Port CLWinogradConvolutionLayer with ClWinogradConv2d

Port CLWinogradInputTransformKernel
Port CLWinogradFilterTransformKernel
Port CLWinogradOutputTransformKernel

Resolves: COMPMID-4504

Change-Id: I3177dda0b9c2f56b36cb317027e94abe8d47229e
Signed-off-by: Manuel Bottini <manuel.bottini@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/5680
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>

6 files changed, 945 insertions, 0 deletions

diff --git a/src/core/gpu/cl/kernels/ClWinogradFilterTransformKernel.cpp b/src/core/gpu/cl/kernels/ClWinogradFilterTransformKernel.cpp
new file mode 100644
index 0000000000..381b4bcae9
--- /dev/null
+++ b/src/core/gpu/cl/kernels/ClWinogradFilterTransformKernel.cpp
@@ -0,0 +1,151 @@
+/*
+ * Copyright (c) 2018-2021 Arm Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#include "src/core/gpu/cl/kernels/ClWinogradFilterTransformKernel.h"
+
+#include "arm_compute/core/CL/CLHelpers.h"
+#include "arm_compute/core/CL/CLKernelLibrary.h"
+#include "arm_compute/core/CL/ICLTensor.h"
+#include "arm_compute/core/Helpers.h"
+#include "arm_compute/core/TensorInfo.h"
+#include "arm_compute/core/Types.h"
+#include "arm_compute/core/Utils.h"
+#include "arm_compute/core/Validate.h"
+#include "arm_compute/core/Window.h"
+#include "arm_compute/core/utils/misc/ShapeCalculator.h"
+#include "src/core/CL/CLValidate.h"
+#include "src/core/helpers/AutoConfiguration.h"
+#include "src/core/helpers/WindowHelpers.h"
+#include "support/Cast.h"
+#include "support/StringSupport.h"
+
+using namespace arm_compute::misc::shape_calculator;
+
+namespace arm_compute
+{
+namespace opencl
+{
+namespace kernels
+{
+namespace
+{
+Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, const WinogradInfo &winograd_info)
+{
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32, DataType::F16);
+    ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input);
+
+    const Size2D kernel_size      = winograd_info.kernel_size;
+    const Size2D output_tile_size = winograd_info.output_tile_size;
+
+    const size_t idx_w = get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::WIDTH);
+    const size_t idx_h = get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::HEIGHT);
+
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG(!cl_winograd_convolution_layer_supported(output_tile_size, kernel_size, input->data_layout()), "Winograd filter transform not supported");
+    ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(idx_w) != kernel_size.width || input->dimension(idx_h) != kernel_size.height);
+    ARM_COMPUTE_RETURN_ERROR_ON(input->num_dimensions() > 4);
+
+    // Checks performed when output is configured
+    if(output->total_size() != 0)
+    {
+        const TensorInfo tensor_info_output = input->clone()->set_tensor_shape(compute_winograd_filter_transform_shape(*input, winograd_info));
+
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(output, &tensor_info_output);
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
+    }
+
+    return Status{};
+}
+
+std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *output)
+{
+    ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
+    ARM_COMPUTE_UNUSED(output);
+
+    const unsigned int num_elems_processed_per_iteration_x = input->data_layout() == DataLayout::NCHW ? input->dimension(0) : 1;
+    const unsigned int num_elems_processed_per_iteration_y = input->dimension(1);
+    const unsigned int num_elems_read_per_iteration_z      = input->data_layout() == DataLayout::NCHW ? 1 : input->dimension(2);
+
+    Window win           = calculate_max_window(*input, Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y, num_elems_read_per_iteration_z));
+    Window win_collapsed = win.collapse(win, Window::DimZ);
+    return std::make_pair(Status{}, win_collapsed);
+}
+} // namespace
+
+void ClWinogradFilterTransformKernel::configure(const ClCompileContext &compile_context, ITensorInfo *src, ITensorInfo *dst, const WinogradInfo &winograd_info)
+{
+    ARM_COMPUTE_ERROR_ON_NULLPTR(src, dst);
+
+    // Output auto initialization if not yet initialized
+    auto_init_if_empty(*dst, src->clone()->set_tensor_shape(compute_winograd_filter_transform_shape(*src, winograd_info)));
+
+    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(src, dst, winograd_info));
+    auto padding_info = get_padding_info({ src, dst });
+
+    // Set build options
+    CLBuildOptions build_opts;
+    build_opts.add_option("-DSRC_DIM_Z=" + support::cpp11::to_string(src->dimension(2)));
+    build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(src->data_type()));
+    build_opts.add_option_if(winograd_info.kernel_size.height == 1, "-DWINOGRAD_FILTER_TRANSFORM_HORIZONTAL");
+    build_opts.add_option_if(winograd_info.kernel_size.width == 1, "-DWINOGRAD_FILTER_TRANSFORM_VERTICAL");
+    const Size2D kernel_size      = winograd_info.kernel_size;
+    const Size2D output_tile_size = winograd_info.output_tile_size;
+
+    // Create kernel
