COMPMID-471 Implement Deconvolution on OpenCL

Change-Id: Ie00c6b08a51d30c5ce2637d40ee3d165b8a68686
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/110311
Reviewed-by: Pablo Tello <pablo.tello@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
Tested-by: Jenkins <bsgcomp@arm.com>

6 files changed, 245 insertions, 78 deletions

diff --git a/tests/datasets/ShapeDatasets.h b/tests/datasets/ShapeDatasets.h
index 58fba07bf8..a5e03c737f 100644
--- a/tests/datasets/ShapeDatasets.h
+++ b/tests/datasets/ShapeDatasets.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017 ARM Limited.
+ * Copyright (c) 2017, 2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -239,7 +239,7 @@ public:
     SmallDeconvolutionShapes()
         : ShapeDataset("InputShape",
     {
-        TensorShape{ 2U, 3U, 3U, 2U },
+        TensorShape{ 4U, 3U, 3U, 2U },
                      TensorShape{ 5U, 5U, 3U },
                      TensorShape{ 11U, 13U, 4U, 3U }
     })
diff --git a/tests/validation/CL/DeconvolutionLayer.cpp b/tests/validation/CL/DeconvolutionLayer.cpp
new file mode 100644
index 0000000000..59e85537e5
--- /dev/null
+++ b/tests/validation/CL/DeconvolutionLayer.cpp
@@ -0,0 +1,192 @@
+/*
+ * Copyright (c) 2017, 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/CL/kernels/CLFillBorderKernel.h"
+#include "arm_compute/core/Types.h"
+#include "arm_compute/runtime/CL/CLTensor.h"
+#include "arm_compute/runtime/CL/CLTensorAllocator.h"
+#include "arm_compute/runtime/CL/functions/CLDeconvolutionLayer.h"
+#include "tests/CL/CLAccessor.h"
+#include "tests/PaddingCalculator.h"
+#include "tests/datasets/ShapeDatasets.h"
+#include "tests/framework/Asserts.h"
+#include "tests/framework/Macros.h"
+#include "tests/framework/datasets/Datasets.h"
+#include "tests/validation/Validation.h"
+#include "tests/validation/fixtures/DeconvolutionLayerFixture.h"
+
+namespace arm_compute
+{
+namespace test
+{
+namespace validation
+{
+namespace
+{
+constexpr AbsoluteTolerance<float> tolerance_fp32(0.001f); /**< Tolerance for floating point tests */
+
+const auto data3x3 = datasets::SmallDeconvolutionShapes() * framework::dataset::make("StrideX", 1, 4) * framework::dataset::make("StrideY", 1, 4) * framework::dataset::make("PadX", 0, 2)
+                     * framework::dataset::make("PadY", 0, 2) * framework::dataset::make("ax", 0) * framework::dataset::make("ay", 0) * framework::dataset::make("NumKernels", { 1, 3 });
+
+const auto data1x1 = datasets::SmallDeconvolutionShapes() * framework::dataset::make("StrideX", 1, 4) * framework::dataset::make("StrideY", 1, 4) * framework::dataset::make("PadX", 0, 1)
+                     * framework::dataset::make("PadY", 0, 1) * framework::dataset::make("ax", 0) * framework::dataset::make("ay", 0) * framework::dataset::make("NumKernels", { 1, 3 });
+
+} // namespace
+
+TEST_SUITE(CL)
+TEST_SUITE(DeconvolutionLayer)
+
+DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, (combine(datasets::SmallDeconvolutionShapes(), framework::dataset::make("DataType", DataType::F32))),
+               input_shape, data_type)
+{
+    // Create shapes
+    const unsigned int kernel_size_x = 3;
+    const unsigned int kernel_size_y = 3;
+    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);
+    TensorShape        output_shape = deconvolution_output_shape(out_dim, input_shape, weights_shape);
+
+    // Create tensors
+    CLTensor src     = create_tensor<CLTensor>(input_shape, data_type, 1);
+    CLTensor weights = create_tensor<CLTensor>(weights_shape, data_type, 1);
+    CLTensor bias    = create_tensor<CLTensor>(bias_shape, data_type, 1);
+    CLTensor dst     = create_tensor<CLTensor>(output_shape, data_type, 1);
+
