COMPMID-1336: Add CLArithmeticAddition support for QASYMM8

Change-Id: Ice2bb644841fdea4e776872ff5481eb927e66bd1
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/137714
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>
Tested-by: Jenkins <bsgcomp@arm.com>

8 files changed, 235 insertions, 32 deletions

diff --git a/arm_compute/core/CL/kernels/CLArithmeticAdditionKernel.h b/arm_compute/core/CL/kernels/CLArithmeticAdditionKernel.h
index 5112476aae..a33cbf321f 100644
--- a/arm_compute/core/CL/kernels/CLArithmeticAdditionKernel.h
+++ b/arm_compute/core/CL/kernels/CLArithmeticAdditionKernel.h
@@ -53,17 +53,17 @@ public:
     ~CLArithmeticAdditionKernel() = default;
     /** Initialise the kernel's inputs, output and convertion policy.
      *
-     * @param[in]  input1 First tensor input. Data types supported: U8/QS8/QS16/S16/F16/F32.
-     * @param[in]  input2 Second tensor input. Data types supported: U8/QS8 (only if @p input1 is QS8), QS16 (only if @p input1 is QS16), S16/F16/F32.
-     * @param[out] output Output tensor. Data types supported: U8 (Only if both inputs are U8), QS8 (only if both inputs are QS8), QS16 (only if both inputs are QS16), S16/F16/F32.
+     * @param[in]  input1 First tensor input. Data types supported: U8/QS8/QASYMM8/QS16/S16/F16/F32.
+     * @param[in]  input2 Second tensor input. Data types supported: U8/QS8 (only if @p input1 is QS8), QASYMM8 (only if @p input1 is QASYMM8), QS16 (only if @p input1 is QS16), S16/F16/F32.
+     * @param[out] output Output tensor. Data types supported: U8 (Only if both inputs are U8), QS8 (only if both inputs are QS8),QASYMM8 (only if @p input1 is QASYMM8),  QS16 (only if both inputs are QS16), S16/F16/F32.
      * @param[in]  policy Policy to use to handle overflow.
      */
     void configure(const ICLTensor *input1, const ICLTensor *input2, ICLTensor *output, ConvertPolicy policy);
     /** Static function to check if given info will lead to a valid configuration of @ref CLArithmeticAdditionKernel
      *
-     * @param[in] input1 First tensor input info. Data types supported: U8/QS8/QS16/S16/F16/F32.
-     * @param[in] input2 Second tensor input info. Data types supported: U8/QS8 (only if @p input1 is QS8), QS16 (only if @p input1 is QS16), S16/F16/F32.
-     * @param[in] output Output tensor info. Data types supported: U8 (Only if both inputs are U8), QS8 (only if both inputs are QS8), QS16 (only if both inputs are QS16), S16/F16/F32.
+     * @param[in] input1 First tensor input info. Data types supported: U8/QS8/QASYMM8/QS16/S16/F16/F32.
+     * @param[in] input2 Second tensor input info. Data types supported: U8/QS8 (only if @p input1 is QS8), QASYMM8 (only if @p input1 is QASYMM8), QS16 (only if @p input1 is QS16), S16/F16/F32.
+     * @param[in] output Output tensor info. Data types supported: U8 (Only if both inputs are U8), QS8 (only if both inputs are QS8), QASYMM8 (only if both inputs are QASYMM8), QS16 (only if both inputs are QS16), S16/F16/F32.
      * @param[in] policy Policy to use to handle overflow.
      *
      * @return a status
diff --git a/arm_compute/runtime/CL/functions/CLArithmeticAddition.h b/arm_compute/runtime/CL/functions/CLArithmeticAddition.h
