COMPMID-1621 Deconvolution wrong output calculation

Change-Id: Ida71312bcf6dbd854f2ab1efc65f74910c79e152
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/151510
Tested-by: bsgcomp <bsgcomp@arm.com>
Reviewed-by: Michele DiGiorgio <michele.digiorgio@arm.com>

16 files changed, 361 insertions, 77 deletions

diff --git a/arm_compute/core/CPP/CPPKernels.h b/arm_compute/core/CPP/CPPKernels.h
index a0c5707a79..bf24a94aa7 100644
--- a/arm_compute/core/CPP/CPPKernels.h
+++ b/arm_compute/core/CPP/CPPKernels.h
@@ -27,6 +27,7 @@
 /* Header regrouping all the CPP kernels */
 #include "arm_compute/core/CPP/kernels/CPPCornerCandidatesKernel.h"
 #include "arm_compute/core/CPP/kernels/CPPDetectionWindowNonMaximaSuppressionKernel.h"
+#include "arm_compute/core/CPP/kernels/CPPFlipWeightsKernel.h"
 #include "arm_compute/core/CPP/kernels/CPPPermuteKernel.h"
 #include "arm_compute/core/CPP/kernels/CPPSortEuclideanDistanceKernel.h"
 #include "arm_compute/core/CPP/kernels/CPPUpsampleKernel.h"
diff --git a/arm_compute/core/CPP/kernels/CPPFlipWeightsKernel.h b/arm_compute/core/CPP/kernels/CPPFlipWeightsKernel.h
new file mode 100644
index 0000000000..801934159d
--- /dev/null
+++ b/arm_compute/core/CPP/kernels/CPPFlipWeightsKernel.h
@@ -0,0 +1,85 @@
+/*
+ * Copyright (c) 2018 ARM Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#ifndef __ARM_COMPUTE_CPP_FLIP_WEIGHTS_KERNEL_H__
+#define __ARM_COMPUTE_CPP_FLIP_WEIGHTS_KERNEL_H__
+
+#include "arm_compute/core/CPP/ICPPKernel.h"
+
+namespace arm_compute
+{
+class ITensor;
+
+/** CPP kernel to perform 180 degrees flipping on deconvolution weights. */
+class CPPFlipWeightsKernel : public ICPPKernel
+{
+public:
+    const char *name() const override
+    {
+        return "CPPFlipWeightsKernel";
+    }
+    /** Default constructor */
+    CPPFlipWeightsKernel();
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    CPPFlipWeightsKernel(const CPPFlipWeightsKernel &) = delete;
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    CPPFlipWeightsKernel &operator=(const CPPFlipWeightsKernel &) = delete;
+    /** Allow instances of this class to be moved */
+    CPPFlipWeightsKernel(CPPFlipWeightsKernel &&) = default;
+    /** Allow instances of this class to be moved */
+    CPPFlipWeightsKernel &operator=(CPPFlipWeightsKernel &&) = default;
+    /** Default destructor */
+    ~CPPFlipWeightsKernel() = default;
+
+    /** Set the input and output of the kernel.
+     *
+     * @param[in]  input  The input tensor to flip. Data types supported: QASYMM8/F16/F32
+     * @param[out] output The output tensor. Data types supported: Same as @p input
+     */
+    void configure(const ITensor *input, ITensor *output);
+
+    // Inherited methods overridden:
+    void run(const Window &window, const ThreadInfo &info) override;
+
+    /** Function to perform flipping.
+     *
+     * @param[in] window_input Input region on which to execute the kernel.
+     * @param[in] window       Output region on which to execute the kernel.
+     */
+    template <typename T>
+    void flip_weights(const Window &window_input, const Window &window);
+
+    /** Common signature for all the specialised Flip functions
+     *
+     * @param[in] window_input Input region on which to execute the kernel.
+     * @param[in] window       Output region on which to execute the kernel.
+     */
+    using FlipWeightsFunction = void (CPPFlipWeightsKernel::*)(const Window &window_input, const Window &window);
+
+private:
+    const ITensor      *_input;
+    ITensor            *_output;
+    FlipWeightsFunction _func;
+};
+} // namespace arm_compute
+#endif /*__ARM_COMPUTE_CPP_FLIP_WEIGHTS_KERNEL_H__ */
diff --git a/arm_compute/core/Utils.h b/arm_compute/core/Utils.h
index c742ebc50e..7ee24e2736 100644
--- a/arm_compute/core/Utils.h
+++ b/arm_compute/core/Utils.h
@@ -827,22 +827,20 @@ TensorShape deconvolution_output_shape(const std::pair<unsigned int, unsigned in
 
 /** Returns expected width and height of the deconvolution's output tensor.
  *
- * @param[in] in_width           Width of input tensor (Number of columns)
- * @param[in] in_height          Height of input tensor (Number of rows)
- * @param[in] kernel_width       Kernel width.
- * @param[in] kernel_height      Kernel height.
- * @param[in] padx               X axis padding.
- * @param[in] pady               Y axis padding.
- * @param[in] inner_border_right The number of zeros added to right edge of the input.
- * @param[in] inner_border_top   The number of zeros added to top edge of the input.
- * @param[in] stride_x           X axis input stride.
- * @param[in] stride_y           Y axis input stride.
+ * @param[in] in_width      Width of input tensor (Number of columns)
+ * @param[in] in_height     Height of input tensor (Number of rows)
+ * @param[in] kernel_width  Kernel width.
+ * @param[in] kernel_height Kernel height.
+ * @param[in] padx          X axis padding.
+ * @param[in] pady          Y axis padding.
+ * @param[in] stride_x      X axis input stride.
+ * @param[in] stride_y      Y axis input stride.
  *
  * @return A pair with the new width in the first position and the new height in the second.
  */
 const std::pair<unsigned int, unsigned int> 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 inner_border_right, unsigned int inner_border_top,
+                                                                            unsigned int padx, unsigned int pady,
                                                                             unsigned int stride_x, unsigned int stride_y);
 
