asm x2x4 gemm

	.globl	<gemm_f16::gemm::f16::GEMM as lazy_static::LazyStatic>::initialize
	.p2align	2
<gemm_f16::gemm::f16::GEMM as lazy_static::LazyStatic>::initialize:
Lfunc_begin18:
	.cfi_startproc
	sub sp, sp, #48
	.cfi_def_cfa_offset 48
	stp x29, x30, [sp, #32]
	add x29, sp, #32
	.cfi_def_cfa w29, 16
	.cfi_offset w30, -8
	.cfi_offset w29, -16
	.cfi_remember_state
Lloh101:
	adrp x8, <gemm_f16::gemm::f16::GEMM as core::ops::deref::Deref>::deref::__stability::LAZY@PAGE
Lloh102:
	add x8, x8, <gemm_f16::gemm::f16::GEMM as core::ops::deref::Deref>::deref::__stability::LAZY@PAGEOFF
	str x8, [sp, #8]
	add x8, x8, #8
	ldapr x8, [x8]
	cmp x8, #3
	b.ne LBB18_2
LBB18_1:
	.cfi_def_cfa wsp, 48
	ldp x29, x30, [sp, #32]
	add sp, sp, #48
	.cfi_def_cfa_offset 0
	.cfi_restore w30
	.cfi_restore w29
	ret
LBB18_2:
	.cfi_restore_state
	add x8, sp, #8
	str x8, [sp, #16]
	add x8, sp, #16
	stur x8, [x29, #-8]
Lloh103:
	adrp x0, <gemm_f16::gemm::f16::GEMM as core::ops::deref::Deref>::deref::__stability::LAZY@PAGE+8
Lloh104:
	add x0, x0, <gemm_f16::gemm::f16::GEMM as core::ops::deref::Deref>::deref::__stability::LAZY@PAGEOFF+8
Lloh105:
	adrp x3, l___unnamed_3@PAGE
Lloh106:
	add x3, x3, l___unnamed_3@PAGEOFF
Lloh107:
	adrp x4, l___unnamed_25@PAGE
Lloh108:
	add x4, x4, l___unnamed_25@PAGEOFF
	sub x2, x29, #8
	mov w1, #0
	bl std::sys_common::once::queue::Once::call
	b LBB18_1
	.loh AdrpAdd	Lloh101, Lloh102
	.loh AdrpAdd	Lloh107, Lloh108
	.loh AdrpAdd	Lloh105, Lloh106
	.loh AdrpAdd	Lloh103, Lloh104
Lfunc_end18:
	.cfi_endproc

	.p2align	2
gemm_f16::microkernel::neon::f16::x1x1:
Lfunc_begin19:
	.cfi_startproc
	sub sp, sp, #16
	.cfi_def_cfa_offset 16
	movi.2d v0, #0000000000000000
	cmp x2, #2
	b.hs LBB19_2
	mov x8, x5
	b LBB19_4
LBB19_2:
	ldp x10, x8, [sp, #16]
	lsr x9, x2, #1
	lsl x10, x10, #1
	lsl x11, x8, #1
	lsl x12, x11, #1
	mov x8, x5
LBB19_3:
	ld1r.8h { v1 }, [x8], x12
	ldr q2, [x4]
	; InlineAsm Start
	fmla.8h v0, v2, v1
	; InlineAsm End
	add x13, x5, x11
	ld1r.8h { v1 }, [x13]
	ldr q2, [x4, x10]
	; InlineAsm Start
	fmla.8h v0, v2, v1
	; InlineAsm End
	add x4, x4, x10, lsl #1
	mov x5, x8
	subs x9, x9, #1
	b.ne LBB19_3
LBB19_4:
	tbz w2, #0, LBB19_6
	ld1r.8h { v1 }, [x8]
	ldr q2, [x4]
	; InlineAsm Start
	fmla.8h v0, v2, v1
	; InlineAsm End

	ldrh w8, [sp, #42]
	ldrh w9, [sp, #40]
	ldrb w10, [sp, #44]
	cmp x0, #8
	b.ne LBB19_12
	cmp x1, #1
	b.ne LBB19_12
	cmp x7, #1
	b.ne LBB19_12
	dup.8h v1, w8
	and w8, w10, #0xff
	cmp w8, #1
	b.eq LBB19_20
	cmp w8, #2
	b.ne LBB19_21
	dup.8h v2, w9
	ldr q3, [x3]
	; InlineAsm Start
	fmul.8h v4, v2, v3
	; InlineAsm End
	; InlineAsm Start
	fmul.8h v2, v1, v0
	; InlineAsm End
	; InlineAsm Start
	fadd.8h v0, v4, v2
	; InlineAsm End
	str q0, [x3]
	b LBB19_35
LBB19_12:
	str q0, [sp]
	and w10, w10, #0xff
	cmp w10, #2
	b.eq LBB19_23
	cmp w10, #1
	b.ne LBB19_29
	cbz x1, LBB19_35
	cbz x0, LBB19_35
	mov x9, #0
	lsl x10, x6, #1
	lsl x11, x7, #1
	mov x12, sp
LBB19_17:
	mov x13, x0
	mov x14, x3
	mov x15, x12
LBB19_18:
	ldr h0, [x15], #2
	fmov s1, w8
	; InlineAsm Start
	fmul h2, h1, h0
	; InlineAsm End
	ldr h0, [x14]
	; InlineAsm Start
	fadd h1, h0, h2
	; InlineAsm End
	str h1, [x14]
	add x14, x14, x11
	subs x13, x13, #1
	b.ne LBB19_18
	add x9, x9, #1
	add x12, x12, #16
	add x3, x3, x10
	cmp x9, x1
	b.ne LBB19_17
	b LBB19_35
LBB19_20:
	ldr q2, [x3]
	; InlineAsm Start
	fmla.8h v2, v1, v0
	; InlineAsm End
	b LBB19_22
LBB19_21:
	; InlineAsm Start
	fmul.8h v2, v1, v0
	; InlineAsm End
LBB19_22:
	str q2, [x3]
	b LBB19_35
LBB19_23:
	cbz x1, LBB19_35
	cbz x0, LBB19_35
	mov x10, #0
	lsl x11, x6, #1
	lsl x12, x7, #1
	mov x13, sp
LBB19_26:
	mov x14, x0
	mov x15, x3
	mov x16, x13
LBB19_27:
	ldr h0, [x15]
	fmov s1, w9
	; InlineAsm Start
	fmul h2, h1, h0
	; InlineAsm End
	ldr h0, [x16], #2
	fmov s1, w8
	; InlineAsm Start
	fmul h3, h1, h0
	; InlineAsm End
	; InlineAsm Start
	fadd h0, h2, h3
	; InlineAsm End
	str h0, [x15]
	add x15, x15, x12
	subs x14, x14, #1
	b.ne LBB19_27
	add x10, x10, #1
	add x13, x13, #16
	add x3, x3, x11
	cmp x10, x1
	b.ne LBB19_26
	b LBB19_35
LBB19_29:
	cbz x1, LBB19_35
	cbz x0, LBB19_35
	mov x9, #0
	lsl x10, x6, #1
	lsl x11, x7, #1
	mov x12, sp
LBB19_32:
	mov x13, x0
	mov x14, x3
	mov x15, x12
LBB19_33:
	ldr h0, [x15], #2
	fmov s1, w8
	; InlineAsm Start
	fmul h2, h1, h0
	; InlineAsm End
	str h2, [x14]
	add x14, x14, x11
	subs x13, x13, #1
	b.ne LBB19_33
	add x9, x9, #1
	add x12, x12, #16
	add x3, x3, x10
	cmp x9, x1
	b.ne LBB19_32
LBB19_35:
	add sp, sp, #16
	.cfi_def_cfa_offset 0
	ret
Lfunc_end19:
	.cfi_endproc

	.p2align	2
gemm_f16::microkernel::neon::f16::x1x2:
Lfunc_begin20:
	.cfi_startproc
	sub sp, sp, #32
	.cfi_def_cfa_offset 32
	ldr x8, [sp, #48]
	movi.2d v0, #0000000000000000
	lsl x8, x8, #1
	cmp x2, #2
	b.hs LBB20_2
	movi.2d v1, #0000000000000000
	mov x9, x5
	b LBB20_4
LBB20_2:
	ldp x11, x9, [sp, #32]
	lsr x10, x2, #1
	lsl x11, x11, #1
	lsl x12, x9, #1
	add x13, x8, x12
	lsl x14, x12, #1
	movi.2d v1, #0000000000000000
	mov x9, x5
LBB20_3:
	ld1r.8h { v2 }, [x9], x14
	ldr q3, [x4]
	; InlineAsm Start
	fmla.8h v1, v3, v2
	; InlineAsm End
	add x15, x5, x8
	ld1r.8h { v2 }, [x15]
	; InlineAsm Start
	fmla.8h v0, v3, v2
	; InlineAsm End
	add x15, x5, x12
	ld1r.8h { v2 }, [x15]
	ldr q3, [x4, x11]
	; InlineAsm Start
	fmla.8h v1, v3, v2
	; InlineAsm End
	add x15, x5, x13
	ld1r.8h { v2 }, [x15]
	; InlineAsm Start
	fmla.8h v0, v3, v2
	; InlineAsm End
	add x4, x4, x11, lsl #1
	mov x5, x9
	subs x10, x10, #1
	b.ne LBB20_3
LBB20_4:
	tbz w2, #0, LBB20_6
	ld1r.8h { v2 }, [x9], x8
	ldr q3, [x4]
	; InlineAsm Start
	fmla.8h v1, v3, v2
	; InlineAsm End
	ld1r.8h { v2 }, [x9]
	; InlineAsm Start
	fmla.8h v0, v3, v2
	; InlineAsm End

	ldrh w8, [sp, #58]
	ldrh w9, [sp, #56]
	ldrb w10, [sp, #60]
	cmp x0, #8
	b.ne LBB20_12
	cmp x1, #2
	b.ne LBB20_12
	cmp x7, #1
	b.ne LBB20_12
	dup.8h v2, w8
	and w8, w10, #0xff
	cmp w8, #1
	b.eq LBB20_20
	cmp w8, #2
	b.ne LBB20_21
	dup.8h v3, w9
	ldr q4, [x3]
	; InlineAsm Start
	fmul.8h v5, v3, v4
	; InlineAsm End
	; InlineAsm Start
	fmul.8h v4, v2, v1
	; InlineAsm End
	; InlineAsm Start
	fadd.8h v1, v5, v4
	; InlineAsm End
	str q1, [x3]
	lsl x8, x6, #1
	ldr q1, [x3, x8]
	; InlineAsm Start
	fmul.8h v4, v3, v1
	; InlineAsm End
	; InlineAsm Start
	fmul.8h v1, v2, v0
	; InlineAsm End
	; InlineAsm Start
	fadd.8h v0, v4, v1
	; InlineAsm End
	str q0, [x3, x8]
	b LBB20_35
LBB20_12:
	stp q1, q0, [sp]
	and w10, w10, #0xff
	cmp w10, #2
	b.eq LBB20_23
	cmp w10, #1
	b.ne LBB20_29
	cbz x1, LBB20_35
	cbz x0, LBB20_35
	mov x9, #0
	lsl x10, x6, #1
	lsl x11, x7, #1
	mov x12, sp
LBB20_17:
	mov x13, x0
	mov x14, x3
	mov x15, x12
LBB20_18:
	ldr h0, [x15], #2
	fmov s1, w8
	; InlineAsm Start
	fmul h2, h1, h0
	; InlineAsm End
	ldr h0, [x14]
	; InlineAsm Start
	fadd h1, h0, h2
	; InlineAsm End
	str h1, [x14]
	add x14, x14, x11
	subs x13, x13, #1
	b.ne LBB20_18
	add x9, x9, #1
	add x12, x12, #16
	add x3, x3, x10
	cmp x9, x1
	b.ne LBB20_17
	b LBB20_35
LBB20_20:
	ldr q3, [x3]
	; InlineAsm Start
	fmla.8h v3, v2, v1
	; InlineAsm End
	str q3, [x3]
	lsl x8, x6, #1
	ldr q1, [x3, x8]
	; InlineAsm Start
	fmla.8h v1, v2, v0
	; InlineAsm End
	b LBB20_22
LBB20_21:
	; InlineAsm Start
	fmul.8h v3, v2, v1
	; InlineAsm End
	str q3, [x3]
	lsl x8, x6, #1
	; InlineAsm Start
	fmul.8h v1, v2, v0
	; InlineAsm End
LBB20_22:
	str q1, [x3, x8]
	b LBB20_35
LBB20_23:
	cbz x1, LBB20_35
	cbz x0, LBB20_35
	mov x10, #0
	lsl x11, x6, #1
	lsl x12, x7, #1
	mov x13, sp
LBB20_26:
	mov x14, x0
	mov x15, x3
	mov x16, x13
LBB20_27:
	ldr h0, [x15]
	fmov s1, w9
	; InlineAsm Start
	fmul h2, h1, h0
	; InlineAsm End
	ldr h0, [x16], #2
	fmov s1, w8
	; InlineAsm Start
	fmul h3, h1, h0
	; InlineAsm End
	; InlineAsm Start
	fadd h0, h2, h3
	; InlineAsm End
	str h0, [x15]
	add x15, x15, x12
	subs x14, x14, #1
	b.ne LBB20_27
	add x10, x10, #1
	add x13, x13, #16
	add x3, x3, x11
	cmp x10, x1
	b.ne LBB20_26
	b LBB20_35
LBB20_29:
	cbz x1, LBB20_35
	cbz x0, LBB20_35
	mov x9, #0
	lsl x10, x6, #1
	lsl x11, x7, #1
	mov x12, sp
LBB20_32:
	mov x13, x0
	mov x14, x3
	mov x15, x12
LBB20_33:
	ldr h0, [x15], #2
	fmov s1, w8
	; InlineAsm Start
	fmul h2, h1, h0
	; InlineAsm End
	str h2, [x14]
	add x14, x14, x11
	subs x13, x13, #1
	b.ne LBB20_33
	add x9, x9, #1
	add x12, x12, #16
	add x3, x3, x10
	cmp x9, x1
	b.ne LBB20_32
LBB20_35:
	add sp, sp, #32
	.cfi_def_cfa_offset 0
	ret
Lfunc_end20:
	.cfi_endproc

