COMPMID-470: Neon Deconvolution.

Implemented by up-sampling the input with zeros insertions between the input samples and convolving the Deconvolution kernels on the up-sampled result.

The upsampling is performed by the function NEDeconvolutionLayerUpsample.
Convolving is done by NEDirectConvolutionLayer.

Change-Id: I25f7ba7c6b99cd9310797972ede40aeff4a54900
Reviewed-on: http://mpd-gerrit.cambridge.arm.com/85319
Tested-by: Kaizen <jeremy.johnson+kaizengerrit@arm.com>
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>

6 files changed, 275 insertions, 0 deletions

diff --git a/arm_compute/core/NEON/NEKernels.h b/arm_compute/core/NEON/NEKernels.h
index bbb440f591..5839d82ef0 100644
--- a/arm_compute/core/NEON/NEKernels.h
+++ b/arm_compute/core/NEON/NEKernels.h
@@ -43,6 +43,7 @@
 #include "arm_compute/core/NEON/kernels/NEColorConvertKernel.h"
 #include "arm_compute/core/NEON/kernels/NEConvolutionKernel.h"
 #include "arm_compute/core/NEON/kernels/NECumulativeDistributionKernel.h"
+#include "arm_compute/core/NEON/kernels/NEDeconvolutionLayerUpsampleKernel.h"
 #include "arm_compute/core/NEON/kernels/NEDepthConcatenateKernel.h"
 #include "arm_compute/core/NEON/kernels/NEDepthConvertKernel.h"
 #include "arm_compute/core/NEON/kernels/NEDequantizationLayerKernel.h"
diff --git a/arm_compute/core/NEON/kernels/NEDeconvolutionLayerUpsampleKernel.h b/arm_compute/core/NEON/kernels/NEDeconvolutionLayerUpsampleKernel.h
new file mode 100644
index 0000000000..707564683f
--- /dev/null
+++ b/arm_compute/core/NEON/kernels/NEDeconvolutionLayerUpsampleKernel.h
@@ -0,0 +1,72 @@
+/*
+ * Copyright (c) 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.
+ */
+#ifndef __ARM_COMPUTE_NEDECONVOLUTIONLAYERKERNEL_H__
+#define __ARM_COMPUTE_NEDECONVOLUTIONLAYERKERNEL_H__
+
+#include "arm_compute/core/NEON/INEKernel.h"
+#include "arm_compute/core/Types.h"
+
+namespace arm_compute
+{
+class ITensor;
+
+/** NEON kernel to perform scaling on a tensor */
+class NEDeconvolutionLayerUpsampleKernel : public INEKernel
+{
+public:
+    /** Default constructor */
+    NEDeconvolutionLayerUpsampleKernel();
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    NEDeconvolutionLayerUpsampleKernel(const NEDeconvolutionLayerUpsampleKernel &) = delete;
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    NEDeconvolutionLayerUpsampleKernel &operator=(const NEDeconvolutionLayerUpsampleKernel &) = delete;
+    /** Allow instances of this class to be moved */
+    NEDeconvolutionLayerUpsampleKernel(NEDeconvolutionLayerUpsampleKernel &&) = default;
+    /** Allow instances of this class to be moved */
+    NEDeconvolutionLayerUpsampleKernel &operator=(NEDeconvolutionLayerUpsampleKernel &&) = default;
+    /** Default destructor */
+    ~NEDeconvolutionLayerUpsampleKernel() = default;
+
+    /** Initialise the kernel's inputs, output and interpolation policy
+     *
+     * @param[in]  input   Source tensor. Data types supported: F32.
+     * @param[in]  offsets Offset to access the pixel with NEAREST interpolation or the top-left pixel with BILINEAR interpolation in the input tensor. Data type supported: S32.
+     * @param[out] output  Destination tensor. Data types supported: F32. All but the lowest two dimensions must be the same size as in the input tensor, i.e. scaling is only performed within the XY-plane.
+     */
+    void configure(const ITensor *input, const ITensor *offsets, ITensor *output);
+
+    // Inherited methods overridden:
+    void run(const Window &window, const ThreadInfo &info) override;
+    BorderSize border_size() const override;
+
+private:
+    /** Function to perform scale using nearest interpolation on the given window */
+    void scale_nearest(const Window &window);
+
+    const ITensor *_offsets;
+    const ITensor *_input;
+    ITensor       *_output;
+};
+} // arm_compute
+#endif /*__ARM_COMPUTE_NEDECONVOLUTIONLAYERKERNEL_H__ */
diff --git a/arm_compute/core/Utils.h b/arm_compute/core/Utils.h
index 06d674644b..7f53bec2c5 100644
--- a/arm_compute/core/Utils.h
+++ b/arm_compute/core/Utils.h
@@ -562,6 +562,38 @@ inline DataType data_type_for_convolution_matrix(const int16_t *conv, size_t siz
     }
 }
 
