diff --git a/kernel/arm64/sgemm_kernel_sve_v2x8.S b/kernel/arm64/sgemm_kernel_sve_v2x8.S
new file mode 100644
index 000000000..1cdd8253e
--- /dev/null
+++ b/kernel/arm64/sgemm_kernel_sve_v2x8.S
@@ -0,0 +1,1683 @@
+/*******************************************************************************
+Copyright (c) 2015, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*******************************************************************************/
+
+/* This is an SVE sgemm kernel with size 2*SVE_LEN x 8.
+However, the data layout is the same as for the kernel 1*SVE_LEN x 8.
+This means that we sweep two panels of packed A when iterating in a loop over K.
+With this approach, we can reuse sgemm_n|tcopy_sve_v1.c packing functions. */
+
+#define ASSEMBLER
+#include "common.h"
+
+/*                   X0          X1          X2          s0         X3        x4       x5           x6 */
+/*int CNAME(BLASLONG bm,BLASLONG bn,BLASLONG bk,FLOAT alpha0,FLOAT* ba,FLOAT* bb,FLOAT* C,BLASLONG ldc )*/
+
+#define origM		x0
+#define origN		x1
+#define origK		x2
+#define origPA		x3
+#define origPB		x4
+#define pC		x5
+#define LDC		x6
+#define temp		x7
+#define counterL	x8
+#define counterI	x9
+#define counterJ	x10
+#define pB		x11
+#define pCRow0		x12
+#define pCRow1		x13
+#define pCRow2		x14
+
+#define lanes		x15
+#define pA1	    	x16
+#define pA2	    	x17
+#define alpha		w18
+#define vec_len		x19
+#define vec_lenx2   x20
+
+#define alpha0		s10
+#define alphaZ		z7.s
+
+#define A_PRE_SIZE	1536
+#define B_PRE_SIZE	512
+#define C_PRE_SIZE	128
+
+// 00 origM
+// 01 origN
+// 02 origK
+// 03 origPA
+// 04 origPB
+// 05 pC
+// 06 origLDC -> LDC
+// 07 temp
+// 08 counterL
+// 09 counterI
+// 10 counterJ
+// 11 pB
+// 12 pCRow0
+// 13 pCRow1
+// 14 pCRow2
+// 15 lanes
+// 16 pA1
+// 17 pA1
+// 18 must save alpha
+// 19 must save vec_len
+// 20 must save
+// 21 must save
+// 22 must save
+// 23 must save
+// 24 must save
+// 25 must save
+// 26 must save
+// 27 must save
+// 28 must save
+// 29 frame
+// 30 link
+// 31 sp
+
+//v00 ALPHA -> pA10_0
+//v01 pA10_1
+//v02 pA20_0
+//v03 pA20_1
+//v04 
+//v05 
+//v06 
+//v07 ALPHA0
+//v08 must save pB0_0
+//v09 must save pB0_1
+//v10 must save pB0_2 
+//v11 must save pB0_3
+//v12 must save pB0_4
+//v13 must save pB0_5
+//v14 must save pB0_6
+//v15 must save pB0_7
+//v16 must save C0
+//v17 must save C1
+//v18 must save C2
+//v19 must save C3
+//v20 must save C4
+//v21 must save C5
+//v22 must save C6
+//v23 must save C7
+//v24 must save C8
+//v25 must save C9
+//v26 must save C10
+//v27 must save C11
+//v28 must save C12
+//v29 must save C13
+//v30 must save C14
+//v31 must save C15
+
+/*******************************************************************************
+* Macro definitions
+*******************************************************************************/
+
+.macro INITv2x8
+    dup         z16.s, #0
+    dup         z17.s, #0
+    dup         z18.s, #0
+    dup         z19.s, #0
+    dup         z20.s, #0
+    dup         z21.s, #0
+    dup         z22.s, #0
+    dup         z23.s, #0
+    dup         z24.s, #0
+    dup         z25.s, #0
+    dup         z26.s, #0
+    dup         z27.s, #0
+    dup         z28.s, #0
+    dup         z29.s, #0
+    dup         z30.s, #0
+    dup         z31.s, #0
+.endm
+
+.macro KERNELv2x8_I
+    ld1w  z0.s, p0/z, [pA1] 
+    ld1w  z1.s, p0/z, [pA2]   
+    ld1w  z2.s, p0/z, [pA1, vec_len, lsl #2] 
+    ld1w  z3.s, p0/z, [pA2, vec_len, lsl #2]   
+	add	pA1, pA1, vec_len, lsl #3	// pA1 = pA1 + vec_len * 4 *2
+	add	pA2, pA2, vec_len, lsl #3	// pA1 = pA1 + vec_len * 4 *2
+
+
+    ld1rw  z8.s, p0/z,  [pB]
+    ld1rw  z9.s, p0/z,  [pB, 4]
+    ld1rw  z10.s, p0/z, [pB, 8]
+    ld1rw  z11.s, p0/z, [pB, 12]
+    ld1rw  z12.s, p0/z, [pB, 16]
+    ld1rw  z13.s, p0/z, [pB, 20]
+    ld1rw  z14.s, p0/z, [pB, 24]
+    ld1rw  z15.s, p0/z, [pB, 28]
+
+    add pB, pB, 32
+
+    fmla z16.s, p0/m, z0.s, z8.s
+    fmla z17.s, p0/m, z1.s, z8.s
+    ld1rw  z8.s, p0/z,  [pB]
+    fmla z18.s, p0/m, z0.s, z9.s
+    fmla z19.s, p0/m, z1.s, z9.s
+    ld1rw  z9.s, p0/z,  [pB, 4]
+    fmla z20.s, p0/m, z0.s, z10.s
+    fmla z21.s, p0/m, z1.s, z10.s
+    ld1rw  z10.s, p0/z, [pB, 8]
+    fmla z22.s, p0/m, z0.s, z11.s
+    fmla z23.s, p0/m, z1.s, z11.s
+    ld1rw  z11.s, p0/z, [pB, 12]
+    fmla z24.s, p0/m, z0.s, z12.s
+    fmla z25.s, p0/m, z1.s, z12.s
+	prfm	PLDL1KEEP, [pA1, #A_PRE_SIZE]
+    ld1rw  z12.s, p0/z, [pB, 16]
+    fmla z26.s, p0/m, z0.s, z13.s
+    fmla z27.s, p0/m, z1.s, z13.s
+	prfm	PLDL1KEEP, [pA2, #A_PRE_SIZE]
+    ld1rw  z13.s, p0/z, [pB, 20]
+    fmla z28.s, p0/m, z0.s, z14.s
+    fmla z29.s, p0/m, z1.s, z14.s
+    ld1rw  z14.s, p0/z, [pB, 24]
+    fmla z30.s, p0/m, z0.s, z15.s
+    fmla z31.s, p0/m, z1.s, z15.s
+	prfm	PLDL1KEEP, [pA1, #A_PRE_SIZE+64]
+    ld1rw  z15.s, p0/z, [pB, 28]
+	prfm	PLDL1KEEP, [pA2, #A_PRE_SIZE+64]
+
+    add pB, pB, 32
+.endm
+
+.macro KERNELv2x8_M1
+    ld1w  z2.s, p0/z, [pA1] 
+    ld1w  z3.s, p0/z, [pA2]   
+	add	pA1, pA1, vec_len, lsl #2	// pA1 = pA1 + vec_len * 4
+	add	pA2, pA2, vec_len, lsl #2	// pA1 = pA1 + vec_len * 4
+
+    fmla z16.s, p0/m, z0.s, z8.s
+    fmla z17.s, p0/m, z1.s, z8.s
+    ld1rw  z8.s, p0/z,  [pB]
