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>

6 files changed, 461 insertions, 295 deletions

diff --git a/arm_compute/core/Types.h b/arm_compute/core/Types.h
index 9641089e7b..d7b47ac512 100644
--- a/arm_compute/core/Types.h
+++ b/arm_compute/core/Types.h
@@ -139,6 +139,13 @@ enum class ConvolutionMethod
     FFT       /**< Convolution using FFT */
 };
 
+/** Available DepthwiseConvolutionFunction*/
+enum class DepthwiseConvolutionFunction
+{
+    OPTIMIZED, /**< Optimized Depthwise Convolution */
+    GENERIC,   /**< Generic Depthwise Convolution */
+};
+
 /** Available DeconvolutionMethod*/
 enum class DeconvolutionMethod
 {
diff --git a/arm_compute/graph/TypePrinter.h b/arm_compute/graph/TypePrinter.h
index e4188125b9..131fd39277 100644
--- a/arm_compute/graph/TypePrinter.h
+++ b/arm_compute/graph/TypePrinter.h
@@ -251,9 +251,6 @@ inline ::std::ostream &operator<<(::std::ostream &os, const DepthwiseConvolution
         case DepthwiseConvolutionMethod::Default:
             os << "DEFAULT";
             break;
-        case DepthwiseConvolutionMethod::GEMV:
-            os << "GEMV";
-            break;
         case DepthwiseConvolutionMethod::Optimized3x3:
             os << "Optimized3x3";
             break;
diff --git a/arm_compute/graph/backends/FunctionHelpers.h b/arm_compute/graph/backends/FunctionHelpers.h
index 94b385e81e..ee257e3abf 100644
--- a/arm_compute/graph/backends/FunctionHelpers.h
+++ b/arm_compute/graph/backends/FunctionHelpers.h
@@ -538,7 +538,7 @@ std::unique_ptr<IFunction> create_deconvolution_layer(DeconvolutionLayerNode &no
  *
  * @return Backend depth-wise convolution layer function
  */
-template <typename DepthwiseConvolutionLayerFunctions, typename TargetInfo>
+template <typename DepthwiseConvolutionLayer, typename TargetInfo>
 std::unique_ptr<IFunction> create_depthwise_convolution_layer(DepthwiseConvolutionLayerNode &node)
 {
     validate_node<TargetInfo>(node, 3 /* expected inputs */, 1 /* expected outputs */);
@@ -556,26 +556,17 @@ std::unique_ptr<IFunction> create_depthwise_convolution_layer(DepthwiseConvoluti
         biases->info()->set_data_type(DataType::S32);
     }
 
-    const PadStrideInfo              conv_info        = node.convolution_info();
-    const DepthwiseConvolutionMethod dwc_algorithm    = node.depthwise_convolution_method();
-    const unsigned int               depth_multiplier = node.depth_multiplier();
-    const ActivationLayerInfo        fused_act        = node.fused_activation();
+    const PadStrideInfo       conv_info        = node.convolution_info();
+    const unsigned int        depth_multiplier = node.depth_multiplier();
+    const ActivationLayerInfo fused_act        = node.fused_activation();
 
     // Create and configure function (we assume that functions have been validated before creation)
     std::unique_ptr<IFunction> func;
     std::string                func_name;
-    if(dwc_algorithm == DepthwiseConvolutionMethod::Optimized3x3)
-    {
-        std::tie(func, func_name) = create_named_function<typename DepthwiseConvolutionLayerFunctions::OptimizedDepthwiseConvolutionLayer>(
-                                        std::string("DepthwiseConvolutionLayer3x3"),
-                                        input, weights, biases, output, conv_info, depth_multiplier, fused_act);
-    }
-    else
-    {
-        std::tie(func, func_name) = create_named_function<typename DepthwiseConvolutionLayerFunctions::GenericDepthwiseConvolutionLayer>(
-                                        std::string("DepthwiseConvolutionLayer"),
-                                        input, weights, biases, output, conv_info, depth_multiplier, fused_act);
-    }
+
+    std::tie(func, func_name) = create_named_function<DepthwiseConvolutionLayer>(
+                                    std::string("DepthwiseConvolutionLayer"),
+                                    input, weights, biases, output, conv_info, depth_multiplier, fused_act);
 
     // Log info
     std::ostringstream qss;
diff --git a/arm_compute/graph/backends/ValidateHelpers.h b/arm_compute/graph/backends/ValidateHelpers.h
index 13de273bdf..9170006d9c 100644
--- a/arm_compute/graph/backends/ValidateHelpers.h
+++ b/arm_compute/graph/backends/ValidateHelpers.h
@@ -163,13 +163,12 @@ Status validate_convolution_layer(ConvolutionLayerNode &node)
 /** Validates a Depthwise Convolution layer node
  *
  * @tparam DepthwiseConvolutionLayer    Default Depthwise Convolution layer type
- * @tparam DepthwiseConvolutionLayer3x3 Optimized 3x3 Depthwise Convolution layer type
  *
  * @param[in] node Node to validate
  *
  * @return Status
  */
-template <typename DepthwiseConvolutionLayer, typename DepthwiseConvolutionLayer3x3>
+template <typename DepthwiseConvolutionLayer>
 Status validate_depthwise_convolution_layer(DepthwiseConvolutionLayerNode &node)
 {
     ARM_COMPUTE_LOG_GRAPH_VERBOSE("Validating DepthwiseConvolutionLayer node with ID : " << node.id() << " and Name: " << node.name() << std::endl);
@@ -191,11 +190,8 @@ Status validate_depthwise_convolution_layer(DepthwiseConvolutionLayerNode &node)
     switch(dwc_algorithm)
     {
         case DepthwiseConvolutionMethod::Default:
-        case DepthwiseConvolutionMethod::GEMV:
-            status = DepthwiseConvolutionLayer::validate(input, weights, biases, output, conv_info, depth_multiplier);
-            break;
         case DepthwiseConvolutionMethod::Optimized3x3:
-            status = DepthwiseConvolutionLayer3x3::validate(input, weights, biases, output, conv_info, depth_multiplier);
+            status = DepthwiseConvolutionLayer::validate(input, weights, biases, output, conv_info, depth_multiplier);
             break;
         default:
             ARM_COMPUTE_RETURN_ERROR_MSG("Unsupported depthwise convolution method");
diff --git a/arm_compute/runtime/CL/functions/CLDepthwiseConvolutionLayer.h b/arm_compute/runtime/CL/functions/CLDepthwiseConvolutionLayer.h
index 98581a21fe..b8b11f08b2 100644
--- a/arm_compute/runtime/CL/functions/CLDepthwiseConvolutionLayer.h
+++ b/arm_compute/runtime/CL/functions/CLDepthwiseConvolutionLayer.h
@@ -40,151 +40,301 @@ namespace arm_compute
 {
 class ICLTensor;
 
