COMPMID-3374: Remove memory state from NEConcatenateLayer kernels

* Allow the following kernels to accept backing memory at run-time:
  * NEBatchConcatenateLayerKernel
  * NEDepthConcatenateLayerKernel
  * NEHeightConcatenateLayerKernel
  * NEWidthConcatenateLayerKernel
* Allow the following functions to accept backing memory at run-time:
  * NEConcatenateLayer

Signed-off-by: Georgios Pinitas <georgios.pinitas@arm.com>
Change-Id: Ib0b6714cff7f06a52dc74d294bc3e0d72a1c2419
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/3569
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Michalis Spyrou <michalis.spyrou@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>

10 files changed, 156 insertions, 101 deletions

diff --git a/src/core/NEON/kernels/NEBatchConcatenateLayerKernel.cpp b/src/core/NEON/kernels/NEBatchConcatenateLayerKernel.cpp
index 65789160f6..c597afd804 100644
--- a/src/core/NEON/kernels/NEBatchConcatenateLayerKernel.cpp
+++ b/src/core/NEON/kernels/NEBatchConcatenateLayerKernel.cpp
@@ -141,21 +141,19 @@ Status validate_arguments(const ITensorInfo *input, unsigned int batch_offset, c
 } // namespace
 
 NEBatchConcatenateLayerKernel::NEBatchConcatenateLayerKernel()
-    : _func(nullptr), _input(nullptr), _output(nullptr), _batch_offset(0)
+    : _func(nullptr), _batch_offset(0)
 {
 }
 
-void NEBatchConcatenateLayerKernel::configure(const ITensor *input, unsigned int batch_offset, ITensor *output)
+void NEBatchConcatenateLayerKernel::configure(const ITensorInfo *input, unsigned int batch_offset, ITensorInfo *output)
 {
     ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
-    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), batch_offset, output->info()));
+    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input, batch_offset, output));
 
     _func         = nullptr;
-    _input        = input;
-    _output       = output;
     _batch_offset = batch_offset;
 
-    switch(input->info()->data_type())
+    switch(input->data_type())
     {
         case DataType::S8:
         case DataType::U8:
@@ -178,10 +176,10 @@ void NEBatchConcatenateLayerKernel::configure(const ITensor *input, unsigned int
     }
 
     // Configure kernel window
-    Window      win = calculate_max_window(*output->info(), Steps());
+    Window      win = calculate_max_window(*output, Steps());
     Coordinates coord;
-    coord.set_num_dimensions(output->info()->num_dimensions());
-    output->info()->set_valid_region(ValidRegion(coord, output->info()->tensor_shape()));
+    coord.set_num_dimensions(output->num_dimensions());
+    output->set_valid_region(ValidRegion(coord, output->tensor_shape()));
     INEKernel::configure(win);
 }
 
@@ -193,13 +191,14 @@ Status NEBatchConcatenateLayerKernel::validate(const arm_compute::ITensorInfo *i
     return Status{};
 }
 
-void NEBatchConcatenateLayerKernel::run(const Window &window, const ThreadInfo &info)
+void NEBatchConcatenateLayerKernel::run_op(const InputTensorMap &inputs, const OutputTensorMap &outputs,
+                                           const Window &window, const ThreadInfo &info)
 {
     ARM_COMPUTE_UNUSED(info);
     ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
     ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window);
     ARM_COMPUTE_ERROR_ON(_func == nullptr);
 
-    (*_func)(_input, _output, _batch_offset, window);
+    (*_func)(inputs.at(TensorType::ACL_SRC), outputs.at(TensorType::ACL_DST), _batch_offset, window);
 }
 } // namespace arm_compute
diff --git a/src/core/NEON/kernels/NEDepthConcatenateLayerKernel.cpp b/src/core/NEON/kernels/NEDepthConcatenateLayerKernel.cpp
index a95d711f43..49e10de94e 100644
--- a/src/core/NEON/kernels/NEDepthConcatenateLayerKernel.cpp
+++ b/src/core/NEON/kernels/NEDepthConcatenateLayerKernel.cpp
@@ -142,21 +142,19 @@ Status validate_arguments(const ITensorInfo *input, unsigned int depth_offset, c
 } // namespace
 