index 5b2fc8c145..9eea846d97 100644
--- a/arm_compute/runtime/CL/functions/CLArithmeticAddition.h
+++ b/arm_compute/runtime/CL/functions/CLArithmeticAddition.h
@@ -41,19 +41,19 @@ class CLArithmeticAddition : public ICLSimpleFunction
 public:
     /** Initialise the kernel's inputs, output and convertion policy.
      *
-     * @param[in, out] input1 First tensor input. Data types supported: U8/QS8/QS16/S16/F16/F32.
+     * @param[in, out] input1 First tensor input. Data types supported: U8/QS8/QASYMM8/QS16/S16/F16/F32.
      *                        The input tensor is [in, out] because its TensorInfo might be modified inside the kernel in case of broadcasting of dimension 0.
-     * @param[in, out] input2 Second tensor input. Data types supported: U8, QS8 (only if @p input1 is QS8), QS16 (only if @p input1 is QS16), S16/F16/F32.
+     * @param[in, out] input2 Second tensor input. Data types supported: U8, QS8 (only if @p input1 is QS8), QASYMM8 (only if @p input1 is QASYMM8), QS16 (only if @p input1 is QS16), S16/F16/F32.
      *                        The input tensor is [in, out] because its TensorInfo might be modified inside the kernel in case of broadcasting of dimension 0.
-     * @param[out]     output Output tensor. Data types supported: U8 (Only if both inputs are U8), QS8 (only if both inputs are QS8), QS16 (only if both inputs are QS16), S16/F16/F32.
+     * @param[out]     output Output tensor. Data types supported: U8 (Only if both inputs are U8), QS8 (only if both inputs are QS8), QASYMM8 (only if both inputs are QASYMM8), QS16 (only if both inputs are QS16), S16/F16/F32.
      * @param[in]      policy Policy to use to handle overflow.
      */
     void configure(ICLTensor *input1, ICLTensor *input2, ICLTensor *output, ConvertPolicy policy);
     /** Static function to check if given info will lead to a valid configuration of @ref CLArithmeticAddition
      *
-     * @param[in] input1 First tensor input info. Data types supported: U8/QS8/QS16/S16/F16/F32.
-     * @param[in] input2 Second tensor input info. Data types supported: U8/QS8 (only if @p input1 is QS8), QS16 (only if @p input1 is QS16), S16/F16/F32.
-     * @param[in] output Output tensor info. Data types supported: U8 (Only if both inputs are U8), QS8 (only if both inputs are QS8), QS16 (only if both inputs are QS16), S16/F16/F32.
+     * @param[in] input1 First tensor input info. Data types supported: U8/QS8/QASYMM8/QS16/S16/F16/F32.
+     * @param[in] input2 Second tensor input info. Data types supported: U8/QS8 (only if @p input1 is QS8), QASYMM8 (only if @p input1 is QASYMM8), QS16 (only if @p input1 is QS16), S16/F16/F32.
+     * @param[in] output Output tensor info. Data types supported: U8 (Only if both inputs are U8), QS8 (only if both inputs are QS8), QASYMM8 ( only if both inputs are QASYMM8), QS16 (only if both inputs are QS16), S16/F16/F32.
      * @param[in] policy Policy to use to handle overflow.
      *
      * @return a status
diff --git a/src/core/CL/CLKernelLibrary.cpp b/src/core/CL/CLKernelLibrary.cpp
index 712a1179a6..aa11edf9ec 100644
