diff --git a/arm_compute/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NCHWKernel.h b/arm_compute/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NCHWKernel.h
index b1c730d..3b7fc7b 100644
--- a/arm_compute/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NCHWKernel.h
+++ b/arm_compute/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NCHWKernel.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2018 ARM Limited.
+ * Copyright (c) 2018-2019 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -41,31 +41,33 @@
      *
      * @param[in]  input            Source tensor. DataType supported: QASYMM8/F16/F32.
      * @param[in]  weights          Weights tensor. A 3D tensor with dimensions [3, 3, IFM]. Data type supported: Same as @p input.
-     * @param[in]  biases           (Optional) Biases tensor. A 1D tensor with dimensions [IFM]. Must be nullptr if not needed.
+     * @param[in]  biases           Biases tensor. A 1D tensor with dimensions [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 QASYMM8 supported.
+     * @param[in]  dilation         (Optional) Dilation, in elements, across x and y. Defaults to (1, 1).
      */
     void configure(const ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const PadStrideInfo &conv_info, unsigned int depth_multiplier,
-                   ActivationLayerInfo act_info) override;
+                   ActivationLayerInfo act_info, const Size2D &dilation) override;
     /** Static function to check if given info will lead to a valid configuration of @ref CLDepthwiseConvolutionLayer3x3NCHWKernel
      *
-     * @param[in] input            Source tensor. DataType supported: F16/F32/QASYMM8.
-     * @param[in] weights          Weights tensor. A 3D tensor with dimensions [3, 3, IFM]. Data type supported: Same as @p input.
-     * @param[in] biases           Biases tensor. A 1D tensor with dimensions [IFM]. Must be nullptr if not needed.
+     * @param[in] input            Source tensor info. DataType supported: F16/F32/QASYMM8.
+     * @param[in] weights          Weights tensor info. A 3D tensor with dimensions [3, 3, IFM]. Data type supported: Same as @p input.
+     * @param[in] biases           Biases tensor info. A 1D tensor with dimensions [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] 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         (Optional) Activation layer information in case of a fused activation. Only RELU, BOUNDED_RELU and LU_BOUNDED_RELU are 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,
-                           ActivationLayerInfo act_info = ActivationLayerInfo(), GPUTarget gpu_target = GPUTarget::MIDGARD);
+                           ActivationLayerInfo act_info = ActivationLayerInfo(), GPUTarget gpu_target = GPUTarget::MIDGARD, const Size2D &dilation = Size2D(1U, 1U));
 
     void run(const Window &window, cl::CommandQueue &queue) override;
     BorderSize border_size() const override;
diff --git a/arm_compute/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NHWCKernel.h b/arm_compute/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NHWCKernel.h
index 2fc9780..7d0ecec 100644
--- a/arm_compute/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NHWCKernel.h
+++ b/arm_compute/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NHWCKernel.h
@@ -42,30 +42,32 @@
      *
      * @param[in]  input            Source tensor. DataType supported: QASYMM8.
      * @param[in]  weights          Weights tensor. A 3D tensor with dimensions [IFM, 3, 3]. Data type supported: Same as @p input.
-     * @param[in]  biases           (Optional) Biases tensor. A 1D tensor with dimensions [IFM]. Must be nullptr if not needed.
+     * @param[in]  biases           Biases tensor. A 1D tensor with dimensions [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 are supported.
+     * @param[in]  dilation         (Optional) Dilation, in elements, across x and y. Defaults to (1, 1).
      */
     void configure(const ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const PadStrideInfo &conv_info, unsigned int depth_multiplier,
-                   ActivationLayerInfo act_info) override;
+                   ActivationLayerInfo act_info, const Size2D &dilation) override;
     /** Static function to check if given info will lead to a valid configuration of @ref CLDepthwiseConvolutionLayer3x3NHWCKernel
      *
      * @param[in] input            Source tensor info. DataType supported: QASYMM8.
      * @param[in] weights          Weights tensor info. A 3D tensor with dimensions [IFM, 3, 3]. Data type supported: Same as @p input.
-     * @param[in] biases           (Optional) Biases tensor info. A 1D tensor with dimensions [IFM]. Must be nullptr if not needed.
+     * @param[in] biases           Biases tensor info. A 1D tensor with dimensions [IFM]. Must be nullptr if not needed.
      *                             Data type supported: Same as @p input.
      * @param[in] output           Destination tensor info. 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 are 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,
-                           ActivationLayerInfo act_info = ActivationLayerInfo());
+                           ActivationLayerInfo act_info = ActivationLayerInfo(), const Size2D &dilation = Size2D(1U, 1U));
 
     // Inherited methods overridden:
     void run(const Window &window, cl::CommandQueue &queue) override;
diff --git a/arm_compute/core/CL/kernels/CLDepthwiseIm2ColKernel.h b/arm_compute/core/CL/kernels/CLDepthwiseIm2ColKernel.h
index 00d9cb6..1579847 100644
--- a/arm_compute/core/CL/kernels/CLDepthwiseIm2ColKernel.h
+++ b/arm_compute/core/CL/kernels/CLDepthwiseIm2ColKernel.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017-2018 ARM Limited.
+ * Copyright (c) 2017-2019 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -58,22 +58,26 @@
      * @param[in]  conv_info        Contains padding and stride information described in @ref PadStrideInfo.
      * @param[in]  has_bias         Boolean that specifies if the depthwise convolution has bias.
      * @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]  dilation         (Optional) Dilation, in elements, across x and y. Defaults to (1, 1).
      */
-    void configure(const ICLTensor *input, ICLTensor *output, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias = false, unsigned int depth_multiplier = 1);
+    void configure(const ICLTensor *input, ICLTensor *output, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias = false, unsigned int depth_multiplier = 1,
+                   const Size2D &dilation = Size2D(1U, 1U));
     /** Static function to check if given info will lead to a valid configuration of @ref CLDepthwiseIm2ColKernel
      *
-     * @param[in] input            The input tensor to convert. 3 lower dimensions represent a single input [width, height, IFM],
+     * @param[in] input            The input tensor info to convert. 3 lower dimensions represent a single input [width, height, IFM],
      *                             while every optional dimension from 4 and above represent a batch of inputs. Data types supported: QASYMM8/F32
-     * @param[in] output           The output tensor. First 3 lower dimensions represent a transform of each 3D input,
+     * @param[in] output           The output tensor info. First 3 lower dimensions represent a transform of each 3D input,
      *                             while every dimension above 3 represents a batch. Data types supported: Same as @p input
      * @param[in] kernel_dims      The kernel dimensions (width and height).
      * @param[in] conv_info        Contains padding and stride information described in @ref PadStrideInfo.
      * @param[in] has_bias         Boolean that specifies if the depthwise convolution has bias.
-     * @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] depth_multiplier Multiplier to apply to the input's depth in order to retrieve the output's depth. Defaults to 1.
+     * @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 *output, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias, unsigned int depth_multiplier);
+    static Status validate(const ITensorInfo *input, const ITensorInfo *output, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias, unsigned int depth_multiplier,
+                           const Size2D &dilation = Size2D(1U, 1U));
 
     // Inherited methods overridden:
     void run(const Window &window, cl::CommandQueue &queue) override;
diff --git a/arm_compute/core/CL/kernels/ICLDepthwiseConvolutionLayer3x3Kernel.h b/arm_compute/core/CL/kernels/ICLDepthwiseConvolutionLayer3x3Kernel.h
index 3396de2..92eca89 100644
--- a/arm_compute/core/CL/kernels/ICLDepthwiseConvolutionLayer3x3Kernel.h
+++ b/arm_compute/core/CL/kernels/ICLDepthwiseConvolutionLayer3x3Kernel.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017-2018 ARM Limited.
+ * Copyright (c) 2017-2019 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -52,15 +52,16 @@
      *
      * @param[in]  input            Source tensor. DataType supported: QASYMM8/F16/F32.
      * @param[in]  weights          Weights tensor. A 3D tensor with dimensions [3, 3, IFM]. Data type supported: Same as @p input.
-     * @param[in]  biases           (Optional) Biases tensor. A 1D tensor with dimensions [IFM]. Must be nullptr if not needed.
+     * @param[in]  biases           Biases tensor. A 1D tensor with dimensions [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 QASYMM8 supported.
+     * @param[in]  dilation         (Optional) Dilation, in elements, across x and y. Defaults to (1, 1).
      */
     virtual void configure(const ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const PadStrideInfo &conv_info, unsigned int depth_multiplier = 1,
-                           ActivationLayerInfo act_info = ActivationLayerInfo()) = 0;
+                           ActivationLayerInfo act_info = ActivationLayerInfo(), const Size2D &dilation = Size2D(1U, 1U)) = 0;
 
 protected:
     BorderSize       _border_size;
diff --git a/arm_compute/core/utils/misc/ShapeCalculator.h b/arm_compute/core/utils/misc/ShapeCalculator.h
index 26b337d..b46b1b2 100644
--- a/arm_compute/core/utils/misc/ShapeCalculator.h
+++ b/arm_compute/core/utils/misc/ShapeCalculator.h
@@ -402,10 +402,12 @@
  * @param[in] weights          Weights tensor info
  * @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.
+ * @param[in] dilation         Dilation, in elements, across x and y. Defaults to (1, 1).
  *
  * @return the calculated shape
  */
-inline TensorShape compute_depthwise_convolution_shape(const ITensorInfo &input, const ITensorInfo &weights, PadStrideInfo conv_info, unsigned int depth_multiplier)
+inline TensorShape compute_depthwise_convolution_shape(const ITensorInfo &input, const ITensorInfo &weights, PadStrideInfo conv_info, unsigned int depth_multiplier, const Size2D &dilation = Size2D(1U,
+                                                       1U))
 {
     const TensorShape input_shape{ input.tensor_shape() };
     const TensorShape weights_shape{ weights.tensor_shape() };
@@ -415,43 +417,15 @@
     const int        height_idx  = get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT);
     const int        channel_idx = get_data_layout_dimension_index(data_layout, DataLayoutDimension::CHANNEL);
 
-    unsigned int output_width  = 0;
-    unsigned int output_height = 0;
-    std::tie(output_width, output_height) = scaled_dimensions(input_shape[width_idx], input_shape[height_idx],
-                                                              weights_shape[width_idx], weights_shape[height_idx],
-                                                              conv_info);
-
-    TensorShape output_shape{ input_shape };
-    output_shape.set(width_idx, output_width);
-    output_shape.set(height_idx, output_height);
-    output_shape.set(channel_idx, input_shape[channel_idx] * depth_multiplier);
-
-    return output_shape;
-}
-
-/** Calculate the depthwise convolution output shape of a tensor
- *
- * @param[in] input            Input tensor info
- * @param[in] weights_width    Weights width
- * @param[in] weights_height   Weights height
- * @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.
- *
- * @return the calculated shape
- */
-inline TensorShape compute_depthwise_convolution_shape(const ITensorInfo &input, int weights_width, int weights_height, PadStrideInfo conv_info, unsigned int depth_multiplier)
-{
-    const TensorShape input_shape{ input.tensor_shape() };
-
-    const DataLayout data_layout = input.data_layout();
-    const int        width_idx   = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH);
-    const int        height_idx  = get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT);
-    const int        channel_idx = get_data_layout_dimension_index(data_layout, DataLayoutDimension::CHANNEL);
+    const DataLayout weights_data_layout = weights.data_layout();
+    const int        weights_width_idx   = get_data_layout_dimension_index(weights_data_layout, DataLayoutDimension::WIDTH);
+    const int        weights_height_idx  = get_data_layout_dimension_index(weights_data_layout, DataLayoutDimension::HEIGHT);
 
     unsigned int output_width  = 0;
     unsigned int output_height = 0;
     std::tie(output_width, output_height) = scaled_dimensions(input_shape[width_idx], input_shape[height_idx],
-                                                              weights_width, weights_width, conv_info);
+                                                              weights_shape[weights_width_idx], weights_shape[weights_height_idx],
+                                                              conv_info, dilation);
 
     TensorShape output_shape{ input_shape };
     output_shape.set(width_idx, output_width);
diff --git a/arm_compute/runtime/CL/functions/CLDepthwiseConvolutionLayer.h b/arm_compute/runtime/CL/functions/CLDepthwiseConvolutionLayer.h
index 23034c2..b25c36a 100644
--- a/arm_compute/runtime/CL/functions/CLDepthwiseConvolutionLayer.h
+++ b/arm_compute/runtime/CL/functions/CLDepthwiseConvolutionLayer.h
@@ -70,32 +70,34 @@
      *
      * @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           (Optional) Biases tensor. A 1D tensor with shape [IFM]. Must be nullptr if not needed.
+     * @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());
+                   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. Data type supported: QASYMM8 for all layouts, F16/F32 for NCHW.
-     * @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.
+     * @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] 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);
+                           ActivationLayerInfo act_info = ActivationLayerInfo(), GPUTarget gpu_target = GPUTarget::MIDGARD, const Size2D &dilation = Size2D(1U, 1U));
     // Inherited methods overriden:
     void run() override;
     void prepare() override;
@@ -142,31 +144,33 @@
      *
      * @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           (Optional) Biases tensor. A 1D tensor with shape [IFM]. Must be nullptr if not needed.
+     * @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());
+                   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 CLDepthwiseConvolutionLayer
      *
-     * @param[in] input            Source tensor. Data type supported: QASYMM8/F32.
-     * @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.
+     * @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());
+                           unsigned int depth_multiplier = 1, const ActivationLayerInfo &act_info = ActivationLayerInfo(), const Size2D &dilation = Size2D(1U, 1U));
 
     // Inherited methods overriden:
     void run() override;
diff --git a/arm_compute/runtime/GLES_COMPUTE/functions/GCDepthwiseConvolutionLayer.h b/arm_compute/runtime/GLES_COMPUTE/functions/GCDepthwiseConvolutionLayer.h
index 5eccc4d..28db5e1 100644
--- a/arm_compute/runtime/GLES_COMPUTE/functions/GCDepthwiseConvolutionLayer.h
+++ b/arm_compute/runtime/GLES_COMPUTE/functions/GCDepthwiseConvolutionLayer.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017-2018 ARM Limited.
+ * Copyright (c) 2017-2019 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -50,15 +50,16 @@
      *
      * @param[in, out] input            Source tensor. Data type supported: F16. (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           (Optional) Biases tensor. A 1D tensor with shape [IFM]. Must be nullptr if not needed.
+     * @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). Currently supports (1,1) only.
      */
     void configure(IGCTensor *input, const IGCTensor *weights, const IGCTensor *biases, IGCTensor *output, const PadStrideInfo &conv_info,
-                   unsigned int depth_multiplier = 1, const ActivationLayerInfo &act_info = ActivationLayerInfo());
+                   unsigned int depth_multiplier = 1, const ActivationLayerInfo &act_info = ActivationLayerInfo(), const Size2D &dilation = Size2D(1U, 1U));
 
     // Inherited methods overridden:
     void run() override final;
diff --git a/arm_compute/runtime/NEON/functions/NEDepthwiseConvolutionLayer.h b/arm_compute/runtime/NEON/functions/NEDepthwiseConvolutionLayer.h
index 28f0560..c602336 100644
--- a/arm_compute/runtime/NEON/functions/NEDepthwiseConvolutionLayer.h
+++ b/arm_compute/runtime/NEON/functions/NEDepthwiseConvolutionLayer.h
@@ -67,31 +67,33 @@
      *
      * @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           (Optional) Biases tensor. A 1D tensor with shape [IFM]. Must be nullptr if not needed.
+     * @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). Currently supports (1,1) only.
      */
     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());
+                   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 [3, 3, 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.
+     * @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). Currently supports (1,1) only.
      *
      * @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());
+                           unsigned int depth_multiplier = 1, const ActivationLayerInfo &act_info = ActivationLayerInfo(), const Size2D &dilation = Size2D(1U, 1U));
 
     // Inherited methods overriden:
     void run() override;
@@ -102,7 +104,7 @@
      *
      * @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           (Optional) Biases tensor. A 1D tensor with shape [IFM]. Must be nullptr if not needed.
+     * @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.
@@ -115,7 +117,7 @@
      *
      * @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           (Optional) Biases tensor. A 1D tensor with shape [IFM]. Must be nullptr if not needed.
+     * @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.
@@ -184,9 +186,10 @@
      * @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). Currently supports (1,1) only.
      */
     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());
+                   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
      *
@@ -198,11 +201,12 @@
      * @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). Currently supports (1,1) only.
      *
      * @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());
+                           unsigned int depth_multiplier = 1, const ActivationLayerInfo &act_info = ActivationLayerInfo(), const Size2D &dilation = Size2D(1U, 1U));
 
     // Inherited methods overriden:
     void run() override;
diff --git a/src/core/CL/cl_kernels/depthwise_convolution.cl b/src/core/CL/cl_kernels/depthwise_convolution.cl
index 4f6fdfa..8ee0185 100644
--- a/src/core/CL/cl_kernels/depthwise_convolution.cl
+++ b/src/core/CL/cl_kernels/depthwise_convolution.cl
@@ -51,13 +51,18 @@
                                       const float           middle_coeff,
                                       const float           right_coeff)
 {
+#if(DILATION_X == 1 && DILATION_Y == 1)
     float4 temp = vload4(0, (__global float *)left_pixel);
 
     float2 left   = CONVERT(temp.s01, float2);
     float2 middle = CONVERT(temp.s12, float2);
     float2 right  = CONVERT(temp.s23, float2);
-
     return left * (float2)left_coeff + middle * (float2)middle_coeff + right * (float2)right_coeff;
+#else  /* DILATION_X==1 && DILATION_Y==1 */
+    return vload2(0, (__global float *)left_pixel) * (float2)left_coeff
+           + vload2(0, (__global float *)(left_pixel) + DILATION_X) * (float2)middle_coeff
+           + vload2(0, (__global float *)(left_pixel) + 2 * DILATION_X) * (float2)right_coeff;
+#endif /* DILATION_X==1 && DILATION_Y==1 */
 }
 
 /** Compute a 1D horizontal convolution of size 3 and stride 2 for floating point type.
@@ -74,6 +79,7 @@
                                       const float           middle_coeff,
                                       const float           right_coeff)
 {
+#if(DILATION_X == 1 && DILATION_Y == 1)
     float4 temp0 = vload4(0, (__global float *)left_pixel);
     float  temp1 = *((__global float *)(left_pixel + 4 * sizeof(float)));
 
@@ -82,6 +88,14 @@
     float2 right  = CONVERT((float2)(temp0.s2, temp1), float2);
 
     return left * (float2)left_coeff + middle * (float2)middle_coeff + right * (float2)right_coeff;
+#else /* DILATION_X==1 && DILATION_Y==1 */
+    __global float *left_pixel_float = (__global float *)left_pixel;
+
+    return vload4(0, left_pixel_float).s02 * (float2)left_coeff
+           + vload4(0, left_pixel_float + DILATION_X).s02 * (float2)middle_coeff
+           + vload4(0, left_pixel_float + DILATION_X * 2).s02 * (float2)right_coeff;
+
+#endif /* DILATION_X==1 && DILATION_Y==1 */
 }
 
 /** Compute a 1D horizontal convolution of size 3 and stride 3 for floating point type.
@@ -98,6 +112,7 @@
                                       const float           middle_coeff,
                                       const float           right_coeff)
 {
+#if(DILATION_X == 1 && DILATION_Y == 1)
     float4 temp0 = vload4(0, (__global float *)left_pixel);
     float2 temp1 = vload2(0, (__global float *)(left_pixel + 4 * sizeof(float)));
 
@@ -106,6 +121,13 @@
     float2 right  = CONVERT((float2)(temp0.s2, temp1.s1), float2);
 
     return left * (float2)left_coeff + middle * (float2)middle_coeff + right * (float2)right_coeff;
+#else  /* DILATION_X==1 && DILATION_Y==1 */
+    __global float *left_pixel_float = (__global float *)left_pixel;
+
+    return (float2)(*left_pixel_float, *(left_pixel_float + 3)) * (float2)left_coeff
+           + (float2)(*(left_pixel_float + DILATION_X), *(left_pixel_float + DILATION_X + 3)) * (float2)middle_coeff
+           + (float2)(*(left_pixel_float + DILATION_X * 2), *(left_pixel_float + DILATION_X * 2 + 3)) * (float2)right_coeff;
+#endif /* DILATION_X==1 && DILATION_Y==1 */
 }
 
