COMPMID-2600: Implement a new and generic depthwise convolution for CL QASYMM8 NHWC

The NCHW case is supported at function level by permuting the
inputs/outputs to NHWC.

This patch also removes CLDirectConvolutionLayerOutputStageKernel which
is deprecated and some kernels which were only used in the generic case
of depthwise convolution.

Change-Id: I91e0f02d0a2f4a4a352e08c248e648944137fe68
Signed-off-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Reviewed-on: https://review.mlplatform.org/c/2056
Reviewed-by: Giorgio Arena <giorgio.arena@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>

24 files changed, 381 insertions, 1483 deletions

diff --git a/arm_compute/core/CL/CLKernels.h b/arm_compute/core/CL/CLKernels.h
index 3d9b2c81cd..c3c485db7c 100644
--- a/arm_compute/core/CL/CLKernels.h
+++ b/arm_compute/core/CL/CLKernels.h
@@ -56,15 +56,11 @@
 #include "arm_compute/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NCHWKernel.h"
 #include "arm_compute/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NHWCKernel.h"
 #include "arm_compute/core/CL/kernels/CLDepthwiseConvolutionLayerNativeKernel.h"
-#include "arm_compute/core/CL/kernels/CLDepthwiseConvolutionLayerReshapeWeightsGenericKernel.h"
 #include "arm_compute/core/CL/kernels/CLDepthwiseConvolutionLayerReshapeWeightsKernel.h"
-#include "arm_compute/core/CL/kernels/CLDepthwiseIm2ColKernel.h"
-#include "arm_compute/core/CL/kernels/CLDepthwiseVectorToTensorKernel.h"
 #include "arm_compute/core/CL/kernels/CLDequantizationLayerKernel.h"
 #include "arm_compute/core/CL/kernels/CLDerivativeKernel.h"
 #include "arm_compute/core/CL/kernels/CLDilateKernel.h"
 #include "arm_compute/core/CL/kernels/CLDirectConvolutionLayerKernel.h"
-#include "arm_compute/core/CL/kernels/CLDirectConvolutionLayerOutputStageKernel.h"
 #include "arm_compute/core/CL/kernels/CLElementWiseUnaryLayerKernel.h"
 #include "arm_compute/core/CL/kernels/CLElementwiseOperationKernel.h"
 #include "arm_compute/core/CL/kernels/CLErodeKernel.h"
diff --git a/arm_compute/core/CL/kernels/CLDepthwiseConvolutionLayerNativeKernel.h b/arm_compute/core/CL/kernels/CLDepthwiseConvolutionLayerNativeKernel.h
index f8c841ab6a..31ec871123 100644
--- a/arm_compute/core/CL/kernels/CLDepthwiseConvolutionLayerNativeKernel.h
+++ b/arm_compute/core/CL/kernels/CLDepthwiseConvolutionLayerNativeKernel.h
@@ -49,7 +49,7 @@ public:
     CLDepthwiseConvolutionLayerNativeKernel &operator=(CLDepthwiseConvolutionLayerNativeKernel &&) = default;
     /** Initialize the function's source, destination and parameters
      *
-     * @param[in]  input            Source tensor. Data type supported: FP32/FP16. Data layout supported: NHWC
+     * @param[in]  input            Source tensor. Data type supported: QASYMM8/FP32/FP16. Data layout supported: NHWC
      * @param[in]  weights          Weights tensor. A 3D tensor with dimensions [IFM, N, M]. Data type supported: Same as @p input.
      * @param[in]  biases           Biases tensor. A 1D tensor with dimensions [IFM]. Must be nullptr if not needed.
      *                              Data type supported: Same as @p input.
@@ -64,7 +64,7 @@ public:
                    const PadStrideInfo &conv_info, 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 CLDepthwiseConvolutionLayerNativeKernel
      *
-     * @param[in] input            Source tensor info. Data type supported: FP32/FP16. Data layout supported: NHWC
+     * @param[in] input            Source tensor info. Data type supported: QASYMM8/FP32/FP16. Data layout supported: NHWC
      * @param[in] weights          Weights tensor info. A 3D tensor with dimensions [IFM, N, M]. 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.
diff --git a/arm_compute/core/CL/kernels/CLDepthwiseConvolutionLayerReshapeWeightsGenericKernel.h b/arm_compute/core/CL/kernels/CLDepthwiseConvolutionLayerReshapeWeightsGenericKernel.h
deleted file mode 100644
index 3f969957e1..0000000000
--- a/arm_compute/core/CL/kernels/CLDepthwiseConvolutionLayerReshapeWeightsGenericKernel.h
+++ /dev/null
@@ -1,79 +0,0 @@
-/*
- * Copyright (c) 2017-2019 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_CLDEPTHWISECONVOLUTIONLAYERRESHAPEWEIGHTSGENERICKERNEL_H__
-#define __ARM_COMPUTE_CLDEPTHWISECONVOLUTIONLAYERRESHAPEWEIGHTSGENERICKERNEL_H__
-
-#include "arm_compute/core/CL/ICLKernel.h"
-
-namespace arm_compute
-{
-class ICLTensor;
-
-/** Interface for the depthwise weights reshape kernel.
- *  This kernel reshape original weights' low 2D dimensions into a single row and
- *  have the second dimension as the original depth size.
- *
- **/
-class CLDepthwiseConvolutionLayerReshapeWeightsGenericKernel : public ICLKernel
-{
-public:
-    /** Default constructor */
-    CLDepthwiseConvolutionLayerReshapeWeightsGenericKernel();
-    /** Prevent instances of this class from being copied (As this class contains pointers) */
-    CLDepthwiseConvolutionLayerReshapeWeightsGenericKernel(const CLDepthwiseConvolutionLayerReshapeWeightsGenericKernel &) = delete;
-    /** Prevent instances of this class from being copied (As this class contains pointers) */
-    CLDepthwiseConvolutionLayerReshapeWeightsGenericKernel &operator=(const CLDepthwiseConvolutionLayerReshapeWeightsGenericKernel &) = delete;
-    /** Allow instances of this class to be moved */
-    CLDepthwiseConvolutionLayerReshapeWeightsGenericKernel(CLDepthwiseConvolutionLayerReshapeWeightsGenericKernel &&) = default;
-    /** Allow instances of this class to be moved */
-    CLDepthwiseConvolutionLayerReshapeWeightsGenericKernel &operator=(CLDepthwiseConvolutionLayerReshapeWeightsGenericKernel &&) = default;
-    /** Set the input and output of the kernel.
-     *
-     * @param[in]  input  The input tensor to convert. 3 lower dimensions represent a single input [width, height, IFM].
-     *                    Data type supported: QASYMM8/F16/F32.
-     * @param[out] output The output tensor. Data type supported: same as @p input.
-     * @param[in]  biases (Optional) The input biases to add. Shape [IFM]. Data type supported: same as @p input.
-     */
-    void configure(const ICLTensor *input, ICLTensor *output, const ICLTensor *biases = nullptr);
-    /** Static function to check if given info will lead to a valid configuration of @ref CLDepthwiseConvolutionLayerReshapeWeightsGenericKernel
-     *
-     * @param[in] input  The input tensor to convert. 3 lower dimensions represent a single input [width, height, IFM].
-     *                   Data type supported: QASYMM8/F32.
-     * @param[in] output The output tensor. Data type supported: same as @p input.
-     * @param[in] biases (Optional) The input biases to add. Shape [IFM]. Data type supported: same as @p input.
-     *
-     * @return a status
-     */
-    static Status validate(const ITensorInfo *input, const ITensorInfo *output, const ITensorInfo *biases = nullptr);
-
-    // Inherited methods overridden:
-    void run(const Window &window, cl::CommandQueue &queue) override;
-
-private:
-    const ICLTensor *_input;
-    const ICLTensor *_biases;
-    ICLTensor       *_output;
-};
-} // arm_compute
-#endif /*__ARM_COMPUTE_CLDEPTHWISECONVOLUTIONLAYERRESHAPEWEIGHTSGENERICKERNEL_H__ */
diff --git a/arm_compute/core/CL/kernels/CLDepthwiseIm2ColKernel.h b/arm_compute/core/CL/kernels/CLDepthwiseIm2ColKernel.h
deleted file mode 100644
index 15798471a8..0000000000
--- a/arm_compute/core/CL/kernels/CLDepthwiseIm2ColKernel.h
+++ /dev/null
@@ -1,90 +0,0 @@
-/*
- * Copyright (c) 2017-2019 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_CLDEPTHWISEIM2COLKERNEL_H__
-#define __ARM_COMPUTE_CLDEPTHWISEIM2COLKERNEL_H__
-
-#include "arm_compute/core/CL/ICLKernel.h"
-#include "arm_compute/core/Size2D.h"
-
-namespace arm_compute
-{
-class ICLTensor;
-
-/** Interface for the depthwise im2col reshape kernel.
- *  This kernel reshape the input low 3 dimensions to a new 3D shape  where the output's first dimension is
- *  the linear patch size (FILTER_WIDTH * FILTER_HEIGHT) and second dimension is number of patches per image and third dimension unchanged .
- **/
-class CLDepthwiseIm2ColKernel : public ICLKernel
-{
-public:
-    /** Default constructor */
-    CLDepthwiseIm2ColKernel();
-    /** Prevent instances of this class from being copied (As this class contains pointers) */
-    CLDepthwiseIm2ColKernel(const CLDepthwiseIm2ColKernel &) = delete;
-    /** Prevent instances of this class from being copied (As this class contains pointers) */
-    CLDepthwiseIm2ColKernel &operator=(const CLDepthwiseIm2ColKernel &) = delete;
-    /** Allow instances of this class to be moved */
-    CLDepthwiseIm2ColKernel(CLDepthwiseIm2ColKernel &&) = default;
-    /** Allow instances of this class to be moved */
-    CLDepthwiseIm2ColKernel &operator=(CLDepthwiseIm2ColKernel &&) = default;
-    /** Set the input and output of the kernel.
-     *
-     * @param[in]  input            The input tensor 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/F16/F32
-     * @param[out] output           The output tensor. 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]  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,
-                   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 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 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 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,
-                           const Size2D &dilation = Size2D(1U, 1U));
-
-    // Inherited methods overridden:
-    void run(const Window &window, cl::CommandQueue &queue) override;
-
-private:
-    const ICLTensor *_input;
-    ICLTensor       *_output;
-};
-} // arm_compute
-#endif /*__ARM_COMPUTE_CLDEPTHWISEIM2COLKERNEL_H__ */
diff --git a/arm_compute/core/CL/kernels/CLDepthwiseVectorToTensorKernel.h b/arm_compute/core/CL/kernels/CLDepthwiseVectorToTensorKernel.h
deleted file mode 100644
index c9ec8e13bf..0000000000
--- a/arm_compute/core/CL/kernels/CLDepthwiseVectorToTensorKernel.h
+++ /dev/null
@@ -1,79 +0,0 @@
-/*
- * Copyright (c) 2017-2018 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_CLDEPTHWISEVECTORTOTENSORKERNEL_H__
-#define __ARM_COMPUTE_CLDEPTHWISEVECTORTOTENSORKERNEL_H__
-
-#include "arm_compute/core/CL/ICLKernel.h"
-
-namespace arm_compute
-{
-class ICLTensor;
-
-/** Interface for the depthwise vector to tensor kernel.
- *
- *  This kernel takes the 1D tensor that's been produced by the MatrixVectorMultiply
- *  kernel and reshapes it to given width and height (previously calculated, based
- *  on input/weights dimensions and convolution strides and padding).
- *
- **/
-class CLDepthwiseVectorToTensorKernel : public ICLKernel
-{
-public:
-    /** Default constructor */
-    CLDepthwiseVectorToTensorKernel();
-    /** Prevent instances of this class from being copied (As this class contains pointers) */
-    CLDepthwiseVectorToTensorKernel(const CLDepthwiseVectorToTensorKernel &) = delete;
-    /** Prevent instances of this class from being copied (As this class contains pointers) */
-    CLDepthwiseVectorToTensorKernel &operator=(const CLDepthwiseVectorToTensorKernel &) = delete;
-    /** Allow instances of this class to be moved */
-    CLDepthwiseVectorToTensorKernel(CLDepthwiseVectorToTensorKernel &&) = default;
-    /** Allow instances of this class to be moved */
-    CLDepthwiseVectorToTensorKernel &operator=(CLDepthwiseVectorToTensorKernel &&) = default;
-    /** Set the input and output of the kernel.
-     *
-     * @param[in]  input  The input vector to convert. Data type supported: QASYMM8/S32/F16/F32.
-     * @param[out] output The output tensor. 3 lower dimensions represent a single input [width, height, IFM]. Data type supported: same as @p input.
-     * @param[in]  conv_w The converted tensor's width.
-     * @param[in]  conv_h The converted tensor's height.
-     */
-    void configure(const ICLTensor *input, ICLTensor *output, size_t conv_w, size_t conv_h);
-    /** Static function to check if given info will lead to a valid configuration of @ref CLDepthwiseVectorToTensorKernel
-     *
-     * @param[in] input  The input vector to convert. Data type supported: QASYMM8/S32/F16/F32.
-     * @param[in] output The output tensor. 3 lower dimensions represent a single input [width, height, IFM]. Data type supported: same as @p input.
-     * @param[in] conv_w The converted tensor's width.
-     * @param[in] conv_h The converted tensor's height.
-     *
-     * @return a status
-     */
-    static Status validate(const ITensorInfo *input, const ITensorInfo *output, size_t conv_w, size_t conv_h);
-    // Inherited methods overridden:
-    void run(const Window &window, cl::CommandQueue &queue) override;
-
-private:
-    const ICLTensor *_input;
-    ICLTensor       *_output;
-};
-} // arm_compute
-#endif /*__ARM_COMPUTE_CLDEPTHWISEVECTORTOTENSORKERNEL_H__ */
diff --git a/arm_compute/core/CL/kernels/CLDirectConvolutionLayerOutputStageKernel.h b/arm_compute/core/CL/kernels/CLDirectConvolutionLayerOutputStageKernel.h
deleted file mode 100644
index 80bc012d9f..0000000000
--- a/arm_compute/core/CL/kernels/CLDirectConvolutionLayerOutputStageKernel.h
+++ /dev/null
@@ -1,91 +0,0 @@
-/*
- * Copyright (c) 2018-2019 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_CLDIRECTCONVOLUTIONLAYEROUTPUTSTAGEKERNEL_H__
-#define __ARM_COMPUTE_CLDIRECTCONVOLUTIONLAYEROUTPUTSTAGEKERNEL_H__
-
-#include "arm_compute/core/CL/ICLKernel.h"
-
-namespace arm_compute
-{
-class ITensor;
-/** OpenCL kernel to accumulate the biases, if provided, or downscale in case of quantized input.
- *
- * @deprecated This kernel is deprecated and will be removed in release 19.05
- *
- * @note We assume bias to be shared
- *
- */
-class CLDirectConvolutionLayerOutputStageKernel : public ICLKernel
-{
-public:
-    /** Default constructor */
-    CLDirectConvolutionLayerOutputStageKernel();
-    /** Prevent instances of this class from being copied (As this class contains pointers) */
-    CLDirectConvolutionLayerOutputStageKernel(const CLDirectConvolutionLayerOutputStageKernel &) = delete;
-    /** Prevent instances of this class from being copied (As this class contains pointers) */
-    CLDirectConvolutionLayerOutputStageKernel &operator=(const CLDirectConvolutionLayerOutputStageKernel &) = delete;
-    /** Allow instances of this class to be moved */
-    CLDirectConvolutionLayerOutputStageKernel(CLDirectConvolutionLayerOutputStageKernel &&) = default;
-    /** Allow instances of this class to be moved */
-    CLDirectConvolutionLayerOutputStageKernel &operator=(CLDirectConvolutionLayerOutputStageKernel &&) = default;
-    /** Default destructor */
-    ~CLDirectConvolutionLayerOutputStageKernel() = default;
-    /** Set the accumulate buffer and the biases of the kernel.
-     *
-     * @param[in, out] input                        Input to add the bias to. If @p output is not specified then accumulation is done in-place.
-     *                                              Data type supported: S32/F16/F32
-     * @param[in]      bias                         (Optional) The shared bias tensor to add. It must be 1D Tensor. Data type supported: Same as @p input
-     * @param[out]     output                       (Optional) If the output tensor is specified the accumulation is done out-of-place. (Defaults to nullptr)
-     *                                              Required parameter if output is of QASYMM8 type.
-     *                                              Data types supported: QASYMM8/F16/F32
-     * @param[in]      result_fixedpoint_multiplier (Optional)Fixed point value to be multiplied to each element of the input matrix when once the result_offset has been add
-     * @param[in]      result_shift                 (Optional)Integer value used to round to nearest division by a power-of-two the result after the fixed point multiplication
-     * @param[in]      result_offset_after_shift    (Optional)Offset to be applied to result before converting it back to QASYMM8
-     */
-    void configure(ICLTensor *input, const ICLTensor *bias = nullptr, ICLTensor *output = nullptr,
-                   int result_fixedpoint_multiplier = 0, int result_shift = 0, int result_offset_after_shift = 0);
-    /** Static function to check if given info will lead to a valid configuration of @ref CLDirectConvolutionLayerOutputStageKernel
-     *
-     * @param[in] input  Input to add the bias to. If @p output is not specified then accumulation is done in-place.
-     *                   Data type supported: F16/F32
-     * @param[in] bias   (Optional) The shared bias tensor to add. It must be 1D Tensor. Data type supported: Same as @p input
-     * @param[in] output (Optional) If the output tensor is specified the accumulation is done out-of-place. (Defaults to nullptr)
-     *                         Data type supported: F16/F32
-     * @return a status
-     */
-    static Status validate(const ITensorInfo *input, const ITensorInfo *bias = nullptr, const ITensorInfo *output = nullptr);
-
-    // Inherited methods overridden:
-    void run(const Window &window, cl::CommandQueue &queue) override;
-
-private:
-    ICLTensor       *_input;
-    const ICLTensor *_bias;
-    ICLTensor       *_output;
-    int              _result_fixedpoint_multiplier;
-    int              _result_shift;
-    int              _result_offset_after_shift;
-};
-} // namespace arm_compute
-#endif /*__ARM_COMPUTE_CLDIRECTCONVOLUTIONLAYEROUTPUTSTAGEKERNEL_H__ */
diff --git a/arm_compute/runtime/CL/functions/CLDepthwiseConvolutionLayer.h b/arm_compute/runtime/CL/functions/CLDepthwiseConvolutionLayer.h
index d177f4505a..98581a21fe 100644
--- a/arm_compute/runtime/CL/functions/CLDepthwiseConvolutionLayer.h
+++ b/arm_compute/runtime/CL/functions/CLDepthwiseConvolutionLayer.h
@@ -26,17 +26,12 @@
 
