COMPMID-2515: Merge optimized depthwise convolution to the generic depthwise convolution function

3RDPARTY_UPDATE

Change-Id: Iff9e915c5329c617527b6f5042979f4e21a8b2b8
Signed-off-by: Manuel Bottini <manuel.bottini@arm.com>
Reviewed-on: https://review.mlplatform.org/c/2022
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Giorgio Arena <giorgio.arena@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>

8 files changed, 484 insertions, 621 deletions

diff --git a/src/graph/TypeLoader.cpp b/src/graph/TypeLoader.cpp
index b63672b39b..81a405b961 100644
--- a/src/graph/TypeLoader.cpp
+++ b/src/graph/TypeLoader.cpp
@@ -131,7 +131,6 @@ DepthwiseConvolutionMethod depthwise_convolution_method_from_name(const std::str
     static const std::map<std::string, DepthwiseConvolutionMethod> methods =
     {
         { "default", DepthwiseConvolutionMethod::Default },
-        { "gemv", DepthwiseConvolutionMethod::GEMV },
         { "optimized3x3", DepthwiseConvolutionMethod::Optimized3x3 },
     };
 
diff --git a/src/graph/backends/CL/CLFunctionsFactory.cpp b/src/graph/backends/CL/CLFunctionsFactory.cpp
index 6d231f2ef3..d53b634bb1 100644
--- a/src/graph/backends/CL/CLFunctionsFactory.cpp
+++ b/src/graph/backends/CL/CLFunctionsFactory.cpp
@@ -56,13 +56,6 @@ struct CLConvolutionLayerFunctions
     using WinogradConvolutionLayer = CLWinogradConvolutionLayer;
 };
 
-/** Collection of CL depthwise convolution functions */
-struct CLDepthwiseConvolutionLayerFunctions
-{
-    using GenericDepthwiseConvolutionLayer   = CLDepthwiseConvolutionLayer;
-    using OptimizedDepthwiseConvolutionLayer = CLDepthwiseConvolutionLayer;
-};
-
 /** Collection of CL element-wise functions */
 struct CLEltwiseFunctions
 {
@@ -249,7 +242,7 @@ std::unique_ptr<IFunction> CLFunctionFactory::create(INode *node, GraphContext &
         case NodeType::ConcatenateLayer:
             return detail::create_concatenate_layer<CLConcatenateLayer, CLTargetInfo>(*polymorphic_downcast<ConcatenateLayerNode *>(node));
         case NodeType::DepthwiseConvolutionLayer:
-            return detail::create_depthwise_convolution_layer<CLDepthwiseConvolutionLayerFunctions, CLTargetInfo>(*polymorphic_downcast<DepthwiseConvolutionLayerNode *>(node));
+            return detail::create_depthwise_convolution_layer<CLDepthwiseConvolutionLayer, CLTargetInfo>(*polymorphic_downcast<DepthwiseConvolutionLayerNode *>(node));
         case NodeType::DetectionOutputLayer:
             return detail::create_detection_output_layer<CPPDetectionOutputLayer, CLTargetInfo>(*polymorphic_downcast<DetectionOutputLayerNode *>(node));
         case NodeType::DetectionPostProcessLayer:
diff --git a/src/graph/backends/CL/CLNodeValidator.cpp b/src/graph/backends/CL/CLNodeValidator.cpp
index 40ec508767..a2786187a2 100644
--- a/src/graph/backends/CL/CLNodeValidator.cpp
+++ b/src/graph/backends/CL/CLNodeValidator.cpp
@@ -58,8 +58,7 @@ Status CLNodeValidator::validate(INode *node)
                    CLGEMMConvolutionLayer,
                    CLWinogradConvolutionLayer>(*polymorphic_downcast<ConvolutionLayerNode *>(node));
         case NodeType::DepthwiseConvolutionLayer:
-            return detail::validate_depthwise_convolution_layer<CLDepthwiseConvolutionLayer,
-                   CLDepthwiseConvolutionLayer3x3>(*polymorphic_downcast<DepthwiseConvolutionLayerNode *>(node));
+            return detail::validate_depthwise_convolution_layer<CLDepthwiseConvolutionLayer>(*polymorphic_downcast<DepthwiseConvolutionLayerNode *>(node));
         case NodeType::DetectionOutputLayer:
             return detail::validate_detection_output_layer<CPPDetectionOutputLayer>(*polymorphic_downcast<DetectionOutputLayerNode *>(node));
         case NodeType::DetectionPostProcessLayer:
diff --git a/src/graph/backends/GLES/GCNodeValidator.cpp b/src/graph/backends/GLES/GCNodeValidator.cpp
index 9cbb9a12ef..9d848ab3b1 100644
--- a/src/graph/backends/GLES/GCNodeValidator.cpp
+++ b/src/graph/backends/GLES/GCNodeValidator.cpp
@@ -58,7 +58,6 @@ Status validate_depthwise_convolution_layer(DepthwiseConvolutionLayerNode &node)
     // TODO (geopin01) : Switch when validation is implemented
     // Validate function
     ARM_COMPUTE_RETURN_ERROR_ON_MSG(weights->tensor_shape().x() != 3 && weights->tensor_shape().y() != 3, "Unsupported depthwise convolution");
-    node.set_depthwise_convolution_method(DepthwiseConvolutionMethod::Optimized3x3);
 
     return Status{};
 }
diff --git a/src/graph/backends/NEON/NEFunctionFactory.cpp b/src/graph/backends/NEON/NEFunctionFactory.cpp
index 45e9727133..d8b0ae92ea 100644
--- a/src/graph/backends/NEON/NEFunctionFactory.cpp
+++ b/src/graph/backends/NEON/NEFunctionFactory.cpp
@@ -62,13 +62,6 @@ struct NEConvolutionLayerFunctions
     using WinogradConvolutionLayer = NEWinogradConvolutionLayer;
 };
 
-/** Collection of CL depthwise convolution functions */
-struct NEDepthwiseConvolutionLayerFunctions
-{
-    using GenericDepthwiseConvolutionLayer   = NEDepthwiseConvolutionLayer;
-    using OptimizedDepthwiseConvolutionLayer = NEDepthwiseConvolutionLayerOptimized;
-};
-
 /** Collection of CL element-wise functions */
 struct NEEltwiseFunctions
 {
@@ -213,7 +206,7 @@ std::unique_ptr<IFunction> NEFunctionFactory::create(INode *node, GraphContext &
         case NodeType::ConcatenateLayer:
             return detail::create_concatenate_layer<NEConcatenateLayer, NETargetInfo>(*polymorphic_downcast<ConcatenateLayerNode *>(node));
         case NodeType::DepthwiseConvolutionLayer:
-            return detail::create_depthwise_convolution_layer<NEDepthwiseConvolutionLayerFunctions, NETargetInfo>(*polymorphic_downcast<DepthwiseConvolutionLayerNode *>(node));
+            return detail::create_depthwise_convolution_layer<NEDepthwiseConvolutionLayer, NETargetInfo>(*polymorphic_downcast<DepthwiseConvolutionLayerNode *>(node));
         case NodeType::DetectionOutputLayer:
             return detail::create_detection_output_layer<CPPDetectionOutputLayer, NETargetInfo>(*polymorphic_downcast<DetectionOutputLayerNode *>(node));
         case NodeType::DetectionPostProcessLayer:
diff --git a/src/graph/backends/NEON/NENodeValidator.cpp b/src/graph/backends/NEON/NENodeValidator.cpp
index 734b3401f7..0b53657c42 100644
--- a/src/graph/backends/NEON/NENodeValidator.cpp
+++ b/src/graph/backends/NEON/NENodeValidator.cpp
@@ -58,8 +58,7 @@ Status NENodeValidator::validate(INode *node)
                    NEGEMMConvolutionLayer,
                    NEWinogradConvolutionLayer>(*polymorphic_downcast<ConvolutionLayerNode *>(node));
         case NodeType::DepthwiseConvolutionLayer:
-            return detail::validate_depthwise_convolution_layer<NEDepthwiseConvolutionLayer,
-                   NEDepthwiseConvolutionLayer3x3>(*polymorphic_downcast<DepthwiseConvolutionLayerNode *>(node));
+            return detail::validate_depthwise_convolution_layer<NEDepthwiseConvolutionLayer>(*polymorphic_downcast<DepthwiseConvolutionLayerNode *>(node));
         case NodeType::DetectionOutputLayer:
             return detail::validate_detection_output_layer<CPPDetectionOutputLayer>(*polymorphic_downcast<DetectionOutputLayerNode *>(node));
         case NodeType::DetectionPostProcessLayer:
diff --git a/src/runtime/CL/functions/CLDepthwiseConvolutionLayer.cpp b/src/runtime/CL/functions/CLDepthwiseConvolutionLayer.cpp
index 5ac7a7a7c6..168d7d5c84 100644
--- a/src/runtime/CL/functions/CLDepthwiseConvolutionLayer.cpp
+++ b/src/runtime/CL/functions/CLDepthwiseConvolutionLayer.cpp
@@ -38,15 +38,262 @@ namespace arm_compute
 using namespace arm_compute::misc;
 using namespace arm_compute::misc::shape_calculator;
 