+    ARM_COMPUTE_EXPECT(src.info()->is_resizable(), framework::LogLevel::ERRORS);
+    ARM_COMPUTE_EXPECT(weights.info()->is_resizable(), framework::LogLevel::ERRORS);
+    ARM_COMPUTE_EXPECT(bias.info()->is_resizable(), framework::LogLevel::ERRORS);
+    ARM_COMPUTE_EXPECT(dst.info()->is_resizable(), framework::LogLevel::ERRORS);
+
+    // Create and configure function
+    CLDeconvolutionLayer deconv;
+    deconv.configure(&src, &weights, &bias, &dst, PadStrideInfo(1, 1, 1, 1, DimensionRoundingType::CEIL), 0, 0);
+
+    // Validate valid region
+    const ValidRegion src_valid_region     = shape_to_valid_region(input_shape);
+    const ValidRegion weights_valid_region = shape_to_valid_region(weights_shape);
+    const ValidRegion bias_valid_region    = shape_to_valid_region(bias_shape);
+    const ValidRegion dst_valid_region     = shape_to_valid_region(output_shape);
+
+    validate(src.info()->valid_region(), src_valid_region);
+    validate(weights.info()->valid_region(), weights_valid_region);
+    validate(bias.info()->valid_region(), bias_valid_region);
+    validate(dst.info()->valid_region(), dst_valid_region);
+}
+
+// *INDENT-OFF*
+// clang-format off
+DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(zip(zip(zip(zip(
+    framework::dataset::make("InputInfo", { TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32, 0),   // Mismatching data type
+                                            TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32, 0),   // Invalid weights shape
+                                            TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::QS8, 4),   // Non supported data type
+                                            TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32, 11),  // Invalid bias shape
+                                            TensorInfo(TensorShape(13U, 11U, 4U, 3U), 1, DataType::F32, 0), // Window shrink
+                                            TensorInfo(TensorShape(32U, 16U, 2U), 1, DataType::F32, 0),
+                                          }),
+    framework::dataset::make("WeightsInfo", { TensorInfo(TensorShape(3U, 3U, 2U, 2U), 1, DataType::F16, 0),
+                                            TensorInfo(TensorShape(3U, 3U, 2U, 4U), 1, DataType::F32, 0),
+                                            TensorInfo(TensorShape(3U, 3U, 2U, 2U), 1, DataType::QS8, 5),
+                                            TensorInfo(TensorShape(3U, 2U, 2U, 2U), 1, DataType::F32, 11),
+                                            TensorInfo(TensorShape(3U, 3U, 4U), 1, DataType::F32, 0),
+                                              TensorInfo(TensorShape(1U, 1U, 2U, 4U), 1, DataType::F32, 0),
+                                          })),
+    framework::dataset::make("BiasInfo",  { TensorInfo(TensorShape(1U), 1, DataType::F16, 0),
+                                            TensorInfo(TensorShape(1U), 1, DataType::F32, 0),
+                                            TensorInfo(TensorShape(1U), 1, DataType::F32, 5),
+                                            TensorInfo(TensorShape(25U, 11U), 1, DataType::F32, 11),
+                                            TensorInfo(TensorShape(1U), 1, DataType::F32, 0),
+                                            TensorInfo(TensorShape(4U), 1, DataType::F32, 0),
+                                          })),
+    framework::dataset::make("OutputInfo",{ TensorInfo(TensorShape(25U, 11U, 2U), 1, DataType::F16, 0),
+                                            TensorInfo(TensorShape(25U, 10U, 2U), 1, DataType::F32, 0),
+                                            TensorInfo(TensorShape(25U, 11U, 2U), 1, DataType::F32, 5),
