Add new operator AddMulAdd for Neon™ backend for Float/Quantized types

This is a fused operator that merges Add + Mul + Add [+ Relu-based-Activation] layers and have an intermediate output after the first Add. It's supported for FP16/32/QASYMM8/QASYMM8_SIGNED data types.

The subsequent Add and Mul are intended for scaling and the coefficients only have one dimension (per channel).

The inputs are
     - input1    : nD tensor   [X, Y, Z, W, ..]
     - input2    : nD tensor   [X, Y, Z, W, ..]
     - add_coef  : 1D tensor   [X]
     - mul_coef  : 1D tensor   [X]

The outputs are
     - out1          : nD tensor (intermediate output)  [X, Y, Z, W, ..]
     - out2          : nD tensor (final output)  [X, Y, Z, W, ..]

The operation can be summarized as follows:
     out1 <- input1 + input2
     out2 <- Act(out1 * mul_coef + add_coef)

The activation function can be Identity, Relu, Bounded Relu or Lower/Upper Bounded Relu. The intermediate output can be skipped by providing a nullptr.

The reason of providing this operator is to be able to fuse in case of Residual network patterns and save computations by reducing memory back and forward.

Resolves: COMPMID-5463

Signed-off-by: Gunes Bayir <gunes.bayir@arm.com>
Change-Id: I8ef577aa623b036e9a9f655cc088493fd19a6109
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/9055
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Jakub Sujak <jakub.sujak@arm.com>
Reviewed-by: Viet-Hoa Do <viet-hoa.do@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>

