COMPMID-1694: Fuse offset contribution with the output stage when we use NEGEMMLowpMatrixMultiplyCore

Change-Id: Ic1a681e4cc03e1eba3bf8485d9cdb17b3e926047
Signed-off-by: giuros01 <giuseppe.rossini@arm.com>
Reviewed-on: https://review.mlplatform.org/c/561
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>

4 files changed, 368 insertions, 58 deletions

diff --git a/tests/datasets/GEMMLowpFusedOffsetOutputDataset.h b/tests/datasets/GEMMLowpFusedOffsetOutputDataset.h
new file mode 100644
index 0000000000..c94019e3d5
--- /dev/null
+++ b/tests/datasets/GEMMLowpFusedOffsetOutputDataset.h
@@ -0,0 +1,201 @@
+/*
+ * Copyright (c) 2019 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_GEMMLOWPOUTPUT_DATASET
+#define ARM_COMPUTE_TEST_GEMMLOWPOUTPUT_DATASET
+
+#include "utils/TypePrinter.h"
+
+#include "arm_compute/core/TensorShape.h"
+#include "arm_compute/core/Utils.h"
+
+using namespace arm_compute;
+
+namespace arm_compute
+{
+namespace test
+{
+namespace datasets
+{
+class GEMMLowpFusedOffsetOutputDataset
+{
+public:
+    using type = std::tuple<TensorShape, TensorShape, TensorShape, int32_t, int32_t, GEMMLowpOutputStageInfo>;
+
+    struct iterator
+    {
+        iterator(std::vector<TensorShape>::const_iterator             a_it,
+                 std::vector<TensorShape>::const_iterator             b_it,
+                 std::vector<TensorShape>::const_iterator             c_it,
+                 std::vector<int32_t>::const_iterator                 a_offset_it,
+                 std::vector<int32_t>::const_iterator                 b_offset_it,
+                 std::vector<GEMMLowpOutputStageInfo>::const_iterator output_stage_it)
+            : _a_it{ std::move(a_it) },
+              _b_it{ std::move(b_it) },
+              _c_it{ std::move(c_it) },
+              _a_offset_it{ std::move(a_offset_it) },
+              _b_offset_it{ std::move(b_offset_it) },
+              _output_stage_it{ std::move(output_stage_it) }
+        {
+        }
+
+        std::string description() const
+        {
+            std::stringstream description;
+            description << "A=" << *_a_it << ":";
+            description << "B=" << *_b_it << ":";
+            description << "C=" << *_c_it << ":";
+            description << "a_offset=" << *_a_offset_it << ":";
+            description << "b_offset=" << *_b_offset_it << ":";
+            description << "output_type=" << string_from_gemmlowp_output_stage((*_output_stage_it).type) << ":";
+            description << "output_offset=" << (*_output_stage_it).gemmlowp_offset << ":";
+            description << "output_multiplier=" << (*_output_stage_it).gemmlowp_multiplier << ":";
+            description << "output_shift=" << (*_output_stage_it).gemmlowp_shift << ":";
+            description << "output_min=" << (*_output_stage_it).gemmlowp_min_bound << ":";
+            description << "output_max=" << (*_output_stage_it).gemmlowp_max_bound << ":";
+
+            return description.str();
+        }
+
+        GEMMLowpFusedOffsetOutputDataset::type operator*() const
+        {
+            return std::make_tuple(*_a_it, *_b_it, *_c_it, *_a_offset_it, *_b_offset_it, *_output_stage_it);
+        }
+
+        iterator &operator++()
+        {
+            ++_a_it;
+            ++_b_it;
+            ++_c_it;
+            ++_a_offset_it;
+            ++_b_offset_it;
+            ++_output_stage_it;
+
+            return *this;
+        }
+
+    private:
+        std::vector<TensorShape>::const_iterator             _a_it;
+        std::vector<TensorShape>::const_iterator             _b_it;
+        std::vector<TensorShape>::const_iterator             _c_it;
+        std::vector<int32_t>::const_iterator                 _a_offset_it;
+        std::vector<int32_t>::const_iterator                 _b_offset_it;
+        std::vector<GEMMLowpOutputStageInfo>::const_iterator _output_stage_it;
+    };
+
+    iterator begin() const
+    {
+        return iterator(_a_shapes.begin(), _b_shapes.begin(), _c_shapes.begin(), _a_offset.begin(), _b_offset.begin(), _output_stage.begin());
+    }
+
+    int size() const
+    {
+        return std::min(_a_shapes.size(), std::min(_b_shapes.size(), std::min(_c_shapes.size(), std::min(_a_offset.size(), std::min(_b_offset.size(), _output_stage.size())))));