+namespace
+{
+Status validate_arguments_3x3(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info,
+                              unsigned int depth_multiplier, ActivationLayerInfo act_info, GPUTarget gpu_target, const Size2D &dilation)
+{
+    // This function should be removed and incorporated inside CLDepthwiseConvolutionLayerInternal3x3 once CLDepthwiseConvolutionLayer3x3 is properly removed
+    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, weights, output);
+    ARM_COMPUTE_RETURN_ERROR_ON(input->data_layout() == DataLayout::UNKNOWN);
+
+    const bool                      is_quantized           = is_data_type_quantized_asymmetric(input->data_type());
+    const bool                      is_nhwc                = input->data_layout() == DataLayout::NHWC;
+    const bool                      needs_permute          = is_nhwc && (depth_multiplier > 1);
+    const bool                      needs_weights_reshape  = is_nhwc && (depth_multiplier == 1) && is_quantized;
+    const bool                      is_stride_1            = ((conv_info.stride().first == conv_info.stride().second) && (conv_info.stride().first == 1));
+    const bool                      is_stride_1_dilation_1 = (is_stride_1 && dilation.x() == 1 && dilation.y() == 1);
+    const bool                      is_dot8_supported      = dot8_supported(CLKernelLibrary::get().get_device());
+    DepthwiseConvolutionReshapeInfo info;
+    info.c0        = 4;
+    info.transpose = is_stride_1_dilation_1 && is_dot8_supported;
+
+    if(is_quantized)
+    {
+        const UniformQuantizationInfo iq_info = input->quantization_info().uniform();
+        const UniformQuantizationInfo wq_info = weights->quantization_info().uniform();
+        const UniformQuantizationInfo oq_info = (output->total_size() == 0) ? iq_info : output->quantization_info().uniform();
+
+        const float multiplier = iq_info.scale * wq_info.scale / oq_info.scale;
+        ARM_COMPUTE_UNUSED(multiplier);
+        ARM_COMPUTE_RETURN_ERROR_ON(multiplier > 1.0f);
+    }
+
+    if(needs_permute)
+    {
+        TensorShape permuted_input_shape   = input->tensor_shape();
+        TensorShape permuted_weights_shape = weights->tensor_shape();
+        TensorShape permuted_output_shape  = shape_calculator::compute_depthwise_convolution_shape(*input, *weights, conv_info, depth_multiplier, dilation);
+
+        permute(permuted_input_shape, PermutationVector(1U, 2U, 0U));
+        permute(permuted_weights_shape, PermutationVector(1U, 2U, 0U));
+        permute(permuted_output_shape, PermutationVector(1U, 2U, 0U));
+
+        const TensorInfo permuted_input   = input->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(permuted_input_shape).set_data_layout(DataLayout::NCHW);
+        const TensorInfo permuted_weights = weights->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(permuted_weights_shape).set_data_layout(DataLayout::NCHW);
+        const TensorInfo permuted_output  = output->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(permuted_output_shape).set_data_layout(DataLayout::NCHW);
+
+        ARM_COMPUTE_RETURN_ON_ERROR(CLDepthwiseConvolutionLayer3x3NCHWKernel::validate(&permuted_input, &permuted_weights, biases, &permuted_output, conv_info, depth_multiplier, act_info, gpu_target,
+                                                                                       dilation));
+    }
+    else if(is_nhwc)
+    {
+        if(needs_weights_reshape)
+        {
+            auto reshaped_weights_shape = arm_compute::misc::shape_calculator::compute_reshaped_depthwise_weights_shape(*weights, info);
+            ARM_COMPUTE_RETURN_ON_ERROR(CLDepthwiseConvolutionLayer3x3NHWCKernel::validate(input, &weights->clone()->set_tensor_shape(reshaped_weights_shape), biases, output, conv_info, depth_multiplier,
+                                                                                           act_info, dilation));
+        }
+        else
+        {
+            ARM_COMPUTE_RETURN_ON_ERROR(CLDepthwiseConvolutionLayer3x3NHWCKernel::validate(input, weights, biases, output, conv_info, depth_multiplier, act_info, dilation));
+        }
+    }
+    else
+    {
+        ARM_COMPUTE_RETURN_ON_ERROR(CLDepthwiseConvolutionLayer3x3NCHWKernel::validate(input, weights, biases, output, conv_info, depth_multiplier, act_info, gpu_target, dilation));
+    }
+    return Status{};
+}
+} // namespace
+
 CLDepthwiseConvolutionLayer3x3::CLDepthwiseConvolutionLayer3x3(std::shared_ptr<IMemoryManager> memory_manager)
-    : _memory_group(std::move(memory_manager)), _kernel(nullptr), _border_handler(), _permute_input_to_nchw(), _permute_weights_to_nchw(), _permute_output_to_nhwc(), _reshape_weights(), _permuted_input(),
-      _permuted_weights(), _permuted_output(), _original_weights(nullptr), _needs_permute(false), _needs_weights_reshape(false), _is_prepared(false)
+    : _func(std::move(memory_manager))
 {
 }
 
 void CLDepthwiseConvolutionLayer3x3::configure(ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const PadStrideInfo &conv_info, unsigned int depth_multiplier,
                                                ActivationLayerInfo act_info, const Size2D &dilation)
 {
+    _func.configure(input, weights, biases, output, conv_info, depth_multiplier, act_info, dilation);
+}
+
+Status CLDepthwiseConvolutionLayer3x3::validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info,
+                                                unsigned int depth_multiplier, ActivationLayerInfo act_info, GPUTarget gpu_target, const Size2D &dilation)
+{
+    return validate_arguments_3x3(input, weights, biases, output, conv_info, depth_multiplier, act_info, gpu_target, dilation);
+}
+
+void CLDepthwiseConvolutionLayer3x3::run()
+{
+    _func.run();
+}
+
+void CLDepthwiseConvolutionLayer3x3::prepare()
+{
+    _func.prepare();
+}
+
+CLDepthwiseConvolutionLayer::CLDepthwiseConvolutionLayerGeneric::CLDepthwiseConvolutionLayerGeneric(std::shared_ptr<IMemoryManager> memory_manager)
+    : _memory_group(std::move(memory_manager)),
+      _dwc_native_kernel(),
+      _permute_input_to_nhwc(),
+      _permute_weights_to_nhwc(),
+      _permute_output_to_nchw(),
+      _permuted_input(),
+      _permuted_weights(),
+      _permuted_output(),
+      _original_weights(),
+      _needs_permute(false),
+      _is_prepared(false)
+{
+}
+
+void CLDepthwiseConvolutionLayer::CLDepthwiseConvolutionLayerGeneric::configure(ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const PadStrideInfo &conv_info,
+                                                                                unsigned int depth_multiplier, const ActivationLayerInfo &act_info, const Size2D &dilation)
+{
+    ARM_COMPUTE_ERROR_ON_NULLPTR(input, weights, output);
+    ARM_COMPUTE_ERROR_THROW_ON(CLDepthwiseConvolutionLayer::validate(input->info(),
+                                                                     weights->info(),
+                                                                     biases != nullptr ? biases->info() : nullptr,
+                                                                     output->info(),
+                                                                     conv_info,
+                                                                     depth_multiplier,
+                                                                     act_info,
+                                                                     dilation));
+
+    _is_prepared      = false;
+    _original_weights = weights;
+    _needs_permute    = input->info()->data_layout() == DataLayout::NCHW;
+
+    ICLTensor       *input_to_use   = input;
+    const ICLTensor *weights_to_use = weights;
+    ICLTensor       *output_to_use  = output;
+    if(_needs_permute)
+    {
+        _memory_group.manage(&_permuted_input);
+        _memory_group.manage(&_permuted_output);
+
+        // Configure the function to transform the input tensor from NCHW -> NHWC
+        _permute_input_to_nhwc.configure(input, &_permuted_input, PermutationVector(2U, 0U, 1U));
+        _permuted_input.info()->set_data_layout(DataLayout::NHWC);
+
+        // Configure the function to transform the weights tensor from IHW -> HWI
+        _permute_weights_to_nhwc.configure(weights, &_permuted_weights, PermutationVector(2U, 0U, 1U));
+        _permuted_weights.info()->set_data_layout(DataLayout::NHWC);
+
+        // Set output quantization info before dwc kernel configure
+        _permuted_output.info()->set_quantization_info(output->info()->quantization_info());
+
+        input_to_use   = &_permuted_input;
+        weights_to_use = &_permuted_weights;
+        output_to_use  = &_permuted_output;
+    }
+
+    DWCWeightsKernelInfo dwc_weights_info;
+    dwc_weights_info.n0 = (depth_multiplier == 1) ? 8 : 1;
+    DWCKernelInfo dwc_info;
+    dwc_info.activation_info = act_info;
+    _dwc_native_kernel.configure(input_to_use, weights_to_use, biases, output_to_use, dwc_weights_info, dwc_info, conv_info, depth_multiplier, dilation);
+
+    if(_needs_permute)
+    {
+        _permuted_input.allocator()->allocate();
+
+        // Configure the function to transform the convoluted output to NCHW format
+        _permuted_output.info()->set_data_layout(DataLayout::NCHW);
+        _permute_output_to_nchw.configure(&_permuted_output, output, PermutationVector(1U, 2U, 0U));
+        _permuted_output.allocator()->allocate();
+    }
+}
+
+Status CLDepthwiseConvolutionLayer::CLDepthwiseConvolutionLayerGeneric::validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output,
+                                                                                 const PadStrideInfo &conv_info,
+                                                                                 unsigned int depth_multiplier, const ActivationLayerInfo &act_info, const Size2D &dilation)
+{
+    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_LAYOUT(input, output);
+    const size_t idx_w = get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::WIDTH);
+    const size_t idx_h = get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::HEIGHT);
+
+    ARM_COMPUTE_RETURN_ERROR_ON(weights->dimension(idx_w) + (weights->dimension(idx_w) - 1) * (dilation.x() - 1) > input->dimension(idx_w) + conv_info.pad_left() + conv_info.pad_right());
+    ARM_COMPUTE_RETURN_ERROR_ON(weights->dimension(idx_h) + (weights->dimension(idx_h) - 1) * (dilation.y() - 1) > input->dimension(idx_h) + conv_info.pad_top() + conv_info.pad_bottom());
+
+    DWCWeightsKernelInfo dwc_weights_info;
+    dwc_weights_info.n0 = (depth_multiplier == 1) ? 8 : 1;
+    DWCKernelInfo dwc_info;
+    dwc_info.activation_info = act_info;
+
+    const bool needs_permute = input->data_layout() == DataLayout::NCHW;
+
+    if(needs_permute)
+    {
+        TensorShape permuted_input_shape   = input->tensor_shape();
+        TensorShape permuted_weights_shape = weights->tensor_shape();
+        TensorShape permuted_output_shape  = shape_calculator::compute_depthwise_convolution_shape(*input, *weights, conv_info, depth_multiplier, dilation);
+
+        permute(permuted_input_shape, PermutationVector(2U, 0U, 1U));
+        permute(permuted_weights_shape, PermutationVector(2U, 0U, 1U));
+        permute(permuted_output_shape, PermutationVector(2U, 0U, 1U));
+
+        const TensorInfo permuted_input   = input->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(permuted_input_shape).set_data_layout(DataLayout::NHWC);
+        const TensorInfo permuted_weights = weights->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(permuted_weights_shape).set_data_layout(DataLayout::NHWC);
+        const TensorInfo permuted_output  = output->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(permuted_output_shape).set_data_layout(DataLayout::NHWC);
+
+        ARM_COMPUTE_RETURN_ON_ERROR(CLPermute::validate(input, &permuted_input, PermutationVector(2U, 0U, 1U)));
+        ARM_COMPUTE_RETURN_ON_ERROR(CLPermute::validate(weights, &permuted_weights, PermutationVector(2U, 0U, 1U)));
+        ARM_COMPUTE_RETURN_ON_ERROR(CLDepthwiseConvolutionLayerNativeKernel::validate(&permuted_input, &permuted_weights, biases, &permuted_output, dwc_weights_info,
+                                                                                      dwc_info, conv_info, depth_multiplier, dilation));
+        ARM_COMPUTE_RETURN_ON_ERROR(CLPermute::validate(&permuted_output, output, PermutationVector(1U, 2U, 0U)));
+    }
+    else
+    {
+        ARM_COMPUTE_RETURN_ON_ERROR(CLDepthwiseConvolutionLayerNativeKernel::validate(input, weights, biases, output, dwc_weights_info, dwc_info, conv_info, depth_multiplier, dilation));
+    }
+    return Status{};
+}
+
+void CLDepthwiseConvolutionLayer::CLDepthwiseConvolutionLayerGeneric::run()
+{
+    prepare();
+
+    MemoryGroupResourceScope scope_mg(_memory_group);
+
+    if(_needs_permute)
+    {
+        _permute_input_to_nhwc.run();
+    }
+    CLScheduler::get().enqueue(_dwc_native_kernel);
+    if(_needs_permute)
+    {
+        _permute_output_to_nchw.run();
+    }
+}
+
+void CLDepthwiseConvolutionLayer::CLDepthwiseConvolutionLayerGeneric::prepare()
+{
+    if(!_is_prepared)
+    {
+        if(_needs_permute)
+        {
+            ARM_COMPUTE_ERROR_ON(!_original_weights->is_used());
+
+            _permuted_weights.allocator()->allocate();
+            _permute_weights_to_nhwc.run();
+            _original_weights->mark_as_unused();
+        }
+        _is_prepared = true;
+    }
+}
+
+CLDepthwiseConvolutionLayer::CLDepthwiseConvolutionLayerInternal3x3::CLDepthwiseConvolutionLayerInternal3x3(std::shared_ptr<IMemoryManager> memory_manager)
+    : _memory_group(std::move(memory_manager)), _kernel(nullptr), _border_handler(), _permute_input_to_nchw(), _permute_weights_to_nchw(), _permute_output_to_nhwc(), _reshape_weights(), _permuted_input(),
+      _permuted_weights(), _permuted_output(), _original_weights(nullptr), _needs_permute(false), _needs_weights_reshape(false), _is_prepared(false)
+{
+}
+
+void CLDepthwiseConvolutionLayer::CLDepthwiseConvolutionLayerInternal3x3::configure(ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output,
+                                                                                    const PadStrideInfo &conv_info, unsigned int depth_multiplier, ActivationLayerInfo act_info, const Size2D &dilation)
+{
     ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::F16, DataType::F32);
     ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights);
     // idx_w and idx_h only used for validation
@@ -136,73 +383,13 @@ void CLDepthwiseConvolutionLayer3x3::configure(ICLTensor *input, const ICLTensor
     _border_handler.configure(input_to_use, _kernel->border_size(), BorderMode::CONSTANT, zero_value);
 }
 
