Port Arm(R) Neon(TM) Scale to new API

Partially resolves: COMPMID-4190

Change-Id: I0c1e32ff6176775c9b7bf547899a791fd318ba0a
Signed-off-by: Manuel Bottini <manuel.bottini@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/5192
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: TeresaARM <teresa.charlinreyes@arm.com>
Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Reviewed-by: Sheri Zhang <sheri.zhang@arm.com>

24 files changed, 712 insertions, 468 deletions

diff --git a/Android.bp b/Android.bp
index 3468bd90a1..6dce0e40bc 100644
--- a/Android.bp
+++ b/Android.bp
@@ -232,7 +232,6 @@ cc_library_static {
         "src/core/NEON/kernels/NERemapKernel.cpp",
         "src/core/NEON/kernels/NEReorgLayerKernel.cpp",
         "src/core/NEON/kernels/NEReverseKernel.cpp",
-        "src/core/NEON/kernels/NEScaleKernel.cpp",
         "src/core/NEON/kernels/NESelectKernel.cpp",
         "src/core/NEON/kernels/NESpaceToBatchLayerKernel.cpp",
         "src/core/NEON/kernels/NESpaceToDepthLayerKernel.cpp",
@@ -305,15 +304,6 @@ cc_library_static {
         "src/core/NEON/kernels/convolution/winograd/winograd_transforms/weights_4x4_3x3_fp16_fp16_integers.cpp",
         "src/core/NEON/kernels/convolution/winograd/winograd_transforms/weights_4x4_3x3_fp32_fp32_integers.cpp",
         "src/core/NEON/kernels/convolution/winograd/winograd_transforms/weights_6_3_fp32_fp32_integers.cpp",
-        "src/core/NEON/kernels/scale/impl/NEON/fp16.cpp",
-        "src/core/NEON/kernels/scale/impl/NEON/integer.cpp",
-        "src/core/NEON/kernels/scale/impl/NEON/qasymm8.cpp",
-        "src/core/NEON/kernels/scale/impl/NEON/qasymm8_signed.cpp",
-        "src/core/NEON/kernels/scale/impl/SVE/fp16.cpp",
-        "src/core/NEON/kernels/scale/impl/SVE/fp32.cpp",
-        "src/core/NEON/kernels/scale/impl/SVE/integer.cpp",
-        "src/core/NEON/kernels/scale/impl/SVE/qasymm8.cpp",
-        "src/core/NEON/kernels/scale/impl/SVE/qasymm8_signed.cpp",
         "src/core/PyramidInfo.cpp",
         "src/core/Rounding.cpp",
         "src/core/Size2D.cpp",
@@ -339,6 +329,7 @@ cc_library_static {
         "src/core/cpu/kernels/CpuPoolingKernel.cpp",
         "src/core/cpu/kernels/CpuQuantizationKernel.cpp",
         "src/core/cpu/kernels/CpuReshapeKernel.cpp",
+        "src/core/cpu/kernels/CpuScaleKernel.cpp",
         "src/core/cpu/kernels/CpuSoftmaxKernel.cpp",
         "src/core/cpu/kernels/CpuSubKernel.cpp",
         "src/core/cpu/kernels/activation/NEON/fp16.cpp",
@@ -366,6 +357,15 @@ cc_library_static {
         "src/core/cpu/kernels/pooling/neon/nchw/all.cpp",
         "src/core/cpu/kernels/pooling/neon/qasymm8.cpp",
         "src/core/cpu/kernels/pooling/neon/qasymm8_signed.cpp",
+        "src/core/cpu/kernels/scale/neon/fp16.cpp",
+        "src/core/cpu/kernels/scale/neon/integer.cpp",
+        "src/core/cpu/kernels/scale/neon/qasymm8.cpp",
+        "src/core/cpu/kernels/scale/neon/qasymm8_signed.cpp",
+        "src/core/cpu/kernels/scale/sve/fp16.cpp",
+        "src/core/cpu/kernels/scale/sve/fp32.cpp",
+        "src/core/cpu/kernels/scale/sve/integer.cpp",
+        "src/core/cpu/kernels/scale/sve/qasymm8.cpp",
+        "src/core/cpu/kernels/scale/sve/qasymm8_signed.cpp",
         "src/core/cpu/kernels/sub/neon/integer.cpp",
         "src/core/cpu/kernels/sub/neon/qasymm8.cpp",
         "src/core/cpu/kernels/sub/neon/qasymm8_signed.cpp",
@@ -667,6 +667,7 @@ cc_library_static {
         "src/runtime/cpu/operators/CpuPoolingAssemblyDispatch.cpp",
         "src/runtime/cpu/operators/CpuQuantization.cpp",
         "src/runtime/cpu/operators/CpuReshape.cpp",
+        "src/runtime/cpu/operators/CpuScale.cpp",
         "src/runtime/cpu/operators/CpuSoftmax.cpp",
         "src/runtime/cpu/operators/CpuSub.cpp",
         "src/runtime/gpu/cl/operators/ClActivation.cpp",
diff --git a/arm_compute/runtime/NEON/functions/NECropResize.h b/arm_compute/runtime/NEON/functions/NECropResize.h
index 5c3733f8ee..7dcf925650 100644
--- a/arm_compute/runtime/NEON/functions/NECropResize.h
+++ b/arm_compute/runtime/NEON/functions/NECropResize.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2019-2020 Arm Limited.
+ * Copyright (c) 2019-2021 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -31,6 +31,7 @@
 namespace arm_compute
 {
 // Forward Declarations
+class Tensor;
 class ITensor;
 class NECropKernel;
 
diff --git a/arm_compute/runtime/NEON/functions/NEScale.h b/arm_compute/runtime/NEON/functions/NEScale.h
index fceda83510..45658a7cd3 100644
--- a/arm_compute/runtime/NEON/functions/NEScale.h
+++ b/arm_compute/runtime/NEON/functions/NEScale.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020 Arm Limited.
+ * Copyright (c) 2016-2021 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -24,24 +24,28 @@
 #ifndef ARM_COMPUTE_NESCALEIMAGE_H
 #define ARM_COMPUTE_NESCALEIMAGE_H
 
+#include "arm_compute/runtime/IFunction.h"
+
 #include "arm_compute/core/KernelDescriptors.h"
 #include "arm_compute/core/Types.h"
-#include "arm_compute/runtime/NEON/INESimpleFunctionNoBorder.h"
-#include "arm_compute/runtime/Tensor.h"
+#include "src/core/common/Macros.h"
+
+#include <memory>
 
 namespace arm_compute
 {
 class ITensor;
+class ITensorInfo;
 
-/** Basic function to run @ref NEScaleKernel */
-class NEScale : public INESimpleFunctionNoBorder
+/** Basic function to compute Scale */
+class NEScale : public IFunction
 {
 public:
-    /** Constructor
-     *
-     * Initialize NEScale
-     */
+    /** Default Constructor */
     NEScale();
+    /** Default Destructor */
+    ~NEScale();
+    ARM_COMPUTE_DISALLOW_COPY_ALLOW_MOVE(NEScale);
     /** Initialize the function's source, destination, interpolation type and border_mode.
      *
      * @param[in, out] input  Source tensor. Data type supported: QASYMM8/QASYMM8_SIGNED/U8/S16/F16/F32. (Written to only for @p border_mode != UNDEFINED)
@@ -59,10 +63,12 @@ public:
      */
     static Status validate(const ITensorInfo *input, const ITensorInfo *output, const ScaleKernelInfo &info);
 
+    // Inherited methods overridden:
+    void run() override;
+
 private:
-    Tensor _offsets; /**< Offset to access the element with NEAREST interpolation or the top-left element with BILINEAR interpolation in the input tensor */
-    Tensor _dx;      /**< Element's distance between the X real coordinate and the smallest X following integer */
-    Tensor _dy;      /**< Element's distance between the Y real coordinate and the smallest Y following integer */
+    struct Impl;
+    std::unique_ptr<Impl> _impl;
 };
 } // namespace arm_compute
 #endif /*ARM_COMPUTE_NESCALEIMAGE_H */
diff --git a/docs/00_introduction.dox b/docs/00_introduction.dox
index 1bab6e57d1..d877236c66 100644
--- a/docs/00_introduction.dox
+++ b/docs/00_introduction.dox
@@ -180,7 +180,7 @@ v21.02 Public major release
  - Add macOS support
  - Add Armv8-R AArch64 architecture support
  - Add SVE/SVE2 support for:
-   - @ref NEScaleKernel
+   - NEScaleKernel
    - @ref NEActivationLayer
    - @ref NEArithmeticAddition
    - @ref NEBatchNormalizationLayerKernel
@@ -260,7 +260,7 @@ v20.11 Public major release
    - @ref NERemapKernel
    - @ref NEGEMMInterleave4x4Kernel
    - @ref NEDirectConvolutionLayerKernel
-   - @ref NEScaleKernel
+   - NEScaleKernel
    - NELocallyConnectedMatrixMultiplyKernel
    - @ref NEGEMMLowpOffsetContributionKernel
    - @ref NEGEMMTranspose1xWKernel
@@ -534,7 +534,7 @@ v20.08 Public major release
    - @ref NECropKernel
    - CLCropKernel
  - Added aligh_corner support for nearest neighbor interpolation in:
-   - @ref NEScaleKernel
+   - NEScaleKernel
    - CLScaleKernel
  - New OpenCL kernels / functions:
    - @ref CLMaxUnpoolingLayerKernel
@@ -621,7 +621,7 @@ v20.05 Public major release
      - @ref CLReductionOperation
      - @ref CLReduceMean
      - @ref NEScale
-     - @ref NEScaleKernel
+     - NEScaleKernel
      - NEUpsampleLayer
      - @ref NECast
      - @ref NEReductionOperation
@@ -1139,7 +1139,7 @@ v18.05 Public major release
  - Added support for the memory manager in the graph API.
  - Enabled Winograd Convolution method in the graph API.
  - Added support for grouped convolutions in the graph API.
- - Replaced NEDeconvolutionLayerUpsampleKernel with @ref NEScaleKernel in @ref NEDeconvolutionLayer.
+ - Replaced NEDeconvolutionLayerUpsampleKernel with NEScaleKernel in @ref NEDeconvolutionLayer.
  - Added fast maths flag in @ref CLConvolutionLayer.
  - Added new tests and benchmarks in validation and benchmark frameworks
  - Merge Activation layer with Convolution Layer (Neon. CL, GLES)
diff --git a/src/core/NEON/NEKernels.h b/src/core/NEON/NEKernels.h
index 6a7e856d8b..973c0f2d31 100644
--- a/src/core/NEON/NEKernels.h
+++ b/src/core/NEON/NEKernels.h
@@ -82,7 +82,6 @@
 #include "src/core/NEON/kernels/NERemapKernel.h"
 #include "src/core/NEON/kernels/NEReorgLayerKernel.h"
 #include "src/core/NEON/kernels/NEReverseKernel.h"
-#include "src/core/NEON/kernels/NEScaleKernel.h"
 #include "src/core/NEON/kernels/NESelectKernel.h"
 #include "src/core/NEON/kernels/NESpaceToBatchLayerKernel.h"
 #include "src/core/NEON/kernels/NESpaceToDepthLayerKernel.h"
diff --git a/src/core/NEON/kernels/NEScaleKernel.h b/src/core/NEON/kernels/NEScaleKernel.h
deleted file mode 100644
index 32fa8d7fb2..0000000000
--- a/src/core/NEON/kernels/NEScaleKernel.h
+++ /dev/null
@@ -1,122 +0,0 @@
-/*
- * Copyright (c) 2016-2021 Arm Limited.
- *
- * SPDX-License-Identifier: MIT
- *
- * Permission is hereby granted, free of charge, to any person obtaining a copy
- * of this software and associated documentation files (the "Software"), to
- * deal in the Software without restriction, including without limitation the
- * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
- * sell copies of the Software, and to permit persons to whom the Software is
- * furnished to do so, subject to the following conditions:
- *
- * The above copyright notice and this permission notice shall be included in all
- * copies or substantial portions of the Software.
- *
- * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
- * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
- * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
- * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
- * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
- * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
- * SOFTWARE.
- */
-#ifndef ARM_COMPUTE_NESCALEKERNEL_H
-#define ARM_COMPUTE_NESCALEKERNEL_H
-
-#include "arm_compute/core/KernelDescriptors.h"
-#include "src/core/NEON/INEKernel.h"
-
-namespace arm_compute
-{
-class ITensor;
-
-/** Neon kernel to perform scaling on a tensor */
-class NEScaleKernel : public INEKernel
-{
-public:
-    const char *name() const override
-    {
-        return "NEScaleKernel";
-    }
-    /** Default constructor */
-    NEScaleKernel();
-    /** Prevent instances of this class from being copied (As this class contains pointers) */
-    NEScaleKernel(const NEScaleKernel &) = delete;
-    /** Prevent instances of this class from being copied (As this class contains pointers) */
-    NEScaleKernel &operator=(const NEScaleKernel &) = delete;
-    /** Allow instances of this class to be moved */
-    NEScaleKernel(NEScaleKernel &&) = default;
-    /** Allow instances of this class to be moved */
-    NEScaleKernel &operator=(NEScaleKernel &&) = default;
-    /** Default destructor */
-    ~NEScaleKernel() = default;
-
-    /** Initialise the kernel's inputs, output and interpolation policy
-     *
-     * @note dx, dy and offsets have the same dimensions (width and height) of the output tensor
-     * @note Using @p policy Area only supports data layout NCHW and input data type U8.
-     *
-     * @param[in]  input   Source tensor. Data types supported: QASYMM8/QASYMM8_SIGNED/U8/S16/F16/F32.
-     * @param[in]  dx      Pixel's distance between the X real coordinate and the smallest X following integer. Data type supported: F32
-     * @param[in]  dy      Pixel's distance between the Y real coordinate and the smallest Y following integer. Data type supported: F32
-     * @param[in]  offsets Offset to access the pixel with NEAREST interpolation or the top-left pixel with BILINEAR interpolation in the input tensor. Data type supported: S32.
-     * @param[out] output  Destination tensor. Data types supported: Same as @p input. All but the lowest two dimensions must be the same size as in the input tensor, i.e. scaling is only performed within the XY-plane.
-     * @param[in]  info    @ref ScaleKernelInfo to use for configuration
-     */
-    void configure(const ITensor *input, const ITensor *dx, const ITensor *dy, const ITensor *offsets, ITensor *output,
-                   const ScaleKernelInfo &info);
-    /** Static function to check if given info will lead to a valid configuration of @ref NEScaleKernel
-     *
-     * @note dx, dy and offsets have the same dimensions (width and height) of the output tensor
-     * @note Using @p policy Area only supports data layout NCHW and input data type U8.
-     *
-     * @param[in] input   Source tensor. Data types supported: QASYMM8/QASYMM8_SIGNED/U8/S16/F16/F32.
-     * @param[in] dx      Pixel's distance between the X real coordinate and the smallest X following integer. Data type supported: F32
-     * @param[in] dy      Pixel's distance between the Y real coordinate and the smallest Y following integer. Data type supported: F32
-     * @param[in] offsets Offset to access the pixel with NEAREST interpolation or the top-left pixel with BILINEAR interpolation in the input tensor. Data type supported: S32.
-     * @param[in] output  Destination tensor. Data types supported: Same as @p input. All but the lowest two dimensions must be the same size as in the input tensor, i.e. scaling is only performed within the XY-plane.
-     * @param[in] info    @ref ScaleKernelInfo to use for validation
-     */
-    static Status validate(const ITensorInfo *input, const ITensorInfo *dx, const ITensorInfo *dy, const ITensorInfo *offsets, ITensorInfo *output,
-                           const ScaleKernelInfo &info);
-
-    // Inherited methods overridden:
-    void run(const Window &window, const ThreadInfo &info) override;
-
-private:
-    /** function to perform scale using area interpolation on the given window
-     *
-     *  @note Used only in case down-sampling.
-     */
-    void scale_area_nchw_u8(const Window &window);
-
-    /** function to perform scale using bilinear interpolation on the given window */
-    template <typename T>
-    void scale_bilinear_nchw(const Window &window);
-    /** function to perform scale using bilinear interpolation on the given window */
-    template <typename T>
-    void scale_bilinear_qasymm(const Window &window);
-
-    /** function to perform scale using nearest neighbour on the given window */
-    template <typename T>
-    void scale_nearest_nchw(const Window &window);
-
-    /** Scale function to use for the particular function to use */
-    using ScaleFunctionPtr = void (NEScaleKernel::*)(const Window &window);
-
-    ScaleFunctionPtr    _func;
-    const ITensor      *_offsets;
-    const ITensor      *_dx;
-    const ITensor      *_dy;
-    const ITensor      *_input;
-    ITensor            *_output;
-    InterpolationPolicy _policy;
-    BorderMode          _border_mode;
-    PixelValue          _constant_border_value;
-    float               _sampling_offset;
-    bool                _align_corners;
-    DataLayout          _data_layout;
-};
-} // namespace arm_compute
-#endif /*ARM_COMPUTE_NESCALEKERNEL_H */
diff --git a/src/core/NEON/kernels/NEScaleKernel.cpp b/src/core/cpu/kernels/CpuScaleKernel.cpp
index 6b9aa51aaa..22d1332a55 100644
--- a/src/core/NEON/kernels/NEScaleKernel.cpp
+++ b/src/core/cpu/kernels/CpuScaleKernel.cpp
@@ -21,27 +21,32 @@
  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  * SOFTWARE.
  */
-#include "src/core/NEON/kernels/NEScaleKernel.h"
+#include "src/core/cpu/kernels/CpuScaleKernel.h"
 
