//
// 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[12];
	.reg .b64 	%SP;
	.reg .b64 	%SPL;
	.reg .pred 	%p<138>;
	.reg .b16 	%rs<162>;
	.reg .f32 	%f<1483>;
	.reg .b32 	%r<251>;
	.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 	%f1375, 0f00000000;
	mov.f32 	%f1376, %f1375;
	mov.f32 	%f1377, %f1375;
	@%p8 bra 	BB0_2;

	ld.u8 	%rs12, [%rd8+1];
	and.b16  	%rs14, %rs10, 255;
	cvt.rn.f32.u16	%f297, %rs14;
	div.rn.f32 	%f298, %f297, 0f437F0000;
	fma.rn.f32 	%f299, %f298, 0f40000000, 0fBF800000;
	cvt.rn.f32.u16	%f300, %rs12;
	div.rn.f32 	%f301, %f300, 0f437F0000;
	fma.rn.f32 	%f302, %f301, 0f40000000, 0fBF800000;
	cvt.rn.f32.u16	%f303, %rs9;
	div.rn.f32 	%f304, %f303, 0f437F0000;
	fma.rn.f32 	%f305, %f304, 0f40000000, 0fBF800000;
	mul.f32 	%f306, %f302, %f302;
	fma.rn.f32 	%f307, %f299, %f299, %f306;
	fma.rn.f32 	%f308, %f305, %f305, %f307;
	sqrt.rn.f32 	%f309, %f308;
	rcp.rn.f32 	%f310, %f309;
	mul.f32 	%f1375, %f299, %f310;
	mul.f32 	%f1376, %f302, %f310;
	mul.f32 	%f1377, %f305, %f310;

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, %f1376, 0f00000000;
	setp.eq.f32	%p10, %f1375, 0f00000000;
	and.pred  	%p11, %p10, %p9;
	setp.eq.f32	%p12, %f1377, 0f00000000;
	and.pred  	%p13, %p11, %p12;
	@%p13 bra 	BB0_111;
	bra.uni 	BB0_3;

BB0_111:
	ld.global.u32 	%r250, [imageEnabled];
	and.b32  	%r198, %r250, 1;
	setp.eq.b32	%p129, %r198, 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 	%r250, [imageEnabled];

BB0_113:
	and.b32  	%r201, %r250, 8;
	setp.eq.s32	%p130, %r201, 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 	%f1336, 0f00000000;
	cvt.rzi.u32.f32	%r204, %f1336;
	cvt.u16.u32	%rs94, %r204;
	mov.u16 	%rs95, 0;
	st.v2.u8 	[%rd164], {%rs94, %rs95};
	ld.global.u32 	%r250, [imageEnabled];

BB0_115:
	cvt.u64.u32	%rd6, %r2;
	cvt.u64.u32	%rd7, %r3;
	and.b32  	%r205, %r250, 4;
	setp.eq.s32	%p131, %r205, 0;
	@%p131 bra 	BB0_119;

	ld.global.u32 	%r206, [additive];
	setp.eq.s32	%p132, %r206, 0;
	@%p132 bra 	BB0_118;

	mov.u64 	%rd183, image_HDR;
	cvta.global.u64 	%rd172, %rd183;
	mov.u32 	%r210, 8;
	// inline asm
	call (%rd171), _rt_buffer_get_64, (%rd172, %r28, %r210, %rd6, %rd7, %rd13, %rd13);
	// inline asm
	ld.v4.u16 	{%rs102, %rs103, %rs104, %rs105}, [%rd171];
	// inline asm
	{  cvt.f32.f16 %f1337, %rs102;}

	// inline asm
	// inline asm
	{  cvt.f32.f16 %f1338, %rs103;}

	// inline asm
	// inline asm
	{  cvt.f32.f16 %f1339, %rs104;}

	// inline asm
	// inline asm
	call (%rd177), _rt_buffer_get_64, (%rd172, %r28, %r210, %rd6, %rd7, %rd13, %rd13);
	// inline asm
	add.f32 	%f1340, %f1337, 0f00000000;
	add.f32 	%f1341, %f1338, 0f00000000;
	add.f32 	%f1342, %f1339, 0f00000000;
	// inline asm
	{  cvt.rn.f16.f32 %rs101, %f1342;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs100, %f1341;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs99, %f1340;}

	// 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 	%f327, %f7, 0f3456BF95;
	mul.f32 	%f328, %f8, 0f3456BF95;
	mul.f32 	%f329, %f9, 0f3456BF95;
	abs.f32 	%f330, %f1375;
	div.rn.f32 	%f331, %f327, %f330;
	abs.f32 	%f332, %f1376;
	div.rn.f32 	%f333, %f328, %f332;
	abs.f32 	%f334, %f1377;
	div.rn.f32 	%f335, %f329, %f334;
	abs.f32 	%f336, %f331;
	abs.f32 	%f337, %f333;
	abs.f32 	%f338, %f335;
	mov.f32 	%f339, 0f38D1B717;
	max.f32 	%f340, %f336, %f339;
	max.f32 	%f341, %f337, %f339;
	max.f32 	%f342, %f338, %f339;
	fma.rn.f32 	%f10, %f1375, %f340, %f7;
	fma.rn.f32 	%f11, %f1376, %f341, %f8;
	fma.rn.f32 	%f12, %f1377, %f342, %f9;
	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 	%f1378, 0f00000000;
	mov.f32 	%f1379, %f1378;
	mov.f32 	%f1380, %f1378;
	mov.f32 	%f1381, %f1378;
	mov.f32 	%f1382, %f1378;
	mov.f32 	%f1383, %f1378;
	mov.f32 	%f1384, %f1378;
	mov.f32 	%f1385, %f1378;
	mov.f32 	%f1386, %f1378;
	mov.f32 	%f1387, %f1378;
	mov.f32 	%f1388, %f1378;
	mov.f32 	%f1389, %f1378;
	mov.f32 	%f1390, %f1378;
	mov.f32 	%f1391, %f1378;
	mov.f32 	%f1392, %f1378;
	mov.f32 	%f1393, %f1378;
	@%p14 bra 	BB0_47;

	mov.f32 	%f359, 0f40000000;
	cvt.rzi.f32.f32	%f360, %f359;
	add.f32 	%f361, %f360, %f360;
	mov.f32 	%f362, 0f40800000;
	sub.f32 	%f363, %f362, %f361;
	abs.f32 	%f13, %f363;
	mul.f32 	%f14, %f10, 0f3456BF95;
	mul.f32 	%f15, %f11, 0f3456BF95;
	mul.f32 	%f16, %f12, 0f3456BF95;
	mov.f32 	%f358, 0f00000000;
	mov.u32 	%r242, 0;
	abs.f32 	%f560, %f14;
	abs.f32 	%f561, %f15;
	max.f32 	%f562, %f560, %f561;
	abs.f32 	%f563, %f16;
	max.f32 	%f564, %f562, %f563;
	mov.f32 	%f1378, %f358;
	mov.f32 	%f1379, %f358;
	mov.f32 	%f1380, %f358;
	mov.f32 	%f1381, %f358;
	mov.f32 	%f1382, %f358;
	mov.f32 	%f1383, %f358;
	mov.f32 	%f1384, %f358;
	mov.f32 	%f1385, %f358;
	mov.f32 	%f1386, %f358;
	mov.f32 	%f1387, %f358;
	mov.f32 	%f1388, %f358;
	mov.f32 	%f1389, %f358;
	mov.f32 	%f1390, %f358;
	mov.f32 	%f1391, %f358;
	mov.f32 	%f1392, %f358;
	mov.f32 	%f1393, %f358;

BB0_5:
	cvt.u64.u32	%rd30, %r242;
	// inline asm
	call (%rd28), _rt_buffer_get_64, (%rd25, %r45, %r46, %rd30, %rd13, %rd13, %rd13);
	// inline asm
	ld.v4.f32 	{%f366, %f367, %f368, %f369}, [%rd28+80];
	ld.v4.f32 	{%f370, %f371, %f372, %f373}, [%rd28+64];
	ld.v4.f32 	{%f374, %f375, %f376, %f377}, [%rd28+48];
	ld.v4.f32 	{%f378, %f1402, %f1403, %f381}, [%rd28+32];
	ld.v4.f32 	{%f382, %f383, %f384, %f385}, [%rd28+16];
	ld.v4.f32 	{%f386, %f387, %f388, %f389}, [%rd28];
	mov.b32 	 %r6, %f369;
	sub.f32 	%f57, %f387, %f7;
	sub.f32 	%f59, %f388, %f8;
	sub.f32 	%f61, %f389, %f9;
	mul.f32 	%f391, %f59, %f59;
	fma.rn.f32 	%f392, %f57, %f57, %f391;
	fma.rn.f32 	%f393, %f61, %f61, %f392;
	sqrt.rn.f32 	%f62, %f393;
	rcp.rn.f32 	%f394, %f62;
	mul.f32 	%f63, %f57, %f394;
	mul.f32 	%f64, %f59, %f394;
	mul.f32 	%f65, %f61, %f394;
	mul.f32 	%f66, %f62, %f385;
	abs.f32 	%f67, %f66;
	setp.lt.f32	%p15, %f67, 0f00800000;
	mul.f32 	%f395, %f67, 0f4B800000;
	selp.f32	%f396, 0fC3170000, 0fC2FE0000, %p15;
	selp.f32	%f397, %f395, %f67, %p15;
	mov.b32 	 %r54, %f397;
	and.b32  	%r55, %r54, 8388607;
	or.b32  	%r56, %r55, 1065353216;
	mov.b32 	 %f398, %r56;
	shr.u32 	%r57, %r54, 23;
	cvt.rn.f32.u32	%f399, %r57;
	add.f32 	%f400, %f396, %f399;
	setp.gt.f32	%p16, %f398, 0f3FB504F3;
	mul.f32 	%f401, %f398, 0f3F000000;
	add.f32 	%f402, %f400, 0f3F800000;
	selp.f32	%f403, %f401, %f398, %p16;
	selp.f32	%f404, %f402, %f400, %p16;
	add.f32 	%f405, %f403, 0fBF800000;
	add.f32 	%f365, %f403, 0f3F800000;
	// inline asm
	rcp.approx.ftz.f32 %f364,%f365;
	// inline asm
	add.f32 	%f406, %f405, %f405;
	mul.f32 	%f407, %f364, %f406;
	mul.f32 	%f408, %f407, %f407;
	mov.f32 	%f409, 0f3C4CAF63;
	mov.f32 	%f410, 0f3B18F0FE;
	fma.rn.f32 	%f411, %f410, %f408, %f409;
	mov.f32 	%f412, 0f3DAAAABD;
	fma.rn.f32 	%f413, %f411, %f408, %f412;
	mul.rn.f32 	%f414, %f413, %f408;
	mul.rn.f32 	%f415, %f414, %f407;
	sub.f32 	%f416, %f405, %f407;
	neg.f32 	%f417, %f407;
	add.f32 	%f418, %f416, %f416;
	fma.rn.f32 	%f419, %f417, %f405, %f418;
	mul.rn.f32 	%f420, %f364, %f419;
	add.f32 	%f421, %f415, %f407;
	sub.f32 	%f422, %f407, %f421;
	add.f32 	%f423, %f415, %f422;
	add.f32 	%f424, %f420, %f423;
	add.f32 	%f425, %f421, %f424;
	sub.f32 	%f426, %f421, %f425;
	add.f32 	%f427, %f424, %f426;
	mov.f32 	%f428, 0f3F317200;
	mul.rn.f32 	%f429, %f404, %f428;
	mov.f32 	%f430, 0f35BFBE8E;
	mul.rn.f32 	%f431, %f404, %f430;
	add.f32 	%f432, %f429, %f425;
	sub.f32 	%f433, %f429, %f432;
	add.f32 	%f434, %f425, %f433;
	add.f32 	%f435, %f427, %f434;
	add.f32 	%f436, %f431, %f435;
	add.f32 	%f437, %f432, %f436;
	sub.f32 	%f438, %f432, %f437;
	add.f32 	%f439, %f436, %f438;
	mul.rn.f32 	%f68, %f362, %f437;
	neg.f32 	%f441, %f68;
	fma.rn.f32 	%f442, %f362, %f437, %f441;
	fma.rn.f32 	%f443, %f362, %f439, %f442;
	fma.rn.f32 	%f69, %f358, %f437, %f443;
	add.rn.f32 	%f70, %f68, %f69;
	mov.b32 	 %r58, %f70;
	setp.eq.s32	%p1, %r58, 1118925336;
	add.s32 	%r59, %r58, -1;
	mov.b32 	 %f445, %r59;
	selp.f32	%f446, %f445, %f70, %p1;
	mul.f32 	%f447, %f446, 0f3FB8AA3B;
	cvt.rzi.f32.f32	%f448, %f447;
	mov.f32 	%f449, 0fBF317200;
	fma.rn.f32 	%f450, %f448, %f449, %f446;
	mov.f32 	%f451, 0fB5BFBE8E;
	fma.rn.f32 	%f452, %f448, %f451, %f450;
	mul.f32 	%f453, %f452, 0f3FB8AA3B;
	ex2.approx.ftz.f32 	%f454, %f453;
	add.f32 	%f455, %f448, 0f00000000;
	ex2.approx.f32 	%f456, %f455;
	mul.f32 	%f457, %f454, %f456;
	setp.lt.f32	%p17, %f446, 0fC2D20000;
	selp.f32	%f458, 0f00000000, %f457, %p17;
	setp.gt.f32	%p18, %f446, 0f42D20000;
	selp.f32	%f1394, 0f7F800000, %f458, %p18;
	setp.eq.f32	%p19, %f1394, 0f7F800000;
	@%p19 bra 	BB0_7;

	neg.f32 	%f459, %f70;
	add.rn.f32 	%f460, %f68, %f459;
	add.rn.f32 	%f461, %f460, %f69;
	add.f32 	%f462, %f461, 0f37000000;
	selp.f32	%f463, %f462, %f461, %p1;
	fma.rn.f32 	%f1394, %f1394, %f463, %f1394;

