diff --git a/arm_compute/core/NEON/kernels/arm64/NEGEMMLowpAArch64Kernel.h b/arm_compute/core/NEON/kernels/arm64/NEGEMMLowpAArch64Kernel.h
index e8e71cf..a93df03 100644
--- a/arm_compute/core/NEON/kernels/arm64/NEGEMMLowpAArch64Kernel.h
+++ b/arm_compute/core/NEON/kernels/arm64/NEGEMMLowpAArch64Kernel.h
@@ -37,11 +37,19 @@
 class NEGEMMLowpAArch64Kernel : public NEGEMMAssemblyBaseKernel
 {
 public:
+    /** Default constructor */
+    NEGEMMLowpAArch64Kernel();
+
     // Inherited methods overridden:
     void run(const Window &window, const ThreadInfo &info) override;
 
 protected:
     void internal_configure(const ITensor *input0, const ITensor *input1, ITensor *output, ITensor *workspace, float alpha, float beta, bool transform_0, bool transform_1) override;
+
+private:
+    using NEGEMMLowpAArch64 = void(const ITensor *input0, const ITensor *input1, ITensor *output, ITensor *workspace, float alpha, float beta, bool transform_0, bool transform_1, const Window &window,
+                                   const ThreadInfo &info);
+    NEGEMMLowpAArch64 *_func;
 };
 } // namespace arm_compute
 #endif /* ARM_COMPUTE_AARCH64_V8A */
diff --git a/arm_compute/core/NEON/kernels/assembly/kernels/a64_gemm_u8_4x4.hpp b/arm_compute/core/NEON/kernels/assembly/kernels/a64_gemm_u8_4x4.hpp
new file mode 100644
index 0000000..3561bfe
--- /dev/null
+++ b/arm_compute/core/NEON/kernels/assembly/kernels/a64_gemm_u8_4x4.hpp
@@ -0,0 +1,61 @@
+/*
+ * Copyright (c) 2017 ARM Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#pragma once
+
+#ifdef __aarch64__
+
+// Load the actual kernel
+#include "a64_gemm_u8_4x4/generic.hpp"
+
+class gemm_u8_4x4 {
+public:
+    typedef uint8_t operand_type;
+    typedef uint32_t result_type;
+
+    typedef void (*kern_type)(const uint8_t *, const uint8_t *, uint32_t *, int, int, int);
+
+    /* Describes the data layout for A input */
+    static const int A_interleave = 4;
+    static const int A_block = 16;
+    static const bool A_transpose = false;
+
+    /* Same for B input */
+    static const int B_interleave = 4;
+    static const int B_block = 16;
+    static const bool B_transpose = true;
+
+    /* Kernel blocking parameters */
+    static const int out_width = 4;
+    static const int out_height = 4;
+    static const int k_unroll = 16;
+
+    kern_type kernel = nullptr;
+
+    gemm_u8_4x4(const CPUInfo *ci) {
+        kernel = a64_gemm_u8_4x4;
+    }
+};
+
+#endif // __aarch64__
+
diff --git a/arm_compute/core/NEON/kernels/assembly/kernels/a64_gemm_u8_4x4/generic.hpp b/arm_compute/core/NEON/kernels/assembly/kernels/a64_gemm_u8_4x4/generic.hpp
new file mode 100644
index 0000000..e48c373
--- /dev/null
+++ b/arm_compute/core/NEON/kernels/assembly/kernels/a64_gemm_u8_4x4/generic.hpp
@@ -0,0 +1,465 @@
+/*
+ * Copyright (c) 2017 ARM Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#pragma once
+
+#ifdef __aarch64__
+
+#include <arm_neon.h>
+
+inline void a64_gemm_u8_4x4(const uint8_t *Apanel, const uint8_t *Bpanel, uint32_t *Cpanel, int ablocks, int bblocks, int K) {
+    const uint8_t *a_ptr = Apanel;
+    uint32_t *c_ptr = Cpanel;
+    K /= 16;
+    int oddk = (K & 1);
+
+    for (int yb=0; yb<ablocks; yb++) {
+        const uint8_t *a_ptr0 = a_ptr;
+        const uint8_t *b_ptr = Bpanel;
+
+        for (int xb=0; xb<bblocks; xb++) {
+            a_ptr = a_ptr0;
+
+            int k = ((K+1)/2)-1;
+
+            register uint8x16_t b0  asm("v4");
+            register uint8x16_t b1  asm("v5");
+            register uint8x16_t b2  asm("v6");
+            register uint8x16_t b3  asm("v7");
+            register uint8x16_t b0a asm("v8");
+            register uint8x16_t b1a asm("v9");
+            register uint8x16_t b2a asm("v10");
+            register uint8x16_t b3a asm("v11");
+
+            __asm __volatile (
+                "movi	v16.4s, #0x0\n"
+                "ldr	q0, [%[a_ptr]]\n"
+                "movi	v17.4s, #0x0\n"
+                "ldr	%q[b0], [%[b_ptr]]\n"
+                "movi	v18.4s, #0x0\n"
+                "ldr	%q[b1], [%[b_ptr], #16]\n"
+                "movi	v19.4s, #0x0\n"
+                "ldr	%q[b2], [%[b_ptr], #32]\n"
+                "movi	v20.4s, #0x0\n"
+                "ldr	%q[b3], [%[b_ptr], #48]\n"
+                "movi	v21.4s, #0x0\n"
+                "ldr	q1, [%[a_ptr], #16]\n"
+                "movi	v22.4s, #0x0\n"
+                "ldr	q2, [%[a_ptr], #32]\n"
+                "movi	v23.4s, #0x0\n"
+                "ldr	q3, [%[a_ptr], #48]\n"
+                "movi	v24.4s, #0x0\n"
+                ASM_PREFETCH("[%[b_ptr], #64]")
+                "movi	v25.4s, #0x0\n"
+                ASM_PREFETCH("[%[a_ptr], #64]")
+                "movi	v26.4s, #0x0\n"
+                ASM_PREFETCH("[%[b_ptr], #128]")
+                "movi	v27.4s, #0x0\n"
+                ASM_PREFETCH("[%[a_ptr], #128]")
+                "movi	v28.4s, #0x0\n"
+                ASM_PREFETCH("[%[b_ptr], #192]")
+                "movi	v29.4s, #0x0\n"
+                ASM_PREFETCH("[%[a_ptr], #192]")
+                "movi	v30.4s, #0x0\n"
+                ASM_PREFETCH("[%[b_ptr], #256]")
+                "movi	v31.4s, #0x0\n"
+                ASM_PREFETCH("[%[a_ptr], #256]")
+
+                // Loop structure optimized for A57 (after r0).
+
+                // Unavoidably, the multiply will "dribble" if
+                // dual issued with an add.
+
+                // Minimize the effect of this by making sure
+                // there are 2 adds to run under the dribbled
+                // multiply.
+
+                // Pipeline in blocks of 8 multiplies - combine
+                // this iteration's multiplies with adds from
+                // the previous iteration.
+