 /** Apply a 3x3 convolution matrix to a single channel F32 input image and return the result.
@@ -139,8 +161,8 @@
     float2 pixels;
 
     pixels = convolution1x3(offset(src, 0, 0), mat0, mat1, mat2);
-    pixels += convolution1x3(offset(src, 0, 1), mat3, mat4, mat5);
-    pixels += convolution1x3(offset(src, 0, 2), mat6, mat7, mat8);
+    pixels += convolution1x3(offset(src, 0, DILATION_Y), mat3, mat4, mat5);
+    pixels += convolution1x3(offset(src, 0, DILATION_Y * 2), mat6, mat7, mat8);
 
     return pixels;
 }
@@ -216,6 +238,8 @@
 }
 #endif //defined(CONV_STRIDE_X)
 
+#if(DILATION_X == 1 && DILATION_Y == 1)
+
 #define CONVOLUTION1x3_BIFROST2X1_STRIDE1(acc, src0, weights_row0) \
     ({                                                             \
         acc.s0 = fma(src0.s0, weights_row0.s0, acc.s0);            \
@@ -268,6 +292,227 @@
         acc.s3 = fma(src1.s0, weights_row0.s2, acc.s3);                  \
     })
 
+#else /* DILATION_X==1 && DILATION_Y==1 */
+
+#define CONVOLUTION1x3_BIFROST2X1_STRIDE1(acc, src0_left, src0_mid, src0_right, weights_row0) \
+    ({                                                                                        \
+        acc.s0 = fma(src0_left.s0, weights_row0.s0, acc.s0);                                  \
+        acc.s0 = fma(src0_mid.s0, weights_row0.s1, acc.s0);                                   \
+        acc.s0 = fma(src0_right.s0, weights_row0.s2, acc.s0);                                 \
+        acc.s1 = fma(src0_left.s1, weights_row0.s0, acc.s1);                                  \
+        acc.s1 = fma(src0_mid.s1, weights_row0.s1, acc.s1);                                   \
+        acc.s1 = fma(src0_right.s1, weights_row0.s2, acc.s1);                                 \
+    })
+
+#define CONVOLUTION1x3_BIFROST2X1_STRIDE2(acc, src0_left, src0_mid, src0_right, weights_row0) \
+    ({                                                                                        \
+        acc.s0 = fma(src0_left.s0, weights_row0.s0, acc.s0);                                  \
+        acc.s0 = fma(src0_mid.s0, weights_row0.s1, acc.s0);                                   \
+        acc.s0 = fma(src0_right.s0, weights_row0.s2, acc.s0);                                 \
+        acc.s1 = fma(src0_left.s2, weights_row0.s0, acc.s1);                                  \
+        acc.s1 = fma(src0_mid.s2, weights_row0.s1, acc.s1);                                   \
+        acc.s1 = fma(src0_right.s2, weights_row0.s2, acc.s1);                                 \
+    })
+
+#define CONVOLUTION1x3_BIFROST4X1_STRIDE1(acc, src0_left, src0_mid, src0_right, weights_row0) \
+    ({                                                                                        \
+        acc.s0 = fma(src0_left.s0, weights_row0.s0, acc.s0);                                  \
+        acc.s0 = fma(src0_mid.s0, weights_row0.s1, acc.s0);                                   \
+        acc.s0 = fma(src0_right.s0, weights_row0.s2, acc.s0);                                 \
+        acc.s1 = fma(src0_left.s1, weights_row0.s0, acc.s1);                                  \
+        acc.s1 = fma(src0_mid.s1, weights_row0.s1, acc.s1);                                   \
+        acc.s1 = fma(src0_right.s1, weights_row0.s2, acc.s1);                                 \
+        acc.s2 = fma(src0_left.s2, weights_row0.s0, acc.s2);                                  \
+        acc.s2 = fma(src0_mid.s2, weights_row0.s1, acc.s2);                                   \
+        acc.s2 = fma(src0_right.s2, weights_row0.s2, acc.s2);                                 \
+        acc.s3 = fma(src0_left.s3, weights_row0.s0, acc.s3);                                  \
+        acc.s3 = fma(src0_mid.s3, weights_row0.s1, acc.s3);                                   \
+        acc.s3 = fma(src0_right.s3, weights_row0.s2, acc.s3);                                 \
+    })
+
+#define CONVOLUTION1x3_BIFROST4X1_STRIDE2(acc, src0_left, src0_mid, src0_right, weights_row0) \
+    ({                                                                                        \
+        acc.s0 = fma(src0_left.s0, weights_row0.s0, acc.s0);                                  \
+        acc.s0 = fma(src0_mid.s0, weights_row0.s1, acc.s0);                                   \
+        acc.s0 = fma(src0_right.s0, weights_row0.s2, acc.s0);                                 \
+        acc.s1 = fma(src0_left.s2, weights_row0.s0, acc.s1);                                  \
+        acc.s1 = fma(src0_mid.s2, weights_row0.s1, acc.s1);                                   \
+        acc.s1 = fma(src0_right.s2, weights_row0.s2, acc.s1);                                 \
+        acc.s2 = fma(src0_left.s4, weights_row0.s0, acc.s2);                                  \
+        acc.s2 = fma(src0_mid.s4, weights_row0.s1, acc.s2);                                   \
+        acc.s2 = fma(src0_right.s4, weights_row0.s2, acc.s2);                                 \
+        acc.s3 = fma(src0_left.s6, weights_row0.s0, acc.s3);                                  \
+        acc.s3 = fma(src0_mid.s6, weights_row0.s1, acc.s3);                                   \
+        acc.s3 = fma(src0_right.s6, weights_row0.s2, acc.s3);                                 \
+    })
+
+/** Get the pointer position at a certain offset in x and y direction.
+ *
+ * @param[in] ptr          Pointer to the starting position of the buffer
+ * @param[in] x            Relative X position
+ * @param[in] y            Relative Y position
+ * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes)
+ * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes)
+ */
+inline __global uchar *ptr_offset(__global uchar *ptr, const int x, const int y, const int stride_x, const int stride_y)
+{
+    return ptr + x * stride_x + y * stride_y;
+}
+
+/** Perform 3x3 convolution for stride_x=1 and stride_y=1 when DILATION_X>1 and DILATION_Y>1 for F32
+ *
+ * @param[in] src_addr         Pointer to the starting position of where to perform the convolution
+ * @param[in] src_stride_x     Stride of the source tensor in X dimension (in bytes)
+ * @param[in] src_stride_y     Stride of the source tensor in Y dimension (in bytes)
+ * @param[in] y_offset         Offset from the source tensor from which to start convolution
+ * @param[in] weights_addr     Pointer from where to get weights
+ * @param[in] weights_stride_y Stride of weights tesnsor in Y dimension
+ */
+inline float2 convolution_3x3_dilation_stridex1_stridey1_bifrost_f32(__global uchar *src_addr, const int stride_x_bytes, const int stride_y_bytes,
+                                                                     const int y_offset, __global uchar *weights_addr, const int weights_stride_y)
+{
+    // Load the weights
+    float3 weights_row0 = vload3(0, (__global float *)(weights_addr + 0 * weights_stride_y));
+    float3 weights_row1 = vload3(0, (__global float *)(weights_addr + 1 * weights_stride_y));
+    float3 weights_row2 = vload3(0, (__global float *)(weights_addr + 2 * weights_stride_y));
+
+    float2 pixels0 = 0.0f;
+
+    float2 src00_left  = vload2(0, (__global float *)ptr_offset(src_addr, 0, y_offset, stride_x_bytes, stride_y_bytes)); // Row0
+    float2 src00_mid   = vload2(0, (__global float *)ptr_offset(src_addr, DILATION_X, y_offset, stride_x_bytes, stride_y_bytes));
+    float2 src00_right = vload2(0, (__global float *)ptr_offset(src_addr, 2 * DILATION_X, y_offset, stride_x_bytes, stride_y_bytes));
+
+    float2 src10_left  = vload2(0, (__global float *)ptr_offset(src_addr, 0, y_offset + DILATION_Y, stride_x_bytes, stride_y_bytes)); // Row1
+    float2 src10_mid   = vload2(0, (__global float *)ptr_offset(src_addr, DILATION_X, y_offset + DILATION_Y, stride_x_bytes, stride_y_bytes));
+    float2 src10_right = vload2(0, (__global float *)ptr_offset(src_addr, 2 * DILATION_X, y_offset + DILATION_Y, stride_x_bytes, stride_y_bytes));
+
+    float2 src20_left  = vload2(0, (__global float *)ptr_offset(src_addr, 0, y_offset + DILATION_Y * 2, stride_x_bytes, stride_y_bytes)); // Row2
+    float2 src20_mid   = vload2(0, (__global float *)ptr_offset(src_addr, DILATION_X, y_offset + DILATION_Y * 2, stride_x_bytes, stride_y_bytes));
+    float2 src20_right = vload2(0, (__global float *)ptr_offset(src_addr, 2 * DILATION_X, y_offset + DILATION_Y * 2, stride_x_bytes, stride_y_bytes));
+
+    CONVOLUTION1x3_BIFROST2X1_STRIDE1(pixels0, src00_left, src00_mid, src00_right, weights_row0);
+    CONVOLUTION1x3_BIFROST2X1_STRIDE1(pixels0, src10_left, src10_mid, src10_right, weights_row1);
+    CONVOLUTION1x3_BIFROST2X1_STRIDE1(pixels0, src20_left, src20_mid, src20_right, weights_row2);
+
+    return pixels0;
+}
+
+/** Perform 3x3 convolution for stride_x=2 and stride_y=2 when DILATION_X>1 and DILATION_Y>1 for F32
+ *
+ * @param[in] src_addr         Pointer to the starting position of where to perform the convolution
+ * @param[in] src_stride_x     Stride of the source tensor in X dimension (in bytes)
+ * @param[in] src_stride_y     Stride of the source tensor in Y dimension (in bytes)
+ * @param[in] y_offset         Offset from the source tensor from which to start convolution
+ * @param[in] weights_addr     Pointer from where to get weights
+ * @param[in] weights_stride_y Stride of weights tesnsor in Y dimension
+ */
+inline float2 convolution_3x3_dilation_stridex2_stridey2_bifrost_f32(__global uchar *src_addr, const int stride_x_bytes, const int stride_y_bytes,
+                                                                     const int y_offset, __global uchar *weights_addr, const int weights_stride_y)
+{
+    // Load the weights
+    float3 weights_row0 = vload3(0, (__global float *)(weights_addr + 0 * weights_stride_y));
+    float3 weights_row1 = vload3(0, (__global float *)(weights_addr + 1 * weights_stride_y));
+    float3 weights_row2 = vload3(0, (__global float *)(weights_addr + 2 * weights_stride_y));
+
+    float2 pixels0 = 0.0f;
+
+    float3 src00_left  = vload3(0, (__global float *)ptr_offset(src_addr, 0, y_offset, stride_x_bytes, stride_y_bytes)); // Row0
+    float3 src00_mid   = vload3(0, (__global float *)ptr_offset(src_addr, DILATION_X, y_offset, stride_x_bytes, stride_y_bytes));
+    float3 src00_right = vload3(0, (__global float *)ptr_offset(src_addr, 2 * DILATION_X, y_offset, stride_x_bytes, stride_y_bytes));
+
+    float3 src10_left  = vload3(0, (__global float *)ptr_offset(src_addr, 0, y_offset + DILATION_Y, stride_x_bytes, stride_y_bytes)); // Row1
+    float3 src10_mid   = vload3(0, (__global float *)ptr_offset(src_addr, DILATION_X, y_offset + DILATION_Y, stride_x_bytes, stride_y_bytes));
+    float3 src10_right = vload3(0, (__global float *)ptr_offset(src_addr, 2 * DILATION_X, y_offset + DILATION_Y, stride_x_bytes, stride_y_bytes));
+
+    float3 src20_left  = vload3(0, (__global float *)ptr_offset(src_addr, 0, y_offset + DILATION_Y * 2, stride_x_bytes, stride_y_bytes)); // Row2
+    float3 src20_mid   = vload3(0, (__global float *)ptr_offset(src_addr, DILATION_X, y_offset + DILATION_Y * 2, stride_x_bytes, stride_y_bytes));
+    float3 src20_right = vload3(0, (__global float *)ptr_offset(src_addr, 2 * DILATION_X, y_offset + DILATION_Y * 2, stride_x_bytes, stride_y_bytes));
+
+    CONVOLUTION1x3_BIFROST2X1_STRIDE2(pixels0, src00_left, src00_mid, src00_right, weights_row0);
+    CONVOLUTION1x3_BIFROST2X1_STRIDE2(pixels0, src10_left, src10_mid, src10_right, weights_row1);
+    CONVOLUTION1x3_BIFROST2X1_STRIDE2(pixels0, src20_left, src20_mid, src20_right, weights_row2);
+
+    return pixels0;
+}
+
+/** Perform 3x3 convolution for stride_x=1 and stride_y=1 when DILATION_X>1 and DILATION_Y>1 for f16
+ *
+ * @param[in] src_addr         Pointer to the starting position of where to perform the convolution
+ * @param[in] src_stride_x     Stride of the source tensor in X dimension (in bytes)
+ * @param[in] src_stride_y     Stride of the source tensor in Y dimension (in bytes)
+ * @param[in] y_offset         Offset from the source tensor from which to start convolution
+ * @param[in] weights_addr     Pointer from where to get weights
+ * @param[in] weights_stride_y Stride of weights tesnsor in Y dimension
+ */
+inline half4 convolution_3x3_dilation_stridex1_stridey1_bifrost_f16(__global uchar *src_addr, const int stride_x_bytes, const int stride_y_bytes,
+                                                                    const int y_offset, __global uchar *weights_addr, const int weights_stride_y)
+{
+    // Load the weights
+    half3 weights_row0 = vload3(0, (__global half *)(weights_addr + 0 * weights_stride_y));
+    half3 weights_row1 = vload3(0, (__global half *)(weights_addr + 1 * weights_stride_y));
+    half3 weights_row2 = vload3(0, (__global half *)(weights_addr + 2 * weights_stride_y));
+
+    half4 pixels0 = 0.0f;
+
+    half4 src00_left  = vload4(0, (__global half *)ptr_offset(src_addr, 0, y_offset, stride_x_bytes, stride_y_bytes)); // Row0
+    half4 src00_mid   = vload4(0, (__global half *)ptr_offset(src_addr, DILATION_X, y_offset, stride_x_bytes, stride_y_bytes));
+    half4 src00_right = vload4(0, (__global half *)ptr_offset(src_addr, 2 * DILATION_X, y_offset, stride_x_bytes, stride_y_bytes));
+
+    half4 src10_left  = vload4(0, (__global half *)ptr_offset(src_addr, 0, y_offset + DILATION_Y, stride_x_bytes, stride_y_bytes)); // Row1
+    half4 src10_mid   = vload4(0, (__global half *)ptr_offset(src_addr, DILATION_X, y_offset + DILATION_Y, stride_x_bytes, stride_y_bytes));
+    half4 src10_right = vload4(0, (__global half *)ptr_offset(src_addr, 2 * DILATION_X, y_offset + DILATION_Y, stride_x_bytes, stride_y_bytes));
+
+    half4 src20_left  = vload4(0, (__global half *)ptr_offset(src_addr, 0, y_offset + DILATION_Y * 2, stride_x_bytes, stride_y_bytes)); // Row2
+    half4 src20_mid   = vload4(0, (__global half *)ptr_offset(src_addr, DILATION_X, y_offset + DILATION_Y * 2, stride_x_bytes, stride_y_bytes));
+    half4 src20_right = vload4(0, (__global half *)ptr_offset(src_addr, 2 * DILATION_X, y_offset + DILATION_Y * 2, stride_x_bytes, stride_y_bytes));
+
+    CONVOLUTION1x3_BIFROST4X1_STRIDE1(pixels0, src00_left, src00_mid, src00_right, weights_row0);
+    CONVOLUTION1x3_BIFROST4X1_STRIDE1(pixels0, src10_left, src10_mid, src10_right, weights_row1);
+    CONVOLUTION1x3_BIFROST4X1_STRIDE1(pixels0, src20_left, src20_mid, src20_right, weights_row2);
+
+    return pixels0;
+}
+
+/** Perform 3x3 convolution for stride_x=2 and stride_y=2 when DILATION_X>1 and DILATION_Y>1 for F16
+ *
+ * @param[in] src_addr         Pointer to the starting position of where to perform the convolution
+ * @param[in] src_stride_x     Stride of the source tensor in X dimension (in bytes)
+ * @param[in] src_stride_y     Stride of the source tensor in Y dimension (in bytes)
+ * @param[in] y_offset         Offset from the source tensor from which to start convolution
+ * @param[in] weights_addr     Pointer from where to get weights
+ * @param[in] weights_stride_y Stride of weights tesnsor in Y dimension
+ */
+inline half4 convolution_3x3_dilation_stridex2_stridey2_bifrost_f16(__global uchar *src_addr, const int stride_x_bytes, const int stride_y_bytes,
+                                                                    const int y_offset, __global uchar *weights_addr, const int weights_stride_y)
+{
+    // Load the weights
+    half3 weights_row0 = vload3(0, (__global half *)(weights_addr + 0 * weights_stride_y));
+    half3 weights_row1 = vload3(0, (__global half *)(weights_addr + 1 * weights_stride_y));
+    half3 weights_row2 = vload3(0, (__global half *)(weights_addr + 2 * weights_stride_y));
+
+    half4 pixels0 = 0.0f;
+
+    half8 src00_left  = vload8(0, (__global half *)ptr_offset(src_addr, 0, y_offset, stride_x_bytes, stride_y_bytes)); // Row0
+    half8 src00_mid   = vload8(0, (__global half *)ptr_offset(src_addr, DILATION_X, y_offset, stride_x_bytes, stride_y_bytes));
+    half8 src00_right = vload8(0, (__global half *)ptr_offset(src_addr, 2 * DILATION_X, y_offset, stride_x_bytes, stride_y_bytes));
+
+    half8 src10_left  = vload8(0, (__global half *)ptr_offset(src_addr, 0, y_offset + DILATION_Y, stride_x_bytes, stride_y_bytes)); // Row1
+    half8 src10_mid   = vload8(0, (__global half *)ptr_offset(src_addr, DILATION_X, y_offset + DILATION_Y, stride_x_bytes, stride_y_bytes));
+    half8 src10_right = vload8(0, (__global half *)ptr_offset(src_addr, 2 * DILATION_X, y_offset + DILATION_Y, stride_x_bytes, stride_y_bytes));
+
+    half8 src20_left  = vload8(0, (__global half *)ptr_offset(src_addr, 0, y_offset + DILATION_Y * 2, stride_x_bytes, stride_y_bytes)); // Row2
+    half8 src20_mid   = vload8(0, (__global half *)ptr_offset(src_addr, DILATION_X, y_offset + DILATION_Y * 2, stride_x_bytes, stride_y_bytes));
+    half8 src20_right = vload8(0, (__global half *)ptr_offset(src_addr, 2 * DILATION_X, y_offset + DILATION_Y * 2, stride_x_bytes, stride_y_bytes));
+
+    CONVOLUTION1x3_BIFROST4X1_STRIDE2(pixels0, src00_left, src00_mid, src00_right, weights_row0);
+    CONVOLUTION1x3_BIFROST4X1_STRIDE2(pixels0, src10_left, src10_mid, src10_right, weights_row1);
+    CONVOLUTION1x3_BIFROST4X1_STRIDE2(pixels0, src20_left, src20_mid, src20_right, weights_row2);
+
+    return pixels0;
+}
+
+#endif /* DILATION_X==1 && DILATION_Y==1 */
+
 /** This OpenCL kernel is optimized for Bifrost architectures and computes the depthwise convolution 3x3 when both
  * stride_x and stride_y are equal to 1
  *
@@ -326,6 +571,7 @@
     __global uchar *weights_addr = weights.ptr + get_global_id(0) * weights_step_x + get_global_id(1) * weights_step_y + channel * weights_step_z;
     __global uchar *src_addr     = src.ptr - batch * (DST_CHANNELS / DEPTH_MULTIPLIER) * (DEPTH_MULTIPLIER - 1) * src_step_z - (channel - (channel / DEPTH_MULTIPLIER)) * src_step_z;
 
+#if(DILATION_X == 1 && DILATION_Y == 1)
     // Load the weights
     float3 weights_row0 = vload3(0, (__global float *)(weights_addr + 0 * weights_stride_y));
     float3 weights_row1 = vload3(0, (__global float *)(weights_addr + 1 * weights_stride_y));
@@ -352,6 +598,19 @@
     CONVOLUTION1x3_BIFROST2X1_STRIDE1(pixels3, src40, weights_row1);
     CONVOLUTION1x3_BIFROST2X1_STRIDE1(pixels3, src50, weights_row2);
 
