COMPMID-464 Implement Depthwise separable convolution on NEON

Change-Id: Iccd686be18381e96bcf09b14c7017c6dda0f38d8
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/109824
Tested-by: BSG Visual Compute Jenkins server to access repositories on http://mpd-gerrit.cambridge.arm.com <bsgcomp@arm.com>
Reviewed-by: Pablo Tello <pablo.tello@arm.com>

14 files changed, 1051 insertions, 3 deletions

diff --git a/arm_compute/core/NEON/NEKernels.h b/arm_compute/core/NEON/NEKernels.h
index ece3ad2c3a..3ad1931ed1 100644
--- a/arm_compute/core/NEON/NEKernels.h
+++ b/arm_compute/core/NEON/NEKernels.h
@@ -47,6 +47,9 @@
 #include "arm_compute/core/NEON/kernels/NEDepthConcatenateKernel.h"
 #include "arm_compute/core/NEON/kernels/NEDepthConvertKernel.h"
 #include "arm_compute/core/NEON/kernels/NEDepthwiseConvolution3x3Kernel.h"
+#include "arm_compute/core/NEON/kernels/NEDepthwiseIm2ColKernel.h"
+#include "arm_compute/core/NEON/kernels/NEDepthwiseVectorToTensorKernel.h"
+#include "arm_compute/core/NEON/kernels/NEDepthwiseWeightsReshapeKernel.h"
 #include "arm_compute/core/NEON/kernels/NEDequantizationLayerKernel.h"
 #include "arm_compute/core/NEON/kernels/NEDerivativeKernel.h"
 #include "arm_compute/core/NEON/kernels/NEDilateKernel.h"
@@ -68,6 +71,7 @@
 #include "arm_compute/core/NEON/kernels/NEGEMMMatrixAccumulateBiasesKernel.h"
 #include "arm_compute/core/NEON/kernels/NEGEMMMatrixAdditionKernel.h"
 #include "arm_compute/core/NEON/kernels/NEGEMMMatrixMultiplyKernel.h"
+#include "arm_compute/core/NEON/kernels/NEGEMMMatrixVectorMultiplyKernel.h"
 #include "arm_compute/core/NEON/kernels/NEGEMMTranspose1xWKernel.h"
 #include "arm_compute/core/NEON/kernels/NEGaussian3x3Kernel.h"
 #include "arm_compute/core/NEON/kernels/NEGaussian5x5Kernel.h"
diff --git a/arm_compute/core/NEON/kernels/NEDepthwiseIm2ColKernel.h b/arm_compute/core/NEON/kernels/NEDepthwiseIm2ColKernel.h
new file mode 100644
index 0000000000..fde474d1f5
--- /dev/null
+++ b/arm_compute/core/NEON/kernels/NEDepthwiseIm2ColKernel.h
@@ -0,0 +1,74 @@
+/*
+ * Copyright (c) 2017 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_NEDEPTHWISEIM2COLKERNEL_H__
+#define __ARM_COMPUTE_NEDEPTHWISEIM2COLKERNEL_H__
+
+#include "arm_compute/core/NEON/INEKernel.h"
+#include "arm_compute/core/Size2D.h"
+
+namespace arm_compute
+{
+class ITensor;
+
+/** 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 in per image and third dimension unchanged .
+ **/
+class NEDepthwiseIm2ColKernel : public INEKernel
+{
+public:
+    /** Default constructor */
+    NEDepthwiseIm2ColKernel();
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    NEDepthwiseIm2ColKernel(const NEDepthwiseIm2ColKernel &) = delete;
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    NEDepthwiseIm2ColKernel &operator=(const NEDepthwiseIm2ColKernel &) = delete;
+    /** Allow instances of this class to be moved */
+    NEDepthwiseIm2ColKernel(NEDepthwiseIm2ColKernel &&) = default;
+    /** Allow instances of this class to be moved */
+    NEDepthwiseIm2ColKernel &operator=(NEDepthwiseIm2ColKernel &&) = 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: 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.
+     */
+    void configure(const ITensor *input, ITensor *output, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias = false);
+
+    // Inherited methods overridden:
+    void run(const Window &window, const ThreadInfo &info) override;
+
+private:
+    const ITensor *_input;
+    ITensor       *_output;
+    Size2D         _kernel_dims;
+    PadStrideInfo  _conv_info;
+    bool           _has_bias;
+};
+} // arm_compute
+#endif /*__ARM_COMPUTE_NEDEPTHWISEIM2COLKERNEL_H__ */
diff --git a/arm_compute/core/NEON/kernels/NEDepthwiseVectorToTensorKernel.h b/arm_compute/core/NEON/kernels/NEDepthwiseVectorToTensorKernel.h
new file mode 100644
index 0000000000..8b33fae6f3
--- /dev/null
+++ b/arm_compute/core/NEON/kernels/NEDepthwiseVectorToTensorKernel.h
@@ -0,0 +1,71 @@
+/*
+ * Copyright (c) 2017 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_NEDEPTHWISEVECTORTOTENSORKERNEL_H__
+#define __ARM_COMPUTE_NEDEPTHWISEVECTORTOTENSORKERNEL_H__
+
+#include "arm_compute/core/NEON/INEKernel.h"
+
+namespace arm_compute
+{
+class ITensor;
+
+/** 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 NEDepthwiseVectorToTensorKernel : public INEKernel
+{
+public:
+    /** Default constructor */
+    NEDepthwiseVectorToTensorKernel();
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    NEDepthwiseVectorToTensorKernel(const NEDepthwiseVectorToTensorKernel &) = delete;
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    NEDepthwiseVectorToTensorKernel &operator=(const NEDepthwiseVectorToTensorKernel &) = delete;
+    /** Allow instances of this class to be moved */
+    NEDepthwiseVectorToTensorKernel(NEDepthwiseVectorToTensorKernel &&) = default;
+    /** Allow instances of this class to be moved */
+    NEDepthwiseVectorToTensorKernel &operator=(NEDepthwiseVectorToTensorKernel &&) = default;
+    /** Set the input and output of the kernel.
