COMPMID-594: Implement reference and CL/NEON validation for LocallyConnected

Change-Id: I01e7abcf3f1b19458128e277044af850ad9fa224
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/118610
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>
Tested-by: Jenkins <bsgcomp@arm.com>

10 files changed, 795 insertions, 189 deletions

diff --git a/src/core/NEON/kernels/NELocallyConnectedMatrixMultiplyKernel.cpp b/src/core/NEON/kernels/NELocallyConnectedMatrixMultiplyKernel.cpp
index 58da0402bc..35beb82689 100644
--- a/src/core/NEON/kernels/NELocallyConnectedMatrixMultiplyKernel.cpp
+++ b/src/core/NEON/kernels/NELocallyConnectedMatrixMultiplyKernel.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017, 2018 ARM Limited.
+ * Copyright (c) 2017-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -186,7 +186,7 @@ void vector_matrix_multiply_f32(const ITensor *input0, const ITensor *input1, IT
     win_out.set(Window::DimX, Window::Dimension(window_start_x, window_end_x, window_step_x));
 
     Window win_a(window);
-    win_a.set(Window::DimX, Window::Dimension(0, 1, 1));
+    win_a.set(Window::DimX, Window::Dimension(0, 0, 0));
 
     Iterator ina(input0, win_a);
     Iterator out(output, win_out);
@@ -234,7 +234,7 @@ void vector_matrix_multiply_f32(const ITensor *input0, const ITensor *input1, IT
             asm volatile("PLD [%0, #128*1]" ::"r"(reinterpret_cast<const uint8_t *>(matrix_b + 2 * in_b_stride)));
             asm volatile("PLD [%0, #128*1]" ::"r"(reinterpret_cast<const uint8_t *>(matrix_b + 3 * in_b_stride)));
             asm volatile("PLD [%0, #128*1]" ::"r"(reinterpret_cast<const uint8_t *>(matrix_b + 4 * in_b_stride)));
-#endif /* __arm __ */
+#endif /* __arm__ */
 
             acc0 = vmlaq_lane_f32(acc0, b00, a0l, 0);
             acc1 = vmlaq_lane_f32(acc1, b01, a0l, 0);
diff --git a/src/runtime/CL/functions/CLLocallyConnectedLayer.cpp b/src/runtime/CL/functions/CLLocallyConnectedLayer.cpp
index 9120aadf17..d284949323 100644
--- a/src/runtime/CL/functions/CLLocallyConnectedLayer.cpp
+++ b/src/runtime/CL/functions/CLLocallyConnectedLayer.cpp
@@ -67,10 +67,13 @@ void CLLocallyConnectedLayer::configure(const ICLTensor *input, const ICLTensor
     std::tie(stride_x, stride_y) = conv_info.stride();
     std::tie(pad_x, pad_y)       = conv_info.pad();
 
+    const unsigned int kernel_width  = weights->info()->dimension(0);
+    const unsigned int kernel_height = weights->info()->dimension(1);
+
     // Get convolved dimensions
     unsigned int conv_w = 0;
     unsigned int conv_h = 0;
-    std::tie(conv_w, conv_h) = scaled_dimensions(input->info()->dimension(0), input->info()->dimension(1), weights->info()->dimension(0), weights->info()->dimension(1),
+    std::tie(conv_w, conv_h) = scaled_dimensions(input->info()->dimension(0), input->info()->dimension(1), kernel_width, kernel_height,
                                                  conv_info);
 
