COMPMID-811 Add NHWC data format support for CL depthwise convolution QASYMM8

Change-Id: I89de432f3fbcba7abf9e1d4f8396a4334b4fa2c2
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/118324
Tested-by: Jenkins <bsgcomp@arm.com>
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>

17 files changed, 1005 insertions, 84 deletions

diff --git a/arm_compute/core/CL/CLKernels.h b/arm_compute/core/CL/CLKernels.h
index 9481e144df..67b7a8bb8f 100644
--- a/arm_compute/core/CL/CLKernels.h
+++ b/arm_compute/core/CL/CLKernels.h
@@ -46,7 +46,8 @@
 #include "arm_compute/core/CL/kernels/CLDeconvolutionLayerUpsampleKernel.h"
 #include "arm_compute/core/CL/kernels/CLDepthConcatenateLayerKernel.h"
 #include "arm_compute/core/CL/kernels/CLDepthConvertLayerKernel.h"
-#include "arm_compute/core/CL/kernels/CLDepthwiseConvolutionLayer3x3Kernel.h"
+#include "arm_compute/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NCHWKernel.h"
+#include "arm_compute/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NHWCKernel.h"
 #include "arm_compute/core/CL/kernels/CLDepthwiseIm2ColKernel.h"
 #include "arm_compute/core/CL/kernels/CLDepthwiseVectorToTensorKernel.h"
 #include "arm_compute/core/CL/kernels/CLDepthwiseWeightsReshapeKernel.h"
@@ -113,5 +114,6 @@
 #include "arm_compute/core/CL/kernels/CLWinogradFilterTransformKernel.h"
 #include "arm_compute/core/CL/kernels/CLWinogradInputTransformKernel.h"
 #include "arm_compute/core/CL/kernels/CLWinogradOutputTransformKernel.h"
+#include "arm_compute/core/CL/kernels/ICLDepthwiseConvolutionLayer3x3Kernel.h"
 
 #endif /* __ARM_COMPUTE_CLKERNELS_H__ */
diff --git a/arm_compute/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NCHWKernel.h b/arm_compute/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NCHWKernel.h
new file mode 100644
index 0000000000..0f3f4bfc76
--- /dev/null
+++ b/arm_compute/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NCHWKernel.h
@@ -0,0 +1,61 @@
+/*
+ * Copyright (c) 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_CLDEPTHWISECONVOLUTIONNCHWKERNEL3x3_H__
+#define __ARM_COMPUTE_CLDEPTHWISECONVOLUTIONNCHWKERNEL3x3_H__
+
+#include "arm_compute/core/CL/kernels/ICLDepthwiseConvolutionLayer3x3Kernel.h"
+
+namespace arm_compute
+{
+class ICLTensor;
+
+/** Interface for the kernel to run a 3x3 depthwise convolution on a tensor when the data layout is NCHW.
+ */
+class CLDepthwiseConvolutionLayer3x3NCHWKernel : public ICLDepthwiseConvolutionLayer3x3Kernel
+{
+public:
+    /** Default constructor */
+    CLDepthwiseConvolutionLayer3x3NCHWKernel();
+    /** Initialize the function's source, destination, conv and border_size.
+     *
+     * @param[in]  input     Source tensor. DataType supported: QASYMM8/F16/F32.
+     * @param[in]  weights   Weights tensor. A 3D tensor with dimensions [3, 3, IFM]. Data type supported: Same as @p input.
+     * @param[in]  biases    (Optional) Biases tensor. A 1D tensor with dimensions [IFM]. Must be nullptr if not needed.
+     *                       Data type supported: Same as @p input.
+     * @param[out] output    Destination tensor. Data type supported: Same as @p input.
+     * @param[in]  conv_info Padding and stride information to use for the convolution.
+     * @param[in]  act_info  (Optional) Activation layer information in case of a fused activation. Only RELU, BOUNDED_RELU and LU_BOUNDED_RELU for QASYMM8 supported.
+     */
+    void configure(const ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const PadStrideInfo &conv_info,
+                   ActivationLayerInfo act_info) override;
+
+    void run(const Window &window, cl::CommandQueue &queue) override;
+    BorderSize border_size() const override;
+
+private:
+    unsigned int _conv_stride_x;
+    unsigned int _conv_pad_top;
+};
+} // namespace arm_compute
+#endif /*__ARM_COMPUTE_CLDEPTHWISECONVOLUTIONNCHWKERNEL3x3_H__ */
diff --git a/arm_compute/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NHWCKernel.h b/arm_compute/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NHWCKernel.h
new file mode 100644
index 0000000000..4ecc07af6a
--- /dev/null
+++ b/arm_compute/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NHWCKernel.h
@@ -0,0 +1,75 @@
+/*
+ * Copyright (c) 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_CLDEPTHWISECONVOLUTIONNHWCKERNEL3x3_H__
+#define __ARM_COMPUTE_CLDEPTHWISECONVOLUTIONNHWCKERNEL3x3_H__
+
+#include "arm_compute/core/CL/kernels/ICLDepthwiseConvolutionLayer3x3Kernel.h"
+
+namespace arm_compute
+{
+class ICLTensor;
+
+/** Interface for the kernel to run a 3x3 depthwise convolution on a tensor when the data layout is NHWC.
+ */
+class CLDepthwiseConvolutionLayer3x3NHWCKernel : public ICLDepthwiseConvolutionLayer3x3Kernel
+{
+public:
+    /** Default constructor */
+    CLDepthwiseConvolutionLayer3x3NHWCKernel();
+    /** Default move assignment operator. */
+    /** Initialize the function's source, destination, conv and border_size.
+     *
+     * @param[in]  input     Source tensor. DataType supported: QASYMM8.
+     * @param[in]  weights   Weights tensor. A 3D tensor with dimensions [IFM, 3, 3]. Data type supported: Same as @p input.
+     * @param[in]  biases    (Optional) Biases tensor. A 1D tensor with dimensions [IFM]. Must be nullptr if not needed.
+     *                       Data type supported: Same as @p input.
+     * @param[out] output    Destination tensor. Data type supported: Same as @p input.
+     * @param[in]  conv_info Padding and stride information to use for the convolution.
+     * @param[in]  act_info  (Optional) Activation layer information in case of a fused activation. Only RELU, BOUNDED_RELU and LU_BOUNDED_RELU are supported.
+     */
+    void configure(const ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const PadStrideInfo &conv_info,
+                   ActivationLayerInfo act_info) override;
+    /** Static function to check if given info will lead to a valid configuration of @ref CLDepthwiseConvolutionLayer3x3NHWCKernel
+     *
+     * @param[in] input     Source tensor. DataType supported: QASYMM8.
+     * @param[in] weights   Weights tensor. A 3D tensor with dimensions [IFM, 3, 3]. Data type supported: Same as @p input.
+     * @param[in] biases    (Optional) Biases tensor. A 1D tensor with dimensions [IFM]. Must be nullptr if not needed.
+     *                      Data type supported: Same as @p input.
+     * @param[in] output    Destination tensor. Data type supported: Same as @p input.
+     * @param[in] conv_info Padding and stride information to use for the convolution.
+     * @param[in] act_info  (Optional) Activation layer information in case of a fused activation. Only RELU, BOUNDED_RELU and LU_BOUNDED_RELU are supported.
+     *
+     * @return a status
+     */
+    static Status validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info,
+                           ActivationLayerInfo act_info = ActivationLayerInfo());
+
+    void run(const Window &window, cl::CommandQueue &queue) override;
+    BorderSize border_size() const override;
+
+private:
+    unsigned int _num_rows_processed_per_iteration;
+};
+} // namespace arm_compute
+#endif /*__ARM_COMPUTE_CLDEPTHWISECONVOLUTIONNHWCKERNEL3x3_H__ */
diff --git a/arm_compute/core/CL/kernels/CLDepthwiseConvolutionLayer3x3Kernel.h b/arm_compute/core/CL/kernels/ICLDepthwiseConvolutionLayer3x3Kernel.h
index 5f72cf70ed..f02ba331b5 100644
--- a/arm_compute/core/CL/kernels/CLDepthwiseConvolutionLayer3x3Kernel.h
+++ b/arm_compute/core/CL/kernels/ICLDepthwiseConvolutionLayer3x3Kernel.h
@@ -21,8 +21,8 @@
  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  * SOFTWARE.
  */
-#ifndef __ARM_COMPUTE_CLDEPTHWISECONVOLUTIONKERNEL3x3_H__
-#define __ARM_COMPUTE_CLDEPTHWISECONVOLUTIONKERNEL3x3_H__
+#ifndef __ARM_COMPUTE_ICLDEPTHWISECONVOLUTIONKERNEL3x3_H__
+#define __ARM_COMPUTE_ICLDEPTHWISECONVOLUTIONKERNEL3x3_H__
 
 #include "arm_compute/core/CL/ICLKernel.h"
 
@@ -32,19 +32,22 @@ class ICLTensor;
 