+                // So the first block doesn't have any adds to
+                // do - but because all the adds are at the
+                // start of the block it's only the first couple
+                // of multiplies that need to be pulled out.
+
+                // Start of unroll 0 (first iteration)
+                "umull	v12.8h, v0.8b, %[b0].8b\n"
+                "umull	v13.8h, v0.8b, %[b1].8b\n"
+
+                // Skip loop if we are doing zero iterations of it.
+                "cbz	%w[k], 4f\n"
+
+                // Unroll 0 continuation (branch target)
+                "1:\n"
+                "umull	v14.8h, v0.8b, %[b2].8b\n"
+                "subs	%w[k], %w[k], #1\n"
+                "umull	v15.8h, v0.8b, %[b3].8b\n"
+                "ldr	%q[b0a], [%[b_ptr], #64]\n"
+                "umlal2	v12.8h, v0.16b, %[b0].16b\n"
+                "umlal2	v13.8h, v0.16b, %[b1].16b\n"
+                "ldr	%q[b1a], [%[b_ptr], #80]\n"
+                "umlal2	v14.8h, v0.16b, %[b2].16b\n"
+                "umlal2	v15.8h, v0.16b, %[b3].16b\n"
+                "ldr 	q0, [%[a_ptr], #64]\n"
+
+                "uadalp	v16.4s, v12.8h\n"
+                "umull	v12.8h, v1.8b, %[b0].8b\n"
+                "uadalp	v17.4s, v13.8h\n"
+                "uadalp	v18.4s, v14.8h\n"
+                "umull	v13.8h, v1.8b, %[b1].8b\n"
+                "uadalp	v19.4s, v15.8h\n"
+                "umull	v14.8h, v1.8b, %[b2].8b\n"
+                "ldr	%q[b2a], [%[b_ptr], #96]\n"
+                "umull	v15.8h, v1.8b, %[b3].8b\n"
+                "umlal2	v12.8h, v1.16b, %[b0].16b\n"
+                "ldr	%q[b3a], [%[b_ptr], #112]\n"
+                "umlal2	v13.8h, v1.16b, %[b1].16b\n"
+                "add	%[b_ptr], %[b_ptr], #128\n"
+                "umlal2	v14.8h, v1.16b, %[b2].16b\n"
+                "umlal2	v15.8h, v1.16b, %[b3].16b\n"
+                "ldr 	q1, [%[a_ptr], #80]\n"
+
+                "uadalp	v20.4s, v12.8h\n"
+                "umull	v12.8h, v2.8b, %[b0].8b\n"
+                "uadalp	v21.4s, v13.8h\n"
+                "uadalp	v22.4s, v14.8h\n"
+                "umull	v13.8h, v2.8b, %[b1].8b\n"
+                "uadalp	v23.4s, v15.8h\n"
+                "umull	v14.8h, v2.8b, %[b2].8b\n"
+                "umull	v15.8h, v2.8b, %[b3].8b\n"
+                "umlal2	v12.8h, v2.16b, %[b0].16b\n"
+                ASM_PREFETCH("[%[b_ptr], #192]")
+                "umlal2	v13.8h, v2.16b, %[b1].16b\n"
+                "umlal2	v14.8h, v2.16b, %[b2].16b\n"
+                ASM_PREFETCH("[%[a_ptr], #320]")
+                "umlal2	v15.8h, v2.16b, %[b3].16b\n"
+                "ldr 	q2, [%[a_ptr], #96]\n"
+
+                "uadalp	v24.4s, v12.8h\n"
+                "umull	v12.8h, v3.8b, %[b0].8b\n"
+                "uadalp	v25.4s, v13.8h\n"
+                "uadalp	v26.4s, v14.8h\n"
+                "umull	v13.8h, v3.8b, %[b1].8b\n"
+                "uadalp	v27.4s, v15.8h\n"
+                "umull	v14.8h, v3.8b, %[b2].8b\n"
+                "umull	v15.8h, v3.8b, %[b3].8b\n"
+                "umlal2	v12.8h, v3.16b, %[b0].16b\n"
+                "ldr 	%q[b0], [%[b_ptr], #0]\n"
+                "umlal2	v13.8h, v3.16b, %[b1].16b\n"
+                "umlal2	v14.8h, v3.16b, %[b2].16b\n"
+                "umlal2	v15.8h, v3.16b, %[b3].16b\n"
+                "ldr 	q3, [%[a_ptr], #112]\n"
+
+                // Unroll 1
+                "uadalp	v28.4s, v12.8h\n"
+                "umull	v12.8h, v0.8b, %[b0a].8b\n"
+                "uadalp	v29.4s, v13.8h\n"
+                "uadalp	v30.4s, v14.8h\n"
+                "umull	v13.8h, v0.8b, %[b1a].8b\n"
+                "uadalp	v31.4s, v15.8h\n"
+                "umull	v14.8h, v0.8b, %[b2a].8b\n"
+                "umull	v15.8h, v0.8b, %[b3a].8b\n"
+                "ldr 	%q[b1], [%[b_ptr], #16]\n"
+                "umlal2	v12.8h, v0.16b, %[b0a].16b\n"
+                "umlal2	v13.8h, v0.16b, %[b1a].16b\n"
+                "ldr 	%q[b2], [%[b_ptr], #32]\n"
+                "umlal2	v14.8h, v0.16b, %[b2a].16b\n"
+                "umlal2	v15.8h, v0.16b, %[b3a].16b\n"
+                "ldr 	q0, [%[a_ptr], #128]\n"
+
+                "uadalp	v16.4s, v12.8h\n"
+                "umull	v12.8h, v1.8b, %[b0a].8b\n"
+                "uadalp	v17.4s, v13.8h\n"
+                "uadalp	v18.4s, v14.8h\n"
+                "umull	v13.8h, v1.8b, %[b1a].8b\n"
+                "uadalp	v19.4s, v15.8h\n"
+                "add	%[a_ptr], %[a_ptr], #128\n"
+                "umull	v14.8h, v1.8b, %[b2a].8b\n"
+                "umull	v15.8h, v1.8b, %[b3a].8b\n"
+                "ldr 	%q[b3], [%[b_ptr], #48]\n"
+                "umlal2	v12.8h, v1.16b, %[b0a].16b\n"
+                "umlal2	v13.8h, v1.16b, %[b1a].16b\n"
+                "umlal2	v14.8h, v1.16b, %[b2a].16b\n"
+                "umlal2	v15.8h, v1.16b, %[b3a].16b\n"
+                "ldr 	q1, [%[a_ptr], #16]\n"
+
+                "uadalp	v20.4s, v12.8h\n"
+                "umull	v12.8h, v2.8b, %[b0a].8b\n"
+                "uadalp	v21.4s, v13.8h\n"
+                "uadalp	v22.4s, v14.8h\n"
+                "umull	v13.8h, v2.8b, %[b1a].8b\n"
+                "uadalp	v23.4s, v15.8h\n"
+                "umull	v14.8h, v2.8b, %[b2a].8b\n"
+                "umull	v15.8h, v2.8b, %[b3a].8b\n"
+                "umlal2	v12.8h, v2.16b, %[b0a].16b\n"
+                ASM_PREFETCH("[%[b_ptr], #256]")
+                "umlal2	v13.8h, v2.16b, %[b1a].16b\n"
+                "umlal2	v14.8h, v2.16b, %[b2a].16b\n"
+                ASM_PREFETCH("[%[a_ptr], #256]")
+                "umlal2	v15.8h, v2.16b, %[b3a].16b\n"
+                "ldr 	q2, [%[a_ptr], #32]\n"
+
+                "uadalp	v24.4s, v12.8h\n"
+                "umull	v12.8h, v3.8b, %[b0a].8b\n"
+                "uadalp	v25.4s, v13.8h\n"