+                                            TensorInfo(TensorShape(13U, 13U, 2U), 1, DataType::F32, 0),
+                                            TensorInfo(TensorShape(11U, 9U, 1U, 3U), 1, DataType::F32, 0),
+                                            TensorInfo(TensorShape(32U, 16U, 4U), 1, DataType::F32, 0),
+                                          })),
+    framework::dataset::make("PadStrideInfo", { PadStrideInfo(1, 1, 0, 0),
+                                                PadStrideInfo(1, 1, 0, 0),
+                                                PadStrideInfo(1, 1, 0, 0),
+                                                PadStrideInfo(1, 1, 0, 0),
+                                                PadStrideInfo(1, 1, 1, 1),
+                                                PadStrideInfo(1, 1, 0, 0),
+                                           })),
+    framework::dataset::make("ax",          {   1U,
+                                                1U,
+                                                1U,
+                                                1U,
+                                                0U,
+                                                0U,
+                                            })),
+   framework::dataset::make("ay",           {   1U,
+                                                1U,
+                                                1U,
+                                                1U,
+                                                0U,
+                                                0U,
+                                            })),
+    framework::dataset::make("Expected", { false, false, false, false, false, true })),
+    input_info, weights_info, bias_info, output_info, pad_info, ax, ay, expected)
+{
+    bool is_valid = bool(CLDeconvolutionLayer::validate(&input_info.clone()->set_is_resizable(false), &weights_info.clone()->set_is_resizable(false), &bias_info.clone()->set_is_resizable(false), &output_info.clone()->set_is_resizable(false), pad_info, ax, ay));
+    ARM_COMPUTE_EXPECT(is_valid == expected, framework::LogLevel::ERRORS);
+}
+// clang-format on
+// *INDENT-ON*
+
+template <typename T>
+using CLDeconvolutionLayerFixture3x3 = DeconvolutionValidationFixture<CLTensor, CLAccessor, CLDeconvolutionLayer, T, 3, 3>;
+
+template <typename T>
+using CLDeconvolutionLayerFixture1x1 = DeconvolutionValidationFixture<CLTensor, CLAccessor, CLDeconvolutionLayer, T, 1, 1>;
+
+TEST_SUITE(Float)
+
+TEST_SUITE(FP32)
+TEST_SUITE(W3x3)
+
+FIXTURE_DATA_TEST_CASE(Run, CLDeconvolutionLayerFixture3x3<float>, framework::DatasetMode::ALL, combine(data3x3, framework::dataset::make("DataType", DataType::F32)))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference, tolerance_fp32);
+}
+TEST_SUITE_END()
+
+TEST_SUITE(W1x1)
+FIXTURE_DATA_TEST_CASE(Run, CLDeconvolutionLayerFixture1x1<float>, framework::DatasetMode::ALL, combine(data1x1, framework::dataset::make("DataType", DataType::F32)))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference, tolerance_fp32);
+}
+TEST_SUITE_END()
+
+TEST_SUITE_END()
+TEST_SUITE_END()
+
+TEST_SUITE_END()
+TEST_SUITE_END()
+} // namespace validation
+} // namespace test
+} // namespace arm_compute
diff --git a/tests/validation/NEON/DeconvolutionLayer.cpp b/tests/validation/NEON/DeconvolutionLayer.cpp
index 751a96558a..9573784d86 100644
--- a/tests/validation/NEON/DeconvolutionLayer.cpp
+++ b/tests/validation/NEON/DeconvolutionLayer.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017 ARM Limited.
+ * Copyright (c) 2017, 2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -44,15 +44,11 @@ namespace
 {
 constexpr AbsoluteTolerance<float> tolerance_fp32(0.001f); /**< Tolerance for floating point tests */
 
