COMPMID-1944 Add support for "reflect" padding mode in NEPad

Change-Id: I56c42524497d37d44708648571fa211ac1afbd98
Signed-off-by: Usama Arif <usama.arif@arm.com>
Reviewed-on: https://review.mlplatform.org/c/885
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Pablo Marquez <pablo.tello@arm.com>

10 files changed, 487 insertions, 116 deletions

diff --git a/arm_compute/core/Helpers.h b/arm_compute/core/Helpers.h
index c7c7110ef5..235657a38a 100644
--- a/arm_compute/core/Helpers.h
+++ b/arm_compute/core/Helpers.h
@@ -707,6 +707,15 @@ inline int coords2index(const TensorShape &shape, const Coordinates &coord);
  */
 inline size_t get_data_layout_dimension_index(const DataLayout data_layout, const DataLayoutDimension data_layout_dimension);
 
+/** Get the DataLayoutDimension of a given index and layout.
+ *
+ * @param[in] data_layout The data layout.
+ * @param[in] index       The data layout index.
+ *
+ * @return The dimension which this index is requested for.
+ */
+inline DataLayoutDimension get_index_data_layout_dimension(const DataLayout data_layout, const size_t index);
+
 /** Calculate the normalization dimension index for a given normalization type
  *
  * @param[in] layout Data layout of the input and output tensor
diff --git a/arm_compute/core/Helpers.inl b/arm_compute/core/Helpers.inl
index c0e4ab8d7d..aeb290b23e 100644
--- a/arm_compute/core/Helpers.inl
+++ b/arm_compute/core/Helpers.inl
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2018 ARM Limited.
+ * Copyright (c) 2016-2019 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -351,4 +351,33 @@ inline size_t get_data_layout_dimension_index(const DataLayout data_layout, cons
             break;
     }
 }
+
+inline DataLayoutDimension get_index_data_layout_dimension(const DataLayout data_layout, const size_t index)
+{
+    ARM_COMPUTE_ERROR_ON_MSG(data_layout == DataLayout::UNKNOWN, "Cannot retrieve the dimension index for an unknown layout!");
+
+    /* Return the index based on the data layout
+    * [N C H W]
+    * [3 2 1 0]
+    * [N H W C]
+    */
+    switch(index)
+    {
+        case 0:
+            return (data_layout == DataLayout::NCHW) ? DataLayoutDimension::WIDTH : DataLayoutDimension::CHANNEL;
+            break;
+        case 1:
+            return (data_layout == DataLayout::NCHW) ? DataLayoutDimension::HEIGHT : DataLayoutDimension::WIDTH;
+            break;
+        case 2:
+            return (data_layout == DataLayout::NCHW) ? DataLayoutDimension::CHANNEL : DataLayoutDimension::HEIGHT;
+            break;
+        case 3:
+            return DataLayoutDimension::BATCHES;
+            break;
+        default:
+            ARM_COMPUTE_ERROR("Index value not supported!");
+            break;
+    }
+}
 } // namespace arm_compute
diff --git a/arm_compute/runtime/CL/functions/CLPadLayer.h b/arm_compute/runtime/CL/functions/CLPadLayer.h
index 1ecf82fa7c..0179441af2 100644
--- a/arm_compute/runtime/CL/functions/CLPadLayer.h
+++ b/arm_compute/runtime/CL/functions/CLPadLayer.h
@@ -53,9 +53,12 @@ public:
      * @param[out] output         Output tensor. Data type supported: same as @p input
      * @param[in]  padding        The padding for each spatial dimension of the input tensor. The pair padding[i]
      *                            specifies the front and the end padding in the i-th dimension.
-     * @param[in]  constant_value (Optional) Constant value to be used for the padding
+     * @param[in]  constant_value (Optional) Constant value to be used for the padding.
+     * @param[in]  mode           (Optional) Controls whether the padding should be filled with @p constant_value using CONSTANT,
+     *                            or reflect the input, either including the border values (SYMMETRIC) or not (REFLECT). Only CONSTANT
+     *                              is currently supported.
      */
-    void configure(ICLTensor *input, ICLTensor *output, const PaddingList &padding, PixelValue constant_value = PixelValue());
+    void configure(ICLTensor *input, ICLTensor *output, const PaddingList &padding, PixelValue constant_value = PixelValue(), PaddingMode mode = PaddingMode::CONSTANT);
 
     /**  Static function to check if given info will lead to a valid configuration of @ref CLPadLayer.
      *
@@ -64,8 +67,11 @@ public:
      * @param[in] padding        The padding for each spatial dimension of the input tensor. The pair padding[i]
      *                           specifies the front and the end padding in the i-th dimension.
      * @param[in] constant_value (Optional) Constant value to be used for the padding
+     * @param[in] mode           (Optional) Controls whether the padding should be filled with @p constant_value using CONSTANT,
+     *                            or reflect the input, either including the border values (SYMMETRIC) or not (REFLECT). Only CONSTANT
+     *                              is currently supported.
      */
-    static Status validate(const ITensorInfo *input, const ITensorInfo *output, const PaddingList &padding, PixelValue constant_value = PixelValue());
+    static Status validate(const ITensorInfo *input, const ITensorInfo *output, const PaddingList &padding, PixelValue constant_value = PixelValue(), PaddingMode mode = PaddingMode::CONSTANT);
 
     // Inherited methods overridden:
     void run() override;
diff --git a/arm_compute/runtime/NEON/functions/NEPadLayer.h b/arm_compute/runtime/NEON/functions/NEPadLayer.h
index 3a0863802a..78dbc1f1f9 100644
--- a/arm_compute/runtime/NEON/functions/NEPadLayer.h
+++ b/arm_compute/runtime/NEON/functions/NEPadLayer.h
@@ -25,17 +25,17 @@
 #define __ARM_COMPUTE_NEPADLAYER_H__
 
