COMPMID-3874: Create ArithmeticAddition SVE/SVE2

Change-Id: I4ec7561a7f6a42a22b8187968ae302dbe75023bc
Signed-off-by: Michalis Spyrou <michalis.spyrou@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/4753
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Sang-Hoon Park <sang-hoon.park@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>

22 files changed, 2027 insertions, 849 deletions

diff --git a/Android.bp b/Android.bp
index 1032950f3e..6e9756ec96 100644
--- a/Android.bp
+++ b/Android.bp
@@ -353,6 +353,14 @@ cc_library_static {
         "src/core/NEON/kernels/activation/impl/SVE/qasymm8.cpp",
         "src/core/NEON/kernels/activation/impl/SVE/qasymm8_signed.cpp",
         "src/core/NEON/kernels/activation/impl/SVE/qsymm16.cpp",
+        "src/core/NEON/kernels/arithmetic_addition/impl/NEON/integer.cpp",
+        "src/core/NEON/kernels/arithmetic_addition/impl/NEON/qasymm8.cpp",
+        "src/core/NEON/kernels/arithmetic_addition/impl/NEON/qasymm8_signed.cpp",
+        "src/core/NEON/kernels/arithmetic_addition/impl/NEON/qsymm16.cpp",
+        "src/core/NEON/kernels/arithmetic_addition/impl/SVE/integer.cpp",
+        "src/core/NEON/kernels/arithmetic_addition/impl/SVE/qasymm8.cpp",
+        "src/core/NEON/kernels/arithmetic_addition/impl/SVE/qasymm8_signed.cpp",
+        "src/core/NEON/kernels/arithmetic_addition/impl/SVE/qsymm16.cpp",
         "src/core/NEON/kernels/arm_gemm/gemm_bf16.cpp",
         "src/core/NEON/kernels/arm_gemm/gemm_fp16.cpp",
         "src/core/NEON/kernels/arm_gemm/gemm_fp32.cpp",
diff --git a/SConscript b/SConscript
index 2b70ca18b8..724208c306 100644
--- a/SConscript
+++ b/SConscript
@@ -256,6 +256,8 @@ if env['neon']:
         core_files += Glob('src/core/NEON/kernels/*/impl/*/qasymm8_signed.cpp')
     if any(i in env['data_type_support'] for i in ['all', 'qsymm16']):
         core_files += Glob('src/core/NEON/kernels/*/impl/*/qsymm16.cpp')
+    if any(i in env['data_type_support'] for i in ['all', 'integer']):
+        core_files += Glob('src/core/NEON/kernels/*/impl/*/integer.cpp')
 
     runtime_files += Glob('src/runtime/NEON/*.cpp')
     runtime_files += Glob('src/runtime/NEON/functions/*.cpp')
diff --git a/SConstruct b/SConstruct
index 6b20ac2177..e19d855210 100644
--- a/SConstruct
+++ b/SConstruct
@@ -66,7 +66,7 @@ vars.AddVariables(
     PathVariable("linker_script", "Use an external linker script", "", PathVariable.PathAccept),
     PathVariable("external_tests_dir", "Add examples, benchmarks and tests to the tests suite", "", PathVariable.PathAccept),
     ListVariable("custom_options", "Custom options that can be used to turn on/off features", "none", ["disable_mmla_fp"]),
-    ListVariable("data_type_support", "Enable a list of data types to support", "all", ["qasymm8", "qasymm8_signed", "qsymm16", "fp16", "fp32"]),
+    ListVariable("data_type_support", "Enable a list of data types to support", "all", ["qasymm8", "qasymm8_signed", "qsymm16", "fp16", "fp32", "integer"]),
     ("toolchain_prefix", "Override the toolchain prefix", ""),
     ("compiler_prefix", "Override the compiler prefix", ""),
     ("extra_cxx_flags", "Extra CXX flags to be appended to the build command", ""),
@@ -306,6 +306,8 @@ if env['data_type_support']:
         env.Append(CXXFLAGS = ['-DENABLE_QASYMM8_SIGNED_KERNELS'])
     if any(i in env['data_type_support'] for i in ['all', 'qsymm16']):
         env.Append(CXXFLAGS = ['-DENABLE_QSYMM16_KERNELS'])
+    if any(i in env['data_type_support'] for i in ['all', 'integer']):
+        env.Append(CXXFLAGS = ['-DENABLE_INTEGER_KERNELS'])
 
 if env['standalone']:
     env.Append(CXXFLAGS = ['-fPIC'])
diff --git a/src/core/NEON/SVEMath.h b/src/core/NEON/SVEMath.h
index 490759c789..2b30e20e8d 100644
--- a/src/core/NEON/SVEMath.h
+++ b/src/core/NEON/SVEMath.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2020 Arm Limited.
+ * Copyright (c) 2020-2021 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -36,7 +36,7 @@ namespace arm_compute
 {
 /** Calculate exponent.
  *
- * @param[in] pg  Input reciprocal.
+ * @param[in] pg  Input predicate.
  * @param[in] val Input vector value in F32 format.
  *
  * @return The calculated exponent.
@@ -45,7 +45,7 @@ svfloat32_t svexp_f32_z(svbool_t pg, svfloat32_t val);
 
 /** Calculate reciprocal.
  *
- * @param[in] pg Input reciprocal.
+ * @param[in] pg Input predicate.
  * @param[in] x  Input value.
  *
  * @return The calculated reciprocal.
@@ -54,7 +54,7 @@ svfloat32_t svinv_f32_z(svbool_t pg, svfloat32_t x);
 
 /** Calculate logarithm
  *
- * @param[in] pg Input reciprocal.
+ * @param[in] pg Input predicate.
  * @param[in] x  Input vector value in F32 format.
  *
  * @return The calculated logarithm.
@@ -67,7 +67,7 @@ svfloat32_t svlog_f32_z(svbool_t pg, svfloat32_t x);
  *
  * @note We clamp x to [-5,5] to avoid overflowing issues.
  *
- * @param[in] pg  Input reciprocal.
+ * @param[in] pg  Input predicate.
  * @param[in] val Input vector value in F32 format.
  *
  * @return The calculated Hyperbolic Tangent.
@@ -80,7 +80,7 @@ svfloat32_t svtanh_f32_z(svbool_t pg, svfloat32_t val);
  *
  * @note We clamp x to [-5,5] to avoid overflowing issues.
  *
- * @param[in] pg  Input reciprocal.
+ * @param[in] pg  Input predicate.
  * @param[in] val Input vector value in F16 format.
  *
  * @return The calculated Hyperbolic Tangent.
@@ -89,7 +89,7 @@ svfloat16_t svtanh_f16_z(svbool_t pg, svfloat16_t val);
 
 /** Calculate exponential
  *
- * @param[in] pg Input reciprocal.
+ * @param[in] pg Input predicate.
  * @param[in] x  Input vector value in F16 format.
  *
  * @return The calculated exponent.
@@ -98,7 +98,7 @@ svfloat16_t svexp_f16_z(svbool_t pg, svfloat16_t x);
 
 /** Calculate reciprocal.
  *
- * @param[in] pg Input reciprocal.
+ * @param[in] pg Input predicate.
  * @param[in] x  Input value.
  *
  * @return The calculated reciprocal.
@@ -107,7 +107,7 @@ svfloat16_t svinv_f16_z(svbool_t pg, svfloat16_t x);
 
 /** Calculate logarithm
  *
- * @param[in] pg Input reciprocal.
+ * @param[in] pg Input predicate.
  * @param[in] x  Input vector value in F32 format.
  *
  * @return The calculated logarithm.
@@ -116,7 +116,7 @@ svfloat16_t svlog_f16_z(svbool_t pg, svfloat16_t x);
 
 /** Calculate inverse square root.
  *
- * @param[in] pg  Input reciprocal.
+ * @param[in] pg  Input predicate.
  * @param[in] val Input value.
  *
  * @return The calculated inverse square root.
@@ -132,7 +132,7 @@ inline VectorType svinvsqrt(svbool_t pg, VectorType val)
 
 /** Calculate sine.
  *
- * @param[in] pg  Input reciprocal.
+ * @param[in] pg  Input predicate.
  * @param[in] val Input vector value in radians, F32 format.
  *
  * @return The calculated sine.
@@ -141,7 +141,7 @@ svfloat32_t svsin_f32_z(svbool_t pg, svfloat32_t val);
 
 /** Calculate sine.
  *
- * @param[in] pg  Input reciprocal.
+ * @param[in] pg  Input predicate.
  * @param[in] val Input vector value in radians, F16 format.
  *
  * @return The calculated sine.
@@ -152,7 +152,7 @@ svfloat16_t svsin_f16_z(svbool_t pg, svfloat16_t val);
  *
  * pow(x,n) = e^(n*log(x))
  *
- * @param[in] pg Input reciprocal.
+ * @param[in] pg Input predicate.
  * @param[in] a  Input vector value in F32 format.
  * @param[in] b  Powers to raise the input to.
  *
@@ -164,7 +164,7 @@ svfloat32_t svpow_f32_z(svbool_t pg, svfloat32_t a, svfloat32_t b);
  *
  * pow(x,n) = e^(n*log(x))
  *
- * @param[in] pg Input reciprocal.
+ * @param[in] pg Input predicate.
  * @param[in] a  Input vector value in F16 format.
  * @param[in] b  Powers to raise the input to.
  *
diff --git a/src/core/NEON/SVEMath.inl b/src/core/NEON/SVEMath.inl
index 86592f6dc3..fbf90f9b04 100644
--- a/src/core/NEON/SVEMath.inl
+++ b/src/core/NEON/SVEMath.inl
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2020 Arm Limited.
+ * Copyright (c) 2020-2021 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
diff --git a/src/core/NEON/kernels/NEArithmeticAdditionKernel.cpp b/src/core/NEON/kernels/NEArithmeticAdditionKernel.cpp
index aa7af54e9c..f706ee5694 100644
--- a/src/core/NEON/kernels/NEArithmeticAdditionKernel.cpp
+++ b/src/core/NEON/kernels/NEArithmeticAdditionKernel.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020 Arm Limited.
+ * Copyright (c) 2016-2021 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -28,7 +28,11 @@
 #include "arm_compute/core/ITensor.h"
 #include "arm_compute/core/Validate.h"
 #include "src/core/CPP/Validate.h"
+#include "src/core/NEON/kernels/arithmetic_addition/impl/NEON/list.h"
+#include "src/core/NEON/kernels/arithmetic_addition/impl/SVE/list.h"
 #include "src/core/NEON/wrapper/wrapper.h"
+#include "src/core/common/Registrars.h"
+#include "src/core/common/StdTypes.h"
 #include "src/core/helpers/AutoConfiguration.h"
 #include "src/core/helpers/WindowHelpers.h"
 
@@ -39,788 +43,156 @@ namespace arm_compute
 {
 namespace
 {
-template <typename T>
-void add_same(const ITensor *in1, const ITensor *in2, ITensor *out, const ConvertPolicy policy, const Window &window)
+struct ArithmeticAdditionSelectorData
 {
-    /** NEON vector tag type. */
-    using ExactTagType = typename wrapper::traits::neon_bitvector_tag_t<T, wrapper::traits::BitWidth::W128>;
+    DataType dt1;
+    DataType dt2;
+    DataType dt3;
+};
 
-    // Create input windows
-    Window input1_win = window.broadcast_if_dimension_le_one(in1->info()->tensor_shape());
-    Window input2_win = window.broadcast_if_dimension_le_one(in2->info()->tensor_shape());
+using ArithmeticAdditionSelectorPtr = std::add_pointer<bool(const ArithmeticAdditionSelectorData &data)>::type;
 
-    // Clear X Dimension on execution window as we handle manually
-    Window win = window;
-    win.set(Window::DimX, Window::Dimension(0, 1, 1));
-
-    constexpr int window_step_x         = 16 / sizeof(T);
-    const auto    window_start_x        = static_cast<int>(window.x().start());
-    const auto    window_end_x          = static_cast<int>(window.x().end());
-    const bool    is_broadcast_across_x = in1->info()->tensor_shape().x() != in2->info()->tensor_shape().x();
+struct ArithmeticAdditionKernel
+{
+    const char                                             *name;
+    const ArithmeticAdditionSelectorPtr                     is_selected;
+    NEArithmeticAdditionKernel::ArithmeticAdditionKernelPtr ukernel;
+};
 