	.p2align	2
gemm_f16::microkernel::neon::f16::x1x3:
Lfunc_begin21:
	.cfi_startproc
	sub sp, sp, #64
	.cfi_def_cfa_offset 64
	stp x20, x19, [sp, #48]
	.cfi_offset w19, -8
	.cfi_offset w20, -16
	ldr x8, [sp, #80]
	movi.2d v0, #0000000000000000
	lsl x9, x8, #2
	cmp x2, #2
	b.hs LBB21_2
	movi.2d v1, #0000000000000000
	movi.2d v2, #0000000000000000
	mov x10, x5
	b LBB21_4
LBB21_2:
	lsr x11, x2, #1
	ldp x13, x10, [sp, #64]
	lsl x12, x8, #1
	lsl x13, x13, #1
	lsl x14, x10, #1
	add x15, x9, x14
	add x16, x12, x14
	lsl x17, x14, #1
	movi.2d v1, #0000000000000000
	movi.2d v2, #0000000000000000
	mov x10, x5
LBB21_3:
	ld1r.8h { v3 }, [x10], x17
	ldr q4, [x4]
	; InlineAsm Start
	fmla.8h v2, v4, v3
	; InlineAsm End
	add x19, x5, x12
	ld1r.8h { v3 }, [x19]
	; InlineAsm Start
	fmla.8h v1, v4, v3
	; InlineAsm End
	add x19, x5, x9
	ld1r.8h { v3 }, [x19]
	; InlineAsm Start
	fmla.8h v0, v4, v3
	; InlineAsm End
	add x19, x5, x14
	ld1r.8h { v3 }, [x19]
	ldr q4, [x4, x13]
	; InlineAsm Start
	fmla.8h v2, v4, v3
	; InlineAsm End
	add x19, x5, x16
	ld1r.8h { v3 }, [x19]
	; InlineAsm Start
	fmla.8h v1, v4, v3
	; InlineAsm End
	add x5, x5, x15
	ld1r.8h { v3 }, [x5]
	; InlineAsm Start
	fmla.8h v0, v4, v3
	; InlineAsm End
	add x4, x4, x13, lsl #1
	mov x5, x10
	subs x11, x11, #1
	b.ne LBB21_3
LBB21_4:
	tbz w2, #0, LBB21_6
	mov x11, x10
	ld1r.8h { v3 }, [x11], x9
	ldr q4, [x4]
	; InlineAsm Start
	fmla.8h v2, v4, v3
	; InlineAsm End
	add x8, x10, x8, lsl #1
	ld1r.8h { v3 }, [x8]
	; InlineAsm Start
	fmla.8h v1, v4, v3
	; InlineAsm End
	ld1r.8h { v3 }, [x11]
	; InlineAsm Start
	fmla.8h v0, v4, v3
	; InlineAsm End

	ldrh w8, [sp, #90]
	ldrh w9, [sp, #88]
	ldrb w10, [sp, #92]
	cmp x0, #8
	b.ne LBB21_12
	cmp x1, #3
	b.ne LBB21_12
	cmp x7, #1
	b.ne LBB21_12
	dup.8h v3, w8
	and w8, w10, #0xff
	cmp w8, #1
	b.eq LBB21_20
	cmp w8, #2
	b.ne LBB21_21
	dup.8h v4, w9
	ldr q5, [x3]
	; InlineAsm Start
	fmul.8h v6, v4, v5
	; InlineAsm End
	; InlineAsm Start
	fmul.8h v5, v3, v2
	; InlineAsm End
	; InlineAsm Start
	fadd.8h v2, v6, v5
	; InlineAsm End
	str q2, [x3]
	lsl x8, x6, #1
	ldr q2, [x3, x8]
	; InlineAsm Start
	fmul.8h v5, v4, v2
	; InlineAsm End
	; InlineAsm Start
	fmul.8h v2, v3, v1
	; InlineAsm End
	; InlineAsm Start
	fadd.8h v1, v5, v2
	; InlineAsm End
	str q1, [x3, x8]
	lsl x8, x6, #2
	ldr q1, [x3, x8]
	; InlineAsm Start
	fmul.8h v2, v4, v1
	; InlineAsm End
	; InlineAsm Start
	fmul.8h v1, v3, v0
	; InlineAsm End
	; InlineAsm Start
	fadd.8h v0, v2, v1
	; InlineAsm End
	str q0, [x3, x8]
	b LBB21_35
LBB21_12:
	stp q2, q1, [sp]
	str q0, [sp, #32]
	and w10, w10, #0xff
	cmp w10, #2
	b.eq LBB21_23
	cmp w10, #1
	b.ne LBB21_29
	cbz x1, LBB21_35
	cbz x0, LBB21_35
	mov x9, #0
	lsl x10, x6, #1
	lsl x11, x7, #1
	mov x12, sp
LBB21_17:
	mov x13, x0
	mov x14, x3
	mov x15, x12
LBB21_18:
	ldr h0, [x15], #2
	fmov s1, w8
	; InlineAsm Start
	fmul h2, h1, h0
	; InlineAsm End
	ldr h0, [x14]
	; InlineAsm Start
	fadd h1, h0, h2
	; InlineAsm End
	str h1, [x14]
	add x14, x14, x11
	subs x13, x13, #1
	b.ne LBB21_18
	add x9, x9, #1
	add x12, x12, #16
	add x3, x3, x10
	cmp x9, x1
	b.ne LBB21_17
	b LBB21_35
LBB21_20:
	ldr q4, [x3]
	; InlineAsm Start
	fmla.8h v4, v3, v2
	; InlineAsm End
	str q4, [x3]
	lsl x8, x6, #1
	ldr q2, [x3, x8]
	; InlineAsm Start
	fmla.8h v2, v3, v1
	; InlineAsm End
	str q2, [x3, x8]
	lsl x8, x6, #2
	ldr q1, [x3, x8]
	; InlineAsm Start
	fmla.8h v1, v3, v0
	; InlineAsm End
	b LBB21_22
LBB21_21:
	; InlineAsm Start
	fmul.8h v4, v3, v2
	; InlineAsm End
	str q4, [x3]
	lsl x8, x6, #1
	; InlineAsm Start
	fmul.8h v2, v3, v1
	; InlineAsm End
	str q2, [x3, x8]
	lsl x8, x6, #2
	; InlineAsm Start
	fmul.8h v1, v3, v0
	; InlineAsm End
LBB21_22:
	str q1, [x3, x8]
	b LBB21_35
LBB21_23:
	cbz x1, LBB21_35
	cbz x0, LBB21_35
	mov x10, #0
	lsl x11, x6, #1
	lsl x12, x7, #1
	mov x13, sp
LBB21_26:
	mov x14, x0
	mov x15, x3
	mov x16, x13
LBB21_27:
	ldr h0, [x15]
	fmov s1, w9
	; InlineAsm Start
	fmul h2, h1, h0
	; InlineAsm End
	ldr h0, [x16], #2
	fmov s1, w8
	; InlineAsm Start
	fmul h3, h1, h0
	; InlineAsm End
	; InlineAsm Start
	fadd h0, h2, h3
	; InlineAsm End
	str h0, [x15]
	add x15, x15, x12
	subs x14, x14, #1
	b.ne LBB21_27
	add x10, x10, #1
	add x13, x13, #16
	add x3, x3, x11
	cmp x10, x1
	b.ne LBB21_26
	b LBB21_35
LBB21_29:
	cbz x1, LBB21_35
	cbz x0, LBB21_35
	mov x9, #0
	lsl x10, x6, #1
	lsl x11, x7, #1
	mov x12, sp
LBB21_32:
	mov x13, x0
	mov x14, x3
	mov x15, x12
LBB21_33:
	ldr h0, [x15], #2
	fmov s1, w8
	; InlineAsm Start
	fmul h2, h1, h0
	; InlineAsm End
	str h2, [x14]
	add x14, x14, x11
	subs x13, x13, #1
	b.ne LBB21_33
	add x9, x9, #1
	add x12, x12, #16
	add x3, x3, x10
	cmp x9, x1
	b.ne LBB21_32
LBB21_35:
	ldp x20, x19, [sp, #48]
	add sp, sp, #64
	.cfi_def_cfa_offset 0
	.cfi_restore w19
	.cfi_restore w20
	ret
Lfunc_end21:
	.cfi_endproc

	.p2align	2
gemm_f16::microkernel::neon::f16::x1x4:
Lfunc_begin22:
	.cfi_startproc
	sub sp, sp, #96
	.cfi_def_cfa_offset 96
	stp x22, x21, [sp, #64]
	stp x20, x19, [sp, #80]
	.cfi_offset w19, -8
	.cfi_offset w20, -16
	.cfi_offset w21, -24
	.cfi_offset w22, -32
	ldp x11, x9, [sp, #104]
	ldr x12, [sp, #96]
	lsr x8, x2, #1
	movi.2d v0, #0000000000000000
	cmp x9, #1
	b.ne LBB22_3
	cmp x2, #2
	b.hs LBB22_5
	movi.2d v1, #0000000000000000
	movi.2d v2, #0000000000000000
	movi.2d v3, #0000000000000000
	b LBB22_7
LBB22_3:
	lsl x10, x9, #1
	cmp x2, #2
	b.hs LBB22_9
	movi.2d v1, #0000000000000000
	movi.2d v2, #0000000000000000
	movi.2d v3, #0000000000000000
	mov x11, x5
	b LBB22_11
LBB22_5:
	lsl x9, x12, #1
	lsl x10, x11, #1
	movi.2d v1, #0000000000000000
	movi.2d v2, #0000000000000000
	movi.2d v3, #0000000000000000
LBB22_6:
	ldr q4, [x5]
	dup.8h v5, v4[0]
	ldr q6, [x4]
	; InlineAsm Start
	fmla.8h v3, v6, v5
	; InlineAsm End
	dup.8h v5, v4[1]
	; InlineAsm Start
	fmla.8h v2, v6, v5
	; InlineAsm End
	dup.8h v5, v4[2]
	; InlineAsm Start
	fmla.8h v1, v6, v5
	; InlineAsm End
	dup.8h v4, v4[3]
	; InlineAsm Start
	fmla.8h v0, v6, v4
	; InlineAsm End
	ldr q4, [x5, x10]
	dup.8h v5, v4[0]
	ldr q6, [x4, x9]
	; InlineAsm Start
	fmla.8h v3, v6, v5
	; InlineAsm End
	dup.8h v5, v4[1]
	; InlineAsm Start
	fmla.8h v2, v6, v5
	; InlineAsm End
	dup.8h v5, v4[2]
	; InlineAsm Start
	fmla.8h v1, v6, v5
	; InlineAsm End
	dup.8h v4, v4[3]
	; InlineAsm Start
	fmla.8h v0, v6, v4
	; InlineAsm End
	add x4, x4, x9, lsl #1
	add x5, x5, x10, lsl #1
	subs x8, x8, #1
	b.ne LBB22_6
LBB22_7:
	tbz w2, #0, LBB22_13
	ldr q4, [x5]
	dup.8h v5, v4[0]
	ldr q6, [x4]
	; InlineAsm Start
	fmla.8h v3, v6, v5
	; InlineAsm End
	dup.8h v5, v4[1]
	; InlineAsm Start
	fmla.8h v2, v6, v5
	; InlineAsm End
	dup.8h v5, v4[2]
	; InlineAsm Start
	fmla.8h v1, v6, v5
	; InlineAsm End
	dup.8h v4, v4[3]
	; InlineAsm Start
	fmla.8h v0, v6, v4
	; InlineAsm End
	b LBB22_13
LBB22_9:
	lsl x12, x12, #1
	lsl x13, x11, #1
	add x11, x10, x9
	lsl x14, x11, #1
	add x15, x14, x13
	lsl x16, x9, #2
	add x17, x16, x13
	add x19, x10, x13
	lsl x20, x13, #1
	movi.2d v1, #0000000000000000
	movi.2d v2, #0000000000000000
	movi.2d v3, #0000000000000000
	mov x11, x5
LBB22_10:
	ld1r.8h { v4 }, [x11], x20
	ldr q5, [x4]
	; InlineAsm Start
	fmla.8h v3, v5, v4
	; InlineAsm End
	add x21, x5, x10
	ld1r.8h { v4 }, [x21]
	; InlineAsm Start
	fmla.8h v2, v5, v4
	; InlineAsm End
	add x21, x5, x16
	ld1r.8h { v4 }, [x21]
	; InlineAsm Start
	fmla.8h v1, v5, v4
	; InlineAsm End
	add x21, x5, x14
	ld1r.8h { v4 }, [x21]
	; InlineAsm Start
	fmla.8h v0, v5, v4
	; InlineAsm End
	add x21, x5, x13
	ld1r.8h { v4 }, [x21]
	ldr q5, [x4, x12]
	; InlineAsm Start
	fmla.8h v3, v5, v4
	; InlineAsm End
	add x21, x5, x19
	ld1r.8h { v4 }, [x21]
	; InlineAsm Start
	fmla.8h v2, v5, v4
	; InlineAsm End
	add x21, x5, x17
	ld1r.8h { v4 }, [x21]
	; InlineAsm Start
	fmla.8h v1, v5, v4
	; InlineAsm End
	add x5, x5, x15
	ld1r.8h { v4 }, [x5]
	; InlineAsm Start
	fmla.8h v0, v5, v4
	; InlineAsm End
	add x4, x4, x12, lsl #1
	mov x5, x11
	subs x8, x8, #1
	b.ne LBB22_10
LBB22_11:
	tbz w2, #0, LBB22_13
	add x8, x10, x9
	lsl x8, x8, #1
	mov x12, x11
	ld1r.8h { v4 }, [x12], x8
	ldr q5, [x4]
	; InlineAsm Start
	fmla.8h v3, v5, v4
	; InlineAsm End
	add x8, x11, x10
	ld1r.8h { v4 }, [x8]
	; InlineAsm Start
	fmla.8h v2, v5, v4
	; InlineAsm End
	add x8, x11, x9, lsl #2
	ld1r.8h { v4 }, [x8]
	; InlineAsm Start
	fmla.8h v1, v5, v4
	; InlineAsm End
	ld1r.8h { v4 }, [x12]
	; InlineAsm Start
	fmla.8h v0, v5, v4
	; InlineAsm End
LBB22_13:
	ldrh w8, [sp, #122]
	ldrh w9, [sp, #120]
	ldrb w10, [sp, #124]
	cmp x0, #8
	b.ne LBB22_19
	cmp x1, #4
	b.ne LBB22_19
	cmp x7, #1
	b.ne LBB22_19
	dup.8h v4, w8
	and w8, w10, #0xff
	cmp w8, #1
	b.eq LBB22_27
	cmp w8, #2
	b.ne LBB22_28
	dup.8h v5, w9
	ldr q6, [x3]
	; InlineAsm Start
	fmul.8h v7, v5, v6
	; InlineAsm End
	; InlineAsm Start
	fmul.8h v6, v4, v3
	; InlineAsm End
	; InlineAsm Start
	fadd.8h v3, v7, v6
	; InlineAsm End
	str q3, [x3]
	lsl x8, x6, #1
	ldr q3, [x3, x8]
	; InlineAsm Start
	fmul.8h v6, v5, v3
	; InlineAsm End
	; InlineAsm Start
	fmul.8h v3, v4, v2
	; InlineAsm End
	; InlineAsm Start
	fadd.8h v2, v6, v3
	; InlineAsm End
	str q2, [x3, x8]
	lsl x8, x6, #2
	ldr q2, [x3, x8]
	; InlineAsm Start
	fmul.8h v3, v5, v2
	; InlineAsm End
	; InlineAsm Start
	fmul.8h v2, v4, v1
	; InlineAsm End
	; InlineAsm Start
	fadd.8h v1, v3, v2
	; InlineAsm End
	str q1, [x3, x8]
	mov w8, #6
	mul x8, x6, x8
	ldr q1, [x3, x8]
	; InlineAsm Start
	fmul.8h v2, v5, v1
	; InlineAsm End
	; InlineAsm Start
	fmul.8h v1, v4, v0
	; InlineAsm End
	; InlineAsm Start
	fadd.8h v0, v2, v1
	; InlineAsm End
	str q0, [x3, x8]
	b LBB22_42
LBB22_19:
	stp q3, q2, [sp]
	stp q1, q0, [sp, #32]
	and w10, w10, #0xff
	cmp w10, #2
	b.eq LBB22_30
	cmp w10, #1
	b.ne LBB22_36
	cbz x1, LBB22_42
	cbz x0, LBB22_42
	mov x9, #0
	lsl x10, x6, #1
	lsl x11, x7, #1
	mov x12, sp
LBB22_24:
	mov x13, x0
	mov x14, x3
	mov x15, x12
LBB22_25:
	ldr h0, [x15], #2
	fmov s1, w8
	; InlineAsm Start
	fmul h2, h1, h0
	; InlineAsm End
	ldr h0, [x14]
	; InlineAsm Start
	fadd h1, h0, h2
	; InlineAsm End
	str h1, [x14]
	add x14, x14, x11
	subs x13, x13, #1
	b.ne LBB22_25
	add x9, x9, #1
	add x12, x12, #16
	add x3, x3, x10
	cmp x9, x1
	b.ne LBB22_24
	b LBB22_42
LBB22_27:
	ldr q5, [x3]
	; InlineAsm Start
	fmla.8h v5, v4, v3
	; InlineAsm End
	str q5, [x3]
	lsl x8, x6, #1
	ldr q3, [x3, x8]
	; InlineAsm Start
	fmla.8h v3, v4, v2
	; InlineAsm End
	str q3, [x3, x8]
	lsl x8, x6, #2
	ldr q2, [x3, x8]
	; InlineAsm Start
	fmla.8h v2, v4, v1
	; InlineAsm End
	str q2, [x3, x8]
	mov w8, #6
	mul x8, x6, x8
	ldr q1, [x3, x8]
	; InlineAsm Start
	fmla.8h v1, v4, v0
	; InlineAsm End
	b LBB22_29
LBB22_28:
	; InlineAsm Start
	fmul.8h v5, v4, v3
	; InlineAsm End
	str q5, [x3]
	lsl x8, x6, #1
	; InlineAsm Start
	fmul.8h v3, v4, v2
	; InlineAsm End
	str q3, [x3, x8]
	lsl x8, x6, #2
	; InlineAsm Start
	fmul.8h v2, v4, v1
	; InlineAsm End
	str q2, [x3, x8]
	mov w8, #6
	mul x8, x6, x8
	; InlineAsm Start
	fmul.8h v1, v4, v0
	; InlineAsm End
LBB22_29:
	str q1, [x3, x8]
	b LBB22_42
LBB22_30:
	cbz x1, LBB22_42
	cbz x0, LBB22_42
	mov x10, #0
	lsl x11, x6, #1
	lsl x12, x7, #1
	mov x13, sp
LBB22_33:
	mov x14, x0
	mov x15, x3
	mov x16, x13
LBB22_34:
	ldr h0, [x15]
	fmov s1, w9
	; InlineAsm Start
	fmul h2, h1, h0
	; InlineAsm End
	ldr h0, [x16], #2
	fmov s1, w8
	; InlineAsm Start
	fmul h3, h1, h0
	; InlineAsm End
	; InlineAsm Start
	fadd h0, h2, h3
	; InlineAsm End
	str h0, [x15]
	add x15, x15, x12
	subs x14, x14, #1
	b.ne LBB22_34
	add x10, x10, #1
	add x13, x13, #16
	add x3, x3, x11
	cmp x10, x1
	b.ne LBB22_33
	b LBB22_42
LBB22_36:
	cbz x1, LBB22_42
	cbz x0, LBB22_42
	mov x9, #0
	lsl x10, x6, #1
	lsl x11, x7, #1
	mov x12, sp
LBB22_39:
	mov x13, x0
	mov x14, x3
	mov x15, x12
LBB22_40:
	ldr h0, [x15], #2
	fmov s1, w8
	; InlineAsm Start
	fmul h2, h1, h0
	; InlineAsm End
	str h2, [x14]
	add x14, x14, x11
	subs x13, x13, #1
	b.ne LBB22_40
	add x9, x9, #1
	add x12, x12, #16
	add x3, x3, x10
	cmp x9, x1
	b.ne LBB22_39
LBB22_42:
	ldp x20, x19, [sp, #80]
	ldp x22, x21, [sp, #64]
	add sp, sp, #96
	.cfi_def_cfa_offset 0
	.cfi_restore w19
	.cfi_restore w20
	.cfi_restore w21
	.cfi_restore w22
	ret
Lfunc_end22:
	.cfi_endproc