+                "uadalp	v26.4s, v14.8h\n"
+                "umull	v13.8h, v3.8b, %[b1a].8b\n"
+                "uadalp	v27.4s, v15.8h\n"
+                "umull	v14.8h, v3.8b, %[b2a].8b\n"
+                "umull	v15.8h, v3.8b, %[b3a].8b\n"
+                "umlal2	v12.8h, v3.16b, %[b0a].16b\n"
+                "umlal2	v13.8h, v3.16b, %[b1a].16b\n"
+                "umlal2	v14.8h, v3.16b, %[b2a].16b\n"
+                "umlal2	v15.8h, v3.16b, %[b3a].16b\n"
+                "ldr 	q3, [%[a_ptr], #48]\n"
+
+                // Start of unroll 0 for next iteration.
+                "uadalp	v28.4s, v12.8h\n"
+                "umull	v12.8h, v0.8b, %[b0].8b\n"
+                "uadalp	v29.4s, v13.8h\n"
+                "uadalp	v30.4s, v14.8h\n"
+                "umull	v13.8h, v0.8b, %[b1].8b\n"
+                "uadalp	v31.4s, v15.8h\n"
+                "bne	1b\n"
+
+                // Target to use when K=1 or 2 (i.e. zero iterations of main loop)
+                "4:\n"
+
+                // Branch to alternative tail for odd K
+                "cbnz	%w[oddk], 2f\n"
+
+                // Detached final iteration (even K)
+                "umull	v14.8h, v0.8b, %[b2].8b\n"
+                "umull	v15.8h, v0.8b, %[b3].8b\n"
+                "ldr	%q[b0a], [%[b_ptr], #64]\n"
+                "umlal2	v12.8h, v0.16b, %[b0].16b\n"
+                "umlal2	v13.8h, v0.16b, %[b1].16b\n"
+                "ldr	%q[b1a], [%[b_ptr], #80]\n"
+                "umlal2	v14.8h, v0.16b, %[b2].16b\n"
+                "umlal2	v15.8h, v0.16b, %[b3].16b\n"
+                "ldr 	q0, [%[a_ptr], #64]\n"
+
+                "uadalp	v16.4s, v12.8h\n"
+                "umull	v12.8h, v1.8b, %[b0].8b\n"
+                "uadalp	v17.4s, v13.8h\n"
+                "uadalp	v18.4s, v14.8h\n"
+                "umull	v13.8h, v1.8b, %[b1].8b\n"
+                "uadalp	v19.4s, v15.8h\n"
+                "umull	v14.8h, v1.8b, %[b2].8b\n"
+                "ldr	%q[b2a], [%[b_ptr], #96]\n"
+                "umull	v15.8h, v1.8b, %[b3].8b\n"
+                "umlal2	v12.8h, v1.16b, %[b0].16b\n"
+                "ldr	%q[b3a], [%[b_ptr], #112]\n"
+                "umlal2	v13.8h, v1.16b, %[b1].16b\n"
+                "add	%[b_ptr], %[b_ptr], #128\n"
+                "umlal2	v14.8h, v1.16b, %[b2].16b\n"
+                "umlal2	v15.8h, v1.16b, %[b3].16b\n"
+                "ldr 	q1, [%[a_ptr], #80]\n"
+
+                "uadalp	v20.4s, v12.8h\n"
+                "umull	v12.8h, v2.8b, %[b0].8b\n"
+                "uadalp	v21.4s, v13.8h\n"
+                "uadalp	v22.4s, v14.8h\n"
+                "umull	v13.8h, v2.8b, %[b1].8b\n"
+                "uadalp	v23.4s, v15.8h\n"
+                "umull	v14.8h, v2.8b, %[b2].8b\n"
+                "umull	v15.8h, v2.8b, %[b3].8b\n"
+                "umlal2	v12.8h, v2.16b, %[b0].16b\n"
+                "umlal2	v13.8h, v2.16b, %[b1].16b\n"
+                "umlal2	v14.8h, v2.16b, %[b2].16b\n"
+                "umlal2	v15.8h, v2.16b, %[b3].16b\n"
+                "ldr 	q2, [%[a_ptr], #96]\n"
+
+                "uadalp	v24.4s, v12.8h\n"
+                "umull	v12.8h, v3.8b, %[b0].8b\n"
+                "uadalp	v25.4s, v13.8h\n"
+                "uadalp	v26.4s, v14.8h\n"
+                "umull	v13.8h, v3.8b, %[b1].8b\n"
+                "uadalp	v27.4s, v15.8h\n"
+                "umull	v14.8h, v3.8b, %[b2].8b\n"
+                "umull	v15.8h, v3.8b, %[b3].8b\n"
+                "umlal2	v12.8h, v3.16b, %[b0].16b\n"
+                "umlal2	v13.8h, v3.16b, %[b1].16b\n"
+                "umlal2	v14.8h, v3.16b, %[b2].16b\n"
+                "umlal2	v15.8h, v3.16b, %[b3].16b\n"
+                "ldr 	q3, [%[a_ptr], #112]\n"
+
+                // Unroll 1
+                "uadalp	v28.4s, v12.8h\n"
+                "umull	v12.8h, v0.8b, %[b0a].8b\n"
+                "uadalp	v29.4s, v13.8h\n"
+                "uadalp	v30.4s, v14.8h\n"
+                "umull	v13.8h, v0.8b, %[b1a].8b\n"
+                "uadalp	v31.4s, v15.8h\n"
+                "umull	v14.8h, v0.8b, %[b2a].8b\n"
+                "add	%[a_ptr], %[a_ptr], #128\n"
+                "umull	v15.8h, v0.8b, %[b3a].8b\n"
+                "umlal2	v12.8h, v0.16b, %[b0a].16b\n"
+                "umlal2	v13.8h, v0.16b, %[b1a].16b\n"
+                "umlal2	v14.8h, v0.16b, %[b2a].16b\n"
+                "umlal2	v15.8h, v0.16b, %[b3a].16b\n"
+
+                "uadalp	v16.4s, v12.8h\n"
+                "umull	v12.8h, v1.8b, %[b0a].8b\n"
+                "uadalp	v17.4s, v13.8h\n"
+                "uadalp	v18.4s, v14.8h\n"
+                "umull	v13.8h, v1.8b, %[b1a].8b\n"
+                "uadalp	v19.4s, v15.8h\n"
+                "umull	v14.8h, v1.8b, %[b2a].8b\n"
+                "umull	v15.8h, v1.8b, %[b3a].8b\n"
+                "umlal2	v12.8h, v1.16b, %[b0a].16b\n"
+                "addp	v16.4s, v16.4s, v17.4s\n"
+                "umlal2	v13.8h, v1.16b, %[b1a].16b\n"
+                "addp	v17.4s, v18.4s, v19.4s\n"
+                "umlal2	v14.8h, v1.16b, %[b2a].16b\n"
+                "umlal2	v15.8h, v1.16b, %[b3a].16b\n"
+
+                "uadalp	v20.4s, v12.8h\n"
+                "umull	v12.8h, v2.8b, %[b0a].8b\n"
+                "uadalp	v21.4s, v13.8h\n"
+                "uadalp	v22.4s, v14.8h\n"
+                "umull	v13.8h, v2.8b, %[b1a].8b\n"
+                "uadalp	v23.4s, v15.8h\n"
+                "addp	v16.4s, v16.4s, v17.4s\n"
+                "umull	v14.8h, v2.8b, %[b2a].8b\n"
+                "addp	v18.4s, v20.4s, v21.4s\n"
+                "addp	v19.4s, v22.4s, v23.4s\n"
+                "umull	v15.8h, v2.8b, %[b3a].8b\n"
+                "umlal2	v12.8h, v2.16b, %[b0a].16b\n"
+                "str	q16, [%[c_ptr]]\n"
+                "umlal2	v13.8h, v2.16b, %[b1a].16b\n"
