Add SVE support and decouple data type for NEScaleKernel

- Decouple data type for NEON NHWC implementation, supported data types are: fp32, fp16, u8, s16, qasymm8, qasymm8_signed.

- Add SVE support for NHWC and all six data types showed above.

Resolves: COMPMID-3873

Change-Id: I097de119f4667b28b025a78cadf7185afa5f15f0
Signed-off-by: Sheri Zhang <sheri.zhang@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/4766
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Pablo Marquez Tello <pablo.tello@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>

16 files changed, 2355 insertions, 184 deletions

diff --git a/Android.bp b/Android.bp
index 580f0db1cb..2675068e91 100644
--- a/Android.bp
+++ b/Android.bp
@@ -405,6 +405,16 @@ 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/fp32.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",
diff --git a/src/core/NEON/kernels/NEScaleKernel.cpp b/src/core/NEON/kernels/NEScaleKernel.cpp
index 5a6d49bf07..1fbd5fb60a 100644
--- a/src/core/NEON/kernels/NEScaleKernel.cpp
+++ b/src/core/NEON/kernels/NEScaleKernel.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020 Arm Limited.
+ * Copyright (c) 2016-2021 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -28,13 +28,14 @@
 #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/list.h"
 #include "src/core/NEON/wrapper/wrapper.h"
+#include "src/core/common/Registrars.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>
 
@@ -42,23 +43,113 @@ namespace arm_compute
 {
 namespace
 {
-inline float compute_bilinear(float a00, float a01, float a10, float a11, float dx_val, float dy_val)
+struct ScaleSelectorData
+{
+    DataType dt;
+};
+using ScaleSelectorPtr = std::add_pointer<bool(const ScaleSelectorData &data)>::type;
+using ScaleKernelPtr   = std::add_pointer<void(const ITensor *, ITensor *, const ITensor *, const ITensor *, const ITensor *,
+                                               InterpolationPolicy, BorderMode, PixelValue, float, bool, const Window &)>::type;
+struct ScaleKernel
+{
+    const char            *name;
+    const ScaleSelectorPtr is_selected;
+    ScaleKernelPtr         ukernel;
+};
+
+static const ScaleKernel available_kernels[] =
 {
-    const float dx1_val = 1.0f - dx_val;
-    const float dy1_val = 1.0f - dy_val;
-
-    const float w1 = dx1_val * dy1_val;
-    const float w2 = dx_val * dy1_val;
-    const float w3 = dx1_val * dy_val;
-    const float w4 = dx_val * dy_val;
-    return a00 * w1 + a01 * w2 + a10 * w3 + a11 * w4;
+#if defined(__ARM_FEATURE_SVE)
+    {
+        "fp16_sve_scale",
+        [](const ScaleSelectorData & data) { return data.dt == DataType::F16; },
+        REGISTER_FP16_NEON(arm_compute::cpu::fp16_sve_scale)
+    },
+    {
+        "f32_sve_scale",
+        [](const ScaleSelectorData & data) { return data.dt == DataType::F32; },
+        REGISTER_FP32_NEON(arm_compute::cpu::fp32_sve_scale)
+    },
+    {
+        "qasymm8_sve_scale",
+        [](const ScaleSelectorData & data) { return data.dt == DataType::QASYMM8; },
+        REGISTER_QASYMM8_NEON(arm_compute::cpu::qasymm8_sve_scale)
+    },
+    {
+        "qasymm8_signed_sve_scale",
+        [](const ScaleSelectorData & data) { return data.dt == DataType::QASYMM8_SIGNED; },
+        REGISTER_QASYMM8_SIGNED_NEON(arm_compute::cpu::qasymm8_signed_sve_scale)
+    },
+    {
+        "u8_sve_scale",
+        [](const ScaleSelectorData & data) { return data.dt == DataType::U8; },
+        REGISTER_INTEGER_NEON(arm_compute::cpu::u8_sve_scale)
+    },
+    {
+        "s16_sve_scale",
+        [](const ScaleSelectorData & data) { return data.dt == DataType::S16; },
+        REGISTER_INTEGER_NEON(arm_compute::cpu::s16_sve_scale)
+    },
+#else /* !defined(__ARM_FEATURE_SVE) */
+#if defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC)
+    {
+        "fp16_neon_scale",
+        [](const ScaleSelectorData & data) { return data.dt == DataType::F16; },
+        REGISTER_FP16_NEON(arm_compute::cpu::fp16_neon_scale)
+    },
+#endif /* !defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) */
+    {
+        "f32_neon_scale",
+        [](const ScaleSelectorData & data) { return data.dt == DataType::F32; },
+        REGISTER_FP32_NEON(arm_compute::cpu::fp32_neon_scale)
+    },
+    {
+        "qasymm8_neon_scale",
+        [](const ScaleSelectorData & data) { return data.dt == DataType::QASYMM8; },
+        REGISTER_QASYMM8_NEON(arm_compute::cpu::qasymm8_neon_scale)
+    },
+    {
+        "qasymm8_signed_neon_scale",
+        [](const ScaleSelectorData & data) { return data.dt == DataType::QASYMM8_SIGNED; },
+        REGISTER_QASYMM8_SIGNED_NEON(arm_compute::cpu::qasymm8_signed_neon_scale)
+    },
+    {
+        "u8_neon_scale",
+        [](const ScaleSelectorData & data) { return data.dt == DataType::U8; },
+        REGISTER_INTEGER_NEON(arm_compute::cpu::u8_neon_scale)
+    },
+    {
+        "s16_neon_scale",
+        [](const ScaleSelectorData & data) { return data.dt == DataType::S16; },
+        REGISTER_INTEGER_NEON(arm_compute::cpu::s16_neon_scale)
+    },
+#endif /* !defined(__ARM_FEATURE_SVE) */
+};
+
+/** Micro-kernel selector
+ *
+ * @param[in] data Selection data passed to help pick the appropriate micro-kernel
+ *
+ * @return A matching micro-kernel else nullptr
+ */
+const ScaleKernel *get_implementation(const ScaleSelectorData &data)
+{
+    for(const auto &uk : available_kernels)
+    {
+        if(uk.is_selected(data))
+        {
+            return &uk;
+        }
+    }
+    return nullptr;
 }
 
 Status validate_arguments(const ITensorInfo *input, const ITensorInfo *dx, const ITensorInfo *dy,
                           const ITensorInfo *offsets, ITensorInfo *output, const ScaleKernelInfo &info)
 {
-    ARM_COMPUTE_RETURN_ERROR_ON_CPU_F16_UNSUPPORTED(input);
-    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::U8, DataType::S16, DataType::F16, DataType::F32, DataType::QASYMM8, DataType::QASYMM8_SIGNED);
+    const auto *uk = get_implementation(ScaleSelectorData{ input->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);
@@ -141,64 +232,51 @@ void NEScaleKernel::configure(const ITensor *input, const ITensor *dx, const ITe
     const auto hr = scale_utils::calculate_resize_ratio(input->info()->dimension(idx_height), output->info()->dimension(idx_height), _align_corners);
 
     // 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 : _policy;
+    _policy = (_policy == InterpolationPolicy::AREA && wr <= 1.f && hr <= 1.f) ? InterpolationPolicy::NEAREST_NEIGHBOR : _policy;
 
     if(_border_mode == BorderMode::UNDEFINED)
     {
         _border_mode           = BorderMode::CONSTANT;
         _constant_border_value = PixelValue();
     }
-    std::string function_to_call("scale_");
-    function_to_call += string_from_data_type(_input->info()->data_type()) + "_";
-    function_to_call += string_from_data_layout(_input->info()->data_layout()) + "_";
-    function_to_call += string_from_interpolation_policy(policy_to_use);
 
-    static std::map<std::string, ScaleFunctionPtr> map_function =
+    // Configure scale function to run
+    if(_input->info()->data_layout() == DataLayout::NCHW)
     {
-        { "scale_U8_NCHW_AREA_CONSTANT", &NEScaleKernel::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_NHWC_BILINEAR", &NEScaleKernel::scale_bilinear_nhwc<uint8_t> },
-        { "scale_U8_NHWC_NEAREST_NEIGHBOUR", &NEScaleKernel::scale_nearest_nhwc<uint8_t> },
+        std::string function_to_call("scale_");
+        function_to_call += string_from_data_type(_input->info()->data_type()) + "_";
+        function_to_call += string_from_data_layout(_input->info()->data_layout()) + "_";
+        function_to_call += string_from_interpolation_policy(_policy);
 
-        { "scale_QASYMM8_NCHW_BILINEAR", &NEScaleKernel::scale_bilinear_qasymm<uint8_t> },
-        { "scale_QASYMM8_NCHW_NEAREST_NEIGHBOUR", &NEScaleKernel::scale_nearest_nchw<uint8_t> },
-
-        { "scale_QASYMM8_NHWC_BILINEAR", &NEScaleKernel::scale_bilinear_qasymm<uint8_t> },
-        { "scale_QASYMM8_NHWC_NEAREST_NEIGHBOUR", &NEScaleKernel::scale_nearest_nhwc<uint8_t> },
+        static std::map<std::string, ScaleFunctionPtr> map_function =
+        {
+            { "scale_U8_NCHW_AREA_CONSTANT", &NEScaleKernel::scale_area_nchw_u8 },
 
-        { "scale_QASYMM8_SIGNED_NCHW_BILINEAR", &NEScaleKernel::scale_bilinear_qasymm<int8_t> },
-        { "scale_QASYMM8_SIGNED_NCHW_NEAREST_NEIGHBOUR", &NEScaleKernel::scale_nearest_nchw<uint8_t> },
+            { "scale_U8_NCHW_BILINEAR", &NEScaleKernel::scale_bilinear_nchw<uint8_t> },
+            { "scale_U8_NCHW_NEAREST_NEIGHBOUR", &NEScaleKernel::scale_nearest_nchw<uint8_t> },
 
-        { "scale_QASYMM8_SIGNED_NHWC_BILINEAR", &NEScaleKernel::scale_bilinear_qasymm<int8_t> },
-        { "scale_QASYMM8_SIGNED_NHWC_NEAREST_NEIGHBOUR", &NEScaleKernel::scale_nearest_nhwc<uint8_t> },
+            { "scale_QASYMM8_NCHW_BILINEAR", &NEScaleKernel::scale_bilinear_qasymm<uint8_t> },
+            { "scale_QASYMM8_NCHW_NEAREST_NEIGHBOUR", &NEScaleKernel::scale_nearest_nchw<uint8_t> },
 
-        { "scale_S16_NCHW_BILINEAR", &NEScaleKernel::scale_bilinear_nchw<int16_t> },
-        { "scale_S16_NCHW_NEAREST_NEIGHBOUR", &NEScaleKernel::scale_nearest_nchw<uint16_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_S16_NHWC_BILINEAR", &NEScaleKernel::scale_bilinear_nhwc<int16_t> },
-        { "scale_S16_NHWC_NEAREST_NEIGHBOUR", &NEScaleKernel::scale_nearest_nhwc<uint16_t> },
+            { "scale_S16_NCHW_BILINEAR", &NEScaleKernel::scale_bilinear_nchw<int16_t> },
+            { "scale_S16_NCHW_NEAREST_NEIGHBOUR", &NEScaleKernel::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<uint16_t> },
-
-        { "scale_F16_NHWC_BILINEAR", &NEScaleKernel::scale_bilinear_nhwc<float16_t> },
-        { "scale_F16_NHWC_NEAREST_NEIGHBOUR", &NEScaleKernel::scale_nearest_nhwc<uint16_t> },
+            { "scale_F16_NCHW_BILINEAR", &NEScaleKernel::scale_bilinear_nchw<float16_t> },
+            { "scale_F16_NCHW_NEAREST_NEIGHBOUR", &NEScaleKernel::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_NHWC_BILINEAR", &NEScaleKernel::scale_bilinear_nhwc<float> },
-        { "scale_F32_NHWC_NEAREST_NEIGHBOUR", &NEScaleKernel::scale_nearest_nhwc<float> },
-    };
-    auto it = map_function.find(function_to_call);
-    if(it != map_function.end())
-    {
-        _func = it->second;
+            { "scale_F32_NCHW_BILINEAR", &NEScaleKernel::scale_bilinear_nchw<float> },
+            { "scale_F32_NCHW_NEAREST_NEIGHBOUR", &NEScaleKernel::scale_nearest_nchw<float> },
+        };
+        auto it = map_function.find(function_to_call);
+        if(it != map_function.end())
+        {
+            _func = it->second;
+        }
     }
 
     // Configure window
@@ -303,7 +381,7 @@ 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>(compute_bilinear(a00, a01, a10, a11, dx_val, dy_val));
+            *reinterpret_cast<T *>(out.ptr()) = static_cast<T>(scale_helpers::delta_bilinear(a00, a01, a10, a11, dx_val, dy_val));
         },
         in, offsets, dx, dy, out);
     }
@@ -327,7 +405,7 @@ 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>(compute_bilinear(a00, a01, a10, a11, dx_val, dy_val));
+            *reinterpret_cast<T *>(out.ptr()) = static_cast<T>(scale_helpers::delta_bilinear(a00, a01, a10, a11, dx_val, dy_val));
         },
         in, offsets, dx, dy, out);
     }
@@ -389,121 +467,6 @@ void NEScaleKernel::scale_area_nchw_u8(const Window &window)
 }
 
