COMPMID-784: Fixed SAME paddding in WinogradLayer

There were mismatches when using kernel size 5 and padding = SAME

Change-Id: Id834e96ebcf665616f99c995b48e302dcff8dc48
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/121144
Tested-by: Jenkins <bsgcomp@arm.com>
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>

3 files changed, 87 insertions, 84 deletions

diff --git a/arm_compute/core/NEON/kernels/convolution/common/utils.hpp b/arm_compute/core/NEON/kernels/convolution/common/utils.hpp
index d8b9c3b7d3..a22809fb58 100644
--- a/arm_compute/core/NEON/kernels/convolution/common/utils.hpp
+++ b/arm_compute/core/NEON/kernels/convolution/common/utils.hpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017 ARM Limited.
+ * Copyright (c) 2017-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -25,13 +25,15 @@
 #pragma once
 
 double TimeInUs(void);
-void PrintMatrix(const float* const m, const int M, const int N, const int row_stride);
+void PrintMatrix(const float *const m, const int M, const int N, const int row_stride);
 
-inline int iceildiv(const int a, const int b) {
-  return (a + b - 1) / b;
+inline int iceildiv(const int a, const int b)
+{
+    return (a + b - 1) / b;
 }
 
 template <typename T>
-inline T roundup(const T a, const T b) {
-  return a + b - (a % b);
+inline T roundup(const T a, const T b)
+{
+    return b * iceildiv(a, b);
 }
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
index d9ebe8b7cd..fd30b6118e 100644
--- 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
@@ -21,7 +21,6 @@
  * 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"
@@ -35,6 +34,7 @@ template <>
 template <>
 int Transform::ops_performed(const Tensor4DShape &input_shape)
 {
+  (void) input_shape;
   return 0;  // TODO
 }
 
@@ -45,8 +45,8 @@ int Transform::ops_performed(const Tensor4DShape &input_shape)
 * different padding combinations which may be required. These padding
 * constraints are the space:
 *
-*     Padding top in {0, 1}
-*     Padding left in {0, 1}
+*     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}
 */
@@ -340,115 +340,115 @@ const Transform::TileFn Transform::tile_fns[2][2][max_pad_bottom][max_pad_right]
     },
     {
       {
-        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, 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, 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, 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, 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, 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, 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, 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, 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<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<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<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<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<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<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<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<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<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<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<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<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<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<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<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<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<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<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<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<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>,
+        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>,
       }
     }
   }
diff --git a/tests/datasets/SmallConvolutionLayerDataset.h b/tests/datasets/SmallConvolutionLayerDataset.h
index d88efbd21b..adb61de8e2 100644
--- a/tests/datasets/SmallConvolutionLayerDataset.h
+++ b/tests/datasets/SmallConvolutionLayerDataset.h
@@ -51,6 +51,7 @@ public:
 
         // Kernel size 5
         add_config(TensorShape(8U, 8U, 2U), TensorShape(5U, 5U, 2U, 1U), TensorShape(1U), TensorShape(4U, 4U, 1U), PadStrideInfo(1, 1, 0, 0));
+        add_config(TensorShape(8U, 8U, 2U), TensorShape(5U, 5U, 2U), TensorShape(1U), TensorShape(8U, 8U, 1U), PadStrideInfo(1, 1, 2, 2));
     }
 };