-const auto data3x3 = datasets::SmallDeconvolutionShapes() * framework::dataset::make("StrideX", 1, 4) * framework::dataset::make("StrideY", 1, 4) * framework::dataset::make("PadX", 0,
-                     2)
-                     * framework::dataset::make("PadY", 0, 2) * framework::dataset::make("ax", 1, 3) * framework::dataset::make("ay", 1, 3) * framework::dataset::make("NumKernels", { 1, 3 })
-                     *framework::dataset::make("ux", 1, 4) *framework::dataset::make("uy", 1, 4);
+const auto data3x3 = datasets::SmallDeconvolutionShapes() * framework::dataset::make("StrideX", 1, 4) * framework::dataset::make("StrideY", 1, 4) * framework::dataset::make("PadX", 0, 2)
+                     * framework::dataset::make("PadY", 0, 2) * framework::dataset::make("ax", 0) * framework::dataset::make("ay", 0) * framework::dataset::make("NumKernels", { 1, 3 });
 
-const auto data1x1 = datasets::SmallDeconvolutionShapes() * framework::dataset::make("StrideX", 1, 4) * framework::dataset::make("StrideY", 1, 4) * framework::dataset::make("PadX", 0,
-                     1)
-                     * framework::dataset::make("PadY", 0, 1) * framework::dataset::make("ax", 1, 3) * framework::dataset::make("ay", 1, 3) * framework::dataset::make("NumKernels", { 1, 3 })
-                     *framework::dataset::make("ux", 1, 4) *framework::dataset::make("uy", 1, 4);
+const auto data1x1 = datasets::SmallDeconvolutionShapes() * framework::dataset::make("StrideX", 1, 4) * framework::dataset::make("StrideY", 1, 4) * framework::dataset::make("PadX", 0, 1)
+                     * framework::dataset::make("PadY", 0, 1) * framework::dataset::make("ax", 0) * framework::dataset::make("ay", 0) * framework::dataset::make("NumKernels", { 1, 3 });
 
 } // namespace
 
diff --git a/tests/validation/fixtures/DeconvolutionLayerFixture.h b/tests/validation/fixtures/DeconvolutionLayerFixture.h
index e98f5e93c0..f2455f31ac 100644
--- a/tests/validation/fixtures/DeconvolutionLayerFixture.h
+++ b/tests/validation/fixtures/DeconvolutionLayerFixture.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017 ARM Limited.
+ * Copyright (c) 2017, 2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -43,20 +43,15 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ
 class DeconvolutionLayerFixtureBase : public framework::Fixture
 {
 public:
-    /*
-     *
-     * @param[in] a The number of zeros added to right and bottom edges of the input.
-     * @param[in] u How much to scale the X and Y axis.
-     */
     template <typename...>
     void setup(TensorShape input_shape, TensorShape weights_shape, TensorShape bias_shape, TensorShape output_shape, PadStrideInfo info,
-               const std::pair<unsigned int, unsigned int> &a, const std::pair<unsigned int, unsigned int> &u, DataType data_type, int fractional_bits)
+               const std::pair<unsigned int, unsigned int> &inner_border, DataType data_type, int fractional_bits)
     {
         _fractional_bits = fractional_bits;
         _data_type       = data_type;
 
-        _target    = compute_target(input_shape, weights_shape, bias_shape, output_shape, info, a, u, data_type, fractional_bits);
-        _reference = compute_reference(input_shape, weights_shape, bias_shape, output_shape, info, a, data_type, fractional_bits);
+        _target    = compute_target(input_shape, weights_shape, bias_shape, output_shape, info, inner_border, data_type, fractional_bits);
+        _reference = compute_reference(input_shape, weights_shape, bias_shape, output_shape, info, inner_border, data_type, fractional_bits);
     }
 
 protected:
@@ -75,13 +70,9 @@ protected:
                 library->fill_tensor_uniform(tensor, i);
         }
     }
-    /*
-     *
-     * @param[in] a The number of zeros added to right and bottom edges of the input.
-     * @param[in] u How much to scale the X and Y axis.
-     */
+
     TensorType compute_target(const TensorShape &input_shape, const TensorShape &weights_shape, const TensorShape &bias_shape, const TensorShape &output_shape,
-                              const PadStrideInfo &info, const std::pair<unsigned int, unsigned int> &a, const std::pair<float, float> &u, DataType data_type, int fixed_point_position)
+                              const PadStrideInfo &info, const std::pair<unsigned int, unsigned int> &inner_border, DataType data_type, int fixed_point_position)
     {
         // Create tensors
         TensorType src     = create_tensor<TensorType>(input_shape, data_type, 1, fixed_point_position);
@@ -91,7 +82,7 @@ protected:
 
         // Create and configure function
         FunctionType conv;
-        conv.configure(&src, &weights, &bias, &dst, info, a.first, a.second, u.first, u.second);
+        conv.configure(&src, &weights, &bias, &dst, info, inner_border.first, inner_border.second);
 
         ARM_COMPUTE_EXPECT(src.info()->is_resizable(), framework::LogLevel::ERRORS);
         ARM_COMPUTE_EXPECT(weights.info()->is_resizable(), framework::LogLevel::ERRORS);
@@ -121,7 +112,7 @@ protected:
     }
 
