COMPMID-1550: Winograd integrate RSH changes.

Refactors the transforms to make use of partial specialization.

Change-Id: Idff68d22817a00a7ee9eef5351a5a9fd33147540
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/146635
Tested-by: Jenkins <bsgcomp@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>

8 files changed, 999 insertions, 1203 deletions

diff --git a/arm_compute/core/NEON/kernels/convolution/winograd/transforms/input.hpp b/arm_compute/core/NEON/kernels/convolution/winograd/transforms/input.hpp
index 369c2ff48f..473a13c3b0 100644
--- a/arm_compute/core/NEON/kernels/convolution/winograd/transforms/input.hpp
+++ b/arm_compute/core/NEON/kernels/convolution/winograd/transforms/input.hpp
@@ -29,10 +29,8 @@ namespace winograd
 {
   /***************************************************************************/
   /* Instance-less API */
-  template <int output_tile_rows, int output_tile_cols,
-            int kernel_rows, int kernel_cols>
-  template <typename T>
-  void WinogradGEMM<output_tile_rows, output_tile_cols, kernel_rows, kernel_cols>::InputTransform<T>::execute(
+  template <int KernelRows, int KernelCols, int InnerTileRows, int InnerTileCols, typename T>
+  void InputTransformImpl<KernelRows, KernelCols, InnerTileRows, InnerTileCols, T>::execute(
     const T* const input,        /** Input tensor data */
     const int n_batches,         /** Number of batches in input tensor. */
     const int in_batch_stride,   /** Stride between batches of the input. */
@@ -50,26 +48,9 @@ namespace winograd
     const int matrix_row_stride  /** Stride within matrices. */
   )
   {
-    // If an Nx1 kernel then transpose and redirect to the 1xN implementation
-    if (kernel_cols == 1)
-    {
-      WinogradGEMM<output_tile_cols, output_tile_rows, kernel_cols, kernel_rows>::
-        template InputTransform<T>::execute(
-          input,
-          n_batches, in_batch_stride,
-          n_cols, in_col_stride,
-          n_rows, in_row_stride,
-          n_channels, padding,
-          tile_N, tile_M,
-          output, matrix_stride, matrix_batch_stride, matrix_row_stride
-        );
-      return;
-    }
-
     // Compute the padding required on each edge of the image
-    const int pad_top = (padding == PADDING_SAME) ? (kernel_rows - 1) / 2 : 0;
-    const int pad_left = (padding == PADDING_SAME) ? (kernel_cols - 1) / 2 : 0;
-    const int tile_overlap = kernel_rows - 1;
+    const int pad_top = (padding == PADDING_SAME) ? (KernelRows - 1) / 2 : 0;
+    const int pad_left = (padding == PADDING_SAME) ? (KernelCols - 1) / 2 : 0;
 
     // Compute striding values (assuming NHWC ordered data)
     const int output_col_stride = matrix_row_stride;
@@ -85,19 +66,19 @@ namespace winograd
       // Loop over rows of tiles
       for (int tile_i = 0; tile_i < tile_M; tile_i++)
       {
+        // Padding (top + bottom) for the row
+        const int row_top = tile_i*(InnerTileRows - overlap_rows) - pad_top;
+        const int row_bottom = row_top + InnerTileRows;
+        const int row_pad_top = std::max(0, pad_top - tile_i*(InnerTileRows - overlap_rows));
+        const int row_pad_bottom = (row_bottom <= n_rows) ? 0 : row_bottom - n_rows;
+
         // Pointer to the row
-        const int row_offset = (tile_i == 0) ? 0 : pad_top;
+        const int row_offset = std::min(0, row_pad_top - pad_top);
         const T* const input_base_row = (
-          input_base_batch + ((inner_tile_rows - (kernel_rows - 1))*tile_i - row_offset)*in_row_stride
+          input_base_batch + ((InnerTileRows - overlap_rows)*tile_i + row_offset)*in_row_stride
         );
         T* const outptr_base_row = outptr_base_batch + tile_i*output_row_stride;
 
-        // Padding (top + bottom) for the row
-        const int row_top = tile_i*(inner_tile_rows - tile_overlap) - pad_top;
-        const int row_bottom = row_top + inner_tile_rows;
-        const int row_pad_top = (tile_i == 0) ? pad_top : 0;
-        const int row_pad_bottom = (row_bottom <= n_rows) ? 0 : row_bottom - n_rows;
-
         // Process the row
         process_tile_row(
           tile_N, n_channels,
@@ -109,10 +90,40 @@ namespace winograd
     }
   }
 
-  template <int output_tile_rows, int output_tile_cols,
-            int kernel_rows, int kernel_cols>
-  template <typename T>
-  void WinogradGEMM<output_tile_rows, output_tile_cols, kernel_rows, kernel_cols>::InputTransform<T>::process_tile_row(
+
+  template <int KernelRows, int InnerTileRows, typename T>
+  void InputTransformImpl<KernelRows, 1, InnerTileRows, 1, T>::execute(
+    const T* const input,        /** Input tensor data */
+    const int n_batches,         /** Number of batches in input tensor. */
+    const int in_batch_stride,   /** Stride between batches of the input. */
+    const int n_rows,            /** Number of rows in input tensor. */
+    const int in_row_stride,     /** Stride between rows of the input. */
+    const int n_cols,            /** Number of columns in input tensor. */
+    const int in_col_stride,     /** Stride between columns of the input. */
+    const int n_channels,        /** Number of channels in input tensor. */
+    const PaddingType padding,   /** Padding type. */
+    const int tile_M,
+    const int tile_N,
+    T* const output,             /** Base of output matrices. */
+    const int matrix_stride,     /** Stride between output matrices. */
+    const int matrix_batch_stride,  /** Stride between batches within the matrix. */
+    const int matrix_row_stride  /** Stride within matrices. */
+  )
+  {
+    // If an Nx1 kernel then transpose and redirect to the 1xN implementation
+    InputTransformImpl<1, KernelRows, 1, InnerTileRows, T>::execute(
+      input,
+      n_batches, in_batch_stride,
+      n_cols, in_col_stride,
+      n_rows, in_row_stride,
+      n_channels, padding,
+      tile_N, tile_M,
+      output, matrix_stride, matrix_batch_stride, matrix_row_stride
+    );
+  }
+
+  template <int KernelRows, int KernelCols, int InnerTileRows, int InnerTileCols, typename T>
+  void InputTransformImpl<KernelRows, KernelCols, InnerTileRows, InnerTileCols, T>::process_tile_row(
     const int tile_N,
     int n_channels,
     const T* const input_base,
@@ -127,33 +138,25 @@ namespace winograd
     const int n_cols
   )
   {
-    if (kernel_cols == 1)
-    {
-      // If an Nx1 implementation then this should never be reached.
-      return;
-    }
-
-    constexpr int tile_overlap = kernel_cols - 1;
-
     // Loop over columns of tiles
     for (int tile_j = 0; tile_j < tile_N; tile_j++)
     {
       // Padding (left + right) for the tile
-      const int t_pad_left = (tile_j == 0) ? row_pad_left : 0;
-      const int t_start = tile_j*(inner_tile_cols - tile_overlap) - row_pad_left;
-      const int t_end = t_start + inner_tile_cols;
+      const int t_start = tile_j*(InnerTileCols - overlap_cols) - row_pad_left;
+      const int t_end = t_start + InnerTileCols;
+      const int t_pad_left = std::max(0, row_pad_left - tile_j*(InnerTileCols - overlap_cols));
       const int t_pad_right = (t_end <= n_cols) ? 0 : t_end - n_cols;
 
       // Get pointers into the inputs and outputs
-      const int col_offset = (tile_j == 0) ? 0 : row_pad_left;
+      const int col_offset = std::min(0, t_pad_left - row_pad_left);
       const T* const input_base_col = (
-        input_base + ((inner_tile_cols - tile_overlap)*tile_j - col_offset)*input_col_stride
+        input_base + ((InnerTileCols - overlap_cols)*tile_j + col_offset)*input_col_stride
       );
       T* const outptr = matrix_base + tile_j*matrix_row_stride;
 
       // Apply the specific tile processing function
-      const int f_pad_top = pad_top ? 1 : 0;
-      const int f_pad_left = t_pad_left ? 1 : 0;
+      const int f_pad_top = iceildiv(pad_top, 2);
+      const int f_pad_left = iceildiv(t_pad_left, 2);
       tile_fns[f_pad_top][f_pad_left][pad_bottom][t_pad_right](
         n_channels,
         input_base_col,
@@ -166,9 +169,8 @@ namespace winograd
   }
 
   /***************************************************************************/
-  template <int otr, int otc, int kr, int kc>
-  template <typename T>
-  WinogradGEMM<otr, otc, kr, kc>::InputTransform<T>::InputTransform(
+  template <int KernelRows, int KernelCols, int InnerTileRows, int InnerTileCols, typename T>
+  InputTransform<KernelRows, KernelCols, InnerTileRows, InnerTileCols, T>::InputTransform(
     const T* 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. */
@@ -184,10 +186,10 @@ namespace winograd
   ) : _inptr(input), _outptr(output),
       _n_batches(n_batches), _n_rows(n_rows), _n_cols(n_cols), _n_channels(n_channels),
       _matrix_stride(matrix_stride), _matrix_row_stride(matrix_row_stride),
-      _tiles_M(iceildiv((padding == PADDING_SAME) ? n_rows : n_rows - kr + 1,
-                        output_tile_rows)),
-      _tiles_N(iceildiv((padding == PADDING_SAME) ? n_cols : n_cols - kc + 1,
-                        output_tile_cols)),
+      _tiles_M(iceildiv((padding == PADDING_SAME) ? n_rows : n_rows - KernelRows + 1,
+                        InnerTileRows - KernelRows + 1)),
+      _tiles_N(iceildiv((padding == PADDING_SAME) ? n_cols : n_cols - KernelCols + 1,
+                        InnerTileCols - KernelCols + 1)),
       _in_col_stride(in_col_stride ? in_col_stride : n_channels),
       _in_row_stride(in_row_stride ? in_row_stride : n_cols * _in_col_stride),
       _in_batch_stride(in_batch_stride ? in_batch_stride : n_rows * _in_row_stride),
@@ -195,18 +197,16 @@ namespace winograd
   {
   }
 
-  template <int otr, int otc, int kr, int kc>
-  template <typename T>
-  unsigned int WinogradGEMM<otr, otc, kr, kc>::InputTransform<T>::get_window() const
+  template <int KernelRows, int KernelCols, int InnerTileRows, int InnerTileCols, typename T>
+  unsigned int InputTransform<KernelRows, KernelCols, InnerTileRows, InnerTileCols, T>::get_window() const
   {
     // The final window includes the tail, all other windows will be a multiple
     // of the window block in size.
     return iceildiv(_n_channels, WINDOW_BLOCK);
   }
 
-  template <int otr, int otc, int kr, int kc>
-  template <typename T>
-  void WinogradGEMM<otr, otc, kr, kc>::InputTransform<T>::run(
+  template <int KernelRows, int KernelCols, int InnerTileRows, int InnerTileCols, typename T>
+  void InputTransform<KernelRows, KernelCols, InnerTileRows, InnerTileCols, T>::run(
     const unsigned int start, const unsigned int stop
   )
   {
@@ -238,4 +238,30 @@ namespace winograd
       _matrix_row_stride
     );
   }
+
+  template <int KernelRows, int KernelCols, int InnerTileRows, int InnerTileCols, typename T>
+  void InputTransform<KernelRows, KernelCols, InnerTileRows, InnerTileCols, T>::execute(
+    const T* const input,        /** Input tensor data */
+    const int n_batches,         /** Number of batches in input tensor. */
+    const int in_batch_stride,   /** Stride between batches of the input. */
+    const int n_rows,            /** Number of rows in input tensor. */
+    const int in_row_stride,     /** Stride between rows of the input. */
+    const int n_cols,            /** Number of columns in input tensor. */
+    const int in_col_stride,     /** Stride between columns of the input. */
+    const int n_channels,        /** Number of channels in input tensor. */
+    const PaddingType padding,   /** Padding type. */
+    const int tile_M,
+    const int tile_N,
+    T* const output,             /** Base of output matrices. */
+    const int matrix_stride,     /** Stride between output matrices. */
+    const int matrix_batch_stride,  /** Stride between batches within the matrix. */
+    const int matrix_row_stride  /** Stride within matrices. */
+  )
+  {
+    Transform::execute(
+      input, n_batches, in_batch_stride, n_rows, in_row_stride, n_cols,
+      in_col_stride, n_channels, padding, tile_M, tile_N, output,
+      matrix_stride, matrix_batch_stride, matrix_row_stride
+    );
+  }
 }
diff --git a/arm_compute/core/NEON/kernels/convolution/winograd/winograd_gemm.hpp b/arm_compute/core/NEON/kernels/convolution/winograd/winograd_gemm.hpp
index 7098fc48a1..31aee35fab 100644
--- a/arm_compute/core/NEON/kernels/convolution/winograd/winograd_gemm.hpp
+++ b/arm_compute/core/NEON/kernels/convolution/winograd/winograd_gemm.hpp
@@ -30,7 +30,7 @@
 #include "arm_compute/core/NEON/kernels/convolution/common/shims.hpp"
 #include "arm_compute/core/NEON/kernels/convolution/common/tensor.hpp"
 #include "arm_compute/core/NEON/kernels/convolution/common/utils.hpp"
-
+#include "winograd_input_transform.hpp"
 