+    fmla z18.s, p0/m, z0.s, z9.s
+    fmla z19.s, p0/m, z1.s, z9.s
+    ld1rw  z9.s, p0/z,  [pB, 4]
+    fmla z20.s, p0/m, z0.s, z10.s
+    fmla z21.s, p0/m, z1.s, z10.s
+    ld1rw  z10.s, p0/z, [pB, 8]
+    fmla z22.s, p0/m, z0.s, z11.s
+    fmla z23.s, p0/m, z1.s, z11.s
+    ld1rw  z11.s, p0/z, [pB, 12]
+    fmla z24.s, p0/m, z0.s, z12.s
+    fmla z25.s, p0/m, z1.s, z12.s
+	prfm	PLDL1KEEP, [pA1, #A_PRE_SIZE]
+    ld1rw  z12.s, p0/z, [pB, 16]
+    fmla z26.s, p0/m, z0.s, z13.s
+    fmla z27.s, p0/m, z1.s, z13.s
+	prfm	PLDL1KEEP, [pA2, #A_PRE_SIZE]
+    ld1rw  z13.s, p0/z, [pB, 20]
+    fmla z28.s, p0/m, z0.s, z14.s
+    fmla z29.s, p0/m, z1.s, z14.s
+    ld1rw  z14.s, p0/z, [pB, 24]
+    fmla z30.s, p0/m, z0.s, z15.s
+	prfm	PLDL1KEEP, [pA1, #A_PRE_SIZE+64]
+    fmla z31.s, p0/m, z1.s, z15.s
+	prfm	PLDL1KEEP, [pA2, #A_PRE_SIZE+64]
+    ld1rw  z15.s, p0/z, [pB, 28]
+
+    add pB, pB, 32
+.endm
+
+.macro KERNELv2x8_M2
+    ld1w  z0.s, p0/z, [pA1] 
+    ld1w  z1.s, p0/z, [pA2]   
+	add	pA1, pA1, vec_len, lsl #2	// pA1 = pA1 + vec_len * 2 * 4
+	add	pA2, pA2, vec_len, lsl #2	// pA1 = pA1 + vec_len * 2 * 4
+
+    fmla z16.s, p0/m, z2.s, z8.s
+    fmla z17.s, p0/m, z3.s, z8.s 
+    ld1rw  z8.s, p0/z,  [pB]
+    fmla z18.s, p0/m, z2.s, z9.s
+    fmla z19.s, p0/m, z3.s, z9.s
+    ld1rw  z9.s, p0/z,  [pB, 4]
+    fmla z20.s, p0/m, z2.s, z10.s
+    fmla z21.s, p0/m, z3.s, z10.s
+    ld1rw  z10.s, p0/z, [pB, 8]
+    fmla z22.s, p0/m, z2.s, z11.s
+    fmla z23.s, p0/m, z3.s, z11.s
+    ld1rw  z11.s, p0/z, [pB, 12]
+    fmla z24.s, p0/m, z2.s, z12.s
+    fmla z25.s, p0/m, z3.s, z12.s
+	prfm	PLDL1KEEP, [pB, #B_PRE_SIZE]
+    ld1rw  z12.s, p0/z, [pB, 16]
+    fmla z26.s, p0/m, z2.s, z13.s
+    fmla z27.s, p0/m, z3.s, z13.s
+    ld1rw  z13.s, p0/z, [pB, 20]
+    fmla z28.s, p0/m, z2.s, z14.s
+    fmla z29.s, p0/m, z3.s, z14.s
+    ld1rw  z14.s, p0/z, [pB, 24]
+    fmla z30.s, p0/m, z2.s, z15.s
+    fmla z31.s, p0/m, z3.s, z15.s
+    ld1rw  z15.s, p0/z, [pB, 28]
+
+    add pB, pB, 32
+.endm
+
+.macro KERNELv2x8_E
+    fmla z16.s, p0/m, z2.s, z8.s
+    fmla z17.s, p0/m, z3.s, z8.s
+    fmla z18.s, p0/m, z2.s, z9.s
+    fmla z19.s, p0/m, z3.s, z9.s
+    fmla z20.s, p0/m, z2.s, z10.s
+    fmla z21.s, p0/m, z3.s, z10.s
+    fmla z22.s, p0/m, z2.s, z11.s
+    fmla z23.s, p0/m, z3.s, z11.s
+    fmla z24.s, p0/m, z2.s, z12.s
+    fmla z25.s, p0/m, z3.s, z12.s
+	prfm	PLDL1KEEP, [pB, #B_PRE_SIZE]
+    fmla z26.s, p0/m, z2.s, z13.s
+    fmla z27.s, p0/m, z3.s, z13.s
+    fmla z28.s, p0/m, z2.s, z14.s
+    fmla z29.s, p0/m, z3.s, z14.s
+    fmla z30.s, p0/m, z2.s, z15.s
+    fmla z31.s, p0/m, z3.s, z15.s
+.endm
+
+.macro KERNELv2x8_SUB
+    ld1w  z0.s, p0/z, [pA1] 
+    ld1w  z1.s, p0/z, [pA2]   
+	add	pA1, pA1, vec_len, lsl #2	// pA1 = pA1 + vec_len * 4
+	add	pA2, pA2, vec_len, lsl #2	// pA1 = pA1 + vec_len * 4
+
+    ld1rw  z8.s, p0/z,  [pB]
+    ld1rw  z9.s, p0/z,  [pB, 4]
+    ld1rw  z10.s, p0/z, [pB, 8]
+    ld1rw  z11.s, p0/z, [pB, 12]
+    ld1rw  z12.s, p0/z, [pB, 16]
+    ld1rw  z13.s, p0/z, [pB, 20]
+    ld1rw  z14.s, p0/z, [pB, 24]
+    ld1rw  z15.s, p0/z, [pB, 28]
+
+    add pB, pB, 32
+
+    fmla z16.s, p0/m, z0.s, z8.s
+    fmla z17.s, p0/m, z1.s, z8.s
+    fmla z18.s, p0/m, z0.s, z9.s
+    fmla z19.s, p0/m, z1.s, z9.s
+    fmla z20.s, p0/m, z0.s, z10.s
+	prfm	PLDL1KEEP, [pA1, #A_PRE_SIZE]
+    fmla z21.s, p0/m, z1.s, z10.s
+    fmla z22.s, p0/m, z0.s, z11.s
+    fmla z23.s, p0/m, z1.s, z11.s
+    fmla z24.s, p0/m, z0.s, z12.s
+	prfm	PLDL1KEEP, [pA2, #A_PRE_SIZE]
+    fmla z25.s, p0/m, z1.s, z12.s
+    fmla z26.s, p0/m, z0.s, z13.s
+    fmla z27.s, p0/m, z1.s, z13.s
+    fmla z28.s, p0/m, z0.s, z14.s
+    fmla z29.s, p0/m, z1.s, z14.s
+	prfm	PLDL1KEEP, [pB, #B_PRE_SIZE]
+    fmla z30.s, p0/m, z0.s, z15.s
+    fmla z31.s, p0/m, z1.s, z15.s
+.endm
+
+.macro SAVEv2x8
+
+	prfm	PLDL2KEEP, [pCRow0, #C_PRE_SIZE]
+
+	add	pCRow1, pCRow0, LDC
+    ld1w  z8.s, p0/z, [pCRow0] 
+    ld1w  z9.s, p0/z, [pCRow0, #1, mul vl] 
+    fmla z8.s, p0/m, z16.s, alphaZ
+    fmla z9.s, p0/m, z17.s, alphaZ
+    st1w  z8.s, p0, [pCRow0]
+    st1w  z9.s, p0, [pCRow0, #1, mul vl]
+	prfm	PLDL2KEEP, [pCRow1, #C_PRE_SIZE]
+
+	add	pCRow2, pCRow1, LDC
+    ld1w  z10.s, p0/z, [pCRow1] 
+    ld1w  z11.s, p0/z, [pCRow1, #1, mul vl] 
+    fmla z10.s, p0/m, z18.s, alphaZ
+    fmla z11.s, p0/m, z19.s, alphaZ
+    st1w  z10.s, p0, [pCRow1]
+    st1w  z11.s, p0, [pCRow1, #1, mul vl]
+	prfm	PLDL2KEEP, [pCRow2, #C_PRE_SIZE]
+
+	add	pCRow1, pCRow2, LDC
+    ld1w  z12.s, p0/z, [pCRow2] 
+    ld1w  z13.s, p0/z, [pCRow2, #1, mul vl] 
+    fmla z12.s, p0/m, z20.s, alphaZ
+    fmla z13.s, p0/m, z21.s, alphaZ
+    st1w  z12.s, p0, [pCRow2]
+    st1w  z13.s, p0, [pCRow2, #1, mul vl]
+	prfm	PLDL2KEEP, [pCRow1, #C_PRE_SIZE]
+
+	add	pCRow2, pCRow1, LDC
+    ld1w  z14.s, p0/z, [pCRow1] 
+    ld1w  z15.s, p0/z, [pCRow1, #1, mul vl] 
+    fmla z14.s, p0/m, z22.s, alphaZ
+    fmla z15.s, p0/m, z23.s, alphaZ
+    st1w  z14.s, p0, [pCRow1]
+    st1w  z15.s, p0, [pCRow1, #1, mul vl]
+	prfm	PLDL2KEEP, [pCRow2, #C_PRE_SIZE]
+
+	add	pCRow1, pCRow2, LDC
+    ld1w  z8.s, p0/z, [pCRow2] 
+    ld1w  z9.s, p0/z, [pCRow2, #1, mul vl] 
+    fmla z8.s, p0/m, z24.s, alphaZ
+    fmla z9.s, p0/m, z25.s, alphaZ
