IVGCVSW-863 Broadcast support in CL/NEON Arithmetic Add

Also, added instrumentation to support generic tensor broadcasting for
NEON and CL backends.

Change-Id: I1bc5747a286e1a4b464c209067581e103d473b9a
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/114201
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>
Tested-by: Jenkins <bsgcomp@arm.com>

31 files changed, 754 insertions, 373 deletions

diff --git a/arm_compute/core/CL/ICLKernel.h b/arm_compute/core/CL/ICLKernel.h
index a1bc3eb8d2..e660ae55a0 100644
--- a/arm_compute/core/CL/ICLKernel.h
+++ b/arm_compute/core/CL/ICLKernel.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016, 2017 ARM Limited.
+ * Copyright (c) 2016-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -41,14 +41,40 @@ class Window;
 /** Common interface for all the OpenCL kernels */
 class ICLKernel : public IKernel
 {
+private:
+    /** Returns the number of arguments enqueued per array object.
+     *
+     * @return The number of arguments enqueued per array object.
+     */
+    template <unsigned int        dimension_size>
+    constexpr static unsigned int num_arguments_per_array()
+    {
+        return num_arguments_per_tensor<dimension_size>();
+    }
+    /** Returns the number of arguments enqueued per tensor object.
+     *
+     * @return The number of arguments enqueued per tensor object.
+     */
+    template <unsigned int        dimension_size>
+    constexpr static unsigned int num_arguments_per_tensor()
+    {
+        return 2 + 2 * dimension_size;
+    }
+
 public:
     /** Constructor */
-    ICLKernel();
+    ICLKernel()
+        : _kernel(nullptr), _lws_hint(CLKernelLibrary::get().default_ndrange()), _target(GPUTarget::MIDGARD), _config_id(arm_compute::default_config_id), _max_workgroup_size(0)
+    {
+    }
     /** Returns a reference to the OpenCL kernel of this object.
      *
      * @return A reference to the OpenCL kernel of this object.
      */
-    cl::Kernel &kernel();
+    cl::Kernel &kernel()
+    {
+        return _kernel;
+    }
     /** Add the passed 1D array's parameters to the object's kernel's arguments starting from the index idx.
      *
      * @param[in,out] idx            Index at which to start adding the array's arguments. Will be incremented by the number of kernel arguments set.
@@ -58,60 +84,90 @@ public:
      * @param[in]     window         Window the kernel will be executed on.
      */
     template <typename T>
-    void add_1D_array_argument(unsigned int &idx, const ICLArray<T> *array, const Strides &strides, unsigned int num_dimensions, const Window &window);
+    void add_1D_array_argument(unsigned int &idx, const ICLArray<T> *array, const Strides &strides, unsigned int num_dimensions, const Window &window)
+    {
+        add_array_argument<T, 1>(idx, array, strides, num_dimensions, window);
+    }
     /** Add the passed 1D tensor's parameters to the object's kernel's arguments starting from the index idx.
      *
      * @param[in,out] idx    Index at which to start adding the tensor's arguments. Will be incremented by the number of kernel arguments set.
      * @param[in]     tensor Tensor to set as an argument of the object's kernel.
      * @param[in]     window Window the kernel will be executed on.
      */
-    void add_1D_tensor_argument(unsigned int &idx, const ICLTensor *tensor, const Window &window);
+    void add_1D_tensor_argument(unsigned int &idx, const ICLTensor *tensor, const Window &window)
+    {
+        add_tensor_argument<1>(idx, tensor, window);
+    }
     /** Add the passed 2D tensor's parameters to the object's kernel's arguments starting from the index idx.
      *
      * @param[in,out] idx    Index at which to start adding the tensor's arguments. Will be incremented by the number of kernel arguments set.
      * @param[in]     tensor Tensor to set as an argument of the object's kernel.
      * @param[in]     window Window the kernel will be executed on.
      */
-    void add_2D_tensor_argument(unsigned int &idx, const ICLTensor *tensor, const Window &window);
+    void add_2D_tensor_argument(unsigned int &idx, const ICLTensor *tensor, const Window &window)
+    {
+        add_tensor_argument<2>(idx, tensor, window);
+    }
     /** Add the passed 3D tensor's parameters to the object's kernel's arguments starting from the index idx.
      *
      * @param[in,out] idx    Index at which to start adding the tensor's arguments. Will be incremented by the number of kernel arguments set.
      * @param[in]     tensor Tensor to set as an argument of the object's kernel.
      * @param[in]     window Window the kernel will be executed on.
      */
-    void add_3D_tensor_argument(unsigned int &idx, const ICLTensor *tensor, const Window &window);
+    void add_3D_tensor_argument(unsigned int &idx, const ICLTensor *tensor, const Window &window)
+    {
+        add_tensor_argument<3>(idx, tensor, window);
+    }
     /** Add the passed 4D tensor's parameters to the object's kernel's arguments starting from the index idx.
      *
      * @param[in,out] idx    Index at which to start adding the tensor's arguments. Will be incremented by the number of kernel arguments set.
      * @param[in]     tensor Tensor to set as an argument of the object's kernel.
      * @param[in]     window Window the kernel will be executed on.
      */
-    void add_4D_tensor_argument(unsigned int &idx, const ICLTensor *tensor, const Window &window);
+    void add_4D_tensor_argument(unsigned int &idx, const ICLTensor *tensor, const Window &window)
+    {
+        add_tensor_argument<4>(idx, tensor, window);
+    }
     /** Returns the number of arguments enqueued per 1D array object.
      *
      * @return The number of arguments enqueues per 1D array object.
      */
-    unsigned int num_arguments_per_1D_array() const;
+    constexpr static unsigned int num_arguments_per_1D_array()
+    {
+        return num_arguments_per_array<1>();
+    }
     /** Returns the number of arguments enqueued per 1D tensor object.
      *
      * @return The number of arguments enqueues per 1D tensor object.
      */
-    unsigned int num_arguments_per_1D_tensor() const;
+    constexpr static unsigned int num_arguments_per_1D_tensor()
+    {
+        return num_arguments_per_tensor<1>();
+    }
     /** Returns the number of arguments enqueued per 2D tensor object.
      *
      * @return The number of arguments enqueues per 2D tensor object.
      */
-    unsigned int num_arguments_per_2D_tensor() const;
+    constexpr static unsigned int num_arguments_per_2D_tensor()
+    {
+        return num_arguments_per_tensor<2>();
+    }
     /** Returns the number of arguments enqueued per 3D tensor object.
      *
      * @return The number of arguments enqueues per 3D tensor object.
      */
-    unsigned int num_arguments_per_3D_tensor() const;
+    constexpr static unsigned int num_arguments_per_3D_tensor()
+    {
+        return num_arguments_per_tensor<3>();
+    }
     /** Returns the number of arguments enqueued per 4D tensor object.
      *
      * @return The number of arguments enqueues per 4D tensor object.
      */
-    unsigned int num_arguments_per_4D_tensor() const;
+    constexpr static unsigned int num_arguments_per_4D_tensor()
+    {
+        return num_arguments_per_tensor<4>();
+    }
     /** Enqueue the OpenCL kernel to process the given window  on the passed OpenCL command queue.
      *
      * @note The queue is *not* flushed by this method, and therefore the kernel will not have been executed by the time this method returns.
@@ -161,7 +217,10 @@ public:
      *
      * @param[in] target The targeted GPU architecture
      */
-    void set_target(GPUTarget target);
+    void set_target(GPUTarget target)
+    {
+        _target = target;
+    }
 
     /** Set the targeted GPU architecture according to the CL device
      *
@@ -173,7 +232,10 @@ public:
      *
      * @return The targeted GPU architecture.
      */
-    GPUTarget get_target() const;
+    GPUTarget get_target() const
+    {
+        return _target;
+    }
 
     /** Get the maximum workgroup size for the device the CLKernelLibrary uses.
      *
@@ -207,18 +269,6 @@ private:
      */
     template <unsigned int dimension_size>
     void add_tensor_argument(unsigned int &idx, const ICLTensor *tensor, const Window &window);
-    /** Returns the number of arguments enqueued per array object.
-     *
-     * @return The number of arguments enqueued per array object.
-     */
-    template <unsigned int dimension_size>
-    unsigned int           num_arguments_per_array() const;
-    /** Returns the number of arguments enqueued per tensor object.
-     *
-     * @return The number of arguments enqueued per tensor object.
-     */
-    template <unsigned int dimension_size>
-    unsigned int           num_arguments_per_tensor() const;
 
 protected:
     cl::Kernel  _kernel;             /**< OpenCL kernel to run */
@@ -246,6 +296,8 @@ void enqueue(cl::CommandQueue &queue, ICLKernel &kernel, const Window &window, c
 template <typename T, unsigned int dimension_size>
 void ICLKernel::add_array_argument(unsigned &idx, const ICLArray<T> *array, const Strides &strides, unsigned int num_dimensions, const Window &window)
 {
+    ARM_COMPUTE_ERROR_ON(array == nullptr);
+
     // Calculate offset to the start of the window
     unsigned int offset_first_element = 0;
 
@@ -269,23 +321,5 @@ void ICLKernel::add_array_argument(unsigned &idx, const ICLArray<T> *array, cons
                              "add_%dD_array_argument() is supposed to add exactly %d arguments to the kernel", dimension_size, num_arguments_per_array<dimension_size>());
     ARM_COMPUTE_UNUSED(idx_start);
 }
-
-template <typename T>
-void ICLKernel::add_1D_array_argument(unsigned int &idx, const ICLArray<T> *array, const Strides &strides, unsigned int num_dimensions, const Window &window)
-{
-    add_array_argument<T, 1>(idx, array, strides, num_dimensions, window);
-}
-
-template <unsigned int dimension_size>
-unsigned int           ICLKernel::num_arguments_per_array() const
-{
-    return num_arguments_per_tensor<dimension_size>();
-}
-
-template <unsigned int dimension_size>
-unsigned int           ICLKernel::num_arguments_per_tensor() const
-{
-    return 2 + 2 * dimension_size;
-}
 }
 #endif /*__ARM_COMPUTE_ICLKERNEL_H__ */
diff --git a/arm_compute/core/CL/kernels/CLArithmeticAdditionKernel.h b/arm_compute/core/CL/kernels/CLArithmeticAdditionKernel.h
index 96b8dc8d48..5112476aae 100644
--- a/arm_compute/core/CL/kernels/CLArithmeticAdditionKernel.h
+++ b/arm_compute/core/CL/kernels/CLArithmeticAdditionKernel.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016, 2017 ARM Limited.
+ * Copyright (c) 2016-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -72,6 +72,7 @@ public:
 
     // Inherited methods overridden:
     void run(const Window &window, cl::CommandQueue &queue) override;
+    BorderSize border_size() const override;
 
 private:
     const ICLTensor *_input1; /**< Source tensor 1 */
diff --git a/arm_compute/core/Dimensions.h b/arm_compute/core/Dimensions.h
index ae8d6c3503..58ffd7ff3c 100644
--- a/arm_compute/core/Dimensions.h
+++ b/arm_compute/core/Dimensions.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017 ARM Limited.
+ * Copyright (c) 2017-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -135,23 +135,24 @@ public:
      * @param[in] n     Number of dimensions to collapse into @p first.
      * @param[in] first Dimensions into which the following @p n are collapsed.
      */
-    void collapse(size_t n, size_t first = 0)
+    void collapse(const size_t n, const size_t first = 0)
     {
         ARM_COMPUTE_ERROR_ON(first + n > _id.size());
 
