//
// 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 image_RNM4[1];
.global .align 1 .b8 image_RNM5[1];
.global .align 1 .b8 image_RNM6[1];
.global .align 1 .b8 image_RNM7[1];
.global .align 1 .b8 image_RNM8[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<148>;
	.reg .b16 	%rs<302>;
	.reg .f32 	%f<1812>;
	.reg .b32 	%r<327>;
	.reg .b64 	%rd<474>;


	mov.u64 	%rd473, __local_depot0;
	cvta.local.u64 	%SP, %rd473;
	ld.global.v2.u32 	{%r29, %r30}, [pixelID];
	cvt.u64.u32	%rd9, %r29;
	cvt.u64.u32	%rd10, %r30;
	mov.u64 	%rd13, uvnormal;
	cvta.global.u64 	%rd8, %rd13;
	mov.u32 	%r27, 2;
	mov.u32 	%r28, 4;
	mov.u64 	%rd12, 0;
	// inline asm
	call (%rd7), _rt_buffer_get_64, (%rd8, %r27, %r28, %rd9, %rd10, %rd12, %rd12);
	// inline asm
	ld.u32 	%r1, [%rd7];
	shr.u32 	%r33, %r1, 16;
	cvt.u16.u32	%rs1, %r33;
	and.b16  	%rs14, %rs1, 255;
	cvt.u16.u32	%rs15, %r1;
	or.b16  	%rs16, %rs15, %rs14;
	setp.eq.s16	%p8, %rs16, 0;
	mov.f32 	%f1644, 0f00000000;
	mov.f32 	%f1645, %f1644;
	mov.f32 	%f1646, %f1644;
	@%p8 bra 	BB0_2;

	ld.u8 	%rs17, [%rd7+1];
	and.b16  	%rs19, %rs15, 255;
	cvt.rn.f32.u16	%f409, %rs19;
	div.rn.f32 	%f410, %f409, 0f437F0000;
	fma.rn.f32 	%f411, %f410, 0f40000000, 0fBF800000;
	cvt.rn.f32.u16	%f412, %rs17;
	div.rn.f32 	%f413, %f412, 0f437F0000;
	fma.rn.f32 	%f414, %f413, 0f40000000, 0fBF800000;
	cvt.rn.f32.u16	%f415, %rs14;
	div.rn.f32 	%f416, %f415, 0f437F0000;
	fma.rn.f32 	%f417, %f416, 0f40000000, 0fBF800000;
	mul.f32 	%f418, %f414, %f414;
	fma.rn.f32 	%f419, %f411, %f411, %f418;
	fma.rn.f32 	%f420, %f417, %f417, %f419;
	sqrt.rn.f32 	%f421, %f420;
	rcp.rn.f32 	%f422, %f421;
	mul.f32 	%f1644, %f411, %f422;
	mul.f32 	%f1645, %f414, %f422;
	mul.f32 	%f1646, %f417, %f422;

BB0_2:
	ld.global.v2.u32 	{%r34, %r35}, [pixelID];
	ld.global.v2.u32 	{%r37, %r38}, [tileInfo];
	add.s32 	%r2, %r34, %r37;
	add.s32 	%r3, %r35, %r38;
	setp.eq.f32	%p9, %f1645, 0f00000000;
	setp.eq.f32	%p10, %f1644, 0f00000000;
	and.pred  	%p11, %p10, %p9;
	setp.eq.f32	%p12, %f1646, 0f00000000;
	and.pred  	%p13, %p11, %p12;
	@%p13 bra 	BB0_126;
	bra.uni 	BB0_3;

BB0_126:
	ld.global.u32 	%r326, [imageEnabled];
	and.b32  	%r239, %r326, 1;
	setp.eq.b32	%p134, %r239, 1;
	@!%p134 bra 	BB0_128;
	bra.uni 	BB0_127;

BB0_127:
	cvt.u64.u32	%rd261, %r2;
	cvt.u64.u32	%rd262, %r3;
	mov.u64 	%rd265, image;
	cvta.global.u64 	%rd260, %rd265;
	// inline asm
	call (%rd259), _rt_buffer_get_64, (%rd260, %r27, %r28, %rd261, %rd262, %rd12, %rd12);
	// inline asm
	mov.u16 	%rs168, 0;
	st.v4.u8 	[%rd259], {%rs168, %rs168, %rs168, %rs168};
	ld.global.u32 	%r326, [imageEnabled];

BB0_128:
	and.b32  	%r242, %r326, 8;
	setp.eq.s32	%p135, %r242, 0;
	@%p135 bra 	BB0_130;

	cvt.u64.u32	%rd269, %r3;
	cvt.u64.u32	%rd268, %r2;
	mov.u64 	%rd272, image_Mask;
	cvta.global.u64 	%rd267, %rd272;
	// inline asm
	call (%rd266), _rt_buffer_get_64, (%rd267, %r27, %r27, %rd268, %rd269, %rd12, %rd12);
	// inline asm
	mov.f32 	%f1570, 0f00000000;
	cvt.rzi.u32.f32	%r245, %f1570;
	cvt.u16.u32	%rs169, %r245;
	mov.u16 	%rs170, 0;
	st.v2.u8 	[%rd266], {%rs169, %rs170};
	ld.global.u32 	%r326, [imageEnabled];

BB0_130:
	cvt.u64.u32	%rd5, %r2;
	cvt.u64.u32	%rd6, %r3;
	and.b32  	%r246, %r326, 4;
	setp.eq.s32	%p136, %r246, 0;
	@%p136 bra 	BB0_134;

	ld.global.u32 	%r247, [additive];
	setp.eq.s32	%p137, %r247, 0;
	@%p137 bra 	BB0_133;

	mov.u64 	%rd285, image_HDR;
	cvta.global.u64 	%rd274, %rd285;
	mov.u32 	%r251, 8;
	// inline asm
	call (%rd273), _rt_buffer_get_64, (%rd274, %r27, %r251, %rd5, %rd6, %rd12, %rd12);
	// inline asm
	ld.v4.u16 	{%rs177, %rs178, %rs179, %rs180}, [%rd273];
	// inline asm
	{  cvt.f32.f16 %f1571, %rs177;}

	// inline asm
	// inline asm
	{  cvt.f32.f16 %f1572, %rs178;}

	// inline asm
	// inline asm
	{  cvt.f32.f16 %f1573, %rs179;}

	// inline asm
	// inline asm
	call (%rd279), _rt_buffer_get_64, (%rd274, %r27, %r251, %rd5, %rd6, %rd12, %rd12);
	// inline asm
	add.f32 	%f1574, %f1571, 0f00000000;
	add.f32 	%f1575, %f1572, 0f00000000;
	add.f32 	%f1576, %f1573, 0f00000000;
	// inline asm
	{  cvt.rn.f16.f32 %rs176, %f1576;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs175, %f1575;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs174, %f1574;}

	// inline asm
	mov.u16 	%rs181, 0;
	st.v4.u16 	[%rd279], {%rs174, %rs175, %rs176, %rs181};
	bra.uni 	BB0_134;

BB0_3:
	ld.global.v2.u32 	{%r46, %r47}, [pixelID];
	cvt.u64.u32	%rd16, %r46;
	cvt.u64.u32	%rd17, %r47;
	mov.u64 	%rd25, uvpos;
	cvta.global.u64 	%rd15, %rd25;
	mov.u32 	%r43, 12;
	// inline asm
	call (%rd14), _rt_buffer_get_64, (%rd15, %r27, %r43, %rd16, %rd17, %rd12, %rd12);
	// inline asm
	ld.f32 	%f9, [%rd14+8];
	ld.f32 	%f8, [%rd14+4];
	ld.f32 	%f7, [%rd14];
	mul.f32 	%f454, %f7, 0f3456BF95;
	mul.f32 	%f455, %f8, 0f3456BF95;
	mul.f32 	%f456, %f9, 0f3456BF95;
	abs.f32 	%f457, %f1644;
	div.rn.f32 	%f458, %f454, %f457;
	abs.f32 	%f459, %f1645;
	div.rn.f32 	%f460, %f455, %f459;
	abs.f32 	%f461, %f1646;
	div.rn.f32 	%f462, %f456, %f461;
	abs.f32 	%f463, %f458;
	abs.f32 	%f464, %f460;
	abs.f32 	%f465, %f462;
	mov.f32 	%f466, 0f38D1B717;
	max.f32 	%f467, %f463, %f466;
	max.f32 	%f468, %f464, %f466;
	max.f32 	%f469, %f465, %f466;
	fma.rn.f32 	%f10, %f1644, %f467, %f7;
	fma.rn.f32 	%f11, %f1645, %f468, %f8;
	fma.rn.f32 	%f12, %f1646, %f469, %f9;
	ld.global.v2.u32 	{%r50, %r51}, [pixelID];
	cvt.rn.f32.u32	%f13, %r50;
	cvt.rn.f32.u32	%f14, %r51;
	mov.u64 	%rd26, localLights;
	cvta.global.u64 	%rd24, %rd26;
	mov.u32 	%r44, 1;
	mov.u32 	%r45, 96;
	// inline asm
	call (%rd20, %rd21, %rd22, %rd23), _rt_buffer_get_size_64, (%rd24, %r44, %r45);
	// inline asm
	cvt.u32.u64	%r4, %rd20;
	setp.eq.s32	%p14, %r4, 0;
	mov.f32 	%f1647, 0f00000000;
	mov.f32 	%f1648, %f1647;
	mov.f32 	%f1649, %f1647;
	mov.f32 	%f1650, %f1647;
	mov.f32 	%f1651, %f1647;
	mov.f32 	%f1652, %f1647;
	mov.f32 	%f1653, %f1647;
	mov.f32 	%f1654, %f1647;
	mov.f32 	%f1655, %f1647;
	mov.f32 	%f1656, %f1647;
	mov.f32 	%f1657, %f1647;
	mov.f32 	%f1658, %f1647;
	mov.f32 	%f1659, %f1647;
	mov.f32 	%f1660, %f1647;
	mov.f32 	%f1661, %f1647;
	mov.f32 	%f1662, %f1647;
	mov.f32 	%f1663, %f1647;
	mov.f32 	%f1664, %f1647;
	mov.f32 	%f1665, %f1647;
	mov.f32 	%f1666, %f1647;
	mov.f32 	%f1667, %f1647;
	mov.f32 	%f1668, %f1647;
	mov.f32 	%f1669, %f1647;
	mov.f32 	%f1670, %f1647;
	mov.f32 	%f1671, %f1647;
	mov.f32 	%f1672, %f1647;
	mov.f32 	%f1673, %f1647;
	mov.f32 	%f1674, %f1647;
	mov.f32 	%f1675, %f1647;
	mov.f32 	%f1676, %f1647;
	mov.f32 	%f1677, %f1647;
	@%p14 bra 	BB0_47;

	mov.f32 	%f501, 0f40000000;
	cvt.rzi.f32.f32	%f502, %f501;
	add.f32 	%f503, %f502, %f502;
	mov.f32 	%f504, 0f40800000;
	sub.f32 	%f505, %f504, %f503;
	abs.f32 	%f15, %f505;
	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 	%f500, 0f00000000;
	mov.u32 	%r318, 0;
	abs.f32 	%f726, %f16;
	abs.f32 	%f727, %f17;
	max.f32 	%f728, %f726, %f727;
	abs.f32 	%f729, %f18;
	max.f32 	%f730, %f728, %f729;
	mov.f32 	%f1647, %f500;
	mov.f32 	%f1648, %f500;
	mov.f32 	%f1649, %f500;
	mov.f32 	%f1650, %f500;
	mov.f32 	%f1651, %f500;
	mov.f32 	%f1652, %f500;
	mov.f32 	%f1653, %f500;
	mov.f32 	%f1654, %f500;
	mov.f32 	%f1655, %f500;
	mov.f32 	%f1656, %f500;
	mov.f32 	%f1657, %f500;
	mov.f32 	%f1658, %f500;
	mov.f32 	%f1659, %f500;
	mov.f32 	%f1660, %f500;
	mov.f32 	%f1661, %f500;
	mov.f32 	%f1662, %f500;
	mov.f32 	%f1663, %f500;
	mov.f32 	%f1664, %f500;
	mov.f32 	%f1665, %f500;
	mov.f32 	%f1666, %f500;
	mov.f32 	%f1667, %f500;
	mov.f32 	%f1668, %f500;
	mov.f32 	%f1669, %f500;
	mov.f32 	%f1670, %f500;
	mov.f32 	%f1671, %f500;
	mov.f32 	%f1672, %f500;
	mov.f32 	%f1673, %f500;
	mov.f32 	%f1674, %f500;
	mov.f32 	%f1675, %f500;
	mov.f32 	%f1676, %f500;
	mov.f32 	%f1677, %f500;

BB0_5:
	cvt.u64.u32	%rd29, %r318;
	// inline asm
	call (%rd27), _rt_buffer_get_64, (%rd24, %r44, %r45, %rd29, %rd12, %rd12, %rd12);
	// inline asm
	ld.v4.f32 	{%f508, %f509, %f510, %f511}, [%rd27+80];
	ld.v4.f32 	{%f512, %f513, %f514, %f515}, [%rd27+64];
	ld.v4.f32 	{%f516, %f517, %f518, %f519}, [%rd27+48];
	ld.v4.f32 	{%f520, %f1686, %f1687, %f523}, [%rd27+32];
	ld.v4.f32 	{%f524, %f525, %f526, %f527}, [%rd27+16];
	ld.v4.f32 	{%f528, %f529, %f530, %f531}, [%rd27];
	mov.b32 	 %r6, %f511;
	sub.f32 	%f76, %f529, %f7;
	sub.f32 	%f78, %f530, %f8;
	sub.f32 	%f80, %f531, %f9;
	mul.f32 	%f533, %f78, %f78;
	fma.rn.f32 	%f534, %f76, %f76, %f533;
	fma.rn.f32 	%f535, %f80, %f80, %f534;
	sqrt.rn.f32 	%f81, %f535;
	rcp.rn.f32 	%f536, %f81;
	mul.f32 	%f82, %f76, %f536;
	mul.f32 	%f83, %f78, %f536;
	mul.f32 	%f84, %f80, %f536;
	mul.f32 	%f85, %f81, %f527;
	abs.f32 	%f86, %f85;
	setp.lt.f32	%p15, %f86, 0f00800000;
	mul.f32 	%f537, %f86, 0f4B800000;
	selp.f32	%f538, 0fC3170000, 0fC2FE0000, %p15;
	selp.f32	%f539, %f537, %f86, %p15;
	mov.b32 	 %r57, %f539;
	and.b32  	%r58, %r57, 8388607;
	or.b32  	%r59, %r58, 1065353216;
	mov.b32 	 %f540, %r59;
	shr.u32 	%r60, %r57, 23;
	cvt.rn.f32.u32	%f541, %r60;
	add.f32 	%f542, %f538, %f541;
	setp.gt.f32	%p16, %f540, 0f3FB504F3;
	mul.f32 	%f543, %f540, 0f3F000000;
	add.f32 	%f544, %f542, 0f3F800000;
	selp.f32	%f545, %f543, %f540, %p16;
	selp.f32	%f546, %f544, %f542, %p16;
	add.f32 	%f547, %f545, 0fBF800000;
	add.f32 	%f507, %f545, 0f3F800000;
	// inline asm
	rcp.approx.ftz.f32 %f506,%f507;
	// inline asm
	add.f32 	%f548, %f547, %f547;
	mul.f32 	%f549, %f506, %f548;
	mul.f32 	%f550, %f549, %f549;
	mov.f32 	%f551, 0f3C4CAF63;
	mov.f32 	%f552, 0f3B18F0FE;
	fma.rn.f32 	%f553, %f552, %f550, %f551;
	mov.f32 	%f554, 0f3DAAAABD;
	fma.rn.f32 	%f555, %f553, %f550, %f554;
	mul.rn.f32 	%f556, %f555, %f550;
	mul.rn.f32 	%f557, %f556, %f549;
	sub.f32 	%f558, %f547, %f549;
	neg.f32 	%f559, %f549;
	add.f32 	%f560, %f558, %f558;
	fma.rn.f32 	%f561, %f559, %f547, %f560;
	mul.rn.f32 	%f562, %f506, %f561;
	add.f32 	%f563, %f557, %f549;
	sub.f32 	%f564, %f549, %f563;
	add.f32 	%f565, %f557, %f564;
	add.f32 	%f566, %f562, %f565;
	add.f32 	%f567, %f563, %f566;
	sub.f32 	%f568, %f563, %f567;
	add.f32 	%f569, %f566, %f568;
	mov.f32 	%f570, 0f3F317200;
	mul.rn.f32 	%f571, %f546, %f570;
	mov.f32 	%f572, 0f35BFBE8E;
	mul.rn.f32 	%f573, %f546, %f572;
	add.f32 	%f574, %f571, %f567;
	sub.f32 	%f575, %f571, %f574;
	add.f32 	%f576, %f567, %f575;
	add.f32 	%f577, %f569, %f576;
	add.f32 	%f578, %f573, %f577;
	add.f32 	%f579, %f574, %f578;
	sub.f32 	%f580, %f574, %f579;
	add.f32 	%f581, %f578, %f580;
	mul.rn.f32 	%f87, %f504, %f579;
	neg.f32 	%f583, %f87;
	fma.rn.f32 	%f584, %f504, %f579, %f583;
	fma.rn.f32 	%f585, %f504, %f581, %f584;
	fma.rn.f32 	%f88, %f500, %f579, %f585;
	add.rn.f32 	%f89, %f87, %f88;
	mov.b32 	 %r61, %f89;
	setp.eq.s32	%p1, %r61, 1118925336;
	add.s32 	%r62, %r61, -1;
	mov.b32 	 %f587, %r62;
	selp.f32	%f588, %f587, %f89, %p1;
	mul.f32 	%f589, %f588, 0f3FB8AA3B;
	cvt.rzi.f32.f32	%f590, %f589;
	mov.f32 	%f591, 0fBF317200;
	fma.rn.f32 	%f592, %f590, %f591, %f588;
	mov.f32 	%f593, 0fB5BFBE8E;
	fma.rn.f32 	%f594, %f590, %f593, %f592;
	mul.f32 	%f595, %f594, 0f3FB8AA3B;
	ex2.approx.ftz.f32 	%f596, %f595;
	add.f32 	%f597, %f590, 0f00000000;
	ex2.approx.f32 	%f598, %f597;
	mul.f32 	%f599, %f596, %f598;
	setp.lt.f32	%p17, %f588, 0fC2D20000;
	selp.f32	%f600, 0f00000000, %f599, %p17;
	setp.gt.f32	%p18, %f588, 0f42D20000;
	selp.f32	%f1678, 0f7F800000, %f600, %p18;
	setp.eq.f32	%p19, %f1678, 0f7F800000;
	@%p19 bra 	BB0_7;

	neg.f32 	%f601, %f89;
	add.rn.f32 	%f602, %f87, %f601;
	add.rn.f32 	%f603, %f602, %f88;
	add.f32 	%f604, %f603, 0f37000000;
	selp.f32	%f605, %f604, %f603, %p1;
	fma.rn.f32 	%f1678, %f1678, %f605, %f1678;