-Status CLDepthwiseConvolutionLayer3x3::validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info,
-                                                unsigned int depth_multiplier, ActivationLayerInfo act_info, GPUTarget gpu_target, const Size2D &dilation)
+Status CLDepthwiseConvolutionLayer::CLDepthwiseConvolutionLayerInternal3x3::validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output,
+                                                                                     const PadStrideInfo &conv_info, unsigned int depth_multiplier, ActivationLayerInfo act_info, GPUTarget gpu_target, const Size2D &dilation)
 {
-    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, weights, output);
-    ARM_COMPUTE_RETURN_ERROR_ON(input->data_layout() == DataLayout::UNKNOWN);
-
-    const bool                      is_quantized           = is_data_type_quantized_asymmetric(input->data_type());
-    const bool                      is_nhwc                = input->data_layout() == DataLayout::NHWC;
-    const bool                      needs_permute          = is_nhwc && (depth_multiplier > 1);
-    const bool                      needs_weights_reshape  = is_nhwc && (depth_multiplier == 1) && is_quantized;
-    const bool                      is_stride_1            = ((conv_info.stride().first == conv_info.stride().second) && (conv_info.stride().first == 1));
-    const bool                      is_stride_1_dilation_1 = (is_stride_1 && dilation.x() == 1 && dilation.y() == 1);
-    const bool                      is_dot8_supported      = dot8_supported(CLKernelLibrary::get().get_device());
-    DepthwiseConvolutionReshapeInfo info;
-    info.c0        = 4;
-    info.transpose = is_stride_1_dilation_1 && is_dot8_supported;
-
-    if(is_quantized)
-    {
-        const UniformQuantizationInfo iq_info = input->quantization_info().uniform();
-        const UniformQuantizationInfo wq_info = weights->quantization_info().uniform();
-        const UniformQuantizationInfo oq_info = (output->total_size() == 0) ? iq_info : output->quantization_info().uniform();
-
-        const float multiplier = iq_info.scale * wq_info.scale / oq_info.scale;
-        ARM_COMPUTE_UNUSED(multiplier);
-        ARM_COMPUTE_RETURN_ERROR_ON(multiplier > 1.0f);
-    }
-
-    if(needs_permute)
-    {
-        TensorShape permuted_input_shape   = input->tensor_shape();
-        TensorShape permuted_weights_shape = weights->tensor_shape();
-        TensorShape permuted_output_shape  = shape_calculator::compute_depthwise_convolution_shape(*input, *weights, conv_info, depth_multiplier, dilation);
-
-        permute(permuted_input_shape, PermutationVector(1U, 2U, 0U));
-        permute(permuted_weights_shape, PermutationVector(1U, 2U, 0U));
-        permute(permuted_output_shape, PermutationVector(1U, 2U, 0U));
-
-        const TensorInfo permuted_input   = input->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(permuted_input_shape).set_data_layout(DataLayout::NCHW);
-        const TensorInfo permuted_weights = weights->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(permuted_weights_shape).set_data_layout(DataLayout::NCHW);
-        const TensorInfo permuted_output  = output->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(permuted_output_shape).set_data_layout(DataLayout::NCHW);
-
-        ARM_COMPUTE_RETURN_ON_ERROR(CLDepthwiseConvolutionLayer3x3NCHWKernel::validate(&permuted_input, &permuted_weights, biases, &permuted_output, conv_info, depth_multiplier, act_info, gpu_target,
-                                                                                       dilation));
-    }
-    else if(is_nhwc)
-    {
-        if(needs_weights_reshape)
-        {
-            auto reshaped_weights_shape = arm_compute::misc::shape_calculator::compute_reshaped_depthwise_weights_shape(*weights, info);
-            ARM_COMPUTE_RETURN_ON_ERROR(CLDepthwiseConvolutionLayer3x3NHWCKernel::validate(input, &weights->clone()->set_tensor_shape(reshaped_weights_shape), biases, output, conv_info, depth_multiplier,
-                                                                                           act_info, dilation));
-        }
-        else
-        {
-            ARM_COMPUTE_RETURN_ON_ERROR(CLDepthwiseConvolutionLayer3x3NHWCKernel::validate(input, weights, biases, output, conv_info, depth_multiplier, act_info, dilation));
-        }
-    }
-    else
-    {
-        ARM_COMPUTE_RETURN_ON_ERROR(CLDepthwiseConvolutionLayer3x3NCHWKernel::validate(input, weights, biases, output, conv_info, depth_multiplier, act_info, gpu_target, dilation));
-    }
-
-    return Status{};
+    return validate_arguments_3x3(input, weights, biases, output, conv_info, depth_multiplier, act_info, gpu_target, dilation);
 }
 
-void CLDepthwiseConvolutionLayer3x3::run()
+void CLDepthwiseConvolutionLayer::CLDepthwiseConvolutionLayerInternal3x3::run()
 {
     prepare();
 
