128ELi4EEvP6__halfPKS0_S3_PKjPKhS3_iiiiffbE12temp_storage has been demoted
.global .align 1 .b8 _ZN37_INTERNAL_36609967_7_gemv_cu_f89788766thrust6system6detail10sequential3seqE[1];
.global .align 1 .b8 _ZN37_INTERNAL_36609967_7_gemv_cu_f89788766thrust12placeholders2_1E[1];
.global .align 1 .b8 _ZN37_INTERNAL_36609967_7_gemv_cu_f89788766thrust12placeholders2_2E[1];
.global .align 1 .b8 _ZN37_INTERNAL_36609967_7_gemv_cu_f89788766thrust12placeholders2_3E[1];
.global .align 1 .b8 _ZN37_INTERNAL_36609967_7_gemv_cu_f89788766thrust12placeholders2_4E[1];
.global .align 1 .b8 _ZN37_INTERNAL_36609967_7_gemv_cu_f89788766thrust12placeholders2_5E[1];
.global .align 1 .b8 _ZN37_INTERNAL_36609967_7_gemv_cu_f89788766thrust12placeholders2_6E[1];
.global .align 1 .b8 _ZN37_INTERNAL_36609967_7_gemv_cu_f89788766thrust12placeholders2_7E[1];
.global .align 1 .b8 _ZN37_INTERNAL_36609967_7_gemv_cu_f89788766thrust12placeholders2_8E[1];
.global .align 1 .b8 _ZN37_INTERNAL_36609967_7_gemv_cu_f89788766thrust12placeholders2_9E[1];
.global .align 1 .b8 _ZN37_INTERNAL_36609967_7_gemv_cu_f89788766thrust12placeholders3_10E[1];

.visible .entry _Z28dequant_gemv_group128_batch423DequantGemvKernelParams(
	.param .align 8 .b8 _Z28dequant_gemv_group128_batch423DequantGemvKernelParams_param_0[80]
)
{
	.reg .pred 	%p<18>;
	.reg .b16 	%rs<526>;
	.reg .f32 	%f<852>;
	.reg .b32 	%r<399>;
	.reg .b64 	%rd<54>;
	// demoted variable
	.shared .align 16 .b8 _ZZ9gemv_int4ILi4ELi128ELi4EEvP6__halfPKS0_S3_PKjPKhS3_iiiiffbE12temp_storage[2560];

	ld.param.v2.u32 	{%r45, %r46}, [_Z28dequant_gemv_group128_batch423DequantGemvKernelParams_param_0+48];
	ld.param.v2.u32 	{%r47, %r48}, [_Z28dequant_gemv_group128_batch423DequantGemvKernelParams_param_0+56];
	ld.param.v2.f32 	{%f41, %f42}, [_Z28dequant_gemv_group128_batch423DequantGemvKernelParams_param_0+64];
	ld.param.v4.u8 	{%rs50, %rs51, %rs52, %rs53}, [_Z28dequant_gemv_group128_batch423DequantGemvKernelParams_param_0+72];
	ld.param.u64 	%rd23, [_Z28dequant_gemv_group128_batch423DequantGemvKernelParams_param_0+40];
	ld.param.u64 	%rd22, [_Z28dequant_gemv_group128_batch423DequantGemvKernelParams_param_0+32];
	ld.param.u64 	%rd21, [_Z28dequant_gemv_group128_batch423DequantGemvKernelParams_param_0+24];
	ld.param.u64 	%rd20, [_Z28dequant_gemv_group128_batch423DequantGemvKernelParams_param_0+16];
	mov.u32 	%r398, %tid.y;
	shl.b32 	%r49, %r398, 5;
	mov.u32 	%r50, %tid.x;
	add.s32 	%r397, %r49, %r50;
	shl.b32 	%r396, %r397, 2;
	setp.ge.u32 	%p1, %r396, %r47;
	mov.f32 	%f840, 0f00000000;
	mov.f32 	%f841, %f840;
	mov.f32 	%f842, %f840;
	mov.f32 	%f843, %f840;
	@%p1 bra 	$L__BB0_9;

	cvta.to.global.u64 	%rd2, %rd20;
	mov.u32 	%r51, %ctaid.x;
	mul.lo.s32 	%r6, %r48, %r51;
	shl.b16 	%rs2, %rs50, 3;
	cvta.to.global.u64 	%rd3, %rd21;
	cvta.to.global.u64 	%rd4, %rd23;

$L__BB0_2:
	mad.lo.s32 	%r53, %r47, %r51, %r396;
	mul.wide.u32 	%rd24, %r53, 4;
	add.s64 	%rd25, %rd3, %rd24;
	ld.global.v4.u32 	{%r54, %r55, %r56, %r57}, [%rd25];
	shr.u32 	%r59, %r50, 2;
	shl.b32 	%r60, %r398, 3;
	add.s32 	%r14, %r60, %r59;
	add.s32 	%r15, %r14, %r6;
	mul.wide.s32 	%rd26, %r15, 2;
	add.s64 	%rd27, %rd4, %rd26;
	ld.global.u16 	%rs58, [%rd27];
	// begin inline asm
	{  cvt.f32.f16 %f47, %rs58;}

	// end inline asm
	setp.eq.s64 	%p2, %rd22, 0;
	mov.u16 	%rs525, %rs2;
	@%p2 bra 	$L__BB0_4;

	shr.u32 	%r61, %r15, 31;
	add.s32 	%r62, %r15, %r61;
	shr.s32 	%r63, %r62, 1;
	cvt.s64.s32 	%rd28, %r63;
	cvta.to.global.u64 	%rd29, %rd22;
	add.s64 	%rd30, %rd29, %rd28;
	ld.global.u8 	%r64, [%rd30];
	shl.b32 	%r65, %r14, 2;
	and.b32  	%r66, %r65, 4;
	shr.u32 	%r67, %r64, %r66;
	cvt.u16.u32 	%rs59, %r67;
	and.b16  	%rs525, %rs59, 15;

$L__BB0_4:
	shl.b32 	%r16, %r397, 5;
	setp.ge.s32 	%p3, %r16, %r45;
	@%p3 bra 	$L__BB0_8;

	setp.eq.s16 	%p4, %rs50, 0;
	shr.u16 	%rs61, %rs525, 3;
	and.b16  	%rs62, %rs61, 1;
	setp.eq.b16 	%p5, %rs62, 1;
	and.pred  	%p6, %p4, %p5;
	selp.b16 	%rs63, -16, 0, %p6;
	or.b16  	%rs64, %rs63, %rs525;
	cvt.s16.s8 	%rs65, %rs64;
	cvt.rn.f32.s16 	%f6, %rs65;
	mul.wide.s32 	%rd31, %r16, 2;
	add.s64 	%rd5, %rd2, %rd31;
	ld.global.v4.u32 	{%r68, %r69, %r70, %r71}, [%rd5];
	mul.wide.s32 	%rd32, %r45, 2;
	add.s64 	%rd33, %rd5, %rd32;
	ld.global.v4.u32 	{%r72, %r73, %r74, %r75}, [%rd33];
	add.s32 	%r76, %r16, %r45;
	add.s32 	%r77, %r76, %r45;
	mul.wide.s32 	%rd34, %r77, 2;
	add.s64 	%rd6, %rd2, %rd34;
	ld.global.v4.u32 	{%r78, %r79, %r80, %r81}, [%rd6];
	add.s32 	%r82, %r77, %r45;
	mul.wide.s32 	%rd35, %r82, 2;
	add.s64 	%rd7, %rd2, %rd35;
	ld.global.v4.u32 	{%r83, %r84, %r85, %r86}, [%rd7];
	cvt.u16.u32 	%rs66, %r54;
	shr.u16 	%rs67, %rs66, 4;
	and.b16  	%rs5, %rs67, 15;
	mov.b32 	{%rs68, %rs6}, %r83;
	shr.u16 	%rs69, %rs66, 8;
	and.b16  	%rs7, %rs69, 15;
	shr.u16 	%rs8, %rs66, 12;
	shr.u32 	%r87, %r54, 16;
	cvt.u16.u32 	%rs70, %r87;
	and.b16  	%rs9, %rs70, 15;
	shr.u32 	%r88, %r54, 20;
	cvt.u16.u32 	%rs71, %r88;
	and.b16  	%rs10, %rs71, 15;
	shr.u32 	%r89, %r54, 24;
	cvt.u16.u32 	%rs72, %r89;
	and.b16  	%rs11, %rs72, 15;
	add.s32 	%r90, %r76, 8;
	mul.wide.s32 	%rd36, %r90, 2;
	add.s64 	%rd8, %rd2, %rd36;
	cvt.u16.u32 	%rs73, %r55;
	shr.u16 	%rs74, %rs73, 4;
	and.b16  	%rs12, %rs74, 15;
	shr.u16 	%rs75, %rs73, 8;
	and.b16  	%rs13, %rs75, 15;
	shr.u16 	%rs14, %rs73, 12;
	shr.u32 	%r91, %r55, 16;
	cvt.u16.u32 	%rs76, %r91;
	and.b16  	%rs15, %rs76, 15;
	shr.u32 	%r92, %r55, 20;
	cvt.u16.u32 	%rs77, %r92;
	and.b16  	%rs16, %rs77, 15;
	shr.u32 	%r93, %r55, 24;
	cvt.u16.u32 	%rs78, %r93;
	and.b16  	%rs17, %rs78, 15;
	cvt.u16.u32 	%rs79, %r56;
	shr.u16 	%rs80, %rs79, 4;
	and.b16  	%rs18, %rs80, 15;
	shr.u16 	%rs81, %rs79, 8;
	and.b16  	%rs19, %rs81, 15;
	shr.u16 	%rs20, %rs79, 12;
	shr.u32 	%r94, %r56, 16;
	cvt.u16.u32 	%rs21, %r94;
	and.b16  	%rs22, %rs21, 15;
	shr.u32 	%r95, %r56, 20;
	cvt.u16.u32 	%rs23, %r95;
	and.b16  	%rs24, %rs23, 15;
	shr.u32 	%r96, %r56, 24;
	cvt.u16.u32 	%rs25, %r96;
	and.b16  	%rs26, %rs25, 15;
	shr.u32 	%r97, %r56, 28;
	cvt.u16.u32 	%rs27, %r97;
	cvt.u16.u32 	%rs82, %r57;
	and.b16  	%rs28, %rs82, 15;
	shr.u32 	%r98, %r57, 4;
	cvt.u16.u32 	%rs29, %r98;
	and.b16  	%rs30, %rs29, 15;
	shr.u32 	%r99, %r57, 8;
	cvt.u16.u32 	%rs31, %r99;
	and.b16  	%rs32, %rs31, 15;
	shr.u32 	%r100, %r57, 12;
	cvt.u16.u32 	%rs33, %r100;
	and.b16  	%rs34, %rs33, 15;
	shr.u32 	%r101, %r57, 16;
	cvt.u16.u32 	%rs35, %r101;
	and.b16  	%rs36, %rs35, 15;
	shr.u32 	%r102, %r57, 20;
	cvt.u16.u32 	%rs37, %r102;
	and.b16  	%rs38, %rs37, 15;
	shr.u32 	%r103, %r57, 24;
	cvt.u16.u32 	%rs39, %r103;
	and.b16  	%rs40, %rs39, 15;
	shr.u32 	%r104, %r57, 28;
	cvt.u16.u32 	%rs41, %r104;
	@%p4 bra 	$L__BB0_7;

	and.b16  	%rs212, %rs66, 15;
	cvt.rn.f32.s16 	%f176, %rs212;
	sub.ftz.f32 	%f177, %f176, %f6;
	mul.ftz.f32 	%f178, %f47, %f177;
	mov.b32 	{%rs83, %rs87}, %r68;
	// begin inline asm
	{  cvt.f32.f16 %f48, %rs83;}

	// end inline asm
	fma.rn.ftz.f32 	%f179, %f178, %f48, %f843;
	mov.b32 	{%rs84, %rs88}, %r72;
	// begin inline asm
	{  cvt.f32.f16 %f49, %rs84;}

	// end inline asm
	fma.rn.ftz.f32 	%f180, %f178, %f49, %f842;
	mov.b32 	{%rs85, %rs89}, %r78;
	// begin inline asm
	{  cvt.f32.f16 %f50, %rs85;}

	// end inline asm
	fma.rn.ftz.f32 	%f181, %f178, %f50, %f841;
	cvt.u16.u32 	%rs86, %r83;
	// begin inline asm
	{  cvt.f32.f16 %f51, %rs86;}

	// end inline asm
	fma.rn.ftz.f32 	%f182, %f178, %f51, %f840;
	cvt.rn.f32.s16 	%f183, %rs5;
	sub.ftz.f32 	%f184, %f183, %f6;
	mul.ftz.f32 	%f185, %f47, %f184;
	// begin inline asm
	{  cvt.f32.f16 %f52, %rs87;}

