//
// Generated by NVIDIA NVVM Compiler
//
// Compiler Build ID: CL-23083092
// Cuda compilation tools, release 9.1, V9.1.85
// Based on LLVM 3.4svn
//

.version 6.1
.target sm_30
.address_size 64

	// .globl	_Z6oxMainv
.global .align 8 .b8 pixelID[8];
.global .align 8 .b8 resolution[8];
.global .align 4 .b8 normal[12];
.global .align 4 .b8 camPos[12];
.global .align 4 .b8 root[4];
.global .align 4 .u32 imageEnabled;
.global .texref lightmap;
.global .align 16 .b8 tileInfo[16];
.global .align 4 .u32 additive;
.global .align 1 .b8 image[1];
.global .align 1 .b8 image_HDR[1];
.global .align 1 .b8 image_HDR2[1];
.global .align 1 .b8 image_Mask[1];
.global .align 1 .b8 image_RNM0[1];
.global .align 1 .b8 image_RNM1[1];
.global .align 1 .b8 image_RNM2[1];
.global .align 1 .b8 image_RNM3[1];
.global .align 1 .b8 uvpos[1];
.global .align 1 .b8 uvnormal[1];
.global .align 4 .u32 ignoreNormal;
.global .align 1 .b8 localLights[1];
.global .align 4 .b8 _ZN21rti_internal_typeinfo7pixelIDE[8] = {82, 97, 121, 0, 8, 0, 0, 0};
.global .align 4 .b8 _ZN21rti_internal_typeinfo10resolutionE[8] = {82, 97, 121, 0, 8, 0, 0, 0};
.global .align 4 .b8 _ZN21rti_internal_typeinfo6normalE[8] = {82, 97, 121, 0, 12, 0, 0, 0};
.global .align 4 .b8 _ZN21rti_internal_typeinfo6camPosE[8] = {82, 97, 121, 0, 12, 0, 0, 0};
.global .align 4 .b8 _ZN21rti_internal_typeinfo4rootE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
.global .align 4 .b8 _ZN21rti_internal_typeinfo12imageEnabledE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
.global .align 4 .b8 _ZN21rti_internal_typeinfo8tileInfoE[8] = {82, 97, 121, 0, 16, 0, 0, 0};
.global .align 4 .b8 _ZN21rti_internal_typeinfo8additiveE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
.global .align 4 .b8 _ZN21rti_internal_typeinfo12ignoreNormalE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
.global .align 8 .u64 _ZN21rti_internal_register20reg_bitness_detectorE;
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail0E;
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail1E;
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail2E;
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail3E;
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail4E;
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail5E;
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail6E;
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail7E;
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail8E;
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail9E;
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail0E;
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail1E;
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail2E;
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail3E;
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail4E;
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail5E;
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail6E;
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail7E;
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail8E;
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail9E;
.global .align 4 .u32 _ZN21rti_internal_register14reg_rayIndex_xE;
.global .align 4 .u32 _ZN21rti_internal_register14reg_rayIndex_yE;
.global .align 4 .u32 _ZN21rti_internal_register14reg_rayIndex_zE;
.global .align 8 .b8 _ZN21rti_internal_typename7pixelIDE[6] = {117, 105, 110, 116, 50, 0};
.global .align 8 .b8 _ZN21rti_internal_typename10resolutionE[6] = {117, 105, 110, 116, 50, 0};
.global .align 8 .b8 _ZN21rti_internal_typename6normalE[7] = {102, 108, 111, 97, 116, 51, 0};
.global .align 8 .b8 _ZN21rti_internal_typename6camPosE[7] = {102, 108, 111, 97, 116, 51, 0};
.global .align 16 .b8 _ZN21rti_internal_typename4rootE[9] = {114, 116, 79, 98, 106, 101, 99, 116, 0};
.global .align 4 .b8 _ZN21rti_internal_typename12imageEnabledE[4] = {105, 110, 116, 0};
.global .align 8 .b8 _ZN21rti_internal_typename8tileInfoE[6] = {117, 105, 110, 116, 52, 0};
.global .align 4 .b8 _ZN21rti_internal_typename8additiveE[4] = {105, 110, 116, 0};
.global .align 4 .b8 _ZN21rti_internal_typename12ignoreNormalE[4] = {105, 110, 116, 0};
.global .align 4 .u32 _ZN21rti_internal_typeenum7pixelIDE = 4919;
.global .align 4 .u32 _ZN21rti_internal_typeenum10resolutionE = 4919;
.global .align 4 .u32 _ZN21rti_internal_typeenum6normalE = 4919;
.global .align 4 .u32 _ZN21rti_internal_typeenum6camPosE = 4919;
.global .align 4 .u32 _ZN21rti_internal_typeenum4rootE = 4919;
.global .align 4 .u32 _ZN21rti_internal_typeenum12imageEnabledE = 4919;
.global .align 4 .u32 _ZN21rti_internal_typeenum8tileInfoE = 4919;
.global .align 4 .u32 _ZN21rti_internal_typeenum8additiveE = 4919;
.global .align 4 .u32 _ZN21rti_internal_typeenum12ignoreNormalE = 4919;
.global .align 16 .b8 _ZN21rti_internal_semantic7pixelIDE[14] = {114, 116, 76, 97, 117, 110, 99, 104, 73, 110, 100, 101, 120, 0};
.global .align 16 .b8 _ZN21rti_internal_semantic10resolutionE[12] = {114, 116, 76, 97, 117, 110, 99, 104, 68, 105, 109, 0};
.global .align 16 .b8 _ZN21rti_internal_semantic6normalE[17] = {97, 116, 116, 114, 105, 98, 117, 116, 101, 32, 110, 111, 114, 109, 97, 108, 0};
.global .align 1 .b8 _ZN21rti_internal_semantic6camPosE[1];
.global .align 1 .b8 _ZN21rti_internal_semantic4rootE[1];
.global .align 1 .b8 _ZN21rti_internal_semantic12imageEnabledE[1];
.global .align 1 .b8 _ZN21rti_internal_semantic8tileInfoE[1];
.global .align 1 .b8 _ZN21rti_internal_semantic8additiveE[1];
.global .align 1 .b8 _ZN21rti_internal_semantic12ignoreNormalE[1];
.global .align 1 .b8 _ZN23rti_internal_annotation7pixelIDE[1];
.global .align 1 .b8 _ZN23rti_internal_annotation10resolutionE[1];
.global .align 1 .b8 _ZN23rti_internal_annotation6normalE[1];
.global .align 1 .b8 _ZN23rti_internal_annotation6camPosE[1];
.global .align 1 .b8 _ZN23rti_internal_annotation4rootE[1];
.global .align 1 .b8 _ZN23rti_internal_annotation12imageEnabledE[1];
.global .align 1 .b8 _ZN23rti_internal_annotation8tileInfoE[1];
.global .align 1 .b8 _ZN23rti_internal_annotation8additiveE[1];
.global .align 1 .b8 _ZN23rti_internal_annotation12ignoreNormalE[1];

.visible .entry _Z6oxMainv(

)
{
	.local .align 4 .b8 	__local_depot0[4];
	.reg .b64 	%SP;
	.reg .b64 	%SPL;
	.reg .pred 	%p<138>;
	.reg .b16 	%rs<162>;
	.reg .f32 	%f<1453>;
	.reg .b32 	%r<254>;
	.reg .b64 	%rd<272>;


	mov.u64 	%rd271, __local_depot0;
	cvta.local.u64 	%SP, %rd271;
	ld.global.v2.u32 	{%r30, %r31}, [pixelID];
	cvt.u64.u32	%rd10, %r30;
	cvt.u64.u32	%rd11, %r31;
	mov.u64 	%rd14, uvnormal;
	cvta.global.u64 	%rd9, %rd14;
	mov.u32 	%r28, 2;
	mov.u32 	%r29, 4;
	mov.u64 	%rd13, 0;
	// inline asm
	call (%rd8), _rt_buffer_get_64, (%rd9, %r28, %r29, %rd10, %rd11, %rd13, %rd13);
	// inline asm
	ld.u32 	%r1, [%rd8];
	shr.u32 	%r34, %r1, 16;
	cvt.u16.u32	%rs1, %r34;
	and.b16  	%rs9, %rs1, 255;
	cvt.u16.u32	%rs10, %r1;
	or.b16  	%rs11, %rs10, %rs9;
	setp.eq.s16	%p8, %rs11, 0;
	mov.f32 	%f1357, 0f00000000;
	mov.f32 	%f1358, %f1357;
	mov.f32 	%f1359, %f1357;
	@%p8 bra 	BB0_2;

	ld.u8 	%rs12, [%rd8+1];
	and.b16  	%rs14, %rs10, 255;
	cvt.rn.f32.u16	%f279, %rs14;
	div.rn.f32 	%f280, %f279, 0f437F0000;
	fma.rn.f32 	%f281, %f280, 0f40000000, 0fBF800000;
	cvt.rn.f32.u16	%f282, %rs12;
	div.rn.f32 	%f283, %f282, 0f437F0000;
	fma.rn.f32 	%f284, %f283, 0f40000000, 0fBF800000;
	cvt.rn.f32.u16	%f285, %rs9;
	div.rn.f32 	%f286, %f285, 0f437F0000;
	fma.rn.f32 	%f287, %f286, 0f40000000, 0fBF800000;
	mul.f32 	%f288, %f284, %f284;
	fma.rn.f32 	%f289, %f281, %f281, %f288;
	fma.rn.f32 	%f290, %f287, %f287, %f289;
	sqrt.rn.f32 	%f291, %f290;
	rcp.rn.f32 	%f292, %f291;
	mul.f32 	%f1357, %f281, %f292;
	mul.f32 	%f1358, %f284, %f292;
	mul.f32 	%f1359, %f287, %f292;

BB0_2:
	ld.global.v2.u32 	{%r35, %r36}, [pixelID];
	ld.global.v2.u32 	{%r38, %r39}, [tileInfo];
	add.s32 	%r2, %r35, %r38;
	add.s32 	%r3, %r36, %r39;
	setp.eq.f32	%p9, %f1358, 0f00000000;
	setp.eq.f32	%p10, %f1357, 0f00000000;
	and.pred  	%p11, %p10, %p9;
	setp.eq.f32	%p12, %f1359, 0f00000000;
	and.pred  	%p13, %p11, %p12;
	@%p13 bra 	BB0_111;
	bra.uni 	BB0_3;

BB0_111:
	ld.global.u32 	%r253, [imageEnabled];
	and.b32  	%r202, %r253, 1;
	setp.eq.b32	%p129, %r202, 1;
	@!%p129 bra 	BB0_113;
	bra.uni 	BB0_112;

BB0_112:
	cvt.u64.u32	%rd159, %r2;
	cvt.u64.u32	%rd160, %r3;
	mov.u64 	%rd163, image;
	cvta.global.u64 	%rd158, %rd163;
	// inline asm
	call (%rd157), _rt_buffer_get_64, (%rd158, %r28, %r29, %rd159, %rd160, %rd13, %rd13);
	// inline asm
	mov.u16 	%rs93, 0;
	st.v4.u8 	[%rd157], {%rs93, %rs93, %rs93, %rs93};
	ld.global.u32 	%r253, [imageEnabled];

BB0_113:
	and.b32  	%r205, %r253, 8;
	setp.eq.s32	%p130, %r205, 0;
	@%p130 bra 	BB0_115;

	cvt.u64.u32	%rd167, %r3;
	cvt.u64.u32	%rd166, %r2;
	mov.u64 	%rd170, image_Mask;
	cvta.global.u64 	%rd165, %rd170;
	// inline asm
	call (%rd164), _rt_buffer_get_64, (%rd165, %r28, %r28, %rd166, %rd167, %rd13, %rd13);
	// inline asm
	mov.f32 	%f1318, 0f00000000;
	cvt.rzi.u32.f32	%r208, %f1318;
	cvt.u16.u32	%rs94, %r208;
	mov.u16 	%rs95, 0;
	st.v2.u8 	[%rd164], {%rs94, %rs95};
	ld.global.u32 	%r253, [imageEnabled];

BB0_115:
	cvt.u64.u32	%rd6, %r2;
	cvt.u64.u32	%rd7, %r3;
	and.b32  	%r209, %r253, 4;
	setp.eq.s32	%p131, %r209, 0;
	@%p131 bra 	BB0_119;

	ld.global.u32 	%r210, [additive];
	setp.eq.s32	%p132, %r210, 0;
	@%p132 bra 	BB0_118;

	mov.u64 	%rd183, image_HDR;
	cvta.global.u64 	%rd172, %rd183;
	mov.u32 	%r214, 8;
	// inline asm
	call (%rd171), _rt_buffer_get_64, (%rd172, %r28, %r214, %rd6, %rd7, %rd13, %rd13);
	// inline asm
	ld.v4.u16 	{%rs102, %rs103, %rs104, %rs105}, [%rd171];
	// inline asm
	{  cvt.f32.f16 %f1319, %rs102;}

	// inline asm
	// inline asm
	{  cvt.f32.f16 %f1320, %rs103;}

	// inline asm
	// inline asm
	{  cvt.f32.f16 %f1321, %rs104;}

	// inline asm
	// inline asm
	call (%rd177), _rt_buffer_get_64, (%rd172, %r28, %r214, %rd6, %rd7, %rd13, %rd13);
	// inline asm
	add.f32 	%f1322, %f1319, 0f00000000;
	add.f32 	%f1323, %f1320, 0f00000000;
	add.f32 	%f1324, %f1321, 0f00000000;
	// inline asm
	{  cvt.rn.f16.f32 %rs101, %f1324;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs100, %f1323;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs99, %f1322;}

	// inline asm
	mov.u16 	%rs106, 0;
	st.v4.u16 	[%rd177], {%rs99, %rs100, %rs101, %rs106};
	bra.uni 	BB0_119;