 #include "arm_compute/runtime/IFunction.h"
+#include "arm_compute/runtime/NEON/functions/NEConcatenateLayer.h"
+#include "arm_compute/runtime/NEON/functions/NEStridedSlice.h"
 #include "arm_compute/runtime/SubTensor.h"
 
 #include "arm_compute/core/NEON/kernels/NECopyKernel.h"
 #include "arm_compute/core/NEON/kernels/NEMemsetKernel.h"
 #include "arm_compute/core/Types.h"
+#include "arm_compute/runtime/Tensor.h"
 
 namespace arm_compute
 {
-// Forward declarations
-class ITensor;
-
 /** Basic function to pad a tensor. This function calls the following NEON kernels:
  *
  *  -# @ref NEMemsetKernel
@@ -53,8 +53,10 @@ public:
      * @param[in]  padding        The padding for each spatial dimension of the input tensor. The pair padding[i]
      *                            specifies the front and the end padding in the i-th dimension.
      * @param[in]  constant_value (Optional) Constant value to be used for the padding
+     * @param[in]  mode           (Optional) Controls whether the padding should be filled with @p constant_value using CONSTANT,
+     *                            or reflect the input, either including the border values (SYMMETRIC) or not (REFLECT).
      */
-    void configure(ITensor *input, ITensor *output, const PaddingList &padding, PixelValue constant_value = PixelValue());
+    void configure(ITensor *input, ITensor *output, const PaddingList &padding, const PixelValue constant_value = PixelValue(), const PaddingMode mode = PaddingMode::CONSTANT);
     /**  Static function to check if given info will lead to a valid configuration of @ref NEPadLayer.
      *
      * @param[in] input          Source tensor info. Data types supported: U8/S8/QASYMM8/U16/S16/F16/U32/S32/F32.
@@ -62,18 +64,44 @@ public:
      * @param[in] padding        The padding for each spatial dimension of the input tensor. The pair padding[i]
      *                           specifies the front and the end padding in the i-th dimension.
      * @param[in] constant_value (Optional) Constant value to be used for the padding
+     * @param[in] mode           (Optional) Controls whether the padding should be filled with @p constant_value using CONSTANT,
+     *                     or reflect the input, either including the border values (SYMMETRIC) or not (REFLECT).
      *
      * @return a status
      */
-    static Status validate(const ITensorInfo *input, const ITensorInfo *output, const PaddingList &padding, PixelValue constant_value = PixelValue());
+    static Status validate(const ITensorInfo *input, const ITensorInfo *output, const PaddingList &padding, const PixelValue constant_value = PixelValue(), const PaddingMode mode = PaddingMode::CONSTANT);
 
     // Inherited methods overridden:
     void run() override;
 
 private:
-    NEMemsetKernel _memset_kernel;
-    NECopyKernel   _copy_kernel;
-    SubTensor      _output_subtensor;
+    /** Configure kernels for when constant padding is used.
+     *
+     * @param[in]  input          Source tensor. Data types supported: U8/S8/QASYMM8/U16/S16/F16/U32/S32/F32.
+     * @param[out] output         Output tensor. Data type supported: same as @p input
+     * @param[in]  padding        The padding for each spatial dimension of the input tensor. The pair padding[i]
+     *                            specifies the front and the end padding in the i-th dimension.
+     * @param[in]  constant_value Constant value to be used for the padding
+     */
+    void configure_constant_mode(ITensor *input, ITensor *output, const PaddingList &padding, const PixelValue constant_value);
+    /** Configure functions for when reflect or symmetric padding is used.
+     *
+     * @param[in]  input  Source tensor. Data types supported: U8/S8/QASYMM8/U16/S16/F16/U32/S32/F32.
+     * @param[out] output Output tensor. Data type supported: same as @p input
+     */
+    void configure_reflect_symmetric_mode(ITensor *input, ITensor *output);
+
+private:
+    NECopyKernel                          _copy_kernel;
+    PaddingMode                           _mode;
+    PaddingList                           _padding;
+    NEMemsetKernel                        _memset_kernel;
+    uint32_t                              _num_dimensions;
+    std::unique_ptr<NEStridedSlice[]>     _slice_functions;
+    std::unique_ptr<NEConcatenateLayer[]> _concat_functions;
+    std::unique_ptr<Tensor[]>             _slice_results;
+    std::unique_ptr<Tensor[]>             _concat_results;
+    SubTensor                             _output_subtensor;
 };
 } // namespace arm_compute
 #endif /*__ARM_COMPUTE_NEPADLAYER_H__ */
diff --git a/src/runtime/CL/functions/CLPadLayer.cpp b/src/runtime/CL/functions/CLPadLayer.cpp
index 3aa1b1e1a0..fac2364ae5 100644
--- a/src/runtime/CL/functions/CLPadLayer.cpp
+++ b/src/runtime/CL/functions/CLPadLayer.cpp
@@ -34,8 +34,9 @@ CLPadLayer::CLPadLayer()
 {
 }
 
-void CLPadLayer::configure(ICLTensor *input, ICLTensor *output, const PaddingList &padding, PixelValue constant_value)
+void CLPadLayer::configure(ICLTensor *input, ICLTensor *output, const PaddingList &padding, PixelValue constant_value, PaddingMode mode)
 {
+    ARM_COMPUTE_UNUSED(mode);
     // Copy the input to the output
     _copy_kernel.configure(input, output, padding);
 
@@ -46,10 +47,11 @@ void CLPadLayer::configure(ICLTensor *input, ICLTensor *output, const PaddingLis
     _fillborder_kernel.configure(input, input->info()->padding(), BorderMode::CONSTANT, constant_value);
 }
 