-    if(is_broadcast_across_x)
+static const ArithmeticAdditionKernel available_kernels[] =
+{
+#if defined(__ARM_FEATURE_SVE)
     {
-        const bool     is_broadcast_input_2 = input2_win.x().step() == 0;
-        Window         broadcast_win        = is_broadcast_input_2 ? input2_win : input1_win;
-        Window         non_broadcast_win    = !is_broadcast_input_2 ? input2_win : input1_win;
-        const ITensor *broadcast_tensor     = is_broadcast_input_2 ? in2 : in1;
-        const ITensor *non_broadcast_tensor = !is_broadcast_input_2 ? in2 : in1;
-
-        // Clear X Dimension on execution window as we handle manually
-        non_broadcast_win.set(Window::DimX, Window::Dimension(0, 1, 1));
-
-        Iterator broadcast_input(broadcast_tensor, broadcast_win);
-        Iterator non_broadcast_input(non_broadcast_tensor, non_broadcast_win);
-        Iterator output(out, win);
-
-        execute_window_loop(win, [&](const Coordinates &)
-        {
-            const auto non_broadcast_input_ptr = reinterpret_cast<const T *>(non_broadcast_input.ptr());
-            const auto output_ptr              = reinterpret_cast<T *>(output.ptr());
-
-            const T    broadcast_value     = *reinterpret_cast<const T *>(broadcast_input.ptr());
-            const auto broadcast_value_vec = wrapper::vdup_n(broadcast_value, ExactTagType{});
-
-            // Compute S elements per iteration
-            int x = window_start_x;
-            for(; x <= (window_end_x - window_step_x); x += window_step_x)
-            {
-                const auto non_broadcast_v = wrapper::vloadq(non_broadcast_input_ptr + x);
-                const auto res             = (policy == ConvertPolicy::SATURATE) ? wrapper::vqadd(broadcast_value_vec, non_broadcast_v) : wrapper::vadd(broadcast_value_vec, non_broadcast_v);
-                wrapper::vstore(output_ptr + x, res);
-            }
-
-            // Compute left-over elements
-            for(; x < window_end_x; ++x)
-            {
-                const auto non_broadcast_v = *(non_broadcast_input_ptr + x);
-                *(output_ptr + x)          = (policy == ConvertPolicy::SATURATE) ? wrapper::add_sat(broadcast_value, non_broadcast_v) : broadcast_value + non_broadcast_v;
-            }
-        },
-        broadcast_input, non_broadcast_input, output);
-    }
-    else
+        "arithmetic_addition_same_sve",
+        [](const ArithmeticAdditionSelectorData & data) { return ((data.dt1 == data.dt2) && (data.dt1 == DataType::F32)); },
+        REGISTER_FP32_SVE(arm_compute::cpu::arithmetic_addition_same_sve<float>)
+    },
     {
-        // Clear X Dimension on execution window as we handle manually
-        input1_win.set(Window::DimX, Window::Dimension(0, 1, 1));
-        input2_win.set(Window::DimX, Window::Dimension(0, 1, 1));
-
-        Iterator input1(in1, input1_win);
-        Iterator input2(in2, input2_win);
-        Iterator output(out, win);
-
-        execute_window_loop(win, [&](const Coordinates &)
-        {
-            const auto input1_ptr = reinterpret_cast<const T *>(input1.ptr());
-            const auto input2_ptr = reinterpret_cast<const T *>(input2.ptr());
-            const auto output_ptr = reinterpret_cast<T *>(output.ptr());
-
-            // Compute S elements per iteration
-            int x = window_start_x;
-            for(; x <= (window_end_x - window_step_x); x += window_step_x)
-            {
-                const auto val1 = wrapper::vloadq(input1_ptr + x);
-                const auto val2 = wrapper::vloadq(input2_ptr + x);
-                const auto res  = (policy == ConvertPolicy::SATURATE) ? wrapper::vqadd(val1, val2) : wrapper::vadd(val1, val2);
-                wrapper::vstore(output_ptr + x, res);
-            }
-
-            // Compute left-over elements
-            for(; x < window_end_x; ++x)
-            {
-                const auto val1   = *(input1_ptr + x);
-                const auto val2   = *(input2_ptr + x);
-                *(output_ptr + x) = (policy == ConvertPolicy::SATURATE) ? wrapper::add_sat(val1, val2) : val1 + val2;
-            }
-        },
-        input1, input2, output);
-    }
-}
-
-void add_QASYMM8_QASYMM8_QASYMM8(const ITensor *in1, const ITensor *in2, ITensor *out, ConvertPolicy policy, const Window &window)
-{
-    ARM_COMPUTE_UNUSED(policy);
-
-    // Create input windows
-    Window input1_win = window.broadcast_if_dimension_le_one(in1->info()->tensor_shape());
-    Window input2_win = window.broadcast_if_dimension_le_one(in2->info()->tensor_shape());
-
-    // Clear X Dimension on execution window as we handle manually
-    Window win = window;
-    win.set(Window::DimX, Window::Dimension(0, 1, 1));
-
-    const int  window_step_x         = 16;
-    const auto window_start_x        = static_cast<int>(window.x().start());
-    const auto window_end_x          = static_cast<int>(window.x().end());
-    const bool is_broadcast_across_x = in1->info()->tensor_shape().x() != in2->info()->tensor_shape().x();
-
-    const UniformQuantizationInfo iq1_info = in1->info()->quantization_info().uniform();
-    const UniformQuantizationInfo iq2_info = in2->info()->quantization_info().uniform();
-    const UniformQuantizationInfo oq_info  = out->info()->quantization_info().uniform();
-
-    const float32x4_t invvscaleo = vdupq_n_f32(1.f / oq_info.scale);
-    const float32x4_t voffseto   = vdupq_n_f32(oq_info.offset);
-
-    if(is_broadcast_across_x)
+        "arithmetic_addition_same_sve",
+        [](const ArithmeticAdditionSelectorData & data) { return ((data.dt1 == data.dt2) && (data.dt1 == DataType::F16)); },
+        REGISTER_FP16_SVE(arm_compute::cpu::arithmetic_addition_same_sve<float16_t>)
+    },
     {
-        const bool                    is_broadcast_input_2 = input2_win.x().step() == 0;
-        Window                        broadcast_win        = is_broadcast_input_2 ? input2_win : input1_win;
-        Window                        non_broadcast_win    = !is_broadcast_input_2 ? input2_win : input1_win;
-        const ITensor                *broadcast_tensor     = is_broadcast_input_2 ? in2 : in1;
-        const ITensor                *non_broadcast_tensor = !is_broadcast_input_2 ? in2 : in1;
-        const UniformQuantizationInfo broadcast_qinfo      = broadcast_tensor->info()->quantization_info().uniform();
-        const UniformQuantizationInfo non_broadcast_qinfo  = non_broadcast_tensor->info()->quantization_info().uniform();
-
-        const float32x4_t vscale1  = is_broadcast_input_2 ? vdupq_n_f32(iq1_info.scale) : vdupq_n_f32(iq2_info.scale);
-        const float32x4_t vscale2  = is_broadcast_input_2 ? vdupq_n_f32(iq2_info.scale) : vdupq_n_f32(iq1_info.scale);
-        const int32x4_t   voffset1 = is_broadcast_input_2 ? vdupq_n_s32(iq1_info.offset) : vdupq_n_s32(iq2_info.offset);
-        const int32x4_t   voffset2 = is_broadcast_input_2 ? vdupq_n_s32(iq2_info.offset) : vdupq_n_s32(iq1_info.offset);
-
-        // Clear X Dimension on execution window as we handle manually
-        non_broadcast_win.set(Window::DimX, Window::Dimension(0, 1, 1));
-
-        Iterator broadcast_input(broadcast_tensor, broadcast_win);
-        Iterator non_broadcast_input(non_broadcast_tensor, non_broadcast_win);
-        Iterator output(out, win);
-
-        execute_window_loop(win, [&](const Coordinates &)
-        {
-            const auto non_broadcast_input_ptr = reinterpret_cast<const uint8_t *>(non_broadcast_input.ptr());
-            const auto output_ptr              = reinterpret_cast<uint8_t *>(output.ptr());
-
-            const uint8_t    broadcast_value     = *reinterpret_cast<const uint8_t *>(broadcast_input.ptr());
-            const uint8x16_t broadcast_value_vec = vdupq_n_u8(broadcast_value);
-
-            const float32x4x4_t bf =
-            {
-                {
-                    vmulq_f32(vcvtq_f32_s32(vsubq_s32(vreinterpretq_s32_u32(vmovl_u16(vget_low_u16(vmovl_u8(vget_low_u8(broadcast_value_vec))))), voffset2)), vscale2),
-                    vmulq_f32(vcvtq_f32_s32(vsubq_s32(vreinterpretq_s32_u32(vmovl_u16(vget_high_u16(vmovl_u8(vget_low_u8(broadcast_value_vec))))), voffset2)), vscale2),
-                    vmulq_f32(vcvtq_f32_s32(vsubq_s32(vreinterpretq_s32_u32(vmovl_u16(vget_low_u16(vmovl_u8(vget_high_u8(broadcast_value_vec))))), voffset2)), vscale2),
-                    vmulq_f32(vcvtq_f32_s32(vsubq_s32(vreinterpretq_s32_u32(vmovl_u16(vget_high_u16(vmovl_u8(vget_high_u8(broadcast_value_vec))))), voffset2)), vscale2),
-                }
-            };
-            const float bfs = static_cast<int32_t>(broadcast_value - broadcast_qinfo.offset) * broadcast_qinfo.scale;
-
-            // Compute S elements per iteration
-            int x = window_start_x;
-            for(; x <= (window_end_x - window_step_x); x += window_step_x)
-            {
-                const uint8x16_t    a = vld1q_u8(non_broadcast_input_ptr + x);
-                const float32x4x4_t af =
-                {
-                    {
-                        vmulq_f32(vcvtq_f32_s32(vsubq_s32(vreinterpretq_s32_u32(vmovl_u16(vget_low_u16(vmovl_u8(vget_low_u8(a))))), voffset1)), vscale1),
-                        vmulq_f32(vcvtq_f32_s32(vsubq_s32(vreinterpretq_s32_u32(vmovl_u16(vget_high_u16(vmovl_u8(vget_low_u8(a))))), voffset1)), vscale1),
-                        vmulq_f32(vcvtq_f32_s32(vsubq_s32(vreinterpretq_s32_u32(vmovl_u16(vget_low_u16(vmovl_u8(vget_high_u8(a))))), voffset1)), vscale1),
-                        vmulq_f32(vcvtq_f32_s32(vsubq_s32(vreinterpretq_s32_u32(vmovl_u16(vget_high_u16(vmovl_u8(vget_high_u8(a))))), voffset1)), vscale1),
-                    }
-                };
-
-                const int32x4x4_t rf =
-                {
-                    {
-#ifdef __aarch64__
-                        vcvtnq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af.val[0], bf.val[0]), invvscaleo)),
-                        vcvtnq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af.val[1], bf.val[1]), invvscaleo)),
-                        vcvtnq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af.val[2], bf.val[2]), invvscaleo)),
-                        vcvtnq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af.val[3], bf.val[3]), invvscaleo)),
-#else  //__aarch64__
-                        vcvtq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af.val[0], bf.val[0]), invvscaleo)),
-                        vcvtq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af.val[1], bf.val[1]), invvscaleo)),
-                        vcvtq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af.val[2], bf.val[2]), invvscaleo)),
-                        vcvtq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af.val[3], bf.val[3]), invvscaleo)),
-#endif //__aarch64__
-                    }
-                };
-
-                const uint8x8_t pa = vqmovun_s16(vcombine_s16(vqmovn_s32(rf.val[0]), vqmovn_s32(rf.val[1])));
-                const uint8x8_t pb = vqmovun_s16(vcombine_s16(vqmovn_s32(rf.val[2]), vqmovn_s32(rf.val[3])));
-                vst1q_u8(output_ptr + x, vcombine_u8(pa, pb));
-            }
-
-            // Compute left-over elements
-            for(; x < window_end_x; ++x)
-            {
-                const float afs   = static_cast<int32_t>(*(non_broadcast_input_ptr + x) - non_broadcast_qinfo.offset) * non_broadcast_qinfo.scale;
-                *(output_ptr + x) = quantize_qasymm8((afs + bfs), oq_info);
-            }
-        },
-        broadcast_input, non_broadcast_input, output);
-    }
-    else
+        "arithmetic_addition_same_sve",
+        [](const ArithmeticAdditionSelectorData & data) { return ((data.dt1 == data.dt2) && (data.dt1 == data.dt3) && (data.dt1 == DataType::U8)); },
+        REGISTER_INTEGER_SVE(arm_compute::cpu::arithmetic_addition_same_sve<uint8_t>)
+    },
     {
-        // Clear X Dimension on execution window as we handle manually
-        input1_win.set(Window::DimX, Window::Dimension(0, 1, 1));
-        input2_win.set(Window::DimX, Window::Dimension(0, 1, 1));
-
-        Iterator input1(in1, input1_win);
-        Iterator input2(in2, input2_win);
-        Iterator output(out, win);
-
-        const float32x4_t vscale1  = vdupq_n_f32(iq1_info.scale);
-        const float32x4_t vscale2  = vdupq_n_f32(iq2_info.scale);
-        const int32x4_t   voffset1 = vdupq_n_s32(iq1_info.offset);
-        const int32x4_t   voffset2 = vdupq_n_s32(iq2_info.offset);
-
-        execute_window_loop(win, [&](const Coordinates &)
-        {
-            const auto input1_ptr = reinterpret_cast<const uint8_t *>(input1.ptr());
-            const auto input2_ptr = reinterpret_cast<const uint8_t *>(input2.ptr());
-            const auto output_ptr = reinterpret_cast<uint8_t *>(output.ptr());
-
-            // Compute S elements per iteration
-            int x = window_start_x;
-            for(; x <= (window_end_x - window_step_x); x += window_step_x)
-            {
-                const uint8x16_t a = vld1q_u8(input1_ptr + x);
-                const uint8x16_t b = vld1q_u8(input2_ptr + x);
-
-                const float32x4x4_t af =
-                {
-                    {
-                        vmulq_f32(vcvtq_f32_s32(vsubq_s32(vreinterpretq_s32_u32(vmovl_u16(vget_low_u16(vmovl_u8(vget_low_u8(a))))), voffset1)), vscale1),
-                        vmulq_f32(vcvtq_f32_s32(vsubq_s32(vreinterpretq_s32_u32(vmovl_u16(vget_high_u16(vmovl_u8(vget_low_u8(a))))), voffset1)), vscale1),
-                        vmulq_f32(vcvtq_f32_s32(vsubq_s32(vreinterpretq_s32_u32(vmovl_u16(vget_low_u16(vmovl_u8(vget_high_u8(a))))), voffset1)), vscale1),
-                        vmulq_f32(vcvtq_f32_s32(vsubq_s32(vreinterpretq_s32_u32(vmovl_u16(vget_high_u16(vmovl_u8(vget_high_u8(a))))), voffset1)), vscale1),
-                    }
-                };
-
-                const float32x4x4_t bf =
-                {
-                    {
-                        vmulq_f32(vcvtq_f32_s32(vsubq_s32(vreinterpretq_s32_u32(vmovl_u16(vget_low_u16(vmovl_u8(vget_low_u8(b))))), voffset2)), vscale2),
-                        vmulq_f32(vcvtq_f32_s32(vsubq_s32(vreinterpretq_s32_u32(vmovl_u16(vget_high_u16(vmovl_u8(vget_low_u8(b))))), voffset2)), vscale2),
-                        vmulq_f32(vcvtq_f32_s32(vsubq_s32(vreinterpretq_s32_u32(vmovl_u16(vget_low_u16(vmovl_u8(vget_high_u8(b))))), voffset2)), vscale2),
-                        vmulq_f32(vcvtq_f32_s32(vsubq_s32(vreinterpretq_s32_u32(vmovl_u16(vget_high_u16(vmovl_u8(vget_high_u8(b))))), voffset2)), vscale2),
-                    }
-                };
-
-                const int32x4x4_t rf =
-                {
-                    {
-#ifdef __aarch64__
-                        vcvtnq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af.val[0], bf.val[0]), invvscaleo)),
-                        vcvtnq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af.val[1], bf.val[1]), invvscaleo)),
-                        vcvtnq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af.val[2], bf.val[2]), invvscaleo)),
-                        vcvtnq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af.val[3], bf.val[3]), invvscaleo)),
-#else  //__aarch64__
-                        vcvtq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af.val[0], bf.val[0]), invvscaleo)),
-                        vcvtq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af.val[1], bf.val[1]), invvscaleo)),
-                        vcvtq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af.val[2], bf.val[2]), invvscaleo)),
-                        vcvtq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af.val[3], bf.val[3]), invvscaleo)),
-#endif //__aarch64__
-                    }
-                };
-
-                const uint8x8_t pa = vqmovun_s16(vcombine_s16(vqmovn_s32(rf.val[0]), vqmovn_s32(rf.val[1])));
-                const uint8x8_t pb = vqmovun_s16(vcombine_s16(vqmovn_s32(rf.val[2]), vqmovn_s32(rf.val[3])));
-                vst1q_u8(output_ptr + x, vcombine_u8(pa, pb));
-            }
-
-            // Compute left-over elements
-            for(; x < window_end_x; ++x)
-            {
-                const float afs   = static_cast<int32_t>((*(input1_ptr + x)) - iq1_info.offset) * iq1_info.scale;
-                const float bfs   = static_cast<int32_t>((*(input2_ptr + x)) - iq2_info.offset) * iq2_info.scale;
-                *(output_ptr + x) = quantize_qasymm8((afs + bfs), out->info()->quantization_info());
-            }
-        },
-        input1, input2, output);
-    }
-}
-
-void add_QASYMM8_SIGNED_QASYMM8_SIGNED_QASYMM8_SIGNED(const ITensor *in1, const ITensor *in2, ITensor *out, ConvertPolicy policy, const Window &window)
-{
-    ARM_COMPUTE_UNUSED(policy);
-
-    // Create input windows
-    Window input1_win = window.broadcast_if_dimension_le_one(in1->info()->tensor_shape());
-    Window input2_win = window.broadcast_if_dimension_le_one(in2->info()->tensor_shape());
-
-    // Clear X Dimension on execution window as we handle manually
-    Window win = window;
-    win.set(Window::DimX, Window::Dimension(0, 1, 1));
-
-    const int  window_step_x         = 16;
-    const auto window_start_x        = static_cast<int>(window.x().start());
-    const auto window_end_x          = static_cast<int>(window.x().end());
-    const bool is_broadcast_across_x = in1->info()->tensor_shape().x() != in2->info()->tensor_shape().x();
-
-    const UniformQuantizationInfo iq1_info = in1->info()->quantization_info().uniform();
-    const UniformQuantizationInfo iq2_info = in2->info()->quantization_info().uniform();
-    const UniformQuantizationInfo oq_info  = out->info()->quantization_info().uniform();
-
-    const float32x4_t invvscaleo = vdupq_n_f32(1.f / oq_info.scale);
-    const float32x4_t voffseto   = vdupq_n_f32(oq_info.offset);
-
-    if(is_broadcast_across_x)
+        "arithmetic_addition_same_sve",
+        [](const ArithmeticAdditionSelectorData & data) { return ((data.dt1 == data.dt2) && (data.dt1 == data.dt3) && (data.dt1 == DataType::S16)); },
+        REGISTER_INTEGER_SVE(arm_compute::cpu::arithmetic_addition_same_sve<int16_t>)
+    },
     {
-        const bool                    is_broadcast_input_2 = input2_win.x().step() == 0;
-        Window                        broadcast_win        = is_broadcast_input_2 ? input2_win : input1_win;
-        Window                        non_broadcast_win    = !is_broadcast_input_2 ? input2_win : input1_win;
-        const ITensor                *broadcast_tensor     = is_broadcast_input_2 ? in2 : in1;
-        const ITensor                *non_broadcast_tensor = !is_broadcast_input_2 ? in2 : in1;
-        const UniformQuantizationInfo broadcast_qinfo      = broadcast_tensor->info()->quantization_info().uniform();
-        const UniformQuantizationInfo non_broadcast_qinfo  = non_broadcast_tensor->info()->quantization_info().uniform();
-
-        const float32x4_t vscale1  = is_broadcast_input_2 ? vdupq_n_f32(iq1_info.scale) : vdupq_n_f32(iq2_info.scale);
-        const float32x4_t vscale2  = is_broadcast_input_2 ? vdupq_n_f32(iq2_info.scale) : vdupq_n_f32(iq1_info.scale);
-        const int32x4_t   voffset1 = is_broadcast_input_2 ? vdupq_n_s32(iq1_info.offset) : vdupq_n_s32(iq2_info.offset);
-        const int32x4_t   voffset2 = is_broadcast_input_2 ? vdupq_n_s32(iq2_info.offset) : vdupq_n_s32(iq1_info.offset);
-
-        // Clear X Dimension on execution window as we handle manually
-        non_broadcast_win.set(Window::DimX, Window::Dimension(0, 1, 1));
-
-        Iterator broadcast_input(broadcast_tensor, broadcast_win);
-        Iterator non_broadcast_input(non_broadcast_tensor, non_broadcast_win);
-        Iterator output(out, win);
-
-        execute_window_loop(win, [&](const Coordinates &)
-        {
-            const auto non_broadcast_input_ptr = reinterpret_cast<const int8_t *>(non_broadcast_input.ptr());
-            const auto output_ptr              = reinterpret_cast<int8_t *>(output.ptr());
-
-            const int8_t    broadcast_value     = *reinterpret_cast<const int8_t *>(broadcast_input.ptr());
-            const int8x16_t broadcast_value_vec = vdupq_n_s8(broadcast_value);
-
-            const float32x4x4_t bf =
-            {
-                {
-                    vmulq_f32(vcvtq_f32_s32(vsubq_s32(vmovl_s16(vget_low_s16(vmovl_s8(vget_low_s8(broadcast_value_vec)))), voffset2)), vscale2),
-                    vmulq_f32(vcvtq_f32_s32(vsubq_s32(vmovl_s16(vget_high_s16(vmovl_s8(vget_low_s8(broadcast_value_vec)))), voffset2)), vscale2),
-                    vmulq_f32(vcvtq_f32_s32(vsubq_s32(vmovl_s16(vget_low_s16(vmovl_s8(vget_high_s8(broadcast_value_vec)))), voffset2)), vscale2),
-                    vmulq_f32(vcvtq_f32_s32(vsubq_s32(vmovl_s16(vget_high_s16(vmovl_s8(vget_high_s8(broadcast_value_vec)))), voffset2)), vscale2),
-                }
-            };
-            const float bfs = static_cast<int32_t>(broadcast_value - broadcast_qinfo.offset) * broadcast_qinfo.scale;
-
-            // Compute S elements per iteration
-            int x = window_start_x;
-            for(; x <= (window_end_x - window_step_x); x += window_step_x)
-            {
-                const int8x16_t     a = vld1q_s8(non_broadcast_input_ptr + x);
-                const float32x4x4_t af =
-                {
-                    {
-                        vmulq_f32(vcvtq_f32_s32(vsubq_s32(vmovl_s16(vget_low_s16(vmovl_s8(vget_low_s8(a)))), voffset1)), vscale1),
-                        vmulq_f32(vcvtq_f32_s32(vsubq_s32(vmovl_s16(vget_high_s16(vmovl_s8(vget_low_s8(a)))), voffset1)), vscale1),
-                        vmulq_f32(vcvtq_f32_s32(vsubq_s32(vmovl_s16(vget_low_s16(vmovl_s8(vget_high_s8(a)))), voffset1)), vscale1),
-                        vmulq_f32(vcvtq_f32_s32(vsubq_s32(vmovl_s16(vget_high_s16(vmovl_s8(vget_high_s8(a)))), voffset1)), vscale1),
-                    }
-                };
-
-                const int32x4x4_t rf =
-                {
-                    {
-#ifdef __aarch64__
-                        vcvtnq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af.val[0], bf.val[0]), invvscaleo)),
-                        vcvtnq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af.val[1], bf.val[1]), invvscaleo)),
-                        vcvtnq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af.val[2], bf.val[2]), invvscaleo)),
-                        vcvtnq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af.val[3], bf.val[3]), invvscaleo)),
-#else  //__aarch64__
-                        vcvtq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af.val[0], bf.val[0]), invvscaleo)),
-                        vcvtq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af.val[1], bf.val[1]), invvscaleo)),
-                        vcvtq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af.val[2], bf.val[2]), invvscaleo)),
-                        vcvtq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af.val[3], bf.val[3]), invvscaleo)),
-#endif //__aarch64__
-                    }
-                };
-
-                const int8x8_t pa = vqmovn_s16(vcombine_s16(vqmovn_s32(rf.val[0]), vqmovn_s32(rf.val[1])));
-                const int8x8_t pb = vqmovn_s16(vcombine_s16(vqmovn_s32(rf.val[2]), vqmovn_s32(rf.val[3])));
-                vst1q_s8(output_ptr + x, vcombine_s8(pa, pb));
-            }
-
-            // Compute left-over elements
-            for(; x < window_end_x; ++x)
-            {
-                const float afs   = static_cast<int32_t>(*(non_broadcast_input_ptr + x) - non_broadcast_qinfo.offset) * non_broadcast_qinfo.scale;
-                *(output_ptr + x) = quantize_qasymm8_signed((afs + bfs), oq_info);
-            }
-        },
-        broadcast_input, non_broadcast_input, output);
-    }
-    else
+        "arithmetic_addition_same_sve",
+        [](const ArithmeticAdditionSelectorData & data) { return ((data.dt1 == data.dt2) && (data.dt1 == data.dt3) && (data.dt1 == DataType::S32)); },
+        REGISTER_INTEGER_SVE(arm_compute::cpu::arithmetic_addition_same_sve<int32_t>)
+    },
     {
-        // Clear X Dimension on execution window as we handle manually
-        input1_win.set(Window::DimX, Window::Dimension(0, 1, 1));
-        input2_win.set(Window::DimX, Window::Dimension(0, 1, 1));
-
-        Iterator input1(in1, input1_win);
-        Iterator input2(in2, input2_win);
-        Iterator output(out, win);
-
-        const float32x4_t vscale1  = vdupq_n_f32(iq1_info.scale);
-        const float32x4_t vscale2  = vdupq_n_f32(iq2_info.scale);
-        const int32x4_t   voffset1 = vdupq_n_s32(iq1_info.offset);
-        const int32x4_t   voffset2 = vdupq_n_s32(iq2_info.offset);
-        execute_window_loop(win, [&](const Coordinates &)
-        {
-            const auto input1_ptr = reinterpret_cast<const int8_t *>(input1.ptr());
-            const auto input2_ptr = reinterpret_cast<const int8_t *>(input2.ptr());
-            const auto output_ptr = reinterpret_cast<int8_t *>(output.ptr());
-
-            // Compute S elements per iteration
-            int x = window_start_x;
-            for(; x <= (window_end_x - window_step_x); x += window_step_x)
-            {
-                const int8x16_t a = vld1q_s8(input1_ptr + x);
-                const int8x16_t b = vld1q_s8(input2_ptr + x);
-
-                const float32x4x4_t af =
-                {
-                    {
-                        vmulq_f32(vcvtq_f32_s32(vsubq_s32(vmovl_s16(vget_low_s16(vmovl_s8(vget_low_s8(a)))), voffset1)), vscale1),
-                        vmulq_f32(vcvtq_f32_s32(vsubq_s32(vmovl_s16(vget_high_s16(vmovl_s8(vget_low_s8(a)))), voffset1)), vscale1),
-                        vmulq_f32(vcvtq_f32_s32(vsubq_s32(vmovl_s16(vget_low_s16(vmovl_s8(vget_high_s8(a)))), voffset1)), vscale1),
-                        vmulq_f32(vcvtq_f32_s32(vsubq_s32(vmovl_s16(vget_high_s16(vmovl_s8(vget_high_s8(a)))), voffset1)), vscale1),
-                    }
-                };
-
-                const float32x4x4_t bf =
-                {
-                    {
-                        vmulq_f32(vcvtq_f32_s32(vsubq_s32(vmovl_s16(vget_low_s16(vmovl_s8(vget_low_s8(b)))), voffset2)), vscale2),
-                        vmulq_f32(vcvtq_f32_s32(vsubq_s32(vmovl_s16(vget_high_s16(vmovl_s8(vget_low_s8(b)))), voffset2)), vscale2),
-                        vmulq_f32(vcvtq_f32_s32(vsubq_s32(vmovl_s16(vget_low_s16(vmovl_s8(vget_high_s8(b)))), voffset2)), vscale2),
-                        vmulq_f32(vcvtq_f32_s32(vsubq_s32(vmovl_s16(vget_high_s16(vmovl_s8(vget_high_s8(b)))), voffset2)), vscale2),
-                    }
-                };
-
-                const int32x4x4_t rf =
-                {
-                    {
-#ifdef __aarch64__
-                        vcvtnq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af.val[0], bf.val[0]), invvscaleo)),
-                        vcvtnq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af.val[1], bf.val[1]), invvscaleo)),
-                        vcvtnq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af.val[2], bf.val[2]), invvscaleo)),
-                        vcvtnq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af.val[3], bf.val[3]), invvscaleo)),
-#else  //__aarch64__
-                        vcvtq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af.val[0], bf.val[0]), invvscaleo)),
-                        vcvtq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af.val[1], bf.val[1]), invvscaleo)),
-                        vcvtq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af.val[2], bf.val[2]), invvscaleo)),
-                        vcvtq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af.val[3], bf.val[3]), invvscaleo)),
-#endif //__aarch64__
-                    }
-                };
-
-                const int8x8_t pa = vqmovn_s16(vcombine_s16(vqmovn_s32(rf.val[0]), vqmovn_s32(rf.val[1])));
-                const int8x8_t pb = vqmovn_s16(vcombine_s16(vqmovn_s32(rf.val[2]), vqmovn_s32(rf.val[3])));
-                vst1q_s8(output_ptr + x, vcombine_s8(pa, pb));
-            }
-
-            // Compute left-over elements
-            for(; x < window_end_x; ++x)
-            {
-                const float afs   = static_cast<int32_t>((*(input1_ptr + x)) - iq1_info.offset) * iq1_info.scale;
-                const float bfs   = static_cast<int32_t>((*(input2_ptr + x)) - iq2_info.offset) * iq2_info.scale;
-                *(output_ptr + x) = quantize_qasymm8_signed((afs + bfs), out->info()->quantization_info());
-            }
-        },
-        input1, input2, output);
-    }
-}
-
-void add_QSYMM16_QSYMM16_QSYMM16(const ITensor *in1, const ITensor *in2, ITensor *out, ConvertPolicy policy, const Window &window)
-{
-    ARM_COMPUTE_UNUSED(policy);
-
-    // Create input windows
-    Window input1_win = window.broadcast_if_dimension_le_one(in1->info()->tensor_shape());
-    Window input2_win = window.broadcast_if_dimension_le_one(in2->info()->tensor_shape());
-
-    // Clear X Dimension on execution window as we handle manually
-    Window win = window;
-    win.set(Window::DimX, Window::Dimension(0, 1, 1));
-
-    const int  window_step_x         = 8;
-    const auto window_start_x        = static_cast<int>(window.x().start());
-    const auto window_end_x          = static_cast<int>(window.x().end());
-    const bool is_broadcast_across_x = in1->info()->tensor_shape().x() != in2->info()->tensor_shape().x();
-
-    const UniformQuantizationInfo iq1_info = in1->info()->quantization_info().uniform();
-    const UniformQuantizationInfo iq2_info = in2->info()->quantization_info().uniform();
-    const UniformQuantizationInfo oq_info  = out->info()->quantization_info().uniform();
-
-    const float32x4_t vscale1    = vdupq_n_f32(iq1_info.scale);
-    const float32x4_t vscale2    = vdupq_n_f32(iq2_info.scale);
-    const float32x4_t invvscaleo = vdupq_n_f32(1.f / oq_info.scale);
-
-    if(is_broadcast_across_x)
+        "arithmetic_addition_U8_S16_S16_sve",
+        [](const ArithmeticAdditionSelectorData & data) { return ((data.dt1 == DataType::U8) && (data.dt2 == DataType::S16)); },
+        REGISTER_INTEGER_SVE(arm_compute::cpu::arithmetic_addition_U8_S16_S16_sve)
+    },
     {
-        const bool                    is_broadcast_input_2 = input2_win.x().step() == 0;
-        Window                        broadcast_win        = is_broadcast_input_2 ? input2_win : input1_win;
-        Window                        non_broadcast_win    = !is_broadcast_input_2 ? input2_win : input1_win;
-        const ITensor                *broadcast_tensor     = is_broadcast_input_2 ? in2 : in1;
-        const ITensor                *non_broadcast_tensor = !is_broadcast_input_2 ? in2 : in1;
-        const UniformQuantizationInfo broadcast_qinfo      = broadcast_tensor->info()->quantization_info().uniform();
-        const UniformQuantizationInfo non_broadcast_qinfo  = non_broadcast_tensor->info()->quantization_info().uniform();
-
-        // Clear X Dimension on execution window as we handle manually
-        non_broadcast_win.set(Window::DimX, Window::Dimension(0, 1, 1));
-
-        Iterator broadcast_input(broadcast_tensor, broadcast_win);
-        Iterator non_broadcast_input(non_broadcast_tensor, non_broadcast_win);
-        Iterator output(out, win);
-
-        execute_window_loop(win, [&](const Coordinates &)
-        {
-            const auto non_broadcast_input_ptr = reinterpret_cast<const int16_t *>(non_broadcast_input.ptr());
-            const auto output_ptr              = reinterpret_cast<int16_t *>(output.ptr());
-
-            const int16_t   broadcast_value     = *reinterpret_cast<const int16_t *>(broadcast_input.ptr());
-            const int16x8_t broadcast_value_vec = vdupq_n_s16(broadcast_value);
-
-            const float32x4x2_t bf =
-            {
-                {
-                    vmulq_f32(vcvtq_f32_s32(vmovl_s16(vget_low_s16(broadcast_value_vec))), vscale2),
-                    vmulq_f32(vcvtq_f32_s32(vmovl_s16(vget_high_s16(broadcast_value_vec))), vscale2),
-                }
-            };
-            const float bfs = static_cast<int32_t>(broadcast_value) * broadcast_qinfo.scale;
-
-            // Compute S elements per iteration
-            int x = window_start_x;
-            for(; x <= (window_end_x - window_step_x); x += window_step_x)
-            {
-                const int16x8_t     a = vld1q_s16(non_broadcast_input_ptr + x);
-                const float32x4x2_t af =
-                {
-                    {
-                        vmulq_f32(vcvtq_f32_s32(vmovl_s16(vget_low_s16(a))), vscale1),
-                        vmulq_f32(vcvtq_f32_s32(vmovl_s16(vget_high_s16(a))), vscale1),
-                    }
-                };
-
-                const int32x4x4_t rf =
-                {
-                    {
-#ifdef __aarch64__
-                        vcvtnq_s32_f32(vmulq_f32(vaddq_f32(af.val[0], bf.val[0]), invvscaleo)),
-                        vcvtnq_s32_f32(vmulq_f32(vaddq_f32(af.val[1], bf.val[1]), invvscaleo)),
-#else  //__aarch64__
-                        vcvtq_s32_f32(vmulq_f32(vaddq_f32(af.val[0], bf.val[0]), invvscaleo)),
-                        vcvtq_s32_f32(vmulq_f32(vaddq_f32(af.val[1], bf.val[1]), invvscaleo)),
-#endif //__aarch64__
-                    }
-                };
-
-                const int16x8_t pa = vcombine_s16(vqmovn_s32(rf.val[0]), vqmovn_s32(rf.val[1]));
-                vst1q_s16(output_ptr + x, pa);
-            }
-
-            // Compute left-over elements
-            for(; x < window_end_x; ++x)
-            {
-                const float afs   = static_cast<int32_t>(*(non_broadcast_input_ptr + x)) * non_broadcast_qinfo.scale;
-                *(output_ptr + x) = quantize_qsymm16((afs + bfs), oq_info);
-            }
-        },
-        broadcast_input, non_broadcast_input, output);
-    }
-    else
+        "arithmetic_addition_S16_U8_S16_sve",
+        [](const ArithmeticAdditionSelectorData & data) { return ((data.dt1 == DataType::S16) && (data.dt2 == DataType::U8)); },
+        REGISTER_INTEGER_SVE(arm_compute::cpu::arithmetic_addition_S16_U8_S16_sve)
+    },
     {
-        // Clear X Dimension on execution window as we handle manually
-        input1_win.set(Window::DimX, Window::Dimension(0, 1, 1));
-        input2_win.set(Window::DimX, Window::Dimension(0, 1, 1));
-
-        Iterator input1(in1, input1_win);
-        Iterator input2(in2, input2_win);
-        Iterator output(out, win);
-
-        execute_window_loop(win, [&](const Coordinates &)
-        {
-            const auto input1_ptr = reinterpret_cast<const int16_t *>(input1.ptr());
-            const auto input2_ptr = reinterpret_cast<const int16_t *>(input2.ptr());
-            const auto output_ptr = reinterpret_cast<int16_t *>(output.ptr());
-
-            // Compute S elements per iteration
-            int x = window_start_x;
-            for(; x <= (window_end_x - window_step_x); x += window_step_x)
-            {
-                const int16x8_t a = vld1q_s16(input1_ptr + x);
-                const int16x8_t b = vld1q_s16(input2_ptr + x);
-
-                const float32x4x2_t af =
-                {
-                    {
-                        vmulq_f32(vcvtq_f32_s32(vmovl_s16(vget_low_s16(a))), vscale1),
-                        vmulq_f32(vcvtq_f32_s32(vmovl_s16(vget_high_s16(a))), vscale1),
-                    }
-                };
-
-                const float32x4x2_t bf =
-                {
-                    {
-                        vmulq_f32(vcvtq_f32_s32(vmovl_s16(vget_low_s16(b))), vscale2),
-                        vmulq_f32(vcvtq_f32_s32(vmovl_s16(vget_high_s16(b))), vscale2),
-                    }
-                };
-
-                const int32x4x2_t rf =
-                {
-                    {
-#ifdef __aarch64__
-                        vcvtnq_s32_f32(vmulq_f32(vaddq_f32(af.val[0], bf.val[0]), invvscaleo)),
-                        vcvtnq_s32_f32(vmulq_f32(vaddq_f32(af.val[1], bf.val[1]), invvscaleo)),
-#else  //__aarch64__
-                        vcvtq_s32_f32(vmulq_f32(vaddq_f32(af.val[0], bf.val[0]), invvscaleo)),
-                        vcvtq_s32_f32(vmulq_f32(vaddq_f32(af.val[1], bf.val[1]), invvscaleo)),
-#endif //__aarch64__
-                    }
-                };
-
-                const int16x8_t pa = vcombine_s16(vqmovn_s32(rf.val[0]), vqmovn_s32(rf.val[1]));
-                vst1q_s16(output_ptr + x, pa);
-            }
-
-            // Compute left-over elements
-            for(; x < window_end_x; ++x)
-            {
-                const float afs   = static_cast<int32_t>((*(input1_ptr + x))) * iq1_info.scale;
-                const float bfs   = static_cast<int32_t>((*(input2_ptr + x))) * iq2_info.scale;
-                *(output_ptr + x) = quantize_qsymm16((afs + bfs), out->info()->quantization_info());
-            }
-        },
-        input1, input2, output);
-    }
-}
-
-void add_S16_U8_S16(const ITensor *in1, const ITensor *in2, ITensor *out, ConvertPolicy policy, const Window &window)
-{
-    // Create input windows
-    Window win        = window;
-    Window input1_win = window.broadcast_if_dimension_le_one(in1->info()->tensor_shape());
-    Window input2_win = window.broadcast_if_dimension_le_one(in2->info()->tensor_shape());
-
-    // Clear X Dimension on execution window as we handle manually
-    win.set(Window::DimX, Window::Dimension(0, 1, 1));
-    input1_win.set(Window::DimX, Window::Dimension(0, 1, 1));
-    input2_win.set(Window::DimX, Window::Dimension(0, 1, 1));
-
-    Iterator input1(in1, input1_win);
-    Iterator input2(in2, input2_win);
-    Iterator output(out, win);
-
-    const int  window_step_x  = 8;
-    const auto window_start_x = static_cast<int>(window.x().start());
-    const auto window_end_x   = static_cast<int>(window.x().end());
-
-    execute_window_loop(win, [&](const Coordinates &)
+        "arithmetic_addition_U8_U8_S16_sve",
+        [](const ArithmeticAdditionSelectorData & data) { return ((data.dt1 == data.dt2) && (data.dt3 == DataType::S16)); },
+        REGISTER_INTEGER_SVE(arm_compute::cpu::arithmetic_addition_U8_U8_S16_sve)
+    },
+#else /* !defined(__ARM_FEATURE_SVE) */
     {
-        const auto input1_ptr = reinterpret_cast<const int16_t *>(input1.ptr());
-        const auto input2_ptr = reinterpret_cast<const uint8_t *>(input2.ptr());
-        const auto output_ptr = reinterpret_cast<int16_t *>(output.ptr());
-
-        if(policy == ConvertPolicy::WRAP)
-        {
-            // Compute S elements per iteration
-            int x = window_start_x;
-            for(; x <= (window_end_x - window_step_x); x += window_step_x)
-            {
-                const auto vin1 = wrapper::vloadq(input1_ptr + x);
-                const auto vin2 = vreinterpretq_s16_u16(wrapper::vmovl(wrapper::vload(input2_ptr + x)));
-                wrapper::vstore(output_ptr + x, wrapper::vadd(vin1, vin2));
-            }
-
-            // Compute left-over elements
-            for(; x < window_end_x; ++x)
-            {
-                *(output_ptr + x) = *(input1_ptr + x) + static_cast<int16_t>(*(input2_ptr + x));
-            }
-        }
-        else
-        {
-            // Compute S elements per iteration
-            int x = window_start_x;
-            for(; x <= (window_end_x - window_step_x); x += window_step_x)
-            {
-                const auto vin1 = wrapper::vloadq(input1_ptr + x);
-                const auto vin2 = vreinterpretq_s16_u16(wrapper::vmovl(wrapper::vload(input2_ptr + x)));
-                wrapper::vstore(output_ptr + x, wrapper::vqadd(vin1, vin2));
-            }
+        "arithmetic_addition_same_neon",
+        [](const ArithmeticAdditionSelectorData & data) { return ((data.dt1 == data.dt2) && (data.dt1 == DataType::F32)); },
+        REGISTER_FP32_NEON(arm_compute::cpu::arithmetic_addition_same_neon<float>)
+    },
+#if defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC)
+    {
+        "arithmetic_addition_same_neon",
+        [](const ArithmeticAdditionSelectorData & data) { return ((data.dt1 == data.dt2) && (data.dt1 == DataType::F16)); },
+        REGISTER_FP16_NEON(arm_compute::cpu::arithmetic_addition_same_neon<float16_t>)
+    },
+#endif /* defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) */
+    {
+        "arithmetic_addition_same_neon",
+        [](const ArithmeticAdditionSelectorData & data) { return ((data.dt1 == data.dt2) && (data.dt1 == data.dt3) && (data.dt1 == DataType::U8)); },
+        REGISTER_INTEGER_NEON(arm_compute::cpu::arithmetic_addition_same_neon<uint8_t>)
+    },
+    {
+        "arithmetic_addition_same_neon",
+        [](const ArithmeticAdditionSelectorData & data) { return ((data.dt1 == data.dt2) && (data.dt1 == data.dt3) && (data.dt1 == DataType::S16)); },
+        REGISTER_INTEGER_NEON(arm_compute::cpu::arithmetic_addition_same_neon<int16_t>)
+    },
+    {
+        "arithmetic_addition_same_neon",
+        [](const ArithmeticAdditionSelectorData & data) { return ((data.dt1 == data.dt2) && (data.dt1 == data.dt3) && (data.dt1 == DataType::S32)); },
+        REGISTER_INTEGER_NEON(arm_compute::cpu::arithmetic_addition_same_neon<int32_t>)
+    },
+    {
+        "arithmetic_addition_U8_S16_S16_neon",
+        [](const ArithmeticAdditionSelectorData & data) { return ((data.dt1 == DataType::U8) && (data.dt2 == DataType::S16)); },
+        REGISTER_INTEGER_NEON(arm_compute::cpu::arithmetic_addition_U8_S16_S16_neon)
+    },
+    {
+        "arithmetic_addition_S16_U8_S16_neon",
+        [](const ArithmeticAdditionSelectorData & data) { return ((data.dt1 == DataType::S16) && (data.dt2 == DataType::U8)); },
+        REGISTER_INTEGER_NEON(arm_compute::cpu::arithmetic_addition_S16_U8_S16_neon)
+    },
+    {
+        "arithmetic_addition_U8_U8_S16_neon",
+        [](const ArithmeticAdditionSelectorData & data) { return ((data.dt1 == data.dt2) && (data.dt3 == DataType::S16)); },
+        REGISTER_INTEGER_NEON(arm_compute::cpu::arithmetic_addition_U8_U8_S16_neon)
+    },
+#endif /* defined(__ARM_FEATURE_SVE) */
 
