COMPMID-470: Neon Deconvolution.

Implemented by up-sampling the input with zeros insertions between the input samples and convolving the Deconvolution kernels on the up-sampled result.

The upsampling is performed by the function NEDeconvolutionLayerUpsample.
Convolving is done by NEDirectConvolutionLayer.

Change-Id: I25f7ba7c6b99cd9310797972ede40aeff4a54900
Reviewed-on: http://mpd-gerrit.cambridge.arm.com/85319
Tested-by: Kaizen <jeremy.johnson+kaizengerrit@arm.com>
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>

5 files changed, 441 insertions, 2 deletions

diff --git a/src/core/NEON/kernels/NEDeconvolutionLayerUpsampleKernel.cpp b/src/core/NEON/kernels/NEDeconvolutionLayerUpsampleKernel.cpp
new file mode 100644
index 0000000000..71db2e9782
--- /dev/null
+++ b/src/core/NEON/kernels/NEDeconvolutionLayerUpsampleKernel.cpp
@@ -0,0 +1,165 @@
+/*
+ * Copyright (c) 2016, 2017 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/NEON/kernels/NEDeconvolutionLayerUpsampleKernel.h"
+
+#include "arm_compute/core/AccessWindowStatic.h"
+#include "arm_compute/core/Coordinates.h"
+#include "arm_compute/core/Error.h"
+#include "arm_compute/core/Helpers.h"
+#include "arm_compute/core/ITensor.h"
+#include "arm_compute/core/TensorInfo.h"
+#include "arm_compute/core/Validate.h"
+#include "arm_compute/core/Window.h"
+
+#include <arm_neon.h>
+#include <cstddef>
+#include <cstdint>
+
+using namespace arm_compute;
+
+NEDeconvolutionLayerUpsampleKernel::NEDeconvolutionLayerUpsampleKernel()
+    : _offsets(nullptr), _input(nullptr), _output(nullptr)
+{
+}
+
+BorderSize NEDeconvolutionLayerUpsampleKernel::border_size() const
+{
+    return BorderSize(1);
+}
+
+void NEDeconvolutionLayerUpsampleKernel::configure(const ITensor *input, const ITensor *offsets, ITensor *output)
+{
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32);
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::F32);
+    ARM_COMPUTE_ERROR_ON(output->info()->dimension(0) == 0);
+    ARM_COMPUTE_ERROR_ON(output->info()->dimension(1) == 0);
+
+    for(size_t i = 2; i < Coordinates::num_max_dimensions; ++i)
+    {
+        ARM_COMPUTE_ERROR_ON(input->info()->dimension(i) != output->info()->dimension(i));
+    }
+
+    _input   = input;
+    _output  = output;
+    _offsets = offsets;
+
+    constexpr unsigned int num_elems_processed_per_iteration = 16;
+    const int              border_offset                     = border_size().left;
+
+    // Configure kernel window
+    Window win = calculate_max_window(*output->info(), Steps(num_elems_processed_per_iteration));
+
+    AccessWindowRectangle  input_access(input->info(), -border_offset, -border_offset, input->info()->dimension(0) + border_offset, input->info()->dimension(1) + border_offset);
+    AccessWindowHorizontal offsets_access(offsets->info(), 0, num_elems_processed_per_iteration);
+    AccessWindowHorizontal output_access(output->info(), 0, num_elems_processed_per_iteration);
+
+    update_window_and_padding(win, input_access, offsets_access, output_access);
+
+    output_access.set_valid_region(win, ValidRegion(Coordinates(), output->info()->tensor_shape()));
+
+    INEKernel::configure(win);
+}
+
+void NEDeconvolutionLayerUpsampleKernel::scale_nearest(const Window &window)
+{
+    const size_t input_stride = _input->info()->strides_in_bytes()[1];
+
+    // Compute the ratio between source height and destination height
+    const auto hr = static_cast<float>(_input->info()->dimension(1)) / static_cast<float>(_output->info()->dimension(1));
+
+    // 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(Window::DimX, Window::Dimension(0, 0, 0));
+    win_in.set(Window::DimY, Window::Dimension(0, 0, 0));
+
+    Window win_off;
+    win_off.set(Window::DimX, window[Window::DimX]);
+    win_off.set(Window::DimY, window[Window::DimY]);
+
+    for(size_t d = Window::DimZ; d < _offsets->info()->num_dimensions(); ++d)
+    {
+        win_off.set(d, Window::Dimension(0, 0, 0));
+    }
+
+    Iterator in(_input, win_in);
+    Iterator out(_output, window);
+    Iterator offsets(_offsets, win_off);
+
+    switch(_input->info()->data_type())
+    {
+        case DataType::F32:
+        {
+            float32x4x4_t tmp =
+            {
+                {
+                    vdupq_n_f32(0),