-Status CLPadLayer::validate(const ITensorInfo *input, const ITensorInfo *output, const PaddingList &padding, PixelValue constant_value)
+Status CLPadLayer::validate(const ITensorInfo *input, const ITensorInfo *output, const PaddingList &padding, PixelValue constant_value, PaddingMode mode)
 {
     ARM_COMPUTE_RETURN_ON_ERROR(CLMemsetKernel::validate(input, constant_value));
     ARM_COMPUTE_RETURN_ON_ERROR(CLCopyKernel::validate(input, output, padding));
+    ARM_COMPUTE_RETURN_ERROR_ON(mode != PaddingMode::CONSTANT);
 
     return Status{};
 }
diff --git a/src/runtime/NEON/functions/NEPadLayer.cpp b/src/runtime/NEON/functions/NEPadLayer.cpp
index f5c2718cec..62a7d4559b 100644
--- a/src/runtime/NEON/functions/NEPadLayer.cpp
+++ b/src/runtime/NEON/functions/NEPadLayer.cpp
@@ -25,7 +25,6 @@
 
 #include "arm_compute/runtime/NEON/NEScheduler.h"
 
-#include "arm_compute/core/ITensor.h"
 #include "arm_compute/core/Types.h"
 #include "arm_compute/core/utils/misc/ShapeCalculator.h"
 
@@ -61,18 +60,29 @@ Coordinates get_subtensor_coords(const PaddingList &paddings)
 
     return coords;
 }
+
+uint32_t last_padding_dimension(const PaddingList &padding)
+{
+    int last_padding_dim = padding.size() - 1;
+    for(; last_padding_dim >= 0; --last_padding_dim)
+    {
+        if(padding[last_padding_dim].first > 0 || padding[last_padding_dim].second > 0)
+        {
+            break;
+        }
+    }
+    return static_cast<uint32_t>(last_padding_dim);
+}
 } // namespace
 
 NEPadLayer::NEPadLayer()
-    : _memset_kernel(), _copy_kernel(), _output_subtensor()
+    : _copy_kernel(), _mode(), _padding(), _memset_kernel(), _num_dimensions(0), _slice_functions(nullptr), _concat_functions(nullptr), _slice_results(nullptr), _concat_results(nullptr),
+      _output_subtensor()
 {
 }
 
-void NEPadLayer::configure(ITensor *input, ITensor *output, const PaddingList &padding, PixelValue constant_value)
+void NEPadLayer::configure_constant_mode(ITensor *input, ITensor *output, const PaddingList &padding, const PixelValue constant_value)
 {
-    ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
-    ARM_COMPUTE_THROW_ON_ERROR(NEPadLayer::validate(input->info(), output->info(), padding, constant_value));
-
     // Auto-init
     auto_init_if_empty(*output->info(), get_expected_output_tensorinfo(*input->info(), padding));
 
@@ -86,23 +96,230 @@ void NEPadLayer::configure(ITensor *input, ITensor *output, const PaddingList &p
     _copy_kernel.configure(input, &_output_subtensor);
 }
 