BB0_7:
	setp.lt.f32	%p20, %f85, 0f00000000;
	setp.eq.f32	%p21, %f15, 0f3F800000;
	and.pred  	%p2, %p20, %p21;
	mov.b32 	 %r63, %f1678;
	xor.b32  	%r64, %r63, -2147483648;
	mov.b32 	 %f606, %r64;
	selp.f32	%f1680, %f606, %f1678, %p2;
	setp.eq.f32	%p22, %f85, 0f00000000;
	@%p22 bra 	BB0_10;
	bra.uni 	BB0_8;

BB0_10:
	add.f32 	%f609, %f85, %f85;
	selp.f32	%f1680, %f609, 0f00000000, %p21;
	bra.uni 	BB0_11;

BB0_8:
	setp.geu.f32	%p23, %f85, 0f00000000;
	@%p23 bra 	BB0_11;

	cvt.rzi.f32.f32	%f608, %f504;
	setp.neu.f32	%p24, %f608, 0f40800000;
	selp.f32	%f1680, 0f7FFFFFFF, %f1680, %p24;

BB0_11:
	add.f32 	%f610, %f86, 0f40800000;
	mov.b32 	 %r65, %f610;
	setp.lt.s32	%p26, %r65, 2139095040;
	@%p26 bra 	BB0_16;

	setp.gtu.f32	%p27, %f86, 0f7F800000;
	@%p27 bra 	BB0_15;
	bra.uni 	BB0_13;

BB0_15:
	add.f32 	%f1680, %f85, 0f40800000;
	bra.uni 	BB0_16;

BB0_13:
	setp.neu.f32	%p28, %f86, 0f7F800000;
	@%p28 bra 	BB0_16;

	selp.f32	%f1680, 0fFF800000, 0f7F800000, %p2;

BB0_16:
	mul.f32 	%f611, %f81, %f525;
	mov.f32 	%f1734, 0f3F800000;
	sub.f32 	%f613, %f1734, %f1680;
	setp.eq.f32	%p29, %f85, 0f3F800000;
	selp.f32	%f614, 0f00000000, %f613, %p29;
	cvt.sat.f32.f32	%f615, %f614;
	fma.rn.f32 	%f616, %f611, %f611, %f526;
	div.rn.f32 	%f115, %f615, %f616;
	mul.f32 	%f617, %f1645, %f83;
	fma.rn.f32 	%f618, %f1644, %f82, %f617;
	fma.rn.f32 	%f116, %f1646, %f84, %f618;
	setp.eq.f32	%p30, %f528, 0f3F800000;
	@%p30 bra 	BB0_23;
	bra.uni 	BB0_17;

BB0_23:
	setp.leu.f32	%p34, %f523, 0f00000000;
	@%p34 bra 	BB0_19;

	setp.geu.f32	%p35, %f519, 0f00000000;
	@%p35 bra 	BB0_25;

	sub.f32 	%f1643, %f531, %f9;
	sub.f32 	%f1642, %f529, %f7;
	sub.f32 	%f1641, %f530, %f8;
	mul.f32 	%f641, %f509, %f1641;
	fma.rn.f32 	%f642, %f508, %f1642, %f641;
	fma.rn.f32 	%f643, %f510, %f1643, %f642;
	rcp.rn.f32 	%f644, %f643;
	mul.f32 	%f1681, %f1642, %f644;
	mul.f32 	%f1682, %f1641, %f644;
	mul.f32 	%f1683, %f1643, %f644;
	neg.f32 	%f519, %f519;
	bra.uni 	BB0_27;

BB0_17:
	setp.eq.f32	%p31, %f528, 0f40000000;
	@%p31 bra 	BB0_21;
	bra.uni 	BB0_18;

BB0_21:
	setp.leu.f32	%p33, %f523, 0f00000000;
	@%p33 bra 	BB0_19;

	mul.f32 	%f635, %f517, %f83;
	fma.rn.f32 	%f636, %f516, %f82, %f635;
	mul.f32 	%f637, %f513, %f83;
	fma.rn.f32 	%f638, %f512, %f82, %f637;
	mul.f32 	%f639, %f509, %f83;
	fma.rn.f32 	%f640, %f508, %f82, %f639;
	fma.rn.f32 	%f632, %f518, %f84, %f636;
	fma.rn.f32 	%f633, %f514, %f84, %f638;
	fma.rn.f32 	%f634, %f510, %f84, %f640;
	cvt.rzi.s32.f32	%r66, %f523;
	mov.u32 	%r67, 6;
	mov.u32 	%r68, 0;
	// inline asm
	call (%f628, %f629, %f630, %f631), _rt_texture_get_base_id, (%r66, %r67, %f632, %f633, %f634, %r68);
	// inline asm
	mul.f32 	%f1685, %f520, %f628;
	mul.f32 	%f1686, %f1686, %f629;
	mul.f32 	%f1687, %f1687, %f630;
	bra.uni 	BB0_28;

BB0_18:
	setp.neu.f32	%p32, %f528, 0f40800000;
	@%p32 bra 	BB0_19;

	mul.f32 	%f619, %f508, %f82;
	mul.f32 	%f620, %f509, %f83;
	neg.f32 	%f621, %f620;
	sub.f32 	%f622, %f621, %f619;
	mul.f32 	%f623, %f510, %f84;
	sub.f32 	%f624, %f622, %f623;
	fma.rn.f32 	%f625, %f523, %f624, %f519;
	cvt.sat.f32.f32	%f626, %f625;
	mul.f32 	%f627, %f626, %f626;
	mul.f32 	%f1688, %f115, %f627;
	mov.f32 	%f1685, %f520;
	bra.uni 	BB0_29;

BB0_19:
	mov.f32 	%f1685, %f520;
	bra.uni 	BB0_28;

BB0_25:
	mov.f32 	%f1681, %f82;
	mov.f32 	%f1682, %f83;
	mov.f32 	%f1683, %f84;

BB0_27:
	mul.f32 	%f653, %f510, %f84;
	mul.f32 	%f654, %f509, %f83;
	neg.f32 	%f655, %f654;
	mul.f32 	%f656, %f508, %f82;
	sub.f32 	%f657, %f655, %f656;
	sub.f32 	%f658, %f657, %f653;
	setp.gt.f32	%p36, %f658, 0f00000000;
	selp.f32	%f659, 0f3F800000, 0f00000000, %p36;
	mul.f32 	%f660, %f517, %f1682;
	fma.rn.f32 	%f661, %f516, %f1681, %f660;
	mul.f32 	%f662, %f513, %f1682;
	fma.rn.f32 	%f663, %f512, %f1681, %f662;
	fma.rn.f32 	%f664, %f518, %f1683, %f661;
	fma.rn.f32 	%f665, %f514, %f1683, %f663;
	fma.rn.f32 	%f666, %f519, %f664, 0f3F000000;
	mov.f32 	%f667, 0f3F800000;
	sub.f32 	%f649, %f667, %f666;
	fma.rn.f32 	%f650, %f519, %f665, 0f3F000000;
	cvt.rzi.s32.f32	%r69, %f523;
	mov.f32 	%f652, 0f00000000;
	// inline asm
	call (%f645, %f646, %f647, %f648), _rt_texture_get_f_id, (%r69, %r27, %f649, %f650, %f652, %f652);
	// inline asm
	mul.f32 	%f668, %f659, %f645;
	mul.f32 	%f669, %f659, %f646;
	mul.f32 	%f670, %f659, %f647;
	mul.f32 	%f1685, %f520, %f668;
	mul.f32 	%f1686, %f1686, %f669;
	mul.f32 	%f1687, %f1687, %f670;

BB0_28:
	mov.f32 	%f1688, %f115;

BB0_29:
	max.f32 	%f701, %f1685, %f1686;
	max.f32 	%f702, %f701, %f1687;
	mul.f32 	%f703, %f1688, %f702;
	setp.lt.f32	%p38, %f703, 0f3727C5AC;
	mov.pred 	%p147, -1;
	mov.f32 	%f166, 0f00000000;
	mov.f32 	%f1690, %f166;
	mov.f32 	%f1691, %f166;
	mov.f32 	%f1692, %f166;
	mov.f32 	%f1693, %f166;
	mov.f32 	%f1694, %f166;
	mov.f32 	%f1695, %f166;
	mov.f32 	%f1696, %f166;
	mov.f32 	%f1697, %f166;
	mov.f32 	%f1698, %f166;
	mov.f32 	%f1699, %f166;
	mov.f32 	%f1700, %f166;
	mov.f32 	%f1701, %f166;
	mov.f32 	%f1702, %f166;
	mov.f32 	%f1703, %f166;
	mov.f32 	%f1704, %f166;
	mov.f32 	%f1705, %f166;
	mov.f32 	%f1706, %f166;
	mov.f32 	%f1707, %f166;
	mov.f32 	%f1708, %f166;
	mov.f32 	%f1709, %f166;
	mov.f32 	%f1710, %f166;
	mov.f32 	%f1711, %f166;
	mov.f32 	%f1712, %f166;
	mov.f32 	%f1713, %f166;
	mov.f32 	%f1714, %f166;
	mov.f32 	%f1715, %f166;
	mov.f32 	%f1716, %f166;
	mov.f32 	%f1717, %f166;
	mov.f32 	%f1718, %f166;
	@%p38 bra 	BB0_31;

	ld.global.u32 	%r71, [ignoreNormal];
	setp.eq.s32	%p40, %r71, 0;
	selp.f32	%f704, %f116, 0f3F800000, %p40;
	cvt.sat.f32.f32	%f705, %f704;
	mul.f32 	%f706, %f1688, %f705;
	mul.f32 	%f166, %f1685, %f706;
	mul.f32 	%f1690, %f1686, %f706;
	mul.f32 	%f1691, %f1687, %f706;
	mul.f32 	%f707, %f1685, %f1688;
	add.f32 	%f1692, %f707, 0f00000000;
	mul.f32 	%f708, %f1686, %f1688;
	add.f32 	%f1693, %f708, 0f00000000;
	mul.f32 	%f709, %f1687, %f1688;
	add.f32 	%f1694, %f709, 0f00000000;
	fma.rn.f32 	%f1695, %f83, %f707, 0f00000000;
	fma.rn.f32 	%f1696, %f83, %f708, 0f00000000;
	fma.rn.f32 	%f1697, %f83, %f709, 0f00000000;
	fma.rn.f32 	%f1698, %f84, %f707, 0f00000000;
	fma.rn.f32 	%f1699, %f84, %f708, 0f00000000;
	fma.rn.f32 	%f1700, %f84, %f709, 0f00000000;
	fma.rn.f32 	%f1701, %f82, %f707, 0f00000000;
	fma.rn.f32 	%f1702, %f82, %f708, 0f00000000;
	fma.rn.f32 	%f1703, %f82, %f709, 0f00000000;
	mul.f32 	%f710, %f82, %f83;
	fma.rn.f32 	%f1704, %f710, %f707, 0f00000000;
	fma.rn.f32 	%f1705, %f710, %f708, 0f00000000;
	fma.rn.f32 	%f1706, %f710, %f709, 0f00000000;
	mul.f32 	%f711, %f83, %f84;
	fma.rn.f32 	%f1707, %f711, %f707, 0f00000000;
	fma.rn.f32 	%f1708, %f711, %f708, 0f00000000;
	fma.rn.f32 	%f1709, %f711, %f709, 0f00000000;
	fma.rn.f32 	%f712, %f84, %f84, 0fBEAAAAAB;
	fma.rn.f32 	%f1710, %f712, %f707, 0f00000000;
	fma.rn.f32 	%f1711, %f712, %f708, 0f00000000;
	fma.rn.f32 	%f1712, %f712, %f709, 0f00000000;
	mul.f32 	%f713, %f82, %f84;
	fma.rn.f32 	%f1713, %f713, %f707, 0f00000000;
	fma.rn.f32 	%f1714, %f713, %f708, 0f00000000;
	fma.rn.f32 	%f1715, %f713, %f709, 0f00000000;
	mul.f32 	%f714, %f83, %f83;
	mul.f32 	%f715, %f82, %f82;
	sub.f32 	%f716, %f715, %f714;
	fma.rn.f32 	%f1716, %f716, %f707, 0f00000000;
	fma.rn.f32 	%f1717, %f716, %f708, 0f00000000;
	fma.rn.f32 	%f1718, %f716, %f709, 0f00000000;
	mov.pred 	%p147, 0;

BB0_31:
	@%p147 bra 	BB0_46;

	setp.eq.s32	%p41, %r6, 0;
	mov.u16 	%rs301, 0;
	mov.f32 	%f1735, %f1734;
	mov.f32 	%f1736, %f1734;
	@%p41 bra 	BB0_43;

	abs.s32 	%r8, %r6;
	mov.f32 	%f1731, 0f00000000;
	setp.lt.s32	%p42, %r8, 1;
	mov.f32 	%f1732, %f1731;
	mov.f32 	%f1733, %f1731;
	@%p42 bra 	BB0_42;

	max.f32 	%f197, %f730, %f466;
	and.b32  	%r73, %r8, 3;
	setp.eq.s32	%p43, %r73, 0;
	mov.f32 	%f1731, 0f00000000;
	mov.u32 	%r322, 0;
	mov.f32 	%f1732, %f1731;
	mov.f32 	%f1733, %f1731;
	@%p43 bra 	BB0_40;

	setp.eq.s32	%p44, %r73, 1;
	mov.f32 	%f1722, 0f00000000;
	mov.u32 	%r320, 0;
	mov.f32 	%f1723, %f1722;
	mov.f32 	%f1724, %f1722;
	@%p44 bra 	BB0_39;

	setp.eq.s32	%p45, %r73, 2;
	mov.f32 	%f1719, 0f00000000;
	mov.u32 	%r319, 0;
	mov.f32 	%f1720, %f1719;
	mov.f32 	%f1721, %f1719;
	@%p45 bra 	BB0_38;

	cvt.rmi.f32.f32	%f746, %f19;
	sub.f32 	%f747, %f19, %f746;
	cvt.rmi.f32.f32	%f748, %f20;
	sub.f32 	%f749, %f20, %f748;
	add.f32 	%f750, %f749, 0f420551EC;
	add.f32 	%f751, %f747, 0f420551EC;
	mul.f32 	%f752, %f747, %f750;
	fma.rn.f32 	%f753, %f749, %f751, %f752;
	add.f32 	%f754, %f753, 0f00000000;
	add.f32 	%f755, %f747, %f754;
	add.f32 	%f756, %f749, %f754;
	add.f32 	%f757, %f755, %f756;
	mul.f32 	%f758, %f754, %f757;
	cvt.rmi.f32.f32	%f759, %f758;
	sub.f32 	%f760, %f758, %f759;
	add.f32 	%f761, %f755, %f755;
	mul.f32 	%f762, %f756, %f761;
	cvt.rmi.f32.f32	%f763, %f762;
	sub.f32 	%f764, %f762, %f763;
	mul.f32 	%f765, %f755, %f757;
	cvt.rmi.f32.f32	%f766, %f765;
	sub.f32 	%f767, %f765, %f766;
	fma.rn.f32 	%f768, %f760, 0f40000000, 0fBF800000;
	fma.rn.f32 	%f769, %f764, 0f40000000, 0fBF800000;
	fma.rn.f32 	%f770, %f767, 0f40000000, 0fBF800000;
	fma.rn.f32 	%f771, %f524, %f768, %f529;
	fma.rn.f32 	%f772, %f524, %f769, %f530;
	fma.rn.f32 	%f773, %f524, %f770, %f531;
	sub.f32 	%f774, %f771, %f7;
	sub.f32 	%f775, %f772, %f8;
	sub.f32 	%f776, %f773, %f9;
	mul.f32 	%f777, %f775, %f775;
	fma.rn.f32 	%f778, %f774, %f774, %f777;
	fma.rn.f32 	%f779, %f776, %f776, %f778;
	sqrt.rn.f32 	%f745, %f779;
	rcp.rn.f32 	%f780, %f745;
	mul.f32 	%f741, %f780, %f774;
	mul.f32 	%f742, %f780, %f775;
	mul.f32 	%f743, %f780, %f776;
	ld.global.u32 	%r82, [imageEnabled];
	and.b32  	%r83, %r82, 32;
	setp.eq.s32	%p46, %r83, 0;
	selp.f32	%f781, 0f3F800000, 0f41200000, %p46;
	mul.f32 	%f744, %f781, %f197;
	add.u64 	%rd34, %SP, 0;
	cvta.to.local.u64 	%rd35, %rd34;
	mov.u32 	%r84, 1065353216;
	st.local.u32 	[%rd35+8], %r84;
	st.local.u32 	[%rd35+4], %r84;
	st.local.u32 	[%rd35], %r84;
	ld.global.u32 	%r78, [root];
	// inline asm
	call _rt_trace_64, (%r78, %f10, %f11, %f12, %f741, %f742, %f743, %r44, %f744, %f745, %rd34, %r43);
	// inline asm
	ld.local.f32 	%f782, [%rd35];
	add.f32 	%f1721, %f782, 0f00000000;
	ld.local.f32 	%f783, [%rd35+4];
	add.f32 	%f1720, %f783, 0f00000000;
	ld.local.f32 	%f784, [%rd35+8];
	add.f32 	%f1719, %f784, 0f00000000;
	mov.u32 	%r319, %r44;