 /** Returns expected width and height of output scaled tensor depending on dimensions rounding mode.
diff --git a/arm_compute/core/utils/misc/ShapeCalculator.h b/arm_compute/core/utils/misc/ShapeCalculator.h
index 804ff3c709..f68401c1b9 100644
--- a/arm_compute/core/utils/misc/ShapeCalculator.h
+++ b/arm_compute/core/utils/misc/ShapeCalculator.h
@@ -229,11 +229,20 @@ inline TensorShape compute_depthwise_convolution_shape(const ITensorInfo &input,
     return output_shape;
 }
 
-inline TensorShape compute_deconvolution_shape(const ITensorInfo &input, unsigned int sx, unsigned int sy, unsigned int inner_border_right, unsigned int inner_border_top, const PadStrideInfo &info)
+inline TensorShape compute_deconvolution_shape(const ITensorInfo &input, const ITensorInfo &weights, unsigned int sx, unsigned int sy, unsigned int inner_border_right, unsigned int inner_border_top,
+                                               std::pair<unsigned int, unsigned int> &out_dims)
 {
-    TensorShape        scale_out_shape(input.tensor_shape());
-    const unsigned int out_x = input.dimension(0) + (input.dimension(0) - 1) * (sx - 1) + inner_border_right + 2 * info.pad().first;
-    const unsigned int out_y = input.dimension(1) + (input.dimension(1) - 1) * (sy - 1) + inner_border_top + 2 * info.pad().second;
+    // Find the upsampled dimensions
+    unsigned int out_x = (input.dimension(0) - 1) * sx + inner_border_right + 1;
+    unsigned int out_y = (input.dimension(1) - 1) * sy + inner_border_top + 1;
+
+    // Find the padding needed for the convolution with stride 1 in order to match output shape
+    unsigned int padx = out_dims.first - (out_x - weights.dimension(0) + 1);
+    unsigned int pady = out_dims.second - (out_y - weights.dimension(1) + 1);
+    out_x += padx;
+    out_y += pady;
+
+    TensorShape scale_out_shape(input.tensor_shape());
     scale_out_shape.set(0, out_x);
     scale_out_shape.set(1, out_y);
 
diff --git a/arm_compute/graph/nodes/DeconvolutionLayerNode.h b/arm_compute/graph/nodes/DeconvolutionLayerNode.h
index 73210a299e..19501482c6 100644
--- a/arm_compute/graph/nodes/DeconvolutionLayerNode.h
+++ b/arm_compute/graph/nodes/DeconvolutionLayerNode.h
@@ -55,14 +55,12 @@ public:
      * @param[in] input_descriptor   Input descriptor
      * @param[in] weights_descriptor Weights descriptor
      * @param[in] info               Convolution operation attributes
-     * @param[in] inner_border       Inner border (right, top)
      *
      * @return Output descriptor
      */
     static TensorDescriptor compute_output_descriptor(const TensorDescriptor &input_descriptor,
                                                       const TensorDescriptor &weights_descriptor,
-                                                      const PadStrideInfo    &info,
-                                                      const Size2D           &inner_border);
+                                                      const PadStrideInfo    &info);
 
     // Inherited overridden methods:
     NodeType         type() const override;
diff --git a/arm_compute/runtime/CL/functions/CLDeconvolutionLayer.h b/arm_compute/runtime/CL/functions/CLDeconvolutionLayer.h
index 4dce1e1801..6716cd6fdd 100644
--- a/arm_compute/runtime/CL/functions/CLDeconvolutionLayer.h
+++ b/arm_compute/runtime/CL/functions/CLDeconvolutionLayer.h
@@ -27,6 +27,8 @@
 #include "arm_compute/runtime/CL/functions/CLConvolutionLayer.h"
 #include "arm_compute/runtime/CL/functions/CLDeconvolutionLayerUpsample.h"
 