-Status NEPadLayer::validate(const ITensorInfo *input, const ITensorInfo *output, const PaddingList &padding, PixelValue constant_value)
+void NEPadLayer::configure_reflect_symmetric_mode(ITensor *input, ITensor *output)
+{
+    // Reflecting can be performed by effectively unfolding the input as follows:
+    // For each dimension starting at DimX:
+    //      For before and after:
+    //          Use strided slice to extract and reverse the part of the
+    //          input / previously produced tensor required for the padding.
+    //      Concatenate the before and after padding with the input / previously
+    //      produced tensor along the current dimension.
+
+    // Two strided slice functions will be required for each dimension padded as well as a
+    // concatenate function and the tensors to hold the temporary results.
+    _slice_functions  = arm_compute::support::cpp14::make_unique<NEStridedSlice[]>(2 * _num_dimensions);
+    _slice_results    = arm_compute::support::cpp14::make_unique<Tensor[]>(2 * _num_dimensions);
+    _concat_functions = arm_compute::support::cpp14::make_unique<NEConcatenateLayer[]>(_num_dimensions);
+    _concat_results   = arm_compute::support::cpp14::make_unique<Tensor[]>(_num_dimensions - 1);
+    Coordinates starts_before, ends_before, starts_after, ends_after, strides;
+    ITensor    *prev = input;
+    for(uint32_t i = 0; i < _num_dimensions; ++i)
+    {
+        // Values in strides from the previous dimensions need to be set to 1 to avoid reversing again.
+        if(i > 0)
+        {
+            strides.set(i - 1, 1);
+        }
+
+        if(_padding[i].first > 0 || _padding[i].second > 0)
+        {
+            // Set the starts, ends, and strides values for the current dimension.
+            // Due to the bit masks passed to strided slice, the values below the current dimension in
+            // starts and ends will be ignored so do not need to be modified.
+            if(_mode == PaddingMode::REFLECT)
+            {
+                starts_before.set(i, _padding[i].first);
+                ends_before.set(i, 0);
+                starts_after.set(i, input->info()->dimension(i) - 2);
+                ends_after.set(i, input->info()->dimension(i) - _padding[i].second - 2);
+                strides.set(i, -1);
+            }
+            else
+            {
+                starts_before.set(i, _padding[i].first - 1);
+                ends_before.set(i, -1);
+                starts_after.set(i, input->info()->dimension(i) - 1);
+                ends_after.set(i, input->info()->dimension(i) - _padding[i].second - 1);
+                strides.set(i, -1);
+            }
+
+            // Strided slice wraps negative indexes around to the end of the range,
+            // instead this should indicate use of the full range and so the bit mask will be modified.
+            const int32_t begin_mask_before = starts_before[i] < 0 ? ~0 : ~(1u << i);
+            const int32_t end_mask_before   = ends_before[i] < 0 ? ~0 : ~(1u << i);
+            const int32_t begin_mask_after  = starts_after[i] < 0 ? ~0 : ~(1u << i);
+            const int32_t end_mask_after    = ends_after[i] < 0 ? ~0 : ~(1u << i);
+
+            // Reflect the input values for the padding before and after the input.
+            std::vector<ITensor *> concat_vector;
+            if(_padding[i].first > 0)
+            {
+                if(i < prev->info()->num_dimensions())
+                {
+                    _slice_functions[2 * i].configure(prev, &_slice_results[2 * i], starts_before, ends_before, strides, begin_mask_before, end_mask_before);
+                    concat_vector.push_back(&_slice_results[2 * i]);
+                }
+                else
+                {
+                    // Performing the slice is unnecessary if the result would simply be a copy of the tensor.
+                    concat_vector.push_back(prev);
+                }
+            }
+            concat_vector.push_back(prev);
+            if(_padding[i].second > 0)
+            {
+                if(i < prev->info()->num_dimensions())
+                {
+                    _slice_functions[2 * i + 1].configure(prev, &_slice_results[2 * i + 1], starts_after, ends_after, strides, begin_mask_after, end_mask_after);
+                    concat_vector.push_back(&_slice_results[2 * i + 1]);
+                }
+                else
+                {
+                    // Performing the slice is unnecessary if the result would simply be a copy of the tensor.
+                    concat_vector.push_back(prev);
+                }
+            }
+            // Concatenate the padding before and after with the input.
+            ITensor *out = (i == _num_dimensions - 1) ? output : &_concat_results[i];
+            _concat_functions[i].configure(concat_vector, out, get_index_data_layout_dimension(input->info()->data_layout(), i));
+            if(i != _num_dimensions - 1)
+            {
+                _concat_results[i].allocator()->allocate();
+            }
+            prev = out;
+        }
+        _slice_results[2 * i].allocator()->allocate();
+        _slice_results[2 * i + 1].allocator()->allocate();
+    }
+}
+
+void NEPadLayer::configure(ITensor *input, ITensor *output, const PaddingList &padding, const PixelValue constant_value, const PaddingMode mode)
+{
+    ARM_COMPUTE_ERROR_THROW_ON(validate(input->info(), output->info(), padding, constant_value, mode));
+
+    _padding = padding;
+    _mode    = mode;
+
+    const TensorShape padded_shape = misc::shape_calculator::compute_padded_shape(input->info()->tensor_shape(), _padding);
+
+    auto_init_if_empty(*output->info(), input->info()->clone()->set_tensor_shape(padded_shape));
+
+    // Find the last dimension requiring padding so that it is known when to write to output and whether any padding is applied.
+    _num_dimensions = last_padding_dimension(padding) + 1;
+    if(_num_dimensions > 0)
+    {
+        switch(_mode)
+        {
+            case PaddingMode::CONSTANT:
+            {
+                configure_constant_mode(input, output, padding, constant_value);
+                break;
+            }
+            case PaddingMode::REFLECT:
+            case PaddingMode::SYMMETRIC:
+            {
+                configure_reflect_symmetric_mode(input, output);
+                break;
+            }
+            default:
+                ARM_COMPUTE_ERROR("Padding mode not supported.");
+        }
+    }
+    else
+    {
+        // Copy the input to the whole output if no padding is applied
+        _copy_kernel.configure(input, output);
+    }
+}
+
+Status NEPadLayer::validate(const ITensorInfo *input, const ITensorInfo *output, const PaddingList &padding, const PixelValue constant_value, const PaddingMode mode)
 {
     ARM_COMPUTE_UNUSED(constant_value);
-    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output);
 
-    auto output_clone = output->clone();
+    const TensorShape padded_shape = misc::shape_calculator::compute_padded_shape(input->tensor_shape(), padding);
 
-    SubTensorInfo output_subtensor_info(output_clone.get(), input->tensor_shape(), get_subtensor_coords(padding), true);
-    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(*input, *output_clone, padding));
-    ARM_COMPUTE_RETURN_ON_ERROR(NECopyKernel::validate(input, &output_subtensor_info));
+    if(output->total_size() > 0)
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(output->tensor_shape(), padded_shape);
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
+    }
 
+    switch(mode)
+    {
+        case PaddingMode::CONSTANT:
+        {
+            auto          output_clone = output->clone();
+            SubTensorInfo output_subtensor_info(output_clone.get(), input->tensor_shape(), get_subtensor_coords(padding), true);
+            ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(*input, *output_clone, padding));
+            ARM_COMPUTE_RETURN_ON_ERROR(NECopyKernel::validate(input, &output_subtensor_info));
+            break;
+        }
+        case PaddingMode::REFLECT:
+        case PaddingMode::SYMMETRIC:
+        {
+            for(uint32_t i = 0; i < padding.size(); ++i)
+            {
+                if(mode == PaddingMode::REFLECT)
+                {
+                    ARM_COMPUTE_RETURN_ERROR_ON(padding[i].first >= input->dimension(i));
+                    ARM_COMPUTE_RETURN_ERROR_ON(padding[i].second >= input->dimension(i));
+                }
+                else
+                {
+                    ARM_COMPUTE_RETURN_ERROR_ON(padding[i].first > input->dimension(i));
+                    ARM_COMPUTE_RETURN_ERROR_ON(padding[i].second > input->dimension(i));
+                }
+            }
+            break;
+        }
+        default:
+        {
+            ARM_COMPUTE_ERROR("Invalid mode");
+        }
+    }
     return Status{};
 }
 