 /** Interface for the kernel to run a 3x3 depthwise convolution on a tensor.
  */
-class CLDepthwiseConvolutionLayer3x3Kernel : public ICLKernel
+class ICLDepthwiseConvolutionLayer3x3Kernel : public ICLKernel
 {
 public:
     /** Default constructor */
-    CLDepthwiseConvolutionLayer3x3Kernel();
+    ICLDepthwiseConvolutionLayer3x3Kernel()
+        : _border_size(0), _input(), _output(), _weights(), _biases(), _conv_stride_y(1), _conv_pad_left(0)
+    {
+    }
     /** Prevent instances of this class from being copied (As this class contains pointers) */
-    CLDepthwiseConvolutionLayer3x3Kernel(const CLDepthwiseConvolutionLayer3x3Kernel &) = delete;
+    ICLDepthwiseConvolutionLayer3x3Kernel(const ICLDepthwiseConvolutionLayer3x3Kernel &) = delete;
     /** Prevent instances of this class from being copied (As this class contains pointers) */
-    CLDepthwiseConvolutionLayer3x3Kernel &operator=(const CLDepthwiseConvolutionLayer3x3Kernel &) = delete;
+    ICLDepthwiseConvolutionLayer3x3Kernel &operator=(const ICLDepthwiseConvolutionLayer3x3Kernel &) = delete;
     /** Default Move Constructor. */
-    CLDepthwiseConvolutionLayer3x3Kernel(CLDepthwiseConvolutionLayer3x3Kernel &&) = default;
+    ICLDepthwiseConvolutionLayer3x3Kernel(ICLDepthwiseConvolutionLayer3x3Kernel &&) = default;
     /** Default move assignment operator */
-    CLDepthwiseConvolutionLayer3x3Kernel &operator=(CLDepthwiseConvolutionLayer3x3Kernel &&) = default;
+    ICLDepthwiseConvolutionLayer3x3Kernel &operator=(ICLDepthwiseConvolutionLayer3x3Kernel &&) = default;
     /** Initialize the function's source, destination, conv and border_size.
      *
      * @param[in]  input     Source tensor. DataType supported: QASYMM8/F16/F32.
@@ -55,22 +58,17 @@ public:
      * @param[in]  conv_info Padding and stride information to use for the convolution.
      * @param[in]  act_info  (Optional) Activation layer information in case of a fused activation. Only RELU, BOUNDED_RELU and LU_BOUNDED_RELU for QASYMM8 supported.
      */
-    void configure(const ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const PadStrideInfo &conv_info, ActivationLayerInfo act_info = ActivationLayerInfo());
+    virtual void configure(const ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const PadStrideInfo &conv_info,
+                           ActivationLayerInfo act_info = ActivationLayerInfo()) = 0;
 
-    // Inherited methods overridden:
-    void run(const Window &window, cl::CommandQueue &queue) override;
-    BorderSize border_size() const override;
-
-private:
+protected:
     BorderSize       _border_size;
     const ICLTensor *_input;
     ICLTensor       *_output;
     const ICLTensor *_weights;
     const ICLTensor *_biases;
-    unsigned int     _conv_stride_x;
     unsigned int     _conv_stride_y;
     unsigned int     _conv_pad_left;
-    unsigned int     _conv_pad_top;
 };
 } // namespace arm_compute
-#endif /*__ARM_COMPUTE_CLDEPTHWISECONVOLUTIONKERNEL3x3_H__ */
+#endif /*__ARM_COMPUTE_ICLDEPTHWISECONVOLUTIONKERNEL3x3_H__ */
diff --git a/arm_compute/core/utils/misc/ShapeCalculator.h b/arm_compute/core/utils/misc/ShapeCalculator.h
index e174227302..b91e52a657 100644
--- a/arm_compute/core/utils/misc/ShapeCalculator.h
+++ b/arm_compute/core/utils/misc/ShapeCalculator.h
@@ -130,15 +130,19 @@ inline TensorShape compute_depthwise_convolution_shape(const ITensorInfo &input,
     const TensorShape input_shape{ input.tensor_shape() };
     const TensorShape weights_shape{ weights.tensor_shape() };
 
+    const DataLayout data_layout = input.data_layout();
+    const int        width_idx   = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH);
+    const int        height_idx  = get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT);
+
     unsigned int output_width  = 0;
     unsigned int output_height = 0;
-    std::tie(output_width, output_height) = scaled_dimensions(input_shape.x(), input_shape.y(),
-                                                              weights_shape.x(), weights_shape.y(),
+    std::tie(output_width, output_height) = scaled_dimensions(input_shape[width_idx], input_shape[height_idx],
+                                                              weights_shape[width_idx], weights_shape[height_idx],
                                                               conv_info);
 
     TensorShape output_shape{ input_shape };
-    output_shape.set(0, output_width);
-    output_shape.set(1, output_height);
+    output_shape.set(width_idx, output_width);
+    output_shape.set(height_idx, output_height);
 
     return output_shape;
 }
diff --git a/arm_compute/runtime/CL/functions/CLDepthwiseConvolutionLayer.h b/arm_compute/runtime/CL/functions/CLDepthwiseConvolutionLayer.h
index d6fc8f0fcc..82947bc7e6 100644
--- a/arm_compute/runtime/CL/functions/CLDepthwiseConvolutionLayer.h
+++ b/arm_compute/runtime/CL/functions/CLDepthwiseConvolutionLayer.h
@@ -24,13 +24,15 @@
 #ifndef __ARM_COMPUTE_CLDEPTHWISECONVOLUTION_H__
 #define __ARM_COMPUTE_CLDEPTHWISECONVOLUTION_H__
 
-#include "arm_compute/core/CL/kernels/CLDepthwiseConvolutionLayer3x3Kernel.h"
+#include "arm_compute/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NCHWKernel.h"
+#include "arm_compute/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NHWCKernel.h"
 #include "arm_compute/core/CL/kernels/CLDepthwiseIm2ColKernel.h"
 #include "arm_compute/core/CL/kernels/CLDepthwiseVectorToTensorKernel.h"
 #include "arm_compute/core/CL/kernels/CLDepthwiseWeightsReshapeKernel.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/IFunction.h"
@@ -39,9 +41,10 @@ namespace arm_compute
 {
 class ICLTensor;
 
-/** Basic function to execute a depthwise convolution for kernel size 3x3xC. This function calls the following OpenCL kernels:
+/** Basic function to execute a depthwise convolution for kernel size 3x3xC (when data layout NCHW) or Cx3x3 (when data layout NHWC). This function calls the following OpenCL kernels:
  *
- * -# @ref CLDepthwiseConvolutionLayer3x3Kernel
+ * -# @ref CLDepthwiseConvolutionLayer3x3NCHWKernel (if data_layout == NCHW)
+ * -# @ref CLDepthwiseConvolutionLayer3x3NHWCKernel (if data_layout == NHWC)
  * -# @ref CLFillBorderKernel (if pad_x or pad_y > 0)
  *
  */
@@ -66,8 +69,8 @@ public:
     void run() override;
 
 private:
-    CLDepthwiseConvolutionLayer3x3Kernel _kernel;
-    CLFillBorderKernel                   _border_handler;
+    std::unique_ptr<ICLDepthwiseConvolutionLayer3x3Kernel> _kernel;
+    CLFillBorderKernel                                     _border_handler;
 };
 
 /** Basic function to execute a generic depthwise convolution. This function calls the following OpenCL kernels:
diff --git a/docs/00_introduction.dox b/docs/00_introduction.dox
index 17d1bdea63..39a7ee1a79 100644
--- a/docs/00_introduction.dox
+++ b/docs/00_introduction.dox
@@ -356,7 +356,7 @@ v17.09 Public major release
     - @ref NEReshapeLayerKernel / @ref NEReshapeLayer
 
  - New OpenCL kernels / functions:
-    - @ref CLDepthwiseConvolutionLayer3x3Kernel @ref CLDepthwiseIm2ColKernel @ref CLDepthwiseVectorToTensorKernel @ref CLDepthwiseWeightsReshapeKernel / @ref CLDepthwiseConvolutionLayer3x3 @ref CLDepthwiseConvolutionLayer @ref CLDepthwiseSeparableConvolutionLayer
+    - @ref CLDepthwiseConvolutionLayer3x3NCHWKernel @ref CLDepthwiseConvolutionLayer3x3NHWCKernel @ref CLDepthwiseIm2ColKernel @ref CLDepthwiseVectorToTensorKernel @ref CLDepthwiseWeightsReshapeKernel / @ref CLDepthwiseConvolutionLayer3x3 @ref CLDepthwiseConvolutionLayer @ref 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 8405754f52..0509f1e785 100644
--- a/src/core/CL/CLKernelLibrary.cpp
+++ b/src/core/CL/CLKernelLibrary.cpp
@@ -191,7 +191,9 @@ const std::map<std::string, std::string> CLKernelLibrary::_kernel_program_map =
     { "deconvolution_upsample", "deconvolution_layer.cl" },
     { "depthwise_convolution_3x3", "depthwise_convolution.cl" },
     { "depthwise_convolution_3x3_f16", "depthwise_convolution.cl" },
-    { "depthwise_convolution_3x3_quantized", "depthwise_convolution_quantized.cl" },
+    { "depthwise_convolution_3x3_quantized_nchw", "depthwise_convolution_quantized.cl" },
+    { "depthwise_convolution_3x3_quantized_nhwc_stride1", "depthwise_convolution_quantized.cl" },
+    { "depthwise_convolution_3x3_quantized_nhwc_stride2", "depthwise_convolution_quantized.cl" },
     { "depthwise_convolution_3x3_stridex1_stridey1_bifrost_f16", "depthwise_convolution.cl" },
     { "depthwise_convolution_3x3_stridex2_stridey2_bifrost_f16", "depthwise_convolution.cl" },
     { "depthwise_convolution_3x3_stridex1_stridey1_bifrost_f32", "depthwise_convolution.cl" },
diff --git a/src/core/CL/cl_kernels/depthwise_convolution_quantized.cl b/src/core/CL/cl_kernels/depthwise_convolution_quantized.cl
index 21daee8230..635bc9d50b 100644
--- a/src/core/CL/cl_kernels/depthwise_convolution_quantized.cl
+++ b/src/core/CL/cl_kernels/depthwise_convolution_quantized.cl
@@ -24,17 +24,21 @@
 
 #include "helpers_asymm.h"
 
-#if defined(CONV_STRIDE_X) && defined(CONV_STRIDE_Y) && defined(WEIGHTS_OFFSET) && defined(INPUT_OFFSET) && defined(K_OFFSET) && defined(OUTPUT_OFFSET) && defined(OUTPUT_MULTIPLIER) && defined(OUTPUT_SHIFT)
+#if defined(WEIGHTS_OFFSET) && defined(INPUT_OFFSET) && defined(K_OFFSET) && defined(OUTPUT_OFFSET) && defined(OUTPUT_MULTIPLIER) && defined(OUTPUT_SHIFT)
 
 #if defined(FUSED_ACTIVATION)
 #define DATA_TYPE uchar
+#ifndef VEC_SIZE
 #define VEC_SIZE 8
+#endif /* VEC_SIZE */
 #include "activation_layer_qa8.cl"
 #define ACTIVATION_FUNC(x) PERFORM_ACTIVATION_QA8(FUSED_ACTIVATION, x)
 #else /* defined(FUSED_ACTIVATION) */
 #define ACTIVATION_FUNC(x) (x)
 #endif /* defined(FUSED_ACTIVATION) */
 
