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>

1 files changed, 35 insertions, 26 deletions

diff --git a/arm_compute/runtime/NEON/functions/NEDeconvolutionLayer.h b/arm_compute/runtime/NEON/functions/NEDeconvolutionLayer.h
index 8757bc63aa..091a928db6 100644
--- a/arm_compute/runtime/NEON/functions/NEDeconvolutionLayer.h
+++ b/arm_compute/runtime/NEON/functions/NEDeconvolutionLayer.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017 ARM Limited.
+ * Copyright (c) 2017, 2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -24,7 +24,6 @@
 #ifndef __ARM_COMPUTE_NEDECONVOLUTIONLAYER_H__
 #define __ARM_COMPUTE_NEDECONVOLUTIONLAYER_H__
 
-#include "arm_compute/runtime/NEON/functions/NEDeconvolutionLayerUpsample.h"
 #include "arm_compute/runtime/NEON/functions/NEDirectConvolutionLayer.h"
 
 #include "arm_compute/core/Types.h"
@@ -39,13 +38,13 @@ namespace arm_compute
 {
 /** Function to run the deconvolution layer.
  *
- *  The operation is similar to convolution but it's implemented by up-sampling the inputs with zeros insertions between the inputs and convolving
- *  the kernels on the up-sampled result.
+ * Deconvolution Layer is the backward pass of Convolution Layer. First we transform the input depending on the stride and pad info and then perfrom a 1x1
+ * convolution pass. Input stride defines how many zeroes we should put between each element of the input, pad is the amount of padding and finaly a is a user
+ * specified value where a < stride - 1 that increases the padding top and right of the input image.
  *
- *  Before the Deconvolution is done, up-scaling the first 2D with zeros is performed. The relation between input to
- *  output is as follows:
- *      width_output = round((width_input − 1) ∗ upscale_x − 2 ∗ padding_x + kernel_x + a_x )
- *      height_output = round((height_input − 1) ∗ upscale_y − 2 ∗ padding_y + kernel_y + a_y )
+ *  The relation between input to output is as follows:
+ *       width_output = round((width_input − 1) ∗ (stride_x - 1) − 2 ∗ padding_x + kernel_x + inner_border_right )
+ *       height_output = round((height_input − 1) ∗ (stride_y - 1) − 2 ∗ padding_y + kernel_y + inner_border_top )
  *
  *  where
  *      width is the size of the first input dimension.
@@ -53,44 +52,54 @@ namespace arm_compute
  *      width_output is the size of the first output dimension.
  *      height_output is the size of the second output dimension.
  *      kernel_x and kernel_y are the convolution sizes in x and y.
- *      ax and ay the number of zeros added to the top and right edges of the input.
- *      upscale_x and upscale_y how much to scale the X and Y axis.
+ *      inner_border_right and inner_border_top the number of zeros added to the top and right edges of the input.
+ *      stride_x and stride_y is the input stride of the first and second dimension.
  *
  *  This function calls the following NEON kernels:
  *
- * -# @ref NEDeconvolutionLayerUpsampleKernel
  * -# @ref NEDirectConvolutionLayer
  *
  */
 class NEDeconvolutionLayer : public IFunction
 {
 public:
-    /** Constructor */
+    /** Default constructor */
     NEDeconvolutionLayer(std::shared_ptr<IMemoryManager> memory_manager = nullptr);
+
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    NEDeconvolutionLayer(const NEDeconvolutionLayer &) = delete;
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    NEDeconvolutionLayer &operator=(const NEDeconvolutionLayer &) = delete;
+    /** Allow instances of this class to be moved */
+    NEDeconvolutionLayer(NEDeconvolutionLayer &&) = default;
+    /** Allow instances of this class to be moved */
+    NEDeconvolutionLayer &operator=(NEDeconvolutionLayer &&) = default;
+    /** Default destructor */
+    virtual ~NEDeconvolutionLayer() = default;
     /** Set the input, weights, biases and output tensors.
      *
-     * @param[in,out] input    Input tensor. 3 lower dimensions represent a single input, and an optional 4th dimension for batch of inputs. Data types supported: F32.
-     * @param[in]     weights  The 4d weights with dimensions [width, height, OFM, IFM]. Data type supported: Same as @p input.
-     * @param[in]     bias     Optional, ignored if NULL. The biases have one dimension. Data type supported: Same as @p input.
-     * @param[out]    output   Output tensor. The output has the same number of dimensions as the @p input.
-     * @param[in]     info     Contains padding and policies to be used in the deconvolution, this is decribed in @ref PadStrideInfo.
-     * @param[in]     ax       The number of zeros added to right edge of the input.
-     * @param[in]     ay       The number of zeros added to top edge of the input.
-     * @param[in]     upscalex How much to scale the X axis.
-     * @param[in]     upscaley How much to scale the Y axis.
+     * @param[in,out] input              Input tensor. 3 lower dimensions represent a single input, and an optional 4th dimension for batch of inputs. Data types supported: F32.
+     * @param[in]     weights            The 4d weights with dimensions [width, height, OFM, IFM]. Data type supported: Same as @p input.
+     * @param[in]     bias               Optional, ignored if NULL. The biases have one dimension. Data type supported: Same as @p input.
+     * @param[out]    output             Output tensor. The output has the same number of dimensions as the @p input.
+     * @param[in]     info               Contains padding and policies to be used in the deconvolution, this is decribed in @ref PadStrideInfo.
+     * @param[in]     inner_border_right The number of zeros added to right edge of the input.
+     * @param[in]     inner_border_top   The number of zeros added to top edge of the input.
      *
      */
     void 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);
 
     // Inherited methods overridden:
     void run() override;
 
 private:
-    MemoryGroup                  _memory_group;
-    NEDeconvolutionLayerUpsample _scale_f;
-    NEDirectConvolutionLayer     _conv_f;
-    Tensor                       _scaled_output;
+    MemoryGroup              _memory_group;
+    NEDirectConvolutionLayer _conv_f;
+    Tensor                   _scaled_output;
+    ITensor                 *_input;
+    PadStrideInfo            _info;
+    std::pair<unsigned int, unsigned int> _inner_border;
 };
 } // arm_compute
 #endif /* __ARM_COMPUTE_NEDECONVOLUTIONLAYER_H__ */