+                "umlal2	v14.8h, v2.16b, %[b2a].16b\n"
+                "umlal2	v15.8h, v2.16b, %[b3a].16b\n"
+
+                "uadalp	v24.4s, v12.8h\n"
+                "umull	v12.8h, v3.8b, %[b0a].8b\n"
+                "uadalp	v25.4s, v13.8h\n"
+                "uadalp	v26.4s, v14.8h\n"
+                "umull	v13.8h, v3.8b, %[b1a].8b\n"
+                "uadalp	v27.4s, v15.8h\n"
+                "addp	v17.4s, v18.4s, v19.4s\n"
+                "umull	v14.8h, v3.8b, %[b2a].8b\n"
+                "addp	v20.4s, v24.4s, v25.4s\n"
+                "addp	v21.4s, v26.4s, v27.4s\n"
+                "umull	v15.8h, v3.8b, %[b3a].8b\n"
+                "umlal2	v12.8h, v3.16b, %[b0a].16b\n"
+                "str	q17, [%[c_ptr], #16]\n"
+                "umlal2	v13.8h, v3.16b, %[b1a].16b\n"
+                "umlal2	v14.8h, v3.16b, %[b2a].16b\n"
+                "addp	v18.4s, v20.4s, v21.4s\n"
+                "umlal2	v15.8h, v3.16b, %[b3a].16b\n"
+                "b	3f\n"
+
+                // Detached final iteration (odd K)
+                "2:\n"
+                "umull	v14.8h, v0.8b, %[b2].8b\n"
+                "add	%[a_ptr], %[a_ptr], #64\n"
+                "umull	v15.8h, v0.8b, %[b3].8b\n"
+                "add	%[b_ptr], %[b_ptr], #64\n"
+                "umlal2	v12.8h, v0.16b, %[b0].16b\n"
+                "umlal2	v13.8h, v0.16b, %[b1].16b\n"
+                "umlal2	v14.8h, v0.16b, %[b2].16b\n"
+                "umlal2	v15.8h, v0.16b, %[b3].16b\n"
+
+                "uadalp	v16.4s, v12.8h\n"
+                "umull	v12.8h, v1.8b, %[b0].8b\n"
+                "uadalp	v17.4s, v13.8h\n"
+                "uadalp	v18.4s, v14.8h\n"
+                "umull	v13.8h, v1.8b, %[b1].8b\n"
+                "uadalp	v19.4s, v15.8h\n"
+                "umull	v14.8h, v1.8b, %[b2].8b\n"
+                "umull	v15.8h, v1.8b, %[b3].8b\n"
+                "umlal2	v12.8h, v1.16b, %[b0].16b\n"
+                "addp	v16.4s, v16.4s, v17.4s\n"
+                "umlal2	v13.8h, v1.16b, %[b1].16b\n"
+                "addp	v17.4s, v18.4s, v19.4s\n"
+                "umlal2	v14.8h, v1.16b, %[b2].16b\n"
+                "umlal2	v15.8h, v1.16b, %[b3].16b\n"
+
+                "uadalp	v20.4s, v12.8h\n"
+                "umull	v12.8h, v2.8b, %[b0].8b\n"
+                "uadalp	v21.4s, v13.8h\n"
+                "uadalp	v22.4s, v14.8h\n"
+                "umull	v13.8h, v2.8b, %[b1].8b\n"
+                "uadalp	v23.4s, v15.8h\n"
+                "addp	v16.4s, v16.4s, v17.4s\n"
+                "umull	v14.8h, v2.8b, %[b2].8b\n"
+                "addp	v18.4s, v20.4s, v21.4s\n"
+                "addp	v19.4s, v22.4s, v23.4s\n"
+                "umull	v15.8h, v2.8b, %[b3].8b\n"
+                "umlal2	v12.8h, v2.16b, %[b0].16b\n"
+                "str	q16, [%[c_ptr]]\n"
+                "umlal2	v13.8h, v2.16b, %[b1].16b\n"
+                "umlal2	v14.8h, v2.16b, %[b2].16b\n"
+                "umlal2	v15.8h, v2.16b, %[b3].16b\n"
+
+                "uadalp	v24.4s, v12.8h\n"
+                "umull	v12.8h, v3.8b, %[b0].8b\n"
+                "uadalp	v25.4s, v13.8h\n"
+                "uadalp	v26.4s, v14.8h\n"
+                "umull	v13.8h, v3.8b, %[b1].8b\n"
+                "uadalp	v27.4s, v15.8h\n"
+                "addp	v17.4s, v18.4s, v19.4s\n"
+                "umull	v14.8h, v3.8b, %[b2].8b\n"
+                "addp	v20.4s, v24.4s, v25.4s\n"
+                "addp	v21.4s, v26.4s, v27.4s\n"
+                "umull	v15.8h, v3.8b, %[b3].8b\n"
+                "umlal2	v12.8h, v3.16b, %[b0].16b\n"
+                "str	q17, [%[c_ptr], #16]\n"
+                "umlal2	v13.8h, v3.16b, %[b1].16b\n"
+                "umlal2	v14.8h, v3.16b, %[b2].16b\n"
+                "addp	v18.4s, v20.4s, v21.4s\n"
+                "umlal2	v15.8h, v3.16b, %[b3].16b\n"
+
+                "3:\n"
+
+                // Final additions
+                "uadalp	v28.4s, v12.8h\n"
+                "str	q18, [%[c_ptr], #32]\n"
+                "uadalp	v29.4s, v13.8h\n"
+                "uadalp	v30.4s, v14.8h\n"
+                "uadalp	v31.4s, v15.8h\n"
+
+                // Horizontal reduction, phase 1
+                "addp	v22.4s, v28.4s, v29.4s\n"
+                "addp	v23.4s, v30.4s, v31.4s\n"
+
+                // Horizontal reduction, phase 2
+                "addp	v19.4s, v22.4s, v23.4s\n"
+                "str	q19, [%[c_ptr], #48]\n"
+                "add	%[c_ptr], %[c_ptr], #64\n"
+
+            :
+              [a_ptr] "+r" (a_ptr), [b_ptr] "+r" (b_ptr), [c_ptr] "+r" (c_ptr),
+              [b0] "+w" (b0), [b1] "+w" (b1), [b2] "+w" (b2), [b3] "+w" (b3),
+              [b0a] "+w" (b0a), [b1a] "+w" (b1a), [b2a] "+w" (b2a), [b3a] "+w" (b3a),
+              [k] "+r" (k)
+            : [oddk] "r" (oddk)
+            : "x20", "x21", "v0","v1","v2","v3","v12","v13","v14","v15","v16","v17","v18","v19",
+              "v20","v21","v22","v23","v24","v25","v26","v27","v28","v29","v30","v31", "cc");
+        }
+    }
+}
+
+#endif // __aarch64__
diff --git a/src/core/NEON/kernels/NEGEMMLowpMatrixMultiplyKernel.cpp b/src/core/NEON/kernels/NEGEMMLowpMatrixMultiplyKernel.cpp
index 2bc251e..208a60c 100644
--- a/src/core/NEON/kernels/NEGEMMLowpMatrixMultiplyKernel.cpp
+++ b/src/core/NEON/kernels/NEGEMMLowpMatrixMultiplyKernel.cpp
@@ -49,7 +49,7 @@
 {
 Error validate_arguments(const ITensorInfo *input0, const ITensorInfo *input1, const ITensorInfo *output)
 {
-    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input0, 1, DataType::QASYMM8, DataType::S8);
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input0, 1, DataType::QASYMM8, DataType::S8, DataType::U8);
     ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input0, input1);
     ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::S32);
 