+/** Returns expected shape for the deconvolution output tensor.
+ *
+ * @param[in] out_dims widht and height of the output tensor, these values can be obtained with the function deconvolution_output_dimensions.
+ * @param[in] input    Shape of the input tensor.
+ * @param[in] weights  Shape of the weights tensor.
+ *
+ * @return Deconvolution output tensor shape.
+ */
+TensorShape deconvolution_output_shape(const std::pair<unsigned int, unsigned int> &out_dims, TensorShape input, TensorShape weights);
+
+/** Returns expected width and height of the deconvolution's output tensor.
+ *
+ * @param[in] in_width      Width of input tensor (Number of columns)
+ * @param[in] in_height     Height of input tensor (Number of rows)
+ * @param[in] kernel_width  Kernel width.
+ * @param[in] kernel_height Kernel height.
+ * @param[in] padx          X axis padding.
+ * @param[in] pady          Y axis padding.
+ * @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] round         Rounding policy to be used when computing the output's dimensions.
+ *
+ * @return A pair with the new width in the first position and the new height in the second.
+ */
+
+const std::pair<unsigned int, unsigned int> deconvolution_output_dimensions(unsigned int in_width, unsigned int in_height,
+                                                                            unsigned int kernel_width, unsigned int kernel_height,
+                                                                            unsigned int padx, unsigned int pady, unsigned int ax, unsigned int ay,
+                                                                            float upscalex, float upscaley, DimensionRoundingType round);
+
 /** Returns expected width and height of output scaled tensor depending on dimensions rounding mode.
  *
  * @param[in] width           Width of input tensor (Number of columns)
diff --git a/arm_compute/runtime/NEON/NEFunctions.h b/arm_compute/runtime/NEON/NEFunctions.h
index 40bff978aa..4e8833eed6 100644
--- a/arm_compute/runtime/NEON/NEFunctions.h
+++ b/arm_compute/runtime/NEON/NEFunctions.h
@@ -42,6 +42,8 @@
 #include "arm_compute/runtime/NEON/functions/NEColorConvert.h"
 #include "arm_compute/runtime/NEON/functions/NEConvolution.h"
 #include "arm_compute/runtime/NEON/functions/NEConvolutionLayer.h"
+#include "arm_compute/runtime/NEON/functions/NEDeconvolutionLayer.h"
+#include "arm_compute/runtime/NEON/functions/NEDeconvolutionLayerUpsample.h"
 #include "arm_compute/runtime/NEON/functions/NEDepthConcatenate.h"
 #include "arm_compute/runtime/NEON/functions/NEDepthConvert.h"
 #include "arm_compute/runtime/NEON/functions/NEDequantizationLayer.h"
diff --git a/arm_compute/runtime/NEON/functions/NEDeconvolutionLayer.h b/arm_compute/runtime/NEON/functions/NEDeconvolutionLayer.h
new file mode 100644
index 0000000000..3433e77ba1
--- /dev/null
+++ b/arm_compute/runtime/NEON/functions/NEDeconvolutionLayer.h
@@ -0,0 +1,96 @@
+/*
+ * Copyright (c) 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.
+ */
+#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"
+#include "arm_compute/runtime/IFunction.h"
+#include "arm_compute/runtime/IMemoryManager.h"
+#include "arm_compute/runtime/MemoryGroup.h"
+#include "arm_compute/runtime/Tensor.h"
+
+#include <memory>
+
+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.
+ *
+ *  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 )
