COMPMID-415: Move FullyConnectedLayer to new validation

Change-Id: I7f60d6fb484d3962b88874e1531cec734c11e416
Reviewed-on: http://mpd-gerrit.cambridge.arm.com/83556
Tested-by: Kaizen <jeremy.johnson+kaizengerrit@arm.com>
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>

5 files changed, 842 insertions, 0 deletions

diff --git a/tests/validation_new/CL/FullyConnectedLayer.cpp b/tests/validation_new/CL/FullyConnectedLayer.cpp
new file mode 100644
index 0000000000..9bf3a75d88
--- /dev/null
+++ b/tests/validation_new/CL/FullyConnectedLayer.cpp
@@ -0,0 +1,205 @@
+/*
+ * 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.
+ */
+#include "arm_compute/core/Types.h"
+#include "arm_compute/runtime/CL/CLTensor.h"
+#include "arm_compute/runtime/CL/CLTensorAllocator.h"
+#include "arm_compute/runtime/CL/functions/CLFullyConnectedLayer.h"
+#include "framework/Asserts.h"
+#include "framework/Macros.h"
+#include "framework/datasets/Datasets.h"
+#include "tests/CL/CLAccessor.h"
+#include "tests/PaddingCalculator.h"
+#include "tests/datasets_new/FullyConnectedLayerDataset.h"
+#include "tests/validation_new/Validation.h"
+#include "tests/validation_new/fixtures/FullyConnectedLayerFixture.h"
+#include "tests/validation_new/half.h"
+
+namespace arm_compute
+{
+namespace test
+{
+namespace validation
+{
+namespace
+{
+/** Tolerance for float operations */
+constexpr AbsoluteTolerance<float> tolerance_f32(0.001f);
+constexpr AbsoluteTolerance<float> tolerance_f16(0.4f);
+/** Tolerance for fixed point operations */
+constexpr AbsoluteTolerance<float> tolerance_fixed_point(1.f);
+
+/** CNN data types */
+const auto CNNDataTypes = framework::dataset::make("DataType",
+{
+    DataType::F16,
+    DataType::F32,
+    DataType::QS8,
+    DataType::QS16,
+});
+
+const auto FullyConnectedParameters = combine(framework::dataset::make("TransposeWeights", { false, true }), framework::dataset::make("ReshapeWeights", { false, true }));
+} // namespace
+
+TEST_SUITE(CL)
+TEST_SUITE(FullyConnectedLayer)
+
+DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(combine(framework::dataset::concat(datasets::SmallFullyConnectedLayerDataset(), datasets::LargeFullyConnectedLayerDataset()),
+                                                                           FullyConnectedParameters),
+                                                                   CNNDataTypes),
+               src_shape, weights_shape, bias_shape, dst_shape, transpose_weights, reshape_weights, data_type)
+{
+    // Set fixed point position data type allowed
+    int fixed_point_position = is_data_type_fixed_point(data_type) ? 3 : 0;
+
+    TensorShape ws(weights_shape);
+
+    // Transpose weights if not done in the function
+    if(!reshape_weights || !transpose_weights)
+    {
+        const size_t shape_x = ws.x();
+        ws.set(0, ws.y());
+        ws.set(1, shape_x);
+
+        // Weights have to be passed reshaped
+        // Transpose 1xW for batched version
+        if(!reshape_weights && dst_shape.y() > 1)
+        {
+            const float  transpose_width = 16.0f / data_size_from_type(data_type);
+            const size_t shape_x         = ws.x();
+            ws.set(0, ws.y() * static_cast<unsigned int>(transpose_width));
+            ws.set(1, static_cast<unsigned int>(std::ceil(shape_x / transpose_width)));
+        }
+    }
+
+    // Create tensors
+    CLTensor src     = create_tensor<CLTensor>(src_shape, data_type, 1, fixed_point_position);
+    CLTensor weights = create_tensor<CLTensor>(ws, data_type, 1, fixed_point_position);
+    CLTensor bias    = create_tensor<CLTensor>(bias_shape, data_type, 1, fixed_point_position);
+    CLTensor dst     = create_tensor<CLTensor>(dst_shape, data_type, 1, fixed_point_position);
+
+    ARM_COMPUTE_EXPECT(src.info()->is_resizable(), framework::LogLevel::ERRORS);
+    ARM_COMPUTE_EXPECT(weights.info()->is_resizable(), framework::LogLevel::ERRORS);
+    ARM_COMPUTE_EXPECT(bias.info()->is_resizable(), framework::LogLevel::ERRORS);
+    ARM_COMPUTE_EXPECT(dst.info()->is_resizable(), framework::LogLevel::ERRORS);
+
+    // Create and configure function.
