Rename functions/classes for elementwise operations

* Create CpuElementwise operator
* Rename kernel classes
* Make the kernels stateless

Partially implements: COMPMID-4003

Change-Id: I4ef9c61a3acc3ac5dbe46463d62dcb88a5face21
Signed-off-by: Sang-Hoon Park <sang-hoon.park@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/4881
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Sheri Zhang <sheri.zhang@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>

6 files changed, 2470 insertions, 0 deletions

diff --git a/src/core/cpu/kernels/CpuElementwiseKernel.cpp b/src/core/cpu/kernels/CpuElementwiseKernel.cpp
new file mode 100644
index 0000000000..ab915b9d72
--- /dev/null
+++ b/src/core/cpu/kernels/CpuElementwiseKernel.cpp
@@ -0,0 +1,356 @@
+/*
+ * Copyright (c) 2018-2021 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/CpuElementwiseKernel.h"
+
+#include "arm_compute/core/Helpers.h"
+#include "arm_compute/core/IAccessWindow.h"
+#include "src/core/CPP/Validate.h"
+#include "src/core/common/Registrars.h"
+#include "src/core/cpu/kernels/elementwise/neon/elementwise_list.h"
+#include "src/core/cpu/kernels/elementwise/neon/elementwise_quantized_list.h"
+#include "src/core/cpu/kernels/elementwise/sve/elementwise_list.h"
+#include "src/core/cpu/kernels/elementwise/sve/elementwise_quantized_list.h"
+#include "src/core/helpers/AutoConfiguration.h"
+#include "src/core/helpers/WindowHelpers.h"
+
+#include <arm_neon.h>
+
+namespace arm_compute
+{
+namespace cpu
+{
+namespace kernels
+{
+namespace
+{
+using ElementwiseSelector = std::add_pointer<bool(DataType)>::type;
+using UKernelType         = CpuElementwiseKernel::ElementwiseFunction;
+struct ElementwiseKernel
+{
+    const char               *name;
+    const ElementwiseSelector is_selected;
+    UKernelType              *ukernel;
+};
+
+template <DataType dt>
+inline bool is_selected(DataType data_type)
+{
+    return dt == data_type;
+}
+
+template <DataType input_data_type, DataType output_data_type = input_data_type>
+static ElementwiseKernel generate_kernel(UKernelType *ukernel)
+{
+    std::string kernel_name("op_");
+    kernel_name += string_from_data_type(input_data_type) + "_";
+    kernel_name += string_from_data_type(input_data_type) + "_";
+    kernel_name += string_from_data_type(output_data_type);
+
+    return { kernel_name.c_str(), is_selected<input_data_type>, ukernel };
+}
+
+template <ArithmeticOperation op>
+std::function<void(const ITensor *, const ITensor *, ITensor *, const Window &)>
+configure_arithm_func(const ITensorInfo *input1, const ITensorInfo *input2, ITensorInfo *output)
+{
+    ARM_COMPUTE_UNUSED(input2, output);
+    static ElementwiseKernel kernels[] =
+    {
+#if defined(__ARM_FEATURE_SVE)
+        generate_kernel<DataType::F32>(REGISTER_FP32_SVE((arm_compute::cpu::sve::elementwise_arithmetic_op<op, float32_t>))),
+        generate_kernel<DataType::S32>(REGISTER_INTEGER_SVE((arm_compute::cpu::sve::elementwise_arithmetic_op<op, int32_t>))),
+#else  /* defined(__ARM_FEATURE_SVE) */
+        generate_kernel<DataType::F32>(REGISTER_FP32_NEON((arm_compute::cpu::elementwise_arithm_op<op, typename wrapper::traits::neon_vector<float, 4>>))),
+        generate_kernel<DataType::S32>(REGISTER_INTEGER_NEON((arm_compute::cpu::elementwise_arithm_op<op, typename wrapper::traits::neon_vector<int32_t, 4>>))),
+#endif /* defined(__ARM_FEATURE_SVE) */
+#if defined(__ARM_FEATURE_SVE2)
+        generate_kernel<DataType::QASYMM8>(REGISTER_QASYMM8_SVE((arm_compute::cpu::sve::elementwise_arithmetic_quantized_op<op, uint8_t>))),
+        generate_kernel<DataType::QASYMM8_SIGNED>(REGISTER_QASYMM8_SIGNED_SVE((arm_compute::cpu::sve::elementwise_arithmetic_quantized_op<op, int8_t>))),
+#else  /* defined(__ARM_FEATURE_SVE2) */
+        generate_kernel<DataType::QASYMM8>(REGISTER_QASYMM8_NEON((arm_compute::cpu::elementwise_arithm_op_quantized<op>))),
+        generate_kernel<DataType::QASYMM8_SIGNED>(REGISTER_QASYMM8_SIGNED_NEON((arm_compute::cpu::elementwise_arithm_op_quantized_signed<op>))),
+#endif /* defined(__ARM_FEATURE_SVE2) */
+#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
+#if defined(__ARM_FEATURE_SVE)
+        generate_kernel<DataType::F16>(REGISTER_FP16_SVE((arm_compute::cpu::sve::elementwise_arithmetic_op<op, float16_t>))),
+#else  /* defined(__ARM_FEATURE_SVE) */
+        generate_kernel<DataType::F16>(REGISTER_FP16_NEON((arm_compute::cpu::elementwise_arithm_op<op, typename wrapper::traits::neon_vector<float16_t, 8>>))),
+#endif /* defined(__ARM_FEATURE_SVE) */
+#endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */
+        generate_kernel<DataType::S16>(REGISTER_INTEGER_NEON((arm_compute::cpu::elementwise_arithm_op<op, typename wrapper::traits::neon_vector<int16_t, 8>>))),
+    };
+
+    for(const auto &uk : kernels)
+    {
+        if(uk.is_selected(input1->data_type()))
+        {
+            return uk.ukernel;
+        }
+    }
+
+    return nullptr;
+}
+
+template <ComparisonOperation op>
+std::function<void(const ITensor *input1, const ITensor *input2, ITensor *output, const Window &window)>
+configure_comp_func(const ITensorInfo *input1, const ITensorInfo *input2, ITensorInfo *output)
+{
+    ARM_COMPUTE_UNUSED(input2, output);
+    static ElementwiseKernel kernels[] =
+    {
+#if defined(__ARM_FEATURE_SVE)
+        generate_kernel<DataType::U8, DataType::U8>(REGISTER_INTEGER_SVE((arm_compute::cpu::sve::elementwise_comparison_op<op, uint8_t>))),
+        generate_kernel<DataType::F32, DataType::U8>(REGISTER_FP32_SVE((arm_compute::cpu::sve::elementwise_comparison_op<op, float>))),
+        generate_kernel<DataType::S16, DataType::U8>(REGISTER_INTEGER_SVE((arm_compute::cpu::sve::elementwise_comparison_op<op, int16_t>))),
+        generate_kernel<DataType::S32, DataType::U8>(REGISTER_INTEGER_SVE((arm_compute::cpu::sve::elementwise_comparison_op<op, int32_t>))),
+#else  /* defined(__ARM_FEATURE_SVE) */
+        generate_kernel<DataType::U8, DataType::U8>(REGISTER_INTEGER_NEON((arm_compute::cpu::elementwise_comp_op_8<op, uint8_t, uint8x16_t>))),
+        generate_kernel<DataType::F32, DataType::U8>(REGISTER_FP32_NEON((arm_compute::cpu::elementwise_comp_op_32<op, float, float32x4_t>))),
+        generate_kernel<DataType::S16, DataType::U8>(REGISTER_INTEGER_NEON((arm_compute::cpu::elementwise_comp_op_16<op, int16_t, int16x8_t>))),
+        generate_kernel<DataType::S32, DataType::U8>(REGISTER_INTEGER_NEON((arm_compute::cpu::elementwise_comp_op_32<op, int32_t, int32x4_t>))),
+#endif /* defined(__ARM_FEATURE_SVE) */
+#if defined(__ARM_FEATURE_SVE2)
+        generate_kernel<DataType::QASYMM8_SIGNED, DataType::U8>(REGISTER_QASYMM8_SIGNED_SVE((arm_compute::cpu::sve::elementwise_comparison_quantized_op<op, int8_t>))),
+        generate_kernel<DataType::QASYMM8, DataType::U8>(REGISTER_QASYMM8_SVE((arm_compute::cpu::sve::elementwise_comparison_quantized_op<op, uint8_t>))),
+#else  /* defined(__ARM_FEATURE_SVE2) */
+        generate_kernel<DataType::QASYMM8_SIGNED, DataType::U8>(REGISTER_QASYMM8_SIGNED_NEON((arm_compute::cpu::elementwise_comp_op_quantized_signed<op>))),
+        generate_kernel<DataType::QASYMM8, DataType::U8>(REGISTER_QASYMM8_NEON((arm_compute::cpu::elementwise_comp_op_quantized<op>))),
+#endif /* defined(__ARM_FEATURE_SVE2) */
+#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
+#if defined(__ARM_FEATURE_SVE)
+        generate_kernel<DataType::F16, DataType::U8>(REGISTER_FP16_SVE((arm_compute::cpu::sve::elementwise_comparison_op<op, float16_t>))),
+#else  /* defined(__ARM_FEATURE_SVE) */
+        generate_kernel<DataType::F16, DataType::U8>(REGISTER_FP16_NEON((arm_compute::cpu::elementwise_comp_op_16<op, float16_t, float16x8_t>))),
+#endif /* defined(__ARM_FEATURE_SVE) */
+#endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */
+    };
+
+    for(const auto &uk : kernels)
+    {
+        if(uk.is_selected(input1->data_type()))
+        {
+            return uk.ukernel;
+        }
+    }
+
+    return nullptr;
+}
+} // namespace
+
+Status CpuElementwiseKernel::validate_arguments_common(const ITensorInfo &input1, const ITensorInfo &input2, const ITensorInfo &output)
+{
+    ARM_COMPUTE_RETURN_ERROR_ON_CPU_F16_UNSUPPORTED(&input1);
+    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(&input1, &input2);
+
+    const TensorShape out_shape = TensorShape::broadcast_shape(input1.tensor_shape(), input2.tensor_shape());
+
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG(out_shape.total_size() == 0, "Inputs are not broadcast compatible");
+
+    // Validate in case of configured output
+    if(output.total_size() > 0)
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON_MSG(detail::have_different_dimensions(out_shape, output.tensor_shape(), 0),
+                                        "Wrong shape for output");
+    }
+
+    return Status{};
+}
+
+void CpuElementwiseKernel::configure_common(const ITensorInfo *input1, const ITensorInfo *input2, ITensorInfo *output)
+{
+    ARM_COMPUTE_ERROR_ON_NULLPTR(input1, input2, output);
+
+    // Configure kernel window
+    const std::pair<TensorShape, ValidRegion> broadcast_pair = ITensorInfo::broadcast_shape_and_valid_region(*input1, *input2);
+    const TensorShape &out_shape    = broadcast_pair.first;
+    const ValidRegion &valid_region = broadcast_pair.second;
+
+    // Auto initialize output if not initialized
+    auto_init_if_empty(*output, out_shape, 1, input1->data_type());
+
+    Window win = calculate_max_window(valid_region);
+
+    ICpuKernel::configure(win);
+}
+
+void CpuElementwiseKernel::run_op(ITensorPack &tensors, const Window &window, const ThreadInfo &info)
+{
+    ARM_COMPUTE_UNUSED(info, window);
+    ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
+    ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICpuKernel::window(), window);
+
+    auto src0 = tensors.get_const_tensor(TensorType::ACL_SRC_0);
+    auto src1 = tensors.get_const_tensor(TensorType::ACL_SRC_1);
+    auto dst  = tensors.get_tensor(TensorType::ACL_DST);
+
+    auto function = get_implementation(src0->info(), src1->info(), dst->info());
+    ARM_COMPUTE_ERROR_ON(function == nullptr);
+    function(src0, src1, dst, window);
+}
+
+/** Arithmetic operators (min, max, squared_diff) */
+void CpuArithmeticKernel::configure(ArithmeticOperation op, const ITensorInfo *input1, const ITensorInfo *input2, ITensorInfo *output)
+{
+    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(*input1, *input2, *output));
+    configure_common(input1, input2, output);
+    _op = op;
+}
+
+Status CpuArithmeticKernel::validate_arguments(const ITensorInfo &input1, const ITensorInfo &input2, const ITensorInfo &output)
+{
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(&input1, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED, DataType::S16, DataType::F16, DataType::S32, DataType::F32);
+    // Validate in case of configured output
+    if(output.total_size() > 0)
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(&input1, &output);
+    }
+    return validate_arguments_common(input1, input2, output);
+}
+
+Status CpuArithmeticKernel::validate(ArithmeticOperation op, const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *output)
+{
+    ARM_COMPUTE_UNUSED(op);
+    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input1, input2, output);
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(*input1, *input2, *output));
+    return Status{};
+}
+
+std::function<CpuElementwiseKernel::ElementwiseFunction>
+CpuArithmeticKernel::get_implementation(const ITensorInfo *input1, const ITensorInfo *input2, ITensorInfo *output)
+{
+    switch(_op)
+    {
+        case ArithmeticOperation::MAX:
+            return configure_arithm_func<ArithmeticOperation::MAX>(input1, input2, output);
+        case ArithmeticOperation::MIN:
+            return configure_arithm_func<ArithmeticOperation::MIN>(input1, input2, output);
+        case ArithmeticOperation::SQUARED_DIFF:
+            return configure_arithm_func<ArithmeticOperation::SQUARED_DIFF>(input1, input2, output);
+        case ArithmeticOperation::PRELU:
+            return configure_arithm_func<ArithmeticOperation::PRELU>(input1, input2, output);
+        case ArithmeticOperation::DIV:
+            return configure_arithm_func<ArithmeticOperation::DIV>(input1, input2, output);
+        case ArithmeticOperation::POWER:
+            return configure_arithm_func<ArithmeticOperation::POWER>(input1, input2, output);
+        default:
+            ARM_COMPUTE_ERROR("NOT_SUPPORTED!");
+    }
+    return nullptr;
+}
+
+/** The division operator */
+
+void CpuDivisionKernel::configure(const ITensorInfo *input1, const ITensorInfo *input2, ITensorInfo *output)
+{
+    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(*input1, *input2, *output));
+    configure_common(input1, input2, output);
+    _op = ArithmeticOperation::DIV;
+}
+
+Status CpuDivisionKernel::validate_arguments(const ITensorInfo &input1, const ITensorInfo &input2, const ITensorInfo &output)
+{
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(&input1, 1, DataType::S32, DataType::F16, DataType::F32);
+    return CpuArithmeticKernel::validate_arguments(input1, input2, output);
+}
+
+Status CpuDivisionKernel::validate(const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *output)
+{
+    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input1, input2, output);
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(*input1, *input2, *output));
+    return Status{};
+}
+
+/** The power operator */
+void CpuPowerKernel::configure(const ITensorInfo *input1, const ITensorInfo *input2, ITensorInfo *output)
+{
+    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(*input1, *input2, *output));
+    configure_common(input1, input2, output);
+    _op = ArithmeticOperation::POWER;
+}
+
+Status CpuPowerKernel::validate_arguments(const ITensorInfo &input1, const ITensorInfo &input2, const ITensorInfo &output)
+{
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(&input1, 1, DataType::F16, DataType::F32);
+    return CpuArithmeticKernel::validate_arguments(input1, input2, output);
+}
+
+Status CpuPowerKernel::validate(const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *output)
+{
+    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input1, input2, output);
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(*input1, *input2, *output));
+    return Status{};
+}
+
+/** Comparison operators (equal, not equal, less than, greater than, less than or equal, greater than or equal) */
+void CpuComparisonKernel::configure(ComparisonOperation op, const ITensorInfo *input1, const ITensorInfo *input2, ITensorInfo *output)
+{
+    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(*input1, *input2, *output));
+    configure_common(input1, input2, output);
+    _op = op;
+}
+
+Status CpuComparisonKernel::validate_arguments(const ITensorInfo &input1, const ITensorInfo &input2, const ITensorInfo &output)
+{
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(&input1, 1, DataType::U8, DataType::QASYMM8, DataType::QASYMM8_SIGNED, DataType::S16, DataType::F16, DataType::S32, DataType::F32);
+    // Validate in case of configured output
+    if(output.total_size() > 0)
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(&output, 1, DataType::U8);
+    }
+    return validate_arguments_common(input1, input2, output);
+}
+
+Status CpuComparisonKernel::validate(ComparisonOperation op, const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *output)
+{
+    ARM_COMPUTE_UNUSED(op);
+    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input1, input2, output);
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(*input1, *input2, *output));
+    return Status{};
+}
+
+std::function<CpuElementwiseKernel::ElementwiseFunction>
+CpuComparisonKernel::get_implementation(const ITensorInfo *input1, const ITensorInfo *input2, ITensorInfo *output)
+{
+    switch(_op)
+    {
+        case ComparisonOperation::Equal:
+            return configure_comp_func<ComparisonOperation::Equal>(input1, input2, output);
+        case ComparisonOperation::NotEqual:
+            return configure_comp_func<ComparisonOperation::NotEqual>(input1, input2, output);
+        case ComparisonOperation::Greater:
+            return configure_comp_func<ComparisonOperation::Greater>(input1, input2, output);
+        case ComparisonOperation::GreaterEqual:
+            return configure_comp_func<ComparisonOperation::GreaterEqual>(input1, input2, output);
+        case ComparisonOperation::Less:
+            return configure_comp_func<ComparisonOperation::Less>(input1, input2, output);
+        case ComparisonOperation::LessEqual:
+            return configure_comp_func<ComparisonOperation::LessEqual>(input1, input2, output);
+        default:
+            ARM_COMPUTE_ERROR("NOT_SUPPORTED!");
+    }
+    return nullptr;
+}
+} // namespace kernels
+} // namespace cpu
+} // namespace arm_compute
diff --git a/src/core/cpu/kernels/CpuElementwiseKernel.h b/src/core/cpu/kernels/CpuElementwiseKernel.h
new file mode 100644
index 0000000000..92cf880172
--- /dev/null
+++ b/src/core/cpu/kernels/CpuElementwiseKernel.h
@@ -0,0 +1,239 @@
+/*
+ * 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.