BB0_38:
	cvt.rn.f32.s32	%f793, %r319;
	add.f32 	%f794, %f13, %f793;
	sub.f32 	%f795, %f14, %f793;
	mul.f32 	%f796, %f794, 0f3DD32618;
	cvt.rmi.f32.f32	%f797, %f796;
	sub.f32 	%f798, %f796, %f797;
	mul.f32 	%f799, %f795, 0f3DD2F1AA;
	cvt.rmi.f32.f32	%f800, %f799;
	sub.f32 	%f801, %f799, %f800;
	mul.f32 	%f802, %f793, 0f3DC74539;
	cvt.rmi.f32.f32	%f803, %f802;
	sub.f32 	%f804, %f802, %f803;
	add.f32 	%f805, %f801, 0f420551EC;
	add.f32 	%f806, %f798, 0f420551EC;
	add.f32 	%f807, %f804, 0f420551EC;
	mul.f32 	%f808, %f798, %f805;
	fma.rn.f32 	%f809, %f801, %f806, %f808;
	fma.rn.f32 	%f810, %f804, %f807, %f809;
	add.f32 	%f811, %f798, %f810;
	add.f32 	%f812, %f801, %f810;
	add.f32 	%f813, %f804, %f810;
	add.f32 	%f814, %f811, %f812;
	mul.f32 	%f815, %f813, %f814;
	cvt.rmi.f32.f32	%f816, %f815;
	sub.f32 	%f817, %f815, %f816;
	add.f32 	%f818, %f811, %f811;
	mul.f32 	%f819, %f812, %f818;
	cvt.rmi.f32.f32	%f820, %f819;
	sub.f32 	%f821, %f819, %f820;
	mul.f32 	%f822, %f811, %f814;
	cvt.rmi.f32.f32	%f823, %f822;
	sub.f32 	%f824, %f822, %f823;
	fma.rn.f32 	%f825, %f817, 0f40000000, 0fBF800000;
	fma.rn.f32 	%f826, %f821, 0f40000000, 0fBF800000;
	fma.rn.f32 	%f827, %f824, 0f40000000, 0fBF800000;
	fma.rn.f32 	%f828, %f524, %f825, %f529;
	fma.rn.f32 	%f829, %f524, %f826, %f530;
	fma.rn.f32 	%f830, %f524, %f827, %f531;
	sub.f32 	%f831, %f828, %f7;
	sub.f32 	%f832, %f829, %f8;
	sub.f32 	%f833, %f830, %f9;
	mul.f32 	%f834, %f832, %f832;
	fma.rn.f32 	%f835, %f831, %f831, %f834;
	fma.rn.f32 	%f836, %f833, %f833, %f835;
	sqrt.rn.f32 	%f792, %f836;
	rcp.rn.f32 	%f837, %f792;
	mul.f32 	%f788, %f837, %f831;
	mul.f32 	%f789, %f837, %f832;
	mul.f32 	%f790, %f837, %f833;
	ld.global.u32 	%r88, [imageEnabled];
	and.b32  	%r89, %r88, 32;
	setp.eq.s32	%p47, %r89, 0;
	selp.f32	%f838, 0f3F800000, 0f41200000, %p47;
	mul.f32 	%f791, %f838, %f197;
	add.u64 	%rd36, %SP, 0;
	cvta.to.local.u64 	%rd37, %rd36;
	mov.u32 	%r90, 1065353216;
	st.local.u32 	[%rd37+8], %r90;
	st.local.u32 	[%rd37+4], %r90;
	st.local.u32 	[%rd37], %r90;
	ld.global.u32 	%r85, [root];
	// inline asm
	call _rt_trace_64, (%r85, %f10, %f11, %f12, %f788, %f789, %f790, %r44, %f791, %f792, %rd36, %r43);
	// inline asm
	ld.local.f32 	%f839, [%rd37];
	add.f32 	%f1724, %f1721, %f839;
	ld.local.f32 	%f840, [%rd37+4];
	add.f32 	%f1723, %f1720, %f840;
	ld.local.f32 	%f841, [%rd37+8];
	add.f32 	%f1722, %f1719, %f841;
	add.s32 	%r320, %r319, 1;

BB0_39:
	cvt.rn.f32.s32	%f850, %r320;
	add.f32 	%f851, %f13, %f850;
	sub.f32 	%f852, %f14, %f850;
	mul.f32 	%f853, %f851, 0f3DD32618;
	cvt.rmi.f32.f32	%f854, %f853;
	sub.f32 	%f855, %f853, %f854;
	mul.f32 	%f856, %f852, 0f3DD2F1AA;
	cvt.rmi.f32.f32	%f857, %f856;
	sub.f32 	%f858, %f856, %f857;
	mul.f32 	%f859, %f850, 0f3DC74539;
	cvt.rmi.f32.f32	%f860, %f859;
	sub.f32 	%f861, %f859, %f860;
	add.f32 	%f862, %f858, 0f420551EC;
	add.f32 	%f863, %f855, 0f420551EC;
	add.f32 	%f864, %f861, 0f420551EC;
	mul.f32 	%f865, %f855, %f862;
	fma.rn.f32 	%f866, %f858, %f863, %f865;
	fma.rn.f32 	%f867, %f861, %f864, %f866;
	add.f32 	%f868, %f855, %f867;
	add.f32 	%f869, %f858, %f867;
	add.f32 	%f870, %f861, %f867;
	add.f32 	%f871, %f868, %f869;
	mul.f32 	%f872, %f870, %f871;
	cvt.rmi.f32.f32	%f873, %f872;
	sub.f32 	%f874, %f872, %f873;
	add.f32 	%f875, %f868, %f868;
	mul.f32 	%f876, %f869, %f875;
	cvt.rmi.f32.f32	%f877, %f876;
	sub.f32 	%f878, %f876, %f877;
	mul.f32 	%f879, %f868, %f871;
	cvt.rmi.f32.f32	%f880, %f879;
	sub.f32 	%f881, %f879, %f880;
	fma.rn.f32 	%f882, %f874, 0f40000000, 0fBF800000;
	fma.rn.f32 	%f883, %f878, 0f40000000, 0fBF800000;
	fma.rn.f32 	%f884, %f881, 0f40000000, 0fBF800000;
	fma.rn.f32 	%f885, %f524, %f882, %f529;
	fma.rn.f32 	%f886, %f524, %f883, %f530;
	fma.rn.f32 	%f887, %f524, %f884, %f531;
	sub.f32 	%f888, %f885, %f7;
	sub.f32 	%f889, %f886, %f8;
	sub.f32 	%f890, %f887, %f9;
	mul.f32 	%f891, %f889, %f889;
	fma.rn.f32 	%f892, %f888, %f888, %f891;
	fma.rn.f32 	%f893, %f890, %f890, %f892;
	sqrt.rn.f32 	%f849, %f893;
	rcp.rn.f32 	%f894, %f849;
	mul.f32 	%f845, %f894, %f888;
	mul.f32 	%f846, %f894, %f889;
	mul.f32 	%f847, %f894, %f890;
	ld.global.u32 	%r94, [imageEnabled];
	and.b32  	%r95, %r94, 32;
	setp.eq.s32	%p48, %r95, 0;
	selp.f32	%f895, 0f3F800000, 0f41200000, %p48;
	mul.f32 	%f848, %f895, %f197;
	add.u64 	%rd38, %SP, 0;
	cvta.to.local.u64 	%rd39, %rd38;
	mov.u32 	%r96, 1065353216;
	st.local.u32 	[%rd39+8], %r96;
	st.local.u32 	[%rd39+4], %r96;
	st.local.u32 	[%rd39], %r96;
	ld.global.u32 	%r91, [root];
	mov.u32 	%r92, 1;
	// inline asm
	call _rt_trace_64, (%r91, %f10, %f11, %f12, %f845, %f846, %f847, %r92, %f848, %f849, %rd38, %r43);
	// inline asm
	ld.local.f32 	%f896, [%rd39];
	add.f32 	%f1733, %f1724, %f896;
	ld.local.f32 	%f897, [%rd39+4];
	add.f32 	%f1732, %f1723, %f897;
	ld.local.f32 	%f898, [%rd39+8];
	add.f32 	%f1731, %f1722, %f898;
	add.s32 	%r322, %r320, 1;

