COMPMID-2008: Add support for "reflect" padding mode in CLPad

Change-Id: I469f8173d5c4a1b6f03b52b9ddd33928dacd1e7b
Signed-off-by: Manuel Bottini <manuel.bottini@arm.com>
Reviewed-on: https://review.mlplatform.org/c/869
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Giuseppe Rossini <giuseppe.rossini@arm.com>

3 files changed, 321 insertions, 63 deletions

diff --git a/arm_compute/runtime/CL/functions/CLPadLayer.h b/arm_compute/runtime/CL/functions/CLPadLayer.h
index 0179441af2..33b09d60a2 100644
--- a/arm_compute/runtime/CL/functions/CLPadLayer.h
+++ b/arm_compute/runtime/CL/functions/CLPadLayer.h
@@ -25,10 +25,13 @@
 #define __ARM_COMPUTE_CLPADLAYER_H__
 
 #include "arm_compute/core/CL/kernels/CLCopyKernel.h"
-#include "arm_compute/core/CL/kernels/CLFillBorderKernel.h"
 #include "arm_compute/core/CL/kernels/CLMemsetKernel.h"
 #include "arm_compute/core/Types.h"
+#include "arm_compute/runtime/CL/functions/CLConcatenateLayer.h"
+
 #include "arm_compute/runtime/CL/CLScheduler.h"
+#include "arm_compute/runtime/CL/CLTensor.h"
+#include "arm_compute/runtime/CL/functions/CLStridedSlice.h"
 #include "arm_compute/runtime/IFunction.h"
 
 namespace arm_compute
@@ -77,9 +80,18 @@ public:
     void run() override;
 
 private:
-    CLCopyKernel       _copy_kernel;
-    CLFillBorderKernel _fillborder_kernel;
-    CLMemsetKernel     _memset_kernel;
+    void configure_constant_mode(ICLTensor *input, ICLTensor *output, const PaddingList &padding, const PixelValue constant_value);
+    void configure_reflect_symmetric_mode(ICLTensor *input, ICLTensor *output);
+
+    CLCopyKernel                          _copy_kernel;
+    PaddingMode                           _mode;
+    PaddingList                           _padding;
+    CLMemsetKernel                        _memset_kernel;
+    size_t                                _num_dimensions;
+    std::unique_ptr<CLStridedSlice[]>     _slice_functions;
+    std::unique_ptr<CLConcatenateLayer[]> _concat_functions;
+    std::unique_ptr<CLTensor[]>           _slice_results;
+    std::unique_ptr<CLTensor[]>           _concat_results;
 };
 } // namespace arm_compute
 #endif /*__ARM_COMPUTE_PADLAYER_H__ */
diff --git a/src/runtime/CL/functions/CLPadLayer.cpp b/src/runtime/CL/functions/CLPadLayer.cpp
index fac2364ae5..f88cb388be 100644
--- a/src/runtime/CL/functions/CLPadLayer.cpp
+++ b/src/runtime/CL/functions/CLPadLayer.cpp
@@ -25,41 +25,288 @@
 
 #include "arm_compute/core/CL/ICLTensor.h"
 #include "arm_compute/core/Types.h"
+#include "arm_compute/core/utils/misc/ShapeCalculator.h"
 #include "support/ToolchainSupport.h"
 
 namespace arm_compute
 {
 CLPadLayer::CLPadLayer()
-    : _copy_kernel(), _fillborder_kernel(), _memset_kernel()
+    : _copy_kernel(), _mode(), _padding(), _memset_kernel(), _num_dimensions(0), _slice_functions(nullptr), _concat_functions(nullptr), _slice_results(nullptr), _concat_results(nullptr)
 {
 }
 