+    std::string kernel_name = "winograd_filter_transform_" + output_tile_size.to_string() + "_" + kernel_size.to_string() + "_" + lower_string(string_from_data_layout(src->data_layout()));
+    _kernel                 = create_kernel(compile_context, kernel_name, build_opts.options());
+
+    // Configure kernel window
+    auto win_config = validate_and_configure_window(src, dst);
+    ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
+    IClKernel::configure_internal(win_config.second);
+    ARM_COMPUTE_ERROR_ON(has_padding_changed(padding_info));
+}
+
+Status ClWinogradFilterTransformKernel::validate(const ITensorInfo *src, const ITensorInfo *dst, const WinogradInfo &winograd_info)
+{
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(src, dst, winograd_info));
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(src->clone().get(), dst->clone().get()).first);
+
+    return Status{};
+}
+
+void ClWinogradFilterTransformKernel::run_op(ITensorPack &tensors, const Window &window, cl::CommandQueue &queue)
+{
+    ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
+    ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(IClKernel::window(), window);
+
+    auto src = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC));
+    auto dst = utils::cast::polymorphic_downcast<ICLTensor *>(tensors.get_tensor(TensorType::ACL_DST));
+
+    // Setup output window
+    Window window_out;
+    window_out.use_tensor_dimensions(dst->info()->tensor_shape(), 0);
+
+    unsigned int idx = 0;
+    add_4D_tensor_argument(idx, src, window);
+    add_3D_tensor_argument(idx, dst, window_out);
+    enqueue(queue, *this, window, lws_hint());
+}
+} // namespace kernels
+} // namespace opencl
+} // namespace arm_compute
\ No newline at end of file
diff --git a/src/core/gpu/cl/kernels/ClWinogradFilterTransformKernel.h b/src/core/gpu/cl/kernels/ClWinogradFilterTransformKernel.h
new file mode 100644
index 0000000000..2bc2ceb36e
--- /dev/null
+++ b/src/core/gpu/cl/kernels/ClWinogradFilterTransformKernel.h
@@ -0,0 +1,78 @@
+/*
+ * Copyright (c) 2018-2021 Arm Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#ifndef ARM_COMPUTE_CL_WINOGRAD_FILTER_TRANSFORM_KERNEL_H
+#define ARM_COMPUTE_CL_WINOGRAD_FILTER_TRANSFORM_KERNEL_H
+
+#include "arm_compute/core/KernelDescriptors.h"
+#include "src/core/common/Macros.h"
+#include "src/core/gpu/cl/ClCompileContext.h"
+#include "src/core/gpu/cl/IClKernel.h"
+
+namespace arm_compute
+{
+namespace opencl
+{
+namespace kernels
+{
+/** Interface for the Winograd filter transform kernel. */
+class ClWinogradFilterTransformKernel : public IClKernel
+{
+public:
+    /** Default constructor */
+    ClWinogradFilterTransformKernel() = default;
+    ARM_COMPUTE_DISALLOW_COPY_ALLOW_MOVE(ClWinogradFilterTransformKernel);
+    /** Set the input and output tensor.
+     *
+     * @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), F(4x1, 5x1), F(1x4, 1x5)
+     *
+     *       Strides: only unit strides
+     *
+     * @param[in]  compile_context The compile context to be used.
+     * @param[in]  src             Source tensor info. 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: F16/F32.
+     * @param[out] dst             The output tensor info. 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
+     * @param[in]  winograd_info   Contains Winograd's information described in @ref WinogradInfo
+     */
+    void configure(const ClCompileContext &compile_context, ITensorInfo *src, ITensorInfo *dst, const WinogradInfo &winograd_info);
+    /** Static function to check if given info will lead to a valid configuration
+     *
+     * Similar to ClWinogradFilterTransformKernel::configure()
+     *
+     * @return a status
+     */
+    static Status validate(const ITensorInfo *src, const ITensorInfo *dst, const WinogradInfo &winograd_info);
+
+    // Inherited methods overridden:
+    void run_op(ITensorPack &tensors, const Window &window, cl::CommandQueue &queue) override;
+};
+} // namespace kernels
+} // namespace opencl
+} // namespace arm_compute
+#endif /*ARM_COMPUTE_CL_WINOGRAD_FILTER_TRANSFORM_KERNEL_H */
diff --git a/src/core/gpu/cl/kernels/ClWinogradInputTransformKernel.cpp b/src/core/gpu/cl/kernels/ClWinogradInputTransformKernel.cpp
new file mode 100644
index 0000000000..17f0eb9e2c
--- /dev/null
+++ b/src/core/gpu/cl/kernels/ClWinogradInputTransformKernel.cpp
@@ -0,0 +1,278 @@
+/*
+ * Copyright (c) 2018-2021 Arm Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#include "src/core/gpu/cl/kernels/ClWinogradInputTransformKernel.h"
+
+#include "arm_compute/core/CL/CLHelpers.h"
+#include "arm_compute/core/CL/CLKernelLibrary.h"
+#include "arm_compute/core/CL/ICLTensor.h"
+#include "arm_compute/core/CL/OpenCL.h"
+#include "arm_compute/core/Error.h"
+#include "arm_compute/core/Helpers.h"
+#include "arm_compute/core/Types.h"
+#include "arm_compute/core/Utils.h"
+#include "arm_compute/core/utils/misc/ShapeCalculator.h"
+#include "src/core/AccessWindowStatic.h"
