Port CLConvertFullyConnectedWeights to new API

* Replace ICLKernel by IClKernel in other unrelated kernels

Resolves partially: COMPMID-4187

Signed-off-by: Teresa Charlin <teresa.charlinreyes@arm.com>
Change-Id: I173b8f2ac645dbfd7d412f4b058c5c9655c229ee
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/5402
Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>

18 files changed, 373 insertions, 234 deletions

diff --git a/Android.bp b/Android.bp
index dd8509ebcd..7e2da8fe53 100644
--- a/Android.bp
+++ b/Android.bp
@@ -89,7 +89,6 @@ cc_library_static {
         "src/core/CL/kernels/CLChannelShuffleLayerKernel.cpp",
         "src/core/CL/kernels/CLCol2ImKernel.cpp",
         "src/core/CL/kernels/CLComparisonKernel.cpp",
-        "src/core/CL/kernels/CLConvertFullyConnectedWeightsKernel.cpp",
         "src/core/CL/kernels/CLDeconvolutionLayerUpsampleKernel.cpp",
         "src/core/CL/kernels/CLDeconvolutionReshapeOutputKernel.cpp",
         "src/core/CL/kernels/CLDepthConvertLayerKernel.cpp",
@@ -356,6 +355,7 @@ cc_library_static {
         "src/core/cpu/kernels/sub/neon/qsymm16.cpp",
         "src/core/gpu/cl/kernels/ClActivationKernel.cpp",
         "src/core/gpu/cl/kernels/ClBatchConcatenateKernel.cpp",
+        "src/core/gpu/cl/kernels/ClConvertFullyConnectedWeightsKernel.cpp",
         "src/core/gpu/cl/kernels/ClCopyKernel.cpp",
         "src/core/gpu/cl/kernels/ClCropKernel.cpp",
         "src/core/gpu/cl/kernels/ClDepthConcatenateKernel.cpp",
@@ -647,6 +647,7 @@ cc_library_static {
         "src/runtime/gpu/cl/operators/ClActivation.cpp",
         "src/runtime/gpu/cl/operators/ClAdd.cpp",
         "src/runtime/gpu/cl/operators/ClConcatenate.cpp",
+        "src/runtime/gpu/cl/operators/ClConvertFullyConnectedWeights.cpp",
         "src/runtime/gpu/cl/operators/ClCopy.cpp",
         "src/runtime/gpu/cl/operators/ClCrop.cpp",
         "src/runtime/gpu/cl/operators/ClDequantization.cpp",
diff --git a/arm_compute/runtime/CL/functions/CLConvertFullyConnectedWeights.h b/arm_compute/runtime/CL/functions/CLConvertFullyConnectedWeights.h
index 75a3d3213e..8892dbad6c 100644
--- a/arm_compute/runtime/CL/functions/CLConvertFullyConnectedWeights.h
+++ b/arm_compute/runtime/CL/functions/CLConvertFullyConnectedWeights.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2018-2020 Arm Limited.
+ * Copyright (c) 2018-2021 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -26,20 +26,33 @@
 
 #include "arm_compute/core/CL/CLKernelLibrary.h"
 #include "arm_compute/runtime/CL/CLTensor.h"
-#include "arm_compute/runtime/CL/ICLSimpleFunction.h"
+#include "arm_compute/runtime/IFunction.h"
 #include "arm_compute/runtime/ITransformWeights.h"
 
+#include <memory>
+
 namespace arm_compute
 {
 class CLCompileContext;
-class CLConvertFullyConnectedWeightsKernel;
 class ICLTensor;
 class ITensorInfo;
 
-/** Basic function to run @ref CLConvertFullyConnectedWeightsKernel. */
-class CLConvertFullyConnectedWeights : public ICLSimpleFunction
+/** Basic function to run an @ref opencl::kernels::ClConvertFullyConnectedWeightsKernel. */
+class CLConvertFullyConnectedWeights : public IFunction
 {
 public:
+    /** Constructor */
+    CLConvertFullyConnectedWeights();
+    /** Destructor */
+    ~CLConvertFullyConnectedWeights();
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    CLConvertFullyConnectedWeights(const CLConvertFullyConnectedWeights &) = delete;
+    /** Default move constructor */
+    CLConvertFullyConnectedWeights(CLConvertFullyConnectedWeights &&) = default;
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    CLConvertFullyConnectedWeights &operator=(const CLConvertFullyConnectedWeights &) = delete;
+    /** Default move assignment operator */
+    CLConvertFullyConnectedWeights &operator=(CLConvertFullyConnectedWeights &&) = default;
     /** Initialize the function.