+#if defined(CONV_STRIDE_Y) && defined(CONV_STRIDE_X)
+
 #if CONV_STRIDE_X > 3
 #error "Stride X not supported"
 #endif /* CONV_STRIDE_X > 3 */
@@ -71,7 +75,7 @@
     })
 #endif /* CONV_STRIDE_X */
 
-/** This function computes the horizontal integral of the image and adds offsets.
+/** This function computes the depthwise convolution quantized.
  *
  * @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)
@@ -103,7 +107,7 @@
  * @param[in] biases_offset_first_element_in_bytes  (Optional) The offset of the first element in the biases vector
  */
 
-__kernel void depthwise_convolution_3x3_quantized(
+__kernel void depthwise_convolution_3x3_quantized_nchw(
     TENSOR3D_DECLARATION(src),
     TENSOR3D_DECLARATION(dst),
     TENSOR3D_DECLARATION(weights)
@@ -244,4 +248,478 @@ __kernel void depthwise_convolution_3x3_quantized(
 #endif /* CONV_STRIDE_Y == 1 */
 }
 
-#endif /* defined(CONV_STRIDE_X) && defined(CONV_STRIDE_Y) && defined(WEIGHTS_OFFSET) && defined(INPUT_OFFSET) && defined(K_OFFSET) && defined(OUTPUT_OFFSET) && defined(OUTPUT_MULTIPLIER) && defined(OUTPUT_SHIFT) */
+#endif /* defined(CONV_STRIDE_Y) && defined(CONV_STRIDE_X) */
+
+#if defined(VEC_SIZE) && defined(SRC_DEPTH) && defined(CONV_PAD_TOP) && defined(ROWS_READ)
+
+#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_UCHAR VEC_DATA_TYPE(uchar, VEC_SIZE)
+
+#define BIFROST_MAD_4(acc, x, y)               \
+    ({                                         \
+        acc.s0 += (ushort)x.s0 * (ushort)y.s0; \
+        acc.s1 += (ushort)x.s1 * (ushort)y.s1; \
+        acc.s2 += (ushort)x.s2 * (ushort)y.s2; \
+        acc.s3 += (ushort)x.s3 * (ushort)y.s3; \
+    })
+
+#if WEIGHTS_OFFSET != 0
+#define BIFROST_MAD_ACC_4(acc, sum, x, y) \
+    ({                                    \
+        sum += CONVERT(x, VEC_INT);       \
+        BIFROST_MAD_4(acc, x, y);         \
+    })
+#else /* WEIGHTS_OFFSET != 0 */
+#define BIFROST_MAD_ACC_4(acc, sum, x, y) BIFROST_MAD_4(acc, x, y)
+#endif /* WEIGHTS_OFFSET != 0 */
+
+/** This function computes the depthwise convolution quantized.
+ *
+ * @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] 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: QASYMM8
+ * @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_3x3_quantized_nhwc_stride1(
+    TENSOR3D_DECLARATION(src),
+    TENSOR3D_DECLARATION(dst),
+    TENSOR3D_DECLARATION(weights),
+#if defined(HAS_BIAS)
+    VECTOR_DECLARATION(biases)
+#endif /* defined(HAS_BIAS) */
+)
+{
+    Image  dst     = CONVERT_TENSOR3D_TO_IMAGE_STRUCT(dst);
+    Vector weights = CONVERT_TO_VECTOR_STRUCT(weights);
+#if defined(HAS_BIAS)
+    Vector biases = CONVERT_TO_VECTOR_STRUCT(biases);
+
+    VEC_INT bias_values = VLOAD(VEC_SIZE)(0, (__global int *)biases.ptr);
+#endif /* defined(HAS_BIAS) */
+
+    __global uchar *first_elem = src_ptr + src_offset_first_element_in_bytes;
+
+    const int z         = get_global_id(2);
+    const int pad_offs  = -ROWS_READ * src_stride_y;
+    const int src_offs0 = get_global_id(0) * src_step_x + get_global_id(1) * src_step_y + z * src_step_z - CONV_PAD_TOP * src_stride_z;
+    const int src_offs1 = src_offs0 + src_stride_z;
+    const int src_offs2 = src_offs1 + src_stride_z;
+
+    const int cond_top    = z - CONV_PAD_TOP < 0;
+    const int cond_bottom = z * (src_step_z / src_stride_z) + 2 >= SRC_DEPTH;
+
+    __global uchar *src_addr0 = first_elem + select(src_offs0, pad_offs, cond_top);
+    __global uchar *src_addr1 = first_elem + src_offs1;
+    __global uchar *src_addr2 = first_elem + select(src_offs2, pad_offs, cond_bottom);
+
+    VEC_INT sum_we = 0;
+    VEC_INT acc0 = 0, acc1 = 0, acc2 = 0, acc3 = 0;
+    VEC_INT sum0 = 0, sum1 = 0, sum2 = 0, sum3 = 0;
+
+    // z == 0
+    VEC_UCHAR w0, w1, w2;
+    w0 = VLOAD(VEC_SIZE)(0, weights.ptr + 0 * weights_stride_y);
+    w1 = VLOAD(VEC_SIZE)(0, weights.ptr + 1 * weights_stride_y);
+    w2 = VLOAD(VEC_SIZE)(0, weights.ptr + 2 * weights_stride_y);
+
+#if INPUT_OFFSET != 0
+    sum_we += CONVERT(w0, VEC_INT) + CONVERT(w1, VEC_INT) + CONVERT(w2, VEC_INT);
+#endif /* INPUT_OFFSET != 0 */
+
+    VEC_UCHAR values = VLOAD(VEC_SIZE)(0, src_addr0);
+    BIFROST_MAD_ACC_4(acc0, sum0, values, w0);
+
+    src_addr0 += src_stride_y;
+    values = VLOAD(VEC_SIZE)(0, src_addr0);
+    BIFROST_MAD_ACC_4(acc0, sum0, values, w1);
+    BIFROST_MAD_ACC_4(acc1, sum1, values, w0);
+
+    src_addr0 += src_stride_y;
+    values = VLOAD(VEC_SIZE)(0, src_addr0);
+    BIFROST_MAD_ACC_4(acc0, sum0, values, w2);
+    BIFROST_MAD_ACC_4(acc1, sum1, values, w1);
+    BIFROST_MAD_ACC_4(acc2, sum2, values, w0);
+
+    src_addr0 += src_stride_y;
+    values = VLOAD(VEC_SIZE)(0, src_addr0);
+    BIFROST_MAD_ACC_4(acc1, sum1, values, w2);
+    BIFROST_MAD_ACC_4(acc2, sum2, values, w1);
+    BIFROST_MAD_ACC_4(acc3, sum3, values, w0);
+
+    src_addr0 += src_stride_y;
+    values = VLOAD(VEC_SIZE)(0, src_addr0);
+    BIFROST_MAD_ACC_4(acc2, sum2, values, w2);
+    BIFROST_MAD_ACC_4(acc3, sum3, values, w1);
+
+    src_addr0 += src_stride_y;
+    values = VLOAD(VEC_SIZE)(0, src_addr0);
+    BIFROST_MAD_ACC_4(acc3, sum3, values, w2);
+
+    weights.ptr += weights_stride_z;
+
+    // z == 1
+    w0 = VLOAD(VEC_SIZE)(0, weights.ptr + 0 * weights_stride_y);
+    w1 = VLOAD(VEC_SIZE)(0, weights.ptr + 1 * weights_stride_y);
+    w2 = VLOAD(VEC_SIZE)(0, weights.ptr + 2 * weights_stride_y);
+
+#if INPUT_OFFSET != 0
+    sum_we += CONVERT(w0, VEC_INT) + CONVERT(w1, VEC_INT) + CONVERT(w2, VEC_INT);
+#endif /* INPUT_OFFSET != 0 */
+
+    values = VLOAD(VEC_SIZE)(0, src_addr1);
+    BIFROST_MAD_ACC_4(acc0, sum0, values, w0);
+
+    src_addr1 += src_stride_y;
+    values = VLOAD(VEC_SIZE)(0, src_addr1);
+    BIFROST_MAD_ACC_4(acc0, sum0, values, w1);
+    BIFROST_MAD_ACC_4(acc1, sum1, values, w0);
+
+    src_addr1 += src_stride_y;
+    values = VLOAD(VEC_SIZE)(0, src_addr1);
+    BIFROST_MAD_ACC_4(acc0, sum0, values, w2);
+    BIFROST_MAD_ACC_4(acc1, sum1, values, w1);
+    BIFROST_MAD_ACC_4(acc2, sum2, values, w0);
+
+    src_addr1 += src_stride_y;
+    values = VLOAD(VEC_SIZE)(0, src_addr1);
+    BIFROST_MAD_ACC_4(acc1, sum1, values, w2);
+    BIFROST_MAD_ACC_4(acc2, sum2, values, w1);
+    BIFROST_MAD_ACC_4(acc3, sum3, values, w0);
+
+    src_addr1 += src_stride_y;
+    values = VLOAD(VEC_SIZE)(0, src_addr1);
+    BIFROST_MAD_ACC_4(acc2, sum2, values, w2);
+    BIFROST_MAD_ACC_4(acc3, sum3, values, w1);
+
+    src_addr1 += src_stride_y;
+    values = VLOAD(VEC_SIZE)(0, src_addr1);
+    BIFROST_MAD_ACC_4(acc3, sum3, values, w2);
+
+    weights.ptr += weights_stride_z;
+
+    // z == 2
+    w0 = VLOAD(VEC_SIZE)(0, weights.ptr + 0 * weights_stride_y);
+    w1 = VLOAD(VEC_SIZE)(0, weights.ptr + 1 * weights_stride_y);
+    w2 = VLOAD(VEC_SIZE)(0, weights.ptr + 2 * weights_stride_y);
+
+#if INPUT_OFFSET != 0
+    sum_we += CONVERT(w0, VEC_INT) + CONVERT(w1, VEC_INT) + CONVERT(w2, VEC_INT);
+#endif /* INPUT_OFFSET != 0 */
+
+    values = VLOAD(VEC_SIZE)(0, src_addr2);
+    BIFROST_MAD_ACC_4(acc0, sum0, values, w0);
+
+    src_addr2 += src_stride_y;
+    values = VLOAD(VEC_SIZE)(0, src_addr2);
+    BIFROST_MAD_ACC_4(acc0, sum0, values, w1);
+    BIFROST_MAD_ACC_4(acc1, sum1, values, w0);
+
+    src_addr2 += src_stride_y;
+    values = VLOAD(VEC_SIZE)(0, src_addr2);
+    BIFROST_MAD_ACC_4(acc0, sum0, values, w2);
+    BIFROST_MAD_ACC_4(acc1, sum1, values, w1);
+    BIFROST_MAD_ACC_4(acc2, sum2, values, w0);
+
+    src_addr2 += src_stride_y;
+    values = VLOAD(VEC_SIZE)(0, src_addr2);
+    BIFROST_MAD_ACC_4(acc1, sum1, values, w2);
+    BIFROST_MAD_ACC_4(acc2, sum2, values, w1);
+    BIFROST_MAD_ACC_4(acc3, sum3, values, w0);
+
+    src_addr2 += src_stride_y;
+    values = VLOAD(VEC_SIZE)(0, src_addr2);
+    BIFROST_MAD_ACC_4(acc2, sum2, values, w2);
+    BIFROST_MAD_ACC_4(acc3, sum3, values, w1);
+
+    src_addr2 += src_stride_y;
+    values = VLOAD(VEC_SIZE)(0, src_addr2);
+    BIFROST_MAD_ACC_4(acc3, sum3, values, w2);
+
+#if defined(HAS_BIAS)
+    acc0 += bias_values;
+    acc1 += bias_values;
+    acc2 += bias_values;
+    acc3 += bias_values;
+#endif /* defined(HAS_BIAS) */
+
+#if WEIGHTS_OFFSET != 0
+    acc0 += WEIGHTS_OFFSET * sum0;
+    acc1 += WEIGHTS_OFFSET * sum1;
+    acc2 += WEIGHTS_OFFSET * sum2;
+    acc3 += WEIGHTS_OFFSET * sum3;
+#endif /* WEIGHTS_OFFSET != 0 */
+
+#if INPUT_OFFSET != 0
+    VEC_INT offs = INPUT_OFFSET * sum_we;
+
+    acc0 += offs;
+    acc1 += offs;
