MLCE-129: NEPad 30x slower than TensorFlow's implementation

Change-Id: I44770e6a3134c70c4bd58f890d06cb43c9bd8bff
Signed-off-by: Manuel Bottini <manuel.bottini@arm.com>
Reviewed-on: https://review.mlplatform.org/c/1853
Reviewed-by: Giorgio Arena <giorgio.arena@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>

2 files changed, 263 insertions, 46 deletions

diff --git a/src/core/NEON/kernels/NEPadLayerKernel.cpp b/src/core/NEON/kernels/NEPadLayerKernel.cpp
new file mode 100644
index 0000000000..88a1c2ec83
--- /dev/null
+++ b/src/core/NEON/kernels/NEPadLayerKernel.cpp
@@ -0,0 +1,259 @@
+/*
+ * Copyright (c) 2019 ARM Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#include "arm_compute/core/NEON/kernels/NEPadLayerKernel.h"
+
+#include "arm_compute/core/Error.h"
+#include "arm_compute/core/Helpers.h"
+#include "arm_compute/core/ITensor.h"
+#include "arm_compute/core/NEON/wrapper/wrapper.h"
+#include "arm_compute/core/TensorInfo.h"
+#include "arm_compute/core/Types.h"
+#include "arm_compute/core/Validate.h"
+#include "arm_compute/core/utils/misc/ShapeCalculator.h"
+
+namespace arm_compute
+{
+namespace
+{
+Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, const PaddingList &paddings, const PaddingMode mode)
+{
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG(mode != PaddingMode::CONSTANT, "Only constant padding mode is supported");
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG(paddings.size() > 4, "Padding list bigger than 4 dimensions");
+    if(output->total_size() != 0)
+    {
+        const TensorShape expected_output_shape = arm_compute::misc::shape_calculator::compute_padded_shape(input->tensor_shape(), paddings);
+        const TensorInfo  expected_output_info  = input->clone()->set_tensor_shape(expected_output_shape);
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(output, &expected_output_info);
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
+    }
+    return Status{};
+}
+} // namespace
+
+template <typename T>
+void NEPadLayerKernel::run_pad_constant(const Window &window)
+{
+    Window output_window{ window };
+    output_window.set(Window::DimX, Window::Dimension(0, 1, 1));
+
+    const size_t element_size = _input->info()->element_size();
+    Iterator     output_it(_output, output_window);
+    execute_window_loop(output_window, [&](const Coordinates & id)
+    {
+        Coordinates idin{ id };
+        for(size_t dim = _padding.size() - 1; dim > 0; --dim)
+        {
+            idin[dim] -= _padding[dim].first;
+            if(idin[dim] < 0 || static_cast<int>(_input->info()->dimension(dim)) - 1 < idin[dim])
+            {
+                std::fill_n(reinterpret_cast<T *>(output_it.ptr()), _output->info()->dimension(0), _constant_value.get<T>());
+                return;
+            }
+        }
+        T *input_it_ptr  = reinterpret_cast<T *>(_input->ptr_to_element(idin));
+        T *output_it_ptr = reinterpret_cast<T *>(output_it.ptr());
+        std::fill_n(output_it_ptr, _padding[0].first, _constant_value.get<T>());
+        memcpy(output_it_ptr + _padding[0].first, input_it_ptr, _input->info()->dimension(0) * element_size);
+        std::fill_n(output_it_ptr + _padding[0].first + _input->info()->dimension(0), _padding[0].second, _constant_value.get<T>());
+    },
+    output_it);
+}
+
+void NEPadLayerKernel::run_pad_constant_uint8_3Dinput_3Dpad(const Window &window)
+{
+    ARM_COMPUTE_UNUSED(window);
+
+    const size_t start_plane = window.z().start();
+    const size_t end_plane   = window.z().end();
+
+    const size_t start_plane_input = start_plane - (_padding.size() > 2 && start_plane >= _padding[2].first ? _padding[2].first : 0);
+
+    const int output_plane_size = _output->info()->dimension(0) * _output->info()->dimension(1);
+    const int input_plane_size  = (_input->info()->dimension(0) + _input->info()->padding().right + _input->info()->padding().left) * (_input->info()->dimension(
+                                      1)
+                                  + _input->info()->padding().top + _input->info()->padding().bottom);
+
+    const int pad_y_elems_top = (_padding.size() > 1 ? _padding[1].first : 0) * _output->info()->dimension(0);
+    const int pad_y_elems_bot = (_padding.size() > 1 ? _padding[1].second : 0) * _output->info()->dimension(0);
+
+    const size_t jump_to_next_row_input   = _input->info()->dimension(0) + _input->info()->padding().right + _input->info()->padding().left;
+    const size_t jump_to_next_row_output  = _padding[0].first + _padding[0].second;
+    const size_t jump_to_next_plane_input = _input->info()->padding().empty() ? 0 : _input->info()->dimension(0) * (_input->info()->padding().right + _input->info()->padding().top);
+
+    uint8_t       *output_row_ptr = _output->buffer() + start_plane * output_plane_size;
+    const uint8_t *input_it_ptr   = _input->buffer() + _input->info()->offset_first_element_in_bytes() + start_plane_input * input_plane_size;
+    const auto     pad_value      = _constant_value.get<uint8_t>();
+
+    for(size_t z_i = start_plane; z_i < end_plane; ++z_i)
+    {
+        if(_padding.size() > 2 && z_i < _padding[2].first)
+        {
+            memset(output_row_ptr, pad_value, output_plane_size);
+            output_row_ptr += output_plane_size;
+        }
+        else if(_padding.size() > 2 && z_i > _input->info()->dimension(2) + _padding[2].first - 1)
+        {
+            memset(output_row_ptr, pad_value, output_plane_size);
+            output_row_ptr += output_plane_size;
+        }
+        else
+        {
+            memset(output_row_ptr, pad_value, pad_y_elems_top);
+            output_row_ptr += pad_y_elems_top;
+            size_t y_i = _input->info()->dimension(1);
+            // Basic loop unrolling
+            for(; y_i > 3; y_i -= 4)
+            {
+                memset(output_row_ptr, pad_value, _padding[0].first);
+                output_row_ptr += _padding[0].first;
+
+                memcpy(output_row_ptr, input_it_ptr, _input->info()->dimension(0));
+                output_row_ptr += _input->info()->dimension(0);
+                input_it_ptr += jump_to_next_row_input;
+
+                memset(output_row_ptr, pad_value, _padding[0].second + _padding[0].first);
