COMPMID-784: Winograd tramsforms refactoring

1) Removed the example files winograd_layer.hpp/cpp
2) Teplatized winograd transform kernels

Change-Id: I7045fa0b801b9d30a11275914aaa2dafd254aed2
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/118332
Tested-by: Jenkins <bsgcomp@arm.com>
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>

7 files changed, 324 insertions, 617 deletions

diff --git a/arm_compute/core/NEON/kernels/NEWinogradLayerKernel.h b/arm_compute/core/NEON/kernels/NEWinogradLayerKernel.h
index ea6c8d813d..97532f3574 100644
--- a/arm_compute/core/NEON/kernels/NEWinogradLayerKernel.h
+++ b/arm_compute/core/NEON/kernels/NEWinogradLayerKernel.h
@@ -25,104 +25,93 @@
 #define __ARM_COMPUTE_NEGEMMWINOGRADLAYERKERNEL_H__
 
 #include "arm_compute/core/NEON/INEKernel.h"
+#include "arm_compute/core/NEON/kernels/winograd/batched_blocked_gemm.hpp"
 #include "arm_compute/core/NEON/kernels/winograd/convolution.hpp"
 #include "arm_compute/core/NEON/kernels/winograd/tensor.hpp"
+#include "arm_compute/core/NEON/kernels/winograd/winograd_gemm.hpp"
 
 namespace arm_compute
 {
 class ITensor;
-class NEWinogradLayerKernel;
-class NEWinogradLayerTransformInputKernel;
-class NEWinogradLayerTransformWeightsKernel;
 
-class Winograd3x3F32 final
+template <int OutputTileRows, int OutputTileCols, int KernelRows, int KernelCols>
+class NEWinogradLayerTransformInputKernel : public INEKernel
 {
 public:
-    /** Create a new Winograd convolution layer.
+    /** Determine how much memory (in units of TIn) to allocate for the
+     * transformed input.
      *
-     * @param[in]  n_batches         Number of batches in the input and output tensors.
-     * @param[in]  n_input_channels  Number of feature maps in a batch of the input tensor.
-     * @param[in]  n_input_rows      Number of rows in a feature map of the input tensor.
-     * @param[in]  n_input_cols      Number of columns in a feature map of the input tensor.
-     * @param[in]  n_output_channels Number of feature maps in the output tensor.
-     * @param[in]  same_padding      Use "SAME" padding, otherwise use "VALID".
-     * @param[in]  weights           Pointer to weight tensor in spatial domain. Must be ordered as "Height x Rows x Input Feature Maps x Output Feature Maps.
-     * @param[out] weights_storage   Pointer to storage for weight tensor in the Winograd domain. Must be at least the size returned by `get_weight_storage_size
-     * @param[in]  input             Pointer to NHWC ordered input tensor, in the spatial domain.
-     * @param[out] winograd_input    Pointer to working space for the input tensor in the Winograd domain. Must be at least the size returned by `get_input_storage_size`.
-     * @param[in]  biases            Pointer to the biases vector.
-     * @param[out] output            Pointer to NHWC ordered output tensor, in the spatial domain.
-     * @param[out] winograd_output   Pointer to working space for the output tensor in the Winograd domain. Must be at least the size returned by `get_output_storage_size`.
+     * @param[in] n_batches    Number of batches in the input tensor.
+     * @param[in] n_channels   Number of feature maps in the input tensor.
+     * @param[in] n_rows       Number of rows in each feature map.
+     * @param[in] n_cols       Number of columns in each feature map.
+     * @param[in] same_padding Use "SAME" padding, otherwise use "VALID".
      */
-    friend class NEWinogradLayerKernel;
-    friend class NEWinogradLayerTransformInputKernel;
-    friend class NEWinogradLayerTransformOutputKernel;
-    friend class NEWinogradLayerTransformWeightsKernel;
+    static unsigned int get_input_storage_size(
+        int  n_batches,
+        int  n_channels,
+        int  n_rows,
+        int  n_cols,
+        bool same_padding);
 
-    Winograd3x3F32(
-        const int          n_batches,
-        const int          n_input_channels,
-        const int          n_input_rows,
-        const int          n_input_cols,
-        const int          n_output_channels,
-        const bool         same_padding,
-        const float *const weights,
-        float *const       weights_storage,
+    NEWinogradLayerTransformInputKernel();
+    const char *name() const override
+    {
+        return "NEWinogradLayerTransformInputKernel";
+    }
+
+    /** Configure the output transform kernel.
+     *
+     * @param[in]  input         Input tensor data
+     * @param[in]  n_batches     Number of batches in input tensor.
+     * @param[in]  n_rows        Number of rows in input tensor.
+     * @param[in]  n_cols        Number of columns in input tensor.
+     * @param[in]  n_channels    Number of channels in input tensor.
+     * @param[in]  padding       Padding type.
+     * @param[out] output        Base of output matrices.
+     * @param[in]  matrix_stride Stride between output matrices.
+     */
+    void configure(
         const float *const input,
-        float *const       winograd_input,
+        const int          n_batches,
+        const int          n_rows,
+        const int          n_cols,
+        const int          n_channels,
+        const PaddingType  padding,
         float *const       output,
-        float *const       winograd_output);
+        const int          matrix_stride);
 
-    ~Winograd3x3F32();
+    // Inherited methods overridden:
+    void run(const Window &window, const ThreadInfo &info) override;
+    bool is_parallelisable() const override;
 
 private:
-    class Private;
-    std::unique_ptr<Private> _pimpl;
+    using WinogradBase   = winograd::WinogradGEMM<OutputTileRows, OutputTileCols, KernelCols, KernelCols>;
+    using WinogradConv   = typename WinogradBase::template Convolution<float, float>;
+    using InputTransform = typename WinogradBase::template InputTransform<float>;
+    std::unique_ptr<InputTransform> _transform;
 };
 
-class INEWinogradLayerTransformKernel : public INEKernel
+template <int OutputTileRows, int OutputTileCols, int KernelRows, int KernelCols>
+class NEWinogradLayerTransformOutputKernel : public INEKernel
 {
 public:
-    /** Constructor */
-    INEWinogradLayerTransformKernel();
-
-    /** Prevent instances of this class from being copied (As this class contains pointers) */
-    INEWinogradLayerTransformKernel(const INEWinogradLayerTransformKernel &) = delete;
-    /** Prevent instances of this class from being copied (As this class contains pointers) */
-    INEWinogradLayerTransformKernel &operator=(const INEWinogradLayerTransformKernel &) = delete;
-    /** Allow instances of this class to be moved */
-    INEWinogradLayerTransformKernel(INEWinogradLayerTransformKernel &&) = default;
-    /** Allow instances of this class to be moved */
-    INEWinogradLayerTransformKernel &operator=(INEWinogradLayerTransformKernel &&) = default;
-
-    virtual ~INEWinogradLayerTransformKernel() = default;
-
-    /** Initialise the kernel
+    /** Determine how much memory (in units of TOut) to allocate for the
+     * (Winograd domain) output.
      *
-     * @param[in] convolver A pointer to the winograd convolver, this object must have been configured and is ready to execute 16 GEMMS .
+     * @param[in] n_batches         Number of batches in the output tensor.
+     * @param[in] n_rows            Number of rows in each feature map of the input tensor.
+     * @param[in] n_cols            Number of columns in each feature map of the input tensor.
+     * @param[in] n_output_channels Number of feature maps in the output tensor.
+     * @param[in] same_padding      Use "SAME" padding, otherwise use "VALID".
      */
-    virtual void configure(Winograd3x3F32 *convolver);
-
-protected:
-    Winograd3x3F32 *_convolver;
-};
-
-class NEWinogradLayerTransformInputKernel final : public INEWinogradLayerTransformKernel
-{
-public:
-    const char *name() const override
-    {
-        return "NEWinogradLayerTransformInputKernel";
-    }
-    // Inherited methods overridden:
-    void configure(Winograd3x3F32 *convolver) override;
-    void run(const Window &window, const ThreadInfo &info) override;
-    bool is_parallelisable() const override;
-};
+    static unsigned int get_output_storage_size(
+        int  n_batches,
+        int  n_rows,
+        int  n_cols,
+        int  n_output_channels,
+        bool same_padding);
 