+    acc2 += offs;
+    acc3 += offs;
+#endif /* INPUT_OFFSET != 0 */
+
+#if K_OFFSET != 0
+    acc0 += (VEC_INT)K_OFFSET;
+    acc1 += (VEC_INT)K_OFFSET;
+    acc2 += (VEC_INT)K_OFFSET;
+    acc3 += (VEC_INT)K_OFFSET;
+#endif /* K_OFFSET != 0 */
+
+    acc0 = asymm_mult_by_quant_multiplier_less_than_one(acc0, OUTPUT_MULTIPLIER, OUTPUT_SHIFT);
+    acc1 = asymm_mult_by_quant_multiplier_less_than_one(acc1, OUTPUT_MULTIPLIER, OUTPUT_SHIFT);
+    acc2 = asymm_mult_by_quant_multiplier_less_than_one(acc2, OUTPUT_MULTIPLIER, OUTPUT_SHIFT);
+    acc3 = asymm_mult_by_quant_multiplier_less_than_one(acc3, OUTPUT_MULTIPLIER, OUTPUT_SHIFT);
+
+    acc0 += (VEC_INT)OUTPUT_OFFSET;
+    acc1 += (VEC_INT)OUTPUT_OFFSET;
+    acc2 += (VEC_INT)OUTPUT_OFFSET;
+    acc3 += (VEC_INT)OUTPUT_OFFSET;
+
+    VEC_UCHAR res0 = CONVERT_SAT(acc0, VEC_UCHAR);
+    VEC_UCHAR res1 = CONVERT_SAT(acc1, VEC_UCHAR);
+    VEC_UCHAR res2 = CONVERT_SAT(acc2, VEC_UCHAR);
+    VEC_UCHAR res3 = CONVERT_SAT(acc3, VEC_UCHAR);
+
+    res0 = CLAMP(res0, (VEC_UCHAR)0, (VEC_UCHAR)255);
+    res1 = CLAMP(res1, (VEC_UCHAR)0, (VEC_UCHAR)255);
+    res2 = CLAMP(res2, (VEC_UCHAR)0, (VEC_UCHAR)255);
+    res3 = CLAMP(res3, (VEC_UCHAR)0, (VEC_UCHAR)255);
+
+    VSTORE(VEC_SIZE)
+    (res0, 0, dst.ptr + 0 * dst_stride_y);
+    VSTORE(VEC_SIZE)
+    (res1, 0, dst.ptr + 1 * dst_stride_y);
+    VSTORE(VEC_SIZE)
+    (res2, 0, dst.ptr + 2 * dst_stride_y);
+    VSTORE(VEC_SIZE)
+    (res3, 0, dst.ptr + 3 * dst_stride_y);
+}
+
+/** This function computes the depthwise convolution quantized.
+ *
+ * @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] 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: QASYMM8
+ * @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_3x3_quantized_nhwc_stride2(
+    TENSOR3D_DECLARATION(src),
+    TENSOR3D_DECLARATION(dst),
+    TENSOR3D_DECLARATION(weights),
+#if defined(HAS_BIAS)
+    VECTOR_DECLARATION(biases)
+#endif /* defined(HAS_BIAS) */
+)
+{
+    Image  dst     = CONVERT_TENSOR3D_TO_IMAGE_STRUCT(dst);
+    Vector weights = CONVERT_TO_VECTOR_STRUCT(weights);
+#if defined(HAS_BIAS)
+    Vector biases = CONVERT_TO_VECTOR_STRUCT(biases);
+
+    VEC_INT bias_values = VLOAD(VEC_SIZE)(0, (__global int *)biases.ptr);
+#endif /* defined(HAS_BIAS) */
+
+    __global uchar *first_elem = src_ptr + src_offset_first_element_in_bytes;
+
+    const int z         = get_global_id(2);
+    const int pad_offs  = -ROWS_READ * src_stride_y;
+    const int src_offs0 = get_global_id(0) * src_step_x + get_global_id(1) * src_step_y + z * src_step_z - CONV_PAD_TOP * src_stride_z;
+    const int src_offs1 = src_offs0 + src_stride_z;
+    const int src_offs2 = src_offs1 + src_stride_z;
+
+    const int cond_top    = z - CONV_PAD_TOP < 0;
+    const int cond_bottom = z * (src_step_z / src_stride_z) + 2 >= SRC_DEPTH;
+    ;
+
+    __global uchar *src_addr0 = first_elem + select(src_offs0, pad_offs, cond_top);
+    __global uchar *src_addr1 = first_elem + src_offs1;
+    __global uchar *src_addr2 = first_elem + select(src_offs2, pad_offs, cond_bottom);
+
+    VEC_INT sum_we = 0;
+    VEC_INT acc0 = 0, acc2 = 0;
+    VEC_INT sum0 = 0, sum2 = 0;
+
+    // z == 0
+    VEC_UCHAR w0, w1, w2;
+    w0 = VLOAD(VEC_SIZE)(0, weights.ptr + 0 * weights_stride_y);
+    w1 = VLOAD(VEC_SIZE)(0, weights.ptr + 1 * weights_stride_y);
+    w2 = VLOAD(VEC_SIZE)(0, weights.ptr + 2 * weights_stride_y);
+
+#if INPUT_OFFSET != 0
+    sum_we += CONVERT(w0, VEC_INT) + CONVERT(w1, VEC_INT) + CONVERT(w2, VEC_INT);
+#endif /* INPUT_OFFSET != 0 */
+
+    VEC_UCHAR values = VLOAD(VEC_SIZE)(0, src_addr0);
+    BIFROST_MAD_ACC_4(acc0, sum0, values, w0);
+
+    src_addr0 += src_stride_y;
+    values = VLOAD(VEC_SIZE)(0, src_addr0);
+    BIFROST_MAD_ACC_4(acc0, sum0, values, w1);
+
+    src_addr0 += src_stride_y;
+    values = VLOAD(VEC_SIZE)(0, src_addr0);
+    BIFROST_MAD_ACC_4(acc0, sum0, values, w2);
+    BIFROST_MAD_ACC_4(acc2, sum2, values, w0);
+
+    src_addr0 += src_stride_y;
+    values = VLOAD(VEC_SIZE)(0, src_addr0);
+    BIFROST_MAD_ACC_4(acc2, sum2, values, w1);
+
+    src_addr0 += src_stride_y;
+    values = VLOAD(VEC_SIZE)(0, src_addr0);
+    BIFROST_MAD_ACC_4(acc2, sum2, values, w2);
+
+    weights.ptr += weights_stride_z;
+
+    // z == 1
+    w0 = VLOAD(VEC_SIZE)(0, weights.ptr + 0 * weights_stride_y);
+    w1 = VLOAD(VEC_SIZE)(0, weights.ptr + 1 * weights_stride_y);
+    w2 = VLOAD(VEC_SIZE)(0, weights.ptr + 2 * weights_stride_y);
+
+#if INPUT_OFFSET != 0
+    sum_we += CONVERT(w0, VEC_INT) + CONVERT(w1, VEC_INT) + CONVERT(w2, VEC_INT);
+#endif /* INPUT_OFFSET != 0 */
+
+    values = VLOAD(VEC_SIZE)(0, src_addr1);
+    BIFROST_MAD_ACC_4(acc0, sum0, values, w0);
+
+    src_addr1 += src_stride_y;
+    values = VLOAD(VEC_SIZE)(0, src_addr1);
+    BIFROST_MAD_ACC_4(acc0, sum0, values, w1);
+
+    src_addr1 += src_stride_y;
+    values = VLOAD(VEC_SIZE)(0, src_addr1);
+    BIFROST_MAD_ACC_4(acc0, sum0, values, w2);
+    BIFROST_MAD_ACC_4(acc2, sum2, values, w0);
+
+    src_addr1 += src_stride_y;
+    values = VLOAD(VEC_SIZE)(0, src_addr1);
+    BIFROST_MAD_ACC_4(acc2, sum2, values, w1);
+
+    src_addr1 += src_stride_y;
+    values = VLOAD(VEC_SIZE)(0, src_addr1);
+    BIFROST_MAD_ACC_4(acc2, sum2, values, w2);
+
+    weights.ptr += weights_stride_z;
+
+    // z == 2
+    w0 = VLOAD(VEC_SIZE)(0, weights.ptr + 0 * weights_stride_y);
+    w1 = VLOAD(VEC_SIZE)(0, weights.ptr + 1 * weights_stride_y);
+    w2 = VLOAD(VEC_SIZE)(0, weights.ptr + 2 * weights_stride_y);
+
+#if INPUT_OFFSET != 0
+    sum_we += CONVERT(w0, VEC_INT) + CONVERT(w1, VEC_INT) + CONVERT(w2, VEC_INT);
+#endif /* INPUT_OFFSET != 0 */
+
+    values = VLOAD(VEC_SIZE)(0, src_addr2);
+    BIFROST_MAD_ACC_4(acc0, sum0, values, w0);
+
+    src_addr2 += src_stride_y;
+    values = VLOAD(VEC_SIZE)(0, src_addr2);
+    BIFROST_MAD_ACC_4(acc0, sum0, values, w1);
+
+    src_addr2 += src_stride_y;
+    values = VLOAD(VEC_SIZE)(0, src_addr2);
+    BIFROST_MAD_ACC_4(acc0, sum0, values, w2);
+    BIFROST_MAD_ACC_4(acc2, sum2, values, w0);
+
+    src_addr2 += src_stride_y;
+    values = VLOAD(VEC_SIZE)(0, src_addr2);
+    BIFROST_MAD_ACC_4(acc2, sum2, values, w1);
+
+    src_addr2 += src_stride_y;
+    values = VLOAD(VEC_SIZE)(0, src_addr2);
+    BIFROST_MAD_ACC_4(acc2, sum2, values, w2);
+
+#if defined(HAS_BIAS)
+    acc0 += bias_values;
+    acc2 += bias_values;
+#endif /* defined(HAS_BIAS) */
+
+#if WEIGHTS_OFFSET != 0
+    acc0 += WEIGHTS_OFFSET * sum0;
+    acc2 += WEIGHTS_OFFSET * sum2;
+#endif /* WEIGHTS_OFFSET != 0 */
+
+#if INPUT_OFFSET != 0
+    VEC_INT offs = INPUT_OFFSET * sum_we;
+
+    acc0 += offs;
+    acc2 += offs;
+#endif /* INPUT_OFFSET != 0 */
+
+#if K_OFFSET != 0
+    acc0 += (VEC_INT)K_OFFSET;
+    acc2 += (VEC_INT)K_OFFSET;
+#endif /* K_OFFSET != 0 */
+
+    acc0 = asymm_mult_by_quant_multiplier_less_than_one(acc0, OUTPUT_MULTIPLIER, OUTPUT_SHIFT);
+    acc2 = asymm_mult_by_quant_multiplier_less_than_one(acc2, OUTPUT_MULTIPLIER, OUTPUT_SHIFT);
+    acc0 += (VEC_INT)OUTPUT_OFFSET;
+    acc2 += (VEC_INT)OUTPUT_OFFSET;
+    VEC_UCHAR res0 = CONVERT_SAT(acc0, VEC_UCHAR);
+    VEC_UCHAR res2 = CONVERT_SAT(acc2, VEC_UCHAR);
+    res0           = CLAMP(res0, (VEC_UCHAR)0, (VEC_UCHAR)255);
+    res2           = CLAMP(res2, (VEC_UCHAR)0, (VEC_UCHAR)255);
+
+    VSTORE(VEC_SIZE)
+    (res0, 0, dst.ptr + 0 * dst_stride_y);
+    VSTORE(VEC_SIZE)
+    (res2, 0, dst.ptr + 1 * dst_stride_y);
+}
+
+#endif /* defined(VEC_SIZE) && defined(SRC_DEPTH) && defined(CONV_PAD_TOP) && defined(ROWS_READ) */
+
+#endif /* defined(WEIGHTS_OFFSET) && defined(INPUT_OFFSET) && defined(K_OFFSET) && defined(OUTPUT_OFFSET) && defined(OUTPUT_MULTIPLIER) && defined(OUTPUT_SHIFT) */
diff --git a/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3Kernel.cpp b/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NCHWKernel.cpp
index c9fe1cfc0b..de68ceda11 100644
--- a/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3Kernel.cpp
+++ b/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NCHWKernel.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017-2018 ARM Limited.
+ * Copyright (c) 2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -21,7 +21,7 @@
  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  * SOFTWARE.
  */
-#include "arm_compute/core/CL/kernels/CLDepthwiseConvolutionLayer3x3Kernel.h"
+#include "arm_compute/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NCHWKernel.h"
 