2 files changed, 498 insertions, 0 deletions

diff --git a/tests/validation/NEON/AddMulAdd.cpp b/tests/validation/NEON/AddMulAdd.cpp
new file mode 100644
index 0000000000..f0aba7833b
--- /dev/null
+++ b/tests/validation/NEON/AddMulAdd.cpp
@@ -0,0 +1,230 @@
+/*
+ * Copyright (c) 2023 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 "arm_compute/core/Types.h"
+#include "arm_compute/runtime/NEON/functions/NEAddMulAdd.h"
+#include "arm_compute/runtime/Tensor.h"
+#include "arm_compute/runtime/TensorAllocator.h"
+
+#include "tests/NEON/Accessor.h"
+#include "tests/datasets/ShapeDatasets.h"
+#include "tests/framework/Asserts.h"
+#include "tests/framework/Macros.h"
+#include "tests/framework/datasets/Datasets.h"
+#include "tests/validation/Validation.h"
+#include "tests/validation/fixtures/AddMulAddFixture.h"
+
+namespace arm_compute
+{
+namespace test
+{
+namespace validation
+{
+namespace
+{
+constexpr AbsoluteTolerance<float> tolerance_fp32(0.001f);     /**< Tolerance for floating point tests */
+const AbsoluteTolerance<half>      tolerance_fp16(half(0.1f)); /**< Tolerance for 16-bit floating point tests */
+constexpr AbsoluteTolerance<float> tolerance_quant(1);         /**< Tolerance for quantized tests */
+
+const auto ActivationFunctionsDataset = framework::dataset::make("ActivationInfo",
+{
+    ActivationLayerInfo(),
+    ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU),
+
+    // Boundaries are aligned with Quantized Data ranges -- DOUBLE check before changing
+    ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, 6.f),
+    ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, 8.f, -2.f)
+});
+
+// QASYMM8 test quantizations
+const auto qasymm8_input1_qinfo_set = framework::dataset::make("Input1QInfo", { QuantizationInfo(0.1, 10) }); // Representable Range: [-1, 24.5]
+const auto qasymm8_input2_qinfo_set = framework::dataset::make("Input2QInfo", { QuantizationInfo(0.2, 60) }); // Representable Range: [-12, 39]
+const auto qasymm8_bn_mul_qinfo_set = framework::dataset::make("BnMulInfo", { QuantizationInfo(0.001, 55) }); // Representable Range: [-0.11, 0.2]
+const auto qasymm8_bn_add_qinfo_set = framework::dataset::make("BnAddInfo", { QuantizationInfo(0.02, 20) });  // Representable Range: [-0.4, 4.7]
+
+// Representable Range: [-9.36, 51.84], Expected F32 range: [-13, 63.5], leaving some space for saturation
+const auto qasymm8_add_output_qinfo_set = framework::dataset::make("AddOutputInfo", { QuantizationInfo(0.24, 39) });
+
+// Representable Range: [-4.8, 10.5], Expected FP32 range: [-6.985, 12.7], leaving some space for saturation
+// This range also makes sense with the activation boundaries above, i.e. [-2, 8] for LU_BOUNDED_RELU and [0, 6] for BOUNDED_RELU
+const auto qasymm8_final_output_qinfo_set = framework::dataset::make("FinalOutputInfo", { QuantizationInfo(0.06, 80) });
+
+// QASYMM8_SIGNED test quantizations
+const auto qasymm8_signed_input1_qinfo_set = framework::dataset::make("Input1QInfo", { QuantizationInfo(0.1, 10) });  // Representable Range: [-13.8, 11.7]
+const auto qasymm8_signed_input2_qinfo_set = framework::dataset::make("Input2QInfo", { QuantizationInfo(0.2, -60) }); // Representable Range: [-13.6, 39.4]
+const auto qasymm8_signed_bn_mul_qinfo_set = framework::dataset::make("BnMulInfo", { QuantizationInfo(0.001, 55) });  // Representable Range: [-0.183, 0.072]
+const auto qasymm8_signed_bn_add_qinfo_set = framework::dataset::make("BnAddInfo", { QuantizationInfo(0.4, -120) });  // Representable Range: [-0.32, 9.08]
+
+// Representable Range: [-21.36, 39.84], Expected F32 range: [-27.4, 51.1], leaving some space for saturation
+const auto qasymm8_signed_add_output_qinfo_set = framework::dataset::make("AddOutputInfo", { QuantizationInfo(0.24, -39) });
+
+// Representable Range: [-4.8, 10.5], Expected FP32 range: [-9.6713, 14.0942], leaving some space for saturation
+// This range also makes sense with the activation boundaries above, i.e. [-2, 8] for LU_BOUNDED_RELU and [0, 6] for BOUNDED_RELU
+const auto qasymm8_signed_final_output_qinfo_set = framework::dataset::make("FinalOutputInfo", { QuantizationInfo(0.06, -48) });
+
+} // namespace
+
+TEST_SUITE(NEON)
+TEST_SUITE(AddMulAdd)
+
+template <typename T>
+using NEAddMulAddFloatFixture = AddMulAddFloatValidationFixture<Tensor, Accessor, NEAddMulAdd, T, true>;
+
+template <typename T>
+using NEAddMulAddFloatFixtureWoIntermOut = AddMulAddFloatValidationFixture<Tensor, Accessor, NEAddMulAdd, T, false>;
+
+TEST_SUITE(Float)
+
+TEST_SUITE(F32)
+FIXTURE_DATA_TEST_CASE(RunSmall, NEAddMulAddFloatFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallShapes(),
+                                                                                                                    framework::dataset::make("DataType", DataType::F32)),
+                                                                                                            ActivationFunctionsDataset))
+{
+    // Validate outputs
+    validate(Accessor(_interm_target), _interm_reference); // Arithmetic Addition has more strict tolerance
+    validate(Accessor(_target), _reference, tolerance_fp32);