-/** Basic function to execute a depthwise convolution for kernel size 3x3xC (when data layout NCHW) or Cx3x3 (when data layout NHWC). This function calls the following OpenCL kernels:
- *
- * -# @ref CLDepthwiseConvolutionLayer3x3NCHWKernel (if data_layout == NCHW)
- * -# @ref CLDepthwiseConvolutionLayer3x3NHWCKernel (if data_layout == NHWC)
- * -# @ref CLDepthwiseConvolutionLayerReshapeWeightsKernel (if data_layout == NHWC)
- * -# @ref CLFillBorderKernel (if pad_x or pad_y > 0)
- *
+/** Function to execute a depthwise convolution
  */
-class CLDepthwiseConvolutionLayer3x3 : public IFunction
+class CLDepthwiseConvolutionLayer : public IFunction
 {
 public:
     /** Default constructor */
-    CLDepthwiseConvolutionLayer3x3(std::shared_ptr<IMemoryManager> memory_manager = nullptr);
+    CLDepthwiseConvolutionLayer(std::shared_ptr<IMemoryManager> memory_manager = nullptr);
     /** Prevent instances of this class from being copied (As this class contains pointers) */
-    CLDepthwiseConvolutionLayer3x3(const CLDepthwiseConvolutionLayer3x3 &) = delete;
+    CLDepthwiseConvolutionLayer(const CLDepthwiseConvolutionLayer &) = delete;
     /** Default move constructor */
-    CLDepthwiseConvolutionLayer3x3(CLDepthwiseConvolutionLayer3x3 &&) = default;
+    CLDepthwiseConvolutionLayer(CLDepthwiseConvolutionLayer &&) = default;
     /** Prevent instances of this class from being copied (As this class contains pointers) */
-    CLDepthwiseConvolutionLayer3x3 &operator=(const CLDepthwiseConvolutionLayer3x3 &) = delete;
+    CLDepthwiseConvolutionLayer &operator=(const CLDepthwiseConvolutionLayer &) = delete;
     /** Default move assignment operator */
-    CLDepthwiseConvolutionLayer3x3 &operator=(CLDepthwiseConvolutionLayer3x3 &&) = default;
-    /** Initialize the function's source, destination, conv and border_size.
+    CLDepthwiseConvolutionLayer &operator=(CLDepthwiseConvolutionLayer &&) = default;
+    /** Initialize the function's source, destination, weights and convolution information.
      *
-     * @param[in, out] input            Source tensor. Data type supported: QASYMM8/F16/F32. (Written to only for border filling).
-     * @param[in]      weights          Weights tensor. A 3D tensor with shape [3, 3, IFM]. Data type supported: Same as @p input.
+     * @param[in, out] input            Source tensor. Data type supported: QASYMM8/FP16/FP32. Data layout supported: NHWC, NCHW
+     * @param[in]      weights          Weights tensor. These are 3D tensors with shape [kernel_x, kernel_y, IFM]. Data type supported: Same as @p input.
      * @param[in]      biases           Biases tensor. A 1D tensor with shape [IFM]. Must be nullptr if not needed.
-     *                                  Data type supported: Same as @p input.
+     *                                  Data type supported: Same as @p input, S32 when input is QASYMM8.
      * @param[out]     output           Destination tensor. Data type supported: same as @p input.
      * @param[in]      conv_info        Padding and stride information to use for the convolution.
      * @param[in]      depth_multiplier (Optional) Multiplier to apply to the input's depth in order to retrieve the output's depth. Defaults to 1.
-     * @param[in]      act_info         (Optional) Activation layer information in case of a fused activation. Only RELU, BOUNDED_RELU and LU_BOUNDED_RELU for 3x3 QASYMM8 supported.
+     * @param[in]      act_info         (Optional) Activation layer information in case of a fused activation.
      * @param[in]      dilation         (Optional) Dilation, in elements, across x and y. Defaults to (1, 1).
      */
     void configure(ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const PadStrideInfo &conv_info, unsigned int depth_multiplier = 1,
                    ActivationLayerInfo act_info = ActivationLayerInfo(), const Size2D &dilation = Size2D(1U, 1U));
 
-    /** Static function to check if given info will lead to a valid configuration of @ref CLDepthwiseConvolutionLayer3x3
+    /** Static function to check if given info will lead to a valid configuration of @ref CLDepthwiseConvolutionLayer
      *
-     * @param[in] input            Source tensor info. Data type supported: QASYMM8 for all layouts, F16/F32 for NCHW.
-     * @param[in] weights          Weights tensor info. A 3D tensor with shape [3, 3, IFM]. Data type supported: Same as @p input.
+     * @param[in] input            Source tensor info. Data type supported: QASYMM8/FP16/FP32. Data layout supported: NHWC, NCHW
+     * @param[in] weights          Weights tensor info. These are 3D tensors with shape [kernel_x, kernel_y, IFM]. Data type supported: Same as @p input.
      * @param[in] biases           Biases tensor info. A 1D tensor with shape [IFM]. Must be nullptr if not needed.
      *                             Data type supported: Same as @p input, S32 when input is QASYMM8.
      * @param[in] output           Destination tensor. Data type supported: same as @p input.
      * @param[in] conv_info        Padding and stride information to use for the convolution.
      * @param[in] depth_multiplier (Optional) Multiplier to apply to the input's depth in order to retrieve the output's depth. Defaults to 1.
      * @param[in] act_info         (Optional) Activation layer information in case of a fused activation. Only RELU, BOUNDED_RELU and LU_BOUNDED_RELU for 3x3 QASYMM8 supported.
-     * @param[in] gpu_target       (Optional) GPU target to validate the kernel for. Defaults to midgard.
      * @param[in] dilation         (Optional) Dilation, in elements, across x and y. Defaults to (1, 1).
      *
      * @return a status
      */
     static Status validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info, unsigned int depth_multiplier = 1,
-                           ActivationLayerInfo act_info = ActivationLayerInfo(), GPUTarget gpu_target = GPUTarget::MIDGARD, const Size2D &dilation = Size2D(1U, 1U));
+                           ActivationLayerInfo act_info = ActivationLayerInfo(), const Size2D &dilation = Size2D(1U, 1U));
+
     // Inherited methods overriden:
     void run() override;
     void prepare() override;
 