 #include "arm_compute/core/AccessWindowStatic.h"
 #include "arm_compute/core/CL/CLHelpers.h"
@@ -39,18 +39,18 @@
 using namespace arm_compute;
 using namespace arm_compute::misc::shape_calculator;
 
-CLDepthwiseConvolutionLayer3x3Kernel::CLDepthwiseConvolutionLayer3x3Kernel()
-    : _border_size(0), _input(), _output(), _weights(), _biases(), _conv_stride_x(0), _conv_stride_y(0), _conv_pad_left(0), _conv_pad_top(0)
+CLDepthwiseConvolutionLayer3x3NCHWKernel::CLDepthwiseConvolutionLayer3x3NCHWKernel()
+    : _conv_stride_x(0), _conv_pad_top(0)
 {
 }
 
-BorderSize CLDepthwiseConvolutionLayer3x3Kernel::border_size() const
+BorderSize CLDepthwiseConvolutionLayer3x3NCHWKernel::border_size() const
 {
     return _border_size;
 }
 
-void CLDepthwiseConvolutionLayer3x3Kernel::configure(const ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const PadStrideInfo &conv_info,
-                                                     ActivationLayerInfo act_info)
+void CLDepthwiseConvolutionLayer3x3NCHWKernel::configure(const ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const PadStrideInfo &conv_info,
+                                                         ActivationLayerInfo act_info)
 {
     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);
@@ -144,8 +144,6 @@ void CLDepthwiseConvolutionLayer3x3Kernel::configure(const ICLTensor *input, con
         }
     }
 
-    // Configure the local work size for Bifrost with a value obtained
-    // via exhaustive autotuning for the MobileNets tensor shapes.
     const GPUTarget gpu_target = get_target();
     const bool      is_bifrost = gpu_target_is_in(gpu_target, GPUTarget::G71, GPUTarget::G72, GPUTarget::G51, GPUTarget::G51BIG, GPUTarget::G51LIT, GPUTarget::TNOX);
 
@@ -228,7 +226,7 @@ void CLDepthwiseConvolutionLayer3x3Kernel::configure(const ICLTensor *input, con
     }
     else
     {
-        kernel_name                       = is_qasymm ? "depthwise_convolution_3x3_quantized" : "depthwise_convolution_3x3";
+        kernel_name                       = is_qasymm ? "depthwise_convolution_3x3_quantized_nchw" : "depthwise_convolution_3x3";
         num_elems_written_per_iteration_x = 8 / data_size_from_type(input->info()->data_type());
         num_elems_written_per_iteration_y = (is_qasymm && _conv_stride_y < 3) ? (2 / _conv_stride_y) : 1;
         num_elems_read_per_iteration_x    = 3 + (num_elems_written_per_iteration_x - 1) * _conv_stride_x;
@@ -268,7 +266,7 @@ void CLDepthwiseConvolutionLayer3x3Kernel::configure(const ICLTensor *input, con
     _config_id += support::cpp11::to_string(output->info()->dimension(1));
 }
 