+#include "src/core/CL/CLValidate.h"
+#include "src/core/helpers/AutoConfiguration.h"
+#include "src/core/helpers/WindowHelpers.h"
+#include "support/Cast.h"
+#include "support/StringSupport.h"
+
+namespace arm_compute
+{
+namespace opencl
+{
+namespace kernels
+{
+namespace
+{
+Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, const WinogradInfo &winograd_info)
+{
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32, DataType::F16);
+    ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input);
+
+    const PadStrideInfo conv_info        = winograd_info.convolution_info;
+    const Size2D        output_tile_size = winograd_info.output_tile_size;
+    const Size2D        kernel_size      = winograd_info.kernel_size;
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG(conv_info.stride().first != 1 || conv_info.stride().second != 1, "Winograd input transform only supports unit strides");
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG(!cl_winograd_convolution_layer_supported(output_tile_size, kernel_size, input->data_layout()), "Winograd input transform not supported");
+
+    ARM_COMPUTE_UNUSED(conv_info);
+    ARM_COMPUTE_UNUSED(output_tile_size);
+    ARM_COMPUTE_UNUSED(kernel_size);
+
+    // Validate configured output
+    if(output->total_size() != 0)
+    {
+        const TensorShape output_shape = misc::shape_calculator::compute_winograd_input_transform_shape(*input, winograd_info);
+
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(output->tensor_shape(), output_shape);
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
+    }
+
+    return Status{};
+}
+
+std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *output, const WinogradInfo &winograd_info)
+{
+    ARM_COMPUTE_UNUSED(output);
+    ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
+
+    bool   window_changed = false;
+    Window win            = calculate_max_window(*input, Steps(1, 1));
+
+    if(input->data_layout() == DataLayout::NCHW)
+    {
+        const PadStrideInfo conv_info        = winograd_info.convolution_info;
+        const Size2D        output_tile_size = winograd_info.output_tile_size;
+        const Size2D        kernel_size      = winograd_info.kernel_size;
+
+        unsigned int num_elems_read_per_iteration_x = output_tile_size.width + kernel_size.width - 1;
+        unsigned int num_elems_read_per_iteration_y = output_tile_size.height + kernel_size.height - 1;
+
+        AccessWindowRectangle input_access(input, -conv_info.pad_left(), -conv_info.pad_top(), num_elems_read_per_iteration_x, num_elems_read_per_iteration_y);
+        window_changed = update_window_and_padding(win, input_access);
+    }
+
+    Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{};
+    return std::make_pair(err, win);
+}
+} // namespace
+
+ClWinogradInputTransformKernel::ClWinogradInputTransformKernel()
+    : _border_size(0), _data_layout(DataLayout::UNKNOWN), _num_tiles_x(0), _num_tiles_y(0), _step_z(1)
+{
+}
+
+BorderSize ClWinogradInputTransformKernel::border_size() const
+{
+    return _border_size;
+}
+
+void ClWinogradInputTransformKernel::configure(const ClCompileContext &compile_context, ITensorInfo *src, ITensorInfo *dst, const WinogradInfo &winograd_info)
+{
+    ARM_COMPUTE_ERROR_ON_NULLPTR(src, dst);
+    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(src, dst, winograd_info));
+
+    auto padding_info = get_padding_info({ src, dst });
+
+    const PadStrideInfo conv_info        = winograd_info.convolution_info;
+    const Size2D        output_tile_size = winograd_info.output_tile_size;
+    const Size2D        kernel_size      = winograd_info.kernel_size;
+
+    _data_layout = src->data_layout();
+
+    const size_t idx_w = get_data_layout_dimension_index(_data_layout, DataLayoutDimension::WIDTH);
+    const size_t idx_h = get_data_layout_dimension_index(_data_layout, DataLayoutDimension::HEIGHT);
+
+    // Compute the number of output tiles along the x and y direction of size "output_tile_size"
+    const Size2D num_tiles = compute_winograd_convolution_tiles(Size2D(src->dimension(idx_w), src->dimension(idx_h)),
+                                                                kernel_size,
+                                                                output_tile_size,
+                                                                conv_info);
+
+    _num_tiles_x = num_tiles.width;
+    _num_tiles_y = num_tiles.height;
+
+    const TensorShape output_shape = misc::shape_calculator::compute_winograd_input_transform_shape(*src, winograd_info);
+
+    // Output auto initialization if not yet initialized
+    auto_init_if_empty(*dst, src->clone()->set_tensor_shape(output_shape));
+
+    ARM_COMPUTE_ERROR_ON(_num_tiles_x * _num_tiles_y != static_cast<int>(dst->dimension(1)));
+    const size_t total_batches = src->tensor_shape().total_size_upper(3);
+
+    CLBuildOptions build_opts;
+    if(_data_layout == DataLayout::NHWC)
+    {
+        build_opts.add_option("-DNHWC");
