Port ClTranspose to new API

Partially Resolves: COMPMID-4277 (1/2)

Signed-off-by: Teresa Charlin <teresa.charlinreyes@arm.com>
Change-Id: I704c2303135cbe1ba46d2fd5642c84c562204bc7
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/5194
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>

19 files changed, 477 insertions, 287 deletions

diff --git a/Android.bp b/Android.bp
index dcf4b802fa..90e28a67fc 100644
--- a/Android.bp
+++ b/Android.bp
@@ -146,7 +146,6 @@ cc_library_static {
         "src/core/CL/kernels/CLStackLayerKernel.cpp",
         "src/core/CL/kernels/CLStridedSliceKernel.cpp",
         "src/core/CL/kernels/CLTileKernel.cpp",
-        "src/core/CL/kernels/CLTransposeKernel.cpp",
         "src/core/CL/kernels/CLWeightsReshapeKernel.cpp",
         "src/core/CL/kernels/CLWinogradFilterTransformKernel.cpp",
         "src/core/CL/kernels/CLWinogradInputTransformKernel.cpp",
@@ -387,6 +386,7 @@ cc_library_static {
         "src/core/gpu/cl/kernels/ClReshapeKernel.cpp",
         "src/core/gpu/cl/kernels/ClScaleKernel.cpp",
         "src/core/gpu/cl/kernels/ClSoftmaxKernel.cpp",
+        "src/core/gpu/cl/kernels/ClTransposeKernel.cpp",
         "src/core/gpu/cl/kernels/ClWidthConcatenate2TensorsKernel.cpp",
         "src/core/gpu/cl/kernels/ClWidthConcatenate4TensorsKernel.cpp",
         "src/core/gpu/cl/kernels/ClWidthConcatenateKernel.cpp",
@@ -691,6 +691,7 @@ cc_library_static {
         "src/runtime/gpu/cl/operators/ClScale.cpp",
         "src/runtime/gpu/cl/operators/ClSoftmax.cpp",
         "src/runtime/gpu/cl/operators/ClSub.cpp",
+        "src/runtime/gpu/cl/operators/ClTranspose.cpp",
         "utils/CommonGraphOptions.cpp",
         "utils/GraphUtils.cpp",
         "utils/Utils.cpp",
diff --git a/arm_compute/runtime/CL/functions/CLFullyConnectedLayer.h b/arm_compute/runtime/CL/functions/CLFullyConnectedLayer.h
index 3f17e4a921..a640e344d4 100644
--- a/arm_compute/runtime/CL/functions/CLFullyConnectedLayer.h
+++ b/arm_compute/runtime/CL/functions/CLFullyConnectedLayer.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017-2020 Arm Limited.
+ * Copyright (c) 2017-2021 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -36,15 +36,27 @@
 
 namespace arm_compute
 {
-/** Basic function to reshape the weights of Fully Connected layer with OpenCL. This function calls the following kernels:
+/** Function to reshape the weights of Fully Connected layer with OpenCL by transposing input tensor. This function calls the following kernel:
  *
- *  -# @ref CLTransposeKernel
+ *  -# @ref opencl::kernels::ClTransposeKernel
  *
  * @note  The fully connected layer accepts "weights" tensors only with 2 dimensions.
  */
-class CLFullyConnectedLayerReshapeWeights : public ICLSimpleFunction
+class CLFullyConnectedLayerReshapeWeights : public IFunction
 {
 public:
+    /** Constructor */
+    CLFullyConnectedLayerReshapeWeights();
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    CLFullyConnectedLayerReshapeWeights(const CLFullyConnectedLayerReshapeWeights &) = delete;
+    /** Default move constructor */
+    CLFullyConnectedLayerReshapeWeights(CLFullyConnectedLayerReshapeWeights &&) = default;
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    CLFullyConnectedLayerReshapeWeights &operator=(const CLFullyConnectedLayerReshapeWeights &) = delete;
+    /** Default move assignment operator */
+    CLFullyConnectedLayerReshapeWeights &operator=(CLFullyConnectedLayerReshapeWeights &&) = default;
+    /** Default destructor */
+    ~CLFullyConnectedLayerReshapeWeights();
     /** Set the input and output tensors.
      *
      * @param[in]  input  Weights tensor. The weights must be 2 dimensional. Data types supported: QASYMM8/QASYMM8_SIGNED/F16/F32.
@@ -66,6 +78,13 @@ public:
      * @return a status
      */
     static Status validate(const ITensorInfo *input, const ITensorInfo *output);
+
+    // Inherited methods overridden
+    void run() override;
+
+private:
+    struct Impl;
+    std::unique_ptr<Impl> _impl;
 };
 
 namespace weights_transformations
diff --git a/arm_compute/runtime/CL/functions/CLLSTMLayer.h b/arm_compute/runtime/CL/functions/CLLSTMLayer.h
index 20b068316c..38a24d030b 100644
--- a/arm_compute/runtime/CL/functions/CLLSTMLayer.h
+++ b/arm_compute/runtime/CL/functions/CLLSTMLayer.h
@@ -46,8 +46,14 @@
 namespace arm_compute
 {
 class CLCompileContext;
-class CLTransposeKernel;
 class ICLTensor;
+namespace opencl
+{
+namespace kernels
+{
+class ClTransposeKernel;
+}
+}
 