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

BB0_41:
	cvt.rn.f32.s32	%f931, %r322;
	add.f32 	%f932, %f13, %f931;
	sub.f32 	%f933, %f14, %f931;
	mul.f32 	%f934, %f932, 0f3DD32618;
	cvt.rmi.f32.f32	%f935, %f934;
	sub.f32 	%f936, %f934, %f935;
	mul.f32 	%f937, %f933, 0f3DD2F1AA;
	cvt.rmi.f32.f32	%f938, %f937;
	sub.f32 	%f939, %f937, %f938;
	mul.f32 	%f940, %f931, 0f3DC74539;
	cvt.rmi.f32.f32	%f941, %f940;
	sub.f32 	%f942, %f940, %f941;
	add.f32 	%f943, %f939, 0f420551EC;
	add.f32 	%f944, %f936, 0f420551EC;
	add.f32 	%f945, %f942, 0f420551EC;
	mul.f32 	%f946, %f936, %f943;
	fma.rn.f32 	%f947, %f939, %f944, %f946;
	fma.rn.f32 	%f948, %f942, %f945, %f947;
	add.f32 	%f949, %f936, %f948;
	add.f32 	%f950, %f939, %f948;
	add.f32 	%f951, %f942, %f948;
	add.f32 	%f952, %f949, %f950;
	mul.f32 	%f953, %f951, %f952;
	cvt.rmi.f32.f32	%f954, %f953;
	sub.f32 	%f955, %f953, %f954;
	add.f32 	%f956, %f949, %f949;
	mul.f32 	%f957, %f950, %f956;
	cvt.rmi.f32.f32	%f958, %f957;
	sub.f32 	%f959, %f957, %f958;
	mul.f32 	%f960, %f949, %f952;
	cvt.rmi.f32.f32	%f961, %f960;
	sub.f32 	%f962, %f960, %f961;
	fma.rn.f32 	%f963, %f955, 0f40000000, 0fBF800000;
	fma.rn.f32 	%f964, %f959, 0f40000000, 0fBF800000;
	fma.rn.f32 	%f965, %f962, 0f40000000, 0fBF800000;
	fma.rn.f32 	%f966, %f524, %f963, %f529;
	fma.rn.f32 	%f967, %f524, %f964, %f530;
	fma.rn.f32 	%f968, %f524, %f965, %f531;
	sub.f32 	%f969, %f966, %f7;
	sub.f32 	%f970, %f967, %f8;
	sub.f32 	%f971, %f968, %f9;
	mul.f32 	%f972, %f970, %f970;
	fma.rn.f32 	%f973, %f969, %f969, %f972;
	fma.rn.f32 	%f974, %f971, %f971, %f973;
	sqrt.rn.f32 	%f906, %f974;
	rcp.rn.f32 	%f975, %f906;
	mul.f32 	%f902, %f975, %f969;
	mul.f32 	%f903, %f975, %f970;
	mul.f32 	%f904, %f975, %f971;
	ld.global.u32 	%r109, [imageEnabled];
	and.b32  	%r110, %r109, 32;
	setp.eq.s32	%p50, %r110, 0;
	selp.f32	%f976, 0f3F800000, 0f41200000, %p50;
	mul.f32 	%f905, %f976, %f197;
	add.u64 	%rd43, %SP, 0;
	cvta.to.local.u64 	%rd44, %rd43;
	mov.u32 	%r111, 1065353216;
	st.local.u32 	[%rd44+8], %r111;
	st.local.u32 	[%rd44+4], %r111;
	st.local.u32 	[%rd44], %r111;
	ld.global.u32 	%r97, [root];
	mov.u32 	%r107, 1;
	// inline asm
	call _rt_trace_64, (%r97, %f10, %f11, %f12, %f902, %f903, %f904, %r107, %f905, %f906, %rd43, %r43);
	// inline asm
	ld.local.f32 	%f977, [%rd44];
	add.f32 	%f978, %f1733, %f977;
	ld.local.f32 	%f979, [%rd44+4];
	add.f32 	%f980, %f1732, %f979;
	ld.local.f32 	%f981, [%rd44+8];
	add.f32 	%f982, %f1731, %f981;
	add.s32 	%r112, %r322, 1;
	cvt.rn.f32.s32	%f983, %r112;
	add.f32 	%f984, %f13, %f983;
	sub.f32 	%f985, %f14, %f983;
	mul.f32 	%f986, %f984, 0f3DD32618;
	cvt.rmi.f32.f32	%f987, %f986;
	sub.f32 	%f988, %f986, %f987;
	mul.f32 	%f989, %f985, 0f3DD2F1AA;
	cvt.rmi.f32.f32	%f990, %f989;
	sub.f32 	%f991, %f989, %f990;
	mul.f32 	%f992, %f983, 0f3DC74539;
	cvt.rmi.f32.f32	%f993, %f992;
	sub.f32 	%f994, %f992, %f993;
	add.f32 	%f995, %f991, 0f420551EC;
	add.f32 	%f996, %f988, 0f420551EC;
	add.f32 	%f997, %f994, 0f420551EC;
	mul.f32 	%f998, %f988, %f995;
	fma.rn.f32 	%f999, %f991, %f996, %f998;
	fma.rn.f32 	%f1000, %f994, %f997, %f999;
	add.f32 	%f1001, %f988, %f1000;
	add.f32 	%f1002, %f991, %f1000;
	add.f32 	%f1003, %f994, %f1000;
	add.f32 	%f1004, %f1001, %f1002;
	mul.f32 	%f1005, %f1003, %f1004;
	cvt.rmi.f32.f32	%f1006, %f1005;
	sub.f32 	%f1007, %f1005, %f1006;
	add.f32 	%f1008, %f1001, %f1001;
	mul.f32 	%f1009, %f1002, %f1008;
	cvt.rmi.f32.f32	%f1010, %f1009;
	sub.f32 	%f1011, %f1009, %f1010;
	mul.f32 	%f1012, %f1001, %f1004;
	cvt.rmi.f32.f32	%f1013, %f1012;
	sub.f32 	%f1014, %f1012, %f1013;
	fma.rn.f32 	%f1015, %f1007, 0f40000000, 0fBF800000;
	fma.rn.f32 	%f1016, %f1011, 0f40000000, 0fBF800000;
	fma.rn.f32 	%f1017, %f1014, 0f40000000, 0fBF800000;
	fma.rn.f32 	%f1018, %f524, %f1015, %f529;
	fma.rn.f32 	%f1019, %f524, %f1016, %f530;
	fma.rn.f32 	%f1020, %f524, %f1017, %f531;
	sub.f32 	%f1021, %f1018, %f7;
	sub.f32 	%f1022, %f1019, %f8;
	sub.f32 	%f1023, %f1020, %f9;
	mul.f32 	%f1024, %f1022, %f1022;
	fma.rn.f32 	%f1025, %f1021, %f1021, %f1024;
	fma.rn.f32 	%f1026, %f1023, %f1023, %f1025;
	sqrt.rn.f32 	%f914, %f1026;
	rcp.rn.f32 	%f1027, %f914;
	mul.f32 	%f910, %f1027, %f1021;
	mul.f32 	%f911, %f1027, %f1022;
	mul.f32 	%f912, %f1027, %f1023;
	ld.global.u32 	%r113, [imageEnabled];
	and.b32  	%r114, %r113, 32;
	setp.eq.s32	%p51, %r114, 0;
	selp.f32	%f1028, 0f3F800000, 0f41200000, %p51;
	mul.f32 	%f913, %f1028, %f197;
	st.local.u32 	[%rd44+8], %r111;
	st.local.u32 	[%rd44+4], %r111;
	st.local.u32 	[%rd44], %r111;
	ld.global.u32 	%r100, [root];
	// inline asm
	call _rt_trace_64, (%r100, %f10, %f11, %f12, %f910, %f911, %f912, %r107, %f913, %f914, %rd43, %r43);
	// inline asm
	ld.local.f32 	%f1029, [%rd44];
	add.f32 	%f1030, %f978, %f1029;
	ld.local.f32 	%f1031, [%rd44+4];
	add.f32 	%f1032, %f980, %f1031;
	ld.local.f32 	%f1033, [%rd44+8];
	add.f32 	%f1034, %f982, %f1033;
	add.s32 	%r115, %r322, 2;
	cvt.rn.f32.s32	%f1035, %r115;
	add.f32 	%f1036, %f13, %f1035;
	sub.f32 	%f1037, %f14, %f1035;
	mul.f32 	%f1038, %f1036, 0f3DD32618;
	cvt.rmi.f32.f32	%f1039, %f1038;
	sub.f32 	%f1040, %f1038, %f1039;
	mul.f32 	%f1041, %f1037, 0f3DD2F1AA;
	cvt.rmi.f32.f32	%f1042, %f1041;
	sub.f32 	%f1043, %f1041, %f1042;
	mul.f32 	%f1044, %f1035, 0f3DC74539;
	cvt.rmi.f32.f32	%f1045, %f1044;
	sub.f32 	%f1046, %f1044, %f1045;
	add.f32 	%f1047, %f1043, 0f420551EC;
	add.f32 	%f1048, %f1040, 0f420551EC;
	add.f32 	%f1049, %f1046, 0f420551EC;
	mul.f32 	%f1050, %f1040, %f1047;
	fma.rn.f32 	%f1051, %f1043, %f1048, %f1050;
	fma.rn.f32 	%f1052, %f1046, %f1049, %f1051;
	add.f32 	%f1053, %f1040, %f1052;
	add.f32 	%f1054, %f1043, %f1052;
	add.f32 	%f1055, %f1046, %f1052;
	add.f32 	%f1056, %f1053, %f1054;
	mul.f32 	%f1057, %f1055, %f1056;
	cvt.rmi.f32.f32	%f1058, %f1057;
	sub.f32 	%f1059, %f1057, %f1058;
	add.f32 	%f1060, %f1053, %f1053;
	mul.f32 	%f1061, %f1054, %f1060;
	cvt.rmi.f32.f32	%f1062, %f1061;
	sub.f32 	%f1063, %f1061, %f1062;
	mul.f32 	%f1064, %f1053, %f1056;
	cvt.rmi.f32.f32	%f1065, %f1064;
	sub.f32 	%f1066, %f1064, %f1065;
	fma.rn.f32 	%f1067, %f1059, 0f40000000, 0fBF800000;
	fma.rn.f32 	%f1068, %f1063, 0f40000000, 0fBF800000;
	fma.rn.f32 	%f1069, %f1066, 0f40000000, 0fBF800000;
	fma.rn.f32 	%f1070, %f524, %f1067, %f529;
	fma.rn.f32 	%f1071, %f524, %f1068, %f530;
	fma.rn.f32 	%f1072, %f524, %f1069, %f531;
	sub.f32 	%f1073, %f1070, %f7;
	sub.f32 	%f1074, %f1071, %f8;
	sub.f32 	%f1075, %f1072, %f9;
	mul.f32 	%f1076, %f1074, %f1074;
	fma.rn.f32 	%f1077, %f1073, %f1073, %f1076;
	fma.rn.f32 	%f1078, %f1075, %f1075, %f1077;
	sqrt.rn.f32 	%f922, %f1078;
	rcp.rn.f32 	%f1079, %f922;
	mul.f32 	%f918, %f1079, %f1073;
	mul.f32 	%f919, %f1079, %f1074;
	mul.f32 	%f920, %f1079, %f1075;
	ld.global.u32 	%r116, [imageEnabled];
	and.b32  	%r117, %r116, 32;
	setp.eq.s32	%p52, %r117, 0;
	selp.f32	%f1080, 0f3F800000, 0f41200000, %p52;
	mul.f32 	%f921, %f1080, %f197;
	st.local.u32 	[%rd44+8], %r111;
	st.local.u32 	[%rd44+4], %r111;
	st.local.u32 	[%rd44], %r111;
	ld.global.u32 	%r103, [root];
	// inline asm
	call _rt_trace_64, (%r103, %f10, %f11, %f12, %f918, %f919, %f920, %r107, %f921, %f922, %rd43, %r43);
	// inline asm
	ld.local.f32 	%f1081, [%rd44];
	add.f32 	%f1082, %f1030, %f1081;
	ld.local.f32 	%f1083, [%rd44+4];
	add.f32 	%f1084, %f1032, %f1083;
	ld.local.f32 	%f1085, [%rd44+8];
	add.f32 	%f1086, %f1034, %f1085;
	add.s32 	%r118, %r322, 3;
	cvt.rn.f32.s32	%f1087, %r118;
	add.f32 	%f1088, %f13, %f1087;
	sub.f32 	%f1089, %f14, %f1087;
	mul.f32 	%f1090, %f1088, 0f3DD32618;
	cvt.rmi.f32.f32	%f1091, %f1090;
	sub.f32 	%f1092, %f1090, %f1091;
	mul.f32 	%f1093, %f1089, 0f3DD2F1AA;
	cvt.rmi.f32.f32	%f1094, %f1093;
	sub.f32 	%f1095, %f1093, %f1094;
	mul.f32 	%f1096, %f1087, 0f3DC74539;
	cvt.rmi.f32.f32	%f1097, %f1096;
	sub.f32 	%f1098, %f1096, %f1097;
	add.f32 	%f1099, %f1095, 0f420551EC;
	add.f32 	%f1100, %f1092, 0f420551EC;
	add.f32 	%f1101, %f1098, 0f420551EC;
	mul.f32 	%f1102, %f1092, %f1099;
	fma.rn.f32 	%f1103, %f1095, %f1100, %f1102;
	fma.rn.f32 	%f1104, %f1098, %f1101, %f1103;
	add.f32 	%f1105, %f1092, %f1104;
	add.f32 	%f1106, %f1095, %f1104;
	add.f32 	%f1107, %f1098, %f1104;
	add.f32 	%f1108, %f1105, %f1106;
	mul.f32 	%f1109, %f1107, %f1108;
	cvt.rmi.f32.f32	%f1110, %f1109;
	sub.f32 	%f1111, %f1109, %f1110;
	add.f32 	%f1112, %f1105, %f1105;
	mul.f32 	%f1113, %f1106, %f1112;
	cvt.rmi.f32.f32	%f1114, %f1113;
	sub.f32 	%f1115, %f1113, %f1114;
	mul.f32 	%f1116, %f1105, %f1108;
	cvt.rmi.f32.f32	%f1117, %f1116;
	sub.f32 	%f1118, %f1116, %f1117;
	fma.rn.f32 	%f1119, %f1111, 0f40000000, 0fBF800000;
	fma.rn.f32 	%f1120, %f1115, 0f40000000, 0fBF800000;
	fma.rn.f32 	%f1121, %f1118, 0f40000000, 0fBF800000;
	fma.rn.f32 	%f1122, %f524, %f1119, %f529;
	fma.rn.f32 	%f1123, %f524, %f1120, %f530;
	fma.rn.f32 	%f1124, %f524, %f1121, %f531;
	sub.f32 	%f1125, %f1122, %f7;
	sub.f32 	%f1126, %f1123, %f8;
	sub.f32 	%f1127, %f1124, %f9;
	mul.f32 	%f1128, %f1126, %f1126;
	fma.rn.f32 	%f1129, %f1125, %f1125, %f1128;
	fma.rn.f32 	%f1130, %f1127, %f1127, %f1129;
	sqrt.rn.f32 	%f930, %f1130;
	rcp.rn.f32 	%f1131, %f930;
	mul.f32 	%f926, %f1131, %f1125;
	mul.f32 	%f927, %f1131, %f1126;
	mul.f32 	%f928, %f1131, %f1127;
	ld.global.u32 	%r119, [imageEnabled];
	and.b32  	%r120, %r119, 32;
	setp.eq.s32	%p53, %r120, 0;
	selp.f32	%f1132, 0f3F800000, 0f41200000, %p53;
	mul.f32 	%f929, %f1132, %f197;
	st.local.u32 	[%rd44+8], %r111;
	st.local.u32 	[%rd44+4], %r111;
	st.local.u32 	[%rd44], %r111;
	ld.global.u32 	%r106, [root];
	// inline asm
	call _rt_trace_64, (%r106, %f10, %f11, %f12, %f926, %f927, %f928, %r107, %f929, %f930, %rd43, %r43);
	// inline asm
	ld.local.f32 	%f1133, [%rd44];
	add.f32 	%f1733, %f1082, %f1133;
	ld.local.f32 	%f1134, [%rd44+4];
	add.f32 	%f1732, %f1084, %f1134;
	ld.local.f32 	%f1135, [%rd44+8];
	add.f32 	%f1731, %f1086, %f1135;
	add.s32 	%r322, %r322, 4;
	setp.lt.s32	%p54, %r322, %r8;
	@%p54 bra 	BB0_41;

BB0_42:
	cvt.rn.f32.s32	%f1136, %r8;
	rcp.rn.f32 	%f1137, %f1136;
	mul.f32 	%f1734, %f1733, %f1137;
	mul.f32 	%f1735, %f1732, %f1137;
	mul.f32 	%f1736, %f1731, %f1137;
	shr.u32 	%r121, %r6, 31;
	cvt.u16.u32	%rs301, %r121;

BB0_43:
	fma.rn.f32 	%f1677, %f166, %f1734, %f1677;
	fma.rn.f32 	%f1676, %f1690, %f1735, %f1676;
	fma.rn.f32 	%f1675, %f1691, %f1736, %f1675;
	fma.rn.f32 	%f1674, %f1692, %f1734, %f1674;
	fma.rn.f32 	%f1673, %f1693, %f1735, %f1673;
	fma.rn.f32 	%f1672, %f1694, %f1736, %f1672;
	fma.rn.f32 	%f1671, %f1695, %f1734, %f1671;
	fma.rn.f32 	%f1670, %f1696, %f1735, %f1670;
	fma.rn.f32 	%f1669, %f1697, %f1736, %f1669;
	fma.rn.f32 	%f1668, %f1698, %f1734, %f1668;
	fma.rn.f32 	%f1667, %f1699, %f1735, %f1667;
	fma.rn.f32 	%f1666, %f1700, %f1736, %f1666;
	fma.rn.f32 	%f1665, %f1701, %f1734, %f1665;
	fma.rn.f32 	%f1664, %f1702, %f1735, %f1664;
	fma.rn.f32 	%f1663, %f1703, %f1736, %f1663;
	fma.rn.f32 	%f1662, %f1704, %f1734, %f1662;
	fma.rn.f32 	%f1661, %f1705, %f1735, %f1661;
	fma.rn.f32 	%f1660, %f1706, %f1736, %f1660;
	fma.rn.f32 	%f1659, %f1707, %f1734, %f1659;
	fma.rn.f32 	%f1658, %f1708, %f1735, %f1658;
	fma.rn.f32 	%f1657, %f1709, %f1736, %f1657;
	fma.rn.f32 	%f1656, %f1710, %f1734, %f1656;
	fma.rn.f32 	%f1655, %f1711, %f1735, %f1655;
	fma.rn.f32 	%f1654, %f1712, %f1736, %f1654;
	fma.rn.f32 	%f1653, %f1713, %f1734, %f1653;
	fma.rn.f32 	%f1652, %f1714, %f1735, %f1652;
	fma.rn.f32 	%f1651, %f1715, %f1736, %f1651;
	fma.rn.f32 	%f1650, %f1716, %f1734, %f1650;
	fma.rn.f32 	%f1649, %f1717, %f1735, %f1649;
	fma.rn.f32 	%f1648, %f1718, %f1736, %f1648;
	setp.eq.s16	%p55, %rs301, 0;
	@%p55 bra 	BB0_45;

	div.rn.f32 	%f1138, %f166, %f520;
	div.rn.f32 	%f1139, %f1138, %f115;
	cvt.sat.f32.f32	%f1140, %f1139;
	mul.f32 	%f1734, %f1734, %f1140;

BB0_45:
	add.f32 	%f1647, %f1647, %f1734;

BB0_46:
	add.s32 	%r318, %r318, 1;
	setp.lt.u32	%p56, %r318, %r4;
	@%p56 bra 	BB0_5;

BB0_47:
	ld.global.u32 	%r324, [imageEnabled];
	and.b32  	%r122, %r324, 8;
	setp.eq.s32	%p57, %r122, 0;
	@%p57 bra 	BB0_60;

	cvt.sat.f32.f32	%f326, %f1647;
	cvt.u64.u32	%rd48, %r3;
	cvt.u64.u32	%rd47, %r2;
	mov.u64 	%rd51, image_Mask;
	cvta.global.u64 	%rd46, %rd51;
	// inline asm
	call (%rd45), _rt_buffer_get_64, (%rd46, %r27, %r27, %rd47, %rd48, %rd12, %rd12);
	// inline asm
	mov.f32 	%f1143, 0f3E68BA2E;
	cvt.rzi.f32.f32	%f1144, %f1143;
	fma.rn.f32 	%f1145, %f1144, 0fC0000000, 0f3EE8BA2E;
	abs.f32 	%f327, %f1145;
	abs.f32 	%f328, %f326;
	setp.lt.f32	%p58, %f328, 0f00800000;
	mul.f32 	%f1146, %f328, 0f4B800000;
	selp.f32	%f1147, 0fC3170000, 0fC2FE0000, %p58;
	selp.f32	%f1148, %f1146, %f328, %p58;
	mov.b32 	 %r125, %f1148;
	and.b32  	%r126, %r125, 8388607;
	or.b32  	%r127, %r126, 1065353216;
	mov.b32 	 %f1149, %r127;
	shr.u32 	%r128, %r125, 23;
	cvt.rn.f32.u32	%f1150, %r128;
	add.f32 	%f1151, %f1147, %f1150;
	setp.gt.f32	%p59, %f1149, 0f3FB504F3;
	mul.f32 	%f1152, %f1149, 0f3F000000;
	add.f32 	%f1153, %f1151, 0f3F800000;
	selp.f32	%f1154, %f1152, %f1149, %p59;
	selp.f32	%f1155, %f1153, %f1151, %p59;
	add.f32 	%f1156, %f1154, 0fBF800000;
	add.f32 	%f1142, %f1154, 0f3F800000;
	// inline asm
	rcp.approx.ftz.f32 %f1141,%f1142;
	// inline asm
	add.f32 	%f1157, %f1156, %f1156;
	mul.f32 	%f1158, %f1141, %f1157;
	mul.f32 	%f1159, %f1158, %f1158;
	mov.f32 	%f1160, 0f3C4CAF63;
	mov.f32 	%f1161, 0f3B18F0FE;
	fma.rn.f32 	%f1162, %f1161, %f1159, %f1160;
	mov.f32 	%f1163, 0f3DAAAABD;
	fma.rn.f32 	%f1164, %f1162, %f1159, %f1163;
	mul.rn.f32 	%f1165, %f1164, %f1159;
	mul.rn.f32 	%f1166, %f1165, %f1158;
	sub.f32 	%f1167, %f1156, %f1158;
	neg.f32 	%f1168, %f1158;
	add.f32 	%f1169, %f1167, %f1167;
	fma.rn.f32 	%f1170, %f1168, %f1156, %f1169;
	mul.rn.f32 	%f1171, %f1141, %f1170;
	add.f32 	%f1172, %f1166, %f1158;
	sub.f32 	%f1173, %f1158, %f1172;
	add.f32 	%f1174, %f1166, %f1173;
	add.f32 	%f1175, %f1171, %f1174;
	add.f32 	%f1176, %f1172, %f1175;
	sub.f32 	%f1177, %f1172, %f1176;
	add.f32 	%f1178, %f1175, %f1177;
	mov.f32 	%f1179, 0f3F317200;
	mul.rn.f32 	%f1180, %f1155, %f1179;
	mov.f32 	%f1181, 0f35BFBE8E;
	mul.rn.f32 	%f1182, %f1155, %f1181;
	add.f32 	%f1183, %f1180, %f1176;
	sub.f32 	%f1184, %f1180, %f1183;
	add.f32 	%f1185, %f1176, %f1184;
	add.f32 	%f1186, %f1178, %f1185;
	add.f32 	%f1187, %f1182, %f1186;
	add.f32 	%f1188, %f1183, %f1187;
	sub.f32 	%f1189, %f1183, %f1188;
	add.f32 	%f1190, %f1187, %f1189;
	mov.f32 	%f1191, 0f3EE8BA2E;
	mul.rn.f32 	%f1192, %f1191, %f1188;
	neg.f32 	%f1193, %f1192;
	fma.rn.f32 	%f1194, %f1191, %f1188, %f1193;
	fma.rn.f32 	%f1195, %f1191, %f1190, %f1194;
	mov.f32 	%f1196, 0f00000000;
	fma.rn.f32 	%f1197, %f1196, %f1188, %f1195;
	add.rn.f32 	%f1198, %f1192, %f1197;
	neg.f32 	%f1199, %f1198;
	add.rn.f32 	%f1200, %f1192, %f1199;
	add.rn.f32 	%f1201, %f1200, %f1197;
	mov.b32 	 %r129, %f1198;
	setp.eq.s32	%p60, %r129, 1118925336;
	add.s32 	%r130, %r129, -1;
	mov.b32 	 %f1202, %r130;
	add.f32 	%f1203, %f1201, 0f37000000;
	selp.f32	%f1204, %f1202, %f1198, %p60;
	selp.f32	%f329, %f1203, %f1201, %p60;
	mul.f32 	%f1205, %f1204, 0f3FB8AA3B;
	cvt.rzi.f32.f32	%f1206, %f1205;
	mov.f32 	%f1207, 0fBF317200;
	fma.rn.f32 	%f1208, %f1206, %f1207, %f1204;
	mov.f32 	%f1209, 0fB5BFBE8E;
	fma.rn.f32 	%f1210, %f1206, %f1209, %f1208;
	mul.f32 	%f1211, %f1210, 0f3FB8AA3B;
	ex2.approx.ftz.f32 	%f1212, %f1211;
	add.f32 	%f1213, %f1206, 0f00000000;
	ex2.approx.f32 	%f1214, %f1213;
	mul.f32 	%f1215, %f1212, %f1214;
	setp.lt.f32	%p61, %f1204, 0fC2D20000;
	selp.f32	%f1216, 0f00000000, %f1215, %p61;
	setp.gt.f32	%p62, %f1204, 0f42D20000;
	selp.f32	%f1800, 0f7F800000, %f1216, %p62;
	setp.eq.f32	%p63, %f1800, 0f7F800000;
	@%p63 bra 	BB0_50;

	fma.rn.f32 	%f1800, %f1800, %f329, %f1800;