+        build_opts.add_option("-DSRC_WIDTH=" + support::cpp11::to_string(src->dimension(idx_w)));
+        build_opts.add_option("-DSRC_HEIGHT=" + support::cpp11::to_string(src->dimension(idx_h)));
+        build_opts.add_option("-DNUM_TILES_X=" + support::cpp11::to_string(_num_tiles_x));
+        build_opts.add_option("-DNUM_TILES_Y=" + support::cpp11::to_string(_num_tiles_y));
+        build_opts.add_option("-DPAD_LEFT=" + support::cpp11::to_string(conv_info.pad_left()));
+        build_opts.add_option("-DPAD_TOP=" + support::cpp11::to_string(conv_info.pad_top()));
+        build_opts.add_option("-DOUTPUT_TILE_W=" + support::cpp11::to_string(output_tile_size.width));
+        build_opts.add_option("-DOUTPUT_TILE_H=" + support::cpp11::to_string(output_tile_size.height));
+        build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(src->data_type()));
+        build_opts.add_option_if(winograd_info.kernel_size.height == 1, "-DWINOGRAD_INPUT_TRANSFORM_HORIZONTAL");
+        build_opts.add_option_if(winograd_info.kernel_size.width == 1, "-DWINOGRAD_INPUT_TRANSFORM_VERTICAL");
+    }
+    else
+    {
+        build_opts.add_option("-DNUM_TILES_X=" + support::cpp11::to_string(_num_tiles_x));
+        build_opts.add_option("-DPAD_LEFT=" + support::cpp11::to_string(conv_info.pad_left()));
+        build_opts.add_option("-DPAD_TOP=" + support::cpp11::to_string(conv_info.pad_top()));
+        build_opts.add_option("-DOUTPUT_TILE_W=" + support::cpp11::to_string(output_tile_size.width));
+        build_opts.add_option("-DOUTPUT_TILE_H=" + support::cpp11::to_string(output_tile_size.height));
+        build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(src->data_type()));
+        build_opts.add_option_if(winograd_info.kernel_size.height == 1, "-DWINOGRAD_INPUT_TRANSFORM_HORIZONTAL");
+        build_opts.add_option_if(winograd_info.kernel_size.width == 1, "-DWINOGRAD_INPUT_TRANSFORM_VERTICAL");
+        build_opts.add_option_if(total_batches > 1, "-DSRC_DEPTH=" + support::cpp11::to_string(src->dimension(2)));
+    }
+
+    // Create kernel
+    std::string kernel_name = "winograd_input_transform_" + output_tile_size.to_string() + "_" + kernel_size.to_string();
+
+    // Get the maximum dimension from the tile size
+    const unsigned int tile_max_dim = std::max(output_tile_size.width, output_tile_size.height);
+
+    // Check optimized kernel if output_dims == 2x2
+    if((tile_max_dim == 2) && (_data_layout == DataLayout::NCHW))
+    {
+        _step_z = (src->dimension(2) % 2) != 0 ? 1 : 2;
+    }
+
+    // Append stepz and data layout
+    kernel_name += "_stepz";
+    kernel_name += support::cpp11::to_string(_step_z);
+    kernel_name += "_" + lower_string(string_from_data_layout(_data_layout));
+
+    _kernel = create_kernel(compile_context, kernel_name, build_opts.options());
+
+    // Create window and update padding
+    auto win_config = validate_and_configure_window(src, dst, winograd_info);
+    ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
+    IClKernel::configure_internal(win_config.second, cl::NDRange(1, 1, 8));
+
+    _border_size = BorderSize(src->padding());
+
+    ARM_COMPUTE_ERROR_ON((src->data_layout() == DataLayout::NHWC) && has_padding_changed(padding_info));
+
+    _config_id = kernel_name;
+    _config_id += support::cpp11::to_string(src->dimension(0));
+    _config_id += "_";
+    _config_id += support::cpp11::to_string(src->dimension(1));
+    _config_id += "_";
+    _config_id += support::cpp11::to_string(src->dimension(2));
+    _config_id += "_";
+    _config_id += support::cpp11::to_string(conv_info.pad_left());
+    _config_id += "_";
+    _config_id += support::cpp11::to_string(conv_info.pad_top());
+    _config_id += "_";
+    _config_id += lower_string(string_from_data_layout(_data_layout));
+}
+
+Status ClWinogradInputTransformKernel::validate(const ITensorInfo *src, const ITensorInfo *dst, const WinogradInfo &winograd_info)
+{
+    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src, dst);
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(src, dst, winograd_info));
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(src->clone().get(), dst->clone().get(), winograd_info).first);
+    return Status{};
+}
+
+void ClWinogradInputTransformKernel::run_op(ITensorPack &tensors, const Window &window, cl::CommandQueue &queue)
+{
+    ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
+    ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(IKernel::window(), window);
+
+    auto src = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC));
+    auto dst = utils::cast::polymorphic_downcast<ICLTensor *>(tensors.get_tensor(TensorType::ACL_DST));
+
+    const size_t idx_w         = get_data_layout_dimension_index(_data_layout, DataLayoutDimension::WIDTH);
+    const size_t idx_h         = get_data_layout_dimension_index(_data_layout, DataLayoutDimension::HEIGHT);
+    const size_t idx_c         = get_data_layout_dimension_index(_data_layout, DataLayoutDimension::CHANNEL);
