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>

27 files changed, 1010 insertions, 1527 deletions

diff --git a/Android.bp b/Android.bp
index 19645c0c26..7dc764de93 100644
--- a/Android.bp
+++ b/Android.bp
@@ -169,10 +169,6 @@ cc_library_static {
         "src/core/NEON/kernels/NEGEMMLowpMatrixMultiplyKernel.cpp",
         "src/core/NEON/kernels/NEGEMMLowpOffsetContributionKernel.cpp",
         "src/core/NEON/kernels/NEGEMMLowpOffsetContributionOutputStageKernel.cpp",
-        "src/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ScaleKernel.cpp",
-        "src/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel.cpp",
-        "src/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel.cpp",
-        "src/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel.cpp",
         "src/core/NEON/kernels/NEGEMMLowpReductionKernel.cpp",
         "src/core/NEON/kernels/NEGEMMMatrixAdditionKernel.cpp",
         "src/core/NEON/kernels/NEGEMMMatrixMultiplyKernel.cpp",
@@ -286,6 +282,10 @@ cc_library_static {
         "src/core/cpu/kernels/CpuElementwiseUnaryKernel.cpp",
         "src/core/cpu/kernels/CpuFillKernel.cpp",
         "src/core/cpu/kernels/CpuFloorKernel.cpp",
+        "src/core/cpu/kernels/CpuGemmLowpQuantizeDownInt32ScaleKernel.cpp",
+        "src/core/cpu/kernels/CpuGemmLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel.cpp",
+        "src/core/cpu/kernels/CpuGemmLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel.cpp",
+        "src/core/cpu/kernels/CpuGemmLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel.cpp",
         "src/core/cpu/kernels/CpuMulKernel.cpp",
         "src/core/cpu/kernels/CpuPermuteKernel.cpp",
         "src/core/cpu/kernels/CpuPool2dKernel.cpp",
@@ -645,6 +645,7 @@ cc_library_static {
         "src/runtime/cpu/operators/CpuFlatten.cpp",
         "src/runtime/cpu/operators/CpuFloor.cpp",
         "src/runtime/cpu/operators/CpuGemmDirectConv2d.cpp",
+        "src/runtime/cpu/operators/CpuGemmLowpOutputStage.cpp",
         "src/runtime/cpu/operators/CpuMul.cpp",
         "src/runtime/cpu/operators/CpuPermute.cpp",
         "src/runtime/cpu/operators/CpuPool2d.cpp",
diff --git a/arm_compute/runtime/NEON/functions/NEFullyConnectedLayer.h b/arm_compute/runtime/NEON/functions/NEFullyConnectedLayer.h
index 9727e108a5..d2cd60e576 100644
--- a/arm_compute/runtime/NEON/functions/NEFullyConnectedLayer.h
+++ b/arm_compute/runtime/NEON/functions/NEFullyConnectedLayer.h
@@ -80,7 +80,7 @@ private:
  *  -# @ref NEIm2ColKernel (called when the input comes from a convolutional layer)
  *  -# @ref NETranspose (if @p are_weights_reshaped is set to false and transpose_weights is set to true ) (called once)
  *  -# @ref NEGEMMMatrixMultiplyKernel or @ref NEGEMMLowpMatrixMultiplyCore (if quantized asymmetric)
- *  -# @ref NEGEMMMatrixAdditionKernel or @ref NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPoint (if quantized asymmetric) (if @p biases is not equal to nullptr)
+ *  -# @ref NEGEMMMatrixAdditionKernel or @ref NEGEMMLowpOutputStage (if quantized asymmetric) (if @p biases is not equal to nullptr)
  *
  * @note  The fully connected layer accepts "weights" tensors only with 2 dimensions.
  */
diff --git a/arm_compute/runtime/NEON/functions/NEGEMMConvolutionLayer.h b/arm_compute/runtime/NEON/functions/NEGEMMConvolutionLayer.h
index e89eae1d31..edb58e956a 100644
--- a/arm_compute/runtime/NEON/functions/NEGEMMConvolutionLayer.h
+++ b/arm_compute/runtime/NEON/functions/NEGEMMConvolutionLayer.h
@@ -155,7 +155,7 @@ private:
  * -# @ref NEIm2ColKernel
  * -# @ref NEGEMM (if the data type is BFLOAT16/FP16/FP32)
  * -# @ref NEGEMMLowpMatrixMultiplyCore (if the data type is QASYMM8/QASYMM8_SIGNED)
- * -# @ref NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPoint (if the data type is QASYMM8/QASYMM8_SIGNED)
+ * -# @ref NEGEMMLowpOutputStage (if the data type is QASYMM8/QASYMM8_SIGNED)
  * -# @ref NEArithmeticAddition (if biases != nullptr and we have a 1x1 convolution with the NHWC data layout)
  * -# @ref NECol2ImKernel (if NCHW data layout)
  *
diff --git a/arm_compute/runtime/NEON/functions/NEGEMMLowpOutputStage.h b/arm_compute/runtime/NEON/functions/NEGEMMLowpOutputStage.h
index fa5f5e3826..232344e5c2 100644
--- a/arm_compute/runtime/NEON/functions/NEGEMMLowpOutputStage.h
+++ b/arm_compute/runtime/NEON/functions/NEGEMMLowpOutputStage.h
@@ -25,7 +25,7 @@
 #define ARM_COMPUTE_NEGEMMLOWPOUTPUTSTAGE_H
 
 #include "arm_compute/core/Types.h"
-#include "arm_compute/runtime/NEON/INESimpleFunctionNoBorder.h"
+#include "arm_compute/runtime/IFunction.h"
 