 NEDepthConcatenateLayerKernel::NEDepthConcatenateLayerKernel()
-    : _func(nullptr), _input(nullptr), _output(nullptr), _depth_offset(0)
+    : _func(nullptr), _depth_offset(0)
 {
 }
 
-void NEDepthConcatenateLayerKernel::configure(const ITensor *input, unsigned int depth_offset, ITensor *output)
+void NEDepthConcatenateLayerKernel::configure(const ITensorInfo *input, unsigned int depth_offset, ITensorInfo *output)
 {
     ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
-    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), depth_offset, output->info()));
+    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input, depth_offset, output));
 
     _func         = nullptr;
-    _input        = input;
-    _output       = output;
     _depth_offset = depth_offset;
 
-    switch(input->info()->data_type())
+    switch(input->data_type())
     {
         case DataType::QASYMM8:
             _func = &depth_concat<uint8_t>;
@@ -175,11 +173,11 @@ void NEDepthConcatenateLayerKernel::configure(const ITensor *input, unsigned int
     }
 
     // Configure kernel window
-    Window      win = calculate_max_window(*output->info(), Steps());
+    Window      win = calculate_max_window(*output, Steps());
     Coordinates coord;
-    coord.set_num_dimensions(output->info()->num_dimensions());
+    coord.set_num_dimensions(output->num_dimensions());
 
-    output->info()->set_valid_region(ValidRegion(coord, output->info()->tensor_shape()));
+    output->set_valid_region(ValidRegion(coord, output->tensor_shape()));
     INEKernel::configure(win);
 }
 
@@ -191,13 +189,14 @@ Status NEDepthConcatenateLayerKernel::validate(const arm_compute::ITensorInfo *i
     return Status{};
 }
 
-void NEDepthConcatenateLayerKernel::run(const Window &window, const ThreadInfo &info)
+void NEDepthConcatenateLayerKernel::run_op(const InputTensorMap &inputs, const OutputTensorMap &outputs,
+                                           const Window &window, const ThreadInfo &info)
 {
     ARM_COMPUTE_UNUSED(info);
     ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
     ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window);
     ARM_COMPUTE_ERROR_ON(_func == nullptr);
 
-    (*_func)(_input, _output, _depth_offset, window);
+    (*_func)(inputs.at(TensorType::ACL_SRC), outputs.at(TensorType::ACL_DST), _depth_offset, window);
 }
 } // namespace arm_compute
diff --git a/src/core/NEON/kernels/NEHeightConcatenateLayerKernel.cpp b/src/core/NEON/kernels/NEHeightConcatenateLayerKernel.cpp
index 0adf996cca..d4043e02b7 100644
--- a/src/core/NEON/kernels/NEHeightConcatenateLayerKernel.cpp
+++ b/src/core/NEON/kernels/NEHeightConcatenateLayerKernel.cpp
@@ -58,24 +58,23 @@ Status validate_arguments(const ITensorInfo *input, unsigned int height_offset,
 } // namespace
 
 NEHeightConcatenateLayerKernel::NEHeightConcatenateLayerKernel()
-    : _input(nullptr), _output(nullptr), _height_offset(0)
+    : _height_offset(0)
 {
 }
 
-void NEHeightConcatenateLayerKernel::configure(const ITensor *input, unsigned int height_offset, ITensor *output)
+void NEHeightConcatenateLayerKernel::configure(const ITensorInfo *input, unsigned int height_offset, ITensorInfo *output)
 {
+    ARM_COMPUTE_UNUSED(input);
     ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
-    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), height_offset, output->info()));
+    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input, height_offset, output));
 
