KleidiAI micro-kernel tables

This page provides an overview of the micro-kernels offered by KleidiAI.

Micro-kernel categories:

Matmul micro-kernels
Indirect matmul micro-kernels
Depthwise convolution micro-kernels
Packing micro-kernels.

Matmul micro-kernels

Output type	Output quantization	LHS type	RHS type	LHS quantization	RHS quantization	Block size	SIMD	Feature	Uarch	Micro-kernel	Packing micro-kernels
bf16	-	i8	i4	dimensionwise asymmetric	blockwise symmetric	1x4	Advanced SIMD	dotprod	-	`kai_matmul_clamp_bf16_qai8dxp1x8_qsi4c32p4x8_1x4_neon_dotprod`	LHS: `kai_lhs_quant_pack_qai8dxp_bf16_neon` RHS: `kai_rhs_pack_nxk_qsi4c32p_qsu4c32s1s0`, `kai_rhs_pack_kxn_qsi4c32p_qsu4c32s1s0`
bf16	-	i8	i4	dimensionwise asymmetric	channelwise symmetric	1x8	Advanced SIMD	dotprod	-	`kai_matmul_clamp_bf16_qai8dxp1x8_qsi4cxp8x8_1x8_neon_dotprod`	LHS: `kai_lhs_quant_pack_qai8dxp_bf16_neon` RHS: `kai_rhs_pack_nxk_qsi4cxp_qs4cxs1s0`, `kai_rhs_pack_kxn_qsi4cxp_qs4cxs1s0`
bf16	-	i8	i4	dimensionwise asymmetric	blockwise symmetric	16x4	Advanced SIMD	i8mm	-	`kai_matmul_clamp_bf16_qai8dxp4x8_qsi4c32p4x8_16x4_neon_i8mm`	LHS: `kai_lhs_quant_pack_qai8dxp_bf16_neon` RHS: `kai_rhs_pack_nxk_qsi4c32p_qsu4c32s1s0`, `kai_rhs_pack_kxn_qsi4c32p_qsu4c32s1s0`
bf16	-	i8	i4	dimensionwise asymmetric	channelwise symmetric	8x8	Advanced SIMD	i8mm	-	`kai_matmul_clamp_bf16_qai8dxp4x8_qsi4cxp8x8_8x8_neon_i8mm`	LHS: `kai_lhs_quant_pack_qai8dxp_bf16_neon` RHS: `kai_rhs_pack_nxk_qsi4cxp_qs4cxs1s0`, `kai_rhs_pack_kxn_qsi4cxp_qs4cxs1s0`
f16	-	bf16	bf16	-	-	8x12	Advanced SIMD	mmla	-	`kai_matmul_clamp_f16_bf16p8x4_bf16p12x4b_8x12_neon_mmla`	LHS: `kai_lhs_pack_bf16p8x4_f16_neon`, `kai_lhs_quant_pack_bf16p8x4_f32_neon` RHS: `kai_rhs_pack_kxn_bf16p12x4biasf16_f16_neon`, `kai_rhs_pack_kxn_bf16p12x4biasf32_f16_neon`
f16	-	f16	f16	-	-	6x16x8	Advanced SIMD	mla	-	`kai_matmul_clamp_f16_f16_f16p16x1biasf16_6x16x8_neon_mla`	RHS: `kai_rhs_pack_kxn_f16p16x1biasf16_f16_f16_neon`
f16	-	f16	f16	-	-	1x16vl	SME2	dot	-	`kai_matmul_clamp_f16_f16_f16p2vlx2b_1x16vl_sme2_dot`	RHS: `kai_rhs_pack_kxn_x16p2vlx2b_x16_x16_sme`
f16	-	f16	f16	-	-	1x8vl	SME	mla	-	`kai_matmul_clamp_f16_f16_f16p2vlx2b_1x8vl_sme_mla`	RHS: `kai_rhs_pack_kxn_x16p2vlx2b_x16_x16_sme`, `kai_rhs_pack_nxk_x16p2vlx2b_x16_x16_sme`
f16	-	f16	f16	-	-	6x32	Advanced SIMD	mla	-	`kai_matmul_clamp_f16_f16_f16p32x1b_6x32_neon_mla`	RHS: `kai_rhs_pack_kxn_x16p32x1b_x16_x16_neon`
f16	-	f16	f16	-	-	6x32	Advanced SIMD	mla	cortexa55	`kai_matmul_clamp_f16_f16_f16p32x1b_6x32_neon_mla_cortexa55`	RHS: `kai_rhs_pack_kxn_x16p32x1b_x16_x16_neon`
f16	-	f16	f16	-	-	2vlx2vl	SME2	mopa	-	`kai_matmul_clamp_f16_f16p2vlx2_f16p2vlx2_2vlx2vl_sme2_mopa`	LHS: `kai_lhs_pack_x16p2vlx2_x16_sme` RHS: `kai_rhs_pack_kxn_x16p2vlx2b_x16_x16_sme`, `kai_rhs_pack_nxk_x16p2vlx2b_x16_x16_sme`
f16	-	f16	f16	-	-	2vlx2vl	SME	mopa	-	`kai_matmul_clamp_f16_f16p2vlx2_f16p2vlx2b_2vlx2vl_sme_mopa`	LHS: `kai_lhs_pack_x16p2vlx2_x16_sme` RHS: `kai_rhs_pack_kxn_x16p2vlx2b_x16_x16_sme`, `kai_rhs_pack_nxk_x16p2vlx2b_x16_x16_sme`
f16	-	i8	i4	dimensionwise asymmetric	channelwise symmetric	1x4	Advanced SIMD	dotprod	-	`kai_matmul_clamp_f16_qai8dxp1x4_qsi4cxp4x4_1x4_neon_dotprod`	LHS: `kai_lhs_quant_pack_qai8dxp_f16_neon` RHS: `kai_rhs_pack_nxk_qsi4cxp_qs4cxs1s0`, `kai_rhs_pack_kxn_qsi4cxp_qs4cxs1s0`
f16	-	i8	i8	dimensionwise asymmetric	channelwise symmetric	1x4	Advanced SIMD	dotprod	-	`kai_matmul_clamp_f16_qai8dxp1x4_qsi8cxp4x4_1x4_neon_dotprod`	LHS: `kai_lhs_quant_pack_qai8dxp_f16_neon` RHS: `kai_rhs_pack_nxk_qsi8cxp_qsi8cx_neon`, `kai_rhs_pack_kxn_qsi8cxp_qsi8cx_neon`
f16	-	i8	i4	dimensionwise asymmetric	channelwise symmetric	1x4	Advanced SIMD	dotprod	-	`kai_matmul_clamp_f16_qai8dxp1x8_qsi4cxp4x8_1x4_neon_dotprod`	LHS: `kai_lhs_quant_pack_qai8dxp_f16_neon` RHS: `kai_rhs_pack_nxk_qsi4cxp_qs4cxs1s0`, `kai_rhs_pack_kxn_qsi4cxp_qs4cxs1s0`
f16	-	i8	i8	dimensionwise asymmetric	channelwise symmetric	1x4	Advanced SIMD	dotprod	-	`kai_matmul_clamp_f16_qai8dxp1x8_qsi8cxp4x8_1x4_neon_dotprod`	LHS: `kai_lhs_quant_pack_qai8dxp_f16_neon` RHS: `kai_rhs_pack_nxk_qsi8cxp_qsi8cx_neon`, `kai_rhs_pack_kxn_qsi8cxp_qsi8cx_neon`
f16	-	i8	i4	dimensionwise asymmetric	channelwise symmetric	16x4	Advanced SIMD	dotprod	-	`kai_matmul_clamp_f16_qai8dxp4x4_qsi4cxp4x4_16x4_neon_dotprod`	LHS: `kai_lhs_quant_pack_qai8dxp_f16_neon` RHS: `kai_rhs_pack_nxk_qsi4cxp_qs4cxs1s0`, `kai_rhs_pack_kxn_qsi4cxp_qs4cxs1s0`