-        if(n == 0)
+        const size_t last = std::min(_num_dimensions, first + n);
+
+        if(last > (first + 1))
         {
-            return;
+            // Collapse dimensions into the first
+            _id[first] = std::accumulate(&_id[first], &_id[last], 1, std::multiplies<T>());
+            // Shift the remaining dimensions down
+            std::copy(&_id[last], &_id[_num_dimensions], &_id[first + 1]);
+            // Reduce the number of dimensions
+            const size_t old_num_dimensions = _num_dimensions;
+            _num_dimensions -= last - first - 1;
+            // Fill the now empty dimensions with zero
+            std::fill(&_id[_num_dimensions], &_id[old_num_dimensions], 0);
         }
-
-        // Collapse dimensions into the first
-        _id[first] = std::accumulate(_id.cbegin() + first, _id.cbegin() + first + n, 1, std::multiplies<T>());
-        // Shift the remaining dimensions down
-        std::copy(_id.begin() + first + n, _id.end(), _id.begin() + first + 1);
-        // Reduce the number of dimensions
-        _num_dimensions -= std::min(n, _num_dimensions) - 1;
-        // Fill the now empty dimensions with zero
-        std::fill(_id.begin() + _num_dimensions, _id.end(), 0);
     }
 
     /** Collapse dimensions starting from a given point
diff --git a/arm_compute/core/Helpers.inl b/arm_compute/core/Helpers.inl
index 6d0f8b0104..8b86c22676 100644
--- a/arm_compute/core/Helpers.inl
+++ b/arm_compute/core/Helpers.inl
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016, 2018 ARM Limited.
+ * Copyright (c) 2016-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -123,6 +123,11 @@ inline void execute_window_loop(const Window &w, L &&lambda_function, Ts &&... i
 {
     w.validate();
 
+    for(unsigned int i = 0; i < Coordinates::num_max_dimensions; ++i)
+    {
+        ARM_COMPUTE_ERROR_ON(w[i].step() == 0);
+    }
+
     Coordinates id;
     ForEachDimension<Coordinates::num_max_dimensions>::unroll(w, id, std::forward<L>(lambda_function), std::forward<Ts>(iterators)...);
 }
@@ -136,9 +141,10 @@ inline Iterator::Iterator(const ITensor *tensor, const Window &win)
     : Iterator()
 {
     ARM_COMPUTE_ERROR_ON(tensor == nullptr);
-    const ITensorInfo *info = tensor->info();
-    ARM_COMPUTE_ERROR_ON(info == nullptr);
-    const Strides &strides = info->strides_in_bytes();
+    ARM_COMPUTE_ERROR_ON(tensor->info() == nullptr);
+
+    const ITensorInfo *info    = tensor->info();
+    const Strides     &strides = info->strides_in_bytes();
 
     _ptr = tensor->buffer() + info->offset_first_element_in_bytes();
 
diff --git a/arm_compute/core/IAccessWindow.h b/arm_compute/core/IAccessWindow.h
index 583041a48b..4bbcbb3a40 100644
--- a/arm_compute/core/IAccessWindow.h
+++ b/arm_compute/core/IAccessWindow.h
@@ -139,8 +139,8 @@ public:
     }
 
     AccessWindowRectangle(const AccessWindowRectangle &) = delete;
+    AccessWindowRectangle(AccessWindowRectangle &&)      = delete;
     AccessWindowRectangle &operator=(const AccessWindowRectangle &) = delete;
-    AccessWindowRectangle(AccessWindowRectangle &&)                 = default;
     AccessWindowRectangle &operator=(AccessWindowRectangle &&) = default;
     ~AccessWindowRectangle()                                   = default;
 
diff --git a/arm_compute/core/ITensorInfo.h b/arm_compute/core/ITensorInfo.h
index 9112f3ea18..b5677dffd6 100644
--- a/arm_compute/core/ITensorInfo.h
+++ b/arm_compute/core/ITensorInfo.h
@@ -30,6 +30,7 @@
 #include "arm_compute/core/Types.h"
 #include "arm_compute/core/Utils.h"
 #include "arm_compute/core/utils/misc/ICloneable.h"
+#include "arm_compute/core/utils/misc/utility.h"
 
 #include <cstddef>
 
@@ -221,6 +222,51 @@ public:
     * @return A QuantizationInfo containing the scale and offset.
     */
     virtual QuantizationInfo quantization_info() const = 0;
+
+    /** If infos are broadcast compatible tensor info's, return the broadcasted shape and the intersection of
+     * the broadcasted valid regions of the tensors.
+     *
+     * Two tensor info's are broadcast compatible if their shapes are broadcast compatible.
+     *
+     * Two tensor shapes are broadcast compatible if for each dimension, they're equal or one of them is 1.
+     *
+     * If two shapes are compatible, each dimension in the broadcasted shape is the max of the original dimensions.
+     *
+     * @param[in] infos Tensor info's.
+     *
+     * @return The broadcasted shape and valid region, or an empty shape and valid region if the info's are
+     * not broadcast compatible.
+     */
+    template <typename... Infos>
+    static std::pair<TensorShape, ValidRegion> broadcast_shape_and_valid_region(const Infos &... infos)
+    {
+        TensorShape bc_shape = TensorShape::broadcast_shape(infos.tensor_shape()...);
+        ValidRegion bc_valid_region{ Coordinates(), bc_shape };
+
+        auto broadcast_valid_region = [&bc_valid_region](const ITensorInfo & info)
+        {
+            if(info.num_dimensions() != 0)
+            {
+                for(size_t d = 0; d < bc_valid_region.shape.num_dimensions(); ++d)
+                {
+                    const bool is_broadcast = (info.tensor_shape()[d] == 1);
+
+                    const int    anchor_max = std::max(bc_valid_region.anchor[d], info.valid_region().anchor[d]);
+                    const size_t valid_min  = std::min(bc_valid_region.shape[d], info.valid_region().shape[d]);
+
+                    if(!is_broadcast || (valid_min == 0))
+                    {
+                        bc_valid_region.anchor.set(d, anchor_max);
+                        bc_valid_region.shape.set(d, valid_min);
+                    }
+                }
+            }
+        };
+
+        utility::for_each(broadcast_valid_region, infos...);
+
+        return std::pair<TensorShape, ValidRegion>(bc_shape, bc_valid_region);
+    }
 };
 }
 #endif /*__ARM_COMPUTE_TENSORINFO_H__ */
diff --git a/arm_compute/core/NEON/kernels/NEArithmeticAdditionKernel.h b/arm_compute/core/NEON/kernels/NEArithmeticAdditionKernel.h
index eedecfb524..155e792f5d 100644
--- a/arm_compute/core/NEON/kernels/NEArithmeticAdditionKernel.h
+++ b/arm_compute/core/NEON/kernels/NEArithmeticAdditionKernel.h
@@ -85,6 +85,7 @@ public:
 
     // Inherited methods overridden:
     void run(const Window &window, const ThreadInfo &info) override;
+    BorderSize border_size() const override;
 
 private:
     /** Common signature for all the specialised add functions
diff --git a/arm_compute/core/TensorShape.h b/arm_compute/core/TensorShape.h
index 50f1211c18..dc836c98da 100644
--- a/arm_compute/core/TensorShape.h
+++ b/arm_compute/core/TensorShape.h
@@ -26,6 +26,7 @@
 
 #include "arm_compute/core/Dimensions.h"
 #include "arm_compute/core/Error.h"
+#include "arm_compute/core/utils/misc/utility.h"
 
 #include <algorithm>
 #include <array>
@@ -132,6 +133,19 @@ public:
         std::fill(_id.begin() + _num_dimensions, _id.end(), 1);
     }
 
+    /** Return a copy with collapsed dimensions starting from a given point.
+     *
+     * @param[in] start Starting point of collapsing dimensions.
+     *
+     * @return A copy with collapse dimensions starting from start.
+     */
+    TensorShape collapsed_from(size_t start) const
+    {
+        TensorShape copy(*this);
+        copy.collapse(num_dimensions(), start);
+        return copy;
+    }
+
     /** Collapses all dimensions to a single linear total size.
      *
      * @return The total tensor size in terms of elements.
@@ -164,6 +178,50 @@ public:
         return std::accumulate(_id.begin(), _id.begin() + dimension, 1, std::multiplies<size_t>());
     }
 
+    /** If shapes are broadcast compatible, return the broadcasted shape.
+     *
+     * Two tensor shapes are broadcast compatible if for each dimension, they're equal or one of them is 1.
+     *
+     * If two shapes are compatible, each dimension in the broadcasted shape is the max of the original dimensions.
+     *
+     * @param[in] shapes Tensor shapes.
+     *
+     * @return The broadcasted shape or an empty shape if the shapes are not broadcast compatible.
+     */
+    template <typename... Shapes>
+    static TensorShape broadcast_shape(const Shapes &... shapes)
+    {
+        TensorShape bc_shape;
+
+        auto broadcast = [&bc_shape](const TensorShape & other)
+        {
+            if(bc_shape.num_dimensions() == 0)
+            {
+                bc_shape = other;
+            }
+            else if(other.num_dimensions() != 0)
+            {
+                for(size_t d = 0; d < TensorShape::num_max_dimensions; ++d)
+                {
+                    const size_t dim_min = std::min(bc_shape[d], other[d]);
+                    const size_t dim_max = std::max(bc_shape[d], other[d]);
+
+                    if((dim_min != 1) && (dim_min != dim_max))
+                    {
+                        bc_shape = TensorShape{ 0U };
+                        break;
+                    }
+
+                    bc_shape.set(d, dim_max);
+                }
+            }
+        };
+
+        utility::for_each(broadcast, shapes...);
+
+        return bc_shape;
+    }
+
 private:
     /** Remove trailing dimensions of size 1 from the reported number of dimensions. */
     void apply_dimension_correction()
diff --git a/arm_compute/core/Window.h b/arm_compute/core/Window.h
index c890bf8f9e..cca12c9efe 100644
--- a/arm_compute/core/Window.h
+++ b/arm_compute/core/Window.h
@@ -104,6 +104,14 @@ public:
         {
             _step = step;
         }
+        /** Set the dimension's end
+         *
+         * @param[in] end The new end
+         */
+        void set_end(int end)
+        {
+            _end = end;
+        }
 
     private:
         int _start; /**< Start of the dimension */
@@ -302,27 +310,64 @@ public:
         return slide_window_slice<4>(slice);
     }
 
+    /* Collapse the dimensions between @p first and @p last if possible.
+     *
+     * A dimension is collapsable if it starts from 0 and matches the corresponding dimension in the full_window
+     *
+     * @param[in]  full_window   Full window @p window has been created from.
+     * @param[in]  first         Start dimension into which the following are collapsed.
+     * @param[in]  last          End (exclusive) dimension to collapse.
+     * @param[out] has_collapsed (Optional) Whether the window was collapsed.
+     *
+     * @return Collapsed window.
+     */
+    Window collapse_if_possible(const Window &full_window, size_t first, size_t last, bool *has_collapsed = nullptr) const;
+
     /* Collapse the dimensions higher than @p first if possible.
      *
      * A dimension is collapsable if it starts from 0 and matches the corresponding dimension in the full_window
      *
-     * @param[in] full_window Full window @p window has been created from.
-     * @param[in] first       Dimensions into which the following are collapsed.
+     * @param[in]  full_window   Full window @p window has been created from.
+     * @param[in]  first         Start dimension into which the following are collapsed.
+     * @param[out] has_collapsed (Optional) Whether the window was collapsed.
      *
      * @return Collapsed window.
      */
-    Window collapse_if_possible(const Window &full_window, size_t first) const;
+    Window collapse_if_possible(const Window &full_window, size_t first, bool *has_collapsed = nullptr) const
+    {
+        return collapse_if_possible(full_window, first, Coordinates::num_max_dimensions, has_collapsed);
+    }
 