 template <typename T>
-void NEScaleKernel::scale_nearest_nhwc(const Window &window)
-{
-    const size_t in_stride_c  = _input->info()->dimension(0) + _input->info()->padding().left + _input->info()->padding().right;
-    const size_t in_stride_w  = _input->info()->dimension(1) + _input->info()->padding().top + _input->info()->padding().bottom;
-    const size_t in_stride_wc = in_stride_w * in_stride_c;
-    const size_t in_dim_h     = _input->info()->dimension(2);
-
-    // Compute the ratio between source height and destination height
-    const auto hr             = scale_utils::calculate_resize_ratio(in_dim_h, _output->info()->dimension(2), _align_corners);
-    const auto window_start_x = static_cast<int32_t>(window.x().start());
-    const auto window_end_x   = static_cast<int32_t>(window.x().end());
-    const int  window_step_x  = 16 / sizeof(T);
-
-    Window win(window);
-    win.set(Window::DimX, Window::Dimension(0, 1, 1));
-    Iterator out(_output, win);
-
-    const uint8_t     *in_ptr_start        = _input->buffer() + _input->info()->offset_first_element_in_bytes();
-    const unsigned int in_stride_bytes_hwc = _input->info()->strides_in_bytes()[3];
-
-    execute_window_loop(win, [&](const Coordinates & id)
-    {
-        const int32_t offset     = *reinterpret_cast<const int32_t *>(_offsets->ptr_to_element(Coordinates(id.y(), id.z()))) * in_stride_c;
-        const auto    in_hi      = static_cast<int>(_align_corners ? utils::rounding::round_half_away_from_zero((id.z() + _sampling_offset) * hr) : std::floor((id.z() + _sampling_offset) * hr));
-        const int     offset_row = in_hi * in_stride_wc;
-        int32_t       x          = window_start_x;
-        const T      *in_ptr     = reinterpret_cast<const T *>(in_ptr_start + in_stride_bytes_hwc * id[3]);
-
-        for(; x <= window_end_x - window_step_x; x += window_step_x)
-        {
-            wrapper::vstore(reinterpret_cast<T *>(out.ptr()) + x,
-                            wrapper::vloadq(in_ptr + offset + offset_row + x));
-        }
-        for(; x < window_end_x; ++x)
-        {
-            *(reinterpret_cast<T *>(out.ptr()) + x) = *(in_ptr + offset + offset_row + x);
-        }
-    },
-    out);
-}
-
-template <typename T>
-void NEScaleKernel::scale_bilinear_nhwc(const Window &window)
-{
-    // Compute the ratio between source height and destination height
-    const auto hr = scale_utils::calculate_resize_ratio(_input->info()->dimension(2), _output->info()->dimension(2), _align_corners);
-
-    Iterator  out(_output, window);
-    const int in_stride_c  = _input->info()->dimension(0) + _input->info()->padding().left + _input->info()->padding().right;
-    const int in_dim_w     = _input->info()->dimension(1);
-    const int in_dim_h     = _input->info()->dimension(2);
-    const int in_stride_wc = in_stride_c * (in_dim_w + _input->info()->padding().top + _input->info()->padding().bottom);
-
-    // Don't increment in Y and Z 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(Window::DimY, Window::Dimension(0, 0, 0));
-    win_in.set(Window::DimZ, Window::Dimension(0, 0, 0));
-    Iterator in(_input, win_in);
-
-    if(_border_mode == BorderMode::CONSTANT)
-    {
-#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
-        using ConstType = typename std::conditional<std::is_same<T, float16_t>::value, half, T>::type;
-#else  /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */
-        using ConstType = T;
-#endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */
-        const T const_border_value = static_cast<T>(_constant_border_value.get<ConstType>());
-        execute_window_loop(window, [&](const Coordinates & id)
-        {
-            const auto    offset = *reinterpret_cast<const int32_t *>(_offsets->ptr_to_element(Coordinates(id.y(), id.z())));
-            const auto    dx_val = *reinterpret_cast<const float *>(_dx->ptr_to_element(Coordinates(id.y(), id.z())));
-            const auto    dy_val = *reinterpret_cast<const float *>(_dy->ptr_to_element(Coordinates(id.y(), id.z())));
-            const int32_t in_hi  = std::floor((id.z() + _sampling_offset) * hr - _sampling_offset);
-            const T      *in_ptr = reinterpret_cast<const T *>(in.ptr()) + offset * in_stride_c + in_hi * in_stride_wc;
-
-            const auto a00 = (0 <= offset && offset < in_dim_w && 0 <= in_hi && in_hi < in_dim_h) ? *in_ptr : const_border_value;
-            const auto a01 = (-1 <= offset && offset < in_dim_w - 1 && 0 <= in_hi && in_hi < in_dim_h) ? *(in_ptr + in_stride_c) : const_border_value;
-            const auto a10 = (0 <= offset && offset < in_dim_w && -1 <= in_hi && in_hi < in_dim_h - 1) ? *(in_ptr + in_stride_wc) : const_border_value;
-            const auto a11 = (-1 <= offset && offset < in_dim_w - 1 && -1 <= in_hi && in_hi < in_dim_h - 1) ? *(in_ptr + in_stride_c + in_stride_wc) : const_border_value;
-
-            *reinterpret_cast<T *>(out.ptr()) = static_cast<T>(compute_bilinear(a00, a01, a10, a11, dx_val, dy_val));
-        },
-        in, out);
-    }
-    else if(_border_mode == BorderMode::REPLICATE)
-    {
-        execute_window_loop(window, [&](const Coordinates & id)
-        {
-            const auto offset = *reinterpret_cast<const int32_t *>(_offsets->ptr_to_element(Coordinates(id.y(), id.z())));
-            const auto dx_val = *reinterpret_cast<const float *>(_dx->ptr_to_element(Coordinates(id.y(), id.z())));
-            const auto dy_val = *reinterpret_cast<const float *>(_dy->ptr_to_element(Coordinates(id.y(), id.z())));
-            const int  in_hi  = std::floor((id.z() + _sampling_offset) * hr - _sampling_offset);
-
-            auto clamped_w  = utility::clamp<int>(offset, 0, in_dim_w - 1);
-            auto clamped_w1 = utility::clamp<int>(offset + 1, 0, in_dim_w - 1);
-            auto clamped_h  = utility::clamp<int>(in_hi, 0, in_dim_h - 1);
-            auto clamped_h1 = utility::clamp<int>(in_hi + 1, 0, in_dim_h - 1);
-
-            const auto a00 = *(reinterpret_cast<const T *>(in.ptr()) + clamped_w * in_stride_c + clamped_h * in_stride_wc);
-            const auto a01 = *(reinterpret_cast<const T *>(in.ptr()) + clamped_w1 * in_stride_c + clamped_h * in_stride_wc);
-            const auto a10 = *(reinterpret_cast<const T *>(in.ptr()) + clamped_w * in_stride_c + clamped_h1 * in_stride_wc);
-            const auto a11 = *(reinterpret_cast<const T *>(in.ptr()) + clamped_w1 * in_stride_c + clamped_h1 * in_stride_wc);
-
-            *reinterpret_cast<T *>(out.ptr()) = static_cast<T>(compute_bilinear(a00, a01, a10, a11, dx_val, dy_val));
-        },
-        in, out);
-    }
-    else
-    {
-        ARM_COMPUTE_ERROR("Not implemented");
-    }
-}
-
-template <typename T>
 void NEScaleKernel::scale_bilinear_qasymm(const Window &window)
 {
     // Get data layout and width/height indices
@@ -572,7 +535,7 @@ void NEScaleKernel::scale_bilinear_qasymm(const Window &window)
             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(compute_bilinear(inp00, inp01, inp10, inp11, dx_val, dy_val), oq_info);
+            *reinterpret_cast<T *>(out.ptr()) = Qasymm8QuantizationHelper<T>::quantize(scale_helpers::delta_bilinear(inp00, inp01, inp10, inp11, dx_val, dy_val), oq_info);
         },
         in, out);
     }
@@ -600,7 +563,7 @@ void NEScaleKernel::scale_bilinear_qasymm(const Window &window)
             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(compute_bilinear(inp00, inp01, inp10, inp11, dx_val, dy_val), oq_info);
+            *reinterpret_cast<T *>(out.ptr()) = Qasymm8QuantizationHelper<T>::quantize(scale_helpers::delta_bilinear(inp00, inp01, inp10, inp11, dx_val, dy_val), oq_info);
         },
         in, out);
     }
@@ -622,8 +585,16 @@ void NEScaleKernel::run(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(_func == nullptr);
+    ARM_COMPUTE_ERROR_ON(_func == nullptr && _input->info()->data_layout() == DataLayout::NCHW);
 
-    (this->*_func)(window);
+    if(_input->info()->data_layout() == DataLayout::NCHW)
+    {
+        (this->*_func)(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);
+    }
 }
 } // namespace arm_compute
diff --git a/src/core/NEON/kernels/NEScaleKernel.h b/src/core/NEON/kernels/NEScaleKernel.h
index a3786db5b7..b93a213e99 100644
--- a/src/core/NEON/kernels/NEScaleKernel.h
+++ b/src/core/NEON/kernels/NEScaleKernel.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020 Arm Limited.
+ * Copyright (c) 2016-2021 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -96,17 +96,11 @@ private:
     void scale_bilinear_nchw(const Window &window);
     /** function to perform scale using bilinear interpolation on the given window */
     template <typename T>
-    void scale_bilinear_nhwc(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);
-    /** function to perform scale using nearest neighbour on the given window */
-    template <typename T>
-    void scale_nearest_nhwc(const Window &window);
 