+    const size_t total_batches = window.shape().total_size_upper(3);
+
+    // Collapse window
+    Window window_collapsed = window.collapse_if_possible(IClKernel::window(), Window::DimZ);
+
+    if(_data_layout == DataLayout::NHWC)
+    {
+        Window slice = window_collapsed.first_slice_window_3D();
+        slice.set(1, Window::Dimension(0, _num_tiles_x * _num_tiles_y, 1));
+        slice.set(2, Window::Dimension(0, total_batches, 1));
+
+        unsigned int idx = 0;
+        add_4D_tensor_argument(idx, src, slice);
+        add_4D_tensor_argument(idx, dst, slice);
+        enqueue(queue, *this, slice, lws_hint());
+    }
+    else
+    {
+        Window slice = window_collapsed.first_slice_window_3D();
+        slice.set(idx_w, Window::Dimension(0, _num_tiles_x, 1));
+        slice.set(idx_h, Window::Dimension(0, _num_tiles_y, 1));
+
+        ARM_COMPUTE_ERROR_ON(((slice[idx_c].end() - slice[idx_c].start()) % _step_z) != 0);
+        slice.set(idx_c, Window::Dimension(slice[idx_c].start(), slice[idx_c].end(), _step_z));
+
+        unsigned int idx = 2 * num_arguments_per_3D_tensor();
+        _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(src->info()->strides_in_bytes()[3]));
+        _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(dst->info()->strides_in_bytes()[3]));
+
+        do
+        {
+            unsigned int idx = 0;
+            add_3D_tensor_argument(idx, src, slice);
+            add_3D_tensor_argument(idx, dst, slice);
+
+            enqueue(queue, *this, slice, lws_hint());
+        }
+        while(window_collapsed.slide_window_slice_3D(slice));
+    }
+}
+} // namespace kernels
+} // namespace opencl
+} // namespace arm_compute
\ No newline at end of file
diff --git a/src/core/gpu/cl/kernels/ClWinogradInputTransformKernel.h b/src/core/gpu/cl/kernels/ClWinogradInputTransformKernel.h
new file mode 100644
index 0000000000..76b45279a4
--- /dev/null
+++ b/src/core/gpu/cl/kernels/ClWinogradInputTransformKernel.h
@@ -0,0 +1,88 @@
+/*
+ * Copyright (c) 2018-2021 Arm Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#ifndef ARM_COMPUTE_CL_WINOGRAD_INPUT_TRANSFORM_KERNEL_H
+#define ARM_COMPUTE_CL_WINOGRAD_INPUT_TRANSFORM_KERNEL_H
+
+#include "arm_compute/core/KernelDescriptors.h"
+#include "src/core/common/Macros.h"
+#include "src/core/gpu/cl/ClCompileContext.h"
+#include "src/core/gpu/cl/IClKernel.h"
+
+namespace arm_compute
+{
+namespace opencl
+{
+namespace kernels
+{
+/** OpenCL kernel to perform Winograd input transform.*/
+class ClWinogradInputTransformKernel : public IClKernel
+{
+public:
+    /** Default constructor */
+    ClWinogradInputTransformKernel();
+    ARM_COMPUTE_DISALLOW_COPY_ALLOW_MOVE(ClWinogradInputTransformKernel);
+    /** Set the input and output of the kernel.
+     *
+     * @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), F(4x1, 5x1), F(1x4, 1x5)
+     *
+     *       Strides: only unit strides
+     *
+     * @param[in] compile_context The compile context to be used.
+     * @param[in] src             The input tensor info to transform. Data types supported: F16/F32
+     * @param[in] dst             The output tensor info. 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
+     * @param[in] winograd_info   Contains Winograd's information described in @ref WinogradInfo.
+     */
+    void configure(const ClCompileContext &compile_context, ITensorInfo *src, ITensorInfo *dst, const WinogradInfo &winograd_info);
+    /** Static function to check if given info will lead to a valid configuration
+     *
+     * Similar to ClWinogradInputTransformKernel::configure()
+     *
+     * @return a status
+     */
+    static Status validate(const ITensorInfo *src, const ITensorInfo *dst, const WinogradInfo &winograd_info);
+
+    // Inherited methods overridden:
+    void run_op(ITensorPack &tensors, const Window &window, cl::CommandQueue &queue) override;
+    BorderSize border_size() const override;
+
+private:
+    using WinogradKey = std::pair<std::pair<int, int>, std::pair<int, int>>;
+
+    BorderSize   _border_size;
+    DataLayout   _data_layout;
+    int          _num_tiles_x;
+    int          _num_tiles_y;
+    unsigned int _step_z;
+};
+} // namespace kernels
+} // namespace opencl
+} // namespace arm_compute
+#endif /*ARM_COMPUTE_CL_WINOGRAD_INPUT_TRANSFORM_KERNEL_H */
diff --git a/src/core/gpu/cl/kernels/ClWinogradOutputTransformKernel.cpp b/src/core/gpu/cl/kernels/ClWinogradOutputTransformKernel.cpp
new file mode 100644
index 0000000000..a6c05420ed
--- /dev/null
+++ b/src/core/gpu/cl/kernels/ClWinogradOutputTransformKernel.cpp
@@ -0,0 +1,263 @@
+/*
+ * Copyright (c) 2018-2021 Arm Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#include "src/core/gpu/cl/kernels/ClWinogradOutputTransformKernel.h"
+
+#include "arm_compute/core/CL/CLHelpers.h"
+#include "arm_compute/core/CL/CLKernelLibrary.h"
+#include "arm_compute/core/CL/ICLTensor.h"
+#include "arm_compute/core/Helpers.h"