+#else /* DILATION_X==1 && DILATION_Y==1 */
+
+    //3x3 Convolution of elements starting in 0th row
+    pixels0 = convolution_3x3_dilation_stridex1_stridey1_bifrost_f32(src_addr, src.stride_x, src.stride_y, 0, weights_addr, weights_stride_y);
+    //3x3 Convolution of elements starting in 1st row
+    pixels1 = convolution_3x3_dilation_stridex1_stridey1_bifrost_f32(src_addr, src.stride_x, src.stride_y, 1, weights_addr, weights_stride_y);
+    //3x3 Convolution of elements starting in 2nd row
+    pixels2 = convolution_3x3_dilation_stridex1_stridey1_bifrost_f32(src_addr, src.stride_x, src.stride_y, 2, weights_addr, weights_stride_y);
+    //3x3 Convolution of elements starting in 3rd row
+    pixels3 = convolution_3x3_dilation_stridex1_stridey1_bifrost_f32(src_addr, src.stride_x, src.stride_y, 3, weights_addr, weights_stride_y);
+
+#endif /* DILATION_X==1 && DILATION_Y==1 */
+
 #ifdef HAS_BIAS
     Vector biases = CONVERT_TO_VECTOR_STRUCT_NO_STEP(biases);
 
@@ -425,6 +684,8 @@
     __global uchar *weights_addr = weights.ptr + get_global_id(0) * weights_step_x + get_global_id(1) * weights_step_y + channel * weights_step_z;
     __global uchar *src_addr     = src.ptr - batch * (DST_CHANNELS / DEPTH_MULTIPLIER) * (DEPTH_MULTIPLIER - 1) * src_step_z - (channel - (channel / DEPTH_MULTIPLIER)) * src_step_z;
 
+#if(DILATION_X == 1 && DILATION_Y == 1)
+
     // Load the weights
     float3 weights_row0 = vload3(0, (__global float *)(weights_addr + 0 * weights_stride_y));
     float3 weights_row1 = vload3(0, (__global float *)(weights_addr + 1 * weights_stride_y));
@@ -449,6 +710,14 @@
     CONVOLUTION1x3_BIFROST2X1_STRIDE2(pixels1, src30, src31, weights_row1);
     CONVOLUTION1x3_BIFROST2X1_STRIDE2(pixels1, src40, src41, weights_row2);
 
+#else  /* DILATION_X==1 && DILATION_Y==1 */
+
+    //3x3 Convolution of elements starting in 0th row
+    pixels0 = convolution_3x3_dilation_stridex2_stridey2_bifrost_f32(src_addr, src.stride_x, src.stride_y, 0, weights_addr, weights_stride_y);
+    //3x3 Convolution of elements starting in 2nd row
+    pixels1 = convolution_3x3_dilation_stridex2_stridey2_bifrost_f32(src_addr, src.stride_x, src.stride_y, 2, weights_addr, weights_stride_y);
+#endif /* DILATION_X==1 && DILATION_Y==1 */
+
 #ifdef HAS_BIAS
     Vector biases = CONVERT_TO_VECTOR_STRUCT_NO_STEP(biases);
 
@@ -632,11 +901,12 @@
 }
 #endif //defined(SRC_WIDTH) && defined(DATA_TYPE)
 
-#if defined(STRIDE_X) && defined(STRIDE_Y) && defined(PAD_LEFT) && defined(PAD_TOP) && defined(PAD_RIGHT) && defined(PAD_BOTTOM) && defined(KERNEL_WIDTH) && defined(KERNEL_HEIGHT) && defined(SRC_WIDTH) && defined(SRC_HEIGHT) && defined(DATA_TYPE) && defined(PAD_VALUE) && defined(DEPTH_MULTIPLIER)
+#if defined(STRIDE_X) && defined(STRIDE_Y) && defined(PAD_LEFT) && defined(PAD_TOP) && defined(PAD_RIGHT) && defined(PAD_BOTTOM) && defined(KERNEL_WIDTH) && defined(KERNEL_HEIGHT) && defined(SRC_WIDTH) && defined(SRC_HEIGHT) && defined(DATA_TYPE) && defined(PAD_VALUE) && defined(DEPTH_MULTIPLIER) && defined(DILATION_X) && defined(DILATION_Y)
 /** This kernel performs a reshaping of the input tensor to a tensor used to perform depthwise convolution using vector to matrix multiplication.
  *
  * @note The data type must be passed at compile time using -DDATA_TYPE: e.g. -DDATA_TYPE=float
  * @note The convolution information must be passed at compile time using -DSTRIDE_X, -DSTRIDE_Y, -DPAD_LEFT, -DPAD_TOP, -DPAD_RIGHT, -DPAD_BOTTOM, -DKERNEL_WIDHT, -DKERNEL_HEIGHT, -DSRC_WIDTH, -DSRC_HEIGHT, -DDEPTH_MULTIPLIER
+ * @note The dilation_x and dilation_y must be passed at compile time using -DDILATION_X and -DDILATION_Y: e.g. -DDILATION_X=1, -DDILATION_Y=1
  *
  * @param[in]  src_ptr                           Pointer to the source tensor. Supported data types: F16/F32
  * @param[in]  src_stride_x                      Stride of the source tensor in X dimension (in bytes)
@@ -661,7 +931,7 @@
 
     const int src_pixel_linear = get_global_id(1) * STRIDE_X;
     const int full_length      = SRC_WIDTH + PAD_LEFT + PAD_RIGHT;
-    const int max_initial_x    = STRIDE_X * (((full_length - KERNEL_WIDTH) / STRIDE_X) + 1);
+    const int max_initial_x    = STRIDE_X * (((full_length - (KERNEL_WIDTH + (KERNEL_WIDTH - 1) * (DILATION_X - 1))) / STRIDE_X) + 1);
 
     const int src_x = -PAD_LEFT + src_pixel_linear % max_initial_x;
     const int src_y = -PAD_TOP + src_pixel_linear / max_initial_x * STRIDE_Y;
@@ -670,9 +940,9 @@
     __global uchar *input_ptr      = src_ptr + src_offset_first_element_in_bytes + src_z * in_stride_z;
     __global DATA_TYPE *output_ptr = ((__global DATA_TYPE *)(dst.ptr));
 
-    for(int y = src_y; y < src_y + KERNEL_HEIGHT; ++y)
+    for(int y = src_y; y < src_y + KERNEL_HEIGHT + (KERNEL_HEIGHT - 1) * (DILATION_Y - 1); y += DILATION_Y)
     {
-        for(int x = src_x; x < src_x + KERNEL_WIDTH; ++x, ++output_ptr)
+        for(int x = src_x; x < src_x + KERNEL_WIDTH + (KERNEL_WIDTH - 1) * (DILATION_X - 1); x += DILATION_X, ++output_ptr)
         {
             if(x < 0 || x >= SRC_WIDTH || y < 0 || y >= SRC_HEIGHT)
             {
@@ -754,6 +1024,8 @@
                                          const half            middle_coeff,
                                          const half            right_coeff)
 {
+#if(DILATION_X == 1 && DILATION_Y == 1)
+
     half8 temp = vload8(0, (__global half *)left_pixel);
 
     half4 left   = CONVERT(temp.s0123, half4);
@@ -761,6 +1033,12 @@
     half4 right  = CONVERT(temp.s2345, half4);
 
     return left * (half4)left_coeff + middle * (half4)middle_coeff + right * (half4)right_coeff;
+#else /* DILATION_X==1 && DILATION_Y==1 */
+    return vload4(0, (__global half *)left_pixel) * (half4)left_coeff
+           + vload4(0, (__global half *)(left_pixel) + DILATION_X) * (half4)middle_coeff
+           + vload4(0, (__global half *)(left_pixel) + 2 * DILATION_X) * (half4)right_coeff;
+
+#endif /* DILATION_X==1 && DILATION_Y==1 */
 }
 
 /** Compute a 1D horizontal convolution of size 3 and stride 2 for 16bit floating point type.
@@ -777,6 +1055,8 @@
                                          const half            middle_coeff,
                                          const half            right_coeff)
 {
+#if(DILATION_X == 1 && DILATION_Y == 1)
+
     half8 temp0 = vload8(0, (__global half *)left_pixel);
     half temp1  = *((__global half *)(left_pixel + 8 * sizeof(half)));
 
@@ -785,6 +1065,15 @@
     half4 right  = CONVERT((half4)(temp0.s246, temp1), half4);
 
     return left * (half4)left_coeff + middle * (half4)middle_coeff + right * (half4)right_coeff;
+#else /* DILATION_X==1 && DILATION_Y==1 */
+
+    __global half *left_pixel_float = (__global half *)left_pixel;
+
+    return (half4)(*left_pixel_float, *(left_pixel_float + 2), *(left_pixel_float + 4), *(left_pixel_float + 6)) * (half4)left_coeff
+           + (half4)(*(left_pixel_float + DILATION_X), *(left_pixel_float + DILATION_X + 2), *(left_pixel_float + DILATION_X + 4), *(left_pixel_float + DILATION_X + 6)) * (half4)middle_coeff
+           + (half4)(*(left_pixel_float + DILATION_X * 2), *(left_pixel_float + DILATION_X * 2 + 2), *(left_pixel_float + DILATION_X * 2 + 4), *(left_pixel_float + DILATION_X * 2 + 6)) * (half4)right_coeff;
+
+#endif /* DILATION_X==1 && DILATION_Y==1 */
 }
 
 /** Compute a 1D horizontal convolution of size 3 and stride 3 for 16bit floating point type.
@@ -801,6 +1090,8 @@
                                          const half            middle_coeff,
                                          const half            right_coeff)
 {
+#if(DILATION_X == 1 && DILATION_Y == 1)
+
     half16 temp0 = vload16(0, (__global half *)left_pixel);
 
     half4 left   = CONVERT(temp0.s0369, half4);
@@ -808,6 +1099,15 @@
     half4 right  = CONVERT(temp0.s258B, half4);
 
     return left * (half4)left_coeff + middle * (half4)middle_coeff + right * (half4)right_coeff;
+#else /* DILATION_X==1 && DILATION_Y==1 */
+
+    __global half *left_pixel_float = (__global half *)left_pixel;
+
+    return (half4)(*left_pixel_float, *(left_pixel_float + 3), *(left_pixel_float + 6), *(left_pixel_float + 9)) * (half4)left_coeff
+           + (half4)(*(left_pixel_float + DILATION_X), *(left_pixel_float + DILATION_X + 3), *(left_pixel_float + DILATION_X + 6), *(left_pixel_float + DILATION_X + 9)) * (half4)middle_coeff
+           + (half4)(*(left_pixel_float + DILATION_X * 2), *(left_pixel_float + DILATION_X * 2 + 3), *(left_pixel_float + DILATION_X * 2 + 6), *(left_pixel_float + DILATION_X * 2 + 9)) * (half4)right_coeff;
+
+#endif /* DILATION_X==1 && DILATION_Y==1 */
 }
 
 /** Apply a 3x3 convolution matrix to a single channel F16 input image and return the result.
@@ -841,8 +1141,8 @@
     half4 pixels;
 
     pixels = convolution1x3_f16(offset(src, 0, 0), mat0, mat1, mat2);
-    pixels += convolution1x3_f16(offset(src, 0, 1), mat3, mat4, mat5);
-    pixels += convolution1x3_f16(offset(src, 0, 2), mat6, mat7, mat8);
+    pixels += convolution1x3_f16(offset(src, 0, DILATION_Y), mat3, mat4, mat5);
+    pixels += convolution1x3_f16(offset(src, 0, DILATION_Y * 2), mat6, mat7, mat8);
 
     return pixels;
 }
@@ -986,6 +1286,7 @@
     __global uchar *weights_addr = weights.ptr + get_global_id(0) * weights_step_x + get_global_id(1) * weights_step_y + channel * weights_step_z;
     __global uchar *src_addr     = src.ptr - batch * (DST_CHANNELS / DEPTH_MULTIPLIER) * (DEPTH_MULTIPLIER - 1) * src_step_z - (channel - (channel / DEPTH_MULTIPLIER)) * src_step_z;
 
+#if(DILATION_X == 1 && DILATION_Y == 1)
     // Load the weights
     half3 weights_row0 = vload3(0, (__global half *)(weights_addr + 0 * weights_stride_y));
     half3 weights_row1 = vload3(0, (__global half *)(weights_addr + 1 * weights_stride_y));
@@ -1012,6 +1313,19 @@
     CONVOLUTION1x3_BIFROST4X1_STRIDE1(pixels3, src40, weights_row1);
     CONVOLUTION1x3_BIFROST4X1_STRIDE1(pixels3, src50, weights_row2);
 
+#else /* DILATION_X==1 && DILATION_Y==1 */
+
+    //3x3 Convolution of elements starting in 0th row
+    pixels0 = convolution_3x3_dilation_stridex1_stridey1_bifrost_f16(src_addr, src.stride_x, src.stride_y, 0, weights_addr, weights_stride_y);
+    //3x3 Convolution of elements starting in 1st row
+    pixels1 = convolution_3x3_dilation_stridex1_stridey1_bifrost_f16(src_addr, src.stride_x, src.stride_y, 1, weights_addr, weights_stride_y);
+    //3x3 Convolution of elements starting in 2nd row
+    pixels2 = convolution_3x3_dilation_stridex1_stridey1_bifrost_f16(src_addr, src.stride_x, src.stride_y, 2, weights_addr, weights_stride_y);
+    //3x3 Convolution of elements starting in 3rd row
+    pixels3 = convolution_3x3_dilation_stridex1_stridey1_bifrost_f16(src_addr, src.stride_x, src.stride_y, 3, weights_addr, weights_stride_y);
+
+#endif /* DILATION_X==1 && DILATION_Y==1 */
+
 #ifdef HAS_BIAS
     pixels0 += (half4)bias;
     pixels1 += (half4)bias;
@@ -1088,6 +1402,8 @@
     __global uchar *weights_addr = weights.ptr + get_global_id(0) * weights_step_x + get_global_id(1) * weights_step_y + channel * weights_step_z;
     __global uchar *src_addr     = src.ptr - batch * (DST_CHANNELS / DEPTH_MULTIPLIER) * (DEPTH_MULTIPLIER - 1) * src_step_z - (channel - (channel / DEPTH_MULTIPLIER)) * src_step_z;
 
+#if(DILATION_X == 1 && DILATION_Y == 1)
+
     // Load the weights
     half3 weights_row0 = vload3(0, (__global half *)(weights_addr + 0 * weights_stride_y));
     half3 weights_row1 = vload3(0, (__global half *)(weights_addr + 1 * weights_stride_y));
@@ -1112,6 +1428,13 @@
     CONVOLUTION1x3_BIFROST4X1_STRIDE2(pixels1, src30, src31, weights_row1);
     CONVOLUTION1x3_BIFROST4X1_STRIDE2(pixels1, src40, src41, weights_row2);
 
+#else  /* DILATION_X==1 && DILATION_Y==1 */
+    //3x3 Convolution of elements starting in 0th row
+    pixels0 = convolution_3x3_dilation_stridex2_stridey2_bifrost_f16(src_addr, src.stride_x, src.stride_y, 0, weights_addr, weights_stride_y);
+    //3x3 Convolution of elements starting in 2nd row
+    pixels1 = convolution_3x3_dilation_stridex2_stridey2_bifrost_f16(src_addr, src.stride_x, src.stride_y, 2, weights_addr, weights_stride_y);
+#endif /* DILATION_X==1 && DILATION_Y==1 */
+
 #ifdef HAS_BIAS
     pixels0 += (half4)bias;
     pixels1 += (half4)bias;
@@ -1189,9 +1512,9 @@
 #if defined(DST_DEPTH)
     int z = get_global_id(2) % (int)DST_DEPTH; // spatial coordinate y
     int b = get_global_id(2) / (int)DST_DEPTH; // batch
-#else // defined(DST_DEPTH)
-    int z = get_global_id(2); // spatial coordinate y
-#endif // defined(DST_DEPTH)
+#else                                          // defined(DST_DEPTH)
+    int      z               = get_global_id(2); // spatial coordinate y
+#endif                                         // defined(DST_DEPTH)
 
     Vector weights = CONVERT_TO_VECTOR_STRUCT(weights);
 
@@ -1203,7 +1526,7 @@
 
     int  z_coord  = 0;
     int4 offset   = 0;
-    int4 y_offset = ((int4)(y * CONV_STRIDE_X) + (int4)(0, 1, 2, 3) - CONV_PAD_LEFT) * (int4)src_stride_y;
+    int4 y_offset = ((int4)(y * CONV_STRIDE_X) + (int4)(0, DILATION_X * 1, DILATION_X * 2, DILATION_X * 3) - CONV_PAD_LEFT) * (int4)src_stride_y;
 
     // We compute 2x1x1 [C,W,H] elements
     VEC_FLOAT acc = 0;
@@ -1236,16 +1559,16 @@
     // z == 1
     // z_coord can be only negative for z = 0 so we do not need to clamp it
     // Moreover z_coord cannot be out-of-bound for z = 1 so we do not need to clamp the offset
-    z_coord           = z * CONV_STRIDE_Y - (int)CONV_PAD_TOP + 1;
+    z_coord           = z * CONV_STRIDE_Y - (int)CONV_PAD_TOP + DILATION_Y;
     offset            = y_offset + (int4)(z_coord * src_stride_z);
     VEC_FLOAT values3 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s0));
     VEC_FLOAT values4 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s1));
     VEC_FLOAT values5 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s2));
 
     // z == 2
-    // After z = 1 we can simply add src_stride_z to offset without updating z_coord
-    // However offset can be out-of-bound so we need to check if it is greater than max_offset
-    offset += (int4)src_stride_z;
+    // Offset can be out-of-bound so we need to check if it is greater than max_offset
+    z_coord           = z * CONV_STRIDE_Y - (int)CONV_PAD_TOP + DILATION_Y * 2;
+    offset            = y_offset + (int4)(z_coord * src_stride_z);
     offset            = min(offset, (int4)max_offset);
     VEC_FLOAT values6 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s0));
     VEC_FLOAT values7 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s1));
@@ -1338,9 +1661,9 @@
 #if defined(DST_DEPTH)
     int z = get_global_id(2) % (int)DST_DEPTH; // spatial coordinate y
     int b = get_global_id(2) / (int)DST_DEPTH; // batch
-#else // defined(DST_DEPTH)
-    int z = get_global_id(2); // spatial coordinate y
-#endif // defined(DST_DEPTH)
+#else                                          // defined(DST_DEPTH)
+    int             z        = get_global_id(2); // spatial coordinate y
+#endif                                         // defined(DST_DEPTH)
 
     Vector weights = CONVERT_TO_VECTOR_STRUCT(weights);
 
diff --git a/src/core/CL/cl_kernels/depthwise_convolution_quantized.cl b/src/core/CL/cl_kernels/depthwise_convolution_quantized.cl
index 503aa7e..8d145a0 100644
--- a/src/core/CL/cl_kernels/depthwise_convolution_quantized.cl
+++ b/src/core/CL/cl_kernels/depthwise_convolution_quantized.cl
@@ -53,6 +53,8 @@
 
 #if !(defined(ARM_COMPUTE_OPENCL_DOT8_ENABLED) && defined(cl_arm_integer_dot_product_int8))
 
+#if DILATION_X == 1
+
 #if CONV_STRIDE_X == 1
 #define GET_VALUES(first_value, left, middle, right)                              \
     ({                                                                            \
@@ -85,6 +87,46 @@
     })
 #endif /* CONV_STRIDE_X */
 