 #include "arm_compute/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NCHWKernel.h"
 #include "arm_compute/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NHWCKernel.h"
-#include "arm_compute/core/CL/kernels/CLDepthwiseConvolutionLayerReshapeWeightsGenericKernel.h"
+#include "arm_compute/core/CL/kernels/CLDepthwiseConvolutionLayerNativeKernel.h"
 #include "arm_compute/core/CL/kernels/CLDepthwiseConvolutionLayerReshapeWeightsKernel.h"
-#include "arm_compute/core/CL/kernels/CLDepthwiseIm2ColKernel.h"
-#include "arm_compute/core/CL/kernels/CLDepthwiseVectorToTensorKernel.h"
-#include "arm_compute/core/CL/kernels/CLDirectConvolutionLayerOutputStageKernel.h"
 #include "arm_compute/core/CL/kernels/CLFillBorderKernel.h"
-#include "arm_compute/core/CL/kernels/CLGEMMMatrixVectorMultiplyKernel.h"
 #include "arm_compute/core/CL/kernels/ICLDepthwiseConvolutionLayer3x3Kernel.h"
 #include "arm_compute/core/Types.h"
 #include "arm_compute/runtime/CL/CLTensor.h"
-#include "arm_compute/runtime/CL/functions/CLActivationLayer.h"
 #include "arm_compute/runtime/CL/functions/CLPermute.h"
 #include "arm_compute/runtime/IFunction.h"
 #include "arm_compute/runtime/MemoryGroup.h"
@@ -121,17 +116,15 @@ private:
 
 /** Basic function to execute a generic depthwise convolution. This function calls the following OpenCL kernels:
  *
- * -# @ref CLDepthwiseIm2ColKernel
- * -# @ref CLGEMMMatrixVectorMultiplyKernel
- * -# @ref CLDepthwiseConvolutionLayerReshapeWeightsGenericKernel
- * -# @ref CLFillBorderKernel (if pad_x or pad_y > 0)
+ * -# @ref CLDepthwiseConvolutionLayerNativeKernel
+ * -# @ref CLPermute (x 3) if the data layout is NCHW
  *
  */
 class CLDepthwiseConvolutionLayer : public IFunction
 {
 public:
     /** Default constructor */
-    CLDepthwiseConvolutionLayer();
+    CLDepthwiseConvolutionLayer(std::shared_ptr<IMemoryManager> memory_manager = nullptr);
     /** Prevent instances of this class from being copied (As this class contains pointers) */
     CLDepthwiseConvolutionLayer(const CLDepthwiseConvolutionLayer &) = delete;
     /** Default move constructor */
@@ -177,23 +170,21 @@ public:
     void prepare() override;
 
 private:
-    CLDepthwiseIm2ColKernel                                _im2col_kernel;
-    CLDepthwiseConvolutionLayerReshapeWeightsGenericKernel _weights_reshape_kernel;
-    CLGEMMMatrixVectorMultiplyKernel                       _v2mm_kernel;
-    CLDepthwiseVectorToTensorKernel                        _vector_to_tensor_kernel;
-    CLDirectConvolutionLayerOutputStageKernel              _output_stage_kernel;
-    CLActivationLayer                                      _activationlayer_function;
-    CLFillBorderKernel                                     _v2mm_input_fill_border;
-    CLFillBorderKernel                                     _v2mm_weights_fill_border;
-    CLTensor                                               _input_reshaped;
-    CLTensor                                               _weights_reshaped;
-    CLTensor                                               _v2mm_output;
-    CLTensor                                               _output_reshaped;
-    bool                                                   _is_prepared;
-    bool                                                   _is_quantized;
-    bool                                                   _is_activationlayer_enabled;
-    const ICLTensor                                       *_original_weights;
-    std::unique_ptr<IFunction>                             _optimised_function;
+    MemoryGroup _memory_group;
+
+    std::unique_ptr<IFunction>              _optimised_function;
+    CLDepthwiseConvolutionLayerNativeKernel _dwc_native_kernel;
+    CLPermute                               _permute_input_to_nhwc;
+    CLPermute                               _permute_weights_to_nhwc;
+    CLPermute                               _permute_output_to_nchw;
+
+    CLTensor       _permuted_input;
+    CLTensor       _permuted_weights;
+    CLTensor       _permuted_output;
+    const ITensor *_original_weights;
+
+    bool _needs_permute;
+    bool _is_prepared;
 };
 } // namespace arm_compute
 #endif /*__ARM_COMPUTE_CLDEPTHWISECONVOLUTION_H__ */
diff --git a/docs/00_introduction.dox b/docs/00_introduction.dox
index 6e014e39c9..1210b2bdec 100644
--- a/docs/00_introduction.dox
+++ b/docs/00_introduction.dox
@@ -236,6 +236,13 @@ If there is more than one release in a month then an extra sequential number is
 
 @subsection S2_2_changelog Changelog
 
+v19.11 Public major release
+ - Deprecated OpenCL kernels / functions
+    - CLDepthwiseConvolutionLayerReshapeWeightsGenericKernel
+    - CLDepthwiseIm2ColKernel
+    - CLDepthwiseVectorToTensorKernel
+    - CLDirectConvolutionLayerOutputStageKernel
+
 v19.08 Public major release
  - Various bug fixes.
  - Various optimisations.
@@ -624,7 +631,7 @@ v18.02 Public major release
  - Added fused batched normalization and activation to @ref CLBatchNormalizationLayer and @ref NEBatchNormalizationLayer
  - Added support for non-square pooling to @ref NEPoolingLayer and @ref CLPoolingLayer
  - New OpenCL kernels / functions:
-    - @ref CLDirectConvolutionLayerOutputStageKernel
+    - CLDirectConvolutionLayerOutputStageKernel
  - New NEON kernels / functions
     - Added name() method to all kernels.
     - Added support for Winograd 5x5.
@@ -746,7 +753,7 @@ v17.09 Public major release
     - @ref NEReshapeLayerKernel / @ref NEReshapeLayer
 
  - New OpenCL kernels / functions:
-    - @ref CLDepthwiseConvolutionLayer3x3NCHWKernel @ref CLDepthwiseConvolutionLayer3x3NHWCKernel @ref CLDepthwiseIm2ColKernel @ref CLDepthwiseVectorToTensorKernel CLDepthwiseWeightsReshapeKernel / @ref CLDepthwiseConvolutionLayer3x3 @ref CLDepthwiseConvolutionLayer CLDepthwiseSeparableConvolutionLayer
+    - @ref CLDepthwiseConvolutionLayer3x3NCHWKernel @ref CLDepthwiseConvolutionLayer3x3NHWCKernel CLDepthwiseIm2ColKernel CLDepthwiseVectorToTensorKernel CLDepthwiseWeightsReshapeKernel / @ref CLDepthwiseConvolutionLayer3x3 @ref CLDepthwiseConvolutionLayer CLDepthwiseSeparableConvolutionLayer
     - @ref CLDequantizationLayerKernel / @ref CLDequantizationLayer
     - @ref CLDirectConvolutionLayerKernel / @ref CLDirectConvolutionLayer
     - @ref CLFlattenLayer
diff --git a/src/core/CL/CLKernelLibrary.cpp b/src/core/CL/CLKernelLibrary.cpp
index fa5193fde2..7b7263fca7 100644
--- a/src/core/CL/CLKernelLibrary.cpp
+++ b/src/core/CL/CLKernelLibrary.cpp
@@ -222,6 +222,7 @@ const std::map<std::string, std::string> CLKernelLibrary::_kernel_program_map =
     { "depthwise_convolution_3x3_nhwc", "depthwise_convolution.cl" },
     { "depthwise_convolution_3x3_nhwc_stride1", "depthwise_convolution.cl" },
     { "dwc_MxN_native_fp_nhwc", "depthwise_convolution.cl" },
+    { "dwc_MxN_native_quantized8_nhwc", "depthwise_convolution_quantized.cl" },
     { "dwc_3x3_native_qasymm8_nchw", "depthwise_convolution_quantized.cl" },
     { "dwc_3x3_native_qasymm8_dot8_nchw", "depthwise_convolution_quantized.cl" },
     { "dwc_3x3_reshaped_qasymm8_nhwc", "depthwise_convolution_quantized.cl" },
@@ -234,9 +235,6 @@ const std::map<std::string, std::string> CLKernelLibrary::_kernel_program_map =
     { "depthwise_convolution_3x3_stridex1_stridey1_bifrost_f32", "depthwise_convolution.cl" },
     { "depthwise_convolution_3x3_stridex2_stridey2_bifrost_f32", "depthwise_convolution.cl" },
     { "depthwise_convolution_reshape_weights", "depthwise_convolution.cl" },
-    { "depthwise_convolution_reshape_weights_generic", "depthwise_convolution.cl" },
-    { "depthwise_im2col", "depthwise_convolution.cl" },
-    { "depthwise_vector_to_tensor", "depthwise_convolution.cl" },
     { "dequantization_layer", "dequantization_layer.cl" },
     { "dequantization_layer_per_channel_nhwc", "dequantization_layer.cl" },
     { "dequantization_layer_per_channel_nchw", "dequantization_layer.cl" },
@@ -413,7 +411,6 @@ const std::map<std::string, std::string> CLKernelLibrary::_kernel_program_map =
     { "NV21_to_RGB888_bt709", "color_convert.cl" },
     { "NV21_to_RGBA8888_bt709", "color_convert.cl" },
     { "NV21_to_YUV444_bt709", "color_convert.cl" },
-    { "output_stage_quantized", "direct_convolution_1x1_3x3_5x5_quantized.cl" },
     { "pad_layer_constant", "pad_layer.cl" },
     { "pad_layer_symmetric_reflect", "pad_layer.cl" },
     { "permute", "permute.cl" },
diff --git a/src/core/CL/cl_kernels/activation_quant_helpers.h b/src/core/CL/cl_kernels/activation_quant_helpers.h
index 402e7ac41f..0e4eb2b32e 100644
--- a/src/core/CL/cl_kernels/activation_quant_helpers.h
+++ b/src/core/CL/cl_kernels/activation_quant_helpers.h
@@ -41,7 +41,7 @@ inline TYPE relu_op(TYPE x)
 // Bounded RELU Activation
 inline TYPE brelu_op(TYPE x)
 {
-    return min((TYPE)A_VAL, max(CONST_0, x));
+    return min((TYPE)A_VAL, max((TYPE)CONST_0, x));
 }
 // Lower Upper Bounded RELU Activation
 inline TYPE lu_brelu_op(TYPE x)
diff --git a/src/core/CL/cl_kernels/depthwise_convolution.cl b/src/core/CL/cl_kernels/depthwise_convolution.cl
index 1b2f5cccaa..3a227282ff 100644
--- a/src/core/CL/cl_kernels/depthwise_convolution.cl
+++ b/src/core/CL/cl_kernels/depthwise_convolution.cl
@@ -782,173 +782,6 @@ __kernel void depthwise_convolution_reshape_weights(
 }
 #endif // defined(VEC_SIZE) && defined(DATA_TYPE) && defined(DST_WIDTH)
 