      *
      * @param[in]  input                Source weights tensor to convert. Must be 2 dimensional. Data types supported: All.
@@ -69,11 +82,18 @@ public:
      * @param[in] data_layout          The data layout the weights have been trained in.
      */
     static Status validate(const ITensorInfo *input, const ITensorInfo *output, const TensorShape &original_input_shape, DataLayout data_layout);
+
+    // Inherited methods overridden:
+    void run() override;
+
+private:
+    struct Impl;
+    std::unique_ptr<Impl> _impl;
 };
 
 namespace weights_transformations
 {
-/** Basic function to run @ref CLConvertFullyConnectedWeightsKernel. */
+/** Basic function to manage @ref CLConvertFullyConnectedWeights. */
 class CLConvertFullyConnectedWeightsManaged : public ITransformWeights
 {
 public:
diff --git a/docs/00_introduction.dox b/docs/00_introduction.dox
index 112254e82a..9f6af6da50 100644
--- a/docs/00_introduction.dox
+++ b/docs/00_introduction.dox
@@ -1116,7 +1116,7 @@ v18.05 Public major release
  - Added support for dilated convolutions in @ref NEConvolutionLayer and @ref CLConvolutionLayer.
  - New OpenCL kernels / functions:
     - @ref CLChannelShuffleLayer / @ref CLChannelShuffleLayerKernel
-    - @ref CLConvertFullyConnectedWeightsKernel / @ref CLConvertFullyConnectedWeights
+    - CLConvertFullyConnectedWeightsKernel / @ref CLConvertFullyConnectedWeights
     - @ref CLCopy / CLCopyKernel
     - @ref CLLSTMLayer
     - @ref CLRNNLayer
diff --git a/src/core/CL/CLKernels.h b/src/core/CL/CLKernels.h
index 9676fa6018..f29c768fa3 100644
--- a/src/core/CL/CLKernels.h
+++ b/src/core/CL/CLKernels.h
@@ -33,7 +33,6 @@
 #include "src/core/CL/kernels/CLChannelShuffleLayerKernel.h"
 #include "src/core/CL/kernels/CLCol2ImKernel.h"
 #include "src/core/CL/kernels/CLComparisonKernel.h"
-#include "src/core/CL/kernels/CLConvertFullyConnectedWeightsKernel.h"
 #include "src/core/CL/kernels/CLDeconvolutionLayerUpsampleKernel.h"
 #include "src/core/CL/kernels/CLDeconvolutionReshapeOutputKernel.h"
 #include "src/core/CL/kernels/CLDepthConvertLayerKernel.h"
diff --git a/src/core/CL/kernels/CLConvertFullyConnectedWeightsKernel.cpp b/src/core/CL/kernels/CLConvertFullyConnectedWeightsKernel.cpp
deleted file mode 100644
index dcf4e6662e..0000000000
--- a/src/core/CL/kernels/CLConvertFullyConnectedWeightsKernel.cpp
+++ /dev/null
@@ -1,121 +0,0 @@
-/*
- * Copyright (c) 2018-2020 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/CL/kernels/CLConvertFullyConnectedWeightsKernel.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 "src/core/CL/CLValidate.h"
-#include "src/core/helpers/AutoConfiguration.h"
-#include "src/core/helpers/WindowHelpers.h"
-#include "support/StringSupport.h"
-
-namespace arm_compute
-{
-CLConvertFullyConnectedWeightsKernel::CLConvertFullyConnectedWeightsKernel()
-    : _input(nullptr), _output(nullptr)
-{
-}
-
-void CLConvertFullyConnectedWeightsKernel::configure(const ICLTensor *input, ICLTensor *output, const TensorShape &original_input_shape,
-                                                     DataLayout data_layout)
-{
-    configure(CLKernelLibrary::get().get_compile_context(), input, output, original_input_shape, data_layout);
-}
-
-void CLConvertFullyConnectedWeightsKernel::configure(const CLCompileContext &compile_context, const ICLTensor *input, ICLTensor *output, const TensorShape &original_input_shape,
-                                                     DataLayout data_layout)
-{
-    ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
-
-    // Output tensor auto initialisation if not yet initialized
-    auto_init_if_empty(*output->info(), *input->info()->clone());
-
-    auto padding_info = get_padding_info({ input, output });
-
-    ARM_COMPUTE_ERROR_THROW_ON(CLConvertFullyConnectedWeightsKernel::validate(input->info(), output->info(), original_input_shape, data_layout));
