COMPMID-1364: Add support for NHWC in NEDepthConcatenateLayer

Change-Id: I4f8e46d1c79afa9284f2c6dc00383c453a8e7bd5
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/140165
Reviewed-by: Giorgio Arena <giorgio.arena@arm.com>
Reviewed-by: Pablo Tello <pablo.tello@arm.com>
Tested-by: Jenkins <bsgcomp@arm.com>

7 files changed, 402 insertions, 51 deletions

diff --git a/src/core/NEON/kernels/NEDepthConcatenateLayerKernel.cpp b/src/core/NEON/kernels/NEDepthConcatenateLayerKernel.cpp
index 38443ca4a8..1b937b5be8 100644
--- a/src/core/NEON/kernels/NEDepthConcatenateLayerKernel.cpp
+++ b/src/core/NEON/kernels/NEDepthConcatenateLayerKernel.cpp
@@ -28,37 +28,18 @@
 #include "arm_compute/core/IAccessWindow.h"
 #include "arm_compute/core/ITensor.h"
 #include "arm_compute/core/NEON/NEFixedPoint.h"
+#include "arm_compute/core/NEON/wrapper/wrapper.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"
 
-#include <arm_neon.h>
 #include <cstdint>
 
 using namespace arm_compute;
 
 namespace
 {
-// Overloads of 128-bit vector loads
-uint16x8_t loadq(const uint16_t *ptr)
-{
-    return vld1q_u16(ptr);
-}
-uint32x4_t loadq(const uint32_t *ptr)
-{
-    return vld1q_u32(ptr);
-}
-// Overloads of 128-bit vector stores
-void storeq(uint16_t *ptr, uint16x8_t val)
-{
-    return vst1q_u16(ptr, val);
-}
-void storeq(uint32_t *ptr, uint32x4_t val)
-{
-    return vst1q_u32(ptr, val);
-}
-
 template <typename T>
 void depth_concat(const ITensor *in, ITensor *out, std::pair<int, int> start_xy, int depth_offset, const Window &window)
 {
@@ -81,10 +62,54 @@ void depth_concat(const ITensor *in, ITensor *out, std::pair<int, int> start_xy,
         const auto in_ptr  = reinterpret_cast<const T *>(input_ptr + input.offset());
         const auto out_ptr = reinterpret_cast<T *>(output_ptr + output.offset());
 
-        storeq(out_ptr, loadq(in_ptr));
+        wrapper::vstore(out_ptr, wrapper::vloadq(in_ptr));
     },
     input, output);
 }
+
+std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, unsigned int depth_offset, ITensorInfo *output)
+{
+    ARM_COMPUTE_UNUSED(depth_offset);
+
+    // Configure kernel window
+    const int left_right = (output->dimension(0) - input->dimension(0)) / 2;
+    const int top_bottom = (output->dimension(1) - input->dimension(1)) / 2;
+
+    const unsigned int num_elems_processed_per_iteration = 16 / input->element_size();
+    const unsigned int num_elems_read_per_iteration      = 16 / input->element_size();
+    const unsigned int num_rows_read_per_iteration       = 1;
+
+    // The window needs to be based on input as we copy all the depths of input
+    Window win = calculate_max_window(*output, Steps(num_elems_processed_per_iteration));
+    win.set(Window::DimZ, Window::Dimension(0, input->tensor_shape().z(), 1));
+
+    AccessWindowRectangle  input_access(input, -left_right, -top_bottom, num_elems_read_per_iteration, num_rows_read_per_iteration);
+    AccessWindowHorizontal output_access(output, 0, num_elems_processed_per_iteration);
+    bool                   window_changed = update_window_and_padding(win, input_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);
+}
+
+Status validate_arguments(const ITensorInfo *input, unsigned int depth_offset, const ITensorInfo *output)
+{
+    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output);
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::F16, DataType::F32);
+    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
+
+    ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(2) + depth_offset > output->dimension(2));
+    ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(0) > output->dimension(0));
+    ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(1) > output->dimension(1));
+    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(3, input, output);
+
+    // The gaps between the two lowest dimensions of input and output need to be divisible by 2
+    // Otherwise it is not clear how the padding should be added onto the input tensor
+    ARM_COMPUTE_RETURN_ERROR_ON((output->dimension(0) - input->dimension(0)) % 2);
+    ARM_COMPUTE_RETURN_ERROR_ON((output->dimension(1) - input->dimension(1)) % 2);
+
+    return Status{};
+}
 } // namespace
 