@@ -221,7 +408,7 @@ void CLDepthwiseConvolutionLayer3x3::run()
     }
 }
 
-void CLDepthwiseConvolutionLayer3x3::prepare()
+void CLDepthwiseConvolutionLayer::CLDepthwiseConvolutionLayerInternal3x3::prepare()
 {
     if(!_is_prepared)
     {
@@ -247,194 +434,91 @@ void CLDepthwiseConvolutionLayer3x3::prepare()
 }
 
 CLDepthwiseConvolutionLayer::CLDepthwiseConvolutionLayer(std::shared_ptr<IMemoryManager> memory_manager)
-    : _memory_group(std::move(memory_manager)),
-      _optimised_function(nullptr),
-      _dwc_native_kernel(),
-      _permute_input_to_nhwc(),
-      _permute_weights_to_nhwc(),
-      _permute_output_to_nchw(),
-      _permuted_input(),
-      _permuted_weights(),
-      _permuted_output(),
-      _original_weights(),
-      _needs_permute(false),
-      _is_prepared(false)
+    : _memory_manager(std::move(memory_manager)), _depth_conv_func(DepthwiseConvolutionFunction::GENERIC), _func_3x3(), _func_generic()
 {
 }
 
-void CLDepthwiseConvolutionLayer::configure(ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const PadStrideInfo &conv_info,
-                                            unsigned int depth_multiplier, const ActivationLayerInfo &act_info, const Size2D &dilation)
+void CLDepthwiseConvolutionLayer::configure(ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const PadStrideInfo &conv_info, unsigned int depth_multiplier,
+                                            ActivationLayerInfo act_info, const Size2D &dilation)
 {
-    ARM_COMPUTE_ERROR_ON_NULLPTR(input, weights, output);
-    ARM_COMPUTE_ERROR_THROW_ON(CLDepthwiseConvolutionLayer::validate(input->info(),
-                                                                     weights->info(),
-                                                                     biases != nullptr ? biases->info() : nullptr,
-                                                                     output->info(),
-                                                                     conv_info,
-                                                                     depth_multiplier,
-                                                                     act_info,
-                                                                     dilation));
-
-    const size_t idx_w = get_data_layout_dimension_index(input->info()->data_layout(), DataLayoutDimension::WIDTH);
-    const size_t idx_h = get_data_layout_dimension_index(input->info()->data_layout(), DataLayoutDimension::HEIGHT);
-
-    const GPUTarget gpu_target                   = CLScheduler::get().target();
-    const bool      can_run_optimised_3x3_kernel = (weights->info()->dimension(idx_w) == 3) && (weights->info()->dimension(idx_h) == 3) && (is_data_type_float(input->info()->data_type())
-                                                   || (get_arch_from_target(gpu_target) == GPUTarget::MIDGARD));
-
-    _needs_permute    = false;
-    _is_prepared      = false;
-    _original_weights = weights;
-
-    if(bool(can_run_optimised_3x3_kernel))
+    const GPUTarget gpu_target = CLScheduler::get().target();
+    _depth_conv_func           = get_depthwiseconvolution_function(input->info(), weights->info(), (biases != nullptr) ? biases->info() : nullptr, output->info(), conv_info, depth_multiplier, act_info,
+                                                                   dilation, gpu_target);
+    switch(_depth_conv_func)
     {
-        auto f = arm_compute::support::cpp14::make_unique<CLDepthwiseConvolutionLayer3x3>();
-        f->configure(input, weights, biases, output, conv_info, depth_multiplier, act_info, dilation);
-        _optimised_function = std::move(f);
-    }
-    else
-    {
-        _needs_permute = input->info()->data_layout() == DataLayout::NCHW;
-
-        ICLTensor       *input_to_use   = input;
-        const ICLTensor *weights_to_use = weights;
-        ICLTensor       *output_to_use  = output;
-        if(_needs_permute)
-        {
-            _memory_group.manage(&_permuted_input);
-            _memory_group.manage(&_permuted_output);
-
-            // Configure the function to transform the input tensor from NCHW -> NHWC
-            _permute_input_to_nhwc.configure(input, &_permuted_input, PermutationVector(2U, 0U, 1U));
-            _permuted_input.info()->set_data_layout(DataLayout::NHWC);
-
-            // Configure the function to transform the weights tensor from IHW -> HWI
-            _permute_weights_to_nhwc.configure(weights, &_permuted_weights, PermutationVector(2U, 0U, 1U));
-            _permuted_weights.info()->set_data_layout(DataLayout::NHWC);
-
-            // Set output quantization info before dwc kernel configure
-            _permuted_output.info()->set_quantization_info(output->info()->quantization_info());
-
-            input_to_use   = &_permuted_input;
-            weights_to_use = &_permuted_weights;
-            output_to_use  = &_permuted_output;
-        }
-
-        DWCWeightsKernelInfo dwc_weights_info;
-        dwc_weights_info.n0 = (depth_multiplier == 1) ? 8 : 1;
-        DWCKernelInfo dwc_info;
-        dwc_info.activation_info = act_info;
-        _dwc_native_kernel.configure(input_to_use, weights_to_use, biases, output_to_use, dwc_weights_info, dwc_info, conv_info, depth_multiplier, dilation);
-
-        if(_needs_permute)
+        case DepthwiseConvolutionFunction::OPTIMIZED:
+            _func_3x3.set_memory_group(_memory_manager);
+            _func_3x3.configure(input, weights, biases, output, conv_info, depth_multiplier, act_info, dilation);
+            break;
+        case DepthwiseConvolutionFunction::GENERIC:
         {
-            _permuted_input.allocator()->allocate();
-
-            // Configure the function to transform the convoluted output to NCHW format
-            _permuted_output.info()->set_data_layout(DataLayout::NCHW);
-            _permute_output_to_nchw.configure(&_permuted_output, output, PermutationVector(1U, 2U, 0U));
-            _permuted_output.allocator()->allocate();
+            _func_generic.set_memory_group(_memory_manager);
+            _func_generic.configure(input, weights, biases, output, conv_info, depth_multiplier, act_info, dilation);
         }
+        break;
+        default:
+            ARM_COMPUTE_ERROR("Unsupported DepthwiseConvolutionFunction");
     }
 }
 
 Status CLDepthwiseConvolutionLayer::validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info,
-                                             unsigned int depth_multiplier, const ActivationLayerInfo &act_info, const Size2D &dilation)
+                                             unsigned int depth_multiplier, ActivationLayerInfo act_info, const Size2D &dilation)
 {
-    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_LAYOUT(input, output);
-
-    const size_t idx_w = get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::WIDTH);
-    const size_t idx_h = get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::HEIGHT);
-
-    ARM_COMPUTE_RETURN_ERROR_ON(weights->dimension(idx_w) + (weights->dimension(idx_w) - 1) * (dilation.x() - 1) > input->dimension(idx_w) + conv_info.pad_left() + conv_info.pad_right());
-    ARM_COMPUTE_RETURN_ERROR_ON(weights->dimension(idx_h) + (weights->dimension(idx_h) - 1) * (dilation.y() - 1) > input->dimension(idx_h) + conv_info.pad_top() + conv_info.pad_bottom());
-
-    const GPUTarget gpu_target                   = CLScheduler::get().target();
-    const bool      can_run_optimised_3x3_kernel = (weights->dimension(idx_w) == 3) && (weights->dimension(idx_h) == 3) && (is_data_type_float(input->data_type())
-                                                                                                                            || (get_arch_from_target(gpu_target) == GPUTarget::MIDGARD));
-
-    if(!can_run_optimised_3x3_kernel)
+    const GPUTarget              gpu_target      = CLScheduler::get().target();
+    DepthwiseConvolutionFunction depth_conv_func = get_depthwiseconvolution_function(input, weights, biases, output, conv_info, depth_multiplier, act_info, dilation, gpu_target);
+    switch(depth_conv_func)
     {
-        DWCWeightsKernelInfo dwc_weights_info;
-        dwc_weights_info.n0 = (depth_multiplier == 1) ? 8 : 1;
-        DWCKernelInfo dwc_info;
-        dwc_info.activation_info = act_info;
-
-        const bool needs_permute = input->data_layout() == DataLayout::NCHW;
+        case DepthwiseConvolutionFunction::OPTIMIZED:
+            return CLDepthwiseConvolutionLayerInternal3x3::validate(input, weights, biases, output, conv_info, depth_multiplier, act_info, gpu_target, dilation);
+        case DepthwiseConvolutionFunction::GENERIC:
+            return CLDepthwiseConvolutionLayerGeneric::validate(input, weights, biases, output, conv_info, depth_multiplier, act_info, dilation);
+        default:
+            ARM_COMPUTE_ERROR("Unsupported DepthwiseConvolutionFunction");
+    }
+}
 