-
-    _input  = input;
-    _output = output;
-
-    const DataLayout input_data_layout = (data_layout == DataLayout::NCHW) ? DataLayout::NHWC : DataLayout::NCHW;
-
-    const int width_idx   = get_data_layout_dimension_index(input_data_layout, DataLayoutDimension::WIDTH);
-    const int height_idx  = get_data_layout_dimension_index(input_data_layout, DataLayoutDimension::HEIGHT);
-    const int channel_idx = get_data_layout_dimension_index(input_data_layout, DataLayoutDimension::CHANNEL);
-
-    const unsigned int num_elems_per_input_plane = original_input_shape[width_idx] * original_input_shape[height_idx];
-    const unsigned int num_channels              = original_input_shape[channel_idx];
-
-    const unsigned int factor_1 = (data_layout == DataLayout::NCHW) ? num_elems_per_input_plane : num_channels;
-    const unsigned int factor_2 = (data_layout == DataLayout::NCHW) ? num_channels : num_elems_per_input_plane;
-
-    // Set build options
-    CLBuildOptions build_opts;
-    build_opts.add_option("-DDATA_TYPE=" + get_cl_unsigned_type_from_element_size(input->info()->element_size()));
-    build_opts.add_option("-DFACTOR_1=" + support::cpp11::to_string(factor_1));
-    build_opts.add_option("-DFACTOR_2=" + support::cpp11::to_string(factor_2));
-
-    // Create kernel
-    _kernel = create_kernel(compile_context, "convert_fc_weights", build_opts.options());
-
-    // Configure kernel window
-    Window win = calculate_max_window(*input->info(), Steps());
-    ICLKernel::configure_internal(win);
-
-    ARM_COMPUTE_ERROR_ON(has_padding_changed(padding_info));
-}
-
-Status CLConvertFullyConnectedWeightsKernel::validate(const ITensorInfo *input, const ITensorInfo *output, const TensorShape &original_input_shape,
-                                                      DataLayout data_layout)
-{
-    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output);
-    ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input);
-    ARM_COMPUTE_RETURN_ERROR_ON(input->data_type() == DataType::UNKNOWN);
-    ARM_COMPUTE_RETURN_ERROR_ON(input->num_dimensions() != 2);
-    ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(1) != original_input_shape.total_size_lower(3));
-    ARM_COMPUTE_RETURN_ERROR_ON(data_layout == DataLayout::UNKNOWN);
-
-    // Checks performed when output is configured
-    if(output->total_size() != 0)
-    {
-        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
-        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input, output);
-    }
-
-    return Status{};
-}
-
-void CLConvertFullyConnectedWeightsKernel::run(const Window &window, cl::CommandQueue &queue)
-{
-    ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
-    ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window);
-
-    unsigned int idx = 0;
-    add_2D_tensor_argument(idx, _input, window);
-    add_2D_tensor_argument(idx, _output, window);
-    enqueue(queue, *this, window, lws_hint());
-}
-} // namespace arm_compute
diff --git a/src/core/CL/kernels/CLConvertFullyConnectedWeightsKernel.h b/src/core/CL/kernels/CLConvertFullyConnectedWeightsKernel.h
deleted file mode 100644
index d1da793df2..0000000000
--- a/src/core/CL/kernels/CLConvertFullyConnectedWeightsKernel.h
+++ /dev/null
@@ -1,90 +0,0 @@
-/*
- * Copyright (c) 2018-2020 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_CLCONVERTFULLYCONNECTEDWEIGHTSKERNEL_H
-#define ARM_COMPUTE_CLCONVERTFULLYCONNECTEDWEIGHTSKERNEL_H
-
-#include "src/core/CL/ICLKernel.h"
-
-namespace arm_compute
-{
-class ICLTensor;
-
-/** Interface to convert the 2D Fully Connected weights from NCHW to NHWC or vice versa.
- *
- * @note This function can be applied to the 2D weights used by a Fully Connected layer if:
- *       - It follows a Convolution layer
- *       - The data layout used by the network does not match the one the model has been trained in.