 void NEPadLayer::run()
 {
-    NEScheduler::get().schedule(&_memset_kernel, Window::DimY);
-    NEScheduler::get().schedule(&_copy_kernel, Window::DimY);
+    if(_num_dimensions > 0)
+    {
+        switch(_mode)
+        {
+            case PaddingMode::CONSTANT:
+            {
+                NEScheduler::get().schedule(&_memset_kernel, Window::DimY);
+                NEScheduler::get().schedule(&_copy_kernel, Window::DimY);
+                break;
+            }
+            case PaddingMode::REFLECT:
+            case PaddingMode::SYMMETRIC:
+            {
+                for(uint32_t i = 0; i < _num_dimensions; ++i)
+                {
+                    if(_padding[i].first > 0 || _padding[i].second > 0)
+                    {
+                        if(_padding[i].first > 0 && _slice_results[2 * i].info()->total_size() > 0)
+                        {
+                            _slice_functions[2 * i].run();
+                        }
+                        if(_padding[i].second > 0 && _slice_results[2 * i + 1].info()->total_size() > 0)
+                        {
+                            _slice_functions[2 * i + 1].run();
+                        }
+                        _concat_functions[i].run();
+                    }
+                }
+                break;
+            }
+            default:
+                ARM_COMPUTE_ERROR("Padding mode not supported.");
+        }
+    }
+    else
+    {
+        NEScheduler::get().schedule(&_copy_kernel, Window::DimY);
+    }
 }
 } // namespace arm_compute
diff --git a/tests/validation/CL/PadLayer.cpp b/tests/validation/CL/PadLayer.cpp
index 4bbd7b8e14..9430b1212b 100644
--- a/tests/validation/CL/PadLayer.cpp
+++ b/tests/validation/CL/PadLayer.cpp
@@ -94,8 +94,9 @@ TEST_SUITE(Float)
 TEST_SUITE(FP32)
 FIXTURE_DATA_TEST_CASE(RunPadding, CLPaddingFixture<float>, framework::DatasetMode::ALL,
                        combine(
-                           combine(datasets::SmallShapes(), framework::dataset::make("DataType", { DataType::F32 })),
-                           PaddingSizesDataset))
+                           combine(combine(datasets::SmallShapes(), framework::dataset::make("DataType", { DataType::F32 })),
+                                   PaddingSizesDataset),
+                           framework::dataset::make("PaddingMode", { PaddingMode::CONSTANT })))
 {
     // Validate output
     validate(CLAccessor(_target), _reference);
@@ -105,8 +106,9 @@ TEST_SUITE_END() // FP32
 TEST_SUITE(FP16)
 FIXTURE_DATA_TEST_CASE(RunPadding, CLPaddingFixture<half>, framework::DatasetMode::ALL,
                        combine(
-                           combine(datasets::SmallShapes(), framework::dataset::make("DataType", { DataType::F16 })),
-                           PaddingSizesDataset))
+                           combine(combine(datasets::SmallShapes(), framework::dataset::make("DataType", { DataType::F16 })),
+                                   PaddingSizesDataset),
+                           framework::dataset::make("PaddingMode", { PaddingMode::CONSTANT })))
 {
     // Validate output
     validate(CLAccessor(_target), _reference);
@@ -118,8 +120,9 @@ TEST_SUITE(Integer)
 TEST_SUITE(S8)
 FIXTURE_DATA_TEST_CASE(RunPadding, CLPaddingFixture<int8_t>, framework::DatasetMode::ALL,
                        combine(
-                           combine(datasets::SmallShapes(), framework::dataset::make("DataType", { DataType::S8 })),
-                           PaddingSizesDataset))
+                           combine(combine(datasets::SmallShapes(), framework::dataset::make("DataType", { DataType::S8 })),
+                                   PaddingSizesDataset),
+                           framework::dataset::make("PaddingMode", { PaddingMode::CONSTANT })))
 {
     // Validate output
     validate(CLAccessor(_target), _reference);
@@ -131,8 +134,9 @@ TEST_SUITE(Quantized)
 TEST_SUITE(QASYMM8)
 FIXTURE_DATA_TEST_CASE(RunPadding, CLPaddingFixture<uint8_t>, framework::DatasetMode::ALL,
                        combine(
-                           combine(datasets::SmallShapes(), framework::dataset::make("DataType", { DataType::QASYMM8 })),
-                           PaddingSizesDataset))
+                           combine(combine(datasets::SmallShapes(), framework::dataset::make("DataType", { DataType::QASYMM8 })),
+                                   PaddingSizesDataset),
+                           framework::dataset::make("PaddingMode", { PaddingMode::CONSTANT })))
 {
     // Validate output
     validate(CLAccessor(_target), _reference);
diff --git a/tests/validation/NEON/PadLayer.cpp b/tests/validation/NEON/PadLayer.cpp
index 90d3ae98d8..5049347f27 100644
--- a/tests/validation/NEON/PadLayer.cpp
+++ b/tests/validation/NEON/PadLayer.cpp
@@ -42,12 +42,14 @@ namespace validation
 {
 namespace
 {
-const auto PaddingSizesDataset = framework::dataset::make("PaddingSize", { PaddingList{ { 0, 0 } },
+const auto PaddingSizesDataset = framework::dataset::make("PaddingSize",
+{
+    PaddingList{ { 0, 0 } },
     PaddingList{ { 1, 1 } },
     PaddingList{ { 1, 1 }, { 2, 2 } },
-    PaddingList{ { 1, 1 }, { 1, 1 }, { 1, 1 }, { 1, 1 } },
-    PaddingList{ { 0, 0 }, { 1, 0 }, { 0, 1 }, { 1, 2 } },
-    PaddingList{ { 0, 0 }, { 0, 0 }, { 0, 0 }, { 1, 1 } }
+    PaddingList{ { 1, 1 }, { 1, 1 }, { 1, 1 } },
+    PaddingList{ { 0, 0 }, { 1, 0 }, { 0, 1 } },
+    PaddingList{ { 0, 1 }, { 1, 0 }, { 0, 1 } },
 });
 } // namespace
 