-        if(needs_permute)
-        {
-            TensorShape permuted_input_shape   = input->tensor_shape();
-            TensorShape permuted_weights_shape = weights->tensor_shape();
-            TensorShape permuted_output_shape  = shape_calculator::compute_depthwise_convolution_shape(*input, *weights, conv_info, depth_multiplier, dilation);
-
-            permute(permuted_input_shape, PermutationVector(2U, 0U, 1U));
-            permute(permuted_weights_shape, PermutationVector(2U, 0U, 1U));
-            permute(permuted_output_shape, PermutationVector(2U, 0U, 1U));
-
-            const TensorInfo permuted_input   = input->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(permuted_input_shape).set_data_layout(DataLayout::NHWC);
-            const TensorInfo permuted_weights = weights->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(permuted_weights_shape).set_data_layout(DataLayout::NHWC);
-            const TensorInfo permuted_output  = output->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(permuted_output_shape).set_data_layout(DataLayout::NHWC);
-
-            ARM_COMPUTE_RETURN_ON_ERROR(CLPermute::validate(input, &permuted_input, PermutationVector(2U, 0U, 1U)));
-            ARM_COMPUTE_RETURN_ON_ERROR(CLPermute::validate(weights, &permuted_weights, PermutationVector(2U, 0U, 1U)));
-            ARM_COMPUTE_RETURN_ON_ERROR(CLDepthwiseConvolutionLayerNativeKernel::validate(&permuted_input, &permuted_weights, biases, &permuted_output, dwc_weights_info,
-                                                                                          dwc_info, conv_info, depth_multiplier, dilation));
-            ARM_COMPUTE_RETURN_ON_ERROR(CLPermute::validate(&permuted_output, output, PermutationVector(1U, 2U, 0U)));
-        }
-        else
-        {
-            ARM_COMPUTE_RETURN_ON_ERROR(CLDepthwiseConvolutionLayerNativeKernel::validate(input, weights, biases, output, dwc_weights_info, dwc_info, conv_info, depth_multiplier, dilation));
-        }
+DepthwiseConvolutionFunction CLDepthwiseConvolutionLayer::get_depthwiseconvolution_function(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output,
+                                                                                            const PadStrideInfo &conv_info,
+                                                                                            unsigned int depth_multiplier, ActivationLayerInfo act_info, const Size2D &dilation, GPUTarget gpu_target)
+{
+    if(bool(CLDepthwiseConvolutionLayerInternal3x3::validate(input, weights, biases, output, conv_info, depth_multiplier, act_info, gpu_target, dilation)) && (is_data_type_float(input->data_type())
+            || get_arch_from_target(gpu_target) == GPUTarget::MIDGARD))
+    {
+        return DepthwiseConvolutionFunction::OPTIMIZED;
     }
     else
     {
-        ARM_COMPUTE_RETURN_ON_ERROR(CLDepthwiseConvolutionLayer3x3::validate(input, weights, biases, output, conv_info, depth_multiplier, act_info, GPUTarget::MIDGARD, dilation));
+        return DepthwiseConvolutionFunction::GENERIC;
     }
-    return Status{};
 }
 
 void CLDepthwiseConvolutionLayer::run()
 {
-    prepare();
-
-    MemoryGroupResourceScope scope_mg(_memory_group);
-
-    if(_optimised_function != nullptr)
+    switch(_depth_conv_func)
     {
-        _optimised_function->run();
-    }
-    else
-    {
-        if(_needs_permute)
-        {
-            _permute_input_to_nhwc.run();
-        }
-        CLScheduler::get().enqueue(_dwc_native_kernel);
-        if(_needs_permute)
-        {
-            _permute_output_to_nchw.run();
-        }
+        case DepthwiseConvolutionFunction::OPTIMIZED:
+            _func_3x3.run();
+            break;
+        case DepthwiseConvolutionFunction::GENERIC:
+            _func_generic.run();
+            break;
+        default:
+            ARM_COMPUTE_ERROR("DepthwiseConvolutionFunction not properly configured");
     }
 }
 
 void CLDepthwiseConvolutionLayer::prepare()
 {
-    if(_optimised_function != nullptr)
+    switch(_depth_conv_func)
     {
-        _optimised_function->prepare();
-    }
-    else if(!_is_prepared)
-    {
-        if(_needs_permute)
-        {
-            ARM_COMPUTE_ERROR_ON(!_original_weights->is_used());
-
-            _permuted_weights.allocator()->allocate();
-            _permute_weights_to_nhwc.run();
-            _original_weights->mark_as_unused();
-        }
-        _is_prepared = true;
+        case DepthwiseConvolutionFunction::OPTIMIZED:
+            _func_3x3.prepare();
+            break;
+        case DepthwiseConvolutionFunction::GENERIC:
+            _func_generic.prepare();
+            break;
+        default:
+            ARM_COMPUTE_ERROR("DepthwiseConvolutionFunction not properly configured");
     }
 }
 } // namespace arm_compute
diff --git a/src/runtime/NEON/functions/NEDepthwiseConvolutionLayer.cpp b/src/runtime/NEON/functions/NEDepthwiseConvolutionLayer.cpp
index 76ae1fba3a..6cf7b97e66 100644
--- a/src/runtime/NEON/functions/NEDepthwiseConvolutionLayer.cpp
+++ b/src/runtime/NEON/functions/NEDepthwiseConvolutionLayer.cpp
@@ -33,203 +33,10 @@ using namespace arm_compute::misc::shape_calculator;
 
 namespace arm_compute
 {
-NEDepthwiseConvolutionLayer3x3::NEDepthwiseConvolutionLayer3x3(std::shared_ptr<IMemoryManager> memory_manager)
-    : _memory_group(memory_manager), _dwc_kernel(), _dwc_optimized_func(memory_manager), _output_stage_kernel(), _border_handler(), _permute_input(), _permute_weights(), _permute_output(),
-      _activationlayer_function(), _accumulator(), _permuted_input(), _permuted_weights(), _permuted_output(), _original_weights(nullptr), _has_bias(false), _is_quantized(false), _is_optimized(false),
-      _is_nchw(true), _permute(false), _is_activationlayer_enabled(false), _is_prepared(false)
-{
-}
-
-void NEDepthwiseConvolutionLayer3x3::configure_generic(ITensor                   *input,
-                                                       const ITensor             *weights,
-                                                       const ITensor             *biases,
-                                                       ITensor                   *output,
-                                                       const PadStrideInfo       &conv_info,
-                                                       unsigned int               depth_multiplier,
-                                                       const ActivationLayerInfo &act_info,
-                                                       const Size2D              &dilation)
+namespace
 {
-    ARM_COMPUTE_UNUSED(act_info);
-
-    PixelValue zero_value(0.f);
-
-    // Initialize the intermediate accumulator tensor in case of quantized input
-    if(_is_quantized)
-    {
-        TensorShape accum_shape  = output->info()->tensor_shape();
-        DataLayout  accum_layout = output->info()->data_layout();
-        if(!_is_nchw)
-        {
-            permute(accum_shape, PermutationVector(1U, 2U, 0U));
-            accum_layout = DataLayout::NCHW;
-        }
-
-        _memory_group.manage(&_accumulator);
-        _accumulator.allocator()->init(TensorInfo(accum_shape, 1, DataType::S32, output->info()->quantization_info()));
-        _accumulator.info()->set_data_layout(accum_layout);
-        zero_value = PixelValue(static_cast<uint32_t>(input->info()->quantization_info().uniform().offset));
-    }
-
-    if(!_is_nchw)
-    {
-        _memory_group.manage(&_permuted_input);
-        _memory_group.manage(&_permuted_output);
-
-        // Configure the function to transform the input tensor from NHWC -> NCHW
-        _permute_input.configure(input, &_permuted_input, PermutationVector(1U, 2U, 0U));
-        _permuted_input.info()->set_data_layout(DataLayout::NCHW);
-
-        // Configure the function to transform the weights tensor from HWI -> IHW
-        _permute_weights.configure(weights, &_permuted_weights, PermutationVector(1U, 2U, 0U));
-        _permuted_weights.info()->set_data_layout(DataLayout::NCHW);
-        _permuted_output.info()->set_quantization_info(output->info()->quantization_info());
-
-        // Configure depthwise
-        _dwc_kernel.configure(&_permuted_input, &_permuted_weights, (_is_quantized) ? &_accumulator : &_permuted_output, conv_info, depth_multiplier, dilation);
-
-        // Configure border handler
-        _border_handler.configure(&_permuted_input, _dwc_kernel.border_size(), BorderMode::CONSTANT, zero_value);
-
-        // Allocate tensors
-        _permuted_input.allocator()->allocate();
-    }
-    else
-    {
-        // Configure depthwise convolution kernel
-        _dwc_kernel.configure(input, weights, (_is_quantized) ? &_accumulator : output, conv_info, depth_multiplier, dilation);
-
-        // Configure border handler
-        _border_handler.configure(input, _dwc_kernel.border_size(), BorderMode::CONSTANT, zero_value);
-    }
-
-    // Configure biases accumulation
-    if(_is_quantized)
-    {
-        const UniformQuantizationInfo iq_info = input->info()->quantization_info().uniform();
-        const UniformQuantizationInfo wq_info = weights->info()->quantization_info().uniform();
-        const UniformQuantizationInfo oq_info = (output->info()->total_size() == 0) ? iq_info : output->info()->quantization_info().uniform();
-
-        float multiplier = (iq_info.scale * wq_info.scale) / oq_info.scale;
-        int   output_multiplier;
-        int   output_shift;
-        quantization::calculate_quantized_multiplier_less_than_one(multiplier, &output_multiplier, &output_shift);
-        _output_stage_kernel.configure(&_accumulator, biases, _is_nchw ? output : &_permuted_output, output_multiplier, output_shift, oq_info.offset);
-        _accumulator.allocator()->allocate();
-    }
-    else if(_has_bias)
-    {
-        _output_stage_kernel.configure(_is_nchw ? output : &_permuted_output, biases);
-    }
-
-    // Permute output
-    if(!_is_nchw)
-    {
-        // Configure the function to transform the convoluted output to NHWC
-        _permute_output.configure(&_permuted_output, output, PermutationVector(2U, 0U, 1U));
-        _permuted_output.allocator()->allocate();
-    }
-}
-
-void NEDepthwiseConvolutionLayer3x3::configure_optimized(const ITensor             *input,
-                                                         const ITensor             *weights,
-                                                         const ITensor             *biases,
-                                                         ITensor                   *output,
-                                                         const PadStrideInfo       &conv_info,
-                                                         unsigned int               depth_multiplier,
-                                                         const ActivationLayerInfo &act_info)
-{
-    ActivationLayerInfo act_info_to_use = ActivationLayerInfo();
-    const bool          is_relu         = arm_compute::utils::info_helpers::is_relu(act_info);
-    const bool          is_relu6        = arm_compute::utils::info_helpers::is_relu6(act_info);
-    _is_activationlayer_enabled         = act_info.enabled() && !(is_relu || is_relu6);
-    if(!_is_activationlayer_enabled)
-    {
-        act_info_to_use = act_info;
-    }
-
-    if(_is_nchw)
-    {
-        _memory_group.manage(&_permuted_input);
-        _memory_group.manage(&_permuted_output);
-
-        // Configure the function to transform the input tensor from NCHW -> NHWC
-        _permute_input.configure(input, &_permuted_input, PermutationVector(2U, 0U, 1U));
-        _permuted_input.info()->set_data_layout(DataLayout::NHWC);
-
-        // Configure the function to transform the weights tensor from IHW -> HWI
-        _permute_weights.configure(weights, &_permuted_weights, PermutationVector(2U, 0U, 1U));
-        _permuted_weights.info()->set_data_layout(DataLayout::NHWC);
-
-        _permuted_output.info()->set_data_layout(DataLayout::NHWC);
-        _permuted_output.info()->set_quantization_info(output->info()->quantization_info());
-
-        // Configure optimized depthwise
-        _dwc_optimized_func.configure(&_permuted_input, &_permuted_weights, biases, &_permuted_output, conv_info, depth_multiplier, act_info_to_use);
-
-        // Configure the function to transform the convoluted output to ACL's native ordering format NCHW
-        _permuted_output.info()->set_data_layout(DataLayout::NHWC);
-        _permute_output.configure(&_permuted_output, output, PermutationVector(1U, 2U, 0U));
-
-        // Allocate tensors
-        _permuted_input.allocator()->allocate();
-        _permuted_output.allocator()->allocate();
-    }
-    else
-    {
-        _dwc_optimized_func.configure(input, weights, biases, output, conv_info, depth_multiplier, act_info_to_use);
-    }
-}
-
-void NEDepthwiseConvolutionLayer3x3::configure(ITensor       *input,
-                                               const ITensor *weights,
-                                               const ITensor *biases,
-                                               ITensor *output, const PadStrideInfo &conv_info,
-                                               unsigned int               depth_multiplier,
-                                               const ActivationLayerInfo &act_info,
-                                               const Size2D              &dilation)
-{
-    ARM_COMPUTE_ERROR_ON_NULLPTR(input, weights, output);
-    // Perform validation step
-    ARM_COMPUTE_ERROR_THROW_ON(NEDepthwiseConvolutionLayer3x3::validate(input->info(), weights->info(), (biases == nullptr) ? nullptr : biases->info(),
-                                                                        output->info(), conv_info, depth_multiplier, act_info, dilation));
-
-    _original_weights = weights;
-    _is_quantized     = is_data_type_quantized_asymmetric(input->info()->data_type());
-    _has_bias         = biases != nullptr;
-    _is_optimized     = NEDepthwiseConvolutionAssemblyDispatch::is_optimized_supported(input->info(),
-                                                                                       weights->info(),
-                                                                                       conv_info,
-                                                                                       depth_multiplier, dilation);
-    _is_nchw                    = input->info()->data_layout() == DataLayout::NCHW;
-    _permute                    = _is_optimized == _is_nchw;
-    _is_prepared                = false;
-    _is_activationlayer_enabled = act_info.enabled();
-
-    // Configure appropriate pipeline
-    if(_is_optimized)
-    {
-        configure_optimized(input, weights, biases, output, conv_info, depth_multiplier, act_info);
-    }
-    else
-    {
-        configure_generic(input, weights, biases, output, conv_info, depth_multiplier, act_info, dilation);
-    }
-
-    // Configure activation
-    if(_is_activationlayer_enabled)
-    {
-        _activationlayer_function.configure(output, nullptr, act_info);
-    }
-}
-
-Status NEDepthwiseConvolutionLayer3x3::validate(const ITensorInfo         *input,
-                                                const ITensorInfo         *weights,
-                                                const ITensorInfo         *biases,
-                                                const ITensorInfo         *output,
-                                                const PadStrideInfo       &conv_info,
-                                                unsigned int               depth_multiplier,
-                                                const ActivationLayerInfo &act_info,
-                                                const Size2D              &dilation)
+Status validate_arguments_optimized(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info,
+                                    unsigned int depth_multiplier, const ActivationLayerInfo &act_info, const Size2D &dilation)
 {
     ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, weights, output);
     ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::F16, DataType::F32);
@@ -248,28 +55,32 @@ Status NEDepthwiseConvolutionLayer3x3::validate(const ITensorInfo         *input
         ARM_COMPUTE_RETURN_ERROR_ON(biases->dimension(0) != weights->dimension(channel_idx));
     }
 