- *
- * @note This function assumes the weights are already reshaped (transposed)
- */
-class CLConvertFullyConnectedWeightsKernel : public ICLKernel
-{
-public:
-    /** Default constructor */
-    CLConvertFullyConnectedWeightsKernel();
-    /** Prevent instances of this class from being copied (As this class contains pointers) */
-    CLConvertFullyConnectedWeightsKernel(const CLConvertFullyConnectedWeightsKernel &) = delete;
-    /** Prevent instances of this class from being copied (As this class contains pointers) */
-    CLConvertFullyConnectedWeightsKernel &operator=(const CLConvertFullyConnectedWeightsKernel &) = delete;
-    /** Allow instances of this class to be moved */
-    CLConvertFullyConnectedWeightsKernel(CLConvertFullyConnectedWeightsKernel &&) = default;
-    /** Allow instances of this class to be moved */
-    CLConvertFullyConnectedWeightsKernel &operator=(CLConvertFullyConnectedWeightsKernel &&) = default;
-    /** Default destructor */
-    ~CLConvertFullyConnectedWeightsKernel() = default;
-    /** Set the input and output tensor.
-     *
-     * @param[in]  input                Source weights tensor to convert. Must be 2 dimensional. Data types supported: All.
-     * @param[out] output               The converted weights tensor. Shape and Data Type: Same as @p input.
-     * @param[in]  original_input_shape Shape of the original input tensor (the one entering fully connected layer).
-     * @param[in]  data_layout          The data layout the weights have been trained in.
-     */
-    void configure(const ICLTensor *input, ICLTensor *output, const TensorShape &original_input_shape, DataLayout data_layout);
-    /** Set the input and output tensor.
-     *
-     * @param[in]  compile_context      The compile context to be used.
-     * @param[in]  input                Source weights tensor to convert. Must be 2 dimensional. Data types supported: All.
-     * @param[out] output               The converted weights tensor. Shape and Data Type: Same as @p input.
-     * @param[in]  original_input_shape Shape of the original input tensor (the one entering fully connected layer).
-     * @param[in]  data_layout          The data layout the weights have been trained in.
-     */
-    void configure(const CLCompileContext &compile_context, const ICLTensor *input, ICLTensor *output, const TensorShape &original_input_shape, DataLayout data_layout);
-    /** Static function to check if given info will lead to a valid configuration of @ref CLConvertFullyConnectedWeightsKernel
-     *
-     * @param[in] input                Source weights tensor info to convert. Must be 2 dimensional. Data types supported: All.
-     * @param[in] output               The converted weights tensor info. Shape and Data Type: Same as @p input.
-     * @param[in] original_input_shape Shape of the original input tensor (the one entering fully connected layer).
-     * @param[in] data_layout          The data layout the weights have been trained in.
-     */
-    static Status validate(const ITensorInfo *input, const ITensorInfo *output, const TensorShape &original_input_shape, DataLayout data_layout);
-
-    // Inherited methods overridden:
-    void run(const Window &window, cl::CommandQueue &queue) override;
-
-private:
-    const ICLTensor *_input;
-    ICLTensor       *_output;
-};
-} // namespace arm_compute
-#endif /*ARM_COMPUTE_CLCONVERTFULLYCONNECTEDWEIGHTSKERNEL_H */
diff --git a/src/core/gpu/cl/kernels/ClConvertFullyConnectedWeightsKernel.cpp b/src/core/gpu/cl/kernels/ClConvertFullyConnectedWeightsKernel.cpp
new file mode 100644
index 0000000000..49f2f68a76
--- /dev/null
+++ b/src/core/gpu/cl/kernels/ClConvertFullyConnectedWeightsKernel.cpp
@@ -0,0 +1,119 @@
+/*
+ * 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/ClConvertFullyConnectedWeightsKernel.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/Utils.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
+{
+void ClConvertFullyConnectedWeightsKernel::configure(const CLCompileContext &compile_context, const ITensorInfo *src, ITensorInfo *dst, const TensorShape &original_src_shape,
+                                                     DataLayout data_layout)
+{
+    ARM_COMPUTE_ERROR_ON_NULLPTR(src, dst);
+
+    // Output tensor auto initialisation if not yet initialized
+    auto_init_if_empty(*dst, *src->clone());
+
+    auto padding_info = get_padding_info({ src, dst });
+
+    ARM_COMPUTE_ERROR_THROW_ON(ClConvertFullyConnectedWeightsKernel::validate(src, dst, original_src_shape, data_layout));