f16	-	i8	i8	dimensionwise asymmetric	channelwise symmetric	16x4	Advanced SIMD	dotprod	-	`kai_matmul_clamp_f16_qai8dxp4x4_qsi8cxp4x4_16x4_neon_dotprod`	LHS: `kai_lhs_quant_pack_qai8dxp_f16_neon` RHS: `kai_rhs_pack_nxk_qsi8cxp_qsi8cx_neon`, `kai_rhs_pack_kxn_qsi8cxp_qsi8cx_neon`
f16	-	i8	i4	dimensionwise asymmetric	channelwise symmetric	16x4	Advanced SIMD	i8mm	-	`kai_matmul_clamp_f16_qai8dxp4x8_qsi4cxp4x8_16x4_neon_i8mm`	LHS: `kai_lhs_quant_pack_qai8dxp_f16_neon` RHS: `kai_rhs_pack_nxk_qsi4cxp_qs4cxs1s0`, `kai_rhs_pack_kxn_qsi4cxp_qs4cxs1s0`
f16	-	i8	i8	dimensionwise asymmetric	channelwise symmetric	16x4	Advanced SIMD	i8mm	-	`kai_matmul_clamp_f16_qai8dxp4x8_qsi8cxp4x8_16x4_neon_i8mm`	LHS: `kai_lhs_quant_pack_qai8dxp_f16_neon` RHS: `kai_rhs_pack_nxk_qsi8cxp_qsi8cx_neon`, `kai_rhs_pack_kxn_qsi8cxp_qsi8cx_neon`
f16	-	i8	i8	dimensionwise asymmetric	channelwise symmetric	1vlx4vl	SME2	mopa	-	`kai_matmul_clamp_f16_qai8dxp1vlx4_qsi8cxp4vlx4_1vlx4vl_sme2_mopa`	LHS: `kai_lhs_quant_pack_qai8dxp_f32` RHS: `kai_rhs_pack_nxk_qsi8cxp_qsi8cx_neon`, `kai_rhs_pack_kxn_qsi8cxp_qsi8cx_neon`
f16	-	i8	i8	dimensionwise asymmetric	channelwise symmetric	1x4vl	SME2	dot	-	`kai_matmul_clamp_f16_qai8dxp1x4_qsi8cxp4vlx4_1x4vl_sme2_dot`	LHS: `kai_lhs_quant_pack_qai8dxp_f32` RHS: `kai_rhs_pack_nxk_qsi8cxp_qsi8cx_neon`, `kai_rhs_pack_kxn_qsi8cxp_qsi8cx_neon`
f16	-	i8	i4	blockwise per dimension symmetric	blockwise asymmetric	1vlx4vl	SME2	mopa	-	`kai_matmul_clamp_f16_qsi8d32p1vlx4_qai4c32p4vlx4_1vlx4vl_sme2_mopa`	LHS: `kai_lhs_quant_pack_qsi8d32pscalef32_f16_neon` RHS: `kai_rhs_pack_nxk_qai4c32ps1s0nrx4_qau4c32s1s0_f32_f32_f32_neon`, `kai_rhs_pack_nxk_qai4c32ps1s0nrx4_qau4c32s0s1_f32_f32_f32_neon`
f16	-	i8	i4	blockwise per dimension symmetric	blockwise asymmetric	1x4vl	SME2	dot	-	`kai_matmul_clamp_f16_qsi8d32p1x4_qai4c32p4vlx4_1x4vl_sme2_dot`	LHS: `kai_lhs_quant_pack_qsi8d32pscalef32_f16_neon` RHS: `kai_rhs_pack_nxk_qai4c32ps1s0nrx4_qau4c32s1s0_f32_f32_f32_neon`, `kai_rhs_pack_nxk_qai4c32ps1s0nrx4_qau4c32s0s1_f32_f32_f32_neon`
f16	-	i8	i4	blockwise per dimension symmetric	blockwise asymmetric	1x4	Advanced SIMD	dotprod	-	`kai_matmul_clamp_f16_qsi8d32p1x4_qai4c32p4x4_1x4_neon_dotprod`	LHS: `kai_lhs_quant_pack_qsi8d32pscalef32_f16_neon` RHS: `kai_rhs_pack_nxk_qai4c32p_qau4c32s0s1_f32_f32_f32_neon`
f16	-	i8	i4	blockwise per dimension symmetric	blockwise asymmetric	1x4	Advanced SIMD	dotprod	-	`kai_matmul_clamp_f16_qsi8d32p1x8_qai4c32p4x8_1x4_neon_dotprod`	LHS: `kai_lhs_quant_pack_qsi8d32pscalef32_f16_neon` RHS: `kai_rhs_pack_nxk_qai4c32p_qau4c32s0s1_f32_f32_f32_neon`
f16	-	i8	i4	blockwise per dimension symmetric	blockwise asymmetric	8x4	Advanced SIMD	dotprod	-	`kai_matmul_clamp_f16_qsi8d32p4x4_qai4c32p4x4_8x4_neon_dotprod`	LHS: `kai_lhs_quant_pack_qsi8d32pscalef32_f16_neon` RHS: `kai_rhs_pack_nxk_qai4c32p_qau4c32s0s1_f32_f32_f32_neon`
f16	-	i8	i4	blockwise per dimension symmetric	blockwise asymmetric	8x4	Advanced SIMD	i8mm	-	`kai_matmul_clamp_f16_qsi8d32p4x8_qai4c32p4x8_8x4_neon_i8mm`	LHS: `kai_lhs_quant_pack_qsi8d32pscalef32_f16_neon` RHS: `kai_rhs_pack_nxk_qai4c32p_qau4c32s0s1_f32_f32_f32_neon`
f32	-	bf16	bf16	-	-	1x36	Advanced SIMD	dot	-	`kai_matmul_clamp_f32_bf16p1x4_bf16p12x4b_1x36_neon_dot`	LHS: `kai_lhs_quant_pack_bf16p1x4_f32_neon` RHS: `kai_rhs_quant_pack_kxn_bf16p12x4biasf32_f32_neon`
f32	-	bf16	bf16	-	-	2vlx2vl	SME2	mopa	-	`kai_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa`	LHS: `kai_lhs_pack_bf16p2vlx2_f32_sme` RHS: `kai_rhs_pack_kxn_bf16p2vlx2b_f32_x32_sme`
f32	-	bf16	bf16	-	-	8x12	Advanced SIMD	mmla	-	`kai_matmul_clamp_f32_bf16p8x4_bf16p12x4b_8x12_neon_mmla`	-
f32	-	f16	i4	-	blockwise symmetric	1vlx4vl	SME2	mopa	-	`kai_matmul_clamp_f32_f16p1vlx2_qsi4c32p4vlx2_1vlx4vl_sme2_mopa`	LHS: `kai_lhs_pack_f16pmrx2_f32_neon` RHS: `kai_rhs_pack_nxk_qsi4c32ps1s0scalef16_qsu4c32s16s0_neon`
f32	-	f32	f32	-	-	1x16vl	SME2	mla	-	`kai_matmul_clamp_f32_f32_f32p16vlx1b_1x16vl_sme2_mla`	RHS: `kai_rhs_pack_kxn_f32p16vlx1b_f32_f32_sme`
f32	-	f32	f32	-	-	6x16	Advanced SIMD	mla	-	`kai_matmul_clamp_f32_f32_f32p16x1b_6x16_neon_mla`	RHS: `kai_rhs_pack_kxn_x32p16x1b_x32_x32_neon`
f32	-	f32	f32	-	-	6x16	Advanced SIMD	mla	cortexa55	`kai_matmul_clamp_f32_f32_f32p16x1b_6x16_neon_mla_cortexa55`	RHS: `kai_rhs_pack_kxn_x32p16x1b_x32_x32_neon`
