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>

29 files changed, 637 insertions, 659 deletions

diff --git a/arm_compute/core/CL/kernels/CLArithmeticAdditionKernel.h b/arm_compute/core/CL/kernels/CLArithmeticAdditionKernel.h
index f4275f4153..48e72f3c13 100644
--- a/arm_compute/core/CL/kernels/CLArithmeticAdditionKernel.h
+++ b/arm_compute/core/CL/kernels/CLArithmeticAdditionKernel.h
@@ -51,7 +51,7 @@ public:
     CLArithmeticAdditionKernel &operator=(CLArithmeticAdditionKernel &&) = default;
     /** Default destructor */
     ~CLArithmeticAdditionKernel() = default;
-    /** Initialise the kernel's inputs, output and convertion policy.
+    /** Initialise the kernel's inputs, output and conversion policy.
      *
      * @param[in]  input1 First tensor input. Data types supported: U8/QASYMM8/S16/F16/F32.
      * @param[in]  input2 Second tensor input. Data types supported: U8, QASYMM8 (only if @p input1 is QASYMM8), S16/F16/F32.
diff --git a/arm_compute/core/CL/kernels/CLArithmeticSubtractionKernel.h b/arm_compute/core/CL/kernels/CLArithmeticSubtractionKernel.h
index 35b918fe4b..9875ac7a31 100644
--- a/arm_compute/core/CL/kernels/CLArithmeticSubtractionKernel.h
+++ b/arm_compute/core/CL/kernels/CLArithmeticSubtractionKernel.h
@@ -53,19 +53,19 @@ public:
     /** Default destructor */
     ~CLArithmeticSubtractionKernel() = default;
 
-    /** Initialise the kernel's inputs, output and convertion policy.
+    /** Initialise the kernel's inputs, output and conversion policy.
      *
-     * @param[in]  input1 First tensor input. Data types supported: U8/S16/F16/F32.
-     * @param[in]  input2 Second tensor input. Data types supported: U8/S16/F16/F32.
-     * @param[out] output Output tensor. Data types supported: U8 (Only if both inputs are U8), S16/F16/F32.
+     * @param[in]  input1 First tensor input. Data types supported: U8/QASYMM8/S16/F16/F32.
+     * @param[in]  input2 Second tensor input. Data types supported: U8/QASYMM8/S16/F16/F32.
+     * @param[out] output Output tensor. Data types supported: U8 (Only if both inputs are U8), QASYMM8/S16/F16/F32.
      * @param[in]  policy Policy to use to handle overflow.
      */
     void configure(const ICLTensor *input1, const ICLTensor *input2, ICLTensor *output, ConvertPolicy policy);
     /** Static function to check if given info will lead to a valid configuration of @ref CLArithmeticSubtractionKernel
      *
-     * @param[in] input1 First tensor input info. Data types supported: U8/S16/F16/F32.
-     * @param[in] input2 Second tensor input info. Data types supported: U8/S16/F16/F32.
-     * @param[in] output Output tensor info. Data types supported: U8 (Only if both inputs are U8), S16/F16/F32.
+     * @param[in] input1 First tensor input info. Data types supported: U8/QASYMM8/S16/F16/F32.
+     * @param[in] input2 Second tensor input info. Data types supported: U8/QASYMM8/S16/F16/F32.
+     * @param[in] output Output tensor info. Data types supported: U8 (Only if both inputs are U8), QASYMM8/S16/F16/F32.
      * @param[in] policy Policy to use to handle overflow.
      *
      * @return a status
@@ -74,6 +74,7 @@ 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/NEON/kernels/NEArithmeticSubtractionKernel.h b/arm_compute/core/NEON/kernels/NEArithmeticSubtractionKernel.h
index 3e93922b65..64ad6e072d 100644
--- a/arm_compute/core/NEON/kernels/NEArithmeticSubtractionKernel.h
+++ b/arm_compute/core/NEON/kernels/NEArithmeticSubtractionKernel.h
@@ -83,6 +83,7 @@ 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 sub functions
diff --git a/arm_compute/runtime/CL/functions/CLArithmeticSubtraction.h b/arm_compute/runtime/CL/functions/CLArithmeticSubtraction.h
index b9690806d7..2940044ed9 100644
--- a/arm_compute/runtime/CL/functions/CLArithmeticSubtraction.h
+++ b/arm_compute/runtime/CL/functions/CLArithmeticSubtraction.h
@@ -36,6 +36,10 @@ class ICLTensor;
  *
  * @note The tensor data type for the inputs must be U8/S16/F16/F32.
  * @note The function performs an arithmetic subtraction between two tensors.
+ *
+ *  This function calls the following kernels:
+ * -# @ref CLFillBorderKernel (In case of broadcasting, in the input being broadcasted)
+ * -# @ref CLArithmeticSubtractionKernel
  */
 class CLArithmeticSubtraction : public ICLSimpleFunction
 {
@@ -47,7 +51,7 @@ public:
      * @param[out] output Output tensor. Data types supported: U8 (Only if both inputs are U8), S16/F16/F32.
      * @param[in]  policy Policy to use to handle overflow.
      */
-    void configure(const ICLTensor *input1, const ICLTensor *input2, ICLTensor *output, ConvertPolicy policy);
+    void configure(ICLTensor *input1, ICLTensor *input2, ICLTensor *output, ConvertPolicy policy);
     /** Static function to check if given info will lead to a valid configuration of @ref CLArithmeticSubtraction
      *
      * @param[in] input1 First tensor input info. Data types supported: U8/S16/F16/F32.
diff --git a/arm_compute/runtime/NEON/functions/NEArithmeticSubtraction.h b/arm_compute/runtime/NEON/functions/NEArithmeticSubtraction.h
index 9b460c1031..541756cd2c 100644
--- a/arm_compute/runtime/NEON/functions/NEArithmeticSubtraction.h
+++ b/arm_compute/runtime/NEON/functions/NEArithmeticSubtraction.h
@@ -31,7 +31,15 @@ namespace arm_compute
 {
 class ITensor;
 
-/** Basic function to run @ref NEArithmeticSubtractionKernel */
+/** Basic function to run @ref NEArithmeticSubtractionKernel
+ *
+ * @note The tensor data type for the inputs must be U8/S16/F16/F32.
+ * @note The function performs an arithmetic subtraction between two tensors.
+ *
+ *  This function calls the following kernels:
+ * -# @ref NEFillBorderKernel (In case of broadcasting, in the input being broadcasted)
+ * -# @ref NEArithmeticSubtractionKernel
+ */
 class NEArithmeticSubtraction : public INESimpleFunction
 {
 public:
@@ -42,7 +50,7 @@ public:
      * @param[out] output Output tensor. Data types supported: U8/S16/F16/F32
      * @param[in]  policy Policy to use to handle overflow.
      */
-    void configure(const ITensor *input1, const ITensor *input2, ITensor *output, ConvertPolicy policy);
+    void configure(ITensor *input1, ITensor *input2, ITensor *output, ConvertPolicy policy);
     /** Static function to check if given info will lead to a valid configuration of @ref NEArithmeticSubtraction
      *
      * @param[in] input1 First tensor input. Data types supported: U8/S16/F16/F32
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
diff --git a/tests/validation/CL/ArithmeticAddition.cpp b/tests/validation/CL/ArithmeticAddition.cpp
index bd27bc4043..e327769397 100644
--- a/tests/validation/CL/ArithmeticAddition.cpp
+++ b/tests/validation/CL/ArithmeticAddition.cpp
@@ -33,7 +33,7 @@
 #include "tests/framework/Macros.h"
 #include "tests/framework/datasets/Datasets.h"
 #include "tests/validation/Validation.h"
-#include "tests/validation/fixtures/ArithmeticAdditionFixture.h"
+#include "tests/validation/fixtures/ArithmeticOperationsFixture.h"
 
 namespace arm_compute
 {
diff --git a/tests/validation/CL/ArithmeticSubtraction.cpp b/tests/validation/CL/ArithmeticSubtraction.cpp
index b19d963515..cd13f42ec4 100644
--- a/tests/validation/CL/ArithmeticSubtraction.cpp
+++ b/tests/validation/CL/ArithmeticSubtraction.cpp
@@ -18,7 +18,7 @@
  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
  * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
- * OUT OF OR IN CONCLCTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  * SOFTWARE.
  */
 #include "arm_compute/core/Types.h"
@@ -33,7 +33,7 @@
 #include "tests/framework/Macros.h"
 #include "tests/framework/datasets/Datasets.h"
 #include "tests/validation/Validation.h"
-#include "tests/validation/fixtures/ArithmeticSubtractionFixture.h"
+#include "tests/validation/fixtures/ArithmeticOperationsFixture.h"
 
