COMPMID-1047 Extract Flatten function from Im2Col for NEON

Change-Id: I80f3aaadc8cae8c9ca1a5a239e79bda302b89bd8
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/144813
Tested-by: Jenkins <bsgcomp@arm.com>
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>

7 files changed, 141 insertions, 65 deletions

diff --git a/arm_compute/core/NEON/NEKernels.h b/arm_compute/core/NEON/NEKernels.h
index d99e5b4d3c..8664c7732c 100644
--- a/arm_compute/core/NEON/NEKernels.h
+++ b/arm_compute/core/NEON/NEKernels.h
@@ -61,6 +61,7 @@
 #include "arm_compute/core/NEON/kernels/NEFillArrayKernel.h"
 #include "arm_compute/core/NEON/kernels/NEFillBorderKernel.h"
 #include "arm_compute/core/NEON/kernels/NEFillInnerBorderKernel.h"
+#include "arm_compute/core/NEON/kernels/NEFlattenLayerKernel.h"
 #include "arm_compute/core/NEON/kernels/NEFloorKernel.h"
 #include "arm_compute/core/NEON/kernels/NEGEMMAssemblyBaseKernel.h"
 #include "arm_compute/core/NEON/kernels/NEGEMMInterleave4x4Kernel.h"
diff --git a/arm_compute/core/NEON/kernels/NEFlattenLayerKernel.h b/arm_compute/core/NEON/kernels/NEFlattenLayerKernel.h
new file mode 100644
index 0000000000..e5c388a95a
--- /dev/null
+++ b/arm_compute/core/NEON/kernels/NEFlattenLayerKernel.h
@@ -0,0 +1,81 @@
+/*
+ * Copyright (c) 2018 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_NEFLATTENLAYERKERNEL_H__
+#define __ARM_COMPUTE_NEFLATTENLAYERKERNEL_H__
+
+#include "arm_compute/core/NEON/INEKernel.h"
+
+namespace arm_compute
+{
+class ITensor;
+
+/** Interface for the flatten layer kernel. */
+class NEFlattenLayerKernel : public INEKernel
+{
+public:
+    const char *name() const override
+    {
+        return "NEFlattenLayerKernel";
+    }
+    /** Default constructor */
+    NEFlattenLayerKernel();
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    NEFlattenLayerKernel(const NEFlattenLayerKernel &) = delete;
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    NEFlattenLayerKernel &operator=(const NEFlattenLayerKernel &) = delete;
+    /** Allow instances of this class to be moved */
+    NEFlattenLayerKernel(NEFlattenLayerKernel &&) = default;
+    /** Allow instances of this class to be moved */
+    NEFlattenLayerKernel &operator=(NEFlattenLayerKernel &&) = default;
+    /** Default destructor */
+    ~NEFlattenLayerKernel() = default;
+
+    /** Set the input and output of the kernel.
+     *
+     * @param[in]  input  First input tensor to flatten with at least 3 dimensions.
+     *                    The dimensions above the third will be interpreted as batches. Data types supported: U8/S8/QASYMM8/U16/S16/F16/U32/S32/F32
+     * @param[out] output Output tensor with shape [w*h*d, input_batches] where:
+     *                    w = width input tensor, h = height input tensor and d = depth input tensor. Data type supported: same as @p input
+     */
+    void configure(const ITensor *input, ITensor *output);
+    /** Static function to check if given info will lead to a valid configuration of @ref NEFlattenLayerKernel
+     *
+     * @param[in]  input  First input tensor to flatten with at least 3 dimensions.
+     *                    The dimensions above the third will be interpreted as batches. Data types supported: U8/S8/QASYMM8/U16/S16/F16/U32/S32/F32
+     * @param[out] output Output tensor with shape [w*h*d, input_batches] where:
+     *                    w = width input tensor, h = height input tensor and d = depth input tensor. Data type supported: same as @p input
+     *
+     * @return a status
+     */
+    static Status validate(const ITensorInfo *input, const ITensorInfo *output);
+
+    // Inherited methods overridden:
+    void run(const Window &window, const ThreadInfo &info) override;
+
+private:
+    const ITensor *_input;
+    ITensor       *_output;
+};
+} // namespace arm_compute
+#endif /*__ARM_COMPUTE_NEFLATTENLAYERKERNEL_H__ */
diff --git a/arm_compute/core/NEON/kernels/NEIm2ColKernel.h b/arm_compute/core/NEON/kernels/NEIm2ColKernel.h
index 38bdff0e98..ec89f0f713 100644
--- a/arm_compute/core/NEON/kernels/NEIm2ColKernel.h
+++ b/arm_compute/core/NEON/kernels/NEIm2ColKernel.h
@@ -76,55 +76,46 @@ public:
 