     ARM_COMPUTE_ERROR_ON_MSG((output->info()->dimension(0) != conv_w) || (output->info()->dimension(1) != conv_h), "Output shape does not match the expected one");
@@ -106,7 +109,7 @@ void CLLocallyConnectedLayer::configure(const ICLTensor *input, const ICLTensor
     _memory_group.manage(&_gemm_output);
 
     // Configure kernels
-    _input_im2col_kernel.configure(input, &_input_im2col_reshaped, Size2D(conv_w, conv_h), conv_info, _has_bias);
+    _input_im2col_kernel.configure(input, &_input_im2col_reshaped, Size2D(kernel_width, kernel_height), conv_info, _has_bias);
     _weights_reshape_kernel.configure(weights, biases, &_weights_reshaped);
     _mm_kernel.configure(&_input_im2col_reshaped, &_weights_reshaped, &_gemm_output);
     _output_col2im_kernel.configure(&_gemm_output, output, std::make_pair(conv_w, conv_h));
diff --git a/tests/datasets/LocallyConnectedDataset.h b/tests/datasets/LocallyConnectedDataset.h
new file mode 100644
index 0000000000..cc2fa88f02
--- /dev/null
+++ b/tests/datasets/LocallyConnectedDataset.h
@@ -0,0 +1,79 @@
+/*
+ * Copyright (c) 2017-2018 ARM Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#ifndef ARM_COMPUTE_TEST_LOCALLYCONNECTED_DATASET
+#define ARM_COMPUTE_TEST_LOCALLYCONNECTED_DATASET
+
+#include "utils/TypePrinter.h"
+
+#include "arm_compute/core/TensorShape.h"
+#include "arm_compute/core/Types.h"
+
+#include "tests/datasets/ConvolutionLayerDataset.h"
+
+namespace arm_compute
+{
+namespace test
+{
+namespace datasets
+{
+class SmallLocallyConnectedDataset final : public ConvolutionLayerDataset
+{
+public:
+    SmallLocallyConnectedDataset()
+    {
+        // Batch size 1
+        add_config(TensorShape(23U, 27U, 5U), TensorShape(3U, 3U, 5U, 21U, 275U), TensorShape(21U, 275U), TensorShape(11U, 25U, 21U), PadStrideInfo(2, 1, 0, 0));
+        add_config(TensorShape(17U, 31U, 2U), TensorShape(5U, 5U, 2U, 19U, 225U), TensorShape(19U, 225U), TensorShape(15U, 15U, 19U), PadStrideInfo(1, 2, 1, 1));
+        add_config(TensorShape(17U, 31U, 2U), TensorShape(5U, 3U, 2U, 19U, 240U), TensorShape(19U, 240U), TensorShape(15U, 16U, 19U), PadStrideInfo(1, 2, 1, 1));
+        // Batch size 4
+        add_config(TensorShape(23U, 27U, 5U, 4U), TensorShape(3U, 3U, 5U, 21U, 275U), TensorShape(21U, 275U), TensorShape(11U, 25U, 21U, 4U), PadStrideInfo(2, 1, 0, 0));
+        add_config(TensorShape(17U, 31U, 2U, 4U), TensorShape(5U, 5U, 2U, 19U, 225U), TensorShape(19U, 225U), TensorShape(15U, 15U, 19U, 4U), PadStrideInfo(1, 2, 1, 1));
+        add_config(TensorShape(17U, 31U, 2U, 4U), TensorShape(5U, 3U, 2U, 19U, 240U), TensorShape(19U, 240U), TensorShape(15U, 16U, 19U, 4U), PadStrideInfo(1, 2, 1, 1));
+        // FC convolution
+        add_config(TensorShape(1U, 1U, 1024U), TensorShape(1U, 1U, 1024U, 1001U, 1U), TensorShape(1001U, 1U), TensorShape(1U, 1U, 1001U), PadStrideInfo(1, 1, 0, 0));
+    }
+};
+
+class LargeLocallyConnectedDataset final : public ConvolutionLayerDataset
+{
+public:
+    LargeLocallyConnectedDataset()
+    {
+        // Batch size 1
+        add_config(TensorShape(23U, 27U, 5U), TensorShape(3U, 1U, 5U, 21U, 297U), TensorShape(21U, 297U), TensorShape(11U, 27U, 21U), PadStrideInfo(2, 1, 0, 0));
+        add_config(TensorShape(33U, 27U, 7U), TensorShape(5U, 5U, 7U, 16U, 132U), TensorShape(16U, 132U), TensorShape(11U, 12U, 16U), PadStrideInfo(3, 2, 1, 0));
+        add_config(TensorShape(33U, 27U, 7U), TensorShape(5U, 7U, 7U, 16U, 121U), TensorShape(16U, 121U), TensorShape(11U, 11U, 16U), PadStrideInfo(3, 2, 1, 0));
+        // Batch size 4
+        add_config(TensorShape(23U, 27U, 5U, 4U), TensorShape(3U, 1U, 5U, 21U, 297U), TensorShape(21U, 297U), TensorShape(11U, 27U, 21U, 4U), PadStrideInfo(2, 1, 0, 0));
+        add_config(TensorShape(33U, 27U, 7U, 4U), TensorShape(5U, 5U, 7U, 16U, 132U), TensorShape(16U, 132U), TensorShape(11U, 12U, 16U, 4U), PadStrideInfo(3, 2, 1, 0));
+        add_config(TensorShape(33U, 27U, 7U, 4U), TensorShape(5U, 7U, 7U, 16U, 121U), TensorShape(16U, 121U), TensorShape(11U, 11U, 16U, 4U), PadStrideInfo(3, 2, 1, 0));
+        // Arbitrary batch size
+        add_config(TensorShape(33U, 27U, 7U, 5U), TensorShape(5U, 7U, 7U, 16U, 121U), TensorShape(16U, 121U), TensorShape(11U, 11U, 16U, 5U), PadStrideInfo(3, 2, 1, 0));
+    }
+};
+
+} // namespace datasets
+} // namespace test
+} // namespace arm_compute
+#endif /* ARM_COMPUTE_TEST_LOCALLYCONNECTED_DATASET */
diff --git a/tests/validation/CL/LocallyConnected.cpp b/tests/validation/CL/LocallyConnected.cpp
new file mode 100644
index 0000000000..05cab29226
--- /dev/null
+++ b/tests/validation/CL/LocallyConnected.cpp
@@ -0,0 +1,95 @@
+/*
+ * Copyright (c) 2017-2018 ARM Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#include "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/CLLocallyConnectedLayer.h"
+#include "tests/CL/CLAccessor.h"
+#include "tests/PaddingCalculator.h"
+#include "tests/datasets/LocallyConnectedDataset.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/LocallyConnectedFixture.h"
+
+namespace arm_compute
+{
+namespace test
+{
+namespace validation
+{
+namespace
+{
+} // namespace
+
+TEST_SUITE(CL)
+TEST_SUITE(LocallyConnected)
+
+DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(framework::dataset::concat(datasets::SmallLocallyConnectedDataset(), datasets::LargeLocallyConnectedDataset()),
+                                                                   framework::dataset::make("DataType", DataType::F32)),
+               src_shape, weights_shape, bias_shape, dst_shape, info, data_type)
+{
+    // Create tensors
+    CLTensor src     = create_tensor<CLTensor>(src_shape, data_type);
+    CLTensor weights = create_tensor<CLTensor>(weights_shape, data_type);
+    CLTensor bias    = create_tensor<CLTensor>(bias_shape, data_type);
+    CLTensor dst     = create_tensor<CLTensor>(dst_shape, data_type);
+
+    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.
