COMPMID-1566: Add broadcast to CLArithmeticSubtraction

Change-Id: I05d21f9a92013ecfd1128d12cf1561cfd6e5c5e9
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/147983
Tested-by: bsgcomp <bsgcomp@arm.com>
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>

9 files changed, 326 insertions, 141 deletions

diff --git a/src/core/CL/CLKernelLibrary.cpp b/src/core/CL/CLKernelLibrary.cpp
index dfc41da09f..8a309ec757 100644
--- a/src/core/CL/CLKernelLibrary.cpp
+++ b/src/core/CL/CLKernelLibrary.cpp
@@ -152,6 +152,7 @@ const std::map<std::string, std::string> CLKernelLibrary::_kernel_program_map =
     { "arithmetic_add_quantized", "arithmetic_op_quantized.cl" },
     { "arithmetic_add", "arithmetic_op.cl" },
     { "arithmetic_sub", "arithmetic_op.cl" },
+    { "arithmetic_sub_quantized", "arithmetic_op_quantized.cl" },
     { "arithmetic_div", "arithmetic_op.cl" },
     { "batch_to_space_nchw", "batch_to_space.cl" },
     { "batch_to_space_static_nchw", "batch_to_space.cl" },
diff --git a/src/core/CL/cl_kernels/arithmetic_op_quantized.cl b/src/core/CL/cl_kernels/arithmetic_op_quantized.cl
index 082317ba11..5f31c56250 100644
--- a/src/core/CL/cl_kernels/arithmetic_op_quantized.cl
+++ b/src/core/CL/cl_kernels/arithmetic_op_quantized.cl
@@ -31,12 +31,16 @@
 #define SUB(x, y) (x) - (y)
 #endif /* SATURATE */
 