+}
+
+// This test is to stress the case when there is no intermediate output required (i.e. nullptr)
+FIXTURE_DATA_TEST_CASE(RunSmallWithoutIntermOutput, NEAddMulAddFloatFixtureWoIntermOut<float>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallShapes(),
+                       framework::dataset::make("DataType", DataType::F32)),
+                       framework::dataset::make("ActivationInfo", { ActivationLayerInfo() })))
+{
+    // Validate outputs
+    validate(Accessor(_target), _reference, tolerance_fp32);
+}
+
+FIXTURE_DATA_TEST_CASE(RunLarge, NEAddMulAddFloatFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(datasets::LargeShapes(),
+                                                                                                                  framework::dataset::make("DataType", DataType::F32)),
+                                                                                                          ActivationFunctionsDataset))
+{
+    // Validate outputs
+    validate(Accessor(_interm_target), _interm_reference); // Arithmetic Addition has more strict tolerance
+    validate(Accessor(_target), _reference, tolerance_fp32);
+}
+
+TEST_SUITE_END() // F32
+
+#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
+TEST_SUITE(F16)
+FIXTURE_DATA_TEST_CASE(RunSmall, NEAddMulAddFloatFixture<half>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallShapes(),
+                                                                                                                   framework::dataset::make("DataType", DataType::F16)),
+                                                                                                           ActivationFunctionsDataset))
+{
+    // Validate outputs
+    validate(Accessor(_interm_target), _interm_reference); // Arithmetic Addition has more strict tolerance
+    validate(Accessor(_target), _reference, tolerance_fp16);
+}
+
+FIXTURE_DATA_TEST_CASE(RunLarge, NEAddMulAddFloatFixture<half>, framework::DatasetMode::NIGHTLY, combine(combine(datasets::LargeShapes(),
+                                                                                                                 framework::dataset::make("DataType", DataType::F16)),
+                                                                                                         ActivationFunctionsDataset))
+{
+    // Validate outputs
+    validate(Accessor(_interm_target), _interm_reference); // Arithmetic Addition has more strict tolerance
+    validate(Accessor(_target), _reference, tolerance_fp16);
+}
+TEST_SUITE_END() // F16
+#endif           // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
+
+TEST_SUITE_END() // Float
+
+template <typename T>
+using NEAddMulQuantizedFixture = AddMulAddQuantizedValidationFixture<Tensor, Accessor, NEAddMulAdd, T, true>;
+
+template <typename T>
+using NEAddMulAddQuantizedFixtureWoIntermOut = AddMulAddQuantizedValidationFixture<Tensor, Accessor, NEAddMulAdd, T, false>;
+
+TEST_SUITE(Quantized)
+
+TEST_SUITE(QASYMM8)
+FIXTURE_DATA_TEST_CASE(RunSmall, NEAddMulQuantizedFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(combine(combine(combine(combine(datasets::SmallShapes(),
+                                                                                                                       framework::dataset::make("DataType", DataType::QASYMM8)),
+                                                                                                                       ActivationFunctionsDataset),
+                                                                                                                       qasymm8_input1_qinfo_set),
+                                                                                                                       qasymm8_input2_qinfo_set),
+                                                                                                                       qasymm8_bn_mul_qinfo_set),
+                                                                                                                       qasymm8_bn_add_qinfo_set),
+                                                                                                                       qasymm8_add_output_qinfo_set),
+                                                                                                               qasymm8_final_output_qinfo_set))
+{
+    // Validate outputs
+    validate(Accessor(_interm_target), _interm_reference, tolerance_quant);
+    validate(Accessor(_target), _reference, tolerance_quant);
+}
+
+FIXTURE_DATA_TEST_CASE(RunLarge, NEAddMulQuantizedFixture<uint8_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(combine(combine(combine(combine(combine(datasets::LargeShapes(),
+                                                                                                                     framework::dataset::make("DataType", DataType::QASYMM8)),
+                                                                                                                     ActivationFunctionsDataset),