 private:
-    MemoryGroup                                            _memory_group;
-    std::unique_ptr<ICLDepthwiseConvolutionLayer3x3Kernel> _kernel;
-    CLFillBorderKernel                                     _border_handler;
-    CLPermute                                              _permute_input_to_nchw;
-    CLPermute                                              _permute_weights_to_nchw;
-    CLPermute                                              _permute_output_to_nhwc;
-    CLDepthwiseConvolutionLayerReshapeWeightsKernel        _reshape_weights;
-    CLTensor                                               _permuted_input;
-    CLTensor                                               _permuted_weights;
-    CLTensor                                               _permuted_output;
-    const ITensor                                         *_original_weights;
-    bool                                                   _needs_permute;
-    bool                                                   _needs_weights_reshape;
-    bool                                                   _is_prepared;
+    /** Static function to choose the best depthwise convolution function for @ref CLDepthwiseConvolutionLayer
+     *
+     * @param[in] input            Source tensor info. Data type supported: QASYMM8/FP16/FP32. Data layout supported: NHWC, NCHW
+     * @param[in] weights          Weights tensor info. These are 3D tensors with shape [kernel_x, kernel_y, IFM]. Data type supported: Same as @p input.
+     * @param[in] biases           Biases tensor info. A 1D tensor with shape [IFM]. Must be nullptr if not needed.
+     *                             Data type supported: Same as @p input, S32 when input is QASYMM8.
+     * @param[in] output           Destination tensor. Data type supported: same as @p input.
+     * @param[in] conv_info        Padding and stride information to use for the convolution.
+     * @param[in] depth_multiplier (Optional) Multiplier to apply to the input's depth in order to retrieve the output's depth. Defaults to 1.
+     * @param[in] act_info         (Optional) Activation layer information in case of a fused activation. Only RELU, BOUNDED_RELU and LU_BOUNDED_RELU for 3x3 QASYMM8 supported.
+     * @param[in] dilation         (Optional) Dilation, in elements, across x and y. Defaults to (1, 1).
+     * @param[in] gpu_target       (Optional) GPU target to validate the kernel for. Defaults to midgard.
+     *
+     * @return a Depthwise Convolution Function
+     */
+    static DepthwiseConvolutionFunction get_depthwiseconvolution_function(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output,
+                                                                          const PadStrideInfo &conv_info, unsigned int depth_multiplier = 1,
+                                                                          ActivationLayerInfo act_info = ActivationLayerInfo(), const Size2D &dilation = Size2D(1U, 1U), GPUTarget gpu_target = GPUTarget::MIDGARD);
+
+    /** Basic function to execute a depthwise convolution for kernel size 3x3xC (when data layout NCHW) or Cx3x3 (when data layout NHWC). This function calls the following OpenCL kernels:
+    *
+    * -# @ref CLDepthwiseConvolutionLayer3x3NCHWKernel (if data_layout == NCHW)
+    * -# @ref CLDepthwiseConvolutionLayer3x3NHWCKernel (if data_layout == NHWC)
+    * -# @ref CLDepthwiseConvolutionLayerReshapeWeightsKernel (if data_layout == NHWC)
+    * -# @ref CLFillBorderKernel (if pad_x or pad_y > 0)
+    *
+    */
+    class CLDepthwiseConvolutionLayerInternal3x3 : public IFunction
+    {
+    public:
+        /** Default constructor */
+        CLDepthwiseConvolutionLayerInternal3x3(std::shared_ptr<IMemoryManager> memory_manager = nullptr);
+        /** Prevent instances of this class from being copied (As this class contains pointers) */
+        CLDepthwiseConvolutionLayerInternal3x3(const CLDepthwiseConvolutionLayerInternal3x3 &) = delete;
+        /** Default move constructor */
+        CLDepthwiseConvolutionLayerInternal3x3(CLDepthwiseConvolutionLayerInternal3x3 &&) = default;
+        /** Prevent instances of this class from being copied (As this class contains pointers) */
+        CLDepthwiseConvolutionLayerInternal3x3 &operator=(const CLDepthwiseConvolutionLayerInternal3x3 &) = delete;
+        /** Default move assignment operator */
+        CLDepthwiseConvolutionLayerInternal3x3 &operator=(CLDepthwiseConvolutionLayerInternal3x3 &&) = default;
+        /** Initialize the function's source, destination, conv and border_size.
+         *
+         * @param[in, out] input            Source tensor. Data type supported: QASYMM8/F16/F32. (Written to only for border filling).
+         * @param[in]      weights          Weights tensor. A 3D tensor with shape [3, 3, IFM]. Data type supported: Same as @p input.
+         * @param[in]      biases           Biases tensor. A 1D tensor with shape [IFM]. Must be nullptr if not needed.
+         *                                  Data type supported: Same as @p input.
+         * @param[out]     output           Destination tensor. Data type supported: same as @p input.
+         * @param[in]      conv_info        Padding and stride information to use for the convolution.
+         * @param[in]      depth_multiplier (Optional) Multiplier to apply to the input's depth in order to retrieve the output's depth. Defaults to 1.
+         * @param[in]      act_info         (Optional) Activation layer information in case of a fused activation. Only RELU, BOUNDED_RELU and LU_BOUNDED_RELU for 3x3 QASYMM8 supported.
+         * @param[in]      dilation         (Optional) Dilation, in elements, across x and y. Defaults to (1, 1).
+         */
+        void configure(ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const PadStrideInfo &conv_info, unsigned int depth_multiplier = 1,
+                       ActivationLayerInfo act_info = ActivationLayerInfo(), const Size2D &dilation = Size2D(1U, 1U));
+
+        /** Static function to check if given info will lead to a valid configuration of @ref CLDepthwiseConvolutionLayer3x3
+         *
+         * @param[in] input            Source tensor info. Data type supported: QASYMM8 for all layouts, F16/F32 for NCHW.
+         * @param[in] weights          Weights tensor info. A 3D tensor with shape [3, 3, IFM]. Data type supported: Same as @p input.
+         * @param[in] biases           Biases tensor info. A 1D tensor with shape [IFM]. Must be nullptr if not needed.
+         *                             Data type supported: Same as @p input, S32 when input is QASYMM8.
+         * @param[in] output           Destination tensor. Data type supported: same as @p input.
+         * @param[in] conv_info        Padding and stride information to use for the convolution.
+         * @param[in] depth_multiplier (Optional) Multiplier to apply to the input's depth in order to retrieve the output's depth. Defaults to 1.
+         * @param[in] act_info         (Optional) Activation layer information in case of a fused activation. Only RELU, BOUNDED_RELU and LU_BOUNDED_RELU for 3x3 QASYMM8 supported.
+         * @param[in] dilation         (Optional) Dilation, in elements, across x and y. Defaults to (1, 1).
+         *
+         * @return a status
+         */
+        static Status validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info, unsigned int depth_multiplier = 1,
+                               ActivationLayerInfo act_info = ActivationLayerInfo(), GPUTarget gpu_target = GPUTarget::MIDGARD, const Size2D &dilation = Size2D(1U, 1U));
+
+        // Inherited methods overriden:
+        void run() override;
+        void prepare() override;
+
+        void set_memory_group(std::shared_ptr<IMemoryManager> memory_manager)
+        {
+            _memory_group = MemoryGroup(std::move(memory_manager));
+        };
+
+    private:
+        MemoryGroup                                            _memory_group;
+        std::unique_ptr<ICLDepthwiseConvolutionLayer3x3Kernel> _kernel;
+        CLFillBorderKernel                                     _border_handler;
+        CLPermute                                              _permute_input_to_nchw;
+        CLPermute                                              _permute_weights_to_nchw;
+        CLPermute                                              _permute_output_to_nhwc;
+        CLDepthwiseConvolutionLayerReshapeWeightsKernel        _reshape_weights;
+        CLTensor                                               _permuted_input;
+        CLTensor                                               _permuted_weights;
+        CLTensor                                               _permuted_output;
+        const ITensor                                         *_original_weights;
+        bool                                                   _needs_permute;
+        bool                                                   _needs_weights_reshape;
+        bool                                                   _is_prepared;
+    };
+
+    /** Basic function to execute a generic depthwise convolution. This function calls the following OpenCL kernels:
+     *
+     * -# @ref CLDepthwiseConvolutionLayerNativeKernel
+     * -# @ref CLPermute (x 3) if the data layout is NCHW
+     *
+     */
+    class CLDepthwiseConvolutionLayerGeneric : public IFunction
+    {
+    public:
+        /** Default constructor */
+        CLDepthwiseConvolutionLayerGeneric(std::shared_ptr<IMemoryManager> memory_manager = nullptr);
+        /** Prevent instances of this class from being copied (As this class contains pointers) */
+        CLDepthwiseConvolutionLayerGeneric(const CLDepthwiseConvolutionLayerGeneric &) = delete;
+        /** Default move constructor */
+        CLDepthwiseConvolutionLayerGeneric(CLDepthwiseConvolutionLayerGeneric &&) = default;
+        /** Prevent instances of this class from being copied (As this class contains pointers) */
+        CLDepthwiseConvolutionLayerGeneric &operator=(const CLDepthwiseConvolutionLayerGeneric &) = delete;
+        /** Default move assignment operator */
+        CLDepthwiseConvolutionLayerGeneric &operator=(CLDepthwiseConvolutionLayerGeneric &&) = default;
+        /** Initialize the function's source, destination, weights and convolution information.
+         *
+         * @param[in, out] input            Source tensor. Data type supported: QASYMM8/F32. (Written to only for border filling).
+         * @param[in]      weights          Weights tensor. These are 3D tensors with shape [kernel_x, kernel_y, IFM]. Data type supported: Same as @p input.
+         * @param[in]      biases           Biases tensor. A 1D tensor with shape [IFM]. Must be nullptr if not needed.
+         *                                  Data type supported: Same as @p input, S32 when input is QASYMM8.
+         * @param[out]     output           Destination tensor. Data type supported: same as @p input.
+         * @param[in]      conv_info        Padding and stride information to use for the convolution.
+         * @param[in]      depth_multiplier (Optional) Multiplier to apply to the input's depth in order to retrieve the output's depth. Defaults to 1.
+         * @param[in]      act_info         (Optional) Activation layer information in case of a fused activation.
+         * @param[in]      dilation         (Optional) Dilation, in elements, across x and y. Defaults to (1, 1).
+         */
+        void configure(ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const PadStrideInfo &conv_info,
+                       unsigned int depth_multiplier = 1, const ActivationLayerInfo &act_info = ActivationLayerInfo(), const Size2D &dilation = Size2D(1U, 1U));
+
+        /** Static function to check if given info will lead to a valid configuration of @ref CLDepthwiseConvolutionLayerGeneric
+         *
+         * @param[in] input            Source tensor info. Data type supported: QASYMM8/F32.
+         * @param[in] weights          Weights tensor info. These are 3D tensors with shape [kernel_x, kernel_y, IFM]. Data type supported: Same as @p input.
+         * @param[in] biases           Biases tensor info. A 1D tensor with shape [IFM]. Must be nullptr if not needed.
+         *                             Data type supported: Same as @p input, S32 when input is QASYMM8.
+         * @param[in] output           Destination tensor. Data type supported: same as @p input.
+         * @param[in] conv_info        Padding and stride information to use for the convolution.
+         * @param[in] depth_multiplier (Optional) Multiplier to apply to the input's depth in order to retrieve the output's depth. Defaults to 1.
+         * @param[in] act_info         (Optional) Activation layer information in case of a fused activation.
+         * @param[in] dilation         (Optional) Dilation, in elements, across x and y. Defaults to (1, 1).
+         *
+         * @return a status
+         */
+        static Status validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info,
+                               unsigned int depth_multiplier = 1, const ActivationLayerInfo &act_info = ActivationLayerInfo(), const Size2D &dilation = Size2D(1U, 1U));
+
+        // Inherited methods overriden:
+        void run() override;
+        void prepare() override;
+
+        void set_memory_group(std::shared_ptr<IMemoryManager> memory_manager)
+        {
+            _memory_group = MemoryGroup(std::move(memory_manager));
+        };
+
+    private:
+        MemoryGroup _memory_group;
+
+        CLDepthwiseConvolutionLayerNativeKernel _dwc_native_kernel;
+        CLPermute                               _permute_input_to_nhwc;
+        CLPermute                               _permute_weights_to_nhwc;
+        CLPermute                               _permute_output_to_nchw;
+
+        CLTensor       _permuted_input;
+        CLTensor       _permuted_weights;
+        CLTensor       _permuted_output;
+        const ITensor *_original_weights;
+
+        bool _needs_permute;
+        bool _is_prepared;
+    };
+
+    std::shared_ptr<IMemoryManager> _memory_manager;
+
+    DepthwiseConvolutionFunction           _depth_conv_func;
+    CLDepthwiseConvolutionLayerInternal3x3 _func_3x3;
+    CLDepthwiseConvolutionLayerGeneric     _func_generic;
 };
 