-class NEWinogradLayerTransformOutputKernel final : public INEKernel
-{
-public:
     const char *name() const override
     {
         return "NEWinogradLayerTransformOutputKernel";
@@ -167,6 +156,10 @@ public:
     bool is_parallelisable() const override;
 
 private:
+    using WinogradBase    = winograd::WinogradGEMM<OutputTileRows, OutputTileCols, KernelRows, KernelCols>;
+    using WinogradConv    = typename WinogradBase::template Convolution<float, float>;
+    using OutputTransform = typename WinogradBase::template OutputTransform<float>;
+
     const ITensor *_biases;
     const float   *_output_workspace;
     int            _matrix_stride;
@@ -178,22 +171,61 @@ private:
     int            _n_channels;
 };
 
-class NEWinogradLayerTransformWeightsKernel final : public INEWinogradLayerTransformKernel
+template <int OutputTileRows, int OutputTileCols, int KernelRows, int KernelCols>
+class NEWinogradLayerTransformWeightsKernel final : public INEKernel
 {
 public:
+    /** Determine how much memory (in units of TIn) to allocate for the
+     * transformed weights.
+     *
+     * @param[in] n_output_channels Number of output feature maps.
+     * @param[in] n_input_channels  Number of input feature maps.
+     */
+    static unsigned int get_weight_storage_size(int n_output_channels, int n_input_channels);
+
+    NEWinogradLayerTransformWeightsKernel();
     const char *name() const override
     {
         return "NEWinogradLayerTransformWeightsKernel";
     }
+    /** Configure the output transform kernel.
+     *
+     * @param[in] weights_hwio      Pointer to the weights tensor
+     * @param[in] output            Pointer to working space for the output tensor in the Winograd domain.
+     * @param[in] matrix_stride     Stride across matrices in the output workspace.
+     * @param[in] n_output_channels Number of filters.
+     * @param[in] n_input_channels  Number of channels in each filter.
+     */
+    void configure(
+        const ITensor *weights_hwio,
+        float *const   output,
+        const int      matrix_stride,
+        const int      n_output_channels,
+        const int      n_input_channels);
+
     // Inherited methods overridden:
-    void configure(Winograd3x3F32 *convolver) override;
+
     void run(const Window &window, const ThreadInfo &info) override;
     bool is_parallelisable() const override;
+
+private:
+    using WinogradBase     = winograd::WinogradGEMM<OutputTileRows, OutputTileCols, KernelRows, KernelCols>;
+    using WinogradConv     = typename WinogradBase::template Convolution<float, float>;
+    using WeightsTransform = typename WinogradBase::template WeightsTransform<float>;
+    std::unique_ptr<WeightsTransform> _transform;
 };
 
-class NEWinogradLayerKernel final : public INEKernel
+template <int OutputTileRows, int OutputTileCols, int KernelRows, int KernelCols>
+class NEWinogradLayerKernel : public INEKernel
 {
 public:
+    using WinogradBase = winograd::WinogradGEMM<OutputTileRows, OutputTileCols, KernelRows, KernelCols>;
+    using WinogradConv = typename WinogradBase::template Convolution<float, float>;
+    using MultiGEMM    = winograd::BatchedBlockedGemm<WinogradConv::M_BLOCK, WinogradConv::N_BLOCK, float, float>;
+
+    static const int _output_tile_rows = OutputTileRows;
+    static const int _output_tile_cols = OutputTileCols;
+
     const char *name() const override
     {
         return "NEWinogradLayerKernel";
@@ -214,57 +246,38 @@ public:
 
     /** Initialise the kernel
      *
-     * @param[in] convolver A pointer to the winograd convolver, this object must have been configured and is ready to execute 16 GEMMS .
+     * @param[in]  n_gemms         Number of GEMMs to compute.
+     * @param[in]  M               in_shape.n_batches * tile_rows * tile_cols.
+     * @param[in]  K               Number of channels in the input tensor.
+     * @param[in]  N               Number of channels in the output tensor.
+     * @param[in]  a_matrix_stride Stride between input matrices.
+     * @param[in]  a_row_stride    Row stride inside input matrix.
+     * @param[in]  b_matrix_stride Stride between weights matrices.
+     * @param[in]  b_row_stride    Row stride inside the weights matrix.
+     * @param[in]  c_matrix_stride Stride between output matrices.
+     * @param[in]  c_row_stride    Row stride inside the output matrix.
+     * @param[out] a_ptr           Input workspace.
+     * @param[out] b_ptr           Kernel workspace.
+     * @param[out] c_ptr           Output workspace.
      */
-    void configure(Winograd3x3F32 *convolver);
+    void configure(
+        const unsigned int n_gemms,
+        const int M, const int K, const int N,
+        const int          a_matrix_stride,
+        const int          a_row_stride,
+        const int          b_matrix_stride,
+        const int          b_row_stride,
+        const int          c_matrix_stride,
+        const int          c_row_stride,
+        const float *const a_ptr,
+        const float *const b_ptr,
+        float *const       c_ptr);
 
     // Inherited methods overridden:
     void run(const Window &window, const ThreadInfo &info) override;
 
-    /** Determine how much memory (in units of TIn) to allocate for the
-     * transformed weights.
-     *
-     * @param[in] n_output_channels Number of output feature maps.
-     * @param[in] n_input_channels  Number of input feature maps.
-     */
-    static unsigned int get_weight_storage_size(
-        const int n_output_channels,
-        const int n_input_channels);
-
-    /** Determine how much memory (in units of TIn) to allocate for the
-     * transformed input.
-     *
-     * @param[in] n_batches    Number of batches in the input tensor.
-     * @param[in] n_channels   Number of feature maps in the input tensor.
-     * @param[in] n_rows       Number of rows in each feature map.
-     * @param[in] n_cols       Number of columns in each feature map.
-     * @param[in] same_padding Use "SAME" padding, otherwise use "VALID".
-     */
-    static unsigned int get_input_storage_size(
-        const int  n_batches,
-        const int  n_channels,
-        const int  n_rows,
-        const int  n_cols,
-        const bool same_padding);
-
-    /** Determine how much memory (in units of TOut) to allocate for the
-     * (Winograd domain) output.
-     *
-     * @param[in] n_batches         Number of batches in the output tensor.
-     * @param[in] n_rows            Number of rows in each feature map of the input tensor.
-     * @param[in] n_cols            Number of columns in each feature map of the input tensor.
-     * @param[in] n_output_channels Number of feature maps in the output tensor.
-     * @param[in] same_padding      Use "SAME" padding, otherwise use "VALID".
-     */
-    static unsigned int get_output_storage_size(
-        const int  n_batches,
-        const int  n_rows,
-        const int  n_cols,
-        const int  n_output_channels,
-        const bool same_padding);
-
-protected:
-    Winograd3x3F32 *_convolver;
+private:
+    std::unique_ptr<MultiGEMM> _gemms;
 };
 