diff --git a/src/core/NEON/kernels/arm64/NEGEMMLowpAArch64Kernel.cpp b/src/core/NEON/kernels/arm64/NEGEMMLowpAArch64Kernel.cpp
index b75a8ab..db37201 100644
--- a/src/core/NEON/kernels/arm64/NEGEMMLowpAArch64Kernel.cpp
+++ b/src/core/NEON/kernels/arm64/NEGEMMLowpAArch64Kernel.cpp
@@ -39,6 +39,7 @@
 {
 #include "arm_compute/core/NEON/kernels/assembly/gemm_interleaved.hpp"
 #include "arm_compute/core/NEON/kernels/assembly/kernels/a64_gemm_s8_4x4.hpp"
+#include "arm_compute/core/NEON/kernels/assembly/kernels/a64_gemm_u8_4x4.hpp"
 } // namespace arm_compute
 
 #include <arm_neon.h>
@@ -50,10 +51,101 @@
 
 namespace arm_compute
 {
+NEGEMMLowpAArch64Kernel::NEGEMMLowpAArch64Kernel()
+    : _func(nullptr)
+{
+}
+
+void gemm_interleaved_s8(const ITensor *input0, const ITensor *input1, ITensor *output, ITensor *workspace, float alpha, float beta, bool transform_0, bool transform_1, const Window &window,
+                         const ThreadInfo &info)
+{
+    const int lda = input0->info()->strides_in_bytes().y();
+    const int ldb = input1->info()->strides_in_bytes().y();
+    const int ldc = output->info()->strides_in_bytes().y() / sizeof(int32_t);
+
+    const auto in1_ptr = reinterpret_cast<const int8_t *>(input1->buffer());
+
+    const int M = std::min(output->info()->tensor_shape().y(), static_cast<size_t>(window.y().end())) - window.y().start();
+    const int N = output->info()->tensor_shape().x();
+    const int K = input0->info()->tensor_shape().x();
+
+    // Only iterate over batches
+    Window win(window);
+    win.set(0, Window::Dimension(0, 1, 1));
+    win.set(1, Window::Dimension(0, 1, 1));
+
+    Iterator in0(input0, window);
+    Iterator out(output, window);
+
+    GemmInterleaved<gemm_s8_4x4, int8_t, int32_t> gemm(&info.cpu_info, M, N, K, !transform_1, !transform_1);
+
+    constexpr size_t alignment      = 4096;
+    const size_t     offset         = (gemm.get_working_size() + alignment - 1) * info.thread_id;
+    void            *_workspace     = workspace->buffer() + offset;
+    size_t           workspace_size = workspace->info()->total_size();
+
+    if(support::cpp11::align(alignment, gemm.get_working_size(), _workspace, workspace_size) == nullptr)
+    {
+        ARM_COMPUTE_ERROR("Not enough space to align buffer!");
+    }
+
+    execute_window_loop(win, [&](const Coordinates & id)
+    {
+        gemm.execute(reinterpret_cast<const int8_t *>(in0.ptr()), lda,
+                     reinterpret_cast<const int8_t *>(in1_ptr), ldb,
+                     reinterpret_cast<int32_t *>(out.ptr()), ldc,
+                     alpha, beta, _workspace);
+    },
+    in0, out);
+}
+
+void gemm_interleaved_u8(const ITensor *input0, const ITensor *input1, ITensor *output, ITensor *workspace, float alpha, float beta, bool transform_0, bool transform_1, const Window &window,
+                         const ThreadInfo &info)
+{
+    const int lda = input0->info()->strides_in_bytes().y();
+    const int ldb = input1->info()->strides_in_bytes().y();
+    const int ldc = output->info()->strides_in_bytes().y() / sizeof(uint32_t);
+
+    const auto in1_ptr = reinterpret_cast<const uint8_t *>(input1->buffer());
+
+    const int M = std::min(output->info()->tensor_shape().y(), static_cast<size_t>(window.y().end())) - window.y().start();
+    const int N = output->info()->tensor_shape().x();
+    const int K = input0->info()->tensor_shape().x();
+
+    // Only iterate over batches
+    Window win(window);
+    win.set(0, Window::Dimension(0, 1, 1));
+    win.set(1, Window::Dimension(0, 1, 1));
+
+    Iterator in0(input0, window);
+    Iterator out(output, window);
+
+    GemmInterleaved<gemm_u8_4x4, uint8_t, uint32_t> gemm(&info.cpu_info, M, N, K, !transform_1, !transform_1);
+
+    constexpr size_t alignment      = 4096;
+    const size_t     offset         = (gemm.get_working_size() + alignment - 1) * info.thread_id;
+    void            *_workspace     = workspace->buffer() + offset;
+    size_t           workspace_size = workspace->info()->total_size();
+
+    if(support::cpp11::align(alignment, gemm.get_working_size(), _workspace, workspace_size) == nullptr)
+    {
+        ARM_COMPUTE_ERROR("Not enough space to align buffer!");
+    }
+
+    execute_window_loop(win, [&](const Coordinates & id)
+    {
+        gemm.execute(reinterpret_cast<const uint8_t *>(in0.ptr()), lda,
+                     reinterpret_cast<const uint8_t *>(in1_ptr), ldb,
+                     reinterpret_cast<uint32_t *>(out.ptr()), ldc,
+                     alpha, beta, _workspace);
+    },
+    in0, out);
+}
+
 void NEGEMMLowpAArch64Kernel::internal_configure(const ITensor *input0, const ITensor *input1, ITensor *output, ITensor *workspace, float alpha, float beta, bool transform_0, bool transform_1)
 {
-    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input0, 1, DataType::S8);
-    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::S32);
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input0, 1, DataType::S8, DataType::U8);
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::S32, DataType::U32);
     ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input0, input1);
 
