506thrust12placeholders2_3E[1];
.global .align 1 .b8 _ZN37_INTERNAL_36609967_7_gemv_cu_2054f3506thrust12placeholders2_4E[1];
.global .align 1 .b8 _ZN37_INTERNAL_36609967_7_gemv_cu_2054f3506thrust12placeholders2_5E[1];
.global .align 1 .b8 _ZN37_INTERNAL_36609967_7_gemv_cu_2054f3506thrust12placeholders2_6E[1];
.global .align 1 .b8 _ZN37_INTERNAL_36609967_7_gemv_cu_2054f3506thrust12placeholders2_7E[1];
.global .align 1 .b8 _ZN37_INTERNAL_36609967_7_gemv_cu_2054f3506thrust12placeholders2_8E[1];
.global .align 1 .b8 _ZN37_INTERNAL_36609967_7_gemv_cu_2054f3506thrust12placeholders2_9E[1];
.global .align 1 .b8 _ZN37_INTERNAL_36609967_7_gemv_cu_2054f3506thrust12placeholders3_10E[1];

.visible .entry _Z28dequant_gemv_group128_batch323DequantGemvKernelParams(
	.param .align 8 .b8 _Z28dequant_gemv_group128_batch323DequantGemvKernelParams_param_0[80]
)
{
	.reg .pred 	%p<18>;
	.reg .b16 	%rs<453>;
	.reg .f32 	%f<697>;
	.reg .b32 	%r<342>;
	.reg .b64 	%rd<65>;
	// demoted variable
	.shared .align 16 .b8 _ZZ9gemv_int4ILi4ELi128ELi3EEvP6__halfPKS0_S3_PKjPKhS3_iiiiffbE12temp_storage[1920];

	ld.param.v2.u32 	{%r40, %r41}, [_Z28dequant_gemv_group128_batch323DequantGemvKernelParams_param_0+48];
	ld.param.v2.u32 	{%r42, %r43}, [_Z28dequant_gemv_group128_batch323DequantGemvKernelParams_param_0+56];
	ld.param.v2.f32 	{%f40, %f41}, [_Z28dequant_gemv_group128_batch323DequantGemvKernelParams_param_0+64];
	ld.param.v4.u8 	{%rs73, %rs74, %rs75, %rs76}, [_Z28dequant_gemv_group128_batch323DequantGemvKernelParams_param_0+72];
	ld.param.u64 	%rd20, [_Z28dequant_gemv_group128_batch323DequantGemvKernelParams_param_0+40];
	ld.param.u64 	%rd19, [_Z28dequant_gemv_group128_batch323DequantGemvKernelParams_param_0+32];
	ld.param.u64 	%rd18, [_Z28dequant_gemv_group128_batch323DequantGemvKernelParams_param_0+24];
	ld.param.u64 	%rd17, [_Z28dequant_gemv_group128_batch323DequantGemvKernelParams_param_0+16];
	mov.u32 	%r341, %tid.y;
	shl.b32 	%r44, %r341, 5;
	mov.u32 	%r45, %tid.x;
	add.s32 	%r340, %r44, %r45;
	shl.b32 	%r339, %r340, 2;
	setp.ge.u32 	%p1, %r339, %r42;
	mov.f32 	%f685, 0f00000000;
	mov.f32 	%f686, %f685;
	mov.f32 	%f687, %f685;
	@%p1 bra 	$L__BB0_9;

	cvta.to.global.u64 	%rd3, %rd17;
	mov.u32 	%r46, %ctaid.x;
	mul.lo.s32 	%r6, %r43, %r46;

$L__BB0_2:
	mad.lo.s32 	%r52, %r42, %r46, %r339;
	mul.wide.u32 	%rd27, %r52, 4;
	add.s64 	%rd22, %rd18, %rd27;
	// begin inline asm
	createpolicy.fractional.L2::evict_first.L2::evict_first.b64 %rd21, 1.0;
	// end inline asm
	// begin inline asm
	ld.global.L1::no_allocate.L2::cache_hint.v4.u32 {%r47,%r48,%r49,%r50}, [%rd22], %rd21;
	// end inline asm
	shr.u32 	%r54, %r45, 2;
	shl.b32 	%r55, %r341, 3;
	add.s32 	%r14, %r55, %r54;
	add.s32 	%r15, %r14, %r6;
	mul.wide.s32 	%rd28, %r15, 2;
	add.s64 	%rd25, %rd20, %rd28;
	// begin inline asm
	createpolicy.fractional.L2::evict_first.L2::evict_first.b64 %rd24, 1.0;
	// end inline asm
	// begin inline asm
	ld.global.L1::evict_last.L2::cache_hint.u16 %rs81, [%rd25], %rd24;
	// end inline asm
	// begin inline asm
	{  cvt.f32.f16 %f45, %rs81;}

	// end inline asm
	shl.b16 	%rs452, %rs73, 3;
	setp.eq.s64 	%p2, %rd19, 0;
	@%p2 bra 	$L__BB0_4;

	shr.u32 	%r56, %r15, 31;
	add.s32 	%r57, %r15, %r56;
	shr.s32 	%r58, %r57, 1;
	cvt.s64.s32 	%rd32, %r58;
	add.s64 	%rd30, %rd19, %rd32;
	// begin inline asm
	createpolicy.fractional.L2::evict_first.L2::evict_first.b64 %rd29, 1.0;
	// end inline asm
	// begin inline asm
	ld.global.L1::evict_last.L2::cache_hint.u8 %rs83, [%rd30], %rd29;
	// end inline asm
	cvt.u32.u16 	%r59, %rs83;
	and.b32  	%r60, %r59, 255;
	shl.b32 	%r61, %r14, 2;
	and.b32  	%r62, %r61, 4;
	shr.u32 	%r63, %r60, %r62;
	cvt.u16.u32 	%rs84, %r63;
	and.b16  	%rs452, %rs84, 15;

$L__BB0_4:
	shl.b32 	%r16, %r340, 5;
	setp.ge.s32 	%p3, %r16, %r40;
	@%p3 bra 	$L__BB0_8;

	setp.eq.s16 	%p4, %rs73, 0;
	shr.u16 	%rs86, %rs452, 3;
	and.b16  	%rs87, %rs86, 1;
	setp.eq.b16 	%p5, %rs87, 1;
	and.pred  	%p6, %p4, %p5;
	selp.b16 	%rs88, -16, 0, %p6;
	or.b16  	%rs89, %rs88, %rs452;
	cvt.s16.s8 	%rs90, %rs89;
	cvt.rn.f32.s16 	%f5, %rs90;
	mul.wide.s32 	%rd33, %r16, 2;
	add.s64 	%rd7, %rd3, %rd33;
	ld.global.v4.u32 	{%r64, %r65, %r66, %r67}, [%rd7];
	mul.wide.s32 	%rd34, %r40, 2;
	add.s64 	%rd35, %rd7, %rd34;
	ld.global.v4.u32 	{%r68, %r69, %r70, %r71}, [%rd35];
	add.s32 	%r72, %r16, %r40;
	add.s32 	%r73, %r72, %r40;
	mul.wide.s32 	%rd36, %r73, 2;
	add.s64 	%rd8, %rd3, %rd36;
	ld.global.v4.u32 	{%r74, %r75, %r76, %r77}, [%rd8];
	cvt.u16.u32 	%rs5, %r47;
	and.b16  	%rs6, %rs5, 15;
	shr.u32 	%r78, %r47, 4;
	cvt.u16.u32 	%rs7, %r78;
	and.b16  	%rs8, %rs7, 15;
	shr.u32 	%r79, %r47, 8;
	cvt.u16.u32 	%rs9, %r79;
	and.b16  	%rs10, %rs9, 15;
	shr.u32 	%r80, %r47, 12;
	cvt.u16.u32 	%rs11, %r80;
	and.b16  	%rs12, %rs11, 15;
	shr.u32 	%r81, %r47, 16;
	cvt.u16.u32 	%rs13, %r81;
	and.b16  	%rs14, %rs13, 15;
	shr.u32 	%r82, %r47, 20;
	cvt.u16.u32 	%rs15, %r82;
	and.b16  	%rs16, %rs15, 15;
	shr.u32 	%r83, %r47, 24;
	cvt.u16.u32 	%rs17, %r83;
	and.b16  	%rs18, %rs17, 15;
	shr.u32 	%r84, %r47, 28;
	cvt.u16.u32 	%rs19, %r84;
	add.s32 	%r85, %r72, 8;
	mul.wide.s32 	%rd37, %r85, 2;
	add.s64 	%rd9, %rd3, %rd37;
	cvt.u16.u32 	%rs20, %r48;
	and.b16  	%rs21, %rs20, 15;
	shr.u32 	%r86, %r48, 4;
	cvt.u16.u32 	%rs22, %r86;
	and.b16  	%rs23, %rs22, 15;
	shr.u32 	%r87, %r48, 8;
	cvt.u16.u32 	%rs24, %r87;
	and.b16  	%rs25, %rs24, 15;
	shr.u32 	%r88, %r48, 12;
	cvt.u16.u32 	%rs26, %r88;
	and.b16  	%rs27, %rs26, 15;
	shr.u32 	%r89, %r48, 16;
	cvt.u16.u32 	%rs28, %r89;
	and.b16  	%rs29, %rs28, 15;
	shr.u32 	%r90, %r48, 20;
	cvt.u16.u32 	%rs30, %r90;
	and.b16  	%rs31, %rs30, 15;
	shr.u32 	%r91, %r48, 24;
	cvt.u16.u32 	%rs32, %r91;
	and.b16  	%rs33, %rs32, 15;
	shr.u32 	%r92, %r48, 28;
	cvt.u16.u32 	%rs34, %r92;
	cvt.u16.u32 	%rs35, %r49;
	and.b16  	%rs36, %rs35, 15;
	shr.u32 	%r93, %r49, 4;
	cvt.u16.u32 	%rs37, %r93;
	and.b16  	%rs38, %rs37, 15;
	shr.u32 	%r94, %r49, 8;
	cvt.u16.u32 	%rs39, %r94;
	and.b16  	%rs40, %rs39, 15;
	shr.u32 	%r95, %r49, 12;
	cvt.u16.u32 	%rs41, %r95;
	and.b16  	%rs42, %rs41, 15;
	shr.u32 	%r96, %r49, 16;
	cvt.u16.u32 	%rs43, %r96;
	and.b16  	%rs44, %rs43, 15;
	shr.u32 	%r97, %r49, 20;
	cvt.u16.u32 	%rs45, %r97;
	and.b16  	%rs46, %rs45, 15;
	shr.u32 	%r98, %r49, 24;
	cvt.u16.u32 	%rs47, %r98;
	and.b16  	%rs48, %rs47, 15;
	shr.u32 	%r99, %r49, 28;
	cvt.u16.u32 	%rs49, %r99;
	cvt.u16.u32 	%rs50, %r50;
	and.b16  	%rs51, %rs50, 15;
	shr.u32 	%r100, %r50, 4;
	cvt.u16.u32 	%rs52, %r100;
	and.b16  	%rs53, %rs52, 15;
	shr.u32 	%r101, %r50, 8;
	cvt.u16.u32 	%rs54, %r101;
	and.b16  	%rs55, %rs54, 15;
	shr.u32 	%r102, %r50, 12;
	cvt.u16.u32 	%rs56, %r102;
	and.b16  	%rs57, %rs56, 15;
	shr.u32 	%r103, %r50, 16;
	cvt.u16.u32 	%rs58, %r103;
	and.b16  	%rs59, %rs58, 15;
	shr.u32 	%r104, %r50, 20;
	cvt.u16.u32 	%rs60, %r104;
	and.b16  	%rs61, %rs60, 15;
	shr.u32 	%r105, %r50, 24;
	cvt.u16.u32 	%rs62, %r105;
	and.b16  	%rs63, %rs62, 15;
	shr.u32 	%r106, %r50, 28;
	cvt.u16.u32 	%rs64, %r106;
	@%p4 bra 	$L__BB0_7;

	cvt.rn.f32.s16 	%f142, %rs6;
	sub.ftz.f32 	%f143, %f142, %f5;
	mul.ftz.f32 	%f144, %f45, %f143;
	mov.b32 	{%rs91, %rs94}, %r64;
	// begin inline asm
	{  cvt.f32.f16 %f46, %rs91;}