+    CLFullyConnectedLayer fc;
+    fc.configure(&src, &weights, &bias, &dst, transpose_weights, !reshape_weights);
+
+    // Validate valid region
+    const ValidRegion dst_valid_region = shape_to_valid_region(dst_shape);
+    validate(dst.info()->valid_region(), dst_valid_region);
+}
+
+template <typename T>
+using CLFullyConnectedLayerFixture = FullyConnectedLayerValidationFixture<CLTensor, CLAccessor, CLFullyConnectedLayer, T>;
+
+TEST_SUITE(Float)
+TEST_SUITE(FP16)
+FIXTURE_DATA_TEST_CASE(RunSmall, CLFullyConnectedLayerFixture<half_float::half>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallFullyConnectedLayerDataset(),
+                       FullyConnectedParameters),
+                       framework::dataset::make("DataType", DataType::F16)))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference, tolerance_f16);
+}
+FIXTURE_DATA_TEST_CASE(RunLarge, CLFullyConnectedLayerFixture<half_float::half>, framework::DatasetMode::NIGHTLY, combine(combine(datasets::LargeFullyConnectedLayerDataset(),
+                       FullyConnectedParameters),
+                       framework::dataset::make("DataType", DataType::F16)))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference, tolerance_f16);
+}
+TEST_SUITE_END()
+
+TEST_SUITE(FP32)
+FIXTURE_DATA_TEST_CASE(RunSmall, CLFullyConnectedLayerFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallFullyConnectedLayerDataset(), FullyConnectedParameters),
+                                                                                                                 framework::dataset::make("DataType", DataType::F32)))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference, tolerance_f32);
+}
+FIXTURE_DATA_TEST_CASE(RunLarge, CLFullyConnectedLayerFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(datasets::LargeFullyConnectedLayerDataset(), FullyConnectedParameters),
+                                                                                                               framework::dataset::make("DataType", DataType::F32)))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference, tolerance_f32);
+}
+TEST_SUITE_END()
+TEST_SUITE_END()
+
+template <typename T>
+using CLFullyConnectedLayerFixedPointFixture = FullyConnectedLayerValidationFixedPointFixture<CLTensor, CLAccessor, CLFullyConnectedLayer, T>;
+
+TEST_SUITE(Quantized)
+TEST_SUITE(QS8)
+// Testing for fixed point position [1,6) as reciprocal limits the maximum fixed point position to 5
+FIXTURE_DATA_TEST_CASE(RunSmall, CLFullyConnectedLayerFixedPointFixture<int8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallFullyConnectedLayerDataset(),
+                       FullyConnectedParameters),
+                       framework::dataset::make("DataType",
+                                                DataType::QS8)),
+                       framework::dataset::make("FractionalBits", 1, 6)))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference, tolerance_fixed_point);
+}
+FIXTURE_DATA_TEST_CASE(RunLarge, CLFullyConnectedLayerFixedPointFixture<int8_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::LargeFullyConnectedLayerDataset(),
+                       FullyConnectedParameters),
+                       framework::dataset::make("DataType",
+                                                DataType::QS8)),
+                       framework::dataset::make("FractionalBits", 1, 6)))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference, tolerance_fixed_point);
+}
+TEST_SUITE_END()
+
+TEST_SUITE(QS16)
+// Testing for fixed point position [1,14) as reciprocal limits the maximum fixed point position to 14
+FIXTURE_DATA_TEST_CASE(RunSmall, CLFullyConnectedLayerFixedPointFixture<int16_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallFullyConnectedLayerDataset(),
+                       FullyConnectedParameters),
+                       framework::dataset::make("DataType",
+                                                DataType::QS16)),
+                       framework::dataset::make("FractionalBits", 1, 14)))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference, tolerance_fixed_point);
+}
+FIXTURE_DATA_TEST_CASE(RunLarge, CLFullyConnectedLayerFixedPointFixture<int16_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::LargeFullyConnectedLayerDataset(),
+                       FullyConnectedParameters),
+                       framework::dataset::make("DataType",
+                                                DataType::QS16)),
+                       framework::dataset::make("FractionalBits", 1, 14)))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference, tolerance_fixed_point);
+}
+TEST_SUITE_END()
+TEST_SUITE_END()
+
+TEST_SUITE_END()
+TEST_SUITE_END()
+} // namespace validation
+} // namespace test
+} // namespace arm_compute
diff --git a/tests/validation_new/CPP/FullyConnectedLayer.cpp b/tests/validation_new/CPP/FullyConnectedLayer.cpp
new file mode 100644
index 0000000000..7852dab27b
--- /dev/null
+++ b/tests/validation_new/CPP/FullyConnectedLayer.cpp
@@ -0,0 +1,133 @@
+/*
+ * 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.