 namespace arm_compute
 {
@@ -47,14 +47,11 @@ namespace
 const auto ArithmeticSubtractionU8Dataset = combine(combine(framework::dataset::make("DataType", DataType::U8), framework::dataset::make("DataType", DataType::U8)),
                                                     framework::dataset::make("DataType",
                                                                              DataType::U8));
-const auto ArithmeticSubtractionS16Dataset = combine(combine(framework::dataset::make("DataType", DataType::S16), framework::dataset::make("DataType", DataType::S16)),
+const auto ArithmeticSubtractionQASYMM8Dataset = combine(combine(framework::dataset::make("DataType", DataType::QASYMM8), framework::dataset::make("DataType", DataType::QASYMM8)),
+                                                         framework::dataset::make("DataType",
+                                                                                  DataType::QASYMM8));
+const auto ArithmeticSubtractionS16Dataset = combine(combine(framework::dataset::make("DataType", { DataType::U8, DataType::S16 }), framework::dataset::make("DataType", DataType::S16)),
                                                      framework::dataset::make("DataType", DataType::S16));
-const auto ArithmeticSubtractionU8U8S16Dataset = combine(combine(framework::dataset::make("DataType", DataType::U8), framework::dataset::make("DataType", DataType::U8)),
-                                                         framework::dataset::make("DataType", DataType::S16));
-const auto ArithmeticSubtractionS16U8S16Dataset = combine(combine(framework::dataset::make("DataType", DataType::S16), framework::dataset::make("DataType", DataType::U8)),
-                                                          framework::dataset::make("DataType", DataType::S16));
-const auto ArithmeticSubtractionU8S16S16Dataset = combine(combine(framework::dataset::make("DataType", DataType::U8), framework::dataset::make("DataType", DataType::S16)),
-                                                          framework::dataset::make("DataType", DataType::S16));
 const auto ArithmeticSubtractionFP16Dataset = combine(combine(framework::dataset::make("DataType", DataType::F16), framework::dataset::make("DataType", DataType::F16)),
                                                       framework::dataset::make("DataType", DataType::F16));
 const auto ArithmeticSubtractionFP32Dataset = combine(combine(framework::dataset::make("DataType", DataType::F32), framework::dataset::make("DataType", DataType::F32)),
@@ -67,34 +64,34 @@ TEST_SUITE(ArithmeticSubtraction)
 // *INDENT-OFF*
 // clang-format off
 DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(
-               framework::dataset::make("Input1Info", { TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::U8),
-                                                        TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::U8),
-                                                        TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::U8),      // Window shrink
-                                                        TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::U8),      // Invalid data type combination
-                                                        TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F32),     // Mismatching shapes
-                                                      }),
-               framework::dataset::make("Input2Info",{ TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::U8),
-                                                       TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::U8),
-                                                       TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::U8),
-                                                       TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::S16),
-                                                       TensorInfo(TensorShape(48U, 11U, 2U), 1, DataType::F32),
-                                                     })),
-               framework::dataset::make("OutputInfo",{ TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::S16),
-                                                       TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::U8),
-                                                       TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::U8),
-                                                       TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::U8),
-                                                       TensorInfo(TensorShape(48U, 11U, 2U), 1, DataType::F32),
-                                                     })),
-               framework::dataset::make("Expected", { true, true, false, false, false, false, true })),
-               input1_info, input2_info, output_info, expected)
+        framework::dataset::make("Input1Info", { TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::U8),
+                                                 TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::U8),
+                                                 TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::U8),      // Window shrink
+                                                 TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::U8),      // Invalid data type combination
+                                                 TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F32),     // Mismatching shapes
+        }),
+        framework::dataset::make("Input2Info",{ TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::U8),
+                                                TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::U8),
+                                                TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::U8),
+                                                TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::S16),
+                                                TensorInfo(TensorShape(48U, 11U, 2U), 1, DataType::F32),
+        })),
+        framework::dataset::make("OutputInfo",{ TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::S16),
+                                                TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::U8),
+                                                TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::U8),
+                                                TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::U8),
+                                                TensorInfo(TensorShape(48U, 11U, 2U), 1, DataType::F32),
+        })),
+        framework::dataset::make("Expected", { true, true, false, false, false})),
+        input1_info, input2_info, output_info, expected)
 {
     ARM_COMPUTE_EXPECT(bool(CLArithmeticSubtraction::validate(&input1_info.clone()->set_is_resizable(false), &input2_info.clone()->set_is_resizable(false), &output_info.clone()->set_is_resizable(false), ConvertPolicy::WRAP)) == expected, framework::LogLevel::ERRORS);
 }
 // clang-format on
 // *INDENT-ON*
 
-template <typename T1, typename T2 = T1, typename T3 = T1>
-using CLArithmeticSubtractionFixture = ArithmeticSubtractionValidationFixture<CLTensor, CLAccessor, CLArithmeticSubtraction, T1, T2, T3>;
+template <typename T>
+using CLArithmeticSubtractionFixture = ArithmeticSubtractionValidationFixture<CLTensor, CLAccessor, CLArithmeticSubtraction, T>;
 
 TEST_SUITE(U8)
 DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(framework::dataset::concat(datasets::SmallShapes(), datasets::LargeShapes()), framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })),
@@ -120,33 +117,30 @@ DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(framework::da
     validate(dst.info()->padding(), padding);
 }
 
-FIXTURE_DATA_TEST_CASE(RunSmall, CLArithmeticSubtractionFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallShapes(),
-                                                                                                                     ArithmeticSubtractionU8Dataset),
+FIXTURE_DATA_TEST_CASE(RunSmall, CLArithmeticSubtractionFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallShapes(), ArithmeticSubtractionU8Dataset),
                                                                                                                      framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })))
 {
     // Validate output
     validate(CLAccessor(_target), _reference);
 }
-TEST_SUITE_END()
+TEST_SUITE_END() // U8
 
-template <typename T1, typename T2 = T1>
-using CLArithmeticSubtractionToS16Fixture = CLArithmeticSubtractionFixture<T1, T2, int16_t>;
+template <typename T>
+using CLArithmeticSubtractionQuantizedFixture = ArithmeticSubtractionValidationQuantizedFixture<CLTensor, CLAccessor, CLArithmeticSubtraction, T>;
 
-TEST_SUITE(S16)
-DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(combine(combine(framework::dataset::concat(datasets::SmallShapes(), datasets::LargeShapes()),
-                                                                                   framework::dataset::make("DataType", { DataType::U8, DataType::S16 })),
-                                                                           framework::dataset::make("DataType", { DataType::U8, DataType::S16 })),
-                                                                   framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })),
-               shape, data_type1, data_type2, policy)
+TEST_SUITE(Quantized)
+TEST_SUITE(QASYMM8)
+DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(framework::dataset::concat(datasets::SmallShapes(), datasets::LargeShapes()), framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })),
+               shape, policy)
 {
     // Create tensors
-    CLTensor ref_src1 = create_tensor<CLTensor>(shape, data_type1);
-    CLTensor ref_src2 = create_tensor<CLTensor>(shape, data_type2);
-    CLTensor dst      = create_tensor<CLTensor>(shape, DataType::S16);
+    CLTensor ref_src1 = create_tensor<CLTensor>(shape, DataType::QASYMM8);
+    CLTensor ref_src2 = create_tensor<CLTensor>(shape, DataType::QASYMM8);
+    CLTensor dst      = create_tensor<CLTensor>(shape, DataType::QASYMM8);
 
     // Create and Configure function
-    CLArithmeticSubtraction sub;
-    sub.configure(&ref_src1, &ref_src2, &dst, policy);
+    CLArithmeticSubtraction add;
+    add.configure(&ref_src1, &ref_src2, &dst, policy);
 
     // Validate valid region
     const ValidRegion valid_region = shape_to_valid_region(shape);
@@ -158,78 +152,61 @@ DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(combine(combi
     validate(ref_src2.info()->padding(), padding);
     validate(dst.info()->padding(), padding);
 }
-TEST_SUITE(S16_S16_S16)
-FIXTURE_DATA_TEST_CASE(RunSmall, CLArithmeticSubtractionToS16Fixture<int16_t>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallShapes(), ArithmeticSubtractionS16Dataset),
-                       framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })))
-{
-    // Validate output
-    validate(CLAccessor(_target), _reference);
-}
 
-FIXTURE_DATA_TEST_CASE(RunLarge, CLArithmeticSubtractionToS16Fixture<int16_t>, framework::DatasetMode::NIGHTLY, combine(combine(datasets::LargeShapes(), ArithmeticSubtractionS16Dataset),
-                                                                                                                        framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })))
-{
-    // Validate output
-    validate(CLAccessor(_target), _reference);
-}
-TEST_SUITE_END()
+FIXTURE_DATA_TEST_CASE(RunSmall, CLArithmeticSubtractionQuantizedFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(combine(datasets::SmallShapes(),
+                       ArithmeticSubtractionQASYMM8Dataset),
+                       framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })),
+                       framework::dataset::make("QuantizationInfo", { QuantizationInfo(5.f / 255.f, 20) })),
+                       framework::dataset::make("QuantizationInfo", { QuantizationInfo(2.f / 255.f, 10) })),
+                       framework::dataset::make("QuantizationInfo", { QuantizationInfo(1.f / 255.f, 5) }))
 
-TEST_SUITE(U8_U8_S16)
-FIXTURE_DATA_TEST_CASE(RunSmall, CLArithmeticSubtractionToS16Fixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallShapes(),
-                       ArithmeticSubtractionU8U8S16Dataset),
-                       framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })))
+                      )
 {
     // Validate output
     validate(CLAccessor(_target), _reference);
 }
+TEST_SUITE_END() // QASYMM8
+TEST_SUITE_END() // Quantized
 
-FIXTURE_DATA_TEST_CASE(RunLarge, CLArithmeticSubtractionToS16Fixture<uint8_t>, framework::DatasetMode::NIGHTLY, combine(combine(datasets::LargeShapes(),
-                                                                                                                        ArithmeticSubtractionU8U8S16Dataset),
-                                                                                                                        framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })))
+TEST_SUITE(S16)
+DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(combine(framework::dataset::concat(datasets::SmallShapes(), datasets::LargeShapes()), framework::dataset::make("DataType", { DataType::U8, DataType::S16 })),
+                                                                   framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })),
+               shape, data_type, policy)
 {
-    // Validate output
-    validate(CLAccessor(_target), _reference);
-}
-TEST_SUITE_END()
+    // Create tensors
+    CLTensor ref_src1 = create_tensor<CLTensor>(shape, data_type);
+    CLTensor ref_src2 = create_tensor<CLTensor>(shape, DataType::S16);
+    CLTensor dst      = create_tensor<CLTensor>(shape, DataType::S16);
 
-TEST_SUITE(S16_U8_S16)
-using CLAriSubS16U8ToS16Fixture = CLArithmeticSubtractionToS16Fixture<int16_t, uint8_t>;
-FIXTURE_DATA_TEST_CASE(RunSmall, CLAriSubS16U8ToS16Fixture, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallShapes(),
-                                                                                                               ArithmeticSubtractionS16U8S16Dataset),
-                                                                                                       framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })))
-{
-    // Validate output
-    validate(CLAccessor(_target), _reference);
-}
+    // Create and Configure function
+    CLArithmeticSubtraction sub;
+    sub.configure(&ref_src1, &ref_src2, &dst, policy);
 
-FIXTURE_DATA_TEST_CASE(RunLarge, CLAriSubS16U8ToS16Fixture, framework::DatasetMode::NIGHTLY, combine(combine(datasets::LargeShapes(),
-                                                                                                             ArithmeticSubtractionS16U8S16Dataset),
-                                                                                                     framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })))
-{
-    // Validate output
-    validate(CLAccessor(_target), _reference);
+    // Validate valid region
+    const ValidRegion valid_region = shape_to_valid_region(shape);
+    validate(dst.info()->valid_region(), valid_region);
+
+    // Validate padding
+    const PaddingSize padding = PaddingCalculator(shape.x(), 16).required_padding();
+    validate(ref_src1.info()->padding(), padding);
+    validate(ref_src2.info()->padding(), padding);
+    validate(dst.info()->padding(), padding);
 }
-TEST_SUITE_END()
 