BB0_3:
	ld.global.v2.u32 	{%r47, %r48}, [pixelID];
	cvt.u64.u32	%rd17, %r47;
	cvt.u64.u32	%rd18, %r48;
	mov.u64 	%rd26, uvpos;
	cvta.global.u64 	%rd16, %rd26;
	mov.u32 	%r44, 12;
	// inline asm
	call (%rd15), _rt_buffer_get_64, (%rd16, %r28, %r44, %rd17, %rd18, %rd13, %rd13);
	// inline asm
	ld.f32 	%f9, [%rd15+8];
	ld.f32 	%f8, [%rd15+4];
	ld.f32 	%f7, [%rd15];
	mul.f32 	%f309, %f7, 0f3456BF95;
	mul.f32 	%f310, %f8, 0f3456BF95;
	mul.f32 	%f311, %f9, 0f3456BF95;
	abs.f32 	%f312, %f1357;
	div.rn.f32 	%f313, %f309, %f312;
	abs.f32 	%f314, %f1358;
	div.rn.f32 	%f315, %f310, %f314;
	abs.f32 	%f316, %f1359;
	div.rn.f32 	%f317, %f311, %f316;
	abs.f32 	%f318, %f313;
	abs.f32 	%f319, %f315;
	abs.f32 	%f320, %f317;
	mov.f32 	%f321, 0f38D1B717;
	max.f32 	%f322, %f318, %f321;
	max.f32 	%f323, %f319, %f321;
	max.f32 	%f324, %f320, %f321;
	fma.rn.f32 	%f10, %f1357, %f322, %f7;
	fma.rn.f32 	%f11, %f1358, %f323, %f8;
	fma.rn.f32 	%f12, %f1359, %f324, %f9;
	ld.global.v2.u32 	{%r51, %r52}, [pixelID];
	cvt.rn.f32.u32	%f13, %r51;
	cvt.rn.f32.u32	%f14, %r52;
	mov.u64 	%rd27, localLights;
	cvta.global.u64 	%rd25, %rd27;
	mov.u32 	%r45, 1;
	mov.u32 	%r46, 96;
	// inline asm
	call (%rd21, %rd22, %rd23, %rd24), _rt_buffer_get_size_64, (%rd25, %r45, %r46);
	// inline asm
	cvt.u32.u64	%r4, %rd21;
	setp.eq.s32	%p14, %r4, 0;
	mov.f32 	%f1360, 0f00000000;
	mov.f32 	%f1361, %f1360;
	mov.f32 	%f1362, %f1360;
	mov.f32 	%f1363, %f1360;
	mov.f32 	%f1364, %f1360;
	mov.f32 	%f1365, %f1360;
	mov.f32 	%f1366, %f1360;
	mov.f32 	%f1367, %f1360;
	mov.f32 	%f1368, %f1360;
	mov.f32 	%f1369, %f1360;
	mov.f32 	%f1370, %f1360;
	mov.f32 	%f1371, %f1360;
	mov.f32 	%f1372, %f1360;
	mov.f32 	%f1373, %f1360;
	mov.f32 	%f1374, %f1360;
	mov.f32 	%f1375, %f1360;
	@%p14 bra 	BB0_47;

	mov.f32 	%f341, 0f40000000;
	cvt.rzi.f32.f32	%f342, %f341;
	add.f32 	%f343, %f342, %f342;
	mov.f32 	%f344, 0f40800000;
	sub.f32 	%f345, %f344, %f343;
	abs.f32 	%f15, %f345;
	mul.f32 	%f16, %f10, 0f3456BF95;
	mul.f32 	%f17, %f11, 0f3456BF95;
	mul.f32 	%f18, %f12, 0f3456BF95;
	mul.f32 	%f19, %f13, 0f3DD32618;
	mul.f32 	%f20, %f14, 0f3DD2F1AA;
	mov.f32 	%f340, 0f00000000;
	mov.u32 	%r245, 0;
	abs.f32 	%f536, %f16;
	abs.f32 	%f537, %f17;
	max.f32 	%f538, %f536, %f537;
	abs.f32 	%f539, %f18;
	max.f32 	%f540, %f538, %f539;
	mov.f32 	%f1360, %f340;
	mov.f32 	%f1361, %f340;
	mov.f32 	%f1362, %f340;
	mov.f32 	%f1363, %f340;
	mov.f32 	%f1364, %f340;
	mov.f32 	%f1365, %f340;
	mov.f32 	%f1366, %f340;
	mov.f32 	%f1367, %f340;
	mov.f32 	%f1368, %f340;
	mov.f32 	%f1369, %f340;
	mov.f32 	%f1370, %f340;
	mov.f32 	%f1371, %f340;
	mov.f32 	%f1372, %f340;
	mov.f32 	%f1373, %f340;
	mov.f32 	%f1374, %f340;
	mov.f32 	%f1375, %f340;

BB0_5:
	cvt.u64.u32	%rd30, %r245;
	// inline asm
	call (%rd28), _rt_buffer_get_64, (%rd25, %r45, %r46, %rd30, %rd13, %rd13, %rd13);
	// inline asm
	ld.v4.f32 	{%f348, %f349, %f350, %f351}, [%rd28+80];
	ld.v4.f32 	{%f352, %f353, %f354, %f355}, [%rd28+64];
	ld.v4.f32 	{%f356, %f357, %f358, %f359}, [%rd28+48];
	ld.v4.f32 	{%f360, %f1384, %f1385, %f363}, [%rd28+32];
	ld.v4.f32 	{%f364, %f365, %f366, %f367}, [%rd28+16];
	ld.v4.f32 	{%f368, %f369, %f370, %f371}, [%rd28];
	mov.b32 	 %r6, %f351;
	sub.f32 	%f61, %f369, %f7;
	sub.f32 	%f63, %f370, %f8;
	sub.f32 	%f65, %f371, %f9;
	mul.f32 	%f373, %f63, %f63;
	fma.rn.f32 	%f374, %f61, %f61, %f373;
	fma.rn.f32 	%f375, %f65, %f65, %f374;
	sqrt.rn.f32 	%f66, %f375;
	rcp.rn.f32 	%f376, %f66;
	mul.f32 	%f67, %f61, %f376;
	mul.f32 	%f68, %f63, %f376;
	mul.f32 	%f69, %f65, %f376;
	mul.f32 	%f70, %f66, %f367;
	abs.f32 	%f71, %f70;
	setp.lt.f32	%p15, %f71, 0f00800000;
	mul.f32 	%f377, %f71, 0f4B800000;
	selp.f32	%f378, 0fC3170000, 0fC2FE0000, %p15;
	selp.f32	%f379, %f377, %f71, %p15;
	mov.b32 	 %r58, %f379;
	and.b32  	%r59, %r58, 8388607;
	or.b32  	%r60, %r59, 1065353216;
	mov.b32 	 %f380, %r60;
	shr.u32 	%r61, %r58, 23;
	cvt.rn.f32.u32	%f381, %r61;
	add.f32 	%f382, %f378, %f381;
	setp.gt.f32	%p16, %f380, 0f3FB504F3;
	mul.f32 	%f383, %f380, 0f3F000000;
	add.f32 	%f384, %f382, 0f3F800000;
	selp.f32	%f385, %f383, %f380, %p16;
	selp.f32	%f386, %f384, %f382, %p16;
	add.f32 	%f387, %f385, 0fBF800000;
	add.f32 	%f347, %f385, 0f3F800000;
	// inline asm
	rcp.approx.ftz.f32 %f346,%f347;
	// inline asm
	add.f32 	%f388, %f387, %f387;
	mul.f32 	%f389, %f346, %f388;
	mul.f32 	%f390, %f389, %f389;
	mov.f32 	%f391, 0f3C4CAF63;
	mov.f32 	%f392, 0f3B18F0FE;
	fma.rn.f32 	%f393, %f392, %f390, %f391;
	mov.f32 	%f394, 0f3DAAAABD;
	fma.rn.f32 	%f395, %f393, %f390, %f394;
	mul.rn.f32 	%f396, %f395, %f390;
	mul.rn.f32 	%f397, %f396, %f389;
	sub.f32 	%f398, %f387, %f389;
	neg.f32 	%f399, %f389;
	add.f32 	%f400, %f398, %f398;
	fma.rn.f32 	%f401, %f399, %f387, %f400;
	mul.rn.f32 	%f402, %f346, %f401;
	add.f32 	%f403, %f397, %f389;
	sub.f32 	%f404, %f389, %f403;
	add.f32 	%f405, %f397, %f404;
	add.f32 	%f406, %f402, %f405;
	add.f32 	%f407, %f403, %f406;
	sub.f32 	%f408, %f403, %f407;
	add.f32 	%f409, %f406, %f408;
	mov.f32 	%f410, 0f3F317200;
	mul.rn.f32 	%f411, %f386, %f410;
	mov.f32 	%f412, 0f35BFBE8E;
	mul.rn.f32 	%f413, %f386, %f412;
	add.f32 	%f414, %f411, %f407;
	sub.f32 	%f415, %f411, %f414;
	add.f32 	%f416, %f407, %f415;
	add.f32 	%f417, %f409, %f416;
	add.f32 	%f418, %f413, %f417;
	add.f32 	%f419, %f414, %f418;
	sub.f32 	%f420, %f414, %f419;
	add.f32 	%f421, %f418, %f420;
	mul.rn.f32 	%f72, %f344, %f419;
	neg.f32 	%f423, %f72;
	fma.rn.f32 	%f424, %f344, %f419, %f423;
	fma.rn.f32 	%f425, %f344, %f421, %f424;
	fma.rn.f32 	%f73, %f340, %f419, %f425;
	add.rn.f32 	%f74, %f72, %f73;
	mov.b32 	 %r62, %f74;
	setp.eq.s32	%p1, %r62, 1118925336;
	add.s32 	%r63, %r62, -1;
	mov.b32 	 %f427, %r63;
	selp.f32	%f428, %f427, %f74, %p1;
	mul.f32 	%f429, %f428, 0f3FB8AA3B;
	cvt.rzi.f32.f32	%f430, %f429;
	mov.f32 	%f431, 0fBF317200;
	fma.rn.f32 	%f432, %f430, %f431, %f428;
	mov.f32 	%f433, 0fB5BFBE8E;
	fma.rn.f32 	%f434, %f430, %f433, %f432;
	mul.f32 	%f435, %f434, 0f3FB8AA3B;
	ex2.approx.ftz.f32 	%f436, %f435;
	add.f32 	%f437, %f430, 0f00000000;
	ex2.approx.f32 	%f438, %f437;
	mul.f32 	%f439, %f436, %f438;
	setp.lt.f32	%p17, %f428, 0fC2D20000;
	selp.f32	%f440, 0f00000000, %f439, %p17;
	setp.gt.f32	%p18, %f428, 0f42D20000;
	selp.f32	%f1376, 0f7F800000, %f440, %p18;
	setp.eq.f32	%p19, %f1376, 0f7F800000;
	@%p19 bra 	BB0_7;

	neg.f32 	%f441, %f74;
	add.rn.f32 	%f442, %f72, %f441;
	add.rn.f32 	%f443, %f442, %f73;
	add.f32 	%f444, %f443, 0f37000000;
	selp.f32	%f445, %f444, %f443, %p1;
	fma.rn.f32 	%f1376, %f1376, %f445, %f1376;

BB0_7:
	setp.lt.f32	%p20, %f70, 0f00000000;
	setp.eq.f32	%p21, %f15, 0f3F800000;
	and.pred  	%p2, %p20, %p21;
	mov.b32 	 %r64, %f1376;
	xor.b32  	%r65, %r64, -2147483648;
	mov.b32 	 %f446, %r65;
	selp.f32	%f1378, %f446, %f1376, %p2;
	setp.eq.f32	%p22, %f70, 0f00000000;
	@%p22 bra 	BB0_10;
	bra.uni 	BB0_8;

BB0_10:
	add.f32 	%f449, %f70, %f70;
	selp.f32	%f1378, %f449, 0f00000000, %p21;
	bra.uni 	BB0_11;

BB0_8:
	setp.geu.f32	%p23, %f70, 0f00000000;
	@%p23 bra 	BB0_11;

	cvt.rzi.f32.f32	%f448, %f344;
	setp.neu.f32	%p24, %f448, 0f40800000;
	selp.f32	%f1378, 0f7FFFFFFF, %f1378, %p24;

BB0_11:
	add.f32 	%f450, %f71, 0f40800000;
	mov.b32 	 %r66, %f450;
	setp.lt.s32	%p26, %r66, 2139095040;
	@%p26 bra 	BB0_16;

	setp.gtu.f32	%p27, %f71, 0f7F800000;
	@%p27 bra 	BB0_15;
	bra.uni 	BB0_13;

BB0_15:
	add.f32 	%f1378, %f70, 0f40800000;
	bra.uni 	BB0_16;

BB0_13:
	setp.neu.f32	%p28, %f71, 0f7F800000;
	@%p28 bra 	BB0_16;

	selp.f32	%f1378, 0fFF800000, 0f7F800000, %p2;

BB0_16:
	mul.f32 	%f451, %f66, %f365;
	mov.f32 	%f1407, 0f3F800000;
	sub.f32 	%f453, %f1407, %f1378;
	setp.eq.f32	%p29, %f70, 0f3F800000;
	selp.f32	%f454, 0f00000000, %f453, %p29;
	cvt.sat.f32.f32	%f455, %f454;
	fma.rn.f32 	%f456, %f451, %f451, %f366;
	div.rn.f32 	%f100, %f455, %f456;
	mul.f32 	%f457, %f1358, %f68;
	fma.rn.f32 	%f458, %f1357, %f67, %f457;
	fma.rn.f32 	%f101, %f1359, %f69, %f458;
	setp.eq.f32	%p30, %f368, 0f3F800000;
	@%p30 bra 	BB0_23;
	bra.uni 	BB0_17;

BB0_23:
	setp.leu.f32	%p34, %f363, 0f00000000;
	@%p34 bra 	BB0_19;

	setp.geu.f32	%p35, %f359, 0f00000000;
	@%p35 bra 	BB0_25;

	sub.f32 	%f1356, %f371, %f9;
	sub.f32 	%f1355, %f369, %f7;
	sub.f32 	%f1354, %f370, %f8;
	mul.f32 	%f481, %f349, %f1354;
	fma.rn.f32 	%f482, %f348, %f1355, %f481;
	fma.rn.f32 	%f483, %f350, %f1356, %f482;
	rcp.rn.f32 	%f484, %f483;
	mul.f32 	%f1379, %f1355, %f484;
	mul.f32 	%f1380, %f1354, %f484;
	mul.f32 	%f1381, %f1356, %f484;
	neg.f32 	%f359, %f359;
	bra.uni 	BB0_27;

BB0_17:
	setp.eq.f32	%p31, %f368, 0f40000000;
	@%p31 bra 	BB0_21;
	bra.uni 	BB0_18;