f32	-	f32	f32	-	-	1x16vl	SME2	mla	-	`kai_matmul_clamp_f32_f32_f32p2vlx1b_1x16vl_sme2_mla`	RHS: `kai_rhs_pack_kxn_f32p2vlx1biasf32_f32_f32_sme`
f32	-	f32	f32	-	-	1x8vl	SME	mla	-	`kai_matmul_clamp_f32_f32_f32p2vlx1b_1x8vl_sme_mla`	RHS: `kai_rhs_pack_kxn_f32p2vlx1biasf32_f32_f32_sme`, `kai_rhs_pack_nxk_f32p2vlx1biasf32_f32_f32_sme`
f32	-	f32	f32	-	-	6x4vl	SVE	mla	-	`kai_matmul_clamp_f32_f32_f32p4vlx1b_6x4vl_sve_mla`	RHS: `kai_rhs_pack_kxn_x32p4vlx1b_x32_x32_sve`
f32	-	f32	f32	-	-	6x8x4	Advanced SIMD	mla	-	`kai_matmul_clamp_f32_f32_f32p8x1biasf32_6x8x4_neon_mla`	RHS: `kai_rhs_pack_kxn_f32p8x1biasf32_f32_f32_neon`
f32	-	f32	f32	-	-	2vlx2vl	SME	mopa	-	`kai_matmul_clamp_f32_f32p2vlx1_f32p2vlx1b_2vlx2vl_sme_mopa`	LHS: `kai_lhs_pack_f32p2vlx1_f32_sme` RHS: `kai_rhs_pack_kxn_f32p2vlx1biasf32_f32_f32_sme`, `kai_rhs_pack_nxk_f32p2vlx1biasf32_f32_f32_sme`
f32	-	f32	f32	-	-	-	SME2	mopa	-	`kai_matmul_clamp_f32_f32p2vlx1_f32p2vlx1biasf32_sme2_mopa`	LHS: `kai_lhs_pack_f32p2vlx1_f32_sme` RHS: `kai_rhs_pack_kxn_f32p2vlx1biasf32_f32_f32_sme`
f32	-	f32	f32	-	-	4vsx4vs	SME2	mopa	-	`kai_matmul_clamp_f32_f32p4vsx1_f32p4vsx1bf32_4vsx4vs_sme2_mopa`	LHS: `kai_matmul_pack_lhs_mxk_x32p4vsx1_x32_sme` RHS: `kai_matmul_pack_rhs_kxn_x32p4vsx1bx32_x32_x32_sme`, `kai_matmul_pack_rhs_nxk_x32p4vsx1bx32_x32_x32_sme`
f32	-	f32	f32	-	-	4vsx8vs	SME2	mopa	-	`kai_matmul_clamp_f32_f32p4vsx1_f32p4vsx1bf32_4vsx8vs_sme2_mopa`	LHS: `kai_matmul_pack_lhs_mxk_x32p4vsx1_x32_sme` RHS: `kai_matmul_pack_rhs_kxn_x32p4vsx1bx32_x32_x32_sme`, `kai_matmul_pack_rhs_nxk_x32p4vsx1bx32_x32_x32_sme`
f32	-	f32	f32	-	-	4vsx16vs	SME2	mopa	-	`kai_matmul_clamp_f32_f32p4vsx1_f32p4vsx1bf32_4vsx16vs_sme2_mopa`	LHS: `kai_matmul_pack_lhs_mxk_x32p4vsx1_x32_sme` RHS: `kai_matmul_pack_rhs_kxn_x32p4vsx1bx32_x32_x32_sme`, `kai_matmul_pack_rhs_nxk_x32p4vsx1bx32_x32_x32_sme`
f32	-	f32	f32	-	-	8vsx4vs	SME2	mopa	-	`kai_matmul_clamp_f32_f32p4vsx1_f32p4vsx1bf32_8vsx4vs_sme2_mopa`	LHS: `kai_matmul_pack_lhs_mxk_x32p4vsx1_x32_sme` RHS: `kai_matmul_pack_rhs_kxn_x32p4vsx1bx32_x32_x32_sme`, `kai_matmul_pack_rhs_nxk_x32p4vsx1bx32_x32_x32_sme`
f32	-	f32	f32	-	-	8vsx8vs	SME2	mopa	-	`kai_matmul_clamp_f32_f32p4vsx1_f32p4vsx1bf32_8vsx8vs_sme2_mopa`	LHS: `kai_matmul_pack_lhs_mxk_x32p4vsx1_x32_sme` RHS: `kai_matmul_pack_rhs_kxn_x32p4vsx1bx32_x32_x32_sme`, `kai_matmul_pack_rhs_nxk_x32p4vsx1bx32_x32_x32_sme`
f32	-	f32	f32	-	-	16vsx4vs	SME2	mopa	-	`kai_matmul_clamp_f32_f32p4vsx1_f32p4vsx1bf32_16vsx4vs_sme2_mopa`	LHS: `kai_matmul_pack_lhs_mxk_x32p4vsx1_x32_sme` RHS: `kai_matmul_pack_rhs_kxn_x32p4vsx1bx32_x32_x32_sme`, `kai_matmul_pack_rhs_nxk_x32p4vsx1bx32_x32_x32_sme`
f32	-	i8	i4	dimensionwise asymmetric	blockwise symmetric	1vlx4vl	SME2	mopa	-	`kai_matmul_clamp_f32_qai8dxp1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa`	LHS: `kai_lhs_quant_pack_qai8dxp_f32` RHS: `kai_rhs_pack_nxk_qsi4c32ps1s0nrx4_qsu4c32s1s0_neon`, `kai_rhs_pack_kxn_qsi4c32ps1s0nrx4_qsu4c32s1s0_neon`
f32	-	i8	i8	dimensionwise asymmetric	channelwise symmetric	1vlx4vl	SME2	mopa	-	`kai_matmul_clamp_f32_qai8dxp1vlx4_qsi8cxp4vlx4_1vlx4vl_sme2_mopa`	LHS: `kai_lhs_quant_pack_qai8dxp_f32` RHS: `kai_rhs_pack_nxk_qsi8cxp_qsi8cx_neon`, `kai_rhs_pack_kxn_qsi8cxp_qsi8cx_neon`
f32	-	i8	i8	dimensionwise asymmetric	channelwise symmetric	1vlx4vl	SME	mopa	-	`kai_matmul_clamp_f32_qai8dxp1vlx4_qsi8cxp4vlx4_1vlx4vl_sme_mopa`	LHS: `kai_lhs_quant_pack_qai8dxp_f32` RHS: `kai_rhs_pack_nxk_qsi8cxp_qsi8cx_neon`, `kai_rhs_pack_kxn_qsi8cxp_qsi8cx_neon`
f32	-	i8	u2	dimensionwise asymmetric	channelwise symmetric	1vlx4vl	SME2	mopa	-	`kai_matmul_clamp_f32_qai8dxp1vlx4_qsu2cxp4vlx4_1vlx4vl_sme2_mopa`	LHS: `kai_lhs_quant_pack_qai8dxp_f32` RHS: `kai_rhs_pack_nxk_qsu2cxp4vlx4_qsu2cx_neon`
f32	-	i8	i4	dimensionwise asymmetric	channelwise symmetric	1vlx4vl	SME2	mopa	-	`kai_matmul_clamp_f32_qai8dxp1vlx8_qsi4cxp4vlx8_1vlx4vl_sme2_mopa`	LHS: `kai_lhs_quant_pack_qai8dxp_f32` RHS: `kai_rhs_pack_nxk_qsi4cxps1s0_qsu4cxs1s0_neon`
f32	-	i8	i4	dimensionwise asymmetric	channelwise symmetric	1vlx4vl	SME	mopa	-	`kai_matmul_clamp_f32_qai8dxp1vlx4_qsi4cxp4vlx4_1vlx4vl_sme_mopa`	LHS: `kai_lhs_quant_pack_qai8dxp_f32` RHS: `kai_rhs_pack_nxk_qsi4cxp_qs4cxs1s0`
f32	-	i8	i4	dimensionwise asymmetric	blockwise symmetric	1x4vl	SME2	dot	-	`kai_matmul_clamp_f32_qai8dxp1x4_qsi4c32p4vlx4_1x4vl_sme2_dot`	LHS: `kai_lhs_quant_pack_qai8dxp_f32` RHS: `kai_rhs_pack_nxk_qsi4c32ps1s0nrx4_qsu4c32s1s0_neon`, `kai_rhs_pack_kxn_qsi4c32ps1s0nrx4_qsu4c32s1s0_neon`