     /** Set the input and output of the kernel.
      *
-     * @param[in]  input              The input tensor to convert. 3 lower dimensions represent a single input [width, height, IFM],
-     *                                while every optional dimension from 4 and above represent a batch of inputs. Data types supported: QASYMM8/F16/F32
-     *                                Note: QASYMM8 works only for has_bias = false
-     * @param[out] output             The output tensor. Data types supported: Same as @p input
-     * @param[in]  kernel_dims        The kernel dimensions (width and height).
-     * @param[in]  conv_info          Contains padding and stride information described in @ref PadStrideInfo.
-     * @param[in]  has_bias           In case biases are provided expands the matrix with 1.
-     * @param[in]  dilation           (Optional) Dilation, in elements, across x and y. Defaults to (1, 1).
-     * @param[in]  num_groups         (Optional) Number of groups when performing a grouped convolution. num_groups != 1 is not supported
-     * @param[in]  is_fully_connected (Optional) Determines whether this kernel will be called by @ref NEFullyConnectedLayer in order to validate the arguments
-     * @param[in]  is_flatten         (Optional) Determines whether this kernel will be called by @ref NEFlattenLayer in order to validate the arguments
+     * @param[in]  input       The input tensor to convert. 3 lower dimensions represent a single input [width, height, IFM],
+     *                         while every optional dimension from 4 and above represent a batch of inputs. Data types supported: QASYMM8/F16/F32
+     *                         Note: QASYMM8 works only for has_bias = false
+     * @param[out] output      The output tensor. Data types supported: Same as @p input
+     * @param[in]  kernel_dims The kernel dimensions (width and height).
+     * @param[in]  conv_info   Contains padding and stride information described in @ref PadStrideInfo.
+     * @param[in]  has_bias    In case biases are provided expands the matrix with 1.
+     * @param[in]  dilation    (Optional) Dilation, in elements, across x and y. Defaults to (1, 1).
+     * @param[in]  num_groups  (Optional) Number of groups when performing a grouped convolution. num_groups != 1 is not supported
      */
     void configure(const ITensor *input, ITensor *output, const Size2D &kernel_dims, const PadStrideInfo &conv_info,
-                   bool has_bias, const Size2D &dilation = Size2D(1U, 1U), unsigned int num_groups = 1, bool is_fully_connected = false, bool is_flatten = false);
+                   bool has_bias, const Size2D &dilation = Size2D(1U, 1U), unsigned int num_groups = 1);
     /** Static function to check if given info will lead to a valid configuration of @ref NEIm2ColKernel
      *
-     * @param[in] input              The input tensor to convert. 3 lower dimensions represent a single input [width, height, IFM],
-     *                               while every optional dimension from 4 and above represent a batch of inputs. Data types supported: QASYMM8/F16/F32
-     *                               Note: QASYMM8 works only for has_bias = false
-     * @param[in] output             The output tensor. Data types supported: Same as @p input
-     * @param[in] kernel_dims        The kernel dimensions (width and height).
-     * @param[in] conv_info          Contains padding and stride information described in @ref PadStrideInfo.
-     * @param[in] has_bias           In case biases are provided expands the matrix with 1.
-     * @param[in] dilation           (Optional) Dilation, in elements, across x and y. Defaults to (1, 1).
-     * @param[in] num_groups         (Optional) Number of groups when performing a grouped convolution. num_groups != 1 is not supported
-     * @param[in] is_fully_connected (Optional)Determines whether this kernel will be called by @ref NEFullyConnectedLayer in order to validate the arguments
-     * @param[in] is_flatten         (Optional) Determines whether this kernel will be called by @ref NEFlattenLayer in order to validate the arguments
+     * @param[in] input       The input tensor to convert. 3 lower dimensions represent a single input [width, height, IFM],
+     *                        while every optional dimension from 4 and above represent a batch of inputs. Data types supported: QASYMM8/F16/F32
+     *                        Note: QASYMM8 works only for has_bias = false
+     * @param[in] output      The output tensor. Data types supported: Same as @p input
+     * @param[in] kernel_dims The kernel dimensions (width and height).
+     * @param[in] conv_info   Contains padding and stride information described in @ref PadStrideInfo.
+     * @param[in] has_bias    In case biases are provided expands the matrix with 1.
+     * @param[in] dilation    (Optional) Dilation, in elements, across x and y. Defaults to (1, 1).
+     * @param[in] num_groups  (Optional) Number of groups when performing a grouped convolution. num_groups != 1 is not supported
      *
      * @return a status
      */
     static Status validate(const ITensorInfo *input, const ITensorInfo *output, const Size2D &kernel_dims, const PadStrideInfo &conv_info,
-                           bool has_bias, const Size2D &dilation = Size2D(1U, 1U), unsigned int num_groups = 1, bool is_fully_connected = false, bool is_flatten = false);
+                           bool has_bias, const Size2D &dilation = Size2D(1U, 1U), unsigned int num_groups = 1);
 