 /** This file contains all available output stages for GEMMLowp.
  *
@@ -39,237 +39,17 @@ namespace arm_compute
 {
 class ITensor;
 class ITensorInfo;
-
-/** Basic function to execute NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPoint.
- *
- *  NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPoint depends on 3 parameters:
- *
- *  result_fixedpoint_multiplier, result_shift, result_offset_after_shift
- *
- * The final result is:
- *
- * (FixedPointMul(input[i][k], result_fixedpoint_multiplier) >> result_shift) + result_offset_after_shift
- *
- * where FixedPointMul(x, y) is the nearest integer to the following
- * mathematical expression, evaluated without overflow or intermediate rounding:
- *
- * (x * y) / 2^31
- *
- * For more information: https://github.com/google/gemmlowp/blob/master/public/output_stages.h#L68
- *
- * In case the bias tensor is provided, the final result is:
- *
- * ((FixedPointMul(input[i][k] + bias[k], result_fixedpoint_multiplier)) >> result_shift) + result_offset_after_shift
- *
- *  This function calls the following kernels:
- *
- * -# @ref NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel
- *
- * @note The function accepts also 2 optional input arguments (min and max) which can be used to implement "rectified linear unit" activation functions
- *       after the result is shifted right by result_shift
-*/
-class NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPoint : public INESimpleFunctionNoBorder
-{
-public:
-    /** Constructor */
-    NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPoint() = default;
-    /** Prevent instances of this class from being copied (As this class contains pointers) */
-    NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPoint(const NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPoint &) = delete;
-    /** Prevent instances of this class from being copied (As this class contains pointers) */
-    NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPoint &operator=(const NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPoint &) = delete;
-    /** Prevent instances of this class from being moved (As this class contains non movable objects) */
-    NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPoint(NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPoint &&) = delete;
-    /** Prevent instances of this class from being moved (As this class contains non movable objects) */
-    NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPoint &operator=(NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPoint &&) = delete;
-    /** Default destructor */
-    ~NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPoint();
-    /** Initialise the kernel's inputs, 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                 Number of bits to shift right 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. Defaults to the minimum possible 32-bit signed integer.
-     * @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. Defaults to the maximum possible 32-bit signed integer.
-     */
-    void configure(const ITensor *input, const ITensor *bias, ITensor *output, int result_fixedpoint_multiplier, int result_shift, int result_offset_after_shift,
-                   int min = std::numeric_limits<int32_t>::lowest(), int max = std::numeric_limits<int32_t>::max());
-    /** Static function to check if given info will lead to a valid configuration of @ref NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPoint
-     *
-     * @param[in] input  Input tensor. It is the output of @ref NEGEMMLowpMatrixMultiplyCore function. Data type supported: S32
-     * @param[in] bias   Biases tensor. Only shared biases supported and it can be a nullptr if the addition of biases 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. Defaults to the minimum possible 32-bit signed integer.
-     * @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. Defaults to the maximum possible 32-bit signed integer.
-     *
-     * @return a status
-     */
-    static Status validate(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, int min = std::numeric_limits<int32_t>::lowest(), int max = std::numeric_limits<int32_t>::max());
-};
-/** Basic function to execute NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPoint.
- *
- *  NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPoint depends on 3 parameters:
- *
- *  result_fixedpoint_multiplier, result_shift, result_offset_after_shift
- *
- * The final result is:
- *
- * (FixedPointMul(input[i][k], result_fixedpoint_multiplier) >> result_shift) + result_offset_after_shift
- *
- * where FixedPointMul(x, y) is the nearest integer to the following
- * mathematical expression, evaluated without overflow or intermediate rounding:
- *
- * (x * y) / 2^31
- *
- * For more information: https://github.com/google/gemmlowp/blob/master/public/output_stages.h#L68
- *
- * In case the bias tensor is provided, the final result is:
- *
- * ((FixedPointMul(input[i][k] + bias[k], result_fixedpoint_multiplier)) >> result_shift) + result_offset_after_shift
- *
- *  This function calls the following kernels:
- *
- * -# @ref NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel
- *
- * @note The function accepts also 2 optional input arguments (min and max) which can be used to implement "rectified linear unit" activation functions
- *       after the result is shifted right by result_shift
-*/
-class NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPoint : public INESimpleFunctionNoBorder
-{
-public:
-    /** Constructor */
-    NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPoint() = default;
-    /** Prevent instances of this class from being copied (As this class contains pointers) */
-    NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPoint(const NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPoint &) = delete;
-    /** Prevent instances of this class from being copied (As this class contains pointers) */
-    NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPoint &operator=(const NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPoint &) = delete;
-    /** Prevent instances of this class from being moved (As this class contains non movable objects) */
-    NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPoint(NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPoint &&) = delete;
-    /** Prevent instances of this class from being moved (As this class contains non movable objects) */
-    NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPoint &operator=(NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPoint &&) = delete;
-    /** Default destructor */
-    ~NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPoint();
-    /** Initialise the kernel's inputs, 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                 Number of bits to shift right 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. Defaults to the minimum possible 32-bit signed integer.
-     * @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. Defaults to the maximum possible 32-bit signed integer.
-     */
-    void configure(const ITensor *input, const ITensor *bias, ITensor *output, int result_fixedpoint_multiplier, int result_shift, int result_offset_after_shift,
-                   int min = std::numeric_limits<int32_t>::lowest(), int max = std::numeric_limits<int32_t>::max());
-    /** Static function to check if given info will lead to a valid configuration of @ref NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPoint
-     *
-     * @param[in] input  Input tensor. It is the output of @ref NEGEMMLowpMatrixMultiplyCore function. Data type supported: S32
-     * @param[in] bias   Biases tensor. Only shared biases supported and it can be a nullptr if the addition of biases 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. Defaults to the minimum possible 32-bit signed integer.
-     * @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. Defaults to the maximum possible 32-bit signed integer.
-     *
-     * @return a status
-     */
-    static Status validate(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, int min = std::numeric_limits<int32_t>::lowest(), int max = std::numeric_limits<int32_t>::max());
-};
-/** Basic function to execute NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPoint.
- *
- *  NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPoint depends on 2 parameters:
- *
- *  result_fixedpoint_multiplier, result_shift
- *
- * The final result is:
- *
- * (FixedPointMul(input[i][k], result_fixedpoint_multiplier) >> result_shift)
- *
- * where FixedPointMul(x, y) is the nearest integer to the following
- * mathematical expression, evaluated without overflow or intermediate rounding:
- *
- * (x * y) / 2^31
- *
- * For more information: https://github.com/google/gemmlowp/blob/master/public/output_stages.h#L68
- *
- * In case the bias tensor is provided, the final result is:
- *
- * ((FixedPointMul(input[i][k] + bias[k], result_fixedpoint_multiplier)) >> result_shift) + result_offset_after_shift
- *
- *  This function calls the following kernels:
- *
- * -# @ref NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel
- *
- * @note The function accepts also 2 optional input arguments (min and max) which can be used to implement "rectified linear unit" activation functions
- *       after the result is shifted right by result_shift
-*/
-class NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPoint : public INESimpleFunctionNoBorder
-{
-public:
-    /** Constructor */
-    NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPoint() = default;
-    /** Prevent instances of this class from being copied (As this class contains pointers) */
-    NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPoint(const NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPoint &) = delete;
-    /** Prevent instances of this class from being copied (As this class contains pointers) */
-    NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPoint &operator=(const NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPoint &) = delete;
-    /** Prevent instances of this class from being moved (As this class contains non movable objects) */
-    NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPoint(NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPoint &&) = delete;
-    /** Prevent instances of this class from being moved (As this class contains non movable objects) */
-    NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPoint &operator=(NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPoint &&) = delete;
-    /** Default destructor */
-    ~NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPoint();
-    /** Initialise the kernel's inputs, 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                 Number of bits to shift right 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 the minimum possible 32-bit signed integer.
-     * @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 the maximum possible 32-bit signed integer.
-     */
-    void configure(const ITensor *input, const ITensor *bias, ITensor *output, int result_fixedpoint_multiplier, int result_shift, int min = std::numeric_limits<int32_t>::lowest(),
-                   int max = std::numeric_limits<int32_t>::max());
-    /** Static function to check if given info will lead to a valid configuration of @ref NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPoint
-     *
-     * @param[in] input  Input tensor info. It is the output of @ref NEGEMMLowpMatrixMultiplyCore function. Data type supported: S32
-     * @param[in] bias   Biases tensor info. Only shared biases supported and it can be a nullptr if the addition of biases is not required.
-     *                   Biases are 1D tensor 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 the minimum possible 32-bit signed integer.
-     * @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 the maximum possible 32-bit signed integer.
-     *
-     * @return a status
-     */
-    static Status validate(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, int min = std::numeric_limits<int32_t>::lowest(), int max = std::numeric_limits<int32_t>::max());
-};
-
 /** Basic function to execute GEMMLowpQuantizeDown kernels.
  *
- *  This function calls the following kernels:
+ *  This function calls the following operators:
  *
- * -# @ref NEGEMMLowpQuantizeDownInt32ScaleKernel
- * -# @ref NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel
- * -# @ref NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel
- * -# @ref NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel
+ * -# @ref cpu::CpuGemmLowpOutputStage
 */
-class NEGEMMLowpOutputStage : public INESimpleFunctionNoBorder
+class NEGEMMLowpOutputStage : public IFunction
 {
 public:
     /** Constructor */
-    NEGEMMLowpOutputStage() = default;
+    NEGEMMLowpOutputStage();
     /** Prevent instances of this class from being copied (As this class contains pointers) */
     NEGEMMLowpOutputStage(const NEGEMMLowpOutputStage &) = delete;
     /** Prevent instances of this class from being copied (As this class contains pointers) */
@@ -310,6 +90,13 @@ public:
      * @return a status
      */
     static Status validate(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, const GEMMLowpOutputStageInfo &info);
+
+    // Inherited methods overridden:
+    void run() override;
+
+private:
+    struct Impl;
+    std::unique_ptr<Impl> _impl;
 };
 } // namespace arm_compute
 #endif /*ARM_COMPUTE_NEGEMMLOWPOUTPUTSTAGE_H */
diff --git a/arm_compute/runtime/NEON/functions/NELSTMLayerQuantized.h b/arm_compute/runtime/NEON/functions/NELSTMLayerQuantized.h
index 2f0c753691..bcb89d997d 100644
--- a/arm_compute/runtime/NEON/functions/NELSTMLayerQuantized.h
+++ b/arm_compute/runtime/NEON/functions/NELSTMLayerQuantized.h
@@ -50,7 +50,7 @@ class ITensor;
  * This function calls the following functions/kernels:
  *
  * -# @ref NEGEMMLowpMatrixMultiplyCore                          Quantized matrix multiplication core. Accumulators are 32-bit integers
- * -# @ref NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPoint   Convert 32-bit integers into QSYMM16
+ * -# @ref NEGEMMLowpOutputStage                                 Convert 32-bit integers into QSYMM16
  * -# @ref NETranspose                                           Matrix transpose
  * -# @ref NEConcatenateLayer                                    Tensor concatenation
  * -# @ref NEActivationLayer                                     Activation functions (tanh and logistic)
@@ -147,30 +147,30 @@ private:
     MemoryGroup _memory_group;
 