 NEDepthConcatenateLayerKernel::NEDepthConcatenateLayerKernel()
@@ -99,17 +124,8 @@ BorderSize NEDepthConcatenateLayerKernel::border_size() const
 
 void NEDepthConcatenateLayerKernel::configure(const ITensor *input, unsigned int depth_offset, ITensor *output)
 {
-    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(input->info()->dimension(2) + depth_offset > output->info()->dimension(2));
-    ARM_COMPUTE_ERROR_ON(input->info()->dimension(0) > output->info()->dimension(0));
-    ARM_COMPUTE_ERROR_ON(input->info()->dimension(1) > output->info()->dimension(1));
-    ARM_COMPUTE_ERROR_ON_MISMATCHING_SHAPES(3, input, output);
-
-    // The gaps between the two lowest dimensions of input and output need to be divisible by 2
-    // Otherwise it is not clear how the padding should be added onto the input tensor
-    ARM_COMPUTE_ERROR_ON((output->info()->dimension(0) - input->info()->dimension(0)) % 2);
-    ARM_COMPUTE_ERROR_ON((output->info()->dimension(1) - input->info()->dimension(1)) % 2);
+    ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
+    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), depth_offset, output->info()));
 
     _func         = nullptr;
     _input        = input;
@@ -120,6 +136,9 @@ void NEDepthConcatenateLayerKernel::configure(const ITensor *input, unsigned int
 
     switch(input->info()->data_type())
     {
+        case DataType::QASYMM8:
+            _func = &depth_concat<uint8_t>;
+            break;
         case DataType::F16:
             _func = &depth_concat<uint16_t>;
             break;
@@ -130,20 +149,20 @@ void NEDepthConcatenateLayerKernel::configure(const ITensor *input, unsigned int
             ARM_COMPUTE_ERROR("Unsupported data type.");
     }
 
-    const unsigned int num_elems_processed_per_iteration = 16 / input->info()->element_size();
-    const unsigned int num_elems_read_per_iteration      = 16 / input->info()->element_size();
-    const unsigned int num_rows_read_per_iteration       = 1;
+    // Configure kernel window
+    auto win_config = validate_and_configure_window(input->info(), depth_offset, output->info());
+    ARM_COMPUTE_ERROR_THROW_ON(std::get<0>(win_config));
 
-    // The window needs to be based on input as we copy all the depths of input
-    Window win = calculate_max_window(*output->info(), Steps(num_elems_processed_per_iteration));
-    win.set(Window::DimZ, Window::Dimension(0, input->info()->tensor_shape().z(), 1));
-
-    AccessWindowRectangle  input_access(input->info(), -_left_right, -_top_bottom, num_elems_read_per_iteration, num_rows_read_per_iteration);
-    AccessWindowHorizontal output_access(output->info(), 0, num_elems_processed_per_iteration);
-    update_window_and_padding(win, input_access, output_access);
-    output_access.set_valid_region(win, ValidRegion(Coordinates(), output->info()->tensor_shape()));
+    INEKernel::configure(std::get<1>(win_config));
+}
 