+    st1w  z8.s, p0, [pCRow2]
+    st1w  z9.s, p0, [pCRow2, #1, mul vl]
+	prfm	PLDL2KEEP, [pCRow1, #C_PRE_SIZE]
+
+	add	pCRow2, pCRow1, LDC
+    ld1w  z10.s, p0/z, [pCRow1] 
+    ld1w  z11.s, p0/z, [pCRow1, #1, mul vl] 
+    fmla z10.s, p0/m, z26.s, alphaZ
+    fmla z11.s, p0/m, z27.s, alphaZ
+    st1w  z10.s, p0, [pCRow1]
+    st1w  z11.s, p0, [pCRow1, #1, mul vl]
+	prfm	PLDL2KEEP, [pCRow2, #C_PRE_SIZE]
+
+	add	pCRow1, pCRow2, LDC
+    ld1w  z12.s, p0/z, [pCRow2] 
+    ld1w  z13.s, p0/z, [pCRow2, #1, mul vl] 
+    fmla z12.s, p0/m, z28.s, alphaZ
+    fmla z13.s, p0/m, z29.s, alphaZ
+    st1w  z12.s, p0, [pCRow2]
+    st1w  z13.s, p0, [pCRow2, #1, mul vl]
+	prfm	PLDL2KEEP, [pCRow1, #C_PRE_SIZE]
+
+    ld1w  z14.s, p0/z, [pCRow1] 
+    ld1w  z15.s, p0/z, [pCRow1, #1, mul vl] 
+    fmla z14.s, p0/m, z30.s, alphaZ
+    fmla z15.s, p0/m, z31.s, alphaZ
+    st1w  z14.s, p0, [pCRow1]
+    st1w  z15.s, p0, [pCRow1, #1, mul vl]
+
+	add	pCRow0, pCRow0, vec_len, lsl #3	// pC = pC + vec_len  * 4 * 2
+
+.endm
+
+.macro INITv2x4
+    dup         z16.s, #0
+    dup         z17.s, #0
+    dup         z18.s, #0
+    dup         z19.s, #0
+    dup         z20.s, #0
+    dup         z21.s, #0
+    dup         z22.s, #0
+    dup         z23.s, #0
+.endm
+
+.macro KERNELv2x4_SUB
+    ld1w  z0.s, p0/z, [pA1] 
+    ld1w  z1.s, p0/z, [pA2]   
+	add	pA1, pA1, vec_len, lsl #2	// pA1 = pA1 + vec_len * 4
+	add	pA2, pA2, vec_len, lsl #2	// pA1 = pA1 + vec_len * 4
+
+    ld1rw  z8.s, p0/z,  [pB]
+    ld1rw  z9.s, p0/z,  [pB, 4]
+    ld1rw  z10.s, p0/z, [pB, 8]
+    ld1rw  z11.s, p0/z, [pB, 12]
+
+    add pB, pB, 16
+
+    fmla z16.s, p0/m, z0.s, z8.s
+    fmla z17.s, p0/m, z1.s, z8.s
+    fmla z18.s, p0/m, z0.s, z9.s
+	prfm	PLDL1KEEP, [pA1, #A_PRE_SIZE]
+    fmla z19.s, p0/m, z1.s, z9.s
+    fmla z20.s, p0/m, z0.s, z10.s
+	prfm	PLDL1KEEP, [pA2, #A_PRE_SIZE]
+    fmla z21.s, p0/m, z1.s, z10.s
+    fmla z22.s, p0/m, z0.s, z11.s
+    fmla z23.s, p0/m, z1.s, z11.s
+.endm
+
+.macro SAVEv2x4
+
+	prfm	PLDL2KEEP, [pCRow0, #C_PRE_SIZE]
+
+	add	pCRow1, pCRow0, LDC
+    ld1w  z8.s, p0/z, [pCRow0] 
+    ld1w  z9.s, p0/z, [pCRow0, #1, mul vl] 
+    fmla z8.s, p0/m, z16.s, alphaZ
+    fmla z9.s, p0/m, z17.s, alphaZ
+    st1w  z8.s, p0, [pCRow0]
+    st1w  z9.s, p0, [pCRow0, #1, mul vl]
+	prfm	PLDL2KEEP, [pCRow1, #C_PRE_SIZE]
+
+	add	pCRow2, pCRow1, LDC
+    ld1w  z10.s, p0/z, [pCRow1] 
+    ld1w  z11.s, p0/z, [pCRow1, #1, mul vl] 
+    fmla z10.s, p0/m, z18.s, alphaZ
+    fmla z11.s, p0/m, z19.s, alphaZ
+    st1w  z10.s, p0, [pCRow1]
+    st1w  z11.s, p0, [pCRow1, #1, mul vl]
+	prfm	PLDL2KEEP, [pCRow2, #C_PRE_SIZE]
+
+	add	pCRow1, pCRow2, LDC
+    ld1w  z12.s, p0/z, [pCRow2] 
+    ld1w  z13.s, p0/z, [pCRow2, #1, mul vl] 
+    fmla z12.s, p0/m, z20.s, alphaZ
+    fmla z13.s, p0/m, z21.s, alphaZ
+    st1w  z12.s, p0, [pCRow2]
+    st1w  z13.s, p0, [pCRow2, #1, mul vl]
+	prfm	PLDL2KEEP, [pCRow1, #C_PRE_SIZE]
+
+    ld1w  z14.s, p0/z, [pCRow1] 
+    ld1w  z15.s, p0/z, [pCRow1, #1, mul vl] 
+    fmla z14.s, p0/m, z22.s, alphaZ
+    fmla z15.s, p0/m, z23.s, alphaZ
+    st1w  z14.s, p0, [pCRow1]
+    st1w  z15.s, p0, [pCRow1, #1, mul vl]
+
+	add	pCRow0, pCRow0, vec_len, lsl #3	// pC = pC + vec_len  * 4 * 2
+
+.endm
+
+.macro INITv2x2
+    dup         z16.s, #0
+    dup         z17.s, #0
+    dup         z18.s, #0
+    dup         z19.s, #0
+.endm
+
+.macro KERNELv2x2_SUB
+    ld1w  z0.s, p0/z, [pA1] 
+    ld1w  z1.s, p0/z, [pA2]   
+	add	pA1, pA1, vec_len, lsl #2	// pA1 = pA1 + vec_len * 4
+	add	pA2, pA2, vec_len, lsl #2	// pA1 = pA1 + vec_len * 4
+
+    ld1rw  z8.s, p0/z,  [pB]
+    ld1rw  z9.s, p0/z,  [pB, 4]
+
+    add pB, pB, 8
+
+    fmla z16.s, p0/m, z0.s, z8.s
+    fmla z17.s, p0/m, z1.s, z8.s
+	prfm	PLDL1KEEP, [pA1, #A_PRE_SIZE]
+    fmla z18.s, p0/m, z0.s, z9.s
+    fmla z19.s, p0/m, z1.s, z9.s
+	prfm	PLDL1KEEP, [pA2, #A_PRE_SIZE]
+.endm
+
+.macro SAVEv2x2
+
+	prfm	PLDL2KEEP, [pCRow0, #C_PRE_SIZE]
+
+	add	pCRow1, pCRow0, LDC
+    ld1w  z8.s, p0/z, [pCRow0] 
+    ld1w  z9.s, p0/z, [pCRow0, #1, mul vl] 
+    fmla z8.s, p0/m, z16.s, alphaZ
+    fmla z9.s, p0/m, z17.s, alphaZ
+    st1w  z8.s, p0, [pCRow0]
+    st1w  z9.s, p0, [pCRow0, #1, mul vl]
+	prfm	PLDL2KEEP, [pCRow1, #C_PRE_SIZE]
+
+    ld1w  z10.s, p0/z, [pCRow1] 
+    ld1w  z11.s, p0/z, [pCRow1, #1, mul vl] 
+    fmla z10.s, p0/m, z18.s, alphaZ
+    fmla z11.s, p0/m, z19.s, alphaZ
+    st1w  z10.s, p0, [pCRow1]
+    st1w  z11.s, p0, [pCRow1, #1, mul vl]
+	prfm	PLDL2KEEP, [pCRow2, #C_PRE_SIZE]
+
+
+
+	add	pCRow0, pCRow0, vec_len, lsl #3	// pC = pC + vec_len  * 4 * 2
+.endm
+
+.macro INITv2x1
+    dup         z16.s, #0
+    dup         z17.s, #0
+.endm
+
+.macro KERNELv2x1_SUB
+    ld1w  z0.s, p0/z, [pA1] 
+    ld1w  z1.s, p0/z, [pA2]   
+	add	pA1, pA1, vec_len, lsl #2	// pA1 = pA1 + vec_len * 4
+	add	pA2, pA2, vec_len, lsl #2	// pA1 = pA1 + vec_len * 4
+
+    ld1rw  z8.s, p0/z,  [pB]
+
+    add pB, pB, 4
+
+    fmla z16.s, p0/m, z0.s, z8.s
+    fmla z17.s, p0/m, z1.s, z8.s
+	prfm	PLDL1KEEP, [pA1, #A_PRE_SIZE]
+.endm
+
+.macro SAVEv2x1
+
+	prfm	PLDL2KEEP, [pCRow0, #C_PRE_SIZE]
+
+	add	pCRow1, pCRow0, LDC
+    ld1w  z8.s, p0/z, [pCRow0] 
+    ld1w  z9.s, p0/z, [pCRow0, #1, mul vl] 