	// end inline asm
	fma.rn.ftz.f32 	%f145, %f144, %f46, %f687;
	mov.b32 	{%rs92, %rs95}, %r68;
	// begin inline asm
	{  cvt.f32.f16 %f47, %rs92;}

	// end inline asm
	fma.rn.ftz.f32 	%f146, %f144, %f47, %f686;
	mov.b32 	{%rs93, %rs96}, %r74;
	// begin inline asm
	{  cvt.f32.f16 %f48, %rs93;}

	// end inline asm
	fma.rn.ftz.f32 	%f147, %f144, %f48, %f685;
	cvt.rn.f32.s16 	%f148, %rs8;
	sub.ftz.f32 	%f149, %f148, %f5;
	mul.ftz.f32 	%f150, %f45, %f149;
	// begin inline asm
	{  cvt.f32.f16 %f49, %rs94;}

	// end inline asm
	fma.rn.ftz.f32 	%f151, %f150, %f49, %f145;
	// begin inline asm
	{  cvt.f32.f16 %f50, %rs95;}

	// end inline asm
	fma.rn.ftz.f32 	%f152, %f150, %f50, %f146;
	// begin inline asm
	{  cvt.f32.f16 %f51, %rs96;}

	// end inline asm
	fma.rn.ftz.f32 	%f153, %f150, %f51, %f147;
	cvt.rn.f32.s16 	%f154, %rs10;
	sub.ftz.f32 	%f155, %f154, %f5;
	mul.ftz.f32 	%f156, %f45, %f155;
	mov.b32 	{%rs97, %rs100}, %r65;
	// begin inline asm
	{  cvt.f32.f16 %f52, %rs97;}

	// end inline asm
	fma.rn.ftz.f32 	%f157, %f156, %f52, %f151;
	mov.b32 	{%rs98, %rs101}, %r69;
	// begin inline asm
	{  cvt.f32.f16 %f53, %rs98;}

	// end inline asm
	fma.rn.ftz.f32 	%f158, %f156, %f53, %f152;
	mov.b32 	{%rs99, %rs102}, %r75;
	// begin inline asm
	{  cvt.f32.f16 %f54, %rs99;}

	// end inline asm
	fma.rn.ftz.f32 	%f159, %f156, %f54, %f153;
	cvt.rn.f32.s16 	%f160, %rs12;
	sub.ftz.f32 	%f161, %f160, %f5;
	mul.ftz.f32 	%f162, %f45, %f161;
	// begin inline asm
	{  cvt.f32.f16 %f55, %rs100;}

	// end inline asm
	fma.rn.ftz.f32 	%f163, %f162, %f55, %f157;
	// begin inline asm
	{  cvt.f32.f16 %f56, %rs101;}

	// end inline asm
	fma.rn.ftz.f32 	%f164, %f162, %f56, %f158;
	// begin inline asm
	{  cvt.f32.f16 %f57, %rs102;}

	// end inline asm
	fma.rn.ftz.f32 	%f165, %f162, %f57, %f159;
	cvt.rn.f32.s16 	%f166, %rs14;
	sub.ftz.f32 	%f167, %f166, %f5;
	mul.ftz.f32 	%f168, %f45, %f167;
	mov.b32 	{%rs103, %rs106}, %r66;
	// begin inline asm
	{  cvt.f32.f16 %f58, %rs103;}

	// end inline asm
	fma.rn.ftz.f32 	%f169, %f168, %f58, %f163;
	mov.b32 	{%rs104, %rs107}, %r70;
	// begin inline asm
	{  cvt.f32.f16 %f59, %rs104;}

	// end inline asm
	fma.rn.ftz.f32 	%f170, %f168, %f59, %f164;
	mov.b32 	{%rs105, %rs108}, %r76;
	// begin inline asm
	{  cvt.f32.f16 %f60, %rs105;}

	// end inline asm
	fma.rn.ftz.f32 	%f171, %f168, %f60, %f165;
	cvt.rn.f32.s16 	%f172, %rs16;
	sub.ftz.f32 	%f173, %f172, %f5;
	mul.ftz.f32 	%f174, %f45, %f173;
	// begin inline asm
	{  cvt.f32.f16 %f61, %rs106;}

	// end inline asm
	fma.rn.ftz.f32 	%f175, %f174, %f61, %f169;
	// begin inline asm
	{  cvt.f32.f16 %f62, %rs107;}

	// end inline asm
	fma.rn.ftz.f32 	%f176, %f174, %f62, %f170;
	// begin inline asm
	{  cvt.f32.f16 %f63, %rs108;}

	// end inline asm
	fma.rn.ftz.f32 	%f177, %f174, %f63, %f171;
	cvt.rn.f32.s16 	%f178, %rs18;
	sub.ftz.f32 	%f179, %f178, %f5;
	mul.ftz.f32 	%f180, %f45, %f179;
	mov.b32 	{%rs109, %rs112}, %r67;
	// begin inline asm
	{  cvt.f32.f16 %f64, %rs109;}

	// end inline asm
	fma.rn.ftz.f32 	%f181, %f180, %f64, %f175;
	mov.b32 	{%rs110, %rs113}, %r71;
	// begin inline asm
	{  cvt.f32.f16 %f65, %rs110;}

	// end inline asm
	fma.rn.ftz.f32 	%f182, %f180, %f65, %f176;
	mov.b32 	{%rs111, %rs114}, %r77;
	// begin inline asm
	{  cvt.f32.f16 %f66, %rs111;}

	// end inline asm
	fma.rn.ftz.f32 	%f183, %f180, %f66, %f177;
	cvt.rn.f32.s16 	%f184, %rs19;
	sub.ftz.f32 	%f185, %f184, %f5;
	mul.ftz.f32 	%f186, %f45, %f185;
	// begin inline asm
	{  cvt.f32.f16 %f67, %rs112;}

	// end inline asm
	fma.rn.ftz.f32 	%f187, %f186, %f67, %f181;
	// begin inline asm
	{  cvt.f32.f16 %f68, %rs113;}

	// end inline asm
	fma.rn.ftz.f32 	%f188, %f186, %f68, %f182;
	// begin inline asm
	{  cvt.f32.f16 %f69, %rs114;}

	// end inline asm
	fma.rn.ftz.f32 	%f189, %f186, %f69, %f183;
	ld.global.v4.u32 	{%r107, %r108, %r109, %r110}, [%rd7+16];
	ld.global.v4.u32 	{%r115, %r116, %r117, %r118}, [%rd9];
	ld.global.v4.u32 	{%r123, %r124, %r125, %r126}, [%rd8+16];
	cvt.rn.f32.s16 	%f190, %rs21;
	sub.ftz.f32 	%f191, %f190, %f5;
	mul.ftz.f32 	%f192, %f45, %f191;
	mov.b32 	{%rs115, %rs118}, %r107;
	// begin inline asm
	{  cvt.f32.f16 %f70, %rs115;}

	// end inline asm
	fma.rn.ftz.f32 	%f193, %f192, %f70, %f187;
	mov.b32 	{%rs116, %rs119}, %r115;
	// begin inline asm
	{  cvt.f32.f16 %f71, %rs116;}

	// end inline asm
	fma.rn.ftz.f32 	%f194, %f192, %f71, %f188;
	mov.b32 	{%rs117, %rs120}, %r123;
	// begin inline asm
	{  cvt.f32.f16 %f72, %rs117;}

	// end inline asm
	fma.rn.ftz.f32 	%f195, %f192, %f72, %f189;
	cvt.rn.f32.s16 	%f196, %rs23;
	sub.ftz.f32 	%f197, %f196, %f5;
	mul.ftz.f32 	%f198, %f45, %f197;
	// begin inline asm
	{  cvt.f32.f16 %f73, %rs118;}

	// end inline asm
	fma.rn.ftz.f32 	%f199, %f198, %f73, %f193;
	// begin inline asm
	{  cvt.f32.f16 %f74, %rs119;}

	// end inline asm
	fma.rn.ftz.f32 	%f200, %f198, %f74, %f194;
	// begin inline asm
	{  cvt.f32.f16 %f75, %rs120;}

	// end inline asm
	fma.rn.ftz.f32 	%f201, %f198, %f75, %f195;
	cvt.rn.f32.s16 	%f202, %rs25;
	sub.ftz.f32 	%f203, %f202, %f5;
	mul.ftz.f32 	%f204, %f45, %f203;
	mov.b32 	{%rs121, %rs124}, %r108;
	// begin inline asm
	{  cvt.f32.f16 %f76, %rs121;}

	// end inline asm
	fma.rn.ftz.f32 	%f205, %f204, %f76, %f199;
	mov.b32 	{%rs122, %rs125}, %r116;
	// begin inline asm
	{  cvt.f32.f16 %f77, %rs122;}

	// end inline asm
	fma.rn.ftz.f32 	%f206, %f204, %f77, %f200;
	mov.b32 	{%rs123, %rs126}, %r124;
	// begin inline asm
	{  cvt.f32.f16 %f78, %rs123;}

	// end inline asm
	fma.rn.ftz.f32 	%f207, %f204, %f78, %f201;
	cvt.rn.f32.s16 	%f208, %rs27;
	sub.ftz.f32 	%f209, %f208, %f5;
	mul.ftz.f32 	%f210, %f45, %f209;
	// begin inline asm
	{  cvt.f32.f16 %f79, %rs124;}