+void CLPadLayer::configure_constant_mode(ICLTensor *input, ICLTensor *output, const PaddingList &padding, const PixelValue constant_value)
+{
+    // Set the pages of the output to the constant_value.
+    _memset_kernel.configure(output, constant_value);
+
+    // Fill out padding list with zeroes.
+    PaddingList padding_extended = padding;
+    for(size_t i = padding.size(); i < TensorShape::num_max_dimensions; i++)
+    {
+        padding_extended.emplace_back(PaddingInfo{ 0, 0 });
+    }
+
+    // Create a window within the output tensor where the input will be copied.
+    Window copy_window = Window();
+    for(uint32_t i = 0; i < output->info()->num_dimensions(); ++i)
+    {
+        copy_window.set(i, Window::Dimension(padding_extended[i].first, padding_extended[i].first + input->info()->dimension(i), 1));
+    }
+    // Copy the input to the output, leaving the padding filled with the constant_value.
+    _copy_kernel.configure(input, output, PaddingList(), &copy_window);
+}
+
+void CLPadLayer::configure_reflect_symmetric_mode(ICLTensor *input, ICLTensor *output)
+{
+    int64_t last_padding_dimension = _padding.size() - 1;
+    // Reflecting can be performed by effectively unfolding the input as follows:
+    // For each dimension starting at DimX:
+    //      Create a before and after slice, which values depend on the selected padding mode
+    //      Concatenate the before and after padding with the tensor to be padded
+
+    // 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<CLStridedSlice[]>(2 * _num_dimensions);
+    _slice_results    = arm_compute::support::cpp14::make_unique<CLTensor[]>(2 * _num_dimensions);
+    _concat_functions = arm_compute::support::cpp14::make_unique<CLConcatenateLayer[]>(_num_dimensions);
+    _concat_results   = arm_compute::support::cpp14::make_unique<CLTensor[]>(_num_dimensions - 1);
+    Coordinates starts_before, ends_before, starts_after, ends_after, strides;
+    ICLTensor *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<ICLTensor *> 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.
+            ICLTensor *out = (static_cast<int32_t>(i) == last_padding_dimension) ? output : &_concat_results[i];
+            _concat_functions[i].configure(concat_vector, out, get_index_data_layout_dimension(prev->info()->data_layout(), i));
+            prev = out;
+        }
+    }
+    for(uint32_t i = 0; i < _num_dimensions; ++i)
+    {
+        if((static_cast<int32_t>(i) != last_padding_dimension))
+        {
+            _concat_results[i].allocator()->allocate();
+        }
+        _slice_results[2 * i].allocator()->allocate();
+        _slice_results[2 * i + 1].allocator()->allocate();
+    }
+}
+
 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);
+    ARM_COMPUTE_ERROR_THROW_ON(validate(input->info(), output->info(), padding, constant_value, mode));
 
-    // Set the pages of the output to zero
-    _memset_kernel.configure(output, constant_value);
+    _padding = padding;
+    _mode    = mode;
+
+    TensorShape padded_shape = misc::shape_calculator::compute_padded_shape(input->info()->tensor_shape(), _padding);
 
-    // Fill padding on the first two dimensions with zeros
-    _fillborder_kernel.configure(input, input->info()->padding(), BorderMode::CONSTANT, constant_value);
+    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.
+    int64_t last_padding_dimension = _padding.size() - 1;
+    for(; last_padding_dimension >= 0; --last_padding_dimension)
+    {
+        if(_padding[last_padding_dimension].first > 0 || _padding[last_padding_dimension].second > 0)
+        {
+            break;
+        }
+    }
+    _num_dimensions = last_padding_dimension + 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 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);
+    ARM_COMPUTE_RETURN_ERROR_ON(padding.size() > input->num_dimensions());
+
+    TensorShape padded_shape = misc::shape_calculator::compute_padded_shape(input->tensor_shape(), padding);
+
+    // Use CLCopyKernel and CLMemsetKernel to validate all padding modes as this includes all of the shape and info validation.
+    PaddingList padding_extended = padding;
+    for(size_t i = padding.size(); i < TensorShape::num_max_dimensions; i++)
+    {
+        padding_extended.emplace_back(PaddingInfo{ 0, 0 });
+    }
+
+    Window copy_window = Window();
+    for(uint32_t i = 0; i < padded_shape.num_dimensions(); ++i)
+    {
+        copy_window.set(i, Window::Dimension(padding_extended[i].first, padding_extended[i].first + input->dimension(i), 1));
+    }
+    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(output, input);
+        ARM_COMPUTE_RETURN_ON_ERROR(CLCopyKernel::validate(input, output, PaddingList(), &copy_window));
+        ARM_COMPUTE_RETURN_ON_ERROR(CLMemsetKernel::validate(output, constant_value));
+    }
+    else
+    {
+        ARM_COMPUTE_RETURN_ON_ERROR(CLCopyKernel::validate(input, &input->clone()->set_tensor_shape(padded_shape), PaddingList(), &copy_window));
+        ARM_COMPUTE_RETURN_ON_ERROR(CLMemsetKernel::validate(&input->clone()->set_tensor_shape(padded_shape), constant_value));
+    }
 
