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>

1 files changed, 261 insertions, 14 deletions

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