+     *
+     * @param[in]  input  The input vector to convert. Data type supported: 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 ITensor *input, ITensor *output, size_t conv_w, size_t conv_h);
+
+    // Inherited methods overridden:
+    void run(const Window &window, const ThreadInfo &info) override;
+
+private:
+    const ITensor *_input;
+    ITensor       *_output;
+    std::pair<size_t, size_t> _conv_dims;
+};
+} // arm_compute
+#endif /*__ARM_COMPUTE_NEDEPTHWISEVECTORTOTENSORKERNEL_H__ */
diff --git a/arm_compute/core/NEON/kernels/NEDepthwiseWeightsReshapeKernel.h b/arm_compute/core/NEON/kernels/NEDepthwiseWeightsReshapeKernel.h
new file mode 100644
index 0000000000..2e986117df
--- /dev/null
+++ b/arm_compute/core/NEON/kernels/NEDepthwiseWeightsReshapeKernel.h
@@ -0,0 +1,67 @@
+/*
+ * Copyright (c) 2017 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_NEDEPTHWISEWEIGHTSRESHAPEKERNEL_H__
+#define __ARM_COMPUTE_NEDEPTHWISEWEIGHTSRESHAPEKERNEL_H__
+
+#include "arm_compute/core/NEON/INEKernel.h"
+
+namespace arm_compute
+{
+class ITensor;
+
+/** Interface for the depthwise weights reshape kernel.
+ *  This kernel reshape original weights' low 2D dimensions into a single col and
+ *  have the second dimension as the original depth size.
+ **/
+class NEDepthwiseWeightsReshapeKernel : public INEKernel
+{
+public:
+    /** Default constructor */
+    NEDepthwiseWeightsReshapeKernel();
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    NEDepthwiseWeightsReshapeKernel(const NEDepthwiseWeightsReshapeKernel &) = delete;
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    NEDepthwiseWeightsReshapeKernel &operator=(const NEDepthwiseWeightsReshapeKernel &) = delete;
+    /** Allow instances of this class to be moved */
+    NEDepthwiseWeightsReshapeKernel(NEDepthwiseWeightsReshapeKernel &&) = default;
+    /** Allow instances of this class to be moved */
+    NEDepthwiseWeightsReshapeKernel &operator=(NEDepthwiseWeightsReshapeKernel &&) = 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: 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 ITensor *input, ITensor *output, const ITensor *biases);
+
+    // Inherited methods overridden:
+    void run(const Window &window, const ThreadInfo &info) override;
+
+private:
+    const ITensor *_input;
+    ITensor       *_output;
+    const ITensor *_biases;
+};
+} // arm_compute
+#endif /*__ARM_COMPUTE_NEDEPTHWISEWEIGHTSRESHAPEKERNEL_H__ */
diff --git a/arm_compute/core/NEON/kernels/NEGEMMMatrixVectorMultiplyKernel.h b/arm_compute/core/NEON/kernels/NEGEMMMatrixVectorMultiplyKernel.h
new file mode 100644
index 0000000000..d844af5d54
--- /dev/null
+++ b/arm_compute/core/NEON/kernels/NEGEMMMatrixVectorMultiplyKernel.h
@@ -0,0 +1,63 @@
+/*
+ * Copyright (c) 2016, 2017 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_NEGEMMMATRIXVECTORMULTIPLYKERNEL_H_
+#define __ARM_COMPUTE_NEGEMMMATRIXVECTORMULTIPLYKERNEL_H_
+
+#include "arm_compute/core/NEON/INESimpleKernel.h"
+
+namespace arm_compute
+{
+class ITensor;
+
+class NEGEMMMatrixVectorMultiplyKernel : public INESimpleKernel
+{
+public:
+    /** Default constructor */
+    NEGEMMMatrixVectorMultiplyKernel();
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    NEGEMMMatrixVectorMultiplyKernel(const NEGEMMMatrixVectorMultiplyKernel &) = delete;
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    NEGEMMMatrixVectorMultiplyKernel &operator=(const NEGEMMMatrixVectorMultiplyKernel &) = delete;
+    /** Allow instances of this class to be moved */
+    NEGEMMMatrixVectorMultiplyKernel(NEGEMMMatrixVectorMultiplyKernel &&) = default;
+    /** Allow instances of this class to be moved */
+    NEGEMMMatrixVectorMultiplyKernel &operator=(NEGEMMMatrixVectorMultiplyKernel &&) = default;
+    /** Initialise the kernel's input and output.
+     *
+     * @param[in]  input0 First Input tensor. Data types supported: F16/F32
+     * @param[in]  input1 Second Input tensor. Data types supported: same as @p input.
+     * @param[out] output Output tensor which stores the interleaved matrix. Data type supported: same as @p input.