-    _input         = input;
-    _output        = output;
     _height_offset = height_offset;
 
     // Configure kernel window
-    Window      win = calculate_max_window(*output->info(), Steps());
+    Window      win = calculate_max_window(*output, Steps());
     Coordinates coord;
-    coord.set_num_dimensions(output->info()->num_dimensions());
-    output->info()->set_valid_region(ValidRegion(coord, output->info()->tensor_shape()));
+    coord.set_num_dimensions(output->num_dimensions());
+    output->set_valid_region(ValidRegion(coord, output->tensor_shape()));
     INEKernel::configure(win);
 }
 
@@ -85,30 +84,34 @@ Status NEHeightConcatenateLayerKernel::validate(const ITensorInfo *input, unsign
     return Status{};
 }
 
-void NEHeightConcatenateLayerKernel::run(const Window &window, const ThreadInfo &info)
+void NEHeightConcatenateLayerKernel::run_op(const InputTensorMap &inputs, const OutputTensorMap &outputs,
+                                            const Window &window, const ThreadInfo &info)
 {
     ARM_COMPUTE_UNUSED(info);
     ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
     ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window);
 
+    const auto src = inputs.at(TensorType::ACL_SRC);
+    auto       dst = outputs.at(TensorType::ACL_DST);
+
     // Offset output pointer to the correct position
-    uint8_t *output_ptr = _output->buffer() + _output->info()->offset_first_element_in_bytes() + _height_offset * _output->info()->strides_in_bytes()[Window::DimY];
+    uint8_t *output_ptr = dst->buffer() + dst->info()->offset_first_element_in_bytes() + _height_offset * dst->info()->strides_in_bytes()[Window::DimY];
 
     const auto window_start_x = static_cast<int>(window.x().start());
-    const auto window_end_x   = static_cast<int>(window.x().end()) * static_cast<int>(_output->info()->element_size()); 
-    const int window_step_x = 16;
+    const auto window_end_x   = static_cast<int>(window.x().end()) * static_cast<int>(dst->info()->element_size());
+    const int  window_step_x  = 16;
 
     Window win{ window };
     win.set(Window::DimX, Window::Dimension(0, 1, 1));
-    win.set(Window::DimY, Window::Dimension(0, _input->info()->tensor_shape().y(), 1));
+    win.set(Window::DimY, Window::Dimension(0, src->info()->tensor_shape().y(), 1));
 
     // Create iterators
-    Iterator input(_input, win);
-    Iterator output(_output, win);
+    Iterator input(src, win);
+    Iterator output(dst, win);
 
-    const DataType                 dt           = _input->info()->data_type();
-    const UniformQuantizationInfo &input_qinfo  = _input->info()->quantization_info().uniform();
-    const UniformQuantizationInfo &output_qinfo = _output->info()->quantization_info().uniform();
+    const DataType                 dt           = src->info()->data_type();
+    const UniformQuantizationInfo &input_qinfo  = src->info()->quantization_info().uniform();
+    const UniformQuantizationInfo &output_qinfo = dst->info()->quantization_info().uniform();
     if(dt == DataType::QASYMM8 && input_qinfo != output_qinfo)
     {
         execute_window_loop(win, [&](const Coordinates &)
diff --git a/src/core/NEON/kernels/NEWidthConcatenateLayerKernel.cpp b/src/core/NEON/kernels/NEWidthConcatenateLayerKernel.cpp
index f5bdfae5d6..1b32e3614e 100644
--- a/src/core/NEON/kernels/NEWidthConcatenateLayerKernel.cpp
+++ b/src/core/NEON/kernels/NEWidthConcatenateLayerKernel.cpp
@@ -58,24 +58,22 @@ Status validate_arguments(const ITensorInfo *input, unsigned int width_offset, c
 } // namespace
 
 NEWidthConcatenateLayerKernel::NEWidthConcatenateLayerKernel()
-    : _input(nullptr), _output(nullptr), _width_offset(0)
+    : _width_offset(0)
 {
 }
 