 #include "arm_compute/core/Helpers.h"
 #include "arm_compute/core/Window.h"
 #include "arm_compute/core/utils/misc/Utility.h"
 #include "src/core/AccessWindowStatic.h"
 #include "src/core/CPP/Validate.h"
-#include "src/core/NEON/kernels/scale/impl/NEON/list.h"
-#include "src/core/NEON/kernels/scale/impl/SVE/list.h"
 #include "src/core/NEON/wrapper/wrapper.h"
 #include "src/core/common/Registrars.h"
+#include "src/core/cpu/kernels/scale/neon/list.h"
+#include "src/core/cpu/kernels/scale/sve/list.h"
 #include "src/core/helpers/AutoConfiguration.h"
 #include "src/core/helpers/ScaleHelpers.h"
 #include "src/core/helpers/WindowHelpers.h"
 #include "src/core/utils/ScaleUtils.h"
 #include "support/Rounding.h"
+
 #include <arm_neon.h>
 #include <map>
 
 namespace arm_compute
 {
+namespace cpu
+{
+namespace kernels
+{
 namespace
 {
 struct ScaleSelectorData
@@ -145,24 +150,24 @@ const ScaleKernel *get_implementation(const ScaleSelectorData &data)
     return nullptr;
 }
 
-Status validate_arguments(const ITensorInfo *input, const ITensorInfo *dx, const ITensorInfo *dy,
-                          const ITensorInfo *offsets, ITensorInfo *output, const ScaleKernelInfo &info)
+Status validate_arguments(const ITensorInfo *src, const ITensorInfo *dx, const ITensorInfo *dy,
+                          const ITensorInfo *offsets, ITensorInfo *dst, const ScaleKernelInfo &info)
 {
-    const auto *uk = get_implementation(ScaleSelectorData{ input->data_type() });
+    const auto *uk = get_implementation(ScaleSelectorData{ src->data_type() });
     ARM_COMPUTE_RETURN_ERROR_ON(uk == nullptr || uk->ukernel == nullptr);
 
-    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(output);
-    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
-    ARM_COMPUTE_RETURN_ERROR_ON(output == input);
+    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(dst);
+    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src, dst);
+    ARM_COMPUTE_RETURN_ERROR_ON(dst == src);
     ARM_COMPUTE_RETURN_ERROR_ON(info.sampling_policy != SamplingPolicy::CENTER && info.sampling_policy != SamplingPolicy::TOP_LEFT);
     ARM_COMPUTE_UNUSED(info.constant_border_value);
     ARM_COMPUTE_RETURN_ERROR_ON_MSG(info.use_padding, "Padding is not supported");
 
-    const DataLayout data_layout   = info.data_layout == DataLayout::UNKNOWN ? input->data_layout() : info.data_layout;
+    const DataLayout data_layout   = info.data_layout == DataLayout::UNKNOWN ? src->data_layout() : info.data_layout;
     const auto       width_index   = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH);
     const auto       height_index  = get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT);
-    const auto       output_width  = output->dimension(width_index);
-    const auto       output_height = output->dimension(height_index);
+    const auto       output_width  = dst->dimension(width_index);
+    const auto       output_height = dst->dimension(height_index);
     ARM_COMPUTE_RETURN_ERROR_ON(output_width == 0);
     ARM_COMPUTE_RETURN_ERROR_ON(output_height == 0);
 
@@ -183,41 +188,36 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *dx, const
     if(info.interpolation_policy == InterpolationPolicy::AREA)
     {
         ARM_COMPUTE_RETURN_ERROR_ON(data_layout != DataLayout::NCHW);
-        ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::U8);
+        ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src, 1, DataType::U8);
     }
 
     return Status{};
 }
 } // namespace
 
-NEScaleKernel::NEScaleKernel()
-    : _func(nullptr), _offsets(nullptr), _dx(nullptr), _dy(nullptr), _input(nullptr), _output(nullptr), _policy(), _border_mode(), _constant_border_value(PixelValue()), _sampling_offset(0),
-      _align_corners(false), _data_layout(DataLayout::UNKNOWN)
+CpuScaleKernel::CpuScaleKernel()
+    : _func(nullptr), _policy(), _border_mode(), _constant_border_value(PixelValue()), _sampling_offset(0), _align_corners(false), _data_layout(DataLayout::UNKNOWN)
 {
 }
 
-void NEScaleKernel::configure(const ITensor *input, const ITensor *dx, const ITensor *dy, const ITensor *offsets,
-                              ITensor *output, const ScaleKernelInfo &info)
+void CpuScaleKernel::configure(const ITensorInfo *src, const ITensorInfo *dx, const ITensorInfo *dy, const ITensorInfo *offsets,
+                               ITensorInfo *dst, const ScaleKernelInfo &info)
 {
-    ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
+    ARM_COMPUTE_UNUSED(dx, dy, offsets);
+    ARM_COMPUTE_ERROR_ON_NULLPTR(src, dst);
     // Perform validation step
-    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(),
-                                                  dx != nullptr ? dx->info() : nullptr,
-                                                  dy != nullptr ? dy->info() : nullptr,
-                                                  offsets != nullptr ? offsets->info() : nullptr,
-                                                  output->info(),
+    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(src,
+                                                  dx,
+                                                  dy,
+                                                  offsets,
+                                                  dst,
                                                   info));
 
     // Get data layout and width/height indices
-    _data_layout         = info.data_layout == DataLayout::UNKNOWN ? input->info()->data_layout() : info.data_layout;
+    _data_layout         = info.data_layout == DataLayout::UNKNOWN ? src->data_layout() : info.data_layout;
     const int idx_width  = get_data_layout_dimension_index(_data_layout, DataLayoutDimension::WIDTH);
     const int idx_height = get_data_layout_dimension_index(_data_layout, DataLayoutDimension::HEIGHT);
 
-    _input                 = input;
-    _output                = output;
-    _offsets               = offsets;
-    _dx                    = dx;
-    _dy                    = dy;
     _policy                = info.interpolation_policy;
     _border_mode           = info.border_mode;
     _constant_border_value = info.constant_border_value;
@@ -229,8 +229,8 @@ void NEScaleKernel::configure(const ITensor *input, const ITensor *dx, const ITe
     }
 
     // Compute the ratio between source width/height and destination width/height
-    const auto wr = scale_utils::calculate_resize_ratio(input->info()->dimension(idx_width), output->info()->dimension(idx_width), _align_corners);
-    const auto hr = scale_utils::calculate_resize_ratio(input->info()->dimension(idx_height), output->info()->dimension(idx_height), _align_corners);
+    const auto wr = scale_utils::calculate_resize_ratio(src->dimension(idx_width), dst->dimension(idx_width), _align_corners);
+    const auto hr = scale_utils::calculate_resize_ratio(src->dimension(idx_height), dst->dimension(idx_height), _align_corners);
 
     // Area interpolation behaves as Nearest Neighbour in case of up-sampling
     _policy = (_policy == InterpolationPolicy::AREA && wr <= 1.f && hr <= 1.f) ? InterpolationPolicy::NEAREST_NEIGHBOR : _policy;
@@ -241,37 +241,38 @@ void NEScaleKernel::configure(const ITensor *input, const ITensor *dx, const ITe
         _constant_border_value = PixelValue();
     }
 
+#ifdef ENABLE_NCHW_KERNELS
     // Configure scale function to run
     if(_data_layout == DataLayout::NCHW)
     {
         std::string function_to_call("scale_");
-        function_to_call += string_from_data_type(_input->info()->data_type()) + "_";
+        function_to_call += string_from_data_type(src->data_type()) + "_";
         function_to_call += string_from_data_layout(_data_layout) + "_";
         function_to_call += string_from_interpolation_policy(_policy);
 
         static std::map<std::string, ScaleFunctionPtr> map_function =
         {
-            { "scale_U8_NCHW_AREA_CONSTANT", &NEScaleKernel::scale_area_nchw_u8 },
+            { "scale_U8_NCHW_AREA_CONSTANT", &CpuScaleKernel::scale_area_nchw_u8 },
 