+    fmla z8.s, p0/m, z16.s, alphaZ
+    fmla z9.s, p0/m, z17.s, alphaZ
+    st1w  z8.s, p0, [pCRow0]
+    st1w  z9.s, p0, [pCRow0, #1, mul vl]
+
+	add	pCRow0, pCRow0, vec_len, lsl #3	// pC = pC + vec_len  * 4 * 2
+
+.endm
+
+.macro INITv1x8
+    dup         z16.s, #0
+    dup         z17.s, #0
+    dup         z18.s, #0
+    dup         z19.s, #0
+    dup         z20.s, #0
+    dup         z21.s, #0
+    dup         z22.s, #0
+    dup         z23.s, #0
+.endm
+
+.macro KERNELv1x8_I
+    ld1w  z0.s, p1/z, [pA1] 
+    ld1w  z1.s, p1/z, [pA1, lanes, lsl #2]   // next one
+	add	pA1, pA1, lanes, lsl #3	// pA1 = pA1 + lanes * 2 * 4
+
+    ld1rw  z8.s, p0/z,  [pB]
+    ld1rw  z9.s, p0/z,  [pB, 4]
+    ld1rw  z10.s, p0/z, [pB, 8]
+    ld1rw  z11.s, p0/z, [pB, 12]
+    ld1rw  z12.s, p0/z, [pB, 16]
+    ld1rw  z13.s, p0/z, [pB, 20]
+    ld1rw  z14.s, p0/z, [pB, 24]
+    ld1rw  z15.s, p0/z, [pB, 28]
+
+    add pB, pB, 32
+
+    fmla z16.s, p1/m, z0.s, z8.s
+    ld1rw  z8.s, p0/z,  [pB]
+    fmla z17.s, p1/m, z0.s, z9.s
+    ld1rw  z9.s, p0/z,  [pB, 4]
+    fmla z18.s, p1/m, z0.s, z10.s
+    ld1rw  z10.s, p0/z, [pB, 8]
+    fmla z19.s, p1/m, z0.s, z11.s
+    ld1rw  z11.s, p0/z, [pB, 12]
+    fmla z20.s, p1/m, z0.s, z12.s
+	prfm	PLDL1KEEP, [pA1, #A_PRE_SIZE]
+    ld1rw  z12.s, p0/z, [pB, 16]
+    fmla z21.s, p1/m, z0.s, z13.s
+    ld1rw  z13.s, p0/z, [pB, 20]
+    fmla z22.s, p1/m, z0.s, z14.s
+    ld1rw  z14.s, p0/z, [pB, 24]
+    fmla z23.s, p1/m, z0.s, z15.s
+	prfm	PLDL1KEEP, [pA1, #A_PRE_SIZE+64]
+    ld1rw  z15.s, p0/z, [pB, 28]
+
+    add pB, pB, 32
+.endm
+
+.macro KERNELv1x8_M1
+    ld1w  z1.s, p1/z, [pA1] 
+	add	pA1, pA1, lanes, lsl #2	// pA1 = pA1 + lanes  * 4
+
+    fmla z16.s, p1/m, z0.s, z8.s
+    ld1rw  z8.s, p0/z,  [pB]
+    fmla z17.s, p1/m, z0.s, z9.s
+    ld1rw  z9.s, p0/z,  [pB, 4]
+    fmla z18.s, p1/m, z0.s, z10.s
+    ld1rw  z10.s, p0/z, [pB, 8]
+    fmla z19.s, p1/m, z0.s, z11.s
+    ld1rw  z11.s, p0/z, [pB, 12]
+    fmla z20.s, p1/m, z0.s, z12.s
+	prfm	PLDL1KEEP, [pA1, #A_PRE_SIZE]
+    ld1rw  z12.s, p0/z, [pB, 16]
+    fmla z21.s, p1/m, z0.s, z13.s
+    ld1rw  z13.s, p0/z, [pB, 20]
+    fmla z22.s, p1/m, z0.s, z14.s
+    ld1rw  z14.s, p0/z, [pB, 24]
+    fmla z23.s, p1/m, z0.s, z15.s
+	prfm	PLDL1KEEP, [pA1, #A_PRE_SIZE+64]
+    ld1rw  z15.s, p0/z, [pB, 28]
+
+    add pB, pB, 32
+.endm
+
+.macro KERNELv1x8_M2
+    ld1w  z0.s, p1/z, [pA1] 
+	add	pA1, pA1, lanes, lsl #2	// pA1 = pA1 + lanes  * 4
+
+    fmla z16.s, p1/m, z1.s, z8.s
+    ld1rw  z8.s, p0/z,  [pB]
+    fmla z17.s, p1/m, z1.s, z9.s
+    ld1rw  z9.s, p0/z,  [pB, 4]
+    fmla z18.s, p1/m, z1.s, z10.s
+    ld1rw  z10.s, p0/z, [pB, 8]
+    fmla z19.s, p1/m, z1.s, z11.s
+    ld1rw  z11.s, p0/z, [pB, 12]
+    fmla z20.s, p1/m, z1.s, z12.s
+    ld1rw  z12.s, p0/z, [pB, 16]
+	prfm	PLDL1KEEP, [pB, #B_PRE_SIZE]
+    fmla z21.s, p1/m, z1.s, z13.s
+    ld1rw  z13.s, p0/z, [pB, 20]
+    fmla z22.s, p1/m, z1.s, z14.s
+    ld1rw  z14.s, p0/z, [pB, 24]
+    fmla z23.s, p1/m, z1.s, z15.s
+    ld1rw  z15.s, p0/z, [pB, 28]
+
+    add pB, pB, 32
+.endm
+
+.macro KERNELv1x8_E
+    fmla z16.s, p1/m, z1.s, z8.s
+    fmla z17.s, p1/m, z1.s, z9.s
+    fmla z18.s, p1/m, z1.s, z10.s
+    fmla z19.s, p1/m, z1.s, z11.s
+    fmla z20.s, p1/m, z1.s, z12.s
+	prfm	PLDL1KEEP, [pB, #B_PRE_SIZE]
+    fmla z21.s, p1/m, z1.s, z13.s
+    fmla z22.s, p1/m, z1.s, z14.s
+    fmla z23.s, p1/m, z1.s, z15.s
+.endm
+
+.macro KERNELv1x8_SUB
+    ld1w  z0.s, p1/z, [pA1] 
+	add	pA1, pA1, lanes, lsl #2	// pA1 = pA1 + lanes  * 4
+
+    ld1rw  z8.s, p0/z,  [pB]
+    ld1rw  z9.s, p0/z,  [pB, 4]
+    ld1rw  z10.s, p0/z, [pB, 8]
+    ld1rw  z11.s, p0/z, [pB, 12]
+    ld1rw  z12.s, p0/z, [pB, 16]
+    ld1rw  z13.s, p0/z, [pB, 20]
+    ld1rw  z14.s, p0/z, [pB, 24]
+    ld1rw  z15.s, p0/z, [pB, 28]
+
+    add pB, pB, 32
+
+    fmla z16.s, p1/m, z0.s, z8.s
+    fmla z17.s, p1/m, z0.s, z9.s
+    fmla z18.s, p1/m, z0.s, z10.s
+	prfm	PLDL1KEEP, [pA1, #A_PRE_SIZE]
+    fmla z19.s, p1/m, z0.s, z11.s
+    fmla z20.s, p1/m, z0.s, z12.s
+    fmla z21.s, p1/m, z0.s, z13.s
+	prfm	PLDL1KEEP, [pB, #B_PRE_SIZE]
+    fmla z22.s, p1/m, z0.s, z14.s
+    fmla z23.s, p1/m, z0.s, z15.s
+
+
+.endm
+
+.macro SAVEv1x8
+
+	prfm	PLDL2KEEP, [pCRow0, #C_PRE_SIZE]
+
+	add	pCRow1, pCRow0, LDC
+    ld1w  z24.s, p1/z, [pCRow0] 
+    fmla z24.s, p1/m, z16.s, alphaZ
+    st1w  z24.s, p1, [pCRow0]
+	prfm	PLDL2KEEP, [pCRow1, #C_PRE_SIZE]
+
+	add	pCRow2, pCRow1, LDC
+    ld1w  z25.s, p1/z, [pCRow1] 
+    fmla z25.s, p1/m, z17.s, alphaZ
+    st1w  z25.s, p1, [pCRow1]
+	prfm	PLDL2KEEP, [pCRow2, #C_PRE_SIZE]
+
+	add	pCRow1, pCRow2, LDC
+    ld1w  z26.s, p1/z, [pCRow2] 
+    fmla z26.s, p1/m, z18.s, alphaZ
+    st1w z26.s, p1, [pCRow2]
+	prfm	PLDL2KEEP, [pCRow1, #C_PRE_SIZE]
+
+	add	pCRow2, pCRow1, LDC
+    ld1w  z27.s, p1/z, [pCRow1] 
+    fmla z27.s, p1/m, z19.s, alphaZ
+    st1w  z27.s, p1, [pCRow1]
+	prfm	PLDL2KEEP, [pCRow2, #C_PRE_SIZE]
+
+	add	pCRow1, pCRow2, LDC
+    ld1w  z28.s, p1/z, [pCRow2] 
+    fmla z28.s, p1/m, z20.s, alphaZ
+    st1w  z28.s, p1, [pCRow2]
+	prfm	PLDL2KEEP, [pCRow1, #C_PRE_SIZE]
+
+	add	pCRow2, pCRow1, LDC
+    ld1w  z29.s, p1/z, [pCRow1] 
+    fmla z29.s, p1/m, z21.s, alphaZ
+    st1w  z29.s, p1, [pCRow1]