 #include <thread>
 #include <utility>
@@ -114,121 +114,12 @@ class WinogradGEMM
     /** Transform input feature maps from the spatial to the Winograd domain.
      */
     template <typename T>
-    struct InputTransform
-    {
-      /** Get the bytes read during the transform. */
-      static size_t bytes_read(const Tensor4DShape &shape)
-      {
-        return shape.size() * sizeof(T);
-      }
-
-      /** Get the bytes written during the transform. */
-      static size_t bytes_written(const Tensor4DShape &shape)
-      {
-        const int M = iceildiv(shape.n_rows, inner_tile_rows) *
-                      iceildiv(shape.n_cols, inner_tile_cols);
-        const int K = shape.n_channels;
-        return inner_tile_rows * inner_tile_cols * M * K * sizeof(T);
-      }
-
-      /** Get the count of operations performed by the transform. */
-      static int ops_performed(const Tensor4DShape &shape);
-
-      /** Apply the transform to a tensor. */
-      static void execute(
-          const T* const input,        /** Input tensor data */
-          const int n_batches,         /** Number of batches in input tensor. */
-          const int in_batch_stride,   /** Stride between batches of the input. */
-          const int n_rows,            /** Number of rows in input tensor. */
-          const int in_row_stride,     /** Stride between rows of the input. */
-          const int n_cols,            /** Number of columns in input tensor. */
-          const int in_col_stride,     /** Stride between columns of the input. */
-          const int n_channels,        /** Number of channels in input tensor. */
-          const PaddingType padding,   /** Padding type. */
-          const int tile_M,
-          const int tile_N,
-          T* const output,             /** Base of output matrices. */
-          const int matrix_stride,     /** Stride between output matrices. */
-          const int matrix_batch_stride,  /** Stride between batches within the matrix. */
-          const int matrix_row_stride  /** Stride within matrices. */
-      );
-
-      /***********************************************************************/
-      /** Create an InputTransform operator fixed on a given problem and set of
-       * pointers.
-       */
-      InputTransform(
-          const T* 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. */
-          T* const output,             /** Base of output matrices. */
-          const int matrix_stride,     /** Stride between output matrices. */
-          const int matrix_row_stride, /** Stride within matrices. */
-          const int in_batch_stride=0, /** Stride between input batches. */
-          const int in_row_stride=0,   /** Stride between input rows. */
-          const int in_col_stride=0    /** Stride between input columns. */
-      );
-
-      /** Get the window of work a given operator can perform. */
-      unsigned int get_window() const;
-      static constexpr unsigned int WINDOW_BLOCK = 16;  // Base size of window
-
-      /** Perform work upon a window of the input. */
-      void run(const unsigned int start, const unsigned int stop);
-      /***********************************************************************/
-
-      private:
-        static void process_tile_row(
-          const int tile_N,
-          int n_channels,
-          const T* const input_base,
-          const int input_row_stride,
-          const int input_col_stride,
-          T* const matrix_base,
-          const int matrix_stride,
-          const int matrix_row_stride,
-          const int row_pad_top,
-          const int row_pad_left,
-          const int row_pad_bottom,
-          const int n_cols
-        );
-
-        // Tile overlaps
-        static constexpr int overlap_rows = kernel_rows - 1;
-        static constexpr int overlap_cols = kernel_cols - 1;
-
-        // Maximum padding and number of distinct paddings
-        static constexpr int max_pad_top = kernel_rows / 2;
-        static constexpr int n_pad_top = 1 + iceildiv(max_pad_top, inner_tile_rows - overlap_rows);
-
-        static constexpr int max_pad_left = kernel_cols / 2;
-        static constexpr int n_pad_left = 1 + iceildiv(max_pad_left, inner_tile_cols - overlap_cols);
-
-        static constexpr int n_pad_bottom = inner_tile_rows;
-        static constexpr int n_pad_right = inner_tile_cols;
-
-
-
-        /** Process a single tile of the input tensor. */
-        template <int pad_top, int pad_left, int pad_bottom, int pad_right>
-        static void process_tile(int, const T*, int, int, T*, int);
-
-        // Array of methods to transform tiles of the input tensor.
-        typedef void (*TileFn)(int, const T*, int, int, T*, int);
-        static const TileFn
-           tile_fns[n_pad_top][n_pad_left][n_pad_bottom][n_pad_right];
-
-        /* Member values for instance-based API. */
-        const T* const _inptr;
-        T* const _outptr;
-        const int _n_batches, _n_rows, _n_cols, _n_channels, _matrix_stride,
-                  _matrix_row_stride, _tiles_M, _tiles_N;
-        const int _in_col_stride, _in_row_stride, _in_batch_stride;
-        const PaddingType _padding_type;
-    };
+    using InputTransform = InputTransform<
+      KernelRows, KernelCols,
+      (OutputTileRows + KernelRows - 1),
+      (OutputTileCols + KernelCols - 1),
+      T
+    >;
 
     /** Transform output feature maps from the Winograd to the spatial domain.
      */
diff --git a/arm_compute/core/NEON/kernels/convolution/winograd/winograd_input_transform.hpp b/arm_compute/core/NEON/kernels/convolution/winograd/winograd_input_transform.hpp
new file mode 100644
index 0000000000..abcda53534
--- /dev/null
+++ b/arm_compute/core/NEON/kernels/convolution/winograd/winograd_input_transform.hpp
@@ -0,0 +1,184 @@
+/*
+ * 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.
+ */
+
+#pragma once
+
+namespace winograd
+{
+
+namespace
+{
+
+template <int KernelRows, int KernelCols, int InnerTileRows, int InnerTileCols, typename T>
+class InputTransformImpl
+{
+  public:
+    /** Apply the transform to a tensor. */
+    static void execute(
+        const T* const input,        /** Input tensor data */
+        const int n_batches,         /** Number of batches in input tensor. */
+        const int in_batch_stride,   /** Stride between batches of the input. */
+        const int n_rows,            /** Number of rows in input tensor. */
+        const int in_row_stride,     /** Stride between rows of the input. */
+        const int n_cols,            /** Number of columns in input tensor. */
+        const int in_col_stride,     /** Stride between columns of the input. */
+        const int n_channels,        /** Number of channels in input tensor. */
+        const PaddingType padding,   /** Padding type. */
+        const int tile_M,
+        const int tile_N,
+        T* const output,             /** Base of output matrices. */
+        const int matrix_stride,     /** Stride between output matrices. */
+        const int matrix_batch_stride,  /** Stride between batches within the matrix. */
+        const int matrix_row_stride  /** Stride within matrices. */
+    );
+
+  private:
+    static void process_tile_row(
+      const int tile_N,
+      int n_channels,
+      const T* const input_base,
+      const int input_row_stride,
+      const int input_col_stride,
+      T* const matrix_base,
+      const int matrix_stride,
+      const int matrix_row_stride,
+      const int row_pad_top,
+      const int row_pad_left,
+      const int row_pad_bottom,
+      const int n_cols
+    );
+
+    // Tile overlaps
+    static constexpr int overlap_rows = KernelRows - 1;
+    static constexpr int overlap_cols = KernelCols - 1;
+
+    // Maximum padding and number of distinct paddings
+    static constexpr int max_pad_top = KernelRows / 2;
+    static constexpr int n_pad_top = 1 + iceildiv(max_pad_top, InnerTileRows - overlap_rows);
+
+    static constexpr int max_pad_left = KernelCols / 2;
+    static constexpr int n_pad_left = 1 + iceildiv(max_pad_left, InnerTileCols - overlap_cols);
+
+    static constexpr int n_pad_bottom = InnerTileRows;
+    static constexpr int n_pad_right = InnerTileCols;
+
+    /** Process a single tile of the input tensor. */
+    template <int pad_top, int pad_left, int pad_bottom, int pad_right>
+    static void process_tile(int, const T*, int, int, T*, int);
+
+    // Array of methods to transform tiles of the input tensor.
+    typedef void (*TileFn)(int, const T*, int, int, T*, int);
+    static const TileFn
+      tile_fns[n_pad_top][n_pad_left][n_pad_bottom][n_pad_right];
+};
+
+
+template <int KernelRows, int InnerTileRows, typename T>
+class InputTransformImpl<KernelRows, 1, InnerTileRows, 1, T>
+{
+  public:
+    /** Apply the transform to a tensor. */
+    static void execute(
+        const T* const input,        /** Input tensor data */
+        const int n_batches,         /** Number of batches in input tensor. */
+        const int in_batch_stride,   /** Stride between batches of the input. */
+        const int n_rows,            /** Number of rows in input tensor. */
+        const int in_row_stride,     /** Stride between rows of the input. */
+        const int n_cols,            /** Number of columns in input tensor. */
+        const int in_col_stride,     /** Stride between columns of the input. */
+        const int n_channels,        /** Number of channels in input tensor. */
+        const PaddingType padding,   /** Padding type. */
+        const int tile_M,
+        const int tile_N,
+        T* const output,             /** Base of output matrices. */
+        const int matrix_stride,     /** Stride between output matrices. */
+        const int matrix_batch_stride,  /** Stride between batches within the matrix. */
+        const int matrix_row_stride  /** Stride within matrices. */
+    );
+};
+
+}  // namespace (anonymous)
+
+template <int KernelRows, int KernelCols, int InnerTileRows, int InnerTileCols, typename T>
+class InputTransform
+{
+  public:
+  /***********************************************************************/
+  /** Create an InputTransform operator fixed on a given problem and set of
+   * pointers.
+   */
+  InputTransform(
+      const T* 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. */
+      T* const output,             /** Base of output matrices. */
+      const int matrix_stride,     /** Stride between output matrices. */
+      const int matrix_row_stride, /** Stride within matrices. */
+      const int in_batch_stride=0, /** Stride between input batches. */
+      const int in_row_stride=0,   /** Stride between input rows. */
+      const int in_col_stride=0    /** Stride between input columns. */
+  );
+
+  /** Get the window of work a given operator can perform. */
+  unsigned int get_window() const;
+  static constexpr unsigned int WINDOW_BLOCK = 16;  // Base size of window
+
+  /** Perform work upon a window of the input. */
+  void run(const unsigned int start, const unsigned int stop);
+
+  /** Apply the transform to a tensor. */
+  static void execute(
+      const T* const input,        /** Input tensor data */
+      const int n_batches,         /** Number of batches in input tensor. */
+      const int in_batch_stride,   /** Stride between batches of the input. */
+      const int n_rows,            /** Number of rows in input tensor. */
+      const int in_row_stride,     /** Stride between rows of the input. */
+      const int n_cols,            /** Number of columns in input tensor. */
+      const int in_col_stride,     /** Stride between columns of the input. */
+      const int n_channels,        /** Number of channels in input tensor. */
+      const PaddingType padding,   /** Padding type. */
+      const int tile_M,
+      const int tile_N,
+      T* const output,             /** Base of output matrices. */
+      const int matrix_stride,     /** Stride between output matrices. */
+      const int matrix_batch_stride,  /** Stride between batches within the matrix. */
+      const int matrix_row_stride  /** Stride within matrices. */
+  );
+
+  protected:
+    using Transform = InputTransformImpl<KernelRows, KernelCols, InnerTileRows, InnerTileCols, T>;
+
+    /* Member values for instance-based API. */
+    const T* const _inptr;
+    T* const _outptr;
+    const int _n_batches, _n_rows, _n_cols, _n_channels, _matrix_stride,
+              _matrix_row_stride, _tiles_M, _tiles_N;
+    const int _in_col_stride, _in_row_stride, _in_batch_stride;
+    const PaddingType _padding_type;
+};
+
+}  // namespace winograd
diff --git a/src/core/NEON/kernels/convolution/winograd/transforms/input_6_3_fp32.cpp b/src/core/NEON/kernels/convolution/winograd/transforms/input_1x8_fp32.cpp
index 67e46499cd..042d4debbc 100644
--- a/src/core/NEON/kernels/convolution/winograd/transforms/input_6_3_fp32.cpp
+++ b/src/core/NEON/kernels/convolution/winograd/transforms/input_1x8_fp32.cpp
@@ -26,24 +26,11 @@
 #include "arm_compute/core/NEON/kernels/convolution/winograd/winograd_gemm.hpp"
 #include "arm_compute/core/NEON/kernels/convolution/common/arm.hpp"
 
-
-namespace winograd
-{
-
-using Transform = WinogradGEMM<1, 6, 1, 3>::InputTransform<float>;
-
-template <>
-template <>
-int Transform::ops_performed(const Tensor4DShape &input_shape)
+namespace
 {
-  (void) input_shape;
-  return 0;  // TODO
-}
 
-template <>
-template <>
 template <int pad_top, int pad_left, int pad_bottom, int pad_right>
-void Transform::process_tile(
+void winograd_input_transform_1x8_fp32_process_tile(
   int n_channels,
   const float* const input_base,
   const int input_row_stride,
@@ -53,23 +40,23 @@ void Transform::process_tile(
 )
 {
   (void) input_row_stride;  // No rows over which to stride
-  constexpr int inner_tile_j = 8;
-  constexpr int cells_j = inner_tile_j - pad_right;
+  constexpr int inner_tile_cols = 8;
+  constexpr int cells_j = inner_tile_cols - pad_right;
 
   float *outptr = matrix_base;
 
   // Get pointers into the input tile
-  const float *x_ptrs[inner_tile_j];
+  const float *x_ptrs[inner_tile_cols];
   for (int j = pad_left, xj = 0; j < cells_j; j++, xj++)
   {
     x_ptrs[j] = input_base + xj*input_col_stride;
   }
 
   // Vectors used/computed in this kernel.
-  float x[inner_tile_j];
-  float U[inner_tile_j];
+  float x[inner_tile_cols];
+  float U[inner_tile_cols];
 
-  for (int j = 0; j < inner_tile_j; j++)
+  for (int j = 0; j < inner_tile_cols; j++)
   {
     x[j] = 0.0f;
   }
@@ -80,7 +67,7 @@ void Transform::process_tile(
 #ifdef __arm_any__
   for (; channels_remaining >= 4; channels_remaining -= 4)
   {
-    float32x4_t x[inner_tile_j], U[inner_tile_j];
+    float32x4_t x[inner_tile_cols], U[inner_tile_cols];
     for (int j = 0; j < inner_tile_cols; j++)
     {
       x[j] = vdupq_n_f32(0.0f);
@@ -104,16 +91,15 @@ void Transform::process_tile(
     U[7] = vmlaq_n_f32(vmlaq_n_f32(vmlaq_n_f32(vmulq_n_f32(x[7], 1), x[3], 49), x[5], -14), x[1], -36);
 
     // Store the transformed vector
-    for (int j = 0; j < inner_tile_j; j++)
+    for (int j = 0; j < inner_tile_cols; j++)
     {
       vst1q_f32(outptr + j*matrix_stride, U[j]);
     }
     outptr += 4;
   }
-
   for (; channels_remaining >= 2; channels_remaining -= 2)
   {
-    float32x2_t x[inner_tile_j], U[inner_tile_j];
+    float32x2_t x[inner_tile_cols], U[inner_tile_cols];
     for (int j = 0; j < inner_tile_cols; j++)
     {
       x[j] = vdup_n_f32(0.0f);
@@ -137,7 +123,7 @@ void Transform::process_tile(
     U[7] = vmla_n_f32(vmla_n_f32(vmla_n_f32(vmul_n_f32(x[7], 1), x[3], 49), x[5], -14), x[1], -36);
 
     // Store the transformed vector
-    for (int j = 0; j < inner_tile_j; j++)
+    for (int j = 0; j < inner_tile_cols; j++)
     {
       vst1_f32(outptr + j*matrix_stride, U[j]);
     }
@@ -163,7 +149,7 @@ void Transform::process_tile(
     U[7] = x[1]*-36 + x[5]*-14 + x[3]*49 + x[7]*1;
 
     // Store the transformed vector
-    for (int j = 0; j < inner_tile_j; j++)
+    for (int j = 0; j < inner_tile_cols; j++)
     {
       *(outptr + j*matrix_stride) = U[j];
     }
@@ -171,56 +157,118 @@ void Transform::process_tile(
   }
 }
 