	// end inline asm
	fma.rn.ftz.f32 	%f186, %f185, %f52, %f179;
	// begin inline asm
	{  cvt.f32.f16 %f53, %rs88;}

	// end inline asm
	fma.rn.ftz.f32 	%f187, %f185, %f53, %f180;
	// begin inline asm
	{  cvt.f32.f16 %f54, %rs89;}

	// end inline asm
	fma.rn.ftz.f32 	%f188, %f185, %f54, %f181;
	// begin inline asm
	{  cvt.f32.f16 %f55, %rs6;}

	// end inline asm
	fma.rn.ftz.f32 	%f189, %f185, %f55, %f182;
	cvt.rn.f32.s16 	%f190, %rs7;
	sub.ftz.f32 	%f191, %f190, %f6;
	mul.ftz.f32 	%f192, %f47, %f191;
	mov.b32 	{%rs91, %rs95}, %r69;
	// begin inline asm
	{  cvt.f32.f16 %f56, %rs91;}

	// end inline asm
	fma.rn.ftz.f32 	%f193, %f192, %f56, %f186;
	mov.b32 	{%rs92, %rs96}, %r73;
	// begin inline asm
	{  cvt.f32.f16 %f57, %rs92;}

	// end inline asm
	fma.rn.ftz.f32 	%f194, %f192, %f57, %f187;
	mov.b32 	{%rs93, %rs97}, %r79;
	// begin inline asm
	{  cvt.f32.f16 %f58, %rs93;}

	// end inline asm
	fma.rn.ftz.f32 	%f195, %f192, %f58, %f188;
	mov.b32 	{%rs94, %rs98}, %r84;
	// begin inline asm
	{  cvt.f32.f16 %f59, %rs94;}

	// end inline asm
	fma.rn.ftz.f32 	%f196, %f192, %f59, %f189;
	cvt.rn.f32.s16 	%f197, %rs8;
	sub.ftz.f32 	%f198, %f197, %f6;
	mul.ftz.f32 	%f199, %f47, %f198;
	// begin inline asm
	{  cvt.f32.f16 %f60, %rs95;}

	// end inline asm
	fma.rn.ftz.f32 	%f200, %f199, %f60, %f193;
	// begin inline asm
	{  cvt.f32.f16 %f61, %rs96;}

	// end inline asm
	fma.rn.ftz.f32 	%f201, %f199, %f61, %f194;
	// begin inline asm
	{  cvt.f32.f16 %f62, %rs97;}

	// end inline asm
	fma.rn.ftz.f32 	%f202, %f199, %f62, %f195;
	// begin inline asm
	{  cvt.f32.f16 %f63, %rs98;}

	// end inline asm
	fma.rn.ftz.f32 	%f203, %f199, %f63, %f196;
	cvt.rn.f32.s16 	%f204, %rs9;
	sub.ftz.f32 	%f205, %f204, %f6;
	mul.ftz.f32 	%f206, %f47, %f205;
	mov.b32 	{%rs99, %rs103}, %r70;
	// begin inline asm
	{  cvt.f32.f16 %f64, %rs99;}

	// end inline asm
	fma.rn.ftz.f32 	%f207, %f206, %f64, %f200;
	mov.b32 	{%rs100, %rs104}, %r74;
	// begin inline asm
	{  cvt.f32.f16 %f65, %rs100;}

	// end inline asm
	fma.rn.ftz.f32 	%f208, %f206, %f65, %f201;
	mov.b32 	{%rs101, %rs105}, %r80;
	// begin inline asm
	{  cvt.f32.f16 %f66, %rs101;}

	// end inline asm
	fma.rn.ftz.f32 	%f209, %f206, %f66, %f202;
	mov.b32 	{%rs102, %rs106}, %r85;
	// begin inline asm
	{  cvt.f32.f16 %f67, %rs102;}

	// end inline asm
	fma.rn.ftz.f32 	%f210, %f206, %f67, %f203;
	cvt.rn.f32.s16 	%f211, %rs10;
	sub.ftz.f32 	%f212, %f211, %f6;
	mul.ftz.f32 	%f213, %f47, %f212;
	// begin inline asm
	{  cvt.f32.f16 %f68, %rs103;}

	// end inline asm
	fma.rn.ftz.f32 	%f214, %f213, %f68, %f207;
	// begin inline asm
	{  cvt.f32.f16 %f69, %rs104;}

	// end inline asm
	fma.rn.ftz.f32 	%f215, %f213, %f69, %f208;
	// begin inline asm
	{  cvt.f32.f16 %f70, %rs105;}

	// end inline asm
	fma.rn.ftz.f32 	%f216, %f213, %f70, %f209;
	// begin inline asm
	{  cvt.f32.f16 %f71, %rs106;}

	// end inline asm
	fma.rn.ftz.f32 	%f217, %f213, %f71, %f210;
	cvt.rn.f32.s16 	%f218, %rs11;
	sub.ftz.f32 	%f219, %f218, %f6;
	mul.ftz.f32 	%f220, %f47, %f219;
	mov.b32 	{%rs107, %rs111}, %r71;
	// begin inline asm
	{  cvt.f32.f16 %f72, %rs107;}

	// end inline asm
	fma.rn.ftz.f32 	%f221, %f220, %f72, %f214;
	mov.b32 	{%rs108, %rs112}, %r75;
	// begin inline asm
	{  cvt.f32.f16 %f73, %rs108;}

	// end inline asm
	fma.rn.ftz.f32 	%f222, %f220, %f73, %f215;
	mov.b32 	{%rs109, %rs113}, %r81;
	// begin inline asm
	{  cvt.f32.f16 %f74, %rs109;}

	// end inline asm
	fma.rn.ftz.f32 	%f223, %f220, %f74, %f216;
	mov.b32 	{%rs110, %rs114}, %r86;
	// begin inline asm
	{  cvt.f32.f16 %f75, %rs110;}

	// end inline asm
	fma.rn.ftz.f32 	%f224, %f220, %f75, %f217;
	shr.u32 	%r105, %r54, 28;
	cvt.u16.u32 	%rs213, %r105;
	cvt.rn.f32.s16 	%f225, %rs213;
	sub.ftz.f32 	%f226, %f225, %f6;
	mul.ftz.f32 	%f227, %f47, %f226;
	// begin inline asm
	{  cvt.f32.f16 %f76, %rs111;}

	// end inline asm
	fma.rn.ftz.f32 	%f228, %f227, %f76, %f221;
	// begin inline asm
	{  cvt.f32.f16 %f77, %rs112;}

	// end inline asm
	fma.rn.ftz.f32 	%f229, %f227, %f77, %f222;
	// begin inline asm
	{  cvt.f32.f16 %f78, %rs113;}

	// end inline asm
	fma.rn.ftz.f32 	%f230, %f227, %f78, %f223;
	// begin inline asm
	{  cvt.f32.f16 %f79, %rs114;}

	// end inline asm
	fma.rn.ftz.f32 	%f231, %f227, %f79, %f224;
	ld.global.v4.u32 	{%r106, %r107, %r108, %r109}, [%rd5+16];
	ld.global.v4.u32 	{%r114, %r115, %r116, %r117}, [%rd8];
	ld.global.v4.u32 	{%r122, %r123, %r124, %r125}, [%rd6+16];
	ld.global.v4.u32 	{%r130, %r131, %r132, %r133}, [%rd7+16];
	and.b16  	%rs215, %rs73, 15;
	cvt.rn.f32.s16 	%f232, %rs215;
	sub.ftz.f32 	%f233, %f232, %f6;
	mul.ftz.f32 	%f234, %f47, %f233;
	mov.b32 	{%rs115, %rs119}, %r106;
	// begin inline asm
	{  cvt.f32.f16 %f80, %rs115;}

	// end inline asm
	fma.rn.ftz.f32 	%f235, %f234, %f80, %f228;
	mov.b32 	{%rs116, %rs120}, %r114;
	// begin inline asm
	{  cvt.f32.f16 %f81, %rs116;}

	// end inline asm
	fma.rn.ftz.f32 	%f236, %f234, %f81, %f229;
	mov.b32 	{%rs117, %rs121}, %r122;
	// begin inline asm
	{  cvt.f32.f16 %f82, %rs117;}

	// end inline asm
	fma.rn.ftz.f32 	%f237, %f234, %f82, %f230;
	mov.b32 	{%rs118, %rs122}, %r130;
	// begin inline asm
	{  cvt.f32.f16 %f83, %rs118;}

	// end inline asm
	fma.rn.ftz.f32 	%f238, %f234, %f83, %f231;
	cvt.rn.f32.s16 	%f239, %rs12;
	sub.ftz.f32 	%f240, %f239, %f6;
	mul.ftz.f32 	%f241, %f47, %f240;
	// begin inline asm
	{  cvt.f32.f16 %f84, %rs119;}

	// end inline asm
	fma.rn.ftz.f32 	%f242, %f241, %f84, %f235;
	// begin inline asm
	{  cvt.f32.f16 %f85, %rs120;}

	// end inline asm
	fma.rn.ftz.f32 	%f243, %f241, %f85, %f236;
	// begin inline asm
	{  cvt.f32.f16 %f86, %rs121;}

	// end inline asm
	fma.rn.ftz.f32 	%f244, %f241, %f86, %f237;
	// begin inline asm
	{  cvt.f32.f16 %f87, %rs122;}

	// end inline asm
	fma.rn.ftz.f32 	%f245, %f241, %f87, %f238;
	cvt.rn.f32.s16 	%f246, %rs13;
	sub.ftz.f32 	%f247, %f246, %f6;
	mul.ftz.f32 	%f248, %f47, %f247;
	mov.b32 	{%rs123, %rs127}, %r107;
	// begin inline asm
	{  cvt.f32.f16 %f88, %rs123;}

	// end inline asm
	fma.rn.ftz.f32 	%f249, %f248, %f88, %f242;
	mov.b32 	{%rs124, %rs128}, %r115;
	// begin inline asm
	{  cvt.f32.f16 %f89, %rs124;}

	// end inline asm
	fma.rn.ftz.f32 	%f250, %f248, %f89, %f243;
	mov.b32 	{%rs125, %rs129}, %r123;
	// begin inline asm
	{  cvt.f32.f16 %f90, %rs125;}

	// end inline asm
	fma.rn.ftz.f32 	%f251, %f248, %f90, %f244;
	mov.b32 	{%rs126, %rs130}, %r131;
	// begin inline asm
	{  cvt.f32.f16 %f91, %rs126;}

	// end inline asm
	fma.rn.ftz.f32 	%f252, %f248, %f91, %f245;
	cvt.rn.f32.s16 	%f253, %rs14;
	sub.ftz.f32 	%f254, %f253, %f6;
	mul.ftz.f32 	%f255, %f47, %f254;
	// begin inline asm
	{  cvt.f32.f16 %f92, %rs127;}

	// end inline asm
	fma.rn.ftz.f32 	%f256, %f255, %f92, %f249;
	// begin inline asm
	{  cvt.f32.f16 %f93, %rs128;}

	// end inline asm
	fma.rn.ftz.f32 	%f257, %f255, %f93, %f250;
	// begin inline asm
	{  cvt.f32.f16 %f94, %rs129;}

	// end inline asm
	fma.rn.ftz.f32 	%f258, %f255, %f94, %f251;
	// begin inline asm
	{  cvt.f32.f16 %f95, %rs130;}

	// end inline asm
	fma.rn.ftz.f32 	%f259, %f255, %f95, %f252;
	cvt.rn.f32.s16 	%f260, %rs15;
	sub.ftz.f32 	%f261, %f260, %f6;
	mul.ftz.f32 	%f262, %f47, %f261;
	mov.b32 	{%rs131, %rs135}, %r108;
	// begin inline asm
	{  cvt.f32.f16 %f96, %rs131;}

	// end inline asm
	fma.rn.ftz.f32 	%f263, %f262, %f96, %f256;
	mov.b32 	{%rs132, %rs136}, %r116;
	// begin inline asm
	{  cvt.f32.f16 %f97, %rs132;}

	// end inline asm
	fma.rn.ftz.f32 	%f264, %f262, %f97, %f257;
	mov.b32 	{%rs133, %rs137}, %r124;
	// begin inline asm
	{  cvt.f32.f16 %f98, %rs133;}

	// end inline asm
	fma.rn.ftz.f32 	%f265, %f262, %f98, %f258;
	mov.b32 	{%rs134, %rs138}, %r132;
	// begin inline asm
	{  cvt.f32.f16 %f99, %rs134;}

	// end inline asm
	fma.rn.ftz.f32 	%f266, %f262, %f99, %f259;
	cvt.rn.f32.s16 	%f267, %rs16;
	sub.ftz.f32 	%f268, %f267, %f6;
	mul.ftz.f32 	%f269, %f47, %f268;
	// begin inline asm
	{  cvt.f32.f16 %f100, %rs135;}