-void CLDepthwiseConvolutionLayer3x3Kernel::run(const Window &window, cl::CommandQueue &queue)
+void CLDepthwiseConvolutionLayer3x3NCHWKernel::run(const Window &window, cl::CommandQueue &queue)
 {
     ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
     ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(IKernel::window(), window);
diff --git a/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NHWCKernel.cpp b/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NHWCKernel.cpp
new file mode 100644
index 0000000000..d783b9e159
--- /dev/null
+++ b/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NHWCKernel.cpp
@@ -0,0 +1,254 @@
+/*
+ * Copyright (c) 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.
+ */
+#include "arm_compute/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NHWCKernel.h"
+
+#include "arm_compute/core/AccessWindowStatic.h"
+#include "arm_compute/core/CL/CLHelpers.h"
+#include "arm_compute/core/CL/CLKernelLibrary.h"
+#include "arm_compute/core/CL/ICLKernel.h"
+#include "arm_compute/core/CL/ICLTensor.h"
+#include "arm_compute/core/Error.h"
+#include "arm_compute/core/Helpers.h"
+#include "arm_compute/core/TensorInfo.h"
+#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"
+
+using namespace arm_compute;
+using namespace arm_compute::misc::shape_calculator;
+
+namespace
+{
+Status validate_arguments(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info,
+                          const ActivationLayerInfo &act_info)
+{
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8);
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG((act_info.enabled()) && (act_info.activation() != ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU)
+                                    && (act_info.activation() != ActivationLayerInfo::ActivationFunction::BOUNDED_RELU)
+                                    && (act_info.activation() != ActivationLayerInfo::ActivationFunction::RELU),
+                                    "For QASYMM8 only relu, lower bounded relu and lower-upper bounded relu are supported");
+    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights);
+    ARM_COMPUTE_RETURN_ERROR_ON(weights->dimension(1) != 3 || weights->dimension(2) != 3);
+
+    if(biases != nullptr)
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(biases, 1, DataType::S32);
+        ARM_COMPUTE_RETURN_ERROR_ON(biases->dimension(0) != weights->dimension(0));
+        ARM_COMPUTE_RETURN_ERROR_ON(biases->num_dimensions() > 1);
+    }
+
+    if(output->total_size() != 0)
+    {
+        const TensorShape output_shape = compute_depthwise_convolution_shape(*input, *weights, conv_info);
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(output->tensor_shape(), output_shape);
+    }
+
+    return Status{};
+}
+
+std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *weights, ITensorInfo *output, const PadStrideInfo &conv_info)
+{
+    const unsigned int num_rows_processed_per_iteration = 4;
+    const unsigned int num_elems_accessed_per_iteration = 4;
+    const unsigned int num_rows_read_per_iteration      = num_rows_processed_per_iteration + 2;
+    const unsigned int num_rows_written_per_iteration   = num_rows_processed_per_iteration / conv_info.stride().first;
+
+    const BorderSize border_size(conv_info.pad_left() + num_rows_read_per_iteration * std::max(conv_info.pad_top(), conv_info.pad_bottom()), 0, conv_info.pad_right(), 0);
+
+    // Configure kernel window
+    Window win = calculate_max_window(*output, Steps(num_elems_accessed_per_iteration, num_rows_written_per_iteration));
+
+    AccessWindowStatic input_access(input, 0, -border_size.top, ceil_to_multiple(input->dimension(0), num_elems_accessed_per_iteration),
+                                    ceil_to_multiple(input->dimension(1) + border_size.bottom, num_rows_read_per_iteration));
+    AccessWindowRectangle  output_access(output, 0, 0, num_elems_accessed_per_iteration, num_rows_written_per_iteration);
+    AccessWindowHorizontal weights_access(weights, 0, num_elems_accessed_per_iteration);
+
+    bool window_changed = update_window_and_padding(win, input_access, weights_access, output_access);
+
+    output_access.set_valid_region(win, ValidRegion(Coordinates(), output->tensor_shape()));
+
+    Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{};
+    return std::make_pair(err, win);
+}
+} // namespace
+
+CLDepthwiseConvolutionLayer3x3NHWCKernel::CLDepthwiseConvolutionLayer3x3NHWCKernel()
+    : _num_rows_processed_per_iteration(1)
+{
+}
+
+BorderSize CLDepthwiseConvolutionLayer3x3NHWCKernel::border_size() const
+{
+    return _border_size;
+}
+
+void CLDepthwiseConvolutionLayer3x3NHWCKernel::configure(const ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const PadStrideInfo &conv_info,
+                                                         ActivationLayerInfo act_info)
+{
+    ARM_COMPUTE_ERROR_ON_NULLPTR(input, weights, output);
+
+    // Get convolved dimensions
+    const TensorShape output_shape = compute_depthwise_convolution_shape(*input->info(), *weights->info(), conv_info);
+
+    // Output auto inizialitation if not yet initialized
+    auto_init_if_empty(*output->info(),
+                       output_shape,
+                       1,
+                       input->info()->data_type(),
+                       input->info()->fixed_point_position(),
+                       input->info()->quantization_info());
+
+    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), weights->info(), (biases != nullptr) ? biases->info() : nullptr, output->info(), conv_info, act_info));
+
+    const unsigned int conv_stride_x = conv_info.stride().first;
+    ARM_COMPUTE_ERROR_ON(conv_stride_x < 1 || conv_stride_x > 2);
+    ARM_COMPUTE_ERROR_ON(std::max(conv_info.pad_top(), conv_info.pad_bottom()) > 1);
+
+    _input                            = input;
+    _output                           = output;
+    _weights                          = weights;
+    _biases                           = biases;
+    _conv_stride_y                    = conv_info.stride().second;
+    _conv_pad_left                    = conv_info.pad_left();
+    _num_rows_processed_per_iteration = 4;
+
+    const unsigned int num_elems_accessed_per_iteration = 4;
+    const unsigned int num_rows_read_per_iteration      = _num_rows_processed_per_iteration + 2;
+
+    _border_size = BorderSize(_conv_pad_left + num_rows_read_per_iteration * std::max(conv_info.pad_top(), conv_info.pad_bottom()), 0, conv_info.pad_right(), 0);
+
+    float multiplier        = _input->info()->quantization_info().scale * _weights->info()->quantization_info().scale / _output->info()->quantization_info().scale;
+    int   output_multiplier = 0;
+    int   output_shift      = 0;
+    quantization::calculate_quantized_multiplier_less_than_one(multiplier, &output_multiplier, &output_shift);
+
+    CLBuildOptions build_opts;
+    build_opts.add_option_if(_biases != nullptr, "-DHAS_BIAS");
+    build_opts.add_option("-DINPUT_OFFSET=" + support::cpp11::to_string(-_input->info()->quantization_info().offset));
+    build_opts.add_option("-DWEIGHTS_OFFSET=" + support::cpp11::to_string(-_weights->info()->quantization_info().offset));
+    build_opts.add_option("-DOUTPUT_OFFSET=" + support::cpp11::to_string(_output->info()->quantization_info().offset));
+    build_opts.add_option("-DK_OFFSET=" + support::cpp11::to_string(9 * input->info()->quantization_info().offset * weights->info()->quantization_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));
+    build_opts.add_option("-DVEC_SIZE=" + support::cpp11::to_string(num_elems_accessed_per_iteration));
+    build_opts.add_option("-DSRC_DEPTH=" + support::cpp11::to_string(_input->info()->dimension(2)));
+    build_opts.add_option("-DCONV_PAD_TOP=" + support::cpp11::to_string(conv_info.pad_top()));
+    build_opts.add_option("-DROWS_READ=" + support::cpp11::to_string(num_rows_read_per_iteration));
+
+    if(act_info.enabled())
+    {
+        const int a_val = input->info()->quantization_info().quantize(act_info.a(), RoundingPolicy::TO_NEAREST_UP);
+        const int b_val = input->info()->quantization_info().quantize(act_info.b(), RoundingPolicy::TO_NEAREST_UP);
+        const int o1    = input->info()->quantization_info().offset;
+
+        build_opts.add_option("-DFUSED_ACTIVATION=" + lower_string(string_from_activation_func(act_info.activation())));
+        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));
+
+        if(output != nullptr)
+        {
+            const float s1 = input->info()->quantization_info().scale;
+            const float s2 = output->info()->quantization_info().scale;
+            const int   o2 = output->info()->quantization_info().offset;
+
+            if(o1 != o2 || s1 != s2)
+            {
+                build_opts.add_option("-DS1_VAL=" + float_to_string_with_full_precision(s1));
+                build_opts.add_option("-DS2_VAL=" + float_to_string_with_full_precision(s2));
+                build_opts.add_option("-DO1_VAL=" + support::cpp11::to_string(o1));
+                build_opts.add_option("-DO2_VAL=" + support::cpp11::to_string(o2));
+            }
+        }
+    }
+
+    // Create kernel
+    std::string kernel_name = std::string("depthwise_convolution_3x3_quantized_nhwc_stride") + support::cpp11::to_string(conv_stride_x);
+    _kernel                 = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel(kernel_name, build_opts.options()));
+
+    // Configure kernel window
+    auto win_config = validate_and_configure_window(input->info(), weights->info(), output->info(), conv_info);
+    ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
+    ICLKernel::configure(win_config.second);
+
+    // Set config_id for enabling LWS tuning
+    _config_id = kernel_name;
+    _config_id += "_";
+    _config_id += support::cpp11::to_string(input->info()->dimension(0));
+    _config_id += "_";
+    _config_id += support::cpp11::to_string(input->info()->dimension(1));
+    _config_id += "_";
+    _config_id += support::cpp11::to_string(input->info()->dimension(2));
+    _config_id += "_";
+    _config_id += support::cpp11::to_string(output->info()->dimension(0));
+    _config_id += "_";
+    _config_id += support::cpp11::to_string(output->info()->dimension(1));
+}
+
+Status CLDepthwiseConvolutionLayer3x3NHWCKernel::validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info,
+                                                          ActivationLayerInfo act_info)
+{
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, weights, biases, output, conv_info, act_info));
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), weights->clone().get(), output->clone().get(), conv_info).first);
+
+    return Status{};
+}
+
+void CLDepthwiseConvolutionLayer3x3NHWCKernel::run(const Window &window, cl::CommandQueue &queue)
+{
+    ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
+    ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(IKernel::window(), window);
+
+    // Create input window and adjust
+    Window win_in = window;
+    win_in.adjust(Window::DimY, -_conv_pad_left, true);
+    win_in.set_dimension_step(Window::DimY, _num_rows_processed_per_iteration);
+    win_in.set_dimension_step(Window::DimZ, _conv_stride_y);
+
+    ARM_COMPUTE_ERROR_ON((win_in.y().step() < window.y().step()) || (win_in.z().step() < window.z().step()));
+
+    Window slice_in  = win_in.first_slice_window_3D();
+    Window slice_out = window.first_slice_window_3D();
+
+    if(_biases != nullptr)
+    {
+        unsigned int idx = 3 * num_arguments_per_3D_tensor();
+        Window       win_biases;
+        win_biases.use_tensor_dimensions(_biases->info()->tensor_shape());
+        win_biases.set_dimension_step(Window::DimX, window.x().step());
+        add_1D_tensor_argument(idx, _biases, win_biases);
+    }
+
+    do
+    {
+        unsigned int idx = 0;
+        add_3D_tensor_argument(idx, _input, slice_in);
+        add_3D_tensor_argument(idx, _output, slice_out);
+        add_3D_tensor_argument(idx, _weights, slice_out);
+
+        enqueue(queue, *this, slice_out, _lws_hint);
+    }
+    while(window.slide_window_slice_3D(slice_out) && win_in.slide_window_slice_3D(slice_in));
+}
diff --git a/src/runtime/CL/functions/CLDepthwiseConvolutionLayer.cpp b/src/runtime/CL/functions/CLDepthwiseConvolutionLayer.cpp
index 8d7c92bdf1..0276b37e09 100644
--- a/src/runtime/CL/functions/CLDepthwiseConvolutionLayer.cpp
+++ b/src/runtime/CL/functions/CLDepthwiseConvolutionLayer.cpp
@@ -35,7 +35,7 @@ using namespace arm_compute::misc;
 using namespace arm_compute::misc::shape_calculator;
 
 CLDepthwiseConvolutionLayer3x3::CLDepthwiseConvolutionLayer3x3()
-    : _kernel(), _border_handler()
+    : _kernel(nullptr), _border_handler()
 {
 }
 