+}
+
+namespace winograd
+{
+template <int x>
+using Transform = InputTransformImpl<1, x, 1, 8, float>;
+
 template <>
-template <>
-const Transform::TileFn Transform::tile_fns[n_pad_top][n_pad_left][n_pad_bottom][n_pad_right] =
+const Transform<3>::TileFn
+  Transform<3>::tile_fns[n_pad_top][n_pad_left][n_pad_bottom][n_pad_right] =
 {
   {
     {
       {
-        Transform::template process_tile<0, 0, 0, 0>,
-        Transform::template process_tile<0, 0, 0, 1>,
-        Transform::template process_tile<0, 0, 0, 2>,
-        Transform::template process_tile<0, 0, 0, 3>,
-        Transform::template process_tile<0, 0, 0, 4>,
-        Transform::template process_tile<0, 0, 0, 5>,
-        Transform::template process_tile<0, 0, 0, 6>,
+        winograd_input_transform_1x8_fp32_process_tile<0, 0, 0, 0>,
+        winograd_input_transform_1x8_fp32_process_tile<0, 0, 0, 1>,
+        winograd_input_transform_1x8_fp32_process_tile<0, 0, 0, 2>,
+        winograd_input_transform_1x8_fp32_process_tile<0, 0, 0, 3>,
+        winograd_input_transform_1x8_fp32_process_tile<0, 0, 0, 4>,
+        winograd_input_transform_1x8_fp32_process_tile<0, 0, 0, 5>,
+        winograd_input_transform_1x8_fp32_process_tile<0, 0, 0, 6>,
       }
     },
     {
       {
-        Transform::template process_tile<0, 1, 0, 0>,
-        Transform::template process_tile<0, 1, 0, 1>,
-        Transform::template process_tile<0, 1, 0, 2>,
-        Transform::template process_tile<0, 1, 0, 3>,
-        Transform::template process_tile<0, 1, 0, 4>,
-        Transform::template process_tile<0, 1, 0, 5>,
-        Transform::template process_tile<0, 1, 0, 6>,
+        winograd_input_transform_1x8_fp32_process_tile<0, 1, 0, 0>,
+        winograd_input_transform_1x8_fp32_process_tile<0, 1, 0, 1>,
+        winograd_input_transform_1x8_fp32_process_tile<0, 1, 0, 2>,
+        winograd_input_transform_1x8_fp32_process_tile<0, 1, 0, 3>,
+        winograd_input_transform_1x8_fp32_process_tile<0, 1, 0, 4>,
+        winograd_input_transform_1x8_fp32_process_tile<0, 1, 0, 5>,
+        winograd_input_transform_1x8_fp32_process_tile<0, 1, 0, 6>,
       }
     }
   }
 };
 
-template <int x, int y>
-using TransformTransposed = typename WinogradGEMM<x, 1, y, 1>::template InputTransform<float>;
-
 template <>
-template <>
-const TransformTransposed<6, 3>::TileFn
-  TransformTransposed<6, 3>::tile_fns[n_pad_top][n_pad_left][n_pad_bottom][n_pad_right] = {};
-
-template <>
-template <>
-const TransformTransposed<4, 5>::TileFn
-  TransformTransposed<4, 5>::tile_fns[n_pad_top][n_pad_left][n_pad_bottom][n_pad_right] = {};
+const Transform<5>::TileFn
+  Transform<5>::tile_fns[n_pad_top][n_pad_left][n_pad_bottom][n_pad_right] =
+{
+  {
+    {
+      {
+        winograd_input_transform_1x8_fp32_process_tile<0, 0, 0, 0>,
+        winograd_input_transform_1x8_fp32_process_tile<0, 0, 0, 1>,
+        winograd_input_transform_1x8_fp32_process_tile<0, 0, 0, 2>,
+        winograd_input_transform_1x8_fp32_process_tile<0, 0, 0, 3>,
+        winograd_input_transform_1x8_fp32_process_tile<0, 0, 0, 4>,
+        winograd_input_transform_1x8_fp32_process_tile<0, 0, 0, 5>,
+        winograd_input_transform_1x8_fp32_process_tile<0, 0, 0, 6>,
+      }
+    },
+    {
+      {
+        winograd_input_transform_1x8_fp32_process_tile<0, 2, 0, 0>,
+        winograd_input_transform_1x8_fp32_process_tile<0, 2, 0, 1>,
+        winograd_input_transform_1x8_fp32_process_tile<0, 2, 0, 2>,
+        winograd_input_transform_1x8_fp32_process_tile<0, 2, 0, 3>,
+        winograd_input_transform_1x8_fp32_process_tile<0, 2, 0, 4>,
+        winograd_input_transform_1x8_fp32_process_tile<0, 2, 0, 5>,
+        winograd_input_transform_1x8_fp32_process_tile<0, 2, 0, 6>,
+      }
+    }
+  }
+};
 
 template <>
-template <>
-const TransformTransposed<2, 7>::TileFn
-  TransformTransposed<2, 7>::tile_fns[n_pad_top][n_pad_left][n_pad_bottom][n_pad_right] = {};
-
-
+const Transform<7>::TileFn
+  Transform<7>::tile_fns[n_pad_top][n_pad_left][n_pad_bottom][n_pad_right] =
+{
+  {
+    {
+      {
+        winograd_input_transform_1x8_fp32_process_tile<0, 0, 0, 0>,
+        winograd_input_transform_1x8_fp32_process_tile<0, 0, 0, 1>,
+        winograd_input_transform_1x8_fp32_process_tile<0, 0, 0, 2>,
+        winograd_input_transform_1x8_fp32_process_tile<0, 0, 0, 3>,
+        winograd_input_transform_1x8_fp32_process_tile<0, 0, 0, 4>,
+        winograd_input_transform_1x8_fp32_process_tile<0, 0, 0, 5>,
+        winograd_input_transform_1x8_fp32_process_tile<0, 0, 0, 6>,
+      }
+    },
+    {
+      {
+        winograd_input_transform_1x8_fp32_process_tile<0, 1, 0, 0>,
+        winograd_input_transform_1x8_fp32_process_tile<0, 1, 0, 1>,
+        winograd_input_transform_1x8_fp32_process_tile<0, 1, 0, 2>,
+        winograd_input_transform_1x8_fp32_process_tile<0, 1, 0, 3>,
+        winograd_input_transform_1x8_fp32_process_tile<0, 1, 0, 4>,
+        winograd_input_transform_1x8_fp32_process_tile<0, 1, 0, 5>,
+        winograd_input_transform_1x8_fp32_process_tile<0, 1, 0, 6>,
+      }
+    },
+    {
+      {
+        winograd_input_transform_1x8_fp32_process_tile<0, 3, 0, 0>,
+        winograd_input_transform_1x8_fp32_process_tile<0, 3, 0, 1>,
+        winograd_input_transform_1x8_fp32_process_tile<0, 3, 0, 2>,
+        winograd_input_transform_1x8_fp32_process_tile<0, 3, 0, 3>,
+        winograd_input_transform_1x8_fp32_process_tile<0, 3, 0, 4>,
+        winograd_input_transform_1x8_fp32_process_tile<0, 3, 0, 5>,
+        winograd_input_transform_1x8_fp32_process_tile<0, 3, 0, 6>,
+      }
+    }
+  }
+};
 
-template struct WinogradGEMM<1, 6, 1, 3>::InputTransform<float>;
-template struct WinogradGEMM<6, 1, 3, 1>::InputTransform<float>;
+template class InputTransform<1, 3, 1, 8, float>;
+template class InputTransform<3, 1, 8, 1, float>;
+template class InputTransform<1, 5, 1, 8, float>;
+template class InputTransform<5, 1, 8, 1, float>;
+template class InputTransform<1, 7, 1, 8, float>;
+template class InputTransform<7, 1, 8, 1, float>;
 }  // namespace winograd
diff --git a/src/core/NEON/kernels/convolution/winograd/transforms/input_2x2_3x3_fp32.cpp b/src/core/NEON/kernels/convolution/winograd/transforms/input_2x2_3x3_fp32.cpp
index 97b2695d69..a9d5d52d15 100644
--- a/src/core/NEON/kernels/convolution/winograd/transforms/input_2x2_3x3_fp32.cpp
+++ b/src/core/NEON/kernels/convolution/winograd/transforms/input_2x2_3x3_fp32.cpp
@@ -29,22 +29,7 @@
 namespace winograd
 {
 
-using Transform = WinogradGEMM<2, 2, 3, 3>::InputTransform<float>;
-
-/******************************************************************************
- * Cost methods for the input transform.
- * =====================================
- */
-template <>
-template <>
-int Transform::ops_performed(const Tensor4DShape &input_shape)
-{
-  // NOTE: Cost in FLOPs rather than instructions or uops.
-  const int tile_M = iceildiv(input_shape.n_rows, inner_tile_rows);
-  const int tile_N = iceildiv(input_shape.n_cols, inner_tile_cols);
-  return 16 * 16 * tile_M * tile_N * input_shape.n_channels;
-}
-/*****************************************************************************/
+using Transform = InputTransformImpl<3, 3, 4, 4, float>;
 
 /*****************************************************************************
 * F(2x2, 3x3) implies the use of a 4x4 input tile. Such tiles can require a
@@ -100,7 +85,6 @@ int Transform::ops_performed(const Tensor4DShape &input_shape)
 *     Padding right in {0, 1, 2}
 */
 template <>
-template <>
 template <int pad_top, int pad_left, int pad_bottom, int pad_right>
 void Transform::process_tile(
   int n_channels,
@@ -328,7 +312,6 @@ void Transform::process_tile(
 }
 
 template <>
-template <>
 const Transform::TileFn Transform::tile_fns[n_pad_top][n_pad_left][n_pad_bottom][n_pad_right] =
 {
   {
@@ -405,5 +388,5 @@ const Transform::TileFn Transform::tile_fns[n_pad_top][n_pad_left][n_pad_bottom]
   }
 };
 