+ */
+#ifndef ARM_COMPUTE_CPU_ELEMENTWISE_KERNEL_H
+#define ARM_COMPUTE_CPU_ELEMENTWISE_KERNEL_H
+
+#include "src/core/common/Macros.h"
+#include "src/core/cpu/ICpuKernel.h"
+
+namespace arm_compute
+{
+class ITensor;
+namespace cpu
+{
+namespace kernels
+{
+/** Interface for an element-wise operation kernel
+ *
+ * Element-wise operation is computed by:
+ * @f[ output(x,y) = OP(input1(x,y), input2(x,y))@f]
+ *
+ */
+class CpuElementwiseKernel : public ICpuKernel
+{
+public:
+    const char *name() const override
+    {
+        return "CpuElementwiseKernel";
+    }
+
+    CpuElementwiseKernel() = default;
+    ARM_COMPUTE_DISALLOW_COPY_ALLOW_MOVE(CpuElementwiseKernel);
+
+    /** Common signature for all the specialised arithmetic functions
+     *
+     * @param[in]  input1 First tensor input info. Data types supported: QASYMM8/S16/F16/S32/F32.
+     * @param[in]  input2 Second tensor input info. Data types supported: Same as @p input1.
+     * @param[out] output Output tensor info. Data types supported: Dependent on subclass.
+     * @param[in]  window Region on which to execute the kernel.
+     */
+    using ElementwiseFunction = void(const ITensor *, const ITensor *, ITensor *, const Window &);
+
+    // Inherited methods overridden:
+    void run_op(ITensorPack &tensors, const Window &window, const ThreadInfo &info) override;
+
+protected:
+    /** Validate the argument passed to the kernel
+     *
+     * @param[in] input1 First tensor input. Data types supported: QASYMM8/S16/F16/S32/F32.
+     * @param[in] input2 Second tensor input. Data types supported: Same as @p input1.
+     * @param[in] output Output tensor. Data types supported: Dependent on subclass.
+     */
+    static Status validate_arguments_common(const ITensorInfo &input1, const ITensorInfo &input2, const ITensorInfo &output);
+
+    /** Commmon configure function for element-wise operators with no additional options (e.g. Min, Max, SquaredDiff)
+     *
+     */
+    void configure_common(const ITensorInfo *input1, const ITensorInfo *input2, ITensorInfo *output);
+
+    /** Function to get the micro kernel implementation
+     *
+     * @param[in] input1 First input tensor information
+     * @param[in] input2 Second input tensor information
+     * @param[in] output Output tensor information
+     *
+     * @return the function instance for the micro kernel
+     */
+    virtual std::function<ElementwiseFunction> get_implementation(const ITensorInfo *input1, const ITensorInfo *input2, ITensorInfo *output) = 0;
+};
+
+class CpuArithmeticKernel : public CpuElementwiseKernel
+{
+public:
+    /** Default constructor */
+    CpuArithmeticKernel() = default;
+
+    /** Configure kernel
+     *
+     * @param[in]  op     Arithmetic operation to be executed.
+     * @param[in]  input1 First tensor input info. Data types supported: QASYMM8/S16/F16/S32/F32.
+     * @param[in]  input2 Second tensor input info. Data types supported: Same as @p input1.
+     * @param[out] output Output tensor info. Data types supported: Same as @p input1.
+     */
+    void configure(ArithmeticOperation op, const ITensorInfo *input1, const ITensorInfo *input2, ITensorInfo *output);
+
+    /** Static function to check if given info will lead to a valid configuration of @ref cpu::kernels::CpuArithmeticKernel
+     *
+     * @param[in] op     Arithmetic operation to be executed.
+     * @param[in] input1 First tensor input info. Data types supported: QASYMM8/S16/F16/S32/F32.
+     * @param[in] input2 Second tensor input info. Data types supported: Same as @p input1.
+     * @param[in] output Output tensor info. Data types supported: Same as @p input1.
+     *
+     * @return a Status
+     */
+    static Status validate(ArithmeticOperation op, const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *output);
+
+protected:
+    // Inherited methods overridden:
+    static Status validate_arguments(const ITensorInfo &input1, const ITensorInfo &input2, const ITensorInfo &output);
+
+    ArithmeticOperation _op{};
+
+private:
+    /** Function to get the micro kernel implementation
+     *
+     * @param[in] input1 First input tensor information
+     * @param[in] input2 Second input tensor information
+     * @param[in] output Output tensor information
+     *
+     * @return the function instance for the micro kernel
+     */
+    std::function<ElementwiseFunction> get_implementation(const ITensorInfo *input1, const ITensorInfo *input2, ITensorInfo *output) override;
+};
+
+class CpuDivisionKernel : public CpuArithmeticKernel
+{
+public:
+    /** Default constructor */
+    CpuDivisionKernel() = default;
+
+    /** Configure kernel
+     *
+     * @param[in]  input1 First tensor input info. Data types supported: S32/F16/F32.
+     * @param[in]  input2 Second tensor input info. Data types supported: Same as @p input1.
+     * @param[out] output Output tensor info. Data types supported: Same as @p input1.
+     */
+    void configure(const ITensorInfo *input1, const ITensorInfo *input2, ITensorInfo *output);
+
+    /** Static function to check if given info will lead to a valid configuration of @ref CpuDivisionKernel
+     *
+     * @param[in] input1 First tensor input info. Data types supported: S32/F16/F32.
+     * @param[in] input2 Second tensor input info. Data types supported: Same as @p input1.
+     * @param[in] output Output tensor info. Data types supported: Same as @p input1.
+     *
+     * @return a Status
+     */
+    static Status validate(const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *output);
+
+protected:
+    // Inherited methods overridden:
+    static Status validate_arguments(const ITensorInfo &input1, const ITensorInfo &input2, const ITensorInfo &output);
+};
+
+class CpuPowerKernel : public CpuArithmeticKernel
+{
+public:
+    /** Default constructor */
+    CpuPowerKernel() = default;
+
+    /** Configure kernel
+     *
+     * @param[in]  input1 First tensor input info. Data types supported: F16/F32.
+     * @param[in]  input2 Second tensor input info. Data types supported: Same as @p input1.
+     * @param[out] output Output tensor info. Data types supported: Same as @p input1.
+     */
+    void configure(const ITensorInfo *input1, const ITensorInfo *input2, ITensorInfo *output);
+
+    /** Static function to check if given info will lead to a valid configuration of @ref CpuPowerKernel
+     *
+     * @param[in] input1 First tensor input info. Data types supported: F16/F32.
+     * @param[in] input2 Second tensor input info. Data types supported: Same as @p input1.
+     * @param[in] output Output tensor info. Data types supported: Same as @p input1.
+     *
+     * @return a Status
+     */
+    static Status validate(const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *output);
+
+protected:
+    // Inherited methods overridden:
+    static Status validate_arguments(const ITensorInfo &input1, const ITensorInfo &input2, const ITensorInfo &output);
+};
+
+class CpuComparisonKernel : public CpuElementwiseKernel
+{
+public:
+    /** Default constructor */
+    CpuComparisonKernel() = default;
+
+    /** Configure kernel
+     *
+     * @param[in]  op     Comparison operation to be executed.
+     * @param[in]  input1 First tensor input info. Data types supported: QASYMM8/QASYMM8_SIGNED/S16/F16/S32/F32.
+     * @param[in]  input2 Second tensor input info. Data types supported: Same as @p input1.
+     * @param[out] output Output tensor info. Data types supported: U8.
+     */
+    void configure(ComparisonOperation op, const ITensorInfo *input1, const ITensorInfo *input2, ITensorInfo *output);
+
+    /** Static function to check if given info will lead to a valid configuration of @ref cpu::kernels::CpuComparisonKernel
+     *
+     * @param[in] op     Comparison operation to be executed.
+     * @param[in] input1 First tensor input info. Data types supported: QASYMM8/QASYMM8_SIGNED/S16/F16/S32/F32.
+     * @param[in] input2 Second tensor input info. Data types supported: Same as @p input1.
+     * @param[in] output Output tensor info. Data types supported: U8.
+     *
+     * @return a Status
+     */
+    static Status validate(ComparisonOperation op, const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *output);
+
+protected:
+    // Inherited methods overridden:
+    static Status validate_arguments(const ITensorInfo &input1, const ITensorInfo &input2, const ITensorInfo &output);
+
+private:
+    /** Function to get the micro kernel implementation
+     *
+     * @param[in] input1 First input tensor information
+     * @param[in] input2 Second input tensor information
+     * @param[in] output Output tensor information
+     *
+     * @return the function instance for the micro kernel
+     */
+    std::function<ElementwiseFunction> get_implementation(const ITensorInfo *input1, const ITensorInfo *input2, ITensorInfo *output) override;
+
+    ComparisonOperation _op{};
+};
+} // namespace kernels
+} // namespace cpu
+} // namespace arm_compute
+#endif /* ARM_COMPUTE_CPU_ELEMENTWISE_KERNEL_H */
\ No newline at end of file
diff --git a/src/core/cpu/kernels/elementwise/neon/elementwise_list.h b/src/core/cpu/kernels/elementwise/neon/elementwise_list.h
new file mode 100644
index 0000000000..43e44be5e2
--- /dev/null
+++ b/src/core/cpu/kernels/elementwise/neon/elementwise_list.h
@@ -0,0 +1,486 @@
+/*
+ * 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.