+ */
+#include "FullyConnectedLayer.h"
+
+#include "tests/validation_new/FixedPoint.h"
+#include "tests/validation_new/half.h"
+
+#include <numeric>
+
+namespace arm_compute
+{
+namespace test
+{
+namespace validation
+{
+namespace reference
+{
+namespace
+{
+// Vector matrix multiply for floating point
+template <typename T, typename std::enable_if<is_floating_point<T>::value, int>::type = 0>
+void vector_matrix_multiply(const T *src, const T *weights, const T *bias, T *dst, int cols_weights, int rows_weights, uint8_t fixed_point_position)
+{
+    ARM_COMPUTE_UNUSED(fixed_point_position);
+
+    for(int y = 0; y < rows_weights; ++y)
+    {
+        dst[y] = std::inner_product(src, src + cols_weights, weights, static_cast<T>(0)) + bias[y];
+        weights += cols_weights;
+    }
+}
+
+// Vector matrix multiply for fixed point type
+template <typename T, typename std::enable_if<std::is_integral<T>::value, int>::type = 0>
+void vector_matrix_multiply(const T *src, const T *weights, const T *bias, T *dst, int cols_weights, int rows_weights, uint8_t fixed_point_position)
+{
+    using namespace fixed_point_arithmetic;
+    using promoted_type = fixed_point_arithmetic::traits::promote_t<T>;
+
+    for(int y = 0; y < rows_weights; ++y)
+    {
+        // Reset accumulator
+        fixed_point<promoted_type> acc(0, fixed_point_position);
+
+        for(int x = 0; x < cols_weights; ++x)
+        {
+            const fixed_point<promoted_type> i_value(src[x], fixed_point_position, true);
+            const fixed_point<promoted_type> w_value(weights[x], fixed_point_position, true);
+            acc = acc + i_value * w_value;
+        }
+
+        // Get the bias
+        const fixed_point<T> b(bias[y], fixed_point_position, true);
+
+        // Convert back and accumulate the bias
+        fixed_point<T> res(acc);
+        res = res + b;
+
+        // Store the result
+        dst[y] = res.raw();
+
+        weights += cols_weights;
+    }
+}
+} // namespace
+
+template <typename T>
+SimpleTensor<T> fully_connected_layer(const SimpleTensor<T> &src, const SimpleTensor<T> &weights, const SimpleTensor<T> &bias, const TensorShape &dst_shape)
+{
+    // Create reference
+    SimpleTensor<T> dst{ TensorShape{ dst_shape }, src.data_type(), 1, src.fixed_point_position() };
+
+    // Sanity checks
+    const int          num_batch_dimensions = std::max(0, static_cast<int>(dst_shape.num_dimensions()) - 1);
+    const int          num_input_dimensions = src.shape().num_dimensions() - num_batch_dimensions;
+    const unsigned int linear_input_size    = src.shape().total_size_lower(num_input_dimensions);
+
+    ARM_COMPUTE_UNUSED(num_batch_dimensions);
+    ARM_COMPUTE_UNUSED(num_input_dimensions);
+    ARM_COMPUTE_UNUSED(linear_input_size);
+    ARM_COMPUTE_ERROR_ON(weights.shape().x() != linear_input_size);
+    ARM_COMPUTE_ERROR_ON(weights.shape().y() != bias.shape().x());
+    ARM_COMPUTE_ERROR_ON(weights.shape().y() != dst.shape().x());
+
+    // Compute reference
+    const int cols_weights = weights.shape().x();
+    const int rows_weights = weights.shape().y();
+    const int num_batches  = dst_shape.total_size_upper(1);
+
+    for(int k = 0; k < num_batches; ++k)
+    {
+        vector_matrix_multiply<T>(src.data() + k * cols_weights,
+                                  weights.data(),
+                                  bias.data(),
+                                  dst.data() + k * rows_weights,
+                                  cols_weights,
+                                  rows_weights,
+                                  src.fixed_point_position());
+    }
+
+    return dst;
+}
+
+template SimpleTensor<float> fully_connected_layer(const SimpleTensor<float> &src, const SimpleTensor<float> &weights, const SimpleTensor<float> &bias, const TensorShape &dst_shape);
+template SimpleTensor<half_float::half> fully_connected_layer(const SimpleTensor<half_float::half> &src, const SimpleTensor<half_float::half> &weights, const SimpleTensor<half_float::half> &bias,
