"	mov.f32 	%f4, 0f00000000;     	\n"
"	mov.f32 	%f5, %f4;\n"
"	mov.f32 	%f6, 0f00000000;     	\n"
"	mov.f32 	%f7, %f6;\n"
"	mov.f32 	%f8, 0f00000000;     	\n"
"	mov.f32 	%f9, %f8;\n"
"	mov.f32 	%f10, 0f00000000;    	\n"
"	mov.f32 	%f11, %f10;\n"
"	mov.f32 	%f12, 0f00000000;    	\n"
"	mov.f32 	%f13, %f12;\n"
"	.loc	15	126	0\n"
"	cvt.u64.s32 	%rd7, %r6;\n"
"	mul.lo.u64 	%rd8, %rd7, 4;\n"
"	ld.param.u64 	%rd9, [__cudaparm_kernel_gayberne_dev_nbor];\n"
"	add.u64 	%rd10, %rd9, %rd8;\n"
"	ld.global.s32 	%r8, [%rd10+0];\n"
"	.loc	15	128	0\n"
"	ld.param.s32 	%r9, [__cudaparm_kernel_gayberne_stride];\n"
"	cvt.u64.s32 	%rd11, %r9;\n"
"	mul.lo.u64 	%rd12, %rd11, 4;\n"
"	add.u64 	%rd13, %rd10, %rd12;\n"
"	ld.global.s32 	%r10, [%rd13+0];\n"
"	.loc	15	129	0\n"
"	add.u64 	%rd14, %rd13, %rd12;\n"
"	mov.s64 	%rd15, %rd14;\n"
"	.loc	15	132	0\n"
"	cvt.u64.s32 	%rd16, %r8;\n"
"	mul.lo.u64 	%rd17, %rd16, 16;\n"
"	ld.param.u64 	%rd18, [__cudaparm_kernel_gayberne_x_];\n"
"	add.u64 	%rd19, %rd17, %rd18;\n"
"	ld.global.v4.f32 	{%f14,%f15,%f16,%f17}, [%rd19+0];\n"
"	.loc	15	135	0\n"
"	cvt.rzi.s32.f32 	%r11, %f17;\n"
"	cvt.u64.s32 	%rd20, %r11;\n"
"	mul.lo.u64 	%rd21, %rd20, 16;\n"
"	ld.param.u64 	%rd22, [__cudaparm_kernel_gayberne_shape];\n"
"	add.u64 	%rd23, %rd21, %rd22;\n"
"	ld.global.v4.f32 	{%f18,%f19,%f20,_}, [%rd23+0];\n"
"	.loc	15	138	0\n"
"	ld.param.u64 	%rd24, [__cudaparm_kernel_gayberne_q];\n"
"	add.u64 	%rd25, %rd17, %rd24;\n"
"	ld.global.v4.f32 	{%f21,%f22,%f23,%f24}, [%rd25+0];\n"
"	.loc	15	141	0\n"
"	ld.param.u64 	%rd26, [__cudaparm_kernel_gayberne_well];\n"
"	add.u64 	%rd27, %rd21, %rd26;\n"
"	ld.global.v4.f32 	{%f25,%f26,%f27,_}, [%rd27+0];\n"
"	mul24.lo.s32 	%r12, %r9, %r10;\n"
"	cvt.s64.s32 	%rd28, %r12;\n"
"	mul.lo.u64 	%rd29, %rd28, 4;\n"
"	add.u64 	%rd30, %rd14, %rd29;\n"
"	ld.param.s32 	%r13, [__cudaparm_kernel_gayberne_vflag];\n"
"	ld.param.s32 	%r14, [__cudaparm_kernel_gayberne_eflag];\n"
"	setp.ge.u64 	%p3, %rd14, %rd30;\n"
"	mov.f32 	%f28, 0f00000000;    	\n"
"	mov.f32 	%f29, 0f00000000;    	\n"
"	mov.f32 	%f30, 0f00000000;    	\n"
"	mov.f32 	%f31, 0f00000000;    	\n"
"	mov.f32 	%f32, 0f00000000;    	\n"
"	mov.f32 	%f33, 0f00000000;    	\n"
"	mov.f32 	%f34, 0f00000000;    	\n"
"	@%p3 bra 	$Lt_0_43266;\n"
"	mov.s32 	%r15, 0;\n"
"	setp.gt.s32 	%p4, %r14, %r15;\n"
"	mov.s32 	%r16, 0;\n"
"	setp.gt.s32 	%p5, %r13, %r16;\n"
"	add.f32 	%f35, %f22, %f22;\n"
"	add.f32 	%f36, %f24, %f24;\n"
"	mul.f32 	%f37, %f21, %f21;\n"
"	mul.f32 	%f38, %f22, %f22;\n"
"	mul.f32 	%f39, %f23, %f23;\n"
"	mul.f32 	%f40, %f24, %f24;\n"
"	add.f32 	%f41, %f23, %f23;\n"
"	ld.param.s32 	%r17, [__cudaparm_kernel_gayberne_ntypes];\n"
"	mul24.lo.s32 	%r18, %r17, %r11;\n"
"	mul.f32 	%f42, %f35, %f23;\n"
"	mul.f32 	%f43, %f35, %f24;\n"
"	mul.f32 	%f44, %f35, %f21;\n"
"	mul.f32 	%f45, %f36, %f21;\n"
"	add.f32 	%f46, %f37, %f38;\n"
"	sub.f32 	%f47, %f37, %f38;\n"
"	mul.f32 	%f48, %f41, %f21;\n"
"	mul.f32 	%f49, %f41, %f24;\n"
"	sub.f32 	%f50, %f42, %f45;\n"
"	add.f32 	%f51, %f42, %f45;\n"
"	sub.f32 	%f52, %f46, %f39;\n"
"	add.f32 	%f53, %f39, %f47;\n"
"	sub.f32 	%f54, %f47, %f39;\n"
"	add.f32 	%f55, %f43, %f48;\n"
"	sub.f32 	%f56, %f43, %f48;\n"
"	sub.f32 	%f57, %f49, %f44;\n"
"	add.f32 	%f58, %f44, %f49;\n"
"	ld.param.u64 	%rd31, [__cudaparm_kernel_gayberne_lshape];\n"
"	mul.lo.u64 	%rd32, %rd20, 4;\n"
"	add.u64 	%rd33, %rd31, %rd32;\n"
"	mul.f32 	%f59, %f50, %f19;\n"
"	mul.f32 	%f60, %f50, %f26;\n"
"	mul.f32 	%f61, %f51, %f18;\n"
"	mul.f32 	%f62, %f51, %f25;\n"
"	sub.f32 	%f63, %f52, %f40;\n"
"	sub.f32 	%f64, %f53, %f40;\n"
"	add.f32 	%f65, %f40, %f54;\n"
"	mul.f32 	%f66, %f55, %f20;\n"
"	add.f32 	%f67, %f55, %f55;\n"
"	mul.f32 	%f68, %f55, %f27;\n"
"	mul.f32 	%f69, %f56, %f18;\n"
"	mul.f32 	%f70, %f56, %f25;\n"
"	mul.f32 	%f71, %f57, %f20;\n"
"	add.f32 	%f72, %f57, %f57;\n"
"	mul.f32 	%f73, %f57, %f27;\n"
"	mul.f32 	%f74, %f58, %f19;\n"
"	mul.f32 	%f75, %f58, %f26;\n"
"	mul.f32 	%f76, %f50, %f59;\n"
"	mul.f32 	%f77, %f58, %f59;\n"
"	mul.f32 	%f78, %f50, %f60;\n"
"	mul.f32 	%f79, %f58, %f60;\n"
"	mul.f32 	%f80, %f63, %f18;\n"
"	mul.f32 	%f81, %f63, %f25;\n"
"	mul.f32 	%f82, %f64, %f19;\n"
"	mul.f32 	%f83, %f59, %f64;\n"
"	mul.f32 	%f84, %f64, %f26;\n"
"	mul.f32 	%f85, %f60, %f64;\n"
"	mul.f32 	%f86, %f65, %f20;\n"
"	add.f32 	%f87, %f65, %f65;\n"
"	mul.f32 	%f88, %f65, %f27;\n"
"	mul.f32 	%f89, %f50, %f74;\n"
"	mul.f32 	%f90, %f64, %f74;\n"
"	mul.f32 	%f91, %f58, %f74;\n"
"	mul.f32 	%f92, %f50, %f75;\n"
"	mul.f32 	%f93, %f64, %f75;\n"
"	mul.f32 	%f94, %f58, %f75;\n"
"	mad.f32 	%f95, %f63, %f80, %f76;\n"
"	mad.f32 	%f96, %f80, %f56, %f77;\n"
"	mad.f32 	%f97, %f63, %f81, %f78;\n"
"	mad.f32 	%f98, %f81, %f56, %f79;\n"
"	mul.f32 	%f99, %f50, %f82;\n"
"	mul.f32 	%f100, %f64, %f82;\n"
"	mul.f32 	%f101, %f58, %f82;\n"
"	mad.f32 	%f102, %f80, %f51, %f83;\n"
"	mul.f32 	%f103, %f50, %f84;\n"
"	mul.f32 	%f104, %f64, %f84;\n"
"	mul.f32 	%f105, %f58, %f84;\n"
"	mad.f32 	%f106, %f81, %f51, %f85;\n"
"	mad.f32 	%f107, %f63, %f69, %f89;\n"
"	mad.f32 	%f108, %f51, %f69, %f90;\n"
"	mad.f32 	%f109, %f56, %f69, %f91;\n"
"	mad.f32 	%f110, %f63, %f70, %f92;\n"
"	mad.f32 	%f111, %f51, %f70, %f93;\n"
"	mad.f32 	%f112, %f56, %f70, %f94;\n"
"	mad.f32 	%f113, %f55, %f66, %f95;\n"
"	mad.f32 	%f114, %f66, %f65, %f96;\n"
"	mad.f32 	%f115, %f55, %f68, %f97;\n"
"	mad.f32 	%f116, %f68, %f65, %f98;\n"
"	mad.f32 	%f117, %f63, %f61, %f99;\n"
"	mad.f32 	%f118, %f51, %f61, %f100;\n"
"	mad.f32 	%f119, %f61, %f56, %f101;\n"
"	mad.f32 	%f120, %f66, %f57, %f102;\n"
"	mad.f32 	%f121, %f63, %f62, %f103;\n"
"	mad.f32 	%f122, %f51, %f62, %f104;\n"
"	mad.f32 	%f123, %f62, %f56, %f105;\n"
"	mad.f32 	%f124, %f68, %f57, %f106;\n"
"	mad.f32 	%f125, %f55, %f86, %f107;\n"
"	mad.f32 	%f126, %f57, %f86, %f108;\n"
"	mad.f32 	%f127, %f65, %f86, %f109;\n"
"	mad.f32 	%f128, %f55, %f88, %f110;\n"
"	mad.f32 	%f129, %f57, %f88, %f111;\n"
"	mad.f32 	%f130, %f65, %f88, %f112;\n"
"	mad.f32 	%f131, %f55, %f71, %f117;\n"
"	mad.f32 	%f132, %f57, %f71, %f118;\n"
"	mad.f32 	%f133, %f71, %f65, %f119;\n"
"	mad.f32 	%f134, %f55, %f73, %f121;\n"
"	mad.f32 	%f135, %f57, %f73, %f122;\n"
"	mad.f32 	%f136, %f73, %f65, %f123;\n"
"	ld.param.u64 	%rd34, [__cudaparm_kernel_gayberne_gum];\n"
"	ld.param.u64 	%rd35, [__cudaparm_kernel_gayberne_sig_eps];\n"
"$Lt_0_32002:\n"
"	.loc	15	148	0\n"
"	ld.global.s32 	%r19, [%rd15+0];\n"
"	.loc	15	149	0\n"
"	shr.s32 	%r20, %r19, 30;\n"
"	cvt.s64.s32 	%rd36, %r20;\n"
"	and.b64 	%rd37, %rd36, 3;\n"
"	mul.lo.u64 	%rd38, %rd37, 4;\n"
"	add.u64 	%rd39, %rd1, %rd38;\n"
"	ld.shared.f32 	%f137, [%rd39+0];\n"
"	.loc	15	152	0\n"
"	and.b32 	%r21, %r19, 1073741823;\n"
"	cvt.u64.s32 	%rd40, %r21;\n"
"	mul.lo.u64 	%rd41, %rd40, 16;\n"
"	add.u64 	%rd42, %rd41, %rd18;\n"
"	ld.global.v4.f32 	{%f138,%f139,%f140,%f141}, [%rd42+0];\n"
"	.loc	15	166	0\n"
"	add.u64 	%rd43, %rd41, %rd24;\n"
"	ld.global.v4.f32 	{%f142,%f143,%f144,%f145}, [%rd43+0];\n"
"	.loc	15	175	0\n"
"	cvt.rzi.s32.f32 	%r22, %f141;\n"
"	cvt.u64.s32 	%rd44, %r22;\n"
"	mul.lo.u64 	%rd45, %rd44, 16;\n"
"	add.u64 	%rd46, %rd45, %rd22;\n"
"	ld.global.v4.f32 	{%f146,%f147,%f148,_}, [%rd46+0];\n"
"	.loc	15	184	0\n"
"	sub.f32 	%f149, %f138, %f14;\n"
"	mov.f32 	%f150, %f149;\n"
"	add.f32 	%f151, %f143, %f143;\n"
"	add.f32 	%f152, %f145, %f145;\n"
"	mul.f32 	%f153, %f142, %f142;\n"
"	mul.f32 	%f154, %f143, %f143;\n"
"	mul.f32 	%f155, %f144, %f144;\n"
"	mul.f32 	%f156, %f145, %f145;\n"
"	add.f32 	%f157, %f144, %f144;\n"
"	mul.f32 	%f158, %f151, %f144;\n"
"	mul.f32 	%f159, %f151, %f145;\n"
"	mul.f32 	%f160, %f152, %f142;\n"
"	add.f32 	%f161, %f153, %f154;\n"
"	mul.f32 	%f162, %f157, %f142;\n"
"	sub.f32 	%f163, %f158, %f160;\n"
"	sub.f32 	%f164, %f161, %f155;\n"
"	add.f32 	%f165, %f159, %f162;\n"
"	mul.f32 	%f166, %f163, %f147;\n"
"	sub.f32 	%f167, %f164, %f156;\n"
"	mul.f32 	%f168, %f165, %f148;\n"
"	mul.f32 	%f169, %f163, %f166;\n"
"	mul.f32 	%f170, %f167, %f146;\n"
"	mad.f32 	%f171, %f167, %f170, %f169;\n"
"	mad.f32 	%f172, %f165, %f168, %f171;\n"
"	add.f32 	%f173, %f113, %f172;\n"
"	mov.f32 	%f174, %f173;\n"
"	mul.f32 	%f175, %f151, %f142;\n"
"	sub.f32 	%f176, %f153, %f154;\n"
"	mul.f32 	%f177, %f157, %f145;\n"
"	add.f32 	%f178, %f158, %f160;\n"
"	add.f32 	%f179, %f155, %f176;\n"
"	sub.f32 	%f180, %f177, %f175;\n"
"	mul.f32 	%f181, %f178, %f146;\n"
"	sub.f32 	%f182, %f179, %f156;\n"
"	mul.f32 	%f183, %f180, %f148;\n"
"	mul.f32 	%f184, %f182, %f147;\n"
"	mul.f32 	%f185, %f163, %f184;\n"
"	mad.f32 	%f186, %f167, %f181, %f185;\n"
"	mad.f32 	%f187, %f165, %f183, %f186;\n"
"	add.f32 	%f188, %f131, %f187;\n"
"	mov.f32 	%f189, %f188;\n"
"	sub.f32 	%f190, %f176, %f155;\n"
"	sub.f32 	%f191, %f159, %f162;\n"
"	add.f32 	%f192, %f175, %f177;\n"
"	add.f32 	%f193, %f156, %f190;\n"
"	mul.f32 	%f194, %f191, %f146;\n"
"	mul.f32 	%f195, %f192, %f147;\n"
"	mul.f32 	%f196, %f193, %f148;\n"
"	mul.f32 	%f197, %f163, %f195;\n"
"	mad.f32 	%f198, %f167, %f194, %f197;\n"
"	mad.f32 	%f199, %f165, %f196, %f198;\n"
"	add.f32 	%f200, %f125, %f199;\n"
"	mov.f32 	%f201, %f200;\n"
"	sub.f32 	%f202, %f139, %f15;\n"
"	mov.f32 	%f203, %f202;\n"
"	mul.f32 	%f204, %f166, %f182;\n"
"	mad.f32 	%f205, %f170, %f178, %f204;\n"
"	mad.f32 	%f206, %f168, %f180, %f205;\n"
"	add.f32 	%f207, %f120, %f206;\n"
"	mov.f32 	%f208, %f207;\n"
"	mul.f32 	%f209, %f182, %f184;\n"
"	mad.f32 	%f210, %f178, %f181, %f209;\n"
"	mad.f32 	%f211, %f180, %f183, %f210;\n"
"	add.f32 	%f212, %f132, %f211;\n"
"	mov.f32 	%f213, %f212;\n"
"	mul.f32 	%f214, %f182, %f195;\n"
"	mad.f32 	%f215, %f178, %f194, %f214;\n"
"	mad.f32 	%f216, %f180, %f196, %f215;\n"
"	add.f32 	%f217, %f126, %f216;\n"
"	mov.f32 	%f218, %f217;\n"
"	sub.f32 	%f219, %f140, %f16;\n"
"	mov.f32 	%f220, %f219;\n"
"	mul.f32 	%f221, %f192, %f166;\n"
"	mad.f32 	%f222, %f170, %f191, %f221;\n"
"	mad.f32 	%f223, %f168, %f193, %f222;\n"
"	add.f32 	%f224, %f114, %f223;\n"
"	mov.f32 	%f225, %f224;\n"
"	mul.f32 	%f226, %f192, %f184;\n"
"	mad.f32 	%f227, %f181, %f191, %f226;\n"
"	mad.f32 	%f228, %f183, %f193, %f227;\n"
"	add.f32 	%f229, %f133, %f228;\n"
"	mov.f32 	%f230, %f229;\n"
"	mul.f32 	%f231, %f192, %f195;\n"
"	mad.f32 	%f232, %f191, %f194, %f231;\n"
"	mad.f32 	%f233, %f193, %f196, %f232;\n"
"	add.f32 	%f234, %f127, %f233;\n"
"	mov.f32 	%f235, %f234;\n"
"	.loc	14	186	0\n"
"	abs.f32 	%f236, %f207;\n"
"	abs.f32 	%f237, %f173;\n"
"	setp.gt.f32 	%p6, %f236, %f237;\n"
"	@!%p6 bra 	$Lt_0_32258;\n"
"	mov.f32 	%f174, %f207;\n"
"	mov.f32 	%f208, %f173;\n"
"	mov.f32 	%f189, %f212;\n"
"	mov.f32 	%f213, %f188;\n"
"	mov.f32 	%f201, %f217;\n"
"	mov.f32 	%f218, %f200;\n"
"	mov.f32 	%f150, %f202;\n"
"	mov.f32 	%f203, %f149;\n"
"$Lt_0_32258:\n"
"	.loc	14	193	0\n"
"	mov.f32 	%f238, %f174;\n"
"	abs.f32 	%f239, %f238;\n"
"	abs.f32 	%f240, %f224;\n"
"	setp.lt.f32 	%p7, %f239, %f240;\n"
"	@!%p7 bra 	$Lt_0_32770;\n"
"	mov.f32 	%f174, %f224;\n"
"	mov.f32 	%f225, %f238;\n"
"	.loc	14	196	0\n"
"	mov.f32 	%f241, %f189;\n"
"	mov.f32 	%f189, %f229;\n"
"	mov.f32 	%f230, %f241;\n"
"	.loc	14	197	0\n"
"	mov.f32 	%f241, %f201;\n"
"	mov.f32 	%f201, %f234;\n"
"	mov.f32 	%f235, %f241;\n"
"	.loc	14	198	0\n"
"	mov.f32 	%f241, %f150;\n"
"	mov.f32 	%f150, %f219;\n"
"	mov.f32 	%f220, %f241;\n"
"$Lt_0_32770:\n"
"	mov.f32 	%f242, %f174;\n"
"	mov.f32 	%f243, 0f00000000;   	\n"
"	setp.neu.f32 	%p8, %f242, %f243;\n"
"	@!%p8 bra 	$Lt_0_33538;\n"