--- a/src/core/CL/CLKernelLibrary.cpp
+++ b/src/core/CL/CLKernelLibrary.cpp
@@ -149,6 +149,7 @@ const std::map<std::string, std::string> CLKernelLibrary::_kernel_program_map =
     { "accumulate_weighted", "accumulate.cl" },
     { "activation_layer", "activation_layer.cl" },
     { "activation_layer_qa8", "activation_layer_qa8.cl" },
+    { "arithmetic_add_quantized", "arithmetic_op_quantized.cl" },
     { "arithmetic_add", "arithmetic_op.cl" },
     { "arithmetic_sub", "arithmetic_op.cl" },
     { "arithmetic_div", "arithmetic_op.cl" },
@@ -416,6 +417,10 @@ const std::map<std::string, std::string> CLKernelLibrary::_program_source_map =
 #include "./cl_kernels/arithmetic_op.clembed"
     },
     {
+        "arithmetic_op_quantized.cl",
+#include "./cl_kernels/arithmetic_op_quantized.clembed"
+    },
+    {
         "bitwise_op.cl",
 #include "./cl_kernels/bitwise_op.clembed"
     },
diff --git a/src/core/CL/cl_kernels/arithmetic_op_quantized.cl b/src/core/CL/cl_kernels/arithmetic_op_quantized.cl
new file mode 100644
index 0000000000..82e92e32a8
--- /dev/null
+++ b/src/core/CL/cl_kernels/arithmetic_op_quantized.cl
@@ -0,0 +1,89 @@
+/*
+ * Copyright (c) 2016-2018 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 "helpers.h"
+
+#ifdef SATURATE
+#define ADD(x, y) add_sat((x), (y))
+#define SUB(x, y) sub_sat((x), (y))
+#else /* SATURATE */
+#define ADD(x, y) (x) + (y)
+#define SUB(x, y) (x) - (y)
+#endif /* SATURATE */
+
+#if defined(OFFSET)
+
+/** This function adds two tensors.
+ *
+ * @attention The quantization offset must be passed at compile time using -DOFFSET, i.e. -DOFFSET=10
+ * @attention 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_add_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);
+
+    // Load values
+    const short16 in_a   = CONVERT(vload16(0, (__global uchar *)in1.ptr), short16);
+    const short16 in_b   = CONVERT(vload16(0, (__global uchar *)in2.ptr), short16);
+    const short16 offset = OFFSET;
+
+    // Calculate result
+    short16 res = ADD(in_a, SUB(in_b, offset));
+
+    res = max((short16)0, min(res, (short16)255));
+
+    // Store result
+    vstore16(CONVERT(res, uchar16), 0, (__global uchar *)out.ptr);
+}
+#endif /* defined(OFFSET) */
diff --git a/src/core/CL/kernels/CLArithmeticAdditionKernel.cpp b/src/core/CL/kernels/CLArithmeticAdditionKernel.cpp
index b0177ab8b6..011807ad88 100644
--- a/src/core/CL/kernels/CLArithmeticAdditionKernel.cpp
+++ b/src/core/CL/kernels/CLArithmeticAdditionKernel.cpp
@@ -37,9 +37,15 @@ Status validate_arguments(const ITensorInfo &input1, const ITensorInfo &input2,
 {
     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::QS8, DataType::QS16, DataType::S16, DataType::F16, DataType::F32);
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(&input1, 1, DataType::U8, DataType::QASYMM8, DataType::QS8, DataType::QS16, 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::QS8, DataType::QS16, DataType::S16, DataType::F16, DataType::F32);
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(&input2, 1, DataType::U8, DataType::QASYMM8, DataType::QS8, DataType::QS16, 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());
 