     /** Scale function to use for the particular function to use */
     using ScaleFunctionPtr = void (NEScaleKernel::*)(const Window &window);
diff --git a/src/core/NEON/kernels/scale/impl/NEON/fp16.cpp b/src/core/NEON/kernels/scale/impl/NEON/fp16.cpp
new file mode 100644
index 0000000000..326723e06d
--- /dev/null
+++ b/src/core/NEON/kernels/scale/impl/NEON/fp16.cpp
@@ -0,0 +1,175 @@
+/*
+ * 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 "arm_compute/core/Helpers.h"
+#include "arm_compute/core/ITensorPack.h"
+#include "arm_compute/core/Window.h"
+#include "src/core/NEON/NEMath.h"
+#include "src/core/NEON/wrapper/wrapper.h"
+#include "src/core/common/Validate.h"
+#include "src/core/helpers/ScaleHelpers.h"
+#include "src/core/utils/ScaleUtils.h"
+#include "support/Rounding.h"
+
+#include <arm_neon.h>
+#include <cmath>
+#include <cstddef>
+
+#if defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && defined(ENABLE_FP16_KERNELS)
+
+namespace arm_compute
+{
+namespace
+{
+void fp16_neon_scale_nearest(const ITensor *src, ITensor *dst, const ITensor *offsets,
+                             float sampling_offset, bool align_corners, const Window &window)
+{
+    const size_t in_stride_c  = src->info()->dimension(0) + src->info()->padding().left + src->info()->padding().right;
+    const size_t in_stride_w  = src->info()->dimension(1) + src->info()->padding().top + src->info()->padding().bottom;
+    const size_t in_stride_wc = in_stride_w * in_stride_c;
+    const size_t in_dim_h     = src->info()->dimension(2);
+
+    // Compute the ratio between source height and destination height
+    const auto hr             = scale_utils::calculate_resize_ratio(in_dim_h, dst->info()->dimension(2), align_corners);
+    const auto window_start_x = static_cast<int32_t>(window.x().start());
+    const auto window_end_x   = static_cast<int32_t>(window.x().end());
+    const int  window_step_x  = 8;
+
+    Window win(window);
+    win.set(Window::DimX, Window::Dimension(0, 1, 1));
+    Iterator out(dst, win);
+
+    const uint8_t     *in_ptr_start        = src->buffer() + src->info()->offset_first_element_in_bytes();
+    const unsigned int in_stride_bytes_hwc = src->info()->strides_in_bytes()[3];
+
+    execute_window_loop(win, [&](const Coordinates & id)
+    {
+        const int32_t    offset     = *reinterpret_cast<const int32_t *>(offsets->ptr_to_element(Coordinates(id.y(), id.z()))) * in_stride_c;
+        const auto       in_hi      = static_cast<int>(align_corners ? utils::rounding::round_half_away_from_zero((id.z() + sampling_offset) * hr) : std::floor((id.z() + sampling_offset) * hr));
+        const int        offset_row = in_hi * in_stride_wc;
+        int32_t          x          = window_start_x;
+        const float16_t *in_ptr     = reinterpret_cast<const float16_t *>(in_ptr_start + in_stride_bytes_hwc * id[3]);
+
+        for(; x <= window_end_x - window_step_x; x += window_step_x)
+        {
+            wrapper::vstore(reinterpret_cast<float16_t *>(out.ptr()) + x,
+                            wrapper::vloadq(in_ptr + offset + offset_row + x));
+        }
+        for(; x < window_end_x; ++x)
+        {
+            *(reinterpret_cast<float16_t *>(out.ptr()) + x) = *(in_ptr + offset + offset_row + x);
+        }
+    },
+    out);
+}
+
+void fp16_neon_scale_bilinear(const ITensor *src, ITensor *dst, const ITensor *offsets, const ITensor *dx, const ITensor *dy,
+                              BorderMode border_mode, PixelValue constant_border_value, float sampling_offset,
+                              bool align_corners, const Window &window)
+{
+    // Compute the ratio between source height and destination height
+    const auto hr = scale_utils::calculate_resize_ratio(src->info()->dimension(2), dst->info()->dimension(2), align_corners);
+
+    Iterator  out(dst, window);
+    const int in_stride_c  = src->info()->dimension(0) + src->info()->padding().left + src->info()->padding().right;
+    const int in_dim_w     = src->info()->dimension(1);
+    const int in_dim_h     = src->info()->dimension(2);
+    const int in_stride_wc = in_stride_c * (in_dim_w + src->info()->padding().top + src->info()->padding().bottom);
+
+    // Don't increment in Y and Z 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(Window::DimY, Window::Dimension(0, 0, 0));
+    win_in.set(Window::DimZ, Window::Dimension(0, 0, 0));
+    Iterator in(src, win_in);
+
+    if(border_mode == BorderMode::CONSTANT)
+    {
+        using ConstType = typename std::conditional<std::is_same<float16_t, float16_t>::value, half, float16_t>::type;
+
+        const float16_t const_border_value = static_cast<float16_t>(constant_border_value.get<ConstType>());
+        execute_window_loop(window, [&](const Coordinates & id)
+        {
+            const auto       offset = *reinterpret_cast<const int32_t *>(offsets->ptr_to_element(Coordinates(id.y(), id.z())));
+            const auto       dx_val = *reinterpret_cast<const float *>(dx->ptr_to_element(Coordinates(id.y(), id.z())));
+            const auto       dy_val = *reinterpret_cast<const float *>(dy->ptr_to_element(Coordinates(id.y(), id.z())));
+            const int32_t    in_hi  = std::floor((id.z() + sampling_offset) * hr - sampling_offset);
+            const float16_t *in_ptr = reinterpret_cast<const float16_t *>(in.ptr()) + offset * in_stride_c + in_hi * in_stride_wc;
+
+            const auto a00 = (0 <= offset && offset < in_dim_w && 0 <= in_hi && in_hi < in_dim_h) ? *in_ptr : const_border_value;
+            const auto a01 = (-1 <= offset && offset < in_dim_w - 1 && 0 <= in_hi && in_hi < in_dim_h) ? *(in_ptr + in_stride_c) : const_border_value;
+            const auto a10 = (0 <= offset && offset < in_dim_w && -1 <= in_hi && in_hi < in_dim_h - 1) ? *(in_ptr + in_stride_wc) : const_border_value;
+            const auto a11 = (-1 <= offset && offset < in_dim_w - 1 && -1 <= in_hi && in_hi < in_dim_h - 1) ? *(in_ptr + in_stride_c + in_stride_wc) : const_border_value;
+
+            *reinterpret_cast<float16_t *>(out.ptr()) = static_cast<float16_t>(scale_helpers::delta_bilinear(a00, a01, a10, a11, dx_val, dy_val));
+        },
+        in, out);
+    }
+    else if(border_mode == BorderMode::REPLICATE)
+    {
+        execute_window_loop(window, [&](const Coordinates & id)
+        {
+            const auto offset = *reinterpret_cast<const int32_t *>(offsets->ptr_to_element(Coordinates(id.y(), id.z())));
+            const auto dx_val = *reinterpret_cast<const float *>(dx->ptr_to_element(Coordinates(id.y(), id.z())));
+            const auto dy_val = *reinterpret_cast<const float *>(dy->ptr_to_element(Coordinates(id.y(), id.z())));
+            const int  in_hi  = std::floor((id.z() + sampling_offset) * hr - sampling_offset);
+
+            auto clamped_w  = utility::clamp<int>(offset, 0, in_dim_w - 1);
+            auto clamped_w1 = utility::clamp<int>(offset + 1, 0, in_dim_w - 1);
+            auto clamped_h  = utility::clamp<int>(in_hi, 0, in_dim_h - 1);
+            auto clamped_h1 = utility::clamp<int>(in_hi + 1, 0, in_dim_h - 1);
+
+            const auto a00 = *(reinterpret_cast<const float16_t *>(in.ptr()) + clamped_w * in_stride_c + clamped_h * in_stride_wc);
+            const auto a01 = *(reinterpret_cast<const float16_t *>(in.ptr()) + clamped_w1 * in_stride_c + clamped_h * in_stride_wc);
+            const auto a10 = *(reinterpret_cast<const float16_t *>(in.ptr()) + clamped_w * in_stride_c + clamped_h1 * in_stride_wc);
+            const auto a11 = *(reinterpret_cast<const float16_t *>(in.ptr()) + clamped_w1 * in_stride_c + clamped_h1 * in_stride_wc);
+
+            *reinterpret_cast<float16_t *>(out.ptr()) = static_cast<float16_t>(scale_helpers::delta_bilinear(a00, a01, a10, a11, dx_val, dy_val));
+        },
+        in, out);
+    }
+    else
+    {
+        ARM_COMPUTE_ERROR("Not implemented");
+    }
+}
+}
+namespace cpu
+{
+void fp16_neon_scale(const ITensor *src, ITensor *dst, const ITensor *offsets, const ITensor *dx, const ITensor *dy,
+                     InterpolationPolicy policy, BorderMode border_mode, PixelValue constant_border_value, float sampling_offset,
+                     bool align_corners, const Window &window)
+{
+    if(policy == InterpolationPolicy::BILINEAR)
+    {
+        fp16_neon_scale_bilinear(src, dst, offsets, dx, dy, border_mode, constant_border_value, sampling_offset, align_corners, window);
+    }
+    else if(policy == InterpolationPolicy::NEAREST_NEIGHBOR)
+    {
+        fp16_neon_scale_nearest(src, dst, offsets, sampling_offset, align_corners, window);
+    }
+}
+} // namespace cpu
+} // namespace arm_compute
+
+#endif /* defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && defined(ENABLE_FP16_KERNELS) */
\ No newline at end of file
diff --git a/src/core/NEON/kernels/scale/impl/NEON/fp32.cpp b/src/core/NEON/kernels/scale/impl/NEON/fp32.cpp
new file mode 100644
index 0000000000..2e8a66123d
--- /dev/null
+++ b/src/core/NEON/kernels/scale/impl/NEON/fp32.cpp
@@ -0,0 +1,169 @@
+/*
+ * 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 "arm_compute/core/Helpers.h"
+#include "arm_compute/core/ITensorPack.h"
+#include "arm_compute/core/Window.h"
+#include "src/core/NEON/NEMath.h"
+#include "src/core/NEON/wrapper/wrapper.h"
+#include "src/core/common/Validate.h"
+#include "src/core/helpers/ScaleHelpers.h"
+#include "src/core/utils/ScaleUtils.h"
+#include "support/Rounding.h"
+
+#include <arm_neon.h>
+#include <cmath>
+#include <cstddef>
+
+namespace arm_compute
+{
+namespace
+{
+void fp32_neon_scale_nearest(const ITensor *src, ITensor *dst, const ITensor *offsets,
+                             float sampling_offset, bool align_corners, const Window &window)
+{
+    const size_t in_stride_c  = src->info()->dimension(0) + src->info()->padding().left + src->info()->padding().right;
+    const size_t in_stride_w  = src->info()->dimension(1) + src->info()->padding().top + src->info()->padding().bottom;
+    const size_t in_stride_wc = in_stride_w * in_stride_c;
+    const size_t in_dim_h     = src->info()->dimension(2);
+
+    // Compute the ratio between source height and destination height
+    const auto hr             = scale_utils::calculate_resize_ratio(in_dim_h, dst->info()->dimension(2), align_corners);
+    const auto window_start_x = static_cast<int32_t>(window.x().start());
+    const auto window_end_x   = static_cast<int32_t>(window.x().end());
+    const int  window_step_x  = 4;
+
+    Window win(window);
+    win.set(Window::DimX, Window::Dimension(0, 1, 1));
+    Iterator out(dst, win);
+
+    const uint8_t     *in_ptr_start        = src->buffer() + src->info()->offset_first_element_in_bytes();
+    const unsigned int in_stride_bytes_hwc = src->info()->strides_in_bytes()[3];
+
+    execute_window_loop(win, [&](const Coordinates & id)
+    {
+        const int32_t offset     = *reinterpret_cast<const int32_t *>(offsets->ptr_to_element(Coordinates(id.y(), id.z()))) * in_stride_c;
+        const auto    in_hi      = static_cast<int>(align_corners ? utils::rounding::round_half_away_from_zero((id.z() + sampling_offset) * hr) : std::floor((id.z() + sampling_offset) * hr));
+        const int     offset_row = in_hi * in_stride_wc;
+        int32_t       x          = window_start_x;
+        const float *in_ptr     = reinterpret_cast<const float *>(in_ptr_start + in_stride_bytes_hwc * id[3]);
+
+        for(; x <= window_end_x - window_step_x; x += window_step_x)
+        {
+            wrapper::vstore(reinterpret_cast<float *>(out.ptr()) + x,
+                            wrapper::vloadq(in_ptr + offset + offset_row + x));
+        }
+        for(; x < window_end_x; ++x)
+        {
+            *(reinterpret_cast<float *>(out.ptr()) + x) = *(in_ptr + offset + offset_row + x);
+        }
+    },
+    out);
+}
+
+void fp32_neon_scale_bilinear(const ITensor *src, ITensor *dst, const ITensor *offsets, const ITensor *dx, const ITensor *dy,
+                              BorderMode border_mode, PixelValue constant_border_value, float sampling_offset,
+                              bool align_corners, const Window &window)
+{
+    // Compute the ratio between source height and destination height
+    const auto hr = scale_utils::calculate_resize_ratio(src->info()->dimension(2), dst->info()->dimension(2), align_corners);
+
+    Iterator  out(dst, window);
+    const int in_stride_c  = src->info()->dimension(0) + src->info()->padding().left + src->info()->padding().right;
+    const int in_dim_w     = src->info()->dimension(1);
+    const int in_dim_h     = src->info()->dimension(2);
+    const int in_stride_wc = in_stride_c * (in_dim_w + src->info()->padding().top + src->info()->padding().bottom);
+
+    // Don't increment in Y and Z 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(Window::DimY, Window::Dimension(0, 0, 0));
+    win_in.set(Window::DimZ, Window::Dimension(0, 0, 0));
+    Iterator in(src, win_in);
+
+    if(border_mode == BorderMode::CONSTANT)
+    {
+        const float const_border_value = static_cast<float>(constant_border_value.get<float>());
+        execute_window_loop(window, [&](const Coordinates & id)
+        {
+            const auto    offset = *reinterpret_cast<const int32_t *>(offsets->ptr_to_element(Coordinates(id.y(), id.z())));
+            const auto    dx_val = *reinterpret_cast<const float *>(dx->ptr_to_element(Coordinates(id.y(), id.z())));
+            const auto    dy_val = *reinterpret_cast<const float *>(dy->ptr_to_element(Coordinates(id.y(), id.z())));
+            const int32_t in_hi  = std::floor((id.z() + sampling_offset) * hr - sampling_offset);
+            const float *in_ptr = reinterpret_cast<const float *>(in.ptr()) + offset * in_stride_c + in_hi * in_stride_wc;
+
+            const auto a00 = (0 <= offset && offset < in_dim_w && 0 <= in_hi && in_hi < in_dim_h) ? *in_ptr : const_border_value;
+            const auto a01 = (-1 <= offset && offset < in_dim_w - 1 && 0 <= in_hi && in_hi < in_dim_h) ? *(in_ptr + in_stride_c) : const_border_value;
+            const auto a10 = (0 <= offset && offset < in_dim_w && -1 <= in_hi && in_hi < in_dim_h - 1) ? *(in_ptr + in_stride_wc) : const_border_value;
+            const auto a11 = (-1 <= offset && offset < in_dim_w - 1 && -1 <= in_hi && in_hi < in_dim_h - 1) ? *(in_ptr + in_stride_c + in_stride_wc) : const_border_value;
+
+            *reinterpret_cast<float *>(out.ptr()) = static_cast<float>(scale_helpers::delta_bilinear(a00, a01, a10, a11, dx_val, dy_val));
+        },
+        in, out);
+    }
+    else if(border_mode == BorderMode::REPLICATE)
+    {
+        execute_window_loop(window, [&](const Coordinates & id)
+        {
+            const auto offset = *reinterpret_cast<const int32_t *>(offsets->ptr_to_element(Coordinates(id.y(), id.z())));
+            const auto dx_val = *reinterpret_cast<const float *>(dx->ptr_to_element(Coordinates(id.y(), id.z())));
+            const auto dy_val = *reinterpret_cast<const float *>(dy->ptr_to_element(Coordinates(id.y(), id.z())));
+            const int  in_hi  = std::floor((id.z() + sampling_offset) * hr - sampling_offset);
+
+            auto clamped_w  = utility::clamp<int>(offset, 0, in_dim_w - 1);
+            auto clamped_w1 = utility::clamp<int>(offset + 1, 0, in_dim_w - 1);
+            auto clamped_h  = utility::clamp<int>(in_hi, 0, in_dim_h - 1);
+            auto clamped_h1 = utility::clamp<int>(in_hi + 1, 0, in_dim_h - 1);
+
+            const auto a00 = *(reinterpret_cast<const float *>(in.ptr()) + clamped_w * in_stride_c + clamped_h * in_stride_wc);
+            const auto a01 = *(reinterpret_cast<const float *>(in.ptr()) + clamped_w1 * in_stride_c + clamped_h * in_stride_wc);
+            const auto a10 = *(reinterpret_cast<const float *>(in.ptr()) + clamped_w * in_stride_c + clamped_h1 * in_stride_wc);
+            const auto a11 = *(reinterpret_cast<const float *>(in.ptr()) + clamped_w1 * in_stride_c + clamped_h1 * in_stride_wc);
+
+            *reinterpret_cast<float *>(out.ptr()) = static_cast<float>(scale_helpers::delta_bilinear(a00, a01, a10, a11, dx_val, dy_val));
+        },
+        in, out);
+    }
+    else
+    {
+        ARM_COMPUTE_ERROR("Not implemented");
+    }
+}
+}
+namespace cpu
+{
+void fp32_neon_scale(const ITensor *src, ITensor *dst, const ITensor *offsets, const ITensor *dx, const ITensor *dy,
+                     InterpolationPolicy policy, BorderMode border_mode, PixelValue constant_border_value, float sampling_offset,
+                     bool align_corners, const Window &window)
+{
+    if(policy == InterpolationPolicy::BILINEAR)
+    {
+        fp32_neon_scale_bilinear(src, dst, offsets, dx, dy, border_mode, constant_border_value, sampling_offset, align_corners, window);
+    }
+    else if(policy == InterpolationPolicy::NEAREST_NEIGHBOR)
+    {
+        fp32_neon_scale_nearest(src, dst, offsets, sampling_offset, align_corners, window);
+    }
+}
+} // namespace cpu
+} // namespace arm_compute
\ No newline at end of file
diff --git a/src/core/NEON/kernels/scale/impl/NEON/integer.cpp b/src/core/NEON/kernels/scale/impl/NEON/integer.cpp
new file mode 100644
index 0000000000..7fc8693e95
--- /dev/null
+++ b/src/core/NEON/kernels/scale/impl/NEON/integer.cpp
@@ -0,0 +1,294 @@
+/*
+ * 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 "arm_compute/core/Helpers.h"
+#include "arm_compute/core/ITensorPack.h"
+#include "arm_compute/core/Window.h"
+#include "src/core/NEON/NEMath.h"
+#include "src/core/NEON/wrapper/wrapper.h"
+#include "src/core/common/Validate.h"
+#include "src/core/helpers/ScaleHelpers.h"
+#include "src/core/utils/ScaleUtils.h"
+#include "support/Rounding.h"
+
+#include <arm_neon.h>
+#include <cmath>
+#include <cstddef>
+
+namespace arm_compute
+{
+namespace
+{
+void u8_neon_scale_nearest(const ITensor *src, ITensor *dst, const ITensor *offsets,
+                           float sampling_offset, bool align_corners, const Window &window)
+{
+    const size_t in_stride_c  = src->info()->dimension(0) + src->info()->padding().left + src->info()->padding().right;
+    const size_t in_stride_w  = src->info()->dimension(1) + src->info()->padding().top + src->info()->padding().bottom;
+    const size_t in_stride_wc = in_stride_w * in_stride_c;
+    const size_t in_dim_h     = src->info()->dimension(2);
+
+    // Compute the ratio between source height and destination height
+    const auto hr             = scale_utils::calculate_resize_ratio(in_dim_h, dst->info()->dimension(2), align_corners);
+    const auto window_start_x = static_cast<int32_t>(window.x().start());
+    const auto window_end_x   = static_cast<int32_t>(window.x().end());
+    const int  window_step_x  = 16;
+
+    Window win(window);
+    win.set(Window::DimX, Window::Dimension(0, 1, 1));
+    Iterator out(dst, win);
+
+    const uint8_t     *in_ptr_start        = src->buffer() + src->info()->offset_first_element_in_bytes();
+    const unsigned int in_stride_bytes_hwc = src->info()->strides_in_bytes()[3];
+
+    execute_window_loop(win, [&](const Coordinates & id)
+    {
+        const int32_t  offset     = *reinterpret_cast<const int32_t *>(offsets->ptr_to_element(Coordinates(id.y(), id.z()))) * in_stride_c;
+        const auto     in_hi      = static_cast<int>(align_corners ? utils::rounding::round_half_away_from_zero((id.z() + sampling_offset) * hr) : std::floor((id.z() + sampling_offset) * hr));
+        const int      offset_row = in_hi * in_stride_wc;
+        int32_t        x          = window_start_x;