"	bra.uni 	$Lt_0_34306;\n"
"$Lt_0_33538:\n"
"	mov.f32 	%f244, 0f00000000;   	\n"
"	setp.neu.f32 	%p9, %f208, %f244;\n"
"	@!%p9 bra 	$Lt_0_34050;\n"
"	mov.f32 	%f174, %f208;\n"
"	mov.f32 	%f208, %f242;\n"
"	.loc	14	213	0\n"
"	mov.f32 	%f245, %f189;\n"
"	mov.f32 	%f189, %f213;\n"
"	mov.f32 	%f213, %f245;\n"
"	.loc	14	214	0\n"
"	mov.f32 	%f246, %f201;\n"
"	mov.f32 	%f201, %f218;\n"
"	mov.f32 	%f218, %f246;\n"
"	.loc	14	215	0\n"
"	mov.f32 	%f247, %f150;\n"
"	mov.f32 	%f150, %f203;\n"
"	mov.f32 	%f203, %f247;\n"
"	bra.uni 	$Lt_0_34306;\n"
"$Lt_0_34050:\n"
"	mov.f32 	%f248, 0f00000000;   	\n"
"	setp.neu.f32 	%p10, %f225, %f248;\n"
"	@!%p10 bra 	$Lt_0_34562;\n"
"	mov.f32 	%f174, %f225;\n"
"	mov.f32 	%f225, %f242;\n"
"	.loc	14	221	0\n"
"	mov.f32 	%f249, %f189;\n"
"	mov.f32 	%f189, %f230;\n"
"	mov.f32 	%f230, %f249;\n"
"	.loc	14	222	0\n"
"	mov.f32 	%f250, %f201;\n"
"	mov.f32 	%f201, %f235;\n"
"	mov.f32 	%f235, %f250;\n"
"	.loc	14	223	0\n"
"	mov.f32 	%f251, %f150;\n"
"	mov.f32 	%f150, %f220;\n"
"	mov.f32 	%f220, %f251;\n"
"	bra.uni 	$Lt_0_34306;\n"
"$Lt_0_34562:\n"
"	.loc	14	226	0\n"
"	mov.s32 	%r23, 2;\n"
"	ld.param.u64 	%rd47, [__cudaparm_kernel_gayberne_err_flag];\n"
"	st.global.s32 	[%rd47+0], %r23;\n"
"$Lt_0_34306:\n"
"$Lt_0_33794:\n"
"$Lt_0_33282:\n"
"	.loc	14	228	0\n"
"	div.approx.f32 	%f252, %f208, %f174;\n"
"	mul.f32 	%f253, %f189, %f252;\n"
"	sub.f32 	%f254, %f213, %f253;\n"
"	mov.f32 	%f213, %f254;\n"
"	mul.f32 	%f255, %f201, %f252;\n"
"	sub.f32 	%f256, %f218, %f255;\n"
"	mov.f32 	%f218, %f256;\n"
"	mul.f32 	%f257, %f150, %f252;\n"
"	sub.f32 	%f258, %f203, %f257;\n"
"	mov.f32 	%f203, %f258;\n"
"	.loc	14	232	0\n"
"	div.approx.f32 	%f259, %f225, %f174;\n"
"	mul.f32 	%f260, %f189, %f259;\n"
"	sub.f32 	%f230, %f230, %f260;\n"
"	mul.f32 	%f261, %f201, %f259;\n"
"	sub.f32 	%f235, %f235, %f261;\n"
"	mul.f32 	%f262, %f150, %f259;\n"
"	sub.f32 	%f220, %f220, %f262;\n"
"	.loc	14	237	0\n"
"	abs.f32 	%f263, %f254;\n"
"	abs.f32 	%f264, %f230;\n"
"	setp.lt.f32 	%p11, %f263, %f264;\n"
"	@!%p11 bra 	$Lt_0_34818;\n"
"	mov.f32 	%f213, %f230;\n"
"	mov.f32 	%f230, %f254;\n"
"	mov.f32 	%f218, %f235;\n"
"	mov.f32 	%f235, %f256;\n"
"	mov.f32 	%f203, %f220;\n"
"	mov.f32 	%f220, %f258;\n"
"$Lt_0_34818:\n"
"	mov.f32 	%f265, %f213;\n"
"	mov.f32 	%f266, 0f00000000;   	\n"
"	setp.neu.f32 	%p12, %f265, %f266;\n"
"	@!%p12 bra 	$Lt_0_35586;\n"
"	bra.uni 	$Lt_0_35842;\n"
"$Lt_0_35586:\n"
"	mov.f32 	%f267, 0f00000000;   	\n"
"	setp.neu.f32 	%p13, %f230, %f267;\n"
"	@!%p13 bra 	$Lt_0_35842;\n"
"	mov.f32 	%f213, %f230;\n"
"	mov.f32 	%f230, %f265;\n"
"	.loc	14	258	0\n"
"	mov.f32 	%f268, %f218;\n"
"	mov.f32 	%f218, %f235;\n"
"	mov.f32 	%f235, %f268;\n"
"	.loc	14	259	0\n"
"	mov.f32 	%f268, %f203;\n"
"	mov.f32 	%f203, %f220;\n"
"	mov.f32 	%f220, %f268;\n"
"$Lt_0_35842:\n"
"$Lt_0_35330:\n"
"	.loc	14	263	0\n"
"	div.approx.f32 	%f269, %f230, %f213;\n"
"	mul.f32 	%f270, %f218, %f269;\n"
"	sub.f32 	%f235, %f235, %f270;\n"
"	mul.f32 	%f271, %f203, %f269;\n"
"	sub.f32 	%f220, %f220, %f271;\n"
"	mov.f32 	%f272, 0f00000000;   	\n"
"	setp.eq.f32 	%p14, %f235, %f272;\n"
"	@!%p14 bra 	$Lt_0_36354;\n"
"	.loc	14	268	0\n"
"	mov.s32 	%r24, 2;\n"
"	ld.param.u64 	%rd48, [__cudaparm_kernel_gayberne_err_flag];\n"
"	st.global.s32 	[%rd48+0], %r24;\n"
"$Lt_0_36354:\n"
"	.loc	15	192	0\n"
"	div.approx.f32 	%f273, %f220, %f235;\n"
"	mul.f32 	%f274, %f202, %f202;\n"
"	mul.f32 	%f275, %f273, %f218;\n"
"	mad.f32 	%f276, %f149, %f149, %f274;\n"
"	sub.f32 	%f277, %f203, %f275;\n"
"	mad.f32 	%f278, %f219, %f219, %f276;\n"
"	div.approx.f32 	%f279, %f277, %f213;\n"
"	rsqrt.approx.f32 	%f280, %f278;\n"
"	mul.f32 	%f281, %f279, %f189;\n"
"	mad.f32 	%f282, %f201, %f273, %f281;\n"
"	sub.f32 	%f283, %f150, %f282;\n"
"	div.approx.f32 	%f284, %f283, %f174;\n"
"	mul.f32 	%f285, %f280, %f284;\n"
"	.loc	15	204	0\n"
"	mul.f32 	%f286, %f279, %f280;\n"
"	mul.f32 	%f287, %f280, %f202;\n"
"	mul.f32 	%f288, %f280, %f149;\n"
"	mul.f32 	%f289, %f280, %f219;\n"
"	mul.f32 	%f290, %f273, %f280;\n"
"	mul.f32 	%f291, %f286, %f287;\n"
"	mad.f32 	%f292, %f288, %f285, %f291;\n"
"	mad.f32 	%f293, %f289, %f290, %f292;\n"
"	mov.f32 	%f294, 0f3f000000;   	\n"
"	mul.f32 	%f295, %f293, %f294;\n"
"	rsqrt.approx.f32 	%f296, %f295;\n"
"	.loc	15	208	0\n"
"	rcp.approx.f32 	%f297, %f280;\n"
"	mul.f32 	%f285, %f297, %f285;\n"
"	.loc	15	213	0\n"
"	add.s32 	%r25, %r22, %r18;\n"
"	cvt.u64.s32 	%rd49, %r25;\n"
"	mul.lo.u64 	%rd50, %rd49, 8;\n"
"	add.u64 	%rd51, %rd35, %rd50;\n"
"	ld.global.v2.f32 	{%f298,%f299}, [%rd51+0];\n"
"	.loc	15	215	0\n"
"	sub.f32 	%f300, %f297, %f296;\n"
"	ld.global.f32 	%f301, [%rd34+0];\n"
"	mad.f32 	%f302, %f301, %f298, %f300;\n"
"	.loc	15	222	0\n"
"	div.approx.f32 	%f303, %f298, %f302;\n"
"	mul.f32 	%f304, %f303, %f303;\n"
"	mul.f32 	%f305, %f303, %f304;\n"
"	mul.f32 	%f306, %f305, %f305;\n"
"	mul.f32 	%f307, %f306, %f306;\n"
"	mul.f32 	%f308, %f303, %f306;\n"
"	add.f32 	%f309, %f307, %f307;\n"
"	mul.f32 	%f310, %f303, %f309;\n"
"	sub.f32 	%f311, %f310, %f308;\n"
"	div.approx.f32 	%f312, %f311, %f298;\n"
"	mov.f32 	%f313, 0f41c00000;   	\n"
"	mul.f32 	%f314, %f312, %f313;\n"
"	mul.f32 	%f315, %f299, %f314;\n"
"	.loc	15	227	0\n"
"	mul.f32 	%f316, %f296, %f315;\n"
"	mul.f32 	%f317, %f296, %f316;\n"
"	mul.f32 	%f318, %f296, %f317;\n"
"	mov.f32 	%f319, 0f3f000000;   	\n"
"	mul.f32 	%f320, %f318, %f319;\n"
"	mul.f32 	%f321, %f320, %f280;\n"
"	mul.f32 	%f322, %f286, %f297;\n"
"	mul.f32 	%f323, %f290, %f297;\n"
"	mul.f32 	%f324, %f280, %f321;\n"
"	mul.f32 	%f325, %f287, %f322;\n"
"	mad.f32 	%f326, %f288, %f285, %f325;\n"
"	mad.f32 	%f327, %f289, %f323, %f326;\n"
"	mul.f32 	%f328, %f288, %f327;\n"
"	sub.f32 	%f329, %f285, %f328;\n"
"	mul.f32 	%f330, %f324, %f329;\n"
"	mad.f32 	%f331, %f288, %f315, %f330;\n"
"	.loc	15	228	0\n"
"	mul.f32 	%f332, %f287, %f315;\n"
"	mul.f32 	%f333, %f327, %f287;\n"
"	sub.f32 	%f334, %f322, %f333;\n"
"	mad.f32 	%f335, %f324, %f334, %f332;\n"
"	.loc	15	229	0\n"
"	mul.f32 	%f336, %f289, %f315;\n"
"	mul.f32 	%f337, %f327, %f289;\n"
"	sub.f32 	%f338, %f323, %f337;\n"
"	mad.f32 	%f339, %f324, %f338, %f336;\n"
"	.loc	15	239	0\n"
"	mul.f32 	%f340, %f126, %f322;\n"
"	mul.f32 	%f341, %f323, %f324;\n"
"	mul.f32 	%f342, %f322, %f324;\n"
"	mul.f32 	%f343, %f322, %f132;\n"
"	mad.f32 	%f344, %f285, %f125, %f340;\n"
"	mad.f32 	%f345, %f285, %f131, %f343;\n"
"	mad.f32 	%f346, %f323, %f127, %f344;\n"
"	mad.f32 	%f347, %f323, %f133, %f345;\n"
"	mul.f32 	%f348, %f341, %f347;\n"
"	neg.f32 	%f349, %f342;\n"
"	mad.f32 	%f350, %f349, %f346, %f348;\n"
"	mul.f32 	%f351, %f120, %f322;\n"
"	mul.f32 	%f352, %f324, %f285;\n"
"	mad.f32 	%f353, %f113, %f285, %f351;\n"
"	mad.f32 	%f354, %f323, %f114, %f353;\n"
"	mul.f32 	%f355, %f352, %f346;\n"
"	neg.f32 	%f356, %f341;\n"
"	mad.f32 	%f357, %f354, %f356, %f355;\n"
"	mul.f32 	%f358, %f342, %f354;\n"
"	neg.f32 	%f359, %f352;\n"
"	mad.f32 	%f360, %f359, %f347, %f358;\n"
"	.loc	15	246	0\n"
"	ld.global.f32 	%f361, [%rd33+0];\n"
"	mul.lo.u64 	%rd52, %rd44, 4;\n"
"	add.u64 	%rd53, %rd31, %rd52;\n"
"	ld.global.f32 	%f362, [%rd53+0];\n"
"	add.f32 	%f363, %f361, %f361;\n"
"	mul.f32 	%f364, %f362, %f363;\n"
"	.loc	15	247	0\n"
"	mul.f32 	%f365, %f207, %f200;\n"
"	mul.f32 	%f366, %f224, %f200;\n"
"	mul.f32 	%f367, %f217, %f173;\n"
"	mul.f32 	%f368, %f207, %f188;\n"
"	mul.f32 	%f369, %f224, %f188;\n"
"	mul.f32 	%f370, %f212, %f173;\n"
"	mul.f32 	%f371, %f367, %f229;\n"
"	mul.f32 	%f372, %f234, %f370;\n"
"	sub.f32 	%f373, %f372, %f371;\n"
"	mul.f32 	%f374, %f234, %f368;\n"
"	sub.f32 	%f375, %f373, %f374;\n"
"	mad.f32 	%f376, %f229, %f365, %f375;\n"
"	mad.f32 	%f377, %f217, %f369, %f376;\n"
"	mul.f32 	%f378, %f212, %f366;\n"
"	sub.f32 	%f379, %f377, %f378;\n"
"	.loc	15	248	0\n"
"	ld.global.f32 	%f380, [%rd34+4];\n"
"	.loc	19	4494	0\n"
"	div.approx.f32 	%f381, %f364, %f379;\n"
"	lg2.approx.f32 	%f382, %f381;\n"
"	mul.f32 	%f383, %f382, %f380;\n"
"	ex2.approx.f32 	%f384, %f383;\n"
"	.loc	15	253	0\n"
"	mul.f32 	%f385, %f229, %f365;\n"
"	sub.f32 	%f386, %f385, %f371;\n"
"	mul.f32 	%f387, %f212, %f366;\n"
"	sub.f32 	%f388, %f386, %f387;\n"
"	mad.f32 	%f389, %f217, %f369, %f388;\n"
"	mul.f32 	%f390, %f234, %f368;\n"
"	sub.f32 	%f391, %f389, %f390;\n"
"	mad.f32 	%f392, %f234, %f370, %f391;\n"
"	rcp.approx.f32 	%f393, %f392;\n"
"	mul.f32 	%f394, %f56, %f224;\n"
"	mul.f32 	%f395, %f56, %f207;\n"
"	add.f32 	%f396, %f217, %f217;\n"
"	mul.f32 	%f397, %f224, %f217;\n"
"	mul.f32 	%f398, %f234, %f207;\n"
"	mul.f32 	%f399, %f217, %f188;\n"
"	add.f32 	%f400, %f212, %f212;\n"
"	mul.f32 	%f401, %f400, %f234;\n"
"	mul.f32 	%f402, %f56, %f399;\n"
"	mad.f32 	%f403, %f63, %f401, %f402;\n"
"	mul.f32 	%f404, %f56, %f212;\n"
"	mul.f32 	%f405, %f200, %f404;\n"
"	sub.f32 	%f406, %f403, %f405;\n"
"	mul.f32 	%f407, %f63, %f396;\n"
"	mul.f32 	%f408, %f229, %f407;\n"
"	sub.f32 	%f409, %f406, %f408;\n"
"	mul.f32 	%f410, %f51, %f200;\n"
"	mad.f32 	%f411, %f229, %f410, %f409;\n"
"	mul.f32 	%f412, %f51, %f188;\n"
"	mul.f32 	%f413, %f234, %f412;\n"
"	sub.f32 	%f414, %f411, %f413;\n"
"	mul.f32 	%f415, %f51, %f398;\n"
"	sub.f32 	%f416, %f414, %f415;\n"
"	mad.f32 	%f417, %f51, %f397, %f416;\n"
"	mad.f32 	%f418, %f229, %f395, %f417;\n"
"	mul.f32 	%f419, %f212, %f394;\n"
"	sub.f32 	%f420, %f418, %f419;\n"
"	mul.f32 	%f421, %f18, %f420;\n"
"	mul.f32 	%f422, %f393, %f421;\n"
"	mul.f32 	%f423, %f56, %f173;\n"
"	add.f32 	%f424, %f224, %f224;\n"
"	mul.f32 	%f425, %f63, %f200;\n"
"	mul.f32 	%f426, %f63, %f234;\n"
"	mul.f32 	%f427, %f188, %f426;\n"
"	mul.f32 	%f428, %f229, %f425;\n"
"	sub.f32 	%f429, %f428, %f427;\n"
"	add.f32 	%f430, %f173, %f173;\n"
"	mul.f32 	%f431, %f234, %f430;\n"
"	mad.f32 	%f432, %f51, %f431, %f429;\n"
"	mul.f32 	%f433, %f217, %f423;\n"
"	sub.f32 	%f434, %f432, %f433;\n"
"	mul.f32 	%f435, %f424, %f200;\n"
"	mul.f32 	%f436, %f51, %f435;\n"
"	sub.f32 	%f437, %f434, %f436;\n"
"	mad.f32 	%f438, %f200, %f395, %f437;\n"
"	mul.f32 	%f439, %f63, %f398;\n"
"	sub.f32 	%f440, %f438, %f439;\n"
"	mul.f32 	%f441, %f63, %f224;\n"
"	mad.f32 	%f442, %f217, %f441, %f440;\n"
"	mad.f32 	%f443, %f188, %f394, %f442;\n"
"	mul.f32 	%f444, %f229, %f423;\n"
"	sub.f32 	%f445, %f443, %f444;\n"
"	mul.f32 	%f446, %f18, %f445;\n"
"	mul.f32 	%f447, %f393, %f446;\n"
"	add.f32 	%f448, %f207, %f207;\n"
"	mul.f32 	%f449, %f224, %f212;\n"
"	mul.f32 	%f450, %f229, %f207;\n"
"	mul.f32 	%f451, %f400, %f173;\n"
"	mul.f32 	%f452, %f425, %f212;\n"
"	mul.f32 	%f453, %f63, %f399;\n"
"	sub.f32 	%f454, %f453, %f452;\n"
"	mul.f32 	%f455, %f51, %f367;\n"
"	sub.f32 	%f456, %f454, %f455;\n"
"	mad.f32 	%f457, %f51, %f365, %f456;\n"
"	mul.f32 	%f458, %f51, %f173;\n"
"	mul.f32 	%f459, %f229, %f458;\n"
"	sub.f32 	%f460, %f457, %f459;\n"
"	mul.f32 	%f461, %f63, %f449;\n"
"	sub.f32 	%f462, %f460, %f461;\n"
"	mad.f32 	%f463, %f56, %f451, %f462;\n"
"	mul.f32 	%f464, %f56, %f448;\n"
"	mul.f32 	%f465, %f188, %f464;\n"
"	sub.f32 	%f466, %f463, %f465;\n"
"	mad.f32 	%f467, %f63, %f450, %f466;\n"
"	mul.f32 	%f468, %f51, %f224;\n"
"	mad.f32 	%f469, %f188, %f468, %f467;\n"
"	mul.f32 	%f470, %f18, %f469;\n"
"	mul.f32 	%f471, %f393, %f470;\n"
"	mul.f32 	%f472, %f58, %f207;\n"
"	mul.f32 	%f473, %f58, %f212;\n"
"	mul.f32 	%f474, %f200, %f473;\n"
"	mul.f32 	%f475, %f50, %f401;\n"
"	sub.f32 	%f476, %f475, %f474;\n"
"	mad.f32 	%f477, %f58, %f399, %f476;\n"
"	mul.f32 	%f478, %f50, %f396;\n"
"	mul.f32 	%f479, %f229, %f478;\n"