     _input0      = input0;
@@ -65,6 +157,19 @@
     _transform_0 = transform_0;
     _transform_1 = transform_1;
 
+    switch(input0->info()->data_type())
+    {
+        case DataType::S8:
+            _func = &gemm_interleaved_s8;
+            break;
+        case DataType::U8:
+            _func = &gemm_interleaved_u8;
+            break;
+        default:
+            ARM_COMPUTE_ERROR("Element size not supported");
+            break;
+    }
+
     // Configure kernel window
     Window win = calculate_max_window(*output->info());
 
@@ -85,45 +190,9 @@
 {
     ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
     ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window);
+    ARM_COMPUTE_ERROR_ON(_func == nullptr);
 
-    const int lda = _input0->info()->strides_in_bytes().y();
-    const int ldb = _input1->info()->strides_in_bytes().y();
-    const int ldc = _output->info()->strides_in_bytes().y() / sizeof(int32_t);
-
-    const auto in1_ptr = reinterpret_cast<const int8_t *>(_input1->buffer());
-
-    const int M = std::min(_output->info()->tensor_shape().y(), static_cast<size_t>(window.y().end())) - window.y().start();
-    const int N = _output->info()->tensor_shape().x();
-    const int K = _input0->info()->tensor_shape().x();
-
-    // Only iterate over batches
-    Window win(window);
-    win.set(0, Window::Dimension(0, 1, 1));
-    win.set(1, Window::Dimension(0, 1, 1));
-
-    Iterator in0(_input0, window);
-    Iterator out(_output, window);
-
-    GemmInterleaved<gemm_s8_4x4, int8_t, int32_t> gemm(&info.cpu_info, M, N, K, !_transform_1, !_transform_1);
-
-    constexpr size_t alignment      = 4096;
-    const size_t     offset         = (gemm.get_working_size() + alignment - 1) * info.thread_id;
-    void            *workspace      = _workspace->buffer() + offset;
-    size_t           workspace_size = _workspace->info()->total_size();
-
-    if(support::cpp11::align(alignment, gemm.get_working_size(), workspace, workspace_size) == nullptr)
-    {
-        ARM_COMPUTE_ERROR("Not enough space to align buffer!");
-    }
-
-    execute_window_loop(win, [&](const Coordinates & id)
-    {
-        gemm.execute(reinterpret_cast<const int8_t *>(in0.ptr()), lda,
-                     reinterpret_cast<const int8_t *>(in1_ptr), ldb,
-                     reinterpret_cast<int32_t *>(out.ptr()), ldc,
-                     _alpha, _beta, workspace);
-    },
-    in0, out);
+    (*_func)(_input0, _input1, _output, _workspace, _alpha, _beta, _transform_0, _transform_1, window, info);
 }
 } // namespace arm_compute
 #endif /* ARM_COMPUTE_AARCH64_V8A */
diff --git a/src/runtime/NEON/functions/NEGEMMLowpAssemblyMatrixMultiplyCore.cpp b/src/runtime/NEON/functions/NEGEMMLowpAssemblyMatrixMultiplyCore.cpp
index 708daeb..1bf437e 100644
--- a/src/runtime/NEON/functions/NEGEMMLowpAssemblyMatrixMultiplyCore.cpp
+++ b/src/runtime/NEON/functions/NEGEMMLowpAssemblyMatrixMultiplyCore.cpp
@@ -43,6 +43,7 @@
 #include "arm_compute/core/NEON/kernels/assembly/gemm_interleaved.hpp"
 #include "arm_compute/core/NEON/kernels/assembly/kernels/a64_gemm_s8_12x8.hpp"
 #include "arm_compute/core/NEON/kernels/assembly/kernels/a64_gemm_s8_4x4.hpp"
+#include "arm_compute/core/NEON/kernels/assembly/kernels/a64_gemm_u8_4x4.hpp"
 
