COMPMID-471 Implement Deconvolution on OpenCL

Change-Id: Ie00c6b08a51d30c5ce2637d40ee3d165b8a68686
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/110311
Reviewed-by: Pablo Tello <pablo.tello@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
Tested-by: Jenkins <bsgcomp@arm.com>

2 files changed, 54 insertions, 172 deletions

diff --git a/src/runtime/NEON/functions/NEDeconvolutionLayer.cpp b/src/runtime/NEON/functions/NEDeconvolutionLayer.cpp
index 7b4e77b296..c4bca11d14 100644
--- a/src/runtime/NEON/functions/NEDeconvolutionLayer.cpp
+++ b/src/runtime/NEON/functions/NEDeconvolutionLayer.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017 ARM Limited.
+ * Copyright (c) 2017, 2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -24,38 +24,41 @@
 #include "arm_compute/runtime/NEON/functions/NEDeconvolutionLayer.h"
 
 #include "arm_compute/core/Helpers.h"
-#include "arm_compute/core/PixelValue.h"
 #include "arm_compute/core/Utils.h"
 #include "arm_compute/core/Validate.h"
+#include "arm_compute/core/utils/misc/ShapeCalculator.h"
 
 using namespace arm_compute;
+using namespace arm_compute::misc::shape_calculator;
 
 NEDeconvolutionLayer::NEDeconvolutionLayer(std::shared_ptr<IMemoryManager> memory_manager) // NOLINT
     : _memory_group(std::move(memory_manager)),
-      _scale_f(),
       _conv_f(),
-      _scaled_output()
+      _scaled_output(),
+      _input(nullptr),
+      _info(),
+      _inner_border()
 {
 }
 
 void NEDeconvolutionLayer::configure(ITensor *input, const ITensor *weights, const ITensor *bias, ITensor *output, const PadStrideInfo &info,
-                                     unsigned int ax, unsigned int ay, float upscalex, float upscaley)
+                                     unsigned int inner_border_right, unsigned int inner_border_top)
 {
     ARM_COMPUTE_ERROR_ON_NULLPTR(output);
     ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32);
     ARM_COMPUTE_ERROR_ON(weights->info()->dimension(0) != weights->info()->dimension(1));
-    ARM_COMPUTE_ERROR_ON(weights->info()->dimension(0) < 1);
+    ARM_COMPUTE_ERROR_ON(weights->info()->dimension(0) != 1 && weights->info()->dimension(0) != 3 && weights->info()->dimension(0) != 5);
 
-    auto out_dims = deconvolution_output_dimensions(input->info()->dimension(0), input->info()->dimension(1), weights->info()->dimension(0), weights->info()->dimension(1),
-                                                    info.pad().first, info.pad().second, ax, ay, upscalex, upscaley, info.round());
+    _input        = input;
+    _info         = info;
+    _inner_border = std::make_pair(inner_border_right, inner_border_top);
 
-    const TensorShape output_shape = deconvolution_output_shape(out_dims, input->info()->tensor_shape(), weights->info()->tensor_shape());
-
-    // Output auto initialization if not yet initialized
-    auto_init_if_empty(*output->info(), output_shape, 1, input->info()->data_type(), input->info()->fixed_point_position());
+    const unsigned int stride_x = info.stride().first;
+    const unsigned int stride_y = info.stride().second;
+    auto               out_dims = deconvolution_output_dimensions(input->info()->dimension(0), input->info()->dimension(1), weights->info()->dimension(0), weights->info()->dimension(1),
+                                                                  info.pad().first, info.pad().second, inner_border_right, inner_border_top, stride_x, stride_y);
 
-    ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, output, weights, bias);
-    ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(input, output, weights, bias);
+    const TensorShape output_shape = deconvolution_output_shape(out_dims, input->info()->tensor_shape(), weights->info()->tensor_shape());
 
     ARM_COMPUTE_ERROR_ON_MSG(output->info()->dimension(Window::DimX) != output_shape.x(), "Output's width is invalid.");
     ARM_COMPUTE_ERROR_ON_MSG(output->info()->dimension(Window::DimY) != output_shape.y(), "Output's height is invalid.");
@@ -64,51 +67,51 @@ void NEDeconvolutionLayer::configure(ITensor *input, const ITensor *weights, con
     _memory_group.manage(&_scaled_output);
 