"	sub.f32 	%f480, %f477, %f479;\n"
"	mul.f32 	%f481, %f64, %f200;\n"
"	mad.f32 	%f482, %f229, %f481, %f480;\n"
"	mul.f32 	%f483, %f64, %f188;\n"
"	mul.f32 	%f484, %f234, %f483;\n"
"	sub.f32 	%f485, %f482, %f484;\n"
"	mul.f32 	%f486, %f64, %f398;\n"
"	sub.f32 	%f487, %f485, %f486;\n"
"	mad.f32 	%f488, %f64, %f397, %f487;\n"
"	mul.f32 	%f489, %f58, %f224;\n"
"	mul.f32 	%f490, %f212, %f489;\n"
"	sub.f32 	%f491, %f488, %f490;\n"
"	mad.f32 	%f492, %f229, %f472, %f491;\n"
"	mul.f32 	%f493, %f19, %f492;\n"
"	mul.f32 	%f494, %f393, %f493;\n"
"	mul.f32 	%f495, %f58, %f173;\n"
"	mul.f32 	%f496, %f50, %f200;\n"
"	mul.f32 	%f497, %f234, %f188;\n"
"	mul.f32 	%f498, %f50, %f497;\n"
"	mul.f32 	%f499, %f229, %f496;\n"
"	sub.f32 	%f500, %f499, %f498;\n"
"	add.f32 	%f501, %f234, %f234;\n"
"	mul.f32 	%f502, %f173, %f501;\n"
"	mad.f32 	%f503, %f64, %f502, %f500;\n"
"	mul.f32 	%f504, %f217, %f495;\n"
"	sub.f32 	%f505, %f503, %f504;\n"
"	mul.f32 	%f506, %f64, %f424;\n"
"	mul.f32 	%f507, %f200, %f506;\n"
"	sub.f32 	%f508, %f505, %f507;\n"
"	mul.f32 	%f509, %f50, %f398;\n"
"	sub.f32 	%f510, %f508, %f509;\n"
"	mad.f32 	%f511, %f50, %f397, %f510;\n"
"	mad.f32 	%f512, %f200, %f472, %f511;\n"
"	mul.f32 	%f513, %f229, %f495;\n"
"	sub.f32 	%f514, %f512, %f513;\n"
"	mul.f32 	%f515, %f58, %f188;\n"
"	mad.f32 	%f516, %f224, %f515, %f514;\n"
"	mul.f32 	%f517, %f19, %f516;\n"
"	mul.f32 	%f518, %f393, %f517;\n"
"	mul.f32 	%f519, %f496, %f212;\n"
"	mul.f32 	%f520, %f50, %f399;\n"
"	sub.f32 	%f521, %f520, %f519;\n"
"	mul.f32 	%f522, %f64, %f367;\n"
"	sub.f32 	%f523, %f521, %f522;\n"
"	mad.f32 	%f524, %f64, %f365, %f523;\n"
"	mad.f32 	%f525, %f58, %f451, %f524;\n"
"	mul.f32 	%f526, %f64, %f173;\n"
"	mul.f32 	%f527, %f229, %f526;\n"
"	sub.f32 	%f528, %f525, %f527;\n"
"	mad.f32 	%f529, %f64, %f369, %f528;\n"
"	mul.f32 	%f530, %f50, %f224;\n"
"	mul.f32 	%f531, %f212, %f530;\n"
"	sub.f32 	%f532, %f529, %f531;\n"
"	mul.f32 	%f533, %f448, %f188;\n"
"	mul.f32 	%f534, %f58, %f533;\n"
"	sub.f32 	%f535, %f532, %f534;\n"
"	mul.f32 	%f536, %f50, %f207;\n"
"	mad.f32 	%f537, %f229, %f536, %f535;\n"
"	mul.f32 	%f538, %f19, %f537;\n"
"	mul.f32 	%f539, %f393, %f538;\n"
"	mul.f32 	%f540, %f212, %f200;\n"
"	mul.f32 	%f541, %f65, %f540;\n"
"	mul.f32 	%f542, %f65, %f399;\n"
"	sub.f32 	%f543, %f542, %f541;\n"
"	mul.f32 	%f544, %f55, %f400;\n"
"	mad.f32 	%f545, %f234, %f544, %f543;\n"
"	mul.f32 	%f546, %f57, %f188;\n"
"	mul.f32 	%f547, %f234, %f546;\n"
"	sub.f32 	%f548, %f545, %f547;\n"
"	mul.f32 	%f549, %f229, %f200;\n"
"	mad.f32 	%f550, %f57, %f549, %f548;\n"
"	mul.f32 	%f551, %f67, %f229;\n"
"	mul.f32 	%f552, %f217, %f551;\n"
"	sub.f32 	%f553, %f550, %f552;\n"
"	mul.f32 	%f554, %f57, %f207;\n"
"	mul.f32 	%f555, %f234, %f554;\n"
"	sub.f32 	%f556, %f553, %f555;\n"
"	mad.f32 	%f557, %f57, %f397, %f556;\n"
"	mul.f32 	%f558, %f65, %f449;\n"
"	sub.f32 	%f559, %f557, %f558;\n"
"	mad.f32 	%f560, %f65, %f450, %f559;\n"
"	mul.f32 	%f561, %f20, %f560;\n"
"	mul.f32 	%f562, %f393, %f561;\n"
"	mul.f32 	%f563, %f55, %f200;\n"
"	mul.f32 	%f564, %f563, %f229;\n"
"	mul.f32 	%f565, %f55, %f497;\n"
"	sub.f32 	%f566, %f565, %f564;\n"
"	mul.f32 	%f567, %f72, %f173;\n"
"	mul.f32 	%f568, %f234, %f567;\n"
"	sub.f32 	%f569, %f566, %f568;\n"
"	mul.f32 	%f570, %f65, %f217;\n"
"	mad.f32 	%f571, %f173, %f570, %f569;\n"
"	mul.f32 	%f572, %f72, %f200;\n"
"	mad.f32 	%f573, %f224, %f572, %f571;\n"
"	mul.f32 	%f574, %f55, %f207;\n"
"	mad.f32 	%f575, %f234, %f574, %f573;\n"
"	mul.f32 	%f576, %f65, %f365;\n"
"	sub.f32 	%f577, %f575, %f576;\n"
"	mul.f32 	%f578, %f55, %f397;\n"
"	sub.f32 	%f579, %f577, %f578;\n"
"	mul.f32 	%f580, %f65, %f173;\n"
"	mad.f32 	%f581, %f229, %f580, %f579;\n"
"	mul.f32 	%f582, %f65, %f188;\n"
"	mul.f32 	%f583, %f224, %f582;\n"
"	sub.f32 	%f584, %f583, %f581;\n"
"	mul.f32 	%f585, %f20, %f584;\n"
"	mul.f32 	%f586, %f393, %f585;\n"
"	mul.f32 	%f587, %f563, %f212;\n"
"	mul.f32 	%f588, %f55, %f399;\n"
"	sub.f32 	%f589, %f588, %f587;\n"
"	mul.f32 	%f590, %f57, %f367;\n"
"	sub.f32 	%f591, %f589, %f590;\n"
"	mad.f32 	%f592, %f57, %f365, %f591;\n"
"	mul.f32 	%f593, %f55, %f449;\n"
"	sub.f32 	%f594, %f592, %f593;\n"
"	mul.f32 	%f595, %f57, %f229;\n"
"	mul.f32 	%f596, %f173, %f595;\n"
"	sub.f32 	%f597, %f594, %f596;\n"
"	mul.f32 	%f598, %f65, %f400;\n"
"	mad.f32 	%f599, %f173, %f598, %f597;\n"
"	mad.f32 	%f600, %f55, %f450, %f599;\n"
"	mad.f32 	%f601, %f57, %f369, %f600;\n"
"	mul.f32 	%f602, %f87, %f207;\n"
"	mul.f32 	%f603, %f188, %f602;\n"
"	sub.f32 	%f604, %f601, %f603;\n"
"	mul.f32 	%f605, %f20, %f604;\n"
"	mul.f32 	%f606, %f393, %f605;\n"
"	.loc	15	254	0\n"
"	mul.f32 	%f607, %f384, %f380;\n"
"	neg.f32 	%f608, %f607;\n"
"	.loc	15	287	0\n"
"	add.u64 	%rd54, %rd45, %rd26;\n"
"	ld.global.v4.f32 	{%f609,%f610,%f611,_}, [%rd54+0];\n"
"	.loc	15	297	0\n"
"	mul.f32 	%f612, %f288, %f297;\n"
"	mov.f32 	%f150, %f612;\n"
"	mul.f32 	%f613, %f163, %f610;\n"
"	mul.f32 	%f614, %f165, %f611;\n"
"	mul.f32 	%f615, %f163, %f613;\n"
"	mul.f32 	%f616, %f167, %f609;\n"
"	mad.f32 	%f617, %f167, %f616, %f615;\n"
"	mad.f32 	%f618, %f165, %f614, %f617;\n"
"	add.f32 	%f619, %f115, %f618;\n"
"	mov.f32 	%f174, %f619;\n"
"	mul.f32 	%f620, %f178, %f609;\n"
"	mul.f32 	%f621, %f180, %f611;\n"
"	mul.f32 	%f622, %f182, %f610;\n"
"	mul.f32 	%f623, %f163, %f622;\n"
"	mad.f32 	%f624, %f167, %f620, %f623;\n"
"	mad.f32 	%f625, %f165, %f621, %f624;\n"
"	add.f32 	%f626, %f134, %f625;\n"
"	mov.f32 	%f189, %f626;\n"
"	mul.f32 	%f627, %f191, %f609;\n"
"	mul.f32 	%f628, %f192, %f610;\n"
"	mul.f32 	%f629, %f193, %f611;\n"
"	mul.f32 	%f630, %f163, %f628;\n"
"	mad.f32 	%f631, %f167, %f627, %f630;\n"
"	mad.f32 	%f632, %f165, %f629, %f631;\n"
"	add.f32 	%f633, %f128, %f632;\n"
"	mov.f32 	%f201, %f633;\n"
"	mul.f32 	%f634, %f287, %f297;\n"
"	mov.f32 	%f203, %f634;\n"
"	mul.f32 	%f635, %f613, %f182;\n"
"	mad.f32 	%f636, %f616, %f178, %f635;\n"
"	mad.f32 	%f637, %f614, %f180, %f636;\n"
"	add.f32 	%f638, %f124, %f637;\n"
"	mov.f32 	%f208, %f638;\n"
"	mul.f32 	%f639, %f182, %f622;\n"
"	mad.f32 	%f640, %f178, %f620, %f639;\n"
"	mad.f32 	%f641, %f180, %f621, %f640;\n"
"	add.f32 	%f213, %f135, %f641;\n"
"	mul.f32 	%f642, %f182, %f628;\n"
"	mad.f32 	%f643, %f178, %f627, %f642;\n"
"	mad.f32 	%f644, %f180, %f629, %f643;\n"
"	add.f32 	%f218, %f129, %f644;\n"
"	mul.f32 	%f645, %f289, %f297;\n"
"	mov.f32 	%f220, %f645;\n"
"	mul.f32 	%f646, %f192, %f613;\n"
"	mad.f32 	%f647, %f616, %f191, %f646;\n"
"	mad.f32 	%f648, %f614, %f193, %f647;\n"
"	add.f32 	%f649, %f116, %f648;\n"
"	mov.f32 	%f225, %f649;\n"
"	mul.f32 	%f650, %f192, %f622;\n"
"	mad.f32 	%f651, %f620, %f191, %f650;\n"
"	mad.f32 	%f652, %f621, %f193, %f651;\n"
"	add.f32 	%f230, %f136, %f652;\n"
"	mul.f32 	%f653, %f192, %f628;\n"
"	mad.f32 	%f654, %f191, %f627, %f653;\n"
"	mad.f32 	%f655, %f193, %f629, %f654;\n"
"	add.f32 	%f235, %f130, %f655;\n"
"	.loc	14	186	0\n"
"	abs.f32 	%f656, %f638;\n"
"	abs.f32 	%f657, %f619;\n"
"	setp.gt.f32 	%p15, %f656, %f657;\n"
"	@!%p15 bra 	$Lt_0_36866;\n"
"	mov.f32 	%f174, %f638;\n"
"	mov.f32 	%f208, %f619;\n"
"	mov.f32 	%f189, %f213;\n"
"	mov.f32 	%f213, %f626;\n"
"	mov.f32 	%f201, %f218;\n"
"	mov.f32 	%f218, %f633;\n"
"	mov.f32 	%f150, %f634;\n"
"	mov.f32 	%f203, %f612;\n"
"$Lt_0_36866:\n"
"	.loc	14	193	0\n"
"	mov.f32 	%f658, %f174;\n"
"	abs.f32 	%f659, %f658;\n"
"	abs.f32 	%f660, %f649;\n"
"	setp.lt.f32 	%p16, %f659, %f660;\n"
"	@!%p16 bra 	$Lt_0_37378;\n"
"	mov.f32 	%f174, %f649;\n"
"	mov.f32 	%f225, %f658;\n"
"	.loc	14	196	0\n"
"	mov.f32 	%f241, %f189;\n"
"	mov.f32 	%f189, %f230;\n"
"	mov.f32 	%f230, %f241;\n"
"	.loc	14	197	0\n"
"	mov.f32 	%f241, %f201;\n"
"	mov.f32 	%f201, %f235;\n"
"	mov.f32 	%f235, %f241;\n"
"	.loc	14	198	0\n"
"	mov.f32 	%f241, %f150;\n"
"	mov.f32 	%f150, %f645;\n"
"	mov.f32 	%f220, %f241;\n"
"$Lt_0_37378:\n"
"	mov.f32 	%f661, %f174;\n"
"	mov.f32 	%f662, 0f00000000;   	\n"
"	setp.neu.f32 	%p17, %f661, %f662;\n"
"	@!%p17 bra 	$Lt_0_38146;\n"
"	bra.uni 	$Lt_0_38914;\n"
"$Lt_0_38146:\n"
"	mov.f32 	%f663, 0f00000000;   	\n"
"	setp.neu.f32 	%p18, %f208, %f663;\n"
"	@!%p18 bra 	$Lt_0_38658;\n"
"	mov.f32 	%f174, %f208;\n"
"	mov.f32 	%f208, %f661;\n"
"	.loc	14	213	0\n"
"	mov.f32 	%f664, %f189;\n"
"	mov.f32 	%f189, %f213;\n"
"	mov.f32 	%f213, %f664;\n"
"	.loc	14	214	0\n"
"	mov.f32 	%f665, %f201;\n"
"	mov.f32 	%f201, %f218;\n"
"	mov.f32 	%f218, %f665;\n"
"	.loc	14	215	0\n"
"	mov.f32 	%f666, %f150;\n"
"	mov.f32 	%f150, %f203;\n"
"	mov.f32 	%f203, %f666;\n"
"	bra.uni 	$Lt_0_38914;\n"
"$Lt_0_38658:\n"
"	mov.f32 	%f667, 0f00000000;   	\n"
"	setp.neu.f32 	%p19, %f225, %f667;\n"
"	@!%p19 bra 	$Lt_0_39170;\n"
"	mov.f32 	%f174, %f225;\n"
"	mov.f32 	%f225, %f661;\n"
"	.loc	14	221	0\n"
"	mov.f32 	%f668, %f189;\n"
"	mov.f32 	%f189, %f230;\n"
"	mov.f32 	%f230, %f668;\n"
"	.loc	14	222	0\n"
"	mov.f32 	%f669, %f201;\n"
"	mov.f32 	%f201, %f235;\n"
"	mov.f32 	%f235, %f669;\n"
"	.loc	14	223	0\n"
"	mov.f32 	%f670, %f150;\n"
"	mov.f32 	%f150, %f220;\n"
"	mov.f32 	%f220, %f670;\n"
"	bra.uni 	$Lt_0_38914;\n"
"$Lt_0_39170:\n"
"	.loc	14	226	0\n"
"	mov.s32 	%r26, 2;\n"
"	ld.param.u64 	%rd55, [__cudaparm_kernel_gayberne_err_flag];\n"
"	st.global.s32 	[%rd55+0], %r26;\n"
"$Lt_0_38914:\n"
"$Lt_0_38402:\n"
"$Lt_0_37890:\n"
"	.loc	14	228	0\n"
"	div.approx.f32 	%f671, %f208, %f174;\n"
"	mul.f32 	%f672, %f189, %f671;\n"
"	sub.f32 	%f673, %f213, %f672;\n"
"	mov.f32 	%f213, %f673;\n"
"	mul.f32 	%f674, %f201, %f671;\n"
"	sub.f32 	%f675, %f218, %f674;\n"
"	mov.f32 	%f218, %f675;\n"
"	mul.f32 	%f676, %f150, %f671;\n"
"	sub.f32 	%f677, %f203, %f676;\n"
"	mov.f32 	%f203, %f677;\n"
"	.loc	14	232	0\n"
"	div.approx.f32 	%f678, %f225, %f174;\n"
"	mul.f32 	%f679, %f189, %f678;\n"
"	sub.f32 	%f230, %f230, %f679;\n"
"	mul.f32 	%f680, %f201, %f678;\n"
"	sub.f32 	%f235, %f235, %f680;\n"
"	mul.f32 	%f681, %f150, %f678;\n"
"	sub.f32 	%f220, %f220, %f681;\n"
"	.loc	14	237	0\n"
"	abs.f32 	%f682, %f673;\n"
"	abs.f32 	%f683, %f230;\n"
"	setp.lt.f32 	%p20, %f682, %f683;\n"
"	@!%p20 bra 	$Lt_0_39426;\n"
"	mov.f32 	%f213, %f230;\n"
"	mov.f32 	%f230, %f673;\n"
"	mov.f32 	%f218, %f235;\n"
"	mov.f32 	%f235, %f675;\n"
"	mov.f32 	%f203, %f220;\n"
"	mov.f32 	%f220, %f677;\n"
"$Lt_0_39426:\n"
"	mov.f32 	%f684, %f213;\n"
"	mov.f32 	%f685, 0f00000000;   	\n"
"	setp.neu.f32 	%p21, %f684, %f685;\n"
"	@!%p21 bra 	$Lt_0_40194;\n"
"	bra.uni 	$Lt_0_40450;\n"
"$Lt_0_40194:\n"
"	mov.f32 	%f686, 0f00000000;   	\n"
"	setp.neu.f32 	%p22, %f230, %f686;\n"
"	@!%p22 bra 	$Lt_0_40450;\n"
"	mov.f32 	%f213, %f230;\n"
"	mov.f32 	%f230, %f684;\n"
"	.loc	14	258	0\n"
"	mov.f32 	%f268, %f218;\n"
"	mov.f32 	%f218, %f235;\n"
"	mov.f32 	%f235, %f268;\n"
"	.loc	14	259	0\n"
"	mov.f32 	%f268, %f203;\n"
"	mov.f32 	%f203, %f220;\n"
"	mov.f32 	%f220, %f268;\n"
"$Lt_0_40450:\n"
"$Lt_0_39938:\n"
"	.loc	14	263	0\n"
"	div.approx.f32 	%f687, %f230, %f213;\n"
"	mul.f32 	%f688, %f218, %f687;\n"
"	sub.f32 	%f235, %f235, %f688;\n"
"	mul.f32 	%f689, %f203, %f687;\n"
"	sub.f32 	%f220, %f220, %f689;\n"
"	mov.f32 	%f690, 0f00000000;   	\n"
"	setp.eq.f32 	%p23, %f235, %f690;\n"
"	@!%p23 bra 	$Lt_0_40962;\n"
"	.loc	14	268	0\n"
"	mov.s32 	%r27, 2;\n"
"	ld.param.u64 	%rd56, [__cudaparm_kernel_gayberne_err_flag];\n"
"	st.global.s32 	[%rd56+0], %r27;\n"
"$Lt_0_40962:\n"
"	.loc	15	299	0\n"
"	div.approx.f32 	%f691, %f220, %f235;\n"
"	mul.f32 	%f692, %f691, %f218;\n"
"	sub.f32 	%f693, %f203, %f692;\n"