	.p2align	2
gemm_f16::microkernel::neon::f16::x1x5:
Lfunc_begin23:
	.cfi_startproc
	sub sp, sp, #80
	.cfi_def_cfa_offset 80
	ldr x8, [sp, #96]
	movi.2d v0, #0000000000000000
	cmp x2, #2
	b.hs LBB23_2
	movi.2d v1, #0000000000000000
	movi.2d v2, #0000000000000000
	movi.2d v3, #0000000000000000
	movi.2d v4, #0000000000000000
	tbnz w2, #0, LBB23_5
	b LBB23_6
LBB23_2:
	lsr x9, x2, #1
	ldp x11, x12, [sp, #80]
	lsl x10, x8, #1
	lsl x11, x11, #1
	lsl x13, x12, #1
	sub x12, x13, x8, lsl #3
	lsl x13, x13, #1
	movi.2d v1, #0000000000000000
	movi.2d v2, #0000000000000000
	movi.2d v3, #0000000000000000
	movi.2d v4, #0000000000000000
LBB23_3:
	mov x14, x5
	ld1r.8h { v5 }, [x14], x13
	ldr q6, [x4]
	; InlineAsm Start
	fmla.8h v4, v6, v5
	; InlineAsm End
	add x5, x5, x10
	ld1r.8h { v5 }, [x5], x10
	; InlineAsm Start
	fmla.8h v3, v6, v5
	; InlineAsm End
	ld1r.8h { v5 }, [x5], x10
	; InlineAsm Start
	fmla.8h v2, v6, v5
	; InlineAsm End
	ld1r.8h { v5 }, [x5], x10
	; InlineAsm Start
	fmla.8h v1, v6, v5
	; InlineAsm End
	ld1r.8h { v5 }, [x5], x12
	; InlineAsm Start
	fmla.8h v0, v6, v5
	; InlineAsm End
	ld1r.8h { v5 }, [x5], x10
	ldr q6, [x4, x11]
	; InlineAsm Start
	fmla.8h v4, v6, v5
	; InlineAsm End
	ld1r.8h { v5 }, [x5], x10
	; InlineAsm Start
	fmla.8h v3, v6, v5
	; InlineAsm End
	ld1r.8h { v5 }, [x5], x10
	; InlineAsm Start
	fmla.8h v2, v6, v5
	; InlineAsm End
	ld1r.8h { v5 }, [x5], x10
	; InlineAsm Start
	fmla.8h v1, v6, v5
	; InlineAsm End
	ld1r.8h { v5 }, [x5], x12
	; InlineAsm Start
	fmla.8h v0, v6, v5
	; InlineAsm End
	add x4, x4, x11, lsl #1
	subs x9, x9, #1
	b.ne LBB23_3
	mov x5, x14
	tbz w2, #0, LBB23_6

	lsl x9, x8, #3
	mov x10, x5
	ld1r.8h { v5 }, [x10], x9
	ldr q6, [x4]
	; InlineAsm Start
	fmla.8h v4, v6, v5
	; InlineAsm End
	add x9, x5, x8, lsl #1
	ld1r.8h { v5 }, [x9]
	; InlineAsm Start
	fmla.8h v3, v6, v5
	; InlineAsm End
	add x9, x5, x8, lsl #2
	ld1r.8h { v5 }, [x9]
	; InlineAsm Start
	fmla.8h v2, v6, v5
	; InlineAsm End
	mov w9, #6
	madd x8, x8, x9, x5
	ld1r.8h { v5 }, [x8]
	; InlineAsm Start
	fmla.8h v1, v6, v5
	; InlineAsm End
	ld1r.8h { v5 }, [x10]
	; InlineAsm Start
	fmla.8h v0, v6, v5
	; InlineAsm End
LBB23_6:
	ldrh w8, [sp, #106]
	ldrh w9, [sp, #104]
	ldrb w10, [sp, #108]
	cmp x0, #8
	b.ne LBB23_12
	cmp x1, #5
	b.ne LBB23_12
	cmp x7, #1
	b.ne LBB23_12
	dup.8h v5, w8
	and w8, w10, #0xff
	cmp w8, #1
	b.eq LBB23_20
	cmp w8, #2
	b.ne LBB23_21
	dup.8h v6, w9
	ldr q7, [x3]
	; InlineAsm Start
	fmul.8h v16, v6, v7
	; InlineAsm End
	; InlineAsm Start
	fmul.8h v7, v5, v4
	; InlineAsm End
	; InlineAsm Start
	fadd.8h v4, v16, v7
	; InlineAsm End
	str q4, [x3]
	lsl x8, x6, #1
	ldr q4, [x3, x8]
	; InlineAsm Start
	fmul.8h v7, v6, v4
	; InlineAsm End
	; InlineAsm Start
	fmul.8h v4, v5, v3
	; InlineAsm End
	; InlineAsm Start
	fadd.8h v3, v7, v4
	; InlineAsm End
	str q3, [x3, x8]
	lsl x8, x6, #2
	ldr q3, [x3, x8]
	; InlineAsm Start
	fmul.8h v4, v6, v3
	; InlineAsm End
	; InlineAsm Start
	fmul.8h v3, v5, v2
	; InlineAsm End
	; InlineAsm Start
	fadd.8h v2, v4, v3
	; InlineAsm End
	str q2, [x3, x8]
	mov w8, #6
	mul x8, x6, x8
	ldr q2, [x3, x8]
	; InlineAsm Start
	fmul.8h v3, v6, v2
	; InlineAsm End
	; InlineAsm Start
	fmul.8h v2, v5, v1
	; InlineAsm End
	; InlineAsm Start
	fadd.8h v1, v3, v2
	; InlineAsm End
	str q1, [x3, x8]
	lsl x8, x6, #3
	ldr q1, [x3, x8]
	; InlineAsm Start
	fmul.8h v2, v6, v1
	; InlineAsm End
	; InlineAsm Start
	fmul.8h v1, v5, v0
	; InlineAsm End
	; InlineAsm Start
	fadd.8h v0, v2, v1
	; InlineAsm End
	str q0, [x3, x8]
	b LBB23_35
LBB23_12:
	stp q4, q3, [sp]
	stp q2, q1, [sp, #32]
	and w10, w10, #0xff
	str q0, [sp, #64]
	cmp w10, #2
	b.eq LBB23_23
	cmp w10, #1
	b.ne LBB23_29
	cbz x1, LBB23_35
	cbz x0, LBB23_35
	mov x9, #0
	lsl x10, x6, #1
	lsl x11, x7, #1
	mov x12, sp
LBB23_17:
	mov x13, x0
	mov x14, x3
	mov x15, x12
LBB23_18:
	ldr h0, [x15], #2
	fmov s1, w8
	; InlineAsm Start
	fmul h2, h1, h0
	; InlineAsm End
	ldr h0, [x14]
	; InlineAsm Start
	fadd h1, h0, h2
	; InlineAsm End
	str h1, [x14]
	add x14, x14, x11
	subs x13, x13, #1
	b.ne LBB23_18
	add x9, x9, #1
	add x12, x12, #16
	add x3, x3, x10
	cmp x9, x1
	b.ne LBB23_17
	b LBB23_35
LBB23_20:
	ldr q6, [x3]
	; InlineAsm Start
	fmla.8h v6, v5, v4
	; InlineAsm End
	str q6, [x3]
	lsl x8, x6, #1
	ldr q4, [x3, x8]
	; InlineAsm Start
	fmla.8h v4, v5, v3
	; InlineAsm End
	str q4, [x3, x8]
	lsl x8, x6, #2
	ldr q3, [x3, x8]
	; InlineAsm Start
	fmla.8h v3, v5, v2
	; InlineAsm End
	str q3, [x3, x8]
	mov w8, #6
	mul x8, x6, x8
	ldr q2, [x3, x8]
	; InlineAsm Start
	fmla.8h v2, v5, v1
	; InlineAsm End
	str q2, [x3, x8]
	lsl x8, x6, #3
	ldr q1, [x3, x8]
	; InlineAsm Start
	fmla.8h v1, v5, v0
	; InlineAsm End
	b LBB23_22
LBB23_21:
	; InlineAsm Start
	fmul.8h v6, v5, v4
	; InlineAsm End
	str q6, [x3]
	lsl x8, x6, #1
	; InlineAsm Start
	fmul.8h v4, v5, v3
	; InlineAsm End
	str q4, [x3, x8]
	lsl x8, x6, #2
	; InlineAsm Start
	fmul.8h v3, v5, v2
	; InlineAsm End
	str q3, [x3, x8]
	mov w8, #6
	mul x8, x6, x8
	; InlineAsm Start
	fmul.8h v2, v5, v1
	; InlineAsm End
	str q2, [x3, x8]
	lsl x8, x6, #3
	; InlineAsm Start
	fmul.8h v1, v5, v0
	; InlineAsm End
LBB23_22:
	str q1, [x3, x8]
	b LBB23_35
LBB23_23:
	cbz x1, LBB23_35
	cbz x0, LBB23_35
	mov x10, #0
	lsl x11, x6, #1
	lsl x12, x7, #1
	mov x13, sp
LBB23_26:
	mov x14, x0
	mov x15, x3
	mov x16, x13
LBB23_27:
	ldr h0, [x15]
	fmov s1, w9
	; InlineAsm Start
	fmul h2, h1, h0
	; InlineAsm End
	ldr h0, [x16], #2
	fmov s1, w8
	; InlineAsm Start
	fmul h3, h1, h0
	; InlineAsm End
	; InlineAsm Start
	fadd h0, h2, h3
	; InlineAsm End
	str h0, [x15]
	add x15, x15, x12
	subs x14, x14, #1
	b.ne LBB23_27
	add x10, x10, #1
	add x13, x13, #16
	add x3, x3, x11
	cmp x10, x1
	b.ne LBB23_26
	b LBB23_35
LBB23_29:
	cbz x1, LBB23_35
	cbz x0, LBB23_35
	mov x9, #0
	lsl x10, x6, #1
	lsl x11, x7, #1
	mov x12, sp
LBB23_32:
	mov x13, x0
	mov x14, x3
	mov x15, x12
LBB23_33:
	ldr h0, [x15], #2
	fmov s1, w8
	; InlineAsm Start
	fmul h2, h1, h0
	; InlineAsm End
	str h2, [x14]
	add x14, x14, x11
	subs x13, x13, #1
	b.ne LBB23_33
	add x9, x9, #1
	add x12, x12, #16
	add x3, x3, x10
	cmp x9, x1
	b.ne LBB23_32
LBB23_35:
	add sp, sp, #80
	.cfi_def_cfa_offset 0
	ret
Lfunc_end23:
	.cfi_endproc

	.p2align	2
gemm_f16::microkernel::neon::f16::x1x6:
Lfunc_begin24:
	.cfi_startproc
	sub sp, sp, #96
	.cfi_def_cfa_offset 96
	ldr x8, [sp, #112]
	movi.2d v0, #0000000000000000
	cmp x2, #2
	b.hs LBB24_2
	movi.2d v1, #0000000000000000
	movi.2d v2, #0000000000000000
	movi.2d v3, #0000000000000000
	movi.2d v4, #0000000000000000
	movi.2d v5, #0000000000000000
	tbnz w2, #0, LBB24_5
	b LBB24_6
LBB24_2:
	ldp x11, x12, [sp, #96]
	lsr x9, x2, #1
	lsl x10, x8, #1
	lsl x11, x11, #1
	lsl x13, x12, #1
	mov w12, #10
	msub x12, x8, x12, x13
	lsl x13, x13, #1
	movi.2d v1, #0000000000000000
	movi.2d v2, #0000000000000000
	movi.2d v3, #0000000000000000
	movi.2d v4, #0000000000000000
	movi.2d v5, #0000000000000000
LBB24_3:
	mov x14, x5
	ld1r.8h { v6 }, [x14], x13
	ldr q7, [x4]
	; InlineAsm Start
	fmla.8h v5, v7, v6
	; InlineAsm End
	add x5, x5, x10
	ld1r.8h { v6 }, [x5], x10
	; InlineAsm Start
	fmla.8h v4, v7, v6
	; InlineAsm End
	ld1r.8h { v6 }, [x5], x10
	; InlineAsm Start
	fmla.8h v3, v7, v6
	; InlineAsm End
	ld1r.8h { v6 }, [x5], x10
	; InlineAsm Start
	fmla.8h v2, v7, v6
	; InlineAsm End
	ld1r.8h { v6 }, [x5], x10
	; InlineAsm Start
	fmla.8h v1, v7, v6
	; InlineAsm End
	ld1r.8h { v6 }, [x5], x12
	; InlineAsm Start
	fmla.8h v0, v7, v6
	; InlineAsm End
	ld1r.8h { v6 }, [x5], x10
	ldr q7, [x4, x11]
	; InlineAsm Start
	fmla.8h v5, v7, v6
	; InlineAsm End
	ld1r.8h { v6 }, [x5], x10
	; InlineAsm Start
	fmla.8h v4, v7, v6
	; InlineAsm End
	ld1r.8h { v6 }, [x5], x10
	; InlineAsm Start
	fmla.8h v3, v7, v6
	; InlineAsm End
	ld1r.8h { v6 }, [x5], x10
	; InlineAsm Start
	fmla.8h v2, v7, v6
	; InlineAsm End
	ld1r.8h { v6 }, [x5], x10
	; InlineAsm Start
	fmla.8h v1, v7, v6
	; InlineAsm End
	ld1r.8h { v6 }, [x5], x12
	; InlineAsm Start
	fmla.8h v0, v7, v6
	; InlineAsm End
	add x4, x4, x11, lsl #1
	subs x9, x9, #1
	b.ne LBB24_3
	mov x5, x14
	tbz w2, #0, LBB24_6