+        const uint8_t *in_ptr     = reinterpret_cast<const uint8_t *>(in_ptr_start + in_stride_bytes_hwc * id[3]);
+
+        for(; x <= window_end_x - window_step_x; x += window_step_x)
+        {
+            wrapper::vstore(reinterpret_cast<uint8_t *>(out.ptr()) + x,
+                            wrapper::vloadq(in_ptr + offset + offset_row + x));
+        }
+        for(; x < window_end_x; ++x)
+        {
+            *(reinterpret_cast<uint8_t *>(out.ptr()) + x) = *(in_ptr + offset + offset_row + x);
+        }
+    },
+    out);
+}
+
+void u8_neon_scale_bilinear(const ITensor *src, ITensor *dst, const ITensor *offsets, const ITensor *dx, const ITensor *dy,
+                            BorderMode border_mode, PixelValue constant_border_value, float sampling_offset,
+                            bool align_corners, const Window &window)
+{
+    // Compute the ratio between source height and destination height
+    const auto hr = scale_utils::calculate_resize_ratio(src->info()->dimension(2), dst->info()->dimension(2), align_corners);
+
+    Iterator  out(dst, window);
+    const int in_stride_c  = src->info()->dimension(0) + src->info()->padding().left + src->info()->padding().right;
+    const int in_dim_w     = src->info()->dimension(1);
+    const int in_dim_h     = src->info()->dimension(2);
+    const int in_stride_wc = in_stride_c * (in_dim_w + src->info()->padding().top + src->info()->padding().bottom);
+
+    // Don't increment in Y and Z 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(Window::DimY, Window::Dimension(0, 0, 0));
+    win_in.set(Window::DimZ, Window::Dimension(0, 0, 0));
+    Iterator in(src, win_in);
+
+    if(border_mode == BorderMode::CONSTANT)
+    {
+        const uint8_t const_border_value = static_cast<uint8_t>(constant_border_value.get<uint8_t>());
+        execute_window_loop(window, [&](const Coordinates & id)
+        {
+            const auto     offset = *reinterpret_cast<const int32_t *>(offsets->ptr_to_element(Coordinates(id.y(), id.z())));
+            const auto     dx_val = *reinterpret_cast<const float *>(dx->ptr_to_element(Coordinates(id.y(), id.z())));
+            const auto     dy_val = *reinterpret_cast<const float *>(dy->ptr_to_element(Coordinates(id.y(), id.z())));
+            const int32_t  in_hi  = std::floor((id.z() + sampling_offset) * hr - sampling_offset);
+            const uint8_t *in_ptr = reinterpret_cast<const uint8_t *>(in.ptr()) + offset * in_stride_c + in_hi * in_stride_wc;
+
+            const auto a00 = (0 <= offset && offset < in_dim_w && 0 <= in_hi && in_hi < in_dim_h) ? *in_ptr : const_border_value;
+            const auto a01 = (-1 <= offset && offset < in_dim_w - 1 && 0 <= in_hi && in_hi < in_dim_h) ? *(in_ptr + in_stride_c) : const_border_value;
+            const auto a10 = (0 <= offset && offset < in_dim_w && -1 <= in_hi && in_hi < in_dim_h - 1) ? *(in_ptr + in_stride_wc) : const_border_value;
+            const auto a11 = (-1 <= offset && offset < in_dim_w - 1 && -1 <= in_hi && in_hi < in_dim_h - 1) ? *(in_ptr + in_stride_c + in_stride_wc) : const_border_value;
+
+            *reinterpret_cast<uint8_t *>(out.ptr()) = static_cast<uint8_t>(scale_helpers::delta_bilinear(a00, a01, a10, a11, dx_val, dy_val));
+        },
+        in, out);
+    }
+    else if(border_mode == BorderMode::REPLICATE)
+    {
+        execute_window_loop(window, [&](const Coordinates & id)
+        {
+            const auto offset = *reinterpret_cast<const int32_t *>(offsets->ptr_to_element(Coordinates(id.y(), id.z())));
+            const auto dx_val = *reinterpret_cast<const float *>(dx->ptr_to_element(Coordinates(id.y(), id.z())));
+            const auto dy_val = *reinterpret_cast<const float *>(dy->ptr_to_element(Coordinates(id.y(), id.z())));
+            const int  in_hi  = std::floor((id.z() + sampling_offset) * hr - sampling_offset);
+
+            auto clamped_w  = utility::clamp<int>(offset, 0, in_dim_w - 1);
+            auto clamped_w1 = utility::clamp<int>(offset + 1, 0, in_dim_w - 1);
+            auto clamped_h  = utility::clamp<int>(in_hi, 0, in_dim_h - 1);
+            auto clamped_h1 = utility::clamp<int>(in_hi + 1, 0, in_dim_h - 1);
+
+            const auto a00 = *(reinterpret_cast<const uint8_t *>(in.ptr()) + clamped_w * in_stride_c + clamped_h * in_stride_wc);
+            const auto a01 = *(reinterpret_cast<const uint8_t *>(in.ptr()) + clamped_w1 * in_stride_c + clamped_h * in_stride_wc);
+            const auto a10 = *(reinterpret_cast<const uint8_t *>(in.ptr()) + clamped_w * in_stride_c + clamped_h1 * in_stride_wc);
+            const auto a11 = *(reinterpret_cast<const uint8_t *>(in.ptr()) + clamped_w1 * in_stride_c + clamped_h1 * in_stride_wc);
+
+            *reinterpret_cast<uint8_t *>(out.ptr()) = static_cast<uint8_t>(scale_helpers::delta_bilinear(a00, a01, a10, a11, dx_val, dy_val));
+        },
+        in, out);
+    }
+    else
+    {
+        ARM_COMPUTE_ERROR("Not implemented");
+    }
+}
+
+void s16_neon_scale_nearest(const ITensor *src, ITensor *dst, const ITensor *offsets,
+                            float sampling_offset, bool align_corners, const Window &window)
+{
+    const size_t in_stride_c  = src->info()->dimension(0) + src->info()->padding().left + src->info()->padding().right;
+    const size_t in_stride_w  = src->info()->dimension(1) + src->info()->padding().top + src->info()->padding().bottom;
+    const size_t in_stride_wc = in_stride_w * in_stride_c;
+    const size_t in_dim_h     = src->info()->dimension(2);
+
+    // Compute the ratio between source height and destination height
+    const auto hr             = scale_utils::calculate_resize_ratio(in_dim_h, dst->info()->dimension(2), align_corners);
+    const auto window_start_x = static_cast<int32_t>(window.x().start());
+    const auto window_end_x   = static_cast<int32_t>(window.x().end());
+    const int  window_step_x  = 8;
+
+    Window win(window);
+    win.set(Window::DimX, Window::Dimension(0, 1, 1));
+    Iterator out(dst, win);
+
+    const uint8_t     *in_ptr_start        = src->buffer() + src->info()->offset_first_element_in_bytes();
+    const unsigned int in_stride_bytes_hwc = src->info()->strides_in_bytes()[3];
+
+    execute_window_loop(win, [&](const Coordinates & id)
+    {
+        const int32_t  offset     = *reinterpret_cast<const int32_t *>(offsets->ptr_to_element(Coordinates(id.y(), id.z()))) * in_stride_c;
+        const auto     in_hi      = static_cast<int>(align_corners ? utils::rounding::round_half_away_from_zero((id.z() + sampling_offset) * hr) : std::floor((id.z() + sampling_offset) * hr));
+        const int      offset_row = in_hi * in_stride_wc;
+        int32_t        x          = window_start_x;
+        const int16_t *in_ptr     = reinterpret_cast<const int16_t *>(in_ptr_start + in_stride_bytes_hwc * id[3]);
+
+        for(; x <= window_end_x - window_step_x; x += window_step_x)
+        {
+            wrapper::vstore(reinterpret_cast<int16_t *>(out.ptr()) + x,
+                            wrapper::vloadq(in_ptr + offset + offset_row + x));
+        }
+        for(; x < window_end_x; ++x)
+        {
+            *(reinterpret_cast<int16_t *>(out.ptr()) + x) = *(in_ptr + offset + offset_row + x);
+        }
+    },
+    out);
+}
+
+void s16_neon_scale_bilinear(const ITensor *src, ITensor *dst, const ITensor *offsets, const ITensor *dx, const ITensor *dy,
+                             BorderMode border_mode, PixelValue constant_border_value, float sampling_offset,
+                             bool align_corners, const Window &window)
+{
+    // Compute the ratio between source height and destination height
+    const auto hr = scale_utils::calculate_resize_ratio(src->info()->dimension(2), dst->info()->dimension(2), align_corners);
+
+    Iterator  out(dst, window);
+    const int in_stride_c  = src->info()->dimension(0) + src->info()->padding().left + src->info()->padding().right;
+    const int in_dim_w     = src->info()->dimension(1);
+    const int in_dim_h     = src->info()->dimension(2);
+    const int in_stride_wc = in_stride_c * (in_dim_w + src->info()->padding().top + src->info()->padding().bottom);
+
+    // Don't increment in Y and Z 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(Window::DimY, Window::Dimension(0, 0, 0));
+    win_in.set(Window::DimZ, Window::Dimension(0, 0, 0));
+    Iterator in(src, win_in);
+
+    if(border_mode == BorderMode::CONSTANT)
+    {
+        const int16_t const_border_value = static_cast<int16_t>(constant_border_value.get<int16_t>());
+        execute_window_loop(window, [&](const Coordinates & id)
+        {
+            const auto     offset = *reinterpret_cast<const int32_t *>(offsets->ptr_to_element(Coordinates(id.y(), id.z())));
+            const auto     dx_val = *reinterpret_cast<const float *>(dx->ptr_to_element(Coordinates(id.y(), id.z())));
+            const auto     dy_val = *reinterpret_cast<const float *>(dy->ptr_to_element(Coordinates(id.y(), id.z())));
+            const int32_t  in_hi  = std::floor((id.z() + sampling_offset) * hr - sampling_offset);
+            const int16_t *in_ptr = reinterpret_cast<const int16_t *>(in.ptr()) + offset * in_stride_c + in_hi * in_stride_wc;
+
+            const auto a00 = (0 <= offset && offset < in_dim_w && 0 <= in_hi && in_hi < in_dim_h) ? *in_ptr : const_border_value;
+            const auto a01 = (-1 <= offset && offset < in_dim_w - 1 && 0 <= in_hi && in_hi < in_dim_h) ? *(in_ptr + in_stride_c) : const_border_value;
+            const auto a10 = (0 <= offset && offset < in_dim_w && -1 <= in_hi && in_hi < in_dim_h - 1) ? *(in_ptr + in_stride_wc) : const_border_value;
+            const auto a11 = (-1 <= offset && offset < in_dim_w - 1 && -1 <= in_hi && in_hi < in_dim_h - 1) ? *(in_ptr + in_stride_c + in_stride_wc) : const_border_value;
+
+            *reinterpret_cast<int16_t *>(out.ptr()) = static_cast<int16_t>(scale_helpers::delta_bilinear(a00, a01, a10, a11, dx_val, dy_val));
+        },
+        in, out);
+    }
+    else if(border_mode == BorderMode::REPLICATE)
+    {
+        execute_window_loop(window, [&](const Coordinates & id)
+        {
+            const auto offset = *reinterpret_cast<const int32_t *>(offsets->ptr_to_element(Coordinates(id.y(), id.z())));
+            const auto dx_val = *reinterpret_cast<const float *>(dx->ptr_to_element(Coordinates(id.y(), id.z())));
+            const auto dy_val = *reinterpret_cast<const float *>(dy->ptr_to_element(Coordinates(id.y(), id.z())));
+            const int  in_hi  = std::floor((id.z() + sampling_offset) * hr - sampling_offset);
+
+            auto clamped_w  = utility::clamp<int>(offset, 0, in_dim_w - 1);
+            auto clamped_w1 = utility::clamp<int>(offset + 1, 0, in_dim_w - 1);
+            auto clamped_h  = utility::clamp<int>(in_hi, 0, in_dim_h - 1);
+            auto clamped_h1 = utility::clamp<int>(in_hi + 1, 0, in_dim_h - 1);
+
+            const auto a00 = *(reinterpret_cast<const int16_t *>(in.ptr()) + clamped_w * in_stride_c + clamped_h * in_stride_wc);
+            const auto a01 = *(reinterpret_cast<const int16_t *>(in.ptr()) + clamped_w1 * in_stride_c + clamped_h * in_stride_wc);
+            const auto a10 = *(reinterpret_cast<const int16_t *>(in.ptr()) + clamped_w * in_stride_c + clamped_h1 * in_stride_wc);
+            const auto a11 = *(reinterpret_cast<const int16_t *>(in.ptr()) + clamped_w1 * in_stride_c + clamped_h1 * in_stride_wc);
+
+            *reinterpret_cast<int16_t *>(out.ptr()) = static_cast<int16_t>(scale_helpers::delta_bilinear(a00, a01, a10, a11, dx_val, dy_val));
+        },
+        in, out);
+    }
+    else
+    {
+        ARM_COMPUTE_ERROR("Not implemented");
+    }
+}
+}
+namespace cpu
+{
+void u8_neon_scale(const ITensor *src, ITensor *dst, const ITensor *offsets, const ITensor *dx, const ITensor *dy,
+                   InterpolationPolicy policy, BorderMode border_mode, PixelValue constant_border_value, float sampling_offset,
+                   bool align_corners, const Window &window)
+{
+    if(policy == InterpolationPolicy::BILINEAR)
+    {
+        u8_neon_scale_bilinear(src, dst, offsets, dx, dy, border_mode, constant_border_value, sampling_offset, align_corners, window);
+    }
+    else if(policy == InterpolationPolicy::NEAREST_NEIGHBOR)
+    {
+        u8_neon_scale_nearest(src, dst, offsets, sampling_offset, align_corners, window);
+    }
+}
+
+void s16_neon_scale(const ITensor *src, ITensor *dst, const ITensor *offsets, const ITensor *dx, const ITensor *dy,
+                    InterpolationPolicy policy, BorderMode border_mode, PixelValue constant_border_value, float sampling_offset,
+                    bool align_corners, const Window &window)
+{
+    if(policy == InterpolationPolicy::BILINEAR)
+    {
+        s16_neon_scale_bilinear(src, dst, offsets, dx, dy, border_mode, constant_border_value, sampling_offset, align_corners, window);
+    }
+    else if(policy == InterpolationPolicy::NEAREST_NEIGHBOR)
+    {
+        s16_neon_scale_nearest(src, dst, offsets, sampling_offset, align_corners, window);
+    }
+}
+} // namespace cpu
+} // namespace arm_compute
\ No newline at end of file
diff --git a/src/core/NEON/kernels/scale/impl/NEON/qasymm8.cpp b/src/core/NEON/kernels/scale/impl/NEON/qasymm8.cpp
new file mode 100644
index 0000000000..2e5d82a65b
--- /dev/null
+++ b/src/core/NEON/kernels/scale/impl/NEON/qasymm8.cpp
@@ -0,0 +1,203 @@
+/*
+ * 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 "arm_compute/core/Helpers.h"
+#include "arm_compute/core/ITensorPack.h"
+#include "arm_compute/core/Window.h"
+#include "src/core/NEON/NEMath.h"
+#include "src/core/NEON/wrapper/wrapper.h"
+#include "src/core/common/Validate.h"
+#include "src/core/helpers/ScaleHelpers.h"
+#include "src/core/utils/ScaleUtils.h"
+#include "support/Rounding.h"
+
+#include <arm_neon.h>
+#include <cmath>
+#include <cstddef>
+
+namespace arm_compute
+{
+namespace
+{
+void qasymm8_neon_scale_nearest(const ITensor *src, ITensor *dst, const ITensor *offsets,
+                                float sampling_offset, bool align_corners, const Window &window)
+{
+    const size_t in_stride_c  = src->info()->dimension(0) + src->info()->padding().left + src->info()->padding().right;
+    const size_t in_stride_w  = src->info()->dimension(1) + src->info()->padding().top + src->info()->padding().bottom;
+    const size_t in_stride_wc = in_stride_w * in_stride_c;
+    const size_t in_dim_h     = src->info()->dimension(2);
+
+    // Compute the ratio between source height and destination height
+    const auto hr             = scale_utils::calculate_resize_ratio(in_dim_h, dst->info()->dimension(2), align_corners);
+    const auto window_start_x = static_cast<int32_t>(window.x().start());
+    const auto window_end_x   = static_cast<int32_t>(window.x().end());
+    const int  window_step_x  = 16;
+
+    Window win(window);
+    win.set(Window::DimX, Window::Dimension(0, 1, 1));
+    Iterator out(dst, win);
+
+    const uint8_t     *in_ptr_start        = src->buffer() + src->info()->offset_first_element_in_bytes();
+    const unsigned int in_stride_bytes_hwc = src->info()->strides_in_bytes()[3];
+
+    execute_window_loop(win, [&](const Coordinates & id)
+    {
+        const int32_t  offset     = *reinterpret_cast<const int32_t *>(offsets->ptr_to_element(Coordinates(id.y(), id.z()))) * in_stride_c;
+        const auto     in_hi      = static_cast<int>(align_corners ? utils::rounding::round_half_away_from_zero((id.z() + sampling_offset) * hr) : std::floor((id.z() + sampling_offset) * hr));
+        const int      offset_row = in_hi * in_stride_wc;
+        int32_t        x          = window_start_x;
+        const uint8_t *in_ptr     = reinterpret_cast<const uint8_t *>(in_ptr_start + in_stride_bytes_hwc * id[3]);
+
+        for(; x <= window_end_x - window_step_x; x += window_step_x)
+        {
+            wrapper::vstore(reinterpret_cast<uint8_t *>(out.ptr()) + x,
+                            wrapper::vloadq(in_ptr + offset + offset_row + x));
+        }
+        for(; x < window_end_x; ++x)
+        {
+            *(reinterpret_cast<uint8_t *>(out.ptr()) + x) = *(in_ptr + offset + offset_row + x);
+        }
+    },
+    out);
+}
+
+void qasymm8_neon_scale_bilinear(const ITensor *src, ITensor *dst, const ITensor *offsets, const ITensor *dx, const ITensor *dy,
+                                 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);
+
+    // 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);
+    Window     win_off;
+    win_off.set(Window::DimX, Window::Dimension(0, 0, 0));
+    win_off.set(Window::DimY, Window::Dimension(0, 0, 0));
+
+    // 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));
+
+    for(size_t d = Window::DimZ; d < offsets->info()->num_dimensions(); ++d)
+    {
+        win_off.set(d, Window::Dimension(0, 0, 0));
+    }
+
+    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 UniformQuantizationInfo iq_info = src->info()->quantization_info().uniform();
+    const UniformQuantizationInfo oq_info = dst->info()->quantization_info().uniform();
+
+    if(border_mode == BorderMode::CONSTANT)
+    {
+        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 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) ?
+                             (*(pixel_row_ptr + index_w * stride_w + index_h * stride_h)) :
+                             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) * stride_w + index_h * stride_h)) :
+                             const_border_value;
+            const auto a10 = (0 <= index_w && index_w < in_dim_w && -1 <= index_h && index_h < in_dim_h - 1) ?
+                             (*(pixel_row_ptr + index_w * stride_w + (index_h + 1) * stride_h)) :
+                             const_border_value;
+            const auto a11 = (-1 <= index_w && index_w < in_dim_w - 1 && -1 <= index_h && index_h < in_dim_h - 1) ?
+                             (*(pixel_row_ptr + (index_w + 1) * stride_w + (index_h + 1) * stride_h)) :
+                             const_border_value;
+
+            const float inp00                       = Qasymm8QuantizationHelper<uint8_t>::dequantize(a00, iq_info);
+            const float inp01                       = Qasymm8QuantizationHelper<uint8_t>::dequantize(a01, iq_info);
+            const float inp10                       = Qasymm8QuantizationHelper<uint8_t>::dequantize(a10, iq_info);
+            const float inp11                       = Qasymm8QuantizationHelper<uint8_t>::dequantize(a11, iq_info);
+            *reinterpret_cast<uint8_t *>(out.ptr()) = Qasymm8QuantizationHelper<uint8_t>::quantize(scale_helpers::delta_bilinear(inp00, inp01, inp10, inp11, dx_val, dy_val), oq_info);
+        },
+        in, out);
+    }
+    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 uint8_t *>(in.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);
+            auto clamped_h  = utility::clamp<int>(index_h, 0, in_dim_h - 1);
+            auto clamped_h1 = utility::clamp<int>(index_h + 1, 0, in_dim_h - 1);
+
+            const auto a00 = *(pixel_row_ptr + clamped_w * stride_w + clamped_h * stride_h);
+            const auto a01 = *(pixel_row_ptr + clamped_w1 * stride_w + clamped_h * stride_h);
+            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<uint8_t>::dequantize(a00, iq_info);
+            const float inp01                       = Qasymm8QuantizationHelper<uint8_t>::dequantize(a01, iq_info);
+            const float inp10                       = Qasymm8QuantizationHelper<uint8_t>::dequantize(a10, iq_info);