-template struct WinogradGEMM<2, 2, 3, 3>::InputTransform<float>;
+template class InputTransform<3, 3, 4, 4, float>;
 }  // namespace winograd
diff --git a/src/core/NEON/kernels/convolution/winograd/transforms/input_2x2_5x5_fp32.cpp b/src/core/NEON/kernels/convolution/winograd/transforms/input_2x2_5x5_fp32.cpp
deleted file mode 100644
index 30c9463bb8..0000000000
--- a/src/core/NEON/kernels/convolution/winograd/transforms/input_2x2_5x5_fp32.cpp
+++ /dev/null
@@ -1,458 +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 "arm_compute/core/NEON/kernels/convolution/winograd/transforms/input.hpp"
-#include "arm_compute/core/NEON/kernels/convolution/winograd/winograd_gemm.hpp"
-#include "arm_compute/core/NEON/kernels/convolution/common/arm.hpp"
-
-namespace winograd
-{
-
-using Transform = WinogradGEMM<2, 2, 5, 5>::InputTransform<float>;
-
-template <>
-template <>
-int Transform::ops_performed(const Tensor4DShape &input_shape)
-{
-  (void) input_shape;
-  return 0;  // TODO
-}
-
-/*****************************************************************************
-* F(2x2, 5x5) implies the use of a 6x6 input tile.
-*
-* Build an array of the specialised methods that deal with each of the
-* different padding combinations which may be required. These padding
-* constraints are the space:
-*
-*     Padding top in {0, 2}
-*     Padding left in {0, 2}
-*     Padding bottom in {0, 1, 2, 3, 4}
-*     Padding right in {0, 1, 2, 3, 4}
-*/
-template <>
-template <>
-template <int pad_top, int pad_left, int pad_bottom, int pad_right>
-void Transform::process_tile(
-  int n_channels,
-  const float* const input_base,
-  const int input_row_stride,
-  const int input_col_stride,
-  float* const matrix_base,
-  const int matrix_stride
-)
-{
-  constexpr int cells_i = 6 - pad_bottom;
-  constexpr int cells_j = 6 - pad_right;
-
-  float *outptr = matrix_base;
-
-  // Get pointers into the input tile
-  const float *x_ptrs[6][6];
-  for (int i = pad_top, xi = 0; i < cells_i; i++, xi++)
-  {
-    // Get a pointer into the row
-    const float* const row_ptr = input_base + xi*input_row_stride;
-
-    for (int j = pad_left, xj = 0; j < cells_j; j++, xj++)
-    {
-      x_ptrs[i][j] = row_ptr + xj*input_col_stride;
-    }
-  }
-
-  // Matrices used/computed in this kernel.
-  float x[6][6], XTx[6][6], U[6][6];
-  for (int i = 0; i < 6; i++)
-  {
-    for (int j = 0; j < 6; j++)
-    {
-      x[i][j] = XTx[i][j] = 0.0f;
-    }
-  }
-
-  // Perform the Winograd input transformation for each channel in the input
-  // tensor.
-  int channels_remaining = n_channels;
-#ifdef __aarch64__
-  for (; channels_remaining >= 4; channels_remaining -= 4)
-  {
-    // Matrices used/computed in this kernel
-    float32x4_t x[6][6], XTx[6][6], U[6][6];
-    for (int i = 0; i < 6; i++)
-    {
-      for (int j = 0; j < 6; j++)
-      {
-        x[i][j] = vdupq_n_f32(0.0f);
-        XTx[i][j] = vdupq_n_f32(0.0f);
-      }
-    }
-
-    // Read a 6x6 tile in the Winograd domain
-    for (int i = pad_top; i < cells_i; i++)
-    {
-      for (int j = pad_left; j < cells_j; j++)
-      {
-        x[i][j] = vld1q_f32(x_ptrs[i][j]);
-        x_ptrs[i][j] += 4;
-      }
-    }
-
-    // Compute XT . x
-    for (int j = pad_left; j < cells_j; j++)
-    {
-      // XTx[0][j] =  4*x[0][j] + -5*x[2][j] +  1*x[4][j];
-      XTx[0][j] = vmlsq_n_f32(vmlaq_n_f32(x[4][j], x[0][j], 4.0f), x[2][j], 5.0f);
-
-      // XTx[1][j] = -4*x[1][j] + -4*x[2][j] +  1*x[3][j] +  1*x[4][j];
-      XTx[1][j] = vmlsq_n_f32(vaddq_f32(x[3][j], x[4][j]), vaddq_f32(x[1][j], x[2][j]), 4.0f);
-
-      // XTx[2][j] =  4*x[1][j] + -4*x[2][j] + -1*x[3][j] +  1*x[4][j];
-      XTx[2][j] = vmlaq_n_f32(vsubq_f32(x[4][j], x[3][j]), vsubq_f32(x[1][j], x[2][j]), 4.0f);
-
-      // XTx[3][j] = -2*x[1][j] + -1*x[2][j] +  2*x[3][j] +  1*x[4][j];
-      XTx[3][j] = vmlaq_n_f32(vsubq_f32(x[4][j], x[2][j]), vsubq_f32(x[3][j], x[1][j]), 2.0f);
-
-      // XTx[4][j] =  2*x[1][j] + -1*x[2][j] + -2*x[3][j] +  1*x[4][j];
-      XTx[4][j] = vmlaq_n_f32(vsubq_f32(x[4][j], x[2][j]), vsubq_f32(x[1][j], x[3][j]), 2.0f);
-
-      // XTx[5][j] =  4*x[1][j] + -5*x[3][j] +  1*x[5][j];
-      XTx[5][j] = vmlsq_n_f32(vmlaq_n_f32(x[5][j], x[1][j], 4.0f), x[3][j], 5.0f);
-    }
-
-    // Compute U = XT . x . X
-    for (int i = 0; i < 6; i++)
-    {
-      // U[i][0] =  4*XTx[i][0] + -5*XTx[i][2] +  1*XTx[i][4];
-      U[i][0] = vmlsq_n_f32(vmlaq_n_f32(XTx[i][4], XTx[i][0], 4.0f), XTx[i][2], 5.0f);
-
-      // U[i][1] = -4*XTx[i][1] + -4*XTx[i][2] +  1*XTx[i][3] +  1*XTx[i][4];
-      U[i][1] = vmlsq_n_f32(vaddq_f32(XTx[i][3], XTx[i][4]), vaddq_f32(XTx[i][1], XTx[i][2]), 4.0f);
-
-      // U[i][2] =  4*XTx[i][1] + -4*XTx[i][2] + -1*XTx[i][3] +  1*XTx[i][4];
-      U[i][2] = vmlaq_n_f32(vsubq_f32(XTx[i][4], XTx[i][3]), vsubq_f32(XTx[i][1], XTx[i][2]), 4.0f);
-
-      // U[i][3] = -2*XTx[i][1] + -1*XTx[i][2] +  2*XTx[i][3] +  1*XTx[i][4];
-      U[i][3] = vmlaq_n_f32(vsubq_f32(XTx[i][4], XTx[i][2]), vsubq_f32(XTx[i][3], XTx[i][1]), 2.0f);
-
-      // U[i][4] =  2*XTx[i][1] + -1*XTx[i][2] + -2*XTx[i][3] +  1*XTx[i][4];
-      U[i][4] = vmlaq_n_f32(vsubq_f32(XTx[i][4], XTx[i][2]), vsubq_f32(XTx[i][1], XTx[i][3]), 2.0f);
-
-      // U[i][5] =  4*XTx[i][1] + -5*XTx[i][3] +  1*XTx[i][5];
-      U[i][5] = vmlsq_n_f32(vmlaq_n_f32(XTx[i][5], XTx[i][1], 4.0f), XTx[i][3], 5.0f);
-    }
-
-    // Store the transformed matrix
-    for (int i = 0, m = 0; i < 6; i++)
-    {
-      for (int j = 0; j < 6; j++, m++)
-      {
-        vst1q_f32(outptr + m*matrix_stride, U[i][j]);
-      }
-    }
-    outptr += 4;
-  }
-#endif  // __aarch64__
-#ifdef __arm_any__
-  for (; channels_remaining >= 2; channels_remaining -= 2)
-  {
-    // Matrices used/computed in this kernel
-    float32x2_t x[6][6], XTx[6][6], U[6][6];
-    for (int i = 0; i < 6; i++)
-    {
-      for (int j = 0; j < 6; j++)
-      {
-        x[i][j] = vdup_n_f32(0.0f);
-        XTx[i][j] = vdup_n_f32(0.0f);
-      }
-    }
-
-    // Read a 6x6 tile in the Winograd domain
-    for (int i = pad_top; i < cells_i; i++)
-    {
-      for (int j = pad_left; j < cells_j; j++)
-      {
-        x[i][j] = vld1_f32(x_ptrs[i][j]);
-        x_ptrs[i][j] += 2;
-      }
-    }
-
-    // Compute XT . x
-    for (int j = pad_left; j < cells_j; j++)
-    {
-      // XTx[0][j] =  4*x[0][j] + -5*x[2][j] +  1*x[4][j];
-      XTx[0][j] = vmls_n_f32(vmla_n_f32(x[4][j], x[0][j], 4.0f), x[2][j], 5.0f);
-
-      // XTx[1][j] = -4*x[1][j] + -4*x[2][j] +  1*x[3][j] +  1*x[4][j];
-      XTx[1][j] = vmls_n_f32(vadd_f32(x[3][j], x[4][j]), vadd_f32(x[1][j], x[2][j]), 4.0f);
-
-      // XTx[2][j] =  4*x[1][j] + -4*x[2][j] + -1*x[3][j] +  1*x[4][j];
-      XTx[2][j] = vmla_n_f32(vsub_f32(x[4][j], x[3][j]), vsub_f32(x[1][j], x[2][j]), 4.0f);
-
-      // XTx[3][j] = -2*x[1][j] + -1*x[2][j] +  2*x[3][j] +  1*x[4][j];
-      XTx[3][j] = vmla_n_f32(vsub_f32(x[4][j], x[2][j]), vsub_f32(x[3][j], x[1][j]), 2.0f);
-
-      // XTx[4][j] =  2*x[1][j] + -1*x[2][j] + -2*x[3][j] +  1*x[4][j];
-      XTx[4][j] = vmla_n_f32(vsub_f32(x[4][j], x[2][j]), vsub_f32(x[1][j], x[3][j]), 2.0f);
-
-      // XTx[5][j] =  4*x[1][j] + -5*x[3][j] +  1*x[5][j];
-      XTx[5][j] = vmls_n_f32(vmla_n_f32(x[5][j], x[1][j], 4.0f), x[3][j], 5.0f);
-    }
-
-    // Compute U = XT . x . X
-    for (int i = 0; i < 6; i++)
-    {
-      // U[i][0] =  4*XTx[i][0] + -5*XTx[i][2] +  1*XTx[i][4];
-      U[i][0] = vmls_n_f32(vmla_n_f32(XTx[i][4], XTx[i][0], 4.0f), XTx[i][2], 5.0f);
-
-      // U[i][1] = -4*XTx[i][1] + -4*XTx[i][2] +  1*XTx[i][3] +  1*XTx[i][4];
-      U[i][1] = vmls_n_f32(vadd_f32(XTx[i][3], XTx[i][4]), vadd_f32(XTx[i][1], XTx[i][2]), 4.0f);
-
-      // U[i][2] =  4*XTx[i][1] + -4*XTx[i][2] + -1*XTx[i][3] +  1*XTx[i][4];
-      U[i][2] = vmla_n_f32(vsub_f32(XTx[i][4], XTx[i][3]), vsub_f32(XTx[i][1], XTx[i][2]), 4.0f);
-
-      // U[i][3] = -2*XTx[i][1] + -1*XTx[i][2] +  2*XTx[i][3] +  1*XTx[i][4];
-      U[i][3] = vmla_n_f32(vsub_f32(XTx[i][4], XTx[i][2]), vsub_f32(XTx[i][3], XTx[i][1]), 2.0f);
-
-      // U[i][4] =  2*XTx[i][1] + -1*XTx[i][2] + -2*XTx[i][3] +  1*XTx[i][4];
-      U[i][4] = vmla_n_f32(vsub_f32(XTx[i][4], XTx[i][2]), vsub_f32(XTx[i][1], XTx[i][3]), 2.0f);
-
-      // U[i][5] =  4*XTx[i][1] + -5*XTx[i][3] +  1*XTx[i][5];
-      U[i][5] = vmls_n_f32(vmla_n_f32(XTx[i][5], XTx[i][1], 4.0f), XTx[i][3], 5.0f);
-    }
-
-    // Store the transformed matrix
-    for (int i = 0, m = 0; i < 6; i++)
-    {
-      for (int j = 0; j < 6; j++, m++)
-      {
-        vst1_f32(outptr + m*matrix_stride, U[i][j]);
-      }
-    }
-    outptr += 2;
-  }
-#endif  // __arm_any__
-  for (; channels_remaining; channels_remaining--)
-  {
-    // Load x
-    for (int i = pad_top; i < cells_i; i++)
-    {
-      for (int j = pad_left; j < cells_j; j++)
-      {
-        x[i][j] = *(x_ptrs[i][j]++);
-      }
-    }
-
-    // Compute XT . x
-    for (int j = pad_left; j < cells_j; j++)
-    {
-      XTx[0][j] =  4*x[0][j] + -5*x[2][j] +  1*x[4][j];
-      XTx[1][j] = -4*x[1][j] + -4*x[2][j] +  1*x[3][j] +  1*x[4][j];
-      XTx[2][j] =  4*x[1][j] + -4*x[2][j] + -1*x[3][j] +  1*x[4][j];
-      XTx[3][j] = -2*x[1][j] + -1*x[2][j] +  2*x[3][j] +  1*x[4][j];
-      XTx[4][j] =  2*x[1][j] + -1*x[2][j] + -2*x[3][j] +  1*x[4][j];
-      XTx[5][j] =  4*x[1][j] + -5*x[3][j] +  1*x[5][j];
-    }
-
-    // Compute U = XT . x . X
-    for (int i = 0; i < 6; i++)
-    {
-      U[i][0] =  4*XTx[i][0] + -5*XTx[i][2] +  1*XTx[i][4];
-      U[i][1] = -4*XTx[i][1] + -4*XTx[i][2] +  1*XTx[i][3] +  1*XTx[i][4];
-      U[i][2] =  4*XTx[i][1] + -4*XTx[i][2] + -1*XTx[i][3] +  1*XTx[i][4];
-      U[i][3] = -2*XTx[i][1] + -1*XTx[i][2] +  2*XTx[i][3] +  1*XTx[i][4];
-      U[i][4] =  2*XTx[i][1] + -1*XTx[i][2] + -2*XTx[i][3] +  1*XTx[i][4];
-      U[i][5] =  4*XTx[i][1] + -5*XTx[i][3] +  1*XTx[i][5];
-    }
-
-    // Store the transformed matrix
-    for (int i = 0, m = 0; i < 6; i++)
-    {
-      for (int j = 0; j < 6; j++, m++)
-      {
-        *(outptr + m*matrix_stride) = U[i][j];
-      }
-    }
-    outptr++;
-  }
-}
-
-template <>
-template <>
-const Transform::TileFn Transform::tile_fns[n_pad_top][n_pad_left][n_pad_bottom][n_pad_right] =
-{
-  {
-    {
-      {
-        Transform::template process_tile<0, 0, 0, 0>,  // No padding
-        Transform::template process_tile<0, 0, 0, 1>,  // Right
-        Transform::template process_tile<0, 0, 0, 2>,  // "   "
-        Transform::template process_tile<0, 0, 0, 3>,  // "   "
-        Transform::template process_tile<0, 0, 0, 4>,  // "   "
-      },
-      {
-        Transform::template process_tile<0, 0, 1, 0>,  // Bottom
-        Transform::template process_tile<0, 0, 1, 1>,  // Bottom right
-        Transform::template process_tile<0, 0, 1, 2>,  // "          "
-        Transform::template process_tile<0, 0, 1, 3>,  // "          "
-        Transform::template process_tile<0, 0, 1, 4>,  // "          "
-      },
-      {
-        Transform::template process_tile<0, 0, 2, 0>,  // Bottom
-        Transform::template process_tile<0, 0, 2, 1>,  // Bottom right
-        Transform::template process_tile<0, 0, 2, 2>,  // "          "
-        Transform::template process_tile<0, 0, 2, 3>,  // "          "
-        Transform::template process_tile<0, 0, 2, 4>,  // "          "
-      },
-      {
-        Transform::template process_tile<0, 0, 3, 0>,  // Bottom
-        Transform::template process_tile<0, 0, 3, 1>,  // Bottom right
-        Transform::template process_tile<0, 0, 3, 2>,  // "          "
-        Transform::template process_tile<0, 0, 3, 3>,  // "          "
-        Transform::template process_tile<0, 0, 3, 4>,  // "          "
-      },
-      {
-        Transform::template process_tile<0, 0, 4, 0>,  // Bottom
-        Transform::template process_tile<0, 0, 4, 1>,  // Bottom right
-        Transform::template process_tile<0, 0, 4, 2>,  // "          "
-        Transform::template process_tile<0, 0, 4, 3>,  // "          "
-        Transform::template process_tile<0, 0, 4, 4>,  // "          "
-      }
-    },
-    {
-      {
-        Transform::template process_tile<0, 2, 0, 0>,  // Left