BB0_21:
	setp.leu.f32	%p33, %f363, 0f00000000;
	@%p33 bra 	BB0_19;

	mul.f32 	%f475, %f357, %f68;
	fma.rn.f32 	%f476, %f356, %f67, %f475;
	mul.f32 	%f477, %f353, %f68;
	fma.rn.f32 	%f478, %f352, %f67, %f477;
	mul.f32 	%f479, %f349, %f68;
	fma.rn.f32 	%f480, %f348, %f67, %f479;
	fma.rn.f32 	%f472, %f358, %f69, %f476;
	fma.rn.f32 	%f473, %f354, %f69, %f478;
	fma.rn.f32 	%f474, %f350, %f69, %f480;
	cvt.rzi.s32.f32	%r67, %f363;
	mov.u32 	%r68, 6;
	mov.u32 	%r69, 0;
	// inline asm
	call (%f468, %f469, %f470, %f471), _rt_texture_get_base_id, (%r67, %r68, %f472, %f473, %f474, %r69);
	// inline asm
	mul.f32 	%f1383, %f360, %f468;
	mul.f32 	%f1384, %f1384, %f469;
	mul.f32 	%f1385, %f1385, %f470;
	bra.uni 	BB0_28;

BB0_18:
	setp.neu.f32	%p32, %f368, 0f40800000;
	@%p32 bra 	BB0_19;

	mul.f32 	%f459, %f348, %f67;
	mul.f32 	%f460, %f349, %f68;
	neg.f32 	%f461, %f460;
	sub.f32 	%f462, %f461, %f459;
	mul.f32 	%f463, %f350, %f69;
	sub.f32 	%f464, %f462, %f463;
	fma.rn.f32 	%f465, %f363, %f464, %f359;
	cvt.sat.f32.f32	%f466, %f465;
	mul.f32 	%f467, %f466, %f466;
	mul.f32 	%f1386, %f100, %f467;
	mov.f32 	%f1383, %f360;
	bra.uni 	BB0_29;

BB0_19:
	mov.f32 	%f1383, %f360;
	bra.uni 	BB0_28;

BB0_25:
	mov.f32 	%f1379, %f67;
	mov.f32 	%f1380, %f68;
	mov.f32 	%f1381, %f69;

BB0_27:
	mul.f32 	%f493, %f350, %f69;
	mul.f32 	%f494, %f349, %f68;
	neg.f32 	%f495, %f494;
	mul.f32 	%f496, %f348, %f67;
	sub.f32 	%f497, %f495, %f496;
	sub.f32 	%f498, %f497, %f493;
	setp.gt.f32	%p36, %f498, 0f00000000;
	selp.f32	%f499, 0f3F800000, 0f00000000, %p36;
	mul.f32 	%f500, %f357, %f1380;
	fma.rn.f32 	%f501, %f356, %f1379, %f500;
	mul.f32 	%f502, %f353, %f1380;
	fma.rn.f32 	%f503, %f352, %f1379, %f502;
	fma.rn.f32 	%f504, %f358, %f1381, %f501;
	fma.rn.f32 	%f505, %f354, %f1381, %f503;
	fma.rn.f32 	%f506, %f359, %f504, 0f3F000000;
	mov.f32 	%f507, 0f3F800000;
	sub.f32 	%f489, %f507, %f506;
	fma.rn.f32 	%f490, %f359, %f505, 0f3F000000;
	cvt.rzi.s32.f32	%r70, %f363;
	mov.f32 	%f492, 0f00000000;
	// inline asm
	call (%f485, %f486, %f487, %f488), _rt_texture_get_f_id, (%r70, %r28, %f489, %f490, %f492, %f492);
	// inline asm
	mul.f32 	%f508, %f499, %f485;
	mul.f32 	%f509, %f499, %f486;
	mul.f32 	%f510, %f499, %f487;
	mul.f32 	%f1383, %f360, %f508;
	mul.f32 	%f1384, %f1384, %f509;
	mul.f32 	%f1385, %f1385, %f510;

BB0_28:
	mov.f32 	%f1386, %f100;

BB0_29:
	max.f32 	%f526, %f1383, %f1384;
	max.f32 	%f527, %f526, %f1385;
	mul.f32 	%f528, %f1386, %f527;
	setp.lt.f32	%p38, %f528, 0f3727C5AC;
	mov.pred 	%p137, -1;
	mov.f32 	%f136, 0f00000000;
	mov.f32 	%f1388, %f136;
	mov.f32 	%f1389, %f136;
	mov.f32 	%f1390, %f136;
	mov.f32 	%f1391, %f136;
	mov.f32 	%f1392, %f136;
	mov.f32 	%f1393, %f136;
	mov.f32 	%f1394, %f136;
	mov.f32 	%f1395, %f136;
	mov.f32 	%f1396, %f136;
	mov.f32 	%f1397, %f136;
	mov.f32 	%f1398, %f136;
	mov.f32 	%f1399, %f136;
	mov.f32 	%f1400, %f136;
	mov.f32 	%f1401, %f136;
	@%p38 bra 	BB0_31;

	ld.global.u32 	%r72, [ignoreNormal];
	setp.eq.s32	%p40, %r72, 0;
	selp.f32	%f529, %f101, 0f3F800000, %p40;
	cvt.sat.f32.f32	%f530, %f529;
	mul.f32 	%f531, %f1386, %f530;
	mul.f32 	%f136, %f1383, %f531;
	mul.f32 	%f1388, %f1384, %f531;
	mul.f32 	%f1389, %f1385, %f531;
	mul.f32 	%f532, %f1386, 0f3E800000;
	mul.f32 	%f1390, %f1383, %f532;
	mul.f32 	%f1391, %f1384, %f532;
	mul.f32 	%f1392, %f1385, %f532;
	mul.f32 	%f1393, %f67, %f1390;
	mul.f32 	%f1394, %f67, %f1391;
	mul.f32 	%f1395, %f67, %f1392;
	mul.f32 	%f1396, %f68, %f1390;
	mul.f32 	%f1397, %f68, %f1391;
	mul.f32 	%f1398, %f68, %f1392;
	mul.f32 	%f1399, %f69, %f1390;
	mul.f32 	%f1400, %f69, %f1391;
	mul.f32 	%f1401, %f69, %f1392;
	mov.pred 	%p137, 0;

BB0_31:
	@%p137 bra 	BB0_46;

	setp.eq.s32	%p41, %r6, 0;
	mov.u16 	%rs161, 0;
	@%p41 bra 	BB0_43;

	abs.s32 	%r8, %r6;
	mov.f32 	%f1406, 0f00000000;
	setp.lt.s32	%p42, %r8, 1;
	@%p42 bra 	BB0_42;

	max.f32 	%f152, %f540, %f321;
	and.b32  	%r9, %r8, 3;
	setp.eq.s32	%p43, %r9, 0;
	add.u64 	%rd35, %SP, 0;
	cvta.to.local.u64 	%rd2, %rd35;
	mov.f32 	%f1406, 0f00000000;
	mov.u32 	%r249, 0;
	@%p43 bra 	BB0_40;

	setp.eq.s32	%p44, %r9, 1;
	mov.f32 	%f1403, 0f00000000;
	mov.u32 	%r247, 0;
	@%p44 bra 	BB0_39;

	setp.eq.s32	%p45, %r9, 2;
	mov.f32 	%f1402, 0f00000000;
	mov.u32 	%r246, 0;
	@%p45 bra 	BB0_38;

	cvt.rmi.f32.f32	%f552, %f19;
	sub.f32 	%f553, %f19, %f552;
	cvt.rmi.f32.f32	%f554, %f20;
	sub.f32 	%f555, %f20, %f554;
	add.f32 	%f556, %f555, 0f420551EC;
	add.f32 	%f557, %f553, 0f420551EC;
	mul.f32 	%f558, %f553, %f556;
	fma.rn.f32 	%f559, %f555, %f557, %f558;
	add.f32 	%f560, %f559, 0f00000000;
	add.f32 	%f561, %f553, %f560;
	add.f32 	%f562, %f555, %f560;
	add.f32 	%f563, %f561, %f562;
	mul.f32 	%f564, %f560, %f563;
	cvt.rmi.f32.f32	%f565, %f564;
	sub.f32 	%f566, %f564, %f565;
	add.f32 	%f567, %f561, %f561;
	mul.f32 	%f568, %f562, %f567;
	cvt.rmi.f32.f32	%f569, %f568;
	sub.f32 	%f570, %f568, %f569;
	mul.f32 	%f571, %f561, %f563;
	cvt.rmi.f32.f32	%f572, %f571;
	sub.f32 	%f573, %f571, %f572;
	fma.rn.f32 	%f574, %f566, 0f40000000, 0fBF800000;
	fma.rn.f32 	%f575, %f570, 0f40000000, 0fBF800000;
	fma.rn.f32 	%f576, %f573, 0f40000000, 0fBF800000;
	fma.rn.f32 	%f577, %f364, %f574, %f369;
	fma.rn.f32 	%f578, %f364, %f575, %f370;
	fma.rn.f32 	%f579, %f364, %f576, %f371;
	sub.f32 	%f580, %f577, %f7;
	sub.f32 	%f581, %f578, %f8;
	sub.f32 	%f582, %f579, %f9;
	mul.f32 	%f583, %f581, %f581;
	fma.rn.f32 	%f584, %f580, %f580, %f583;
	fma.rn.f32 	%f585, %f582, %f582, %f584;
	sqrt.rn.f32 	%f551, %f585;
	rcp.rn.f32 	%f586, %f551;
	mul.f32 	%f547, %f586, %f580;
	mul.f32 	%f548, %f586, %f581;
	mul.f32 	%f549, %f586, %f582;
	ld.global.u32 	%r80, [imageEnabled];
	and.b32  	%r81, %r80, 32;
	setp.eq.s32	%p46, %r81, 0;
	selp.f32	%f587, 0f3F800000, 0f41200000, %p46;
	mul.f32 	%f550, %f587, %f152;
	mov.u32 	%r82, 1065353216;
	st.local.u32 	[%rd2], %r82;
	ld.global.u32 	%r76, [root];
	// inline asm
	call _rt_trace_64, (%r76, %f10, %f11, %f12, %f547, %f548, %f549, %r45, %f550, %f551, %rd35, %r29);
	// inline asm
	ld.local.f32 	%f588, [%rd2];
	add.f32 	%f1402, %f588, 0f00000000;
	mov.u32 	%r246, %r45;

BB0_38:
	cvt.rn.f32.s32	%f597, %r246;
	add.f32 	%f598, %f13, %f597;
	sub.f32 	%f599, %f14, %f597;
	mul.f32 	%f600, %f598, 0f3DD32618;
	cvt.rmi.f32.f32	%f601, %f600;
	sub.f32 	%f602, %f600, %f601;
	mul.f32 	%f603, %f599, 0f3DD2F1AA;
	cvt.rmi.f32.f32	%f604, %f603;
	sub.f32 	%f605, %f603, %f604;
	mul.f32 	%f606, %f597, 0f3DC74539;
	cvt.rmi.f32.f32	%f607, %f606;
	sub.f32 	%f608, %f606, %f607;
	add.f32 	%f609, %f605, 0f420551EC;
	add.f32 	%f610, %f602, 0f420551EC;
	add.f32 	%f611, %f608, 0f420551EC;
	mul.f32 	%f612, %f602, %f609;
	fma.rn.f32 	%f613, %f605, %f610, %f612;
	fma.rn.f32 	%f614, %f608, %f611, %f613;
	add.f32 	%f615, %f602, %f614;
	add.f32 	%f616, %f605, %f614;
	add.f32 	%f617, %f608, %f614;
	add.f32 	%f618, %f615, %f616;
	mul.f32 	%f619, %f617, %f618;
	cvt.rmi.f32.f32	%f620, %f619;
	sub.f32 	%f621, %f619, %f620;
	add.f32 	%f622, %f615, %f615;
	mul.f32 	%f623, %f616, %f622;
	cvt.rmi.f32.f32	%f624, %f623;
	sub.f32 	%f625, %f623, %f624;
	mul.f32 	%f626, %f615, %f618;
	cvt.rmi.f32.f32	%f627, %f626;
	sub.f32 	%f628, %f626, %f627;
	fma.rn.f32 	%f629, %f621, 0f40000000, 0fBF800000;
	fma.rn.f32 	%f630, %f625, 0f40000000, 0fBF800000;
	fma.rn.f32 	%f631, %f628, 0f40000000, 0fBF800000;
	fma.rn.f32 	%f632, %f364, %f629, %f369;
	fma.rn.f32 	%f633, %f364, %f630, %f370;
	fma.rn.f32 	%f634, %f364, %f631, %f371;
	sub.f32 	%f635, %f632, %f7;
	sub.f32 	%f636, %f633, %f8;
	sub.f32 	%f637, %f634, %f9;
	mul.f32 	%f638, %f636, %f636;
	fma.rn.f32 	%f639, %f635, %f635, %f638;
	fma.rn.f32 	%f640, %f637, %f637, %f639;
	sqrt.rn.f32 	%f596, %f640;
	rcp.rn.f32 	%f641, %f596;
	mul.f32 	%f592, %f641, %f635;
	mul.f32 	%f593, %f641, %f636;
	mul.f32 	%f594, %f641, %f637;
	ld.global.u32 	%r86, [imageEnabled];
	and.b32  	%r87, %r86, 32;
	setp.eq.s32	%p47, %r87, 0;
	selp.f32	%f642, 0f3F800000, 0f41200000, %p47;
	mul.f32 	%f595, %f642, %f152;
	mov.u32 	%r88, 1065353216;
	st.local.u32 	[%rd2], %r88;
	ld.global.u32 	%r83, [root];
	// inline asm
	call _rt_trace_64, (%r83, %f10, %f11, %f12, %f592, %f593, %f594, %r45, %f595, %f596, %rd35, %r29);
	// inline asm
	ld.local.f32 	%f643, [%rd2];
	add.f32 	%f1403, %f1402, %f643;
	add.s32 	%r247, %r246, 1;

