COMPMID-1600: Reduce number of tile specialisations.

Change-Id: I4d06eca9404ea6d3df9d0ca52f5d6f5421ab7116
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/150117
Tested-by: bsgcomp <bsgcomp@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>

5 files changed, 419 insertions, 576 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 473a13c3b0..b813bbb25c 100644
--- a/arm_compute/core/NEON/kernels/convolution/winograd/transforms/input.hpp
+++ b/arm_compute/core/NEON/kernels/convolution/winograd/transforms/input.hpp
@@ -155,15 +155,15 @@ namespace winograd
       T* const outptr = matrix_base + tile_j*matrix_row_stride;
 
       // Apply the specific tile processing function
-      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](
+      const typename Tiles::TileFn tilefn = Tiles::get_tile_specialization(
+        pad_top, t_pad_left, pad_bottom, t_pad_right
+      );
+
+      tilefn(
         n_channels,
-        input_base_col,
-        input_row_stride,
-        input_col_stride,
-        outptr,
-        matrix_stride
+        input_base_col, input_row_stride, input_col_stride,
+        outptr, matrix_stride,
+        pad_top, t_pad_left, pad_bottom, t_pad_right
       );
     }
   }
@@ -264,4 +264,86 @@ namespace winograd
       matrix_stride, matrix_batch_stride, matrix_row_stride
     );
   }
+
+  template <int KernelRows, int KernelCols, int InnerTileRows, int InnerTileCols, typename T>
+  typename InputTransformImplTiles<KernelRows, KernelCols, InnerTileRows, InnerTileCols, T>::TileFn
+    InputTransformImplTiles<KernelRows, KernelCols, InnerTileRows, InnerTileCols, T>::
+      get_tile_specialization(
+        const int pad_top,
+        const int pad_left,
+        const int pad_bottom,
+        const int pad_right
+      )
+  {
+    if (!(pad_top || pad_left || pad_bottom || pad_right))
+    {
+      // No padding, return unpadded specialisation
+      return tilefn_unpadded;
+    }
+    else if (pad_top && !(pad_left || pad_bottom || pad_right))
+    {
+      // Top padding only
+      const int index = (pad_top - min_pad_top) / (InnerTileRows - overlap_rows);
+      return tilefn_top_padded[index];
+    }
+    else if (!(pad_top) && pad_left && !(pad_bottom || pad_right))
+    {
+      // Left padding only
+      const int index = (pad_left - min_pad_left) / (InnerTileCols - overlap_cols);
+      return tilefn_left_padded[index];
+    }
+    else if (!(pad_top || pad_left) && pad_bottom && !(pad_right))
+    {
+      // Bottom padding only
+      return tilefn_bottom_padded[pad_bottom - 1];
+    }
+    else if (!(pad_top || pad_left || pad_bottom) && pad_right)
+    {
+      // Right padding only
+      return tilefn_right_padded[pad_right - 1];
+    }
+    else
+    {
+      // Combination of paddings, return an unspecialised method
+      return tilefn_generic;
+    }
+  }
+
+  template <int KernelCols, int InnerTileCols, typename T>
+  typename InputTransformImplTiles<1, KernelCols, 1, InnerTileCols, T>::TileFn
+    InputTransformImplTiles<1, KernelCols, 1, InnerTileCols, T>::
+      get_tile_specialization(
+        const int pad_top,
+        const int pad_left,
+        const int pad_bottom,
+        const int pad_right
+      )
+  {
+    (void) pad_top;
+    (void) pad_bottom;
+
+    if (!(pad_left || pad_right))
+    {
+      // No padding, return unpadded specialisation
+      return tilefn_unpadded;
+    }
+    else if (pad_left && !pad_right)
+    {
+      // Left padding only
+      const int index = (pad_left - min_pad_left) / (InnerTileCols - overlap_cols);
+      return tilefn_left_padded[index];
+    }
+    else if (!pad_left && pad_right)
+    {
+      // Right padding only
+      return tilefn_right_padded[pad_right - 1];
+    }
+    else
+    {
+      // Combination of paddings, return an unspecialised method
+      return tilefn_generic;
+    }
+  }
 }
+
+
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
index abcda53534..995554d7f2 100644
--- a/arm_compute/core/NEON/kernels/convolution/winograd/winograd_input_transform.hpp
+++ b/arm_compute/core/NEON/kernels/convolution/winograd/winograd_input_transform.hpp
@@ -31,6 +31,109 @@ namespace
 {
 