     // Inherited methods overridden:
     void run(const Window &window, const ThreadInfo &info) override;
 
 private:
-    /** Template function to run the im2col optimised for the fully connected and flatten layers case
-     *
-     * @param[in] window Region on which to execute the kernel. (Must be a valid region of the window returned by window()).
-     */
-    template <typename T>
-    void run_reduced(const Window &window);
-    /** Template function to run the im2col used for the convolution layer case
+    /** Template function to run the im2col
      *
      * @param[in] window Region on which to execute the kernel. (Must be a valid region of the window returned by window()).
      */
     template <typename T, bool has_pads>
-    void run_generic(const Window &window);
+    void run_im2col(const Window &window);
+
     /** Common signature for all the specialised im2col functions
      *
      * @param[in] window Region on which to execute the kernel.
diff --git a/arm_compute/runtime/CL/functions/CLFlattenLayer.h b/arm_compute/runtime/CL/functions/CLFlattenLayer.h
index ebc0e5e53f..24ed56f345 100644
--- a/arm_compute/runtime/CL/functions/CLFlattenLayer.h
+++ b/arm_compute/runtime/CL/functions/CLFlattenLayer.h
@@ -47,7 +47,7 @@ public:
      *                    w = width input tensor, h = height input tensor and d = depth input tensor. Data type supported: same as @p input
      */
     void configure(const ICLTensor *input, ICLTensor *output);
-    /** Static function to check if given info will lead to a valid configuration of @ref CLTranspose
+    /** Static function to check if given info will lead to a valid configuration of @ref CLFlattenLayer
      *
      * @param[in]  input  First input tensor to flatten with at least 3 dimensions.
      *                    The dimensions above the third will be interpreted as batches. Data types supported: U8/S8/QASYMM8/U16/S16/F16/U32/S32/F32
diff --git a/arm_compute/runtime/NEON/functions/NEFlattenLayer.h b/arm_compute/runtime/NEON/functions/NEFlattenLayer.h
index 2c259fa178..26d7c7f636 100644
--- a/arm_compute/runtime/NEON/functions/NEFlattenLayer.h
+++ b/arm_compute/runtime/NEON/functions/NEFlattenLayer.h
@@ -31,11 +31,7 @@ namespace arm_compute
 {
 class ITensor;
 