+    CLLocallyConnectedLayer lc;
+    lc.configure(&src, &weights, &bias, &dst, info);
+
+    // 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 CLLocallyConnectedFixture = LocallyConnectedValidationFixture<CLTensor, CLAccessor, CLLocallyConnectedLayer, T>;
+FIXTURE_DATA_TEST_CASE(RunSmall, CLLocallyConnectedFixture<float>, framework::DatasetMode::PRECOMMIT, combine(datasets::SmallLocallyConnectedDataset(),
+                                                                                                              framework::dataset::make("DataType",
+                                                                                                                      DataType::F32)))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference);
+}
+
+FIXTURE_DATA_TEST_CASE(RunLarge, CLLocallyConnectedFixture<float>, framework::DatasetMode::NIGHTLY, combine(datasets::LargeLocallyConnectedDataset(),
+                                                                                                            framework::dataset::make("DataType",
+                                                                                                                    DataType::F32)))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference);
+}
+TEST_SUITE_END()
+TEST_SUITE_END()
+} // namespace validation
+} // namespace test
+} // namespace arm_compute
diff --git a/tests/validation/NEON/LocallyConnected.cpp b/tests/validation/NEON/LocallyConnected.cpp
new file mode 100644
index 0000000000..56430d9650
--- /dev/null
+++ b/tests/validation/NEON/LocallyConnected.cpp
@@ -0,0 +1,96 @@
+/*
+ * Copyright (c) 2017-2018 ARM Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#include "arm_compute/core/Types.h"
+#include "arm_compute/runtime/NEON/functions/NELocallyConnectedLayer.h"
+#include "arm_compute/runtime/Tensor.h"
+#include "arm_compute/runtime/TensorAllocator.h"
+#include "tests/NEON/Accessor.h"
+#include "tests/PaddingCalculator.h"
+#include "tests/datasets/LocallyConnectedDataset.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/LocallyConnectedFixture.h"
+
+namespace arm_compute
+{
+namespace test
+{
+namespace validation
+{
+namespace
+{
+constexpr RelativeTolerance<float> tolerance_f32(0.0001f); /**< Tolerance value for comparing reference's output against implementation's output for DataType::F32 */
+} // namespace
+
+TEST_SUITE(NEON)
+TEST_SUITE(LocallyConnected)
+
+DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(framework::dataset::concat(datasets::SmallLocallyConnectedDataset(), datasets::LargeLocallyConnectedDataset()),
+                                                                   framework::dataset::make("DataType", DataType::F32)),
+               src_shape, weights_shape, bias_shape, dst_shape, info, data_type)
+{
+    // Create tensors
+    Tensor src     = create_tensor<Tensor>(src_shape, data_type);
+    Tensor weights = create_tensor<Tensor>(weights_shape, data_type);
+    Tensor bias    = create_tensor<Tensor>(bias_shape, data_type);
+    Tensor dst     = create_tensor<Tensor>(dst_shape, data_type);
+
+    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.
+    NELocallyConnectedLayer lc;
+    lc.configure(&src, &weights, &bias, &dst, info);
+
+    // 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 NELocallyConnectedFixture = LocallyConnectedValidationFixture<Tensor, Accessor, NELocallyConnectedLayer, T>;
+FIXTURE_DATA_TEST_CASE(RunSmall, NELocallyConnectedFixture<float>, framework::DatasetMode::PRECOMMIT, combine(datasets::SmallLocallyConnectedDataset(),
+                                                                                                              framework::dataset::make("DataType",
+                                                                                                                      DataType::F32)))
+{
+    // Validate output
+    validate(Accessor(_target), _reference, tolerance_f32);
+}
+
+FIXTURE_DATA_TEST_CASE(RunLarge, NELocallyConnectedFixture<float>, framework::DatasetMode::PRECOMMIT, combine(datasets::LargeLocallyConnectedDataset(),
+                                                                                                              framework::dataset::make("DataType",