-        Transform::template process_tile<0, 2, 0, 1>,
-        Transform::template process_tile<0, 2, 0, 2>,
-        Transform::template process_tile<0, 2, 0, 3>,
-        Transform::template process_tile<0, 2, 0, 4>,
-      },
-      {
-        Transform::template process_tile<0, 2, 1, 0>,  // Bottom left
-        Transform::template process_tile<0, 2, 1, 1>,
-        Transform::template process_tile<0, 2, 1, 2>,
-        Transform::template process_tile<0, 2, 1, 3>,
-        Transform::template process_tile<0, 2, 1, 4>,
-      },
-      {
-        Transform::template process_tile<0, 2, 2, 0>,  // "          "
-        Transform::template process_tile<0, 2, 2, 1>,
-        Transform::template process_tile<0, 2, 2, 2>,
-        Transform::template process_tile<0, 2, 2, 3>,
-        Transform::template process_tile<0, 2, 2, 4>,
-      },
-      {
-        Transform::template process_tile<0, 2, 3, 0>,  // "          "
-        Transform::template process_tile<0, 2, 3, 1>,
-        Transform::template process_tile<0, 2, 3, 2>,
-        Transform::template process_tile<0, 2, 3, 3>,
-        Transform::template process_tile<0, 2, 3, 4>,
-      },
-      {
-        Transform::template process_tile<0, 2, 4, 0>,  // "          "
-        Transform::template process_tile<0, 2, 4, 1>,
-        Transform::template process_tile<0, 2, 4, 2>,
-        Transform::template process_tile<0, 2, 4, 3>,
-        Transform::template process_tile<0, 2, 4, 4>,
-      }
-    }
-  },
-  {
-    {
-      {
-        Transform::template process_tile<2, 0, 0, 0>,  // Top
-        Transform::template process_tile<2, 0, 0, 1>,  // Top right
-        Transform::template process_tile<2, 0, 0, 2>,  // "       "
-        Transform::template process_tile<2, 0, 0, 3>,  // "       "
-        Transform::template process_tile<2, 0, 0, 4>,  // "       "
-      },
-      {
-        Transform::template process_tile<2, 0, 1, 0>,
-        Transform::template process_tile<2, 0, 1, 1>,
-        Transform::template process_tile<2, 0, 1, 2>,
-        Transform::template process_tile<2, 0, 1, 3>,
-        Transform::template process_tile<2, 0, 1, 4>,
-      },
-      {
-        Transform::template process_tile<2, 0, 2, 0>,
-        Transform::template process_tile<2, 0, 2, 1>,
-        Transform::template process_tile<2, 0, 2, 2>,
-        Transform::template process_tile<2, 0, 2, 3>,
-        Transform::template process_tile<2, 0, 2, 4>,
-      },
-      {
-        Transform::template process_tile<2, 0, 3, 0>,
-        Transform::template process_tile<2, 0, 3, 1>,
-        Transform::template process_tile<2, 0, 3, 2>,
-        Transform::template process_tile<2, 0, 3, 3>,
-        Transform::template process_tile<2, 0, 3, 4>,
-      },
-      {
-        Transform::template process_tile<2, 0, 4, 0>,
-        Transform::template process_tile<2, 0, 4, 1>,
-        Transform::template process_tile<2, 0, 4, 2>,
-        Transform::template process_tile<2, 0, 4, 3>,
-        Transform::template process_tile<2, 0, 4, 4>,
-      },
-    },
-    {
-      {
-        Transform::template process_tile<2, 2, 0, 0>,  // Top left
-        Transform::template process_tile<2, 2, 0, 1>,
-        Transform::template process_tile<2, 2, 0, 2>,
-        Transform::template process_tile<2, 2, 0, 3>,
-        Transform::template process_tile<2, 2, 0, 4>,
-      },
-      {
-        Transform::template process_tile<2, 2, 1, 0>,
-        Transform::template process_tile<2, 2, 1, 1>,
-        Transform::template process_tile<2, 2, 1, 2>,
-        Transform::template process_tile<2, 2, 1, 3>,
-        Transform::template process_tile<2, 2, 1, 4>,
-      },
-      {
-        Transform::template process_tile<2, 2, 2, 0>,
-        Transform::template process_tile<2, 2, 2, 1>,
-        Transform::template process_tile<2, 2, 2, 2>,
-        Transform::template process_tile<2, 2, 2, 3>,
-        Transform::template process_tile<2, 2, 2, 4>,
-      },
-      {
-        Transform::template process_tile<2, 2, 3, 0>,
-        Transform::template process_tile<2, 2, 3, 1>,
-        Transform::template process_tile<2, 2, 3, 2>,
-        Transform::template process_tile<2, 2, 3, 3>,
-        Transform::template process_tile<2, 2, 3, 4>,
-      },
-      {
-        Transform::template process_tile<2, 2, 4, 0>,
-        Transform::template process_tile<2, 2, 4, 1>,
-        Transform::template process_tile<2, 2, 4, 2>,
-        Transform::template process_tile<2, 2, 4, 3>,
-        Transform::template process_tile<2, 2, 4, 4>,
-      }
-    }
-  }
-};
-
-template struct WinogradGEMM<2, 2, 5, 5>::InputTransform<float>;
-}  // namespace winograd
diff --git a/src/core/NEON/kernels/convolution/winograd/transforms/input_4x4_3x3_fp32.cpp b/src/core/NEON/kernels/convolution/winograd/transforms/input_4x4_3x3_fp32.cpp
deleted file mode 100644
index 7f93187132..0000000000
--- a/src/core/NEON/kernels/convolution/winograd/transforms/input_4x4_3x3_fp32.cpp
+++ /dev/null
@@ -1,486 +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 "arm_compute/core/NEON/kernels/convolution/winograd/transforms/input.hpp"
-#include "arm_compute/core/NEON/kernels/convolution/winograd/winograd_gemm.hpp"
-#include "arm_compute/core/NEON/kernels/convolution/common/arm.hpp"
-
-namespace winograd
-{
-
-using Transform = WinogradGEMM<4, 4, 3, 3>::InputTransform<float>;
-
-template <>
-template <>
-int Transform::ops_performed(const Tensor4DShape &input_shape)
-{
-  // NOTE: Cost in FLOPs rather than instructions or uops.
-  const int tile_M = iceildiv(input_shape.n_rows, inner_tile_rows);
-  const int tile_N = iceildiv(input_shape.n_cols, inner_tile_cols);
-  return 12 * 24 * tile_M * tile_N * input_shape.n_channels;
-}
-
-/* F(4x4, 3x3) implies the use of a 6x6 input tile. Such tiles can require a
-* variety of padding types. For example, tiles at the top and left of an
-* image can require one row or column of padding on their top and left sides
-* if the padding type is SAME (where X represents a padded value):
-*
-*      ___________    ___________
-*     |X X X X X X|  |X X X X X X|
-*     |X          |  |           |
-*     |X          |  |           |
-*     |X          |  |           |
-*     |X          |  |           |
-*     |X__________|  |___________|
-*      ___________
-*     |X          |
-*     |X          |
-*     |X          |
-*     |X          |
-*     |X          |
-*     |X__________|
-*
-* For tiles near the right or bottom of the image it is more complicated.
-* Such tiles might require padding by 0, 1, 2 or 3 rows or columns if the
-* padding type is VALID or 1, 2, 3 or 4 rows or columns if the padding
-* type is SAME.
-*
-* Build an array of the specialised methods that deal with each of the
-* different padding combinations which may be required. These padding
-* constraints are the space:
-*
-*     Padding top in {0, 1}
-*     Padding left in {0, 1}
-*     Padding bottom in {0, 1, 2, 3, 4}
-*     Padding right in {0, 1, 2, 3, 4}
-*/
-template <>
-template <>
-template <int pad_top, int pad_left, int pad_bottom, int pad_right>
-void Transform::process_tile(
-  int n_channels,
-  const float* const input_base,
-  const int input_row_stride,
-  const int input_col_stride,
-  float* const matrix_base,
-  const int matrix_stride
-)
-{
-  constexpr int cells_i = 6 - pad_bottom;
-  constexpr int cells_j = 6 - pad_right;
-
-  float *outptr = matrix_base;
-
-  // Get pointers into the input tile
-  const float *x_ptrs[6][6];
-  for (int i = pad_top, xi = 0; i < cells_i; i++, xi++)
-  {
-    // Get a pointer into the row
-    const float* const row_ptr = input_base + xi*input_row_stride;
-
-    for (int j = pad_left, xj = 0; j < cells_j; j++, xj++)
-    {
-      x_ptrs[i][j] = row_ptr + xj*input_col_stride;
-    }
-  }
-
-  // Matrices used/computed in this kernel.
-  float x[6][6], XTx[6][6], U[6][6];
-  for (int i = 0; i < 6; i++)
-  {
-    for (int j = 0; j < 6; j++)
-    {
-      x[i][j] = XTx[i][j] = 0.0f;
-    }
-  }
-
-  // Perform the Winograd input transformation for each channel in the input
-  // tensor.
-  int channels_remaining = n_channels;
-#ifdef __aarch64__
-  for (; channels_remaining >= 4; channels_remaining -= 4)
-  {
-    // Matrices used/computed in this kernel
-    float32x4_t x[6][6], XTx[6][6], U[6][6];
-    for (int i = 0; i < 6; i++)
-    {
-      for (int j = 0; j < 6; j++)
-      {
-        x[i][j] = vdupq_n_f32(0.0f);
-        XTx[i][j] = vdupq_n_f32(0.0f);
-      }
-    }
-
-    // Read a 6x6 tile in the Winograd domain
-    for (int i = pad_top; i < cells_i; i++)
-    {
-      for (int j = pad_left; j < cells_j; j++)
-      {
-        x[i][j] = vld1q_f32(x_ptrs[i][j]);
-        x_ptrs[i][j] += 4;
-      }
-    }
-
-    // Compute XT . x
-    for (int j = pad_left; j < cells_j; j++)
-    {
-      // XTx[0][j] =  4*x[0][j] + -5*x[2][j] +  1*x[4][j];
-      XTx[0][j] = vmlsq_n_f32(vmlaq_n_f32(x[4][j], x[0][j], 4.0f), x[2][j], 5.0f);
-
-      // XTx[1][j] = -4*x[1][j] + -4*x[2][j] +  1*x[3][j] +  1*x[4][j];
-      XTx[1][j] = vmlsq_n_f32(vaddq_f32(x[3][j], x[4][j]), vaddq_f32(x[1][j], x[2][j]), 4.0f);
-
-      // XTx[2][j] =  4*x[1][j] + -4*x[2][j] + -1*x[3][j] +  1*x[4][j];
-      XTx[2][j] = vmlaq_n_f32(vsubq_f32(x[4][j], x[3][j]), vsubq_f32(x[1][j], x[2][j]), 4.0f);
-
-      // XTx[3][j] = -2*x[1][j] + -1*x[2][j] +  2*x[3][j] +  1*x[4][j];
-      XTx[3][j] = vmlaq_n_f32(vsubq_f32(x[4][j], x[2][j]), vsubq_f32(x[3][j], x[1][j]), 2.0f);
-
-      // XTx[4][j] =  2*x[1][j] + -1*x[2][j] + -2*x[3][j] +  1*x[4][j];
-      XTx[4][j] = vmlaq_n_f32(vsubq_f32(x[4][j], x[2][j]), vsubq_f32(x[1][j], x[3][j]), 2.0f);
-
-      // XTx[5][j] =  4*x[1][j] + -5*x[3][j] +  1*x[5][j];
-      XTx[5][j] = vmlsq_n_f32(vmlaq_n_f32(x[5][j], x[1][j], 4.0f), x[3][j], 5.0f);
-    }
-
-    // Compute U = XT . x . X
-    for (int i = 0; i < 6; i++)
-    {
-      // U[i][0] =  4*XTx[i][0] + -5*XTx[i][2] +  1*XTx[i][4];
-      U[i][0] = vmlsq_n_f32(vmlaq_n_f32(XTx[i][4], XTx[i][0], 4.0f), XTx[i][2], 5.0f);
-
-      // U[i][1] = -4*XTx[i][1] + -4*XTx[i][2] +  1*XTx[i][3] +  1*XTx[i][4];
-      U[i][1] = vmlsq_n_f32(vaddq_f32(XTx[i][3], XTx[i][4]), vaddq_f32(XTx[i][1], XTx[i][2]), 4.0f);
-
-      // U[i][2] =  4*XTx[i][1] + -4*XTx[i][2] + -1*XTx[i][3] +  1*XTx[i][4];
-      U[i][2] = vmlaq_n_f32(vsubq_f32(XTx[i][4], XTx[i][3]), vsubq_f32(XTx[i][1], XTx[i][2]), 4.0f);
-
-      // U[i][3] = -2*XTx[i][1] + -1*XTx[i][2] +  2*XTx[i][3] +  1*XTx[i][4];
-      U[i][3] = vmlaq_n_f32(vsubq_f32(XTx[i][4], XTx[i][2]), vsubq_f32(XTx[i][3], XTx[i][1]), 2.0f);
-
-      // U[i][4] =  2*XTx[i][1] + -1*XTx[i][2] + -2*XTx[i][3] +  1*XTx[i][4];
-      U[i][4] = vmlaq_n_f32(vsubq_f32(XTx[i][4], XTx[i][2]), vsubq_f32(XTx[i][1], XTx[i][3]), 2.0f);
-
-      // U[i][5] =  4*XTx[i][1] + -5*XTx[i][3] +  1*XTx[i][5];
-      U[i][5] = vmlsq_n_f32(vmlaq_n_f32(XTx[i][5], XTx[i][1], 4.0f), XTx[i][3], 5.0f);
-    }
-
-    // Store the transformed matrix
-    for (int i = 0, m = 0; i < 6; i++)
-    {
-      for (int j = 0; j < 6; j++, m++)
-      {
-        vst1q_f32(outptr + m*matrix_stride, U[i][j]);
-      }
-    }
-    outptr += 4;
-  }
-#endif  // __aarch64__
-#ifdef __arm_any__
-  for (; channels_remaining >= 2; channels_remaining -= 2)
-  {
-    // Matrices used/computed in this kernel
-    float32x2_t x[6][6], XTx[6][6], U[6][6];
-    for (int i = 0; i < 6; i++)
-    {
-      for (int j = 0; j < 6; j++)
-      {
-        x[i][j] = vdup_n_f32(0.0f);
-        XTx[i][j] = vdup_n_f32(0.0f);
-      }
-    }
-
-    // Read a 6x6 tile in the Winograd domain
-    for (int i = pad_top; i < cells_i; i++)
-    {
-      for (int j = pad_left; j < cells_j; j++)
-      {
-        x[i][j] = vld1_f32(x_ptrs[i][j]);
-        x_ptrs[i][j] += 2;
-      }
-    }
-
-    // Compute XT . x
-    for (int j = pad_left; j < cells_j; j++)
-    {
-      // XTx[0][j] =  4*x[0][j] + -5*x[2][j] +  1*x[4][j];
-      XTx[0][j] = vmls_n_f32(vmla_n_f32(x[4][j], x[0][j], 4.0f), x[2][j], 5.0f);
-
-      // XTx[1][j] = -4*x[1][j] + -4*x[2][j] +  1*x[3][j] +  1*x[4][j];
-      XTx[1][j] = vmls_n_f32(vadd_f32(x[3][j], x[4][j]), vadd_f32(x[1][j], x[2][j]), 4.0f);
-
