Make Add kernel and operator stateless

- Rename NEArithmeticAdditionKernel to CpuAddKernel Cpu and move files appropriately

- Add CpuAdd under src/runtime/cpu/operators

Partially resolves: COMPMID-4005

Signed-off-by: Sheri Zhang <sheri.zhang@arm.com>
Change-Id: I1d8d406df9773fea198899f50327407d7125c38d
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/4867
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>

12 files changed, 2204 insertions, 0 deletions

diff --git a/src/core/cpu/kernels/CpuAddKernel.cpp b/src/core/cpu/kernels/CpuAddKernel.cpp
new file mode 100644
index 0000000000..31c7b2af60
--- /dev/null
+++ b/src/core/cpu/kernels/CpuAddKernel.cpp
@@ -0,0 +1,347 @@
+/*
+ * Copyright (c) 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 "src/core/cpu/kernels/CpuAddKernel.h"
+
+#include "arm_compute/core/ITensor.h"
+#include "arm_compute/core/TensorInfo.h"
+#include "arm_compute/core/Validate.h"
+#include "src/core/CPP/Validate.h"
+#include "src/core/common/Registrars.h"
+#include "src/core/cpu/kernels/add/neon/list.h"
+#include "src/core/cpu/kernels/add/sve/list.h"
+#include "src/core/helpers/AutoConfiguration.h"
+#include "src/core/helpers/WindowHelpers.h"
+
+#include <array>
+
+namespace arm_compute
+{
+namespace cpu
+{
+namespace kernels
+{
+namespace
+{
+struct AddSelectorData
+{
+    DataType dt1;
+    DataType dt2;
+    DataType dt3;
+};
+
+using AddSelectorPtr = std::add_pointer<bool(const AddSelectorData &data)>::type;
+using AddKernelPtr   = std::add_pointer<void(const ITensor *, const ITensor *, ITensor *, const ConvertPolicy &, const Window &)>::type;
+struct AddKernel
+{
+    const char          *name;
+    const AddSelectorPtr is_selected;
+    AddKernelPtr         ukernel;
+};
+
+static const AddKernel available_kernels[] =
+{
+#if defined(__ARM_FEATURE_SVE)
+    {
+        "add_same_sve",
+        [](const AddSelectorData & data) { return ((data.dt1 == data.dt2) && (data.dt1 == DataType::F32)); },
+        REGISTER_FP32_SVE(arm_compute::cpu::add_same_sve<float>)
+    },
+    {
+        "add_same_sve",
+        [](const AddSelectorData & data) { return ((data.dt1 == data.dt2) && (data.dt1 == DataType::F16)); },
+        REGISTER_FP16_SVE(arm_compute::cpu::add_same_sve<float16_t>)
+    },
+    {
+        "add_same_sve",
+        [](const AddSelectorData & data) { return ((data.dt1 == data.dt2) && (data.dt1 == data.dt3) && (data.dt1 == DataType::U8)); },
+        REGISTER_INTEGER_SVE(arm_compute::cpu::add_same_sve<uint8_t>)
+    },
+    {
+        "add_same_sve",
+        [](const AddSelectorData & data) { return ((data.dt1 == data.dt2) && (data.dt1 == data.dt3) && (data.dt1 == DataType::S16)); },
+        REGISTER_INTEGER_SVE(arm_compute::cpu::add_same_sve<int16_t>)
+    },
+    {
+        "add_same_sve",
+        [](const AddSelectorData & data) { return ((data.dt1 == data.dt2) && (data.dt1 == data.dt3) && (data.dt1 == DataType::S32)); },
+        REGISTER_INTEGER_SVE(arm_compute::cpu::add_same_sve<int32_t>)
+    },
+    {
+        "add_u8_s16_s16_sve",
+        [](const AddSelectorData & data) { return ((data.dt1 == DataType::U8) && (data.dt2 == DataType::S16)); },
+        REGISTER_INTEGER_SVE(arm_compute::cpu::add_u8_s16_s16_sve)
+    },
+    {
+        "add_s16_u8_s16_sve",
+        [](const AddSelectorData & data) { return ((data.dt1 == DataType::S16) && (data.dt2 == DataType::U8)); },
+        REGISTER_INTEGER_SVE(arm_compute::cpu::add_s16_u8_s16_sve)
+    },
+    {
+        "add_u8_u8_s16_sve",
+        [](const AddSelectorData & data) { return ((data.dt1 == data.dt2) && (data.dt3 == DataType::S16)); },
+        REGISTER_INTEGER_SVE(arm_compute::cpu::add_u8_u8_s16_sve)
+    },
+#else /* !defined(__ARM_FEATURE_SVE) */
+    {
+        "add_same_neon",
+        [](const AddSelectorData & data) { return ((data.dt1 == data.dt2) && (data.dt1 == DataType::F32)); },
+        REGISTER_FP32_NEON(arm_compute::cpu::add_same_neon<float>)
+    },
+#if defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC)
+    {
+        "add_same_neon",
+        [](const AddSelectorData & data) { return ((data.dt1 == data.dt2) && (data.dt1 == DataType::F16)); },
+        REGISTER_FP16_NEON(arm_compute::cpu::add_same_neon<float16_t>)
+    },