-    INEKernel::configure(win);
+Status NEDepthConcatenateLayerKernel::validate(const arm_compute::ITensorInfo *input,
+                                               unsigned int                    depth_offset,
+                                               const arm_compute::ITensorInfo *output)
+{
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, depth_offset, output));
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), depth_offset, output->clone().get()).first);
+    return Status{};
 }
 
 void NEDepthConcatenateLayerKernel::run(const Window &window, const ThreadInfo &info)
diff --git a/src/core/NEON/kernels/NEWidthConcatenateLayerKernel.cpp b/src/core/NEON/kernels/NEWidthConcatenateLayerKernel.cpp
new file mode 100644
index 0000000000..1b38677991
--- /dev/null
+++ b/src/core/NEON/kernels/NEWidthConcatenateLayerKernel.cpp
@@ -0,0 +1,125 @@
+/*
+ * 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/NEWidthConcatenateLayerKernel.h"
+
+#include "arm_compute/core/Error.h"
+#include "arm_compute/core/Helpers.h"
+#include "arm_compute/core/IAccessWindow.h"
+#include "arm_compute/core/ITensor.h"
+#include "arm_compute/core/NEON/wrapper/wrapper.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"
+
+#include <cstdint>
+
+using namespace arm_compute;
+
+namespace
+{
+std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, unsigned int width_offset, ITensorInfo *output)
+{
+    const unsigned int num_elems_processed_per_iteration = 16 / output->element_size();
+
+    // The window needs to be based on input as we copy all the widths of input
+    Window                 win = calculate_max_window(*input, Steps(num_elems_processed_per_iteration));
+    AccessWindowHorizontal input_access(input, 0, num_elems_processed_per_iteration);
+    AccessWindowHorizontal output_access(output, width_offset, num_elems_processed_per_iteration);
+    bool                   window_changed = update_window_and_padding(win, input_access, output_access);
+
+    Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{};
+    return std::make_pair(err, win);
+}
+
+Status validate_arguments(const ITensorInfo *input, unsigned int width_offset, const ITensorInfo *output)
+{
+    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output);
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1,
+                                                         DataType::U8, DataType::S8, DataType::QASYMM8,
+                                                         DataType::U16, DataType::S16, DataType::F16,
+                                                         DataType::U32, DataType::F32);
+    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
+    ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(0) + width_offset > output->dimension(0));
+
+    for(size_t i = 1; i < Coordinates::num_max_dimensions; ++i)
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(i) != output->dimension(i));
+    }
+    ARM_COMPUTE_RETURN_ERROR_ON(input->num_dimensions() > 3);
+
+    return Status{};
+}
+} // namespace
+
+NEWidthConcatenateLayerKernel::NEWidthConcatenateLayerKernel()
+    : _input(nullptr), _output(nullptr), _width_offset(0)
+{
+}
+
+void NEWidthConcatenateLayerKernel::configure(const ITensor *input, unsigned int width_offset, ITensor *output)
+{
+    ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
+    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), width_offset, output->info()));
+
+    _input        = input;
+    _output       = output;
+    _width_offset = width_offset;
+
+    // Configure kernel window
+    auto win_config = validate_and_configure_window(input->info(), width_offset, output->info());
+    ARM_COMPUTE_ERROR_THROW_ON(std::get<0>(win_config));
+
+    INEKernel::configure(std::get<1>(win_config));
+}
+
+Status NEWidthConcatenateLayerKernel::validate(const ITensorInfo *input, unsigned int width_offset, const ITensorInfo *output)
+{
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, width_offset, output));
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), width_offset, output->clone().get()).first);
+    return Status{};
+}
+
+void NEWidthConcatenateLayerKernel::run(const Window &window, const ThreadInfo &info)
+{
+    ARM_COMPUTE_UNUSED(info);
+    ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
+    ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window);
+
+    // Offset output pointer to the correct position
+    uint8_t *output_ptr = _output->buffer() + _output->info()->offset_first_element_in_bytes() + _width_offset * _output->info()->strides_in_bytes()[0];
+
+    // Create iterators
+    Iterator input(_input, window);
+    Iterator output(_output, window);
+
+    execute_window_loop(window, [&](const Coordinates & id)
+    {
+        const auto in_ptr  = input.ptr();
+        const auto out_ptr = output_ptr + output.offset();
+
+        wrapper::vstore(out_ptr, wrapper::vloadq(in_ptr));
+    },
+    input, output);
+}
diff --git a/src/runtime/CL/functions/CLConcatenateLayer.cpp b/src/runtime/CL/functions/CLConcatenateLayer.cpp
index f4bc1ff4ac..018c674c83 100644
--- a/src/runtime/CL/functions/CLConcatenateLayer.cpp
+++ b/src/runtime/CL/functions/CLConcatenateLayer.cpp
@@ -39,7 +39,7 @@ CLConcatenateLayer::CLConcatenateLayer()
 {
 }
 