     // configure scale function
-    //Init and allocate intermmidiate tensor for output, same size as input but the first two axis are the same as the output tensor
-    TensorShape scale_out_shape(input->info()->tensor_shape());
-    scale_out_shape.set(0, output->info()->dimension(0));
-    scale_out_shape.set(1, output->info()->dimension(1));
-    TensorInfo scale_out_info(scale_out_shape, 1, input->info()->data_type(), input->info()->fixed_point_position());
+    // Init and allocate intermmidiate tensor for output, same size as input but the first two axis are the same as the output tensor
+    const TensorInfo scale_out_info(compute_deconvolution_shape(*input->info(), stride_x, stride_y, inner_border_right, inner_border_top, info), 1, input->info()->data_type(),
+                                    input->info()->fixed_point_position());
     _scaled_output.allocator()->init(scale_out_info);
-    const unsigned int kernel_size = weights->info()->dimension(0);
-    // Padding for the upsampled image is calculated with the equiation: p' = k - p - 1, where k is kernel size and p is the input padding
-    ARM_COMPUTE_ERROR_ON(info.pad().first > (kernel_size - 1));
-    const unsigned int  tr_px     = kernel_size - info.pad().first - 1;
-    const unsigned int  tr_py     = kernel_size - info.pad().second - 1;
-    const unsigned int  tr_stride = 1;
-    const PadStrideInfo transposed_info(tr_stride, tr_stride, tr_px, tr_py);
-    _scale_f.configure(input, &_scaled_output, std::make_pair(ax, ay), std::make_pair(info.stride().first - 1u, info.stride().second - 1u), transposed_info);
+
     // setup the function to convolve the upscaled output
-    switch(kernel_size)
-    {
-        case 1:
-        {
-            _conv_f.configure(&_scaled_output, weights, bias, output, PadStrideInfo(1, 1, 0, 0, DimensionRoundingType::CEIL));
-            break;
-        }
-        case 3:
-        {
-            _conv_f.configure(&_scaled_output, weights, bias, output, PadStrideInfo(1, 1, 1, 1, DimensionRoundingType::CEIL));
-            break;
-        }
-        case 5:
-        {
-            _conv_f.configure(&_scaled_output, weights, bias, output, PadStrideInfo(1, 1, 2, 2, DimensionRoundingType::CEIL));
-            break;
-        }
-        default:
-        {
-            ARM_COMPUTE_ERROR("Not supported");
-            break;
-        }
-    }
+    const PadStrideInfo conv_info(1, 1, 0, 0, 0, 0, DimensionRoundingType::CEIL);
+    _conv_f.configure(&_scaled_output, weights, bias, output, conv_info);
     _scaled_output.allocator()->allocate();
 }
 