-#if defined(NCHW)
-#define in_stride_x src_stride_x
-#define in_stride_y src_stride_y
-#define in_stride_z src_stride_z
-#define out_stride_x dst_stride_x
-#define out_stride_y dst_stride_y
-#define out_stride_z dst_stride_z
-#else //defined(NCHW)
-#define in_stride_x src_stride_y
-#define in_stride_y src_stride_z
-#define in_stride_z src_stride_x
-#define out_stride_x dst_stride_y
-#define out_stride_y dst_stride_z
-#define out_stride_z dst_stride_x
-#endif //defined(NCHW)
-
-#if defined(SRC_WIDTH) && defined(DATA_TYPE)
-/** This kernel reshapes each of the tensor's low three dimensions to single rows.
- *
- * @note Datatype and source width should be given as a preprocessor argument using -DDATA_TYPE=type and -DSRC_WIDTH=width. e.g. -DSRC_WIDTH=128
- *
- * @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)
- * @param[in]  src_step_x                           src_stride_x * number of elements along X processed per workitem(in bytes)
- * @param[in]  src_stride_y                         Stride of the source tensor in Y dimension (in bytes)
- * @param[in]  src_step_y                           src_stride_y * number of elements along Y processed per workitem(in bytes)
- * @param[in]  src_stride_z                         Stride of the source tensor in Z dimension (in bytes)
- * @param[in]  src_step_z                           src_stride_z * number of elements along Y processed per workitem(in bytes)
- * @param[in]  src_offset_first_element_in_bytes    The offset of the first element in the source tensor
- * @param[out] dst_ptr                              Pointer to the destination tensor. Same as @p src_ptr
- * @param[in]  dst_stride_x                         Stride of the destination tensor in X dimension (in bytes)
- * @param[in]  dst_step_x                           dst_stride_x * number of elements along X processed per workitem(in bytes)
- * @param[in]  dst_stride_y                         Stride of the destination tensor in Y dimension (in bytes)
- * @param[in]  dst_step_y                           dst_stride_y * number of elements along Y processed per workitem(in bytes)
- * @param[in]  dst_offset_first_element_in_bytes    The offset of the first element in the destination tensor
- * @param[in]  biases_ptr                           (Optional) Pointer to the biases vector. Supported data types: F16/F32
- * @param[in]  biases_stride_x                      (Optional) Stride of the biases vector in X dimension (in bytes)
- * @param[in]  biases_step_x                        (Optional) biases_stride_x * number of elements along X processed per workitem(in bytes)
- * @param[in]  biases_offset_first_element_in_bytes (Optional) The offset of the first element in the biases vector
- */
-__kernel void depthwise_convolution_reshape_weights_generic(
-    TENSOR3D_DECLARATION(src),
-    IMAGE_DECLARATION(dst)
-#ifdef HAS_BIAS
-    ,
-    VECTOR_DECLARATION(biases)
-#endif /* HAS_BIAS */
-)
-{
-#ifdef HAS_BIAS
-    Vector biases = CONVERT_TO_VECTOR_STRUCT_NO_STEP(biases);
-#endif /* HAS_BIAS */
-
-    __global uchar *input_ptr  = src_ptr + src_offset_first_element_in_bytes + get_global_id(1) * in_stride_y + get_global_id(2) * in_stride_z;
-    __global uchar *output_ptr = dst_ptr + dst_offset_first_element_in_bytes + get_global_id(1) * SRC_WIDTH * dst_stride_x + get_global_id(2) * dst_stride_y;
-
-    for(int i = 0; i < SRC_WIDTH; ++i, input_ptr += in_stride_x)
-    {
-        *((__global DATA_TYPE *)(output_ptr + i * dst_stride_x)) = *((__global DATA_TYPE *)input_ptr);
-    }
-
-#if defined(HAS_BIAS)
-    if(get_global_id(1) == 0)
-    {
-        *((__global DATA_TYPE *)(output_ptr + SRC_WIDTH * get_global_size(1) * dst_stride_x)) = *((__global DATA_TYPE *)(biases.ptr + get_global_id(2) * biases_stride_x));
-    }
-#endif // defined(HAS_BIAS)
-}
-#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) && 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)
- * @param[in]  src_step_x                        src_stride_x * number of elements along X processed per workitem(in bytes)
- * @param[in]  src_stride_y                      Stride of the source tensor in Y dimension (in bytes)
- * @param[in]  src_step_y                        src_stride_y * number of elements along Y processed per workitem(in bytes)
- * @param[in]  src_stride_z                      Stride of the source tensor in Z dimension (in bytes)
- * @param[in]  src_step_z                        src_stride_z * number of elements along Z processed per workitem(in bytes)
- * @param[in]  src_offset_first_element_in_bytes The offset of the first element in the source tensor
- * @param[out] dst_ptr                           Pointer to the destination tensor. Supported data types: same as @p src_ptr
- * @param[in]  dst_stride_x                      Stride of the destination tensor in X dimension (in bytes)
- * @param[in]  dst_step_x                        dst_stride_x * number of elements along X processed per workitem(in bytes)
- * @param[in]  dst_stride_y                      Stride of the destination tensor in Y dimension (in bytes)
- * @param[in]  dst_step_y                        dst_stride_y * number of elements along Y processed per workitem(in bytes)
- * @param[in]  dst_stride_z                      Stride of the destination tensor in Z dimension (in bytes)
- * @param[in]  dst_step_z                        dst_stride_z * number of elements along Z processed per workitem(in bytes)
- * @param[in]  dst_offset_first_element_in_bytes The offset of the first element in the destination tensor
- */
-__kernel void depthwise_im2col(TENSOR3D_DECLARATION(src), TENSOR3D_DECLARATION(dst))
-{
-    Tensor3D dst = CONVERT_TO_TENSOR3D_STRUCT(dst);
-
-    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 + (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;
-    const int src_z = get_global_id(2) / DEPTH_MULTIPLIER;
-
-    __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 + (KERNEL_HEIGHT - 1) * (DILATION_Y - 1); y += DILATION_Y)
-    {
-        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)
-            {
-                *output_ptr = PAD_VALUE;
-            }
-            else
-            {
-                *output_ptr = *((__global DATA_TYPE *)(input_ptr + x * in_stride_x + y * in_stride_y));
-            }
-        }
-    }
-#if defined(HAS_BIAS)
-    *output_ptr = (DATA_TYPE)(1);
-#endif // defined(HAS_BIAS)
-}
-
-#endif //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(DATA_TYPE) && defined(PAD_VALUE) && defined(DEPTH_MULTIPLIER)
-
-#if defined(CONV_WIDTH) && defined(CONV_HEIGHT) && defined(DATA_TYPE)
-
-/** This kernel performs a reshaping of the output of the depthwise generic convolution.
- *
- * @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 -DCONV_WIDTH, -DCONV_HEIGHT, e.g -DCONV_WIDTH=32, -DCONV_HEIGHT=42
- *
- * @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)
- * @param[in]  src_step_x                        src_stride_x * number of elements along X processed per workitem(in bytes)
- * @param[in]  src_offset_first_element_in_bytes The offset of the first element in the source tensor
- * @param[out] dst_ptr                           Pointer to the destination tensor. Supported data types: same as @p src_ptr
- * @param[in]  dst_stride_x                      Stride of the destination tensor in X dimension (in bytes)
- * @param[in]  dst_step_x                        dst_stride_x * number of elements along X processed per workitem(in bytes)
- * @param[in]  dst_stride_y                      Stride of the destination tensor in Y dimension (in bytes)
- * @param[in]  dst_step_y                        dst_stride_y * number of elements along Y processed per workitem(in bytes)
- * @param[in]  dst_stride_z                      Stride of the destination tensor in Z dimension (in bytes)
- * @param[in]  dst_step_z                        dst_stride_z * number of elements along Z processed per workitem(in bytes)
- * @param[in]  dst_offset_first_element_in_bytes The offset of the first element in the destination tensor
- */
-__kernel void depthwise_vector_to_tensor(
-    VECTOR_DECLARATION(src),
-    TENSOR3D_DECLARATION(dst))
-{
-    Vector src = CONVERT_TO_VECTOR_STRUCT(src);
-
-    const int patch_size = CONV_WIDTH * CONV_HEIGHT;
-    const int id0        = get_global_id(0);
-    const int z          = id0 / patch_size;
-    const int index2D    = id0 - z * patch_size;
-
-    __global uchar *out_ptr          = dst_ptr + dst_offset_first_element_in_bytes + index2D % CONV_WIDTH * out_stride_x + index2D / CONV_WIDTH * out_stride_y + z * out_stride_z;
-    *((__global DATA_TYPE *)out_ptr) = *((__global DATA_TYPE *)src.ptr);
-}
-
-#endif //defined(CONV_WIDTH) && defined(CONV_HEIGHT) && defined(DATA_TYPE)
-
 #if defined(ARM_COMPUTE_OPENCL_FP16_ENABLED) && defined(DEPTH_MULTIPLIER) && defined(DST_CHANNELS) && defined(IS_F16)
 #if defined(CONV_STRIDE_X)
 #if CONV_STRIDE_X == 1
@@ -1478,7 +1311,7 @@ __kernel void depthwise_convolution_3x3_stridex2_stridey2_bifrost_f16(
     //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);
+    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
@@ -1556,23 +1389,17 @@ __kernel void dwc_MxN_native_fp_nhwc(
     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
+    int z                   = get_global_id(2); // spatial coordinate y
 #endif                                         // defined(DST_DEPTH)
 
-    __global uchar *s_addr = src_ptr + src_offset_first_element_in_bytes +
-                             x * sizeof(DATA_TYPE) * (int)N0;
+    __global uchar *s_addr = src_ptr + src_offset_first_element_in_bytes + x * sizeof(DATA_TYPE) * (int)N0;
 
-    __global uchar *d_addr = dst_ptr + dst_offset_first_element_in_bytes +
-                             x * sizeof(DATA_TYPE) * (int)DEPTH_MULTIPLIER * (int)N0 +
-                             y * dst_stride_y +
-                             z * dst_stride_z;
+    __global uchar *d_addr = dst_ptr + dst_offset_first_element_in_bytes + x * sizeof(DATA_TYPE) * (int)DEPTH_MULTIPLIER * (int)N0 + y * dst_stride_y + z * dst_stride_z;
 
-    __global uchar *w_addr = weights_ptr + weights_offset_first_element_in_bytes +
-                             x * sizeof(DATA_TYPE) * (int)DEPTH_MULTIPLIER * (int)N0;
+    __global uchar *w_addr = weights_ptr + weights_offset_first_element_in_bytes + x * sizeof(DATA_TYPE) * (int)DEPTH_MULTIPLIER * (int)N0;
 
 #if defined(HAS_BIAS)
-        __global uchar *b_addr = biases_ptr + biases_offset_first_element_in_bytes +
-                                 x * sizeof(DATA_TYPE) * (int)DEPTH_MULTIPLIER * (int)N0;
+    __global uchar *b_addr = biases_ptr + biases_offset_first_element_in_bytes + x * sizeof(DATA_TYPE) * (int)DEPTH_MULTIPLIER * (int)N0;
 #endif // defined(HAS_BIAS)
 
 #if defined(DST_DEPTH)
@@ -1611,7 +1438,7 @@ __kernel void dwc_MxN_native_fp_nhwc(
 #if GPU_ARCH == GPU_ARCH_MIDGARD
                         res += i * w;
 #else  // GPU_ARCH == GPU_ARCH_MIDGARD
-                        res  = fma(i, w, res);
+                        res = fma(i, w, res);
 #endif // GPU_ARCH == GPU_ARCH_MIDGARD
                     }
                     x_coord_tmp += DILATION_X;
diff --git a/src/core/CL/cl_kernels/depthwise_convolution_quantized.cl b/src/core/CL/cl_kernels/depthwise_convolution_quantized.cl
index 8f2e441693..10872d460a 100644
--- a/src/core/CL/cl_kernels/depthwise_convolution_quantized.cl
+++ b/src/core/CL/cl_kernels/depthwise_convolution_quantized.cl
@@ -24,19 +24,30 @@
 
 #include "helpers_asymm.h"
 
-#if defined(WEIGHTS_OFFSET) && defined(INPUT_OFFSET) && defined(K_OFFSET) && ((defined(OUTPUT_OFFSET) && defined(OUTPUT_MULTIPLIER) && defined(OUTPUT_SHIFT)) || defined(REAL_MULTIPLIER))
-
-#if defined(ACTIVATION_TYPE) && defined(CONST_0)
-#define DATA_TYPE uchar
 #ifndef VEC_SIZE
+#if defined(N0)
+#define VEC_SIZE N0
+#else /* defined(N0) */
 #define VEC_SIZE 8
+#endif /* defined(N0) */
 #endif /* VEC_SIZE */
+
+#if defined(ACTIVATION_TYPE) && defined(CONST_0)
+#define DATA_TYPE uchar
 #include "activation_layer_quant.cl"
 #define ACTIVATION_FUNC(x) PERFORM_ACTIVATION_QUANT(ACTIVATION_TYPE, x)
 #else /* defined(ACTIVATION_TYPE) && defined(CONST_0) */
 #define ACTIVATION_FUNC(x) (x)
 #endif /* defined(ACTIVATION_TYPE) && defined(CONST_0) */
 
+#define VEC_INT VEC_DATA_TYPE(int, VEC_SIZE)
+#define VEC_FLOAT VEC_DATA_TYPE(float, VEC_SIZE)
+#define VEC_UCHAR VEC_DATA_TYPE(uchar, VEC_SIZE)
+#define VEC_USHORT VEC_DATA_TYPE(ushort, VEC_SIZE)
+#define VEC_SHORT VEC_DATA_TYPE(short, VEC_SIZE)
+
+#if defined(WEIGHTS_OFFSET) && defined(INPUT_OFFSET) && defined(K_OFFSET) && ((defined(OUTPUT_OFFSET) && defined(OUTPUT_MULTIPLIER) && defined(OUTPUT_SHIFT)) || defined(REAL_MULTIPLIER))
+
 #if defined(ARM_COMPUTE_OPENCL_DOT8_ENABLED) && defined(cl_arm_integer_dot_product_int8)
 #if defined(ARM_COMPUTE_OPENCL_DOT8_ACC_ENABLED) && defined(cl_arm_integer_dot_product_accumulate_int8)
 #define ARM_DOT(x, y, val) val = arm_dot_acc((x), (y), val);
@@ -635,11 +646,6 @@ __kernel void dwc_3x3_native_qasymm8_dot8_nchw(
 