	// end inline asm
	fma.rn.ftz.f32 	%f270, %f269, %f100, %f263;
	// begin inline asm
	{  cvt.f32.f16 %f101, %rs136;}

	// end inline asm
	fma.rn.ftz.f32 	%f271, %f269, %f101, %f264;
	// begin inline asm
	{  cvt.f32.f16 %f102, %rs137;}

	// end inline asm
	fma.rn.ftz.f32 	%f272, %f269, %f102, %f265;
	// begin inline asm
	{  cvt.f32.f16 %f103, %rs138;}

	// end inline asm
	fma.rn.ftz.f32 	%f273, %f269, %f103, %f266;
	cvt.rn.f32.s16 	%f274, %rs17;
	sub.ftz.f32 	%f275, %f274, %f6;
	mul.ftz.f32 	%f276, %f47, %f275;
	mov.b32 	{%rs139, %rs143}, %r109;
	// begin inline asm
	{  cvt.f32.f16 %f104, %rs139;}

	// end inline asm
	fma.rn.ftz.f32 	%f277, %f276, %f104, %f270;
	mov.b32 	{%rs140, %rs144}, %r117;
	// begin inline asm
	{  cvt.f32.f16 %f105, %rs140;}

	// end inline asm
	fma.rn.ftz.f32 	%f278, %f276, %f105, %f271;
	mov.b32 	{%rs141, %rs145}, %r125;
	// begin inline asm
	{  cvt.f32.f16 %f106, %rs141;}

	// end inline asm
	fma.rn.ftz.f32 	%f279, %f276, %f106, %f272;
	mov.b32 	{%rs142, %rs146}, %r133;
	// begin inline asm
	{  cvt.f32.f16 %f107, %rs142;}

	// end inline asm
	fma.rn.ftz.f32 	%f280, %f276, %f107, %f273;
	shr.u32 	%r138, %r55, 28;
	cvt.u16.u32 	%rs216, %r138;
	cvt.rn.f32.s16 	%f281, %rs216;
	sub.ftz.f32 	%f282, %f281, %f6;
	mul.ftz.f32 	%f283, %f47, %f282;
	// begin inline asm
	{  cvt.f32.f16 %f108, %rs143;}

	// end inline asm
	fma.rn.ftz.f32 	%f284, %f283, %f108, %f277;
	// begin inline asm
	{  cvt.f32.f16 %f109, %rs144;}

	// end inline asm
	fma.rn.ftz.f32 	%f285, %f283, %f109, %f278;
	// begin inline asm
	{  cvt.f32.f16 %f110, %rs145;}

	// end inline asm
	fma.rn.ftz.f32 	%f286, %f283, %f110, %f279;
	// begin inline asm
	{  cvt.f32.f16 %f111, %rs146;}

	// end inline asm
	fma.rn.ftz.f32 	%f287, %f283, %f111, %f280;
	ld.global.v4.u32 	{%r139, %r140, %r141, %r142}, [%rd5+32];
	ld.global.v4.u32 	{%r147, %r148, %r149, %r150}, [%rd8+16];
	ld.global.v4.u32 	{%r155, %r156, %r157, %r158}, [%rd6+32];
	ld.global.v4.u32 	{%r163, %r164, %r165, %r166}, [%rd7+32];
	and.b16  	%rs218, %rs79, 15;
	cvt.rn.f32.s16 	%f288, %rs218;
	sub.ftz.f32 	%f289, %f288, %f6;
	mul.ftz.f32 	%f290, %f47, %f289;
	mov.b32 	{%rs147, %rs151}, %r139;
	// begin inline asm
	{  cvt.f32.f16 %f112, %rs147;}

	// end inline asm
	fma.rn.ftz.f32 	%f291, %f290, %f112, %f284;
	mov.b32 	{%rs148, %rs152}, %r147;
	// begin inline asm
	{  cvt.f32.f16 %f113, %rs148;}

	// end inline asm
	fma.rn.ftz.f32 	%f292, %f290, %f113, %f285;
	mov.b32 	{%rs149, %rs153}, %r155;
	// begin inline asm
	{  cvt.f32.f16 %f114, %rs149;}

	// end inline asm
	fma.rn.ftz.f32 	%f293, %f290, %f114, %f286;
	mov.b32 	{%rs150, %rs154}, %r163;
	// begin inline asm
	{  cvt.f32.f16 %f115, %rs150;}

	// end inline asm
	fma.rn.ftz.f32 	%f294, %f290, %f115, %f287;
	cvt.rn.f32.s16 	%f295, %rs18;
	sub.ftz.f32 	%f296, %f295, %f6;
	mul.ftz.f32 	%f297, %f47, %f296;
	// begin inline asm
	{  cvt.f32.f16 %f116, %rs151;}

	// end inline asm
	fma.rn.ftz.f32 	%f298, %f297, %f116, %f291;
	// begin inline asm
	{  cvt.f32.f16 %f117, %rs152;}

	// end inline asm
	fma.rn.ftz.f32 	%f299, %f297, %f117, %f292;
	// begin inline asm
	{  cvt.f32.f16 %f118, %rs153;}

	// end inline asm
	fma.rn.ftz.f32 	%f300, %f297, %f118, %f293;
	// begin inline asm
	{  cvt.f32.f16 %f119, %rs154;}

	// end inline asm
	fma.rn.ftz.f32 	%f301, %f297, %f119, %f294;
	cvt.rn.f32.s16 	%f302, %rs19;
	sub.ftz.f32 	%f303, %f302, %f6;
	mul.ftz.f32 	%f304, %f47, %f303;
	mov.b32 	{%rs155, %rs159}, %r140;
	// begin inline asm
	{  cvt.f32.f16 %f120, %rs155;}

	// end inline asm
	fma.rn.ftz.f32 	%f305, %f304, %f120, %f298;
	mov.b32 	{%rs156, %rs160}, %r148;
	// begin inline asm
	{  cvt.f32.f16 %f121, %rs156;}

	// end inline asm
	fma.rn.ftz.f32 	%f306, %f304, %f121, %f299;
	mov.b32 	{%rs157, %rs161}, %r156;
	// begin inline asm
	{  cvt.f32.f16 %f122, %rs157;}

	// end inline asm
	fma.rn.ftz.f32 	%f307, %f304, %f122, %f300;
	mov.b32 	{%rs158, %rs162}, %r164;
	// begin inline asm
	{  cvt.f32.f16 %f123, %rs158;}

	// end inline asm
	fma.rn.ftz.f32 	%f308, %f304, %f123, %f301;
	cvt.rn.f32.s16 	%f309, %rs20;
	sub.ftz.f32 	%f310, %f309, %f6;
	mul.ftz.f32 	%f311, %f47, %f310;
	// begin inline asm
	{  cvt.f32.f16 %f124, %rs159;}

	// end inline asm
	fma.rn.ftz.f32 	%f312, %f311, %f124, %f305;
	// begin inline asm
	{  cvt.f32.f16 %f125, %rs160;}

	// end inline asm
	fma.rn.ftz.f32 	%f313, %f311, %f125, %f306;
	// begin inline asm
	{  cvt.f32.f16 %f126, %rs161;}

	// end inline asm
	fma.rn.ftz.f32 	%f314, %f311, %f126, %f307;
	// begin inline asm
	{  cvt.f32.f16 %f127, %rs162;}

	// end inline asm
	fma.rn.ftz.f32 	%f315, %f311, %f127, %f308;
	cvt.rn.f32.s16 	%f316, %rs22;
	sub.ftz.f32 	%f317, %f316, %f6;
	mul.ftz.f32 	%f318, %f47, %f317;
	mov.b32 	{%rs163, %rs167}, %r141;
	// begin inline asm
	{  cvt.f32.f16 %f128, %rs163;}

	// end inline asm
	fma.rn.ftz.f32 	%f319, %f318, %f128, %f312;
	mov.b32 	{%rs164, %rs168}, %r149;
	// begin inline asm
	{  cvt.f32.f16 %f129, %rs164;}

	// end inline asm
	fma.rn.ftz.f32 	%f320, %f318, %f129, %f313;
	mov.b32 	{%rs165, %rs169}, %r157;
	// begin inline asm
	{  cvt.f32.f16 %f130, %rs165;}

	// end inline asm
	fma.rn.ftz.f32 	%f321, %f318, %f130, %f314;
	mov.b32 	{%rs166, %rs170}, %r165;
	// begin inline asm
	{  cvt.f32.f16 %f131, %rs166;}

	// end inline asm
	fma.rn.ftz.f32 	%f322, %f318, %f131, %f315;
	cvt.rn.f32.s16 	%f323, %rs24;
	sub.ftz.f32 	%f324, %f323, %f6;
	mul.ftz.f32 	%f325, %f47, %f324;
	// begin inline asm
	{  cvt.f32.f16 %f132, %rs167;}

	// end inline asm
	fma.rn.ftz.f32 	%f326, %f325, %f132, %f319;
	// begin inline asm
	{  cvt.f32.f16 %f133, %rs168;}

	// end inline asm
	fma.rn.ftz.f32 	%f327, %f325, %f133, %f320;
	// begin inline asm
	{  cvt.f32.f16 %f134, %rs169;}

	// end inline asm
	fma.rn.ftz.f32 	%f328, %f325, %f134, %f321;
	// begin inline asm
	{  cvt.f32.f16 %f135, %rs170;}

	// end inline asm
	fma.rn.ftz.f32 	%f329, %f325, %f135, %f322;
	cvt.rn.f32.s16 	%f330, %rs26;
	sub.ftz.f32 	%f331, %f330, %f6;
	mul.ftz.f32 	%f332, %f47, %f331;
	mov.b32 	{%rs171, %rs175}, %r142;
	// begin inline asm
	{  cvt.f32.f16 %f136, %rs171;}

	// end inline asm
	fma.rn.ftz.f32 	%f333, %f332, %f136, %f326;
	mov.b32 	{%rs172, %rs176}, %r150;
	// begin inline asm
	{  cvt.f32.f16 %f137, %rs172;}

	// end inline asm
	fma.rn.ftz.f32 	%f334, %f332, %f137, %f327;
	mov.b32 	{%rs173, %rs177}, %r158;
	// begin inline asm
	{  cvt.f32.f16 %f138, %rs173;}

	// end inline asm
	fma.rn.ftz.f32 	%f335, %f332, %f138, %f328;
	mov.b32 	{%rs174, %rs178}, %r166;
	// begin inline asm
	{  cvt.f32.f16 %f139, %rs174;}

	// end inline asm
	fma.rn.ftz.f32 	%f336, %f332, %f139, %f329;
	cvt.rn.f32.s16 	%f337, %rs27;
	sub.ftz.f32 	%f338, %f337, %f6;
	mul.ftz.f32 	%f339, %f47, %f338;
	// begin inline asm
	{  cvt.f32.f16 %f140, %rs175;}

	// end inline asm
	fma.rn.ftz.f32 	%f340, %f339, %f140, %f333;
	// begin inline asm
	{  cvt.f32.f16 %f141, %rs176;}

	// end inline asm
	fma.rn.ftz.f32 	%f341, %f339, %f141, %f334;
	// begin inline asm
	{  cvt.f32.f16 %f142, %rs177;}

	// end inline asm
	fma.rn.ftz.f32 	%f342, %f339, %f142, %f335;
	// begin inline asm
	{  cvt.f32.f16 %f143, %rs178;}

	// end inline asm
	fma.rn.ftz.f32 	%f343, %f339, %f143, %f336;
	ld.global.v4.u32 	{%r171, %r172, %r173, %r174}, [%rd5+48];
	ld.global.v4.u32 	{%r179, %r180, %r181, %r182}, [%rd8+32];
	ld.global.v4.u32 	{%r187, %r188, %r189, %r190}, [%rd6+48];
	ld.global.v4.u32 	{%r195, %r196, %r197, %r198}, [%rd7+48];
	cvt.rn.f32.s16 	%f344, %rs28;
	sub.ftz.f32 	%f345, %f344, %f6;
	mul.ftz.f32 	%f346, %f47, %f345;
	mov.b32 	{%rs179, %rs183}, %r171;
	// begin inline asm
	{  cvt.f32.f16 %f144, %rs179;}

	// end inline asm
	fma.rn.ftz.f32 	%f347, %f346, %f144, %f340;
	mov.b32 	{%rs180, %rs184}, %r179;
	// begin inline asm
	{  cvt.f32.f16 %f145, %rs180;}

	// end inline asm
	fma.rn.ftz.f32 	%f348, %f346, %f145, %f341;
	mov.b32 	{%rs181, %rs185}, %r187;
	// begin inline asm
	{  cvt.f32.f16 %f146, %rs181;}

	// end inline asm
	fma.rn.ftz.f32 	%f349, %f346, %f146, %f342;
	mov.b32 	{%rs182, %rs186}, %r195;
	// begin inline asm
	{  cvt.f32.f16 %f147, %rs182;}