-    /* Collapse the dimensions higher than @p first.
+    /* Collapse the dimensions between @p first and @p last.
      *
      * A dimension is collapsable if it starts from 0 and matches the corresponding dimension in the full_window
      *
      * @param[in] full_window Full window @p window has been created from.
-     * @param[in] first       Dimensions into which the following are collapsed.
+     * @param[in] first       Start dimension into which the following are collapsed.
+     * @param[in] last        End (exclusive) dimension to collapse.
      *
      * @return Collapsed window if successful.
      */
-    Window collapse(const Window &full_window, size_t first) const;
+    Window collapse(const Window &full_window, size_t first, size_t last = Coordinates::num_max_dimensions) const;
+
+    /* Don't advance in the dimension where @p shape is less equal to 1.
+     *
+     * @param[in] shape A TensorShape.
+     *
+     * @return Broadcast window.
+     */
+    Window broadcast_if_dimension_le_one(const TensorShape &shape) const;
+
+    /* Don't advance in the dimension where shape of @p info is less equal to 1.
+     *
+     * @param[in] info An ITensorInfo.
+     *
+     * @return Broadcast window.
+     */
+    Window broadcast_if_dimension_le_one(const ITensorInfo &info) const
+    {
+        return broadcast_if_dimension_le_one(info.tensor_shape());
+    }
 
 private:
     /** First slice of the window
diff --git a/arm_compute/core/Window.inl b/arm_compute/core/Window.inl
index 1b21820f90..23b2a8e322 100644
--- a/arm_compute/core/Window.inl
+++ b/arm_compute/core/Window.inl
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016, 2017 ARM Limited.
+ * Copyright (c) 2016-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -37,55 +37,66 @@ inline constexpr const Window::Dimension &Window::operator[](size_t dimension) c
     // Precondition: dimension < Coordinates::num_max_dimensions
     return _dims.at(dimension);
 }
+
 inline void Window::set(size_t dimension, const Window::Dimension &dim)
 {
     ARM_COMPUTE_ERROR_ON(dimension >= Coordinates::num_max_dimensions);
     _dims[dimension] = dim;
 }
 
-inline Window Window::collapse_if_possible(const Window &full_window, size_t first) const
+inline Window Window::collapse_if_possible(const Window &full_window, const size_t first,
+                                           const size_t last, bool *has_collapsed) const
 {
-    bool   is_collapsable = false;
-    Window collapsed;
-    for(size_t d = 0; d < Coordinates::num_max_dimensions; ++d)
+    Window collapsed(*this);
+
+    bool is_collapsable = true;
+    int  collapsed_end  = _dims[first].end();
+
+    for(size_t d = first + 1; is_collapsable && (d < last); ++d)
     {
-        if(is_collapsable)
-        {
-            collapsed.set(first, Window::Dimension(collapsed[first].end() * _dims[d].start(), collapsed[first].end() * _dims[d].end()));
-        }
-        else
-        {
-            collapsed.set(d, _dims[d]);
-        }
+        // The _dims's dimension must match the full _dims dimension to be collapsable:
+        is_collapsable = (_dims[d].start() == 0) && (full_window[d].start() == 0) && (_dims[d].step() <= 1)
+                         && (full_window[d].end() == _dims[d].end());
+        collapsed_end *= _dims[d].end();
+    }
 
-        if(is_collapsable || d == first) // Try to start collapsing from this dimension
-        {
-            // The _dims's dimension must match the full _dims dimension to be collapsable:
-            is_collapsable = _dims[d].start() == 0 && _dims[d].start() == full_window[d].start()
-                             && full_window[d].end() == _dims[d].end();
-        }
-        else
+    if(is_collapsable)
+    {
+        collapsed._dims.at(first).set_end(collapsed_end);
+        for(size_t d = first + 1; is_collapsable && (d < last); ++d)
         {
-            is_collapsable = false;
+            collapsed.set(d, Dimension());
         }
     }
+
+    if(has_collapsed != nullptr)
+    {
+        *has_collapsed = is_collapsable;
+    }
+
     return collapsed;
 }
 
-inline Window Window::collapse(const Window &full_window, size_t first) const
+inline Window Window::collapse(const Window &full_window, const size_t first, const size_t last) const
 {
-    Window collapsed = collapse_if_possible(full_window, first);
+    bool   has_collapsed = false;
+    Window collapsed     = collapse_if_possible(full_window, first, last, &has_collapsed);
     // Make sure that the window has collapsed
-    int end   = _dims[first].end();
-    int start = 0;
-    ARM_COMPUTE_UNUSED(start);
-    for(size_t d = first + 1; d < Coordinates::num_max_dimensions; ++d)
+    ARM_COMPUTE_ERROR_ON(!has_collapsed);
+    return collapsed;
+}
+
+inline Window Window::broadcast_if_dimension_le_one(const TensorShape &shape) const
+{
+    Window broadcastWin(*this);
+    for(size_t d = 0; d < TensorShape::num_max_dimensions; ++d)
     {
-        start = end * _dims[d].start();
-        end *= _dims[d].end();
+        if(shape[d] <= 1)
+        {
+            broadcastWin.set(d, Dimension(0, 0, 0));
+        }
     }
-    ARM_COMPUTE_ERROR_ON((collapsed[first].end() != end) || (collapsed[first].start() != start));
-    return collapsed;
+    return broadcastWin;
 }
 
 inline void Window::shift(size_t dimension, int shift_value)
@@ -129,9 +140,8 @@ inline void Window::validate() const
 {
     for(size_t i = 0; i < Coordinates::num_max_dimensions; ++i)
     {
-        ARM_COMPUTE_ERROR_ON(_dims[i].step() == 0);
         ARM_COMPUTE_ERROR_ON(_dims[i].end() < _dims[i].start());
-        ARM_COMPUTE_ERROR_ON((_dims[i].end() - _dims[i].start()) % _dims[i].step());
+        ARM_COMPUTE_ERROR_ON((_dims[i].step() != 0) && (((_dims[i].end() - _dims[i].start()) % _dims[i].step()) != 0));
     }
 }
 
diff --git a/arm_compute/runtime/CL/functions/CLArithmeticAddition.h b/arm_compute/runtime/CL/functions/CLArithmeticAddition.h
index 1ef3e274c7..921738d0c2 100644
--- a/arm_compute/runtime/CL/functions/CLArithmeticAddition.h
+++ b/arm_compute/runtime/CL/functions/CLArithmeticAddition.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016, 2017 ARM Limited.
+ * Copyright (c) 2016-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -46,7 +46,7 @@ public:
      * @param[out] output Output tensor. Data types supported: U8 (Only if both inputs are U8), QS8 (only if both inputs are QS8), QS16 (only if both inputs are QS16), S16/F16/F32.
      * @param[in]  policy Policy to use to handle overflow.
      */
-    void configure(const ICLTensor *input1, const ICLTensor *input2, ICLTensor *output, ConvertPolicy policy);
+    void configure(ICLTensor *input1, ICLTensor *input2, ICLTensor *output, ConvertPolicy policy);
     /** Static function to check if given info will lead to a valid configuration of @ref CLArithmeticAddition
      *
      * @param[in] input1 First tensor input info. Data types supported: U8/QS8/QS16/S16/F16/F32.
diff --git a/arm_compute/runtime/CL/functions/CLLaplacianReconstruct.h b/arm_compute/runtime/CL/functions/CLLaplacianReconstruct.h
index 4a676c85a0..6905b03652 100644
--- a/arm_compute/runtime/CL/functions/CLLaplacianReconstruct.h
+++ b/arm_compute/runtime/CL/functions/CLLaplacianReconstruct.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017 ARM Limited.
+ * Copyright (c) 2017-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -76,7 +76,7 @@ public:
      * @param[in]  constant_border_value (Optional) Constant value to use for borders if border_mode is set to CONSTANT.
      *
      */
-    void configure(const CLPyramid *pyramid, const ICLTensor *input, ICLTensor *output, BorderMode border_mode, uint8_t constant_border_value);
+    void configure(const CLPyramid *pyramid, ICLTensor *input, ICLTensor *output, BorderMode border_mode, uint8_t constant_border_value);
 
     // Inherited methods overridden:
     void run() override;
diff --git a/arm_compute/runtime/NEON/functions/NEArithmeticAddition.h b/arm_compute/runtime/NEON/functions/NEArithmeticAddition.h
index 3718073937..c72d0b6d61 100644
--- a/arm_compute/runtime/NEON/functions/NEArithmeticAddition.h
+++ b/arm_compute/runtime/NEON/functions/NEArithmeticAddition.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016, 2017 ARM Limited.
+ * Copyright (c) 2016-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -42,7 +42,7 @@ public:
      * @param[out] output Output tensor. Data types supported: U8/QS8/QS16/S16/F16/F32
      * @param[in]  policy Policy to use to handle overflow.
      */
-    void configure(const ITensor *input1, const ITensor *input2, ITensor *output, ConvertPolicy policy);
+    void configure(ITensor *input1, ITensor *input2, ITensor *output, ConvertPolicy policy);
     /** Static function to check if given info will lead to a valid configuration of @ref NEArithmeticAddition
      *
      * @param[in] input1 First tensor input. Data types supported: U8/QS8/QS16/S16/F16/F32
diff --git a/arm_compute/runtime/NEON/functions/NELaplacianReconstruct.h b/arm_compute/runtime/NEON/functions/NELaplacianReconstruct.h
index 3d423607a3..2143042bd3 100644
--- a/arm_compute/runtime/NEON/functions/NELaplacianReconstruct.h
+++ b/arm_compute/runtime/NEON/functions/NELaplacianReconstruct.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016, 2017 ARM Limited.
+ * Copyright (c) 2016-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -76,7 +76,7 @@ public:
      * @param[in]  constant_border_value (Optional) Constant value to use for borders if border_mode is set to CONSTANT.
      *
      */
-    void configure(const IPyramid *pyramid, const ITensor *input, ITensor *output, BorderMode border_mode, uint8_t constant_border_value);
+    void configure(const IPyramid *pyramid, ITensor *input, ITensor *output, BorderMode border_mode, uint8_t constant_border_value);
 
     // Inherited methods overridden:
     void run() override;
diff --git a/src/core/CL/ICLKernel.cpp b/src/core/CL/ICLKernel.cpp
index 7da74381d3..491e0c4b91 100644
--- a/src/core/CL/ICLKernel.cpp
+++ b/src/core/CL/ICLKernel.cpp
@@ -43,10 +43,11 @@ void arm_compute::enqueue(cl::CommandQueue &queue, ICLKernel &kernel, const Wind
         return;
     }
 
-    // Make sure that dimensions > Z are 1
-    for(unsigned int i = 3; i < Coordinates::num_max_dimensions; ++i)
+    for(unsigned int i = 0; i < Coordinates::num_max_dimensions; ++i)
     {
-        ARM_COMPUTE_ERROR_ON((window[i].end() - window[i].start()) != 1);
+        ARM_COMPUTE_ERROR_ON(window[i].step() == 0);
+        // Make sure that dimensions > Z are 1
+        ARM_COMPUTE_ERROR_ON((i >= 3) && ((window[i].end() - window[i].start()) != 1));
     }
 
     cl::NDRange gws = ICLKernel::gws_from_window(window);
@@ -77,16 +78,6 @@ void arm_compute::enqueue(cl::CommandQueue &queue, ICLKernel &kernel, const Wind
     queue.enqueueNDRangeKernel(kernel.kernel(), cl::NullRange, gws, lws);
 }
 
-ICLKernel::ICLKernel()
-    : _kernel(nullptr), _lws_hint(CLKernelLibrary::get().default_ndrange()), _target(GPUTarget::MIDGARD), _config_id(arm_compute::default_config_id), _max_workgroup_size(0)
-{
-}
-
-cl::Kernel &ICLKernel::kernel()
-{
-    return _kernel;
-}
-
 template <unsigned int dimension_size>
 void ICLKernel::add_tensor_argument(unsigned &idx, const ICLTensor *tensor, const Window &window)
 {
@@ -106,10 +97,10 @@ void ICLKernel::add_tensor_argument(unsigned &idx, const ICLTensor *tensor, cons
     unsigned int idx_start = idx;
     _kernel.setArg(idx++, tensor->cl_buffer());
 