-/** Basic function to execute a generic depthwise convolution. This function calls the following OpenCL kernels:
+/** Basic function to execute a depthwise convolution for kernel size 3x3xC (when data layout NCHW) or Cx3x3 (when data layout NHWC). This function calls the following OpenCL kernels:
  *
- * -# @ref CLDepthwiseConvolutionLayerNativeKernel
- * -# @ref CLPermute (x 3) if the data layout is NCHW
+ * -# @ref CLDepthwiseConvolutionLayer3x3NCHWKernel (if data_layout == NCHW)
+ * -# @ref CLDepthwiseConvolutionLayer3x3NHWCKernel (if data_layout == NHWC)
+ * -# @ref CLDepthwiseConvolutionLayerReshapeWeightsKernel (if data_layout == NHWC)
+ * -# @ref CLFillBorderKernel (if pad_x or pad_y > 0)
  *
  */
-class CLDepthwiseConvolutionLayer : public IFunction
+class CLDepthwiseConvolutionLayer3x3 : public IFunction
 {
 public:
     /** Default constructor */
-    CLDepthwiseConvolutionLayer(std::shared_ptr<IMemoryManager> memory_manager = nullptr);
+    CLDepthwiseConvolutionLayer3x3(std::shared_ptr<IMemoryManager> memory_manager = nullptr);
     /** Prevent instances of this class from being copied (As this class contains pointers) */
-    CLDepthwiseConvolutionLayer(const CLDepthwiseConvolutionLayer &) = delete;
+    CLDepthwiseConvolutionLayer3x3(const CLDepthwiseConvolutionLayer3x3 &) = delete;
     /** Default move constructor */
-    CLDepthwiseConvolutionLayer(CLDepthwiseConvolutionLayer &&) = default;
+    CLDepthwiseConvolutionLayer3x3(CLDepthwiseConvolutionLayer3x3 &&) = default;
     /** Prevent instances of this class from being copied (As this class contains pointers) */
-    CLDepthwiseConvolutionLayer &operator=(const CLDepthwiseConvolutionLayer &) = delete;
+    CLDepthwiseConvolutionLayer3x3 &operator=(const CLDepthwiseConvolutionLayer3x3 &) = delete;
     /** Default move assignment operator */
-    CLDepthwiseConvolutionLayer &operator=(CLDepthwiseConvolutionLayer &&) = default;
-    /** Initialize the function's source, destination, weights and convolution information.
+    CLDepthwiseConvolutionLayer3x3 &operator=(CLDepthwiseConvolutionLayer3x3 &&) = default;
+    /** Initialize the function's source, destination, conv and border_size.
      *
-     * @param[in, out] input            Source tensor. Data type supported: QASYMM8/F32. (Written to only for border filling).
-     * @param[in]      weights          Weights tensor. These are 3D tensors with shape [kernel_x, kernel_y, IFM]. Data type supported: Same as @p input.
+     * @param[in, out] input            Source tensor. Data type supported: QASYMM8/F16/F32. (Written to only for border filling).
+     * @param[in]      weights          Weights tensor. A 3D tensor with shape [3, 3, IFM]. Data type supported: Same as @p input.
      * @param[in]      biases           Biases tensor. A 1D tensor with shape [IFM]. Must be nullptr if not needed.
-     *                                  Data type supported: Same as @p input, S32 when input is QASYMM8.
+     *                                  Data type supported: Same as @p input.
      * @param[out]     output           Destination tensor. Data type supported: same as @p input.
      * @param[in]      conv_info        Padding and stride information to use for the convolution.
      * @param[in]      depth_multiplier (Optional) Multiplier to apply to the input's depth in order to retrieve the output's depth. Defaults to 1.
-     * @param[in]      act_info         (Optional) Activation layer information in case of a fused activation.
+     * @param[in]      act_info         (Optional) Activation layer information in case of a fused activation. Only RELU, BOUNDED_RELU and LU_BOUNDED_RELU for 3x3 QASYMM8 supported.
      * @param[in]      dilation         (Optional) Dilation, in elements, across x and y. Defaults to (1, 1).
      */
-    void configure(ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const PadStrideInfo &conv_info,
-                   unsigned int depth_multiplier = 1, const ActivationLayerInfo &act_info = ActivationLayerInfo(), const Size2D &dilation = Size2D(1U, 1U));
+    ARM_COMPUTE_DEPRECATED_REL_REPLACE(20.02, CLDepthwiseConvolutionLayer)
+    void configure(ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const PadStrideInfo &conv_info, unsigned int depth_multiplier = 1,
+                   ActivationLayerInfo act_info = ActivationLayerInfo(), const Size2D &dilation = Size2D(1U, 1U));
 
-    /** Static function to check if given info will lead to a valid configuration of @ref CLDepthwiseConvolutionLayer
+    /** Static function to check if given info will lead to a valid configuration of @ref CLDepthwiseConvolutionLayer3x3
      *
-     * @param[in] input            Source tensor info. Data type supported: QASYMM8/F32.
-     * @param[in] weights          Weights tensor info. These are 3D tensors with shape [kernel_x, kernel_y, IFM]. Data type supported: Same as @p input.
+     * @param[in] input            Source tensor info. Data type supported: QASYMM8 for all layouts, F16/F32 for NCHW.
+     * @param[in] weights          Weights tensor info. A 3D tensor with shape [3, 3, IFM]. Data type supported: Same as @p input.
      * @param[in] biases           Biases tensor info. A 1D tensor with shape [IFM]. Must be nullptr if not needed.
      *                             Data type supported: Same as @p input, S32 when input is QASYMM8.
      * @param[in] output           Destination tensor. Data type supported: same as @p input.
      * @param[in] conv_info        Padding and stride information to use for the convolution.
      * @param[in] depth_multiplier (Optional) Multiplier to apply to the input's depth in order to retrieve the output's depth. Defaults to 1.
-     * @param[in] act_info         (Optional) Activation layer information in case of a fused activation.
+     * @param[in] act_info         (Optional) Activation layer information in case of a fused activation. Only RELU, BOUNDED_RELU and LU_BOUNDED_RELU for 3x3 QASYMM8 supported.
+     * @param[in] gpu_target       (Optional) GPU target to validate the kernel for. Defaults to midgard.
      * @param[in] dilation         (Optional) Dilation, in elements, across x and y. Defaults to (1, 1).
      *
      * @return a status
      */
-    static Status validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info,
-                           unsigned int depth_multiplier = 1, const ActivationLayerInfo &act_info = ActivationLayerInfo(), const Size2D &dilation = Size2D(1U, 1U));
+    static Status validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info, unsigned int depth_multiplier = 1,
+                           ActivationLayerInfo act_info = ActivationLayerInfo(), GPUTarget gpu_target = GPUTarget::MIDGARD, const Size2D &dilation = Size2D(1U, 1U));
 