-            // Compute left-over elements
-            for(; x < window_end_x; ++x)
-            {
-                *(output_ptr + x) = wrapper::add_sat(*(input1_ptr + x), static_cast<int16_t>(*(input2_ptr + x)));
-            }
-        }
+#if defined(__ARM_FEATURE_SVE2)
+    {
+        "arithmetic_addition_qasymm8_sve",
+        [](const ArithmeticAdditionSelectorData & data) { return ((data.dt1 == data.dt2) && (data.dt1 == DataType::QASYMM8)); },
+        REGISTER_QASYMM8_SVE(arm_compute::cpu::arithmetic_addition_qasymm8_sve)
     },
-    input1, input2, output);
-}
+    {
+        "arithmetic_addition_qasymm8_signed_sve",
+        [](const ArithmeticAdditionSelectorData & data) { return ((data.dt1 == data.dt2) && (data.dt1 == DataType::QASYMM8_SIGNED)); },
+        REGISTER_QASYMM8_SIGNED_SVE(arm_compute::cpu::arithmetic_addition_qasymm8_signed_sve)
+    },
+    {
+        "arithmetic_addition_qsymm16_sve",
+        [](const ArithmeticAdditionSelectorData & data) { return ((data.dt1 == data.dt2) && (data.dt1 == DataType::QSYMM16)); },
+        REGISTER_QSYMM16_SVE(arm_compute::cpu::arithmetic_addition_qsymm16_sve)
+    },
+#else  /* !defined(__ARM_FEATURE_SVE2) */
+    {
+        "arithmetic_addition_qasymm8_neon",
+        [](const ArithmeticAdditionSelectorData & data) { return ((data.dt1 == data.dt2) && (data.dt1 == DataType::QASYMM8)); },
+        REGISTER_QASYMM8_NEON(arm_compute::cpu::arithmetic_addition_qasymm8_neon)
+    },
+    {
+        "arithmetic_addition_qasymm8_signed_neon",
+        [](const ArithmeticAdditionSelectorData & data) { return ((data.dt1 == data.dt2) && (data.dt1 == DataType::QASYMM8_SIGNED)); },
+        REGISTER_QASYMM8_SIGNED_NEON(arm_compute::cpu::arithmetic_addition_qasymm8_signed_neon)
+    },
+    {
+        "arithmetic_addition_qsymm16_neon",
+        [](const ArithmeticAdditionSelectorData & data) { return ((data.dt1 == data.dt2) && (data.dt1 == DataType::QSYMM16)); },
+        REGISTER_QSYMM16_NEON(arm_compute::cpu::arithmetic_addition_qsymm16_neon)
+    },
+#endif /* defined(__ARM_FEATURE_SVE2) */
 
-inline void add_U8_S16_S16(const ITensor *input1, const ITensor *input2, ITensor *output, ConvertPolicy policy, const Window &window)
-{
-    // Simply swap the two input buffers:
-    add_S16_U8_S16(input2, input1, output, policy, window);
-}
+};
 
-void add_U8_U8_S16(const ITensor *in1, const ITensor *in2, ITensor *out, ConvertPolicy policy, const Window &window)
+const ArithmeticAdditionKernel *get_implementation(DataType dt1, DataType dt2, DataType dt3)
 {
-    // Create input windows
-    Window win        = window;
-    Window input1_win = window.broadcast_if_dimension_le_one(in1->info()->tensor_shape());
-    Window input2_win = window.broadcast_if_dimension_le_one(in2->info()->tensor_shape());
-
-    // Clear X Dimension on execution window as we handle manually
-    win.set(Window::DimX, Window::Dimension(0, 1, 1));
-    input1_win.set(Window::DimX, Window::Dimension(0, 1, 1));
-    input2_win.set(Window::DimX, Window::Dimension(0, 1, 1));
-
-    Iterator input1(in1, input1_win);
-    Iterator input2(in2, input2_win);
-    Iterator output(out, win);
-
-    const int  window_step_x  = 8;
-    const auto window_start_x = static_cast<int>(window.x().start());
-    const auto window_end_x   = static_cast<int>(window.x().end());
-
-    execute_window_loop(win, [&](const Coordinates &)
+    for(const auto &uk : available_kernels)
     {
-        const auto input1_ptr = reinterpret_cast<const uint8_t *>(input1.ptr());
-        const auto input2_ptr = reinterpret_cast<const uint8_t *>(input2.ptr());
-        const auto output_ptr = reinterpret_cast<int16_t *>(output.ptr());
-
-        if(policy == ConvertPolicy::WRAP)
-        {
-            // Compute S elements per iteration
-            int x = window_start_x;
-            for(; x <= (window_end_x - window_step_x); x += window_step_x)
-            {
-                const auto vin1 = vreinterpretq_s16_u16(wrapper::vmovl(wrapper::vload(input1_ptr + x)));
-                const auto vin2 = vreinterpretq_s16_u16(wrapper::vmovl(wrapper::vload(input2_ptr + x)));
-                wrapper::vstore(output_ptr + x, wrapper::vadd(vin1, vin2));
-            }
-
-            // Compute left-over elements
-            for(; x < window_end_x; ++x)
-            {
-                *(output_ptr + x) = static_cast<int16_t>(*(input1_ptr + x)) + static_cast<int16_t>(*(input2_ptr + x));
-            }
-        }
-        else
+        if(uk.is_selected({ dt1, dt2, dt3 }))
         {
-            // Compute S elements per iteration
-            int x = window_start_x;
-            for(; x <= (window_end_x - window_step_x); x += window_step_x)
-            {
-                const auto vin1 = vreinterpretq_s16_u16(wrapper::vmovl(wrapper::vload(input1_ptr + x)));
-                const auto vin2 = vreinterpretq_s16_u16(wrapper::vmovl(wrapper::vload(input2_ptr + x)));
-                wrapper::vstore(output_ptr + x, wrapper::vqadd(vin1, vin2));
-            }
-
-            // Compute left-over elements
-            for(; x < window_end_x; ++x)
-            {
-                *(output_ptr + x) = wrapper::add_sat(static_cast<int16_t>(*(input1_ptr + x)),
-                                                     static_cast<int16_t>(*(input2_ptr + x)));
-            }
+            return &uk;
         }
-    },
-    input1, input2, output);
+    }
+    return nullptr;
 }
 
 Status validate_arguments(const ITensorInfo &input1, const ITensorInfo &input2, const ITensorInfo &output, ConvertPolicy policy)
@@ -926,53 +298,12 @@ void NEArithmeticAdditionKernel::configure(const ITensorInfo *input1, const ITen
     ARM_COMPUTE_ERROR_ON_NULLPTR(input1, input2, output);
     ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(*input1, *input2, *output, policy));
 