 } // namespace arm_compute
diff --git a/arm_compute/core/NEON/kernels/winograd/winograd_layer.hpp b/arm_compute/core/NEON/kernels/winograd/winograd_layer.hpp
deleted file mode 100644
index 1db63d750b..0000000000
--- a/arm_compute/core/NEON/kernels/winograd/winograd_layer.hpp
+++ /dev/null
@@ -1,129 +0,0 @@
-/*
- * 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.
- */
-
-#pragma once
-
-#include <utility>
-
-#include "batched_blocked_gemm.hpp"
-#include "winograd_gemm.hpp"
-
-/** Example of how to construct an ACL-like interface.
- *
- * Use `get_weight_storage_size`, `get_input_storage_size` and
- * `get_output_storage_size` to allocate memory for the convolution engine.
- * Then create a `WinogradConvolutionLayer`.
- *
- * Initialise the weights using `weights_transform.run(...)`.
- *
- * For each inference:
- *   1. Transform the inputs to the Winograd domain using `input_transform.run(...)`
- *   2. Perform a number of GEMMs using `gemms.run(...)`
- *   3. Transform the output to the spatial domain using `output_transform.run(...)`
- */
-template <int OutputTileRows, int OutputTileCols, int KernelRows, int KernelCols,
-          typename TIn, typename TOut>
-class WinogradConvolutionLayer
-{
-  private:
-    const KernelShape _kernel_shape;
-    const Tensor4DShape _input_shape;
-    const PaddingType _padding;
-    const Tensor4DShape _output_shape;
-    const int _n_output_rows, _n_output_cols;
-    const int _kernel_matrix_stride, _kernel_matrix_row_stride;
-    const int _input_matrix_stride, _input_matrix_row_stride;
-    const int _output_matrix_stride, _output_matrix_row_stride;
-    const int _tile_rows, _tile_cols;
-    const int _m, _k, _n;
-
-  public:
-    using WinogradBase = winograd::WinogradGEMM<OutputTileRows, OutputTileCols, KernelRows, KernelCols>;
-    using WeightsTransform = typename WinogradBase::template WeightsTransform<TIn>;
-    using InputTransform = typename WinogradBase::template InputTransform<TIn>;
-    using WinogradConv = typename WinogradBase::template Convolution<TOut, TIn>;
-    using MultiGEMM = winograd::BatchedBlockedGemm<WinogradConv::M_BLOCK, WinogradConv::N_BLOCK, TIn, TOut>;
-    using OutputTransform = typename WinogradBase::template OutputTransform<TOut>;
-
-    /* Public member variables. */
-    WeightsTransform weights_transform;  /** Operator to transform weights to Winograd domain. */
-    InputTransform input_transform;      /** Operator to transform input to Winograd domain. */
-    MultiGEMM gemms;                     /** Operator to perform multiple GEMMs. */
-    OutputTransform output_transform;    /** Operator to transform output from Winograd domain. */
-
-    /** Determine how much memory (in units of TIn) to allocate for the
-     * transformed weights.
-     */
-    static unsigned int get_weight_storage_size(
-      const int n_output_channels,  /** Number of output feature maps. */
-      const int n_input_channels    /** Number of input feature maps. */
-    );
-
-    /** Determine how much memory (in units of TIn) to allocate for the
-     * transformed input.
-     */
-    static unsigned int get_input_storage_size(
-      const int n_batches,     /** Number of batches in the input tensor. */
-      const int n_channels,    /** Number of feature maps in the input tensor. */
-      const int n_rows,        /** Number of rows in each feature map. */
-      const int n_cols,        /** Number of columns in each feature map. */
-      const bool same_padding  /** Use "SAME" padding, otherwise use "VALID". */
-    );
-
-    /** Determine how much memory (in units of TOut) to allocate for the
-     * (Winograd domain) output.
-     */
-    static unsigned int get_output_storage_size(
-      const int n_batches,          /** Number of batches in the output tensor. */
-      const int n_rows,             /** Number of rows in each feature map of the input tensor. */
-      const int n_cols,             /** Number of columns in each feature map of the input tensor. */
-      const int n_output_channels,  /** Number of feature maps in the output tensor. */
-      const bool same_padding       /** Use "SAME" padding, otherwise use "VALID". */
-    );
-
-    /** Get the shape (rows, cols) of a feature map of the output tensor. */
-    static std::pair<int, int> get_output_feature_map_shape(
-      const int n_input_rows,  /** Number of rows in the input feature map. */
-      const int n_input_cols,  /** Number of columns in the input feature map. */
-      const bool same_padding  /** Use "SAME" padding, otherwise use "VALID". */
-    );
-
-    /** Create a new Winograd convolution layer.
-     */
-    WinogradConvolutionLayer(
-      const int n_batches,          /** Number of batches in the input and output tensors. */
-      const int n_input_channels,   /** Number of feature maps in a batch of the input tensor. */
-      const int n_input_rows,       /** Number of rows in a feature map of the input tensor. */
-      const int n_input_cols,       /** Number of columns in a feature map of the input tensor. */
-      const int n_output_channels,  /** Number of feature maps in the output tensor. */
-      const bool same_padding,      /** Use "SAME" padding, otherwise use "VALID". */
-      const TIn* const weights,     /** Pointer to weight tensor in spatial domain. Must be ordered as "Height x Rows x Input Feature Maps x Output Feature Maps. */
-      TIn* const weights_storage,   /** Pointer to storage for weight tensor in the Winograd domain. Must be at least the size returned by `get_weight_storage_size`. */
-      const TIn* const input,       /** Pointer to NHWC ordered input tensor, in the spatial domain. */
-      TIn* const winograd_input,    /** Pointer to working space for the input tensor in the Winograd domain. Must be at least the size returned by `get_input_storage_size`. */
-      const TOut* const biases,     /** Pointer to biases vector. */
-      TOut* const output,           /** Pointer to NHWC ordered output tensor, in the spatial domain. */
-      TOut* const winograd_output   /** Pointer to working space for the output tensor in the Winograd domain. Must be at least the size returned by `get_output_storage_size`. */
-    );
-};
diff --git a/arm_compute/runtime/NEON/functions/NEWinogradLayer.h b/arm_compute/runtime/NEON/functions/NEWinogradLayer.h
index 1682495f0d..63cac3a3b4 100644
--- a/arm_compute/runtime/NEON/functions/NEWinogradLayer.h
+++ b/arm_compute/runtime/NEON/functions/NEWinogradLayer.h
@@ -68,25 +68,24 @@ public:
     NEWinogradLayer &operator=(const NEWinogradLayer &) = delete;
 
 private:
-    MemoryGroup                           _memory_group;
-    NEWinogradLayerKernel                 _winograd_kernel;
-    NEWinogradLayerTransformInputKernel   _transform_input_kernel;
-    NEWinogradLayerTransformOutputKernel  _transform_output_kernel;
-    NEWinogradLayerTransformWeightsKernel _transform_weights_kernel;
-    CPPPermute                            _permute_input;
-    CPPPermute                            _permute_weights;
-    CPPPermute                            _permute_output;
-    Tensor                                _input_workspace;
-    Tensor                                _output_workspace;
-    Tensor                                _kernel_storage;
-    Tensor                                _input_nhwc;
-    Tensor                                _output_nhwc;
-    Tensor                                _weights_hwio;
-    const ITensor                        *_input;
-    const ITensor                        *_weights;
-    ITensor                              *_output;
-    bool                                  _reshaped_kernel;
-    std::unique_ptr<Winograd3x3F32>       _conv;
+    MemoryGroup _memory_group;
+    NEWinogradLayerKernel<2, 2, 3, 3>                 _winograd_kernel;
+    NEWinogradLayerTransformInputKernel<2, 2, 3, 3>   _transform_input_kernel;
+    NEWinogradLayerTransformOutputKernel<2, 2, 3, 3>  _transform_output_kernel;
+    NEWinogradLayerTransformWeightsKernel<2, 2, 3, 3> _transform_weights_kernel;
+    CPPPermute     _permute_input;
+    CPPPermute     _permute_weights;
+    CPPPermute     _permute_output;
+    Tensor         _input_workspace;
+    Tensor         _output_workspace;
+    Tensor         _kernel_storage;
+    Tensor         _input_nhwc;
+    Tensor         _output_nhwc;
+    Tensor         _weights_hwio;
+    const ITensor *_input;
+    const ITensor *_weights;
+    ITensor       *_output;
+    bool           _reshaped_kernel;
 };
 }
 #endif /* __ARM_COMPUTE_NEWINOGRADLAYER_H__ */
diff --git a/src/core/NEON/kernels/NEWinogradLayerKernel.cpp b/src/core/NEON/kernels/NEWinogradLayerKernel.cpp
index e2e4e40fe4..b0a36ff46a 100644
--- a/src/core/NEON/kernels/NEWinogradLayerKernel.cpp
+++ b/src/core/NEON/kernels/NEWinogradLayerKernel.cpp
@@ -29,173 +29,193 @@
 #include "arm_compute/core/TensorInfo.h"
 #include "support/ToolchainSupport.h"
 