     SimpleTensor<T> compute_reference(const TensorShape &input_shape, const TensorShape &weights_shape, const TensorShape &bias_shape, const TensorShape &output_shape,
-                                      const PadStrideInfo &info, const std::pair<unsigned int, unsigned int> a, DataType data_type, int fixed_point_position)
+                                      const PadStrideInfo &info, const std::pair<unsigned int, unsigned int> inner_border, DataType data_type, int fixed_point_position)
     {
         // Create reference
         SimpleTensor<T> src{ input_shape, data_type, 1, fixed_point_position };
@@ -133,7 +124,7 @@ protected:
         fill(weights, 1);
         fill(bias, 2);
 
-        return reference::deconvolution_layer<T>(src, weights, bias, output_shape, info, a);
+        return reference::deconvolution_layer<T>(src, weights, bias, output_shape, info, inner_border);
     }
 
     TensorType      _target{};
@@ -148,18 +139,16 @@ class DeconvolutionValidationFixture : public DeconvolutionLayerFixtureBase<Tens
 public:
     template <typename...>
     void setup(TensorShape input_shape, unsigned int sx, unsigned int sy, unsigned int padx, unsigned int pady,
-               unsigned int ax, unsigned int ay, unsigned int ux, unsigned int uy, unsigned int num_kernels, DataType data_type)
+               unsigned int inner_border_right, unsigned int inner_border_top, unsigned int num_kernels, DataType data_type)
     {
         ARM_COMPUTE_ERROR_ON_MSG(kernel_size_x != kernel_size_y, "Only square kernels supported");
         const TensorShape   weights_shape(kernel_size_x, kernel_size_y, input_shape.z(), num_kernels);
         const TensorShape   bias_shape(num_kernels);
         const PadStrideInfo info(sx, sy, padx, pady, DimensionRoundingType::CEIL);
-        const std::pair<unsigned int, unsigned int> a(ax, ay);
-        const std::pair<float, float>               u(ux, uy);
-        auto out_dim = deconvolution_output_dimensions(input_shape.x(), input_shape.y(), kernel_size_x, kernel_size_y, padx, pady, a.first, a.second, u.first, u.second,
-                                                       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);
         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, a, u, data_type, 0);
+        DeconvolutionLayerFixtureBase<TensorType, AccessorType, FunctionType, T>::setup(input_shape, weights_shape, bias_shape, output_shape, info, inner_border, data_type, 0);
     }
 };
 
diff --git a/tests/validation/reference/DeconvolutionLayer.cpp b/tests/validation/reference/DeconvolutionLayer.cpp
index 82c2188ade..0cf1087346 100644
--- a/tests/validation/reference/DeconvolutionLayer.cpp
+++ b/tests/validation/reference/DeconvolutionLayer.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017 ARM Limited.
+ * Copyright (c) 2017, 2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -39,26 +39,27 @@ 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;
     TensorShape scaled_shape = src.shape();
-    scaled_shape.set(0, output_shape.x());
-    scaled_shape.set(1, output_shape.y());
+    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.fixed_point_position() };
 