+ */
+#ifndef SRC_CORE_NEON_KERNELS_ELEMENTWISE_LIST_H
+#define SRC_CORE_NEON_KERNELS_ELEMENTWISE_LIST_H
+
+#include "src/core/NEON/NEAsymm.h"
+#include "src/core/NEON/wrapper/wrapper.h"
+#include "src/core/helpers/WindowHelpers.h"
+
+namespace arm_compute
+{
+namespace cpu
+{
+template <typename InputScalarType, typename OutputScalarType, typename InputVectorType>
+void elementwise_op(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window,
+                    OutputScalarType (*scalar_func)(const InputScalarType &, const InputScalarType &),
+                    int (*broadcast_func)(int, int, int, const InputScalarType *, const InputScalarType &, OutputScalarType *, const bool),
+                    int (*neon_func)(int, int, int, const InputScalarType *, const InputScalarType *, OutputScalarType *))
+{
+    // Create input windows
+    Window input1_win = window.broadcast_if_dimension_le_one(in1->info()->tensor_shape());
+    Window input2_win = window.broadcast_if_dimension_le_one(in2->info()->tensor_shape());
+
+    // Clear X Dimension on execution window as we handle manually
+    Window win = window;
+    win.set(Window::DimX, Window::Dimension(0, 1, 1));
+
+    const int  window_step_x         = std::min(16 / static_cast<int>(sizeof(OutputScalarType)), 8);
+    const auto window_start_x        = static_cast<int>(window.x().start());
+    const auto window_end_x          = static_cast<int>(window.x().end());
+    const bool is_broadcast_across_x = in1->info()->tensor_shape().x() != in2->info()->tensor_shape().x();
+
+    if(is_broadcast_across_x)
+    {
+        const bool     is_broadcast_input_2 = input2_win.x().step() == 0;
+        Window         broadcast_win        = is_broadcast_input_2 ? input2_win : input1_win;
+        Window         non_broadcast_win    = !is_broadcast_input_2 ? input2_win : input1_win;
+        const ITensor *broadcast_tensor     = is_broadcast_input_2 ? in2 : in1;
+        const ITensor *non_broadcast_tensor = !is_broadcast_input_2 ? in2 : in1;
+
+        // Clear X Dimension on execution window as we handle manually
+        non_broadcast_win.set(Window::DimX, Window::Dimension(0, 1, 1));
+
+        Iterator broadcast_input(broadcast_tensor, broadcast_win);
+        Iterator non_broadcast_input(non_broadcast_tensor, non_broadcast_win);
+        Iterator output(out, win);
+
+        execute_window_loop(win, [&](const Coordinates &)
+        {
+            auto                  output_ptr              = reinterpret_cast<OutputScalarType *>(output.ptr());
+            const auto            non_broadcast_input_ptr = reinterpret_cast<const InputScalarType *>(non_broadcast_input.ptr());
+            const InputScalarType broadcast_value         = *reinterpret_cast<const InputScalarType *>(broadcast_input.ptr());
+
+            int x = (*broadcast_func)(window_start_x, window_end_x, window_step_x, non_broadcast_input_ptr, broadcast_value, output_ptr, !is_broadcast_input_2);
+            for(; x < window_end_x; ++x)
+            {
+                const auto a      = *(non_broadcast_input_ptr + x);
+                *(output_ptr + x) = (*scalar_func)(!is_broadcast_input_2 ? broadcast_value : a, !is_broadcast_input_2 ? a : broadcast_value);
+            }
+        },
+        broadcast_input, non_broadcast_input, output);
+    }
+    else
+    {
+        // Clear X Dimension on execution window as we handle manually
+        input1_win.set(Window::DimX, Window::Dimension(0, 1, 1));
+        input2_win.set(Window::DimX, Window::Dimension(0, 1, 1));
+
+        Iterator input1(in1, input1_win);
+        Iterator input2(in2, input2_win);
+        Iterator output(out, win);
+
+        execute_window_loop(win, [&](const Coordinates &)
+        {
+            auto       output_ptr = reinterpret_cast<OutputScalarType *>(output.ptr());
+            const auto input1_ptr = reinterpret_cast<const InputScalarType *>(input1.ptr());
+            const auto input2_ptr = reinterpret_cast<const InputScalarType *>(input2.ptr());
+
+            int x = (*neon_func)(window_start_x, window_end_x, window_step_x, input1_ptr, input2_ptr, output_ptr);
+            for(; x < window_end_x; ++x)
+            {
+                const auto a      = *(input1_ptr + x);
+                const auto b      = *(input2_ptr + x);
+                *(output_ptr + x) = (*scalar_func)(a, b);
+            }
+        },
+        input1, input2, output);
+    }
+}
+
+template <ArithmeticOperation op, typename ScalarType>
+inline ScalarType elementwise_arithm_op_scalar(const ScalarType &a, const ScalarType &b)
+{
+    auto res = ScalarType(0);
+
+    switch(op)
+    {
+        case ArithmeticOperation::MAX:
+            res = std::max(a, b);
+            break;
+        case ArithmeticOperation::MIN:
+            res = std::min(a, b);
+            break;
+        case ArithmeticOperation::SQUARED_DIFF:
+        {
+            res = (a - b) * (a - b);
+            break;
+        }
+        case ArithmeticOperation::PRELU:
+        {
+            res = (a > 0 ? a : a * b);
+            break;
+        }
+        case ArithmeticOperation::DIV:
+        {
+            res = a / b;
+            if(std::is_integral<ScalarType>::value)
+            {
+                res = (b == 0) ? 0 : res;
+                if(static_cast<int32_t>(a) % static_cast<int32_t>(b) != 0 && ((a < 0) != (b < 0)))
+                {
+                    --res;
+                }
+            }
+            break;
+        }
+        case ArithmeticOperation::POWER:
+        {
+            res = std::pow(a, b);
+            break;
+        }
+        default:
+            ARM_COMPUTE_ERROR("NOT_SUPPORTED!");
+    }
+    return res;
+}
+
+template <ArithmeticOperation    op, typename VectorType>
+inline typename VectorType::type elementwise_arithm_op(const typename VectorType::type &a, const typename VectorType::type &b)
+{
+    using vec_type    = typename VectorType::type;
+    using scalar_type = typename VectorType::scalar_type;
+    using tag_type    = typename VectorType::tag_type;
+
+    vec_type res = wrapper::vdup_n(static_cast<scalar_type>(0), tag_type{});
+
+    switch(op)
+    {
+        case ArithmeticOperation::MAX:
+            res = wrapper::vmax(a, b);
+            break;
+        case ArithmeticOperation::MIN:
+            res = wrapper::vmin(a, b);
+            break;
+        case ArithmeticOperation::SQUARED_DIFF:
+        {
+            const vec_type tmp = wrapper::vsub(a, b);
+            res                = wrapper::vmul(tmp, tmp);
+            break;
+        }
+        case ArithmeticOperation::PRELU:
+        {
+            const vec_type zero = wrapper::vdup_n(static_cast<scalar_type>(0), tag_type{});
+            const vec_type tmp  = wrapper::vmul(a, b);
+            const auto     gt   = wrapper::vcgt(a, zero);
+
+            res = wrapper::vbsl(gt, a, tmp);
+            break;
+        }
+
+        default:
+            ARM_COMPUTE_ERROR("NOT_SUPPORTED!");
+    }
+
+    return res;
+}
+
+template <>
+inline int32x4_t elementwise_arithm_op<ArithmeticOperation::DIV, typename wrapper::traits::neon_vector<int32_t, 4>>(const int32x4_t &a, const int32x4_t &b)
+{
+    return vcvtq_s32_f32(vfloorq_f32(wrapper::vdiv(vcvtq_f32_s32(a), vcvtq_f32_s32(b))));
+}
+
+template <>
+inline float32x4_t elementwise_arithm_op<ArithmeticOperation::DIV, typename wrapper::traits::neon_vector<float, 4>>(const float32x4_t &a, const float32x4_t &b)
+{
+    return wrapper::vdiv(a, b);
+}
+
+template <>
+inline float32x4_t elementwise_arithm_op<ArithmeticOperation::POWER, typename wrapper::traits::neon_vector<float, 4>>(const float32x4_t &a, const float32x4_t &b)
+{
+    return wrapper::vpow(a, b);
+}
+
+#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
+template <>
+inline float16x8_t elementwise_arithm_op<ArithmeticOperation::DIV, typename wrapper::traits::neon_vector<float16_t, 8>>(const float16x8_t &a, const float16x8_t &b)
+{
+    return wrapper::vdiv(a, b);
+}
+
+template <>
+inline float16x8_t elementwise_arithm_op<ArithmeticOperation::POWER, typename wrapper::traits::neon_vector<float16_t, 8>>(const float16x8_t &a, const float16x8_t &b)
+{
+    return wrapper::vpow(a, b);
+}
+#endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
+
+template <ArithmeticOperation    op, typename ScalarType, typename VectorType>
+inline typename VectorType::type elementwise_arithm_op_broadcast(const typename VectorType::type &a, const ScalarType &broadcast_value, const bool reorder)
+{
+    using tag_type = typename VectorType::tag_type;
+    using vec_type = typename VectorType::type;
+
+    vec_type broadcast_vector = wrapper::vdup_n(broadcast_value, tag_type{});
+    return elementwise_arithm_op<op, VectorType>(reorder ? broadcast_vector : a, reorder ? a : broadcast_vector);
+}
+
+template <ArithmeticOperation op, typename ScalarType, typename VectorType>
+inline int elementwise_arithm_op_loop(int window_start_x, int window_end_x, int window_step_x,
+                                      const ScalarType *input1_ptr, const ScalarType *input2_ptr, ScalarType *output_ptr)
+{
+    int x = window_start_x;
+    for(; x <= (window_end_x - window_step_x); x += window_step_x)
+    {
+        const auto a = wrapper::vloadq(input1_ptr + x);
+        const auto b = wrapper::vloadq(input2_ptr + x);
+        wrapper::vstore(output_ptr + x, elementwise_arithm_op<op, VectorType>(a, b));
+    }
+    return x;
+}
+
+template <ArithmeticOperation op, typename ScalarType, typename VectorType>
+inline int elementwise_arithm_op_broadcast_loop(int window_start_x, int window_end_x, int window_step_x,
+                                                const ScalarType *non_broadcast_input_ptr, const ScalarType &broadcast_value, ScalarType *output_ptr, const bool reorder)
+{
+    int x = window_start_x;
+    for(; x <= (window_end_x - window_step_x); x += window_step_x)
+    {
+        const auto a = wrapper::vloadq((non_broadcast_input_ptr + x));
+        wrapper::vstore(output_ptr + x, elementwise_arithm_op_broadcast<op, ScalarType, VectorType>(a, broadcast_value, reorder));
+    }
+    return x;
+}
+
+template <ArithmeticOperation op, typename VectorType>
+void elementwise_arithm_op(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window)
+{
+    using scalar_type = typename VectorType::scalar_type;
+
+    elementwise_op<scalar_type, scalar_type, VectorType>(in1, in2, out, window,
+                                                         &elementwise_arithm_op_scalar<op, scalar_type>,
+                                                         &elementwise_arithm_op_broadcast_loop<op, scalar_type, VectorType>,
+                                                         &elementwise_arithm_op_loop<op, scalar_type, VectorType>);
+}
+
+template <ComparisonOperation op, typename InputScalarType>
+inline uint8_t elementwise_comp_op_scalar(const InputScalarType &a, const InputScalarType &b)
+{
+    bool res = false;
+
+    switch(op)
+    {
+        case ComparisonOperation::Equal:
+            res = (a == b);
+            break;
+        case ComparisonOperation::NotEqual:
+            res = (a != b);
+            break;
+        case ComparisonOperation::Greater:
+            res = (a > b);
+            break;
+        case ComparisonOperation::GreaterEqual:
+            res = (a >= b);
+            break;
+        case ComparisonOperation::Less:
+            res = (a < b);
+            break;
+        case ComparisonOperation::LessEqual:
+            res = (a <= b);
+            break;
+        default:
+            ARM_COMPUTE_ERROR("NOT_SUPPORTED!");
+    }
+    return res ? ~static_cast<uint8_t>(0) : static_cast<uint8_t>(0);
+}
+
+template <ComparisonOperation op, typename InputVectorType, typename OutputVectorType>
+inline OutputVectorType elementwise_comp_op(const InputVectorType &a, const InputVectorType &b)
+{
+    OutputVectorType res = { 0, 0, 0, 0 };
+
+    switch(op)
+    {
+        case ComparisonOperation::Equal:
+            res = wrapper::vceq(a, b);
+            break;
+        case ComparisonOperation::NotEqual:
+            res = wrapper::vnot(wrapper::vceq(a, b));
+            break;
+        case ComparisonOperation::Greater:
+            res = wrapper::vcgt(a, b);
+            break;
+        case ComparisonOperation::GreaterEqual:
+            res = wrapper::vcge(a, b);
+            break;
+        case ComparisonOperation::Less:
+            res = wrapper::vcgt(b, a);
+            break;
+        case ComparisonOperation::LessEqual:
+            res = wrapper::vcge(b, a);
+            break;
+        default:
+            ARM_COMPUTE_ERROR("NOT_SUPPORTED!");
+    }
+
+    return res;
+}
+
+template <ComparisonOperation op, typename InputScalarType, typename InputVectorType, typename OutputVectorType>
+inline OutputVectorType elementwise_comp_op_broadcast(const InputVectorType &a, const InputScalarType &broadcast_value, const bool reorder)
+{
+    InputVectorType broadcast_vector = wrapper::vdup_n(broadcast_value, wrapper::traits::vector_128_tag());
+    return elementwise_comp_op<op, InputVectorType, OutputVectorType>(reorder ? broadcast_vector : a, reorder ? a : broadcast_vector);
+}
+
+template <ComparisonOperation op, typename InputScalarType, typename InputVectorType>
+inline int elementwise_comp_op_broadcast_8_loop(int window_start_x, int window_end_x, int window_step_x,
+                                                const InputScalarType *non_broadcast_input_ptr, const InputScalarType &broadcast_value, uint8_t *output_ptr, const bool reorder)
+{
+    int x = window_start_x;
+    for(; x <= (window_end_x - window_step_x); x += window_step_x)
+    {
+        const auto a = elementwise_comp_op_broadcast<op, InputScalarType, InputVectorType, uint8x16_t>(wrapper::vloadq((non_broadcast_input_ptr + x)), broadcast_value, reorder);
+        wrapper::vstore(output_ptr + x, a);
+    }
+    return x;
+}
+
+template <ComparisonOperation op, typename InputScalarType, typename InputVectorType>
+inline int elementwise_comp_op_broadcast_16_loop(int window_start_x, int window_end_x, int window_step_x,
+                                                 const InputScalarType *non_broadcast_input_ptr, const InputScalarType &broadcast_value, uint8_t *output_ptr, const bool reorder)
+{
+    int x = window_start_x;
+    for(; x <= (window_end_x - window_step_x); x += window_step_x)
+    {
+        const auto a = elementwise_comp_op_broadcast<op, InputScalarType, InputVectorType, uint16x8_t>(wrapper::vloadq((non_broadcast_input_ptr + x)), broadcast_value, reorder);
+        wrapper::vstore(output_ptr + x, wrapper::vmovn(a));
+    }
+    return x;
+}
+
+template <ComparisonOperation op, typename InputScalarType, typename InputVectorType>
+inline int elementwise_comp_op_broadcast_32_loop(int window_start_x, int window_end_x, int window_step_x,
+                                                 const InputScalarType *non_broadcast_input_ptr, const InputScalarType &broadcast_value, uint8_t *output_ptr, const bool reorder)
+{
+    int x = window_start_x;
+    for(; x <= (window_end_x - window_step_x); x += window_step_x)
+    {
+        const auto a = elementwise_comp_op_broadcast<op, InputScalarType, InputVectorType, uint32x4_t>(wrapper::vloadq(non_broadcast_input_ptr + x), broadcast_value, reorder);
+        const auto b = elementwise_comp_op_broadcast<op, InputScalarType, InputVectorType, uint32x4_t>(wrapper::vloadq(non_broadcast_input_ptr + x + 4), broadcast_value, reorder);
+        wrapper::vstore(output_ptr + x, wrapper::vmovn(wrapper::vcombine(wrapper::vmovn(a), wrapper::vmovn(b))));
+    }
+    if(x <= window_end_x - 4)
+    {
+        const auto a = elementwise_comp_op_broadcast<op, InputScalarType, InputVectorType, uint32x4_t>(wrapper::vloadq((non_broadcast_input_ptr + x)), broadcast_value, reorder);
+        for(int i = 0; i < 4; i++)
+        {
+            *(output_ptr + x + i) = wrapper::vgetlane(a, i);
+        }
+        x = +4;
+    }
+    return x;
+}
+
+template <ComparisonOperation op, typename InputScalarType, typename InputVectorType>
+inline int elementwise_comp_op_8_loop(int window_start_x, int window_end_x, int window_step_x,
+                                      const InputScalarType *input1_ptr, const InputScalarType *input2_ptr, uint8_t *output_ptr)
+{
+    int x = window_start_x;
+    for(; x <= (window_end_x - window_step_x); x += window_step_x)
+    {
+        const auto a   = wrapper::vloadq(input1_ptr + x);
+        const auto b   = wrapper::vloadq(input2_ptr + x);
+        const auto res = elementwise_comp_op<op, InputVectorType, uint8x16_t>(a, b);
+        wrapper::vstore(output_ptr + x, res);
+    }
+    return x;
+}
+
+template <ComparisonOperation op, typename InputScalarType, typename InputVectorType>
+inline int elementwise_comp_op_16_loop(int window_start_x, int window_end_x, int window_step_x,
+                                       const InputScalarType *input1_ptr, const InputScalarType *input2_ptr, uint8_t *output_ptr)
+{
+    int x = window_start_x;
+    for(; x <= (window_end_x - window_step_x); x += window_step_x)
+    {
+        const auto a   = wrapper::vloadq(input1_ptr + x);
+        const auto b   = wrapper::vloadq(input2_ptr + x);
+        const auto res = elementwise_comp_op<op, InputVectorType, uint16x8_t>(a, b);
+        wrapper::vstore(output_ptr + x, wrapper::vmovn(res));
+    }
+    return x;
+}
+
+template <ComparisonOperation op, typename InputScalarType, typename InputVectorType>
+inline int elementwise_comp_op_32_loop(int window_start_x, int window_end_x, int window_step_x,
+                                       const InputScalarType *input1_ptr, const InputScalarType *input2_ptr, uint8_t *output_ptr)
+{
+    int x = window_start_x;
+    for(; x <= (window_end_x - window_step_x); x += window_step_x)
+    {
+        auto       a    = wrapper::vloadq(input1_ptr + x);
+        auto       b    = wrapper::vloadq(input2_ptr + x);
+        const auto res  = elementwise_comp_op<op, InputVectorType, uint32x4_t>(a, b);
+        a               = wrapper::vloadq(input1_ptr + x + 4);
+        b               = wrapper::vloadq(input2_ptr + x + 4);
+        const auto res2 = elementwise_comp_op<op, InputVectorType, uint32x4_t>(a, b);
+        wrapper::vstore(output_ptr + x, wrapper::vmovn(wrapper::vcombine(wrapper::vmovn(res), wrapper::vmovn(res2))));
+    }
+    if(x <= window_end_x - 4)
+    {
+        const auto a   = wrapper::vloadq(input1_ptr + x);
+        const auto b   = wrapper::vloadq(input2_ptr + x);
+        const auto res = elementwise_comp_op<op, InputVectorType, uint32x4_t>(a, b);
+        for(int i = 0; i < 4; i++)
+        {
+            *(output_ptr + x + i) = wrapper::vgetlane(res, i);
+        }
+        x = +4;
+    }
+    return x;
+}
+
+template <ComparisonOperation op, typename InputScalarType, typename InputVectorType>
+void elementwise_comp_op_8(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window)
+{
+    elementwise_op<InputScalarType, uint8_t, InputVectorType>(in1, in2, out, window,
+                                                              &elementwise_comp_op_scalar<op, InputScalarType>,
+                                                              &elementwise_comp_op_broadcast_8_loop<op, InputScalarType, InputVectorType>,
+                                                              &elementwise_comp_op_8_loop<op, InputScalarType, InputVectorType>);
+}
+
+template <ComparisonOperation op, typename InputScalarType, typename InputVectorType>
+void elementwise_comp_op_16(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window)
+{
+    elementwise_op<InputScalarType, uint8_t, InputVectorType>(in1, in2, out, window,
+                                                              &elementwise_comp_op_scalar<op, InputScalarType>,
+                                                              &elementwise_comp_op_broadcast_16_loop<op, InputScalarType, InputVectorType>,
+                                                              &elementwise_comp_op_16_loop<op, InputScalarType, InputVectorType>);
+}
+
+template <ComparisonOperation op, typename InputScalarType, typename InputVectorType>
+void elementwise_comp_op_32(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window)
+{
+    elementwise_op<InputScalarType, uint8_t, InputVectorType>(in1, in2, out, window,
+                                                              &elementwise_comp_op_scalar<op, InputScalarType>,
+                                                              &elementwise_comp_op_broadcast_32_loop<op, InputScalarType, InputVectorType>,
+                                                              &elementwise_comp_op_32_loop<op, InputScalarType, InputVectorType>);
+}
+} // namesapce cpu
+} // namespace arm_compute
+
+#endif /* SRC_CORE_NEON_KERNELS_ELEMENTWISE_LIST_H */
\ No newline at end of file
diff --git a/src/core/cpu/kernels/elementwise/neon/elementwise_quantized_list.h b/src/core/cpu/kernels/elementwise/neon/elementwise_quantized_list.h
new file mode 100644
index 0000000000..1ff4632f5c
--- /dev/null
+++ b/src/core/cpu/kernels/elementwise/neon/elementwise_quantized_list.h
@@ -0,0 +1,654 @@
+/*
+ * 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.