+    }
+
+    void add_config(TensorShape a, TensorShape b, TensorShape c, int32_t a_offset, int32_t b_offset, GEMMLowpOutputStageInfo output_stage)
+    {
+        _a_shapes.emplace_back(std::move(a));
+        _b_shapes.emplace_back(std::move(b));
+        _c_shapes.emplace_back(std::move(c));
+        _a_offset.emplace_back(std::move(a_offset));
+        _b_offset.emplace_back(std::move(b_offset));
+        _output_stage.emplace_back(std::move(output_stage));
+    }
+
+    GEMMLowpOutputStageInfo OutputStageInfo(GEMMLowpOutputStageType type, int32_t offset, int32_t multiplier, int32_t shift, int32_t min, int32_t max)
+    {
+        GEMMLowpOutputStageInfo output_stage = GEMMLowpOutputStageInfo();
+        output_stage.type                    = type;
+        output_stage.gemmlowp_offset         = offset;
+        output_stage.gemmlowp_multiplier     = multiplier;
+        output_stage.gemmlowp_shift          = shift;
+        output_stage.gemmlowp_min_bound      = min;
+        output_stage.gemmlowp_max_bound      = max;
+        return output_stage;
+    }
+
+protected:
+    GEMMLowpFusedOffsetOutputDataset()                                    = default;
+    GEMMLowpFusedOffsetOutputDataset(GEMMLowpFusedOffsetOutputDataset &&) = default;
+
+private:
+    std::vector<TensorShape>             _a_shapes{};
+    std::vector<TensorShape>             _b_shapes{};
+    std::vector<TensorShape>             _c_shapes{};
+    std::vector<int32_t>                 _a_offset{};
+    std::vector<int32_t>                 _b_offset{};
+    std::vector<GEMMLowpOutputStageInfo> _output_stage{};
+};
+
+class SmallGEMMLowpFusedOffsetOutputDataset final : public GEMMLowpFusedOffsetOutputDataset
+{
+public:
+    SmallGEMMLowpFusedOffsetOutputDataset()
+    {
+        add_config(TensorShape(21U, 1U), TensorShape(43U, 21U), TensorShape(43U, 1U), 0, 0, OutputStageInfo(GEMMLowpOutputStageType::QUANTIZE_DOWN, -200, 2, 13, 10, 210));
+        add_config(TensorShape(21U, 13U), TensorShape(33U, 21U), TensorShape(33U, 13U), 0, 0, OutputStageInfo(GEMMLowpOutputStageType::QUANTIZE_DOWN, -100, 2, 13, 10, 210));
+        add_config(TensorShape(31U, 3U), TensorShape(72U, 31U), TensorShape(72U, 3U), -2, 13, OutputStageInfo(GEMMLowpOutputStageType::QUANTIZE_DOWN, 0, 2, 13, 10, 210));
+        add_config(TensorShape(52U, 13U), TensorShape(33U, 52U), TensorShape(33U, 13U), 0, 4, OutputStageInfo(GEMMLowpOutputStageType::QUANTIZE_DOWN, 100, 2, 13, 10, 210));
+        add_config(TensorShape(52U, 26U), TensorShape(33U, 52U), TensorShape(33U, 26U), -2, 0, OutputStageInfo(GEMMLowpOutputStageType::QUANTIZE_DOWN, 0, 2, 13, 10, 210));
+        add_config(TensorShape(31U, 27U), TensorShape(23U, 31U), TensorShape(23U, 27U), 18, 23, OutputStageInfo(GEMMLowpOutputStageType::QUANTIZE_DOWN, 200, 2, 13, 10, 210));
+        add_config(TensorShape(38U, 43U), TensorShape(21U, 38U), TensorShape(21U, 43U), -3, -2, OutputStageInfo(GEMMLowpOutputStageType::QUANTIZE_DOWN, -200, 2, 13, 10, 210));
+        add_config(TensorShape(32U, 72U), TensorShape(17U, 32U), TensorShape(17U, 72U), -9, 1, OutputStageInfo(GEMMLowpOutputStageType::QUANTIZE_DOWN, -100, 2, 13, 10, 210));
+
+        add_config(TensorShape(21U, 1U), TensorShape(43U, 21U), TensorShape(43U, 1U), 0, 0, OutputStageInfo(GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT, -2, 254601600, 10, 10, 210));
+        add_config(TensorShape(21U, 13U), TensorShape(33U, 21U), TensorShape(33U, 13U), 0, 0, OutputStageInfo(GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT, -1, 254601600, 10, 10, 210));
+        add_config(TensorShape(31U, 3U), TensorShape(72U, 31U), TensorShape(72U, 3U), -2, 13, OutputStageInfo(GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT, 0, 254601600, 10, 10, 210));
+        add_config(TensorShape(52U, 26U), TensorShape(33U, 52U), TensorShape(33U, 26U), -2, 0, OutputStageInfo(GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT, 1, 254601600, 10, 10, 210));
+        add_config(TensorShape(31U, 27U), TensorShape(23U, 31U), TensorShape(23U, 27U), 5, 13, OutputStageInfo(GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT, 2, 254601602, 10, 10, 210));