BB0_50:
	setp.lt.f32	%p64, %f326, 0f00000000;
	setp.eq.f32	%p65, %f327, 0f3F800000;
	and.pred  	%p4, %p64, %p65;
	mov.b32 	 %r131, %f1800;
	xor.b32  	%r132, %r131, -2147483648;
	mov.b32 	 %f1217, %r132;
	selp.f32	%f1802, %f1217, %f1800, %p4;
	setp.eq.f32	%p66, %f326, 0f00000000;
	@%p66 bra 	BB0_53;
	bra.uni 	BB0_51;

BB0_53:
	add.f32 	%f1220, %f326, %f326;
	selp.f32	%f1802, %f1220, 0f00000000, %p65;
	bra.uni 	BB0_54;

BB0_133:
	mov.u64 	%rd292, image_HDR;
	cvta.global.u64 	%rd287, %rd292;
	mov.u32 	%r253, 8;
	// inline asm
	call (%rd286), _rt_buffer_get_64, (%rd287, %r27, %r253, %rd5, %rd6, %rd12, %rd12);
	// inline asm
	mov.f32 	%f1577, 0f00000000;
	// inline asm
	{  cvt.rn.f16.f32 %rs182, %f1577;}

	// inline asm
	mov.u16 	%rs183, 0;
	st.v4.u16 	[%rd286], {%rs182, %rs182, %rs182, %rs183};

BB0_134:
	ld.global.u32 	%r254, [additive];
	setp.eq.s32	%p138, %r254, 0;
	@%p138 bra 	BB0_136;

	mov.u64 	%rd305, image_RNM0;
	cvta.global.u64 	%rd294, %rd305;
	mov.u32 	%r258, 8;
	// inline asm
	call (%rd293), _rt_buffer_get_64, (%rd294, %r27, %r258, %rd5, %rd6, %rd12, %rd12);
	// inline asm
	ld.v4.u16 	{%rs190, %rs191, %rs192, %rs193}, [%rd293];
	// inline asm
	{  cvt.f32.f16 %f1578, %rs190;}

	// inline asm
	// inline asm
	{  cvt.f32.f16 %f1579, %rs191;}

	// inline asm
	// inline asm
	{  cvt.f32.f16 %f1580, %rs192;}

	// inline asm
	// inline asm
	call (%rd299), _rt_buffer_get_64, (%rd294, %r27, %r258, %rd5, %rd6, %rd12, %rd12);
	// inline asm
	add.f32 	%f1581, %f1578, 0f00000000;
	add.f32 	%f1582, %f1579, 0f00000000;
	add.f32 	%f1583, %f1580, 0f00000000;
	// inline asm
	{  cvt.rn.f16.f32 %rs189, %f1583;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs188, %f1582;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs187, %f1581;}

	// inline asm
	mov.u16 	%rs194, 0;
	st.v4.u16 	[%rd299], {%rs187, %rs188, %rs189, %rs194};
	bra.uni 	BB0_137;

BB0_136:
	mov.u64 	%rd312, image_RNM0;
	cvta.global.u64 	%rd307, %rd312;
	mov.u32 	%r260, 8;
	// inline asm
	call (%rd306), _rt_buffer_get_64, (%rd307, %r27, %r260, %rd5, %rd6, %rd12, %rd12);
	// inline asm
	mov.f32 	%f1584, 0f00000000;
	// inline asm
	{  cvt.rn.f16.f32 %rs195, %f1584;}

	// inline asm
	mov.u16 	%rs196, 0;
	st.v4.u16 	[%rd306], {%rs195, %rs195, %rs195, %rs196};

BB0_137:
	ld.global.u32 	%r261, [additive];
	setp.eq.s32	%p139, %r261, 0;
	@%p139 bra 	BB0_139;

	mov.u64 	%rd325, image_RNM1;
	cvta.global.u64 	%rd314, %rd325;
	mov.u32 	%r265, 8;
	// inline asm
	call (%rd313), _rt_buffer_get_64, (%rd314, %r27, %r265, %rd5, %rd6, %rd12, %rd12);
	// inline asm
	ld.v4.u16 	{%rs203, %rs204, %rs205, %rs206}, [%rd313];
	// inline asm
	{  cvt.f32.f16 %f1585, %rs203;}

	// inline asm
	// inline asm
	{  cvt.f32.f16 %f1586, %rs204;}

	// inline asm
	// inline asm
	{  cvt.f32.f16 %f1587, %rs205;}

	// inline asm
	// inline asm
	call (%rd319), _rt_buffer_get_64, (%rd314, %r27, %r265, %rd5, %rd6, %rd12, %rd12);
	// inline asm
	add.f32 	%f1588, %f1585, 0f00000000;
	add.f32 	%f1589, %f1586, 0f00000000;
	add.f32 	%f1590, %f1587, 0f00000000;
	// inline asm
	{  cvt.rn.f16.f32 %rs202, %f1590;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs201, %f1589;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs200, %f1588;}

	// inline asm
	mov.u16 	%rs207, 0;
	st.v4.u16 	[%rd319], {%rs200, %rs201, %rs202, %rs207};
	bra.uni 	BB0_140;

BB0_139:
	mov.u64 	%rd332, image_RNM1;
	cvta.global.u64 	%rd327, %rd332;
	mov.u32 	%r267, 8;
	// inline asm
	call (%rd326), _rt_buffer_get_64, (%rd327, %r27, %r267, %rd5, %rd6, %rd12, %rd12);
	// inline asm
	mov.f32 	%f1591, 0f00000000;
	// inline asm
	{  cvt.rn.f16.f32 %rs208, %f1591;}

	// inline asm
	mov.u16 	%rs209, 0;
	st.v4.u16 	[%rd326], {%rs208, %rs208, %rs208, %rs209};

BB0_140:
	ld.global.u32 	%r268, [additive];
	setp.eq.s32	%p140, %r268, 0;
	@%p140 bra 	BB0_142;

	mov.u64 	%rd345, image_RNM2;
	cvta.global.u64 	%rd334, %rd345;
	mov.u32 	%r272, 8;
	// inline asm
	call (%rd333), _rt_buffer_get_64, (%rd334, %r27, %r272, %rd5, %rd6, %rd12, %rd12);
	// inline asm
	ld.v4.u16 	{%rs216, %rs217, %rs218, %rs219}, [%rd333];
	// inline asm
	{  cvt.f32.f16 %f1592, %rs216;}

	// inline asm
	// inline asm
	{  cvt.f32.f16 %f1593, %rs217;}

	// inline asm
	// inline asm
	{  cvt.f32.f16 %f1594, %rs218;}

	// inline asm
	// inline asm
	call (%rd339), _rt_buffer_get_64, (%rd334, %r27, %r272, %rd5, %rd6, %rd12, %rd12);
	// inline asm
	add.f32 	%f1595, %f1592, 0f00000000;
	add.f32 	%f1596, %f1593, 0f00000000;
	add.f32 	%f1597, %f1594, 0f00000000;
	// inline asm
	{  cvt.rn.f16.f32 %rs215, %f1597;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs214, %f1596;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs213, %f1595;}

	// inline asm
	mov.u16 	%rs220, 0;
	st.v4.u16 	[%rd339], {%rs213, %rs214, %rs215, %rs220};
	bra.uni 	BB0_143;

BB0_142:
	mov.u64 	%rd352, image_RNM2;
	cvta.global.u64 	%rd347, %rd352;
	mov.u32 	%r274, 8;
	// inline asm
	call (%rd346), _rt_buffer_get_64, (%rd347, %r27, %r274, %rd5, %rd6, %rd12, %rd12);
	// inline asm
	mov.f32 	%f1598, 0f00000000;
	// inline asm
	{  cvt.rn.f16.f32 %rs221, %f1598;}

	// inline asm
	mov.u16 	%rs222, 0;
	st.v4.u16 	[%rd346], {%rs221, %rs221, %rs221, %rs222};

BB0_143:
	ld.global.u32 	%r275, [additive];
	setp.eq.s32	%p141, %r275, 0;
	@%p141 bra 	BB0_145;

	mov.u64 	%rd365, image_RNM3;
	cvta.global.u64 	%rd354, %rd365;
	mov.u32 	%r279, 8;
	// inline asm
	call (%rd353), _rt_buffer_get_64, (%rd354, %r27, %r279, %rd5, %rd6, %rd12, %rd12);
	// inline asm
	ld.v4.u16 	{%rs229, %rs230, %rs231, %rs232}, [%rd353];
	// inline asm
	{  cvt.f32.f16 %f1599, %rs229;}

	// inline asm
	// inline asm
	{  cvt.f32.f16 %f1600, %rs230;}

	// inline asm
	// inline asm
	{  cvt.f32.f16 %f1601, %rs231;}

	// inline asm
	// inline asm
	call (%rd359), _rt_buffer_get_64, (%rd354, %r27, %r279, %rd5, %rd6, %rd12, %rd12);
	// inline asm
	add.f32 	%f1602, %f1599, 0f00000000;
	add.f32 	%f1603, %f1600, 0f00000000;
	add.f32 	%f1604, %f1601, 0f00000000;
	// inline asm
	{  cvt.rn.f16.f32 %rs228, %f1604;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs227, %f1603;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs226, %f1602;}

	// inline asm
	mov.u16 	%rs233, 0;
	st.v4.u16 	[%rd359], {%rs226, %rs227, %rs228, %rs233};
	bra.uni 	BB0_146;

BB0_145:
	mov.u64 	%rd372, image_RNM3;
	cvta.global.u64 	%rd367, %rd372;
	mov.u32 	%r281, 8;
	// inline asm
	call (%rd366), _rt_buffer_get_64, (%rd367, %r27, %r281, %rd5, %rd6, %rd12, %rd12);
	// inline asm
	mov.f32 	%f1605, 0f00000000;
	// inline asm
	{  cvt.rn.f16.f32 %rs234, %f1605;}

	// inline asm
	mov.u16 	%rs235, 0;
	st.v4.u16 	[%rd366], {%rs234, %rs234, %rs234, %rs235};

BB0_146:
	ld.global.u32 	%r282, [additive];
	setp.eq.s32	%p142, %r282, 0;
	@%p142 bra 	BB0_148;

	mov.u64 	%rd385, image_RNM4;
	cvta.global.u64 	%rd374, %rd385;
	mov.u32 	%r286, 8;
	// inline asm
	call (%rd373), _rt_buffer_get_64, (%rd374, %r27, %r286, %rd5, %rd6, %rd12, %rd12);
	// inline asm
	ld.v4.u16 	{%rs242, %rs243, %rs244, %rs245}, [%rd373];
	// inline asm
	{  cvt.f32.f16 %f1606, %rs242;}

	// inline asm
	// inline asm
	{  cvt.f32.f16 %f1607, %rs243;}

	// inline asm
	// inline asm
	{  cvt.f32.f16 %f1608, %rs244;}

	// inline asm
	// inline asm
	call (%rd379), _rt_buffer_get_64, (%rd374, %r27, %r286, %rd5, %rd6, %rd12, %rd12);
	// inline asm
	add.f32 	%f1609, %f1606, 0f00000000;
	add.f32 	%f1610, %f1607, 0f00000000;
	add.f32 	%f1611, %f1608, 0f00000000;
	// inline asm
	{  cvt.rn.f16.f32 %rs241, %f1611;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs240, %f1610;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs239, %f1609;}

	// inline asm
	mov.u16 	%rs246, 0;
	st.v4.u16 	[%rd379], {%rs239, %rs240, %rs241, %rs246};
	bra.uni 	BB0_149;

BB0_148:
	mov.u64 	%rd392, image_RNM4;
	cvta.global.u64 	%rd387, %rd392;
	mov.u32 	%r288, 8;
	// inline asm
	call (%rd386), _rt_buffer_get_64, (%rd387, %r27, %r288, %rd5, %rd6, %rd12, %rd12);
	// inline asm
	mov.f32 	%f1612, 0f00000000;
	// inline asm
	{  cvt.rn.f16.f32 %rs247, %f1612;}

	// inline asm
	mov.u16 	%rs248, 0;
	st.v4.u16 	[%rd386], {%rs247, %rs247, %rs247, %rs248};

BB0_149:
	ld.global.u32 	%r289, [additive];
	setp.eq.s32	%p143, %r289, 0;
	@%p143 bra 	BB0_151;

	mov.u64 	%rd405, image_RNM5;
	cvta.global.u64 	%rd394, %rd405;
	mov.u32 	%r293, 8;
	// inline asm
	call (%rd393), _rt_buffer_get_64, (%rd394, %r27, %r293, %rd5, %rd6, %rd12, %rd12);
	// inline asm
	ld.v4.u16 	{%rs255, %rs256, %rs257, %rs258}, [%rd393];
	// inline asm
	{  cvt.f32.f16 %f1613, %rs255;}

	// inline asm
	// inline asm
	{  cvt.f32.f16 %f1614, %rs256;}

	// inline asm
	// inline asm
	{  cvt.f32.f16 %f1615, %rs257;}

	// inline asm
	// inline asm
	call (%rd399), _rt_buffer_get_64, (%rd394, %r27, %r293, %rd5, %rd6, %rd12, %rd12);
	// inline asm
	add.f32 	%f1616, %f1613, 0f00000000;
	add.f32 	%f1617, %f1614, 0f00000000;
	add.f32 	%f1618, %f1615, 0f00000000;
	// inline asm
	{  cvt.rn.f16.f32 %rs254, %f1618;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs253, %f1617;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs252, %f1616;}

	// inline asm
	mov.u16 	%rs259, 0;
	st.v4.u16 	[%rd399], {%rs252, %rs253, %rs254, %rs259};
	bra.uni 	BB0_152;

BB0_151:
	mov.u64 	%rd412, image_RNM5;
	cvta.global.u64 	%rd407, %rd412;
	mov.u32 	%r295, 8;
	// inline asm
	call (%rd406), _rt_buffer_get_64, (%rd407, %r27, %r295, %rd5, %rd6, %rd12, %rd12);
	// inline asm
	mov.f32 	%f1619, 0f00000000;
	// inline asm
	{  cvt.rn.f16.f32 %rs260, %f1619;}

	// inline asm
	mov.u16 	%rs261, 0;
	st.v4.u16 	[%rd406], {%rs260, %rs260, %rs260, %rs261};

BB0_152:
	ld.global.u32 	%r296, [additive];
	setp.eq.s32	%p144, %r296, 0;
	@%p144 bra 	BB0_154;

	mov.u64 	%rd425, image_RNM6;
	cvta.global.u64 	%rd414, %rd425;
	mov.u32 	%r300, 8;
	// inline asm
	call (%rd413), _rt_buffer_get_64, (%rd414, %r27, %r300, %rd5, %rd6, %rd12, %rd12);
	// inline asm
	ld.v4.u16 	{%rs268, %rs269, %rs270, %rs271}, [%rd413];
	// inline asm
	{  cvt.f32.f16 %f1620, %rs268;}

	// inline asm
	// inline asm
	{  cvt.f32.f16 %f1621, %rs269;}

	// inline asm
	// inline asm
	{  cvt.f32.f16 %f1622, %rs270;}

	// inline asm
	// inline asm
	call (%rd419), _rt_buffer_get_64, (%rd414, %r27, %r300, %rd5, %rd6, %rd12, %rd12);
	// inline asm
	add.f32 	%f1623, %f1620, 0f00000000;
	add.f32 	%f1624, %f1621, 0f00000000;
	add.f32 	%f1625, %f1622, 0f00000000;
	// inline asm
	{  cvt.rn.f16.f32 %rs267, %f1625;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs266, %f1624;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs265, %f1623;}

	// inline asm
	mov.u16 	%rs272, 0;
	st.v4.u16 	[%rd419], {%rs265, %rs266, %rs267, %rs272};
	bra.uni 	BB0_155;

