COMPMID-1579: Add support for ChannelShuffle operator in NEON

Change-Id: I6d5f91579850906e1eb973ff6c5612195255e631

11 files changed, 567 insertions, 2 deletions

diff --git a/arm_compute/core/NEON/NEKernels.h b/arm_compute/core/NEON/NEKernels.h
index cb2e851b66..76eb5cb5f2 100644
--- a/arm_compute/core/NEON/NEKernels.h
+++ b/arm_compute/core/NEON/NEKernels.h
@@ -39,6 +39,7 @@
 #include "arm_compute/core/NEON/kernels/NECannyEdgeKernel.h"
 #include "arm_compute/core/NEON/kernels/NEChannelCombineKernel.h"
 #include "arm_compute/core/NEON/kernels/NEChannelExtractKernel.h"
+#include "arm_compute/core/NEON/kernels/NEChannelShuffleLayerKernel.h"
 #include "arm_compute/core/NEON/kernels/NECol2ImKernel.h"
 #include "arm_compute/core/NEON/kernels/NEColorConvertKernel.h"
 #include "arm_compute/core/NEON/kernels/NEConvertFullyConnectedWeightsKernel.h"
diff --git a/arm_compute/core/NEON/kernels/NEChannelShuffleLayerKernel.h b/arm_compute/core/NEON/kernels/NEChannelShuffleLayerKernel.h
new file mode 100644
index 0000000000..e9ef2d4a51
--- /dev/null
+++ b/arm_compute/core/NEON/kernels/NEChannelShuffleLayerKernel.h
@@ -0,0 +1,80 @@
+/*
+ * 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_NECHANNELSHUFFLELAYERKERNEL_H__
+#define __ARM_COMPUTE_NECHANNELSHUFFLELAYERKERNEL_H__
+
+#include "arm_compute/core/NEON/INEKernel.h"
+
+namespace arm_compute
+{
+// Forward declarations
+class ITensor;
+
+/** Interface for the channel shuffle kernel */
+class NEChannelShuffleLayerKernel : public INEKernel
+{
+public:
+    const char *name() const override
+    {
+        return "NEChannelShuffleLayerKernel";
+    }
+    /** Default constructor */
+    NEChannelShuffleLayerKernel();
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    NEChannelShuffleLayerKernel(const NEChannelShuffleLayerKernel &) = delete;
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    NEChannelShuffleLayerKernel &operator=(const NEChannelShuffleLayerKernel &) = delete;
+    /** Allow instances of this class to be moved */
+    NEChannelShuffleLayerKernel(NEChannelShuffleLayerKernel &&) = default;
+    /** Allow instances of this class to be moved */
+    NEChannelShuffleLayerKernel &operator=(NEChannelShuffleLayerKernel &&) = default;
+    /** Default destructor */
+    ~NEChannelShuffleLayerKernel() = default;
+    /** Configure function's inputs and outputs.
+     *
+     * @param[in]  input      Input tensor. Data types supported: U8/S8/QASYMM8/U16/S16/F16/U32/S32/F32
+     * @param[out] output     Output tensor. Data type supported: Same as @p input
+     * @param[in]  num_groups Number of groups. Must be greater than 1 and the number of channels of the tensors must be a multiple of the number of groups.
+     */
+    void configure(const ITensor *input, ITensor *output, unsigned int num_groups);
+    /** Static function to check if given info will lead to a valid configuration of @ref NEChannelShuffleLayerKernel
+     *
+     * @param[in]  input      Input tensor. Data types supported: U8/S8/QASYMM8/U16/S16/F16/U32/S32/F32
+     * @param[out] output     Output tensor. Data type supported: Same as @p input
+     * @param[in]  num_groups Number of groups. Must be greater than 1 and the number of channels of the tensors must be a multiple of the number of groups.