+        add_config(TensorShape(38U, 43U), TensorShape(21U, 38U), TensorShape(21U, 43U), -3, -2, OutputStageInfo(GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT, -2, 254601602, 10, 10, 210));
+        add_config(TensorShape(32U, 72U), TensorShape(17U, 32U), TensorShape(17U, 72U), -9, 1, OutputStageInfo(GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT, -1, 254601602, 10, 10, 210));
+    }
+};
+
+class LargeGEMMLowpFusedOffsetOutputDataset final : public GEMMLowpFusedOffsetOutputDataset
+{
+public:
+    LargeGEMMLowpFusedOffsetOutputDataset()
+    {
+        add_config(TensorShape(923U, 1U), TensorShape(871U, 923U), TensorShape(871U, 1U), 0, 0, OutputStageInfo(GEMMLowpOutputStageType::QUANTIZE_DOWN, -200, 2, 18, 10, 210));
+        add_config(TensorShape(923U, 429U), TensorShape(871U, 923U), TensorShape(871U, 429U), 0, 0, OutputStageInfo(GEMMLowpOutputStageType::QUANTIZE_DOWN, -100, 2, 18, 10, 210));
+        add_config(TensorShape(873U, 7U), TensorShape(784U, 873U), TensorShape(784U, 7U), -1, 3, OutputStageInfo(GEMMLowpOutputStageType::QUANTIZE_DOWN, 0, 2, 18, 10, 210));
+        add_config(TensorShape(873U, 513U), TensorShape(784U, 873U), TensorShape(784U, 513U), 0, 4, OutputStageInfo(GEMMLowpOutputStageType::QUANTIZE_DOWN, 100, 2, 18, 10, 210));
+        add_config(TensorShape(697U, 872U), TensorShape(563U, 697U), TensorShape(563U, 872U), -2, 0, OutputStageInfo(GEMMLowpOutputStageType::QUANTIZE_DOWN, 0, 2, 18, 10, 210));
+        add_config(TensorShape(1021U, 973U), TensorShape(783U, 1021U), TensorShape(783U, 973U), 5, 13, OutputStageInfo(GEMMLowpOutputStageType::QUANTIZE_DOWN, 200, 2, 18, 10, 210));
+        add_config(TensorShape(681U, 1023U), TensorShape(213U, 681U), TensorShape(213U, 1023U), -3, -2, OutputStageInfo(GEMMLowpOutputStageType::QUANTIZE_DOWN, -200, 2, 18, 10, 210));
+        add_config(TensorShape(941U, 1011U), TensorShape(623U, 941U), TensorShape(623U, 1011U), -9, 1, OutputStageInfo(GEMMLowpOutputStageType::QUANTIZE_DOWN, -100, 2, 18, 10, 210));
+
+        add_config(TensorShape(923U, 1U), TensorShape(871U, 923U), TensorShape(871U, 1U), 0, 0, OutputStageInfo(GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT, -2, 254601600, 15, 10, 210));
+        add_config(TensorShape(923U, 429U), TensorShape(871U, 923U), TensorShape(871U, 429U), 0, 0, OutputStageInfo(GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT, -1, 254601600, 15, 10, 210));
+        add_config(TensorShape(873U, 7U), TensorShape(784U, 873U), TensorShape(784U, 7U), -1, 3, OutputStageInfo(GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT, 0, 254601600, 15, 10, 210));
+        add_config(TensorShape(873U, 513U), TensorShape(784U, 873U), TensorShape(784U, 513U), 0, 4, OutputStageInfo(GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT, 1, 254601600, 15, 10, 210));
+        add_config(TensorShape(697U, 872U), TensorShape(563U, 697U), TensorShape(563U, 872U), -2, 0, OutputStageInfo(GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT, 2, 254601602, 15, 10, 210));
+        add_config(TensorShape(1021U, 973U), TensorShape(783U, 1021U), TensorShape(783U, 973U), 5, 13, OutputStageInfo(GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT, -2, 254601602, 15, 10, 210));
+        add_config(TensorShape(681U, 1023U), TensorShape(213U, 681U), TensorShape(213U, 1023U), -3, -2, OutputStageInfo(GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT, -1, 254601602, 15, 10, 210));
+    }
+};
+} // namespace datasets
+} // namespace test
+} // namespace arm_compute
+#endif /* ARM_COMPUTE_TEST_GEMMLOWPOUTPUT_DATASET */
diff --git a/tests/validation/CL/GEMMLowp.cpp b/tests/validation/CL/GEMMLowp.cpp
index 08641dbaa3..efefbd645b 100644
--- a/tests/validation/CL/GEMMLowp.cpp
+++ b/tests/validation/CL/GEMMLowp.cpp
@@ -28,6 +28,7 @@
 #include "arm_compute/runtime/CL/functions/CLGEMMLowpOutputStage.h"
 #include "tests/CL/CLAccessor.h"
 #include "tests/PaddingCalculator.h"
+#include "tests/datasets/GEMMLowpFusedOffsetOutputDataset.h"
 #include "tests/datasets/LargeGEMMLowpDataset.h"
 #include "tests/datasets/ShapeDatasets.h"
 #include "tests/datasets/SmallGEMMLowpDataset.h"
@@ -83,6 +84,21 @@ FIXTURE_DATA_TEST_CASE(RunLarge, CLGEMMLowpMatrixMultiplyCoreFixture, framework:
     validate(CLAccessor(_target), _reference);
 }
 