 template <int KernelRows, int KernelCols, int InnerTileRows, int InnerTileCols, typename T>
+class InputTransformImplTiles
+{
+  public:
+    /** Method to transform a tile of the input tensor into the Winograd domain. */
+    typedef void (*TileFn)(
+      const int n_channels,        /** @param[in] Number of channels in the tensor. */
+      const T* const inptr_base,   /** @param[in] Pointer to the base of the input tile. */
+      const int input_row_stride,  /** @param[in] Stride between rows of the input tensor. */
+      const int input_col_stride,  /** @param[in] Stride between columns of the input tensor. */
+      T* const mptr_base,          /** @param[out] Base pointer to transformed input matrices. */
+      const int matrix_stride,     /** @param[in] Stride between matrices in the input space. */
+      const int _pad_top,          /** @param[in] Top padding for unspecialised tiles. */
+      const int _pad_left,         /** @param[in] Left padding for unspecialised tiles. */
+      const int _pad_bottom,       /** @param[in] Bottom padding for unspecialised tiles. */
+      const int _pad_right         /** @param[in] Right padding for unspecialised tiles. */
+    );
+
+    static TileFn get_tile_specialization(
+      const int pad_top,
+      const int pad_left,
+      const int pad_bottom,
+      const int pad_right
+    );
+
+    // Tile overlaps
+    static constexpr int overlap_rows = KernelRows - 1;
+    static constexpr int overlap_cols = KernelCols - 1;
+
+  private:
+
+    // Maximum padding and number of distinct paddings
+    static constexpr int max_pad_top = KernelRows / 2;
+    static constexpr int min_pad_top = KernelRows % (InnerTileRows - overlap_rows);
+    static constexpr int n_pad_top = iceildiv(max_pad_top, InnerTileRows - overlap_rows);
+
+    static constexpr int max_pad_left = KernelCols / 2;
+    static constexpr int min_pad_left = KernelCols % (InnerTileCols - overlap_cols);
+    static constexpr int n_pad_left = iceildiv(max_pad_left, InnerTileCols - overlap_cols);
+
+    static constexpr int n_pad_bottom = InnerTileRows;
+    static constexpr int n_pad_right = InnerTileCols;
+
+    // Pointers to methods implementing a generically padded tile and a totally unpadded tile.
+    static const TileFn tilefn_generic;   /** Generic tile processing function. */
+    static const TileFn tilefn_unpadded;  /** Tile processor for unpadded tiles. */
+
+    // Arrays of methods covering tiles which are padded only on a single side.
+    static const TileFn tilefn_top_padded[n_pad_top];
+    static const TileFn tilefn_left_padded[n_pad_left];
+    static const TileFn tilefn_bottom_padded[n_pad_bottom];
+    static const TileFn tilefn_right_padded[n_pad_right];
+};
+
+
+template < int KernelCols, int InnerTileCols, typename T>
+class InputTransformImplTiles<1, KernelCols, 1, InnerTileCols, T>
+{
+  public:
+    /** Method to transform a tile of the input tensor into the Winograd domain. */
+    typedef void (*TileFn)(
+      const int n_channels,        /** @param[in] Number of channels in the tensor. */
+      const T* const inptr_base,   /** @param[in] Pointer to the base of the input tile. */
+      const int input_row_stride,  /** @param[in] Stride between rows of the input tensor. */
+      const int input_col_stride,  /** @param[in] Stride between columns of the input tensor. */
+      T* const mptr_base,          /** @param[out] Base pointer to transformed input matrices. */
+      const int matrix_stride,     /** @param[in] Stride between matrices in the input space. */
+      const int _pad_top,          /** @param[in] Top padding for unspecialised tiles. */
+      const int _pad_left,         /** @param[in] Left padding for unspecialised tiles. */
+      const int _pad_bottom,       /** @param[in] Bottom padding for unspecialised tiles. */
+      const int _pad_right         /** @param[in] Right padding for unspecialised tiles. */
+    );
+
+    static TileFn get_tile_specialization(
+      const int pad_top,
+      const int pad_left,
+      const int pad_bottom,
+      const int pad_right
+    );
+
+    // Tile overlaps
+    static constexpr int overlap_rows = 0;
+    static constexpr int overlap_cols = KernelCols - 1;
+
+  private:
+    // Maximum padding and number of distinct paddings
+    static constexpr int max_pad_left = KernelCols / 2;
+    static constexpr int min_pad_left = KernelCols % (InnerTileCols - overlap_cols);
+    static constexpr int n_pad_left = iceildiv(max_pad_left, InnerTileCols - overlap_cols);
+
+    static constexpr int n_pad_right = InnerTileCols;
+
+    // Pointers to methods implementing a generically padded tile and a totally unpadded tile.
+    static const TileFn tilefn_generic;   /** Generic tile processing function. */
+    static const TileFn tilefn_unpadded;  /** Tile processor for unpadded tiles. */
+
+    // Arrays of methods covering tiles which are padded only on a single side.
+    static const TileFn tilefn_left_padded[n_pad_left];
+    static const TileFn tilefn_right_padded[n_pad_right];
+};
+
+
+
+template <int KernelRows, int KernelCols, int InnerTileRows, int InnerTileCols, typename T>
 class InputTransformImpl
 {
   public:
@@ -69,29 +172,13 @@ class InputTransformImpl
       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);
+    using Tiles = InputTransformImplTiles<KernelRows, KernelCols, InnerTileRows, InnerTileCols, T>;
 