+ */
+#ifndef SRC_CORE_NEON_KERNELS_ELEMENTWISE_QUANTIZED_LIST_H
+#define SRC_CORE_NEON_KERNELS_ELEMENTWISE_QUANTIZED_LIST_H
+
+#include "src/core/cpu/kernels/elementwise/neon/elementwise_list.h"
+
+namespace arm_compute
+{
+namespace cpu
+{
+float32x4x4_t load_quantized(const uint8_t *input1_ptr, const int32x4_t &offset, const float32x4_t &scale)
+{
+    qasymm8x16_t        x = vld1q_u8(input1_ptr);
+    const float32x4x4_t out =
+    {
+        {
+            vmulq_f32(vcvtq_f32_s32(vsubq_s32(vreinterpretq_s32_u32(vmovl_u16(vget_low_u16(vmovl_u8(vget_low_u8(x))))), offset)), scale),
+            vmulq_f32(vcvtq_f32_s32(vsubq_s32(vreinterpretq_s32_u32(vmovl_u16(vget_high_u16(vmovl_u8(vget_low_u8(x))))), offset)), scale),
+            vmulq_f32(vcvtq_f32_s32(vsubq_s32(vreinterpretq_s32_u32(vmovl_u16(vget_low_u16(vmovl_u8(vget_high_u8(x))))), offset)), scale),
+            vmulq_f32(vcvtq_f32_s32(vsubq_s32(vreinterpretq_s32_u32(vmovl_u16(vget_high_u16(vmovl_u8(vget_high_u8(x))))), offset)), scale),
+        }
+    };
+    return out;
+}
+
+float32x4x4_t load_quantized_signed(const int8_t *input1_ptr, const int32x4_t &offset, const float32x4_t &scale)
+{
+    qasymm8x16_signed_t x = vld1q_s8(input1_ptr);
+    const float32x4x4_t out =
+    {
+        {
+            vmulq_f32(vcvtq_f32_s32(vsubq_s32(vmovl_s16(vget_low_s16(vmovl_s8(vget_low_s8(x)))), offset)), scale),
+            vmulq_f32(vcvtq_f32_s32(vsubq_s32(vmovl_s16(vget_high_s16(vmovl_s8(vget_low_s8(x)))), offset)), scale),
+            vmulq_f32(vcvtq_f32_s32(vsubq_s32(vmovl_s16(vget_low_s16(vmovl_s8(vget_high_s8(x)))), offset)), scale),
+            vmulq_f32(vcvtq_f32_s32(vsubq_s32(vmovl_s16(vget_high_s16(vmovl_s8(vget_high_s8(x)))), offset)), scale),
+        }
+    };
+    return out;
+}
+
+void store_quantized(uint8_t *output_ptr, const uint32x4x4_t &out)
+{
+    const uint8x8_t pa = vqmovn_u16(vcombine_u16(vqmovn_u32(out.val[0]), vqmovn_u32(out.val[1])));
+    const uint8x8_t pb = vqmovn_u16(vcombine_u16(vqmovn_u32(out.val[2]), vqmovn_u32(out.val[3])));
+    vst1q_u8(output_ptr, vcombine_u8(pa, pb));
+}
+
+void store_quantized(uint8_t *output_ptr, const int32x4x4_t &out)
+{
+    const uint8x8_t pa = vqmovun_s16(vcombine_s16(vqmovn_s32(out.val[0]), vqmovn_s32(out.val[1])));
+    const uint8x8_t pb = vqmovun_s16(vcombine_s16(vqmovn_s32(out.val[2]), vqmovn_s32(out.val[3])));
+    vst1q_u8(output_ptr, vcombine_u8(pa, pb));
+}
+
+void store_quantized(uint8_t *output_ptr, const float32x4x4_t &rf, const float32x4_t &offset, const float32x4_t &invscale)
+{
+    int32x4x4_t out =
+    {
+        {
+            vcvtq_s32_f32(vmlaq_f32(offset, rf.val[0], invscale)),
+            vcvtq_s32_f32(vmlaq_f32(offset, rf.val[1], invscale)),
+            vcvtq_s32_f32(vmlaq_f32(offset, rf.val[2], invscale)),
+            vcvtq_s32_f32(vmlaq_f32(offset, rf.val[3], invscale)),
+        }
+    };
+    store_quantized(output_ptr, out);
+}
+
+void store_quantized_signed(int8_t *output_ptr, const int32x4x4_t &out)
+{
+    const int8x8_t pa = vqmovn_s16(vcombine_s16(vqmovn_s32(out.val[0]), vqmovn_s32(out.val[1])));
+    const int8x8_t pb = vqmovn_s16(vcombine_s16(vqmovn_s32(out.val[2]), vqmovn_s32(out.val[3])));
+    vst1q_s8(output_ptr, vcombine_s8(pa, pb));
+}
+
+void store_quantized_signed(int8_t *output_ptr, const float32x4x4_t &rf, const float32x4_t &offset, const float32x4_t &invscale)
+{
+    int32x4x4_t out =
+    {
+        {
+            vcvtq_s32_f32(vmlaq_f32(offset, rf.val[0], invscale)),
+            vcvtq_s32_f32(vmlaq_f32(offset, rf.val[1], invscale)),
+            vcvtq_s32_f32(vmlaq_f32(offset, rf.val[2], invscale)),
+            vcvtq_s32_f32(vmlaq_f32(offset, rf.val[3], invscale)),
+        }
+    };
+    store_quantized_signed(output_ptr, out);
+}
+
+template <ArithmeticOperation op>
+inline uint8_t elementwise_arithm_op_quantized_scalar(const float &a, const float &b, UniformQuantizationInfo qinfo)
+{
+    return quantize_qasymm8(elementwise_arithm_op_scalar<op>(a, b), qinfo);
+}
+
+template <ArithmeticOperation op>
+inline int8_t elementwise_arithm_op_quantized_signed_scalar(const float &a, const float &b, UniformQuantizationInfo qinfo)
+{
+    return quantize_qasymm8_signed(elementwise_arithm_op_scalar<op>(a, b), qinfo);
+}
+
+template <ArithmeticOperation op>
+inline float32x4x4_t elementwise_arithm_op(const float32x4x4_t &a, const float32x4x4_t &b)
+{
+    using neon_vector_float = wrapper::traits::neon_vector<float, 4>;
+    float32x4x4_t out =
+    {
+        {
+            elementwise_arithm_op<op, neon_vector_float>(a.val[0], b.val[0]),
+            elementwise_arithm_op<op, neon_vector_float>(a.val[1], b.val[1]),
+            elementwise_arithm_op<op, neon_vector_float>(a.val[2], b.val[2]),
+            elementwise_arithm_op<op, neon_vector_float>(a.val[3], b.val[3]),
+        }
+    };
+    return out;
+}
+
+template <ComparisonOperation op>
+inline uint8_t elementwise_comp_op_quantized_scalar(const float &a, const float &b, UniformQuantizationInfo qinfo)
+{
+    ARM_COMPUTE_UNUSED(qinfo);
+    return elementwise_comp_op_scalar<op>(a, b);
+}
+
+template <ComparisonOperation op>
+inline uint32x4x4_t elementwise_comp_op(const float32x4x4_t &a, const float32x4x4_t &b)
+{
+    uint32x4x4_t out =
+    {
+        {
+            elementwise_comp_op<op, float32x4_t, uint32x4_t>(a.val[0], b.val[0]),
+            elementwise_comp_op<op, float32x4_t, uint32x4_t>(a.val[1], b.val[1]),
+            elementwise_comp_op<op, float32x4_t, uint32x4_t>(a.val[2], b.val[2]),
+            elementwise_comp_op<op, float32x4_t, uint32x4_t>(a.val[3], b.val[3])
+        }
+    };
+    return out;
+}
+
+template <ArithmeticOperation op>
+inline int elementwise_arithm_op_quantized_loop(int window_start_x, int window_end_x, int window_step_x,
+                                                const uint8_t *input1_ptr, const uint8_t *input2_ptr, uint8_t *output_ptr,
+                                                int32x4_t voffset1, int32x4_t voffset2, float32x4_t vscale1, float32x4_t vscale2,
+                                                float32x4_t voffseto, float32x4_t invvscaleo)
+{
+    int x = window_start_x;
+    for(; x <= (window_end_x - window_step_x); x += window_step_x)
+    {
+        // Get inputs and compute output
+        const float32x4x4_t af = load_quantized(input1_ptr + x, voffset1, vscale1);
+        const float32x4x4_t bf = load_quantized(input2_ptr + x, voffset2, vscale2);
+        const float32x4x4_t rf = elementwise_arithm_op<op>(af, bf);
+        store_quantized(output_ptr + x, rf, voffseto, invvscaleo);
+    }
+    return x;
+}
+
+template <ArithmeticOperation op>
+inline int elementwise_arithm_op_quantized_singed_loop(int window_start_x, int window_end_x, int window_step_x,
+                                                       const int8_t *input1_ptr, const int8_t *input2_ptr, int8_t *output_ptr,
+                                                       int32x4_t voffset1, int32x4_t voffset2, float32x4_t vscale1, float32x4_t vscale2,
+                                                       float32x4_t voffseto, float32x4_t invvscaleo)
+{
+    int x = window_start_x;
+    for(; x <= (window_end_x - window_step_x); x += window_step_x)
+    {
+        // Get inputs and compute output
+        const float32x4x4_t af = load_quantized_signed(input1_ptr + x, voffset1, vscale1);
+        const float32x4x4_t bf = load_quantized_signed(input2_ptr + x, voffset2, vscale2);
+        const float32x4x4_t rf = elementwise_arithm_op<op>(af, bf);
+        store_quantized_signed(output_ptr + x, rf, voffseto, invvscaleo);
+    }
+    return x;
+}
+
+template <ArithmeticOperation op>
+inline int elementwise_arithm_op_quantized_broadcast_loop(int window_start_x, int window_end_x, int window_step_x,
+                                                          const uint8_t *non_broadcast_input_ptr, float32x4x4_t broadcast_vector, uint8_t *output_ptr,
+                                                          int32x4_t voffset_non_broadcast, float32x4_t vscale_non_broadcast,
+                                                          float32x4_t voffseto, float32x4_t invvscaleo, bool reorder)
+{
+    int x = window_start_x;
+    for(; x <= (window_end_x - window_step_x); x += window_step_x)
+    {
+        const float32x4x4_t af = load_quantized(non_broadcast_input_ptr + x, voffset_non_broadcast, vscale_non_broadcast);
+        const float32x4x4_t rf = elementwise_arithm_op<op>(reorder ? broadcast_vector : af, reorder ? af : broadcast_vector);
+        store_quantized(output_ptr + x, rf, voffseto, invvscaleo);
+    }
+    return x;
+}
+template <ArithmeticOperation op>
+inline int elementwise_arithm_op_quantized_signed_broadcast_loop(int window_start_x, int window_end_x, int window_step_x,
+                                                                 const int8_t *non_broadcast_input_ptr, float32x4x4_t broadcast_vector, int8_t *output_ptr,
+                                                                 int32x4_t voffset_non_broadcast, float32x4_t vscale_non_broadcast,
+                                                                 float32x4_t voffseto, float32x4_t invvscaleo, bool reorder)
+{
+    int x = window_start_x;
+    for(; x <= (window_end_x - window_step_x); x += window_step_x)
+    {
+        const float32x4x4_t af = load_quantized_signed(non_broadcast_input_ptr + x, voffset_non_broadcast, vscale_non_broadcast);
+        const float32x4x4_t rf = elementwise_arithm_op<op>(reorder ? broadcast_vector : af, reorder ? af : broadcast_vector);
+        store_quantized_signed(output_ptr + x, rf, voffseto, invvscaleo);
+    }
+    return x;
+}
+
+template <ComparisonOperation op>
+inline int elementwise_comp_op_quantized_loop(int window_start_x, int window_end_x, int window_step_x,
+                                              const uint8_t *input1_ptr, const uint8_t *input2_ptr, uint8_t *output_ptr,
+                                              int32x4_t voffset1, int32x4_t voffset2, float32x4_t vscale1, float32x4_t vscale2,
+                                              float32x4_t voffseto, float32x4_t invvscaleo)
+{
+    ARM_COMPUTE_UNUSED(voffseto, invvscaleo);
+    int x = window_start_x;
+    for(; x <= (window_end_x - window_step_x); x += window_step_x)
+    {
+        const float32x4x4_t af = load_quantized(input1_ptr + x, voffset1, vscale1);
+        const float32x4x4_t bf = load_quantized(input2_ptr + x, voffset2, vscale2);
+        const uint32x4x4_t  rf = elementwise_comp_op<op>(af, bf);
+        store_quantized(output_ptr + x, rf);
+    }
+    return x;
+}
+
+template <ComparisonOperation op>
+inline int elementwise_comp_op_quantized_signed_loop(int window_start_x, int window_end_x, int window_step_x,
+                                                     const int8_t *input1_ptr, const int8_t *input2_ptr, uint8_t *output_ptr,
+                                                     int32x4_t voffset1, int32x4_t voffset2, float32x4_t vscale1, float32x4_t vscale2,
+                                                     float32x4_t voffseto, float32x4_t invvscaleo)
+{
+    ARM_COMPUTE_UNUSED(voffseto, invvscaleo);
+    int x = window_start_x;
+    for(; x <= (window_end_x - window_step_x); x += window_step_x)
+    {
+        const float32x4x4_t af = load_quantized_signed(input1_ptr + x, voffset1, vscale1);
+        const float32x4x4_t bf = load_quantized_signed(input2_ptr + x, voffset2, vscale2);
+        const uint32x4x4_t  rf = elementwise_comp_op<op>(af, bf);
+        store_quantized(output_ptr + x, rf);
+    }
+    return x;
+}
+
+template <ComparisonOperation op>
+inline int elementwise_comp_op_quantized_broadcast_loop(int window_start_x, int window_end_x, int window_step_x,
+                                                        const uint8_t *non_broadcast_input_ptr, float32x4x4_t broadcast_vector, uint8_t *output_ptr,
+                                                        int32x4_t voffset_non_broadcast, float32x4_t vscale_non_broadcast,
+                                                        float32x4_t voffseto, float32x4_t invvscaleo, bool reorder)
+{
+    ARM_COMPUTE_UNUSED(voffseto, invvscaleo);
+    int x = window_start_x;
+    for(; x <= (window_end_x - window_step_x); x += window_step_x)
+    {
+        const float32x4x4_t af = load_quantized(non_broadcast_input_ptr + x, voffset_non_broadcast, vscale_non_broadcast);
+        const uint32x4x4_t  rf = elementwise_comp_op<op>(reorder ? broadcast_vector : af, reorder ? af : broadcast_vector);
+        store_quantized(output_ptr + x, rf);
+    }
+    return x;
+}
+
+template <ComparisonOperation op>
+inline int elementwise_comp_op_quantized_signed_broadcast_loop(int window_start_x, int window_end_x, int window_step_x,
+                                                               const int8_t *non_broadcast_input_ptr, float32x4x4_t broadcast_vector, uint8_t *output_ptr,
+                                                               int32x4_t voffset_non_broadcast, float32x4_t vscale_non_broadcast,
+                                                               float32x4_t voffseto, float32x4_t invvscaleo, bool reorder)
+{
+    ARM_COMPUTE_UNUSED(voffseto, invvscaleo);
+    int x = window_start_x;
+    for(; x <= (window_end_x - window_step_x); x += window_step_x)