+    switch(mode)
+    {
+        case PaddingMode::CONSTANT:
+        {
+            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 CLPadLayer::run()
 {
-    CLScheduler::get().enqueue(_memset_kernel, false);
-    CLScheduler::get().enqueue(_fillborder_kernel, false);
-    CLScheduler::get().enqueue(_copy_kernel, true);
+    if(_num_dimensions > 0)
+    {
+        switch(_mode)
+        {
+            case PaddingMode::CONSTANT:
+            {
+                CLScheduler::get().enqueue(_memset_kernel, false);
+                CLScheduler::get().enqueue(_copy_kernel, true);
+                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();
+                        }
+                        CLScheduler::get().sync();
+                        _concat_functions[i].run();
+                        CLScheduler::get().sync();
+                    }
+                }
+                break;
+            }
+            default:
+                ARM_COMPUTE_ERROR("Padding mode not supported.");
+        }
+    }
+    else
+    {
+        CLScheduler::get().enqueue(_copy_kernel, true);
+    }
 }
 } // namespace arm_compute
diff --git a/tests/validation/CL/PadLayer.cpp b/tests/validation/CL/PadLayer.cpp
index 9430b1212b..2ad29fc0e5 100644
--- a/tests/validation/CL/PadLayer.cpp
+++ b/tests/validation/CL/PadLayer.cpp
@@ -43,9 +43,9 @@ namespace
 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, 0 }, { 0, 0 }, { 0, 0 } }
 });
 } // namespace
 
@@ -55,32 +55,44 @@ TEST_SUITE(PadLayer)
 // *INDENT-OFF*
 // clang-format off
 
-DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(
+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),    // Mismatching shapes with padding
                                                        TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32),
+                                                       TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32),    // Mismatching shapes dimension
                                                        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(32U, 13U), 1, DataType::F32)     // Invalid padding list
                                                      }),
                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(29U, 17U, 2U), 1, DataType::F32),
+                                                       TensorInfo(TensorShape(32U, 13U), 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)
+                                                         PaddingList{{1, 1}},
+                                                         PaddingList{{1, 1}, {2, 2}},
+                                                         PaddingList{{1,1}, {1,1}, {1,1}},
+                                                         PaddingList{{1, 1}, {2, 2}},
+                                                         PaddingList{{0,0}, {0,0}, {1,1}}
+                                                         })),
+               framework::dataset::make("PaddingMode", { PaddingMode::CONSTANT,
+                                                         PaddingMode::CONSTANT,
+                                                         PaddingMode::CONSTANT,
+                                                         PaddingMode::SYMMETRIC,
+                                                         PaddingMode::REFLECT,
+                                                         PaddingMode::REFLECT
+})),
+               framework::dataset::make("Expected", { false,
+                                                   false,
+                                                   true,
+                                                   false,
+                                                   true,
+                                                   false })),
+               input_info, output_info, padding, mode, expected)
 {
-    ARM_COMPUTE_EXPECT(bool(CLPadLayer::validate(&input_info.clone()->set_is_resizable(true), &output_info.clone()->set_is_resizable(true), padding)) == expected, framework::LogLevel::ERRORS);
+    ARM_COMPUTE_EXPECT(bool(CLPadLayer::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
@@ -92,11 +104,9 @@ using CLPaddingFixture = PaddingFixture<CLTensor, CLAccessor, CLPadLayer, T>;
 TEST_SUITE(Float)
 