 /** This function performs a single time step in a Long Short-Term Memory (LSTM) layer.
  *
@@ -212,90 +218,91 @@ public:
     void prepare() override;
 
 private:
-    MemoryGroup                        _memory_group;
-    CLFullyConnectedLayer              _fully_connected_input_gate;
-    CLArithmeticAddition               _accum_input_gate1;
-    CLArithmeticSubtraction            _subtract_input_gate;
-    CLPixelWiseMultiplication          _pixelwise_mul_input_gate;
-    CLActivationLayer                  _activation_input_gate;
-    CLFullyConnectedLayer              _fully_connected_forget_gate;
-    CLArithmeticAddition               _accum_forget_gate1;
-    CLPixelWiseMultiplication          _pixelwise_mul_forget_gate;
-    CLActivationLayer                  _activation_forget_gate;
-    CLFullyConnectedLayer              _fully_connected_cell_state;
-    CLGEMM                             _gemm_cell_state1;
-    std::unique_ptr<CLTransposeKernel> _transpose_cell_state;
-    CLArithmeticAddition               _accum_cell_state1;
-    CLArithmeticAddition               _accum_cell_state2;
-    CLPixelWiseMultiplication          _pixelwise_mul_cell_state1;
-    CLActivationLayer                  _activation_cell_state;
-    CLActivationLayer                  _cell_clip;
-    CLPixelWiseMultiplication          _pixelwise_mul_cell_state2;
-    CLFullyConnectedLayer              _fully_connected_output;
-    CLPixelWiseMultiplication          _pixelwise_mul_output_state1;
-    CLArithmeticAddition               _accum_output1;
-    CLActivationLayer                  _activation_output;
-    CLActivationLayer                  _activation_output_state;
-    CLPixelWiseMultiplication          _pixelwise_mul_output_state2;
-    CLFullyConnectedLayer              _fully_connected_output_state;
-    CLActivationLayer                  _projection_clip;
-    CLCopy                             _copy_cell_state;
-    CLCopy                             _copy_output;
-    CLConcatenateLayer                 _concat_scratch_buffer;
-    CLConcatenateLayer                 _concat_inputs_forget_gate;
-    CLConcatenateLayer                 _concat_weights_forget_gate;
-    CLConcatenateLayer                 _concat_weights_input_gate;
-    CLConcatenateLayer                 _concat_weights_output;
-    CLFill                             _ones_fill;
-    CLMeanStdDevNormalizationLayer     _mean_std_norm_input_gate;
-    CLPixelWiseMultiplication          _pixelwise_mul_input_gate_coeff;
-    CLArithmeticAddition               _accum_input_gate_bias;
-    CLMeanStdDevNormalizationLayer     _mean_std_norm_forget_gate;
-    CLPixelWiseMultiplication          _pixelwise_mul_forget_gate_coeff;
-    CLArithmeticAddition               _accum_forget_gate_bias;
-    CLMeanStdDevNormalizationLayer     _mean_std_norm_cell_gate;
-    CLPixelWiseMultiplication          _pixelwise_mul_cell_gate_coeff;
-    CLArithmeticAddition               _accum_cell_gate_bias;
-    CLMeanStdDevNormalizationLayer     _mean_std_norm_output_gate;
-    CLPixelWiseMultiplication          _pixelwise_mul_output_gate_coeff;
-    CLArithmeticAddition               _accum_output_gate_bias;
-    CLTensor                           _input_gate_out1;
-    CLTensor                           _input_gate_out2;
-    CLTensor                           _input_gate_out3;
-    CLTensor                           _input_gate_out4;
-    CLTensor                           _forget_gate_out1;
-    CLTensor                           _forget_gate_out2;
-    CLTensor                           _forget_gate_out3;
-    CLTensor                           _forget_gate_out4;
-    CLTensor                           _forget_gate_out5;
-    CLTensor                           _forget_gate_out6;
-    CLTensor                           _cell_state_out1;
-    CLTensor                           _cell_state_out2;
-    CLTensor                           _cell_state_out3;
-    CLTensor                           _cell_state_out4;
-    CLTensor                           _cell_state_out5;
-    CLTensor                           _output1;
-    CLTensor                           _output2;
-    CLTensor                           _output3;
-    CLTensor                           _output4;
-    CLTensor                           _cell_state_activation;
-    CLTensor                           _output_state1;
-    CLTensor                           _ones;
-    CLTensor                           _input_layer_norm_out1;
-    CLTensor                           _input_layer_norm_out2;
-    CLTensor                           _forget_layer_norm_out1;
-    CLTensor                           _forget_layer_norm_out2;
-    CLTensor                           _cell_layer_norm_out1;
-    CLTensor                           _cell_layer_norm_out2;
-    CLTensor                           _output_layer_norm_out1;
-    CLTensor                           _output_layer_norm_out2;
-    bool                               _run_peephole_opt;
-    bool                               _run_cifg_opt;
-    bool                               _perform_cell_clipping;
-    bool                               _has_projection_weights;
-    bool                               _perform_projection_clipping;
-    bool                               _is_prepared;
-    bool                               _is_layer_norm_lstm;
+    MemoryGroup                                         _memory_group;
+    CLFullyConnectedLayer                               _fully_connected_input_gate;
+    CLArithmeticAddition                                _accum_input_gate1;
+    CLArithmeticSubtraction                             _subtract_input_gate;
+    CLPixelWiseMultiplication                           _pixelwise_mul_input_gate;
+    CLActivationLayer                                   _activation_input_gate;
+    CLFullyConnectedLayer                               _fully_connected_forget_gate;
+    CLArithmeticAddition                                _accum_forget_gate1;
+    CLPixelWiseMultiplication                           _pixelwise_mul_forget_gate;
+    CLActivationLayer                                   _activation_forget_gate;
+    CLFullyConnectedLayer                               _fully_connected_cell_state;
+    CLGEMM                                              _gemm_cell_state1;
+    std::unique_ptr<opencl::kernels::ClTransposeKernel> _transpose_cell_state;
+    CLArithmeticAddition                                _accum_cell_state1;
+    CLArithmeticAddition                                _accum_cell_state2;
+    CLPixelWiseMultiplication                           _pixelwise_mul_cell_state1;
+    CLActivationLayer                                   _activation_cell_state;
+    CLActivationLayer                                   _cell_clip;
+    CLPixelWiseMultiplication                           _pixelwise_mul_cell_state2;
+    CLFullyConnectedLayer                               _fully_connected_output;
+    CLPixelWiseMultiplication                           _pixelwise_mul_output_state1;
+    CLArithmeticAddition                                _accum_output1;
+    CLActivationLayer                                   _activation_output;
+    CLActivationLayer                                   _activation_output_state;
+    CLPixelWiseMultiplication                           _pixelwise_mul_output_state2;
+    CLFullyConnectedLayer                               _fully_connected_output_state;
+    CLActivationLayer                                   _projection_clip;
+    CLCopy                                              _copy_cell_state;
+    CLCopy                                              _copy_output;
+    CLConcatenateLayer                                  _concat_scratch_buffer;
+    CLConcatenateLayer                                  _concat_inputs_forget_gate;
+    CLConcatenateLayer                                  _concat_weights_forget_gate;
+    CLConcatenateLayer                                  _concat_weights_input_gate;
+    CLConcatenateLayer                                  _concat_weights_output;
+    CLFill                                              _ones_fill;
+    CLMeanStdDevNormalizationLayer                      _mean_std_norm_input_gate;
+    CLPixelWiseMultiplication                           _pixelwise_mul_input_gate_coeff;
+    CLArithmeticAddition                                _accum_input_gate_bias;
+    CLMeanStdDevNormalizationLayer                      _mean_std_norm_forget_gate;
+    CLPixelWiseMultiplication                           _pixelwise_mul_forget_gate_coeff;
+    CLArithmeticAddition                                _accum_forget_gate_bias;
+    CLMeanStdDevNormalizationLayer                      _mean_std_norm_cell_gate;
+    CLPixelWiseMultiplication                           _pixelwise_mul_cell_gate_coeff;
+    CLArithmeticAddition                                _accum_cell_gate_bias;
+    CLMeanStdDevNormalizationLayer                      _mean_std_norm_output_gate;
+    CLPixelWiseMultiplication                           _pixelwise_mul_output_gate_coeff;
+    CLArithmeticAddition                                _accum_output_gate_bias;
+    CLTensor                                            _input_gate_out1;
+    CLTensor                                            _input_gate_out2;
+    CLTensor                                            _input_gate_out3;
+    CLTensor                                            _input_gate_out4;
+    CLTensor                                            _forget_gate_out1;
+    CLTensor                                            _forget_gate_out2;
+    CLTensor                                            _forget_gate_out3;
+    CLTensor                                            _forget_gate_out4;
+    CLTensor                                            _forget_gate_out5;
+    CLTensor                                            _forget_gate_out6;
+    CLTensor                                            _cell_state_out1;
+    CLTensor                                            _cell_state_out2;
+    CLTensor                                            _cell_state_out3;
+    CLTensor                                            _cell_state_out4;
+    CLTensor                                            _cell_state_out5;
+    CLTensor                                            _output1;
+    CLTensor                                            _output2;
+    CLTensor                                            _output3;
+    CLTensor                                            _output4;
+    CLTensor                                            _cell_state_activation;
+    CLTensor                                            _output_state1;
+    CLTensor                                            _ones;
+    CLTensor                                            _input_layer_norm_out1;
+    CLTensor                                            _input_layer_norm_out2;
+    CLTensor                                            _forget_layer_norm_out1;
+    CLTensor                                            _forget_layer_norm_out2;
+    CLTensor                                            _cell_layer_norm_out1;
+    CLTensor                                            _cell_layer_norm_out2;
+    CLTensor                                            _output_layer_norm_out1;
+    CLTensor                                            _output_layer_norm_out2;
+    bool                                                _run_peephole_opt;
+    bool                                                _run_cifg_opt;
+    bool                                                _perform_cell_clipping;
+    bool                                                _has_projection_weights;
+    bool                                                _perform_projection_clipping;
+    bool                                                _is_prepared;
+    bool                                                _is_layer_norm_lstm;
+    const ICLTensor                                    *_recurrent_to_cell_weights{ nullptr };
 };
 } // namespace arm_compute
 #endif /* ARM_COMPUTE_CLLSTMLAYER_H */
diff --git a/arm_compute/runtime/CL/functions/CLPermute.h b/arm_compute/runtime/CL/functions/CLPermute.h
index bcd9566fbf..e59cca2a67 100644
--- a/arm_compute/runtime/CL/functions/CLPermute.h
+++ b/arm_compute/runtime/CL/functions/CLPermute.h
@@ -46,11 +46,11 @@ public:
     /** Prevent instances of this class from being copied (As this class contains pointers) */
     CLPermute(const CLPermute &) = delete;
     /** Default move constructor */
-    CLPermute(CLPermute &&);
+    CLPermute(CLPermute &&) = default;
     /** Prevent instances of this class from being copied (As this class contains pointers) */
     CLPermute &operator=(const CLPermute &) = delete;
     /** Default move assignment operator */
-    CLPermute &operator=(CLPermute &&);
+    CLPermute &operator=(CLPermute &&) = default;
     /** Set the input and output tensors.
      *
      * @note Arbitrary permutation vectors are supported with rank not greater than 4
diff --git a/arm_compute/runtime/CL/functions/CLTranspose.h b/arm_compute/runtime/CL/functions/CLTranspose.h
index 2b7a03f23f..43cebeba90 100644
--- a/arm_compute/runtime/CL/functions/CLTranspose.h
+++ b/arm_compute/runtime/CL/functions/CLTranspose.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017-2020 Arm Limited.
+ * Copyright (c) 2017-2021 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -25,7 +25,9 @@
 #define ARM_COMPUTE_CLTRANSPOSE_H
 