     // Functions used
-    NEGEMMLowpMatrixMultiplyCore                        _gemmlowp;
-    NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPoint _output_stage;
-    NETranspose                                         _transpose_weights;
-    NEConcatenateLayer                                  _concat_input_weights;
-    NEConcatenateLayer                                  _concat_recurrent_weights;
-    NEConcatenateLayer                                  _concat_weights;
-    NEConcatenateLayer                                  _concat_inputs;
-    NEConcatenateLayer                                  _concat_bias;
-    NEActivationLayer                                   _sigmoid_forget_gate;
-    NEActivationLayer                                   _sigmoid_input_gate;
-    NEActivationLayer                                   _sigmoid_output_gate;
-    NEActivationLayer                                   _tanh_modulation_gate;
-    NEActivationLayer                                   _tanh_output_state;
-    NEArithmeticAddition                                _add1;
-    NEArithmeticAddition                                _add2;
-    NEPixelWiseMultiplication                           _mul1;
-    NEPixelWiseMultiplication                           _mul2;
-    NEPixelWiseMultiplication                           _mul3;
-    NESlice                                             _slice_input_tensor;
-    NESlice                                             _slice_forget_tensor;
-    NESlice                                             _slice_cell_tensor;
-    NESlice                                             _slice_output_tensor;
-    NEDequantizationLayer                               _dequantize;
-    NEQuantizationLayer                                 _quantize;
+    NEGEMMLowpMatrixMultiplyCore _gemmlowp;
+    NEGEMMLowpOutputStage        _output_stage;
+    NETranspose                  _transpose_weights;
+    NEConcatenateLayer           _concat_input_weights;
+    NEConcatenateLayer           _concat_recurrent_weights;
+    NEConcatenateLayer           _concat_weights;
+    NEConcatenateLayer           _concat_inputs;
+    NEConcatenateLayer           _concat_bias;
+    NEActivationLayer            _sigmoid_forget_gate;
+    NEActivationLayer            _sigmoid_input_gate;
+    NEActivationLayer            _sigmoid_output_gate;
+    NEActivationLayer            _tanh_modulation_gate;
+    NEActivationLayer            _tanh_output_state;
+    NEArithmeticAddition         _add1;
+    NEArithmeticAddition         _add2;
+    NEPixelWiseMultiplication    _mul1;
+    NEPixelWiseMultiplication    _mul2;
+    NEPixelWiseMultiplication    _mul3;
+    NESlice                      _slice_input_tensor;
+    NESlice                      _slice_forget_tensor;
+    NESlice                      _slice_cell_tensor;
+    NESlice                      _slice_output_tensor;
+    NEDequantizationLayer        _dequantize;
+    NEQuantizationLayer          _quantize;
 
     // Tensor pointers
     const ITensor *_input_to_input_weights;
diff --git a/arm_compute/runtime/NEON/functions/NEQLSTMLayer.h b/arm_compute/runtime/NEON/functions/NEQLSTMLayer.h
index 7c2e9bc5a1..77adffd543 100644
--- a/arm_compute/runtime/NEON/functions/NEQLSTMLayer.h
+++ b/arm_compute/runtime/NEON/functions/NEQLSTMLayer.h
@@ -54,7 +54,7 @@ class NEGEMMLowpMatrixAReductionKernel;
  * -# @ref NEArithmeticSubtraction                               Elementwise subtraction
  * -# @ref NECopy                                                Copy kernel for copying output_state_out to output
  * -# @ref NEGEMMLowpMatrixMultiplyCore                          Quantized matrix multiplication core. Accumulators are 32-bit integers
- * -# @ref NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPoint   Convert 32-bit integers into QSYMM16
+ * -# @ref NEGEMMLowpOutputStage                                 Convert 32-bit integers into QSYMM16
  * -# @ref NEGEMMLowpMatrixAReductionKernel                      For precomputing effective biases to use
  * -# @ref NEPixelWiseMultiplication                             Elementwise multiplication
  * -# @ref NETranspose                                           Transpose function for reshaping the weights
diff --git a/docs/user_guide/release_version_and_change_log.dox b/docs/user_guide/release_version_and_change_log.dox
index 3ffa11b045..b847761b8c 100644
--- a/docs/user_guide/release_version_and_change_log.dox
+++ b/docs/user_guide/release_version_and_change_log.dox
@@ -609,7 +609,7 @@ v20.05 Public major release
      - @ref NEReduceMean
      - @ref NEArgMinMaxLayer
      - @ref NEDeconvolutionLayer
-     - @ref NEGEMMLowpQuantizeDownInt32ScaleKernel
+     - NEGEMMLowpQuantizeDownInt32ScaleKernel
      - @ref CPPBoxWithNonMaximaSuppressionLimit
      - @ref CPPDetectionPostProcessLayer
      - @ref CPPPermuteKernel
@@ -683,7 +683,7 @@ v20.02 Public major release
      - CLGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel / @ref CLGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPoint
  - New Arm® Neon™ kernels / functions:
      - @ref NEFill
-     - @ref NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel / @ref NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPoint
+     - NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel / NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPoint
  - Deprecated Arm® Neon™ functions / interfaces:
      - CLDepthwiseConvolutionLayer3x3
      - NEDepthwiseConvolutionLayerOptimized
@@ -800,7 +800,7 @@ v19.08 Public major release
     - NEBatchConcatenateLayerKernel
     - @ref NEDepthToSpaceLayerKernel / @ref NEDepthToSpaceLayer
     - NEDepthwiseConvolutionLayerNativeKernel
-    - @ref NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel
+    - NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel
     - @ref NEMeanStdDevNormalizationKernel / @ref NEMeanStdDevNormalizationLayer
     - @ref NESpaceToDepthLayerKernel / @ref NESpaceToDepthLayer
  - New OpenCL kernels / functions:
@@ -1243,7 +1243,7 @@ v17.12 Public major release
     - arm_compute::NEHGEMMAArch64FP16Kernel
     - NEDepthwiseConvolutionLayer3x3Kernel / NEDepthwiseIm2ColKernel / NEGEMMMatrixVectorMultiplyKernel / NEDepthwiseVectorToTensorKernel / @ref NEDepthwiseConvolutionLayer
     - @ref NEGEMMLowpOffsetContributionKernel / @ref NEGEMMLowpMatrixAReductionKernel / @ref NEGEMMLowpMatrixBReductionKernel / @ref NEGEMMLowpMatrixMultiplyCore
-    - @ref NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel / @ref NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPoint
+    - NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel / NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPoint
     - NEWinogradLayer / NEWinogradLayerKernel
 
  - New OpenCL kernels / functions
diff --git a/examples/neon_gemm_qasymm8.cpp b/examples/neon_gemm_qasymm8.cpp
index efe1655d45..159636db1d 100644
--- a/examples/neon_gemm_qasymm8.cpp
+++ b/examples/neon_gemm_qasymm8.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2020 Arm Limited.
+ * Copyright (c) 2020-2021 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -26,8 +26,8 @@
 #include "arm_compute/core/utils/quantization/AsymmHelpers.h"
 #include "arm_compute/runtime/NEON/NEFunctions.h"
 #include "arm_compute/runtime/NEON/NEScheduler.h"
-#include "utils/Utils.h"
 #include "support/ToolchainSupport.h"
+#include "utils/Utils.h"
 
 #include <cstdlib>
 
@@ -102,10 +102,10 @@ int main(int argc, char **argv)
     Tensor q_dst0;
     Tensor q_res;
     Tensor q_res_output;
-    size_t M = 4;
-    size_t N = 4;
-    size_t K = 4;
-    bool default_input = true;
+    size_t M             = 4;
+    size_t N             = 4;
+    size_t K             = 4;
+    bool   default_input = true;
 
     // Parse args
     if(argc < 3) /* case default matrix sizes */
@@ -144,15 +144,18 @@ int main(int argc, char **argv)
 
     // Fill in: one is the identity matrix, other is sequential values
     // src1: Identity matrix
-    for(size_t i = 0; i < M * K; i++) {
+    for(size_t i = 0; i < M * K; i++)
+    {
         src1_ptr[i] = 0;
     }
-    for(size_t i = 0; i < M; i++) {
+    for(size_t i = 0; i < M; i++)
+    {
         src1_ptr[i * K + i] = 1.0f;
     }
 
     // src2: Sequential values matrix
-    for(size_t i = 0; i < K * N; i++) {
+    for(size_t i = 0; i < K * N; i++)
+    {
         src2_ptr[i] = i * 1.123f;
     }
 
@@ -217,13 +220,22 @@ int main(int argc, char **argv)
     qgemm.configure(&q_src1, &q_src2, nullptr, &q_res);
 
     // Configure output stage after computing shift and multiplier parameters
-    NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPoint gemmlowp_output_stage;
-    int output_multiplier;
-    int output_shift;
-    float multiplier = (src1_qinfo.uniform().scale * src2_qinfo.uniform().scale) / dst0_qinfo.uniform().scale;
+    NEGEMMLowpOutputStage gemmlowp_output_stage;
+    int                   output_multiplier;
+    int                   output_shift;
+    float                 multiplier = (src1_qinfo.uniform().scale * src2_qinfo.uniform().scale) / dst0_qinfo.uniform().scale;
     quantization::calculate_quantized_multiplier_less_than_one(multiplier, &output_multiplier, &output_shift);
     std::cout << "(q_multiplier, q_shift) = (" << output_multiplier << ", " << output_shift << ")\n\n";
-    gemmlowp_output_stage.configure(&q_res, nullptr, &q_res_output, output_multiplier, output_shift, dst0_qinfo.uniform().offset);
+
+    GEMMLowpOutputStageInfo info;
+    info.type                = GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT;
+    info.gemmlowp_multiplier = output_multiplier;
+    info.gemmlowp_shift      = output_shift;
+    info.gemmlowp_offset     = dst0_qinfo.uniform().offset;
+    info.output_data_type    = DataType::QASYMM8;
+    q_res_output.info()->set_data_type(DataType::QASYMM8);
+    q_res_output.info()->set_num_channels(1);
+    gemmlowp_output_stage.configure(&q_res, nullptr, &q_res_output, info);
 
     // Allocate all tensors
     q_src1.allocator()->allocate();
diff --git a/filelist.json b/filelist.json
index c375b1daef..e30426bd19 100644
--- a/filelist.json
+++ b/filelist.json
@@ -160,6 +160,7 @@
                     "src/runtime/cpu/operators/CpuFlatten.cpp",
                     "src/runtime/cpu/operators/CpuFloor.cpp",
                     "src/runtime/cpu/operators/CpuGemmDirectConv2d.cpp",
+                    "src/runtime/cpu/operators/CpuGemmLowpOutputStage.cpp",
                     "src/runtime/cpu/operators/CpuMul.cpp",
                     "src/runtime/cpu/operators/CpuQuantize.cpp",
                     "src/runtime/cpu/operators/CpuReshape.cpp",
@@ -197,6 +198,10 @@
                     "src/core/cpu/kernels/CpuElementwiseUnaryKernel.cpp",
                     "src/core/cpu/kernels/CpuFillKernel.cpp",
                     "src/core/cpu/kernels/CpuFloorKernel.cpp",
+                    "src/core/cpu/kernels/CpuGemmLowpQuantizeDownInt32ScaleKernel.cpp",
+                    "src/core/cpu/kernels/CpuGemmLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel.cpp",
+                    "src/core/cpu/kernels/CpuGemmLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel.cpp",
+                    "src/core/cpu/kernels/CpuGemmLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel.cpp",
                     "src/core/cpu/kernels/CpuMulKernel.cpp",
                     "src/core/cpu/kernels/CpuQuantizeKernel.cpp",
                     "src/core/cpu/kernels/CpuScaleKernel.cpp",
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 */
diff --git a/src/runtime/NEON/functions/NEGEMMLowpOutputStage.cpp b/src/runtime/NEON/functions/NEGEMMLowpOutputStage.cpp
index 807785a534..8351cc66d0 100644
--- a/src/runtime/NEON/functions/NEGEMMLowpOutputStage.cpp
+++ b/src/runtime/NEON/functions/NEGEMMLowpOutputStage.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017-2020 Arm Limited.
+ * Copyright (c) 2017-2021 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -25,57 +25,23 @@
 
 #include "arm_compute/core/ITensor.h"
 #include "arm_compute/core/Validate.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/runtime/cpu/operators/CpuGemmLowpOutputStage.h"
 
 namespace arm_compute
 {
-NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPoint::~NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPoint() = default;
-
-void NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPoint::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)
-{
-    auto k = std::make_unique<NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel>();
-    k->configure(input, bias, output, result_fixedpoint_multiplier, result_shift, result_offset_after_shift, min, max);
-    _kernel = std::move(k);
-}
-
-Status NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPoint::validate(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, int min, int max)
-{
-    return NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel::validate(input, bias, output, min, max);
-}
-
-NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPoint::~NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPoint() = default;
-
-void NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPoint::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)
-{
-    auto k = std::make_unique<NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel>();
-    k->configure(input, bias, output, result_fixedpoint_multiplier, result_shift, result_offset_after_shift, min, max);
-    _kernel = std::move(k);
-}
-
-Status NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPoint::validate(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, int min, int max)
+struct NEGEMMLowpOutputStage::Impl
 {
-    return NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel::validate(input, bias, output, min, max);
-}
-
-NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPoint::~NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPoint() = default;
-
-void NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPoint::configure(const ITensor *input, const ITensor *bias, ITensor *output, int result_fixedpoint_multiplier, int result_shift, int min, int max)
-{
-    auto k = std::make_unique<NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel>();
-    k->configure(input, bias, output, result_fixedpoint_multiplier, result_shift, min, max);
-    _kernel = std::move(k);
-}
-
-Status NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPoint::validate(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, int min, int max)
+    const ITensor                               *src{ nullptr };
+    const ITensor                               *bias{ nullptr };
+    ITensor                                     *dst{ nullptr };
+    ITensorPack                                  run_pack{};
+    std::unique_ptr<cpu::CpuGemmLowpOutputStage> op{ nullptr };
+};
+
+NEGEMMLowpOutputStage::NEGEMMLowpOutputStage()
+    : _impl(std::make_unique<Impl>())
 {
-    return NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel::validate(input, bias, output, min, max);
 }
-
 NEGEMMLowpOutputStage::~NEGEMMLowpOutputStage() = default;
 
 void NEGEMMLowpOutputStage::configure(const ITensor *input, const ITensor *bias, ITensor *output, const GEMMLowpOutputStageInfo &info)
@@ -83,104 +49,27 @@ void NEGEMMLowpOutputStage::configure(const ITensor *input, const ITensor *bias,
     // Perform validate step
     ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
     ARM_COMPUTE_ERROR_THROW_ON(NEGEMMLowpOutputStage::validate(input->info(), bias != nullptr ? bias->info() : nullptr, output->info(), info));
+    _impl->src  = input;
+    _impl->bias = bias;
+    _impl->dst  = output;
+    _impl->op   = std::make_unique<cpu::CpuGemmLowpOutputStage>();
+    _impl->op->configure(input->info(), (bias == nullptr) ? nullptr : bias->info(), output->info(), info);
 
-    switch(info.type)
+    _impl->run_pack =
     {
-        case GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT:
-        {
-            switch(info.output_data_type)
-            {
-                case DataType::QASYMM8:
-                {
-                    auto k = std::make_unique<NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel>();
-                    k->configure(input, bias, output, info.gemmlowp_multiplier, info.gemmlowp_shift, info.gemmlowp_offset, info.gemmlowp_min_bound, info.gemmlowp_max_bound);
-                    _kernel = std::move(k);
-                    break;
-                }
-                case DataType::QASYMM8_SIGNED:
-                {
-                    auto k = std::make_unique<NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel>();
-                    k->configure(input, bias, output, info.gemmlowp_multiplier, info.gemmlowp_shift, info.gemmlowp_offset, info.gemmlowp_min_bound, info.gemmlowp_max_bound);
-                    _kernel = std::move(k);
-                    break;
-                }
-                case DataType::QSYMM16:
-                {
-                    auto k = std::make_unique<NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel>();
-                    k->configure(input, bias, output, info.gemmlowp_multiplier, info.gemmlowp_shift, info.gemmlowp_min_bound, info.gemmlowp_max_bound);
-                    _kernel = std::move(k);
-                    break;
-                }
-                default:
-                {
-                    ARM_COMPUTE_ERROR("Unsupported output data type.");
-                    break;
-                }
-            }
-            break;
-        }
-        case GEMMLowpOutputStageType::QUANTIZE_DOWN:
-        {
-            switch(info.output_data_type)
-            {
-                case DataType::QASYMM8:
-                case DataType::QASYMM8_SIGNED:
-                {
-                    auto k = std::make_unique<NEGEMMLowpQuantizeDownInt32ScaleKernel>();
-                    k->configure(input, bias, output, &info);
-                    _kernel = std::move(k);
-                    break;
-                }
-                default:
-                {
-                    ARM_COMPUTE_ERROR("Unsupported output data type.");
-                    break;
-                }
-            }
-            break;
-        }
-        default:
-            ARM_COMPUTE_ERROR("Unsupported GEMMLowpOutputStage type.");
-    }
+        { TensorType::ACL_SRC, _impl->src },
+        { TensorType::ACL_BIAS, _impl->bias },
+        { TensorType::ACL_DST, _impl->dst }
+    };
 }
 
 Status NEGEMMLowpOutputStage::validate(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, const GEMMLowpOutputStageInfo &info)
 {
-    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(output);
-    ARM_COMPUTE_RETURN_ERROR_ON_MSG(output->data_type() == DataType::UNKNOWN, "NEGEMMLowpQuantizeDownScaleByFixedPoint cannot be used with UNKNOWN output data type.");
-    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED, DataType::QSYMM16);
-
-    ARM_COMPUTE_RETURN_ERROR_ON((info.type != GEMMLowpOutputStageType::QUANTIZE_DOWN) && (info.type != GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT));
+    return cpu::CpuGemmLowpOutputStage::validate(input, bias, output, info);
+}
 