+                                                                                                                     qasymm8_input1_qinfo_set),
+                                                                                                                     qasymm8_input2_qinfo_set),
+                                                                                                                     qasymm8_bn_mul_qinfo_set),
+                                                                                                                     qasymm8_bn_add_qinfo_set),
+                                                                                                                     qasymm8_add_output_qinfo_set),
+                                                                                                             qasymm8_final_output_qinfo_set))
+{
+    // Validate outputs
+    validate(Accessor(_interm_target), _interm_reference, tolerance_quant);
+    validate(Accessor(_target), _reference, tolerance_quant);
+}
+TEST_SUITE_END() // QASYMM8
+
+TEST_SUITE(QASYMM8_SIGNED)
+FIXTURE_DATA_TEST_CASE(RunSmall, NEAddMulQuantizedFixture<int8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(combine(combine(combine(combine(datasets::SmallShapes(),
+                                                                                                                      framework::dataset::make("DataType", DataType::QASYMM8_SIGNED)),
+                                                                                                                      ActivationFunctionsDataset),
+                                                                                                                      qasymm8_signed_input1_qinfo_set),
+                                                                                                                      qasymm8_signed_input2_qinfo_set),
+                                                                                                                      qasymm8_signed_bn_mul_qinfo_set),
+                                                                                                                      qasymm8_signed_bn_add_qinfo_set),
+                                                                                                                      qasymm8_signed_add_output_qinfo_set),
+                                                                                                              qasymm8_signed_final_output_qinfo_set))
+{
+    // Validate outputs
+    validate(Accessor(_interm_target), _interm_reference, tolerance_quant);
+    validate(Accessor(_target), _reference, tolerance_quant);
+}
+
+FIXTURE_DATA_TEST_CASE(RunLarge, NEAddMulQuantizedFixture<int8_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(combine(combine(combine(combine(combine(datasets::LargeShapes(),
+                                                                                                                    framework::dataset::make("DataType", DataType::QASYMM8_SIGNED)),
+                                                                                                                    ActivationFunctionsDataset),
+                                                                                                                    qasymm8_signed_input1_qinfo_set),
+                                                                                                                    qasymm8_signed_input2_qinfo_set),
+                                                                                                                    qasymm8_signed_bn_mul_qinfo_set),
+                                                                                                                    qasymm8_signed_bn_add_qinfo_set),
+                                                                                                                    qasymm8_signed_add_output_qinfo_set),
+                                                                                                            qasymm8_signed_final_output_qinfo_set))
+{
+    // Validate outputs
+    validate(Accessor(_interm_target), _interm_reference, tolerance_quant);
+    validate(Accessor(_target), _reference, tolerance_quant);
+}
+TEST_SUITE_END() // QASYMM8_SIGNED
+
+TEST_SUITE_END() // Quantized
+
+TEST_SUITE_END() // AddMulAdd
+TEST_SUITE_END() // NEON
+} // namespace validation
+} // namespace test
+} // namespace arm_compute
diff --git a/tests/validation/fixtures/AddMulAddFixture.h b/tests/validation/fixtures/AddMulAddFixture.h
new file mode 100644
index 0000000000..fac2bfe528
--- /dev/null
+++ b/tests/validation/fixtures/AddMulAddFixture.h
@@ -0,0 +1,268 @@
+/*
+ * Copyright (c) 2023 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 TESTS_VALIDATION_FIXTURES_ADDMULADDFIXTURE
+#define TESTS_VALIDATION_FIXTURES_ADDMULADDFIXTURE
+
+#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/ActivationLayer.h"
+#include "tests/validation/reference/ArithmeticOperations.h"
+#include "tests/validation/reference/DequantizationLayer.h"
+#include "tests/validation/reference/PixelWiseMultiplication.h"
+#include "tests/validation/reference/QuantizationLayer.h"