+using CLGEMMLowpMatrixMultiplyCoreFusedOffsetOutputFixture = GEMMLowpMatrixMultiplyCoreFusedOffsetOutputValidationFixture<CLTensor, CLAccessor, CLGEMMLowpMatrixMultiplyCore>;
+TEST_SUITE(FusedOffsetOutput)
+FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMLowpMatrixMultiplyCoreFusedOffsetOutputFixture, framework::DatasetMode::ALL, datasets::SmallGEMMLowpFusedOffsetOutputDataset())
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference);
+}
+
+FIXTURE_DATA_TEST_CASE(RunLarge, CLGEMMLowpMatrixMultiplyCoreFusedOffsetOutputFixture, framework::DatasetMode::NIGHTLY, datasets::LargeGEMMLowpFusedOffsetOutputDataset())
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference);
+}
+TEST_SUITE_END() // FusedOffsetOutput
+
 TEST_SUITE(Output3D)
 using CLGEMMLowpMatrixMultiplyCoreOutput3DFixture = GEMMLowpMatrixMultiplyCoreValidationFixture<CLTensor, CLAccessor, CLGEMMLowpMatrixMultiplyCore, false, true>;
 FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMLowpMatrixMultiplyCoreOutput3DFixture, framework::DatasetMode::PRECOMMIT, datasets::SmallGEMMLowpOutput3DDataset())
diff --git a/tests/validation/NEON/GEMMLowp.cpp b/tests/validation/NEON/GEMMLowp.cpp
index 57067f140f..f0460b4a23 100644
--- a/tests/validation/NEON/GEMMLowp.cpp
+++ b/tests/validation/NEON/GEMMLowp.cpp
@@ -30,6 +30,7 @@
 #include "tests/NEON/Accessor.h"
 #include "tests/NEON/Helper.h"
 #include "tests/PaddingCalculator.h"
+#include "tests/datasets/GEMMLowpFusedOffsetOutputDataset.h"
 #include "tests/datasets/LargeGEMMLowpDataset.h"
 #include "tests/datasets/ShapeDatasets.h"
 #include "tests/datasets/SmallGEMMLowpDataset.h"
@@ -144,6 +145,20 @@ FIXTURE_DATA_TEST_CASE(RunLarge, NEGEMMLowpMatrixMultiplyCoreFixture, framework:
     validate(Accessor(_target), _reference);
 }
 