+	prfm	PLDL2KEEP, [pCRow2, #C_PRE_SIZE]
+
+	add	pCRow1, pCRow2, LDC
+    ld1w  z30.s, p1/z, [pCRow2] 
+    fmla z30.s, p1/m, z22.s, alphaZ
+    st1w  z30.s, p1, [pCRow2]
+	prfm	PLDL2KEEP, [pCRow1, #C_PRE_SIZE]
+
+    ld1w  z31.s, p1/z, [pCRow1] 
+    fmla z31.s, p1/m, z23.s, alphaZ
+    st1w  z31.s, p1, [pCRow1]
+
+	add	pCRow0, pCRow0, lanes, lsl #2	// pC = pC + lanes  * 4
+
+.endm
+
+/******************************************************************************/
+
+.macro INITv1x4
+    dup         z16.s, #0
+    dup         z17.s, #0
+    dup         z18.s, #0
+    dup         z19.s, #0
+.endm
+
+.macro KERNELv1x4_SUB
+    ld1w  z0.s, p1/z, [pA1] 
+	add	pA1, pA1, lanes, lsl #2	// pA1 = pA1 + lanes  * 4
+
+    ld1rw  z8.s, p0/z,  [pB]
+    ld1rw  z9.s, p0/z,  [pB, 4]
+    ld1rw  z10.s, p0/z, [pB, 8]
+    ld1rw  z11.s, p0/z, [pB, 12]
+
+    add pB, pB, 16
+
+    fmla z16.s, p1/m, z0.s, z8.s
+    fmla z17.s, p1/m, z0.s, z9.s
+	prfm	PLDL1KEEP, [pA1, #A_PRE_SIZE]
+    fmla z18.s, p1/m, z0.s, z10.s
+    fmla z19.s, p1/m, z0.s, z11.s
+
+.endm
+
+.macro SAVEv1x4
+
+	prfm	PLDL2KEEP, [pCRow0, #C_PRE_SIZE]
+
+	add	pCRow1, pCRow0, LDC
+    ld1w  z24.s, p1/z, [pCRow0] 
+    fmla z24.s, p1/m, z16.s, alphaZ
+    st1w  z24.s, p1, [pCRow0]
+	prfm	PLDL2KEEP, [pCRow1, #C_PRE_SIZE]
+
+	add	pCRow2, pCRow1, LDC
+    ld1w  z25.s, p1/z, [pCRow1] 
+    fmla z25.s, p1/m, z17.s, alphaZ
+    st1w  z25.s, p1, [pCRow1]
+	prfm	PLDL2KEEP, [pCRow2, #C_PRE_SIZE]
+
+	add	pCRow1, pCRow2, LDC
+    ld1w  z26.s, p1/z, [pCRow2] 
+    fmla z26.s, p1/m, z18.s, alphaZ
+    st1w z26.s, p1, [pCRow2]
+	prfm	PLDL2KEEP, [pCRow1, #C_PRE_SIZE]
+
+    ld1w  z27.s, p1/z, [pCRow1] 
+    fmla z27.s, p1/m, z19.s, alphaZ
+    st1w  z27.s, p1, [pCRow1]
+
+	add	pCRow0, pCRow0, lanes, lsl #2	// pC = pC + lanes  * 4
+
+.endm
+
+/******************************************************************************/
+
+.macro INITv1x2
+    dup         z16.s, #0
+    dup         z17.s, #0
+.endm
+
+.macro KERNELv1x2_SUB
+    ld1w  z0.s, p1/z, [pA1] 
+	add	pA1, pA1, lanes, lsl #2	// pA1 = pA1 + lanes  * 4
+
+    ld1rw  z8.s, p0/z,  [pB]
+    ld1rw  z9.s, p0/z,  [pB, 4]
+
+    add pB, pB, 8
+
+    fmla z16.s, p1/m, z0.s, z8.s
+	prfm	PLDL1KEEP, [pA1, #A_PRE_SIZE]
+    fmla z17.s, p1/m, z0.s, z9.s
+
+.endm
+
+.macro SAVEv1x2
+
+	prfm	PLDL2KEEP, [pCRow0, #C_PRE_SIZE]
+
+	add	pCRow1, pCRow0, LDC
+    ld1w  z24.s, p1/z, [pCRow0] 
+    fmla z24.s, p1/m, z16.s, alphaZ
+    st1w  z24.s, p1, [pCRow0]
+	prfm	PLDL2KEEP, [pCRow1, #C_PRE_SIZE]
+
+    ld1w  z25.s, p1/z, [pCRow1] 
+    fmla z25.s, p1/m, z17.s, alphaZ
+    st1w  z25.s, p1, [pCRow1]
+
+	add	pCRow0, pCRow0, lanes, lsl #2	// pC = pC + lanes  * 4
+
+.endm
+
+/******************************************************************************/
+
+.macro INITv1x1
+    dup         z16.s, #0
+.endm
+
+.macro KERNELv1x1_SUB
+    ld1w  z0.s, p1/z, [pA1] 
+	add	pA1, pA1, lanes, lsl #2	// pA1 = pA1 + lanes  * 4
+
+    ld1rw  z8.s, p0/z,  [pB]
+
+    add pB, pB, 4
+
+    fmla z16.s, p1/m, z0.s, z8.s
+	prfm	PLDL1KEEP, [pA1, #A_PRE_SIZE]
+
+.endm
+
+.macro SAVEv1x1
+
+	prfm	PLDL2KEEP, [pCRow0, #C_PRE_SIZE]
+
+    ld1w  z24.s, p1/z, [pCRow0] 
+    fmla z24.s, p1/m, z16.s, alphaZ
+    st1w  z24.s, p1, [pCRow0]
+
+
+	add	pCRow0, pCRow0, lanes, lsl #2	// pC = pC + lanes  * 4
+
+.endm
+
+
+/*******************************************************************************
+* End of macro definitions
+*******************************************************************************/
+
+	PROLOGUE
+
+	.align 5
+	add	sp, sp, #-(11 * 16)
+	stp	d8, d9, [sp, #(0 * 16)]
+	stp	d10, d11, [sp, #(1 * 16)]
+	stp	d12, d13, [sp, #(2 * 16)]
+	stp	d14, d15, [sp, #(3 * 16)]
+	stp	d16, d17, [sp, #(4 * 16)]
+	stp	x18, x19, [sp, #(5 * 16)]
+	stp	x20, x21, [sp, #(6 * 16)]
+	stp	x22, x23, [sp, #(7 * 16)]
+	stp	x24, x25, [sp, #(8 * 16)]
+	stp	x26, x27, [sp, #(9 * 16)]
+	str	x28, [sp, #(10 * 16)]
+
+	prfm	PLDL1KEEP, [origPB]
+	prfm	PLDL1KEEP, [origPA]
+
+	fmov	alpha, s0
+	dup	alphaZ, alpha
+    cntw vec_len
+    lsl vec_lenx2, vec_len, #1
+
+	lsl	LDC, LDC, #2			// ldc = ldc * 8
+    ptrue p0.s                  // create true predicate 
+
+	mov	pB, origPB
+// Loop over N
+	mov	counterJ, origN
+	asr 	counterJ, counterJ, #3		// J = J / 8
+	cmp 	counterJ, #0
+	ble	.Lsgemm_kernel_L4_BEGIN
+
+/******************************************************************************/
+/* Repeat this as long as there are 8 left in N */
+
+	.align 5
+.Lsgemm_kernel_L8_BEGIN:
+	mov	pCRow0, pC
+
+    add pC, pC, LDC, lsl #3 // add 8 x LDC
+
+	mov	pA1, origPA			// pA1 = start of A array
+
+.Lsgemm_kernel_L8_Mv2_BEGIN:
+
+    mov counterI, #0
+    cmp origM, vec_lenx2        // Check if M < 2*SVE_LEN
+    blt .Lsgemm_kernel_L8_Mv1_BEGIN
+
+    mov counterI, origM
+
+/* Until we have at least 2*SVE_LEN iters left in M, we do them with V2*8 kernel */
+    mul temp, vec_len, origK                // generate address of pA2
+	add	pA2, pA1, temp, lsl #2			// pA1 = start of A array
+	prfm	PLDL1KEEP, [pA2]
+
+	.align 5
+.Lsgemm_kernel_L8_Mv2_20:
+
+	mov	pB, origPB
+    INITv2x8                     // fill with zeros
+
+	asr 	counterL , origK, #3		// L = K / 8
+	cmp	counterL , #2			// is there at least 4 to do?
