Port NEGEMMLowp Part 1

Details:
Port NEGEMMLowpQuantizeDownInt32ScaleKernel to CpuGemmLowpQuantizeDownInt32ScaleKernel
Port NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel to CpuGemmLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel
Port NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel to CpuGemmLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel
Port NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel to CpuGemmLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel
Port NEGEMMLowpOutputStage functions to CpuGemmLowpOutputStage operators

Partially Resolves: COMPMID-4403

Change-Id: I6d5f45e43f35d731d564ed3b5c0e804d2a318fb1
Signed-off-by: Manuel Bottini <manuel.bottini@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/5833
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>

13 files changed, 652 insertions, 704 deletions

diff --git a/src/core/NEON/NEKernels.h b/src/core/NEON/NEKernels.h
index ea15f4eddd..268871a4e8 100644
--- a/src/core/NEON/NEKernels.h
+++ b/src/core/NEON/NEKernels.h
@@ -45,10 +45,6 @@
 #include "src/core/NEON/kernels/NEGEMMLowpMatrixMultiplyKernel.h"
 #include "src/core/NEON/kernels/NEGEMMLowpOffsetContributionKernel.h"
 #include "src/core/NEON/kernels/NEGEMMLowpOffsetContributionOutputStageKernel.h"
-#include "src/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ScaleKernel.h"
-#include "src/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel.h"
-#include "src/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel.h"
-#include "src/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel.h"
 #include "src/core/NEON/kernels/NEGEMMLowpReductionKernel.h"
 #include "src/core/NEON/kernels/NEGEMMMatrixAdditionKernel.h"
 #include "src/core/NEON/kernels/NEGEMMMatrixMultiplyKernel.h"
diff --git a/src/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ScaleKernel.h b/src/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ScaleKernel.h
deleted file mode 100644
index 021ff8e2e0..0000000000
--- a/src/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ScaleKernel.h
+++ /dev/null
@@ -1,114 +0,0 @@
-/*
- * Copyright (c) 2020-2021 Arm Limited.
- *
- * SPDX-License-Identifier: MIT
- *
- * Permission is hereby granted, free of charge, to any person obtaining a copy
- * of this software and associated documentation files (the "Software"), to
- * deal in the Software without restriction, including without limitation the
- * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
- * sell copies of the Software, and to permit persons to whom the Software is
- * furnished to do so, subject to the following conditions:
- *
- * The above copyright notice and this permission notice shall be included in all
- * copies or substantial portions of the Software.
- *
- * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
- * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
- * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
- * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
- * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
- * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
- * SOFTWARE.
- */
-#ifndef ARM_COMPUTE_NEGEMMLOWPQUANTIZEDOWNINT32SCALEKERNEL_H
-#define ARM_COMPUTE_NEGEMMLOWPQUANTIZEDOWNINT32SCALEKERNEL_H
-
-#include "src/core/NEON/INEKernel.h"
-
-namespace arm_compute
-{
-class ITensor;
-
-/** Kernel used to quantize down the int32 accumulator values of GEMMLowp to QASYMM8/QASYMM8_SIGNED
- *
- * This kernel takes a final int32 accumulator value (the output of @ref NEGEMMLowpMatrixMultiplyKernel), and processes it to obtain the final QASYMM8/QASYMM8_SIGNED value.
- * The following computations will be performed by the kernel:
- *
- *  -# Add offset terms to final result
- *  -# Multiply each entry of result by result_mult_int
- *  -# Add bias to final result if bias tensor is not a nullptr
- *  -# Shift the int32 accumulator by result_shift
- *  -# Clamp the value between the specified min and max bounds
- *  -# Clamp the resulting int32 values:
- *  -#  -to the [0..255] range and cast to QASYMM8.
- *  -#  -to the [-128..127] range and cast to QASYMM8_SIGNED.
- *
- */
-class NEGEMMLowpQuantizeDownInt32ScaleKernel : public INEKernel
-{
-public:
-    const char *name() const override
-    {
-        return "NEGEMMLowpQuantizeDownInt32ScaleKernel";
-    }
-    /** Constructor */
-    NEGEMMLowpQuantizeDownInt32ScaleKernel();
-    /** Prevent instances of this class from being copied (As this class contains pointers)*/
-    NEGEMMLowpQuantizeDownInt32ScaleKernel(const NEGEMMLowpQuantizeDownInt32ScaleKernel &) = delete;
-    /** Prevent instances of this class from being copied (As this class contains pointers)*/
-    NEGEMMLowpQuantizeDownInt32ScaleKernel &operator=(const NEGEMMLowpQuantizeDownInt32ScaleKernel &) = delete;
-    /** Allow instances of this class to be moved */
-    NEGEMMLowpQuantizeDownInt32ScaleKernel(NEGEMMLowpQuantizeDownInt32ScaleKernel &&) = default;
-    /** Allow instances of this class to be moved */
-    NEGEMMLowpQuantizeDownInt32ScaleKernel &operator=(NEGEMMLowpQuantizeDownInt32ScaleKernel &&) = default;
-    /** Default destructor */
-    ~NEGEMMLowpQuantizeDownInt32ScaleKernel() = default;
-    /** Initialise the kernel's input and output.
-     *
-     * @param[in]  input        Input tensor. Data type supported: S32
-     * @param[in]  bias         Biases tensor. Only shared biases supported and it can be a nullptr if the biases addition is not required.
-     *                          Biases are 1D tensor with dimensions [OFM]. Data type supported: Same as @p input.
-     * @param[out] output       Output tensor. Data type supported: Data type supported: QASYMM8/QASYMM8_SIGNED
-     * @param[out] output_stage GEMMLowp output stage metadata.
-     */
-    void configure(const ITensor *input, const ITensor *bias, ITensor *output, const GEMMLowpOutputStageInfo *output_stage);
-    /** Static function to check if given info will lead to a valid configuration of @ref NEGEMMLowpQuantizeDownInt32ScaleKernel
-     *
-     * @param[in]  input        Input tensor. Data type supported: S32
-     * @param[in]  bias         Biases tensor. Only shared biases supported and it can be a nullptr if the biases addition is not required.
-     *                          Biases are 1D tensor with dimensions [OFM]. Data type supported: Same as @p input.
-     * @param[in]  output       Output tensor. Data type supported: Data type supported: QASYMM8/QASYMM8_SIGNED
-     * @param[out] output_stage GEMMLowp output stage metadata.
-     *
-     * @return a status
-     */
-    static Status validate(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, const GEMMLowpOutputStageInfo *output_stage);
-
-    // Inherited methods overridden:
-    void run(const Window &window, const ThreadInfo &info) override;
-
-private:
-    /** Template function to run the NEGEMMLowpQuantizeDownInt32ScaleKernel
-     *
-     * @param[in] window Region on which to execute the kernel. (Must be a valid region of the window returned by window()).
-     */
-    template <typename T>
-    void run(const Window &window);
-
-    /** Common signature for all the specialised NEGEMMLowpQuantizeDownInt32ScaleKernel functions
-     *
-     * @param[in] window Region on which to execute the kernel.
-     */
-    using QuantizeDownFunctionPtr = void (NEGEMMLowpQuantizeDownInt32ScaleKernel::*)(const Window &window);
-
-    QuantizeDownFunctionPtr        _func;
-    const ITensor                 *_input;
-    const ITensor                 *_bias;
-    ITensor                       *_output;
-    const GEMMLowpOutputStageInfo *_output_stage;
-    bool                           _is_bounded_relu;
-};
-} // namespace arm_compute
-
-#endif /* ARM_COMPUTE_NEGEMMLOWPQUANTIZEDOWNINT32SCALEKERNEL_H */
diff --git a/src/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel.h b/src/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel.h
deleted file mode 100644
index b01b204a6f..0000000000
--- a/src/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel.h
+++ /dev/null
@@ -1,118 +0,0 @@
-/*
- * Copyright (c) 2019-2021 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_NEGEMMLOWPQUANTIZEDOWNINT32TOINT16SCALEBYFIXEDPOINTKERNEL_H
-#define ARM_COMPUTE_NEGEMMLOWPQUANTIZEDOWNINT32TOINT16SCALEBYFIXEDPOINTKERNEL_H
-
-#include "src/core/NEON/INEKernel.h"
-
-namespace arm_compute
-{
-class ITensor;
-
-/** Kernel used to quantize down the int32 accumulator values of GEMMLowp to QSYMM16
- *
- * This kernel takes a final int32 accumulator value (the output of @ref NEGEMMLowpMatrixMultiplyKernel), and processes it to obtain the final QSYMM16 value.
- * The following computations will be performed by the kernel:
- *
- *  -# Compute fixed point multiplication between each entry of input by result_fixedpoint_multiplier
- *  -# Add bias to final result if bias tensor is not a nullptr
- *  -# Round to nearest division by a power-of-two using result_shift
- *  -# Clamp the value between the specified min and max bounds
- *  -# Clamp the resulting int32 values to the [-32768, 32767] range and cast to QSYMM16.
- *
- */
-class NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel : public INEKernel
-{
-public:
-    const char *name() const override
-    {
-        return "NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel";
-    }
-    /** Constructor */
-    NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel();
-    /** Prevent instances of this class from being copied (As this class contains pointers)*/
-    NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel(const NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel &) = delete;
-    /** Prevent instances of this class from being copied (As this class contains pointers)*/
-    NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel &operator=(const NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel &) = delete;
-    /** Allow instances of this class to be moved */
-    NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel(NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel &&) = default;
-    /** Allow instances of this class to be moved */
-    NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel &operator=(NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel &&) = default;
-    /** Default destructor */
-    ~NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel() = default;
-    /** Initialise the kernel's input and output.
-     *
-     * @param[in]  input                        Input tensor. Data type supported: S32
-     * @param[in]  bias                         Biases tensor. Only shared biases supported and it can be a nullptr if the biases addition is not required.
-     *                                          Biases are 1D tensor with dimensions [OFM]. Data type supported: Same as @p input.
-     * @param[out] output                       Output tensor. Data type supported: Data type supported: QSYMM16
-     * @param[in]  result_fixedpoint_multiplier Fixed point value to be multiplied to each element of the input matrix when once the result_offset has been add
-     * @param[in]  result_shift                 Integer value used to round to nearest division by a power-of-two the result after the fixed point multiplication
-     * @param[in]  min                          (Optional) Min value used to saturate down the output result before converting back to QSYMM16. Defaults to 0.
-     * @param[in]  max                          (Optional) Max value used to saturate up the output result before converting back to QSYMM16.
-     *                                          Along with @p min, this value can be used to implement "rectified linear unit" activation functions. Defaults to 0.
-     */
-    void configure(const ITensor *input, const ITensor *bias, ITensor *output, int result_fixedpoint_multiplier, int result_shift, int min = 0, int max = 0);
-    /** Static function to check if given info will lead to a valid configuration of @ref NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel
-     *
-     * @param[in] input  Input tensor info. Data type supported: S32
-     * @param[in] bias   Biases tensor info. Only shared biases supported and it can be a nullptr if the biases addition is not required.
-     *                   Biases are 1D tensor info with dimensions [OFM]. Data type supported: Same as @p input.
-     * @param[in] output Output tensor info. Data type supported: Data type supported: QSYMM16
-     * @param[in] min    (Optional) Min value used to saturate down the output result before converting back to QSYMM16. Defaults to 0.
-     * @param[in] max    (Optional) Max value used to saturate up the output result before converting back to QSYMM16,
-     *                            Along with @p min, this value can be used to implement "rectified linear unit" activation functions. Defaults to 0.
-     *
-     * @return a status
-     */
-    static Status validate(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, int min = 0, int max = 0);
-
-    // Inherited methods overridden:
-    void run(const Window &window, const ThreadInfo &info) override;
-
-private:
-    /** Template function to run the NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel
-     *
-     * @param[in] window Region on which to execute the kernel. (Must be a valid region of the window returned by window()).
-     */
-    template <bool is_bounded_relu>
-    void run(const Window &window);
-
-    /** Common signature for all the specialised NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel functions
-     *
-     * @param[in] window Region on which to execute the kernel.
-     */
-    using QuantizeDownFunctionPtr = void (NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel::*)(const Window &window);
-
-    QuantizeDownFunctionPtr _func;
-    const ITensor          *_input;
-    const ITensor          *_bias;
-    ITensor                *_output;
-    int                     _result_fixedpoint_multiplier;
-    int                     _result_shift;