-TEST_SUITE(U8_S16_S16)
-using CLAriSubU8S16ToS16Fixture = CLArithmeticSubtractionToS16Fixture<uint8_t, int16_t>;
-FIXTURE_DATA_TEST_CASE(RunSmall, CLAriSubU8S16ToS16Fixture, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallShapes(),
-                                                                                                               ArithmeticSubtractionU8S16S16Dataset),
-                                                                                                       framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })))
+FIXTURE_DATA_TEST_CASE(RunSmall, CLArithmeticSubtractionFixture<int16_t>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallShapes(), ArithmeticSubtractionS16Dataset),
+                                                                                                                     framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })))
 {
     // Validate output
     validate(CLAccessor(_target), _reference);
 }
 
-FIXTURE_DATA_TEST_CASE(RunLarge, CLAriSubU8S16ToS16Fixture, framework::DatasetMode::NIGHTLY, combine(combine(datasets::LargeShapes(),
-                                                                                                             ArithmeticSubtractionU8S16S16Dataset),
-                                                                                                     framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })))
+FIXTURE_DATA_TEST_CASE(RunLarge, CLArithmeticSubtractionFixture<int16_t>, framework::DatasetMode::NIGHTLY, combine(combine(datasets::LargeShapes(), ArithmeticSubtractionS16Dataset),
+                                                                                                                   framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })))
 {
     // Validate output
     validate(CLAccessor(_target), _reference);
 }
-TEST_SUITE_END()
-TEST_SUITE_END()
+TEST_SUITE_END() // S16
 
 TEST_SUITE(Float)
 TEST_SUITE(FP16)
@@ -239,7 +216,7 @@ FIXTURE_DATA_TEST_CASE(RunSmall, CLArithmeticSubtractionFixture<half>, framework
     // Validate output
     validate(CLAccessor(_target), _reference);
 }
-TEST_SUITE_END()
+TEST_SUITE_END() // FP16
 
 TEST_SUITE(FP32)
 DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(framework::dataset::concat(datasets::SmallShapes(), datasets::LargeShapes()), framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })),
@@ -278,11 +255,30 @@ FIXTURE_DATA_TEST_CASE(RunLarge, CLArithmeticSubtractionFixture<float>, framewor
     // Validate output
     validate(CLAccessor(_target), _reference);
 }
-TEST_SUITE_END()
-TEST_SUITE_END()
 
-TEST_SUITE_END()
-TEST_SUITE_END()
+template <typename T>
+using CLArithmeticSubtractionBroadcastFixture = ArithmeticSubtractionBroadcastValidationFixture<CLTensor, CLAccessor, CLArithmeticSubtraction, T>;
+
+FIXTURE_DATA_TEST_CASE(RunSmallBroadcast, CLArithmeticSubtractionBroadcastFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallShapesBroadcast(),
+                       ArithmeticSubtractionFP32Dataset),
+                       framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference);
+}
+
+FIXTURE_DATA_TEST_CASE(RunLargeBroadcast, CLArithmeticSubtractionBroadcastFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(datasets::LargeShapesBroadcast(),
+                       ArithmeticSubtractionFP32Dataset),
+                       framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference);
+}
+TEST_SUITE_END() // FP32
+TEST_SUITE_END() // Float
+
+TEST_SUITE_END() // ArithmeticSubtraction
+TEST_SUITE_END() // CL
 } // namespace validation
 } // namespace test
-} // namespace arm_compute
+} // namespace arm_compute
\ No newline at end of file
diff --git a/tests/validation/GLES_COMPUTE/ArithmeticAddition.cpp b/tests/validation/GLES_COMPUTE/ArithmeticAddition.cpp
index b16503e079..82946fa209 100755
--- a/tests/validation/GLES_COMPUTE/ArithmeticAddition.cpp
+++ b/tests/validation/GLES_COMPUTE/ArithmeticAddition.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017 ARM Limited.
+ * Copyright (c) 2017-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -33,7 +33,7 @@
 #include "tests/framework/Macros.h"
 #include "tests/framework/datasets/Datasets.h"
 #include "tests/validation/Validation.h"
-#include "tests/validation/fixtures/ArithmeticAdditionFixture.h"
+#include "tests/validation/fixtures/ArithmeticOperationsFixture.h"
 
 namespace arm_compute
 {
diff --git a/tests/validation/NEON/ArithmeticAddition.cpp b/tests/validation/NEON/ArithmeticAddition.cpp
index 3632c3c207..4dc56dce54 100644
--- a/tests/validation/NEON/ArithmeticAddition.cpp
+++ b/tests/validation/NEON/ArithmeticAddition.cpp
@@ -33,7 +33,7 @@
 #include "tests/framework/Macros.h"
 #include "tests/framework/datasets/Datasets.h"
 #include "tests/validation/Validation.h"
-#include "tests/validation/fixtures/ArithmeticAdditionFixture.h"
+#include "tests/validation/fixtures/ArithmeticOperationsFixture.h"
 
 namespace arm_compute
 {
diff --git a/tests/validation/NEON/ArithmeticSubtraction.cpp b/tests/validation/NEON/ArithmeticSubtraction.cpp
index 210ed4578f..2e21c14ebc 100644
--- a/tests/validation/NEON/ArithmeticSubtraction.cpp
+++ b/tests/validation/NEON/ArithmeticSubtraction.cpp
@@ -33,7 +33,7 @@
 #include "tests/framework/Macros.h"
 #include "tests/framework/datasets/Datasets.h"
 #include "tests/validation/Validation.h"
-#include "tests/validation/fixtures/ArithmeticSubtractionFixture.h"
+#include "tests/validation/fixtures/ArithmeticOperationsFixture.h"
 
 namespace arm_compute
 {
@@ -49,12 +49,6 @@ const auto ArithmeticSubtractionU8Dataset = combine(combine(framework::dataset::
                                                                              DataType::U8));
 const auto ArithmeticSubtractionS16Dataset = combine(combine(framework::dataset::make("DataType", { DataType::U8, DataType::S16 }), framework::dataset::make("DataType", DataType::S16)),
                                                      framework::dataset::make("DataType", DataType::S16));
-const auto ArithmeticSubtractionU8U8S16Dataset = combine(combine(framework::dataset::make("DataType", DataType::U8), framework::dataset::make("DataType", DataType::U8)),
-                                                         framework::dataset::make("DataType", DataType::S16));
-const auto ArithmeticSubtractionS16U8S16Dataset = combine(combine(framework::dataset::make("DataType", DataType::S16), framework::dataset::make("DataType", DataType::U8)),
-                                                          framework::dataset::make("DataType", DataType::S16));
-const auto ArithmeticSubtractionU8S16S16Dataset = combine(combine(framework::dataset::make("DataType", DataType::U8), framework::dataset::make("DataType", DataType::S16)),
-                                                          framework::dataset::make("DataType", DataType::S16));
 #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
 const auto ArithmeticSubtractionFP16Dataset = combine(combine(framework::dataset::make("DataType", DataType::F16), framework::dataset::make("DataType", DataType::F16)),
                                                       framework::dataset::make("DataType", DataType::F16));
@@ -66,38 +60,38 @@ const auto ArithmeticSubtractionFP32Dataset = combine(combine(framework::dataset
 TEST_SUITE(NEON)
 TEST_SUITE(ArithmeticSubtraction)
 
+template <typename T>
+using NEArithmeticSubtractionFixture = ArithmeticSubtractionValidationFixture<Tensor, Accessor, NEArithmeticSubtraction, T>;
+
 // *INDENT-OFF*
 // clang-format off
 DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(
-               framework::dataset::make("Input1Info", { TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::U8),
-                                                        TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::U8),
-                                                        TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::U8),      // Window shrink
-                                                        TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::U8),      // Invalid data type combination
-                                                        TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F32),     // Mismatching shapes
-                                                      }),
-               framework::dataset::make("Input2Info",{ TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::U8),
-                                                       TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::U8),
-                                                       TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::U8),
-                                                       TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::S16),
-                                                       TensorInfo(TensorShape(48U, 11U, 2U), 1, DataType::F32),
-                                                     })),
-               framework::dataset::make("OutputInfo",{ TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::S16),
-                                                       TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::U8),
-                                                       TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::U8),
-                                                       TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::U8),
-                                                       TensorInfo(TensorShape(48U, 11U, 2U), 1, DataType::F32),
-                                                     })),
-               framework::dataset::make("Expected", { true, true, false, false, false})),
-               input1_info, input2_info, output_info, expected)
+        framework::dataset::make("Input1Info", { TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::U8),
+                                                 TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::U8),
+                                                 TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::U8),      // Window shrink
+                                                 TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::U8),      // Invalid data type combination
+                                                 TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F32),     // Mismatching shapes
+        }),
+        framework::dataset::make("Input2Info",{ TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::U8),
+                                                TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::U8),
+                                                TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::U8),
+                                                TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::S16),
+                                                TensorInfo(TensorShape(48U, 11U, 2U), 1, DataType::F32),
+        })),
+        framework::dataset::make("OutputInfo",{ TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::S16),
+                                                TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::U8),
+                                                TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::U8),
+                                                TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::U8),
+                                                TensorInfo(TensorShape(48U, 11U, 2U), 1, DataType::F32),
+        })),
+        framework::dataset::make("Expected", { true, true, false, false, false})),
+        input1_info, input2_info, output_info, expected)
 {
     ARM_COMPUTE_EXPECT(bool(NEArithmeticSubtraction::validate(&input1_info.clone()->set_is_resizable(false), &input2_info.clone()->set_is_resizable(false), &output_info.clone()->set_is_resizable(false), ConvertPolicy::WRAP)) == expected, framework::LogLevel::ERRORS);
 }
 // clang-format on
 // *INDENT-ON*
 
-template <typename T1, typename T2 = T1, typename T3 = T1>
-using NEArithmeticSubtractionFixture = ArithmeticSubtractionValidationFixture<Tensor, Accessor, NEArithmeticSubtraction, T1, T2, T3>;
-
 TEST_SUITE(U8)
 DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(framework::dataset::concat(datasets::SmallShapes(), datasets::LargeShapes()), framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })),
                shape, policy)
@@ -128,21 +122,16 @@ FIXTURE_DATA_TEST_CASE(RunSmall, NEArithmeticSubtractionFixture<uint8_t>, framew
     // Validate output
     validate(Accessor(_target), _reference);
 }
-TEST_SUITE_END()
-
-template <typename T1, typename T2 = T1>
-using NEArithmeticSubtractionToS16Fixture = NEArithmeticSubtractionFixture<T1, T2, int16_t>;
+TEST_SUITE_END() // U8
 
