Make all CL Concatenate kernels and functions state-less

Resolves COMPMID-3995

Change-Id: I84172bed20924f1d9ae3b4d14d7b321e9494296e
Signed-off-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/4887
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>

4 files changed, 387 insertions, 233 deletions

diff --git a/src/runtime/CL/functions/CLConcatenateLayer.cpp b/src/runtime/CL/functions/CLConcatenateLayer.cpp
index 0c473a79c8..ea96e45bf8 100644
--- a/src/runtime/CL/functions/CLConcatenateLayer.cpp
+++ b/src/runtime/CL/functions/CLConcatenateLayer.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2018-2020 Arm Limited.
+ * Copyright (c) 2018-2021 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -23,242 +23,19 @@
  */
 #include "arm_compute/runtime/CL/functions/CLConcatenateLayer.h"
 
-#include "arm_compute/core/utils/misc/ShapeCalculator.h"
-#include "arm_compute/runtime/CL/CLScheduler.h"
-#include "src/core/CL/kernels/CLDepthConcatenateLayerKernel.h"
-#include "src/core/CL/kernels/CLHeightConcatenateLayerKernel.h"
-#include "src/core/CL/kernels/CLWidthConcatenate2TensorsKernel.h"
-#include "src/core/CL/kernels/CLWidthConcatenate4TensorsKernel.h"
-#include "src/core/CL/kernels/CLWidthConcatenateLayerKernel.h"
-
 #include "arm_compute/core/CL/ICLTensor.h"
-#include "arm_compute/core/Error.h"
-#include "arm_compute/core/TensorInfo.h"
-#include "arm_compute/core/Types.h"
-#include "src/core/CL/kernels/CLBatchConcatenateLayerKernel.h"
-#include "src/core/helpers/AutoConfiguration.h"
+#include "src/core/CL/ICLKernel.h"
+#include "src/runtime/gpu/cl/operators/ClConcatenate.h"
 
 namespace arm_compute
 {
-namespace experimental
-{
-CLConcatenation::CLConcatenation()
-    : _concat_kernels(),
-      _num_inputs(0),
-      _axis(Window::DimX)
-{
-}
-
-void CLConcatenation::configure(const CLCompileContext &compile_context, const std::vector<ITensorInfo *> &inputs_vector, ITensorInfo *output, size_t axis)
-{
-    ARM_COMPUTE_ERROR_ON(output == nullptr);
-    _axis       = axis;
-    _num_inputs = inputs_vector.size();
-
-    TensorShape                      output_shape = arm_compute::misc::shape_calculator::calculate_concatenate_shape(inputs_vector, _axis);
-    std::vector<const ITensorInfo *> const_inputs_vector(inputs_vector.size());
-    std::transform(inputs_vector.begin(), inputs_vector.end(), const_inputs_vector.begin(), [](ITensorInfo * t)
-    {
-        ARM_COMPUTE_ERROR_ON_NULLPTR(t);
-        return t;
-    });
-
-    // Output auto inizialitation if not yet initialized
-    auto_init_if_empty(*output, output_shape, 1, inputs_vector[0]->data_type());
-    ARM_COMPUTE_ERROR_THROW_ON(CLConcatenateLayer::validate(const_inputs_vector, output, axis));
-
-    unsigned int offset = 0;
-    switch(_axis)
-    {
-        case Window::DimX:
-        {
-            switch(_num_inputs)
-            {
-                case 2:
-                {
-                    // Configure WidthConcatenate2Tensors kernel
-                    auto kernel = std::make_unique<CLWidthConcatenate2TensorsKernel>();
-                    kernel->configure(compile_context, inputs_vector.at(0), inputs_vector.at(1), output);
-                    _concat_kernels.emplace_back(std::move(kernel));
-                    break;
-                }
-                case 4:
-                {
-                    // Configure WidthConcatenate4Tensors kernel
-                    auto kernel = std::make_unique<CLWidthConcatenate4TensorsKernel>();
-                    kernel->configure(compile_context, inputs_vector.at(0), inputs_vector.at(1), inputs_vector.at(2), inputs_vector.at(3), output);
-                    _concat_kernels.emplace_back(std::move(kernel));
-                    break;
-                }
-                default:
-                {
-                    // Configure generic case WidthConcatenate kernels
-                    for(unsigned int i = 0; i < _num_inputs; ++i)
-                    {
-                        auto kernel = std::make_unique<CLWidthConcatenateLayerKernel>();
-                        kernel->configure(compile_context, inputs_vector.at(i), offset, output);