+#include "arm_compute/core/CPP/kernels/CPPFlipWeightsKernel.h"
+
 #include "arm_compute/runtime/CL/CLMemoryGroup.h"
 #include "arm_compute/runtime/CL/CLTensor.h"
 #include "arm_compute/runtime/IFunction.h"
@@ -62,6 +64,14 @@ class CLDeconvolutionLayer : public IFunction
 public:
     /** Constructor */
     CLDeconvolutionLayer(std::shared_ptr<IMemoryManager> memory_manager = nullptr);
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    CLDeconvolutionLayer(const CLDeconvolutionLayer &) = delete;
+    /** Default move constructor */
+    CLDeconvolutionLayer(CLDeconvolutionLayer &&) = default;
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    CLDeconvolutionLayer &operator=(const CLDeconvolutionLayer &) = delete;
+    /** Default move assignment operator */
+    CLDeconvolutionLayer &operator=(CLDeconvolutionLayer &&) = default;
     /** Set the input, weights, biases and output tensors.
      *
      * @param[in,out] input              Input tensor. 3 lower dimensions represent a single input, and an optional 4th dimension for batch of inputs. Data types supported: QASYMM8/F16/F32.
@@ -74,7 +84,7 @@ public:
      * @param[in]     weights_info       (Optional) Weights information needed for @ref CLConvolutionLayer, specifies if the weights tensor has been reshaped with @ref CLWeightsReshapeKernel.
      *
      */
-    void configure(ICLTensor *input, const ICLTensor *weights, const ICLTensor *bias, ICLTensor *output, const PadStrideInfo &info,
+    void configure(ICLTensor *input, ICLTensor *weights, const ICLTensor *bias, ICLTensor *output, const PadStrideInfo &info,
                    unsigned int inner_border_right, unsigned int inner_border_top, const WeightsInfo &weights_info = WeightsInfo());
     /** Static function to check if given info will lead to a valid configuration of @ref CLDeconvolutionLayer
      *
@@ -100,7 +110,10 @@ private:
     CLMemoryGroup                _memory_group;
     CLDeconvolutionLayerUpsample _scale_f;
     CLConvolutionLayer           _conv_f;
+    CPPFlipWeightsKernel         _flip_weights;
     CLTensor                     _scaled_output;
+    ICLTensor                   *_weights;
+    CLTensor                     _weights_flipped;
     bool                         _is_prepared;
 };
 }
diff --git a/arm_compute/runtime/NEON/functions/NEDeconvolutionLayer.h b/arm_compute/runtime/NEON/functions/NEDeconvolutionLayer.h
index 3e527168c1..0cca555621 100644
--- a/arm_compute/runtime/NEON/functions/NEDeconvolutionLayer.h
+++ b/arm_compute/runtime/NEON/functions/NEDeconvolutionLayer.h
@@ -28,6 +28,7 @@
 #include "arm_compute/runtime/NEON/functions/NEConvolutionLayer.h"
 #include "arm_compute/runtime/NEON/functions/NEDirectConvolutionLayer.h"
 
+#include "arm_compute/core/CPP/kernels/CPPFlipWeightsKernel.h"
 #include "arm_compute/core/Types.h"
 #include "arm_compute/runtime/IFunction.h"
 #include "arm_compute/runtime/IMemoryManager.h"
@@ -111,12 +112,14 @@ public:
     void prepare() override;
 
 private:
-    MemoryGroup        _memory_group;
-    NEConvolutionLayer _conv_f;
-    CPPUpsample        _upsample_f;
-    Tensor             _scaled_output;
-    ITensor           *_input;
-    PadStrideInfo      _info;
+    MemoryGroup          _memory_group;
+    NEConvolutionLayer   _conv_f;
+    CPPUpsample          _upsample_f;
+    CPPFlipWeightsKernel _flip_weights;
+    Tensor               _scaled_output;
+    Tensor               _weights_flipped;
+    ITensor             *_input;
+    PadStrideInfo        _info;
     std::pair<unsigned int, unsigned int> _inner_border;
     bool _is_prepared;
 };
diff --git a/src/core/CPP/kernels/CPPFlipWeightsKernel.cpp b/src/core/CPP/kernels/CPPFlipWeightsKernel.cpp
new file mode 100644
index 0000000000..741218e4f7
--- /dev/null
+++ b/src/core/CPP/kernels/CPPFlipWeightsKernel.cpp
@@ -0,0 +1,106 @@
+/*
+ * Copyright (c) 2018 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/CPP/kernels/CPPFlipWeightsKernel.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/Types.h"
+#include "arm_compute/core/Validate.h"
+#include "arm_compute/core/utils/misc/ShapeCalculator.h"
+
+#include <cstddef>
+#include <cstdint>
+
+using namespace arm_compute;
+
+CPPFlipWeightsKernel::CPPFlipWeightsKernel()
+    : _input(nullptr), _output(nullptr), _func(nullptr)
+{
+}
+
+template <typename T>
+void CPPFlipWeightsKernel::flip_weights(const Window &window_input, const Window &window)
+{
+    // Create iterators
+    Iterator in(_input, window_input);
+
+    Iterator out(_output, window);
+
+    const int kernel_size = _input->info()->dimension(0);
+
+    execute_window_loop(window_input, [&](const Coordinates & id)
+    {
+        *((reinterpret_cast<T *>(out.ptr()) + kernel_size * (kernel_size - id.y() - 1) + (kernel_size - id.x() - 1))) = *(reinterpret_cast<const T *>(in.ptr()));
+    },
+    in, out);
+}
+
+void CPPFlipWeightsKernel::configure(const ITensor *input, ITensor *output)
+{
+    ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
+
+    _input  = input;
+    _output = output;
+
+    // Configure kernel window
+    Window win = calculate_max_window(*input->info(), Steps());
+
+    // The CPPFlipWeightsKernel doesn't need padding so update_window_and_padding() can be skipped
+    Coordinates coord;
+    coord.set_num_dimensions(output->info()->num_dimensions());
+    output->info()->set_valid_region(ValidRegion(coord, output->info()->tensor_shape()));
+
+    ICPPKernel::configure(win);
+
+    switch(input->info()->data_type())
+    {
+        case DataType::F32:
+            _func = &CPPFlipWeightsKernel::flip_weights<float>;
+            break;
+        case DataType::F16:
+            _func = &CPPFlipWeightsKernel::flip_weights<half>;
+            break;
+        case DataType::QASYMM8:
+            _func = &CPPFlipWeightsKernel::flip_weights<uint8_t>;
+            break;
+        default:
+            ARM_COMPUTE_ERROR("Not supported");
+    }
+}
+
+void CPPFlipWeightsKernel::run(const Window &window, const ThreadInfo &info)
+{
+    ARM_COMPUTE_UNUSED(info);
+    ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
+    ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICPPKernel::window(), window);
+    ARM_COMPUTE_ERROR_ON(_func == nullptr);
+
+    Window out_window{ window };
+    out_window.set(Window::DimX, Window::Dimension(0, 0, 0));
+    out_window.set(Window::DimY, Window::Dimension(0, 0, 0));
+
+    (this->*_func)(window, out_window);
+}
diff --git a/src/core/Utils.cpp b/src/core/Utils.cpp
index 229579d8d9..a6a5771ec1 100644
--- a/src/core/Utils.cpp
+++ b/src/core/Utils.cpp
@@ -334,17 +334,14 @@ TensorShape arm_compute::deconvolution_output_shape(const std::pair<unsigned int
 