-void CLConcatenateLayer::configure(std::vector<ICLTensor *> inputs_vector, ICLTensor *output, DataLayoutDimension axis)
+void CLConcatenateLayer::configure(const std::vector<ICLTensor *> &inputs_vector, ICLTensor *output, DataLayoutDimension axis)
 {
     ARM_COMPUTE_ERROR_ON(output == nullptr);
 
@@ -48,14 +48,14 @@ void CLConcatenateLayer::configure(std::vector<ICLTensor *> inputs_vector, ICLTe
         case 0:
         {
             auto func = support::cpp14::make_unique<CLWidthConcatenateLayer>();
-            func->configure(std::move(inputs_vector), output);
+            func->configure(inputs_vector, output);
             _concat_function = std::move(func);
             break;
         }
         case 2:
         {
             auto func = support::cpp14::make_unique<CLDepthConcatenateLayer>();
-            func->configure(std::move(inputs_vector), output);
+            func->configure(inputs_vector, output);
             _concat_function = std::move(func);
             break;
         }
diff --git a/src/runtime/CL/functions/CLDepthConcatenateLayer.cpp b/src/runtime/CL/functions/CLDepthConcatenateLayer.cpp
index 174be94410..b5e8fd96d0 100644
--- a/src/runtime/CL/functions/CLDepthConcatenateLayer.cpp
+++ b/src/runtime/CL/functions/CLDepthConcatenateLayer.cpp
@@ -43,7 +43,7 @@ CLDepthConcatenateLayer::CLDepthConcatenateLayer() // NOLINT
 {
 }
 
-void CLDepthConcatenateLayer::configure(std::vector<ICLTensor *> inputs_vector, ICLTensor *output) // NOLINT
+void CLDepthConcatenateLayer::configure(const std::vector<ICLTensor *> &inputs_vector, ICLTensor *output) // NOLINT
 {
     _num_inputs = inputs_vector.size();
 