BB0_39:
	cvt.rn.f32.s32	%f652, %r247;
	add.f32 	%f653, %f13, %f652;
	sub.f32 	%f654, %f14, %f652;
	mul.f32 	%f655, %f653, 0f3DD32618;
	cvt.rmi.f32.f32	%f656, %f655;
	sub.f32 	%f657, %f655, %f656;
	mul.f32 	%f658, %f654, 0f3DD2F1AA;
	cvt.rmi.f32.f32	%f659, %f658;
	sub.f32 	%f660, %f658, %f659;
	mul.f32 	%f661, %f652, 0f3DC74539;
	cvt.rmi.f32.f32	%f662, %f661;
	sub.f32 	%f663, %f661, %f662;
	add.f32 	%f664, %f660, 0f420551EC;
	add.f32 	%f665, %f657, 0f420551EC;
	add.f32 	%f666, %f663, 0f420551EC;
	mul.f32 	%f667, %f657, %f664;
	fma.rn.f32 	%f668, %f660, %f665, %f667;
	fma.rn.f32 	%f669, %f663, %f666, %f668;
	add.f32 	%f670, %f657, %f669;
	add.f32 	%f671, %f660, %f669;
	add.f32 	%f672, %f663, %f669;
	add.f32 	%f673, %f670, %f671;
	mul.f32 	%f674, %f672, %f673;
	cvt.rmi.f32.f32	%f675, %f674;
	sub.f32 	%f676, %f674, %f675;
	add.f32 	%f677, %f670, %f670;
	mul.f32 	%f678, %f671, %f677;
	cvt.rmi.f32.f32	%f679, %f678;
	sub.f32 	%f680, %f678, %f679;
	mul.f32 	%f681, %f670, %f673;
	cvt.rmi.f32.f32	%f682, %f681;
	sub.f32 	%f683, %f681, %f682;
	fma.rn.f32 	%f684, %f676, 0f40000000, 0fBF800000;
	fma.rn.f32 	%f685, %f680, 0f40000000, 0fBF800000;
	fma.rn.f32 	%f686, %f683, 0f40000000, 0fBF800000;
	fma.rn.f32 	%f687, %f364, %f684, %f369;
	fma.rn.f32 	%f688, %f364, %f685, %f370;
	fma.rn.f32 	%f689, %f364, %f686, %f371;
	sub.f32 	%f690, %f687, %f7;
	sub.f32 	%f691, %f688, %f8;
	sub.f32 	%f692, %f689, %f9;
	mul.f32 	%f693, %f691, %f691;
	fma.rn.f32 	%f694, %f690, %f690, %f693;
	fma.rn.f32 	%f695, %f692, %f692, %f694;
	sqrt.rn.f32 	%f651, %f695;
	rcp.rn.f32 	%f696, %f651;
	mul.f32 	%f647, %f696, %f690;
	mul.f32 	%f648, %f696, %f691;
	mul.f32 	%f649, %f696, %f692;
	ld.global.u32 	%r92, [imageEnabled];
	and.b32  	%r93, %r92, 32;
	setp.eq.s32	%p48, %r93, 0;
	selp.f32	%f697, 0f3F800000, 0f41200000, %p48;
	mul.f32 	%f650, %f697, %f152;
	mov.u32 	%r94, 1065353216;
	st.local.u32 	[%rd2], %r94;
	ld.global.u32 	%r89, [root];
	mov.u32 	%r90, 1;
	// inline asm
	call _rt_trace_64, (%r89, %f10, %f11, %f12, %f647, %f648, %f649, %r90, %f650, %f651, %rd35, %r29);
	// inline asm
	ld.local.f32 	%f698, [%rd2];
	add.f32 	%f1406, %f1403, %f698;
	add.s32 	%r249, %r247, 1;

BB0_40:
	setp.lt.u32	%p49, %r8, 4;
	@%p49 bra 	BB0_42;

BB0_41:
	cvt.rn.f32.s32	%f731, %r249;
	add.f32 	%f732, %f13, %f731;
	sub.f32 	%f733, %f14, %f731;
	mul.f32 	%f734, %f732, 0f3DD32618;
	cvt.rmi.f32.f32	%f735, %f734;
	sub.f32 	%f736, %f734, %f735;
	mul.f32 	%f737, %f733, 0f3DD2F1AA;
	cvt.rmi.f32.f32	%f738, %f737;
	sub.f32 	%f739, %f737, %f738;
	mul.f32 	%f740, %f731, 0f3DC74539;
	cvt.rmi.f32.f32	%f741, %f740;
	sub.f32 	%f742, %f740, %f741;
	add.f32 	%f743, %f739, 0f420551EC;
	add.f32 	%f744, %f736, 0f420551EC;
	add.f32 	%f745, %f742, 0f420551EC;
	mul.f32 	%f746, %f736, %f743;
	fma.rn.f32 	%f747, %f739, %f744, %f746;
	fma.rn.f32 	%f748, %f742, %f745, %f747;
	add.f32 	%f749, %f736, %f748;
	add.f32 	%f750, %f739, %f748;
	add.f32 	%f751, %f742, %f748;
	add.f32 	%f752, %f749, %f750;
	mul.f32 	%f753, %f751, %f752;
	cvt.rmi.f32.f32	%f754, %f753;
	sub.f32 	%f755, %f753, %f754;
	add.f32 	%f756, %f749, %f749;
	mul.f32 	%f757, %f750, %f756;
	cvt.rmi.f32.f32	%f758, %f757;
	sub.f32 	%f759, %f757, %f758;
	mul.f32 	%f760, %f749, %f752;
	cvt.rmi.f32.f32	%f761, %f760;
	sub.f32 	%f762, %f760, %f761;
	fma.rn.f32 	%f763, %f755, 0f40000000, 0fBF800000;
	fma.rn.f32 	%f764, %f759, 0f40000000, 0fBF800000;
	fma.rn.f32 	%f765, %f762, 0f40000000, 0fBF800000;
	fma.rn.f32 	%f766, %f364, %f763, %f369;
	fma.rn.f32 	%f767, %f364, %f764, %f370;
	fma.rn.f32 	%f768, %f364, %f765, %f371;
	sub.f32 	%f769, %f766, %f7;
	sub.f32 	%f770, %f767, %f8;
	sub.f32 	%f771, %f768, %f9;
	mul.f32 	%f772, %f770, %f770;
	fma.rn.f32 	%f773, %f769, %f769, %f772;
	fma.rn.f32 	%f774, %f771, %f771, %f773;
	sqrt.rn.f32 	%f706, %f774;
	rcp.rn.f32 	%f775, %f706;
	mul.f32 	%f702, %f775, %f769;
	mul.f32 	%f703, %f775, %f770;
	mul.f32 	%f704, %f775, %f771;
	ld.global.u32 	%r107, [imageEnabled];
	and.b32  	%r108, %r107, 32;
	setp.eq.s32	%p50, %r108, 0;
	selp.f32	%f776, 0f3F800000, 0f41200000, %p50;
	mul.f32 	%f705, %f776, %f152;
	mov.u32 	%r109, 1065353216;
	st.local.u32 	[%rd2], %r109;
	ld.global.u32 	%r95, [root];
	mov.u32 	%r105, 1;
	// inline asm
	call _rt_trace_64, (%r95, %f10, %f11, %f12, %f702, %f703, %f704, %r105, %f705, %f706, %rd35, %r29);
	// inline asm
	ld.local.f32 	%f777, [%rd2];
	add.f32 	%f778, %f1406, %f777;
	add.s32 	%r110, %r249, 1;
	cvt.rn.f32.s32	%f779, %r110;
	add.f32 	%f780, %f13, %f779;
	sub.f32 	%f781, %f14, %f779;
	mul.f32 	%f782, %f780, 0f3DD32618;
	cvt.rmi.f32.f32	%f783, %f782;
	sub.f32 	%f784, %f782, %f783;
	mul.f32 	%f785, %f781, 0f3DD2F1AA;
	cvt.rmi.f32.f32	%f786, %f785;
	sub.f32 	%f787, %f785, %f786;
	mul.f32 	%f788, %f779, 0f3DC74539;
	cvt.rmi.f32.f32	%f789, %f788;
	sub.f32 	%f790, %f788, %f789;
	add.f32 	%f791, %f787, 0f420551EC;
	add.f32 	%f792, %f784, 0f420551EC;
	add.f32 	%f793, %f790, 0f420551EC;
	mul.f32 	%f794, %f784, %f791;
	fma.rn.f32 	%f795, %f787, %f792, %f794;
	fma.rn.f32 	%f796, %f790, %f793, %f795;
	add.f32 	%f797, %f784, %f796;
	add.f32 	%f798, %f787, %f796;
	add.f32 	%f799, %f790, %f796;
	add.f32 	%f800, %f797, %f798;
	mul.f32 	%f801, %f799, %f800;
	cvt.rmi.f32.f32	%f802, %f801;
	sub.f32 	%f803, %f801, %f802;
	add.f32 	%f804, %f797, %f797;
	mul.f32 	%f805, %f798, %f804;
	cvt.rmi.f32.f32	%f806, %f805;
	sub.f32 	%f807, %f805, %f806;
	mul.f32 	%f808, %f797, %f800;
	cvt.rmi.f32.f32	%f809, %f808;
	sub.f32 	%f810, %f808, %f809;
	fma.rn.f32 	%f811, %f803, 0f40000000, 0fBF800000;
	fma.rn.f32 	%f812, %f807, 0f40000000, 0fBF800000;
	fma.rn.f32 	%f813, %f810, 0f40000000, 0fBF800000;
	fma.rn.f32 	%f814, %f364, %f811, %f369;
	fma.rn.f32 	%f815, %f364, %f812, %f370;
	fma.rn.f32 	%f816, %f364, %f813, %f371;
	sub.f32 	%f817, %f814, %f7;
	sub.f32 	%f818, %f815, %f8;
	sub.f32 	%f819, %f816, %f9;
	mul.f32 	%f820, %f818, %f818;
	fma.rn.f32 	%f821, %f817, %f817, %f820;
	fma.rn.f32 	%f822, %f819, %f819, %f821;
	sqrt.rn.f32 	%f714, %f822;
	rcp.rn.f32 	%f823, %f714;
	mul.f32 	%f710, %f823, %f817;
	mul.f32 	%f711, %f823, %f818;
	mul.f32 	%f712, %f823, %f819;
	ld.global.u32 	%r111, [imageEnabled];
	and.b32  	%r112, %r111, 32;
	setp.eq.s32	%p51, %r112, 0;
	selp.f32	%f824, 0f3F800000, 0f41200000, %p51;
	mul.f32 	%f713, %f824, %f152;
	st.local.u32 	[%rd2], %r109;
	ld.global.u32 	%r98, [root];
	// inline asm
	call _rt_trace_64, (%r98, %f10, %f11, %f12, %f710, %f711, %f712, %r105, %f713, %f714, %rd35, %r29);
	// inline asm
	ld.local.f32 	%f825, [%rd2];
	add.f32 	%f826, %f778, %f825;
	add.s32 	%r113, %r249, 2;
	cvt.rn.f32.s32	%f827, %r113;
	add.f32 	%f828, %f13, %f827;
	sub.f32 	%f829, %f14, %f827;
	mul.f32 	%f830, %f828, 0f3DD32618;
	cvt.rmi.f32.f32	%f831, %f830;
	sub.f32 	%f832, %f830, %f831;
	mul.f32 	%f833, %f829, 0f3DD2F1AA;
	cvt.rmi.f32.f32	%f834, %f833;
	sub.f32 	%f835, %f833, %f834;
	mul.f32 	%f836, %f827, 0f3DC74539;
	cvt.rmi.f32.f32	%f837, %f836;
	sub.f32 	%f838, %f836, %f837;
	add.f32 	%f839, %f835, 0f420551EC;
	add.f32 	%f840, %f832, 0f420551EC;
	add.f32 	%f841, %f838, 0f420551EC;
	mul.f32 	%f842, %f832, %f839;
	fma.rn.f32 	%f843, %f835, %f840, %f842;
	fma.rn.f32 	%f844, %f838, %f841, %f843;
	add.f32 	%f845, %f832, %f844;
	add.f32 	%f846, %f835, %f844;
	add.f32 	%f847, %f838, %f844;
	add.f32 	%f848, %f845, %f846;
	mul.f32 	%f849, %f847, %f848;
	cvt.rmi.f32.f32	%f850, %f849;
	sub.f32 	%f851, %f849, %f850;
	add.f32 	%f852, %f845, %f845;
	mul.f32 	%f853, %f846, %f852;
	cvt.rmi.f32.f32	%f854, %f853;
	sub.f32 	%f855, %f853, %f854;
	mul.f32 	%f856, %f845, %f848;
	cvt.rmi.f32.f32	%f857, %f856;
	sub.f32 	%f858, %f856, %f857;
	fma.rn.f32 	%f859, %f851, 0f40000000, 0fBF800000;
	fma.rn.f32 	%f860, %f855, 0f40000000, 0fBF800000;
	fma.rn.f32 	%f861, %f858, 0f40000000, 0fBF800000;
	fma.rn.f32 	%f862, %f364, %f859, %f369;
	fma.rn.f32 	%f863, %f364, %f860, %f370;
	fma.rn.f32 	%f864, %f364, %f861, %f371;
	sub.f32 	%f865, %f862, %f7;
	sub.f32 	%f866, %f863, %f8;
	sub.f32 	%f867, %f864, %f9;
	mul.f32 	%f868, %f866, %f866;
	fma.rn.f32 	%f869, %f865, %f865, %f868;
	fma.rn.f32 	%f870, %f867, %f867, %f869;
	sqrt.rn.f32 	%f722, %f870;
	rcp.rn.f32 	%f871, %f722;
	mul.f32 	%f718, %f871, %f865;
	mul.f32 	%f719, %f871, %f866;
	mul.f32 	%f720, %f871, %f867;
	ld.global.u32 	%r114, [imageEnabled];
	and.b32  	%r115, %r114, 32;
	setp.eq.s32	%p52, %r115, 0;
	selp.f32	%f872, 0f3F800000, 0f41200000, %p52;
	mul.f32 	%f721, %f872, %f152;
	st.local.u32 	[%rd2], %r109;
	ld.global.u32 	%r101, [root];
	// inline asm
	call _rt_trace_64, (%r101, %f10, %f11, %f12, %f718, %f719, %f720, %r105, %f721, %f722, %rd35, %r29);
	// inline asm
	ld.local.f32 	%f873, [%rd2];
	add.f32 	%f874, %f826, %f873;
	add.s32 	%r116, %r249, 3;
	cvt.rn.f32.s32	%f875, %r116;
	add.f32 	%f876, %f13, %f875;
	sub.f32 	%f877, %f14, %f875;
	mul.f32 	%f878, %f876, 0f3DD32618;
	cvt.rmi.f32.f32	%f879, %f878;
	sub.f32 	%f880, %f878, %f879;
	mul.f32 	%f881, %f877, 0f3DD2F1AA;
	cvt.rmi.f32.f32	%f882, %f881;
	sub.f32 	%f883, %f881, %f882;
	mul.f32 	%f884, %f875, 0f3DC74539;
	cvt.rmi.f32.f32	%f885, %f884;
	sub.f32 	%f886, %f884, %f885;
	add.f32 	%f887, %f883, 0f420551EC;
	add.f32 	%f888, %f880, 0f420551EC;
	add.f32 	%f889, %f886, 0f420551EC;
	mul.f32 	%f890, %f880, %f887;
	fma.rn.f32 	%f891, %f883, %f888, %f890;
	fma.rn.f32 	%f892, %f886, %f889, %f891;
	add.f32 	%f893, %f880, %f892;
	add.f32 	%f894, %f883, %f892;
	add.f32 	%f895, %f886, %f892;
	add.f32 	%f896, %f893, %f894;
	mul.f32 	%f897, %f895, %f896;
	cvt.rmi.f32.f32	%f898, %f897;
	sub.f32 	%f899, %f897, %f898;
	add.f32 	%f900, %f893, %f893;
	mul.f32 	%f901, %f894, %f900;
	cvt.rmi.f32.f32	%f902, %f901;
	sub.f32 	%f903, %f901, %f902;
	mul.f32 	%f904, %f893, %f896;
	cvt.rmi.f32.f32	%f905, %f904;
	sub.f32 	%f906, %f904, %f905;
	fma.rn.f32 	%f907, %f899, 0f40000000, 0fBF800000;
	fma.rn.f32 	%f908, %f903, 0f40000000, 0fBF800000;
	fma.rn.f32 	%f909, %f906, 0f40000000, 0fBF800000;
	fma.rn.f32 	%f910, %f364, %f907, %f369;
	fma.rn.f32 	%f911, %f364, %f908, %f370;
	fma.rn.f32 	%f912, %f364, %f909, %f371;
	sub.f32 	%f913, %f910, %f7;
	sub.f32 	%f914, %f911, %f8;
	sub.f32 	%f915, %f912, %f9;
	mul.f32 	%f916, %f914, %f914;
	fma.rn.f32 	%f917, %f913, %f913, %f916;
	fma.rn.f32 	%f918, %f915, %f915, %f917;
	sqrt.rn.f32 	%f730, %f918;
	rcp.rn.f32 	%f919, %f730;
	mul.f32 	%f726, %f919, %f913;
	mul.f32 	%f727, %f919, %f914;
	mul.f32 	%f728, %f919, %f915;
	ld.global.u32 	%r117, [imageEnabled];
	and.b32  	%r118, %r117, 32;
	setp.eq.s32	%p53, %r118, 0;
	selp.f32	%f920, 0f3F800000, 0f41200000, %p53;
	mul.f32 	%f729, %f920, %f152;
	st.local.u32 	[%rd2], %r109;
	ld.global.u32 	%r104, [root];
	// inline asm
	call _rt_trace_64, (%r104, %f10, %f11, %f12, %f726, %f727, %f728, %r105, %f729, %f730, %rd35, %r29);
	// inline asm
	ld.local.f32 	%f921, [%rd2];
	add.f32 	%f1406, %f874, %f921;
	add.s32 	%r249, %r249, 4;
	setp.lt.s32	%p54, %r249, %r8;
	@%p54 bra 	BB0_41;