+    const bool is_quantized = is_data_type_quantized_asymmetric(input->data_type());
+
+    if(is_quantized)
+    {
+        const UniformQuantizationInfo iq_info = input->quantization_info().uniform();
+        const UniformQuantizationInfo wq_info = weights->quantization_info().uniform();
+        const UniformQuantizationInfo oq_info = output->quantization_info().uniform();
+
+        float multiplier = (iq_info.scale * wq_info.scale) / oq_info.scale;
+        ARM_COMPUTE_UNUSED(multiplier);
+        ARM_COMPUTE_RETURN_ERROR_ON(multiplier > 1.0f);
+    }
+
     if(!NEDepthwiseConvolutionAssemblyDispatch::is_optimized_supported(input, weights, conv_info, depth_multiplier, dilation))
     {
-        const bool is_quantized = is_data_type_quantized_asymmetric(input->data_type());
-        TensorInfo accumulator  = TensorInfo(output->clone()->set_is_resizable(true).reset_padding().set_data_type(DataType::S32));
-        ARM_COMPUTE_RETURN_ON_ERROR(NEDepthwiseConvolutionLayer3x3Kernel::validate(input, weights, is_quantized ? &accumulator : output, conv_info, depth_multiplier));
+        TensorInfo accumulator = TensorInfo(output->clone()->set_is_resizable(true).reset_padding().set_data_type(DataType::S32));
+        ARM_COMPUTE_RETURN_ON_ERROR(NEDepthwiseConvolutionLayer3x3Kernel::validate(input, weights, is_quantized ? &accumulator : output, conv_info, depth_multiplier, dilation));
 
         if(is_quantized)
         {
-            const UniformQuantizationInfo iq_info = input->quantization_info().uniform();
-            const UniformQuantizationInfo wq_info = weights->quantization_info().uniform();
-            const UniformQuantizationInfo oq_info = output->quantization_info().uniform();
-
-            float multiplier = (iq_info.scale * wq_info.scale) / oq_info.scale;
-            int   output_multiplier;
-            int   output_shift;
-            ARM_COMPUTE_RETURN_ON_ERROR(quantization::calculate_quantized_multiplier_less_than_one(multiplier, &output_multiplier, &output_shift));
-            ARM_COMPUTE_RETURN_ON_ERROR(NEDirectConvolutionLayerOutputStageKernel::validate(&accumulator, biases, output, output_multiplier, output_shift, oq_info.offset));
+            ARM_COMPUTE_RETURN_ON_ERROR(NEDirectConvolutionLayerOutputStageKernel::validate(&accumulator, biases, output));
         }
     }
     else
     {
-        ARM_COMPUTE_RETURN_ON_ERROR(NEDepthwiseConvolutionAssemblyDispatch::validate(input, weights, biases, output, conv_info, depth_multiplier));
+        ARM_COMPUTE_RETURN_ON_ERROR(NEDepthwiseConvolutionAssemblyDispatch::validate(input, weights, biases, output, conv_info, depth_multiplier, act_info, dilation));
     }
 
     //Validate Activation Layer
@@ -280,102 +91,55 @@ Status NEDepthwiseConvolutionLayer3x3::validate(const ITensorInfo         *input
 
     return Status{};
 }
+} // namespace
 
-void NEDepthwiseConvolutionLayer3x3::run_generic()
+NEDepthwiseConvolutionLayerOptimized::NEDepthwiseConvolutionLayerOptimized(std::shared_ptr<IMemoryManager> memory_manager)
+    : _func(std::move(memory_manager))
 {
-    // Fill border
-    NEScheduler::get().schedule(&_border_handler, Window::DimX);
-
-    // Execute depthwise convolution
-    NEScheduler::get().schedule(&_dwc_kernel, Window::DimX);
-
-    // Add biases
-    if(_has_bias || _is_quantized)
-    {
-        NEScheduler::get().schedule(&_output_stage_kernel, Window::DimX);
-    }
-
-    // Permute output
-    if(!_is_nchw)
-    {
-        _permute_output.run();
-    }
 }
 
-void NEDepthwiseConvolutionLayer3x3::run_optimized()
+void NEDepthwiseConvolutionLayerOptimized::configure(ITensor       *input,
+                                                     const ITensor *weights,
+                                                     const ITensor *biases,
+                                                     ITensor *output, const PadStrideInfo &conv_info,
+                                                     unsigned int               depth_multiplier,
+                                                     const ActivationLayerInfo &act_info,
+                                                     const Size2D              &dilation)
 {
-    // Run assembly function
-    _dwc_optimized_func.run();
-
-    // Permute output
-    if(_is_nchw)
-    {
-        _permute_output.run();
-    }
+    _func.configure(input, weights, biases, output, conv_info, depth_multiplier, act_info, dilation);
 }
 