-      // XTx[2][j] =  4*x[1][j] + -4*x[2][j] + -1*x[3][j] +  1*x[4][j];
-      XTx[2][j] = vmla_n_f32(vsub_f32(x[4][j], x[3][j]), vsub_f32(x[1][j], x[2][j]), 4.0f);
-
-      // XTx[3][j] = -2*x[1][j] + -1*x[2][j] +  2*x[3][j] +  1*x[4][j];
-      XTx[3][j] = vmla_n_f32(vsub_f32(x[4][j], x[2][j]), vsub_f32(x[3][j], x[1][j]), 2.0f);
-
-      // XTx[4][j] =  2*x[1][j] + -1*x[2][j] + -2*x[3][j] +  1*x[4][j];
-      XTx[4][j] = vmla_n_f32(vsub_f32(x[4][j], x[2][j]), vsub_f32(x[1][j], x[3][j]), 2.0f);
-
-      // XTx[5][j] =  4*x[1][j] + -5*x[3][j] +  1*x[5][j];
-      XTx[5][j] = vmls_n_f32(vmla_n_f32(x[5][j], x[1][j], 4.0f), x[3][j], 5.0f);
-    }
-
-    // Compute U = XT . x . X
-    for (int i = 0; i < 6; i++)
-    {
-      // U[i][0] =  4*XTx[i][0] + -5*XTx[i][2] +  1*XTx[i][4];
-      U[i][0] = vmls_n_f32(vmla_n_f32(XTx[i][4], XTx[i][0], 4.0f), XTx[i][2], 5.0f);
-
-      // U[i][1] = -4*XTx[i][1] + -4*XTx[i][2] +  1*XTx[i][3] +  1*XTx[i][4];
-      U[i][1] = vmls_n_f32(vadd_f32(XTx[i][3], XTx[i][4]), vadd_f32(XTx[i][1], XTx[i][2]), 4.0f);
-
-      // U[i][2] =  4*XTx[i][1] + -4*XTx[i][2] + -1*XTx[i][3] +  1*XTx[i][4];
-      U[i][2] = vmla_n_f32(vsub_f32(XTx[i][4], XTx[i][3]), vsub_f32(XTx[i][1], XTx[i][2]), 4.0f);
-
-      // U[i][3] = -2*XTx[i][1] + -1*XTx[i][2] +  2*XTx[i][3] +  1*XTx[i][4];
-      U[i][3] = vmla_n_f32(vsub_f32(XTx[i][4], XTx[i][2]), vsub_f32(XTx[i][3], XTx[i][1]), 2.0f);
-
-      // U[i][4] =  2*XTx[i][1] + -1*XTx[i][2] + -2*XTx[i][3] +  1*XTx[i][4];
-      U[i][4] = vmla_n_f32(vsub_f32(XTx[i][4], XTx[i][2]), vsub_f32(XTx[i][1], XTx[i][3]), 2.0f);
-
-      // U[i][5] =  4*XTx[i][1] + -5*XTx[i][3] +  1*XTx[i][5];
-      U[i][5] = vmls_n_f32(vmla_n_f32(XTx[i][5], XTx[i][1], 4.0f), XTx[i][3], 5.0f);
-    }
-
-    // Store the transformed matrix
-    for (int i = 0, m = 0; i < 6; i++)
-    {
-      for (int j = 0; j < 6; j++, m++)
-      {
-        vst1_f32(outptr + m*matrix_stride, U[i][j]);
-      }
-    }
-    outptr += 2;
-  }
-#endif  // __arm_any__
-  for (; channels_remaining; channels_remaining--)
-  {
-    // Load x
-    for (int i = pad_top; i < cells_i; i++)
-    {
-      for (int j = pad_left; j < cells_j; j++)
-      {
-        x[i][j] = *(x_ptrs[i][j]++);
-      }
-    }
-
-    // Compute XT . x
-    for (int j = pad_left; j < cells_j; j++)
-    {
-      XTx[0][j] =  4*x[0][j] + -5*x[2][j] +  1*x[4][j];
-      XTx[1][j] = -4*x[1][j] + -4*x[2][j] +  1*x[3][j] +  1*x[4][j];
-      XTx[2][j] =  4*x[1][j] + -4*x[2][j] + -1*x[3][j] +  1*x[4][j];
-      XTx[3][j] = -2*x[1][j] + -1*x[2][j] +  2*x[3][j] +  1*x[4][j];
-      XTx[4][j] =  2*x[1][j] + -1*x[2][j] + -2*x[3][j] +  1*x[4][j];
-      XTx[5][j] =  4*x[1][j] + -5*x[3][j] +  1*x[5][j];
-    }
-
-    // Compute U = XT . x . X
-    for (int i = 0; i < 6; i++)
-    {
-      U[i][0] =  4*XTx[i][0] + -5*XTx[i][2] +  1*XTx[i][4];
-      U[i][1] = -4*XTx[i][1] + -4*XTx[i][2] +  1*XTx[i][3] +  1*XTx[i][4];
-      U[i][2] =  4*XTx[i][1] + -4*XTx[i][2] + -1*XTx[i][3] +  1*XTx[i][4];
-      U[i][3] = -2*XTx[i][1] + -1*XTx[i][2] +  2*XTx[i][3] +  1*XTx[i][4];
-      U[i][4] =  2*XTx[i][1] + -1*XTx[i][2] + -2*XTx[i][3] +  1*XTx[i][4];
-      U[i][5] =  4*XTx[i][1] + -5*XTx[i][3] +  1*XTx[i][5];
-    }
-
-    // Store the transformed matrix
-    for (int i = 0, m = 0; i < 6; i++)
-    {
-      for (int j = 0; j < 6; j++, m++)
-      {
-        *(outptr + m*matrix_stride) = U[i][j];
-      }
-    }
-    outptr++;
-  }
-}
-
-/* In the below, unusual or especially small tiles are routed via the slow
- * path whereas common or large tiles are routed through a faster path.
- */
-template <>
-template <>
-const Transform::TileFn Transform::tile_fns[n_pad_top][n_pad_left][n_pad_bottom][n_pad_right] =
-{
-  {
-    {
-      {
-        Transform::template process_tile<0, 0, 0, 0>,  // No padding
-        Transform::template process_tile<0, 0, 0, 1>,  // Right
-        Transform::template process_tile<0, 0, 0, 2>,  // "   "
-        Transform::template process_tile<0, 0, 0, 3>,  // "   "
-        Transform::template process_tile<0, 0, 0, 4>,  // "   "
-      },
-      {
-        Transform::template process_tile<0, 0, 1, 0>,  // Bottom
-        Transform::template process_tile<0, 0, 1, 1>,  // Bottom right
-        Transform::template process_tile<0, 0, 1, 2>,  // "          "
-        Transform::template process_tile<0, 0, 1, 3>,  // "          "
-        Transform::template process_tile<0, 0, 1, 4>,  // "          "
-      },
-      {
-        Transform::template process_tile<0, 0, 2, 0>,  // Bottom
-        Transform::template process_tile<0, 0, 2, 1>,  // Bottom right
-        Transform::template process_tile<0, 0, 2, 2>,  // "          "
-        Transform::template process_tile<0, 0, 2, 3>,  // "          "
-        Transform::template process_tile<0, 0, 2, 4>,  // "          "
-      },
-      {
-        Transform::template process_tile<0, 0, 3, 0>,  // Bottom
-        Transform::template process_tile<0, 0, 3, 1>,  // Bottom right
-        Transform::template process_tile<0, 0, 3, 2>,  // "          "
-        Transform::template process_tile<0, 0, 3, 3>,  // "          "
-        Transform::template process_tile<0, 0, 3, 4>,  // "          "
-      },
-      {
-        Transform::template process_tile<0, 0, 4, 0>,  // Bottom
-        Transform::template process_tile<0, 0, 4, 1>,  // Bottom right
-        Transform::template process_tile<0, 0, 4, 2>,  // "          "
-        Transform::template process_tile<0, 0, 4, 3>,  // "          "
-        Transform::template process_tile<0, 0, 4, 4>,  // "          "
-      }
-    },
-    {
-      {
-        Transform::template process_tile<0, 1, 0, 0>,  // Left
-        Transform::template process_tile<0, 1, 0, 1>,
-        Transform::template process_tile<0, 1, 0, 2>,
-        Transform::template process_tile<0, 1, 0, 3>,
-        Transform::template process_tile<0, 1, 0, 4>,
-      },
-      {
-        Transform::template process_tile<0, 1, 1, 0>,  // Bottom left
-        Transform::template process_tile<0, 1, 1, 1>,
-        Transform::template process_tile<0, 1, 1, 2>,
-        Transform::template process_tile<0, 1, 1, 3>,
-        Transform::template process_tile<0, 1, 1, 4>,
-      },
-      {
-        Transform::template process_tile<0, 1, 2, 0>,  // "          "
-        Transform::template process_tile<0, 1, 2, 1>,
-        Transform::template process_tile<0, 1, 2, 2>,
-        Transform::template process_tile<0, 1, 2, 3>,
-        Transform::template process_tile<0, 1, 2, 4>,
-      },
-      {
-        Transform::template process_tile<0, 1, 3, 0>,  // "          "
-        Transform::template process_tile<0, 1, 3, 1>,
-        Transform::template process_tile<0, 1, 3, 2>,
-        Transform::template process_tile<0, 1, 3, 3>,
-        Transform::template process_tile<0, 1, 3, 4>,
-      },
-      {
-        Transform::template process_tile<0, 1, 4, 0>,  // "          "
-        Transform::template process_tile<0, 1, 4, 1>,
-        Transform::template process_tile<0, 1, 4, 2>,
-        Transform::template process_tile<0, 1, 4, 3>,
-        Transform::template process_tile<0, 1, 4, 4>,
-      }
-    }
-  },
-  {
-    {
-      {
-        Transform::template process_tile<1, 0, 0, 0>,  // Top
-        Transform::template process_tile<1, 0, 0, 1>,  // Top right
-        Transform::template process_tile<1, 0, 0, 2>,  // "       "
-        Transform::template process_tile<1, 0, 0, 3>,  // "       "
-        Transform::template process_tile<1, 0, 0, 4>,  // "       "
-      },
-      {
-        Transform::template process_tile<1, 0, 1, 0>,
-        Transform::template process_tile<1, 0, 1, 1>,
-        Transform::template process_tile<1, 0, 1, 2>,
-        Transform::template process_tile<1, 0, 1, 3>,
-        Transform::template process_tile<1, 0, 1, 4>,
-      },
-      {
-        Transform::template process_tile<1, 0, 2, 0>,
-        Transform::template process_tile<1, 0, 2, 1>,
-        Transform::template process_tile<1, 0, 2, 2>,
-        Transform::template process_tile<1, 0, 2, 3>,
-        Transform::template process_tile<1, 0, 2, 4>,
-      },
-      {
-        Transform::template process_tile<1, 0, 3, 0>,
-        Transform::template process_tile<1, 0, 3, 1>,
-        Transform::template process_tile<1, 0, 3, 2>,
-        Transform::template process_tile<1, 0, 3, 3>,
-        Transform::template process_tile<1, 0, 3, 4>,
-      },
-      {
-        Transform::template process_tile<1, 0, 4, 0>,
-        Transform::template process_tile<1, 0, 4, 1>,
-        Transform::template process_tile<1, 0, 4, 2>,
-        Transform::template process_tile<1, 0, 4, 3>,
-        Transform::template process_tile<1, 0, 4, 4>,
-      },
-    },
-    {
-      {
-        Transform::template process_tile<1, 1, 0, 0>,  // Top left
-        Transform::template process_tile<1, 1, 0, 1>,
-        Transform::template process_tile<1, 1, 0, 2>,
-        Transform::template process_tile<1, 1, 0, 3>,
-        Transform::template process_tile<1, 1, 0, 4>,
-      },
-      {
-        Transform::template process_tile<1, 1, 1, 0>,
-        Transform::template process_tile<1, 1, 1, 1>,
-        Transform::template process_tile<1, 1, 1, 2>,
-        Transform::template process_tile<1, 1, 1, 3>,
-        Transform::template process_tile<1, 1, 1, 4>,
-      },
-      {
-        Transform::template process_tile<1, 1, 2, 0>,
-        Transform::template process_tile<1, 1, 2, 1>,
-        Transform::template process_tile<1, 1, 2, 2>,
-        Transform::template process_tile<1, 1, 2, 3>,
-        Transform::template process_tile<1, 1, 2, 4>,
-      },
-      {
-        Transform::template process_tile<1, 1, 3, 0>,
-        Transform::template process_tile<1, 1, 3, 1>,
-        Transform::template process_tile<1, 1, 3, 2>,
-        Transform::template process_tile<1, 1, 3, 3>,
-        Transform::template process_tile<1, 1, 3, 4>,
-      },
-      {
-        Transform::template process_tile<1, 1, 4, 0>,
-        Transform::template process_tile<1, 1, 4, 1>,
-        Transform::template process_tile<1, 1, 4, 2>,
-        Transform::template process_tile<1, 1, 4, 3>,
-        Transform::template process_tile<1, 1, 4, 4>,
-      }
-    }
-  }
-};
-
-template struct WinogradGEMM<4, 4, 3, 3>::InputTransform<float>;
-}  // namespace winograd
diff --git a/src/core/NEON/kernels/convolution/winograd/transforms/input_6x6_fp32.cpp b/src/core/NEON/kernels/convolution/winograd/transforms/input_6x6_fp32.cpp
new file mode 100644
index 0000000000..908613068a
--- /dev/null
+++ b/src/core/NEON/kernels/convolution/winograd/transforms/input_6x6_fp32.cpp
@@ -0,0 +1,608 @@
+/*
+ * 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 "arm_compute/core/NEON/kernels/convolution/winograd/transforms/input.hpp"
+#include "arm_compute/core/NEON/kernels/convolution/winograd/winograd_gemm.hpp"
+#include "arm_compute/core/NEON/kernels/convolution/common/arm.hpp"
+
+namespace
+{
+
+template <int pad_top, int pad_left, int pad_bottom, int pad_right>
+void winograd_input_transform_6x6_fp32_process_tile(
+  int n_channels,
+  const float* const input_base,
+  const int input_row_stride,
+  const int input_col_stride,
+  float* const matrix_base,
+  const int matrix_stride
+)
+{
+  constexpr int inner_tile_rows = 6;
+  constexpr int inner_tile_cols = 6;
+  constexpr int cells_i = inner_tile_rows - pad_bottom;
+  constexpr int cells_j = inner_tile_cols - pad_right;
+
+  float *outptr = matrix_base;
+
+  // Get pointers into the input tile
+  const float *x_ptrs[inner_tile_rows][inner_tile_cols];
+  for (int i = pad_top, xi = 0; i < cells_i; i++, xi++)
+  {
+    // Get a pointer into the row
+    const float* const row_ptr = input_base + xi*input_row_stride;
+
+    for (int j = pad_left, xj = 0; j < cells_j; j++, xj++)
+    {
+      x_ptrs[i][j] = row_ptr + xj*input_col_stride;
+    }
+  }
+
+  // Matrices used/computed in this kernel.
+  float x[inner_tile_rows][inner_tile_cols];
+  float XTx[inner_tile_rows][inner_tile_cols];
+  float U[inner_tile_rows][inner_tile_cols];
+  for (int i = 0; i < inner_tile_rows; i++)
+  {
+    for (int j = 0; j < inner_tile_cols; j++)
+    {
+      x[i][j] = XTx[i][j] = 0.0f;