 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 inner_border_right, unsigned int inner_border_top, unsigned int stride_x, unsigned int stride_y)
+    unsigned int stride_x, unsigned int stride_y)
 {
     ARM_COMPUTE_ERROR_ON(in_width < 1 || in_height < 1);
-    ARM_COMPUTE_ERROR_ON(((in_width - 1) * stride_x + kernel_width + inner_border_right) < 2 * padx);
-    ARM_COMPUTE_ERROR_ON(((in_height - 1) * stride_y + kernel_height + inner_border_top) < 2 * pady);
-    const int padx_deconv = (kernel_width - padx - 1);
-    const int pady_deconv = (kernel_height - pady - 1);
-    ARM_COMPUTE_ERROR_ON(padx_deconv < 0);
-    ARM_COMPUTE_ERROR_ON(pady_deconv < 0);
-    const int w = stride_x * (in_width - 1) + kernel_width + inner_border_right - 2 * padx_deconv;
-    const int h = stride_y * (in_height - 1) + kernel_height + inner_border_top - 2 * pady_deconv;
+    ARM_COMPUTE_ERROR_ON(((in_width - 1) * stride_x + kernel_width) < 2 * padx);
+    ARM_COMPUTE_ERROR_ON(((in_height - 1) * stride_y + kernel_height) < 2 * pady);
+    const int w = stride_x * (in_width - 1) + kernel_width - 2 * padx;
+    const int h = stride_y * (in_height - 1) + kernel_height - 2 * pady;
+
     return std::make_pair<unsigned int, unsigned int>(w, h);
 }
 
diff --git a/src/graph/nodes/DeconvolutionLayerNode.cpp b/src/graph/nodes/DeconvolutionLayerNode.cpp
index 9329ae3c23..e7ccffd04f 100644
--- a/src/graph/nodes/DeconvolutionLayerNode.cpp
+++ b/src/graph/nodes/DeconvolutionLayerNode.cpp
@@ -51,8 +51,7 @@ Size2D DeconvolutionLayerNode::inner_border() const
 
 TensorDescriptor DeconvolutionLayerNode::compute_output_descriptor(const TensorDescriptor &input_descriptor,
                                                                    const TensorDescriptor &weights_descriptor,
-                                                                   const PadStrideInfo    &info,
-                                                                   const Size2D           &inner_border)
+                                                                   const PadStrideInfo    &info)
 {
     unsigned int output_width  = 0;
     unsigned int output_height = 0;
@@ -65,7 +64,6 @@ TensorDescriptor DeconvolutionLayerNode::compute_output_descriptor(const TensorD
     std::tie(output_width, output_height) = deconvolution_output_dimensions(input_width, input_height,
                                                                             kernel_width, kernel_height,
                                                                             info.pad().first, info.pad().second,
-                                                                            inner_border.x(), inner_border.y(),
                                                                             info.stride().first, info.stride().second);
 
     TensorDescriptor output_descriptor = input_descriptor;
@@ -96,7 +94,7 @@ TensorDescriptor DeconvolutionLayerNode::configure_output(size_t idx) const
 
     ARM_COMPUTE_ERROR_ON(src == nullptr || weights == nullptr);
 
-    TensorDescriptor output_info = compute_output_descriptor(src->desc(), weights->desc(), _info, _inner_border);
+    TensorDescriptor output_info = compute_output_descriptor(src->desc(), weights->desc(), _info);
     return output_info;
 }
 
diff --git a/src/runtime/CL/functions/CLDeconvolutionLayer.cpp b/src/runtime/CL/functions/CLDeconvolutionLayer.cpp
index 3f5b8c92dd..26d44e9c96 100644
--- a/src/runtime/CL/functions/CLDeconvolutionLayer.cpp
+++ b/src/runtime/CL/functions/CLDeconvolutionLayer.cpp
@@ -27,6 +27,8 @@
 #include "arm_compute/core/Utils.h"
 #include "arm_compute/core/Validate.h"
 #include "arm_compute/core/utils/misc/ShapeCalculator.h"
+#include "arm_compute/runtime/CL/CLScheduler.h"
+#include "arm_compute/runtime/CPP/CPPScheduler.h"
 
 #include <memory>
 #include <tuple>
@@ -38,7 +40,10 @@ CLDeconvolutionLayer::CLDeconvolutionLayer(std::shared_ptr<IMemoryManager> memor
     : _memory_group(std::move(memory_manager)),
       _scale_f(),
       _conv_f(),
+      _flip_weights(),
       _scaled_output(),
+      _weights(),
+      _weights_flipped(),
       _is_prepared(false)
 {
 }
@@ -59,7 +64,7 @@ Status CLDeconvolutionLayer::validate(const ITensorInfo *input, const ITensorInf
     ARM_COMPUTE_RETURN_ERROR_ON_MSG(inner_border_top > stride_y - 1, "inner_border_top must be smaller than stride_y");
 
     auto out_dims = deconvolution_output_dimensions(input->dimension(0), input->dimension(1), weights->dimension(0), weights->dimension(1),
-                                                    info.pad().first, info.pad().second, inner_border_right, inner_border_top, stride_x, stride_y);
+                                                    info.pad().first, info.pad().second, stride_x, stride_y);
 
     const TensorShape output_shape = deconvolution_output_shape(out_dims, input->tensor_shape(), weights->tensor_shape());
 
