COMPMID-1339 - Implementing Winograd Convolution Layer 1x5 and 5x1 kernels on OpenCL NCHW

Change-Id: Ia293cd89651146a0e27e5f7c74ca9c924807e83c
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/138707
Tested-by: Jenkins <bsgcomp@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>

3 files changed, 48 insertions, 18 deletions

diff --git a/arm_compute/core/CL/kernels/CLWinogradFilterTransformKernel.h b/arm_compute/core/CL/kernels/CLWinogradFilterTransformKernel.h
index 5e3d815d8c..9d0833d695 100644
--- a/arm_compute/core/CL/kernels/CLWinogradFilterTransformKernel.h
+++ b/arm_compute/core/CL/kernels/CLWinogradFilterTransformKernel.h
@@ -48,10 +48,15 @@ public:
     ~CLWinogradFilterTransformKernel() = default;
     /** Set the input and output tensor.
      *
-     * @note Winograd filter transform supports the following configurations:
-     *       F(output tile, kernel size):F(2x2, 3x3), F(4x4, 3x3), F(4x4, 5x5)
+     * @note Winograd filter transform supports the following configurations for NCWH data layout
+     *       F(output tile, kernel size):F(2x2, 3x3), F(2x1, 3x1), F(1x2, 1x3),
+     *                                   F(4x4, 3x3), F(4x1, 3x1), F(1x4, 1x3),
+     *                                   F(4x4, 5x5), F(4x1, 5x1), F(1x4, 1x5)
+     *
+     * @note Winograd filter transform supports the following configurations for NHWC data layout
+     *       F(output tile, kernel size):F(4x4, 3x3), F(4x1, 3x1), F(1x4, 1x3)
+     *                                   F(4x4, 5x5)
      *       Strides: only unit strides
-     *       Data Layout: NCHW for all configurations, NHWC for F(4x4, 3x3) and F(4x4, 5x5)
      *
      * @param[in]  input         Source tensor. The input is a 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM] (NCHW data layout) or [IFM, kernel_x, kernel_y, OFM] (NHWC data layout). Data types supported: F32.
      * @param[out] output        The output tensor. The shape for this tensor can be calculated using the utility function @p compute_winograd_filter_transform_shape. Data types supported: Same as @p input
@@ -60,10 +65,15 @@ public:
     void configure(const ICLTensor *input, ICLTensor *output, const WinogradInfo &winograd_info);
     /** Static function to check if given info will lead to a valid configuration of @ref CLWinogradFilterTransformKernel
      *
-     * @note Winograd filter transform supports the following configurations:
-     *       F(output tile, kernel size):F(2x2, 3x3), F(4x4, 3x3), F(4x4, 5x5)
+     * @note Winograd filter transform supports the following configurations for NCWH data layout
+     *       F(output tile, kernel size):F(2x2, 3x3), F(2x1, 3x1), F(1x2, 1x3),
+     *                                   F(4x4, 3x3), F(4x1, 3x1), F(1x4, 1x3),
+     *                                   F(4x4, 5x5), F(4x1, 5x1), F(1x4, 1x5)
+     *
+     * @note Winograd filter transform supports the following configurations for NHWC data layout
+     *       F(output tile, kernel size):F(4x4, 3x3),
+     *                                   F(4x4, 5x5)
      *       Strides: only unit strides
-     *       Data Layout: NCHW for all configurations, NHWC for F(4x4, 3x3) and F(4x4, 5x5)
      *
      * @param[in]  input         Source tensor. The input is a 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM] (NCHW data layout) or [IFM, kernel_x, kernel_y, OFM] (NHWC data layout). Data types supported: F32.
      * @param[out] output        The output tensor. The shape for this tensor can be calculated using the utility function @p compute_winograd_filter_transform_shape. Data types supported: Same as @p input
diff --git a/arm_compute/core/CL/kernels/CLWinogradInputTransformKernel.h b/arm_compute/core/CL/kernels/CLWinogradInputTransformKernel.h
index ddf07200d8..410e8ba765 100644
--- a/arm_compute/core/CL/kernels/CLWinogradInputTransformKernel.h
+++ b/arm_compute/core/CL/kernels/CLWinogradInputTransformKernel.h
@@ -46,10 +46,15 @@ public:
     CLWinogradInputTransformKernel &operator=(CLWinogradInputTransformKernel &&) = default;
     /** Set the input and output of the kernel.
      *
-     * @note Winograd input transform supports the following configurations:
-     *       F(output tile, kernel size):F(2x2, 3x3), F(4x4, 3x3), F(4x4, 5x5)
+     * @note Winograd input transform supports the following configurations for NCWH data layout
+     *       F(output tile, kernel size):F(2x2, 3x3), F(2x1, 3x1), F(1x2, 1x3),
+     *                                   F(4x4, 3x3), F(4x1, 3x1), F(1x4, 1x3),
+     *                                   F(4x4, 5x5), F(4x1, 5x1), F(1x4, 1x5)
+     *
+     * @note Winograd input transform supports the following configurations for NHWC data layout
+     *       F(output tile, kernel size):F(4x4, 3x3), F(4x1, 3x1), F(1x4, 1x3)
+     *                                   F(4x4, 5x5)