+     *
+     * @return a status
+     */
+    static Status validate(const ITensorInfo *input, const ITensorInfo *output, unsigned int num_groups);
+
+    // Inherited methods overridden:
+    void run(const Window &window, const ThreadInfo &info) override;
+
+private:
+    const ITensor *_input;
+    ITensor       *_output;
+    unsigned int   _num_groups;
+};
+} // namespace arm_compute
+#endif /* __ARM_COMPUTE_NECHANNELSHUFFLELAYERKERNEL_H__ */
diff --git a/arm_compute/core/Validate.h b/arm_compute/core/Validate.h
index 918c8e5fc3..dab4221a3b 100644
--- a/arm_compute/core/Validate.h
+++ b/arm_compute/core/Validate.h
@@ -693,6 +693,58 @@ inline arm_compute::Status error_on_data_type_not_in(const char *function, const
 #define ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_NOT_IN(t, ...) \
     ARM_COMPUTE_RETURN_ON_ERROR(::arm_compute::error_on_data_type_not_in(__func__, __FILE__, __LINE__, t, __VA_ARGS__))
 
+/** Return an error if the data layout of the passed tensor info does not match any of the data layouts provided.
+ *
+ * @param[in] function    Function in which the error occurred.
+ * @param[in] file        Name of the file where the error occurred.
+ * @param[in] line        Line on which the error occurred.
+ * @param[in] tensor_info Tensor info to validate.
+ * @param[in] dl          First data layout allowed.
+ * @param[in] dls         (Optional) Further allowed data layouts.
+ *
+ * @return Status
+ */
+template <typename T, typename... Ts>
+inline arm_compute::Status error_on_data_layout_not_in(const char *function, const char *file, const int line,
+                                                       const ITensorInfo *tensor_info, T &&dl, Ts &&... dls)
+{
+    ARM_COMPUTE_RETURN_ERROR_ON_LOC(tensor_info == nullptr, function, file, line);
+
+    const DataLayout &tensor_dl = tensor_info->data_layout(); //NOLINT
+    ARM_COMPUTE_RETURN_ERROR_ON_LOC(tensor_dl == DataLayout::UNKNOWN, function, file, line);
+
+    const std::array<T, sizeof...(Ts)> dls_array{ { std::forward<Ts>(dls)... } };
+    ARM_COMPUTE_RETURN_ERROR_ON_LOC_MSG(tensor_dl != dl && std::none_of(dls_array.begin(), dls_array.end(), [&](const T & l)
+    {
+        return l == tensor_dl;
+    }),
+    function, file, line, "ITensor data layout %s not supported by this kernel", string_from_data_layout(tensor_dl).c_str());
+    return arm_compute::Status{};
+}
+/** Return an error if the data layout of the passed tensor does not match any of the data layout provided.
+ *
+ * @param[in] function Function in which the error occurred.
+ * @param[in] file     Name of the file where the error occurred.
+ * @param[in] line     Line on which the error occurred.
+ * @param[in] tensor   Tensor to validate.
+ * @param[in] dl       First data layout allowed.
+ * @param[in] dls      (Optional) Further allowed data layouts.
+ *
+ * @return Status
+ */
+template <typename T, typename... Ts>
+inline arm_compute::Status error_on_data_layout_not_in(const char *function, const char *file, const int line,
+                                                       const ITensor *tensor, T &&dl, Ts &&... dls)
+{
+    ARM_COMPUTE_RETURN_ERROR_ON_LOC(tensor == nullptr, function, file, line);
+    ARM_COMPUTE_RETURN_ON_ERROR(::arm_compute::error_on_data_layout_not_in(function, file, line, tensor->info(), std::forward<T>(dl), std::forward<Ts>(dls)...));
+    return arm_compute::Status{};
+}
+#define ARM_COMPUTE_ERROR_ON_DATA_LAYOUT_NOT_IN(t, ...) \
+    ARM_COMPUTE_ERROR_THROW_ON(::arm_compute::error_on_data_layout_not_in(__func__, __FILE__, __LINE__, t, __VA_ARGS__))
+#define ARM_COMPUTE_RETURN_ERROR_ON_DATA_LAYOUT_NOT_IN(t, ...) \
+    ARM_COMPUTE_RETURN_ON_ERROR(::arm_compute::error_on_data_layout_not_in(__func__, __FILE__, __LINE__, t, __VA_ARGS__))
+
 /** Return an error if the data type or the number of channels of the passed tensor info does not match any of the data types and number of channels provided.