 #define asymm_mult_by_quant_multiplier_less_than_one(x, y, z) ASYMM_MULT_BY_QUANT_MULTIPLIER_LESS_THAN_ONE(x, y, z, VEC_SIZE)
 
-#define VEC_INT VEC_DATA_TYPE(int, VEC_SIZE)
-#define VEC_FLOAT VEC_DATA_TYPE(float, VEC_SIZE)
-#define VEC_UCHAR VEC_DATA_TYPE(uchar, VEC_SIZE)
-#define VEC_USHORT VEC_DATA_TYPE(ushort, VEC_SIZE)
-
 #define MULTIPLY_ADD(x, y, acc) acc += CONVERT(CONVERT(x, VEC_USHORT) * CONVERT(y, VEC_USHORT), VEC_INT)
 
 #if WEIGHTS_OFFSET != 0
@@ -1375,3 +1381,144 @@ __kernel void dwc_3x3_reshaped_qasymm8_dot8_stride1_nhwc(
 #endif // defined(VEC_SIZE) && defined(SRC_DIM_1) && defined(SRC_DIM_2) && defined(CONV_PAD_TOP) && defined(CONV_PAD_LEFT)
 
 #endif // defined(WEIGHTS_OFFSET) && defined(INPUT_OFFSET) && defined(K_OFFSET) && ((defined(OUTPUT_OFFSET) && defined(OUTPUT_MULTIPLIER) && defined(OUTPUT_SHIFT)) || defined(REAL_MULTIPLIER))
+
+#if defined(SRC_DIM1) && defined(SRC_DIM2) && defined(KERNEL_WIDTH) && defined(KERNEL_HEIGHT) && defined(N0) && defined(DILATION_X) && defined(DILATION_Y) && defined(CONV_STRIDE_X) && defined(CONV_STRIDE_Y) && defined(CONV_PAD_LEFT) && defined(CONV_PAD_TOP) && defined(INPUT_OFFSET) && defined(WEIGHTS_OFFSET) && defined(OUTPUT_OFFSET) && defined(OUTPUT_MULTIPLIER) && defined(OUTPUT_SHIFT)
+/** This function computes the depthwise convolution for NHWC data layout. This kernel assumes that the weights tensor is NOT reshaped
+ *
+ * @note The number of elements processed must be passed at compile time using -DN0 (e.g. -DN0=2)
+ * @note The depth multiplier must be passed at compile time using -DDEPTH_MULTIPLIER (e.g. -DDEPTH_MULTIPLIER=1)
+ * @note The first dimension of the input tensor must be passed at compile time using -DSRC_DIM1 (e.g. -DSRC_DIM1=112)
+ * @note The second dimension of the input tensor must be passed at compile time using -DSRC_DIM2 (e.g. -DSRC_DIM2=80)
+ * @note The kernel width must be passed at compile time using -DKERNEL_WIDTH (e.g. -DKERNEL_WIDTH=5)
+ * @note The kernel height must be passed at compile time using -DKERNEL_HEIGHT (e.g. -DKERNEL_HEIGHT=5)
+ * @note The convolution pad top must be passed at compile time using -DCONV_PAD_TOP (e.g. -DCONV_PAD_TOP=1)
+ * @note The convolution pad top must be passed at compile time using -DCONV_PAD_LEFT (e.g. -DCONV_PAD_LEFT=1)
+ * @note The convolution stride along the width must be passed at compile time using -DCONV_STRIDE_X (e.g. -DCONV_STRIDE_Y=X)
+ * @note The convolution stride along the height must be passed at compile time using -DCONV_STRIDE_Y (e.g. -DCONV_STRIDE_Y=1)
+ * @note It is possible to select the activation function to apply using -DACTIVATION_TYPE e.g. -DACTIVATION_TYPE=relu
+ * @note A, B variables required by some activation functions are set using -DA_VAL= and -DB_VAL= respectively
+ *
+ * @param[in] src_ptr                               Pointer to the source tensor. Supported data types: QASYMM8
+ * @param[in] src_stride_x                          Stride of the source tensor in X dimension (in bytes)
+ * @param[in] src_step_x                            src_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in] src_stride_y                          Stride of the source tensor in Y dimension (in bytes)
+ * @param[in] src_step_y                            src_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in] src_stride_z                          Stride of the source tensor in Z dimension (in bytes)
+ * @param[in] src_step_z                            src_stride_y * number of elements along Z processed per workitem(in bytes)
+ * @param[in] src_stride_w                          Stride of the source tensor in W dimension (in bytes)
+ * @param[in] src_step_w                            src_stride_w * number of elements along W processed per workitem(in bytes)
+ * @param[in] src_offset_first_element_in_bytes     The offset of the first element in the source tensor
+ * @param[in] dst_ptr                               Pointer to the destination tensor. Supported data types: same as src_ptr
+ * @param[in] dst_stride_x                          Stride of the destination tensor in X dimension (in bytes)
+ * @param[in] dst_step_x                            dst_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in] dst_stride_y                          Stride of the destination tensor in Y dimension (in bytes)
+ * @param[in] dst_step_y                            dst_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in] dst_stride_z                          Stride of the destination tensor in Z dimension (in bytes)
+ * @param[in] dst_step_z                            dst_stride_z * number of elements along Y processed per workitem(in bytes)
+ * @param[in] dst_stride_w                          Stride of the destination tensor in W dimension (in bytes)
+ * @param[in] dst_step_w                            dst_stride_w * number of elements along W processed per workitem(in bytes)
+ * @param[in] dst_offset_first_element_in_bytes     The offset of the first element in the destination tensor
+ * @param[in] weights_ptr                           Pointer to the weights tensor. Supported data types: QASYMM8
+ * @param[in] weights_stride_x                      Stride of the weights tensor in X dimension (in bytes)
+ * @param[in] weights_step_x                        weights_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in] weights_stride_y                      Stride of the weights tensor in Y dimension (in bytes)
+ * @param[in] weights_step_y                        weights_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in] weights_stride_z                      Stride of the weights tensor in Z dimension (in bytes)
+ * @param[in] weights_step_z                        weights_stride_z * number of elements along Y processed per workitem(in bytes)
+ * @param[in] weights_offset_first_element_in_bytes The offset of the first element in the weights tensor
+ * @param[in] biases_ptr                            (Optional) Pointer to the biases vector. Supported data types: same as src_ptr
+ * @param[in] biases_stride_x                       (Optional) Stride of the biases vector in X dimension (in bytes)
+ * @param[in] biases_step_x                         (Optional) biases_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in] biases_offset_first_element_in_bytes  (Optional) The offset of the first element in the biases vector
+ */
+__kernel void dwc_MxN_native_quantized8_nhwc(
+    TENSOR4D_DECLARATION(src),
+    TENSOR4D_DECLARATION(dst),
+    TENSOR3D_DECLARATION(weights),
+#if defined(HAS_BIAS)
+    VECTOR_DECLARATION(biases)
+#endif // defined(HAS_BIAS)
+)
+{
+    int x = get_global_id(0); // channels
+    int y = get_global_id(1); // spatial coordinate x
+#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)
+
+    __global uchar *s_addr = src_ptr + src_offset_first_element_in_bytes + x * sizeof(uchar) * (int)N0;
+
+    __global uchar *d_addr = dst_ptr + dst_offset_first_element_in_bytes + x * sizeof(uchar) * (int)DEPTH_MULTIPLIER * (int)N0 + y * dst_stride_y + z * dst_stride_z;
+
+    __global uchar *w_addr = weights_ptr + weights_offset_first_element_in_bytes + x * sizeof(uchar) * (int)DEPTH_MULTIPLIER * (int)N0;
+
+#if defined(HAS_BIAS)
+    __global uchar *b_addr = biases_ptr + biases_offset_first_element_in_bytes + x * sizeof(int) * (int)DEPTH_MULTIPLIER * (int)N0;
+#endif // defined(HAS_BIAS)
+
+#if defined(DST_DEPTH)
+    s_addr += b * src_stride_w;
+    d_addr += b * dst_stride_w;
+#endif // defined(DST_DEPTH)
+
+#if DEPTH_MULTIPLIER > 1
+    for(int d = 0; d < (int)DEPTH_MULTIPLIER; ++d)
+    {
+#endif // DEPTH_MULTIPLIER > 1
+        // Each work-item computes N0x1x1 elements
+        VEC_SHORT res = 0;
+
+        int x_coord = y * CONV_STRIDE_X - (int)CONV_PAD_LEFT;
+        int y_coord = z * CONV_STRIDE_Y - (int)CONV_PAD_TOP;
+
+        for(int yk = 0; yk < KERNEL_HEIGHT; ++yk)
+        {
+            if(y_coord >= 0 && y_coord < SRC_DIM2)
+            {
+                int x_coord_tmp = x_coord;
+
+                for(int xk = 0; xk < KERNEL_WIDTH; ++xk)
+                {
+                    if(x_coord_tmp >= 0 && x_coord_tmp < SRC_DIM1)
+                    {
+                        int s_offset = x_coord_tmp * (int)src_stride_y + y_coord * (int)src_stride_z;
+                        int w_offset = xk * weights_stride_y + yk * weights_stride_z;
+
+                        // Load input and weights values
+                        VEC_SHORT i = CONVERT(VLOAD(N0)(0, (__global uchar *)(s_addr + s_offset)), VEC_SHORT);
+                        VEC_SHORT w = CONVERT(VLOAD(N0)(0, (__global uchar *)(w_addr + w_offset)), VEC_SHORT);
+
+                        res += (i + (VEC_SHORT)INPUT_OFFSET) * (w + (VEC_SHORT)WEIGHTS_OFFSET);
+                    }
+                    x_coord_tmp += DILATION_X;
+                }
+            }
+            y_coord += DILATION_Y;
+        }
+
+#if defined(HAS_BIAS)
+        VEC_SHORT bias = CONVERT(VLOAD(N0)(0, (__global int *)(b_addr)), VEC_SHORT);
+        res += bias;
+#endif // defined(HAS_BIAS)
+
+        res = CONVERT(ASYMM_MULT_BY_QUANT_MULTIPLIER_LESS_THAN_ONE(CONVERT(res, VEC_INT), OUTPUT_MULTIPLIER, OUTPUT_SHIFT, N0), VEC_SHORT);
+        res += (VEC_SHORT)OUTPUT_OFFSET;
+
+        VEC_UCHAR res1 = CONVERT_SAT(res, VEC_UCHAR);
+
+        VSTORE(N0)
+        (ACTIVATION_FUNC(res1), 0, (__global uchar *)(d_addr));
+
+#if DEPTH_MULTIPLIER > 1
+        w_addr += sizeof(uchar);
+        d_addr += sizeof(uchar);
+#if defined(HAS_BIAS)
+        b_addr += sizeof(int);
+#endif // defined(HAS_BIAS)
+    }
+#endif // DEPTH_MULTIPLIER > 1
+}
+#endif // defined(SRC_DIM1) && defined(SRC_DIM2) && defined(KERNEL_WIDTH) && defined(KERNEL_HEIGHT) && defiend(N0) && defined(DILATION_X) && defined(DILATION_Y) && defined(CONV_STRIDE_X) && defined(CONV_STRIDE_Y) && defined(CONV_PAD_LEFT) && defined(CONV_PAD_TOP) && defined(INPUT_OFFSET) && defined(WEIGHTS_OFFSET) && defined(OUTPUT_OFFSET) && defined(OUTPUT_MULTIPLIER) && defined(OUTPUT_SHIFT)
diff --git a/src/core/CL/cl_kernels/direct_convolution_1x1_3x3_5x5_quantized.cl b/src/core/CL/cl_kernels/direct_convolution_1x1_3x3_5x5_quantized.cl
index 83da76785b..5ad9afb23c 100644
--- a/src/core/CL/cl_kernels/direct_convolution_1x1_3x3_5x5_quantized.cl
+++ b/src/core/CL/cl_kernels/direct_convolution_1x1_3x3_5x5_quantized.cl
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017-2018 ARM Limited.
+ * Copyright (c) 2017-2019 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -247,79 +247,3 @@ __kernel void direct_convolution_1x1_3x3_5x5_quantized(
     vstore8(convert_uchar8_sat(pixels0), 0, (__global uchar *)dst.ptr);
 }
 #endif // defined(DATA_TYPE) && defined(STRIDE_X) && defined(WEIGHTS_DEPTH)
-
-#if defined(VEC_SIZE)
-
-#define VEC_INT VEC_DATA_TYPE(int, VEC_SIZE)
-#define CONVERT_SAT_UCHAR_STR(x, size) (convert_uchar##size##_sat((x)))
-#define CONVERT_SAT_UCHAR(x, size) CONVERT_SAT_UCHAR_STR(x, size)
-
-/** This function computes the output stage of a depthwise convolution.
- *
- * @param[in] src_ptr                            Pointer to the source image. Supported data types: QASYMM8
- * @param[in] src_stride_x                       Stride of the source image in X dimension (in bytes)
- * @param[in] src_step_x                         src_stride_x * number of elements along X processed per workitem(in bytes)
- * @param[in] src_stride_y                       Stride of the source image in Y dimension (in bytes)
- * @param[in] src_step_y                         src_stride_y * number of elements along Y processed per workitem(in bytes)
- * @param[in] src_offset_first_element_in_bytes  The offset of the first element in the source image
- * @param[in] src_stride_z                       Stride of the source tensor in Z dimension (in bytes)
- * @param[in] src_step_z                         src_stride_z * number of elements along Y processed per workitem(in bytes)
- * @param[in] dst_ptr                            Pointer to the destination tensor. Supported data types: QASYMM8
- * @param[in] dst_stride_x                       Stride of the destination tensor in X dimension (in bytes)
- * @param[in] dst_step_x                         dst_stride_x * number of elements along X processed per workitem(in bytes)
- * @param[in] dst_stride_y                       Stride of the destination tensor in Y dimension (in bytes)
- * @param[in] dst_step_y                         dst_stride_y * number of elements along Y processed per workitem(in bytes)
- * @param[in] dst_stride_z                       Stride of the destination tensor in Z dimension (in bytes)
- * @param[in] dst_step_z                         dst_stride_z * number of elements along Y processed per workitem(in bytes)
- * @param[in] dst_offset_first_element_in_bytes  The offset of the first element in the destination tensor
- * @param[in] bias_ptr                           (Optional) Pointer to the biases vector. Supported data types: S32
- * @param[in] bias_stride_x                      (Optional) Stride of the biases vector in X dimension (in bytes)
- * @param[in] bias_step_x                        (Optional) bias_stride_x * number of elements along X processed per workitem(in bytes)
- * @param[in] bias_offset_first_element_in_bytes (Optional) The offset of the first element in the biases vector
- * @param[in] output_offset                      Quantized offset of zero point of the output tensor data range
- * @param[in] output_multiplier                  Output scale multiplier
- * @param[in] output_shift                       Output scale divisor exponent
- */
-__kernel void output_stage_quantized(
-    TENSOR3D_DECLARATION(src),
-    TENSOR3D_DECLARATION(dst),
-#if defined(HAS_BIAS)
-    VECTOR_DECLARATION(bias),
-#endif //defined(HAS_BIAS)
-    int output_offset,
-    int output_multiplier,
-    int output_shift)
-{
-    Image src = CONVERT_TENSOR3D_TO_IMAGE_STRUCT(src);
-    Image dst = CONVERT_TENSOR3D_TO_IMAGE_STRUCT(dst);
-#if defined(HAS_BIAS)
-    Vector bias = CONVERT_TO_VECTOR_STRUCT_NO_STEP(bias);
-#endif //defined(HAS_BIAS)
-
-    // Load input
-    VEC_INT vals = VLOAD(VEC_SIZE)(0, (__global int *)(src.ptr));
-
-#if defined(HAS_BIAS)
-    // Load and add bias
-#if defined(NCHW)
-    int bias_value = *((__global int *)(vector_offset(&bias, get_global_id(2))));
-#else  // defined(NCHW)
-    VEC_INT bias_value = VLOAD(VEC_SIZE)(0, ((__global int *)(vector_offset(&bias, get_global_id(0) * VEC_SIZE))));
-#endif // defined(NCHW)
-
-    vals += (VEC_INT)(bias_value);
-#endif //defined(HAS_BIAS)
-
-    vals = ASYMM_MULT_BY_QUANT_MULTIPLIER_LESS_THAN_ONE(vals, output_multiplier, output_shift, VEC_SIZE);
-    vals = vals + output_offset;
-
-    // Store result in dst
-    VSTORE(VEC_SIZE)
-    (CONVERT_SAT_UCHAR(vals, VEC_SIZE), 0, (__global uchar *)dst.ptr);
-}
-
-#undef VEC_INT
-#undef CONVERT_SAT_UCHAR_STR
-#undef CONVERT_SAT_UCHAR
-
-#endif // defined(VEC_SIZE)
diff --git a/src/core/CL/cl_kernels/helpers.h b/src/core/CL/cl_kernels/helpers.h
index f7f208529a..8117c1e519 100644
--- a/src/core/CL/cl_kernels/helpers.h
+++ b/src/core/CL/cl_kernels/helpers.h
@@ -150,6 +150,28 @@
 #define convert_half8_sat convert_half8
 #define convert_half16_sat convert_half16
 