-    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 overlap_rows = Tiles::overlap_rows;
+    static constexpr int overlap_cols = Tiles::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>
diff --git a/src/core/NEON/kernels/convolution/winograd/transforms/input_1x8_fp32.cpp b/src/core/NEON/kernels/convolution/winograd/transforms/input_1x8_fp32.cpp
index 042d4debbc..e66300d39a 100644
--- a/src/core/NEON/kernels/convolution/winograd/transforms/input_1x8_fp32.cpp
+++ b/src/core/NEON/kernels/convolution/winograd/transforms/input_1x8_fp32.cpp
@@ -29,19 +29,30 @@
 namespace
 {
 
-template <int pad_top, int pad_left, int pad_bottom, int pad_right>
+template <bool Specialized, int PadTop=0, int PadLeft=0, int PadBottom=0, int PadRight=0>
 void winograd_input_transform_1x8_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
+  const int matrix_stride,
+     const int _pad_top,
+     const int _pad_left,
+     const int _pad_bottom,
+    const int _pad_right
 )
 {
   (void) input_row_stride;  // No rows over which to stride
+ (void) _pad_top;  // Never any top padding
+  (void) _pad_bottom;  // Never any bottom padding
+
+  // Extract padding arguments
+  const int pad_left = Specialized ? PadLeft : _pad_left;
+  const int pad_right = Specialized ? PadRight : _pad_right;
+
   constexpr int inner_tile_cols = 8;
-  constexpr int cells_j = inner_tile_cols - pad_right;
+  const int cells_j = inner_tile_cols - pad_right;
 
   float *outptr = matrix_base;
 
@@ -162,109 +173,85 @@ void winograd_input_transform_1x8_fp32_process_tile(
 namespace winograd
 {
 template <int x>
-using Transform = InputTransformImpl<1, x, 1, 8, float>;
+using Tiles = InputTransformImplTiles<1, x, 1, 8, float>;
 
+/*****************************************************************************/
+// 1x3 specialisations
 template <>
-const Transform<3>::TileFn
-  Transform<3>::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>,
-      }
-    }
-  }
+const Tiles<3>::TileFn Tiles<3>::tilefn_generic = winograd_input_transform_1x8_fp32_process_tile<false>;
+
+template <>
+const Tiles<3>::TileFn Tiles<3>::tilefn_unpadded = winograd_input_transform_1x8_fp32_process_tile<true>;
+
+template <>
+const Tiles<3>::TileFn Tiles<3>::tilefn_left_padded[n_pad_left] = {
+  winograd_input_transform_1x8_fp32_process_tile<true, 0, 1, 0, 0>,
 };
 
 template <>
-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>,
-      }
-    }
-  }
+const Tiles<3>::TileFn Tiles<3>::tilefn_right_padded[n_pad_right] = {
+  winograd_input_transform_1x8_fp32_process_tile<true, 0, 0, 0, 1>,
+  winograd_input_transform_1x8_fp32_process_tile<true, 0, 0, 0, 2>,
+  winograd_input_transform_1x8_fp32_process_tile<true, 0, 0, 0, 3>,
+  winograd_input_transform_1x8_fp32_process_tile<true, 0, 0, 0, 4>,
+  winograd_input_transform_1x8_fp32_process_tile<true, 0, 0, 0, 5>,
+  winograd_input_transform_1x8_fp32_process_tile<true, 0, 0, 0, 6>,
+  winograd_input_transform_1x8_fp32_process_tile<true, 0, 0, 0, 7>,
 };
+/*****************************************************************************/
 