+
+    const DataLayout src_data_layout = (data_layout == DataLayout::NCHW) ? DataLayout::NHWC : DataLayout::NCHW;
+
+    const int width_idx   = get_data_layout_dimension_index(src_data_layout, DataLayoutDimension::WIDTH);
+    const int height_idx  = get_data_layout_dimension_index(src_data_layout, DataLayoutDimension::HEIGHT);
+    const int channel_idx = get_data_layout_dimension_index(src_data_layout, DataLayoutDimension::CHANNEL);
+
+    const unsigned int num_elems_per_src_plane = original_src_shape[width_idx] * original_src_shape[height_idx];
+    const unsigned int num_channels            = original_src_shape[channel_idx];
+
+    const unsigned int factor_1 = (data_layout == DataLayout::NCHW) ? num_elems_per_src_plane : num_channels;
+    const unsigned int factor_2 = (data_layout == DataLayout::NCHW) ? num_channels : num_elems_per_src_plane;
+
+    // Set build options
+    CLBuildOptions build_opts;
+    build_opts.add_option("-DDATA_TYPE=" + get_cl_unsigned_type_from_element_size(src->element_size()));
+    build_opts.add_option("-DFACTOR_1=" + support::cpp11::to_string(factor_1));
+    build_opts.add_option("-DFACTOR_2=" + support::cpp11::to_string(factor_2));
+
+    // Create kernel
+    _kernel = create_kernel(compile_context, "convert_fc_weights", build_opts.options());
+
+    // Configure kernel window
+    Window win = calculate_max_window(*src, Steps());
+    ICLKernel::configure_internal(win);
+
+    ARM_COMPUTE_ERROR_ON(has_padding_changed(padding_info));
+}
+
+Status ClConvertFullyConnectedWeightsKernel::validate(const ITensorInfo *src, const ITensorInfo *dst, const TensorShape &original_src_shape,
+                                                      DataLayout data_layout)
+{
+    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src, dst);
+    ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(src);
+    ARM_COMPUTE_RETURN_ERROR_ON(src->data_type() == DataType::UNKNOWN);
+    ARM_COMPUTE_RETURN_ERROR_ON(src->num_dimensions() != 2);
+    ARM_COMPUTE_RETURN_ERROR_ON(src->dimension(1) != original_src_shape.total_size_lower(3));
+    ARM_COMPUTE_RETURN_ERROR_ON(data_layout == DataLayout::UNKNOWN);
+
+    // Checks performed when dst is configured
+    if(dst->total_size() != 0)
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src, dst);
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(src, dst);
+    }
+
+    return Status{};
+}
+
+void ClConvertFullyConnectedWeightsKernel::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);
+
+    const 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));
+    ARM_COMPUTE_ERROR_ON_NULLPTR(src, dst);
+
+    unsigned int idx = 0;
+    add_2D_tensor_argument(idx, src, window);
+    add_2D_tensor_argument(idx, dst, window);
+    enqueue(queue, *this, window, lws_hint());
+}
+} // namespace kernels
+} // namespace opencl
+} // namespace arm_compute
diff --git a/src/core/gpu/cl/kernels/ClConvertFullyConnectedWeightsKernel.h b/src/core/gpu/cl/kernels/ClConvertFullyConnectedWeightsKernel.h
new file mode 100644
index 0000000000..11ab4d2a0d
--- /dev/null
+++ b/src/core/gpu/cl/kernels/ClConvertFullyConnectedWeightsKernel.h
@@ -0,0 +1,76 @@
+/*
+ * 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_CLCONVERTFULLYCONNECTEDWEIGHTSKERNEL_H
+#define ARM_COMPUTE_CLCONVERTFULLYCONNECTEDWEIGHTSKERNEL_H
+
+#include "src/core/common/Macros.h"
+#include "src/core/gpu/cl/ClCompileContext.h"
+#include "src/core/gpu/cl/IClKernel.h"
+
+namespace arm_compute
+{
+class ICLTensor;
+
+/** Interface to convert the 2D Fully Connected weights from NCHW to NHWC or vice versa.
+ *
+ * @note This function can be applied to the 2D weights used by a Fully Connected layer if:
+ *       - It follows a Convolution layer
+ *       - The data layout used by the network does not match the one the model has been trained in.
+ *
+ * @note This function assumes the weights are already reshaped (transposed)
+ */
+namespace opencl
+{
+namespace kernels
+{
+class ClConvertFullyConnectedWeightsKernel : public IClKernel
+{
+public:
+    ClConvertFullyConnectedWeightsKernel() = default;
+    ARM_COMPUTE_DISALLOW_COPY_ALLOW_MOVE(ClConvertFullyConnectedWeightsKernel);
+    /** Set the src and dst tensor.
+     *
+     * @param[in]  compile_context    The compile context to be used.
+     * @param[in]  src                Source weights tensor info to convert. Must be 2 dimensional. Data types supported: All.
+     * @param[out] dst                The converted weights tensor info. Shape and Data Type: Same as @p src.