+                                                              const TensorShape &dst_shape);
+template SimpleTensor<qint8_t> fully_connected_layer(const SimpleTensor<qint8_t> &src, const SimpleTensor<qint8_t> &weights, const SimpleTensor<qint8_t> &bias, const TensorShape &dst_shape);
+template SimpleTensor<qint16_t> fully_connected_layer(const SimpleTensor<qint16_t> &src, const SimpleTensor<qint16_t> &weights, const SimpleTensor<qint16_t> &bias, const TensorShape &dst_shape);
+} // namespace reference
+} // namespace validation
+} // namespace test
+} // namespace arm_compute
diff --git a/tests/validation_new/CPP/FullyConnectedLayer.h b/tests/validation_new/CPP/FullyConnectedLayer.h
new file mode 100644
index 0000000000..5d62179f57
--- /dev/null
+++ b/tests/validation_new/CPP/FullyConnectedLayer.h
@@ -0,0 +1,44 @@
+/*
+ * 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_FULLY_CONNECTED_LAYER_H__
+#define __ARM_COMPUTE_TEST_FULLY_CONNECTED_LAYER_H__
+
+#include "tests/SimpleTensor.h"
+#include "tests/validation_new/Helpers.h"
+
+namespace arm_compute
+{
+namespace test
+{
+namespace validation
+{
+namespace reference
+{
+template <typename T>
+SimpleTensor<T> fully_connected_layer(const SimpleTensor<T> &src, const SimpleTensor<T> &weights, const SimpleTensor<T> &bias, const TensorShape &dst_shape);
+} // namespace reference
+} // namespace validation
+} // namespace test
+} // namespace arm_compute
+#endif /* __ARM_COMPUTE_TEST_FULLY_CONNECTED_LAYER_H__ */
diff --git a/tests/validation_new/NEON/FullyConnectedLayer.cpp b/tests/validation_new/NEON/FullyConnectedLayer.cpp
new file mode 100644
index 0000000000..6eb18ebc6a
--- /dev/null
+++ b/tests/validation_new/NEON/FullyConnectedLayer.cpp
@@ -0,0 +1,211 @@
+/*
+ * 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.
+ */
+#include "arm_compute/core/Types.h"
+#include "arm_compute/runtime/NEON/functions/NEFullyConnectedLayer.h"
+#include "arm_compute/runtime/Tensor.h"
+#include "arm_compute/runtime/TensorAllocator.h"
+#include "framework/Asserts.h"
+#include "framework/Macros.h"
+#include "framework/datasets/Datasets.h"
+#include "tests/NEON/Accessor.h"
+#include "tests/PaddingCalculator.h"
+#include "tests/datasets_new/FullyConnectedLayerDataset.h"
+#include "tests/validation_new/Validation.h"
+#include "tests/validation_new/fixtures/FullyConnectedLayerFixture.h"
+#include "tests/validation_new/half.h"
+
+namespace arm_compute
+{
+namespace test
+{
+namespace validation
+{
+namespace
+{
+/** Tolerance for float operations */
+constexpr AbsoluteTolerance<float> tolerance_f32(0.001f);
+#ifdef ARM_COMPUTE_ENABLE_FP16
+constexpr AbsoluteTolerance<float> tolerance_f16(0.01f);
+#endif /* ARM_COMPUTE_ENABLE_FP16*/
+/** Tolerance for fixed point operations */
+constexpr AbsoluteTolerance<float> tolerance_fixed_point(1.f);
+
+/** CNN data types */
+const auto CNNDataTypes = framework::dataset::make("DataType",
+{
+#ifdef ARM_COMPUTE_ENABLE_FP16
+    DataType::F16,
+#endif /* ARM_COMPUTE_ENABLE_FP16 */
+    DataType::F32,
+    DataType::QS8,
+    DataType::QS16,
+});
+
+const auto FullyConnectedParameters = combine(framework::dataset::make("TransposeWeights", { false, true }), framework::dataset::make("ReshapeWeights", { false, true }));
+} // namespace
+
+TEST_SUITE(NEON)
+TEST_SUITE(FullyConnectedLayer)
+
+DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(combine(framework::dataset::concat(datasets::SmallFullyConnectedLayerDataset(), datasets::LargeFullyConnectedLayerDataset()),
+                                                                           FullyConnectedParameters),
+                                                                   CNNDataTypes),
+               src_shape, weights_shape, bias_shape, dst_shape, transpose_weights, reshape_weights, data_type)
+{
+    // Set fixed point position data type allowed
+    int fixed_point_position = is_data_type_fixed_point(data_type) ? 3 : 0;