     // Inherited methods overriden:
     void run() override;
     void prepare() override;
 
 private:
-    MemoryGroup _memory_group;
-
-    std::unique_ptr<IFunction>              _optimised_function;
-    CLDepthwiseConvolutionLayerNativeKernel _dwc_native_kernel;
-    CLPermute                               _permute_input_to_nhwc;
-    CLPermute                               _permute_weights_to_nhwc;
-    CLPermute                               _permute_output_to_nchw;
-
-    CLTensor       _permuted_input;
-    CLTensor       _permuted_weights;
-    CLTensor       _permuted_output;
-    const ITensor *_original_weights;
-
-    bool _needs_permute;
-    bool _is_prepared;
+    CLDepthwiseConvolutionLayer _func;
 };
 } // namespace arm_compute
 #endif /*__ARM_COMPUTE_CLDEPTHWISECONVOLUTION_H__ */
diff --git a/arm_compute/runtime/NEON/functions/NEDepthwiseConvolutionLayer.h b/arm_compute/runtime/NEON/functions/NEDepthwiseConvolutionLayer.h
index ea3ef9bf38..8fe9644963 100644
--- a/arm_compute/runtime/NEON/functions/NEDepthwiseConvolutionLayer.h
+++ b/arm_compute/runtime/NEON/functions/NEDepthwiseConvolutionLayer.h
@@ -37,48 +37,43 @@ namespace arm_compute
 // Forward declarations
 class ITensor;
 
-/** Basic function to execute a depthwise convolution for kernel size 3x3xC. This function calls the following NEON kernels:
- *
- * -# @ref NEDepthwiseConvolutionLayer3x3
- * -# @ref NEFillBorderKernel (if pad_x or pad_y > 0)
- *
+/** Function to execute a depthwise convolution.
  */
-class NEDepthwiseConvolutionLayer3x3 : public IFunction
+class NEDepthwiseConvolutionLayer : public IFunction
 {
 public:
     /** Default constructor */
-    NEDepthwiseConvolutionLayer3x3(std::shared_ptr<IMemoryManager> memory_manager = nullptr);
+    NEDepthwiseConvolutionLayer(std::shared_ptr<IMemoryManager> memory_manager = nullptr);
     /** Prevent instances of this class from being copied (As this class contains pointers) */
-    NEDepthwiseConvolutionLayer3x3(const NEDepthwiseConvolutionLayer3x3 &) = delete;
+    NEDepthwiseConvolutionLayer(const NEDepthwiseConvolutionLayer &) = delete;
     /** Default move constructor */
-    NEDepthwiseConvolutionLayer3x3(NEDepthwiseConvolutionLayer3x3 &&) = default;
+    NEDepthwiseConvolutionLayer(NEDepthwiseConvolutionLayer &&) = default;
     /** Prevent instances of this class from being copied (As this class contains pointers) */
-    NEDepthwiseConvolutionLayer3x3 &operator=(const NEDepthwiseConvolutionLayer3x3 &) = delete;
+    NEDepthwiseConvolutionLayer &operator=(const NEDepthwiseConvolutionLayer &) = delete;
     /** Default move assignment operator */
-    NEDepthwiseConvolutionLayer3x3 &operator=(NEDepthwiseConvolutionLayer3x3 &&) = default;
-    /** Initialize the function's source, destination, kernels and border_size.
+    NEDepthwiseConvolutionLayer &operator=(NEDepthwiseConvolutionLayer &&) = default;
+    /** Initialize the function's source, destination, weights and convolution information.
      *
-     * @param[in, out] input            Source tensor. Data type supported: QASYMM8/F16/F32. (Written to only for border filling).
-     * @param[in]      weights          Weights tensor. These are 3D tensors with shape [3, 3, IFM]. Data type supported: Same as @p input.
+     * @param[in, out] input            Source tensor. Data type supported: QASYMM8/F16/F32
+     * @param[out]     output           Destination tensor. Data type supported: same as @p input.
+     * @param[in]      weights          Weights tensor. These are 3D tensors with shape [kernel_x, kernel_y, IFM]. Data type supported: Same as @p input.
      * @param[in]      biases           Biases tensor. A 1D tensor with shape [IFM]. Must be nullptr if not needed.
      *                                  Data type supported: Same as @p input.
-     * @param[out]     output           Destination tensor. Data type supported: same as @p input.
      * @param[in]      conv_info        Padding and stride information to use for the convolution.
      * @param[in]      depth_multiplier (Optional) Multiplier to apply to the input's depth in order to retrieve the output's depth. Defaults to 1.
      * @param[in]      act_info         (Optional) Activation layer information in case of a fused activation.
      * @param[in]      dilation         (Optional) Dilation, in elements, across x and y. Defaults to (1, 1).
      */
-    ARM_COMPUTE_DEPRECATED_REL_REPLACE(19.08, NEDepthwiseConvolutionLayerOptimized)
     void configure(ITensor *input, const ITensor *weights, const ITensor *biases, ITensor *output, const PadStrideInfo &conv_info,
                    unsigned int depth_multiplier = 1, const ActivationLayerInfo &act_info = ActivationLayerInfo(), const Size2D &dilation = Size2D(1U, 1U));
 
-    /** Static function to check if given info will lead to a valid configuration of @ref NEDepthwiseConvolutionLayer3x3
+    /** Static function to check if given info will lead to a valid configuration of @ref NEDepthwiseConvolutionLayer
      *
-     * @param[in] input            Source tensor. Data type supported: QASYMM8/F16/F32. (Written to only for border filling).
-     * @param[in] weights          Weights tensor. These are 3D tensors with shape [3, 3, IFM]. Data type supported: Same as @p input.
+     * @param[in] input            Source tensor. Data type supported: QASYMM8/F16/F32
+     * @param[in] output           Destination tensor. Data type supported: same as @p input.
+     * @param[in] weights          Weights tensor. These are 3D tensors with shape [kernel_x, kernel_y, IFM]. Data type supported: Same as @p input.
      * @param[in] biases           Biases tensor. A 1D tensor with shape [IFM]. Must be nullptr if not needed.
      *                             Data type supported: Same as @p input.
-     * @param[in] output           Destination tensor. Data type supported: same as @p input.
      * @param[in] conv_info        Padding and stride information to use for the convolution.
      * @param[in] depth_multiplier (Optional) Multiplier to apply to the input's depth in order to retrieve the output's depth. Defaults to 1.
      * @param[in] act_info         (Optional) Activation layer information in case of a fused activation.
@@ -94,61 +89,214 @@ public:
     void prepare() override;
 