+                output_row_ptr += jump_to_next_row_output;
+
+                memcpy(output_row_ptr, input_it_ptr, _input->info()->dimension(0));
+                output_row_ptr += _input->info()->dimension(0);
+                input_it_ptr += jump_to_next_row_input;
+
+                memset(output_row_ptr, pad_value, _padding[0].second + _padding[0].first);
+                output_row_ptr += jump_to_next_row_output;
+
+                memcpy(output_row_ptr, input_it_ptr, _input->info()->dimension(0));
+                output_row_ptr += _input->info()->dimension(0);
+                input_it_ptr += jump_to_next_row_input;
+
+                memset(output_row_ptr, pad_value, _padding[0].second + _padding[0].first);
+                output_row_ptr += jump_to_next_row_output;
+
+                memcpy(output_row_ptr, input_it_ptr, _input->info()->dimension(0));
+                output_row_ptr += _input->info()->dimension(0);
+                input_it_ptr += jump_to_next_row_input;
+
+                memset(output_row_ptr, pad_value, _padding[0].second);
+                output_row_ptr += _padding[0].second;
+            }
+            for(; y_i > 0; --y_i)
+            {
+                memset(output_row_ptr, pad_value, _padding[0].first);
+                output_row_ptr += _padding[0].first;
+
+                memcpy(output_row_ptr, input_it_ptr, _input->info()->dimension(0));
+                output_row_ptr += _input->info()->dimension(0);
+                input_it_ptr += _input->info()->dimension(0);
+
+                memset(output_row_ptr, pad_value, _padding[0].second);
+                output_row_ptr += _padding[0].second;
+            }
+            input_it_ptr += jump_to_next_plane_input;
+            memset(output_row_ptr, pad_value, pad_y_elems_bot);
+            output_row_ptr += pad_y_elems_bot;
+        }
+    }
+}
+
+NEPadLayerKernel::NEPadLayerKernel()
+    : _func(), _input(nullptr), _output(nullptr), _padding(), _constant_value(), _mode()
+{
+}
+
+void NEPadLayerKernel::configure(ITensor *input, ITensor *output, const PaddingList &padding, const PixelValue constant_value, const PaddingMode mode)
+{
+    ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
+    // Auto-init
+    const TensorShape expected_output_shape = arm_compute::misc::shape_calculator::compute_padded_shape(input->info()->tensor_shape(), padding);
+    const TensorInfo  expected_output_info  = input->info()->clone()->set_tensor_shape(expected_output_shape);
+    auto_init_if_empty(*output->info(), expected_output_info);
+
+    // Perform validation step
+    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info(), padding, mode));
+
+    _input          = input;
+    _output         = output;
+    _padding        = padding;
+    _constant_value = constant_value;
+    _mode           = mode;
+
+    if(_mode == PaddingMode::CONSTANT)
+    {
+        switch(_input->info()->element_size())
+        {
+            case 1:
+                if(_input->info()->num_dimensions() == 3 && padding.size() <= 3)
+                {
+                    _func = &NEPadLayerKernel::run_pad_constant_uint8_3Dinput_3Dpad;
+                }
+                else
+                {
+                    _func = &NEPadLayerKernel::run_pad_constant<uint8_t>;
+                }
+                break;
+            case 2:
+                _func = &NEPadLayerKernel::run_pad_constant<uint16_t>;
+                break;
+            case 4:
+                _func = &NEPadLayerKernel::run_pad_constant<uint32_t>;
+                break;
+            default:
+                ARM_COMPUTE_ERROR("Element size not supported");
+                break;
+        }
+    }
+    else
+    {
+        ARM_COMPUTE_ERROR("Padding mode not supported");
+    }
+
+    // Configure kernel window
+    Window win = calculate_max_window(*output->info(), Steps());
+
+    // The NEPad doesn't need padding so update_window_and_padding() can be skipped
+    Coordinates coord;
+    coord.set_num_dimensions(output->info()->num_dimensions());
+    output->info()->set_valid_region(ValidRegion(coord, output->info()->tensor_shape()));
+
+    ICPPKernel::configure(win);
+}
+
+Status NEPadLayerKernel::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_ON_ERROR(validate_arguments(input, output, padding, mode));
+    return Status{};
+}
+
+void NEPadLayerKernel::run(const Window &window, const ThreadInfo &info)
+{
+    ARM_COMPUTE_UNUSED(info);
+    ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
+    ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window);
+
+    if(_func != nullptr)
+    {
+        (this->*_func)(window);
+    }
+}
+} // namespace arm_compute
diff --git a/src/runtime/NEON/functions/NEPadLayer.cpp b/src/runtime/NEON/functions/NEPadLayer.cpp
index c608edfdee..cf86240cab 100644
--- a/src/runtime/NEON/functions/NEPadLayer.cpp
+++ b/src/runtime/NEON/functions/NEPadLayer.cpp
@@ -34,33 +34,6 @@ namespace arm_compute
 {
 namespace
 {
-TensorInfo get_expected_output_tensorinfo(const ITensorInfo &input, const PaddingList &paddings)
-{
-    const TensorShape expected_output_shape = arm_compute::misc::shape_calculator::compute_padded_shape(input.tensor_shape(), paddings);
-    const TensorInfo  expected_output_info  = input.clone()->set_tensor_shape(expected_output_shape);
-    return expected_output_info;
-}
-
-Status validate_arguments(const ITensorInfo &input, ITensorInfo &output, const PaddingList &paddings)
-{
-    const TensorInfo expected_output_info = get_expected_output_tensorinfo(input, paddings);
-    auto_init_if_empty(output, expected_output_info);
-    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(&output, &expected_output_info);
-
-    return Status{};
-}
-
-Coordinates get_subtensor_coords(const PaddingList &paddings)
-{
-    Coordinates coords;
-    for(unsigned int i = 0; i < paddings.size(); ++i)
-    {
-        coords.set(i, paddings[i].first);
-    }
-
-    return coords;
-}
-
 uint32_t last_padding_dimension(const PaddingList &padding)
 {
     int last_padding_dim = padding.size() - 1;
@@ -76,23 +49,13 @@ uint32_t last_padding_dimension(const PaddingList &padding)
 } // namespace
 