-    switch(info.type)
-    {
-        case GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT:
-        {
-            switch(output->data_type())
-            {
-                case DataType::QASYMM8:
-                    return NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel::validate(input, bias, output, info.gemmlowp_min_bound, info.gemmlowp_max_bound);
-                case DataType::QASYMM8_SIGNED:
-                    return NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel::validate(input, bias, output, info.gemmlowp_min_bound, info.gemmlowp_max_bound);
-                case DataType::QSYMM16:
-                    return NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel::validate(input, bias, output, info.gemmlowp_min_bound, info.gemmlowp_max_bound);
-                default:
-                    return ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Unsupported output data type.");
-            }
-        }
-        case GEMMLowpOutputStageType::QUANTIZE_DOWN:
-        {
-            switch(output->data_type())
-            {
-                case DataType::QASYMM8:
-                case DataType::QASYMM8_SIGNED:
-                    return NEGEMMLowpQuantizeDownInt32ScaleKernel::validate(input, bias, output, &info);
-                default:
-                    return ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Unsupported output data type.");
-            }
-        }
-        default:
-            return ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Unsupported GEMMLowpOutputStage type.");
-    }
+void NEGEMMLowpOutputStage::run()
+{
+    _impl->op->run(_impl->run_pack);
 }
 } // namespace arm_compute
diff --git a/src/runtime/NEON/functions/NELSTMLayerQuantized.cpp b/src/runtime/NEON/functions/NELSTMLayerQuantized.cpp
index 5c0f19a15c..a56e16864a 100644
--- a/src/runtime/NEON/functions/NELSTMLayerQuantized.cpp
+++ b/src/runtime/NEON/functions/NELSTMLayerQuantized.cpp
@@ -152,7 +152,13 @@ void NELSTMLayerQuantized::configure(const ITensor *input,
     quantization::calculate_quantized_multiplier(multiplier, &output_multiplier, &output_shift);
 
     _memory_group.manage(&_output_lowp);
-    _output_stage.configure(&_output_highp, &_bias, &_output_lowp, output_multiplier, output_shift);
+
+    GEMMLowpOutputStageInfo info;
+    info.type                = GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT;
+    info.gemmlowp_multiplier = output_multiplier;
+    info.gemmlowp_shift      = output_shift;
+    info.output_data_type    = DataType::QSYMM16;
+    _output_stage.configure(&_output_highp, &_bias, &_output_lowp, info);
     _output_highp.allocator()->allocate();
     _bias.allocator()->allocate();
 
@@ -358,7 +364,12 @@ Status NELSTMLayerQuantized::validate(const ITensorInfo *input,
     ARM_COMPUTE_RETURN_ON_ERROR(quantization::calculate_quantized_multiplier(multiplier, &output_multiplier, &output_shift));
 
     // _output_stage
-    ARM_COMPUTE_RETURN_ON_ERROR(NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPoint::validate(&output_highp, &bias_concatenated, &output_lowp));
+    GEMMLowpOutputStageInfo info;
+    info.type                = GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT;
+    info.gemmlowp_multiplier = output_multiplier;
+    info.gemmlowp_shift      = output_shift;
+    info.output_data_type    = DataType::QSYMM16;
+    ARM_COMPUTE_RETURN_ON_ERROR(NEGEMMLowpOutputStage::validate(&output_highp, &bias_concatenated, &output_lowp, info));
 