-    int                     _min;
-    int                     _max;
-};
-} // namespace arm_compute
-#endif /* ARM_COMPUTE_NEGEMMLOWPQUANTIZEDOWNINT32TOINT16SCALEBYFIXEDPOINTKERNEL_H */
diff --git a/src/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel.h b/src/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel.h
deleted file mode 100644
index 9e7dc2f599..0000000000
--- a/src/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel.h
+++ /dev/null
@@ -1,121 +0,0 @@
-/*
- * Copyright (c) 2019-2021 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_NEGEMMLOWPQUANTIZEDOWNINT32TOINT8SCALEBYFIXEDPOINTKERNEL_H
-#define ARM_COMPUTE_NEGEMMLOWPQUANTIZEDOWNINT32TOINT8SCALEBYFIXEDPOINTKERNEL_H
-
-#include "src/core/NEON/INEKernel.h"
-
-namespace arm_compute
-{
-class ITensor;
-
-/** Kernel used to quantize down the int32 accumulator values of GEMMLowp to QASYMM8_SIGNED
- *
- * This kernel takes a final int32 accumulator value (the output of @ref NEGEMMLowpMatrixMultiplyKernel), and processes it to obtain the final QASYMM8_SIGNED value.
- * The following computations will be performed by the kernel:
- *
- *  -# Compute fixed point multiplication between each entry of input by result_fixedpoint_multiplier
- *  -# Add bias to final result if bias tensor is not a nullptr
- *  -# Round to nearest division by a power-of-two using result_shift
- *  -# Add offset to each result
- *  -# Clamp the value between the specified min and max bounds
- *  -# Clamp the resulting int32 values to the [-128..127] range and cast to QASYMM8_SIGNED.
- *
- */
-class NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel : public INEKernel
-{
-public:
-    const char *name() const override
-    {
-        return "NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel";
-    }
-    /** Constructor */
-    NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel();
-    /** Prevent instances of this class from being copied (As this class contains pointers)*/
-    NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel(const NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel &) = delete;
-    /** Prevent instances of this class from being copied (As this class contains pointers)*/
-    NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel &operator=(const NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel &) = delete;
-    /** Allow instances of this class to be moved */
-    NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel(NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel &&) = default;
-    /** Allow instances of this class to be moved */
-    NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel &operator=(NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel &&) = default;
-    /** Default destructor */
-    ~NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel() = default;
-    /** Initialise the kernel's input and output.
-     *
-     * @param[in]  input                        Input tensor. Data type supported: S32
-     * @param[in]  bias                         Biases tensor. Only shared biases supported and it can be a nullptr if the biases addition is not required.
-     *                                          Biases are 1D tensor with dimensions [OFM]. Data type supported: Same as @p input.
-     * @param[out] output                       Output tensor. Data type supported: Data type supported: QASYMM8_SIGNED
-     * @param[in]  result_fixedpoint_multiplier Fixed point value to be multiplied to each element of the input matrix when once the result_offset has been add
-     * @param[in]  result_shift                 Integer value used to round to nearest division by a power-of-two the result after the fixed point multiplication
-     * @param[in]  result_offset_after_shift    Offset to be applied to result before converting it back to QASYMM8_SIGNED
-     * @param[in]  min                          (Optional) Min value used to saturate down the output result before converting back to QASYMM8_SIGNED
-     * @param[in]  max                          (Optional) Max value used to saturate up the output result before converting back to QASYMM8_SIGNED,
-     *                                          Along with @p min, this value can be used to implement "rectified linear unit" activation functions
-     */
-    void configure(const ITensor *input, const ITensor *bias, ITensor *output, int result_fixedpoint_multiplier, int result_shift, int result_offset_after_shift, int min = 0, int max = 0);
-    /** Static function to check if given info will lead to a valid configuration of @ref NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel
-     *
-     * @param[in] input  Input tensor. Data type supported: S32
-     * @param[in] bias   Biases tensor. Only shared biases supported and it can be a nullptr if the biases addition is not required.
-     *                   Biases are 1D tensor with dimensions [OFM]. Data type supported: Same as @p input.
-     * @param[in] output Output tensor. Data type supported: Data type supported: QASYMM8_SIGNED
-     * @param[in] min    (Optional) Min value used to saturate down the output result before converting back to QASYMM8_SIGNED
-     * @param[in] max    (Optional) Max value used to saturate up the output result before converting back to QASYMM8_SIGNED,
-     *                            Along with @p min, this value can be used to implement "rectified linear unit" activation functions
-     *
-     * @return a status
-     */
-    static Status validate(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, int min = 0, int max = 0);
-
-    // Inherited methods overridden:
-    void run(const Window &window, const ThreadInfo &info) override;
-
-private:
-    /** Template function to run the NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel
-     *
-     * @param[in] window Region on which to execute the kernel. (Must be a valid region of the window returned by window()).
-     */
-    template <bool is_bounded_relu>
-    void run(const Window &window);
-
-    /** Common signature for all the specialised NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel functions
-     *
-     * @param[in] window Region on which to execute the kernel.
-     */
-    using QuantizeDownFunctionPtr = void (NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel::*)(const Window &window);
-
-    QuantizeDownFunctionPtr _func;
-    const ITensor          *_input;
-    const ITensor          *_bias;
-    ITensor                *_output;
-    int                     _result_fixedpoint_multiplier;
-    int                     _result_shift;
-    int                     _result_offset_after_shift;
-    int                     _min;
-    int                     _max;
-};
-} // namespace arm_compute
-#endif /* ARM_COMPUTE_NEGEMMLOWPQUANTIZEDOWNINT32TOINT8SCALEBYFIXEDPOINTKERNEL_H */
diff --git a/src/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel.h b/src/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel.h
deleted file mode 100644
index def0573967..0000000000
--- a/src/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel.h
+++ /dev/null
@@ -1,121 +0,0 @@
-/*
- * Copyright (c) 2017-2021 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_NEGEMMLOWPQUANTIZEDOWNINT32TOUINT8SCALEBYFIXEDPOINTKERNEL_H
-#define ARM_COMPUTE_NEGEMMLOWPQUANTIZEDOWNINT32TOUINT8SCALEBYFIXEDPOINTKERNEL_H
-
-#include "src/core/NEON/INEKernel.h"
-
-namespace arm_compute
-{
-class ITensor;
-
-/** Kernel used to quantize down the int32 accumulator values of GEMMLowp to QASYMM8
- *
- * This kernel takes a final int32 accumulator value (the output of @ref NEGEMMLowpMatrixMultiplyKernel), and processes it to obtain the final QASYMM8 value.
- * The following computations will be performed by the kernel:
- *
- *  -# Compute fixed point multiplication between each entry of input by result_fixedpoint_multiplier
- *  -# Add bias to final result if bias tensor is not a nullptr
- *  -# Round to nearest division by a power-of-two using result_shift
- *  -# Add offset to each result
- *  -# Clamp the value between the specified min and max bounds
- *  -# Clamp the resulting int32 values to the [0..255] range and cast to QASYMM8.
- *
- */
-class NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel : public INEKernel
-{
-public:
-    const char *name() const override
-    {
-        return "NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel";
-    }
-    /** Constructor */
-    NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel();
-    /** Prevent instances of this class from being copied (As this class contains pointers)*/
-    NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel(const NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel &) = delete;
-    /** Prevent instances of this class from being copied (As this class contains pointers)*/
-    NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel &operator=(const NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel &) = delete;
-    /** Allow instances of this class to be moved */
-    NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel(NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel &&) = default;
-    /** Allow instances of this class to be moved */
-    NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel &operator=(NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel &&) = default;
-    /** Default destructor */
-    ~NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel() = default;
-    /** Initialise the kernel's input and output.
-     *
-     * @param[in]  input                        Input tensor. Data type supported: S32
-     * @param[in]  bias                         Biases tensor. Only shared biases supported and it can be a nullptr if the biases addition is not required.
-     *                                          Biases are 1D tensor with dimensions [OFM]. Data type supported: Same as @p input.
-     * @param[out] output                       Output tensor. Data type supported: Data type supported: QASYMM8
-     * @param[in]  result_fixedpoint_multiplier Fixed point value to be multiplied to each element of the input matrix when once the result_offset has been add
-     * @param[in]  result_shift                 Integer value used to round to nearest division by a power-of-two the result after the fixed point multiplication
-     * @param[in]  result_offset_after_shift    Offset to be applied to result before converting it back to QASYMM8
-     * @param[in]  min                          (Optional) Min value used to saturate down the output result before converting back to QASYMM8
-     * @param[in]  max                          (Optional) Max value used to saturate up the output result before converting back to QASYMM8,
-     *                                          Along with @p min, this value can be used to implement "rectified linear unit" activation functions
-     */
-    void configure(const ITensor *input, const ITensor *bias, ITensor *output, int result_fixedpoint_multiplier, int result_shift, int result_offset_after_shift, int min = 0, int max = 0);
-    /** Static function to check if given info will lead to a valid configuration of @ref NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel
-     *
-     * @param[in] input  Input tensor. Data type supported: S32
-     * @param[in] bias   Biases tensor. Only shared biases supported and it can be a nullptr if the biases addition is not required.
-     *                   Biases are 1D tensor with dimensions [OFM]. Data type supported: Same as @p input.
-     * @param[in] output Output tensor. Data type supported: Data type supported: QASYMM8
-     * @param[in] min    (Optional) Min value used to saturate down the output result before converting back to QASYMM8
-     * @param[in] max    (Optional) Max value used to saturate up the output result before converting back to QASYMM8,
-     *                            Along with @p min, this value can be used to implement "rectified linear unit" activation functions
-     *
-     * @return a status
-     */
-    static Status validate(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, int min = 0, int max = 0);
-
-    // Inherited methods overridden:
-    void run(const Window &window, const ThreadInfo &info) override;
-
-private:
-    /** Template function to run the NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel
-     *
-     * @param[in] window Region on which to execute the kernel. (Must be a valid region of the window returned by window()).
-     */
-    template <bool is_bounded_relu>
-    void run(const Window &window);
-
-    /** Common signature for all the specialised NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel functions
-     *
-     * @param[in] window Region on which to execute the kernel.
-     */
-    using QuantizeDownFunctionPtr = void (NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel::*)(const Window &window);
-
-    QuantizeDownFunctionPtr _func;
-    const ITensor          *_input;
-    const ITensor          *_bias;
-    ITensor                *_output;
-    int                     _result_fixedpoint_multiplier;
-    int                     _result_shift;
-    int                     _result_offset_after_shift;
-    int                     _min;
-    int                     _max;
-};
-} // namespace arm_compute
-#endif /* ARM_COMPUTE_NEGEMMLOWPQUANTIZEDOWNINT32TOUINT8SCALEBYFIXEDPOINTKERNEL_H */
diff --git a/src/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ScaleKernel.cpp b/src/core/cpu/kernels/CpuGemmLowpQuantizeDownInt32ScaleKernel.cpp
index 84365ba25b..f1c797244a 100644
--- a/src/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ScaleKernel.cpp
+++ b/src/core/cpu/kernels/CpuGemmLowpQuantizeDownInt32ScaleKernel.cpp
@@ -21,7 +21,7 @@
  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  * SOFTWARE.
  */
-#include "src/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ScaleKernel.h"
+#include "src/core/cpu/kernels/CpuGemmLowpQuantizeDownInt32ScaleKernel.h"
 