@@ -50,12 +56,16 @@ Status validate_arguments(const ITensorInfo &input1, const ITensorInfo &input2,
     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::QS8, DataType::QS16, DataType::S16, DataType::F16, DataType::F32);
+        ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(&output, 1, DataType::U8, DataType::QS8, DataType::QASYMM8, DataType::QS16, DataType::S16, DataType::F16, DataType::F32);
         ARM_COMPUTE_RETURN_ERROR_ON_MSG((output.data_type() == DataType::U8) && ((input1.data_type() != DataType::U8) || (input2.data_type() != DataType::U8)),
                                         "Output can only be U8 if both inputs are U8");
         ARM_COMPUTE_RETURN_ERROR_ON_MSG(detail::have_different_dimensions(out_shape, output.tensor_shape(), 0),
                                         "Wrong shape for output");
         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(&input1, &output);
+        if(is_qasymm)
+        {
+            ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(&input1, &output);
+        }
     }
 
     return Status{};
@@ -124,6 +134,8 @@ void CLArithmeticAdditionKernel::configure(const ICLTensor *input1, const ICLTen
 
     const bool has_float_out = is_data_type_float(output->info()->data_type());
 
+    std::string kernel_name = "arithmetic_add";
+
     // Set kernel build options
     std::set<std::string> build_opts;
     build_opts.emplace((policy == ConvertPolicy::WRAP || has_float_out) ? "-DWRAP" : "-DSATURATE");
@@ -134,9 +146,14 @@ void CLArithmeticAdditionKernel::configure(const ICLTensor *input1, const ICLTen
     {
         build_opts.emplace("-DFIXED_POINT_POSITION=" + support::cpp11::to_string(input1->info()->fixed_point_position()));
     }
+    else if(is_data_type_quantized_asymmetric(input1->info()->data_type()))
+    {
+        build_opts.emplace("-DOFFSET=" + support::cpp11::to_string(input1->info()->quantization_info().offset));
+        kernel_name += "_quantized";
+    }
 
     // Create kernel
-    _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("arithmetic_add", build_opts));
+    _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel(kernel_name, build_opts));
 
     ICLKernel::configure(win_config.second);
 }
diff --git a/tests/validation/CL/ArithmeticAddition.cpp b/tests/validation/CL/ArithmeticAddition.cpp
index 9323ed2934..0646c05ad7 100644
--- a/tests/validation/CL/ArithmeticAddition.cpp
+++ b/tests/validation/CL/ArithmeticAddition.cpp
@@ -46,6 +46,9 @@ namespace
 /** Input data sets **/
 const auto ArithmeticAdditionU8Dataset = combine(combine(framework::dataset::make("DataType", DataType::U8), framework::dataset::make("DataType", DataType::U8)), framework::dataset::make("DataType",
                                                  DataType::U8));
+const auto ArithmeticAdditionQASYMM8Dataset = combine(combine(framework::dataset::make("DataType", DataType::QASYMM8), framework::dataset::make("DataType", DataType::QASYMM8)),
+                                                      framework::dataset::make("DataType",
+                                                                               DataType::QASYMM8));
 const auto ArithmeticAdditionS16Dataset = combine(combine(framework::dataset::make("DataType", { DataType::U8, DataType::S16 }), framework::dataset::make("DataType", DataType::S16)),
                                                   framework::dataset::make("DataType", DataType::S16));
 const auto ArithmeticAdditionFP16Dataset = combine(combine(framework::dataset::make("DataType", DataType::F16), framework::dataset::make("DataType", DataType::F16)),
@@ -121,6 +124,44 @@ FIXTURE_DATA_TEST_CASE(RunSmall, CLArithmeticAdditionFixture<uint8_t>, framework
 }
 TEST_SUITE_END()
 
+template <typename T>
+using CLArithmeticAdditionQuantizedFixture = ArithmeticAdditionValidationQuantizedFixture<CLTensor, CLAccessor, CLArithmeticAddition, T>;
+
+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, DataType::QASYMM8);
+    CLTensor ref_src2 = create_tensor<CLTensor>(shape, DataType::QASYMM8);
+    CLTensor dst      = create_tensor<CLTensor>(shape, DataType::QASYMM8);
+
+    // Create and Configure function
+    CLArithmeticAddition add;
+    add.configure(&ref_src1, &ref_src2, &dst, policy);
+
+    // 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);
+}
+
+FIXTURE_DATA_TEST_CASE(RunSmall, CLArithmeticAdditionQuantizedFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallShapes(), ArithmeticAdditionQASYMM8Dataset),
+                       framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })),
+                       framework::dataset::make("QuantizationInfo", { QuantizationInfo(2.f / 255.f, 10) })))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference);
+}
+TEST_SUITE_END()
+TEST_SUITE_END()
+
 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 })),
diff --git a/tests/validation/fixtures/ArithmeticAdditionFixture.h b/tests/validation/fixtures/ArithmeticAdditionFixture.h
index f3888ae565..99a56777a2 100644
--- a/tests/validation/fixtures/ArithmeticAdditionFixture.h
+++ b/tests/validation/fixtures/ArithmeticAdditionFixture.h
@@ -41,14 +41,15 @@ namespace test
 namespace validation
 {
 template <typename TensorType, typename AccessorType, typename FunctionType, typename T>
-class ArithmeticAdditionBroadcastValidationFixedPointFixture : public framework::Fixture
+class ArithmeticAdditionGenericFixture : 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, int fractional_bits)
+    void setup(const TensorShape &shape0, const TensorShape &shape1, DataType data_type0, DataType data_type1, DataType output_data_type, ConvertPolicy convert_policy, int fractional_bits,
+               QuantizationInfo quantization_info)
     {
-        _target    = compute_target(shape0, shape1, data_type0, data_type1, output_data_type, convert_policy, fractional_bits);
-        _reference = compute_reference(shape0, shape1, data_type0, data_type1, output_data_type, convert_policy, fractional_bits);
+        _target    = compute_target(shape0, shape1, data_type0, data_type1, output_data_type, convert_policy, fractional_bits, quantization_info);
+        _reference = compute_reference(shape0, shape1, data_type0, data_type1, output_data_type, convert_policy, fractional_bits, quantization_info);
     }
 