+#endif /* defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) */
+    {
+        "add_same_neon",
+        [](const AddSelectorData & data) { return ((data.dt1 == data.dt2) && (data.dt1 == data.dt3) && (data.dt1 == DataType::U8)); },
+        REGISTER_INTEGER_NEON(arm_compute::cpu::add_same_neon<uint8_t>)
+    },
+    {
+        "add_same_neon",
+        [](const AddSelectorData & data) { return ((data.dt1 == data.dt2) && (data.dt1 == data.dt3) && (data.dt1 == DataType::S16)); },
+        REGISTER_INTEGER_NEON(arm_compute::cpu::add_same_neon<int16_t>)
+    },
+    {
+        "add_same_neon",
+        [](const AddSelectorData & data) { return ((data.dt1 == data.dt2) && (data.dt1 == data.dt3) && (data.dt1 == DataType::S32)); },
+        REGISTER_INTEGER_NEON(arm_compute::cpu::add_same_neon<int32_t>)
+    },
+    {
+        "add_u8_s16_s16_neon",
+        [](const AddSelectorData & data) { return ((data.dt1 == DataType::U8) && (data.dt2 == DataType::S16)); },
+        REGISTER_INTEGER_NEON(arm_compute::cpu::add_u8_s16_s16_neon)
+    },
+    {
+        "add_s16_u8_s16_neon",
+        [](const AddSelectorData & data) { return ((data.dt1 == DataType::S16) && (data.dt2 == DataType::U8)); },
+        REGISTER_INTEGER_NEON(arm_compute::cpu::add_s16_u8_s16_neon)
+    },
+    {
+        "add_u8_u8_s16_neon",
+        [](const AddSelectorData & data) { return ((data.dt1 == data.dt2) && (data.dt3 == DataType::S16)); },
+        REGISTER_INTEGER_NEON(arm_compute::cpu::add_u8_u8_s16_neon)
+    },
+#endif /* defined(__ARM_FEATURE_SVE) */
+
+#if defined(__ARM_FEATURE_SVE2)
+    {
+        "add_qasymm8_sve",
+        [](const AddSelectorData & data) { return ((data.dt1 == data.dt2) && (data.dt1 == DataType::QASYMM8)); },
+        REGISTER_QASYMM8_SVE(arm_compute::cpu::add_qasymm8_sve)
+    },
+    {
+        "add_qasymm8_signed_sve",
+        [](const AddSelectorData & data) { return ((data.dt1 == data.dt2) && (data.dt1 == DataType::QASYMM8_SIGNED)); },
+        REGISTER_QASYMM8_SIGNED_SVE(arm_compute::cpu::add_qasymm8_signed_sve)
+    },
+    {
+        "add_qsymm16_sve",
+        [](const AddSelectorData & data) { return ((data.dt1 == data.dt2) && (data.dt1 == DataType::QSYMM16)); },
+        REGISTER_QSYMM16_SVE(arm_compute::cpu::add_qsymm16_sve)
+    },
+#else  /* !defined(__ARM_FEATURE_SVE2) */
+    {
+        "add_qasymm8_neon",
+        [](const AddSelectorData & data) { return ((data.dt1 == data.dt2) && (data.dt1 == DataType::QASYMM8)); },
+        REGISTER_QASYMM8_NEON(arm_compute::cpu::add_qasymm8_neon)
+    },
+    {
+        "add_qasymm8_signed_neon",
+        [](const AddSelectorData & data) { return ((data.dt1 == data.dt2) && (data.dt1 == DataType::QASYMM8_SIGNED)); },
+        REGISTER_QASYMM8_SIGNED_NEON(arm_compute::cpu::add_qasymm8_signed_neon)
+    },
+    {
+        "add_qsymm16_neon",
+        [](const AddSelectorData & data) { return ((data.dt1 == data.dt2) && (data.dt1 == DataType::QSYMM16)); },
+        REGISTER_QSYMM16_NEON(arm_compute::cpu::add_qsymm16_neon)
+    },
+#endif /* defined(__ARM_FEATURE_SVE2) */
+
+};
+
+/** Micro-kernel selector
+ *
+ * @param[in] data Selection data passed to help pick the appropriate micro-kernel
+ *
+ * @return A matching micro-kernel else nullptr
+ */
+const AddKernel *get_implementation(DataType dt1, DataType dt2, DataType dt3)
+{
+    for(const auto &uk : available_kernels)
+    {
+        if(uk.is_selected({ dt1, dt2, dt3 }))
+        {
+            return &uk;
+        }
+    }
+    return nullptr;
+}
+
+Status validate_arguments(const ITensorInfo &src0, const ITensorInfo &src1, const ITensorInfo &dst, ConvertPolicy policy)
+{
+    ARM_COMPUTE_UNUSED(policy);
+
+    ARM_COMPUTE_RETURN_ERROR_ON_CPU_F16_UNSUPPORTED(&src0);
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(&src0, 1, DataType::U8, DataType::QASYMM8, DataType::QASYMM8_SIGNED,
+                                                         DataType::S16, DataType::QSYMM16, DataType::F16,
+                                                         DataType::S32, DataType::F32);
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(&src1, 1, DataType::U8, DataType::QASYMM8, DataType::QASYMM8_SIGNED,