BB0_7:
	setp.lt.f32	%p20, %f66, 0f00000000;
	setp.eq.f32	%p21, %f13, 0f3F800000;
	and.pred  	%p2, %p20, %p21;
	mov.b32 	 %r60, %f1394;
	xor.b32  	%r61, %r60, -2147483648;
	mov.b32 	 %f464, %r61;
	selp.f32	%f1396, %f464, %f1394, %p2;
	setp.eq.f32	%p22, %f66, 0f00000000;
	@%p22 bra 	BB0_10;
	bra.uni 	BB0_8;

BB0_10:
	add.f32 	%f467, %f66, %f66;
	selp.f32	%f1396, %f467, 0f00000000, %p21;
	bra.uni 	BB0_11;

BB0_8:
	setp.geu.f32	%p23, %f66, 0f00000000;
	@%p23 bra 	BB0_11;

	cvt.rzi.f32.f32	%f466, %f362;
	setp.neu.f32	%p24, %f466, 0f40800000;
	selp.f32	%f1396, 0f7FFFFFFF, %f1396, %p24;

BB0_11:
	add.f32 	%f468, %f67, 0f40800000;
	mov.b32 	 %r62, %f468;
	setp.lt.s32	%p26, %r62, 2139095040;
	@%p26 bra 	BB0_16;

	setp.gtu.f32	%p27, %f67, 0f7F800000;
	@%p27 bra 	BB0_15;
	bra.uni 	BB0_13;

BB0_15:
	add.f32 	%f1396, %f66, 0f40800000;
	bra.uni 	BB0_16;

BB0_13:
	setp.neu.f32	%p28, %f67, 0f7F800000;
	@%p28 bra 	BB0_16;

	selp.f32	%f1396, 0fFF800000, 0f7F800000, %p2;

BB0_16:
	mul.f32 	%f469, %f62, %f383;
	mov.f32 	%f1435, 0f3F800000;
	sub.f32 	%f471, %f1435, %f1396;
	setp.eq.f32	%p29, %f66, 0f3F800000;
	selp.f32	%f472, 0f00000000, %f471, %p29;
	cvt.sat.f32.f32	%f473, %f472;
	fma.rn.f32 	%f474, %f469, %f469, %f384;
	div.rn.f32 	%f96, %f473, %f474;
	mul.f32 	%f475, %f1376, %f64;
	fma.rn.f32 	%f476, %f1375, %f63, %f475;
	fma.rn.f32 	%f97, %f1377, %f65, %f476;
	setp.eq.f32	%p30, %f386, 0f3F800000;
	@%p30 bra 	BB0_23;
	bra.uni 	BB0_17;

BB0_23:
	setp.leu.f32	%p34, %f381, 0f00000000;
	@%p34 bra 	BB0_19;

	setp.geu.f32	%p35, %f377, 0f00000000;
	@%p35 bra 	BB0_25;

	sub.f32 	%f1374, %f389, %f9;
	sub.f32 	%f1373, %f387, %f7;
	sub.f32 	%f1372, %f388, %f8;
	mul.f32 	%f499, %f367, %f1372;
	fma.rn.f32 	%f500, %f366, %f1373, %f499;
	fma.rn.f32 	%f501, %f368, %f1374, %f500;
	rcp.rn.f32 	%f502, %f501;
	mul.f32 	%f1397, %f1373, %f502;
	mul.f32 	%f1398, %f1372, %f502;
	mul.f32 	%f1399, %f1374, %f502;
	neg.f32 	%f377, %f377;
	bra.uni 	BB0_27;

BB0_17:
	setp.eq.f32	%p31, %f386, 0f40000000;
	@%p31 bra 	BB0_21;
	bra.uni 	BB0_18;

BB0_21:
	setp.leu.f32	%p33, %f381, 0f00000000;
	@%p33 bra 	BB0_19;

	mul.f32 	%f493, %f375, %f64;
	fma.rn.f32 	%f494, %f374, %f63, %f493;
	mul.f32 	%f495, %f371, %f64;
	fma.rn.f32 	%f496, %f370, %f63, %f495;
	mul.f32 	%f497, %f367, %f64;
	fma.rn.f32 	%f498, %f366, %f63, %f497;
	fma.rn.f32 	%f490, %f376, %f65, %f494;
	fma.rn.f32 	%f491, %f372, %f65, %f496;
	fma.rn.f32 	%f492, %f368, %f65, %f498;
	cvt.rzi.s32.f32	%r63, %f381;
	mov.u32 	%r64, 6;
	mov.u32 	%r65, 0;
	// inline asm
	call (%f486, %f487, %f488, %f489), _rt_texture_get_base_id, (%r63, %r64, %f490, %f491, %f492, %r65);
	// inline asm
	mul.f32 	%f1401, %f378, %f486;
	mul.f32 	%f1402, %f1402, %f487;
	mul.f32 	%f1403, %f1403, %f488;
	bra.uni 	BB0_28;

BB0_18:
	setp.neu.f32	%p32, %f386, 0f40800000;
	@%p32 bra 	BB0_19;

	mul.f32 	%f477, %f366, %f63;
	mul.f32 	%f478, %f367, %f64;
	neg.f32 	%f479, %f478;
	sub.f32 	%f480, %f479, %f477;
	mul.f32 	%f481, %f368, %f65;
	sub.f32 	%f482, %f480, %f481;
	fma.rn.f32 	%f483, %f381, %f482, %f377;
	cvt.sat.f32.f32	%f484, %f483;
	mul.f32 	%f485, %f484, %f484;
	mul.f32 	%f1404, %f96, %f485;
	mov.f32 	%f1401, %f378;
	bra.uni 	BB0_29;

BB0_19:
	mov.f32 	%f1401, %f378;
	bra.uni 	BB0_28;

BB0_25:
	mov.f32 	%f1397, %f63;
	mov.f32 	%f1398, %f64;
	mov.f32 	%f1399, %f65;

BB0_27:
	mul.f32 	%f511, %f368, %f65;
	mul.f32 	%f512, %f367, %f64;
	neg.f32 	%f513, %f512;
	mul.f32 	%f514, %f366, %f63;
	sub.f32 	%f515, %f513, %f514;
	sub.f32 	%f516, %f515, %f511;
	setp.gt.f32	%p36, %f516, 0f00000000;
	selp.f32	%f517, 0f3F800000, 0f00000000, %p36;
	mul.f32 	%f518, %f375, %f1398;
	fma.rn.f32 	%f519, %f374, %f1397, %f518;
	mul.f32 	%f520, %f371, %f1398;
	fma.rn.f32 	%f521, %f370, %f1397, %f520;
	fma.rn.f32 	%f522, %f376, %f1399, %f519;
	fma.rn.f32 	%f523, %f372, %f1399, %f521;
	fma.rn.f32 	%f524, %f377, %f522, 0f3F000000;
	mov.f32 	%f525, 0f3F800000;
	sub.f32 	%f507, %f525, %f524;
	fma.rn.f32 	%f508, %f377, %f523, 0f3F000000;
	cvt.rzi.s32.f32	%r66, %f381;
	mov.f32 	%f510, 0f00000000;
	// inline asm
	call (%f503, %f504, %f505, %f506), _rt_texture_get_f_id, (%r66, %r28, %f507, %f508, %f510, %f510);
	// inline asm
	mul.f32 	%f526, %f517, %f503;
	mul.f32 	%f527, %f517, %f504;
	mul.f32 	%f528, %f517, %f505;
	mul.f32 	%f1401, %f378, %f526;
	mul.f32 	%f1402, %f1402, %f527;
	mul.f32 	%f1403, %f1403, %f528;

BB0_28:
	mov.f32 	%f1404, %f96;

BB0_29:
	max.f32 	%f544, %f1401, %f1402;
	max.f32 	%f545, %f544, %f1403;
	mul.f32 	%f546, %f1404, %f545;
	setp.lt.f32	%p38, %f546, 0f3727C5AC;
	mov.pred 	%p137, -1;
	mov.f32 	%f132, 0f00000000;
	mov.f32 	%f1406, %f132;
	mov.f32 	%f1407, %f132;
	mov.f32 	%f1408, %f132;
	mov.f32 	%f1409, %f132;
	mov.f32 	%f1410, %f132;
	mov.f32 	%f1411, %f132;
	mov.f32 	%f1412, %f132;
	mov.f32 	%f1413, %f132;
	mov.f32 	%f1414, %f132;
	mov.f32 	%f1415, %f132;
	mov.f32 	%f1416, %f132;
	mov.f32 	%f1417, %f132;
	mov.f32 	%f1418, %f132;
	mov.f32 	%f1419, %f132;
	@%p38 bra 	BB0_31;

	ld.global.u32 	%r68, [ignoreNormal];
	setp.eq.s32	%p40, %r68, 0;
	selp.f32	%f547, %f97, 0f3F800000, %p40;
	cvt.sat.f32.f32	%f548, %f547;
	mul.f32 	%f549, %f1404, %f548;
	mul.f32 	%f132, %f1401, %f549;
	mul.f32 	%f1406, %f1402, %f549;
	mul.f32 	%f1407, %f1403, %f549;
	mul.f32 	%f550, %f1404, 0f3E800000;
	mul.f32 	%f1408, %f1401, %f550;
	mul.f32 	%f1409, %f1402, %f550;
	mul.f32 	%f1410, %f1403, %f550;
	mul.f32 	%f1411, %f63, %f1408;
	mul.f32 	%f1412, %f63, %f1409;
	mul.f32 	%f1413, %f63, %f1410;
	mul.f32 	%f1414, %f64, %f1408;
	mul.f32 	%f1415, %f64, %f1409;
	mul.f32 	%f1416, %f64, %f1410;
	mul.f32 	%f1417, %f65, %f1408;
	mul.f32 	%f1418, %f65, %f1409;
	mul.f32 	%f1419, %f65, %f1410;
	mov.pred 	%p137, 0;