"	div.approx.f32 	%f694, %f693, %f213;\n"
"	mul.f32 	%f695, %f694, %f189;\n"
"	mad.f32 	%f696, %f201, %f691, %f695;\n"
"	sub.f32 	%f697, %f150, %f696;\n"
"	div.approx.f32 	%f698, %f697, %f174;\n"
"	mul.f32 	%f699, %f280, %f698;\n"
"	.loc	15	306	0\n"
"	mul.f32 	%f700, %f694, %f280;\n"
"	mul.f32 	%f701, %f691, %f280;\n"
"	mul.f32 	%f702, %f280, %f612;\n"
"	mul.f32 	%f703, %f280, %f634;\n"
"	mul.f32 	%f704, %f280, %f645;\n"
"	mul.f32 	%f705, %f700, %f703;\n"
"	mad.f32 	%f706, %f702, %f699, %f705;\n"
"	mad.f32 	%f707, %f704, %f701, %f706;\n"
"	add.f32 	%f708, %f707, %f707;\n"
"	ld.global.f32 	%f709, [%rd34+8];\n"
"	.loc	15	309	0\n"
"	mul.f32 	%f699, %f297, %f699;\n"
"	.loc	15	314	0\n"
"	mov.f32 	%f710, 0fbf800000;   	\n"
"	add.f32 	%f711, %f709, %f710;\n"
"	lg2.approx.f32 	%f712, %f708;\n"
"	mul.f32 	%f713, %f712, %f709;\n"
"	ex2.approx.f32 	%f714, %f713;\n"
"	mov.f32 	%f715, 0fc0800000;   	\n"
"	mul.f32 	%f716, %f280, %f715;\n"
"	mul.f32 	%f717, %f280, %f716;\n"
"	lg2.approx.f32 	%f718, %f714;\n"
"	div.approx.f32 	%f719, %f711, %f709;\n"
"	mul.f32 	%f720, %f718, %f719;\n"
"	ex2.approx.f32 	%f721, %f720;\n"
"	mul.f32 	%f722, %f717, %f709;\n"
"	mul.f32 	%f723, %f721, %f722;\n"
"	.loc	15	316	0\n"
"	mul.f32 	%f724, %f700, %f297;\n"
"	mul.f32 	%f725, %f701, %f297;\n"
"	mul.f32 	%f726, %f724, %f703;\n"
"	mad.f32 	%f727, %f702, %f699, %f726;\n"
"	mad.f32 	%f728, %f704, %f725, %f727;\n"
"	mul.f32 	%f729, %f702, %f728;\n"
"	sub.f32 	%f730, %f699, %f729;\n"
"	mul.f32 	%f731, %f723, %f730;\n"
"	.loc	15	317	0\n"
"	mul.f32 	%f732, %f728, %f703;\n"
"	sub.f32 	%f733, %f724, %f732;\n"
"	mul.f32 	%f734, %f723, %f733;\n"
"	.loc	15	318	0\n"
"	mul.f32 	%f735, %f728, %f704;\n"
"	sub.f32 	%f736, %f725, %f735;\n"
"	mul.f32 	%f737, %f723, %f736;\n"
"	.loc	15	323	0\n"
"	mul.f32 	%f738, %f129, %f724;\n"
"	mul.f32 	%f739, %f724, %f135;\n"
"	mad.f32 	%f740, %f699, %f128, %f738;\n"
"	mad.f32 	%f741, %f699, %f134, %f739;\n"
"	mad.f32 	%f742, %f725, %f130, %f740;\n"
"	mad.f32 	%f743, %f725, %f136, %f741;\n"
"	mul.f32 	%f744, %f724, %f742;\n"
"	mul.f32 	%f745, %f743, %f725;\n"
"	sub.f32 	%f746, %f745, %f744;\n"
"	mul.f32 	%f747, %f124, %f724;\n"
"	mad.f32 	%f748, %f115, %f699, %f747;\n"
"	mad.f32 	%f749, %f725, %f116, %f748;\n"
"	mul.f32 	%f750, %f725, %f749;\n"
"	mul.f32 	%f751, %f699, %f742;\n"
"	sub.f32 	%f752, %f751, %f750;\n"
"	mul.f32 	%f753, %f743, %f699;\n"
"	mul.f32 	%f754, %f749, %f724;\n"
"	sub.f32 	%f755, %f754, %f753;\n"
"	.loc	15	325	0\n"
"	mul.f32 	%f746, %f717, %f746;\n"
"	.loc	15	326	0\n"
"	mul.f32 	%f752, %f717, %f752;\n"
"	.loc	15	327	0\n"
"	mul.f32 	%f755, %f717, %f755;\n"
"	.loc	14	270	0\n"
"	mov.f32 	%f756, 0f40800000;   	\n"
"	mul.f32 	%f757, %f299, %f756;\n"
"	.loc	15	149	0\n"
"	ld.shared.f32 	%f137, [%rd39+0];\n"
"	.loc	14	270	0\n"
"	mul.f32 	%f758, %f384, %f137;\n"
"	mul.f32 	%f759, %f758, %f714;\n"
"	sub.f32 	%f760, %f307, %f306;\n"
"	mul.f32 	%f761, %f757, %f760;\n"
"	mad.f32 	%f762, %f761, %f759, %f34;\n"
"	selp.f32 	%f34, %f762, %f34, %p4;\n"
"	mul.f32 	%f763, %f759, %f331;\n"
"	mul.f32 	%f764, %f759, %f335;\n"
"	mul.f32 	%f765, %f759, %f339;\n"
"	mul.f32 	%f766, %f761, %f384;\n"
"	mul.f32 	%f767, %f766, %f137;\n"
"	neg.f32 	%f768, %f767;\n"
"	mul.f32 	%f769, %f731, %f768;\n"
"	sub.f32 	%f770, %f769, %f763;\n"
"	mul.f32 	%f771, %f768, %f734;\n"
"	sub.f32 	%f772, %f771, %f764;\n"
"	mul.f32 	%f773, %f768, %f737;\n"
"	sub.f32 	%f774, %f773, %f765;\n"
"	@!%p5 bra 	$Lt_0_41730;\n"
"	.loc	15	339	0\n"
"	add.f32 	%f33, %f770, %f33;\n"
"	.loc	15	340	0\n"
"	mul.f32 	%f775, %f297, %f702;\n"
"	neg.f32 	%f776, %f775;\n"
"	mov.f32 	%f777, %f3;\n"
"	mad.f32 	%f778, %f776, %f770, %f777;\n"
"	mov.f32 	%f3, %f778;\n"
"	.loc	15	342	0\n"
"	add.f32 	%f32, %f772, %f32;\n"
"	.loc	15	343	0\n"
"	mul.f32 	%f779, %f297, %f703;\n"
"	neg.f32 	%f780, %f779;\n"
"	mov.f32 	%f781, %f5;\n"
"	mad.f32 	%f782, %f772, %f780, %f781;\n"
"	mov.f32 	%f5, %f782;\n"
"	.loc	15	344	0\n"
"	mov.f32 	%f783, %f9;\n"
"	mad.f32 	%f784, %f776, %f772, %f783;\n"
"	mov.f32 	%f9, %f784;\n"
"	.loc	15	346	0\n"
"	add.f32 	%f31, %f774, %f31;\n"
"	.loc	15	347	0\n"
"	mov.f32 	%f785, %f7;\n"
"	mul.f32 	%f786, %f297, %f704;\n"
"	neg.f32 	%f787, %f786;\n"
"	mad.f32 	%f788, %f774, %f787, %f785;\n"
"	mov.f32 	%f7, %f788;\n"
"	.loc	15	348	0\n"
"	mov.f32 	%f789, %f11;\n"
"	mad.f32 	%f790, %f776, %f774, %f789;\n"
"	mov.f32 	%f11, %f790;\n"
"	.loc	15	349	0\n"
"	mad.f32 	%f12, %f774, %f780, %f12;\n"
"	mov.f32 	%f791, %f12;\n"
"	bra.uni 	$Lt_0_41474;\n"
"$Lt_0_41730:\n"
"	.loc	15	351	0\n"
"	add.f32 	%f33, %f770, %f33;\n"
"	.loc	15	352	0\n"
"	add.f32 	%f32, %f772, %f32;\n"
"	.loc	15	353	0\n"
"	add.f32 	%f31, %f774, %f31;\n"
"$Lt_0_41474:\n"
"	.loc	15	360	0\n"
"	mul.f32 	%f792, %f608, %f518;\n"
"	mul.f32 	%f793, %f608, %f539;\n"
"	mul.f32 	%f794, %f608, %f447;\n"
"	mul.f32 	%f795, %f608, %f471;\n"
"	mul.f32 	%f796, %f608, %f586;\n"
"	mul.f32 	%f797, %f608, %f606;\n"
"	mul.f32 	%f798, %f714, %f384;\n"
"	.loc	15	149	0\n"
"	ld.shared.f32 	%f137, [%rd39+0];\n"
"	.loc	15	360	0\n"
"	mul.f32 	%f799, %f798, %f137;\n"
"	neg.f32 	%f800, %f799;\n"
"	mul.f32 	%f801, %f714, %f761;\n"
"	mul.f32 	%f802, %f801, %f137;\n"
"	neg.f32 	%f803, %f802;\n"
"	mul.f32 	%f804, %f767, %f746;\n"
"	mul.f32 	%f805, %f792, %f58;\n"
"	mul.f32 	%f806, %f793, %f64;\n"
"	sub.f32 	%f807, %f806, %f805;\n"
"	mul.f32 	%f808, %f794, %f56;\n"
"	mul.f32 	%f809, %f51, %f795;\n"
"	sub.f32 	%f810, %f809, %f808;\n"
"	add.f32 	%f811, %f807, %f810;\n"
"	mul.f32 	%f812, %f796, %f65;\n"
"	mul.f32 	%f813, %f797, %f57;\n"
"	sub.f32 	%f814, %f813, %f812;\n"
"	add.f32 	%f815, %f811, %f814;\n"
"	mul.f32 	%f816, %f815, %f803;\n"
"	sub.f32 	%f817, %f816, %f804;\n"
"	mad.f32 	%f818, %f350, %f800, %f817;\n"
"	add.f32 	%f30, %f30, %f818;\n"
"	.loc	15	361	0\n"
"	mul.f32 	%f819, %f422, %f608;\n"
"	mul.f32 	%f820, %f608, %f494;\n"
"	mul.f32 	%f821, %f608, %f562;\n"
"	mul.f32 	%f822, %f767, %f752;\n"
"	mul.f32 	%f823, %f793, %f50;\n"
"	mul.f32 	%f824, %f820, %f58;\n"
"	sub.f32 	%f825, %f824, %f823;\n"
"	mul.f32 	%f826, %f795, %f63;\n"
"	mul.f32 	%f827, %f819, %f56;\n"
"	sub.f32 	%f828, %f827, %f826;\n"
"	add.f32 	%f829, %f825, %f828;\n"
"	mul.f32 	%f830, %f797, %f55;\n"
"	mul.f32 	%f831, %f821, %f65;\n"
"	sub.f32 	%f832, %f831, %f830;\n"
"	add.f32 	%f833, %f829, %f832;\n"
"	mul.f32 	%f834, %f803, %f833;\n"
"	sub.f32 	%f835, %f834, %f822;\n"
"	mad.f32 	%f836, %f800, %f357, %f835;\n"
"	add.f32 	%f29, %f29, %f836;\n"
"	.loc	15	362	0\n"
"	mul.f32 	%f837, %f767, %f755;\n"
"	mul.f32 	%f838, %f819, %f51;\n"
"	mul.f32 	%f839, %f63, %f794;\n"
"	sub.f32 	%f840, %f839, %f838;\n"
"	mul.f32 	%f841, %f820, %f64;\n"
"	mul.f32 	%f842, %f792, %f50;\n"
"	sub.f32 	%f843, %f842, %f841;\n"
"	add.f32 	%f844, %f840, %f843;\n"
"	mul.f32 	%f845, %f821, %f57;\n"
"	mul.f32 	%f846, %f796, %f55;\n"
"	sub.f32 	%f847, %f846, %f845;\n"
"	add.f32 	%f848, %f844, %f847;\n"
"	mul.f32 	%f849, %f803, %f848;\n"
"	sub.f32 	%f850, %f849, %f837;\n"
"	mad.f32 	%f851, %f800, %f360, %f850;\n"
"	add.f32 	%f28, %f28, %f851;\n"
"	.loc	15	146	0\n"
"	add.u64 	%rd15, %rd12, %rd15;\n"
"	setp.gt.u64 	%p24, %rd30, %rd15;\n"
"	@%p24 bra 	$Lt_0_32002;\n"
"	bra.uni 	$Lt_0_31490;\n"
"$Lt_0_43266:\n"
"	mov.s32 	%r28, 0;\n"
"	setp.gt.s32 	%p4, %r14, %r28;\n"
"	mov.s32 	%r29, 0;\n"
"	setp.gt.s32 	%p5, %r13, %r29;\n"
"$Lt_0_31490:\n"
"	.loc	15	367	0\n"
"	ld.param.u64 	%rd57, [__cudaparm_kernel_gayberne_engv];\n"
"	add.u64 	%rd58, %rd57, %rd8;\n"
"	ld.param.s32 	%r30, [__cudaparm_kernel_gayberne_astride];\n"
"	@!%p4 bra 	$Lt_0_42242;\n"
"	.loc	15	369	0\n"
"	st.global.f32 	[%rd58+0], %f34;\n"
"	.loc	15	370	0\n"
"	cvt.u64.s32 	%rd59, %r30;\n"
"	mul.lo.u64 	%rd60, %rd59, 4;\n"
"	add.u64 	%rd58, %rd58, %rd60;\n"
"$Lt_0_42242:\n"
"	@!%p5 bra 	$Lt_0_42754;\n"
"	.loc	15	374	0\n"
"	mov.f32 	%f852, %f3;\n"
"	st.global.f32 	[%rd58+0], %f852;\n"
"	.loc	15	375	0\n"
"	cvt.u64.s32 	%rd61, %r30;\n"
"	mul.lo.u64 	%rd62, %rd61, 4;\n"
"	add.u64 	%rd58, %rd62, %rd58;\n"
"	.loc	15	374	0\n"
"	mov.f32 	%f853, %f5;\n"
"	st.global.f32 	[%rd58+0], %f853;\n"
"	.loc	15	375	0\n"
"	add.u64 	%rd58, %rd62, %rd58;\n"
"	.loc	15	374	0\n"
"	mov.f32 	%f854, %f7;\n"
"	st.global.f32 	[%rd58+0], %f854;\n"
"	.loc	15	375	0\n"
"	add.u64 	%rd58, %rd62, %rd58;\n"
"	.loc	15	374	0\n"
"	mov.f32 	%f855, %f9;\n"
"	st.global.f32 	[%rd58+0], %f855;\n"
"	.loc	15	375	0\n"
"	add.u64 	%rd58, %rd62, %rd58;\n"
"	.loc	15	374	0\n"
"	mov.f32 	%f856, %f11;\n"
"	st.global.f32 	[%rd58+0], %f856;\n"
"	add.u64 	%rd63, %rd62, %rd58;\n"
"	st.global.f32 	[%rd63+0], %f12;\n"
"$Lt_0_42754:\n"
"	.loc	15	378	0\n"
"	ld.param.u64 	%rd64, [__cudaparm_kernel_gayberne_ans];\n"
"	mul.lo.u64 	%rd65, %rd7, 16;\n"
"	add.u64 	%rd66, %rd64, %rd65;\n"
"	mov.f32 	%f857, %f858;\n"
"	st.global.v4.f32 	[%rd66+0], {%f33,%f32,%f31,%f857};\n"
"	.loc	15	379	0\n"
"	add.s32 	%r31, %r6, %r30;\n"
"	cvt.s64.s32 	%rd67, %r31;\n"
"	mul.lo.u64 	%rd68, %rd67, 16;\n"
"	add.u64 	%rd69, %rd64, %rd68;\n"
"	mov.f32 	%f859, %f860;\n"
"	st.global.v4.f32 	[%rd69+0], {%f30,%f29,%f28,%f859};\n"
"$Lt_0_30978:\n"
"	.loc	15	381	0\n"
"	exit;\n"
"$LDWend_kernel_gayberne:\n"
"	}\n"
;
const char * gb_gpu_kernel_lj = 
"	.version 1.4\n"
"	.target sm_13\n"
"	.entry kernel_sphere_gb (\n"
"		.param .u64 __cudaparm_kernel_sphere_gb_x_,\n"
"		.param .u64 __cudaparm_kernel_sphere_gb_q,\n"
"		.param .u64 __cudaparm_kernel_sphere_gb_shape,\n"
"		.param .u64 __cudaparm_kernel_sphere_gb_well,\n"
"		.param .u64 __cudaparm_kernel_sphere_gb_gum,\n"
"		.param .u64 __cudaparm_kernel_sphere_gb_sig_eps,\n"
"		.param .s32 __cudaparm_kernel_sphere_gb_ntypes,\n"
"		.param .u64 __cudaparm_kernel_sphere_gb_lshape,\n"
"		.param .u64 __cudaparm_kernel_sphere_gb_dev_nbor,\n"
"		.param .s32 __cudaparm_kernel_sphere_gb_stride,\n"
"		.param .u64 __cudaparm_kernel_sphere_gb_ans,\n"
"		.param .u64 __cudaparm_kernel_sphere_gb_engv,\n"
"		.param .u64 __cudaparm_kernel_sphere_gb_err_flag,\n"
"		.param .s32 __cudaparm_kernel_sphere_gb_eflag,\n"
"		.param .s32 __cudaparm_kernel_sphere_gb_vflag,\n"
"		.param .s32 __cudaparm_kernel_sphere_gb_start,\n"
"		.param .s32 __cudaparm_kernel_sphere_gb_inum,\n"
"		.param .s32 __cudaparm_kernel_sphere_gb_nall)\n"
"	{\n"
"	.reg .u32 %r<33>;\n"
"	.reg .u64 %rd<67>;\n"
"	.reg .f32 %f<401>;\n"
"	.reg .pred %p<26>;\n"
"	.shared .align 4 .b8 __cuda_sp_lj124[16];\n"
"	.loc	15	37	0\n"
"$LBB1_kernel_sphere_gb:\n"
"	cvt.s32.u16 	%r1, %tid.x;\n"
"	mov.u32 	%r2, 3;\n"
"	setp.gt.s32 	%p1, %r1, %r2;\n"
"	@%p1 bra 	$Lt_0_30466;\n"
"	.loc	15	43	0\n"
"	mov.u64 	%rd1, __cuda_sp_lj124;\n"
"	cvt.u64.s32 	%rd2, %r1;\n"
"	mul.lo.u64 	%rd3, %rd2, 4;\n"
"	ld.param.u64 	%rd4, [__cudaparm_kernel_sphere_gb_gum];\n"
"	add.u64 	%rd5, %rd4, %rd3;\n"
"	ld.global.f32 	%f1, [%rd5+12];\n"
"	add.u64 	%rd6, %rd3, %rd1;\n"
"	st.shared.f32 	[%rd6+0], %f1;\n"
"$Lt_0_30466:\n"
"	mov.u64 	%rd1, __cuda_sp_lj124;\n"
"	.loc	15	45	0\n"
"	bar.sync 	0;\n"
"	cvt.s32.u16 	%r3, %ctaid.x;\n"
"	cvt.s32.u16 	%r4, %ntid.x;\n"
"	mul24.lo.s32 	%r5, %r3, %r4;\n"
"	ld.param.s32 	%r6, [__cudaparm_kernel_sphere_gb_start];\n"
"	add.s32 	%r7, %r6, %r5;\n"
"	add.s32 	%r8, %r7, %r1;\n"
"	ld.param.s32 	%r9, [__cudaparm_kernel_sphere_gb_inum];\n"
"	setp.ge.s32 	%p2, %r8, %r9;\n"
"	@%p2 bra 	$Lt_0_30978;\n"
"	.loc	15	56	0\n"
"	mov.f32 	%f2, 0f00000000;     	\n"
"	mov.f32 	%f3, %f2;\n"
"	mov.f32 	%f4, 0f00000000;     	\n"
"	mov.f32 	%f5, %f4;\n"