 private:
-    /** Configure the kernels/functions for the generic pipeline.
+    /** Static function to choose the best depthwise convolution function for @ref NEDepthwiseConvolutionLayer
      *
-     * @param[in, out] input            Source tensor. Data type supported: QASYMM8/F16/F32. (Written to only for border filling).
-     * @param[in]      weights          Weights tensor. These are 3D tensors with shape [3, 3, IFM]. Data type supported: Same as @p input.
-     * @param[in]      biases           Biases tensor. A 1D tensor with shape [IFM]. Must be nullptr if not needed.
-     *                                  Data type supported: Same as @p input.
-     * @param[out]     output           Destination tensor. Data type supported: same as @p input.
-     * @param[in]      conv_info        Padding and stride information to use for the convolution.
-     * @param[in]      depth_multiplier Multiplier to apply to the input's depth in order to retrieve the output's depth. Defaults to 1.
-     * @param[in]      act_info         Activation layer information in case of a fused activation.
-     * @param[in]      dilation         (Optional) Dilation, in elements, across x and y. Defaults to (1, 1).
+     * @param[in] input            Source tensor info. Data type supported: QASYMM8/F16/F32
+     * @param[in] weights          Weights tensor info. These are 3D tensors with shape [kernel_x, kernel_y, IFM]. Data type supported: Same as @p input.
+     * @param[in] biases           Biases tensor info. A 1D tensor with shape [IFM]. Must be nullptr if not needed.
+     *                             Data type supported: Same as @p input.
+     * @param[in] output           Destination tensor. Data type supported: same as @p input.
+     * @param[in] conv_info        Padding and stride information to use for the convolution.
+     * @param[in] depth_multiplier (Optional) Multiplier to apply to the input's depth in order to retrieve the output's depth. Defaults to 1.
+     * @param[in] act_info         (Optional) Activation layer information in case of a fused activation. Only RELU, BOUNDED_RELU and LU_BOUNDED_RELU for 3x3 QASYMM8 supported.
+     * @param[in] dilation         (Optional) Dilation, in elements, across x and y. Defaults to (1, 1).
      *
+     * @return a Depthwise Convolution Function
      */
-    void 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 = Size2D(1U, 1U));
-    /** Configure the kernels/functions for the optimized pipeline.
+    static DepthwiseConvolutionFunction get_depthwiseconvolution_function(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output,
+                                                                          const PadStrideInfo &conv_info, unsigned int depth_multiplier = 1,
+                                                                          ActivationLayerInfo act_info = ActivationLayerInfo(), const Size2D &dilation = Size2D(1U, 1U));
+
+    /** Basic function to execute optimized depthwise convolution routines. This function calls the following NEON kernels:
+    *
+    * @note At the moment 3x3 and 5x5 convolution of stride 1, 2 are supported
+    *
+    * -# @ref NEFillBorderKernel (if pad_x or pad_y > 0) and no assembly kernel implementation is present
+    * -# @ref NEDepthwiseConvolutionLayer3x3Kernel if 3x3 and no assembly kernel implementation is present
+    * -# @ref NEDepthwiseConvolutionAssemblyDispatch if assembly kernel implementation is present
+    * -# @ref NEDirectConvolutionLayerOutputStageKernel if re-quantization of output is required
+    * -# @ref NEActivationLayer if fused activation is required
+    *
+    */
+    class NEDepthwiseConvolutionLayerOptimizedInternal : public IFunction
+    {
+    public:
+        /** Default constructor */
+        NEDepthwiseConvolutionLayerOptimizedInternal(std::shared_ptr<IMemoryManager> memory_manager = nullptr);
+        /** Prevent instances of this class from being copied (As this class contains pointers) */
+        NEDepthwiseConvolutionLayerOptimizedInternal(const NEDepthwiseConvolutionLayerOptimizedInternal &) = delete;
+        /** Default move constructor */
+        NEDepthwiseConvolutionLayerOptimizedInternal(NEDepthwiseConvolutionLayerOptimizedInternal &&) = default;
+        /** Prevent instances of this class from being copied (As this class contains pointers) */
+        NEDepthwiseConvolutionLayerOptimizedInternal &operator=(const NEDepthwiseConvolutionLayerOptimizedInternal &) = delete;
+        /** Default move assignment operator */
+        NEDepthwiseConvolutionLayerOptimizedInternal &operator=(NEDepthwiseConvolutionLayerOptimizedInternal &&) = default;
+        /** Initialize the function's source, destination, kernels and border_size.
+         *
+         * @param[in, out] input            Source tensor. Data type supported: QASYMM8/F16/F32. (Written to only for border filling).
+         * @param[in]      weights          Weights tensor. These are 3D tensors with shape [kernel_x, kernel_y, IFM]. Data type supported: Same as @p input.
+         * @param[in]      biases           Biases tensor. A 1D tensor with shape [IFM]. Must be nullptr if not needed.
+         *                                  Data type supported: Same as @p input.
+         * @param[out]     output           Destination tensor. Data type supported: same as @p input.
+         * @param[in]      conv_info        Padding and stride information to use for the convolution.
+         * @param[in]      depth_multiplier (Optional) Multiplier to apply to the input's depth in order to retrieve the output's depth. Defaults to 1.
+         * @param[in]      act_info         (Optional) Activation layer information in case of a fused activation.
+         * @param[in]      dilation         (Optional) Dilation, in elements, across x and y. Defaults to (1, 1).
+         */
+        void configure(ITensor *input, const ITensor *weights, const ITensor *biases, ITensor *output, const PadStrideInfo &conv_info,
+                       unsigned int depth_multiplier = 1, const ActivationLayerInfo &act_info = ActivationLayerInfo(), const Size2D &dilation = Size2D(1U, 1U));
+
+        /** Static function to check if given info will lead to a valid configuration of @ref NEDepthwiseConvolutionLayer3x3
+         *
+         * @param[in] input            Source tensor. Data type supported: QASYMM8/F16/F32. (Written to only for border filling).
+         * @param[in] weights          Weights tensor. These are 3D tensors with shape [kernel_x, kernel_y, IFM]. Data type supported: Same as @p input.
+         * @param[in] biases           Biases tensor. A 1D tensor with shape [IFM]. Must be nullptr if not needed.
+         *                             Data type supported: Same as @p input.
+         * @param[in] output           Destination tensor. Data type supported: same as @p input.
+         * @param[in] conv_info        Padding and stride information to use for the convolution.
+         * @param[in] depth_multiplier (Optional) Multiplier to apply to the input's depth in order to retrieve the output's depth. Defaults to 1.
+         * @param[in] act_info         (Optional) Activation layer information in case of a fused activation.
+         * @param[in] dilation         (Optional) Dilation, in elements, across x and y. Defaults to (1, 1).
+         *
+         * @return a status
+         */
+        static Status validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info,
+                               unsigned int depth_multiplier = 1, const ActivationLayerInfo &act_info = ActivationLayerInfo(), const Size2D &dilation = Size2D(1U, 1U));
+
+        // Inherited methods overriden:
+        void run() override;
+        void prepare() override;
+
+    private:
+        /** Configure the kernels/functions for the generic pipeline.
+         *
+         * @param[in, out] input            Source tensor. Data type supported: QASYMM8/F16/F32. (Written to only for border filling).
+         * @param[in]      weights          Weights tensor. These are 3D tensors with shape [kernel_x, kernel_y, IFM]. Data type supported: Same as @p input.
+         * @param[in]      biases           Biases tensor. A 1D tensor with shape [IFM]. Must be nullptr if not needed.
+         *                                  Data type supported: Same as @p input.
+         * @param[out]     output           Destination tensor. Data type supported: same as @p input.
+         * @param[in]      conv_info        Padding and stride information to use for the convolution.
+         * @param[in]      depth_multiplier Multiplier to apply to the input's depth in order to retrieve the output's depth. Defaults to 1.
+         * @param[in]      act_info         Activation layer information in case of a fused activation.
+         * @param[in]      dilation         (Optional) Dilation, in elements, across x and y. Defaults to (1, 1).
+         *
+         */
+        void 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 = Size2D(1U, 1U));
+        /** Configure the kernels/functions for the optimized pipeline.
+         *
+         * @param[in]  input            Source tensor. Data type supported: QASYMM8/F16/F32. (Written to only for border filling).
+         * @param[in]  weights          Weights tensor. These are 3D tensors with shape [kernel_x, kernel_y, IFM]. Data type supported: Same as @p input.
+         * @param[in]  biases           Biases tensor. A 1D tensor with shape [IFM]. Must be nullptr if not needed.
+         *                              Data type supported: Same as @p input.
+         * @param[out] output           Destination tensor. Data type supported: same as @p input.
+         * @param[in]  conv_info        Padding and stride information to use for the convolution.
+         * @param[in]  depth_multiplier Multiplier to apply to the input's depth in order to retrieve the output's depth. Defaults to 1.
+         * @param[in]  act_info         Activation layer information in case of a fused activation.
+         */
+        void 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 = Size2D(1U, 1U));
+        /** Run generic kernel */
+        void run_generic();
+        /** Run optimized function */
+        void run_optimized();
+
+        MemoryGroup                               _memory_group;
+        NEDepthwiseConvolutionLayer3x3Kernel      _dwc_kernel;
+        NEDepthwiseConvolutionAssemblyDispatch    _dwc_optimized_func;
+        NEDirectConvolutionLayerOutputStageKernel _output_stage_kernel;
+        NEFillBorderKernel                        _border_handler;
+        NEPermute                                 _permute_input;
+        NEPermute                                 _permute_weights;
+        NEPermute                                 _permute_output;
+        NEActivationLayer                         _activationlayer_function;
+        Tensor                                    _accumulator;
+        Tensor                                    _permuted_input;
+        Tensor                                    _permuted_weights;
+        Tensor                                    _permuted_output;
+        const ITensor                            *_original_weights;
+        bool                                      _has_bias;
+        bool                                      _is_quantized;
+        bool                                      _is_optimized;
+        bool                                      _is_nchw;
+        bool                                      _permute;
+        bool                                      _is_activationlayer_enabled;
+        bool                                      _is_prepared;
+    };
+
+    /** Basic function to execute a generic depthwise convolution. This function calls the following NEON kernel:
+     *
+     * -# @ref NEDepthwiseConvolutionLayerNativeKernel
      *
-     * @param[in]  input            Source tensor. Data type supported: QASYMM8/F16/F32. (Written to only for border filling).
-     * @param[in]  weights          Weights tensor. These are 3D tensors with shape [3, 3, IFM]. Data type supported: Same as @p input.
-     * @param[in]  biases           Biases tensor. A 1D tensor with shape [IFM]. Must be nullptr if not needed.
-     *                              Data type supported: Same as @p input.
-     * @param[out] output           Destination tensor. Data type supported: same as @p input.
-     * @param[in]  conv_info        Padding and stride information to use for the convolution.
-     * @param[in]  depth_multiplier Multiplier to apply to the input's depth in order to retrieve the output's depth. Defaults to 1.
-     * @param[in]  act_info         Activation layer information in case of a fused activation.
      */
-    void 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);
-    /** Run generic kernel */
-    void run_generic();
-    /** Run optimized function */
-    void run_optimized();
+    class NEDepthwiseConvolutionLayerGeneric : public IFunction
+    {
+    public:
+        /** Default constructor */
+        NEDepthwiseConvolutionLayerGeneric();
+        /** Prevent instances of this class from being copied (As this class contains pointers) */
+        NEDepthwiseConvolutionLayerGeneric(const NEDepthwiseConvolutionLayerGeneric &) = delete;
+        /** Default move constructor */
+        NEDepthwiseConvolutionLayerGeneric(NEDepthwiseConvolutionLayerGeneric &&) = default;
+        /** Prevent instances of this class from being copied (As this class contains pointers) */
+        NEDepthwiseConvolutionLayerGeneric &operator=(const NEDepthwiseConvolutionLayerGeneric &) = delete;
+        /** Default move assignment operator */
+        NEDepthwiseConvolutionLayerGeneric &operator=(NEDepthwiseConvolutionLayerGeneric &&) = default;
+        /** Initialize the function's source, destination, weights and convolution information.
+         *
+         * @param[in, out] input            Source tensor. Data type supported: QASYMM8/F16/F32. (Written to only for border filling).
+         * @param[out]     output           Destination tensor. Data type supported: same as @p input.
+         * @param[in]      weights          Weights tensor. These are 3D tensors with shape [kernel_x, kernel_y, IFM]. Data type supported: Same as @p input.
+         * @param[in]      biases           Biases tensor. A 1D tensor with shape [IFM]. Must be nullptr if not needed.
+         *                                  Data type supported: Same as @p input, S32 when input is QASYMM8.
+         * @param[in]      conv_info        Padding and stride information to use for the convolution.
+         * @param[in]      depth_multiplier (Optional) Multiplier to apply to the input's depth in order to retrieve the output's depth. Defaults to 1.
+         * @param[in]      act_info         (Optional) Activation layer information in case of a fused activation.
+         * @param[in]      dilation         (Optional) Dilation, in elements, across x and y. Defaults to (1, 1).
+         */
+        void configure(ITensor *input, const ITensor *weights, const ITensor *biases, ITensor *output, const PadStrideInfo &conv_info,
+                       unsigned int depth_multiplier = 1, const ActivationLayerInfo &act_info = ActivationLayerInfo(), const Size2D &dilation = Size2D(1U, 1U));
 