+            const float inp11                       = Qasymm8QuantizationHelper<uint8_t>::dequantize(a11, iq_info);
+            *reinterpret_cast<uint8_t *>(out.ptr()) = Qasymm8QuantizationHelper<uint8_t>::quantize(scale_helpers::delta_bilinear(inp00, inp01, inp10, inp11, dx_val, dy_val), oq_info);
+        },
+        in, out);
+    }
+    else
+    {
+        ARM_COMPUTE_ERROR("Not implemented");
+    }
+}
+}
+namespace cpu
+{
+void qasymm8_neon_scale(const ITensor *src, ITensor *dst, const ITensor *offsets, const ITensor *dx, const ITensor *dy,
+                        InterpolationPolicy policy, BorderMode border_mode, PixelValue constant_border_value, float sampling_offset,
+                        bool align_corners, const Window &window)
+{
+    if(policy == InterpolationPolicy::BILINEAR)
+    {
+        qasymm8_neon_scale_bilinear(src, dst, offsets, dx, dy, border_mode, constant_border_value, sampling_offset, align_corners, window);
+    }
+    else if(policy == InterpolationPolicy::NEAREST_NEIGHBOR)
+    {
+        qasymm8_neon_scale_nearest(src, dst, offsets, sampling_offset, align_corners, window);
+    }
+}
+} // namespace cpu
+} // namespace arm_compute
\ No newline at end of file
diff --git a/src/core/NEON/kernels/scale/impl/NEON/qasymm8_signed.cpp b/src/core/NEON/kernels/scale/impl/NEON/qasymm8_signed.cpp
new file mode 100644
index 0000000000..c9d71ef8df
--- /dev/null
+++ b/src/core/NEON/kernels/scale/impl/NEON/qasymm8_signed.cpp
@@ -0,0 +1,203 @@
+/*
+ * 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 "arm_compute/core/Helpers.h"
+#include "arm_compute/core/ITensorPack.h"
+#include "arm_compute/core/Window.h"
+#include "src/core/NEON/NEMath.h"
+#include "src/core/NEON/wrapper/wrapper.h"
+#include "src/core/common/Validate.h"
+#include "src/core/helpers/ScaleHelpers.h"
+#include "src/core/utils/ScaleUtils.h"
+#include "support/Rounding.h"
+
+#include <arm_neon.h>
+#include <cmath>
+#include <cstddef>
+
+namespace arm_compute
+{
+namespace
+{
+void qasymm8_signed_neon_scale_nearest(const ITensor *src, ITensor *dst, const ITensor *offsets,
+                                       float sampling_offset, bool align_corners, const Window &window)
+{
+    const size_t in_stride_c  = src->info()->dimension(0) + src->info()->padding().left + src->info()->padding().right;
+    const size_t in_stride_w  = src->info()->dimension(1) + src->info()->padding().top + src->info()->padding().bottom;
+    const size_t in_stride_wc = in_stride_w * in_stride_c;
+    const size_t in_dim_h     = src->info()->dimension(2);
+
+    // Compute the ratio between source height and destination height
+    const auto hr             = scale_utils::calculate_resize_ratio(in_dim_h, dst->info()->dimension(2), align_corners);
+    const auto window_start_x = static_cast<int32_t>(window.x().start());
+    const auto window_end_x   = static_cast<int32_t>(window.x().end());
+    const int  window_step_x  = 16;
+
+    Window win(window);
+    win.set(Window::DimX, Window::Dimension(0, 1, 1));
+    Iterator out(dst, win);
+
+    const uint8_t     *in_ptr_start        = src->buffer() + src->info()->offset_first_element_in_bytes();
+    const unsigned int in_stride_bytes_hwc = src->info()->strides_in_bytes()[3];
+
+    execute_window_loop(win, [&](const Coordinates & id)
+    {
+        const int32_t offset     = *reinterpret_cast<const int32_t *>(offsets->ptr_to_element(Coordinates(id.y(), id.z()))) * in_stride_c;
+        const auto    in_hi      = static_cast<int>(align_corners ? utils::rounding::round_half_away_from_zero((id.z() + sampling_offset) * hr) : std::floor((id.z() + sampling_offset) * hr));
+        const int     offset_row = in_hi * in_stride_wc;
+        int32_t       x          = window_start_x;
+        const int8_t *in_ptr     = reinterpret_cast<const int8_t *>(in_ptr_start + in_stride_bytes_hwc * id[3]);
+
+        for(; x <= window_end_x - window_step_x; x += window_step_x)
+        {
+            wrapper::vstore(reinterpret_cast<int8_t *>(out.ptr()) + x,
+                            wrapper::vloadq(in_ptr + offset + offset_row + x));
+        }
+        for(; x < window_end_x; ++x)
+        {
+            *(reinterpret_cast<int8_t *>(out.ptr()) + x) = *(in_ptr + offset + offset_row + x);
+        }
+    },
+    out);
+}
+
+void qasymm8_signed_neon_scale_bilinear(const ITensor *src, ITensor *dst, const ITensor *offsets, const ITensor *dx, const ITensor *dy,
+                                        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);
+
+    // 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);
+    Window     win_off;
+    win_off.set(Window::DimX, Window::Dimension(0, 0, 0));
+    win_off.set(Window::DimY, Window::Dimension(0, 0, 0));
+
+    // 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));
+
+    for(size_t d = Window::DimZ; d < offsets->info()->num_dimensions(); ++d)
+    {
+        win_off.set(d, Window::Dimension(0, 0, 0));
+    }
+
+    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 UniformQuantizationInfo iq_info = src->info()->quantization_info().uniform();
+    const UniformQuantizationInfo oq_info = dst->info()->quantization_info().uniform();
+
+    if(border_mode == BorderMode::CONSTANT)
+    {
+        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 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) ?
+                             (*(pixel_row_ptr + index_w * stride_w + index_h * stride_h)) :
+                             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) * stride_w + index_h * stride_h)) :
+                             const_border_value;
+            const auto a10 = (0 <= index_w && index_w < in_dim_w && -1 <= index_h && index_h < in_dim_h - 1) ?
+                             (*(pixel_row_ptr + index_w * stride_w + (index_h + 1) * stride_h)) :
+                             const_border_value;
+            const auto a11 = (-1 <= index_w && index_w < in_dim_w - 1 && -1 <= index_h && index_h < in_dim_h - 1) ?
+                             (*(pixel_row_ptr + (index_w + 1) * stride_w + (index_h + 1) * stride_h)) :
+                             const_border_value;
+
+            const float inp00                      = Qasymm8QuantizationHelper<int8_t>::dequantize(a00, iq_info);
+            const float inp01                      = Qasymm8QuantizationHelper<int8_t>::dequantize(a01, iq_info);
+            const float inp10                      = Qasymm8QuantizationHelper<int8_t>::dequantize(a10, iq_info);
+            const float inp11                      = Qasymm8QuantizationHelper<int8_t>::dequantize(a11, iq_info);
+            *reinterpret_cast<int8_t *>(out.ptr()) = Qasymm8QuantizationHelper<int8_t>::quantize(scale_helpers::delta_bilinear(inp00, inp01, inp10, inp11, dx_val, dy_val), oq_info);
+        },
+        in, out);
+    }
+    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 int8_t *>(in.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);
+            auto clamped_h  = utility::clamp<int>(index_h, 0, in_dim_h - 1);
+            auto clamped_h1 = utility::clamp<int>(index_h + 1, 0, in_dim_h - 1);
+
+            const auto a00 = *(pixel_row_ptr + clamped_w * stride_w + clamped_h * stride_h);
+            const auto a01 = *(pixel_row_ptr + clamped_w1 * stride_w + clamped_h * stride_h);
+            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<int8_t>::dequantize(a00, iq_info);
+            const float inp01                      = Qasymm8QuantizationHelper<int8_t>::dequantize(a01, iq_info);
+            const float inp10                      = Qasymm8QuantizationHelper<int8_t>::dequantize(a10, iq_info);
+            const float inp11                      = Qasymm8QuantizationHelper<int8_t>::dequantize(a11, iq_info);
+            *reinterpret_cast<int8_t *>(out.ptr()) = Qasymm8QuantizationHelper<int8_t>::quantize(scale_helpers::delta_bilinear(inp00, inp01, inp10, inp11, dx_val, dy_val), oq_info);
+        },
+        in, out);
+    }
+    else
+    {
+        ARM_COMPUTE_ERROR("Not implemented");
+    }
+}
+}
+namespace cpu
+{
+void qasymm8_signed_neon_scale(const ITensor *src, ITensor *dst, const ITensor *offsets, const ITensor *dx, const ITensor *dy,
+                               InterpolationPolicy policy, BorderMode border_mode, PixelValue constant_border_value, float sampling_offset,
+                               bool align_corners, const Window &window)
+{
+    if(policy == InterpolationPolicy::BILINEAR)
+    {
+        qasymm8_signed_neon_scale_bilinear(src, dst, offsets, dx, dy, border_mode, constant_border_value, sampling_offset, align_corners, window);
+    }
+    else if(policy == InterpolationPolicy::NEAREST_NEIGHBOR)
+    {
+        qasymm8_signed_neon_scale_nearest(src, dst, offsets, sampling_offset, align_corners, window);
+    }
+}
+} // namespace cpu
+} // namespace arm_compute
\ No newline at end of file
diff --git a/src/core/NEON/kernels/scale/impl/SVE/fp16.cpp b/src/core/NEON/kernels/scale/impl/SVE/fp16.cpp
new file mode 100644
index 0000000000..91c3dc3b4e
--- /dev/null
+++ b/src/core/NEON/kernels/scale/impl/SVE/fp16.cpp
@@ -0,0 +1,177 @@
+/*
+ * 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 "arm_compute/core/Helpers.h"
+#include "arm_compute/core/ITensorPack.h"
+#include "arm_compute/core/Window.h"
+#include "src/core/NEON/NEMath.h"
+#include "src/core/NEON/wrapper/wrapper.h"
+#include "src/core/common/Validate.h"
+#include "src/core/helpers/ScaleHelpers.h"
+#include "src/core/utils/ScaleUtils.h"
+#include "support/Rounding.h"
+
+#include <cmath>
+#include <cstddef>
+
+#if defined(__ARM_FEATURE_SVE)
+#include <arm_sve.h>
+
+namespace arm_compute
+{
+namespace
+{
+void fp16_sve_scale_nearest(const ITensor *src, ITensor *dst, const ITensor *offsets,
+                            float sampling_offset, bool align_corners, const Window &window)
+{
+    const size_t in_stride_c  = src->info()->dimension(0) + src->info()->padding().left + src->info()->padding().right;
+    const size_t in_stride_w  = src->info()->dimension(1) + src->info()->padding().top + src->info()->padding().bottom;
+    const size_t in_stride_wc = in_stride_w * in_stride_c;
+    const size_t in_dim_h     = src->info()->dimension(2);
+
+    // Compute the ratio between source height and destination height
+    const auto hr             = scale_utils::calculate_resize_ratio(in_dim_h, dst->info()->dimension(2), align_corners);
+    const auto window_start_x = static_cast<int32_t>(window.x().start());
+    const auto window_end_x   = static_cast<int32_t>(window.x().end());
+
+    Window win(window);
+    win.set(Window::DimX, Window::Dimension(0, 1, 1));
+    Iterator out(dst, win);
+
+    const uint8_t     *in_ptr_start        = src->buffer() + src->info()->offset_first_element_in_bytes();
+    const unsigned int in_stride_bytes_hwc = src->info()->strides_in_bytes()[3];
+
+    execute_window_loop(win, [&](const Coordinates & id)
+    {
+        const int32_t offset     = *reinterpret_cast<const int32_t *>(offsets->ptr_to_element(Coordinates(id.y(), id.z()))) * in_stride_c;
+        const auto    in_hi      = static_cast<int>(align_corners ? utils::rounding::round_half_away_from_zero((id.z() + sampling_offset) * hr) : std::floor((id.z() + sampling_offset) * hr));
+        const int     offset_row = in_hi * in_stride_wc;
+        const auto    in_ptr     = reinterpret_cast<const float16_t *>(in_ptr_start + in_stride_bytes_hwc * id[3]);
+        const auto    out_ptr    = reinterpret_cast<float16_t *>(out.ptr());
+
+        // Compute S elements per iteration
+        int      x  = window_start_x;
+        svbool_t pg = svwhilelt_b16(x, window_end_x);
+        do
+        {
+            // Store results
+            svst1_f16(pg, out_ptr + x, svld1_f16(pg, in_ptr + offset + offset_row + x));
+
+            x += svcntw();
+            pg = svwhilelt_b16(x, window_end_x);
+        }
+        while(svptest_any(svptrue_b16(), pg));
+    },
+    out);
+}
+
+void fp16_sve_scale_bilinear(const ITensor *src, ITensor *dst, const ITensor *offsets, const ITensor *dx, const ITensor *dy,
+                             BorderMode border_mode, PixelValue constant_border_value, float sampling_offset,
+                             bool align_corners, const Window &window)
+{
+    // Compute the ratio between source height and destination height
+    const auto hr = scale_utils::calculate_resize_ratio(src->info()->dimension(2), dst->info()->dimension(2), align_corners);
+
+    Iterator  out(dst, window);
+    const int in_stride_c  = src->info()->dimension(0) + src->info()->padding().left + src->info()->padding().right;
+    const int in_dim_w     = src->info()->dimension(1);
+    const int in_dim_h     = src->info()->dimension(2);
+    const int in_stride_wc = in_stride_c * (in_dim_w + src->info()->padding().top + src->info()->padding().bottom);
+
+    // Don't increment in Y and Z 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(Window::DimY, Window::Dimension(0, 0, 0));
+    win_in.set(Window::DimZ, Window::Dimension(0, 0, 0));
+    Iterator in(src, win_in);
+
+    if(border_mode == BorderMode::CONSTANT)
+    {
+        using ConstType = typename std::conditional<std::is_same<float16_t, float16_t>::value, half, float16_t>::type;
+
+        const float16_t const_border_value = static_cast<float16_t>(constant_border_value.get<ConstType>());
+        execute_window_loop(window, [&](const Coordinates & id)
+        {
+            const auto       offset = *reinterpret_cast<const int32_t *>(offsets->ptr_to_element(Coordinates(id.y(), id.z())));
+            const auto       dx_val = *reinterpret_cast<const float *>(dx->ptr_to_element(Coordinates(id.y(), id.z())));
+            const auto       dy_val = *reinterpret_cast<const float *>(dy->ptr_to_element(Coordinates(id.y(), id.z())));
+            const int32_t    in_hi  = std::floor((id.z() + sampling_offset) * hr - sampling_offset);
+            const float16_t *in_ptr = reinterpret_cast<const float16_t *>(in.ptr()) + offset * in_stride_c + in_hi * in_stride_wc;
+
+            const auto a00 = (0 <= offset && offset < in_dim_w && 0 <= in_hi && in_hi < in_dim_h) ? *in_ptr : const_border_value;
+            const auto a01 = (-1 <= offset && offset < in_dim_w - 1 && 0 <= in_hi && in_hi < in_dim_h) ? *(in_ptr + in_stride_c) : const_border_value;
+            const auto a10 = (0 <= offset && offset < in_dim_w && -1 <= in_hi && in_hi < in_dim_h - 1) ? *(in_ptr + in_stride_wc) : const_border_value;
+            const auto a11 = (-1 <= offset && offset < in_dim_w - 1 && -1 <= in_hi && in_hi < in_dim_h - 1) ? *(in_ptr + in_stride_c + in_stride_wc) : const_border_value;
+
+            *reinterpret_cast<float16_t *>(out.ptr()) = static_cast<float16_t>(scale_helpers::delta_bilinear(a00, a01, a10, a11, dx_val, dy_val));
+        },
+        in, out);
+    }
+    else if(border_mode == BorderMode::REPLICATE)
+    {
+        execute_window_loop(window, [&](const Coordinates & id)
+        {
+            const auto offset = *reinterpret_cast<const int32_t *>(offsets->ptr_to_element(Coordinates(id.y(), id.z())));
+            const auto dx_val = *reinterpret_cast<const float *>(dx->ptr_to_element(Coordinates(id.y(), id.z())));
+            const auto dy_val = *reinterpret_cast<const float *>(dy->ptr_to_element(Coordinates(id.y(), id.z())));
+            const int  in_hi  = std::floor((id.z() + sampling_offset) * hr - sampling_offset);
+
+            auto clamped_w  = utility::clamp<int>(offset, 0, in_dim_w - 1);
+            auto clamped_w1 = utility::clamp<int>(offset + 1, 0, in_dim_w - 1);
+            auto clamped_h  = utility::clamp<int>(in_hi, 0, in_dim_h - 1);
+            auto clamped_h1 = utility::clamp<int>(in_hi + 1, 0, in_dim_h - 1);
+
+            const auto a00 = *(reinterpret_cast<const float16_t *>(in.ptr()) + clamped_w * in_stride_c + clamped_h * in_stride_wc);
+            const auto a01 = *(reinterpret_cast<const float16_t *>(in.ptr()) + clamped_w1 * in_stride_c + clamped_h * in_stride_wc);
+            const auto a10 = *(reinterpret_cast<const float16_t *>(in.ptr()) + clamped_w * in_stride_c + clamped_h1 * in_stride_wc);
+            const auto a11 = *(reinterpret_cast<const float16_t *>(in.ptr()) + clamped_w1 * in_stride_c + clamped_h1 * in_stride_wc);
+
+            *reinterpret_cast<float16_t *>(out.ptr()) = static_cast<float16_t>(scale_helpers::delta_bilinear(a00, a01, a10, a11, dx_val, dy_val));
+        },
+        in, out);
+    }
+    else
+    {
+        ARM_COMPUTE_ERROR("Not implemented");
+    }
+}
+}
+namespace cpu
+{
+void fp16_sve_scale(const ITensor *src, ITensor *dst, const ITensor *offsets, const ITensor *dx, const ITensor *dy,
+                    InterpolationPolicy policy, BorderMode border_mode, PixelValue constant_border_value, float sampling_offset,
+                    bool align_corners, const Window &window)
+{
+    if(policy == InterpolationPolicy::BILINEAR)
+    {
+        fp16_sve_scale_bilinear(src, dst, offsets, dx, dy, border_mode, constant_border_value, sampling_offset, align_corners, window);
+    }
+    else if(policy == InterpolationPolicy::NEAREST_NEIGHBOR)
+    {
+        fp16_sve_scale_nearest(src, dst, offsets, sampling_offset, align_corners, window);
+    }
+}
+} // namespace cpu
+} // namespace arm_compute
+
+#endif // __ARM_FEATURE_SVE
\ No newline at end of file
diff --git a/src/core/NEON/kernels/scale/impl/SVE/fp32.cpp b/src/core/NEON/kernels/scale/impl/SVE/fp32.cpp
new file mode 100644
index 0000000000..abb4faa6c5
--- /dev/null
+++ b/src/core/NEON/kernels/scale/impl/SVE/fp32.cpp
@@ -0,0 +1,175 @@
+/*
+ * 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 "arm_compute/core/Helpers.h"
+#include "arm_compute/core/ITensorPack.h"
+#include "arm_compute/core/Window.h"
+#include "src/core/NEON/NEMath.h"
+#include "src/core/NEON/wrapper/wrapper.h"
+#include "src/core/common/Validate.h"
+#include "src/core/helpers/ScaleHelpers.h"
+#include "src/core/utils/ScaleUtils.h"
+#include "support/Rounding.h"
+
+#include <cmath>
+#include <cstddef>
+
+#if defined(__ARM_FEATURE_SVE)
+#include <arm_sve.h>
+
+namespace arm_compute
+{
+namespace
+{
+void fp32_sve_scale_nearest(const ITensor *src, ITensor *dst, const ITensor *offsets,
+                            float sampling_offset, bool align_corners, const Window &window)
+{
+    const size_t in_stride_c  = src->info()->dimension(0) + src->info()->padding().left + src->info()->padding().right;
+    const size_t in_stride_w  = src->info()->dimension(1) + src->info()->padding().top + src->info()->padding().bottom;
+    const size_t in_stride_wc = in_stride_w * in_stride_c;
+    const size_t in_dim_h     = src->info()->dimension(2);
+
+    // Compute the ratio between source height and destination height
+    const auto hr             = scale_utils::calculate_resize_ratio(in_dim_h, dst->info()->dimension(2), align_corners);
+    const auto window_start_x = static_cast<int32_t>(window.x().start());
+    const auto window_end_x   = static_cast<int32_t>(window.x().end());
+
+    Window win(window);
+    win.set(Window::DimX, Window::Dimension(0, 1, 1));
+    Iterator out(dst, win);
+
+    const uint8_t     *in_ptr_start        = src->buffer() + src->info()->offset_first_element_in_bytes();
+    const unsigned int in_stride_bytes_hwc = src->info()->strides_in_bytes()[3];
+
+    execute_window_loop(win, [&](const Coordinates & id)
+    {
+        const int32_t offset     = *reinterpret_cast<const int32_t *>(offsets->ptr_to_element(Coordinates(id.y(), id.z()))) * in_stride_c;