"	mov.f32 	%f6, 0f00000000;     	\n"
"	mov.f32 	%f7, %f6;\n"
"	mov.f32 	%f8, 0f00000000;     	\n"
"	mov.f32 	%f9, %f8;\n"
"	mov.f32 	%f10, 0f00000000;    	\n"
"	mov.f32 	%f11, %f10;\n"
"	mov.f32 	%f12, 0f00000000;    	\n"
"	mov.f32 	%f13, %f12;\n"
"	.loc	15	59	0\n"
"	cvt.u64.s32 	%rd7, %r8;\n"
"	mul.lo.u64 	%rd8, %rd7, 4;\n"
"	ld.param.u64 	%rd9, [__cudaparm_kernel_sphere_gb_dev_nbor];\n"
"	add.u64 	%rd10, %rd9, %rd8;\n"
"	ld.global.s32 	%r10, [%rd10+0];\n"
"	.loc	15	61	0\n"
"	ld.param.s32 	%r11, [__cudaparm_kernel_sphere_gb_stride];\n"
"	cvt.u64.s32 	%rd11, %r11;\n"
"	mul.lo.u64 	%rd12, %rd11, 4;\n"
"	add.u64 	%rd13, %rd10, %rd12;\n"
"	ld.global.s32 	%r12, [%rd13+0];\n"
"	.loc	15	62	0\n"
"	add.u64 	%rd14, %rd13, %rd12;\n"
"	mov.s64 	%rd15, %rd14;\n"
"	.loc	15	65	0\n"
"	ld.param.u64 	%rd16, [__cudaparm_kernel_sphere_gb_x_];\n"
"	cvt.u64.s32 	%rd17, %r10;\n"
"	mul.lo.u64 	%rd18, %rd17, 16;\n"
"	add.u64 	%rd19, %rd16, %rd18;\n"
"	ld.global.v4.f32 	{%f14,%f15,%f16,%f17}, [%rd19+0];\n"
"	.loc	15	68	0\n"
"	cvt.rzi.s32.f32 	%r13, %f17;\n"
"	cvt.u64.s32 	%rd20, %r13;\n"
"	mul.lo.u64 	%rd21, %rd20, 16;\n"
"	ld.param.u64 	%rd22, [__cudaparm_kernel_sphere_gb_shape];\n"
"	add.u64 	%rd23, %rd21, %rd22;\n"
"	ld.global.f32 	%f18, [%rd23+0];\n"
"	.loc	15	69	0\n"
"	ld.param.u64 	%rd24, [__cudaparm_kernel_sphere_gb_well];\n"
"	add.u64 	%rd25, %rd21, %rd24;\n"
"	ld.global.f32 	%f19, [%rd25+0];\n"
"	mul.lo.s32 	%r14, %r12, %r11;\n"
"	cvt.s64.s32 	%rd26, %r14;\n"
"	mul.lo.u64 	%rd27, %rd26, 4;\n"
"	add.u64 	%rd28, %rd14, %rd27;\n"
"	ld.param.s32 	%r15, [__cudaparm_kernel_sphere_gb_vflag];\n"
"	ld.param.s32 	%r16, [__cudaparm_kernel_sphere_gb_eflag];\n"
"	setp.ge.u64 	%p3, %rd14, %rd28;\n"
"	mov.f32 	%f20, 0f00000000;    	\n"
"	mov.f32 	%f21, 0f00000000;    	\n"
"	mov.f32 	%f22, 0f00000000;    	\n"
"	mov.f32 	%f23, 0f00000000;    	\n"
"	@%p3 bra 	$Lt_0_43266;\n"
"	mov.s32 	%r17, 0;\n"
"	setp.gt.s32 	%p4, %r16, %r17;\n"
"	mov.s32 	%r18, 0;\n"
"	setp.gt.s32 	%p5, %r15, %r18;\n"
"	ld.param.s32 	%r19, [__cudaparm_kernel_sphere_gb_ntypes];\n"
"	mul24.lo.s32 	%r20, %r19, %r13;\n"
"	ld.param.u64 	%rd29, [__cudaparm_kernel_sphere_gb_lshape];\n"
"	mul.lo.u64 	%rd30, %rd20, 4;\n"
"	add.u64 	%rd31, %rd29, %rd30;\n"
"	ld.param.u64 	%rd32, [__cudaparm_kernel_sphere_gb_gum];\n"
"	ld.param.u64 	%rd33, [__cudaparm_kernel_sphere_gb_sig_eps];\n"
"	ld.param.u64 	%rd34, [__cudaparm_kernel_sphere_gb_q];\n"
"$Lt_0_32002:\n"
"	.loc	15	74	0\n"
"	ld.global.s32 	%r21, [%rd15+0];\n"
"	.loc	15	75	0\n"
"	shr.s32 	%r22, %r21, 30;\n"
"	cvt.s64.s32 	%rd35, %r22;\n"
"	and.b64 	%rd36, %rd35, 3;\n"
"	mul.lo.u64 	%rd37, %rd36, 4;\n"
"	add.u64 	%rd38, %rd1, %rd37;\n"
"	ld.shared.f32 	%f24, [%rd38+0];\n"
"	.loc	15	78	0\n"
"	and.b32 	%r23, %r21, 1073741823;\n"
"	cvt.u64.s32 	%rd39, %r23;\n"
"	mul.lo.u64 	%rd40, %rd39, 16;\n"
"	add.u64 	%rd41, %rd40, %rd16;\n"
"	ld.global.v4.f32 	{%f25,%f26,%f27,%f28}, [%rd41+0];\n"
"	.loc	15	97	0\n"
"	add.u64 	%rd42, %rd40, %rd34;\n"
"	ld.global.v4.f32 	{%f29,%f30,%f31,%f32}, [%rd42+0];\n"
"	.loc	15	106	0\n"
"	cvt.rzi.s32.f32 	%r24, %f28;\n"
"	cvt.u64.s32 	%rd43, %r24;\n"
"	mul.lo.u64 	%rd44, %rd43, 16;\n"
"	add.u64 	%rd45, %rd44, %rd22;\n"
"	ld.global.v4.f32 	{%f33,%f34,%f35,_}, [%rd45+0];\n"
"	.loc	15	123	0\n"
"	sub.f32 	%f36, %f25, %f14;\n"
"	mov.f32 	%f37, %f36;\n"
"	add.f32 	%f38, %f30, %f30;\n"
"	add.f32 	%f39, %f32, %f32;\n"
"	mul.f32 	%f40, %f29, %f29;\n"
"	mul.f32 	%f41, %f30, %f30;\n"
"	mul.f32 	%f42, %f31, %f31;\n"
"	mul.f32 	%f43, %f32, %f32;\n"
"	add.f32 	%f44, %f31, %f31;\n"
"	mul.f32 	%f45, %f38, %f31;\n"
"	mul.f32 	%f46, %f38, %f32;\n"
"	mul.f32 	%f47, %f39, %f29;\n"
"	add.f32 	%f48, %f40, %f41;\n"
"	mul.f32 	%f49, %f44, %f29;\n"
"	sub.f32 	%f50, %f45, %f47;\n"
"	sub.f32 	%f51, %f48, %f42;\n"
"	add.f32 	%f52, %f46, %f49;\n"
"	mul.f32 	%f53, %f50, %f34;\n"
"	sub.f32 	%f54, %f51, %f43;\n"
"	mul.f32 	%f55, %f52, %f35;\n"
"	mul.f32 	%f56, %f50, %f53;\n"
"	mul.f32 	%f57, %f54, %f33;\n"
"	mad.f32 	%f58, %f54, %f57, %f56;\n"
"	mad.f32 	%f59, %f52, %f55, %f58;\n"
"	add.f32 	%f60, %f59, %f18;\n"
"	mov.f32 	%f61, %f60;\n"
"	mul.f32 	%f62, %f38, %f29;\n"
"	sub.f32 	%f63, %f40, %f41;\n"
"	mul.f32 	%f64, %f44, %f32;\n"
"	add.f32 	%f65, %f45, %f47;\n"
"	add.f32 	%f66, %f42, %f63;\n"
"	sub.f32 	%f67, %f64, %f62;\n"
"	mul.f32 	%f68, %f65, %f33;\n"
"	sub.f32 	%f69, %f66, %f43;\n"
"	mul.f32 	%f70, %f67, %f35;\n"
"	mul.f32 	%f71, %f69, %f34;\n"
"	mul.f32 	%f72, %f50, %f71;\n"
"	mad.f32 	%f73, %f54, %f68, %f72;\n"
"	mad.f32 	%f74, %f52, %f70, %f73;\n"
"	mov.f32 	%f75, %f74;\n"
"	sub.f32 	%f76, %f63, %f42;\n"
"	sub.f32 	%f77, %f46, %f49;\n"
"	add.f32 	%f78, %f62, %f64;\n"
"	add.f32 	%f79, %f43, %f76;\n"
"	mul.f32 	%f80, %f77, %f33;\n"
"	mul.f32 	%f81, %f78, %f34;\n"
"	mul.f32 	%f82, %f79, %f35;\n"
"	mul.f32 	%f83, %f50, %f81;\n"
"	mad.f32 	%f84, %f54, %f80, %f83;\n"
"	mad.f32 	%f85, %f52, %f82, %f84;\n"
"	mov.f32 	%f86, %f85;\n"
"	sub.f32 	%f87, %f26, %f15;\n"
"	mov.f32 	%f88, %f87;\n"
"	mul.f32 	%f89, %f53, %f69;\n"
"	mad.f32 	%f90, %f57, %f65, %f89;\n"
"	mad.f32 	%f91, %f55, %f67, %f90;\n"
"	mov.f32 	%f92, %f91;\n"
"	mul.f32 	%f93, %f69, %f71;\n"
"	mad.f32 	%f94, %f65, %f68, %f93;\n"
"	mad.f32 	%f95, %f67, %f70, %f94;\n"
"	add.f32 	%f96, %f95, %f18;\n"
"	mov.f32 	%f97, %f96;\n"
"	mul.f32 	%f98, %f69, %f81;\n"
"	mad.f32 	%f99, %f65, %f80, %f98;\n"
"	mad.f32 	%f100, %f67, %f82, %f99;\n"
"	mov.f32 	%f101, %f100;\n"
"	sub.f32 	%f102, %f27, %f16;\n"
"	mov.f32 	%f103, %f102;\n"
"	mul.f32 	%f104, %f78, %f53;\n"
"	mad.f32 	%f105, %f57, %f77, %f104;\n"
"	mad.f32 	%f106, %f55, %f79, %f105;\n"
"	mov.f32 	%f107, %f106;\n"
"	mul.f32 	%f108, %f78, %f71;\n"
"	mad.f32 	%f109, %f68, %f77, %f108;\n"
"	mad.f32 	%f110, %f70, %f79, %f109;\n"
"	mov.f32 	%f111, %f110;\n"
"	mul.f32 	%f112, %f78, %f81;\n"
"	mad.f32 	%f113, %f77, %f80, %f112;\n"
"	mad.f32 	%f114, %f79, %f82, %f113;\n"
"	add.f32 	%f115, %f114, %f18;\n"
"	mov.f32 	%f116, %f115;\n"
"	.loc	14	186	0\n"
"	abs.f32 	%f117, %f91;\n"
"	abs.f32 	%f118, %f60;\n"
"	setp.gt.f32 	%p6, %f117, %f118;\n"
"	@!%p6 bra 	$Lt_0_32258;\n"
"	mov.f32 	%f61, %f91;\n"
"	mov.f32 	%f92, %f60;\n"
"	mov.f32 	%f75, %f96;\n"
"	mov.f32 	%f97, %f74;\n"
"	mov.f32 	%f86, %f100;\n"
"	mov.f32 	%f101, %f85;\n"
"	mov.f32 	%f37, %f87;\n"
"	mov.f32 	%f88, %f36;\n"
"$Lt_0_32258:\n"
"	.loc	14	193	0\n"
"	mov.f32 	%f119, %f61;\n"
"	abs.f32 	%f120, %f119;\n"
"	abs.f32 	%f121, %f106;\n"
"	setp.lt.f32 	%p7, %f120, %f121;\n"
"	@!%p7 bra 	$Lt_0_32770;\n"
"	mov.f32 	%f61, %f106;\n"
"	mov.f32 	%f107, %f119;\n"
"	.loc	14	196	0\n"
"	mov.f32 	%f122, %f75;\n"
"	mov.f32 	%f75, %f110;\n"
"	mov.f32 	%f111, %f122;\n"
"	.loc	14	197	0\n"
"	mov.f32 	%f122, %f86;\n"
"	mov.f32 	%f86, %f115;\n"
"	mov.f32 	%f116, %f122;\n"
"	.loc	14	198	0\n"
"	mov.f32 	%f122, %f37;\n"
"	mov.f32 	%f37, %f102;\n"
"	mov.f32 	%f103, %f122;\n"
"$Lt_0_32770:\n"
"	mov.f32 	%f123, %f61;\n"
"	mov.f32 	%f124, 0f00000000;   	\n"
"	setp.neu.f32 	%p8, %f123, %f124;\n"
"	@!%p8 bra 	$Lt_0_33538;\n"
"	bra.uni 	$Lt_0_34306;\n"
"$Lt_0_33538:\n"
"	mov.f32 	%f125, 0f00000000;   	\n"
"	setp.neu.f32 	%p9, %f92, %f125;\n"
"	@!%p9 bra 	$Lt_0_34050;\n"
"	mov.f32 	%f61, %f92;\n"
"	mov.f32 	%f92, %f123;\n"
"	.loc	14	213	0\n"
"	mov.f32 	%f126, %f75;\n"
"	mov.f32 	%f75, %f97;\n"
"	mov.f32 	%f97, %f126;\n"
"	.loc	14	214	0\n"
"	mov.f32 	%f127, %f86;\n"
"	mov.f32 	%f86, %f101;\n"
"	mov.f32 	%f101, %f127;\n"
"	.loc	14	215	0\n"
"	mov.f32 	%f128, %f37;\n"
"	mov.f32 	%f37, %f88;\n"
"	mov.f32 	%f88, %f128;\n"
"	bra.uni 	$Lt_0_34306;\n"
"$Lt_0_34050:\n"
"	mov.f32 	%f129, 0f00000000;   	\n"
"	setp.neu.f32 	%p10, %f107, %f129;\n"
"	@!%p10 bra 	$Lt_0_34562;\n"
"	mov.f32 	%f61, %f107;\n"
"	mov.f32 	%f107, %f123;\n"
"	.loc	14	221	0\n"
"	mov.f32 	%f130, %f75;\n"
"	mov.f32 	%f75, %f111;\n"
"	mov.f32 	%f111, %f130;\n"
"	.loc	14	222	0\n"
"	mov.f32 	%f131, %f86;\n"
"	mov.f32 	%f86, %f116;\n"
"	mov.f32 	%f116, %f131;\n"
"	.loc	14	223	0\n"
"	mov.f32 	%f132, %f37;\n"
"	mov.f32 	%f37, %f103;\n"
"	mov.f32 	%f103, %f132;\n"
"	bra.uni 	$Lt_0_34306;\n"
"$Lt_0_34562:\n"
"	.loc	14	226	0\n"
"	mov.s32 	%r25, 2;\n"
"	ld.param.u64 	%rd46, [__cudaparm_kernel_sphere_gb_err_flag];\n"
"	st.global.s32 	[%rd46+0], %r25;\n"
"$Lt_0_34306:\n"
"$Lt_0_33794:\n"
"$Lt_0_33282:\n"
"	.loc	14	228	0\n"
"	div.approx.f32 	%f133, %f92, %f61;\n"
"	mul.f32 	%f134, %f75, %f133;\n"
"	sub.f32 	%f135, %f97, %f134;\n"
"	mov.f32 	%f97, %f135;\n"
"	mul.f32 	%f136, %f86, %f133;\n"
"	sub.f32 	%f137, %f101, %f136;\n"
"	mov.f32 	%f101, %f137;\n"
"	mul.f32 	%f138, %f37, %f133;\n"
"	sub.f32 	%f139, %f88, %f138;\n"
"	mov.f32 	%f88, %f139;\n"
"	.loc	14	232	0\n"
"	div.approx.f32 	%f140, %f107, %f61;\n"
"	mul.f32 	%f141, %f75, %f140;\n"
"	sub.f32 	%f111, %f111, %f141;\n"
"	mul.f32 	%f142, %f86, %f140;\n"
"	sub.f32 	%f116, %f116, %f142;\n"
"	mul.f32 	%f143, %f37, %f140;\n"
"	sub.f32 	%f103, %f103, %f143;\n"
"	.loc	14	237	0\n"
"	abs.f32 	%f144, %f135;\n"
"	abs.f32 	%f145, %f111;\n"
"	setp.lt.f32 	%p11, %f144, %f145;\n"
"	@!%p11 bra 	$Lt_0_34818;\n"
"	mov.f32 	%f97, %f111;\n"
"	mov.f32 	%f111, %f135;\n"
"	mov.f32 	%f101, %f116;\n"
"	mov.f32 	%f116, %f137;\n"
"	mov.f32 	%f88, %f103;\n"
"	mov.f32 	%f103, %f139;\n"
"$Lt_0_34818:\n"
"	mov.f32 	%f146, %f97;\n"
"	mov.f32 	%f147, 0f00000000;   	\n"
"	setp.neu.f32 	%p12, %f146, %f147;\n"
"	@!%p12 bra 	$Lt_0_35586;\n"
"	bra.uni 	$Lt_0_35842;\n"
"$Lt_0_35586:\n"
"	mov.f32 	%f148, 0f00000000;   	\n"
"	setp.neu.f32 	%p13, %f111, %f148;\n"
"	@!%p13 bra 	$Lt_0_35842;\n"
"	mov.f32 	%f97, %f111;\n"
"	mov.f32 	%f111, %f146;\n"
"	.loc	14	258	0\n"
"	mov.f32 	%f149, %f101;\n"
"	mov.f32 	%f101, %f116;\n"
"	mov.f32 	%f116, %f149;\n"
"	.loc	14	259	0\n"
"	mov.f32 	%f149, %f88;\n"
"	mov.f32 	%f88, %f103;\n"
"	mov.f32 	%f103, %f149;\n"
"$Lt_0_35842:\n"
"$Lt_0_35330:\n"
"	.loc	14	263	0\n"
"	div.approx.f32 	%f150, %f111, %f97;\n"
"	mul.f32 	%f151, %f101, %f150;\n"
"	sub.f32 	%f116, %f116, %f151;\n"
"	mul.f32 	%f152, %f88, %f150;\n"
"	sub.f32 	%f103, %f103, %f152;\n"
"	mov.f32 	%f153, 0f00000000;   	\n"
"	setp.eq.f32 	%p14, %f116, %f153;\n"
"	@!%p14 bra 	$Lt_0_36354;\n"
"	.loc	14	268	0\n"
"	mov.s32 	%r26, 2;\n"
"	ld.param.u64 	%rd47, [__cudaparm_kernel_sphere_gb_err_flag];\n"
"	st.global.s32 	[%rd47+0], %r26;\n"
"$Lt_0_36354:\n"
"	.loc	15	126	0\n"
"	div.approx.f32 	%f154, %f103, %f116;\n"
"	mul.f32 	%f155, %f87, %f87;\n"
"	mul.f32 	%f156, %f154, %f101;\n"
"	mad.f32 	%f157, %f36, %f36, %f155;\n"
"	sub.f32 	%f158, %f88, %f156;\n"
"	mad.f32 	%f159, %f102, %f102, %f157;\n"
"	div.approx.f32 	%f160, %f158, %f97;\n"
"	rsqrt.approx.f32 	%f161, %f159;\n"
"	mul.f32 	%f162, %f160, %f75;\n"
"	mad.f32 	%f163, %f86, %f154, %f162;\n"
"	sub.f32 	%f164, %f37, %f163;\n"
"	div.approx.f32 	%f165, %f164, %f61;\n"
"	mul.f32 	%f166, %f161, %f165;\n"
"	.loc	15	138	0\n"
"	mul.f32 	%f167, %f160, %f161;\n"
"	mul.f32 	%f168, %f161, %f87;\n"
"	mul.f32 	%f169, %f161, %f36;\n"
"	mul.f32 	%f170, %f161, %f102;\n"
"	mul.f32 	%f171, %f154, %f161;\n"
"	mul.f32 	%f172, %f167, %f168;\n"
"	mad.f32 	%f173, %f169, %f166, %f172;\n"
"	mad.f32 	%f174, %f170, %f171, %f173;\n"
"	mov.f32 	%f175, 0f3f000000;   	\n"
"	mul.f32 	%f176, %f174, %f175;\n"
"	rsqrt.approx.f32 	%f177, %f176;\n"
"	.loc	15	142	0\n"
"	rcp.approx.f32 	%f178, %f161;\n"
"	mul.f32 	%f166, %f178, %f166;\n"
"	.loc	15	147	0\n"
"	add.s32 	%r27, %r24, %r20;\n"
"	cvt.u64.s32 	%rd48, %r27;\n"