	// end inline asm
	fma.rn.ftz.f32 	%f211, %f210, %f79, %f205;
	// begin inline asm
	{  cvt.f32.f16 %f80, %rs125;}

	// end inline asm
	fma.rn.ftz.f32 	%f212, %f210, %f80, %f206;
	// begin inline asm
	{  cvt.f32.f16 %f81, %rs126;}

	// end inline asm
	fma.rn.ftz.f32 	%f213, %f210, %f81, %f207;
	cvt.rn.f32.s16 	%f214, %rs29;
	sub.ftz.f32 	%f215, %f214, %f5;
	mul.ftz.f32 	%f216, %f45, %f215;
	mov.b32 	{%rs127, %rs130}, %r109;
	// begin inline asm
	{  cvt.f32.f16 %f82, %rs127;}

	// end inline asm
	fma.rn.ftz.f32 	%f217, %f216, %f82, %f211;
	mov.b32 	{%rs128, %rs131}, %r117;
	// begin inline asm
	{  cvt.f32.f16 %f83, %rs128;}

	// end inline asm
	fma.rn.ftz.f32 	%f218, %f216, %f83, %f212;
	mov.b32 	{%rs129, %rs132}, %r125;
	// begin inline asm
	{  cvt.f32.f16 %f84, %rs129;}

	// end inline asm
	fma.rn.ftz.f32 	%f219, %f216, %f84, %f213;
	cvt.rn.f32.s16 	%f220, %rs31;
	sub.ftz.f32 	%f221, %f220, %f5;
	mul.ftz.f32 	%f222, %f45, %f221;
	// begin inline asm
	{  cvt.f32.f16 %f85, %rs130;}

	// end inline asm
	fma.rn.ftz.f32 	%f223, %f222, %f85, %f217;
	// begin inline asm
	{  cvt.f32.f16 %f86, %rs131;}

	// end inline asm
	fma.rn.ftz.f32 	%f224, %f222, %f86, %f218;
	// begin inline asm
	{  cvt.f32.f16 %f87, %rs132;}

	// end inline asm
	fma.rn.ftz.f32 	%f225, %f222, %f87, %f219;
	cvt.rn.f32.s16 	%f226, %rs33;
	sub.ftz.f32 	%f227, %f226, %f5;
	mul.ftz.f32 	%f228, %f45, %f227;
	mov.b32 	{%rs133, %rs136}, %r110;
	// begin inline asm
	{  cvt.f32.f16 %f88, %rs133;}

	// end inline asm
	fma.rn.ftz.f32 	%f229, %f228, %f88, %f223;
	mov.b32 	{%rs134, %rs137}, %r118;
	// begin inline asm
	{  cvt.f32.f16 %f89, %rs134;}

	// end inline asm
	fma.rn.ftz.f32 	%f230, %f228, %f89, %f224;
	mov.b32 	{%rs135, %rs138}, %r126;
	// begin inline asm
	{  cvt.f32.f16 %f90, %rs135;}

	// end inline asm
	fma.rn.ftz.f32 	%f231, %f228, %f90, %f225;
	cvt.rn.f32.s16 	%f232, %rs34;
	sub.ftz.f32 	%f233, %f232, %f5;
	mul.ftz.f32 	%f234, %f45, %f233;
	// begin inline asm
	{  cvt.f32.f16 %f91, %rs136;}

	// end inline asm
	fma.rn.ftz.f32 	%f235, %f234, %f91, %f229;
	// begin inline asm
	{  cvt.f32.f16 %f92, %rs137;}

	// end inline asm
	fma.rn.ftz.f32 	%f236, %f234, %f92, %f230;
	// begin inline asm
	{  cvt.f32.f16 %f93, %rs138;}

	// end inline asm
	fma.rn.ftz.f32 	%f237, %f234, %f93, %f231;
	ld.global.v4.u32 	{%r131, %r132, %r133, %r134}, [%rd7+32];
	ld.global.v4.u32 	{%r139, %r140, %r141, %r142}, [%rd9+16];
	ld.global.v4.u32 	{%r147, %r148, %r149, %r150}, [%rd8+32];
	cvt.rn.f32.s16 	%f238, %rs36;
	sub.ftz.f32 	%f239, %f238, %f5;
	mul.ftz.f32 	%f240, %f45, %f239;
	mov.b32 	{%rs139, %rs142}, %r131;
	// begin inline asm
	{  cvt.f32.f16 %f94, %rs139;}

	// end inline asm
	fma.rn.ftz.f32 	%f241, %f240, %f94, %f235;
	mov.b32 	{%rs140, %rs143}, %r139;
	// begin inline asm
	{  cvt.f32.f16 %f95, %rs140;}

	// end inline asm
	fma.rn.ftz.f32 	%f242, %f240, %f95, %f236;
	mov.b32 	{%rs141, %rs144}, %r147;
	// begin inline asm
	{  cvt.f32.f16 %f96, %rs141;}

	// end inline asm
	fma.rn.ftz.f32 	%f243, %f240, %f96, %f237;
	cvt.rn.f32.s16 	%f244, %rs38;
	sub.ftz.f32 	%f245, %f244, %f5;
	mul.ftz.f32 	%f246, %f45, %f245;
	// begin inline asm
	{  cvt.f32.f16 %f97, %rs142;}

	// end inline asm
	fma.rn.ftz.f32 	%f247, %f246, %f97, %f241;
	// begin inline asm
	{  cvt.f32.f16 %f98, %rs143;}

	// end inline asm
	fma.rn.ftz.f32 	%f248, %f246, %f98, %f242;
	// begin inline asm
	{  cvt.f32.f16 %f99, %rs144;}

	// end inline asm
	fma.rn.ftz.f32 	%f249, %f246, %f99, %f243;
	cvt.rn.f32.s16 	%f250, %rs40;
	sub.ftz.f32 	%f251, %f250, %f5;
	mul.ftz.f32 	%f252, %f45, %f251;
	mov.b32 	{%rs145, %rs148}, %r132;
	// begin inline asm
	{  cvt.f32.f16 %f100, %rs145;}

	// end inline asm
	fma.rn.ftz.f32 	%f253, %f252, %f100, %f247;
	mov.b32 	{%rs146, %rs149}, %r140;
	// begin inline asm
	{  cvt.f32.f16 %f101, %rs146;}

	// end inline asm
	fma.rn.ftz.f32 	%f254, %f252, %f101, %f248;
	mov.b32 	{%rs147, %rs150}, %r148;
	// begin inline asm
	{  cvt.f32.f16 %f102, %rs147;}

	// end inline asm
	fma.rn.ftz.f32 	%f255, %f252, %f102, %f249;
	cvt.rn.f32.s16 	%f256, %rs42;
	sub.ftz.f32 	%f257, %f256, %f5;
	mul.ftz.f32 	%f258, %f45, %f257;
	// begin inline asm
	{  cvt.f32.f16 %f103, %rs148;}

	// end inline asm
	fma.rn.ftz.f32 	%f259, %f258, %f103, %f253;
	// begin inline asm
	{  cvt.f32.f16 %f104, %rs149;}

	// end inline asm
	fma.rn.ftz.f32 	%f260, %f258, %f104, %f254;
	// begin inline asm
	{  cvt.f32.f16 %f105, %rs150;}

	// end inline asm
	fma.rn.ftz.f32 	%f261, %f258, %f105, %f255;
	cvt.rn.f32.s16 	%f262, %rs44;
	sub.ftz.f32 	%f263, %f262, %f5;
	mul.ftz.f32 	%f264, %f45, %f263;
	mov.b32 	{%rs151, %rs154}, %r133;
	// begin inline asm
	{  cvt.f32.f16 %f106, %rs151;}

	// end inline asm
	fma.rn.ftz.f32 	%f265, %f264, %f106, %f259;
	mov.b32 	{%rs152, %rs155}, %r141;
	// begin inline asm
	{  cvt.f32.f16 %f107, %rs152;}

	// end inline asm
	fma.rn.ftz.f32 	%f266, %f264, %f107, %f260;
	mov.b32 	{%rs153, %rs156}, %r149;
	// begin inline asm
	{  cvt.f32.f16 %f108, %rs153;}

	// end inline asm
	fma.rn.ftz.f32 	%f267, %f264, %f108, %f261;
	cvt.rn.f32.s16 	%f268, %rs46;
	sub.ftz.f32 	%f269, %f268, %f5;
	mul.ftz.f32 	%f270, %f45, %f269;
	// begin inline asm
	{  cvt.f32.f16 %f109, %rs154;}

	// end inline asm
	fma.rn.ftz.f32 	%f271, %f270, %f109, %f265;
	// begin inline asm
	{  cvt.f32.f16 %f110, %rs155;}

	// end inline asm
	fma.rn.ftz.f32 	%f272, %f270, %f110, %f266;
	// begin inline asm
	{  cvt.f32.f16 %f111, %rs156;}

	// end inline asm
	fma.rn.ftz.f32 	%f273, %f270, %f111, %f267;
	cvt.rn.f32.s16 	%f274, %rs48;
	sub.ftz.f32 	%f275, %f274, %f5;
	mul.ftz.f32 	%f276, %f45, %f275;
	mov.b32 	{%rs157, %rs160}, %r134;
	// begin inline asm
	{  cvt.f32.f16 %f112, %rs157;}

	// end inline asm
	fma.rn.ftz.f32 	%f277, %f276, %f112, %f271;
	mov.b32 	{%rs158, %rs161}, %r142;
	// begin inline asm
	{  cvt.f32.f16 %f113, %rs158;}

	// end inline asm
	fma.rn.ftz.f32 	%f278, %f276, %f113, %f272;
	mov.b32 	{%rs159, %rs162}, %r150;
	// begin inline asm
	{  cvt.f32.f16 %f114, %rs159;}

	// end inline asm
	fma.rn.ftz.f32 	%f279, %f276, %f114, %f273;
	cvt.rn.f32.s16 	%f280, %rs49;
	sub.ftz.f32 	%f281, %f280, %f5;
	mul.ftz.f32 	%f282, %f45, %f281;
	// begin inline asm
	{  cvt.f32.f16 %f115, %rs160;}

	// end inline asm
	fma.rn.ftz.f32 	%f283, %f282, %f115, %f277;
	// begin inline asm
	{  cvt.f32.f16 %f116, %rs161;}

	// end inline asm
	fma.rn.ftz.f32 	%f284, %f282, %f116, %f278;
	// begin inline asm
	{  cvt.f32.f16 %f117, %rs162;}

	// end inline asm
	fma.rn.ftz.f32 	%f285, %f282, %f117, %f279;
	ld.global.v4.u32 	{%r155, %r156, %r157, %r158}, [%rd7+48];
	ld.global.v4.u32 	{%r163, %r164, %r165, %r166}, [%rd9+32];
	ld.global.v4.u32 	{%r171, %r172, %r173, %r174}, [%rd8+48];
	cvt.rn.f32.s16 	%f286, %rs51;
	sub.ftz.f32 	%f287, %f286, %f5;
	mul.ftz.f32 	%f288, %f45, %f287;
	mov.b32 	{%rs163, %rs166}, %r155;
	// begin inline asm
	{  cvt.f32.f16 %f118, %rs163;}