-#include "arm_compute/core/NEON/kernels/winograd/winograd_layer.hpp"
-
-namespace
-{
-using T = WinogradConvolutionLayer<2, 2, 3, 3, float, float>;
-} // namespace
-
 namespace arm_compute
 {
-class Winograd3x3F32::Private
-{
-public:
-    Private(
-        const int          n_batches,         /** Number of batches in the input and output tensors. */
-        const int          n_input_channels,  /** Number of feature maps in a batch of the input tensor. */
-        const int          n_input_rows,      /** Number of rows in a feature map of the input tensor. */
-        const int          n_input_cols,      /** Number of columns in a feature map of the input tensor. */
-        const int          n_output_channels, /** Number of feature maps in the output tensor. */
-        const bool         same_padding,      /** Use "SAME" padding, otherwise use "VALID". */
-        const float *const weights,           /** Pointer to weight tensor in spatial domain. Must be ordered as "Height x Rows x Input Feature Maps x Output Feature Maps. */
-        float *const       weights_storage,   /** Pointer to storage for weight tensor in the Winograd domain. Must be at least the size returned by `get_weight_storage_size`. */
-        const float *const input,             /** Pointer to NHWC ordered input tensor, in the spatial domain. */
-        float *const       winograd_input,    /** Pointer to working space for the input tensor in the Winograd domain. Must be at least the size returned by `get_input_storage_size`. */
-        float *const       output,            /** Pointer to NHWC ordered output tensor, in the spatial domain. */
-        float *const       winograd_output    /** Pointer to working space for the output tensor in the Winograd domain. Must be at least the size returned by `get_output_storage_size`. */
-    )
-        : convolver(n_batches, n_input_channels, n_input_rows, n_input_cols, n_output_channels, same_padding, weights, weights_storage, input, winograd_input, nullptr, output, winograd_output)
-    {
-    }
-    T convolver;
-};
-
-Winograd3x3F32::~Winograd3x3F32()
-{
-}
-
-Winograd3x3F32::Winograd3x3F32(
-    const int          n_batches,         /** Number of batches in the input and output tensors. */
-    const int          n_input_channels,  /** Number of feature maps in a batch of the input tensor. */
-    const int          n_input_rows,      /** Number of rows in a feature map of the input tensor. */
-    const int          n_input_cols,      /** Number of columns in a feature map of the input tensor. */
-    const int          n_output_channels, /** Number of feature maps in the output tensor. */
-    const bool         same_padding,      /** Use "SAME" padding, otherwise use "VALID". */
-    const float *const weights,           /** Pointer to weight tensor in spatial domain. Must be ordered as "Height x Rows x Input Feature Maps x Output Feature Maps. */
-    float *const       weights_storage,   /** Pointer to storage for weight tensor in the Winograd domain. Must be at least the size returned by `get_weight_storage_size`. */
-    const float *const input,             /** Pointer to NHWC ordered input tensor, in the spatial domain. */
-    float *const       winograd_input,    /** Pointer to working space for the input tensor in the Winograd domain. Must be at least the size returned by `get_input_storage_size`. */
-    float *const       output,            /** Pointer to NHWC ordered output tensor, in the spatial domain. */
-    float *const       winograd_output    /** Pointer to working space for the output tensor in the Winograd domain. Must be at least the size returned by `get_output_storage_size`. */
-)
-    : _pimpl(support::cpp14::make_unique<Private>(n_batches, n_input_channels, n_input_rows, n_input_cols, n_output_channels, same_padding, weights, weights_storage, input, winograd_input, output,
-                                                  winograd_output))
-{
-}
-
-unsigned int NEWinogradLayerKernel::get_input_storage_size(const int n_batches, const int n_channels, const int n_rows, const int n_cols, const bool same_padding)
-{
-    return T::get_input_storage_size(n_batches, n_channels, n_rows, n_cols, same_padding);
-}
-
-unsigned int NEWinogradLayerKernel::get_output_storage_size(
-    const int  n_batches,         /** Number of batches in the output tensor. */
-    const int  n_rows,            /** Number of rows in each feature map of the input tensor. */
-    const int  n_cols,            /** Number of columns in each feature map of the input tensor. */
-    const int  n_output_channels, /** Number of feature maps in the output tensor. */
-    const bool same_padding       /** Use "SAME" padding, otherwise use "VALID". */
-)
+//Batched Gemms
+template <int OutputTileRows, int OutputTileCols, int KernelRows, int KernelCols>
+NEWinogradLayerKernel<OutputTileRows, OutputTileCols, KernelRows, KernelCols>::NEWinogradLayerKernel()
+    : _gemms()
 {
-    return T::get_output_storage_size(n_batches, n_rows, n_cols, n_output_channels, same_padding);
 }
 
-unsigned int NEWinogradLayerKernel::get_weight_storage_size(const int n_output_channels, const int n_input_channels)
+template <int OutputTileRows, int OutputTileCols, int KernelRows, int KernelCols>
+void NEWinogradLayerKernel<OutputTileRows, OutputTileCols, KernelRows, KernelCols>::configure(
+    const unsigned int n_gemms,
+    const int M, const int K, const int N,
+    const int          a_matrix_stride,
+    const int          a_row_stride,
+    const int          b_matrix_stride,
+    const int          b_row_stride,
+    const int          c_matrix_stride,
+    const int          c_row_stride,
+    const float *const a_ptr,
+    const float *const b_ptr,
+    float *const       c_ptr)
 {
-    return T::get_weight_storage_size(n_output_channels, n_input_channels);
-}
-
-NEWinogradLayerKernel::NEWinogradLayerKernel()
-    : _convolver(nullptr)
-{
-}
-
-void NEWinogradLayerKernel::configure(Winograd3x3F32 *convolver)
-{
-    ARM_COMPUTE_ERROR_ON_NULLPTR(convolver);
-    _convolver = convolver;
+    _gemms = support::cpp14::make_unique<MultiGEMM>(n_gemms, M, K, N, a_matrix_stride, a_row_stride, b_matrix_stride, b_row_stride, c_matrix_stride, c_row_stride, a_ptr, b_ptr, c_ptr);
     Window win;
-    auto   win_last = _convolver->_pimpl->convolver.gemms.get_window();
+    auto   win_last = _gemms->get_window();
     win.set(Window::DimX, Window::Dimension(0, win_last, 1));
     INEKernel::configure(win);
 }
 
-void NEWinogradLayerKernel::run(const Window &window, const ThreadInfo &info)
+template <int OutputTileRows, int OutputTileCols, int KernelRows, int KernelCols>
+void NEWinogradLayerKernel<OutputTileRows, OutputTileCols, KernelRows, KernelCols>::run(const Window &window, const ThreadInfo &info)
 {
     ARM_COMPUTE_UNUSED(info);
     ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
     const size_t first_gemm = window.x().start();
     const size_t last_gemm  = window.x().end();
-    _convolver->_pimpl->convolver.gemms.run(first_gemm, last_gemm);
+    _gemms->run(first_gemm, last_gemm);
 }
 