f32	-	i8	i4	dimensionwise asymmetric	blockwise symmetric	1x4	Advanced SIMD	dotprod	-	`kai_matmul_clamp_f32_qai8dxp1x4_qsi4c32p4x4_1x4_neon_dotprod`	LHS: `kai_lhs_quant_pack_qai8dxp_f32` RHS: `kai_rhs_pack_nxk_qsi4c32p_qsu4c32s1s0`, `kai_rhs_pack_kxn_qsi4c32p_qsu4c32s1s0`
f32	-	i8	i4	dimensionwise asymmetric	blockwise symmetric	1x8	Advanced SIMD	dotprod	-	`kai_matmul_clamp_f32_qai8dxp1x4_qsi4c32p8x4_1x8_neon_dotprod`	LHS: `kai_lhs_quant_pack_qai8dxp_f32` RHS: `kai_rhs_pack_nxk_qsi4c32p_qsu4c32s1s0`, `kai_rhs_pack_kxn_qsi4c32p_qsu4c32s1s0`
f32	-	i8	i4	dimensionwise asymmetric	channelwise symmetric	1x4vl	SME2	sdot	-	`kai_matmul_clamp_f32_qai8dxp1x4_qsi4cxp4vlx4_1x4vl_sme2_sdot`	LHS: `kai_lhs_quant_pack_qai8dxp_f32` RHS: `kai_rhs_pack_nxk_qsi4cxps1s0_qsu4cxs1s0_neon`
f32	-	i8	i4	dimensionwise asymmetric	channelwise symmetric	1x4vl	SME	sdot	-	`kai_matmul_clamp_f32_qai8dxp1x4_qsi4cxp4vlx4_1x4vl_sme_dot`	LHS: `kai_lhs_quant_pack_qai8dxp_f32` RHS: `kai_rhs_pack_nxk_qsi4cxp_qs4cxs1s0`
f32	-	i8	i4	dimensionwise asymmetric	channelwise symmetric	1x4	Advanced SIMD	dotprod	-	`kai_matmul_clamp_f32_qai8dxp1x4_qsi4cxp4x4_1x4_neon_dotprod`	LHS: `kai_lhs_quant_pack_qai8dxp_f32` RHS: `kai_rhs_pack_nxk_qsi4cxp_qs4cxs1s0`, `kai_rhs_pack_kxn_qsi4cxp_qs4cxs1s0`
f32	-	i8	i8	dimensionwise asymmetric	channelwise symmetric	1x4vl	SME2	dot	-	`kai_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4vlx4_1x4vl_sme2_dot`	LHS: `kai_lhs_quant_pack_qai8dxp_f32` RHS: `kai_rhs_pack_nxk_qsi8cxp_qsi8cx_neon`, `kai_rhs_pack_kxn_qsi8cxp_qsi8cx_neon`
f32	-	i8	i8	dimensionwise asymmetric	channelwise symmetric	1x4vl	SME	dot	-	`kai_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4vlx4_1x4vl_sme_dot`	LHS: `kai_lhs_quant_pack_qai8dxp_f32` RHS: `kai_rhs_pack_nxk_qsi8cxp_qsi8cx_neon`, `kai_rhs_pack_kxn_qsi8cxp_qsi8cx_neon`
f32	-	i8	i8	dimensionwise asymmetric	channelwise symmetric	1x4	Advanced SIMD	dotprod	-	`kai_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4x4_1x4_neon_dotprod`	LHS: `kai_lhs_quant_pack_qai8dxp_f32` RHS: `kai_rhs_pack_nxk_qsi8cxp_qsi8cx_neon`, `kai_rhs_pack_kxn_qsi8cxp_qsi8cx_neon`
f32	-	i8	u2	dimensionwise asymmetric	channelwise symmetric	1x4vl	SME2	dot	-	`kai_matmul_clamp_f32_qai8dxp1x4_qsu2cxp4vlx4_1x4vl_sme2_dot`	LHS: `kai_lhs_quant_pack_qai8dxp_f32` RHS: `kai_rhs_pack_nxk_qsu2cxp4vlx4_qsu2cx_neon`
f32	-	i8	i4	dimensionwise asymmetric	blockwise symmetric	1x4x32	Advanced SIMD	dotprod	-	`kai_matmul_clamp_f32_qai8dxp1x8_qsi4c32p4x8_1x4x32_neon_dotprod`	LHS: `kai_lhs_quant_pack_qai8dxp_f32` RHS: `kai_rhs_pack_nxk_qsi4c32p_qsu4c32s1s0`, `kai_rhs_pack_kxn_qsi4c32p_qsu4c32s1s0`
f32	-	i8	i4	dimensionwise asymmetric	blockwise symmetric	1x8	Advanced SIMD	dotprod	-	`kai_matmul_clamp_f32_qai8dxp1x8_qsi4c32p8x8_1x8_neon_dotprod`	LHS: `kai_lhs_quant_pack_qai8dxp_f32` RHS: `kai_rhs_pack_nxk_qsi4c32p_qsu4c32s1s0`, `kai_rhs_pack_kxn_qsi4c32p_qsu4c32s1s0`
f32	-	i8	i4	dimensionwise asymmetric	blockwise symmetric	1x8x32	Advanced SIMD	dotprod	-	`kai_matmul_clamp_f32_qai8dxp1x8_qsi4c32p8x8_1x8x32_neon_dotprod`	LHS: `kai_lhs_quant_pack_qai8dxp_f32` RHS: `kai_rhs_pack_nxk_qsi4c32p_qsu4c32s1s0`, `kai_rhs_pack_kxn_qsi4c32p_qsu4c32s1s0`
f32	-	i8	i4	dimensionwise asymmetric	channelwise symmetric	1x4x32	Advanced SIMD	dotprod	-	`kai_matmul_clamp_f32_qai8dxp1x8_qsi4cxp4x8_1x4x32_neon_dotprod`	LHS: `kai_lhs_quant_pack_qai8dxp_f32` RHS: `kai_rhs_pack_kxn_qsi4cxp_qs4cxs1s0`, `kai_rhs_pack_nxk_qsi4cxp_qs4cxs1s0`
f32	-	i8	i4	dimensionwise asymmetric	channelwise symmetric	1x8x32	Advanced SIMD	dotprod	-	`kai_matmul_clamp_f32_qai8dxp1x8_qsi4cxp8x8_1x8x32_neon_dotprod`	LHS: `kai_lhs_quant_pack_qai8dxp_f32` RHS: `kai_rhs_pack_kxn_qsi4cxp_qs4cxs1s0`, `kai_rhs_pack_nxk_qsi4cxp_qs4cxs1s0`
f32	-	i8	i8	dimensionwise asymmetric	channelwise symmetric	1x4	Advanced SIMD	dotprod	-	`kai_matmul_clamp_f32_qai8dxp1x8_qsi8cxp4x8_1x4_neon_dotprod`	LHS: `kai_lhs_quant_pack_qai8dxp_f32` RHS: `kai_rhs_pack_nxk_qsi8cxp_qsi8cx_neon`, `kai_rhs_pack_kxn_qsi8cxp_qsi8cx_neon`
f32	-	i8	i4	dimensionwise asymmetric	blockwise symmetric	16x4	Advanced SIMD	dotprod	-	`kai_matmul_clamp_f32_qai8dxp4x4_qsi4c32p4x4_16x4_neon_dotprod`	LHS: `kai_lhs_quant_pack_qai8dxp_f32` RHS: `kai_rhs_pack_nxk_qsi4c32p_qsu4c32s1s0`, `kai_rhs_pack_kxn_qsi4c32p_qsu4c32s1s0`