@@ -81,8 +86,9 @@ Status CLDeconvolutionLayer::validate(const ITensorInfo *input, const ITensorInf
     ARM_COMPUTE_RETURN_ERROR_ON_MSG(output->dimension(Window::DimY) != output_shape.y(), "Output's height is invalid.");
     ARM_COMPUTE_RETURN_ERROR_ON_MSG(output->dimension(Window::DimZ) != output_shape.z(), "Output's depth is invalid.");
 
-    TensorInfo scale_out_info(input->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(compute_deconvolution_shape(*input, stride_x, stride_y, inner_border_right, inner_border_top,
-                                                                                                      info)));
+    TensorInfo scale_out_info(input->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(compute_deconvolution_shape(*input, *weights, stride_x, stride_y, inner_border_right,
+                                                                                                      inner_border_top,
+                                                                                                      out_dims)));
     const PadStrideInfo conv_info(1, 1, 0, 0, 0, 0, DimensionRoundingType::CEIL);
 
     ARM_COMPUTE_RETURN_ON_ERROR(CLDeconvolutionLayerUpsample::validate(input, &scale_out_info, BorderSize(inner_border_right, inner_border_top), info));
@@ -91,7 +97,7 @@ Status CLDeconvolutionLayer::validate(const ITensorInfo *input, const ITensorInf
     return Status{};
 }
 
-void CLDeconvolutionLayer::configure(ICLTensor *input, const ICLTensor *weights, const ICLTensor *bias, ICLTensor *output, const PadStrideInfo &info,
+void CLDeconvolutionLayer::configure(ICLTensor *input, ICLTensor *weights, const ICLTensor *bias, ICLTensor *output, const PadStrideInfo &info,
                                      unsigned int inner_border_right, unsigned int inner_border_top, const WeightsInfo &weights_info)
 {
     ARM_COMPUTE_ERROR_ON_NULLPTR(input, weights, output);
@@ -99,8 +105,12 @@ void CLDeconvolutionLayer::configure(ICLTensor *input, const ICLTensor *weights,
     const unsigned int stride_x = info.stride().first;
     const unsigned int stride_y = info.stride().second;
 
+    _weights = weights;
+    _weights_flipped.allocator()->init(TensorInfo(weights->info()->tensor_shape(), 1, weights->info()->data_type()));
+    _flip_weights.configure(weights, &_weights_flipped);
+
     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, inner_border_right, inner_border_top, stride_x, stride_y);
+                                                    info.pad().first, info.pad().second, stride_x, stride_y);
 
     const TensorShape output_shape = deconvolution_output_shape(out_dims, input->info()->tensor_shape(), weights->info()->tensor_shape());
 
@@ -113,22 +123,31 @@ void CLDeconvolutionLayer::configure(ICLTensor *input, const ICLTensor *weights,
     _is_prepared = false;
 
     _memory_group.manage(&_scaled_output);
+    _memory_group.manage(&_weights_flipped);
 
-    // configure scale function
-    // Init and allocate intermediate 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());
-    const unsigned int out_x = input->info()->dimension(0) + (input->info()->dimension(0) - 1) * (stride_x - 1) + inner_border_right + 2 * info.pad().first;
-    const unsigned int out_y = input->info()->dimension(1) + (input->info()->dimension(1) - 1) * (stride_y - 1) + inner_border_top + 2 * info.pad().second;
+    // Find the upsampled dimensions
+    unsigned int out_x = (input->info()->dimension(0) - 1) * stride_x + inner_border_right + 1;
+    unsigned int out_y = (input->info()->dimension(1) - 1) * stride_y + inner_border_top + 1;
+
+    // Find the padding needed for the convolution with stride 1 in order to match output shape
+    unsigned int padx = out_dims.first - (out_x - weights->info()->dimension(0) + 1);
+    unsigned int pady = out_dims.second - (out_y - weights->info()->dimension(1) + 1);
+    out_x += padx;
+    out_y += pady;
+
+    TensorShape scale_out_shape(input->info()->tensor_shape());
     scale_out_shape.set(0, out_x);
     scale_out_shape.set(1, out_y);
     TensorInfo scale_out_info(scale_out_shape, 1, input->info()->data_type(), input->info()->quantization_info());
     _scaled_output.allocator()->init(scale_out_info);
 
-    _scale_f.configure(input, &_scaled_output, BorderSize(inner_border_top, inner_border_right), info);
+    // configure scale function
+    const PadStrideInfo upsample_info(stride_x, stride_y, padx / 2, pady / 2);
+    _scale_f.configure(input, &_scaled_output, BorderSize(inner_border_top, inner_border_right), upsample_info);
 
     // setup the function to convolve the upscaled output
     const PadStrideInfo conv_info(1, 1, 0, 0, 0, 0, DimensionRoundingType::CEIL);
-    _conv_f.configure(&_scaled_output, weights, bias, output, conv_info, weights_info);
+    _conv_f.configure(&_scaled_output, &_weights_flipped, bias, output, conv_info, weights_info);
     _scaled_output.allocator()->allocate();
 }
 