-INEWinogradLayerTransformKernel::INEWinogradLayerTransformKernel()
-    : _convolver(nullptr)
+template class NEWinogradLayerKernel<2, 2, 3, 3>;
+
+// Weights transform
+
+template <int OutputTileRows, int OutputTileCols, int KernelRows, int KernelCols>
+unsigned int NEWinogradLayerTransformWeightsKernel<OutputTileRows, OutputTileCols, KernelRows, KernelCols>::get_weight_storage_size(int n_output_channels, int n_input_channels)
 {
+    const KernelShape shape(n_output_channels, KernelRows, KernelCols, n_input_channels);
+    return static_cast<unsigned int>(
+               // WinogradConv returns the size in bytes, we divide by `sizeof(float)` to
+               // express that in units of float.
+               WinogradConv::get_kernel_storage_size(shape) / sizeof(float));
 }
 
-void INEWinogradLayerTransformKernel::configure(Winograd3x3F32 *convolver)
+template <int OutputTileRows, int OutputTileCols, int KernelRows, int KernelCols>
+NEWinogradLayerTransformWeightsKernel<OutputTileRows, OutputTileCols, KernelRows, KernelCols>::NEWinogradLayerTransformWeightsKernel()
+    : _transform()
 {
-    ARM_COMPUTE_ERROR_ON_NULLPTR(convolver);
-    _convolver = convolver;
 }
 
-// Weights transform
-
-void NEWinogradLayerTransformWeightsKernel::configure(Winograd3x3F32 *convolver)
+template <int OutputTileRows, int OutputTileCols, int KernelRows, int KernelCols>
+void NEWinogradLayerTransformWeightsKernel<OutputTileRows, OutputTileCols, KernelRows, KernelCols>::configure(
+    const ITensor *weights_hwio,
+    float *const   output,
+    const int      matrix_stride,     /** Stride across matrices in the output. */
+    const int      n_output_channels, /** Number of filters. */
+    const int      n_input_channels)  /** Number of channels in each filter. */
 {
-    INEWinogradLayerTransformKernel::configure(convolver);
+    const int matrix_row_stride = roundup(n_output_channels, WinogradConv::N_BLOCK);
+    _transform                  = support::cpp14::make_unique<WeightsTransform>(reinterpret_cast<float *>(weights_hwio->buffer()), output, matrix_stride, matrix_row_stride, n_output_channels,
+                                                                                n_input_channels);
     Window win;
-    auto   win_last = _convolver->_pimpl->convolver.weights_transform.get_window();
+    auto   win_last = _transform->get_window();
     win.set(Window::DimX, Window::Dimension(0, win_last, 1));
     INEKernel::configure(win);
 }
 
-void NEWinogradLayerTransformWeightsKernel::run(const Window &window, const ThreadInfo &info)
+template <int OutputTileRows, int OutputTileCols, int KernelRows, int KernelCols>
+void NEWinogradLayerTransformWeightsKernel<OutputTileRows, OutputTileCols, KernelRows, KernelCols>::run(const Window &window, const ThreadInfo &info)
 {
     ARM_COMPUTE_UNUSED(info);
     ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
     const size_t fst = window.x().start();
     const size_t lst = window.x().end();
-    _convolver->_pimpl->convolver.weights_transform.run(fst, lst);
+    _transform->run(fst, lst);
 }
 
-bool NEWinogradLayerTransformWeightsKernel::is_parallelisable() const
+template <int OutputTileRows, int OutputTileCols, int KernelRows, int KernelCols>
+bool NEWinogradLayerTransformWeightsKernel<OutputTileRows, OutputTileCols, KernelRows, KernelCols>::is_parallelisable() const
 {
     return false;
 }
 
+template class NEWinogradLayerTransformWeightsKernel<2, 2, 3, 3>;
+
 // Input transform
 
-void NEWinogradLayerTransformInputKernel::configure(Winograd3x3F32 *convolver)
+template <int OutputTileRows, int OutputTileCols, int KernelRows, int KernelCols>
+unsigned int NEWinogradLayerTransformInputKernel<OutputTileRows, OutputTileCols, KernelRows, KernelCols>::get_input_storage_size(
+    int  n_batches,   /** Number of batches in the input tensor. */
+    int  n_channels,  /** Number of feature maps in the input tensor. */
+    int  n_rows,      /** Number of rows in each feature map. */
+    int  n_cols,      /** Number of columns in each feature map. */
+    bool same_padding /** Use "SAME" padding, otherwise use "VALID". */
+)
+{
+    // Construct shapes for the input and kernel tensors.
+    const Tensor4DShape input_shape(n_batches, n_rows, n_cols, n_channels);
+    const KernelShape   kern_shape(1, KernelRows, KernelCols, n_channels);
+    const PaddingType   padding = (same_padding) ? PADDING_SAME : PADDING_VALID;
+    // Return the size, converted into units of TIn
+    return static_cast<unsigned int>(
+               WinogradConv::get_input_storage_size(kern_shape, input_shape, padding) / sizeof(float));
+}
+
+template <int OutputTileRows, int OutputTileCols, int KernelRows, int KernelCols>
+NEWinogradLayerTransformInputKernel<OutputTileRows, OutputTileCols, KernelRows, KernelCols>::NEWinogradLayerTransformInputKernel()
+    : _transform()
 {
-    INEWinogradLayerTransformKernel::configure(convolver);
+}
+
+template <int OutputTileRows, int OutputTileCols, int KernelRows, int KernelCols>
+void NEWinogradLayerTransformInputKernel<OutputTileRows, OutputTileCols, KernelRows, KernelCols>::configure(
+    const float *const input,         /** Input tensor data */
+    const int          n_batches,     /** Number of batches in input tensor. */
+    const int          n_rows,        /** Number of rows in input tensor. */
+    const int          n_cols,        /** Number of columns in input tensor. */
+    const int          n_channels,    /** Number of channels in input tensor. */
+    const PaddingType  padding,       /** Padding type. */
+    float *const       output,        /** Base of output matrices. */
+    const int          matrix_stride) /** Stride between output matrices. */
+{
+    //  _input_matrix_row_stride(n_input_channels),
+    _transform = support::cpp14::make_unique<InputTransform>(input, n_batches, n_rows, n_cols, n_channels, padding, output, matrix_stride, n_channels);
     Window win;
-    auto   win_last = _convolver->_pimpl->convolver.input_transform.get_window();
+    auto   win_last = _transform->get_window();
     win.set(Window::DimX, Window::Dimension(0, win_last, 1));
     INEKernel::configure(win);
 }
 
-void NEWinogradLayerTransformInputKernel::run(const Window &window, const ThreadInfo &info)
+template <int OutputTileRows, int OutputTileCols, int KernelRows, int KernelCols>
+void NEWinogradLayerTransformInputKernel<OutputTileRows, OutputTileCols, KernelRows, KernelCols>::run(const Window &window, const ThreadInfo &info)
 {
     ARM_COMPUTE_UNUSED(info);
     ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
     const size_t fst = window.x().start();
     const size_t lst = window.x().end();
-    _convolver->_pimpl->convolver.input_transform.run(fst, lst);
+    _transform->run(fst, lst);
 }
-bool NEWinogradLayerTransformInputKernel::is_parallelisable() const
+
+template <int OutputTileRows, int OutputTileCols, int KernelRows, int KernelCols>
+bool NEWinogradLayerTransformInputKernel<OutputTileRows, OutputTileCols, KernelRows, KernelCols>::is_parallelisable() const
 {
     return false;
 }
 
