COMPMID-2592 Create a new kernel for CLPad with SYMMETRIC and REFLECT mode

Change-Id: Icaf0516f490b2ddca6d1ea03a5cf26cc7d43041f
Signed-off-by: Giorgio Arena <giorgio.arena@arm.com>
Reviewed-on: https://review.mlplatform.org/c/1872
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>

1 files changed, 20 insertions, 225 deletions

diff --git a/src/runtime/CL/functions/CLPadLayer.cpp b/src/runtime/CL/functions/CLPadLayer.cpp
index 88b1b77a0d..8f36a69866 100644
--- a/src/runtime/CL/functions/CLPadLayer.cpp
+++ b/src/runtime/CL/functions/CLPadLayer.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2018-2019 ARM Limited.
+ * Copyright (c) 2019 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -23,155 +23,25 @@
  */
 #include "arm_compute/runtime/CL/functions/CLPadLayer.h"
 
-#include "arm_compute/core/utils/misc/ShapeCalculator.h"
-
 namespace arm_compute
 {
 CLPadLayer::CLPadLayer()
-    : _pad_kernel(), _copy_kernel(), _mode(), _padding(), _num_dimensions(0), _slice_functions(), _concat_functions(), _slice_results(), _concat_results()
+    : _pad_kernel(), _copy_kernel(), _perform_pad(false)
 {
 }
-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.resize(2 * _num_dimensions);
-    _slice_results.resize(2 * _num_dimensions);
-    _concat_functions.resize(_num_dimensions);
-    _concat_results.resize(_num_dimensions - 1);
-
-    Coordinates starts_before{};
-    Coordinates ends_before{};
-    Coordinates starts_after{};
-    Coordinates ends_after{};
-    Coordinates 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, 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)
 {
-    _padding = padding;
-    _mode    = mode;
-
-    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));
     ARM_COMPUTE_ERROR_THROW_ON(validate(input->info(), output->info(), padding, constant_value, mode));
 
-    // 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)
+    _perform_pad = std::any_of(padding.begin(), padding.end(), [](PaddingInfo info)
     {
-        if(_padding[last_padding_dimension].first > 0 || _padding[last_padding_dimension].second > 0)
-        {
-            break;
-        }
-    }
-    _num_dimensions = last_padding_dimension + 1;
-    if(_num_dimensions > 0)
+        return info.first > 0 || info.second > 0;
+    });
+
+    if(_perform_pad)
     {
-        switch(_mode)
-        {
-            case PaddingMode::CONSTANT:
-            {
-                _pad_kernel.configure(input, output, padding, constant_value, mode);
-                break;
-            }
-            case PaddingMode::REFLECT:
-            case PaddingMode::SYMMETRIC:
-            {
-                configure_reflect_symmetric_mode(input, output);
-                break;
-            }
-            default:
-                ARM_COMPUTE_ERROR("Padding mode not supported.");
-        }
+        _pad_kernel.configure(input, output, padding, constant_value, mode);
     }
     else
     {
@@ -179,109 +49,34 @@ void CLPadLayer::configure(ICLTensor *input, ICLTensor *output, const PaddingLis
         _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_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++)
+    bool perform_pad = std::any_of(padding.begin(), padding.end(), [](PaddingInfo info)
     {
-        padding_extended.emplace_back(PaddingInfo{ 0, 0 });
-    }
+        return info.first > 0 || info.second > 0;
+    });
 
-    Window copy_window = Window();
-    for(uint32_t i = 0; i < padded_shape.num_dimensions(); ++i)
+    if(perform_pad)
     {
-        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(CLPadLayerKernel::validate(input, output, padding, constant_value, mode));
     }
     else
     {
-        ARM_COMPUTE_RETURN_ON_ERROR(CLCopyKernel::validate(input, &input->clone()->set_tensor_shape(padded_shape), PaddingList(), &copy_window));
-    }
-
-    switch(mode)
-    {
-        case PaddingMode::CONSTANT:
-        {
-            ARM_COMPUTE_RETURN_ERROR_ON(padding.size() > 4);
-            ARM_COMPUTE_RETURN_ON_ERROR(CLPadLayerKernel::validate(input, output, padding, constant_value, mode));
-            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");
-        }
+        Window copy_window = Window();
+        copy_window.use_tensor_dimensions(output->tensor_shape());
+        ARM_COMPUTE_RETURN_ON_ERROR(CLCopyKernel::validate(input, output, PaddingList(), &copy_window));
     }
     return Status{};
 }
-
 void CLPadLayer::run()
 {
-    if(_num_dimensions > 0)
+    if(_perform_pad)
     {
-        switch(_mode)
-        {
-            case PaddingMode::CONSTANT:
-            {
-                CLScheduler::get().enqueue(_pad_kernel, false);
-                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.");
-        }
+        CLScheduler::get().enqueue(_pad_kernel);
     }
     else
     {
-        CLScheduler::get().enqueue(_copy_kernel, true);
+        CLScheduler::get().enqueue(_copy_kernel);
     }
 }
-} // namespace arm_compute
+} // namespace arm_compute
\ No newline at end of file