+                                                         DataType::S16, DataType::QSYMM16, DataType::F16,
+                                                         DataType::S32, DataType::F32);
+
+    const TensorShape out_shape = TensorShape::broadcast_shape(src0.tensor_shape(), src1.tensor_shape());
+
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG(out_shape.total_size() == 0, "Inputs are not broadcast compatible");
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG((src0.tensor_shape().x() != src1.tensor_shape().x()) && ((src0.data_type() != src1.data_type()) || (src0.data_type() != dst.data_type())
+                                                                                             || (src1.data_type() != dst.data_type())),
+                                    "Broadcasting across width is supported on configurations where all tensors have the same data type");
+
+    // Validate in case of configured dst
+    if(dst.total_size() > 0)
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON_MSG(
+            !(src0.data_type() == DataType::U8 && src1.data_type() == DataType::U8 && dst.data_type() == DataType::U8)
+            && !(src0.data_type() == DataType::U8 && src1.data_type() == DataType::U8 && dst.data_type() == DataType::S16)
+            && !(src0.data_type() == DataType::U8 && src1.data_type() == DataType::S16 && dst.data_type() == DataType::S16)
+            && !(src0.data_type() == DataType::S16 && src1.data_type() == DataType::U8 && dst.data_type() == DataType::S16)
+            && !(src0.data_type() == DataType::S16 && src1.data_type() == DataType::S16 && dst.data_type() == DataType::S16)
+            && !(src0.data_type() == DataType::S32 && src1.data_type() == DataType::S32 && dst.data_type() == DataType::S32)
+            && !(src0.data_type() == DataType::F32 && src1.data_type() == DataType::F32 && dst.data_type() == DataType::F32)
+            && !(src0.data_type() == DataType::F16 && src1.data_type() == DataType::F16 && dst.data_type() == DataType::F16)
+            && !(src0.data_type() == DataType::QASYMM8 && src1.data_type() == DataType::QASYMM8 && dst.data_type() == DataType::QASYMM8)
+            && !(src0.data_type() == DataType::QASYMM8_SIGNED && src1.data_type() == DataType::QASYMM8_SIGNED && dst.data_type() == DataType::QASYMM8_SIGNED)
+            && !(src0.data_type() == DataType::QSYMM16 && src1.data_type() == DataType::QSYMM16 && dst.data_type() == DataType::QSYMM16),
+            "You called addition with the wrong image formats");
+
+        ARM_COMPUTE_RETURN_ERROR_ON_MSG(detail::have_different_dimensions(out_shape, dst.tensor_shape(), 0),
+                                        "Wrong shape for dst");
+    }
+
+    const auto *uk = get_implementation(src0.data_type(), src1.data_type(), dst.data_type());
+    ARM_COMPUTE_RETURN_ERROR_ON(uk == nullptr || uk->ukernel == nullptr);
+
+    return Status{};
+}
+
+std::pair<Status, Window> validate_and_configure_window(const ITensorInfo &src0, const ITensorInfo &src1, ITensorInfo &dst)
+{
+    const std::pair<TensorShape, ValidRegion> broadcast_pair = ITensorInfo::broadcast_shape_and_valid_region(src0, src1);
+    const TensorShape &out_shape    = broadcast_pair.first;
+    const ValidRegion &valid_region = broadcast_pair.second;
+
+    // Auto initialize dst if not initialized
+    {
+        set_shape_if_empty(dst, out_shape);
+
+        if(src0.data_type() == DataType::S16 || src1.data_type() == DataType::S16)
+        {
+            set_format_if_unknown(dst, Format::S16);
+        }
+        if(src0.data_type() == DataType::S32 || src1.data_type() == DataType::S32)
+        {
+            set_format_if_unknown(dst, Format::S32);
+        }
+        else if(src0.data_type() == DataType::F16 || src1.data_type() == DataType::F16)
+        {
+            set_format_if_unknown(dst, Format::F16);
+        }
+        else if(src0.data_type() == DataType::F32 || src1.data_type() == DataType::F32)
+        {
+            set_format_if_unknown(dst, Format::F32);
+        }
+        else if(src0.data_type() == DataType::QASYMM8 || src1.data_type() == DataType::QASYMM8)
+        {
+            set_data_type_if_unknown(dst, DataType::QASYMM8);
+        }