+#define convert_float1 convert_float
+#define convert_half1 convert_half
+#define convert_char1 convert_char
+#define convert_uchar1 convert_uchar
+#define convert_short1 convert_short
+#define convert_ushort1 convert_ushort
+#define convert_int1 convert_int
+#define convert_uint1 convert_uint
+#define convert_long1 convert_long
+#define convert_ulong1 convert_ulong
+#define convert_double1 convert_double
+
+#define convert_char1_sat convert_char_sat
+#define convert_uchar1_sat convert_uchar_sat
+#define convert_short1_sat convert_short_sat
+#define convert_ushort1_sat convert_ushort_sat
+#define convert_int1_sat convert_int_sat
+#define convert_uint1_sat convert_uint_sat
+#define convert_long1_sat convert_long_sat
+#define convert_ulong1_sat convert_ulong_sat
+#define convert_double1_sat convert_double_sat
+
 #define VEC_DATA_TYPE_STR(type, size) type##size
 #define VEC_DATA_TYPE(type, size) VEC_DATA_TYPE_STR(type, size)
 
diff --git a/src/core/CL/cl_kernels/helpers_asymm.h b/src/core/CL/cl_kernels/helpers_asymm.h
index 53e6719cd7..57ecccc2b2 100644
--- a/src/core/CL/cl_kernels/helpers_asymm.h
+++ b/src/core/CL/cl_kernels/helpers_asymm.h
@@ -381,11 +381,13 @@ DEQUANTIZE_IMPL(uchar, 4)
 DEQUANTIZE_IMPL(ushort, 4)
 DEQUANTIZE_IMPL(short, 4)
 
+ASYMM_ROUNDING_DIVIDE_BY_POW2_IMPL(1)
 ASYMM_ROUNDING_DIVIDE_BY_POW2_IMPL(2)
 ASYMM_ROUNDING_DIVIDE_BY_POW2_IMPL(4)
 ASYMM_ROUNDING_DIVIDE_BY_POW2_IMPL(8)
 ASYMM_ROUNDING_DIVIDE_BY_POW2_IMPL(16)
 
+ASYMM_MULT_IMPL(1)
 ASYMM_MULT_IMPL(2)
 ASYMM_MULT_IMPL(4)
 ASYMM_MULT_IMPL(8)
diff --git a/src/core/CL/kernels/CLDepthwiseConvolutionLayerNativeKernel.cpp b/src/core/CL/kernels/CLDepthwiseConvolutionLayerNativeKernel.cpp
index 8b624bb2cb..2115fc614d 100644
--- a/src/core/CL/kernels/CLDepthwiseConvolutionLayerNativeKernel.cpp
+++ b/src/core/CL/kernels/CLDepthwiseConvolutionLayerNativeKernel.cpp
@@ -35,6 +35,7 @@
 #include "arm_compute/core/Types.h"
 #include "arm_compute/core/Utils.h"
 #include "arm_compute/core/utils/misc/ShapeCalculator.h"
+#include "arm_compute/core/utils/quantization/AsymmHelpers.h"
 
 namespace arm_compute
 {
@@ -46,7 +47,7 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *weights,
     ARM_COMPUTE_UNUSED(dwc_info);
     ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input);
     ARM_COMPUTE_RETURN_ERROR_ON_DATA_LAYOUT_NOT_IN(input, DataLayout::NHWC);
-    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F16, DataType::F32);
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::F16, DataType::F32);
     ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights);
     ARM_COMPUTE_RETURN_ERROR_ON(depth_multiplier > 1 && dwc_weights_info.n0 != 1);
     ARM_COMPUTE_RETURN_ERROR_ON(conv_info.stride().first < 1);
@@ -59,8 +60,16 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *weights,
     if(biases != nullptr)
     {
         ARM_COMPUTE_RETURN_ERROR_ON(biases->dimension(0) != weights->dimension(0));
-        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(weights, biases);
         ARM_COMPUTE_RETURN_ERROR_ON(biases->num_dimensions() > 1);
+
+        if(is_data_type_quantized(input->data_type()))
+        {
+            ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(biases, 1, DataType::S32);
+        }
+        else
+        {
+            ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(weights, biases);
+        }
     }
 
     if(output->total_size() != 0)
@@ -137,6 +146,7 @@ void CLDepthwiseConvolutionLayerNativeKernel::configure(const ICLTensor *input,
     const size_t idx_h          = get_data_layout_dimension_index(input->info()->data_layout(), DataLayoutDimension::HEIGHT);
     const size_t weights_width  = weights->info()->dimension(idx_w);
     const size_t weights_height = weights->info()->dimension(idx_h);
+    const bool   is_quantized   = is_data_type_quantized(input->info()->data_type());
 
     CLBuildOptions build_opts;
     build_opts.add_option_if(_biases != nullptr, "-DHAS_BIAS");
@@ -155,10 +165,46 @@ void CLDepthwiseConvolutionLayerNativeKernel::configure(const ICLTensor *input,
     build_opts.add_option("-DCONV_STRIDE_Y=" + support::cpp11::to_string(conv_info.stride().second));
     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(dwc_info.activation_info.enabled(), "-DA_VAL=" + float_to_string_with_full_precision(dwc_info.activation_info.a()));
-    build_opts.add_option_if(dwc_info.activation_info.enabled(), "-DB_VAL=" + float_to_string_with_full_precision(dwc_info.activation_info.b()));
 
-    std::string kernel_name("dwc_MxN_native_fp_nhwc");
+    std::string kernel_name = (is_quantized) ? "dwc_MxN_native_quantized8_nhwc" : "dwc_MxN_native_fp_nhwc";
+
+    if(is_quantized)
+    {
+        const UniformQuantizationInfo iq_info = _input->info()->quantization_info().uniform();
+        const UniformQuantizationInfo wq_info = _weights->info()->quantization_info().uniform();
+        const UniformQuantizationInfo oq_info = _output->info()->quantization_info().uniform();
+
+        float multiplier        = iq_info.scale * wq_info.scale / oq_info.scale;
+        int   output_multiplier = 0;
+        int   output_shift      = 0;
+        quantization::calculate_quantized_multiplier_less_than_one(multiplier, &output_multiplier, &output_shift);
+
+        build_opts.add_option("-DINPUT_OFFSET=" + support::cpp11::to_string(-iq_info.offset));
+        build_opts.add_option("-DWEIGHTS_OFFSET=" + support::cpp11::to_string(-wq_info.offset));
+        build_opts.add_option("-DOUTPUT_OFFSET=" + support::cpp11::to_string(oq_info.offset));
+        build_opts.add_option("-DOUTPUT_MULTIPLIER=" + support::cpp11::to_string(output_multiplier));
+        build_opts.add_option("-DOUTPUT_SHIFT=" + support::cpp11::to_string(output_shift));
+
+        if(dwc_info.activation_info.enabled())
+        {
+            const int a_val = quantize_qasymm8(dwc_info.activation_info.a(), oq_info);
+            const int b_val = quantize_qasymm8(dwc_info.activation_info.b(), oq_info);
+            const int o1    = oq_info.offset;
+
+            build_opts.add_option("-DA_VAL=" + support::cpp11::to_string(a_val));
+            build_opts.add_option("-DB_VAL=" + support::cpp11::to_string(b_val));
+            build_opts.add_option("-DCONST_0=" + support::cpp11::to_string(o1));
+
+            const float s1 = iq_info.scale;
+            build_opts.add_option("-DS1_VAL=" + float_to_string_with_full_precision(s1));
+            build_opts.add_option("-DO1_VAL=" + support::cpp11::to_string(o1));
+        }
+    }
+    else
+    {
+        build_opts.add_option_if(dwc_info.activation_info.enabled(), "-DA_VAL=" + float_to_string_with_full_precision(dwc_info.activation_info.a()));
+        build_opts.add_option_if(dwc_info.activation_info.enabled(), "-DB_VAL=" + float_to_string_with_full_precision(dwc_info.activation_info.b()));
+    }
 