@@ -44,8 +44,17 @@ void CLDepthwiseConvolutionLayer3x3::configure(ICLTensor *input, const ICLTensor
     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);
 
-    _kernel.set_target(CLScheduler::get().target());
-    _kernel.configure(input, weights, biases, output, conv_info, act_info);
+    if(input->info()->data_layout() == DataLayout::NCHW)
+    {
+        _kernel = arm_compute::support::cpp14::make_unique<CLDepthwiseConvolutionLayer3x3NCHWKernel>();
+    }
+    else
+    {
+        _kernel = arm_compute::support::cpp14::make_unique<CLDepthwiseConvolutionLayer3x3NHWCKernel>();
+    }
+
+    _kernel->set_target(CLScheduler::get().target());
+    _kernel->configure(input, weights, biases, output, conv_info, act_info);
 
     // Configure border handler
     PixelValue &&zero_value(0.f);
@@ -53,13 +62,13 @@ void CLDepthwiseConvolutionLayer3x3::configure(ICLTensor *input, const ICLTensor
     {
         zero_value = PixelValue(static_cast<uint8_t>(input->info()->quantization_info().offset));
     }
-    _border_handler.configure(input, _kernel.border_size(), BorderMode::CONSTANT, zero_value);
+    _border_handler.configure(input, _kernel->border_size(), BorderMode::CONSTANT, zero_value);
 }
 
 void CLDepthwiseConvolutionLayer3x3::run()
 {
     CLScheduler::get().enqueue(_border_handler);
-    CLScheduler::get().enqueue(_kernel);
+    CLScheduler::get().enqueue(*_kernel);
 }
 
 CLDepthwiseConvolutionLayer::CLDepthwiseConvolutionLayer()
diff --git a/tests/datasets/DepthwiseConvolutionLayerDataset.h b/tests/datasets/DepthwiseConvolutionLayerDataset.h
index 0413cd0c94..1e77a0c8dd 100644
--- a/tests/datasets/DepthwiseConvolutionLayerDataset.h
+++ b/tests/datasets/DepthwiseConvolutionLayerDataset.h
@@ -160,11 +160,20 @@ public:
     {
         add_config(TensorShape(3U, 3U, 2U), TensorShape(3U, 3U, 2U), TensorShape(1U, 1U, 2U), PadStrideInfo(1, 1, 0, 0));
         add_config(TensorShape(7U, 7U, 3U, 2U), TensorShape(3U, 3U, 3U), TensorShape(5U, 5U, 3U, 2U), PadStrideInfo(1, 1, 0, 0));
-        add_config(TensorShape(33U, 27U, 11U), TensorShape(3U, 3U, 11U), TensorShape(11U, 14U, 11U), PadStrideInfo(3, 2, 1, 1));
         add_config(TensorShape(21U, 31U, 9U, 4U), TensorShape(3U, 3U, 9U), TensorShape(21U, 15U, 9U, 4U), PadStrideInfo(1, 2, 1, 0));
         add_config(TensorShape(33U, 27U, 11U, 3U), TensorShape(3U, 3U, 11U), TensorShape(31U, 14U, 11U, 3U), PadStrideInfo(1, 2, 0, 1));
         // Asymmetric padding
         add_config(TensorShape(33U, 27U, 11U), TensorShape(3U, 3U, 11U), TensorShape(16U, 13U, 11U), PadStrideInfo(2, 2, 0, 1, 0, 1, DimensionRoundingType::FLOOR));
+    }
+};
+
+class SmallDepthwiseConvolutionLayerDataset3x3NCHW final : public DepthwiseConvolutionLayerDataset
+{
+public:
+    SmallDepthwiseConvolutionLayerDataset3x3NCHW()
+    {
+        add_config(TensorShape(33U, 27U, 11U), TensorShape(3U, 3U, 11U), TensorShape(11U, 14U, 11U), PadStrideInfo(3, 2, 1, 1));
+        // Asymmetric padding
         add_config(TensorShape(33U, 27U, 11U), TensorShape(3U, 3U, 11U), TensorShape(18U, 14U, 11U), PadStrideInfo(2, 2, 3, 1, 2, 1, DimensionRoundingType::FLOOR));
     }
 };
@@ -176,7 +185,6 @@ public:
     LargeDepthwiseConvolutionLayerDataset3x3()
     {
         add_config(TensorShape(233U, 277U, 55U, 3U), TensorShape(3U, 3U, 55U), TensorShape(116U, 275U, 55U, 3U), PadStrideInfo(2, 1, 0, 0));
-        add_config(TensorShape(333U, 277U, 77U), TensorShape(3U, 3U, 77U), TensorShape(111U, 138U, 77U), PadStrideInfo(3, 2, 1, 0));
         add_config(TensorShape(177U, 311U, 22U), TensorShape(3U, 3U, 22U), TensorShape(177U, 156U, 22U), PadStrideInfo(1, 2, 1, 1));
         add_config(TensorShape(233U, 277U, 55U), TensorShape(3U, 3U, 55U), TensorShape(231U, 138U, 55U), PadStrideInfo(1, 2, 0, 0));
         add_config(TensorShape(333U, 277U, 77U, 5U), TensorShape(3U, 3U, 77U), TensorShape(166U, 93U, 77U, 5U), PadStrideInfo(2, 3, 0, 1));
diff --git a/tests/validation/CL/DepthwiseConvolutionLayer.cpp b/tests/validation/CL/DepthwiseConvolutionLayer.cpp
index 1779ff1aee..450bb21e77 100644
--- a/tests/validation/CL/DepthwiseConvolutionLayer.cpp
+++ b/tests/validation/CL/DepthwiseConvolutionLayer.cpp
@@ -76,10 +76,12 @@ using CLDepthwiseConvolutionLayerFixture3x3 = DepthwiseConvolutionLayerValidatio
 TEST_SUITE(Float)
 TEST_SUITE(F16)
 TEST_SUITE(W3x3)
-FIXTURE_DATA_TEST_CASE(RunSmall, CLDepthwiseConvolutionLayerFixture3x3<half>, framework::DatasetMode::ALL, combine(combine(datasets::SmallDepthwiseConvolutionLayerDataset3x3(),
-                                                                                                                   framework::dataset::make("DataType",
-                                                                                                                           DataType::F16)),
-                                                                                                                   framework::dataset::make("DataLayout", DataLayout::NCHW)))
+FIXTURE_DATA_TEST_CASE(RunSmall, CLDepthwiseConvolutionLayerFixture3x3<half>, framework::DatasetMode::ALL,
+                       combine(combine(framework::dataset::concat(datasets::SmallDepthwiseConvolutionLayerDataset3x3(),
+                                                                  datasets::SmallDepthwiseConvolutionLayerDataset3x3NCHW()),
+                                       framework::dataset::make("DataType",
+                                                                DataType::F16)),
+                               framework::dataset::make("DataLayout", DataLayout::NCHW)))
 {
     validate(CLAccessor(_target), _reference, tolerance_f16);
 }
@@ -95,10 +97,12 @@ TEST_SUITE_END()
 