+	blt	.Lsgemm_kernel_L8_Mv2_32
+
+	KERNELv2x8_I
+	KERNELv2x8_M2
+	KERNELv2x8_M1
+	KERNELv2x8_M2
+	KERNELv2x8_M1
+	KERNELv2x8_M2
+	KERNELv2x8_M1
+	KERNELv2x8_M2
+
+	subs	counterL, counterL, #2		// subtract 2
+	ble	.Lsgemm_kernel_L8_Mv2_22a
+
+	.align 5
+.Lsgemm_kernel_L8_Mv2_22:
+
+	KERNELv2x8_M1
+	KERNELv2x8_M2
+	KERNELv2x8_M1
+	KERNELv2x8_M2
+	KERNELv2x8_M1
+	KERNELv2x8_M2
+	KERNELv2x8_M1
+	KERNELv2x8_M2
+
+	subs	counterL, counterL, #1
+	bgt	.Lsgemm_kernel_L8_Mv2_22
+
+	.align 5
+.Lsgemm_kernel_L8_Mv2_22a:
+
+	KERNELv2x8_M1
+	KERNELv2x8_M2
+	KERNELv2x8_M1
+	KERNELv2x8_M2
+	KERNELv2x8_M1
+	KERNELv2x8_M2
+	KERNELv2x8_M1
+	KERNELv2x8_E
+
+	b	 .Lsgemm_kernel_L8_Mv2_44
+
+	.align 5
+.Lsgemm_kernel_L8_Mv2_32:
+
+	tst	counterL, #1
+	ble	.Lsgemm_kernel_L8_Mv2_40
+
+	KERNELv2x8_I
+	KERNELv2x8_M2
+	KERNELv2x8_M1
+	KERNELv2x8_M2
+	KERNELv2x8_M1
+	KERNELv2x8_M2
+	KERNELv2x8_M1
+	KERNELv2x8_E
+
+
+	b	.Lsgemm_kernel_L8_Mv2_44
+
+.Lsgemm_kernel_L8_Mv2_40:
+
+	INITv2x8
+
+.Lsgemm_kernel_L8_Mv2_44:
+
+	ands	counterL , origK, #7
+	ble	.Lsgemm_kernel_L8_Mv2_100
+
+	.align 5
+.Lsgemm_kernel_L8_Mv2_46:
+
+	KERNELv2x8_SUB
+
+	subs	counterL, counterL, #1
+	bne	.Lsgemm_kernel_L8_Mv2_46
+
+.Lsgemm_kernel_L8_Mv2_100:
+	prfm	PLDL1KEEP, [pA1]
+	prfm	PLDL1KEEP, [pA1, #64]
+	prfm	PLDL1KEEP, [pA2]
+	prfm	PLDL1KEEP, [pA2, #64]
+	prfm	PLDL1KEEP, [origPB]
+
+	SAVEv2x8
+    mov pA1, pA2                            // pA1 = pA2
+    mul temp, vec_len, origK                // generate address of pA2
+	add	pA2, pA1, temp, lsl #2			    // 
+
+.Lsgemm_kernel_L8_Mv2_END:
+    sub counterI, counterI, vec_lenx2
+    cmp counterI, vec_lenx2
+    bge .Lsgemm_kernel_L8_Mv2_20
+    sub counterI, origM, counterI
+
+    cmp counterI, origM
+    beq .Lsgemm_kernel_L8_END
+
+//////////////////////////////////////////
+// We have less than 2*SVE_LEN left. We do this with V1x8 kernel.
+.Lsgemm_kernel_L8_Mv1_BEGIN:
+
+    whilelt p1.s, counterI, origM               //SVE instruction
+    cntp lanes, p0, p1.s                        // lanes contain number of active SVE lanes in M dimension
+
+	.align 5
+.Lsgemm_kernel_L8_Mv1_20:
+
+	mov	pB, origPB
+    INITv1x8                     // fill with zeros
+
+	asr 	counterL , origK, #3		// L = K / 8
+	cmp	counterL , #2			// is there at least 4 to do?
+	blt	.Lsgemm_kernel_L8_Mv1_32
+
+	KERNELv1x8_I
+	KERNELv1x8_M2
+	KERNELv1x8_M1
+	KERNELv1x8_M2
+	KERNELv1x8_M1
+	KERNELv1x8_M2
+	KERNELv1x8_M1
+	KERNELv1x8_M2
+
+	subs	counterL, counterL, #2		// subtract 2
+	ble	.Lsgemm_kernel_L8_Mv1_22a
+
+	.align 5
+.Lsgemm_kernel_L8_Mv1_22:
+
+	KERNELv1x8_M1
+	KERNELv1x8_M2
+	KERNELv1x8_M1
+	KERNELv1x8_M2
+	KERNELv1x8_M1
+	KERNELv1x8_M2
+	KERNELv1x8_M1
+	KERNELv1x8_M2
+
+	subs	counterL, counterL, #1
+	bgt	.Lsgemm_kernel_L8_Mv1_22
+
+	.align 5
+.Lsgemm_kernel_L8_Mv1_22a:
+
+	KERNELv1x8_M1
+	KERNELv1x8_M2
+	KERNELv1x8_M1
+	KERNELv1x8_M2
+	KERNELv1x8_M1
+	KERNELv1x8_M2
+	KERNELv1x8_M1
+	KERNELv1x8_E
+
+	b	 .Lsgemm_kernel_L8_Mv1_44
+
+	.align 5
+.Lsgemm_kernel_L8_Mv1_32:
+
+	tst	counterL, #1
+	ble	.Lsgemm_kernel_L8_Mv1_40
+
+	KERNELv1x8_I
+	KERNELv1x8_M2
+	KERNELv1x8_M1
+	KERNELv1x8_M2
+	KERNELv1x8_M1
+	KERNELv1x8_M2
+	KERNELv1x8_M1
+	KERNELv1x8_E
+
+
+	b	.Lsgemm_kernel_L8_Mv1_44
+
+.Lsgemm_kernel_L8_Mv1_40:
+
+	INITv1x8
+
+.Lsgemm_kernel_L8_Mv1_44:
+
+	ands	counterL , origK, #7
+	ble	.Lsgemm_kernel_L8_Mv1_100
+
+	.align 5
+.Lsgemm_kernel_L8_Mv1_46:
+
+	KERNELv1x8_SUB
+
+	subs	counterL, counterL, #1
+	bne	.Lsgemm_kernel_L8_Mv1_46
+
+.Lsgemm_kernel_L8_Mv1_100:
+	prfm	PLDL1KEEP, [pA1]
+	prfm	PLDL1KEEP, [pA1, #64]
+	prfm	PLDL1KEEP, [origPB]
+
+	SAVEv1x8
+
+.Lsgemm_kernel_L8_Mv1_END:
+
+    incw    counterI
+    whilelt p1.s, counterI, origM             //SVE instruction
+    cntp lanes, p0, p1.s                        // lanes contain number of active SVE lanes in M dimension
+    b.any   .Lsgemm_kernel_L8_Mv1_20   
+
+.Lsgemm_kernel_L8_END:
+
+	lsl	temp, origK, #5 
+	add	origPB, origPB, temp		// B = B + K * 8 * 4
+
+	subs	counterJ, counterJ , #1		// j--
+	bgt	.Lsgemm_kernel_L8_BEGIN
+
+/******************************************************************************/
+/* Repeat the same thing if 4 left in N */
+
+	.align 5
+.Lsgemm_kernel_L4_BEGIN:
+
+	mov	counterJ , origN
+	tst	counterJ , #4
+	ble	.Lsgemm_kernel_L2_BEGIN
+
+
+	mov	pCRow0, pC
+
+    add pC, pC, LDC, lsl #2 // add 4 x LDC
+
+	mov	pA1, origPA			// pA1 = start of A array
+
+.Lsgemm_kernel_L4_Mv2_BEGIN:
+
+    mov counterI, #0
+    cmp origM, vec_lenx2
+    blt .Lsgemm_kernel_L4_Mv1_BEGIN
+
+    mov counterI, origM
+
+    mul temp, vec_len, origK                // generate address of pA2
+	add	pA2, pA1, temp, lsl #2			// pA1 = start of A array
+
+	.align 5
+.Lsgemm_kernel_L4_Mv2_20:
+
+	mov	pB, origPB
+    INITv2x4                     // fill with zeros
+
+	asr 	counterL , origK, #3		// L = K / 8
+	cmp	counterL , #0			// is there at least 4 to do?