-private:
-    MemoryGroup                               _memory_group;
-    NEDepthwiseConvolutionLayer3x3Kernel      _dwc_kernel;
-    NEDepthwiseConvolutionAssemblyDispatch    _dwc_optimized_func;
-    NEDirectConvolutionLayerOutputStageKernel _output_stage_kernel;
-    NEFillBorderKernel                        _border_handler;
-    NEPermute                                 _permute_input;
-    NEPermute                                 _permute_weights;
-    NEPermute                                 _permute_output;
-    NEActivationLayer                         _activationlayer_function;
-    Tensor                                    _accumulator;
-    Tensor                                    _permuted_input;
-    Tensor                                    _permuted_weights;
-    Tensor                                    _permuted_output;
-    const ITensor                            *_original_weights;
-    bool                                      _has_bias;
-    bool                                      _is_quantized;
-    bool                                      _is_optimized;
-    bool                                      _is_nchw;
-    bool                                      _permute;
-    bool                                      _is_activationlayer_enabled;
-    bool                                      _is_prepared;
+        /** Static function to check if given info will lead to a valid configuration of @ref NEDepthwiseConvolutionLayerGeneric
+         *
+         * @param[in] input            Source tensor. Data type supported: QASYMM8/F16/F32. (Written to only for border filling).
+         * @param[in] output           Destination tensor. Data type supported: same as @p input.
+         * @param[in] weights          Weights tensor. These are 3D tensors with shape [kernel_x, kernel_y, IFM]. Data type supported: Same as @p input.
+         * @param[in] biases           Biases tensor. A 1D tensor with shape [IFM]. Must be nullptr if not needed.
+         *                             Data type supported: Same as @p input, S32 when input is QASYMM8.
+         * @param[in] conv_info        Padding and stride information to use for the convolution.
+         * @param[in] depth_multiplier (Optional) Multiplier to apply to the input's depth in order to retrieve the output's depth. Defaults to 1.
+         * @param[in] act_info         (Optional) Activation layer information in case of a fused activation.
+         * @param[in] dilation         (Optional) Dilation, in elements, across x and y. Defaults to (1, 1).
+         *
+         * @return a status
+         */
+        static Status validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info,
+                               unsigned int depth_multiplier = 1, const ActivationLayerInfo &act_info = ActivationLayerInfo(), const Size2D &dilation = Size2D(1U, 1U));
+
+        // Inherited methods overriden:
+        void run() override;
+        void prepare() override;
+
+    private:
+        NEDepthwiseConvolutionLayerNativeKernel _depthwise_conv_kernel;
+        NEFillBorderKernel                      _fill_border;
+        NEPermute                               _permute_input;
+        NEPermute                               _permute_weights;
+        NEPermute                               _permute_output;
+        NEActivationLayer                       _activationlayer_function;
+        Tensor                                  _permuted_input;
+        Tensor                                  _permuted_weights;
+        Tensor                                  _permuted_output;
+        bool                                    _is_prepared;
+        bool                                    _is_nchw;
+        bool                                    _is_activationlayer_enabled;
+        const ITensor                          *_original_weights;
+    };
+
+    DepthwiseConvolutionFunction                 _depth_conv_func;
+    NEDepthwiseConvolutionLayerOptimizedInternal _func_optimized;
+    NEDepthwiseConvolutionLayerGeneric           _func_generic;
 };
 
 /** Basic function to execute optimized depthwise convolution routines. This function calls the following NEON kernels:
@@ -187,10 +335,11 @@ public:
      * @param[in]      act_info         (Optional) Activation layer information in case of a fused activation.
      * @param[in]      dilation         (Optional) Dilation, in elements, across x and y. Defaults to (1, 1).
      */
+    ARM_COMPUTE_DEPRECATED_REL_REPLACE(20.02, NEDepthwiseConvolutionLayer)
     void configure(ITensor *input, const ITensor *weights, const ITensor *biases, ITensor *output, const PadStrideInfo &conv_info,
                    unsigned int depth_multiplier = 1, const ActivationLayerInfo &act_info = ActivationLayerInfo(), const Size2D &dilation = Size2D(1U, 1U));
 