	// end inline asm
	fma.rn.ftz.f32 	%f289, %f288, %f118, %f283;
	mov.b32 	{%rs164, %rs167}, %r163;
	// begin inline asm
	{  cvt.f32.f16 %f119, %rs164;}

	// end inline asm
	fma.rn.ftz.f32 	%f290, %f288, %f119, %f284;
	mov.b32 	{%rs165, %rs168}, %r171;
	// begin inline asm
	{  cvt.f32.f16 %f120, %rs165;}

	// end inline asm
	fma.rn.ftz.f32 	%f291, %f288, %f120, %f285;
	cvt.rn.f32.s16 	%f292, %rs53;
	sub.ftz.f32 	%f293, %f292, %f5;
	mul.ftz.f32 	%f294, %f45, %f293;
	// begin inline asm
	{  cvt.f32.f16 %f121, %rs166;}

	// end inline asm
	fma.rn.ftz.f32 	%f295, %f294, %f121, %f289;
	// begin inline asm
	{  cvt.f32.f16 %f122, %rs167;}

	// end inline asm
	fma.rn.ftz.f32 	%f296, %f294, %f122, %f290;
	// begin inline asm
	{  cvt.f32.f16 %f123, %rs168;}

	// end inline asm
	fma.rn.ftz.f32 	%f297, %f294, %f123, %f291;
	cvt.rn.f32.s16 	%f298, %rs55;
	sub.ftz.f32 	%f299, %f298, %f5;
	mul.ftz.f32 	%f300, %f45, %f299;
	mov.b32 	{%rs169, %rs172}, %r156;
	// begin inline asm
	{  cvt.f32.f16 %f124, %rs169;}

	// end inline asm
	fma.rn.ftz.f32 	%f301, %f300, %f124, %f295;
	mov.b32 	{%rs170, %rs173}, %r164;
	// begin inline asm
	{  cvt.f32.f16 %f125, %rs170;}

	// end inline asm
	fma.rn.ftz.f32 	%f302, %f300, %f125, %f296;
	mov.b32 	{%rs171, %rs174}, %r172;
	// begin inline asm
	{  cvt.f32.f16 %f126, %rs171;}

	// end inline asm
	fma.rn.ftz.f32 	%f303, %f300, %f126, %f297;
	cvt.rn.f32.s16 	%f304, %rs57;
	sub.ftz.f32 	%f305, %f304, %f5;
	mul.ftz.f32 	%f306, %f45, %f305;
	// begin inline asm
	{  cvt.f32.f16 %f127, %rs172;}

	// end inline asm
	fma.rn.ftz.f32 	%f307, %f306, %f127, %f301;
	// begin inline asm
	{  cvt.f32.f16 %f128, %rs173;}

	// end inline asm
	fma.rn.ftz.f32 	%f308, %f306, %f128, %f302;
	// begin inline asm
	{  cvt.f32.f16 %f129, %rs174;}

	// end inline asm
	fma.rn.ftz.f32 	%f309, %f306, %f129, %f303;
	cvt.rn.f32.s16 	%f310, %rs59;
	sub.ftz.f32 	%f311, %f310, %f5;
	mul.ftz.f32 	%f312, %f45, %f311;
	mov.b32 	{%rs175, %rs178}, %r157;
	// begin inline asm
	{  cvt.f32.f16 %f130, %rs175;}

	// end inline asm
	fma.rn.ftz.f32 	%f313, %f312, %f130, %f307;
	mov.b32 	{%rs176, %rs179}, %r165;
	// begin inline asm
	{  cvt.f32.f16 %f131, %rs176;}

	// end inline asm
	fma.rn.ftz.f32 	%f314, %f312, %f131, %f308;
	mov.b32 	{%rs177, %rs180}, %r173;
	// begin inline asm
	{  cvt.f32.f16 %f132, %rs177;}

	// end inline asm
	fma.rn.ftz.f32 	%f315, %f312, %f132, %f309;
	cvt.rn.f32.s16 	%f316, %rs61;
	sub.ftz.f32 	%f317, %f316, %f5;
	mul.ftz.f32 	%f318, %f45, %f317;
	// begin inline asm
	{  cvt.f32.f16 %f133, %rs178;}

	// end inline asm
	fma.rn.ftz.f32 	%f319, %f318, %f133, %f313;
	// begin inline asm
	{  cvt.f32.f16 %f134, %rs179;}

	// end inline asm
	fma.rn.ftz.f32 	%f320, %f318, %f134, %f314;
	// begin inline asm
	{  cvt.f32.f16 %f135, %rs180;}

	// end inline asm
	fma.rn.ftz.f32 	%f321, %f318, %f135, %f315;
	cvt.rn.f32.s16 	%f322, %rs63;
	sub.ftz.f32 	%f323, %f322, %f5;
	mul.ftz.f32 	%f324, %f45, %f323;
	mov.b32 	{%rs181, %rs184}, %r158;
	// begin inline asm
	{  cvt.f32.f16 %f136, %rs181;}

	// end inline asm
	fma.rn.ftz.f32 	%f325, %f324, %f136, %f319;
	mov.b32 	{%rs182, %rs185}, %r166;
	// begin inline asm
	{  cvt.f32.f16 %f137, %rs182;}

	// end inline asm
	fma.rn.ftz.f32 	%f326, %f324, %f137, %f320;
	mov.b32 	{%rs183, %rs186}, %r174;
	// begin inline asm
	{  cvt.f32.f16 %f138, %rs183;}

	// end inline asm
	fma.rn.ftz.f32 	%f327, %f324, %f138, %f321;
	cvt.rn.f32.s16 	%f328, %rs64;
	sub.ftz.f32 	%f329, %f328, %f5;
	mul.ftz.f32 	%f330, %f45, %f329;
	// begin inline asm
	{  cvt.f32.f16 %f139, %rs184;}

	// end inline asm
	fma.rn.ftz.f32 	%f687, %f330, %f139, %f325;
	// begin inline asm
	{  cvt.f32.f16 %f140, %rs185;}

	// end inline asm
	fma.rn.ftz.f32 	%f686, %f330, %f140, %f326;
	// begin inline asm
	{  cvt.f32.f16 %f141, %rs186;}

	// end inline asm
	fma.rn.ftz.f32 	%f685, %f330, %f141, %f327;
	bra.uni 	$L__BB0_8;

$L__BB0_7:
	shl.b16 	%rs283, %rs5, 4;
	cvt.s16.s8 	%rs284, %rs283;
	shr.s16 	%rs285, %rs284, 7;
	and.b16  	%rs286, %rs285, -16;
	or.b16  	%rs287, %rs286, %rs6;
	cvt.rn.f32.s16 	%f427, %rs287;
	sub.ftz.f32 	%f428, %f427, %f5;
	mul.ftz.f32 	%f429, %f45, %f428;
	mov.b32 	{%rs187, %rs190}, %r64;
	// begin inline asm
	{  cvt.f32.f16 %f331, %rs187;}

	// end inline asm
	fma.rn.ftz.f32 	%f430, %f429, %f331, %f687;
	mov.b32 	{%rs188, %rs191}, %r68;
	// begin inline asm
	{  cvt.f32.f16 %f332, %rs188;}

	// end inline asm
	fma.rn.ftz.f32 	%f431, %f429, %f332, %f686;
	mov.b32 	{%rs189, %rs192}, %r74;
	// begin inline asm
	{  cvt.f32.f16 %f333, %rs189;}

	// end inline asm
	fma.rn.ftz.f32 	%f432, %f429, %f333, %f685;
	shl.b16 	%rs288, %rs7, 4;
	cvt.s16.s8 	%rs289, %rs288;
	shr.s16 	%rs290, %rs289, 7;
	and.b16  	%rs291, %rs290, -16;
	or.b16  	%rs292, %rs291, %rs8;
	cvt.rn.f32.s16 	%f433, %rs292;
	sub.ftz.f32 	%f434, %f433, %f5;
	mul.ftz.f32 	%f435, %f45, %f434;
	// begin inline asm
	{  cvt.f32.f16 %f334, %rs190;}

	// end inline asm
	fma.rn.ftz.f32 	%f436, %f435, %f334, %f430;
	// begin inline asm
	{  cvt.f32.f16 %f335, %rs191;}

	// end inline asm
	fma.rn.ftz.f32 	%f437, %f435, %f335, %f431;
	// begin inline asm
	{  cvt.f32.f16 %f336, %rs192;}

	// end inline asm
	fma.rn.ftz.f32 	%f438, %f435, %f336, %f432;
	shl.b16 	%rs293, %rs9, 4;
	cvt.s16.s8 	%rs294, %rs293;
	shr.s16 	%rs295, %rs294, 7;
	and.b16  	%rs296, %rs295, -16;
	or.b16  	%rs297, %rs296, %rs10;
	cvt.rn.f32.s16 	%f439, %rs297;
	sub.ftz.f32 	%f440, %f439, %f5;
	mul.ftz.f32 	%f441, %f45, %f440;
	mov.b32 	{%rs193, %rs196}, %r65;
	// begin inline asm
	{  cvt.f32.f16 %f337, %rs193;}

	// end inline asm
	fma.rn.ftz.f32 	%f442, %f441, %f337, %f436;
	mov.b32 	{%rs194, %rs197}, %r69;
	// begin inline asm
	{  cvt.f32.f16 %f338, %rs194;}

	// end inline asm
	fma.rn.ftz.f32 	%f443, %f441, %f338, %f437;
	mov.b32 	{%rs195, %rs198}, %r75;
	// begin inline asm
	{  cvt.f32.f16 %f339, %rs195;}

	// end inline asm
	fma.rn.ftz.f32 	%f444, %f441, %f339, %f438;
	shl.b16 	%rs298, %rs11, 4;
	cvt.s16.s8 	%rs299, %rs298;
	shr.s16 	%rs300, %rs299, 7;
	and.b16  	%rs301, %rs300, -16;
	or.b16  	%rs302, %rs301, %rs12;
	cvt.rn.f32.s16 	%f445, %rs302;
	sub.ftz.f32 	%f446, %f445, %f5;
	mul.ftz.f32 	%f447, %f45, %f446;
	// begin inline asm
	{  cvt.f32.f16 %f340, %rs196;}

	// end inline asm
	fma.rn.ftz.f32 	%f448, %f447, %f340, %f442;
	// begin inline asm
	{  cvt.f32.f16 %f341, %rs197;}

	// end inline asm
	fma.rn.ftz.f32 	%f449, %f447, %f341, %f443;
	// begin inline asm
	{  cvt.f32.f16 %f342, %rs198;}