-#if defined(OFFSET_IN1)
-
+#if defined(OFFSET_IN1) && defined(OFFSET_IN2) && defined(OFFSET_OUT) && defined(SCALE_IN1) && defined(SCALE_IN2) && defined(SCALE_OUT)
 /** This function adds two tensors.
  *
- * @attention The quantization offset must be passed at compile time using -DOFFSET_IN1, i.e. -DOFFSET_IN1=10
- * @attention To perform saturating operation -DSATURATE has to be passed to the compiler otherwise wrapping policy will be used.
+ * @note The quantization offset of the first operand must be passed at compile time using -DOFFSET_IN1, i.e. -DOFFSET_IN1=10
+ * @note The quantization offset of the second operand must be passed at compile time using -DOFFSET_IN2, i.e. -DOFFSET_IN2=10
+ * @note The quantization offset of the output must be passed at compile time using -DOFFSET_OUT, i.e. -DOFFSET_OUT=10
+ * @note The quantization scale of the first operand must be passed at compile time using -DSCALE_IN1, i.e. -DSCALE_IN1=10
+ * @note The quantization scale of the second operand must be passed at compile time using -DSCALE_IN2, i.e. -DSCALE_IN2=10
+ * @note The quantization scale of the output must be passed at compile time using -DSCALE_OUT, i.e. -DSCALE_OUT=10
+ * @note To perform saturating operation -DSATURATE has to be passed to the compiler otherwise wrapping policy will be used.
  *
  * @param[in]  in1_ptr                           Pointer to the source tensor. Supported data types: QASYMM8
  * @param[in]  in1_stride_x                      Stride of the source tensor in X dimension (in bytes)
@@ -87,4 +91,64 @@ __kernel void arithmetic_add_quantized(
     // Store result
     vstore16(res, 0, (__global uchar *)out.ptr);
 }
-#endif /* defined(OFFSET) */
+
+/** This function subtracts two tensors.
+ *
+ * @note The quantization offset of the first operand must be passed at compile time using -DOFFSET_IN1, i.e. -DOFFSET_IN1=10
+ * @note The quantization offset of the second operand must be passed at compile time using -DOFFSET_IN2, i.e. -DOFFSET_IN2=10
+ * @note The quantization offset of the output must be passed at compile time using -DOFFSET_OUT, i.e. -DOFFSET_OUT=10
+ * @note The quantization scale of the first operand must be passed at compile time using -DSCALE_IN1, i.e. -DSCALE_IN1=10
+ * @note The quantization scale of the second operand must be passed at compile time using -DSCALE_IN2, i.e. -DSCALE_IN2=10
+ * @note The quantization scale of the output must be passed at compile time using -DSCALE_OUT, i.e. -DSCALE_OUT=10
+ * @note To perform saturating operation -DSATURATE has to be passed to the compiler otherwise wrapping policy will be used.
+ *
+ * @param[in]  in1_ptr                           Pointer to the source tensor. Supported data types: QASYMM8
+ * @param[in]  in1_stride_x                      Stride of the source tensor in X dimension (in bytes)
+ * @param[in]  in1_step_x                        in1_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  in1_stride_y                      Stride of the source tensor in Y dimension (in bytes)
+ * @param[in]  in1_step_y                        in1_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  in1_stride_z                      Stride of the source tensor in Z dimension (in bytes)
+ * @param[in]  in1_step_z                        in1_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  in1_offset_first_element_in_bytes The offset of the first element in the source tensor
+ * @param[in]  in2_ptr                           Pointer to the source tensor. Supported data types: same as @p in1_ptr
+ * @param[in]  in2_stride_x                      Stride of the source tensor in X dimension (in bytes)
+ * @param[in]  in2_step_x                        in2_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  in2_stride_y                      Stride of the source tensor in Y dimension (in bytes)
+ * @param[in]  in2_step_y                        in2_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  in2_stride_z                      Stride of the source tensor in Z dimension (in bytes)
+ * @param[in]  in2_step_z                        in2_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  in2_offset_first_element_in_bytes The offset of the first element in the source tensor
+ * @param[out] out_ptr                           Pointer to the destination tensor. Supported data types: same as @p in1_ptr
+ * @param[in]  out_stride_x                      Stride of the destination tensor in X dimension (in bytes)
+ * @param[in]  out_step_x                        out_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  out_stride_y                      Stride of the destination tensor in Y dimension (in bytes)
+ * @param[in]  out_step_y                        out_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  out_stride_z                      Stride of the source tensor in Z dimension (in bytes)
+ * @param[in]  out_step_z                        out_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  out_offset_first_element_in_bytes The offset of the first element in the destination tensor
+ */
+__kernel void arithmetic_sub_quantized(
+    TENSOR3D_DECLARATION(in1),
+    TENSOR3D_DECLARATION(in2),
+    TENSOR3D_DECLARATION(out))
+{
+    // Get pixels pointer
+    Tensor3D in1 = CONVERT_TO_TENSOR3D_STRUCT(in1);
+    Tensor3D in2 = CONVERT_TO_TENSOR3D_STRUCT(in2);
+    Tensor3D out = CONVERT_TO_TENSOR3D_STRUCT(out);
+
+    int16 in_a = CONVERT(vload16(0, (__global uchar *)in1.ptr), int16);
+    int16 in_b = CONVERT(vload16(0, (__global uchar *)in2.ptr), int16);
+
+    in_a = SUB(in_a, (int16)((int)OFFSET_IN1));
+    in_b = SUB(in_b, (int16)((int)OFFSET_IN2));
+
+    const float16 in1f32  = convert_float16(in_a) * (float16)((float)SCALE_IN1);
+    const float16 in2f32  = convert_float16(in_b) * (float16)((float)SCALE_IN2);
+    const float16 qresf32 = (in1f32 - in2f32) / ((float16)(float)SCALE_OUT) + ((float16)((float16)OFFSET_OUT));
+    const uchar16 res     = convert_uchar16_sat(convert_int16_rte(qresf32));
+
+    // Store result
+    vstore16(res, 0, (__global uchar *)out.ptr);
+}
+#endif /* defined(OFFSET_IN1) && defined(OFFSET_IN2) && defined(OFFSET_OUT) && defined(SCALE_IN1) && defined(SCALE_IN2) && defined(SCALE_OUT) */
\ No newline at end of file
diff --git a/src/core/CL/kernels/CLArithmeticAdditionKernel.cpp b/src/core/CL/kernels/CLArithmeticAdditionKernel.cpp
index 2372d458cf..de14f00856 100644
--- a/src/core/CL/kernels/CLArithmeticAdditionKernel.cpp
+++ b/src/core/CL/kernels/CLArithmeticAdditionKernel.cpp
@@ -159,7 +159,7 @@ void CLArithmeticAdditionKernel::configure(const ICLTensor *input1, const ICLTen
 