+     * @param[in]  original_src_shape Shape of the original src tensor (the one entering fully connected layer).
+     * @param[in]  data_layout        The data layout the weights have been trained in.
+     */
+    void configure(const CLCompileContext &compile_context, const ITensorInfo *src, ITensorInfo *dst, const TensorShape &original_src_shape, DataLayout data_layout);
+    /** Static function to check if given info will lead to a valid configuration of @ref ClConvertFullyConnectedWeightsKernel
+     *
+     * @param[in] src                Source weights tensor info to convert. Must be 2 dimensional. Data types supported: All.
+     * @param[in] dst                The converted weights tensor info. Shape and Data Type: Same as @p src.
+     * @param[in] original_src_shape Shape of the original src tensor (the one entering fully connected layer).
+     * @param[in] data_layout        The data layout the weights have been trained in.
+     */
+    static Status validate(const ITensorInfo *src, const ITensorInfo *dst, const TensorShape &original_src_shape, DataLayout data_layout);
+
+    // 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_CLCONVERTFULLYCONNECTEDWEIGHTSKERNEL_H */
diff --git a/src/core/gpu/cl/kernels/ClCopyKernel.h b/src/core/gpu/cl/kernels/ClCopyKernel.h
index 9d5457d72d..b1b9672bcb 100644
--- a/src/core/gpu/cl/kernels/ClCopyKernel.h
+++ b/src/core/gpu/cl/kernels/ClCopyKernel.h
@@ -35,7 +35,7 @@ namespace opencl
 namespace kernels
 {
 /** OpenCL kernel to perform a copy between two tensors */
-class ClCopyKernel : public ICLKernel
+class ClCopyKernel : public IClKernel
 {
 public:
     ClCopyKernel() = default;
diff --git a/src/core/gpu/cl/kernels/ClCropKernel.h b/src/core/gpu/cl/kernels/ClCropKernel.h
index 92f94d471b..ec0f8e58da 100644
--- a/src/core/gpu/cl/kernels/ClCropKernel.h
+++ b/src/core/gpu/cl/kernels/ClCropKernel.h
@@ -35,7 +35,7 @@ namespace opencl
 namespace kernels
 {
 /** OpenCL kernel to perform a copy between two tensors */
-class ClCropKernel : public ICLKernel
+class ClCropKernel : public IClKernel
 {
 public:
     ClCropKernel() = default;
diff --git a/src/core/gpu/cl/kernels/ClDepthConcatenateKernel.h b/src/core/gpu/cl/kernels/ClDepthConcatenateKernel.h
index 5acfb33199..103ef00695 100644
--- a/src/core/gpu/cl/kernels/ClDepthConcatenateKernel.h
+++ b/src/core/gpu/cl/kernels/ClDepthConcatenateKernel.h
@@ -37,7 +37,7 @@ namespace kernels
 /** Interface for the depth concatenate kernel.
  *  The src tensor will be concatenated into the dst tensor.
  */
-class ClDepthConcatenateKernel : public ICLKernel
+class ClDepthConcatenateKernel : public IClKernel
 {
 public:
     /** Default constructor */
diff --git a/src/core/gpu/cl/kernels/ClDirectConvolutionKernel.h b/src/core/gpu/cl/kernels/ClDirectConvolutionKernel.h
index ff2f5619db..384b561003 100644
--- a/src/core/gpu/cl/kernels/ClDirectConvolutionKernel.h
+++ b/src/core/gpu/cl/kernels/ClDirectConvolutionKernel.h
@@ -36,7 +36,7 @@ namespace kernels
 {
 /** Interface for the  direct convolution kernel.