	// end inline asm
	fma.rn.ftz.f32 	%f450, %f447, %f342, %f444;
	shl.b16 	%rs303, %rs13, 4;
	cvt.s16.s8 	%rs304, %rs303;
	shr.s16 	%rs305, %rs304, 7;
	and.b16  	%rs306, %rs305, -16;
	or.b16  	%rs307, %rs306, %rs14;
	cvt.rn.f32.s16 	%f451, %rs307;
	sub.ftz.f32 	%f452, %f451, %f5;
	mul.ftz.f32 	%f453, %f45, %f452;
	mov.b32 	{%rs199, %rs202}, %r66;
	// begin inline asm
	{  cvt.f32.f16 %f343, %rs199;}

	// end inline asm
	fma.rn.ftz.f32 	%f454, %f453, %f343, %f448;
	mov.b32 	{%rs200, %rs203}, %r70;
	// begin inline asm
	{  cvt.f32.f16 %f344, %rs200;}

	// end inline asm
	fma.rn.ftz.f32 	%f455, %f453, %f344, %f449;
	mov.b32 	{%rs201, %rs204}, %r76;
	// begin inline asm
	{  cvt.f32.f16 %f345, %rs201;}

	// end inline asm
	fma.rn.ftz.f32 	%f456, %f453, %f345, %f450;
	shl.b16 	%rs308, %rs15, 4;
	cvt.s16.s8 	%rs309, %rs308;
	shr.s16 	%rs310, %rs309, 7;
	and.b16  	%rs311, %rs310, -16;
	or.b16  	%rs312, %rs311, %rs16;
	cvt.rn.f32.s16 	%f457, %rs312;
	sub.ftz.f32 	%f458, %f457, %f5;
	mul.ftz.f32 	%f459, %f45, %f458;
	// begin inline asm
	{  cvt.f32.f16 %f346, %rs202;}

	// end inline asm
	fma.rn.ftz.f32 	%f460, %f459, %f346, %f454;
	// begin inline asm
	{  cvt.f32.f16 %f347, %rs203;}

	// end inline asm
	fma.rn.ftz.f32 	%f461, %f459, %f347, %f455;
	// begin inline asm
	{  cvt.f32.f16 %f348, %rs204;}

	// end inline asm
	fma.rn.ftz.f32 	%f462, %f459, %f348, %f456;
	shl.b16 	%rs313, %rs17, 4;
	cvt.s16.s8 	%rs314, %rs313;
	shr.s16 	%rs315, %rs314, 7;
	and.b16  	%rs316, %rs315, -16;
	or.b16  	%rs317, %rs316, %rs18;
	cvt.rn.f32.s16 	%f463, %rs317;
	sub.ftz.f32 	%f464, %f463, %f5;
	mul.ftz.f32 	%f465, %f45, %f464;
	mov.b32 	{%rs205, %rs208}, %r67;
	// begin inline asm
	{  cvt.f32.f16 %f349, %rs205;}

	// end inline asm
	fma.rn.ftz.f32 	%f466, %f465, %f349, %f460;
	mov.b32 	{%rs206, %rs209}, %r71;
	// begin inline asm
	{  cvt.f32.f16 %f350, %rs206;}

	// end inline asm
	fma.rn.ftz.f32 	%f467, %f465, %f350, %f461;
	mov.b32 	{%rs207, %rs210}, %r77;
	// begin inline asm
	{  cvt.f32.f16 %f351, %rs207;}

	// end inline asm
	fma.rn.ftz.f32 	%f468, %f465, %f351, %f462;
	shl.b16 	%rs318, %rs19, 4;
	cvt.s16.s8 	%rs319, %rs318;
	shr.s16 	%rs320, %rs319, 7;
	and.b16  	%rs321, %rs320, -16;
	or.b16  	%rs322, %rs321, %rs19;
	cvt.rn.f32.s16 	%f469, %rs322;
	sub.ftz.f32 	%f470, %f469, %f5;
	mul.ftz.f32 	%f471, %f45, %f470;
	// begin inline asm
	{  cvt.f32.f16 %f352, %rs208;}

	// end inline asm
	fma.rn.ftz.f32 	%f472, %f471, %f352, %f466;
	// begin inline asm
	{  cvt.f32.f16 %f353, %rs209;}

	// end inline asm
	fma.rn.ftz.f32 	%f473, %f471, %f353, %f467;
	// begin inline asm
	{  cvt.f32.f16 %f354, %rs210;}

	// end inline asm
	fma.rn.ftz.f32 	%f474, %f471, %f354, %f468;
	ld.global.v4.u32 	{%r179, %r180, %r181, %r182}, [%rd7+16];
	ld.global.v4.u32 	{%r187, %r188, %r189, %r190}, [%rd9];
	ld.global.v4.u32 	{%r195, %r196, %r197, %r198}, [%rd8+16];
	shl.b16 	%rs323, %rs20, 4;
	cvt.s16.s8 	%rs324, %rs323;
	shr.s16 	%rs325, %rs324, 7;
	and.b16  	%rs326, %rs325, -16;
	or.b16  	%rs327, %rs326, %rs21;
	cvt.rn.f32.s16 	%f475, %rs327;
	sub.ftz.f32 	%f476, %f475, %f5;
	mul.ftz.f32 	%f477, %f45, %f476;
	mov.b32 	{%rs211, %rs214}, %r179;
	// begin inline asm
	{  cvt.f32.f16 %f355, %rs211;}

	// end inline asm
	fma.rn.ftz.f32 	%f478, %f477, %f355, %f472;
	mov.b32 	{%rs212, %rs215}, %r187;
	// begin inline asm
	{  cvt.f32.f16 %f356, %rs212;}

	// end inline asm
	fma.rn.ftz.f32 	%f479, %f477, %f356, %f473;
	mov.b32 	{%rs213, %rs216}, %r195;
	// begin inline asm
	{  cvt.f32.f16 %f357, %rs213;}

	// end inline asm
	fma.rn.ftz.f32 	%f480, %f477, %f357, %f474;
	shl.b16 	%rs328, %rs22, 4;
	cvt.s16.s8 	%rs329, %rs328;
	shr.s16 	%rs330, %rs329, 7;
	and.b16  	%rs331, %rs330, -16;
	or.b16  	%rs332, %rs331, %rs23;
	cvt.rn.f32.s16 	%f481, %rs332;
	sub.ftz.f32 	%f482, %f481, %f5;
	mul.ftz.f32 	%f483, %f45, %f482;
	// begin inline asm
	{  cvt.f32.f16 %f358, %rs214;}

	// end inline asm
	fma.rn.ftz.f32 	%f484, %f483, %f358, %f478;
	// begin inline asm
	{  cvt.f32.f16 %f359, %rs215;}

	// end inline asm
	fma.rn.ftz.f32 	%f485, %f483, %f359, %f479;
	// begin inline asm
	{  cvt.f32.f16 %f360, %rs216;}

	// end inline asm
	fma.rn.ftz.f32 	%f486, %f483, %f360, %f480;
	shl.b16 	%rs333, %rs24, 4;
	cvt.s16.s8 	%rs334, %rs333;
	shr.s16 	%rs335, %rs334, 7;
	and.b16  	%rs336, %rs335, -16;
	or.b16  	%rs337, %rs336, %rs25;
	cvt.rn.f32.s16 	%f487, %rs337;
	sub.ftz.f32 	%f488, %f487, %f5;
	mul.ftz.f32 	%f489, %f45, %f488;
	mov.b32 	{%rs217, %rs220}, %r180;
	// begin inline asm
	{  cvt.f32.f16 %f361, %rs217;}

	// end inline asm
	fma.rn.ftz.f32 	%f490, %f489, %f361, %f484;
	mov.b32 	{%rs218, %rs221}, %r188;
	// begin inline asm
	{  cvt.f32.f16 %f362, %rs218;}

	// end inline asm
	fma.rn.ftz.f32 	%f491, %f489, %f362, %f485;
	mov.b32 	{%rs219, %rs222}, %r196;
	// begin inline asm
	{  cvt.f32.f16 %f363, %rs219;}

	// end inline asm
	fma.rn.ftz.f32 	%f492, %f489, %f363, %f486;
	shl.b16 	%rs338, %rs26, 4;
	cvt.s16.s8 	%rs339, %rs338;
	shr.s16 	%rs340, %rs339, 7;
	and.b16  	%rs341, %rs340, -16;
	or.b16  	%rs342, %rs341, %rs27;
	cvt.rn.f32.s16 	%f493, %rs342;
	sub.ftz.f32 	%f494, %f493, %f5;
	mul.ftz.f32 	%f495, %f45, %f494;
	// begin inline asm
	{  cvt.f32.f16 %f364, %rs220;}

	// end inline asm
	fma.rn.ftz.f32 	%f496, %f495, %f364, %f490;
	// begin inline asm
	{  cvt.f32.f16 %f365, %rs221;}

	// end inline asm
	fma.rn.ftz.f32 	%f497, %f495, %f365, %f491;
	// begin inline asm
	{  cvt.f32.f16 %f366, %rs222;}

	// end inline asm
	fma.rn.ftz.f32 	%f498, %f495, %f366, %f492;
	shl.b16 	%rs343, %rs28, 4;
	cvt.s16.s8 	%rs344, %rs343;
	shr.s16 	%rs345, %rs344, 7;
	and.b16  	%rs346, %rs345, -16;
	or.b16  	%rs347, %rs346, %rs29;
	cvt.rn.f32.s16 	%f499, %rs347;
	sub.ftz.f32 	%f500, %f499, %f5;
	mul.ftz.f32 	%f501, %f45, %f500;
	mov.b32 	{%rs223, %rs226}, %r181;
	// begin inline asm
	{  cvt.f32.f16 %f367, %rs223;}

	// end inline asm
	fma.rn.ftz.f32 	%f502, %f501, %f367, %f496;
	mov.b32 	{%rs224, %rs227}, %r189;
	// begin inline asm
	{  cvt.f32.f16 %f368, %rs224;}

	// end inline asm
	fma.rn.ftz.f32 	%f503, %f501, %f368, %f497;
	mov.b32 	{%rs225, %rs228}, %r197;
	// begin inline asm
	{  cvt.f32.f16 %f369, %rs225;}

	// end inline asm
	fma.rn.ftz.f32 	%f504, %f501, %f369, %f498;
	shl.b16 	%rs348, %rs30, 4;
	cvt.s16.s8 	%rs349, %rs348;
	shr.s16 	%rs350, %rs349, 7;
	and.b16  	%rs351, %rs350, -16;
	or.b16  	%rs352, %rs351, %rs31;
	cvt.rn.f32.s16 	%f505, %rs352;
	sub.ftz.f32 	%f506, %f505, %f5;
	mul.ftz.f32 	%f507, %f45, %f506;
	// begin inline asm
	{  cvt.f32.f16 %f370, %rs226;}

	// end inline asm
	fma.rn.ftz.f32 	%f508, %f507, %f370, %f502;
	// begin inline asm
	{  cvt.f32.f16 %f371, %rs227;}