BB0_31:
	@%p137 bra 	BB0_46;

	setp.eq.s32	%p41, %r6, 0;
	mov.u16 	%rs161, 0;
	mov.f32 	%f1436, %f1435;
	mov.f32 	%f1437, %f1435;
	@%p41 bra 	BB0_43;

	abs.s32 	%r8, %r6;
	mov.f32 	%f1432, 0f00000000;
	setp.lt.s32	%p42, %r8, 1;
	mov.f32 	%f1433, %f1432;
	mov.f32 	%f1434, %f1432;
	@%p42 bra 	BB0_42;

	max.f32 	%f148, %f564, %f339;
	and.b32  	%r9, %r8, 3;
	setp.eq.s32	%p43, %r9, 0;
	add.u64 	%rd35, %SP, 0;
	cvta.to.local.u64 	%rd2, %rd35;
	mov.f32 	%f1432, 0f00000000;
	mov.u32 	%r246, 0;
	mov.f32 	%f1433, %f1432;
	mov.f32 	%f1434, %f1432;
	@%p43 bra 	BB0_40;

	setp.eq.s32	%p44, %r9, 1;
	mov.f32 	%f1423, 0f00000000;
	mov.u32 	%r244, 0;
	mov.f32 	%f1424, %f1423;
	mov.f32 	%f1425, %f1423;
	@%p44 bra 	BB0_39;

	setp.eq.s32	%p45, %r9, 2;
	mov.f32 	%f1420, 0f00000000;
	mov.u32 	%r243, 0;
	mov.f32 	%f1421, %f1420;
	mov.f32 	%f1422, %f1420;
	@%p45 bra 	BB0_38;

	sub.f32 	%f580, %f387, %f382;
	sub.f32 	%f581, %f388, %f382;
	sub.f32 	%f582, %f389, %f382;
	sub.f32 	%f583, %f580, %f7;
	sub.f32 	%f584, %f581, %f8;
	sub.f32 	%f585, %f582, %f9;
	mul.f32 	%f586, %f584, %f584;
	fma.rn.f32 	%f587, %f583, %f583, %f586;
	fma.rn.f32 	%f588, %f585, %f585, %f587;
	sqrt.rn.f32 	%f579, %f588;
	rcp.rn.f32 	%f589, %f579;
	mul.f32 	%f575, %f589, %f583;
	mul.f32 	%f576, %f589, %f584;
	mul.f32 	%f577, %f589, %f585;
	ld.global.u32 	%r76, [imageEnabled];
	and.b32  	%r77, %r76, 32;
	setp.eq.s32	%p46, %r77, 0;
	selp.f32	%f590, 0f3F800000, 0f41200000, %p46;
	mul.f32 	%f578, %f590, %f148;
	mov.u32 	%r78, 1065353216;
	st.local.u32 	[%rd2+8], %r78;
	st.local.u32 	[%rd2+4], %r78;
	st.local.u32 	[%rd2], %r78;
	ld.global.u32 	%r72, [root];
	// inline asm
	call _rt_trace_64, (%r72, %f10, %f11, %f12, %f575, %f576, %f577, %r45, %f578, %f579, %rd35, %r44);
	// inline asm
	ld.local.f32 	%f591, [%rd2];
	add.f32 	%f1422, %f591, 0f00000000;
	ld.local.f32 	%f592, [%rd2+4];
	add.f32 	%f1421, %f592, 0f00000000;
	ld.local.f32 	%f593, [%rd2+8];
	add.f32 	%f1420, %f593, 0f00000000;
	mov.u32 	%r243, %r45;

BB0_38:
	cvt.rn.f32.s32	%f602, %r243;
	mul.f32 	%f603, %f602, 0f3DD32618;
	cvt.rmi.f32.f32	%f604, %f603;
	sub.f32 	%f605, %f603, %f604;
	mul.f32 	%f606, %f602, 0f3DD2F1AA;
	cvt.rmi.f32.f32	%f607, %f606;
	sub.f32 	%f608, %f606, %f607;
	mul.f32 	%f609, %f602, 0f3DC74539;
	cvt.rmi.f32.f32	%f610, %f609;
	sub.f32 	%f611, %f609, %f610;
	add.f32 	%f612, %f608, 0f4199851F;
	add.f32 	%f613, %f611, 0f4199851F;
	add.f32 	%f614, %f605, 0f4199851F;
	mul.f32 	%f615, %f608, %f613;
	fma.rn.f32 	%f616, %f605, %f612, %f615;
	fma.rn.f32 	%f617, %f614, %f611, %f616;
	add.f32 	%f618, %f605, %f617;
	add.f32 	%f619, %f608, %f617;
	add.f32 	%f620, %f611, %f617;
	add.f32 	%f621, %f618, %f619;
	mul.f32 	%f622, %f620, %f621;
	cvt.rmi.f32.f32	%f623, %f622;
	sub.f32 	%f624, %f622, %f623;
	add.f32 	%f625, %f618, %f620;
	mul.f32 	%f626, %f619, %f625;
	cvt.rmi.f32.f32	%f627, %f626;
	sub.f32 	%f628, %f626, %f627;
	add.f32 	%f629, %f619, %f620;
	mul.f32 	%f630, %f618, %f629;
	cvt.rmi.f32.f32	%f631, %f630;
	sub.f32 	%f632, %f630, %f631;
	fma.rn.f32 	%f633, %f624, 0f40000000, 0fBF800000;
	fma.rn.f32 	%f634, %f628, 0f40000000, 0fBF800000;
	fma.rn.f32 	%f635, %f632, 0f40000000, 0fBF800000;
	fma.rn.f32 	%f636, %f382, %f633, %f387;
	fma.rn.f32 	%f637, %f382, %f634, %f388;
	fma.rn.f32 	%f638, %f382, %f635, %f389;
	sub.f32 	%f639, %f636, %f7;
	sub.f32 	%f640, %f637, %f8;
	sub.f32 	%f641, %f638, %f9;
	mul.f32 	%f642, %f640, %f640;
	fma.rn.f32 	%f643, %f639, %f639, %f642;
	fma.rn.f32 	%f644, %f641, %f641, %f643;
	sqrt.rn.f32 	%f601, %f644;
	rcp.rn.f32 	%f645, %f601;
	mul.f32 	%f597, %f645, %f639;
	mul.f32 	%f598, %f645, %f640;
	mul.f32 	%f599, %f645, %f641;
	ld.global.u32 	%r82, [imageEnabled];
	and.b32  	%r83, %r82, 32;
	setp.eq.s32	%p47, %r83, 0;
	selp.f32	%f646, 0f3F800000, 0f41200000, %p47;
	mul.f32 	%f600, %f646, %f148;
	mov.u32 	%r84, 1065353216;
	st.local.u32 	[%rd2+8], %r84;
	st.local.u32 	[%rd2+4], %r84;
	st.local.u32 	[%rd2], %r84;
	ld.global.u32 	%r79, [root];
	// inline asm
	call _rt_trace_64, (%r79, %f10, %f11, %f12, %f597, %f598, %f599, %r45, %f600, %f601, %rd35, %r44);
	// inline asm
	ld.local.f32 	%f647, [%rd2];
	add.f32 	%f1425, %f1422, %f647;
	ld.local.f32 	%f648, [%rd2+4];
	add.f32 	%f1424, %f1421, %f648;
	ld.local.f32 	%f649, [%rd2+8];
	add.f32 	%f1423, %f1420, %f649;
	add.s32 	%r244, %r243, 1;

BB0_39:
	cvt.rn.f32.s32	%f658, %r244;
	mul.f32 	%f659, %f658, 0f3DD32618;
	cvt.rmi.f32.f32	%f660, %f659;
	sub.f32 	%f661, %f659, %f660;
	mul.f32 	%f662, %f658, 0f3DD2F1AA;
	cvt.rmi.f32.f32	%f663, %f662;
	sub.f32 	%f664, %f662, %f663;
	mul.f32 	%f665, %f658, 0f3DC74539;
	cvt.rmi.f32.f32	%f666, %f665;
	sub.f32 	%f667, %f665, %f666;
	add.f32 	%f668, %f664, 0f4199851F;
	add.f32 	%f669, %f667, 0f4199851F;
	add.f32 	%f670, %f661, 0f4199851F;
	mul.f32 	%f671, %f664, %f669;
	fma.rn.f32 	%f672, %f661, %f668, %f671;
	fma.rn.f32 	%f673, %f670, %f667, %f672;
	add.f32 	%f674, %f661, %f673;
	add.f32 	%f675, %f664, %f673;
	add.f32 	%f676, %f667, %f673;
	add.f32 	%f677, %f674, %f675;
	mul.f32 	%f678, %f676, %f677;
	cvt.rmi.f32.f32	%f679, %f678;
	sub.f32 	%f680, %f678, %f679;
	add.f32 	%f681, %f674, %f676;
	mul.f32 	%f682, %f675, %f681;
	cvt.rmi.f32.f32	%f683, %f682;
	sub.f32 	%f684, %f682, %f683;
	add.f32 	%f685, %f675, %f676;
	mul.f32 	%f686, %f674, %f685;
	cvt.rmi.f32.f32	%f687, %f686;
	sub.f32 	%f688, %f686, %f687;
	fma.rn.f32 	%f689, %f680, 0f40000000, 0fBF800000;
	fma.rn.f32 	%f690, %f684, 0f40000000, 0fBF800000;
	fma.rn.f32 	%f691, %f688, 0f40000000, 0fBF800000;
	fma.rn.f32 	%f692, %f382, %f689, %f387;
	fma.rn.f32 	%f693, %f382, %f690, %f388;
	fma.rn.f32 	%f694, %f382, %f691, %f389;
	sub.f32 	%f695, %f692, %f7;
	sub.f32 	%f696, %f693, %f8;
	sub.f32 	%f697, %f694, %f9;
	mul.f32 	%f698, %f696, %f696;
	fma.rn.f32 	%f699, %f695, %f695, %f698;
	fma.rn.f32 	%f700, %f697, %f697, %f699;
	sqrt.rn.f32 	%f657, %f700;
	rcp.rn.f32 	%f701, %f657;
	mul.f32 	%f653, %f701, %f695;
	mul.f32 	%f654, %f701, %f696;
	mul.f32 	%f655, %f701, %f697;
	ld.global.u32 	%r88, [imageEnabled];
	and.b32  	%r89, %r88, 32;
	setp.eq.s32	%p48, %r89, 0;
	selp.f32	%f702, 0f3F800000, 0f41200000, %p48;
	mul.f32 	%f656, %f702, %f148;
	mov.u32 	%r90, 1065353216;
	st.local.u32 	[%rd2+8], %r90;
	st.local.u32 	[%rd2+4], %r90;
	st.local.u32 	[%rd2], %r90;
	ld.global.u32 	%r85, [root];
	mov.u32 	%r86, 1;
	// inline asm
	call _rt_trace_64, (%r85, %f10, %f11, %f12, %f653, %f654, %f655, %r86, %f656, %f657, %rd35, %r44);
	// inline asm
	ld.local.f32 	%f703, [%rd2];
	add.f32 	%f1434, %f1425, %f703;
	ld.local.f32 	%f704, [%rd2+4];
	add.f32 	%f1433, %f1424, %f704;
	ld.local.f32 	%f705, [%rd2+8];
	add.f32 	%f1432, %f1423, %f705;
	add.s32 	%r246, %r244, 1;