 TEST_SUITE(S16)
-DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(combine(combine(framework::dataset::concat(datasets::SmallShapes(), datasets::LargeShapes()),
-                                                                                   framework::dataset::make("DataType", { DataType::U8, DataType::S16 })),
-                                                                           framework::dataset::make("DataType", { DataType::U8, DataType::S16 })),
+DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(combine(framework::dataset::concat(datasets::SmallShapes(), datasets::LargeShapes()), framework::dataset::make("DataType", { DataType::U8, DataType::S16 })),
                                                                    framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })),
-               shape, data_type1, data_type2, policy)
+               shape, data_type, policy)
 {
     // Create tensors
-    Tensor ref_src1 = create_tensor<Tensor>(shape, data_type1);
-    Tensor ref_src2 = create_tensor<Tensor>(shape, data_type2);
+    Tensor ref_src1 = create_tensor<Tensor>(shape, data_type);
+    Tensor ref_src2 = create_tensor<Tensor>(shape, DataType::S16);
     Tensor dst      = create_tensor<Tensor>(shape, DataType::S16);
 
     // Create and Configure function
@@ -160,92 +149,34 @@ DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(combine(combi
     validate(dst.info()->padding(), padding);
 }
 
-TEST_SUITE(S16_S16_S16)
-FIXTURE_DATA_TEST_CASE(RunSmall, NEArithmeticSubtractionToS16Fixture<int16_t>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallShapes(), ArithmeticSubtractionS16Dataset),
-                       framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })))
-{
-    // Validate output
-    validate(Accessor(_target), _reference);
-}
-
-FIXTURE_DATA_TEST_CASE(RunLarge, NEArithmeticSubtractionToS16Fixture<int16_t>, framework::DatasetMode::NIGHTLY, combine(combine(datasets::LargeShapes(), ArithmeticSubtractionS16Dataset),
-                                                                                                                        framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })))
-{
-    // Validate output
-    validate(Accessor(_target), _reference);
-}
-TEST_SUITE_END()
-
-TEST_SUITE(U8_U8_S16)
-FIXTURE_DATA_TEST_CASE(RunSmall, NEArithmeticSubtractionToS16Fixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallShapes(),
-                       ArithmeticSubtractionU8U8S16Dataset),
-                       framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })))
-{
-    // Validate output
-    validate(Accessor(_target), _reference);
-}
-
-FIXTURE_DATA_TEST_CASE(RunLarge, NEArithmeticSubtractionToS16Fixture<uint8_t>, framework::DatasetMode::NIGHTLY, combine(combine(datasets::LargeShapes(),
-                                                                                                                        ArithmeticSubtractionU8U8S16Dataset),
-                                                                                                                        framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })))
-{
-    // Validate output
-    validate(Accessor(_target), _reference);
-}
-TEST_SUITE_END()
-
-TEST_SUITE(S16_U8_S16)
-using NEAriSubS16U8ToS16Fixture = NEArithmeticSubtractionToS16Fixture<int16_t, uint8_t>;
-FIXTURE_DATA_TEST_CASE(RunSmall, NEAriSubS16U8ToS16Fixture, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallShapes(),
-                                                                                                               ArithmeticSubtractionS16U8S16Dataset),
-                                                                                                       framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })))
-{
-    // Validate output
-    validate(Accessor(_target), _reference);
-}
-
-FIXTURE_DATA_TEST_CASE(RunLarge, NEAriSubS16U8ToS16Fixture, framework::DatasetMode::NIGHTLY, combine(combine(datasets::LargeShapes(),
-                                                                                                             ArithmeticSubtractionS16U8S16Dataset),
-                                                                                                     framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })))
-{
-    // Validate output
-    validate(Accessor(_target), _reference);
-}
-TEST_SUITE_END()
-
-TEST_SUITE(U8_S16_S16)
-using NEAriSubU8S16ToS16Fixture = NEArithmeticSubtractionToS16Fixture<uint8_t, int16_t>;
-FIXTURE_DATA_TEST_CASE(RunSmall, NEAriSubU8S16ToS16Fixture, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallShapes(),
-                                                                                                               ArithmeticSubtractionU8S16S16Dataset),
-                                                                                                       framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })))
+FIXTURE_DATA_TEST_CASE(RunSmall, NEArithmeticSubtractionFixture<int16_t>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallShapes(), ArithmeticSubtractionS16Dataset),
+                                                                                                                     framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })))
 {
     // Validate output
     validate(Accessor(_target), _reference);
 }
 
-FIXTURE_DATA_TEST_CASE(RunLarge, NEAriSubU8S16ToS16Fixture, framework::DatasetMode::NIGHTLY, combine(combine(datasets::LargeShapes(),
-                                                                                                             ArithmeticSubtractionU8S16S16Dataset),
-                                                                                                     framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })))
+FIXTURE_DATA_TEST_CASE(RunLarge, NEArithmeticSubtractionFixture<int16_t>, framework::DatasetMode::NIGHTLY, combine(combine(datasets::LargeShapes(), ArithmeticSubtractionS16Dataset),
+                                                                                                                   framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })))
 {
     // Validate output
     validate(Accessor(_target), _reference);
 }
-TEST_SUITE_END()
-TEST_SUITE_END()
+TEST_SUITE_END() // S16
 
 TEST_SUITE(Float)
 #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
-TEST_SUITE(FP16)
+TEST_SUITE(F16)
 FIXTURE_DATA_TEST_CASE(RunSmall, NEArithmeticSubtractionFixture<half>, framework::DatasetMode::ALL, combine(combine(datasets::SmallShapes(), ArithmeticSubtractionFP16Dataset),
                                                                                                             framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })))
 {
     // Validate output
     validate(Accessor(_target), _reference);
 }
-TEST_SUITE_END()
-#endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */
+TEST_SUITE_END() // F16
+#endif           /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */
 
-TEST_SUITE(FP32)
+TEST_SUITE(F32)
 DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(framework::dataset::concat(datasets::SmallShapes(), datasets::LargeShapes()), framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })),
                shape, policy)
 {
@@ -282,11 +213,30 @@ FIXTURE_DATA_TEST_CASE(RunLarge, NEArithmeticSubtractionFixture<float>, framewor
     // Validate output
     validate(Accessor(_target), _reference);
 }
-TEST_SUITE_END()
-TEST_SUITE_END()
 
-TEST_SUITE_END()
-TEST_SUITE_END()
+template <typename T>
+using NEArithmeticSubtractionBroadcastFixture = ArithmeticSubtractionBroadcastValidationFixture<Tensor, Accessor, NEArithmeticSubtraction, T>;
+
+FIXTURE_DATA_TEST_CASE(RunSmallBroadcast, NEArithmeticSubtractionBroadcastFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallShapesBroadcast(),
+                       ArithmeticSubtractionFP32Dataset),
+                       framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })))
+{
+    // Validate output
+    validate(Accessor(_target), _reference);
+}
+
+FIXTURE_DATA_TEST_CASE(RunLargeBroadcast, NEArithmeticSubtractionBroadcastFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(datasets::LargeShapesBroadcast(),
+                       ArithmeticSubtractionFP32Dataset),
+                       framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })))
+{
+    // Validate output
+    validate(Accessor(_target), _reference);
+}
+TEST_SUITE_END() // F32
+TEST_SUITE_END() // Float
+
+TEST_SUITE_END() // ArithmeticSubtraction
+TEST_SUITE_END() // NEON
 } // namespace validation
 } // namespace test
 } // namespace arm_compute
diff --git a/tests/validation/fixtures/ArithmeticAdditionFixture.h b/tests/validation/fixtures/ArithmeticOperationsFixture.h
index 66875934de..fb46a5185e 100644
--- a/tests/validation/fixtures/ArithmeticAdditionFixture.h
+++ b/tests/validation/fixtures/ArithmeticOperationsFixture.h
@@ -21,8 +21,8 @@
  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  * SOFTWARE.
  */
-#ifndef ARM_COMPUTE_TEST_ARITHMETIC_ADDITION_FIXTURE
-#define ARM_COMPUTE_TEST_ARITHMETIC_ADDITION_FIXTURE
+#ifndef ARM_COMPUTE_TEST_ARITHMETIC_OPERATIONS_FIXTURE
+#define ARM_COMPUTE_TEST_ARITHMETIC_OPERATIONS_FIXTURE
 
 #include "arm_compute/core/TensorShape.h"
 #include "arm_compute/core/Types.h"
@@ -32,7 +32,7 @@
 #include "tests/framework/Asserts.h"
 #include "tests/framework/Fixture.h"
 #include "tests/validation/Helpers.h"
-#include "tests/validation/reference/ArithmeticAddition.h"
+#include "tests/validation/reference/ArithmeticOperations.h"
 
 namespace arm_compute
 {
@@ -41,13 +41,15 @@ namespace test
 namespace validation
 {
 template <typename TensorType, typename AccessorType, typename FunctionType, typename T>
-class ArithmeticAdditionGenericFixture : public framework::Fixture
+class ArithmeticOperationGenericFixture : public framework::Fixture
 {
 public:
     template <typename...>
-    void setup(const TensorShape &shape0, const TensorShape &shape1, DataType data_type0, DataType data_type1, DataType output_data_type, ConvertPolicy convert_policy,
+    void setup(reference::ArithmeticOperation op, const TensorShape &shape0, const TensorShape &shape1,
+               DataType data_type0, DataType data_type1, DataType output_data_type, ConvertPolicy convert_policy,
                QuantizationInfo qinfo0, QuantizationInfo qinfo1, QuantizationInfo qinfo_out)
     {
+        _op        = op;
         _target    = compute_target(shape0, shape1, data_type0, data_type1, output_data_type, convert_policy, qinfo0, qinfo1, qinfo_out);
         _reference = compute_reference(shape0, shape1, data_type0, data_type1, output_data_type, convert_policy, qinfo0, qinfo1, qinfo_out);
     }
@@ -68,8 +70,8 @@ protected:
         TensorType dst      = create_tensor<TensorType>(TensorShape::broadcast_shape(shape0, shape1), output_data_type, 1, qinfo_out);
 
         // Create and configure function
-        FunctionType add;
-        add.configure(&ref_src1, &ref_src2, &dst, convert_policy);
+        FunctionType arith_op;
+        arith_op.configure(&ref_src1, &ref_src2, &dst, convert_policy);
 
         ARM_COMPUTE_EXPECT(ref_src1.info()->is_resizable(), framework::LogLevel::ERRORS);
         ARM_COMPUTE_EXPECT(ref_src2.info()->is_resizable(), framework::LogLevel::ERRORS);
@@ -89,7 +91,7 @@ protected:
         fill(AccessorType(ref_src2), 1);
 