  */
-class ClDirectConvolutionKernel : public ICLKernel
+class ClDirectConvolutionKernel : public IClKernel
 {
 public:
     ClDirectConvolutionKernel() = default;
diff --git a/src/core/gpu/cl/kernels/ClFillKernel.h b/src/core/gpu/cl/kernels/ClFillKernel.h
index 136c1243fe..b439eac0de 100644
--- a/src/core/gpu/cl/kernels/ClFillKernel.h
+++ b/src/core/gpu/cl/kernels/ClFillKernel.h
@@ -35,7 +35,7 @@ namespace opencl
 namespace kernels
 {
 /** Interface for filling the planes of a tensor */
-class ClFillKernel : public ICLKernel
+class ClFillKernel : public IClKernel
 {
 public:
     ClFillKernel() = default;
diff --git a/src/core/gpu/cl/kernels/ClPixelWiseMultiplicationKernel.h b/src/core/gpu/cl/kernels/ClPixelWiseMultiplicationKernel.h
index 64b6aa1eda..5889b84938 100644
--- a/src/core/gpu/cl/kernels/ClPixelWiseMultiplicationKernel.h
+++ b/src/core/gpu/cl/kernels/ClPixelWiseMultiplicationKernel.h
@@ -35,7 +35,7 @@ namespace opencl
 namespace kernels
 {
 /** Interface for the pixelwise multiplication kernel. */
-class ClPixelWiseMultiplicationKernel : public ICLKernel
+class ClPixelWiseMultiplicationKernel : public IClKernel
 {
 public:
     /** Default constructor */
diff --git a/src/core/gpu/cl/kernels/ClReshapeKernel.h b/src/core/gpu/cl/kernels/ClReshapeKernel.h
index ee835c0fd3..0501b93f40 100644
--- a/src/core/gpu/cl/kernels/ClReshapeKernel.h
+++ b/src/core/gpu/cl/kernels/ClReshapeKernel.h
@@ -35,7 +35,7 @@ namespace opencl
 namespace kernels
 {
 /** Interface for the kernel to perform tensor reshaping */
-class ClReshapeKernel : public ICLKernel
+class ClReshapeKernel : public IClKernel
 {
 public:
     ClReshapeKernel() = default;
diff --git a/src/runtime/CL/functions/CLConvertFullyConnectedWeights.cpp b/src/runtime/CL/functions/CLConvertFullyConnectedWeights.cpp
index bbe9b487e5..8189eee402 100644
--- a/src/runtime/CL/functions/CLConvertFullyConnectedWeights.cpp
+++ b/src/runtime/CL/functions/CLConvertFullyConnectedWeights.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2018-2020 Arm Limited.
+ * Copyright (c) 2018-2021 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -22,11 +22,28 @@
  * SOFTWARE.
  */
 #include "arm_compute/runtime/CL/functions/CLConvertFullyConnectedWeights.h"
-#include "src/core/CL/kernels/CLConvertFullyConnectedWeightsKernel.h"
-#include "src/core/CL/kernels/CLFillBorderKernel.h"
+
+#include "arm_compute/core/CL/CLKernelLibrary.h"
+#include "arm_compute/core/CL/ICLTensor.h"
+#include "arm_compute/core/Types.h"
+#include "arm_compute/core/Validate.h"
+#include "src/core/CL/ICLKernel.h"
+#include "src/runtime/gpu/cl/operators/ClConvertFullyConnectedWeights.h"
 
 namespace arm_compute
 {
+struct CLConvertFullyConnectedWeights::Impl
+{
+    const ICLTensor                                        *src{ nullptr };
+    ICLTensor                                              *dst{ nullptr };
+    std::unique_ptr<opencl::ClConvertFullyConnectedWeights> op{ nullptr };
+};
+CLConvertFullyConnectedWeights::CLConvertFullyConnectedWeights()
+    : _impl(std::make_unique<Impl>())
+{
+}
+CLConvertFullyConnectedWeights::~CLConvertFullyConnectedWeights() = default;
+
 void CLConvertFullyConnectedWeights::configure(const ICLTensor *input, ICLTensor *output, const TensorShape &original_input_shape,
                                                DataLayout data_layout)
 {
@@ -36,14 +53,25 @@ void CLConvertFullyConnectedWeights::configure(const ICLTensor *input, ICLTensor
 void CLConvertFullyConnectedWeights::configure(const CLCompileContext &compile_context, const ICLTensor *input, ICLTensor *output, const TensorShape &original_input_shape,
                                                DataLayout data_layout)
 {
-    auto k = std::make_unique<CLConvertFullyConnectedWeightsKernel>();
-    k->configure(compile_context, input, output, original_input_shape, data_layout);
-    _kernel = std::move(k);
+    ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
+    _impl->src = input;
+    _impl->dst = output;
+    _impl->op  = std::make_unique<opencl::ClConvertFullyConnectedWeights>();
+    _impl->op->configure(compile_context, _impl->src->info(), _impl->dst->info(), original_input_shape, data_layout);
 }
 
 Status CLConvertFullyConnectedWeights::validate(const ITensorInfo *input, const ITensorInfo *output, const TensorShape &original_input_shape,