-    const int          width_in      = src.shape().x();
-    const int          height_in     = src.shape().y();
-    const int          width_scaled  = scaled.shape().x();
-    const int          height_scaled = scaled.shape().y();
-    const int          num_2d_slices = src.shape().total_size() / (width_in * height_in);
-    const float        width_ratio   = static_cast<float>(width_in) / static_cast<float>(width_scaled);
-    const float        height_ratio  = static_cast<float>(height_in) / static_cast<float>(height_scaled);
-    const int          ax            = a.first;  // The number of zeros added to right edge of the input.
-    const int          ay            = a.second; // The number of zeros added to bottom edge of the input.
-    const unsigned int kernel_size   = weights.shape().x();
-    ARM_COMPUTE_ERROR_ON(info.pad().first > (kernel_size - 1));
-    const int transposed_convolution_padx = kernel_size - info.pad().first - 1;
-    const int transposed_convolution_pady = kernel_size - info.pad().second - 1;
-    const int stridex                     = info.stride().first;
-    const int stridey                     = info.stride().second;
+    const int width_in      = src.shape().x();
+    const int height_in     = src.shape().y();
+    const int width_scaled  = scaled.shape().x();
+    const int height_scaled = scaled.shape().y();
+    const int num_2d_slices = src.shape().total_size() / (width_in * height_in);
+    const int ax            = a.first;  // The number of zeros added to right edge of the input.
+    const int ay            = a.second; // The number of zeros added to top edge of the input.
+    ARM_COMPUTE_ERROR_ON(info.pad().first > (weights.shape().x() - 1));
+
+    ARM_COMPUTE_ERROR_ON_MSG(ax > stride_x - 1, "ax must be smaller than stride_x");
+    ARM_COMPUTE_ERROR_ON_MSG(ay > stride_y - 1, "ay must be smaller than stride_y");
+
     for(int j = 0; j < scaled.num_elements(); ++j)
     {
         scaled[j] = T(0);
@@ -68,34 +69,23 @@ SimpleTensor<T> deconvolution_layer(const SimpleTensor<T> &src, const SimpleTens
     {
         const int offset_slice_in  = slice * width_in * height_in;
         const int offset_slice_out = slice * width_scaled * height_scaled;
-        for(int yi = ay; yi < height_scaled; yi += stridey)
+        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;
+
+        for(int yi = start_y, in_y = 0; yi < end_y; yi += stride_y, in_y++)
         {
-            for(int xi = transposed_convolution_padx; xi < width_scaled; xi += stridex)
+            for(int xi = start_x, in_x = 0; xi < end_x; xi += stride_x, in_x++)
             {
-                const float x_src     = (xi + 0.5f) * width_ratio - 0.5f;
-                const float y_src     = (yi + 0.5f) * height_ratio - 0.5f;
-                T          *out       = scaled.data() + offset_slice_out + xi + yi * width_scaled;
-                const bool  in_bounds = x_src > -1 && y_src > -1 && x_src < width_in && y_src < height_in;
-                const bool  in_axy    = xi < transposed_convolution_padx || xi >= (width_scaled - ax)  // this is checking if the x coordinate is in the padded left/right area
-                                        || yi < ay || yi >= (height_scaled - transposed_convolution_pady); // like above but top and bottom padding in the upscaled XY plane
-                if(!in_axy)
-                {
-                    if(in_bounds)
-                    {
-                        const int in_scaled_x = (x_src < 0.f) ? static_cast<int>(x_src - 0.5f) : static_cast<int>(x_src + 0.5f);
-                        const int in_scaled_y = (y_src < 0.f) ? static_cast<int>(y_src - 0.5f) : static_cast<int>(y_src + 0.5f);
-                        const T *in          = src.data() + offset_slice_in + in_scaled_x + in_scaled_y * width_in;
-                        *out                  = *in;
-                    }
-                    else
-                    {
-                        *out = T(0);
-                    }
-                }
+                const T *in  = src.data() + offset_slice_in + in_y * width_in + in_x;
+                T       *out = scaled.data() + offset_slice_out + xi + yi * width_scaled;
+                *out         = *in;
             }
         }
     }
-    const PadStrideInfo conv_info(1, 1, 1, 1, DimensionRoundingType::CEIL);
+
+    const PadStrideInfo conv_info(1, 1, 0, 0, 0, 0, DimensionRoundingType::CEIL);
     return convolution_layer(scaled, weights, bias, output_shape, conv_info);
 }
 
diff --git a/tests/validation/reference/DeconvolutionLayer.h b/tests/validation/reference/DeconvolutionLayer.h
index 8222e32027..c0bc1fa928 100644
--- a/tests/validation/reference/DeconvolutionLayer.h
+++ b/tests/validation/reference/DeconvolutionLayer.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017 ARM Limited.
+ * Copyright (c) 2017, 2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -42,7 +42,7 @@ namespace reference
  * bias             Optional, ignored if NULL. The biases have one dimension. Data type supported: Same as @p input.
  * output_shape     Output tensor shape. The output has the same number of dimensions as the @p input.
  * info             Contains padding and policies to be used in the deconvolution, this is decribed in @ref PadStrideInfo.
- * a                The number of zeros added to right edge of the input.
+ * a                The number of zeros added to right and top edges of the input.
  *
  */
 template <typename T>