         // Compute function
-        add.run();
+        arith_op.run();
 
         return dst;
     }
@@ -107,39 +109,40 @@ protected:
         fill(ref_src1, 0);
         fill(ref_src2, 1);
 
-        return reference::arithmetic_addition<T>(ref_src1, ref_src2, ref_dst, convert_policy);
+        return reference::arithmetic_operation<T>(_op, ref_src1, ref_src2, ref_dst, convert_policy);
     }
 
-    TensorType      _target{};
-    SimpleTensor<T> _reference{};
+    TensorType                     _target{};
+    SimpleTensor<T>                _reference{};
+    reference::ArithmeticOperation _op{ reference::ArithmeticOperation::ADD };
 };
 
 template <typename TensorType, typename AccessorType, typename FunctionType, typename T>
-class ArithmeticAdditionBroadcastValidationFixture : public ArithmeticAdditionGenericFixture<TensorType, AccessorType, FunctionType, T>
+class ArithmeticAdditionBroadcastValidationFixture : public ArithmeticOperationGenericFixture<TensorType, AccessorType, FunctionType, T>
 {
 public:
     template <typename...>
     void setup(const TensorShape &shape0, const TensorShape &shape1, DataType data_type0, DataType data_type1, DataType output_data_type, ConvertPolicy convert_policy)
     {
-        ArithmeticAdditionGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(shape0, shape1, data_type0, data_type1,
-                                                                                           output_data_type, convert_policy, QuantizationInfo(), QuantizationInfo(), QuantizationInfo());
+        ArithmeticOperationGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(reference::ArithmeticOperation::ADD, shape0, shape1, data_type0, data_type1,
+                                                                                            output_data_type, convert_policy, QuantizationInfo(), QuantizationInfo(), QuantizationInfo());
     }
 };
 
 template <typename TensorType, typename AccessorType, typename FunctionType, typename T>
-class ArithmeticAdditionValidationFixture : public ArithmeticAdditionGenericFixture<TensorType, AccessorType, FunctionType, T>
+class ArithmeticAdditionValidationFixture : public ArithmeticOperationGenericFixture<TensorType, AccessorType, FunctionType, T>
 {
 public:
     template <typename...>
     void setup(const TensorShape &shape, DataType data_type0, DataType data_type1, DataType output_data_type, ConvertPolicy convert_policy)
     {
-        ArithmeticAdditionGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(shape, shape, data_type0, data_type1,
-                                                                                           output_data_type, convert_policy, QuantizationInfo(), QuantizationInfo(), QuantizationInfo());
+        ArithmeticOperationGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(reference::ArithmeticOperation::ADD, shape, shape, data_type0, data_type1,
+                                                                                            output_data_type, convert_policy, QuantizationInfo(), QuantizationInfo(), QuantizationInfo());
     }
 };
 
 template <typename TensorType, typename AccessorType, typename FunctionType, typename T>
-class ArithmeticAdditionValidationQuantizedFixture : public ArithmeticAdditionGenericFixture<TensorType, AccessorType, FunctionType, T>
+class ArithmeticAdditionValidationQuantizedFixture : public ArithmeticOperationGenericFixture<TensorType, AccessorType, FunctionType, T>
 {
 public:
     template <typename...>
@@ -147,11 +150,52 @@ public:
                QuantizationInfo qinfo0, QuantizationInfo qinfo1, QuantizationInfo qinfo_out)
 
     {
-        ArithmeticAdditionGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(shape, shape, data_type0, data_type1,
-                                                                                           output_data_type, convert_policy, qinfo0, qinfo1, qinfo_out);
+        ArithmeticOperationGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(reference::ArithmeticOperation::ADD, shape, shape, data_type0, data_type1,
+                                                                                            output_data_type, convert_policy, qinfo0, qinfo1, qinfo_out);
+    }
+};
+
+template <typename TensorType, typename AccessorType, typename FunctionType, typename T>
+class ArithmeticSubtractionBroadcastValidationFixture : public ArithmeticOperationGenericFixture<TensorType, AccessorType, FunctionType, T>
+{
+public:
+    template <typename...>
+    void setup(const TensorShape &shape0, const TensorShape &shape1, DataType data_type0, DataType data_type1, DataType output_data_type, ConvertPolicy convert_policy)
+    {
+        ArithmeticOperationGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(reference::ArithmeticOperation::SUB, shape0, shape1,
+                                                                                            data_type0, data_type1, output_data_type, convert_policy,
+                                                                                            QuantizationInfo(), QuantizationInfo(), QuantizationInfo());
+    }
+};
+
+template <typename TensorType, typename AccessorType, typename FunctionType, typename T>
+class ArithmeticSubtractionValidationFixture : public ArithmeticOperationGenericFixture<TensorType, AccessorType, FunctionType, T>
+{
+public:
+    template <typename...>
+    void setup(const TensorShape &shape, DataType data_type0, DataType data_type1, DataType output_data_type, ConvertPolicy convert_policy)
+    {
+        ArithmeticOperationGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(reference::ArithmeticOperation::SUB, shape, shape,
+                                                                                            data_type0, data_type1, output_data_type, convert_policy,
+                                                                                            QuantizationInfo(), QuantizationInfo(), QuantizationInfo());
+    }
+};
+
+template <typename TensorType, typename AccessorType, typename FunctionType, typename T>
+class ArithmeticSubtractionValidationQuantizedFixture : public ArithmeticOperationGenericFixture<TensorType, AccessorType, FunctionType, T>
+{
+public:
+    template <typename...>
+    void setup(const TensorShape &shape, DataType data_type0, DataType data_type1, DataType output_data_type, ConvertPolicy convert_policy,
+               QuantizationInfo qinfo0, QuantizationInfo qinfo1, QuantizationInfo qinfo_out)
+
+    {
+        ArithmeticOperationGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(reference::ArithmeticOperation::SUB, shape, shape,
+                                                                                            data_type0, data_type1, output_data_type,
+                                                                                            convert_policy, qinfo0, qinfo1, qinfo_out);
     }
 };
 } // namespace validation
 } // namespace test
 } // namespace arm_compute
-#endif /* ARM_COMPUTE_TEST_ARITHMETIC_ADDITION_FIXTURE */
+#endif /* ARM_COMPUTE_TEST_ARITHMETIC_OPERATIONS_FIXTURE */
diff --git a/tests/validation/fixtures/ArithmeticSubtractionFixture.h b/tests/validation/fixtures/ArithmeticSubtractionFixture.h
deleted file mode 100644
index 9897677691..0000000000
--- a/tests/validation/fixtures/ArithmeticSubtractionFixture.h
+++ /dev/null
@@ -1,114 +0,0 @@
-/*
- * Copyright (c) 2017-2018 ARM Limited.
- *
- * SPDX-License-Identifier: MIT
- *
- * Permission is hereby granted, free of charge, to any person obtaining a copy
- * of this software and associated documentation files (the "Software"), to
- * deal in the Software without restriction, including without limitation the
- * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
- * sell copies of the Software, and to permit persons to whom the Software is
- * furnished to do so, subject to the following conditions:
- *
- * The above copyright notice and this permission notice shall be included in all
- * copies or substantial portions of the Software.
- *
- * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
- * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
- * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
- * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
- * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
- * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
- * SOFTWARE.
- */
-#ifndef ARM_COMPUTE_TEST_ARITHMETIC_SUBTRACTION_FIXTURE
-#define ARM_COMPUTE_TEST_ARITHMETIC_SUBTRACTION_FIXTURE
-
-#include "arm_compute/core/TensorShape.h"
-#include "arm_compute/core/Types.h"
-#include "tests/AssetsLibrary.h"
-#include "tests/Globals.h"
-#include "tests/IAccessor.h"
-#include "tests/framework/Asserts.h"
-#include "tests/framework/Fixture.h"
-#include "tests/validation/Helpers.h"
-#include "tests/validation/reference/ArithmeticSubtraction.h"
-
-namespace arm_compute
-{
-namespace test
-{
-namespace validation
-{
-template <typename TensorType, typename AccessorType, typename FunctionType, typename T1, typename T2 = T1, typename T3 = T1>
-class ArithmeticSubtractionValidationFixture : public framework::Fixture
-{
-public:
-    template <typename...>
-    void setup(TensorShape shape, DataType data_type0, DataType data_type1, DataType output_data_type, ConvertPolicy convert_policy)
-    {
-        _target    = compute_target(shape, data_type0, data_type1, output_data_type, convert_policy);
-        _reference = compute_reference(shape, data_type0, data_type1, output_data_type, convert_policy);
-    }
-
-protected:
-    template <typename U>
-    void fill(U &&tensor, int i)
-    {
-        library->fill_tensor_uniform(tensor, i);
-    }
-
-    TensorType compute_target(const TensorShape &shape, DataType data_type0, DataType data_type1, DataType output_data_type, ConvertPolicy convert_policy)
-    {
-        // Create tensors
-        TensorType ref_src1 = create_tensor<TensorType>(shape, data_type0, 1);
-        TensorType ref_src2 = create_tensor<TensorType>(shape, data_type1, 1);
-        TensorType dst      = create_tensor<TensorType>(shape, output_data_type, 1);
-
-        // Create and configure function
-        FunctionType sub;
-        sub.configure(&ref_src1, &ref_src2, &dst, convert_policy);
-
-        ARM_COMPUTE_EXPECT(ref_src1.info()->is_resizable(), framework::LogLevel::ERRORS);
-        ARM_COMPUTE_EXPECT(ref_src2.info()->is_resizable(), framework::LogLevel::ERRORS);
-        ARM_COMPUTE_EXPECT(dst.info()->is_resizable(), framework::LogLevel::ERRORS);
-
-        // Allocate tensors
-        ref_src1.allocator()->allocate();
-        ref_src2.allocator()->allocate();
-        dst.allocator()->allocate();
-
-        ARM_COMPUTE_EXPECT(!ref_src1.info()->is_resizable(), framework::LogLevel::ERRORS);
-        ARM_COMPUTE_EXPECT(!ref_src2.info()->is_resizable(), framework::LogLevel::ERRORS);
-        ARM_COMPUTE_EXPECT(!dst.info()->is_resizable(), framework::LogLevel::ERRORS);
-
-        // Fill tensors
-        fill(AccessorType(ref_src1), 0);
-        fill(AccessorType(ref_src2), 1);
-
-        // Compute function
-        sub.run();
-
-        return dst;
-    }
-
-    SimpleTensor<T3> compute_reference(const TensorShape &shape, DataType data_type0, DataType data_type1, DataType output_data_type, ConvertPolicy convert_policy)
-    {
-        // Create reference
-        SimpleTensor<T1> ref_src1{ shape, data_type0, 1 };
-        SimpleTensor<T2> ref_src2{ shape, data_type1, 1 };
-
-        // Fill reference
-        fill(ref_src1, 0);
-        fill(ref_src2, 1);
-
-        return reference::arithmetic_subtraction<T1, T2, T3>(ref_src1, ref_src2, output_data_type, convert_policy);
-    }
-
-    TensorType       _target{};
-    SimpleTensor<T3> _reference{};
-};
-} // namespace validation
-} // namespace test
-} // namespace arm_compute
-#endif /* ARM_COMPUTE_TEST_ARITHMETIC_SUBTRACTION_FIXTURE */
diff --git a/tests/validation/fixtures/LSTMLayerFixture.h b/tests/validation/fixtures/LSTMLayerFixture.h
index 20df855242..bc892bfecf 100644
--- a/tests/validation/fixtures/LSTMLayerFixture.h
+++ b/tests/validation/fixtures/LSTMLayerFixture.h
@@ -28,8 +28,7 @@
 #include "tests/framework/Asserts.h"
 #include "tests/framework/Fixture.h"
 #include "tests/validation/reference/ActivationLayer.h"
-#include "tests/validation/reference/ArithmeticAddition.h"
-#include "tests/validation/reference/ArithmeticSubtraction.h"
+#include "tests/validation/reference/ArithmeticOperations.h"
 #include "tests/validation/reference/FullyConnectedLayer.h"
 #include "tests/validation/reference/GEMM.h"
 #include "tests/validation/reference/PixelWiseMultiplication.h"
@@ -333,12 +332,12 @@ protected:
         SimpleTensor<T> fully_connected_forget = reference::fully_connected_layer(input, input_to_forget_w, forget_gate_bias, output_cell_shape);
         SimpleTensor<T> transposed_weights     = reference::transpose(recurrent_to_forget_w);
         SimpleTensor<T> gemm                   = reference::gemm(output_state_in, transposed_weights, cell_state_in, 1.f, 0.f);
-        SimpleTensor<T> forget_gate            = reference::arithmetic_addition(fully_connected_forget, gemm, data_type, ConvertPolicy::SATURATE);
+        SimpleTensor<T> forget_gate            = reference::arithmetic_operation(reference::ArithmeticOperation::ADD, fully_connected_forget, gemm, data_type, ConvertPolicy::SATURATE);
 