 #include "arm_compute/core/Error.h"
 #include "arm_compute/core/Helpers.h"
@@ -37,14 +37,19 @@
 #include "src/core/helpers/WindowHelpers.h"
 
 #include <arm_neon.h>
-#include <cstddef>
-#include <cstdint>
 
 namespace arm_compute
 {
-Status validate_arguments(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, const GEMMLowpOutputStageInfo *output_stage)
+namespace cpu
 {
-    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::S32);
+namespace kernels
+{
+namespace
+{
+Status validate_arguments(const ITensorInfo *src, const ITensorInfo *bias, const ITensorInfo *dst, const GEMMLowpOutputStageInfo *output_stage)
+{
+    ARM_COMPUTE_ERROR_ON_NULLPTR(src, dst);
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src, 1, DataType::S32);
 
     ARM_COMPUTE_RETURN_ERROR_ON(output_stage->gemmlowp_max_bound > std::get<1>(quantization::get_min_max_values_from_quantized_data_type(output_stage->output_data_type)));
     ARM_COMPUTE_RETURN_ERROR_ON(output_stage->gemmlowp_min_bound < std::get<0>(quantization::get_min_max_values_from_quantized_data_type(output_stage->output_data_type))
@@ -53,19 +58,19 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *bias, con
     // Check biases if exist
     if(bias != nullptr)
     {
-        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, bias);
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src, bias);
         ARM_COMPUTE_RETURN_ERROR_ON(bias->num_dimensions() > 1);
-        ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(0) != bias->dimension(0));
+        ARM_COMPUTE_RETURN_ERROR_ON(src->dimension(0) != bias->dimension(0));
     }
 
-    if(output->total_size() != 0)
+    if(dst->total_size() != 0)
     {
-        if(output->data_type() != output_stage->output_data_type && (output_stage->output_data_type == DataType::QASYMM8 || output_stage->output_data_type == DataType::QASYMM8_SIGNED))
+        if(dst->data_type() != output_stage->output_data_type && (output_stage->output_data_type == DataType::QASYMM8 || output_stage->output_data_type == DataType::QASYMM8_SIGNED))
         {
             ARM_COMPUTE_RETURN_ERROR_MSG("Mismatching data types");
         }
 
-        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input, output);
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(src, dst);
     }
 
     return Status{};
@@ -129,11 +134,10 @@ inline typename wrapper::traits::neon_vector<T, 16>::type finalize_quantization(
 
     return out;
 }
-
-class Coordinates;
+} // namespace
 
 template <typename T>
-void NEGEMMLowpQuantizeDownInt32ScaleKernel::run(const Window &window)
+void CpuGemmLowpQuantizeDownInt32ScaleKernel::run_internal(const ITensor *src, const ITensor *bias, ITensor *dst, const Window &window)
 {
     using VectorType = typename wrapper::traits::neon_vector<T, 16>::type;
 
@@ -152,16 +156,16 @@ void NEGEMMLowpQuantizeDownInt32ScaleKernel::run(const Window &window)
     Window win(window);
     win.set(Window::DimX, Window::Dimension(0, 1, 1));
 
-    Iterator in(_input, win);
-    Iterator out(_output, win);
+    Iterator in(src, win);
+    Iterator out(dst, win);
 
-    if(_bias != nullptr)
+    if(bias != nullptr)
     {
         Window win_biases;
         win_biases.set(Window::DimX, Window::Dimension(0, 1, 1));
         win_biases.set(Window::DimY, Window::Dimension(0, 1, 1));
 
-        Iterator bias(_bias, win_biases);
+        Iterator bias_i(bias, win_biases);
         execute_window_loop(win, [&](const Coordinates &)
         {
             // Compute 16 elements per iteration
@@ -181,10 +185,10 @@ void NEGEMMLowpQuantizeDownInt32ScaleKernel::run(const Window &window)
                 const int32x4x4_t bias_s32 =
                 {
                     {
-                        vld1q_s32(reinterpret_cast<const int32_t *>(bias.ptr()) + x + 0),
-                        vld1q_s32(reinterpret_cast<const int32_t *>(bias.ptr()) + x + 4),
-                        vld1q_s32(reinterpret_cast<const int32_t *>(bias.ptr()) + x + 8),
-                        vld1q_s32(reinterpret_cast<const int32_t *>(bias.ptr()) + x + 12)
+                        vld1q_s32(reinterpret_cast<const int32_t *>(bias_i.ptr()) + x + 0),
+                        vld1q_s32(reinterpret_cast<const int32_t *>(bias_i.ptr()) + x + 4),
+                        vld1q_s32(reinterpret_cast<const int32_t *>(bias_i.ptr()) + x + 8),
+                        vld1q_s32(reinterpret_cast<const int32_t *>(bias_i.ptr()) + x + 12)
                     }
                 };
 
@@ -203,7 +207,7 @@ void NEGEMMLowpQuantizeDownInt32ScaleKernel::run(const Window &window)
             // Compute left-over elements
             for(; x < window_end_x; ++x)
             {
-                const int bias_value = *(reinterpret_cast<const int *>(bias.ptr()) + x);
+                const int bias_value = *(reinterpret_cast<const int *>(bias_i.ptr()) + x);
                 int       in_value   = *(reinterpret_cast<const int *>(in.ptr()) + x);
 
                 // Quantize
@@ -213,7 +217,7 @@ void NEGEMMLowpQuantizeDownInt32ScaleKernel::run(const Window &window)
                 *(out.ptr() + x) = static_cast<T>(utility::clamp<int>(in_value, clamp_min, clamp_max));
             }
         },
-        in, bias, out);
+        in, bias_i, out);
     }
     else
     {
@@ -255,33 +259,26 @@ void NEGEMMLowpQuantizeDownInt32ScaleKernel::run(const Window &window)
     }
 }
 
-NEGEMMLowpQuantizeDownInt32ScaleKernel::NEGEMMLowpQuantizeDownInt32ScaleKernel()
-    : _func(nullptr), _input(nullptr), _bias(nullptr), _output(nullptr), _output_stage(nullptr), _is_bounded_relu(false)
-{
-}
-
-void NEGEMMLowpQuantizeDownInt32ScaleKernel::configure(const ITensor *input, const ITensor *bias, ITensor *output, const GEMMLowpOutputStageInfo *output_stage)
+void CpuGemmLowpQuantizeDownInt32ScaleKernel::configure(ITensorInfo *src, ITensorInfo *bias, ITensorInfo *dst, const GEMMLowpOutputStageInfo *output_stage)
 {
+    ARM_COMPUTE_UNUSED(bias);
     // Perform validate step
-    ARM_COMPUTE_ERROR_ON_NULLPTR(input, output, output_stage);
+    ARM_COMPUTE_ERROR_ON_NULLPTR(src, dst, output_stage);
 