f32	-	i8	i4	dimensionwise asymmetric	blockwise symmetric	4x8	Advanced SIMD	dotprod	-	`kai_matmul_clamp_f32_qai8dxp4x4_qsi4c32p8x4_4x8_neon_dotprod`	LHS: `kai_lhs_quant_pack_qai8dxp_f32` RHS: `kai_rhs_pack_nxk_qsi4c32p_qsu4c32s1s0`, `kai_rhs_pack_kxn_qsi4c32p_qsu4c32s1s0`
f32	-	i8	i4	dimensionwise asymmetric	channelwise symmetric	8x8x32	Advanced SIMD	dotprod	-	`kai_matmul_clamp_f32_qai8dxp4x4_qsi4cxp8x4_8x8x32_neon_dotprod`	LHS: `kai_lhs_quant_pack_qai8dxp_f32` RHS: `kai_rhs_pack_kxn_qsi4cxp_qs4cxs1s0`, `kai_rhs_pack_nxk_qsi4cxp_qs4cxs1s0`
f32	-	i8	i8	dimensionwise asymmetric	channelwise symmetric	16x4	Advanced SIMD	dotprod	-	`kai_matmul_clamp_f32_qai8dxp4x4_qsi8cxp4x4_16x4_neon_dotprod`	LHS: `kai_lhs_quant_pack_qai8dxp_f32` RHS: `kai_rhs_pack_nxk_qsi8cxp_qsi8cx_neon`, `kai_rhs_pack_kxn_qsi8cxp_qsi8cx_neon`
f32	-	i8	i4	dimensionwise asymmetric	blockwise symmetric	16x4x32	Advanced SIMD	i8mm	-	`kai_matmul_clamp_f32_qai8dxp4x8_qsi4c32p4x8_16x4x32_neon_i8mm`	LHS: `kai_lhs_quant_pack_qai8dxp_f32` RHS: `kai_rhs_pack_nxk_qsi4c32p_qsu4c32s1s0`, `kai_rhs_pack_kxn_qsi4c32p_qsu4c32s1s0`
f32	-	i8	i4	dimensionwise asymmetric	blockwise symmetric	8x4x32	Advanced SIMD	i8mm	-	`kai_matmul_clamp_f32_qai8dxp4x8_qsi4c32p4x8_8x4x32_neon_i8mm`	LHS: `kai_lhs_quant_pack_qai8dxp_f32` RHS: `kai_rhs_pack_nxk_qsi4c32p_qsu4c32s1s0`, `kai_rhs_pack_kxn_qsi4c32p_qsu4c32s1s0`
f32	-	i8	i4	dimensionwise asymmetric	blockwise symmetric	4x8	Advanced SIMD	i8mm	-	`kai_matmul_clamp_f32_qai8dxp4x8_qsi4c32p8x8_4x8_neon_i8mm`	LHS: `kai_lhs_quant_pack_qai8dxp_f32` RHS: `kai_rhs_pack_nxk_qsi4c32p_qsu4c32s1s0`, `kai_rhs_pack_kxn_qsi4c32p_qsu4c32s1s0`
f32	-	i8	i4	dimensionwise asymmetric	blockwise symmetric	4x8x32	Advanced SIMD	i8mm	-	`kai_matmul_clamp_f32_qai8dxp4x8_qsi4c32p8x8_4x8x32_neon_i8mm`	LHS: `kai_lhs_quant_pack_qai8dxp_f32` RHS: `kai_rhs_pack_nxk_qsi4c32p_qsu4c32s1s0`, `kai_rhs_pack_kxn_qsi4c32p_qsu4c32s1s0`
f32	-	i8	i4	dimensionwise asymmetric	channelwise symmetric	16x4x32	Advanced SIMD	dotprod	-	`kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp4x4_16x4x32_neon_dotprod`	LHS: `kai_lhs_quant_pack_qai8dxp_f32` RHS: `kai_rhs_pack_kxn_qsi4cxp_qs4cxs1s0`, `kai_rhs_pack_nxk_qsi4cxp_qs4cxs1s0`
f32	-	i8	i4	dimensionwise asymmetric	channelwise symmetric	4x4x32	Advanced SIMD	i8mm	-	`kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp4x8_4x4x32_neon_i8mm`	LHS: `kai_lhs_quant_pack_qai8dxp_f32` RHS: `kai_rhs_pack_kxn_qsi4cxp_qs4cxs1s0`, `kai_rhs_pack_nxk_qsi4cxp_qs4cxs1s0`
f32	-	i8	i4	dimensionwise asymmetric	channelwise symmetric	8x4x32	Advanced SIMD	i8mm	-	`kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp4x8_8x4x32_neon_i8mm`	LHS: `kai_lhs_quant_pack_qai8dxp_f32` RHS: `kai_rhs_pack_kxn_qsi4cxp_qs4cxs1s0`, `kai_rhs_pack_nxk_qsi4cxp_qs4cxs1s0`
f32	-	i8	i4	dimensionwise asymmetric	channelwise symmetric	4x8x32	Advanced SIMD	i8mm	-	`kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp8x8_4x8x32_neon_i8mm`	LHS: `kai_lhs_quant_pack_qai8dxp_f32` RHS: `kai_rhs_pack_kxn_qsi4cxp_qs4cxs1s0`, `kai_rhs_pack_nxk_qsi4cxp_qs4cxs1s0`
f32	-	i8	i4	dimensionwise asymmetric	channelwise symmetric	8x8x32	Advanced SIMD	i8mm	-	`kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp8x8_8x8x32_neon_i8mm`	LHS: `kai_lhs_quant_pack_qai8dxp_f32` RHS: `kai_rhs_pack_kxn_qsi4cxp_qs4cxs1s0`, `kai_rhs_pack_nxk_qsi4cxp_qs4cxs1s0`
f32	-	i8	i8	dimensionwise asymmetric	channelwise symmetric	16x4	Advanced SIMD	i8mm	-	`kai_matmul_clamp_f32_qai8dxp4x8_qsi8cxp4x8_16x4_neon_i8mm`	LHS: `kai_lhs_quant_pack_qai8dxp_f32` RHS: `kai_rhs_pack_nxk_qsi8cxp_qsi8cx_neon`, `kai_rhs_pack_kxn_qsi8cxp_qsi8cx_neon`
f32	-	i8	i4	blockwise per dimension symmetric	blockwise asymmetric	1vlx4vl	SME2	mopa	-	`kai_matmul_clamp_f32_qsi8d32p1vlx4_qai4c32p4vlx4_1vlx4vl_sme2_mopa`	LHS: `kai_lhs_quant_pack_qsi8d32pscalef32_f32_neon` RHS: `kai_rhs_pack_nxk_qai4c32ps1s0nrx4_qau4c32s1s0_f32_f32_f32_neon`, `kai_rhs_pack_nxk_qai4c32ps1s0nrx4_qau4c32s0s1_f32_f32_f32_neon`
f32	-	i8	i4	blockwise per dimension symmetric	blockwise symmetric	1vlx4vl	SME2	mopa	-	`kai_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa`	LHS: `kai_lhs_quant_pack_qsi8d32p_f32_neon` RHS: `kai_rhs_pack_nxk_qsi4c32ps1s0scalef16_qsu4c32s16s0_neon`
f32	-	i8	i4	blockwise per dimension symmetric	blockwise asymmetric	1x4vl	SME2	dot	-	`kai_matmul_clamp_f32_qsi8d32p1x4_qai4c32p4vlx4_1x4vl_sme2_dot`	LHS: `kai_lhs_quant_pack_qsi8d32pscalef32_f32_neon` RHS: `kai_rhs_pack_nxk_qai4c32ps1s0nrx4_qau4c32s1s0_f32_f32_f32_neon`, `kai_rhs_pack_nxk_qai4c32ps1s0nrx4_qau4c32s0s1_f32_f32_f32_neon`