+                                                                                                                      DataType::F32)))
+{
+    // Validate output
+    validate(Accessor(_target), _reference, tolerance_f32);
+}
+TEST_SUITE_END()
+TEST_SUITE_END()
+} // namespace validation
+} // namespace test
+} // namespace arm_compute
diff --git a/tests/validation/fixtures/LocallyConnectedFixture.h b/tests/validation/fixtures/LocallyConnectedFixture.h
new file mode 100644
index 0000000000..ab9819e56f
--- /dev/null
+++ b/tests/validation/fixtures/LocallyConnectedFixture.h
@@ -0,0 +1,133 @@
+/*
+ * Copyright (c) 2017-2018 ARM Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#ifndef ARM_COMPUTE_TEST_LOCALLY_CONNECTED_FIXTURE
+#define ARM_COMPUTE_TEST_LOCALLY_CONNECTED_FIXTURE
+
+#include "arm_compute/core/TensorShape.h"
+#include "arm_compute/core/Types.h"
+#include "arm_compute/runtime/NEON/NEScheduler.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/LocallyConnected.h"
+#include "tests/validation/reference/Utils.h"
+
+#include <random>
+
+namespace arm_compute
+{
+class NELocallyConnected;
+
+namespace test
+{
+namespace validation
+{
+template <typename TensorType, typename AccessorType, typename FunctionType, typename T>
+class LocallyConnectedValidationFixture : public framework::Fixture
+{
+public:
+    using TBias = typename std::conditional<std::is_same<typename std::decay<T>::type, uint8_t>::value, int32_t, T>::type;
+
+public:
+    template <typename...>
+    void setup(TensorShape input_shape, TensorShape weights_shape, TensorShape bias_shape, TensorShape output_shape, PadStrideInfo info, DataType data_type)
+    {
+        _data_type      = data_type;
+        _bias_data_type = data_type;
+
+        _target    = compute_target(input_shape, weights_shape, bias_shape, output_shape, info);
+        _reference = compute_reference(input_shape, weights_shape, bias_shape, output_shape, info);
+    }
+
+protected:
+    template <typename U>
+    void fill(U &&tensor, int i)
+    {
+        std::uniform_real_distribution<> distribution(-1.0f, 1.0f);
+        library->fill(tensor, distribution, i);
+    }
+
+    TensorType compute_target(const TensorShape &input_shape, const TensorShape &weights_shape, const TensorShape &bias_shape, const TensorShape &output_shape, const PadStrideInfo &info)
+    {
+        TensorShape reshaped_weights_shape(weights_shape);
+
+        // Create tensors
+        TensorType src     = create_tensor<TensorType>(input_shape, _data_type);
+        TensorType weights = create_tensor<TensorType>(reshaped_weights_shape, _data_type);
+        TensorType bias    = create_tensor<TensorType>(bias_shape, _bias_data_type);
+        TensorType dst     = create_tensor<TensorType>(output_shape, _data_type);
+
+        // Create and configure function
+        FunctionType locally_connected;
+        locally_connected.configure(&src, &weights, &bias, &dst, info);
+
+        // 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(weights), 1);
+        fill(AccessorType(bias), 2);
+
+        locally_connected.run();
+
+        return dst;
+    }
+
+    SimpleTensor<T> compute_reference(const TensorShape &input_shape, const TensorShape &weights_shape, const TensorShape &bias_shape, const TensorShape &output_shape, const PadStrideInfo &info)
+    {
+        // Create reference
+        SimpleTensor<T>     src(input_shape, _data_type);
+        SimpleTensor<T>     weights(weights_shape, _data_type);
+        SimpleTensor<TBias> bias(bias_shape, _bias_data_type);
+
+        // Fill reference
+        fill(src, 0);
+        fill(weights, 1);
+        fill(bias, 2);
+
+        return reference::locally_connected<T>(src, weights, bias, output_shape, info);
+    }
+
+    TensorType      _target{};
+    SimpleTensor<T> _reference{};
+    DataType        _data_type{};
+    DataType        _bias_data_type{};
+};
+
+} // namespace validation
+} // namespace test
+} // namespace arm_compute
+#endif /* ARM_COMPUTE_TEST_LOCALLY_CONNECTED_FIXTURE */
diff --git a/tests/validation/reference/Convolution3d.h b/tests/validation/reference/Convolution3d.h
new file mode 100644