+ *
+ *  where
+ *      width is the size of the first input dimension.
+ *      height is the size of the second input dimension.
+ *      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.
+ *
+ *  This function calls the following NEON kernels:
+ *
+ * -# @ref NEDeconvolutionLayerUpsampleKernel
+ * -# @ref NEDirectConvolutionLayer
+ *
+ */
+class NEDeconvolutionLayer : public IFunction
+{
+public:
+    /** Constructor */
+    NEDeconvolutionLayer(std::shared_ptr<IMemoryManager> memory_manager = nullptr);
+    /** 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.
+      *
+      */
+    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);
+
+    // Inherited methods overridden:
+    void run() override;
+
+private:
+    MemoryGroup                  _memory_group;
+    NEDeconvolutionLayerUpsample _scale_f;
+    NEDirectConvolutionLayer     _conv_f;
+    Tensor                       _scaled_output;
+};
+} // arm_compute
+#endif /* __ARM_COMPUTE_NEDECONVOLUTIONLAYER_H__ */
diff --git a/arm_compute/runtime/NEON/functions/NEDeconvolutionLayerUpsample.h b/arm_compute/runtime/NEON/functions/NEDeconvolutionLayerUpsample.h
new file mode 100644
index 0000000000..d2ac12a58a
--- /dev/null
+++ b/arm_compute/runtime/NEON/functions/NEDeconvolutionLayerUpsample.h
@@ -0,0 +1,72 @@
+/*
+ * 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.
+ */
+#ifndef __ARM_COMPUTE_NEDECONVOLUTIONUPSAMPLE_H__
+#define __ARM_COMPUTE_NEDECONVOLUTIONUPSAMPLE_H__
+
+#include "arm_compute/core/NEON/kernels/NEDeconvolutionLayerUpsampleKernel.h"
+#include "arm_compute/core/NEON/kernels/NEFillBorderKernel.h"
+#include "arm_compute/core/Types.h"
+#include "arm_compute/runtime/IFunction.h"
+#include "arm_compute/runtime/IMemoryManager.h"
+#include "arm_compute/runtime/MemoryGroup.h"
+#include "arm_compute/runtime/Tensor.h"
+
+#include <cstdint>
+#include <memory>
+
+namespace arm_compute
+{
+class ITensor;
+
+/** Basic function to run @ref NEDeconvolutionLayerUpsampleKernel */
+class NEDeconvolutionLayerUpsample : public IFunction
+{
+public:
+    /** Constructor
+     *
+     * Initialize NEDeconvolutionLayerUpsample
+     */
+    NEDeconvolutionLayerUpsample(std::shared_ptr<IMemoryManager> memory_manager = nullptr);
+    /** Initialize the function's source, destination, interpolation type and border_mode.
+     *
+     * @param[in, out] input  Source tensor. Data type supported: F32.
+     * @param[out]     output Destination tensor. Data type supported: F32.
+     * @param[in]      a      Top and right inner border sizes. These rows and columns will be filled with zero.
+     * @param[in]      iz     The number of zeros to be inserted between each input sample
+     * @param[in]      info   Contains padding and policies to be used in the deconvolution, this is decribed in @ref PadStrideInfo.
+     */
+    void configure(ITensor *input, ITensor *output, const std::pair<unsigned int, unsigned int> &a,
+                   const std::pair<unsigned int, unsigned int> &iz, const PadStrideInfo &info);
+
+    // Inherited methods overridden:
+    void run() override;
+
+private:
+    MemoryGroup                        _memory_group;
+    Tensor                             _offsets;
+    NEFillBorderKernel                 _border_handler;
+    NEDeconvolutionLayerUpsampleKernel _upsample;
+};
+} // arm_compute
+#endif /*__ARM_COMPUTE_NEDECONVOLUTIONUPSAMPLE_H__ */