-                        offset += inputs_vector.at(i)->dimension(_axis);
-                        _concat_kernels.emplace_back(std::move(kernel));
-                    }
-                    break;
-                }
-            }
-            break;
-        }
-        case Window::DimY:
-        {
-            for(unsigned int i = 0; i < _num_inputs; ++i)
-            {
-                auto kernel = std::make_unique<CLHeightConcatenateLayerKernel>();
-                kernel->configure(compile_context, inputs_vector.at(i), offset, output);
-                offset += inputs_vector.at(i)->dimension(_axis);
-                _concat_kernels.emplace_back(std::move(kernel));
-            }
-            break;
-        }
-        case Window::DimZ:
-        {
-            for(unsigned int i = 0; i < _num_inputs; ++i)
-            {
-                auto kernel = std::make_unique<CLDepthConcatenateLayerKernel>();
-                kernel->configure(compile_context, inputs_vector.at(i), offset, output);
-                offset += inputs_vector.at(i)->dimension(_axis);
-                _concat_kernels.emplace_back(std::move(kernel));
-            }
-            break;
-        }
-        case 3:
-        {
-            for(unsigned int i = 0; i < _num_inputs; ++i)
-            {
-                auto kernel = std::make_unique<CLBatchConcatenateLayerKernel>();
-                kernel->configure(compile_context, inputs_vector.at(i), offset, output);
-                offset += inputs_vector.at(i)->dimension(_axis);
-                _concat_kernels.emplace_back(std::move(kernel));
-            }
-            break;
-        }
-        default:
-            ARM_COMPUTE_ERROR("Axis not supported");
-    }
-}
-
-Status CLConcatenation::validate(const std::vector<const ITensorInfo *> &inputs_vector, const ITensorInfo *output, size_t axis)
-{
-    ARM_COMPUTE_RETURN_ERROR_ON(output == nullptr);
-    const unsigned int num_inputs = inputs_vector.size();
-
-    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(output);
-    ARM_COMPUTE_RETURN_ERROR_ON(num_inputs < 2);
-
-    unsigned int offset = 0;
-    switch(axis)
-    {
-        case Window::DimX:
-        {
-            switch(num_inputs)
-            {
-                case 2:
-                    // Validate WidthConcatenate2Tensors kernels if there are 2 inputs
-                    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(inputs_vector[0], inputs_vector[1]);
-                    ARM_COMPUTE_RETURN_ON_ERROR(CLWidthConcatenate2TensorsKernel::validate(inputs_vector[0], inputs_vector[1], output));
-                    break;
-                case 4:
-                    // Validate WidthConcatenate4Tensors kernels if there are 4 inputs
-                    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(inputs_vector[0], inputs_vector[1], inputs_vector[2], inputs_vector[3]);
-                    ARM_COMPUTE_RETURN_ON_ERROR(CLWidthConcatenate4TensorsKernel::validate(inputs_vector[0], inputs_vector[1], inputs_vector[2], inputs_vector[3], output));
-                    break;
-                default:
-                    // Validate generic case of WidthConcatenate kernel
-                    for(const auto &input : inputs_vector)
-                    {
-                        ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input);
-                        ARM_COMPUTE_RETURN_ON_ERROR(CLWidthConcatenateLayerKernel::validate(input, offset, output));
-                        offset += input->dimension(axis);
-                    }
-                    break;
-            }
-            break;
-        }
-        case Window::DimY:
-        {
-            for(const auto &input : inputs_vector)
-            {
-                ARM_COMPUTE_RETURN_ON_ERROR(CLHeightConcatenateLayerKernel::validate(input, offset, output));
-                offset += input->dimension(axis);
-            }
-            break;
-        }
-        case Window::DimZ:
-        {
-            for(const auto &input : inputs_vector)
-            {
-                ARM_COMPUTE_RETURN_ON_ERROR(CLDepthConcatenateLayerKernel::validate(input, offset, output));
-                offset += input->dimension(axis);
-            }
-            break;
-        }
-        case 3:
-        {
-            for(const auto &input : inputs_vector)
-            {
-                ARM_COMPUTE_RETURN_ON_ERROR(CLBatchConcatenateLayerKernel::validate(input, offset, output));
-                offset += input->dimension(axis);
-            }
-            break;
-        }
-        default:
-            ARM_COMPUTE_ERROR("Axis not supported");
-    }
-
-    if(output->total_size() != 0)
-    {
-        TensorShape output_shape = arm_compute::misc::shape_calculator::calculate_concatenate_shape(inputs_vector, axis);
-        ARM_COMPUTE_RETURN_ERROR_ON(output_shape.total_size() != output->tensor_shape().total_size());
-    }
-
-    return Status{};
-}
-
-void CLConcatenation::run(ITensorPack &tensors)
-{
-    if(tensors.empty())
-    {
-        ARM_COMPUTE_ERROR("No inputs provided");
-    }
-
-    if(static_cast<int>(tensors.size()) - 1 != static_cast<int>(_num_inputs))
-    {
-        ARM_COMPUTE_ERROR("Configured with different number of inputs");
-    }
-
-    if(_axis == Window::DimX && (_num_inputs == 2 || _num_inputs == 4))
-    {
-        ARM_COMPUTE_ERROR_ON(_concat_kernels.empty());
-        CLScheduler::get().enqueue_op(*_concat_kernels.at(0), tensors, true);
-    }
-    else
-    {
-        int i = 0;
-        for(auto &k : _concat_kernels)
-        {
-            ITensorPack pack;
-            pack.add_tensor(TensorType::ACL_SRC, tensors.get_const_tensor(ACL_SRC_VEC + i));
-            pack.add_tensor(TensorType::ACL_DST, tensors.get_tensor(ACL_DST));
-            CLScheduler::get().enqueue_op(*k, pack, true);
-            ++i;
-        }
-    }
-}
-} // namespace experimental
-
 struct CLConcatenateLayer::Impl
 {
-    std::vector<const ICLTensor *>                 srcs{};
-    ICLTensor                                     *dst{ nullptr };
-    unsigned int                                   num_inputs{ 0 };
-    unsigned int                                   axis{ 0 };
-    std::unique_ptr<experimental::CLConcatenation> op{ nullptr };
+    std::vector<const ICLTensor *>         srcs{};
+    ICLTensor                             *dst{ nullptr };
+    unsigned int                           num_inputs{ 0 };
+    unsigned int                           axis{ 0 };
+    std::unique_ptr<opencl::ClConcatenate> op{ nullptr };
 };
 
 CLConcatenateLayer::CLConcatenateLayer()
@@ -285,7 +62,7 @@ void CLConcatenateLayer::configure(const CLCompileContext &compile_context, std:
     _impl->dst        = output;
     _impl->axis       = axis;
     _impl->num_inputs = inputs_vector.size();
-    _impl->op         = std::make_unique<experimental::CLConcatenation>();
+    _impl->op         = std::make_unique<opencl::ClConcatenate>();
 
     std::vector<ITensorInfo *> inputs_vector_info;
     for(unsigned int i = 0; i < inputs_vector.size(); ++i)
@@ -298,7 +75,7 @@ void CLConcatenateLayer::configure(const CLCompileContext &compile_context, std:
 
 Status CLConcatenateLayer::validate(const std::vector<const ITensorInfo *> &inputs_vector, const ITensorInfo *output, size_t axis)
 {
-    return experimental::CLConcatenation::validate(inputs_vector, output, axis);
+    return opencl::ClConcatenate::validate(inputs_vector, output, axis);
 }
 
 void CLConcatenateLayer::run()
diff --git a/src/runtime/gpu/cl/IClOperator.h b/src/runtime/gpu/cl/IClOperator.h
new file mode 100644
index 0000000000..049bf05dc1
--- /dev/null
+++ b/src/runtime/gpu/cl/IClOperator.h
@@ -0,0 +1,37 @@
+/*
+ * 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_ICL_OPERATOR_H
+#define ARM_COMPUTE_ICL_OPERATOR_H
+
+#include "arm_compute/core/ITensorInfo.h"
+#include "arm_compute/runtime/CL/ICLOperator.h"
+
+namespace arm_compute
+{
+namespace opencl
+{
+using IClOperator = experimental::ICLOperator;
+} // namespace opencl
+} // namespace arm_compute
+#endif /* ARM_COMPUTE_ICL_OPERATOR_H */
diff --git a/src/runtime/gpu/cl/operators/ClConcatenate.cpp b/src/runtime/gpu/cl/operators/ClConcatenate.cpp
new file mode 100644
index 0000000000..4385fcfaed
--- /dev/null
+++ b/src/runtime/gpu/cl/operators/ClConcatenate.cpp
@@ -0,0 +1,254 @@
+/*
+ * 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/ClConcatenate.h"
+
+#include "arm_compute/core/utils/misc/ShapeCalculator.h"
+#include "arm_compute/runtime/CL/CLScheduler.h"
+
+#include "src/core/gpu/cl/kernels/ClBatchConcatenateKernel.h"
+#include "src/core/gpu/cl/kernels/ClDepthConcatenateKernel.h"
+#include "src/core/gpu/cl/kernels/ClHeightConcatenateKernel.h"
+#include "src/core/gpu/cl/kernels/ClWidthConcatenate2TensorsKernel.h"
+#include "src/core/gpu/cl/kernels/ClWidthConcatenate4TensorsKernel.h"
+#include "src/core/gpu/cl/kernels/ClWidthConcatenateKernel.h"
+
+#include "arm_compute/core/Error.h"
+#include "arm_compute/core/TensorInfo.h"
+#include "arm_compute/core/Types.h"
+#include "src/core/helpers/AutoConfiguration.h"
+
+namespace arm_compute
+{
+namespace opencl
+{
+ClConcatenate::ClConcatenate()
+    : _concat_kernels(),
+      _num_inputs(0),
+      _axis(Window::DimX)
+{
+}
+
+void ClConcatenate::configure(const CLCompileContext &compile_context, const std::vector<ITensorInfo *> &src_vector, ITensorInfo *dst, size_t axis)
+{
+    ARM_COMPUTE_ERROR_ON(dst == nullptr);
+    _axis       = axis;
+    _num_inputs = src_vector.size();
+
+    TensorShape                      dst_shape = arm_compute::misc::shape_calculator::calculate_concatenate_shape(src_vector, _axis);
+    std::vector<const ITensorInfo *> const_src_vector(src_vector.size());
+    std::transform(src_vector.begin(), src_vector.end(), const_src_vector.begin(), [](ITensorInfo * t)
+    {
+        ARM_COMPUTE_ERROR_ON_NULLPTR(t);
+        return t;
+    });
+
+    // dst auto inizialitation if not yet initialized
+    auto_init_if_empty(*dst, dst_shape, 1, src_vector[0]->data_type());
+    ARM_COMPUTE_ERROR_THROW_ON(ClConcatenate::validate(const_src_vector, dst, axis));
+
+    unsigned int offset = 0;
+    switch(_axis)
+    {
+        case Window::DimX:
+        {
+            switch(_num_inputs)
+            {
+                case 2:
+                {
+                    // Configure WidthConcatenate2Tensors kernel
+                    auto kernel = std::make_unique<kernels::ClWidthConcatenate2TensorsKernel>();
+                    kernel->configure(compile_context, src_vector.at(0), src_vector.at(1), dst);
+                    _concat_kernels.emplace_back(std::move(kernel));
+                    break;
+                }
+                case 4:
+                {
+                    // Configure WidthConcatenate4Tensors kernel
+                    auto kernel = std::make_unique<kernels::ClWidthConcatenate4TensorsKernel>();
+                    kernel->configure(compile_context, src_vector.at(0), src_vector.at(1), src_vector.at(2), src_vector.at(3), dst);
+                    _concat_kernels.emplace_back(std::move(kernel));
+                    break;
+                }
+                default:
+                {
+                    // Configure generic case WidthConcatenate kernels
+                    for(unsigned int i = 0; i < _num_inputs; ++i)
+                    {
+                        auto kernel = std::make_unique<kernels::ClWidthConcatenateKernel>();
+                        kernel->configure(compile_context, src_vector.at(i), offset, dst);
+                        offset += src_vector.at(i)->dimension(_axis);
+                        _concat_kernels.emplace_back(std::move(kernel));
+                    }
+                    break;
+                }
+            }
+            break;
+        }
+        case Window::DimY:
+        {
+            for(unsigned int i = 0; i < _num_inputs; ++i)
+            {
+                auto kernel = std::make_unique<kernels::ClHeightConcatenateKernel>();
+                kernel->configure(compile_context, src_vector.at(i), offset, dst);
+                offset += src_vector.at(i)->dimension(_axis);
+                _concat_kernels.emplace_back(std::move(kernel));
+            }
+            break;
+        }
+        case Window::DimZ:
+        {
+            for(unsigned int i = 0; i < _num_inputs; ++i)
+            {
+                auto kernel = std::make_unique<kernels::ClDepthConcatenateKernel>();