+
+    TensorShape ws(weights_shape);
+
+    // Transpose weights if not done in the function
+    if(!reshape_weights || !transpose_weights)
+    {
+        const size_t shape_x = ws.x();
+        ws.set(0, ws.y());
+        ws.set(1, shape_x);
+
+        // Weights have to be passed reshaped
+        // Transpose 1xW for batched version
+        if(!reshape_weights && dst_shape.y() > 1)
+        {
+            const float  transpose_width = 16.0f / data_size_from_type(data_type);
+            const size_t shape_x         = ws.x();
+            ws.set(0, ws.y() * static_cast<unsigned int>(transpose_width));
+            ws.set(1, static_cast<unsigned int>(std::ceil(shape_x / transpose_width)));
+        }
+    }
+
+    // Create tensors
+    Tensor src     = create_tensor<Tensor>(src_shape, data_type, 1, fixed_point_position);
+    Tensor weights = create_tensor<Tensor>(ws, data_type, 1, fixed_point_position);
+    Tensor bias    = create_tensor<Tensor>(bias_shape, data_type, 1, fixed_point_position);
+    Tensor dst     = create_tensor<Tensor>(dst_shape, data_type, 1, fixed_point_position);
+
+    ARM_COMPUTE_EXPECT(src.info()->is_resizable(), framework::LogLevel::ERRORS);
+    ARM_COMPUTE_EXPECT(weights.info()->is_resizable(), framework::LogLevel::ERRORS);
+    ARM_COMPUTE_EXPECT(bias.info()->is_resizable(), framework::LogLevel::ERRORS);
+    ARM_COMPUTE_EXPECT(dst.info()->is_resizable(), framework::LogLevel::ERRORS);
+
+    // Create and configure function.
+    NEFullyConnectedLayer fc;
+    fc.configure(&src, &weights, &bias, &dst, transpose_weights, !reshape_weights);
+
+    // Validate valid region
+    const ValidRegion dst_valid_region = shape_to_valid_region(dst_shape);
+    validate(dst.info()->valid_region(), dst_valid_region);
+}
+
+template <typename T>
+using NEFullyConnectedLayerFixture = FullyConnectedLayerValidationFixture<Tensor, Accessor, NEFullyConnectedLayer, T>;
+
+TEST_SUITE(Float)
+#ifdef ARM_COMPUTE_ENABLE_FP16
+TEST_SUITE(FP16)
+FIXTURE_DATA_TEST_CASE(RunSmall, NEFullyConnectedLayerFixture<half_float::half>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallFullyConnectedLayerDataset(),
+                       FullyConnectedParameters),
+                       framework::dataset::make("DataType", DataType::F16)))
+{
+    // Validate output
+    validate(Accessor(_target), _reference, tolerance_f16);
+}
+FIXTURE_DATA_TEST_CASE(RunLarge, NEFullyConnectedLayerFixture<half_float::half>, framework::DatasetMode::NIGHTLY, combine(combine(datasets::LargeFullyConnectedLayerDataset(),
+                       FullyConnectedParameters),
+                       framework::dataset::make("DataType", DataType::F16)))
+{
+    // Validate output
+    validate(Accessor(_target), _reference, tolerance_f16);
+}
+TEST_SUITE_END()
+#endif /* ARM_COMPUTE_ENABLE_FP16 */
+
+TEST_SUITE(FP32)
+FIXTURE_DATA_TEST_CASE(RunSmall, NEFullyConnectedLayerFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallFullyConnectedLayerDataset(), FullyConnectedParameters),
+                                                                                                                 framework::dataset::make("DataType", DataType::F32)))
+{
+    // Validate output
+    validate(Accessor(_target), _reference, tolerance_f32);
+}
+FIXTURE_DATA_TEST_CASE(RunLarge, NEFullyConnectedLayerFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(datasets::LargeFullyConnectedLayerDataset(), FullyConnectedParameters),
+                                                                                                               framework::dataset::make("DataType", DataType::F32)))
+{
+    // Validate output
+    validate(Accessor(_target), _reference, tolerance_f32);
+}
+TEST_SUITE_END()
+TEST_SUITE_END()
+
+template <typename T>
+using NEFullyConnectedLayerFixedPointFixture = FullyConnectedLayerValidationFixedPointFixture<Tensor, Accessor, NEFullyConnectedLayer, T>;