"	mul.lo.u64 	%rd49, %rd48, 8;\n"
"	add.u64 	%rd50, %rd33, %rd49;\n"
"	ld.global.v2.f32 	{%f179,%f180}, [%rd50+0];\n"
"	.loc	15	149	0\n"
"	sub.f32 	%f181, %f178, %f177;\n"
"	ld.global.f32 	%f182, [%rd32+0];\n"
"	mad.f32 	%f183, %f182, %f179, %f181;\n"
"	.loc	15	156	0\n"
"	div.approx.f32 	%f184, %f179, %f183;\n"
"	mul.f32 	%f185, %f184, %f184;\n"
"	mul.f32 	%f186, %f184, %f185;\n"
"	mul.f32 	%f187, %f186, %f186;\n"
"	mul.f32 	%f188, %f187, %f187;\n"
"	mul.f32 	%f189, %f184, %f187;\n"
"	add.f32 	%f190, %f188, %f188;\n"
"	mul.f32 	%f191, %f184, %f190;\n"
"	sub.f32 	%f192, %f191, %f189;\n"
"	div.approx.f32 	%f193, %f192, %f179;\n"
"	mov.f32 	%f194, 0f41c00000;   	\n"
"	mul.f32 	%f195, %f193, %f194;\n"
"	mul.f32 	%f196, %f180, %f195;\n"
"	.loc	15	161	0\n"
"	mul.f32 	%f197, %f177, %f196;\n"
"	mul.f32 	%f198, %f177, %f197;\n"
"	mul.f32 	%f199, %f177, %f198;\n"
"	mov.f32 	%f200, 0f3f000000;   	\n"
"	mul.f32 	%f201, %f199, %f200;\n"
"	mul.f32 	%f202, %f201, %f161;\n"
"	mul.f32 	%f203, %f167, %f178;\n"
"	mul.f32 	%f204, %f171, %f178;\n"
"	mul.f32 	%f205, %f161, %f202;\n"
"	mul.f32 	%f206, %f168, %f203;\n"
"	mad.f32 	%f207, %f169, %f166, %f206;\n"
"	mad.f32 	%f208, %f170, %f204, %f207;\n"
"	mul.f32 	%f209, %f169, %f208;\n"
"	sub.f32 	%f210, %f166, %f209;\n"
"	mul.f32 	%f211, %f205, %f210;\n"
"	mad.f32 	%f212, %f169, %f196, %f211;\n"
"	.loc	15	162	0\n"
"	mul.f32 	%f213, %f168, %f196;\n"
"	mul.f32 	%f214, %f208, %f168;\n"
"	sub.f32 	%f215, %f203, %f214;\n"
"	mad.f32 	%f216, %f205, %f215, %f213;\n"
"	.loc	15	163	0\n"
"	mul.f32 	%f217, %f170, %f196;\n"
"	mul.f32 	%f218, %f208, %f170;\n"
"	sub.f32 	%f219, %f204, %f218;\n"
"	mad.f32 	%f220, %f205, %f219, %f217;\n"
"	.loc	15	170	0\n"
"	ld.global.f32 	%f221, [%rd31+0];\n"
"	mul.lo.u64 	%rd51, %rd43, 4;\n"
"	add.u64 	%rd52, %rd29, %rd51;\n"
"	ld.global.f32 	%f222, [%rd52+0];\n"
"	add.f32 	%f223, %f221, %f221;\n"
"	mul.f32 	%f224, %f222, %f223;\n"
"	.loc	15	171	0\n"
"	mul.f32 	%f225, %f100, %f60;\n"
"	mul.f32 	%f226, %f110, %f225;\n"
"	mul.f32 	%f227, %f96, %f60;\n"
"	mul.f32 	%f228, %f115, %f227;\n"
"	sub.f32 	%f229, %f228, %f226;\n"
"	mul.f32 	%f230, %f91, %f74;\n"
"	mul.f32 	%f231, %f115, %f230;\n"
"	sub.f32 	%f232, %f229, %f231;\n"
"	mul.f32 	%f233, %f91, %f85;\n"
"	mad.f32 	%f234, %f110, %f233, %f232;\n"
"	mul.f32 	%f235, %f106, %f74;\n"
"	mad.f32 	%f236, %f100, %f235, %f234;\n"
"	mul.f32 	%f237, %f106, %f85;\n"
"	mul.f32 	%f238, %f96, %f237;\n"
"	sub.f32 	%f239, %f236, %f238;\n"
"	.loc	15	172	0\n"
"	ld.global.f32 	%f240, [%rd32+4];\n"
"	div.approx.f32 	%f241, %f224, %f239;\n"
"	lg2.approx.f32 	%f242, %f241;\n"
"	mul.f32 	%f243, %f240, %f242;\n"
"	ex2.approx.f32 	%f224, %f243;\n"
"	.loc	15	183	0\n"
"	add.u64 	%rd53, %rd44, %rd24;\n"
"	ld.global.v4.f32 	{%f244,%f245,%f246,_}, [%rd53+0];\n"
"	.loc	15	198	0\n"
"	mov.f32 	%f37, %f36;\n"
"	mul.f32 	%f247, %f50, %f245;\n"
"	mul.f32 	%f248, %f52, %f246;\n"
"	mul.f32 	%f249, %f50, %f247;\n"
"	mul.f32 	%f250, %f54, %f244;\n"
"	mad.f32 	%f251, %f54, %f250, %f249;\n"
"	mad.f32 	%f252, %f52, %f248, %f251;\n"
"	add.f32 	%f253, %f252, %f19;\n"
"	mov.f32 	%f61, %f253;\n"
"	mul.f32 	%f254, %f65, %f244;\n"
"	mul.f32 	%f255, %f67, %f246;\n"
"	mul.f32 	%f256, %f69, %f245;\n"
"	mul.f32 	%f257, %f50, %f256;\n"
"	mad.f32 	%f258, %f54, %f254, %f257;\n"
"	mad.f32 	%f259, %f52, %f255, %f258;\n"
"	mov.f32 	%f75, %f259;\n"
"	mul.f32 	%f260, %f77, %f244;\n"
"	mul.f32 	%f261, %f78, %f245;\n"
"	mul.f32 	%f262, %f79, %f246;\n"
"	mul.f32 	%f263, %f50, %f261;\n"
"	mad.f32 	%f264, %f54, %f260, %f263;\n"
"	mad.f32 	%f265, %f52, %f262, %f264;\n"
"	mov.f32 	%f86, %f265;\n"
"	mov.f32 	%f88, %f87;\n"
"	mul.f32 	%f266, %f247, %f69;\n"
"	mad.f32 	%f267, %f250, %f65, %f266;\n"
"	mad.f32 	%f268, %f248, %f67, %f267;\n"
"	mov.f32 	%f92, %f268;\n"
"	mul.f32 	%f269, %f69, %f256;\n"
"	mad.f32 	%f270, %f65, %f254, %f269;\n"
"	mad.f32 	%f271, %f67, %f255, %f270;\n"
"	add.f32 	%f97, %f19, %f271;\n"
"	mul.f32 	%f272, %f69, %f261;\n"
"	mad.f32 	%f273, %f65, %f260, %f272;\n"
"	mad.f32 	%f101, %f67, %f262, %f273;\n"
"	mov.f32 	%f103, %f102;\n"
"	mul.f32 	%f274, %f78, %f247;\n"
"	mad.f32 	%f275, %f250, %f77, %f274;\n"
"	mad.f32 	%f276, %f248, %f79, %f275;\n"
"	mov.f32 	%f107, %f276;\n"
"	mul.f32 	%f277, %f78, %f256;\n"
"	mad.f32 	%f278, %f254, %f77, %f277;\n"
"	mad.f32 	%f111, %f255, %f79, %f278;\n"
"	mul.f32 	%f279, %f78, %f261;\n"
"	mad.f32 	%f280, %f77, %f260, %f279;\n"
"	mad.f32 	%f281, %f79, %f262, %f280;\n"
"	add.f32 	%f116, %f19, %f281;\n"
"	.loc	14	186	0\n"
"	abs.f32 	%f282, %f268;\n"
"	abs.f32 	%f283, %f253;\n"
"	setp.gt.f32 	%p15, %f282, %f283;\n"
"	@!%p15 bra 	$Lt_0_36866;\n"
"	mov.f32 	%f61, %f268;\n"
"	mov.f32 	%f92, %f253;\n"
"	mov.f32 	%f75, %f97;\n"
"	mov.f32 	%f97, %f259;\n"
"	mov.f32 	%f86, %f101;\n"
"	mov.f32 	%f101, %f265;\n"
"	mov.f32 	%f37, %f87;\n"
"	mov.f32 	%f88, %f36;\n"
"$Lt_0_36866:\n"
"	.loc	14	193	0\n"
"	mov.f32 	%f284, %f61;\n"
"	abs.f32 	%f285, %f284;\n"
"	abs.f32 	%f286, %f276;\n"
"	setp.lt.f32 	%p16, %f285, %f286;\n"
"	@!%p16 bra 	$Lt_0_37378;\n"
"	mov.f32 	%f61, %f276;\n"
"	mov.f32 	%f107, %f284;\n"
"	.loc	14	196	0\n"
"	mov.f32 	%f122, %f75;\n"
"	mov.f32 	%f75, %f111;\n"
"	mov.f32 	%f111, %f122;\n"
"	.loc	14	197	0\n"
"	mov.f32 	%f122, %f86;\n"
"	mov.f32 	%f86, %f116;\n"
"	mov.f32 	%f116, %f122;\n"
"	.loc	14	198	0\n"
"	mov.f32 	%f122, %f37;\n"
"	mov.f32 	%f37, %f102;\n"
"	mov.f32 	%f103, %f122;\n"
"$Lt_0_37378:\n"
"	mov.f32 	%f287, %f61;\n"
"	mov.f32 	%f288, 0f00000000;   	\n"
"	setp.neu.f32 	%p17, %f287, %f288;\n"
"	@!%p17 bra 	$Lt_0_38146;\n"
"	bra.uni 	$Lt_0_38914;\n"
"$Lt_0_38146:\n"
"	mov.f32 	%f289, 0f00000000;   	\n"
"	setp.neu.f32 	%p18, %f92, %f289;\n"
"	@!%p18 bra 	$Lt_0_38658;\n"
"	mov.f32 	%f61, %f92;\n"
"	mov.f32 	%f92, %f287;\n"
"	.loc	14	213	0\n"
"	mov.f32 	%f290, %f75;\n"
"	mov.f32 	%f75, %f97;\n"
"	mov.f32 	%f97, %f290;\n"
"	.loc	14	214	0\n"
"	mov.f32 	%f291, %f86;\n"
"	mov.f32 	%f86, %f101;\n"
"	mov.f32 	%f101, %f291;\n"
"	.loc	14	215	0\n"
"	mov.f32 	%f292, %f37;\n"
"	mov.f32 	%f37, %f88;\n"
"	mov.f32 	%f88, %f292;\n"
"	bra.uni 	$Lt_0_38914;\n"
"$Lt_0_38658:\n"
"	mov.f32 	%f293, 0f00000000;   	\n"
"	setp.neu.f32 	%p19, %f107, %f293;\n"
"	@!%p19 bra 	$Lt_0_39170;\n"
"	mov.f32 	%f61, %f107;\n"
"	mov.f32 	%f107, %f287;\n"
"	.loc	14	221	0\n"
"	mov.f32 	%f294, %f75;\n"
"	mov.f32 	%f75, %f111;\n"
"	mov.f32 	%f111, %f294;\n"
"	.loc	14	222	0\n"
"	mov.f32 	%f295, %f86;\n"
"	mov.f32 	%f86, %f116;\n"
"	mov.f32 	%f116, %f295;\n"
"	.loc	14	223	0\n"
"	mov.f32 	%f296, %f37;\n"
"	mov.f32 	%f37, %f103;\n"
"	mov.f32 	%f103, %f296;\n"
"	bra.uni 	$Lt_0_38914;\n"
"$Lt_0_39170:\n"
"	.loc	14	226	0\n"
"	mov.s32 	%r28, 2;\n"
"	ld.param.u64 	%rd54, [__cudaparm_kernel_sphere_gb_err_flag];\n"
"	st.global.s32 	[%rd54+0], %r28;\n"
"$Lt_0_38914:\n"
"$Lt_0_38402:\n"
"$Lt_0_37890:\n"
"	.loc	14	228	0\n"
"	div.approx.f32 	%f297, %f92, %f61;\n"
"	mul.f32 	%f298, %f75, %f297;\n"
"	sub.f32 	%f299, %f97, %f298;\n"
"	mov.f32 	%f97, %f299;\n"
"	mul.f32 	%f300, %f86, %f297;\n"
"	sub.f32 	%f301, %f101, %f300;\n"
"	mov.f32 	%f101, %f301;\n"
"	mul.f32 	%f302, %f37, %f297;\n"
"	sub.f32 	%f303, %f88, %f302;\n"
"	mov.f32 	%f88, %f303;\n"
"	.loc	14	232	0\n"
"	div.approx.f32 	%f304, %f107, %f61;\n"
"	mul.f32 	%f305, %f75, %f304;\n"
"	sub.f32 	%f111, %f111, %f305;\n"
"	mul.f32 	%f306, %f86, %f304;\n"
"	sub.f32 	%f116, %f116, %f306;\n"
"	mul.f32 	%f307, %f37, %f304;\n"
"	sub.f32 	%f103, %f103, %f307;\n"
"	.loc	14	237	0\n"
"	abs.f32 	%f308, %f299;\n"
"	abs.f32 	%f309, %f111;\n"
"	setp.lt.f32 	%p20, %f308, %f309;\n"
"	@!%p20 bra 	$Lt_0_39426;\n"
"	mov.f32 	%f97, %f111;\n"
"	mov.f32 	%f111, %f299;\n"
"	mov.f32 	%f101, %f116;\n"
"	mov.f32 	%f116, %f301;\n"
"	mov.f32 	%f88, %f103;\n"
"	mov.f32 	%f103, %f303;\n"
"$Lt_0_39426:\n"
"	mov.f32 	%f310, %f97;\n"
"	mov.f32 	%f311, 0f00000000;   	\n"
"	setp.neu.f32 	%p21, %f310, %f311;\n"
"	@!%p21 bra 	$Lt_0_40194;\n"
"	bra.uni 	$Lt_0_40450;\n"
"$Lt_0_40194:\n"
"	mov.f32 	%f312, 0f00000000;   	\n"
"	setp.neu.f32 	%p22, %f111, %f312;\n"
"	@!%p22 bra 	$Lt_0_40450;\n"
"	mov.f32 	%f97, %f111;\n"
"	mov.f32 	%f111, %f310;\n"
"	.loc	14	258	0\n"
"	mov.f32 	%f149, %f101;\n"
"	mov.f32 	%f101, %f116;\n"
"	mov.f32 	%f116, %f149;\n"
"	.loc	14	259	0\n"
"	mov.f32 	%f149, %f88;\n"
"	mov.f32 	%f88, %f103;\n"
"	mov.f32 	%f103, %f149;\n"
"$Lt_0_40450:\n"
"$Lt_0_39938:\n"
"	.loc	14	263	0\n"
"	div.approx.f32 	%f313, %f111, %f97;\n"
"	mul.f32 	%f314, %f101, %f313;\n"
"	sub.f32 	%f116, %f116, %f314;\n"
"	mul.f32 	%f315, %f88, %f313;\n"
"	sub.f32 	%f103, %f103, %f315;\n"
"	mov.f32 	%f316, 0f00000000;   	\n"
"	setp.eq.f32 	%p23, %f116, %f316;\n"
"	@!%p23 bra 	$Lt_0_40962;\n"
"	.loc	14	268	0\n"
"	mov.s32 	%r29, 2;\n"
"	ld.param.u64 	%rd55, [__cudaparm_kernel_sphere_gb_err_flag];\n"
"	st.global.s32 	[%rd55+0], %r29;\n"
"$Lt_0_40962:\n"
"	.loc	15	200	0\n"
"	div.approx.f32 	%f317, %f103, %f116;\n"
"	mul.f32 	%f318, %f317, %f101;\n"
"	sub.f32 	%f319, %f88, %f318;\n"
"	div.approx.f32 	%f320, %f319, %f97;\n"
"	mul.f32 	%f321, %f320, %f75;\n"
"	mad.f32 	%f322, %f86, %f317, %f321;\n"
"	sub.f32 	%f323, %f37, %f322;\n"
"	div.approx.f32 	%f324, %f323, %f61;\n"
"	mul.f32 	%f325, %f161, %f324;\n"
"	.loc	15	204	0\n"
"	ld.global.f32 	%f326, [%rd32+8];\n"
"	.loc	19	4494	0\n"
"	mul.f32 	%f327, %f320, %f161;\n"
"	mul.f32 	%f328, %f317, %f161;\n"
"	mul.f32 	%f329, %f327, %f168;\n"
"	mad.f32 	%f330, %f169, %f325, %f329;\n"
"	mad.f32 	%f331, %f170, %f328, %f330;\n"
"	add.f32 	%f332, %f331, %f331;\n"
"	lg2.approx.f32 	%f333, %f332;\n"
"	.loc	15	204	0\n"
"	mul.f32 	%f334, %f333, %f326;\n"
"	ex2.approx.f32 	%f335, %f334;\n"
"	.loc	15	207	0\n"
"	mul.f32 	%f325, %f178, %f325;\n"
"	.loc	15	212	0\n"
"	mov.f32 	%f336, 0fbf800000;   	\n"
"	add.f32 	%f337, %f326, %f336;\n"
"	.loc	19	4494	0\n"
"	lg2.approx.f32 	%f333, %f335;\n"
"	.loc	15	212	0\n"
"	div.approx.f32 	%f338, %f337, %f326;\n"
"	mul.f32 	%f339, %f333, %f338;\n"
"	ex2.approx.f32 	%f340, %f339;\n"
"	mov.f32 	%f341, 0fc0800000;   	\n"
"	mul.f32 	%f342, %f161, %f341;\n"
"	mul.f32 	%f343, %f161, %f342;\n"
"	mul.f32 	%f344, %f326, %f343;\n"
"	mul.f32 	%f345, %f340, %f344;\n"
"	.loc	15	213	0\n"
"	mul.f32 	%f346, %f327, %f178;\n"
"	mul.f32 	%f347, %f328, %f178;\n"
"	mul.f32 	%f348, %f168, %f346;\n"
"	mad.f32 	%f349, %f169, %f325, %f348;\n"
"	mad.f32 	%f350, %f170, %f347, %f349;\n"
"	mul.f32 	%f351, %f169, %f350;\n"
"	sub.f32 	%f352, %f325, %f351;\n"
"	mul.f32 	%f353, %f345, %f352;\n"
"	.loc	15	214	0\n"
"	mul.f32 	%f354, %f350, %f168;\n"
"	sub.f32 	%f355, %f346, %f354;\n"
"	mul.f32 	%f356, %f345, %f355;\n"
"	.loc	15	215	0\n"
"	mul.f32 	%f357, %f350, %f170;\n"
"	sub.f32 	%f358, %f347, %f357;\n"
"	mul.f32 	%f359, %f345, %f358;\n"
"	.loc	14	270	0\n"
"	mov.f32 	%f360, 0f40800000;   	\n"
"	mul.f32 	%f361, %f180, %f360;\n"
"	.loc	15	75	0\n"
"	ld.shared.f32 	%f24, [%rd38+0];\n"
"	.loc	14	270	0\n"
"	mul.f32 	%f362, %f24, %f224;\n"
"	mul.f32 	%f363, %f362, %f335;\n"
"	sub.f32 	%f364, %f188, %f187;\n"
"	mul.f32 	%f365, %f361, %f364;\n"
"	mad.f32 	%f366, %f365, %f363, %f23;\n"
"	selp.f32 	%f23, %f366, %f23, %p4;\n"
"	mul.f32 	%f367, %f363, %f212;\n"
"	mul.f32 	%f368, %f363, %f216;\n"
"	mul.f32 	%f369, %f363, %f220;\n"
"	mul.f32 	%f370, %f365, %f224;\n"
"	mul.f32 	%f371, %f370, %f24;\n"
"	neg.f32 	%f372, %f371;\n"
"	mul.f32 	%f373, %f353, %f372;\n"
"	sub.f32 	%f374, %f373, %f367;\n"