+    {
+        const float32x4x4_t af = load_quantized_signed(non_broadcast_input_ptr + x, voffset_non_broadcast, vscale_non_broadcast);
+        const uint32x4x4_t  rf = elementwise_comp_op<op>(reorder ? broadcast_vector : af, reorder ? af : broadcast_vector);
+        store_quantized(output_ptr + x, rf);
+    }
+    return x;
+}
+
+void elementwise_op_quantized(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window,
+                              uint8_t (*scalar_func)(const float &, const float &, UniformQuantizationInfo),
+                              int (*broadcast_func)(int, int, int, const uint8_t *, float32x4x4_t, uint8_t *, int32x4_t, float32x4_t,
+                                                    float32x4_t, float32x4_t, const bool),
+                              int (*neon_func)(int, int, int, const uint8_t *, const uint8_t *, uint8_t *,
+                                               int32x4_t, int32x4_t, float32x4_t, float32x4_t,
+                                               float32x4_t, float32x4_t))
+{
+    // Create input windows
+    Window input1_win = window.broadcast_if_dimension_le_one(in1->info()->tensor_shape());
+    Window input2_win = window.broadcast_if_dimension_le_one(in2->info()->tensor_shape());
+
+    // Clear X Dimension on execution window as we handle manually
+    Window win = window;
+    win.set(Window::DimX, Window::Dimension(0, 1, 1));
+
+    const int  window_step_x         = 16;
+    const auto window_start_x        = static_cast<int>(window.x().start());
+    const auto window_end_x          = static_cast<int>(window.x().end());
+    const bool is_broadcast_across_x = in1->info()->tensor_shape().x() != in2->info()->tensor_shape().x();
+
+    const UniformQuantizationInfo output_qinfo = out->info()->quantization_info().uniform();
+
+    // Output quantization info (add 0.5 to round toward the nearest integer - 0.5 rounds away from zero)
+    const float32x4_t voffseto   = vdupq_n_f32(output_qinfo.offset + 0.5f);
+    const float32x4_t invvscaleo = vdupq_n_f32(1.f / output_qinfo.scale);
+
+    if(is_broadcast_across_x)
+    {
+        // Select the broadcast input on the X axis
+        const bool     is_broadcast_input_2 = input2_win.x().step() == 0;
+        Window         broadcast_win        = is_broadcast_input_2 ? input2_win : input1_win;
+        Window         non_broadcast_win    = !is_broadcast_input_2 ? input2_win : input1_win;
+        const ITensor *broadcast_tensor     = is_broadcast_input_2 ? in2 : in1;
+        const ITensor *non_broadcast_tensor = !is_broadcast_input_2 ? in2 : in1;
+
+        const UniformQuantizationInfo broadcast_qinfo     = broadcast_tensor->info()->quantization_info().uniform();
+        const UniformQuantizationInfo non_broadcast_qinfo = non_broadcast_tensor->info()->quantization_info().uniform();
+
+        const int32x4_t   voffset_non_broadcast = vdupq_n_s32(non_broadcast_qinfo.offset);
+        const float32x4_t vscale_non_broadcast  = vdupq_n_f32(non_broadcast_qinfo.scale);
+
+        // Clear X Dimension on execution window as we handle manually
+        non_broadcast_win.set(Window::DimX, Window::Dimension(0, 1, 1));
+
+        Iterator broadcast_input(broadcast_tensor, broadcast_win);
+        Iterator non_broadcast_input(non_broadcast_tensor, non_broadcast_win);
+        Iterator output(out, win);
+
+        execute_window_loop(win, [&](const Coordinates &)
+        {
+            const auto non_broadcast_input_ptr = reinterpret_cast<const uint8_t *>(non_broadcast_input.ptr());
+            const auto output_ptr              = reinterpret_cast<uint8_t *>(output.ptr());
+
+            const uint8_t       broadcast_value  = *reinterpret_cast<const uint8_t *>(broadcast_input.ptr());
+            const float32x4x4_t broadcast_vector = vdequantize(vdupq_n_u8(broadcast_value), broadcast_qinfo);
+
+            int x = (*broadcast_func)(window_start_x, window_end_x, window_step_x, non_broadcast_input_ptr, broadcast_vector, output_ptr,
+                                      voffset_non_broadcast, vscale_non_broadcast, voffseto, invvscaleo, !is_broadcast_input_2);
+            for(; x < window_end_x; ++x)
+            {
+                const float afs   = dequantize_qasymm8(*(non_broadcast_input_ptr + x), non_broadcast_qinfo);
+                const float bfs   = dequantize_qasymm8(broadcast_value, broadcast_qinfo);
+                *(output_ptr + x) = (*scalar_func)(!is_broadcast_input_2 ? bfs : afs, !is_broadcast_input_2 ? afs : bfs, output_qinfo);
+            }
+        },
+        broadcast_input, non_broadcast_input, output);
+    }
+    else
+    {
+        const UniformQuantizationInfo input1_qinfo = in1->info()->quantization_info().uniform();
+        const UniformQuantizationInfo input2_qinfo = in2->info()->quantization_info().uniform();
+
+        // Input1 quantization info
+        const int32x4_t   voffset1 = vdupq_n_s32(input1_qinfo.offset);
+        const float32x4_t vscale1  = vdupq_n_f32(input1_qinfo.scale);
+
+        // Input2 quantization info
+        const int32x4_t   voffset2 = vdupq_n_s32(input2_qinfo.offset);
+        const float32x4_t vscale2  = vdupq_n_f32(input2_qinfo.scale);
+
+        // Clear X Dimension on execution window as we handle manually
+        input1_win.set(Window::DimX, Window::Dimension(0, 1, 1));
+        input2_win.set(Window::DimX, Window::Dimension(0, 1, 1));
+
+        Iterator input1(in1, input1_win);
+        Iterator input2(in2, input2_win);
+        Iterator output(out, win);
+
+        execute_window_loop(win, [&](const Coordinates &)
+        {
+            const auto input1_ptr = reinterpret_cast<const 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 = (*neon_func)(window_start_x, window_end_x, window_step_x, input1_ptr, input2_ptr, output_ptr, voffset1, voffset2,
+                                 vscale1, vscale2, voffseto, invvscaleo);
+            for(; x < window_end_x; ++x)
+            {
+                const float afs   = dequantize_qasymm8(*(input1_ptr + x), input1_qinfo);
+                const float bfs   = dequantize_qasymm8(*(input2_ptr + x), input2_qinfo);
+                *(output_ptr + x) = (*scalar_func)(afs, bfs, output_qinfo);
+            }
+        },
+        input1, input2, output);
+    }
+}
+
+void elementwise_comp_quantized_signed(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window,
+                                       uint8_t (*scalar_func)(const float &, const float &, UniformQuantizationInfo),
+                                       int (*broadcast_func)(int, int, int, const int8_t *, float32x4x4_t, uint8_t *, int32x4_t, float32x4_t,
+                                                             float32x4_t, float32x4_t, const bool),
+                                       int (*neon_func)(int, int, int, const int8_t *, const int8_t *, uint8_t *,
+                                                        int32x4_t, int32x4_t, float32x4_t, float32x4_t,
+                                                        float32x4_t, float32x4_t))
+{
+    // Create input windows
+    Window input1_win = window.broadcast_if_dimension_le_one(in1->info()->tensor_shape());
+    Window input2_win = window.broadcast_if_dimension_le_one(in2->info()->tensor_shape());
+
+    // Clear X Dimension on execution window as we handle manually
+    Window win = window;
+    win.set(Window::DimX, Window::Dimension(0, 1, 1));
+
+    const int  window_step_x         = 16;
+    const auto window_start_x        = static_cast<int>(window.x().start());
+    const auto window_end_x          = static_cast<int>(window.x().end());
+    const bool is_broadcast_across_x = in1->info()->tensor_shape().x() != in2->info()->tensor_shape().x();
+
+    const UniformQuantizationInfo output_qinfo = out->info()->quantization_info().uniform();
+
+    const float32x4_t voffseto   = vdupq_n_f32(output_qinfo.offset);
+    const float32x4_t invvscaleo = vdupq_n_f32(1.f / output_qinfo.scale);
+
+    if(is_broadcast_across_x)
+    {
+        // Select the broadcast input on the X axis
+        const bool     is_broadcast_input_2 = input2_win.x().step() == 0;
+        Window         broadcast_win        = is_broadcast_input_2 ? input2_win : input1_win;
+        Window         non_broadcast_win    = !is_broadcast_input_2 ? input2_win : input1_win;
+        const ITensor *broadcast_tensor     = is_broadcast_input_2 ? in2 : in1;
+        const ITensor *non_broadcast_tensor = !is_broadcast_input_2 ? in2 : in1;
+
+        const UniformQuantizationInfo broadcast_qinfo     = broadcast_tensor->info()->quantization_info().uniform();
+        const UniformQuantizationInfo non_broadcast_qinfo = non_broadcast_tensor->info()->quantization_info().uniform();
+
+        const int32x4_t   voffset_non_broadcast = vdupq_n_s32(non_broadcast_qinfo.offset);
+        const float32x4_t vscale_non_broadcast  = vdupq_n_f32(non_broadcast_qinfo.scale);
+
+        // Clear X Dimension on execution window as we handle manually
+        non_broadcast_win.set(Window::DimX, Window::Dimension(0, 1, 1));
+
+        Iterator broadcast_input(broadcast_tensor, broadcast_win);
+        Iterator non_broadcast_input(non_broadcast_tensor, non_broadcast_win);
+        Iterator output(out, win);
+
+        execute_window_loop(win, [&](const Coordinates &)
+        {
+            const auto non_broadcast_input_ptr = reinterpret_cast<const int8_t *>(non_broadcast_input.ptr());
+            const auto output_ptr              = reinterpret_cast<uint8_t *>(output.ptr());
+
+            const int8_t        broadcast_value  = *reinterpret_cast<const int8_t *>(broadcast_input.ptr());
+            const float32x4x4_t broadcast_vector = vdequantize(vdupq_n_s8(broadcast_value), broadcast_qinfo);
+
+            int x = (*broadcast_func)(window_start_x, window_end_x, window_step_x, non_broadcast_input_ptr, broadcast_vector, output_ptr,
+                                      voffset_non_broadcast, vscale_non_broadcast, voffseto, invvscaleo, !is_broadcast_input_2);
+            for(; x < window_end_x; ++x)
+            {
+                const float afs   = dequantize_qasymm8_signed(*(non_broadcast_input_ptr + x), non_broadcast_qinfo);
+                const float bfs   = dequantize_qasymm8_signed(broadcast_value, broadcast_qinfo);
+                *(output_ptr + x) = (*scalar_func)(!is_broadcast_input_2 ? bfs : afs, !is_broadcast_input_2 ? afs : bfs, output_qinfo);
+            }
+        },
+        broadcast_input, non_broadcast_input, output);
+    }
+    else
+    {
+        const UniformQuantizationInfo input1_qinfo = in1->info()->quantization_info().uniform();
+        const UniformQuantizationInfo input2_qinfo = in2->info()->quantization_info().uniform();
+
+        // Input1 quantization info
+        const int32x4_t   voffset1 = vdupq_n_s32(input1_qinfo.offset);
+        const float32x4_t vscale1  = vdupq_n_f32(input1_qinfo.scale);
+
+        // Input2 quantization info
+        const int32x4_t   voffset2 = vdupq_n_s32(input2_qinfo.offset);
+        const float32x4_t vscale2  = vdupq_n_f32(input2_qinfo.scale);
+
+        // Clear X Dimension on execution window as we handle manually
+        input1_win.set(Window::DimX, Window::Dimension(0, 1, 1));
+        input2_win.set(Window::DimX, Window::Dimension(0, 1, 1));
+
+        Iterator input1(in1, input1_win);
+        Iterator input2(in2, input2_win);
+        Iterator output(out, win);
+
+        execute_window_loop(win, [&](const Coordinates &)
+        {
+            const auto input1_ptr = reinterpret_cast<const int8_t *>(input1.ptr());
+            const auto input2_ptr = reinterpret_cast<const int8_t *>(input2.ptr());
+            const auto output_ptr = reinterpret_cast<uint8_t *>(output.ptr());
+
+            int x = (*neon_func)(window_start_x, window_end_x, window_step_x, input1_ptr, input2_ptr, output_ptr, voffset1, voffset2,
+                                 vscale1, vscale2, voffseto, invvscaleo);
+            for(; x < window_end_x; ++x)
+            {
+                const float afs   = dequantize_qasymm8_signed(*(input1_ptr + x), input1_qinfo);
+                const float bfs   = dequantize_qasymm8_signed(*(input2_ptr + x), input2_qinfo);
+                *(output_ptr + x) = (*scalar_func)(afs, bfs, output_qinfo);
+            }
+        },
+        input1, input2, output);
+    }
+}
+
+void elementwise_op_quantized_signed(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window,
+                                     int8_t (*scalar_func)(const float &, const float &, UniformQuantizationInfo),
+                                     int (*broadcast_func)(int, int, int, const int8_t *, float32x4x4_t, int8_t *, int32x4_t, float32x4_t,
+                                                           float32x4_t, float32x4_t, const bool),
+                                     int (*neon_func)(int, int, int, const int8_t *, const int8_t *, int8_t *,
+                                                      int32x4_t, int32x4_t, float32x4_t, float32x4_t,
+                                                      float32x4_t, float32x4_t))
+{
+    // Create input windows
+    Window input1_win = window.broadcast_if_dimension_le_one(in1->info()->tensor_shape());
+    Window input2_win = window.broadcast_if_dimension_le_one(in2->info()->tensor_shape());