diff --git a/src/runtime/NEON/functions/NEConcatenateLayer.cpp b/src/runtime/NEON/functions/NEConcatenateLayer.cpp
new file mode 100644
index 0000000000..21ab47d3fe
--- /dev/null
+++ b/src/runtime/NEON/functions/NEConcatenateLayer.cpp
@@ -0,0 +1,90 @@
+/*
+ * 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/NEConcatenateLayer.h"
+
+#include "arm_compute/runtime/NEON/functions/NEDepthConcatenateLayer.h"
+#include "arm_compute/runtime/NEON/functions/NEWidthConcatenateLayer.h"
+
+#include "arm_compute/core/Error.h"
+#include "arm_compute/core/ITensor.h"
+#include "arm_compute/core/TensorInfo.h"
+#include "arm_compute/core/Types.h"
+#include "support/ToolchainSupport.h"
+
+namespace arm_compute
+{
+NEConcatenateLayer::NEConcatenateLayer()
+    : _concat_function(nullptr)
+{
+}
+
+void NEConcatenateLayer::configure(const std::vector<ITensor *> &inputs_vector, ITensor *output, DataLayoutDimension axis)
+{
+    ARM_COMPUTE_ERROR_ON(output == nullptr);
+
+    switch(get_data_layout_dimension_index(output->info()->data_layout(), axis))
+    {
+        case 0:
+        {
+            auto func = support::cpp14::make_unique<NEWidthConcatenateLayer>();
+            func->configure(inputs_vector, output);
+            _concat_function = std::move(func);
+            break;
+        }
+        case 2:
+        {
+            auto func = support::cpp14::make_unique<NEDepthConcatenateLayer>();
+            func->configure(inputs_vector, output);
+            _concat_function = std::move(func);
+            break;
+        }
+        default:
+            ARM_COMPUTE_ERROR("Concatenation is supported across width and depth only!");
+    }
+}
+
+Status NEConcatenateLayer::validate(const std::vector<ITensorInfo *> &inputs_vector, const ITensorInfo *output, DataLayoutDimension axis)
+{
+    ARM_COMPUTE_RETURN_ERROR_ON(output == nullptr);
+
+    switch(get_data_layout_dimension_index(output->data_layout(), axis))
+    {
+        case 0:
+            ARM_COMPUTE_RETURN_ON_ERROR(NEWidthConcatenateLayer::validate(inputs_vector, output));
+            break;
+        case 2:
+            ARM_COMPUTE_RETURN_ON_ERROR(NEDepthConcatenateLayer::validate(inputs_vector, output));
+            break;
+        default:
+            ARM_COMPUTE_RETURN_ERROR_MSG("Concatenation is supported across width and depth only!");
+    }
+    return Status{};
+}
+
+void NEConcatenateLayer::run()
+{
+    ARM_COMPUTE_ERROR_ON(_concat_function == nullptr);
+    _concat_function->run();
+}
+} // namespace arm_compute
diff --git a/src/runtime/NEON/functions/NEDepthConcatenateLayer.cpp b/src/runtime/NEON/functions/NEDepthConcatenateLayer.cpp
index cb0157664b..49db855f21 100644
--- a/src/runtime/NEON/functions/NEDepthConcatenateLayer.cpp
+++ b/src/runtime/NEON/functions/NEDepthConcatenateLayer.cpp
@@ -27,6 +27,7 @@
 #include "arm_compute/core/Helpers.h"
 #include "arm_compute/core/ITensor.h"
 #include "arm_compute/core/PixelValue.h"
+#include "arm_compute/core/TensorInfo.h"
 #include "arm_compute/core/Types.h"
 #include "arm_compute/core/utils/misc/ShapeCalculator.h"
 #include "arm_compute/runtime/NEON/NEScheduler.h"
@@ -42,10 +43,8 @@ NEDepthConcatenateLayer::NEDepthConcatenateLayer() // NOLINT
 {
 }
 
-void NEDepthConcatenateLayer::configure(std::vector<ITensor *> inputs_vector, ITensor *output) // NOLINT
+void NEDepthConcatenateLayer::configure(const std::vector<ITensor *> &inputs_vector, ITensor *output) // NOLINT
 {
-    ARM_COMPUTE_ERROR_ON(inputs_vector.size() < 2);
-
     _num_inputs             = inputs_vector.size();
     _concat_kernels_vector  = arm_compute::support::cpp14::make_unique<NEDepthConcatenateLayerKernel[]>(_num_inputs);
     _border_handlers_vector = arm_compute::support::cpp14::make_unique<NEFillBorderKernel[]>(_num_inputs);
@@ -59,6 +58,7 @@ void NEDepthConcatenateLayer::configure(std::vector<ITensor *> inputs_vector, IT
 
     // Output auto inizialitation if not yet initialized
     auto_init_if_empty(*output->info(), output_shape, 1, inputs_vector[0]->info()->data_type());
+    ARM_COMPUTE_ERROR_THROW_ON(NEDepthConcatenateLayer::validate(inputs_vector_info, output->info()));
 
     unsigned int depth_offset = 0;
     for(unsigned int i = 0; i < _num_inputs; ++i)
@@ -73,6 +73,27 @@ void NEDepthConcatenateLayer::configure(std::vector<ITensor *> inputs_vector, IT
     output->info()->set_valid_region(ValidRegion(Coordinates(), output_shape));
 }
 
