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>

4 files changed, 375 insertions, 139 deletions

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 */