+        else if(src0.data_type() == DataType::QASYMM8_SIGNED || src1.data_type() == DataType::QASYMM8_SIGNED)
+        {
+            set_data_type_if_unknown(dst, DataType::QASYMM8_SIGNED);
+        }
+        else if(src0.data_type() == DataType::QSYMM16 || src1.data_type() == DataType::QSYMM16)
+        {
+            set_data_type_if_unknown(dst, DataType::QSYMM16);
+        }
+    }
+
+    Window win = calculate_max_window(valid_region, Steps());
+
+    // CpuAddKernel doesn't need padding so update_window_and_padding() can be skipped
+    Coordinates coord;
+    coord.set_num_dimensions(dst.num_dimensions());
+    dst.set_valid_region(valid_region);
+    return std::make_pair(Status{}, win);
+}
+} // namespace
+
+void CpuAddKernel::configure(const ITensorInfo *src0, const ITensorInfo *src1, ITensorInfo *dst, ConvertPolicy policy)
+{
+    ARM_COMPUTE_ERROR_ON_NULLPTR(src0, src1, dst);
+    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(*src0, *src1, *dst, policy));
+
+    _policy = policy;
+
+    // Configure kernel window
+    auto win_config = validate_and_configure_window(*src0, *src1, *dst);
+    ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
+    ICpuKernel::configure(win_config.second);
+}
+
+Status CpuAddKernel::validate(const ITensorInfo *src0, const ITensorInfo *src1, const ITensorInfo *dst, ConvertPolicy policy)
+{
+    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src0, src1, dst);
+
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(*src0, *src1, *dst, policy));
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(*src0->clone(), *src1->clone(), *dst->clone()).first);
+
+    return Status{};
+}
+
+void CpuAddKernel::run_op(ITensorPack &tensors, const Window &window, const ThreadInfo &info)
+{
+    ARM_COMPUTE_UNUSED(info);
+    ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
+    ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICpuKernel::window(), window);
+
+    ARM_COMPUTE_ERROR_ON(tensors.empty());
+
+    const ITensor *src0 = tensors.get_const_tensor(TensorType::ACL_SRC_0);
+    const ITensor *src1 = tensors.get_const_tensor(TensorType::ACL_SRC_1);
+    ITensor       *dst  = tensors.get_tensor(TensorType::ACL_DST);
+
+    const auto *uk = get_implementation(src0->info()->data_type(), src1->info()->data_type(), dst->info()->data_type());
+    ARM_COMPUTE_ERROR_ON(uk == nullptr || uk->ukernel == nullptr);
+
+    uk->ukernel(src0, src1, dst, _policy, window);
+}
+
+const char *CpuAddKernel::name() const
+{
+    return "CpuAddKernel";
+}
+} // namespace kernels
+} // namespace cpu
+} // namespace arm_compute
diff --git a/src/core/cpu/kernels/CpuAddKernel.h b/src/core/cpu/kernels/CpuAddKernel.h
new file mode 100644
index 0000000000..a36ec7ad65
--- /dev/null
+++ b/src/core/cpu/kernels/CpuAddKernel.h
@@ -0,0 +1,85 @@
+/*
+ * Copyright (c) 2016-2021 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 ARM_COMPUTE_CPUADDKERNEL_H
+#define ARM_COMPUTE_CPUADDKERNEL_H
+
+#include "src/core/common/Macros.h"
+#include "src/core/cpu/ICpuKernel.h"
+
+namespace arm_compute
+{
+namespace cpu
+{
+namespace kernels
+{
+/** Interface for the kernel to perform addition between two tensors */
+class CpuAddKernel : public ICpuKernel
+{
+public:
+    CpuAddKernel() = default;
+    ARM_COMPUTE_DISALLOW_COPY_ALLOW_MOVE(CpuAddKernel);
+    /** Initialise the kernel's input, dst and border mode.
+     *
+     * Valid configurations (src0,src1) -> dst :
+     *
+     *   - (U8,U8)           -> U8
+     *   - (U8,U8)           -> S16
+     *   - (S16,U8)          -> S16
+     *   - (U8,S16)          -> S16
+     *   - (S16,S16)         -> S16
+     *   - (S32,S32)         -> S32
+     *   - (F16,F16)         -> F16
+     *   - (F32,F32)         -> F32
+     *   - (QASYMM8,QASYMM8) -> QASYMM8
+     *   - (QASYMM8_SIGNED,QASYMM8_SIGNED) -> QASYMM8_SIGNED
+     *   - (QSYMM16,QSYMM16) -> QSYMM16
+     *
+     * @param[in]  src0   First input tensor info. Data types supported: U8/QASYMM8/QASYMM8_SIGNED/S16/QSYMM16/F16/S32/F32
+     * @param[in]  src1   Second input tensor info. Data types supported: U8/QASYMM8/QASYMM8_SIGNED/S16/QSYMM16/F16/S32/F32
+     * @param[out] dst    The dst tensor info. Data types supported: U8/QASYMM8/QASYMM8_SIGNED/S16/QSYMM16/F16/S32/F32.