+     */
+    void configure(const ITensor *input0, const ITensor *input1, ITensor *output);
+
+    // Inherited methods overridden:
+    void run(const Window &window, const ThreadInfo &info) override;
+
+private:
+    const ITensor *_input0;
+    const ITensor *_input1;
+    ITensor       *_output;
+};
+} // namespace arm_compute
+#endif /*__ARM_COMPUTE_NEGEMMMATRIXVECTORMULTIPLYKERNEL_H_*/
diff --git a/arm_compute/core/NEON/kernels/convolution/NEDirectConvolution3x3.h b/arm_compute/core/NEON/kernels/convolution/NEDirectConvolution3x3.h
new file mode 100644
index 0000000000..7f39e5ee8d
--- /dev/null
+++ b/arm_compute/core/NEON/kernels/convolution/NEDirectConvolution3x3.h
@@ -0,0 +1,172 @@
+/*
+ * Copyright (c) 2017 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_NECONVOLUTIONKERNEL3x3_H__
+#define __ARM_COMPUTE_NECONVOLUTIONKERNEL3x3_H__
+
+#include <arm_neon.h>
+
+namespace arm_compute
+{
+namespace detail
+{
+inline float32x4x3_t load_matrix_row(const float *ptr)
+{
+    const float32x4x3_t r =
+    {
+        {
+            vld1q_dup_f32(ptr),
+            vld1q_dup_f32(1 + ptr),
+            vld1q_dup_f32(2 + ptr)
+        }
+    };
+    return r;
+}
+
+template <unsigned int stridex>
+float32x4x2_t convolve_3x3(const float *in_top, const float *in_mid, const float *in_low, const float32x4x3_t &m0, const float32x4x3_t &m1, const float32x4x3_t &m2, int fixed_point_position);
+
+template <>
+inline float32x4x2_t convolve_3x3<1>(const float *in_top, const float *in_mid, const float *in_low, const float32x4x3_t &m0, const float32x4x3_t &m1, const float32x4x3_t &m2, int fixed_point_position)
+{
+    ARM_COMPUTE_UNUSED(fixed_point_position);
+
+    const float32x4x3_t vtop =
+    {
+        {
+            vld1q_f32(in_top),
+            vld1q_f32(in_top + 4),
+            vld1q_f32(in_top + 8)
+        }
+    };
+    const float32x4x3_t vmid =
+    {
+        {
+            vld1q_f32(in_mid),
+            vld1q_f32(in_mid + 4),
+            vld1q_f32(in_mid + 8)
+        }
+    };
+    const float32x4x3_t vlow =
+    {
+        {
+            vld1q_f32(in_low),
+            vld1q_f32(in_low + 4),
+            vld1q_f32(in_low + 8)
+        }
+    };
+    float32x4x2_t out =
+    {
+        {
+            vmulq_f32(vtop.val[0], m0.val[0]),
+            vmulq_f32(vtop.val[1], m0.val[0])
+        }
+    };
+    out.val[0] = vmlaq_f32(out.val[0], vextq_f32(vtop.val[0], vtop.val[1], 1), m0.val[1]);
+    out.val[0] = vmlaq_f32(out.val[0], vextq_f32(vtop.val[0], vtop.val[1], 2), m0.val[2]);
+
+    out.val[0] = vmlaq_f32(out.val[0], vmid.val[0], m1.val[0]);
+    out.val[0] = vmlaq_f32(out.val[0], vextq_f32(vmid.val[0], vmid.val[1], 1), m1.val[1]);
+    out.val[0] = vmlaq_f32(out.val[0], vextq_f32(vmid.val[0], vmid.val[1], 2), m1.val[2]);
+
+    out.val[0] = vmlaq_f32(out.val[0], vlow.val[0], m2.val[0]);
+    out.val[0] = vmlaq_f32(out.val[0], vextq_f32(vlow.val[0], vlow.val[1], 1), m2.val[1]);
+    out.val[0] = vmlaq_f32(out.val[0], vextq_f32(vlow.val[0], vlow.val[1], 2), m2.val[2]);
+
+    out.val[1] = vmlaq_f32(out.val[1], vextq_f32(vtop.val[1], vtop.val[2], 1), m0.val[1]);
+    out.val[1] = vmlaq_f32(out.val[1], vextq_f32(vtop.val[1], vtop.val[2], 2), m0.val[2]);
+
+    out.val[1] = vmlaq_f32(out.val[1], vmid.val[1], m1.val[0]);
+    out.val[1] = vmlaq_f32(out.val[1], vextq_f32(vmid.val[1], vmid.val[2], 1), m1.val[1]);
+    out.val[1] = vmlaq_f32(out.val[1], vextq_f32(vmid.val[1], vmid.val[2], 2), m1.val[2]);
+
+    out.val[1] = vmlaq_f32(out.val[1], vlow.val[1], m2.val[0]);
+    out.val[1] = vmlaq_f32(out.val[1], vextq_f32(vlow.val[1], vlow.val[2], 1), m2.val[1]);
+    out.val[1] = vmlaq_f32(out.val[1], vextq_f32(vlow.val[1], vlow.val[2], 2), m2.val[2]);
+    return out;
+}
+
+template <>
+inline float32x4x2_t convolve_3x3<2>(const float *in_top, const float *in_mid, const float *in_low, const float32x4x3_t &m0, const float32x4x3_t &m1, const float32x4x3_t &m2, int fixed_point_position)
+{
+    float32x4x2_t out = convolve_3x3<1>(in_top, in_mid, in_low, m0, m1, m2, fixed_point_position);
+    out.val[0]        = vsetq_lane_f32(vgetq_lane_f32(out.val[0], 2), out.val[0], 1);
+    out.val[0]        = vsetq_lane_f32(vgetq_lane_f32(out.val[1], 0), out.val[0], 2);
+    out.val[0]        = vsetq_lane_f32(vgetq_lane_f32(out.val[1], 2), out.val[0], 3);
+    return out;
+}
+
+template <>
+inline float32x4x2_t convolve_3x3<3>(const float *in_top, const float *in_mid, const float *in_low, const float32x4x3_t &m0, const float32x4x3_t &m1, const float32x4x3_t &m2, int fixed_point_position)
+{
+    float32x4x2_t out = convolve_3x3<1>(in_top, in_mid, in_low, m0, m1, m2, fixed_point_position);
+    out.val[0]        = vsetq_lane_f32(vgetq_lane_f32(out.val[0], 3), out.val[0], 1);
+    return out;
+}
+
+template <unsigned int stridex>
+void store_results(float *buffer, const float32x4x2_t &values);
+
+template <>
+void store_results<1>(float *buffer, const float32x4x2_t &values)
+{
+    vst1q_f32(buffer, values.val[0]);
+    vst1q_f32(buffer + 4, values.val[1]);
+}
+
+template <>
+void store_results<2>(float *buffer, const float32x4x2_t &values)
+{
+    vst1q_f32(buffer, values.val[0]);
+}
+
+template <>
+void store_results<3>(float *buffer, const float32x4x2_t &values)
+{
+    vst1_f32(buffer, vget_low_f32(values.val[0]));
+}
+
+template <unsigned int stridex>