 Status CLArithmeticAdditionKernel::validate(const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *output, ConvertPolicy policy)
 {
-    ARM_COMPUTE_ERROR_ON_NULLPTR(input1, input2, output);
+    ARM_COMPUTE_RETURN_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);
diff --git a/src/core/CL/kernels/CLArithmeticSubtractionKernel.cpp b/src/core/CL/kernels/CLArithmeticSubtractionKernel.cpp
index 299ac553e9..95d201104d 100644
--- a/src/core/CL/kernels/CLArithmeticSubtractionKernel.cpp
+++ b/src/core/CL/kernels/CLArithmeticSubtractionKernel.cpp
@@ -36,45 +36,82 @@
 #include <set>
 #include <string>
 
-using namespace arm_compute;
-
+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_F16_UNSUPPORTED(input1);
-    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input1, 1, DataType::U8, DataType::S16, DataType::F16, DataType::F32);
-    ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input2);
-    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input2, 1, DataType::U8, DataType::S16, DataType::F16, DataType::F32);
-    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input1, input2);
+    ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(&input1);
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(&input1, 1, DataType::U8, DataType::QASYMM8, DataType::S16, DataType::F16, DataType::F32);
+    ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(&input2);
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(&input2, 1, DataType::U8, DataType::QASYMM8, DataType::S16, DataType::F16, DataType::F32);
+    const bool is_qasymm = is_data_type_quantized_asymmetric(input1.data_type()) || is_data_type_quantized_asymmetric(input2.data_type());
+    if(is_qasymm)
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(&input1, &input2);
+    }
+
+    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");
 
     // Validate in case of configured output
-    if((output != nullptr) && (output->total_size() != 0))
+    if(output.total_size() > 0)
     {
-        ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(output);
-        ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::U8, 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_F16_UNSUPPORTED(&output);
+        ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(&output, 1, DataType::U8, DataType::QASYMM8, 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_MSG(detail::have_different_dimensions(out_shape, output.tensor_shape(), 0),
+                                        "Wrong shape for output");
+        if(is_qasymm)
+        {
+            ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(&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));
-    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);
+    // Auto initialize output if not initialized
+    {
+        set_shape_if_empty(output, out_shape);
 
-    bool window_changed = update_window_and_padding(win, input1_access, input2_access, output_access);
+        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);
 
-    ValidRegion valid_region = intersect_valid_regions(input1->valid_region(),
-                                                       input2->valid_region());
+    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);
+
+    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);
 
@@ -91,22 +128,11 @@ CLArithmeticSubtractionKernel::CLArithmeticSubtractionKernel()
 void CLArithmeticSubtractionKernel::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);
-        }
-    }
-
-    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;
@@ -114,26 +140,39 @@ void CLArithmeticSubtractionKernel::configure(const ICLTensor *input1, const ICL
 
     bool has_float_out = is_data_type_float(output->info()->data_type());
 
+    // Setup kernel
+    std::string kernel_name = "arithmetic_sub";
+
     // Set kernel build options
-    std::set<std::string> build_opts;
-    build_opts.emplace((policy == ConvertPolicy::WRAP || has_float_out) ? "-DWRAP" : "-DSATURATE");
-    build_opts.emplace("-DDATA_TYPE_IN1=" + get_cl_type_from_data_type(input1->info()->data_type()));
-    build_opts.emplace("-DDATA_TYPE_IN2=" + get_cl_type_from_data_type(input2->info()->data_type()));
-    build_opts.emplace("-DDATA_TYPE_OUT=" + get_cl_type_from_data_type(output->info()->data_type()));
+    CLBuildOptions build_opts;
+    build_opts.add_option_if_else(policy == ConvertPolicy::WRAP || has_float_out, "-DWRAP", "-DSATURATE");
+    build_opts.add_option("-DDATA_TYPE_IN1=" + get_cl_type_from_data_type(input1->info()->data_type()));
+    build_opts.add_option("-DDATA_TYPE_IN2=" + get_cl_type_from_data_type(input2->info()->data_type()));
+    build_opts.add_option("-DDATA_TYPE_OUT=" + get_cl_type_from_data_type(output->info()->data_type()));
+    if(is_data_type_quantized_asymmetric(input1->info()->data_type()))
+    {
+        build_opts.add_option("-DOFFSET_IN1=" + support::cpp11::to_string(input1->info()->quantization_info().offset));
+        build_opts.add_option("-DOFFSET_IN2=" + support::cpp11::to_string(input2->info()->quantization_info().offset));
+        build_opts.add_option("-DOFFSET_OUT=" + support::cpp11::to_string(output->info()->quantization_info().offset));
+        build_opts.add_option("-DSCALE_IN1=" + support::cpp11::to_string(input1->info()->quantization_info().scale));
+        build_opts.add_option("-DSCALE_IN2=" + support::cpp11::to_string(input2->info()->quantization_info().scale));
+        build_opts.add_option("-DSCALE_OUT=" + support::cpp11::to_string(output->info()->quantization_info().scale));
+        kernel_name += "_quantized";
+    }
 
     // Create kernel
-    _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("arithmetic_sub", build_opts));
+    _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel(kernel_name, build_opts.options()));
 
     // 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_internal(win_config.second);
 }
 
 Status CLArithmeticSubtractionKernel::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_RETURN_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{};
 }
@@ -143,16 +182,51 @@ void CLArithmeticSubtractionKernel::run(const Window &window, cl::CommandQueue &
     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();
+
+    // Collapse only if broadcast dimensions is less than 2, or in case of no broadcasting
+    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 CLArithmeticSubtractionKernel::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);
+}
+} // namespace arm_compute
\ No newline at end of file
diff --git a/src/core/NEON/kernels/NEArithmeticAdditionKernel.cpp b/src/core/NEON/kernels/NEArithmeticAdditionKernel.cpp
index a6102b159f..169554f87a 100644
--- a/src/core/NEON/kernels/NEArithmeticAdditionKernel.cpp
+++ b/src/core/NEON/kernels/NEArithmeticAdditionKernel.cpp
@@ -456,7 +456,7 @@ void NEArithmeticAdditionKernel::configure(const ITensor *input1, const ITensor
 
 Status NEArithmeticAdditionKernel::validate(const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *output, ConvertPolicy policy)
 {
-    ARM_COMPUTE_ERROR_ON_NULLPTR(input1, input2, output);
+    ARM_COMPUTE_RETURN_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);
diff --git a/src/core/NEON/kernels/NEArithmeticSubtractionKernel.cpp b/src/core/NEON/kernels/NEArithmeticSubtractionKernel.cpp
index 3c76548b0a..ff8fb84958 100644
--- a/src/core/NEON/kernels/NEArithmeticSubtractionKernel.cpp
+++ b/src/core/NEON/kernels/NEArithmeticSubtractionKernel.cpp
@@ -46,10 +46,12 @@ class Coordinates;
 
 namespace
 {
+constexpr unsigned int num_elems_processed_per_iteration = 16;
+
 void sub_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)
@@ -64,8 +66,8 @@ void sub_wrap_U8_U8_U8(const ITensor *in1, const ITensor *in2, ITensor *out, con
 
 void sub_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)
@@ -80,8 +82,8 @@ void sub_saturate_U8_U8_U8(const ITensor *in1, const ITensor *in2, ITensor *out,
 
 void sub_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)
@@ -104,8 +106,8 @@ void sub_wrap_S16_S16_S16(const ITensor *in1, const ITensor *in2, ITensor *out,
 
 void sub_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)
@@ -144,8 +146,8 @@ inline float16x8x2_t vsub2q_f16(const float16x8x2_t &a, const float16x8x2_t &b)
 void sub_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)
@@ -167,8 +169,8 @@ void sub_F16_F16_F16(const ITensor *in1, const ITensor *in2, ITensor *out, const
 
 void sub_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)
@@ -192,8 +194,8 @@ void sub_F32_F32_F32(const ITensor *in1, const ITensor *in2, ITensor *out, const
 }
 void sub_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)
@@ -213,8 +215,8 @@ void sub_wrap_S16_U8_S16(const ITensor *in1, const ITensor *in2, ITensor *out, c
 
 void sub_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)
@@ -234,8 +236,8 @@ void sub_saturate_S16_U8_S16(const ITensor *in1, const ITensor *in2, ITensor *ou
 
 void sub_wrap_U8_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)
@@ -255,8 +257,8 @@ void sub_wrap_U8_S16_S16(const ITensor *in1, const ITensor *in2, ITensor *out, c
 
 void sub_saturate_U8_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)
@@ -276,8 +278,8 @@ void sub_saturate_U8_S16_S16(const ITensor *in1, const ITensor *in2, ITensor *ou
 
 void sub_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)
@@ -298,8 +300,8 @@ void sub_wrap_U8_U8_S16(const ITensor *in1, const ITensor *in2, ITensor *out, co
 
 void sub_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)
@@ -318,43 +320,71 @@ void sub_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)
+inline Status validate_arguments(const ITensorInfo &input1, const ITensorInfo &input2, const ITensorInfo &output, ConvertPolicy policy)
 {
     ARM_COMPUTE_UNUSED(policy);
-    ARM_COMPUTE_RETURN_ERROR_ON_CPU_F16_UNSUPPORTED(input1);
-    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input1, input2, output);
-    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input1, 1, DataType::U8, DataType::S16, DataType::F16, DataType::F32);
-    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input2, 1, DataType::U8, DataType::S16, DataType::F16, DataType::F32);
-    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::U8, DataType::S16, DataType::F16, DataType::F32);
-
-    ARM_COMPUTE_RETURN_ERROR_ON_MSG(
-        !(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::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 subtract with the wrong image formats");
+    ARM_COMPUTE_RETURN_ERROR_ON_CPU_F16_UNSUPPORTED(&input1);
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(&input1, 1, DataType::U8, DataType::S16, DataType::F16, DataType::F32);
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(&input2, 1, DataType::U8, DataType::S16, DataType::F16, DataType::F32);
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(&output, 1, DataType::U8, 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");
 
+    // Validate in case of configured output
+    if(output.total_size() > 0)
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON_MSG(
+            !(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::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 subtract 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");
+    }
     return Status{};
 }
 
-inline std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input1, ITensorInfo *input2, ITensorInfo *output)
+inline 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);
 
-    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);
+        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);
 
-    ValidRegion valid_region = intersect_valid_regions(input1->valid_region(),
-                                                       input2->valid_region());
+    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);
+
+    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);
 
@@ -371,26 +401,11 @@ NEArithmeticSubtractionKernel::NEArithmeticSubtractionKernel()
 void NEArithmeticSubtractionKernel::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, NEArithmeticSubtractionKernel::SubFunction *> map_function =
     {
@@ -427,16 +442,15 @@ void NEArithmeticSubtractionKernel::configure(const ITensor *input1, const ITens
         _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 NEArithmeticSubtractionKernel::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_RETURN_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{};
 }
@@ -450,3 +464,10 @@ void NEArithmeticSubtractionKernel::run(const Window &window, const ThreadInfo &
 