+	ble	.Lsgemm_kernel_L4_Mv2_44
+
+	.align 5
+.Lsgemm_kernel_L4_Mv2_22:
+
+	prfm	PLDL1KEEP, [pB, #B_PRE_SIZE]
+	KERNELv2x4_SUB
+	KERNELv2x4_SUB
+	prfm	PLDL1KEEP, [pB, #B_PRE_SIZE]
+	KERNELv2x4_SUB
+	KERNELv2x4_SUB
+	prfm	PLDL1KEEP, [pB, #B_PRE_SIZE]
+	KERNELv2x4_SUB
+	KERNELv2x4_SUB
+	prfm	PLDL1KEEP, [pB, #B_PRE_SIZE]
+	KERNELv2x4_SUB
+	KERNELv2x4_SUB
+
+	subs	counterL, counterL, #1
+	bgt	.Lsgemm_kernel_L4_Mv2_22
+
+.Lsgemm_kernel_L4_Mv2_44:
+
+	ands	counterL , origK, #7
+	ble	.Lsgemm_kernel_L4_Mv2_100
+
+	.align 5
+.Lsgemm_kernel_L4_Mv2_46:
+
+	prfm	PLDL1KEEP, [pB, #B_PRE_SIZE]
+	KERNELv2x4_SUB
+
+	subs	counterL, counterL, #1
+	bne	.Lsgemm_kernel_L4_Mv2_46
+
+.Lsgemm_kernel_L4_Mv2_100:
+	prfm	PLDL1KEEP, [pA1]
+	prfm	PLDL1KEEP, [pA1, #64]
+	prfm	PLDL1KEEP, [pA2]
+	prfm	PLDL1KEEP, [pA2, #64]
+	prfm	PLDL1KEEP, [origPB]
+
+	SAVEv2x4
+    mov pA1, pA2                            // pA1 = pA2
+    mul temp, vec_len, origK                // generate address of pA2
+	add	pA2, pA1, temp, lsl #2			    // 
+
+.Lsgemm_kernel_L4_Mv2_END:
+    sub counterI, counterI, vec_lenx2
+    cmp counterI, vec_lenx2
+    bge .Lsgemm_kernel_L4_Mv2_20
+    sub counterI, origM, counterI
+
+    cmp counterI, origM
+    beq .Lsgemm_kernel_L4_END
+
+//////////////////////////////////
+// We have less than 2*SVE_LEN left. We do this with V1x4 kernel.
+.Lsgemm_kernel_L4_Mv1_BEGIN:
+
+    whilelt p1.s, counterI, origM               //SVE instruction
+    cntp lanes, p0, p1.s                        // lanes contain number of active SVE lanes in M dimension
+
+	.align 5
+.Lsgemm_kernel_L4_Mv1_20:
+
+	mov	pB, origPB
+    INITv1x4                     // fill with zeros
+
+	asr 	counterL , origK, #3		// L = K / 8
+	cmp	counterL , #0			// is there at least 4 to do?
+	ble	.Lsgemm_kernel_L4_Mv1_44
+
+	.align 5
+.Lsgemm_kernel_L4_Mv1_22:
+
+	prfm	PLDL1KEEP, [pB, #B_PRE_SIZE]
+	KERNELv1x4_SUB
+	KERNELv1x4_SUB
+	prfm	PLDL1KEEP, [pB, #B_PRE_SIZE]
+	KERNELv1x4_SUB
+	KERNELv1x4_SUB
+	prfm	PLDL1KEEP, [pB, #B_PRE_SIZE]
+	KERNELv1x4_SUB
+	KERNELv1x4_SUB
+	prfm	PLDL1KEEP, [pB, #B_PRE_SIZE]
+	KERNELv1x4_SUB
+	KERNELv1x4_SUB
+
+	subs	counterL, counterL, #1
+	bgt	.Lsgemm_kernel_L4_Mv1_22
+
+.Lsgemm_kernel_L4_Mv1_44:
+
+	ands	counterL , origK, #7
+	ble	.Lsgemm_kernel_L4_Mv1_100
+
+	.align 5
+.Lsgemm_kernel_L4_Mv1_46:
+
+	prfm	PLDL1KEEP, [pB, #B_PRE_SIZE]
+	KERNELv1x4_SUB
+
+	subs	counterL, counterL, #1
+	bne	.Lsgemm_kernel_L4_Mv1_46
+
+.Lsgemm_kernel_L4_Mv1_100:
+	prfm	PLDL1KEEP, [pA1]
+	prfm	PLDL1KEEP, [pA1, #64]
+	prfm	PLDL1KEEP, [origPB]
+
+	SAVEv1x4
+
+.Lsgemm_kernel_L4_Mv1_END:
+
+    incw    counterI
+    whilelt p1.s, counterI, origM             //SVE instruction
+    cntp lanes, p0, p1.s
+    b.any   .Lsgemm_kernel_L4_Mv1_20   
+
+
+.Lsgemm_kernel_L4_END:
+	lsl	temp, origK, #4 
+	add	origPB, origPB, temp	// B = B + K * 4 * 4
+
+/******************************************************************************/
+/* Repeat the same thing if 2 left in N */
+
+	.align 5
+.Lsgemm_kernel_L2_BEGIN:
+
+	mov	counterJ , origN
+	tst	counterJ , #2
+	ble	.Lsgemm_kernel_L1_BEGIN
+
+	mov	pCRow0, pC
+
+    add pC, pC, LDC, lsl #1 // add 2 x LDC
+
+	mov	pA1, origPA			// pA1 = start of A array
+
+.Lsgemm_kernel_L2_Mv2_BEGIN:
+
+    mov counterI, #0
+    cmp origM, vec_lenx2
+    blt .Lsgemm_kernel_L2_Mv1_BEGIN
+
+    mov counterI, origM
+
+    mul temp, vec_len, origK                // generate address of pA2
+	add	pA2, pA1, temp, lsl #2			// pA1 = start of A array
+
+	.align 5
+.Lsgemm_kernel_L2_Mv2_20:
+
+	mov	pB, origPB
+    INITv2x2                     // fill with zeros
+
+	asr 	counterL , origK, #3		// L = K / 8
+	cmp	counterL , #0			// is there at least 4 to do?
+	ble	.Lsgemm_kernel_L2_Mv2_44
+
+	.align 5
+.Lsgemm_kernel_L2_Mv2_22:
+
+	prfm	PLDL1KEEP, [pB, #B_PRE_SIZE]
+	KERNELv2x2_SUB
+	KERNELv2x2_SUB
+	KERNELv2x2_SUB
+	KERNELv2x2_SUB
+	prfm	PLDL1KEEP, [pB, #B_PRE_SIZE]
+	KERNELv2x2_SUB
+	KERNELv2x2_SUB
+	KERNELv2x2_SUB
+	KERNELv2x2_SUB
+
+	subs	counterL, counterL, #1
+	bgt	.Lsgemm_kernel_L2_Mv2_22
+
+.Lsgemm_kernel_L2_Mv2_44:
+
+	ands	counterL , origK, #7
+	ble	.Lsgemm_kernel_L2_Mv2_100
+
+	.align 5
+.Lsgemm_kernel_L2_Mv2_46:
+
+	prfm	PLDL1KEEP, [pB, #B_PRE_SIZE]
+	KERNELv2x2_SUB
+
+	subs	counterL, counterL, #1
+	bne	.Lsgemm_kernel_L2_Mv2_46
+
+.Lsgemm_kernel_L2_Mv2_100:
+	prfm	PLDL1KEEP, [pA1]
+	prfm	PLDL1KEEP, [pA1, #64]
+	prfm	PLDL1KEEP, [pA2]
+	prfm	PLDL1KEEP, [pA2, #64]
+	prfm	PLDL1KEEP, [origPB]
+
+	SAVEv2x2
+    mov pA1, pA2                            // pA1 = pA2
+    mul temp, vec_len, origK                // generate address of pA2
+	add	pA2, pA1, temp, lsl #2			    // 
+
+.Lsgemm_kernel_L2_Mv2_END:
+    sub counterI, counterI, vec_lenx2
+    cmp counterI, vec_lenx2
+    bge .Lsgemm_kernel_L2_Mv2_20
+    sub counterI, origM, counterI
+
+    cmp counterI, origM
+    beq .Lsgemm_kernel_L2_END
+
+
+//////////////////////////////////
+// We have less than 2*SVE_LEN left. We do this with V1x2 kernel.