+     * @param[in]  policy Overflow policy.
+     */
+    void configure(const ITensorInfo *src0, const ITensorInfo *src1, ITensorInfo *dst, ConvertPolicy policy);
+    /** Static function to check if given info will lead to a valid configuration of @ref CpuAddKernel
+     *
+     * @param[in] src0   First input tensor info. Data types supported: U8/QASYMM8/QASYMM8_SIGNED/S16/QSYMM16/F16/S32/F32
+     * @param[in] src1   Second input tensor info. Data types supported: U8/QASYMM8/QASYMM8_SIGNED/S16/QSYMM16/F16/S32/F32
+     * @param[in] dst    The dst tensor info. Data types supported: U8/QASYMM8/QASYMM8_SIGNED/S16/QSYMM16/F16/S32/F32.
+     * @param[in] policy Overflow policy.
+     *
+     * @return a status
+     */
+    static Status validate(const ITensorInfo *src0, const ITensorInfo *src1, const ITensorInfo *dst, ConvertPolicy policy);
+
+    // Inherited methods overridden:
+    void run_op(ITensorPack &tensors, const Window &window, const ThreadInfo &info) override;
+    const char *name() const override;
+
+private:
+    ConvertPolicy _policy{};
+};
+} // namespace kernels
+} // namespace cpu
+} // namespace arm_compute
+#endif /*ARM_COMPUTE_CPUADDKERNEL_H */
diff --git a/src/core/cpu/kernels/add/neon/integer.cpp b/src/core/cpu/kernels/add/neon/integer.cpp
new file mode 100644
index 0000000000..24a0ac3b7c
--- /dev/null
+++ b/src/core/cpu/kernels/add/neon/integer.cpp
@@ -0,0 +1,170 @@
+/*
+ * 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/helpers/WindowHelpers.h"
+
+namespace arm_compute
+{
+namespace cpu
+{
+void add_u8_u8_s16_neon(const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window)
+{
+    // Create input windows
+    Window win        = window;
+    Window input1_win = window.broadcast_if_dimension_le_one(src0->info()->tensor_shape());
+    Window input2_win = window.broadcast_if_dimension_le_one(src1->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(src0, input1_win);
+    Iterator input2(src1, input2_win);
+    Iterator output(dst, 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 add_s16_u8_s16_neon(const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window)
+{
+    // Create input windows
+    Window win        = window;
+    Window input1_win = window.broadcast_if_dimension_le_one(src0->info()->tensor_shape());
+    Window input2_win = window.broadcast_if_dimension_le_one(src1->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(src0, input1_win);
+    Iterator input2(src1, input2_win);
+    Iterator output(dst, 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 add_u8_s16_s16_neon(const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window)
+{
+    // Simply swap the two input buffers:
+    add_s16_u8_s16_neon(src1, src0, dst, policy, window);
+}
+} // namespace cpu
+} // namespace arm_compute
\ No newline at end of file
diff --git a/src/core/cpu/kernels/add/neon/list.h b/src/core/cpu/kernels/add/neon/list.h
new file mode 100644
index 0000000000..53ea81e284
--- /dev/null
+++ b/src/core/cpu/kernels/add/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_ADD_LIST_H
+#define SRC_CORE_NEON_KERNELS_ADD_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_ADD_KERNEL(func_name) \
+    void func_name(const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window)
+
+DECLARE_ADD_KERNEL(add_qasymm8_neon);
+DECLARE_ADD_KERNEL(add_qasymm8_signed_neon);
+DECLARE_ADD_KERNEL(add_qsymm16_neon);
+DECLARE_ADD_KERNEL(add_s16_u8_s16_neon);
+DECLARE_ADD_KERNEL(add_u8_s16_s16_neon);
+DECLARE_ADD_KERNEL(add_u8_u8_s16_neon);
+
+#undef DECLARE_ADD_KERNEL
+
+template <typename ScalarType>
+void add_same_neon(const ITensor *src0, const ITensor *src1, ITensor *dst, 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(src0->info()->tensor_shape());
+    Window input2_win = window.broadcast_if_dimension_le_one(src1->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 = src0->info()->tensor_shape().x() != src1->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 ? src1 : src0;
+        const ITensor *non_broadcast_tensor = !is_broadcast_input_2 ? src1 : src0;
+
+        // 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(dst, 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(src0, input1_win);
+        Iterator input2(src1, input2_win);
+        Iterator output(dst, 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_ADD_LIST_H
\ No newline at end of file
diff --git a/src/core/cpu/kernels/add/neon/qasymm8.cpp b/src/core/cpu/kernels/add/neon/qasymm8.cpp
new file mode 100644
index 0000000000..cc97f0067c
--- /dev/null
+++ b/src/core/cpu/kernels/add/neon/qasymm8.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/helpers/WindowHelpers.h"
+
+namespace arm_compute
+{
+namespace cpu
+{
+void add_qasymm8_neon(const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window)