+                    vdupq_n_f32(0)
+                }
+            };
+            execute_window_loop(window, [&](const Coordinates & id)
+            {
+                const auto offsets_ptr = reinterpret_cast<const int32_t *>(offsets.ptr());
+
+                const size_t in_yi      = (id.y() + 0.5f) * hr;
+                const size_t offset_row = in_yi * input_stride;
+
+                tmp.val[0] = vsetq_lane_f32(*reinterpret_cast<const float *>(in.ptr() + offsets_ptr[0] + offset_row), tmp.val[0], 0);
+                tmp.val[0] = vsetq_lane_f32(*reinterpret_cast<const float *>(in.ptr() + offsets_ptr[4] + offset_row), tmp.val[0], 1);
+                tmp.val[0] = vsetq_lane_f32(*reinterpret_cast<const float *>(in.ptr() + offsets_ptr[8] + offset_row), tmp.val[0], 2);
+                tmp.val[0] = vsetq_lane_f32(*reinterpret_cast<const float *>(in.ptr() + offsets_ptr[12] + offset_row), tmp.val[0], 3);
+
+                tmp.val[1] = vsetq_lane_f32(*reinterpret_cast<const float *>(in.ptr() + offsets_ptr[1] + offset_row), tmp.val[1], 0);
+                tmp.val[1] = vsetq_lane_f32(*reinterpret_cast<const float *>(in.ptr() + offsets_ptr[5] + offset_row), tmp.val[1], 1);
+                tmp.val[1] = vsetq_lane_f32(*reinterpret_cast<const float *>(in.ptr() + offsets_ptr[9] + offset_row), tmp.val[1], 2);
+                tmp.val[1] = vsetq_lane_f32(*reinterpret_cast<const float *>(in.ptr() + offsets_ptr[13] + offset_row), tmp.val[1], 3);
+
+                tmp.val[2] = vsetq_lane_f32(*reinterpret_cast<const float *>(in.ptr() + offsets_ptr[2] + offset_row), tmp.val[2], 0);
+                tmp.val[2] = vsetq_lane_f32(*reinterpret_cast<const float *>(in.ptr() + offsets_ptr[6] + offset_row), tmp.val[2], 1);
+                tmp.val[2] = vsetq_lane_f32(*reinterpret_cast<const float *>(in.ptr() + offsets_ptr[10] + offset_row), tmp.val[2], 2);
+                tmp.val[2] = vsetq_lane_f32(*reinterpret_cast<const float *>(in.ptr() + offsets_ptr[14] + offset_row), tmp.val[2], 3);
+
+                tmp.val[3] = vsetq_lane_f32(*reinterpret_cast<const float *>(in.ptr() + offsets_ptr[3] + offset_row), tmp.val[3], 0);
+                tmp.val[3] = vsetq_lane_f32(*reinterpret_cast<const float *>(in.ptr() + offsets_ptr[7] + offset_row), tmp.val[3], 1);
+                tmp.val[3] = vsetq_lane_f32(*reinterpret_cast<const float *>(in.ptr() + offsets_ptr[11] + offset_row), tmp.val[3], 2);
+                tmp.val[3] = vsetq_lane_f32(*reinterpret_cast<const float *>(in.ptr() + offsets_ptr[15] + offset_row), tmp.val[3], 3);
+
+                vst4q_f32(reinterpret_cast<float *>(out.ptr()), tmp);
+            },
+            in, offsets, out);
+            break;
+        }
+        default:
+            ARM_COMPUTE_ERROR("Not supported");
+            break;
+    }
+}
+
+void NEDeconvolutionLayerUpsampleKernel::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);
+    scale_nearest(window);
+}
diff --git a/src/core/NEON/kernels/NEScaleKernel.cpp b/src/core/NEON/kernels/NEScaleKernel.cpp
index 6634d4b13c..b1ced7e38d 100644
--- a/src/core/NEON/kernels/NEScaleKernel.cpp
+++ b/src/core/NEON/kernels/NEScaleKernel.cpp
@@ -180,8 +180,10 @@ void NEScaleKernel::scale_nearest(const Window &window)
                 const auto           offsets_ptr = reinterpret_cast<const int32_t *>(offsets.ptr());
                 const uint8_t *const in_ptr      = in.ptr();
 
-                const int in_yi      = std::floor((id.y() + 0.5f) * hr);
-                const int offset_row = in_yi * input_stride;
+                const int in_yi         = std::floor((id.y() + 0.5f) * hr);
+                const int in_yi_clamped = std::min(static_cast<int>(_input->info()->dimension(1)), std::max(in_yi, -1));
+                ARM_COMPUTE_ERROR_ON(in_yi_clamped < -1 || in_yi_clamped > static_cast<int>(_input->info()->dimension(1)));
+                const int offset_row = in_yi_clamped * input_stride;
 
                 tmp = vsetq_lane_u8(in_ptr[offsets_ptr[0] + offset_row], tmp, 0);
                 tmp = vsetq_lane_u8(in_ptr[offsets_ptr[1] + offset_row], tmp, 1);
diff --git a/src/core/Utils.cpp b/src/core/Utils.cpp
index 99d39569c7..d5ce1ea027 100644
--- a/src/core/Utils.cpp
+++ b/src/core/Utils.cpp