 #include "arm_compute/core/Error.h"
-#include "arm_compute/runtime/CL/ICLSimpleFunction.h"
+#include "arm_compute/runtime/IFunction.h"
+
+#include <memory>
 
 namespace arm_compute
 {
@@ -33,14 +35,22 @@ class CLCompileContext;
 class ICLTensor;
 class ITensorInfo;
 
-/** Basic function to transpose a matrix on OpenCL. This function calls the following OpenCL kernel:
- *
- *  -# @ref CLTransposeKernel
- *
- */
-class CLTranspose : public ICLSimpleFunction
+/** Basic function to execute an @ref opencl::kernels::ClTransposeKernel. */
+class CLTranspose : public IFunction
 {
 public:
+    /** Constructor */
+    CLTranspose();
+    /** Destructor */
+    ~CLTranspose();
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    CLTranspose(const CLTranspose &) = delete;
+    /** Default move constructor */
+    CLTranspose(CLTranspose &&) = default;
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    CLTranspose &operator=(const CLTranspose &) = delete;
+    /** Default move assignment operator */
+    CLTranspose &operator=(CLTranspose &&) = default;
     /** Initialise the kernel's inputs and output
      *
      * @param[in]  input  Input tensor. Data types supported: All.
@@ -62,6 +72,13 @@ public:
      * @return a status
      */
     static Status validate(const ITensorInfo *input, const ITensorInfo *output);
+
+    // Inherited methods overridden:
+    void run() override;
+
+private:
+    struct Impl;
+    std::unique_ptr<Impl> _impl;
 };
 }
 
diff --git a/docs/00_introduction.dox b/docs/00_introduction.dox
index 00a5e93ce5..b5bf34644a 100644
--- a/docs/00_introduction.dox
+++ b/docs/00_introduction.dox
@@ -197,7 +197,7 @@ v21.02 Public major release
    - @ref CLSelectKernel
    - @ref CLBitwiseKernel
    - @ref opencl::kernels::ClFloorKernel
-   - @ref CLTransposeKernel
+   - CLTransposeKernel
  - Deprecate functions in CLTuner:
     - add_lws_to_table
     - import_lws_table
@@ -1390,7 +1390,7 @@ v17.03 Sources preview
    - CLGradientKernel, CLEdgeNonMaxSuppressionKernel, CLEdgeTraceKernel / CLCannyEdge
    - GEMM refactoring + FP16 support: CLGEMMInterleave4x4Kernel, CLGEMMTranspose1xWKernel, @ref CLGEMMMatrixMultiplyKernel, CLGEMMMatrixAdditionKernel / @ref CLGEMM
    - CLGEMMMatrixAccumulateBiasesKernel / @ref CLFullyConnectedLayer
-   - @ref CLTransposeKernel / @ref CLTranspose
+   - CLTransposeKernel / @ref CLTranspose
    - @ref CLLKTrackerInitKernel, @ref CLLKTrackerStage0Kernel, @ref CLLKTrackerStage1Kernel, @ref CLLKTrackerFinalizeKernel / CLOpticalFlow
    - @ref CLNormalizationLayerKernel / @ref CLNormalizationLayer
    - CLLaplacianPyramid, CLLaplacianReconstruct
diff --git a/src/core/CL/CLKernels.h b/src/core/CL/CLKernels.h
index eaac415bc4..d81b60b663 100644
--- a/src/core/CL/CLKernels.h
+++ b/src/core/CL/CLKernels.h
@@ -89,7 +89,6 @@
 #include "src/core/CL/kernels/CLStackLayerKernel.h"
 #include "src/core/CL/kernels/CLStridedSliceKernel.h"
 #include "src/core/CL/kernels/CLTileKernel.h"
-#include "src/core/CL/kernels/CLTransposeKernel.h"
 #include "src/core/CL/kernels/CLWeightsReshapeKernel.h"
 #include "src/core/CL/kernels/CLWinogradFilterTransformKernel.h"
 #include "src/core/CL/kernels/CLWinogradInputTransformKernel.h"
diff --git a/src/core/CL/kernels/CLTransposeKernel.cpp b/src/core/CL/kernels/CLTransposeKernel.cpp
deleted file mode 100644
index 56ff48be1f..0000000000
--- a/src/core/CL/kernels/CLTransposeKernel.cpp
+++ /dev/null
@@ -1,116 +0,0 @@
-/*
- * Copyright (c) 2017-2021 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/CL/kernels/CLTransposeKernel.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/Helpers.h"
-#include "arm_compute/core/TensorInfo.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/StringSupport.h"
-
-namespace arm_compute
-{
-namespace
-{
-TensorShape transposed_tensor_shape(const TensorShape &in)
-{
-    TensorShape  output_shape{ in };
-    const size_t w_out = in[1];
-    const size_t h_out = in[0];
-    output_shape.set(0, w_out);
-    output_shape.set(1, h_out);
-
-    return output_shape;
-}
-
-Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output)
-{
-    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);
-
-    if(output->total_size() != 0)
-    {
-        const TensorInfo tensor_info = input->clone()->set_tensor_shape(transposed_tensor_shape(input->tensor_shape()));
-
-        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(output, &tensor_info);
-        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
-        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_QUANTIZATION_INFO(input, output);
-    }
-
-    return Status{};
-}
-} // namespace
-
-Status CLTransposeKernel::validate(const ITensorInfo *input, const ITensorInfo *output)
-{
-    ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
-    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output));
-    return Status{};
-}
-
-void CLTransposeKernel::configure(const ICLTensor *input, ICLTensor *output)
-{
-    configure(CLKernelLibrary::get().get_compile_context(), input, output);
-}
-
-void CLTransposeKernel::configure(const CLCompileContext &compile_context, const ICLTensor *input, ICLTensor *output)
-{
-    ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
-
-    // Output tensor auto inizialitation if not yet initialized
-    auto_init_if_empty(*output->info(), input->info()->clone()->set_tensor_shape(transposed_tensor_shape(input->info()->tensor_shape())));
-
-    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info()));
-    auto padding_info = get_padding_info({ input, output });
-
-    _input  = input;
-    _output = output;
-
-    const unsigned int vec_size_x           = adjust_vec_size(max_cl_vector_width / input->info()->element_size(), input->info()->dimension(0));
-    const int          vec_size_x_leftovers = input->info()->dimension(0) % vec_size_x;
-    const unsigned int vec_size_y           = adjust_vec_size(max_cl_vector_width / input->info()->element_size(), input->info()->dimension(1));
-    const int          vec_size_y_leftovers = input->info()->dimension(1) % vec_size_y;
-
-    CLBuildOptions build_opts;
-    build_opts.add_option("-DDATA_TYPE_IN_BYTES=" + support::cpp11::to_string(input->info()->element_size()));
-    build_opts.add_option("-DVEC_SIZE_X=" + support::cpp11::to_string(vec_size_x));
-    build_opts.add_option("-DVEC_SIZE_LEFTOVER_X=" + support::cpp11::to_string(vec_size_x_leftovers));
-    build_opts.add_option("-DVEC_SIZE_Y=" + support::cpp11::to_string(vec_size_y));
-    build_opts.add_option("-DVEC_SIZE_LEFTOVER_Y=" + support::cpp11::to_string(vec_size_y_leftovers));
-
-    _kernel = create_kernel(compile_context, "transpose", build_opts.options());
-
-    // Configure kernel window
-    Window win = calculate_max_window(*input->info(), Steps(vec_size_x, vec_size_y));
-    ICLKernel::configure_internal(win, cl::NDRange(2, 8));
-    ARM_COMPUTE_ERROR_ON(has_padding_changed(padding_info));
-}
-} // namespace arm_compute
\ No newline at end of file
diff --git a/src/core/gpu/cl/kernels/ClPermuteKernel.cpp b/src/core/gpu/cl/kernels/ClPermuteKernel.cpp
index 04e649b911..ffc13060a8 100644
--- a/src/core/gpu/cl/kernels/ClPermuteKernel.cpp
+++ b/src/core/gpu/cl/kernels/ClPermuteKernel.cpp
@@ -57,7 +57,7 @@ Status validate_arguments(const ITensorInfo *src, const ITensorInfo *dst, const
     ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src, dst);
     ARM_COMPUTE_RETURN_ERROR_ON(src->data_type() == DataType::UNKNOWN);
     ARM_COMPUTE_RETURN_ERROR_ON_MSG(src->num_dimensions() < 1 || src->num_dimensions() > 4,
-                                    "Permutation upto 4-D src tensor is supported");
+                                    "Permutation up to 4-D src tensor is supported");
     ARM_COMPUTE_RETURN_ERROR_ON_MSG(perm.num_dimensions() < 1 || perm.num_dimensions() > 4,
                                     "Permutation vector size should be less than or equal to 4");
     for(const auto &p : perm)
@@ -77,17 +77,12 @@ Status validate_arguments(const ITensorInfo *src, const ITensorInfo *dst, const
 }
 } // namespace
 
-void ClPermuteKernel::configure(const ITensorInfo *src, ITensorInfo *dst, const PermutationVector &perm)
-{
-    configure(CLKernelLibrary::get().get_compile_context(), src, dst, perm);
-}
-
 void ClPermuteKernel::configure(const CLCompileContext &compile_context, const ITensorInfo *src, ITensorInfo *dst, const PermutationVector &perm)
 {
     ARM_COMPUTE_ERROR_ON_NULLPTR(src, dst);
     auto              padding_info = get_padding_info({ src, dst });
     const TensorShape dst_shape    = get_dst_shape(src, perm);
-    // Output auto inizialitation if not yet initialized
+    // Output auto initialization if not yet initialized
     auto_init_if_empty(*dst, src->clone()->set_tensor_shape(dst_shape));
 
     ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(src, dst, perm));
diff --git a/src/core/gpu/cl/kernels/ClPermuteKernel.h b/src/core/gpu/cl/kernels/ClPermuteKernel.h
index ae3996fca1..b844214595 100644
--- a/src/core/gpu/cl/kernels/ClPermuteKernel.h
+++ b/src/core/gpu/cl/kernels/ClPermuteKernel.h
@@ -38,7 +38,7 @@ namespace kernels
  *
  * Permutes given a permutation vector
  */
-class ClPermuteKernel : public ICLKernel
+class ClPermuteKernel : public IClKernel
 {
 public:
     /** Default constructor */
@@ -48,15 +48,6 @@ public:
      *
      * @note Arbitrary permutation vectors are supported with rank not greater than 4
      *
-     * @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] perm Permutation vector
-     */
-    void configure(const ITensorInfo *src, ITensorInfo *dst, const PermutationVector &perm);
-    /** Set the src and dst of the kernel.
-     *
-     * @note Arbitrary permutation vectors are supported with rank not greater than 4
-     *
      * @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
diff --git a/src/core/gpu/cl/kernels/ClTransposeKernel.cpp b/src/core/gpu/cl/kernels/ClTransposeKernel.cpp
new file mode 100644
index 0000000000..704d0152cf
--- /dev/null
+++ b/src/core/gpu/cl/kernels/ClTransposeKernel.cpp
@@ -0,0 +1,119 @@
+/*
+ * Copyright (c) 2017-2021 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/ClTransposeKernel.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/Utils.h"
+#include "arm_compute/core/Validate.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"
+
+namespace arm_compute
+{
+namespace opencl
+{
+namespace kernels
+{
+void ClTransposeKernel::configure(const CLCompileContext &compile_context, const ITensorInfo *src, ITensorInfo *dst)
+{
+    ARM_COMPUTE_ERROR_ON_NULLPTR(src, dst);
+
+    // Output auto initialization if not yet initialized
+    const TensorShape dst_shape = misc::shape_calculator::compute_transposed_shape(*src);
+    auto_init_if_empty(*dst, src->clone()->set_tensor_shape(dst_shape));
+
+    ARM_COMPUTE_ERROR_THROW_ON(ClTransposeKernel::validate(src, dst));
+    auto padding_info = get_padding_info({ src, dst });
+
+    // Create kernel
+    const unsigned int vec_size_x           = adjust_vec_size(max_cl_vector_width / src->element_size(), src->dimension(0));
+    const int          vec_size_x_leftovers = src->dimension(0) % vec_size_x;
+    const unsigned int vec_size_y           = adjust_vec_size(max_cl_vector_width / src->element_size(), src->dimension(1));
+    const int          vec_size_y_leftovers = src->dimension(1) % vec_size_y;
+
+    CLBuildOptions build_opts;
+    build_opts.add_option("-DDATA_TYPE_IN_BYTES=" + support::cpp11::to_string(src->element_size()));
+    build_opts.add_option("-DVEC_SIZE_X=" + support::cpp11::to_string(vec_size_x));
+    build_opts.add_option("-DVEC_SIZE_LEFTOVER_X=" + support::cpp11::to_string(vec_size_x_leftovers));
+    build_opts.add_option("-DVEC_SIZE_Y=" + support::cpp11::to_string(vec_size_y));
+    build_opts.add_option("-DVEC_SIZE_LEFTOVER_Y=" + support::cpp11::to_string(vec_size_y_leftovers));
+
+    _kernel = create_kernel(compile_context, "transpose", build_opts.options());
+
+    // Configure kernel window
+    Window win = calculate_max_window(*src, Steps(vec_size_x, vec_size_y));
+    ICLKernel::configure_internal(win, cl::NDRange(2, 8));
+    ARM_COMPUTE_ERROR_ON(has_padding_changed(padding_info));
+}
+
+Status ClTransposeKernel::validate(const ITensorInfo *src, const ITensorInfo *dst)
+{
+    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_MSG(src->num_dimensions() > 2, "Transpose up to 2-D src tensor is supported");
+
+    // Validate configured dst
+    if(dst->total_size() != 0)
+    {
+        const TensorInfo dst_info = src->clone()->set_tensor_shape(misc::shape_calculator::compute_transposed_shape(*src));
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(dst, &dst_info);
+
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_QUANTIZATION_INFO(src, dst);
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src, dst);
+    }
+
+    return Status{};
+}
+
+void ClTransposeKernel::run_op(ITensorPack &tensors, const Window &window, cl::CommandQueue &queue)
+{
+    ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
+    ARM_COMPUTE_ERROR_ON_MISMATCHING_WINDOWS(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));
+
+    Window slice = window.first_slice_window_2D();
+
+    do
+    {
+        unsigned int idx = 0;
+        add_2D_tensor_argument(idx, src, slice);
+        add_2D_tensor_argument(idx, dst, slice);
+        enqueue(queue, *this, slice, lws_hint());
+    }
+    while(window.slide_window_slice_2D(slice));
+}
+} // namespace kernels
+} // namespace opencl
+} // namespace arm_compute
\ No newline at end of file
diff --git a/src/core/CL/kernels/CLTransposeKernel.h b/src/core/gpu/cl/kernels/ClTransposeKernel.h
index 0c4b7b4aff..21d4fd41f5 100644
--- a/src/core/CL/kernels/CLTransposeKernel.h
+++ b/src/core/gpu/cl/kernels/ClTransposeKernel.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017-2020 Arm Limited.
+ * Copyright (c) 2017-2021 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -21,44 +21,45 @@
  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  * SOFTWARE.
  */
-#ifndef ARM_COMPUTE_CLTRANSPOSEKERNEL_H
-#define ARM_COMPUTE_CLTRANSPOSEKERNEL_H
+#ifndef ARM_COMPUTE_CL_TRANSPOSE_KERNEL_H
+#define ARM_COMPUTE_CL_TRANSPOSE_KERNEL_H
 
-#include "src/core/CL/ICLSimple2DKernel.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;
-
-/** OpenCL kernel which transposes the elements of a matrix.
- *
- * [width, height, batch] -> [height, width, batch]
- *
- */
-class CLTransposeKernel : public ICLSimple2DKernel
+namespace opencl
+{
+namespace kernels
+{
+/** OpenCL kernel to transpose a 2D tensor. */
+class ClTransposeKernel : public IClKernel
 {
 public:
-    /** Initialise the kernel's input and output.
-     *
-     * @param[in]  input  Input tensor. Data types supported: All.
-     * @param[out] output Output tensor. Data type supported: Same as @p input
-     */
-    void configure(const ICLTensor *input, ICLTensor *output);
-    /** Initialise the kernel's input and output.
+    ClTransposeKernel() = default;
+    ARM_COMPUTE_DISALLOW_COPY_ALLOW_MOVE(ClTransposeKernel);
+    /** Set the src and dst of the kernel.
      *
-     * @param[in]  compile_context The compile context to be used.
-     * @param[in]  input           Input tensor. Data types supported: All.
-     * @param[out] output          Output tensor. Data type supported: Same as @p input
+     * @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
      */
-    void configure(const CLCompileContext &compile_context, const ICLTensor *input, ICLTensor *output);
-    /** Static function to check if given info will lead to a valid configuration of @ref CLTransposeKernel
+    void configure(const CLCompileContext &compile_context, const ITensorInfo *src, ITensorInfo *dst);
+    /** Static function to check if given info will lead to a valid configuration of @ref ClTransposeKernel
      *
-     * @param[in] input  Input tensor. Data types supported: All.
-     * @param[in] output Output tensor. Data type supported: Same as @p input
+     * @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.
      *
      * @return a status
      */
-    static Status validate(const ITensorInfo *input, const ITensorInfo *output);
+    static Status validate(const ITensorInfo *src, const ITensorInfo *dst);
+
+    // 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_CLTRANSPOSEKERNEL_H */
+#endif /*ARM_COMPUTE_CL_TRANSPOSE_KERNEL_H */
diff --git a/src/runtime/CL/functions/CLFullyConnectedLayer.cpp b/src/runtime/CL/functions/CLFullyConnectedLayer.cpp
index 46a90a54b7..176c5bebd5 100644
--- a/src/runtime/CL/functions/CLFullyConnectedLayer.cpp
+++ b/src/runtime/CL/functions/CLFullyConnectedLayer.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017-2020 Arm Limited.
+ * Copyright (c) 2017-2021 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -40,7 +40,7 @@
 #include "src/core/CL/kernels/CLGEMMMatrixMultiplyReshapedOnlyRHSKernel.h"
 #include "src/core/CL/kernels/CLGEMMReshapeLHSMatrixKernel.h"
 #include "src/core/CL/kernels/CLGEMMReshapeRHSMatrixKernel.h"
-#include "src/core/CL/kernels/CLTransposeKernel.h"
+#include "src/core/gpu/cl/kernels/ClTransposeKernel.h"
 #include "support/Cast.h"
 
 #include <algorithm>
@@ -141,6 +141,20 @@ Status validate_mm(const ITensorInfo &input, const ITensorInfo &weights, const I
 }
 } // namespace
 
+struct CLFullyConnectedLayerReshapeWeights::Impl
+{
+    const ITensor                                      *src{ nullptr };
+    ITensor                                            *dst{ nullptr };
+    std::unique_ptr<opencl::kernels::ClTransposeKernel> op{ nullptr };
+};
+
+CLFullyConnectedLayerReshapeWeights::CLFullyConnectedLayerReshapeWeights()
+    : _impl(std::make_unique<Impl>())
+{
+}
+
+CLFullyConnectedLayerReshapeWeights::~CLFullyConnectedLayerReshapeWeights() = default;
+
 void CLFullyConnectedLayerReshapeWeights::configure(const ICLTensor *input, ICLTensor *output)
 {
     configure(CLKernelLibrary::get().get_compile_context(), input, output);
@@ -148,14 +162,23 @@ void CLFullyConnectedLayerReshapeWeights::configure(const ICLTensor *input, ICLT
 
 void CLFullyConnectedLayerReshapeWeights::configure(const CLCompileContext &compile_context, const ICLTensor *input, ICLTensor *output)
 {
-    auto k = std::make_unique<CLTransposeKernel>();
-    k->configure(compile_context, input, output);
-    _kernel = std::move(k);
+    _impl->src = input;
+    _impl->dst = output;
+    _impl->op  = std::make_unique<opencl::kernels::ClTransposeKernel>();
+    _impl->op->configure(compile_context, _impl->src->info(), _impl->dst->info());
 }
 
 Status CLFullyConnectedLayerReshapeWeights::validate(const ITensorInfo *input, const ITensorInfo *output)
 {
-    return CLTransposeKernel::validate(input, output);
+    return opencl::kernels::ClTransposeKernel::validate(input, output);
+}
+
+void CLFullyConnectedLayerReshapeWeights::run()
+{
+    ITensorPack pack{};
+    pack.add_tensor(TensorType::ACL_SRC, _impl->src);
+    pack.add_tensor(TensorType::ACL_DST, _impl->dst);
+    CLScheduler::get().enqueue_op(*_impl->op.get(), pack, false);
 }
 
 CLFullyConnectedLayer::CLFullyConnectedLayer(std::shared_ptr<IMemoryManager> memory_manager, IWeightsManager *weights_manager)
@@ -164,6 +187,7 @@ CLFullyConnectedLayer::CLFullyConnectedLayer(std::shared_ptr<IMemoryManager> mem
       _are_weights_reshaped(true), _is_fc_after_conv(true), _is_quantized(false), _is_prepared(false), _original_weights(nullptr)
 {
 }
+
 void CLFullyConnectedLayer::configure_mm(const CLCompileContext &compile_context, const ICLTensor *input, const ICLTensor *weights, const ICLTensor *bias, ICLTensor *output,
                                          const FullyConnectedLayerInfo &fc_info)
 {
diff --git a/src/runtime/CL/functions/CLLSTMLayer.cpp b/src/runtime/CL/functions/CLLSTMLayer.cpp
index 5036126aea..05d459c899 100644
--- a/src/runtime/CL/functions/CLLSTMLayer.cpp
+++ b/src/runtime/CL/functions/CLLSTMLayer.cpp
@@ -41,7 +41,7 @@
 #include "src/core/CL/kernels/CLGEMMMatrixMultiplyReshapedOnlyRHSKernel.h"
 #include "src/core/CL/kernels/CLGEMMReshapeLHSMatrixKernel.h"
 #include "src/core/CL/kernels/CLGEMMReshapeRHSMatrixKernel.h"
-#include "src/core/CL/kernels/CLTransposeKernel.h"
+#include "src/core/gpu/cl/kernels/ClTransposeKernel.h"
 
 namespace arm_compute
 {
@@ -51,7 +51,7 @@ using namespace arm_compute::utils::info_helpers;
 CLLSTMLayer::CLLSTMLayer(std::shared_ptr<IMemoryManager> memory_manager)
     : _memory_group(std::move(memory_manager)), _fully_connected_input_gate(), _accum_input_gate1(), _subtract_input_gate(), _pixelwise_mul_input_gate(), _activation_input_gate(),
       _fully_connected_forget_gate(), _accum_forget_gate1(), _pixelwise_mul_forget_gate(), _activation_forget_gate(), _fully_connected_cell_state(), _gemm_cell_state1(),
-      _transpose_cell_state(std::make_unique<CLTransposeKernel>()), _accum_cell_state1(), _accum_cell_state2(), _pixelwise_mul_cell_state1(), _activation_cell_state(), _cell_clip(),
+      _transpose_cell_state(std::make_unique<opencl::kernels::ClTransposeKernel>()), _accum_cell_state1(), _accum_cell_state2(), _pixelwise_mul_cell_state1(), _activation_cell_state(), _cell_clip(),
       _pixelwise_mul_cell_state2(), _fully_connected_output(), _pixelwise_mul_output_state1(), _accum_output1(), _activation_output(), _activation_output_state(), _pixelwise_mul_output_state2(),
       _fully_connected_output_state(), _projection_clip(), _copy_cell_state(), _copy_output(), _concat_scratch_buffer(), _concat_inputs_forget_gate(), _concat_weights_forget_gate(),
       _concat_weights_input_gate(), _concat_weights_output(), _ones_fill(), _mean_std_norm_input_gate(), _pixelwise_mul_input_gate_coeff(), _accum_input_gate_bias(), _mean_std_norm_forget_gate(),
@@ -256,7 +256,8 @@ void CLLSTMLayer::configure(const CLCompileContext &compile_context, const ICLTe
     _memory_group.manage(&_cell_state_out1);
     _fully_connected_cell_state.configure(compile_context, input, input_to_cell_weights, (_is_layer_norm_lstm) ? nullptr : cell_bias, &_cell_state_out1);
     _memory_group.manage(&_cell_state_out2);
-    _transpose_cell_state->configure(compile_context, recurrent_to_cell_weights, &_cell_state_out2);
+    _transpose_cell_state->configure(compile_context, recurrent_to_cell_weights->info(), _cell_state_out2.info());
+    _recurrent_to_cell_weights = recurrent_to_cell_weights;
     _memory_group.manage(&_cell_state_out3);
     _gemm_cell_state1.configure(compile_context, output_state_in, &_cell_state_out2, nullptr, &_cell_state_out3, 1.f, 0.f);
     _cell_state_out2.allocator()->allocate();
@@ -680,7 +681,12 @@ void CLLSTMLayer::run()
     }
 