  *
  * @param[in] function     Function in which the error occurred.
diff --git a/arm_compute/runtime/NEON/NEFunctions.h b/arm_compute/runtime/NEON/NEFunctions.h
index 57bd5859f9..555d7d7551 100644
--- a/arm_compute/runtime/NEON/NEFunctions.h
+++ b/arm_compute/runtime/NEON/NEFunctions.h
@@ -39,6 +39,7 @@
 #include "arm_compute/runtime/NEON/functions/NECannyEdge.h"
 #include "arm_compute/runtime/NEON/functions/NEChannelCombine.h"
 #include "arm_compute/runtime/NEON/functions/NEChannelExtract.h"
+#include "arm_compute/runtime/NEON/functions/NEChannelShuffleLayer.h"
 #include "arm_compute/runtime/NEON/functions/NECol2Im.h"
 #include "arm_compute/runtime/NEON/functions/NEColorConvert.h"
 #include "arm_compute/runtime/NEON/functions/NEConcatenateLayer.h"
diff --git a/arm_compute/runtime/NEON/functions/NEChannelShuffleLayer.h b/arm_compute/runtime/NEON/functions/NEChannelShuffleLayer.h
new file mode 100644
index 0000000000..20fe483604
--- /dev/null
+++ b/arm_compute/runtime/NEON/functions/NEChannelShuffleLayer.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_NECHANNELSHUFFLELAYER_H__
+#define __ARM_COMPUTE_NECHANNELSHUFFLELAYER_H__
+
+#include "arm_compute/runtime/NEON/INESimpleFunction.h"
+
+namespace arm_compute
+{
+// Forward declarations
+class ITensor;
+
+/** Basic function to run @ref NEChannelShuffleLayerKernel
+ *
+ * @note The function performs a channel shuffle operation on the input tensor. Given NCHW tensor with group G, it will
+ * first divide the channels into G groups, C = (G * C'), and perform a transpose of the channel, which gives C = (C' * G).
+ * for more details see: https://arxiv.org/pdf/1707.01083.pdf
+ */
+class NEChannelShuffleLayer : public INESimpleFunction
+{
+public:
+    /** Initialize the function
+     *
+     * @param[in]  input      Input tensor. Data types supported: U8/S8/QASYMM8/U16/S16/F16/U32/S32/F32
+     * @param[out] output     Output tensor. Data type supported: Same as @p input
+     * @param[in]  num_groups Number of groups. Must be greater than 1 and the number of channels of the tensors must be a multiple of the number of groups.
+     */
+    void configure(const ITensor *input, ITensor *output, unsigned int num_groups);
+    /** Static function to check if given info will lead to a valid configuration of @ref NEChannelShuffleLayer
+     *
+     * @param[in]  input      Input tensor. Data types supported: U8/S8/QASYMM8/U16/S16/F16/U32/S32/F32
+     * @param[out] output     Output tensor. Data type supported: Same as @p input
+     * @param[in]  num_groups Number of groups. Must be greater than 1 and the number of channels of the tensors must be a multiple of the number of groups.