+int get_input_num_elems_processed(unsigned int num_elems_written_per_iteration);
+
+template <>
+int get_input_num_elems_processed<1>(unsigned int num_elems_written_per_iteration)
+{
+    return num_elems_written_per_iteration;
+}
+
+template <>
+int get_input_num_elems_processed<2>(unsigned int num_elems_written_per_iteration)
+{
+    return num_elems_written_per_iteration << 1;
+}
+
+template <>
+int get_input_num_elems_processed<3>(unsigned int num_elems_written_per_iteration)
+{
+    return num_elems_written_per_iteration * 3;
+}
+}
+} // namespace arm_compute
+#endif /* __ARM_COMPUTE_NECONVOLUTIONKERNEL3x3_H__ */
\ No newline at end of file
diff --git a/arm_compute/runtime/NEON/functions/NEDepthwiseConvolution.h b/arm_compute/runtime/NEON/functions/NEDepthwiseConvolution.h
index acf66f8dd4..f2c209cd80 100644
--- a/arm_compute/runtime/NEON/functions/NEDepthwiseConvolution.h
+++ b/arm_compute/runtime/NEON/functions/NEDepthwiseConvolution.h
@@ -25,8 +25,12 @@
 #define __ARM_COMPUTE_NEDEPTHWISECONVOLUTION_H__
 
 #include "arm_compute/core/NEON/kernels/NEDepthwiseConvolution3x3Kernel.h"
+#include "arm_compute/core/NEON/kernels/NEDepthwiseIm2ColKernel.h"
+#include "arm_compute/core/NEON/kernels/NEDepthwiseVectorToTensorKernel.h"
+#include "arm_compute/core/NEON/kernels/NEDepthwiseWeightsReshapeKernel.h"
 #include "arm_compute/core/NEON/kernels/NEDirectConvolutionLayerBiasAccumulateKernel.h"
 #include "arm_compute/core/NEON/kernels/NEFillBorderKernel.h"
+#include "arm_compute/core/NEON/kernels/NEGEMMMatrixVectorMultiplyKernel.h"
 #include "arm_compute/core/Types.h"
 #include "arm_compute/runtime/IFunction.h"
 #include "arm_compute/runtime/IMemoryManager.h"
@@ -68,5 +72,42 @@ private:
     NEFillBorderKernel                           _border_handler;
     bool                                         _has_bias;
 };
+
+/** Basic function to execute a generic depthwise convolution. This function calls the following OpenCL kernels:
+ *
+ * -# @ref NEDepthwiseIm2ColKernel
+ * -# @ref NEDepthwiseWeightsReshapeKernel
+ * -# @ref NEGEMMMatrixVectorMultiplyKernel
+ * -# @ref NEFillBorderKernel (if pad_x or pad_y > 0)
+ *
+ */
+class NEDepthwiseConvolution : public IFunction
+{
+public:
+    /** Default constructor */
+    NEDepthwiseConvolution();
+    /** Initialize the function's source, destination, weights and convolution information.
+     *
+     * @param[in, out] input     Source tensor. Data type supported: F32. (Written to only for border filling).
+     * @param[out]     output    Destination tensor. Data type supported: same as @p input.
+     * @param[in]      weights   Weights tensor. These are 3D tensors with shape [kernel_x, kernel_y, IFM]. Data type supported: Same as @p input.
+     * @param[in]      biases    (Optional) Biases tensor. A 1D tensor with shape [IFM]. Must be nullptr if not needed.
+     *                           Data type supported: Same as @p input.
+     * @param[in]      conv_info Padding and stride information to use for the convolution.
+     */
+    void configure(ITensor *input, const ITensor *weights, const ITensor *biases, ITensor *output, const PadStrideInfo &conv_info);
+
+    // Inherited methods overriden:
+    void run() override;
+
+private:
+    NEDepthwiseIm2ColKernel          _im2col_kernel;
+    NEDepthwiseWeightsReshapeKernel  _weights_reshape_kernel;
+    NEGEMMMatrixVectorMultiplyKernel _v2mm_kernel;
+    NEDepthwiseVectorToTensorKernel  _vector_to_tensor_kernel;
+    Tensor                           _input_reshaped;
+    Tensor                           _weights_reshaped;
+    Tensor                           _v2mm_output;
+};
 }
 #endif /* __ARM_COMPUTE_NEDEPTHWISECONVOLUTION_H__ */
\ No newline at end of file
diff --git a/src/core/CL/kernels/CLDepthwiseIm2ColKernel.cpp b/src/core/CL/kernels/CLDepthwiseIm2ColKernel.cpp
index 743cd4a38f..c23941426e 100644
--- a/src/core/CL/kernels/CLDepthwiseIm2ColKernel.cpp
+++ b/src/core/CL/kernels/CLDepthwiseIm2ColKernel.cpp
@@ -32,6 +32,7 @@
 #include "arm_compute/core/Types.h"
 #include "support/ToolchainSupport.h"
 
+#include "arm_compute/runtime/CL/CLScheduler.h"
 #include <tuple>
 
 using namespace arm_compute;
diff --git a/src/core/NEON/kernels/NEDepthwiseIm2ColKernel.cpp b/src/core/NEON/kernels/NEDepthwiseIm2ColKernel.cpp
new file mode 100644
index 0000000000..2ceb39d217
--- /dev/null
+++ b/src/core/NEON/kernels/NEDepthwiseIm2ColKernel.cpp
@@ -0,0 +1,138 @@
+/*
+ * Copyright (c) 2017 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/NEON/kernels/NEDepthwiseIm2ColKernel.h"
+
+#include "arm_compute/core/AccessWindowTranspose.h"
+#include "arm_compute/core/Coordinates.h"
+#include "arm_compute/core/Error.h"
+#include "arm_compute/core/Helpers.h"
+#include "arm_compute/core/ITensor.h"
+#include "arm_compute/core/NEON/INEKernel.h"
+#include "arm_compute/core/TensorInfo.h"
+#include "arm_compute/core/TensorShape.h"
+#include "arm_compute/core/Types.h"
+#include "arm_compute/core/Validate.h"
+#include "arm_compute/core/Window.h"
+
+using namespace arm_compute;
+
+NEDepthwiseIm2ColKernel::NEDepthwiseIm2ColKernel()
+    : _input(nullptr), _output(nullptr), _kernel_dims(), _conv_info(), _has_bias()
+{
+}
+
+void NEDepthwiseIm2ColKernel::configure(const ITensor *input, ITensor *output, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias)
+{
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F16, DataType::F32);