	// end inline asm
	fma.rn.ftz.f32 	%f350, %f346, %f147, %f343;
	cvt.rn.f32.s16 	%f351, %rs30;
	sub.ftz.f32 	%f352, %f351, %f6;
	mul.ftz.f32 	%f353, %f47, %f352;
	// begin inline asm
	{  cvt.f32.f16 %f148, %rs183;}

	// end inline asm
	fma.rn.ftz.f32 	%f354, %f353, %f148, %f347;
	// begin inline asm
	{  cvt.f32.f16 %f149, %rs184;}

	// end inline asm
	fma.rn.ftz.f32 	%f355, %f353, %f149, %f348;
	// begin inline asm
	{  cvt.f32.f16 %f150, %rs185;}

	// end inline asm
	fma.rn.ftz.f32 	%f356, %f353, %f150, %f349;
	// begin inline asm
	{  cvt.f32.f16 %f151, %rs186;}

	// end inline asm
	fma.rn.ftz.f32 	%f357, %f353, %f151, %f350;
	cvt.rn.f32.s16 	%f358, %rs32;
	sub.ftz.f32 	%f359, %f358, %f6;
	mul.ftz.f32 	%f360, %f47, %f359;
	mov.b32 	{%rs187, %rs191}, %r172;
	// begin inline asm
	{  cvt.f32.f16 %f152, %rs187;}

	// end inline asm
	fma.rn.ftz.f32 	%f361, %f360, %f152, %f354;
	mov.b32 	{%rs188, %rs192}, %r180;
	// begin inline asm
	{  cvt.f32.f16 %f153, %rs188;}

	// end inline asm
	fma.rn.ftz.f32 	%f362, %f360, %f153, %f355;
	mov.b32 	{%rs189, %rs193}, %r188;
	// begin inline asm
	{  cvt.f32.f16 %f154, %rs189;}

	// end inline asm
	fma.rn.ftz.f32 	%f363, %f360, %f154, %f356;
	mov.b32 	{%rs190, %rs194}, %r196;
	// begin inline asm
	{  cvt.f32.f16 %f155, %rs190;}

	// end inline asm
	fma.rn.ftz.f32 	%f364, %f360, %f155, %f357;
	cvt.rn.f32.s16 	%f365, %rs34;
	sub.ftz.f32 	%f366, %f365, %f6;
	mul.ftz.f32 	%f367, %f47, %f366;
	// begin inline asm
	{  cvt.f32.f16 %f156, %rs191;}

	// end inline asm
	fma.rn.ftz.f32 	%f368, %f367, %f156, %f361;
	// begin inline asm
	{  cvt.f32.f16 %f157, %rs192;}

	// end inline asm
	fma.rn.ftz.f32 	%f369, %f367, %f157, %f362;
	// begin inline asm
	{  cvt.f32.f16 %f158, %rs193;}

	// end inline asm
	fma.rn.ftz.f32 	%f370, %f367, %f158, %f363;
	// begin inline asm
	{  cvt.f32.f16 %f159, %rs194;}

	// end inline asm
	fma.rn.ftz.f32 	%f371, %f367, %f159, %f364;
	cvt.rn.f32.s16 	%f372, %rs36;
	sub.ftz.f32 	%f373, %f372, %f6;
	mul.ftz.f32 	%f374, %f47, %f373;
	mov.b32 	{%rs195, %rs199}, %r173;
	// begin inline asm
	{  cvt.f32.f16 %f160, %rs195;}

	// end inline asm
	fma.rn.ftz.f32 	%f375, %f374, %f160, %f368;
	mov.b32 	{%rs196, %rs200}, %r181;
	// begin inline asm
	{  cvt.f32.f16 %f161, %rs196;}

	// end inline asm
	fma.rn.ftz.f32 	%f376, %f374, %f161, %f369;
	mov.b32 	{%rs197, %rs201}, %r189;
	// begin inline asm
	{  cvt.f32.f16 %f162, %rs197;}

	// end inline asm
	fma.rn.ftz.f32 	%f377, %f374, %f162, %f370;
	mov.b32 	{%rs198, %rs202}, %r197;
	// begin inline asm
	{  cvt.f32.f16 %f163, %rs198;}

	// end inline asm
	fma.rn.ftz.f32 	%f378, %f374, %f163, %f371;
	cvt.rn.f32.s16 	%f379, %rs38;
	sub.ftz.f32 	%f380, %f379, %f6;
	mul.ftz.f32 	%f381, %f47, %f380;
	// begin inline asm
	{  cvt.f32.f16 %f164, %rs199;}

	// end inline asm
	fma.rn.ftz.f32 	%f382, %f381, %f164, %f375;
	// begin inline asm
	{  cvt.f32.f16 %f165, %rs200;}

	// end inline asm
	fma.rn.ftz.f32 	%f383, %f381, %f165, %f376;
	// begin inline asm
	{  cvt.f32.f16 %f166, %rs201;}

	// end inline asm
	fma.rn.ftz.f32 	%f384, %f381, %f166, %f377;
	// begin inline asm
	{  cvt.f32.f16 %f167, %rs202;}

	// end inline asm
	fma.rn.ftz.f32 	%f385, %f381, %f167, %f378;
	cvt.rn.f32.s16 	%f386, %rs40;
	sub.ftz.f32 	%f387, %f386, %f6;
	mul.ftz.f32 	%f388, %f47, %f387;
	mov.b32 	{%rs203, %rs207}, %r174;
	// begin inline asm
	{  cvt.f32.f16 %f168, %rs203;}

	// end inline asm
	fma.rn.ftz.f32 	%f389, %f388, %f168, %f382;
	mov.b32 	{%rs204, %rs208}, %r182;
	// begin inline asm
	{  cvt.f32.f16 %f169, %rs204;}

	// end inline asm
	fma.rn.ftz.f32 	%f390, %f388, %f169, %f383;
	mov.b32 	{%rs205, %rs209}, %r190;
	// begin inline asm
	{  cvt.f32.f16 %f170, %rs205;}

	// end inline asm
	fma.rn.ftz.f32 	%f391, %f388, %f170, %f384;
	mov.b32 	{%rs206, %rs210}, %r198;
	// begin inline asm
	{  cvt.f32.f16 %f171, %rs206;}

	// end inline asm
	fma.rn.ftz.f32 	%f392, %f388, %f171, %f385;
	cvt.rn.f32.s16 	%f393, %rs41;
	sub.ftz.f32 	%f394, %f393, %f6;
	mul.ftz.f32 	%f395, %f47, %f394;
	// begin inline asm
	{  cvt.f32.f16 %f172, %rs207;}

	// end inline asm
	fma.rn.ftz.f32 	%f843, %f395, %f172, %f389;
	// begin inline asm
	{  cvt.f32.f16 %f173, %rs208;}

	// end inline asm
	fma.rn.ftz.f32 	%f842, %f395, %f173, %f390;
	// begin inline asm
	{  cvt.f32.f16 %f174, %rs209;}

	// end inline asm
	fma.rn.ftz.f32 	%f841, %f395, %f174, %f391;
	// begin inline asm
	{  cvt.f32.f16 %f175, %rs210;}

	// end inline asm
	fma.rn.ftz.f32 	%f840, %f395, %f175, %f392;
	bra.uni 	$L__BB0_8;

$L__BB0_7:
	shl.b16 	%rs348, %rs66, 4;
	cvt.s16.s8 	%rs349, %rs348;
	shr.s16 	%rs350, %rs349, 7;
	and.b16  	%rs351, %rs350, -16;
	and.b16  	%rs352, %rs66, 15;
	or.b16  	%rs353, %rs351, %rs352;
	cvt.rn.f32.s16 	%f524, %rs353;
	sub.ftz.f32 	%f525, %f524, %f6;
	mul.ftz.f32 	%f526, %f47, %f525;
	mov.b32 	{%rs219, %rs223}, %r68;
	// begin inline asm
	{  cvt.f32.f16 %f396, %rs219;}

	// end inline asm
	fma.rn.ftz.f32 	%f527, %f526, %f396, %f843;
	mov.b32 	{%rs220, %rs224}, %r72;
	// begin inline asm
	{  cvt.f32.f16 %f397, %rs220;}

	// end inline asm
	fma.rn.ftz.f32 	%f528, %f526, %f397, %f842;
	mov.b32 	{%rs221, %rs225}, %r78;
	// begin inline asm
	{  cvt.f32.f16 %f398, %rs221;}

	// end inline asm
	fma.rn.ftz.f32 	%f529, %f526, %f398, %f841;
	cvt.u16.u32 	%rs222, %r83;
	// begin inline asm
	{  cvt.f32.f16 %f399, %rs222;}

	// end inline asm
	fma.rn.ftz.f32 	%f530, %f526, %f399, %f840;
	cvt.s16.s8 	%rs354, %rs66;
	shr.s16 	%rs355, %rs354, 7;
	and.b16  	%rs356, %rs355, -16;
	or.b16  	%rs357, %rs356, %rs5;
	cvt.rn.f32.s16 	%f531, %rs357;
	sub.ftz.f32 	%f532, %f531, %f6;
	mul.ftz.f32 	%f533, %f47, %f532;
	// begin inline asm
	{  cvt.f32.f16 %f400, %rs223;}

	// end inline asm
	fma.rn.ftz.f32 	%f534, %f533, %f400, %f527;
	// begin inline asm
	{  cvt.f32.f16 %f401, %rs224;}

	// end inline asm
	fma.rn.ftz.f32 	%f535, %f533, %f401, %f528;
	// begin inline asm
	{  cvt.f32.f16 %f402, %rs225;}

	// end inline asm
	fma.rn.ftz.f32 	%f536, %f533, %f402, %f529;
	// begin inline asm
	{  cvt.f32.f16 %f403, %rs6;}

	// end inline asm
	fma.rn.ftz.f32 	%f537, %f533, %f403, %f530;
	cvt.s16.s8 	%rs359, %rs67;
	shr.s16 	%rs360, %rs359, 7;
	and.b16  	%rs361, %rs360, -16;
	or.b16  	%rs362, %rs361, %rs7;
	cvt.rn.f32.s16 	%f538, %rs362;
	sub.ftz.f32 	%f539, %f538, %f6;
	mul.ftz.f32 	%f540, %f47, %f539;
	mov.b32 	{%rs227, %rs231}, %r69;
	// begin inline asm
	{  cvt.f32.f16 %f404, %rs227;}

	// end inline asm
	fma.rn.ftz.f32 	%f541, %f540, %f404, %f534;
	mov.b32 	{%rs228, %rs232}, %r73;
	// begin inline asm
	{  cvt.f32.f16 %f405, %rs228;}

	// end inline asm
	fma.rn.ftz.f32 	%f542, %f540, %f405, %f535;
	mov.b32 	{%rs229, %rs233}, %r79;
	// begin inline asm
	{  cvt.f32.f16 %f406, %rs229;}

	// end inline asm
	fma.rn.ftz.f32 	%f543, %f540, %f406, %f536;
	mov.b32 	{%rs230, %rs234}, %r84;
	// begin inline asm
	{  cvt.f32.f16 %f407, %rs230;}

	// end inline asm
	fma.rn.ftz.f32 	%f544, %f540, %f407, %f537;
	shr.s16 	%rs363, %rs66, 15;
	and.b16  	%rs364, %rs363, -16;
	or.b16  	%rs365, %rs364, %rs8;
	cvt.rn.f32.s16 	%f545, %rs365;
	sub.ftz.f32 	%f546, %f545, %f6;
	mul.ftz.f32 	%f547, %f47, %f546;
	// begin inline asm
	{  cvt.f32.f16 %f408, %rs231;}

	// end inline asm
	fma.rn.ftz.f32 	%f548, %f547, %f408, %f541;
	// begin inline asm
	{  cvt.f32.f16 %f409, %rs232;}

	// end inline asm
	fma.rn.ftz.f32 	%f549, %f547, %f409, %f542;
	// begin inline asm
	{  cvt.f32.f16 %f410, %rs233;}

	// end inline asm
	fma.rn.ftz.f32 	%f550, %f547, %f410, %f543;
	// begin inline asm
	{  cvt.f32.f16 %f411, %rs234;}

	// end inline asm
	fma.rn.ftz.f32 	%f551, %f547, %f411, %f544;
	shl.b16 	%rs367, %rs70, 4;
	cvt.s16.s8 	%rs368, %rs367;
	shr.s16 	%rs369, %rs368, 7;
	and.b16  	%rs370, %rs369, -16;
	or.b16  	%rs371, %rs370, %rs9;
	cvt.rn.f32.s16 	%f552, %rs371;
	sub.ftz.f32 	%f553, %f552, %f6;
	mul.ftz.f32 	%f554, %f47, %f553;
	mov.b32 	{%rs235, %rs239}, %r70;
	// begin inline asm
	{  cvt.f32.f16 %f412, %rs235;}

	// end inline asm
	fma.rn.ftz.f32 	%f555, %f554, %f412, %f548;
	mov.b32 	{%rs236, %rs240}, %r74;
	// begin inline asm
	{  cvt.f32.f16 %f413, %rs236;}