 void NEDeconvolutionLayer::run()
 {
     _memory_group.acquire();
-    _scale_f.run();
+
+    // Initialize _scaled_output buffer
+    const int width_in      = _input->info()->dimension(0);
+    const int height_in     = _input->info()->dimension(1);
+    const int width_scaled  = _scaled_output.info()->dimension(0);
+    const int height_scaled = _scaled_output.info()->dimension(1);
+    const int num_2d_slices = _input->info()->tensor_shape().total_size() / (width_in * height_in);
+    const int stride_x      = _info.stride().first;
+    const int stride_y      = _info.stride().second;
+
+    std::fill_n(reinterpret_cast<float *>(_scaled_output.buffer()), _scaled_output.info()->tensor_shape().total_size(), 0.f);
+
+    // scaled_output is the input for the forward convolution. We copy the input elements to scaled_output
+    // and insert rows and columns with zeroes depending on the stride values.
+    for(int slice = 0; slice < num_2d_slices; ++slice)
+    {
+        const int start_x = _info.pad().first;
+        const int start_y = _inner_border.second + _info.pad().second;
+        const int end_y   = height_scaled - _info.pad().second;
+        const int end_x   = width_scaled - _inner_border.first - _info.pad().first;
+
+        for(int yi = start_y, in_y = 0; yi < end_y; yi += stride_y, in_y++)
+        {
+            for(int xi = start_x, in_x = 0; xi < end_x; xi += stride_x, in_x++)
+            {
+                const auto in = *(reinterpret_cast<float *>(_input->buffer() + _input->info()->offset_element_in_bytes(Coordinates(in_x, in_y, slice))));
+                *(reinterpret_cast<float *>(_scaled_output.buffer() + _scaled_output.info()->offset_element_in_bytes(Coordinates(xi, yi, slice)))) = in;
+            }
+        }
+    }
+
     _conv_f.run();
     _memory_group.release();
 }
diff --git a/src/runtime/NEON/functions/NEDeconvolutionLayerUpsample.cpp b/src/runtime/NEON/functions/NEDeconvolutionLayerUpsample.cpp
deleted file mode 100644
index 63f17bcb5a..0000000000
--- a/src/runtime/NEON/functions/NEDeconvolutionLayerUpsample.cpp
+++ /dev/null
@@ -1,121 +0,0 @@
-/*
- * Copyright (c) 2016, 2017 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/runtime/NEON/functions/NEDeconvolutionLayerUpsample.h"
-
-#include "arm_compute/core/Coordinates.h"
-#include "arm_compute/core/Error.h"
-#include "arm_compute/core/Helpers.h"
-#include "arm_compute/core/ITensor.h"
-#include "arm_compute/core/NEON/kernels/NEDeconvolutionLayerUpsampleKernel.h"
-#include "arm_compute/core/PixelValue.h"
-#include "arm_compute/core/TensorInfo.h"
-#include "arm_compute/core/Window.h"
-#include "arm_compute/runtime/NEON/NEScheduler.h"
-#include "arm_compute/runtime/TensorAllocator.h"
-#include "support/ToolchainSupport.h"
-
-#include <cmath>
-#include <cstddef>
-#include <utility>
-
-using namespace arm_compute;
-
-namespace
-{
-inline void precompute_offsets(ITensor *offsets, float wr, size_t input_element_size, const std::pair<unsigned int, unsigned int> &a,
-                               const std::pair<unsigned int, unsigned int> &iz, const PadStrideInfo &info)
-{
-    ARM_COMPUTE_ERROR_ON(nullptr == offsets);
-    Window    win;
-    const int padx          = info.pad().first;
-    const int pady          = info.pad().second;
-    const int ax            = a.first;
-    const int ay            = a.second;
-    const int offset_width  = offsets->info()->dimension(0);
-    const int offset_height = offsets->info()->dimension(1);
-    // The values of ax and ay denote the number of ZEROS to be added on the top and right inner border of the image.
-    // Step value along the XY axis will depend on the number of zeros to be inserted between samples (number of zeros + 1).
-    // Pre-compute the X offset, Y's stride is unknown at this point so we can't precompute Y's offsets
-    for(int yi = ay; yi < (offset_height - pady); yi += (1 + iz.second))
-    {
-        for(int xi = padx; xi < (offset_width - ax); xi += (1 + iz.first))
-        {
-            int         *ptr                  = reinterpret_cast<int *>(offsets->ptr_to_element(Coordinates(xi, yi)));
-            const size_t in_xi                = (xi + 0.5f) * wr;
-            *reinterpret_cast<int32_t *>(ptr) = in_xi * input_element_size;
-        }
-    }
-}
-} // namespace
-
-NEDeconvolutionLayerUpsample::NEDeconvolutionLayerUpsample(std::shared_ptr<IMemoryManager> memory_manager) // NOLINT
-    : _memory_group(std::move(memory_manager)),
-      _offsets(),
-      _border_handler(),
-      _upsample()
-{
-}
-
-void NEDeconvolutionLayerUpsample::configure(ITensor *input, ITensor *output, const std::pair<unsigned int, unsigned int> &a,
-                                             const std::pair<unsigned int, unsigned int> &iz, const PadStrideInfo &info)
-{
-    ARM_COMPUTE_ERROR_ON(nullptr == input);
-    ARM_COMPUTE_ERROR_ON(nullptr == output);
-
-    for(size_t i = 2; i < Coordinates::num_max_dimensions; ++i)
-    {
-        ARM_COMPUTE_ERROR_ON(input->info()->dimension(i) != output->info()->dimension(i));
-    }
-
-    // Get the tensor shape
-    const TensorShape shape(output->info()->dimension(0), output->info()->dimension(1));
-
-    // Compute the ratio between source width/height and destination width/height
-    const auto wr = static_cast<float>(input->info()->dimension(0)) / static_cast<float>(output->info()->dimension(0));
-    const auto hr = static_cast<float>(input->info()->dimension(1)) / static_cast<float>(output->info()->dimension(1));
-    ARM_COMPUTE_UNUSED(hr);
-    // Get the element size of the input image
-    const size_t input_element_size = input->info()->element_size();
-
-    TensorInfo tensor_info_offsets(shape, Format::S32);
-    _offsets.allocator()->init(tensor_info_offsets);
-
-    _upsample.configure(input, &_offsets, output);
-
-    // Allocate once the configure methods have been called
-    _offsets.allocator()->allocate();
-    // Pre-compute offsets for nearest interpolation
-    std::fill_n(reinterpret_cast<int32_t *>(_offsets.buffer()), _offsets.info()->total_size() / sizeof(int32_t), -1 * input_element_size);
-    precompute_offsets(&_offsets, wr, input_element_size, a, iz, info);
-
-    _border_handler.configure(input, _upsample.border_size(), BorderMode::CONSTANT, PixelValue(0));
-}
-
-void NEDeconvolutionLayerUpsample::run()
-{
-    NEScheduler::get().schedule(&_border_handler, Window::DimZ);
-    _memory_group.acquire();
-    NEScheduler::get().schedule(&_upsample, Window::DimY);
-    _memory_group.release();
-}