+    _policy = policy;
+    _func   = get_implementation(input1->data_type(), input2->data_type(), output->data_type())->ukernel;
+
     // Configure kernel window
     auto win_config = validate_and_configure_window(*input1, *input2, *output);
     ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
-
-    static std::map<std::string, AddFunction *> map_function =
-    {
-        { "add_wrap_QASYMM8_QASYMM8_QASYMM8", &add_QASYMM8_QASYMM8_QASYMM8 },
-        { "add_saturate_QASYMM8_QASYMM8_QASYMM8", &add_QASYMM8_QASYMM8_QASYMM8 },
-        { "add_wrap_QASYMM8_SIGNED_QASYMM8_SIGNED_QASYMM8_SIGNED", &add_QASYMM8_SIGNED_QASYMM8_SIGNED_QASYMM8_SIGNED },
-        { "add_saturate_QASYMM8_SIGNED_QASYMM8_SIGNED_QASYMM8_SIGNED", &add_QASYMM8_SIGNED_QASYMM8_SIGNED_QASYMM8_SIGNED },
-        { "add_wrap_QSYMM16_QSYMM16_QSYMM16", &add_QSYMM16_QSYMM16_QSYMM16 },
-        { "add_saturate_QSYMM16_QSYMM16_QSYMM16", &add_QSYMM16_QSYMM16_QSYMM16 },
-        { "add_wrap_U8_U8_U8", &add_same<uint8_t> },
-        { "add_saturate_U8_U8_U8", &add_same<uint8_t> },
-        { "add_wrap_S16_U8_S16", &add_S16_U8_S16 },
-        { "add_saturate_S16_U8_S16", &add_S16_U8_S16 },
-        { "add_wrap_U8_S16_S16", &add_U8_S16_S16 },
-        { "add_saturate_U8_S16_S16", &add_U8_S16_S16 },
-        { "add_wrap_U8_U8_S16", &add_U8_U8_S16 },
-        { "add_saturate_U8_U8_S16", &add_U8_U8_S16 },
-        { "add_wrap_S16_S16_S16", &add_same<int16_t> },
-        { "add_saturate_S16_S16_S16", &add_same<int16_t> },
-        { "add_wrap_S32_S32_S32", &add_same<int32_t> },
-        { "add_saturate_S32_S32_S32", &add_same<int32_t> },
-        { "add_wrap_F32_F32_F32", &add_same<float> },
-        { "add_saturate_F32_F32_F32", &add_same<float> },
-#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
-        { "add_wrap_F16_F16_F16", &add_same<float16_t> },
-        { "add_saturate_F16_F16_F16", &add_same<float16_t> },
-#endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */
-    };
-
-    _policy = policy;
-
-    std::string function_to_call("add_");
-    function_to_call += policy == ConvertPolicy::WRAP ? "wrap_" : "saturate_";
-    function_to_call += string_from_data_type(input1->data_type()) + "_";
-    function_to_call += string_from_data_type(input2->data_type()) + "_";
-    function_to_call += string_from_data_type(output->data_type());
-
-    auto it = map_function.find(function_to_call);
-
-    if(it != map_function.end())
-    {
-        _func = it->second;
-    }
-
     INEKernel::configure(win_config.second);
 }
 
diff --git a/src/core/NEON/kernels/NEArithmeticAdditionKernel.h b/src/core/NEON/kernels/NEArithmeticAdditionKernel.h
index 2072ad91bd..b88fc8aa74 100644
--- a/src/core/NEON/kernels/NEArithmeticAdditionKernel.h
+++ b/src/core/NEON/kernels/NEArithmeticAdditionKernel.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020 Arm Limited.
+ * Copyright (c) 2016-2021 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -88,7 +88,6 @@ public:
     // Inherited methods overridden:
     void run_op(ITensorPack &tensors, const Window &window, const ThreadInfo &info) override;
 
-private:
     /** Common signature for all the specialised add functions
      *
      * @param[in]  input1 First input tensor. Data types supported: U8/QASYMM8/S16/QSYMM16/F16/S32/F32
@@ -97,10 +96,12 @@ private:
      * @param[in]  policy Overflow policy.
      * @param[in]  window Region on which to execute the kernel.
      */
-    using AddFunction = void(const ITensor *input1, const ITensor *input2, ITensor *output, ConvertPolicy policy, const Window &window);
+    using ArithmeticAdditionKernelPtr = std::add_pointer<void(const ITensor *, const ITensor *, ITensor *, const ConvertPolicy &, const Window &)>::type;
+
+private:
     /** Add function to use for the particular tensor types passed to configure() */
-    AddFunction *_func;
-    ConvertPolicy _policy;
+    ArithmeticAdditionKernelPtr _func;
+    ConvertPolicy               _policy;
 };
 } // namespace arm_compute
 #endif /*ARM_COMPUTE_NEARITHMETICADDITIONKERNEL_H */
diff --git a/src/core/NEON/kernels/arithmetic_addition/impl/NEON/integer.cpp b/src/core/NEON/kernels/arithmetic_addition/impl/NEON/integer.cpp
new file mode 100644
index 0000000000..8dd58cec6d
--- /dev/null
+++ b/src/core/NEON/kernels/arithmetic_addition/impl/NEON/integer.cpp
@@ -0,0 +1,171 @@
+/*
+ * Copyright (c) 2020-2021 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.
+ */
+#include "arm_compute/core/Helpers.h"
+#include "arm_compute/core/ITensor.h"
+#include "arm_compute/core/Types.h"
+#include "arm_compute/core/utils/misc/Traits.h"
+#include "src/core/NEON/wrapper/wrapper.h"
+#include "src/core/common/StdTypes.h"
+#include "src/core/helpers/WindowHelpers.h"
+
+namespace arm_compute
+{
+namespace cpu
+{
+void arithmetic_addition_U8_U8_S16_neon(const ITensor *in1, const ITensor *in2, ITensor *out, const ConvertPolicy &policy, const Window &window)
+{
+    // Create input windows
+    Window win        = window;
+    Window input1_win = window.broadcast_if_dimension_le_one(in1->info()->tensor_shape());
+    Window input2_win = window.broadcast_if_dimension_le_one(in2->info()->tensor_shape());
+
+    // Clear X Dimension on execution window as we handle manually
+    win.set(Window::DimX, Window::Dimension(0, 1, 1));
+    input1_win.set(Window::DimX, Window::Dimension(0, 1, 1));
+    input2_win.set(Window::DimX, Window::Dimension(0, 1, 1));
+
+    Iterator input1(in1, input1_win);
+    Iterator input2(in2, input2_win);
+    Iterator output(out, win);
+
+    const int  window_step_x  = 8;
+    const auto window_start_x = static_cast<int>(window.x().start());
+    const auto window_end_x   = static_cast<int>(window.x().end());
+
+    execute_window_loop(win, [&](const Coordinates &)
+    {
+        const auto input1_ptr = reinterpret_cast<const uint8_t *>(input1.ptr());
+        const auto input2_ptr = reinterpret_cast<const uint8_t *>(input2.ptr());
+        const auto output_ptr = reinterpret_cast<int16_t *>(output.ptr());
+
+        if(policy == ConvertPolicy::WRAP)
+        {
+            // Compute S elements per iteration
+            int x = window_start_x;
+            for(; x <= (window_end_x - window_step_x); x += window_step_x)
+            {
+                const auto vin1 = vreinterpretq_s16_u16(wrapper::vmovl(wrapper::vload(input1_ptr + x)));
+                const auto vin2 = vreinterpretq_s16_u16(wrapper::vmovl(wrapper::vload(input2_ptr + x)));
+                wrapper::vstore(output_ptr + x, wrapper::vadd(vin1, vin2));
+            }
+
+            // Compute left-over elements
+            for(; x < window_end_x; ++x)
+            {
+                *(output_ptr + x) = static_cast<int16_t>(*(input1_ptr + x)) + static_cast<int16_t>(*(input2_ptr + x));
+            }
+        }
+        else
+        {
+            // Compute S elements per iteration
+            int x = window_start_x;
+            for(; x <= (window_end_x - window_step_x); x += window_step_x)
+            {
+                const auto vin1 = vreinterpretq_s16_u16(wrapper::vmovl(wrapper::vload(input1_ptr + x)));
+                const auto vin2 = vreinterpretq_s16_u16(wrapper::vmovl(wrapper::vload(input2_ptr + x)));
+                wrapper::vstore(output_ptr + x, wrapper::vqadd(vin1, vin2));
+            }
+
+            // Compute left-over elements
+            for(; x < window_end_x; ++x)
+            {
+                *(output_ptr + x) = wrapper::add_sat(static_cast<int16_t>(*(input1_ptr + x)),
+                                                     static_cast<int16_t>(*(input2_ptr + x)));
+            }
+        }
+    },
+    input1, input2, output);
+}
+
+void arithmetic_addition_S16_U8_S16_neon(const ITensor *in1, const ITensor *in2, ITensor *out, const ConvertPolicy &policy, const Window &window)
+{
+    // Create input windows
+    Window win        = window;
+    Window input1_win = window.broadcast_if_dimension_le_one(in1->info()->tensor_shape());
+    Window input2_win = window.broadcast_if_dimension_le_one(in2->info()->tensor_shape());
+
+    // Clear X Dimension on execution window as we handle manually
+    win.set(Window::DimX, Window::Dimension(0, 1, 1));
+    input1_win.set(Window::DimX, Window::Dimension(0, 1, 1));
+    input2_win.set(Window::DimX, Window::Dimension(0, 1, 1));
+
+    Iterator input1(in1, input1_win);
+    Iterator input2(in2, input2_win);
+    Iterator output(out, win);
+
+    const int  window_step_x  = 8;
+    const auto window_start_x = static_cast<int>(window.x().start());
+    const auto window_end_x   = static_cast<int>(window.x().end());
+
+    execute_window_loop(win, [&](const Coordinates &)
+    {
+        const auto input1_ptr = reinterpret_cast<const int16_t *>(input1.ptr());
+        const auto input2_ptr = reinterpret_cast<const uint8_t *>(input2.ptr());
+        const auto output_ptr = reinterpret_cast<int16_t *>(output.ptr());
+
+        if(policy == ConvertPolicy::WRAP)
+        {
+            // Compute S elements per iteration
+            int x = window_start_x;
+            for(; x <= (window_end_x - window_step_x); x += window_step_x)
+            {
+                const auto vin1 = wrapper::vloadq(input1_ptr + x);
+                const auto vin2 = vreinterpretq_s16_u16(wrapper::vmovl(wrapper::vload(input2_ptr + x)));
+                wrapper::vstore(output_ptr + x, wrapper::vadd(vin1, vin2));
+            }
+
+            // Compute left-over elements
+            for(; x < window_end_x; ++x)
+            {
+                *(output_ptr + x) = *(input1_ptr + x) + static_cast<int16_t>(*(input2_ptr + x));
+            }
+        }
+        else
+        {
+            // Compute S elements per iteration
+            int x = window_start_x;
+            for(; x <= (window_end_x - window_step_x); x += window_step_x)
+            {
+                const auto vin1 = wrapper::vloadq(input1_ptr + x);
+                const auto vin2 = vreinterpretq_s16_u16(wrapper::vmovl(wrapper::vload(input2_ptr + x)));
+                wrapper::vstore(output_ptr + x, wrapper::vqadd(vin1, vin2));
+            }
+
+            // Compute left-over elements
+            for(; x < window_end_x; ++x)
+            {
+                *(output_ptr + x) = wrapper::add_sat(*(input1_ptr + x), static_cast<int16_t>(*(input2_ptr + x)));
+            }
+        }
+    },
+    input1, input2, output);
+}
+
+void arithmetic_addition_U8_S16_S16_neon(const ITensor *input1, const ITensor *input2, ITensor *output, const ConvertPolicy &policy, const Window &window)
+{
+    // Simply swap the two input buffers:
+    arithmetic_addition_S16_U8_S16_neon(input2, input1, output, policy, window);
+}
+} // namespace cpu
+} // namespace arm_compute
\ No newline at end of file
diff --git a/src/core/NEON/kernels/arithmetic_addition/impl/NEON/list.h b/src/core/NEON/kernels/arithmetic_addition/impl/NEON/list.h
new file mode 100644
index 0000000000..a8ab0910fd
--- /dev/null
+++ b/src/core/NEON/kernels/arithmetic_addition/impl/NEON/list.h
@@ -0,0 +1,146 @@
+/*
+ * Copyright (c) 2020-2021 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.
+ */
+#ifndef SRC_CORE_NEON_KERNELS_ARITHMETIC_ADDITION_LIST_H
+#define SRC_CORE_NEON_KERNELS_ARITHMETIC_ADDITION_LIST_H
+
+#include "arm_compute/core/Types.h"
+#include "arm_compute/core/utils/misc/Traits.h"
+#include "src/core/NEON/wrapper/wrapper.h"
+
+namespace arm_compute
+{
+namespace cpu
+{
+#define DECLARE_ARITHMETIC_ADDITION_KERNEL(func_name) \
+    void func_name(const ITensor *in1, const ITensor *in2, ITensor *out, const ConvertPolicy &policy, const Window &window)
+
+DECLARE_ARITHMETIC_ADDITION_KERNEL(arithmetic_addition_qasymm8_neon);
+DECLARE_ARITHMETIC_ADDITION_KERNEL(arithmetic_addition_qasymm8_signed_neon);
+DECLARE_ARITHMETIC_ADDITION_KERNEL(arithmetic_addition_qsymm16_neon);
+DECLARE_ARITHMETIC_ADDITION_KERNEL(arithmetic_addition_S16_U8_S16_neon);
+DECLARE_ARITHMETIC_ADDITION_KERNEL(arithmetic_addition_U8_S16_S16_neon);
+DECLARE_ARITHMETIC_ADDITION_KERNEL(arithmetic_addition_U8_U8_S16_neon);
+
+#undef DECLARE_ARITHMETIC_ADDITION_KERNEL
+
+template <typename ScalarType>
+void arithmetic_addition_same_neon(const ITensor *in1, const ITensor *in2, ITensor *out, const ConvertPolicy &policy, const Window &window)
+{
+    /** NEON vector tag type. */
+    using ExactTagType = typename wrapper::traits::neon_bitvector_tag_t<ScalarType, wrapper::traits::BitWidth::W128>;
+
+    // Create input windows
+    Window input1_win = window.broadcast_if_dimension_le_one(in1->info()->tensor_shape());
+    Window input2_win = window.broadcast_if_dimension_le_one(in2->info()->tensor_shape());
+
+    // Clear X Dimension on execution window as we handle manually
+    Window win = window;
+    win.set(Window::DimX, Window::Dimension(0, 1, 1));
+
+    constexpr int window_step_x         = 16 / sizeof(ScalarType);
+    const auto    window_start_x        = static_cast<int>(window.x().start());
+    const auto    window_end_x          = static_cast<int>(window.x().end());
+    const bool    is_broadcast_across_x = in1->info()->tensor_shape().x() != in2->info()->tensor_shape().x();
+
+    if(is_broadcast_across_x)
+    {
+        const bool     is_broadcast_input_2 = input2_win.x().step() == 0;
+        Window         broadcast_win        = is_broadcast_input_2 ? input2_win : input1_win;
+        Window         non_broadcast_win    = !is_broadcast_input_2 ? input2_win : input1_win;
+        const ITensor *broadcast_tensor     = is_broadcast_input_2 ? in2 : in1;
+        const ITensor *non_broadcast_tensor = !is_broadcast_input_2 ? in2 : in1;
+
+        // Clear X Dimension on execution window as we handle manually
+        non_broadcast_win.set(Window::DimX, Window::Dimension(0, 1, 1));
+
+        Iterator broadcast_input(broadcast_tensor, broadcast_win);
+        Iterator non_broadcast_input(non_broadcast_tensor, non_broadcast_win);
+        Iterator output(out, win);
+
+        execute_window_loop(win, [&](const Coordinates &)
+        {
+            const auto non_broadcast_input_ptr = reinterpret_cast<const ScalarType *>(non_broadcast_input.ptr());
+            const auto output_ptr              = reinterpret_cast<ScalarType *>(output.ptr());
+
+            const ScalarType broadcast_value     = *reinterpret_cast<const ScalarType *>(broadcast_input.ptr());
+            const auto       broadcast_value_vec = wrapper::vdup_n(broadcast_value, ExactTagType{});
+
+            // Compute S elements per iteration
+            int x = window_start_x;
+            for(; x <= (window_end_x - window_step_x); x += window_step_x)
+            {
+                const auto non_broadcast_v = wrapper::vloadq(non_broadcast_input_ptr + x);
+                const auto res             = (policy == ConvertPolicy::SATURATE) ? wrapper::vqadd(broadcast_value_vec, non_broadcast_v) : wrapper::vadd(broadcast_value_vec, non_broadcast_v);
+                wrapper::vstore(output_ptr + x, res);
+            }
+
+            // Compute left-over elements
+            for(; x < window_end_x; ++x)
+            {
+                const auto non_broadcast_v = *(non_broadcast_input_ptr + x);
+                *(output_ptr + x)          = (policy == ConvertPolicy::SATURATE) ? wrapper::add_sat(broadcast_value, non_broadcast_v) : broadcast_value + non_broadcast_v;
+            }
+        },
+        broadcast_input, non_broadcast_input, output);
+    }
+    else
+    {
+        // Clear X Dimension on execution window as we handle manually
+        input1_win.set(Window::DimX, Window::Dimension(0, 1, 1));
+        input2_win.set(Window::DimX, Window::Dimension(0, 1, 1));
+
+        Iterator input1(in1, input1_win);
+        Iterator input2(in2, input2_win);
+        Iterator output(out, win);
+
+        execute_window_loop(win, [&](const Coordinates &)
+        {
+            const auto input1_ptr = reinterpret_cast<const ScalarType *>(input1.ptr());
+            const auto input2_ptr = reinterpret_cast<const ScalarType *>(input2.ptr());
+            const auto output_ptr = reinterpret_cast<ScalarType *>(output.ptr());
+
+            // Compute S elements per iteration
+            int x = window_start_x;
+            for(; x <= (window_end_x - window_step_x); x += window_step_x)
+            {
+                const auto val1 = wrapper::vloadq(input1_ptr + x);
+                const auto val2 = wrapper::vloadq(input2_ptr + x);
+                const auto res  = (policy == ConvertPolicy::SATURATE) ? wrapper::vqadd(val1, val2) : wrapper::vadd(val1, val2);
+                wrapper::vstore(output_ptr + x, res);
+            }
+
+            // Compute left-over elements
+            for(; x < window_end_x; ++x)
+            {
+                const auto val1   = *(input1_ptr + x);
+                const auto val2   = *(input2_ptr + x);
+                *(output_ptr + x) = (policy == ConvertPolicy::SATURATE) ? wrapper::add_sat(val1, val2) : val1 + val2;
+            }
+        },
+        input1, input2, output);
+    }
+}
+} // namespace cpu
+} // namespace arm_compute
+#endif // SRC_CORE_NEON_KERNELS_ARITHMETIC_ADDITION_LIST_H
\ No newline at end of file
diff --git a/src/core/NEON/kernels/arithmetic_addition/impl/NEON/qasymm8.cpp b/src/core/NEON/kernels/arithmetic_addition/impl/NEON/qasymm8.cpp
new file mode 100644
index 0000000000..b93dad20f1
--- /dev/null
+++ b/src/core/NEON/kernels/arithmetic_addition/impl/NEON/qasymm8.cpp
@@ -0,0 +1,210 @@
+/*
+ * Copyright (c) 2020-2021 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.
+ */
+#include "arm_compute/core/Helpers.h"
+#include "arm_compute/core/ITensor.h"
+#include "arm_compute/core/Types.h"
+#include "arm_compute/core/utils/misc/Traits.h"
+#include "src/core/NEON/wrapper/intrinsics/intrinsics.h"
+#include "src/core/common/StdTypes.h"
+#include "src/core/helpers/WindowHelpers.h"
+
+namespace arm_compute
+{
+namespace cpu
+{
+void arithmetic_addition_qasymm8_neon(const ITensor *in1, const ITensor *in2, ITensor *out, const ConvertPolicy &policy, const Window &window)
+{
+    ARM_COMPUTE_UNUSED(policy);
+
+    // Create input windows
+    Window input1_win = window.broadcast_if_dimension_le_one(in1->info()->tensor_shape());
+    Window input2_win = window.broadcast_if_dimension_le_one(in2->info()->tensor_shape());
+
+    // Clear X Dimension on execution window as we handle manually
+    Window win = window;
+    win.set(Window::DimX, Window::Dimension(0, 1, 1));
+
+    const int  window_step_x         = 16;
+    const auto window_start_x        = static_cast<int>(window.x().start());
+    const auto window_end_x          = static_cast<int>(window.x().end());
+    const bool is_broadcast_across_x = in1->info()->tensor_shape().x() != in2->info()->tensor_shape().x();
+
+    const UniformQuantizationInfo iq1_info = in1->info()->quantization_info().uniform();
+    const UniformQuantizationInfo iq2_info = in2->info()->quantization_info().uniform();
+    const UniformQuantizationInfo oq_info  = out->info()->quantization_info().uniform();
+
+    const float32x4_t invvscaleo = vdupq_n_f32(1.f / oq_info.scale);
+    const float32x4_t voffseto   = vdupq_n_f32(oq_info.offset);
+
+    if(is_broadcast_across_x)
+    {
+        const bool                    is_broadcast_input_2 = input2_win.x().step() == 0;
+        Window                        broadcast_win        = is_broadcast_input_2 ? input2_win : input1_win;
+        Window                        non_broadcast_win    = !is_broadcast_input_2 ? input2_win : input1_win;
+        const ITensor                *broadcast_tensor     = is_broadcast_input_2 ? in2 : in1;
+        const ITensor                *non_broadcast_tensor = !is_broadcast_input_2 ? in2 : in1;
+        const UniformQuantizationInfo broadcast_qinfo      = broadcast_tensor->info()->quantization_info().uniform();
+        const UniformQuantizationInfo non_broadcast_qinfo  = non_broadcast_tensor->info()->quantization_info().uniform();
+
+        const float32x4_t vscale1  = is_broadcast_input_2 ? vdupq_n_f32(iq1_info.scale) : vdupq_n_f32(iq2_info.scale);
+        const float32x4_t vscale2  = is_broadcast_input_2 ? vdupq_n_f32(iq2_info.scale) : vdupq_n_f32(iq1_info.scale);
+        const int32x4_t   voffset1 = is_broadcast_input_2 ? vdupq_n_s32(iq1_info.offset) : vdupq_n_s32(iq2_info.offset);
+        const int32x4_t   voffset2 = is_broadcast_input_2 ? vdupq_n_s32(iq2_info.offset) : vdupq_n_s32(iq1_info.offset);
+
+        // Clear X Dimension on execution window as we handle manually
+        non_broadcast_win.set(Window::DimX, Window::Dimension(0, 1, 1));
+
+        Iterator broadcast_input(broadcast_tensor, broadcast_win);
+        Iterator non_broadcast_input(non_broadcast_tensor, non_broadcast_win);
+        Iterator output(out, win);
+
+        execute_window_loop(win, [&](const Coordinates &)
+        {
+            const auto non_broadcast_input_ptr = reinterpret_cast<const uint8_t *>(non_broadcast_input.ptr());
+            const auto output_ptr              = reinterpret_cast<uint8_t *>(output.ptr());
+
+            const uint8_t    broadcast_value     = *reinterpret_cast<const uint8_t *>(broadcast_input.ptr());
+            const uint8x16_t broadcast_value_vec = vdupq_n_u8(broadcast_value);
+
+            const auto bf_0 = vmulq_f32(vcvtq_f32_s32(vsubq_s32(vreinterpretq_s32_u32(vmovl_u16(vget_low_u16(vmovl_u8(vget_low_u8(broadcast_value_vec))))), voffset2)), vscale2);
+            const auto bf_1 = vmulq_f32(vcvtq_f32_s32(vsubq_s32(vreinterpretq_s32_u32(vmovl_u16(vget_high_u16(vmovl_u8(vget_low_u8(broadcast_value_vec))))), voffset2)), vscale2);
+            const auto bf_2 = vmulq_f32(vcvtq_f32_s32(vsubq_s32(vreinterpretq_s32_u32(vmovl_u16(vget_low_u16(vmovl_u8(vget_high_u8(broadcast_value_vec))))), voffset2)), vscale2);
+            const auto bf_3 = vmulq_f32(vcvtq_f32_s32(vsubq_s32(vreinterpretq_s32_u32(vmovl_u16(vget_high_u16(vmovl_u8(vget_high_u8(broadcast_value_vec))))), voffset2)), vscale2);
+
+            const float bfs = static_cast<int32_t>(broadcast_value - broadcast_qinfo.offset) * broadcast_qinfo.scale;
+
+            // Compute S elements per iteration
+            int x = window_start_x;
+            for(; x <= (window_end_x - window_step_x); x += window_step_x)
+            {
+                const uint8x16_t a    = vld1q_u8(non_broadcast_input_ptr + x);
+                const auto       af_0 = vmulq_f32(vcvtq_f32_s32(vsubq_s32(vreinterpretq_s32_u32(vmovl_u16(vget_low_u16(vmovl_u8(vget_low_u8(a))))), voffset1)), vscale1);
+                const auto       af_1 = vmulq_f32(vcvtq_f32_s32(vsubq_s32(vreinterpretq_s32_u32(vmovl_u16(vget_high_u16(vmovl_u8(vget_low_u8(a))))), voffset1)), vscale1);
+                const auto       af_2 = vmulq_f32(vcvtq_f32_s32(vsubq_s32(vreinterpretq_s32_u32(vmovl_u16(vget_low_u16(vmovl_u8(vget_high_u8(a))))), voffset1)), vscale1);
+                const auto       af_3 = vmulq_f32(vcvtq_f32_s32(vsubq_s32(vreinterpretq_s32_u32(vmovl_u16(vget_high_u16(vmovl_u8(vget_high_u8(a))))), voffset1)), vscale1);
+
+                int32x4_t rf_0{};
+                int32x4_t rf_1{};
+                int32x4_t rf_2{};
+                int32x4_t rf_3{};
+
+#ifdef __aarch64__
+                rf_0 = vcvtnq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af_0, bf_0), invvscaleo));
+                rf_1 = vcvtnq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af_1, bf_1), invvscaleo));
+                rf_2 = vcvtnq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af_2, bf_2), invvscaleo));
+                rf_3 = vcvtnq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af_3, bf_3), invvscaleo));
+#else  //__aarch64__
+                rf_0 = vcvtq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af_0, bf_0), invvscaleo));
+                rf_1 = vcvtq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af_1, bf_1), invvscaleo));
+                rf_2 = vcvtq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af_2, bf_2), invvscaleo));
+                rf_3 = vcvtq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af_3, bf_3), invvscaleo));
+#endif //__aarch64__
+
+                const uint8x8_t pa = vqmovun_s16(vcombine_s16(vqmovn_s32(rf_0), vqmovn_s32(rf_1)));
+                const uint8x8_t pb = vqmovun_s16(vcombine_s16(vqmovn_s32(rf_2), vqmovn_s32(rf_3)));
+                vst1q_u8(output_ptr + x, vcombine_u8(pa, pb));
+            }
+
+            // Compute left-over elements
+            for(; x < window_end_x; ++x)
+            {
+                const float afs   = static_cast<int32_t>(*(non_broadcast_input_ptr + x) - non_broadcast_qinfo.offset) * non_broadcast_qinfo.scale;
+                *(output_ptr + x) = quantize_qasymm8((afs + bfs), oq_info);
+            }
+        },
+        broadcast_input, non_broadcast_input, output);
+    }
+    else
+    {
+        // Clear X Dimension on execution window as we handle manually
+        input1_win.set(Window::DimX, Window::Dimension(0, 1, 1));
+        input2_win.set(Window::DimX, Window::Dimension(0, 1, 1));
+
+        Iterator input1(in1, input1_win);
+        Iterator input2(in2, input2_win);
+        Iterator output(out, win);
+
+        const float32x4_t vscale1  = vdupq_n_f32(iq1_info.scale);
+        const float32x4_t vscale2  = vdupq_n_f32(iq2_info.scale);
+        const int32x4_t   voffset1 = vdupq_n_s32(iq1_info.offset);
+        const int32x4_t   voffset2 = vdupq_n_s32(iq2_info.offset);
+
+        execute_window_loop(win, [&](const Coordinates &)
+        {
+            const auto input1_ptr = reinterpret_cast<const uint8_t *>(input1.ptr());
+            const auto input2_ptr = reinterpret_cast<const uint8_t *>(input2.ptr());
+            const auto output_ptr = reinterpret_cast<uint8_t *>(output.ptr());
+
+            // Compute S elements per iteration
+            int x = window_start_x;
+            for(; x <= (window_end_x - window_step_x); x += window_step_x)
+            {
+                const uint8x16_t a = vld1q_u8(input1_ptr + x);
+                const uint8x16_t b = vld1q_u8(input2_ptr + x);
+
+                const auto af_0 = vmulq_f32(vcvtq_f32_s32(vsubq_s32(vreinterpretq_s32_u32(vmovl_u16(vget_low_u16(vmovl_u8(vget_low_u8(a))))), voffset1)), vscale1);
+                const auto af_1 = vmulq_f32(vcvtq_f32_s32(vsubq_s32(vreinterpretq_s32_u32(vmovl_u16(vget_high_u16(vmovl_u8(vget_low_u8(a))))), voffset1)), vscale1);
+                const auto af_2 = vmulq_f32(vcvtq_f32_s32(vsubq_s32(vreinterpretq_s32_u32(vmovl_u16(vget_low_u16(vmovl_u8(vget_high_u8(a))))), voffset1)), vscale1);
+                const auto af_3 = vmulq_f32(vcvtq_f32_s32(vsubq_s32(vreinterpretq_s32_u32(vmovl_u16(vget_high_u16(vmovl_u8(vget_high_u8(a))))), voffset1)), vscale1);
+
+                const auto bf_0 = vmulq_f32(vcvtq_f32_s32(vsubq_s32(vreinterpretq_s32_u32(vmovl_u16(vget_low_u16(vmovl_u8(vget_low_u8(b))))), voffset2)), vscale2);
+                const auto bf_1 = vmulq_f32(vcvtq_f32_s32(vsubq_s32(vreinterpretq_s32_u32(vmovl_u16(vget_high_u16(vmovl_u8(vget_low_u8(b))))), voffset2)), vscale2);
+                const auto bf_2 = vmulq_f32(vcvtq_f32_s32(vsubq_s32(vreinterpretq_s32_u32(vmovl_u16(vget_low_u16(vmovl_u8(vget_high_u8(b))))), voffset2)), vscale2);
+                const auto bf_3 = vmulq_f32(vcvtq_f32_s32(vsubq_s32(vreinterpretq_s32_u32(vmovl_u16(vget_high_u16(vmovl_u8(vget_high_u8(b))))), voffset2)), vscale2);
+
+                int32x4_t rf_0{};
+                int32x4_t rf_1{};
+                int32x4_t rf_2{};
+                int32x4_t rf_3{};
+
+#ifdef __aarch64__
+                rf_0 = vcvtnq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af_0, bf_0), invvscaleo));
+                rf_1 = vcvtnq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af_1, bf_1), invvscaleo));
+                rf_2 = vcvtnq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af_2, bf_2), invvscaleo));
+                rf_3 = vcvtnq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af_3, bf_3), invvscaleo));
+#else  //__aarch64__
+                rf_0 = vcvtq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af_0, bf_0), invvscaleo));
+                rf_1 = vcvtq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af_1, bf_1), invvscaleo));
+                rf_2 = vcvtq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af_2, bf_2), invvscaleo));
+                rf_3 = vcvtq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af_3, bf_3), invvscaleo));
+#endif //__aarch64__
+
+                const uint8x8_t pa = vqmovun_s16(vcombine_s16(vqmovn_s32(rf_0), vqmovn_s32(rf_1)));
+                const uint8x8_t pb = vqmovun_s16(vcombine_s16(vqmovn_s32(rf_2), vqmovn_s32(rf_3)));
+                vst1q_u8(output_ptr + x, vcombine_u8(pa, pb));
+            }
+
+            // Compute left-over elements
+            for(; x < window_end_x; ++x)
+            {
+                const float afs   = static_cast<int32_t>((*(input1_ptr + x)) - iq1_info.offset) * iq1_info.scale;
+                const float bfs   = static_cast<int32_t>((*(input2_ptr + x)) - iq2_info.offset) * iq2_info.scale;
+                *(output_ptr + x) = quantize_qasymm8((afs + bfs), out->info()->quantization_info());
+            }
+        },
+        input1, input2, output);
+    }
+}
+} // namespace cpu
+} // namespace arm_compute
\ No newline at end of file
diff --git a/src/core/NEON/kernels/arithmetic_addition/impl/NEON/qasymm8_signed.cpp b/src/core/NEON/kernels/arithmetic_addition/impl/NEON/qasymm8_signed.cpp
new file mode 100644
index 0000000000..ba81cfcc03
--- /dev/null
+++ b/src/core/NEON/kernels/arithmetic_addition/impl/NEON/qasymm8_signed.cpp
@@ -0,0 +1,209 @@
+/*
+ * Copyright (c) 2020-2021 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.
+ */
+#include "arm_compute/core/Helpers.h"
+#include "arm_compute/core/ITensor.h"
+#include "arm_compute/core/Types.h"
+#include "arm_compute/core/utils/misc/Traits.h"
+#include "src/core/NEON/wrapper/intrinsics/intrinsics.h"
+#include "src/core/common/StdTypes.h"
+#include "src/core/helpers/WindowHelpers.h"
+
+namespace arm_compute
+{
+namespace cpu
+{
+void arithmetic_addition_qasymm8_signed_neon(const ITensor *in1, const ITensor *in2, ITensor *out, const ConvertPolicy &policy, const Window &window)
+{
+    ARM_COMPUTE_UNUSED(policy);
+
+    // Create input windows
+    Window input1_win = window.broadcast_if_dimension_le_one(in1->info()->tensor_shape());
+    Window input2_win = window.broadcast_if_dimension_le_one(in2->info()->tensor_shape());
+
+    // Clear X Dimension on execution window as we handle manually
+    Window win = window;
+    win.set(Window::DimX, Window::Dimension(0, 1, 1));
+
+    const int  window_step_x         = 16;
+    const auto window_start_x        = static_cast<int>(window.x().start());
+    const auto window_end_x          = static_cast<int>(window.x().end());
+    const bool is_broadcast_across_x = in1->info()->tensor_shape().x() != in2->info()->tensor_shape().x();
+
+    const UniformQuantizationInfo iq1_info = in1->info()->quantization_info().uniform();
+    const UniformQuantizationInfo iq2_info = in2->info()->quantization_info().uniform();
+    const UniformQuantizationInfo oq_info  = out->info()->quantization_info().uniform();
+
+    const float32x4_t invvscaleo = vdupq_n_f32(1.f / oq_info.scale);
+    const float32x4_t voffseto   = vdupq_n_f32(oq_info.offset);
+
+    if(is_broadcast_across_x)
+    {
+        const bool                    is_broadcast_input_2 = input2_win.x().step() == 0;
+        Window                        broadcast_win        = is_broadcast_input_2 ? input2_win : input1_win;
+        Window                        non_broadcast_win    = !is_broadcast_input_2 ? input2_win : input1_win;
+        const ITensor                *broadcast_tensor     = is_broadcast_input_2 ? in2 : in1;
+        const ITensor                *non_broadcast_tensor = !is_broadcast_input_2 ? in2 : in1;
+        const UniformQuantizationInfo broadcast_qinfo      = broadcast_tensor->info()->quantization_info().uniform();
+        const UniformQuantizationInfo non_broadcast_qinfo  = non_broadcast_tensor->info()->quantization_info().uniform();
+
+        const float32x4_t vscale1  = is_broadcast_input_2 ? vdupq_n_f32(iq1_info.scale) : vdupq_n_f32(iq2_info.scale);
+        const float32x4_t vscale2  = is_broadcast_input_2 ? vdupq_n_f32(iq2_info.scale) : vdupq_n_f32(iq1_info.scale);
+        const int32x4_t   voffset1 = is_broadcast_input_2 ? vdupq_n_s32(iq1_info.offset) : vdupq_n_s32(iq2_info.offset);
+        const int32x4_t   voffset2 = is_broadcast_input_2 ? vdupq_n_s32(iq2_info.offset) : vdupq_n_s32(iq1_info.offset);
+
+        // Clear X Dimension on execution window as we handle manually
+        non_broadcast_win.set(Window::DimX, Window::Dimension(0, 1, 1));
+
+        Iterator broadcast_input(broadcast_tensor, broadcast_win);
+        Iterator non_broadcast_input(non_broadcast_tensor, non_broadcast_win);
+        Iterator output(out, win);
+
+        execute_window_loop(win, [&](const Coordinates &)
+        {
+            const auto non_broadcast_input_ptr = reinterpret_cast<const int8_t *>(non_broadcast_input.ptr());
+            const auto output_ptr              = reinterpret_cast<int8_t *>(output.ptr());
+
+            const int8_t    broadcast_value     = *reinterpret_cast<const int8_t *>(broadcast_input.ptr());
+            const int8x16_t broadcast_value_vec = vdupq_n_s8(broadcast_value);
+
+            const auto  bf_0 = vmulq_f32(vcvtq_f32_s32(vsubq_s32(vmovl_s16(vget_low_s16(vmovl_s8(vget_low_s8(broadcast_value_vec)))), voffset2)), vscale2);
+            const auto  bf_1 = vmulq_f32(vcvtq_f32_s32(vsubq_s32(vmovl_s16(vget_high_s16(vmovl_s8(vget_low_s8(broadcast_value_vec)))), voffset2)), vscale2);
+            const auto  bf_2 = vmulq_f32(vcvtq_f32_s32(vsubq_s32(vmovl_s16(vget_low_s16(vmovl_s8(vget_high_s8(broadcast_value_vec)))), voffset2)), vscale2);
+            const auto  bf_3 = vmulq_f32(vcvtq_f32_s32(vsubq_s32(vmovl_s16(vget_high_s16(vmovl_s8(vget_high_s8(broadcast_value_vec)))), voffset2)), vscale2);
+            const float bfs  = static_cast<int32_t>(broadcast_value - broadcast_qinfo.offset) * broadcast_qinfo.scale;
+
+            // Compute S elements per iteration
+            int x = window_start_x;
+            for(; x <= (window_end_x - window_step_x); x += window_step_x)
+            {
+                const int8x16_t a = vld1q_s8(non_broadcast_input_ptr + x);
+
+                const auto af_0 = vmulq_f32(vcvtq_f32_s32(vsubq_s32(vmovl_s16(vget_low_s16(vmovl_s8(vget_low_s8(a)))), voffset1)), vscale1);
+                const auto af_1 = vmulq_f32(vcvtq_f32_s32(vsubq_s32(vmovl_s16(vget_high_s16(vmovl_s8(vget_low_s8(a)))), voffset1)), vscale1);
+                const auto af_2 = vmulq_f32(vcvtq_f32_s32(vsubq_s32(vmovl_s16(vget_low_s16(vmovl_s8(vget_high_s8(a)))), voffset1)), vscale1);
+                const auto af_3 = vmulq_f32(vcvtq_f32_s32(vsubq_s32(vmovl_s16(vget_high_s16(vmovl_s8(vget_high_s8(a)))), voffset1)), vscale1);
+
+                int32x4_t rf_0{};
+                int32x4_t rf_1{};
+                int32x4_t rf_2{};
+                int32x4_t rf_3{};
+
+#ifdef __aarch64__
+                rf_0 = vcvtnq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af_0, bf_0), invvscaleo));
+                rf_1 = vcvtnq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af_1, bf_1), invvscaleo));
+                rf_2 = vcvtnq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af_2, bf_2), invvscaleo));
+                rf_3 = vcvtnq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af_3, bf_3), invvscaleo));
+#else  //__aarch64__
+                rf_0 = vcvtq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af_0, bf_0), invvscaleo));
+                rf_1 = vcvtq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af_1, bf_1), invvscaleo));
+                rf_2 = vcvtq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af_2, bf_2), invvscaleo));
+                rf_3 = vcvtq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af_3, bf_3), invvscaleo));
+#endif //__aarch64__
+
+                const int8x8_t pa = vqmovn_s16(vcombine_s16(vqmovn_s32(rf_0), vqmovn_s32(rf_1)));
+                const int8x8_t pb = vqmovn_s16(vcombine_s16(vqmovn_s32(rf_2), vqmovn_s32(rf_3)));
+                vst1q_s8(output_ptr + x, vcombine_s8(pa, pb));
+            }
+
+            // Compute left-over elements
+            for(; x < window_end_x; ++x)
+            {
+                const float afs   = static_cast<int32_t>(*(non_broadcast_input_ptr + x) - non_broadcast_qinfo.offset) * non_broadcast_qinfo.scale;
+                *(output_ptr + x) = quantize_qasymm8_signed((afs + bfs), oq_info);
+            }
+        },
+        broadcast_input, non_broadcast_input, output);
+    }
+    else
+    {
+        // Clear X Dimension on execution window as we handle manually
+        input1_win.set(Window::DimX, Window::Dimension(0, 1, 1));
+        input2_win.set(Window::DimX, Window::Dimension(0, 1, 1));
+
+        Iterator input1(in1, input1_win);
+        Iterator input2(in2, input2_win);
+        Iterator output(out, win);
+
+        const float32x4_t vscale1  = vdupq_n_f32(iq1_info.scale);
+        const float32x4_t vscale2  = vdupq_n_f32(iq2_info.scale);
+        const int32x4_t   voffset1 = vdupq_n_s32(iq1_info.offset);
+        const int32x4_t   voffset2 = vdupq_n_s32(iq2_info.offset);
+        execute_window_loop(win, [&](const Coordinates &)
+        {
+            const auto input1_ptr = reinterpret_cast<const int8_t *>(input1.ptr());
+            const auto input2_ptr = reinterpret_cast<const int8_t *>(input2.ptr());
+            const auto output_ptr = reinterpret_cast<int8_t *>(output.ptr());
+
+            // Compute S elements per iteration
+            int x = window_start_x;
+            for(; x <= (window_end_x - window_step_x); x += window_step_x)
+            {
+                const int8x16_t a = vld1q_s8(input1_ptr + x);
+                const int8x16_t b = vld1q_s8(input2_ptr + x);
+
+                const auto af_0 = vmulq_f32(vcvtq_f32_s32(vsubq_s32(vmovl_s16(vget_low_s16(vmovl_s8(vget_low_s8(a)))), voffset1)), vscale1);
+                const auto af_1 = vmulq_f32(vcvtq_f32_s32(vsubq_s32(vmovl_s16(vget_high_s16(vmovl_s8(vget_low_s8(a)))), voffset1)), vscale1);
+                const auto af_2 = vmulq_f32(vcvtq_f32_s32(vsubq_s32(vmovl_s16(vget_low_s16(vmovl_s8(vget_high_s8(a)))), voffset1)), vscale1);
+                const auto af_3 = vmulq_f32(vcvtq_f32_s32(vsubq_s32(vmovl_s16(vget_high_s16(vmovl_s8(vget_high_s8(a)))), voffset1)), vscale1);
+
+                const auto bf_0 = vmulq_f32(vcvtq_f32_s32(vsubq_s32(vmovl_s16(vget_low_s16(vmovl_s8(vget_low_s8(b)))), voffset2)), vscale2);
+                const auto bf_1 = vmulq_f32(vcvtq_f32_s32(vsubq_s32(vmovl_s16(vget_high_s16(vmovl_s8(vget_low_s8(b)))), voffset2)), vscale2);
+                const auto bf_2 = vmulq_f32(vcvtq_f32_s32(vsubq_s32(vmovl_s16(vget_low_s16(vmovl_s8(vget_high_s8(b)))), voffset2)), vscale2);
+                const auto bf_3 = vmulq_f32(vcvtq_f32_s32(vsubq_s32(vmovl_s16(vget_high_s16(vmovl_s8(vget_high_s8(b)))), voffset2)), vscale2);
+
+                int32x4_t rf_0{};
+                int32x4_t rf_1{};
+                int32x4_t rf_2{};
+                int32x4_t rf_3{};
+
+#ifdef __aarch64__
+                rf_0 = vcvtnq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af_0, bf_0), invvscaleo));
+                rf_1 = vcvtnq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af_1, bf_1), invvscaleo));
+                rf_2 = vcvtnq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af_2, bf_2), invvscaleo));
+                rf_3 = vcvtnq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af_3, bf_3), invvscaleo));
+#else  //__aarch64__
+                rf_0 = vcvtq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af_0, bf_0), invvscaleo));
+                rf_1 = vcvtq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af_1, bf_1), invvscaleo));
+                rf_2 = vcvtq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af_2, bf_2), invvscaleo));
+                rf_3 = vcvtq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af_3, bf_3), invvscaleo));
+#endif //__aarch64__
+
+                const int8x8_t pa = vqmovn_s16(vcombine_s16(vqmovn_s32(rf_0), vqmovn_s32(rf_1)));
+                const int8x8_t pb = vqmovn_s16(vcombine_s16(vqmovn_s32(rf_2), vqmovn_s32(rf_3)));
+                vst1q_s8(output_ptr + x, vcombine_s8(pa, pb));
+            }
+
+            // Compute left-over elements
+            for(; x < window_end_x; ++x)
+            {
+                const float afs   = static_cast<int32_t>((*(input1_ptr + x)) - iq1_info.offset) * iq1_info.scale;
+                const float bfs   = static_cast<int32_t>((*(input2_ptr + x)) - iq2_info.offset) * iq2_info.scale;
+                *(output_ptr + x) = quantize_qasymm8_signed((afs + bfs), out->info()->quantization_info());
+            }
+        },
+        input1, input2, output);
+    }
+}
+} // namespace cpu
+} // namespace arm_compute
\ No newline at end of file
diff --git a/src/core/NEON/kernels/arithmetic_addition/impl/NEON/qsymm16.cpp b/src/core/NEON/kernels/arithmetic_addition/impl/NEON/qsymm16.cpp
new file mode 100644
index 0000000000..538c600187
--- /dev/null
+++ b/src/core/NEON/kernels/arithmetic_addition/impl/NEON/qsymm16.cpp
@@ -0,0 +1,175 @@
+/*
+ * Copyright (c) 2020-2021 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.
+ */
+#include "arm_compute/core/Helpers.h"
+#include "arm_compute/core/ITensor.h"
+#include "arm_compute/core/Types.h"
+#include "arm_compute/core/utils/misc/Traits.h"
+#include "src/core/NEON/wrapper/intrinsics/intrinsics.h"
+#include "src/core/common/StdTypes.h"
+#include "src/core/helpers/WindowHelpers.h"
+
+namespace arm_compute
+{
+namespace cpu
+{
+void arithmetic_addition_qsymm16_neon(const ITensor *in1, const ITensor *in2, ITensor *out, const ConvertPolicy &policy, const Window &window)
+{
+    ARM_COMPUTE_UNUSED(policy);
+
+    // Create input windows
+    Window input1_win = window.broadcast_if_dimension_le_one(in1->info()->tensor_shape());
+    Window input2_win = window.broadcast_if_dimension_le_one(in2->info()->tensor_shape());
+
+    // Clear X Dimension on execution window as we handle manually
+    Window win = window;
+    win.set(Window::DimX, Window::Dimension(0, 1, 1));
+
+    const int  window_step_x         = 8;
+    const auto window_start_x        = static_cast<int>(window.x().start());
+    const auto window_end_x          = static_cast<int>(window.x().end());
+    const bool is_broadcast_across_x = in1->info()->tensor_shape().x() != in2->info()->tensor_shape().x();
+
+    const UniformQuantizationInfo iq1_info = in1->info()->quantization_info().uniform();
+    const UniformQuantizationInfo iq2_info = in2->info()->quantization_info().uniform();
+    const UniformQuantizationInfo oq_info  = out->info()->quantization_info().uniform();
+
+    const float32x4_t vscale1    = vdupq_n_f32(iq1_info.scale);
+    const float32x4_t vscale2    = vdupq_n_f32(iq2_info.scale);
+    const float32x4_t invvscaleo = vdupq_n_f32(1.f / oq_info.scale);
+
+    if(is_broadcast_across_x)
+    {
+        const bool                    is_broadcast_input_2 = input2_win.x().step() == 0;
+        Window                        broadcast_win        = is_broadcast_input_2 ? input2_win : input1_win;
+        Window                        non_broadcast_win    = !is_broadcast_input_2 ? input2_win : input1_win;
+        const ITensor                *broadcast_tensor     = is_broadcast_input_2 ? in2 : in1;
+        const ITensor                *non_broadcast_tensor = !is_broadcast_input_2 ? in2 : in1;
+        const UniformQuantizationInfo broadcast_qinfo      = broadcast_tensor->info()->quantization_info().uniform();
+        const UniformQuantizationInfo non_broadcast_qinfo  = non_broadcast_tensor->info()->quantization_info().uniform();
+
+        // Clear X Dimension on execution window as we handle manually
+        non_broadcast_win.set(Window::DimX, Window::Dimension(0, 1, 1));
+
+        Iterator broadcast_input(broadcast_tensor, broadcast_win);
+        Iterator non_broadcast_input(non_broadcast_tensor, non_broadcast_win);
+        Iterator output(out, win);
+
+        execute_window_loop(win, [&](const Coordinates &)
+        {
+            const auto non_broadcast_input_ptr = reinterpret_cast<const int16_t *>(non_broadcast_input.ptr());
+            const auto output_ptr              = reinterpret_cast<int16_t *>(output.ptr());
+
+            const int16_t   broadcast_value     = *reinterpret_cast<const int16_t *>(broadcast_input.ptr());
+            const int16x8_t broadcast_value_vec = vdupq_n_s16(broadcast_value);
+
+            const auto  bf_0 = vmulq_f32(vcvtq_f32_s32(vmovl_s16(vget_low_s16(broadcast_value_vec))), vscale2);
+            const auto  bf_1 = vmulq_f32(vcvtq_f32_s32(vmovl_s16(vget_high_s16(broadcast_value_vec))), vscale2);
+            const float bfs  = static_cast<int32_t>(broadcast_value) * broadcast_qinfo.scale;
+
+            // Compute S elements per iteration
+            int x = window_start_x;
+            for(; x <= (window_end_x - window_step_x); x += window_step_x)
+            {
+                const int16x8_t a    = vld1q_s16(non_broadcast_input_ptr + x);
+                const auto      af_0 = vmulq_f32(vcvtq_f32_s32(vmovl_s16(vget_low_s16(a))), vscale1);
+                const auto      af_1 = vmulq_f32(vcvtq_f32_s32(vmovl_s16(vget_high_s16(a))), vscale1);
+
+                int32x4_t rf_0{};
+                int32x4_t rf_1{};
+#ifdef __aarch64__
+                rf_0 = vcvtnq_s32_f32(vmulq_f32(vaddq_f32(af_0, bf_0), invvscaleo));
+                rf_1 = vcvtnq_s32_f32(vmulq_f32(vaddq_f32(af_1, bf_1), invvscaleo));
+#else  //__aarch64__
+                rf_0 = vcvtq_s32_f32(vmulq_f32(vaddq_f32(af_0, bf_0), invvscaleo));
+                rf_1 = vcvtq_s32_f32(vmulq_f32(vaddq_f32(af_1, bf_1), invvscaleo));
+#endif //__aarch64__
+
+                const int16x8_t pa = vcombine_s16(vqmovn_s32(rf_0), vqmovn_s32(rf_1));
+                vst1q_s16(output_ptr + x, pa);
+            }
+
+            // Compute left-over elements
+            for(; x < window_end_x; ++x)
+            {
+                const float afs   = static_cast<int32_t>(*(non_broadcast_input_ptr + x)) * non_broadcast_qinfo.scale;
+                *(output_ptr + x) = quantize_qsymm16((afs + bfs), oq_info);
+            }
+        },
+        broadcast_input, non_broadcast_input, output);
+    }
+    else
+    {
+        // Clear X Dimension on execution window as we handle manually
+        input1_win.set(Window::DimX, Window::Dimension(0, 1, 1));
+        input2_win.set(Window::DimX, Window::Dimension(0, 1, 1));
+
+        Iterator input1(in1, input1_win);
+        Iterator input2(in2, input2_win);
+        Iterator output(out, win);
+
+        execute_window_loop(win, [&](const Coordinates &)
+        {
+            const auto input1_ptr = reinterpret_cast<const int16_t *>(input1.ptr());
+            const auto input2_ptr = reinterpret_cast<const int16_t *>(input2.ptr());
+            const auto output_ptr = reinterpret_cast<int16_t *>(output.ptr());
+
+            // Compute S elements per iteration
+            int x = window_start_x;
+            for(; x <= (window_end_x - window_step_x); x += window_step_x)
+            {
+                const int16x8_t a = vld1q_s16(input1_ptr + x);
+                const int16x8_t b = vld1q_s16(input2_ptr + x);
+
+                const auto af_0 = vmulq_f32(vcvtq_f32_s32(vmovl_s16(vget_low_s16(a))), vscale1);
+                const auto af_1 = vmulq_f32(vcvtq_f32_s32(vmovl_s16(vget_high_s16(a))), vscale1);
+                const auto bf_0 = vmulq_f32(vcvtq_f32_s32(vmovl_s16(vget_low_s16(b))), vscale2);
+                const auto bf_1 = vmulq_f32(vcvtq_f32_s32(vmovl_s16(vget_high_s16(b))), vscale2);
+
+                int32x4_t rf_0{};
+                int32x4_t rf_1{};
+#ifdef __aarch64__
+                rf_0 = vcvtnq_s32_f32(vmulq_f32(vaddq_f32(af_0, bf_0), invvscaleo));
+                rf_1 = vcvtnq_s32_f32(vmulq_f32(vaddq_f32(af_1, bf_1), invvscaleo));
+#else  //__aarch64__
+                rf_0 = vcvtq_s32_f32(vmulq_f32(vaddq_f32(af_0, bf_0), invvscaleo));
+                rf_1 = vcvtq_s32_f32(vmulq_f32(vaddq_f32(af_1, bf_1), invvscaleo));
+#endif //__aarch64__
+
+                const int16x8_t pa = vcombine_s16(vqmovn_s32(rf_0), vqmovn_s32(rf_1));
+                vst1q_s16(output_ptr + x, pa);
+            }
+
+            // Compute left-over elements
+            for(; x < window_end_x; ++x)
+            {
+                const float afs   = static_cast<int32_t>((*(input1_ptr + x))) * iq1_info.scale;
+                const float bfs   = static_cast<int32_t>((*(input2_ptr + x))) * iq2_info.scale;
+                *(output_ptr + x) = quantize_qsymm16((afs + bfs), out->info()->quantization_info());
+            }
+        },
+        input1, input2, output);
+    }
+}
+} // namespace cpu
+} // namespace arm_compute
\ No newline at end of file
diff --git a/src/core/NEON/kernels/arithmetic_addition/impl/SVE/integer.cpp b/src/core/NEON/kernels/arithmetic_addition/impl/SVE/integer.cpp
new file mode 100644
index 0000000000..c502a0235e
--- /dev/null
+++ b/src/core/NEON/kernels/arithmetic_addition/impl/SVE/integer.cpp
@@ -0,0 +1,201 @@
+/*
+ * Copyright (c) 2020-2021 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.
+ */
+#include "arm_compute/core/Helpers.h"
+#include "arm_compute/core/ITensor.h"
+#include "arm_compute/core/Types.h"
+#include "arm_compute/core/utils/misc/Traits.h"
+#include "src/core/NEON/wrapper/intrinsics/intrinsics.h"
+#if defined(__ARM_FEATURE_SVE)
+#include "src/core/NEON/SVEMath.h"
+#include <arm_sve.h>
+
+namespace arm_compute
+{
+namespace cpu
+{
+void arithmetic_addition_U8_U8_S16_sve(const ITensor *in1, const ITensor *in2, ITensor *out, const ConvertPolicy &policy, const Window &window)
+{
+    // Create input windows
+    Window win        = window;
+    Window input1_win = window.broadcast_if_dimension_le_one(in1->info()->tensor_shape());
+    Window input2_win = window.broadcast_if_dimension_le_one(in2->info()->tensor_shape());
+
+    // Clear X Dimension on execution window as we handle manually
+    win.set(Window::DimX, Window::Dimension(0, 1, 1));
+    input1_win.set(Window::DimX, Window::Dimension(0, 1, 1));
+    input2_win.set(Window::DimX, Window::Dimension(0, 1, 1));
+
+    Iterator input1(in1, input1_win);
+    Iterator input2(in2, input2_win);
+    Iterator output(out, win);
+
+    const auto window_start_x = static_cast<int>(window.x().start());
+    const auto window_end_x   = static_cast<int>(window.x().end());
+    const auto all_true_pg    = svptrue_b8();
+
+    execute_window_loop(win, [&](const Coordinates &)
+    {
+        const auto input1_ptr = reinterpret_cast<const uint8_t *>(input1.ptr());
+        const auto input2_ptr = reinterpret_cast<const uint8_t *>(input2.ptr());
+        const auto output_ptr = reinterpret_cast<int16_t *>(output.ptr());
+
+        if(policy == ConvertPolicy::WRAP)
+        {
+            int      x    = window_start_x;
+            svbool_t pg_u = svwhilelt_b8(x, window_end_x);
+            svbool_t pg_0 = svwhilelt_b16(x, window_end_x);
+            svbool_t pg_1 = svwhilelt_b16(x, static_cast<int>(window_end_x + svcnth()));
+            do
+            {
+                const auto vin1 = svld1(pg_u, input1_ptr + x);
+                const auto vin2 = svld1(pg_u, input2_ptr + x);
+
+                const auto vin1_lo = svreinterpret_s16_u16(svunpklo(vin1));
+                const auto vin1_hi = svreinterpret_s16_u16(svunpkhi(vin1));
+                const auto vin2_lo = svreinterpret_s16_u16(svunpklo(vin2));
+                const auto vin2_hi = svreinterpret_s16_u16(svunpkhi(vin2));
+                svst1(pg_0, output_ptr + x, svqadd(vin1_lo, vin2_lo));
+                svst1(pg_1, output_ptr + x + svcnth(), svqadd(vin1_hi, vin2_hi));
+
+                x += svcntb();
+                pg_u = svwhilelt_b8(x, window_end_x);
+                pg_0 = svwhilelt_b16(x, window_end_x);
+                pg_1 = svwhilelt_b16(x, static_cast<int>(window_end_x + svcnth()));
+            }
+            while(svptest_any(all_true_pg, pg_u));
+        }
+        else
+        {
+            int      x    = window_start_x;
+            svbool_t pg_u = svwhilelt_b8(x, window_end_x);
+            svbool_t pg_0 = svwhilelt_b16(x, window_end_x);
+            svbool_t pg_1 = svwhilelt_b16(x, static_cast<int>(window_end_x + svcnth()));
+            do
+            {
+                const auto vin1 = svld1(pg_u, input1_ptr + x);
+                const auto vin2 = svld1(pg_u, input2_ptr + x);
+
+                const auto vin1_lo = svreinterpret_s16_u16(svunpklo(vin1));
+                const auto vin1_hi = svreinterpret_s16_u16(svunpkhi(vin1));
+                const auto vin2_lo = svreinterpret_s16_u16(svunpklo(vin2));
+                const auto vin2_hi = svreinterpret_s16_u16(svunpkhi(vin2));
+                svst1(pg_0, output_ptr + x, svqadd(vin1_lo, vin2_lo));
+                svst1(pg_1, output_ptr + x + svcnth(), svqadd(vin1_hi, vin2_hi));
+
+                x += svcntb();
+                pg_u = svwhilelt_b8(x, window_end_x);
+                pg_0 = svwhilelt_b16(x, window_end_x);
+                pg_1 = svwhilelt_b16(x, static_cast<int>(window_end_x + svcnth()));
+            }
+            while(svptest_any(all_true_pg, pg_u));
+        }
+    },
+    input1, input2, output);
+}
+
+void arithmetic_addition_S16_U8_S16_sve(const ITensor *in1, const ITensor *in2, ITensor *out, const ConvertPolicy &policy, const Window &window)
+{
+    // Create input windows
+    Window win        = window;
+    Window input1_win = window.broadcast_if_dimension_le_one(in1->info()->tensor_shape());
+    Window input2_win = window.broadcast_if_dimension_le_one(in2->info()->tensor_shape());
+
+    // Clear X Dimension on execution window as we handle manually
+    win.set(Window::DimX, Window::Dimension(0, 1, 1));
+    input1_win.set(Window::DimX, Window::Dimension(0, 1, 1));
+    input2_win.set(Window::DimX, Window::Dimension(0, 1, 1));
+
+    Iterator input1(in1, input1_win);
+    Iterator input2(in2, input2_win);
+    Iterator output(out, win);
+
+    const auto window_start_x = static_cast<int>(window.x().start());
+    const auto window_end_x   = static_cast<int>(window.x().end());
+    const auto all_true_pg    = svptrue_b8();
+
+    execute_window_loop(win, [&](const Coordinates &)
+    {
+        const auto input1_ptr = reinterpret_cast<const int16_t *>(input1.ptr());
+        const auto input2_ptr = reinterpret_cast<const uint8_t *>(input2.ptr());
+        const auto output_ptr = reinterpret_cast<int16_t *>(output.ptr());
+
+        if(policy == ConvertPolicy::WRAP)
+        {
+            int      x    = window_start_x;
+            svbool_t pg_u = svwhilelt_b8(x, window_end_x);
+            svbool_t pg_0 = svwhilelt_b16(x, window_end_x);
+            svbool_t pg_1 = svwhilelt_b16(x + static_cast<int>(svcnth()), window_end_x);
+            do
+            {
+                const auto vin1_0  = svld1_s16(pg_0, input1_ptr + x);
+                const auto vin1_1  = svld1_s16(pg_1, input1_ptr + x + svcnth());
+                const auto vin2_u8 = svld1_u8(pg_u, input2_ptr + x);
+                const auto vin2_0  = svreinterpret_s16_u16(svunpklo(vin2_u8));
+                const auto vin2_1  = svreinterpret_s16_u16(svunpkhi(vin2_u8));
+                svst1_s16(pg_0, output_ptr + x, svadd_s16_z(pg_0, vin1_0, vin2_0));
+                svst1_s16(pg_1, output_ptr + x, svadd_s16_z(pg_1, vin1_1, vin2_1));
+
+                x += svcnth();
+                pg_u = svwhilelt_b8(x, window_end_x);
+                pg_0 = svwhilelt_b16(x, window_end_x);
+                pg_1 = svwhilelt_b16(x + static_cast<int>(svcnth()), window_end_x);
+            }
+            while(svptest_any(all_true_pg, pg_u));
+        }
+        else
+        {
+            int      x    = window_start_x;
+            svbool_t pg_u = svwhilelt_b8(x, window_end_x);
+            svbool_t pg_0 = svwhilelt_b16(x, window_end_x);
+            svbool_t pg_1 = svwhilelt_b16(x + static_cast<int>(svcnth()), window_end_x);
+            do
+            {
+                const auto vin1_0  = svld1_s16(pg_0, input1_ptr + x);
+                const auto vin1_1  = svld1_s16(pg_1, input1_ptr + x);
+                const auto vin2_u8 = svld1_u8(pg_u, input2_ptr + x);
+                const auto vin2_0  = svreinterpret_s16_u16(svunpklo(vin2_u8));
+                const auto vin2_1  = svreinterpret_s16_u16(svunpkhi(vin2_u8));
+
+                svst1_s16(pg_0, output_ptr + x, svqadd(vin1_0, vin2_0));
+                svst1_s16(pg_1, output_ptr + x, svqadd(vin1_1, vin2_1));
+
+                x += svcnth();
+                pg_u = svwhilelt_b8(x, window_end_x);
+                pg_0 = svwhilelt_b16(x, window_end_x);
+                pg_1 = svwhilelt_b16(x + static_cast<int>(svcnth()), window_end_x);
+            }
+            while(svptest_any(all_true_pg, pg_u));
+        }
+    },
+    input1, input2, output);
+}
+
+void arithmetic_addition_U8_S16_S16_sve(const ITensor *input1, const ITensor *input2, ITensor *output, const ConvertPolicy &policy, const Window &window)
+{
+    // Simply swap the two input buffers:
+    arithmetic_addition_S16_U8_S16_sve(input2, input1, output, policy, window);
+}
+} // namespace cpu
+} // namespace arm_compute
+#endif /* defined(__ARM_FEATURE_SVE) */
\ No newline at end of file
diff --git a/src/core/NEON/kernels/arithmetic_addition/impl/SVE/list.h b/src/core/NEON/kernels/arithmetic_addition/impl/SVE/list.h
new file mode 100644
index 0000000000..3e238c40d0
--- /dev/null
+++ b/src/core/NEON/kernels/arithmetic_addition/impl/SVE/list.h
@@ -0,0 +1,145 @@
+/*
+ * Copyright (c) 2020-2021 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.
+ */
+#ifndef SRC_CORE_SVE_KERNELS_ARITHMETIC_ADDITION_LIST_H
+#define SRC_CORE_SVE_KERNELS_ARITHMETIC_ADDITION_LIST_H
+
+#if defined(__ARM_FEATURE_SVE)
+#include "arm_compute/core/Types.h"
+#include "arm_compute/core/utils/misc/Traits.h"
+#include "src/core/NEON/SVEMath.h"
+#include "src/core/NEON/wrapper/intrinsics/intrinsics.h"
+#include <arm_sve.h>
+
+namespace arm_compute
+{
+namespace cpu
+{
+#define DECLARE_ARITHMETIC_ADDITION_KERNEL(func_name) \
+    void func_name(const ITensor *in1, const ITensor *in2, ITensor *out, const ConvertPolicy &policy, const Window &window)
+
+DECLARE_ARITHMETIC_ADDITION_KERNEL(arithmetic_addition_qasymm8_sve);
+DECLARE_ARITHMETIC_ADDITION_KERNEL(arithmetic_addition_qasymm8_signed_sve);
+DECLARE_ARITHMETIC_ADDITION_KERNEL(arithmetic_addition_qsymm16_sve);
+DECLARE_ARITHMETIC_ADDITION_KERNEL(arithmetic_addition_S16_U8_S16_sve);
+DECLARE_ARITHMETIC_ADDITION_KERNEL(arithmetic_addition_U8_S16_S16_sve);
+DECLARE_ARITHMETIC_ADDITION_KERNEL(arithmetic_addition_U8_U8_S16_sve);
+
+#undef DECLARE_ARITHMETIC_ADDITION_KERNEL
+
+template <typename ScalarType>
+void arithmetic_addition_same_sve(const ITensor *in1, const ITensor *in2, ITensor *out, const ConvertPolicy &policy, const Window &window)
+{
+    const auto all_true_pg           = wrapper::svptrue<ScalarType>();
+    const auto window_start_x        = static_cast<int>(window.x().start());
+    const auto window_end_x          = static_cast<int>(window.x().end());
+    const bool is_broadcast_across_x = in1->info()->tensor_shape().x() != in2->info()->tensor_shape().x();
+    const bool is_sat                = (policy == ConvertPolicy::SATURATE);
+
+    // Clear X Dimension on execution window as we handle manually
+    Window win = window;
+    win.set(Window::DimX, Window::Dimension(0, 1, 1));
+
+    // Create input windows
+    Window input1_win = window.broadcast_if_dimension_le_one(in1->info()->tensor_shape());
+    Window input2_win = window.broadcast_if_dimension_le_one(in2->info()->tensor_shape());
+
+    Iterator input1(in1, window.broadcast_if_dimension_le_one(in1->info()->tensor_shape()));
+    Iterator input2(in2, window.broadcast_if_dimension_le_one(in2->info()->tensor_shape()));
+    Iterator output(out, window);
+
+    if(is_broadcast_across_x)
+    {
+        const bool     is_broadcast_input_2 = input2_win.x().step() == 0;
+        Window         broadcast_win        = is_broadcast_input_2 ? input2_win : input1_win;
+        Window         non_broadcast_win    = !is_broadcast_input_2 ? input2_win : input1_win;
+        const ITensor *broadcast_tensor     = is_broadcast_input_2 ? in2 : in1;
+        const ITensor *non_broadcast_tensor = !is_broadcast_input_2 ? in2 : in1;
+
+        // Clear X Dimension on execution window as we handle manually
+        non_broadcast_win.set(Window::DimX, Window::Dimension(0, 1, 1));
+
+        Iterator broadcast_input(broadcast_tensor, broadcast_win);
+        Iterator non_broadcast_input(non_broadcast_tensor, non_broadcast_win);
+        Iterator output(out, win);
+
+        execute_window_loop(win, [&](const Coordinates &)
+        {
+            const auto non_broadcast_input_ptr = reinterpret_cast<const ScalarType *>(non_broadcast_input.ptr());
+            const auto output_ptr              = reinterpret_cast<ScalarType *>(output.ptr());
+
+            const ScalarType broadcast_value     = *reinterpret_cast<const ScalarType *>(broadcast_input.ptr());
+            const auto       broadcast_value_vec = wrapper::svdup_n(broadcast_value);
+
+            int      x  = window_start_x;
+            svbool_t pg = wrapper::svwhilelt<ScalarType>(x, window_end_x);
+            do
+            {
+                const auto non_broadcast_v = svld1(pg, non_broadcast_input_ptr + x);
+                auto       res             = is_sat ? wrapper::svqadd(broadcast_value_vec, non_broadcast_v) : svadd_z(pg, broadcast_value_vec, non_broadcast_v);
+                svst1(pg, output_ptr + x, res);
+
+                x += wrapper::svcnt<ScalarType>();
+                pg = wrapper::svwhilelt<ScalarType>(x, window_end_x);
+            }
+            while(svptest_any(all_true_pg, pg));
+        },
+        broadcast_input, non_broadcast_input, output);
+    }
+    else
+    {
+        // Clear X Dimension on execution window as we handle manually
+        input1_win.set(Window::DimX, Window::Dimension(0, 1, 1));
+        input2_win.set(Window::DimX, Window::Dimension(0, 1, 1));
+
+        Iterator input1(in1, input1_win);
+        Iterator input2(in2, input2_win);
+        Iterator output(out, win);
+
+        execute_window_loop(win, [&](const Coordinates &)
+        {
+            const auto input1_ptr = reinterpret_cast<const ScalarType *>(input1.ptr());
+            const auto input2_ptr = reinterpret_cast<const ScalarType *>(input2.ptr());
+            const auto output_ptr = reinterpret_cast<ScalarType *>(output.ptr());
+
+            int      x  = window_start_x;
+            svbool_t pg = wrapper::svwhilelt<ScalarType>(x, window_end_x);
+            do
+            {
+                const auto val1 = svld1(pg, input1_ptr + x);
+                const auto val2 = svld1(pg, input2_ptr + x);
+                const auto res  = is_sat ? wrapper::svqadd(val1, val2) : svadd_z(pg, val1, val2);
+                svst1(pg, output_ptr + x, res);
+
+                x += wrapper::svcnt<ScalarType>();
+                pg = wrapper::svwhilelt<ScalarType>(x, window_end_x);
+            }
+            while(svptest_any(all_true_pg, pg));
+        },
+        input1, input2, output);
+    }
+}
+} // namespace cpu
+} // namespace arm_compute
+#endif // defined(__ARM_FEATURE_SVE)
+#endif // SRC_CORE_SVE_KERNELS_ARITHMETIC_ADDITION_LIST_H
\ No newline at end of file
diff --git a/src/core/NEON/kernels/arithmetic_addition/impl/SVE/qasymm8.cpp b/src/core/NEON/kernels/arithmetic_addition/impl/SVE/qasymm8.cpp
new file mode 100644
index 0000000000..871ee23ded
--- /dev/null
+++ b/src/core/NEON/kernels/arithmetic_addition/impl/SVE/qasymm8.cpp
@@ -0,0 +1,182 @@
+/*
+ * Copyright (c) 2020-2021 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.
+ */
+#include "arm_compute/core/Helpers.h"
+#include "arm_compute/core/ITensor.h"
+#include "arm_compute/core/Types.h"
+#include "arm_compute/core/utils/misc/Traits.h"
+#include "src/core/NEON/wrapper/intrinsics/intrinsics.h"
+#if defined(__ARM_FEATURE_SVE2)
+#include "src/core/NEON/SVEMath.h"
+#include <arm_sve.h>
+
+namespace arm_compute
+{
+namespace cpu
+{
+void arithmetic_addition_qasymm8_sve(const ITensor *in1, const ITensor *in2, ITensor *out, const ConvertPolicy &policy, const Window &window)
+{
+    ARM_COMPUTE_UNUSED(policy);
+
+    // Create input windows
+    Window input1_win = window.broadcast_if_dimension_le_one(in1->info()->tensor_shape());
+    Window input2_win = window.broadcast_if_dimension_le_one(in2->info()->tensor_shape());
+
+    // Clear X Dimension on execution window as we handle manually
+    Window win = window;
+    win.set(Window::DimX, Window::Dimension(0, 1, 1));
+
+    const auto window_start_x        = static_cast<int>(window.x().start());
+    const auto window_end_x          = static_cast<int>(window.x().end());
+    const bool is_broadcast_across_x = in1->info()->tensor_shape().x() != in2->info()->tensor_shape().x();
+    const auto all_true_pg           = svptrue_b8();
+
+    const UniformQuantizationInfo iq1_info = in1->info()->quantization_info().uniform();
+    const UniformQuantizationInfo iq2_info = in2->info()->quantization_info().uniform();
+    const UniformQuantizationInfo oq_info  = out->info()->quantization_info().uniform();
+
+    const auto invvscaleo = svdup_n_f32(1.f / oq_info.scale);
+    const auto voffseto   = svdup_n_f32(oq_info.offset);
+
+    if(is_broadcast_across_x)
+    {
+        const bool     is_broadcast_input_2 = input2_win.x().step() == 0;
+        Window         broadcast_win        = is_broadcast_input_2 ? input2_win : input1_win;
+        Window         non_broadcast_win    = !is_broadcast_input_2 ? input2_win : input1_win;
+        const ITensor *broadcast_tensor     = is_broadcast_input_2 ? in2 : in1;
+        const ITensor *non_broadcast_tensor = !is_broadcast_input_2 ? in2 : in1;
+
+        const svfloat32_t vscale1  = is_broadcast_input_2 ? svdup_n_f32(iq1_info.scale) : svdup_n_f32(iq2_info.scale);
+        const svfloat32_t vscale2  = is_broadcast_input_2 ? svdup_n_f32(iq2_info.scale) : svdup_n_f32(iq1_info.scale);
+        const svint32_t   voffset1 = is_broadcast_input_2 ? svdup_n_s32(iq1_info.offset) : svdup_n_s32(iq2_info.offset);
+        const svint32_t   voffset2 = is_broadcast_input_2 ? svdup_n_s32(iq2_info.offset) : svdup_n_s32(iq1_info.offset);
+
+        // Clear X Dimension on execution window as we handle manually
+        non_broadcast_win.set(Window::DimX, Window::Dimension(0, 1, 1));
+
+        Iterator broadcast_input(broadcast_tensor, broadcast_win);
+        Iterator non_broadcast_input(non_broadcast_tensor, non_broadcast_win);
+        Iterator output(out, win);
+
+        execute_window_loop(win, [&](const Coordinates &)
+        {
+            const auto non_broadcast_input_ptr = reinterpret_cast<const uint8_t *>(non_broadcast_input.ptr());
+            const auto output_ptr              = reinterpret_cast<uint8_t *>(output.ptr());
+
+            const uint8_t   broadcast_value     = *reinterpret_cast<const uint8_t *>(broadcast_input.ptr());
+            const svuint8_t broadcast_value_vec = svdup_n_u8(broadcast_value);
+
+            int      x  = window_start_x;
+            svbool_t pg = svwhilelt_b8(x, window_end_x);
+
+            const auto bf_0 = svmul_f32_z(pg, svcvt_f32_s32_z(pg, svsub_s32_z(pg, svreinterpret_s32_u32(svmovlb_u32(svmovlb_u16(broadcast_value_vec))), voffset2)), vscale2);
+            const auto bf_1 = svmul_f32_z(pg, svcvt_f32_s32_z(pg, svsub_s32_z(pg, svreinterpret_s32_u32(svmovlt_u32(svmovlb_u16(broadcast_value_vec))), voffset2)), vscale2);
+            const auto bf_2 = svmul_f32_z(pg, svcvt_f32_s32_z(pg, svsub_s32_z(pg, svreinterpret_s32_u32(svmovlb_u32(svmovlt_u16(broadcast_value_vec))), voffset2)), vscale2);
+            const auto bf_3 = svmul_f32_z(pg, svcvt_f32_s32_z(pg, svsub_s32_z(pg, svreinterpret_s32_u32(svmovlt_u32(svmovlt_u16(broadcast_value_vec))), voffset2)), vscale2);
+
+            do
+            {
+                const svuint8_t a = svld1_u8(pg, non_broadcast_input_ptr + x);
+
+                const auto af_0 = svmul_f32_z(pg, svcvt_f32_s32_z(pg, svsub_s32_z(pg, svreinterpret_s32_u32(svmovlb_u32(svmovlb_u16(a))), voffset1)), vscale1);
+                const auto af_1 = svmul_f32_z(pg, svcvt_f32_s32_z(pg, svsub_s32_z(pg, svreinterpret_s32_u32(svmovlt_u32(svmovlb_u16(a))), voffset1)), vscale1);
+                const auto af_2 = svmul_f32_z(pg, svcvt_f32_s32_z(pg, svsub_s32_z(pg, svreinterpret_s32_u32(svmovlb_u32(svmovlt_u16(a))), voffset1)), vscale1);
+                const auto af_3 = svmul_f32_z(pg, svcvt_f32_s32_z(pg, svsub_s32_z(pg, svreinterpret_s32_u32(svmovlt_u32(svmovlt_u16(a))), voffset1)), vscale1);
+
+                const auto rf_0 = svcvt_u32_f32_z(pg, svmla_f32_z(pg, voffseto, svadd_f32_z(pg, af_0, bf_0), invvscaleo));
+                const auto rf_1 = svcvt_u32_f32_z(pg, svmla_f32_z(pg, voffseto, svadd_f32_z(pg, af_1, bf_1), invvscaleo));
+                const auto rf_2 = svcvt_u32_f32_z(pg, svmla_f32_z(pg, voffseto, svadd_f32_z(pg, af_2, bf_2), invvscaleo));
+                const auto rf_3 = svcvt_u32_f32_z(pg, svmla_f32_z(pg, voffseto, svadd_f32_z(pg, af_3, bf_3), invvscaleo));
+
+                const auto pa = svqxtnt_u32(svqxtnb_u32(rf_0), rf_1);
+                const auto pb = svqxtnt_u32(svqxtnb_u32(rf_2), rf_3);
+
+                const auto res = svqxtnt_u16(svqxtnb_u16(pa), pb);
+                svst1_u8(pg, output_ptr + x, res);
+
+                x += svcntb();
+                pg = svwhilelt_b8(x, window_end_x);
+            }
+            while(svptest_any(all_true_pg, pg));
+        },
+        broadcast_input, non_broadcast_input, output);
+    }
+    else
+    {
+        // Clear X Dimension on execution window as we handle manually
+        input1_win.set(Window::DimX, Window::Dimension(0, 1, 1));
+        input2_win.set(Window::DimX, Window::Dimension(0, 1, 1));
+
+        Iterator input1(in1, input1_win);
+        Iterator input2(in2, input2_win);
+        Iterator output(out, win);
+
+        const auto vscale1  = svdup_n_f32(iq1_info.scale);
+        const auto vscale2  = svdup_n_f32(iq2_info.scale);
+        const auto voffset1 = svdup_n_s32(iq1_info.offset);
+        const auto voffset2 = svdup_n_s32(iq2_info.offset);
+
+        execute_window_loop(win, [&](const Coordinates &)
+        {
+            const auto input1_ptr = reinterpret_cast<const uint8_t *>(input1.ptr());
+            const auto input2_ptr = reinterpret_cast<const uint8_t *>(input2.ptr());
+            const auto output_ptr = reinterpret_cast<uint8_t *>(output.ptr());
+
+            int      x  = window_start_x;
+            svbool_t pg = svwhilelt_b8(x, window_end_x);
+            do
+            {
+                const auto a    = svld1_u8(pg, input1_ptr + x);
+                const auto b    = svld1_u8(pg, input2_ptr + x);
+                const auto af_0 = svmul_f32_z(pg, svcvt_f32_s32_z(pg, svsub_s32_z(pg, svreinterpret_s32_u32(svmovlb_u32(svmovlb_u16(a))), voffset1)), vscale1);
+                const auto af_1 = svmul_f32_z(pg, svcvt_f32_s32_z(pg, svsub_s32_z(pg, svreinterpret_s32_u32(svmovlt_u32(svmovlb_u16(a))), voffset1)), vscale1);
+                const auto af_2 = svmul_f32_z(pg, svcvt_f32_s32_z(pg, svsub_s32_z(pg, svreinterpret_s32_u32(svmovlb_u32(svmovlt_u16(a))), voffset1)), vscale1);
+                const auto af_3 = svmul_f32_z(pg, svcvt_f32_s32_z(pg, svsub_s32_z(pg, svreinterpret_s32_u32(svmovlt_u32(svmovlt_u16(a))), voffset1)), vscale1);
+
+                const auto bf_0 = svmul_f32_z(pg, svcvt_f32_s32_z(pg, svsub_s32_z(pg, svreinterpret_s32_u32(svmovlb_u32(svmovlb_u16(b))), voffset2)), vscale2);
+                const auto bf_1 = svmul_f32_z(pg, svcvt_f32_s32_z(pg, svsub_s32_z(pg, svreinterpret_s32_u32(svmovlt_u32(svmovlb_u16(b))), voffset2)), vscale2);
+                const auto bf_2 = svmul_f32_z(pg, svcvt_f32_s32_z(pg, svsub_s32_z(pg, svreinterpret_s32_u32(svmovlb_u32(svmovlt_u16(b))), voffset2)), vscale2);
+                const auto bf_3 = svmul_f32_z(pg, svcvt_f32_s32_z(pg, svsub_s32_z(pg, svreinterpret_s32_u32(svmovlt_u32(svmovlt_u16(b))), voffset2)), vscale2);
+
+                const auto rf_0 = svcvt_u32_f32_z(pg, svmla_f32_z(pg, voffseto, svadd_f32_z(pg, af_0, bf_0), invvscaleo));
+                const auto rf_1 = svcvt_u32_f32_z(pg, svmla_f32_z(pg, voffseto, svadd_f32_z(pg, af_1, bf_1), invvscaleo));
+                const auto rf_2 = svcvt_u32_f32_z(pg, svmla_f32_z(pg, voffseto, svadd_f32_z(pg, af_2, bf_2), invvscaleo));
+                const auto rf_3 = svcvt_u32_f32_z(pg, svmla_f32_z(pg, voffseto, svadd_f32_z(pg, af_3, bf_3), invvscaleo));
+
+                const auto pa  = svqxtnt_u32(svqxtnb_u32(rf_0), rf_1);
+                const auto pb  = svqxtnt_u32(svqxtnb_u32(rf_2), rf_3);
+                const auto res = svqxtnt_u16(svqxtnb_u16(pa), pb);
+
+                svst1_u8(pg, output_ptr + x, res);
+
+                x += svcntb();
+                pg = svwhilelt_b8(x, window_end_x);
+            }
+            while(svptest_any(all_true_pg, pg));
+        },
+        input1, input2, output);
+    }
+}
+} // namespace cpu
+} // namespace arm_compute
+#endif /* defined(__ARM_FEATURE_SVE2) */
\ No newline at end of file
diff --git a/src/core/NEON/kernels/arithmetic_addition/impl/SVE/qasymm8_signed.cpp b/src/core/NEON/kernels/arithmetic_addition/impl/SVE/qasymm8_signed.cpp
new file mode 100644
index 0000000000..2ba5d39400
--- /dev/null
+++ b/src/core/NEON/kernels/arithmetic_addition/impl/SVE/qasymm8_signed.cpp
@@ -0,0 +1,181 @@
+/*
+ * Copyright (c) 2020-2021 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.
+ */
+#include "arm_compute/core/Helpers.h"
+#include "arm_compute/core/ITensor.h"
+#include "arm_compute/core/Types.h"
+#include "arm_compute/core/utils/misc/Traits.h"
+#include "src/core/NEON/wrapper/intrinsics/intrinsics.h"
+#if defined(__ARM_FEATURE_SVE2)
+#include "src/core/NEON/SVEMath.h"
+#include <arm_sve.h>
+
+namespace arm_compute
+{
+namespace cpu
+{
+void arithmetic_addition_qasymm8_signed_sve(const ITensor *in1, const ITensor *in2, ITensor *out, const ConvertPolicy &policy, const Window &window)
+{
+    ARM_COMPUTE_UNUSED(policy);
+
+    // Create input windows
+    Window input1_win = window.broadcast_if_dimension_le_one(in1->info()->tensor_shape());
+    Window input2_win = window.broadcast_if_dimension_le_one(in2->info()->tensor_shape());
+
+    // Clear X Dimension on execution window as we handle manually
+    Window win = window;
+    win.set(Window::DimX, Window::Dimension(0, 1, 1));
+
+    const auto window_start_x        = static_cast<int>(window.x().start());
+    const auto window_end_x          = static_cast<int>(window.x().end());
+    const bool is_broadcast_across_x = in1->info()->tensor_shape().x() != in2->info()->tensor_shape().x();
+
+    const UniformQuantizationInfo iq1_info = in1->info()->quantization_info().uniform();
+    const UniformQuantizationInfo iq2_info = in2->info()->quantization_info().uniform();
+    const UniformQuantizationInfo oq_info  = out->info()->quantization_info().uniform();
+
+    const auto invvscaleo = svdup_n_f32(1.f / oq_info.scale);
+    const auto voffseto   = svdup_n_f32(oq_info.offset);
+
+    if(is_broadcast_across_x)
+    {
+        const bool     is_broadcast_input_2 = input2_win.x().step() == 0;
+        Window         broadcast_win        = is_broadcast_input_2 ? input2_win : input1_win;
+        Window         non_broadcast_win    = !is_broadcast_input_2 ? input2_win : input1_win;
+        const ITensor *broadcast_tensor     = is_broadcast_input_2 ? in2 : in1;
+        const ITensor *non_broadcast_tensor = !is_broadcast_input_2 ? in2 : in1;
+        const auto     all_true_pg          = svptrue_b8();
+
+        const auto vscale1  = is_broadcast_input_2 ? svdup_n_f32(iq1_info.scale) : svdup_n_f32(iq2_info.scale);
+        const auto vscale2  = is_broadcast_input_2 ? svdup_n_f32(iq2_info.scale) : svdup_n_f32(iq1_info.scale);
+        const auto voffset1 = is_broadcast_input_2 ? svdup_n_s32(iq1_info.offset) : svdup_n_s32(iq2_info.offset);
+        const auto voffset2 = is_broadcast_input_2 ? svdup_n_s32(iq2_info.offset) : svdup_n_s32(iq1_info.offset);
+
+        // Clear X Dimension on execution window as we handle manually
+        non_broadcast_win.set(Window::DimX, Window::Dimension(0, 1, 1));
+
+        Iterator broadcast_input(broadcast_tensor, broadcast_win);
+        Iterator non_broadcast_input(non_broadcast_tensor, non_broadcast_win);
+        Iterator output(out, win);
+
+        execute_window_loop(win, [&](const Coordinates &)
+        {
+            const auto non_broadcast_input_ptr = reinterpret_cast<const int8_t *>(non_broadcast_input.ptr());
+            const auto output_ptr              = reinterpret_cast<int8_t *>(output.ptr());
+
+            const int8_t broadcast_value     = *reinterpret_cast<const int8_t *>(broadcast_input.ptr());
+            const auto   broadcast_value_vec = svdup_n_s8(broadcast_value);
+
+            int        x    = window_start_x;
+            svbool_t   pg   = svwhilelt_b8(x, window_end_x);
+            const auto bf_0 = svmul_f32_z(pg, svcvt_f32_s32_z(pg, svsub_s32_z(pg, svmovlb_s32(svmovlb_s16(broadcast_value_vec)), voffset2)), vscale2);
+            const auto bf_1 = svmul_f32_z(pg, svcvt_f32_s32_z(pg, svsub_s32_z(pg, svmovlt_s32(svmovlb_s16(broadcast_value_vec)), voffset2)), vscale2);
+            const auto bf_2 = svmul_f32_z(pg, svcvt_f32_s32_z(pg, svsub_s32_z(pg, svmovlb_s32(svmovlt_s16(broadcast_value_vec)), voffset2)), vscale2);
+            const auto bf_3 = svmul_f32_z(pg, svcvt_f32_s32_z(pg, svsub_s32_z(pg, svmovlt_s32(svmovlt_s16(broadcast_value_vec)), voffset2)), vscale2);
+
+            do
+            {
+                const auto a    = svld1_s8(pg, non_broadcast_input_ptr + x);
+                const auto af_0 = svmul_f32_z(pg, svcvt_f32_s32_z(pg, svsub_s32_z(pg, svmovlb_s32(svmovlb_s16(a)), voffset1)), vscale1);
+                const auto af_1 = svmul_f32_z(pg, svcvt_f32_s32_z(pg, svsub_s32_z(pg, svmovlt_s32(svmovlb_s16(a)), voffset1)), vscale1);
+                const auto af_2 = svmul_f32_z(pg, svcvt_f32_s32_z(pg, svsub_s32_z(pg, svmovlb_s32(svmovlt_s16(a)), voffset1)), vscale1);
+                const auto af_3 = svmul_f32_z(pg, svcvt_f32_s32_z(pg, svsub_s32_z(pg, svmovlt_s32(svmovlt_s16(a)), voffset1)), vscale1);
+
+                const auto rf_0 = svcvt_s32_f32_z(pg, svmla_f32_z(pg, voffseto, svadd_f32_z(pg, af_0, bf_0), invvscaleo));
+                const auto rf_1 = svcvt_s32_f32_z(pg, svmla_f32_z(pg, voffseto, svadd_f32_z(pg, af_1, bf_1), invvscaleo));
+                const auto rf_2 = svcvt_s32_f32_z(pg, svmla_f32_z(pg, voffseto, svadd_f32_z(pg, af_2, bf_2), invvscaleo));
+                const auto rf_3 = svcvt_s32_f32_z(pg, svmla_f32_z(pg, voffseto, svadd_f32_z(pg, af_3, bf_3), invvscaleo));
+
+                const auto pa  = svqxtnt_s32(svqxtnb_s32(rf_0), rf_1);
+                const auto pb  = svqxtnt_s32(svqxtnb_s32(rf_2), rf_3);
+                const auto res = svqxtnt_s16(svqxtnb_s16(pa), pb);
+
+                svst1_s8(pg, output_ptr + x, res);
+
+                x += svcntb();
+                pg = svwhilelt_b8(x, window_end_x);
+            }
+            while(svptest_any(all_true_pg, pg));
+        },
+        broadcast_input, non_broadcast_input, output);
+    }
+    else
+    {
+        // Clear X Dimension on execution window as we handle manually
+        input1_win.set(Window::DimX, Window::Dimension(0, 1, 1));
+        input2_win.set(Window::DimX, Window::Dimension(0, 1, 1));
+
+        Iterator input1(in1, input1_win);
+        Iterator input2(in2, input2_win);
+        Iterator output(out, win);
+
+        const auto vscale1  = svdup_n_f32(iq1_info.scale);
+        const auto vscale2  = svdup_n_f32(iq2_info.scale);
+        const auto voffset1 = svdup_n_s32(iq1_info.offset);
+        const auto voffset2 = svdup_n_s32(iq2_info.offset);
+
+        execute_window_loop(win, [&](const Coordinates &)
+        {
+            const auto input1_ptr = reinterpret_cast<const int8_t *>(input1.ptr());
+            const auto input2_ptr = reinterpret_cast<const int8_t *>(input2.ptr());
+            const auto output_ptr = reinterpret_cast<int8_t *>(output.ptr());
+
+            int      x  = window_start_x;
+            svbool_t pg = svwhilelt_b8(x, window_end_x);
+            do
+            {
+                const auto a = svld1_s8(pg, input1_ptr + x);
+                const auto b = svld1_s8(pg, input2_ptr + x);
+
+                const auto af_0 = svmul_f32_z(pg, svcvt_f32_s32_z(pg, svsub_s32_z(pg, svmovlb_s32(svmovlb_s16(a)), voffset1)), vscale1);
+                const auto af_1 = svmul_f32_z(pg, svcvt_f32_s32_z(pg, svsub_s32_z(pg, svmovlt_s32(svmovlb_s16(a)), voffset1)), vscale1);
+                const auto af_2 = svmul_f32_z(pg, svcvt_f32_s32_z(pg, svsub_s32_z(pg, svmovlb_s32(svmovlt_s16(a)), voffset1)), vscale1);
+                const auto af_3 = svmul_f32_z(pg, svcvt_f32_s32_z(pg, svsub_s32_z(pg, svmovlt_s32(svmovlt_s16(a)), voffset1)), vscale1);
+
+                const auto bf_0 = svmul_f32_z(pg, svcvt_f32_s32_z(pg, svsub_s32_z(pg, svmovlb_s32(svmovlb_s16(b)), voffset2)), vscale2);
+                const auto bf_1 = svmul_f32_z(pg, svcvt_f32_s32_z(pg, svsub_s32_z(pg, svmovlt_s32(svmovlb_s16(b)), voffset2)), vscale2);
+                const auto bf_2 = svmul_f32_z(pg, svcvt_f32_s32_z(pg, svsub_s32_z(pg, svmovlb_s32(svmovlt_s16(b)), voffset2)), vscale2);
+                const auto bf_3 = svmul_f32_z(pg, svcvt_f32_s32_z(pg, svsub_s32_z(pg, svmovlt_s32(svmovlt_s16(b)), voffset2)), vscale2);
+
+                const auto rf_0 = svcvt_s32_f32_z(pg, svmla_f32_z(pg, voffseto, svadd_f32_z(pg, af_0, bf_0), invvscaleo));
+                const auto rf_1 = svcvt_s32_f32_z(pg, svmla_f32_z(pg, voffseto, svadd_f32_z(pg, af_1, bf_1), invvscaleo));
+                const auto rf_2 = svcvt_s32_f32_z(pg, svmla_f32_z(pg, voffseto, svadd_f32_z(pg, af_2, bf_2), invvscaleo));
+                const auto rf_3 = svcvt_s32_f32_z(pg, svmla_f32_z(pg, voffseto, svadd_f32_z(pg, af_3, bf_3), invvscaleo));
+
+                const auto pa  = svqxtnt_s32(svqxtnb_s32(rf_0), rf_1);
+                const auto pb  = svqxtnt_s32(svqxtnb_s32(rf_2), rf_3);
+                const auto res = svqxtnt_s16(svqxtnb_s16(pa), pb);
+
+                svst1_s8(pg, output_ptr + x, res);
+
+                x += svcntb();
+                pg = svwhilelt_b8(x, window_end_x);
+            }
+            while(svptest_any(svptrue_b8(), pg));
+        },
+        input1, input2, output);
+    }
+}
+} // namespace cpu
+} // namespace arm_compute
+#endif /* defined(__ARM_FEATURE_SVE2) */
\ No newline at end of file
diff --git a/src/core/NEON/kernels/arithmetic_addition/impl/SVE/qsymm16.cpp b/src/core/NEON/kernels/arithmetic_addition/impl/SVE/qsymm16.cpp
new file mode 100644
index 0000000000..c072cdb249
--- /dev/null
+++ b/src/core/NEON/kernels/arithmetic_addition/impl/SVE/qsymm16.cpp
@@ -0,0 +1,156 @@
+/*
+ * Copyright (c) 2020-2021 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.
+ */
+#include "arm_compute/core/Helpers.h"
+#include "arm_compute/core/ITensor.h"
+#include "arm_compute/core/Types.h"
+#include "arm_compute/core/utils/misc/Traits.h"
+#include "src/core/NEON/wrapper/intrinsics/intrinsics.h"
+#if defined(__ARM_FEATURE_SVE2)
+#include "src/core/NEON/SVEMath.h"
+#include <arm_sve.h>
+
+namespace arm_compute
+{
+namespace cpu
+{
+void arithmetic_addition_qsymm16_sve(const ITensor *in1, const ITensor *in2, ITensor *out, const ConvertPolicy &policy, const Window &window)
+{
+    ARM_COMPUTE_UNUSED(policy);
+
+    // Create input windows
+    Window input1_win = window.broadcast_if_dimension_le_one(in1->info()->tensor_shape());
+    Window input2_win = window.broadcast_if_dimension_le_one(in2->info()->tensor_shape());
+
+    // Clear X Dimension on execution window as we handle manually
+    Window win = window;
+    win.set(Window::DimX, Window::Dimension(0, 1, 1));
+
+    const auto window_start_x        = static_cast<int>(window.x().start());
+    const auto window_end_x          = static_cast<int>(window.x().end());
+    const bool is_broadcast_across_x = in1->info()->tensor_shape().x() != in2->info()->tensor_shape().x();
+
+    const UniformQuantizationInfo iq1_info = in1->info()->quantization_info().uniform();
+    const UniformQuantizationInfo iq2_info = in2->info()->quantization_info().uniform();
+    const UniformQuantizationInfo oq_info  = out->info()->quantization_info().uniform();
+
+    const auto vscale1     = svdup_n_f32(iq1_info.scale);
+    const auto vscale2     = svdup_n_f32(iq2_info.scale);
+    const auto invvscaleo  = svdup_n_f32(1.f / oq_info.scale);
+    const auto all_true_pg = svptrue_b16();
+
+    if(is_broadcast_across_x)
+    {
+        const bool     is_broadcast_input_2 = input2_win.x().step() == 0;
+        Window         broadcast_win        = is_broadcast_input_2 ? input2_win : input1_win;
+        Window         non_broadcast_win    = !is_broadcast_input_2 ? input2_win : input1_win;
+        const ITensor *broadcast_tensor     = is_broadcast_input_2 ? in2 : in1;
+        const ITensor *non_broadcast_tensor = !is_broadcast_input_2 ? in2 : in1;
+
+        // Clear X Dimension on execution window as we handle manually
+        non_broadcast_win.set(Window::DimX, Window::Dimension(0, 1, 1));
+
+        Iterator broadcast_input(broadcast_tensor, broadcast_win);
+        Iterator non_broadcast_input(non_broadcast_tensor, non_broadcast_win);
+        Iterator output(out, win);
+
+        execute_window_loop(win, [&](const Coordinates &)
+        {
+            const auto non_broadcast_input_ptr = reinterpret_cast<const int16_t *>(non_broadcast_input.ptr());
+            const auto output_ptr              = reinterpret_cast<int16_t *>(output.ptr());
+
+            const int16_t broadcast_value     = *reinterpret_cast<const int16_t *>(broadcast_input.ptr());
+            const auto    broadcast_value_vec = svdup_n_s16(broadcast_value);
+
+            int      x  = window_start_x;
+            svbool_t pg = svwhilelt_b16(x, window_end_x);
+
+            const auto bf_0 = svmul_f32_z(pg, svcvt_f32_s32_z(pg, svmovlb_s32(broadcast_value_vec)), vscale2);
+            const auto bf_1 = svmul_f32_z(pg, svcvt_f32_s32_z(pg, svmovlt_s32(broadcast_value_vec)), vscale2);
+
+            do
+            {
+                const auto a    = svld1_s16(pg, non_broadcast_input_ptr + x);
+                const auto af_0 = svmul_f32_z(pg, svcvt_f32_s32_z(pg, svmovlb_s32(a)), vscale1);
+                const auto af_1 = svmul_f32_z(pg, svcvt_f32_s32_z(pg, svmovlt_s32(a)), vscale1);
+
+                const auto rf_0 = svcvt_s32_f32_z(pg, svmul_f32_z(pg, svadd_f32_z(pg, af_0, bf_0), invvscaleo));
+                const auto rf_1 = svcvt_s32_f32_z(pg, svmul_f32_z(pg, svadd_f32_z(pg, af_1, bf_1), invvscaleo));
+
+                const auto res = svqxtnt_s32(svqxtnb_s32(rf_0), rf_1);
+
+                svst1_s16(pg, output_ptr + x, res);
+
+                x += svcnth();
+                pg = svwhilelt_b16(x, window_end_x);
+            }
+            while(svptest_any(all_true_pg, pg));
+        },
+        broadcast_input, non_broadcast_input, output);
+    }
+    else
+    {
+        // Clear X Dimension on execution window as we handle manually
+        input1_win.set(Window::DimX, Window::Dimension(0, 1, 1));
+        input2_win.set(Window::DimX, Window::Dimension(0, 1, 1));
+
+        Iterator input1(in1, input1_win);
+        Iterator input2(in2, input2_win);
+        Iterator output(out, win);
+
+        execute_window_loop(win, [&](const Coordinates &)
+        {
+            const auto input1_ptr = reinterpret_cast<const int16_t *>(input1.ptr());
+            const auto input2_ptr = reinterpret_cast<const int16_t *>(input2.ptr());
+            const auto output_ptr = reinterpret_cast<int16_t *>(output.ptr());
+
+            int      x  = window_start_x;
+            svbool_t pg = svwhilelt_b16(x, window_end_x);
+            do
+            {
+                auto a = svld1_s16(pg, input1_ptr + x);
+                auto b = svld1_s16(pg, input2_ptr + x);
+
+                const auto af_0 = svmul_f32_z(pg, svcvt_f32_s32_z(pg, svmovlb_s32(a)), vscale1);
+                const auto af_1 = svmul_f32_z(pg, svcvt_f32_s32_z(pg, svmovlt_s32(a)), vscale1);
+
+                const auto bf_0 = svmul_f32_z(pg, svcvt_f32_s32_z(pg, svmovlb_s32(b)), vscale2);
+                const auto bf_1 = svmul_f32_z(pg, svcvt_f32_s32_z(pg, svmovlt_s32(b)), vscale2);
+
+                const auto rf_0 = svcvt_s32_f32_z(pg, svmul_f32_z(pg, svadd_f32_z(pg, af_0, bf_0), invvscaleo));
+                const auto rf_1 = svcvt_s32_f32_z(pg, svmul_f32_z(pg, svadd_f32_z(pg, af_1, bf_1), invvscaleo));
+
+                const auto res = svqxtnt_s32(svqxtnb_s32(rf_0), rf_1);
+                svst1_s16(pg, output_ptr + x, res);
+
+                x += svcnth();
+                pg = svwhilelt_b16(x, window_end_x);
+            }
+            while(svptest_any(all_true_pg, pg));
+        },
+        input1, input2, output);
+    }
+}
+} // namespace cpu
+} // namespace arm_compute
+#endif /* defined(__ARM_FEATURE_SVE2) */
\ No newline at end of file
diff --git a/src/core/NEON/wrapper/intrinsics/intrinsics.h b/src/core/NEON/wrapper/intrinsics/intrinsics.h
index 6cf7f9d287..4c7b674e2e 100644
--- a/src/core/NEON/wrapper/intrinsics/intrinsics.h
+++ b/src/core/NEON/wrapper/intrinsics/intrinsics.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2018-2020 Arm Limited.
+ * Copyright (c) 2018-2021 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -80,6 +80,7 @@
 #include "src/core/NEON/wrapper/intrinsics/svexp.h"
 #include "src/core/NEON/wrapper/intrinsics/svlog.h"
 #include "src/core/NEON/wrapper/intrinsics/svptrue.h"
+#include "src/core/NEON/wrapper/intrinsics/svqadd.h"
 #include "src/core/NEON/wrapper/intrinsics/svsin.h"
 #include "src/core/NEON/wrapper/intrinsics/svwhilelt.h"
 #endif /* defined(__ARM_FEATURE_SVE) */
diff --git a/src/core/NEON/wrapper/intrinsics/svqadd.h b/src/core/NEON/wrapper/intrinsics/svqadd.h
new file mode 100644
index 0000000000..fd45d82104
--- /dev/null
+++ b/src/core/NEON/wrapper/intrinsics/svqadd.h
@@ -0,0 +1,60 @@
+/*
+ * Copyright (c) 2020-2021 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.
+ */
+#ifndef SRC_CORE_NEON_WRAPPER_INTRINSICS_SVQADD_H
+#define SRC_CORE_NEON_WRAPPER_INTRINSICS_SVQADD_H
+#if defined(__ARM_FEATURE_SVE)
+#include <arm_sve.h>
+namespace arm_compute
+{
+namespace wrapper
+{
+#define SVQADD_IMPL_F(type, postfix, svppostfix)                        \
+    inline type svqadd(const type &val1, const type &val2)              \
+    {                                                                   \
+        return svadd_##postfix##_z(svptrue_##svppostfix(), val1, val2); \
+    }
+
+SVQADD_IMPL_F(svfloat32_t, f32, b32)
+SVQADD_IMPL_F(svfloat16_t, f16, b16)
+#undef SVQADD_IMPL_F
+
+#define SVQADD_IMPL(type, postfix)                         \
+    inline type svqadd(const type &val1, const type &val2) \
+    {                                                      \
+        return svqadd_##postfix(val1, val2);               \
+    }
+
+SVQADD_IMPL(svint32_t, s32)
+SVQADD_IMPL(svint16_t, s16)
+SVQADD_IMPL(svint8_t, s8)
+SVQADD_IMPL(svuint32_t, u32)
+SVQADD_IMPL(svuint16_t, u16)
+SVQADD_IMPL(svuint8_t, u8)
+
+#undef SVQADD_IMPL
+} // namespace wrapper
+} // namespace arm_compute
+
+#endif /* defined(__ARM_FEATURE_SVE) */
+#endif /* SRC_CORE_NEON_WRAPPER_INTRINSICS_SVQADD_H */
\ No newline at end of file
diff --git a/src/core/common/Registrars.h b/src/core/common/Registrars.h
index 649fe468a3..112c83ad94 100644
--- a/src/core/common/Registrars.h
+++ b/src/core/common/Registrars.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2020 Arm Limited.
+ * Copyright (c) 2020-2021 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -83,4 +83,14 @@
 #define REGISTER_QSYMM16_SVE(func_name) nullptr
 #endif /* defined(ENABLE_QSYMM16_KERNELS) */
 