	lsl x9, x8, #2
	add x10, x9, x8
	lsl x10, x10, #1
	mov x11, x5
	ld1r.8h { v6 }, [x11], x10
	ldr q7, [x4]
	; InlineAsm Start
	fmla.8h v5, v7, v6
	; InlineAsm End
	add x10, x5, x8, lsl #1
	ld1r.8h { v6 }, [x10]
	; InlineAsm Start
	fmla.8h v4, v7, v6
	; InlineAsm End
	add x9, x5, x9
	ld1r.8h { v6 }, [x9]
	; InlineAsm Start
	fmla.8h v3, v7, v6
	; InlineAsm End
	mov w9, #6
	madd x9, x8, x9, x5
	ld1r.8h { v6 }, [x9]
	; InlineAsm Start
	fmla.8h v2, v7, v6
	; InlineAsm End
	add x8, x5, x8, lsl #3
	ld1r.8h { v6 }, [x8]
	; InlineAsm Start
	fmla.8h v1, v7, v6
	; InlineAsm End
	ld1r.8h { v6 }, [x11]
	; InlineAsm Start
	fmla.8h v0, v7, v6
	; InlineAsm End
LBB24_6:
	ldrh w8, [sp, #122]
	ldrh w9, [sp, #120]
	ldrb w10, [sp, #124]
	cmp x0, #8
	b.ne LBB24_12
	cmp x1, #6
	b.ne LBB24_12
	cmp x7, #1
	b.ne LBB24_12
	dup.8h v6, w8
	and w8, w10, #0xff
	cmp w8, #1
	b.eq LBB24_20
	cmp w8, #2
	b.ne LBB24_21
	dup.8h v7, w9
	ldr q16, [x3]
	; InlineAsm Start
	fmul.8h v17, v7, v16
	; InlineAsm End
	; InlineAsm Start
	fmul.8h v16, v6, v5
	; InlineAsm End
	; InlineAsm Start
	fadd.8h v5, v17, v16
	; InlineAsm End
	str q5, [x3]
	lsl x8, x6, #1
	ldr q5, [x3, x8]
	; InlineAsm Start
	fmul.8h v16, v7, v5
	; InlineAsm End
	; InlineAsm Start
	fmul.8h v5, v6, v4
	; InlineAsm End
	; InlineAsm Start
	fadd.8h v4, v16, v5
	; InlineAsm End
	str q4, [x3, x8]
	lsl x8, x6, #2
	ldr q4, [x3, x8]
	; InlineAsm Start
	fmul.8h v5, v7, v4
	; InlineAsm End
	; InlineAsm Start
	fmul.8h v4, v6, v3
	; InlineAsm End
	; InlineAsm Start
	fadd.8h v3, v5, v4
	; InlineAsm End
	str q3, [x3, x8]
	mov w8, #6
	mul x8, x6, x8
	ldr q3, [x3, x8]
	; InlineAsm Start
	fmul.8h v4, v7, v3
	; InlineAsm End
	; InlineAsm Start
	fmul.8h v3, v6, v2
	; InlineAsm End
	; InlineAsm Start
	fadd.8h v2, v4, v3
	; InlineAsm End
	str q2, [x3, x8]
	lsl x8, x6, #3
	ldr q2, [x3, x8]
	; InlineAsm Start
	fmul.8h v3, v7, v2
	; InlineAsm End
	; InlineAsm Start
	fmul.8h v2, v6, v1
	; InlineAsm End
	; InlineAsm Start
	fadd.8h v1, v3, v2
	; InlineAsm End
	str q1, [x3, x8]
	mov w8, #10
	mul x8, x6, x8
	ldr q1, [x3, x8]
	; InlineAsm Start
	fmul.8h v2, v7, v1
	; InlineAsm End
	; InlineAsm Start
	fmul.8h v1, v6, v0
	; InlineAsm End
	; InlineAsm Start
	fadd.8h v0, v2, v1
	; InlineAsm End
	str q0, [x3, x8]
	b LBB24_35
LBB24_12:
	stp q5, q4, [sp]
	stp q3, q2, [sp, #32]
	and w10, w10, #0xff
	stp q1, q0, [sp, #64]
	cmp w10, #2
	b.eq LBB24_23
	cmp w10, #1
	b.ne LBB24_29
	cbz x1, LBB24_35
	cbz x0, LBB24_35
	mov x9, #0
	lsl x10, x6, #1
	lsl x11, x7, #1
	mov x12, sp
LBB24_17:
	mov x13, x0
	mov x14, x3
	mov x15, x12
LBB24_18:
	ldr h0, [x15], #2
	fmov s1, w8
	; InlineAsm Start
	fmul h2, h1, h0
	; InlineAsm End
	ldr h0, [x14]
	; InlineAsm Start
	fadd h1, h0, h2
	; InlineAsm End
	str h1, [x14]
	add x14, x14, x11
	subs x13, x13, #1
	b.ne LBB24_18
	add x9, x9, #1
	add x12, x12, #16
	add x3, x3, x10
	cmp x9, x1
	b.ne LBB24_17
	b LBB24_35
LBB24_20:
	ldr q7, [x3]
	; InlineAsm Start
	fmla.8h v7, v6, v5
	; InlineAsm End
	str q7, [x3]
	lsl x8, x6, #1
	ldr q5, [x3, x8]
	; InlineAsm Start
	fmla.8h v5, v6, v4
	; InlineAsm End
	str q5, [x3, x8]
	lsl x8, x6, #2
	ldr q4, [x3, x8]
	; InlineAsm Start
	fmla.8h v4, v6, v3
	; InlineAsm End
	str q4, [x3, x8]
	mov w8, #6
	mul x8, x6, x8
	ldr q3, [x3, x8]
	; InlineAsm Start
	fmla.8h v3, v6, v2
	; InlineAsm End
	str q3, [x3, x8]
	lsl x8, x6, #3
	ldr q2, [x3, x8]
	; InlineAsm Start
	fmla.8h v2, v6, v1
	; InlineAsm End
	str q2, [x3, x8]
	mov w8, #10
	mul x8, x6, x8
	ldr q1, [x3, x8]
	; InlineAsm Start
	fmla.8h v1, v6, v0
	; InlineAsm End
	b LBB24_22
LBB24_21:
	; InlineAsm Start
	fmul.8h v7, v6, v5
	; InlineAsm End
	str q7, [x3]
	lsl x8, x6, #1
	; InlineAsm Start
	fmul.8h v5, v6, v4
	; InlineAsm End
	str q5, [x3, x8]
	lsl x8, x6, #2
	; InlineAsm Start
	fmul.8h v4, v6, v3
	; InlineAsm End
	str q4, [x3, x8]
	mov w8, #6
	mul x8, x6, x8
	; InlineAsm Start
	fmul.8h v3, v6, v2
	; InlineAsm End
	str q3, [x3, x8]
	lsl x8, x6, #3
	; InlineAsm Start
	fmul.8h v2, v6, v1
	; InlineAsm End
	str q2, [x3, x8]
	mov w8, #10
	mul x8, x6, x8
	; InlineAsm Start
	fmul.8h v1, v6, v0
	; InlineAsm End
LBB24_22:
	str q1, [x3, x8]
	b LBB24_35
LBB24_23:
	cbz x1, LBB24_35
	cbz x0, LBB24_35
	mov x10, #0
	lsl x11, x6, #1
	lsl x12, x7, #1
	mov x13, sp
LBB24_26:
	mov x14, x0
	mov x15, x3
	mov x16, x13
LBB24_27:
	ldr h0, [x15]
	fmov s1, w9
	; InlineAsm Start
	fmul h2, h1, h0
	; InlineAsm End
	ldr h0, [x16], #2
	fmov s1, w8
	; InlineAsm Start
	fmul h3, h1, h0
	; InlineAsm End
	; InlineAsm Start
	fadd h0, h2, h3
	; InlineAsm End
	str h0, [x15]
	add x15, x15, x12
	subs x14, x14, #1
	b.ne LBB24_27
	add x10, x10, #1
	add x13, x13, #16
	add x3, x3, x11
	cmp x10, x1
	b.ne LBB24_26
	b LBB24_35
LBB24_29:
	cbz x1, LBB24_35
	cbz x0, LBB24_35
	mov x9, #0
	lsl x10, x6, #1
	lsl x11, x7, #1
	mov x12, sp
LBB24_32:
	mov x13, x0
	mov x14, x3
	mov x15, x12
LBB24_33:
	ldr h0, [x15], #2
	fmov s1, w8
	; InlineAsm Start
	fmul h2, h1, h0
	; InlineAsm End
	str h2, [x14]
	add x14, x14, x11
	subs x13, x13, #1
	b.ne LBB24_33
	add x9, x9, #1
	add x12, x12, #16
	add x3, x3, x10
	cmp x9, x1
	b.ne LBB24_32
LBB24_35:
	add sp, sp, #96
	.cfi_def_cfa_offset 0
	ret
Lfunc_end24:
	.cfi_endproc

	.p2align	2
gemm_f16::microkernel::neon::f16::x1x7:
Lfunc_begin25:
	.cfi_startproc
	sub sp, sp, #112
	.cfi_def_cfa_offset 112
	ldr x8, [sp, #128]
	movi.2d v0, #0000000000000000
	lsl x9, x8, #1
	cmp x2, #2
	b.hs LBB25_2
	movi.2d v1, #0000000000000000
	movi.2d v2, #0000000000000000
	movi.2d v3, #0000000000000000
	movi.2d v4, #0000000000000000
	movi.2d v5, #0000000000000000
	movi.2d v6, #0000000000000000
	tbnz w2, #0, LBB25_5
	b LBB25_6
LBB25_2:
	ldp x11, x12, [sp, #112]
	lsr x10, x2, #1
	lsl x11, x11, #1
	lsl x13, x12, #1
	mov w12, #12
	msub x12, x8, x12, x13
	lsl x13, x13, #1
	movi.2d v1, #0000000000000000
	movi.2d v2, #0000000000000000
	movi.2d v3, #0000000000000000
	movi.2d v4, #0000000000000000
	movi.2d v5, #0000000000000000
	movi.2d v6, #0000000000000000
LBB25_3:
	mov x14, x5
	ld1r.8h { v7 }, [x14], x13
	ldr q16, [x4]
	; InlineAsm Start
	fmla.8h v6, v16, v7
	; InlineAsm End
	add x5, x5, x9
	ld1r.8h { v7 }, [x5], x9
	; InlineAsm Start
	fmla.8h v5, v16, v7
	; InlineAsm End
	ld1r.8h { v7 }, [x5], x9
	; InlineAsm Start
	fmla.8h v4, v16, v7
	; InlineAsm End
	ld1r.8h { v7 }, [x5], x9
	; InlineAsm Start
	fmla.8h v3, v16, v7
	; InlineAsm End
	ld1r.8h { v7 }, [x5], x9
	; InlineAsm Start
	fmla.8h v2, v16, v7
	; InlineAsm End
	ld1r.8h { v7 }, [x5], x9
	; InlineAsm Start
	fmla.8h v1, v16, v7
	; InlineAsm End
	ld1r.8h { v7 }, [x5], x12
	; InlineAsm Start
	fmla.8h v0, v16, v7
	; InlineAsm End
	ld1r.8h { v7 }, [x5], x9
	ldr q16, [x4, x11]
	; InlineAsm Start
	fmla.8h v6, v16, v7
	; InlineAsm End
	ld1r.8h { v7 }, [x5], x9
	; InlineAsm Start
	fmla.8h v5, v16, v7
	; InlineAsm End
	ld1r.8h { v7 }, [x5], x9
	; InlineAsm Start
	fmla.8h v4, v16, v7
	; InlineAsm End
	ld1r.8h { v7 }, [x5], x9
	; InlineAsm Start
	fmla.8h v3, v16, v7
	; InlineAsm End
	ld1r.8h { v7 }, [x5], x9
	; InlineAsm Start
	fmla.8h v2, v16, v7
	; InlineAsm End
	ld1r.8h { v7 }, [x5], x9
	; InlineAsm Start
	fmla.8h v1, v16, v7
	; InlineAsm End
	ld1r.8h { v7 }, [x5], x12
	; InlineAsm Start
	fmla.8h v0, v16, v7
	; InlineAsm End
	add x4, x4, x11, lsl #1
	subs x10, x10, #1
	b.ne LBB25_3
	mov x5, x14
	tbz w2, #0, LBB25_6

	add x10, x9, x8
	lsl x10, x10, #2
	mov x11, x5
	ld1r.8h { v7 }, [x11], x10
	ldr q16, [x4]
	; InlineAsm Start
	fmla.8h v6, v16, v7
	; InlineAsm End
	add x9, x5, x9
	ld1r.8h { v7 }, [x9]
	; InlineAsm Start
	fmla.8h v5, v16, v7
	; InlineAsm End
	add x9, x5, x8, lsl #2
	ld1r.8h { v7 }, [x9]
	; InlineAsm Start
	fmla.8h v4, v16, v7
	; InlineAsm End
	mov w9, #6
	madd x9, x8, x9, x5
	ld1r.8h { v7 }, [x9]
	; InlineAsm Start
	fmla.8h v3, v16, v7
	; InlineAsm End
	add x9, x5, x8, lsl #3
	ld1r.8h { v7 }, [x9]
	; InlineAsm Start
	fmla.8h v2, v16, v7
	; InlineAsm End
	mov w9, #10
	madd x8, x8, x9, x5
	ld1r.8h { v7 }, [x8]
	; InlineAsm Start
	fmla.8h v1, v16, v7
	; InlineAsm End
	ld1r.8h { v7 }, [x11]
	; InlineAsm Start
	fmla.8h v0, v16, v7
	; InlineAsm End
LBB25_6:
	ldrh w8, [sp, #138]
	ldrh w9, [sp, #136]
	ldrb w10, [sp, #140]
	cmp x0, #8
	b.ne LBB25_12
	cmp x1, #7
	b.ne LBB25_12
	cmp x7, #1
	b.ne LBB25_12
	dup.8h v7, w8
	and w8, w10, #0xff
	cmp w8, #1
	b.eq LBB25_20
	cmp w8, #2
	b.ne LBB25_21
	dup.8h v16, w9
	ldr q17, [x3]
	; InlineAsm Start
	fmul.8h v18, v16, v17
	; InlineAsm End
	; InlineAsm Start
	fmul.8h v17, v7, v6
	; InlineAsm End
	; InlineAsm Start
	fadd.8h v6, v18, v17
	; InlineAsm End
	str q6, [x3]
	lsl x8, x6, #1
	ldr q6, [x3, x8]
	; InlineAsm Start
	fmul.8h v17, v16, v6
	; InlineAsm End
	; InlineAsm Start
	fmul.8h v6, v7, v5
	; InlineAsm End
	; InlineAsm Start
	fadd.8h v5, v17, v6
	; InlineAsm End
	str q5, [x3, x8]
	lsl x8, x6, #2
	ldr q5, [x3, x8]
	; InlineAsm Start
	fmul.8h v6, v16, v5
	; InlineAsm End
	; InlineAsm Start
	fmul.8h v5, v7, v4
	; InlineAsm End
	; InlineAsm Start
	fadd.8h v4, v6, v5
	; InlineAsm End
	str q4, [x3, x8]
	mov w8, #6
	mul x8, x6, x8
	ldr q4, [x3, x8]
	; InlineAsm Start
	fmul.8h v5, v16, v4
	; InlineAsm End
	; InlineAsm Start
	fmul.8h v4, v7, v3
	; InlineAsm End
	; InlineAsm Start
	fadd.8h v3, v5, v4
	; InlineAsm End
	str q3, [x3, x8]
	lsl x8, x6, #3
	ldr q3, [x3, x8]
	; InlineAsm Start
	fmul.8h v4, v16, v3
	; InlineAsm End
	; InlineAsm Start
	fmul.8h v3, v7, v2
	; InlineAsm End
	; InlineAsm Start
	fadd.8h v2, v4, v3
	; InlineAsm End
	str q2, [x3, x8]
	mov w8, #10
	mul x8, x6, x8
	ldr q2, [x3, x8]
	; InlineAsm Start
	fmul.8h v3, v16, v2
	; InlineAsm End
	; InlineAsm Start
	fmul.8h v2, v7, v1
	; InlineAsm End
	; InlineAsm Start
	fadd.8h v1, v3, v2
	; InlineAsm End
	str q1, [x3, x8]
	mov w8, #12
	mul x8, x6, x8
	ldr q1, [x3, x8]
	; InlineAsm Start
	fmul.8h v2, v16, v1
	; InlineAsm End
	; InlineAsm Start
	fmul.8h v1, v7, v0
	; InlineAsm End
	; InlineAsm Start
	fadd.8h v0, v2, v1
	; InlineAsm End
	str q0, [x3, x8]
	b LBB25_35
LBB25_12:
	stp q6, q5, [sp]
	stp q4, q3, [sp, #32]
	stp q2, q1, [sp, #64]
	and w10, w10, #0xff
	str q0, [sp, #96]
	cmp w10, #2
	b.eq LBB25_23
	cmp w10, #1
	b.ne LBB25_29
	cbz x1, LBB25_35
	cbz x0, LBB25_35
	mov x9, #0
	lsl x10, x6, #1
	lsl x11, x7, #1
	mov x12, sp
LBB25_17:
	mov x13, x0
	mov x14, x3
	mov x15, x12
LBB25_18:
	ldr h0, [x15], #2
	fmov s1, w8
	; InlineAsm Start
	fmul h2, h1, h0
	; InlineAsm End
	ldr h0, [x14]
	; InlineAsm Start
	fadd h1, h0, h2
	; InlineAsm End
	str h1, [x14]
	add x14, x14, x11
	subs x13, x13, #1
	b.ne LBB25_18
	add x9, x9, #1
	add x12, x12, #16
	add x3, x3, x10
	cmp x9, x1
	b.ne LBB25_17
	b LBB25_35
LBB25_20:
	ldr q16, [x3]
	; InlineAsm Start
	fmla.8h v16, v7, v6
	; InlineAsm End
	str q16, [x3]
	lsl x8, x6, #1
	ldr q6, [x3, x8]
	; InlineAsm Start
	fmla.8h v6, v7, v5
	; InlineAsm End
	str q6, [x3, x8]
	lsl x8, x6, #2
	ldr q5, [x3, x8]
	; InlineAsm Start
	fmla.8h v5, v7, v4
	; InlineAsm End
	str q5, [x3, x8]
	mov w8, #6
	mul x8, x6, x8
	ldr q4, [x3, x8]
	; InlineAsm Start
	fmla.8h v4, v7, v3
	; InlineAsm End
	str q4, [x3, x8]
	lsl x8, x6, #3
	ldr q3, [x3, x8]
	; InlineAsm Start
	fmla.8h v3, v7, v2
	; InlineAsm End
	str q3, [x3, x8]
	mov w8, #10
	mul x8, x6, x8
	ldr q2, [x3, x8]
	; InlineAsm Start
	fmla.8h v2, v7, v1
	; InlineAsm End
	str q2, [x3, x8]
	mov w8, #12
	mul x8, x6, x8
	ldr q1, [x3, x8]
	; InlineAsm Start
	fmla.8h v1, v7, v0
	; InlineAsm End
	b LBB25_22
LBB25_21:
	; InlineAsm Start
	fmul.8h v16, v7, v6
	; InlineAsm End
	str q16, [x3]
	lsl x8, x6, #1
	; InlineAsm Start
	fmul.8h v6, v7, v5
	; InlineAsm End
	str q6, [x3, x8]
	lsl x8, x6, #2
	; InlineAsm Start
	fmul.8h v5, v7, v4
	; InlineAsm End
	str q5, [x3, x8]
	mov w8, #6
	mul x8, x6, x8
	; InlineAsm Start
	fmul.8h v4, v7, v3
	; InlineAsm End
	str q4, [x3, x8]
	lsl x8, x6, #3
	; InlineAsm Start
	fmul.8h v3, v7, v2
	; InlineAsm End
	str q3, [x3, x8]
	mov w8, #10
	mul x8, x6, x8
	; InlineAsm Start
	fmul.8h v2, v7, v1
	; InlineAsm End
	str q2, [x3, x8]
	mov w8, #12
	mul x8, x6, x8
	; InlineAsm Start
	fmul.8h v1, v7, v0
	; InlineAsm End
LBB25_22:
	str q1, [x3, x8]
	b LBB25_35
LBB25_23:
	cbz x1, LBB25_35
	cbz x0, LBB25_35
	mov x10, #0
	lsl x11, x6, #1
	lsl x12, x7, #1
	mov x13, sp
LBB25_26:
	mov x14, x0
	mov x15, x3
	mov x16, x13
LBB25_27:
	ldr h0, [x15]
	fmov s1, w9
	; InlineAsm Start
	fmul h2, h1, h0
	; InlineAsm End
	ldr h0, [x16], #2
	fmov s1, w8
	; InlineAsm Start
	fmul h3, h1, h0
	; InlineAsm End
	; InlineAsm Start
	fadd h0, h2, h3
	; InlineAsm End
	str h0, [x15]
	add x15, x15, x12
	subs x14, x14, #1
	b.ne LBB25_27
	add x10, x10, #1
	add x13, x13, #16
	add x3, x3, x11
	cmp x10, x1
	b.ne LBB25_26
	b LBB25_35
LBB25_29:
	cbz x1, LBB25_35
	cbz x0, LBB25_35
	mov x9, #0
	lsl x10, x6, #1
	lsl x11, x7, #1
	mov x12, sp
LBB25_32:
	mov x13, x0
	mov x14, x3
	mov x15, x12
LBB25_33:
	ldr h0, [x15], #2
	fmov s1, w8
	; InlineAsm Start
	fmul h2, h1, h0
	; InlineAsm End
	str h2, [x14]
	add x14, x14, x11
	subs x13, x13, #1
	b.ne LBB25_33
	add x9, x9, #1
	add x12, x12, #16
	add x3, x3, x10
	cmp x9, x1
	b.ne LBB25_32
LBB25_35:
	add sp, sp, #112
	.cfi_def_cfa_offset 0
	ret
Lfunc_end25:
	.cfi_endproc