+    ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
+    ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(input, output);
+    ARM_COMPUTE_ERROR_ON(input->info()->dimension(2) != output->info()->dimension(2));
+    ARM_COMPUTE_ERROR_ON(output->info()->dimension(0) != (kernel_dims.width * kernel_dims.height + ((has_bias) ? 1 : 0)));
+
+    _input       = input;
+    _output      = output;
+    _kernel_dims = kernel_dims;
+    _conv_info   = conv_info;
+    _has_bias    = has_bias;
+
+    // Configure kernel window
+    Window win = calculate_max_window(*input->info(), Steps());
+
+    // The NEDepthwiseIm2ColKernel doesn't need padding so update_window_and_padding() can be skipped
+    output->info()->set_valid_region(ValidRegion(Coordinates(), output->info()->tensor_shape()));
+
+    INEKernel::configure(win);
+}
+
+void NEDepthwiseIm2ColKernel::run(const Window &window, const ThreadInfo &info)
+{
+    ARM_COMPUTE_UNUSED(info);
+    ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
+
+    //const int kernel_depth   = _input->info()->dimension(2);
+    const int input_w        = _input->info()->dimension(0);
+    const int input_h        = _input->info()->dimension(1);
+    const int input_stride_x = _input->info()->strides_in_bytes().x();
+    const int input_stride_y = _input->info()->strides_in_bytes().y();
+    const int input_stride_z = _input->info()->strides_in_bytes().z();
+    const int stride_x       = _conv_info.stride().first;
+    const int stride_y       = _conv_info.stride().second;
+
+    const int pad_left  = _conv_info.pad_left();
+    const int pad_right = _conv_info.pad_right();
+    const int pad_top   = _conv_info.pad_top();
+
+    Window window_in(window);
+    // The first three dimensions of the input are increased by the inner loops
+    window_in.set(Window::DimX, Window::Dimension(0, 0, 0));
+    window_in.set(Window::DimY, Window::Dimension(0, 0, 0));
+    window_in.set(Window::DimZ, Window::Dimension(0, 0, 0));
+
+    // Setup output window
+    Window window_out(window);
+    window_out.set(Window::DimX, Window::Dimension(0, _output->info()->dimension(0), _output->info()->dimension(0)));
+    window_out.set(Window::DimY, Window::Dimension(0, _output->info()->dimension(1), 1));
+    window_out.set(Window::DimZ, Window::Dimension(0, _output->info()->dimension(2), 1));
+
+    Iterator in(_input, window_in);
+    Iterator out(_output, window_out);
+
+    const int full_length   = input_w + pad_left + pad_right;
+    const int max_initial_x = stride_x * (((full_length - _kernel_dims.width) / stride_x) + 1);
+
+    execute_window_loop(window_out, [&](const Coordinates & id)
+    {
+        const int src_pixel_linear = id.y() * stride_x;
+
+        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;
+
+        // Get pointers
+        const uint8_t *const input_ptr  = in.ptr() + id.z() * input_stride_z;
+        auto                 output_ptr = reinterpret_cast<float *>(out.ptr());
+        const int            height     = src_y + _kernel_dims.height;
+        const int            width      = src_x + _kernel_dims.width;
+
+        for(int y = src_y; y < height; ++y)
+        {
+            for(int x = src_x; x < width; ++x, ++output_ptr)
+            {
+                if(x < 0 || x >= input_w || y < 0 || y >= input_h)
+                {
+                    *output_ptr = 0;
+                }
+                else
+                {
+                    *output_ptr = *(reinterpret_cast<const float *>(input_ptr + x * input_stride_x + y * input_stride_y));
+                }
+            }
+        }
+
+        if(_has_bias)
+        {
+            *output_ptr = static_cast<float>(1);
+        }
+    },
+    in, out);
+}
diff --git a/src/core/NEON/kernels/NEDepthwiseVectorToTensorKernel.cpp b/src/core/NEON/kernels/NEDepthwiseVectorToTensorKernel.cpp
new file mode 100644
index 0000000000..6deda506ab
--- /dev/null
+++ b/src/core/NEON/kernels/NEDepthwiseVectorToTensorKernel.cpp
@@ -0,0 +1,95 @@
+/*
+ * Copyright (c) 2017 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/NEON/kernels/NEDepthwiseVectorToTensorKernel.h"
+
+#include "arm_compute/core/AccessWindowTranspose.h"
+#include "arm_compute/core/Coordinates.h"
+#include "arm_compute/core/Error.h"
+#include "arm_compute/core/Helpers.h"
+#include "arm_compute/core/ITensor.h"
+#include "arm_compute/core/NEON/INEKernel.h"
+#include "arm_compute/core/TensorInfo.h"
+#include "arm_compute/core/TensorShape.h"
+#include "arm_compute/core/Types.h"
+#include "arm_compute/core/Validate.h"
+#include "arm_compute/core/Window.h"
+
+using namespace arm_compute;
+
+NEDepthwiseVectorToTensorKernel::NEDepthwiseVectorToTensorKernel()
+    : _input(nullptr), _output(nullptr), _conv_dims()
+{
+}
+
+void NEDepthwiseVectorToTensorKernel::configure(const ITensor *input, ITensor *output, size_t conv_w, size_t conv_h)
+{
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F16, DataType::F32);
+    ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
+    ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(input, output);
+
+    _input     = input;
+    _output    = output;
+    _conv_dims = std::pair<size_t, size_t>(conv_w, conv_h);
+
+    // Configure  kernel window
+    Window win = calculate_max_window(*input->info(), Steps());