+        const auto    in_hi      = static_cast<int>(align_corners ? utils::rounding::round_half_away_from_zero((id.z() + sampling_offset) * hr) : std::floor((id.z() + sampling_offset) * hr));
+        const int     offset_row = in_hi * in_stride_wc;
+        const auto    in_ptr     = reinterpret_cast<const float *>(in_ptr_start + in_stride_bytes_hwc * id[3]);
+        const auto    out_ptr    = reinterpret_cast<float *>(out.ptr());
+
+        // Compute S elements per iteration
+        int      x  = window_start_x;
+        svbool_t pg = svwhilelt_b32(x, window_end_x);
+        do
+        {
+            // Store results
+            svst1_f32(pg, out_ptr + x, svld1_f32(pg, in_ptr + offset + offset_row + x));
+
+            x += svcntw();
+            pg = svwhilelt_b32(x, window_end_x);
+        }
+        while(svptest_any(svptrue_b32(), pg));
+    },
+    out);
+}
+
+void fp32_sve_scale_bilinear(const ITensor *src, ITensor *dst, const ITensor *offsets, const ITensor *dx, const ITensor *dy,
+                             BorderMode border_mode, PixelValue constant_border_value, float sampling_offset,
+                             bool align_corners, const Window &window)
+{
+    // Compute the ratio between source height and destination height
+    const auto hr = scale_utils::calculate_resize_ratio(src->info()->dimension(2), dst->info()->dimension(2), align_corners);
+
+    Iterator  out(dst, window);
+    const int in_stride_c  = src->info()->dimension(0) + src->info()->padding().left + src->info()->padding().right;
+    const int in_dim_w     = src->info()->dimension(1);
+    const int in_dim_h     = src->info()->dimension(2);
+    const int in_stride_wc = in_stride_c * (in_dim_w + src->info()->padding().top + src->info()->padding().bottom);
+
+    // Don't increment in Y and Z 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(Window::DimY, Window::Dimension(0, 0, 0));
+    win_in.set(Window::DimZ, Window::Dimension(0, 0, 0));
+    Iterator in(src, win_in);
+
+    if(border_mode == BorderMode::CONSTANT)
+    {
+        const float const_border_value = static_cast<float>(constant_border_value.get<float>());
+        execute_window_loop(window, [&](const Coordinates & id)
+        {
+            const auto    offset = *reinterpret_cast<const int32_t *>(offsets->ptr_to_element(Coordinates(id.y(), id.z())));
+            const auto    dx_val = *reinterpret_cast<const float *>(dx->ptr_to_element(Coordinates(id.y(), id.z())));
+            const auto    dy_val = *reinterpret_cast<const float *>(dy->ptr_to_element(Coordinates(id.y(), id.z())));
+            const int32_t in_hi  = std::floor((id.z() + sampling_offset) * hr - sampling_offset);
+            const float *in_ptr = reinterpret_cast<const float *>(in.ptr()) + offset * in_stride_c + in_hi * in_stride_wc;
+
+            const auto a00 = (0 <= offset && offset < in_dim_w && 0 <= in_hi && in_hi < in_dim_h) ? *in_ptr : const_border_value;
+            const auto a01 = (-1 <= offset && offset < in_dim_w - 1 && 0 <= in_hi && in_hi < in_dim_h) ? *(in_ptr + in_stride_c) : const_border_value;
+            const auto a10 = (0 <= offset && offset < in_dim_w && -1 <= in_hi && in_hi < in_dim_h - 1) ? *(in_ptr + in_stride_wc) : const_border_value;
+            const auto a11 = (-1 <= offset && offset < in_dim_w - 1 && -1 <= in_hi && in_hi < in_dim_h - 1) ? *(in_ptr + in_stride_c + in_stride_wc) : const_border_value;
+
+            *reinterpret_cast<float *>(out.ptr()) = static_cast<float>(scale_helpers::delta_bilinear(a00, a01, a10, a11, dx_val, dy_val));
+        },
+        in, out);
+    }
+    else if(border_mode == BorderMode::REPLICATE)
+    {
+        execute_window_loop(window, [&](const Coordinates & id)
+        {
+            const auto offset = *reinterpret_cast<const int32_t *>(offsets->ptr_to_element(Coordinates(id.y(), id.z())));
+            const auto dx_val = *reinterpret_cast<const float *>(dx->ptr_to_element(Coordinates(id.y(), id.z())));
+            const auto dy_val = *reinterpret_cast<const float *>(dy->ptr_to_element(Coordinates(id.y(), id.z())));
+            const int  in_hi  = std::floor((id.z() + sampling_offset) * hr - sampling_offset);
+
+            auto clamped_w  = utility::clamp<int>(offset, 0, in_dim_w - 1);
+            auto clamped_w1 = utility::clamp<int>(offset + 1, 0, in_dim_w - 1);
+            auto clamped_h  = utility::clamp<int>(in_hi, 0, in_dim_h - 1);
+            auto clamped_h1 = utility::clamp<int>(in_hi + 1, 0, in_dim_h - 1);
+
+            const auto a00 = *(reinterpret_cast<const float *>(in.ptr()) + clamped_w * in_stride_c + clamped_h * in_stride_wc);
+            const auto a01 = *(reinterpret_cast<const float *>(in.ptr()) + clamped_w1 * in_stride_c + clamped_h * in_stride_wc);
+            const auto a10 = *(reinterpret_cast<const float *>(in.ptr()) + clamped_w * in_stride_c + clamped_h1 * in_stride_wc);
+            const auto a11 = *(reinterpret_cast<const float *>(in.ptr()) + clamped_w1 * in_stride_c + clamped_h1 * in_stride_wc);
+
+            *reinterpret_cast<float *>(out.ptr()) = static_cast<float>(scale_helpers::delta_bilinear(a00, a01, a10, a11, dx_val, dy_val));
+        },
+        in, out);
+    }
+    else
+    {
+        ARM_COMPUTE_ERROR("Not implemented");
+    }
+}
+}
+namespace cpu
+{
+void fp32_sve_scale(const ITensor *src, ITensor *dst, const ITensor *offsets, const ITensor *dx, const ITensor *dy,
+                    InterpolationPolicy policy, BorderMode border_mode, PixelValue constant_border_value, float sampling_offset,
+                    bool align_corners, const Window &window)
+{
+    if(policy == InterpolationPolicy::BILINEAR)
+    {
+        fp32_sve_scale_bilinear(src, dst, offsets, dx, dy, border_mode, constant_border_value, sampling_offset, align_corners, window);
+    }
+    else if(policy == InterpolationPolicy::NEAREST_NEIGHBOR)
+    {
+        fp32_sve_scale_nearest(src, dst, offsets, sampling_offset, align_corners, window);
+    }
+}
+} // namespace cpu
+} // namespace arm_compute
+
+#endif // __ARM_FEATURE_SVE
\ No newline at end of file
diff --git a/src/core/NEON/kernels/scale/impl/SVE/integer.cpp b/src/core/NEON/kernels/scale/impl/SVE/integer.cpp
new file mode 100644
index 0000000000..5f5263c5d3
--- /dev/null
+++ b/src/core/NEON/kernels/scale/impl/SVE/integer.cpp
@@ -0,0 +1,302 @@
+/*
+ * 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 "arm_compute/core/Helpers.h"
+#include "arm_compute/core/ITensorPack.h"
+#include "arm_compute/core/Window.h"
+#include "src/core/NEON/NEMath.h"
+#include "src/core/NEON/wrapper/wrapper.h"
+#include "src/core/common/Validate.h"
+#include "src/core/helpers/ScaleHelpers.h"
+#include "src/core/utils/ScaleUtils.h"
+#include "support/Rounding.h"
+
+#include <cmath>
+#include <cstddef>
+
+#if defined(__ARM_FEATURE_SVE)
+#include <arm_sve.h>
+
+namespace arm_compute
+{
+namespace
+{
+void u8_sve_scale_nearest(const ITensor *src, ITensor *dst, const ITensor *offsets,
+                          float sampling_offset, bool align_corners, const Window &window)
+{
+    const size_t in_stride_c  = src->info()->dimension(0) + src->info()->padding().left + src->info()->padding().right;
+    const size_t in_stride_w  = src->info()->dimension(1) + src->info()->padding().top + src->info()->padding().bottom;
+    const size_t in_stride_wc = in_stride_w * in_stride_c;
+    const size_t in_dim_h     = src->info()->dimension(2);
+
+    // Compute the ratio between source height and destination height
+    const auto hr             = scale_utils::calculate_resize_ratio(in_dim_h, dst->info()->dimension(2), align_corners);
+    const auto window_start_x = static_cast<int32_t>(window.x().start());
+    const auto window_end_x   = static_cast<int32_t>(window.x().end());
+
+    Window win(window);
+    win.set(Window::DimX, Window::Dimension(0, 1, 1));
+    Iterator out(dst, win);
+
+    const uint8_t     *in_ptr_start        = src->buffer() + src->info()->offset_first_element_in_bytes();
+    const unsigned int in_stride_bytes_hwc = src->info()->strides_in_bytes()[3];
+
+    execute_window_loop(win, [&](const Coordinates & id)
+    {
+        const int32_t offset     = *reinterpret_cast<const int32_t *>(offsets->ptr_to_element(Coordinates(id.y(), id.z()))) * in_stride_c;
+        const auto    in_hi      = static_cast<int>(align_corners ? utils::rounding::round_half_away_from_zero((id.z() + sampling_offset) * hr) : std::floor((id.z() + sampling_offset) * hr));
+        const int     offset_row = in_hi * in_stride_wc;
+        const auto    in_ptr     = reinterpret_cast<const uint8_t *>(in_ptr_start + in_stride_bytes_hwc * id[3]);
+        const auto    out_ptr    = reinterpret_cast<uint8_t *>(out.ptr());
+
+        // Compute S elements per iteration
+        int      x  = window_start_x;
+        svbool_t pg = svwhilelt_b8(x, window_end_x);
+        do
+        {
+            // Store results
+            svst1_u8(pg, out_ptr + x, svld1_u8(pg, in_ptr + offset + offset_row + x));
+
+            x += svcntw();
+            pg = svwhilelt_b8(x, window_end_x);
+        }
+        while(svptest_any(svptrue_b8(), pg));
+    },
+    out);
+}
+
+void u8_sve_scale_bilinear(const ITensor *src, ITensor *dst, const ITensor *offsets, const ITensor *dx, const ITensor *dy,
+                           BorderMode border_mode, PixelValue constant_border_value, float sampling_offset,
+                           bool align_corners, const Window &window)
+{
+    // Compute the ratio between source height and destination height
+    const auto hr = scale_utils::calculate_resize_ratio(src->info()->dimension(2), dst->info()->dimension(2), align_corners);
+
+    Iterator  out(dst, window);
+    const int in_stride_c  = src->info()->dimension(0) + src->info()->padding().left + src->info()->padding().right;
+    const int in_dim_w     = src->info()->dimension(1);
+    const int in_dim_h     = src->info()->dimension(2);
+    const int in_stride_wc = in_stride_c * (in_dim_w + src->info()->padding().top + src->info()->padding().bottom);
+
+    // Don't increment in Y and Z 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(Window::DimY, Window::Dimension(0, 0, 0));
+    win_in.set(Window::DimZ, Window::Dimension(0, 0, 0));
+    Iterator in(src, win_in);
+
+    if(border_mode == BorderMode::CONSTANT)
+    {
+        const uint8_t const_border_value = static_cast<uint8_t>(constant_border_value.get<uint8_t>());
+        execute_window_loop(window, [&](const Coordinates & id)
+        {
+            const auto     offset = *reinterpret_cast<const int32_t *>(offsets->ptr_to_element(Coordinates(id.y(), id.z())));
+            const auto     dx_val = *reinterpret_cast<const float *>(dx->ptr_to_element(Coordinates(id.y(), id.z())));
+            const auto     dy_val = *reinterpret_cast<const float *>(dy->ptr_to_element(Coordinates(id.y(), id.z())));
+            const int32_t  in_hi  = std::floor((id.z() + sampling_offset) * hr - sampling_offset);
+            const uint8_t *in_ptr = reinterpret_cast<const uint8_t *>(in.ptr()) + offset * in_stride_c + in_hi * in_stride_wc;
+
+            const auto a00 = (0 <= offset && offset < in_dim_w && 0 <= in_hi && in_hi < in_dim_h) ? *in_ptr : const_border_value;
+            const auto a01 = (-1 <= offset && offset < in_dim_w - 1 && 0 <= in_hi && in_hi < in_dim_h) ? *(in_ptr + in_stride_c) : const_border_value;
+            const auto a10 = (0 <= offset && offset < in_dim_w && -1 <= in_hi && in_hi < in_dim_h - 1) ? *(in_ptr + in_stride_wc) : const_border_value;
+            const auto a11 = (-1 <= offset && offset < in_dim_w - 1 && -1 <= in_hi && in_hi < in_dim_h - 1) ? *(in_ptr + in_stride_c + in_stride_wc) : const_border_value;
+
+            *reinterpret_cast<uint8_t *>(out.ptr()) = static_cast<uint8_t>(scale_helpers::delta_bilinear(a00, a01, a10, a11, dx_val, dy_val));
+        },
+        in, out);
+    }
+    else if(border_mode == BorderMode::REPLICATE)
+    {
+        execute_window_loop(window, [&](const Coordinates & id)
+        {
+            const auto offset = *reinterpret_cast<const int32_t *>(offsets->ptr_to_element(Coordinates(id.y(), id.z())));
+            const auto dx_val = *reinterpret_cast<const float *>(dx->ptr_to_element(Coordinates(id.y(), id.z())));
+            const auto dy_val = *reinterpret_cast<const float *>(dy->ptr_to_element(Coordinates(id.y(), id.z())));
+            const int  in_hi  = std::floor((id.z() + sampling_offset) * hr - sampling_offset);
+
+            auto clamped_w  = utility::clamp<int>(offset, 0, in_dim_w - 1);
+            auto clamped_w1 = utility::clamp<int>(offset + 1, 0, in_dim_w - 1);
+            auto clamped_h  = utility::clamp<int>(in_hi, 0, in_dim_h - 1);
+            auto clamped_h1 = utility::clamp<int>(in_hi + 1, 0, in_dim_h - 1);
+
+            const auto a00 = *(reinterpret_cast<const uint8_t *>(in.ptr()) + clamped_w * in_stride_c + clamped_h * in_stride_wc);
+            const auto a01 = *(reinterpret_cast<const uint8_t *>(in.ptr()) + clamped_w1 * in_stride_c + clamped_h * in_stride_wc);
+            const auto a10 = *(reinterpret_cast<const uint8_t *>(in.ptr()) + clamped_w * in_stride_c + clamped_h1 * in_stride_wc);
+            const auto a11 = *(reinterpret_cast<const uint8_t *>(in.ptr()) + clamped_w1 * in_stride_c + clamped_h1 * in_stride_wc);
+
+            *reinterpret_cast<uint8_t *>(out.ptr()) = static_cast<uint8_t>(scale_helpers::delta_bilinear(a00, a01, a10, a11, dx_val, dy_val));
+        },
+        in, out);
+    }
+    else
+    {
+        ARM_COMPUTE_ERROR("Not implemented");
+    }
+}
+
+void s16_sve_scale_nearest(const ITensor *src, ITensor *dst, const ITensor *offsets,
+                           float sampling_offset, bool align_corners, const Window &window)
+{
+    const size_t in_stride_c  = src->info()->dimension(0) + src->info()->padding().left + src->info()->padding().right;
+    const size_t in_stride_w  = src->info()->dimension(1) + src->info()->padding().top + src->info()->padding().bottom;
+    const size_t in_stride_wc = in_stride_w * in_stride_c;
+    const size_t in_dim_h     = src->info()->dimension(2);
+
+    // Compute the ratio between source height and destination height
+    const auto hr             = scale_utils::calculate_resize_ratio(in_dim_h, dst->info()->dimension(2), align_corners);
+    const auto window_start_x = static_cast<int32_t>(window.x().start());
+    const auto window_end_x   = static_cast<int32_t>(window.x().end());
+
+    Window win(window);
+    win.set(Window::DimX, Window::Dimension(0, 1, 1));
+    Iterator out(dst, win);
+
+    const uint8_t     *in_ptr_start        = src->buffer() + src->info()->offset_first_element_in_bytes();
+    const unsigned int in_stride_bytes_hwc = src->info()->strides_in_bytes()[3];
+
+    execute_window_loop(win, [&](const Coordinates & id)
+    {
+        const int32_t offset     = *reinterpret_cast<const int32_t *>(offsets->ptr_to_element(Coordinates(id.y(), id.z()))) * in_stride_c;
+        const auto    in_hi      = static_cast<int>(align_corners ? utils::rounding::round_half_away_from_zero((id.z() + sampling_offset) * hr) : std::floor((id.z() + sampling_offset) * hr));
+        const int     offset_row = in_hi * in_stride_wc;
+        const auto    in_ptr     = reinterpret_cast<const int16_t *>(in_ptr_start + in_stride_bytes_hwc * id[3]);
+        const auto    out_ptr    = reinterpret_cast<int16_t *>(out.ptr());
+
+        // Compute S elements per iteration
+        int      x  = window_start_x;
+        svbool_t pg = svwhilelt_b16(x, window_end_x);
+        do
+        {
+            // Store results
+            svst1_s16(pg, out_ptr + x, svld1_s16(pg, in_ptr + offset + offset_row + x));
+
+            x += svcntw();
+            pg = svwhilelt_b16(x, window_end_x);
+        }
+        while(svptest_any(svptrue_b16(), pg));
+    },
+    out);
+}
+
+void s16_sve_scale_bilinear(const ITensor *src, ITensor *dst, const ITensor *offsets, const ITensor *dx, const ITensor *dy,
+                            BorderMode border_mode, PixelValue constant_border_value, float sampling_offset,
+                            bool align_corners, const Window &window)
+{
+    // Compute the ratio between source height and destination height
+    const auto hr = scale_utils::calculate_resize_ratio(src->info()->dimension(2), dst->info()->dimension(2), align_corners);
+
+    Iterator  out(dst, window);
+    const int in_stride_c  = src->info()->dimension(0) + src->info()->padding().left + src->info()->padding().right;
+    const int in_dim_w     = src->info()->dimension(1);
+    const int in_dim_h     = src->info()->dimension(2);
+    const int in_stride_wc = in_stride_c * (in_dim_w + src->info()->padding().top + src->info()->padding().bottom);
+
+    // Don't increment in Y and Z 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(Window::DimY, Window::Dimension(0, 0, 0));
+    win_in.set(Window::DimZ, Window::Dimension(0, 0, 0));
+    Iterator in(src, win_in);
+
+    if(border_mode == BorderMode::CONSTANT)
+    {
+        const int16_t const_border_value = static_cast<int16_t>(constant_border_value.get<int16_t>());
+        execute_window_loop(window, [&](const Coordinates & id)
+        {
+            const auto     offset = *reinterpret_cast<const int32_t *>(offsets->ptr_to_element(Coordinates(id.y(), id.z())));
+            const auto     dx_val = *reinterpret_cast<const float *>(dx->ptr_to_element(Coordinates(id.y(), id.z())));
+            const auto     dy_val = *reinterpret_cast<const float *>(dy->ptr_to_element(Coordinates(id.y(), id.z())));
+            const int32_t  in_hi  = std::floor((id.z() + sampling_offset) * hr - sampling_offset);
+            const int16_t *in_ptr = reinterpret_cast<const int16_t *>(in.ptr()) + offset * in_stride_c + in_hi * in_stride_wc;
+
+            const auto a00 = (0 <= offset && offset < in_dim_w && 0 <= in_hi && in_hi < in_dim_h) ? *in_ptr : const_border_value;
+            const auto a01 = (-1 <= offset && offset < in_dim_w - 1 && 0 <= in_hi && in_hi < in_dim_h) ? *(in_ptr + in_stride_c) : const_border_value;
+            const auto a10 = (0 <= offset && offset < in_dim_w && -1 <= in_hi && in_hi < in_dim_h - 1) ? *(in_ptr + in_stride_wc) : const_border_value;
+            const auto a11 = (-1 <= offset && offset < in_dim_w - 1 && -1 <= in_hi && in_hi < in_dim_h - 1) ? *(in_ptr + in_stride_c + in_stride_wc) : const_border_value;
+
+            *reinterpret_cast<int16_t *>(out.ptr()) = static_cast<int16_t>(scale_helpers::delta_bilinear(a00, a01, a10, a11, dx_val, dy_val));
+        },
+        in, out);
+    }
+    else if(border_mode == BorderMode::REPLICATE)
+    {
+        execute_window_loop(window, [&](const Coordinates & id)
+        {
+            const auto offset = *reinterpret_cast<const int32_t *>(offsets->ptr_to_element(Coordinates(id.y(), id.z())));
+            const auto dx_val = *reinterpret_cast<const float *>(dx->ptr_to_element(Coordinates(id.y(), id.z())));
+            const auto dy_val = *reinterpret_cast<const float *>(dy->ptr_to_element(Coordinates(id.y(), id.z())));
+            const int  in_hi  = std::floor((id.z() + sampling_offset) * hr - sampling_offset);
+
+            auto clamped_w  = utility::clamp<int>(offset, 0, in_dim_w - 1);
+            auto clamped_w1 = utility::clamp<int>(offset + 1, 0, in_dim_w - 1);
+            auto clamped_h  = utility::clamp<int>(in_hi, 0, in_dim_h - 1);
+            auto clamped_h1 = utility::clamp<int>(in_hi + 1, 0, in_dim_h - 1);
+
+            const auto a00 = *(reinterpret_cast<const int16_t *>(in.ptr()) + clamped_w * in_stride_c + clamped_h * in_stride_wc);