	.p2align	2
gemm_f16::microkernel::neon::f16::x1x8:
Lfunc_begin26:
	.cfi_startproc
	sub sp, sp, #128
	.cfi_def_cfa_offset 128
	ldp x12, x9, [sp, #136]
	ldr x11, [sp, #128]
	lsr x8, x2, #1
	movi.2d v0, #0000000000000000
	cmp x9, #1
	b.ne LBB26_3
	cmp x2, #2
	b.hs LBB26_5
	movi.2d v1, #0000000000000000
	movi.2d v2, #0000000000000000
	movi.2d v3, #0000000000000000
	movi.2d v4, #0000000000000000
	movi.2d v5, #0000000000000000
	movi.2d v6, #0000000000000000
	movi.2d v7, #0000000000000000
	tbnz w2, #0, LBB26_8
	b LBB26_13
LBB26_3:
	lsl x10, x9, #1
	cmp x2, #2
	b.hs LBB26_9
	movi.2d v1, #0000000000000000
	movi.2d v2, #0000000000000000
	movi.2d v3, #0000000000000000
	movi.2d v4, #0000000000000000
	movi.2d v5, #0000000000000000
	movi.2d v6, #0000000000000000
	movi.2d v7, #0000000000000000
	tbnz w2, #0, LBB26_12
	b LBB26_13
LBB26_5:
	lsl x9, x11, #1
	movi.2d v1, #0000000000000000
	lsl x10, x12, #1
	movi.2d v2, #0000000000000000
	movi.2d v3, #0000000000000000
	movi.2d v4, #0000000000000000
	movi.2d v5, #0000000000000000
	movi.2d v6, #0000000000000000
	movi.2d v7, #0000000000000000
LBB26_6:
	mov x11, x5
	ldr q16, [x5]
	dup.8h v17, v16[0]
	ldr q18, [x4]
	; InlineAsm Start
	fmla.8h v7, v18, v17
	; InlineAsm End
	dup.8h v17, v16[1]
	; InlineAsm Start
	fmla.8h v6, v18, v17
	; InlineAsm End
	dup.8h v17, v16[2]
	; InlineAsm Start
	fmla.8h v5, v18, v17
	; InlineAsm End
	dup.8h v16, v16[3]
	; InlineAsm Start
	fmla.8h v4, v18, v16
	; InlineAsm End
	ldur q16, [x5, #8]
	dup.8h v17, v16[0]
	; InlineAsm Start
	fmla.8h v3, v18, v17
	; InlineAsm End
	dup.8h v17, v16[1]
	; InlineAsm Start
	fmla.8h v2, v18, v17
	; InlineAsm End
	dup.8h v17, v16[2]
	; InlineAsm Start
	fmla.8h v1, v18, v17
	; InlineAsm End
	dup.8h v16, v16[3]
	; InlineAsm Start
	fmla.8h v0, v18, v16
	; InlineAsm End
	add x12, x5, x10
	add x5, x12, x10
	ldr q16, [x12]
	dup.8h v17, v16[0]
	ldr q18, [x4, x9]
	; InlineAsm Start
	fmla.8h v7, v18, v17
	; InlineAsm End
	dup.8h v17, v16[1]
	; InlineAsm Start
	fmla.8h v6, v18, v17
	; InlineAsm End
	dup.8h v17, v16[2]
	; InlineAsm Start
	fmla.8h v5, v18, v17
	; InlineAsm End
	dup.8h v16, v16[3]
	; InlineAsm Start
	fmla.8h v4, v18, v16
	; InlineAsm End
	ldur q16, [x12, #8]
	dup.8h v17, v16[0]
	; InlineAsm Start
	fmla.8h v3, v18, v17
	; InlineAsm End
	dup.8h v17, v16[1]
	; InlineAsm Start
	fmla.8h v2, v18, v17
	; InlineAsm End
	dup.8h v17, v16[2]
	; InlineAsm Start
	fmla.8h v1, v18, v17
	; InlineAsm End
	dup.8h v16, v16[3]
	; InlineAsm Start
	fmla.8h v0, v18, v16
	; InlineAsm End
	add x4, x4, x9, lsl #1
	subs x8, x8, #1
	b.ne LBB26_6
	add x5, x11, x10, lsl #1
	tbz w2, #0, LBB26_13

	ldr q16, [x5]
	dup.8h v17, v16[0]
	ldr q18, [x4]
	; InlineAsm Start
	fmla.8h v7, v18, v17
	; InlineAsm End
	dup.8h v17, v16[1]
	dup.8h v19, v16[2]
	; InlineAsm Start
	fmla.8h v6, v18, v17
	; InlineAsm End
	; InlineAsm Start
	fmla.8h v5, v18, v19
	; InlineAsm End
	dup.8h v16, v16[3]
	; InlineAsm Start
	fmla.8h v4, v18, v16
	; InlineAsm End
	ldur q16, [x5, #8]
	dup.8h v17, v16[0]
	; InlineAsm Start
	fmla.8h v3, v18, v17
	; InlineAsm End
	dup.8h v17, v16[1]
	; InlineAsm Start
	fmla.8h v2, v18, v17
	; InlineAsm End
	dup.8h v17, v16[2]
	; InlineAsm Start
	fmla.8h v1, v18, v17
	; InlineAsm End
	dup.8h v16, v16[3]
	; InlineAsm Start
	fmla.8h v0, v18, v16
	; InlineAsm End
	b LBB26_13
LBB26_9:
	lsl x11, x11, #1
	lsl x13, x12, #1
	mov w12, #14
	movi.2d v1, #0000000000000000
	msub x12, x9, x12, x13
	lsl x13, x13, #1
	movi.2d v2, #0000000000000000
	movi.2d v3, #0000000000000000
	movi.2d v4, #0000000000000000
	movi.2d v5, #0000000000000000
	movi.2d v6, #0000000000000000
	movi.2d v7, #0000000000000000
LBB26_10:
	mov x14, x5
	ld1r.8h { v16 }, [x14], x13
	ldr q17, [x4]
	; InlineAsm Start
	fmla.8h v7, v17, v16
	; InlineAsm End
	add x5, x5, x10
	ld1r.8h { v16 }, [x5], x10
	; InlineAsm Start
	fmla.8h v6, v17, v16
	; InlineAsm End
	ld1r.8h { v16 }, [x5], x10
	; InlineAsm Start
	fmla.8h v5, v17, v16
	; InlineAsm End
	ld1r.8h { v16 }, [x5], x10
	; InlineAsm Start
	fmla.8h v4, v17, v16
	; InlineAsm End
	ld1r.8h { v16 }, [x5], x10
	; InlineAsm Start
	fmla.8h v3, v17, v16
	; InlineAsm End
	ld1r.8h { v16 }, [x5], x10
	; InlineAsm Start
	fmla.8h v2, v17, v16
	; InlineAsm End
	ld1r.8h { v16 }, [x5], x10
	; InlineAsm Start
	fmla.8h v1, v17, v16
	; InlineAsm End
	ld1r.8h { v16 }, [x5], x12
	; InlineAsm Start
	fmla.8h v0, v17, v16
	; InlineAsm End
	ld1r.8h { v16 }, [x5], x10
	ldr q17, [x4, x11]
	; InlineAsm Start
	fmla.8h v7, v17, v16
	; InlineAsm End
	ld1r.8h { v16 }, [x5], x10
	; InlineAsm Start
	fmla.8h v6, v17, v16
	; InlineAsm End
	ld1r.8h { v16 }, [x5], x10
	; InlineAsm Start
	fmla.8h v5, v17, v16
	; InlineAsm End
	ld1r.8h { v16 }, [x5], x10
	; InlineAsm Start
	fmla.8h v4, v17, v16
	; InlineAsm End
	ld1r.8h { v16 }, [x5], x10
	; InlineAsm Start
	fmla.8h v3, v17, v16
	; InlineAsm End
	ld1r.8h { v16 }, [x5], x10
	; InlineAsm Start
	fmla.8h v2, v17, v16
	; InlineAsm End
	ld1r.8h { v16 }, [x5], x10
	; InlineAsm Start
	fmla.8h v1, v17, v16
	; InlineAsm End
	ld1r.8h { v16 }, [x5], x12
	; InlineAsm Start
	fmla.8h v0, v17, v16
	; InlineAsm End
	add x4, x4, x11, lsl #1
	subs x8, x8, #1
	b.ne LBB26_10
	mov x5, x14
	tbz w2, #0, LBB26_13