+    // The NEDepthwisevectorToTensorKernel doesn't need padding so update_window_and_padding() can be skipped
+    output->info()->set_valid_region(ValidRegion(Coordinates(), output->info()->tensor_shape()));
+
+    INEKernel::configure(win);
+}
+
+void NEDepthwiseVectorToTensorKernel::run(const Window &window, const ThreadInfo &info)
+{
+    ARM_COMPUTE_UNUSED(info);
+    ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
+
+    // const int input_w         = _input->info()->dimension(0);
+    const int output_stride_x = _output->info()->strides_in_bytes().x();
+    const int output_stride_y = _output->info()->strides_in_bytes().y();
+    const int output_stride_z = _output->info()->strides_in_bytes().z();
+
+    // Setup output window
+    Window window_out(window);
+    window_out.set(Window::DimX, Window::Dimension(0, 0, 0));
+    window_out.set(Window::DimY, Window::Dimension(0, 0, 0));
+    window_out.set(Window::DimZ, Window::Dimension(0, 0, 0));
+
+    Iterator in(_input, window);
+    Iterator out(_output, window_out);
+
+    const int patch_size = _conv_dims.first * _conv_dims.second;
+
+    execute_window_loop(window, [&](const Coordinates & id)
+    {
+        const int z       = id.x() / patch_size;
+        const int index2D = id.x() - z * patch_size;
+
+        auto input_ptr  = reinterpret_cast<float *>(in.ptr());
+        auto output_ptr = reinterpret_cast<float *>(out.ptr() + index2D % _conv_dims.first * output_stride_x + index2D / _conv_dims.first * output_stride_y + z * output_stride_z);
+
+        *output_ptr = *input_ptr;
+    },
+    in, out);
+}
diff --git a/src/core/NEON/kernels/NEDepthwiseWeightsReshapeKernel.cpp b/src/core/NEON/kernels/NEDepthwiseWeightsReshapeKernel.cpp
new file mode 100644
index 0000000000..6585fdb8b8
--- /dev/null
+++ b/src/core/NEON/kernels/NEDepthwiseWeightsReshapeKernel.cpp
@@ -0,0 +1,112 @@
+/*
+ * Copyright (c) 2017 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/NEON/kernels/NEDepthwiseWeightsReshapeKernel.h"
+
+#include "arm_compute/core/AccessWindowTranspose.h"
+#include "arm_compute/core/Coordinates.h"
+#include "arm_compute/core/Error.h"
+#include "arm_compute/core/Helpers.h"
+#include "arm_compute/core/ITensor.h"
+#include "arm_compute/core/NEON/INEKernel.h"
+#include "arm_compute/core/TensorInfo.h"
+#include "arm_compute/core/TensorShape.h"
+#include "arm_compute/core/Types.h"
+#include "arm_compute/core/Validate.h"
+#include "arm_compute/core/Window.h"
+
+using namespace arm_compute;
+
+NEDepthwiseWeightsReshapeKernel::NEDepthwiseWeightsReshapeKernel()
+    : _input(nullptr), _output(nullptr), _biases(nullptr)
+{
+}
+
+void NEDepthwiseWeightsReshapeKernel::configure(const ITensor *input, ITensor *output, const ITensor *biases)
+{
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F16, DataType::F32);
+    ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
+    ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(input, output);
+    ARM_COMPUTE_ERROR_ON(input->info()->dimension(2) != output->info()->dimension(1));
+    ARM_COMPUTE_ERROR_ON(output->info()->dimension(0) != (input->info()->dimension(0) * input->info()->dimension(1) + ((biases != nullptr) ? 1 : 0)));
+
+    if(biases != nullptr)
+    {
+        ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, biases);
+        ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(input, biases);
+        ARM_COMPUTE_ERROR_ON(biases->info()->dimension(0) != input->info()->dimension(2));
+        ARM_COMPUTE_ERROR_ON(biases->info()->num_dimensions() > 1);
+    }
+
+    _input  = input;
+    _output = output;
+    _biases = biases;
+
+    // Configure  kernel window
+    Window win = calculate_max_window(*input->info(), Steps());
+    // The NEDepthwiseWeightsReshapeKernel doesn't need padding so update_window_and_padding() can be skipped
+    output->info()->set_valid_region(ValidRegion(Coordinates(), output->info()->tensor_shape()));
+
+    INEKernel::configure(win);
+}
+
+void NEDepthwiseWeightsReshapeKernel::run(const Window &window, const ThreadInfo &info)
+{
+    ARM_COMPUTE_UNUSED(info);
+    ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
+
+    const int input_w         = _input->info()->dimension(0);
+    const int output_stride_x = _output->info()->strides_in_bytes().x();
+    const int output_stride_y = _output->info()->strides_in_bytes().y();
+
+    Window window_in(window);
+    // The first three dimensions of the input are increased by the inner loops
+    window_in.set(Window::DimX, Window::Dimension(0, _input->info()->dimension(0), _input->info()->dimension(0)));
+    window_in.set(Window::DimY, Window::Dimension(0, _input->info()->dimension(1), 1));
+    window_in.set(Window::DimZ, Window::Dimension(0, _input->info()->dimension(2), 1));
+
+    // Setup output window
+    Window window_out;
+    window_out.set(Window::DimX, Window::Dimension(0, 0, 0));
+    window_out.set(Window::DimY, Window::Dimension(0, 0, 0));
+
+    Iterator in(_input, window_in);
+    Iterator out(_output, window_out);
+
+    execute_window_loop(window_in, [&](const Coordinates & id)
+    {
+        auto input_ptr  = reinterpret_cast<float *>(in.ptr());
+        auto output_ptr = reinterpret_cast<float *>(out.ptr() + id.y() * input_w * output_stride_x + id.z() * output_stride_y);
+
+        for(int i = 0; i < input_w; ++i, ++input_ptr)
+        {