+using NEGEMMLowpMatrixMultiplyCoreFusedOffsetOutputFixture = GEMMLowpMatrixMultiplyCoreFusedOffsetOutputValidationFixture<Tensor, Accessor, NEGEMMLowpMatrixMultiplyCore>;
+TEST_SUITE(FusedOffsetOutput)
+FIXTURE_DATA_TEST_CASE(RunSmall, NEGEMMLowpMatrixMultiplyCoreFusedOffsetOutputFixture, framework::DatasetMode::ALL, datasets::SmallGEMMLowpFusedOffsetOutputDataset())
+{
+    // Validate output
+    validate(Accessor(_target), _reference);
+}
+
+FIXTURE_DATA_TEST_CASE(RunLarge, NEGEMMLowpMatrixMultiplyCoreFusedOffsetOutputFixture, framework::DatasetMode::NIGHTLY, datasets::LargeGEMMLowpFusedOffsetOutputDataset())
+{
+    // Validate output
+    validate(Accessor(_target), _reference);
+}
+TEST_SUITE_END() // FusedOffsetOutput
 TEST_SUITE_END() // MatrixMultiplyCore
 
 TEST_SUITE(OutputStage)
diff --git a/tests/validation/fixtures/GEMMLowpFixture.h b/tests/validation/fixtures/GEMMLowpFixture.h
index 836f8eddfe..90a4b5cf40 100644
--- a/tests/validation/fixtures/GEMMLowpFixture.h
+++ b/tests/validation/fixtures/GEMMLowpFixture.h
@@ -42,86 +42,164 @@ namespace test
 {
 namespace validation
 {
-template <typename TensorType, typename AccessorType, typename FunctionType, bool reinterpret_input_as_3d = false, bool reinterpret_output_as_3d = false>
-class GEMMLowpMatrixMultiplyCoreValidationFixture : public framework::Fixture
+namespace
 {
-public:
-    template <typename...>
-    void setup(TensorShape shape_a, TensorShape shape_b, TensorShape shape_c, int32_t a_offset, int32_t b_offset)
+template <typename U>
+void fill(U &&tensor, int i)
+{
+    // Between 1 and 254 in order to avoid having -128 and 128 for the DOT product path
+    std::uniform_int_distribution<> distribution(1, 254);
+    library->fill(tensor, distribution, i);
+}
+
+template <typename TensorType, typename AccessorType, typename FunctionType, bool reinterpret_input_as_3d, bool reinterpret_output_as_3d, typename OutputType, bool is_fused = false>
+TensorType compute_gemmlowp_target(const TensorShape &shape_a, const TensorShape &shape_b, const TensorShape &shape_output, int32_t a_offset, int32_t b_offset,
+                                   GEMMLowpOutputStageInfo output_stage = GEMMLowpOutputStageInfo())
+{
+    // Create tensors
+    TensorType a      = create_tensor<TensorType>(shape_a, DataType::QASYMM8, 1);
+    TensorType b      = create_tensor<TensorType>(shape_b, DataType::QASYMM8, 1);
+    TensorType output = create_tensor<TensorType>(shape_output, output_stage.type == GEMMLowpOutputStageType::NONE ? DataType::S32 : DataType::QASYMM8, 1);
+
+    a.info()->set_quantization_info(QuantizationInfo(1.0f / 255, a_offset));
+    b.info()->set_quantization_info(QuantizationInfo(1.0f / 255, b_offset));
+
+    TensorType bias;
+    if(is_fused)
     {
-        _target    = compute_target(shape_a, shape_b, shape_c, a_offset, b_offset);
-        _reference = compute_reference(shape_a, shape_b, shape_c, a_offset, b_offset);
+        TensorShape bias_shape(shape_b[0]);
+        bias = create_tensor<TensorType>(bias_shape, DataType::S32, 1);
     }
 