-            { "scale_U8_NCHW_BILINEAR", &NEScaleKernel::scale_bilinear_nchw<uint8_t> },
-            { "scale_U8_NCHW_NEAREST_NEIGHBOUR", &NEScaleKernel::scale_nearest_nchw<uint8_t> },
+            { "scale_U8_NCHW_BILINEAR", &CpuScaleKernel::scale_bilinear_nchw<uint8_t> },
+            { "scale_U8_NCHW_NEAREST_NEIGHBOUR", &CpuScaleKernel::scale_nearest_nchw<uint8_t> },
 
-            { "scale_QASYMM8_NCHW_BILINEAR", &NEScaleKernel::scale_bilinear_qasymm<uint8_t> },
-            { "scale_QASYMM8_NCHW_NEAREST_NEIGHBOUR", &NEScaleKernel::scale_nearest_nchw<uint8_t> },
+            { "scale_QASYMM8_NCHW_BILINEAR", &CpuScaleKernel::scale_bilinear_qasymm<uint8_t> },
+            { "scale_QASYMM8_NCHW_NEAREST_NEIGHBOUR", &CpuScaleKernel::scale_nearest_nchw<uint8_t> },
 
-            { "scale_QASYMM8_SIGNED_NCHW_BILINEAR", &NEScaleKernel::scale_bilinear_qasymm<int8_t> },
-            { "scale_QASYMM8_SIGNED_NCHW_NEAREST_NEIGHBOUR", &NEScaleKernel::scale_nearest_nchw<int8_t> },
+            { "scale_QASYMM8_SIGNED_NCHW_BILINEAR", &CpuScaleKernel::scale_bilinear_qasymm<int8_t> },
+            { "scale_QASYMM8_SIGNED_NCHW_NEAREST_NEIGHBOUR", &CpuScaleKernel::scale_nearest_nchw<int8_t> },
 
-            { "scale_S16_NCHW_BILINEAR", &NEScaleKernel::scale_bilinear_nchw<int16_t> },
-            { "scale_S16_NCHW_NEAREST_NEIGHBOUR", &NEScaleKernel::scale_nearest_nchw<int16_t> },
+            { "scale_S16_NCHW_BILINEAR", &CpuScaleKernel::scale_bilinear_nchw<int16_t> },
+            { "scale_S16_NCHW_NEAREST_NEIGHBOUR", &CpuScaleKernel::scale_nearest_nchw<int16_t> },
 
 #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
-            { "scale_F16_NCHW_BILINEAR", &NEScaleKernel::scale_bilinear_nchw<float16_t> },
-            { "scale_F16_NCHW_NEAREST_NEIGHBOUR", &NEScaleKernel::scale_nearest_nchw<float16_t> },
+            { "scale_F16_NCHW_BILINEAR", &CpuScaleKernel::scale_bilinear_nchw<float16_t> },
+            { "scale_F16_NCHW_NEAREST_NEIGHBOUR", &CpuScaleKernel::scale_nearest_nchw<float16_t> },
 #endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */
 
-            { "scale_F32_NCHW_BILINEAR", &NEScaleKernel::scale_bilinear_nchw<float> },
-            { "scale_F32_NCHW_NEAREST_NEIGHBOUR", &NEScaleKernel::scale_nearest_nchw<float> },
+            { "scale_F32_NCHW_BILINEAR", &CpuScaleKernel::scale_bilinear_nchw<float> },
+            { "scale_F32_NCHW_NEAREST_NEIGHBOUR", &CpuScaleKernel::scale_nearest_nchw<float> },
         };
         auto it = map_function.find(function_to_call);
         if(it != map_function.end())
@@ -279,19 +280,22 @@ void NEScaleKernel::configure(const ITensor *input, const ITensor *dx, const ITe
             _func = it->second;
         }
     }
+#endif // ENABLE_NCHW_KERNELS
 
     // Configure window
-    Window win = calculate_max_window(*output->info(), Steps());
-    INEKernel::configure(win);
+    Window win = calculate_max_window(*dst, Steps());
+    ICpuKernel::configure(win);
 }
 
+#ifdef ENABLE_NCHW_KERNELS
 template <typename T>
-void NEScaleKernel::scale_nearest_nchw(const Window &window)
+void CpuScaleKernel::scale_nearest_nchw(const ITensor *src, ITensor *dst, const ITensor *dx, const ITensor *dy, const ITensor *offsets, const Window &window)
 {
-    const size_t in_stride_x = _input->info()->dimension(0) + _input->info()->padding().left + _input->info()->padding().right;
+    ARM_COMPUTE_UNUSED(dx, dy);
+    const size_t in_stride_x = src->info()->dimension(0) + src->info()->padding().left + src->info()->padding().right;
 
     // Compute the ratio between source height and destination height
-    const auto hr = scale_utils::calculate_resize_ratio(_input->info()->dimension(1), _output->info()->dimension(1), _align_corners);
+    const auto hr = scale_utils::calculate_resize_ratio(src->info()->dimension(1), dst->info()->dimension(1), _align_corners);
 
     // Don't increment in X and Y direction for the input tensor
     // A pointer to the start of this plane is needed as base for the precomputed offsets
@@ -303,30 +307,32 @@ void NEScaleKernel::scale_nearest_nchw(const Window &window)
     Window win_off;
     win_off.set(Window::DimX, window[Window::DimX]);
     win_off.set(Window::DimY, window[Window::DimY]);
-    for(size_t d = Window::DimZ; d < _offsets->info()->num_dimensions(); ++d)
+    for(size_t d = Window::DimZ; d < offsets->info()->num_dimensions(); ++d)
     {
         win_off.set(d, Window::Dimension(0, 0, 0));
     }
 
     // Create iterators
-    Iterator in(_input, win_in);
-    Iterator out(_output, window);
-    Iterator offsets(_offsets, win_off);
+    Iterator src_i(src, win_in);
+    Iterator dst_i(dst, window);
+    Iterator offsets_i(offsets, win_off);
     execute_window_loop(window, [&](const Coordinates & id)
     {
-        const auto    offsets_ptr         = reinterpret_cast<const int32_t *>(offsets.ptr());
-        const auto    in_yi               = static_cast<int32_t>(_align_corners ? utils::rounding::round_half_away_from_zero((id.y() + _sampling_offset) * hr) : std::floor((id.y() + _sampling_offset) * hr));
-        const int32_t offset_row          = in_yi * in_stride_x;
-        *reinterpret_cast<T *>(out.ptr()) = *(reinterpret_cast<const T *>(in.ptr()) + offsets_ptr[0] + offset_row);
+        const auto offsets_ptr = reinterpret_cast<const int32_t *>(offsets_i.ptr());
+        const auto in_yi       = static_cast<int32_t>(_align_corners ? utils::rounding::round_half_away_from_zero((id.y() + _sampling_offset) * hr) : std::floor((
+                                                          id.y() + _sampling_offset)
+                                                      * hr));
+        const int32_t offset_row            = in_yi * in_stride_x;
+        *reinterpret_cast<T *>(dst_i.ptr()) = *(reinterpret_cast<const T *>(src_i.ptr()) + offsets_ptr[0] + offset_row);
     },
-    in, offsets, out);
+    src_i, offsets_i, dst_i);
 }
 
 template <typename T>
-void NEScaleKernel::scale_bilinear_nchw(const Window &window)
+void CpuScaleKernel::scale_bilinear_nchw(const ITensor *src, ITensor *dst, const ITensor *dx, const ITensor *dy, const ITensor *offsets, const Window &window)
 {
     // Compute the ratio between source height and destination height
-    const auto hr = scale_utils::calculate_resize_ratio(_input->info()->dimension(1), _output->info()->dimension(1), _align_corners);
+    const auto hr = scale_utils::calculate_resize_ratio(src->info()->dimension(1), dst->info()->dimension(1), _align_corners);
     Window     win_off;
     win_off.set(Window::DimX, window.x());
     win_off.set(Window::DimY, window.y());
@@ -337,20 +343,20 @@ void NEScaleKernel::scale_bilinear_nchw(const Window &window)
     win_in.set(Window::DimX, Window::Dimension(0, 0, 0));
     win_in.set(Window::DimY, Window::Dimension(0, 0, 0));
 
-    for(size_t d = Window::DimZ; d < _offsets->info()->num_dimensions(); ++d)
+    for(size_t d = Window::DimZ; d < offsets->info()->num_dimensions(); ++d)
     {
         win_off.set(d, Window::Dimension(0, 0, 0));
     }
 
-    Iterator in(_input, win_in);
-    Iterator out(_output, window);
-    Iterator offsets(_offsets, win_off);
-    Iterator dx(_dx, win_off);
-    Iterator dy(_dy, win_off);
+    Iterator src_i(src, win_in);
+    Iterator dst_i(dst, window);
+    Iterator offsets_i(offsets, win_off);
+    Iterator dx_i(dx, win_off);
+    Iterator dy_i(dy, win_off);
 
-    const int32_t in_dim_w    = _input->info()->dimension(0);
-    const int32_t in_dim_h    = _input->info()->dimension(1);
-    const int32_t in_stride_w = in_dim_w + _input->info()->padding().left + _input->info()->padding().right;
+    const int32_t in_dim_w    = src->info()->dimension(0);
+    const int32_t in_dim_h    = src->info()->dimension(1);
+    const int32_t in_stride_w = in_dim_w + src->info()->padding().left + src->info()->padding().right;
 
     if(_border_mode == BorderMode::CONSTANT)
     {
@@ -363,10 +369,10 @@ void NEScaleKernel::scale_bilinear_nchw(const Window &window)
         execute_window_loop(window, [&](const Coordinates & id)
         {
             const int32_t index_h       = std::floor((id.y() + _sampling_offset) * hr - _sampling_offset);
-            const auto    index_w       = *(reinterpret_cast<const int32_t *>(offsets.ptr()));
-            const auto    dx_val        = *(reinterpret_cast<const float *>(dx.ptr()));
-            const auto    dy_val        = *(reinterpret_cast<const float *>(dy.ptr()));
-            const auto    pixel_row_ptr = reinterpret_cast<const T *>(in.ptr());
+            const auto    index_w       = *(reinterpret_cast<const int32_t *>(offsets_i.ptr()));
+            const auto    dx_val        = *(reinterpret_cast<const float *>(dx_i.ptr()));
+            const auto    dy_val        = *(reinterpret_cast<const float *>(dy_i.ptr()));
+            const auto    pixel_row_ptr = reinterpret_cast<const T *>(src_i.ptr());
 
             const auto a00 = (0 <= index_w && index_w < in_dim_w && 0 <= index_h && index_h < in_dim_h) ? (*(pixel_row_ptr + index_w + index_h * in_stride_w)) : const_border_value;
             const auto a01 = (-1 <= index_w && index_w < in_dim_w - 1 && 0 <= index_h && index_h < in_dim_h) ? (*(pixel_row_ptr + index_w + 1 + index_h * in_stride_w)) : const_border_value;
@@ -379,19 +385,19 @@ void NEScaleKernel::scale_bilinear_nchw(const Window &window)
                              (*(pixel_row_ptr + index_w + 1 + index_h * in_stride_w + in_stride_w)) :
                              const_border_value;
 
-            *reinterpret_cast<T *>(out.ptr()) = static_cast<T>(scale_helpers::delta_bilinear(a00, a01, a10, a11, dx_val, dy_val));
+            *reinterpret_cast<T *>(dst_i.ptr()) = static_cast<T>(scale_helpers::delta_bilinear(a00, a01, a10, a11, dx_val, dy_val));
         },
-        in, offsets, dx, dy, out);
+        src_i, offsets_i, dx_i, dy_i, dst_i);
     }
     else if(_border_mode == BorderMode::REPLICATE)
     {
         execute_window_loop(window, [&](const Coordinates & id)
         {
             const int  index_h       = std::floor((id.y() + _sampling_offset) * hr - _sampling_offset);
-            const auto index_w       = *(reinterpret_cast<const int32_t *>(offsets.ptr()));
-            const auto dx_val        = *(reinterpret_cast<const float *>(dx.ptr()));
-            const auto dy_val        = *(reinterpret_cast<const float *>(dy.ptr()));
-            const auto pixel_row_ptr = reinterpret_cast<const T *>(in.ptr());
+            const auto index_w       = *(reinterpret_cast<const int32_t *>(offsets_i.ptr()));
+            const auto dx_val        = *(reinterpret_cast<const float *>(dx_i.ptr()));
+            const auto dy_val        = *(reinterpret_cast<const float *>(dy_i.ptr()));
+            const auto pixel_row_ptr = reinterpret_cast<const T *>(src_i.ptr());
 
             auto clamped_x  = utility::clamp<int>(index_w, 0, in_dim_w - 1);
             auto clamped_x1 = utility::clamp<int>(index_w + 1, 0, in_dim_w - 1);
@@ -403,9 +409,9 @@ void NEScaleKernel::scale_bilinear_nchw(const Window &window)
             const auto a10 = *(pixel_row_ptr + clamped_x + clamped_y1 * in_stride_w);
             const auto a11 = *(pixel_row_ptr + clamped_x1 + clamped_y1 * in_stride_w);
 