+#else /* DILATION_X == 1 */
+
+#if CONV_STRIDE_X == 1
+#define GET_VALUES(first_value, left, middle, right)                                     \
+    ({                                                                                   \
+        left   = CONVERT(vload8(0, first_value), int8);                                  \
+        middle = CONVERT(vload8(0, first_value + DILATION_X * sizeof(uchar)), int8);     \
+        right  = CONVERT(vload8(0, first_value + 2 * DILATION_X * sizeof(uchar)), int8); \
+    })
+#elif CONV_STRIDE_X == 2
+#define GET_VALUES(first_value, left, middle, right)                                      \
+    ({                                                                                    \
+        int16 temp0 = CONVERT(vload16(0, first_value), int16);                            \
+        left        = CONVERT(temp0.s02468ace, int8);                                     \
+        \
+        temp0  = CONVERT(vload16(0, first_value + DILATION_X * sizeof(uchar)), int16);    \
+        middle = CONVERT(temp0.s02468ace, int8);                                          \
+        \
+        temp0 = CONVERT(vload16(0, first_value + 2 * DILATION_X * sizeof(uchar)), int16); \
+        right = CONVERT(temp0.s02468ace, int8);                                           \
+    })
+#else /* CONV_STRIDE_X */
+#define GET_VALUES(first_value, left, middle, right)                                             \
+    ({                                                                                           \
+        int16 temp0 = CONVERT(vload16(0, first_value), int16);                                   \
+        int8  temp1 = CONVERT(vload8(0, (first_value + 16 * sizeof(uchar))), int8);              \
+        left        = CONVERT((int8)(temp0.s0369, temp0.scf, temp1.s25), int8);                  \
+        \
+        temp0  = CONVERT(vload16(0, first_value + DILATION_X * sizeof(uchar)), int16);           \
+        temp1  = CONVERT(vload8(0, (first_value + (16 + DILATION_X) * sizeof(uchar))), int8);    \
+        middle = CONVERT((int8)(temp0.s0369, temp0.scf, temp1.s25), int8);                       \
+        \
+        temp0 = CONVERT(vload16(0, first_value + 2 * DILATION_X * sizeof(uchar)), int16);        \
+        temp1 = CONVERT(vload8(0, (first_value + (16 + 2 * DILATION_X) * sizeof(uchar))), int8); \
+        right = CONVERT((int8)(temp0.s0369, temp0.scf, temp1.s25), int8);                        \
+    })
+
+#endif /* CONV_STRIDE_X */
+#endif /* DILATION_X==1 */
+
 /** This function computes the depthwise convolution quantized.
  *
  * @param[in] src_ptr                               Pointer to the source tensor. Supported data types: QASYMM8
@@ -151,10 +193,10 @@
 
     int8 values0 = 0;
     int8 sum0    = 0;
-#if CONV_STRIDE_Y == 1
+#if CONV_STRIDE_Y == 1 && DILATION_Y == 1
     int8 values1 = 0;
     int8 sum1    = 0;
-#endif /* CONV_STRIDE_Y */
+#endif /* CONV_STRIDE_Y &&DILATION_Y==1 */
 
     // Row0
     int8 left, middle, right;
@@ -168,44 +210,44 @@
 #endif /* WEIGHTS_OFFSET != 0 */
 
     // Row1
-    GET_VALUES(src.ptr + 1 * src_stride_y, left, middle, right);
+    GET_VALUES(src.ptr + DILATION_Y * src_stride_y, left, middle, right);
     values0 += left * (int8)(w1.s0);
     values0 += middle * (int8)(w1.s1);
     values0 += right * (int8)(w1.s2);
-#if CONV_STRIDE_Y == 1
+#if CONV_STRIDE_Y == 1 && DILATION_Y == 1
     values1 += left * (int8)(w0.s0);
     values1 += middle * (int8)(w0.s1);
     values1 += right * (int8)(w0.s2);
-#endif /* CONV_STRIDE_Y == 1 */
+#endif /* CONV_STRIDE_Y && DILATION_Y== 1 */
 
 #if WEIGHTS_OFFSET != 0
     int8 tmp = left + middle + right;
     sum0 += tmp;
-#if CONV_STRIDE_Y == 1
+#if CONV_STRIDE_Y == 1 && DILATION_Y == 1
     sum1 += tmp;
-#endif /* CONV_STRIDE_Y == 1 */
+#endif /* CONV_STRIDE_Y &&DILATION_Y== 1 */
 #endif /* WEIGHTS_OFFSET != 0 */
 
     // Row2
-    GET_VALUES(src.ptr + 2 * src_stride_y, left, middle, right);
+    GET_VALUES(src.ptr + 2 * DILATION_Y * src_stride_y, left, middle, right);
     values0 += left * (int8)(w2.s0);
     values0 += middle * (int8)(w2.s1);
     values0 += right * (int8)(w2.s2);
-#if CONV_STRIDE_Y == 1
+#if CONV_STRIDE_Y == 1 && DILATION_Y == 1
     values1 += left * (int8)(w1.s0);
     values1 += middle * (int8)(w1.s1);
     values1 += right * (int8)(w1.s2);
-#endif /* CONV_STRIDE_Y == 1 */
+#endif /* CONV_STRIDE_Y &&DILATION_Y == 1 */
 
 #if WEIGHTS_OFFSET != 0
     tmp = left + middle + right;
     sum0 += tmp;
-#if CONV_STRIDE_Y == 1
+#if CONV_STRIDE_Y == 1 && DILATION_Y == 1
     sum1 += tmp;
-#endif /* CONV_STRIDE_Y == 1 */
+#endif /* CONV_STRIDE_Y == 1 && DILATION_Y==1 */
 #endif /* WEIGHTS_OFFSET != 0 */
 
-#if CONV_STRIDE_Y == 1
+#if CONV_STRIDE_Y == 1 && DILATION_Y == 1
     // Row3
     GET_VALUES(src.ptr + 3 * src_stride_y, left, middle, right);
     values1 += left * (int8)(w2.s0);
@@ -215,20 +257,20 @@
 #if WEIGHTS_OFFSET != 0
     sum1 += left + middle + right;
 #endif /* WEIGHTS_OFFSET != 0 */
-#endif /* CONV_STRIDE_Y == 1 */
+#endif /* CONV_STRIDE_Y && DILATION_Y == 1 */
 
 #if defined(HAS_BIAS)
     values0 += (int8)(bias_value);
-#if CONV_STRIDE_Y == 1
+#if CONV_STRIDE_Y == 1 && DILATION_Y == 1
     values1 += (int8)(bias_value);
-#endif /* CONV_STRIDE_Y == 1 */
+#endif /* CONV_STRIDE_Y & &DILATION_Y == 1 */
 #endif //defined(HAS_BIAS)
 
 #if WEIGHTS_OFFSET != 0
     values0 += sum0 * (int8)(WEIGHTS_OFFSET);
-#if CONV_STRIDE_Y == 1
+#if CONV_STRIDE_Y == 1 && DILATION_Y == 1
     values1 += sum1 * (int8)(WEIGHTS_OFFSET);
-#endif /* CONV_STRIDE_Y == 1 */
+#endif /* CONV_STRIDE_Y == 1 && DILATION_Y==1 */
 #endif /* WEIGHTS_OFFSET != 0 */
 
 #if INPUT_OFFSET != 0
@@ -236,16 +278,16 @@
     ushort3 tmp_we      = convert_ushort3(w0) + convert_ushort3(w1) + convert_ushort3(w2);
     sum_weights += tmp_we.s0 + tmp_we.s1 + tmp_we.s2;
     values0 += sum_weights * (int8)(INPUT_OFFSET);
-#if CONV_STRIDE_Y == 1
+#if CONV_STRIDE_Y == 1 && DILATION_Y == 1
     values1 += sum_weights * (int8)(INPUT_OFFSET);
-#endif /* CONV_STRIDE_Y == 1 */
+#endif /* CONV_STRIDE_Y == 1 && DILATION_Y==1 */
 #endif /* INPUT_OFFSET != 0 */
 
 #if K_OFFSET != 0
     values0 += (int8)(K_OFFSET);
-#if CONV_STRIDE_Y == 1
+#if CONV_STRIDE_Y == 1 && DILATION_Y == 1
     values1 += (int8)(K_OFFSET);
-#endif /* CONV_STRIDE_Y == 1 */
+#endif /* CONV_STRIDE_Y == 1 && DILATION_Y==1*/
 #endif /* K_OFFSET != 0 */
 
 #if defined(REAL_MULTIPLIER)
@@ -264,7 +306,7 @@
     res0        = min(res0, (uchar8)255);
 
     vstore8(ACTIVATION_FUNC(res0), 0, dst.ptr);
-#if CONV_STRIDE_Y == 1
+#if CONV_STRIDE_Y == 1 && DILATION_Y == 1
 #if defined(REAL_MULTIPLIER)
 
     values1 = CONVERT(round(CONVERT(values1, float8) * (float8)REAL_MULTIPLIER), int8);
@@ -281,11 +323,11 @@
     res1        = min(res1, (uchar8)255);
 
     vstore8(ACTIVATION_FUNC(res1), 0, dst.ptr + dst_stride_y);
-#endif /* CONV_STRIDE_Y == 1 */
+#endif /* CONV_STRIDE_Y == 1 && DILATION_Y==1*/
 }
 
 #else // !(defined(ARM_COMPUTE_OPENCL_DOT8_ENABLED) && defined(cl_arm_integer_dot_product_int8))
-
+#if DILATION_X == 1
 #if CONV_STRIDE_X == 1
 #define GET_VALUES(first_value, left, middle, right)                 \
     ({                                                               \
@@ -317,6 +359,43 @@
         right  = (uchar8)(temp0.s258b, temp0.se, temp1.s147);          \
     })
 #endif /* CONV_STRIDE_X */
+#else  /*DILATION_X==1*/
+
+#if CONV_STRIDE_X == 1
+#define GET_VALUES(first_value, left, middle, right)                      \
+    ({                                                                    \
+        left   = vload8(0, first_value);                                  \
+        middle = vload8(0, first_value + DILATION_X * sizeof(uchar));     \
+        right  = vload8(0, first_value + 2 * DILATION_X * sizeof(uchar)); \
+    })
+#elif CONV_STRIDE_X == 2
+#define GET_VALUES(first_value, left, middle, right)                              \
+    ({                                                                            \
+        uchar16 temp0 = vload16(0, first_value);                                  \
+        left          = temp0.s02468ace;                                          \
+        temp0         = vload16(0, first_value + DILATION_X * sizeof(uchar));     \
+        middle        = temp0.s02468ace;                                          \
+        temp0         = vload16(0, first_value + 2 * DILATION_X * sizeof(uchar)); \
+        right         = temp0.s02468ace;                                          \
+    })
+#else /* CONV_STRIDE_X */
+#define GET_VALUES(first_value, left, middle, right)                              \
+    ({                                                                            \
+        uchar16 temp0 = vload16(0, first_value);                                  \
+        uchar8  temp1 = vload8(0, (first_value + 16 * sizeof(uchar)));            \
+        left          = (uchar8)(temp0.s0369, temp0.scf, temp1.s25);              \
+        \
+        temp0  = vload16(0, first_value + DILATION_X * sizeof(uchar));            \
+        temp1  = vload8(0, (first_value + (16 + DILATION_X) * sizeof(uchar)));    \
+        middle = (uchar8)(temp0.s0369, temp0.scf, temp1.s25);                     \
+        \
+        temp0 = vload16(0, first_value + 2 * DILATION_X * sizeof(uchar));         \
+        temp1 = vload8(0, (first_value + (16 + 2 * DILATION_X) * sizeof(uchar))); \
+        right = (uchar8)(temp0.s0369, temp0.scf, temp1.s25);                      \
+    })
+
+#endif /* CONV_STRIDE_X */
+#endif /*DILATION_X==1*/
 /** This function computes the depthwise convolution quantized using dot product when the data layout is NCHW.
  *
  * @param[in] src_ptr                               Pointer to the source tensor. Supported data types: QASYMM8
@@ -389,8 +468,8 @@
     int8 sum0    = 0;
 
     GET_VALUES(src.ptr + 0 * src_stride_y, left0, middle0, right0);
-    GET_VALUES(src.ptr + 1 * src_stride_y, left1, middle1, right1);
-    GET_VALUES(src.ptr + 2 * src_stride_y, left2, middle2, right2);
+    GET_VALUES(src.ptr + DILATION_Y * src_stride_y, left1, middle1, right1);
+    GET_VALUES(src.ptr + 2 * DILATION_Y * src_stride_y, left2, middle2, right2);
 
 #if WEIGHTS_OFFSET != 0
     sum0 += convert_int8(left0) + convert_int8(middle0) + convert_int8(right0);
@@ -398,7 +477,7 @@
     sum0 += convert_int8(left2) + convert_int8(middle2) + convert_int8(right2);
 #endif /* WEIGHTS_OFFSET != 0 */
 
-#if CONV_STRIDE_Y == 1
+#if CONV_STRIDE_Y == 1 && DILATION_Y == 1
     // If conv_stride_y is equals to 1, we compute two output rows
 
     uchar8 left3, middle3, right3;
@@ -412,7 +491,7 @@
     sum1 += convert_int8(left2) + convert_int8(middle2) + convert_int8(right2);
     sum1 += convert_int8(left3) + convert_int8(middle3) + convert_int8(right3);
 #endif /* WEIGHTS_OFFSET != 0 */
-#endif // CONV_STRIDE_Y == 1
+#endif // CONV_STRIDE_Y == 1 && DILATION_Y==1
 
     ARM_DOT((uchar4)(left0.s0, middle0.s0, right0.s0, left1.s0), (uchar4)(w0.s0, w0.s1, w0.s2, w1.s0), values0.s0);
     ARM_DOT((uchar4)(middle1.s0, right1.s0, left2.s0, middle2.s0), (uchar4)(w1.s1, w1.s2, w2.s0, w2.s1), values0.s0);
@@ -446,7 +525,7 @@
     ARM_DOT((uchar4)(middle1.s7, right1.s7, left2.s7, middle2.s7), (uchar4)(w1.s1, w1.s2, w2.s0, w2.s1), values0.s7);
     values0.s7 += right2.s7 * w2.s2;
 
-#if CONV_STRIDE_Y == 1
+#if CONV_STRIDE_Y == 1 && DILATION_Y == 1
     ARM_DOT((uchar4)(left1.s0, middle1.s0, right1.s0, left2.s0), (uchar4)(w0.s0, w0.s1, w0.s2, w1.s0), values1.s0);
     ARM_DOT((uchar4)(middle2.s0, right2.s0, left3.s0, middle3.s0), (uchar4)(w1.s1, w1.s2, w2.s0, w2.s1), values1.s0);
     values1.s0 += right3.s0 * w2.s2;
@@ -478,20 +557,20 @@
     ARM_DOT((uchar4)(left1.s7, middle1.s7, right1.s7, left2.s7), (uchar4)(w0.s0, w0.s1, w0.s2, w1.s0), values1.s7);
     ARM_DOT((uchar4)(middle2.s7, right2.s7, left3.s7, middle3.s7), (uchar4)(w1.s1, w1.s2, w2.s0, w2.s1), values1.s7);
     values1.s7 += right3.s7 * w2.s2;
-#endif // CONV_STRIDE_Y == 1
+#endif // CONV_STRIDE_Y == 1 && DILATION_Y==1
 
 #if defined(HAS_BIAS)
     values0 += (int8)(bias_value);
-#if CONV_STRIDE_Y == 1
+#if CONV_STRIDE_Y == 1 && DILATION_Y == 1
     values1 += (int8)(bias_value);
-#endif /* CONV_STRIDE_Y == 1 */
+#endif /* CONV_STRIDE_Y == 1 && DILATION_Y==1 */
 #endif //defined(HAS_BIAS)
 
 #if WEIGHTS_OFFSET != 0
     values0 += sum0 * (int8)(WEIGHTS_OFFSET);
-#if CONV_STRIDE_Y == 1
+#if CONV_STRIDE_Y == 1 && DILATION_Y == 1
     values1 += sum1 * (int8)(WEIGHTS_OFFSET);
-#endif /* CONV_STRIDE_Y == 1 */
+#endif /* CONV_STRIDE_Y == 1 && DILATION_Y==1 */
 #endif /* WEIGHTS_OFFSET != 0 */
 
 #if INPUT_OFFSET != 0
@@ -499,16 +578,16 @@
     ushort3 tmp_we      = convert_ushort3(w0) + convert_ushort3(w1) + convert_ushort3(w2);
     sum_weights += tmp_we.s0 + tmp_we.s1 + tmp_we.s2;
     values0 += sum_weights * (int8)(INPUT_OFFSET);
-#if CONV_STRIDE_Y == 1
+#if CONV_STRIDE_Y == 1 && DILATION_Y == 1
     values1 += sum_weights * (int8)(INPUT_OFFSET);
-#endif /* CONV_STRIDE_Y == 1 */
+#endif /* CONV_STRIDE_Y == 1 && DILATION_Y==1*/
 #endif /* INPUT_OFFSET != 0 */
 
 #if K_OFFSET != 0
     values0 += (int8)(K_OFFSET);
-#if CONV_STRIDE_Y == 1
+#if CONV_STRIDE_Y == 1 && DILATION_Y == 1
     values1 += (int8)(K_OFFSET);
-#endif /* CONV_STRIDE_Y == 1 */
+#endif /* CONV_STRIDE_Y == 1 && DILATION_Y==1*/
 #endif /* K_OFFSET != 0 */
 
 #if defined(REAL_MULTIPLIER)
@@ -527,7 +606,7 @@
     res0        = min(res0, (uchar8)255);
 
     vstore8(ACTIVATION_FUNC(res0), 0, dst.ptr);
-#if CONV_STRIDE_Y == 1
+#if CONV_STRIDE_Y == 1 && DILATION_Y == 1
 
 #if defined(REAL_MULTIPLIER)
 
@@ -545,7 +624,7 @@
     res1        = min(res1, (uchar8)255);
 
     vstore8(ACTIVATION_FUNC(res1), 0, dst.ptr + dst_stride_y);
-#endif /* CONV_STRIDE_Y == 1 */
+#endif /* CONV_STRIDE_Y == 1 && DILATION_Y==1*/
 }
 
 #endif // defined(ARM_COMPUTE_OPENCL_DOT8_ENABLED) && defined(cl_arm_integer_dot_product_int8)
@@ -669,7 +748,7 @@
 
     int  z_coord = 0;
     int4 offset  = 0;
-    int4 y_coord = ((int4)(y * CONV_STRIDE_X) + (int4)(0, 1, 2, 3)) - (int)CONV_PAD_LEFT;
+    int4 y_coord = ((int4)(y * CONV_STRIDE_X) + (int4)(0, DILATION_X * 1, DILATION_X * 2, DILATION_X * 3)) - (int)CONV_PAD_LEFT;
 
     // Only for y = 0 we can have a negative coordinate. If so, we convert it to SRC_DIM_1
     y_coord.s0 = min((uint)y_coord.s0, (uint)SRC_DIM_1);
@@ -720,16 +799,16 @@
     // z == 1
     // z_coord can be only negative for z = 0 so we do not need to clamp it
     // Moreover z_coord cannot be out-of-bound for z = 1 so we do not need to clamp the offset
-    z_coord           = z * (int)CONV_STRIDE_Y - (int)CONV_PAD_TOP + 1;
+    z_coord           = z * (int)CONV_STRIDE_Y - (int)CONV_PAD_TOP + DILATION_Y;
     offset            = y_offset + (int4)(z_coord * src_stride_z);
     VEC_UCHAR values3 = VLOAD(VEC_SIZE)(0, src_addr + offset.s0);
     VEC_UCHAR values4 = VLOAD(VEC_SIZE)(0, src_addr + offset.s1);
     VEC_UCHAR values5 = VLOAD(VEC_SIZE)(0, src_addr + offset.s2);
 
     // z == 2
-    // After z = 1 we can simply add src_stride_z to offset without updating z_coord
-    // However offset can be out-of-bound so we need to check if it is greater than max_offset
-    offset += (int4)src_stride_z;
+    // Offset can be out-of-bound so we need to check if it is greater than max_offset
+    z_coord           = z * (int)CONV_STRIDE_Y - (int)CONV_PAD_TOP + DILATION_Y * 2;
+    offset            = y_offset + (int4)(z_coord * src_stride_z);
     offset            = min(offset, (int4)max_offset);
     VEC_UCHAR values6 = VLOAD(VEC_SIZE)(0, src_addr + offset.s0);
     VEC_UCHAR values7 = VLOAD(VEC_SIZE)(0, src_addr + offset.s1);
diff --git a/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NCHWKernel.cpp b/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NCHWKernel.cpp
index 83fac16..ec27e41 100644
--- a/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NCHWKernel.cpp
+++ b/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NCHWKernel.cpp
@@ -43,7 +43,7 @@
 namespace
 {
 Status validate_arguments(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 ActivationLayerInfo &act_info, const Size2D dilation)
 {
     ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input);
     ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::F16, DataType::F32);
@@ -56,6 +56,8 @@
     ARM_COMPUTE_RETURN_ERROR_ON(weights->dimension(0) != 3 || weights->dimension(1) != 3);
     ARM_COMPUTE_RETURN_ERROR_ON(conv_info.stride().first < 1 || conv_info.stride().first > 3);
 
+    ARM_COMPUTE_RETURN_ERROR_ON((dilation.x() < 1) || (dilation.y() < 1));
+
     const bool is_qasymm = is_data_type_quantized_asymmetric(input->data_type());
 
     if(biases != nullptr)
@@ -74,7 +76,7 @@
 
     if(output->total_size() != 0)
     {
-        const TensorShape output_shape = compute_depthwise_convolution_shape(*input, *weights, conv_info, depth_multiplier);
+        const TensorShape output_shape = compute_depthwise_convolution_shape(*input, *weights, conv_info, depth_multiplier, dilation);
         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(output->tensor_shape(), output_shape);
     }
 