	// end inline asm
	fma.rn.ftz.f32 	%f556, %f554, %f413, %f549;
	mov.b32 	{%rs237, %rs241}, %r80;
	// begin inline asm
	{  cvt.f32.f16 %f414, %rs237;}

	// end inline asm
	fma.rn.ftz.f32 	%f557, %f554, %f414, %f550;
	mov.b32 	{%rs238, %rs242}, %r85;
	// begin inline asm
	{  cvt.f32.f16 %f415, %rs238;}

	// end inline asm
	fma.rn.ftz.f32 	%f558, %f554, %f415, %f551;
	shl.b16 	%rs373, %rs71, 4;
	cvt.s16.s8 	%rs374, %rs373;
	shr.s16 	%rs375, %rs374, 7;
	and.b16  	%rs376, %rs375, -16;
	or.b16  	%rs377, %rs376, %rs10;
	cvt.rn.f32.s16 	%f559, %rs377;
	sub.ftz.f32 	%f560, %f559, %f6;
	mul.ftz.f32 	%f561, %f47, %f560;
	// begin inline asm
	{  cvt.f32.f16 %f416, %rs239;}

	// end inline asm
	fma.rn.ftz.f32 	%f562, %f561, %f416, %f555;
	// begin inline asm
	{  cvt.f32.f16 %f417, %rs240;}

	// end inline asm
	fma.rn.ftz.f32 	%f563, %f561, %f417, %f556;
	// begin inline asm
	{  cvt.f32.f16 %f418, %rs241;}

	// end inline asm
	fma.rn.ftz.f32 	%f564, %f561, %f418, %f557;
	// begin inline asm
	{  cvt.f32.f16 %f419, %rs242;}

	// end inline asm
	fma.rn.ftz.f32 	%f565, %f561, %f419, %f558;
	shl.b16 	%rs379, %rs72, 4;
	cvt.s16.s8 	%rs380, %rs379;
	shr.s16 	%rs381, %rs380, 7;
	and.b16  	%rs382, %rs381, -16;
	or.b16  	%rs383, %rs382, %rs11;
	cvt.rn.f32.s16 	%f566, %rs383;
	sub.ftz.f32 	%f567, %f566, %f6;
	mul.ftz.f32 	%f568, %f47, %f567;
	mov.b32 	{%rs243, %rs247}, %r71;
	// begin inline asm
	{  cvt.f32.f16 %f420, %rs243;}

	// end inline asm
	fma.rn.ftz.f32 	%f569, %f568, %f420, %f562;
	mov.b32 	{%rs244, %rs248}, %r75;
	// begin inline asm
	{  cvt.f32.f16 %f421, %rs244;}

	// end inline asm
	fma.rn.ftz.f32 	%f570, %f568, %f421, %f563;
	mov.b32 	{%rs245, %rs249}, %r81;
	// begin inline asm
	{  cvt.f32.f16 %f422, %rs245;}

	// end inline asm
	fma.rn.ftz.f32 	%f571, %f568, %f422, %f564;
	mov.b32 	{%rs246, %rs250}, %r86;
	// begin inline asm
	{  cvt.f32.f16 %f423, %rs246;}

	// end inline asm
	fma.rn.ftz.f32 	%f572, %f568, %f423, %f565;
	shr.u32 	%r206, %r54, 28;
	cvt.u16.u32 	%rs384, %r206;
	shl.b16 	%rs385, %rs384, 4;
	cvt.s16.s8 	%rs386, %rs385;
	shr.s16 	%rs387, %rs386, 7;
	and.b16  	%rs388, %rs387, -16;
	or.b16  	%rs389, %rs388, %rs384;
	cvt.rn.f32.s16 	%f573, %rs389;
	sub.ftz.f32 	%f574, %f573, %f6;
	mul.ftz.f32 	%f575, %f47, %f574;
	// begin inline asm
	{  cvt.f32.f16 %f424, %rs247;}

	// end inline asm
	fma.rn.ftz.f32 	%f576, %f575, %f424, %f569;
	// begin inline asm
	{  cvt.f32.f16 %f425, %rs248;}

	// end inline asm
	fma.rn.ftz.f32 	%f577, %f575, %f425, %f570;
	// begin inline asm
	{  cvt.f32.f16 %f426, %rs249;}

	// end inline asm
	fma.rn.ftz.f32 	%f578, %f575, %f426, %f571;
	// begin inline asm
	{  cvt.f32.f16 %f427, %rs250;}

	// end inline asm
	fma.rn.ftz.f32 	%f579, %f575, %f427, %f572;
	ld.global.v4.u32 	{%r207, %r208, %r209, %r210}, [%rd5+16];
	ld.global.v4.u32 	{%r215, %r216, %r217, %r218}, [%rd8];
	ld.global.v4.u32 	{%r223, %r224, %r225, %r226}, [%rd6+16];
	ld.global.v4.u32 	{%r231, %r232, %r233, %r234}, [%rd7+16];
	shl.b16 	%rs391, %rs73, 4;
	cvt.s16.s8 	%rs392, %rs391;
	shr.s16 	%rs393, %rs392, 7;
	and.b16  	%rs394, %rs393, -16;
	and.b16  	%rs395, %rs73, 15;
	or.b16  	%rs396, %rs394, %rs395;
	cvt.rn.f32.s16 	%f580, %rs396;
	sub.ftz.f32 	%f581, %f580, %f6;
	mul.ftz.f32 	%f582, %f47, %f581;
	mov.b32 	{%rs251, %rs255}, %r207;
	// begin inline asm
	{  cvt.f32.f16 %f428, %rs251;}

	// end inline asm
	fma.rn.ftz.f32 	%f583, %f582, %f428, %f576;
	mov.b32 	{%rs252, %rs256}, %r215;
	// begin inline asm
	{  cvt.f32.f16 %f429, %rs252;}

	// end inline asm
	fma.rn.ftz.f32 	%f584, %f582, %f429, %f577;
	mov.b32 	{%rs253, %rs257}, %r223;
	// begin inline asm
	{  cvt.f32.f16 %f430, %rs253;}

	// end inline asm
	fma.rn.ftz.f32 	%f585, %f582, %f430, %f578;
	mov.b32 	{%rs254, %rs258}, %r231;
	// begin inline asm
	{  cvt.f32.f16 %f431, %rs254;}

	// end inline asm
	fma.rn.ftz.f32 	%f586, %f582, %f431, %f579;
	cvt.s16.s8 	%rs397, %rs73;
	shr.s16 	%rs398, %rs397, 7;
	and.b16  	%rs399, %rs398, -16;
	or.b16  	%rs400, %rs399, %rs12;
	cvt.rn.f32.s16 	%f587, %rs400;
	sub.ftz.f32 	%f588, %f587, %f6;
	mul.ftz.f32 	%f589, %f47, %f588;
	// begin inline asm
	{  cvt.f32.f16 %f432, %rs255;}

	// end inline asm
	fma.rn.ftz.f32 	%f590, %f589, %f432, %f583;
	// begin inline asm
	{  cvt.f32.f16 %f433, %rs256;}

	// end inline asm
	fma.rn.ftz.f32 	%f591, %f589, %f433, %f584;
	// begin inline asm
	{  cvt.f32.f16 %f434, %rs257;}

	// end inline asm
	fma.rn.ftz.f32 	%f592, %f589, %f434, %f585;
	// begin inline asm
	{  cvt.f32.f16 %f435, %rs258;}

	// end inline asm
	fma.rn.ftz.f32 	%f593, %f589, %f435, %f586;
	cvt.s16.s8 	%rs402, %rs74;
	shr.s16 	%rs403, %rs402, 7;
	and.b16  	%rs404, %rs403, -16;
	or.b16  	%rs405, %rs404, %rs13;
	cvt.rn.f32.s16 	%f594, %rs405;
	sub.ftz.f32 	%f595, %f594, %f6;
	mul.ftz.f32 	%f596, %f47, %f595;
	mov.b32 	{%rs259, %rs263}, %r208;
	// begin inline asm
	{  cvt.f32.f16 %f436, %rs259;}

	// end inline asm
	fma.rn.ftz.f32 	%f597, %f596, %f436, %f590;
	mov.b32 	{%rs260, %rs264}, %r216;
	// begin inline asm
	{  cvt.f32.f16 %f437, %rs260;}

	// end inline asm
	fma.rn.ftz.f32 	%f598, %f596, %f437, %f591;
	mov.b32 	{%rs261, %rs265}, %r224;
	// begin inline asm
	{  cvt.f32.f16 %f438, %rs261;}

	// end inline asm
	fma.rn.ftz.f32 	%f599, %f596, %f438, %f592;
	mov.b32 	{%rs262, %rs266}, %r232;
	// begin inline asm
	{  cvt.f32.f16 %f439, %rs262;}

	// end inline asm
	fma.rn.ftz.f32 	%f600, %f596, %f439, %f593;
	shr.s16 	%rs406, %rs73, 15;
	and.b16  	%rs407, %rs406, -16;
	or.b16  	%rs408, %rs407, %rs14;
	cvt.rn.f32.s16 	%f601, %rs408;
	sub.ftz.f32 	%f602, %f601, %f6;
	mul.ftz.f32 	%f603, %f47, %f602;
	// begin inline asm
	{  cvt.f32.f16 %f440, %rs263;}

	// end inline asm
	fma.rn.ftz.f32 	%f604, %f603, %f440, %f597;
	// begin inline asm
	{  cvt.f32.f16 %f441, %rs264;}

	// end inline asm
	fma.rn.ftz.f32 	%f605, %f603, %f441, %f598;
	// begin inline asm
	{  cvt.f32.f16 %f442, %rs265;}

	// end inline asm
	fma.rn.ftz.f32 	%f606, %f603, %f442, %f599;
	// begin inline asm
	{  cvt.f32.f16 %f443, %rs266;}

	// end inline asm
	fma.rn.ftz.f32 	%f607, %f603, %f443, %f600;
	shl.b16 	%rs410, %rs76, 4;
	cvt.s16.s8 	%rs411, %rs410;
	shr.s16 	%rs412, %rs411, 7;
	and.b16  	%rs413, %rs412, -16;
	or.b16  	%rs414, %rs413, %rs15;
	cvt.rn.f32.s16 	%f608, %rs414;
	sub.ftz.f32 	%f609, %f608, %f6;
	mul.ftz.f32 	%f610, %f47, %f609;
	mov.b32 	{%rs267, %rs271}, %r209;
	// begin inline asm
	{  cvt.f32.f16 %f444, %rs267;}

	// end inline asm
	fma.rn.ftz.f32 	%f611, %f610, %f444, %f604;
	mov.b32 	{%rs268, %rs272}, %r217;
	// begin inline asm
	{  cvt.f32.f16 %f445, %rs268;}

	// end inline asm
	fma.rn.ftz.f32 	%f612, %f610, %f445, %f605;
	mov.b32 	{%rs269, %rs273}, %r225;
	// begin inline asm
	{  cvt.f32.f16 %f446, %rs269;}

	// end inline asm
	fma.rn.ftz.f32 	%f613, %f610, %f446, %f606;
	mov.b32 	{%rs270, %rs274}, %r233;
	// begin inline asm
	{  cvt.f32.f16 %f447, %rs270;}

	// end inline asm
	fma.rn.ftz.f32 	%f614, %f610, %f447, %f607;
	shl.b16 	%rs416, %rs77, 4;
	cvt.s16.s8 	%rs417, %rs416;
	shr.s16 	%rs418, %rs417, 7;
	and.b16  	%rs419, %rs418, -16;
	or.b16  	%rs420, %rs419, %rs16;
	cvt.rn.f32.s16 	%f615, %rs420;
	sub.ftz.f32 	%f616, %f615, %f6;
	mul.ftz.f32 	%f617, %f47, %f616;
	// begin inline asm
	{  cvt.f32.f16 %f448, %rs271;}

	// end inline asm
	fma.rn.ftz.f32 	%f618, %f617, %f448, %f611;
	// begin inline asm
	{  cvt.f32.f16 %f449, %rs272;}

	// end inline asm
	fma.rn.ftz.f32 	%f619, %f617, %f449, %f612;
	// begin inline asm
	{  cvt.f32.f16 %f450, %rs273;}

	// end inline asm
	fma.rn.ftz.f32 	%f620, %f617, %f450, %f613;
	// begin inline asm
	{  cvt.f32.f16 %f451, %rs274;}

	// end inline asm
	fma.rn.ftz.f32 	%f621, %f617, %f451, %f614;
	shl.b16 	%rs422, %rs78, 4;
	cvt.s16.s8 	%rs423, %rs422;
	shr.s16 	%rs424, %rs423, 7;
	and.b16  	%rs425, %rs424, -16;
	or.b16  	%rs426, %rs425, %rs17;
	cvt.rn.f32.s16 	%f622, %rs426;
	sub.ftz.f32 	%f623, %f622, %f6;
	mul.ftz.f32 	%f624, %f47, %f623;
	mov.b32 	{%rs275, %rs279}, %r210;
	// begin inline asm
	{  cvt.f32.f16 %f452, %rs275;}