BB0_42:
	cvt.rn.f32.s32	%f922, %r8;
	div.rn.f32 	%f1407, %f1406, %f922;
	shr.u32 	%r119, %r6, 31;
	cvt.u16.u32	%rs161, %r119;

BB0_43:
	fma.rn.f32 	%f1375, %f136, %f1407, %f1375;
	fma.rn.f32 	%f1374, %f1388, %f1407, %f1374;
	fma.rn.f32 	%f1373, %f1389, %f1407, %f1373;
	fma.rn.f32 	%f1372, %f1390, %f1407, %f1372;
	fma.rn.f32 	%f1371, %f1391, %f1407, %f1371;
	fma.rn.f32 	%f1370, %f1392, %f1407, %f1370;
	fma.rn.f32 	%f1369, %f1393, %f1407, %f1369;
	fma.rn.f32 	%f1368, %f1394, %f1407, %f1368;
	fma.rn.f32 	%f1367, %f1395, %f1407, %f1367;
	fma.rn.f32 	%f1366, %f1396, %f1407, %f1366;
	fma.rn.f32 	%f1365, %f1397, %f1407, %f1365;
	fma.rn.f32 	%f1364, %f1398, %f1407, %f1364;
	fma.rn.f32 	%f1363, %f1399, %f1407, %f1363;
	fma.rn.f32 	%f1362, %f1400, %f1407, %f1362;
	fma.rn.f32 	%f1361, %f1401, %f1407, %f1361;
	setp.eq.s16	%p55, %rs161, 0;
	@%p55 bra 	BB0_45;

	div.rn.f32 	%f923, %f136, %f360;
	div.rn.f32 	%f924, %f923, %f100;
	cvt.sat.f32.f32	%f925, %f924;
	mul.f32 	%f1407, %f1407, %f925;

BB0_45:
	add.f32 	%f1360, %f1360, %f1407;

BB0_46:
	add.s32 	%r245, %r245, 1;
	setp.lt.u32	%p56, %r245, %r4;
	@%p56 bra 	BB0_5;

BB0_47:
	ld.global.u32 	%r251, [imageEnabled];
	and.b32  	%r120, %r251, 8;
	setp.eq.s32	%p57, %r120, 0;
	@%p57 bra 	BB0_60;

	cvt.sat.f32.f32	%f214, %f1360;
	cvt.u64.u32	%rd46, %r3;
	cvt.u64.u32	%rd45, %r2;
	mov.u64 	%rd49, image_Mask;
	cvta.global.u64 	%rd44, %rd49;
	// inline asm
	call (%rd43), _rt_buffer_get_64, (%rd44, %r28, %r28, %rd45, %rd46, %rd13, %rd13);
	// inline asm
	mov.f32 	%f928, 0f3E68BA2E;
	cvt.rzi.f32.f32	%f929, %f928;
	fma.rn.f32 	%f930, %f929, 0fC0000000, 0f3EE8BA2E;
	abs.f32 	%f215, %f930;
	abs.f32 	%f216, %f214;
	setp.lt.f32	%p58, %f216, 0f00800000;
	mul.f32 	%f931, %f216, 0f4B800000;
	selp.f32	%f932, 0fC3170000, 0fC2FE0000, %p58;
	selp.f32	%f933, %f931, %f216, %p58;
	mov.b32 	 %r123, %f933;
	and.b32  	%r124, %r123, 8388607;
	or.b32  	%r125, %r124, 1065353216;
	mov.b32 	 %f934, %r125;
	shr.u32 	%r126, %r123, 23;
	cvt.rn.f32.u32	%f935, %r126;
	add.f32 	%f936, %f932, %f935;
	setp.gt.f32	%p59, %f934, 0f3FB504F3;
	mul.f32 	%f937, %f934, 0f3F000000;
	add.f32 	%f938, %f936, 0f3F800000;
	selp.f32	%f939, %f937, %f934, %p59;
	selp.f32	%f940, %f938, %f936, %p59;
	add.f32 	%f941, %f939, 0fBF800000;
	add.f32 	%f927, %f939, 0f3F800000;
	// inline asm
	rcp.approx.ftz.f32 %f926,%f927;
	// inline asm
	add.f32 	%f942, %f941, %f941;
	mul.f32 	%f943, %f926, %f942;
	mul.f32 	%f944, %f943, %f943;
	mov.f32 	%f945, 0f3C4CAF63;
	mov.f32 	%f946, 0f3B18F0FE;
	fma.rn.f32 	%f947, %f946, %f944, %f945;
	mov.f32 	%f948, 0f3DAAAABD;
	fma.rn.f32 	%f949, %f947, %f944, %f948;
	mul.rn.f32 	%f950, %f949, %f944;
	mul.rn.f32 	%f951, %f950, %f943;
	sub.f32 	%f952, %f941, %f943;
	neg.f32 	%f953, %f943;
	add.f32 	%f954, %f952, %f952;
	fma.rn.f32 	%f955, %f953, %f941, %f954;
	mul.rn.f32 	%f956, %f926, %f955;
	add.f32 	%f957, %f951, %f943;
	sub.f32 	%f958, %f943, %f957;
	add.f32 	%f959, %f951, %f958;
	add.f32 	%f960, %f956, %f959;
	add.f32 	%f961, %f957, %f960;
	sub.f32 	%f962, %f957, %f961;
	add.f32 	%f963, %f960, %f962;
	mov.f32 	%f964, 0f3F317200;
	mul.rn.f32 	%f965, %f940, %f964;
	mov.f32 	%f966, 0f35BFBE8E;
	mul.rn.f32 	%f967, %f940, %f966;
	add.f32 	%f968, %f965, %f961;
	sub.f32 	%f969, %f965, %f968;
	add.f32 	%f970, %f961, %f969;
	add.f32 	%f971, %f963, %f970;
	add.f32 	%f972, %f967, %f971;
	add.f32 	%f973, %f968, %f972;
	sub.f32 	%f974, %f968, %f973;
	add.f32 	%f975, %f972, %f974;
	mov.f32 	%f976, 0f3EE8BA2E;
	mul.rn.f32 	%f977, %f976, %f973;
	neg.f32 	%f978, %f977;
	fma.rn.f32 	%f979, %f976, %f973, %f978;
	fma.rn.f32 	%f980, %f976, %f975, %f979;
	mov.f32 	%f981, 0f00000000;
	fma.rn.f32 	%f982, %f981, %f973, %f980;
	add.rn.f32 	%f983, %f977, %f982;
	neg.f32 	%f984, %f983;
	add.rn.f32 	%f985, %f977, %f984;
	add.rn.f32 	%f986, %f985, %f982;
	mov.b32 	 %r127, %f983;
	setp.eq.s32	%p60, %r127, 1118925336;
	add.s32 	%r128, %r127, -1;
	mov.b32 	 %f987, %r128;
	add.f32 	%f988, %f986, 0f37000000;
	selp.f32	%f989, %f987, %f983, %p60;
	selp.f32	%f217, %f988, %f986, %p60;
	mul.f32 	%f990, %f989, 0f3FB8AA3B;
	cvt.rzi.f32.f32	%f991, %f990;
	mov.f32 	%f992, 0fBF317200;
	fma.rn.f32 	%f993, %f991, %f992, %f989;
	mov.f32 	%f994, 0fB5BFBE8E;
	fma.rn.f32 	%f995, %f991, %f994, %f993;
	mul.f32 	%f996, %f995, 0f3FB8AA3B;
	ex2.approx.ftz.f32 	%f997, %f996;
	add.f32 	%f998, %f991, 0f00000000;
	ex2.approx.f32 	%f999, %f998;
	mul.f32 	%f1000, %f997, %f999;
	setp.lt.f32	%p61, %f989, 0fC2D20000;
	selp.f32	%f1001, 0f00000000, %f1000, %p61;
	setp.gt.f32	%p62, %f989, 0f42D20000;
	selp.f32	%f1441, 0f7F800000, %f1001, %p62;
	setp.eq.f32	%p63, %f1441, 0f7F800000;
	@%p63 bra 	BB0_50;

	fma.rn.f32 	%f1441, %f1441, %f217, %f1441;

BB0_50:
	setp.lt.f32	%p64, %f214, 0f00000000;
	setp.eq.f32	%p65, %f215, 0f3F800000;
	and.pred  	%p4, %p64, %p65;
	mov.b32 	 %r129, %f1441;
	xor.b32  	%r130, %r129, -2147483648;
	mov.b32 	 %f1002, %r130;
	selp.f32	%f1443, %f1002, %f1441, %p4;
	setp.eq.f32	%p66, %f214, 0f00000000;
	@%p66 bra 	BB0_53;
	bra.uni 	BB0_51;

BB0_53:
	add.f32 	%f1005, %f214, %f214;
	selp.f32	%f1443, %f1005, 0f00000000, %p65;
	bra.uni 	BB0_54;

BB0_118:
	mov.u64 	%rd190, image_HDR;
	cvta.global.u64 	%rd185, %rd190;
	mov.u32 	%r216, 8;
	// inline asm
	call (%rd184), _rt_buffer_get_64, (%rd185, %r28, %r216, %rd6, %rd7, %rd13, %rd13);
	// inline asm
	mov.f32 	%f1325, 0f00000000;
	// inline asm
	{  cvt.rn.f16.f32 %rs107, %f1325;}

	// inline asm
	mov.u16 	%rs108, 0;
	st.v4.u16 	[%rd184], {%rs107, %rs107, %rs107, %rs108};

BB0_119:
	ld.global.u32 	%r217, [additive];
	setp.eq.s32	%p133, %r217, 0;
	@%p133 bra 	BB0_121;

	mov.u64 	%rd203, image_RNM0;
	cvta.global.u64 	%rd192, %rd203;
	mov.u32 	%r221, 8;
	// inline asm
	call (%rd191), _rt_buffer_get_64, (%rd192, %r28, %r221, %rd6, %rd7, %rd13, %rd13);
	// inline asm
	ld.v4.u16 	{%rs115, %rs116, %rs117, %rs118}, [%rd191];
	// inline asm
	{  cvt.f32.f16 %f1326, %rs115;}

	// inline asm
	// inline asm
	{  cvt.f32.f16 %f1327, %rs116;}

	// inline asm
	// inline asm
	{  cvt.f32.f16 %f1328, %rs117;}

	// inline asm
	// inline asm
	call (%rd197), _rt_buffer_get_64, (%rd192, %r28, %r221, %rd6, %rd7, %rd13, %rd13);
	// inline asm
	add.f32 	%f1329, %f1326, 0f00000000;
	add.f32 	%f1330, %f1327, 0f00000000;
	add.f32 	%f1331, %f1328, 0f00000000;
	// inline asm
	{  cvt.rn.f16.f32 %rs114, %f1331;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs113, %f1330;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs112, %f1329;}

	// inline asm
	mov.u16 	%rs119, 0;
	st.v4.u16 	[%rd197], {%rs112, %rs113, %rs114, %rs119};
	bra.uni 	BB0_122;

BB0_121:
	mov.u64 	%rd210, image_RNM0;
	cvta.global.u64 	%rd205, %rd210;
	mov.u32 	%r223, 8;
	// inline asm
	call (%rd204), _rt_buffer_get_64, (%rd205, %r28, %r223, %rd6, %rd7, %rd13, %rd13);
	// inline asm
	mov.f32 	%f1332, 0f00000000;
	// inline asm
	{  cvt.rn.f16.f32 %rs120, %f1332;}

	// inline asm
	mov.u16 	%rs121, 0;
	st.v4.u16 	[%rd204], {%rs120, %rs120, %rs120, %rs121};