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

BB0_41:
	cvt.rn.f32.s32	%f738, %r246;
	mul.f32 	%f739, %f738, 0f3DD32618;
	cvt.rmi.f32.f32	%f740, %f739;
	sub.f32 	%f741, %f739, %f740;
	mul.f32 	%f742, %f738, 0f3DD2F1AA;
	cvt.rmi.f32.f32	%f743, %f742;
	sub.f32 	%f744, %f742, %f743;
	mul.f32 	%f745, %f738, 0f3DC74539;
	cvt.rmi.f32.f32	%f746, %f745;
	sub.f32 	%f747, %f745, %f746;
	add.f32 	%f748, %f744, 0f4199851F;
	add.f32 	%f749, %f747, 0f4199851F;
	add.f32 	%f750, %f741, 0f4199851F;
	mul.f32 	%f751, %f744, %f749;
	fma.rn.f32 	%f752, %f741, %f748, %f751;
	fma.rn.f32 	%f753, %f750, %f747, %f752;
	add.f32 	%f754, %f741, %f753;
	add.f32 	%f755, %f744, %f753;
	add.f32 	%f756, %f747, %f753;
	add.f32 	%f757, %f754, %f755;
	mul.f32 	%f758, %f756, %f757;
	cvt.rmi.f32.f32	%f759, %f758;
	sub.f32 	%f760, %f758, %f759;
	add.f32 	%f761, %f754, %f756;
	mul.f32 	%f762, %f755, %f761;
	cvt.rmi.f32.f32	%f763, %f762;
	sub.f32 	%f764, %f762, %f763;
	add.f32 	%f765, %f755, %f756;
	mul.f32 	%f766, %f754, %f765;
	cvt.rmi.f32.f32	%f767, %f766;
	sub.f32 	%f768, %f766, %f767;
	fma.rn.f32 	%f769, %f760, 0f40000000, 0fBF800000;
	fma.rn.f32 	%f770, %f764, 0f40000000, 0fBF800000;
	fma.rn.f32 	%f771, %f768, 0f40000000, 0fBF800000;
	fma.rn.f32 	%f772, %f382, %f769, %f387;
	fma.rn.f32 	%f773, %f382, %f770, %f388;
	fma.rn.f32 	%f774, %f382, %f771, %f389;
	sub.f32 	%f775, %f772, %f7;
	sub.f32 	%f776, %f773, %f8;
	sub.f32 	%f777, %f774, %f9;
	mul.f32 	%f778, %f776, %f776;
	fma.rn.f32 	%f779, %f775, %f775, %f778;
	fma.rn.f32 	%f780, %f777, %f777, %f779;
	sqrt.rn.f32 	%f713, %f780;
	rcp.rn.f32 	%f781, %f713;
	mul.f32 	%f709, %f781, %f775;
	mul.f32 	%f710, %f781, %f776;
	mul.f32 	%f711, %f781, %f777;
	ld.global.u32 	%r103, [imageEnabled];
	and.b32  	%r104, %r103, 32;
	setp.eq.s32	%p50, %r104, 0;
	selp.f32	%f782, 0f3F800000, 0f41200000, %p50;
	mul.f32 	%f712, %f782, %f148;
	mov.u32 	%r105, 1065353216;
	st.local.u32 	[%rd2+8], %r105;
	st.local.u32 	[%rd2+4], %r105;
	st.local.u32 	[%rd2], %r105;
	ld.global.u32 	%r91, [root];
	mov.u32 	%r101, 1;
	// inline asm
	call _rt_trace_64, (%r91, %f10, %f11, %f12, %f709, %f710, %f711, %r101, %f712, %f713, %rd35, %r44);
	// inline asm
	ld.local.f32 	%f783, [%rd2];
	add.f32 	%f784, %f1434, %f783;
	ld.local.f32 	%f785, [%rd2+4];
	add.f32 	%f786, %f1433, %f785;
	ld.local.f32 	%f787, [%rd2+8];
	add.f32 	%f788, %f1432, %f787;
	add.s32 	%r106, %r246, 1;
	cvt.rn.f32.s32	%f789, %r106;
	mul.f32 	%f790, %f789, 0f3DD32618;
	cvt.rmi.f32.f32	%f791, %f790;
	sub.f32 	%f792, %f790, %f791;
	mul.f32 	%f793, %f789, 0f3DD2F1AA;
	cvt.rmi.f32.f32	%f794, %f793;
	sub.f32 	%f795, %f793, %f794;
	mul.f32 	%f796, %f789, 0f3DC74539;
	cvt.rmi.f32.f32	%f797, %f796;
	sub.f32 	%f798, %f796, %f797;
	add.f32 	%f799, %f795, 0f4199851F;
	add.f32 	%f800, %f798, 0f4199851F;
	add.f32 	%f801, %f792, 0f4199851F;
	mul.f32 	%f802, %f795, %f800;
	fma.rn.f32 	%f803, %f792, %f799, %f802;
	fma.rn.f32 	%f804, %f801, %f798, %f803;
	add.f32 	%f805, %f792, %f804;
	add.f32 	%f806, %f795, %f804;
	add.f32 	%f807, %f798, %f804;
	add.f32 	%f808, %f805, %f806;
	mul.f32 	%f809, %f807, %f808;
	cvt.rmi.f32.f32	%f810, %f809;
	sub.f32 	%f811, %f809, %f810;
	add.f32 	%f812, %f805, %f807;
	mul.f32 	%f813, %f806, %f812;
	cvt.rmi.f32.f32	%f814, %f813;
	sub.f32 	%f815, %f813, %f814;
	add.f32 	%f816, %f806, %f807;
	mul.f32 	%f817, %f805, %f816;
	cvt.rmi.f32.f32	%f818, %f817;
	sub.f32 	%f819, %f817, %f818;
	fma.rn.f32 	%f820, %f811, 0f40000000, 0fBF800000;
	fma.rn.f32 	%f821, %f815, 0f40000000, 0fBF800000;
	fma.rn.f32 	%f822, %f819, 0f40000000, 0fBF800000;
	fma.rn.f32 	%f823, %f382, %f820, %f387;
	fma.rn.f32 	%f824, %f382, %f821, %f388;
	fma.rn.f32 	%f825, %f382, %f822, %f389;
	sub.f32 	%f826, %f823, %f7;
	sub.f32 	%f827, %f824, %f8;
	sub.f32 	%f828, %f825, %f9;
	mul.f32 	%f829, %f827, %f827;
	fma.rn.f32 	%f830, %f826, %f826, %f829;
	fma.rn.f32 	%f831, %f828, %f828, %f830;
	sqrt.rn.f32 	%f721, %f831;
	rcp.rn.f32 	%f832, %f721;
	mul.f32 	%f717, %f832, %f826;
	mul.f32 	%f718, %f832, %f827;
	mul.f32 	%f719, %f832, %f828;
	ld.global.u32 	%r107, [imageEnabled];
	and.b32  	%r108, %r107, 32;
	setp.eq.s32	%p51, %r108, 0;
	selp.f32	%f833, 0f3F800000, 0f41200000, %p51;
	mul.f32 	%f720, %f833, %f148;
	st.local.u32 	[%rd2+8], %r105;
	st.local.u32 	[%rd2+4], %r105;
	st.local.u32 	[%rd2], %r105;
	ld.global.u32 	%r94, [root];
	// inline asm
	call _rt_trace_64, (%r94, %f10, %f11, %f12, %f717, %f718, %f719, %r101, %f720, %f721, %rd35, %r44);
	// inline asm
	ld.local.f32 	%f834, [%rd2];
	add.f32 	%f835, %f784, %f834;
	ld.local.f32 	%f836, [%rd2+4];
	add.f32 	%f837, %f786, %f836;
	ld.local.f32 	%f838, [%rd2+8];
	add.f32 	%f839, %f788, %f838;
	add.s32 	%r109, %r246, 2;
	cvt.rn.f32.s32	%f840, %r109;
	mul.f32 	%f841, %f840, 0f3DD32618;
	cvt.rmi.f32.f32	%f842, %f841;
	sub.f32 	%f843, %f841, %f842;
	mul.f32 	%f844, %f840, 0f3DD2F1AA;
	cvt.rmi.f32.f32	%f845, %f844;
	sub.f32 	%f846, %f844, %f845;
	mul.f32 	%f847, %f840, 0f3DC74539;
	cvt.rmi.f32.f32	%f848, %f847;
	sub.f32 	%f849, %f847, %f848;
	add.f32 	%f850, %f846, 0f4199851F;
	add.f32 	%f851, %f849, 0f4199851F;
	add.f32 	%f852, %f843, 0f4199851F;
	mul.f32 	%f853, %f846, %f851;
	fma.rn.f32 	%f854, %f843, %f850, %f853;
	fma.rn.f32 	%f855, %f852, %f849, %f854;
	add.f32 	%f856, %f843, %f855;
	add.f32 	%f857, %f846, %f855;
	add.f32 	%f858, %f849, %f855;
	add.f32 	%f859, %f856, %f857;
	mul.f32 	%f860, %f858, %f859;
	cvt.rmi.f32.f32	%f861, %f860;
	sub.f32 	%f862, %f860, %f861;
	add.f32 	%f863, %f856, %f858;
	mul.f32 	%f864, %f857, %f863;
	cvt.rmi.f32.f32	%f865, %f864;
	sub.f32 	%f866, %f864, %f865;
	add.f32 	%f867, %f857, %f858;
	mul.f32 	%f868, %f856, %f867;
	cvt.rmi.f32.f32	%f869, %f868;
	sub.f32 	%f870, %f868, %f869;
	fma.rn.f32 	%f871, %f862, 0f40000000, 0fBF800000;
	fma.rn.f32 	%f872, %f866, 0f40000000, 0fBF800000;
	fma.rn.f32 	%f873, %f870, 0f40000000, 0fBF800000;
	fma.rn.f32 	%f874, %f382, %f871, %f387;
	fma.rn.f32 	%f875, %f382, %f872, %f388;
	fma.rn.f32 	%f876, %f382, %f873, %f389;
	sub.f32 	%f877, %f874, %f7;
	sub.f32 	%f878, %f875, %f8;
	sub.f32 	%f879, %f876, %f9;
	mul.f32 	%f880, %f878, %f878;
	fma.rn.f32 	%f881, %f877, %f877, %f880;
	fma.rn.f32 	%f882, %f879, %f879, %f881;
	sqrt.rn.f32 	%f729, %f882;
	rcp.rn.f32 	%f883, %f729;
	mul.f32 	%f725, %f883, %f877;
	mul.f32 	%f726, %f883, %f878;
	mul.f32 	%f727, %f883, %f879;
	ld.global.u32 	%r110, [imageEnabled];
	and.b32  	%r111, %r110, 32;
	setp.eq.s32	%p52, %r111, 0;
	selp.f32	%f884, 0f3F800000, 0f41200000, %p52;
	mul.f32 	%f728, %f884, %f148;
	st.local.u32 	[%rd2+8], %r105;
	st.local.u32 	[%rd2+4], %r105;
	st.local.u32 	[%rd2], %r105;
	ld.global.u32 	%r97, [root];
	// inline asm
	call _rt_trace_64, (%r97, %f10, %f11, %f12, %f725, %f726, %f727, %r101, %f728, %f729, %rd35, %r44);
	// inline asm
	ld.local.f32 	%f885, [%rd2];
	add.f32 	%f886, %f835, %f885;
	ld.local.f32 	%f887, [%rd2+4];
	add.f32 	%f888, %f837, %f887;
	ld.local.f32 	%f889, [%rd2+8];
	add.f32 	%f890, %f839, %f889;
	add.s32 	%r112, %r246, 3;
	cvt.rn.f32.s32	%f891, %r112;
	mul.f32 	%f892, %f891, 0f3DD32618;
	cvt.rmi.f32.f32	%f893, %f892;
	sub.f32 	%f894, %f892, %f893;
	mul.f32 	%f895, %f891, 0f3DD2F1AA;
	cvt.rmi.f32.f32	%f896, %f895;
	sub.f32 	%f897, %f895, %f896;
	mul.f32 	%f898, %f891, 0f3DC74539;
	cvt.rmi.f32.f32	%f899, %f898;
	sub.f32 	%f900, %f898, %f899;
	add.f32 	%f901, %f897, 0f4199851F;
	add.f32 	%f902, %f900, 0f4199851F;
	add.f32 	%f903, %f894, 0f4199851F;
	mul.f32 	%f904, %f897, %f902;
	fma.rn.f32 	%f905, %f894, %f901, %f904;
	fma.rn.f32 	%f906, %f903, %f900, %f905;
	add.f32 	%f907, %f894, %f906;
	add.f32 	%f908, %f897, %f906;
	add.f32 	%f909, %f900, %f906;
	add.f32 	%f910, %f907, %f908;
	mul.f32 	%f911, %f909, %f910;
	cvt.rmi.f32.f32	%f912, %f911;
	sub.f32 	%f913, %f911, %f912;
	add.f32 	%f914, %f907, %f909;
	mul.f32 	%f915, %f908, %f914;
	cvt.rmi.f32.f32	%f916, %f915;
	sub.f32 	%f917, %f915, %f916;
	add.f32 	%f918, %f908, %f909;
	mul.f32 	%f919, %f907, %f918;
	cvt.rmi.f32.f32	%f920, %f919;
	sub.f32 	%f921, %f919, %f920;
	fma.rn.f32 	%f922, %f913, 0f40000000, 0fBF800000;
	fma.rn.f32 	%f923, %f917, 0f40000000, 0fBF800000;
	fma.rn.f32 	%f924, %f921, 0f40000000, 0fBF800000;
	fma.rn.f32 	%f925, %f382, %f922, %f387;
	fma.rn.f32 	%f926, %f382, %f923, %f388;
	fma.rn.f32 	%f927, %f382, %f924, %f389;
	sub.f32 	%f928, %f925, %f7;
	sub.f32 	%f929, %f926, %f8;
	sub.f32 	%f930, %f927, %f9;
	mul.f32 	%f931, %f929, %f929;
	fma.rn.f32 	%f932, %f928, %f928, %f931;
	fma.rn.f32 	%f933, %f930, %f930, %f932;
	sqrt.rn.f32 	%f737, %f933;
	rcp.rn.f32 	%f934, %f737;
	mul.f32 	%f733, %f934, %f928;
	mul.f32 	%f734, %f934, %f929;
	mul.f32 	%f735, %f934, %f930;
	ld.global.u32 	%r113, [imageEnabled];
	and.b32  	%r114, %r113, 32;
	setp.eq.s32	%p53, %r114, 0;
	selp.f32	%f935, 0f3F800000, 0f41200000, %p53;
	mul.f32 	%f736, %f935, %f148;
	st.local.u32 	[%rd2+8], %r105;
	st.local.u32 	[%rd2+4], %r105;
	st.local.u32 	[%rd2], %r105;
	ld.global.u32 	%r100, [root];
	// inline asm
	call _rt_trace_64, (%r100, %f10, %f11, %f12, %f733, %f734, %f735, %r101, %f736, %f737, %rd35, %r44);
	// inline asm
	ld.local.f32 	%f936, [%rd2];
	add.f32 	%f1434, %f886, %f936;
	ld.local.f32 	%f937, [%rd2+4];
	add.f32 	%f1433, %f888, %f937;
	ld.local.f32 	%f938, [%rd2+8];
	add.f32 	%f1432, %f890, %f938;
	add.s32 	%r246, %r246, 4;
	setp.lt.s32	%p54, %r246, %r8;
	@%p54 bra 	BB0_41;