-protected:
-    template <typename U>
-    void fill(U &&tensor, int i)
+    // Create and configure function
+    // The GEMMinfo includes the values of the depth in case of reinterpreted 3d input/output
+    FunctionType gemmlowp;
+    // TODO (COMPMID-1672) - Extending the test to validate add bias in offset contribution
+    gemmlowp.configure(&a, &b, is_fused ? &bias : nullptr, &output, GEMMInfo(false, false, false, (reinterpret_output_as_3d ? shape_output[2] : 0), reinterpret_input_as_3d, false, output_stage));
+
+    ARM_COMPUTE_EXPECT(a.info()->is_resizable(), framework::LogLevel::ERRORS);
+    ARM_COMPUTE_EXPECT(b.info()->is_resizable(), framework::LogLevel::ERRORS);
+    ARM_COMPUTE_EXPECT(output.info()->is_resizable(), framework::LogLevel::ERRORS);
+
+    // Allocate tensors
+    a.allocator()->allocate();
+    b.allocator()->allocate();
+    output.allocator()->allocate();
+
+    ARM_COMPUTE_EXPECT(!a.info()->is_resizable(), framework::LogLevel::ERRORS);
+    ARM_COMPUTE_EXPECT(!b.info()->is_resizable(), framework::LogLevel::ERRORS);
+    ARM_COMPUTE_EXPECT(!output.info()->is_resizable(), framework::LogLevel::ERRORS);
+
+    // Fill tensors
+    fill(AccessorType(a), 0);
+    fill(AccessorType(b), 1);
+
+    if(is_fused)
     {
-        // Between 1 and 254 in order to avoid having -128 and 128 for the DOT product path
-        std::uniform_int_distribution<> distribution(1, 254);
-        library->fill(tensor, distribution, i);
+        ARM_COMPUTE_EXPECT(bias.info()->is_resizable(), framework::LogLevel::ERRORS);
+        bias.allocator()->allocate();
+        ARM_COMPUTE_EXPECT(!bias.info()->is_resizable(), framework::LogLevel::ERRORS);
+        fill(AccessorType(bias), 2);
     }
 
-    TensorType compute_target(const TensorShape &shape_a, const TensorShape &shape_b, const TensorShape &shape_c, int32_t a_offset, int32_t b_offset)
+    // Compute GEMM function
+    gemmlowp.run();
+    return output;
+}
+
+template <bool        reinterpret_input_as_3d>
+SimpleTensor<int32_t> compute_gemmlowp_reference(const TensorShape &shape_a, const TensorShape &shape_b, const TensorShape &shape_output, int32_t a_offset, int32_t b_offset)
+{
+    TensorShape shape_a_to_use = shape_a;
+    if(reinterpret_input_as_3d)
     {
-        // Create tensors
-        TensorType a = create_tensor<TensorType>(shape_a, DataType::QASYMM8, 1);
-        TensorType b = create_tensor<TensorType>(shape_b, DataType::QASYMM8, 1);
-        TensorType c = create_tensor<TensorType>(shape_c, DataType::S32, 1);
+        // Collapse the second and third dimension if the input is 3D
+        shape_a_to_use.collapse(2U, 1U);
+    }
 