BB0_122:
	ld.global.u32 	%r224, [additive];
	setp.eq.s32	%p134, %r224, 0;
	@%p134 bra 	BB0_124;

	mov.u64 	%rd223, image_RNM1;
	cvta.global.u64 	%rd212, %rd223;
	mov.u32 	%r228, 8;
	// inline asm
	call (%rd211), _rt_buffer_get_64, (%rd212, %r28, %r228, %rd6, %rd7, %rd13, %rd13);
	// inline asm
	ld.v4.u16 	{%rs128, %rs129, %rs130, %rs131}, [%rd211];
	// inline asm
	{  cvt.f32.f16 %f1333, %rs128;}

	// inline asm
	// inline asm
	{  cvt.f32.f16 %f1334, %rs129;}

	// inline asm
	// inline asm
	{  cvt.f32.f16 %f1335, %rs130;}

	// inline asm
	// inline asm
	call (%rd217), _rt_buffer_get_64, (%rd212, %r28, %r228, %rd6, %rd7, %rd13, %rd13);
	// inline asm
	add.f32 	%f1336, %f1333, 0f00000000;
	add.f32 	%f1337, %f1334, 0f00000000;
	add.f32 	%f1338, %f1335, 0f00000000;
	// inline asm
	{  cvt.rn.f16.f32 %rs127, %f1338;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs126, %f1337;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs125, %f1336;}

	// inline asm
	mov.u16 	%rs132, 0;
	st.v4.u16 	[%rd217], {%rs125, %rs126, %rs127, %rs132};
	bra.uni 	BB0_125;

BB0_124:
	mov.u64 	%rd230, image_RNM1;
	cvta.global.u64 	%rd225, %rd230;
	mov.u32 	%r230, 8;
	// inline asm
	call (%rd224), _rt_buffer_get_64, (%rd225, %r28, %r230, %rd6, %rd7, %rd13, %rd13);
	// inline asm
	mov.f32 	%f1339, 0f00000000;
	// inline asm
	{  cvt.rn.f16.f32 %rs133, %f1339;}

	// inline asm
	mov.u16 	%rs134, 0;
	st.v4.u16 	[%rd224], {%rs133, %rs133, %rs133, %rs134};

BB0_125:
	ld.global.u32 	%r231, [additive];
	setp.eq.s32	%p135, %r231, 0;
	@%p135 bra 	BB0_127;

	mov.u64 	%rd243, image_RNM2;
	cvta.global.u64 	%rd232, %rd243;
	mov.u32 	%r235, 8;
	// inline asm
	call (%rd231), _rt_buffer_get_64, (%rd232, %r28, %r235, %rd6, %rd7, %rd13, %rd13);
	// inline asm
	ld.v4.u16 	{%rs141, %rs142, %rs143, %rs144}, [%rd231];
	// inline asm
	{  cvt.f32.f16 %f1340, %rs141;}

	// inline asm
	// inline asm
	{  cvt.f32.f16 %f1341, %rs142;}

	// inline asm
	// inline asm
	{  cvt.f32.f16 %f1342, %rs143;}

	// inline asm
	// inline asm
	call (%rd237), _rt_buffer_get_64, (%rd232, %r28, %r235, %rd6, %rd7, %rd13, %rd13);
	// inline asm
	add.f32 	%f1343, %f1340, 0f00000000;
	add.f32 	%f1344, %f1341, 0f00000000;
	add.f32 	%f1345, %f1342, 0f00000000;
	// inline asm
	{  cvt.rn.f16.f32 %rs140, %f1345;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs139, %f1344;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs138, %f1343;}

	// inline asm
	mov.u16 	%rs145, 0;
	st.v4.u16 	[%rd237], {%rs138, %rs139, %rs140, %rs145};
	bra.uni 	BB0_128;

BB0_127:
	mov.u64 	%rd250, image_RNM2;
	cvta.global.u64 	%rd245, %rd250;
	mov.u32 	%r237, 8;
	// inline asm
	call (%rd244), _rt_buffer_get_64, (%rd245, %r28, %r237, %rd6, %rd7, %rd13, %rd13);
	// inline asm
	mov.f32 	%f1346, 0f00000000;
	// inline asm
	{  cvt.rn.f16.f32 %rs146, %f1346;}

	// inline asm
	mov.u16 	%rs147, 0;
	st.v4.u16 	[%rd244], {%rs146, %rs146, %rs146, %rs147};

BB0_128:
	ld.global.u32 	%r238, [additive];
	setp.eq.s32	%p136, %r238, 0;
	@%p136 bra 	BB0_130;

	mov.u64 	%rd263, image_RNM3;
	cvta.global.u64 	%rd252, %rd263;
	mov.u32 	%r242, 8;
	// inline asm
	call (%rd251), _rt_buffer_get_64, (%rd252, %r28, %r242, %rd6, %rd7, %rd13, %rd13);
	// inline asm
	ld.v4.u16 	{%rs154, %rs155, %rs156, %rs157}, [%rd251];
	// inline asm
	{  cvt.f32.f16 %f1347, %rs154;}

	// inline asm
	// inline asm
	{  cvt.f32.f16 %f1348, %rs155;}

	// inline asm
	// inline asm
	{  cvt.f32.f16 %f1349, %rs156;}

	// inline asm
	// inline asm
	call (%rd257), _rt_buffer_get_64, (%rd252, %r28, %r242, %rd6, %rd7, %rd13, %rd13);
	// inline asm
	add.f32 	%f1350, %f1347, 0f00000000;
	add.f32 	%f1351, %f1348, 0f00000000;
	add.f32 	%f1352, %f1349, 0f00000000;
	// inline asm
	{  cvt.rn.f16.f32 %rs153, %f1352;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs152, %f1351;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs151, %f1350;}

	// inline asm
	mov.u16 	%rs158, 0;
	st.v4.u16 	[%rd257], {%rs151, %rs152, %rs153, %rs158};
	bra.uni 	BB0_131;

BB0_130:
	mov.u64 	%rd270, image_RNM3;
	cvta.global.u64 	%rd265, %rd270;
	mov.u32 	%r244, 8;
	// inline asm
	call (%rd264), _rt_buffer_get_64, (%rd265, %r28, %r244, %rd6, %rd7, %rd13, %rd13);
	// inline asm
	mov.f32 	%f1353, 0f00000000;
	// inline asm
	{  cvt.rn.f16.f32 %rs159, %f1353;}

	// inline asm
	mov.u16 	%rs160, 0;
	st.v4.u16 	[%rd264], {%rs159, %rs159, %rs159, %rs160};
	bra.uni 	BB0_131;

BB0_51:
	setp.geu.f32	%p67, %f214, 0f00000000;
	@%p67 bra 	BB0_54;

	cvt.rzi.f32.f32	%f1004, %f976;
	setp.neu.f32	%p68, %f1004, 0f3EE8BA2E;
	selp.f32	%f1443, 0f7FFFFFFF, %f1443, %p68;

BB0_54:
	add.f32 	%f1006, %f216, 0f3EE8BA2E;
	mov.b32 	 %r131, %f1006;
	setp.lt.s32	%p70, %r131, 2139095040;
	@%p70 bra 	BB0_59;

	setp.gtu.f32	%p71, %f216, 0f7F800000;
	@%p71 bra 	BB0_58;
	bra.uni 	BB0_56;

BB0_58:
	add.f32 	%f1443, %f214, 0f3EE8BA2E;
	bra.uni 	BB0_59;

BB0_56:
	setp.neu.f32	%p72, %f216, 0f7F800000;
	@%p72 bra 	BB0_59;

	selp.f32	%f1443, 0fFF800000, 0f7F800000, %p4;

BB0_59:
	mul.f32 	%f1007, %f1443, 0f437F0000;
	setp.eq.f32	%p73, %f214, 0f3F800000;
	selp.f32	%f1008, 0f437F0000, %f1007, %p73;
	cvt.rzi.u32.f32	%r132, %f1008;
	cvt.u16.u32	%rs17, %r132;
	mov.u16 	%rs18, 255;
	st.v2.u8 	[%rd43], {%rs17, %rs18};
	ld.global.u32 	%r251, [imageEnabled];

BB0_60:
	and.b32  	%r133, %r251, 1;
	setp.eq.b32	%p74, %r133, 1;
	@!%p74 bra 	BB0_95;
	bra.uni 	BB0_61;

BB0_61:
	mov.f32 	%f1011, 0f3E666666;
	cvt.rzi.f32.f32	%f1012, %f1011;
	fma.rn.f32 	%f1013, %f1012, 0fC0000000, 0f3EE66666;
	abs.f32 	%f228, %f1013;
	abs.f32 	%f229, %f1375;
	setp.lt.f32	%p75, %f229, 0f00800000;
	mul.f32 	%f1014, %f229, 0f4B800000;
	selp.f32	%f1015, 0fC3170000, 0fC2FE0000, %p75;
	selp.f32	%f1016, %f1014, %f229, %p75;
	mov.b32 	 %r134, %f1016;
	and.b32  	%r135, %r134, 8388607;
	or.b32  	%r136, %r135, 1065353216;
	mov.b32 	 %f1017, %r136;
	shr.u32 	%r137, %r134, 23;
	cvt.rn.f32.u32	%f1018, %r137;
	add.f32 	%f1019, %f1015, %f1018;
	setp.gt.f32	%p76, %f1017, 0f3FB504F3;
	mul.f32 	%f1020, %f1017, 0f3F000000;
	add.f32 	%f1021, %f1019, 0f3F800000;
	selp.f32	%f1022, %f1020, %f1017, %p76;
	selp.f32	%f1023, %f1021, %f1019, %p76;
	add.f32 	%f1024, %f1022, 0fBF800000;
	add.f32 	%f1010, %f1022, 0f3F800000;
	// inline asm
	rcp.approx.ftz.f32 %f1009,%f1010;
	// inline asm
	add.f32 	%f1025, %f1024, %f1024;
	mul.f32 	%f1026, %f1009, %f1025;
	mul.f32 	%f1027, %f1026, %f1026;
	mov.f32 	%f1028, 0f3C4CAF63;
	mov.f32 	%f1029, 0f3B18F0FE;
	fma.rn.f32 	%f1030, %f1029, %f1027, %f1028;
	mov.f32 	%f1031, 0f3DAAAABD;
	fma.rn.f32 	%f1032, %f1030, %f1027, %f1031;
	mul.rn.f32 	%f1033, %f1032, %f1027;
	mul.rn.f32 	%f1034, %f1033, %f1026;
	sub.f32 	%f1035, %f1024, %f1026;
	neg.f32 	%f1036, %f1026;
	add.f32 	%f1037, %f1035, %f1035;
	fma.rn.f32 	%f1038, %f1036, %f1024, %f1037;
	mul.rn.f32 	%f1039, %f1009, %f1038;
	add.f32 	%f1040, %f1034, %f1026;
	sub.f32 	%f1041, %f1026, %f1040;
	add.f32 	%f1042, %f1034, %f1041;
	add.f32 	%f1043, %f1039, %f1042;
	add.f32 	%f1044, %f1040, %f1043;
	sub.f32 	%f1045, %f1040, %f1044;
	add.f32 	%f1046, %f1043, %f1045;
	mov.f32 	%f1047, 0f3F317200;
	mul.rn.f32 	%f1048, %f1023, %f1047;
	mov.f32 	%f1049, 0f35BFBE8E;
	mul.rn.f32 	%f1050, %f1023, %f1049;
	add.f32 	%f1051, %f1048, %f1044;
	sub.f32 	%f1052, %f1048, %f1051;
	add.f32 	%f1053, %f1044, %f1052;
	add.f32 	%f1054, %f1046, %f1053;
	add.f32 	%f1055, %f1050, %f1054;
	add.f32 	%f1056, %f1051, %f1055;
	sub.f32 	%f1057, %f1051, %f1056;
	add.f32 	%f1058, %f1055, %f1057;
	mov.f32 	%f1059, 0f3EE66666;
	mul.rn.f32 	%f1060, %f1059, %f1056;
	neg.f32 	%f1061, %f1060;
	fma.rn.f32 	%f1062, %f1059, %f1056, %f1061;
	fma.rn.f32 	%f1063, %f1059, %f1058, %f1062;
	mov.f32 	%f1064, 0f00000000;
	fma.rn.f32 	%f1065, %f1064, %f1056, %f1063;
	add.rn.f32 	%f1066, %f1060, %f1065;
	neg.f32 	%f1067, %f1066;
	add.rn.f32 	%f1068, %f1060, %f1067;
	add.rn.f32 	%f1069, %f1068, %f1065;
	mov.b32 	 %r138, %f1066;
	setp.eq.s32	%p77, %r138, 1118925336;
	add.s32 	%r139, %r138, -1;
	mov.b32 	 %f1070, %r139;
	add.f32 	%f1071, %f1069, 0f37000000;
	selp.f32	%f1072, %f1070, %f1066, %p77;
	selp.f32	%f230, %f1071, %f1069, %p77;
	mul.f32 	%f1073, %f1072, 0f3FB8AA3B;
	cvt.rzi.f32.f32	%f1074, %f1073;
	mov.f32 	%f1075, 0fBF317200;
	fma.rn.f32 	%f1076, %f1074, %f1075, %f1072;
	mov.f32 	%f1077, 0fB5BFBE8E;
	fma.rn.f32 	%f1078, %f1074, %f1077, %f1076;
	mul.f32 	%f1079, %f1078, 0f3FB8AA3B;
	ex2.approx.ftz.f32 	%f1080, %f1079;
	add.f32 	%f1081, %f1074, 0f00000000;
	ex2.approx.f32 	%f1082, %f1081;
	mul.f32 	%f1083, %f1080, %f1082;
	setp.lt.f32	%p78, %f1072, 0fC2D20000;
	selp.f32	%f1084, 0f00000000, %f1083, %p78;
	setp.gt.f32	%p79, %f1072, 0f42D20000;
	selp.f32	%f1444, 0f7F800000, %f1084, %p79;
	setp.eq.f32	%p80, %f1444, 0f7F800000;
	@%p80 bra 	BB0_63;

	fma.rn.f32 	%f1444, %f1444, %f230, %f1444;

BB0_63:
	setp.lt.f32	%p81, %f1375, 0f00000000;
	setp.eq.f32	%p82, %f228, 0f3F800000;
	and.pred  	%p5, %p81, %p82;
	mov.b32 	 %r140, %f1444;
	xor.b32  	%r141, %r140, -2147483648;
	mov.b32 	 %f1085, %r141;
	selp.f32	%f1446, %f1085, %f1444, %p5;
	setp.eq.f32	%p83, %f1375, 0f00000000;
	@%p83 bra 	BB0_66;
	bra.uni 	BB0_64;