+            const auto a01 = *(reinterpret_cast<const int16_t *>(in.ptr()) + clamped_w1 * in_stride_c + clamped_h * in_stride_wc);
+            const auto a10 = *(reinterpret_cast<const int16_t *>(in.ptr()) + clamped_w * in_stride_c + clamped_h1 * in_stride_wc);
+            const auto a11 = *(reinterpret_cast<const int16_t *>(in.ptr()) + clamped_w1 * in_stride_c + clamped_h1 * in_stride_wc);
+
+            *reinterpret_cast<int16_t *>(out.ptr()) = static_cast<int16_t>(scale_helpers::delta_bilinear(a00, a01, a10, a11, dx_val, dy_val));
+        },
+        in, out);
+    }
+    else
+    {
+        ARM_COMPUTE_ERROR("Not implemented");
+    }
+}
+}
+namespace cpu
+{
+void u8_sve_scale(const ITensor *src, ITensor *dst, const ITensor *offsets, const ITensor *dx, const ITensor *dy,
+                  InterpolationPolicy policy, BorderMode border_mode, PixelValue constant_border_value, float sampling_offset,
+                  bool align_corners, const Window &window)
+{
+    if(policy == InterpolationPolicy::BILINEAR)
+    {
+        u8_sve_scale_bilinear(src, dst, offsets, dx, dy, border_mode, constant_border_value, sampling_offset, align_corners, window);
+    }
+    else if(policy == InterpolationPolicy::NEAREST_NEIGHBOR)
+    {
+        u8_sve_scale_nearest(src, dst, offsets, sampling_offset, align_corners, window);
+    }
+}
+
+void s16_sve_scale(const ITensor *src, ITensor *dst, const ITensor *offsets, const ITensor *dx, const ITensor *dy,
+                   InterpolationPolicy policy, BorderMode border_mode, PixelValue constant_border_value, float sampling_offset,
+                   bool align_corners, const Window &window)
+{
+    if(policy == InterpolationPolicy::BILINEAR)
+    {
+        s16_sve_scale_bilinear(src, dst, offsets, dx, dy, border_mode, constant_border_value, sampling_offset, align_corners, window);
+    }
+    else if(policy == InterpolationPolicy::NEAREST_NEIGHBOR)
+    {
+        s16_sve_scale_nearest(src, dst, offsets, sampling_offset, align_corners, window);
+    }
+}
+} // namespace cpu
+} // namespace arm_compute
+
+#endif // __ARM_FEATURE_SVE
\ No newline at end of file
diff --git a/src/core/NEON/kernels/scale/impl/SVE/qasymm8.cpp b/src/core/NEON/kernels/scale/impl/SVE/qasymm8.cpp
new file mode 100644
index 0000000000..fc65ff44f0
--- /dev/null
+++ b/src/core/NEON/kernels/scale/impl/SVE/qasymm8.cpp
@@ -0,0 +1,210 @@
+/*
+ * 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 "arm_compute/core/Helpers.h"
+#include "arm_compute/core/ITensorPack.h"
+#include "arm_compute/core/Window.h"
+#include "src/core/NEON/NEMath.h"
+#include "src/core/NEON/wrapper/wrapper.h"
+#include "src/core/common/Validate.h"
+#include "src/core/helpers/ScaleHelpers.h"
+#include "src/core/helpers/ScaleHelpers.h"
+#include "src/core/utils/ScaleUtils.h"
+#include "support/Rounding.h"
+
+#include <cmath>
+#include <cstddef>
+
+#if defined(__ARM_FEATURE_SVE)
+#include <arm_sve.h>
+
+namespace arm_compute
+{
+namespace
+{
+void qasymm8_sve_scale_nearest(const ITensor *src, ITensor *dst, const ITensor *offsets,
+                               float sampling_offset, bool align_corners, const Window &window)
+{
+    const size_t in_stride_c  = src->info()->dimension(0) + src->info()->padding().left + src->info()->padding().right;
+    const size_t in_stride_w  = src->info()->dimension(1) + src->info()->padding().top + src->info()->padding().bottom;
+    const size_t in_stride_wc = in_stride_w * in_stride_c;
+    const size_t in_dim_h     = src->info()->dimension(2);
+
+    // Compute the ratio between source height and destination height
+    const auto hr             = scale_utils::calculate_resize_ratio(in_dim_h, dst->info()->dimension(2), align_corners);
+    const auto window_start_x = static_cast<int32_t>(window.x().start());
+    const auto window_end_x   = static_cast<int32_t>(window.x().end());
+
+    Window win(window);
+    win.set(Window::DimX, Window::Dimension(0, 1, 1));
+    Iterator out(dst, win);
+
+    const uint8_t     *in_ptr_start        = src->buffer() + src->info()->offset_first_element_in_bytes();
+    const unsigned int in_stride_bytes_hwc = src->info()->strides_in_bytes()[3];
+
+    execute_window_loop(win, [&](const Coordinates & id)
+    {
+        const int32_t offset     = *reinterpret_cast<const int32_t *>(offsets->ptr_to_element(Coordinates(id.y(), id.z()))) * in_stride_c;
+        const auto    in_hi      = static_cast<int>(align_corners ? utils::rounding::round_half_away_from_zero((id.z() + sampling_offset) * hr) : std::floor((id.z() + sampling_offset) * hr));
+        const int     offset_row = in_hi * in_stride_wc;
+        const auto    in_ptr     = reinterpret_cast<const uint8_t *>(in_ptr_start + in_stride_bytes_hwc * id[3]);
+        const auto    out_ptr    = reinterpret_cast<uint8_t *>(out.ptr());
+
+        // Compute S elements per iteration
+        int      x  = window_start_x;
+        svbool_t pg = svwhilelt_b8(x, window_end_x);
+        do
+        {
+            // Store results
+            svst1_u8(pg, out_ptr + x, svld1_u8(pg, in_ptr + offset + offset_row + x));
+
+            x += svcntw();
+            pg = svwhilelt_b8(x, window_end_x);
+        }
+        while(svptest_any(svptrue_b8(), pg));
+    },
+    out);
+}
+
+void qasymm8_sve_scale_bilinear(const ITensor *src, ITensor *dst, const ITensor *offsets, const ITensor *dx, const ITensor *dy,
+                                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);
+
+    // 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);
+    Window     win_off;
+    win_off.set(Window::DimX, Window::Dimension(0, 0, 0));
+    win_off.set(Window::DimY, Window::Dimension(0, 0, 0));
+
+    // 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));
+
+    for(size_t d = Window::DimZ; d < offsets->info()->num_dimensions(); ++d)
+    {
+        win_off.set(d, Window::Dimension(0, 0, 0));
+    }
+
+    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 UniformQuantizationInfo iq_info = src->info()->quantization_info().uniform();
+    const UniformQuantizationInfo oq_info = dst->info()->quantization_info().uniform();
+
+    if(border_mode == BorderMode::CONSTANT)
+    {
+        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 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) ?
+                             (*(pixel_row_ptr + index_w * stride_w + index_h * stride_h)) :
+                             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) * stride_w + index_h * stride_h)) :
+                             const_border_value;
+            const auto a10 = (0 <= index_w && index_w < in_dim_w && -1 <= index_h && index_h < in_dim_h - 1) ?
+                             (*(pixel_row_ptr + index_w * stride_w + (index_h + 1) * stride_h)) :
+                             const_border_value;
+            const auto a11 = (-1 <= index_w && index_w < in_dim_w - 1 && -1 <= index_h && index_h < in_dim_h - 1) ?
+                             (*(pixel_row_ptr + (index_w + 1) * stride_w + (index_h + 1) * stride_h)) :
+                             const_border_value;
+
+            const float inp00                       = Qasymm8QuantizationHelper<uint8_t>::dequantize(a00, iq_info);
+            const float inp01                       = Qasymm8QuantizationHelper<uint8_t>::dequantize(a01, iq_info);
+            const float inp10                       = Qasymm8QuantizationHelper<uint8_t>::dequantize(a10, iq_info);
+            const float inp11                       = Qasymm8QuantizationHelper<uint8_t>::dequantize(a11, iq_info);
+            *reinterpret_cast<uint8_t *>(out.ptr()) = Qasymm8QuantizationHelper<uint8_t>::quantize(scale_helpers::delta_bilinear(inp00, inp01, inp10, inp11, dx_val, dy_val), oq_info);
+        },
+        in, out);
+    }
+    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 uint8_t *>(in.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);
+            auto clamped_h  = utility::clamp<int>(index_h, 0, in_dim_h - 1);
+            auto clamped_h1 = utility::clamp<int>(index_h + 1, 0, in_dim_h - 1);
+
+            const auto a00 = *(pixel_row_ptr + clamped_w * stride_w + clamped_h * stride_h);
+            const auto a01 = *(pixel_row_ptr + clamped_w1 * stride_w + clamped_h * stride_h);
+            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<uint8_t>::dequantize(a00, iq_info);
+            const float inp01                       = Qasymm8QuantizationHelper<uint8_t>::dequantize(a01, iq_info);
+            const float inp10                       = Qasymm8QuantizationHelper<uint8_t>::dequantize(a10, iq_info);
+            const float inp11                       = Qasymm8QuantizationHelper<uint8_t>::dequantize(a11, iq_info);
+            *reinterpret_cast<uint8_t *>(out.ptr()) = Qasymm8QuantizationHelper<uint8_t>::quantize(scale_helpers::delta_bilinear(inp00, inp01, inp10, inp11, dx_val, dy_val), oq_info);
+        },
+        in, out);
+    }
+    else
+    {
+        ARM_COMPUTE_ERROR("Not implemented");
+    }
+}
+}
+namespace cpu
+{
+void qasymm8_sve_scale(const ITensor *src, ITensor *dst, const ITensor *offsets, const ITensor *dx, const ITensor *dy,
+                       InterpolationPolicy policy, BorderMode border_mode, PixelValue constant_border_value, float sampling_offset,
+                       bool align_corners, const Window &window)
+{
+    if(policy == InterpolationPolicy::BILINEAR)
+    {
+        qasymm8_sve_scale_bilinear(src, dst, offsets, dx, dy, border_mode, constant_border_value, sampling_offset, align_corners, window);
+    }
+    else if(policy == InterpolationPolicy::NEAREST_NEIGHBOR)
+    {
+        qasymm8_sve_scale_nearest(src, dst, offsets, sampling_offset, align_corners, window);
+    }
+}
+} // namespace cpu
+} // namespace arm_compute
+
+#endif // __ARM_FEATURE_SVE
\ No newline at end of file
diff --git a/src/core/NEON/kernels/scale/impl/SVE/qasymm8_signed.cpp b/src/core/NEON/kernels/scale/impl/SVE/qasymm8_signed.cpp
new file mode 100644
index 0000000000..676ca94fb0
--- /dev/null
+++ b/src/core/NEON/kernels/scale/impl/SVE/qasymm8_signed.cpp
@@ -0,0 +1,210 @@
+/*
+ * 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 "arm_compute/core/Helpers.h"
+#include "arm_compute/core/ITensorPack.h"
+#include "arm_compute/core/Window.h"
+#include "src/core/NEON/NEMath.h"
+#include "src/core/NEON/wrapper/wrapper.h"
+#include "src/core/common/Validate.h"
+#include "src/core/helpers/ScaleHelpers.h"
+#include "src/core/helpers/ScaleHelpers.h"
+#include "src/core/utils/ScaleUtils.h"
+#include "support/Rounding.h"
+
+#include <cmath>
+#include <cstddef>
+
+#if defined(__ARM_FEATURE_SVE)
+#include <arm_sve.h>
+
+namespace arm_compute
+{
+namespace
+{
+void qasymm8_signed_sve_scale_nearest(const ITensor *src, ITensor *dst, const ITensor *offsets,
+                                      float sampling_offset, bool align_corners, const Window &window)
+{
+    const size_t in_stride_c  = src->info()->dimension(0) + src->info()->padding().left + src->info()->padding().right;
+    const size_t in_stride_w  = src->info()->dimension(1) + src->info()->padding().top + src->info()->padding().bottom;
+    const size_t in_stride_wc = in_stride_w * in_stride_c;
+    const size_t in_dim_h     = src->info()->dimension(2);
+
+    // Compute the ratio between source height and destination height
+    const auto hr             = scale_utils::calculate_resize_ratio(in_dim_h, dst->info()->dimension(2), align_corners);
+    const auto window_start_x = static_cast<int32_t>(window.x().start());
+    const auto window_end_x   = static_cast<int32_t>(window.x().end());
+
+    Window win(window);
+    win.set(Window::DimX, Window::Dimension(0, 1, 1));
+    Iterator out(dst, win);
+
+    const uint8_t     *in_ptr_start        = src->buffer() + src->info()->offset_first_element_in_bytes();
+    const unsigned int in_stride_bytes_hwc = src->info()->strides_in_bytes()[3];
+
+    execute_window_loop(win, [&](const Coordinates & id)
+    {
+        const int32_t offset     = *reinterpret_cast<const int32_t *>(offsets->ptr_to_element(Coordinates(id.y(), id.z()))) * in_stride_c;
+        const auto    in_hi      = static_cast<int>(align_corners ? utils::rounding::round_half_away_from_zero((id.z() + sampling_offset) * hr) : std::floor((id.z() + sampling_offset) * hr));
+        const int     offset_row = in_hi * in_stride_wc;
+        const auto    in_ptr     = reinterpret_cast<const int8_t *>(in_ptr_start + in_stride_bytes_hwc * id[3]);
+        const auto    out_ptr    = reinterpret_cast<int8_t *>(out.ptr());
+
+        // Compute S elements per iteration
+        int      x  = window_start_x;
+        svbool_t pg = svwhilelt_b8(x, window_end_x);
+        do
+        {
+            // Store results
+            svst1_s8(pg, out_ptr + x, svld1_s8(pg, in_ptr + offset + offset_row + x));
+
+            x += svcntw();
+            pg = svwhilelt_b8(x, window_end_x);
+        }
+        while(svptest_any(svptrue_b8(), pg));
+    },
+    out);
+}
+
+void qasymm8_signed_sve_scale_bilinear(const ITensor *src, ITensor *dst, const ITensor *offsets, const ITensor *dx, const ITensor *dy,
+                                       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);
+
+    // 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);
+    Window     win_off;
+    win_off.set(Window::DimX, Window::Dimension(0, 0, 0));
+    win_off.set(Window::DimY, Window::Dimension(0, 0, 0));
+
+    // 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));
+
+    for(size_t d = Window::DimZ; d < offsets->info()->num_dimensions(); ++d)
+    {
+        win_off.set(d, Window::Dimension(0, 0, 0));
+    }
+
+    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 UniformQuantizationInfo iq_info = src->info()->quantization_info().uniform();
+    const UniformQuantizationInfo oq_info = dst->info()->quantization_info().uniform();
+
+    if(border_mode == BorderMode::CONSTANT)
+    {
+        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 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) ?
+                             (*(pixel_row_ptr + index_w * stride_w + index_h * stride_h)) :
+                             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) * stride_w + index_h * stride_h)) :
+                             const_border_value;
+            const auto a10 = (0 <= index_w && index_w < in_dim_w && -1 <= index_h && index_h < in_dim_h - 1) ?
+                             (*(pixel_row_ptr + index_w * stride_w + (index_h + 1) * stride_h)) :
+                             const_border_value;
+            const auto a11 = (-1 <= index_w && index_w < in_dim_w - 1 && -1 <= index_h && index_h < in_dim_h - 1) ?
+                             (*(pixel_row_ptr + (index_w + 1) * stride_w + (index_h + 1) * stride_h)) :
+                             const_border_value;
+
+            const float inp00                      = Qasymm8QuantizationHelper<int8_t>::dequantize(a00, iq_info);
+            const float inp01                      = Qasymm8QuantizationHelper<int8_t>::dequantize(a01, iq_info);
+            const float inp10                      = Qasymm8QuantizationHelper<int8_t>::dequantize(a10, iq_info);
+            const float inp11                      = Qasymm8QuantizationHelper<int8_t>::dequantize(a11, iq_info);
+            *reinterpret_cast<int8_t *>(out.ptr()) = Qasymm8QuantizationHelper<int8_t>::quantize(scale_helpers::delta_bilinear(inp00, inp01, inp10, inp11, dx_val, dy_val), oq_info);
+        },
+        in, out);
+    }
+    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 int8_t *>(in.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);
+            auto clamped_h  = utility::clamp<int>(index_h, 0, in_dim_h - 1);
+            auto clamped_h1 = utility::clamp<int>(index_h + 1, 0, in_dim_h - 1);
+
+            const auto a00 = *(pixel_row_ptr + clamped_w * stride_w + clamped_h * stride_h);
+            const auto a01 = *(pixel_row_ptr + clamped_w1 * stride_w + clamped_h * stride_h);
+            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<int8_t>::dequantize(a00, iq_info);
+            const float inp01                      = Qasymm8QuantizationHelper<int8_t>::dequantize(a01, iq_info);
+            const float inp10                      = Qasymm8QuantizationHelper<int8_t>::dequantize(a10, iq_info);
+            const float inp11                      = Qasymm8QuantizationHelper<int8_t>::dequantize(a11, iq_info);
+            *reinterpret_cast<int8_t *>(out.ptr()) = Qasymm8QuantizationHelper<int8_t>::quantize(scale_helpers::delta_bilinear(inp00, inp01, inp10, inp11, dx_val, dy_val), oq_info);
+        },
+        in, out);
+    }
+    else
+    {
+        ARM_COMPUTE_ERROR("Not implemented");
+    }
+}
+}
+namespace cpu
+{
+void qasymm8_signed_sve_scale(const ITensor *src, ITensor *dst, const ITensor *offsets, const ITensor *dx, const ITensor *dy,
+                              InterpolationPolicy policy, BorderMode border_mode, PixelValue constant_border_value, float sampling_offset,
+                              bool align_corners, const Window &window)
+{
+    if(policy == InterpolationPolicy::BILINEAR)
+    {
+        qasymm8_signed_sve_scale_bilinear(src, dst, offsets, dx, dy, border_mode, constant_border_value, sampling_offset, align_corners, window);
+    }
+    else if(policy == InterpolationPolicy::NEAREST_NEIGHBOR)
+    {
+        qasymm8_signed_sve_scale_nearest(src, dst, offsets, sampling_offset, align_corners, window);
+    }
+}
+} // namespace cpu
+} // namespace arm_compute
+
+#endif // __ARM_FEATURE_SVE
\ No newline at end of file
diff --git a/src/core/NEON/kernels/scale/impl/list.h b/src/core/NEON/kernels/scale/impl/list.h
new file mode 100644
index 0000000000..eb3ba8b392
--- /dev/null
+++ b/src/core/NEON/kernels/scale/impl/list.h
@@ -0,0 +1,53 @@
+/*
+ * Copyright (c) 2021 Arm Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#ifndef SRC_CORE_NEON_KERNELS_SCALE_LIST_H
+#define SRC_CORE_NEON_KERNELS_SCALE_LIST_H
+
+namespace arm_compute
+{
+namespace cpu
+{
+#define DECLARE_SCALE_KERNEL(func_name)                                                                                         \
+    void func_name(const ITensor *src, ITensor *dst, const ITensor *offsets, const ITensor *dx, const ITensor *dy,              \
+                   InterpolationPolicy policy, BorderMode border_mode, PixelValue constant_border_value, float sampling_offset, \
+                   bool align_corners, const Window &window)
+
+DECLARE_SCALE_KERNEL(fp16_neon_scale);
+DECLARE_SCALE_KERNEL(fp16_sve_scale);
+DECLARE_SCALE_KERNEL(fp32_neon_scale);
+DECLARE_SCALE_KERNEL(fp32_sve_scale);
+DECLARE_SCALE_KERNEL(s16_neon_scale);
+DECLARE_SCALE_KERNEL(s16_sve_scale);
+DECLARE_SCALE_KERNEL(u8_neon_scale);
+DECLARE_SCALE_KERNEL(u8_sve_scale);
+DECLARE_SCALE_KERNEL(qasymm8_neon_scale);
+DECLARE_SCALE_KERNEL(qasymm8_sve_scale);
+DECLARE_SCALE_KERNEL(qasymm8_signed_neon_scale);
+DECLARE_SCALE_KERNEL(qasymm8_signed_sve_scale);
+
+#undef DECLARE_ACTIVATION_KERNEL
+} // namespace cpu
+} // namespace arm_compute
+
+#endif /* SRC_CORE_NEON_KERNELS_SCALE_LIST_H */
diff --git a/src/core/cpu/kernels/activation/SVE/qasymm8.cpp b/src/core/cpu/kernels/activation/SVE/qasymm8.cpp
index 9eea3ace9e..3d9476ac56 100644
--- a/src/core/cpu/kernels/activation/SVE/qasymm8.cpp
+++ b/src/core/cpu/kernels/activation/SVE/qasymm8.cpp
@@ -26,7 +26,6 @@
 #include "arm_compute/core/Window.h"
 #include "src/core/common/Validate.h"
 