     TensorInfo input_gate_input;
     TensorInfo forget_gate_input;
diff --git a/src/runtime/cpu/operators/CpuGemmLowpOutputStage.cpp b/src/runtime/cpu/operators/CpuGemmLowpOutputStage.cpp
new file mode 100644
index 0000000000..e17f854a21
--- /dev/null
+++ b/src/runtime/cpu/operators/CpuGemmLowpOutputStage.cpp
@@ -0,0 +1,147 @@
+/*
+ * Copyright (c) 2021 Arm Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#include "src/runtime/cpu/operators/CpuGemmLowpOutputStage.h"
+
+#include "arm_compute/core/ITensor.h"
+#include "arm_compute/core/Validate.h"
+#include "arm_compute/runtime/NEON/NEScheduler.h"
+#include "src/core/cpu/kernels/CpuGemmLowpQuantizeDownInt32ScaleKernel.h"
+#include "src/core/cpu/kernels/CpuGemmLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel.h"
+#include "src/core/cpu/kernels/CpuGemmLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel.h"
+#include "src/core/cpu/kernels/CpuGemmLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel.h"
+
+namespace arm_compute
+{
+namespace cpu
+{
+void CpuGemmLowpOutputStage::configure(ITensorInfo *src, ITensorInfo *bias, ITensorInfo *dst, const GEMMLowpOutputStageInfo &info)
+{
+    // Perform validate step
+    ARM_COMPUTE_ERROR_THROW_ON(CpuGemmLowpOutputStage::validate(src, bias, dst, info));
+
+    switch(info.type)
+    {
+        case GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT:
+        {
+            switch(info.output_data_type)
+            {
+                case DataType::QASYMM8:
+                {
+                    auto k = std::make_unique<kernels::CpuGemmLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel>();
+                    k->configure(src, bias, dst, info.gemmlowp_multiplier, info.gemmlowp_shift, info.gemmlowp_offset, info.gemmlowp_min_bound, info.gemmlowp_max_bound);
+                    _kernel = std::move(k);
+                    break;
+                }
+                case DataType::QASYMM8_SIGNED:
+                {
+                    auto k = std::make_unique<kernels::CpuGemmLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel>();
+                    k->configure(src, bias, dst, info.gemmlowp_multiplier, info.gemmlowp_shift, info.gemmlowp_offset, info.gemmlowp_min_bound, info.gemmlowp_max_bound);
+                    _kernel = std::move(k);
+                    break;
+                }
+                case DataType::QSYMM16:
+                {
+                    auto k = std::make_unique<kernels::CpuGemmLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel>();
+                    k->configure(src, bias, dst, info.gemmlowp_multiplier, info.gemmlowp_shift, info.gemmlowp_min_bound, info.gemmlowp_max_bound);
+                    _kernel = std::move(k);
+                    break;
+                }
+                default:
+                {
+                    ARM_COMPUTE_ERROR("Unsupported output data type.");
+                    break;
+                }
+            }
+            break;
+        }
+        case GEMMLowpOutputStageType::QUANTIZE_DOWN:
+        {
+            switch(info.output_data_type)
+            {
+                case DataType::QASYMM8:
+                case DataType::QASYMM8_SIGNED:
+                {
+                    auto k = std::make_unique<kernels::CpuGemmLowpQuantizeDownInt32ScaleKernel>();
+                    k->configure(src, bias, dst, &info);
+                    _kernel = std::move(k);
+                    break;
+                }
+                default:
+                {
+                    ARM_COMPUTE_ERROR("Unsupported output data type.");
+                    break;
+                }
+            }
+            break;
+        }
+        default:
+            ARM_COMPUTE_ERROR("Unsupported GEMMLowpOutputStage type.");
+    }
+}
+
+Status CpuGemmLowpOutputStage::validate(const ITensorInfo *src, const ITensorInfo *bias, const ITensorInfo *dst, const GEMMLowpOutputStageInfo &info)
+{
+    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src, dst);
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG(dst->data_type() == DataType::UNKNOWN, "CpuGemmLowpOutputStage cannot be used with UNKNOWN output data type.");
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(dst, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED, DataType::QSYMM16);
+    ARM_COMPUTE_RETURN_ERROR_ON((info.type != GEMMLowpOutputStageType::QUANTIZE_DOWN) && (info.type != GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT));
+
+    switch(info.type)
+    {
+        case GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT:
+        {
+            switch(dst->data_type())
+            {
+                case DataType::QASYMM8:
+                    return kernels::CpuGemmLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel::validate(src, bias, dst, info.gemmlowp_min_bound, info.gemmlowp_max_bound);
+                case DataType::QASYMM8_SIGNED:
+                    return kernels::CpuGemmLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel::validate(src, bias, dst, info.gemmlowp_min_bound, info.gemmlowp_max_bound);
+                case DataType::QSYMM16:
+                    return kernels::CpuGemmLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel::validate(src, bias, dst, info.gemmlowp_min_bound, info.gemmlowp_max_bound);
+                default:
+                    return ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Unsupported output data type.");
+            }
+        }
+        case GEMMLowpOutputStageType::QUANTIZE_DOWN:
+        {
+            switch(dst->data_type())
+            {
+                case DataType::QASYMM8:
+                case DataType::QASYMM8_SIGNED:
+                    return kernels::CpuGemmLowpQuantizeDownInt32ScaleKernel::validate(src, bias, dst, &info);
+                default:
+                    return ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Unsupported output data type.");
+            }
+        }
+        default:
+            return ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Unsupported GEMMLowpOutputStage type.");
+    }
+}
+
+void CpuGemmLowpOutputStage::run(ITensorPack &tensors)
+{
+    NEScheduler::get().schedule_op(_kernel.get(), Window::DimY, _kernel->window(), tensors);
+}
+} // namespace cpu
+} // namespace arm_compute
\ No newline at end of file
diff --git a/src/runtime/cpu/operators/CpuGemmLowpOutputStage.h b/src/runtime/cpu/operators/CpuGemmLowpOutputStage.h
new file mode 100644
index 0000000000..848aaea3e7
--- /dev/null
+++ b/src/runtime/cpu/operators/CpuGemmLowpOutputStage.h
@@ -0,0 +1,90 @@
+/*
+ * Copyright (c) 2021 Arm Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#ifndef ARM_COMPUTE_CPU_GEMMLOWP_OUTPUTSTAGE_H
+#define ARM_COMPUTE_CPU_GEMMLOWP_OUTPUTSTAGE_H
+
+#include "arm_compute/core/Types.h"
+#include "src/runtime/cpu/ICpuOperator.h"
+
+/** This file contains all available output stages for GEMMLowp.
+ *
+ *  In gemmlowp, the "output stage" is the process that takes a final int32 accumulator value (the output of @ref NEGEMMLowpMatrixMultiplyCore),
+ *  and processes it to obtain the final ASYMM8 value.
+ *
+ *  More information about the GEMMLowp output stage can be found at https://github.com/google/gemmlowp/blob/master/doc/output.md
+ */
+
+namespace arm_compute
+{
+namespace cpu
+{
+/** Basic function to execute GEMMLowpQuantizeDown kernels.
+ *
+ *  This function calls the following kernels:
+ *
+ * -# @ref kernels::CpuGemmLowpQuantizeDownInt32ScaleKernel
+ * -# @ref kernels::CpuGemmLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel
+ * -# @ref kernels::CpuGemmLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel
+ * -# @ref kernels::CpuGemmLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel
+*/
+class CpuGemmLowpOutputStage : public ICpuOperator
+{
+public:
+    /** Constructor */
+    CpuGemmLowpOutputStage() = default;
+    /** Default destructor */
+    ~CpuGemmLowpOutputStage() = default;
+    /** Initialise the kernel's inputs, output
+     *
+     * Valid data layouts:
+     * - All
+     *
+     * Valid data type configurations:
+     * |src0           |src1          |dst           |
+     * |:--------------|:-------------|:-------------|
+     * |S32            |S32           |QASYMM8       |
+     * |S32            |S32           |QASYMM8_SIGNED|
+     * |S32            |S32           |QSYMM16       |
+     *
+     * @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/QSYMM16
+     * @param[in]  info GEMMLowp output stage metadata.
+     */
+    void configure(ITensorInfo *src, ITensorInfo *bias, ITensorInfo *dst, const GEMMLowpOutputStageInfo &info);
+    /** Static function to check if given info will lead to a valid configuration
+     *
+     * Similar to CpuGemmLowpOutputStage::configure()
+     *
+     * @return a status
+     */
+    static Status validate(const ITensorInfo *src, const ITensorInfo *bias, const ITensorInfo *dst, const GEMMLowpOutputStageInfo &info);
+
+    // Inherited methods overridden:
+    void run(ITensorPack &tensors) override;
+};
+} // namespace cpu
+} // namespace arm_compute
+#endif /* ARM_COMPUTE_CPU_GEMMLOWP_OUTPUTSTAGE_H */
diff --git a/tests/validation/NEON/GEMMLowp.cpp b/tests/validation/NEON/GEMMLowp.cpp
index 9d075e12c1..5d87330982 100644
--- a/tests/validation/NEON/GEMMLowp.cpp
+++ b/tests/validation/NEON/GEMMLowp.cpp
@@ -138,415 +138,8 @@ FIXTURE_DATA_TEST_CASE(RunLarge, NEGEMMLowpMatrixMultiplyCoreFusedOffsetOutputFi
 }
 TEST_SUITE_END() // FusedOffsetOutput
 TEST_SUITE_END() // MatrixMultiplyCore
-
-TEST_SUITE(OutputStage)
-
-TEST_SUITE(QuantizeDownInt32Scale)
-
-TEST_SUITE(QASYMM8)
-
-const auto quantize_down_int32_to_uint8_scale_cases = framework::dataset::make("result_offset", -2, 1) * framework::dataset::make("result_mult_int", 1, 2) * framework::dataset::make("result_shift", 2,
-                                                      3)
-                                                      * framework::dataset::make("min", 0) * framework::dataset::make("max", 255) * framework::dataset::make("addBias", { false, true });
-
-const auto quantize_down_int32_to_uint8_scale_relu_cases = framework::dataset::make("result_offset", -2, 1) * framework::dataset::make("result_mult_int", 1,
-                                                           2)
-                                                           * framework::dataset::make("result_shift", 2, 3) * framework::dataset::make("min", 0, 2) * framework::dataset::make("max", 171, 174) * framework::dataset::make("addBias", { false, true });
-
-using NEGEMMLowpQuantizeDownInt32ScaleFixture = GEMMLowpQuantizeDownInt32ToUint8ScaleValidationFixture<Tensor, Accessor, NEGEMMLowpOutputStage>;
-
-// *INDENT-OFF*
-// clang-format off
-DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(zip(zip(
-    framework::dataset::make("InputAInfo", { TensorInfo(TensorShape(21U, 13U), 1, DataType::S32), // Input not a multiple of 16
-                                             TensorInfo(TensorShape(20U, 13U), 1, DataType::S32), // Wrong output data type
-                                          }),
-    framework::dataset::make("InputBInfo",{ TensorInfo(TensorShape(21U), 1, DataType::S32),
-                                            TensorInfo(TensorShape(20U), 1, DataType::S32),
-                                          })),
-    framework::dataset::make("OutputInfo",{ TensorInfo(TensorShape(21U, 13U), 1, DataType::QASYMM8),
-                                            TensorInfo(TensorShape(20U, 13U), 1, DataType::S32),
-                                           })),
-    framework::dataset::make("Min",{        0,