+{
+    ARM_COMPUTE_UNUSED(policy);
+
+    // Create input windows
+    Window input1_win = window.broadcast_if_dimension_le_one(src0->info()->tensor_shape());
+    Window input2_win = window.broadcast_if_dimension_le_one(src1->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 = src0->info()->tensor_shape().x() != src1->info()->tensor_shape().x();
+
+    const UniformQuantizationInfo iq1_info = src0->info()->quantization_info().uniform();
+    const UniformQuantizationInfo iq2_info = src1->info()->quantization_info().uniform();
+    const UniformQuantizationInfo oq_info  = dst->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 ? src1 : src0;
+        const ITensor                *non_broadcast_tensor = !is_broadcast_input_2 ? src1 : src0;
+        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(dst, 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(src0, input1_win);
+        Iterator input2(src1, input2_win);
+        Iterator output(dst, 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), dst->info()->quantization_info());
+            }
+        },
+        input1, input2, output);
+    }
+}
+} // namespace cpu
+} // namespace arm_compute
\ No newline at end of file
diff --git a/src/core/cpu/kernels/add/neon/qasymm8_signed.cpp b/src/core/cpu/kernels/add/neon/qasymm8_signed.cpp
new file mode 100644
index 0000000000..d62d0739f5
--- /dev/null
+++ b/src/core/cpu/kernels/add/neon/qasymm8_signed.cpp
@@ -0,0 +1,208 @@
+/*
+ * 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/helpers/WindowHelpers.h"
+
+namespace arm_compute
+{
+namespace cpu
+{
+void add_qasymm8_signed_neon(const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window)
+{
+    ARM_COMPUTE_UNUSED(policy);
+
+    // Create input windows
+    Window input1_win = window.broadcast_if_dimension_le_one(src0->info()->tensor_shape());
+    Window input2_win = window.broadcast_if_dimension_le_one(src1->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 = src0->info()->tensor_shape().x() != src1->info()->tensor_shape().x();
+
+    const UniformQuantizationInfo iq1_info = src0->info()->quantization_info().uniform();
+    const UniformQuantizationInfo iq2_info = src1->info()->quantization_info().uniform();
+    const UniformQuantizationInfo oq_info  = dst->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 ? src1 : src0;
+        const ITensor                *non_broadcast_tensor = !is_broadcast_input_2 ? src1 : src0;
+        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(dst, 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(src0, input1_win);
+        Iterator input2(src1, input2_win);
+        Iterator output(dst, 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), dst->info()->quantization_info());
+            }
+        },
+        input1, input2, output);
+    }
+}
+} // namespace cpu
+} // namespace arm_compute
\ No newline at end of file
diff --git a/src/core/cpu/kernels/add/neon/qsymm16.cpp b/src/core/cpu/kernels/add/neon/qsymm16.cpp
new file mode 100644
index 0000000000..e76e408d6e
--- /dev/null
+++ b/src/core/cpu/kernels/add/neon/qsymm16.cpp
@@ -0,0 +1,174 @@
+/*
+ * 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/helpers/WindowHelpers.h"
+
+namespace arm_compute
+{
+namespace cpu
+{
+void add_qsymm16_neon(const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window)
+{
+    ARM_COMPUTE_UNUSED(policy);
+
+    // Create input windows
+    Window input1_win = window.broadcast_if_dimension_le_one(src0->info()->tensor_shape());
+    Window input2_win = window.broadcast_if_dimension_le_one(src1->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 = src0->info()->tensor_shape().x() != src1->info()->tensor_shape().x();
+
+    const UniformQuantizationInfo iq1_info = src0->info()->quantization_info().uniform();
+    const UniformQuantizationInfo iq2_info = src1->info()->quantization_info().uniform();
+    const UniformQuantizationInfo oq_info  = dst->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 ? src1 : src0;
+        const ITensor                *non_broadcast_tensor = !is_broadcast_input_2 ? src1 : src0;
+        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(dst, 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(src0, input1_win);
+        Iterator input2(src1, input2_win);
+        Iterator output(dst, 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), dst->info()->quantization_info());
+            }
+        },
+        input1, input2, output);
+    }
+}
+} // namespace cpu
+} // namespace arm_compute
\ No newline at end of file
diff --git a/src/core/cpu/kernels/add/sve/integer.cpp b/src/core/cpu/kernels/add/sve/integer.cpp
new file mode 100644
index 0000000000..5bd2e12665
--- /dev/null