	// end inline asm
	fma.rn.ftz.f32 	%f509, %f507, %f371, %f503;
	// begin inline asm
	{  cvt.f32.f16 %f372, %rs228;}

	// end inline asm
	fma.rn.ftz.f32 	%f510, %f507, %f372, %f504;
	shl.b16 	%rs353, %rs32, 4;
	cvt.s16.s8 	%rs354, %rs353;
	shr.s16 	%rs355, %rs354, 7;
	and.b16  	%rs356, %rs355, -16;
	or.b16  	%rs357, %rs356, %rs33;
	cvt.rn.f32.s16 	%f511, %rs357;
	sub.ftz.f32 	%f512, %f511, %f5;
	mul.ftz.f32 	%f513, %f45, %f512;
	mov.b32 	{%rs229, %rs232}, %r182;
	// begin inline asm
	{  cvt.f32.f16 %f373, %rs229;}

	// end inline asm
	fma.rn.ftz.f32 	%f514, %f513, %f373, %f508;
	mov.b32 	{%rs230, %rs233}, %r190;
	// begin inline asm
	{  cvt.f32.f16 %f374, %rs230;}

	// end inline asm
	fma.rn.ftz.f32 	%f515, %f513, %f374, %f509;
	mov.b32 	{%rs231, %rs234}, %r198;
	// begin inline asm
	{  cvt.f32.f16 %f375, %rs231;}

	// end inline asm
	fma.rn.ftz.f32 	%f516, %f513, %f375, %f510;
	shl.b16 	%rs358, %rs34, 4;
	cvt.s16.s8 	%rs359, %rs358;
	shr.s16 	%rs360, %rs359, 7;
	and.b16  	%rs361, %rs360, -16;
	or.b16  	%rs362, %rs361, %rs34;
	cvt.rn.f32.s16 	%f517, %rs362;
	sub.ftz.f32 	%f518, %f517, %f5;
	mul.ftz.f32 	%f519, %f45, %f518;
	// begin inline asm
	{  cvt.f32.f16 %f376, %rs232;}

	// end inline asm
	fma.rn.ftz.f32 	%f520, %f519, %f376, %f514;
	// begin inline asm
	{  cvt.f32.f16 %f377, %rs233;}

	// end inline asm
	fma.rn.ftz.f32 	%f521, %f519, %f377, %f515;
	// begin inline asm
	{  cvt.f32.f16 %f378, %rs234;}

	// end inline asm
	fma.rn.ftz.f32 	%f522, %f519, %f378, %f516;
	ld.global.v4.u32 	{%r203, %r204, %r205, %r206}, [%rd7+32];
	ld.global.v4.u32 	{%r211, %r212, %r213, %r214}, [%rd9+16];
	ld.global.v4.u32 	{%r219, %r220, %r221, %r222}, [%rd8+32];
	shl.b16 	%rs363, %rs35, 4;
	cvt.s16.s8 	%rs364, %rs363;
	shr.s16 	%rs365, %rs364, 7;
	and.b16  	%rs366, %rs365, -16;
	or.b16  	%rs367, %rs366, %rs36;
	cvt.rn.f32.s16 	%f523, %rs367;
	sub.ftz.f32 	%f524, %f523, %f5;
	mul.ftz.f32 	%f525, %f45, %f524;
	mov.b32 	{%rs235, %rs238}, %r203;
	// begin inline asm
	{  cvt.f32.f16 %f379, %rs235;}

	// end inline asm
	fma.rn.ftz.f32 	%f526, %f525, %f379, %f520;
	mov.b32 	{%rs236, %rs239}, %r211;
	// begin inline asm
	{  cvt.f32.f16 %f380, %rs236;}

	// end inline asm
	fma.rn.ftz.f32 	%f527, %f525, %f380, %f521;
	mov.b32 	{%rs237, %rs240}, %r219;
	// begin inline asm
	{  cvt.f32.f16 %f381, %rs237;}

	// end inline asm
	fma.rn.ftz.f32 	%f528, %f525, %f381, %f522;
	shl.b16 	%rs368, %rs37, 4;
	cvt.s16.s8 	%rs369, %rs368;
	shr.s16 	%rs370, %rs369, 7;
	and.b16  	%rs371, %rs370, -16;
	or.b16  	%rs372, %rs371, %rs38;
	cvt.rn.f32.s16 	%f529, %rs372;
	sub.ftz.f32 	%f530, %f529, %f5;
	mul.ftz.f32 	%f531, %f45, %f530;
	// begin inline asm
	{  cvt.f32.f16 %f382, %rs238;}

	// end inline asm
	fma.rn.ftz.f32 	%f532, %f531, %f382, %f526;
	// begin inline asm
	{  cvt.f32.f16 %f383, %rs239;}

	// end inline asm
	fma.rn.ftz.f32 	%f533, %f531, %f383, %f527;
	// begin inline asm
	{  cvt.f32.f16 %f384, %rs240;}

	// end inline asm
	fma.rn.ftz.f32 	%f534, %f531, %f384, %f528;
	shl.b16 	%rs373, %rs39, 4;
	cvt.s16.s8 	%rs374, %rs373;
	shr.s16 	%rs375, %rs374, 7;
	and.b16  	%rs376, %rs375, -16;
	or.b16  	%rs377, %rs376, %rs40;
	cvt.rn.f32.s16 	%f535, %rs377;
	sub.ftz.f32 	%f536, %f535, %f5;
	mul.ftz.f32 	%f537, %f45, %f536;
	mov.b32 	{%rs241, %rs244}, %r204;
	// begin inline asm
	{  cvt.f32.f16 %f385, %rs241;}

	// end inline asm
	fma.rn.ftz.f32 	%f538, %f537, %f385, %f532;
	mov.b32 	{%rs242, %rs245}, %r212;
	// begin inline asm
	{  cvt.f32.f16 %f386, %rs242;}

	// end inline asm
	fma.rn.ftz.f32 	%f539, %f537, %f386, %f533;
	mov.b32 	{%rs243, %rs246}, %r220;
	// begin inline asm
	{  cvt.f32.f16 %f387, %rs243;}

	// end inline asm
	fma.rn.ftz.f32 	%f540, %f537, %f387, %f534;
	shl.b16 	%rs378, %rs41, 4;
	cvt.s16.s8 	%rs379, %rs378;
	shr.s16 	%rs380, %rs379, 7;
	and.b16  	%rs381, %rs380, -16;
	or.b16  	%rs382, %rs381, %rs42;
	cvt.rn.f32.s16 	%f541, %rs382;
	sub.ftz.f32 	%f542, %f541, %f5;
	mul.ftz.f32 	%f543, %f45, %f542;
	// begin inline asm
	{  cvt.f32.f16 %f388, %rs244;}

	// end inline asm
	fma.rn.ftz.f32 	%f544, %f543, %f388, %f538;
	// begin inline asm
	{  cvt.f32.f16 %f389, %rs245;}

	// end inline asm
	fma.rn.ftz.f32 	%f545, %f543, %f389, %f539;
	// begin inline asm
	{  cvt.f32.f16 %f390, %rs246;}

	// end inline asm
	fma.rn.ftz.f32 	%f546, %f543, %f390, %f540;
	shl.b16 	%rs383, %rs43, 4;
	cvt.s16.s8 	%rs384, %rs383;
	shr.s16 	%rs385, %rs384, 7;
	and.b16  	%rs386, %rs385, -16;
	or.b16  	%rs387, %rs386, %rs44;
	cvt.rn.f32.s16 	%f547, %rs387;
	sub.ftz.f32 	%f548, %f547, %f5;
	mul.ftz.f32 	%f549, %f45, %f548;
	mov.b32 	{%rs247, %rs250}, %r205;
	// begin inline asm
	{  cvt.f32.f16 %f391, %rs247;}

	// end inline asm
	fma.rn.ftz.f32 	%f550, %f549, %f391, %f544;
	mov.b32 	{%rs248, %rs251}, %r213;
	// begin inline asm
	{  cvt.f32.f16 %f392, %rs248;}

	// end inline asm
	fma.rn.ftz.f32 	%f551, %f549, %f392, %f545;
	mov.b32 	{%rs249, %rs252}, %r221;
	// begin inline asm
	{  cvt.f32.f16 %f393, %rs249;}

	// end inline asm
	fma.rn.ftz.f32 	%f552, %f549, %f393, %f546;
	shl.b16 	%rs388, %rs45, 4;
	cvt.s16.s8 	%rs389, %rs388;
	shr.s16 	%rs390, %rs389, 7;
	and.b16  	%rs391, %rs390, -16;
	or.b16  	%rs392, %rs391, %rs46;
	cvt.rn.f32.s16 	%f553, %rs392;
	sub.ftz.f32 	%f554, %f553, %f5;
	mul.ftz.f32 	%f555, %f45, %f554;
	// begin inline asm
	{  cvt.f32.f16 %f394, %rs250;}

	// end inline asm
	fma.rn.ftz.f32 	%f556, %f555, %f394, %f550;
	// begin inline asm
	{  cvt.f32.f16 %f395, %rs251;}

	// end inline asm
	fma.rn.ftz.f32 	%f557, %f555, %f395, %f551;
	// begin inline asm
	{  cvt.f32.f16 %f396, %rs252;}

	// end inline asm
	fma.rn.ftz.f32 	%f558, %f555, %f396, %f552;
	shl.b16 	%rs393, %rs47, 4;
	cvt.s16.s8 	%rs394, %rs393;
	shr.s16 	%rs395, %rs394, 7;
	and.b16  	%rs396, %rs395, -16;
	or.b16  	%rs397, %rs396, %rs48;
	cvt.rn.f32.s16 	%f559, %rs397;
	sub.ftz.f32 	%f560, %f559, %f5;
	mul.ftz.f32 	%f561, %f45, %f560;
	mov.b32 	{%rs253, %rs256}, %r206;
	// begin inline asm
	{  cvt.f32.f16 %f397, %rs253;}

	// end inline asm
	fma.rn.ftz.f32 	%f562, %f561, %f397, %f556;
	mov.b32 	{%rs254, %rs257}, %r214;
	// begin inline asm
	{  cvt.f32.f16 %f398, %rs254;}

	// end inline asm
	fma.rn.ftz.f32 	%f563, %f561, %f398, %f557;
	mov.b32 	{%rs255, %rs258}, %r222;
	// begin inline asm
	{  cvt.f32.f16 %f399, %rs255;}

	// end inline asm
	fma.rn.ftz.f32 	%f564, %f561, %f399, %f558;
	shl.b16 	%rs398, %rs49, 4;
	cvt.s16.s8 	%rs399, %rs398;
	shr.s16 	%rs400, %rs399, 7;
	and.b16  	%rs401, %rs400, -16;
	or.b16  	%rs402, %rs401, %rs49;
	cvt.rn.f32.s16 	%f565, %rs402;
	sub.ftz.f32 	%f566, %f565, %f5;
	mul.ftz.f32 	%f567, %f45, %f566;
	// begin inline asm
	{  cvt.f32.f16 %f400, %rs256;}