-    /** Static function to check if given info will lead to a valid configuration of @ref NEDepthwiseConvolutionLayer3x3
+    /** Static function to check if given info will lead to a valid configuration of @ref NEDepthwiseConvolutionLayerOptimized
      *
      * @param[in] input            Source tensor. Data type supported: QASYMM8/F16/F32. (Written to only for border filling).
      * @param[in] weights          Weights tensor. These are 3D tensors with shape [W, H, IFM]. Data type supported: Same as @p input.
@@ -212,131 +361,7 @@ public:
     void prepare() override;
 
 private:
-    /** Configure the kernels/functions for the generic pipeline.
-     *
-     * @param[in, out] input            Source tensor. Data type supported: QASYMM8/F16/F32. (Written to only for border filling).
-     * @param[in]      weights          Weights tensor. These are 3D tensors with shape [W, H, IFM]. Data type supported: Same as @p input.
-     * @param[in]      biases           Biases tensor. A 1D tensor with shape [IFM]. Must be nullptr if not needed.
-     *                                  Data type supported: Same as @p input.
-     * @param[out]     output           Destination tensor. Data type supported: same as @p input.
-     * @param[in]      conv_info        Padding and stride information to use for the convolution.
-     * @param[in]      depth_multiplier Multiplier to apply to the input's depth in order to retrieve the output's depth. Defaults to 1.
-     * @param[in]      act_info         Activation layer information in case of a fused activation.
-     * @param[in]      dilation         (Optional) Dilation, in elements, across x and y. Defaults to (1, 1).
-     *
-     */
-    void 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 = Size2D(1U, 1U));
-    /** Configure the kernels/functions for the optimized pipeline.
-     *
-     * @param[in]  input            Source tensor. Data type supported: QASYMM8/F16/F32. (Written to only for border filling).
-     * @param[in]  weights          Weights tensor. These are 3D tensors with shape [W, H, IFM]. Data type supported: Same as @p input.
-     * @param[in]  biases           Biases tensor. A 1D tensor with shape [IFM]. Must be nullptr if not needed.
-     *                              Data type supported: Same as @p input.
-     * @param[out] output           Destination tensor. Data type supported: same as @p input.
-     * @param[in]  conv_info        Padding and stride information to use for the convolution.
-     * @param[in]  depth_multiplier Multiplier to apply to the input's depth in order to retrieve the output's depth. Defaults to 1.
-     * @param[in]  act_info         Activation layer information in case of a fused activation.
-     */
-    void 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 = Size2D(1U, 1U));
-    /** Run generic kernel */
-    void run_generic();
-    /** Run optimized function */
-    void run_optimized();
-
-private:
-    MemoryGroup                               _memory_group;
-    NEDepthwiseConvolutionLayer3x3Kernel      _dwc_kernel;
-    NEDepthwiseConvolutionAssemblyDispatch    _dwc_optimized_func;
-    NEDirectConvolutionLayerOutputStageKernel _output_stage_kernel;
-    NEFillBorderKernel                        _border_handler;
-    NEPermute                                 _permute_input;
-    NEPermute                                 _permute_weights;
-    NEPermute                                 _permute_output;
-    NEActivationLayer                         _activationlayer_function;
-    Tensor                                    _accumulator;
-    Tensor                                    _permuted_input;
-    Tensor                                    _permuted_weights;
-    Tensor                                    _permuted_output;
-    const ITensor                            *_original_weights;
-    bool                                      _has_bias;
-    bool                                      _is_quantized;
-    bool                                      _is_optimized;
-    bool                                      _is_nchw;
-    bool                                      _permute;
-    bool                                      _is_activationlayer_enabled;
-    bool                                      _is_prepared;
-};
-
-/** Basic function to execute a generic depthwise convolution. This function calls the following NEON kernel:
- *
- * -# @ref NEDepthwiseConvolutionLayerNativeKernel
- *
- */
-class NEDepthwiseConvolutionLayer : public IFunction
-{
-public:
-    /** Default constructor */
-    NEDepthwiseConvolutionLayer();
-    /** Prevent instances of this class from being copied (As this class contains pointers) */
-    NEDepthwiseConvolutionLayer(const NEDepthwiseConvolutionLayer &) = delete;
-    /** Default move constructor */
-    NEDepthwiseConvolutionLayer(NEDepthwiseConvolutionLayer &&) = default;
-    /** Prevent instances of this class from being copied (As this class contains pointers) */
-    NEDepthwiseConvolutionLayer &operator=(const NEDepthwiseConvolutionLayer &) = delete;
-    /** Default move assignment operator */
-    NEDepthwiseConvolutionLayer &operator=(NEDepthwiseConvolutionLayer &&) = default;
-    /** Initialize the function's source, destination, weights and convolution information.
-     *
-     * @param[in, out] input            Source tensor. Data type supported: QASYMM8/F16/F32. (Written to only for border filling).
-     * @param[out]     output           Destination tensor. Data type supported: same as @p input.
-     * @param[in]      weights          Weights tensor. These are 3D tensors with shape [kernel_x, kernel_y, IFM]. Data type supported: Same as @p input.
-     * @param[in]      biases           (Optional) Biases tensor. A 1D tensor with shape [IFM]. Must be nullptr if not needed.
-     *                                  Data type supported: Same as @p input, S32 when input is QASYMM8.
-     * @param[in]      conv_info        Padding and stride information to use for the convolution.
-     * @param[in]      depth_multiplier (Optional) Multiplier to apply to the input's depth in order to retrieve the output's depth. Defaults to 1.
-     * @param[in]      act_info         (Optional) Activation layer information in case of a fused activation.
-     * @param[in]      dilation         (Optional) Dilation, in elements, across x and y. Defaults to (1, 1).
-     */
-    void configure(ITensor *input, const ITensor *weights, const ITensor *biases, ITensor *output, const PadStrideInfo &conv_info,
-                   unsigned int depth_multiplier = 1, const ActivationLayerInfo &act_info = ActivationLayerInfo(), const Size2D &dilation = Size2D(1U, 1U));
-
-    /** Static function to check if given info will lead to a valid configuration of @ref NEDepthwiseConvolutionLayer
-     *
-     * @param[in] input            Source tensor. Data type supported: QASYMM8/F16/F32. (Written to only for border filling).
-     * @param[in] output           Destination tensor. Data type supported: same as @p input.
-     * @param[in] weights          Weights tensor. These are 3D tensors with shape [kernel_x, kernel_y, IFM]. Data type supported: Same as @p input.
-     * @param[in] biases           (Optional) Biases tensor. A 1D tensor with shape [IFM]. Must be nullptr if not needed.
-     *                             Data type supported: Same as @p input, S32 when input is QASYMM8.
-     * @param[in] conv_info        Padding and stride information to use for the convolution.
-     * @param[in] depth_multiplier (Optional) Multiplier to apply to the input's depth in order to retrieve the output's depth. Defaults to 1.
-     * @param[in] act_info         (Optional) Activation layer information in case of a fused activation.
-     * @param[in] dilation         (Optional) Dilation, in elements, across x and y. Defaults to (1, 1).
-     *
-     * @return a status
-     */
-    static Status validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info,
-                           unsigned int depth_multiplier = 1, const ActivationLayerInfo &act_info = ActivationLayerInfo(), const Size2D &dilation = Size2D(1U, 1U));
-
-    // Inherited methods overriden:
-    void run() override;
-    void prepare() override;
-
-private:
-    NEDepthwiseConvolutionLayerNativeKernel _depthwise_conv_kernel;
-    NEFillBorderKernel                      _fill_border;
-    NEPermute                               _permute_input;
-    NEPermute                               _permute_weights;
-    NEPermute                               _permute_output;
-    NEActivationLayer                       _activationlayer_function;
-    Tensor                                  _permuted_input;
-    Tensor                                  _permuted_weights;
-    Tensor                                  _permuted_output;
-    bool                                    _is_prepared;
-    bool                                    _is_nchw;
-    bool                                    _is_activationlayer_enabled;
-    const ITensor                          *_original_weights;
+    NEDepthwiseConvolutionLayer _func;
 };
 } // namespace arm_compute
 #endif /* __ARM_COMPUTE_NEDEPTHWISECONVOLUTION_H__ */
\ No newline at end of file