     ICLKernel::configure_internal(win_config.second);
     _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel(kernel_name, build_opts.options()));
diff --git a/src/core/CL/kernels/CLDepthwiseConvolutionLayerReshapeWeightsGenericKernel.cpp b/src/core/CL/kernels/CLDepthwiseConvolutionLayerReshapeWeightsGenericKernel.cpp
deleted file mode 100644
index 582c600c61..0000000000
--- a/src/core/CL/kernels/CLDepthwiseConvolutionLayerReshapeWeightsGenericKernel.cpp
+++ /dev/null
@@ -1,147 +0,0 @@
-/*
- * Copyright (c) 2017-2019 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.
- */
-#include "arm_compute/core/CL/kernels/CLDepthwiseConvolutionLayerReshapeWeightsGenericKernel.h"
-
-#include "arm_compute/core/CL/CLHelpers.h"
-#include "arm_compute/core/CL/CLKernelLibrary.h"
-#include "arm_compute/core/CL/CLValidate.h"
-#include "arm_compute/core/CL/ICLTensor.h"
-#include "arm_compute/core/CL/OpenCL.h"
-#include "arm_compute/core/Error.h"
-#include "arm_compute/core/Helpers.h"
-#include "arm_compute/core/Types.h"
-#include "support/ToolchainSupport.h"
-
-using namespace arm_compute;
-
-namespace
-{
-Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, const ITensorInfo *biases)
-{
-    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);
-    const size_t idx_c = get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::CHANNEL);
-
-    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);
-    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
-    ARM_COMPUTE_RETURN_ERROR_ON(is_data_type_quantized_asymmetric(input->data_type()) && (biases != nullptr));
-    ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(idx_c) != output->dimension(1));
-    ARM_COMPUTE_RETURN_ERROR_ON(output->dimension(0) != (input->dimension(idx_w) * input->dimension(idx_h) + ((biases != nullptr) ? 1 : 0)));
-    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_QUANTIZATION_INFO(input, output);
-
-    if(biases != nullptr)
-    {
-        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, biases);
-        ARM_COMPUTE_RETURN_ERROR_ON(biases->dimension(0) != input->dimension(idx_c));
-        ARM_COMPUTE_RETURN_ERROR_ON(biases->num_dimensions() > 1);
-    }
-
-    return Status{};
-}
-} // namespace
-
-CLDepthwiseConvolutionLayerReshapeWeightsGenericKernel::CLDepthwiseConvolutionLayerReshapeWeightsGenericKernel()
-    : _input(nullptr), _biases(nullptr), _output(nullptr)
-{
-}
-
-void CLDepthwiseConvolutionLayerReshapeWeightsGenericKernel::configure(const ICLTensor *input, ICLTensor *output, const ICLTensor *biases)
-{
-    ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
-    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info(), (biases != nullptr) ? biases->info() : nullptr));
-
-    _input  = input;
-    _biases = biases;
-    _output = output;
-
-    const size_t idx_w = get_data_layout_dimension_index(input->info()->data_layout(), DataLayoutDimension::WIDTH);
-
-    // Create kernel
-    std::set<std::string> build_opts;
-
-    build_opts.emplace("-DDATA_TYPE=" + get_cl_type_from_data_type(input->info()->data_type()));
-    build_opts.emplace("-DSRC_WIDTH=" + support::cpp11::to_string(input->info()->dimension(idx_w)));
-    build_opts.emplace("-D" + string_from_data_layout(input->info()->data_layout()));
-    if(_biases != nullptr)
-    {
-        build_opts.emplace("-DHAS_BIAS");
-    }
-
-    _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("depthwise_convolution_reshape_weights_generic", build_opts));
-
-    // Configure  kernel window
-    Window win = calculate_max_window(*input->info(), Steps());
-    // The CLDepthwiseConvolutionLayerReshapeWeightsGenericKernel doesn't need padding so update_window_and_padding() can be skipped
-    output->info()->set_valid_region(ValidRegion(Coordinates(), output->info()->tensor_shape()));
-
-    ICLKernel::configure_internal(win);
-}
-
-Status CLDepthwiseConvolutionLayerReshapeWeightsGenericKernel::validate(const ITensorInfo *input, const ITensorInfo *output, const ITensorInfo *biases)
-{
-    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output, biases));
-    return Status{};
-}
-
-void CLDepthwiseConvolutionLayerReshapeWeightsGenericKernel::run(const Window &window, cl::CommandQueue &queue)
-{
-    ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
-    ARM_COMPUTE_ERROR_ON_MISMATCHING_WINDOWS(ICLKernel::window(), window);
-
-    Window slice     = window.first_slice_window_3D();
-    Window slice_out = window.first_slice_window_2D();
-
-    const size_t idx_w = get_data_layout_dimension_index(_input->info()->data_layout(), DataLayoutDimension::WIDTH);
-    const size_t idx_h = get_data_layout_dimension_index(_input->info()->data_layout(), DataLayoutDimension::HEIGHT);
-    const size_t idx_c = get_data_layout_dimension_index(_input->info()->data_layout(), DataLayoutDimension::CHANNEL);
-
-    // Setup slice
-    slice.set(Window::DimX, Window::Dimension(0, _input->info()->dimension(idx_w), _input->info()->dimension(idx_w)));
-    slice.set(Window::DimY, Window::Dimension(0, _input->info()->dimension(idx_h), 1));
-    slice.set(Window::DimZ, Window::Dimension(0, _input->info()->dimension(idx_c), 1));
-
-    // Setup output slice
-    // The first two dimensions of the output are increased by the inner loops
-    slice_out.set(Window::DimX, Window::Dimension(0, 0, 0));
-    slice_out.set(Window::DimY, Window::Dimension(0, 0, 0));
-
-    // Set biases
-    if(_biases != nullptr)
-    {
-        unsigned int idx = num_arguments_per_3D_tensor() + num_arguments_per_2D_tensor();
-        Window       slice_biases;
-        slice_biases.use_tensor_dimensions(_biases->info()->tensor_shape());
-        add_1D_tensor_argument(idx, _biases, slice_biases);
-    }
-
-    do
-    {
-        unsigned int idx = 0;
-        add_3D_tensor_argument(idx, _input, slice);
-        add_2D_tensor_argument(idx, _output, slice_out);
-        enqueue(queue, *this, slice, lws_hint());
-    }
-    while(window.slide_window_slice_3D(slice) && window.slide_window_slice_2D(slice_out));
-}
diff --git a/src/core/CL/kernels/CLDepthwiseIm2ColKernel.cpp b/src/core/CL/kernels/CLDepthwiseIm2ColKernel.cpp
deleted file mode 100644
index 0312a57664..0000000000
--- a/src/core/CL/kernels/CLDepthwiseIm2ColKernel.cpp
+++ /dev/null
@@ -1,148 +0,0 @@
-/*
- * Copyright (c) 2017-2019 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.
- */
-#include "arm_compute/core/CL/kernels/CLDepthwiseIm2ColKernel.h"
-
-#include "arm_compute/core/CL/CLHelpers.h"
-#include "arm_compute/core/CL/CLKernelLibrary.h"
-#include "arm_compute/core/CL/CLValidate.h"
-#include "arm_compute/core/CL/ICLTensor.h"
-#include "arm_compute/core/CL/OpenCL.h"
-#include "arm_compute/core/Error.h"
-#include "arm_compute/core/Helpers.h"
-#include "arm_compute/core/Types.h"
-#include "support/ToolchainSupport.h"
-
-#include <tuple>
-
-using namespace arm_compute;
-
-CLDepthwiseIm2ColKernel::CLDepthwiseIm2ColKernel()
-    : _input(nullptr), _output(nullptr)
-{
-}
-
-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,
-                          const Size2D &dilation)
-{
-    const size_t idx_c = get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::CHANNEL);
-
-    ARM_COMPUTE_UNUSED(conv_info);
-    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);
-    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
-    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,
-                                        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, dilation));
-
-    _input  = input;
-    _output = output;
-
-    const DataLayout              data_layout = input->info()->data_layout();
-    const size_t                  idx_w       = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH);
-    const size_t                  idx_h       = get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT);
-    const UniformQuantizationInfo qinfo       = input->info()->quantization_info().uniform();
-
-    // Create kernel
-    CLBuildOptions build_opts;
-
-    build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(input->info()->data_type()));
-    build_opts.add_option("-DSTRIDE_X=" + support::cpp11::to_string(conv_info.stride().first));
-    build_opts.add_option("-DSTRIDE_Y=" + support::cpp11::to_string(conv_info.stride().second));
-    build_opts.add_option("-DPAD_LEFT=" + support::cpp11::to_string(conv_info.pad_left()));
-    build_opts.add_option("-DPAD_TOP=" + support::cpp11::to_string(conv_info.pad_top()));
-    build_opts.add_option("-DPAD_RIGHT=" + support::cpp11::to_string(conv_info.pad_right()));
-    build_opts.add_option("-DPAD_BOTTOM=" + support::cpp11::to_string(conv_info.pad_bottom()));
-    build_opts.add_option("-DSRC_WIDTH=" + support::cpp11::to_string(input->info()->dimension(idx_w)));
-    build_opts.add_option("-DSRC_HEIGHT=" + support::cpp11::to_string(input->info()->dimension(idx_h)));
-    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()),
-                                  "-DPAD_VALUE=" + support::cpp11::to_string(qinfo.offset),
-                                  "-DPAD_VALUE=0");
-
-    _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("depthwise_im2col", build_opts.options()));
-
-    // Configure  kernel window
-    Window win = calculate_max_window(*output->info(), Steps());
-    // CLDepthwiseIm2ColKernel doesn't need padding so update_window_and_padding() can be skipped
-    output->info()->set_valid_region(ValidRegion(Coordinates(), output->info()->tensor_shape()));
-
-    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,
-                                         const Size2D &dilation)
-{
-    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output, kernel_dims, conv_info, has_bias, depth_multiplier, dilation));
-
-    return Status{};
-}
-
-void CLDepthwiseIm2ColKernel::run(const Window &window, cl::CommandQueue &queue)
-{
-    ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
-    ARM_COMPUTE_ERROR_ON_MISMATCHING_WINDOWS(ICLKernel::window(), window);
-
-    Window slice    = window.first_slice_window_3D();
-    Window slice_in = window.first_slice_window_3D();
-
-    // Setup slice
-    slice.set(Window::DimX, Window::Dimension(0, _output->info()->dimension(0), _output->info()->dimension(0)));
-    slice.set(Window::DimY, Window::Dimension(0, _output->info()->dimension(1), 1));
-    slice.set(Window::DimZ, Window::Dimension(0, _output->info()->dimension(2), 1));
-
-    // Setup input slice
-    // The first three dimensions of the input are increased by the inner loops
-    slice_in.set(Window::DimX, Window::Dimension(0, 0, 0));
-    slice_in.set(Window::DimY, Window::Dimension(0, 0, 0));
-    slice_in.set(Window::DimZ, Window::Dimension(0, 0, 0));
-
-    do
-    {
-        unsigned int idx = 0;
-        add_3D_tensor_argument(idx, _input, slice_in);
-        add_3D_tensor_argument(idx, _output, slice);
-        enqueue(queue, *this, slice, lws_hint());
-    }
-    while(window.slide_window_slice_3D(slice) && window.slide_window_slice_3D(slice_in));
-}
diff --git a/src/core/CL/kernels/CLDepthwiseVectorToTensorKernel.cpp b/src/core/CL/kernels/CLDepthwiseVectorToTensorKernel.cpp
deleted file mode 100644
index 0f029fda74..0000000000
--- a/src/core/CL/kernels/CLDepthwiseVectorToTensorKernel.cpp
+++ /dev/null
@@ -1,125 +0,0 @@
-/*
- * Copyright (c) 2017-2019 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.
- */
-#include "arm_compute/core/CL/kernels/CLDepthwiseVectorToTensorKernel.h"
-
-#include "arm_compute/core/CL/CLHelpers.h"
-#include "arm_compute/core/CL/CLKernelLibrary.h"
-#include "arm_compute/core/CL/CLValidate.h"
-#include "arm_compute/core/CL/ICLTensor.h"
-#include "arm_compute/core/CL/OpenCL.h"
-#include "arm_compute/core/Error.h"
-#include "arm_compute/core/Helpers.h"
-#include "arm_compute/core/Types.h"
-#include "arm_compute/core/utils/misc/ShapeCalculator.h"
-#include "support/ToolchainSupport.h"
-
-using namespace arm_compute;
-using namespace arm_compute::misc::shape_calculator;
-
-namespace
-{
-Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, size_t conv_w, size_t conv_h)
-{
-    ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input);
-    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::S32, DataType::F16, DataType::F32);
-
-    if(output->total_size() != 0)
-    {
-        TensorShape output_shape = compute_vector_to_tensor_output_shape(input->tensor_shape(), conv_w, conv_h, output->data_layout());
-        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(output->tensor_shape(), output_shape);
-        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
-        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_QUANTIZATION_INFO(input, output);
-    }
-
-    return Status{};
-}
-} // namespace
-
-CLDepthwiseVectorToTensorKernel::CLDepthwiseVectorToTensorKernel()
-    : _input(nullptr), _output(nullptr)
-{
-}
-
-void CLDepthwiseVectorToTensorKernel::configure(const ICLTensor *input, ICLTensor *output, size_t conv_w, size_t conv_h)
-{
-    ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
-
-    // Output auto inizialitation if not yet initialized
-    TensorShape output_shape = compute_vector_to_tensor_output_shape(input->info()->tensor_shape(), conv_w, conv_h, output->info()->data_layout());
-    auto_init_if_empty(*output->info(), input->info()->clone()->set_tensor_shape(output_shape));
-
-    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info(), conv_w, conv_h));
-
-    _input  = input;
-    _output = output;
-
-    // Create kernel
-    CLBuildOptions build_opts;
-    build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(input->info()->data_type()));
-    build_opts.add_option("-DCONV_WIDTH=" + support::cpp11::to_string(conv_w));
-    build_opts.add_option("-DCONV_HEIGHT=" + support::cpp11::to_string(conv_h));
-    build_opts.add_option("-D" + string_from_data_layout(output->info()->data_layout()));
-
-    _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("depthwise_vector_to_tensor", build_opts.options()));
-
-    // Configure  kernel window
-    Window win = calculate_max_window(*input->info(), Steps());
-    // The CLDepthwisevectorToTensorKernel doesn't need padding so update_window_and_padding() can be skipped
-    output->info()->set_valid_region(ValidRegion(Coordinates(), output->info()->tensor_shape()));
-
-    ICLKernel::configure_internal(win);
-}
-
-Status CLDepthwiseVectorToTensorKernel::validate(const ITensorInfo *input, const ITensorInfo *output, size_t conv_w, size_t conv_h)
-{
-    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output, conv_w, conv_h));
-    return Status{};
-}
-
-void CLDepthwiseVectorToTensorKernel::run(const Window &window, cl::CommandQueue &queue)
-{
-    ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
-    ARM_COMPUTE_ERROR_ON_MISMATCHING_WINDOWS(ICLKernel::window(), window);
-
-    Window slice     = window.first_slice_window_1D();
-    Window slice_out = window.first_slice_window_3D();
-
-    // Setup slice
-    slice.set(Window::DimX, Window::Dimension(0, _input->info()->dimension(0), 1));
-
-    // Setup output slice
-    // The first three dimensions of the output are increased by the inner loops
-    slice_out.set(Window::DimX, Window::Dimension(0, 0, 0));
-    slice_out.set(Window::DimY, Window::Dimension(0, 0, 0));
-    slice_out.set(Window::DimZ, Window::Dimension(0, 0, 0));
-
-    do
-    {
-        unsigned int idx = 0;
-        add_1D_tensor_argument(idx, _input, slice);
-        add_3D_tensor_argument(idx, _output, slice_out);
-        enqueue(queue, *this, slice, lws_hint());
-    }
-    while(window.slide_window_slice_1D(slice) && window.slide_window_slice_3D(slice_out));
-}
diff --git a/src/core/CL/kernels/CLDirectConvolutionOutputStageKernel.cpp b/src/core/CL/kernels/CLDirectConvolutionOutputStageKernel.cpp
deleted file mode 100644
index 515321cdfc..0000000000
--- a/src/core/CL/kernels/CLDirectConvolutionOutputStageKernel.cpp
+++ /dev/null
@@ -1,209 +0,0 @@
-/*
- * Copyright (c) 2018-2019 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.
- */
-#include "arm_compute/core/CL/kernels/CLDirectConvolutionLayerOutputStageKernel.h"
-
-#include "arm_compute/core/AccessWindowStatic.h"
-#include "arm_compute/core/CL/CLValidate.h"
-#include "arm_compute/core/CL/ICLTensor.h"
-#include "arm_compute/core/Error.h"
-#include "arm_compute/core/Helpers.h"
-#include "arm_compute/core/Types.h"
-#include "arm_compute/core/Window.h"
-
-#include <cstddef>
-#include <cstdint>
-
-namespace arm_compute
-{
-namespace
-{
-Status validate_arguments(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output)
-{
-    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input);
-    ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input);
-    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::S32, DataType::F16,
-                                                         DataType::F32);
-
-    if(bias != nullptr)
-    {
-        ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(bias);
-        ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(bias, 1, DataType::S32, DataType::F16, DataType::F32);
-
-        if(is_data_type_quantized_asymmetric(input->data_type()))
-        {
-            ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(bias, 1, DataType::S32);
-        }
-        else
-        {
-            ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, bias);
-        }
-
-        ARM_COMPUTE_RETURN_ERROR_ON(bias->num_dimensions() > 1);
-    }
-    else
-    {
-        ARM_COMPUTE_RETURN_ERROR_ON_MSG(is_data_type_float(input->data_type()),
-                                        "Calling output stage kernel with floating point arguments");
-    }
-
-    // Checks performed on output
-    if(input->data_type() == DataType::S32)
-    {
-        // Quantized configuration checks
-        ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::QASYMM8);
-    }
-    else
-    {
-        // In case of out-of-place computation (supported for non-quantized configurations)
-        if((output != nullptr) && (output->total_size() != 0))
-        {
-            ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
-        }
-    }
-
-    return Status{};
-}
-
-std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *bias, ITensorInfo *output)
-{
-    bool               window_changed                    = false;
-    const unsigned int num_elems_processed_per_iteration = 16 / element_size_from_data_type(input->data_type());
-
-    // Configure kernel window
-    Window win = calculate_max_window(*input, Steps(num_elems_processed_per_iteration));
-
-    // Input window
-    AccessWindowHorizontal input_access(input, 0, num_elems_processed_per_iteration);
-    window_changed = window_changed || update_window_and_padding(win, input_access);
-
-    // Bias window
-    if(bias != nullptr)
-    {
-        AccessWindowStatic bias_access(bias, 0, 0, ceil_to_multiple(bias->dimension(0), num_elems_processed_per_iteration), bias->dimension(1));
-        window_changed = window_changed || update_window_and_padding(win, bias_access);
-    }
-
-    // Output window
-    if(output != nullptr && (output->total_size() != 0))
-    {
-        AccessWindowHorizontal output_access(output, 0, num_elems_processed_per_iteration);
-        window_changed = window_changed || update_window_and_padding(win, output_access);
-        output_access.set_valid_region(win, ValidRegion(Coordinates(), output->tensor_shape()));
-    }
-    else
-    {
-        input_access.set_valid_region(win, ValidRegion(Coordinates(), input->tensor_shape()));
-    }
-
-    Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{};
-    return std::make_pair(err, win);
-}
-} // namespace
-
-CLDirectConvolutionLayerOutputStageKernel::CLDirectConvolutionLayerOutputStageKernel()
-    : _input(nullptr), _bias(nullptr), _output(nullptr), _result_fixedpoint_multiplier(0), _result_shift(0), _result_offset_after_shift(0)
-{
-}
-
-void CLDirectConvolutionLayerOutputStageKernel::configure(ICLTensor *input, const ICLTensor *bias, ICLTensor *output,
-                                                          int result_fixedpoint_multiplier, int result_shift, int result_offset_after_shift)
-{
-    ARM_COMPUTE_ERROR_ON_NULLPTR(input);
-
-    // Auto-initialize output if required
-    if(output != nullptr)
-    {
-        // Work out expected output data type
-        const DataType output_dt = (input->info()->data_type() == DataType::S32) ? DataType::QASYMM8 : input->info()->data_type();
-        // Output tensor auto initialization if not yet initialized
-        auto_init_if_empty(*output->info(), input->info()->clone()->set_data_type(output_dt));
-    }
-
-    // Perform validation step
-    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), (bias == nullptr) ? nullptr : bias->info(), (output == nullptr) ? nullptr : output->info()));
-
-    _bias                         = bias;
-    _input                        = input;
-    _output                       = output;
-    _result_fixedpoint_multiplier = result_fixedpoint_multiplier;
-    _result_shift                 = result_shift;
-    _result_offset_after_shift    = result_offset_after_shift;
-
-    const unsigned int num_elems_accessed_per_iteration = 16 / element_size_from_data_type(input->info()->data_type());
-
-    // Create kernel
-    CLBuildOptions build_opts;
-    build_opts.add_option_if(bias != nullptr, "-DHAS_BIAS");
-    build_opts.add_option("-D" + string_from_data_layout(input->info()->data_layout()));
-    build_opts.add_option("-DVEC_SIZE=" + support::cpp11::to_string(num_elems_accessed_per_iteration));
-    _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("output_stage_quantized", build_opts.options()));
-
-    // Set static kernel arguments
-    int idx = 2 * num_arguments_per_3D_tensor() + ((bias != nullptr) ? num_arguments_per_1D_tensor() : 0);
-    _kernel.setArg<int>(idx++, _result_offset_after_shift);
-    _kernel.setArg<int>(idx++, _result_fixedpoint_multiplier);
-    _kernel.setArg<int>(idx++, _result_shift);
-
-    // Configure kernel window
-    auto win_config = validate_and_configure_window(input->info(), (bias == nullptr) ? nullptr : bias->info(), (output == nullptr) ? nullptr : output->info());
-    ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
-    ICLKernel::configure_internal(win_config.second);
-}
-
-Status CLDirectConvolutionLayerOutputStageKernel::validate(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output)
-{
-    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, bias, output));
-    ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), bias == nullptr ? nullptr : bias->clone().get(), output == nullptr ? nullptr : output->clone().get()).first);
-
-    return Status{};
-}
-
-void CLDirectConvolutionLayerOutputStageKernel::run(const Window &window, cl::CommandQueue &queue)
-{
-    ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
-    ARM_COMPUTE_ERROR_ON_MISMATCHING_WINDOWS(ICLKernel::window(), window);
-
-    Window slice = window.first_slice_window_3D();
-
-    // Set bias vector
-    if(_bias != nullptr)
-    {
-        unsigned int idx1 = 2 * num_arguments_per_3D_tensor();
-        Window       slice_biases;
-        slice_biases.use_tensor_dimensions(_bias->info()->tensor_shape());
-        add_1D_tensor_argument(idx1, _bias, slice_biases);
-    }
-
-    // Run kernel
-    do
-    {
-        // Set arguments
-        unsigned int idx = 0;
-        add_3D_tensor_argument(idx, _input, slice);
-        add_3D_tensor_argument(idx, _output, slice);
-        enqueue(queue, *this, slice, lws_hint());
-    }
-    while(window.slide_window_slice_3D(slice));
-}
-} // namespace arm_compute
diff --git a/src/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFloatKernel.cpp b/src/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFloatKernel.cpp
index 2967a73866..ae096f295c 100644
--- a/src/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFloatKernel.cpp
+++ b/src/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFloatKernel.cpp
@@ -35,8 +35,6 @@
 