-void NEWidthConcatenateLayerKernel::configure(const ITensor *input, unsigned int width_offset, ITensor *output)
+void NEWidthConcatenateLayerKernel::configure(const ITensorInfo *input, unsigned int width_offset, ITensorInfo *output)
 {
     ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
-    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), width_offset, output->info()));
+    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input, width_offset, output));
 
-    _input        = input;
-    _output       = output;
     _width_offset = width_offset;
 
     // Configure kernel window
-    Window      win = calculate_max_window(*input->info(), Steps());
+    Window      win = calculate_max_window(*input, Steps());
     Coordinates coord;
-    coord.set_num_dimensions(output->info()->num_dimensions());
-    output->info()->set_valid_region(ValidRegion(coord, output->info()->tensor_shape()));
+    coord.set_num_dimensions(output->num_dimensions());
+    output->set_valid_region(ValidRegion(coord, output->tensor_shape()));
 
     INEKernel::configure(win);
 }
@@ -86,28 +84,32 @@ Status NEWidthConcatenateLayerKernel::validate(const ITensorInfo *input, unsigne
     return Status{};
 }
 
-void NEWidthConcatenateLayerKernel::run(const Window &window, const ThreadInfo &info)
+void NEWidthConcatenateLayerKernel::run_op(const InputTensorMap &inputs, const OutputTensorMap &outputs,
+                                           const Window &window, const ThreadInfo &info)
 {
     ARM_COMPUTE_UNUSED(info);
     ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
     ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window);
 
+    const auto src = inputs.at(TensorType::ACL_SRC);
+    auto       dst = outputs.at(TensorType::ACL_DST);
+
     // Offset output pointer to the correct position
-    uint8_t *output_ptr = _output->buffer() + _output->info()->offset_first_element_in_bytes() + _width_offset * _output->info()->strides_in_bytes()[0];
+    uint8_t *output_ptr = dst->buffer() + dst->info()->offset_first_element_in_bytes() + _width_offset * dst->info()->strides_in_bytes()[0];
 
     const auto    window_start_x = static_cast<int>(window.x().start());
-    const auto    window_end_x   = static_cast<int>(window.x().end()) * static_cast<int>(_output->info()->element_size());
+    const auto    window_end_x   = static_cast<int>(window.x().end()) * static_cast<int>(dst->info()->element_size());
     constexpr int window_step_x  = 16;
 
     Window win{ window };
     win.set(Window::DimX, Window::Dimension(0, 1, 1));
 
     // Create iterators
-    Iterator                       input(_input, win);
-    Iterator                       output(_output, win);
-    const DataType                 dt           = _input->info()->data_type();
-    const UniformQuantizationInfo &input_qinfo  = _input->info()->quantization_info().uniform();
-    const UniformQuantizationInfo &output_qinfo = _output->info()->quantization_info().uniform();
+    Iterator                       input(src, win);
+    Iterator                       output(dst, win);
+    const DataType                 dt           = src->info()->data_type();
+    const UniformQuantizationInfo &input_qinfo  = src->info()->quantization_info().uniform();
+    const UniformQuantizationInfo &output_qinfo = dst->info()->quantization_info().uniform();
     if(dt == DataType::QASYMM8 && input_qinfo != output_qinfo)
     {
         execute_window_loop(win, [&](const Coordinates &)
diff --git a/src/graph/backends/CL/CLFunctionsFactory.cpp b/src/graph/backends/CL/CLFunctionsFactory.cpp
index 917741a2b7..ceff6e5cef 100644
--- a/src/graph/backends/CL/CLFunctionsFactory.cpp
+++ b/src/graph/backends/CL/CLFunctionsFactory.cpp
@@ -42,6 +42,7 @@ namespace backends
 struct CLTargetInfo
 {
     using TensorType         = arm_compute::ICLTensor;
+    using SrcTensorType      = TensorType;
     using TensorConcreteType = CLTensor;
     static Target TargetType;
 };
diff --git a/src/graph/backends/GLES/GCFunctionsFactory.cpp b/src/graph/backends/GLES/GCFunctionsFactory.cpp
index a78f51cdbd..8ecb593e11 100644
--- a/src/graph/backends/GLES/GCFunctionsFactory.cpp
+++ b/src/graph/backends/GLES/GCFunctionsFactory.cpp
@@ -40,7 +40,8 @@ namespace backends
 /** Target specific information structure used to pass information to the layer templates */
 struct GCTargetInfo
 {
-    using TensorType = arm_compute::IGCTensor;
+    using TensorType    = arm_compute::IGCTensor;
+    using SrcTensorType = TensorType;
     static Target TargetType;
 };
 
