COMPMID-846: Create a ConvolutionLayer for NEON

Change-Id: I98bbef40bfac5b05134be4ef9fb54d14c0c9e8e8
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/118806
Tested-by: Jenkins <bsgcomp@arm.com>
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>

1 files changed, 30 insertions, 102 deletions

diff --git a/arm_compute/runtime/NEON/functions/NEConvolutionLayer.h b/arm_compute/runtime/NEON/functions/NEConvolutionLayer.h
index f80f67d944..6ab1350b25 100644
--- a/arm_compute/runtime/NEON/functions/NEConvolutionLayer.h
+++ b/arm_compute/runtime/NEON/functions/NEConvolutionLayer.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017-2018 ARM Limited.
+ * Copyright (c) 2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -26,79 +26,27 @@
 
 #include "arm_compute/runtime/IFunction.h"
 
-#include "arm_compute/core/NEON/kernels/NECol2ImKernel.h"
-#include "arm_compute/core/NEON/kernels/NEFillBorderKernel.h"
-#include "arm_compute/core/NEON/kernels/NEGEMMAssemblyBaseKernel.h"
-#include "arm_compute/core/NEON/kernels/NEGEMMInterleave4x4Kernel.h"
-#include "arm_compute/core/NEON/kernels/NEGEMMMatrixMultiplyKernel.h"
-#include "arm_compute/core/NEON/kernels/NEGEMMTranspose1xWKernel.h"
-#include "arm_compute/core/NEON/kernels/NEIm2ColKernel.h"
-#include "arm_compute/core/NEON/kernels/NEWeightsReshapeKernel.h"
 #include "arm_compute/core/Types.h"
 #include "arm_compute/runtime/MemoryGroup.h"
-#include "arm_compute/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.h"
-#include "arm_compute/runtime/NEON/functions/NEGEMMLowpOutputStage.h"
-#include "arm_compute/runtime/Tensor.h"
+#include "arm_compute/runtime/NEON/functions/NEDirectConvolutionLayer.h"
+#include "arm_compute/runtime/NEON/functions/NEGEMMConvolutionLayer.h"
+#include "arm_compute/runtime/NEON/functions/NEWinogradLayer.h"
 #include <memory>
 
 namespace arm_compute
 {
 class ITensor;
 
-/** Function to reshape and perform 1xW transposition on the weights. This function calls the following kernels:
- * -# @ref NEWeightsReshapeKernel
- * -# @ref NEGEMMTranspose1xWKernel (executed in case GEMM is required for the operation)
- */
-class NEConvolutionLayerReshapeWeights : public IFunction
-{
-public:
-    /** Constructor */
-    NEConvolutionLayerReshapeWeights(std::shared_ptr<IMemoryManager> memory_manager = nullptr);
-    /** Set the input and output tensors.
-     *
-     * @param[in]  weights      Weights tensor. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM]. Data type supported: QS8/QASYMM8/QS16/F32.
-     * @param[in]  biases       Biases tensor. Shared biases supported. Biases are 1D tensor with dimensions [OFM]. Data type supported: Same as @p weights.
-     * @param[out] output       Destination tensor. Data types supported: Same as @p weights.
-     * @param[in]  transpose1xW True if the weights are to undergo a 1xW transposition after reshaping (in case of GEMM operation), false otherwise.
-     *                          Data types supported: Same as @p weights.
-     */
-    void configure(const ITensor *weights, const ITensor *biases, ITensor *output, bool transpose1xW);
-    /** Static function to check if given info will lead to a valid configuration of @ref NEConvolutionLayerReshapeWeights
-     *
-     * @param[in] weights      Weights tensor. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM]. Data type supported: QS8/QASYMM8/QS16/F16/F32.
-     * @param[in] biases       Biases tensor. Shared biases supported. Biases are 1D tensor with dimensions [OFM]. Data type supported: Same as @p weights.
-     * @param[in] output       Destination tensor. Data types supported: Same as @p weights.
-     * @param[in] transpose1xW True if the weights are to undergo a 1xW transposition after reshaping (in case of GEMM operation), false otherwise.
-     *                         Data types supported: Same as @p weights.
-     *
-     * @return an error status
-     */
-    static Status validate(const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, bool transpose1xW);
-
-    // Inherited methods overridden:
-    void run() override;
-
-private:
-    MemoryGroup              _memory_group;
-    NEWeightsReshapeKernel   _weights_reshape_kernel;
-    NEGEMMTranspose1xWKernel _weights_transposed_kernel;
-    Tensor                   _weights_reshaped;
-    bool                     _transpose1xW;
-};
-
-/** Basic function to simulate a convolution layer. This function calls the following NEON kernels:
- * -# @ref NEWeightsReshapeKernel   (executed only once for each configuration)
- * -# @ref NEIm2ColKernel
- * -# @ref NEGEMMInterleave4x4Kernel (executed only in case GEMM is required for the operation)
- * -# @ref NEGEMMMatrixMultiplyKernel or @ref NEGEMMLowpMatrixMultiplyCore (if quantized asymmetric)
- * -# @ref NEGEMMLowpQuantizeDownInt32ToUint8Scale (if quantized asymmetric)
- * -# @ref NECol2ImKernel
+/** Basic function to simulate a convolution layer. This function calls one of the following NEON functions:
+ * -# @ref NEGEMMConvolutionLayer   (executed only in case GEMM is required for the operation)
+ * -# @ref NEWinogradLayer (executed only in case Winograd is required for the operation)
+ * -# @ref NEDirectConvolutionLayer (executed only in case Direct Convolution is required for the operation)
  */
 class NEConvolutionLayer : public IFunction
 {
 public:
     /** Constructor */
-    NEConvolutionLayer(const std::shared_ptr<IMemoryManager> &memory_manager = nullptr);
+    NEConvolutionLayer(std::shared_ptr<IMemoryManager> memory_manager = nullptr);
 