     (*_func)(_input1, _input2, _output, window);
 }
+
+BorderSize NEArithmeticSubtractionKernel::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);
+}
\ No newline at end of file
diff --git a/src/runtime/CL/functions/CLArithmeticSubtraction.cpp b/src/runtime/CL/functions/CLArithmeticSubtraction.cpp
index 5fca30c4f9..e661f6adc1 100644
--- a/src/runtime/CL/functions/CLArithmeticSubtraction.cpp
+++ b/src/runtime/CL/functions/CLArithmeticSubtraction.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/CLArithmeticSubtraction.h"
 
+#include "arm_compute/core/CL/ICLTensor.h"
 #include "arm_compute/core/CL/kernels/CLArithmeticSubtractionKernel.h"
 #include "support/ToolchainSupport.h"
 
@@ -30,11 +31,21 @@
 
 using namespace arm_compute;
 
-void CLArithmeticSubtraction::configure(const ICLTensor *input1, const ICLTensor *input2, ICLTensor *output, ConvertPolicy policy)
+void CLArithmeticSubtraction::configure(ICLTensor *input1, ICLTensor *input2, ICLTensor *output, ConvertPolicy policy)
 {
     auto k = arm_compute::support::cpp14::make_unique<CLArithmeticSubtractionKernel>();
     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 CLArithmeticSubtraction::validate(const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *output, ConvertPolicy policy)
diff --git a/src/runtime/NEON/functions/NEArithmeticAddition.cpp b/src/runtime/NEON/functions/NEArithmeticAddition.cpp
index 7d8e3cff1c..677e9f676f 100644
--- a/src/runtime/NEON/functions/NEArithmeticAddition.cpp
+++ b/src/runtime/NEON/functions/NEArithmeticAddition.cpp
@@ -29,8 +29,8 @@
 
 #include <utility>
 
-using namespace arm_compute;
-
+namespace arm_compute
+{
 void NEArithmeticAddition::configure(ITensor *input1, ITensor *input2, ITensor *output, ConvertPolicy policy)
 {
     auto k = arm_compute::support::cpp14::make_unique<NEArithmeticAdditionKernel>();
@@ -51,3 +51,4 @@ Status NEArithmeticAddition::validate(const ITensorInfo *input1, const ITensorIn
 {
     return NEArithmeticAdditionKernel::validate(input1, input2, output, policy);
 }
+} // namespace arm_compute
diff --git a/src/runtime/NEON/functions/NEArithmeticSubtraction.cpp b/src/runtime/NEON/functions/NEArithmeticSubtraction.cpp
index 5c0491ec6f..ceb4b496bc 100644
--- a/src/runtime/NEON/functions/NEArithmeticSubtraction.cpp
+++ b/src/runtime/NEON/functions/NEArithmeticSubtraction.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017 ARM Limited.
+ * Copyright (c) 2017-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -23,20 +23,33 @@
  */
 #include "arm_compute/runtime/NEON/functions/NEArithmeticSubtraction.h"
 
+#include "arm_compute/core/ITensor.h"
 #include "arm_compute/core/NEON/kernels/NEArithmeticSubtractionKernel.h"
 #include "support/ToolchainSupport.h"
 
 #include <utility>
 
-using namespace arm_compute;
-
-void NEArithmeticSubtraction::configure(const ITensor *input1, const ITensor *input2, ITensor *output, ConvertPolicy policy)
+namespace arm_compute
+{
+void NEArithmeticSubtraction::configure(ITensor *input1, ITensor *input2, ITensor *output, ConvertPolicy policy)
 {
     auto k = arm_compute::support::cpp14::make_unique<NEArithmeticSubtractionKernel>();
     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 NEArithmeticSubtraction::validate(const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *output, ConvertPolicy policy)
 {
     return NEArithmeticSubtractionKernel::validate(input1, input2, output, policy);
 }
+} // namespace arm_compute