BB0_42:
	cvt.rn.f32.s32	%f939, %r8;
	rcp.rn.f32 	%f940, %f939;
	mul.f32 	%f1435, %f1434, %f940;
	mul.f32 	%f1436, %f1433, %f940;
	mul.f32 	%f1437, %f1432, %f940;
	shr.u32 	%r115, %r6, 31;
	cvt.u16.u32	%rs161, %r115;

BB0_43:
	fma.rn.f32 	%f1393, %f132, %f1435, %f1393;
	fma.rn.f32 	%f1392, %f1406, %f1436, %f1392;
	fma.rn.f32 	%f1391, %f1407, %f1437, %f1391;
	fma.rn.f32 	%f1390, %f1408, %f1435, %f1390;
	fma.rn.f32 	%f1389, %f1409, %f1436, %f1389;
	fma.rn.f32 	%f1388, %f1410, %f1437, %f1388;
	fma.rn.f32 	%f1387, %f1411, %f1435, %f1387;
	fma.rn.f32 	%f1386, %f1412, %f1436, %f1386;
	fma.rn.f32 	%f1385, %f1413, %f1437, %f1385;
	fma.rn.f32 	%f1384, %f1414, %f1435, %f1384;
	fma.rn.f32 	%f1383, %f1415, %f1436, %f1383;
	fma.rn.f32 	%f1382, %f1416, %f1437, %f1382;
	fma.rn.f32 	%f1381, %f1417, %f1435, %f1381;
	fma.rn.f32 	%f1380, %f1418, %f1436, %f1380;
	fma.rn.f32 	%f1379, %f1419, %f1437, %f1379;
	setp.eq.s16	%p55, %rs161, 0;
	@%p55 bra 	BB0_45;

	div.rn.f32 	%f941, %f132, %f378;
	div.rn.f32 	%f942, %f941, %f96;
	cvt.sat.f32.f32	%f943, %f942;
	mul.f32 	%f1435, %f1435, %f943;

BB0_45:
	add.f32 	%f1378, %f1378, %f1435;

BB0_46:
	add.s32 	%r242, %r242, 1;
	setp.lt.u32	%p56, %r242, %r4;
	@%p56 bra 	BB0_5;

BB0_47:
	ld.global.u32 	%r248, [imageEnabled];
	and.b32  	%r116, %r248, 8;
	setp.eq.s32	%p57, %r116, 0;
	@%p57 bra 	BB0_60;

	cvt.sat.f32.f32	%f232, %f1378;
	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 	%f946, 0f3E68BA2E;
	cvt.rzi.f32.f32	%f947, %f946;
	fma.rn.f32 	%f948, %f947, 0fC0000000, 0f3EE8BA2E;
	abs.f32 	%f233, %f948;
	abs.f32 	%f234, %f232;
	setp.lt.f32	%p58, %f234, 0f00800000;
	mul.f32 	%f949, %f234, 0f4B800000;
	selp.f32	%f950, 0fC3170000, 0fC2FE0000, %p58;
	selp.f32	%f951, %f949, %f234, %p58;
	mov.b32 	 %r119, %f951;
	and.b32  	%r120, %r119, 8388607;
	or.b32  	%r121, %r120, 1065353216;
	mov.b32 	 %f952, %r121;
	shr.u32 	%r122, %r119, 23;
	cvt.rn.f32.u32	%f953, %r122;
	add.f32 	%f954, %f950, %f953;
	setp.gt.f32	%p59, %f952, 0f3FB504F3;
	mul.f32 	%f955, %f952, 0f3F000000;
	add.f32 	%f956, %f954, 0f3F800000;
	selp.f32	%f957, %f955, %f952, %p59;
	selp.f32	%f958, %f956, %f954, %p59;
	add.f32 	%f959, %f957, 0fBF800000;
	add.f32 	%f945, %f957, 0f3F800000;
	// inline asm
	rcp.approx.ftz.f32 %f944,%f945;
	// inline asm
	add.f32 	%f960, %f959, %f959;
	mul.f32 	%f961, %f944, %f960;
	mul.f32 	%f962, %f961, %f961;
	mov.f32 	%f963, 0f3C4CAF63;
	mov.f32 	%f964, 0f3B18F0FE;
	fma.rn.f32 	%f965, %f964, %f962, %f963;
	mov.f32 	%f966, 0f3DAAAABD;
	fma.rn.f32 	%f967, %f965, %f962, %f966;
	mul.rn.f32 	%f968, %f967, %f962;
	mul.rn.f32 	%f969, %f968, %f961;
	sub.f32 	%f970, %f959, %f961;
	neg.f32 	%f971, %f961;
	add.f32 	%f972, %f970, %f970;
	fma.rn.f32 	%f973, %f971, %f959, %f972;
	mul.rn.f32 	%f974, %f944, %f973;
	add.f32 	%f975, %f969, %f961;
	sub.f32 	%f976, %f961, %f975;
	add.f32 	%f977, %f969, %f976;
	add.f32 	%f978, %f974, %f977;
	add.f32 	%f979, %f975, %f978;
	sub.f32 	%f980, %f975, %f979;
	add.f32 	%f981, %f978, %f980;
	mov.f32 	%f982, 0f3F317200;
	mul.rn.f32 	%f983, %f958, %f982;
	mov.f32 	%f984, 0f35BFBE8E;
	mul.rn.f32 	%f985, %f958, %f984;
	add.f32 	%f986, %f983, %f979;
	sub.f32 	%f987, %f983, %f986;
	add.f32 	%f988, %f979, %f987;
	add.f32 	%f989, %f981, %f988;
	add.f32 	%f990, %f985, %f989;
	add.f32 	%f991, %f986, %f990;
	sub.f32 	%f992, %f986, %f991;
	add.f32 	%f993, %f990, %f992;
	mov.f32 	%f994, 0f3EE8BA2E;
	mul.rn.f32 	%f995, %f994, %f991;
	neg.f32 	%f996, %f995;
	fma.rn.f32 	%f997, %f994, %f991, %f996;
	fma.rn.f32 	%f998, %f994, %f993, %f997;
	mov.f32 	%f999, 0f00000000;
	fma.rn.f32 	%f1000, %f999, %f991, %f998;
	add.rn.f32 	%f1001, %f995, %f1000;
	neg.f32 	%f1002, %f1001;
	add.rn.f32 	%f1003, %f995, %f1002;
	add.rn.f32 	%f1004, %f1003, %f1000;
	mov.b32 	 %r123, %f1001;
	setp.eq.s32	%p60, %r123, 1118925336;
	add.s32 	%r124, %r123, -1;
	mov.b32 	 %f1005, %r124;
	add.f32 	%f1006, %f1004, 0f37000000;
	selp.f32	%f1007, %f1005, %f1001, %p60;
	selp.f32	%f235, %f1006, %f1004, %p60;
	mul.f32 	%f1008, %f1007, 0f3FB8AA3B;
	cvt.rzi.f32.f32	%f1009, %f1008;
	mov.f32 	%f1010, 0fBF317200;
	fma.rn.f32 	%f1011, %f1009, %f1010, %f1007;
	mov.f32 	%f1012, 0fB5BFBE8E;
	fma.rn.f32 	%f1013, %f1009, %f1012, %f1011;
	mul.f32 	%f1014, %f1013, 0f3FB8AA3B;
	ex2.approx.ftz.f32 	%f1015, %f1014;
	add.f32 	%f1016, %f1009, 0f00000000;
	ex2.approx.f32 	%f1017, %f1016;
	mul.f32 	%f1018, %f1015, %f1017;
	setp.lt.f32	%p61, %f1007, 0fC2D20000;
	selp.f32	%f1019, 0f00000000, %f1018, %p61;
	setp.gt.f32	%p62, %f1007, 0f42D20000;
	selp.f32	%f1471, 0f7F800000, %f1019, %p62;
	setp.eq.f32	%p63, %f1471, 0f7F800000;
	@%p63 bra 	BB0_50;

	fma.rn.f32 	%f1471, %f1471, %f235, %f1471;

BB0_50:
	setp.lt.f32	%p64, %f232, 0f00000000;
	setp.eq.f32	%p65, %f233, 0f3F800000;
	and.pred  	%p4, %p64, %p65;
	mov.b32 	 %r125, %f1471;
	xor.b32  	%r126, %r125, -2147483648;
	mov.b32 	 %f1020, %r126;
	selp.f32	%f1473, %f1020, %f1471, %p4;
	setp.eq.f32	%p66, %f232, 0f00000000;
	@%p66 bra 	BB0_53;
	bra.uni 	BB0_51;

BB0_53:
	add.f32 	%f1023, %f232, %f232;
	selp.f32	%f1473, %f1023, 0f00000000, %p65;
	bra.uni 	BB0_54;

BB0_118:
	mov.u64 	%rd190, image_HDR;
	cvta.global.u64 	%rd185, %rd190;
	mov.u32 	%r212, 8;
	// inline asm
	call (%rd184), _rt_buffer_get_64, (%rd185, %r28, %r212, %rd6, %rd7, %rd13, %rd13);
	// inline asm
	mov.f32 	%f1343, 0f00000000;
	// inline asm
	{  cvt.rn.f16.f32 %rs107, %f1343;}

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

BB0_119:
	ld.global.u32 	%r213, [additive];
	setp.eq.s32	%p133, %r213, 0;
	@%p133 bra 	BB0_121;

	mov.u64 	%rd203, image_RNM0;
	cvta.global.u64 	%rd192, %rd203;
	mov.u32 	%r217, 8;
	// inline asm
	call (%rd191), _rt_buffer_get_64, (%rd192, %r28, %r217, %rd6, %rd7, %rd13, %rd13);
	// inline asm
	ld.v4.u16 	{%rs115, %rs116, %rs117, %rs118}, [%rd191];
	// inline asm
	{  cvt.f32.f16 %f1344, %rs115;}

	// inline asm
	// inline asm
	{  cvt.f32.f16 %f1345, %rs116;}

	// inline asm
	// inline asm
	{  cvt.f32.f16 %f1346, %rs117;}

	// inline asm
	// inline asm
	call (%rd197), _rt_buffer_get_64, (%rd192, %r28, %r217, %rd6, %rd7, %rd13, %rd13);
	// inline asm
	add.f32 	%f1347, %f1344, 0f00000000;
	add.f32 	%f1348, %f1345, 0f00000000;
	add.f32 	%f1349, %f1346, 0f00000000;
	// inline asm
	{  cvt.rn.f16.f32 %rs114, %f1349;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs113, %f1348;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs112, %f1347;}

	// 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 	%r219, 8;
	// inline asm
	call (%rd204), _rt_buffer_get_64, (%rd205, %r28, %r219, %rd6, %rd7, %rd13, %rd13);
	// inline asm
	mov.f32 	%f1350, 0f00000000;
	// inline asm
	{  cvt.rn.f16.f32 %rs120, %f1350;}

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

BB0_122:
	ld.global.u32 	%r220, [additive];
	setp.eq.s32	%p134, %r220, 0;
	@%p134 bra 	BB0_124;

	mov.u64 	%rd223, image_RNM1;
	cvta.global.u64 	%rd212, %rd223;
	mov.u32 	%r224, 8;
	// inline asm
	call (%rd211), _rt_buffer_get_64, (%rd212, %r28, %r224, %rd6, %rd7, %rd13, %rd13);
	// inline asm
	ld.v4.u16 	{%rs128, %rs129, %rs130, %rs131}, [%rd211];
	// inline asm
	{  cvt.f32.f16 %f1351, %rs128;}

	// inline asm
	// inline asm
	{  cvt.f32.f16 %f1352, %rs129;}

	// inline asm
	// inline asm
	{  cvt.f32.f16 %f1353, %rs130;}

	// inline asm
	// inline asm
	call (%rd217), _rt_buffer_get_64, (%rd212, %r28, %r224, %rd6, %rd7, %rd13, %rd13);
	// inline asm
	add.f32 	%f1354, %f1351, 0f00000000;
	add.f32 	%f1355, %f1352, 0f00000000;
	add.f32 	%f1356, %f1353, 0f00000000;
	// inline asm
	{  cvt.rn.f16.f32 %rs127, %f1356;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs126, %f1355;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs125, %f1354;}

	// 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 	%r226, 8;
	// inline asm
	call (%rd224), _rt_buffer_get_64, (%rd225, %r28, %r226, %rd6, %rd7, %rd13, %rd13);
	// inline asm
	mov.f32 	%f1357, 0f00000000;
	// inline asm
	{  cvt.rn.f16.f32 %rs133, %f1357;}