+/*****************************************************************************/
+// 1x5 specialisations
 template <>
-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>,
-      }
-    }
-  }
+const Tiles<5>::TileFn Tiles<5>::tilefn_generic = winograd_input_transform_1x8_fp32_process_tile<false>;
+
+template <>
+const Tiles<5>::TileFn Tiles<5>::tilefn_unpadded = winograd_input_transform_1x8_fp32_process_tile<true>;
+
+template <>
+const Tiles<5>::TileFn Tiles<5>::tilefn_left_padded[n_pad_left] = {
+  winograd_input_transform_1x8_fp32_process_tile<true, 0, 2, 0, 0>,
 };
 
+template <>
+const Tiles<5>::TileFn Tiles<5>::tilefn_right_padded[n_pad_right] = {
+  winograd_input_transform_1x8_fp32_process_tile<true, 0, 0, 0, 1>,
+  winograd_input_transform_1x8_fp32_process_tile<true, 0, 0, 0, 2>,
+  winograd_input_transform_1x8_fp32_process_tile<true, 0, 0, 0, 3>,
+  winograd_input_transform_1x8_fp32_process_tile<true, 0, 0, 0, 4>,
+  winograd_input_transform_1x8_fp32_process_tile<true, 0, 0, 0, 5>,
+  winograd_input_transform_1x8_fp32_process_tile<true, 0, 0, 0, 6>,
+  winograd_input_transform_1x8_fp32_process_tile<true, 0, 0, 0, 7>,
+};
+/*****************************************************************************/
+
+/*****************************************************************************/
+// 1x7 specialisations
+template <>
+const Tiles<7>::TileFn Tiles<7>::tilefn_generic = winograd_input_transform_1x8_fp32_process_tile<false>;
+
+template <>
+const Tiles<7>::TileFn Tiles<7>::tilefn_unpadded = winograd_input_transform_1x8_fp32_process_tile<true>;
+
+template <>
+const Tiles<7>::TileFn Tiles<7>::tilefn_left_padded[n_pad_left] = {
+  winograd_input_transform_1x8_fp32_process_tile<true, 0, 1, 0, 0>,
+  winograd_input_transform_1x8_fp32_process_tile<true, 0, 3, 0, 0>,
+};
+
+template <>
+const Tiles<7>::TileFn Tiles<7>::tilefn_right_padded[n_pad_right] = {
+  winograd_input_transform_1x8_fp32_process_tile<true, 0, 0, 0, 1>,
+  winograd_input_transform_1x8_fp32_process_tile<true, 0, 0, 0, 2>,
+  winograd_input_transform_1x8_fp32_process_tile<true, 0, 0, 0, 3>,
+  winograd_input_transform_1x8_fp32_process_tile<true, 0, 0, 0, 4>,
+  winograd_input_transform_1x8_fp32_process_tile<true, 0, 0, 0, 5>,
+  winograd_input_transform_1x8_fp32_process_tile<true, 0, 0, 0, 6>,
+  winograd_input_transform_1x8_fp32_process_tile<true, 0, 0, 0, 7>,
+};
+/*****************************************************************************/
+
+
 template class InputTransform<1, 3, 1, 8, float>;
 template class InputTransform<3, 1, 8, 1, float>;
 template class InputTransform<1, 5, 1, 8, float>;
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 a9d5d52d15..4203945dd3 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
@@ -29,75 +29,36 @@
 namespace winograd
 {
 