+#if defined(ENABLE_INTEGER_KERNELS)
+#if defined(__ARM_FEATURE_SVE)
+#define REGISTER_INTEGER_SVE(func_name) &(func_name)
+#endif /* defined(__ARM_FEATURE_SVE) */
+#define REGISTER_INTEGER_NEON(func_name) &(func_name)
+#else /* defined(ENABLE_INTEGER_KERNELS) */
+#define REGISTER_INTEGER_NEON(func_name) nullptr
+#define REGISTER_INTEGER_SVE(func_name) nullptr
+#endif /* defined(ENABLE_INTEGER_KERNELS) */
+
 #endif /* SRC_CORE_COMMON_REGISTRARS_H */
diff --git a/tests/validation/Helpers.h b/tests/validation/Helpers.h
index 325cc0042e..604840b33e 100644
--- a/tests/validation/Helpers.h
+++ b/tests/validation/Helpers.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017-2020 Arm Limited.
+ * Copyright (c) 2017-2021 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -30,6 +30,7 @@
 #include "tests/Globals.h"
 #include "tests/SimpleTensor.h"
 
+#include <math.h>
 #include <random>
 #include <type_traits>
 #include <utility>
diff --git a/tests/validation/NEON/ArithmeticAddition.cpp b/tests/validation/NEON/ArithmeticAddition.cpp
index 7b3d4f9ac0..5598a1106b 100644
--- a/tests/validation/NEON/ArithmeticAddition.cpp
+++ b/tests/validation/NEON/ArithmeticAddition.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017-2020 Arm Limited.
+ * Copyright (c) 2017-2021 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -43,9 +43,11 @@ namespace validation
 {
 namespace
 {
-#ifndef __aarch64__
+#if !defined(__aarch64__) || defined(__ARM_FEATURE_SVE)
 constexpr AbsoluteTolerance<float> tolerance_quant(1); /**< Tolerance value for comparing reference's output against implementation's output for quantized data types */
-#endif                                                 //__aarch64__
+#else                                                  // !defined(__aarch64__) || defined(__ARM_FEATURE_SVE)
+constexpr AbsoluteTolerance<float> tolerance_quant(0);
+#endif                                                 // !defined(__aarch64__) || defined(__ARM_FEATURE_SVE)
 