@@ -57,33 +59,62 @@ TEST_SUITE(PadLayer)
 // *INDENT-OFF*
 // clang-format off
 
-DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(
-        framework::dataset::make("InputInfo", { TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32),     // Mismatching data type input/output
-                                                TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32),     // Mismatching shapes
-                                                TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32),
-                                                TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32),
-                                                TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32),
-                                                TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F32)
-        }),
-        framework::dataset::make("OutputInfo",{ TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F16),
-                                                TensorInfo(TensorShape(28U, 11U, 2U), 1, DataType::F32),
-                                                TensorInfo(TensorShape(29U, 17U, 2U), 1, DataType::F32),
-                                                TensorInfo(TensorShape(29U, 15U, 4U, 3U), 1, DataType::F32),
-                                                TensorInfo(TensorShape(27U, 14U, 3U, 4U), 1, DataType::F32),
-                                                TensorInfo(TensorShape(32U, 13U, 2U, 3U), 1, DataType::F32)
-        })),
-        framework::dataset::make("PaddingSize", { PaddingList{{0, 0}},
-                                                  PaddingList{{1, 1}},
-                                                  PaddingList{{1, 1}, {2, 2}},
-                                                  PaddingList{{1,1}, {1,1}, {1,1}, {1,1}},
-                                                  PaddingList{{0,0}, {1,0}, {0,1}, {1,2}},
-                                                  PaddingList{{0,0}, {0,0}, {0,0}, {1,1}}
-        })),
-        framework::dataset::make("Expected", { false, false, true, true, true, true })),
-        input_info, output_info, padding, expected)
+DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(zip(
+               framework::dataset::make("InputInfo", { TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32),     // Mismatching data type input/output
+                                                       TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32),     // Mismatching shapes
+                                                       TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32),
+                                                       TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32),
+                                                       TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32),
+                                                       TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F32),
+                                                       TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32),     // Mismatching data type input/output
+                                                       TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32),     // Mismatching shapes
+                                                       TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32),
+                                                       TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32),
+                                                       TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32),
+                                                       TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F32)
+                                                     }),
+               framework::dataset::make("OutputInfo",{ TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F16),
+                                                       TensorInfo(TensorShape(28U, 11U, 2U), 1, DataType::F32),
+                                                       TensorInfo(TensorShape(29U, 17U, 2U), 1, DataType::F32),
+                                                       TensorInfo(TensorShape(29U, 15U, 4U, 3U), 1, DataType::F32),
+                                                       TensorInfo(TensorShape(27U, 14U, 3U, 4U), 1, DataType::F32),
+                                                       TensorInfo(TensorShape(32U, 13U, 2U, 3U), 1, DataType::F32),
+                                                       TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F16),
+                                                       TensorInfo(TensorShape(28U, 11U, 2U), 1, DataType::F32),
+                                                       TensorInfo(TensorShape(29U, 17U, 2U), 1, DataType::F32),
+                                                       TensorInfo(TensorShape(29U, 15U, 4U, 3U), 1, DataType::F32),
+                                                       TensorInfo(TensorShape(27U, 14U, 3U, 4U), 1, DataType::F32),
+                                                       TensorInfo(TensorShape(32U, 13U, 2U, 3U), 1, DataType::F32)
+                                                     })),
+               framework::dataset::make("PaddingSize", { PaddingList{{0, 0}},
+                                                         PaddingList{{1, 1}},
+                                                         PaddingList{{1, 1}, {2, 2}},
+                                                         PaddingList{{1,1}, {1,1}, {1,1}, {1,1}},
+                                                         PaddingList{{0,0}, {1,0}, {0,1}, {1,2}},
+                                                         PaddingList{{0,0}, {0,0}, {0,0}, {1,1}},
+                                                         PaddingList{{0, 0}},
+                                                         PaddingList{{1, 1}},
+                                                         PaddingList{{1, 1}, {2, 2}},
+                                                         PaddingList{{1,1}, {1,1}, {1,1}, {1,1}},
+                                                         PaddingList{{0,0}, {1,0}, {0,1}, {1,2}},
+                                                         PaddingList{{0,0}, {0,0}, {0,0}, {1,1}}
+                                                         })),
+               framework::dataset::make("PaddingMode", { PaddingMode::CONSTANT,
+                                                         PaddingMode::CONSTANT,
+                                                         PaddingMode::CONSTANT,
+                                                         PaddingMode::CONSTANT,
+                                                         PaddingMode::CONSTANT,
+                                                         PaddingMode::CONSTANT,
+                                                         PaddingMode::REFLECT,
+                                                         PaddingMode::REFLECT,
+                                                         PaddingMode::REFLECT,
+                                                         PaddingMode::REFLECT,
+                                                         PaddingMode::REFLECT,
+                                                         PaddingMode::SYMMETRIC })),
+               framework::dataset::make("Expected", { false, false, true, true, true, true, false, false, true, false, false, true })),
+               input_info, output_info, padding, mode, expected)
 {
-    Status s = NEPadLayer::validate(&input_info.clone()->set_is_resizable(true), &output_info.clone()->set_is_resizable(true), padding);
-    ARM_COMPUTE_EXPECT(bool(s) == expected, framework::LogLevel::ERRORS);
+    ARM_COMPUTE_EXPECT(bool(NEPadLayer::validate(&input_info.clone()->set_is_resizable(true), &output_info.clone()->set_is_resizable(true), padding, PixelValue(), mode)) == expected, framework::LogLevel::ERRORS);
 }
 