-            *reinterpret_cast<T *>(out.ptr()) = static_cast<T>(scale_helpers::delta_bilinear(a00, a01, a10, a11, dx_val, dy_val));
+            *reinterpret_cast<T *>(dst_i.ptr()) = static_cast<T>(scale_helpers::delta_bilinear(a00, a01, a10, a11, dx_val, dy_val));
         },
-        in, offsets, dx, dy, out);
+        src_i, offsets_i, dx_i, dy_i, dst_i);
     }
     else
     {
@@ -413,11 +419,12 @@ void NEScaleKernel::scale_bilinear_nchw(const Window &window)
     }
 }
 
-void NEScaleKernel::scale_area_nchw_u8(const Window &window)
+void CpuScaleKernel::scale_area_nchw_u8(const ITensor *src, ITensor *dst, const ITensor *dx, const ITensor *dy, const ITensor *offsets, const Window &window)
 {
+    ARM_COMPUTE_UNUSED(dx, dy, offsets);
     using namespace scale_helpers;
 
-    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(_input, 1, DataType::U8);
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src, 1, DataType::U8);
 
     // Don't increment in width/height/channels for the input tensor
     // A pointer to the start of this plane is needed as base for the precomputed offsets
@@ -426,18 +433,18 @@ void NEScaleKernel::scale_area_nchw_u8(const Window &window)
     win_in.set(Window::DimY, Window::Dimension(0, 0, 0));
     win_in.set(Window::DimZ, Window::Dimension(0, 0, 0));
 
-    Iterator in(_input, win_in);
-    Iterator out(_output, window);
+    Iterator src_i(src, win_in);
+    Iterator dst_i(dst, window);
 
-    const auto   wr        = scale_utils::calculate_resize_ratio(_input->info()->dimension(0), _output->info()->dimension(0), _align_corners);
-    const auto   hr        = scale_utils::calculate_resize_ratio(_input->info()->dimension(1), _output->info()->dimension(1), _align_corners);
-    const auto   w         = _input->info()->dimension(0);
-    const auto   h         = _input->info()->dimension(1);
-    const size_t in_stride = _input->info()->strides_in_bytes()[1];
+    const auto   wr        = scale_utils::calculate_resize_ratio(src->info()->dimension(0), dst->info()->dimension(0), _align_corners);
+    const auto   hr        = scale_utils::calculate_resize_ratio(src->info()->dimension(1), dst->info()->dimension(1), _align_corners);
+    const auto   w         = src->info()->dimension(0);
+    const auto   h         = src->info()->dimension(1);
+    const size_t in_stride = src->info()->strides_in_bytes()[1];
 
     execute_window_loop(window, [&](const Coordinates & id)
     {
-        const auto in_ptr = reinterpret_cast<const uint8_t *>(in.ptr());
+        const auto in_ptr = reinterpret_cast<const uint8_t *>(src_i.ptr());
 
         uint8x8_t tmp0 = vdup_n_u8(0);
         tmp0           = vset_lane_u8(pixel_area_c1u8_clamp(in_ptr, in_stride, w, h, wr, hr, id.x(), id.y()), tmp0, 0);
@@ -459,20 +466,20 @@ void NEScaleKernel::scale_area_nchw_u8(const Window &window)
         tmp1           = vset_lane_u8(pixel_area_c1u8_clamp(in_ptr, in_stride, w, h, wr, hr, id.x() + 14, id.y()), tmp1, 6);
         tmp1           = vset_lane_u8(pixel_area_c1u8_clamp(in_ptr, in_stride, w, h, wr, hr, id.x() + 15, id.y()), tmp1, 7);
 
-        vst1q_u8(out.ptr(), vcombine_u8(tmp0, tmp1));
+        vst1q_u8(dst_i.ptr(), vcombine_u8(tmp0, tmp1));
     },
-    in, out);
+    src_i, dst_i);
 }
 
 template <typename T>
-void NEScaleKernel::scale_bilinear_qasymm(const Window &window)
+void CpuScaleKernel::scale_bilinear_qasymm(const ITensor *src, ITensor *dst, const ITensor *dx, const ITensor *dy, const ITensor *offsets, const Window &window)
 {
     // Get data layout and width/height indices
     const int idx_width  = get_data_layout_dimension_index(_data_layout, DataLayoutDimension::WIDTH);
     const int idx_height = get_data_layout_dimension_index(_data_layout, DataLayoutDimension::HEIGHT);
 
     // Compute the ratio between source height and destination height
-    const auto hr = scale_utils::calculate_resize_ratio(_input->info()->dimension(idx_height), _output->info()->dimension(idx_height), _align_corners);
+    const auto hr = scale_utils::calculate_resize_ratio(src->info()->dimension(idx_height), dst->info()->dimension(idx_height), _align_corners);
     Window     win_off;
     win_off.set(Window::DimX, Window::Dimension(0, 0, 0));
     win_off.set(Window::DimY, Window::Dimension(0, 0, 0));
@@ -483,21 +490,21 @@ void NEScaleKernel::scale_bilinear_qasymm(const Window &window)
     win_in.set(idx_width, Window::Dimension(0, 0, 0));
     win_in.set(idx_height, Window::Dimension(0, 0, 0));
 
-    for(size_t d = Window::DimZ; d < _offsets->info()->num_dimensions(); ++d)
+    for(size_t d = Window::DimZ; d < offsets->info()->num_dimensions(); ++d)
     {
         win_off.set(d, Window::Dimension(0, 0, 0));
     }
 
-    Iterator in(_input, win_in);
-    Iterator out(_output, window);
+    Iterator src_i(src, win_in);
+    Iterator dst_i(dst, window);
 
-    const int32_t in_dim_w = _input->info()->dimension(idx_width);
-    const int32_t in_dim_h = _input->info()->dimension(idx_height);
-    const int32_t stride_w = _input->info()->strides_in_bytes()[idx_width];
-    const int32_t stride_h = _input->info()->strides_in_bytes()[idx_height];
+    const int32_t in_dim_w = src->info()->dimension(idx_width);
+    const int32_t in_dim_h = src->info()->dimension(idx_height);
+    const int32_t stride_w = src->info()->strides_in_bytes()[idx_width];
+    const int32_t stride_h = src->info()->strides_in_bytes()[idx_height];
 
-    const UniformQuantizationInfo iq_info = _input->info()->quantization_info().uniform();
-    const UniformQuantizationInfo oq_info = _output->info()->quantization_info().uniform();
+    const UniformQuantizationInfo iq_info = src->info()->quantization_info().uniform();
+    const UniformQuantizationInfo oq_info = dst->info()->quantization_info().uniform();
 
     if(_border_mode == BorderMode::CONSTANT)
     {
@@ -510,10 +517,10 @@ void NEScaleKernel::scale_bilinear_qasymm(const Window &window)
         execute_window_loop(window, [&](const Coordinates & id)
         {
             const int32_t index_h       = std::floor((id[idx_height] + _sampling_offset) * hr - _sampling_offset);
-            const int32_t index_w       = *(reinterpret_cast<const int32_t *>(_offsets->ptr_to_element(Coordinates(id[idx_width], id[idx_height]))));
-            const auto    dx_val        = *(reinterpret_cast<const float *>(_dx->ptr_to_element(Coordinates(id[idx_width], id[idx_height]))));
-            const auto    dy_val        = *(reinterpret_cast<const float *>(_dy->ptr_to_element(Coordinates(id[idx_width], id[idx_height]))));
-            const auto    pixel_row_ptr = reinterpret_cast<const T *>(in.ptr());
+            const int32_t index_w       = *(reinterpret_cast<const int32_t *>(offsets->ptr_to_element(Coordinates(id[idx_width], id[idx_height]))));
+            const auto    dx_val        = *(reinterpret_cast<const float *>(dx->ptr_to_element(Coordinates(id[idx_width], id[idx_height]))));
+            const auto    dy_val        = *(reinterpret_cast<const float *>(dy->ptr_to_element(Coordinates(id[idx_width], id[idx_height]))));
+            const auto    pixel_row_ptr = reinterpret_cast<const T *>(src_i.ptr());
 
             const auto a00 = (0 <= index_w && index_w < in_dim_w && 0 <= index_h && index_h < in_dim_h) ?
                              (*(pixel_row_ptr + index_w * stride_w + index_h * stride_h)) :
@@ -528,23 +535,23 @@ void NEScaleKernel::scale_bilinear_qasymm(const Window &window)
                              (*(pixel_row_ptr + (index_w + 1) * stride_w + (index_h + 1) * stride_h)) :
                              const_border_value;
 
-            const float inp00                 = Qasymm8QuantizationHelper<T>::dequantize(a00, iq_info);
-            const float inp01                 = Qasymm8QuantizationHelper<T>::dequantize(a01, iq_info);
-            const float inp10                 = Qasymm8QuantizationHelper<T>::dequantize(a10, iq_info);
-            const float inp11                 = Qasymm8QuantizationHelper<T>::dequantize(a11, iq_info);
-            *reinterpret_cast<T *>(out.ptr()) = Qasymm8QuantizationHelper<T>::quantize(scale_helpers::delta_bilinear(inp00, inp01, inp10, inp11, dx_val, dy_val), oq_info);
+            const float inp00                   = Qasymm8QuantizationHelper<T>::dequantize(a00, iq_info);
+            const float inp01                   = Qasymm8QuantizationHelper<T>::dequantize(a01, iq_info);
+            const float inp10                   = Qasymm8QuantizationHelper<T>::dequantize(a10, iq_info);
+            const float inp11                   = Qasymm8QuantizationHelper<T>::dequantize(a11, iq_info);
+            *reinterpret_cast<T *>(dst_i.ptr()) = Qasymm8QuantizationHelper<T>::quantize(scale_helpers::delta_bilinear(inp00, inp01, inp10, inp11, dx_val, dy_val), oq_info);
         },
-        in, out);
+        src_i, dst_i);
     }
     else if(_border_mode == BorderMode::REPLICATE)
     {
         execute_window_loop(window, [&](const Coordinates & id)
         {
             const int     index_h       = std::floor((id[idx_height] + _sampling_offset) * hr - _sampling_offset);
-            const int32_t index_w       = *(reinterpret_cast<const int32_t *>(_offsets->ptr_to_element(Coordinates(id[idx_width], id[idx_height]))));
-            const auto    dx_val        = *(reinterpret_cast<const float *>(_dx->ptr_to_element(Coordinates(id[idx_width], id[idx_height]))));
-            const auto    dy_val        = *(reinterpret_cast<const float *>(_dy->ptr_to_element(Coordinates(id[idx_width], id[idx_height]))));
-            const auto    pixel_row_ptr = reinterpret_cast<const T *>(in.ptr());
+            const int32_t index_w       = *(reinterpret_cast<const int32_t *>(offsets->ptr_to_element(Coordinates(id[idx_width], id[idx_height]))));
+            const auto    dx_val        = *(reinterpret_cast<const float *>(dx->ptr_to_element(Coordinates(id[idx_width], id[idx_height]))));
+            const auto    dy_val        = *(reinterpret_cast<const float *>(dy->ptr_to_element(Coordinates(id[idx_width], id[idx_height]))));
+            const auto    pixel_row_ptr = reinterpret_cast<const T *>(src_i.ptr());
 
             auto clamped_w  = utility::clamp<int>(index_w, 0, in_dim_w - 1);
             auto clamped_w1 = utility::clamp<int>(index_w + 1, 0, in_dim_w - 1);
@@ -556,42 +563,56 @@ void NEScaleKernel::scale_bilinear_qasymm(const Window &window)
             const auto a10 = *(pixel_row_ptr + clamped_w * stride_w + clamped_h1 * stride_h);
             const auto a11 = *(pixel_row_ptr + clamped_w1 * stride_w + clamped_h1 * stride_h);
 
-            const float inp00                 = Qasymm8QuantizationHelper<T>::dequantize(a00, iq_info);
-            const float inp01                 = Qasymm8QuantizationHelper<T>::dequantize(a01, iq_info);
-            const float inp10                 = Qasymm8QuantizationHelper<T>::dequantize(a10, iq_info);
-            const float inp11                 = Qasymm8QuantizationHelper<T>::dequantize(a11, iq_info);
-            *reinterpret_cast<T *>(out.ptr()) = Qasymm8QuantizationHelper<T>::quantize(scale_helpers::delta_bilinear(inp00, inp01, inp10, inp11, dx_val, dy_val), oq_info);
+            const float inp00                   = Qasymm8QuantizationHelper<T>::dequantize(a00, iq_info);
+            const float inp01                   = Qasymm8QuantizationHelper<T>::dequantize(a01, iq_info);
+            const float inp10                   = Qasymm8QuantizationHelper<T>::dequantize(a10, iq_info);
+            const float inp11                   = Qasymm8QuantizationHelper<T>::dequantize(a11, iq_info);
+            *reinterpret_cast<T *>(dst_i.ptr()) = Qasymm8QuantizationHelper<T>::quantize(scale_helpers::delta_bilinear(inp00, inp01, inp10, inp11, dx_val, dy_val), oq_info);
         },
-        in, out);
+        src_i, dst_i);
     }
     else
     {
         ARM_COMPUTE_ERROR("Not implemented");
     }
 }
+#endif // ENABLE_NCHW_KERNELS
 