	lsl x8, x9, #4
	sub x8, x8, x10
	mov x11, x5
	ld1r.8h { v16 }, [x11], x8
	ldr q17, [x4]
	; InlineAsm Start
	fmla.8h v7, v17, v16
	; InlineAsm End
	add x8, x5, x10
	ld1r.8h { v16 }, [x8]
	; InlineAsm Start
	fmla.8h v6, v17, v16
	; InlineAsm End
	add x8, x5, x9, lsl #2
	ld1r.8h { v16 }, [x8]
	; InlineAsm Start
	fmla.8h v5, v17, v16
	; InlineAsm End
	mov w8, #6
	madd x8, x9, x8, x5
	ld1r.8h { v16 }, [x8]
	; InlineAsm Start
	fmla.8h v4, v17, v16
	; InlineAsm End
	add x8, x5, x9, lsl #3
	ld1r.8h { v16 }, [x8]
	; InlineAsm Start
	fmla.8h v3, v17, v16
	; InlineAsm End
	mov w8, #10
	madd x8, x9, x8, x5
	ld1r.8h { v16 }, [x8]
	; InlineAsm Start
	fmla.8h v2, v17, v16
	; InlineAsm End
	mov w8, #12
	madd x8, x9, x8, x5
	ld1r.8h { v16 }, [x8]
	; InlineAsm Start
	fmla.8h v1, v17, v16
	; InlineAsm End
	ld1r.8h { v16 }, [x11]
	; InlineAsm Start
	fmla.8h v0, v17, v16
	; InlineAsm End
LBB26_13:
	ldrh w8, [sp, #154]
	ldrh w9, [sp, #152]
	ldrb w10, [sp, #156]
	cmp x0, #8
	b.ne LBB26_19
	cmp x1, #8
	b.ne LBB26_19
	cmp x7, #1
	b.ne LBB26_19
	dup.8h v16, w8
	and w8, w10, #0xff
	cmp w8, #1
	b.eq LBB26_27
	cmp w8, #2
	b.ne LBB26_28
	dup.8h v17, w9
	ldr q18, [x3]
	; InlineAsm Start
	fmul.8h v19, v17, v18
	; InlineAsm End
	; InlineAsm Start
	fmul.8h v18, v16, v7
	; InlineAsm End
	; InlineAsm Start
	fadd.8h v7, v19, v18
	; InlineAsm End
	str q7, [x3]
	lsl x8, x6, #1
	ldr q7, [x3, x8]
	; InlineAsm Start
	fmul.8h v18, v17, v7
	; InlineAsm End
	; InlineAsm Start
	fmul.8h v7, v16, v6
	; InlineAsm End
	; InlineAsm Start
	fadd.8h v6, v18, v7
	; InlineAsm End
	str q6, [x3, x8]
	lsl x8, x6, #2
	ldr q6, [x3, x8]
	; InlineAsm Start
	fmul.8h v7, v17, v6
	; InlineAsm End
	; InlineAsm Start
	fmul.8h v6, v16, v5
	; InlineAsm End
	; InlineAsm Start
	fadd.8h v5, v7, v6
	; InlineAsm End
	str q5, [x3, x8]
	mov w8, #6
	mul x8, x6, x8
	ldr q5, [x3, x8]
	; InlineAsm Start
	fmul.8h v6, v17, v5
	; InlineAsm End
	; InlineAsm Start
	fmul.8h v5, v16, v4
	; InlineAsm End
	; InlineAsm Start
	fadd.8h v4, v6, v5
	; InlineAsm End
	str q4, [x3, x8]
	lsl x8, x6, #3
	ldr q4, [x3, x8]
	; InlineAsm Start
	fmul.8h v5, v17, v4
	; InlineAsm End
	; InlineAsm Start
	fmul.8h v4, v16, v3
	; InlineAsm End
	; InlineAsm Start
	fadd.8h v3, v5, v4
	; InlineAsm End
	str q3, [x3, x8]
	mov w8, #10
	mul x8, x6, x8
	ldr q3, [x3, x8]
	; InlineAsm Start
	fmul.8h v4, v17, v3
	; InlineAsm End
	; InlineAsm Start
	fmul.8h v3, v16, v2
	; InlineAsm End
	; InlineAsm Start
	fadd.8h v2, v4, v3
	; InlineAsm End
	str q2, [x3, x8]
	mov w8, #12
	mul x8, x6, x8
	ldr q2, [x3, x8]
	; InlineAsm Start
	fmul.8h v3, v17, v2
	; InlineAsm End
	; InlineAsm Start
	fmul.8h v2, v16, v1
	; InlineAsm End
	; InlineAsm Start
	fadd.8h v1, v3, v2
	; InlineAsm End
	str q1, [x3, x8]
	mov w8, #14
	mul x8, x6, x8
	ldr q1, [x3, x8]
	; InlineAsm Start
	fmul.8h v2, v17, v1
	; InlineAsm End
	; InlineAsm Start
	fmul.8h v1, v16, v0
	; InlineAsm End
	; InlineAsm Start
	fadd.8h v0, v2, v1
	; InlineAsm End
	str q0, [x3, x8]
	b LBB26_42
LBB26_19:
	stp q7, q6, [sp]
	stp q5, q4, [sp, #32]
	stp q3, q2, [sp, #64]
	and w10, w10, #0xff
	stp q1, q0, [sp, #96]
	cmp w10, #2
	b.eq LBB26_30
	cmp w10, #1
	b.ne LBB26_36
	cbz x1, LBB26_42
	cbz x0, LBB26_42
	mov x9, #0
	lsl x10, x6, #1
	lsl x11, x7, #1
	mov x12, sp
LBB26_24:
	mov x13, x0
	mov x14, x3
	mov x15, x12
LBB26_25:
	ldr h0, [x15], #2
	fmov s1, w8
	; InlineAsm Start
	fmul h2, h1, h0
	; InlineAsm End
	ldr h0, [x14]
	; InlineAsm Start
	fadd h1, h0, h2
	; InlineAsm End
	str h1, [x14]
	add x14, x14, x11
	subs x13, x13, #1
	b.ne LBB26_25
	add x9, x9, #1
	add x12, x12, #16
	add x3, x3, x10
	cmp x9, x1
	b.ne LBB26_24
	b LBB26_42
LBB26_27:
	ldr q17, [x3]
	; InlineAsm Start
	fmla.8h v17, v16, v7
	; InlineAsm End
	str q17, [x3]
	lsl x8, x6, #1
	ldr q7, [x3, x8]
	; InlineAsm Start
	fmla.8h v7, v16, v6
	; InlineAsm End
	str q7, [x3, x8]
	lsl x8, x6, #2
	ldr q6, [x3, x8]
	; InlineAsm Start
	fmla.8h v6, v16, v5
	; InlineAsm End
	str q6, [x3, x8]
	mov w8, #6
	mul x8, x6, x8
	ldr q5, [x3, x8]
	; InlineAsm Start
	fmla.8h v5, v16, v4
	; InlineAsm End
	str q5, [x3, x8]
	lsl x8, x6, #3
	ldr q4, [x3, x8]
	; InlineAsm Start
	fmla.8h v4, v16, v3
	; InlineAsm End
	str q4, [x3, x8]
	mov w8, #10
	mul x8, x6, x8
	ldr q3, [x3, x8]
	; InlineAsm Start
	fmla.8h v3, v16, v2
	; InlineAsm End
	str q3, [x3, x8]
	mov w8, #12
	mul x8, x6, x8
	ldr q2, [x3, x8]
	; InlineAsm Start
	fmla.8h v2, v16, v1
	; InlineAsm End
	str q2, [x3, x8]
	mov w8, #14
	mul x8, x6, x8
	ldr q1, [x3, x8]
	; InlineAsm Start
	fmla.8h v1, v16, v0
	; InlineAsm End
	b LBB26_29
LBB26_28:
	; InlineAsm Start
	fmul.8h v17, v16, v7
	; InlineAsm End
	str q17, [x3]
	lsl x8, x6, #1
	; InlineAsm Start
	fmul.8h v7, v16, v6
	; InlineAsm End
	str q7, [x3, x8]
	lsl x8, x6, #2
	; InlineAsm Start
	fmul.8h v6, v16, v5
	; InlineAsm End
	str q6, [x3, x8]
	mov w8, #6
	mul x8, x6, x8
	; InlineAsm Start
	fmul.8h v5, v16, v4
	; InlineAsm End
	str q5, [x3, x8]
	lsl x8, x6, #3
	; InlineAsm Start
	fmul.8h v4, v16, v3
	; InlineAsm End
	str q4, [x3, x8]
	mov w8, #10
	mul x8, x6, x8
	; InlineAsm Start
	fmul.8h v3, v16, v2
	; InlineAsm End
	str q3, [x3, x8]
	mov w8, #12
	mul x8, x6, x8
	; InlineAsm Start
	fmul.8h v2, v16, v1
	; InlineAsm End
	str q2, [x3, x8]
	mov w8, #14
	mul x8, x6, x8
	; InlineAsm Start
	fmul.8h v1, v16, v0
	; InlineAsm End
LBB26_29:
	str q1, [x3, x8]
	b LBB26_42
LBB26_30:
	cbz x1, LBB26_42
	cbz x0, LBB26_42
	mov x10, #0
	lsl x11, x6, #1
	lsl x12, x7, #1
	mov x13, sp
LBB26_33:
	mov x14, x0
	mov x15, x3
	mov x16, x13
LBB26_34:
	ldr h0, [x15]
	fmov s1, w9
	; InlineAsm Start
	fmul h2, h1, h0
	; InlineAsm End
	ldr h0, [x16], #2
	fmov s1, w8
	; InlineAsm Start
	fmul h3, h1, h0
	; InlineAsm End
	; InlineAsm Start
	fadd h0, h2, h3
	; InlineAsm End
	str h0, [x15]
	add x15, x15, x12
	subs x14, x14, #1
	b.ne LBB26_34
	add x10, x10, #1
	add x13, x13, #16
	add x3, x3, x11
	cmp x10, x1
	b.ne LBB26_33
	b LBB26_42
LBB26_36:
	cbz x1, LBB26_42
	cbz x0, LBB26_42
	mov x9, #0
	lsl x10, x6, #1
	lsl x11, x7, #1
	mov x12, sp
LBB26_39:
	mov x13, x0
	mov x14, x3
	mov x15, x12
LBB26_40:
	ldr h0, [x15], #2
	fmov s1, w8
	; InlineAsm Start
	fmul h2, h1, h0
	; InlineAsm End
	str h2, [x14]
	add x14, x14, x11
	subs x13, x13, #1
	b.ne LBB26_40
	add x9, x9, #1
	add x12, x12, #16
	add x3, x3, x10
	cmp x9, x1
	b.ne LBB26_39
LBB26_42:
	add sp, sp, #128
	.cfi_def_cfa_offset 0
	ret
Lfunc_end26:
	.cfi_endproc

	.p2align	2
gemm_f16::microkernel::neon::f16::x2x1:
Lfunc_begin27:
	.cfi_startproc
	sub sp, sp, #32
	.cfi_def_cfa_offset 32
	movi.2d v0, #0000000000000000
	cmp x2, #2
	b.hs LBB27_2
	movi.2d v1, #0000000000000000
	tbnz w2, #0, LBB27_5
	b LBB27_6
LBB27_2:
	ldp x9, x10, [sp, #32]
	lsr x8, x2, #1
	lsl x9, x9, #1
	lsl x10, x10, #1
	lsl x11, x10, #1
	movi.2d v1, #0000000000000000
	mov x12, x5
LBB27_3:
	mov x13, x4
	ld1r.8h { v2 }, [x12], x11
	ldp q3, q4, [x4]
	; InlineAsm Start
	fmla.8h v1, v3, v2
	; InlineAsm End
	; InlineAsm Start
	fmla.8h v0, v4, v2
	; InlineAsm End
	add x14, x5, x10
	ld1r.8h { v2 }, [x14]
	add x14, x4, x9
	add x4, x14, x9
	ldp q3, q4, [x14]
	; InlineAsm Start
	fmla.8h v1, v3, v2
	; InlineAsm End
	; InlineAsm Start
	fmla.8h v0, v4, v2
	; InlineAsm End
	mov x5, x12
	subs x8, x8, #1
	b.ne LBB27_3
	add x4, x13, x9, lsl #1
	tbz w2, #0, LBB27_6

	ld1r.8h { v2 }, [x5]
	ldp q3, q4, [x4]
	; InlineAsm Start
	fmla.8h v1, v3, v2
	; InlineAsm End
	; InlineAsm Start
	fmla.8h v0, v4, v2
	; InlineAsm End
LBB27_6:
	ldrh w8, [sp, #58]
	ldrh w9, [sp, #56]
	ldrb w10, [sp, #60]
	cmp x0, #16
	b.ne LBB27_12
	cmp x1, #1
	b.ne LBB27_12
	cmp x7, #1
	b.ne LBB27_12
	dup.8h v2, w8
	and w8, w10, #0xff
	cmp w8, #1
	b.eq LBB27_20
	cmp w8, #2
	b.ne LBB27_21
	dup.8h v3, w9
	ldp q4, q5, [x3]
	; InlineAsm Start
	fmul.8h v6, v3, v4
	; InlineAsm End
	; InlineAsm Start
	fmul.8h v4, v2, v1
	; InlineAsm End
	; InlineAsm Start
	fadd.8h v1, v6, v4
	; InlineAsm End
	; InlineAsm Start
	fmul.8h v4, v3, v5
	; InlineAsm End
	; InlineAsm Start
	fmul.8h v3, v2, v0
	; InlineAsm End
	; InlineAsm Start
	fadd.8h v0, v4, v3
	; InlineAsm End
	stp q1, q0, [x3]
	b LBB27_35
LBB27_12:
	stp q1, q0, [sp]
	and w10, w10, #0xff
	cmp w10, #2
	b.eq LBB27_23
	cmp w10, #1
	b.ne LBB27_29
	cbz x1, LBB27_35
	cbz x0, LBB27_35
	mov x9, #0
	lsl x10, x6, #1
	lsl x11, x7, #1
	mov x12, sp
LBB27_17:
	mov x13, x0
	mov x14, x3
	mov x15, x12
LBB27_18:
	ldr h0, [x15], #2
	fmov s1, w8
	; InlineAsm Start
	fmul h2, h1, h0
	; InlineAsm End
	ldr h0, [x14]
	; InlineAsm Start
	fadd h1, h0, h2
	; InlineAsm End
	str h1, [x14]
	add x14, x14, x11
	subs x13, x13, #1
	b.ne LBB27_18
	add x9, x9, #1
	add x12, x12, #32
	add x3, x3, x10
	cmp x9, x1
	b.ne LBB27_17
	b LBB27_35
LBB27_20:
	ldr q3, [x3]
	; InlineAsm Start
	fmla.8h v3, v2, v1
	; InlineAsm End
	str q3, [x3]
	ldr q1, [x3, #16]
	; InlineAsm Start
	fmla.8h v1, v2, v0
	; InlineAsm End
	b LBB27_22
LBB27_21:
	; InlineAsm Start
	fmul.8h v3, v2, v1
	; InlineAsm End
	; InlineAsm Start
	fmul.8h v1, v2, v0
	; InlineAsm End
	str q3, [x3]
LBB27_22:
	str q1, [x3, #16]
	b LBB27_35
LBB27_23:
	cbz x1, LBB27_35
	cbz x0, LBB27_35
	mov x10, #0
	lsl x11, x6, #1
	lsl x12, x7, #1
	mov x13, sp
LBB27_26:
	mov x14, x0
	mov x15, x3
	mov x16, x13
LBB27_27:
	ldr h0, [x15]
	fmov s1, w9
	; InlineAsm Start
	fmul h2, h1, h0
	; InlineAsm End
	ldr h0, [x16], #2
	fmov s1, w8
	; InlineAsm Start
	fmul h3, h1, h0
	; InlineAsm End
	; InlineAsm Start
	fadd h0, h2, h3
	; InlineAsm End
	str h0, [x15]
	add x15, x15, x12
	subs x14, x14, #1
	b.ne LBB27_27
	add x10, x10, #1
	add x13, x13, #32
	add x3, x3, x11
	cmp x10, x1
	b.ne LBB27_26
	b LBB27_35
LBB27_29:
	cbz x1, LBB27_35
	cbz x0, LBB27_35
	mov x9, #0
	lsl x10, x6, #1
	lsl x11, x7, #1
	mov x12, sp
LBB27_32:
	mov x13, x0
	mov x14, x3
	mov x15, x12
LBB27_33:
	ldr h0, [x15], #2
	fmov s1, w8
	; InlineAsm Start
	fmul h2, h1, h0
	; InlineAsm End
	str h2, [x14]
	add x14, x14, x11
	subs x13, x13, #1
	b.ne LBB27_33
	add x9, x9, #1
	add x12, x12, #32
	add x3, x3, x10
	cmp x9, x1
	b.ne LBB27_32
LBB27_35:
	add sp, sp, #32
	.cfi_def_cfa_offset 0
	ret
Lfunc_end27:
	.cfi_endproc

	.p2align	2
gemm_f16::microkernel::neon::f16::x2x2:
Lfunc_begin28:
	.cfi_startproc
	sub sp, sp, #64
	.cfi_def_cfa_offset 64
	ldr x8, [sp, #80]
	movi.2d v0, #0000000000000000
	lsl x8, x8, #1
	cmp x2, #2
	b.hs LBB28_2
	movi.2d v1, #0000000000000000
	movi.2d v2, #0000000000000000
	movi.2d v3, #0000000000000000
	tbnz w2, #0, LBB28_5
	b LBB28_6
LBB28_2:
	ldp x10, x11, [sp, #64]
	lsr x9, x2, #1
	lsl x10, x10, #1
	lsl x11, x11, #1
	add x12, x8, x11
	lsl x13, x11, #1
	movi.2d v1, #0000000000000000
	movi.2d v2, #0000000000000000
	movi.2d v3, #0000000000000000
	mov x14, x5
LBB28_3:
	mov x15, x4
	ld1r.8h { v4 }, [x14], x13
	ldp q5, q6, [x4]
	; InlineAsm Start
	fmla.8h v3, v5, v4
	; InlineAsm End
	; InlineAsm Start
	fmla.8h v2, v6, v4
	; InlineAsm End
	add x16, x5, x8
	ld1r.8h { v4 }, [x16]
	; InlineAsm Start
	fmla.8h v1, v5, v4
	; InlineAsm End
	; InlineAsm Start
	fmla.8h v0, v6, v4
	; InlineAsm End
	add x16, x5, x11
	ld1r.8h { v4 }, [x16]
	add x16, x4, x10
	add x4, x16, x10
	ldp q5, q6, [x16]
	; InlineAsm Start
	fmla.8h v3, v5, v4
	; InlineAsm End
	; InlineAsm Start
	fmla.8h v2, v6, v4
	; InlineAsm End
	add x16, x5, x12
	ld1r.8h { v4 }, [x16]
	; InlineAsm Start
	fmla.8h v1, v5, v4
	; InlineAsm End
	; InlineAsm Start
	fmla.8h v0, v6, v4
	; InlineAsm End
	mov x5, x14
	subs x9, x9, #1
	b.ne LBB28_3
	add x4, x15, x10, lsl #1
	tbz w2, #0, LBB28_6