     /** Set the input and output tensors.
      *
@@ -114,7 +62,7 @@ public:
      * @param[in]  weights_info Specifies if the weights tensor has been reshaped with NEWeightsReshapeKernel. If this is not part of the fully connected layer the weights
      *                          tensor has also been transposed with NEGEMMTranspose1xWKernel. Data type supported: Same as @p input.
      */
-    void configure(const ITensor *input, const ITensor *weights, const ITensor *biases, ITensor *output, const PadStrideInfo &conv_info, const WeightsInfo &weights_info = WeightsInfo());
+    void configure(ITensor *input, const ITensor *weights, const ITensor *biases, ITensor *output, const PadStrideInfo &conv_info, const WeightsInfo &weights_info = WeightsInfo());
     /** Static function to check if given info will lead to a valid configuration of @ref NEConvolutionLayer
      *
      * @param[in] input        Source tensor. 3 lower dimensions represent a single input [width, height, IFM],
@@ -133,51 +81,31 @@ public:
      */
     static Status validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info,
                            const WeightsInfo &weights_info = WeightsInfo());
-
-    // Inherited methods overridden:
-    void run() override;
-
-private:
-    /** Configures the appropriate matrix multiply routine
+    /** Static function to check if given info will return the convolution called by @ref NEConvolutionLayer
      *
-     * @param[in]  input   Input tensor. Data types supported: QS8/QASYMM8/QS16/F16/F32.
-     * @param[in]  weights Weights tensor. Data type supported: Same as @p input.
-     * @param[out] output  Output tensor. Data types supported: Same as @p input,
-     *                     except for input of QASYMM8 type where output should be of S32 type.
-     */
-    void configure_mm(const ITensor *input, const ITensor *weights, ITensor *output);
-    /** Prepare the appropriate assembly optimized kernel
+     * @param[in] input        Source tensor. 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: QS8/QASYMM8/QS16/F16/F32.
+     * @param[in] weights      Weights tensor. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM]. Data type supported:Same as @p input.
+     * @param[in] biases       Biases tensor. Shared biases supported. Biases are 1D tensor with dimensions [OFM].
+     *                         Data type supported: Should match @p input data type, except for input of QASYMM8 type where biases should be of S32 type.
+     * @param[in] output       Destination tensor. 3 lower dimensions represent a single output [width, height, OFM], while the rest represent batch of outputs.
+     *                         Data types supported: Same as @p input.
+     * @param[in] conv_info    Contains padding and stride information described in @ref PadStrideInfo.
+     * @param[in] weights_info Specifies if the weights tensor has been reshaped with NEWeightsReshapeKernel. If this is not part of the fully connected layer the weights
+     *                         tensor has also been transposed with NEGEMMTranspose1xWKernel. Data type supported: Same as @p input.
      *
-     * @param[in] ci CPU information
-     * @param[in] M  M parameter of matrix multiplication
-     * @param[in] N  N parameter of matrix multiplication
-     * @param[in] K  K parameter of matrix multiplication
+     * @return the Convolution Method Hint
      */
-    void configure_asm_mm(const struct CPUInfo &ci, int M, int N, int K);
+    static ConvolutionMethod get_convolution_method(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info,
+                                                    const WeightsInfo &weights_info = WeightsInfo());
 
-private:
-    MemoryGroup                                         _memory_group;
-    NEIm2ColKernel                                      _input_im2col_kernel;
-    NEGEMMInterleave4x4Kernel                           _input_interleave_kernel;
-    NEConvolutionLayerReshapeWeights                    _reshape_weights;
-    NEGEMMMatrixMultiplyKernel                          _mm_kernel;
-    std::unique_ptr<NEGEMMAssemblyBaseKernel>           _mm_optimised_kernel;
-    NEGEMMLowpMatrixMultiplyCore                        _mm_gemmlowp;
-    NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPoint _gemmlowp_output_stage;
-    NECol2ImKernel                                      _output_col2im_kernel;
-
-    Tensor _input_im2col_reshaped;
-    Tensor _input_interleaved_reshaped;
-    Tensor _weights_reshaped;
-    Tensor _gemm_output;
-    Tensor _tmp_output;
-    Tensor _workspace;
+    // Inherited methods overridden:
+    void run() override;
 
-    bool _append_bias;
-    bool _is_fully_connected_convolution;
-    bool _are_weights_reshaped;
-    bool _is_quantized;
-    bool _is_interleaved_transposed;
+private:
+    std::shared_ptr<IMemoryManager> _memory_manager;
+    std::unique_ptr<IFunction>      _function; /**< Function to run */
 };
 }
 #endif /* __ARM_COMPUTE_NECONVOLUTIONLAYER_H__ */