 } // namespace arm_compute
 
@@ -55,8 +56,8 @@
 
 void NEGEMMLowpAssemblyMatrixMultiplyCore::configure(const ITensor *a, const ITensor *b, ITensor *output)
 {
-    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(a, 1, DataType::S8);
-    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::S32);
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(a, 1, DataType::U8, DataType::S8);
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::U32, DataType::S32);
     ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(a, b);
     ARM_COMPUTE_ERROR_ON_MSG((a)->info()->dimension(0) != (b)->info()->dimension(1), "The product AB is defined only if the number of columns in A is equal to the number of rows in B");
     ARM_COMPUTE_ERROR_ON_MSG((a)->info()->dimension(1) != (output)->info()->dimension(1), "The output matrix must have the same number of rows as the matrix A");
@@ -92,9 +93,25 @@
 #elif defined(ARM_COMPUTE_AARCH64_V8A)
     if(1)
     {
-        // Configure matrix multiply kernel
-        GemmInterleaved<gemm_s8_4x4, int8_t, int32_t> gemm(&ci, M, N, K, false, false);
-        _workspace.allocator()->init(TensorInfo(TensorShape{ (gemm.get_working_size() + workspace_alignment - 1) * NEScheduler::get().num_threads() }, 1, DataType::U8));
+        switch(a->info()->data_type())
+        {
+            case DataType::S8:
+            {
+                // Configure matrix multiply kernel
+                GemmInterleaved<gemm_s8_4x4, int8_t, int32_t> gemm(&ci, M, N, K, false, false);
+                _workspace.allocator()->init(TensorInfo(TensorShape{ (gemm.get_working_size() + workspace_alignment - 1) * NEScheduler::get().num_threads() }, 1, DataType::U8));
+            }
+            break;
+            case DataType::U8:
+            {
+                // Configure matrix multiply kernel
+                GemmInterleaved<gemm_u8_4x4, uint8_t, uint32_t> gemm(&ci, M, N, K, false, false);
+                _workspace.allocator()->init(TensorInfo(TensorShape{ (gemm.get_working_size() + workspace_alignment - 1) * NEScheduler::get().num_threads() }, 1, DataType::U8));
+            }
+            break;
+            default:
+                ARM_COMPUTE_ERROR("Datatype not supported");
+        }
         _memory_group.manage(&_workspace);
         // Configure matrix multiplication kernel
         auto k = arm_compute::support::cpp14::make_unique<NEGEMMLowpAArch64Kernel>();
diff --git a/tests/validation/CPP/GEMMLowp.cpp b/tests/validation/CPP/GEMMLowp.cpp
index bf002cf..35b8a64 100644
--- a/tests/validation/CPP/GEMMLowp.cpp
+++ b/tests/validation/CPP/GEMMLowp.cpp
@@ -63,19 +63,21 @@
 }
 } // namespace
 
-template <typename T>
-SimpleTensor<int32_t> gemmlowp_matrix_multiply_core(const SimpleTensor<T> &a, const SimpleTensor<T> &b, int32_t a_offset, int32_t b_offset)
+template <typename T_out, typename T_in>
+SimpleTensor<T_out> gemmlowp_matrix_multiply_core(const SimpleTensor<T_in> &a, const SimpleTensor<T_in> &b, int32_t a_offset, int32_t b_offset)
 {
-    TensorShape shape(b.shape()[0], a.shape()[1]);
+    static_assert(std::is_same<typename std::decay<T_out>::type, int32_t>::value, "Only int32_t is allowed for the output");
 
-    SimpleTensor<int32_t> c(shape, DataType::S32);
+    TensorShape         shape(b.shape()[0], a.shape()[1]);
+    DataType            dt = std::is_same<T_out, int32_t>::value ? DataType::S32 : DataType::U32;
+    SimpleTensor<T_out> c(shape, dt);
 
     const int K       = a.shape().x();
     const int b_width = b.shape().x();
     const int rows    = c.shape().y(); //M
     const int cols    = c.shape().x(); //N
 
-    std::vector<int32_t> acc;
+    std::vector<T_out> acc;
     acc.resize(cols);
 
     for(int i = 0; i < rows; ++i)
@@ -86,11 +88,11 @@
         }
         for(int k = 0; k < K; ++k)
         {
-            const int32_t tmp_a = a_offset + static_cast<int32_t>(a[k + i * K]);
+            const T_out tmp_a = a_offset + static_cast<T_out>(a[k + i * K]);
             for(int j = 0; j < b_width; ++j)
             {
-                const int32_t tmp_b       = b_offset + static_cast<int32_t>(b[j + k * b_width]);
-                const int32_t mult_as_int = tmp_a * tmp_b;
+                const T_out tmp_b       = b_offset + static_cast<T_out>(b[j + k * b_width]);
+                const T_out mult_as_int = tmp_a * tmp_b;
                 acc[j] += mult_as_int;
             }
         }
@@ -104,9 +106,10 @@
 }
 
 // used to validate assembly kernels which don't know anything about offsets
-SimpleTensor<int32_t> gemmlowp(const SimpleTensor<int8_t> &a, const SimpleTensor<int8_t> &b)
+template <typename T1, typename T2>
+SimpleTensor<T1> gemmlowp(const SimpleTensor<T2> &a, const SimpleTensor<T2> &b)
 {
-    return gemmlowp_matrix_multiply_core(a, b, 0, 0);
+    return gemmlowp_matrix_multiply_core<T1, T2>(a, b, 0, 0);
 }
 
 template <typename T>
@@ -130,11 +133,14 @@
     return dst;
 }
 
-template SimpleTensor<int32_t> gemmlowp_matrix_multiply_core(const SimpleTensor<uint8_t> &a, const SimpleTensor<uint8_t> &b, int32_t a_offset, int32_t b_offset);
 template SimpleTensor<uint8_t> gemmlowp_quantize_down_int32_to_uint8_scale(const SimpleTensor<int32_t> &a, int32_t result_offset, int32_t result_mult_int, int32_t result_shift, int32_t min,
                                                                            int32_t max);
 template SimpleTensor<uint8_t> gemmlowp_quantize_down_int32_to_uint8_scale(const SimpleTensor<int32_t> &a, const SimpleTensor<int32_t> &b, int32_t result_offset, int32_t result_mult_int,
                                                                            int32_t result_shift, int32_t min, int32_t max);
+template SimpleTensor<int32_t> gemmlowp_matrix_multiply_core(const SimpleTensor<int8_t> &a, const SimpleTensor<int8_t> &b, int32_t a_offset, int32_t b_offset);
+template SimpleTensor<int32_t> gemmlowp_matrix_multiply_core(const SimpleTensor<uint8_t> &a, const SimpleTensor<uint8_t> &b, int32_t a_offset, int32_t b_offset);
+template SimpleTensor<int32_t> gemmlowp(const SimpleTensor<int8_t> &a, const SimpleTensor<int8_t> &b);
+template SimpleTensor<int32_t> gemmlowp(const SimpleTensor<uint8_t> &a, const SimpleTensor<uint8_t> &b);
 } // namespace reference
 } // namespace validation
 } // namespace test
diff --git a/tests/validation/CPP/GEMMLowp.h b/tests/validation/CPP/GEMMLowp.h
index ee33d8e..6c72b56 100644
--- a/tests/validation/CPP/GEMMLowp.h
+++ b/tests/validation/CPP/GEMMLowp.h
@@ -35,13 +35,16 @@
 {
 namespace reference
 {
-SimpleTensor<int32_t> gemmlowp(const SimpleTensor<int8_t> &a, const SimpleTensor<int8_t> &b);
-
-template <typename T>
-SimpleTensor<int32_t> gemmlowp_matrix_multiply_core(const SimpleTensor<T> &a, const SimpleTensor<T> &b, int32_t a_offset, int32_t b_offset);
-
 template <typename T>
 SimpleTensor<uint8_t> gemmlowp_quantize_down_int32_to_uint8_scale(const SimpleTensor<T> &in, int32_t result_offset, int32_t result_mult_int, int32_t result_shift, int32_t min = 0, int32_t max = 0);
+template <typename T1, typename T2>
+SimpleTensor<T1> gemmlowp_matrix_multiply_core(const SimpleTensor<T2> &a, const SimpleTensor<T2> &b, int32_t a_offset, int32_t b_offset);
+
+template <typename T>
+SimpleTensor<uint8_t> gemmlowp_quantize_down_int32_to_uint8_scale(const SimpleTensor<T> &in, int32_t result_offset, int32_t result_mult_int, int32_t result_shift);
+
+template <typename T1, typename T2>
+SimpleTensor<T1> gemmlowp(const SimpleTensor<T2> &a, const SimpleTensor<T2> &b);
 