-void NEDepthwiseConvolutionLayer3x3::run()
+Status NEDepthwiseConvolutionLayerOptimized::validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info,
+                                                      unsigned int depth_multiplier, const ActivationLayerInfo &act_info, const Size2D &dilation)
 {
-    prepare();
-
-    MemoryGroupResourceScope scope_mg(_memory_group);
-
-    // Permute input
-    if(_permute)
-    {
-        _permute_input.run();
-    }
-
-    _is_optimized ? run_optimized() : run_generic();
-
-    // Run activation
-    if(_is_activationlayer_enabled)
-    {
-        _activationlayer_function.run();
-    }
+    return validate_arguments_optimized(input, weights, biases, output, conv_info, depth_multiplier, act_info, dilation);
 }
 
-void NEDepthwiseConvolutionLayer3x3::prepare()
+void NEDepthwiseConvolutionLayerOptimized::run()
 {
-    if(!_is_prepared)
-    {
-        // Permute weights
-        if(_permute)
-        {
-            _permuted_weights.allocator()->allocate();
-            _permute_weights.run();
-            _original_weights->mark_as_unused();
-        }
-
-        // Prepare optimized function
-        if(_is_optimized)
-        {
-            _dwc_optimized_func.prepare();
-            if(!_permuted_weights.is_used())
-            {
-                _permuted_weights.allocator()->free();
-            }
-        }
+    _func.run();
+}
 
-        _is_prepared = true;
-    }
+void NEDepthwiseConvolutionLayerOptimized::prepare()
+{
+    _func.prepare();
 }
 
-NEDepthwiseConvolutionLayerOptimized::NEDepthwiseConvolutionLayerOptimized(std::shared_ptr<IMemoryManager> memory_manager)
+NEDepthwiseConvolutionLayer::NEDepthwiseConvolutionLayerOptimizedInternal::NEDepthwiseConvolutionLayerOptimizedInternal(std::shared_ptr<IMemoryManager> memory_manager)
     : _memory_group(memory_manager), _dwc_kernel(), _dwc_optimized_func(memory_manager), _output_stage_kernel(), _border_handler(), _permute_input(), _permute_weights(), _permute_output(),
       _activationlayer_function(), _accumulator(), _permuted_input(), _permuted_weights(), _permuted_output(), _original_weights(nullptr), _has_bias(false), _is_quantized(false), _is_optimized(false),
       _is_nchw(true), _permute(false), _is_activationlayer_enabled(false), _is_prepared(false)
 {
 }
 
-void NEDepthwiseConvolutionLayerOptimized::configure_generic(ITensor                   *input,
-                                                             const ITensor             *weights,
-                                                             const ITensor             *biases,
-                                                             ITensor                   *output,
-                                                             const PadStrideInfo       &conv_info,
-                                                             unsigned int               depth_multiplier,
-                                                             const ActivationLayerInfo &act_info,
-                                                             const Size2D              &dilation)
+void NEDepthwiseConvolutionLayer::NEDepthwiseConvolutionLayerOptimizedInternal::configure_generic(ITensor                   *input,
+                                                                                                  const ITensor             *weights,
+                                                                                                  const ITensor             *biases,
+                                                                                                  ITensor                   *output,
+                                                                                                  const PadStrideInfo       &conv_info,
+                                                                                                  unsigned int               depth_multiplier,
+                                                                                                  const ActivationLayerInfo &act_info,
+                                                                                                  const Size2D              &dilation)
 {
     ARM_COMPUTE_UNUSED(act_info);
 
@@ -458,14 +222,14 @@ void NEDepthwiseConvolutionLayerOptimized::configure_generic(ITensor
     }
 }
 
-void NEDepthwiseConvolutionLayerOptimized::configure_optimized(const ITensor             *input,
-                                                               const ITensor             *weights,
-                                                               const ITensor             *biases,
-                                                               ITensor                   *output,
-                                                               const PadStrideInfo       &conv_info,
-                                                               unsigned int               depth_multiplier,
-                                                               const ActivationLayerInfo &act_info,
-                                                               const Size2D              &dilation)
+void NEDepthwiseConvolutionLayer::NEDepthwiseConvolutionLayerOptimizedInternal::configure_optimized(const ITensor             *input,
+                                                                                                    const ITensor             *weights,
+                                                                                                    const ITensor             *biases,
+                                                                                                    ITensor                   *output,
+                                                                                                    const PadStrideInfo       &conv_info,
+                                                                                                    unsigned int               depth_multiplier,
+                                                                                                    const ActivationLayerInfo &act_info,
+                                                                                                    const Size2D              &dilation)
 {
     ActivationLayerInfo act_info_to_use = ActivationLayerInfo();
     const bool          is_relu         = arm_compute::utils::info_helpers::is_relu(act_info);
@@ -509,18 +273,18 @@ void NEDepthwiseConvolutionLayerOptimized::configure_optimized(const ITensor
     }
 }
 
-void NEDepthwiseConvolutionLayerOptimized::configure(ITensor       *input,
-                                                     const ITensor *weights,
-                                                     const ITensor *biases,
-                                                     ITensor *output, const PadStrideInfo &conv_info,
-                                                     unsigned int               depth_multiplier,
-                                                     const ActivationLayerInfo &act_info,
-                                                     const Size2D              &dilation)
+void NEDepthwiseConvolutionLayer::NEDepthwiseConvolutionLayerOptimizedInternal::configure(ITensor       *input,
+                                                                                          const ITensor *weights,
+                                                                                          const ITensor *biases,
+                                                                                          ITensor *output, const PadStrideInfo &conv_info,
+                                                                                          unsigned int               depth_multiplier,
+                                                                                          const ActivationLayerInfo &act_info,
+                                                                                          const Size2D              &dilation)
 {
     ARM_COMPUTE_ERROR_ON_NULLPTR(input, weights, output);
     // Perform validation step
-    ARM_COMPUTE_ERROR_THROW_ON(NEDepthwiseConvolutionLayerOptimized::validate(input->info(), weights->info(), (biases == nullptr) ? nullptr : biases->info(),
-                                                                              output->info(), conv_info, depth_multiplier, act_info, dilation));
+    ARM_COMPUTE_ERROR_THROW_ON(NEDepthwiseConvolutionLayerOptimizedInternal::validate(input->info(), weights->info(), (biases == nullptr) ? nullptr : biases->info(),
+                                                                                      output->info(), conv_info, depth_multiplier, act_info, dilation));
 
     _original_weights = weights;
     _is_quantized     = is_data_type_quantized_asymmetric(input->info()->data_type());
@@ -552,70 +316,19 @@ void NEDepthwiseConvolutionLayerOptimized::configure(ITensor       *input,
     }
 }
 
-Status NEDepthwiseConvolutionLayerOptimized::validate(const ITensorInfo         *input,
-                                                      const ITensorInfo         *weights,
-                                                      const ITensorInfo         *biases,
-                                                      const ITensorInfo         *output,
-                                                      const PadStrideInfo       &conv_info,
-                                                      unsigned int               depth_multiplier,
-                                                      const ActivationLayerInfo &act_info,
-                                                      const Size2D              &dilation)
+Status NEDepthwiseConvolutionLayer::NEDepthwiseConvolutionLayerOptimizedInternal::validate(const ITensorInfo         *input,
+                                                                                           const ITensorInfo         *weights,
+                                                                                           const ITensorInfo         *biases,
+                                                                                           const ITensorInfo         *output,
+                                                                                           const PadStrideInfo       &conv_info,
+                                                                                           unsigned int               depth_multiplier,
+                                                                                           const ActivationLayerInfo &act_info,
+                                                                                           const Size2D              &dilation)
 {
-    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, weights, output);
-    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::F16, DataType::F32);
-    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights);
-    ARM_COMPUTE_RETURN_ERROR_ON(input->data_layout() == DataLayout::UNKNOWN);
-    ARM_COMPUTE_RETURN_ERROR_ON(dilation.x() < 1 || dilation.y() < 1);
-    const size_t idx_w = get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::WIDTH);
-    const size_t idx_h = get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::HEIGHT);
-    ARM_COMPUTE_RETURN_ERROR_ON(weights->dimension(idx_w) + (weights->dimension(idx_w) - 1) * (dilation.x() - 1) > input->dimension(idx_w) + conv_info.pad_left() + conv_info.pad_right());
-    ARM_COMPUTE_RETURN_ERROR_ON(weights->dimension(idx_h) + (weights->dimension(idx_h) - 1) * (dilation.y() - 1) > input->dimension(idx_h) + conv_info.pad_top() + conv_info.pad_bottom());
-
-    if(biases != nullptr)
-    {
-        const unsigned int channel_idx = get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::CHANNEL);
-        ARM_COMPUTE_RETURN_ERROR_ON(biases->num_dimensions() > 1);
-        ARM_COMPUTE_RETURN_ERROR_ON(biases->dimension(0) != weights->dimension(channel_idx));
-    }
-
-    const bool is_quantized = is_data_type_quantized_asymmetric(input->data_type());
-
-    if(is_quantized)
-    {
-        const UniformQuantizationInfo iq_info = input->quantization_info().uniform();
-        const UniformQuantizationInfo wq_info = weights->quantization_info().uniform();
-        const UniformQuantizationInfo oq_info = output->quantization_info().uniform();
-
-        float multiplier = (iq_info.scale * wq_info.scale) / oq_info.scale;
-        ARM_COMPUTE_UNUSED(multiplier);
-        ARM_COMPUTE_RETURN_ERROR_ON(multiplier > 1.0f);
-    }
-
-    if(!NEDepthwiseConvolutionAssemblyDispatch::is_optimized_supported(input, weights, conv_info, depth_multiplier, dilation))
-    {
-        TensorInfo accumulator = TensorInfo(output->clone()->set_is_resizable(true).reset_padding().set_data_type(DataType::S32));
-        ARM_COMPUTE_RETURN_ON_ERROR(NEDepthwiseConvolutionLayer3x3Kernel::validate(input, weights, is_quantized ? &accumulator : output, conv_info, depth_multiplier, dilation));
-
-        if(is_quantized)
-        {
-            ARM_COMPUTE_RETURN_ON_ERROR(NEDirectConvolutionLayerOutputStageKernel::validate(&accumulator, biases, output));
-        }
-    }
-    else
-    {
-        ARM_COMPUTE_RETURN_ON_ERROR(NEDepthwiseConvolutionAssemblyDispatch::validate(input, weights, biases, output, conv_info, depth_multiplier, act_info, dilation));
-    }
-
-    //Validate Activation Layer
-    if(act_info.enabled())
-    {
-        ARM_COMPUTE_RETURN_ON_ERROR(NEActivationLayer::validate(output, nullptr, act_info));
-    }
-
-    return Status{};
+    return validate_arguments_optimized(input, weights, biases, output, conv_info, depth_multiplier, act_info, dilation);
 }
 