	ld1r.8h { v4 }, [x5], x8
	ldp q5, q6, [x4]
	; InlineAsm Start
	fmla.8h v3, v5, v4
	; InlineAsm End
	; InlineAsm Start
	fmla.8h v2, v6, v4
	; InlineAsm End
	ld1r.8h { v4 }, [x5]
	; InlineAsm Start
	fmla.8h v1, v5, v4
	; InlineAsm End
	; InlineAsm Start
	fmla.8h v0, v6, v4
	; InlineAsm End
LBB28_6:
	ldrh w8, [sp, #90]
	ldrh w9, [sp, #88]
	ldrb w10, [sp, #92]
	cmp x0, #16
	b.ne LBB28_12
	cmp x1, #2
	b.ne LBB28_12
	cmp x7, #1
	b.ne LBB28_12
	dup.8h v4, w8
	and w8, w10, #0xff
	cmp w8, #1
	b.eq LBB28_20
	cmp w8, #2
	b.ne LBB28_21
	dup.8h v5, w9
	ldp q6, q7, [x3]
	; InlineAsm Start
	fmul.8h v16, v5, v6
	; InlineAsm End
	; InlineAsm Start
	fmul.8h v6, v4, v3
	; InlineAsm End
	; InlineAsm Start
	fadd.8h v3, v16, v6
	; InlineAsm End
	; InlineAsm Start
	fmul.8h v6, v5, v7
	; InlineAsm End
	; InlineAsm Start
	fmul.8h v7, v4, v2
	; InlineAsm End
	; InlineAsm Start
	fadd.8h v2, v6, v7
	; InlineAsm End
	stp q3, q2, [x3]
	add x8, x3, x6, lsl #1
	ldp q2, q3, [x8]
	; InlineAsm Start
	fmul.8h v6, v5, v2
	; InlineAsm End
	; InlineAsm Start
	fmul.8h v2, v4, v1
	; InlineAsm End
	; InlineAsm Start
	fadd.8h v1, v6, v2
	; InlineAsm End
	; InlineAsm Start
	fmul.8h v2, v5, v3
	; InlineAsm End
	; InlineAsm Start
	fmul.8h v3, v4, v0
	; InlineAsm End
	; InlineAsm Start
	fadd.8h v0, v2, v3
	; InlineAsm End
	stp q1, q0, [x8]
	b LBB28_35
LBB28_12:
	stp q3, q2, [sp]
	stp q1, q0, [sp, #32]
	and w10, w10, #0xff
	cmp w10, #2
	b.eq LBB28_23
	cmp w10, #1
	b.ne LBB28_29
	cbz x1, LBB28_35
	cbz x0, LBB28_35
	mov x9, #0
	lsl x10, x6, #1
	lsl x11, x7, #1
	mov x12, sp
LBB28_17:
	mov x13, x0
	mov x14, x3
	mov x15, x12
LBB28_18:
	ldr h0, [x15], #2
	fmov s1, w8
	; InlineAsm Start
	fmul h2, h1, h0
	; InlineAsm End
	ldr h0, [x14]
	; InlineAsm Start
	fadd h1, h0, h2
	; InlineAsm End
	str h1, [x14]
	add x14, x14, x11
	subs x13, x13, #1
	b.ne LBB28_18
	add x9, x9, #1
	add x12, x12, #32
	add x3, x3, x10
	cmp x9, x1
	b.ne LBB28_17
	b LBB28_35
LBB28_20:
	ldr q5, [x3]
	; InlineAsm Start
	fmla.8h v5, v4, v3
	; InlineAsm End
	ldr q3, [x3, #16]
	; InlineAsm Start
	fmla.8h v3, v4, v2
	; InlineAsm End
	stp q5, q3, [x3]
	add x8, x3, x6, lsl #1
	ldr q2, [x8]
	; InlineAsm Start
	fmla.8h v2, v4, v1
	; InlineAsm End
	str q2, [x8]
	ldr q1, [x8, #16]
	; InlineAsm Start
	fmla.8h v1, v4, v0
	; InlineAsm End
	b LBB28_22
LBB28_21:
	; InlineAsm Start
	fmul.8h v5, v4, v3
	; InlineAsm End
	; InlineAsm Start
	fmul.8h v3, v4, v2
	; InlineAsm End
	stp q5, q3, [x3]
	add x8, x3, x6, lsl #1
	; InlineAsm Start
	fmul.8h v2, v4, v1
	; InlineAsm End
	; InlineAsm Start
	fmul.8h v1, v4, v0
	; InlineAsm End
	str q2, [x8]
LBB28_22:
	str q1, [x8, #16]
	b LBB28_35
LBB28_23:
	cbz x1, LBB28_35
	cbz x0, LBB28_35
	mov x10, #0
	lsl x11, x6, #1
	lsl x12, x7, #1
	mov x13, sp
LBB28_26:
	mov x14, x0
	mov x15, x3
	mov x16, x13
LBB28_27:
	ldr h0, [x15]
	fmov s1, w9
	; InlineAsm Start
	fmul h2, h1, h0
	; InlineAsm End
	ldr h0, [x16], #2
	fmov s1, w8
	; InlineAsm Start
	fmul h3, h1, h0
	; InlineAsm End
	; InlineAsm Start
	fadd h0, h2, h3
	; InlineAsm End
	str h0, [x15]
	add x15, x15, x12
	subs x14, x14, #1
	b.ne LBB28_27
	add x10, x10, #1
	add x13, x13, #32
	add x3, x3, x11
	cmp x10, x1
	b.ne LBB28_26
	b LBB28_35
LBB28_29:
	cbz x1, LBB28_35
	cbz x0, LBB28_35
	mov x9, #0
	lsl x10, x6, #1
	lsl x11, x7, #1
	mov x12, sp
LBB28_32:
	mov x13, x0
	mov x14, x3
	mov x15, x12
LBB28_33:
	ldr h0, [x15], #2
	fmov s1, w8
	; InlineAsm Start
	fmul h2, h1, h0
	; InlineAsm End
	str h2, [x14]
	add x14, x14, x11
	subs x13, x13, #1
	b.ne LBB28_33
	add x9, x9, #1
	add x12, x12, #32
	add x3, x3, x10
	cmp x9, x1
	b.ne LBB28_32
LBB28_35:
	add sp, sp, #64
	.cfi_def_cfa_offset 0
	ret
Lfunc_end28:
	.cfi_endproc

	.p2align	2
gemm_f16::microkernel::neon::f16::x2x3:
Lfunc_begin29:
	.cfi_startproc
	sub sp, sp, #112
	.cfi_def_cfa_offset 112
	stp x20, x19, [sp, #96]
	.cfi_offset w19, -8
	.cfi_offset w20, -16
	ldr x8, [sp, #128]
	movi.2d v0, #0000000000000000
	lsl x9, x8, #2
	cmp x2, #2
	b.hs LBB29_2
	movi.2d v1, #0000000000000000
	movi.2d v2, #0000000000000000
	movi.2d v3, #0000000000000000
	movi.2d v4, #0000000000000000
	movi.2d v5, #0000000000000000
	tbnz w2, #0, LBB29_5
	b LBB29_6
LBB29_2:
	ldp x12, x13, [sp, #112]
	lsr x10, x2, #1
	lsl x11, x8, #1
	lsl x12, x12, #1
	lsl x13, x13, #1
	add x14, x9, x13
	add x15, x11, x13
	lsl x16, x13, #1
	movi.2d v1, #0000000000000000
	movi.2d v2, #0000000000000000
	movi.2d v3, #0000000000000000
	movi.2d v4, #0000000000000000
	movi.2d v5, #0000000000000000
	mov x17, x5
LBB29_3:
	mov x19, x4
	ld1r.8h { v6 }, [x17], x16
	ldp q7, q16, [x4]
	; InlineAsm Start
	fmla.8h v5, v7, v6
	; InlineAsm End
	; InlineAsm Start
	fmla.8h v4, v16, v6
	; InlineAsm End
	add x4, x5, x11
	ld1r.8h { v6 }, [x4]
	; InlineAsm Start
	fmla.8h v3, v7, v6
	; InlineAsm End
	; InlineAsm Start
	fmla.8h v2, v16, v6
	; InlineAsm End
	add x4, x5, x9
	ld1r.8h { v6 }, [x4]
	; InlineAsm Start
	fmla.8h v1, v7, v6
	; InlineAsm End
	; InlineAsm Start
	fmla.8h v0, v16, v6
	; InlineAsm End
	add x4, x5, x13
	ld1r.8h { v6 }, [x4]
	add x20, x19, x12
	add x4, x20, x12
	ldp q7, q16, [x20]
	; InlineAsm Start
	fmla.8h v5, v7, v6
	; InlineAsm End
	; InlineAsm Start
	fmla.8h v4, v16, v6
	; InlineAsm End
	add x20, x5, x15
	ld1r.8h { v6 }, [x20]
	; InlineAsm Start
	fmla.8h v3, v7, v6
	; InlineAsm End
	; InlineAsm Start
	fmla.8h v2, v16, v6
	; InlineAsm End
	add x5, x5, x14
	ld1r.8h { v6 }, [x5]
	; InlineAsm Start
	fmla.8h v1, v7, v6
	; InlineAsm End
	; InlineAsm Start
	fmla.8h v0, v16, v6
	; InlineAsm End
	mov x5, x17
	subs x10, x10, #1
	b.ne LBB29_3
	add x4, x19, x12, lsl #1
	tbz w2, #0, LBB29_6

	mov x10, x5
	ld1r.8h { v6 }, [x10], x9
	ldp q7, q16, [x4]
	; InlineAsm Start
	fmla.8h v5, v7, v6
	; InlineAsm End
	; InlineAsm Start
	fmla.8h v4, v16, v6
	; InlineAsm End
	add x8, x5, x8, lsl #1
	ld1r.8h { v6 }, [x8]
	; InlineAsm Start
	fmla.8h v3, v7, v6
	; InlineAsm End
	; InlineAsm Start
	fmla.8h v2, v16, v6
	; InlineAsm End
	ld1r.8h { v6 }, [x10]
	; InlineAsm Start
	fmla.8h v1, v7, v6
	; InlineAsm End
	; InlineAsm Start
	fmla.8h v0, v16, v6
	; InlineAsm End
LBB29_6:
	ldrh w8, [sp, #138]
	ldrh w9, [sp, #136]
	ldrb w10, [sp, #140]
	cmp x0, #16
	b.ne LBB29_12
	cmp x1, #3
	b.ne LBB29_12
	cmp x7, #1
	b.ne LBB29_12
	dup.8h v6, w8
	and w8, w10, #0xff
	cmp w8, #1
	b.eq LBB29_20
	cmp w8, #2
	b.ne LBB29_21
	dup.8h v7, w9
	ldp q16, q17, [x3]
	; InlineAsm Start
	fmul.8h v18, v7, v16
	; InlineAsm End
	; InlineAsm Start
	fmul.8h v16, v6, v5
	; InlineAsm End
	; InlineAsm Start
	fadd.8h v5, v18, v16
	; InlineAsm End
	; InlineAsm Start
	fmul.8h v16, v7, v17
	; InlineAsm End
	; InlineAsm Start
	fmul.8h v17, v6, v4
	; InlineAsm End
	; InlineAsm Start
	fadd.8h v4, v16, v17
	; InlineAsm End
	stp q5, q4, [x3]
	add x8, x3, x6, lsl #1
	ldp q4, q5, [x8]
	; InlineAsm Start
	fmul.8h v16, v7, v4
	; InlineAsm End
	; InlineAsm Start
	fmul.8h v4, v6, v3
	; InlineAsm End
	; InlineAsm Start
	fadd.8h v3, v16, v4
	; InlineAsm End
	; InlineAsm Start
	fmul.8h v4, v7, v5
	; InlineAsm End
	; InlineAsm Start
	fmul.8h v5, v6, v2
	; InlineAsm End
	; InlineAsm Start
	fadd.8h v2, v4, v5
	; InlineAsm End
	stp q3, q2, [x8]
	add x8, x3, x6, lsl #2
	ldp q2, q3, [x8]
	; InlineAsm Start
	fmul.8h v4, v7, v2
	; InlineAsm End
	; InlineAsm Start
	fmul.8h v2, v6, v1
	; InlineAsm End
	; InlineAsm Start
	fadd.8h v1, v4, v2
	; InlineAsm End
	; InlineAsm Start
	fmul.8h v2, v7, v3
	; InlineAsm End
	; InlineAsm Start
	fmul.8h v3, v6, v0
	; InlineAsm End
	; InlineAsm Start
	fadd.8h v0, v2, v3
	; InlineAsm End
	stp q1, q0, [x8]
	b LBB29_35
LBB29_12:
	stp q5, q4, [sp]
	stp q3, q2, [sp, #32]
	and w10, w10, #0xff
	stp q1, q0, [sp, #64]
	cmp w10, #2
	b.eq LBB29_23
	cmp w10, #1
	b.ne LBB29_29
	cbz x1, LBB29_35
	cbz x0, LBB29_35
	mov x9, #0
	lsl x10, x6, #1
	lsl x11, x7, #1
	mov x12, sp
LBB29_17:
	mov x13, x0
	mov x14, x3
	mov x15, x12
LBB29_18:
	ldr h0, [x15], #2
	fmov s1, w8
	; InlineAsm Start
	fmul h2, h1, h0
	; InlineAsm End
	ldr h0, [x14]
	; InlineAsm Start
	fadd h1, h0, h2
	; InlineAsm End
	str h1, [x14]
	add x14, x14, x11
	subs x13, x13, #1
	b.ne LBB29_18
	add x9, x9, #1
	add x12, x12, #32
	add x3, x3, x10
	cmp x9, x1
	b.ne LBB29_17
	b LBB29_35
LBB29_20:
	ldr q7, [x3]
	; InlineAsm Start
	fmla.8h v7, v6, v5
	; InlineAsm End
	ldr q5, [x3, #16]
	; InlineAsm Start
	fmla.8h v5, v6, v4
	; InlineAsm End
	stp q7, q5, [x3]
	add x8, x3, x6, lsl #1
	ldr q4, [x8]
	; InlineAsm Start
	fmla.8h v4, v6, v3
	; InlineAsm End
	ldr q3, [x8, #16]
	; InlineAsm Start
	fmla.8h v3, v6, v2
	; InlineAsm End
	stp q4, q3, [x8]
	add x8, x3, x6, lsl #2
	ldr q2, [x8]
	; InlineAsm Start
	fmla.8h v2, v6, v1
	; InlineAsm End
	str q2, [x8]
	ldr q1, [x8, #16]
	; InlineAsm Start
	fmla.8h v1, v6, v0
	; InlineAsm End
	b LBB29_22
LBB29_21:
	; InlineAsm Start
	fmul.8h v7, v6, v5
	; InlineAsm End
	; InlineAsm Start
	fmul.8h v5, v6, v4
	; InlineAsm End
	stp q7, q5, [x3]
	add x8, x3, x6, lsl #1
	; InlineAsm Start
	fmul.8h v4, v6, v3
	; InlineAsm End
	; InlineAsm Start
	fmul.8h v3, v6, v2
	; InlineAsm End
	stp q4, q3, [x8]
	add x8, x3, x6, lsl #2
	; InlineAsm Start
	fmul.8h v2, v6, v1
	; InlineAsm End
	; InlineAsm Start
	fmul.8h v1, v6, v0
	; InlineAsm End
	str q2, [x8]
LBB29_22:
	str q1, [x8, #16]
	b LBB29_35
LBB29_23:
	cbz x1, LBB29_35
	cbz x0, LBB29_35
	mov x10, #0
	lsl x11, x6, #1
	lsl x12, x7, #1
	mov x13, sp
LBB29_26:
	mov x14, x0
	mov x15, x3
	mov x16, x13
LBB29_27:
	ldr h0, [x15]
	fmov s1, w9
	; InlineAsm Start
	fmul h2, h1, h0
	; InlineAsm End
	ldr h0, [x16], #2
	fmov s1, w8
	; InlineAsm Start
	fmul h3, h1, h0
	; InlineAsm End
	; InlineAsm Start
	fadd h0, h2, h3
	; InlineAsm End
	str h0, [x15]
	add x15, x15, x12
	subs x14, x14, #1
	b.ne LBB29_27
	add x10, x10, #1
	add x13, x13, #32
	add x3, x3, x11
	cmp x10, x1
	b.ne LBB29_26
	b LBB29_35
LBB29_29:
	cbz x1, LBB29_35
	cbz x0, LBB29_35
	mov x9, #0
	lsl x10, x6, #1
	lsl x11, x7, #1
	mov x12, sp
LBB29_32:
	mov x13, x0
	mov x14, x3
	mov x15, x12
LBB29_33:
	ldr h0, [x15], #2
	fmov s1, w8
	; InlineAsm Start
	fmul h2, h1, h0
	; InlineAsm End
	str h2, [x14]
	add x14, x14, x11
	subs x13, x13, #1
	b.ne LBB29_33
	add x9, x9, #1
	add x12, x12, #32
	add x3, x3, x10
	cmp x9, x1
	b.ne LBB29_32
LBB29_35:
	ldp x20, x19, [sp, #96]
	add sp, sp, #112
	.cfi_def_cfa_offset 0
	.cfi_restore w19
	.cfi_restore w20
	ret
Lfunc_end29:
	.cfi_endproc