      *       Strides: only unit strides
-     *       Data Layout: NCHW for all configurations, NHWC for F(4x4, 3x3), F(4x4, 5x5)
      *
      * @param[in] input         The input tensor to transform. Data types supported: F32
      * @param[in] output        The output tensor. The shape for this tensor can be calculated using the utility function @p compute_winograd_input_transform_shape. Data types supported: Same as @p input
@@ -58,10 +63,15 @@ public:
     void configure(const ICLTensor *input, ICLTensor *output, const WinogradInfo &winograd_info);
     /** Static function to check if given info will lead to a valid configuration of @ref CLWinogradInputTransformKernel
      *
-     * @note Winograd input transform supports the following configurations:
-     *       F(output tile, kernel size):F(2x2, 3x3), F(4x4, 3x3), F(4x4, 5x5)
+     * @note Winograd input transform supports the following configurations for NCWH data layout
+     *       F(output tile, kernel size):F(2x2, 3x3), F(2x1, 3x1), F(1x2, 1x3),
+     *                                   F(4x4, 3x3), F(4x1, 3x1), F(1x4, 1x3),
+     *                                   F(4x4, 5x5), F(4x1, 5x1), F(1x4, 1x5)
+     *
+     * @note Winograd input transform supports the following configurations for NHWC data layout
+     *       F(output tile, kernel size):F(4x4, 3x3),
+     *                                   F(4x4, 5x5)
      *       Strides: only unit strides
-     *       Data Layout: NCHW for all configurations, NHWC for F(4x4, 3x3), F(4x4, 5x5)
      *
      * @param[in] input         The input tensor to transform. Data types supported: F32
      * @param[in] output        The output tensor. The shape for this tensor can be calculated using the utility function @p compute_winograd_input_transform_shape. Data types supported: Same as @p input
diff --git a/arm_compute/core/CL/kernels/CLWinogradOutputTransformKernel.h b/arm_compute/core/CL/kernels/CLWinogradOutputTransformKernel.h
index cd46e9813e..0798172ba7 100644
--- a/arm_compute/core/CL/kernels/CLWinogradOutputTransformKernel.h
+++ b/arm_compute/core/CL/kernels/CLWinogradOutputTransformKernel.h
@@ -48,10 +48,15 @@ public:
     ~CLWinogradOutputTransformKernel() = default;
     /** Set the input and output tensor.
      *
-     * @note Winograd output transform supports the following configurations:
-     *       F(output tile, kernel size):F(2x2, 3x3), F(4x4, 3x3), F(4x4, 5x5)
+     * @note Winograd output transform supports the following configurations for NCWH data layout
+     *       F(output tile, kernel size):F(2x2, 3x3), F(2x1, 3x1), F(1x2, 1x3),
+     *                                   F(4x4, 3x3), F(4x1, 3x1), F(1x4, 1x3),
+     *                                   F(4x4, 5x5), F(4x1, 5x1), F(1x4, 1x5)
+     *
+     * @note Winograd output transform supports the following configurations for NHWC data layout
+     *       F(output tile, kernel size):F(4x4, 3x3),
+     *                                   F(4x4, 5x5)
      *       Strides: only unit strides
-     *       Data Layout: NCHW for all configurations, NHWC for F(4x4, 3x3) and F(4x4, 5x5)
      *
      * @param[in]  input         Source tensor with shape [C, N, K, batches]. Data types supported: F32.
      * @param[in]  bias          Biases tensor. Shared biases supported. Biases are 1D tensor with dimensions [OFM]. It can be a nullptr. Data type supported: as @p input
@@ -61,10 +66,15 @@ public:
     void configure(const ICLTensor *input, const ICLTensor *bias, ICLTensor *output, const WinogradInfo &winograd_info);
     /** Static function to check if given info will lead to a valid configuration of @ref CLWinogradOutputTransformKernel
      *
-     * @note Winograd output transform supports the following configurations:
-     *       F(output tile, kernel size):F(2x2, 3x3), F(4x4, 3x3), F(4x4, 5x5)
+     * @note Winograd output transform supports the following configurations for NCWH data layout
+     *       F(output tile, kernel size):F(2x2, 3x3), F(2x1, 3x1), F(1x2, 1x3),
+     *                                   F(4x4, 3x3), F(4x1, 3x1), F(1x4, 1x3),
+     *                                   F(4x4, 5x5), F(4x1, 5x1), F(1x4, 1x5)
+     *
+     * @note Winograd output transform supports the following configurations for NHWC data layout
+     *       F(output tile, kernel size):F(4x4, 3x3), F(4x1, 3x1), F(1x4, 1x3)
+     *                                   F(4x4, 5x5)
      *       Strides: only unit strides
-     *       Data Layout: NCHW for all configurations, NHWC for F(4x4, 3x3) and F(4x4, 5x5)
      *
      * @param[in]  input         Source tensor with shape [C, N, K, batches]. Data types supported: F32.
      * @param[in]  bias          Biases tensor. Shared biases supported. Biases are 1D tensor with dimensions [OFM]. It can be a nullptr. Data type supported: as @p input