COMPMID-1009 Support 4x4 output tile for Winograd Filter Transform on OpenCL.

Change-Id: I68c6453e0f192de659582404f109a89616b9fbb9
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/124811
Tested-by: Jenkins <bsgcomp@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>

3 files changed, 31 insertions, 19 deletions

diff --git a/arm_compute/core/CL/kernels/CLWinogradFilterTransformKernel.h b/arm_compute/core/CL/kernels/CLWinogradFilterTransformKernel.h
index ec5e51482a..c4ae5745b8 100644
--- a/arm_compute/core/CL/kernels/CLWinogradFilterTransformKernel.h
+++ b/arm_compute/core/CL/kernels/CLWinogradFilterTransformKernel.h
@@ -48,20 +48,22 @@ public:
     ~CLWinogradFilterTransformKernel() = default;
     /** Set the input and output tensor.
      *
-     * @param[in]  input  Source tensor. The input is a 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM] (NCHW data layout).
-     *                    kernel_x must be 3 and equal to kernel_y. Data types supported: F32.
-     * @param[out] output Destination tensor. The output is a 3D tensor with dimensions [OFM, IFM, 16]. Data type supported: same as @p input
+     * @param[in]  input       Source tensor. The input is a 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM] (NCHW data layout).
+     *                         kernel_x must be 3 and equal to kernel_y. Data types supported: F32.
+     * @param[out] output      Destination tensor. The output is a 3D tensor with dimensions [OFM, IFM, 16]. Data type supported: same as @p input
+     * @param[in]  output_tile Output tile. Currently only 2x2 and 4x4 tiles are supported.
      */
-    void configure(const ICLTensor *input, ICLTensor *output);
+    void configure(const ICLTensor *input, ICLTensor *output, const Size2D &output_tile);
     /** Static function to check if given info will lead to a valid configuration of @ref CLWinogradFilterTransformKernel
      *
-     * @param[in] input  Source tensor info. The input is a 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM] (NCHW data layout).
-     *                   kernel_x must be 3 and equal to kernel_y. Data types supported: F32.
-     * @param[in] output Destination tensor info. The output is a 3D tensor with dimensions [OFM, IFM, 16]. Data type supported: same as @p input
+     * @param[in] input       Source tensor info. The input is a 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM] (NCHW data layout).
+     *                        kernel_x must be 3 and equal to kernel_y. Data types supported: F32.
+     * @param[in] output      Destination tensor info. The output is a 3D tensor with dimensions [OFM, IFM, 16]. Data type supported: same as @p input
+     * @param[in] output_tile Output tile. Currently only 2x2 and 4x4 tiles are supported.
      *
      * @return a status
      */
-    static Status validate(const ITensorInfo *input, const ITensorInfo *output);
+    static Status validate(const ITensorInfo *input, const ITensorInfo *output, const Size2D &output_tile);
 
     // Inherited methods overridden:
     void run(const Window &window, cl::CommandQueue &queue) override;
diff --git a/arm_compute/core/Size2D.h b/arm_compute/core/Size2D.h
index 3840771cd1..37c4ebd041 100644
--- a/arm_compute/core/Size2D.h
+++ b/arm_compute/core/Size2D.h
@@ -24,6 +24,7 @@
 #ifndef __ARM_COMPUTE_SIZE2D_H__
 #define __ARM_COMPUTE_SIZE2D_H__
 
+#include "support/ToolchainSupport.h"
 #include <cstddef>
 #include <utility>
 
@@ -54,6 +55,21 @@ public:
         return (width * height);
     }
 
+    bool operator==(const Size2D &other) const
+    {
+        return (width == other.width) && (height == other.height);
+    }
+
+    bool operator!=(const Size2D &other) const
+    {
+        return !(*this == other);
+    }
+
+    std::string to_string() const
+    {
+        return support::cpp11::to_string(width) + std::string("x") + support::cpp11::to_string(height);
+    }
+
 public:
     size_t width  = {}; /**< Width of the image region or rectangle */
     size_t height = {}; /**< Height of the image region or rectangle */
diff --git a/arm_compute/core/utils/misc/ShapeCalculator.h b/arm_compute/core/utils/misc/ShapeCalculator.h
index 5344ce7e74..383fc6cda6 100644
--- a/arm_compute/core/utils/misc/ShapeCalculator.h
+++ b/arm_compute/core/utils/misc/ShapeCalculator.h
@@ -196,23 +196,17 @@ inline TensorShape compute_fully_connected_reshaped_weights_shape(const ITensorI
     return output_shape;
 }
 
-inline TensorShape compute_winograd_filter_transform_shape(const ITensorInfo &input)
+inline TensorShape compute_winograd_filter_transform_shape(const ITensorInfo &input, const Size2D &output_tile)
 {
-    // COMPMID-984 (giaiod01)
     TensorShape tensor_shape{ input.tensor_shape() };
 
+    tensor_shape.remove_dimension(get_data_layout_dimension_index(input.data_layout(), DataLayoutDimension::WIDTH));
+    tensor_shape.set(Window::DimY, input.dimension(2));
+    tensor_shape.set(Window::DimZ, (output_tile.width == 2) ? 16 : 36);
+
     if(input.data_layout() == DataLayout::NCHW)
     {
-        tensor_shape.remove_dimension(0);
         tensor_shape.set(Window::DimX, input.dimension(3));
-        tensor_shape.set(Window::DimY, input.dimension(2));
-        tensor_shape.set(Window::DimZ, 16);
-    }
-    else
-    {
-        tensor_shape.remove_dimension(1);
-        tensor_shape.set(Window::DimY, input.dimension(2));
-        tensor_shape.set(Window::DimZ, 16);
     }
 
     return tensor_shape;