+
+TEST_SUITE(Quantized)
+TEST_SUITE(QS8)
+// Testing for fixed point position [1,6) as reciprocal limits the maximum fixed point position to 5
+FIXTURE_DATA_TEST_CASE(RunSmall, NEFullyConnectedLayerFixedPointFixture<int8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallFullyConnectedLayerDataset(),
+                       FullyConnectedParameters),
+                       framework::dataset::make("DataType",
+                                                DataType::QS8)),
+                       framework::dataset::make("FractionalBits", 1, 6)))
+{
+    // Validate output
+    validate(Accessor(_target), _reference, tolerance_fixed_point);
+}
+FIXTURE_DATA_TEST_CASE(RunLarge, NEFullyConnectedLayerFixedPointFixture<int8_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::LargeFullyConnectedLayerDataset(),
+                       FullyConnectedParameters),
+                       framework::dataset::make("DataType",
+                                                DataType::QS8)),
+                       framework::dataset::make("FractionalBits", 1, 6)))
+{
+    // Validate output
+    validate(Accessor(_target), _reference, tolerance_fixed_point);
+}
+TEST_SUITE_END()
+
+TEST_SUITE(QS16)
+// Testing for fixed point position [1,14) as reciprocal limits the maximum fixed point position to 14
+FIXTURE_DATA_TEST_CASE(RunSmall, NEFullyConnectedLayerFixedPointFixture<int16_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallFullyConnectedLayerDataset(),
+                       FullyConnectedParameters),
+                       framework::dataset::make("DataType",
+                                                DataType::QS16)),
+                       framework::dataset::make("FractionalBits", 1, 14)))
+{
+    // Validate output
+    validate(Accessor(_target), _reference, tolerance_fixed_point);
+}
+FIXTURE_DATA_TEST_CASE(RunLarge, NEFullyConnectedLayerFixedPointFixture<int16_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::LargeFullyConnectedLayerDataset(),
+                       FullyConnectedParameters),
+                       framework::dataset::make("DataType",
+                                                DataType::QS16)),
+                       framework::dataset::make("FractionalBits", 1, 14)))
+{
+    // Validate output
+    validate(Accessor(_target), _reference, tolerance_fixed_point);
+}
+TEST_SUITE_END()
+TEST_SUITE_END()
+
+TEST_SUITE_END()
+TEST_SUITE_END()
+} // namespace validation
+} // namespace test
+} // namespace arm_compute
diff --git a/tests/validation_new/fixtures/FullyConnectedLayerFixture.h b/tests/validation_new/fixtures/FullyConnectedLayerFixture.h
new file mode 100644
index 0000000000..eb4aad8952
--- /dev/null
+++ b/tests/validation_new/fixtures/FullyConnectedLayerFixture.h
@@ -0,0 +1,249 @@
+/*
+ * 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_FULLY_CONNECTED_LAYER_FIXTURE
+#define ARM_COMPUTE_TEST_FULLY_CONNECTED_LAYER_FIXTURE
+
+#include "arm_compute/core/TensorShape.h"
+#include "arm_compute/core/Types.h"
+#include "arm_compute/core/Utils.h"
+#include "framework/Asserts.h"
+#include "framework/Fixture.h"
+#include "tests/AssetsLibrary.h"
+#include "tests/Globals.h"
+#include "tests/IAccessor.h"
+#include "tests/RawTensor.h"
+#include "tests/validation_new/CPP/FullyConnectedLayer.h"
+#include "tests/validation_new/Helpers.h"
+
+#include <random>
+
+namespace arm_compute
+{
+namespace test
+{
+namespace validation
+{
+namespace
+{
+RawTensor transpose(const RawTensor &src, int interleave = 1)
+{
+    // Create reference
+    TensorShape dst_shape(src.shape());
+    dst_shape.set(0, src.shape().y() * interleave);
+    dst_shape.set(1, std::ceil(src.shape().x() / static_cast<float>(interleave)));
+
+    RawTensor dst{ dst_shape, src.data_type() };
+
+    // Compute reference
+    uint8_t *out_ptr = dst.data();
+
+    for(int i = 0; i < dst.num_elements(); i += interleave)
+    {
+        Coordinates coord   = index2coord(dst.shape(), i);