-    for(unsigned int dimension = 0; dimension < dimension_size; dimension++)
+    for(unsigned int d = 0; d < dimension_size; ++d)
     {
-        _kernel.setArg<cl_uint>(idx++, strides[dimension]);
-        _kernel.setArg<cl_uint>(idx++, strides[dimension] * window[dimension].step());
+        _kernel.setArg<cl_uint>(idx++, strides[d]);
+        _kernel.setArg<cl_uint>(idx++, strides[d] * window[d].step());
     }
 
     _kernel.setArg<cl_uint>(idx++, offset_first_element);
@@ -119,66 +110,16 @@ void ICLKernel::add_tensor_argument(unsigned &idx, const ICLTensor *tensor, cons
     ARM_COMPUTE_UNUSED(idx_start);
 }
 
-void ICLKernel::add_1D_tensor_argument(unsigned int &idx, const ICLTensor *tensor, const Window &window)
-{
-    add_tensor_argument<1>(idx, tensor, window);
-}
-
-void ICLKernel::add_2D_tensor_argument(unsigned int &idx, const ICLTensor *tensor, const Window &window)
-{
-    add_tensor_argument<2>(idx, tensor, window);
-}
-
-void ICLKernel::add_3D_tensor_argument(unsigned int &idx, const ICLTensor *tensor, const Window &window)
-{
-    add_tensor_argument<3>(idx, tensor, window);
-}
-
-void ICLKernel::add_4D_tensor_argument(unsigned int &idx, const ICLTensor *tensor, const Window &window)
-{
-    add_tensor_argument<4>(idx, tensor, window);
-}
-
-unsigned int ICLKernel::num_arguments_per_1D_array() const
-{
-    return num_arguments_per_array<1>();
-}
-
-unsigned int ICLKernel::num_arguments_per_1D_tensor() const
-{
-    return num_arguments_per_tensor<1>();
-}
-
-unsigned int ICLKernel::num_arguments_per_2D_tensor() const
-{
-    return num_arguments_per_tensor<2>();
-}
-
-unsigned int ICLKernel::num_arguments_per_3D_tensor() const
-{
-    return num_arguments_per_tensor<3>();
-}
-
-unsigned int ICLKernel::num_arguments_per_4D_tensor() const
-{
-    return num_arguments_per_tensor<4>();
-}
+template void ICLKernel::add_tensor_argument<1>(unsigned &idx, const ICLTensor *tensor, const Window &window);
+template void ICLKernel::add_tensor_argument<2>(unsigned &idx, const ICLTensor *tensor, const Window &window);
+template void ICLKernel::add_tensor_argument<3>(unsigned &idx, const ICLTensor *tensor, const Window &window);
+template void ICLKernel::add_tensor_argument<4>(unsigned &idx, const ICLTensor *tensor, const Window &window);
 
 void ICLKernel::set_target(cl::Device &device)
 {
     _target = get_target_from_device(device);
 }
 
-void ICLKernel::set_target(GPUTarget target)
-{
-    _target = target;
-}
-
-GPUTarget ICLKernel::get_target() const
-{
-    return _target;
-}
-
 size_t ICLKernel::get_max_workgroup_size()
 {
     if(_max_workgroup_size == 0)
diff --git a/src/core/CL/kernels/CLArithmeticAdditionKernel.cpp b/src/core/CL/kernels/CLArithmeticAdditionKernel.cpp
index 75701ee011..c4904ecbe1 100644
--- a/src/core/CL/kernels/CLArithmeticAdditionKernel.cpp
+++ b/src/core/CL/kernels/CLArithmeticAdditionKernel.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016, 2017 ARM Limited.
+ * Copyright (c) 2016-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -24,60 +24,75 @@
 #include "arm_compute/core/CL/kernels/CLArithmeticAdditionKernel.h"
 
 #include "arm_compute/core/CL/CLHelpers.h"
-#include "arm_compute/core/CL/CLKernelLibrary.h"
 #include "arm_compute/core/CL/ICLTensor.h"
-#include "arm_compute/core/CL/OpenCL.h"
-#include "arm_compute/core/Error.h"
-#include "arm_compute/core/Helpers.h"
-#include "arm_compute/core/IAccessWindow.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 <cstddef>
-#include <set>
-#include <string>
 
 using namespace arm_compute;
 
 namespace
 {
-Status validate_arguments(const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *output, ConvertPolicy policy)
+constexpr unsigned int num_elems_processed_per_iteration = 16;
+
+Status validate_arguments(const ITensorInfo &input1, const ITensorInfo &input2, const ITensorInfo &output, ConvertPolicy policy)
 {
     ARM_COMPUTE_UNUSED(policy);
-    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input1, 1, DataType::U8, DataType::QS8, DataType::QS16, DataType::S16, DataType::F16, DataType::F32);
-    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input2, 1, DataType::U8, DataType::QS8, DataType::QS16, DataType::S16, DataType::F16, DataType::F32);
-    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input1, input2);
-    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input1, input2);
+
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(&input1, 1, DataType::U8, DataType::QS8, DataType::QS16, DataType::S16, DataType::F16, DataType::F32);
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(&input2, 1, DataType::U8, DataType::QS8, DataType::QS16, DataType::S16, DataType::F16, DataType::F32);
+
+    const TensorShape out_shape = TensorShape::broadcast_shape(input1.tensor_shape(), input2.tensor_shape());
+
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG(out_shape.total_size() == 0, "Inputs are not broadcast compatible");
+    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(&input1, &input2);
 
     // Validate in case of configured output
-    if((output != nullptr) && (output->total_size() != 0))
+    if(output.total_size() > 0)
     {
-        ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::U8, DataType::QS8, DataType::QS16, DataType::S16, DataType::F16, DataType::F32);
-        ARM_COMPUTE_RETURN_ERROR_ON_MSG(output->data_type() == DataType::U8 && (input1->data_type() != DataType::U8 || input2->data_type() != DataType::U8),
+        ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(&output, 1, DataType::U8, DataType::QS8, DataType::QS16, DataType::S16, DataType::F16, DataType::F32);
+        ARM_COMPUTE_RETURN_ERROR_ON_MSG((output.data_type() == DataType::U8) && ((input1.data_type() != DataType::U8) || (input2.data_type() != DataType::U8)),
                                         "Output can only be U8 if both inputs are U8");
-        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input1, output);
-        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input1, output);
+        ARM_COMPUTE_RETURN_ERROR_ON_MSG(detail::have_different_dimensions(out_shape, output.tensor_shape(), 0),
+                                        "Wrong shape for output");
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(&input1, &output);
     }
 
     return Status{};
 }
 
-std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input1, ITensorInfo *input2, ITensorInfo *output)
+std::pair<Status, Window> validate_and_configure_window(ITensorInfo &input1, ITensorInfo &input2, ITensorInfo &output)
 {
-    constexpr unsigned int num_elems_processed_per_iteration = 16;
+    const std::pair<TensorShape, ValidRegion> broadcast_pair = ITensorInfo::broadcast_shape_and_valid_region(input1, input2);
+    const TensorShape &out_shape    = broadcast_pair.first;
+    const ValidRegion &valid_region = broadcast_pair.second;
 
-    Window win = calculate_max_window(*input1, Steps(num_elems_processed_per_iteration));
+    // Auto initialize output if not initialized
+    {
+        set_shape_if_empty(output, out_shape);
 
-    AccessWindowHorizontal input1_access(input1, 0, num_elems_processed_per_iteration);
-    AccessWindowHorizontal input2_access(input2, 0, num_elems_processed_per_iteration);
-    AccessWindowHorizontal output_access(output, 0, num_elems_processed_per_iteration);
+        if(input1.data_type() == DataType::S16 || input2.data_type() == DataType::S16)
+        {
+            set_format_if_unknown(output, Format::S16);
+        }
+        else if(input1.data_type() == DataType::F16 && input2.data_type() == DataType::F16)
+        {
+            set_format_if_unknown(output, Format::F16);
+        }
+        else if(input1.data_type() == DataType::F32 || input2.data_type() == DataType::F32)
+        {
+            set_format_if_unknown(output, Format::F32);
+        }
+    }
+
+    Window win        = calculate_max_window(valid_region, Steps(num_elems_processed_per_iteration));
+    Window win_input1 = win.broadcast_if_dimension_le_one(input1);
+    Window win_input2 = win.broadcast_if_dimension_le_one(input2);
 
-    bool window_changed = update_window_and_padding(win, input1_access, input2_access, output_access);
+    AccessWindowHorizontal input1_access(&input1, 0, num_elems_processed_per_iteration);
+    AccessWindowHorizontal input2_access(&input2, 0, num_elems_processed_per_iteration);
+    AccessWindowHorizontal output_access(&output, 0, num_elems_processed_per_iteration);
 
-    ValidRegion valid_region = intersect_valid_regions(input1->valid_region(),
-                                                       input2->valid_region());
+    bool window_changed = update_window_and_padding(win_input1, input1_access)
+                          || update_window_and_padding(win_input2, input2_access)
+                          || update_window_and_padding(win, output_access);
 
     output_access.set_valid_region(win, valid_region);
 
@@ -94,26 +109,11 @@ CLArithmeticAdditionKernel::CLArithmeticAdditionKernel()
 void CLArithmeticAdditionKernel::configure(const ICLTensor *input1, const ICLTensor *input2, ICLTensor *output, ConvertPolicy policy)
 {
     ARM_COMPUTE_ERROR_ON_NULLPTR(input1, input2, output);
+    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(*input1->info(), *input2->info(), *output->info(), policy));
 
-    // Auto initialize output if not initialized
-    {
-        set_shape_if_empty(*output->info(), input1->info()->tensor_shape());
-
-        if(input1->info()->data_type() == DataType::S16 || input2->info()->data_type() == DataType::S16)
-        {
-            set_format_if_unknown(*output->info(), Format::S16);
-        }
-        else if(input1->info()->data_type() == DataType::F32 || input2->info()->data_type() == DataType::F32)
-        {
-            set_format_if_unknown(*output->info(), Format::F32);
-        }
-        else if(input1->info()->data_type() == DataType::F16 && input2->info()->data_type() == DataType::F16)
-        {
-            set_format_if_unknown(*output->info(), Format::F16);
-        }
-    }
-
-    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input1->info(), input2->info(), output->info(), policy));
+    // Configure kernel window
+    auto win_config = validate_and_configure_window(*input1->info(), *input2->info(), *output->info());
+    ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
 
     _input1 = input1;
     _input2 = input2;
@@ -135,16 +135,15 @@ void CLArithmeticAdditionKernel::configure(const ICLTensor *input1, const ICLTen
     // Create kernel
     _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("arithmetic_add", build_opts));
 
-    // Configure kernel window
-    auto win_config = validate_and_configure_window(input1->info(), input2->info(), output->info());
-    ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
     ICLKernel::configure(win_config.second);
 }
 
 Status CLArithmeticAdditionKernel::validate(const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *output, ConvertPolicy policy)
 {
-    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input1, input2, output, policy));
-    ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input1->clone().get(), input2->clone().get(), output->clone().get()).first);
+    ARM_COMPUTE_ERROR_ON_NULLPTR(input1, input2, output);
+
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(*input1, *input2, *output, policy));
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(*input1->clone(), *input2->clone(), *output->clone()).first);
 
     return Status{};
 }
@@ -154,16 +153,49 @@ void CLArithmeticAdditionKernel::run(const Window &window, cl::CommandQueue &que
     ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
     ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window);
 