 NEPadLayer::NEPadLayer()
-    : _copy_kernel(), _mode(), _padding(), _memset_kernel(), _num_dimensions(0), _slice_functions(), _concat_functions(), _slice_results(), _concat_results(), _output_subtensor()
+    : _copy_kernel(), _pad_kernel(), _mode(), _padding(), _num_dimensions(0), _slice_functions(), _concat_functions(), _slice_results(), _concat_results()
 {
 }
 
 void NEPadLayer::configure_constant_mode(ITensor *input, ITensor *output, const PaddingList &padding, const PixelValue constant_value)
 {
-    // Auto-init
-    auto_init_if_empty(*output->info(), get_expected_output_tensorinfo(*input->info(), padding));
-
-    // Create SubTensor (Can use sub-tensor as the kernels to be executed do not require padding)
-    _output_subtensor = SubTensor(output, input->info()->tensor_shape(), get_subtensor_coords(padding), true);
-
-    // Set the pages of the output to the specified value
-    _memset_kernel.configure(output, constant_value);
-
-    // Copy the input to the output
-    _copy_kernel.configure(input, &_output_subtensor);
+    _pad_kernel.configure(input, output, padding, constant_value, PaddingMode::CONSTANT);
 }
 
 void NEPadLayer::configure_reflect_symmetric_mode(ITensor *input, ITensor *output)
@@ -253,11 +216,7 @@ Status NEPadLayer::validate(const ITensorInfo *input, const ITensorInfo *output,
     {
         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;
+            return NEPadLayerKernel::validate(input, output, padding, constant_value, mode);
         }
         case PaddingMode::REFLECT:
         case PaddingMode::SYMMETRIC:
@@ -293,8 +252,7 @@ void NEPadLayer::run()
         {
             case PaddingMode::CONSTANT:
             {
-                NEScheduler::get().schedule(&_memset_kernel, Window::DimY);
-                NEScheduler::get().schedule(&_copy_kernel, Window::DimY);
+                NEScheduler::get().schedule(&_pad_kernel, Window::DimZ);
                 break;
             }
             case PaddingMode::REFLECT: