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>

10 files changed, 305 insertions, 103 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));
     }
 }