+    }
+  }
+
+  // Perform the Winograd input transformation for each channel in the input
+  // tensor.
+  int channels_remaining = n_channels;
+#ifdef __aarch64__
+  for (; channels_remaining >= 4; channels_remaining -= 4)
+  {
+    // Matrices used/computed in this kernel
+    float32x4_t x[inner_tile_rows][inner_tile_cols];
+    float32x4_t XTx[inner_tile_rows][inner_tile_cols];
+    float32x4_t U[inner_tile_rows][inner_tile_cols];
+    for (int i = 0; i < inner_tile_rows; i++)
+    {
+      for (int j = 0; j < inner_tile_cols; j++)
+      {
+        x[i][j] = vdupq_n_f32(0.0f);
+        XTx[i][j] = vdupq_n_f32(0.0f);
+      }
+    }
+
+    // Read a 6x6 tile in the Winograd domain
+    for (int i = pad_top; i < cells_i; i++)
+    {
+      for (int j = pad_left; j < cells_j; j++)
+      {
+        x[i][j] = vld1q_f32(x_ptrs[i][j]);
+        x_ptrs[i][j] += 4;
+      }
+    }
+
+    // Compute XT . x
+    for (int j = pad_left; j < cells_j; j++)
+    {
+      // XTx[0][j] =  4*x[0][j] + -5*x[2][j] +  1*x[4][j];
+      XTx[0][j] = vmlsq_n_f32(vmlaq_n_f32(x[4][j], x[0][j], 4.0f), x[2][j], 5.0f);
+
+      // XTx[1][j] = -4*x[1][j] + -4*x[2][j] +  1*x[3][j] +  1*x[4][j];
+      XTx[1][j] = vmlsq_n_f32(vaddq_f32(x[3][j], x[4][j]), vaddq_f32(x[1][j], x[2][j]), 4.0f);
+
+      // XTx[2][j] =  4*x[1][j] + -4*x[2][j] + -1*x[3][j] +  1*x[4][j];
+      XTx[2][j] = vmlaq_n_f32(vsubq_f32(x[4][j], x[3][j]), vsubq_f32(x[1][j], x[2][j]), 4.0f);
+
+      // XTx[3][j] = -2*x[1][j] + -1*x[2][j] +  2*x[3][j] +  1*x[4][j];
+      XTx[3][j] = vmlaq_n_f32(vsubq_f32(x[4][j], x[2][j]), vsubq_f32(x[3][j], x[1][j]), 2.0f);
+
+      // XTx[4][j] =  2*x[1][j] + -1*x[2][j] + -2*x[3][j] +  1*x[4][j];
+      XTx[4][j] = vmlaq_n_f32(vsubq_f32(x[4][j], x[2][j]), vsubq_f32(x[1][j], x[3][j]), 2.0f);
+
+      // XTx[5][j] =  4*x[1][j] + -5*x[3][j] +  1*x[5][j];
+      XTx[5][j] = vmlsq_n_f32(vmlaq_n_f32(x[5][j], x[1][j], 4.0f), x[3][j], 5.0f);
+    }
+
+    // Compute U = XT . x . X
+    for (int i = 0; i < inner_tile_rows; i++)
+    {
+      // U[i][0] =  4*XTx[i][0] + -5*XTx[i][2] +  1*XTx[i][4];
+      U[i][0] = vmlsq_n_f32(vmlaq_n_f32(XTx[i][4], XTx[i][0], 4.0f), XTx[i][2], 5.0f);
+
+      // U[i][1] = -4*XTx[i][1] + -4*XTx[i][2] +  1*XTx[i][3] +  1*XTx[i][4];
+      U[i][1] = vmlsq_n_f32(vaddq_f32(XTx[i][3], XTx[i][4]), vaddq_f32(XTx[i][1], XTx[i][2]), 4.0f);
+
+      // U[i][2] =  4*XTx[i][1] + -4*XTx[i][2] + -1*XTx[i][3] +  1*XTx[i][4];
+      U[i][2] = vmlaq_n_f32(vsubq_f32(XTx[i][4], XTx[i][3]), vsubq_f32(XTx[i][1], XTx[i][2]), 4.0f);
+
+      // U[i][3] = -2*XTx[i][1] + -1*XTx[i][2] +  2*XTx[i][3] +  1*XTx[i][4];
+      U[i][3] = vmlaq_n_f32(vsubq_f32(XTx[i][4], XTx[i][2]), vsubq_f32(XTx[i][3], XTx[i][1]), 2.0f);
+
+      // U[i][4] =  2*XTx[i][1] + -1*XTx[i][2] + -2*XTx[i][3] +  1*XTx[i][4];
+      U[i][4] = vmlaq_n_f32(vsubq_f32(XTx[i][4], XTx[i][2]), vsubq_f32(XTx[i][1], XTx[i][3]), 2.0f);
+
+      // U[i][5] =  4*XTx[i][1] + -5*XTx[i][3] +  1*XTx[i][5];
+      U[i][5] = vmlsq_n_f32(vmlaq_n_f32(XTx[i][5], XTx[i][1], 4.0f), XTx[i][3], 5.0f);
+    }
+
+    // Store the transformed matrix
+    for (int i = 0, m = 0; i < inner_tile_rows; i++)
+    {
+      for (int j = 0; j < inner_tile_cols; j++, m++)
+      {
+        vst1q_f32(outptr + m*matrix_stride, U[i][j]);
+      }
+    }
+    outptr += 4;
+  }
+#endif  // __aarch64__
+#ifdef __arm_any__
+  for (; channels_remaining >= 2; channels_remaining -= 2)
+  {
+    // Matrices used/computed in this kernel
+    float32x2_t x[inner_tile_rows][inner_tile_cols];
+    float32x2_t XTx[inner_tile_rows][inner_tile_cols];
+    float32x2_t U[inner_tile_rows][inner_tile_cols];
+    for (int i = 0; i < inner_tile_rows; i++)
+    {
+      for (int j = 0; j < inner_tile_cols; j++)
+      {
+        x[i][j] = vdup_n_f32(0.0f);
+        XTx[i][j] = vdup_n_f32(0.0f);
+      }
+    }
+
+    // Read a 6x6 tile in the Winograd domain
+    for (int i = pad_top; i < cells_i; i++)
+    {
+      for (int j = pad_left; j < cells_j; j++)
+      {
+        x[i][j] = vld1_f32(x_ptrs[i][j]);
+        x_ptrs[i][j] += 2;
+      }
+    }
+
+    // Compute XT . x
+    for (int j = pad_left; j < cells_j; j++)
+    {
+      // XTx[0][j] =  4*x[0][j] + -5*x[2][j] +  1*x[4][j];
+      XTx[0][j] = vmls_n_f32(vmla_n_f32(x[4][j], x[0][j], 4.0f), x[2][j], 5.0f);
+
+      // XTx[1][j] = -4*x[1][j] + -4*x[2][j] +  1*x[3][j] +  1*x[4][j];
+      XTx[1][j] = vmls_n_f32(vadd_f32(x[3][j], x[4][j]), vadd_f32(x[1][j], x[2][j]), 4.0f);
+
+      // XTx[2][j] =  4*x[1][j] + -4*x[2][j] + -1*x[3][j] +  1*x[4][j];
+      XTx[2][j] = vmla_n_f32(vsub_f32(x[4][j], x[3][j]), vsub_f32(x[1][j], x[2][j]), 4.0f);
+
+      // XTx[3][j] = -2*x[1][j] + -1*x[2][j] +  2*x[3][j] +  1*x[4][j];
+      XTx[3][j] = vmla_n_f32(vsub_f32(x[4][j], x[2][j]), vsub_f32(x[3][j], x[1][j]), 2.0f);
+
+      // XTx[4][j] =  2*x[1][j] + -1*x[2][j] + -2*x[3][j] +  1*x[4][j];
+      XTx[4][j] = vmla_n_f32(vsub_f32(x[4][j], x[2][j]), vsub_f32(x[1][j], x[3][j]), 2.0f);
+
+      // XTx[5][j] =  4*x[1][j] + -5*x[3][j] +  1*x[5][j];
+      XTx[5][j] = vmls_n_f32(vmla_n_f32(x[5][j], x[1][j], 4.0f), x[3][j], 5.0f);
+    }
+
+    // Compute U = XT . x . X
+    for (int i = 0; i < inner_tile_rows; i++)
+    {
+      // U[i][0] =  4*XTx[i][0] + -5*XTx[i][2] +  1*XTx[i][4];
+      U[i][0] = vmls_n_f32(vmla_n_f32(XTx[i][4], XTx[i][0], 4.0f), XTx[i][2], 5.0f);
+
+      // U[i][1] = -4*XTx[i][1] + -4*XTx[i][2] +  1*XTx[i][3] +  1*XTx[i][4];
+      U[i][1] = vmls_n_f32(vadd_f32(XTx[i][3], XTx[i][4]), vadd_f32(XTx[i][1], XTx[i][2]), 4.0f);
+
+      // U[i][2] =  4*XTx[i][1] + -4*XTx[i][2] + -1*XTx[i][3] +  1*XTx[i][4];
+      U[i][2] = vmla_n_f32(vsub_f32(XTx[i][4], XTx[i][3]), vsub_f32(XTx[i][1], XTx[i][2]), 4.0f);
+
+      // U[i][3] = -2*XTx[i][1] + -1*XTx[i][2] +  2*XTx[i][3] +  1*XTx[i][4];
+      U[i][3] = vmla_n_f32(vsub_f32(XTx[i][4], XTx[i][2]), vsub_f32(XTx[i][3], XTx[i][1]), 2.0f);
+
+      // U[i][4] =  2*XTx[i][1] + -1*XTx[i][2] + -2*XTx[i][3] +  1*XTx[i][4];
+      U[i][4] = vmla_n_f32(vsub_f32(XTx[i][4], XTx[i][2]), vsub_f32(XTx[i][1], XTx[i][3]), 2.0f);
+
+      // U[i][5] =  4*XTx[i][1] + -5*XTx[i][3] +  1*XTx[i][5];
+      U[i][5] = vmls_n_f32(vmla_n_f32(XTx[i][5], XTx[i][1], 4.0f), XTx[i][3], 5.0f);
+    }
+
+    // Store the transformed matrix
+    for (int i = 0, m = 0; i < inner_tile_rows; i++)
+    {
+      for (int j = 0; j < inner_tile_cols; j++, m++)
+      {
+        vst1_f32(outptr + m*matrix_stride, U[i][j]);
+      }
+    }
+    outptr += 2;
+  }
+#endif  // __arm_any__
+  for (; channels_remaining; channels_remaining--)
+  {
+    // Load x
+    for (int i = pad_top; i < cells_i; i++)
+    {
+      for (int j = pad_left; j < cells_j; j++)
+      {
+        x[i][j] = *(x_ptrs[i][j]++);
+      }
+    }
+
+    // Compute XT . x
+    for (int j = pad_left; j < cells_j; j++)
+    {
+      XTx[0][j] =  4*x[0][j] + -5*x[2][j] +  1*x[4][j];
+      XTx[1][j] = -4*x[1][j] + -4*x[2][j] +  1*x[3][j] +  1*x[4][j];
+      XTx[2][j] =  4*x[1][j] + -4*x[2][j] + -1*x[3][j] +  1*x[4][j];
+      XTx[3][j] = -2*x[1][j] + -1*x[2][j] +  2*x[3][j] +  1*x[4][j];
+      XTx[4][j] =  2*x[1][j] + -1*x[2][j] + -2*x[3][j] +  1*x[4][j];
+      XTx[5][j] =  4*x[1][j] + -5*x[3][j] +  1*x[5][j];
+    }
+
+    // Compute U = XT . x . X
+    for (int i = 0; i < inner_tile_rows; i++)
+    {
+      U[i][0] =  4*XTx[i][0] + -5*XTx[i][2] +  1*XTx[i][4];
+      U[i][1] = -4*XTx[i][1] + -4*XTx[i][2] +  1*XTx[i][3] +  1*XTx[i][4];
+      U[i][2] =  4*XTx[i][1] + -4*XTx[i][2] + -1*XTx[i][3] +  1*XTx[i][4];
+      U[i][3] = -2*XTx[i][1] + -1*XTx[i][2] +  2*XTx[i][3] +  1*XTx[i][4];
+      U[i][4] =  2*XTx[i][1] + -1*XTx[i][2] + -2*XTx[i][3] +  1*XTx[i][4];
+      U[i][5] =  4*XTx[i][1] + -5*XTx[i][3] +  1*XTx[i][5];
+    }
+
+    // Store the transformed matrix
+    for (int i = 0, m = 0; i < inner_tile_rows; i++)
+    {
+      for (int j = 0; j < inner_tile_cols; j++, m++)
+      {
+        *(outptr + m*matrix_stride) = U[i][j];
+      }
+    }
+    outptr++;
+  }
+}
+}
+
+namespace winograd
+{
+template <int k>
+using Transform = InputTransformImpl<k, k, 6, 6, float>;
+
+template <>
+const Transform<3>::TileFn
+  Transform<3>::tile_fns[n_pad_top][n_pad_left][n_pad_bottom][n_pad_right] =
+{
+  {
+    {
+      {
+        winograd_input_transform_6x6_fp32_process_tile<0, 0, 0, 0>,  // No padding
+        winograd_input_transform_6x6_fp32_process_tile<0, 0, 0, 1>,  // Right
+        winograd_input_transform_6x6_fp32_process_tile<0, 0, 0, 2>,  // "   "
+        winograd_input_transform_6x6_fp32_process_tile<0, 0, 0, 3>,  // "   "
+        winograd_input_transform_6x6_fp32_process_tile<0, 0, 0, 4>,  // "   "
+      },
+      {
+        winograd_input_transform_6x6_fp32_process_tile<0, 0, 1, 0>,  // Bottom
+        winograd_input_transform_6x6_fp32_process_tile<0, 0, 1, 1>,  // Bottom right
+        winograd_input_transform_6x6_fp32_process_tile<0, 0, 1, 2>,  // "          "
+        winograd_input_transform_6x6_fp32_process_tile<0, 0, 1, 3>,  // "          "
+        winograd_input_transform_6x6_fp32_process_tile<0, 0, 1, 4>,  // "          "
+      },
+      {
+        winograd_input_transform_6x6_fp32_process_tile<0, 0, 2, 0>,  // Bottom
+        winograd_input_transform_6x6_fp32_process_tile<0, 0, 2, 1>,  // Bottom right
+        winograd_input_transform_6x6_fp32_process_tile<0, 0, 2, 2>,  // "          "
+        winograd_input_transform_6x6_fp32_process_tile<0, 0, 2, 3>,  // "          "
+        winograd_input_transform_6x6_fp32_process_tile<0, 0, 2, 4>,  // "          "
+      },
+      {
+        winograd_input_transform_6x6_fp32_process_tile<0, 0, 3, 0>,  // Bottom
+        winograd_input_transform_6x6_fp32_process_tile<0, 0, 3, 1>,  // Bottom right
+        winograd_input_transform_6x6_fp32_process_tile<0, 0, 3, 2>,  // "          "
+        winograd_input_transform_6x6_fp32_process_tile<0, 0, 3, 3>,  // "          "
+        winograd_input_transform_6x6_fp32_process_tile<0, 0, 3, 4>,  // "          "
+      },
+      {
+        winograd_input_transform_6x6_fp32_process_tile<0, 0, 4, 0>,  // Bottom
+        winograd_input_transform_6x6_fp32_process_tile<0, 0, 4, 1>,  // Bottom right
+        winograd_input_transform_6x6_fp32_process_tile<0, 0, 4, 2>,  // "          "
+        winograd_input_transform_6x6_fp32_process_tile<0, 0, 4, 3>,  // "          "
+        winograd_input_transform_6x6_fp32_process_tile<0, 0, 4, 4>,  // "          "
+      }
+    },
+    {
+      {
+        winograd_input_transform_6x6_fp32_process_tile<0, 1, 0, 0>,  // Left
+        winograd_input_transform_6x6_fp32_process_tile<0, 1, 0, 1>,
+        winograd_input_transform_6x6_fp32_process_tile<0, 1, 0, 2>,
+        winograd_input_transform_6x6_fp32_process_tile<0, 1, 0, 3>,
+        winograd_input_transform_6x6_fp32_process_tile<0, 1, 0, 4>,
+      },
+      {
+        winograd_input_transform_6x6_fp32_process_tile<0, 1, 1, 0>,  // Bottom left
+        winograd_input_transform_6x6_fp32_process_tile<0, 1, 1, 1>,
+        winograd_input_transform_6x6_fp32_process_tile<0, 1, 1, 2>,
+        winograd_input_transform_6x6_fp32_process_tile<0, 1, 1, 3>,
+        winograd_input_transform_6x6_fp32_process_tile<0, 1, 1, 4>,
+      },
+      {
+        winograd_input_transform_6x6_fp32_process_tile<0, 1, 2, 0>,  // "          "
+        winograd_input_transform_6x6_fp32_process_tile<0, 1, 2, 1>,
+        winograd_input_transform_6x6_fp32_process_tile<0, 1, 2, 2>,
+        winograd_input_transform_6x6_fp32_process_tile<0, 1, 2, 3>,
+        winograd_input_transform_6x6_fp32_process_tile<0, 1, 2, 4>,
+      },
+      {
+        winograd_input_transform_6x6_fp32_process_tile<0, 1, 3, 0>,  // "          "
+        winograd_input_transform_6x6_fp32_process_tile<0, 1, 3, 1>,
+        winograd_input_transform_6x6_fp32_process_tile<0, 1, 3, 2>,
+        winograd_input_transform_6x6_fp32_process_tile<0, 1, 3, 3>,
+        winograd_input_transform_6x6_fp32_process_tile<0, 1, 3, 4>,
+      },
+      {
+        winograd_input_transform_6x6_fp32_process_tile<0, 1, 4, 0>,  // "          "
+        winograd_input_transform_6x6_fp32_process_tile<0, 1, 4, 1>,