-        a.info()->set_quantization_info(QuantizationInfo(1.0f / 255, a_offset));
-        b.info()->set_quantization_info(QuantizationInfo(1.0f / 255, b_offset));
+    // Create reference
+    SimpleTensor<uint8_t> a{ shape_a_to_use, DataType::QASYMM8, 1 };
+    SimpleTensor<uint8_t> b{ shape_b, DataType::QASYMM8, 1 };
 
-        // Create and configure function
-        // The GEMMinfo includes the values of the depth in case of reinterpreted 3d input/output
-        FunctionType gemmlowp;
-        // TODO (COMPMID-1672) - Extending the test to validate add bias in offset contribution
-        gemmlowp.configure(&a, &b, nullptr, &c, GEMMInfo(false, false, false, (reinterpret_output_as_3d ? shape_c[2] : 0), reinterpret_input_as_3d));
+    // Fill reference
+    fill(a, 0);
+    fill(b, 1);
 
-        ARM_COMPUTE_EXPECT(a.info()->is_resizable(), framework::LogLevel::ERRORS);
-        ARM_COMPUTE_EXPECT(b.info()->is_resizable(), framework::LogLevel::ERRORS);
-        ARM_COMPUTE_EXPECT(c.info()->is_resizable(), framework::LogLevel::ERRORS);
+    return reference::gemmlowp_matrix_multiply_core<int32_t, uint8_t>(a, b, shape_output, a_offset, b_offset);
+}
+}
 
-        // Allocate tensors
-        a.allocator()->allocate();
-        b.allocator()->allocate();
-        c.allocator()->allocate();
+template <typename TensorType, typename AccessorType, typename FunctionType, bool reinterpret_input_as_3d = false, bool reinterpret_output_as_3d = false>
+class GEMMLowpMatrixMultiplyCoreValidationFixture : public framework::Fixture
+{
+public:
+    template <typename...>
+    void setup(TensorShape shape_a, TensorShape shape_b, TensorShape shape_output, int32_t a_offset, int32_t b_offset)
+    {
+        _target    = compute_target(shape_a, shape_b, shape_output, a_offset, b_offset);
+        _reference = compute_reference(shape_a, shape_b, shape_output, a_offset, b_offset);
+    }
 
-        ARM_COMPUTE_EXPECT(!a.info()->is_resizable(), framework::LogLevel::ERRORS);
-        ARM_COMPUTE_EXPECT(!b.info()->is_resizable(), framework::LogLevel::ERRORS);
-        ARM_COMPUTE_EXPECT(!c.info()->is_resizable(), framework::LogLevel::ERRORS);
+protected:
+    TensorType compute_target(const TensorShape &shape_a, const TensorShape &shape_b, const TensorShape &shape_output, int32_t a_offset, int32_t b_offset)
+    {
+        return compute_gemmlowp_target<TensorType, AccessorType, FunctionType, reinterpret_input_as_3d, reinterpret_output_as_3d, int32_t>(shape_a, shape_b, shape_output, a_offset, b_offset);
+    }
 
-        // Fill tensors
-        fill(AccessorType(a), 0);
-        fill(AccessorType(b), 1);
+    SimpleTensor<int32_t> compute_reference(const TensorShape &shape_a, const TensorShape &shape_b, const TensorShape &shape_output, int32_t a_offset, int32_t b_offset)
+    {
+        return compute_gemmlowp_reference<reinterpret_input_as_3d>(shape_a, shape_b, shape_output, a_offset, b_offset);
+    }
 
-        // Compute GEMM function
-        gemmlowp.run();
-        return c;
+    TensorType            _target{};
+    SimpleTensor<int32_t> _reference{};
+};
+
+template <typename TensorType, typename AccessorType, typename FunctionType, bool reinterpret_input_as_3d = false, bool reinterpret_output_as_3d = false>
+class GEMMLowpMatrixMultiplyCoreFusedOffsetOutputValidationFixture : public framework::Fixture
+{
+public:
+    template <typename...>
+    void setup(TensorShape shape_a, TensorShape shape_b, TensorShape shape_output, int32_t a_offset, int32_t b_offset, GEMMLowpOutputStageInfo output_stage)
+    {
+        ARM_COMPUTE_EXPECT(output_stage.type != GEMMLowpOutputStageType::NONE, framework::LogLevel::ERRORS);
+        _target    = compute_target(shape_a, shape_b, shape_output, a_offset, b_offset, output_stage);
+        _reference = compute_reference(shape_a, shape_b, shape_output, a_offset, b_offset, output_stage);
     }
 