-using Transform = InputTransformImpl<3, 3, 4, 4, float>;
-
-/*****************************************************************************
-* F(2x2, 3x3) implies the use of a 4x4 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______|
-*
-* For tiles near the right or bottom of the image it is more complicated.  Such
-* tiles might require padding by 0 or 1 rows or columns if the padding type is
-* VALID or 1 or 2 rows or columns if the padding type is SAME:
-*
-*      _______    _______    _______    _______
-*     |X X X X|  |X X X X|  |X X X X|  |X X X X|
-*     |X      |  |       |  |      X|  |    X X|
-*     |X      |  |       |  |      X|  |    X X|
-*     |X______|  |_______|  |______X|  |____X_X|
-*      _______    _______    _______    _______
-*     |X      |  |       |  |      X|  |    X X|
-*     |X      |  |       |  |      X|  |    X X|
-*     |X      |  |       |  |      X|  |    X X|
-*     |X______|  |_______|  |______X|  |____X_X|
-*      _______    _______    _______    _______
-*     |X      |  |       |  |      X|  |    X X|
-*     |X      |  |       |  |      X|  |    X X|
-*     |X      |  |       |  |      X|  |    X X|
-*     |X_X_X_X|  |X_X_X_X|  |X_X_X_X|  |X_X_X_X|
-*      _______    _______    _______    _______
-*     |X      |  |       |  |      X|  |    X X|
-*     |X      |  |       |  |      X|  |    X X|
-*     |X X X X|  |X X X X|  |X X X X|  |X X X X|
-*     |X_X_X_X|  |X_X_X_X|  |X_X_X_X|  |X_X_X_X|
-*
-* Additional tiles are required for especially small input images.
-*
-* 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}
-*     Padding right in {0, 1, 2}
-*/
-template <>
-template <int pad_top, int pad_left, int pad_bottom, int pad_right>
-void Transform::process_tile(
+using Tiles = InputTransformImplTiles<3, 3, 4, 4, float>;
+
+namespace
+{
+
+
+template <bool Specialized, int PadTop=0, int PadLeft=0, int PadBottom=0, int PadRight=0>
+void winograd_input_transform_4x4_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
-)
+    const int matrix_stride,
+     const int _pad_top,
+     const int _pad_left,
+     const int _pad_bottom,
+     const int _pad_right
+  )
 {
+const int pad_top = Specialized ? PadTop : _pad_top;
+  const int pad_left = Specialized ? PadLeft : _pad_left;
+  const int pad_bottom = Specialized ? PadBottom : _pad_bottom;
+  const int pad_right = Specialized ? PadRight : _pad_right;
+
   constexpr int inner_tile_i = 4, inner_tile_j = 4;
-  constexpr int cells_i = inner_tile_i - pad_bottom;
-  constexpr int cells_j = inner_tile_i - pad_right;
+  const int cells_i = inner_tile_i - pad_bottom;
+  const int cells_j = inner_tile_i - pad_right;
+
+
 
   float *outptr = matrix_base;
 