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

BB0_125:
	ld.global.u32 	%r227, [additive];
	setp.eq.s32	%p135, %r227, 0;
	@%p135 bra 	BB0_127;

	mov.u64 	%rd243, image_RNM2;
	cvta.global.u64 	%rd232, %rd243;
	mov.u32 	%r231, 8;
	// inline asm
	call (%rd231), _rt_buffer_get_64, (%rd232, %r28, %r231, %rd6, %rd7, %rd13, %rd13);
	// inline asm
	ld.v4.u16 	{%rs141, %rs142, %rs143, %rs144}, [%rd231];
	// inline asm
	{  cvt.f32.f16 %f1358, %rs141;}

	// inline asm
	// inline asm
	{  cvt.f32.f16 %f1359, %rs142;}

	// inline asm
	// inline asm
	{  cvt.f32.f16 %f1360, %rs143;}

	// inline asm
	// inline asm
	call (%rd237), _rt_buffer_get_64, (%rd232, %r28, %r231, %rd6, %rd7, %rd13, %rd13);
	// inline asm
	add.f32 	%f1361, %f1358, 0f00000000;
	add.f32 	%f1362, %f1359, 0f00000000;
	add.f32 	%f1363, %f1360, 0f00000000;
	// inline asm
	{  cvt.rn.f16.f32 %rs140, %f1363;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs139, %f1362;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs138, %f1361;}

	// 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 	%r233, 8;
	// inline asm
	call (%rd244), _rt_buffer_get_64, (%rd245, %r28, %r233, %rd6, %rd7, %rd13, %rd13);
	// inline asm
	mov.f32 	%f1364, 0f00000000;
	// inline asm
	{  cvt.rn.f16.f32 %rs146, %f1364;}

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

BB0_128:
	ld.global.u32 	%r234, [additive];
	setp.eq.s32	%p136, %r234, 0;
	@%p136 bra 	BB0_130;

	mov.u64 	%rd263, image_RNM3;
	cvta.global.u64 	%rd252, %rd263;
	mov.u32 	%r238, 8;
	// inline asm
	call (%rd251), _rt_buffer_get_64, (%rd252, %r28, %r238, %rd6, %rd7, %rd13, %rd13);
	// inline asm
	ld.v4.u16 	{%rs154, %rs155, %rs156, %rs157}, [%rd251];
	// inline asm
	{  cvt.f32.f16 %f1365, %rs154;}

	// inline asm
	// inline asm
	{  cvt.f32.f16 %f1366, %rs155;}

	// inline asm
	// inline asm
	{  cvt.f32.f16 %f1367, %rs156;}

	// inline asm
	// inline asm
	call (%rd257), _rt_buffer_get_64, (%rd252, %r28, %r238, %rd6, %rd7, %rd13, %rd13);
	// inline asm
	add.f32 	%f1368, %f1365, 0f00000000;
	add.f32 	%f1369, %f1366, 0f00000000;
	add.f32 	%f1370, %f1367, 0f00000000;
	// inline asm
	{  cvt.rn.f16.f32 %rs153, %f1370;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs152, %f1369;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs151, %f1368;}

	// 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 	%r240, 8;
	// inline asm
	call (%rd264), _rt_buffer_get_64, (%rd265, %r28, %r240, %rd6, %rd7, %rd13, %rd13);
	// inline asm
	mov.f32 	%f1371, 0f00000000;
	// inline asm
	{  cvt.rn.f16.f32 %rs159, %f1371;}

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

BB0_51:
	setp.geu.f32	%p67, %f232, 0f00000000;
	@%p67 bra 	BB0_54;

	cvt.rzi.f32.f32	%f1022, %f994;
	setp.neu.f32	%p68, %f1022, 0f3EE8BA2E;
	selp.f32	%f1473, 0f7FFFFFFF, %f1473, %p68;

BB0_54:
	add.f32 	%f1024, %f234, 0f3EE8BA2E;
	mov.b32 	 %r127, %f1024;
	setp.lt.s32	%p70, %r127, 2139095040;
	@%p70 bra 	BB0_59;

	setp.gtu.f32	%p71, %f234, 0f7F800000;
	@%p71 bra 	BB0_58;
	bra.uni 	BB0_56;

BB0_58:
	add.f32 	%f1473, %f232, 0f3EE8BA2E;
	bra.uni 	BB0_59;

BB0_56:
	setp.neu.f32	%p72, %f234, 0f7F800000;
	@%p72 bra 	BB0_59;

	selp.f32	%f1473, 0fFF800000, 0f7F800000, %p4;

BB0_59:
	mul.f32 	%f1025, %f1473, 0f437F0000;
	setp.eq.f32	%p73, %f232, 0f3F800000;
	selp.f32	%f1026, 0f437F0000, %f1025, %p73;
	cvt.rzi.u32.f32	%r128, %f1026;
	cvt.u16.u32	%rs17, %r128;
	mov.u16 	%rs18, 255;
	st.v2.u8 	[%rd43], {%rs17, %rs18};
	ld.global.u32 	%r248, [imageEnabled];

BB0_60:
	and.b32  	%r129, %r248, 1;
	setp.eq.b32	%p74, %r129, 1;
	@!%p74 bra 	BB0_95;
	bra.uni 	BB0_61;

BB0_61:
	mov.f32 	%f1029, 0f3E666666;
	cvt.rzi.f32.f32	%f1030, %f1029;
	fma.rn.f32 	%f1031, %f1030, 0fC0000000, 0f3EE66666;
	abs.f32 	%f246, %f1031;
	abs.f32 	%f247, %f1393;
	setp.lt.f32	%p75, %f247, 0f00800000;
	mul.f32 	%f1032, %f247, 0f4B800000;
	selp.f32	%f1033, 0fC3170000, 0fC2FE0000, %p75;
	selp.f32	%f1034, %f1032, %f247, %p75;
	mov.b32 	 %r130, %f1034;
	and.b32  	%r131, %r130, 8388607;
	or.b32  	%r132, %r131, 1065353216;
	mov.b32 	 %f1035, %r132;
	shr.u32 	%r133, %r130, 23;
	cvt.rn.f32.u32	%f1036, %r133;
	add.f32 	%f1037, %f1033, %f1036;
	setp.gt.f32	%p76, %f1035, 0f3FB504F3;
	mul.f32 	%f1038, %f1035, 0f3F000000;
	add.f32 	%f1039, %f1037, 0f3F800000;
	selp.f32	%f1040, %f1038, %f1035, %p76;
	selp.f32	%f1041, %f1039, %f1037, %p76;
	add.f32 	%f1042, %f1040, 0fBF800000;
	add.f32 	%f1028, %f1040, 0f3F800000;
	// inline asm
	rcp.approx.ftz.f32 %f1027,%f1028;
	// inline asm
	add.f32 	%f1043, %f1042, %f1042;
	mul.f32 	%f1044, %f1027, %f1043;
	mul.f32 	%f1045, %f1044, %f1044;
	mov.f32 	%f1046, 0f3C4CAF63;
	mov.f32 	%f1047, 0f3B18F0FE;
	fma.rn.f32 	%f1048, %f1047, %f1045, %f1046;
	mov.f32 	%f1049, 0f3DAAAABD;
	fma.rn.f32 	%f1050, %f1048, %f1045, %f1049;
	mul.rn.f32 	%f1051, %f1050, %f1045;
	mul.rn.f32 	%f1052, %f1051, %f1044;
	sub.f32 	%f1053, %f1042, %f1044;
	neg.f32 	%f1054, %f1044;
	add.f32 	%f1055, %f1053, %f1053;
	fma.rn.f32 	%f1056, %f1054, %f1042, %f1055;
	mul.rn.f32 	%f1057, %f1027, %f1056;
	add.f32 	%f1058, %f1052, %f1044;
	sub.f32 	%f1059, %f1044, %f1058;
	add.f32 	%f1060, %f1052, %f1059;
	add.f32 	%f1061, %f1057, %f1060;
	add.f32 	%f1062, %f1058, %f1061;
	sub.f32 	%f1063, %f1058, %f1062;
	add.f32 	%f1064, %f1061, %f1063;
	mov.f32 	%f1065, 0f3F317200;
	mul.rn.f32 	%f1066, %f1041, %f1065;
	mov.f32 	%f1067, 0f35BFBE8E;
	mul.rn.f32 	%f1068, %f1041, %f1067;
	add.f32 	%f1069, %f1066, %f1062;
	sub.f32 	%f1070, %f1066, %f1069;
	add.f32 	%f1071, %f1062, %f1070;
	add.f32 	%f1072, %f1064, %f1071;
	add.f32 	%f1073, %f1068, %f1072;
	add.f32 	%f1074, %f1069, %f1073;
	sub.f32 	%f1075, %f1069, %f1074;
	add.f32 	%f1076, %f1073, %f1075;
	mov.f32 	%f1077, 0f3EE66666;
	mul.rn.f32 	%f1078, %f1077, %f1074;
	neg.f32 	%f1079, %f1078;
	fma.rn.f32 	%f1080, %f1077, %f1074, %f1079;
	fma.rn.f32 	%f1081, %f1077, %f1076, %f1080;
	mov.f32 	%f1082, 0f00000000;
	fma.rn.f32 	%f1083, %f1082, %f1074, %f1081;
	add.rn.f32 	%f1084, %f1078, %f1083;
	neg.f32 	%f1085, %f1084;
	add.rn.f32 	%f1086, %f1078, %f1085;
	add.rn.f32 	%f1087, %f1086, %f1083;
	mov.b32 	 %r134, %f1084;
	setp.eq.s32	%p77, %r134, 1118925336;
	add.s32 	%r135, %r134, -1;
	mov.b32 	 %f1088, %r135;
	add.f32 	%f1089, %f1087, 0f37000000;
	selp.f32	%f1090, %f1088, %f1084, %p77;
	selp.f32	%f248, %f1089, %f1087, %p77;
	mul.f32 	%f1091, %f1090, 0f3FB8AA3B;
	cvt.rzi.f32.f32	%f1092, %f1091;
	mov.f32 	%f1093, 0fBF317200;
	fma.rn.f32 	%f1094, %f1092, %f1093, %f1090;
	mov.f32 	%f1095, 0fB5BFBE8E;
	fma.rn.f32 	%f1096, %f1092, %f1095, %f1094;
	mul.f32 	%f1097, %f1096, 0f3FB8AA3B;
	ex2.approx.ftz.f32 	%f1098, %f1097;
	add.f32 	%f1099, %f1092, 0f00000000;
	ex2.approx.f32 	%f1100, %f1099;
	mul.f32 	%f1101, %f1098, %f1100;
	setp.lt.f32	%p78, %f1090, 0fC2D20000;
	selp.f32	%f1102, 0f00000000, %f1101, %p78;
	setp.gt.f32	%p79, %f1090, 0f42D20000;
	selp.f32	%f1474, 0f7F800000, %f1102, %p79;
	setp.eq.f32	%p80, %f1474, 0f7F800000;
	@%p80 bra 	BB0_63;

	fma.rn.f32 	%f1474, %f1474, %f248, %f1474;

BB0_63:
	setp.lt.f32	%p81, %f1393, 0f00000000;
	setp.eq.f32	%p82, %f246, 0f3F800000;
	and.pred  	%p5, %p81, %p82;
	mov.b32 	 %r136, %f1474;
	xor.b32  	%r137, %r136, -2147483648;
	mov.b32 	 %f1103, %r137;
	selp.f32	%f1476, %f1103, %f1474, %p5;
	setp.eq.f32	%p83, %f1393, 0f00000000;
	@%p83 bra 	BB0_66;
	bra.uni 	BB0_64;

BB0_66:
	add.f32 	%f1106, %f1393, %f1393;
	selp.f32	%f1476, %f1106, 0f00000000, %p82;
	bra.uni 	BB0_67;

BB0_64:
	setp.geu.f32	%p84, %f1393, 0f00000000;
	@%p84 bra 	BB0_67;

	cvt.rzi.f32.f32	%f1105, %f1077;
	setp.neu.f32	%p85, %f1105, 0f3EE66666;
	selp.f32	%f1476, 0f7FFFFFFF, %f1476, %p85;

BB0_67:
	add.f32 	%f1107, %f247, 0f3EE66666;
	mov.b32 	 %r138, %f1107;
	setp.lt.s32	%p87, %r138, 2139095040;
	@%p87 bra 	BB0_72;

	setp.gtu.f32	%p88, %f247, 0f7F800000;
	@%p88 bra 	BB0_71;
	bra.uni 	BB0_69;

BB0_71:
	add.f32 	%f1476, %f1393, 0f3EE66666;
	bra.uni 	BB0_72;

BB0_69:
	setp.neu.f32	%p89, %f247, 0f7F800000;
	@%p89 bra 	BB0_72;

	selp.f32	%f1476, 0fFF800000, 0f7F800000, %p5;