+#include "arm_compute/core/IAccessWindow.h"
+#include "arm_compute/core/TensorInfo.h"
+#include "arm_compute/core/Types.h"
+#include "arm_compute/core/Utils.h"
+#include "arm_compute/core/Validate.h"
+#include "arm_compute/core/Window.h"
+#include "arm_compute/core/utils/misc/ShapeCalculator.h"
+#include "src/core/AccessWindowStatic.h"
+#include "src/core/CL/CLValidate.h"
+#include "src/core/helpers/AutoConfiguration.h"
+#include "src/core/helpers/WindowHelpers.h"
+#include "support/Cast.h"
+#include "support/StringSupport.h"
+
+#include <cmath>
+
+using namespace arm_compute::misc::shape_calculator;
+
+namespace arm_compute
+{
+namespace opencl
+{
+namespace kernels
+{
+namespace
+{
+Status validate_arguments(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, const WinogradInfo &winograd_info, const ActivationLayerInfo &act_info)
+{
+    ARM_COMPUTE_UNUSED(act_info);
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32, DataType::F16);
+    ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input);
+
+    ARM_COMPUTE_RETURN_ERROR_ON(output->data_layout() != winograd_info.output_data_layout);
+
+    const PadStrideInfo conv_info        = winograd_info.convolution_info;
+    const Size2D        output_tile_size = winograd_info.output_tile_size;
+    const Size2D        kernel_size      = winograd_info.kernel_size;
+    const Size2D        input_dimensions = winograd_info.input_dimensions;
+    const unsigned int  num_channels     = (winograd_info.kernel_size.width + winograd_info.output_tile_size.width - 1) * (winograd_info.kernel_size.height + winograd_info.output_tile_size.height - 1);
+
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG(!cl_winograd_convolution_layer_supported(output_tile_size, kernel_size, winograd_info.output_data_layout), "Winograd output transform not supported");
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG(input->dimension(2) != num_channels, "Wrong number of channels");
+
+    // Compute number of elements to process in the X and Y direction
+    // Compute the number of output tiles along the x and y direction of size "output_tile_size"
+    const Size2D num_tiles = compute_winograd_convolution_tiles(input_dimensions,
+                                                                kernel_size,
+                                                                output_tile_size,
+                                                                conv_info);
+
+    ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(1) != static_cast<unsigned int>((num_tiles.area())));
+
+    if(bias != nullptr)
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, bias);
+        ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(0) != bias->dimension(0));
+    }
+
+    // Checks performed when output is configured
+    if(output->total_size() != 0)
+    {
+        const TensorInfo tensor_info_output = input->clone()->set_tensor_shape(compute_winograd_output_transform_shape(*input, winograd_info));
+
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(output, &tensor_info_output);
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
+    }
+
+    return Status{};
+}
+
+std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *bias, ITensorInfo *output, const Size2D &output_tile_size)
+{
+    ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
+    ARM_COMPUTE_UNUSED(bias);
+
+    constexpr unsigned int num_elems_processed_per_iteration = 1;
+
+    Window win            = calculate_max_window(*input, Steps(num_elems_processed_per_iteration));
+    bool   window_changed = false;
+
+    if(output->data_layout() == DataLayout::NCHW)
+    {
+        const int output_static_window_end_x = ceil_to_multiple(output->dimension(0), output_tile_size.width);
+        const int output_static_window_end_y = ceil_to_multiple(output->dimension(1), output_tile_size.height);
+
+        AccessWindowRectangle input_access(input, 0, 0, num_elems_processed_per_iteration, num_elems_processed_per_iteration);
+        AccessWindowStatic    output_access(output, 0, 0, output_static_window_end_x, output_static_window_end_y);
+        window_changed = update_window_and_padding(win, input_access, output_access);
+    }
+
+    Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{};
+    return std::make_pair(err, win);
+}
+} // namespace
+
+void ClWinogradOutputTransformKernel::configure(const ClCompileContext &compile_context, ITensorInfo *src, ITensorInfo *bias, ITensorInfo *dst, const WinogradInfo &winograd_info,
+                                                const ActivationLayerInfo &act_info)
+{
+    ARM_COMPUTE_ERROR_ON_NULLPTR(src, dst);
+
+    // Output tensor auto initialization if not yet initialized
+    auto_init_if_empty(*dst, src->clone()->set_tensor_shape(compute_winograd_output_transform_shape(*src, winograd_info)));
+
+    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(src, bias, dst, winograd_info, act_info));
+
+    // Configure kernel window