-/** Basic function to execute flatten. This function calls the following NEON kernel:
-*
-* -# @ref NEIm2ColKernel
-*
-*/
+/** Basic function to execute flatten layer kernel. */
 class NEFlattenLayer : public INESimpleFunction
 {
 public:
@@ -46,6 +42,17 @@ public:
      *             w = width input tensor, h = height input tensor and d = depth input tensor. Data type supported: same as @p input
      */
     void configure(const ITensor *input, ITensor *output);
+
+    /** Static function to check if given info will lead to a valid configuration of @ref NEFlattenLayer
+     *
+     * @param[in]  input  First input tensor to flatten with at least 3 dimensions.
+     *                    The dimensions above the third will be interpreted as batches. Data types supported: U8/S8/QASYMM8/U16/S16/F16/U32/S32/F32
+     * @param[out] output Output tensor with shape [w*h*d, input_batches] where:
+     *                    w = width input tensor, h = height input tensor and d = depth input tensor. Data type supported: same as @p input
+     *
+     * @return a status
+     */
+    static Status validate(const ITensorInfo *input, const ITensorInfo *output);
 };
 } // namespace arm_compute
 
diff --git a/arm_compute/runtime/NEON/functions/NEFullyConnectedLayer.h b/arm_compute/runtime/NEON/functions/NEFullyConnectedLayer.h
index fe0f2f03f7..9c9074ceec 100644
--- a/arm_compute/runtime/NEON/functions/NEFullyConnectedLayer.h
+++ b/arm_compute/runtime/NEON/functions/NEFullyConnectedLayer.h
@@ -26,8 +26,8 @@
 
 #include "arm_compute/runtime/IFunction.h"
 
+#include "arm_compute/core/NEON/kernels/NEFlattenLayerKernel.h"
 #include "arm_compute/core/NEON/kernels/NEGEMMMatrixAccumulateBiasesKernel.h"
-#include "arm_compute/core/NEON/kernels/NEIm2ColKernel.h"
 #include "arm_compute/core/NEON/kernels/NETransposeKernel.h"
 #include "arm_compute/runtime/MemoryGroup.h"
 #include "arm_compute/runtime/NEON/functions/NEConvertFullyConnectedWeights.h"
@@ -129,14 +129,14 @@ private:
     void configure_mm(const ITensor *input, const ITensor *weights, ITensor *output);
 