+template class NEWinogradLayerTransformInputKernel<2, 2, 3, 3>;
+
 // Output transform
-NEWinogradLayerTransformOutputKernel::NEWinogradLayerTransformOutputKernel()
+
+template <int OutputTileRows, int OutputTileCols, int KernelRows, int KernelCols>
+unsigned int NEWinogradLayerTransformOutputKernel<OutputTileRows, OutputTileCols, KernelRows, KernelCols>::get_output_storage_size(
+    int  n_batches,         /** Number of batches in the output tensor. */
+    int  n_rows,            /** Number of rows in each feature map of the input tensor. */
+    int  n_cols,            /** Number of columns in each feature map of the input tensor. */
+    int  n_output_channels, /** Number of feature maps in the output tensor. */
+    bool same_padding       /** Use "SAME" padding, otherwise use "VALID". */
+)
+{
+    // Construct shapes for the input and kernel tensors.
+    const Tensor4DShape input_shape(n_batches, n_rows, n_cols, 1);
+    const KernelShape   kern_shape(n_output_channels, KernelRows, KernelCols, 1);
+    const PaddingType   padding = (same_padding) ? PADDING_SAME : PADDING_VALID;
+
+    // Return the size, converted into units of TOut
+    return static_cast<unsigned int>(
+               WinogradConv::get_output_storage_size(kern_shape, input_shape, padding) / sizeof(float));
+}
+
+template <int OutputTileRows, int OutputTileCols, int KernelRows, int KernelCols>
+NEWinogradLayerTransformOutputKernel<OutputTileRows, OutputTileCols, KernelRows, KernelCols>::NEWinogradLayerTransformOutputKernel()
     : _biases(nullptr), _output_workspace(nullptr), _matrix_stride(0), _matrix_row_stride(0), _output(nullptr), _n_batches(0), _n_rows(0), _n_cols(0), _n_channels(0)
 {
 }
 
-void NEWinogradLayerTransformOutputKernel::configure(
+template <int OutputTileRows, int OutputTileCols, int KernelRows, int KernelCols>
+void NEWinogradLayerTransformOutputKernel<OutputTileRows, OutputTileCols, KernelRows, KernelCols>::configure(
     const ITensor     *biases,
     const float *const output_workingspace,
     const int          matrix_stride,
@@ -205,13 +225,10 @@ void NEWinogradLayerTransformOutputKernel::configure(
     const int          n_cols,
     const int          n_channels)
 {
-    using WinogradBase    = winograd::WinogradGEMM<2, 2, 3, 3>;
-    using OutputTransform = typename WinogradBase::template OutputTransform<float>;
-
     _biases            = biases;
     _output_workspace  = output_workingspace;
     _matrix_stride     = matrix_stride;
-    _matrix_row_stride = roundup(n_channels, WinogradBase::Convolution<float, float>::N_BLOCK);
+    _matrix_row_stride = roundup(n_channels, WinogradConv::N_BLOCK);
     _output            = output;
     _n_batches         = n_batches;
     _n_rows            = n_rows;
@@ -226,7 +243,8 @@ void NEWinogradLayerTransformOutputKernel::configure(
     INEKernel::configure(win);
 }
 
-void NEWinogradLayerTransformOutputKernel::run(const Window &window, const ThreadInfo &info)
+template <int OutputTileRows, int OutputTileCols, int KernelRows, int KernelCols>
+void NEWinogradLayerTransformOutputKernel<OutputTileRows, OutputTileCols, KernelRows, KernelCols>::run(const Window &window, const ThreadInfo &info)
 {
     ARM_COMPUTE_UNUSED(info);
     ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
@@ -234,9 +252,6 @@ void NEWinogradLayerTransformOutputKernel::run(const Window &window, const Threa
     ARM_COMPUTE_ERROR_ON_NULLPTR(_output_workspace);
     ARM_COMPUTE_ERROR_ON_NULLPTR(_output);
 
-    using WinogradBase    = winograd::WinogradGEMM<2, 2, 3, 3>;
-    using OutputTransform = typename WinogradBase::template OutputTransform<float>;
-
     OutputTransform output_transform(_output_workspace, _matrix_stride, _matrix_row_stride,
                                      reinterpret_cast<float *>(_biases->buffer()), _output,
                                      _n_batches, _n_rows, _n_cols, _n_channels);
@@ -247,9 +262,12 @@ void NEWinogradLayerTransformOutputKernel::run(const Window &window, const Threa
     output_transform.run(fst, lst);
 }
 
-bool NEWinogradLayerTransformOutputKernel::is_parallelisable() const
+template <int OutputTileRows, int OutputTileCols, int KernelRows, int KernelCols>
+bool NEWinogradLayerTransformOutputKernel<OutputTileRows, OutputTileCols, KernelRows, KernelCols>::is_parallelisable() const
 {
     return false;
 }
 