-Status NEScaleKernel::validate(const ITensorInfo *input, const ITensorInfo *dx, const ITensorInfo *dy,
-                               const ITensorInfo *offsets, ITensorInfo *output, const ScaleKernelInfo &info)
+Status CpuScaleKernel::validate(const ITensorInfo *input, const ITensorInfo *dx, const ITensorInfo *dy,
+                                const ITensorInfo *offsets, ITensorInfo *output, const ScaleKernelInfo &info)
 {
     ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, dx, dy, offsets, output, info));
     return Status{};
 }
 
-void NEScaleKernel::run(const Window &window, const ThreadInfo &info)
+void CpuScaleKernel::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(_func == nullptr && _data_layout == DataLayout::NCHW);
 
+    const auto src     = tensors.get_const_tensor(TensorType::ACL_SRC);
+    auto       dst     = tensors.get_tensor(TensorType::ACL_DST);
+    const auto dx      = tensors.get_const_tensor(TensorType::ACL_INT_0);
+    const auto dy      = tensors.get_const_tensor(TensorType::ACL_INT_1);
+    const auto offsets = tensors.get_const_tensor(TensorType::ACL_INT_2);
+
     if(_data_layout == DataLayout::NCHW)
     {
-        (this->*_func)(window);
+        (this->*_func)(src, dst, dx, dy, offsets, window);
     }
     else
     {
-        const auto *uk = get_implementation(ScaleSelectorData{ _input->info()->data_type() });
-        uk->ukernel(_input, _output, _offsets, _dx, _dy, _policy, _border_mode, _constant_border_value, _sampling_offset, _align_corners, window);
+        const auto *uk = get_implementation(ScaleSelectorData{ src->info()->data_type() });
+        uk->ukernel(src, dst, offsets, dx, dy, _policy, _border_mode, _constant_border_value, _sampling_offset, _align_corners, window);
     }
 }
+
+const char *CpuScaleKernel::name() const
+{
+    return "CpuScaleKernel";
+}
+} // namespace kernels
+} // namespace cpu
 } // namespace arm_compute
diff --git a/src/core/cpu/kernels/CpuScaleKernel.h b/src/core/cpu/kernels/CpuScaleKernel.h
new file mode 100644
index 0000000000..c1de8e0736
--- /dev/null
+++ b/src/core/cpu/kernels/CpuScaleKernel.h
@@ -0,0 +1,111 @@
+/*
+ * 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_SCALEKERNEL_H
+#define ARM_COMPUTE_CPU_SCALEKERNEL_H
+
+#include "arm_compute/core/KernelDescriptors.h"
+#include "src/core/common/Macros.h"
+#include "src/core/cpu/ICpuKernel.h"
+
+namespace arm_compute
+{
+namespace cpu
+{
+namespace kernels
+{
+/** Arm(R) Neon(TM) kernel to perform scaling on a tensor */
+class CpuScaleKernel : public ICpuKernel
+{
+public:
+    /** Default constructor */
+    CpuScaleKernel();
+    ARM_COMPUTE_DISALLOW_COPY_ALLOW_MOVE(CpuScaleKernel);
+    /** Initialise the kernel's inputs, output and interpolation policy
+     *
+     * @note dx, dy and offsets have the same dimensions (width and height) of the output tensor
+     * @note Using @p policy Area only supports data layout NCHW and input data type U8.
+     *
+     * @param[in]  src     Source tensor info. Data types supported: QASYMM8/QASYMM8_SIGNED/U8/S16/F16/F32.
+     * @param[in]  dx      Distance x tensor info. Pixel's distance between the X real coordinate and the smallest X following integer. Data type supported: F32
+     * @param[in]  dy      Distance y tensor info. Pixel's distance between the Y real coordinate and the smallest Y following integer. Data type supported: F32
+     * @param[in]  offsets Offset tensor info. Offset to access the pixel with NEAREST interpolation or the top-left pixel with BILINEAR interpolation in the input tensor. Data type supported: S32.
+     * @param[out] dst     Destination tensor info. Data types supported: Same as @p input. All but the lowest two dimensions must be the same size as in the input tensor, i.e. scaling is only performed within the XY-plane.
+     * @param[in]  info    @ref ScaleKernelInfo to use for configuration
+     */
+    void configure(const ITensorInfo *src, const ITensorInfo *dx, const ITensorInfo *dy, const ITensorInfo *offsets, ITensorInfo *dst,
+                   const ScaleKernelInfo &info);
+    /** Static function to check if given info will lead to a valid configuration of @ref CpuScaleKernel
+     *
+     * @note dx, dy and offsets have the same dimensions (width and height) of the output tensor
+     * @note Using @p policy Area only supports data layout NCHW and input data type U8.
+     *
+     * @param[in] src     Source tensor. Data types supported: QASYMM8/QASYMM8_SIGNED/U8/S16/F16/F32.
+     * @param[in] dx      Distance x tensor info. Pixel's distance between the X real coordinate and the smallest X following integer. Data type supported: F32
+     * @param[in] dy      Distance y tensor info. Pixel's distance between the Y real coordinate and the smallest Y following integer. Data type supported: F32
+     * @param[in] offsets Offset tensor info. Offset to access the pixel with NEAREST interpolation or the top-left pixel with BILINEAR interpolation in the input tensor. Data type supported: S32.
+     * @param[in] dst     Destination tensor info. Data types supported: Same as @p input. All but the lowest two dimensions must be the same size as in the input tensor, i.e. scaling is only performed within the XY-plane.
+     * @param[in] info    @ref ScaleKernelInfo to use for validation
+     */
+    static Status validate(const ITensorInfo *src, const ITensorInfo *dx, const ITensorInfo *dy, const ITensorInfo *offsets, ITensorInfo *dst,
+                           const ScaleKernelInfo &info);
+
+    // Inherited methods overridden:
+    void run_op(ITensorPack &tensors, const Window &window, const ThreadInfo &info) override;
+    const char *name() const override;
+
+private:
+#ifdef ENABLE_NCHW_KERNELS
+    /** function to perform scale using area interpolation on the given window
+     *
+     *  @note Used only in case down-sampling.
+     */
+    void scale_area_nchw_u8(const ITensor *src, ITensor *dst, const ITensor *dx, const ITensor *dy, const ITensor *offsets, const Window &window);
+
+    /** function to perform scale using bilinear interpolation on the given window */
+    template <typename T>
+    void scale_bilinear_nchw(const ITensor *src, ITensor *dst, const ITensor *dx, const ITensor *dy, const ITensor *offsets, const Window &window);
+    /** function to perform scale using bilinear interpolation on the given window */
+    template <typename T>
+    void scale_bilinear_qasymm(const ITensor *src, ITensor *dst, const ITensor *dx, const ITensor *dy, const ITensor *offsets, const Window &window);
+
+    /** function to perform scale using nearest neighbour on the given window */
+    template <typename T>
+    void scale_nearest_nchw(const ITensor *src, ITensor *dst, const ITensor *dx, const ITensor *dy, const ITensor *offsets, const Window &window);
+#endif // ENABLE_NCHW_KERNELS
+
+    /** Scale function to use for the particular function to use */
+    using ScaleFunctionPtr = void (CpuScaleKernel::*)(const ITensor *, ITensor *, const ITensor *, const ITensor *, const ITensor *, const Window &window);
+
+    ScaleFunctionPtr    _func;
+    InterpolationPolicy _policy;
+    BorderMode          _border_mode;
+    PixelValue          _constant_border_value;
+    float               _sampling_offset;
+    bool                _align_corners;
+    DataLayout          _data_layout;
+};
+} // namespace kernels
+} // namespace cpu
+} // namespace arm_compute
+#endif /*ARM_COMPUTE_CPU_SCALEKERNEL_H */
diff --git a/src/core/NEON/kernels/scale/impl/NEON/fp16.cpp b/src/core/cpu/kernels/scale/neon/fp16.cpp
index 0ad66cab1c..0ad66cab1c 100644
--- a/src/core/NEON/kernels/scale/impl/NEON/fp16.cpp
+++ b/src/core/cpu/kernels/scale/neon/fp16.cpp
diff --git a/src/core/NEON/kernels/scale/impl/NEON/integer.cpp b/src/core/cpu/kernels/scale/neon/integer.cpp
index a2359aac94..a2359aac94 100644
--- a/src/core/NEON/kernels/scale/impl/NEON/integer.cpp
+++ b/src/core/cpu/kernels/scale/neon/integer.cpp
diff --git a/src/core/NEON/kernels/scale/impl/NEON/list.h b/src/core/cpu/kernels/scale/neon/list.h
index c91242f5b2..c91242f5b2 100644
--- a/src/core/NEON/kernels/scale/impl/NEON/list.h
+++ b/src/core/cpu/kernels/scale/neon/list.h
diff --git a/src/core/NEON/kernels/scale/impl/NEON/qasymm8.cpp b/src/core/cpu/kernels/scale/neon/qasymm8.cpp
index 536ad2cd17..90302ce889 100644
--- a/src/core/NEON/kernels/scale/impl/NEON/qasymm8.cpp
+++ b/src/core/cpu/kernels/scale/neon/qasymm8.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/scale/impl/NEON/list.h"
+#include "src/core/cpu/kernels/scale/neon/list.h"
 
 namespace arm_compute
 {
@@ -31,13 +31,9 @@ void qasymm8_neon_scale_bilinear(const ITensor *src, ITensor *dst, const ITensor
                                  BorderMode border_mode, PixelValue constant_border_value, float sampling_offset,
                                  bool align_corners, const Window &window)
 {
-    // Get data layout and width/height indices
-    const DataLayout data_layout = src->info()->data_layout();
-    const int        idx_width   = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH);
-    const int        idx_height  = get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT);
-
+    // Data layout is NHWC
     // Compute the ratio between source height and destination height
-    const auto hr = scale_utils::calculate_resize_ratio(src->info()->dimension(idx_height), dst->info()->dimension(idx_height), align_corners);
+    const auto hr = scale_utils::calculate_resize_ratio(src->info()->dimension(2), dst->info()->dimension(2), align_corners);
     Window     win_off;
     win_off.set(Window::DimX, Window::Dimension(0, 0, 0));
     win_off.set(Window::DimY, Window::Dimension(0, 0, 0));
@@ -45,8 +41,8 @@ void qasymm8_neon_scale_bilinear(const ITensor *src, ITensor *dst, const ITensor
     // Don't increment in X and Y direction for the input tensor
     // A pointer to the start of this plane is needed as base for the precomputed offsets
     Window win_in(window);
-    win_in.set(idx_width, Window::Dimension(0, 0, 0));
-    win_in.set(idx_height, Window::Dimension(0, 0, 0));
+    win_in.set(1, Window::Dimension(0, 0, 0));
+    win_in.set(2, Window::Dimension(0, 0, 0));
 
     for(size_t d = Window::DimZ; d < offsets->info()->num_dimensions(); ++d)
     {
@@ -56,10 +52,10 @@ void qasymm8_neon_scale_bilinear(const ITensor *src, ITensor *dst, const ITensor
     Iterator in(src, win_in);
     Iterator out(dst, window);
 