     // Output auto inizialitation if not yet initialized
-    auto_init_if_empty(*output->info(), input->info()->clone()->set_data_type(output_stage->output_data_type));
+    auto_init_if_empty(*dst, src->clone()->set_data_type(output_stage->output_data_type));
 
-    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(),
-                                                  (bias != nullptr) ? bias->info() : nullptr,
-                                                  output->info(),
+    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(src,
+                                                  bias,
+                                                  dst,
                                                   output_stage));
 
-    _input        = input;
-    _bias         = bias;
-    _output       = output;
     _output_stage = output_stage;
 
     // Configure kernel window
-    Window win = calculate_max_window(*input->info(), Steps());
+    Window win = calculate_max_window(*src, Steps());
 
-    INEKernel::configure(win);
+    ICpuKernel::configure(win);
 
     // Check if we need to clamp the result using min and max
     _is_bounded_relu = ((_output_stage->gemmlowp_min_bound != _output_stage->gemmlowp_max_bound)
@@ -289,11 +286,11 @@ void NEGEMMLowpQuantizeDownInt32ScaleKernel::configure(const ITensor *input, con
                              && _output_stage->gemmlowp_max_bound == std::get<1>(quantization::get_min_max_values_from_quantized_data_type(output_stage->output_data_type))));
     if(_output_stage->output_data_type == DataType::QASYMM8)
     {
-        _func = &NEGEMMLowpQuantizeDownInt32ScaleKernel::run<uint8_t>;
+        _func = &CpuGemmLowpQuantizeDownInt32ScaleKernel::run_internal<uint8_t>;
     }
     else if(_output_stage->output_data_type == DataType::QASYMM8_SIGNED)
     {
-        _func = &NEGEMMLowpQuantizeDownInt32ScaleKernel::run<int8_t>;
+        _func = &CpuGemmLowpQuantizeDownInt32ScaleKernel::run_internal<int8_t>;
     }
     else
     {
@@ -301,20 +298,29 @@ void NEGEMMLowpQuantizeDownInt32ScaleKernel::configure(const ITensor *input, con
     }
 }
 
-Status NEGEMMLowpQuantizeDownInt32ScaleKernel::validate(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, const GEMMLowpOutputStageInfo *output_stage)
+Status CpuGemmLowpQuantizeDownInt32ScaleKernel::validate(const ITensorInfo *src, const ITensorInfo *bias, const ITensorInfo *dst, const GEMMLowpOutputStageInfo *output_stage)
 {
-    ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
-    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, bias, output, output_stage));
-
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(src, bias, dst, output_stage));
     return Status{};
 }
 
-void NEGEMMLowpQuantizeDownInt32ScaleKernel::run(const Window &window, const ThreadInfo &info)
+void CpuGemmLowpQuantizeDownInt32ScaleKernel::run_op(ITensorPack &tensors, const Window &window, const ThreadInfo &info)
 {
     ARM_COMPUTE_UNUSED(info);
     ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
-    ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window);
+    ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICpuKernel::window(), window);
+    ARM_COMPUTE_ERROR_ON_MSG(tensors.empty(), "No inputs provided");
 
-    (this->*_func)(window);
+    auto src  = tensors.get_const_tensor(TensorType::ACL_SRC);
+    auto bias = tensors.get_const_tensor(TensorType::ACL_BIAS);
+    auto dst  = tensors.get_tensor(TensorType::ACL_DST);
+    (this->*_func)(src, bias, dst, window);
+}
+
+const char *CpuGemmLowpQuantizeDownInt32ScaleKernel::name() const
+{
+    return "CpuGemmLowpQuantizeDownInt32ScaleKernel";
 }
+} // namespace kernels
+} // namespace cpu
 } // namespace arm_compute
\ No newline at end of file
diff --git a/src/core/cpu/kernels/CpuGemmLowpQuantizeDownInt32ScaleKernel.h b/src/core/cpu/kernels/CpuGemmLowpQuantizeDownInt32ScaleKernel.h
new file mode 100644
index 0000000000..646242dc7e
--- /dev/null
+++ b/src/core/cpu/kernels/CpuGemmLowpQuantizeDownInt32ScaleKernel.h
@@ -0,0 +1,107 @@
+/*
+ * Copyright (c) 2020-2021 Arm Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#ifndef ARM_COMPUTE_CPU_GEMMLOWP_QUANTIZEDOWNINT32_SCALE_KERNEL_H
+#define ARM_COMPUTE_CPU_GEMMLOWP_QUANTIZEDOWNINT32_SCALE_KERNEL_H
+
+#include "arm_compute/core/KernelDescriptors.h"
+#include "src/core/common/Macros.h"
+#include "src/core/cpu/ICpuKernel.h"
+
+namespace arm_compute
+{
+class ITensor;
+namespace cpu
+{
+namespace kernels
+{
+/** Kernel used to quantize down the int32 accumulator values of GEMMLowp to QASYMM8/QASYMM8_SIGNED
+ *
+ * This kernel takes a final int32 accumulator value (the output of @ref NEGEMMLowpMatrixMultiplyKernel), and processes it to obtain the final QASYMM8/QASYMM8_SIGNED value.
+ * The following computations will be performed by the kernel:
+ *
+ *  -# Add offset terms to final result
+ *  -# Multiply each entry of result by result_mult_int
+ *  -# Add bias to final result if bias tensor is not a nullptr
+ *  -# Shift the int32 accumulator by result_shift
+ *  -# Clamp the value between the specified min and max bounds
+ *  -# Clamp the resulting int32 values:
+ *  -#  -to the [0..255] range and cast to QASYMM8.
+ *  -#  -to the [-128..127] range and cast to QASYMM8_SIGNED.
+ *
+ */
+class CpuGemmLowpQuantizeDownInt32ScaleKernel : public ICpuKernel
+{
+public:
+    /** Default constructor */
+    CpuGemmLowpQuantizeDownInt32ScaleKernel() = default;
+    ARM_COMPUTE_DISALLOW_COPY_ALLOW_MOVE(CpuGemmLowpQuantizeDownInt32ScaleKernel);
+    /** Initialise the kernel's input and output.
+     *
+     * @param[in]  src          Input tensor info. Data type supported: S32
+     * @param[in]  bias         Biases tensor info. Only shared biases supported and it can be a nullptr if the biases addition is not required.
+     *                          Biases are 1D tensor with dimensions [OFM]. Data type supported: Same as @p input.
+     * @param[out] dst          Output tensor info. Data type supported: Data type supported: QASYMM8/QASYMM8_SIGNED
+     * @param[out] output_stage GEMMLowp output stage metadata.
+     */
+    void configure(ITensorInfo *src, ITensorInfo *bias, ITensorInfo *dst, const GEMMLowpOutputStageInfo *output_stage);
+    /** Static function to check if given info will lead to a valid configuration
+     *
+     * Similar to CpuGemmLowpQuantizeDownInt32ScaleKernel::configure()
+     *
+     * @return a status
+     */
+    static Status validate(const ITensorInfo *src, const ITensorInfo *bias, const ITensorInfo *dst, const GEMMLowpOutputStageInfo *output_stage);
+
+    // Inherited methods overridden:
+    void run_op(ITensorPack &tensors, const Window &window, const ThreadInfo &info) override;
+    const char *name() const override;
+
+private:
+    /** Template function to run the NEGEMMLowpQuantizeDownInt32ScaleKernel
+     *
+     * @param[in]  src    Input tensor info
+     * @param[in]  bias   Biases tensor info
+     * @param[out] dst    Output tensor info
+     * @param[in]  window Region on which to execute the kernel. (Must be a valid region of the window returned by window())
+     */
+    template <typename T>
+    void run_internal(const ITensor *src, const ITensor *bias, ITensor *dst, const Window &window);
+
+    /** Common signature for all the specialised CpuGemmLowpQuantizeDownInt32ScaleKernel functions
+     *
+     * @param[in]  src    Input tensor info
+     * @param[in]  bias   Biases tensor info
+     * @param[out] dst    Output tensor info
+     * @param[in]  window Region on which to execute the kernel.
+     */
+    using QuantizeDownFunctionPtr = void (CpuGemmLowpQuantizeDownInt32ScaleKernel::*)(const ITensor *src, const ITensor *bias, ITensor *dst, const Window &window);
+
+    QuantizeDownFunctionPtr        _func{ nullptr };
+    const GEMMLowpOutputStageInfo *_output_stage{ nullptr };
+    bool                           _is_bounded_relu{ false };
+};
+} // namespace kernels
+} // namespace cpu
+} // namespace arm_compute
+#endif /* ARM_COMPUTE_CPU_GEMMLOWP_QUANTIZEDOWNINT32_SCALE_KERNEL_H */
diff --git a/src/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel.cpp b/src/core/cpu/kernels/CpuGemmLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel.cpp
index aa54b80436..390e269cbb 100644
--- a/src/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel.cpp
+++ b/src/core/cpu/kernels/CpuGemmLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel.cpp
@@ -21,7 +21,7 @@
  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  * SOFTWARE.
  */
-#include "src/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel.h"
+#include "src/core/cpu/kernels/CpuGemmLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel.h"
 
 #include "arm_compute/core/Error.h"
 #include "arm_compute/core/Helpers.h"
@@ -37,53 +37,41 @@
 #include "src/core/helpers/WindowHelpers.h"
 
 #include <arm_neon.h>
-#include <cstddef>
-#include <cstdint>
 
 namespace arm_compute
 {
+namespace cpu
+{
+namespace kernels
+{
 namespace
 {
-Status validate_arguments(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, int min, int max)
+Status validate_arguments(const ITensorInfo *src, const ITensorInfo *bias, const ITensorInfo *dst, int min, int max)
 {
-    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::S32);
+    ARM_COMPUTE_ERROR_ON_NULLPTR(src, dst);
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src, 1, DataType::S32);
     ARM_COMPUTE_RETURN_ERROR_ON(min > max);
 