diff --git a/src/graph/backends/NEON/NEFunctionFactory.cpp b/src/graph/backends/NEON/NEFunctionFactory.cpp
index 2f313081e0..4fee630192 100644
--- a/src/graph/backends/NEON/NEFunctionFactory.cpp
+++ b/src/graph/backends/NEON/NEFunctionFactory.cpp
@@ -47,6 +47,7 @@ namespace backends
 struct NETargetInfo
 {
     using TensorType         = arm_compute::ITensor;
+    using SrcTensorType      = const arm_compute::ITensor;
     using TensorConcreteType = arm_compute::Tensor;
     static Target TargetType;
 };
diff --git a/src/runtime/NEON/functions/NEConcatenateLayer.cpp b/src/runtime/NEON/functions/NEConcatenateLayer.cpp
index 9c480a0d50..37cdd15529 100644
--- a/src/runtime/NEON/functions/NEConcatenateLayer.cpp
+++ b/src/runtime/NEON/functions/NEConcatenateLayer.cpp
@@ -39,58 +39,31 @@
 
 namespace arm_compute
 {
-NEConcatenateLayer::NEConcatenateLayer()
-    : _concat_kernels(),
-      _num_inputs(0),
-      _axis(Window::DimX)
-{
-}
-
-void NEConcatenateLayer::configure(std::vector<ITensor *> inputs_vector, ITensor *output, size_t axis)
-{
-    configure_internal(std::move(inputs_vector), output, axis);
-}
-
-void NEConcatenateLayer::configure(std::vector<const ITensor *> inputs_vector, ITensor *output, size_t axis)
+namespace experimental
 {
-    configure_internal(std::move(inputs_vector), output, axis);
-}
-
-Status NEConcatenateLayer::validate(const std::vector<ITensorInfo *> &inputs_vector, const ITensorInfo *output, size_t axis)
-{
-    return validate_internal(inputs_vector, output, axis);
-}
-
-Status NEConcatenateLayer::validate(const std::vector<const ITensorInfo *> &inputs_vector, const ITensorInfo *output, size_t axis)
+NEConcatenateLayer::NEConcatenateLayer()
+    : _concat_kernels(), _num_inputs(0), _axis(0)
 {
-    return validate_internal(inputs_vector, output, axis);
 }
 
-template <typename TensorType, typename>
-void NEConcatenateLayer::configure_internal(std::vector<TensorType *> &&inputs_vector, ITensor *output, size_t axis)
+void NEConcatenateLayer::configure(const std::vector<const ITensorInfo *> &inputs_vector, ITensorInfo *output, size_t axis)
 {
     ARM_COMPUTE_ERROR_ON(output == nullptr);
+
     _axis       = axis;
     _num_inputs = inputs_vector.size();
 