@@ -148,7 +167,14 @@ void CLDeconvolutionLayer::prepare()
 {
     if(!_is_prepared)
     {
+        _weights_flipped.allocator()->allocate();
+        _weights_flipped.map(true);
+        _weights->map(CLScheduler::get().queue(), true);
+        CPPScheduler::get().schedule(&_flip_weights, Window::DimZ);
+        _weights_flipped.unmap();
+        _weights->unmap(CLScheduler::get().queue());
         _conv_f.prepare();
+
         _is_prepared = true;
     }
 }
diff --git a/src/runtime/NEON/functions/NEDeconvolutionLayer.cpp b/src/runtime/NEON/functions/NEDeconvolutionLayer.cpp
index fda9f57499..6ca60c66a4 100644
--- a/src/runtime/NEON/functions/NEDeconvolutionLayer.cpp
+++ b/src/runtime/NEON/functions/NEDeconvolutionLayer.cpp
@@ -27,6 +27,7 @@
 #include "arm_compute/core/Utils.h"
 #include "arm_compute/core/Validate.h"
 #include "arm_compute/core/utils/misc/ShapeCalculator.h"
+#include "arm_compute/runtime/CPP/CPPScheduler.h"
 
 using namespace arm_compute;
 using namespace arm_compute::misc::shape_calculator;
@@ -35,7 +36,9 @@ NEDeconvolutionLayer::NEDeconvolutionLayer(std::shared_ptr<IMemoryManager> memor
     : _memory_group(std::move(memory_manager)),
       _conv_f(),
       _upsample_f(),
+      _flip_weights(),
       _scaled_output(),
+      _weights_flipped(),
       _input(nullptr),
       _info(),
       _inner_border(),
@@ -60,7 +63,7 @@ Status NEDeconvolutionLayer::validate(const ITensorInfo *input, const ITensorInf
     ARM_COMPUTE_RETURN_ERROR_ON_MSG(inner_border_top > stride_y - 1, "inner_border_top must be smaller than stride_y");
 
     auto out_dims = deconvolution_output_dimensions(input->dimension(0), input->dimension(1), weights->dimension(0), weights->dimension(1),
-                                                    info.pad().first, info.pad().second, inner_border_right, inner_border_top, stride_x, stride_y);
+                                                    info.pad().first, info.pad().second, stride_x, stride_y);
 
     ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights, bias);
 
@@ -74,14 +77,14 @@ Status NEDeconvolutionLayer::validate(const ITensorInfo *input, const ITensorInf
         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
 
         const TensorShape output_shape = deconvolution_output_shape(out_dims, input->tensor_shape(), weights->tensor_shape());
-
         ARM_COMPUTE_RETURN_ERROR_ON_MSG(output->dimension(Window::DimX) != output_shape.x(), "Output's width is invalid.");
         ARM_COMPUTE_RETURN_ERROR_ON_MSG(output->dimension(Window::DimY) != output_shape.y(), "Output's height is invalid.");
         ARM_COMPUTE_RETURN_ERROR_ON_MSG(output->dimension(Window::DimZ) != output_shape.z(), "Output's depth is invalid.");
     }
 
-    TensorInfo scale_out_info(input->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(compute_deconvolution_shape(*input, stride_x, stride_y, inner_border_right, inner_border_top,
-                                                                                                      info)));
+    TensorInfo scale_out_info(input->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(compute_deconvolution_shape(*input, *weights, stride_x, stride_y, inner_border_right,
+                                                                                                      inner_border_top,
+                                                                                                      out_dims)));
     const PadStrideInfo conv_info(1, 1, 0, 0, 0, 0, DimensionRoundingType::CEIL);
 
     for(size_t i = 2; i < Coordinates::num_max_dimensions; ++i)
@@ -107,25 +110,44 @@ void NEDeconvolutionLayer::configure(ITensor *input, const ITensor *weights, con
     const unsigned int stride_x = info.stride().first;
     const unsigned int stride_y = info.stride().second;
 
+    _weights_flipped.allocator()->init(TensorInfo(weights->info()->tensor_shape(), 1, weights->info()->data_type()));
+    _flip_weights.configure(weights, &_weights_flipped);
+
+    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, stride_x, stride_y);
+
+    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()->quantization_info());
+
     // Perform validation step
     ARM_COMPUTE_ERROR_THROW_ON(NEDeconvolutionLayer::validate(input->info(), weights->info(), bias == nullptr ? nullptr : bias->info(), output->info(), info, inner_border_right, inner_border_top));
 
     _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
-    const TensorInfo scale_out_info(compute_deconvolution_shape(*input->info(), stride_x, stride_y, inner_border_right, inner_border_top, info), 1, input->info()->data_type());
+    // Find the upsampled dimensions
+    unsigned int out_x = (input->info()->dimension(0) - 1) * stride_x + inner_border_right + 1;
+    unsigned int out_y = (input->info()->dimension(1) - 1) * stride_y + inner_border_top + 1;
+
+    // Find the padding needed for the convolution with stride 1 in order to match output shape
+    unsigned int padx = out_dims.first - (out_x - weights->info()->dimension(0) + 1);
+    unsigned int pady = out_dims.second - (out_y - weights->info()->dimension(1) + 1);
+    out_x += padx;
+    out_y += pady;
+
+    TensorShape scale_out_shape(input->info()->tensor_shape());
+    scale_out_shape.set(0, out_x);
+    scale_out_shape.set(1, out_y);
+    TensorInfo scale_out_info(scale_out_shape, 1, input->info()->data_type(), input->info()->quantization_info());
     _scaled_output.allocator()->init(scale_out_info);
 