     MemoryGroup                                         _memory_group;
-    NEIm2ColKernel                                      _im2col_kernel;
+    NEFlattenLayerKernel                                _flatten_kernel;
     NEConvertFullyConnectedWeights                      _convert_weights;
     NEFullyConnectedLayerReshapeWeights                 _reshape_weights_function;
     NEGEMM                                              _mm_gemm;
     NEGEMMLowpMatrixMultiplyCore                        _mm_gemmlowp;
     NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPoint _gemmlowp_output_stage;
     NEGEMMMatrixAccumulateBiasesKernel                  _accumulate_biases_kernel;
-    Tensor                                              _im2col_output;
+    Tensor                                              _flatten_output;
     Tensor                                              _gemmlowp_output;
     Tensor                                              _converted_weights_output;
     Tensor                                              _reshape_weights_output;
diff --git a/arm_compute/runtime/NEON/functions/NEIm2Col.h b/arm_compute/runtime/NEON/functions/NEIm2Col.h
index 9df4f070d8..de4780f8f0 100644
--- a/arm_compute/runtime/NEON/functions/NEIm2Col.h
+++ b/arm_compute/runtime/NEON/functions/NEIm2Col.h
@@ -42,38 +42,34 @@ public:
     NEIm2Col();
     /** Configure the im2col NEON kernel
      *
-     * @param[in]  input              The input tensor to convert. 3 lower dimensions represent a single input [width, height, IFM],
-     *                                while every optional dimension from 4 and above represent a batch of inputs. Data types supported: QASYMM8/F16/F32
-     *                                Note: QASYMM8 works only for has_bias = false
-     * @param[out] output             The output tensor. Data types supported: Same as @p input
-     * @param[in]  kernel_dims        The kernel dimensions (width and height).
-     * @param[in]  conv_info          Contains padding and stride information described in @ref PadStrideInfo.
-     * @param[in]  has_bias           In case biases are provided expands the matrix with 1.
-     * @param[in]  dilation           (Optional) Dilation, in elements, across x and y. Defaults to (1, 1).
-     * @param[in]  num_groups         (Optional) Number of groups when performing a grouped convolution
-     * @param[in]  is_fully_connected (Optional) Determines whether this function will be called by @ref NEFullyConnectedLayer in order to validate the arguments
-     * @param[in]  is_flatten         (Optional) Determines whether this function will be called by @ref NEFlattenLayer in order to validate the arguments
+     * @param[in]  input       The input tensor to convert. 3 lower dimensions represent a single input [width, height, IFM],
+     *                         while every optional dimension from 4 and above represent a batch of inputs. Data types supported: QASYMM8/F16/F32
+     *                         Note: QASYMM8 works only for has_bias = false
+     * @param[out] output      The output tensor. Data types supported: Same as @p input
+     * @param[in]  kernel_dims The kernel dimensions (width and height).
+     * @param[in]  conv_info   Contains padding and stride information described in @ref PadStrideInfo.
+     * @param[in]  has_bias    In case biases are provided expands the matrix with 1.
+     * @param[in]  dilation    (Optional) Dilation, in elements, across x and y. Defaults to (1, 1).
+     * @param[in]  num_groups  (Optional) Number of groups when performing a grouped convolution
      */
     void configure(const ITensor *input, ITensor *output, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias, const Size2D &dilation = Size2D(1U, 1U),
-                   unsigned int num_groups = 1, bool is_fully_connected = false, bool is_flatten = false);
+                   unsigned int num_groups = 1);
     /** Static function to check if given info will lead to a valid configuration of @ref NEIm2Col
      *
-     * @param[in] input              The input tensor to convert. 3 lower dimensions represent a single input [width, height, IFM],
-     *                               while every optional dimension from 4 and above represent a batch of inputs. Data types supported: QASYMM8/F16/F32
-     *                               Note: QASYMM8 works only for has_bias = false
-     * @param[in] output             The output tensor. Data types supported: Same as @p input
-     * @param[in] kernel_dims        The kernel dimensions (width and height).
-     * @param[in] conv_info          Contains padding and stride information described in @ref PadStrideInfo.
-     * @param[in] has_bias           In case biases are provided expands the matrix with 1.
-     * @param[in] dilation           (Optional) Dilation, in elements, across x and y. Defaults to (1, 1).
-     * @param[in] num_groups         (Optional) Number of groups when performing a grouped convolution
-     * @param[in] is_fully_connected (Optional) Determines whether this function will be called by @ref NEFullyConnectedLayer in order to validate the arguments
-     * @param[in] is_flatten         (Optional) Determines whether this function will be called by @ref NEFlattenLayer in order to validate the arguments
+     * @param[in] input       The input tensor to convert. 3 lower dimensions represent a single input [width, height, IFM],
+     *                        while every optional dimension from 4 and above represent a batch of inputs. Data types supported: QASYMM8/F16/F32
+     *                        Note: QASYMM8 works only for has_bias = false
+     * @param[in] output      The output tensor. Data types supported: Same as @p input
+     * @param[in] kernel_dims The kernel dimensions (width and height).
+     * @param[in] conv_info   Contains padding and stride information described in @ref PadStrideInfo.
+     * @param[in] has_bias    In case biases are provided expands the matrix with 1.
+     * @param[in] dilation    (Optional) Dilation, in elements, across x and y. Defaults to (1, 1).
+     * @param[in] num_groups  (Optional) Number of groups when performing a grouped convolution
      *
      * @return a status
      */
     static Status validate(const ITensorInfo *input, const ITensorInfo *output, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias, const Size2D &dilation = Size2D(1U, 1U),
-                           unsigned int num_groups = 1, bool is_fully_connected = false, bool is_flatten = false);
+                           unsigned int num_groups = 1);
 
     // Inherited methods overridden:
     void run() override;