COMPMID-1623: NEWinograd reduce the number of output tiles.

Change-Id: I4d9240924fe483d2dd127ad6a4ae6f8066f61bd1
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/151893
Tested-by: bsgcomp <bsgcomp@arm.com>
Reviewed-by: Andrew Mundy <andrew.mundy@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>

1 files changed, 26 insertions, 33 deletions

diff --git a/src/core/NEON/kernels/convolution/winograd/transforms/output_4_5_fp32.cpp b/src/core/NEON/kernels/convolution/winograd/transforms/output_4_5_fp32.cpp
index 2417f527bf..911759b128 100644
--- a/src/core/NEON/kernels/convolution/winograd/transforms/output_4_5_fp32.cpp
+++ b/src/core/NEON/kernels/convolution/winograd/transforms/output_4_5_fp32.cpp
@@ -23,38 +23,32 @@
  */
 
 #include "arm_compute/core/NEON/kernels/convolution/winograd/transforms/output.hpp"
-#include "arm_compute/core/NEON/kernels/convolution/winograd/winograd_gemm.hpp"
+#include "arm_compute/core/NEON/kernels/convolution/winograd/winograd_output_transform.hpp"
 #include "arm_compute/core/NEON/kernels/convolution/common/arm.hpp"
 
-namespace winograd
-{
-
-using Transform = WinogradGEMM<1, 4, 1, 5>::OutputTransform<float>;
-using TransformTransposed = WinogradGEMM<4, 1, 5, 1>::OutputTransform<float>;
-
-template <>
-template <>
-int Transform::ops_performed(const Tensor4DShape &shape)
+namespace
 {
-  (void) shape;
-  return 0;  // TODO
-}
 
-template <>
-template <>
-template <int pad_bottom, int pad_right>
-void Transform::process_tile(
+template <bool Specialized, int PadRight=0>
+void winograd_output_transform_4_5_fp32_process_tile(
   const int n_channels,
   const float* const matrix_base,
   const int matrix_stride,
   const float* const biases,
   float* const output,
   const int output_row_stride,
-  const int output_col_stride
+  const int output_col_stride,
+  const int _pad_bottom,
+  const int _pad_right
 )
 {
   (void) output_row_stride;
-  constexpr int cells_j = output_tile_cols - pad_right;
+  (void) _pad_bottom;
+  constexpr int output_tile_cols = 4;
+  constexpr int inner_tile_cols = 8;
+
+  const int pad_right = Specialized ? PadRight : _pad_right;
+  const int cells_j = output_tile_cols - pad_right;
 
   // Construct a map to the output cells
   float *outptrs[cells_j];
@@ -156,23 +150,22 @@ void Transform::process_tile(
   }
 }
 
-template <>
-template <>
-const Transform::TileFn Transform::tile_fns[max_pad_bottom][max_pad_right] =
+}  // namespace (anonymous)
+
+namespace winograd
 {
-  {
-    Transform::template process_tile<0, 0>,
-    Transform::template process_tile<0, 1>,
-    Transform::template process_tile<0, 2>,
-    Transform::template process_tile<0, 3>,
-  },
-};
+using Tiles = OutputTransformImplTiles<1, 5, 1, 8, float>;
 
 template <>
-template <>
-const TransformTransposed::TileFn TransformTransposed::tile_fns[max_pad_bottom][max_pad_right] = {};
+const Tiles::TileFn Tiles::tilefn_unpadded = winograd_output_transform_4_5_fp32_process_tile<true>;
 
+template <>
+const Tiles::TileFn Tiles::tilefn_right_padded[n_pad_right] = {
+  winograd_output_transform_4_5_fp32_process_tile<true, 1>,
+  winograd_output_transform_4_5_fp32_process_tile<true, 2>,
+  winograd_output_transform_4_5_fp32_process_tile<true, 3>
+};
 
-template struct WinogradGEMM<1, 4, 1, 5>::OutputTransform<float>;
-template struct WinogradGEMM<4, 1, 5, 1>::OutputTransform<float>;
+template class OutputTransform<1, 5, 1, 8, float>;
+template class OutputTransform<5, 1, 8, 1, float>;
 }  // namespace winograd