f32	-	i8	i4	blockwise per dimension symmetric	blockwise asymmetric	1x4	Advanced SIMD	dotprod	-	`kai_matmul_clamp_f32_qsi8d32p1x4_qai4c32p4x4_1x4_neon_dotprod`	LHS: `kai_lhs_quant_pack_qsi8d32pscalef32_f32_neon` RHS: `kai_rhs_pack_nxk_qai4c32p_qau4c32s0s1_f32_f32_f32_neon`
f32	-	i8	i4	blockwise per dimension symmetric	blockwise symmetric	1x4vl	SME2	sdot	-	`kai_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot`	LHS: `kai_lhs_quant_pack_qsi8d32p_f32_neon` RHS: `kai_rhs_pack_nxk_qsi4c32ps1s0scalef16_qsu4c32s16s0_neon`
f32	-	i8	i4	blockwise per dimension symmetric	blockwise symmetric	1x4	Advanced SIMD	dotprod	-	`kai_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod`	LHS: `kai_lhs_quant_pack_qsi8d32p_f32` RHS: `kai_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0`
f32	-	i8	i4	blockwise per dimension symmetric	blockwise symmetric	1x8	SVE	dotprod	-	`kai_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p8x4_1x8_sve_dotprod`	LHS: `kai_lhs_quant_pack_qsi8d32p_f32_neon` RHS: `kai_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0`
f32	-	i8	i4	blockwise per dimension symmetric	blockwise asymmetric	1x4	Advanced SIMD	dotprod	-	`kai_matmul_clamp_f32_qsi8d32p1x8_qai4c32p4x8_1x4_neon_dotprod`	LHS: `kai_lhs_quant_pack_qsi8d32pscalef32_f32_neon` RHS: `kai_rhs_pack_nxk_qai4c32p_qau4c32s0s1_f32_f32_f32_neon`
f32	-	i8	i4	blockwise per dimension symmetric	blockwise symmetric	1x4x32	Advanced SIMD	dotprod	-	`kai_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod`	LHS: `kai_lhs_quant_pack_qsi8d32p_f32` RHS: `kai_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0`
f32	-	i8	i4	blockwise per dimension symmetric	blockwise symmetric	1x8	SVE	dotprod	-	`kai_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p8x8_1x8_sve_dotprod`	LHS: `kai_lhs_quant_pack_qsi8d32p_f32_neon` RHS: `kai_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0`
f32	-	i8	i4	blockwise per dimension symmetric	blockwise asymmetric	8x4	Advanced SIMD	dotprod	-	`kai_matmul_clamp_f32_qsi8d32p4x4_qai4c32p4x4_8x4_neon_dotprod`	LHS: `kai_lhs_quant_pack_qsi8d32pscalef32_f32_neon` RHS: `kai_rhs_pack_nxk_qai4c32p_qau4c32s0s1_f32_f32_f32_neon`
f32	-	i8	i4	blockwise per dimension symmetric	blockwise symmetric	16x4	Advanced SIMD	dotprod	-	`kai_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod`	LHS: `kai_lhs_quant_pack_qsi8d32p_f32` RHS: `kai_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0`
f32	-	i8	i4	blockwise per dimension symmetric	blockwise asymmetric	8x4	Advanced SIMD	i8mm	-	`kai_matmul_clamp_f32_qsi8d32p4x8_qai4c32p4x8_8x4_neon_i8mm`	LHS: `kai_lhs_quant_pack_qsi8d32pscalef32_f32_neon` RHS: `kai_rhs_pack_nxk_qai4c32p_qau4c32s0s1_f32_f32_f32_neon`
f32	-	i8	i4	blockwise per dimension symmetric	blockwise symmetric	16x4	Advanced SIMD	i8mm	-	`kai_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm`	LHS: `kai_lhs_quant_pack_qsi8d32p_f32` RHS: `kai_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0`
f32	-	i8	i4	blockwise per dimension symmetric	blockwise symmetric	8x4x32	Advanced SIMD	i8mm	-	`kai_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_8x4x32_neon_i8mm`	LHS: `kai_lhs_quant_pack_qsi8d32p_f32` RHS: `kai_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0`
f32	-	i8	i4	blockwise per dimension symmetric	blockwise symmetric	16x8	SVE	i8mm	-	`kai_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p8x8_16x8_sve_i8mm`	LHS: `kai_lhs_quant_pack_qsi8d32p_f32_neon` RHS: `kai_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0`
i8	tensorwise asymmetric	i8	i8	tensorwise asymmetric	channelwise symmetric	1x16vl	SME2	dot	-	`kai_matmul_clamp_qai8_qai8_qsi8cxp2vlx4sb_1x16vl_sme2_dot`	RHS: `kai_rhs_pack_kxn_qsi8cxp2vlx4sb_qs8cx_f32_i32_sme`
i8	tensorwise asymmetric	i8	i8	tensorwise asymmetric	channelwise symmetric	2vlx2vl	SME	mopa	-	`kai_matmul_clamp_qai8_qai8p2vlx4_qsi8cxp2vlx4sb_2vlx2vl_sme_mopa`	LHS: `kai_lhs_pack_x8p2vlx4_x8_sme` RHS: `kai_rhs_pack_kxn_qsi8cxp2vlx4sb_qs8cx_f32_i32_sme`
i8	tensorwise asymmetric	i8	i8	tensorwise asymmetric	channelwise symmetric	2vlx2vl	SME2	mopa	-	`kai_matmul_clamp_qai8_qai8p2vlx4_qsi8cxpsb2vlx4_2vlx2vl_sme2_mopa`	LHS: `kai_lhs_pack_x8p2vlx4_x8_sme` RHS: `kai_rhs_pack_kxn_qsi8cxp2vlx4sb_qs8cx_f32_i32_sme`