@@ -247,6 +247,43 @@ std::string arm_compute::lower_string(const std::string &val)
     return res;
 }
 
+TensorShape arm_compute::deconvolution_output_shape(const std::pair<unsigned int, unsigned int> &out_dims, TensorShape input, TensorShape weights)
+{
+    TensorShape out_shape(input);
+    out_shape.set(0, out_dims.first);
+    out_shape.set(1, out_dims.second);
+    out_shape.set(2, weights[3]);
+    return out_shape;
+}
+
+const std::pair<unsigned int, unsigned int> arm_compute::deconvolution_output_dimensions(
+    unsigned int in_width, unsigned int in_height, unsigned int kernel_width, unsigned int kernel_height, unsigned int padx, unsigned int pady,
+    unsigned int ax, unsigned int ay, float upscalex, float upscaley, DimensionRoundingType round)
+{
+    ARM_COMPUTE_ERROR_ON(in_width < 1 || in_height < 1);
+    ARM_COMPUTE_ERROR_ON(((in_width - 1) * upscalex + kernel_width + ax) < 2.f * padx);
+    ARM_COMPUTE_ERROR_ON(((in_height - 1) * upscaley + kernel_height + ay) < 2.f * pady);
+    const float fw = (in_width - 1) * upscalex - 2.f * padx + kernel_width + ax;
+    const float fh = (in_height - 1) * upscaley - 2.f * pady + kernel_height + ay;
+    int         w  = 0;
+    int         h  = 0;
+    switch(round)
+    {
+        case DimensionRoundingType::FLOOR:
+            w = std::floor(fw);
+            h = std::floor(fh);
+            break;
+        case DimensionRoundingType::CEIL:
+            w = std::ceil(fw);
+            h = std::ceil(fh);
+            break;
+        default:
+            ARM_COMPUTE_ERROR("Not supported");
+            break;
+    }
+    return std::make_pair<unsigned int, unsigned int>(w, h);
+}
+
 const std::pair<unsigned int, unsigned int> arm_compute::scaled_dimensions(unsigned int width, unsigned int height,
                                                                            unsigned int kernel_width, unsigned int kernel_height,
                                                                            const PadStrideInfo &pad_stride_info)
diff --git a/src/runtime/NEON/functions/NEDeconvolutionLayer.cpp b/src/runtime/NEON/functions/NEDeconvolutionLayer.cpp
new file mode 100644
index 0000000000..7b4e77b296
--- /dev/null
+++ b/src/runtime/NEON/functions/NEDeconvolutionLayer.cpp
@@ -0,0 +1,114 @@
+/*
+ * Copyright (c) 2017 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/runtime/NEON/functions/NEDeconvolutionLayer.h"
+
+#include "arm_compute/core/Helpers.h"
+#include "arm_compute/core/PixelValue.h"
+#include "arm_compute/core/Utils.h"
+#include "arm_compute/core/Validate.h"
+
+using namespace arm_compute;
+
+NEDeconvolutionLayer::NEDeconvolutionLayer(std::shared_ptr<IMemoryManager> memory_manager) // NOLINT
+    : _memory_group(std::move(memory_manager)),
+      _scale_f(),
+      _conv_f(),
+      _scaled_output()
+{
+}
+
+void NEDeconvolutionLayer::configure(ITensor *input, const ITensor *weights, const ITensor *bias, ITensor *output, const PadStrideInfo &info,
+                                     unsigned int ax, unsigned int ay, float upscalex, float upscaley)
+{
+    ARM_COMPUTE_ERROR_ON_NULLPTR(output);
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32);
+    ARM_COMPUTE_ERROR_ON(weights->info()->dimension(0) != weights->info()->dimension(1));
+    ARM_COMPUTE_ERROR_ON(weights->info()->dimension(0) < 1);
+
+    auto out_dims = deconvolution_output_dimensions(input->info()->dimension(0), input->info()->dimension(1), weights->info()->dimension(0), weights->info()->dimension(1),
+                                                    info.pad().first, info.pad().second, ax, ay, upscalex, upscaley, info.round());
+
+    const TensorShape output_shape = deconvolution_output_shape(out_dims, input->info()->tensor_shape(), weights->info()->tensor_shape());
+
+    // Output auto initialization if not yet initialized
+    auto_init_if_empty(*output->info(), output_shape, 1, input->info()->data_type(), input->info()->fixed_point_position());
+
+    ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, output, weights, bias);
+    ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(input, output, weights, bias);
+
+    ARM_COMPUTE_ERROR_ON_MSG(output->info()->dimension(Window::DimX) != output_shape.x(), "Output's width is invalid.");