-#include <arm_neon.h>
 #include <cmath>
 #include <cstddef>
 
diff --git a/src/core/helpers/ScaleHelpers.h b/src/core/helpers/ScaleHelpers.h
index 827bbef4cd..f19a8b8f60 100644
--- a/src/core/helpers/ScaleHelpers.h
+++ b/src/core/helpers/ScaleHelpers.h
@@ -1,5 +1,5 @@
 /*
-* Copyright (c) 2020 Arm Limited.
+* Copyright (c) 2020-2021 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -325,6 +325,32 @@ pixel_area_c1u8_clamp(const uint8_t *first_pixel_ptr, size_t stride, size_t widt
     // Return average
     return sum / (x_elements * y_elements);
 }
+
+/** Computes bilinear interpolation using the top-left, top-right, bottom-left, bottom-right pixels and the pixel's distance between
+ * the real coordinates and the smallest following integer coordinates.
+ *
+ * @param[in] a00 The top-left pixel value.
+ * @param[in] a01 The top-right pixel value.
+ * @param[in] a10 The bottom-left pixel value.
+ * @param[in] a11 The bottom-right pixel value.
+ * @param[in] dx  Pixel's distance between the X real coordinate and the smallest X following integer
+ * @param[in] dy  Pixel's distance between the Y real coordinate and the smallest Y following integer
+ *
+ * @note dx and dy must be in the range [0, 1.0]
+ *
+ * @return The bilinear interpolated pixel value
+ */
+inline float delta_bilinear(float a00, float a01, float a10, float a11, float dx_val, float dy_val)
+{
+    const float dx1_val = 1.0f - dx_val;
+    const float dy1_val = 1.0f - dy_val;
+
+    const float w1 = dx1_val * dy1_val;
+    const float w2 = dx_val * dy1_val;
+    const float w3 = dx1_val * dy_val;
+    const float w4 = dx_val * dy_val;
+    return a00 * w1 + a01 * w2 + a10 * w3 + a11 * w4;
+}
 } // namespace scale_helpers
 } // namespace arm_compute