	// end inline asm
	fma.rn.ftz.f32 	%f625, %f624, %f452, %f618;
	mov.b32 	{%rs276, %rs280}, %r218;
	// begin inline asm
	{  cvt.f32.f16 %f453, %rs276;}

	// end inline asm
	fma.rn.ftz.f32 	%f626, %f624, %f453, %f619;
	mov.b32 	{%rs277, %rs281}, %r226;
	// begin inline asm
	{  cvt.f32.f16 %f454, %rs277;}

	// end inline asm
	fma.rn.ftz.f32 	%f627, %f624, %f454, %f620;
	mov.b32 	{%rs278, %rs282}, %r234;
	// begin inline asm
	{  cvt.f32.f16 %f455, %rs278;}

	// end inline asm
	fma.rn.ftz.f32 	%f628, %f624, %f455, %f621;
	shr.u32 	%r242, %r55, 28;
	cvt.u16.u32 	%rs427, %r242;
	shl.b16 	%rs428, %rs427, 4;
	cvt.s16.s8 	%rs429, %rs428;
	shr.s16 	%rs430, %rs429, 7;
	and.b16  	%rs431, %rs430, -16;
	or.b16  	%rs432, %rs431, %rs427;
	cvt.rn.f32.s16 	%f629, %rs432;
	sub.ftz.f32 	%f630, %f629, %f6;
	mul.ftz.f32 	%f631, %f47, %f630;
	// begin inline asm
	{  cvt.f32.f16 %f456, %rs279;}

	// end inline asm
	fma.rn.ftz.f32 	%f632, %f631, %f456, %f625;
	// begin inline asm
	{  cvt.f32.f16 %f457, %rs280;}

	// end inline asm
	fma.rn.ftz.f32 	%f633, %f631, %f457, %f626;
	// begin inline asm
	{  cvt.f32.f16 %f458, %rs281;}

	// end inline asm
	fma.rn.ftz.f32 	%f634, %f631, %f458, %f627;
	// begin inline asm
	{  cvt.f32.f16 %f459, %rs282;}

	// end inline asm
	fma.rn.ftz.f32 	%f635, %f631, %f459, %f628;
	ld.global.v4.u32 	{%r243, %r244, %r245, %r246}, [%rd5+32];
	ld.global.v4.u32 	{%r251, %r252, %r253, %r254}, [%rd8+16];
	ld.global.v4.u32 	{%r259, %r260, %r261, %r262}, [%rd6+32];
	ld.global.v4.u32 	{%r267, %r268, %r269, %r270}, [%rd7+32];
	shl.b16 	%rs434, %rs79, 4;
	cvt.s16.s8 	%rs435, %rs434;
	shr.s16 	%rs436, %rs435, 7;
	and.b16  	%rs437, %rs436, -16;
	and.b16  	%rs438, %rs79, 15;
	or.b16  	%rs439, %rs437, %rs438;
	cvt.rn.f32.s16 	%f636, %rs439;
	sub.ftz.f32 	%f637, %f636, %f6;
	mul.ftz.f32 	%f638, %f47, %f637;
	mov.b32 	{%rs283, %rs287}, %r243;
	// begin inline asm
	{  cvt.f32.f16 %f460, %rs283;}

	// end inline asm
	fma.rn.ftz.f32 	%f639, %f638, %f460, %f632;
	mov.b32 	{%rs284, %rs288}, %r251;
	// begin inline asm
	{  cvt.f32.f16 %f461, %rs284;}

	// end inline asm
	fma.rn.ftz.f32 	%f640, %f638, %f461, %f633;
	mov.b32 	{%rs285, %rs289}, %r259;
	// begin inline asm
	{  cvt.f32.f16 %f462, %rs285;}

	// end inline asm
	fma.rn.ftz.f32 	%f641, %f638, %f462, %f634;
	mov.b32 	{%rs286, %rs290}, %r267;
	// begin inline asm
	{  cvt.f32.f16 %f463, %rs286;}

	// end inline asm
	fma.rn.ftz.f32 	%f642, %f638, %f463, %f635;
	cvt.s16.s8 	%rs440, %rs79;
	shr.s16 	%rs441, %rs440, 7;
	and.b16  	%rs442, %rs441, -16;
	or.b16  	%rs443, %rs442, %rs18;
	cvt.rn.f32.s16 	%f643, %rs443;
	sub.ftz.f32 	%f644, %f643, %f6;
	mul.ftz.f32 	%f645, %f47, %f644;
	// begin inline asm
	{  cvt.f32.f16 %f464, %rs287;}

	// end inline asm
	fma.rn.ftz.f32 	%f646, %f645, %f464, %f639;
	// begin inline asm
	{  cvt.f32.f16 %f465, %rs288;}

	// end inline asm
	fma.rn.ftz.f32 	%f647, %f645, %f465, %f640;
	// begin inline asm
	{  cvt.f32.f16 %f466, %rs289;}

	// end inline asm
	fma.rn.ftz.f32 	%f648, %f645, %f466, %f641;
	// begin inline asm
	{  cvt.f32.f16 %f467, %rs290;}

	// end inline asm
	fma.rn.ftz.f32 	%f649, %f645, %f467, %f642;
	cvt.s16.s8 	%rs445, %rs80;
	shr.s16 	%rs446, %rs445, 7;
	and.b16  	%rs447, %rs446, -16;
	or.b16  	%rs448, %rs447, %rs19;
	cvt.rn.f32.s16 	%f650, %rs448;
	sub.ftz.f32 	%f651, %f650, %f6;
	mul.ftz.f32 	%f652, %f47, %f651;
	mov.b32 	{%rs291, %rs295}, %r244;
	// begin inline asm
	{  cvt.f32.f16 %f468, %rs291;}

	// end inline asm
	fma.rn.ftz.f32 	%f653, %f652, %f468, %f646;
	mov.b32 	{%rs292, %rs296}, %r252;
	// begin inline asm
	{  cvt.f32.f16 %f469, %rs292;}

	// end inline asm
	fma.rn.ftz.f32 	%f654, %f652, %f469, %f647;
	mov.b32 	{%rs293, %rs297}, %r260;
	// begin inline asm
	{  cvt.f32.f16 %f470, %rs293;}

	// end inline asm
	fma.rn.ftz.f32 	%f655, %f652, %f470, %f648;
	mov.b32 	{%rs294, %rs298}, %r268;
	// begin inline asm
	{  cvt.f32.f16 %f471, %rs294;}

	// end inline asm
	fma.rn.ftz.f32 	%f656, %f652, %f471, %f649;
	shr.s16 	%rs449, %rs79, 15;
	and.b16  	%rs450, %rs449, -16;
	or.b16  	%rs451, %rs450, %rs20;
	cvt.rn.f32.s16 	%f657, %rs451;
	sub.ftz.f32 	%f658, %f657, %f6;
	mul.ftz.f32 	%f659, %f47, %f658;
	// begin inline asm
	{  cvt.f32.f16 %f472, %rs295;}

	// end inline asm
	fma.rn.ftz.f32 	%f660, %f659, %f472, %f653;
	// begin inline asm
	{  cvt.f32.f16 %f473, %rs296;}

	// end inline asm
	fma.rn.ftz.f32 	%f661, %f659, %f473, %f654;
	// begin inline asm
	{  cvt.f32.f16 %f474, %rs297;}

	// end inline asm
	fma.rn.ftz.f32 	%f662, %f659, %f474, %f655;
	// begin inline asm
	{  cvt.f32.f16 %f475, %rs298;}

	// end inline asm
	fma.rn.ftz.f32 	%f663, %f659, %f475, %f656;
	shl.b16 	%rs452, %rs21, 4;
	cvt.s16.s8 	%rs453, %rs452;
	shr.s16 	%rs454, %rs453, 7;
	and.b16  	%rs455, %rs454, -16;
	or.b16  	%rs456, %rs455, %rs22;
	cvt.rn.f32.s16 	%f664, %rs456;
	sub.ftz.f32 	%f665, %f664, %f6;
	mul.ftz.f32 	%f666, %f47, %f665;
	mov.b32 	{%rs299, %rs303}, %r245;
	// begin inline asm
	{  cvt.f32.f16 %f476, %rs299;}

	// end inline asm
	fma.rn.ftz.f32 	%f667, %f666, %f476, %f660;
	mov.b32 	{%rs300, %rs304}, %r253;
	// begin inline asm
	{  cvt.f32.f16 %f477, %rs300;}

	// end inline asm
	fma.rn.ftz.f32 	%f668, %f666, %f477, %f661;
	mov.b32 	{%rs301, %rs305}, %r261;
	// begin inline asm
	{  cvt.f32.f16 %f478, %rs301;}

	// end inline asm
	fma.rn.ftz.f32 	%f669, %f666, %f478, %f662;
	mov.b32 	{%rs302, %rs306}, %r269;
	// begin inline asm
	{  cvt.f32.f16 %f479, %rs302;}

	// end inline asm
	fma.rn.ftz.f32 	%f670, %f666, %f479, %f663;
	shl.b16 	%rs457, %rs23, 4;
	cvt.s16.s8 	%rs458, %rs457;
	shr.s16 	%rs459, %rs458, 7;
	and.b16  	%rs460, %rs459, -16;
	or.b16  	%rs461, %rs460, %rs24;
	cvt.rn.f32.s16 	%f671, %rs461;
	sub.ftz.f32 	%f672, %f671, %f6;
	mul.ftz.f32 	%f673, %f47, %f672;
	// begin inline asm
	{  cvt.f32.f16 %f480, %rs303;}

	// end inline asm
	fma.rn.ftz.f32 	%f674, %f673, %f480, %f667;
	// begin inline asm
	{  cvt.f32.f16 %f481, %rs304;}

	// end inline asm
	fma.rn.ftz.f32 	%f675, %f673, %f481, %f668;
	// begin inline asm
	{  cvt.f32.f16 %f482, %rs305;}

	// end inline asm
	fma.rn.ftz.f32 	%f676, %f673, %f482, %f669;
	// begin inline asm
	{  cvt.f32.f16 %f483, %rs306;}

	// end inline asm
	fma.rn.ftz.f32 	%f677, %f673, %f483, %f670;
	shl.b16 	%rs462, %rs25, 4;
	cvt.s16.s8 	%rs463, %rs462;
	shr.s16 	%rs464, %rs463, 7;
	and.b16  	%rs465, %rs464, -16;
	or.b16  	%rs466, %rs465, %rs26;
	cvt.rn.f32.s16 	%f678, %rs466;
	sub.ftz.f32 	%f679, %f678, %f6;
	mul.ftz.f32 	%f680, %f47, %f679;
	mov.b32 	{%rs307, %rs311}, %r246;
	// begin inline asm
	{  cvt.f32.f16 %f484, %rs307;}

	// end inline asm
	fma.rn.ftz.f32 	%f681, %f680, %f484, %f674;
	mov.b32 	{%rs308, %rs312}, %r254;
	// begin inline asm
	{  cvt.f32.f16 %f485, %rs308;}

	// end inline asm
	fma.rn.ftz.f32 	%f682, %f680, %f485, %f675;
	mov.b32 	{%rs309, %rs313}, %r262;
	// begin inline asm
	{  cvt.f32.f16 %f486, %rs309;}

	// end inline asm
	fma.rn.ftz.f32 	%f683, %f680, %f486, %f676;
	mov.b32 	{%rs310, %rs314}, %r270;
	// begin inline asm
	{  cvt.f32.f16 %f487, %rs310;}

	// end inline asm
	fma.rn.ftz.f32 	%f684, %f680, %f487, %f677;
	shl.b16 	%rs467, %rs27, 4;
	cvt.s16.s8 	%rs468, %rs467;
	shr.s16 	%rs469, %rs468, 7;
	and.b16  	%rs470, %rs469, -16;
	or.b16  	%rs471, %rs470, %rs27;
	cvt.rn.f32.s16 	%f685, %rs471;
	sub.ftz.f32 	%f686, %f685, %f6;
	mul.ftz.f32 	%f687, %f47, %f686;
	// begin inline asm
	{  cvt.f32.f16 %f488, %rs311;}

	// end inline asm
	fma.rn.ftz.f32 	%f688, %f687, %f488, %f681;
	// begin inline asm
	{  cvt.f32.f16 %f489, %rs312;}

	// end inline asm
	fma.rn.ftz.f32 	%f689, %f687, %f489, %f682;
	// begin inline asm
	{  cvt.f32.f16 %f490, %rs313;}

	// end inline asm
	fma.rn.ftz.f32 	%f690, %f687, %f490, %f683;
	// begin inline asm
	{  cvt.f32.f16 %f491, %rs314;}