+        size_t      coord_x = coord.x();
+        coord.set(0, coord.y() * interleave);
+        coord.set(1, coord_x / interleave);
+
+        const int num_elements = std::min<int>(interleave, src.shape().x() - coord.x());
+
+        std::copy_n(static_cast<const uint8_t *>(src(coord)), num_elements * src.element_size(), out_ptr);
+
+        out_ptr += interleave * dst.element_size();
+    }
+
+    return dst;
+}
+} // namespace
+
+template <typename TensorType, typename AccessorType, typename FunctionType, typename T>
+class FullyConnectedLayerValidationFixedPointFixture : public framework::Fixture
+{
+public:
+    template <typename...>
+    void setup(TensorShape input_shape, TensorShape weights_shape, TensorShape bias_shape, TensorShape output_shape, bool transpose_weights, bool reshape_weights, DataType data_type, int fractional_bits)
+    {
+        ARM_COMPUTE_UNUSED(weights_shape);
+        ARM_COMPUTE_UNUSED(bias_shape);
+
+        _fractional_bits = fractional_bits;
+        _data_type       = data_type;
+
+        _target    = compute_target(input_shape, weights_shape, bias_shape, output_shape, transpose_weights, reshape_weights, data_type, fractional_bits);
+        _reference = compute_reference(input_shape, weights_shape, bias_shape, output_shape, transpose_weights, reshape_weights, data_type, fractional_bits);
+    }
+
+protected:
+    template <typename U>
+    void fill(U &&tensor, int i)
+    {
+        if(is_data_type_float(_data_type))
+        {
+            std::uniform_real_distribution<> distribution(0.5f, 1.f);
+            library->fill(tensor, distribution, i);
+        }
+        else
+        {
+            library->fill_tensor_uniform(tensor, i);
+        }
+    }
+
+    TensorType compute_target(const TensorShape &input_shape, const TensorShape &weights_shape, const TensorShape &bias_shape, const TensorShape &output_shape, bool transpose_weights,
+                              bool reshape_weights, DataType data_type, int fixed_point_position)
+    {
+        TensorShape reshaped_weights_shape(weights_shape);
+
+        // Test actions depending on the target settings
+        //
+        //            | reshape   | !reshape
+        // -----------+-----------+---------------------------
+        //  transpose |           | ***
+        // -----------+-----------+---------------------------
+        // !transpose | transpose | transpose &
+        //            |           | transpose1xW (if required)
+        //
+        // ***: That combination is invalid. But we can ignore the transpose flag and handle all !reshape the same
+        if(!reshape_weights || !transpose_weights)
+        {
+            const size_t shape_x = reshaped_weights_shape.x();
+            reshaped_weights_shape.set(0, reshaped_weights_shape.y());
+            reshaped_weights_shape.set(1, shape_x);
+
+            // Weights have to be passed reshaped
+            // Transpose 1xW for batched version
+            if(!reshape_weights && output_shape.y() > 1)
+            {
+                const int   transpose_width = 16 / data_size_from_type(data_type);
+                const float shape_x         = reshaped_weights_shape.x();
+                reshaped_weights_shape.set(0, reshaped_weights_shape.y() * transpose_width);
+                reshaped_weights_shape.set(1, static_cast<unsigned int>(std::ceil(shape_x / transpose_width)));
+            }
+        }
+
+        // Create tensors
+        TensorType src     = create_tensor<TensorType>(input_shape, data_type, 1, fixed_point_position);
+        TensorType weights = create_tensor<TensorType>(reshaped_weights_shape, data_type, 1, fixed_point_position);
+        TensorType bias    = create_tensor<TensorType>(bias_shape, data_type, 1, fixed_point_position);
+        TensorType dst     = create_tensor<TensorType>(output_shape, data_type, 1, fixed_point_position);
+
+        // Create and configure function.