     _fully_connected_cell_state.run();
-    CLScheduler::get().enqueue(*_transpose_cell_state);
+    ITensorPack pack;
+    pack.add_tensor(TensorType::ACL_SRC, _recurrent_to_cell_weights);
+    pack.add_tensor(TensorType::ACL_DST, &_cell_state_out2);
+    CLScheduler::get().enqueue_op(*_transpose_cell_state,
+                                  pack,
+                                  false);
     _gemm_cell_state1.run();
     _accum_cell_state1.run();
     if(_is_layer_norm_lstm)
diff --git a/src/runtime/CL/functions/CLPermute.cpp b/src/runtime/CL/functions/CLPermute.cpp
index eeb0169241..c1da2a9eca 100644
--- a/src/runtime/CL/functions/CLPermute.cpp
+++ b/src/runtime/CL/functions/CLPermute.cpp
@@ -43,9 +43,8 @@ CLPermute::CLPermute()
     : _impl(std::make_unique<Impl>())
 {
 }
-CLPermute::CLPermute(CLPermute &&) = default;
-CLPermute &CLPermute::operator=(CLPermute &&) = default;
-CLPermute::~CLPermute()                       = default;
+
+CLPermute::~CLPermute() = default;
 
 void CLPermute::configure(const ICLTensor *input, ICLTensor *output, const PermutationVector &perm)
 {
diff --git a/src/runtime/CL/functions/CLTranspose.cpp b/src/runtime/CL/functions/CLTranspose.cpp
index 67151be9a1..142cf73259 100644
--- a/src/runtime/CL/functions/CLTranspose.cpp
+++ b/src/runtime/CL/functions/CLTranspose.cpp
@@ -23,12 +23,27 @@
  */
 #include "arm_compute/runtime/CL/functions/CLTranspose.h"
 
-#include "src/core/CL/kernels/CLTransposeKernel.h"
-
-#include <utility>
+#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/ClTranspose.h"
 
 namespace arm_compute
 {
+struct CLTranspose::Impl
+{
+    const ICLTensor                     *src{ nullptr };
+    ICLTensor                           *dst{ nullptr };
+    std::unique_ptr<opencl::ClTranspose> op{ nullptr };
+};
+CLTranspose::CLTranspose()
+    : _impl(std::make_unique<Impl>())
+{
+}
+CLTranspose::~CLTranspose() = default;
+
 void CLTranspose::configure(const ICLTensor *input, ICLTensor *output)
 {
     configure(CLKernelLibrary::get().get_compile_context(), input, output);
@@ -36,13 +51,23 @@ void CLTranspose::configure(const ICLTensor *input, ICLTensor *output)
 
 void CLTranspose::configure(const CLCompileContext &compile_context, const ICLTensor *input, ICLTensor *output)
 {
-    auto k = std::make_unique<CLTransposeKernel>();
-    k->configure(compile_context, input, output);
-    _kernel = std::move(k);
+    ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
+    _impl->src = input;
+    _impl->dst = output;
+    _impl->op  = std::make_unique<opencl::ClTranspose>();
+    _impl->op->configure(compile_context, _impl->src->info(), _impl->dst->info());
 }
 
 Status CLTranspose::validate(const ITensorInfo *input, const ITensorInfo *output)
 {
-    return CLTransposeKernel::validate(input, output);
+    return opencl::ClTranspose::validate(input, output);
+}
+
+void CLTranspose::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/ClPermute.h b/src/runtime/gpu/cl/operators/ClPermute.h
index 65648c1434..20e7a32428 100644
--- a/src/runtime/gpu/cl/operators/ClPermute.h
+++ b/src/runtime/gpu/cl/operators/ClPermute.h
@@ -21,8 +21,8 @@
  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  * SOFTWARE.
  */
-#ifndef ARM_COMPUTE_CL_RESHAPE_H
-#define ARM_COMPUTE_CL_RESHAPE_H
+#ifndef ARM_COMPUTE_CL_PERMUTE_H
+#define ARM_COMPUTE_CL_PERMUTE_H
 
 #include "src/core/gpu/cl/ClCompileContext.h"
 #include "src/runtime/gpu/cl/IClOperator.h"
@@ -61,4 +61,4 @@ public:
 };
 } // namespace opencl
 } // namespace arm_compute
-#endif /* ARM_COMPUTE_CL_RESHAPE_H */
\ No newline at end of file
+#endif /* ARM_COMPUTE_CL_PERMUTE_H */
\ No newline at end of file
diff --git a/src/runtime/gpu/cl/operators/ClTranspose.cpp b/src/runtime/gpu/cl/operators/ClTranspose.cpp
new file mode 100644
index 0000000000..48f44282e8
--- /dev/null
+++ b/src/runtime/gpu/cl/operators/ClTranspose.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/ClTranspose.h"
+
+#include "src/core/gpu/cl/ClCompileContext.h"
+#include "src/core/gpu/cl/kernels/ClTransposeKernel.h"
+
+namespace arm_compute
+{
+namespace opencl
+{
+void ClTranspose::configure(const ClCompileContext &compile_context, const ITensorInfo *src, ITensorInfo *dst)
+{
+    auto k = std::make_unique<kernels::ClTransposeKernel>();
+    k->configure(compile_context, src, dst);
+    _kernel = std::move(k);
+}
+
+Status ClTranspose::validate(const ITensorInfo *src, const ITensorInfo *dst)
+{
+    return kernels::ClTransposeKernel::validate(src, dst);
+}
+} // namespace opencl
+} // namespace arm_compute
\ No newline at end of file
diff --git a/src/runtime/gpu/cl/operators/ClTranspose.h b/src/runtime/gpu/cl/operators/ClTranspose.h
new file mode 100644
index 0000000000..d898f677ca
--- /dev/null
+++ b/src/runtime/gpu/cl/operators/ClTranspose.h
@@ -0,0 +1,58 @@
+/*
+ * 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_TRANSPOSE_H
+#define ARM_COMPUTE_CL_TRANSPOSE_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::ClTransposeKernel */
+class ClTranspose : public IClOperator
+{
+public:
+    /** Constructor */
+    ClTranspose() = 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
+     */
+    void configure(const CLCompileContext &compile_context, const ITensorInfo *src, ITensorInfo *dst);
+    /**  Static function to check if given info will lead to a valid configuration of @ref kernels::ClTransposeKernel.
+     *
+     * @param[in] src First tensor src info. Data types supported: All.
+     * @param[in] dst Output tensor info. Data types supported: same as @p src.
+     *
+     * @return a status
+     */
+    static Status validate(const ITensorInfo *src, const ITensorInfo *dst);
+};
+} // namespace opencl
+} // namespace arm_compute
+#endif /* ARM_COMPUTE_CL_TRANSPOSE_H */