     // Check biases if exist
     if(bias != nullptr)
     {
-        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, bias);
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src, bias);
         ARM_COMPUTE_RETURN_ERROR_ON(bias->num_dimensions() > 1);
-        ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(0) != bias->dimension(0));
+        ARM_COMPUTE_RETURN_ERROR_ON(src->dimension(0) != bias->dimension(0));
     }
 
-    if(output->total_size() != 0)
+    if(dst->total_size() != 0)
     {
-        ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::QSYMM16);
-        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(output, input);
+        ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(dst, 1, DataType::QSYMM16);
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(dst, src);
     }
 
     return Status{};
 }
-
-std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *output)
-{
-    // Output auto inizialitation if not yet initialized
-    auto_init_if_empty(*output, input->clone()->set_data_type(DataType::QSYMM16));
-
-    // Configure kernel window
-    Window win = calculate_max_window(*input, Steps());
-
-    // NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel doesn't need padding so update_window_and_padding() can be skipped
-
-    return std::make_pair(Status{}, win);
-}
 } // namespace
 
-class Coordinates;
-
 template <bool is_bounded_relu>
-void NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel::run(const Window &window)
+void CpuGemmLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel::run_internal(const ITensor *src, const ITensor *bias, ITensor *dst, const Window &window)
 {
     const int16x8_t min_s16 = vdupq_n_s16(static_cast<int16_t>(_min));
     const int16x8_t max_s16 = vdupq_n_s16(static_cast<int16_t>(_max));
@@ -98,15 +86,15 @@ void NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel::run(const Window
     Window win_collapsed = window.collapse_if_possible(window, Window::DimZ);
     win_collapsed.set(Window::DimX, Window::Dimension(0, 1, 1));
 
-    Iterator in(_input, win_collapsed);
-    Iterator out(_output, win_collapsed);
-    if(_bias != nullptr)
+    Iterator in(src, win_collapsed);
+    Iterator out(dst, win_collapsed);
+    if(bias != nullptr)
     {
         Window win_biases;
         win_biases.set(Window::DimX, Window::Dimension(0, 1, 1));
         win_biases.set(Window::DimY, Window::Dimension(0, 1, 1));
 
-        Iterator bias(_bias, win_biases);
+        Iterator bias_i(bias, win_biases);
         execute_window_loop(win_collapsed, [&](const Coordinates &)
         {
             // Compute 16 elements per iteration
@@ -124,8 +112,8 @@ void NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel::run(const Window
                 const int32x4x2_t bias_s32 =
                 {
                     {
-                        vld1q_s32(reinterpret_cast<const int32_t *>(bias.ptr()) + x + 0),
-                        vld1q_s32(reinterpret_cast<const int32_t *>(bias.ptr()) + x + 4)
+                        vld1q_s32(reinterpret_cast<const int32_t *>(bias_i.ptr()) + x + 0),
+                        vld1q_s32(reinterpret_cast<const int32_t *>(bias_i.ptr()) + x + 4)
                     }
                 };
 
@@ -139,7 +127,7 @@ void NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel::run(const Window
             // Compute left-over elements
             for(; x < window_end_x; ++x)
             {
-                const int32_t bias_value = *(reinterpret_cast<const int32_t *>(bias.ptr()) + x);
+                const int32_t bias_value = *(reinterpret_cast<const int32_t *>(bias_i.ptr()) + x);
                 int32_t       in_value   = *(reinterpret_cast<const int32_t *>(in.ptr()) + x);
 
                 // Add bias
@@ -149,7 +137,7 @@ void NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel::run(const Window
                                                                                                              static_cast<int16_t>(_max));
             }
         },
-        in, out, bias);
+        in, out, bias_i);
     }
     else
     {
@@ -184,51 +172,56 @@ void NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel::run(const Window
     }
 }
 
-NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel::NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel()
-    : _func(nullptr), _input(nullptr), _bias(nullptr), _output(nullptr), _result_fixedpoint_multiplier(0), _result_shift(0), _min(0), _max(0)
-{
-}
-
-void NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel::configure(const ITensor *input, const ITensor *bias, ITensor *output, int result_fixedpoint_multiplier, int result_shift,
-                                                                          int min, int max)
+void CpuGemmLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel::configure(ITensorInfo *src, ITensorInfo *bias, ITensorInfo *dst, int result_fixedpoint_multiplier, int result_shift,
+                                                                           int min, int max)
 {
     // Perform validate step
-    ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
-    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), (bias != nullptr) ? bias->info() : nullptr, output->info(), min, max));
+    ARM_COMPUTE_UNUSED(bias, dst);
+    ARM_COMPUTE_ERROR_ON_NULLPTR(src, dst);
+    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(src, bias, dst, min, max));
 
-    _input                        = input;
-    _bias                         = bias;
-    _output                       = output;
     _result_fixedpoint_multiplier = result_fixedpoint_multiplier;
     _result_shift                 = result_shift;
     _min                          = min;
     _max                          = max;
 
+    // Output auto inizialitation if not yet initialized
+    auto_init_if_empty(*src, src->clone()->set_data_type(DataType::QSYMM16));
     // Configure kernel window
-    auto win_config = validate_and_configure_window(input->info(), output->info());
-    ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
-    INEKernel::configure(win_config.second);
+    Window win_config = calculate_max_window(*src, Steps());
+    ICpuKernel::configure(win_config);
 
     // Check if we need to clamp the result using min and max
     const bool is_bounded_relu = !(min <= -32768 && max >= 32767);
-    _func                      = is_bounded_relu ? &NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel::run<true> : &NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel::run<false>;
+    _func                      = is_bounded_relu ? &CpuGemmLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel::run_internal<true> :
+                                 &CpuGemmLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel::run_internal<false>;
 }
 
-Status NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel::validate(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, int min, int max)
+Status CpuGemmLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel::validate(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, int min, int max)
 {
     ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
     ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, bias, output, min, max));
-    ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), output->clone().get()).first);
-
     return Status{};
 }
 
-void NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel::run(const Window &window, const ThreadInfo &info)
+void CpuGemmLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel::run_op(ITensorPack &tensors, const Window &window, const ThreadInfo &info)
 {
     ARM_COMPUTE_UNUSED(info);
     ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
-    ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window);
+    ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICpuKernel::window(), window);
+    ARM_COMPUTE_ERROR_ON_MSG(tensors.empty(), "No inputs provided");
+
+    auto src  = tensors.get_const_tensor(TensorType::ACL_SRC);
+    auto bias = tensors.get_const_tensor(TensorType::ACL_BIAS);
+    auto dst  = tensors.get_tensor(TensorType::ACL_DST);
 
-    (this->*_func)(window);
+    (this->*_func)(src, bias, dst, window);
+}
+
+const char *CpuGemmLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel::name() const
+{
+    return "CpuGemmLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel";
 }
+} // namespace kernels
+} // namespace cpu
 } // namespace arm_compute
diff --git a/src/core/cpu/kernels/CpuGemmLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel.h b/src/core/cpu/kernels/CpuGemmLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel.h
new file mode 100644
index 0000000000..4d743e9a0a
--- /dev/null
+++ b/src/core/cpu/kernels/CpuGemmLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel.h
@@ -0,0 +1,111 @@
+/*
+ * Copyright (c) 2019-2021 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_GEMMLOWP_QUANTIZEDOWNINT32TOINT16_SCALEBYFIXEDPOINT_KERNEL_H
+#define ARM_COMPUTE_CPU_GEMMLOWP_QUANTIZEDOWNINT32TOINT16_SCALEBYFIXEDPOINT_KERNEL_H
+
+#include "arm_compute/core/KernelDescriptors.h"
+#include "src/core/common/Macros.h"
+#include "src/core/cpu/ICpuKernel.h"
+
+namespace arm_compute
+{
+class ITensor;
+namespace cpu
+{
+namespace kernels
+{
+/** Kernel used to quantize down the int32 accumulator values of GEMMLowp to QSYMM16
+ *
+ * This kernel takes a final int32 accumulator value (the output of @ref NEGEMMLowpMatrixMultiplyKernel), and processes it to obtain the final QSYMM16 value.
+ * The following computations will be performed by the kernel:
+ *
+ *  -# Compute fixed point multiplication between each entry of input by result_fixedpoint_multiplier
+ *  -# Add bias to final result if bias tensor is not a nullptr
+ *  -# Round to nearest division by a power-of-two using result_shift
+ *  -# Clamp the value between the specified min and max bounds
+ *  -# Clamp the resulting int32 values to the [-32768, 32767] range and cast to QSYMM16.
+ *
+ */
+class CpuGemmLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel : public ICpuKernel
+{
+public:
+    /** Default constructor */
+    CpuGemmLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel() = default;
+    ARM_COMPUTE_DISALLOW_COPY_ALLOW_MOVE(CpuGemmLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel);
+    /** Initialise the kernel's input and output.
+     *
+     * @param[in]  src                          Input tensor info. Data type supported: S32
+     * @param[in]  bias                         Biases tensor info. Only shared biases supported and it can be a nullptr if the biases addition is not required.
+     *                                          Biases are 1D tensor with dimensions [OFM]. Data type supported: Same as @p input.
+     * @param[out] dst                          Output tensor info. Data type supported: Data type supported: QSYMM16
+     * @param[in]  result_fixedpoint_multiplier Fixed point value to be multiplied to each element of the input matrix when once the result_offset has been add
+     * @param[in]  result_shift                 Integer value used to round to nearest division by a power-of-two the result after the fixed point multiplication
+     * @param[in]  min                          (Optional) Min value used to saturate down the output result before converting back to QSYMM16. Defaults to 0.
+     * @param[in]  max                          (Optional) Max value used to saturate up the output result before converting back to QSYMM16.
+     *                                          Along with @p min, this value can be used to implement "rectified linear unit" activation functions. Defaults to 0.
+     */
+    void configure(ITensorInfo *src, ITensorInfo *bias, ITensorInfo *dst, int result_fixedpoint_multiplier, int result_shift, int min = 0, int max = 0);
+    /** Static function to check if given info will lead to a valid configuration
+     *
+     * Similar to CpuGemmLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel::configure()
+     *
+     * @return a status
+     */
+    static Status validate(const ITensorInfo *src, const ITensorInfo *bias, const ITensorInfo *dst, int min = 0, int max = 0);
+
+    // Inherited methods overridden:
+    void run_op(ITensorPack &tensors, const Window &window, const ThreadInfo &info) override;
+    const char *name() const override;
+
+private:
+    /** Template function to run the CpuGemmLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel
+     *
+     * @param[in]  src    Input tensor info
+     * @param[in]  bias   Bias tensor info
+     * @param[out] dst    Output tensor info
+     * @param[in]  window Region on which to execute the kernel. (Must be a valid region of the window returned by window()).
+     */
+    template <bool is_bounded_relu>
+    void run_internal(const ITensor *src, const ITensor *bias, ITensor *dst, const Window &window);
+
+    /** Common signature for all the specialised CpuGemmLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel functions
+     *
+     * @param[in]  src    Input tensor info
+     * @param[in]  bias   Bias tensor info
+     * @param[out] dst    Output tensor info
+     * @param[in]  window Region on which to execute the kernel.
+     */
+    using QuantizeDownFunctionPtr = void (CpuGemmLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel::*)(
+                                        const ITensor *src, const ITensor *bias, ITensor *dst, const Window &window);
+
+    QuantizeDownFunctionPtr _func{ nullptr };
+    int                     _result_fixedpoint_multiplier{ 0 };
+    int                     _result_shift{ 0 };
+    int                     _min{ 0 };
+    int                     _max{ 0 };
+};
+} // namespace kernels
+} // namespace cpu
+} // namespace arm_compute
+#endif /* ARM_COMPUTE_CPU_GEMMLOWP_QUANTIZEDOWNINT32TOINT16_SCALEBYFIXEDPOINT_KERNEL_H */
diff --git a/src/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel.cpp b/src/core/cpu/kernels/CpuGemmLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel.cpp
index 9ed85e62aa..318b6a06f8 100644
--- a/src/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel.cpp
+++ b/src/core/cpu/kernels/CpuGemmLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel.cpp
@@ -21,7 +21,7 @@
  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  * SOFTWARE.
  */
-#include "src/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel.h"
+#include "src/core/cpu/kernels/CpuGemmLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel.h"
 