-    std::vector<ITensorInfo *> inputs_vector_info;
-    inputs_vector_info.reserve(_num_inputs);
-    for(unsigned int i = 0; i < _num_inputs; ++i)
-    {
-        ARM_COMPUTE_ERROR_ON_NULLPTR(inputs_vector.at(i));
-        inputs_vector_info.emplace_back(inputs_vector.at(i)->info());
-    }
-    TensorShape output_shape = arm_compute::misc::shape_calculator::calculate_concatenate_shape(inputs_vector, _axis);
+    TensorShape output_shape = arm_compute::misc::shape_calculator::calculate_concatenate_shape(inputs_vector, axis);
 
     // Output auto inizialitation if not yet initialized
-    auto_init_if_empty(*output->info(), output_shape, 1, inputs_vector[0]->info()->data_type());
-    ARM_COMPUTE_ERROR_THROW_ON(NEConcatenateLayer::validate(inputs_vector_info, output->info(), axis));
+    auto_init_if_empty(*output, output_shape, 1, inputs_vector[0]->data_type());
+    ARM_COMPUTE_ERROR_THROW_ON(NEConcatenateLayer::validate(inputs_vector, output, axis));
 
     unsigned int offset = 0;
 
     for(unsigned int i = 0; i < _num_inputs; ++i)
     {
-        switch(_axis)
+        switch(axis)
         {
             case Window::DimX:
             {
@@ -123,12 +96,11 @@ void NEConcatenateLayer::configure_internal(std::vector<TensorType *> &&inputs_v
             default:
                 ARM_COMPUTE_ERROR("Axis not supported");
         }
-        offset += inputs_vector.at(i)->info()->dimension(_axis);
+        offset += inputs_vector.at(i)->dimension(axis);
     }
 }
 
-template <typename TensorInfoType, typename>
-Status NEConcatenateLayer::validate_internal(const std::vector<TensorInfoType *> &inputs_vector, const ITensorInfo *output, size_t axis)
+Status NEConcatenateLayer::validate(const std::vector<const ITensorInfo *> &inputs_vector, const ITensorInfo *output, size_t axis)
 {
     ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(output);
     ARM_COMPUTE_RETURN_ERROR_ON(inputs_vector.size() < 2);
@@ -174,11 +146,88 @@ Status NEConcatenateLayer::validate_internal(const std::vector<TensorInfoType *>
     return Status{};
 }
 