 TEST_SUITE(FP32)
-FIXTURE_DATA_TEST_CASE(RunPadding, CLPaddingFixture<float>, framework::DatasetMode::ALL,
-                       combine(
-                           combine(combine(datasets::SmallShapes(), framework::dataset::make("DataType", { DataType::F32 })),
-                                   PaddingSizesDataset),
-                           framework::dataset::make("PaddingMode", { PaddingMode::CONSTANT })))
+FIXTURE_DATA_TEST_CASE(RunSmall, CLPaddingFixture<float>, framework::DatasetMode::ALL,
+                       combine(combine(combine(datasets::Small3DShapes(), framework::dataset::make("DataType", { DataType::F32 })), PaddingSizesDataset),
+                               framework::dataset::make("PaddingMode", { PaddingMode::CONSTANT, PaddingMode::REFLECT, PaddingMode::SYMMETRIC })))
 {
     // Validate output
     validate(CLAccessor(_target), _reference);
@@ -104,11 +114,9 @@ FIXTURE_DATA_TEST_CASE(RunPadding, CLPaddingFixture<float>, framework::DatasetMo
 TEST_SUITE_END() // FP32
 
 TEST_SUITE(FP16)
-FIXTURE_DATA_TEST_CASE(RunPadding, CLPaddingFixture<half>, framework::DatasetMode::ALL,
-                       combine(
-                           combine(combine(datasets::SmallShapes(), framework::dataset::make("DataType", { DataType::F16 })),
-                                   PaddingSizesDataset),
-                           framework::dataset::make("PaddingMode", { PaddingMode::CONSTANT })))
+FIXTURE_DATA_TEST_CASE(RunLarge, CLPaddingFixture<half>, framework::DatasetMode::NIGHTLY,
+                       combine(combine(combine(datasets::Large3DShapes(), framework::dataset::make("DataType", { DataType::F16 })), PaddingSizesDataset),
+                               framework::dataset::make("PaddingMode", { PaddingMode::CONSTANT, PaddingMode::REFLECT })))
 {
     // Validate output
     validate(CLAccessor(_target), _reference);
@@ -116,27 +124,18 @@ FIXTURE_DATA_TEST_CASE(RunPadding, CLPaddingFixture<half>, framework::DatasetMod
 TEST_SUITE_END() // FP16
 TEST_SUITE_END() // Float
 
-TEST_SUITE(Integer)
-TEST_SUITE(S8)
-FIXTURE_DATA_TEST_CASE(RunPadding, CLPaddingFixture<int8_t>, framework::DatasetMode::ALL,
-                       combine(
-                           combine(combine(datasets::SmallShapes(), framework::dataset::make("DataType", { DataType::S8 })),
-                                   PaddingSizesDataset),
-                           framework::dataset::make("PaddingMode", { PaddingMode::CONSTANT })))
+TEST_SUITE(Quantized)
+TEST_SUITE(QASYMM8)
+FIXTURE_DATA_TEST_CASE(RunSmall, CLPaddingFixture<uint8_t>, framework::DatasetMode::PRECOMMIT,
+                       combine(combine(combine(datasets::Small3DShapes(), framework::dataset::make("DataType", { DataType::QASYMM8 })), PaddingSizesDataset),
+                               framework::dataset::make("PaddingMode", { PaddingMode::CONSTANT, PaddingMode::REFLECT })))
 {
     // Validate output
     validate(CLAccessor(_target), _reference);
 }
-TEST_SUITE_END() // S8
-TEST_SUITE_END() // Integer
-
-TEST_SUITE(Quantized)
-TEST_SUITE(QASYMM8)
-FIXTURE_DATA_TEST_CASE(RunPadding, CLPaddingFixture<uint8_t>, framework::DatasetMode::ALL,
-                       combine(
-                           combine(combine(datasets::SmallShapes(), framework::dataset::make("DataType", { DataType::QASYMM8 })),
-                                   PaddingSizesDataset),
-                           framework::dataset::make("PaddingMode", { PaddingMode::CONSTANT })))
+FIXTURE_DATA_TEST_CASE(RunLarge, CLPaddingFixture<uint8_t>, framework::DatasetMode::NIGHTLY,
+                       combine(combine(combine(datasets::Large3DShapes(), framework::dataset::make("DataType", { DataType::QASYMM8 })), PaddingSizesDataset),
+                               framework::dataset::make("PaddingMode", { PaddingMode::CONSTANT, PaddingMode::REFLECT })))
 {
     // Validate output
     validate(CLAccessor(_target), _reference);