 #include "arm_compute/core/Error.h"
 #include "arm_compute/core/Helpers.h"
@@ -37,51 +37,41 @@
 #include "src/core/helpers/WindowHelpers.h"
 
 #include <arm_neon.h>
-#include <cstddef>
-#include <cstdint>
 
 namespace arm_compute
 {
+namespace cpu
+{
+namespace kernels
+{
 namespace
 {
-Status validate_arguments(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, int min, int max)
+Status validate_arguments(const ITensorInfo *src, const ITensorInfo *bias, const ITensorInfo *dst, int min, int max)
 {
-    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::S32);
+    ARM_COMPUTE_ERROR_ON_NULLPTR(src, dst);
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src, 1, DataType::S32);
     ARM_COMPUTE_RETURN_ERROR_ON(min > max);
 
     // Check biases if exist
     if(bias != nullptr)
     {
-        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, bias);
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src, bias);
         ARM_COMPUTE_RETURN_ERROR_ON(bias->num_dimensions() > 1);
-        ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(0) != bias->dimension(0));
+        ARM_COMPUTE_RETURN_ERROR_ON(src->dimension(0) != bias->dimension(0));
     }
 
-    if(output->total_size() != 0)
+    if(dst->total_size() != 0)
     {
-        ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::QASYMM8_SIGNED);
-        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(output, input);
+        ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(dst, 1, DataType::QASYMM8_SIGNED);
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(dst, src);
     }
 
     return Status{};
 }
-
-std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *output)
-{
-    // Output auto initialization if not yet initialized
-    auto_init_if_empty(*output, input->clone()->set_data_type(DataType::QASYMM8_SIGNED));
-
-    // Configure kernel window
-    Window win = calculate_max_window(*input, Steps());
-
-    // NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel doesn't need padding so update_window_and_padding() can be skipped
-
-    return std::make_pair(Status{}, win);
-}
 } // namespace
 
 template <bool is_bounded_relu>
-void NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel::run(const Window &window)
+void CpuGemmLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel::run_internal(const ITensor *src, const ITensor *bias, ITensor *dst, const Window &window)
 {
     const int32x4_t result_offset_after_shift_s32 = vdupq_n_s32(_result_offset_after_shift);
     const int8x16_t min_s8                        = vdupq_n_s8(static_cast<int8_t>(_min));
@@ -96,15 +86,15 @@ void NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel::run(const Window
     Window win_collapsed = window.collapse_if_possible(window, Window::DimZ);
     win_collapsed.set(Window::DimX, Window::Dimension(0, 1, 1));
 
-    Iterator in(_input, win_collapsed);
-    Iterator out(_output, win_collapsed);
-    if(_bias != nullptr)
+    Iterator in(src, win_collapsed);
+    Iterator out(dst, win_collapsed);
+    if(bias != nullptr)
     {
         Window win_biases;
         win_biases.set(Window::DimX, Window::Dimension(0, 1, 1));
         win_biases.set(Window::DimY, Window::Dimension(0, 1, 1));
 
-        Iterator bias(_bias, win_biases);
+        Iterator bias_i(bias, win_biases);
         execute_window_loop(win_collapsed, [&](const Coordinates &)
         {
             // Compute 16 elements per iteration
@@ -124,10 +114,10 @@ void NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel::run(const Window
                 const int32x4x4_t bias_s32 =
                 {
                     {
-                        vld1q_s32(reinterpret_cast<const int32_t *>(bias.ptr()) + x + 0),
-                        vld1q_s32(reinterpret_cast<const int32_t *>(bias.ptr()) + x + 4),
-                        vld1q_s32(reinterpret_cast<const int32_t *>(bias.ptr()) + x + 8),
-                        vld1q_s32(reinterpret_cast<const int32_t *>(bias.ptr()) + x + 12)
+                        vld1q_s32(reinterpret_cast<const int32_t *>(bias_i.ptr()) + x + 0),
+                        vld1q_s32(reinterpret_cast<const int32_t *>(bias_i.ptr()) + x + 4),
+                        vld1q_s32(reinterpret_cast<const int32_t *>(bias_i.ptr()) + x + 8),
+                        vld1q_s32(reinterpret_cast<const int32_t *>(bias_i.ptr()) + x + 12)
                     }
                 };
 
@@ -144,7 +134,7 @@ void NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel::run(const Window
             // Compute left-over elements
             for(; x < window_end_x; ++x)
             {
-                const int32_t bias_value = *(reinterpret_cast<const int32_t *>(bias.ptr()) + x);
+                const int32_t bias_value = *(reinterpret_cast<const int32_t *>(bias_i.ptr()) + x);
                 int32_t       in_value   = *(reinterpret_cast<const int32_t *>(in.ptr()) + x);
 
                 // Add bias
@@ -154,7 +144,7 @@ void NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel::run(const Window
                                                                                    static_cast<int8_t>(_min), static_cast<int8_t>(_max), is_bounded_relu);
             }
         },
-        in, out, bias);
+        in, out, bias_i);
     }
     else
     {
@@ -192,52 +182,58 @@ void NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel::run(const Window
     }
 }
 
-NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel::NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel()
-    : _func(nullptr), _input(nullptr), _bias(nullptr), _output(nullptr), _result_fixedpoint_multiplier(0), _result_shift(0), _result_offset_after_shift(0), _min(0), _max(0)
-{
-}
-
-void NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel::configure(const ITensor *input, const ITensor *bias, ITensor *output, int result_fixedpoint_multiplier, int result_shift,
-                                                                         int result_offset_after_shift, int min, int max)
+void CpuGemmLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel::configure(ITensorInfo *src, ITensorInfo *bias, ITensorInfo *dst, int result_fixedpoint_multiplier, int result_shift,
+                                                                          int result_offset_after_shift, int min, int max)
 {
+    ARM_COMPUTE_UNUSED(bias);
     // Perform validate step
-    ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
-    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), (bias != nullptr) ? bias->info() : nullptr, output->info(), min, max));
+    ARM_COMPUTE_ERROR_ON_NULLPTR(src, dst);
+    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(src, bias, dst, min, max));
 
-    _input                        = input;
-    _bias                         = bias;
-    _output                       = output;
     _result_fixedpoint_multiplier = result_fixedpoint_multiplier;
     _result_shift                 = result_shift;
     _result_offset_after_shift    = result_offset_after_shift;
     _min                          = min;
     _max                          = max;
 
+    // Output auto initialization if not yet initialized
+    auto_init_if_empty(*dst, src->clone()->set_data_type(DataType::QASYMM8_SIGNED));
+
     // Configure kernel window
-    auto win_config = validate_and_configure_window(input->info(), output->info());
-    ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
-    INEKernel::configure(win_config.second);
+    Window win_config = calculate_max_window(*src, Steps());
+    ICpuKernel::configure(win_config);
 
     // Check if we need to clamp the result using min and max
     const bool is_bounded_relu = !(min <= -128 && max >= 127);
-    _func                      = is_bounded_relu ? &NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel::run<true> : &NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel::run<false>;
+    _func                      = is_bounded_relu ? &CpuGemmLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel::run_internal<true> :
+                                 &CpuGemmLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel::run_internal<false>;
 }
 
-Status NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel::validate(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, int min, int max)
+Status CpuGemmLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel::validate(const ITensorInfo *src, const ITensorInfo *bias, const ITensorInfo *dst, int min, int max)
 {
-    ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
-    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, bias, output, min, max));
-    ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), output->clone().get()).first);
-
+    ARM_COMPUTE_ERROR_ON_NULLPTR(src, dst);
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(src, bias, dst, min, max));
     return Status{};
 }
 