+template class NEWinogradLayerTransformOutputKernel<2, 2, 3, 3>;
+
 } // namespace arm_compute
diff --git a/src/core/NEON/kernels/winograd/winograd_gemm.cpp b/src/core/NEON/kernels/winograd/winograd_gemm.cpp
index b45f6f55d9..05426450a6 100644
--- a/src/core/NEON/kernels/winograd/winograd_gemm.cpp
+++ b/src/core/NEON/kernels/winograd/winograd_gemm.cpp
@@ -36,8 +36,8 @@ Tensor4DShape WinogradGEMM<kr, kc, itr, itc>::Convolution<TOut, TIn>::get_output
 {
   return Tensor4DShape {
     in_shape.n_batches,
-    (padding == PADDING_SAME) ? in_shape.n_rows : in_shape.n_rows - (kernel_rows - 2),
-    (padding == PADDING_SAME) ? in_shape.n_cols : in_shape.n_cols - (kernel_cols - 2),
+  (padding == PADDING_SAME) ? in_shape.n_rows : in_shape.n_rows - (kernel_rows - 1),
+  (padding == PADDING_SAME) ? in_shape.n_cols : in_shape.n_cols - (kernel_cols - 1),
     kernel_shape.n_output_channels,
     in_shape.ordering
   };
diff --git a/src/core/NEON/kernels/winograd/winograd_layer.cpp b/src/core/NEON/kernels/winograd/winograd_layer.cpp
deleted file mode 100644
index f16d62c0ef..0000000000
--- a/src/core/NEON/kernels/winograd/winograd_layer.cpp
+++ /dev/null
@@ -1,206 +0,0 @@
-/*
- * 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.
- */
-
-#include "convolution.hpp"
-#include "winograd_layer.hpp"
-#include "tensor.hpp"
-
-
-/** Determine how much memory (in units of TIn) to allocate for the transformed
- * weights.
- */
-template <
-  int OutputTileRows, int OutputTileCols, int KernelRows, int KernelCols,
-  typename TIn, typename TOut
->
-unsigned int WinogradConvolutionLayer<
-  OutputTileRows, OutputTileCols, KernelRows, KernelCols, TIn, TOut
->::get_weight_storage_size(
-  const int n_output_channels,  /** Number of output feature maps. */
-  const int n_input_channels    /** Number of input feature maps. */
-)
-{
-  const KernelShape shape(
-    n_output_channels, KernelRows, KernelCols, n_input_channels
-  );
-  return static_cast<unsigned int>(
-    // WinogradConv returns the size in bytes, we divide by `sizeof(TIn)` to
-    // express that in units of TIn.
-    WinogradConv::get_kernel_storage_size(shape) / sizeof(TIn)
-  );
-}
-
-
-/** Determine how much memory (in units of TIn) to allocate for the transformed
- * input.
- */
-template <
-  int OutputTileRows, int OutputTileCols, int KernelRows, int KernelCols,
-  typename TIn, typename TOut
->
-unsigned int WinogradConvolutionLayer<
-  OutputTileRows, OutputTileCols, KernelRows, KernelCols, TIn, TOut
->::get_input_storage_size(
-  const int n_batches,     /** Number of batches in the input tensor. */
-  const int n_channels,    /** Number of feature maps in the input tensor. */
-  const int n_rows,        /** Number of rows in each feature map. */
-  const int n_cols,        /** Number of columns in each feature map. */
-  const bool same_padding  /** Use "SAME" padding, otherwise use "VALID". */
-)
-{
-  // Construct shapes for the input and kernel tensors.
-  const Tensor4DShape input_shape(n_batches, n_rows, n_cols, n_channels);
-  const KernelShape kern_shape(1, KernelRows, KernelCols, n_channels);
-  const PaddingType padding = (same_padding) ? PADDING_SAME : PADDING_VALID;
-
-  // Return the size, converted into units of TIn
-  return static_cast<unsigned int>(
-    WinogradConv::get_input_storage_size(kern_shape, input_shape, padding) /
-    sizeof(TIn)
-  );
-}
-
-
-/** Determine how much memory (in units of TOut) to allocate for the (Winograd
- * domain) output.
- */
-template <
-  int OutputTileRows, int OutputTileCols, int KernelRows, int KernelCols,
-  typename TIn, typename TOut
->
-unsigned int WinogradConvolutionLayer<
-  OutputTileRows, OutputTileCols, KernelRows, KernelCols, TIn, TOut
->::get_output_storage_size(
-  const int n_batches,          /** Number of batches in the output tensor. */
-  const int n_rows,             /** Number of rows in each feature map of the input tensor. */
-  const int n_cols,             /** Number of columns in each feature map of the input tensor. */
-  const int n_output_channels,  /** Number of feature maps in the output tensor. */
-  const bool same_padding       /** Use "SAME" padding, otherwise use "VALID". */
-)
-{
-  // Construct shapes for the input and kernel tensors.
-  const Tensor4DShape input_shape(n_batches, n_rows, n_cols, 1);
-  const KernelShape kern_shape(n_output_channels, KernelRows, KernelCols, 1);
-  const PaddingType padding = (same_padding) ? PADDING_SAME : PADDING_VALID;
-
-  // Return the size, converted into units of TOut
-  return static_cast<unsigned int>(
-    WinogradConv::get_output_storage_size(kern_shape, input_shape, padding) /
-    sizeof(TOut)
-  );
-}
-
-
-/** Get the shape (rows, cols) of a feature map of the output tensor. */
-template <
-  int OutputTileRows, int OutputTileCols, int KernelRows, int KernelCols,
-  typename TIn, typename TOut
->
-std::pair<int, int> WinogradConvolutionLayer<
-  OutputTileRows, OutputTileCols, KernelRows, KernelCols, TIn, TOut
->::get_output_feature_map_shape(
-  const int n_input_rows,  /** Number of rows in the input feature map. */
-  const int n_input_cols,  /** Number of columns in the input feature map. */
-  const bool same_padding  /** Use "SAME" padding, otherwise use "VALID". */
-)
-{
-  // Construct shapes for the input and kernel tensors.
-  const Tensor4DShape input_shape(1, n_input_rows, n_input_cols, 1);
-  const KernelShape kern_shape(1, KernelRows, KernelCols, 1);
-  const PaddingType padding = (same_padding) ? PADDING_SAME : PADDING_VALID;
-
-  // Compute the new shape
-  const auto output_shape = WinogradConv::get_output_shape(
-    kern_shape, input_shape, padding
-  );
-
-  return std::make_pair(output_shape.n_rows, output_shape.n_cols);
-}
-
-
-/** Create a new Winograd convolution layer.
- */
-template <
-  int OutputTileRows, int OutputTileCols, int KernelRows, int KernelCols,
-  typename TIn, typename TOut
->
-WinogradConvolutionLayer<OutputTileRows, OutputTileCols, KernelRows, KernelCols, TIn, TOut>::
-WinogradConvolutionLayer(
-  const int n_batches,          /** Number of batches in the input and output tensors. */
-  const int n_input_channels,   /** Number of feature maps in a batch of the input tensor. */
-  const int n_input_rows,       /** Number of rows in a feature map of the input tensor. */
-  const int n_input_cols,       /** Number of columns in a feature map of the input tensor. */
-  const int n_output_channels,  /** Number of feature maps in the output tensor. */
-  const bool same_padding,      /** Use "SAME" padding, otherwise use "VALID". */
-  const TIn* const weights,     /** Pointer to weight tensor in spatial domain. Must be ordered as "Height x Rows x Input Feature Maps x Output Feature Maps. */
-  TIn* const winograd_weights,  /** Pointer to storage for weight tensor in the Winograd domain. Must be at least the size returned by `get_weight_storage_size`. */
-  const TIn* const input,       /** Pointer to NHWC ordered input tensor, in the spatial domain. */
-  TIn* const winograd_input,    /** Pointer to working space for the input tensor in the Winograd domain. Must be at least the size returned by `get_input_storage_size`. */
-  const TOut* const biases,     /** Pointer to biases vector. */
-  TOut* const output,           /** Pointer to NHWC ordered output tensor, in the spatial domain. */
-  TOut* const winograd_output   /** Pointer to working space for the output tensor in the Winograd domain. Must be at least the size returned by `get_output_storage_size`. */
-) : _kernel_shape(n_output_channels, KernelRows, KernelCols, n_input_channels),
-    _input_shape(n_batches, n_input_rows, n_input_cols, n_input_channels),
-    _padding(same_padding ? PADDING_SAME : PADDING_VALID),
-    _output_shape(WinogradConv::get_output_shape(_kernel_shape, _input_shape, _padding)),
-    _n_output_rows(_output_shape.n_rows),
-    _n_output_cols(_output_shape.n_cols),
-    _kernel_matrix_stride(WinogradConv::get_kernel_matrix_stride(_kernel_shape)),
-    _kernel_matrix_row_stride(roundup(n_output_channels, WinogradConv::N_BLOCK)),
-    _input_matrix_stride(WinogradConv::get_input_matrix_stride(_kernel_shape, _input_shape, _padding)),
-    _input_matrix_row_stride(n_input_channels),
-    _output_matrix_stride(WinogradConv::get_output_matrix_stride(_kernel_shape, _input_shape, _padding)),
-    _output_matrix_row_stride(_kernel_matrix_row_stride),
-    _tile_rows(iceildiv(_n_output_rows, OutputTileRows)),
-    _tile_cols(iceildiv(_n_output_cols, OutputTileCols)),
-    _m(n_batches * _tile_rows * _tile_cols),
-    _k(n_input_channels),
-    _n(n_output_channels),
-    weights_transform(
-      weights, winograd_weights,
-      _kernel_matrix_stride, _kernel_matrix_row_stride,
-      n_output_channels, n_input_channels
-    ),
-    input_transform(
-      input, n_batches, n_input_rows, n_input_cols, n_input_channels, _padding,
-      winograd_input, _input_matrix_stride, _input_matrix_row_stride
-    ),
-    gemms(
-      WinogradBase::N_GEMMS, _m, _k, _n,
-      _input_matrix_stride, _input_matrix_row_stride,
-      _kernel_matrix_stride, _kernel_matrix_row_stride,
-      _output_matrix_stride, _output_matrix_row_stride,
-      winograd_input, winograd_weights, winograd_output
-    ),
-    output_transform(
-      winograd_output, _output_matrix_stride, _output_matrix_row_stride, biases,
-      output, n_batches, _n_output_rows, _n_output_cols, n_output_channels
-    )
-{
-}
-
-// Instantiate valid implementations.
-template class WinogradConvolutionLayer<2, 2, 3, 3, float, float>;
-template class WinogradConvolutionLayer<4, 4, 3, 3, float, float>;
-template class WinogradConvolutionLayer<2, 2, 5, 5, float, float>;
diff --git a/src/runtime/NEON/functions/NEWinogradLayer.cpp b/src/runtime/NEON/functions/NEWinogradLayer.cpp
index e8c77412a2..6196c514a8 100644
--- a/src/runtime/NEON/functions/NEWinogradLayer.cpp
+++ b/src/runtime/NEON/functions/NEWinogradLayer.cpp
@@ -46,7 +46,7 @@ namespace arm_compute
 {
 NEWinogradLayer::NEWinogradLayer(std::shared_ptr<IMemoryManager> memory_manager)
     : _memory_group(std::move(memory_manager)), _winograd_kernel(), _transform_input_kernel(), _transform_output_kernel(), _transform_weights_kernel(), _permute_input(), _permute_weights(),
-      _permute_output(), _input_workspace(), _output_workspace(), _kernel_storage(), _input_nhwc(), _output_nhwc(), _weights_hwio(), _input(), _weights(), _output(), _reshaped_kernel(false), _conv()
+      _permute_output(), _input_workspace(), _output_workspace(), _kernel_storage(), _input_nhwc(), _output_nhwc(), _weights_hwio(), _input(), _weights(), _output(), _reshaped_kernel(false)
 {
 } /* arm_compute */
 