BB0_72:
	setp.eq.f32	%p90, %f1393, 0f3F800000;
	selp.f32	%f259, 0f3F800000, %f1476, %p90;
	abs.f32 	%f260, %f1392;
	setp.lt.f32	%p91, %f260, 0f00800000;
	mul.f32 	%f1110, %f260, 0f4B800000;
	selp.f32	%f1111, 0fC3170000, 0fC2FE0000, %p91;
	selp.f32	%f1112, %f1110, %f260, %p91;
	mov.b32 	 %r139, %f1112;
	and.b32  	%r140, %r139, 8388607;
	or.b32  	%r141, %r140, 1065353216;
	mov.b32 	 %f1113, %r141;
	shr.u32 	%r142, %r139, 23;
	cvt.rn.f32.u32	%f1114, %r142;
	add.f32 	%f1115, %f1111, %f1114;
	setp.gt.f32	%p92, %f1113, 0f3FB504F3;
	mul.f32 	%f1116, %f1113, 0f3F000000;
	add.f32 	%f1117, %f1115, 0f3F800000;
	selp.f32	%f1118, %f1116, %f1113, %p92;
	selp.f32	%f1119, %f1117, %f1115, %p92;
	add.f32 	%f1120, %f1118, 0fBF800000;
	add.f32 	%f1109, %f1118, 0f3F800000;
	// inline asm
	rcp.approx.ftz.f32 %f1108,%f1109;
	// inline asm
	add.f32 	%f1121, %f1120, %f1120;
	mul.f32 	%f1122, %f1108, %f1121;
	mul.f32 	%f1123, %f1122, %f1122;
	fma.rn.f32 	%f1126, %f1047, %f1123, %f1046;
	fma.rn.f32 	%f1128, %f1126, %f1123, %f1049;
	mul.rn.f32 	%f1129, %f1128, %f1123;
	mul.rn.f32 	%f1130, %f1129, %f1122;
	sub.f32 	%f1131, %f1120, %f1122;
	neg.f32 	%f1132, %f1122;
	add.f32 	%f1133, %f1131, %f1131;
	fma.rn.f32 	%f1134, %f1132, %f1120, %f1133;
	mul.rn.f32 	%f1135, %f1108, %f1134;
	add.f32 	%f1136, %f1130, %f1122;
	sub.f32 	%f1137, %f1122, %f1136;
	add.f32 	%f1138, %f1130, %f1137;
	add.f32 	%f1139, %f1135, %f1138;
	add.f32 	%f1140, %f1136, %f1139;
	sub.f32 	%f1141, %f1136, %f1140;
	add.f32 	%f1142, %f1139, %f1141;
	mul.rn.f32 	%f1144, %f1119, %f1065;
	mul.rn.f32 	%f1146, %f1119, %f1067;
	add.f32 	%f1147, %f1144, %f1140;
	sub.f32 	%f1148, %f1144, %f1147;
	add.f32 	%f1149, %f1140, %f1148;
	add.f32 	%f1150, %f1142, %f1149;
	add.f32 	%f1151, %f1146, %f1150;
	add.f32 	%f1152, %f1147, %f1151;
	sub.f32 	%f1153, %f1147, %f1152;
	add.f32 	%f1154, %f1151, %f1153;
	mul.rn.f32 	%f1156, %f1077, %f1152;
	neg.f32 	%f1157, %f1156;
	fma.rn.f32 	%f1158, %f1077, %f1152, %f1157;
	fma.rn.f32 	%f1159, %f1077, %f1154, %f1158;
	fma.rn.f32 	%f1161, %f1082, %f1152, %f1159;
	add.rn.f32 	%f1162, %f1156, %f1161;
	neg.f32 	%f1163, %f1162;
	add.rn.f32 	%f1164, %f1156, %f1163;
	add.rn.f32 	%f1165, %f1164, %f1161;
	mov.b32 	 %r143, %f1162;
	setp.eq.s32	%p93, %r143, 1118925336;
	add.s32 	%r144, %r143, -1;
	mov.b32 	 %f1166, %r144;
	add.f32 	%f1167, %f1165, 0f37000000;
	selp.f32	%f1168, %f1166, %f1162, %p93;
	selp.f32	%f261, %f1167, %f1165, %p93;
	mul.f32 	%f1169, %f1168, 0f3FB8AA3B;
	cvt.rzi.f32.f32	%f1170, %f1169;
	fma.rn.f32 	%f1172, %f1170, %f1093, %f1168;
	fma.rn.f32 	%f1174, %f1170, %f1095, %f1172;
	mul.f32 	%f1175, %f1174, 0f3FB8AA3B;
	ex2.approx.ftz.f32 	%f1176, %f1175;
	add.f32 	%f1177, %f1170, 0f00000000;
	ex2.approx.f32 	%f1178, %f1177;
	mul.f32 	%f1179, %f1176, %f1178;
	setp.lt.f32	%p94, %f1168, 0fC2D20000;
	selp.f32	%f1180, 0f00000000, %f1179, %p94;
	setp.gt.f32	%p95, %f1168, 0f42D20000;
	selp.f32	%f1477, 0f7F800000, %f1180, %p95;
	setp.eq.f32	%p96, %f1477, 0f7F800000;
	@%p96 bra 	BB0_74;

	fma.rn.f32 	%f1477, %f1477, %f261, %f1477;

BB0_74:
	setp.lt.f32	%p97, %f1392, 0f00000000;
	and.pred  	%p6, %p97, %p82;
	mov.b32 	 %r145, %f1477;
	xor.b32  	%r146, %r145, -2147483648;
	mov.b32 	 %f1181, %r146;
	selp.f32	%f1479, %f1181, %f1477, %p6;
	setp.eq.f32	%p99, %f1392, 0f00000000;
	@%p99 bra 	BB0_77;
	bra.uni 	BB0_75;

BB0_77:
	add.f32 	%f1184, %f1392, %f1392;
	selp.f32	%f1479, %f1184, 0f00000000, %p82;
	bra.uni 	BB0_78;

BB0_75:
	setp.geu.f32	%p100, %f1392, 0f00000000;
	@%p100 bra 	BB0_78;

	cvt.rzi.f32.f32	%f1183, %f1077;
	setp.neu.f32	%p101, %f1183, 0f3EE66666;
	selp.f32	%f1479, 0f7FFFFFFF, %f1479, %p101;

BB0_78:
	add.f32 	%f1185, %f260, 0f3EE66666;
	mov.b32 	 %r147, %f1185;
	setp.lt.s32	%p103, %r147, 2139095040;
	@%p103 bra 	BB0_83;

	setp.gtu.f32	%p104, %f260, 0f7F800000;
	@%p104 bra 	BB0_82;
	bra.uni 	BB0_80;

BB0_82:
	add.f32 	%f1479, %f1392, 0f3EE66666;
	bra.uni 	BB0_83;

BB0_80:
	setp.neu.f32	%p105, %f260, 0f7F800000;
	@%p105 bra 	BB0_83;

	selp.f32	%f1479, 0fFF800000, 0f7F800000, %p6;

BB0_83:
	setp.eq.f32	%p106, %f1392, 0f3F800000;
	selp.f32	%f272, 0f3F800000, %f1479, %p106;
	abs.f32 	%f273, %f1391;
	setp.lt.f32	%p107, %f273, 0f00800000;
	mul.f32 	%f1188, %f273, 0f4B800000;
	selp.f32	%f1189, 0fC3170000, 0fC2FE0000, %p107;
	selp.f32	%f1190, %f1188, %f273, %p107;
	mov.b32 	 %r148, %f1190;
	and.b32  	%r149, %r148, 8388607;
	or.b32  	%r150, %r149, 1065353216;
	mov.b32 	 %f1191, %r150;
	shr.u32 	%r151, %r148, 23;
	cvt.rn.f32.u32	%f1192, %r151;
	add.f32 	%f1193, %f1189, %f1192;
	setp.gt.f32	%p108, %f1191, 0f3FB504F3;
	mul.f32 	%f1194, %f1191, 0f3F000000;
	add.f32 	%f1195, %f1193, 0f3F800000;
	selp.f32	%f1196, %f1194, %f1191, %p108;
	selp.f32	%f1197, %f1195, %f1193, %p108;
	add.f32 	%f1198, %f1196, 0fBF800000;
	add.f32 	%f1187, %f1196, 0f3F800000;
	// inline asm
	rcp.approx.ftz.f32 %f1186,%f1187;
	// inline asm
	add.f32 	%f1199, %f1198, %f1198;
	mul.f32 	%f1200, %f1186, %f1199;
	mul.f32 	%f1201, %f1200, %f1200;
	fma.rn.f32 	%f1204, %f1047, %f1201, %f1046;
	fma.rn.f32 	%f1206, %f1204, %f1201, %f1049;
	mul.rn.f32 	%f1207, %f1206, %f1201;
	mul.rn.f32 	%f1208, %f1207, %f1200;
	sub.f32 	%f1209, %f1198, %f1200;
	neg.f32 	%f1210, %f1200;
	add.f32 	%f1211, %f1209, %f1209;
	fma.rn.f32 	%f1212, %f1210, %f1198, %f1211;
	mul.rn.f32 	%f1213, %f1186, %f1212;
	add.f32 	%f1214, %f1208, %f1200;
	sub.f32 	%f1215, %f1200, %f1214;
	add.f32 	%f1216, %f1208, %f1215;
	add.f32 	%f1217, %f1213, %f1216;
	add.f32 	%f1218, %f1214, %f1217;
	sub.f32 	%f1219, %f1214, %f1218;
	add.f32 	%f1220, %f1217, %f1219;
	mul.rn.f32 	%f1222, %f1197, %f1065;
	mul.rn.f32 	%f1224, %f1197, %f1067;
	add.f32 	%f1225, %f1222, %f1218;
	sub.f32 	%f1226, %f1222, %f1225;
	add.f32 	%f1227, %f1218, %f1226;
	add.f32 	%f1228, %f1220, %f1227;
	add.f32 	%f1229, %f1224, %f1228;
	add.f32 	%f1230, %f1225, %f1229;
	sub.f32 	%f1231, %f1225, %f1230;
	add.f32 	%f1232, %f1229, %f1231;
	mul.rn.f32 	%f1234, %f1077, %f1230;
	neg.f32 	%f1235, %f1234;
	fma.rn.f32 	%f1236, %f1077, %f1230, %f1235;
	fma.rn.f32 	%f1237, %f1077, %f1232, %f1236;
	fma.rn.f32 	%f1239, %f1082, %f1230, %f1237;
	add.rn.f32 	%f1240, %f1234, %f1239;
	neg.f32 	%f1241, %f1240;
	add.rn.f32 	%f1242, %f1234, %f1241;
	add.rn.f32 	%f1243, %f1242, %f1239;
	mov.b32 	 %r152, %f1240;
	setp.eq.s32	%p109, %r152, 1118925336;
	add.s32 	%r153, %r152, -1;
	mov.b32 	 %f1244, %r153;
	add.f32 	%f1245, %f1243, 0f37000000;
	selp.f32	%f1246, %f1244, %f1240, %p109;
	selp.f32	%f274, %f1245, %f1243, %p109;
	mul.f32 	%f1247, %f1246, 0f3FB8AA3B;
	cvt.rzi.f32.f32	%f1248, %f1247;
	fma.rn.f32 	%f1250, %f1248, %f1093, %f1246;
	fma.rn.f32 	%f1252, %f1248, %f1095, %f1250;
	mul.f32 	%f1253, %f1252, 0f3FB8AA3B;
	ex2.approx.ftz.f32 	%f1254, %f1253;
	add.f32 	%f1255, %f1248, 0f00000000;
	ex2.approx.f32 	%f1256, %f1255;
	mul.f32 	%f1257, %f1254, %f1256;
	setp.lt.f32	%p110, %f1246, 0fC2D20000;
	selp.f32	%f1258, 0f00000000, %f1257, %p110;
	setp.gt.f32	%p111, %f1246, 0f42D20000;
	selp.f32	%f1480, 0f7F800000, %f1258, %p111;
	setp.eq.f32	%p112, %f1480, 0f7F800000;
	@%p112 bra 	BB0_85;

	fma.rn.f32 	%f1480, %f1480, %f274, %f1480;

BB0_85:
	setp.lt.f32	%p113, %f1391, 0f00000000;
	and.pred  	%p7, %p113, %p82;
	mov.b32 	 %r154, %f1480;
	xor.b32  	%r155, %r154, -2147483648;
	mov.b32 	 %f1259, %r155;
	selp.f32	%f1482, %f1259, %f1480, %p7;
	setp.eq.f32	%p115, %f1391, 0f00000000;
	@%p115 bra 	BB0_88;
	bra.uni 	BB0_86;

BB0_88:
	add.f32 	%f1262, %f1391, %f1391;
	selp.f32	%f1482, %f1262, 0f00000000, %p82;
	bra.uni 	BB0_89;