Indirect matmul micro-kernels

Output type	Output quantization	LHS type	RHS type	LHS quantization	RHS quantization	Block size	SIMD	Feature	Uarch	Micro-kernel	Packing micro-kernels
f16	-	f16	f16	-	-	2vlx2vl	SME2	mopa	-	`kai_imatmul_clamp_f16_f16p2vlx2_f16p2vlx2_2vlx2vl_sme2_mopa`	LHS: `kai_lhs_imatmul_pack_x16p2vlx2_x16p_sme` RHS: `kai_rhs_imatmul_pack_kxn_x16p2vlx2b_x16_x16_sme`
f16	-	f16	f16	-	-	2vlx2vl	SME	mopa	-	`kai_imatmul_clamp_f16_f16p2vlx2_f16p2vlx2b_2vlx2vl_sme_mopa`	LHS: `kai_lhs_imatmul_pack_x16p2vlx2_x16p_sme` RHS: `kai_rhs_imatmul_pack_kxn_x16p2vlx2b_x16_x16_sme`
f32	-	f32	f32	-	-	6x4vl	SVE	mla	-	`kai_imatmul_clamp_f32_f32_f32p4vlx1b_6x4vl_sve_mla`	RHS: `kai_rhs_imatmul_pack_kxn_x32p4vlx1b_x32_x32_sve`
f32	-	f32	f32	-	-	2vlx2vl	SME2	mopa	-	`kai_imatmul_clamp_f32_f32p2vlx1_f32p2vlx1b_2vlx2vl_sme2_mopa`	LHS: `kai_lhs_imatmul_pack_x32p2vlx1_x32p_sme` RHS: `kai_rhs_imatmul_pack_kxn_x32p2vlx1b_x32_x32_sme`
f32	-	f32	f32	-	-	2vlx2vl	SME	mopa	-	`kai_imatmul_clamp_f32_f32p2vlx1_f32p2vlx1b_2vlx2vl_sme_mopa`	LHS: `kai_lhs_imatmul_pack_x32p2vlx1_x32p_sme` RHS: `kai_rhs_imatmul_pack_kxn_x32p2vlx1b_x32_x32_sme`
i8	tensorwise asymmetric	i8	i8	tensorwise asymmetric	channelwise symmetric	2vlx2vl	SME	mopa	-	`kai_imatmul_clamp_qai8_qai8p2vlx4_qsi8cxp2vlx4sb_2vlx2vl_sme_mopa`	LHS: `kai_lhs_imatmul_pack_x8p2vlx4_x8p_sme` RHS: `kai_rhs_imatmul_pack_kxn_qsi8cxp2vlx4sb_qs8cx_f32_i32_sme`
i8	tensorwise asymmetric	i8	i8	tensorwise asymmetric	channelwise symmetric	2vlx2vl	SME2	mopa	-	`kai_imatmul_clamp_qai8_qai8p2vlx4_qsi8cxpsb2vlx4_2vlx2vl_sme2_mopa`	LHS: `kai_lhs_imatmul_pack_x8p2vlx4_x8p_sme` RHS: `kai_rhs_imatmul_pack_kxn_qsi8cxp2vlx4sb_qs8cx_f32_i32_sme`

Depthwise convolution micro-kernels

Output type	LHS type	RHS type	LHS quantization	RHS quantization	Filter size	Block size	SIMD	Feature	Uarch	Micro-kernel	Packing micro-kernels
f32	f32	f32	-	-	3x3 stride 1	4 rows, planar	SME2	mla	-	`kai_dwconv_clamp_f32_f32_f32p1vlx1b_3x3_s1_4xc_sme2_mla`	RHS: `kai_rhs_dwconv_pack_x32p1vlx1b_x32_x32_sme`