+    const PadStrideInfo upsample_info(stride_x, stride_y, padx / 2, pady / 2);
+    _upsample_f.configure(input, &_scaled_output, upsample_info, inner_border_right, inner_border_top);
+
     // setup the function to convolve the upscaled output
     const PadStrideInfo conv_info(1, 1, 0, 0, 0, 0, DimensionRoundingType::CEIL);
-    _conv_f.configure(&_scaled_output, weights, bias, output, conv_info);
-
-    // Allocate auxiliary tensors
+    _conv_f.configure(&_scaled_output, &_weights_flipped, bias, output, conv_info);
     _scaled_output.allocator()->allocate();
-
-    // configure upsample function
-    _upsample_f.configure(input, &_scaled_output, info, inner_border_right, inner_border_top);
 }
 
 void NEDeconvolutionLayer::run()
@@ -144,6 +166,8 @@ void NEDeconvolutionLayer::prepare()
 {
     if(!_is_prepared)
     {
+        _weights_flipped.allocator()->allocate();
+        CPPScheduler::get().schedule(&_flip_weights, Window::DimZ);
         _conv_f.prepare();
         _is_prepared = true;
     }
diff --git a/tests/validation/CL/DeconvolutionLayer.cpp b/tests/validation/CL/DeconvolutionLayer.cpp
index 2aa7cfe7c1..84a2b01797 100644
--- a/tests/validation/CL/DeconvolutionLayer.cpp
+++ b/tests/validation/CL/DeconvolutionLayer.cpp
@@ -48,7 +48,7 @@ RelativeTolerance<half_float::half> tolerance_f16(half_float::half(0.2)); /**< T
 constexpr AbsoluteTolerance<float>  tolerance_qasymm8(1.0);               /**< Tolerance value for comparing reference's output against implementation's output for quantized data types */
 constexpr float                     tolerance_num = 0.07f;                /**< Tolerance number */
 
-const auto data4x4 = datasets::SmallDeconvolutionShapes() * framework::dataset::make("StrideX", 1, 4) * framework::dataset::make("StrideY", 1, 4) * framework::dataset::make("PadX", 0, 3)
+const auto data4x4 = datasets::SmallDeconvolutionShapes() * framework::dataset::make("StrideX", 1, 5) * framework::dataset::make("StrideY", 1, 5) * framework::dataset::make("PadX", 0, 3)
                      * framework::dataset::make("PadY", 0, 3) * framework::dataset::make("ax", 0) * framework::dataset::make("ay", 0) * framework::dataset::make("NumKernels", { 1, 3 });
 
 const auto data3x3 = datasets::SmallDeconvolutionShapes() * framework::dataset::make("StrideX", 1, 4) * framework::dataset::make("StrideY", 1, 4) * framework::dataset::make("PadX", 0, 2)
@@ -71,7 +71,7 @@ DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, (combine(datasets::Sm
     const unsigned int num_kernels   = 1;
     const TensorShape  weights_shape(kernel_size_x, kernel_size_y, input_shape.z(), num_kernels);
     const TensorShape  bias_shape(num_kernels);
-    auto               out_dim      = deconvolution_output_dimensions(input_shape.x(), input_shape.y(), kernel_size_x, kernel_size_y, 1, 1, 0, 0, 1, 1);
+    auto               out_dim      = deconvolution_output_dimensions(input_shape.x(), input_shape.y(), kernel_size_x, kernel_size_y, 1, 1, 1, 1);
     TensorShape        output_shape = deconvolution_output_shape(out_dim, input_shape, weights_shape);
 
     // Create tensors
diff --git a/tests/validation/NEON/DeconvolutionLayer.cpp b/tests/validation/NEON/DeconvolutionLayer.cpp
index 277953badb..eb643b8e7c 100644
--- a/tests/validation/NEON/DeconvolutionLayer.cpp
+++ b/tests/validation/NEON/DeconvolutionLayer.cpp
@@ -67,7 +67,7 @@ DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, (combine(datasets::Sm
     const unsigned int num_kernels   = 1;
     const TensorShape  weights_shape(kernel_size_x, kernel_size_y, input_shape.z(), num_kernels);
     const TensorShape  bias_shape(num_kernels);
-    auto               out_dim      = deconvolution_output_dimensions(input_shape.x(), input_shape.y(), kernel_size_x, kernel_size_y, 1, 1, 0, 0, 1, 1);
+    auto               out_dim      = deconvolution_output_dimensions(input_shape.x(), input_shape.y(), kernel_size_x, kernel_size_y, 1, 1, 1, 1);
     TensorShape        output_shape = deconvolution_output_shape(out_dim, input_shape, weights_shape);
 