BB0_66:
	add.f32 	%f1088, %f1375, %f1375;
	selp.f32	%f1446, %f1088, 0f00000000, %p82;
	bra.uni 	BB0_67;

BB0_64:
	setp.geu.f32	%p84, %f1375, 0f00000000;
	@%p84 bra 	BB0_67;

	cvt.rzi.f32.f32	%f1087, %f1059;
	setp.neu.f32	%p85, %f1087, 0f3EE66666;
	selp.f32	%f1446, 0f7FFFFFFF, %f1446, %p85;

BB0_67:
	add.f32 	%f1089, %f229, 0f3EE66666;
	mov.b32 	 %r142, %f1089;
	setp.lt.s32	%p87, %r142, 2139095040;
	@%p87 bra 	BB0_72;

	setp.gtu.f32	%p88, %f229, 0f7F800000;
	@%p88 bra 	BB0_71;
	bra.uni 	BB0_69;

BB0_71:
	add.f32 	%f1446, %f1375, 0f3EE66666;
	bra.uni 	BB0_72;

BB0_69:
	setp.neu.f32	%p89, %f229, 0f7F800000;
	@%p89 bra 	BB0_72;

	selp.f32	%f1446, 0fFF800000, 0f7F800000, %p5;

BB0_72:
	setp.eq.f32	%p90, %f1375, 0f3F800000;
	selp.f32	%f241, 0f3F800000, %f1446, %p90;
	abs.f32 	%f242, %f1374;
	setp.lt.f32	%p91, %f242, 0f00800000;
	mul.f32 	%f1092, %f242, 0f4B800000;
	selp.f32	%f1093, 0fC3170000, 0fC2FE0000, %p91;
	selp.f32	%f1094, %f1092, %f242, %p91;
	mov.b32 	 %r143, %f1094;
	and.b32  	%r144, %r143, 8388607;
	or.b32  	%r145, %r144, 1065353216;
	mov.b32 	 %f1095, %r145;
	shr.u32 	%r146, %r143, 23;
	cvt.rn.f32.u32	%f1096, %r146;
	add.f32 	%f1097, %f1093, %f1096;
	setp.gt.f32	%p92, %f1095, 0f3FB504F3;
	mul.f32 	%f1098, %f1095, 0f3F000000;
	add.f32 	%f1099, %f1097, 0f3F800000;
	selp.f32	%f1100, %f1098, %f1095, %p92;
	selp.f32	%f1101, %f1099, %f1097, %p92;
	add.f32 	%f1102, %f1100, 0fBF800000;
	add.f32 	%f1091, %f1100, 0f3F800000;
	// inline asm
	rcp.approx.ftz.f32 %f1090,%f1091;
	// inline asm
	add.f32 	%f1103, %f1102, %f1102;
	mul.f32 	%f1104, %f1090, %f1103;
	mul.f32 	%f1105, %f1104, %f1104;
	fma.rn.f32 	%f1108, %f1029, %f1105, %f1028;
	fma.rn.f32 	%f1110, %f1108, %f1105, %f1031;
	mul.rn.f32 	%f1111, %f1110, %f1105;
	mul.rn.f32 	%f1112, %f1111, %f1104;
	sub.f32 	%f1113, %f1102, %f1104;
	neg.f32 	%f1114, %f1104;
	add.f32 	%f1115, %f1113, %f1113;
	fma.rn.f32 	%f1116, %f1114, %f1102, %f1115;
	mul.rn.f32 	%f1117, %f1090, %f1116;
	add.f32 	%f1118, %f1112, %f1104;
	sub.f32 	%f1119, %f1104, %f1118;
	add.f32 	%f1120, %f1112, %f1119;
	add.f32 	%f1121, %f1117, %f1120;
	add.f32 	%f1122, %f1118, %f1121;
	sub.f32 	%f1123, %f1118, %f1122;
	add.f32 	%f1124, %f1121, %f1123;
	mul.rn.f32 	%f1126, %f1101, %f1047;
	mul.rn.f32 	%f1128, %f1101, %f1049;
	add.f32 	%f1129, %f1126, %f1122;
	sub.f32 	%f1130, %f1126, %f1129;
	add.f32 	%f1131, %f1122, %f1130;
	add.f32 	%f1132, %f1124, %f1131;
	add.f32 	%f1133, %f1128, %f1132;
	add.f32 	%f1134, %f1129, %f1133;
	sub.f32 	%f1135, %f1129, %f1134;
	add.f32 	%f1136, %f1133, %f1135;
	mul.rn.f32 	%f1138, %f1059, %f1134;
	neg.f32 	%f1139, %f1138;
	fma.rn.f32 	%f1140, %f1059, %f1134, %f1139;
	fma.rn.f32 	%f1141, %f1059, %f1136, %f1140;
	fma.rn.f32 	%f1143, %f1064, %f1134, %f1141;
	add.rn.f32 	%f1144, %f1138, %f1143;
	neg.f32 	%f1145, %f1144;
	add.rn.f32 	%f1146, %f1138, %f1145;
	add.rn.f32 	%f1147, %f1146, %f1143;
	mov.b32 	 %r147, %f1144;
	setp.eq.s32	%p93, %r147, 1118925336;
	add.s32 	%r148, %r147, -1;
	mov.b32 	 %f1148, %r148;
	add.f32 	%f1149, %f1147, 0f37000000;
	selp.f32	%f1150, %f1148, %f1144, %p93;
	selp.f32	%f243, %f1149, %f1147, %p93;
	mul.f32 	%f1151, %f1150, 0f3FB8AA3B;
	cvt.rzi.f32.f32	%f1152, %f1151;
	fma.rn.f32 	%f1154, %f1152, %f1075, %f1150;
	fma.rn.f32 	%f1156, %f1152, %f1077, %f1154;
	mul.f32 	%f1157, %f1156, 0f3FB8AA3B;
	ex2.approx.ftz.f32 	%f1158, %f1157;
	add.f32 	%f1159, %f1152, 0f00000000;
	ex2.approx.f32 	%f1160, %f1159;
	mul.f32 	%f1161, %f1158, %f1160;
	setp.lt.f32	%p94, %f1150, 0fC2D20000;
	selp.f32	%f1162, 0f00000000, %f1161, %p94;
	setp.gt.f32	%p95, %f1150, 0f42D20000;
	selp.f32	%f1447, 0f7F800000, %f1162, %p95;
	setp.eq.f32	%p96, %f1447, 0f7F800000;
	@%p96 bra 	BB0_74;

	fma.rn.f32 	%f1447, %f1447, %f243, %f1447;

BB0_74:
	setp.lt.f32	%p97, %f1374, 0f00000000;
	and.pred  	%p6, %p97, %p82;
	mov.b32 	 %r149, %f1447;
	xor.b32  	%r150, %r149, -2147483648;
	mov.b32 	 %f1163, %r150;
	selp.f32	%f1449, %f1163, %f1447, %p6;
	setp.eq.f32	%p99, %f1374, 0f00000000;
	@%p99 bra 	BB0_77;
	bra.uni 	BB0_75;

BB0_77:
	add.f32 	%f1166, %f1374, %f1374;
	selp.f32	%f1449, %f1166, 0f00000000, %p82;
	bra.uni 	BB0_78;

BB0_75:
	setp.geu.f32	%p100, %f1374, 0f00000000;
	@%p100 bra 	BB0_78;

	cvt.rzi.f32.f32	%f1165, %f1059;
	setp.neu.f32	%p101, %f1165, 0f3EE66666;
	selp.f32	%f1449, 0f7FFFFFFF, %f1449, %p101;

BB0_78:
	add.f32 	%f1167, %f242, 0f3EE66666;
	mov.b32 	 %r151, %f1167;
	setp.lt.s32	%p103, %r151, 2139095040;
	@%p103 bra 	BB0_83;

	setp.gtu.f32	%p104, %f242, 0f7F800000;
	@%p104 bra 	BB0_82;
	bra.uni 	BB0_80;

BB0_82:
	add.f32 	%f1449, %f1374, 0f3EE66666;
	bra.uni 	BB0_83;

BB0_80:
	setp.neu.f32	%p105, %f242, 0f7F800000;
	@%p105 bra 	BB0_83;

	selp.f32	%f1449, 0fFF800000, 0f7F800000, %p6;

BB0_83:
	setp.eq.f32	%p106, %f1374, 0f3F800000;
	selp.f32	%f254, 0f3F800000, %f1449, %p106;
	abs.f32 	%f255, %f1373;
	setp.lt.f32	%p107, %f255, 0f00800000;
	mul.f32 	%f1170, %f255, 0f4B800000;
	selp.f32	%f1171, 0fC3170000, 0fC2FE0000, %p107;
	selp.f32	%f1172, %f1170, %f255, %p107;
	mov.b32 	 %r152, %f1172;
	and.b32  	%r153, %r152, 8388607;
	or.b32  	%r154, %r153, 1065353216;
	mov.b32 	 %f1173, %r154;
	shr.u32 	%r155, %r152, 23;
	cvt.rn.f32.u32	%f1174, %r155;
	add.f32 	%f1175, %f1171, %f1174;
	setp.gt.f32	%p108, %f1173, 0f3FB504F3;
	mul.f32 	%f1176, %f1173, 0f3F000000;
	add.f32 	%f1177, %f1175, 0f3F800000;
	selp.f32	%f1178, %f1176, %f1173, %p108;
	selp.f32	%f1179, %f1177, %f1175, %p108;
	add.f32 	%f1180, %f1178, 0fBF800000;
	add.f32 	%f1169, %f1178, 0f3F800000;
	// inline asm
	rcp.approx.ftz.f32 %f1168,%f1169;
	// inline asm
	add.f32 	%f1181, %f1180, %f1180;
	mul.f32 	%f1182, %f1168, %f1181;
	mul.f32 	%f1183, %f1182, %f1182;
	fma.rn.f32 	%f1186, %f1029, %f1183, %f1028;
	fma.rn.f32 	%f1188, %f1186, %f1183, %f1031;
	mul.rn.f32 	%f1189, %f1188, %f1183;
	mul.rn.f32 	%f1190, %f1189, %f1182;
	sub.f32 	%f1191, %f1180, %f1182;
	neg.f32 	%f1192, %f1182;
	add.f32 	%f1193, %f1191, %f1191;
	fma.rn.f32 	%f1194, %f1192, %f1180, %f1193;
	mul.rn.f32 	%f1195, %f1168, %f1194;
	add.f32 	%f1196, %f1190, %f1182;
	sub.f32 	%f1197, %f1182, %f1196;
	add.f32 	%f1198, %f1190, %f1197;
	add.f32 	%f1199, %f1195, %f1198;
	add.f32 	%f1200, %f1196, %f1199;
	sub.f32 	%f1201, %f1196, %f1200;
	add.f32 	%f1202, %f1199, %f1201;
	mul.rn.f32 	%f1204, %f1179, %f1047;
	mul.rn.f32 	%f1206, %f1179, %f1049;
	add.f32 	%f1207, %f1204, %f1200;
	sub.f32 	%f1208, %f1204, %f1207;
	add.f32 	%f1209, %f1200, %f1208;
	add.f32 	%f1210, %f1202, %f1209;
	add.f32 	%f1211, %f1206, %f1210;
	add.f32 	%f1212, %f1207, %f1211;
	sub.f32 	%f1213, %f1207, %f1212;
	add.f32 	%f1214, %f1211, %f1213;
	mul.rn.f32 	%f1216, %f1059, %f1212;
	neg.f32 	%f1217, %f1216;
	fma.rn.f32 	%f1218, %f1059, %f1212, %f1217;
	fma.rn.f32 	%f1219, %f1059, %f1214, %f1218;
	fma.rn.f32 	%f1221, %f1064, %f1212, %f1219;
	add.rn.f32 	%f1222, %f1216, %f1221;
	neg.f32 	%f1223, %f1222;
	add.rn.f32 	%f1224, %f1216, %f1223;
	add.rn.f32 	%f1225, %f1224, %f1221;
	mov.b32 	 %r156, %f1222;
	setp.eq.s32	%p109, %r156, 1118925336;
	add.s32 	%r157, %r156, -1;
	mov.b32 	 %f1226, %r157;
	add.f32 	%f1227, %f1225, 0f37000000;
	selp.f32	%f1228, %f1226, %f1222, %p109;
	selp.f32	%f256, %f1227, %f1225, %p109;
	mul.f32 	%f1229, %f1228, 0f3FB8AA3B;
	cvt.rzi.f32.f32	%f1230, %f1229;
	fma.rn.f32 	%f1232, %f1230, %f1075, %f1228;
	fma.rn.f32 	%f1234, %f1230, %f1077, %f1232;
	mul.f32 	%f1235, %f1234, 0f3FB8AA3B;
	ex2.approx.ftz.f32 	%f1236, %f1235;
	add.f32 	%f1237, %f1230, 0f00000000;
	ex2.approx.f32 	%f1238, %f1237;
	mul.f32 	%f1239, %f1236, %f1238;
	setp.lt.f32	%p110, %f1228, 0fC2D20000;
	selp.f32	%f1240, 0f00000000, %f1239, %p110;
	setp.gt.f32	%p111, %f1228, 0f42D20000;
	selp.f32	%f1450, 0f7F800000, %f1240, %p111;
	setp.eq.f32	%p112, %f1450, 0f7F800000;
	@%p112 bra 	BB0_85;

	fma.rn.f32 	%f1450, %f1450, %f256, %f1450;

BB0_85:
	setp.lt.f32	%p113, %f1373, 0f00000000;
	and.pred  	%p7, %p113, %p82;
	mov.b32 	 %r158, %f1450;
	xor.b32  	%r159, %r158, -2147483648;
	mov.b32 	 %f1241, %r159;
	selp.f32	%f1452, %f1241, %f1450, %p7;
	setp.eq.f32	%p115, %f1373, 0f00000000;
	@%p115 bra 	BB0_88;
	bra.uni 	BB0_86;

BB0_88:
	add.f32 	%f1244, %f1373, %f1373;
	selp.f32	%f1452, %f1244, 0f00000000, %p82;
	bra.uni 	BB0_89;

BB0_86:
	setp.geu.f32	%p116, %f1373, 0f00000000;
	@%p116 bra 	BB0_89;

	cvt.rzi.f32.f32	%f1243, %f1059;
	setp.neu.f32	%p117, %f1243, 0f3EE66666;
	selp.f32	%f1452, 0f7FFFFFFF, %f1452, %p117;