	// end inline asm
	fma.rn.ftz.f32 	%f568, %f567, %f400, %f562;
	// begin inline asm
	{  cvt.f32.f16 %f401, %rs257;}

	// end inline asm
	fma.rn.ftz.f32 	%f569, %f567, %f401, %f563;
	// begin inline asm
	{  cvt.f32.f16 %f402, %rs258;}

	// end inline asm
	fma.rn.ftz.f32 	%f570, %f567, %f402, %f564;
	ld.global.v4.u32 	{%r227, %r228, %r229, %r230}, [%rd7+48];
	ld.global.v4.u32 	{%r235, %r236, %r237, %r238}, [%rd9+32];
	ld.global.v4.u32 	{%r243, %r244, %r245, %r246}, [%rd8+48];
	shl.b16 	%rs403, %rs50, 4;
	cvt.s16.s8 	%rs404, %rs403;
	shr.s16 	%rs405, %rs404, 7;
	and.b16  	%rs406, %rs405, -16;
	or.b16  	%rs407, %rs406, %rs51;
	cvt.rn.f32.s16 	%f571, %rs407;
	sub.ftz.f32 	%f572, %f571, %f5;
	mul.ftz.f32 	%f573, %f45, %f572;
	mov.b32 	{%rs259, %rs262}, %r227;
	// begin inline asm
	{  cvt.f32.f16 %f403, %rs259;}

	// end inline asm
	fma.rn.ftz.f32 	%f574, %f573, %f403, %f568;
	mov.b32 	{%rs260, %rs263}, %r235;
	// begin inline asm
	{  cvt.f32.f16 %f404, %rs260;}

	// end inline asm
	fma.rn.ftz.f32 	%f575, %f573, %f404, %f569;
	mov.b32 	{%rs261, %rs264}, %r243;
	// begin inline asm
	{  cvt.f32.f16 %f405, %rs261;}

	// end inline asm
	fma.rn.ftz.f32 	%f576, %f573, %f405, %f570;
	shl.b16 	%rs408, %rs52, 4;
	cvt.s16.s8 	%rs409, %rs408;
	shr.s16 	%rs410, %rs409, 7;
	and.b16  	%rs411, %rs410, -16;
	or.b16  	%rs412, %rs411, %rs53;
	cvt.rn.f32.s16 	%f577, %rs412;
	sub.ftz.f32 	%f578, %f577, %f5;
	mul.ftz.f32 	%f579, %f45, %f578;
	// begin inline asm
	{  cvt.f32.f16 %f406, %rs262;}

	// end inline asm
	fma.rn.ftz.f32 	%f580, %f579, %f406, %f574;
	// begin inline asm
	{  cvt.f32.f16 %f407, %rs263;}

	// end inline asm
	fma.rn.ftz.f32 	%f581, %f579, %f407, %f575;
	// begin inline asm
	{  cvt.f32.f16 %f408, %rs264;}

	// end inline asm
	fma.rn.ftz.f32 	%f582, %f579, %f408, %f576;
	shl.b16 	%rs413, %rs54, 4;
	cvt.s16.s8 	%rs414, %rs413;
	shr.s16 	%rs415, %rs414, 7;
	and.b16  	%rs416, %rs415, -16;
	or.b16  	%rs417, %rs416, %rs55;
	cvt.rn.f32.s16 	%f583, %rs417;
	sub.ftz.f32 	%f584, %f583, %f5;
	mul.ftz.f32 	%f585, %f45, %f584;
	mov.b32 	{%rs265, %rs268}, %r228;
	// begin inline asm
	{  cvt.f32.f16 %f409, %rs265;}

	// end inline asm
	fma.rn.ftz.f32 	%f586, %f585, %f409, %f580;
	mov.b32 	{%rs266, %rs269}, %r236;
	// begin inline asm
	{  cvt.f32.f16 %f410, %rs266;}

	// end inline asm
	fma.rn.ftz.f32 	%f587, %f585, %f410, %f581;
	mov.b32 	{%rs267, %rs270}, %r244;
	// begin inline asm
	{  cvt.f32.f16 %f411, %rs267;}

	// end inline asm
	fma.rn.ftz.f32 	%f588, %f585, %f411, %f582;
	shl.b16 	%rs418, %rs56, 4;
	cvt.s16.s8 	%rs419, %rs418;
	shr.s16 	%rs420, %rs419, 7;
	and.b16  	%rs421, %rs420, -16;
	or.b16  	%rs422, %rs421, %rs57;
	cvt.rn.f32.s16 	%f589, %rs422;
	sub.ftz.f32 	%f590, %f589, %f5;
	mul.ftz.f32 	%f591, %f45, %f590;
	// begin inline asm
	{  cvt.f32.f16 %f412, %rs268;}

	// end inline asm
	fma.rn.ftz.f32 	%f592, %f591, %f412, %f586;
	// begin inline asm
	{  cvt.f32.f16 %f413, %rs269;}

	// end inline asm
	fma.rn.ftz.f32 	%f593, %f591, %f413, %f587;
	// begin inline asm
	{  cvt.f32.f16 %f414, %rs270;}

	// end inline asm
	fma.rn.ftz.f32 	%f594, %f591, %f414, %f588;
	shl.b16 	%rs423, %rs58, 4;
	cvt.s16.s8 	%rs424, %rs423;
	shr.s16 	%rs425, %rs424, 7;
	and.b16  	%rs426, %rs425, -16;
	or.b16  	%rs427, %rs426, %rs59;
	cvt.rn.f32.s16 	%f595, %rs427;
	sub.ftz.f32 	%f596, %f595, %f5;
	mul.ftz.f32 	%f597, %f45, %f596;
	mov.b32 	{%rs271, %rs274}, %r229;
	// begin inline asm
	{  cvt.f32.f16 %f415, %rs271;}

	// end inline asm
	fma.rn.ftz.f32 	%f598, %f597, %f415, %f592;
	mov.b32 	{%rs272, %rs275}, %r237;
	// begin inline asm
	{  cvt.f32.f16 %f416, %rs272;}

	// end inline asm
	fma.rn.ftz.f32 	%f599, %f597, %f416, %f593;
	mov.b32 	{%rs273, %rs276}, %r245;
	// begin inline asm
	{  cvt.f32.f16 %f417, %rs273;}

	// end inline asm
	fma.rn.ftz.f32 	%f600, %f597, %f417, %f594;
	shl.b16 	%rs428, %rs60, 4;
	cvt.s16.s8 	%rs429, %rs428;
	shr.s16 	%rs430, %rs429, 7;
	and.b16  	%rs431, %rs430, -16;
	or.b16  	%rs432, %rs431, %rs61;
	cvt.rn.f32.s16 	%f601, %rs432;
	sub.ftz.f32 	%f602, %f601, %f5;
	mul.ftz.f32 	%f603, %f45, %f602;
	// begin inline asm
	{  cvt.f32.f16 %f418, %rs274;}

	// end inline asm
	fma.rn.ftz.f32 	%f604, %f603, %f418, %f598;
	// begin inline asm
	{  cvt.f32.f16 %f419, %rs275;}

	// end inline asm
	fma.rn.ftz.f32 	%f605, %f603, %f419, %f599;
	// begin inline asm
	{  cvt.f32.f16 %f420, %rs276;}

	// end inline asm
	fma.rn.ftz.f32 	%f606, %f603, %f420, %f600;
	shl.b16 	%rs433, %rs62, 4;
	cvt.s16.s8 	%rs434, %rs433;
	shr.s16 	%rs435, %rs434, 7;
	and.b16  	%rs436, %rs435, -16;
	or.b16  	%rs437, %rs436, %rs63;
	cvt.rn.f32.s16 	%f607, %rs437;
	sub.ftz.f32 	%f608, %f607, %f5;
	mul.ftz.f32 	%f609, %f45, %f608;
	mov.b32 	{%rs277, %rs280}, %r230;
	// begin inline asm
	{  cvt.f32.f16 %f421, %rs277;}

	// end inline asm
	fma.rn.ftz.f32 	%f610, %f609, %f421, %f604;
	mov.b32 	{%rs278, %rs281}, %r238;
	// begin inline asm
	{  cvt.f32.f16 %f422, %rs278;}

	// end inline asm
	fma.rn.ftz.f32 	%f611, %f609, %f422, %f605;
	mov.b32 	{%rs279, %rs282}, %r246;
	// begin inline asm
	{  cvt.f32.f16 %f423, %rs279;}

	// end inline asm
	fma.rn.ftz.f32 	%f612, %f609, %f423, %f606;
	shl.b16 	%rs438, %rs64, 4;
	cvt.s16.s8 	%rs439, %rs438;
	shr.s16 	%rs440, %rs439, 7;
	and.b16  	%rs441, %rs440, -16;
	or.b16  	%rs442, %rs441, %rs64;
	cvt.rn.f32.s16 	%f613, %rs442;
	sub.ftz.f32 	%f614, %f613, %f5;
	mul.ftz.f32 	%f615, %f45, %f614;
	// begin inline asm
	{  cvt.f32.f16 %f424, %rs280;}

	// end inline asm
	fma.rn.ftz.f32 	%f687, %f615, %f424, %f610;
	// begin inline asm
	{  cvt.f32.f16 %f425, %rs281;}

	// end inline asm
	fma.rn.ftz.f32 	%f686, %f615, %f425, %f611;
	// begin inline asm
	{  cvt.f32.f16 %f426, %rs282;}

	// end inline asm
	fma.rn.ftz.f32 	%f685, %f615, %f426, %f612;

$L__BB0_8:
	add.s32 	%r341, %r341, 4;
	shl.b32 	%r251, %r341, 5;
	add.s32 	%r340, %r251, %r45;
	shl.b32 	%r339, %r340, 2;
	setp.lt.u32 	%p7, %r339, %r42;
	@%p7 bra 	$L__BB0_2;

$L__BB0_9:
	mov.u32 	%r253, %tid.y;
	shl.b32 	%r254, %r253, 5;
	add.s32 	%r32, %r254, %r45;
	setp.lt.u32 	%p8, %r32, 32;
	shl.b32 	%r256, %r32, 2;
	mov.u32 	%r257, _ZZ9gemv_int4ILi4ELi128ELi3EEvP6__halfPKS0_S3_PKjPKhS3_iiiiffbE12temp_storage;
	add.s32 	%r258, %r257, %r256;
	@%p8 bra 	$L__BB0_11;

	add.s32 	%r336, %r258, -112;
	st.shared.f32 	[%r336], %f687;