@@ -82,10 +84,10 @@
 }
 
 std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *weights, ITensorInfo *output, const PadStrideInfo &conv_info, unsigned int depth_multiplier,
-                                                        GPUTarget gpu_target, std::string &kernel_name)
+                                                        GPUTarget gpu_target, std::string &kernel_name, const Size2D dilation)
 {
     // Output auto inizialitation if not yet initialized
-    const TensorShape output_shape = compute_depthwise_convolution_shape(*input, *weights, conv_info, depth_multiplier);
+    const TensorShape output_shape = compute_depthwise_convolution_shape(*input, *weights, conv_info, depth_multiplier, dilation);
     auto_init_if_empty(*output, input->clone()->set_tensor_shape(output_shape));
 
     const unsigned int conv_stride_x = conv_info.stride().first;
@@ -176,10 +178,12 @@
         kernel_name += (is_qasymm ? "_nchw" : "");
 
         num_elems_written_per_iteration_x = 8 / data_size_from_type(input->data_type());
-        num_elems_written_per_iteration_y = (is_qasymm && conv_stride_y == 1) ? 2 : 1;
+        num_elems_written_per_iteration_y = (is_qasymm && conv_stride_y == 1 && dilation.y() == 1) ? 2 : 1;
         num_elems_read_per_iteration_x    = 3 + (num_elems_written_per_iteration_x - 1) * conv_stride_x;
         num_elems_read_per_iteration_y    = num_elems_written_per_iteration_y + 2;
     }
+    num_elems_read_per_iteration_x += (num_elems_read_per_iteration_x - 1) * (dilation.x() - 1);
+    num_elems_read_per_iteration_y += (num_elems_read_per_iteration_y - 1) * (dilation.y() - 1);
 
     // Create window and update padding
     Window win = calculate_max_window(*output, Steps(num_elems_written_per_iteration_x, num_elems_written_per_iteration_y));
@@ -210,10 +214,10 @@
 }
 
 void CLDepthwiseConvolutionLayer3x3NCHWKernel::configure(const ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const PadStrideInfo &conv_info,
-                                                         unsigned int depth_multiplier, ActivationLayerInfo act_info)
+                                                         unsigned int depth_multiplier, ActivationLayerInfo act_info, const Size2D &dilation)
 {
     ARM_COMPUTE_ERROR_ON_NULLPTR(input, weights, output);
-    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), weights->info(), (biases != nullptr) ? biases->info() : nullptr, output->info(), conv_info, depth_multiplier, act_info));
+    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), weights->info(), (biases != nullptr) ? biases->info() : nullptr, output->info(), conv_info, depth_multiplier, act_info, dilation));
 
     bool is_qasymm = is_data_type_quantized_asymmetric(input->info()->data_type());
 
@@ -231,7 +235,7 @@
     std::string     kernel_name;
     const GPUTarget gpu_target = get_target();
 
-    auto win_config = validate_and_configure_window(input->info(), weights->info(), output->info(), conv_info, depth_multiplier, gpu_target, kernel_name);
+    auto win_config = validate_and_configure_window(input->info(), weights->info(), output->info(), conv_info, depth_multiplier, gpu_target, kernel_name, dilation);
     ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
     ICLKernel::configure_internal(win_config.second);
 
@@ -240,6 +244,8 @@
     build_opts.add_option("-DDST_CHANNELS=" + support::cpp11::to_string(_output->info()->tensor_shape().z()));
     build_opts.add_option("-DDEPTH_MULTIPLIER=" + support::cpp11::to_string(depth_multiplier));
     build_opts.add_option("-DCONV_STRIDE_X=" + support::cpp11::to_string(_conv_stride_x));
+    build_opts.add_option("-DDILATION_X=" + support::cpp11::to_string(dilation.x()));
+    build_opts.add_option("-DDILATION_Y=" + support::cpp11::to_string(dilation.y()));
     build_opts.add_option_if(_biases != nullptr, "-DHAS_BIAS");
 
     if(is_qasymm)
@@ -292,12 +298,11 @@
 }
 
 Status CLDepthwiseConvolutionLayer3x3NCHWKernel::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)
+                                                          unsigned int depth_multiplier, ActivationLayerInfo act_info, GPUTarget gpu_target, const Size2D &dilation)
 {
     std::string kernel_name;
-    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, weights, biases, output, conv_info, depth_multiplier, act_info));
-    ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), weights->clone().get(), output->clone().get(), conv_info, depth_multiplier, gpu_target, kernel_name).first);
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, weights, biases, output, conv_info, depth_multiplier, act_info, dilation));
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), weights->clone().get(), output->clone().get(), conv_info, depth_multiplier, gpu_target, kernel_name, dilation).first);
 
     return Status{};
 }
diff --git a/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NHWCKernel.cpp b/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NHWCKernel.cpp
index 431039c..86d186b 100644
--- a/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NHWCKernel.cpp
+++ b/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NHWCKernel.cpp
@@ -42,7 +42,7 @@
 namespace
 {
 Status validate_arguments(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 ActivationLayerInfo &act_info, const Size2D &dilation)
 {
     ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input);
     ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F16, DataType::F32, DataType::QASYMM8);
@@ -57,6 +57,8 @@
     ARM_COMPUTE_RETURN_ERROR_ON(conv_info.stride().first < 1);
     ARM_COMPUTE_RETURN_ERROR_ON(std::max(conv_info.pad_top(), conv_info.pad_bottom()) > 1);
 
+    ARM_COMPUTE_RETURN_ERROR_ON((dilation.x() < 1) || (dilation.y() < 1));
+
     const bool   is_qasymm      = is_data_type_quantized_asymmetric(input->data_type());
     const size_t weights_width  = 3;
     const size_t weights_height = 3;
@@ -89,7 +91,8 @@
 
     if(output->total_size() != 0)
     {
-        const TensorShape output_shape = arm_compute::misc::shape_calculator::compute_depthwise_convolution_shape(*input, weights_width, weights_height, conv_info, depth_multiplier);
+        const TensorShape output_shape = arm_compute::misc::shape_calculator::compute_depthwise_convolution_shape(
+                                             *input, TensorInfo(TensorShape(weights_width, weights_height), 1, weights->data_type()).set_data_layout(DataLayout::NCHW), conv_info, depth_multiplier, dilation);
         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(output->tensor_shape(), output_shape);
     }
 
@@ -97,13 +100,14 @@
 }
 
 std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *weights, ITensorInfo *bias, ITensorInfo *output,
-                                                        const PadStrideInfo &conv_info, unsigned int depth_multiplier)
+                                                        const PadStrideInfo &conv_info, unsigned int depth_multiplier, const Size2D &dilation)
 {
     const size_t weights_width  = 3;
     const size_t weights_height = 3;
 
     // Get convolved dimensions
-    const TensorShape output_shape = arm_compute::misc::shape_calculator::compute_depthwise_convolution_shape(*input, weights_width, weights_height, conv_info, depth_multiplier);
+    const TensorShape output_shape = arm_compute::misc::shape_calculator::compute_depthwise_convolution_shape(
+                                         *input, TensorInfo(TensorShape(weights_width, weights_height), 1, weights->data_type()).set_data_layout(DataLayout::NCHW), conv_info, depth_multiplier, dilation);
 
     // Output auto inizialitation if not yet initialized
     auto_init_if_empty(*output,
@@ -112,10 +116,10 @@
                        input->data_type(),
                        input->quantization_info());
 
-    const bool is_qasymm   = is_data_type_quantized_asymmetric(input->data_type());
-    const bool is_stride_1 = ((conv_info.stride().first == conv_info.stride().second) && (conv_info.stride().first == 1));
+    const bool is_qasymm              = is_data_type_quantized_asymmetric(input->data_type());
+    const bool is_stride_1_dilation_1 = ((conv_info.stride().first == conv_info.stride().second) && (conv_info.stride().first == 1) && dilation.x() == 1 && dilation.y() == 1);
 
-    const unsigned int num_rows_processed_per_iteration = is_stride_1 ? 2 : 1;
+    const unsigned int num_rows_processed_per_iteration = is_stride_1_dilation_1 ? 2 : 1;
     const unsigned int num_elems_accessed_per_iteration = is_qasymm ? 4 : (8 / input->element_size());
     const unsigned int num_rows_read_per_iteration      = num_rows_processed_per_iteration + 2;
     const unsigned int num_rows_written_per_iteration   = std::ceil(num_rows_processed_per_iteration / static_cast<float>(conv_info.stride().first));
@@ -166,15 +170,17 @@
 }
 
 void CLDepthwiseConvolutionLayer3x3NHWCKernel::configure(const ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const PadStrideInfo &conv_info,
-                                                         unsigned int depth_multiplier, ActivationLayerInfo act_info)
+                                                         unsigned int depth_multiplier, ActivationLayerInfo act_info, const Size2D &dilation)
 {
     ARM_COMPUTE_ERROR_ON_NULLPTR(input, weights, output);
-    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), weights->info(), (biases != nullptr) ? biases->info() : nullptr, output->info(), conv_info, depth_multiplier, act_info));
-    auto win_config = validate_and_configure_window(input->info(), weights->info(), biases != nullptr ? biases->info() : nullptr, output->info(), conv_info, depth_multiplier);
+    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), weights->info(), (biases != nullptr) ? biases->info() : nullptr, output->info(), conv_info, depth_multiplier, act_info, dilation));
+    auto win_config = validate_and_configure_window(input->info(), weights->info(), biases != nullptr ? biases->info() : nullptr, output->info(), conv_info, depth_multiplier, dilation);
     ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
 
-    const bool is_qasymm         = is_data_type_quantized_asymmetric(input->info()->data_type());
-    const bool is_stride_1       = ((conv_info.stride().first == conv_info.stride().second) && (conv_info.stride().first == 1));
+    const bool is_qasymm              = is_data_type_quantized_asymmetric(input->info()->data_type());
+    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());
 
     _input                              = input;
@@ -182,8 +188,8 @@
     _weights                            = weights;
     _biases                             = biases;
     _conv_stride_y                      = conv_info.stride().second;
-    _num_rows_processed_per_iteration   = is_stride_1 ? 2 : 1;
-    _num_planes_processed_per_iteration = is_stride_1 ? 2 : 1;
+    _num_rows_processed_per_iteration   = is_stride_1_dilation_1 ? 2 : 1;
+    _num_planes_processed_per_iteration = is_stride_1_dilation_1 ? 2 : 1;
 
     // If QASYMM8 and the 8 bit dot product is available, force _num_planes_processed_per_iteration to 1
     if(is_dot8_supported && is_qasymm)
@@ -201,6 +207,8 @@
     build_opts.add_option("-DSRC_DIM_2=" + support::cpp11::to_string(_input->info()->dimension(2)));
     build_opts.add_option("-DCONV_PAD_TOP=" + support::cpp11::to_string(conv_info.pad_top()));
     build_opts.add_option("-DCONV_PAD_LEFT=" + support::cpp11::to_string(conv_info.pad_left()));
+    build_opts.add_option("-DDILATION_X=" + support::cpp11::to_string(dilation.x()));
+    build_opts.add_option("-DDILATION_Y=" + support::cpp11::to_string(dilation.y()));
 
     if(is_qasymm)
     {
@@ -238,7 +246,7 @@
         build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(_input->info()->data_type()));
     }
 
-    if(is_stride_1)
+    if(is_stride_1_dilation_1)
     {
         build_opts.add_option("-DNUM_ROWS_PROCESSED=" + support::cpp11::to_string(_num_rows_processed_per_iteration));
         build_opts.add_option("-DNUM_PLANES_PROCESSED=" + support::cpp11::to_string(_num_planes_processed_per_iteration));
@@ -257,14 +265,14 @@
     if(is_qasymm)
     {
         kernel_name = std::string("dwc_3x3_reshaped_qasymm8");
-        kernel_name += (is_dot8_supported && is_stride_1 ? "_dot8" : "");
-        kernel_name += (is_stride_1 ? "_stride1" : "");
+        kernel_name += (is_dot8_supported && is_stride_1_dilation_1 ? "_dot8" : "");
+        kernel_name += (is_stride_1_dilation_1 ? "_stride1" : "");
         kernel_name += "_nhwc";
     }
     else
     {
         kernel_name = std::string("depthwise_convolution_3x3_nhwc");
-        kernel_name += (is_stride_1 ? "_stride1" : "");
+        kernel_name += (is_stride_1_dilation_1 ? "_stride1" : "");
     }
 
     ICLKernel::configure_internal(win_config.second);
@@ -287,13 +295,12 @@
 }
 
 Status CLDepthwiseConvolutionLayer3x3NHWCKernel::validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info,
-                                                          unsigned int        depth_multiplier,
-                                                          ActivationLayerInfo act_info)
+                                                          unsigned int depth_multiplier, ActivationLayerInfo act_info, const Size2D &dilation)
 {
-    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, weights, biases, output, conv_info, depth_multiplier, act_info));
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, weights, biases, output, conv_info, depth_multiplier, act_info, dilation));
     ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), weights->clone().get(),
                                                               biases != nullptr ? biases->clone().get() : nullptr,
-                                                              output->clone().get(), conv_info, depth_multiplier)
+                                                              output->clone().get(), conv_info, depth_multiplier, dilation)
                                 .first);
 
     return Status{};
diff --git a/src/core/CL/kernels/CLDepthwiseIm2ColKernel.cpp b/src/core/CL/kernels/CLDepthwiseIm2ColKernel.cpp
index beff7ae..28d4ff2 100644
--- a/src/core/CL/kernels/CLDepthwiseIm2ColKernel.cpp
+++ b/src/core/CL/kernels/CLDepthwiseIm2ColKernel.cpp
@@ -44,7 +44,8 @@
 
 namespace
 {
-Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias, unsigned int depth_multiplier)
+Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias, unsigned int depth_multiplier,
+                          const Size2D &dilation)
 {
     const size_t idx_c = get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::CHANNEL);
 
@@ -55,16 +56,18 @@
     ARM_COMPUTE_RETURN_ERROR_ON(is_data_type_quantized_asymmetric(input->data_type()) && has_bias);
     ARM_COMPUTE_RETURN_ERROR_ON((input->dimension(idx_c) * depth_multiplier) != output->dimension(2));
     ARM_COMPUTE_RETURN_ERROR_ON(output->dimension(0) != (kernel_dims.width * kernel_dims.height + ((has_bias) ? 1 : 0)));
+    ARM_COMPUTE_RETURN_ERROR_ON((dilation.x() < 1) || dilation.y() < 1);
     ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_QUANTIZATION_INFO(input, output);
 
     return Status{};
 }
 } // namespace
 
-void CLDepthwiseIm2ColKernel::configure(const ICLTensor *input, ICLTensor *output, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias, unsigned int depth_multiplier)
+void CLDepthwiseIm2ColKernel::configure(const ICLTensor *input, ICLTensor *output, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias, unsigned int depth_multiplier,
+                                        const Size2D &dilation)
 {
     ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
-    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info(), kernel_dims, conv_info, has_bias, depth_multiplier));
+    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info(), kernel_dims, conv_info, has_bias, depth_multiplier, dilation));
 
     _input  = input;
     _output = output;
@@ -88,6 +91,8 @@
     build_opts.add_option("-DKERNEL_WIDTH=" + support::cpp11::to_string(kernel_dims.width));
     build_opts.add_option("-DKERNEL_HEIGHT=" + support::cpp11::to_string(kernel_dims.height));
     build_opts.add_option("-DDEPTH_MULTIPLIER=" + support::cpp11::to_string(depth_multiplier));
+    build_opts.add_option("-DDILATION_X=" + support::cpp11::to_string(dilation.x()));
+    build_opts.add_option("-DDILATION_Y=" + support::cpp11::to_string(dilation.y()));
     build_opts.add_option("-D" + string_from_data_layout(input->info()->data_layout()));
     build_opts.add_option_if(has_bias, "-DHAS_BIAS");
     build_opts.add_option_if_else(is_data_type_quantized_asymmetric(input->info()->data_type()),
@@ -104,9 +109,10 @@
     ICLKernel::configure_internal(win);
 }
 
-Status CLDepthwiseIm2ColKernel::validate(const ITensorInfo *input, const ITensorInfo *output, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias, unsigned int depth_multiplier)
+Status CLDepthwiseIm2ColKernel::validate(const ITensorInfo *input, const ITensorInfo *output, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias, unsigned int depth_multiplier,
+                                         const Size2D &dilation)
 {
-    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output, kernel_dims, conv_info, has_bias, depth_multiplier));
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output, kernel_dims, conv_info, has_bias, depth_multiplier, dilation));
 
     return Status{};
 }
diff --git a/src/runtime/CL/functions/CLDepthwiseConvolutionLayer.cpp b/src/runtime/CL/functions/CLDepthwiseConvolutionLayer.cpp
index 65d3f5f..23c3f81 100644
--- a/src/runtime/CL/functions/CLDepthwiseConvolutionLayer.cpp
+++ b/src/runtime/CL/functions/CLDepthwiseConvolutionLayer.cpp
@@ -45,7 +45,7 @@
 }
 
 void CLDepthwiseConvolutionLayer3x3::configure(ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const PadStrideInfo &conv_info, unsigned int depth_multiplier,
-                                               ActivationLayerInfo act_info)
+                                               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);
@@ -62,11 +62,13 @@
     const ICLTensor *weights_to_use = weights;
     ICLTensor       *output_to_use  = output;
 
-    const bool                      is_stride_1       = ((conv_info.stride().first == conv_info.stride().second) && (conv_info.stride().first == 1));
-    const bool                      is_dot8_supported = dot8_supported(CLKernelLibrary::get().get_device());
+    const bool is_stride_1            = ((conv_info.stride().first == conv_info.stride().second) && (conv_info.stride().first == 1));
+    const bool is_dot8_supported      = dot8_supported(CLKernelLibrary::get().get_device());
+    const bool is_stride_1_dilation_1 = (is_stride_1 && dilation.x() == 1 && dilation.y() == 1);
+
     DepthwiseConvolutionReshapeInfo info;
     info.c0        = 4;
-    info.transpose = is_stride_1 && is_dot8_supported;
+    info.transpose = is_stride_1_dilation_1 && is_dot8_supported;
 
     if(_needs_permute)
     {
@@ -103,7 +105,7 @@
 
     // Configure kernel
     _kernel->set_target(CLScheduler::get().target());
-    _kernel->configure(input_to_use, weights_to_use, biases, output_to_use, conv_info, depth_multiplier, act_info);
+    _kernel->configure(input_to_use, weights_to_use, biases, output_to_use, conv_info, depth_multiplier, act_info, dilation);
 
     // Permute output if needed
     if(_needs_permute)
@@ -126,26 +128,26 @@
 }
 
 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)
+                                                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_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);
-    const bool                      is_stride_1           = ((conv_info.stride().first == conv_info.stride().second) && (conv_info.stride().first == 1));
-    const bool                      is_dot8_supported     = dot8_supported(CLKernelLibrary::get().get_device());
+    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);
+    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 && is_dot8_supported;
+    info.transpose = is_stride_1_dilation_1 && is_dot8_supported;
 
     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);
+        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));
@@ -155,7 +157,8 @@
         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));
+        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)
     {
@@ -163,13 +166,13 @@
         {
             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));
+                                                                                           act_info, dilation));
         }
-        ARM_COMPUTE_RETURN_ON_ERROR(CLDepthwiseConvolutionLayer3x3NHWCKernel::validate(input, weights, biases, output, conv_info, depth_multiplier, act_info));
+        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));
+        ARM_COMPUTE_RETURN_ON_ERROR(CLDepthwiseConvolutionLayer3x3NCHWKernel::validate(input, weights, biases, output, conv_info, depth_multiplier, act_info, gpu_target, dilation));
     }
 
     return Status{};
@@ -227,7 +230,7 @@
 }
 
 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)
+                                            unsigned int depth_multiplier, const 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);
@@ -241,7 +244,7 @@
     if(bool(can_run_optimised_3x3_kernel))
     {
         auto f = arm_compute::support::cpp14::make_unique<CLDepthwiseConvolutionLayer3x3>();
-        f->configure(input, weights, biases, output, conv_info, depth_multiplier, act_info);
+        f->configure(input, weights, biases, output, conv_info, depth_multiplier, act_info, dilation);
         _optimised_function = std::move(f);
     }
     else
@@ -260,7 +263,7 @@
         const GPUTarget gpu_target  = CLScheduler::get().target();
 
         // Calculate output shape
-        TensorShape output_shape = shape_calculator::compute_depthwise_convolution_shape(*input->info(), *weights->info(), conv_info, depth_multiplier);
+        TensorShape output_shape = shape_calculator::compute_depthwise_convolution_shape(*input->info(), *weights->info(), conv_info, depth_multiplier, dilation);
 