-    Window collapsed = window.collapse_if_possible(ICLKernel::window(), Window::DimZ);
-    Window slice     = collapsed.first_slice_window_3D();
+    const TensorShape &in_shape1 = _input1->info()->tensor_shape();
+    const TensorShape &in_shape2 = _input2->info()->tensor_shape();
+    const TensorShape &out_shape = _output->info()->tensor_shape();
+
+    bool can_collapse = true;
+    if(std::min(in_shape1.total_size(), in_shape2.total_size()) > 1)
+    {
+        can_collapse = (std::min(in_shape1.num_dimensions(), in_shape2.num_dimensions()) > Window::DimZ);
+        for(size_t d = Window::DimZ; can_collapse && (d < out_shape.num_dimensions()); d++)
+        {
+            can_collapse = (in_shape1[d] == in_shape2[d]);
+        }
+    }
+
+    bool   has_collapsed = false;
+    Window collapsed     = can_collapse ? window.collapse_if_possible(ICLKernel::window(), Window::DimZ, &has_collapsed) : window;
+
+    const TensorShape &in_shape1_collapsed = has_collapsed ? in_shape1.collapsed_from(Window::DimZ) : in_shape1;
+    const TensorShape &in_shape2_collapsed = has_collapsed ? in_shape2.collapsed_from(Window::DimZ) : in_shape2;
+
+    Window slice        = collapsed.first_slice_window_3D();
+    Window slice_input1 = slice.broadcast_if_dimension_le_one(in_shape1_collapsed);
+    Window slice_input2 = slice.broadcast_if_dimension_le_one(in_shape2_collapsed);
 
     do
     {
         unsigned int idx = 0;
-        add_3D_tensor_argument(idx, _input1, slice);
-        add_3D_tensor_argument(idx, _input2, slice);
+
+        add_3D_tensor_argument(idx, _input1, slice_input1);
+        add_3D_tensor_argument(idx, _input2, slice_input2);
         add_3D_tensor_argument(idx, _output, slice);
+
         enqueue(queue, *this, slice);
+
+        collapsed.slide_window_slice_3D(slice_input1);
+        collapsed.slide_window_slice_3D(slice_input2);
     }
     while(collapsed.slide_window_slice_3D(slice));
 }
+
+BorderSize CLArithmeticAdditionKernel::border_size() const
+{
+    const unsigned int replicateSize = _output->info()->dimension(0) - std::min(_input1->info()->dimension(0), _input2->info()->dimension(0));
+    const unsigned int border        = std::min<unsigned int>(num_elems_processed_per_iteration - 1U, replicateSize);
+    return BorderSize(0, border, 0, 0);
+}
diff --git a/src/core/CL/kernels/CLPermuteKernel.cpp b/src/core/CL/kernels/CLPermuteKernel.cpp
index 132de60b68..1f36445732 100644
--- a/src/core/CL/kernels/CLPermuteKernel.cpp
+++ b/src/core/CL/kernels/CLPermuteKernel.cpp
@@ -106,10 +106,10 @@ void CLPermuteKernel::run(const Window &window, cl::CommandQueue &queue)
     ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
     ARM_COMPUTE_ERROR_ON_MISMATCHING_WINDOWS(ICLKernel::window(), window);
 
-    Window slice_in = window.first_slice_window_4D();
-    Window slice_out(slice_in);
+    Window slice_in = window.first_slice_window_4D().collapse(ICLKernel::window(), 2, 4);
 
     // Setup output slice
+    Window slice_out(slice_in);
     slice_out.set(Window::DimX, Window::Dimension(0, 0, 0));
     slice_out.set(Window::DimY, Window::Dimension(0, 0, 0));
     slice_out.set(Window::DimZ, Window::Dimension(0, 0, 0));
@@ -117,12 +117,10 @@ void CLPermuteKernel::run(const Window &window, cl::CommandQueue &queue)
 
     do
     {
-        auto         collapsed_slice_in  = slice_in.collapse(ICLKernel::window(), 2);
-        auto         collapsed_slice_out = slice_out.collapse(ICLKernel::window(), 2);
-        unsigned int idx                 = 0;
-        add_4D_tensor_argument(idx, _input, collapsed_slice_in);
-        add_4D_tensor_argument(idx, _output, collapsed_slice_out);
-        enqueue(queue, *this, collapsed_slice_in);
+        unsigned int idx = 0;
+        add_4D_tensor_argument(idx, _input, slice_in);
+        add_4D_tensor_argument(idx, _output, slice_out);
+        enqueue(queue, *this, slice_in);
     }
     while(window.slide_window_slice_4D(slice_in) && window.slide_window_slice_4D(slice_out));
 }
diff --git a/src/core/IAccessWindow.cpp b/src/core/IAccessWindow.cpp
index 7dfe5db5c5..c73f4e7bb2 100644
--- a/src/core/IAccessWindow.cpp
+++ b/src/core/IAccessWindow.cpp
@@ -207,8 +207,8 @@ bool AccessWindowRectangle::update_padding_if_needed(const Window &window)
         return false;
     }
 
-    ARM_COMPUTE_ERROR_ON(window.x().step() * _scale_x == 0);
-    ARM_COMPUTE_ERROR_ON(window.y().step() * _scale_y == 0);
+    ARM_COMPUTE_ERROR_ON(_scale_x == 0);
+    ARM_COMPUTE_ERROR_ON(_scale_y == 0);
 
     const int min_x = window.x().start() * _scale_x + _x;
     const int max_x = (window.x().end() - window.x().step()) * _scale_x + _x + _width;
diff --git a/src/core/NEON/kernels/NEArithmeticAdditionKernel.cpp b/src/core/NEON/kernels/NEArithmeticAdditionKernel.cpp
index 8a98cf7cbc..a487090a98 100644
--- a/src/core/NEON/kernels/NEArithmeticAdditionKernel.cpp
+++ b/src/core/NEON/kernels/NEArithmeticAdditionKernel.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016, 2017 ARM Limited.
+ * Copyright (c) 2016-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -46,10 +46,12 @@ class Coordinates;
 
 namespace
 {
+constexpr unsigned int num_elems_processed_per_iteration = 16;
+
 void add_wrap_QS8_QS8_QS8(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window)
 {
-    Iterator input1(in1, window);
-    Iterator input2(in2, window);
+    Iterator input1(in1, window.broadcast_if_dimension_le_one(in1->info()->tensor_shape()));
+    Iterator input2(in2, window.broadcast_if_dimension_le_one(in2->info()->tensor_shape()));
     Iterator output(out, window);
 
     execute_window_loop(window, [&](const Coordinates & id)
@@ -64,8 +66,8 @@ void add_wrap_QS8_QS8_QS8(const ITensor *in1, const ITensor *in2, ITensor *out,
 
 void add_saturate_QS8_QS8_QS8(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window)
 {
-    Iterator input1(in1, window);
-    Iterator input2(in2, window);
+    Iterator input1(in1, window.broadcast_if_dimension_le_one(in1->info()->tensor_shape()));
+    Iterator input2(in2, window.broadcast_if_dimension_le_one(in2->info()->tensor_shape()));
     Iterator output(out, window);
 
     execute_window_loop(window, [&](const Coordinates & id)
@@ -80,8 +82,8 @@ void add_saturate_QS8_QS8_QS8(const ITensor *in1, const ITensor *in2, ITensor *o
 
 void add_wrap_U8_U8_U8(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window)
 {
-    Iterator input1(in1, window);
-    Iterator input2(in2, window);
+    Iterator input1(in1, window.broadcast_if_dimension_le_one(in1->info()->tensor_shape()));
+    Iterator input2(in2, window.broadcast_if_dimension_le_one(in2->info()->tensor_shape()));
     Iterator output(out, window);
 
     execute_window_loop(window, [&](const Coordinates & id)
@@ -93,8 +95,8 @@ void add_wrap_U8_U8_U8(const ITensor *in1, const ITensor *in2, ITensor *out, con
 
 void add_saturate_U8_U8_U8(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window)
 {
-    Iterator input1(in1, window);
-    Iterator input2(in2, window);
+    Iterator input1(in1, window.broadcast_if_dimension_le_one(in1->info()->tensor_shape()));
+    Iterator input2(in2, window.broadcast_if_dimension_le_one(in2->info()->tensor_shape()));
     Iterator output(out, window);
 
     execute_window_loop(window, [&](const Coordinates & id)
@@ -163,8 +165,8 @@ inline float16x8x2_t vadd2q_f16(const float16x8x2_t &a, const float16x8x2_t &b)
 void add_F16_F16_F16(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window)
 {
 #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
-    Iterator input1(in1, window);
-    Iterator input2(in2, window);
+    Iterator input1(in1, window.broadcast_if_dimension_le_one(in1->info()->tensor_shape()));
+    Iterator input2(in2, window.broadcast_if_dimension_le_one(in2->info()->tensor_shape()));
     Iterator output(out, window);
 
     execute_window_loop(window, [&](const Coordinates & id)
@@ -186,8 +188,8 @@ void add_F16_F16_F16(const ITensor *in1, const ITensor *in2, ITensor *out, const
 
 void add_F32_F32_F32(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window)
 {
-    Iterator input1(in1, window);
-    Iterator input2(in2, window);
+    Iterator input1(in1, window.broadcast_if_dimension_le_one(in1->info()->tensor_shape()));
+    Iterator input2(in2, window.broadcast_if_dimension_le_one(in2->info()->tensor_shape()));
     Iterator output(out, window);
 
     execute_window_loop(window, [&](const Coordinates & id)
@@ -202,8 +204,8 @@ void add_F32_F32_F32(const ITensor *in1, const ITensor *in2, ITensor *out, const
 
 void add_wrap_S16_S16_S16(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window)
 {
-    Iterator input1(in1, window);
-    Iterator input2(in2, window);
+    Iterator input1(in1, window.broadcast_if_dimension_le_one(in1->info()->tensor_shape()));
+    Iterator input2(in2, window.broadcast_if_dimension_le_one(in2->info()->tensor_shape()));
     Iterator output(out, window);
 
     execute_window_loop(window, [&](const Coordinates & id)
@@ -218,8 +220,8 @@ void add_wrap_S16_S16_S16(const ITensor *in1, const ITensor *in2, ITensor *out,
 
 void add_saturate_S16_S16_S16(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window)
 {
-    Iterator input1(in1, window);
-    Iterator input2(in2, window);
+    Iterator input1(in1, window.broadcast_if_dimension_le_one(in1->info()->tensor_shape()));
+    Iterator input2(in2, window.broadcast_if_dimension_le_one(in2->info()->tensor_shape()));
     Iterator output(out, window);
 
     execute_window_loop(window, [&](const Coordinates & id)
@@ -234,8 +236,8 @@ void add_saturate_S16_S16_S16(const ITensor *in1, const ITensor *in2, ITensor *o
 
 void add_wrap_S16_U8_S16(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window)
 {
-    Iterator input1(in1, window);
-    Iterator input2(in2, window);
+    Iterator input1(in1, window.broadcast_if_dimension_le_one(in1->info()->tensor_shape()));
+    Iterator input2(in2, window.broadcast_if_dimension_le_one(in2->info()->tensor_shape()));
     Iterator output(out, window);
 
     execute_window_loop(window, [&](const Coordinates & id)
@@ -257,8 +259,8 @@ void add_wrap_S16_U8_S16(const ITensor *in1, const ITensor *in2, ITensor *out, c
 
 void add_saturate_S16_U8_S16(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window)
 {
-    Iterator input1(in1, window);
-    Iterator input2(in2, window);
+    Iterator input1(in1, window.broadcast_if_dimension_le_one(in1->info()->tensor_shape()));
+    Iterator input2(in2, window.broadcast_if_dimension_le_one(in2->info()->tensor_shape()));
     Iterator output(out, window);
 
     execute_window_loop(window, [&](const Coordinates & id)
@@ -292,8 +294,8 @@ inline void add_saturate_U8_S16_S16(const ITensor *input1, const ITensor *input2
 
 void add_wrap_U8_U8_S16(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window)
 {
-    Iterator input1(in1, window);
-    Iterator input2(in2, window);
+    Iterator input1(in1, window.broadcast_if_dimension_le_one(in1->info()->tensor_shape()));
+    Iterator input2(in2, window.broadcast_if_dimension_le_one(in2->info()->tensor_shape()));
     Iterator output(out, window);
 
     execute_window_loop(window, [&](const Coordinates & id)
@@ -325,8 +327,8 @@ void add_wrap_U8_U8_S16(const ITensor *in1, const ITensor *in2, ITensor *out, co
 
 void add_saturate_U8_U8_S16(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window)
 {
-    Iterator input1(in1, window);
-    Iterator input2(in2, window);
+    Iterator input1(in1, window.broadcast_if_dimension_le_one(in1->info()->tensor_shape()));
+    Iterator input2(in2, window.broadcast_if_dimension_le_one(in2->info()->tensor_shape()));
     Iterator output(out, window);
 
     execute_window_loop(window, [&](const Coordinates & id)
@@ -356,50 +358,84 @@ void add_saturate_U8_U8_S16(const ITensor *in1, const ITensor *in2, ITensor *out
     input1, input2, output);
 }
 