+        FunctionType fc;
+        fc.configure(&src, &weights, &bias, &dst, transpose_weights, !reshape_weights);
+
+        ARM_COMPUTE_EXPECT(src.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(weights.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(bias.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(dst.info()->is_resizable(), framework::LogLevel::ERRORS);
+
+        // Allocate tensors
+        src.allocator()->allocate();
+        weights.allocator()->allocate();
+        bias.allocator()->allocate();
+        dst.allocator()->allocate();
+
+        ARM_COMPUTE_EXPECT(!src.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(!weights.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(!bias.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(!dst.info()->is_resizable(), framework::LogLevel::ERRORS);
+
+        // Fill tensors
+        fill(AccessorType(src), 0);
+        fill(AccessorType(bias), 2);
+
+        if(!reshape_weights || !transpose_weights)
+        {
+            TensorShape tmp_shape(weights_shape);
+            RawTensor   tmp(tmp_shape, data_type, 1, fixed_point_position);
+
+            // Fill with original shape
+            fill(tmp, 1);
+
+            // Transpose elementwise
+            tmp = transpose(tmp);
+
+            // Reshape weights for batched runs
+            if(!reshape_weights && output_shape.y() > 1)
+            {
+                // Transpose with interleave
+                const int interleave_size = 16 / tmp.element_size();
+                tmp                       = transpose(tmp, interleave_size);
+            }
+
+            AccessorType weights_accessor(weights);
+
+            for(int i = 0; i < tmp.num_elements(); ++i)
+            {
+                Coordinates coord = index2coord(tmp.shape(), i);
+                std::copy_n(static_cast<const RawTensor::value_type *>(tmp(coord)),
+                            tmp.element_size(),
+                            static_cast<RawTensor::value_type *>(weights_accessor(coord)));
+            }
+        }
+        else
+        {
+            fill(AccessorType(weights), 1);
+        }
+
+        // Compute NEFullyConnectedLayer function
+        fc.run();
+
+        return dst;
+    }
+
+    SimpleTensor<T> compute_reference(const TensorShape &input_shape, const TensorShape &weights_shape, const TensorShape &bias_shape, const TensorShape &output_shape, bool transpose_weights,
+                                      bool reshape_weights, DataType data_type, int fixed_point_position = 0)
+    {
+        // Create reference
+        SimpleTensor<T> src{ input_shape, data_type, 1, fixed_point_position };
+        SimpleTensor<T> weights{ weights_shape, data_type, 1, fixed_point_position };
+        SimpleTensor<T> bias{ bias_shape, data_type, 1, fixed_point_position };
+
+        // Fill reference
+        fill(src, 0);
+        fill(weights, 1);
+        fill(bias, 2);
+
+        return reference::fully_connected_layer<T>(src, weights, bias, output_shape);
+    }
+
+    TensorType      _target{};
+    SimpleTensor<T> _reference{};
+    int             _fractional_bits{};
+    DataType        _data_type{};
+};
+
+template <typename TensorType, typename AccessorType, typename FunctionType, typename T>
+class FullyConnectedLayerValidationFixture : public FullyConnectedLayerValidationFixedPointFixture<TensorType, AccessorType, FunctionType, T>
+{
+public:
+    template <typename...>
+    void setup(TensorShape input_shape, TensorShape weights_shape, TensorShape bias_shape, TensorShape output_shape, bool transpose_weights, bool reshape_weights, DataType data_type)
+    {
+        FullyConnectedLayerValidationFixedPointFixture<TensorType, AccessorType, FunctionType, T>::setup(input_shape, weights_shape, bias_shape, output_shape, transpose_weights, reshape_weights, data_type,
+                                                                                                         0);
+    }
+};
+} // namespace validation
+} // namespace test
+} // namespace arm_compute
+#endif /* ARM_COMPUTE_TEST_FULLY_CONNECTED_LAYER_FIXTURE */