+++ b/src/core/cpu/kernels/add/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 add_u8_u8_s16_sve(const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window)
+{
+    // Create input windows
+    Window win        = window;
+    Window input1_win = window.broadcast_if_dimension_le_one(src0->info()->tensor_shape());
+    Window input2_win = window.broadcast_if_dimension_le_one(src1->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(src0, input1_win);
+    Iterator input2(src1, input2_win);
+    Iterator output(dst, 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 vsrc0 = svld1(pg_u, input1_ptr + x);
+                const auto vsrc1 = svld1(pg_u, input2_ptr + x);
+
+                const auto vsrc0_lo = svreinterpret_s16_u16(svunpklo(vsrc0));
+                const auto vsrc0_hi = svreinterpret_s16_u16(svunpkhi(vsrc0));
+                const auto vsrc1_lo = svreinterpret_s16_u16(svunpklo(vsrc1));
+                const auto vsrc1_hi = svreinterpret_s16_u16(svunpkhi(vsrc1));
+                svst1(pg_0, output_ptr + x, svqadd(vsrc0_lo, vsrc1_lo));
+                svst1(pg_1, output_ptr + x + svcnth(), svqadd(vsrc0_hi, vsrc1_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 vsrc0 = svld1(pg_u, input1_ptr + x);
+                const auto vsrc1 = svld1(pg_u, input2_ptr + x);
+
+                const auto vsrc0_lo = svreinterpret_s16_u16(svunpklo(vsrc0));
+                const auto vsrc0_hi = svreinterpret_s16_u16(svunpkhi(vsrc0));
+                const auto vsrc1_lo = svreinterpret_s16_u16(svunpklo(vsrc1));
+                const auto vsrc1_hi = svreinterpret_s16_u16(svunpkhi(vsrc1));
+                svst1(pg_0, output_ptr + x, svqadd(vsrc0_lo, vsrc1_lo));
+                svst1(pg_1, output_ptr + x + svcnth(), svqadd(vsrc0_hi, vsrc1_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 add_s16_u8_s16_sve(const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window)
+{
+    // Create input windows
+    Window win        = window;
+    Window input1_win = window.broadcast_if_dimension_le_one(src0->info()->tensor_shape());
+    Window input2_win = window.broadcast_if_dimension_le_one(src1->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(src0, input1_win);
+    Iterator input2(src1, input2_win);
+    Iterator output(dst, 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 vsrc0_0  = svld1_s16(pg_0, input1_ptr + x);
+                const auto vsrc0_1  = svld1_s16(pg_1, input1_ptr + x + svcnth());
+                const auto vsrc1_u8 = svld1_u8(pg_u, input2_ptr + x);
+                const auto vsrc1_0  = svreinterpret_s16_u16(svunpklo(vsrc1_u8));
+                const auto vsrc1_1  = svreinterpret_s16_u16(svunpkhi(vsrc1_u8));
+                svst1_s16(pg_0, output_ptr + x, svadd_s16_z(pg_0, vsrc0_0, vsrc1_0));
+                svst1_s16(pg_1, output_ptr + x, svadd_s16_z(pg_1, vsrc0_1, vsrc1_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 vsrc0_0  = svld1_s16(pg_0, input1_ptr + x);
+                const auto vsrc0_1  = svld1_s16(pg_1, input1_ptr + x);
+                const auto vsrc1_u8 = svld1_u8(pg_u, input2_ptr + x);
+                const auto vsrc1_0  = svreinterpret_s16_u16(svunpklo(vsrc1_u8));
+                const auto vsrc1_1  = svreinterpret_s16_u16(svunpkhi(vsrc1_u8));
+
+                svst1_s16(pg_0, output_ptr + x, svqadd(vsrc0_0, vsrc1_0));
+                svst1_s16(pg_1, output_ptr + x, svqadd(vsrc0_1, vsrc1_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 add_u8_s16_s16_sve(const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window)
+{
+    // Simply swap the two input buffers:
+    add_s16_u8_s16_sve(src1, src0, dst, policy, window);
+}
+} // namespace cpu
+} // namespace arm_compute
+#endif /* defined(__ARM_FEATURE_SVE) */
\ No newline at end of file
diff --git a/src/core/cpu/kernels/add/sve/list.h b/src/core/cpu/kernels/add/sve/list.h
new file mode 100644
index 0000000000..71dd875ad8
--- /dev/null
+++ b/src/core/cpu/kernels/add/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_ADD_LIST_H
+#define SRC_CORE_SVE_KERNELS_ADD_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_ADD_KERNEL(func_name) \
+    void func_name(const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window)
+
+DECLARE_ADD_KERNEL(add_qasymm8_sve);
+DECLARE_ADD_KERNEL(add_qasymm8_signed_sve);
+DECLARE_ADD_KERNEL(add_qsymm16_sve);
+DECLARE_ADD_KERNEL(add_s16_u8_s16_sve);
+DECLARE_ADD_KERNEL(add_u8_s16_s16_sve);
+DECLARE_ADD_KERNEL(add_u8_u8_s16_sve);
+
+#undef DECLARE_ADD_KERNEL
+
+template <typename ScalarType>
+void add_same_sve(const ITensor *src0, const ITensor *src1, ITensor *dst, 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 = src0->info()->tensor_shape().x() != src1->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(src0->info()->tensor_shape());