+MemoryRequirements NEConcatenateLayer::workspace() const
+{
+    return MemoryRequirements{};
+}
+
+void NEConcatenateLayer::run(InputTensorMap inputs, OutputTensorMap outputs, OperatorTensorMap workspace)
+{
+    ARM_COMPUTE_UNUSED(workspace);
+
+    if(inputs.empty() || outputs.empty())
+    {
+        ARM_COMPUTE_ERROR("No inputs provided");
+    }
+
+    if(inputs.size() != _num_inputs)
+    {
+        ARM_COMPUTE_ERROR("Configured with different number of inputs");
+    }
+
+    int i = 0;
+    for(auto &k : _concat_kernels)
+    {
+        const InputTensorMap input = { { TensorType::ACL_SRC, inputs.at(ACL_SRC_VEC + i) } };
+        NEScheduler::get().schedule_op(k.get(), Window::DimY, input, outputs);
+        ++i;
+    }
+}
+} // namespace experimental
+
+struct NEConcatenateLayer::Impl
+{
+    std::vector<const ITensor *>                      srcs{};
+    ITensor                                          *dst{ nullptr };
+    unsigned int                                      num_inputs{ 0 };
+    unsigned int                                      axis{ 0 };
+    std::unique_ptr<experimental::NEConcatenateLayer> op{ nullptr };
+};
+
+NEConcatenateLayer::NEConcatenateLayer()
+    : _impl(support::cpp14::make_unique<Impl>())
+{
+}
+
+NEConcatenateLayer::NEConcatenateLayer(NEConcatenateLayer &&) = default;
+
+NEConcatenateLayer &NEConcatenateLayer::operator=(NEConcatenateLayer &&) = default;
+
+NEConcatenateLayer::~NEConcatenateLayer() = default;
+
+void NEConcatenateLayer::configure(std::vector<const ITensor *> inputs_vector, ITensor *output, size_t axis)
+{
+    ARM_COMPUTE_ERROR_ON(output == nullptr);
+
+    _impl->srcs       = inputs_vector;
+    _impl->dst        = output;
+    _impl->axis       = axis;
+    _impl->num_inputs = inputs_vector.size();
+    _impl->op         = arm_compute::support::cpp14::make_unique<experimental::NEConcatenateLayer>();
+
+    std::vector<const ITensorInfo *> inputs_vector_info;
+    for(unsigned int i = 0; i < inputs_vector.size(); ++i)
+    {
+        ARM_COMPUTE_ERROR_ON_NULLPTR(inputs_vector.at(i));
+        inputs_vector_info.emplace_back(inputs_vector.at(i)->info());
+    }
+    _impl->op->configure(inputs_vector_info, _impl->dst->info(), axis);
+}
+
+Status NEConcatenateLayer::validate(const std::vector<const ITensorInfo *> &inputs_vector, const ITensorInfo *output, size_t axis)
+{
+    return experimental::NEConcatenateLayer::validate(inputs_vector, output, axis);
+}
+
 void NEConcatenateLayer::run()
 {
-    for(auto &kernel : _concat_kernels)
+    InputTensorMap srcs;
+    for(unsigned i = 0; i < _impl->num_inputs; ++i)
     {
-        NEScheduler::get().schedule(kernel.get(), Window::DimY);
+        srcs.insert(std::make_pair(TensorType::ACL_SRC_VEC + i, _impl->srcs.at(i)));
     }
+    const OutputTensorMap dst{ { TensorType::ACL_DST, _impl->dst } };
+
+    _impl->op->run(srcs, dst, {});
 }
 } // namespace arm_compute
diff --git a/src/runtime/NEON/functions/NELSTMLayer.cpp b/src/runtime/NEON/functions/NELSTMLayer.cpp
index f89b3e999c..dca274acd2 100644
--- a/src/runtime/NEON/functions/NELSTMLayer.cpp
+++ b/src/runtime/NEON/functions/NELSTMLayer.cpp
@@ -347,7 +347,7 @@ void NELSTMLayer::configure(const ITensor *input,
     _copy_output.configure(output_state_out, output);
 
     // Vector for holding the tensors to store in scratch buffer
-    std::vector<ITensor *> scratch_inputs;
+    std::vector<const ITensor *> scratch_inputs;
     if(!lstm_params.has_cifg_opt())
     {
         scratch_inputs.emplace_back(input_gate_out);
@@ -579,7 +579,7 @@ Status NELSTMLayer::validate(const ITensorInfo *input,
     ARM_COMPUTE_RETURN_ON_ERROR(NECopyKernel::validate(output_state_out, output));
 
     // Validate scratch concatenation
-    std::vector<ITensorInfo *> inputs_vector_info_raw;
+    std::vector<const ITensorInfo *> inputs_vector_info_raw;
     if(!lstm_params.has_cifg_opt())
     {
         inputs_vector_info_raw.push_back(&input_gate);
diff --git a/src/runtime/NEON/functions/NEPadLayer.cpp b/src/runtime/NEON/functions/NEPadLayer.cpp
index da9a425d9b..21c349ba95 100644
--- a/src/runtime/NEON/functions/NEPadLayer.cpp
+++ b/src/runtime/NEON/functions/NEPadLayer.cpp
@@ -117,7 +117,7 @@ void NEPadLayer::configure_reflect_symmetric_mode(ITensor *input, ITensor *outpu
             const int32_t end_mask_after    = ends_after[i] < 0 ? ~0 : ~(1u << i);
 
             // Reflect the input values for the padding before and after the input.
-            std::vector<ITensor *> concat_vector;
+            std::vector<const ITensor *> concat_vector;
             if(_padding[i].first > 0)
             {
                 if(i < prev->info()->num_dimensions())