+
+    // Clear X Dimension on execution window as we handle manually
+    Window win = window;
+    win.set(Window::DimX, Window::Dimension(0, 1, 1));
+
+    const int  window_step_x         = 16;
+    const auto window_start_x        = static_cast<int>(window.x().start());
+    const auto window_end_x          = static_cast<int>(window.x().end());
+    const bool is_broadcast_across_x = in1->info()->tensor_shape().x() != in2->info()->tensor_shape().x();
+
+    const UniformQuantizationInfo output_qinfo = out->info()->quantization_info().uniform();
+
+    const float32x4_t voffseto   = vdupq_n_f32(output_qinfo.offset);
+    const float32x4_t invvscaleo = vdupq_n_f32(1.f / output_qinfo.scale);
+
+    if(is_broadcast_across_x)
+    {
+        // Select the broadcast input on the X axis
+        const bool     is_broadcast_input_2 = input2_win.x().step() == 0;
+        Window         broadcast_win        = is_broadcast_input_2 ? input2_win : input1_win;
+        Window         non_broadcast_win    = !is_broadcast_input_2 ? input2_win : input1_win;
+        const ITensor *broadcast_tensor     = is_broadcast_input_2 ? in2 : in1;
+        const ITensor *non_broadcast_tensor = !is_broadcast_input_2 ? in2 : in1;
+
+        const UniformQuantizationInfo broadcast_qinfo     = broadcast_tensor->info()->quantization_info().uniform();
+        const UniformQuantizationInfo non_broadcast_qinfo = non_broadcast_tensor->info()->quantization_info().uniform();
+
+        const int32x4_t   voffset_non_broadcast = vdupq_n_s32(non_broadcast_qinfo.offset);
+        const float32x4_t vscale_non_broadcast  = vdupq_n_f32(non_broadcast_qinfo.scale);
+
+        // Clear X Dimension on execution window as we handle manually
+        non_broadcast_win.set(Window::DimX, Window::Dimension(0, 1, 1));
+
+        Iterator broadcast_input(broadcast_tensor, broadcast_win);
+        Iterator non_broadcast_input(non_broadcast_tensor, non_broadcast_win);
+        Iterator output(out, win);
+
+        execute_window_loop(win, [&](const Coordinates &)
+        {
+            const auto non_broadcast_input_ptr = reinterpret_cast<const int8_t *>(non_broadcast_input.ptr());
+            const auto output_ptr              = reinterpret_cast<int8_t *>(output.ptr());
+
+            const int8_t        broadcast_value  = *reinterpret_cast<const int8_t *>(broadcast_input.ptr());
+            const float32x4x4_t broadcast_vector = vdequantize(vdupq_n_s8(broadcast_value), broadcast_qinfo);
+
+            int x = (*broadcast_func)(window_start_x, window_end_x, window_step_x, non_broadcast_input_ptr, broadcast_vector, output_ptr,
+                                      voffset_non_broadcast, vscale_non_broadcast, voffseto, invvscaleo, !is_broadcast_input_2);
+            for(; x < window_end_x; ++x)
+            {
+                const float afs   = dequantize_qasymm8_signed(*(non_broadcast_input_ptr + x), non_broadcast_qinfo);
+                const float bfs   = dequantize_qasymm8_signed(broadcast_value, broadcast_qinfo);
+                *(output_ptr + x) = (*scalar_func)(!is_broadcast_input_2 ? bfs : afs, !is_broadcast_input_2 ? afs : bfs, output_qinfo);
+            }
+        },
+        broadcast_input, non_broadcast_input, output);
+    }
+    else
+    {
+        const UniformQuantizationInfo input1_qinfo = in1->info()->quantization_info().uniform();
+        const UniformQuantizationInfo input2_qinfo = in2->info()->quantization_info().uniform();
+
+        // Input1 quantization info
+        const int32x4_t   voffset1 = vdupq_n_s32(input1_qinfo.offset);
+        const float32x4_t vscale1  = vdupq_n_f32(input1_qinfo.scale);
+
+        // Input2 quantization info
+        const int32x4_t   voffset2 = vdupq_n_s32(input2_qinfo.offset);
+        const float32x4_t vscale2  = vdupq_n_f32(input2_qinfo.scale);
+
+        // Clear X Dimension on execution window as we handle manually
+        input1_win.set(Window::DimX, Window::Dimension(0, 1, 1));
+        input2_win.set(Window::DimX, Window::Dimension(0, 1, 1));
+
+        Iterator input1(in1, input1_win);
+        Iterator input2(in2, input2_win);
+        Iterator output(out, win);
+
+        execute_window_loop(win, [&](const Coordinates &)
+        {
+            const auto input1_ptr = reinterpret_cast<const 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 = (*neon_func)(window_start_x, window_end_x, window_step_x, input1_ptr, input2_ptr, output_ptr, voffset1, voffset2,
+                                 vscale1, vscale2, voffseto, invvscaleo);
+            for(; x < window_end_x; ++x)
+            {
+                const float afs   = dequantize_qasymm8_signed(*(input1_ptr + x), input1_qinfo);
+                const float bfs   = dequantize_qasymm8_signed(*(input2_ptr + x), input2_qinfo);
+                *(output_ptr + x) = (*scalar_func)(afs, bfs, output_qinfo);
+            }
+        },
+        input1, input2, output);
+    }
+}
+
+template <ArithmeticOperation op>
+void elementwise_arithm_op_quantized(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window)
+{
+    elementwise_op_quantized(in1, in2, out, window, &elementwise_arithm_op_quantized_scalar<op>,
+                             &elementwise_arithm_op_quantized_broadcast_loop<op>,
+                             &elementwise_arithm_op_quantized_loop<op>);
+}
+template <ArithmeticOperation op>
+void elementwise_arithm_op_quantized_signed(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window)
+{
+    elementwise_op_quantized_signed(in1, in2, out, window, &elementwise_arithm_op_quantized_signed_scalar<op>,
+                                    &elementwise_arithm_op_quantized_signed_broadcast_loop<op>,
+                                    &elementwise_arithm_op_quantized_singed_loop<op>);
+}
+
+template <ComparisonOperation op>
+void elementwise_comp_op_quantized(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window)
+{
+    elementwise_op_quantized(in1, in2, out, window, &elementwise_comp_op_quantized_scalar<op>,
+                             &elementwise_comp_op_quantized_broadcast_loop<op>,
+                             &elementwise_comp_op_quantized_loop<op>);
+}
+
+template <ComparisonOperation op>
+void elementwise_comp_op_quantized_signed(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window)
+{
+    elementwise_comp_quantized_signed(in1, in2, out, window, &elementwise_comp_op_quantized_scalar<op>,
+                                      &elementwise_comp_op_quantized_signed_broadcast_loop<op>,
+                                      &elementwise_comp_op_quantized_signed_loop<op>);
+}
+} // namespace cpu
+} // namespace arm_compute
+
+#endif /* SRC_CORE_NEON_KERNELS_ELEMENTWISE_QUANTIZED_LIST_H */
diff --git a/src/core/cpu/kernels/elementwise/sve/elementwise_list.h b/src/core/cpu/kernels/elementwise/sve/elementwise_list.h
new file mode 100644
index 0000000000..83c3355de4
--- /dev/null
+++ b/src/core/cpu/kernels/elementwise/sve/elementwise_list.h
@@ -0,0 +1,366 @@
+/*
+ * 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.
+ */
+#ifndef SRC_CORE_SVE_KERNELS_ELEMENTWISE_LIST_H
+#define SRC_CORE_SVE_KERNELS_ELEMENTWISE_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 "src/core/NEON/wrapper/svtraits.h"
+#include <arm_sve.h>
+
+namespace arm_compute
+{
+namespace cpu
+{
+namespace sve
+{
+using namespace arm_compute::wrapper;
+
+template <typename VectorType>
+inline VectorType elementwise_pow(svbool_t &pg, const VectorType &a, const VectorType &b)
+{
+    return svpow_z(pg, a, b);
+}
+
+template <>
+inline svint32_t elementwise_pow<svint32_t>(svbool_t &pg, const svint32_t &a, const svint32_t &b)
+{
+    return svcvt_s32_z(pg, svpow_z(pg, svcvt_f32_z(pg, a), svcvt_f32_z(pg, b)));
+}
+
+template <typename VectorType>
+inline VectorType elementwise_div(svbool_t &pg, const VectorType &a, const VectorType &b)
+{
+    return svdiv_z(pg, a, b);
+}
+
+template <>
+inline svint32_t elementwise_div<svint32_t>(svbool_t &pg, const svint32_t &a, const svint32_t &b)
+{
+    return svcvt_s32_z(pg, svdiv_z(pg, svcvt_f32_z(pg, a), svcvt_f32_z(pg, b)));
+}
+
+template <typename VectorType>
+inline VectorType elementwise_arithmetic_op(svbool_t &pg, const VectorType &a, const VectorType &b, ArithmeticOperation op)
+{
+    using ScalarType = typename sve_scalar<VectorType>::type;
+    VectorType res{};
+
+    switch(op)
+    {
+        case ArithmeticOperation::MAX:
+            res = svmax_z(pg, a, b);
+            break;
+        case ArithmeticOperation::MIN:
+            res = svmin_z(pg, a, b);
+            break;
+        case ArithmeticOperation::SQUARED_DIFF:
+        {
+            const auto tmp = svsub_z(pg, a, b);
+            res            = svmul_z(pg, tmp, tmp);
+            break;
+        }
+        case ArithmeticOperation::PRELU:
+        {
+            const auto zero = svdup_n(ScalarType(0));
+            const auto tmp  = svmul_z(pg, a, b);
+            const auto gt   = svcmpgt(pg, a, zero);
+            res             = svsel(gt, a, tmp);
+            break;
+        }
+        case ArithmeticOperation::DIV:
+        {
+            res = elementwise_div(pg, a, b);
+            break;
+        }
+        case ArithmeticOperation::POWER:
+        {
+            res = elementwise_pow(pg, a, b);
+            break;
+        }
+        default:
+            ARM_COMPUTE_ERROR("NOT_SUPPORTED!");
+    }
+
+    return res;
+}
+
+template <uint32_t bytewidth>
+inline svbool_t narrow_to_byte_predicate(svbool_t pg)
+{
+    const auto all_false = svpfalse();
+
+    switch(bytewidth)
+    {
+        case 8:
+            pg = svuzp1_b32(pg, all_false);
+        /* fall through */
+        case 4:
+            pg = svuzp1_b16(pg, all_false);
+        /* fall through */
+        case 2:
+            pg = svuzp1_b8(pg, all_false);
+        /* fall through */
+        default:
+            break;
+    }
+    return pg;
+}
+
+template <typename InputVectorType, typename OutputVectorType>
+inline OutputVectorType elementwise_comparison_op(svbool_t &pg, const InputVectorType &a, const InputVectorType &b, ComparisonOperation op)
+{
+    svbool_t selection_vector{};
+
+    switch(op)
+    {
+        case ComparisonOperation::Equal:
+            selection_vector = svcmpeq(pg, a, b);
+            break;
+        case ComparisonOperation::NotEqual:
+            selection_vector = svcmpne(pg, a, b);
+            break;
+        case ComparisonOperation::Greater:
+            selection_vector = svcmpgt(pg, a, b);
+            break;
+        case ComparisonOperation::GreaterEqual:
+            selection_vector = svcmpge(pg, a, b);
+            break;
+        case ComparisonOperation::Less:
+            selection_vector = svcmplt(pg, a, b);
+            break;
+        case ComparisonOperation::LessEqual:
+            selection_vector = svcmple(pg, a, b);
+            break;
+        default:
+            ARM_COMPUTE_ERROR("NOT_SUPPORTED!");
+    }
+
+    using InputScalarType = typename sve_scalar<InputVectorType>::type;
+    selection_vector      = narrow_to_byte_predicate<sizeof(InputScalarType)>(selection_vector);
+
+    using OutputScalarType  = typename sve_scalar<OutputVectorType>::type;
+    const auto false_vector = svdup_n(static_cast<OutputScalarType>((uint32_t)0));
+    const auto true_vector  = svdup_n(static_cast<OutputScalarType>(~(uint32_t)0));
+    auto       ret          = svsel(selection_vector, true_vector, false_vector);
+
+    return ret;
+}
+
+template <typename InputScalarType, typename OutputScalarType, typename OperatorType>
+struct LoopArguments
+{
+    OperatorType           op;
+    const InputScalarType *input1_ptr;
+    const InputScalarType *input2_ptr;
+    OutputScalarType      *output_ptr;
+};
+
+template <typename InputScalarType, typename OutputScalarType, typename OperatorType>
+struct BroadcastLoopArguments
+{
+    OperatorType           op;
+    const InputScalarType *input1_ptr;
+    InputScalarType        broadcast_value;
+    OutputScalarType      *output_ptr;
+    bool                   reorder;
+};
+
+template <typename InputScalarType, typename OutputScalarType>
+inline void arithmetic_op_loop(svbool_t pg, const LoopArguments<InputScalarType, OutputScalarType, ArithmeticOperation> &args)
+{
+    const auto in1 = svld1(pg, args.input1_ptr);
+    const auto in2 = svld1(pg, args.input2_ptr);
+    const auto res = elementwise_arithmetic_op<typename sve_vector<InputScalarType>::type>(pg, in1, in2, args.op);
+    svst1(pg, args.output_ptr, res);
+}
+
+template <typename InputScalarType, typename OutputScalarType>
+inline void arithmetic_op_broadcast_loop(svbool_t pg, const BroadcastLoopArguments<InputScalarType, OutputScalarType, ArithmeticOperation> &args)
+{
+    const auto non_broadcast_vector = svld1(pg, args.input1_ptr);
+    const auto broadcast_vector     = svdup_n(args.broadcast_value);
+    const auto in1                  = args.reorder ? broadcast_vector : non_broadcast_vector;
+    const auto in2                  = args.reorder ? non_broadcast_vector : broadcast_vector;
+    const auto res                  = elementwise_arithmetic_op<typename sve_vector<InputScalarType>::type>(pg, in1, in2, args.op);
+    svst1(pg, args.output_ptr, res);
+}
+
+template <typename InputScalarType, typename OutputScalarType>