+    auto win_config = validate_and_configure_window(src, bias, dst, winograd_info.output_tile_size);
+    ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
+    IClKernel::configure_internal(win_config.second);
+
+    auto padding_info = get_padding_info({ src, bias, dst });
+
+    _is_nhwc = winograd_info.output_data_layout == DataLayout::NHWC;
+
+    // Compute num_tiles_x
+    const Size2D        input_dimensions = winograd_info.input_dimensions;
+    const Size2D        kernel_size      = winograd_info.kernel_size;
+    const Size2D        output_tile_size = winograd_info.output_tile_size;
+    const PadStrideInfo conv_info        = winograd_info.convolution_info;
+    const int           idx_width        = get_data_layout_dimension_index(winograd_info.output_data_layout, DataLayoutDimension::WIDTH);
+    const int           idx_height       = get_data_layout_dimension_index(winograd_info.output_data_layout, DataLayoutDimension::HEIGHT);
+
+    // Compute the number of output tiles along the x and y direction of size "output_tile_size"
+    const Size2D num_tiles = compute_winograd_convolution_tiles(input_dimensions,
+                                                                kernel_size,
+                                                                output_tile_size,
+                                                                conv_info);
+    const size_t total_batches = dst->tensor_shape().total_size_upper(3);
+
+    // Set build options
+    CLBuildOptions build_opts;
+    build_opts.add_option("-DACTIVATION_TYPE=" + lower_string(string_from_activation_func(act_info.activation())));
+    build_opts.add_option_if(act_info.enabled(), "-DA_VAL=" + float_to_string_with_full_precision(act_info.a()));
+    build_opts.add_option_if(act_info.enabled(), "-DB_VAL=" + float_to_string_with_full_precision(act_info.b()));
+
+    if((output_tile_size.x() == 2) || (output_tile_size.x() == 1 && output_tile_size.y() == 2))
+    {
+        build_opts.add_option("-DVEC_SIZE=2");
+    }
+    else if((output_tile_size.x() == 4) || (output_tile_size.x() == 1 && output_tile_size.y() == 4))
+    {
+        build_opts.add_option("-DVEC_SIZE=4");
+    }
+
+    build_opts.add_option_if(bias != nullptr, std::string("-DHAS_BIAS"));
+    build_opts.add_option("-cl-fast-relaxed-math");
+    build_opts.add_option("-DN0=" + support::cpp11::to_string(win_config.second.x().step()));
+    build_opts.add_option("-DNUM_TILES_X=" + support::cpp11::to_string(num_tiles.width));
+    build_opts.add_option("-DOUTPUT_TILE_W=" + support::cpp11::to_string(output_tile_size.width));
+    build_opts.add_option("-DOUTPUT_TILE_H=" + support::cpp11::to_string(output_tile_size.height));
+    build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(src->data_type()));
+    build_opts.add_option("-DSRC_HEIGHT=" + support::cpp11::to_string(src->dimension(1)));
+    build_opts.add_option("-DDST_WIDTH=" + support::cpp11::to_string(dst->dimension(idx_width)));
+    build_opts.add_option("-DDST_HEIGHT=" + support::cpp11::to_string(dst->dimension(idx_height)));
+    build_opts.add_option_if(total_batches > 1, "-DSRC_DEPTH=" + support::cpp11::to_string(src->dimension(2)));
+    build_opts.add_option_if(winograd_info.kernel_size.height == 1, "-DWINOGRAD_OUTPUT_TRANSFORM_HORIZONTAL");
+    build_opts.add_option_if(winograd_info.kernel_size.width == 1, "-DWINOGRAD_OUTPUT_TRANSFORM_VERTICAL");
+
+    // Create kernel
+    std::string kernel_name = "winograd_output_transform_" + output_tile_size.to_string() + "_" + kernel_size.to_string() + "_" + lower_string(string_from_data_layout(winograd_info.output_data_layout));
+    _kernel                 = create_kernel(compile_context, kernel_name, build_opts.options());
+
+    // Set config_id for enabling LWS tuning
+    _config_id = kernel_name;
+    _config_id += "_";
+    _config_id += lower_string(string_from_data_type(src->data_type()));
+    _config_id += "_";
+    _config_id += support::cpp11::to_string(src->dimension(0));
+    _config_id += "_";
+    _config_id += support::cpp11::to_string(src->dimension(1));
+    _config_id += "_";
+    _config_id += support::cpp11::to_string(dst->dimension(0));
+    _config_id += "_";
+    _config_id += support::cpp11::to_string(dst->dimension(1));
+    _config_id += "_";
+    _config_id += lower_string(string_from_data_layout(winograd_info.output_data_layout));
+
+    ARM_COMPUTE_ERROR_ON(has_padding_changed(padding_info) && _is_nhwc);
+}
+
+Status ClWinogradOutputTransformKernel::validate(const ITensorInfo *src, const ITensorInfo *bias, const ITensorInfo *dst, const WinogradInfo &winograd_info, const ActivationLayerInfo &act_info)
+{
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(src, (bias != nullptr ? bias->clone().get() : nullptr), dst, winograd_info, act_info));
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(src->clone().get(), (bias != nullptr ? bias->clone().get() : nullptr), dst->clone().get(), winograd_info.output_tile_size).first);