+    ARM_COMPUTE_ERROR_ON_MSG(output->info()->dimension(Window::DimY) != output_shape.y(), "Output's height is invalid.");
+    ARM_COMPUTE_ERROR_ON_MSG(output->info()->dimension(Window::DimZ) != output_shape.z(), "Output's depth is invalid.");
+
+    _memory_group.manage(&_scaled_output);
+
+    // configure scale function
+    //Init and allocate intermmidiate tensor for output, same size as input but the first two axis are the same as the output tensor
+    TensorShape scale_out_shape(input->info()->tensor_shape());
+    scale_out_shape.set(0, output->info()->dimension(0));
+    scale_out_shape.set(1, output->info()->dimension(1));
+    TensorInfo scale_out_info(scale_out_shape, 1, input->info()->data_type(), input->info()->fixed_point_position());
+    _scaled_output.allocator()->init(scale_out_info);
+    const unsigned int kernel_size = weights->info()->dimension(0);
+    // Padding for the upsampled image is calculated with the equiation: p' = k - p - 1, where k is kernel size and p is the input padding
+    ARM_COMPUTE_ERROR_ON(info.pad().first > (kernel_size - 1));
+    const unsigned int  tr_px     = kernel_size - info.pad().first - 1;
+    const unsigned int  tr_py     = kernel_size - info.pad().second - 1;
+    const unsigned int  tr_stride = 1;
+    const PadStrideInfo transposed_info(tr_stride, tr_stride, tr_px, tr_py);
+    _scale_f.configure(input, &_scaled_output, std::make_pair(ax, ay), std::make_pair(info.stride().first - 1u, info.stride().second - 1u), transposed_info);
+    // setup the function to convolve the upscaled output
+    switch(kernel_size)
+    {
+        case 1:
+        {
+            _conv_f.configure(&_scaled_output, weights, bias, output, PadStrideInfo(1, 1, 0, 0, DimensionRoundingType::CEIL));
+            break;
+        }
+        case 3:
+        {
+            _conv_f.configure(&_scaled_output, weights, bias, output, PadStrideInfo(1, 1, 1, 1, DimensionRoundingType::CEIL));
+            break;
+        }
+        case 5:
+        {
+            _conv_f.configure(&_scaled_output, weights, bias, output, PadStrideInfo(1, 1, 2, 2, DimensionRoundingType::CEIL));
+            break;
+        }
+        default:
+        {
+            ARM_COMPUTE_ERROR("Not supported");
+            break;
+        }
+    }
+    _scaled_output.allocator()->allocate();
+}
+
+void NEDeconvolutionLayer::run()
+{
+    _memory_group.acquire();
+    _scale_f.run();
+    _conv_f.run();
+    _memory_group.release();
+}
diff --git a/src/runtime/NEON/functions/NEDeconvolutionLayerUpsample.cpp b/src/runtime/NEON/functions/NEDeconvolutionLayerUpsample.cpp
new file mode 100644
index 0000000000..63f17bcb5a
--- /dev/null
+++ b/src/runtime/NEON/functions/NEDeconvolutionLayerUpsample.cpp
@@ -0,0 +1,121 @@
+/*
+ * Copyright (c) 2016, 2017 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/runtime/NEON/functions/NEDeconvolutionLayerUpsample.h"
+
+#include "arm_compute/core/Coordinates.h"
+#include "arm_compute/core/Error.h"
+#include "arm_compute/core/Helpers.h"
+#include "arm_compute/core/ITensor.h"
+#include "arm_compute/core/NEON/kernels/NEDeconvolutionLayerUpsampleKernel.h"
+#include "arm_compute/core/PixelValue.h"
+#include "arm_compute/core/TensorInfo.h"
+#include "arm_compute/core/Window.h"
+#include "arm_compute/runtime/NEON/NEScheduler.h"
+#include "arm_compute/runtime/TensorAllocator.h"
+#include "support/ToolchainSupport.h"
+
+#include <cmath>
+#include <cstddef>
+#include <utility>
+
+using namespace arm_compute;
+
+namespace
+{
+inline void precompute_offsets(ITensor *offsets, float wr, size_t input_element_size, const std::pair<unsigned int, unsigned int> &a,
+                               const std::pair<unsigned int, unsigned int> &iz, const PadStrideInfo &info)
+{
+    ARM_COMPUTE_ERROR_ON(nullptr == offsets);
+    Window    win;
+    const int padx          = info.pad().first;
+    const int pady          = info.pad().second;
+    const int ax            = a.first;
+    const int ay            = a.second;
+    const int offset_width  = offsets->info()->dimension(0);
+    const int offset_height = offsets->info()->dimension(1);
+    // The values of ax and ay denote the number of ZEROS to be added on the top and right inner border of the image.