	// end inline asm
	fma.rn.ftz.f32 	%f691, %f687, %f491, %f684;
	ld.global.v4.u32 	{%r275, %r276, %r277, %r278}, [%rd5+48];
	ld.global.v4.u32 	{%r283, %r284, %r285, %r286}, [%rd8+32];
	ld.global.v4.u32 	{%r291, %r292, %r293, %r294}, [%rd6+48];
	ld.global.v4.u32 	{%r299, %r300, %r301, %r302}, [%rd7+48];
	shl.b16 	%rs473, %rs82, 4;
	cvt.s16.s8 	%rs474, %rs473;
	shr.s16 	%rs475, %rs474, 7;
	and.b16  	%rs476, %rs475, -16;
	or.b16  	%rs477, %rs476, %rs28;
	cvt.rn.f32.s16 	%f692, %rs477;
	sub.ftz.f32 	%f693, %f692, %f6;
	mul.ftz.f32 	%f694, %f47, %f693;
	mov.b32 	{%rs315, %rs319}, %r275;
	// begin inline asm
	{  cvt.f32.f16 %f492, %rs315;}

	// end inline asm
	fma.rn.ftz.f32 	%f695, %f694, %f492, %f688;
	mov.b32 	{%rs316, %rs320}, %r283;
	// begin inline asm
	{  cvt.f32.f16 %f493, %rs316;}

	// end inline asm
	fma.rn.ftz.f32 	%f696, %f694, %f493, %f689;
	mov.b32 	{%rs317, %rs321}, %r291;
	// begin inline asm
	{  cvt.f32.f16 %f494, %rs317;}

	// end inline asm
	fma.rn.ftz.f32 	%f697, %f694, %f494, %f690;
	mov.b32 	{%rs318, %rs322}, %r299;
	// begin inline asm
	{  cvt.f32.f16 %f495, %rs318;}

	// end inline asm
	fma.rn.ftz.f32 	%f698, %f694, %f495, %f691;
	shl.b16 	%rs478, %rs29, 4;
	cvt.s16.s8 	%rs479, %rs478;
	shr.s16 	%rs480, %rs479, 7;
	and.b16  	%rs481, %rs480, -16;
	or.b16  	%rs482, %rs481, %rs30;
	cvt.rn.f32.s16 	%f699, %rs482;
	sub.ftz.f32 	%f700, %f699, %f6;
	mul.ftz.f32 	%f701, %f47, %f700;
	// begin inline asm
	{  cvt.f32.f16 %f496, %rs319;}

	// end inline asm
	fma.rn.ftz.f32 	%f702, %f701, %f496, %f695;
	// begin inline asm
	{  cvt.f32.f16 %f497, %rs320;}

	// end inline asm
	fma.rn.ftz.f32 	%f703, %f701, %f497, %f696;
	// begin inline asm
	{  cvt.f32.f16 %f498, %rs321;}

	// end inline asm
	fma.rn.ftz.f32 	%f704, %f701, %f498, %f697;
	// begin inline asm
	{  cvt.f32.f16 %f499, %rs322;}

	// end inline asm
	fma.rn.ftz.f32 	%f705, %f701, %f499, %f698;
	shl.b16 	%rs483, %rs31, 4;
	cvt.s16.s8 	%rs484, %rs483;
	shr.s16 	%rs485, %rs484, 7;
	and.b16  	%rs486, %rs485, -16;
	or.b16  	%rs487, %rs486, %rs32;
	cvt.rn.f32.s16 	%f706, %rs487;
	sub.ftz.f32 	%f707, %f706, %f6;
	mul.ftz.f32 	%f708, %f47, %f707;
	mov.b32 	{%rs323, %rs327}, %r276;
	// begin inline asm
	{  cvt.f32.f16 %f500, %rs323;}

	// end inline asm
	fma.rn.ftz.f32 	%f709, %f708, %f500, %f702;
	mov.b32 	{%rs324, %rs328}, %r284;
	// begin inline asm
	{  cvt.f32.f16 %f501, %rs324;}

	// end inline asm
	fma.rn.ftz.f32 	%f710, %f708, %f501, %f703;
	mov.b32 	{%rs325, %rs329}, %r292;
	// begin inline asm
	{  cvt.f32.f16 %f502, %rs325;}

	// end inline asm
	fma.rn.ftz.f32 	%f711, %f708, %f502, %f704;
	mov.b32 	{%rs326, %rs330}, %r300;
	// begin inline asm
	{  cvt.f32.f16 %f503, %rs326;}

	// end inline asm
	fma.rn.ftz.f32 	%f712, %f708, %f503, %f705;
	shl.b16 	%rs488, %rs33, 4;
	cvt.s16.s8 	%rs489, %rs488;
	shr.s16 	%rs490, %rs489, 7;
	and.b16  	%rs491, %rs490, -16;
	or.b16  	%rs492, %rs491, %rs34;
	cvt.rn.f32.s16 	%f713, %rs492;
	sub.ftz.f32 	%f714, %f713, %f6;
	mul.ftz.f32 	%f715, %f47, %f714;
	// begin inline asm
	{  cvt.f32.f16 %f504, %rs327;}

	// end inline asm
	fma.rn.ftz.f32 	%f716, %f715, %f504, %f709;
	// begin inline asm
	{  cvt.f32.f16 %f505, %rs328;}

	// end inline asm
	fma.rn.ftz.f32 	%f717, %f715, %f505, %f710;
	// begin inline asm
	{  cvt.f32.f16 %f506, %rs329;}

	// end inline asm
	fma.rn.ftz.f32 	%f718, %f715, %f506, %f711;
	// begin inline asm
	{  cvt.f32.f16 %f507, %rs330;}

	// end inline asm
	fma.rn.ftz.f32 	%f719, %f715, %f507, %f712;
	shl.b16 	%rs493, %rs35, 4;
	cvt.s16.s8 	%rs494, %rs493;
	shr.s16 	%rs495, %rs494, 7;
	and.b16  	%rs496, %rs495, -16;
	or.b16  	%rs497, %rs496, %rs36;
	cvt.rn.f32.s16 	%f720, %rs497;
	sub.ftz.f32 	%f721, %f720, %f6;
	mul.ftz.f32 	%f722, %f47, %f721;
	mov.b32 	{%rs331, %rs335}, %r277;
	// begin inline asm
	{  cvt.f32.f16 %f508, %rs331;}

	// end inline asm
	fma.rn.ftz.f32 	%f723, %f722, %f508, %f716;
	mov.b32 	{%rs332, %rs336}, %r285;
	// begin inline asm
	{  cvt.f32.f16 %f509, %rs332;}

	// end inline asm
	fma.rn.ftz.f32 	%f724, %f722, %f509, %f717;
	mov.b32 	{%rs333, %rs337}, %r293;
	// begin inline asm
	{  cvt.f32.f16 %f510, %rs333;}

	// end inline asm
	fma.rn.ftz.f32 	%f725, %f722, %f510, %f718;
	mov.b32 	{%rs334, %rs338}, %r301;
	// begin inline asm
	{  cvt.f32.f16 %f511, %rs334;}

	// end inline asm
	fma.rn.ftz.f32 	%f726, %f722, %f511, %f719;
	shl.b16 	%rs498, %rs37, 4;
	cvt.s16.s8 	%rs499, %rs498;
	shr.s16 	%rs500, %rs499, 7;
	and.b16  	%rs501, %rs500, -16;
	or.b16  	%rs502, %rs501, %rs38;
	cvt.rn.f32.s16 	%f727, %rs502;
	sub.ftz.f32 	%f728, %f727, %f6;
	mul.ftz.f32 	%f729, %f47, %f728;
	// begin inline asm
	{  cvt.f32.f16 %f512, %rs335;}

	// end inline asm
	fma.rn.ftz.f32 	%f730, %f729, %f512, %f723;
	// begin inline asm
	{  cvt.f32.f16 %f513, %rs336;}

	// end inline asm
	fma.rn.ftz.f32 	%f731, %f729, %f513, %f724;
	// begin inline asm
	{  cvt.f32.f16 %f514, %rs337;}

	// end inline asm
	fma.rn.ftz.f32 	%f732, %f729, %f514, %f725;
	// begin inline asm
	{  cvt.f32.f16 %f515, %rs338;}

	// end inline asm
	fma.rn.ftz.f32 	%f733, %f729, %f515, %f726;
	shl.b16 	%rs503, %rs39, 4;
	cvt.s16.s8 	%rs504, %rs503;
	shr.s16 	%rs505, %rs504, 7;
	and.b16  	%rs506, %rs505, -16;
	or.b16  	%rs507, %rs506, %rs40;
	cvt.rn.f32.s16 	%f734, %rs507;
	sub.ftz.f32 	%f735, %f734, %f6;
	mul.ftz.f32 	%f736, %f47, %f735;
	mov.b32 	{%rs339, %rs343}, %r278;
	// begin inline asm
	{  cvt.f32.f16 %f516, %rs339;}

	// end inline asm
	fma.rn.ftz.f32 	%f737, %f736, %f516, %f730;
	mov.b32 	{%rs340, %rs344}, %r286;
	// begin inline asm
	{  cvt.f32.f16 %f517, %rs340;}

	// end inline asm
	fma.rn.ftz.f32 	%f738, %f736, %f517, %f731;
	mov.b32 	{%rs341, %rs345}, %r294;
	// begin inline asm
	{  cvt.f32.f16 %f518, %rs341;}

	// end inline asm
	fma.rn.ftz.f32 	%f739, %f736, %f518, %f732;
	mov.b32 	{%rs342, %rs346}, %r302;
	// begin inline asm
	{  cvt.f32.f16 %f519, %rs342;}

	// end inline asm
	fma.rn.ftz.f32 	%f740, %f736, %f519, %f733;
	shl.b16 	%rs508, %rs41, 4;
	cvt.s16.s8 	%rs509, %rs508;
	shr.s16 	%rs510, %rs509, 7;
	and.b16  	%rs511, %rs510, -16;
	or.b16  	%rs512, %rs511, %rs41;
	cvt.rn.f32.s16 	%f741, %rs512;
	sub.ftz.f32 	%f742, %f741, %f6;
	mul.ftz.f32 	%f743, %f47, %f742;
	// begin inline asm
	{  cvt.f32.f16 %f520, %rs343;}

	// end inline asm
	fma.rn.ftz.f32 	%f843, %f743, %f520, %f737;
	// begin inline asm
	{  cvt.f32.f16 %f521, %rs344;}

	// end inline asm
	fma.rn.ftz.f32 	%f842, %f743, %f521, %f738;
	// begin inline asm
	{  cvt.f32.f16 %f522, %rs345;}

	// end inline asm
	fma.rn.ftz.f32 	%f841, %f743, %f522, %f739;
	// begin inline asm
	{  cvt.f32.f16 %f523, %rs346;}

	// end inline asm
	fma.rn.ftz.f32 	%f840, %f743, %f523, %f740;

$L__BB0_8:
	add.s32 	%r398, %r398, 4;
	shl.b32 	%r307, %r398, 5;
	add.s32 	%r397, %r307, %r50;
	shl.b32 	%r396, %r397, 2;
	setp.lt.u32 	%p7, %r396, %r47;
	@%p7 bra 	$L__BB0_2;

$L__BB0_9:
	mov.u32 	%r391, %tid.y;
	mov.u32 	%r390, %tid.x;
	shl.b32 	%r389, %r391, 5;
	add.s32 	%r388, %r389, %r390;
	shl.b32 	%r309, %r388, 2;
	mov.u32 	%r310, _ZZ9gemv_int4ILi4ELi128ELi4EEvP6__halfPKS0_S3_PKjPKhS3_iiiiffbE12temp_storage;
	add.s32 	%r311, %r310, %r309;
	setp.lt.u32 	%p8, %r388, 32;
	@%p8 bra 	$L__BB0_11;

	add.s32 	%r384, %r311, -112;
	st.shared.f32 	[%r384], %f843;