-    const int32_t in_dim_w = src->info()->dimension(idx_width);
-    const int32_t in_dim_h = src->info()->dimension(idx_height);
-    const int32_t stride_w = src->info()->strides_in_bytes()[idx_width];
-    const int32_t stride_h = src->info()->strides_in_bytes()[idx_height];
+    const int32_t in_dim_w = src->info()->dimension(1);
+    const int32_t in_dim_h = src->info()->dimension(2);
+    const int32_t stride_w = src->info()->strides_in_bytes()[1];
+    const int32_t stride_h = src->info()->strides_in_bytes()[2];
 
     const UniformQuantizationInfo iq_info = src->info()->quantization_info().uniform();
     const UniformQuantizationInfo oq_info = dst->info()->quantization_info().uniform();
@@ -69,10 +65,10 @@ void qasymm8_neon_scale_bilinear(const ITensor *src, ITensor *dst, const ITensor
         const uint8_t const_border_value = static_cast<uint8_t>(constant_border_value.get<uint8_t>());
         execute_window_loop(window, [&](const Coordinates & id)
         {
-            const int32_t index_h       = std::floor((id[idx_height] + sampling_offset) * hr - sampling_offset);
-            const int32_t index_w       = *(reinterpret_cast<const int32_t *>(offsets->ptr_to_element(Coordinates(id[idx_width], id[idx_height]))));
-            const auto    dx_val        = *(reinterpret_cast<const float *>(dx->ptr_to_element(Coordinates(id[idx_width], id[idx_height]))));
-            const auto    dy_val        = *(reinterpret_cast<const float *>(dy->ptr_to_element(Coordinates(id[idx_width], id[idx_height]))));
+            const int32_t index_h       = std::floor((id[2] + sampling_offset) * hr - sampling_offset);
+            const int32_t index_w       = *(reinterpret_cast<const int32_t *>(offsets->ptr_to_element(Coordinates(id[1], id[2]))));
+            const auto    dx_val        = *(reinterpret_cast<const float *>(dx->ptr_to_element(Coordinates(id[1], id[2]))));
+            const auto    dy_val        = *(reinterpret_cast<const float *>(dy->ptr_to_element(Coordinates(id[1], id[2]))));
             const auto    pixel_row_ptr = reinterpret_cast<const uint8_t *>(in.ptr());
 
             const auto a00 = (0 <= index_w && index_w < in_dim_w && 0 <= index_h && index_h < in_dim_h) ?
@@ -100,10 +96,10 @@ void qasymm8_neon_scale_bilinear(const ITensor *src, ITensor *dst, const ITensor
     {
         execute_window_loop(window, [&](const Coordinates & id)
         {
-            const int     index_h       = std::floor((id[idx_height] + sampling_offset) * hr - sampling_offset);
-            const int32_t index_w       = *(reinterpret_cast<const int32_t *>(offsets->ptr_to_element(Coordinates(id[idx_width], id[idx_height]))));
-            const auto    dx_val        = *(reinterpret_cast<const float *>(dx->ptr_to_element(Coordinates(id[idx_width], id[idx_height]))));
-            const auto    dy_val        = *(reinterpret_cast<const float *>(dy->ptr_to_element(Coordinates(id[idx_width], id[idx_height]))));
+            const int     index_h       = std::floor((id[2] + sampling_offset) * hr - sampling_offset);
+            const int32_t index_w       = *(reinterpret_cast<const int32_t *>(offsets->ptr_to_element(Coordinates(id[1], id[2]))));
+            const auto    dx_val        = *(reinterpret_cast<const float *>(dx->ptr_to_element(Coordinates(id[1], id[2]))));
+            const auto    dy_val        = *(reinterpret_cast<const float *>(dy->ptr_to_element(Coordinates(id[1], id[2]))));
             const auto    pixel_row_ptr = reinterpret_cast<const uint8_t *>(in.ptr());
 
             auto clamped_w  = utility::clamp<int>(index_w, 0, in_dim_w - 1);
diff --git a/src/core/NEON/kernels/scale/impl/NEON/qasymm8_signed.cpp b/src/core/cpu/kernels/scale/neon/qasymm8_signed.cpp
index 149cdf478f..07d6c6ef03 100644
--- a/src/core/NEON/kernels/scale/impl/NEON/qasymm8_signed.cpp
+++ b/src/core/cpu/kernels/scale/neon/qasymm8_signed.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/scale/impl/NEON/list.h"
+#include "src/core/cpu/kernels/scale/neon/list.h"
 
 namespace arm_compute
 {
@@ -31,13 +31,9 @@ void qasymm8_signed_neon_scale_bilinear(const ITensor *src, ITensor *dst, const
                                         BorderMode border_mode, PixelValue constant_border_value, float sampling_offset,
                                         bool align_corners, const Window &window)
 {
-    // Get data layout and width/height indices
-    const DataLayout data_layout = src->info()->data_layout();
-    const int        idx_width   = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH);
-    const int        idx_height  = get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT);
-
+    // Data layout is NHWC
     // Compute the ratio between source height and destination height
-    const auto hr = scale_utils::calculate_resize_ratio(src->info()->dimension(idx_height), dst->info()->dimension(idx_height), align_corners);
+    const auto hr = scale_utils::calculate_resize_ratio(src->info()->dimension(2), dst->info()->dimension(2), align_corners);
     Window     win_off;
     win_off.set(Window::DimX, Window::Dimension(0, 0, 0));
     win_off.set(Window::DimY, Window::Dimension(0, 0, 0));
@@ -45,8 +41,8 @@ void qasymm8_signed_neon_scale_bilinear(const ITensor *src, ITensor *dst, const
     // Don't increment in X and Y direction for the input tensor
     // A pointer to the start of this plane is needed as base for the precomputed offsets
     Window win_in(window);
-    win_in.set(idx_width, Window::Dimension(0, 0, 0));
-    win_in.set(idx_height, Window::Dimension(0, 0, 0));
+    win_in.set(1, Window::Dimension(0, 0, 0));
+    win_in.set(2, Window::Dimension(0, 0, 0));
 
     for(size_t d = Window::DimZ; d < offsets->info()->num_dimensions(); ++d)
     {
@@ -56,10 +52,10 @@ void qasymm8_signed_neon_scale_bilinear(const ITensor *src, ITensor *dst, const
     Iterator in(src, win_in);
     Iterator out(dst, window);
 
-    const int32_t in_dim_w = src->info()->dimension(idx_width);
-    const int32_t in_dim_h = src->info()->dimension(idx_height);
-    const int32_t stride_w = src->info()->strides_in_bytes()[idx_width];
-    const int32_t stride_h = src->info()->strides_in_bytes()[idx_height];
+    const int32_t in_dim_w = src->info()->dimension(1);
+    const int32_t in_dim_h = src->info()->dimension(2);
+    const int32_t stride_w = src->info()->strides_in_bytes()[1];
+    const int32_t stride_h = src->info()->strides_in_bytes()[2];
 
     const UniformQuantizationInfo iq_info = src->info()->quantization_info().uniform();
     const UniformQuantizationInfo oq_info = dst->info()->quantization_info().uniform();
@@ -69,10 +65,10 @@ void qasymm8_signed_neon_scale_bilinear(const ITensor *src, ITensor *dst, const
         const int8_t const_border_value = static_cast<int8_t>(constant_border_value.get<int8_t>());
         execute_window_loop(window, [&](const Coordinates & id)
         {
-            const int32_t index_h       = std::floor((id[idx_height] + sampling_offset) * hr - sampling_offset);
-            const int32_t index_w       = *(reinterpret_cast<const int32_t *>(offsets->ptr_to_element(Coordinates(id[idx_width], id[idx_height]))));
-            const auto    dx_val        = *(reinterpret_cast<const float *>(dx->ptr_to_element(Coordinates(id[idx_width], id[idx_height]))));
-            const auto    dy_val        = *(reinterpret_cast<const float *>(dy->ptr_to_element(Coordinates(id[idx_width], id[idx_height]))));
+            const int32_t index_h       = std::floor((id[2] + sampling_offset) * hr - sampling_offset);
+            const int32_t index_w       = *(reinterpret_cast<const int32_t *>(offsets->ptr_to_element(Coordinates(id[1], id[2]))));
+            const auto    dx_val        = *(reinterpret_cast<const float *>(dx->ptr_to_element(Coordinates(id[1], id[2]))));
+            const auto    dy_val        = *(reinterpret_cast<const float *>(dy->ptr_to_element(Coordinates(id[1], id[2]))));
             const auto    pixel_row_ptr = reinterpret_cast<const int8_t *>(in.ptr());
 
             const auto a00 = (0 <= index_w && index_w < in_dim_w && 0 <= index_h && index_h < in_dim_h) ?
@@ -100,10 +96,10 @@ void qasymm8_signed_neon_scale_bilinear(const ITensor *src, ITensor *dst, const
     {
         execute_window_loop(window, [&](const Coordinates & id)
         {
-            const int     index_h       = std::floor((id[idx_height] + sampling_offset) * hr - sampling_offset);
-            const int32_t index_w       = *(reinterpret_cast<const int32_t *>(offsets->ptr_to_element(Coordinates(id[idx_width], id[idx_height]))));
-            const auto    dx_val        = *(reinterpret_cast<const float *>(dx->ptr_to_element(Coordinates(id[idx_width], id[idx_height]))));
-            const auto    dy_val        = *(reinterpret_cast<const float *>(dy->ptr_to_element(Coordinates(id[idx_width], id[idx_height]))));
+            const int     index_h       = std::floor((id[2] + sampling_offset) * hr - sampling_offset);
+            const int32_t index_w       = *(reinterpret_cast<const int32_t *>(offsets->ptr_to_element(Coordinates(id[1], id[2]))));
+            const auto    dx_val        = *(reinterpret_cast<const float *>(dx->ptr_to_element(Coordinates(id[1], id[2]))));
+            const auto    dy_val        = *(reinterpret_cast<const float *>(dy->ptr_to_element(Coordinates(id[1], id[2]))));
             const auto    pixel_row_ptr = reinterpret_cast<const int8_t *>(in.ptr());
 
             auto clamped_w  = utility::clamp<int>(index_w, 0, in_dim_w - 1);
diff --git a/src/core/NEON/kernels/scale/impl/SVE/fp16.cpp b/src/core/cpu/kernels/scale/sve/fp16.cpp
index 99f08dbdf9..99f08dbdf9 100644
--- a/src/core/NEON/kernels/scale/impl/SVE/fp16.cpp
+++ b/src/core/cpu/kernels/scale/sve/fp16.cpp
diff --git a/src/core/NEON/kernels/scale/impl/SVE/fp32.cpp b/src/core/cpu/kernels/scale/sve/fp32.cpp
index 94055ae953..94055ae953 100644
--- a/src/core/NEON/kernels/scale/impl/SVE/fp32.cpp
+++ b/src/core/cpu/kernels/scale/sve/fp32.cpp
diff --git a/src/core/NEON/kernels/scale/impl/SVE/integer.cpp b/src/core/cpu/kernels/scale/sve/integer.cpp
index 2a724ece31..2a724ece31 100644
--- a/src/core/NEON/kernels/scale/impl/SVE/integer.cpp
+++ b/src/core/cpu/kernels/scale/sve/integer.cpp
diff --git a/src/core/NEON/kernels/scale/impl/SVE/list.h b/src/core/cpu/kernels/scale/sve/list.h
index b9c3a10a78..b9c3a10a78 100644
--- a/src/core/NEON/kernels/scale/impl/SVE/list.h
+++ b/src/core/cpu/kernels/scale/sve/list.h
diff --git a/src/core/NEON/kernels/scale/impl/SVE/qasymm8.cpp b/src/core/cpu/kernels/scale/sve/qasymm8.cpp
index c475ad615c..c475ad615c 100644
--- a/src/core/NEON/kernels/scale/impl/SVE/qasymm8.cpp
+++ b/src/core/cpu/kernels/scale/sve/qasymm8.cpp
diff --git a/src/core/NEON/kernels/scale/impl/SVE/qasymm8_signed.cpp b/src/core/cpu/kernels/scale/sve/qasymm8_signed.cpp
index b39b75abba..b39b75abba 100644
--- a/src/core/NEON/kernels/scale/impl/SVE/qasymm8_signed.cpp
+++ b/src/core/cpu/kernels/scale/sve/qasymm8_signed.cpp
diff --git a/src/runtime/NEON/functions/NECropResize.cpp b/src/runtime/NEON/functions/NECropResize.cpp
index af85cac7da..1e1070d961 100644
--- a/src/runtime/NEON/functions/NECropResize.cpp
+++ b/src/runtime/NEON/functions/NECropResize.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2019-2020 Arm Limited.
+ * Copyright (c) 2019-2021 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -24,6 +24,7 @@
 #include "arm_compute/runtime/NEON/NEScheduler.h"
 
 #include "arm_compute/runtime/NEON/functions/NECropResize.h"
+#include "arm_compute/runtime/Tensor.h"
 #include "src/core/NEON/kernels/NECropKernel.h"
 
 #include <cstddef>
diff --git a/src/runtime/NEON/functions/NEScale.cpp b/src/runtime/NEON/functions/NEScale.cpp
index f91de32191..0fbad07d0f 100644
--- a/src/runtime/NEON/functions/NEScale.cpp
+++ b/src/runtime/NEON/functions/NEScale.cpp
@@ -23,191 +23,99 @@
  */
 #include "arm_compute/runtime/NEON/functions/NEScale.h"
 
-#include "arm_compute/core/Coordinates.h"
-#include "arm_compute/core/Error.h"
-#include "arm_compute/core/Helpers.h"
-#include "arm_compute/core/ITensor.h"
-#include "arm_compute/core/PixelValue.h"
-#include "arm_compute/core/TensorInfo.h"
-#include "arm_compute/core/Window.h"
-#include "arm_compute/runtime/NEON/NEScheduler.h"
-#include "arm_compute/runtime/TensorAllocator.h"
-#include "src/core/NEON/kernels/NEScaleKernel.h"
-
+#include "arm_compute/core/Validate.h"
+#include "arm_compute/runtime/Tensor.h"
 #include "src/core/utils/ScaleUtils.h"
-
+#include "src/runtime/cpu/operators/CpuScale.h"
 #include "support/Rounding.h"
 