-inline Status validate_arguments(const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *output, ConvertPolicy policy)
+Status validate_arguments(const ITensorInfo &input1, const ITensorInfo &input2, const ITensorInfo &output, ConvertPolicy policy)
 {
     ARM_COMPUTE_UNUSED(policy);
-    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input1, input2, output);
-    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input1, 1, DataType::QS8, DataType::U8, DataType::QS16, DataType::S16, DataType::F16, DataType::F32);
-    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input2, 1, DataType::QS8, DataType::U8, DataType::QS16, DataType::S16, DataType::F16, DataType::F32);
-    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::QS8, DataType::U8, DataType::QS16, DataType::S16, DataType::F16, DataType::F32);
 
-    if(is_data_type_fixed_point(input1->data_type()) || is_data_type_fixed_point(input2->data_type()) || is_data_type_fixed_point(output->data_type()))
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(&input1, 1, DataType::U8, DataType::QS8, DataType::QS16, DataType::S16, DataType::F16, DataType::F32);
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(&input2, 1, DataType::U8, DataType::QS8, DataType::QS16, DataType::S16, DataType::F16, DataType::F32);
+
+    const TensorShape out_shape = TensorShape::broadcast_shape(input1.tensor_shape(), input2.tensor_shape());
+
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG(out_shape.total_size() == 0, "Inputs are not broadcast compatible");
+
+    if(is_data_type_fixed_point(input1.data_type()) || is_data_type_fixed_point(input2.data_type()))
     {
-        // Check that all data types are the same and all fixed-point positions are the same
-        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input1, input2, output);
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(&input1, &input2);
     }
 
-    ARM_COMPUTE_RETURN_ERROR_ON_MSG(
-        !(input1->data_type() == DataType::QS8 && input2->data_type() == DataType::QS8 && output->data_type() == DataType::QS8)
-        && !(input1->data_type() == DataType::U8 && input2->data_type() == DataType::U8 && output->data_type() == DataType::U8)
-        && !(input1->data_type() == DataType::U8 && input2->data_type() == DataType::U8 && output->data_type() == DataType::S16)
-        && !(input1->data_type() == DataType::U8 && input2->data_type() == DataType::S16 && output->data_type() == DataType::S16)
-        && !(input1->data_type() == DataType::S16 && input2->data_type() == DataType::U8 && output->data_type() == DataType::S16)
-        && !(input1->data_type() == DataType::QS16 && input2->data_type() == DataType::QS16 && output->data_type() == DataType::QS16)
-        && !(input1->data_type() == DataType::S16 && input2->data_type() == DataType::S16 && output->data_type() == DataType::S16)
-        && !(input1->data_type() == DataType::F32 && input2->data_type() == DataType::F32 && output->data_type() == DataType::F32)
-        && !(input1->data_type() == DataType::F16 && input2->data_type() == DataType::F16 && output->data_type() == DataType::F16),
-        "You called addition with the wrong image formats");
+    // Validate in case of configured output
+    if(output.total_size() > 0)
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON_MSG(
+            !(input1.data_type() == DataType::QS8 && input2.data_type() == DataType::QS8 && output.data_type() == DataType::QS8)
+            && !(input1.data_type() == DataType::U8 && input2.data_type() == DataType::U8 && output.data_type() == DataType::U8)
+            && !(input1.data_type() == DataType::U8 && input2.data_type() == DataType::U8 && output.data_type() == DataType::S16)
+            && !(input1.data_type() == DataType::U8 && input2.data_type() == DataType::S16 && output.data_type() == DataType::S16)
+            && !(input1.data_type() == DataType::S16 && input2.data_type() == DataType::U8 && output.data_type() == DataType::S16)
+            && !(input1.data_type() == DataType::QS16 && input2.data_type() == DataType::QS16 && output.data_type() == DataType::QS16)
+            && !(input1.data_type() == DataType::S16 && input2.data_type() == DataType::S16 && output.data_type() == DataType::S16)
+            && !(input1.data_type() == DataType::F32 && input2.data_type() == DataType::F32 && output.data_type() == DataType::F32)
+            && !(input1.data_type() == DataType::F16 && input2.data_type() == DataType::F16 && output.data_type() == DataType::F16),
+            "You called addition with the wrong image formats");
+
+        ARM_COMPUTE_RETURN_ERROR_ON_MSG(detail::have_different_dimensions(out_shape, output.tensor_shape(), 0),
+                                        "Wrong shape for output");
+
+        if(is_data_type_fixed_point(input1.data_type()) || is_data_type_fixed_point(output.data_type()))
+        {
+            ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(&input1, &output);
+        }
+    }
 
     return Status{};
 }
 
-inline std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input1, ITensorInfo *input2, ITensorInfo *output)
+std::pair<Status, Window> validate_and_configure_window(ITensorInfo &input1, ITensorInfo &input2, ITensorInfo &output)
 {
-    constexpr unsigned int num_elems_processed_per_iteration = 16;
+    const std::pair<TensorShape, ValidRegion> broadcast_pair = ITensorInfo::broadcast_shape_and_valid_region(input1, input2);
+    const TensorShape &out_shape    = broadcast_pair.first;
+    const ValidRegion &valid_region = broadcast_pair.second;
 
-    // Configure kernel window
-    Window                 win = calculate_max_window(*input1, Steps(num_elems_processed_per_iteration));
-    AccessWindowHorizontal output_access(output, 0, num_elems_processed_per_iteration);
+    // Auto initialize output if not initialized
+    {
+        set_shape_if_empty(output, out_shape);
+
+        if(input1.data_type() == DataType::S16 || input2.data_type() == DataType::S16)
+        {
+            set_format_if_unknown(output, Format::S16);
+        }
+        else if(input1.data_type() == DataType::F16 && input2.data_type() == DataType::F16)
+        {
+            set_format_if_unknown(output, Format::F16);
+        }
+        else if(input1.data_type() == DataType::F32 || input2.data_type() == DataType::F32)
+        {
+            set_format_if_unknown(output, Format::F32);
+        }
+    }
+
+    Window win        = calculate_max_window(valid_region, Steps(num_elems_processed_per_iteration));
+    Window win_input1 = win.broadcast_if_dimension_le_one(input1);
+    Window win_input2 = win.broadcast_if_dimension_le_one(input2);
 
-    bool window_changed = update_window_and_padding(win,
-                                                    AccessWindowHorizontal(input1, 0, num_elems_processed_per_iteration),
-                                                    AccessWindowHorizontal(input2, 0, num_elems_processed_per_iteration),
-                                                    output_access);
+    AccessWindowHorizontal input1_access(&input1, 0, num_elems_processed_per_iteration);
+    AccessWindowHorizontal input2_access(&input2, 0, num_elems_processed_per_iteration);
+    AccessWindowHorizontal output_access(&output, 0, num_elems_processed_per_iteration);
 
-    ValidRegion valid_region = intersect_valid_regions(input1->valid_region(),
-                                                       input2->valid_region());
+    bool window_changed = update_window_and_padding(win_input1, input1_access)
+                          || update_window_and_padding(win_input2, input2_access)
+                          || update_window_and_padding(win, output_access);
 
     output_access.set_valid_region(win, valid_region);
 
@@ -416,26 +452,11 @@ NEArithmeticAdditionKernel::NEArithmeticAdditionKernel()
 void NEArithmeticAdditionKernel::configure(const ITensor *input1, const ITensor *input2, ITensor *output, ConvertPolicy policy)
 {
     ARM_COMPUTE_ERROR_ON_NULLPTR(input1, input2, output);
+    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(*input1->info(), *input2->info(), *output->info(), policy));
 
-    // Auto initialize output if not initialized
-    {
-        set_shape_if_empty(*output->info(), input1->info()->tensor_shape());
-
-        if(input1->info()->data_type() == DataType::S16 || input2->info()->data_type() == DataType::S16)
-        {
-            set_format_if_unknown(*output->info(), Format::S16);
-        }
-        else if(input1->info()->data_type() == DataType::F16 || input2->info()->data_type() == DataType::F16)
-        {
-            set_format_if_unknown(*output->info(), Format::F16);
-        }
-        else if(input1->info()->data_type() == DataType::F32 || input2->info()->data_type() == DataType::F32)
-        {
-            set_format_if_unknown(*output->info(), Format::F32);
-        }
-    }
-
-    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input1->info(), input2->info(), output->info(), policy));
+    // Configure kernel window
+    auto win_config = validate_and_configure_window(*input1->info(), *input2->info(), *output->info());
+    ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
 
     static std::map<std::string, AddFunction *> map_function =
     {
@@ -476,16 +497,15 @@ void NEArithmeticAdditionKernel::configure(const ITensor *input1, const ITensor
         _func = it->second;
     }
 
-    // Configure kernel window
-    auto win_config = validate_and_configure_window(input1->info(), input2->info(), output->info());
-    ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
     INEKernel::configure(win_config.second);
 }
 
 Status NEArithmeticAdditionKernel::validate(const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *output, ConvertPolicy policy)
 {
-    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input1, input2, output, policy));
-    ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input1->clone().get(), input2->clone().get(), output->clone().get()).first);
+    ARM_COMPUTE_ERROR_ON_NULLPTR(input1, input2, output);
+
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(*input1, *input2, *output, policy));
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(*input1->clone(), *input2->clone(), *output->clone()).first);
 
     return Status{};
 }
@@ -499,3 +519,10 @@ void NEArithmeticAdditionKernel::run(const Window &window, const ThreadInfo &inf
 
     (*_func)(_input1, _input2, _output, window);
 }
+
+BorderSize NEArithmeticAdditionKernel::border_size() const
+{
+    const unsigned int replicateSize = _output->info()->dimension(0) - std::min(_input1->info()->dimension(0), _input2->info()->dimension(0));
+    const unsigned int border        = std::min<unsigned int>(num_elems_processed_per_iteration - 1U, replicateSize);
+    return BorderSize(0, border, 0, 0);
+}
diff --git a/src/core/NEON/kernels/NEConvolutionKernel.cpp b/src/core/NEON/kernels/NEConvolutionKernel.cpp
index 7468f58ca5..0a10546b7b 100644
--- a/src/core/NEON/kernels/NEConvolutionKernel.cpp
+++ b/src/core/NEON/kernels/NEConvolutionKernel.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016, 2018 ARM Limited.
+ * Copyright (c) 2016-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -1456,8 +1456,8 @@ void NEConvolutionRectangleKernel::configure(const ITensor *input, ITensor *outp
     constexpr unsigned int num_elems_read_per_iteration      = 16;
     constexpr unsigned int num_elems_written_per_iteration   = 8;
 