@@ -81,19 +81,23 @@ void NEWinogradLayer::configure(const ITensor *input, const ITensor *weights, co
     const size_t        data_type_size = input->info()->element_size();
     // Get the memory required to instantiate a new Winograd operator.
     constexpr size_t storage_alignment   = 64;
-    const size_t     kernel_storage_size = NEWinogradLayerKernel::get_weight_storage_size(out_channels, in_channels) * data_type_size;
+    const size_t     kernel_storage_size = NEWinogradLayerTransformWeightsKernel<2, 2, 3, 3>::get_weight_storage_size(out_channels, in_channels) * data_type_size;
     _kernel_storage.allocator()->init(TensorInfo(TensorShape{ (kernel_storage_size + storage_alignment - 1) }, 1, DataType::U8));
     _memory_group.manage(&_kernel_storage);
     _memory_group.manage(&_input_nhwc);
     _kernel_storage.allocator()->allocate();
     // Input storage
-    const size_t input_storage_size = NEWinogradLayerKernel::get_input_storage_size(in_shape.n_batches, in_shape.n_channels, in_shape.n_rows, in_shape.n_cols, false) * data_type_size;
+    const size_t input_storage_size = NEWinogradLayerTransformInputKernel<2, 2, 3, 3>::get_input_storage_size(in_shape.n_batches, in_shape.n_channels, in_shape.n_rows, in_shape.n_cols,
+                                                                                                              false)
+                                      * data_type_size;
     _input_workspace.allocator()->init(TensorInfo(TensorShape{ (input_storage_size + storage_alignment - 1) }, 1, DataType::U8));
     _memory_group.manage(&_input_workspace);
     _input_workspace.allocator()->allocate();
 
     // Output storage
-    const size_t output_storage_size = NEWinogradLayerKernel::get_output_storage_size(in_shape.n_batches, in_shape.n_rows, in_shape.n_cols, out_channels, false) * data_type_size;
+    const size_t output_storage_size = NEWinogradLayerTransformOutputKernel<2, 2, 3, 3>::get_output_storage_size(in_shape.n_batches, in_shape.n_rows, in_shape.n_cols, out_channels,
+                                                                                                                 false)
+                                       * data_type_size;
     _output_workspace.allocator()->init(TensorInfo(TensorShape{ (output_storage_size + storage_alignment - 1) }, 1, DataType::U8));
     _memory_group.manage(&_output_workspace);
     _output_workspace.allocator()->allocate();
@@ -132,38 +136,46 @@ void NEWinogradLayer::configure(const ITensor *input, const ITensor *weights, co
     _permute_input.configure(input, &_input_nhwc, PermutationVector(2U, 0U, 1U));
 
     _input_nhwc.allocator()->allocate();
-    // Create Winograd operator object
-    _conv = support::cpp14::make_unique<Winograd3x3F32>(
-                in_shape.n_batches,
-                in_shape.n_channels,
-                in_shape.n_rows,
-                in_shape.n_cols,
-                out_channels,
-                false,
-                reinterpret_cast<const float *>(_weights_hwio.buffer()),
-                reinterpret_cast<float *>(_kernel_storage.buffer()),
-                reinterpret_cast<float *>(_input_nhwc.buffer()),
-                reinterpret_cast<float *>(_input_workspace.buffer()),
-                reinterpret_cast<float *>(_output_nhwc.buffer()),
-                reinterpret_cast<float *>(_output_workspace.buffer()));
-
-    // Configure the kernel, padding not needed so it's safe to call configure after allocare
-    _winograd_kernel.configure(_conv.get());
-    _transform_input_kernel.configure(_conv.get());
-    _transform_weights_kernel.configure(_conv.get());
 
-    //The biases tensor has not been allocated at this point in time, the output transform will add the biases to the final result in the run() method
     using T                          = winograd::WinogradGEMM<2, 2, 3, 3>::Convolution<float, float>;
     const int         weights_width  = weights->info()->dimension(0);
     const int         weights_height = weights->info()->dimension(1);
     const KernelShape kernel_shape({ out_channels, weights_height, weights_width, in_channels });
-    const int         output_matrix_stride = T::get_output_matrix_stride(kernel_shape, in_shape, PADDING_VALID);
-    const auto        output_shape(T::get_output_shape(kernel_shape, in_shape, PADDING_VALID));
+
+    // Configure the InputTransform
+    const int input_matrix_stride = T::get_input_matrix_stride(kernel_shape, in_shape, PADDING_VALID);
+    _transform_input_kernel.configure(reinterpret_cast<float *>(_input_nhwc.buffer()), in_shape.n_batches, in_shape.n_rows, in_shape.n_cols, in_shape.n_channels, PADDING_VALID,
+                                      reinterpret_cast<float *>(_input_workspace.buffer()), input_matrix_stride);
+
+    // Configure WeightsTransform
+    const int kernel_matrix_stride = T::get_kernel_matrix_stride(kernel_shape);
+    _transform_weights_kernel.configure(&_weights_hwio, reinterpret_cast<float *>(_kernel_storage.buffer()), kernel_matrix_stride, out_channels, in_channels);
+
+    // Configure OutputTransform
+    //The biases tensor has not been allocated at this point in time, the output transform will add the biases to the final result in the run() method
+    const int  output_matrix_stride = T::get_output_matrix_stride(kernel_shape, in_shape, PADDING_VALID);
+    const auto output_shape(T::get_output_shape(kernel_shape, in_shape, PADDING_VALID));
 
     _transform_output_kernel.configure(biases, reinterpret_cast<float *>(_output_workspace.buffer()),
                                        output_matrix_stride, reinterpret_cast<float *>(_output_nhwc.buffer()),
                                        in_shape.n_batches, output_shape.n_rows, output_shape.n_cols, out_channels);
 
+    // Configure Batched GEMMs
+    const int tile_rows                = iceildiv(output_shape.n_rows, NEWinogradLayerKernel<2, 2, 3, 3>::_output_tile_rows);
+    const int tile_cols                = iceildiv(output_shape.n_cols, NEWinogradLayerKernel<2, 2, 3, 3>::_output_tile_cols);
+    const int m                        = in_shape.n_batches * tile_rows * tile_cols;
+    const int k                        = in_shape.n_channels;
+    const int n                        = out_channels;
+    const int input_matrix_row_stride  = in_shape.n_channels;
+    const int kernel_matrix_row_stride = roundup(out_channels, NEWinogradLayerKernel<2, 2, 3, 3>::WinogradConv::N_BLOCK);
+    const int output_matrix_row_stride = kernel_matrix_row_stride;
+
+    _winograd_kernel.configure(NEWinogradLayerKernel<2, 2, 3, 3>::WinogradBase::N_GEMMS, m, k, n,
+                               input_matrix_stride, input_matrix_row_stride,
+                               kernel_matrix_stride, kernel_matrix_row_stride,
+                               output_matrix_stride, output_matrix_row_stride,
+                               reinterpret_cast<float *>(_input_workspace.buffer()), reinterpret_cast<float *>(_kernel_storage.buffer()), reinterpret_cast<float *>(_output_workspace.buffer()));
+
     // Reorder the convoluted output to ACL's ordering NCHW
     _permute_output.configure(&_output_nhwc, _output, PermutationVector(1U, 2U, 0U));
 }