 template <typename T>
 SimpleTensor<uint8_t> gemmlowp_quantize_down_int32_to_uint8_scale(const SimpleTensor<T> &in, const SimpleTensor<T> &bias, int32_t result_offset, int32_t result_mult_int, int32_t result_shift,
diff --git a/tests/validation/NEON/GEMMLowp.cpp b/tests/validation/NEON/GEMMLowp.cpp
index 1418578..6366223 100644
--- a/tests/validation/NEON/GEMMLowp.cpp
+++ b/tests/validation/NEON/GEMMLowp.cpp
@@ -58,14 +58,27 @@
 
 TEST_SUITE(NEON)
 TEST_SUITE(ASSEMBLY_MATRIX_MULTIPLY)
-using NEGEMMAssemblyFixture = GEMMLowpAssemblyFixture<Tensor, Accessor, NEGEMMLowpAssemblyMatrixMultiplyCore>;
-FIXTURE_DATA_TEST_CASE(RunSmall, NEGEMMAssemblyFixture, framework::DatasetMode::PRECOMMIT, data_matrix_multiply)
+
+using NEGEMMAssemblyFixture_S8 = GEMMLowpAssemblyFixture<Tensor, Accessor, NEGEMMLowpAssemblyMatrixMultiplyCore, int8_t>;
+using NEGEMMAssemblyFixture_U8 = GEMMLowpAssemblyFixture<Tensor, Accessor, NEGEMMLowpAssemblyMatrixMultiplyCore, uint8_t>;
+
+TEST_SUITE(S8)
+FIXTURE_DATA_TEST_CASE(RunSmall, NEGEMMAssemblyFixture_S8, framework::DatasetMode::PRECOMMIT, data_matrix_multiply)
 {
     // Validate output
     validate(Accessor(_target), _reference);
 }
 TEST_SUITE_END()
 
+TEST_SUITE(U8)
+FIXTURE_DATA_TEST_CASE(RunSmall, NEGEMMAssemblyFixture_U8, framework::DatasetMode::PRECOMMIT, data_matrix_multiply)
+{
+    // Validate output
+    validate(Accessor(_target), _reference);
+}
+TEST_SUITE_END()
+TEST_SUITE_END()
+
 TEST_SUITE(GEMMLowp)
 
 TEST_SUITE(INTERLEAVE_BLOCKED)
diff --git a/tests/validation/fixtures/GEMMLowpAssemblyFixture.h b/tests/validation/fixtures/GEMMLowpAssemblyFixture.h
index a258744..38e08f7 100644
--- a/tests/validation/fixtures/GEMMLowpAssemblyFixture.h
+++ b/tests/validation/fixtures/GEMMLowpAssemblyFixture.h
@@ -42,7 +42,7 @@
 {
 namespace validation
 {
-template <typename TensorType, typename AccessorType, typename FunctionType>
+template <typename TensorType, typename AccessorType, typename FunctionType, typename T2>
 class GEMMLowpAssemblyFixture : public framework::Fixture
 {
 public:
@@ -66,9 +66,11 @@
 
     TensorType compute_target(const TensorShape &shape_a, const TensorShape &shape_b, const TensorShape &shape_c)
     {
+        DataType dt_in = std::is_same<T2, int8_t>::value ? DataType::S8 : DataType::U8;
+
         // Create tensors
-        TensorType a = create_tensor<TensorType>(shape_a, DataType::S8, 1);
-        TensorType b = create_tensor<TensorType>(shape_b, DataType::S8, 1);
+        TensorType a = create_tensor<TensorType>(shape_a, dt_in, 1);
+        TensorType b = create_tensor<TensorType>(shape_b, dt_in, 1);
         TensorType c = create_tensor<TensorType>(shape_c, DataType::S32, 1);
 
         // Create and configure function
@@ -89,8 +91,16 @@
         ARM_COMPUTE_EXPECT(!c.info()->is_resizable(), framework::LogLevel::ERRORS);
 
         // Fill tensors
-        fill(AccessorType(a), 0, -128, 127);
-        fill(AccessorType(b), 1, -128, 127);
+        if(dt_in == DataType::S8)
+        {
+            fill(AccessorType(a), 0, -128, 127);
+            fill(AccessorType(b), 1, -128, 127);
+        }
+        else
+        {
+            fill(AccessorType(a), 0, 0, 128);
+            fill(AccessorType(b), 1, 0, 128);
+        }
         fill(AccessorType(c), 2, 0, 0);
 
         // Compute GEMM function
@@ -100,15 +110,25 @@
 
     SimpleTensor<int32_t> compute_reference(const TensorShape &shape_a, const TensorShape &shape_b, const TensorShape &shape_c)
     {
+        DataType dt = std::is_same<T2, int8_t>::value ? DataType::S8 : DataType::U8;
+
         // Create reference
-        SimpleTensor<int8_t> a{ shape_a, DataType::S8, 1 };
-        SimpleTensor<int8_t> b{ shape_b, DataType::S8, 1 };
+        SimpleTensor<T2> a{ shape_a, dt, 1 };
+        SimpleTensor<T2> b{ shape_b, dt, 1 };
 
         // Fill reference
-        fill(a, 0, -128, 127);
-        fill(b, 1, -128, 127);
+        if(dt == DataType::S8)
+        {
+            fill(a, 0, -128, 127);
+            fill(b, 1, -128, 127);
+        }
+        else
+        {
+            fill(a, 0, 0, 128);
+            fill(b, 1, 0, 128);
+        }
 
-        return reference::gemmlowp(a, b);
+        return reference::gemmlowp<int32_t, T2>(a, b);
     }
 
     TensorType            _target{};
diff --git a/tests/validation/fixtures/GEMMLowpFixture.h b/tests/validation/fixtures/GEMMLowpFixture.h
index a99e932..60b89bc 100644
--- a/tests/validation/fixtures/GEMMLowpFixture.h
+++ b/tests/validation/fixtures/GEMMLowpFixture.h
@@ -110,7 +110,7 @@
         fill(a, 0);
         fill(b, 1);
 
-        return reference::gemmlowp_matrix_multiply_core<uint8_t>(a, b, a_offset, b_offset);
+        return reference::gemmlowp_matrix_multiply_core<int32_t, uint8_t>(a, b, a_offset, b_offset);
     }
 
     TensorType            _target{};