 protected:
@@ -59,12 +60,12 @@ protected:
     }
 
     TensorType compute_target(const TensorShape &shape0, const TensorShape &shape1, DataType data_type0, DataType data_type1, DataType output_data_type, ConvertPolicy convert_policy,
-                              int fixed_point_position)
+                              int fixed_point_position, QuantizationInfo quantization_info)
     {
         // Create tensors
-        TensorType ref_src1 = create_tensor<TensorType>(shape0, data_type0, 1, fixed_point_position);
-        TensorType ref_src2 = create_tensor<TensorType>(shape1, data_type1, 1, fixed_point_position);
-        TensorType dst      = create_tensor<TensorType>(TensorShape::broadcast_shape(shape0, shape1), output_data_type, 1, fixed_point_position);
+        TensorType ref_src1 = create_tensor<TensorType>(shape0, data_type0, 1, fixed_point_position, quantization_info);
+        TensorType ref_src2 = create_tensor<TensorType>(shape1, data_type1, 1, fixed_point_position, quantization_info);
+        TensorType dst      = create_tensor<TensorType>(TensorShape::broadcast_shape(shape0, shape1), output_data_type, 1, fixed_point_position, quantization_info);
 
         // Create and configure function
         FunctionType add;
@@ -94,11 +95,11 @@ protected:
     }
 
     SimpleTensor<T> compute_reference(const TensorShape &shape0, const TensorShape &shape1, DataType data_type0, DataType data_type1, DataType output_data_type, ConvertPolicy convert_policy,
-                                      int fixed_point_position)
+                                      int fixed_point_position, QuantizationInfo quantization_info)
     {
         // Create reference
-        SimpleTensor<T> ref_src1{ shape0, data_type0, 1, fixed_point_position };
-        SimpleTensor<T> ref_src2{ shape1, data_type1, 1, fixed_point_position };
+        SimpleTensor<T> ref_src1{ shape0, data_type0, 1, fixed_point_position, quantization_info };
+        SimpleTensor<T> ref_src2{ shape1, data_type1, 1, fixed_point_position, quantization_info };
 
         // Fill reference
         fill(ref_src1, 0);
@@ -112,24 +113,35 @@ protected:
 };
 
 template <typename TensorType, typename AccessorType, typename FunctionType, typename T>
-class ArithmeticAdditionBroadcastValidationFixture : public ArithmeticAdditionBroadcastValidationFixedPointFixture<TensorType, AccessorType, FunctionType, T>
+class ArithmeticAdditionBroadcastValidationFixedPointFixture : public ArithmeticAdditionGenericFixture<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, int fractional_bits)
+    {
+        ArithmeticAdditionGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(shape0, shape1, data_type0, data_type1, output_data_type, convert_policy, 0, QuantizationInfo());
+    }
+};
+
+template <typename TensorType, typename AccessorType, typename FunctionType, typename T>
+class ArithmeticAdditionBroadcastValidationFixture : public ArithmeticAdditionGenericFixture<TensorType, AccessorType, FunctionType, T>
 {
 public:
     template <typename...>
     void setup(const TensorShape &shape0, const TensorShape &shape1, DataType data_type0, DataType data_type1, DataType output_data_type, ConvertPolicy convert_policy)
     {
-        ArithmeticAdditionBroadcastValidationFixedPointFixture<TensorType, AccessorType, FunctionType, T>::setup(shape0, shape1, data_type0, data_type1, output_data_type, convert_policy, 0);
+        ArithmeticAdditionGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(shape0, shape1, data_type0, data_type1, output_data_type, convert_policy, 0, QuantizationInfo());
     }
 };
 