         // Output auto inizialitation if not yet initialized
         auto_init_if_empty(*output->info(), input->info()->clone()->set_tensor_shape(output_shape));
@@ -281,7 +284,7 @@
         shape_im2col.set(2, weights_z);
         _input_reshaped.allocator()->init(input->info()->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(shape_im2col));
         _im2col_kernel.set_target(gpu_target);
-        _im2col_kernel.configure(input, &_input_reshaped, Size2D(weights_w, weights_h), conv_info, append_bias, depth_multiplier);
+        _im2col_kernel.configure(input, &_input_reshaped, Size2D(weights_w, weights_h), conv_info, append_bias, depth_multiplier, dilation);
         CLScheduler::get().tune_kernel_static(_im2col_kernel);
 
         // Weights reshape configuration
@@ -343,11 +346,14 @@
 }
 
 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)
+                                             unsigned int depth_multiplier, const ActivationLayerInfo &act_info, const Size2D &dilation)
 {
     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 bool can_run_optimised_3x3_kernel = (weights->dimension(idx_w) == 3) && (weights->dimension(idx_h) == 3);
 
     if(can_run_optimised_3x3_kernel)
@@ -359,7 +365,7 @@
 
         const bool         is_quantized = is_data_type_quantized_asymmetric(input->data_type());
         const bool         append_bias  = (biases != nullptr) && !is_quantized;
-        const TensorShape  output_shape = shape_calculator::compute_depthwise_convolution_shape(*input, *weights, conv_info, depth_multiplier);
+        const TensorShape  output_shape = shape_calculator::compute_depthwise_convolution_shape(*input, *weights, conv_info, depth_multiplier, dilation);
         const size_t       weights_w    = weights->dimension(idx_w);
         const size_t       weights_h    = weights->dimension(idx_h);
         const size_t       weights_z    = weights->dimension(idx_c);
@@ -373,7 +379,7 @@
         shape_im2col.set(1, conv_size);
         shape_im2col.set(2, weights_z);
         TensorInfo input_reshaped(input->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(shape_im2col));
-        ARM_COMPUTE_RETURN_ON_ERROR(CLDepthwiseIm2ColKernel::validate(input, &input_reshaped, Size2D(weights_w, weights_h), conv_info, append_bias, depth_multiplier));
+        ARM_COMPUTE_RETURN_ON_ERROR(CLDepthwiseIm2ColKernel::validate(input, &input_reshaped, Size2D(weights_w, weights_h), conv_info, append_bias, depth_multiplier, dilation));
 
         const TensorShape shape_weights_reshape(patch_size, weights_z);
         TensorInfo        weights_reshaped(weights->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(shape_weights_reshape));
@@ -403,7 +409,7 @@
     }
     else
     {
-        CLDepthwiseConvolutionLayer3x3::validate(input, weights, biases, output, conv_info, depth_multiplier, act_info);
+        CLDepthwiseConvolutionLayer3x3::validate(input, weights, biases, output, conv_info, depth_multiplier, act_info, GPUTarget::MIDGARD, dilation);
     }
     return Status{};
 }
diff --git a/src/runtime/GLES_COMPUTE/functions/GCDepthwiseConvolutionLayer.cpp b/src/runtime/GLES_COMPUTE/functions/GCDepthwiseConvolutionLayer.cpp
index ba05838..0f772bd 100644
--- a/src/runtime/GLES_COMPUTE/functions/GCDepthwiseConvolutionLayer.cpp
+++ b/src/runtime/GLES_COMPUTE/functions/GCDepthwiseConvolutionLayer.cpp
@@ -36,8 +36,10 @@
 }
 
 void GCDepthwiseConvolutionLayer3x3::configure(IGCTensor *input, const IGCTensor *weights, const IGCTensor *biases, IGCTensor *output, const PadStrideInfo &conv_info,
-                                               unsigned int depth_multiplier, const ActivationLayerInfo &act_info)
+                                               unsigned int depth_multiplier, const ActivationLayerInfo &act_info, const Size2D &dilation)
 {
+    ARM_COMPUTE_ERROR_ON(dilation.x() != 1 || dilation.y() != 1);
+    ARM_COMPUTE_UNUSED(dilation);
     auto k = arm_compute::support::cpp14::make_unique<GCDepthwiseConvolutionLayer3x3Kernel>();
     k->configure(input, weights, biases, output, conv_info, depth_multiplier);
     _kernel = std::move(k);
diff --git a/src/runtime/NEON/functions/NEDepthwiseConvolutionLayer.cpp b/src/runtime/NEON/functions/NEDepthwiseConvolutionLayer.cpp
index 4f632a2..4c602b3 100644
--- a/src/runtime/NEON/functions/NEDepthwiseConvolutionLayer.cpp
+++ b/src/runtime/NEON/functions/NEDepthwiseConvolutionLayer.cpp
@@ -182,8 +182,11 @@
                                                const ITensor *biases,
                                                ITensor *output, const PadStrideInfo &conv_info,
                                                unsigned int               depth_multiplier,
-                                               const ActivationLayerInfo &act_info)
+                                               const ActivationLayerInfo &act_info,
+                                               const Size2D              &dilation)
 {
+    ARM_COMPUTE_ERROR_ON(dilation.x() != 1 || dilation.y() != 1);
+    ARM_COMPUTE_UNUSED(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);
 
@@ -222,10 +225,12 @@
                                                 const ITensorInfo         *output,
                                                 const PadStrideInfo       &conv_info,
                                                 unsigned int               depth_multiplier,
-                                                const ActivationLayerInfo &act_info)
+                                                const ActivationLayerInfo &act_info,
+                                                const Size2D              &dilation)
 {
     ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, weights, output);
     ARM_COMPUTE_RETURN_ERROR_ON(input->data_layout() == DataLayout::UNKNOWN);
+    ARM_COMPUTE_RETURN_ERROR_ON(dilation.x() != 1 || dilation.y() != 1);
 
     if(biases != nullptr)
     {
@@ -347,10 +352,12 @@
 }
 
 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)
+                                            unsigned int depth_multiplier, const ActivationLayerInfo &act_info, const Size2D &dilation)
 {
     const unsigned int channel_idx = get_data_layout_dimension_index(input->info()->data_layout(), DataLayoutDimension::CHANNEL);
     ARM_COMPUTE_UNUSED(channel_idx);
+    ARM_COMPUTE_ERROR_ON(dilation.x() != 1 || dilation.y() != 1);
+    ARM_COMPUTE_UNUSED(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);
@@ -480,10 +487,11 @@
 }
 
 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)
+                                             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(input->data_layout() == DataLayout::UNKNOWN);
+    ARM_COMPUTE_RETURN_ERROR_ON(dilation.x() != 1 || dilation.y() != 1);
 
     const unsigned int width_idx  = get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::WIDTH);
     const unsigned int height_idx = get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::HEIGHT);
diff --git a/tests/benchmark/fixtures/DepthwiseConvolutionLayerFixture.h b/tests/benchmark/fixtures/DepthwiseConvolutionLayerFixture.h
index 48ea038..33753bc 100644
--- a/tests/benchmark/fixtures/DepthwiseConvolutionLayerFixture.h
+++ b/tests/benchmark/fixtures/DepthwiseConvolutionLayerFixture.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017-2018 ARM Limited.
+ * Copyright (c) 2017-2019 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -45,7 +45,7 @@
 {
 public:
     template <typename...>
-    void setup(TensorShape src_shape, Size2D kernel_size, PadStrideInfo info, DataType data_type, int batches)
+    void setup(TensorShape src_shape, Size2D kernel_size, PadStrideInfo info, Size2D Dilation, DataType data_type, int batches)
     {
         // Get shapes
         TensorShape weights_shape(kernel_size.width, kernel_size.height);
diff --git a/tests/datasets/DepthwiseConvolutionLayerDataset.h b/tests/datasets/DepthwiseConvolutionLayerDataset.h
index 8a1f3b6..4c78eb8 100644
--- a/tests/datasets/DepthwiseConvolutionLayerDataset.h
+++ b/tests/datasets/DepthwiseConvolutionLayerDataset.h
@@ -38,16 +38,18 @@
 class DepthwiseConvolutionLayerDataset
 {
 public:
-    using type = std::tuple<TensorShape, Size2D, PadStrideInfo>;
+    using type = std::tuple<TensorShape, Size2D, PadStrideInfo, Size2D>;
 
     struct iterator
     {
         iterator(std::vector<TensorShape>::const_iterator   src_it,
                  std::vector<Size2D>::const_iterator        weights_it,
-                 std::vector<PadStrideInfo>::const_iterator infos_it)
+                 std::vector<PadStrideInfo>::const_iterator infos_it,
+                 std::vector<Size2D>::const_iterator        dilation_it)
             : _src_it{ std::move(src_it) },
               _weights_it{ std::move(weights_it) },
-              _infos_it{ std::move(infos_it) }
+              _infos_it{ std::move(infos_it) },
+              _dilation_it{ std::move(dilation_it) }
         {
         }
 
@@ -56,13 +58,14 @@
             std::stringstream description;
             description << "In=" << *_src_it << ":";
             description << "Weights=" << *_weights_it << ":";
-            description << "Info=" << *_infos_it;
+            description << "Info=" << *_infos_it << ":";
+            description << "Dilation=" << *_dilation_it;
             return description.str();
         }
 
         DepthwiseConvolutionLayerDataset::type operator*() const
         {
-            return std::make_tuple(*_src_it, *_weights_it, *_infos_it);
+            return std::make_tuple(*_src_it, *_weights_it, *_infos_it, *_dilation_it);
         }
 
         iterator &operator++()
@@ -70,6 +73,7 @@
             ++_src_it;
             ++_weights_it;
             ++_infos_it;
+            ++_dilation_it;
 
             return *this;
         }
@@ -78,23 +82,25 @@
         std::vector<TensorShape>::const_iterator   _src_it;
         std::vector<Size2D>::const_iterator        _weights_it;
         std::vector<PadStrideInfo>::const_iterator _infos_it;
+        std::vector<Size2D>::const_iterator        _dilation_it;
     };
 
     iterator begin() const
     {
-        return iterator(_src_shapes.begin(), _weight_shapes.begin(), _infos.begin());
+        return iterator(_src_shapes.begin(), _weight_shapes.begin(), _infos.begin(), _dilations.begin());
     }
 
     int size() const
     {
-        return std::min(_src_shapes.size(), std::min(_weight_shapes.size(), _infos.size()));
+        return std::min(_src_shapes.size(), std::min(_weight_shapes.size(), std::min(_infos.size(), _dilations.size())));
     }
 
-    void add_config(TensorShape src, Size2D weights, PadStrideInfo info)
+    void add_config(TensorShape src, Size2D weights, PadStrideInfo info, Size2D dilation = Size2D(1U, 1U))
     {
         _src_shapes.emplace_back(std::move(src));
         _weight_shapes.emplace_back(std::move(weights));
         _infos.emplace_back(std::move(info));
+        _dilations.emplace_back(std::move(dilation));
     }
 
 protected:
@@ -105,6 +111,7 @@
     std::vector<TensorShape>   _src_shapes{};
     std::vector<Size2D>        _weight_shapes{};
     std::vector<PadStrideInfo> _infos{};
+    std::vector<Size2D>        _dilations{};
 };
 
 /** Dataset containing small, generic depthwise convolution shapes. */
diff --git a/tests/datasets/DilatedDepthwiseConvolutionLayerDataset.h b/tests/datasets/DilatedDepthwiseConvolutionLayerDataset.h
new file mode 100644
index 0000000..df054de
--- /dev/null
+++ b/tests/datasets/DilatedDepthwiseConvolutionLayerDataset.h
@@ -0,0 +1,139 @@
+/*
+ * Copyright (c) 2019 ARM Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#ifndef ARM_COMPUTE_TEST_DILATED_CONVOLUTION_LAYER_DATASET
+#define ARM_COMPUTE_TEST_DILATED_CONVOLUTION_LAYER_DATASET
+
+#include "utils/TypePrinter.h"
+
+#include "arm_compute/core/TensorShape.h"
+#include "arm_compute/core/Types.h"
+#include "tests/datasets/DepthwiseConvolutionLayerDataset.h"
+
+namespace arm_compute
+{
+namespace test
+{
+namespace datasets
+{
+/** Dataset containing small, generic depthwise convolution shapes with dilation. */
+class SmallDepthwiseDilatedConvolutionLayerDataset final : public DepthwiseConvolutionLayerDataset
+{
+public:
+    SmallDepthwiseDilatedConvolutionLayerDataset()
+    {
+        // Different strides and dilations
+        add_config(TensorShape(7U, 7U, 1U), Size2D(3U, 2U), PadStrideInfo(1, 1, 0, 0), Size2D(2U, 2U));
+        add_config(TensorShape(7U, 7U, 1U), Size2D(3U, 2U), PadStrideInfo(1, 2, 0, 0), Size2D(2U, 1U));
+        add_config(TensorShape(7U, 7U, 1U), Size2D(3U, 2U), PadStrideInfo(1, 1, 0, 0), Size2D(2U, 1U));
+        add_config(TensorShape(7U, 7U, 1U), Size2D(3U, 2U), PadStrideInfo(2, 1, 0, 0), Size2D(2U, 2U));
+        add_config(TensorShape(7U, 7U, 1U), Size2D(3U, 2U), PadStrideInfo(2, 2, 0, 0), Size2D(1U, 2U));
+
+        add_config(TensorShape(7U, 8U, 1U), Size2D(2U, 3U), PadStrideInfo(1, 2, 0, 0), Size2D(2U, 2U));
+        add_config(TensorShape(23U, 27U, 5U), Size2D(3U, 5U), PadStrideInfo(2, 1, 0, 0), Size2D(2U, 1U));
+        add_config(TensorShape(33U, 27U, 7U), Size2D(7U, 3U), PadStrideInfo(3, 2, 1, 0), Size2D(1U, 2U));
+        // Asymmetric padding
+        add_config(TensorShape(33U, 27U, 7U), Size2D(5U, 7U), PadStrideInfo(3, 2, 1, 1, 2, 0, DimensionRoundingType::FLOOR), Size2D(2U, 2U));
+        add_config(TensorShape(33U, 27U, 7U), Size2D(5U, 7U), PadStrideInfo(3, 2, 1, 1, 0, 2, DimensionRoundingType::FLOOR), Size2D(2U, 2U));
+    }
+};
+
+/** Dataset containing small, 3x3 depthwise convolution shapes with dilation. */
+class SmallDepthwiseDilatedConvolutionLayerDataset3x3 final : public DepthwiseConvolutionLayerDataset
+{
+public:
+    SmallDepthwiseDilatedConvolutionLayerDataset3x3()
+    {
+        add_config(TensorShape(7U, 7U, 1U), Size2D(3U, 3U), PadStrideInfo(1, 1, 1, 0), Size2D(2U, 2U));
+        add_config(TensorShape(7U, 7U, 1U), Size2D(3U, 3U), PadStrideInfo(1, 1, 2, 0), Size2D(2U, 2U));
+        add_config(TensorShape(7U, 7U, 1U), Size2D(3U, 3U), PadStrideInfo(1, 1, 3, 0), Size2D(2U, 2U));
+
+        // Different strides and dilations
+        add_config(TensorShape(7U, 7U, 1U), Size2D(3U, 3U), PadStrideInfo(1, 1, 0, 0), Size2D(2U, 2U));
+        add_config(TensorShape(7U, 7U, 1U), Size2D(3U, 3U), PadStrideInfo(1, 2, 0, 0), Size2D(2U, 1U));
+        add_config(TensorShape(7U, 7U, 1U), Size2D(3U, 3U), PadStrideInfo(2, 1, 0, 0), Size2D(2U, 2U));
+        add_config(TensorShape(7U, 7U, 1U), Size2D(3U, 3U), PadStrideInfo(2, 2, 0, 0), Size2D(1U, 2U));
+
+        add_config(TensorShape(11U, 11U, 1U), Size2D(3U, 3U), PadStrideInfo(3, 3, 0, 0), Size2D(2U, 2U));
+        add_config(TensorShape(7U, 7U, 3U, 2U), Size2D(3U, 3U), PadStrideInfo(1, 1, 0, 0), Size2D(2U, 2U));
+
+        add_config(TensorShape(21U, 31U, 9U, 4U), Size2D(3U, 3U), PadStrideInfo(1, 1, 1, 0), Size2D(1U, 1U));
+
+        add_config(TensorShape(21U, 31U, 9U, 4U), Size2D(3U, 3U), PadStrideInfo(1, 1, 1, 0), Size2D(2U, 2U));
+        // Asymmetric padding
+        add_config(TensorShape(33U, 27U, 11U), Size2D(3U, 3U), PadStrideInfo(2, 2, 0, 1, 0, 1, DimensionRoundingType::FLOOR), Size2D(2U, 2U));
+    }
+};
+
+/** Dataset containing large, generic depthwise convolution shapes with dilation. */
+class LargeDepthwiseDilatedConvolutionLayerDataset final : public DepthwiseConvolutionLayerDataset
+{
+public:
+    LargeDepthwiseDilatedConvolutionLayerDataset()
+    {
+        add_config(TensorShape(33U, 27U, 11U), Size2D(3U, 3U), PadStrideInfo(1, 2, 0, 1), Size2D(2U, 1U));
+        add_config(TensorShape(17U, 31U, 2U), Size2D(5U, 9U), PadStrideInfo(1, 2, 1, 1), Size2D(1U, 2U));
+        add_config(TensorShape(23U, 27U, 5U), Size2D(11U, 3U), PadStrideInfo(1, 2, 0, 0), Size2D(3U, 3U));
+        add_config(TensorShape(17U, 31U, 2U, 3U), Size2D(5U, 9U), PadStrideInfo(1, 2, 1, 1), Size2D(2U, 2U));
+        add_config(TensorShape(233U, 277U, 55U), Size2D(3U, 3U), PadStrideInfo(2, 1, 0, 0), Size2D(2U, 2U));
+        add_config(TensorShape(333U, 277U, 77U), Size2D(3U, 3U), PadStrideInfo(3, 2, 1, 0), Size2D(3U, 2U));
+        add_config(TensorShape(177U, 311U, 22U), Size2D(3U, 3U), PadStrideInfo(1, 2, 1, 1), Size2D(2U, 2U));
+        add_config(TensorShape(233U, 277U, 55U), Size2D(3U, 3U), PadStrideInfo(1, 2, 0, 0), Size2D(5U, 2U));
+        add_config(TensorShape(333U, 277U, 77U), Size2D(3U, 3U), PadStrideInfo(2, 3, 0, 1), Size2D(2U, 2U));
+        add_config(TensorShape(177U, 311U, 22U), Size2D(3U, 3U), PadStrideInfo(2, 1, 1, 1), Size2D(2U, 5U));
+        // Asymmetric padding
+        add_config(TensorShape(33U, 27U, 7U), Size2D(5U, 7U), PadStrideInfo(3, 2, 2, 1, 2, 0, DimensionRoundingType::FLOOR), Size2D(3U, 2U));
+        add_config(TensorShape(33U, 27U, 7U), Size2D(5U, 7U), PadStrideInfo(3, 2, 1, 3, 0, 2, DimensionRoundingType::FLOOR), Size2D(4U, 4U));
+        add_config(TensorShape(33U, 27U, 7U), Size2D(5U, 7U), PadStrideInfo(3, 2, 1, 0, 1, 0, DimensionRoundingType::FLOOR), Size2D(2U, 2U));
+        add_config(TensorShape(33U, 27U, 7U), Size2D(5U, 7U), PadStrideInfo(3, 2, 0, 1, 0, 1, DimensionRoundingType::FLOOR), Size2D(3U, 3U));
+    }
+};
+
+/** Dataset containing large, 3x3 depthwise convolution shapes with dilation. */
+class LargeDepthwiseDilatedConvolutionLayerDataset3x3 final : public DepthwiseConvolutionLayerDataset
+{
+public:
+    LargeDepthwiseDilatedConvolutionLayerDataset3x3()
+    {
+        add_config(TensorShape(33U, 27U, 11U, 3U), Size2D(3U, 3U), PadStrideInfo(1, 1, 0, 1), Size2D(2U, 1U));
+        add_config(TensorShape(33U, 27U, 11U, 3U), Size2D(3U, 3U), PadStrideInfo(1, 1, 1, 1), Size2D(2U, 2U));
+        add_config(TensorShape(21U, 31U, 9U, 4U), Size2D(3U, 3U), PadStrideInfo(1, 2, 1, 0), Size2D(2U, 2U));
+        add_config(TensorShape(33U, 27U, 11U, 3U), Size2D(3U, 3U), PadStrideInfo(1, 2, 0, 1), Size2D(2U, 1U));
+        add_config(TensorShape(33U, 27U, 11U, 3U), Size2D(3U, 3U), PadStrideInfo(1, 2, 1, 1), Size2D(2U, 3U));
+        add_config(TensorShape(21U, 31U, 9U, 4U), Size2D(3U, 3U), PadStrideInfo(2, 1, 1, 0), Size2D(2U, 1U));
+        add_config(TensorShape(33U, 27U, 11U, 3U), Size2D(3U, 3U), PadStrideInfo(2, 1, 0, 1), Size2D(3U, 3U));
+        add_config(TensorShape(33U, 27U, 11U, 3U), Size2D(3U, 3U), PadStrideInfo(2, 1, 1, 1), Size2D(2U, 2U));
+        add_config(TensorShape(21U, 31U, 9U, 4U), Size2D(3U, 3U), PadStrideInfo(2, 2, 1, 0), Size2D(2U, 2U));
+        add_config(TensorShape(33U, 27U, 11U, 3U), Size2D(3U, 3U), PadStrideInfo(2, 2, 0, 1), Size2D(4U, 4U));
+        add_config(TensorShape(33U, 27U, 11U, 3U), Size2D(3U, 3U), PadStrideInfo(2, 2, 1, 1), Size2D(2U, 5U));
+        add_config(TensorShape(233U, 277U, 55U, 3U), Size2D(3U, 3U), PadStrideInfo(2, 1, 0, 0), Size2D(3U, 3U));
+        add_config(TensorShape(177U, 311U, 22U), Size2D(3U, 3U), PadStrideInfo(1, 2, 1, 1), Size2D(4U, 4U));
+        add_config(TensorShape(233U, 277U, 55U), Size2D(3U, 3U), PadStrideInfo(1, 2, 0, 0), Size2D(5U, 5U));
+        add_config(TensorShape(333U, 277U, 77U, 5U), Size2D(3U, 3U), PadStrideInfo(2, 3, 0, 1), Size2D(4U, 4U));
+        add_config(TensorShape(177U, 311U, 22U), Size2D(3U, 3U), PadStrideInfo(2, 1, 1, 1), Size2D(3U, 3U));
+    }
+};
+} // namespace datasets
+} // namespace test
+} // namespace arm_compute
+#endif /* ARM_COMPUTE_TEST_DILATED_CONVOLUTION_LAYER_DATASET */
diff --git a/tests/validation/CL/DepthwiseConvolutionLayer.cpp b/tests/validation/CL/DepthwiseConvolutionLayer.cpp
index dd2d9f3..94f64e1 100644
--- a/tests/validation/CL/DepthwiseConvolutionLayer.cpp
+++ b/tests/validation/CL/DepthwiseConvolutionLayer.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017-2018 ARM Limited.
+ * Copyright (c) 2017-2019 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -18,7 +18,7 @@
  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
  * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
- * OUT OF OR IN CONCLCTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  * SOFTWARE.
  */
 #include "arm_compute/core/Types.h"
@@ -28,6 +28,7 @@
 #include "tests/CL/CLAccessor.h"
 #include "tests/PaddingCalculator.h"
 #include "tests/datasets/DepthwiseConvolutionLayerDataset.h"
+#include "tests/datasets/DilatedDepthwiseConvolutionLayerDataset.h"
 #include "tests/framework/Asserts.h"
 #include "tests/framework/Macros.h"
 #include "tests/framework/datasets/Datasets.h"
@@ -55,7 +56,7 @@
 