+     *
+     * @return a status
+     */
+    static Status validate(const ITensorInfo *input, const ITensorInfo *output, unsigned int num_groups);
+};
+} // namespace arm_compute
+#endif /* __ARM_COMPUTE_NECHANNELSHUFFLELAYER_H__ */
diff --git a/examples/graph_cl_shufflenet.cpp b/examples/graph_shufflenet.cpp
index 0db34d6123..af8d11beb9 100644
--- a/examples/graph_cl_shufflenet.cpp
+++ b/examples/graph_shufflenet.cpp
@@ -62,7 +62,7 @@ public:
 
         // Checks
         ARM_COMPUTE_EXIT_ON_MSG(arm_compute::is_data_type_quantized_asymmetric(common_params.data_type), "QASYMM8 not supported for this graph");
-        ARM_COMPUTE_EXIT_ON_MSG(common_params.target != Target::CL, "Only CL backend is supported for this network");
+        ARM_COMPUTE_EXIT_ON_MSG(common_params.data_type == DataType::F16 && common_params.target == Target::NEON, "F16 NEON not supported for this graph");
 
         // Print parameter values
         std::cout << common_params << std::endl;
diff --git a/src/core/NEON/kernels/NEChannelShuffleLayerKernel.cpp b/src/core/NEON/kernels/NEChannelShuffleLayerKernel.cpp
new file mode 100644
index 0000000000..f8217d3505
--- /dev/null
+++ b/src/core/NEON/kernels/NEChannelShuffleLayerKernel.cpp
@@ -0,0 +1,193 @@
+/*
+ * 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/NEON/kernels/NEChannelShuffleLayerKernel.h"
+
+#include "arm_compute/core/CPP/Validate.h"
+#include "arm_compute/core/Error.h"
+#include "arm_compute/core/Helpers.h"
+#include "arm_compute/core/ITensor.h"
+#include "arm_compute/core/TensorInfo.h"
+#include "arm_compute/core/Utils.h"
+#include "arm_compute/core/Validate.h"
+#include "arm_compute/core/Window.h"
+
+namespace arm_compute
+{
+namespace
+{
+Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, unsigned int num_groups)
+{
+    ARM_COMPUTE_RETURN_ERROR_ON_CPU_F16_UNSUPPORTED(input);
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input,
+                                                         1,
+                                                         DataType::U8, DataType::S8, DataType::QASYMM8,
+                                                         DataType::U16, DataType::S16,
+                                                         DataType::U32, DataType::S32,
+                                                         DataType::F16, DataType::F32);
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_LAYOUT_NOT_IN(input, DataLayout::NCHW, DataLayout::NHWC);
+
+    const unsigned int channels = input->dimension(get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::CHANNEL));
+
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG(num_groups < 2, "Channel shuffling with less than 2 groups would be inefficient");
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG(num_groups == channels, "Channel shuffling with same number of groups as number of channels would be inefficient");
+    ARM_COMPUTE_RETURN_ERROR_ON(num_groups > channels); // There cannot be more groups than channels
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG((channels % num_groups) != 0, "The number of channels must be a multiple of the number of groups");
+
+    // Checks performed when output is configured
+    if(output->total_size() != 0)
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input, output);
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_LAYOUT(input, output);
+    }
+
+    return Status{};
+}
+void channel_shuffle_nhwc(const ITensor *input, ITensor *output, unsigned int num_groups, const Window &window)
+{
+    const DataLayout   data_layout = input->info()->data_layout();
+    const unsigned int channel_idx = get_data_layout_dimension_index(data_layout, DataLayoutDimension::CHANNEL);
+
+    const size_t       element_size = input->info()->element_size();
+    const unsigned int K            = input->info()->dimension(channel_idx) / num_groups;
+    const float        rK           = 1.f / K;
+
+    Iterator in(input, window);
+
+    execute_window_loop(window, [&](const Coordinates & id)
+    {
+        // Shuffle channel
+        const unsigned int curr_channel = id.x();
+        const unsigned int group_id     = curr_channel * rK;
+        const unsigned int r            = group_id * K;
+        const unsigned int channel_id   = curr_channel - r;
+
+        // Calculate output coordinates
+        Coordinates out_coords = id;
+        out_coords.set(Window::DimX, channel_id * num_groups + group_id);