+    Window input2_win = window.broadcast_if_dimension_le_one(src1->info()->tensor_shape());
+
+    Iterator input1(src0, window.broadcast_if_dimension_le_one(src0->info()->tensor_shape()));
+    Iterator input2(src1, window.broadcast_if_dimension_le_one(src1->info()->tensor_shape()));
+    Iterator output(dst, 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 ? src1 : src0;
+        const ITensor *non_broadcast_tensor = !is_broadcast_input_2 ? src1 : src0;
+
+        // 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(dst, 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(src0, input1_win);
+        Iterator input2(src1, input2_win);
+        Iterator output(dst, 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_ADD_LIST_H
\ No newline at end of file
diff --git a/src/core/cpu/kernels/add/sve/qasymm8.cpp b/src/core/cpu/kernels/add/sve/qasymm8.cpp
new file mode 100644
index 0000000000..c47b5abf8a
--- /dev/null
+++ b/src/core/cpu/kernels/add/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 add_qasymm8_sve(const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window)
+{
+    ARM_COMPUTE_UNUSED(policy);
+
+    // Create input windows
+    Window input1_win = window.broadcast_if_dimension_le_one(src0->info()->tensor_shape());
+    Window input2_win = window.broadcast_if_dimension_le_one(src1->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 = src0->info()->tensor_shape().x() != src1->info()->tensor_shape().x();
+    const auto all_true_pg           = svptrue_b8();
+
+    const UniformQuantizationInfo iq1_info = src0->info()->quantization_info().uniform();
+    const UniformQuantizationInfo iq2_info = src1->info()->quantization_info().uniform();
+    const UniformQuantizationInfo oq_info  = dst->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 ? src1 : src0;
+        const ITensor *non_broadcast_tensor = !is_broadcast_input_2 ? src1 : src0;
+
+        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(dst, 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(src0, input1_win);
+        Iterator input2(src1, input2_win);
+        Iterator output(dst, 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/cpu/kernels/add/sve/qasymm8_signed.cpp b/src/core/cpu/kernels/add/sve/qasymm8_signed.cpp
new file mode 100644
index 0000000000..75d0f75a65
--- /dev/null
+++ b/src/core/cpu/kernels/add/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 add_qasymm8_signed_sve(const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window)
+{
+    ARM_COMPUTE_UNUSED(policy);
+
+    // Create input windows
+    Window input1_win = window.broadcast_if_dimension_le_one(src0->info()->tensor_shape());
+    Window input2_win = window.broadcast_if_dimension_le_one(src1->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 = src0->info()->tensor_shape().x() != src1->info()->tensor_shape().x();
+
+    const UniformQuantizationInfo iq1_info = src0->info()->quantization_info().uniform();
+    const UniformQuantizationInfo iq2_info = src1->info()->quantization_info().uniform();
+    const UniformQuantizationInfo oq_info  = dst->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 ? src1 : src0;
+        const ITensor *non_broadcast_tensor = !is_broadcast_input_2 ? src1 : src0;
+        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(dst, 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(src0, input1_win);
+        Iterator input2(src1, input2_win);
+        Iterator output(dst, 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/cpu/kernels/add/sve/qsymm16.cpp b/src/core/cpu/kernels/add/sve/qsymm16.cpp
new file mode 100644
index 0000000000..c3b72a5e65
--- /dev/null
+++ b/src/core/cpu/kernels/add/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 add_qsymm16_sve(const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window)
+{
+    ARM_COMPUTE_UNUSED(policy);
+
+    // Create input windows
+    Window input1_win = window.broadcast_if_dimension_le_one(src0->info()->tensor_shape());
+    Window input2_win = window.broadcast_if_dimension_le_one(src1->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 = src0->info()->tensor_shape().x() != src1->info()->tensor_shape().x();
+
+    const UniformQuantizationInfo iq1_info = src0->info()->quantization_info().uniform();
+    const UniformQuantizationInfo iq2_info = src1->info()->quantization_info().uniform();
+    const UniformQuantizationInfo oq_info  = dst->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 ? src1 : src0;
+        const ITensor *non_broadcast_tensor = !is_broadcast_input_2 ? src1 : src0;
+
+        // 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(dst, 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(src0, input1_win);
+        Iterator input2(src1, input2_win);
+        Iterator output(dst, 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