 #include "support/ToolchainSupport.h"
 
-using namespace arm_compute;
-
 namespace arm_compute
 {
 namespace
@@ -101,8 +99,6 @@ std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITen
 } // namespace
 
 class Coordinates;
-} // namespace arm_compute
-
 CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFloatKernel::CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFloatKernel()
     : _input(nullptr), _bias(nullptr), _output(nullptr)
 {
@@ -177,3 +173,4 @@ void CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFloatKernel::run(const Window &win
     }
     while(collapsed.slide_window_slice_3D(slice));
 }
+} // namespace arm_compute
diff --git a/src/runtime/CL/functions/CLDepthwiseConvolutionLayer.cpp b/src/runtime/CL/functions/CLDepthwiseConvolutionLayer.cpp
index f01b58a8b3..d9c21150df 100644
--- a/src/runtime/CL/functions/CLDepthwiseConvolutionLayer.cpp
+++ b/src/runtime/CL/functions/CLDepthwiseConvolutionLayer.cpp
@@ -246,28 +246,44 @@ void CLDepthwiseConvolutionLayer3x3::prepare()
     }
 }
 
-CLDepthwiseConvolutionLayer::CLDepthwiseConvolutionLayer()
-    : _im2col_kernel(), _weights_reshape_kernel(), _v2mm_kernel(), _vector_to_tensor_kernel(), _output_stage_kernel(), _activationlayer_function(), _v2mm_input_fill_border(), _v2mm_weights_fill_border(),
-      _input_reshaped(), _weights_reshaped(), _v2mm_output(), _output_reshaped(), _is_prepared(false), _is_quantized(false), _is_activationlayer_enabled(false), _original_weights(nullptr),
-      _optimised_function(nullptr)
+CLDepthwiseConvolutionLayer::CLDepthwiseConvolutionLayer(std::shared_ptr<IMemoryManager> memory_manager)
+    : _memory_group(std::move(memory_manager)),
+      _optimised_function(nullptr),
+      _dwc_native_kernel(),
+      _permute_input_to_nhwc(),
+      _permute_weights_to_nhwc(),
+      _permute_output_to_nchw(),
+      _permuted_input(),
+      _permuted_weights(),
+      _permuted_output(),
+      _original_weights(),
+      _needs_permute(false),
+      _is_prepared(false)
 {
 }
 
 void CLDepthwiseConvolutionLayer::configure(ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const PadStrideInfo &conv_info,
                                             unsigned int depth_multiplier, const ActivationLayerInfo &act_info, const Size2D &dilation)
 {
-    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);
-    ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_LAYOUT(input, output);
+    ARM_COMPUTE_ERROR_ON_NULLPTR(input, weights, output);
+    ARM_COMPUTE_ERROR_THROW_ON(CLDepthwiseConvolutionLayer::validate(input->info(),
+                                                                     weights->info(),
+                                                                     biases != nullptr ? biases->info() : nullptr,
+                                                                     output->info(),
+                                                                     conv_info,
+                                                                     depth_multiplier,
+                                                                     act_info,
+                                                                     dilation));
 
     const size_t idx_w = get_data_layout_dimension_index(input->info()->data_layout(), DataLayoutDimension::WIDTH);
     const size_t idx_h = get_data_layout_dimension_index(input->info()->data_layout(), DataLayoutDimension::HEIGHT);
 
-    ARM_COMPUTE_ERROR_ON(weights->info()->dimension(idx_w) + (weights->info()->dimension(idx_w) - 1) * (dilation.x() - 1) > input->info()->dimension(idx_w) + conv_info.pad_left() + conv_info.pad_right());
-    ARM_COMPUTE_ERROR_ON(weights->info()->dimension(idx_h) + (weights->info()->dimension(idx_h) - 1) * (dilation.y() - 1) > input->info()->dimension(idx_h) + conv_info.pad_top() + conv_info.pad_bottom());
-
     const bool can_run_optimised_3x3_kernel = (weights->info()->dimension(idx_w) == 3) && (weights->info()->dimension(idx_h) == 3);
 
+    _needs_permute    = false;
+    _is_prepared      = false;
+    _original_weights = weights;
+
     if(bool(can_run_optimised_3x3_kernel))
     {
         auto f = arm_compute::support::cpp14::make_unique<CLDepthwiseConvolutionLayer3x3>();
@@ -276,103 +292,46 @@ void CLDepthwiseConvolutionLayer::configure(ICLTensor *input, const ICLTensor *w
     }
     else
     {
-        const size_t idx_c = get_data_layout_dimension_index(input->info()->data_layout(), DataLayoutDimension::CHANNEL);
-
-        const size_t weights_w = weights->info()->dimension(idx_w);
-        const size_t weights_h = weights->info()->dimension(idx_h);
-        const size_t weights_z = weights->info()->dimension(idx_c);
-
-        _is_prepared      = false;
-        _original_weights = weights;
-        _is_quantized     = is_data_type_quantized_asymmetric(input->info()->data_type());
-
-        bool            append_bias = (biases != nullptr) && !_is_quantized;
-        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, dilation);
-
-        // Output auto inizialitation if not yet initialized
-        auto_init_if_empty(*output->info(), input->info()->clone()->set_tensor_shape(output_shape));
-        ARM_COMPUTE_ERROR_ON_MISMATCHING_DIMENSIONS(output->info()->tensor_shape(), output_shape);
-
-        // Output width and height
-        const unsigned int conv_w = output_shape[idx_w];
-        const unsigned int conv_h = output_shape[idx_h];
-
-        // Set up intermediate tensors
-        const size_t patch_size = weights_w * weights_h + ((append_bias) ? 1 : 0);
-        const size_t conv_size  = conv_w * conv_h;
-
-        const UniformQuantizationInfo iq_info = input->info()->quantization_info().uniform();
-        const UniformQuantizationInfo wq_info = weights->info()->quantization_info().uniform();
-        const UniformQuantizationInfo oq_info = output->info()->quantization_info().uniform();
-
-        // Im2Col configuration
-        TensorShape shape_im2col = input->info()->tensor_shape();
-        shape_im2col.set(0, patch_size);
-        shape_im2col.set(1, conv_size);
-        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, dilation);
-        CLScheduler::get().tune_kernel_static(_im2col_kernel);
-
-        // Weights reshape configuration
-        const TensorShape shape_weights_reshape(patch_size, weights_z);
-        _weights_reshaped.allocator()->init(weights->info()->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(shape_weights_reshape));
-        _weights_reshape_kernel.configure(weights, &_weights_reshaped, append_bias ? biases : nullptr);
-
-        // GEMV configuration
-        DataType    v2mm_dt        = (input->info()->data_type() == DataType::QASYMM8) ? DataType::S32 : input->info()->data_type();
-        TensorShape shape_v2mm_out = input->info()->tensor_shape();
-        shape_v2mm_out.set(0, conv_size * weights_z);
-        shape_v2mm_out.set(1, 1);
-        shape_v2mm_out.set(2, 1);
-        _v2mm_output.allocator()->init(input->info()->clone()->set_is_resizable(true).reset_padding().set_data_type(v2mm_dt).set_tensor_shape(shape_v2mm_out));
-        _v2mm_kernel.set_target(gpu_target);
-        _v2mm_kernel.configure(&_input_reshaped, &_weights_reshaped, &_v2mm_output);
-        CLScheduler::get().tune_kernel_static(_v2mm_kernel);
-        _output_reshaped.allocator()->init(_v2mm_output.info()->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(output_shape));
-        _vector_to_tensor_kernel.configure(&_v2mm_output, (_is_quantized) ? &_output_reshaped : output, conv_w, conv_h);
-
-        // Output staged configuration
-        if(_is_quantized)
-        {
-            const UniformQuantizationInfo output_quant_info = (output->info()->total_size() == 0) ? iq_info : oq_info;
-
-            int         output_multiplier = 0;
-            int         output_shift      = 0;
-            const float multiplier        = iq_info.scale * wq_info.scale / output_quant_info.scale;
-            quantization::calculate_quantized_multiplier_less_than_one(multiplier, &output_multiplier, &output_shift);
-            _output_stage_kernel.configure(&_output_reshaped, biases, output, output_multiplier, output_shift, output_quant_info.offset);
-            _output_reshaped.allocator()->allocate();
-        }
+        _needs_permute = input->info()->data_layout() == DataLayout::NCHW;
 
-        // Fill borders on inputs
-        PixelValue zero_in(static_cast<int32_t>(0));
-        PixelValue zero_w(static_cast<int32_t>(0));
-        if(_is_quantized)
+        ICLTensor       *input_to_use   = input;
+        const ICLTensor *weights_to_use = weights;
+        ICLTensor       *output_to_use  = output;
+        if(_needs_permute)
         {
-            zero_in = PixelValue(static_cast<int32_t>(iq_info.offset));
-            zero_w  = PixelValue(static_cast<int32_t>(wq_info.offset));
-        }
-        BorderSize border_size = _v2mm_kernel.border_size();
-        _v2mm_input_fill_border.configure(&_input_reshaped, border_size, BorderMode::CONSTANT, zero_in);
+            _memory_group.manage(&_permuted_input);
+            _memory_group.manage(&_permuted_output);
 
-        border_size.bottom = 0;
-        _v2mm_weights_fill_border.configure(&_weights_reshaped, border_size, BorderMode::CONSTANT, zero_w);
+            // Configure the function to transform the input tensor from NCHW -> NHWC
+            _permute_input_to_nhwc.configure(input, &_permuted_input, PermutationVector(2U, 0U, 1U));
+            _permuted_input.info()->set_data_layout(DataLayout::NHWC);
 
-        // Allocate intermediate tensors
-        _input_reshaped.allocator()->allocate();
-        _v2mm_output.allocator()->allocate();
+            // Configure the function to transform the weights tensor from IHW -> HWI
+            _permute_weights_to_nhwc.configure(weights, &_permuted_weights, PermutationVector(2U, 0U, 1U));
+            _permuted_weights.info()->set_data_layout(DataLayout::NHWC);
+
+            // Set output quantization info before dwc kernel configure
+            _permuted_output.info()->set_quantization_info(output->info()->quantization_info());
+
+            input_to_use   = &_permuted_input;
+            weights_to_use = &_permuted_weights;
+            output_to_use  = &_permuted_output;
+        }
 
-        //Configure Activation Layer
-        _is_activationlayer_enabled = act_info.enabled();
+        DWCWeightsKernelInfo dwc_weights_info;
+        dwc_weights_info.n0 = (depth_multiplier == 1) ? 8 : 1;
+        DWCKernelInfo dwc_info;
+        dwc_info.activation_info = act_info;
+        _dwc_native_kernel.configure(input_to_use, weights_to_use, biases, output_to_use, dwc_weights_info, dwc_info, conv_info, depth_multiplier, dilation);
 