 // clang-format on
@@ -96,17 +127,17 @@ TEST_SUITE(Float)
 
 TEST_SUITE(FP32)
 FIXTURE_DATA_TEST_CASE(RunSmall, NEPaddingFixture<float>, framework::DatasetMode::ALL,
-                       combine(
-                           combine(datasets::SmallShapes(), framework::dataset::make("DataType", { DataType::F32 })),
-                           PaddingSizesDataset))
+                       combine(combine(combine(datasets::Small3DShapes(), framework::dataset::make("DataType", { DataType::F32 })),
+                                       PaddingSizesDataset),
+                               framework::dataset::make("PaddingMode", { PaddingMode::CONSTANT, PaddingMode::REFLECT })))
 {
     // Validate output
     validate(Accessor(_target), _reference);
 }
 FIXTURE_DATA_TEST_CASE(RunLarge, NEPaddingFixture<float>, framework::DatasetMode::NIGHTLY,
-                       combine(
-                           combine(datasets::LargeShapes(), framework::dataset::make("DataType", { DataType::F32 })),
-                           PaddingSizesDataset))
+                       combine(combine(combine(datasets::Large3DShapes(), framework::dataset::make("DataType", { DataType::F32 })),
+                                       PaddingSizesDataset),
+                               framework::dataset::make("PaddingMode", { PaddingMode::CONSTANT, PaddingMode::REFLECT, PaddingMode::SYMMETRIC })))
 {
     // Validate output
     validate(Accessor(_target), _reference);
@@ -116,17 +147,17 @@ TEST_SUITE_END() // FP32
 #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
 TEST_SUITE(FP16)
 FIXTURE_DATA_TEST_CASE(RunSmall, NEPaddingFixture<half>, framework::DatasetMode::ALL,
-                       combine(
-                           combine(datasets::SmallShapes(), framework::dataset::make("DataType", { DataType::F16 })),
-                           PaddingSizesDataset))
+                       combine(combine(combine(datasets::Small3DShapes(), framework::dataset::make("DataType", { DataType::F16 })),
+                                       PaddingSizesDataset),
+                               framework::dataset::make("PaddingMode", { PaddingMode::CONSTANT, PaddingMode::REFLECT })))
 {
     // Validate output
     validate(Accessor(_target), _reference);
 }
 FIXTURE_DATA_TEST_CASE(RunLarge, NEPaddingFixture<half>, framework::DatasetMode::NIGHTLY,
-                       combine(
-                           combine(datasets::LargeShapes(), framework::dataset::make("DataType", { DataType::F16 })),
-                           PaddingSizesDataset))
+                       combine(combine(combine(datasets::Large3DShapes(), framework::dataset::make("DataType", { DataType::F16 })),
+                                       PaddingSizesDataset),
+                               framework::dataset::make("PaddingMode", { PaddingMode::CONSTANT, PaddingMode::REFLECT, PaddingMode::SYMMETRIC })))
 {
     // Validate output
     validate(Accessor(_target), _reference);
@@ -135,41 +166,20 @@ TEST_SUITE_END() // FP16
 #endif           /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */
 TEST_SUITE_END() // Float
 
-TEST_SUITE(Integer)
-TEST_SUITE(S8)
-FIXTURE_DATA_TEST_CASE(RunSmall, NEPaddingFixture<int8_t>, framework::DatasetMode::ALL,
-                       combine(
-                           combine(datasets::SmallShapes(), framework::dataset::make("DataType", { DataType::S8 })),
-                           PaddingSizesDataset))
-{
-    // Validate output
-    validate(Accessor(_target), _reference);
-}
-FIXTURE_DATA_TEST_CASE(RunLarge, NEPaddingFixture<int8_t>, framework::DatasetMode::NIGHTLY,
-                       combine(
-                           combine(datasets::LargeShapes(), framework::dataset::make("DataType", { DataType::S8 })),
-                           PaddingSizesDataset))
-{
-    // Validate output
-    validate(Accessor(_target), _reference);
-}
-TEST_SUITE_END() // S8
-TEST_SUITE_END() // Integer
-
 TEST_SUITE(Quantized)
 TEST_SUITE(QASYMM8)
 FIXTURE_DATA_TEST_CASE(RunSmall, NEPaddingFixture<uint8_t>, framework::DatasetMode::ALL,
-                       combine(
-                           combine(datasets::SmallShapes(), framework::dataset::make("DataType", { DataType::QASYMM8 })),
-                           PaddingSizesDataset))
+                       combine(combine(combine(datasets::Small3DShapes(), framework::dataset::make("DataType", { DataType::QASYMM8 })),
+                                       PaddingSizesDataset),
+                               framework::dataset::make("PaddingMode", { PaddingMode::CONSTANT, PaddingMode::REFLECT })))
 {
     // Validate output
     validate(Accessor(_target), _reference);
 }
 FIXTURE_DATA_TEST_CASE(RunLarge, NEPaddingFixture<uint8_t>, framework::DatasetMode::NIGHTLY,
-                       combine(
-                           combine(datasets::LargeShapes(), framework::dataset::make("DataType", { DataType::QASYMM8 })),
-                           PaddingSizesDataset))
+                       combine(combine(combine(datasets::Large3DShapes(), framework::dataset::make("DataType", { DataType::QASYMM8 })),
+                                       PaddingSizesDataset),
+                               framework::dataset::make("PaddingMode", { PaddingMode::CONSTANT, PaddingMode::REFLECT, PaddingMode::SYMMETRIC })))
 {
     // Validate output
     validate(Accessor(_target), _reference);
diff --git a/tests/validation/fixtures/PadLayerFixture.h b/tests/validation/fixtures/PadLayerFixture.h
index 839313a118..3538cabfeb 100644
--- a/tests/validation/fixtures/PadLayerFixture.h
+++ b/tests/validation/fixtures/PadLayerFixture.h
@@ -45,30 +45,54 @@ class PaddingFixture : public framework::Fixture
 {
 public:
     template <typename...>
-    void setup(TensorShape shape, DataType data_type, const PaddingList &padding)
+    void setup(TensorShape shape, DataType data_type, const PaddingList &padding, const PaddingMode mode)
     {
-        _target    = compute_target(shape, data_type, padding);
-        _reference = compute_reference(shape, data_type, padding);
+        PaddingList clamped_padding = padding;
+        if(mode != PaddingMode::CONSTANT)
+        {
+            // Clamp padding to prevent applying more than is possible.
+            for(uint32_t i = 0; i < padding.size(); ++i)
+            {
+                if(mode == PaddingMode::REFLECT)
+                {
+                    clamped_padding[i].first  = std::min(static_cast<uint64_t>(padding[i].first), static_cast<uint64_t>(shape[i] - 1));
+                    clamped_padding[i].second = std::min(static_cast<uint64_t>(padding[i].second), static_cast<uint64_t>(shape[i] - 1));
+                }
+                else
+                {
+                    clamped_padding[i].first  = std::min(static_cast<uint64_t>(padding[i].first), static_cast<uint64_t>(shape[i]));
+                    clamped_padding[i].second = std::min(static_cast<uint64_t>(padding[i].second), static_cast<uint64_t>(shape[i]));
+                }
+            }
+        }
+        _target    = compute_target(shape, data_type, clamped_padding, mode);
+        _reference = compute_reference(shape, data_type, clamped_padding, mode);
     }
 