+inline void comparison_op_loop(svbool_t pg, const LoopArguments<InputScalarType, OutputScalarType, ComparisonOperation> &args)
+{
+    const auto     in1       = svld1(pg, args.input1_ptr);
+    const auto     in2       = svld1(pg, args.input2_ptr);
+    const auto     res       = elementwise_comparison_op<typename sve_vector<InputScalarType>::type, typename sve_vector<OutputScalarType>::type>(pg, in1, in2, args.op);
+    const svbool_t output_pg = narrow_to_byte_predicate<sizeof(InputScalarType)>(pg);
+    svst1(output_pg, args.output_ptr, res);
+}
+
+template <typename InputScalarType, typename OutputScalarType>
+inline void comparison_op_broadcast_loop(svbool_t pg, const BroadcastLoopArguments<InputScalarType, OutputScalarType, ComparisonOperation> &args)
+{
+    const auto     non_broadcast_vector = svld1(pg, args.input1_ptr);
+    const auto     broadcast_vector     = svdup_n(args.broadcast_value);
+    const auto     in1                  = args.reorder ? broadcast_vector : non_broadcast_vector;
+    const auto     in2                  = args.reorder ? non_broadcast_vector : broadcast_vector;
+    const auto     res                  = elementwise_comparison_op<typename sve_vector<InputScalarType>::type, typename sve_vector<OutputScalarType>::type>(pg, in1, in2, args.op);
+    const svbool_t output_pg            = narrow_to_byte_predicate<sizeof(InputScalarType)>(pg);
+    svst1(output_pg, args.output_ptr, res);
+}
+
+template <typename InputScalarType, typename OutputScalarType, typename OperatorType>
+using LoopFuncType = void (*)(svbool_t, const LoopArguments<InputScalarType, OutputScalarType, OperatorType> &);
+
+template <typename InputScalarType, typename OutputScalarType, typename OperatorType>
+using BroadcastLoopFuncType = void (*)(svbool_t, const BroadcastLoopArguments<InputScalarType, OutputScalarType, OperatorType> &);
+
+template <typename InputVectorType, typename OutputVectorType, typename OperatorType,
+          typename InputScalarType  = typename sve_scalar<InputVectorType>::type,
+          typename OutputScalarType = typename sve_scalar<OutputVectorType>::type>
+void elementwise_op(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window,
+                    OperatorType op,
+                    LoopFuncType<InputScalarType, OutputScalarType, OperatorType>          func,
+                    BroadcastLoopFuncType<InputScalarType, OutputScalarType, OperatorType> broadcast_func)
+{
+    const auto all_true_pg = svptrue<InputScalarType>();
+
+    // Create input windows
+    Window input1_win = window.broadcast_if_dimension_le_one(in1->info()->tensor_shape());
+    Window input2_win = window.broadcast_if_dimension_le_one(in2->info()->tensor_shape());
+
+    // Clear X Dimension on execution window as we handle manually
+    Window win = window;
+    win.set(Window::DimX, Window::Dimension(0, 1, 1));
+
+    const auto window_start_x        = static_cast<int>(window.x().start());
+    const auto window_end_x          = static_cast<int>(window.x().end());
+    const bool is_broadcast_across_x = in1->info()->tensor_shape().x() != in2->info()->tensor_shape().x();
+
+    if(is_broadcast_across_x)
+    {
+        const bool     is_broadcast_input_2 = input2_win.x().step() == 0;
+        Window         broadcast_win        = is_broadcast_input_2 ? input2_win : input1_win;
+        Window         non_broadcast_win    = !is_broadcast_input_2 ? input2_win : input1_win;
+        const ITensor *broadcast_tensor     = is_broadcast_input_2 ? in2 : in1;
+        const ITensor *non_broadcast_tensor = !is_broadcast_input_2 ? in2 : in1;
+
+        // Clear X Dimension on execution window as we handle manually
+        non_broadcast_win.set(Window::DimX, Window::Dimension(0, 1, 1));
+
+        Iterator broadcast_input(broadcast_tensor, broadcast_win);
+        Iterator non_broadcast_input(non_broadcast_tensor, non_broadcast_win);
+        Iterator output(out, win);
+
+        execute_window_loop(win, [&](const Coordinates &)
+        {
+            auto                  output_ptr              = reinterpret_cast<OutputScalarType *>(output.ptr());
+            const auto            non_broadcast_input_ptr = reinterpret_cast<const InputScalarType *>(non_broadcast_input.ptr());
+            const InputScalarType broadcast_value         = *reinterpret_cast<const InputScalarType *>(broadcast_input.ptr());
+
+            int x = window_start_x;
+
+            svbool_t pg = svwhilelt<InputScalarType>(x, window_end_x);
+            do
+            {
+                broadcast_func(pg,
+                {
+                    op,
+                    non_broadcast_input_ptr + x,
+                    broadcast_value,
+                    output_ptr + x,
+                    !is_broadcast_input_2
+                });
+                x += svcnt<InputScalarType>();
+                pg = svwhilelt<InputScalarType>(x, window_end_x);
+            }
+            while(svptest_any(all_true_pg, pg));
+        },
+        broadcast_input, non_broadcast_input, output);
+    }
+    else
+    {
+        // Clear X Dimension on execution window as we handle manually
+        input1_win.set(Window::DimX, Window::Dimension(0, 1, 1));
+        input2_win.set(Window::DimX, Window::Dimension(0, 1, 1));
+
+        Iterator input1(in1, input1_win);
+        Iterator input2(in2, input2_win);
+        Iterator output(out, win);
+
+        execute_window_loop(win, [&](const Coordinates &)
+        {
+            auto       output_ptr = reinterpret_cast<OutputScalarType *>(output.ptr());
+            const auto input1_ptr = reinterpret_cast<const InputScalarType *>(input1.ptr());
+            const auto input2_ptr = reinterpret_cast<const InputScalarType *>(input2.ptr());
+
+            int x = window_start_x;
+
+            svbool_t pg = svwhilelt<InputScalarType>(x, window_end_x);
+            do
+            {
+                func(pg,
+                {
+                    op,
+                    input1_ptr + x,
+                    input2_ptr + x,
+                    output_ptr + x
+                });
+                x += svcnt<InputScalarType>();
+                pg = svwhilelt<InputScalarType>(x, window_end_x);
+            }
+            while(svptest_any(all_true_pg, pg));
+        },
+        input1, input2, output);
+    }
+}
+
+template <ArithmeticOperation op, typename ScalarType>
+void elementwise_arithmetic_op(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window)
+{
+    using VectorType = typename sve_vector<ScalarType>::type;
+
+    elementwise_op<VectorType, VectorType, ArithmeticOperation>(in1, in2, out, window, op,
+                                                                &arithmetic_op_loop<ScalarType, ScalarType>,
+                                                                &arithmetic_op_broadcast_loop<ScalarType, ScalarType>);
+}
+
+template <ComparisonOperation op, typename InputScalarType, typename OutputScalarType = uint8_t>
+void elementwise_comparison_op(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window)
+{
+    static_assert(sizeof(InputScalarType) >= sizeof(OutputScalarType), "input data type's width should be equal to or greater than output data type's width");
+    using InputVectorType  = typename sve_vector<InputScalarType>::type;
+    using OutputVectorType = typename sve_vector<OutputScalarType>::type;
+
+    elementwise_op<InputVectorType, OutputVectorType, ComparisonOperation>(in1, in2, out, window, op,
+                                                                           &comparison_op_loop<InputScalarType, OutputScalarType>,
+                                                                           &comparison_op_broadcast_loop<InputScalarType, OutputScalarType>);
+}
+
+} // namespace sve
+} // namespace cpu
+} // namespace arm_compute
+#endif // defined(__ARM_FEATURE_SVE)
+#endif /* SRC_CORE_SVE_KERNELS_ELEMENTWISE_LIST_H */
diff --git a/src/core/cpu/kernels/elementwise/sve/elementwise_quantized_list.h b/src/core/cpu/kernels/elementwise/sve/elementwise_quantized_list.h
new file mode 100644
index 0000000000..b6342c727c
--- /dev/null
+++ b/src/core/cpu/kernels/elementwise/sve/elementwise_quantized_list.h
@@ -0,0 +1,369 @@
+/*
+ * 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.
+ */
+#ifndef SRC_CORE_SVE_KERNELS_ELEMENTWISE_QUANTIZED_LIST_H
+#define SRC_CORE_SVE_KERNELS_ELEMENTWISE_QUANTIZED_LIST_H
+
+#if defined(__ARM_FEATURE_SVE2)
+
+#include "src/core/cpu/kernels/elementwise/sve/elementwise_list.h"
+
+namespace arm_compute
+{
+namespace cpu
+{
+namespace sve
+{
+using namespace arm_compute::wrapper;
+
+template <typename InputScalarType, typename OutputScalarType, typename OperatorType>
+struct QuantizedLoopArguments
+{
+    OperatorType           op;
+    const InputScalarType *input1_ptr;
+    const InputScalarType *input2_ptr;
+    OutputScalarType      *output_ptr;
+
+    const svint32_t   &in1_offset;
+    const svint32_t   &in2_offset;
+    const svint32_t   &out_offset;
+    const svfloat32_t &in1_scale;
+    const svfloat32_t &in2_scale;
+    const svfloat32_t &out_scale;
+};
+
+template <typename InputScalarType, typename OutputScalarType, typename OperatorType>
+struct BroadcastQuantizedLoopArguments
+{
+    OperatorType           op;
+    const InputScalarType *input1_ptr;
+    float                  broadcast_value;
+    OutputScalarType      *output_ptr;
+    bool                   reorder;
+
+    const svint32_t   &in1_offset;
+    const svint32_t   &out_offset;
+    const svfloat32_t &in1_scale;
+    const svfloat32_t &out_scale;
+};
+
+svfloat32x4_t load_quantized(const int8_t *ptr, svbool_t pg, const svint32_t &offset, const svfloat32_t &scale)
+{
+    auto x = svld1(pg, ptr);
+
+    const auto widened = svcreate4(
+                             svmovlb(svmovlb(x)),
+                             svmovlt(svmovlb(x)),
+                             svmovlb(svmovlt(x)),
+                             svmovlt(svmovlt(x)));
+
+    pg = svptrue_b8();
+
+    return svcreate4(
+               svmul_z(pg, svcvt_f32_z(pg, svsub_z(pg, svget4(widened, 0), offset)), scale),
+               svmul_z(pg, svcvt_f32_z(pg, svsub_z(pg, svget4(widened, 1), offset)), scale),
+               svmul_z(pg, svcvt_f32_z(pg, svsub_z(pg, svget4(widened, 2), offset)), scale),
+               svmul_z(pg, svcvt_f32_z(pg, svsub_z(pg, svget4(widened, 3), offset)), scale));
+}
+
+svfloat32x4_t load_quantized(const uint8_t *ptr, svbool_t pg, const svint32_t &offset, const svfloat32_t &scale)
+{
+    auto x = svld1(pg, ptr);
+
+    //vprint(x);
+
+    const auto widened = svcreate4(
+                             svmovlb(svmovlb(x)),
+                             svmovlt(svmovlb(x)),
+                             svmovlb(svmovlt(x)),
+                             svmovlt(svmovlt(x)));
+
+    pg = svptrue_b8();
+
+    return svcreate4(
+               svmul_z(pg, svcvt_f32_z(pg, svsub_z(pg, svreinterpret_s32(svget4(widened, 0)), offset)), scale),
+               svmul_z(pg, svcvt_f32_z(pg, svsub_z(pg, svreinterpret_s32(svget4(widened, 1)), offset)), scale),
+               svmul_z(pg, svcvt_f32_z(pg, svsub_z(pg, svreinterpret_s32(svget4(widened, 2)), offset)), scale),
+               svmul_z(pg, svcvt_f32_z(pg, svsub_z(pg, svreinterpret_s32(svget4(widened, 3)), offset)), scale));
+}
+
+void store_quantized(uint8_t *ptr, svbool_t pg, svfloat32x4_t data, const svint32_t &offset, const svfloat32_t &inv_scale)
+{
+    const auto quantized = svcreate4(
+                               svadd_z(pg, svcvt_s32_z(pg, svrinta_z(pg, svmul_z(pg, svget4(data, 0), inv_scale))), offset),
+                               svadd_z(pg, svcvt_s32_z(pg, svrinta_z(pg, svmul_z(pg, svget4(data, 1), inv_scale))), offset),
+                               svadd_z(pg, svcvt_s32_z(pg, svrinta_z(pg, svmul_z(pg, svget4(data, 2), inv_scale))), offset),
+                               svadd_z(pg, svcvt_s32_z(pg, svrinta_z(pg, svmul_z(pg, svget4(data, 3), inv_scale))), offset));
+
+    const auto narrowed_bottom = svqxtunt(svqxtunb(svget4(quantized, 0)), svget4(quantized, 1));
+    const auto narrowed_top    = svqxtunt(svqxtunb(svget4(quantized, 2)), svget4(quantized, 3));
+    const auto narrowed        = svqxtnt(svqxtnb(narrowed_bottom), narrowed_top);
+    svst1(pg, ptr, narrowed);
+}
+
+void store_quantized(int8_t *ptr, svbool_t pg, svfloat32x4_t data, const svint32_t &offset, const svfloat32_t &inv_scale)
+{
+    const auto quantized = svcreate4(
+                               svadd_z(pg, svcvt_s32_z(pg, svrinta_z(pg, svmul_z(pg, svget4(data, 0), inv_scale))), offset),
+                               svadd_z(pg, svcvt_s32_z(pg, svrinta_z(pg, svmul_z(pg, svget4(data, 1), inv_scale))), offset),
+                               svadd_z(pg, svcvt_s32_z(pg, svrinta_z(pg, svmul_z(pg, svget4(data, 2), inv_scale))), offset),
+                               svadd_z(pg, svcvt_s32_z(pg, svrinta_z(pg, svmul_z(pg, svget4(data, 3), inv_scale))), offset));
+
+    const auto narrowed_bottom = svqxtnt(svqxtnb(svget4(quantized, 0)), svget4(quantized, 1));
+    const auto narrowed_top    = svqxtnt(svqxtnb(svget4(quantized, 2)), svget4(quantized, 3));
+    const auto narrowed        = svqxtnt(svqxtnb(narrowed_bottom), narrowed_top);
+
+    svst1(pg, ptr, narrowed);
+}
+
+template <typename InputScalarType, typename OutputScalarType>