     // Create tensors
diff --git a/tests/validation/fixtures/DeconvolutionLayerFixture.h b/tests/validation/fixtures/DeconvolutionLayerFixture.h
index 6a8b4f220c..d3a7be74b0 100644
--- a/tests/validation/fixtures/DeconvolutionLayerFixture.h
+++ b/tests/validation/fixtures/DeconvolutionLayerFixture.h
@@ -160,7 +160,7 @@ public:
         const TensorShape   bias_shape(num_kernels);
         const PadStrideInfo info(sx, sy, padx, pady, DimensionRoundingType::CEIL);
         const std::pair<unsigned int, unsigned int> inner_border(inner_border_right, inner_border_top);
-        auto        out_dim      = deconvolution_output_dimensions(input_shape.x(), input_shape.y(), kernel_size_x, kernel_size_y, padx, pady, inner_border.first, inner_border.second, sx, sy);
+        auto        out_dim      = deconvolution_output_dimensions(input_shape.x(), input_shape.y(), kernel_size_x, kernel_size_y, padx, pady, sx, sy);
         TensorShape output_shape = deconvolution_output_shape(out_dim, input_shape, weights_shape);
         DeconvolutionLayerFixtureBase<TensorType, AccessorType, FunctionType, T>::setup(input_shape, weights_shape, bias_shape, output_shape, info, inner_border, data_type, QuantizationInfo());
     }
@@ -179,7 +179,7 @@ public:
         const TensorShape   bias_shape(num_kernels);
         const PadStrideInfo info(sx, sy, padx, pady, DimensionRoundingType::CEIL);
         const std::pair<unsigned int, unsigned int> inner_border(inner_border_right, inner_border_top);
-        auto        out_dim      = deconvolution_output_dimensions(input_shape.x(), input_shape.y(), kernel_size_x, kernel_size_y, padx, pady, inner_border.first, inner_border.second, sx, sy);
+        auto        out_dim      = deconvolution_output_dimensions(input_shape.x(), input_shape.y(), kernel_size_x, kernel_size_y, padx, pady, sx, sy);
         TensorShape output_shape = deconvolution_output_shape(out_dim, input_shape, weights_shape);
         DeconvolutionLayerFixtureBase<TensorType, AccessorType, FunctionType, T>::setup(input_shape, weights_shape, bias_shape, output_shape, info, inner_border, data_type, quantization_info);
     }
diff --git a/tests/validation/reference/DeconvolutionLayer.cpp b/tests/validation/reference/DeconvolutionLayer.cpp
index ba28b46d3a..5ca3b44baa 100644
--- a/tests/validation/reference/DeconvolutionLayer.cpp
+++ b/tests/validation/reference/DeconvolutionLayer.cpp
@@ -38,11 +38,23 @@ SimpleTensor<T> deconvolution_layer(const SimpleTensor<T> &src, const SimpleTens
                                     const PadStrideInfo &info, const std::pair<unsigned int, unsigned int> &a)
 {
     // Create reference
-    const int   stride_x     = info.stride().first;
-    const int   stride_y     = info.stride().second;
+    const int stride_x           = info.stride().first;
+    const int stride_y           = info.stride().second;
+    const int weights_width      = weights.shape().x();
+    const int weights_height     = weights.shape().y();
+    const int weights_upper_dims = weights.shape().total_size() / (weights_width * weights_height);
+
+    // Find the upsampled dimensions
+    unsigned int out_x = (src.shape().x() - 1) * stride_x + a.first + 1;
+    unsigned int out_y = (src.shape().y() - 1) * stride_y + a.second + 1;
+
+    // Find the padding needed for the convolution with stride 1 in order to match output shape
+    unsigned int padx = output_shape.x() - (out_x - weights_width + 1);
+    unsigned int pady = output_shape.y() - (out_y - weights_height + 1);
+    out_x += padx;
+    out_y += pady;
+
     TensorShape scaled_shape = src.shape();
-    int         out_x        = src.shape().x() + (src.shape().x() - 1) * (stride_x - 1) + a.first + 2 * info.pad().first;
-    int         out_y        = src.shape().y() + (src.shape().y() - 1) * (stride_y - 1) + a.second + 2 * info.pad().second;
     scaled_shape.set(0, out_x);
     scaled_shape.set(1, out_y);
     SimpleTensor<T> scaled{ scaled_shape, src.data_type(), 1, src.quantization_info() };
@@ -69,14 +81,28 @@ SimpleTensor<T> deconvolution_layer(const SimpleTensor<T> &src, const SimpleTens
         std::fill_n(scaled.data(), scaled.num_elements(), T(0));
     }
 
+    // Flip weights by 180 degrees
+    SimpleTensor<T> weights_flipped{ weights.shape(), weights.data_type(), 1, weights.quantization_info() };
+    for(int ud = 0; ud < weights_upper_dims; ++ud)
+    {
+        const int offset = ud * weights_width * weights_height;
+        for(int y = 0; y < weights_height; ++y)
+        {
+            for(int x = 0; x < weights_width; ++x)
+            {
+                weights_flipped[offset + (weights_height - 1 - y) * weights_width + (weights_width - 1 - x)] = weights[offset + y * weights_width + x];
+            }
+        }
+    }
+
     for(int slice = 0; slice < num_2d_slices; ++slice)
     {
         const int offset_slice_in  = slice * width_in * height_in;
         const int offset_slice_out = slice * width_scaled * height_scaled;
-        const int start_x          = info.pad().first;
-        const int start_y          = ay + info.pad().second;
-        const int end_y            = height_scaled - info.pad().second;
-        const int end_x            = width_scaled - ax - info.pad().first;
+        const int start_x          = padx / 2;
+        const int start_y          = ay + pady / 2;
+        const int end_y            = height_scaled - pady / 2;
+        const int end_x            = width_scaled - ax - padx / 2;
 
         for(int yi = start_y, in_y = 0; yi < end_y; yi += stride_y, in_y++)
         {
@@ -90,7 +116,7 @@ SimpleTensor<T> deconvolution_layer(const SimpleTensor<T> &src, const SimpleTens
     }
 
     const PadStrideInfo conv_info(1, 1, 0, 0, 0, 0, DimensionRoundingType::CEIL);
-    return convolution_layer(scaled, weights, bias, output_shape, conv_info);
+    return convolution_layer(scaled, weights_flipped, bias, output_shape, conv_info);
 }
 
 template SimpleTensor<uint8_t> deconvolution_layer(const SimpleTensor<uint8_t> &src, const SimpleTensor<uint8_t> &weights, const SimpleTensor<int32_t> &bias, const TensorShape &output_shape,