-void NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel::run(const Window &window, const ThreadInfo &info)
+void CpuGemmLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel::run_op(ITensorPack &tensors, const Window &window, const ThreadInfo &info)
 {
     ARM_COMPUTE_UNUSED(info);
     ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
-    ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window);
+    ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICpuKernel::window(), window);
+    ARM_COMPUTE_ERROR_ON_MSG(tensors.empty(), "No inputs provided");
+
+    auto src  = tensors.get_const_tensor(TensorType::ACL_SRC);
+    auto bias = tensors.get_const_tensor(TensorType::ACL_BIAS);
+    auto dst  = tensors.get_tensor(TensorType::ACL_DST);
 
-    (this->*_func)(window);
+    (this->*_func)(src, bias, dst, window);
+}
+
+const char *CpuGemmLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel::name() const
+{
+    return "CpuGemmLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel";
 }
+} // namespace kernels
+} // namespace cpu
 } // namespace arm_compute
diff --git a/src/core/cpu/kernels/CpuGemmLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel.h b/src/core/cpu/kernels/CpuGemmLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel.h
new file mode 100644
index 0000000000..a941f1f542
--- /dev/null
+++ b/src/core/cpu/kernels/CpuGemmLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel.h
@@ -0,0 +1,114 @@
+/*
+ * Copyright (c) 2019-2021 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_GEMMLOWP_QUANTIZEDOWNINT32TOINT8_SCALEBYFIXEDPOINT_KERNEL_H
+#define ARM_COMPUTE_CPU_GEMMLOWP_QUANTIZEDOWNINT32TOINT8_SCALEBYFIXEDPOINT_KERNEL_H
+
+#include "arm_compute/core/KernelDescriptors.h"
+#include "src/core/common/Macros.h"
+#include "src/core/cpu/ICpuKernel.h"
+
+namespace arm_compute
+{
+class ITensor;
+namespace cpu
+{
+namespace kernels
+{
+/** Kernel used to quantize down the int32 accumulator values of GEMMLowp to QASYMM8_SIGNED
+ *
+ * This kernel takes a final int32 accumulator value (the output of @ref NEGEMMLowpMatrixMultiplyKernel), and processes it to obtain the final QASYMM8_SIGNED value.
+ * The following computations will be performed by the kernel:
+ *
+ *  -# Compute fixed point multiplication between each entry of input by result_fixedpoint_multiplier
+ *  -# Add bias to final result if bias tensor is not a nullptr
+ *  -# Round to nearest division by a power-of-two using result_shift
+ *  -# Add offset to each result
+ *  -# Clamp the value between the specified min and max bounds
+ *  -# Clamp the resulting int32 values to the [-128..127] range and cast to QASYMM8_SIGNED.
+ *
+ */
+class CpuGemmLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel : public ICpuKernel
+{
+public:
+    /** Default constructor */
+    CpuGemmLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel() = default;
+    ARM_COMPUTE_DISALLOW_COPY_ALLOW_MOVE(CpuGemmLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel);
+    /** Initialise the kernel's input and output.
+     *
+     * @param[in]  src                          Input tensor info. Data type supported: S32
+     * @param[in]  bias                         Biases tensor info. Only shared biases supported and it can be a nullptr if the biases addition is not required.
+     *                                          Biases are 1D tensor with dimensions [OFM]. Data type supported: Same as @p input.
+     * @param[out] dst                          Output tensor info. Data type supported: Data type supported: QASYMM8_SIGNED
+     * @param[in]  result_fixedpoint_multiplier Fixed point value to be multiplied to each element of the input matrix when once the result_offset has been add
+     * @param[in]  result_shift                 Integer value used to round to nearest division by a power-of-two the result after the fixed point multiplication
+     * @param[in]  result_offset_after_shift    Offset to be applied to result before converting it back to QASYMM8_SIGNED
+     * @param[in]  min                          (Optional) Min value used to saturate down the output result before converting back to QASYMM8_SIGNED
+     * @param[in]  max                          (Optional) Max value used to saturate up the output result before converting back to QASYMM8_SIGNED,
+     *                                          Along with @p min, this value can be used to implement "rectified linear unit" activation functions
+     */
+    void configure(ITensorInfo *src, ITensorInfo *bias, ITensorInfo *dst, int result_fixedpoint_multiplier, int result_shift, int result_offset_after_shift, int min = 0, int max = 0);
+    /** Static function to check if given info will lead to a valid configuration
+     *
+     * Similar to CpuGemmLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel::configure()
+     *
+     * @return a status
+     */
+    static Status validate(const ITensorInfo *src, const ITensorInfo *bias, const ITensorInfo *dst, int min = 0, int max = 0);
+
+    // Inherited methods overridden:
+    void run_op(ITensorPack &tensors, const Window &window, const ThreadInfo &info) override;
+    const char *name() const override;
+
+private:
+    /** Template function to run the CpuGemmLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel
+     *
+     * @param[in]  src    Input tensor info
+     * @param[in]  bias   Bias tensor info
+     * @param[out] dst    Output tensor info
+     * @param[in]  window Region on which to execute the kernel. (Must be a valid region of the window returned by window()).
+     */
+    template <bool is_bounded_relu>
+    void run_internal(const ITensor *src, const ITensor *bias, ITensor *dst, const Window &window);
+
+    /** Common signature for all the specialised CpuGemmLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel functions
+     *
+     * @param[in]  src    Input tensor info
+     * @param[in]  bias   Bias tensor info
+     * @param[out] dst    Output tensor info
+     * @param[in]  window Region on which to execute the kernel.
+     */
+    using QuantizeDownFunctionPtr = void (CpuGemmLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel::*)(
+                                        const ITensor *src, const ITensor *bias, ITensor *dst, const Window &window);
+
+    QuantizeDownFunctionPtr _func{ nullptr };
+    int                     _result_fixedpoint_multiplier{ 0 };
+    int                     _result_shift{ 0 };
+    int                     _result_offset_after_shift{ 0 };
+    int                     _min{ 0 };
+    int                     _max{ 0 };
+};
+} // namespace kernels
+} // namespace cpu
+} // namespace arm_compute
+#endif /* ARM_COMPUTE_CPU_GEMMLOWP_QUANTIZEDOWNINT32TOINT8_SCALEBYFIXEDPOINT_KERNEL_H */
diff --git a/src/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel.cpp b/src/core/cpu/kernels/CpuGemmLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel.cpp
index 83ca6f944d..6631a4fc67 100644
--- a/src/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel.cpp
+++ b/src/core/cpu/kernels/CpuGemmLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel.cpp
@@ -21,7 +21,7 @@
  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  * SOFTWARE.
  */
-#include "src/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel.h"
+#include "src/core/cpu/kernels/CpuGemmLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel.h"
 
 #include "arm_compute/core/Error.h"
 #include "arm_compute/core/Helpers.h"
@@ -37,56 +37,41 @@
 #include "src/core/helpers/WindowHelpers.h"
 
 #include <arm_neon.h>
-#include <cstddef>
-#include <cstdint>
 
 namespace arm_compute
 {
+namespace cpu
+{
+namespace kernels
+{
 namespace
 {
-Status validate_arguments(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, int min, int max)
+Status validate_arguments(const ITensorInfo *src, const ITensorInfo *bias, const ITensorInfo *dst, int min, int max)
 {
-    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::S32);
+    ARM_COMPUTE_ERROR_ON_NULLPTR(src, dst);
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src, 1, DataType::S32);
     ARM_COMPUTE_RETURN_ERROR_ON(min > max);
 
     // Check biases if exist
     if(bias != nullptr)
     {
-        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, bias);
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src, bias);
         ARM_COMPUTE_RETURN_ERROR_ON(bias->num_dimensions() > 1);
-        ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(0) != bias->dimension(0));
+        ARM_COMPUTE_RETURN_ERROR_ON(src->dimension(0) != bias->dimension(0));
     }
 
-    if(output->total_size() != 0)
+    if(dst->total_size() != 0)
     {
-        ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::QASYMM8);
-        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(output, input);
+        ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(dst, 1, DataType::QASYMM8);
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(dst, src);
     }
 
     return Status{};
 }
-
-std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *output)
-{
-    // Output auto inizialitation if not yet initialized
-    auto_init_if_empty(*output, input->clone()->set_data_type(DataType::QASYMM8));
-
-    // Configure kernel window
-    Window win = calculate_max_window(*input, Steps());
-
-    // NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel doesn't need padding so update_window_and_padding() can be skipped
-
-    return std::make_pair(Status{}, win);
-}
 } // namespace
 
-namespace arm_compute
-{
-class Coordinates;
-} // namespace arm_compute
-
 template <bool is_bounded_relu>
-void NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel::run(const Window &window)
+void CpuGemmLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel::run_internal(const ITensor *src, const ITensor *bias, ITensor *dst, const Window &window)
 {
     const int32x4_t  result_offset_after_shift_s32 = vdupq_n_s32(_result_offset_after_shift);
     const uint8x16_t min_u8                        = vdupq_n_u8(static_cast<uint8_t>(_min));
@@ -102,15 +87,15 @@ void NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel::run(const Window
     Window win_collapsed = window.collapse_if_possible(window, Window::DimZ);
     win_collapsed.set(Window::DimX, Window::Dimension(0, 1, 1));
 
-    Iterator in(_input, win_collapsed);
-    Iterator out(_output, win_collapsed);
-    if(_bias != nullptr)
+    Iterator in(src, win_collapsed);
+    Iterator out(dst, win_collapsed);
+    if(bias != nullptr)
     {
         Window win_biases;
         win_biases.set(Window::DimX, Window::Dimension(0, 1, 1));
         win_biases.set(Window::DimY, Window::Dimension(0, 1, 1));
 