@@ -311,81 +272,39 @@ void Transform::process_tile(
   }
 }
 
+}  // namespace (anonymous)
+
 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>,  // Right
-      },
-      {
-        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>,  // Bottom-right
-      },
-      {
-        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>,  // Bottom-right
-      }
-    },
-    {
-      {
-        Transform::template process_tile<0, 1, 0, 0>,  // Left
-        Transform::template process_tile<0, 1, 0, 1>,  // Left AND right
-        Transform::template process_tile<0, 1, 0, 2>,  // Left AND right
-      },
-      {
-        Transform::template process_tile<0, 1, 1, 0>,  // Left-bottom
-        Transform::template process_tile<0, 1, 1, 1>,  // Left, bottom AND right
-        Transform::template process_tile<0, 1, 1, 2>,  // Left, bottom AND right
-      },
-      {
-        Transform::template process_tile<0, 1, 2, 0>,  // Left-bottom
-        Transform::template process_tile<0, 1, 2, 1>,  // Left, bottom AND right
-        Transform::template process_tile<0, 1, 2, 2>,  // Left, bottom AND right
-      }
-    },
-  },
-  {
-    {
-      {
-        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>,  // Top-right
-      },
-      {
-        Transform::template process_tile<1, 0, 1, 0>,  // Top AND bottom
-        Transform::template process_tile<1, 0, 1, 1>,  // Top, bottom AND right
-        Transform::template process_tile<1, 0, 1, 2>,  // Top, bottom AND right
-      },
-      {
-        Transform::template process_tile<1, 0, 2, 0>,  // Top AND bottom
-        Transform::template process_tile<1, 0, 2, 1>,  // Top, bottom AND right
-        Transform::template process_tile<1, 0, 2, 2>,  // Top, bottom AND right
-      }
-    },
-    {
-      {
-        Transform::template process_tile<1, 1, 0, 0>,  // Top-left
-        Transform::template process_tile<1, 1, 0, 1>,  // Top, left AND right
-        Transform::template process_tile<1, 1, 0, 2>,  // Top, left AND right
-      },
-      {
-        Transform::template process_tile<1, 1, 1, 0>,  // Top, left AND bottom
-        Transform::template process_tile<1, 1, 1, 1>,  // All padded
-        Transform::template process_tile<1, 1, 1, 2>,  // All padded
-      },
-      {
-        Transform::template process_tile<1, 1, 2, 0>,  // Top, left AND bottom
-        Transform::template process_tile<1, 1, 2, 1>,  // All padded
-        Transform::template process_tile<1, 1, 2, 2>,  // All padded
-      }
-    }
-  }
+const Tiles::TileFn Tiles::tilefn_generic = winograd_input_transform_4x4_fp32_process_tile<false>;
+
+template <>
+const Tiles::TileFn Tiles::tilefn_unpadded = winograd_input_transform_4x4_fp32_process_tile<true>;
+
+
+template <>
+const Tiles::TileFn Tiles::tilefn_top_padded[n_pad_top] = {
+  winograd_input_transform_4x4_fp32_process_tile<true, 1, 0, 0, 0>,
+};
+
+template <>
+const Tiles::TileFn Tiles::tilefn_left_padded[n_pad_left] = {
+  winograd_input_transform_4x4_fp32_process_tile<true, 0, 1, 0, 0>,
+};
+
+template <>
+const Tiles::TileFn Tiles::tilefn_bottom_padded[n_pad_bottom] = {
+  winograd_input_transform_4x4_fp32_process_tile<true, 0, 0, 1, 0>,
+  winograd_input_transform_4x4_fp32_process_tile<true, 0, 0, 2, 0>,
+  winograd_input_transform_4x4_fp32_process_tile<true, 0, 0, 3, 0>,
+  winograd_input_transform_4x4_fp32_process_tile<true, 0, 0, 4, 0>,
+};
+
+template <>
+const Tiles::TileFn Tiles::tilefn_right_padded[n_pad_right] = {
+  winograd_input_transform_4x4_fp32_process_tile<true, 0, 0, 0, 1>,
+  winograd_input_transform_4x4_fp32_process_tile<true, 0, 0, 0, 2>,
+  winograd_input_transform_4x4_fp32_process_tile<true, 0, 0, 0, 3>,
+  winograd_input_transform_4x4_fp32_process_tile<true, 0, 0, 0, 4>,
 };
 
 template class InputTransform<3, 3, 4, 4, float>;
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
index 908613068a..893122cc45 100644
--- a/src/core/NEON/kernels/convolution/winograd/transforms/input_6x6_fp32.cpp
+++ b/src/core/NEON/kernels/convolution/winograd/transforms/input_6x6_fp32.cpp
@@ -29,20 +29,30 @@
 namespace
 {
 
-template <int pad_top, int pad_left, int pad_bottom, int pad_right>
+template <bool Specialized, int PadTop=0, int PadLeft=0, int PadBottom=0, int PadRight=0>
 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
+const int matrix_stride,
+     const int _pad_top,
+     const int _pad_left,
+     const int _pad_bottom,
+     const int _pad_right
 )
 {
-  constexpr int inner_tile_rows = 6;
+  const int pad_top = Specialized ? PadTop : _pad_top;
+  const int pad_left = Specialized ? PadLeft : _pad_left;
+  const int pad_bottom = Specialized ? PadBottom : _pad_bottom;
+  const int pad_right = Specialized ? PadRight : _pad_right;
+
+ 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;
+
+  const int cells_i = inner_tile_rows - pad_bottom;
+  const int cells_j = inner_tile_cols - pad_right;
 
   float *outptr = matrix_base;
 
@@ -285,322 +295,80 @@ void winograd_input_transform_6x6_fp32_process_tile(
 namespace winograd
 {
 template <int k>
-using Transform = InputTransformImpl<k, k, 6, 6, float>;
+using Tiles = InputTransformImplTiles<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>,
-      }
-    }
-  }
+const Tiles<3>::TileFn Tiles<3>::tilefn_generic = winograd_input_transform_6x6_fp32_process_tile<false>;
+
+template <>
+const Tiles<3>::TileFn Tiles<3>::tilefn_unpadded = winograd_input_transform_6x6_fp32_process_tile<true>;
+
+template <>
+const Tiles<3>::TileFn Tiles<3>::tilefn_top_padded[n_pad_top] = {
+  winograd_input_transform_6x6_fp32_process_tile<true, 1, 0, 0, 0>,
 };
 