+        winograd_input_transform_6x6_fp32_process_tile<0, 1, 4, 2>,
+        winograd_input_transform_6x6_fp32_process_tile<0, 1, 4, 3>,
+        winograd_input_transform_6x6_fp32_process_tile<0, 1, 4, 4>,
+      }
+    }
+  },
+  {
+    {
+      {
+        winograd_input_transform_6x6_fp32_process_tile<1, 0, 0, 0>,  // Top
+        winograd_input_transform_6x6_fp32_process_tile<1, 0, 0, 1>,  // Top right
+        winograd_input_transform_6x6_fp32_process_tile<1, 0, 0, 2>,  // "       "
+        winograd_input_transform_6x6_fp32_process_tile<1, 0, 0, 3>,  // "       "
+        winograd_input_transform_6x6_fp32_process_tile<1, 0, 0, 4>,  // "       "
+      },
+      {
+        winograd_input_transform_6x6_fp32_process_tile<1, 0, 1, 0>,
+        winograd_input_transform_6x6_fp32_process_tile<1, 0, 1, 1>,
+        winograd_input_transform_6x6_fp32_process_tile<1, 0, 1, 2>,
+        winograd_input_transform_6x6_fp32_process_tile<1, 0, 1, 3>,
+        winograd_input_transform_6x6_fp32_process_tile<1, 0, 1, 4>,
+      },
+      {
+        winograd_input_transform_6x6_fp32_process_tile<1, 0, 2, 0>,
+        winograd_input_transform_6x6_fp32_process_tile<1, 0, 2, 1>,
+        winograd_input_transform_6x6_fp32_process_tile<1, 0, 2, 2>,
+        winograd_input_transform_6x6_fp32_process_tile<1, 0, 2, 3>,
+        winograd_input_transform_6x6_fp32_process_tile<1, 0, 2, 4>,
+      },
+      {
+        winograd_input_transform_6x6_fp32_process_tile<1, 0, 3, 0>,
+        winograd_input_transform_6x6_fp32_process_tile<1, 0, 3, 1>,
+        winograd_input_transform_6x6_fp32_process_tile<1, 0, 3, 2>,
+        winograd_input_transform_6x6_fp32_process_tile<1, 0, 3, 3>,
+        winograd_input_transform_6x6_fp32_process_tile<1, 0, 3, 4>,
+      },
+      {
+        winograd_input_transform_6x6_fp32_process_tile<1, 0, 4, 0>,
+        winograd_input_transform_6x6_fp32_process_tile<1, 0, 4, 1>,
+        winograd_input_transform_6x6_fp32_process_tile<1, 0, 4, 2>,
+        winograd_input_transform_6x6_fp32_process_tile<1, 0, 4, 3>,
+        winograd_input_transform_6x6_fp32_process_tile<1, 0, 4, 4>,
+      },
+    },
+    {
+      {
+        winograd_input_transform_6x6_fp32_process_tile<1, 1, 0, 0>,  // Top left
+        winograd_input_transform_6x6_fp32_process_tile<1, 1, 0, 1>,
+        winograd_input_transform_6x6_fp32_process_tile<1, 1, 0, 2>,
+        winograd_input_transform_6x6_fp32_process_tile<1, 1, 0, 3>,
+        winograd_input_transform_6x6_fp32_process_tile<1, 1, 0, 4>,
+      },
+      {
+        winograd_input_transform_6x6_fp32_process_tile<1, 1, 1, 0>,
+        winograd_input_transform_6x6_fp32_process_tile<1, 1, 1, 1>,
+        winograd_input_transform_6x6_fp32_process_tile<1, 1, 1, 2>,
+        winograd_input_transform_6x6_fp32_process_tile<1, 1, 1, 3>,
+        winograd_input_transform_6x6_fp32_process_tile<1, 1, 1, 4>,
+      },
+      {
+        winograd_input_transform_6x6_fp32_process_tile<1, 1, 2, 0>,
+        winograd_input_transform_6x6_fp32_process_tile<1, 1, 2, 1>,
+        winograd_input_transform_6x6_fp32_process_tile<1, 1, 2, 2>,
+        winograd_input_transform_6x6_fp32_process_tile<1, 1, 2, 3>,
+        winograd_input_transform_6x6_fp32_process_tile<1, 1, 2, 4>,
+      },
+      {
+        winograd_input_transform_6x6_fp32_process_tile<1, 1, 3, 0>,
+        winograd_input_transform_6x6_fp32_process_tile<1, 1, 3, 1>,
+        winograd_input_transform_6x6_fp32_process_tile<1, 1, 3, 2>,
+        winograd_input_transform_6x6_fp32_process_tile<1, 1, 3, 3>,
+        winograd_input_transform_6x6_fp32_process_tile<1, 1, 3, 4>,
+      },
+      {
+        winograd_input_transform_6x6_fp32_process_tile<1, 1, 4, 0>,
+        winograd_input_transform_6x6_fp32_process_tile<1, 1, 4, 1>,
+        winograd_input_transform_6x6_fp32_process_tile<1, 1, 4, 2>,
+        winograd_input_transform_6x6_fp32_process_tile<1, 1, 4, 3>,
+        winograd_input_transform_6x6_fp32_process_tile<1, 1, 4, 4>,
+      }
+    }
+  }
+};
+
+template <>
+const Transform<5>::TileFn
+  Transform<5>::tile_fns[n_pad_top][n_pad_left][n_pad_bottom][n_pad_right] =
+{
+  {
+    {
+      {
+        winograd_input_transform_6x6_fp32_process_tile<0, 0, 0, 0>,  // No padding
+        winograd_input_transform_6x6_fp32_process_tile<0, 0, 0, 1>,  // Right
+        winograd_input_transform_6x6_fp32_process_tile<0, 0, 0, 2>,  // "   "
+        winograd_input_transform_6x6_fp32_process_tile<0, 0, 0, 3>,  // "   "
+        winograd_input_transform_6x6_fp32_process_tile<0, 0, 0, 4>,  // "   "
+      },
+      {
+        winograd_input_transform_6x6_fp32_process_tile<0, 0, 1, 0>,  // Bottom
+        winograd_input_transform_6x6_fp32_process_tile<0, 0, 1, 1>,  // Bottom right
+        winograd_input_transform_6x6_fp32_process_tile<0, 0, 1, 2>,  // "          "
+        winograd_input_transform_6x6_fp32_process_tile<0, 0, 1, 3>,  // "          "
+        winograd_input_transform_6x6_fp32_process_tile<0, 0, 1, 4>,  // "          "
+      },
+      {
+        winograd_input_transform_6x6_fp32_process_tile<0, 0, 2, 0>,  // Bottom
+        winograd_input_transform_6x6_fp32_process_tile<0, 0, 2, 1>,  // Bottom right
+        winograd_input_transform_6x6_fp32_process_tile<0, 0, 2, 2>,  // "          "
+        winograd_input_transform_6x6_fp32_process_tile<0, 0, 2, 3>,  // "          "
+        winograd_input_transform_6x6_fp32_process_tile<0, 0, 2, 4>,  // "          "
+      },
+      {
+        winograd_input_transform_6x6_fp32_process_tile<0, 0, 3, 0>,  // Bottom
+        winograd_input_transform_6x6_fp32_process_tile<0, 0, 3, 1>,  // Bottom right
+        winograd_input_transform_6x6_fp32_process_tile<0, 0, 3, 2>,  // "          "
+        winograd_input_transform_6x6_fp32_process_tile<0, 0, 3, 3>,  // "          "
+        winograd_input_transform_6x6_fp32_process_tile<0, 0, 3, 4>,  // "          "
+      },
+      {
+        winograd_input_transform_6x6_fp32_process_tile<0, 0, 4, 0>,  // Bottom
+        winograd_input_transform_6x6_fp32_process_tile<0, 0, 4, 1>,  // Bottom right
+        winograd_input_transform_6x6_fp32_process_tile<0, 0, 4, 2>,  // "          "
+        winograd_input_transform_6x6_fp32_process_tile<0, 0, 4, 3>,  // "          "
+        winograd_input_transform_6x6_fp32_process_tile<0, 0, 4, 4>,  // "          "
+      }
+    },
+    {
+      {
+        winograd_input_transform_6x6_fp32_process_tile<0, 2, 0, 0>,  // Left
+        winograd_input_transform_6x6_fp32_process_tile<0, 2, 0, 1>,
+        winograd_input_transform_6x6_fp32_process_tile<0, 2, 0, 2>,
+        winograd_input_transform_6x6_fp32_process_tile<0, 2, 0, 3>,
+        winograd_input_transform_6x6_fp32_process_tile<0, 2, 0, 4>,
+      },
+      {
+        winograd_input_transform_6x6_fp32_process_tile<0, 2, 1, 0>,  // Bottom left
+        winograd_input_transform_6x6_fp32_process_tile<0, 2, 1, 1>,
+        winograd_input_transform_6x6_fp32_process_tile<0, 2, 1, 2>,
+        winograd_input_transform_6x6_fp32_process_tile<0, 2, 1, 3>,
+        winograd_input_transform_6x6_fp32_process_tile<0, 2, 1, 4>,
+      },
+      {
+        winograd_input_transform_6x6_fp32_process_tile<0, 2, 2, 0>,  // "          "
+        winograd_input_transform_6x6_fp32_process_tile<0, 2, 2, 1>,
+        winograd_input_transform_6x6_fp32_process_tile<0, 2, 2, 2>,
+        winograd_input_transform_6x6_fp32_process_tile<0, 2, 2, 3>,
+        winograd_input_transform_6x6_fp32_process_tile<0, 2, 2, 4>,
+      },
+      {
+        winograd_input_transform_6x6_fp32_process_tile<0, 2, 3, 0>,  // "          "
+        winograd_input_transform_6x6_fp32_process_tile<0, 2, 3, 1>,
+        winograd_input_transform_6x6_fp32_process_tile<0, 2, 3, 2>,
+        winograd_input_transform_6x6_fp32_process_tile<0, 2, 3, 3>,
+        winograd_input_transform_6x6_fp32_process_tile<0, 2, 3, 4>,
+      },
+      {
+        winograd_input_transform_6x6_fp32_process_tile<0, 2, 4, 0>,  // "          "
+        winograd_input_transform_6x6_fp32_process_tile<0, 2, 4, 1>,
+        winograd_input_transform_6x6_fp32_process_tile<0, 2, 4, 2>,
+        winograd_input_transform_6x6_fp32_process_tile<0, 2, 4, 3>,
+        winograd_input_transform_6x6_fp32_process_tile<0, 2, 4, 4>,
+      }
+    }
+  },
+  {
+    {
+      {
+        winograd_input_transform_6x6_fp32_process_tile<2, 0, 0, 0>,  // Top
+        winograd_input_transform_6x6_fp32_process_tile<2, 0, 0, 1>,  // Top right
+        winograd_input_transform_6x6_fp32_process_tile<2, 0, 0, 2>,  // "       "
+        winograd_input_transform_6x6_fp32_process_tile<2, 0, 0, 3>,  // "       "
+        winograd_input_transform_6x6_fp32_process_tile<2, 0, 0, 4>,  // "       "
+      },
+      {
+        winograd_input_transform_6x6_fp32_process_tile<2, 0, 1, 0>,
+        winograd_input_transform_6x6_fp32_process_tile<2, 0, 1, 1>,
+        winograd_input_transform_6x6_fp32_process_tile<2, 0, 1, 2>,
+        winograd_input_transform_6x6_fp32_process_tile<2, 0, 1, 3>,
+        winograd_input_transform_6x6_fp32_process_tile<2, 0, 1, 4>,
+      },
+      {
+        winograd_input_transform_6x6_fp32_process_tile<2, 0, 2, 0>,
+        winograd_input_transform_6x6_fp32_process_tile<2, 0, 2, 1>,
+        winograd_input_transform_6x6_fp32_process_tile<2, 0, 2, 2>,
+        winograd_input_transform_6x6_fp32_process_tile<2, 0, 2, 3>,
+        winograd_input_transform_6x6_fp32_process_tile<2, 0, 2, 4>,
+      },
+      {
+        winograd_input_transform_6x6_fp32_process_tile<2, 0, 3, 0>,
+        winograd_input_transform_6x6_fp32_process_tile<2, 0, 3, 1>,
+        winograd_input_transform_6x6_fp32_process_tile<2, 0, 3, 2>,
+        winograd_input_transform_6x6_fp32_process_tile<2, 0, 3, 3>,
+        winograd_input_transform_6x6_fp32_process_tile<2, 0, 3, 4>,
+      },
+      {
+        winograd_input_transform_6x6_fp32_process_tile<2, 0, 4, 0>,
+        winograd_input_transform_6x6_fp32_process_tile<2, 0, 4, 1>,
+        winograd_input_transform_6x6_fp32_process_tile<2, 0, 4, 2>,
+        winograd_input_transform_6x6_fp32_process_tile<2, 0, 4, 3>,
+        winograd_input_transform_6x6_fp32_process_tile<2, 0, 4, 4>,
+      },
+    },
+    {
+      {
+        winograd_input_transform_6x6_fp32_process_tile<2, 2, 0, 0>,  // Top left
+        winograd_input_transform_6x6_fp32_process_tile<2, 2, 0, 1>,
+        winograd_input_transform_6x6_fp32_process_tile<2, 2, 0, 2>,
+        winograd_input_transform_6x6_fp32_process_tile<2, 2, 0, 3>,
+        winograd_input_transform_6x6_fp32_process_tile<2, 2, 0, 4>,
+      },
+      {
+        winograd_input_transform_6x6_fp32_process_tile<2, 2, 1, 0>,
+        winograd_input_transform_6x6_fp32_process_tile<2, 2, 1, 1>,
+        winograd_input_transform_6x6_fp32_process_tile<2, 2, 1, 2>,
+        winograd_input_transform_6x6_fp32_process_tile<2, 2, 1, 3>,
+        winograd_input_transform_6x6_fp32_process_tile<2, 2, 1, 4>,
+      },
+      {
+        winograd_input_transform_6x6_fp32_process_tile<2, 2, 2, 0>,
+        winograd_input_transform_6x6_fp32_process_tile<2, 2, 2, 1>,
+        winograd_input_transform_6x6_fp32_process_tile<2, 2, 2, 2>,
+        winograd_input_transform_6x6_fp32_process_tile<2, 2, 2, 3>,
+        winograd_input_transform_6x6_fp32_process_tile<2, 2, 2, 4>,
+      },
+      {
+        winograd_input_transform_6x6_fp32_process_tile<2, 2, 3, 0>,
+        winograd_input_transform_6x6_fp32_process_tile<2, 2, 3, 1>,
+        winograd_input_transform_6x6_fp32_process_tile<2, 2, 3, 2>,
+        winograd_input_transform_6x6_fp32_process_tile<2, 2, 3, 3>,
+        winograd_input_transform_6x6_fp32_process_tile<2, 2, 3, 4>,
+      },
+      {
+        winograd_input_transform_6x6_fp32_process_tile<2, 2, 4, 0>,
+        winograd_input_transform_6x6_fp32_process_tile<2, 2, 4, 1>,
+        winograd_input_transform_6x6_fp32_process_tile<2, 2, 4, 2>,
+        winograd_input_transform_6x6_fp32_process_tile<2, 2, 4, 3>,
+        winograd_input_transform_6x6_fp32_process_tile<2, 2, 4, 4>,
+      }
+    }
+  }
+};
+
+template class InputTransform<3, 3, 6, 6, float>;
+template class InputTransform<5, 5, 6, 6, float>;
+}