"	mul.f32 	%f375, %f372, %f356;\n"
"	sub.f32 	%f376, %f375, %f368;\n"
"	mul.f32 	%f377, %f372, %f359;\n"
"	sub.f32 	%f378, %f377, %f369;\n"
"	@!%p5 bra 	$Lt_0_41730;\n"
"	.loc	15	227	0\n"
"	add.f32 	%f22, %f374, %f22;\n"
"	.loc	15	228	0\n"
"	sub.f32 	%f379, %f14, %f25;\n"
"	mov.f32 	%f380, %f3;\n"
"	mad.f32 	%f381, %f379, %f374, %f380;\n"
"	mov.f32 	%f3, %f381;\n"
"	.loc	15	230	0\n"
"	add.f32 	%f21, %f376, %f21;\n"
"	.loc	15	231	0\n"
"	sub.f32 	%f382, %f15, %f26;\n"
"	mov.f32 	%f383, %f5;\n"
"	mad.f32 	%f384, %f376, %f382, %f383;\n"
"	mov.f32 	%f5, %f384;\n"
"	.loc	15	232	0\n"
"	mov.f32 	%f385, %f9;\n"
"	mad.f32 	%f386, %f379, %f376, %f385;\n"
"	mov.f32 	%f9, %f386;\n"
"	.loc	15	234	0\n"
"	add.f32 	%f20, %f378, %f20;\n"
"	.loc	15	235	0\n"
"	mov.f32 	%f387, %f7;\n"
"	sub.f32 	%f388, %f16, %f27;\n"
"	mad.f32 	%f389, %f378, %f388, %f387;\n"
"	mov.f32 	%f7, %f389;\n"
"	.loc	15	236	0\n"
"	mov.f32 	%f390, %f11;\n"
"	mad.f32 	%f391, %f379, %f378, %f390;\n"
"	mov.f32 	%f11, %f391;\n"
"	.loc	15	237	0\n"
"	mad.f32 	%f12, %f378, %f382, %f12;\n"
"	mov.f32 	%f392, %f12;\n"
"	bra.uni 	$Lt_0_41474;\n"
"$Lt_0_41730:\n"
"	.loc	15	239	0\n"
"	add.f32 	%f22, %f374, %f22;\n"
"	.loc	15	240	0\n"
"	add.f32 	%f21, %f376, %f21;\n"
"	.loc	15	241	0\n"
"	add.f32 	%f20, %f378, %f20;\n"
"$Lt_0_41474:\n"
"	.loc	15	72	0\n"
"	add.u64 	%rd15, %rd12, %rd15;\n"
"	setp.gt.u64 	%p24, %rd28, %rd15;\n"
"	@%p24 bra 	$Lt_0_32002;\n"
"	bra.uni 	$Lt_0_31490;\n"
"$Lt_0_43266:\n"
"	mov.s32 	%r30, 0;\n"
"	setp.gt.s32 	%p4, %r16, %r30;\n"
"	mov.s32 	%r31, 0;\n"
"	setp.gt.s32 	%p5, %r15, %r31;\n"
"$Lt_0_31490:\n"
"	.loc	15	246	0\n"
"	ld.param.u64 	%rd56, [__cudaparm_kernel_sphere_gb_engv];\n"
"	add.u64 	%rd57, %rd56, %rd8;\n"
"	@!%p4 bra 	$Lt_0_42242;\n"
"	.loc	15	248	0\n"
"	st.global.f32 	[%rd57+0], %f23;\n"
"	.loc	15	249	0\n"
"	cvt.u64.s32 	%rd58, %r9;\n"
"	mul.lo.u64 	%rd59, %rd58, 4;\n"
"	add.u64 	%rd57, %rd57, %rd59;\n"
"$Lt_0_42242:\n"
"	@!%p5 bra 	$Lt_0_42754;\n"
"	.loc	15	253	0\n"
"	mov.f32 	%f393, %f3;\n"
"	st.global.f32 	[%rd57+0], %f393;\n"
"	.loc	15	254	0\n"
"	cvt.u64.s32 	%rd60, %r9;\n"
"	mul.lo.u64 	%rd61, %rd60, 4;\n"
"	add.u64 	%rd57, %rd61, %rd57;\n"
"	.loc	15	253	0\n"
"	mov.f32 	%f394, %f5;\n"
"	st.global.f32 	[%rd57+0], %f394;\n"
"	.loc	15	254	0\n"
"	add.u64 	%rd57, %rd61, %rd57;\n"
"	.loc	15	253	0\n"
"	mov.f32 	%f395, %f7;\n"
"	st.global.f32 	[%rd57+0], %f395;\n"
"	.loc	15	254	0\n"
"	add.u64 	%rd57, %rd61, %rd57;\n"
"	.loc	15	253	0\n"
"	mov.f32 	%f396, %f9;\n"
"	st.global.f32 	[%rd57+0], %f396;\n"
"	.loc	15	254	0\n"
"	add.u64 	%rd57, %rd61, %rd57;\n"
"	.loc	15	253	0\n"
"	mov.f32 	%f397, %f11;\n"
"	st.global.f32 	[%rd57+0], %f397;\n"
"	add.u64 	%rd62, %rd61, %rd57;\n"
"	st.global.f32 	[%rd62+0], %f12;\n"
"$Lt_0_42754:\n"
"	.loc	15	257	0\n"
"	ld.param.u64 	%rd63, [__cudaparm_kernel_sphere_gb_ans];\n"
"	mul.lo.u64 	%rd64, %rd7, 16;\n"
"	add.u64 	%rd65, %rd63, %rd64;\n"
"	mov.f32 	%f398, %f399;\n"
"	st.global.v4.f32 	[%rd65+0], {%f22,%f21,%f20,%f398};\n"
"$Lt_0_30978:\n"
"	.loc	15	259	0\n"
"	exit;\n"
"$LDWend_kernel_sphere_gb:\n"
"	}\n"
"	.entry kernel_lj (\n"
"		.param .u64 __cudaparm_kernel_lj_x_,\n"
"		.param .u64 __cudaparm_kernel_lj_lj1,\n"
"		.param .u64 __cudaparm_kernel_lj_lj3,\n"
"		.param .s32 __cudaparm_kernel_lj_lj_types,\n"
"		.param .u64 __cudaparm_kernel_lj_gum,\n"
"		.param .s32 __cudaparm_kernel_lj_stride,\n"
"		.param .u64 __cudaparm_kernel_lj_dev_ij,\n"
"		.param .u64 __cudaparm_kernel_lj_ans,\n"
"		.param .u64 __cudaparm_kernel_lj_engv,\n"
"		.param .u64 __cudaparm_kernel_lj_err_flag,\n"
"		.param .s32 __cudaparm_kernel_lj_eflag,\n"
"		.param .s32 __cudaparm_kernel_lj_vflag,\n"
"		.param .s32 __cudaparm_kernel_lj_start,\n"
"		.param .s32 __cudaparm_kernel_lj_inum,\n"
"		.param .s32 __cudaparm_kernel_lj_nall)\n"
"	{\n"
"	.reg .u32 %r<29>;\n"
"	.reg .u64 %rd<46>;\n"
"	.reg .f32 %f<93>;\n"
"	.reg .pred %p<10>;\n"
"	.shared .align 4 .b8 __cuda_sp_lj244[16];\n"
"	.loc	15	268	0\n"
"$LBB1_kernel_lj:\n"
"	cvt.s32.u16 	%r1, %tid.x;\n"
"	mov.u32 	%r2, 3;\n"
"	setp.gt.s32 	%p1, %r1, %r2;\n"
"	@%p1 bra 	$Lt_1_9474;\n"
"	.loc	15	274	0\n"
"	mov.u64 	%rd1, __cuda_sp_lj244;\n"
"	cvt.u64.s32 	%rd2, %r1;\n"
"	mul.lo.u64 	%rd3, %rd2, 4;\n"
"	ld.param.u64 	%rd4, [__cudaparm_kernel_lj_gum];\n"
"	add.u64 	%rd5, %rd4, %rd3;\n"
"	ld.global.f32 	%f1, [%rd5+12];\n"
"	add.u64 	%rd6, %rd3, %rd1;\n"
"	st.shared.f32 	[%rd6+0], %f1;\n"
"$Lt_1_9474:\n"
"	mov.u64 	%rd1, __cuda_sp_lj244;\n"
"	.loc	15	276	0\n"
"	bar.sync 	0;\n"
"	cvt.s32.u16 	%r3, %ctaid.x;\n"
"	cvt.s32.u16 	%r4, %ntid.x;\n"
"	mul24.lo.s32 	%r5, %r3, %r4;\n"
"	ld.param.s32 	%r6, [__cudaparm_kernel_lj_start];\n"
"	add.s32 	%r7, %r6, %r5;\n"
"	add.s32 	%r8, %r7, %r1;\n"
"	ld.param.s32 	%r9, [__cudaparm_kernel_lj_inum];\n"
"	setp.ge.s32 	%p2, %r8, %r9;\n"
"	@%p2 bra 	$Lt_1_9986;\n"
"	.loc	15	287	0\n"
"	mov.f32 	%f2, 0f00000000;     	\n"
"	mov.f32 	%f3, %f2;\n"
"	mov.f32 	%f4, 0f00000000;     	\n"
"	mov.f32 	%f5, %f4;\n"
"	mov.f32 	%f6, 0f00000000;     	\n"
"	mov.f32 	%f7, %f6;\n"
"	mov.f32 	%f8, 0f00000000;     	\n"
"	mov.f32 	%f9, %f8;\n"
"	mov.f32 	%f10, 0f00000000;    	\n"
"	mov.f32 	%f11, %f10;\n"
"	mov.f32 	%f12, 0f00000000;    	\n"
"	mov.f32 	%f13, %f12;\n"
"	.loc	15	290	0\n"
"	cvt.u64.s32 	%rd7, %r8;\n"
"	mul.lo.u64 	%rd8, %rd7, 4;\n"
"	ld.param.u64 	%rd9, [__cudaparm_kernel_lj_dev_ij];\n"
"	add.u64 	%rd10, %rd9, %rd8;\n"
"	ld.global.s32 	%r10, [%rd10+0];\n"
"	.loc	15	292	0\n"
"	ld.param.s32 	%r11, [__cudaparm_kernel_lj_stride];\n"
"	cvt.u64.s32 	%rd11, %r11;\n"
"	mul.lo.u64 	%rd12, %rd11, 4;\n"
"	add.u64 	%rd13, %rd10, %rd12;\n"
"	ld.global.s32 	%r12, [%rd13+0];\n"
"	.loc	15	293	0\n"
"	add.u64 	%rd14, %rd13, %rd12;\n"
"	mov.s64 	%rd15, %rd14;\n"
"	.loc	15	296	0\n"
"	ld.param.u64 	%rd16, [__cudaparm_kernel_lj_x_];\n"
"	cvt.u64.s32 	%rd17, %r10;\n"
"	mul.lo.u64 	%rd18, %rd17, 16;\n"
"	add.u64 	%rd19, %rd16, %rd18;\n"
"	ld.global.v4.f32 	{%f14,%f15,%f16,%f17}, [%rd19+0];\n"
"	mul24.lo.s32 	%r13, %r11, %r12;\n"
"	cvt.s64.s32 	%rd20, %r13;\n"
"	mul.lo.u64 	%rd21, %rd20, 4;\n"
"	add.u64 	%rd22, %rd14, %rd21;\n"
"	ld.param.s32 	%r14, [__cudaparm_kernel_lj_vflag];\n"
"	ld.param.s32 	%r15, [__cudaparm_kernel_lj_eflag];\n"
"	setp.ge.u64 	%p3, %rd14, %rd22;\n"
"	mov.f32 	%f18, 0f00000000;    	\n"
"	mov.f32 	%f19, 0f00000000;    	\n"
"	mov.f32 	%f20, 0f00000000;    	\n"
"	mov.f32 	%f21, 0f00000000;    	\n"
"	@%p3 bra 	$Lt_1_13570;\n"
"	mov.s32 	%r16, 0;\n"
"	setp.gt.s32 	%p4, %r15, %r16;\n"
"	mov.s32 	%r17, 0;\n"
"	setp.gt.s32 	%p5, %r14, %r17;\n"
"	cvt.rzi.s32.f32 	%r18, %f17;\n"
"	ld.param.s32 	%r19, [__cudaparm_kernel_lj_lj_types];\n"
"	mul.lo.s32 	%r20, %r19, %r18;\n"
"	ld.param.u64 	%rd23, [__cudaparm_kernel_lj_lj1];\n"
"$Lt_1_11010:\n"
"	.loc	15	302	0\n"
"	ld.global.s32 	%r21, [%rd15+0];\n"
"	.loc	15	303	0\n"
"	shr.s32 	%r22, %r21, 30;\n"
"	cvt.s64.s32 	%rd24, %r22;\n"
"	and.b64 	%rd25, %rd24, 3;\n"
"	mul.lo.u64 	%rd26, %rd25, 4;\n"
"	add.u64 	%rd27, %rd1, %rd26;\n"
"	ld.shared.f32 	%f22, [%rd27+0];\n"
"	.loc	15	306	0\n"
"	and.b32 	%r23, %r21, 1073741823;\n"
"	cvt.u64.s32 	%rd28, %r23;\n"
"	mul.lo.u64 	%rd29, %rd28, 16;\n"
"	add.u64 	%rd30, %rd16, %rd29;\n"
"	ld.global.v4.f32 	{%f23,%f24,%f25,%f26}, [%rd30+0];\n"
"	.loc	15	302	0\n"
"	cvt.rzi.s32.f32 	%r24, %f26;\n"
"	sub.f32 	%f27, %f15, %f24;\n"
"	sub.f32 	%f28, %f14, %f23;\n"
"	sub.f32 	%f29, %f16, %f25;\n"
"	mul.f32 	%f30, %f27, %f27;\n"
"	mad.f32 	%f31, %f28, %f28, %f30;\n"
"	mad.f32 	%f32, %f29, %f29, %f31;\n"
"	add.s32 	%r25, %r24, %r20;\n"
"	cvt.u64.s32 	%rd31, %r25;\n"
"	mul.lo.u64 	%rd32, %rd31, 16;\n"
"	add.u64 	%rd33, %rd32, %rd23;\n"
"	ld.global.f32 	%f33, [%rd33+8];\n"
"	setp.gt.f32 	%p6, %f33, %f32;\n"
"	@!%p6 bra 	$Lt_1_13826;\n"
"	ld.global.f32 	%f34, [%rd33+12];\n"
"	mov.f32 	%f35, 0f00000000;    	\n"
"	setp.eq.f32 	%p7, %f34, %f35;\n"
"	@!%p7 bra 	$Lt_1_13826;\n"
"	.loc	15	320	0\n"
"	rcp.approx.f32 	%f36, %f32;\n"
"	mul.f32 	%f37, %f36, %f36;\n"
"	mul.f32 	%f38, %f36, %f37;\n"
"	mul.f32 	%f39, %f36, %f38;\n"
"	ld.global.v2.f32 	{%f40,%f41}, [%rd33+0];\n"
"	mul.f32 	%f42, %f40, %f38;\n"
"	sub.f32 	%f43, %f42, %f41;\n"
"	mul.f32 	%f44, %f39, %f43;\n"
"	.loc	15	303	0\n"
"	ld.shared.f32 	%f22, [%rd27+0];\n"
"	.loc	15	320	0\n"
"	mul.f32 	%f45, %f22, %f44;\n"
"	.loc	15	322	0\n"
"	mad.f32 	%f20, %f28, %f45, %f20;\n"
"	.loc	15	323	0\n"
"	mad.f32 	%f19, %f27, %f45, %f19;\n"
"	.loc	15	324	0\n"
"	mad.f32 	%f18, %f29, %f45, %f18;\n"
"	@!%p4 bra 	$Lt_1_11266;\n"
"	.loc	15	328	0\n"
"	ld.param.u64 	%rd34, [__cudaparm_kernel_lj_lj3];\n"
"	add.u64 	%rd35, %rd34, %rd32;\n"
"	ld.global.v4.f32 	{%f46,%f47,%f48,_}, [%rd35+0];\n"
"	mul.f32 	%f49, %f46, %f38;\n"
"	sub.f32 	%f50, %f49, %f47;\n"
"	mul.f32 	%f51, %f38, %f50;\n"
"	sub.f32 	%f52, %f51, %f48;\n"
"	.loc	15	303	0\n"
"	ld.shared.f32 	%f22, [%rd27+0];\n"
"	.loc	15	328	0\n"
"	mad.f32 	%f21, %f22, %f52, %f21;\n"
"$Lt_1_11266:\n"
"	@!%p5 bra 	$Lt_1_13826;\n"
"	.loc	15	331	0\n"
"	mov.f32 	%f53, %f3;\n"
"	mul.f32 	%f54, %f28, %f28;\n"
"	mad.f32 	%f55, %f45, %f54, %f53;\n"
"	mov.f32 	%f3, %f55;\n"
"	.loc	15	332	0\n"
"	mov.f32 	%f56, %f5;\n"
"	mad.f32 	%f57, %f45, %f30, %f56;\n"
"	mov.f32 	%f5, %f57;\n"
"	.loc	15	333	0\n"
"	mov.f32 	%f58, %f7;\n"
"	mul.f32 	%f59, %f29, %f29;\n"
"	mad.f32 	%f60, %f45, %f59, %f58;\n"
"	mov.f32 	%f7, %f60;\n"
"	.loc	15	334	0\n"
"	mov.f32 	%f61, %f9;\n"
"	mul.f32 	%f62, %f27, %f28;\n"
"	mad.f32 	%f63, %f45, %f62, %f61;\n"
"	mov.f32 	%f9, %f63;\n"
"	.loc	15	335	0\n"
"	mov.f32 	%f64, %f11;\n"
"	mul.f32 	%f65, %f28, %f29;\n"
"	mad.f32 	%f66, %f45, %f65, %f64;\n"
"	mov.f32 	%f11, %f66;\n"
"	.loc	15	336	0\n"
"	mul.f32 	%f67, %f27, %f29;\n"
"	mad.f32 	%f12, %f45, %f67, %f12;\n"
"	mov.f32 	%f68, %f12;\n"
"$Lt_1_13826:\n"
"$L_1_8962:\n"
"	.loc	15	300	0\n"
"	add.u64 	%rd15, %rd12, %rd15;\n"
"	setp.gt.u64 	%p8, %rd22, %rd15;\n"
"	@%p8 bra 	$Lt_1_11010;\n"
"	bra.uni 	$Lt_1_10498;\n"
"$Lt_1_13570:\n"
"	mov.s32 	%r26, 0;\n"
"	setp.gt.s32 	%p4, %r15, %r26;\n"
"	mov.s32 	%r27, 0;\n"
"	setp.gt.s32 	%p5, %r14, %r27;\n"
"$Lt_1_10498:\n"
"	.loc	15	343	0\n"
"	ld.param.u64 	%rd36, [__cudaparm_kernel_lj_engv];\n"
"	add.u64 	%rd37, %rd36, %rd8;\n"
"	@!%p4 bra 	$Lt_1_12546;\n"
"	.loc	15	345	0\n"
"	ld.global.f32 	%f69, [%rd37+0];\n"
"	add.f32 	%f70, %f69, %f21;\n"
"	st.global.f32 	[%rd37+0], %f70;\n"
"	.loc	15	346	0\n"
"	cvt.u64.s32 	%rd38, %r9;\n"
"	mul.lo.u64 	%rd39, %rd38, 4;\n"
"	add.u64 	%rd37, %rd37, %rd39;\n"
"$Lt_1_12546:\n"
"	@!%p5 bra 	$Lt_1_13058;\n"
"	.loc	15	350	0\n"
"	ld.global.f32 	%f71, [%rd37+0];\n"
"	mov.f32 	%f72, %f3;\n"
"	add.f32 	%f73, %f71, %f72;\n"
"	st.global.f32 	[%rd37+0], %f73;\n"
"	.loc	15	351	0\n"
"	cvt.u64.s32 	%rd40, %r9;\n"
"	mul.lo.u64 	%rd41, %rd40, 4;\n"
"	add.u64 	%rd37, %rd41, %rd37;\n"
"	.loc	15	350	0\n"
"	ld.global.f32 	%f74, [%rd37+0];\n"
"	mov.f32 	%f75, %f5;\n"
"	add.f32 	%f76, %f74, %f75;\n"
"	st.global.f32 	[%rd37+0], %f76;\n"
"	.loc	15	351	0\n"
"	add.u64 	%rd37, %rd41, %rd37;\n"
"	.loc	15	350	0\n"
"	ld.global.f32 	%f77, [%rd37+0];\n"
"	mov.f32 	%f78, %f7;\n"
"	add.f32 	%f79, %f77, %f78;\n"
"	st.global.f32 	[%rd37+0], %f79;\n"
"	.loc	15	351	0\n"
"	add.u64 	%rd37, %rd41, %rd37;\n"
"	.loc	15	350	0\n"
"	ld.global.f32 	%f80, [%rd37+0];\n"
"	mov.f32 	%f81, %f9;\n"
"	add.f32 	%f82, %f80, %f81;\n"
"	st.global.f32 	[%rd37+0], %f82;\n"
"	.loc	15	351	0\n"
"	add.u64 	%rd37, %rd41, %rd37;\n"
"	.loc	15	350	0\n"