+            *(output_ptr + i) = *input_ptr;
+        }
+
+        if(_biases != nullptr)
+        {
+            *(output_ptr + input_w) = *(reinterpret_cast<float *>(_biases->ptr_to_element(Coordinates(id.z()))));
+        }
+    },
+    in, out);
+}
diff --git a/src/core/NEON/kernels/NEGEMMMatrixVectorMultiplyKernel.cpp b/src/core/NEON/kernels/NEGEMMMatrixVectorMultiplyKernel.cpp
new file mode 100644
index 0000000000..c28dcf7463
--- /dev/null
+++ b/src/core/NEON/kernels/NEGEMMMatrixVectorMultiplyKernel.cpp
@@ -0,0 +1,131 @@
+/*
+ * Copyright (c) 2016, 2017 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/NEON/kernels/NEGEMMMatrixVectorMultiplyKernel.h"
+
+#include "arm_compute/core/AccessWindowStatic.h"
+#include "arm_compute/core/Error.h"
+#include "arm_compute/core/Helpers.h"
+#include "arm_compute/core/ITensor.h"
+#include "arm_compute/core/NEON/INEKernel.h"
+#include "arm_compute/core/Types.h"
+#include "arm_compute/core/Validate.h"
+#include "arm_compute/core/Window.h"
+
+#include <arm_neon.h>
+#include <cstddef>
+#include <cstdint>
+#include <tuple>
+
+using namespace arm_compute;
+
+NEGEMMMatrixVectorMultiplyKernel::NEGEMMMatrixVectorMultiplyKernel()
+    : _input0(nullptr), _input1(nullptr), _output(nullptr)
+{
+}
+
+void NEGEMMMatrixVectorMultiplyKernel::configure(const ITensor *input0, const ITensor *input1, ITensor *output)
+{
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input0, 1, DataType::F16, DataType::F32);
+    ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input0, input1, output);
+    ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(input0, input1, output);
+    ARM_COMPUTE_ERROR_ON(input0->info()->dimension(2) != input1->info()->dimension(1));
+
+    _input0 = input0;
+    _input1 = input1;
+    _output = output;
+
+    // Configure kernel window
+    const unsigned int num_elems_read_per_iteration = 4;
+
+    Window win = calculate_max_window(*input0->info(), Steps(num_elems_read_per_iteration));
+
+    AccessWindowHorizontal input0_access(input0->info(), 0, num_elems_read_per_iteration);
+    AccessWindowHorizontal input1_access(input1->info(), 0, num_elems_read_per_iteration);
+    AccessWindowHorizontal output_access(output->info(), 0, 1);
+
+    update_window_and_padding(win, input0_access, input1_access, output_access);
+
+    _output->info()->set_valid_region(ValidRegion(Coordinates(), _output->info()->tensor_shape()));
+
+    INEKernel::configure(win);
+}
+
+void NEGEMMMatrixVectorMultiplyKernel::run(const Window &window, const ThreadInfo &info)
+{
+    ARM_COMPUTE_UNUSED(info);
+    ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
+
+    Window window_slice = window.first_slice_window_3D();
+
+    Window window_in(window);
+    Window window_weights(window_slice);
+    Window window_out(window);
+
+    // Setup input0 slice
+    window_in.set(Window::DimX, Window::Dimension(0, _input0->info()->dimension(0), _input0->info()->dimension(0)));
+    window_in.set(Window::DimY, Window::Dimension(0, _input0->info()->dimension(1), 1));
+    window_in.set(Window::DimZ, Window::Dimension(0, _input0->info()->dimension(2), 1));
+
+    // Setup input1 and output slice. Their dimensions are increased in the kernel.
+    window_weights.set(Window::DimX, Window::Dimension(0, 0, 0));
+    window_weights.set(Window::DimY, Window::Dimension(0, 0, 0));
+    window_weights.set(Window::DimZ, Window::Dimension(0, 0, 0));
+
+    window_out.set(Window::DimX, Window::Dimension(0, 0, 0));
+    window_out.set(Window::DimY, Window::Dimension(0, 0, 0));
+    window_out.set(Window::DimZ, Window::Dimension(0, 0, 0));
+
+    Iterator in(_input0, window_in);
+    Iterator in2(_input1, window_weights);
+    Iterator out(_output, window_out);
+
+    const int input_w          = _input0->info()->dimension(0);
+    const int input_h          = _input0->info()->dimension(1);
+    const int input_stride_x   = _input0->info()->strides_in_bytes().x();
+    const int weights_stride_x = _input1->info()->strides_in_bytes().x();
+    const int weights_stride_y = _input1->info()->strides_in_bytes().y();
+    const int output_stride_x  = _output->info()->strides_in_bytes().x();
+
+    execute_window_loop(window_in, [&](const Coordinates & id)
+    {
+        // Get pointers
+        const uint8_t *const input_ptr   = in.ptr();
+        const uint8_t *const weights_ptr = in2.ptr() + id.z() * weights_stride_y;
+        auto                 output_ptr  = reinterpret_cast<float *>(out.ptr() + (id.y() + id.z() * input_h) * output_stride_x);
+
+        float32x4_t row_dot = vdupq_n_f32(0.f);
+        for(int i = 0; i < input_w; i += 4)
+        {
+            const auto input   = vld1q_f32(reinterpret_cast<const float *>(input_ptr + i * input_stride_x));
+            const auto weights = vld1q_f32(reinterpret_cast<const float *>(weights_ptr + i * weights_stride_x));
+            row_dot            = vaddq_f32(row_dot, vmulq_f32(input, weights));
+        }
+
+        auto temp = vadd_f32(vget_high_f32(row_dot), vget_low_f32(row_dot));
+        temp      = vpadd_f32(temp, temp);
+
+        *output_ptr = vget_lane_f32(temp, 0);
+    },
+    in, in2, out);
+}
diff --git a/src/runtime/NEON/functions/NEDepthwiseConvolution.cpp b/src/runtime/NEON/functions/NEDepthwiseConvolution.cpp
index fd8d419fa1..e12bc07464 100644
--- a/src/runtime/NEON/functions/NEDepthwiseConvolution.cpp