+                kernel->configure(compile_context, src_vector.at(i), offset, dst);
+                offset += src_vector.at(i)->dimension(_axis);
+                _concat_kernels.emplace_back(std::move(kernel));
+            }
+            break;
+        }
+        case 3:
+        {
+            for(unsigned int i = 0; i < _num_inputs; ++i)
+            {
+                auto kernel = std::make_unique<kernels::ClBatchConcatenateKernel>();
+                kernel->configure(compile_context, src_vector.at(i), offset, dst);
+                offset += src_vector.at(i)->dimension(_axis);
+                _concat_kernels.emplace_back(std::move(kernel));
+            }
+            break;
+        }
+        default:
+            ARM_COMPUTE_ERROR("Axis not supported");
+    }
+}
+
+Status ClConcatenate::validate(const std::vector<const ITensorInfo *> &src_vector, const ITensorInfo *dst, size_t axis)
+{
+    ARM_COMPUTE_RETURN_ERROR_ON(dst == nullptr);
+    const unsigned int num_inputs = src_vector.size();
+
+    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(dst);
+    ARM_COMPUTE_RETURN_ERROR_ON(num_inputs < 2);
+
+    unsigned int offset = 0;
+    switch(axis)
+    {
+        case Window::DimX:
+        {
+            switch(num_inputs)
+            {
+                case 2:
+                    // Validate WidthConcatenate2Tensors kernels if there are 2 inputs
+                    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src_vector[0], src_vector[1]);
+                    ARM_COMPUTE_RETURN_ON_ERROR(kernels::ClWidthConcatenate2TensorsKernel::validate(src_vector[0], src_vector[1], dst));
+                    break;
+                case 4:
+                    // Validate WidthConcatenate4Tensors kernels if there are 4 inputs
+                    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src_vector[0], src_vector[1], src_vector[2], src_vector[3]);
+                    ARM_COMPUTE_RETURN_ON_ERROR(kernels::ClWidthConcatenate4TensorsKernel::validate(src_vector[0], src_vector[1], src_vector[2], src_vector[3], dst));
+                    break;
+                default:
+                    // Validate generic case of WidthConcatenate kernel
+                    for(const auto &src : src_vector)
+                    {
+                        ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src);
+                        ARM_COMPUTE_RETURN_ON_ERROR(kernels::ClWidthConcatenateKernel::validate(src, offset, dst));
+                        offset += src->dimension(axis);
+                    }
+                    break;
+            }
+            break;
+        }
+        case Window::DimY:
+        {
+            for(const auto &src : src_vector)
+            {
+                ARM_COMPUTE_RETURN_ON_ERROR(kernels::ClHeightConcatenateKernel::validate(src, offset, dst));
+                offset += src->dimension(axis);
+            }
+            break;
+        }
+        case Window::DimZ:
+        {
+            for(const auto &src : src_vector)
+            {
+                ARM_COMPUTE_RETURN_ON_ERROR(kernels::ClDepthConcatenateKernel::validate(src, offset, dst));
+                offset += src->dimension(axis);
+            }
+            break;
+        }
+        case 3:
+        {
+            for(const auto &src : src_vector)
+            {
+                ARM_COMPUTE_RETURN_ON_ERROR(kernels::ClBatchConcatenateKernel::validate(src, offset, dst));
+                offset += src->dimension(axis);
+            }
+            break;
+        }
+        default:
+            ARM_COMPUTE_ERROR("Axis not supported");
+    }
+
+    if(dst->total_size() != 0)
+    {
+        TensorShape dst_shape = arm_compute::misc::shape_calculator::calculate_concatenate_shape(src_vector, axis);
+        ARM_COMPUTE_RETURN_ERROR_ON(dst_shape.total_size() != dst->tensor_shape().total_size());
+    }
+
+    return Status{};
+}
+
+void ClConcatenate::run(ITensorPack &tensors)
+{
+    if(tensors.empty())
+    {
+        ARM_COMPUTE_ERROR("No inputs provided");
+    }
+
+    if(static_cast<int>(tensors.size()) - 1 != static_cast<int>(_num_inputs))