-#include <cmath>
-#include <cstddef>
-#include <utility>
-
 namespace arm_compute
 {
-namespace
-{
-void precompute_dx_dy_offsets(ITensor *dx, ITensor *dy, ITensor *offsets, float wr, float hr, SamplingPolicy sampling_policy, bool align_corners)
+struct NEScale::Impl
 {
-    ARM_COMPUTE_ERROR_ON(nullptr == offsets);
-    ARM_COMPUTE_UNUSED(sampling_policy);
-    float sampling_offset = 0.0f;
-    if(sampling_policy == SamplingPolicy::CENTER)
-    {
-        sampling_offset = 0.5f;
-    }
-
-    Window win;
-    win.set(Window::DimX, Window::Dimension(0, offsets->info()->dimension(0), 1));
-    win.set(Window::DimY, Window::Dimension(0, offsets->info()->dimension(1), 1));
-
-    if(dx != nullptr && dy != nullptr)
-    {
-        // Pre-compute the offset and pixel's distance for BILINEAR interpolation
-        Iterator offsets_it(offsets, win);
-        Iterator dx_it(dx, win);
-        Iterator dy_it(dy, win);
-
-        execute_window_loop(win, [&](const Coordinates & id)
-        {
-            const float in_x  = (id.x() + sampling_offset) * wr - sampling_offset;
-            const float in_y  = (id.y() + sampling_offset) * hr - sampling_offset;
-            const int   in_xi = std::floor(in_x);
-            const int   in_yi = std::floor(in_y);
-
-            *reinterpret_cast<int32_t *>(offsets_it.ptr()) = in_xi;
-            *reinterpret_cast<float *>(dx_it.ptr())        = in_x - in_xi;
-            *reinterpret_cast<float *>(dy_it.ptr())        = in_y - in_yi;
-        },
-        offsets_it, dx_it, dy_it);
-    }
-    else
-    {
-        // Pre-compute the offset for NEAREST interpolation
-        Iterator offsets_it(offsets, win);
-
-        execute_window_loop(win, [&](const Coordinates & id)
-        {
-            const float float_in_xi                        = (id.x() + sampling_offset) * wr;
-            const auto  in_xi                              = static_cast<size_t>(align_corners ? arm_compute::utils::rounding::round_half_away_from_zero(float_in_xi) : std::floor(float_in_xi));
-            *reinterpret_cast<int32_t *>(offsets_it.ptr()) = in_xi;
-        },
-        offsets_it);
-    }
-}
-} // namespace
+    const ITensor                 *src{ nullptr };
+    ITensor                       *dst{ nullptr };
+    Tensor                         dx{ nullptr };      /**< Element's distance between the X real coordinate and the smallest X following integer */
+    Tensor                         dy{ nullptr };      /**< Element's distance between the Y real coordinate and the smallest Y following integer */
+    Tensor                         offsets{ nullptr }; /**< Offset to access the element with NEAREST interpolation or the top-left element with BILINEAR interpolation in the input tensor */
+    std::unique_ptr<cpu::CpuScale> op{ nullptr };
+};
 
 NEScale::NEScale()
-    : _offsets(), _dx(), _dy()
+    : _impl(std::make_unique<Impl>())
 {
 }
+NEScale::~NEScale() = default;
 
 void NEScale::configure(ITensor *input, ITensor *output, const ScaleKernelInfo &info)
 {
-    ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
-    ARM_COMPUTE_ERROR_THROW_ON(NEScale::validate(input->info(), output->info(), info));
-
-    const bool is_align_corners_used = info.align_corners && arm_compute::scale_utils::is_align_corners_allowed_sampling_policy(info.sampling_policy);
+    _impl->src = input;
+    _impl->dst = output;
+    _impl->op  = std::make_unique<cpu::CpuScale>();
+    _impl->op->configure(input->info(), output->info(), info);
 
+    // Configure for size of allocation of internal tensors
     // Get data layout and width/height indices
     const DataLayout data_layout = info.data_layout == DataLayout::UNKNOWN ? input->info()->data_layout() : info.data_layout;
     const int        idx_width   = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH);
     const int        idx_height  = get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT);
 
+    // Compute the ratio between source width/height and destination width/height
+    const bool is_align_corners_used = info.align_corners && arm_compute::scale_utils::is_align_corners_allowed_sampling_policy(info.sampling_policy);
+    const auto wr                    = arm_compute::scale_utils::calculate_resize_ratio(input->info()->dimension(idx_width), output->info()->dimension(idx_width), is_align_corners_used);
+    const auto hr                    = arm_compute::scale_utils::calculate_resize_ratio(input->info()->dimension(idx_height), output->info()->dimension(idx_height), is_align_corners_used);
+
+    // Area interpolation behaves as Nearest Neighbour in case of up-sampling
+    InterpolationPolicy policy_to_use = (info.interpolation_policy == InterpolationPolicy::AREA && wr <= 1.f && hr <= 1.f) ? InterpolationPolicy::NEAREST_NEIGHBOR : info.interpolation_policy;
+
     // Get the tensor shape
     TensorShape shape(output->info()->dimension(idx_width));
     shape.set(1, output->info()->dimension(idx_height), false);
 
-    // Compute the ratio between source width/height and destination width/height
-    const auto wr = arm_compute::scale_utils::calculate_resize_ratio(input->info()->dimension(idx_width), output->info()->dimension(idx_width), is_align_corners_used);
-    const auto hr = arm_compute::scale_utils::calculate_resize_ratio(input->info()->dimension(idx_height), output->info()->dimension(idx_height), is_align_corners_used);
+    const TensorInfo tensor_info_dxdy(shape, Format::F32);
+    const TensorInfo tensor_info_offsets(shape, Format::S32);
 
-    // Area interpolation behaves as Nearest Neighbour in case of up-sampling
-    const auto policy_to_use = (info.interpolation_policy == InterpolationPolicy::AREA && wr <= 1.f && hr <= 1.f) ? InterpolationPolicy::NEAREST_NEIGHBOR : info.interpolation_policy;
-
-    auto scale_kernel = std::make_unique<NEScaleKernel>();
+    _impl->dx.allocator()->init(tensor_info_dxdy);
+    _impl->dy.allocator()->init(tensor_info_dxdy);
+    _impl->offsets.allocator()->init(tensor_info_offsets);
     switch(policy_to_use)
     {
         case InterpolationPolicy::NEAREST_NEIGHBOR:
         {
-            TensorInfo tensor_info_offsets(shape, Format::S32);
-            _offsets.allocator()->init(tensor_info_offsets);
-
-            scale_kernel->configure(input, nullptr, nullptr, &_offsets, output, info);
-
             // Allocate once the configure methods have been called
-            _offsets.allocator()->allocate();
-
-            // Pre-compute offsets for nearest interpolation
-            precompute_dx_dy_offsets(nullptr, nullptr, &_offsets, wr, hr, info.sampling_policy, is_align_corners_used);
+            _impl->offsets.allocator()->allocate();
             break;
         }
         case InterpolationPolicy::BILINEAR:
         {
-            TensorInfo tensor_info_offsets(shape, Format::S32);
-            TensorInfo tensor_info_dxdy(shape, Format::F32);
-
-            _offsets.allocator()->init(tensor_info_offsets);
-            _dx.allocator()->init(tensor_info_dxdy);
-            _dy.allocator()->init(tensor_info_dxdy);
-
-            scale_kernel->configure(input, &_dx, &_dy, &_offsets, output, info);
-
             // Allocate once the configure methods have been called
-            _offsets.allocator()->allocate();
-            _dx.allocator()->allocate();
-            _dy.allocator()->allocate();
-
-            // Pre-compute dx, dy and offsets for bilinear interpolation
-            precompute_dx_dy_offsets(&_dx, &_dy, &_offsets, wr, hr, info.sampling_policy, is_align_corners_used);
+            _impl->dx.allocator()->allocate();
+            _impl->dy.allocator()->allocate();
+            _impl->offsets.allocator()->allocate();
             break;
         }
         case InterpolationPolicy::AREA:
         {
-            scale_kernel->configure(input, nullptr, nullptr, nullptr, output, info);
             break;
         }
         default:
             ARM_COMPUTE_ERROR("Unsupported interpolation mode");
     }
-    _kernel = std::move(scale_kernel);
 }
 
 Status NEScale::validate(const ITensorInfo *input, const ITensorInfo *output, const ScaleKernelInfo &info)
 {
-    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output);
-    ARM_COMPUTE_RETURN_ERROR_ON(info.sampling_policy != SamplingPolicy::CENTER && info.sampling_policy != SamplingPolicy::TOP_LEFT);
-
-    ITensorInfo *offsets = nullptr;
-    ITensorInfo *dx      = nullptr;
-    ITensorInfo *dy      = nullptr;
-
-    // Get data layout and width/height indices
-    const DataLayout data_layout = info.data_layout == DataLayout::UNKNOWN ? input->data_layout() : info.data_layout;
-    const int        idx_width   = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH);
-    const int        idx_height  = get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT);
-
-    // Get the tensor shape of auxilary buffers
-    const TensorShape shape(output->dimension(idx_width), output->dimension(idx_height));
-
-    TensorInfo tensor_info_offsets(shape, Format::S32);
-    TensorInfo tensor_info_dx(shape, Format::F32);
-    TensorInfo tensor_info_dy(shape, Format::F32);
-
-    switch(info.interpolation_policy)
-    {
-        case InterpolationPolicy::NEAREST_NEIGHBOR:
-            offsets = &tensor_info_offsets;
-            break;
-        case InterpolationPolicy::BILINEAR:
-            offsets = &tensor_info_offsets;
-            dx      = &tensor_info_dx;
-            dy      = &tensor_info_dy;
-            break;
-        default:
-            break;
-    }
+    return cpu::CpuScale::validate(input, output, info);
+}
 