+++ b/src/runtime/NEON/functions/NEDepthwiseConvolution.cpp
@@ -59,4 +59,68 @@ void NEDepthwiseConvolution3x3::run()
     {
         NEScheduler::get().schedule(&_bias_kernel, Window::DimX);
     }
+}
+
+NEDepthwiseConvolution::NEDepthwiseConvolution()
+    : _im2col_kernel(), _weights_reshape_kernel(), _v2mm_kernel(), _vector_to_tensor_kernel(), _input_reshaped(), _weights_reshaped(), _v2mm_output()
+{
+}
+
+void NEDepthwiseConvolution::configure(ITensor *input, const ITensor *weights, const ITensor *biases, ITensor *output, const PadStrideInfo &conv_info)
+{
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32);
+    ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights);
+    ARM_COMPUTE_ERROR_ON(input->info()->dimension(2) != weights->info()->dimension(2));
+
+    const size_t weights_w = weights->info()->dimension(0);
+    const size_t weights_h = weights->info()->dimension(1);
+    const size_t weights_z = weights->info()->dimension(2);
+
+    bool has_bias = (biases != nullptr);
+
+    unsigned int conv_w = 0;
+    unsigned int conv_h = 0;
+    std::tie(conv_w, conv_h) = scaled_dimensions(input->info()->dimension(0), input->info()->dimension(1), weights_w, weights_h, conv_info);
+
+    // Set up intermediate tensors
+    const size_t patch_size = weights_w * weights_h + ((has_bias) ? 1 : 0);
+    const size_t conv_size  = conv_w * conv_h;
+
+    // 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);
+    const TensorInfo info_im2col(shape_im2col, 1, input->info()->data_type(), input->info()->fixed_point_position());
+    _input_reshaped.allocator()->init(info_im2col);
+    _im2col_kernel.configure(input, &_input_reshaped, Size2D(weights_w, weights_h), conv_info, has_bias);
+
+    // Weights reshape configuration
+    const TensorShape shape_weights_reshape(patch_size, weights_z);
+    const TensorInfo  info_weights_reshape(shape_weights_reshape, 1, weights->info()->data_type(), weights->info()->fixed_point_position());
+    _weights_reshaped.allocator()->init(info_weights_reshape);
+    _weights_reshape_kernel.configure(weights, &_weights_reshaped, biases);
+
+    // GEMV configuration
+    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);
+    const TensorInfo info_v2mm_out(shape_v2mm_out, 1, input->info()->data_type(), input->info()->fixed_point_position());
+    _v2mm_output.allocator()->init(info_v2mm_out);
+    _v2mm_kernel.configure(&_input_reshaped, &_weights_reshaped, &_v2mm_output);
+    _vector_to_tensor_kernel.configure(&_v2mm_output, output, conv_w, conv_h);
+
+    // Allocate intermediate tensors
+    _input_reshaped.allocator()->allocate();
+    _weights_reshaped.allocator()->allocate();
+    _v2mm_output.allocator()->allocate();
+}
+
+void NEDepthwiseConvolution::run()
+{
+    NEScheduler::get().schedule(&_im2col_kernel, Window::DimX);
+    NEScheduler::get().schedule(&_weights_reshape_kernel, Window::DimX);
+    NEScheduler::get().schedule(&_v2mm_kernel, Window::DimX);
+    NEScheduler::get().schedule(&_vector_to_tensor_kernel, Window::DimX);
 }
\ No newline at end of file
diff --git a/tests/validation/NEON/DepthwiseConvolution.cpp b/tests/validation/NEON/DepthwiseConvolution.cpp
index b6719b58e8..3a4b7aa2e9 100644
--- a/tests/validation/NEON/DepthwiseConvolution.cpp
+++ b/tests/validation/NEON/DepthwiseConvolution.cpp
@@ -84,12 +84,26 @@ DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(framework::da
     validate(dst.info()->padding(), padding);
 }
 
-template <typename T>
-using NEDepthwiseConvolutionFixture3x3 = DepthwiseConvolutionValidationFixture<Tensor, Accessor, NEDepthwiseConvolution3x3, T>;
-
 TEST_SUITE(Float)
 TEST_SUITE(F32)
+TEST_SUITE(Generic)
+template <typename T>
+using NEDepthwiseConvolutionFixture = DepthwiseConvolutionValidationFixture<Tensor, Accessor, NEDepthwiseConvolution, T>;
+FIXTURE_DATA_TEST_CASE(RunSmall, NEDepthwiseConvolutionFixture<float>, framework::DatasetMode::PRECOMMIT, combine(datasets::SmallDepthwiseConvolutionDataset(), framework::dataset::make("DataType",
+                                                                                                                  DataType::F32)))
+{
+    validate(Accessor(_target), _reference, tolerance_f32);
+}
+FIXTURE_DATA_TEST_CASE(RunLarge, NEDepthwiseConvolutionFixture<float>, framework::DatasetMode::NIGHTLY, combine(datasets::LargeDepthwiseConvolutionDataset(), framework::dataset::make("DataType",
+                                                                                                                DataType::F32)))
+{
+    validate(Accessor(_target), _reference, tolerance_f32);
+}
+TEST_SUITE_END()
+
 TEST_SUITE(W3x3)
+template <typename T>
+using NEDepthwiseConvolutionFixture3x3 = DepthwiseConvolutionValidationFixture<Tensor, Accessor, NEDepthwiseConvolution3x3, T>;
 FIXTURE_DATA_TEST_CASE(RunSmall, NEDepthwiseConvolutionFixture3x3<float>, framework::DatasetMode::ALL, combine(datasets::SmallDepthwiseConvolutionDataset3x3(), framework::dataset::make("DataType",
                                                                                                                DataType::F32)))
 {
@@ -101,6 +115,7 @@ FIXTURE_DATA_TEST_CASE(RunLarge, NEDepthwiseConvolutionFixture3x3<float>, framew
     validate(Accessor(_target), _reference, tolerance_f32);
 }
 TEST_SUITE_END()
+
 TEST_SUITE_END()
 TEST_SUITE_END()