-        if(_is_activationlayer_enabled)
+        if(_needs_permute)
         {
-            _activationlayer_function.configure(output, nullptr, act_info);
+            _permuted_input.allocator()->allocate();
+
+            // Configure the function to transform the convoluted output to NCHW format
+            _permuted_output.info()->set_data_layout(DataLayout::NCHW);
+            _permute_output_to_nchw.configure(&_permuted_output, output, PermutationVector(1U, 2U, 0U));
+            _permuted_output.allocator()->allocate();
         }
     }
 }
@@ -380,6 +339,8 @@ void CLDepthwiseConvolutionLayer::configure(ICLTensor *input, const ICLTensor *w
 Status CLDepthwiseConvolutionLayer::validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info,
                                              unsigned int depth_multiplier, const ActivationLayerInfo &act_info, const Size2D &dilation)
 {
+    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_LAYOUT(input, output);
+
     const size_t idx_w = get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::WIDTH);
     const size_t idx_h = get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::HEIGHT);
 
@@ -390,60 +351,36 @@ Status CLDepthwiseConvolutionLayer::validate(const ITensorInfo *input, const ITe
 
     if(!can_run_optimised_3x3_kernel)
     {
-        const size_t idx_c = get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::CHANNEL);
-
-        ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, weights, output);
-        ARM_COMPUTE_RETURN_ERROR_ON((input->dimension(idx_c) * depth_multiplier) != weights->dimension(idx_c));
-
-        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, 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);
-        const unsigned int conv_w       = output_shape[idx_w];
-        const unsigned int conv_h       = output_shape[idx_h];
-        const size_t       patch_size   = weights_w * weights_h + ((append_bias) ? 1 : 0);
-        const size_t       conv_size    = conv_w * conv_h;
-
-        TensorShape shape_im2col = input->tensor_shape();
-        shape_im2col.set(0, patch_size);
-        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, 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));
-        ARM_COMPUTE_RETURN_ON_ERROR(CLDepthwiseConvolutionLayerReshapeWeightsGenericKernel::validate(weights, &weights_reshaped, append_bias ? biases : nullptr));
-
-        DataType    v2mm_dt        = (input->data_type() == DataType::QASYMM8) ? DataType::S32 : input->data_type();
-        TensorShape shape_v2mm_out = input->tensor_shape();
-        shape_v2mm_out.set(0, conv_size * weights_z);
-        shape_v2mm_out.set(1, 1);
-        shape_v2mm_out.set(2, 1);
-        TensorInfo v2mm_output(input->clone()->set_is_resizable(true).reset_padding().set_data_type(v2mm_dt).set_tensor_shape(shape_v2mm_out));
-        ARM_COMPUTE_RETURN_ON_ERROR(CLGEMMMatrixVectorMultiplyKernel::validate(&input_reshaped, &weights_reshaped, &v2mm_output));
-
-        TensorInfo output_reshaped(v2mm_output.clone()->set_is_resizable(true).reset_padding().set_tensor_shape(output_shape));
-        ARM_COMPUTE_RETURN_ON_ERROR(CLDepthwiseVectorToTensorKernel::validate(&v2mm_output, (is_quantized) ? &output_reshaped : output, conv_w, conv_h));
-
-        if(is_quantized)
+        DWCWeightsKernelInfo dwc_weights_info;
+        dwc_weights_info.n0 = (depth_multiplier == 1) ? 8 : 1;
+        DWCKernelInfo dwc_info;
+        dwc_info.activation_info = act_info;
+
+        const bool needs_permute = input->data_layout() == DataLayout::NCHW;
+
+        if(needs_permute)
         {
-            const UniformQuantizationInfo iq_info = input->quantization_info().uniform();
-            const UniformQuantizationInfo wq_info = weights->quantization_info().uniform();
-            const UniformQuantizationInfo oq_info = (output->total_size() == 0) ? iq_info : output->quantization_info().uniform();
-
-            const float multiplier = iq_info.scale * wq_info.scale / oq_info.scale;
-            ARM_COMPUTE_UNUSED(multiplier);
-            ARM_COMPUTE_RETURN_ERROR_ON(multiplier > 1.0f);
-            ARM_COMPUTE_RETURN_ON_ERROR(CLDirectConvolutionLayerOutputStageKernel::validate(&output_reshaped, biases, output));
+            TensorShape permuted_input_shape   = input->tensor_shape();
+            TensorShape permuted_weights_shape = weights->tensor_shape();
+            TensorShape permuted_output_shape  = shape_calculator::compute_depthwise_convolution_shape(*input, *weights, conv_info, depth_multiplier, dilation);
+
+            permute(permuted_input_shape, PermutationVector(2U, 0U, 1U));
+            permute(permuted_weights_shape, PermutationVector(2U, 0U, 1U));
+            permute(permuted_output_shape, PermutationVector(2U, 0U, 1U));
+
+            const TensorInfo permuted_input   = input->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(permuted_input_shape).set_data_layout(DataLayout::NHWC);
+            const TensorInfo permuted_weights = weights->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(permuted_weights_shape).set_data_layout(DataLayout::NHWC);
+            const TensorInfo permuted_output  = output->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(permuted_output_shape).set_data_layout(DataLayout::NHWC);
+
+            ARM_COMPUTE_RETURN_ON_ERROR(CLPermute::validate(input, &permuted_input, PermutationVector(2U, 0U, 1U)));
+            ARM_COMPUTE_RETURN_ON_ERROR(CLPermute::validate(weights, &permuted_weights, PermutationVector(2U, 0U, 1U)));
+            ARM_COMPUTE_RETURN_ON_ERROR(CLDepthwiseConvolutionLayerNativeKernel::validate(&permuted_input, &permuted_weights, biases, &permuted_output, dwc_weights_info,
+                                                                                          dwc_info, conv_info, depth_multiplier, dilation));
+            ARM_COMPUTE_RETURN_ON_ERROR(CLPermute::validate(&permuted_output, output, PermutationVector(1U, 2U, 0U)));
         }
-
-        // Validate Activation Layer
-        if(act_info.enabled())
+        else
         {
-            ARM_COMPUTE_RETURN_ON_ERROR(CLActivationLayer::validate(output, nullptr, act_info));
+            ARM_COMPUTE_RETURN_ON_ERROR(CLDepthwiseConvolutionLayerNativeKernel::validate(input, weights, biases, output, dwc_weights_info, dwc_info, conv_info, depth_multiplier, dilation));
         }
     }
     else
@@ -457,23 +394,22 @@ void CLDepthwiseConvolutionLayer::run()
 {
     prepare();
 
+    MemoryGroupResourceScope scope_mg(_memory_group);
+
     if(_optimised_function != nullptr)
     {
         _optimised_function->run();
     }
     else
     {
-        CLScheduler::get().enqueue(_im2col_kernel);
-        CLScheduler::get().enqueue(_v2mm_input_fill_border);
-        CLScheduler::get().enqueue(_v2mm_kernel);
-        CLScheduler::get().enqueue(_vector_to_tensor_kernel);
-        if(_is_quantized)
+        if(_needs_permute)
         {
-            CLScheduler::get().enqueue(_output_stage_kernel);
+            _permute_input_to_nhwc.run();
         }
-        if(_is_activationlayer_enabled)
+        CLScheduler::get().enqueue(_dwc_native_kernel);
+        if(_needs_permute)
         {
-            _activationlayer_function.run();
+            _permute_output_to_nchw.run();
         }
     }
 }
@@ -484,21 +420,17 @@ void CLDepthwiseConvolutionLayer::prepare()
     {
         _optimised_function->prepare();
     }
-    else
+    else if(!_is_prepared)
     {
-        if(!_is_prepared)
+        if(_needs_permute)
         {
             ARM_COMPUTE_ERROR_ON(!_original_weights->is_used());
 
-            // Run weights reshaping and mark original weights tensor as unused
-            _weights_reshaped.allocator()->allocate();
-            CLScheduler::get().enqueue(_weights_reshape_kernel);
-            CLScheduler::get().enqueue(_v2mm_weights_fill_border);
+            _permuted_weights.allocator()->allocate();
+            _permute_weights_to_nhwc.run();
             _original_weights->mark_as_unused();
-
-            CLScheduler::get().queue().finish();
-            _is_prepared = true;
         }
+        _is_prepared = true;
     }
 }
 } // namespace arm_compute
diff --git a/src/runtime/CL/tuners/BifrostTuner.cpp b/src/runtime/CL/tuners/BifrostTuner.cpp
index 187f52fcf7..5b23baaed3 100644
--- a/src/runtime/CL/tuners/BifrostTuner.cpp
+++ b/src/runtime/CL/tuners/BifrostTuner.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2018 ARM Limited.
+ * Copyright (c) 2018-2019 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -171,24 +171,6 @@ void tune_im2col_kernel(CLIm2ColKernel &k)
     k.set_lws_hint(lws_hint);
 }
 
-void tune_depthwise_im2col_kernel(CLDepthwiseIm2ColKernel &k)
-{
-    cl::NDRange     lws_hint   = k.lws_hint();
-    const GPUTarget gpu_target = k.get_target();
-
-    // Configure the local work size for Bifrost with a value obtained
-    // via exhaustive autotuning for the MobileNets tensor shapes.
-    if(gpu_target_is_in(gpu_target,
-                        GPUTarget::G71, GPUTarget::G72, GPUTarget::G76,
-                        GPUTarget::G51, GPUTarget::G51BIG, GPUTarget::G51LIT,
-                        GPUTarget::G52, GPUTarget::G52LIT))
-    {
-        lws_hint = cl::NDRange(1, 2, 1);
-    }
-
-    k.set_lws_hint(lws_hint);
-}
-
 void tune_gemv_kernel(CLGEMMMatrixVectorMultiplyKernel &k)
 {
     cl::NDRange     lws_hint   = k.lws_hint();
@@ -311,10 +293,6 @@ void BifrostTuner::tune_kernel_static(ICLKernel &kernel)
     {
         tune_im2col_kernel(*utils::cast::polymorphic_downcast<CLIm2ColKernel *>(&kernel));
     }
-    else if(dynamic_cast<CLDepthwiseIm2ColKernel *>(&kernel) != nullptr)
-    {
-        tune_depthwise_im2col_kernel(*utils::cast::polymorphic_downcast<CLDepthwiseIm2ColKernel *>(&kernel));
-    }
     else if(dynamic_cast<CLGEMMMatrixVectorMultiplyKernel *>(&kernel) != nullptr)
     {
         tune_gemv_kernel(*utils::cast::polymorphic_downcast<CLGEMMMatrixVectorMultiplyKernel *>(&kernel));
diff --git a/tests/datasets/DepthwiseConvolutionLayerDataset.h b/tests/datasets/DepthwiseConvolutionLayerDataset.h
index d0617275c0..2990b135d2 100644
--- a/tests/datasets/DepthwiseConvolutionLayerDataset.h
+++ b/tests/datasets/DepthwiseConvolutionLayerDataset.h
@@ -120,7 +120,7 @@ class SmallDepthwiseConvolutionLayerDataset final : public DepthwiseConvolutionL
 public:
     SmallDepthwiseConvolutionLayerDataset()
     {
-        add_config(TensorShape(7U, 7U, 1U), Size2D(3U, 3U), PadStrideInfo(1, 1, 0, 0));
+        add_config(TensorShape(7U, 7U, 1U), Size2D(1U, 1U), PadStrideInfo(1, 1, 0, 0));
         add_config(TensorShape(23U, 27U, 5U), Size2D(3U, 5U), PadStrideInfo(2, 1, 0, 0));
         add_config(TensorShape(33U, 27U, 7U), Size2D(7U, 3U), PadStrideInfo(3, 2, 1, 0));
         // Asymmetric padding
@@ -135,16 +135,16 @@ class LargeDepthwiseConvolutionLayerDataset final : public DepthwiseConvolutionL
 public:
     LargeDepthwiseConvolutionLayerDataset()
     {
-        add_config(TensorShape(33U, 27U, 11U), Size2D(3U, 3U), PadStrideInfo(1, 2, 0, 1));
+        add_config(TensorShape(33U, 27U, 11U), Size2D(3U, 4U), PadStrideInfo(1, 2, 0, 1));
         add_config(TensorShape(17U, 31U, 2U), Size2D(5U, 9U), PadStrideInfo(1, 2, 1, 1));
         add_config(TensorShape(23U, 27U, 5U), Size2D(11U, 3U), PadStrideInfo(1, 2, 0, 0));
         add_config(TensorShape(17U, 31U, 2U, 3U), Size2D(5U, 9U), PadStrideInfo(1, 2, 1, 1));
-        add_config(TensorShape(233U, 277U, 55U), Size2D(3U, 3U), PadStrideInfo(2, 1, 0, 0));
-        add_config(TensorShape(333U, 277U, 77U), Size2D(3U, 3U), PadStrideInfo(3, 2, 1, 0));
-        add_config(TensorShape(177U, 311U, 22U), Size2D(3U, 3U), PadStrideInfo(1, 2, 1, 1));
-        add_config(TensorShape(233U, 277U, 55U), Size2D(3U, 3U), PadStrideInfo(1, 2, 0, 0));
-        add_config(TensorShape(333U, 277U, 77U), Size2D(3U, 3U), PadStrideInfo(2, 3, 0, 1));
-        add_config(TensorShape(177U, 311U, 22U), Size2D(3U, 3U), PadStrideInfo(2, 1, 1, 1));
+        add_config(TensorShape(233U, 277U, 55U), Size2D(1U, 1U), PadStrideInfo(2, 1, 0, 0));
+        add_config(TensorShape(333U, 277U, 77U), Size2D(1U, 1U), PadStrideInfo(3, 2, 1, 0));
+        add_config(TensorShape(177U, 311U, 22U), Size2D(3U, 4U), PadStrideInfo(1, 2, 1, 1));
+        add_config(TensorShape(233U, 277U, 55U), Size2D(3U, 4U), PadStrideInfo(1, 2, 0, 0));
+        add_config(TensorShape(333U, 277U, 77U), Size2D(3U, 4U), PadStrideInfo(2, 3, 0, 1));
+        add_config(TensorShape(177U, 311U, 22U), Size2D(3U, 4U), PadStrideInfo(2, 1, 1, 1));
         // Asymmetric padding
         add_config(TensorShape(33U, 27U, 7U), Size2D(5U, 7U), PadStrideInfo(3, 2, 2, 1, 2, 0, DimensionRoundingType::FLOOR));
         add_config(TensorShape(33U, 27U, 7U), Size2D(5U, 7U), PadStrideInfo(3, 2, 1, 3, 0, 2, DimensionRoundingType::FLOOR));