-    ARM_COMPUTE_RETURN_ON_ERROR(NEScaleKernel::validate(input->clone().get(), dx, dy, offsets, output->clone().get(), info));
-    return Status{};
+void NEScale::run()
+{
+    ITensorPack pack;
+    pack.add_tensor(TensorType::ACL_SRC, _impl->src);
+    pack.add_tensor(TensorType::ACL_DST, _impl->dst);
+    pack.add_tensor(TensorType::ACL_INT_0, &_impl->dx);
+    pack.add_tensor(TensorType::ACL_INT_1, &_impl->dy);
+    pack.add_tensor(TensorType::ACL_INT_2, &_impl->offsets);
+    _impl->op->run(pack);
 }
 } // namespace arm_compute
diff --git a/src/runtime/cpu/operators/CpuScale.cpp b/src/runtime/cpu/operators/CpuScale.cpp
new file mode 100644
index 0000000000..681a15e26c
--- /dev/null
+++ b/src/runtime/cpu/operators/CpuScale.cpp
@@ -0,0 +1,254 @@
+/*
+ * 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/CpuScale.h"
+
+#include "arm_compute/core/Helpers.h"
+#include "arm_compute/core/TensorInfo.h"
+#include "arm_compute/core/Validate.h"
+#include "arm_compute/runtime/NEON/NEScheduler.h"
+#include "src/core/cpu/kernels/CpuScaleKernel.h"
+#include "src/core/utils/ScaleUtils.h"
+#include "support/Rounding.h"
+
+namespace arm_compute
+{
+namespace cpu
+{
+namespace
+{
+void precompute_dx_dy_offsets(ITensor *dx, ITensor *dy, ITensor *offsets, float wr, float hr, SamplingPolicy sampling_policy, bool align_corners)
+{
+    ARM_COMPUTE_ERROR_ON(offsets == nullptr);
+    float sampling_offset = 0.0f;
+    if(sampling_policy == SamplingPolicy::CENTER)
+    {
+        sampling_offset = 0.5f;
+    }
+
+    Window win;
+    win.set(Window::DimX, Window::Dimension(0, offsets->info()->dimension(0), 1));
+    win.set(Window::DimY, Window::Dimension(0, offsets->info()->dimension(1), 1));
+
+    if(dx != nullptr && dy != nullptr)
+    {
+        // Pre-compute the offset and pixel's distance for BILINEAR interpolation
+        Iterator offsets_it(offsets, win);
+        Iterator dx_it(dx, win);
+        Iterator dy_it(dy, win);
+
+        execute_window_loop(win, [&](const Coordinates & id)
+        {
+            const float in_x  = (id.x() + sampling_offset) * wr - sampling_offset;
+            const float in_y  = (id.y() + sampling_offset) * hr - sampling_offset;
+            const int   in_xi = std::floor(in_x);
+            const int   in_yi = std::floor(in_y);
+
+            *reinterpret_cast<int32_t *>(offsets_it.ptr()) = in_xi;
+            *reinterpret_cast<float *>(dx_it.ptr())        = in_x - in_xi;
+            *reinterpret_cast<float *>(dy_it.ptr())        = in_y - in_yi;
+        },
+        offsets_it, dx_it, dy_it);
+    }
+    else
+    {
+        // Pre-compute the offset for NEAREST interpolation
+        Iterator offsets_it(offsets, win);
+
+        execute_window_loop(win, [&](const Coordinates & id)
+        {
+            const float float_in_xi                        = (id.x() + sampling_offset) * wr;
+            const auto  in_xi                              = static_cast<size_t>(align_corners ? arm_compute::utils::rounding::round_half_away_from_zero(float_in_xi) : std::floor(float_in_xi));
+            *reinterpret_cast<int32_t *>(offsets_it.ptr()) = in_xi;
+        },
+        offsets_it);
+    }
+}
+} // namespace
+
+CpuScale::CpuScale()
+    : _scale_info(InterpolationPolicy::NEAREST_NEIGHBOR, BorderMode::UNDEFINED), _data_layout(DataLayout::UNKNOWN), _is_prepared(false)
+{
+}
+
+void CpuScale::configure(ITensorInfo *src, ITensorInfo *dst, const ScaleKernelInfo &info)
+{
+    ARM_COMPUTE_ERROR_ON_NULLPTR(src, dst);
+    ARM_COMPUTE_ERROR_THROW_ON(CpuScale::validate(src, dst, info));
+
+    _scale_info = info;
+
+    // Get data layout and width/height indices
+    _data_layout = _scale_info.data_layout == DataLayout::UNKNOWN ? src->data_layout() : _scale_info.data_layout;
+    const int        idx_width   = get_data_layout_dimension_index(_data_layout, DataLayoutDimension::WIDTH);
+    const int        idx_height  = get_data_layout_dimension_index(_data_layout, DataLayoutDimension::HEIGHT);
+
+    // Compute the ratio between source width/height and destination width/height
+    const bool is_align_corners_used = _scale_info.align_corners && arm_compute::scale_utils::is_align_corners_allowed_sampling_policy(_scale_info.sampling_policy);
+    const auto wr                    = arm_compute::scale_utils::calculate_resize_ratio(src->dimension(idx_width), dst->dimension(idx_width), is_align_corners_used);
+    const auto hr                    = arm_compute::scale_utils::calculate_resize_ratio(src->dimension(idx_height), dst->dimension(idx_height), is_align_corners_used);
+
+    // Area interpolation behaves as Nearest Neighbour in case of up-sampling
+    InterpolationPolicy policy_to_use = (_scale_info.interpolation_policy == InterpolationPolicy::AREA && wr <= 1.f
+                                         && hr <= 1.f) ?
+                                        InterpolationPolicy::NEAREST_NEIGHBOR :
+                                        _scale_info.interpolation_policy;
+
+    // Get the tensor shape
+    TensorShape shape(dst->dimension(idx_width));
+    shape.set(1, dst->dimension(idx_height), false);
+
+    TensorInfo tensor_info_offsets(shape, Format::S32);
+    TensorInfo tensor_info_dxdy(shape, Format::F32);
+
+    auto dx           = std::make_unique<TensorInfo>(tensor_info_dxdy);
+    auto dy           = std::make_unique<TensorInfo>(tensor_info_dxdy);
+    auto offsets      = std::make_unique<TensorInfo>(tensor_info_offsets);
+    auto scale_kernel = std::make_unique<kernels::CpuScaleKernel>();
+    switch(policy_to_use)
+    {
+        case InterpolationPolicy::NEAREST_NEIGHBOR:
+        {
+            scale_kernel->configure(src, nullptr, nullptr, offsets.get(), dst, info);
+            break;
+        }
+        case InterpolationPolicy::BILINEAR:
+        {
+            scale_kernel->configure(src, dx.get(), dy.get(), offsets.get(), dst, info);
+            break;
+        }
+        case InterpolationPolicy::AREA:
+        {
+            scale_kernel->configure(src, nullptr, nullptr, nullptr, dst, info);
+            break;
+        }
+        default:
+            ARM_COMPUTE_ERROR("Unsupported interpolation mode");
+    }
+    _kernel = std::move(scale_kernel);
+}
+
+Status CpuScale::validate(const ITensorInfo *src, const ITensorInfo *dst, const ScaleKernelInfo &info)
+{
+    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src, dst);
+    ARM_COMPUTE_RETURN_ERROR_ON(info.sampling_policy != SamplingPolicy::CENTER && info.sampling_policy != SamplingPolicy::TOP_LEFT);
+
+    ITensorInfo *offsets = nullptr;
+    ITensorInfo *dx      = nullptr;
+    ITensorInfo *dy      = nullptr;
+
+    // Get data layout and width/height indices
+    const DataLayout data_layout = info.data_layout == DataLayout::UNKNOWN ? src->data_layout() : info.data_layout;
+    const int        idx_width   = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH);
+    const int        idx_height  = get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT);
+
+    // Compute the ratio between source width/height and destination width/height
+    const bool is_align_corners_used = info.align_corners && arm_compute::scale_utils::is_align_corners_allowed_sampling_policy(info.sampling_policy);
+    const auto wr                    = arm_compute::scale_utils::calculate_resize_ratio(src->dimension(idx_width), dst->dimension(idx_width), is_align_corners_used);
+    const auto hr                    = arm_compute::scale_utils::calculate_resize_ratio(src->dimension(idx_height), dst->dimension(idx_height), is_align_corners_used);
+
+    // Area interpolation behaves as Nearest Neighbour in case of up-sampling
+    InterpolationPolicy policy_to_use = (info.interpolation_policy == InterpolationPolicy::AREA && wr <= 1.f && hr <= 1.f) ? InterpolationPolicy::NEAREST_NEIGHBOR : info.interpolation_policy;
+
+    // Get the tensor shape of auxilary buffers
+    const TensorShape shape(dst->dimension(idx_width), dst->dimension(idx_height));
+    TensorInfo        tensor_info_offsets(shape, Format::S32);
+    TensorInfo        tensor_info_dx(shape, Format::F32);
+    TensorInfo        tensor_info_dy(shape, Format::F32);
+    switch(policy_to_use)
+    {
+        case InterpolationPolicy::NEAREST_NEIGHBOR:
+            offsets = &tensor_info_offsets;
+            break;
+        case InterpolationPolicy::BILINEAR:
+            offsets = &tensor_info_offsets;
+            dx      = &tensor_info_dx;
+            dy      = &tensor_info_dy;
+            break;
+        default:
+            break;
+    }
+
+    ARM_COMPUTE_RETURN_ON_ERROR(kernels::CpuScaleKernel::validate(src->clone().get(), dx, dy, offsets, dst->clone().get(), info));
+    return Status{};
+}
+
+void CpuScale::prepare(ITensorPack &tensors)
+{
+    if(!_is_prepared)
+    {
+        _is_prepared       = true;
+        const auto src     = tensors.get_const_tensor(TensorType::ACL_SRC);
+        auto       dst     = tensors.get_tensor(TensorType::ACL_DST);
+        auto       dx      = tensors.get_tensor(TensorType::ACL_INT_0);
+        auto       dy      = tensors.get_tensor(TensorType::ACL_INT_1);
+        auto       offsets = tensors.get_tensor(TensorType::ACL_INT_2);
+
+        // Get data layout and width/height indices
+        const int        idx_width   = get_data_layout_dimension_index(_data_layout, DataLayoutDimension::WIDTH);
+        const int        idx_height  = get_data_layout_dimension_index(_data_layout, DataLayoutDimension::HEIGHT);
+
+        // Compute the ratio between source width/height and destination width/height
+        const bool is_align_corners_used = _scale_info.align_corners && arm_compute::scale_utils::is_align_corners_allowed_sampling_policy(_scale_info.sampling_policy);
+        const auto wr                    = arm_compute::scale_utils::calculate_resize_ratio(src->info()->dimension(idx_width), dst->info()->dimension(idx_width), is_align_corners_used);
+        const auto hr                    = arm_compute::scale_utils::calculate_resize_ratio(src->info()->dimension(idx_height), dst->info()->dimension(idx_height), is_align_corners_used);
+
+        // Area interpolation behaves as Nearest Neighbour in case of up-sampling
+        InterpolationPolicy policy_to_use = (_scale_info.interpolation_policy == InterpolationPolicy::AREA && wr <= 1.f
+                                             && hr <= 1.f) ?
+                                            InterpolationPolicy::NEAREST_NEIGHBOR :
+                                            _scale_info.interpolation_policy;
+        const SamplingPolicy sampling_policy = _scale_info.sampling_policy;
+
+        switch(policy_to_use)
+        {
+            case InterpolationPolicy::NEAREST_NEIGHBOR:
+            {
+                // Pre-compute offsets for nearest interpolation
+                precompute_dx_dy_offsets(nullptr, nullptr, offsets, wr, hr, sampling_policy, is_align_corners_used);
+                break;
+            }
+            case InterpolationPolicy::BILINEAR:
+            {
+                // Pre-compute dx, dy and offsets for bilinear interpolation
+                precompute_dx_dy_offsets(dx, dy, offsets, wr, hr, sampling_policy, is_align_corners_used);
+                break;
+            }
+            case InterpolationPolicy::AREA:
+            {
+                break;
+            }
+            default:
+                ARM_COMPUTE_ERROR("Unsupported interpolation mode");
+        }
+    }
+}
+
+void CpuScale::run(ITensorPack &tensors)
+{
+    ARM_COMPUTE_ERROR_ON_MSG(tensors.empty(), "No inputs provided");
+    prepare(tensors);
+    NEScheduler::get().schedule_op(_kernel.get(), Window::DimY, _kernel->window(), tensors);
+}
+} // namespace cpu
+} // namespace arm_compute
diff --git a/src/runtime/cpu/operators/CpuScale.h b/src/runtime/cpu/operators/CpuScale.h
new file mode 100644
index 0000000000..90248a8d59
--- /dev/null
+++ b/src/runtime/cpu/operators/CpuScale.h
@@ -0,0 +1,73 @@
+/*
+ * 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_SCALE_H
+#define ARM_COMPUTE_CPU_SCALE_H
+
+#include "arm_compute/core/ITensorInfo.h"
+#include "arm_compute/core/KernelDescriptors.h"
+#include "arm_compute/core/experimental/Types.h"
+#include "src/core/cpu/ICpuKernel.h"
+#include "src/runtime/cpu/ICpuOperator.h"
+
+#include <memory>
+
+namespace arm_compute
+{
+namespace cpu
+{
+/** Basic function to compute Scale */
+class CpuScale : public ICpuOperator
+{
+public:
+    /** Default Constructor */
+    CpuScale();
+    /** Initialize the function's source, destination, interpolation type and border_mode.
+     *
+     * @param[in, out] src  Source tensor info. Data type supported: QASYMM8/QASYMM8_SIGNED/U8/S16/F16/F32. (Written to only for @p border_mode != UNDEFINED)
+     * @param[out]     dst  Destination tensor info. Data type supported: Same as @p src. All but the lowest two dimensions must be the same size as in the input tensor, i.e. scaling is only performed within the XY-plane.
+     * @param[in]      info @ref ScaleKernelInfo to be used for configuration
+     */
+    void configure(ITensorInfo *src, ITensorInfo *dst, const ScaleKernelInfo &info);
+    /** Static function to check if given info will lead to a valid configuration of @ref NEScale
+     *
+     * @param[in] src  Source tensor info. Data type supported: QASYMM8/QASYMM8_SIGNED/U8/S16/F16/F32. (Written to only for @p border_mode != UNDEFINED)
+     * @param[in] dst  Destination tensor info. Data type supported: Same as @p src. All but the lowest two dimensions must be the same size as in the input tensor, i.e. scaling is only performed within the XY-plane.
+     * @param[in] info @ref ScaleKernelInfo to be used for validation
+     *
+     * @return a status
+     */
+    static Status validate(const ITensorInfo *src, const ITensorInfo *dst, const ScaleKernelInfo &info);
+
+    // Inherited methods overridden:
+    void prepare(ITensorPack &tensors) override;
+    void run(ITensorPack &tensors) override;
+
+private:
+    ScaleKernelInfo _scale_info;
+    DataLayout      _data_layout;
+    bool            _is_prepared;
+};
+} // namespace cpu
+} // namespace arm_compute
+#endif /*ARM_COMPUTE_CPU_SCALE_H */
diff --git a/tests/validation/NEON/Scale.cpp b/tests/validation/NEON/Scale.cpp
index bb1ab936d1..eab241cb88 100644
--- a/tests/validation/NEON/Scale.cpp
+++ b/tests/validation/NEON/Scale.cpp
@@ -50,7 +50,7 @@ using datasets::ScaleSamplingPolicySet;
 using datasets::ScaleAlignCornersSamplingPolicySet;
 
 /** We consider vector size in byte 64 since the maximum size of
- * a vector used by @ref NEScaleKernel is currently 64-byte (float32x4x4).
+ * a vector used by the kernel is currently 64-byte (float32x4x4).
  * There are possibility to reduce test time further by using
  * smaller vector sizes for different data types where applicable.
  */
@@ -94,9 +94,8 @@ TEST_SUITE(Validate)
 
 /** Validate test suite is to test ARM_COMPUTE_RETURN_ON_* macros
  * we use to check the validity of given arguments in @ref NEScale
- * and subsequent call to @ref NEScaleKernel.
  * Since this is using validate() of @ref NEScale, which pre-adjust
- * arguments for @ref NEScaleKernel, the following conditions in
+ * arguments for the kernel, the following conditions in
  * the kernel are not currently tested.
  * - The same input and output
  * - Data type of offset, dx and dy