-void NEDepthwiseConvolutionLayerOptimized::run_generic()
+void NEDepthwiseConvolutionLayer::NEDepthwiseConvolutionLayerOptimizedInternal::run_generic()
 {
     // Fill border
     NEScheduler::get().schedule(&_border_handler, Window::DimX);
@@ -636,7 +349,7 @@ void NEDepthwiseConvolutionLayerOptimized::run_generic()
     }
 }
 
-void NEDepthwiseConvolutionLayerOptimized::run_optimized()
+void NEDepthwiseConvolutionLayer::NEDepthwiseConvolutionLayerOptimizedInternal::run_optimized()
 {
     // Run assembly function
     _dwc_optimized_func.run();
@@ -648,7 +361,7 @@ void NEDepthwiseConvolutionLayerOptimized::run_optimized()
     }
 }
 
-void NEDepthwiseConvolutionLayerOptimized::run()
+void NEDepthwiseConvolutionLayer::NEDepthwiseConvolutionLayerOptimizedInternal::run()
 {
     prepare();
 
@@ -669,7 +382,7 @@ void NEDepthwiseConvolutionLayerOptimized::run()
     }
 }
 
-void NEDepthwiseConvolutionLayerOptimized::prepare()
+void NEDepthwiseConvolutionLayer::NEDepthwiseConvolutionLayerOptimizedInternal::prepare()
 {
     if(!_is_prepared)
     {
@@ -695,14 +408,14 @@ void NEDepthwiseConvolutionLayerOptimized::prepare()
     }
 }
 
-NEDepthwiseConvolutionLayer::NEDepthwiseConvolutionLayer()
+NEDepthwiseConvolutionLayer::NEDepthwiseConvolutionLayerGeneric::NEDepthwiseConvolutionLayerGeneric()
     : _depthwise_conv_kernel(), _fill_border(), _permute_input(), _permute_weights(), _permute_output(), _activationlayer_function(), _permuted_input(), _permuted_weights(), _permuted_output(),
       _is_prepared(false), _is_nchw(false), _is_activationlayer_enabled(false), _original_weights(nullptr)
 {
 }
 
-void NEDepthwiseConvolutionLayer::configure(ITensor *input, const ITensor *weights, const ITensor *biases, ITensor *output, const PadStrideInfo &conv_info,
-                                            unsigned int depth_multiplier, const ActivationLayerInfo &act_info, const Size2D &dilation)
+void NEDepthwiseConvolutionLayer::NEDepthwiseConvolutionLayerGeneric::configure(ITensor *input, const ITensor *weights, const ITensor *biases, ITensor *output, const PadStrideInfo &conv_info,
+                                                                                unsigned int depth_multiplier, const ActivationLayerInfo &act_info, const Size2D &dilation)
 {
     ARM_COMPUTE_ERROR_ON_NULLPTR(input, weights, output);
     ARM_COMPUTE_ERROR_THROW_ON(NEDepthwiseConvolutionLayer::validate(input->info(), weights->info(), (biases == nullptr) ? nullptr : biases->info(),
@@ -750,8 +463,9 @@ void NEDepthwiseConvolutionLayer::configure(ITensor *input, const ITensor *weigh
     }
 }
 
-Status NEDepthwiseConvolutionLayer::validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info,
-                                             unsigned int depth_multiplier, const ActivationLayerInfo &act_info, const Size2D &dilation)
+Status NEDepthwiseConvolutionLayer::NEDepthwiseConvolutionLayerGeneric::validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output,
+                                                                                 const PadStrideInfo &conv_info,
+                                                                                 unsigned int depth_multiplier, const ActivationLayerInfo &act_info, const Size2D &dilation)
 {
     ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, weights, output);
     if(input->data_layout() == DataLayout::NCHW)
@@ -787,7 +501,7 @@ Status NEDepthwiseConvolutionLayer::validate(const ITensorInfo *input, const ITe
     return Status{};
 }
 
-void NEDepthwiseConvolutionLayer::run()
+void NEDepthwiseConvolutionLayer::NEDepthwiseConvolutionLayerGeneric::run()
 {
     if(_is_nchw)
     {
@@ -809,7 +523,7 @@ void NEDepthwiseConvolutionLayer::run()
     }
 }
 
-void NEDepthwiseConvolutionLayer::prepare()
+void NEDepthwiseConvolutionLayer::NEDepthwiseConvolutionLayerGeneric::prepare()
 {
     if(!_is_prepared)
     {
@@ -820,4 +534,87 @@ void NEDepthwiseConvolutionLayer::prepare()
         _is_prepared = true;
     }
 }
+
+NEDepthwiseConvolutionLayer::NEDepthwiseConvolutionLayer(std::shared_ptr<IMemoryManager> memory_manager)
+    : _depth_conv_func(DepthwiseConvolutionFunction::GENERIC), _func_optimized(std::move(memory_manager)), _func_generic()
+{
+}
+
+void NEDepthwiseConvolutionLayer::configure(ITensor *input, const ITensor *weights, const ITensor *biases, ITensor *output, const PadStrideInfo &conv_info, unsigned int depth_multiplier,
+                                            const ActivationLayerInfo &act_info, const Size2D &dilation)
+{
+    _depth_conv_func = get_depthwiseconvolution_function(input->info(), weights->info(), (biases != nullptr) ? biases->info() : nullptr, output->info(), conv_info, depth_multiplier, act_info, dilation);
+    switch(_depth_conv_func)
+    {
+        case DepthwiseConvolutionFunction::OPTIMIZED:
+            _func_optimized.configure(input, weights, biases, output, conv_info, depth_multiplier, act_info, dilation);
+            break;
+        case DepthwiseConvolutionFunction::GENERIC:
+            _func_generic.configure(input, weights, biases, output, conv_info, depth_multiplier, act_info, dilation);
+            break;
+        default:
+            ARM_COMPUTE_ERROR("Unsupported DepthwiseConvolutionFunction");
+    }
+}
+
+Status NEDepthwiseConvolutionLayer::validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info,
+                                             unsigned int depth_multiplier, const ActivationLayerInfo &act_info, const Size2D &dilation)
+{
+    DepthwiseConvolutionFunction depth_conv_func = get_depthwiseconvolution_function(input, weights, biases, output, conv_info, depth_multiplier, act_info, dilation);
+    switch(depth_conv_func)
+    {
+        case DepthwiseConvolutionFunction::OPTIMIZED:
+            return NEDepthwiseConvolutionLayerOptimized::validate(input, weights, biases, output, conv_info, depth_multiplier, act_info, dilation);
+            break;
+        case DepthwiseConvolutionFunction::GENERIC:
+            return NEDepthwiseConvolutionLayerGeneric::validate(input, weights, biases, output, conv_info, depth_multiplier, act_info, dilation);
+            break;
+        default:
+            ARM_COMPUTE_ERROR("Unsupported DepthwiseConvolutionFunction");
+    }
+}
+
+DepthwiseConvolutionFunction NEDepthwiseConvolutionLayer::get_depthwiseconvolution_function(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output,
+                                                                                            const PadStrideInfo &conv_info,
+                                                                                            unsigned int depth_multiplier, ActivationLayerInfo act_info, const Size2D &dilation)
+{
+    if(bool(NEDepthwiseConvolutionLayerOptimized::validate(input, weights, biases, output, conv_info, depth_multiplier, act_info, dilation)))
+    {
+        return DepthwiseConvolutionFunction::OPTIMIZED;
+    }
+    else
+    {
+        return DepthwiseConvolutionFunction::GENERIC;
+    }
+}
+
+void NEDepthwiseConvolutionLayer::run()
+{
+    switch(_depth_conv_func)
+    {
+        case DepthwiseConvolutionFunction::OPTIMIZED:
+            _func_optimized.run();
+            break;
+        case DepthwiseConvolutionFunction::GENERIC:
+            _func_generic.run();
+            break;
+        default:
+            ARM_COMPUTE_ERROR("DepthwiseConvolutionFunction not properly configured");
+    }
+}
+
+void NEDepthwiseConvolutionLayer::prepare()
+{
+    switch(_depth_conv_func)
+    {
+        case DepthwiseConvolutionFunction::OPTIMIZED:
+            _func_optimized.prepare();
+            break;
+        case DepthwiseConvolutionFunction::GENERIC:
+            _func_generic.prepare();
+            break;
+        default:
+            ARM_COMPUTE_ERROR("DepthwiseConvolutionFunction not properly configured");
+    }
+}
 } // namespace arm_compute