 // *INDENT-OFF*
 // clang-format off
-DATA_TEST_CASE(Validate3x3, framework::DatasetMode::ALL, zip(zip(zip(zip(zip(zip(zip(
+DATA_TEST_CASE(Validate3x3, framework::DatasetMode::ALL, zip(zip(zip(zip(zip(zip(zip(zip(
                framework::dataset::make("InputInfo", { TensorInfo(TensorShape(32U, 18U, 2U), 1, DataType::F32),     // Mismatching data type input/weights
                                                        TensorInfo(TensorShape(32U, 18U, 3U), 1, DataType::F32),     // Mismatching input feature maps
                                                        TensorInfo(TensorShape(32U, 18U, 2U), 1, DataType::F32),     // Unsupported weights dimensions
@@ -66,6 +67,8 @@
                                                        TensorInfo(TensorShape(32U, 18U, 2U), 1, DataType::F32),     // Invalid biases dimensions
                                                        TensorInfo(TensorShape(32U, 18U, 2U), 1, DataType::F32),     // Invalid output size
                                                        TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32),     // Window shrink
+                                                       TensorInfo(TensorShape(32U, 18U, 8U), 1, DataType::F32),     // patch size bigger than input width
+                                                       TensorInfo(TensorShape(32U, 18U, 8U), 1, DataType::F32),     // dilation < 1
                                                        TensorInfo(TensorShape(32U, 18U, 8U), 1, DataType::F32),
                                                        TensorInfo(TensorShape(50U, 32U, 8U), 1, DataType::QASYMM8),
                                                      }),
@@ -80,6 +83,8 @@
                                                          TensorInfo(TensorShape(3U, 3U, 2U), 1, DataType::F32),
                                                          TensorInfo(TensorShape(3U, 3U, 2U), 1, DataType::F32),
                                                          TensorInfo(TensorShape(3U, 3U, 16U), 1, DataType::F32),
+                                                         TensorInfo(TensorShape(3U, 3U, 16U), 1, DataType::F32),
+                                                         TensorInfo(TensorShape(3U, 3U, 16U), 1, DataType::F32),
                                                          TensorInfo(TensorShape(3U, 3U, 24U), 1, DataType::QASYMM8),
                                                        })),
                framework::dataset::make("BiasesInfo", { TensorInfo(TensorShape(2U), 1, DataType::F32),
@@ -93,6 +98,8 @@
                                                         TensorInfo(TensorShape(2U), 1, DataType::F32),
                                                         TensorInfo(TensorShape(2U), 1, DataType::F32),
                                                         TensorInfo(TensorShape(16U), 1, DataType::F32),
+                                                        TensorInfo(TensorShape(16U), 1, DataType::F32),
+                                                        TensorInfo(TensorShape(16U), 1, DataType::F32),
                                                         TensorInfo(TensorShape(24U), 1, DataType::S32),
                                                       })),
                framework::dataset::make("OutputInfo", { TensorInfo(TensorShape(30U, 16U, 2U), 1, DataType::F32),
@@ -106,6 +113,8 @@
                                                         TensorInfo(TensorShape(32U, 18U, 2U), 1, DataType::F32),
                                                         TensorInfo(TensorShape(25U, 11U, 2U), 1, DataType::F32),
                                                         TensorInfo(TensorShape(30U, 16U, 16U), 1, DataType::F32),
+                                                        TensorInfo(TensorShape(30U, 16U, 16U), 1, DataType::F32),
+                                                        TensorInfo(TensorShape(30U, 16U, 16U), 1, DataType::F32),
                                                         TensorInfo(TensorShape(48U, 30U, 24U), 1, DataType::QASYMM8),
                                                       })),
                framework::dataset::make("ConvInfo", { PadStrideInfo(1, 1, 0, 0),
@@ -120,6 +129,8 @@
                                                       PadStrideInfo(1, 1, 0, 0),
                                                       PadStrideInfo(1, 1, 0, 0),
                                                       PadStrideInfo(1, 1, 0, 0),
+                                                      PadStrideInfo(1, 1, 0, 0),
+                                                      PadStrideInfo(1, 1, 0, 0),
                                                      })),
                framework::dataset::make("DepthMultiplier", { 1,
                                                              1,
@@ -132,6 +143,8 @@
                                                              1,
                                                              1,
                                                              2,
+                                                             2,
+                                                             2,
                                                              3,
                                                             })),
                 framework::dataset::make("ActivationInfo", { ActivationLayerInfo(),
@@ -145,22 +158,41 @@
                                                              ActivationLayerInfo(),
                                                              ActivationLayerInfo(),
                                                              ActivationLayerInfo(),
+                                                             ActivationLayerInfo(),
+                                                             ActivationLayerInfo(),
                                                              ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU),
                                                            })),
-               framework::dataset::make("Expected", { false, false, false, false, false, false, false, false, false, false, true, true })),
-               input_info, weights_info, biases_info, output_info, conv_info, depth_multiplier, act_info, expected)
+                framework::dataset::make("Dilation", { Size2D(1U, 1U),
+                                                       Size2D(1U, 1U),
+                                                       Size2D(1U, 1U),
+                                                       Size2D(1U, 1U),
+                                                       Size2D(1U, 1U),
+                                                       Size2D(1U, 1U),
+                                                       Size2D(1U, 1U),
+                                                       Size2D(1U, 1U),
+                                                       Size2D(1U, 1U),
+                                                       Size2D(1U, 1U),
+                                                       Size2D(20U, 1U),
+                                                       Size2D(0U, 1U),
+                                                       Size2D(1U, 1U),
+                                                       Size2D(1U, 1U),
+                                                       })),
+               framework::dataset::make("Expected", { false, false, false, false, false, false, false, false, false, false, false, false, true, true })),
+               input_info, weights_info, biases_info, output_info, conv_info, depth_multiplier, act_info, dilation, expected)
 {
-    bool is_valid = bool(CLDepthwiseConvolutionLayer3x3::validate(&input_info.clone()->set_is_resizable(false), &weights_info.clone()->set_is_resizable(false), &biases_info.clone()->set_is_resizable(false), &output_info.clone()->set_is_resizable(false), conv_info, depth_multiplier, act_info));
+    bool is_valid = bool(CLDepthwiseConvolutionLayer3x3::validate(&input_info.clone()->set_is_resizable(false), &weights_info.clone()->set_is_resizable(false), &biases_info.clone()->set_is_resizable(false), &output_info.clone()->set_is_resizable(false), conv_info, depth_multiplier, act_info,GPUTarget::MIDGARD, dilation));
     ARM_COMPUTE_EXPECT(is_valid == expected, framework::LogLevel::ERRORS);
 }
 
-DATA_TEST_CASE(ValidateGeneric, framework::DatasetMode::ALL, zip(zip(zip(zip(zip(zip(
+DATA_TEST_CASE(ValidateGeneric, framework::DatasetMode::ALL, zip(zip(zip(zip(zip(zip(zip(
                 framework::dataset::make("InputInfo", { TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32),    // Mismatching data type input/weights
                                                         TensorInfo(TensorShape(27U, 13U, 3U), 1, DataType::F32),    // Mismatching input feature maps
                                                         TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32),    // Mismatching depth multiplier
                                                         TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32),    // Invalid biases size
                                                         TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32),    // Invalid biases dimensions
                                                         TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32),    // Invalid output size
+                                                        TensorInfo(TensorShape(27U, 13U, 8U), 1, DataType::F32),    // patch size bigger than input width
+                                                        TensorInfo(TensorShape(27U, 13U, 8U), 1, DataType::F32),    // dilation < 1
                                                         TensorInfo(TensorShape(27U, 13U, 8U), 1, DataType::F32),
                                                         TensorInfo(TensorShape(32U, 13U, 8U), 1, DataType::QASYMM8),
                                                       }),
@@ -171,6 +203,8 @@
                                                           TensorInfo(TensorShape(3U, 3U, 2U), 1, DataType::F32),
                                                           TensorInfo(TensorShape(3U, 3U, 2U), 1, DataType::F32),
                                                           TensorInfo(TensorShape(3U, 3U, 16U), 1, DataType::F32),
+                                                          TensorInfo(TensorShape(3U, 3U, 16U), 1, DataType::F32),
+                                                          TensorInfo(TensorShape(3U, 3U, 16U), 1, DataType::F32),
                                                           TensorInfo(TensorShape(3U, 3U, 24U), 1, DataType::QASYMM8),
                                                         })),
                 framework::dataset::make("BiasesInfo", { TensorInfo(TensorShape(2U), 1, DataType::F32),
@@ -180,6 +214,8 @@
                                                          TensorInfo(TensorShape(2U, 2U), 1, DataType::F32),
                                                          TensorInfo(TensorShape(2U), 1, DataType::F32),
                                                          TensorInfo(TensorShape(16U), 1, DataType::F32),
+                                                         TensorInfo(TensorShape(16U), 1, DataType::F32),
+                                                         TensorInfo(TensorShape(16U), 1, DataType::F32),
                                                          TensorInfo(TensorShape(24U), 1, DataType::S32),
                                                        })),
                 framework::dataset::make("OutputInfo", { TensorInfo(TensorShape(25U, 11U, 2U), 1, DataType::F32),
@@ -189,6 +225,8 @@
                                                          TensorInfo(TensorShape(25U, 11U, 2U), 1, DataType::F32),
                                                          TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32),
                                                          TensorInfo(TensorShape(25U, 11U, 16U), 1, DataType::F32),
+                                                         TensorInfo(TensorShape(25U, 11U, 16U), 1, DataType::F32),
+                                                         TensorInfo(TensorShape(25U, 11U, 16U), 1, DataType::F32),
                                                          TensorInfo(TensorShape(32U, 11U, 24U), 1, DataType::QASYMM8),
                                                        })),
                 framework::dataset::make("ConvInfo", { PadStrideInfo(1, 1, 0, 0),
@@ -198,6 +236,8 @@
                                                        PadStrideInfo(1, 1, 0, 0),
                                                        PadStrideInfo(1, 1, 0, 0),
                                                        PadStrideInfo(1, 1, 0, 0),
+                                                       PadStrideInfo(1, 1, 0, 0),
+                                                       PadStrideInfo(1, 1, 0, 0),
                                                        PadStrideInfo(1, 1, 1, 0),
                                                       })),
                 framework::dataset::make("DepthMultiplier", { 1,
@@ -207,12 +247,25 @@
                                                               1,
                                                               1,
                                                               2,
+                                                              2,
+                                                              2,
                                                               3,
                                                              })),
-                framework::dataset::make("Expected", { false, false, false, false, false, false, true, true })),
-                input_info, weights_info, biases_info, output_info, conv_info, depth_multiplier, expected)
+                framework::dataset::make("Dilation", { Size2D(1U, 1U),
+                                                              Size2D(1U, 1U),
+                                                              Size2D(1U, 1U),
+                                                              Size2D(1U, 1U),
+                                                              Size2D(1U, 1U),
+                                                              Size2D(1U, 1U),
+                                                              Size2D(20U, 1U),
+                                                              Size2D(0U, 1U),
+                                                              Size2D(1U, 1U),
+                                                              Size2D(1U, 1U),
+                                                             })),
+                framework::dataset::make("Expected", { false, false, false, false, false, false, false, false, true, true })),
+                input_info, weights_info, biases_info, output_info, conv_info, depth_multiplier, dilation, expected)
 {
-    bool is_valid = bool(CLDepthwiseConvolutionLayer::validate(&input_info.clone()->set_is_resizable(false), &weights_info.clone()->set_is_resizable(false), &biases_info.clone()->set_is_resizable(false), &output_info.clone()->set_is_resizable(false), conv_info, depth_multiplier));
+    bool is_valid = bool(CLDepthwiseConvolutionLayer::validate(&input_info.clone()->set_is_resizable(false), &weights_info.clone()->set_is_resizable(false), &biases_info.clone()->set_is_resizable(false), &output_info.clone()->set_is_resizable(false), conv_info, depth_multiplier,ActivationLayerInfo(), dilation));
     ARM_COMPUTE_EXPECT(is_valid == expected, framework::LogLevel::ERRORS);
 }
 // clang-format on
@@ -243,7 +296,25 @@
 {
     validate(CLAccessor(_target), _reference, tolerance_f16);
 }
-TEST_SUITE_END()
+TEST_SUITE(Dilation)
+FIXTURE_DATA_TEST_CASE(RunSmall, CLDepthwiseConvolutionLayerFixture<half>, framework::DatasetMode::ALL, combine(combine(combine(datasets::SmallDepthwiseDilatedConvolutionLayerDataset3x3(),
+                                                                                                                        depth_multipliers),
+                                                                                                                        framework::dataset::make("DataType",
+                                                                                                                                DataType::F16)),
+                                                                                                                framework::dataset::make("DataLayout", { DataLayout::NCHW })))
+{
+    validate(CLAccessor(_target), _reference, tolerance_f16);
+}
+FIXTURE_DATA_TEST_CASE(RunLarge, CLDepthwiseConvolutionLayerFixture<half>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::LargeDepthwiseDilatedConvolutionLayerDataset3x3(),
+                                                                                                                    depth_multipliers),
+                                                                                                                    framework::dataset::make("DataType",
+                                                                                                                            DataType::F16)),
+                                                                                                                    framework::dataset::make("DataLayout", { DataLayout::NCHW })))
+{
+    validate(CLAccessor(_target), _reference, tolerance_f16);
+}
+TEST_SUITE_END() // Dilation
+TEST_SUITE_END() // NCHW
 
 TEST_SUITE(NHWC)
 FIXTURE_DATA_TEST_CASE(RunSmall, CLDepthwiseConvolutionLayerFixture<half>, framework::DatasetMode::ALL,
@@ -263,8 +334,26 @@
 {
     validate(CLAccessor(_target), _reference, tolerance_f16);
 }
-TEST_SUITE_END()
-TEST_SUITE_END()
+TEST_SUITE(Dilation)
+FIXTURE_DATA_TEST_CASE(RunSmall, CLDepthwiseConvolutionLayerFixture<half>, framework::DatasetMode::ALL, combine(combine(combine(datasets::SmallDepthwiseDilatedConvolutionLayerDataset3x3(),
+                                                                                                                        depth_multipliers),
+                                                                                                                        framework::dataset::make("DataType",
+                                                                                                                                DataType::F16)),
+                                                                                                                framework::dataset::make("DataLayout", { DataLayout::NHWC })))
+{
+    validate(CLAccessor(_target), _reference, tolerance_f16);
+}
+FIXTURE_DATA_TEST_CASE(RunLarge, CLDepthwiseConvolutionLayerFixture<half>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::LargeDepthwiseDilatedConvolutionLayerDataset3x3(),
+                                                                                                                    depth_multipliers),
+                                                                                                                    framework::dataset::make("DataType",
+                                                                                                                            DataType::F16)),
+                                                                                                                    framework::dataset::make("DataLayout", { DataLayout::NHWC })))
+{
+    validate(CLAccessor(_target), _reference, tolerance_f16);
+}
+TEST_SUITE_END() // Dilation
+TEST_SUITE_END() // NHWC
+TEST_SUITE_END() // W3x3
 
 TEST_SUITE(Generic)
 FIXTURE_DATA_TEST_CASE(RunSmall, CLDepthwiseConvolutionLayerFixture<half>, framework::DatasetMode::ALL, combine(combine(combine(datasets::SmallDepthwiseConvolutionLayerDataset(), depth_multipliers),
@@ -282,8 +371,27 @@
 {
     validate(CLAccessor(_target), _reference, tolerance_f16, tolerance_num);
 }
-TEST_SUITE_END()
-TEST_SUITE_END()
+
+TEST_SUITE(Dilation)
+FIXTURE_DATA_TEST_CASE(RunSmall, CLDepthwiseConvolutionLayerFixture<half>, framework::DatasetMode::ALL, combine(combine(combine(datasets::SmallDepthwiseDilatedConvolutionLayerDataset(),
+                                                                                                                        depth_multipliers),
+                                                                                                                        framework::dataset::make("DataType",
+                                                                                                                                DataType::F16)),
+                                                                                                                framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })))
+{
+    validate(CLAccessor(_target), _reference, tolerance_f16, tolerance_num);
+}
+FIXTURE_DATA_TEST_CASE(RunLarge, CLDepthwiseConvolutionLayerFixture<half>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::LargeDepthwiseDilatedConvolutionLayerDataset(),
+                                                                                                                    depth_multipliers),
+                                                                                                                    framework::dataset::make("DataType",
+                                                                                                                            DataType::F16)),
+                                                                                                                    framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })))
+{
+    validate(CLAccessor(_target), _reference, tolerance_f16, tolerance_num);
+}
+TEST_SUITE_END() // Dilation
+TEST_SUITE_END() // Generic
+TEST_SUITE_END() // FP16
 