 template <typename TensorType, typename AccessorType, typename FunctionType, typename T>
-class ArithmeticAdditionValidationFixedPointFixture : public ArithmeticAdditionBroadcastValidationFixedPointFixture<TensorType, AccessorType, FunctionType, T>
+class ArithmeticAdditionValidationFixedPointFixture : public ArithmeticAdditionGenericFixture<TensorType, AccessorType, FunctionType, T>
 {
 public:
     template <typename...>
     void setup(const TensorShape &shape, DataType data_type0, DataType data_type1, DataType output_data_type, ConvertPolicy convert_policy, int fractional_bits)
     {
-        ArithmeticAdditionBroadcastValidationFixedPointFixture<TensorType, AccessorType, FunctionType, T>::setup(shape, shape, data_type0, data_type1, output_data_type, convert_policy, fractional_bits);
+        ArithmeticAdditionGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(shape, shape, data_type0, data_type1, output_data_type, convert_policy, fractional_bits, QuantizationInfo());
     }
 };
 
@@ -143,6 +155,17 @@ public:
         ArithmeticAdditionValidationFixedPointFixture<TensorType, AccessorType, FunctionType, T>::setup(shape, data_type0, data_type1, output_data_type, convert_policy, 0);
     }
 };
+
+template <typename TensorType, typename AccessorType, typename FunctionType, typename T>
+class ArithmeticAdditionValidationQuantizedFixture : public ArithmeticAdditionGenericFixture<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 quantization_info)
+    {
+        ArithmeticAdditionGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(shape, shape, data_type0, data_type1, output_data_type, convert_policy, 0, quantization_info);
+    }
+};
 } // namespace validation
 } // namespace test
 } // namespace arm_compute
diff --git a/tests/validation/reference/ArithmeticAddition.cpp b/tests/validation/reference/ArithmeticAddition.cpp
index 17020a6277..f26838dcb8 100644
--- a/tests/validation/reference/ArithmeticAddition.cpp
+++ b/tests/validation/reference/ArithmeticAddition.cpp
@@ -97,7 +97,35 @@ SimpleTensor<T> arithmetic_addition(const SimpleTensor<T> &src1, const SimpleTen
     return dst;
 }
 
-template SimpleTensor<uint8_t> arithmetic_addition(const SimpleTensor<uint8_t> &src1, const SimpleTensor<uint8_t> &src2, DataType dst_data_type, ConvertPolicy convert_policy);
+template <>
+SimpleTensor<uint8_t> arithmetic_addition(const SimpleTensor<uint8_t> &src1, const SimpleTensor<uint8_t> &src2, DataType dst_data_type, ConvertPolicy convert_policy)
+{
+    if(dst_data_type == DataType::QASYMM8)
+    {
+        SimpleTensor<float> src1_tmp = convert_from_asymmetric(src1);
+        SimpleTensor<float> src2_tmp = convert_from_asymmetric(src2);
+        SimpleTensor<float> dst_tmp(TensorShape::broadcast_shape(src1.shape(), src2.shape()), dst_data_type);
+
+        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);
+
+        SimpleTensor<uint8_t> dst = convert_to_asymmetric(dst_tmp, src1.quantization_info());
+        return dst;
+    }
+    else
+    {
+        // DataType::U8
+        SimpleTensor<uint8_t> dst(TensorShape::broadcast_shape(src1.shape(), src2.shape()), dst_data_type);
+
+        Coordinates id_src1, id_src2, id_dst;
+
+        BroadcastUnroll<Coordinates::num_max_dimensions>::unroll(src1, src2, dst, convert_policy, id_src1, id_src2, id_dst);
+
+        return dst;
+    }
+}
+
 template SimpleTensor<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);