BB0_86:
	setp.geu.f32	%p116, %f1391, 0f00000000;
	@%p116 bra 	BB0_89;

	cvt.rzi.f32.f32	%f1261, %f1077;
	setp.neu.f32	%p117, %f1261, 0f3EE66666;
	selp.f32	%f1482, 0f7FFFFFFF, %f1482, %p117;

BB0_89:
	add.f32 	%f1263, %f273, 0f3EE66666;
	mov.b32 	 %r156, %f1263;
	setp.lt.s32	%p119, %r156, 2139095040;
	@%p119 bra 	BB0_94;

	setp.gtu.f32	%p120, %f273, 0f7F800000;
	@%p120 bra 	BB0_93;
	bra.uni 	BB0_91;

BB0_93:
	add.f32 	%f1482, %f1391, 0f3EE66666;
	bra.uni 	BB0_94;

BB0_91:
	setp.neu.f32	%p121, %f273, 0f7F800000;
	@%p121 bra 	BB0_94;

	selp.f32	%f1482, 0fFF800000, 0f7F800000, %p7;

BB0_94:
	mov.u32 	%r241, 4;
	setp.eq.f32	%p122, %f1391, 0f3F800000;
	selp.f32	%f1264, 0f3F800000, %f1482, %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, %r241, %rd52, %rd53, %rd13, %rd13);
	// inline asm
	cvt.sat.f32.f32	%f1265, %f1264;
	mul.f32 	%f1266, %f1265, 0f437FFD71;
	cvt.rzi.u32.f32	%r159, %f1266;
	cvt.sat.f32.f32	%f1267, %f272;
	mul.f32 	%f1268, %f1267, 0f437FFD71;
	cvt.rzi.u32.f32	%r160, %f1268;
	cvt.sat.f32.f32	%f1269, %f259;
	mul.f32 	%f1270, %f1269, 0f437FFD71;
	cvt.rzi.u32.f32	%r161, %f1270;
	cvt.u16.u32	%rs19, %r159;
	cvt.u16.u32	%rs20, %r161;
	cvt.u16.u32	%rs21, %r160;
	mov.u16 	%rs22, 255;
	st.v4.u8 	[%rd50], {%rs19, %rs21, %rs20, %rs22};
	ld.global.u32 	%r248, [imageEnabled];

BB0_95:
	cvt.u64.u32	%rd4, %r2;
	cvt.u64.u32	%rd5, %r3;
	and.b32  	%r162, %r248, 4;
	setp.eq.s32	%p123, %r162, 0;
	@%p123 bra 	BB0_99;

	ld.global.u32 	%r163, [additive];
	setp.eq.s32	%p124, %r163, 0;
	mov.f32 	%f1271, 0f3F800000;
	// inline asm
	{  cvt.rn.f16.f32 %rs23, %f1271;}

	// inline asm
	@%p124 bra 	BB0_98;

	mov.u64 	%rd69, image_HDR;
	cvta.global.u64 	%rd58, %rd69;
	mov.u32 	%r167, 8;
	// inline asm
	call (%rd57), _rt_buffer_get_64, (%rd58, %r28, %r167, %rd4, %rd5, %rd13, %rd13);
	// inline asm
	ld.v4.u16 	{%rs30, %rs31, %rs32, %rs33}, [%rd57];
	// inline asm
	{  cvt.f32.f16 %f1272, %rs30;}

	// inline asm
	// inline asm
	{  cvt.f32.f16 %f1273, %rs31;}

	// inline asm
	// inline asm
	{  cvt.f32.f16 %f1274, %rs32;}

	// inline asm
	// inline asm
	call (%rd63), _rt_buffer_get_64, (%rd58, %r28, %r167, %rd4, %rd5, %rd13, %rd13);
	// inline asm
	add.f32 	%f1275, %f1393, %f1272;
	add.f32 	%f1276, %f1392, %f1273;
	add.f32 	%f1277, %f1391, %f1274;
	// inline asm
	{  cvt.rn.f16.f32 %rs29, %f1277;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs28, %f1276;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs27, %f1275;}

	// 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 	%r169, 8;
	// inline asm
	call (%rd70), _rt_buffer_get_64, (%rd71, %r28, %r169, %rd4, %rd5, %rd13, %rd13);
	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs36, %f1391;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs35, %f1392;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs34, %f1393;}

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

BB0_99:
	mov.f32 	%f1282, 0f34000000;
	max.f32 	%f1283, %f1390, %f1282;
	div.rn.f32 	%f1284, %f1387, %f1283;
	max.f32 	%f1285, %f1389, %f1282;
	div.rn.f32 	%f1286, %f1386, %f1285;
	max.f32 	%f1287, %f1388, %f1282;
	div.rn.f32 	%f1288, %f1385, %f1287;
	fma.rn.f32 	%f285, %f1284, 0f3F000000, 0f3F000000;
	fma.rn.f32 	%f286, %f1286, 0f3F000000, 0f3F000000;
	fma.rn.f32 	%f287, %f1288, 0f3F000000, 0f3F000000;
	div.rn.f32 	%f1289, %f1384, %f1283;
	div.rn.f32 	%f1290, %f1383, %f1285;
	div.rn.f32 	%f1291, %f1382, %f1287;
	fma.rn.f32 	%f288, %f1289, 0f3F000000, 0f3F000000;
	fma.rn.f32 	%f289, %f1290, 0f3F000000, 0f3F000000;
	fma.rn.f32 	%f290, %f1291, 0f3F000000, 0f3F000000;
	div.rn.f32 	%f1292, %f1381, %f1283;
	div.rn.f32 	%f1293, %f1380, %f1285;
	div.rn.f32 	%f1294, %f1379, %f1287;
	fma.rn.f32 	%f291, %f1292, 0f3F000000, 0f3F000000;
	fma.rn.f32 	%f292, %f1293, 0f3F000000, 0f3F000000;
	fma.rn.f32 	%f293, %f1294, 0f3F000000, 0f3F000000;
	ld.global.u32 	%r170, [additive];
	setp.eq.s32	%p125, %r170, 0;
	mov.f32 	%f1281, 0f3F800000;
	// inline asm
	{  cvt.rn.f16.f32 %rs37, %f1281;}

	// inline asm
	@%p125 bra 	BB0_101;

	mov.u64 	%rd89, image_RNM0;
	cvta.global.u64 	%rd78, %rd89;
	mov.u32 	%r174, 8;
	// inline asm
	call (%rd77), _rt_buffer_get_64, (%rd78, %r28, %r174, %rd4, %rd5, %rd13, %rd13);
	// inline asm
	ld.v4.u16 	{%rs44, %rs45, %rs46, %rs47}, [%rd77];
	// inline asm
	{  cvt.f32.f16 %f1295, %rs44;}

	// inline asm
	// inline asm
	{  cvt.f32.f16 %f1296, %rs45;}

	// inline asm
	// inline asm
	{  cvt.f32.f16 %f1297, %rs46;}

	// inline asm
	// inline asm
	call (%rd83), _rt_buffer_get_64, (%rd78, %r28, %r174, %rd4, %rd5, %rd13, %rd13);
	// inline asm
	add.f32 	%f1298, %f1390, %f1295;
	add.f32 	%f1299, %f1389, %f1296;
	add.f32 	%f1300, %f1388, %f1297;
	// inline asm
	{  cvt.rn.f16.f32 %rs43, %f1300;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs42, %f1299;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs41, %f1298;}

	// 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 	%r176, 8;
	// inline asm
	call (%rd90), _rt_buffer_get_64, (%rd91, %r28, %r176, %rd4, %rd5, %rd13, %rd13);
	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs50, %f1388;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs49, %f1389;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs48, %f1390;}

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

BB0_102:
	ld.global.u32 	%r177, [additive];
	setp.eq.s32	%p126, %r177, 0;
	// inline asm
	{  cvt.rn.f16.f32 %rs51, %f1281;}

	// inline asm
	@%p126 bra 	BB0_104;

	mov.u64 	%rd109, image_RNM1;
	cvta.global.u64 	%rd98, %rd109;
	mov.u32 	%r181, 8;
	// inline asm
	call (%rd97), _rt_buffer_get_64, (%rd98, %r28, %r181, %rd4, %rd5, %rd13, %rd13);
	// inline asm
	ld.v4.u16 	{%rs58, %rs59, %rs60, %rs61}, [%rd97];
	// inline asm
	{  cvt.f32.f16 %f1305, %rs58;}

	// inline asm
	// inline asm
	{  cvt.f32.f16 %f1306, %rs59;}

	// inline asm
	// inline asm
	{  cvt.f32.f16 %f1307, %rs60;}

	// inline asm
	// inline asm
	call (%rd103), _rt_buffer_get_64, (%rd98, %r28, %r181, %rd4, %rd5, %rd13, %rd13);
	// inline asm
	add.f32 	%f1308, %f285, %f1305;
	add.f32 	%f1309, %f286, %f1306;
	add.f32 	%f1310, %f287, %f1307;
	// inline asm
	{  cvt.rn.f16.f32 %rs57, %f1310;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs56, %f1309;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs55, %f1308;}

	// 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 	%r183, 8;
	// inline asm
	call (%rd110), _rt_buffer_get_64, (%rd111, %r28, %r183, %rd4, %rd5, %rd13, %rd13);
	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs64, %f287;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs63, %f286;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs62, %f285;}

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

BB0_105:
	ld.global.u32 	%r184, [additive];
	setp.eq.s32	%p127, %r184, 0;
	// inline asm
	{  cvt.rn.f16.f32 %rs65, %f1281;}

	// inline asm
	@%p127 bra 	BB0_107;

	mov.u64 	%rd129, image_RNM2;
	cvta.global.u64 	%rd118, %rd129;
	mov.u32 	%r188, 8;
	// inline asm
	call (%rd117), _rt_buffer_get_64, (%rd118, %r28, %r188, %rd4, %rd5, %rd13, %rd13);
	// inline asm
	ld.v4.u16 	{%rs72, %rs73, %rs74, %rs75}, [%rd117];
	// inline asm
	{  cvt.f32.f16 %f1315, %rs72;}

	// inline asm
	// inline asm
	{  cvt.f32.f16 %f1316, %rs73;}

	// inline asm
	// inline asm
	{  cvt.f32.f16 %f1317, %rs74;}

	// inline asm
	// inline asm
	call (%rd123), _rt_buffer_get_64, (%rd118, %r28, %r188, %rd4, %rd5, %rd13, %rd13);
	// inline asm
	add.f32 	%f1318, %f288, %f1315;
	add.f32 	%f1319, %f289, %f1316;
	add.f32 	%f1320, %f290, %f1317;
	// inline asm
	{  cvt.rn.f16.f32 %rs71, %f1320;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs70, %f1319;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs69, %f1318;}

	// 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 	%r190, 8;
	// inline asm
	call (%rd130), _rt_buffer_get_64, (%rd131, %r28, %r190, %rd4, %rd5, %rd13, %rd13);
	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs78, %f290;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs77, %f289;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs76, %f288;}

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

BB0_108:
	ld.global.u32 	%r191, [additive];
	setp.eq.s32	%p128, %r191, 0;
	// inline asm
	{  cvt.rn.f16.f32 %rs79, %f1281;}

	// inline asm
	@%p128 bra 	BB0_110;

	mov.u64 	%rd149, image_RNM3;
	cvta.global.u64 	%rd138, %rd149;
	mov.u32 	%r195, 8;
	// inline asm
	call (%rd137), _rt_buffer_get_64, (%rd138, %r28, %r195, %rd4, %rd5, %rd13, %rd13);
	// inline asm
	ld.v4.u16 	{%rs86, %rs87, %rs88, %rs89}, [%rd137];
	// inline asm
	{  cvt.f32.f16 %f1325, %rs86;}

	// inline asm
	// inline asm
	{  cvt.f32.f16 %f1326, %rs87;}

	// inline asm
	// inline asm
	{  cvt.f32.f16 %f1327, %rs88;}

	// inline asm
	// inline asm
	call (%rd143), _rt_buffer_get_64, (%rd138, %r28, %r195, %rd4, %rd5, %rd13, %rd13);
	// inline asm
	add.f32 	%f1328, %f291, %f1325;
	add.f32 	%f1329, %f292, %f1326;
	add.f32 	%f1330, %f293, %f1327;
	// inline asm
	{  cvt.rn.f16.f32 %rs85, %f1330;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs84, %f1329;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs83, %f1328;}

	// 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 	%r197, 8;
	// inline asm
	call (%rd150), _rt_buffer_get_64, (%rd151, %r28, %r197, %rd4, %rd5, %rd13, %rd13);
	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs92, %f293;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs91, %f292;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs90, %f291;}

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

BB0_131:
	ret;
}