$L__BB0_11:
	setp.gt.u32 	%p9, %r32, 31;
	bar.sync 	0;
	mad.lo.s32 	%r34, %r32, 12, %r257;
	@%p9 bra 	$L__BB0_13;

	mov.u32 	%r277, 16;
	ld.shared.f32 	%f631, [%r34+16];
	add.ftz.f32 	%f632, %f687, %f631;
	ld.shared.f32 	%f633, [%r34+20];
	add.ftz.f32 	%f634, %f632, %f633;
	ld.shared.f32 	%f635, [%r34+24];
	add.ftz.f32 	%f618, %f634, %f635;
	mov.u32 	%r265, 1;
	mov.u32 	%r278, 31;
	mov.u32 	%r279, -1;
	// begin inline asm
	{  .reg .f32 r0;  .reg .pred p;  shfl.sync.down.b32 r0|p, %f618, %r265, %r278, %r279;  @p add.f32 r0, r0, %f618;  mov.f32 %f616, r0;}
	// end inline asm
	mov.u32 	%r268, 2;
	// begin inline asm
	{  .reg .f32 r0;  .reg .pred p;  shfl.sync.down.b32 r0|p, %f616, %r268, %r278, %r279;  @p add.f32 r0, r0, %f616;  mov.f32 %f619, r0;}
	// end inline asm
	mov.u32 	%r271, 4;
	// begin inline asm
	{  .reg .f32 r0;  .reg .pred p;  shfl.sync.down.b32 r0|p, %f619, %r271, %r278, %r279;  @p add.f32 r0, r0, %f619;  mov.f32 %f622, r0;}
	// end inline asm
	mov.u32 	%r274, 8;
	// begin inline asm
	{  .reg .f32 r0;  .reg .pred p;  shfl.sync.down.b32 r0|p, %f622, %r274, %r278, %r279;  @p add.f32 r0, r0, %f622;  mov.f32 %f625, r0;}
	// end inline asm
	// begin inline asm
	{  .reg .f32 r0;  .reg .pred p;  shfl.sync.down.b32 r0|p, %f625, %r277, %r278, %r279;  @p add.f32 r0, r0, %f625;  mov.f32 %f687, r0;}
	// end inline asm

$L__BB0_13:
	@%p8 bra 	$L__BB0_15;

	add.s32 	%r337, %r258, -112;
	st.shared.f32 	[%r337+640], %f686;

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

	ld.shared.f32 	%f651, [%r34+656];
	add.ftz.f32 	%f652, %f686, %f651;
	ld.shared.f32 	%f653, [%r34+660];
	add.ftz.f32 	%f654, %f652, %f653;
	ld.shared.f32 	%f655, [%r34+664];
	add.ftz.f32 	%f638, %f654, %f655;
	mov.u32 	%r289, 1;
	mov.u32 	%r302, 31;
	mov.u32 	%r303, -1;
	// begin inline asm
	{  .reg .f32 r0;  .reg .pred p;  shfl.sync.down.b32 r0|p, %f638, %r289, %r302, %r303;  @p add.f32 r0, r0, %f638;  mov.f32 %f636, r0;}
	// end inline asm
	mov.u32 	%r292, 2;
	// begin inline asm
	{  .reg .f32 r0;  .reg .pred p;  shfl.sync.down.b32 r0|p, %f636, %r292, %r302, %r303;  @p add.f32 r0, r0, %f636;  mov.f32 %f639, r0;}
	// end inline asm
	mov.u32 	%r295, 4;
	// begin inline asm
	{  .reg .f32 r0;  .reg .pred p;  shfl.sync.down.b32 r0|p, %f639, %r295, %r302, %r303;  @p add.f32 r0, r0, %f639;  mov.f32 %f642, r0;}
	// end inline asm
	mov.u32 	%r298, 8;
	// begin inline asm
	{  .reg .f32 r0;  .reg .pred p;  shfl.sync.down.b32 r0|p, %f642, %r298, %r302, %r303;  @p add.f32 r0, r0, %f642;  mov.f32 %f645, r0;}
	// end inline asm
	mov.u32 	%r301, 16;
	// begin inline asm
	{  .reg .f32 r0;  .reg .pred p;  shfl.sync.down.b32 r0|p, %f645, %r301, %r302, %r303;  @p add.f32 r0, r0, %f645;  mov.f32 %f686, r0;}
	// end inline asm

$L__BB0_17:
	@%p8 bra 	$L__BB0_19;

	add.s32 	%r338, %r258, -112;
	st.shared.f32 	[%r338+1280], %f685;

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

	ld.shared.f32 	%f671, [%r34+1296];
	add.ftz.f32 	%f672, %f685, %f671;
	ld.shared.f32 	%f673, [%r34+1300];
	add.ftz.f32 	%f674, %f672, %f673;
	ld.shared.f32 	%f675, [%r34+1304];
	add.ftz.f32 	%f658, %f674, %f675;
	mov.u32 	%r313, 1;
	mov.u32 	%r326, 31;
	mov.u32 	%r327, -1;
	// begin inline asm
	{  .reg .f32 r0;  .reg .pred p;  shfl.sync.down.b32 r0|p, %f658, %r313, %r326, %r327;  @p add.f32 r0, r0, %f658;  mov.f32 %f656, r0;}
	// end inline asm
	mov.u32 	%r316, 2;
	// begin inline asm
	{  .reg .f32 r0;  .reg .pred p;  shfl.sync.down.b32 r0|p, %f656, %r316, %r326, %r327;  @p add.f32 r0, r0, %f656;  mov.f32 %f659, r0;}
	// end inline asm
	mov.u32 	%r319, 4;
	// begin inline asm
	{  .reg .f32 r0;  .reg .pred p;  shfl.sync.down.b32 r0|p, %f659, %r319, %r326, %r327;  @p add.f32 r0, r0, %f659;  mov.f32 %f662, r0;}
	// end inline asm
	mov.u32 	%r322, 8;
	// begin inline asm
	{  .reg .f32 r0;  .reg .pred p;  shfl.sync.down.b32 r0|p, %f662, %r322, %r326, %r327;  @p add.f32 r0, r0, %f662;  mov.f32 %f665, r0;}
	// end inline asm
	mov.u32 	%r325, 16;
	// begin inline asm
	{  .reg .f32 r0;  .reg .pred p;  shfl.sync.down.b32 r0|p, %f665, %r325, %r326, %r327;  @p add.f32 r0, r0, %f665;  mov.f32 %f685, r0;}
	// end inline asm

$L__BB0_21:
	or.b32  	%r330, %r45, %r253;
	setp.ne.s32 	%p14, %r330, 0;
	@%p14 bra 	$L__BB0_29;

	ld.param.u64 	%rd56, [_Z28dequant_gemv_group128_batch323DequantGemvKernelParams_param_0+8];
	setp.eq.s64 	%p15, %rd56, 0;
	mul.ftz.f32 	%f694, %f40, %f687;
	mov.u32 	%r331, %ctaid.x;
	cvt.s64.s32 	%rd10, %r331;
	@%p15 bra 	$L__BB0_24;

	ld.param.u64 	%rd60, [_Z28dequant_gemv_group128_batch323DequantGemvKernelParams_param_0+8];
	cvta.to.global.u64 	%rd59, %rd60;
	shl.b64 	%rd38, %rd10, 1;
	add.s64 	%rd39, %rd59, %rd38;
	ld.global.u16 	%rs443, [%rd39];
	// begin inline asm
	{  cvt.f32.f16 %f676, %rs443;}

	// end inline asm
	fma.rn.ftz.f32 	%f694, %f41, %f676, %f694;

$L__BB0_24:
	ld.param.u64 	%rd57, [_Z28dequant_gemv_group128_batch323DequantGemvKernelParams_param_0];
	// begin inline asm
	{  cvt.rn.f16.f32 %rs444, %f694;}

	// end inline asm
	// begin inline asm
	createpolicy.fractional.L2::evict_last.L2::evict_unchanged.b64 %rd40, 1.0;
	// end inline asm
	shl.b64 	%rd43, %rd10, 1;
	add.s64 	%rd41, %rd57, %rd43;
	// begin inline asm
	st.global.L1::evict_normal.L2::cache_hint.u16 [%rd41], %rs444, %rd40;
	// end inline asm
	mul.ftz.f32 	%f695, %f40, %f686;
	add.s32 	%r333, %r41, %r331;
	cvt.s64.s32 	%rd13, %r333;
	@%p15 bra 	$L__BB0_26;

	ld.param.u64 	%rd62, [_Z28dequant_gemv_group128_batch323DequantGemvKernelParams_param_0+8];
	cvta.to.global.u64 	%rd61, %rd62;
	shl.b64 	%rd44, %rd13, 1;
	add.s64 	%rd45, %rd61, %rd44;
	ld.global.u16 	%rs446, [%rd45];
	// begin inline asm
	{  cvt.f32.f16 %f678, %rs446;}

	// end inline asm
	fma.rn.ftz.f32 	%f695, %f41, %f678, %f695;

$L__BB0_26:
	mul.wide.s32 	%rd49, %r41, 2;
	add.s64 	%rd47, %rd41, %rd49;
	// begin inline asm
	{  cvt.rn.f16.f32 %rs447, %f695;}

	// end inline asm
	// begin inline asm
	createpolicy.fractional.L2::evict_last.L2::evict_unchanged.b64 %rd46, 1.0;
	// end inline asm
	// begin inline asm
	st.global.L1::evict_normal.L2::cache_hint.u16 [%rd47], %rs447, %rd46;
	// end inline asm
	mul.ftz.f32 	%f696, %f40, %f685;
	cvt.u32.u64 	%r334, %rd13;
	add.s32 	%r335, %r334, %r41;
	cvt.s64.s32 	%rd14, %r335;
	@%p15 bra 	$L__BB0_28;

	ld.param.u64 	%rd64, [_Z28dequant_gemv_group128_batch323DequantGemvKernelParams_param_0+8];
	cvta.to.global.u64 	%rd63, %rd64;
	shl.b64 	%rd50, %rd14, 1;
	add.s64 	%rd51, %rd63, %rd50;
	ld.global.u16 	%rs449, [%rd51];
	// begin inline asm
	{  cvt.f32.f16 %f680, %rs449;}

	// end inline asm
	fma.rn.ftz.f32 	%f696, %f41, %f680, %f696;

$L__BB0_28:
	ld.param.u64 	%rd58, [_Z28dequant_gemv_group128_batch323DequantGemvKernelParams_param_0];
	// begin inline asm
	{  cvt.rn.f16.f32 %rs450, %f696;}

	// end inline asm
	// begin inline asm
	createpolicy.fractional.L2::evict_last.L2::evict_unchanged.b64 %rd52, 1.0;
	// end inline asm
	shl.b64 	%rd55, %rd14, 1;
	add.s64 	%rd53, %rd58, %rd55;
	// begin inline asm
	st.global.L1::evict_normal.L2::cache_hint.u16 [%rd53], %rs450, %rd52;
	// end inline asm

$L__BB0_29:
	ret;

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



	ret;

}