+        std::copy_n(in.ptr(), element_size, output->ptr_to_element(out_coords));
+    },
+    in);
+}
+void channel_shuffle_nchw(const ITensor *input, ITensor *output, unsigned int num_groups, const Window &window)
+{
+    Window win = window;
+    win.set(Window::DimX, Window::Dimension(0, 1, 1));
+    win.set(Window::DimY, Window::Dimension(0, 1, 1));
+
+    const DataLayout   data_layout = input->info()->data_layout();
+    const unsigned int width_idx   = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH);
+    const unsigned int channel_idx = get_data_layout_dimension_index(data_layout, DataLayoutDimension::CHANNEL);
+
+    const unsigned int height          = input->info()->tensor_shape().y();
+    const size_t       input_stride_y  = input->info()->strides_in_bytes().y();
+    const size_t       output_stride_y = output->info()->strides_in_bytes().y();
+    const size_t       row_size        = input->info()->dimension(width_idx) * input->info()->element_size();
+
+    const unsigned int K  = input->info()->dimension(channel_idx) / num_groups;
+    const float        rK = 1.f / K;
+
+    Iterator in(input, win);
+
+    execute_window_loop(win, [&](const Coordinates & id)
+    {
+        // Shuffle channel
+        const unsigned int curr_channel = id.z();
+        const unsigned int group_id     = curr_channel * rK;
+        const unsigned int r            = group_id * K;
+        const unsigned int channel_id   = curr_channel - r;
+
+        // Calculate output coordinates
+        Coordinates out_coords = id;
+        out_coords.set(Window::DimZ, channel_id * num_groups + group_id);
+        const uint8_t *input_ptr  = in.ptr();
+        uint8_t       *output_ptr = output->ptr_to_element(out_coords);
+
+        // Copy plane
+        for(unsigned int y = 0; y < height; ++y)
+        {
+            std::copy_n(input_ptr, row_size, output_ptr);
+            input_ptr += input_stride_y;
+            output_ptr += output_stride_y;
+        }
+    },
+    in);
+}
+} // namespace
+
+NEChannelShuffleLayerKernel::NEChannelShuffleLayerKernel()
+    : _input(nullptr), _output(nullptr), _num_groups()
+{
+}
+
+void NEChannelShuffleLayerKernel::configure(const ITensor *input, ITensor *output, unsigned int num_groups)
+{
+    ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
+
+    // Output tensor auto initialization if not yet initialized
+    auto_init_if_empty(*output->info(), *input->info()->clone());
+
+    _input      = input;
+    _output     = output;
+    _num_groups = num_groups;
+
+    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info(), num_groups));
+
+    // Configure kernel window
+    Window win = calculate_max_window(*input->info(), Steps());
+
+    // The NEChannelShuffleLayerKernel doesn't need padding so update_window_and_padding() can be skipped
+    Coordinates coord;
+    coord.set_num_dimensions(output->info()->num_dimensions());
+    output->info()->set_valid_region(ValidRegion(coord, output->info()->tensor_shape()));
+
+    INEKernel::configure(win);
+}
+
+Status NEChannelShuffleLayerKernel::validate(const ITensorInfo *input, const ITensorInfo *output, unsigned int num_groups)
+{
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output, num_groups));
+    return Status{};
+}
+
+void NEChannelShuffleLayerKernel::run(const Window &window, const ThreadInfo &info)
+{
+    ARM_COMPUTE_UNUSED(info);
+    ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
+    ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(IKernel::window(), window);
+
+    switch(_input->info()->data_layout())
+    {
+        case DataLayout::NHWC:
+            channel_shuffle_nhwc(_input, _output, _num_groups, window);
+            break;
+        case DataLayout::NCHW:
+            channel_shuffle_nchw(_input, _output, _num_groups, window);
+            break;
+        default:
+            ARM_COMPUTE_ERROR("Unsupported data layout!");
+            break;
+    }
+}
+} // namespace arm_compute
diff --git a/src/graph/backends/NEON/NEFunctionFactory.cpp b/src/graph/backends/NEON/NEFunctionFactory.cpp
index e967c1be61..f03cead2b4 100644
--- a/src/graph/backends/NEON/NEFunctionFactory.cpp
+++ b/src/graph/backends/NEON/NEFunctionFactory.cpp
@@ -189,6 +189,8 @@ std::unique_ptr<IFunction> NEFunctionFactory::create(INode *node, GraphContext &
             return detail::create_activation_layer<NEActivationLayer, NETargetInfo>(*polymorphic_downcast<ActivationLayerNode *>(node));
         case NodeType::BatchNormalizationLayer:
             return detail::create_batch_normalization_layer<NEBatchNormalizationLayer, NETargetInfo>(*polymorphic_downcast<BatchNormalizationLayerNode *>(node));
+        case NodeType::ChannelShuffleLayer:
+            return detail::create_channel_shuffle_layer<NEChannelShuffleLayer, NETargetInfo>(*polymorphic_downcast<ChannelShuffleLayerNode *>(node));
         case NodeType::ConvolutionLayer:
             return detail::create_convolution_layer<NEConvolutionLayerFunctions, NETargetInfo>(*polymorphic_downcast<ConvolutionLayerNode *>(node), ctx);
         case NodeType::DeconvolutionLayer:
diff --git a/src/graph/backends/NEON/NENodeValidator.cpp b/src/graph/backends/NEON/NENodeValidator.cpp
index 06023f0714..325b943e81 100644
--- a/src/graph/backends/NEON/NENodeValidator.cpp
+++ b/src/graph/backends/NEON/NENodeValidator.cpp
@@ -48,7 +48,7 @@ Status NENodeValidator::validate(INode *node)
     switch(type)
     {
         case NodeType::ChannelShuffleLayer:
-            return ARM_COMPUTE_CREATE_ERROR(arm_compute::ErrorCode::RUNTIME_ERROR, "Unsupported operation : ChannelShuffleLayer");
+            return detail::validate_channel_shuffle_layer<NEChannelShuffleLayer>(*polymorphic_downcast<ChannelShuffleLayerNode *>(node));
         case NodeType::ConvolutionLayer:
             return detail::validate_convolution_layer<NEConvolutionLayer,
                    NEDirectConvolutionLayer,
diff --git a/src/runtime/NEON/functions/NEChannelShuffleLayer.cpp b/src/runtime/NEON/functions/NEChannelShuffleLayer.cpp
new file mode 100644
index 0000000000..485abfe269
--- /dev/null
+++ b/src/runtime/NEON/functions/NEChannelShuffleLayer.cpp
@@ -0,0 +1,43 @@
+/*
+ * 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/runtime/NEON/functions/NEChannelShuffleLayer.h"
+
+#include "arm_compute/core/NEON/kernels/NEChannelShuffleLayerKernel.h"
+#include "arm_compute/core/Types.h"
+#include "support/ToolchainSupport.h"
+
+namespace arm_compute
+{
+void NEChannelShuffleLayer::configure(const ITensor *input, ITensor *output, unsigned int num_groups)
+{
+    auto k = arm_compute::support::cpp14::make_unique<NEChannelShuffleLayerKernel>();
+    k->configure(input, output, num_groups);
+    _kernel = std::move(k);
+}
+
+Status NEChannelShuffleLayer::validate(const ITensorInfo *input, const ITensorInfo *output, unsigned int num_groups)
+{
+    return NEChannelShuffleLayerKernel::validate(input, output, num_groups);
+}
+} // namespace arm_compute
diff --git a/tests/validation/NEON/ChannelShuffle.cpp b/tests/validation/NEON/ChannelShuffle.cpp
new file mode 100644
index 0000000000..474fb49d29
--- /dev/null
+++ b/tests/validation/NEON/ChannelShuffle.cpp
@@ -0,0 +1,132 @@
+/*
+ * 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/Types.h"
+#include "arm_compute/runtime/NEON/functions/NEChannelShuffleLayer.h"
+#include "arm_compute/runtime/Tensor.h"
+#include "arm_compute/runtime/TensorAllocator.h"
+#include "tests/NEON/Accessor.h"
+#include "tests/datasets/ChannelShuffleLayerDataset.h"
+#include "tests/datasets/ShapeDatasets.h"
+#include "tests/framework/Asserts.h"
+#include "tests/framework/Macros.h"
+#include "tests/framework/datasets/Datasets.h"
+#include "tests/validation/Validation.h"
+#include "tests/validation/fixtures/ChannelShuffleLayerFixture.h"
+
+namespace arm_compute
+{
+namespace test
+{
+namespace validation
+{
+TEST_SUITE(NEON)
+TEST_SUITE(ChannelShuffle)
+
+DATA_TEST_CASE(Configuration,
+               framework::DatasetMode::ALL,
+               combine(datasets::SmallRandomChannelShuffleLayerDataset(),
+                       framework::dataset::make("DataType", { DataType::S8, DataType::U8, DataType::S16, DataType::U16, DataType::U32, DataType::S32, DataType::F32 })),
+               shape, num_groups, data_type)
+{
+    // Create tensors
+    Tensor ref_src = create_tensor<Tensor>(shape, data_type);
+    Tensor dst     = create_tensor<Tensor>(shape, data_type);
+
+    // Create and Configure function
+    NEChannelShuffleLayer channel_shuffle_func;
+    channel_shuffle_func.configure(&ref_src, &dst, num_groups);
+
+    // Validate valid region
+    const ValidRegion valid_region = shape_to_valid_region(shape);
+    validate(dst.info()->valid_region(), valid_region);
+}
+
+template <typename T>
+using NEChannelShuffleLayerFixture = ChannelShuffleLayerValidationFixture<Tensor, Accessor, NEChannelShuffleLayer, T>;
+
+TEST_SUITE(U8)
+FIXTURE_DATA_TEST_CASE(RunSmall, NEChannelShuffleLayerFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallRandomChannelShuffleLayerDataset(),
+                                                                                                                   framework::dataset::make("DataType", DataType::U8)),
+                                                                                                                   framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })))
+{
+    // Validate output
+    validate(Accessor(_target), _reference);
+}
+FIXTURE_DATA_TEST_CASE(RunLarge, NEChannelShuffleLayerFixture<uint8_t>, framework::DatasetMode::NIGHTLY, combine(combine(datasets::LargeRandomChannelShuffleLayerDataset(),
+                                                                                                                 framework::dataset::make("DataType",
+                                                                                                                         DataType::U8)),
+                                                                                                                 framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })))
+{
+    // Validate output
+    validate(Accessor(_target), _reference);
+}
+TEST_SUITE_END() // U8
+
+TEST_SUITE(Float)
+#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
+TEST_SUITE(FP16)
+FIXTURE_DATA_TEST_CASE(RunSmall, NEChannelShuffleLayerFixture<half>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallRandomChannelShuffleLayerDataset(),
+                                                                                                                        framework::dataset::make("DataType",
+                                                                                                                                DataType::F16)),
+                                                                                                                framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })))
+{
+    // Validate output
+    validate(Accessor(_target), _reference);
+}
+FIXTURE_DATA_TEST_CASE(RunLarge, NEChannelShuffleLayerFixture<half>, framework::DatasetMode::NIGHTLY, combine(combine(datasets::LargeRandomChannelShuffleLayerDataset(),
+                                                                                                                      framework::dataset::make("DataType",
+                                                                                                                              DataType::F16)),
+                                                                                                              framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })))
+{
+    // Validate output
+    validate(Accessor(_target), _reference);
+}
+TEST_SUITE_END() // FP16
+#endif           /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */
+
+TEST_SUITE(FP32)
+FIXTURE_DATA_TEST_CASE(RunSmall, NEChannelShuffleLayerFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallRandomChannelShuffleLayerDataset(),
+                                                                                                                 framework::dataset::make("DataType",
+                                                                                                                         DataType::F32)),
+                                                                                                                 framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })))
+{
+    // Validate output
+    validate(Accessor(_target), _reference);
+}
+FIXTURE_DATA_TEST_CASE(RunLarge, NEChannelShuffleLayerFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(datasets::LargeRandomChannelShuffleLayerDataset(),
+                                                                                                                       framework::dataset::make("DataType",
+                                                                                                                               DataType::F32)),
+                                                                                                               framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })))
+{
+    // Validate output
+    validate(Accessor(_target), _reference);
+}
+TEST_SUITE_END() // FP32
+TEST_SUITE_END() // Float
+
+TEST_SUITE_END() // ChannelShuffle
+TEST_SUITE_END() // NEON
+} // namespace validation
+} // namespace test
+} // namespace arm_compute