+inline void arithmetic_op_quantized_loop(svbool_t pg, const QuantizedLoopArguments<InputScalarType, OutputScalarType, ArithmeticOperation> &args)
+{
+    const auto in1 = load_quantized(args.input1_ptr, pg, args.in1_offset, args.in1_scale);
+    const auto in2 = load_quantized(args.input2_ptr, pg, args.in2_offset, args.in2_scale);
+
+    const auto result = svcreate4(
+                            elementwise_arithmetic_op<svfloat32_t>(pg, svget4(in1, 0), svget4(in2, 0), args.op),
+                            elementwise_arithmetic_op<svfloat32_t>(pg, svget4(in1, 1), svget4(in2, 1), args.op),
+                            elementwise_arithmetic_op<svfloat32_t>(pg, svget4(in1, 2), svget4(in2, 2), args.op),
+                            elementwise_arithmetic_op<svfloat32_t>(pg, svget4(in1, 3), svget4(in2, 3), args.op));
+
+    store_quantized(args.output_ptr, pg, result, args.out_offset, args.out_scale);
+}
+
+template <typename InputScalarType, typename OutputScalarType>
+inline void arithmetic_op_broadcast_quantized_loop(svbool_t pg, const BroadcastQuantizedLoopArguments<InputScalarType, OutputScalarType, ArithmeticOperation> &args)
+{
+    const auto in1 = load_quantized(args.input1_ptr, pg, args.in1_offset, args.in1_scale);
+    const auto in2 = svcreate4(
+                         svdup_n(args.broadcast_value), svdup_n(args.broadcast_value), svdup_n(args.broadcast_value), svdup_n(args.broadcast_value));
+
+    const auto &af = args.reorder ? in2 : in1;
+    const auto &bf = args.reorder ? in1 : in2;
+
+    const auto result = svcreate4(
+                            elementwise_arithmetic_op<svfloat32_t>(pg, svget4(af, 0), svget4(bf, 0), args.op),
+                            elementwise_arithmetic_op<svfloat32_t>(pg, svget4(af, 1), svget4(bf, 1), args.op),
+                            elementwise_arithmetic_op<svfloat32_t>(pg, svget4(af, 2), svget4(bf, 2), args.op),
+                            elementwise_arithmetic_op<svfloat32_t>(pg, svget4(af, 3), svget4(bf, 3), args.op));
+
+    store_quantized(args.output_ptr, pg, result, args.out_offset, args.out_scale);
+}
+
+template <typename InputScalarType, typename OutputScalarType>
+inline void comparison_op_quantized_loop(svbool_t pg, const QuantizedLoopArguments<InputScalarType, OutputScalarType, ComparisonOperation> &args)
+{
+    const auto in1 = load_quantized(args.input1_ptr, pg, args.in1_offset, args.in1_scale);
+    const auto in2 = load_quantized(args.input2_ptr, pg, args.in2_offset, args.in2_scale);
+
+    using OutputVectorType = typename sve_vector<OutputScalarType>::type;
+
+    const auto result = svcreate4(
+                            elementwise_comparison_op<svfloat32_t, OutputVectorType>(pg, svget4(in1, 0), svget4(in2, 0), args.op),
+                            elementwise_comparison_op<svfloat32_t, OutputVectorType>(pg, svget4(in1, 1), svget4(in2, 1), args.op),
+                            elementwise_comparison_op<svfloat32_t, OutputVectorType>(pg, svget4(in1, 2), svget4(in2, 2), args.op),
+                            elementwise_comparison_op<svfloat32_t, OutputVectorType>(pg, svget4(in1, 3), svget4(in2, 3), args.op));
+
+    const auto zipped_bottom = svzip1(svget4(result, 0), svget4(result, 1));
+    const auto zipped_top    = svzip1(svget4(result, 2), svget4(result, 3));
+    const auto zipped        = svzip1(zipped_bottom, zipped_top);
+    svst1(pg, args.output_ptr, zipped);
+}
+
+template <typename InputScalarType, typename OutputScalarType>
+inline void comparison_op_broadcast_quantized_loop(svbool_t pg, const BroadcastQuantizedLoopArguments<InputScalarType, OutputScalarType, ComparisonOperation> &args)
+{
+    const auto in1 = load_quantized(args.input1_ptr, pg, args.in1_offset, args.in1_scale);
+    const auto in2 = svcreate4(
+                         svdup_n(args.broadcast_value), svdup_n(args.broadcast_value), svdup_n(args.broadcast_value), svdup_n(args.broadcast_value));
+
+    const auto &af = args.reorder ? in2 : in1;
+    const auto &bf = args.reorder ? in1 : in2;
+
+    using OutputVectorType = typename sve_vector<OutputScalarType>::type;
+
+    const auto result = svcreate4(
+                            elementwise_comparison_op<svfloat32_t, OutputVectorType>(pg, svget4(af, 0), svget4(bf, 0), args.op),
+                            elementwise_comparison_op<svfloat32_t, OutputVectorType>(pg, svget4(af, 1), svget4(bf, 1), args.op),
+                            elementwise_comparison_op<svfloat32_t, OutputVectorType>(pg, svget4(af, 2), svget4(bf, 2), args.op),
+                            elementwise_comparison_op<svfloat32_t, OutputVectorType>(pg, svget4(af, 3), svget4(bf, 3), args.op));
+
+    const auto zipped_bottom = svzip1(svget4(result, 0), svget4(result, 1));
+    const auto zipped_top    = svzip1(svget4(result, 2), svget4(result, 3));
+    const auto zipped        = svzip1(zipped_bottom, zipped_top);
+    svst1(pg, args.output_ptr, zipped);
+}
+
+template <typename InputScalarType, typename OutputScalarType, typename OperatorType>
+using LoopQuantizedFuncType = void (*)(svbool_t, const QuantizedLoopArguments<InputScalarType, OutputScalarType, OperatorType> &);
+
+template <typename InputScalarType, typename OutputScalarType, typename OperatorType>
+using BroadcastQuantizedLoopFuncType = void (*)(svbool_t, const BroadcastQuantizedLoopArguments<InputScalarType, OutputScalarType, OperatorType> &);
+
+template <typename InputVectorType, typename OutputVectorType, typename OperatorType,
+          typename InputScalarType  = typename sve_scalar<InputVectorType>::type,
+          typename OutputScalarType = typename sve_scalar<OutputVectorType>::type>
+void elementwise_quantized_op(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window,
+                              OperatorType op,
+                              LoopQuantizedFuncType<InputScalarType, OutputScalarType, OperatorType>          func,
+                              BroadcastQuantizedLoopFuncType<InputScalarType, OutputScalarType, OperatorType> broadcast_func)
+{
+    const auto all_true_pg = wrapper::svptrue<InputScalarType>();
+
+    // Create input windows
+    Window input1_win = window.broadcast_if_dimension_le_one(in1->info()->tensor_shape());
+    Window input2_win = window.broadcast_if_dimension_le_one(in2->info()->tensor_shape());
+
+    // Clear X Dimension on execution window as we handle manually
+    Window win = window;
+    win.set(Window::DimX, Window::Dimension(0, 1, 1));
+
+    const auto window_start_x        = static_cast<int>(window.x().start());
+    const auto window_end_x          = static_cast<int>(window.x().end());
+    const bool is_broadcast_across_x = in1->info()->tensor_shape().x() != in2->info()->tensor_shape().x();
+
+    const auto output_voffset = svdup_n(out->info()->quantization_info().uniform().offset);
+    const auto output_vscale  = svdup_n(1.f / out->info()->quantization_info().uniform().scale);
+
+    if(is_broadcast_across_x)
+    {
+        const bool     is_broadcast_input_2 = input2_win.x().step() == 0;
+        Window         broadcast_win        = is_broadcast_input_2 ? input2_win : input1_win;
+        Window         non_broadcast_win    = !is_broadcast_input_2 ? input2_win : input1_win;
+        const ITensor *broadcast_tensor     = is_broadcast_input_2 ? in2 : in1;
+        const ITensor *non_broadcast_tensor = !is_broadcast_input_2 ? in2 : in1;
+
+        const auto non_broadcast_qinfo = is_broadcast_input_2 ? in1->info()->quantization_info() : in2->info()->quantization_info();
+        const auto broadcast_qinfo     = is_broadcast_input_2 ? in2->info()->quantization_info() : in1->info()->quantization_info();
+
+        const auto non_broadcast_voffset = svdup_n(non_broadcast_qinfo.uniform().offset);
+        const auto non_broadcast_vscale  = svdup_n(non_broadcast_qinfo.uniform().scale);
+
+        // Clear X Dimension on execution window as we handle manually
+        non_broadcast_win.set(Window::DimX, Window::Dimension(0, 1, 1));
+
+        Iterator broadcast_input(broadcast_tensor, broadcast_win);
+        Iterator non_broadcast_input(non_broadcast_tensor, non_broadcast_win);
+        Iterator output(out, win);
+
+        execute_window_loop(win, [&](const Coordinates &)
+        {
+            auto                  output_ptr              = reinterpret_cast<OutputScalarType *>(output.ptr());
+            const auto            non_broadcast_input_ptr = reinterpret_cast<const InputScalarType *>(non_broadcast_input.ptr());
+            const InputScalarType broadcast_value         = *reinterpret_cast<const InputScalarType *>(broadcast_input.ptr());
+
+            int x = window_start_x;
+
+            svbool_t pg = wrapper::svwhilelt<InputScalarType>(x, window_end_x);
+            do
+            {
+                const auto args = BroadcastQuantizedLoopArguments<InputScalarType, OutputScalarType, OperatorType>
+                {
+                    op,
+                    non_broadcast_input_ptr + x,
+                    Qasymm8QuantizationHelper<InputScalarType>::dequantize(broadcast_value, broadcast_qinfo),
+                    output_ptr + x,
+                    !is_broadcast_input_2,
+                    non_broadcast_voffset, output_voffset,
+                    non_broadcast_vscale, output_vscale
+                };
+                broadcast_func(pg, args);
+                x += wrapper::svcnt<InputScalarType>();
+                pg = wrapper::svwhilelt<InputScalarType>(x, window_end_x);
+            }
+            while(svptest_any(all_true_pg, pg));
+        },
+        broadcast_input, non_broadcast_input, output);
+    }
+    else
+    {
+        // Clear X Dimension on execution window as we handle manually
+        input1_win.set(Window::DimX, Window::Dimension(0, 1, 1));
+        input2_win.set(Window::DimX, Window::Dimension(0, 1, 1));
+
+        Iterator input1(in1, input1_win);
+        Iterator input2(in2, input2_win);
+        Iterator output(out, win);
+
+        const auto in1_voffset = svdup_n(in1->info()->quantization_info().uniform().offset);
+        const auto in1_vscale  = svdup_n(in1->info()->quantization_info().uniform().scale);
+
+        const auto in2_voffset = svdup_n(in2->info()->quantization_info().uniform().offset);
+        const auto in2_vscale  = svdup_n(in2->info()->quantization_info().uniform().scale);
+
+        execute_window_loop(win, [&](const Coordinates &)
+        {
+            auto       output_ptr = reinterpret_cast<OutputScalarType *>(output.ptr());
+            const auto input1_ptr = reinterpret_cast<const InputScalarType *>(input1.ptr());
+            const auto input2_ptr = reinterpret_cast<const InputScalarType *>(input2.ptr());
+
+            int x = window_start_x;
+
+            svbool_t pg = wrapper::svwhilelt<InputScalarType>(x, window_end_x);
+            do
+            {
+                const auto args = QuantizedLoopArguments<InputScalarType, OutputScalarType, OperatorType>
+                {
+                    op,
+                    input1_ptr + x,
+                    input2_ptr + x,
+                    output_ptr + x,
+                    in1_voffset, in2_voffset, output_voffset,
+                    in1_vscale, in2_vscale, output_vscale
+                };
+                func(pg, args);
+                x += wrapper::svcnt<InputScalarType>();
+                pg = wrapper::svwhilelt<InputScalarType>(x, window_end_x);
+            }
+            while(svptest_any(all_true_pg, pg));
+        },
+        input1, input2, output);
+    }
+}
+
+template <ArithmeticOperation op, typename ScalarType>
+void elementwise_arithmetic_quantized_op(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window)
+{
+    using VectorType = typename sve_vector<ScalarType>::type;
+    elementwise_quantized_op<VectorType, VectorType, ArithmeticOperation>(in1, in2, out, window, op,
+                                                                          &arithmetic_op_quantized_loop<ScalarType, ScalarType>,
+                                                                          &arithmetic_op_broadcast_quantized_loop<ScalarType, ScalarType>);
+}
+
+template <ComparisonOperation op, typename InputScalarType, typename OutputScalarType = uint8_t>
+void elementwise_comparison_quantized_op(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window)
+{
+    static_assert(sizeof(InputScalarType) >= sizeof(OutputScalarType), "input data type's width should be equal to or greater than output data type's width");
+    using InputVectorType  = typename sve_vector<InputScalarType>::type;
+    using OutputVectorType = typename sve_vector<OutputScalarType>::type;
+    elementwise_quantized_op<InputVectorType, OutputVectorType, ComparisonOperation>(in1, in2, out, window, op,
+                                                                                     &comparison_op_quantized_loop<InputScalarType, OutputScalarType>,
+                                                                                     &comparison_op_broadcast_quantized_loop<InputScalarType, OutputScalarType>);
+}
+
+} // namespace sve
+} // namespace cpu
+} // namespace arm_compute
+
+#endif /* defined(__ARM_FEATURE_SVE2) */
+#endif /* SRC_CORE_SVE_KERNELS_ELEMENTWISE_QUANTIZED_LIST_H */
\ No newline at end of file