"	ld.global.f32 	%f83, [%rd37+0];\n"
"	mov.f32 	%f84, %f11;\n"
"	add.f32 	%f85, %f83, %f84;\n"
"	st.global.f32 	[%rd37+0], %f85;\n"
"	.loc	15	351	0\n"
"	add.u64 	%rd37, %rd41, %rd37;\n"
"	.loc	15	350	0\n"
"	ld.global.f32 	%f86, [%rd37+0];\n"
"	add.f32 	%f87, %f86, %f12;\n"
"	st.global.f32 	[%rd37+0], %f87;\n"
"$Lt_1_13058:\n"
"	.loc	15	354	0\n"
"	ld.param.u64 	%rd42, [__cudaparm_kernel_lj_ans];\n"
"	mul.lo.u64 	%rd43, %rd7, 16;\n"
"	add.u64 	%rd44, %rd42, %rd43;\n"
"	ld.global.v4.f32 	{%f88,%f89,%f90,%f91}, [%rd44+0];\n"
"	.loc	15	355	0\n"
"	add.f32 	%f88, %f88, %f20;\n"
"	.loc	15	356	0\n"
"	add.f32 	%f89, %f89, %f19;\n"
"	.loc	15	357	0\n"
"	add.f32 	%f90, %f90, %f18;\n"
"	st.global.v4.f32 	[%rd44+0], {%f88,%f89,%f90,%f91};\n"
"$Lt_1_9986:\n"
"	.loc	15	360	0\n"
"	exit;\n"
"$LDWend_kernel_lj:\n"
"	}\n"
"	.entry kernel_lj_fast (\n"
"		.param .u64 __cudaparm_kernel_lj_fast_x_,\n"
"		.param .u64 __cudaparm_kernel_lj_fast_lj1_in,\n"
"		.param .u64 __cudaparm_kernel_lj_fast_lj3_in,\n"
"		.param .u64 __cudaparm_kernel_lj_fast_gum,\n"
"		.param .s32 __cudaparm_kernel_lj_fast_stride,\n"
"		.param .u64 __cudaparm_kernel_lj_fast_dev_ij,\n"
"		.param .u64 __cudaparm_kernel_lj_fast_ans,\n"
"		.param .u64 __cudaparm_kernel_lj_fast_engv,\n"
"		.param .u64 __cudaparm_kernel_lj_fast_err_flag,\n"
"		.param .s32 __cudaparm_kernel_lj_fast_eflag,\n"
"		.param .s32 __cudaparm_kernel_lj_fast_vflag,\n"
"		.param .s32 __cudaparm_kernel_lj_fast_start,\n"
"		.param .s32 __cudaparm_kernel_lj_fast_inum,\n"
"		.param .s32 __cudaparm_kernel_lj_fast_nall)\n"
"	{\n"
"	.reg .u32 %r<31>;\n"
"	.reg .u64 %rd<54>;\n"
"	.reg .f32 %f<103>;\n"
"	.reg .pred %p<12>;\n"
"	.shared .align 4 .b8 __cuda_sp_lj356[16];\n"
"	.shared .align 16 .b8 __cuda_lj1384[1024];\n"
"	.shared .align 16 .b8 __cuda_lj31408[1024];\n"
"	.loc	15	368	0\n"
"$LBB1_kernel_lj_fast:\n"
"	cvt.s32.u16 	%r1, %tid.x;\n"
"	mov.u32 	%r2, 3;\n"
"	setp.gt.s32 	%p1, %r1, %r2;\n"
"	@%p1 bra 	$Lt_2_11010;\n"
"	.loc	15	375	0\n"
"	mov.u64 	%rd1, __cuda_sp_lj356;\n"
"	cvt.u64.s32 	%rd2, %r1;\n"
"	mul.lo.u64 	%rd3, %rd2, 4;\n"
"	ld.param.u64 	%rd4, [__cudaparm_kernel_lj_fast_gum];\n"
"	add.u64 	%rd5, %rd4, %rd3;\n"
"	ld.global.f32 	%f1, [%rd5+12];\n"
"	add.u64 	%rd6, %rd3, %rd1;\n"
"	st.shared.f32 	[%rd6+0], %f1;\n"
"$Lt_2_11010:\n"
"	mov.u64 	%rd1, __cuda_sp_lj356;\n"
"	mov.u32 	%r3, 63;\n"
"	setp.gt.s32 	%p2, %r1, %r3;\n"
"	@%p2 bra 	$Lt_2_11522;\n"
"	.loc	15	377	0\n"
"	mov.u64 	%rd7, __cuda_lj1384;\n"
"	cvt.u64.s32 	%rd8, %r1;\n"
"	mul.lo.u64 	%rd9, %rd8, 16;\n"
"	ld.param.u64 	%rd10, [__cudaparm_kernel_lj_fast_lj1_in];\n"
"	add.u64 	%rd11, %rd10, %rd9;\n"
"	add.u64 	%rd12, %rd9, %rd7;\n"
"	ld.global.v4.f32 	{%f2,%f3,%f4,%f5}, [%rd11+0];\n"
"	st.shared.f32 	[%rd12+0], %f2;\n"
"	st.shared.f32 	[%rd12+4], %f3;\n"
"	st.shared.f32 	[%rd12+8], %f4;\n"
"	st.shared.f32 	[%rd12+12], %f5;\n"
"	ld.param.s32 	%r4, [__cudaparm_kernel_lj_fast_eflag];\n"
"	mov.u32 	%r5, 0;\n"
"	setp.le.s32 	%p3, %r4, %r5;\n"
"	@%p3 bra 	$Lt_2_12034;\n"
"	.loc	15	379	0\n"
"	mov.u64 	%rd13, __cuda_lj31408;\n"
"	ld.param.u64 	%rd14, [__cudaparm_kernel_lj_fast_lj3_in];\n"
"	add.u64 	%rd15, %rd14, %rd9;\n"
"	add.u64 	%rd16, %rd9, %rd13;\n"
"	ld.global.v4.f32 	{%f6,%f7,%f8,%f9}, [%rd15+0];\n"
"	st.shared.f32 	[%rd16+0], %f6;\n"
"	st.shared.f32 	[%rd16+4], %f7;\n"
"	st.shared.f32 	[%rd16+8], %f8;\n"
"	st.shared.f32 	[%rd16+12], %f9;\n"
"$Lt_2_12034:\n"
"	mov.u64 	%rd13, __cuda_lj31408;\n"
"$Lt_2_11522:\n"
"	mov.u64 	%rd7, __cuda_lj1384;\n"
"	mov.u64 	%rd13, __cuda_lj31408;\n"
"	.loc	15	382	0\n"
"	bar.sync 	0;\n"
"	cvt.s32.u16 	%r6, %ctaid.x;\n"
"	cvt.s32.u16 	%r7, %ntid.x;\n"
"	mul24.lo.s32 	%r8, %r6, %r7;\n"
"	ld.param.s32 	%r9, [__cudaparm_kernel_lj_fast_start];\n"
"	add.s32 	%r10, %r9, %r8;\n"
"	add.s32 	%r11, %r10, %r1;\n"
"	ld.param.s32 	%r12, [__cudaparm_kernel_lj_fast_inum];\n"
"	setp.ge.s32 	%p4, %r11, %r12;\n"
"	@%p4 bra 	$Lt_2_12546;\n"
"	.loc	15	393	0\n"
"	mov.f32 	%f10, 0f00000000;    	\n"
"	mov.f32 	%f11, %f10;\n"
"	mov.f32 	%f12, 0f00000000;    	\n"
"	mov.f32 	%f13, %f12;\n"
"	mov.f32 	%f14, 0f00000000;    	\n"
"	mov.f32 	%f15, %f14;\n"
"	mov.f32 	%f16, 0f00000000;    	\n"
"	mov.f32 	%f17, %f16;\n"
"	mov.f32 	%f18, 0f00000000;    	\n"
"	mov.f32 	%f19, %f18;\n"
"	mov.f32 	%f20, 0f00000000;    	\n"
"	mov.f32 	%f21, %f20;\n"
"	.loc	15	396	0\n"
"	cvt.u64.s32 	%rd17, %r11;\n"
"	mul.lo.u64 	%rd18, %rd17, 4;\n"
"	ld.param.u64 	%rd19, [__cudaparm_kernel_lj_fast_dev_ij];\n"
"	add.u64 	%rd20, %rd19, %rd18;\n"
"	ld.global.s32 	%r13, [%rd20+0];\n"
"	.loc	15	398	0\n"
"	ld.param.s32 	%r14, [__cudaparm_kernel_lj_fast_stride];\n"
"	cvt.u64.s32 	%rd21, %r14;\n"
"	mul.lo.u64 	%rd22, %rd21, 4;\n"
"	add.u64 	%rd23, %rd20, %rd22;\n"
"	ld.global.s32 	%r15, [%rd23+0];\n"
"	.loc	15	399	0\n"
"	add.u64 	%rd24, %rd23, %rd22;\n"
"	mov.s64 	%rd25, %rd24;\n"
"	.loc	15	402	0\n"
"	ld.param.u64 	%rd26, [__cudaparm_kernel_lj_fast_x_];\n"
"	cvt.u64.s32 	%rd27, %r13;\n"
"	mul.lo.u64 	%rd28, %rd27, 16;\n"
"	add.u64 	%rd29, %rd26, %rd28;\n"
"	ld.global.v4.f32 	{%f22,%f23,%f24,%f25}, [%rd29+0];\n"
"	mul24.lo.s32 	%r16, %r14, %r15;\n"
"	cvt.s64.s32 	%rd30, %r16;\n"
"	mul.lo.u64 	%rd31, %rd30, 4;\n"
"	add.u64 	%rd32, %rd24, %rd31;\n"
"	ld.param.s32 	%r17, [__cudaparm_kernel_lj_fast_vflag];\n"
"	ld.param.s32 	%r18, [__cudaparm_kernel_lj_fast_eflag];\n"
"	setp.ge.u64 	%p5, %rd24, %rd32;\n"
"	mov.f32 	%f26, 0f00000000;    	\n"
"	mov.f32 	%f27, 0f00000000;    	\n"
"	mov.f32 	%f28, 0f00000000;    	\n"
"	mov.f32 	%f29, 0f00000000;    	\n"
"	@%p5 bra 	$Lt_2_16130;\n"
"	mov.s32 	%r19, 0;\n"
"	setp.gt.s32 	%p6, %r18, %r19;\n"
"	mov.s32 	%r20, 0;\n"
"	setp.gt.s32 	%p7, %r17, %r20;\n"
"	cvt.rzi.s32.f32 	%r21, %f25;\n"
"	mov.s32 	%r22, 8;\n"
"	mul24.lo.s32 	%r23, %r22, %r21;\n"
"	cvt.rn.f32.s32 	%f30, %r23;\n"
"$Lt_2_13570:\n"
"	.loc	15	409	0\n"
"	ld.global.s32 	%r24, [%rd25+0];\n"
"	.loc	15	410	0\n"
"	shr.s32 	%r25, %r24, 30;\n"
"	cvt.s64.s32 	%rd33, %r25;\n"
"	and.b64 	%rd34, %rd33, 3;\n"
"	mul.lo.u64 	%rd35, %rd34, 4;\n"
"	add.u64 	%rd36, %rd1, %rd35;\n"
"	ld.shared.f32 	%f31, [%rd36+0];\n"
"	.loc	15	413	0\n"
"	and.b32 	%r26, %r24, 1073741823;\n"
"	cvt.u64.s32 	%rd37, %r26;\n"
"	mul.lo.u64 	%rd38, %rd37, 16;\n"
"	add.u64 	%rd39, %rd26, %rd38;\n"
"	ld.global.v4.f32 	{%f32,%f33,%f34,%f35}, [%rd39+0];\n"
"	.loc	15	409	0\n"
"	sub.f32 	%f36, %f23, %f33;\n"
"	sub.f32 	%f37, %f22, %f32;\n"
"	sub.f32 	%f38, %f24, %f34;\n"
"	mul.f32 	%f39, %f36, %f36;\n"
"	mad.f32 	%f40, %f37, %f37, %f39;\n"
"	mad.f32 	%f41, %f38, %f38, %f40;\n"
"	add.f32 	%f42, %f30, %f35;\n"
"	cvt.rzi.s32.f32 	%r27, %f42;\n"
"	cvt.u64.s32 	%rd40, %r27;\n"
"	mul.lo.u64 	%rd41, %rd40, 16;\n"
"	add.u64 	%rd42, %rd41, %rd7;\n"
"	ld.shared.f32 	%f43, [%rd42+8];\n"
"	setp.gt.f32 	%p8, %f43, %f41;\n"
"	@!%p8 bra 	$Lt_2_16386;\n"
"	ld.shared.f32 	%f44, [%rd42+12];\n"
"	mov.f32 	%f45, 0f00000000;    	\n"
"	setp.eq.f32 	%p9, %f44, %f45;\n"
"	@!%p9 bra 	$Lt_2_16386;\n"
"	.loc	15	425	0\n"
"	rcp.approx.f32 	%f46, %f41;\n"
"	mul.f32 	%f47, %f46, %f46;\n"
"	mul.f32 	%f48, %f46, %f47;\n"
"	.loc	15	410	0\n"
"	ld.shared.f32 	%f31, [%rd36+0];\n"
"	.loc	15	425	0\n"
"	mul.f32 	%f49, %f46, %f31;\n"
"	mul.f32 	%f50, %f48, %f49;\n"
"	ld.shared.f32 	%f51, [%rd42+4];\n"
"	ld.shared.f32 	%f52, [%rd42+0];\n"
"	mul.f32 	%f53, %f52, %f48;\n"
"	sub.f32 	%f54, %f53, %f51;\n"
"	mul.f32 	%f55, %f50, %f54;\n"
"	.loc	15	427	0\n"
"	mad.f32 	%f28, %f37, %f55, %f28;\n"
"	.loc	15	428	0\n"
"	mad.f32 	%f27, %f36, %f55, %f27;\n"
"	.loc	15	429	0\n"
"	mad.f32 	%f26, %f38, %f55, %f26;\n"
"	@!%p6 bra 	$Lt_2_13826;\n"
"	.loc	15	432	0\n"
"	add.u64 	%rd43, %rd41, %rd13;\n"
"	ld.shared.f32 	%f56, [%rd43+4];\n"
"	ld.shared.f32 	%f57, [%rd43+0];\n"
"	mul.f32 	%f58, %f57, %f48;\n"
"	sub.f32 	%f59, %f58, %f56;\n"
"	mul.f32 	%f60, %f48, %f59;\n"
"	.loc	15	433	0\n"
"	ld.shared.f32 	%f61, [%rd43+8];\n"
"	sub.f32 	%f62, %f60, %f61;\n"
"	.loc	15	410	0\n"
"	ld.shared.f32 	%f31, [%rd36+0];\n"
"	.loc	15	433	0\n"
"	mad.f32 	%f29, %f31, %f62, %f29;\n"
"$Lt_2_13826:\n"
"	@!%p7 bra 	$Lt_2_16386;\n"
"	.loc	15	436	0\n"
"	mov.f32 	%f63, %f11;\n"
"	mul.f32 	%f64, %f37, %f37;\n"
"	mad.f32 	%f65, %f55, %f64, %f63;\n"
"	mov.f32 	%f11, %f65;\n"
"	.loc	15	437	0\n"
"	mov.f32 	%f66, %f13;\n"
"	mad.f32 	%f67, %f55, %f39, %f66;\n"
"	mov.f32 	%f13, %f67;\n"
"	.loc	15	438	0\n"
"	mov.f32 	%f68, %f15;\n"
"	mul.f32 	%f69, %f38, %f38;\n"
"	mad.f32 	%f70, %f55, %f69, %f68;\n"
"	mov.f32 	%f15, %f70;\n"
"	.loc	15	439	0\n"
"	mov.f32 	%f71, %f17;\n"
"	mul.f32 	%f72, %f36, %f37;\n"
"	mad.f32 	%f73, %f55, %f72, %f71;\n"
"	mov.f32 	%f17, %f73;\n"
"	.loc	15	440	0\n"
"	mov.f32 	%f74, %f19;\n"
"	mul.f32 	%f75, %f37, %f38;\n"
"	mad.f32 	%f76, %f55, %f75, %f74;\n"
"	mov.f32 	%f19, %f76;\n"
"	.loc	15	441	0\n"
"	mul.f32 	%f77, %f36, %f38;\n"
"	mad.f32 	%f20, %f55, %f77, %f20;\n"
"	mov.f32 	%f78, %f20;\n"
"$Lt_2_16386:\n"
"$L_2_10498:\n"
"	.loc	15	407	0\n"
"	add.u64 	%rd25, %rd22, %rd25;\n"
"	setp.gt.u64 	%p10, %rd32, %rd25;\n"
"	@%p10 bra 	$Lt_2_13570;\n"
"	bra.uni 	$Lt_2_13058;\n"
"$Lt_2_16130:\n"
"	mov.s32 	%r28, 0;\n"
"	setp.gt.s32 	%p6, %r18, %r28;\n"
"	mov.s32 	%r29, 0;\n"
"	setp.gt.s32 	%p7, %r17, %r29;\n"
"$Lt_2_13058:\n"
"	.loc	15	448	0\n"
"	ld.param.u64 	%rd44, [__cudaparm_kernel_lj_fast_engv];\n"
"	add.u64 	%rd45, %rd44, %rd18;\n"
"	@!%p6 bra 	$Lt_2_15106;\n"
"	.loc	15	450	0\n"
"	ld.global.f32 	%f79, [%rd45+0];\n"
"	add.f32 	%f80, %f79, %f29;\n"
"	st.global.f32 	[%rd45+0], %f80;\n"
"	.loc	15	451	0\n"
"	cvt.u64.s32 	%rd46, %r12;\n"
"	mul.lo.u64 	%rd47, %rd46, 4;\n"
"	add.u64 	%rd45, %rd45, %rd47;\n"
"$Lt_2_15106:\n"
"	@!%p7 bra 	$Lt_2_15618;\n"
"	.loc	15	455	0\n"
"	ld.global.f32 	%f81, [%rd45+0];\n"
"	mov.f32 	%f82, %f11;\n"
"	add.f32 	%f83, %f81, %f82;\n"
"	st.global.f32 	[%rd45+0], %f83;\n"
"	.loc	15	456	0\n"
"	cvt.u64.s32 	%rd48, %r12;\n"
"	mul.lo.u64 	%rd49, %rd48, 4;\n"
"	add.u64 	%rd45, %rd49, %rd45;\n"
"	.loc	15	455	0\n"
"	ld.global.f32 	%f84, [%rd45+0];\n"
"	mov.f32 	%f85, %f13;\n"
"	add.f32 	%f86, %f84, %f85;\n"
"	st.global.f32 	[%rd45+0], %f86;\n"
"	.loc	15	456	0\n"
"	add.u64 	%rd45, %rd49, %rd45;\n"
"	.loc	15	455	0\n"
"	ld.global.f32 	%f87, [%rd45+0];\n"
"	mov.f32 	%f88, %f15;\n"
"	add.f32 	%f89, %f87, %f88;\n"
"	st.global.f32 	[%rd45+0], %f89;\n"
"	.loc	15	456	0\n"
"	add.u64 	%rd45, %rd49, %rd45;\n"
"	.loc	15	455	0\n"
"	ld.global.f32 	%f90, [%rd45+0];\n"
"	mov.f32 	%f91, %f17;\n"
"	add.f32 	%f92, %f90, %f91;\n"
"	st.global.f32 	[%rd45+0], %f92;\n"
"	.loc	15	456	0\n"
"	add.u64 	%rd45, %rd49, %rd45;\n"
"	.loc	15	455	0\n"
"	ld.global.f32 	%f93, [%rd45+0];\n"
"	mov.f32 	%f94, %f19;\n"
"	add.f32 	%f95, %f93, %f94;\n"
"	st.global.f32 	[%rd45+0], %f95;\n"
"	.loc	15	456	0\n"
"	add.u64 	%rd45, %rd49, %rd45;\n"
"	.loc	15	455	0\n"
"	ld.global.f32 	%f96, [%rd45+0];\n"
"	add.f32 	%f97, %f96, %f20;\n"
"	st.global.f32 	[%rd45+0], %f97;\n"
"$Lt_2_15618:\n"
"	.loc	15	459	0\n"
"	ld.param.u64 	%rd50, [__cudaparm_kernel_lj_fast_ans];\n"
"	mul.lo.u64 	%rd51, %rd17, 16;\n"
"	add.u64 	%rd52, %rd50, %rd51;\n"
"	ld.global.v4.f32 	{%f98,%f99,%f100,%f101}, [%rd52+0];\n"
"	.loc	15	460	0\n"
"	add.f32 	%f98, %f98, %f28;\n"
"	.loc	15	461	0\n"
"	add.f32 	%f99, %f99, %f27;\n"
"	.loc	15	462	0\n"
"	add.f32 	%f100, %f100, %f26;\n"
"	st.global.v4.f32 	[%rd52+0], {%f98,%f99,%f100,%f101};\n"
"$Lt_2_12546:\n"
"	.loc	15	465	0\n"
"	exit;\n"
"$LDWend_kernel_lj_fast:\n"
"	}\n"
;