-    SimpleTensor<int32_t> compute_reference(const TensorShape &shape_a, const TensorShape &shape_b, const TensorShape &shape_c, int32_t a_offset, int32_t b_offset)
+protected:
+    TensorType compute_target(const TensorShape &shape_a, const TensorShape &shape_b, const TensorShape &shape_output, int32_t a_offset, int32_t b_offset, GEMMLowpOutputStageInfo output_stage)
     {
-        TensorShape shape_a_to_use = shape_a;
-        if(reinterpret_input_as_3d)
-        {
-            // Collapse the second and third dimension if the input is 3D
-            shape_a_to_use.collapse(2U, 1U);
-        }
+        return compute_gemmlowp_target<TensorType, AccessorType, FunctionType, reinterpret_input_as_3d, reinterpret_output_as_3d, qasymm8_t, true>(shape_a, shape_b, shape_output, a_offset, b_offset,
+                output_stage);
+    }
 
-        // Create reference
-        SimpleTensor<uint8_t> a{ shape_a_to_use, DataType::QASYMM8, 1 };
-        SimpleTensor<uint8_t> b{ shape_b, DataType::QASYMM8, 1 };
+    SimpleTensor<qasymm8_t> compute_reference(const TensorShape &shape_a, const TensorShape &shape_b, const TensorShape &shape_output, int32_t a_offset, int32_t b_offset,
+                                              GEMMLowpOutputStageInfo output_stage)
+    {
+        SimpleTensor<int32_t> output = compute_gemmlowp_reference<reinterpret_input_as_3d>(shape_a, shape_b, shape_output, a_offset, b_offset);
 
-        // Fill reference
-        fill(a, 0);
-        fill(b, 1);
+        TensorShape           bias_shape(shape_b[0]);
+        SimpleTensor<int32_t> bias{ bias_shape, DataType::S32, 1 };
+        fill(bias, 2);
 
-        return reference::gemmlowp_matrix_multiply_core<int32_t, uint8_t>(a, b, shape_c, a_offset, b_offset);
+        switch(output_stage.type)
+        {
+            case GEMMLowpOutputStageType::QUANTIZE_DOWN:
+                return reference::gemmlowp_quantize_down_int32_to_uint8_scale<int32_t>(output, bias,
+                                                                                       output_stage.gemmlowp_offset, output_stage.gemmlowp_multiplier, output_stage.gemmlowp_shift, output_stage.gemmlowp_min_bound, output_stage.gemmlowp_max_bound);
+                break;
+            case GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT:
+                return reference::gemmlowp_quantize_down_int32_to_uint8_scale_by_fixedpoint<int32_t>(output, bias,
+                                                                                                     output_stage.gemmlowp_multiplier, output_stage.gemmlowp_shift, output_stage.gemmlowp_offset, output_stage.gemmlowp_min_bound, output_stage.gemmlowp_max_bound);
+                break;
+            default:
+                ARM_COMPUTE_ERROR("Not Supported!");
+        }
     }
 
-    TensorType            _target{};
-    SimpleTensor<int32_t> _reference{};
+    TensorType              _target{};
+    SimpleTensor<qasymm8_t> _reference{};
 };
 
 template <typename TensorType, typename AccessorType, typename FunctionType>
@@ -536,4 +614,4 @@ protected:
 } // namespace validation
 } // namespace test
 } // namespace arm_compute
-#endif /* ARM_COMPUTE_TEST_GEMMLOWP_FIXTURE */
\ No newline at end of file
+#endif /* ARM_COMPUTE_TEST_GEMMLOWP_FIXTURE */