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>

11 files changed, 253 insertions, 233 deletions

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);