-                                            13,
-                                           })),
-    framework::dataset::make("Max",{        205,
-                                            180,
-                                           })),
-    framework::dataset::make("Expected", { true, false })),
-    a_info, b_info, output_info, min, max, expected)
-{
-
-    GEMMLowpOutputStageInfo output_stage = GEMMLowpOutputStageInfo();
-    output_stage.type        = GEMMLowpOutputStageType::QUANTIZE_DOWN;
-    output_stage.gemmlowp_min_bound        = min;
-    output_stage.gemmlowp_max_bound        = max;
-    output_stage.output_data_type = DataType::QASYMM8;
-
-    // Lock tensors
-    Status status =  NEGEMMLowpOutputStage::validate(&a_info.clone()->set_is_resizable(false),
-                                                                     &b_info.clone()->set_is_resizable(false),
-                                                                     &output_info.clone()->set_is_resizable(false),
-                                                                     output_stage);
-    ARM_COMPUTE_EXPECT(bool(status) == expected, framework::LogLevel::ERRORS);
-}
-// clang-format on
-// *INDENT-ON*
-
-TEST_CASE(NoPaddingAdded, framework::DatasetMode::PRECOMMIT)
-{
-    Tensor input1 = create_tensor<Tensor>(TensorShape(21U, 13U), DataType::S32);
-    Tensor input2 = create_tensor<Tensor>(TensorShape(21U, 1U), DataType::S32);
-    Tensor output = create_tensor<Tensor>(TensorShape(21U, 13U), DataType::QASYMM8);
-
-    GEMMLowpOutputStageInfo output_stage = GEMMLowpOutputStageInfo();
-    output_stage.type                    = GEMMLowpOutputStageType::QUANTIZE_DOWN;
-    output_stage.gemmlowp_min_bound      = 0;
-    output_stage.gemmlowp_max_bound      = 205;
-    output_stage.output_data_type        = DataType::QASYMM8;
-
-    NEGEMMLowpOutputStage f;
-    f.configure(&input1, &input2, &output, output_stage);
-
-    // Validate padding is zero
-    validate(input1.info()->padding(), PaddingSize());
-    validate(input2.info()->padding(), PaddingSize());
-    validate(output.info()->padding(), PaddingSize());
-}
-
-FIXTURE_DATA_TEST_CASE(RunSmall, NEGEMMLowpQuantizeDownInt32ScaleFixture, framework::DatasetMode::ALL, combine(datasets::SmallShapes(), quantize_down_int32_to_uint8_scale_cases))
-{
-    // Validate output
-    validate(Accessor(_target), _reference);
-}
-
-TEST_SUITE(BoundedReLu)
-FIXTURE_DATA_TEST_CASE(RunSmall, NEGEMMLowpQuantizeDownInt32ScaleFixture, framework::DatasetMode::ALL, combine(datasets::SmallShapes(), quantize_down_int32_to_uint8_scale_relu_cases))
-{
-    // Validate output
-    validate(Accessor(_target), _reference);
-}
-
-TEST_SUITE_END() // BoundedReLu
-
-TEST_SUITE_END() // QASYMM8
-
-TEST_SUITE(QASYMM8_SIGNED)
-
-const auto quantize_down_int32_to_int8_scale_cases = framework::dataset::make("result_offset", -2, 1) * framework::dataset::make("result_mult_int", 1, 2) * framework::dataset::make("result_shift", 2,
-                                                     3)
-                                                     * framework::dataset::make("min", 0) * framework::dataset::make("max", 0) * framework::dataset::make("addBias", { false, true });
-
-const auto quantize_down_int32_to_int8_scale_relu_cases = framework::dataset::make("result_offset", -2, 1) * framework::dataset::make("result_mult_int", 1,
-                                                          2)
-                                                          * framework::dataset::make("result_shift", 2, 3) * framework::dataset::make("min", -100, -98) * framework::dataset::make("max", 71, 74) * framework::dataset::make("addBias", { false, true });
-
-using NEGEMMLowpQuantizeDownInt32ScaleFixture = GEMMLowpQuantizeDownInt32ToInt8ScaleValidationFixture<Tensor, Accessor, NEGEMMLowpOutputStage>;
-
-// *INDENT-OFF*
-// clang-format off
-DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(zip(zip(
-    framework::dataset::make("InputAInfo", { TensorInfo(TensorShape(21U, 13U), 1, DataType::S32), // Input not a multiple of 16
-                                             TensorInfo(TensorShape(21U, 13U), 1, DataType::S32), // Invalid min and max
-                                             TensorInfo(TensorShape(20U, 13U), 1, DataType::S32), // Wrong output data type
-                                          }),
-    framework::dataset::make("InputBInfo",{ TensorInfo(TensorShape(21U), 1, DataType::S32),
-                                            TensorInfo(TensorShape(21U), 1, DataType::S32),
-                                            TensorInfo(TensorShape(20U), 1, DataType::S32),
-                                          })),
-    framework::dataset::make("OutputInfo",{ TensorInfo(TensorShape(21U, 13U), 1, DataType::QASYMM8_SIGNED),
-                                            TensorInfo(TensorShape(21U, 13U), 1, DataType::QASYMM8_SIGNED),
-                                            TensorInfo(TensorShape(20U, 13U), 1, DataType::S32),
-                                           })),
-    framework::dataset::make("Min",{        -10,
-                                            -200,
-                                            -113,
-                                           })),
-    framework::dataset::make("Max",{        105,
-                                            300,
-                                            -18,
-                                           })),
-    framework::dataset::make("Expected", { true, false, false })),
-    a_info, b_info, output_info, min, max, expected)
-{
-    GEMMLowpOutputStageInfo output_stage = GEMMLowpOutputStageInfo();
-    output_stage.type        = GEMMLowpOutputStageType::QUANTIZE_DOWN;
-    output_stage.gemmlowp_min_bound        = min;
-    output_stage.gemmlowp_max_bound        = max;
-    output_stage.output_data_type = DataType::QASYMM8_SIGNED;
-
-    // Lock tensors
-    Status status =  NEGEMMLowpOutputStage::validate(&a_info.clone()->set_is_resizable(false),
-                                                                     &b_info.clone()->set_is_resizable(false),
-                                                                     &output_info.clone()->set_is_resizable(false),
-                                                                     output_stage);
-    ARM_COMPUTE_EXPECT(bool(status) == expected, framework::LogLevel::ERRORS);
-}
-// clang-format on
-// *INDENT-ON*
-
-FIXTURE_DATA_TEST_CASE(RunSmall, NEGEMMLowpQuantizeDownInt32ScaleFixture, framework::DatasetMode::ALL, combine(datasets::SmallShapes(), quantize_down_int32_to_int8_scale_cases))
-{
-    // Validate output
-    validate(Accessor(_target), _reference);
-}
-
-TEST_SUITE(BoundedReLu)
-FIXTURE_DATA_TEST_CASE(RunSmall, NEGEMMLowpQuantizeDownInt32ScaleFixture, framework::DatasetMode::ALL, combine(datasets::SmallShapes(), quantize_down_int32_to_int8_scale_relu_cases))
-{
-    // Validate output
-    validate(Accessor(_target), _reference);
-}
-
-TEST_SUITE_END() // BoundedReLu
-
-TEST_SUITE_END() // QASYMM8_SIGNED
-
-TEST_SUITE_END() // QuantizeDownInt32Scale
-
-TEST_SUITE(QuantizeDownInt32ToUint8ScaleByFixedPoint)
-
-const auto quantize_down_int32_to_uint8_scale_by_fixedpoint_cases = framework::dataset::make("result_fixedpoint_multiplier", 254601600, 254601602) * framework::dataset::make("result_shift", 1,
-                                                                    2)
-                                                                    * framework::dataset::make("result_offset_after_shift", 2, 3) * framework::dataset::make("min", 0) * framework::dataset::make("max", 255) * framework::dataset::make("addBias", { false, true });
-
-const auto quantize_down_int32_to_uint8_scale_by_fixedpoint_relu_cases = framework::dataset::make("result_fixedpoint_multiplier", 254601600, 254601602) * framework::dataset::make("result_shift", 1,
-                                                                         2)
-                                                                         * framework::dataset::make("result_offset_after_shift", 2, 3) * framework::dataset::make("min", 0, 2) * framework::dataset::make("max", 171, 174) * framework::dataset::make("addBias", { false, true });
-
-using NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointFixture =
-    GEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointValidationFixture<Tensor, Accessor, NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPoint>;
-
-using NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointFixture =
-    GEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointValidationFixture<Tensor, Accessor, NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPoint>;
-
-// *INDENT-OFF*
-// clang-format off
-DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(zip(zip(
-    framework::dataset::make("InputAInfo", { TensorInfo(TensorShape(21U, 13U), 1, DataType::S32), // Input not a multiple of 16
-                                             TensorInfo(TensorShape(20U, 13U), 1, DataType::S32), // Wrong output data type
-                                          }),
-    framework::dataset::make("InputBInfo",{ TensorInfo(TensorShape(21U), 1, DataType::S32),
-                                            TensorInfo(TensorShape(20U), 1, DataType::S32),
-                                          })),
-    framework::dataset::make("OutputInfo",{ TensorInfo(TensorShape(21U, 13U), 1, DataType::QASYMM8),
-                                            TensorInfo(TensorShape(20U, 13U), 1, DataType::S32),
-                                           })),
-    framework::dataset::make("Min",{        0,
-                                            13,
-                                           })),
-    framework::dataset::make("Max",{        205,
-                                            180,
-                                           })),
-    framework::dataset::make("Expected", { true, false })),
-    a_info, b_info, output_info, min, max, expected)
-{
-    // Lock tensors
-    Status status =  NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPoint::validate(&a_info.clone()->set_is_resizable(false),
-                                                                                 &b_info.clone()->set_is_resizable(false),
-                                                                                 &output_info.clone()->set_is_resizable(false),
-                                                                                 min,
-                                                                                 max);
-    ARM_COMPUTE_EXPECT(bool(status) == expected, framework::LogLevel::ERRORS);
-}
-// clang-format on
-// *INDENT-ON*
-
-FIXTURE_DATA_TEST_CASE(RunSmall, NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointFixture, framework::DatasetMode::ALL, combine(datasets::SmallShapes(),
-                       quantize_down_int32_to_uint8_scale_by_fixedpoint_cases))
-{
-    // Validate output
-    validate(Accessor(_target), _reference);
-}
-
-FIXTURE_DATA_TEST_CASE(RunLarge, NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointFixture, framework::DatasetMode::NIGHTLY, combine(datasets::LargeShapes(),
-                       quantize_down_int32_to_uint8_scale_by_fixedpoint_cases))
-{
-    // Validate output
-    validate(Accessor(_target), _reference);
-}
-
-TEST_SUITE(BoundedReLu)
-FIXTURE_DATA_TEST_CASE(RunSmall, NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointFixture, framework::DatasetMode::ALL, combine(datasets::SmallShapes(),
-                       quantize_down_int32_to_uint8_scale_by_fixedpoint_relu_cases))