-        Iterator bias(_bias, win_biases);
+        Iterator bias_i(bias, win_biases);
         execute_window_loop(win_collapsed, [&](const Coordinates &)
         {
             // Compute 16 elements per iteration
@@ -130,10 +115,10 @@ void NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel::run(const Window
                 const int32x4x4_t bias_s32 =
                 {
                     {
-                        vld1q_s32(reinterpret_cast<const int32_t *>(bias.ptr()) + x + 0),
-                        vld1q_s32(reinterpret_cast<const int32_t *>(bias.ptr()) + x + 4),
-                        vld1q_s32(reinterpret_cast<const int32_t *>(bias.ptr()) + x + 8),
-                        vld1q_s32(reinterpret_cast<const int32_t *>(bias.ptr()) + x + 12)
+                        vld1q_s32(reinterpret_cast<const int32_t *>(bias_i.ptr()) + x + 0),
+                        vld1q_s32(reinterpret_cast<const int32_t *>(bias_i.ptr()) + x + 4),
+                        vld1q_s32(reinterpret_cast<const int32_t *>(bias_i.ptr()) + x + 8),
+                        vld1q_s32(reinterpret_cast<const int32_t *>(bias_i.ptr()) + x + 12)
                     }
                 };
 
@@ -149,7 +134,7 @@ void NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel::run(const Window
             // Compute left-over elements
             for(; x < window_end_x; ++x)
             {
-                const int32_t bias_value = *(reinterpret_cast<const int32_t *>(bias.ptr()) + x);
+                const int32_t bias_value = *(reinterpret_cast<const int32_t *>(bias_i.ptr()) + x);
                 int32_t       in_value   = *(reinterpret_cast<const int32_t *>(in.ptr()) + x);
 
                 // Add bias
@@ -158,7 +143,7 @@ void NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel::run(const Window
                 *(out.ptr() + x) = finalize_quantization(in_value, _result_fixedpoint_multiplier, _result_shift, _result_offset_after_shift, static_cast<uint8_t>(_min), static_cast<uint8_t>(_max), is_bounded_relu);
             }
         },
-        in, out, bias);
+        in, out, bias_i);
     }
     else
     {
@@ -194,52 +179,58 @@ void NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel::run(const Window
     }
 }
 
-NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel::NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel()
-    : _func(nullptr), _input(nullptr), _bias(nullptr), _output(nullptr), _result_fixedpoint_multiplier(0), _result_shift(0), _result_offset_after_shift(0), _min(0), _max(0)
-{
-}
-
-void NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel::configure(const ITensor *input, const ITensor *bias, ITensor *output, int result_fixedpoint_multiplier, int result_shift,
-                                                                          int result_offset_after_shift, int min, int max)
+void CpuGemmLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel::configure(ITensorInfo *src, ITensorInfo *bias, ITensorInfo *dst, int result_fixedpoint_multiplier, int result_shift,
+                                                                           int result_offset_after_shift, int min, int max)
 {
+    ARM_COMPUTE_UNUSED(bias);
     // Perform validate step
-    ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
-    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), (bias != nullptr) ? bias->info() : nullptr, output->info(), min, max));
+    ARM_COMPUTE_ERROR_ON_NULLPTR(src, dst);
+    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(src, bias, dst, min, max));
 
-    _input                        = input;
-    _bias                         = bias;
-    _output                       = output;
     _result_fixedpoint_multiplier = result_fixedpoint_multiplier;
     _result_shift                 = result_shift;
     _result_offset_after_shift    = result_offset_after_shift;
     _min                          = min;
     _max                          = max;
 
+    // Output auto inizialitation if not yet initialized
+    auto_init_if_empty(*dst, src->clone()->set_data_type(DataType::QASYMM8));
+
     // Configure kernel window
-    auto win_config = validate_and_configure_window(input->info(), output->info());
-    ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
-    INEKernel::configure(win_config.second);
+    auto win_config = calculate_max_window(*src, Steps());
+    ICpuKernel::configure(win_config);
 
     // Check if we need to clamp the result using min and max
     const bool is_bounded_relu = !(min <= 0 && max >= 255);
-    _func                      = is_bounded_relu ? &NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel::run<true> : &NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel::run<false>;
+    _func                      = is_bounded_relu ? &CpuGemmLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel::run_internal<true> :
+                                 &CpuGemmLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel::run_internal<false>;
 }
 
-Status NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel::validate(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, int min, int max)
+Status CpuGemmLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel::validate(const ITensorInfo *src, const ITensorInfo *bias, const ITensorInfo *dst, int min, int max)
 {
-    ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
-    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, bias, output, min, max));
-    ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), output->clone().get()).first);
-
+    ARM_COMPUTE_ERROR_ON_NULLPTR(src, dst);
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(src, bias, dst, min, max));
     return Status{};
 }
 
-void NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel::run(const Window &window, const ThreadInfo &info)
+void CpuGemmLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel::run_op(ITensorPack &tensors, const Window &window, const ThreadInfo &info)
 {
     ARM_COMPUTE_UNUSED(info);
     ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
-    ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window);
+    ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICpuKernel::window(), window);
+    ARM_COMPUTE_ERROR_ON_MSG(tensors.empty(), "No inputs provided");
+
+    auto src  = tensors.get_const_tensor(TensorType::ACL_SRC);
+    auto bias = tensors.get_const_tensor(TensorType::ACL_BIAS);
+    auto dst  = tensors.get_tensor(TensorType::ACL_DST);
 
-    (this->*_func)(window);
+    (this->*_func)(src, bias, dst, window);
+}
+
+const char *CpuGemmLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel::name() const
+{
+    return "CpuGemmLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel";
 }
-} // namespace arm_compute
+} // namespace kernels
+} // namespace cpu
+} // namespace arm_compute
\ No newline at end of file
diff --git a/src/core/cpu/kernels/CpuGemmLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel.h b/src/core/cpu/kernels/CpuGemmLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel.h
new file mode 100644
index 0000000000..9b4c056419
--- /dev/null
+++ b/src/core/cpu/kernels/CpuGemmLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel.h
@@ -0,0 +1,108 @@
+/*
+ * Copyright (c) 2017-2021 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_GEMMLOWP_QUANTIZEDOWNINT32TOUINT8_SCALEBYFIXEDPOINT_KERNEL_H
+#define ARM_COMPUTE_CPU_GEMMLOWP_QUANTIZEDOWNINT32TOUINT8_SCALEBYFIXEDPOINT_KERNEL_H
+
+#include "arm_compute/core/KernelDescriptors.h"
+#include "src/core/common/Macros.h"
+#include "src/core/cpu/ICpuKernel.h"
+
+namespace arm_compute
+{
+class ITensor;
+namespace cpu
+{
+namespace kernels
+{
+/** Kernel used to quantize down the int32 accumulator values of GEMMLowp to QASYMM8
+ *
+ * This kernel takes a final int32 accumulator value (the output of @ref NEGEMMLowpMatrixMultiplyKernel), and processes it to obtain the final QASYMM8 value.
+ * The following computations will be performed by the kernel:
+ *
+ *  -# Compute fixed point multiplication between each entry of input by result_fixedpoint_multiplier
+ *  -# Add bias to final result if bias tensor is not a nullptr
+ *  -# Round to nearest division by a power-of-two using result_shift
+ *  -# Add offset to each result
+ *  -# Clamp the value between the specified min and max bounds
+ *  -# Clamp the resulting int32 values to the [0..255] range and cast to QASYMM8.
+ *
+ */
+class CpuGemmLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel : public ICpuKernel
+{
+public:
+    /** Default constructor */
+    CpuGemmLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel() = default;
+    ARM_COMPUTE_DISALLOW_COPY_ALLOW_MOVE(CpuGemmLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel);
+    /** Initialise the kernel's input and output.
+     *
+     * @param[in]  src                          Input tensor info. Data type supported: S32
+     * @param[in]  bias                         Biases tensor info. Only shared biases supported and it can be a nullptr if the biases addition is not required.
+     *                                          Biases are 1D tensor with dimensions [OFM]. Data type supported: Same as @p input.
+     * @param[out] dst                          Output tensor info. Data type supported: Data type supported: QASYMM8
+     * @param[in]  result_fixedpoint_multiplier Fixed point value to be multiplied to each element of the input matrix when once the result_offset has been add
+     * @param[in]  result_shift                 Integer value used to round to nearest division by a power-of-two the result after the fixed point multiplication
+     * @param[in]  result_offset_after_shift    Offset to be applied to result before converting it back to QASYMM8
+     * @param[in]  min                          (Optional) Min value used to saturate down the output result before converting back to QASYMM8
+     * @param[in]  max                          (Optional) Max value used to saturate up the output result before converting back to QASYMM8,
+     *                                          Along with @p min, this value can be used to implement "rectified linear unit" activation functions
+     */
+    void configure(ITensorInfo *src, ITensorInfo *bias, ITensorInfo *dst, int result_fixedpoint_multiplier, int result_shift, int result_offset_after_shift, int min = 0, int max = 0);
+    /** Static function to check if given info will lead to a valid configuration
+     *
+     * Similar to CpuGemmLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel::configure()
+     *
+     * @return a status
+     */
+    static Status validate(const ITensorInfo *src, const ITensorInfo *bias, const ITensorInfo *dst, int min = 0, int max = 0);
+
+    // Inherited methods overridden:
+    void run_op(ITensorPack &tensors, const Window &window, const ThreadInfo &info) override;
+    const char *name() const override;
+
+private:
+    /** Template function to run the CpuGemmLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel
+     *
+     * @param[in] window Region on which to execute the kernel. (Must be a valid region of the window returned by window()).
+     */
+    template <bool is_bounded_relu>
+    void run_internal(const ITensor *src, const ITensor *bias, ITensor *dst, const Window &window);
+
+    /** Common signature for all the specialised CpuGemmLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel functions
+     *
+     * @param[in] window Region on which to execute the kernel.
+     */
+    using QuantizeDownFunctionPtr = void (CpuGemmLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel::*)(
+                                        const ITensor *src, const ITensor *bias, ITensor *dst, const Window &window);
+
+    QuantizeDownFunctionPtr _func{ nullptr };
+    int                     _result_fixedpoint_multiplier{ 0 };
+    int                     _result_shift{ 0 };
+    int                     _result_offset_after_shift{ 0 };
+    int                     _min{ 0 };
+    int                     _max{ 0 };
+};
+} // namespace kernels
+} // namespace cpu
+} // namespace arm_compute
+#endif /* ARM_COMPUTE_CPU_GEMMLOWP_QUANTIZEDOWNINT32TOUINT8_SCALEBYFIXEDPOINT_KERNEL_H */