 TEST_SUITE(FP32)
 TEST_SUITE(W3x3)
-FIXTURE_DATA_TEST_CASE(RunSmall, CLDepthwiseConvolutionLayerFixture3x3<float>, framework::DatasetMode::ALL, combine(combine(datasets::SmallDepthwiseConvolutionLayerDataset3x3(),
-                                                                                                                    framework::dataset::make("DataType",
-                                                                                                                            DataType::F32)),
-                                                                                                                    framework::dataset::make("DataLayout", DataLayout::NCHW)))
+FIXTURE_DATA_TEST_CASE(RunSmall, CLDepthwiseConvolutionLayerFixture3x3<float>, framework::DatasetMode::ALL,
+                       combine(combine(framework::dataset::concat(datasets::SmallDepthwiseConvolutionLayerDataset3x3(),
+                                                                  datasets::SmallDepthwiseConvolutionLayerDataset3x3NCHW()),
+                                       framework::dataset::make("DataType",
+                                                                DataType::F32)),
+                               framework::dataset::make("DataLayout", DataLayout::NCHW)))
 {
     validate(CLAccessor(_target), _reference, tolerance_f32);
 }
@@ -121,29 +125,34 @@ using CLDepthwiseConvolutionLayerQuantizedFixture3x3 = DepthwiseConvolutionLayer
 TEST_SUITE(Quantized)
 TEST_SUITE(QASYMM8)
 TEST_SUITE(Generic)
-FIXTURE_DATA_TEST_CASE(RunSmall, CLDepthwiseConvolutionLayerQuantizedFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallDepthwiseConvolutionLayerDataset(),
+FIXTURE_DATA_TEST_CASE(RunSmall, CLDepthwiseConvolutionLayerQuantizedFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallDepthwiseConvolutionLayerDataset(),
                        framework::dataset::make("DataType", DataType::QASYMM8)),
-                       framework::dataset::make("QuantizationInfo", { QuantizationInfo(0.5f, 10) })))
+                       framework::dataset::make("QuantizationInfo", { QuantizationInfo(0.5f, 10) })),
+                       framework::dataset::make("DataLayout", DataLayout::NCHW)))
 {
     validate(CLAccessor(_target), _reference, tolerance_qasymm8);
 }
-FIXTURE_DATA_TEST_CASE(RunLarge, CLDepthwiseConvolutionLayerQuantizedFixture<uint8_t>, framework::DatasetMode::NIGHTLY, combine(combine(datasets::LargeDepthwiseConvolutionLayerDataset(),
+FIXTURE_DATA_TEST_CASE(RunLarge, CLDepthwiseConvolutionLayerQuantizedFixture<uint8_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::LargeDepthwiseConvolutionLayerDataset(),
                        framework::dataset::make("DataType", DataType::QASYMM8)),
-                       framework::dataset::make("QuantizationInfo", { QuantizationInfo(0.5f, 10) })))
+                       framework::dataset::make("QuantizationInfo", { QuantizationInfo(0.5f, 10) })),
+                       framework::dataset::make("DataLayout", DataLayout::NCHW)))
 {
     validate(CLAccessor(_target), _reference, tolerance_qasymm8);
 }
 TEST_SUITE_END()
 TEST_SUITE(W3x3)
-FIXTURE_DATA_TEST_CASE(RunSmall, CLDepthwiseConvolutionLayerQuantizedFixture3x3<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallDepthwiseConvolutionLayerDataset3x3(),
-                       framework::dataset::make("DataType", DataType::QASYMM8)),
-                       framework::dataset::make("QuantizationInfo", { QuantizationInfo(0.5f, 10) })))
+FIXTURE_DATA_TEST_CASE(RunSmall, CLDepthwiseConvolutionLayerQuantizedFixture3x3<uint8_t>, framework::DatasetMode::PRECOMMIT,
+                       combine(combine(combine(datasets::SmallDepthwiseConvolutionLayerDataset3x3(),
+                                               framework::dataset::make("DataType", DataType::QASYMM8)),
+                                       framework::dataset::make("QuantizationInfo", { QuantizationInfo(0.5f, 10) })),
+                               framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })))
 {
     validate(CLAccessor(_target), _reference, tolerance_qasymm8);
 }
-FIXTURE_DATA_TEST_CASE(RunLarge, CLDepthwiseConvolutionLayerQuantizedFixture3x3<uint8_t>, framework::DatasetMode::NIGHTLY, combine(combine(datasets::LargeDepthwiseConvolutionLayerDataset3x3(),
+FIXTURE_DATA_TEST_CASE(RunLarge, CLDepthwiseConvolutionLayerQuantizedFixture3x3<uint8_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::LargeDepthwiseConvolutionLayerDataset3x3(),
                        framework::dataset::make("DataType", DataType::QASYMM8)),
-                       framework::dataset::make("QuantizationInfo", { QuantizationInfo(0.5f, 10) })))
+                       framework::dataset::make("QuantizationInfo", { QuantizationInfo(0.5f, 10) })),
+                       framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })))
 {
     validate(CLAccessor(_target), _reference, tolerance_qasymm8);
 }
diff --git a/tests/validation/NEON/DepthwiseConvolutionLayer.cpp b/tests/validation/NEON/DepthwiseConvolutionLayer.cpp
index 49e146c084..236d4bd653 100644
--- a/tests/validation/NEON/DepthwiseConvolutionLayer.cpp
+++ b/tests/validation/NEON/DepthwiseConvolutionLayer.cpp
@@ -148,23 +148,27 @@ using NEDepthwiseConvolutionLayerQuantizedFixture = DepthwiseConvolutionLayerVal
 TEST_SUITE(Quantized)
 TEST_SUITE(QASYMM8)
 TEST_SUITE(Generic)
-FIXTURE_DATA_TEST_CASE(RunSmall, NEDepthwiseConvolutionLayerQuantizedFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallDepthwiseConvolutionLayerDataset(),
+FIXTURE_DATA_TEST_CASE(RunSmall, NEDepthwiseConvolutionLayerQuantizedFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallDepthwiseConvolutionLayerDataset(),
                        framework::dataset::make("DataType", DataType::QASYMM8)),
-                       framework::dataset::make("QuantizationInfo", { QuantizationInfo(0.5f, 10) })))
+                       framework::dataset::make("QuantizationInfo", { QuantizationInfo(0.5f, 10) })),
+                       framework::dataset::make("DataLayout", DataLayout::NCHW)))
 {
     validate(Accessor(_target), _reference, tolerance_qasymm8);
 }
 TEST_SUITE_END()
 TEST_SUITE(W3x3)
-FIXTURE_DATA_TEST_CASE(RunSmall, NEDepthwiseConvolutionLayerQuantizedFixture3x3<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallDepthwiseConvolutionLayerDataset3x3(),
-                       framework::dataset::make("DataType", DataType::QASYMM8)),
-                       framework::dataset::make("QuantizationInfo", { QuantizationInfo(0.5f, 10) })))
+FIXTURE_DATA_TEST_CASE(RunSmall, NEDepthwiseConvolutionLayerQuantizedFixture3x3<uint8_t>, framework::DatasetMode::PRECOMMIT,
+                       combine(combine(combine(datasets::SmallDepthwiseConvolutionLayerDataset3x3(),
+                                               framework::dataset::make("DataType", DataType::QASYMM8)),
+                                       framework::dataset::make("QuantizationInfo", { QuantizationInfo(0.5f, 10) })),
+                               framework::dataset::make("DataLayout", DataLayout::NCHW)))
 {
     validate(Accessor(_target), _reference, tolerance_qasymm8);
 }
-FIXTURE_DATA_TEST_CASE(RunLarge, NEDepthwiseConvolutionLayerQuantizedFixture3x3<uint8_t>, framework::DatasetMode::NIGHTLY, combine(combine(datasets::LargeDepthwiseConvolutionLayerDataset3x3(),
+FIXTURE_DATA_TEST_CASE(RunLarge, NEDepthwiseConvolutionLayerQuantizedFixture3x3<uint8_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::LargeDepthwiseConvolutionLayerDataset3x3(),
                        framework::dataset::make("DataType", DataType::QASYMM8)),
-                       framework::dataset::make("QuantizationInfo", { QuantizationInfo(0.5f, 10) })))
+                       framework::dataset::make("QuantizationInfo", { QuantizationInfo(0.5f, 10) })),
+                       framework::dataset::make("DataLayout", DataLayout::NCHW)))
 {
     validate(Accessor(_target), _reference, tolerance_qasymm8);
 }
diff --git a/tests/validation/fixtures/DepthwiseConvolutionLayerFixture.h b/tests/validation/fixtures/DepthwiseConvolutionLayerFixture.h
index ccdd443999..bb756f806e 100644
--- a/tests/validation/fixtures/DepthwiseConvolutionLayerFixture.h
+++ b/tests/validation/fixtures/DepthwiseConvolutionLayerFixture.h
@@ -177,10 +177,10 @@ class DepthwiseConvolutionLayerValidationQuantizedFixture : public DepthwiseConv
 {
 public:
     template <typename...>
-    void setup(TensorShape in_shape, TensorShape weights_shape, TensorShape out_shape, PadStrideInfo pad_stride_info, DataType data_type, QuantizationInfo quantization_info)
+    void setup(TensorShape in_shape, TensorShape weights_shape, TensorShape out_shape, PadStrideInfo pad_stride_info, DataType data_type, QuantizationInfo quantization_info, DataLayout data_layout)
     {
         DepthwiseConvolutionLayerValidationGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(in_shape, weights_shape, out_shape, pad_stride_info,
-                                                                                                            data_type, quantization_info, DataLayout::NCHW);
+                                                                                                            data_type, quantization_info, data_layout);
     }
 };
 } // namespace validation
diff --git a/tests/validation/reference/DepthwiseConvolutionLayer.cpp b/tests/validation/reference/DepthwiseConvolutionLayer.cpp
index ab61b7dd65..d05da9140b 100644
--- a/tests/validation/reference/DepthwiseConvolutionLayer.cpp
+++ b/tests/validation/reference/DepthwiseConvolutionLayer.cpp
@@ -108,13 +108,9 @@ void depthwise_convolution_nchw(const SimpleTensor<T> &src, const SimpleTensor<T
     }
 }
 
-template <>
-SimpleTensor<uint8_t> depthwise_convolution(const SimpleTensor<uint8_t> &src, const SimpleTensor<uint8_t> &weights, const SimpleTensor<int32_t> &biases, const TensorShape &dst_shape,
-                                            const PadStrideInfo &conv_info)
+void depthwise_convolution_nchw(const SimpleTensor<uint8_t> &src, const SimpleTensor<uint8_t> &weights, const SimpleTensor<int32_t> &biases, SimpleTensor<uint8_t> &dst, const PadStrideInfo &conv_info)
 {
     // Create reference
-    SimpleTensor<uint8_t> dst{ dst_shape, src.data_type(), 1, src.fixed_point_position(), src.quantization_info() };
-
     const int   input_offset   = -src.quantization_info().offset;
     const float input_scale    = src.quantization_info().scale;
     const int   weights_offset = -weights.quantization_info().offset;
@@ -169,8 +165,8 @@ SimpleTensor<uint8_t> depthwise_convolution(const SimpleTensor<uint8_t> &src, co
                         {
                             coords.set(0, i);
                             coords.set(1, j);
-                            auto    in_val = tensor_elem_at<uint8_t>(src, coords, BorderMode::CONSTANT, -input_offset);
-                            uint8_t w_val  = *(weights.data() + filter_offset);
+                            const auto    in_val = tensor_elem_at<uint8_t>(src, coords, BorderMode::CONSTANT, -input_offset);
+                            const uint8_t w_val  = *(weights.data() + filter_offset);
                             val += (in_val + input_offset) * (w_val + weights_offset);
                             ++filter_offset;
                         }
@@ -187,6 +183,26 @@ SimpleTensor<uint8_t> depthwise_convolution(const SimpleTensor<uint8_t> &src, co
             }
         }
     }
+}
+
+template <>
+SimpleTensor<uint8_t> depthwise_convolution(const SimpleTensor<uint8_t> &src, const SimpleTensor<uint8_t> &weights, const SimpleTensor<int32_t> &biases, const TensorShape &dst_shape,
+                                            const PadStrideInfo &conv_info)
+{
+    SimpleTensor<uint8_t> dst{ dst_shape, src.data_type(), 1, src.fixed_point_position(), src.quantization_info() };
+
+    if(src.data_layout() == DataLayout::NHWC)
+    {
+        SimpleTensor<uint8_t> src_nchw     = reference::permute<uint8_t>(src, PermutationVector(1U, 2U, 0U));
+        SimpleTensor<uint8_t> weights_nchw = reference::permute<uint8_t>(weights, PermutationVector(1U, 2U, 0U));
+        SimpleTensor<uint8_t> dst_nchw     = reference::permute<uint8_t>(dst, PermutationVector(1U, 2U, 0U));
+
+        depthwise_convolution_nchw(src_nchw, weights_nchw, biases, dst_nchw, conv_info);
+
+        return reference::permute<uint8_t>(dst_nchw, PermutationVector(2U, 0U, 1U));
+    }
+
+    depthwise_convolution_nchw(src, weights, biases, dst, conv_info);
 
     return dst;
 }