+    {
+        ARM_COMPUTE_ERROR("Configured with different number of inputs");
+    }
+
+    if(_axis == Window::DimX && (_num_inputs == 2 || _num_inputs == 4))
+    {
+        ARM_COMPUTE_ERROR_ON(_concat_kernels.empty());
+        CLScheduler::get().enqueue_op(*_concat_kernels.at(0), tensors, true);
+    }
+    else
+    {
+        int i = 0;
+        for(auto &k : _concat_kernels)
+        {
+            ITensorPack pack;
+            pack.add_tensor(TensorType::ACL_SRC, tensors.get_const_tensor(ACL_SRC_VEC + i));
+            pack.add_tensor(TensorType::ACL_DST, tensors.get_tensor(ACL_DST));
+            CLScheduler::get().enqueue_op(*k, pack, true);
+            ++i;
+        }
+    }
+}
+} // namespace opencl
+} // namespace arm_compute
diff --git a/src/runtime/gpu/cl/operators/ClConcatenate.h b/src/runtime/gpu/cl/operators/ClConcatenate.h
new file mode 100644
index 0000000000..112e2ac6b7
--- /dev/null
+++ b/src/runtime/gpu/cl/operators/ClConcatenate.h
@@ -0,0 +1,86 @@
+/*
+ * 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_CLCONCATENATE_H
+#define ARM_COMPUTE_CLCONCATENATE_H
+
+#include "src/core/gpu/cl/ClCompileContext.h"
+#include "src/core/gpu/cl/IClKernel.h"
+#include "src/runtime/gpu/cl/IClOperator.h"
+
+#include <vector>
+
+namespace arm_compute
+{
+namespace opencl
+{
+/** Basic function to execute concatenate tensors along a given axis. This function calls the following kernels:
+ *
+ * -# @ref kernels::ClWidthConcatenateKernel (if underlying concatenation axis is 0).
+ * -# @ref kernels::ClHeightConcatenateKernel (if underlying concatenation axis is 1).
+ * -# @ref kernels::ClDepthConcatenateKernel (if underlying concatenation axis is 2).
+ * -# @ref kernels::ClBatchConcatenateKernel (if underlying concatenation axis is 3).
+ */
+class ClConcatenate : public IClOperator
+{
+public:
+    /** Default constructor */
+    ClConcatenate();
+    /** Initialise the kernel's inputs vector and dst.
+     *
+     * @note Input and dst tensor dimensions preconditions defer depending on the concatenation axis.
+     * @note Preconditions can be found respectively at @ref kernels::ClWidthConcatenateKernel,
+     *       @ref kernels::ClHeightConcatenateKernel and @ref kernels::ClDepthConcatenateKernel.
+     *
+     *
+     * @param[in]     compile_context The compile context to be used.
+     * @param[in,out] src_vector      The vectors containing all the tensors to concatenate. Data types supported: All
+     * @param[out]    dst             Destination tensor. Data types supported: same as @p src_vector.
+     * @param[in]     axis            Concatenation axis. Supported underlying concatenation axis are 0, 1, 2 and 3.
+     */
+    void configure(const ClCompileContext &compile_context, const std::vector<ITensorInfo *> &src_vector, ITensorInfo *dst, size_t axis);
+    /** Static function to check if given info will lead to a valid configuration of @ref ClConcatenate
+     *
+     * @note Input and dst tensor dimensions preconditions defer depending on the concatenation axis.
+     * @note Preconditions can be found respectively at @ref kernels::ClWidthConcatenateKernel,
+     *       @ref kernels::ClHeightConcatenateKernel and @ref kernels::ClDepthConcatenateKernel.
+     *
+     * @param[in] src_vector The vectors containing all the tensors info to concatenate. Data types supported: All
+     * @param[in] dst        Destination tensor info. Data types supported: same as @p src_vector.
+     * @param[in] axis       Concatenation axis. Supported underlying concatenation axis are 0, 1, 2 and 3.
+     *
+     * @return a status
+     */
+    static Status validate(const std::vector<const ITensorInfo *> &src_vector, const ITensorInfo *dst, size_t axis);
+
+    // Inherited methods overridden:
+    void run(ITensorPack &tensors) override;
+
+private:
+    std::vector<std::unique_ptr<IClKernel>> _concat_kernels;
+    unsigned int                            _num_inputs;
+    unsigned int                            _axis;
+};
+} // namespace opencl
+} // namespace arm_comPUTE
+#endif /* ARM_COMPUTE_CL_CONCATENATE_H */