                                                 DataLayout data_layout)
 {
-    return CLConvertFullyConnectedWeightsKernel::validate(input, output, original_input_shape, data_layout);
+    return opencl::ClConvertFullyConnectedWeights::validate(input, output, original_input_shape, data_layout);
 }
+
+void CLConvertFullyConnectedWeights::run()
+{
+    ITensorPack pack;
+    pack.add_tensor(TensorType::ACL_SRC, _impl->src);
+    pack.add_tensor(TensorType::ACL_DST, _impl->dst);
+    _impl->op->run(pack);
+}
+
 } // namespace arm_compute
\ No newline at end of file
diff --git a/src/runtime/gpu/cl/operators/ClConvertFullyConnectedWeights.cpp b/src/runtime/gpu/cl/operators/ClConvertFullyConnectedWeights.cpp
new file mode 100644
index 0000000000..0d2f2925d3
--- /dev/null
+++ b/src/runtime/gpu/cl/operators/ClConvertFullyConnectedWeights.cpp
@@ -0,0 +1,45 @@
+/*
+ * Copyright (c) 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/runtime/gpu/cl/operators/ClConvertFullyConnectedWeights.h"
+
+#include "src/core/gpu/cl/ClCompileContext.h"
+#include "src/core/gpu/cl/kernels/ClConvertFullyConnectedWeightsKernel.h"
+
+namespace arm_compute
+{
+namespace opencl
+{
+void ClConvertFullyConnectedWeights::configure(const ClCompileContext &compile_context, const ITensorInfo *src, ITensorInfo *dst, const TensorShape &original_src_shape, DataLayout data_layout)
+{
+    auto k = std::make_unique<kernels::ClConvertFullyConnectedWeightsKernel>();
+    k->configure(compile_context, src, dst, original_src_shape, data_layout);
+    _kernel = std::move(k);
+}
+
+Status ClConvertFullyConnectedWeights::validate(const ITensorInfo *src, const ITensorInfo *dst, const TensorShape &original_src_shape, DataLayout data_layout)
+{
+    return kernels::ClConvertFullyConnectedWeightsKernel::validate(src, dst, original_src_shape, data_layout);
+}
+} // namespace opencl
+} // namespace arm_compute
\ No newline at end of file
diff --git a/src/runtime/gpu/cl/operators/ClConvertFullyConnectedWeights.h b/src/runtime/gpu/cl/operators/ClConvertFullyConnectedWeights.h
new file mode 100644
index 0000000000..efedc2fcb7
--- /dev/null
+++ b/src/runtime/gpu/cl/operators/ClConvertFullyConnectedWeights.h
@@ -0,0 +1,62 @@
+/*
+ * Copyright (c) 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_CONVERTFULLYCONNECTEDWEIGHTS_H
+#define ARM_COMPUTE_CL_CONVERTFULLYCONNECTEDWEIGHTS_H
+
+#include "src/core/gpu/cl/ClCompileContext.h"
+#include "src/runtime/gpu/cl/IClOperator.h"
+
+namespace arm_compute
+{
+namespace opencl
+{
+/** Basic function to run @ref kernels::ClConvertFullyConnectedWeightsKernel */
+class ClConvertFullyConnectedWeights : public IClOperator
+{
+public:
+    /** Constructor */
+    ClConvertFullyConnectedWeights() = default;
+    /** Initialise the kernel's inputs and outputs
+     *
+     * @param[in] compile_context    The compile context to be used.
+     * @param[in] src                The src tensor info. Data types supported: All.
+     * @param[in] dst                The dst tensor info. Data types supported: Same as @p src
+     * @param[in] original_src_shape Shape of the original src tensor (the one entering fully connected layer).
+     * @param[in] data_layout        The data layout the weights have been trained in.
+     */
+    void configure(const CLCompileContext &compile_context, const ITensorInfo *src, ITensorInfo *dst, const TensorShape &original_src_shape, DataLayout data_layout);
+    /**  Static function to check if given info will lead to a valid configuration of @ref kernels::ClConvertFullyConnectedWeightsKernel.
+     *
+     * @param[in] src                First tensor src info. Data types supported: All.
+     * @param[in] dst                Output tensor info. Data types supported: same as @p src.
+     * @param[in] original_src_shape Shape of the original src tensor (the one entering fully connected layer).
+     * @param[in] data_layout        The data layout the weights have been trained in.
+     *
+     * @return a status
+     */
+    static Status validate(const ITensorInfo *src, const ITensorInfo *dst, const TensorShape &original_src_shape, DataLayout data_layout);
+};
+} // namespace opencl
+} // namespace arm_compute
+#endif /* ARM_COMPUTE_CL_CONVERTFULLYCONNECTEDWEIGHTS_H */