 /** Input data sets **/
 const auto ArithmeticAdditionU8Dataset = combine(combine(framework::dataset::make("DataType", DataType::U8), framework::dataset::make("DataType", DataType::U8)), framework::dataset::make("DataType",
@@ -225,11 +227,7 @@ FIXTURE_DATA_TEST_CASE(RunSmall,
                                framework::dataset::make("OutQInfo", { QuantizationInfo(1.f / 255.f, 5) })))
 {
     // Validate output
-#ifdef __aarch64__
-    validate(Accessor(_target), _reference);
-#else  //__aarch64__
     validate(Accessor(_target), _reference, tolerance_quant);
-#endif //__aarch64__
 }
 TEST_SUITE_END() // QASYMM8
 
@@ -244,11 +242,7 @@ FIXTURE_DATA_TEST_CASE(RunSmall,
                                framework::dataset::make("OutQInfo", { QuantizationInfo(0.5f, 5) })))
 {
     // Validate output
-#ifdef __aarch64__
-    validate(Accessor(_target), _reference);
-#else  //__aarch64__
     validate(Accessor(_target), _reference, tolerance_quant);
-#endif //__aarch64__
 }
 
 FIXTURE_DATA_TEST_CASE(RunSmallBroadcast, NEArithmeticAdditionQuantizedBroadcastFixture<int8_t>, framework::DatasetMode::ALL, combine(combine(combine(combine(combine(
@@ -259,11 +253,7 @@ FIXTURE_DATA_TEST_CASE(RunSmallBroadcast, NEArithmeticAdditionQuantizedBroadcast
                        framework::dataset::make("OutQInfo", { QuantizationInfo(0.5f, 5) })))
 {
     // Validate output
-#ifdef __aarch64__
-    validate(Accessor(_target), _reference);
-#else  //__aarch64__
     validate(Accessor(_target), _reference, tolerance_quant);
-#endif //__aarch64__
 }
 TEST_SUITE_END() // QASYMM8_SIGNED
 
@@ -278,11 +268,7 @@ FIXTURE_DATA_TEST_CASE(RunSmall,
                                framework::dataset::make("OutQInfo", { QuantizationInfo(5.f / 32768.f, 0) })))
 {
     // Validate output
-#ifdef __aarch64__
-    validate(Accessor(_target), _reference);
-#else  //__aarch64__
     validate(Accessor(_target), _reference, tolerance_quant);
-#endif //__aarch64__
 }
 TEST_SUITE_END() // QSYMM16
 TEST_SUITE_END() // Quantized