 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>,
-      }
-    }
-  }
+const Tiles<3>::TileFn Tiles<3>::tilefn_left_padded[n_pad_left] = {
+  winograd_input_transform_6x6_fp32_process_tile<true, 0, 1, 0, 0>,
+};
+
+template <>
+const Tiles<3>::TileFn Tiles<3>::tilefn_bottom_padded[n_pad_bottom] = {
+  winograd_input_transform_6x6_fp32_process_tile<true, 0, 0, 1, 0>,
+  winograd_input_transform_6x6_fp32_process_tile<true, 0, 0, 2, 0>,
+  winograd_input_transform_6x6_fp32_process_tile<true, 0, 0, 3, 0>,
+  winograd_input_transform_6x6_fp32_process_tile<true, 0, 0, 4, 0>,
+  winograd_input_transform_6x6_fp32_process_tile<true, 0, 0, 5, 0>,
+  winograd_input_transform_6x6_fp32_process_tile<true, 0, 0, 6, 0>,
+};
+
+template <>
+const Tiles<3>::TileFn Tiles<3>::tilefn_right_padded[n_pad_right] = {
+  winograd_input_transform_6x6_fp32_process_tile<true, 0, 0, 0, 1>,
+  winograd_input_transform_6x6_fp32_process_tile<true, 0, 0, 0, 2>,
+  winograd_input_transform_6x6_fp32_process_tile<true, 0, 0, 0, 3>,
+  winograd_input_transform_6x6_fp32_process_tile<true, 0, 0, 0, 4>,
+  winograd_input_transform_6x6_fp32_process_tile<true, 0, 0, 0, 5>,
+  winograd_input_transform_6x6_fp32_process_tile<true, 0, 0, 0, 6>,
+};
+
+template <>
+const Tiles<5>::TileFn Tiles<5>::tilefn_generic = winograd_input_transform_6x6_fp32_process_tile<false>;
+
+
+template <>
+const Tiles<5>::TileFn Tiles<5>::tilefn_unpadded = winograd_input_transform_6x6_fp32_process_tile<true>;
+
+
+template <>
+const Tiles<5>::TileFn Tiles<5>::tilefn_top_padded[n_pad_top] = {
+  winograd_input_transform_6x6_fp32_process_tile<true, 2, 0, 0, 0>,
+};
+
+template <>
+const Tiles<5>::TileFn Tiles<5>::tilefn_left_padded[n_pad_left] = {
+  winograd_input_transform_6x6_fp32_process_tile<true, 0, 2, 0, 0>,
+};
+
+template <>
+const Tiles<5>::TileFn Tiles<5>::tilefn_bottom_padded[n_pad_bottom] = {
+  winograd_input_transform_6x6_fp32_process_tile<true, 0, 0, 1, 0>,
+  winograd_input_transform_6x6_fp32_process_tile<true, 0, 0, 2, 0>,
+  winograd_input_transform_6x6_fp32_process_tile<true, 0, 0, 3, 0>,
+  winograd_input_transform_6x6_fp32_process_tile<true, 0, 0, 4, 0>,
+  winograd_input_transform_6x6_fp32_process_tile<true, 0, 0, 5, 0>,
+  winograd_input_transform_6x6_fp32_process_tile<true, 0, 0, 6, 0>,
+};
+
+template <>
+const Tiles<5>::TileFn Tiles<5>::tilefn_right_padded[n_pad_right] = {
+  winograd_input_transform_6x6_fp32_process_tile<true, 0, 0, 0, 1>,
+  winograd_input_transform_6x6_fp32_process_tile<true, 0, 0, 0, 2>,
+  winograd_input_transform_6x6_fp32_process_tile<true, 0, 0, 0, 3>,
+  winograd_input_transform_6x6_fp32_process_tile<true, 0, 0, 0, 4>,
+  winograd_input_transform_6x6_fp32_process_tile<true, 0, 0, 0, 5>,
+  winograd_input_transform_6x6_fp32_process_tile<true, 0, 0, 0, 6>,
 };
 
 template class InputTransform<3, 3, 6, 6, float>;