-    Window                 win           = calculate_max_window(*input->info(), Steps(num_elems_processed_per_iteration), border_undefined, _border_size);
-    AccessWindowHorizontal output_access = AccessWindowHorizontal(output->info(), 0, num_elems_written_per_iteration);
+    Window                 win = calculate_max_window(*input->info(), Steps(num_elems_processed_per_iteration), border_undefined, _border_size);
+    AccessWindowHorizontal output_access(output->info(), 0, num_elems_written_per_iteration);
 
     update_window_and_padding(win,
                               AccessWindowRectangle(input->info(), -_border_size.left, -_border_size.top, num_elems_read_per_iteration, height),
diff --git a/src/core/Validate.cpp b/src/core/Validate.cpp
index f495e488e2..f5f9f1f8f7 100644
--- a/src/core/Validate.cpp
+++ b/src/core/Validate.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016, 2017 ARM Limited.
+ * Copyright (c) 2016-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -82,7 +82,7 @@ arm_compute::Status arm_compute::error_on_window_dimensions_gte(const char *func
 {
     for(unsigned int i = max_dim; i < arm_compute::Coordinates::num_max_dimensions; ++i)
     {
-        ARM_COMPUTE_RETURN_ERROR_ON_LOC_MSG(win[i].start() != 0 || win[i].end() != win[i].step(),
+        ARM_COMPUTE_RETURN_ERROR_ON_LOC_MSG((win[i].start() != 0) || (win[i].end() != win[i].step()),
                                             function, file, line,
                                             "Maximum number of dimensions expected %u but dimension %u is not empty", max_dim, i);
     }
diff --git a/src/runtime/CL/functions/CLArithmeticAddition.cpp b/src/runtime/CL/functions/CLArithmeticAddition.cpp
index 5c2e582ba2..0b05058c4d 100644
--- a/src/runtime/CL/functions/CLArithmeticAddition.cpp
+++ b/src/runtime/CL/functions/CLArithmeticAddition.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016, 2017 ARM Limited.
+ * Copyright (c) 2016-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -23,6 +23,7 @@
  */
 #include "arm_compute/runtime/CL/functions/CLArithmeticAddition.h"
 
+#include "arm_compute/core/CL/ICLTensor.h"
 #include "arm_compute/core/CL/kernels/CLArithmeticAdditionKernel.h"
 #include "support/ToolchainSupport.h"
 
@@ -30,11 +31,21 @@
 
 using namespace arm_compute;
 
-void CLArithmeticAddition::configure(const ICLTensor *input1, const ICLTensor *input2, ICLTensor *output, ConvertPolicy policy)
+void CLArithmeticAddition::configure(ICLTensor *input1, ICLTensor *input2, ICLTensor *output, ConvertPolicy policy)
 {
     auto k = arm_compute::support::cpp14::make_unique<CLArithmeticAdditionKernel>();
     k->configure(input1, input2, output, policy);
     _kernel = std::move(k);
+
+    if(output->info()->dimension(0) > 1)
+    {
+        ICLTensor *broadcasted_info = (input1->info()->dimension(0) == 1) ? input1 : input2;
+
+        if(broadcasted_info->info()->dimension(0) == 1)
+        {
+            _border_handler.configure(broadcasted_info, _kernel->border_size(), BorderMode::REPLICATE);
+        }
+    }
 }
 
 Status CLArithmeticAddition::validate(const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *output, ConvertPolicy policy)
diff --git a/src/runtime/CL/functions/CLLaplacianReconstruct.cpp b/src/runtime/CL/functions/CLLaplacianReconstruct.cpp
index 678848b82e..911c9b3b27 100644
--- a/src/runtime/CL/functions/CLLaplacianReconstruct.cpp
+++ b/src/runtime/CL/functions/CLLaplacianReconstruct.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017 ARM Limited.
+ * Copyright (c) 2017-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -42,7 +42,7 @@ CLLaplacianReconstruct::CLLaplacianReconstruct() // NOLINT
 {
 }
 
-void CLLaplacianReconstruct::configure(const CLPyramid *pyramid, const ICLTensor *input, ICLTensor *output, BorderMode border_mode, uint8_t constant_border_value)
+void CLLaplacianReconstruct::configure(const CLPyramid *pyramid, ICLTensor *input, ICLTensor *output, BorderMode border_mode, uint8_t constant_border_value)
 {
     ARM_COMPUTE_ERROR_ON(nullptr == pyramid);
     ARM_COMPUTE_ERROR_ON(input == output);
diff --git a/src/runtime/NEON/functions/NEArithmeticAddition.cpp b/src/runtime/NEON/functions/NEArithmeticAddition.cpp
index b5dd4d0d06..7d8e3cff1c 100644
--- a/src/runtime/NEON/functions/NEArithmeticAddition.cpp
+++ b/src/runtime/NEON/functions/NEArithmeticAddition.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017 ARM Limited.
+ * Copyright (c) 2017-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -23,6 +23,7 @@
  */
 #include "arm_compute/runtime/NEON/functions/NEArithmeticAddition.h"
 
+#include "arm_compute/core/ITensor.h"
 #include "arm_compute/core/NEON/kernels/NEArithmeticAdditionKernel.h"
 #include "support/ToolchainSupport.h"
 
@@ -30,11 +31,21 @@
 
 using namespace arm_compute;
 
-void NEArithmeticAddition::configure(const ITensor *input1, const ITensor *input2, ITensor *output, ConvertPolicy policy)
+void NEArithmeticAddition::configure(ITensor *input1, ITensor *input2, ITensor *output, ConvertPolicy policy)
 {
     auto k = arm_compute::support::cpp14::make_unique<NEArithmeticAdditionKernel>();
     k->configure(input1, input2, output, policy);
     _kernel = std::move(k);
+
+    if(output->info()->dimension(0) > 1)
+    {
+        ITensor *broadcasted_info = (input1->info()->dimension(0) == 1) ? input1 : input2;
+
+        if(broadcasted_info->info()->dimension(0) == 1)
+        {
+            _border_handler.configure(broadcasted_info, _kernel->border_size(), BorderMode::REPLICATE);
+        }
+    }
 }
 Status NEArithmeticAddition::validate(const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *output, ConvertPolicy policy)
 {
diff --git a/src/runtime/NEON/functions/NELaplacianReconstruct.cpp b/src/runtime/NEON/functions/NELaplacianReconstruct.cpp
index 0893701cd5..9ad9689b13 100644
--- a/src/runtime/NEON/functions/NELaplacianReconstruct.cpp
+++ b/src/runtime/NEON/functions/NELaplacianReconstruct.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016, 2017 ARM Limited.
+ * Copyright (c) 2016-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -42,7 +42,7 @@ NELaplacianReconstruct::NELaplacianReconstruct() // NOLINT
 {
 }
 
-void NELaplacianReconstruct::configure(const IPyramid *pyramid, const ITensor *input, ITensor *output, BorderMode border_mode, uint8_t constant_border_value)
+void NELaplacianReconstruct::configure(const IPyramid *pyramid, ITensor *input, ITensor *output, BorderMode border_mode, uint8_t constant_border_value)
 {
     ARM_COMPUTE_ERROR_ON(nullptr == pyramid);
     ARM_COMPUTE_ERROR_ON(input == output);
diff --git a/tests/datasets/ShapeDatasets.h b/tests/datasets/ShapeDatasets.h
index 7d4f2b866d..79e052c697 100644
--- a/tests/datasets/ShapeDatasets.h
+++ b/tests/datasets/ShapeDatasets.h
@@ -117,6 +117,34 @@ public:
     }
 };
 
+/** Data set containing pairs of small tensor shapes that are broadcast compatible. */
+class SmallShapesBroadcast final : public framework::dataset::ZipDataset<ShapeDataset, ShapeDataset>
+{
+public:
+    SmallShapesBroadcast()
+        : ZipDataset<ShapeDataset, ShapeDataset>(
+              ShapeDataset("Shape0",
+    {
+        TensorShape{ 9U, 9U },
+                     TensorShape{ 27U, 13U, 2U },
+                     TensorShape{ 128U, 1U, 5U, 3U },
+                     TensorShape{ 9U, 9U, 3U, 4U },
+                     TensorShape{ 27U, 13U, 2U, 4U },
+                     TensorShape{ 1U, 1U, 1U, 5U }
+    }),
+    ShapeDataset("Shape1",
+    {
+        TensorShape{ 9U, 1U, 2U },
+        TensorShape{ 1U, 13U, 2U },
+        TensorShape{ 128U, 64U, 1U, 3U },
+        TensorShape{ 9U, 1U, 3U },
+        TensorShape{ 1U },
+        TensorShape{ 9U, 9U, 3U, 5U }
+    }))
+    {
+    }
+};
+
 /** Data set containing medium tensor shapes. */
 class MediumShapes final : public ShapeDataset
 {
@@ -172,6 +200,30 @@ public:
     }
 };
 
+/** Data set containing pairs of large tensor shapes that are broadcast compatible. */
+class LargeShapesBroadcast final : public framework::dataset::ZipDataset<ShapeDataset, ShapeDataset>
+{
+public:
+    LargeShapesBroadcast()
+        : ZipDataset<ShapeDataset, ShapeDataset>(
+              ShapeDataset("Shape0",
+    {
+        TensorShape{ 1921U, 541U },
+                     TensorShape{ 1U, 485U, 2U, 3U },
+                     TensorShape{ 4159U, 1U },
+                     TensorShape{ 799U }
+    }),
+    ShapeDataset("Shape1",
+    {
+        TensorShape{ 1921U, 1U, 2U },
+        TensorShape{ 641U, 1U, 2U, 3U },
+        TensorShape{ 1U, 127U, 25U },
+        TensorShape{ 799U, 595U, 1U, 4U }
+    }))
+    {
+    }
+};
+
 /** Data set containing large 1D tensor shapes. */
 class Large1DShapes final : public ShapeDataset
 {
diff --git a/tests/framework/datasets/ContainerDataset.h b/tests/framework/datasets/ContainerDataset.h
index bdca97cbac..80616c46fc 100644
--- a/tests/framework/datasets/ContainerDataset.h
+++ b/tests/framework/datasets/ContainerDataset.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017 ARM Limited.
+ * Copyright (c) 2017-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -72,7 +72,8 @@ public:
     {
     }
 
-    ContainerDataset(ContainerDataset &&) = default;
+    ContainerDataset(const ContainerDataset &) = default;
+    ContainerDataset(ContainerDataset &&)      = default;
 
     /** Type of the dataset. */
     using type = std::tuple<container_value_type>;
diff --git a/tests/validation/CL/ArithmeticAddition.cpp b/tests/validation/CL/ArithmeticAddition.cpp
index 787b1b986f..4c19670d50 100644
--- a/tests/validation/CL/ArithmeticAddition.cpp
+++ b/tests/validation/CL/ArithmeticAddition.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017 ARM Limited.
+ * Copyright (c) 2017-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -259,6 +259,25 @@ FIXTURE_DATA_TEST_CASE(RunLarge, CLArithmeticAdditionFixture<float>, framework::
     // Validate output
     validate(CLAccessor(_target), _reference);
 }
+
+template <typename T>
+using CLArithmeticAdditionBroadcastFixture = ArithmeticAdditionBroadcastValidationFixture<CLTensor, CLAccessor, CLArithmeticAddition, T>;
+
+FIXTURE_DATA_TEST_CASE(RunSmallBroadcast, CLArithmeticAdditionBroadcastFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallShapesBroadcast(),
+                       ArithmeticAdditionFP32Dataset),
+                       framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference);
+}
+
+FIXTURE_DATA_TEST_CASE(RunLargeBroadcast, CLArithmeticAdditionBroadcastFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(datasets::LargeShapesBroadcast(),
+                       ArithmeticAdditionFP32Dataset),
+                       framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference);
+}
 TEST_SUITE_END()
 TEST_SUITE_END()
 
diff --git a/tests/validation/NEON/ArithmeticAddition.cpp b/tests/validation/NEON/ArithmeticAddition.cpp
index e20e8df665..32a4ff3a4d 100644
--- a/tests/validation/NEON/ArithmeticAddition.cpp
+++ b/tests/validation/NEON/ArithmeticAddition.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017 ARM Limited.
+ * Copyright (c) 2017-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -263,6 +263,25 @@ FIXTURE_DATA_TEST_CASE(RunLarge, NEArithmeticAdditionFixture<float>, framework::
     // Validate output
     validate(Accessor(_target), _reference);
 }
+
+template <typename T>
+using NEArithmeticAdditionBroadcastFixture = ArithmeticAdditionBroadcastValidationFixture<Tensor, Accessor, NEArithmeticAddition, T>;
+
+FIXTURE_DATA_TEST_CASE(RunSmallBroadcast, NEArithmeticAdditionBroadcastFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallShapesBroadcast(),
+                       ArithmeticAdditionFP32Dataset),
+                       framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })))
+{
+    // Validate output
+    validate(Accessor(_target), _reference);
+}
+
+FIXTURE_DATA_TEST_CASE(RunLargeBroadcast, NEArithmeticAdditionBroadcastFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(datasets::LargeShapesBroadcast(),
+                       ArithmeticAdditionFP32Dataset),
+                       framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })))
+{
+    // Validate output
+    validate(Accessor(_target), _reference);
+}
 TEST_SUITE_END()
 TEST_SUITE_END()
 