+
+namespace arm_compute
+{
+namespace test
+{
+namespace validation
+{
+template <typename TensorType, typename AccessorType, typename FunctionType, typename T>
+class AddMulAddGenericFixture : public framework::Fixture
+{
+public:
+    template <typename...>
+    void setup(const TensorShape &shape, DataType data_type, ActivationLayerInfo &act_info, bool interm_out)
+    {
+        compute_target(shape, data_type, act_info, interm_out);
+    }
+
+protected:
+    template <typename U>
+    void fill(U &&tensor, int i, DataType data_type)
+    {
+        switch(data_type)
+        {
+            case DataType::F32:
+                library->fill_tensor_uniform(tensor, i, -10.f, 10.f);
+                break;
+            case DataType::F16:
+                library->fill_tensor_uniform(tensor, i, -1.f, 1.f);
+                break;
+            default:
+                library->fill_tensor_uniform(tensor, i);
+                break;
+        }
+    }
+
+    void compute_target(const TensorShape &shape, DataType data_type, ActivationLayerInfo &act_info, bool interm_out)
+    {
+        TensorShape b_shape(shape.x());
+
+        // Create tensors
+        TensorType input1       = create_tensor<TensorType>(shape, data_type, 1, _input1_qinfo);
+        TensorType input2       = create_tensor<TensorType>(shape, data_type, 1, _input2_qinfo);
+        TensorType bn_mul       = create_tensor<TensorType>(b_shape, data_type, 1, _bn_mul_qinfo);
+        TensorType bn_add       = create_tensor<TensorType>(b_shape, data_type, 1, _bn_add_qinfo);
+        TensorType add_output   = create_tensor<TensorType>(shape, data_type, 1, _add_output_qinfo);
+        TensorType final_output = create_tensor<TensorType>(shape, data_type, 1, _final_output_qinfo);
+
+        // Create and configure function
+        FunctionType add_mul_add;
+        add_mul_add.configure(&input1, &input2, &bn_mul, &bn_add, interm_out ? &add_output : nullptr, &final_output, ConvertPolicy::SATURATE, act_info);
+
+        // Allocate tensors
+        input1.allocator()->allocate();
+        input2.allocator()->allocate();
+        bn_mul.allocator()->allocate();
+        bn_add.allocator()->allocate();
+
+        if(interm_out)
+        {
+            add_output.allocator()->allocate();
+        }
+
+        final_output.allocator()->allocate();
+
+        // Fill tensors
+        fill(AccessorType(input1), 0, data_type);
+        fill(AccessorType(input2), 1, data_type);
+        fill(AccessorType(bn_mul), 2, data_type);
+        fill(AccessorType(bn_add), 3, data_type);
+
+        // // Compute function
+        add_mul_add.run();
+
+        _target = std::move(final_output);
+
+        if(interm_out)
+        {
+            _interm_target = std::move(add_output);
+        }
+    }
+
+    TensorType      _target{};
+    TensorType      _interm_target{};
+    SimpleTensor<T> _reference{};
+    SimpleTensor<T> _interm_reference{};
+
+    QuantizationInfo _input1_qinfo{};
+    QuantizationInfo _input2_qinfo{};
+    QuantizationInfo _bn_mul_qinfo{};
+    QuantizationInfo _bn_add_qinfo{};
+    QuantizationInfo _add_output_qinfo{};
+    QuantizationInfo _final_output_qinfo{};
+};
+
+template <typename TensorType, typename AccessorType, typename FunctionType, typename T, bool interm_out>
+class AddMulAddFloatValidationFixture : public AddMulAddGenericFixture<TensorType, AccessorType, FunctionType, T>
+{
+public:
+    using Parent = AddMulAddGenericFixture<TensorType, AccessorType, FunctionType, T>;
+
+    template <typename...>
+    void setup(const TensorShape &shape, DataType data_type, ActivationLayerInfo act_info)
+    {
+        Parent::setup(shape, data_type, act_info, interm_out);
+        compute_reference(shape, data_type, act_info);
+    }
+
+    // Compute Reference is moved outside of the generic fixture because with the quantized data types,
+    // it becomes a very different implementation with intermediate tensors' data types being always float.
+    // This way the reference calculations are more readable and the size of the classes will be smaller
+    // due to unrepeated fill() and target() methods.
+    void compute_reference(const TensorShape &shape, DataType data_type, ActivationLayerInfo &act_info)
+    {
+        TensorShape b_shape(shape.x());
+
+        // Create reference
+        SimpleTensor<T> input1{ shape, data_type };
+        SimpleTensor<T> input2{ shape, data_type };
+        SimpleTensor<T> bn_mul{ b_shape, data_type };
+        SimpleTensor<T> bn_add{ b_shape, data_type };
+        SimpleTensor<T> add_output{ shape, data_type, 1 };
+
+        SimpleTensor<T> bn_mul_out{ shape, data_type };
+        SimpleTensor<T> bn_add_out{ shape, data_type };
+
+        // Fill reference
+        Parent::fill(input1, 0, data_type);
+        Parent::fill(input2, 1, data_type);
+        Parent::fill(bn_mul, 2, data_type);
+        Parent::fill(bn_add, 3, data_type);
+
+        reference::arithmetic_operation<T>(reference::ArithmeticOperation::ADD, input1, input2, add_output, ConvertPolicy::SATURATE);