Packing micro-kernels

Output type	Output quantization	Input type	Input quantization	Bias type	Scale type	Zero type	SIMD	Micro-kernel
bf16	-	f16	-	f16	-	-	Advanced SIMD	`kai_rhs_pack_kxn_bf16p12x4biasf16_f16_neon`
bf16	-	f16	-	f32	-	-	Advanced SIMD	`kai_rhs_pack_kxn_bf16p12x4biasf32_f16_neon`
bf16	-	f32	-	f32	-	-	Advanced SIMD	`kai_rhs_quant_pack_kxn_bf16p12x4biasf32_f32_neon`
bf16	-	f32	-	-	-	-	Advanced SIMD	`kai_lhs_quant_pack_bf16p1x4_f32_neon`
bf16	-	f32	-	-	-	-	SME	`kai_lhs_pack_bf16p2vlx2_f32_sme`
bf16	-	f32	-	x32	-	-	SME	`kai_rhs_pack_kxn_bf16p2vlx2b_f32_x32_sme`
bf16	-	f16	-	-	-	-	Advanced SIMD	`kai_lhs_pack_bf16p8x4_f16_neon`
bf16	-	f32	-	-	-	-	Advanced SIMD	`kai_lhs_quant_pack_bf16p8x4_f32_neon`
f16	-	f16	-	f16	-	-	Advanced SIMD	`kai_rhs_pack_kxn_f16p16x1biasf16_f16_f16_neon`
f16	-	f32	-	-	-	-	Advanced SIMD	`kai_lhs_pack_f16pmrx2_f32_neon`
f32	-	f32	-	f32	-	-	SME	`kai_rhs_pack_kxn_f32p16vlx1b_f32_f32_sme`
f32	-	f32	-	-	-	-	SME	`kai_lhs_pack_f32p2vlx1_f32_sme`
f32	-	f32	-	f32	-	-	SME	`kai_rhs_pack_kxn_f32p2vlx1biasf32_f32_f32_sme`
f32	-	f32	-	f32	-	-	SME	`kai_rhs_pack_nxk_f32p2vlx1biasf32_f32_f32_sme`
f32	-	f32	-	f32	-	-	Advanced SIMD	`kai_rhs_pack_kxn_f32p8x1biasf32_f32_f32_neon`
i4	blockwise asymmetric	u4	blockwise asymmetric	f32	f32	f32	Advanced SIMD	`kai_rhs_pack_nxk_qai4c32p_qau4c32s0s1_f32_f32_f32_neon`
i4	blockwise asymmetric	u4	blockwise asymmetric	f32	f32	f32	Advanced SIMD	`kai_rhs_pack_nxk_qai4c32ps1s0nrx4_qau4c32s0s1_f32_f32_f32_neon`
i4	blockwise asymmetric	u4	blockwise asymmetric	f32	f32	f32	Advanced SIMD	`kai_rhs_pack_nxk_qai4c32ps1s0nrx4_qau4c32s1s0_f32_f32_f32_neon`
i8	dimensionwise asymmetric	bf16	-	-	-	-	Advanced SIMD	`kai_lhs_quant_pack_qai8dxp_bf16_neon`
i8	dimensionwise asymmetric	f16	-	-	-	-	Advanced SIMD	`kai_lhs_quant_pack_qai8dxp_f16_neon`
i8	dimensionwise asymmetric	f32	-	-	-	-	-	`kai_lhs_quant_pack_qai8dxp_f32`
i4	blockwise symmetric	u4	blockwise symmetric	-	bf16	-	-	`kai_rhs_pack_kxn_qsi4c32p_qsu4c32s1s0`
i4	blockwise symmetric	u4	blockwise symmetric	-	bf16	-	-	`kai_rhs_pack_nxk_qsi4c32p_qsu4c32s1s0`
i4	blockwise symmetric	u4	blockwise symmetric	f32	bf16	-	Advanced SIMD	`kai_rhs_pack_nxk_qsi4c32pnrx4_qsu4c32s1s0_neon`
i4	blockwise symmetric	u4	blockwise symmetric	f32	bf16	-	Advanced SIMD	`kai_rhs_pack_nxk_qsi4c32pnrx8_qsu4c32s1s0_neon`
i4	blockwise symmetric	u4	blockwise symmetric	f32	bf16	-	Advanced SIMD	`kai_rhs_pack_kxn_qsi4c32ps1s0nrx4_qsu4c32s1s0_neon`
i4	blockwise symmetric	u4	blockwise symmetric	f32	bf16	-	Advanced SIMD	`kai_rhs_pack_nxk_qsi4c32ps1s0nrx4_qsu4c32s1s0_neon`
i4	blockwise symmetric	u4	blockwise symmetric	-	f16	-	Advanced SIMD	`kai_rhs_pack_nxk_qsi4c32ps1s0scalef16_qsu4c32s16s0_neon`
i4	blockwise symmetric	u4	blockwise symmetric	-	f16	-	-	`kai_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0`
i4	channelwise symmetric	i4	channelwise symmetric	-	f32	-	-	`kai_rhs_pack_kxn_qsi4cxp_qs4cxs1s0`
i4	channelwise symmetric	i4	channelwise symmetric	-	f32	-	-	`kai_rhs_pack_nxk_qsi4cxp_qs4cxs1s0`
i4	channelwise symmetric	u4	channelwise symmetric	f32	f32	-	Advanced SIMD	`kai_rhs_pack_nxk_qsi4cxps1s0_qsu4cxs1s0_neon`
i8	channelwise symmetric	i8	channelwise symmetric	f32	f32	-	Advanced SIMD	`kai_rhs_pack_kxn_qsi8cxp_qsi8cx_neon`
i8	channelwise symmetric	i8	channelwise symmetric	f32	f32	-	Advanced SIMD	`kai_rhs_pack_nxk_qsi8cxp_qsi8cx_neon`
i8	channelwise symmetric	i8	channelwise symmetric	i32	f32	-	SME	`kai_rhs_imatmul_pack_kxn_qsi8cxp2vlx4sb_qs8cx_f32_i32_sme`
i8	channelwise symmetric	i8	channelwise symmetric	i32	f32	-	SME	`kai_rhs_pack_kxn_qsi8cxp2vlx4sb_qs8cx_f32_i32_sme`
i8	blockwise per dimension symmetric	f32	-	-	-	-	-	`kai_lhs_quant_pack_qsi8d32p_f32`
i8	blockwise per dimension symmetric	f32	-	-	-	-	Advanced SIMD	`kai_lhs_quant_pack_qsi8d32p_f32_neon`
i8	blockwise per dimension symmetric	f32	-	-	-	-	Advanced SIMD	`kai_lhs_quant_pack_qsi8d32p4x8sb_f32_neon`
i8	blockwise per dimension symmetric	f16	-	-	-	-	Advanced SIMD	`kai_lhs_quant_pack_qsi8d32pscalef32_f16_neon`
i8	blockwise per dimension symmetric	f32	-	-	-	-	Advanced SIMD	`kai_lhs_quant_pack_qsi8d32pscalef32_f32_neon`
u2	channelwise symmetric	u2	channelwise symmetric	f32	f32	-	Advanced SIMD	`kai_rhs_pack_nxk_qsu2cxp4vlx4_qsu2cx_neon`
x16	-	x16	-	-	-	-	SME	`kai_lhs_pack_x16p2vlx2_x16_sme`
x16	-	x16	-	-	-	-	SME	`kai_lhs_imatmul_pack_x16p2vlx2_x16p_sme`
x16	-	x16	-	x16	-	-	SME	`kai_rhs_imatmul_pack_kxn_x16p2vlx2b_x16_x16_sme`
x16	-	x16	-	x16	-	-	SME	`kai_rhs_pack_kxn_x16p2vlx2b_x16_x16_sme`
x16	-	x16	-	x16	-	-	SME	`kai_rhs_pack_nxk_x16p2vlx2b_x16_x16_sme`
x16	-	x16	-	x16	-	-	Advanced SIMD	`kai_rhs_pack_kxn_x16p32x1b_x16_x16_neon`
x32	-	x32	-	x32	-	-	Advanced SIMD	`kai_rhs_pack_kxn_x32p16x1b_x32_x32_neon`
x32	-	x32	-	x32	-	-	SME	`kai_rhs_dwconv_pack_x32p1vlx1b_x32_x32_sme`
x32	-	x32	-	-	-	-	SME	`kai_matmul_pack_lhs_mxk_x32p4vsx1_x32_sme`
x32	-	x32	-	x32	-	-	SME	`kai_matmul_pack_rhs_kxn_x32p4vsx1bx32_x32_x32_sme`
x32	-	x32	-	x32	-	-	SME	`kai_matmul_pack_rhs_nxk_x32p4vsx1bx32_x32_x32_sme`
x32	-	x32	-	-	-	-	SME	`kai_lhs_imatmul_pack_x32p2vlx1_x32p_sme`
x32	-	x32	-	x32	-	-	SME	`kai_rhs_imatmul_pack_kxn_x32p2vlx1b_x32_x32_sme`
x32	-	x32	-	x32	-	-	SVE	`kai_rhs_imatmul_pack_kxn_x32p4vlx1b_x32_x32_sve`
x32	-	x32	-	x32	-	-	SVE	`kai_rhs_pack_kxn_x32p4vlx1b_x32_x32_sve`
x8	-	x8	-	-	-	-	SME	`kai_lhs_pack_x8p2vlx4_x8_sme`
x8	-	x8	-	-	-	-	SME	`kai_lhs_imatmul_pack_x8p2vlx4_x8p_sme`

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

KleidiAI micro-kernel tables

Matmul micro-kernels

Indirect matmul micro-kernels

Depthwise convolution micro-kernels

Packing micro-kernels

FilesExpand file tree

microkernel_tables.md

Latest commit

History

microkernel_tables.md

File metadata and controls

KleidiAI micro-kernel tables

Matmul micro-kernels

Indirect matmul micro-kernels

Depthwise convolution micro-kernels

Packing micro-kernels