$L__BB0_11:
	mov.u32 	%r395, %tid.y;
	mov.u32 	%r394, %tid.x;
	shl.b32 	%r393, %r395, 5;
	add.s32 	%r392, %r393, %r394;
	setp.gt.u32 	%p9, %r392, 31;
	bar.sync 	0;
	mad.lo.s32 	%r40, %r392, 12, %r310;
	@%p9 bra 	$L__BB0_13;

	mov.u32 	%r326, 16;
	ld.shared.f32 	%f759, [%r40+16];
	add.ftz.f32 	%f760, %f843, %f759;
	ld.shared.f32 	%f761, [%r40+20];
	add.ftz.f32 	%f762, %f760, %f761;
	ld.shared.f32 	%f763, [%r40+24];
	add.ftz.f32 	%f746, %f762, %f763;
	mov.u32 	%r314, 1;
	mov.u32 	%r327, 31;
	mov.u32 	%r328, -1;
	// begin inline asm
	{  .reg .f32 r0;  .reg .pred p;  shfl.sync.down.b32 r0|p, %f746, %r314, %r327, %r328;  @p add.f32 r0, r0, %f746;  mov.f32 %f744, r0;}
	// end inline asm
	mov.u32 	%r317, 2;
	// begin inline asm
	{  .reg .f32 r0;  .reg .pred p;  shfl.sync.down.b32 r0|p, %f744, %r317, %r327, %r328;  @p add.f32 r0, r0, %f744;  mov.f32 %f747, r0;}
	// end inline asm
	mov.u32 	%r320, 4;
	// begin inline asm
	{  .reg .f32 r0;  .reg .pred p;  shfl.sync.down.b32 r0|p, %f747, %r320, %r327, %r328;  @p add.f32 r0, r0, %f747;  mov.f32 %f750, r0;}
	// end inline asm
	mov.u32 	%r323, 8;
	// begin inline asm
	{  .reg .f32 r0;  .reg .pred p;  shfl.sync.down.b32 r0|p, %f750, %r323, %r327, %r328;  @p add.f32 r0, r0, %f750;  mov.f32 %f753, r0;}
	// end inline asm
	// begin inline asm
	{  .reg .f32 r0;  .reg .pred p;  shfl.sync.down.b32 r0|p, %f753, %r326, %r327, %r328;  @p add.f32 r0, r0, %f753;  mov.f32 %f843, r0;}
	// end inline asm

$L__BB0_13:
	@%p8 bra 	$L__BB0_15;

	add.s32 	%r385, %r311, -112;
	st.shared.f32 	[%r385+640], %f842;

$L__BB0_15:
	bar.sync 	0;
	@%p9 bra 	$L__BB0_17;

	ld.shared.f32 	%f779, [%r40+656];
	add.ftz.f32 	%f780, %f842, %f779;
	ld.shared.f32 	%f781, [%r40+660];
	add.ftz.f32 	%f782, %f780, %f781;
	ld.shared.f32 	%f783, [%r40+664];
	add.ftz.f32 	%f766, %f782, %f783;
	mov.u32 	%r330, 1;
	mov.u32 	%r343, 31;
	mov.u32 	%r344, -1;
	// begin inline asm
	{  .reg .f32 r0;  .reg .pred p;  shfl.sync.down.b32 r0|p, %f766, %r330, %r343, %r344;  @p add.f32 r0, r0, %f766;  mov.f32 %f764, r0;}
	// end inline asm
	mov.u32 	%r333, 2;
	// begin inline asm
	{  .reg .f32 r0;  .reg .pred p;  shfl.sync.down.b32 r0|p, %f764, %r333, %r343, %r344;  @p add.f32 r0, r0, %f764;  mov.f32 %f767, r0;}
	// end inline asm
	mov.u32 	%r336, 4;
	// begin inline asm
	{  .reg .f32 r0;  .reg .pred p;  shfl.sync.down.b32 r0|p, %f767, %r336, %r343, %r344;  @p add.f32 r0, r0, %f767;  mov.f32 %f770, r0;}
	// end inline asm
	mov.u32 	%r339, 8;
	// begin inline asm
	{  .reg .f32 r0;  .reg .pred p;  shfl.sync.down.b32 r0|p, %f770, %r339, %r343, %r344;  @p add.f32 r0, r0, %f770;  mov.f32 %f773, r0;}
	// end inline asm
	mov.u32 	%r342, 16;
	// begin inline asm
	{  .reg .f32 r0;  .reg .pred p;  shfl.sync.down.b32 r0|p, %f773, %r342, %r343, %r344;  @p add.f32 r0, r0, %f773;  mov.f32 %f842, r0;}
	// end inline asm

$L__BB0_17:
	@%p8 bra 	$L__BB0_19;

	add.s32 	%r386, %r311, -112;
	st.shared.f32 	[%r386+1280], %f841;

$L__BB0_19:
	bar.sync 	0;
	@%p9 bra 	$L__BB0_21;

	ld.shared.f32 	%f799, [%r40+1296];
	add.ftz.f32 	%f800, %f841, %f799;
	ld.shared.f32 	%f801, [%r40+1300];
	add.ftz.f32 	%f802, %f800, %f801;
	ld.shared.f32 	%f803, [%r40+1304];
	add.ftz.f32 	%f786, %f802, %f803;
	mov.u32 	%r346, 1;
	mov.u32 	%r359, 31;
	mov.u32 	%r360, -1;
	// begin inline asm
	{  .reg .f32 r0;  .reg .pred p;  shfl.sync.down.b32 r0|p, %f786, %r346, %r359, %r360;  @p add.f32 r0, r0, %f786;  mov.f32 %f784, r0;}
	// end inline asm
	mov.u32 	%r349, 2;
	// begin inline asm
	{  .reg .f32 r0;  .reg .pred p;  shfl.sync.down.b32 r0|p, %f784, %r349, %r359, %r360;  @p add.f32 r0, r0, %f784;  mov.f32 %f787, r0;}
	// end inline asm
	mov.u32 	%r352, 4;
	// begin inline asm
	{  .reg .f32 r0;  .reg .pred p;  shfl.sync.down.b32 r0|p, %f787, %r352, %r359, %r360;  @p add.f32 r0, r0, %f787;  mov.f32 %f790, r0;}
	// end inline asm
	mov.u32 	%r355, 8;
	// begin inline asm
	{  .reg .f32 r0;  .reg .pred p;  shfl.sync.down.b32 r0|p, %f790, %r355, %r359, %r360;  @p add.f32 r0, r0, %f790;  mov.f32 %f793, r0;}
	// end inline asm
	mov.u32 	%r358, 16;
	// begin inline asm
	{  .reg .f32 r0;  .reg .pred p;  shfl.sync.down.b32 r0|p, %f793, %r358, %r359, %r360;  @p add.f32 r0, r0, %f793;  mov.f32 %f841, r0;}
	// end inline asm

$L__BB0_21:
	@%p8 bra 	$L__BB0_23;

	add.s32 	%r387, %r311, -112;
	st.shared.f32 	[%r387+1920], %f840;

$L__BB0_23:
	bar.sync 	0;
	@%p9 bra 	$L__BB0_25;

	ld.shared.f32 	%f819, [%r40+1936];
	add.ftz.f32 	%f820, %f840, %f819;
	ld.shared.f32 	%f821, [%r40+1940];
	add.ftz.f32 	%f822, %f820, %f821;
	ld.shared.f32 	%f823, [%r40+1944];
	add.ftz.f32 	%f806, %f822, %f823;
	mov.u32 	%r362, 1;
	mov.u32 	%r375, 31;
	mov.u32 	%r376, -1;
	// begin inline asm
	{  .reg .f32 r0;  .reg .pred p;  shfl.sync.down.b32 r0|p, %f806, %r362, %r375, %r376;  @p add.f32 r0, r0, %f806;  mov.f32 %f804, r0;}
	// end inline asm
	mov.u32 	%r365, 2;
	// begin inline asm
	{  .reg .f32 r0;  .reg .pred p;  shfl.sync.down.b32 r0|p, %f804, %r365, %r375, %r376;  @p add.f32 r0, r0, %f804;  mov.f32 %f807, r0;}
	// end inline asm
	mov.u32 	%r368, 4;
	// begin inline asm
	{  .reg .f32 r0;  .reg .pred p;  shfl.sync.down.b32 r0|p, %f807, %r368, %r375, %r376;  @p add.f32 r0, r0, %f807;  mov.f32 %f810, r0;}
	// end inline asm
	mov.u32 	%r371, 8;
	// begin inline asm
	{  .reg .f32 r0;  .reg .pred p;  shfl.sync.down.b32 r0|p, %f810, %r371, %r375, %r376;  @p add.f32 r0, r0, %f810;  mov.f32 %f813, r0;}
	// end inline asm
	mov.u32 	%r374, 16;
	// begin inline asm
	{  .reg .f32 r0;  .reg .pred p;  shfl.sync.down.b32 r0|p, %f813, %r374, %r375, %r376;  @p add.f32 r0, r0, %f813;  mov.f32 %f840, r0;}
	// end inline asm

$L__BB0_25:
	mov.u32 	%r377, %tid.y;
	or.b32  	%r379, %r50, %r377;
	setp.ne.s32 	%p16, %r379, 0;
	@%p16 bra 	$L__BB0_29;

	ld.param.u64 	%rd52, [_Z28dequant_gemv_group128_batch423DequantGemvKernelParams_param_0+8];
	ld.param.u64 	%rd51, [_Z28dequant_gemv_group128_batch423DequantGemvKernelParams_param_0];
	cvta.to.global.u64 	%rd37, %rd51;
	setp.eq.s64 	%p17, %rd52, 0;
	mul.ftz.f32 	%f31, %f41, %f843;
	mov.u32 	%r380, %ctaid.x;
	cvt.s64.s32 	%rd10, %r380;
	mul.wide.s32 	%rd38, %r380, 2;
	add.s64 	%rd11, %rd37, %rd38;
	mul.ftz.f32 	%f32, %f41, %f842;
	add.s32 	%r381, %r46, %r380;
	cvt.s64.s32 	%rd12, %r46;
	mul.wide.s32 	%rd39, %r46, 2;
	add.s64 	%rd13, %rd11, %rd39;
	mul.ftz.f32 	%f33, %f41, %f841;
	add.s32 	%r382, %r381, %r46;
	cvt.s64.s32 	%rd14, %r382;
	mul.wide.s32 	%rd40, %r382, 2;
	add.s64 	%rd15, %rd37, %rd40;
	mul.ftz.f32 	%f34, %f41, %f840;
	add.s32 	%r383, %r382, %r46;
	cvt.s64.s32 	%rd16, %r383;
	mul.wide.s32 	%rd41, %r383, 2;
	add.s64 	%rd17, %rd37, %rd41;
	@%p17 bra 	$L__BB0_28;

	ld.param.u64 	%rd53, [_Z28dequant_gemv_group128_batch423DequantGemvKernelParams_param_0+8];
	cvta.to.global.u64 	%rd42, %rd53;
	shl.b64 	%rd43, %rd10, 1;
	add.s64 	%rd44, %rd42, %rd43;
	ld.global.u16 	%rs513, [%rd44];
	// begin inline asm
	{  cvt.f32.f16 %f824, %rs513;}

	// end inline asm
	fma.rn.ftz.f32 	%f825, %f42, %f824, %f31;
	// begin inline asm
	{  cvt.rn.f16.f32 %rs514, %f825;}

	// end inline asm
	st.global.u16 	[%rd11], %rs514;
	shl.b64 	%rd45, %rd12, 1;
	add.s64 	%rd46, %rd44, %rd45;
	ld.global.u16 	%rs515, [%rd46];
	// begin inline asm
	{  cvt.f32.f16 %f826, %rs515;}

	// end inline asm
	fma.rn.ftz.f32 	%f827, %f42, %f826, %f32;
	// begin inline asm
	{  cvt.rn.f16.f32 %rs516, %f827;}

	// end inline asm
	st.global.u16 	[%rd13], %rs516;
	shl.b64 	%rd47, %rd14, 1;
	add.s64 	%rd48, %rd42, %rd47;
	ld.global.u16 	%rs517, [%rd48];
	// begin inline asm
	{  cvt.f32.f16 %f828, %rs517;}

	// end inline asm
	fma.rn.ftz.f32 	%f829, %f42, %f828, %f33;
	// begin inline asm
	{  cvt.rn.f16.f32 %rs518, %f829;}

	// end inline asm
	st.global.u16 	[%rd15], %rs518;
	shl.b64 	%rd49, %rd16, 1;
	add.s64 	%rd50, %rd42, %rd49;
	ld.global.u16 	%rs519, [%rd50];
	// begin inline asm
	{  cvt.f32.f16 %f830, %rs519;}

	// end inline asm
	fma.rn.ftz.f32 	%f831, %f42, %f830, %f34;
	// begin inline asm
	{  cvt.rn.f16.f32 %rs520, %f831;}

	// end inline asm
	st.global.u16 	[%rd17], %rs520;
	bra.uni 	$L__BB0_29;

$L__BB0_28:
	// begin inline asm
	{  cvt.rn.f16.f32 %rs521, %f31;}

	// end inline asm
	st.global.u16 	[%rd11], %rs521;
	// begin inline asm
	{  cvt.rn.f16.f32 %rs522, %f32;}

	// end inline asm
	st.global.u16 	[%rd13], %rs522;
	// begin inline asm
	{  cvt.rn.f16.f32 %rs523, %f33;}

	// end inline asm
	st.global.u16 	[%rd15], %rs523;
	// begin inline asm
	{  cvt.rn.f16.f32 %rs524, %f34;}

	// end inline asm
	st.global.u16 	[%rd17], %rs524;

$L__BB0_29:
	ret;

}
	// .globl	_ZN3cub11EmptyKernelIvEEvv
.visible .entry _ZN3cub11EmptyKernelIvEEvv()
{



	ret;

}