+        bn_mul_out = reference::pixel_wise_multiplication<T, T, T>(add_output, bn_mul, 1.f, ConvertPolicy::SATURATE, RoundingPolicy::TO_NEAREST_UP, data_type);
+        reference::arithmetic_operation<T>(reference::ArithmeticOperation::ADD, bn_mul_out, bn_add, bn_add_out, ConvertPolicy::SATURATE);
+
+        if(interm_out)
+        {
+            Parent::_interm_reference = std::move(add_output);
+        }
+
+        if(act_info.enabled() && act_info.activation() != ActivationLayerInfo::ActivationFunction::IDENTITY)
+        {
+            Parent::_reference = reference::activation_layer(bn_add_out, act_info);
+        }
+        else
+        {
+            Parent::_reference = std::move(bn_add_out);
+        }
+    }
+};
+
+template <typename TensorType, typename AccessorType, typename FunctionType, typename T, bool interm_out>
+class AddMulAddQuantizedValidationFixture : public AddMulAddGenericFixture<TensorType, AccessorType, FunctionType, T>
+{
+public:
+    using Parent = AddMulAddGenericFixture<TensorType, AccessorType, FunctionType, T>;
+
+    template <typename...>
+    void setup(const TensorShape &shape, DataType data_type, ActivationLayerInfo act_info,
+               QuantizationInfo input1_qinfo, QuantizationInfo input2_qinfo, QuantizationInfo bn_mul_qinfo,
+               QuantizationInfo bn_add_qinfo, QuantizationInfo add_output_qinfo, QuantizationInfo final_output_qinfo)
+    {
+        // Quantization arguments moved to class attributes to prevent long function declerations
+        Parent::_input1_qinfo       = input1_qinfo;
+        Parent::_input2_qinfo       = input2_qinfo;
+        Parent::_bn_mul_qinfo       = bn_mul_qinfo;
+        Parent::_bn_add_qinfo       = bn_add_qinfo;
+        Parent::_add_output_qinfo   = add_output_qinfo;
+        Parent::_final_output_qinfo = final_output_qinfo;
+
+        Parent::setup(shape, data_type, act_info, interm_out);
+        compute_reference(shape, data_type, act_info);
+    }
+
+    // Compute Reference is moved outside of the generic fixture because with the quantized data types,
+    // it becomes a very different implementation with intermediate tensors' data types being always float.
+    // This way the reference calculations are more readable and the size of the classes will be smaller
+    // due to unrepeated fill() and target() methods.
+    void compute_reference(const TensorShape &shape, DataType data_type, ActivationLayerInfo &act_info)
+    {
+        TensorShape b_shape(shape.x());
+
+        // Create reference
+        SimpleTensor<T> input1{ shape, data_type, 1, Parent::_input1_qinfo };
+        SimpleTensor<T> input2{ shape, data_type, 1, Parent::_input2_qinfo };
+        SimpleTensor<T> bn_mul{ b_shape, data_type, 1, Parent::_bn_mul_qinfo };
+        SimpleTensor<T> bn_add{ b_shape, data_type, 1, Parent::_bn_add_qinfo };
+
+        // Fill input tensors
+        Parent::fill(input1, 0, data_type);
+        Parent::fill(input2, 1, data_type);
+        Parent::fill(bn_mul, 2, data_type);
+        Parent::fill(bn_add, 3, data_type);
+
+        SimpleTensor<float> input1_dequantized = reference::dequantization_layer<float>(input1);
+        SimpleTensor<float> input2_dequantized = reference::dequantization_layer<float>(input2);
+        SimpleTensor<float> bn_mul_dequantized = reference::dequantization_layer<float>(bn_mul);
+        SimpleTensor<float> bn_add_dequantized = reference::dequantization_layer<float>(bn_add);
+
+        SimpleTensor<float> add_output_dequantized{ shape, DataType::F32 };
+        SimpleTensor<float> bn_add_out_dequantized{ shape, DataType::F32 };
+
+        reference::arithmetic_operation<float>(reference::ArithmeticOperation::ADD, input1_dequantized, input2_dequantized, add_output_dequantized, ConvertPolicy::SATURATE);
+        SimpleTensor<float> bn_mul_out_dequantized = reference::pixel_wise_multiplication<float, float, float>(add_output_dequantized, bn_mul_dequantized, 1.f, ConvertPolicy::SATURATE,
+                                                                                                               RoundingPolicy::TO_NEAREST_UP, DataType::F32);
+        reference::arithmetic_operation<float>(reference::ArithmeticOperation::ADD, bn_mul_out_dequantized, bn_add_dequantized, bn_add_out_dequantized, ConvertPolicy::SATURATE);
+
+        if(interm_out)
+        {
+            Parent::_interm_reference = reference::quantization_layer<float, T>(add_output_dequantized, data_type, Parent::_add_output_qinfo);
+        }
+
+        if(act_info.enabled() && act_info.activation() != ActivationLayerInfo::ActivationFunction::IDENTITY)
+        {
+            SimpleTensor<T> ref = reference::quantization_layer<float, T>(bn_add_out_dequantized, data_type, Parent::_final_output_qinfo);
+            Parent::_reference  = reference::activation_layer(ref, act_info);
+        }
+        else
+        {
+            Parent::_reference = reference::quantization_layer<float, T>(bn_add_out_dequantized, data_type, Parent::_final_output_qinfo);
+        }
+    }
+};
+} // namespace validation
+} // namespace test
+} // namespace arm_compute
+
+#endif /* TESTS_VALIDATION_FIXTURES_ADDMULADDFIXTURE */