	.p2align	2
gemm_f16::microkernel::neon::f16::x2x4:
Lfunc_begin30:
	.cfi_startproc
	sub sp, sp, #160
	.cfi_def_cfa_offset 160
	stp x22, x21, [sp, #128]
	stp x20, x19, [sp, #144]
	.cfi_offset w19, -8
	.cfi_offset w20, -16
	.cfi_offset w21, -24
	.cfi_offset w22, -32
	ldp x12, x9, [sp, #168]
	ldr x11, [sp, #160]
	lsr x8, x2, #1
	movi.2d v0, #0000000000000000
	cmp x9, #1
	b.ne LBB30_3
	cmp x2, #2
	b.hs LBB30_5
	movi.2d v1, #0000000000000000
	movi.2d v2, #0000000000000000
	movi.2d v3, #0000000000000000
	movi.2d v4, #0000000000000000
	movi.2d v5, #0000000000000000
	movi.2d v6, #0000000000000000
	movi.2d v7, #0000000000000000
	tbnz w2, #0, LBB30_8
	b LBB30_13
LBB30_3:
	lsl x10, x9, #1
	cmp x2, #2
	b.hs LBB30_9
	movi.2d v1, #0000000000000000
	movi.2d v2, #0000000000000000
	movi.2d v3, #0000000000000000
	movi.2d v4, #0000000000000000
	movi.2d v5, #0000000000000000
	movi.2d v6, #0000000000000000
	movi.2d v7, #0000000000000000
	tbnz w2, #0, LBB30_12
	b LBB30_13
LBB30_5:
	lsl x9, x11, #1
	movi.2d v1, #0000000000000000
	lsl x10, x12, #1
	movi.2d v2, #0000000000000000
	movi.2d v3, #0000000000000000
	movi.2d v4, #0000000000000000
	movi.2d v5, #0000000000000000
	movi.2d v6, #0000000000000000
	movi.2d v7, #0000000000000000
LBB30_6:
	mov x11, x4
	ldr q16, [x5]
	dup.8h v17, v16[0]
	ldp q18, q19, [x4]
	; InlineAsm Start
	fmla.8h v7, v18, v17
	; InlineAsm End
	; InlineAsm Start
	fmla.8h v6, v19, v17
	; InlineAsm End
	dup.8h v17, v16[1]
	; InlineAsm Start
	fmla.8h v5, v18, v17
	; InlineAsm End
	; InlineAsm Start
	fmla.8h v4, v19, v17
	; InlineAsm End
	dup.8h v17, v16[2]
	; InlineAsm Start
	fmla.8h v3, v18, v17
	; InlineAsm End
	; InlineAsm Start
	fmla.8h v2, v19, v17
	; InlineAsm End
	dup.8h v16, v16[3]
	; InlineAsm Start
	fmla.8h v1, v18, v16
	; InlineAsm End
	; InlineAsm Start
	fmla.8h v0, v19, v16
	; InlineAsm End
	add x12, x4, x9
	add x4, x12, x9
	ldp q16, q17, [x12]
	ldr q18, [x5, x10]
	dup.8h v19, v18[0]
	; InlineAsm Start
	fmla.8h v7, v16, v19
	; InlineAsm End
	; InlineAsm Start
	fmla.8h v6, v17, v19
	; InlineAsm End
	dup.8h v19, v18[1]
	; InlineAsm Start
	fmla.8h v5, v16, v19
	; InlineAsm End
	; InlineAsm Start
	fmla.8h v4, v17, v19
	; InlineAsm End
	dup.8h v19, v18[2]
	; InlineAsm Start
	fmla.8h v3, v16, v19
	; InlineAsm End
	; InlineAsm Start
	fmla.8h v2, v17, v19
	; InlineAsm End
	dup.8h v18, v18[3]
	; InlineAsm Start
	fmla.8h v1, v16, v18
	; InlineAsm End
	; InlineAsm Start
	fmla.8h v0, v17, v18
	; InlineAsm End
	add x5, x5, x10, lsl #1
	subs x8, x8, #1
	b.ne LBB30_6
	add x4, x11, x9, lsl #1
	tbz w2, #0, LBB30_13

	ldr q16, [x5]
	dup.8h v17, v16[0]
	ldp q18, q19, [x4]
	; InlineAsm Start
	fmla.8h v7, v18, v17
	; InlineAsm End
	; InlineAsm Start
	fmla.8h v6, v19, v17
	; InlineAsm End
	dup.8h v17, v16[1]
	; InlineAsm Start
	fmla.8h v5, v18, v17
	; InlineAsm End
	; InlineAsm Start
	fmla.8h v4, v19, v17
	; InlineAsm End
	dup.8h v17, v16[2]
	; InlineAsm Start
	fmla.8h v3, v18, v17
	; InlineAsm End
	; InlineAsm Start
	fmla.8h v2, v19, v17
	; InlineAsm End
	dup.8h v16, v16[3]
	; InlineAsm Start
	fmla.8h v1, v18, v16
	; InlineAsm End
	; InlineAsm Start
	fmla.8h v0, v19, v16
	; InlineAsm End
	b LBB30_13
LBB30_9:
	lsl x11, x11, #1
	lsl x12, x12, #1
	add x13, x10, x9
	lsl x13, x13, #1
	movi.2d v1, #0000000000000000
	add x14, x13, x12
	lsl x15, x9, #2
	add x16, x15, x12
	add x17, x10, x12
	lsl x19, x12, #1
	movi.2d v2, #0000000000000000
	movi.2d v3, #0000000000000000
	movi.2d v4, #0000000000000000
	movi.2d v5, #0000000000000000
	movi.2d v6, #0000000000000000
	movi.2d v7, #0000000000000000
	mov x20, x5
LBB30_10:
	mov x21, x4
	ld1r.8h { v16 }, [x20], x19
	ldp q17, q18, [x4]
	; InlineAsm Start
	fmla.8h v7, v17, v16
	; InlineAsm End
	; InlineAsm Start
	fmla.8h v6, v18, v16
	; InlineAsm End
	add x4, x5, x10
	ld1r.8h { v16 }, [x4]
	; InlineAsm Start
	fmla.8h v5, v17, v16
	; InlineAsm End
	; InlineAsm Start
	fmla.8h v4, v18, v16
	; InlineAsm End
	add x4, x5, x15
	ld1r.8h { v16 }, [x4]
	; InlineAsm Start
	fmla.8h v3, v17, v16
	; InlineAsm End
	; InlineAsm Start
	fmla.8h v2, v18, v16
	; InlineAsm End
	add x4, x5, x13
	ld1r.8h { v16 }, [x4]
	; InlineAsm Start
	fmla.8h v1, v17, v16
	; InlineAsm End
	; InlineAsm Start
	fmla.8h v0, v18, v16
	; InlineAsm End
	add x4, x5, x12
	ld1r.8h { v16 }, [x4]
	add x22, x21, x11
	add x4, x22, x11
	ldp q17, q18, [x22]
	; InlineAsm Start
	fmla.8h v7, v17, v16
	; InlineAsm End
	; InlineAsm Start
	fmla.8h v6, v18, v16
	; InlineAsm End
	add x22, x5, x17
	ld1r.8h { v16 }, [x22]
	; InlineAsm Start
	fmla.8h v5, v17, v16
	; InlineAsm End
	; InlineAsm Start
	fmla.8h v4, v18, v16
	; InlineAsm End
	add x22, x5, x16
	ld1r.8h { v16 }, [x22]
	; InlineAsm Start
	fmla.8h v3, v17, v16
	; InlineAsm End
	; InlineAsm Start
	fmla.8h v2, v18, v16
	; InlineAsm End
	add x5, x5, x14
	ld1r.8h { v16 }, [x5]
	; InlineAsm Start
	fmla.8h v1, v17, v16
	; InlineAsm End
	; InlineAsm Start
	fmla.8h v0, v18, v16
	; InlineAsm End
	mov x5, x20
	subs x8, x8, #1
	b.ne LBB30_10
	add x4, x21, x11, lsl #1
	tbz w2, #0, LBB30_13

	add x8, x10, x9
	lsl x8, x8, #1
	mov x11, x5
	ld1r.8h { v16 }, [x11], x8
	ldp q17, q18, [x4]
	; InlineAsm Start
	fmla.8h v7, v17, v16
	; InlineAsm End
	; InlineAsm Start
	fmla.8h v6, v18, v16
	; InlineAsm End
	add x8, x5, x10
	ld1r.8h { v16 }, [x8]
	; InlineAsm Start
	fmla.8h v5, v17, v16
	; InlineAsm End
	; InlineAsm Start
	fmla.8h v4, v18, v16
	; InlineAsm End
	add x8, x5, x9, lsl #2
	ld1r.8h { v16 }, [x8]
	; InlineAsm Start
	fmla.8h v3, v17, v16
	; InlineAsm End
	; InlineAsm Start
	fmla.8h v2, v18, v16
	; InlineAsm End
	ld1r.8h { v16 }, [x11]
	; InlineAsm Start
	fmla.8h v1, v17, v16
	; InlineAsm End
	; InlineAsm Start
	fmla.8h v0, v18, v16
	; InlineAsm End
LBB30_13:
	ldrh w8, [sp, #186]
	ldrh w9, [sp, #184]
	ldrb w10, [sp, #188]
	cmp x0, #16
	b.ne LBB30_19
	cmp x1, #4
	b.ne LBB30_19
	cmp x7, #1
	b.ne LBB30_19
	dup.8h v16, w8
	and w8, w10, #0xff
	cmp w8, #1
	b.eq LBB30_27
	cmp w8, #2
	b.ne LBB30_28
	dup.8h v17, w9
	ldp q18, q19, [x3]
	; InlineAsm Start
	fmul.8h v20, v17, v18
	; InlineAsm End
	; InlineAsm Start
	fmul.8h v18, v16, v7
	; InlineAsm End
	; InlineAsm Start
	fadd.8h v7, v20, v18
	; InlineAsm End
	; InlineAsm Start
	fmul.8h v18, v17, v19
	; InlineAsm End
	; InlineAsm Start
	fmul.8h v19, v16, v6
	; InlineAsm End
	; InlineAsm Start
	fadd.8h v6, v18, v19
	; InlineAsm End
	stp q7, q6, [x3]
	add x8, x3, x6, lsl #1
	ldp q6, q7, [x8]
	; InlineAsm Start
	fmul.8h v18, v17, v6
	; InlineAsm End
	; InlineAsm Start
	fmul.8h v6, v16, v5
	; InlineAsm End
	; InlineAsm Start
	fadd.8h v5, v18, v6
	; InlineAsm End
	; InlineAsm Start
	fmul.8h v6, v17, v7
	; InlineAsm End
	; InlineAsm Start
	fmul.8h v7, v16, v4
	; InlineAsm End
	; InlineAsm Start
	fadd.8h v4, v6, v7
	; InlineAsm End
	stp q5, q4, [x8]
	add x8, x3, x6, lsl #2
	ldp q4, q5, [x8]
	; InlineAsm Start
	fmul.8h v6, v17, v4
	; InlineAsm End
	; InlineAsm Start
	fmul.8h v4, v16, v3
	; InlineAsm End
	; InlineAsm Start
	fadd.8h v3, v6, v4
	; InlineAsm End
	; InlineAsm Start
	fmul.8h v4, v17, v5
	; InlineAsm End
	; InlineAsm Start
	fmul.8h v5, v16, v2
	; InlineAsm End
	; InlineAsm Start
	fadd.8h v2, v4, v5
	; InlineAsm End
	stp q3, q2, [x8]
	mov w8, #6
	madd x8, x6, x8, x3
	ldp q2, q3, [x8]
	; InlineAsm Start
	fmul.8h v4, v17, v2
	; InlineAsm End
	; InlineAsm Start
	fmul.8h v2, v16, v1
	; InlineAsm End
	; InlineAsm Start
	fadd.8h v1, v4, v2
	; InlineAsm End
	; InlineAsm Start
	fmul.8h v2, v17, v3
	; InlineAsm End
	; InlineAsm Start
	fmul.8h v3, v16, v0
	; InlineAsm End
	; InlineAsm Start
	fadd.8h v0, v2, v3
	; InlineAsm End
	stp q1, q0, [x8]
	b LBB30_42
LBB30_19:
	stp q7, q6, [sp]
	stp q5, q4, [sp, #32]
	stp q3, q2, [sp, #64]
	and w10, w10, #0xff
	stp q1, q0, [sp, #96]
	cmp w10, #2
	b.eq LBB30_30
	cmp w10, #1
	b.ne LBB30_36
	cbz x1, LBB30_42
	cbz x0, LBB30_42
	mov x9, #0
	lsl x10, x6, #1
	lsl x11, x7, #1
	mov x12, sp
LBB30_24:
	mov x13, x0
	mov x14, x3
	mov x15, x12
LBB30_25:
	ldr h0, [x15], #2
	fmov s1, w8
	; InlineAsm Start
	fmul h2, h1, h0
	; InlineAsm End
	ldr h0, [x14]
	; InlineAsm Start
	fadd h1, h0, h2
	; InlineAsm End
	str h1, [x14]
	add x14, x14, x11
	subs x13, x13, #1
	b.ne LBB30_25
	add x9, x9, #1
	add x12, x12, #32
	add x3, x3, x10
	cmp x9, x1
	b.ne LBB30_24
	b LBB30_42
LBB30_27:
	ldr q17, [x3]
	; InlineAsm Start
	fmla.8h v17, v16, v7
	; InlineAsm End
	ldr q7, [x3, #16]
	; InlineAsm Start
	fmla.8h v7, v16, v6
	; InlineAsm End
	stp q17, q7, [x3]
	add x8, x3, x6, lsl #1
	ldr q6, [x8]
	; InlineAsm Start
	fmla.8h v6, v16, v5
	; InlineAsm End
	ldr q5, [x8, #16]
	; InlineAsm Start
	fmla.8h v5, v16, v4
	; InlineAsm End
	stp q6, q5, [x8]
	add x8, x3, x6, lsl #2
	ldr q4, [x8]
	; InlineAsm Start
	fmla.8h v4, v16, v3
	; InlineAsm End
	ldr q3, [x8, #16]
	; InlineAsm Start
	fmla.8h v3, v16, v2
	; InlineAsm End
	stp q4, q3, [x8]
	mov w8, #6
	madd x8, x6, x8, x3
	ldr q2, [x8]
	; InlineAsm Start
	fmla.8h v2, v16, v1
	; InlineAsm End
	str q2, [x8]
	ldr q1, [x8, #16]
	; InlineAsm Start
	fmla.8h v1, v16, v0
	; InlineAsm End
	b LBB30_29
LBB30_28:
	; InlineAsm Start
	fmul.8h v17, v16, v7
	; InlineAsm End
	; InlineAsm Start
	fmul.8h v7, v16, v6
	; InlineAsm End
	stp q17, q7, [x3]
	add x8, x3, x6, lsl #1
	; InlineAsm Start
	fmul.8h v6, v16, v5
	; InlineAsm End
	; InlineAsm Start
	fmul.8h v5, v16, v4
	; InlineAsm End
	stp q6, q5, [x8]
	add x8, x3, x6, lsl #2
	; InlineAsm Start
	fmul.8h v4, v16, v3
	; InlineAsm End
	; InlineAsm Start
	fmul.8h v3, v16, v2
	; InlineAsm End
	stp q4, q3, [x8]
	mov w8, #6
	madd x8, x6, x8, x3
	; InlineAsm Start
	fmul.8h v2, v16, v1
	; InlineAsm End
	; InlineAsm Start
	fmul.8h v1, v16, v0
	; InlineAsm End
	str q2, [x8]
LBB30_29:
	str q1, [x8, #16]
	b LBB30_42
LBB30_30:
	cbz x1, LBB30_42
	cbz x0, LBB30_42
	mov x10, #0
	lsl x11, x6, #1
	lsl x12, x7, #1
	mov x13, sp
LBB30_33:
	mov x14, x0
	mov x15, x3
	mov x16, x13
LBB30_34:
	ldr h0, [x15]
	fmov s1, w9
	; InlineAsm Start
	fmul h2, h1, h0
	; InlineAsm End
	ldr h0, [x16], #2
	fmov s1, w8
	; InlineAsm Start
	fmul h3, h1, h0
	; InlineAsm End
	; InlineAsm Start
	fadd h0, h2, h3
	; InlineAsm End
	str h0, [x15]
	add x15, x15, x12
	subs x14, x14, #1
	b.ne LBB30_34
	add x10, x10, #1
	add x13, x13, #32
	add x3, x3, x11
	cmp x10, x1
	b.ne LBB30_33
	b LBB30_42
LBB30_36:
	cbz x1, LBB30_42
	cbz x0, LBB30_42
	mov x9, #0
	lsl x10, x6, #1
	lsl x11, x7, #1
	mov x12, sp
LBB30_39:
	mov x13, x0
	mov x14, x3
	mov x15, x12
LBB30_40:
	ldr h0, [x15], #2
	fmov s1, w8
	; InlineAsm Start
	fmul h2, h1, h0
	; InlineAsm End
	str h2, [x14]
	add x14, x14, x11
	subs x13, x13, #1
	b.ne LBB30_40
	add x9, x9, #1
	add x12, x12, #32
	add x3, x3, x10
	cmp x9, x1
	b.ne LBB30_39
LBB30_42:
	ldp x20, x19, [sp, #144]
	ldp x22, x21, [sp, #128]
	add sp, sp, #160
	.cfi_def_cfa_offset 0
	.cfi_restore w19
	.cfi_restore w20
	.cfi_restore w21
	.cfi_restore w22
	ret