+    // Step value along the XY axis will depend on the number of zeros to be inserted between samples (number of zeros + 1).
+    // Pre-compute the X offset, Y's stride is unknown at this point so we can't precompute Y's offsets
+    for(int yi = ay; yi < (offset_height - pady); yi += (1 + iz.second))
+    {
+        for(int xi = padx; xi < (offset_width - ax); xi += (1 + iz.first))
+        {
+            int         *ptr                  = reinterpret_cast<int *>(offsets->ptr_to_element(Coordinates(xi, yi)));
+            const size_t in_xi                = (xi + 0.5f) * wr;
+            *reinterpret_cast<int32_t *>(ptr) = in_xi * input_element_size;
+        }
+    }
+}
+} // namespace
+
+NEDeconvolutionLayerUpsample::NEDeconvolutionLayerUpsample(std::shared_ptr<IMemoryManager> memory_manager) // NOLINT
+    : _memory_group(std::move(memory_manager)),
+      _offsets(),
+      _border_handler(),
+      _upsample()
+{
+}
+
+void NEDeconvolutionLayerUpsample::configure(ITensor *input, ITensor *output, const std::pair<unsigned int, unsigned int> &a,
+                                             const std::pair<unsigned int, unsigned int> &iz, const PadStrideInfo &info)
+{
+    ARM_COMPUTE_ERROR_ON(nullptr == input);
+    ARM_COMPUTE_ERROR_ON(nullptr == output);
+
+    for(size_t i = 2; i < Coordinates::num_max_dimensions; ++i)
+    {
+        ARM_COMPUTE_ERROR_ON(input->info()->dimension(i) != output->info()->dimension(i));
+    }
+
+    // Get the tensor shape
+    const TensorShape shape(output->info()->dimension(0), output->info()->dimension(1));
+
+    // Compute the ratio between source width/height and destination width/height
+    const auto wr = static_cast<float>(input->info()->dimension(0)) / static_cast<float>(output->info()->dimension(0));
+    const auto hr = static_cast<float>(input->info()->dimension(1)) / static_cast<float>(output->info()->dimension(1));
+    ARM_COMPUTE_UNUSED(hr);
+    // Get the element size of the input image
+    const size_t input_element_size = input->info()->element_size();
+
+    TensorInfo tensor_info_offsets(shape, Format::S32);
+    _offsets.allocator()->init(tensor_info_offsets);
+
+    _upsample.configure(input, &_offsets, output);
+
+    // Allocate once the configure methods have been called
+    _offsets.allocator()->allocate();
+    // Pre-compute offsets for nearest interpolation
+    std::fill_n(reinterpret_cast<int32_t *>(_offsets.buffer()), _offsets.info()->total_size() / sizeof(int32_t), -1 * input_element_size);
+    precompute_offsets(&_offsets, wr, input_element_size, a, iz, info);
+
+    _border_handler.configure(input, _upsample.border_size(), BorderMode::CONSTANT, PixelValue(0));
+}
+
+void NEDeconvolutionLayerUpsample::run()
+{
+    NEScheduler::get().schedule(&_border_handler, Window::DimZ);
+    _memory_group.acquire();
+    NEScheduler::get().schedule(&_upsample, Window::DimY);
+    _memory_group.release();
+}