diff --git a/tests/validation/fixtures/ArithmeticAdditionFixture.h b/tests/validation/fixtures/ArithmeticAdditionFixture.h
index c3a51b97d1..f3888ae565 100644
--- a/tests/validation/fixtures/ArithmeticAdditionFixture.h
+++ b/tests/validation/fixtures/ArithmeticAdditionFixture.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017 ARM Limited.
+ * Copyright (c) 2017-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -41,15 +41,14 @@ namespace test
 namespace validation
 {
 template <typename TensorType, typename AccessorType, typename FunctionType, typename T>
-class ArithmeticAdditionValidationFixedPointFixture : public framework::Fixture
+class ArithmeticAdditionBroadcastValidationFixedPointFixture : public framework::Fixture
 {
 public:
     template <typename...>
-    void setup(TensorShape shape, DataType data_type0, DataType data_type1, DataType output_data_type, ConvertPolicy convert_policy, int fractional_bits)
+    void setup(const TensorShape &shape0, const TensorShape &shape1, DataType data_type0, DataType data_type1, DataType output_data_type, ConvertPolicy convert_policy, int fractional_bits)
     {
-        _fractional_bits = fractional_bits;
-        _target          = compute_target(shape, data_type0, data_type1, output_data_type, convert_policy, fractional_bits);
-        _reference       = compute_reference(shape, data_type0, data_type1, output_data_type, convert_policy, fractional_bits);
+        _target    = compute_target(shape0, shape1, data_type0, data_type1, output_data_type, convert_policy, fractional_bits);
+        _reference = compute_reference(shape0, shape1, data_type0, data_type1, output_data_type, convert_policy, fractional_bits);
     }
 
 protected:
@@ -59,12 +58,13 @@ protected:
         library->fill_tensor_uniform(tensor, i);
     }
 
-    TensorType compute_target(const TensorShape &shape, DataType data_type0, DataType data_type1, DataType output_data_type, ConvertPolicy convert_policy, int fixed_point_position)
+    TensorType compute_target(const TensorShape &shape0, const TensorShape &shape1, DataType data_type0, DataType data_type1, DataType output_data_type, ConvertPolicy convert_policy,
+                              int fixed_point_position)
     {
         // Create tensors
-        TensorType ref_src1 = create_tensor<TensorType>(shape, data_type0, 1, fixed_point_position);
-        TensorType ref_src2 = create_tensor<TensorType>(shape, data_type1, 1, fixed_point_position);
-        TensorType dst      = create_tensor<TensorType>(shape, output_data_type, 1, fixed_point_position);
+        TensorType ref_src1 = create_tensor<TensorType>(shape0, data_type0, 1, fixed_point_position);
+        TensorType ref_src2 = create_tensor<TensorType>(shape1, data_type1, 1, fixed_point_position);
+        TensorType dst      = create_tensor<TensorType>(TensorShape::broadcast_shape(shape0, shape1), output_data_type, 1, fixed_point_position);
 
         // Create and configure function
         FunctionType add;
@@ -93,11 +93,12 @@ protected:
         return dst;
     }
 
-    SimpleTensor<T> compute_reference(const TensorShape &shape, DataType data_type0, DataType data_type1, DataType output_data_type, ConvertPolicy convert_policy, int fixed_point_position)
+    SimpleTensor<T> compute_reference(const TensorShape &shape0, const TensorShape &shape1, DataType data_type0, DataType data_type1, DataType output_data_type, ConvertPolicy convert_policy,
+                                      int fixed_point_position)
     {
         // Create reference
-        SimpleTensor<T> ref_src1{ shape, data_type0, 1, fixed_point_position };
-        SimpleTensor<T> ref_src2{ shape, data_type1, 1, fixed_point_position };
+        SimpleTensor<T> ref_src1{ shape0, data_type0, 1, fixed_point_position };
+        SimpleTensor<T> ref_src2{ shape1, data_type1, 1, fixed_point_position };
 
         // Fill reference
         fill(ref_src1, 0);
@@ -108,14 +109,36 @@ protected:
 
     TensorType      _target{};
     SimpleTensor<T> _reference{};
-    int             _fractional_bits{};
 };
+
+template <typename TensorType, typename AccessorType, typename FunctionType, typename T>
+class ArithmeticAdditionBroadcastValidationFixture : public ArithmeticAdditionBroadcastValidationFixedPointFixture<TensorType, AccessorType, FunctionType, T>
+{
+public:
+    template <typename...>
+    void setup(const TensorShape &shape0, const TensorShape &shape1, DataType data_type0, DataType data_type1, DataType output_data_type, ConvertPolicy convert_policy)
+    {
+        ArithmeticAdditionBroadcastValidationFixedPointFixture<TensorType, AccessorType, FunctionType, T>::setup(shape0, shape1, data_type0, data_type1, output_data_type, convert_policy, 0);
+    }
+};
+
+template <typename TensorType, typename AccessorType, typename FunctionType, typename T>
+class ArithmeticAdditionValidationFixedPointFixture : public ArithmeticAdditionBroadcastValidationFixedPointFixture<TensorType, AccessorType, FunctionType, T>
+{
+public:
+    template <typename...>
+    void setup(const TensorShape &shape, DataType data_type0, DataType data_type1, DataType output_data_type, ConvertPolicy convert_policy, int fractional_bits)
+    {
+        ArithmeticAdditionBroadcastValidationFixedPointFixture<TensorType, AccessorType, FunctionType, T>::setup(shape, shape, data_type0, data_type1, output_data_type, convert_policy, fractional_bits);
+    }
+};
+
 template <typename TensorType, typename AccessorType, typename FunctionType, typename T>
 class ArithmeticAdditionValidationFixture : public ArithmeticAdditionValidationFixedPointFixture<TensorType, AccessorType, FunctionType, T>
 {
 public:
     template <typename...>
-    void setup(TensorShape shape, DataType data_type0, DataType data_type1, DataType output_data_type, ConvertPolicy convert_policy)
+    void setup(const TensorShape &shape, DataType data_type0, DataType data_type1, DataType output_data_type, ConvertPolicy convert_policy)
     {
         ArithmeticAdditionValidationFixedPointFixture<TensorType, AccessorType, FunctionType, T>::setup(shape, data_type0, data_type1, output_data_type, convert_policy, 0);
     }
diff --git a/tests/validation/reference/ArithmeticAddition.cpp b/tests/validation/reference/ArithmeticAddition.cpp
index 82dd1437cd..17020a6277 100644
--- a/tests/validation/reference/ArithmeticAddition.cpp
+++ b/tests/validation/reference/ArithmeticAddition.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017 ARM Limited.
+ * Copyright (c) 2017-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -35,27 +35,72 @@ namespace validation
 {
 namespace reference
 {
+namespace
+{
 template <typename T>
-SimpleTensor<T> arithmetic_addition(const SimpleTensor<T> &src1, const SimpleTensor<T> &src2, DataType dst_data_type, ConvertPolicy convert_policy)
+T add(T src1, T src2, ConvertPolicy convert_policy)
 {
-    SimpleTensor<T> result(src1.shape(), dst_data_type);
-
     using intermediate_type = typename common_promoted_signed_type<T>::intermediate_type;
 
-    for(int i = 0; i < src1.num_elements(); ++i)
+    intermediate_type val = static_cast<intermediate_type>(src1) + static_cast<intermediate_type>(src2);
+
+    T result = (convert_policy == ConvertPolicy::SATURATE) ? saturate_cast<T>(val) : static_cast<T>(val);
+
+    return result;
+}
+
+template <size_t dim>
+struct BroadcastUnroll
+{
+    template <typename T>
+    static void unroll(const SimpleTensor<T> &src1, const SimpleTensor<T> &src2, SimpleTensor<T> &dst,
+                       ConvertPolicy convert_policy, Coordinates &id_src1, Coordinates &id_src2, Coordinates &id_dst)
     {
-        intermediate_type val = static_cast<intermediate_type>(src1[i]) + static_cast<intermediate_type>(src2[i]);
-        result[i]             = (convert_policy == ConvertPolicy::SATURATE) ? saturate_cast<T>(val) : static_cast<T>(val);
+        const bool src1_is_broadcast = (src1.shape()[dim - 1] != dst.shape()[dim - 1]);
+        const bool src2_is_broadcast = (src2.shape()[dim - 1] != dst.shape()[dim - 1]);
+
+        id_src1.set(dim - 1, 0);
+        id_src2.set(dim - 1, 0);
+        id_dst.set(dim - 1, 0);
+
+        for(size_t i = 0; i < dst.shape()[dim - 1]; ++i, ++id_dst[dim - 1])
+        {
+            BroadcastUnroll < dim - 1 >::unroll(src1, src2, dst, convert_policy, id_src1, id_src2, id_dst);
+
+            id_src1[dim - 1] += !src1_is_broadcast;
+            id_src2[dim - 1] += !src2_is_broadcast;
+        }
     }
+};
 
-    return result;
+template <>
+struct BroadcastUnroll<0>
+{
+    template <typename T>
+    static void unroll(const SimpleTensor<T> &src1, const SimpleTensor<T> &src2, SimpleTensor<T> &dst,
+                       ConvertPolicy convert_policy, Coordinates &id_src1, Coordinates &id_src2, Coordinates &id_dst)
+    {
+        dst[coord2index(dst.shape(), id_dst)] = add(src1[coord2index(src1.shape(), id_src1)], src2[coord2index(src2.shape(), id_src2)], convert_policy);
+    }
+};
+} // namespace
+
+template <typename T>
+SimpleTensor<T> arithmetic_addition(const SimpleTensor<T> &src1, const SimpleTensor<T> &src2, DataType dst_data_type, ConvertPolicy convert_policy)
+{
+    SimpleTensor<T> dst(TensorShape::broadcast_shape(src1.shape(), src2.shape()), dst_data_type);
+
+    Coordinates id_src1, id_src2, id_dst;
+
+    BroadcastUnroll<Coordinates::num_max_dimensions>::unroll(src1, src2, dst, convert_policy, id_src1, id_src2, id_dst);
+
+    return dst;
 }
 
 template SimpleTensor<uint8_t> arithmetic_addition(const SimpleTensor<uint8_t> &src1, const SimpleTensor<uint8_t> &src2, DataType dst_data_type, ConvertPolicy convert_policy);
 template SimpleTensor<int16_t> arithmetic_addition(const SimpleTensor<int16_t> &src1, const SimpleTensor<int16_t> &src2, DataType dst_data_type, ConvertPolicy convert_policy);
 template SimpleTensor<int8_t> arithmetic_addition(const SimpleTensor<int8_t> &src1, const SimpleTensor<int8_t> &src2, DataType dst_data_type, ConvertPolicy convert_policy);
-template SimpleTensor<half> arithmetic_addition(const SimpleTensor<half> &src1, const SimpleTensor<half> &src2, DataType dst_data_type,
-                                                ConvertPolicy convert_policy);
+template SimpleTensor<half> arithmetic_addition(const SimpleTensor<half> &src1, const SimpleTensor<half> &src2, DataType dst_data_type, ConvertPolicy convert_policy);
 template SimpleTensor<float> arithmetic_addition(const SimpleTensor<float> &src1, const SimpleTensor<float> &src2, DataType dst_data_type, ConvertPolicy convert_policy);
 } // namespace reference
 } // namespace validation