BB0_89:
	add.f32 	%f1245, %f255, 0f3EE66666;
	mov.b32 	 %r160, %f1245;
	setp.lt.s32	%p119, %r160, 2139095040;
	@%p119 bra 	BB0_94;

	setp.gtu.f32	%p120, %f255, 0f7F800000;
	@%p120 bra 	BB0_93;
	bra.uni 	BB0_91;

BB0_93:
	add.f32 	%f1452, %f1373, 0f3EE66666;
	bra.uni 	BB0_94;

BB0_91:
	setp.neu.f32	%p121, %f255, 0f7F800000;
	@%p121 bra 	BB0_94;

	selp.f32	%f1452, 0fFF800000, 0f7F800000, %p7;

BB0_94:
	setp.eq.f32	%p122, %f1373, 0f3F800000;
	selp.f32	%f1246, 0f3F800000, %f1452, %p122;
	cvt.u64.u32	%rd53, %r3;
	cvt.u64.u32	%rd52, %r2;
	mov.u64 	%rd56, image;
	cvta.global.u64 	%rd51, %rd56;
	// inline asm
	call (%rd50), _rt_buffer_get_64, (%rd51, %r28, %r29, %rd52, %rd53, %rd13, %rd13);
	// inline asm
	cvt.sat.f32.f32	%f1247, %f1246;
	mul.f32 	%f1248, %f1247, 0f437FFD71;
	cvt.rzi.u32.f32	%r163, %f1248;
	cvt.sat.f32.f32	%f1249, %f254;
	mul.f32 	%f1250, %f1249, 0f437FFD71;
	cvt.rzi.u32.f32	%r164, %f1250;
	cvt.sat.f32.f32	%f1251, %f241;
	mul.f32 	%f1252, %f1251, 0f437FFD71;
	cvt.rzi.u32.f32	%r165, %f1252;
	cvt.u16.u32	%rs19, %r163;
	cvt.u16.u32	%rs20, %r165;
	cvt.u16.u32	%rs21, %r164;
	mov.u16 	%rs22, 255;
	st.v4.u8 	[%rd50], {%rs19, %rs21, %rs20, %rs22};
	ld.global.u32 	%r251, [imageEnabled];

BB0_95:
	cvt.u64.u32	%rd4, %r2;
	cvt.u64.u32	%rd5, %r3;
	and.b32  	%r166, %r251, 4;
	setp.eq.s32	%p123, %r166, 0;
	@%p123 bra 	BB0_99;

	ld.global.u32 	%r167, [additive];
	setp.eq.s32	%p124, %r167, 0;
	mov.f32 	%f1253, 0f3F800000;
	// inline asm
	{  cvt.rn.f16.f32 %rs23, %f1253;}

	// inline asm
	@%p124 bra 	BB0_98;

	mov.u64 	%rd69, image_HDR;
	cvta.global.u64 	%rd58, %rd69;
	mov.u32 	%r171, 8;
	// inline asm
	call (%rd57), _rt_buffer_get_64, (%rd58, %r28, %r171, %rd4, %rd5, %rd13, %rd13);
	// inline asm
	ld.v4.u16 	{%rs30, %rs31, %rs32, %rs33}, [%rd57];
	// inline asm
	{  cvt.f32.f16 %f1254, %rs30;}

	// inline asm
	// inline asm
	{  cvt.f32.f16 %f1255, %rs31;}

	// inline asm
	// inline asm
	{  cvt.f32.f16 %f1256, %rs32;}

	// inline asm
	// inline asm
	call (%rd63), _rt_buffer_get_64, (%rd58, %r28, %r171, %rd4, %rd5, %rd13, %rd13);
	// inline asm
	add.f32 	%f1257, %f1375, %f1254;
	add.f32 	%f1258, %f1374, %f1255;
	add.f32 	%f1259, %f1373, %f1256;
	// inline asm
	{  cvt.rn.f16.f32 %rs29, %f1259;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs28, %f1258;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs27, %f1257;}

	// inline asm
	st.v4.u16 	[%rd63], {%rs27, %rs28, %rs29, %rs23};
	bra.uni 	BB0_99;

BB0_98:
	mov.u64 	%rd76, image_HDR;
	cvta.global.u64 	%rd71, %rd76;
	mov.u32 	%r173, 8;
	// inline asm
	call (%rd70), _rt_buffer_get_64, (%rd71, %r28, %r173, %rd4, %rd5, %rd13, %rd13);
	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs36, %f1373;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs35, %f1374;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs34, %f1375;}

	// inline asm
	st.v4.u16 	[%rd70], {%rs34, %rs35, %rs36, %rs23};

BB0_99:
	mov.f32 	%f1264, 0f34000000;
	max.f32 	%f1265, %f1372, %f1264;
	div.rn.f32 	%f1266, %f1369, %f1265;
	max.f32 	%f1267, %f1371, %f1264;
	div.rn.f32 	%f1268, %f1368, %f1267;
	max.f32 	%f1269, %f1370, %f1264;
	div.rn.f32 	%f1270, %f1367, %f1269;
	fma.rn.f32 	%f267, %f1266, 0f3F000000, 0f3F000000;
	fma.rn.f32 	%f268, %f1268, 0f3F000000, 0f3F000000;
	fma.rn.f32 	%f269, %f1270, 0f3F000000, 0f3F000000;
	div.rn.f32 	%f1271, %f1366, %f1265;
	div.rn.f32 	%f1272, %f1365, %f1267;
	div.rn.f32 	%f1273, %f1364, %f1269;
	fma.rn.f32 	%f270, %f1271, 0f3F000000, 0f3F000000;
	fma.rn.f32 	%f271, %f1272, 0f3F000000, 0f3F000000;
	fma.rn.f32 	%f272, %f1273, 0f3F000000, 0f3F000000;
	div.rn.f32 	%f1274, %f1363, %f1265;
	div.rn.f32 	%f1275, %f1362, %f1267;
	div.rn.f32 	%f1276, %f1361, %f1269;
	fma.rn.f32 	%f273, %f1274, 0f3F000000, 0f3F000000;
	fma.rn.f32 	%f274, %f1275, 0f3F000000, 0f3F000000;
	fma.rn.f32 	%f275, %f1276, 0f3F000000, 0f3F000000;
	ld.global.u32 	%r174, [additive];
	setp.eq.s32	%p125, %r174, 0;
	mov.f32 	%f1263, 0f3F800000;
	// inline asm
	{  cvt.rn.f16.f32 %rs37, %f1263;}

	// inline asm
	@%p125 bra 	BB0_101;

	mov.u64 	%rd89, image_RNM0;
	cvta.global.u64 	%rd78, %rd89;
	mov.u32 	%r178, 8;
	// inline asm
	call (%rd77), _rt_buffer_get_64, (%rd78, %r28, %r178, %rd4, %rd5, %rd13, %rd13);
	// inline asm
	ld.v4.u16 	{%rs44, %rs45, %rs46, %rs47}, [%rd77];
	// inline asm
	{  cvt.f32.f16 %f1277, %rs44;}

	// inline asm
	// inline asm
	{  cvt.f32.f16 %f1278, %rs45;}

	// inline asm
	// inline asm
	{  cvt.f32.f16 %f1279, %rs46;}

	// inline asm
	// inline asm
	call (%rd83), _rt_buffer_get_64, (%rd78, %r28, %r178, %rd4, %rd5, %rd13, %rd13);
	// inline asm
	add.f32 	%f1280, %f1372, %f1277;
	add.f32 	%f1281, %f1371, %f1278;
	add.f32 	%f1282, %f1370, %f1279;
	// inline asm
	{  cvt.rn.f16.f32 %rs43, %f1282;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs42, %f1281;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs41, %f1280;}

	// inline asm
	st.v4.u16 	[%rd83], {%rs41, %rs42, %rs43, %rs37};
	bra.uni 	BB0_102;

BB0_101:
	mov.u64 	%rd96, image_RNM0;
	cvta.global.u64 	%rd91, %rd96;
	mov.u32 	%r180, 8;
	// inline asm
	call (%rd90), _rt_buffer_get_64, (%rd91, %r28, %r180, %rd4, %rd5, %rd13, %rd13);
	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs50, %f1370;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs49, %f1371;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs48, %f1372;}

	// inline asm
	st.v4.u16 	[%rd90], {%rs48, %rs49, %rs50, %rs37};

BB0_102:
	ld.global.u32 	%r181, [additive];
	setp.eq.s32	%p126, %r181, 0;
	// inline asm
	{  cvt.rn.f16.f32 %rs51, %f1263;}

	// inline asm
	@%p126 bra 	BB0_104;

	mov.u64 	%rd109, image_RNM1;
	cvta.global.u64 	%rd98, %rd109;
	mov.u32 	%r185, 8;
	// inline asm
	call (%rd97), _rt_buffer_get_64, (%rd98, %r28, %r185, %rd4, %rd5, %rd13, %rd13);
	// inline asm
	ld.v4.u16 	{%rs58, %rs59, %rs60, %rs61}, [%rd97];
	// inline asm
	{  cvt.f32.f16 %f1287, %rs58;}

	// inline asm
	// inline asm
	{  cvt.f32.f16 %f1288, %rs59;}

	// inline asm
	// inline asm
	{  cvt.f32.f16 %f1289, %rs60;}

	// inline asm
	// inline asm
	call (%rd103), _rt_buffer_get_64, (%rd98, %r28, %r185, %rd4, %rd5, %rd13, %rd13);
	// inline asm
	add.f32 	%f1290, %f267, %f1287;
	add.f32 	%f1291, %f268, %f1288;
	add.f32 	%f1292, %f269, %f1289;
	// inline asm
	{  cvt.rn.f16.f32 %rs57, %f1292;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs56, %f1291;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs55, %f1290;}

	// inline asm
	st.v4.u16 	[%rd103], {%rs55, %rs56, %rs57, %rs51};
	bra.uni 	BB0_105;

BB0_104:
	mov.u64 	%rd116, image_RNM1;
	cvta.global.u64 	%rd111, %rd116;
	mov.u32 	%r187, 8;
	// inline asm
	call (%rd110), _rt_buffer_get_64, (%rd111, %r28, %r187, %rd4, %rd5, %rd13, %rd13);
	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs64, %f269;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs63, %f268;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs62, %f267;}

	// inline asm
	st.v4.u16 	[%rd110], {%rs62, %rs63, %rs64, %rs51};

BB0_105:
	ld.global.u32 	%r188, [additive];
	setp.eq.s32	%p127, %r188, 0;
	// inline asm
	{  cvt.rn.f16.f32 %rs65, %f1263;}

	// inline asm
	@%p127 bra 	BB0_107;

	mov.u64 	%rd129, image_RNM2;
	cvta.global.u64 	%rd118, %rd129;
	mov.u32 	%r192, 8;
	// inline asm
	call (%rd117), _rt_buffer_get_64, (%rd118, %r28, %r192, %rd4, %rd5, %rd13, %rd13);
	// inline asm
	ld.v4.u16 	{%rs72, %rs73, %rs74, %rs75}, [%rd117];
	// inline asm
	{  cvt.f32.f16 %f1297, %rs72;}

	// inline asm
	// inline asm
	{  cvt.f32.f16 %f1298, %rs73;}

	// inline asm
	// inline asm
	{  cvt.f32.f16 %f1299, %rs74;}

	// inline asm
	// inline asm
	call (%rd123), _rt_buffer_get_64, (%rd118, %r28, %r192, %rd4, %rd5, %rd13, %rd13);
	// inline asm
	add.f32 	%f1300, %f270, %f1297;
	add.f32 	%f1301, %f271, %f1298;
	add.f32 	%f1302, %f272, %f1299;
	// inline asm
	{  cvt.rn.f16.f32 %rs71, %f1302;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs70, %f1301;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs69, %f1300;}

	// inline asm
	st.v4.u16 	[%rd123], {%rs69, %rs70, %rs71, %rs65};
	bra.uni 	BB0_108;

BB0_107:
	mov.u64 	%rd136, image_RNM2;
	cvta.global.u64 	%rd131, %rd136;
	mov.u32 	%r194, 8;
	// inline asm
	call (%rd130), _rt_buffer_get_64, (%rd131, %r28, %r194, %rd4, %rd5, %rd13, %rd13);
	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs78, %f272;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs77, %f271;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs76, %f270;}

	// inline asm
	st.v4.u16 	[%rd130], {%rs76, %rs77, %rs78, %rs65};

BB0_108:
	ld.global.u32 	%r195, [additive];
	setp.eq.s32	%p128, %r195, 0;
	// inline asm
	{  cvt.rn.f16.f32 %rs79, %f1263;}

	// inline asm
	@%p128 bra 	BB0_110;

	mov.u64 	%rd149, image_RNM3;
	cvta.global.u64 	%rd138, %rd149;
	mov.u32 	%r199, 8;
	// inline asm
	call (%rd137), _rt_buffer_get_64, (%rd138, %r28, %r199, %rd4, %rd5, %rd13, %rd13);
	// inline asm
	ld.v4.u16 	{%rs86, %rs87, %rs88, %rs89}, [%rd137];
	// inline asm
	{  cvt.f32.f16 %f1307, %rs86;}

	// inline asm
	// inline asm
	{  cvt.f32.f16 %f1308, %rs87;}

	// inline asm
	// inline asm
	{  cvt.f32.f16 %f1309, %rs88;}

	// inline asm
	// inline asm
	call (%rd143), _rt_buffer_get_64, (%rd138, %r28, %r199, %rd4, %rd5, %rd13, %rd13);
	// inline asm
	add.f32 	%f1310, %f273, %f1307;
	add.f32 	%f1311, %f274, %f1308;
	add.f32 	%f1312, %f275, %f1309;
	// inline asm
	{  cvt.rn.f16.f32 %rs85, %f1312;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs84, %f1311;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs83, %f1310;}

	// inline asm
	st.v4.u16 	[%rd143], {%rs83, %rs84, %rs85, %rs79};
	bra.uni 	BB0_131;

BB0_110:
	mov.u64 	%rd156, image_RNM3;
	cvta.global.u64 	%rd151, %rd156;
	mov.u32 	%r201, 8;
	// inline asm
	call (%rd150), _rt_buffer_get_64, (%rd151, %r28, %r201, %rd4, %rd5, %rd13, %rd13);
	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs92, %f275;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs91, %f274;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs90, %f273;}

	// inline asm
	st.v4.u16 	[%rd150], {%rs90, %rs91, %rs92, %rs79};

BB0_131:
	ret;
}