BB0_154:
	mov.u64 	%rd432, image_RNM6;
	cvta.global.u64 	%rd427, %rd432;
	mov.u32 	%r302, 8;
	// inline asm
	call (%rd426), _rt_buffer_get_64, (%rd427, %r27, %r302, %rd5, %rd6, %rd12, %rd12);
	// inline asm
	mov.f32 	%f1626, 0f00000000;
	// inline asm
	{  cvt.rn.f16.f32 %rs273, %f1626;}

	// inline asm
	mov.u16 	%rs274, 0;
	st.v4.u16 	[%rd426], {%rs273, %rs273, %rs273, %rs274};

BB0_155:
	ld.global.u32 	%r303, [additive];
	setp.eq.s32	%p145, %r303, 0;
	@%p145 bra 	BB0_157;

	mov.u64 	%rd445, image_RNM7;
	cvta.global.u64 	%rd434, %rd445;
	mov.u32 	%r307, 8;
	// inline asm
	call (%rd433), _rt_buffer_get_64, (%rd434, %r27, %r307, %rd5, %rd6, %rd12, %rd12);
	// inline asm
	ld.v4.u16 	{%rs281, %rs282, %rs283, %rs284}, [%rd433];
	// inline asm
	{  cvt.f32.f16 %f1627, %rs281;}

	// inline asm
	// inline asm
	{  cvt.f32.f16 %f1628, %rs282;}

	// inline asm
	// inline asm
	{  cvt.f32.f16 %f1629, %rs283;}

	// inline asm
	// inline asm
	call (%rd439), _rt_buffer_get_64, (%rd434, %r27, %r307, %rd5, %rd6, %rd12, %rd12);
	// inline asm
	add.f32 	%f1630, %f1627, 0f00000000;
	add.f32 	%f1631, %f1628, 0f00000000;
	add.f32 	%f1632, %f1629, 0f00000000;
	// inline asm
	{  cvt.rn.f16.f32 %rs280, %f1632;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs279, %f1631;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs278, %f1630;}

	// inline asm
	mov.u16 	%rs285, 0;
	st.v4.u16 	[%rd439], {%rs278, %rs279, %rs280, %rs285};
	bra.uni 	BB0_158;

BB0_157:
	mov.u64 	%rd452, image_RNM7;
	cvta.global.u64 	%rd447, %rd452;
	mov.u32 	%r309, 8;
	// inline asm
	call (%rd446), _rt_buffer_get_64, (%rd447, %r27, %r309, %rd5, %rd6, %rd12, %rd12);
	// inline asm
	mov.f32 	%f1633, 0f00000000;
	// inline asm
	{  cvt.rn.f16.f32 %rs286, %f1633;}

	// inline asm
	mov.u16 	%rs287, 0;
	st.v4.u16 	[%rd446], {%rs286, %rs286, %rs286, %rs287};

BB0_158:
	ld.global.u32 	%r310, [additive];
	setp.eq.s32	%p146, %r310, 0;
	@%p146 bra 	BB0_160;

	mov.u64 	%rd465, image_RNM8;
	cvta.global.u64 	%rd454, %rd465;
	mov.u32 	%r314, 8;
	// inline asm
	call (%rd453), _rt_buffer_get_64, (%rd454, %r27, %r314, %rd5, %rd6, %rd12, %rd12);
	// inline asm
	ld.v4.u16 	{%rs294, %rs295, %rs296, %rs297}, [%rd453];
	// inline asm
	{  cvt.f32.f16 %f1634, %rs294;}

	// inline asm
	// inline asm
	{  cvt.f32.f16 %f1635, %rs295;}

	// inline asm
	// inline asm
	{  cvt.f32.f16 %f1636, %rs296;}

	// inline asm
	// inline asm
	call (%rd459), _rt_buffer_get_64, (%rd454, %r27, %r314, %rd5, %rd6, %rd12, %rd12);
	// inline asm
	add.f32 	%f1637, %f1634, 0f00000000;
	add.f32 	%f1638, %f1635, 0f00000000;
	add.f32 	%f1639, %f1636, 0f00000000;
	// inline asm
	{  cvt.rn.f16.f32 %rs293, %f1639;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs292, %f1638;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs291, %f1637;}

	// inline asm
	mov.u16 	%rs298, 0;
	st.v4.u16 	[%rd459], {%rs291, %rs292, %rs293, %rs298};
	bra.uni 	BB0_161;

BB0_160:
	mov.u64 	%rd472, image_RNM8;
	cvta.global.u64 	%rd467, %rd472;
	mov.u32 	%r316, 8;
	// inline asm
	call (%rd466), _rt_buffer_get_64, (%rd467, %r27, %r316, %rd5, %rd6, %rd12, %rd12);
	// inline asm
	mov.f32 	%f1640, 0f00000000;
	// inline asm
	{  cvt.rn.f16.f32 %rs299, %f1640;}

	// inline asm
	mov.u16 	%rs300, 0;
	st.v4.u16 	[%rd466], {%rs299, %rs299, %rs299, %rs300};
	bra.uni 	BB0_161;

BB0_51:
	setp.geu.f32	%p67, %f326, 0f00000000;
	@%p67 bra 	BB0_54;

	cvt.rzi.f32.f32	%f1219, %f1191;
	setp.neu.f32	%p68, %f1219, 0f3EE8BA2E;
	selp.f32	%f1802, 0f7FFFFFFF, %f1802, %p68;

BB0_54:
	add.f32 	%f1221, %f328, 0f3EE8BA2E;
	mov.b32 	 %r133, %f1221;
	setp.lt.s32	%p70, %r133, 2139095040;
	@%p70 bra 	BB0_59;

	setp.gtu.f32	%p71, %f328, 0f7F800000;
	@%p71 bra 	BB0_58;
	bra.uni 	BB0_56;

BB0_58:
	add.f32 	%f1802, %f326, 0f3EE8BA2E;
	bra.uni 	BB0_59;

BB0_56:
	setp.neu.f32	%p72, %f328, 0f7F800000;
	@%p72 bra 	BB0_59;

	selp.f32	%f1802, 0fFF800000, 0f7F800000, %p4;

BB0_59:
	mul.f32 	%f1222, %f1802, 0f437F0000;
	setp.eq.f32	%p73, %f326, 0f3F800000;
	selp.f32	%f1223, 0f437F0000, %f1222, %p73;
	cvt.rzi.u32.f32	%r134, %f1223;
	cvt.u16.u32	%rs22, %r134;
	mov.u16 	%rs23, 255;
	st.v2.u8 	[%rd45], {%rs22, %rs23};
	ld.global.u32 	%r324, [imageEnabled];

BB0_60:
	and.b32  	%r135, %r324, 1;
	setp.eq.b32	%p74, %r135, 1;
	@!%p74 bra 	BB0_95;
	bra.uni 	BB0_61;

BB0_61:
	mov.f32 	%f1226, 0f3E666666;
	cvt.rzi.f32.f32	%f1227, %f1226;
	fma.rn.f32 	%f1228, %f1227, 0fC0000000, 0f3EE66666;
	abs.f32 	%f340, %f1228;
	abs.f32 	%f341, %f1677;
	setp.lt.f32	%p75, %f341, 0f00800000;
	mul.f32 	%f1229, %f341, 0f4B800000;
	selp.f32	%f1230, 0fC3170000, 0fC2FE0000, %p75;
	selp.f32	%f1231, %f1229, %f341, %p75;
	mov.b32 	 %r136, %f1231;
	and.b32  	%r137, %r136, 8388607;
	or.b32  	%r138, %r137, 1065353216;
	mov.b32 	 %f1232, %r138;
	shr.u32 	%r139, %r136, 23;
	cvt.rn.f32.u32	%f1233, %r139;
	add.f32 	%f1234, %f1230, %f1233;
	setp.gt.f32	%p76, %f1232, 0f3FB504F3;
	mul.f32 	%f1235, %f1232, 0f3F000000;
	add.f32 	%f1236, %f1234, 0f3F800000;
	selp.f32	%f1237, %f1235, %f1232, %p76;
	selp.f32	%f1238, %f1236, %f1234, %p76;
	add.f32 	%f1239, %f1237, 0fBF800000;
	add.f32 	%f1225, %f1237, 0f3F800000;
	// inline asm
	rcp.approx.ftz.f32 %f1224,%f1225;
	// inline asm
	add.f32 	%f1240, %f1239, %f1239;
	mul.f32 	%f1241, %f1224, %f1240;
	mul.f32 	%f1242, %f1241, %f1241;
	mov.f32 	%f1243, 0f3C4CAF63;
	mov.f32 	%f1244, 0f3B18F0FE;
	fma.rn.f32 	%f1245, %f1244, %f1242, %f1243;
	mov.f32 	%f1246, 0f3DAAAABD;
	fma.rn.f32 	%f1247, %f1245, %f1242, %f1246;
	mul.rn.f32 	%f1248, %f1247, %f1242;
	mul.rn.f32 	%f1249, %f1248, %f1241;
	sub.f32 	%f1250, %f1239, %f1241;
	neg.f32 	%f1251, %f1241;
	add.f32 	%f1252, %f1250, %f1250;
	fma.rn.f32 	%f1253, %f1251, %f1239, %f1252;
	mul.rn.f32 	%f1254, %f1224, %f1253;
	add.f32 	%f1255, %f1249, %f1241;
	sub.f32 	%f1256, %f1241, %f1255;
	add.f32 	%f1257, %f1249, %f1256;
	add.f32 	%f1258, %f1254, %f1257;
	add.f32 	%f1259, %f1255, %f1258;
	sub.f32 	%f1260, %f1255, %f1259;
	add.f32 	%f1261, %f1258, %f1260;
	mov.f32 	%f1262, 0f3F317200;
	mul.rn.f32 	%f1263, %f1238, %f1262;
	mov.f32 	%f1264, 0f35BFBE8E;
	mul.rn.f32 	%f1265, %f1238, %f1264;
	add.f32 	%f1266, %f1263, %f1259;
	sub.f32 	%f1267, %f1263, %f1266;
	add.f32 	%f1268, %f1259, %f1267;
	add.f32 	%f1269, %f1261, %f1268;
	add.f32 	%f1270, %f1265, %f1269;
	add.f32 	%f1271, %f1266, %f1270;
	sub.f32 	%f1272, %f1266, %f1271;
	add.f32 	%f1273, %f1270, %f1272;
	mov.f32 	%f1274, 0f3EE66666;
	mul.rn.f32 	%f1275, %f1274, %f1271;
	neg.f32 	%f1276, %f1275;
	fma.rn.f32 	%f1277, %f1274, %f1271, %f1276;
	fma.rn.f32 	%f1278, %f1274, %f1273, %f1277;
	mov.f32 	%f1279, 0f00000000;
	fma.rn.f32 	%f1280, %f1279, %f1271, %f1278;
	add.rn.f32 	%f1281, %f1275, %f1280;
	neg.f32 	%f1282, %f1281;
	add.rn.f32 	%f1283, %f1275, %f1282;
	add.rn.f32 	%f1284, %f1283, %f1280;
	mov.b32 	 %r140, %f1281;
	setp.eq.s32	%p77, %r140, 1118925336;
	add.s32 	%r141, %r140, -1;
	mov.b32 	 %f1285, %r141;
	add.f32 	%f1286, %f1284, 0f37000000;
	selp.f32	%f1287, %f1285, %f1281, %p77;
	selp.f32	%f342, %f1286, %f1284, %p77;
	mul.f32 	%f1288, %f1287, 0f3FB8AA3B;
	cvt.rzi.f32.f32	%f1289, %f1288;
	mov.f32 	%f1290, 0fBF317200;
	fma.rn.f32 	%f1291, %f1289, %f1290, %f1287;
	mov.f32 	%f1292, 0fB5BFBE8E;
	fma.rn.f32 	%f1293, %f1289, %f1292, %f1291;
	mul.f32 	%f1294, %f1293, 0f3FB8AA3B;
	ex2.approx.ftz.f32 	%f1295, %f1294;
	add.f32 	%f1296, %f1289, 0f00000000;
	ex2.approx.f32 	%f1297, %f1296;
	mul.f32 	%f1298, %f1295, %f1297;
	setp.lt.f32	%p78, %f1287, 0fC2D20000;
	selp.f32	%f1299, 0f00000000, %f1298, %p78;
	setp.gt.f32	%p79, %f1287, 0f42D20000;
	selp.f32	%f1803, 0f7F800000, %f1299, %p79;
	setp.eq.f32	%p80, %f1803, 0f7F800000;
	@%p80 bra 	BB0_63;

	fma.rn.f32 	%f1803, %f1803, %f342, %f1803;

BB0_63:
	setp.lt.f32	%p81, %f1677, 0f00000000;
	setp.eq.f32	%p82, %f340, 0f3F800000;
	and.pred  	%p5, %p81, %p82;
	mov.b32 	 %r142, %f1803;
	xor.b32  	%r143, %r142, -2147483648;
	mov.b32 	 %f1300, %r143;
	selp.f32	%f1805, %f1300, %f1803, %p5;
	setp.eq.f32	%p83, %f1677, 0f00000000;
	@%p83 bra 	BB0_66;
	bra.uni 	BB0_64;

BB0_66:
	add.f32 	%f1303, %f1677, %f1677;
	selp.f32	%f1805, %f1303, 0f00000000, %p82;
	bra.uni 	BB0_67;

BB0_64:
	setp.geu.f32	%p84, %f1677, 0f00000000;
	@%p84 bra 	BB0_67;

	cvt.rzi.f32.f32	%f1302, %f1274;
	setp.neu.f32	%p85, %f1302, 0f3EE66666;
	selp.f32	%f1805, 0f7FFFFFFF, %f1805, %p85;

BB0_67:
	add.f32 	%f1304, %f341, 0f3EE66666;
	mov.b32 	 %r144, %f1304;
	setp.lt.s32	%p87, %r144, 2139095040;
	@%p87 bra 	BB0_72;

	setp.gtu.f32	%p88, %f341, 0f7F800000;
	@%p88 bra 	BB0_71;
	bra.uni 	BB0_69;

BB0_71:
	add.f32 	%f1805, %f1677, 0f3EE66666;
	bra.uni 	BB0_72;

BB0_69:
	setp.neu.f32	%p89, %f341, 0f7F800000;
	@%p89 bra 	BB0_72;

	selp.f32	%f1805, 0fFF800000, 0f7F800000, %p5;

BB0_72:
	setp.eq.f32	%p90, %f1677, 0f3F800000;
	selp.f32	%f353, 0f3F800000, %f1805, %p90;
	abs.f32 	%f354, %f1676;
	setp.lt.f32	%p91, %f354, 0f00800000;
	mul.f32 	%f1307, %f354, 0f4B800000;
	selp.f32	%f1308, 0fC3170000, 0fC2FE0000, %p91;
	selp.f32	%f1309, %f1307, %f354, %p91;
	mov.b32 	 %r145, %f1309;
	and.b32  	%r146, %r145, 8388607;
	or.b32  	%r147, %r146, 1065353216;
	mov.b32 	 %f1310, %r147;
	shr.u32 	%r148, %r145, 23;
	cvt.rn.f32.u32	%f1311, %r148;
	add.f32 	%f1312, %f1308, %f1311;
	setp.gt.f32	%p92, %f1310, 0f3FB504F3;
	mul.f32 	%f1313, %f1310, 0f3F000000;
	add.f32 	%f1314, %f1312, 0f3F800000;
	selp.f32	%f1315, %f1313, %f1310, %p92;
	selp.f32	%f1316, %f1314, %f1312, %p92;
	add.f32 	%f1317, %f1315, 0fBF800000;
	add.f32 	%f1306, %f1315, 0f3F800000;
	// inline asm
	rcp.approx.ftz.f32 %f1305,%f1306;
	// inline asm
	add.f32 	%f1318, %f1317, %f1317;
	mul.f32 	%f1319, %f1305, %f1318;
	mul.f32 	%f1320, %f1319, %f1319;
	fma.rn.f32 	%f1323, %f1244, %f1320, %f1243;
	fma.rn.f32 	%f1325, %f1323, %f1320, %f1246;
	mul.rn.f32 	%f1326, %f1325, %f1320;
	mul.rn.f32 	%f1327, %f1326, %f1319;
	sub.f32 	%f1328, %f1317, %f1319;
	neg.f32 	%f1329, %f1319;
	add.f32 	%f1330, %f1328, %f1328;
	fma.rn.f32 	%f1331, %f1329, %f1317, %f1330;
	mul.rn.f32 	%f1332, %f1305, %f1331;
	add.f32 	%f1333, %f1327, %f1319;
	sub.f32 	%f1334, %f1319, %f1333;
	add.f32 	%f1335, %f1327, %f1334;
	add.f32 	%f1336, %f1332, %f1335;
	add.f32 	%f1337, %f1333, %f1336;
	sub.f32 	%f1338, %f1333, %f1337;
	add.f32 	%f1339, %f1336, %f1338;
	mul.rn.f32 	%f1341, %f1316, %f1262;
	mul.rn.f32 	%f1343, %f1316, %f1264;
	add.f32 	%f1344, %f1341, %f1337;
	sub.f32 	%f1345, %f1341, %f1344;
	add.f32 	%f1346, %f1337, %f1345;
	add.f32 	%f1347, %f1339, %f1346;
	add.f32 	%f1348, %f1343, %f1347;
	add.f32 	%f1349, %f1344, %f1348;
	sub.f32 	%f1350, %f1344, %f1349;
	add.f32 	%f1351, %f1348, %f1350;
	mul.rn.f32 	%f1353, %f1274, %f1349;
	neg.f32 	%f1354, %f1353;
	fma.rn.f32 	%f1355, %f1274, %f1349, %f1354;
	fma.rn.f32 	%f1356, %f1274, %f1351, %f1355;
	fma.rn.f32 	%f1358, %f1279, %f1349, %f1356;
	add.rn.f32 	%f1359, %f1353, %f1358;
	neg.f32 	%f1360, %f1359;
	add.rn.f32 	%f1361, %f1353, %f1360;
	add.rn.f32 	%f1362, %f1361, %f1358;
	mov.b32 	 %r149, %f1359;
	setp.eq.s32	%p93, %r149, 1118925336;
	add.s32 	%r150, %r149, -1;
	mov.b32 	 %f1363, %r150;
	add.f32 	%f1364, %f1362, 0f37000000;
	selp.f32	%f1365, %f1363, %f1359, %p93;
	selp.f32	%f355, %f1364, %f1362, %p93;
	mul.f32 	%f1366, %f1365, 0f3FB8AA3B;
	cvt.rzi.f32.f32	%f1367, %f1366;
	fma.rn.f32 	%f1369, %f1367, %f1290, %f1365;
	fma.rn.f32 	%f1371, %f1367, %f1292, %f1369;
	mul.f32 	%f1372, %f1371, 0f3FB8AA3B;
	ex2.approx.ftz.f32 	%f1373, %f1372;
	add.f32 	%f1374, %f1367, 0f00000000;
	ex2.approx.f32 	%f1375, %f1374;
	mul.f32 	%f1376, %f1373, %f1375;
	setp.lt.f32	%p94, %f1365, 0fC2D20000;
	selp.f32	%f1377, 0f00000000, %f1376, %p94;
	setp.gt.f32	%p95, %f1365, 0f42D20000;
	selp.f32	%f1806, 0f7F800000, %f1377, %p95;
	setp.eq.f32	%p96, %f1806, 0f7F800000;
	@%p96 bra 	BB0_74;

	fma.rn.f32 	%f1806, %f1806, %f355, %f1806;