-{
-    // Validate output
-    validate(Accessor(_target), _reference);
-}
-
-FIXTURE_DATA_TEST_CASE(RunLarge, NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointFixture, framework::DatasetMode::NIGHTLY, combine(datasets::LargeShapes(),
-                       quantize_down_int32_to_uint8_scale_by_fixedpoint_relu_cases))
-{
-    // Validate output
-    validate(Accessor(_target), _reference);
-}
-TEST_SUITE_END() // BoundedReLu
-
-TEST_SUITE_END() // QuantizeDownInt32ToUint8ScaleByFixedPoint
-
-TEST_SUITE(QuantizeDownInt32ToInt8ScaleByFixedPoint)
-
-const auto quantize_down_int32_to_int8_scale_by_fixedpoint_cases = framework::dataset::make("result_fixedpoint_multiplier", 254601600, 254601602) * framework::dataset::make("result_shift", 1,
-                                                                   2)
-                                                                   * framework::dataset::make("result_offset_after_shift", 2, 3) * framework::dataset::make("min", -128) * framework::dataset::make("max", 128) * framework::dataset::make("addBias", { false, true });
-
-const auto quantize_down_int32_to_int8_scale_by_fixedpoint_relu_cases = framework::dataset::make("result_fixedpoint_multiplier", 254601600, 254601602) * framework::dataset::make("result_shift", 1,
-                                                                        2)
-                                                                        * framework::dataset::make("result_offset_after_shift", 2, 3) * framework::dataset::make("min", -2, 0) * framework::dataset::make("max", 1, 3) * framework::dataset::make("addBias", { false, true });
-
-using NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointFixture =
-    GEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointValidationFixture<Tensor, Accessor, NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPoint>;
-
-// *INDENT-OFF*
-// clang-format off
-DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(zip(zip(
-        framework::dataset::make("InputAInfo", { TensorInfo(TensorShape(21U, 13U), 1, DataType::F32), // Invalid input data type
-                                                 TensorInfo(TensorShape(20U, 13U), 1, DataType::S32), // Wrong output data type
-                                                 TensorInfo(TensorShape(21U, 13U), 1, DataType::S32),
-        }),
-        framework::dataset::make("InputBInfo",{ TensorInfo(TensorShape(21U), 1, DataType::S32),
-                                                TensorInfo(TensorShape(20U), 1, DataType::S32),
-                                                TensorInfo(TensorShape(21U), 1, DataType::S32),
-        })),
-        framework::dataset::make("OutputInfo",{ TensorInfo(TensorShape(21U, 13U), 1, DataType::QASYMM8_SIGNED),
-                                                TensorInfo(TensorShape(20U, 13U), 1, DataType::S32),
-                                                TensorInfo(TensorShape(21U, 13U), 1, DataType::QASYMM8_SIGNED),
-        })),
-        framework::dataset::make("Min",{ -110,
-                                         -113,
-                                         -113,
-        })),
-        framework::dataset::make("Max",{ 87,
-                                         97,
-                                         97,
-        })),
-        framework::dataset::make("Expected", { false, false, true })),
-               a_info, b_info, output_info, min, max, expected)
-{
-    // Lock tensors
-    Status status =  NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPoint::validate(&a_info.clone()->set_is_resizable(false),
-                                                                                  &b_info.clone()->set_is_resizable(false),
-                                                                                  &output_info.clone()->set_is_resizable(false),
-                                                                                  min,
-                                                                                  max);
-    ARM_COMPUTE_EXPECT(bool(status) == expected, framework::LogLevel::ERRORS);
-}
-// clang-format on
-// *INDENT-ON*
-
-FIXTURE_DATA_TEST_CASE(RunSmall, NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointFixture, framework::DatasetMode::ALL, combine(datasets::SmallShapes(),
-                       quantize_down_int32_to_int8_scale_by_fixedpoint_cases))
-{
-    // Validate output
-    validate(Accessor(_target), _reference);
-}
-
-TEST_SUITE(BoundedReLu)
-FIXTURE_DATA_TEST_CASE(RunSmall, NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointFixture, framework::DatasetMode::ALL, combine(datasets::SmallShapes(),
-                       quantize_down_int32_to_int8_scale_by_fixedpoint_relu_cases))
-{
-    // Validate output
-    validate(Accessor(_target), _reference);
-}
-TEST_SUITE_END() // BoundedReLu
-TEST_SUITE_END() // QuantizeDownInt32ToInt8ScaleByFixedPoint
-
-TEST_SUITE(QuantizeDownInt32ToInt16ScaleByFixedPoint)
-
-const auto quantize_down_int32_to_int16_scale_by_fixedpoint_cases = framework::dataset::make("result_fixedpoint_multiplier", 254601600, 254601602) * framework::dataset::make("result_shift", 1,
-                                                                    2)
-                                                                    * framework::dataset::make("min", -32768) * framework::dataset::make("max", 32767) * framework::dataset::make("addBias", { false, true });
-
-const auto quantize_down_int32_to_int16_scale_by_fixedpoint_relu_cases = framework::dataset::make("result_fixedpoint_multiplier", 254601600, 254601602) * framework::dataset::make("result_shift", 1,
-                                                                         2)
-                                                                         * framework::dataset::make("min", -2, 0) * framework::dataset::make("max", 1, 3) * framework::dataset::make("addBias", { false, true });
-const auto quantize_down_int32_to_int16_scale_by_fixedpoint_multgreat1_cases = framework::dataset::make("result_fixedpoint_multiplier", 1073741823,
-                                                                                                        1073741825)
-                                                                               * framework::dataset::make("result_shift", -3,
-                                                                                                          -2)
-                                                                               * framework::dataset::make("min", -32768) * framework::dataset::make("max", 32767) * framework::dataset::make("addBias", { false, true });
-
-const auto quantize_down_int32_to_int16_scale_by_fixedpoint_multgreat1_relu_cases = framework::dataset::make("result_fixedpoint_multiplier", 254601600,
-                                                                                                             254601602)
-                                                                                    * framework::dataset::make("result_shift", -3,
-                                                                                                               -1)
-                                                                                    * framework::dataset::make("min", -2, 0) * framework::dataset::make("max", 1, 3) * framework::dataset::make("addBias", { false, true });
-
-using NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointFixture =
-    GEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointValidationFixture<Tensor, Accessor, NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPoint>;
-
-// *INDENT-OFF*
-// clang-format off
-DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(zip(zip(
-    framework::dataset::make("InputAInfo", { TensorInfo(TensorShape(21U, 13U), 1, DataType::S32), // Input not a multiple of 16
-                                             TensorInfo(TensorShape(20U, 13U), 1, DataType::S32), // Wrong output data type
-                                          }),
-    framework::dataset::make("InputBInfo",{ TensorInfo(TensorShape(21U), 1, DataType::S32),
-                                            TensorInfo(TensorShape(20U), 1, DataType::S32),
-                                          })),
-    framework::dataset::make("OutputInfo",{ TensorInfo(TensorShape(21U, 13U), 1, DataType::QSYMM16),
-                                            TensorInfo(TensorShape(20U, 13U), 1, DataType::S32),
-                                           })),
-    framework::dataset::make("Min",{        -205,
-                                            -180,
-                                           })),
-    framework::dataset::make("Max",{        205,
-                                            180,
-                                           })),
-    framework::dataset::make("Expected", { true, false })),
-    a_info, b_info, output_info, min, max, expected)
-{
-    // Lock tensors
-    Status status =  NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPoint::validate(&a_info.clone()->set_is_resizable(false),
-                                                                                 &b_info.clone()->set_is_resizable(false),
-                                                                                 &output_info.clone()->set_is_resizable(false),
-                                                                                 min,
-                                                                                 max);
-    ARM_COMPUTE_EXPECT(bool(status) == expected, framework::LogLevel::ERRORS);
-}
-// clang-format on
-// *INDENT-ON*
-
-TEST_SUITE(NoRelu)
-TEST_SUITE(MultSmallerEq1)
-FIXTURE_DATA_TEST_CASE(RunSmall, NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointFixture, framework::DatasetMode::ALL, combine(datasets::SmallShapes(),
-                       quantize_down_int32_to_int16_scale_by_fixedpoint_cases))
-{
-    // Validate output
-    validate(Accessor(_target), _reference);
-}
-TEST_SUITE_END() // MultSmallerEq1
-TEST_SUITE(MultGreater1)
-FIXTURE_DATA_TEST_CASE(RunSmall, NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointFixture, framework::DatasetMode::ALL, combine(datasets::SmallShapes(),
-                       quantize_down_int32_to_int16_scale_by_fixedpoint_multgreat1_cases))
-{
-    // Validate output
-    validate(Accessor(_target), _reference);
-}
-TEST_SUITE_END() // MultGreater1
-TEST_SUITE_END() // NoRelu
-TEST_SUITE(BoundedReLu)
-TEST_SUITE(MultSmallerEq1)
-FIXTURE_DATA_TEST_CASE(RunSmall, NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointFixture, framework::DatasetMode::ALL, combine(datasets::SmallShapes(),
-                       quantize_down_int32_to_int16_scale_by_fixedpoint_relu_cases))
-{
-    // Validate output
-    validate(Accessor(_target), _reference);
-}
-TEST_SUITE_END() // MultSmallerEq1
-TEST_SUITE(MultGreater1)
-FIXTURE_DATA_TEST_CASE(RunSmall, NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointFixture, framework::DatasetMode::ALL, combine(datasets::SmallShapes(),
-                       quantize_down_int32_to_int16_scale_by_fixedpoint_multgreat1_relu_cases))
-{
-    // Validate output
-    validate(Accessor(_target), _reference);
-}
-TEST_SUITE_END() // MultGreater1
-TEST_SUITE_END() // BoundedReLu
-TEST_SUITE_END() // QuantizeDownInt32ToInt16ScaleByFixedPoint
-TEST_SUITE_END() // OutputStage
 TEST_SUITE_END() // GEMMLowp
-TEST_SUITE_END() // Neon
+TEST_SUITE_END() // NEON
 } // namespace validation
 } // namespace test
 } // namespace arm_compute