 protected:
     template <typename U>
-    void fill(U &&tensor)
+    void fill(U &&tensor, int i)
     {
-        library->fill_tensor_uniform(tensor, 0);
+        library->fill_tensor_uniform(tensor, i);
     }
 
     TensorType compute_target(const TensorShape &shape,
                               DataType           data_type,
-                              const PaddingList &paddings)
+                              const PaddingList &paddings,
+                              const PaddingMode  mode)
     {
         // Create tensors
         TensorType src = create_tensor<TensorType>(shape, data_type);
         TensorType dst;
 
+        TensorType const_val = create_tensor<TensorType>(TensorShape(1), data_type);
+        const_val.allocator()->allocate();
+        fill(AccessorType(const_val), 1);
+        T const_value = *static_cast<T *>(AccessorType(const_val)(Coordinates(0)));
+
         // Create and configure function
         FunctionType padding;
-        padding.configure(&src, &dst, paddings);
+        padding.configure(&src, &dst, paddings, const_value, mode);
 
         ARM_COMPUTE_EXPECT(src.info()->is_resizable(), framework::LogLevel::ERRORS);
         ARM_COMPUTE_EXPECT(dst.info()->is_resizable(), framework::LogLevel::ERRORS);
@@ -81,7 +105,7 @@ protected:
         ARM_COMPUTE_EXPECT(!dst.info()->is_resizable(), framework::LogLevel::ERRORS);
 
         // Fill tensors
-        fill(AccessorType(src));
+        fill(AccessorType(src), 0);
 
         // Compute function
         padding.run();
@@ -90,15 +114,17 @@ protected:
     }
 
     SimpleTensor<T> compute_reference(const TensorShape &shape, DataType data_type,
-                                      const PaddingList &paddings)
+                                      const PaddingList &paddings, const PaddingMode mode)
     {
         // Create reference tensor
         SimpleTensor<T> src{ shape, data_type };
+        SimpleTensor<T> const_val{ TensorShape(1), data_type };
 
         // Fill reference tensor
-        fill(src);
+        fill(src, 0);
+        fill(const_val, 1);
 
-        return reference::pad_layer(src, paddings);
+        return reference::pad_layer(src, paddings, const_val[0], mode);
     }
 
     TensorType      _target{};
diff --git a/utils/TypePrinter.h b/utils/TypePrinter.h
index f2cf606a00..7c23399bc1 100644
--- a/utils/TypePrinter.h
+++ b/utils/TypePrinter.h
@@ -1183,6 +1183,46 @@ inline ::std::ostream &operator<<(::std::ostream &os, const Rectangle &rect)
     return os;
 }
 
+/** Formatted output of the PaddingMode type.
+ *
+ * @param[out] os   Output stream.
+ * @param[in]  mode Type to output.
+ *
+ * @return Modified output stream.
+ */
+inline ::std::ostream &operator<<(::std::ostream &os, const PaddingMode &mode)
+{
+    switch(mode)
+    {
+        case PaddingMode::CONSTANT:
+            os << "CONSTANT";
+            break;
+        case PaddingMode::REFLECT:
+            os << "REFLECT";
+            break;
+        case PaddingMode::SYMMETRIC:
+            os << "SYMMETRIC";
+            break;
+        default:
+            ARM_COMPUTE_ERROR("NOT_SUPPORTED!");
+    }
+
+    return os;
+}
+
+/** Formatted output of the PaddingMode type.
+ *
+ * @param[in] mode Type to output.
+ *
+ * @return Formatted string.
+ */
+inline std::string to_string(const PaddingMode &mode)
+{
+    std::stringstream str;
+    str << mode;
+    return str.str();
+}
+
 /** Formatted output of the PadStrideInfo type.
  *
  * @param[out] os              Output stream.