 TEST_SUITE(FP32)
 TEST_SUITE(W3x3)
@@ -306,7 +414,26 @@
 {
     validate(CLAccessor(_target), _reference, tolerance_f32);
 }
-TEST_SUITE_END()
+TEST_SUITE(Dilation)
+FIXTURE_DATA_TEST_CASE(RunSmall, CLDepthwiseConvolutionLayerFixture<float>, framework::DatasetMode::ALL,
+                       combine(combine(combine(datasets::SmallDepthwiseDilatedConvolutionLayerDataset3x3(), depth_multipliers),
+                                       framework::dataset::make("DataType",
+                                                                DataType::F32)),
+                               framework::dataset::make("DataLayout", DataLayout::NCHW)))
+{
+    validate(CLAccessor(_target), _reference, tolerance_f32);
+}
+FIXTURE_DATA_TEST_CASE(RunLarge, CLDepthwiseConvolutionLayerFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::LargeDepthwiseDilatedConvolutionLayerDataset3x3(),
+                                                                                                                     depth_multipliers),
+                                                                                                                     framework::dataset::make("DataType",
+                                                                                                                             DataType::F32)),
+                                                                                                                     framework::dataset::make("DataLayout", DataLayout::NCHW)))
+{
+    validate(CLAccessor(_target), _reference, tolerance_f32);
+}
+
+TEST_SUITE_END() // Dilation
+TEST_SUITE_END() // NCHW
 TEST_SUITE(NHWC)
 FIXTURE_DATA_TEST_CASE(RunSmall, CLDepthwiseConvolutionLayerFixture<float>, framework::DatasetMode::ALL,
                        combine(combine(combine(datasets::SmallDepthwiseConvolutionLayerDataset3x3(),
@@ -325,8 +452,28 @@
 {
     validate(CLAccessor(_target), _reference, tolerance_f32);
 }
-TEST_SUITE_END()
-TEST_SUITE_END()
+TEST_SUITE(Dilation)
+
+FIXTURE_DATA_TEST_CASE(RunSmall, CLDepthwiseConvolutionLayerFixture<float>, framework::DatasetMode::ALL,
+                       combine(combine(combine(datasets::SmallDepthwiseDilatedConvolutionLayerDataset3x3(),
+                                               depth_multipliers),
+                                       framework::dataset::make("DataType",
+                                                                DataType::F32)),
+                               framework::dataset::make("DataLayout", DataLayout::NHWC)))
+{
+    validate(CLAccessor(_target), _reference, tolerance_f32);
+}
+FIXTURE_DATA_TEST_CASE(RunLarge, CLDepthwiseConvolutionLayerFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::LargeDepthwiseDilatedConvolutionLayerDataset3x3(),
+                                                                                                                     depth_multipliers),
+                                                                                                                     framework::dataset::make("DataType",
+                                                                                                                             DataType::F32)),
+                                                                                                                     framework::dataset::make("DataLayout", DataLayout::NHWC)))
+{
+    validate(CLAccessor(_target), _reference, tolerance_f32);
+}
+TEST_SUITE_END() // Dilation
+TEST_SUITE_END() // NHWC
+TEST_SUITE_END() // W3x3
 
 TEST_SUITE(Generic)
 FIXTURE_DATA_TEST_CASE(RunSmall, CLDepthwiseConvolutionLayerFixture<float>, framework::DatasetMode::ALL, combine(combine(combine(datasets::SmallDepthwiseConvolutionLayerDataset(), depth_multipliers),
@@ -344,9 +491,28 @@
 {
     validate(CLAccessor(_target), _reference, tolerance_f32);
 }
-TEST_SUITE_END()
-TEST_SUITE_END()
-TEST_SUITE_END()
+
+TEST_SUITE(Dilation)
+FIXTURE_DATA_TEST_CASE(RunSmall, CLDepthwiseConvolutionLayerFixture<float>, framework::DatasetMode::ALL, combine(combine(combine(datasets::SmallDepthwiseDilatedConvolutionLayerDataset(),
+                                                                                                                 depth_multipliers),
+                                                                                                                 framework::dataset::make("DataType",
+                                                                                                                         DataType::F32)),
+                                                                                                                 framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })))
+{
+    validate(CLAccessor(_target), _reference, tolerance_f32);
+}
+FIXTURE_DATA_TEST_CASE(RunLarge, CLDepthwiseConvolutionLayerFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::LargeDepthwiseDilatedConvolutionLayerDataset3x3(),
+                                                                                                                     depth_multipliers),
+                                                                                                                     framework::dataset::make("DataType",
+                                                                                                                             DataType::F32)),
+                                                                                                                     framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })))
+{
+    validate(CLAccessor(_target), _reference, tolerance_f32);
+}
+TEST_SUITE_END() // Dilation
+TEST_SUITE_END() // Generic
+TEST_SUITE_END() // FP32
+TEST_SUITE_END() // Float
 
 template <typename T>
 using CLDepthwiseConvolutionLayerQuantizedFixture = DepthwiseConvolutionLayerValidationQuantizedFixture<CLTensor, CLAccessor, CLDepthwiseConvolutionLayer, T>;
@@ -372,7 +538,27 @@
 {
     validate(CLAccessor(_target), _reference, tolerance_qasymm8);
 }
-TEST_SUITE_END()
+TEST_SUITE(Dilation)
+FIXTURE_DATA_TEST_CASE(RunSmall, CLDepthwiseConvolutionLayerQuantizedFixture<uint8_t>, framework::DatasetMode::PRECOMMIT,
+                       combine(combine(combine(combine(datasets::SmallDepthwiseDilatedConvolutionLayerDataset(),
+                                                       depth_multipliers),
+                                               framework::dataset::make("DataType", DataType::QASYMM8)),
+                                       framework::dataset::make("QuantizationInfo", { QuantizationInfo(0.5f, 10) })),
+                               framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })))
+{
+    validate(CLAccessor(_target), _reference, tolerance_qasymm8);
+}
+FIXTURE_DATA_TEST_CASE(RunLarge, CLDepthwiseConvolutionLayerQuantizedFixture<uint8_t>, framework::DatasetMode::NIGHTLY,
+                       combine(combine(combine(combine(datasets::LargeDepthwiseDilatedConvolutionLayerDataset(),
+                                                       depth_multipliers),
+                                               framework::dataset::make("DataType", DataType::QASYMM8)),
+                                       framework::dataset::make("QuantizationInfo", { QuantizationInfo(0.5f, 10) })),
+                               framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })))
+{
+    validate(CLAccessor(_target), _reference, tolerance_qasymm8);
+}
+TEST_SUITE_END() // Dilation
+TEST_SUITE_END() // Generic
 TEST_SUITE(W3x3)
 FIXTURE_DATA_TEST_CASE(RunSmall, CLDepthwiseConvolutionLayerQuantizedFixture<uint8_t>, framework::DatasetMode::PRECOMMIT,
                        combine(combine(combine(combine(datasets::SmallDepthwiseConvolutionLayerDataset3x3(),
@@ -392,12 +578,32 @@
 {
     validate(CLAccessor(_target), _reference, tolerance_qasymm8);
 }
-TEST_SUITE_END()
-TEST_SUITE_END()
-TEST_SUITE_END()
+TEST_SUITE(Dilation)
+FIXTURE_DATA_TEST_CASE(RunSmall, CLDepthwiseConvolutionLayerQuantizedFixture<uint8_t>, framework::DatasetMode::PRECOMMIT,
+                       combine(combine(combine(combine(datasets::SmallDepthwiseDilatedConvolutionLayerDataset3x3(),
+                                                       depth_multipliers),
+                                               framework::dataset::make("DataType", DataType::QASYMM8)),
+                                       framework::dataset::make("QuantizationInfo", { QuantizationInfo(0.5f, 10) })),
+                               framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })))
+{
+    validate(CLAccessor(_target), _reference, tolerance_qasymm8);
+}
+FIXTURE_DATA_TEST_CASE(RunLarge, CLDepthwiseConvolutionLayerQuantizedFixture<uint8_t>, framework::DatasetMode::NIGHTLY,
+                       combine(combine(combine(combine(datasets::LargeDepthwiseDilatedConvolutionLayerDataset3x3(),
+                                                       depth_multipliers),
+                                               framework::dataset::make("DataType", DataType::QASYMM8)),
+                                       framework::dataset::make("QuantizationInfo", { QuantizationInfo(0.5f, 10) })),
+                               framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })))
+{
+    validate(CLAccessor(_target), _reference, tolerance_qasymm8);
+}
+TEST_SUITE_END() // Dilation
+TEST_SUITE_END() // W3x3
+TEST_SUITE_END() // QASYMM8
+TEST_SUITE_END() // Quantized
 
-TEST_SUITE_END()
-TEST_SUITE_END()
+TEST_SUITE_END() // DepthwiseConvolutionLayer
+TEST_SUITE_END() // CL
 } // namespace validation
 } // namespace test
 } // namespace arm_compute
diff --git a/tests/validation/fixtures/DepthwiseConvolutionLayerFixture.h b/tests/validation/fixtures/DepthwiseConvolutionLayerFixture.h
index 5428154..dd8bf23 100644
--- a/tests/validation/fixtures/DepthwiseConvolutionLayerFixture.h
+++ b/tests/validation/fixtures/DepthwiseConvolutionLayerFixture.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017-2018 ARM Limited.
+ * Copyright (c) 2017-2019 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -55,7 +55,8 @@
 
 public:
     template <typename...>
-    void setup(TensorShape in_shape, Size2D kernel_size, PadStrideInfo pad_stride_info, unsigned int depth_multiplier, DataType data_type, QuantizationInfo quantization_info, DataLayout data_layout)
+    void setup(TensorShape in_shape, Size2D kernel_size, PadStrideInfo pad_stride_info, Size2D dilation, unsigned int depth_multiplier, DataType data_type, QuantizationInfo quantization_info,
+               DataLayout data_layout)
     {
         _quantization_info            = quantization_info;
         _data_type                    = data_type;
@@ -65,13 +66,13 @@
 
         const TensorInfo in_info(in_shape, 1, data_type);
         const TensorInfo we_info(weights_shape, 1, data_type);
-        TensorShape      out_shape = compute_depthwise_convolution_shape(in_info, we_info, pad_stride_info, depth_multiplier);
+        const TensorShape      out_shape = compute_depthwise_convolution_shape(in_info, we_info, pad_stride_info, depth_multiplier, dilation);
 
         weights_shape.set(2, out_shape.z());
         const TensorShape biases_shape(weights_shape[2]);
 
-        _target    = compute_target(in_shape, weights_shape, biases_shape, out_shape, pad_stride_info, depth_multiplier, data_type, bias_data_type, quantization_info, data_layout);
-        _reference = compute_reference(in_shape, weights_shape, biases_shape, out_shape, pad_stride_info, depth_multiplier, data_type, bias_data_type, quantization_info);
+        _target    = compute_target(in_shape, weights_shape, biases_shape, out_shape, pad_stride_info, dilation, depth_multiplier, data_type, bias_data_type, quantization_info, data_layout);
+        _reference = compute_reference(in_shape, weights_shape, biases_shape, out_shape, pad_stride_info, dilation, depth_multiplier, data_type, bias_data_type, quantization_info);
     }
 
 protected:
@@ -104,7 +105,8 @@
         }
     }
 
-    TensorType compute_target(TensorShape input_shape, TensorShape weights_shape, TensorShape biases_shape, TensorShape output_shape, PadStrideInfo &pad_stride_info, unsigned int depth_multiplier,
+    TensorType compute_target(TensorShape input_shape, TensorShape weights_shape, TensorShape biases_shape, TensorShape output_shape, PadStrideInfo &pad_stride_info, Size2D dilation,
+                              unsigned int   depth_multiplier,
                               const DataType data_type, const DataType bias_data_type, const QuantizationInfo quantization_info, const DataLayout data_layout)
     {
         if(data_layout == DataLayout::NHWC)
@@ -122,7 +124,7 @@
 
         // Create Depthwise Convolution configure function
         FunctionType dwc;
-        dwc.configure(&src, &weights, &biases, &dst, pad_stride_info, depth_multiplier);
+        dwc.configure(&src, &weights, &biases, &dst, pad_stride_info, depth_multiplier, ActivationLayerInfo(), dilation);
 
         ARM_COMPUTE_EXPECT(src.info()->is_resizable(), framework::LogLevel::ERRORS);
         ARM_COMPUTE_EXPECT(weights.info()->is_resizable(), framework::LogLevel::ERRORS);
@@ -152,7 +154,7 @@
     }
 
     SimpleTensor<T> compute_reference(const TensorShape &in_shape, const TensorShape &weights_shape, const TensorShape &biases_shape, const TensorShape &out_shape, const PadStrideInfo &pad_stride_info,
-                                      unsigned int   depth_multiplier,
+                                      const Size2D &dilation, unsigned int depth_multiplier,
                                       const DataType data_type, const DataType bias_data_type, const QuantizationInfo quantization_info)
     {
         SimpleTensor<T>     src{ in_shape, data_type, 1, quantization_info };
@@ -163,7 +165,7 @@
         fill(weights, 1);
         fill(biases, 2);
 
-        return reference::depthwise_convolution(src, weights, biases, out_shape, pad_stride_info, depth_multiplier);
+        return reference::depthwise_convolution(src, weights, biases, out_shape, pad_stride_info, depth_multiplier, dilation);
     }
 
     TensorType       _target{};
@@ -177,9 +179,9 @@
 {
 public:
     template <typename...>
-    void setup(TensorShape in_shape, Size2D kernel_size, PadStrideInfo pad_stride_info, unsigned int depth_multiplier, DataType data_type, DataLayout data_layout)
+    void setup(TensorShape in_shape, Size2D kernel_size, PadStrideInfo pad_stride_info, Size2D dilation, unsigned int depth_multiplier, DataType data_type, DataLayout data_layout)
     {
-        DepthwiseConvolutionLayerValidationGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(in_shape, kernel_size, pad_stride_info, depth_multiplier,
+        DepthwiseConvolutionLayerValidationGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(in_shape, kernel_size, pad_stride_info, dilation, depth_multiplier,
                                                                                                             data_type, QuantizationInfo(), data_layout);
     }
 };
@@ -189,9 +191,10 @@
 {
 public:
     template <typename...>
-    void setup(TensorShape in_shape, Size2D kernel_size, PadStrideInfo pad_stride_info, unsigned int depth_multiplier, DataType data_type, QuantizationInfo quantization_info, DataLayout data_layout)
+    void setup(TensorShape in_shape, Size2D kernel_size, PadStrideInfo pad_stride_info, Size2D dilation, unsigned int depth_multiplier, DataType data_type, QuantizationInfo quantization_info,
+               DataLayout data_layout)
     {
-        DepthwiseConvolutionLayerValidationGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(in_shape, kernel_size, pad_stride_info, depth_multiplier,
+        DepthwiseConvolutionLayerValidationGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(in_shape, kernel_size, pad_stride_info, dilation, depth_multiplier,
                                                                                                             data_type, quantization_info, data_layout);
     }
 };
diff --git a/tests/validation/reference/DepthwiseConvolutionLayer.cpp b/tests/validation/reference/DepthwiseConvolutionLayer.cpp
index 122dbd4..90ecffb 100644
--- a/tests/validation/reference/DepthwiseConvolutionLayer.cpp
+++ b/tests/validation/reference/DepthwiseConvolutionLayer.cpp
@@ -70,13 +70,19 @@
     const int pad_right  = conv_info.pad_right();
     const int pad_bottom = conv_info.pad_bottom();
 
-    const int patch_half_width  = (filter_width + (dilation.x() - 1) * (filter_width - 1)) / 2;
-    const int patch_half_height = (filter_height + (dilation.y() - 1) * (filter_height - 1)) / 2;
+    const float patch_width  = (filter_width + (dilation.x() - 1) * (filter_width - 1));
+    const float patch_height = (filter_height + (dilation.y() - 1) * (filter_height - 1));
 
-    const int minimum_x = -pad_left + patch_half_width;
-    const int minimum_y = -pad_top + patch_half_height;
-    const int maximum_x = input_width + pad_left + pad_right - patch_half_width * 2;
-    const int maximum_y = input_height + pad_top + pad_bottom - patch_half_height * 2;
+    const int patch_half_width_floor  = patch_width / 2;
+    const int patch_half_height_floor = patch_height / 2;
+
+    const auto patch_half_width_ceil  = static_cast<int>(std::ceil(patch_width / 2));
+    const auto patch_half_height_ceil = static_cast<int>(std::ceil(patch_height / 2));
+
+    const int minimum_x = -pad_left + patch_half_width_floor;
+    const int minimum_y = -pad_top + patch_half_height_floor;
+    const int maximum_x = input_width + pad_left + pad_right - static_cast<int>(patch_width);
+    const int maximum_y = input_height + pad_top + pad_bottom - static_cast<int>(patch_height);
 
     const T border_value(0);
 
@@ -89,21 +95,20 @@
             {
                 const int out_z = z * depth_multiplier + m;
 
-                for(int y = minimum_y; y < minimum_y + maximum_y; y += conv_info.stride().second)
+                for(int y = minimum_y; y <= minimum_y + maximum_y; y += conv_info.stride().second)
                 {
-                    for(int x = minimum_x; x < minimum_x + maximum_x; x += conv_info.stride().first)
+                    for(int x = minimum_x; x <= minimum_x + maximum_x; x += conv_info.stride().first)
                     {
                         Coordinates coords(static_cast<int>(x), static_cast<int>(y), static_cast<int>(z), static_cast<int>(r));
                         size_t      filter_offset = filter_plane * out_z;
 
                         T val(0);
-                        for(int j = y - patch_half_height; j <= y + patch_half_height; j += dilation.y())
+                        for(int j = y - patch_half_height_floor; j < y + patch_half_height_ceil; j += dilation.y())
                         {
-                            for(int i = x - patch_half_width; i <= x + patch_half_width; i += dilation.x())
+                            for(int i = x - patch_half_width_floor; i < x + patch_half_width_ceil; i += dilation.x())
                             {
                                 coords.set(0, i);
                                 coords.set(1, j);
-
                                 val += *(weights.data() + filter_offset) * tensor_elem_at(src, coords, BorderMode::CONSTANT, border_value);
                                 ++filter_offset;
                             }
@@ -157,13 +162,19 @@
     const int pad_right  = conv_info.pad_right();
     const int pad_bottom = conv_info.pad_bottom();
 
-    const int patch_half_width  = (filter_width + (dilation.x() - 1) * (filter_width - 1)) / 2;
-    const int patch_half_height = (filter_height + (dilation.y() - 1) * (filter_height - 1)) / 2;
+    const float patch_width  = (filter_width + (dilation.x() - 1) * (filter_width - 1));
+    const float patch_height = (filter_height + (dilation.y() - 1) * (filter_height - 1));
 
-    const int minimum_x = -pad_left + patch_half_width;
-    const int minimum_y = -pad_top + patch_half_height;
-    const int maximum_x = input_width + pad_left + pad_right - patch_half_width * 2;
-    const int maximum_y = input_height + pad_top + pad_bottom - patch_half_height * 2;
+    const int patch_half_width_floor  = patch_width / 2;
+    const int patch_half_height_floor = patch_height / 2;
+
+    const auto patch_half_width_ceil  = static_cast<int>(std::ceil(patch_width / 2));
+    const auto patch_half_height_ceil = static_cast<int>(std::ceil(patch_height / 2));
+
+    const int minimum_x = -pad_left + patch_half_width_floor;
+    const int minimum_y = -pad_top + patch_half_height_floor;
+    const int maximum_x = input_width + pad_left + pad_right - static_cast<int>(patch_width);
+    const int maximum_y = input_height + pad_top + pad_bottom - static_cast<int>(patch_height);
 
     int out_pos = 0;
     for(int r = 0; r < num_batches; ++r)
@@ -175,17 +186,17 @@
                 const int     out_z    = z * depth_multiplier + m;
                 const int32_t bias_val = *static_cast<const int32_t *>(biases(Coordinates(out_z)));
 
-                for(int y = minimum_y; y < minimum_y + maximum_y; y += conv_info.stride().second)
+                for(int y = minimum_y; y <= minimum_y + maximum_y; y += conv_info.stride().second)
                 {
-                    for(int x = minimum_x; x < minimum_x + maximum_x; x += conv_info.stride().first)
+                    for(int x = minimum_x; x <= minimum_x + maximum_x; x += conv_info.stride().first)
                     {
                         Coordinates coords(x, y, z, r);
                         int         filter_offset = filter_plane * out_z;
 
                         int32_t val = 0;
-                        for(int j = y - patch_half_height; j <= y + patch_half_height; j += dilation.y())
+                        for(int j = y - patch_half_height_floor; j < y + patch_half_height_ceil; j += dilation.y())
                         {
-                            for(int i = x - patch_half_width; i <= x + patch_half_width; i += dilation.x())
+                            for(int i = x - patch_half_width_floor; i < x + patch_half_width_ceil; i += dilation.x())
                             {
                                 coords.set(0, i);
                                 coords.set(1, j);