BB0_74:
	setp.lt.f32	%p97, %f1676, 0f00000000;
	and.pred  	%p6, %p97, %p82;
	mov.b32 	 %r151, %f1806;
	xor.b32  	%r152, %r151, -2147483648;
	mov.b32 	 %f1378, %r152;
	selp.f32	%f1808, %f1378, %f1806, %p6;
	setp.eq.f32	%p99, %f1676, 0f00000000;
	@%p99 bra 	BB0_77;
	bra.uni 	BB0_75;

BB0_77:
	add.f32 	%f1381, %f1676, %f1676;
	selp.f32	%f1808, %f1381, 0f00000000, %p82;
	bra.uni 	BB0_78;

BB0_75:
	setp.geu.f32	%p100, %f1676, 0f00000000;
	@%p100 bra 	BB0_78;

	cvt.rzi.f32.f32	%f1380, %f1274;
	setp.neu.f32	%p101, %f1380, 0f3EE66666;
	selp.f32	%f1808, 0f7FFFFFFF, %f1808, %p101;

BB0_78:
	add.f32 	%f1382, %f354, 0f3EE66666;
	mov.b32 	 %r153, %f1382;
	setp.lt.s32	%p103, %r153, 2139095040;
	@%p103 bra 	BB0_83;

	setp.gtu.f32	%p104, %f354, 0f7F800000;
	@%p104 bra 	BB0_82;
	bra.uni 	BB0_80;

BB0_82:
	add.f32 	%f1808, %f1676, 0f3EE66666;
	bra.uni 	BB0_83;

BB0_80:
	setp.neu.f32	%p105, %f354, 0f7F800000;
	@%p105 bra 	BB0_83;

	selp.f32	%f1808, 0fFF800000, 0f7F800000, %p6;

BB0_83:
	setp.eq.f32	%p106, %f1676, 0f3F800000;
	selp.f32	%f366, 0f3F800000, %f1808, %p106;
	abs.f32 	%f367, %f1675;
	setp.lt.f32	%p107, %f367, 0f00800000;
	mul.f32 	%f1385, %f367, 0f4B800000;
	selp.f32	%f1386, 0fC3170000, 0fC2FE0000, %p107;
	selp.f32	%f1387, %f1385, %f367, %p107;
	mov.b32 	 %r154, %f1387;
	and.b32  	%r155, %r154, 8388607;
	or.b32  	%r156, %r155, 1065353216;
	mov.b32 	 %f1388, %r156;
	shr.u32 	%r157, %r154, 23;
	cvt.rn.f32.u32	%f1389, %r157;
	add.f32 	%f1390, %f1386, %f1389;
	setp.gt.f32	%p108, %f1388, 0f3FB504F3;
	mul.f32 	%f1391, %f1388, 0f3F000000;
	add.f32 	%f1392, %f1390, 0f3F800000;
	selp.f32	%f1393, %f1391, %f1388, %p108;
	selp.f32	%f1394, %f1392, %f1390, %p108;
	add.f32 	%f1395, %f1393, 0fBF800000;
	add.f32 	%f1384, %f1393, 0f3F800000;
	// inline asm
	rcp.approx.ftz.f32 %f1383,%f1384;
	// inline asm
	add.f32 	%f1396, %f1395, %f1395;
	mul.f32 	%f1397, %f1383, %f1396;
	mul.f32 	%f1398, %f1397, %f1397;
	fma.rn.f32 	%f1401, %f1244, %f1398, %f1243;
	fma.rn.f32 	%f1403, %f1401, %f1398, %f1246;
	mul.rn.f32 	%f1404, %f1403, %f1398;
	mul.rn.f32 	%f1405, %f1404, %f1397;
	sub.f32 	%f1406, %f1395, %f1397;
	neg.f32 	%f1407, %f1397;
	add.f32 	%f1408, %f1406, %f1406;
	fma.rn.f32 	%f1409, %f1407, %f1395, %f1408;
	mul.rn.f32 	%f1410, %f1383, %f1409;
	add.f32 	%f1411, %f1405, %f1397;
	sub.f32 	%f1412, %f1397, %f1411;
	add.f32 	%f1413, %f1405, %f1412;
	add.f32 	%f1414, %f1410, %f1413;
	add.f32 	%f1415, %f1411, %f1414;
	sub.f32 	%f1416, %f1411, %f1415;
	add.f32 	%f1417, %f1414, %f1416;
	mul.rn.f32 	%f1419, %f1394, %f1262;
	mul.rn.f32 	%f1421, %f1394, %f1264;
	add.f32 	%f1422, %f1419, %f1415;
	sub.f32 	%f1423, %f1419, %f1422;
	add.f32 	%f1424, %f1415, %f1423;
	add.f32 	%f1425, %f1417, %f1424;
	add.f32 	%f1426, %f1421, %f1425;
	add.f32 	%f1427, %f1422, %f1426;
	sub.f32 	%f1428, %f1422, %f1427;
	add.f32 	%f1429, %f1426, %f1428;
	mul.rn.f32 	%f1431, %f1274, %f1427;
	neg.f32 	%f1432, %f1431;
	fma.rn.f32 	%f1433, %f1274, %f1427, %f1432;
	fma.rn.f32 	%f1434, %f1274, %f1429, %f1433;
	fma.rn.f32 	%f1436, %f1279, %f1427, %f1434;
	add.rn.f32 	%f1437, %f1431, %f1436;
	neg.f32 	%f1438, %f1437;
	add.rn.f32 	%f1439, %f1431, %f1438;
	add.rn.f32 	%f1440, %f1439, %f1436;
	mov.b32 	 %r158, %f1437;
	setp.eq.s32	%p109, %r158, 1118925336;
	add.s32 	%r159, %r158, -1;
	mov.b32 	 %f1441, %r159;
	add.f32 	%f1442, %f1440, 0f37000000;
	selp.f32	%f1443, %f1441, %f1437, %p109;
	selp.f32	%f368, %f1442, %f1440, %p109;
	mul.f32 	%f1444, %f1443, 0f3FB8AA3B;
	cvt.rzi.f32.f32	%f1445, %f1444;
	fma.rn.f32 	%f1447, %f1445, %f1290, %f1443;
	fma.rn.f32 	%f1449, %f1445, %f1292, %f1447;
	mul.f32 	%f1450, %f1449, 0f3FB8AA3B;
	ex2.approx.ftz.f32 	%f1451, %f1450;
	add.f32 	%f1452, %f1445, 0f00000000;
	ex2.approx.f32 	%f1453, %f1452;
	mul.f32 	%f1454, %f1451, %f1453;
	setp.lt.f32	%p110, %f1443, 0fC2D20000;
	selp.f32	%f1455, 0f00000000, %f1454, %p110;
	setp.gt.f32	%p111, %f1443, 0f42D20000;
	selp.f32	%f1809, 0f7F800000, %f1455, %p111;
	setp.eq.f32	%p112, %f1809, 0f7F800000;
	@%p112 bra 	BB0_85;

	fma.rn.f32 	%f1809, %f1809, %f368, %f1809;

BB0_85:
	setp.lt.f32	%p113, %f1675, 0f00000000;
	and.pred  	%p7, %p113, %p82;
	mov.b32 	 %r160, %f1809;
	xor.b32  	%r161, %r160, -2147483648;
	mov.b32 	 %f1456, %r161;
	selp.f32	%f1811, %f1456, %f1809, %p7;
	setp.eq.f32	%p115, %f1675, 0f00000000;
	@%p115 bra 	BB0_88;
	bra.uni 	BB0_86;

BB0_88:
	add.f32 	%f1459, %f1675, %f1675;
	selp.f32	%f1811, %f1459, 0f00000000, %p82;
	bra.uni 	BB0_89;

BB0_86:
	setp.geu.f32	%p116, %f1675, 0f00000000;
	@%p116 bra 	BB0_89;

	cvt.rzi.f32.f32	%f1458, %f1274;
	setp.neu.f32	%p117, %f1458, 0f3EE66666;
	selp.f32	%f1811, 0f7FFFFFFF, %f1811, %p117;

BB0_89:
	add.f32 	%f1460, %f367, 0f3EE66666;
	mov.b32 	 %r162, %f1460;
	setp.lt.s32	%p119, %r162, 2139095040;
	@%p119 bra 	BB0_94;

	setp.gtu.f32	%p120, %f367, 0f7F800000;
	@%p120 bra 	BB0_93;
	bra.uni 	BB0_91;

BB0_93:
	add.f32 	%f1811, %f1675, 0f3EE66666;
	bra.uni 	BB0_94;

BB0_91:
	setp.neu.f32	%p121, %f367, 0f7F800000;
	@%p121 bra 	BB0_94;

	selp.f32	%f1811, 0fFF800000, 0f7F800000, %p7;

BB0_94:
	mov.u32 	%r317, 4;
	setp.eq.f32	%p122, %f1675, 0f3F800000;
	selp.f32	%f1461, 0f3F800000, %f1811, %p122;
	cvt.u64.u32	%rd55, %r3;
	cvt.u64.u32	%rd54, %r2;
	mov.u64 	%rd58, image;
	cvta.global.u64 	%rd53, %rd58;
	// inline asm
	call (%rd52), _rt_buffer_get_64, (%rd53, %r27, %r317, %rd54, %rd55, %rd12, %rd12);
	// inline asm
	cvt.sat.f32.f32	%f1462, %f1461;
	mul.f32 	%f1463, %f1462, 0f437FFD71;
	cvt.rzi.u32.f32	%r165, %f1463;
	cvt.sat.f32.f32	%f1464, %f366;
	mul.f32 	%f1465, %f1464, 0f437FFD71;
	cvt.rzi.u32.f32	%r166, %f1465;
	cvt.sat.f32.f32	%f1466, %f353;
	mul.f32 	%f1467, %f1466, 0f437FFD71;
	cvt.rzi.u32.f32	%r167, %f1467;
	cvt.u16.u32	%rs24, %r165;
	cvt.u16.u32	%rs25, %r167;
	cvt.u16.u32	%rs26, %r166;
	mov.u16 	%rs27, 255;
	st.v4.u8 	[%rd52], {%rs24, %rs26, %rs25, %rs27};
	ld.global.u32 	%r324, [imageEnabled];

BB0_95:
	cvt.u64.u32	%rd3, %r2;
	cvt.u64.u32	%rd4, %r3;
	and.b32  	%r168, %r324, 4;
	setp.eq.s32	%p123, %r168, 0;
	@%p123 bra 	BB0_99;

	ld.global.u32 	%r169, [additive];
	setp.eq.s32	%p124, %r169, 0;
	mov.f32 	%f1468, 0f3F800000;
	// inline asm
	{  cvt.rn.f16.f32 %rs28, %f1468;}

	// inline asm
	@%p124 bra 	BB0_98;

	mov.u64 	%rd71, image_HDR;
	cvta.global.u64 	%rd60, %rd71;
	mov.u32 	%r173, 8;
	// inline asm
	call (%rd59), _rt_buffer_get_64, (%rd60, %r27, %r173, %rd3, %rd4, %rd12, %rd12);
	// inline asm
	ld.v4.u16 	{%rs35, %rs36, %rs37, %rs38}, [%rd59];
	// inline asm
	{  cvt.f32.f16 %f1469, %rs35;}

	// inline asm
	// inline asm
	{  cvt.f32.f16 %f1470, %rs36;}

	// inline asm
	// inline asm
	{  cvt.f32.f16 %f1471, %rs37;}

	// inline asm
	// inline asm
	call (%rd65), _rt_buffer_get_64, (%rd60, %r27, %r173, %rd3, %rd4, %rd12, %rd12);
	// inline asm
	add.f32 	%f1472, %f1677, %f1469;
	add.f32 	%f1473, %f1676, %f1470;
	add.f32 	%f1474, %f1675, %f1471;
	// inline asm
	{  cvt.rn.f16.f32 %rs34, %f1474;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs33, %f1473;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs32, %f1472;}

	// inline asm
	st.v4.u16 	[%rd65], {%rs32, %rs33, %rs34, %rs28};
	bra.uni 	BB0_99;

BB0_98:
	mov.u64 	%rd78, image_HDR;
	cvta.global.u64 	%rd73, %rd78;
	mov.u32 	%r175, 8;
	// inline asm
	call (%rd72), _rt_buffer_get_64, (%rd73, %r27, %r175, %rd3, %rd4, %rd12, %rd12);
	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs41, %f1675;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs40, %f1676;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs39, %f1677;}

	// inline asm
	st.v4.u16 	[%rd72], {%rs39, %rs40, %rs41, %rs28};

BB0_99:
	mul.f32 	%f379, %f1674, 0f3F558750;
	mul.f32 	%f380, %f1673, 0f3F558750;
	mul.f32 	%f381, %f1672, 0f3F558750;
	ld.global.u32 	%r176, [additive];
	setp.eq.s32	%p125, %r176, 0;
	mov.f32 	%f1478, 0f3F800000;
	// inline asm
	{  cvt.rn.f16.f32 %rs42, %f1478;}

	// inline asm
	@%p125 bra 	BB0_101;

	mov.u64 	%rd91, image_RNM0;
	cvta.global.u64 	%rd80, %rd91;
	mov.u32 	%r180, 8;
	// inline asm
	call (%rd79), _rt_buffer_get_64, (%rd80, %r27, %r180, %rd3, %rd4, %rd12, %rd12);
	// inline asm
	ld.v4.u16 	{%rs49, %rs50, %rs51, %rs52}, [%rd79];
	// inline asm
	{  cvt.f32.f16 %f1479, %rs49;}

	// inline asm
	// inline asm
	{  cvt.f32.f16 %f1480, %rs50;}

	// inline asm
	// inline asm
	{  cvt.f32.f16 %f1481, %rs51;}

	// inline asm
	// inline asm
	call (%rd85), _rt_buffer_get_64, (%rd80, %r27, %r180, %rd3, %rd4, %rd12, %rd12);
	// inline asm
	add.f32 	%f1482, %f379, %f1479;
	add.f32 	%f1483, %f380, %f1480;
	add.f32 	%f1484, %f381, %f1481;
	// inline asm
	{  cvt.rn.f16.f32 %rs48, %f1484;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs47, %f1483;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs46, %f1482;}

	// inline asm
	st.v4.u16 	[%rd85], {%rs46, %rs47, %rs48, %rs42};
	bra.uni 	BB0_102;

BB0_101:
	mov.u64 	%rd98, image_RNM0;
	cvta.global.u64 	%rd93, %rd98;
	mov.u32 	%r182, 8;
	// inline asm
	call (%rd92), _rt_buffer_get_64, (%rd93, %r27, %r182, %rd3, %rd4, %rd12, %rd12);
	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs55, %f381;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs54, %f380;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs53, %f379;}

	// inline asm
	st.v4.u16 	[%rd92], {%rs53, %rs54, %rs55, %rs42};

BB0_102:
	mul.f32 	%f382, %f1671, 0f3FB8EBD1;
	mul.f32 	%f383, %f1670, 0f3FB8EBD1;
	mul.f32 	%f384, %f1669, 0f3FB8EBD1;
	ld.global.u32 	%r183, [additive];
	setp.eq.s32	%p126, %r183, 0;
	// inline asm
	{  cvt.rn.f16.f32 %rs56, %f1478;}

	// inline asm
	@%p126 bra 	BB0_104;

	mov.u64 	%rd111, image_RNM1;
	cvta.global.u64 	%rd100, %rd111;
	mov.u32 	%r187, 8;
	// inline asm
	call (%rd99), _rt_buffer_get_64, (%rd100, %r27, %r187, %rd3, %rd4, %rd12, %rd12);
	// inline asm
	ld.v4.u16 	{%rs63, %rs64, %rs65, %rs66}, [%rd99];
	// inline asm
	{  cvt.f32.f16 %f1489, %rs63;}