+.Lsgemm_kernel_L2_Mv1_BEGIN:
+
+    whilelt p1.s, counterI, origM               //SVE instruction
+    cntp lanes, p0, p1.s
+
+	.align 5
+.Lsgemm_kernel_L2_Mv1_20:
+
+	mov	pB, origPB
+    INITv1x2                     // fill with zeros
+
+	asr 	counterL , origK, #3		// L = K / 8
+	cmp	counterL , #0			// is there at least 4 to do?
+	ble	.Lsgemm_kernel_L2_Mv1_44
+
+	.align 5
+.Lsgemm_kernel_L2_Mv1_22:
+
+	prfm	PLDL1KEEP, [pB, #B_PRE_SIZE]
+	KERNELv1x2_SUB
+	KERNELv1x2_SUB
+	KERNELv1x2_SUB
+	KERNELv1x2_SUB
+	prfm	PLDL1KEEP, [pB, #B_PRE_SIZE]
+	KERNELv1x2_SUB
+	KERNELv1x2_SUB
+	KERNELv1x2_SUB
+	KERNELv1x2_SUB
+
+	subs	counterL, counterL, #1
+	bgt	.Lsgemm_kernel_L2_Mv1_22
+
+.Lsgemm_kernel_L2_Mv1_44:
+
+	ands	counterL , origK, #7
+	ble	.Lsgemm_kernel_L2_Mv1_100
+
+	.align 5
+.Lsgemm_kernel_L2_Mv1_46:
+
+	prfm	PLDL1KEEP, [pB, #B_PRE_SIZE]
+	KERNELv1x2_SUB
+
+	subs	counterL, counterL, #1
+	bne	.Lsgemm_kernel_L2_Mv1_46
+
+.Lsgemm_kernel_L2_Mv1_100:
+	prfm	PLDL1KEEP, [pA1]
+	prfm	PLDL1KEEP, [pA1, #64]
+	prfm	PLDL1KEEP, [origPB]
+
+	SAVEv1x2
+
+.Lsgemm_kernel_L2_Mv1_END:
+
+    incw    counterI
+    whilelt p1.s, counterI, origM             //SVE instruction
+    cntp lanes, p0, p1.s
+    b.any   .Lsgemm_kernel_L2_Mv1_20   
+
+
+.Lsgemm_kernel_L2_END:
+	add	origPB, origPB, origK, lsl #3	// B = B + K * 2 * 4
+
+/******************************************************************************/
+/* Repeat the same thing if 1 left in N */
+
+	.align 5
+.Lsgemm_kernel_L1_BEGIN:
+
+	mov	counterJ , origN
+	tst	counterJ , #1
+	ble	.Lsgemm_kernel_L999 // done
+
+	mov	pCRow0, pC
+
+    add pC, pC, LDC // add 1 x LDC
+
+	mov	pA1, origPA			// pA1 = start of A array
+
+.Lsgemm_kernel_L1_Mv2_BEGIN:
+
+    mov counterI, #0
+    cmp origM, vec_lenx2
+    blt .Lsgemm_kernel_L1_Mv1_BEGIN
+
+    mov counterI, origM
+
+    mul temp, vec_len, origK                // generate address of pA2
+	add	pA2, pA1, temp, lsl #2			// pA1 = start of A array
+
+
+	.align 5
+.Lsgemm_kernel_L1_Mv2_20:
+
+	mov	pB, origPB
+    INITv2x1                     // fill with zeros
+
+	asr 	counterL , origK, #3		// L = K / 8
+	cmp	counterL , #0			// is there at least 8 to do?
+	ble	.Lsgemm_kernel_L1_Mv2_44
+
+	.align 5
+.Lsgemm_kernel_L1_Mv2_22:
+
+	prfm	PLDL1KEEP, [pB, #B_PRE_SIZE]
+	KERNELv2x1_SUB
+	KERNELv2x1_SUB
+	KERNELv2x1_SUB
+	KERNELv2x1_SUB
+	KERNELv2x1_SUB
+	KERNELv2x1_SUB
+	KERNELv2x1_SUB
+	KERNELv2x1_SUB
+
+	subs	counterL, counterL, #1
+	bgt	.Lsgemm_kernel_L1_Mv2_22
+
+.Lsgemm_kernel_L1_Mv2_44:
+
+	ands	counterL , origK, #7
+	ble	.Lsgemm_kernel_L1_Mv2_100
+
+	.align 5
+.Lsgemm_kernel_L1_Mv2_46:
+
+	prfm	PLDL1KEEP, [pB, #B_PRE_SIZE]
+	KERNELv2x1_SUB
+
+	subs	counterL, counterL, #1
+	bgt	.Lsgemm_kernel_L1_Mv2_46
+
+.Lsgemm_kernel_L1_Mv2_100:
+	prfm	PLDL1KEEP, [pA1]
+	prfm	PLDL1KEEP, [pA1, #64]
+	prfm	PLDL1KEEP, [origPB]
+
+	SAVEv2x1
+    mov pA1, pA2                            // pA1 = pA2
+    mul temp, vec_len, origK                // generate address of pA2
+	add	pA2, pA1, temp, lsl #2			    // 
+
+.Lsgemm_kernel_L1_Mv2_END:
+    sub counterI, counterI, vec_lenx2
+    cmp counterI, vec_lenx2
+    bge .Lsgemm_kernel_L1_Mv2_20
+    sub counterI, origM, counterI
+
+    cmp counterI, origM
+    beq .Lsgemm_kernel_L1_END
+
+
+//////////////////////////////////
+// We have less than 2*SVE_LEN left. We do this with V1x1 kernel.
+.Lsgemm_kernel_L1_Mv1_BEGIN:
+
+    whilelt p1.s, counterI, origM               //SVE instruction
+    cntp lanes, p0, p1.s
+
+	.align 5
+.Lsgemm_kernel_L1_Mv1_20:
+
+	mov	pB, origPB
+    INITv1x1                     // fill with zeros
+
+	asr 	counterL , origK, #3		// L = K / 8
+	cmp	counterL , #0			// is there at least 8 to do?
+	ble	.Lsgemm_kernel_L1_Mv1_44
+
+	.align 5
+.Lsgemm_kernel_L1_Mv1_22:
+
+	prfm	PLDL1KEEP, [pB, #B_PRE_SIZE]
+	KERNELv1x1_SUB
+	KERNELv1x1_SUB
+	KERNELv1x1_SUB
+	KERNELv1x1_SUB
+	KERNELv1x1_SUB
+	KERNELv1x1_SUB
+	KERNELv1x1_SUB
+	KERNELv1x1_SUB
+
+	subs	counterL, counterL, #1
+	bgt	.Lsgemm_kernel_L1_Mv1_22
+
+.Lsgemm_kernel_L1_Mv1_44:
+
+	ands	counterL , origK, #7
+	ble	.Lsgemm_kernel_L1_Mv1_100
+
+	.align 5
+.Lsgemm_kernel_L1_Mv1_46:
+
+	prfm	PLDL1KEEP, [pB, #B_PRE_SIZE]
+	KERNELv1x1_SUB
+
+	subs	counterL, counterL, #1
+	bgt	.Lsgemm_kernel_L1_Mv1_46
+
+.Lsgemm_kernel_L1_Mv1_100:
+	prfm	PLDL1KEEP, [pA1]
+	prfm	PLDL1KEEP, [pA1, #64]
+	prfm	PLDL1KEEP, [origPB]
+
+	SAVEv1x1
+
+.Lsgemm_kernel_L1_Mv1_END:
+
+    incw    counterI
+    whilelt p1.s, counterI, origM             //SVE instruction
+    cntp lanes, p0, p1.s
+    b.any   .Lsgemm_kernel_L1_Mv1_20   
+
+
+.Lsgemm_kernel_L1_END:
+
+/******************************************************************************/
+
+.Lsgemm_kernel_L999:
+	mov	x0, #0				// set return value
+	ldp	d8, d9, [sp, #(0 * 16)]
+	ldp	d10, d11, [sp, #(1 * 16)]
+	ldp	d12, d13, [sp, #(2 * 16)]
+	ldp	d14, d15, [sp, #(3 * 16)]
+	ldp	d16, d17, [sp, #(4 * 16)]
+	ldp	x18, x19, [sp, #(5 * 16)]
+	ldp	x20, x21, [sp, #(6 * 16)]
+	ldp	x22, x23, [sp, #(7 * 16)]
+	ldp	x24, x25, [sp, #(8 * 16)]
+	ldp	x26, x27, [sp, #(9 * 16)]
+	ldr	x28, [sp, #(10 * 16)]
+	add	sp, sp, #(11*16)
+	ret
+
+	EPILOGUE
+