+Status NEDepthConcatenateLayer::validate(const std::vector<ITensorInfo *> &inputs_vector, const ITensorInfo *output)
+{
+    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(output);
+    ARM_COMPUTE_RETURN_ERROR_ON(inputs_vector.size() < 2);
+
+    // Output auto inizialitation if not yet initialized
+    TensorInfo  tmp_output_info = *output->clone();
+    TensorShape output_shape    = arm_compute::misc::shape_calculator::calculate_depth_concatenate_shape(inputs_vector);
+    auto_init_if_empty(tmp_output_info, output_shape, 1, inputs_vector[0]->data_type());
+
+    unsigned int depth_offset = 0;
+    for(const auto &input : inputs_vector)
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input);
+        ARM_COMPUTE_RETURN_ON_ERROR(NEDepthConcatenateLayerKernel::validate(input, depth_offset, &tmp_output_info));
+        depth_offset += input->dimension(2);
+    }
+
+    return Status{};
+}
+
 void NEDepthConcatenateLayer::run()
 {
     for(unsigned i = 0; i < _num_inputs; ++i)
diff --git a/src/runtime/NEON/functions/NEWidthConcatenateLayer.cpp b/src/runtime/NEON/functions/NEWidthConcatenateLayer.cpp
new file mode 100644
index 0000000000..097605c062
--- /dev/null
+++ b/src/runtime/NEON/functions/NEWidthConcatenateLayer.cpp
@@ -0,0 +1,96 @@
+/*
+ * 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/NEWidthConcatenateLayer.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/misc/ShapeCalculator.h"
+#include "arm_compute/runtime/NEON/NEScheduler.h"
+#include "arm_compute/runtime/Tensor.h"
+#include "support/ToolchainSupport.h"
+
+using namespace arm_compute;
+
+NEWidthConcatenateLayer::NEWidthConcatenateLayer()
+    : _concat_kernels_vector(),
+      _num_inputs(0)
+{
+}
+
+Status NEWidthConcatenateLayer::validate(const std::vector<ITensorInfo *> &inputs_vector, const ITensorInfo *output)
+{
+    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(output);
+    ARM_COMPUTE_RETURN_ERROR_ON(inputs_vector.size() < 2);
+
+    // Output auto inizialitation if not yet initialized
+    TensorInfo  tmp_output_info = *output->clone();
+    TensorShape output_shape    = arm_compute::misc::shape_calculator::calculate_width_concatenate_shape(inputs_vector);
+    auto_init_if_empty(tmp_output_info, output_shape, 1, inputs_vector[0]->data_type());
+
+    unsigned int width_offset = 0;
+    for(const auto &input : inputs_vector)
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input);
+        ARM_COMPUTE_RETURN_ON_ERROR(NEWidthConcatenateLayerKernel::validate(input, width_offset, &tmp_output_info));
+        width_offset += input->dimension(0);
+    }
+
+    return Status{};
+}
+
+void NEWidthConcatenateLayer::configure(std::vector<ITensor *> inputs_vector, ITensor *output)
+{
+    _num_inputs = inputs_vector.size();
+
+    std::vector<ITensorInfo *> inputs_vector_info;
+    for(unsigned int i = 0; i < _num_inputs; i++)
+    {
+        inputs_vector_info.emplace_back(inputs_vector.at(i)->info());
+    }
+    TensorShape output_shape = arm_compute::misc::shape_calculator::calculate_width_concatenate_shape(inputs_vector);
+
+    // Output auto inizialitation if not yet initialized
+    auto_init_if_empty(*output->info(), output_shape, 1, inputs_vector[0]->info()->data_type());
+    ARM_COMPUTE_ERROR_THROW_ON(NEWidthConcatenateLayer::validate(inputs_vector_info, output->info()));
+
+    unsigned int width_offset = 0;
+
+    _concat_kernels_vector = arm_compute::support::cpp14::make_unique<NEWidthConcatenateLayerKernel[]>(_num_inputs);
+
+    for(unsigned int i = 0; i < _num_inputs; i++)
+    {
+        _concat_kernels_vector[i].configure(inputs_vector.at(i), width_offset, output);
+        width_offset += inputs_vector.at(i)->info()->dimension(0);
+    }
+}
+
+void NEWidthConcatenateLayer::run()
+{
+    for(unsigned i = 0; i < _num_inputs; i++)
+    {
+        NEScheduler::get().schedule(_concat_kernels_vector.get() + i, Window::DimY);
+    }
+}