	// inline asm
	// inline asm
	{  cvt.f32.f16 %f1490, %rs64;}

	// inline asm
	// inline asm
	{  cvt.f32.f16 %f1491, %rs65;}

	// inline asm
	// inline asm
	call (%rd105), _rt_buffer_get_64, (%rd100, %r27, %r187, %rd3, %rd4, %rd12, %rd12);
	// inline asm
	add.f32 	%f1492, %f382, %f1489;
	add.f32 	%f1493, %f383, %f1490;
	add.f32 	%f1494, %f384, %f1491;
	// inline asm
	{  cvt.rn.f16.f32 %rs62, %f1494;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs61, %f1493;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs60, %f1492;}

	// inline asm
	st.v4.u16 	[%rd105], {%rs60, %rs61, %rs62, %rs56};
	bra.uni 	BB0_105;

BB0_104:
	mov.u64 	%rd118, image_RNM1;
	cvta.global.u64 	%rd113, %rd118;
	mov.u32 	%r189, 8;
	// inline asm
	call (%rd112), _rt_buffer_get_64, (%rd113, %r27, %r189, %rd3, %rd4, %rd12, %rd12);
	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs69, %f384;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs68, %f383;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs67, %f382;}

	// inline asm
	st.v4.u16 	[%rd112], {%rs67, %rs68, %rs69, %rs56};

BB0_105:
	mul.f32 	%f385, %f1668, 0f3FB8EBD1;
	mul.f32 	%f386, %f1667, 0f3FB8EBD1;
	mul.f32 	%f387, %f1666, 0f3FB8EBD1;
	ld.global.u32 	%r190, [additive];
	setp.eq.s32	%p127, %r190, 0;
	// inline asm
	{  cvt.rn.f16.f32 %rs70, %f1478;}

	// inline asm
	@%p127 bra 	BB0_107;

	mov.u64 	%rd131, image_RNM2;
	cvta.global.u64 	%rd120, %rd131;
	mov.u32 	%r194, 8;
	// inline asm
	call (%rd119), _rt_buffer_get_64, (%rd120, %r27, %r194, %rd3, %rd4, %rd12, %rd12);
	// inline asm
	ld.v4.u16 	{%rs77, %rs78, %rs79, %rs80}, [%rd119];
	// inline asm
	{  cvt.f32.f16 %f1499, %rs77;}

	// inline asm
	// inline asm
	{  cvt.f32.f16 %f1500, %rs78;}

	// inline asm
	// inline asm
	{  cvt.f32.f16 %f1501, %rs79;}

	// inline asm
	// inline asm
	call (%rd125), _rt_buffer_get_64, (%rd120, %r27, %r194, %rd3, %rd4, %rd12, %rd12);
	// inline asm
	add.f32 	%f1502, %f385, %f1499;
	add.f32 	%f1503, %f386, %f1500;
	add.f32 	%f1504, %f387, %f1501;
	// inline asm
	{  cvt.rn.f16.f32 %rs76, %f1504;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs75, %f1503;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs74, %f1502;}

	// inline asm
	st.v4.u16 	[%rd125], {%rs74, %rs75, %rs76, %rs70};
	bra.uni 	BB0_108;

BB0_107:
	mov.u64 	%rd138, image_RNM2;
	cvta.global.u64 	%rd133, %rd138;
	mov.u32 	%r196, 8;
	// inline asm
	call (%rd132), _rt_buffer_get_64, (%rd133, %r27, %r196, %rd3, %rd4, %rd12, %rd12);
	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs83, %f387;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs82, %f386;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs81, %f385;}

	// inline asm
	st.v4.u16 	[%rd132], {%rs81, %rs82, %rs83, %rs70};

BB0_108:
	mul.f32 	%f388, %f1665, 0f3FB8EBD1;
	mul.f32 	%f389, %f1664, 0f3FB8EBD1;
	mul.f32 	%f390, %f1663, 0f3FB8EBD1;
	ld.global.u32 	%r197, [additive];
	setp.eq.s32	%p128, %r197, 0;
	// inline asm
	{  cvt.rn.f16.f32 %rs84, %f1478;}

	// inline asm
	@%p128 bra 	BB0_110;

	mov.u64 	%rd151, image_RNM3;
	cvta.global.u64 	%rd140, %rd151;
	mov.u32 	%r201, 8;
	// inline asm
	call (%rd139), _rt_buffer_get_64, (%rd140, %r27, %r201, %rd3, %rd4, %rd12, %rd12);
	// inline asm
	ld.v4.u16 	{%rs91, %rs92, %rs93, %rs94}, [%rd139];
	// inline asm
	{  cvt.f32.f16 %f1509, %rs91;}

	// inline asm
	// inline asm
	{  cvt.f32.f16 %f1510, %rs92;}

	// inline asm
	// inline asm
	{  cvt.f32.f16 %f1511, %rs93;}

	// inline asm
	// inline asm
	call (%rd145), _rt_buffer_get_64, (%rd140, %r27, %r201, %rd3, %rd4, %rd12, %rd12);
	// inline asm
	add.f32 	%f1512, %f388, %f1509;
	add.f32 	%f1513, %f389, %f1510;
	add.f32 	%f1514, %f390, %f1511;
	// inline asm
	{  cvt.rn.f16.f32 %rs90, %f1514;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs89, %f1513;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs88, %f1512;}

	// inline asm
	st.v4.u16 	[%rd145], {%rs88, %rs89, %rs90, %rs84};
	bra.uni 	BB0_111;

BB0_110:
	mov.u64 	%rd158, image_RNM3;
	cvta.global.u64 	%rd153, %rd158;
	mov.u32 	%r203, 8;
	// inline asm
	call (%rd152), _rt_buffer_get_64, (%rd153, %r27, %r203, %rd3, %rd4, %rd12, %rd12);
	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs97, %f390;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs96, %f389;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs95, %f388;}

	// inline asm
	st.v4.u16 	[%rd152], {%rs95, %rs96, %rs97, %rs84};

BB0_111:
	mul.f32 	%f391, %f1662, 0f404EBF87;
	mul.f32 	%f392, %f1661, 0f404EBF87;
	mul.f32 	%f393, %f1660, 0f404EBF87;
	ld.global.u32 	%r204, [additive];
	setp.eq.s32	%p129, %r204, 0;
	// inline asm
	{  cvt.rn.f16.f32 %rs98, %f1478;}

	// inline asm
	@%p129 bra 	BB0_113;

	mov.u64 	%rd171, image_RNM4;
	cvta.global.u64 	%rd160, %rd171;
	mov.u32 	%r208, 8;
	// inline asm
	call (%rd159), _rt_buffer_get_64, (%rd160, %r27, %r208, %rd3, %rd4, %rd12, %rd12);
	// inline asm
	ld.v4.u16 	{%rs105, %rs106, %rs107, %rs108}, [%rd159];
	// inline asm
	{  cvt.f32.f16 %f1519, %rs105;}

	// inline asm
	// inline asm
	{  cvt.f32.f16 %f1520, %rs106;}

	// inline asm
	// inline asm
	{  cvt.f32.f16 %f1521, %rs107;}

	// inline asm
	// inline asm
	call (%rd165), _rt_buffer_get_64, (%rd160, %r27, %r208, %rd3, %rd4, %rd12, %rd12);
	// inline asm
	add.f32 	%f1522, %f391, %f1519;
	add.f32 	%f1523, %f392, %f1520;
	add.f32 	%f1524, %f393, %f1521;
	// inline asm
	{  cvt.rn.f16.f32 %rs104, %f1524;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs103, %f1523;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs102, %f1522;}

	// inline asm
	st.v4.u16 	[%rd165], {%rs102, %rs103, %rs104, %rs98};
	bra.uni 	BB0_114;

BB0_113:
	mov.u64 	%rd178, image_RNM4;
	cvta.global.u64 	%rd173, %rd178;
	mov.u32 	%r210, 8;
	// inline asm
	call (%rd172), _rt_buffer_get_64, (%rd173, %r27, %r210, %rd3, %rd4, %rd12, %rd12);
	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs111, %f393;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs110, %f392;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs109, %f391;}

	// inline asm
	st.v4.u16 	[%rd172], {%rs109, %rs110, %rs111, %rs98};

BB0_114:
	mul.f32 	%f394, %f1659, 0f404EBF87;
	mul.f32 	%f395, %f1658, 0f404EBF87;
	mul.f32 	%f396, %f1657, 0f404EBF87;
	ld.global.u32 	%r211, [additive];
	setp.eq.s32	%p130, %r211, 0;
	// inline asm
	{  cvt.rn.f16.f32 %rs112, %f1478;}

	// inline asm
	@%p130 bra 	BB0_116;

	mov.u64 	%rd191, image_RNM5;
	cvta.global.u64 	%rd180, %rd191;
	mov.u32 	%r215, 8;
	// inline asm
	call (%rd179), _rt_buffer_get_64, (%rd180, %r27, %r215, %rd3, %rd4, %rd12, %rd12);
	// inline asm
	ld.v4.u16 	{%rs119, %rs120, %rs121, %rs122}, [%rd179];
	// inline asm
	{  cvt.f32.f16 %f1529, %rs119;}

	// inline asm
	// inline asm
	{  cvt.f32.f16 %f1530, %rs120;}

	// inline asm
	// inline asm
	{  cvt.f32.f16 %f1531, %rs121;}

	// inline asm
	// inline asm
	call (%rd185), _rt_buffer_get_64, (%rd180, %r27, %r215, %rd3, %rd4, %rd12, %rd12);
	// inline asm
	add.f32 	%f1532, %f394, %f1529;
	add.f32 	%f1533, %f395, %f1530;
	add.f32 	%f1534, %f396, %f1531;
	// inline asm
	{  cvt.rn.f16.f32 %rs118, %f1534;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs117, %f1533;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs116, %f1532;}

	// inline asm
	st.v4.u16 	[%rd185], {%rs116, %rs117, %rs118, %rs112};
	bra.uni 	BB0_117;

BB0_116:
	mov.u64 	%rd198, image_RNM5;
	cvta.global.u64 	%rd193, %rd198;
	mov.u32 	%r217, 8;
	// inline asm
	call (%rd192), _rt_buffer_get_64, (%rd193, %r27, %r217, %rd3, %rd4, %rd12, %rd12);
	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs125, %f396;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs124, %f395;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs123, %f394;}

	// inline asm
	st.v4.u16 	[%rd192], {%rs123, %rs124, %rs125, %rs112};

BB0_117:
	mul.f32 	%f397, %f1656, 0f40330C94;
	mul.f32 	%f398, %f1655, 0f40330C94;
	mul.f32 	%f399, %f1654, 0f40330C94;
	ld.global.u32 	%r218, [additive];
	setp.eq.s32	%p131, %r218, 0;
	// inline asm
	{  cvt.rn.f16.f32 %rs126, %f1478;}

	// inline asm
	@%p131 bra 	BB0_119;

	mov.u64 	%rd211, image_RNM6;
	cvta.global.u64 	%rd200, %rd211;
	mov.u32 	%r222, 8;
	// inline asm
	call (%rd199), _rt_buffer_get_64, (%rd200, %r27, %r222, %rd3, %rd4, %rd12, %rd12);
	// inline asm
	ld.v4.u16 	{%rs133, %rs134, %rs135, %rs136}, [%rd199];
	// inline asm
	{  cvt.f32.f16 %f1539, %rs133;}

	// inline asm
	// inline asm
	{  cvt.f32.f16 %f1540, %rs134;}

	// inline asm
	// inline asm
	{  cvt.f32.f16 %f1541, %rs135;}

	// inline asm
	// inline asm
	call (%rd205), _rt_buffer_get_64, (%rd200, %r27, %r222, %rd3, %rd4, %rd12, %rd12);
	// inline asm
	add.f32 	%f1542, %f397, %f1539;
	add.f32 	%f1543, %f398, %f1540;
	add.f32 	%f1544, %f399, %f1541;
	// inline asm
	{  cvt.rn.f16.f32 %rs132, %f1544;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs131, %f1543;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs130, %f1542;}

	// inline asm
	st.v4.u16 	[%rd205], {%rs130, %rs131, %rs132, %rs126};
	bra.uni 	BB0_120;

BB0_119:
	mov.u64 	%rd218, image_RNM6;
	cvta.global.u64 	%rd213, %rd218;
	mov.u32 	%r224, 8;
	// inline asm
	call (%rd212), _rt_buffer_get_64, (%rd213, %r27, %r224, %rd3, %rd4, %rd12, %rd12);
	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs139, %f399;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs138, %f398;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs137, %f397;}

	// inline asm
	st.v4.u16 	[%rd212], {%rs137, %rs138, %rs139, %rs126};

BB0_120:
	mul.f32 	%f400, %f1653, 0f404EBF87;
	mul.f32 	%f401, %f1652, 0f404EBF87;
	mul.f32 	%f402, %f1651, 0f404EBF87;
	ld.global.u32 	%r225, [additive];
	setp.eq.s32	%p132, %r225, 0;
	// inline asm
	{  cvt.rn.f16.f32 %rs140, %f1478;}

	// inline asm
	@%p132 bra 	BB0_122;

	mov.u64 	%rd231, image_RNM7;
	cvta.global.u64 	%rd220, %rd231;
	mov.u32 	%r229, 8;
	// inline asm
	call (%rd219), _rt_buffer_get_64, (%rd220, %r27, %r229, %rd3, %rd4, %rd12, %rd12);
	// inline asm
	ld.v4.u16 	{%rs147, %rs148, %rs149, %rs150}, [%rd219];
	// inline asm
	{  cvt.f32.f16 %f1549, %rs147;}

	// inline asm
	// inline asm
	{  cvt.f32.f16 %f1550, %rs148;}

	// inline asm
	// inline asm
	{  cvt.f32.f16 %f1551, %rs149;}

	// inline asm
	// inline asm
	call (%rd225), _rt_buffer_get_64, (%rd220, %r27, %r229, %rd3, %rd4, %rd12, %rd12);
	// inline asm
	add.f32 	%f1552, %f400, %f1549;
	add.f32 	%f1553, %f401, %f1550;
	add.f32 	%f1554, %f402, %f1551;
	// inline asm
	{  cvt.rn.f16.f32 %rs146, %f1554;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs145, %f1553;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs144, %f1552;}

	// inline asm
	st.v4.u16 	[%rd225], {%rs144, %rs145, %rs146, %rs140};
	bra.uni 	BB0_123;

BB0_122:
	mov.u64 	%rd238, image_RNM7;
	cvta.global.u64 	%rd233, %rd238;
	mov.u32 	%r231, 8;
	// inline asm
	call (%rd232), _rt_buffer_get_64, (%rd233, %r27, %r231, %rd3, %rd4, %rd12, %rd12);
	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs153, %f402;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs152, %f401;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs151, %f400;}

	// inline asm
	st.v4.u16 	[%rd232], {%rs151, %rs152, %rs153, %rs140};

BB0_123:
	mul.f32 	%f403, %f1650, 0f3FCEBF87;
	mul.f32 	%f404, %f1649, 0f3FCEBF87;
	mul.f32 	%f405, %f1648, 0f3FCEBF87;
	ld.global.u32 	%r232, [additive];
	setp.eq.s32	%p133, %r232, 0;
	// inline asm
	{  cvt.rn.f16.f32 %rs154, %f1478;}

	// inline asm
	@%p133 bra 	BB0_125;

	mov.u64 	%rd251, image_RNM8;
	cvta.global.u64 	%rd240, %rd251;
	mov.u32 	%r236, 8;
	// inline asm
	call (%rd239), _rt_buffer_get_64, (%rd240, %r27, %r236, %rd3, %rd4, %rd12, %rd12);
	// inline asm
	ld.v4.u16 	{%rs161, %rs162, %rs163, %rs164}, [%rd239];
	// inline asm
	{  cvt.f32.f16 %f1559, %rs161;}

	// inline asm
	// inline asm
	{  cvt.f32.f16 %f1560, %rs162;}

	// inline asm
	// inline asm
	{  cvt.f32.f16 %f1561, %rs163;}

	// inline asm
	// inline asm
	call (%rd245), _rt_buffer_get_64, (%rd240, %r27, %r236, %rd3, %rd4, %rd12, %rd12);
	// inline asm
	add.f32 	%f1562, %f403, %f1559;
	add.f32 	%f1563, %f404, %f1560;
	add.f32 	%f1564, %f405, %f1561;
	// inline asm
	{  cvt.rn.f16.f32 %rs160, %f1564;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs159, %f1563;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs158, %f1562;}

	// inline asm
	st.v4.u16 	[%rd245], {%rs158, %rs159, %rs160, %rs154};
	bra.uni 	BB0_161;

BB0_125:
	mov.u64 	%rd258, image_RNM8;
	cvta.global.u64 	%rd253, %rd258;
	mov.u32 	%r238, 8;
	// inline asm
	call (%rd252), _rt_buffer_get_64, (%rd253, %r27, %r238, %rd3, %rd4, %rd12, %rd12);
	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs167, %f405;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs166, %f404;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs165, %f403;}

	// inline asm
	st.v4.u16 	[%rd252], {%rs165, %rs166, %rs167, %rs154};

BB0_161:
	ret;
}