+    return Status{};
+}
+
+void ClWinogradOutputTransformKernel::run_op(ITensorPack &tensors, const Window &window, cl::CommandQueue &queue)
+{
+    ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
+    ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(IClKernel::window(), window);
+
+    auto src  = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC_0));
+    auto bias = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC_1));
+    auto dst  = utils::cast::polymorphic_downcast<ICLTensor *>(tensors.get_tensor(TensorType::ACL_DST));
+
+    // Collapse window
+    Window window_collapsed = window.collapse_if_possible(IClKernel::window(), Window::DimZ);
+
+    // Get initial windows
+    Window slice = window_collapsed.first_slice_window_4D();
+    slice.set(Window::DimZ, Window::Dimension(0, 1, 1));
+
+    // Setup output slice
+    Window slice_out(slice);
+    slice_out.set(Window::DimX, Window::Dimension(0, 0, 0));
+    slice_out.set(Window::DimY, Window::Dimension(0, 0, 0));
+
+    if(bias != nullptr)
+    {
+        unsigned int idx1 = 2 * num_arguments_per_4D_tensor();
+        Window       slice_biases;
+        slice_biases.use_tensor_dimensions(bias->info()->tensor_shape());
+        add_1D_tensor_argument(idx1, bias, slice_biases);
+    }
+
+    if(_is_nhwc)
+    {
+        unsigned int idx2 = 2 * num_arguments_per_4D_tensor() + ((bias != nullptr) ? num_arguments_per_1D_tensor() : 0);
+        _kernel.setArg(idx2, static_cast<int>(dst->info()->total_size() - dst->info()->strides_in_bytes().y()));
+    }
+
+    do
+    {
+        unsigned int idx = 0;
+        add_4D_tensor_argument(idx, src, slice);
+        add_4D_tensor_argument(idx, dst, slice_out);
+        enqueue(queue, *this, slice, lws_hint());
+    }
+    while(window.slide_window_slice_3D(slice) && window.slide_window_slice_3D(slice_out));
+}
+} // namespace kernels
+} // namespace opencl
+} // namespace arm_compute
diff --git a/src/core/gpu/cl/kernels/ClWinogradOutputTransformKernel.h b/src/core/gpu/cl/kernels/ClWinogradOutputTransformKernel.h
new file mode 100644
index 0000000000..48b27e658c
--- /dev/null
+++ b/src/core/gpu/cl/kernels/ClWinogradOutputTransformKernel.h
@@ -0,0 +1,87 @@
+/*
+ * Copyright (c) 2018-2021 Arm Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#ifndef ARM_COMPUTE_CL_WINOGRAD_OUTPUT_TRANSFORM_KERNEL_H
+#define ARM_COMPUTE_CL_WINOGRAD_OUTPUT_TRANSFORM_KERNEL_H
+
+#include "arm_compute/core/KernelDescriptors.h"
+#include "src/core/common/Macros.h"
+#include "src/core/gpu/cl/ClCompileContext.h"
+#include "src/core/gpu/cl/IClKernel.h"
+
+namespace arm_compute
+{
+namespace opencl
+{
+namespace kernels
+{
+/** Interface for the Winograd output transform kernel. */
+class ClWinogradOutputTransformKernel : public IClKernel
+{
+public:
+    /** Default constructor */
+    ClWinogradOutputTransformKernel() = default;
+    ARM_COMPUTE_DISALLOW_COPY_ALLOW_MOVE(ClWinogradOutputTransformKernel);
+    /** Set the input and output tensor.
+     *
+     * @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), F(4x1, 5x1), F(1x4, 1x5)
+     *
+     *       Strides: only unit strides
+     *
+     * @param[in]  compile_context The compile context to be used.
+     * @param[in]  src             Source tensor info with shape [C, N, K, batches]. Data types supported: F16/F32.
+     * @param[in]  bias            Biases tensor info. Shared biases supported. Biases are 1D tensor with dimensions [OFM]. It can be a nullptr. Data type supported: as @p src
+     * @param[out] dst             The output tensor info. The shape for this tensor can be calculated using the utility function @p compute_winograd_output_transform_shape. Data types supported: Same as @p src
+     * @param[in]  winograd_info   Contains Winograd's information described in @ref WinogradInfo
+     * @param[in]  act_info        (Optional) Activation layer information in case of a fused activation.
+     */
+    void configure(const ClCompileContext &compile_context, ITensorInfo *src, ITensorInfo *bias, ITensorInfo *dst, const WinogradInfo &winograd_info,
+                   const ActivationLayerInfo &act_info = ActivationLayerInfo());
+
+    /** Static function to check if given info will lead to a valid configuration
+     *
+     * Similar to ClWinogradOutputTransformKernel::configure()
+     *
+     * @return a status
+     */
+    static Status validate(const ITensorInfo *src, const ITensorInfo *bias, const ITensorInfo *dst, const WinogradInfo &winograd_info, const ActivationLayerInfo &act_info = ActivationLayerInfo());
+
+    // Inherited methods overridden:
+    void run_op(ITensorPack &tensors, const Window &window, cl::CommandQueue &queue) override;
+
+private:
+    using WinogradKey = std::pair<std::pair<int, int>, std::pair<int, int>>;
+
+    bool _is_nhwc{ false };
+};
+} // namespace kernels
+} // namespace opencl
+} // namespace arm_compute
+#endif /*ARM_COMPUTE_CL_WINOGRAD_OUTPUT_TRANSFORM_KERNEL_H */