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>

2 files changed, 197 insertions, 28 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>