         if(peephole_opt)
         {
             SimpleTensor<T> pixelwise_mul_forget_gate = reference::pixel_wise_multiplication(cell_state_in, cell_to_forget_w, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_NEAREST_EVEN);
-            forget_gate                               = reference::arithmetic_addition(forget_gate, pixelwise_mul_forget_gate, data_type, ConvertPolicy::SATURATE);
+            forget_gate                               = reference::arithmetic_operation(reference::ArithmeticOperation::ADD, forget_gate, pixelwise_mul_forget_gate, data_type, ConvertPolicy::SATURATE);
         }
 
         forget_gate = reference::activation_layer(forget_gate, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC));
@@ -349,18 +348,18 @@ protected:
         {
             SimpleTensor<T> ones{ cell_bias_shape, data_type };
             fill_custom_val(ones, 1.f, 0);
-            input_gate = reference::arithmetic_subtraction<T, T, T>(ones, forget_gate, data_type, ConvertPolicy::SATURATE);
+            input_gate = reference::arithmetic_operation<T>(reference::ArithmeticOperation::SUB, ones, forget_gate, data_type, ConvertPolicy::SATURATE);
         }
         else
         {
             SimpleTensor<T> fully_connected_input = reference::fully_connected_layer(input, input_to_input_w, input_gate_bias, output_cell_shape);
             transposed_weights                    = reference::transpose(recurrent_to_input_w);
             gemm                                  = reference::gemm(output_state_in, transposed_weights, cell_state_in, 1.f, 0.f);
-            input_gate                            = reference::arithmetic_addition(fully_connected_input, gemm, data_type, ConvertPolicy::SATURATE);
+            input_gate                            = reference::arithmetic_operation(reference::ArithmeticOperation::ADD, fully_connected_input, gemm, data_type, ConvertPolicy::SATURATE);
             if(peephole_opt)
             {
                 SimpleTensor<T> pixelwise_mul_input_gate = reference::pixel_wise_multiplication(cell_state_in, cell_to_input_w, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_NEAREST_EVEN);
-                input_gate                               = reference::arithmetic_addition(input_gate, pixelwise_mul_input_gate, data_type, ConvertPolicy::SATURATE);
+                input_gate                               = reference::arithmetic_operation(reference::ArithmeticOperation::ADD, input_gate, pixelwise_mul_input_gate, data_type, ConvertPolicy::SATURATE);
             }
             input_gate = reference::activation_layer(input_gate, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC));
         }
@@ -370,10 +369,10 @@ protected:
         transposed_weights                         = reference::transpose(recurrent_to_cell_w);
         gemm                                       = reference::gemm(output_state_in, transposed_weights, cell_state_out, 1.f, 0.f);
         SimpleTensor<T> pixelwise_mul              = reference::pixel_wise_multiplication(cell_state_in, forget_gate, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_NEAREST_EVEN);
-        cell_state_out                             = reference::arithmetic_addition(fully_connected_cell_state, gemm, data_type, ConvertPolicy::SATURATE);
+        cell_state_out                             = reference::arithmetic_operation(reference::ArithmeticOperation::ADD, fully_connected_cell_state, gemm, data_type, ConvertPolicy::SATURATE);
         cell_state_out                             = reference::activation_layer(cell_state_out, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC));
         cell_state_out                             = reference::pixel_wise_multiplication(cell_state_out, input_gate, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_NEAREST_EVEN);
-        cell_state_out                             = reference::arithmetic_addition(cell_state_out, pixelwise_mul, data_type, ConvertPolicy::SATURATE);
+        cell_state_out                             = reference::arithmetic_operation(reference::ArithmeticOperation::ADD, cell_state_out, pixelwise_mul, data_type, ConvertPolicy::SATURATE);
         if(cell_threshold != 0.f)
         {
             cell_state_out = reference::activation_layer(cell_state_out, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, -cell_threshold, cell_threshold));
@@ -383,11 +382,11 @@ protected:
         SimpleTensor<T> fully_connected_output = reference::fully_connected_layer(input, input_to_output_w, output_gate_bias, output_cell_shape);
         transposed_weights                     = reference::transpose(recurrent_to_output_w);
         gemm                                   = reference::gemm(output_state_in, transposed_weights, cell_state_out, 1.f, 0.f);
-        output                                 = reference::arithmetic_addition(fully_connected_output, gemm, data_type, ConvertPolicy::SATURATE);
+        output                                 = reference::arithmetic_operation(reference::ArithmeticOperation::ADD, fully_connected_output, gemm, data_type, ConvertPolicy::SATURATE);
         if(peephole_opt)
         {
             pixelwise_mul = reference::pixel_wise_multiplication(cell_state_out, cell_to_output_w, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_NEAREST_EVEN);
-            output        = reference::arithmetic_addition(output, pixelwise_mul, data_type, ConvertPolicy::SATURATE);
+            output        = reference::arithmetic_operation(reference::ArithmeticOperation::ADD, output, pixelwise_mul, data_type, ConvertPolicy::SATURATE);
         }
         output = reference::activation_layer(output, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC));
 
diff --git a/tests/validation/fixtures/RNNLayerFixture.h b/tests/validation/fixtures/RNNLayerFixture.h
index 42b99cce1c..2645116b44 100644
--- a/tests/validation/fixtures/RNNLayerFixture.h
+++ b/tests/validation/fixtures/RNNLayerFixture.h
@@ -28,7 +28,7 @@
 #include "tests/framework/Asserts.h"
 #include "tests/framework/Fixture.h"
 #include "tests/validation/reference/ActivationLayer.h"
-#include "tests/validation/reference/ArithmeticAddition.h"
+#include "tests/validation/reference/ArithmeticOperations.h"
 #include "tests/validation/reference/FullyConnectedLayer.h"
 #include "tests/validation/reference/GEMM.h"
 
@@ -132,7 +132,7 @@ protected:
         SimpleTensor<T> out_w{ out_shape, data_type };
         SimpleTensor<T> fully_connected = reference::fully_connected_layer(input, weights, bias, out_shape);
         SimpleTensor<T> gemm            = reference::gemm(hidden_state, recurrent_weights, out_w, 1.f, 0.f);
-        SimpleTensor<T> add_res         = reference::arithmetic_addition(fully_connected, gemm, data_type, ConvertPolicy::SATURATE);
+        SimpleTensor<T> add_res         = reference::arithmetic_operation(reference::ArithmeticOperation::ADD, fully_connected, gemm, data_type, ConvertPolicy::SATURATE);
         return reference::activation_layer(add_res, info);
     }
 
diff --git a/tests/validation/reference/ArithmeticAddition.cpp b/tests/validation/reference/ArithmeticOperations.cpp
index c68c6d4ed6..062be93b95 100644
--- a/tests/validation/reference/ArithmeticAddition.cpp
+++ b/tests/validation/reference/ArithmeticOperations.cpp
@@ -21,7 +21,7 @@
  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  * SOFTWARE.
  */
-#include "ArithmeticAddition.h"
+#include "ArithmeticOperations.h"
 
 #include "arm_compute/core/Types.h"
 #include "tests/validation/Helpers.h"
@@ -37,11 +37,12 @@ namespace reference
 namespace
 {
 template <typename T>
-T add(T src1, T src2, ConvertPolicy convert_policy)
+T arithm_op(ArithmeticOperation op, T src1, T src2, ConvertPolicy convert_policy)
 {
     using intermediate_type = typename common_promoted_signed_type<T>::intermediate_type;
 
-    intermediate_type val = static_cast<intermediate_type>(src1) + static_cast<intermediate_type>(src2);
+    intermediate_type val = (op == ArithmeticOperation::ADD) ? static_cast<intermediate_type>(src1) + static_cast<intermediate_type>(src2) : static_cast<intermediate_type>
+                            (src1) - static_cast<intermediate_type>(src2);
 
     T result = (convert_policy == ConvertPolicy::SATURATE) ? saturate_cast<T>(val) : static_cast<T>(val);
 
@@ -52,7 +53,7 @@ template <size_t dim>
 struct BroadcastUnroll
 {
     template <typename T>
-    static void unroll(const SimpleTensor<T> &src1, const SimpleTensor<T> &src2, SimpleTensor<T> &dst,
+    static void unroll(ArithmeticOperation op, const SimpleTensor<T> &src1, const SimpleTensor<T> &src2, SimpleTensor<T> &dst,
                        ConvertPolicy convert_policy, Coordinates &id_src1, Coordinates &id_src2, Coordinates &id_dst)
     {
         const bool src1_is_broadcast = (src1.shape()[dim - 1] != dst.shape()[dim - 1]);
@@ -64,7 +65,7 @@ struct BroadcastUnroll
 
         for(size_t i = 0; i < dst.shape()[dim - 1]; ++i, ++id_dst[dim - 1])
         {
-            BroadcastUnroll < dim - 1 >::unroll(src1, src2, dst, convert_policy, id_src1, id_src2, id_dst);
+            BroadcastUnroll < dim - 1 >::unroll(op, src1, src2, dst, convert_policy, id_src1, id_src2, id_dst);
 
             id_src1[dim - 1] += !src1_is_broadcast;
             id_src2[dim - 1] += !src2_is_broadcast;
@@ -76,26 +77,26 @@ template <>
 struct BroadcastUnroll<0>
 {
     template <typename T>
-    static void unroll(const SimpleTensor<T> &src1, const SimpleTensor<T> &src2, SimpleTensor<T> &dst,
+    static void unroll(ArithmeticOperation op, const SimpleTensor<T> &src1, const SimpleTensor<T> &src2, SimpleTensor<T> &dst,
                        ConvertPolicy convert_policy, Coordinates &id_src1, Coordinates &id_src2, Coordinates &id_dst)
     {
-        dst[coord2index(dst.shape(), id_dst)] = add(src1[coord2index(src1.shape(), id_src1)], src2[coord2index(src2.shape(), id_src2)], convert_policy);
+        dst[coord2index(dst.shape(), id_dst)] = arithm_op(op, src1[coord2index(src1.shape(), id_src1)], src2[coord2index(src2.shape(), id_src2)], convert_policy);
     }
 };
 } // namespace
 
 template <typename T>
-SimpleTensor<T> arithmetic_addition(const SimpleTensor<T> &src1, const SimpleTensor<T> &src2, SimpleTensor<T> &dst, ConvertPolicy convert_policy)
+SimpleTensor<T> arithmetic_operation(ArithmeticOperation op, const SimpleTensor<T> &src1, const SimpleTensor<T> &src2, SimpleTensor<T> &dst, ConvertPolicy convert_policy)
 {
     Coordinates id_src1, id_src2, id_dst;
 
-    BroadcastUnroll<Coordinates::num_max_dimensions>::unroll(src1, src2, dst, convert_policy, id_src1, id_src2, id_dst);
+    BroadcastUnroll<Coordinates::num_max_dimensions>::unroll(op, src1, src2, dst, convert_policy, id_src1, id_src2, id_dst);
 
     return dst;
 }
 
 template <>
-SimpleTensor<uint8_t> arithmetic_addition(const SimpleTensor<uint8_t> &src1, const SimpleTensor<uint8_t> &src2, SimpleTensor<uint8_t> &dst, ConvertPolicy convert_policy)
+SimpleTensor<uint8_t> arithmetic_operation(ArithmeticOperation op, const SimpleTensor<uint8_t> &src1, const SimpleTensor<uint8_t> &src2, SimpleTensor<uint8_t> &dst, ConvertPolicy convert_policy)
 {
     if(dst.data_type() == DataType::QASYMM8)
     {
@@ -105,7 +106,7 @@ SimpleTensor<uint8_t> arithmetic_addition(const SimpleTensor<uint8_t> &src1, con
 
         Coordinates id_src1, id_src2, id_dst;
 
-        BroadcastUnroll<Coordinates::num_max_dimensions>::unroll(src1_tmp, src2_tmp, dst_tmp, convert_policy, id_src1, id_src2, id_dst);
+        BroadcastUnroll<Coordinates::num_max_dimensions>::unroll(op, src1_tmp, src2_tmp, dst_tmp, convert_policy, id_src1, id_src2, id_dst);
 
         dst = convert_to_asymmetric(dst_tmp, dst.quantization_info());
         return dst;
@@ -115,31 +116,32 @@ SimpleTensor<uint8_t> arithmetic_addition(const SimpleTensor<uint8_t> &src1, con
         // DataType::U8
         Coordinates id_src1, id_src2, id_dst;
 
-        BroadcastUnroll<Coordinates::num_max_dimensions>::unroll(src1, src2, dst, convert_policy, id_src1, id_src2, id_dst);
+        BroadcastUnroll<Coordinates::num_max_dimensions>::unroll(op, src1, src2, dst, convert_policy, id_src1, id_src2, id_dst);
 
         return dst;
     }
 }
 
-template SimpleTensor<int16_t> arithmetic_addition(const SimpleTensor<int16_t> &src1, const SimpleTensor<int16_t> &src2, SimpleTensor<int16_t> &dst, ConvertPolicy convert_policy);
-template SimpleTensor<int8_t> arithmetic_addition(const SimpleTensor<int8_t> &src1, const SimpleTensor<int8_t> &src2, SimpleTensor<int8_t> &dst, ConvertPolicy convert_policy);
-template SimpleTensor<half> arithmetic_addition(const SimpleTensor<half> &src1, const SimpleTensor<half> &src2, SimpleTensor<half> &dst, ConvertPolicy convert_policy);
-template SimpleTensor<float> arithmetic_addition(const SimpleTensor<float> &src1, const SimpleTensor<float> &src2, SimpleTensor<float> &dst, ConvertPolicy convert_policy);
+template SimpleTensor<int16_t> arithmetic_operation(ArithmeticOperation op, const SimpleTensor<int16_t> &src1, const SimpleTensor<int16_t> &src2, SimpleTensor<int16_t> &dst,
+                                                    ConvertPolicy convert_policy);
+template SimpleTensor<int8_t> arithmetic_operation(ArithmeticOperation op, const SimpleTensor<int8_t> &src1, const SimpleTensor<int8_t> &src2, SimpleTensor<int8_t> &dst, ConvertPolicy convert_policy);
+template SimpleTensor<half> arithmetic_operation(ArithmeticOperation op, const SimpleTensor<half> &src1, const SimpleTensor<half> &src2, SimpleTensor<half> &dst, ConvertPolicy convert_policy);
+template SimpleTensor<float> arithmetic_operation(ArithmeticOperation op, const SimpleTensor<float> &src1, const SimpleTensor<float> &src2, SimpleTensor<float> &dst, ConvertPolicy convert_policy);
 
 template <typename T>
-SimpleTensor<T> arithmetic_addition(const SimpleTensor<T> &src1, const SimpleTensor<T> &src2, DataType dst_data_type, ConvertPolicy convert_policy)
+SimpleTensor<T> arithmetic_operation(ArithmeticOperation op, const SimpleTensor<T> &src1, const SimpleTensor<T> &src2, DataType dst_data_type, ConvertPolicy convert_policy)
 {
     ARM_COMPUTE_ERROR_ON_MSG(dst_data_type == DataType::QASYMM8, "For QASYMM8, the quantized output tensor should be passed directly.");
 
     SimpleTensor<T> dst(TensorShape::broadcast_shape(src1.shape(), src2.shape()), dst_data_type);
-    arithmetic_addition<T>(src1, src2, dst, convert_policy);
+    arithmetic_operation<T>(op, src1, src2, dst, convert_policy);
     return dst;
 }
 
-template SimpleTensor<int16_t> arithmetic_addition(const SimpleTensor<int16_t> &src1, const SimpleTensor<int16_t> &src2, DataType dst_data_type, ConvertPolicy convert_policy);
-template SimpleTensor<int8_t> arithmetic_addition(const SimpleTensor<int8_t> &src1, const SimpleTensor<int8_t> &src2, DataType dst_data_type, ConvertPolicy convert_policy);
-template SimpleTensor<half> arithmetic_addition(const SimpleTensor<half> &src1, const SimpleTensor<half> &src2, DataType dst_data_type, ConvertPolicy convert_policy);
-template SimpleTensor<float> arithmetic_addition(const SimpleTensor<float> &src1, const SimpleTensor<float> &src2, DataType dst_data_type, ConvertPolicy convert_policy);
+template SimpleTensor<int16_t> arithmetic_operation(ArithmeticOperation op, const SimpleTensor<int16_t> &src1, const SimpleTensor<int16_t> &src2, DataType dst_data_type, ConvertPolicy convert_policy);
+template SimpleTensor<int8_t> arithmetic_operation(ArithmeticOperation op, const SimpleTensor<int8_t> &src1, const SimpleTensor<int8_t> &src2, DataType dst_data_type, ConvertPolicy convert_policy);
+template SimpleTensor<half> arithmetic_operation(ArithmeticOperation op, const SimpleTensor<half> &src1, const SimpleTensor<half> &src2, DataType dst_data_type, ConvertPolicy convert_policy);
+template SimpleTensor<float> arithmetic_operation(ArithmeticOperation op, const SimpleTensor<float> &src1, const SimpleTensor<float> &src2, DataType dst_data_type, ConvertPolicy convert_policy);
 
 } // namespace reference
 } // namespace validation
diff --git a/tests/validation/reference/ArithmeticAddition.h b/tests/validation/reference/ArithmeticOperations.h
index faeabd7a6f..73634823c4 100644
--- a/tests/validation/reference/ArithmeticAddition.h
+++ b/tests/validation/reference/ArithmeticOperations.h
@@ -21,8 +21,8 @@
  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  * SOFTWARE.
  */
-#ifndef __ARM_COMPUTE_TEST_ARITHMETIC_ADDITION_H__
-#define __ARM_COMPUTE_TEST_ARITHMETIC_ADDITION_H__
+#ifndef __ARM_COMPUTE_TEST_ARITHMETIC_OPERATIONS_H__
+#define __ARM_COMPUTE_TEST_ARITHMETIC_OPERATIONS_H__
 
 #include "tests/SimpleTensor.h"
 #include "tests/validation/Helpers.h"
@@ -35,14 +35,20 @@ namespace validation
 {
 namespace reference
 {
-template <typename T>
-SimpleTensor<T> arithmetic_addition(const SimpleTensor<T> &src1, const SimpleTensor<T> &src2, SimpleTensor<T> &dst, ConvertPolicy convert_policy);
+/** Arithmetic operation types */
+enum class ArithmeticOperation
+{
+    ADD,
+    SUB
+};
 
 template <typename T>
-SimpleTensor<T> arithmetic_addition(const SimpleTensor<T> &src1, const SimpleTensor<T> &src2, DataType dst_data_type, ConvertPolicy convert_policy);
+SimpleTensor<T> arithmetic_operation(ArithmeticOperation op, const SimpleTensor<T> &src1, const SimpleTensor<T> &src2, SimpleTensor<T> &dst, ConvertPolicy convert_policy);
 
+template <typename T>
+SimpleTensor<T> arithmetic_operation(ArithmeticOperation op, const SimpleTensor<T> &src1, const SimpleTensor<T> &src2, DataType dst_data_type, ConvertPolicy convert_policy);
 } // namespace reference
 } // namespace validation
 } // namespace test
 } // namespace arm_compute
-#endif /* __ARM_COMPUTE_TEST_ARITHMETIC_ADDITION_H__ */
+#endif /* __ARM_COMPUTE_TEST_ARITHMETIC_OPERATIONS_H__ */
diff --git a/tests/validation/reference/ArithmeticSubtraction.cpp b/tests/validation/reference/ArithmeticSubtraction.cpp
deleted file mode 100644
index f39d01f9e8..0000000000
--- a/tests/validation/reference/ArithmeticSubtraction.cpp
+++ /dev/null
@@ -1,63 +0,0 @@
-/*
- * Copyright (c) 2017-2018 ARM Limited.
- *
- * SPDX-License-Identifier: MIT
- *
- * Permission is hereby granted, free of charge, to any person obtaining a copy
- * of this software and associated documentation files (the "Software"), to
- * deal in the Software without restriction, including without limitation the
- * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
- * sell copies of the Software, and to permit persons to whom the Software is
- * furnished to do so, subject to the following conditions:
- *
- * The above copyright notice and this permission notice shall be included in all
- * copies or substantial portions of the Software.
- *
- * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
- * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
- * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
- * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
- * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
- * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
- * SOFTWARE.
- */
-#include "ArithmeticSubtraction.h"
-
-#include "tests/validation/Helpers.h"
-
-namespace arm_compute
-{
-namespace test
-{
-namespace validation
-{
-namespace reference
-{
-template <typename T1, typename T2, typename T3>
-SimpleTensor<T3> arithmetic_subtraction(const SimpleTensor<T1> &src1, const SimpleTensor<T2> &src2, DataType dst_data_type, ConvertPolicy convert_policy)
-{
-    SimpleTensor<T3> result(src1.shape(), dst_data_type);
-
-    using intermediate_type = typename common_promoted_signed_type<typename std::conditional<sizeof(T1) >= sizeof(T2), T1, T2>::type >::intermediate_type;
-
-    for(int i = 0; i < src1.num_elements(); ++i)
-    {
-        intermediate_type val = static_cast<intermediate_type>(src1[i]) - static_cast<intermediate_type>(src2[i]);
-        result[i]             = (convert_policy == ConvertPolicy::SATURATE) ? saturate_cast<T3>(val) : static_cast<T3>(val);
-    }
-
-    return result;
-}
-
-template SimpleTensor<uint8_t> arithmetic_subtraction(const SimpleTensor<uint8_t> &src1, const SimpleTensor<uint8_t> &src2, DataType dst_data_type, ConvertPolicy convert_policy);
-template SimpleTensor<int16_t> arithmetic_subtraction(const SimpleTensor<uint8_t> &src1, const SimpleTensor<uint8_t> &src2, DataType dst_data_type, ConvertPolicy convert_policy);
-template SimpleTensor<int16_t> arithmetic_subtraction(const SimpleTensor<uint8_t> &src1, const SimpleTensor<int16_t> &src2, DataType dst_data_type, ConvertPolicy convert_policy);
-template SimpleTensor<int16_t> arithmetic_subtraction(const SimpleTensor<int16_t> &src1, const SimpleTensor<uint8_t> &src2, DataType dst_data_type, ConvertPolicy convert_policy);
-template SimpleTensor<int16_t> arithmetic_subtraction(const SimpleTensor<int16_t> &src1, const SimpleTensor<int16_t> &src2, DataType dst_data_type, ConvertPolicy convert_policy);
-template SimpleTensor<int8_t> arithmetic_subtraction(const SimpleTensor<int8_t> &src1, const SimpleTensor<int8_t> &src2, DataType dst_data_type, ConvertPolicy convert_policy);
-template SimpleTensor<half> arithmetic_subtraction(const SimpleTensor<half> &src1, const SimpleTensor<half> &src2, DataType dst_data_type, ConvertPolicy convert_policy);
-template SimpleTensor<float> arithmetic_subtraction(const SimpleTensor<float> &src1, const SimpleTensor<float> &src2, DataType dst_data_type, ConvertPolicy convert_policy);
-} // namespace reference
-} // namespace validation
-} // namespace test
-} // namespace arm_compute
diff --git a/tests/validation/reference/ArithmeticSubtraction.h b/tests/validation/reference/ArithmeticSubtraction.h
deleted file mode 100644
index 9308314bda..0000000000
--- a/tests/validation/reference/ArithmeticSubtraction.h
+++ /dev/null
@@ -1,44 +0,0 @@
-/*
- * Copyright (c) 2017 ARM Limited.
- *
- * SPDX-License-Identifier: MIT
- *
- * Permission is hereby granted, free of charge, to any person obtaining a copy
- * of this software and associated documentation files (the "Software"), to
- * deal in the Software without restriction, including without limitation the
- * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
- * sell copies of the Software, and to permit persons to whom the Software is
- * furnished to do so, subject to the following conditions:
- *
- * The above copyright notice and this permission notice shall be included in all
- * copies or substantial portions of the Software.
- *
- * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
- * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
- * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
- * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
- * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
- * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
- * SOFTWARE.
- */
-#ifndef __ARM_COMPUTE_TEST_ARITHMETIC_SUBTRACTION_H__
-#define __ARM_COMPUTE_TEST_ARITHMETIC_SUBTRACTION_H__
-
-#include "tests/SimpleTensor.h"
-#include "tests/validation/Helpers.h"
-
-namespace arm_compute
-{
-namespace test
-{
-namespace validation
-{
-namespace reference
-{
-template <typename T1, typename T2, typename T3>
-SimpleTensor<T3> arithmetic_subtraction(const SimpleTensor<T1> &src1, const SimpleTensor<T2> &src2, DataType dst_data_type, ConvertPolicy convert_policy);
-} // namespace reference
-} // namespace validation
-} // namespace test
-} // namespace arm_compute
-#endif /* __ARM_COMPUTE_TEST_ARITHMETIC_SUBTRACTION_H__ */
diff --git a/tests/validation/reference/LaplacianPyramid.cpp b/tests/validation/reference/LaplacianPyramid.cpp
index 566847453b..21ddc1e48e 100644
--- a/tests/validation/reference/LaplacianPyramid.cpp
+++ b/tests/validation/reference/LaplacianPyramid.cpp
@@ -23,7 +23,7 @@
  */
 #include "LaplacianPyramid.h"
 
-#include "tests/validation/reference/ArithmeticSubtraction.h"
+#include "tests/validation/reference/ArithmeticOperations.h"
 #include "tests/validation/reference/DepthConvertLayer.h"
 #include "tests/validation/reference/Gaussian5x5.h"
 #include "tests/validation/reference/GaussianPyramidHalf.h"
@@ -53,7 +53,10 @@ std::vector<SimpleTensor<U>> laplacian_pyramid(const SimpleTensor<T> &src, Simpl
         const SimpleTensor<T> level_filtered = reference::gaussian5x5(gaussian_level_pyramid[i], border_mode, constant_border_value);
         pyramid_conv.push_back(level_filtered);
 
-        const SimpleTensor<U> level_sub = reference::arithmetic_subtraction<T, T, U>(gaussian_level_pyramid[i], level_filtered, dst.data_type(), ConvertPolicy::WRAP);
+        const SimpleTensor<U> level_filtered_converted = depth_convert<T, U>(level_filtered, DataType::S16, ConvertPolicy::WRAP, 0);
+        const SimpleTensor<U> gaussian_level_converted = depth_convert<T, U>(gaussian_level_pyramid[i], DataType::S16, ConvertPolicy::WRAP, 0);
+
+        const SimpleTensor<U> level_sub = reference::arithmetic_operation<U>(reference::ArithmeticOperation::SUB, gaussian_level_converted, level_filtered_converted, dst.data_type(), ConvertPolicy::WRAP);
         pyramid_dst.push_back(level_sub);
     }
 
diff --git a/tests/validation/reference/LaplacianReconstruct.cpp b/tests/validation/reference/LaplacianReconstruct.cpp
index 23468287b0..ef14355bc8 100644
--- a/tests/validation/reference/LaplacianReconstruct.cpp
+++ b/tests/validation/reference/LaplacianReconstruct.cpp
@@ -24,7 +24,7 @@
 #include "LaplacianReconstruct.h"
 
 #include "arm_compute/core/Types.h"
-#include "tests/validation/reference/ArithmeticAddition.h"
+#include "tests/validation/reference/ArithmeticOperations.h"
 #include "tests/validation/reference/DepthConvertLayer.h"
 #include "tests/validation/reference/Scale.h"
 
@@ -45,7 +45,7 @@ SimpleTensor<U> laplacian_reconstruct(const std::vector<SimpleTensor<T>> &pyrami
     const DataType data_type  = low_res.data_type();
 
     // input + L(n-1)
-    tmp_pyramid[last_level] = reference::arithmetic_addition(low_res, pyramid[last_level], data_type, ConvertPolicy::SATURATE);
+    tmp_pyramid[last_level] = reference::arithmetic_operation(reference::ArithmeticOperation::ADD, low_res, pyramid[last_level], data_type, ConvertPolicy::SATURATE);
 
     // Scale levels n-1 to 1, and add levels n-2 to 0
     for(size_t i = last_level; i-- > 0;)
@@ -56,7 +56,7 @@ SimpleTensor<U> laplacian_reconstruct(const std::vector<SimpleTensor<T>> &pyrami
         tmp_pyramid[i] = reference::scale(tmp_pyramid[i + 1], scale_x, scale_y, InterpolationPolicy::NEAREST_NEIGHBOR,
                                           border_mode, constant_border_value, SamplingPolicy::CENTER, false);
 
-        tmp_pyramid[i] = reference::arithmetic_addition(tmp_pyramid[i], pyramid[i], data_type, ConvertPolicy::SATURATE);
+        tmp_pyramid[i] = reference::arithmetic_operation(reference::ArithmeticOperation::ADD, tmp_pyramid[i], pyramid[i], data_type, ConvertPolicy::SATURATE);
     }
 
     return reference::depth_convert<T, U>(tmp_pyramid[0], DataType::U8, ConvertPolicy::SATURATE, 0);