index 0000000000..b99d534635
--- /dev/null
+++ b/tests/validation/reference/Convolution3d.h
@@ -0,0 +1,223 @@
+/*
+ * Copyright (c) 2017-2018 ARM Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *asymm_int_mult
+ * 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, asymm_int_multDAMAGES 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_VALIDATION_CONVOLUTION_H__
+#define __ARM_COMPUTE_TEST_VALIDATION_CONVOLUTION_H__
+
+#include "arm_compute/core/utils/quantization/AsymmHelpers.h"
+#include "tests/validation/FixedPoint.h"
+#include "tests/validation/Helpers.h"
+#include "tests/validation/reference/UtilsQuantizedAsymm.h"
+
+namespace arm_compute
+{
+namespace test
+{
+namespace convolution_3d
+{
+namespace detail
+{
+inline bool is_valid_pixel(int i, int min, int max)
+{
+    return (i >= min && i < max);
+}
+
+// 3D convolution for floating point type
+template < typename T, typename TB, typename std::enable_if < validation::is_floating_point<T>::value &&validation::is_floating_point<TB>::value, int >::type = 0 >
+inline void convolution3d(const SimpleTensor<T> &in, const SimpleTensor<T> &weights, const SimpleTensor<TB> &bias, SimpleTensor<T> &out,
+                          int i_offset, int w_offset, int b_offset, int o_offset,
+                          int xi, int yi, int width_in, int height_in, int depth_in, int width_weights, int height_weights)
+{
+    const T *in_ptr  = in.data() + i_offset;
+    const T *w_ptr   = weights.data() + w_offset;
+    const TB *b_ptr   = bias.data() + b_offset;
+    T        *out_ptr = out.data() + o_offset;
+
+    const int half_width_weights_start  = width_weights / 2;
+    const int half_width_weights_end    = ((width_weights % 2) == 0) ? (half_width_weights_start - 1) : half_width_weights_start;
+    const int half_height_weights_start = height_weights / 2;
+    const int half_height_weights_end   = ((height_weights % 2) == 0) ? (half_height_weights_start - 1) : half_height_weights_start;
+
+    // Reset accumulator
+    T acc(0);
+
+    // Compute a 2D convolution for each IFM and accumulate the result
+    for(int ifm = 0; ifm < depth_in; ++ifm)
+    {
+        // Compute the offset for the input slice
+        const int offset_slice_in = xi + yi * width_in + ifm * width_in * height_in;
+
+        // Compute 2D convolution
+        for(int yk = -half_height_weights_start; yk <= half_height_weights_end; ++yk)
+        {
+            for(int xk = -half_width_weights_start; xk <= half_width_weights_end; ++xk)
+            {
+                // Check if the pixel is out-of-bound
+                if(is_valid_pixel(xi + xk, 0, width_in) && is_valid_pixel(yi + yk, 0, height_in))
+                {
+                    const int idx = xk + half_width_weights_start;
+                    const int idy = yk + half_height_weights_start;
+
+                    const T i_value = in_ptr[offset_slice_in + xk + yk * width_in];
+                    const T w_value = w_ptr[idx + idy * width_weights + ifm * width_weights * height_weights];
+
+                    acc += i_value * w_value;
+                }
+            }
+        }
+    }
+
+    // Accumulate the bias and store the result
+    *out_ptr = acc + (*b_ptr);
+}
+
+// 3D convolution for fixed point type
+template < typename T, typename TB, typename std::enable_if < std::is_integral<T>::value &&std::is_integral<TB>::value, int >::type = 0 >
+inline void convolution3d(const SimpleTensor<T> &in, const SimpleTensor<T> &weights, const SimpleTensor<TB> &bias, SimpleTensor<T> &out,
+                          int i_offset, int w_offset, int b_offset, int o_offset,
+                          int xi, int yi, int width_in, int height_in, int depth_in, int width_weights, int height_weights)
+{
+    const T *in_ptr               = in.data() + i_offset;
+    const T *w_ptr                = weights.data() + w_offset;
+    const T *b_ptr                = bias.data() + b_offset;
+    T       *out_ptr              = out.data() + o_offset;
+    int      fixed_point_position = in.fixed_point_position();
+
+    const int half_width_weights_start  = width_weights / 2;
+    const int half_width_weights_end    = ((width_weights % 2) == 0) ? (half_width_weights_start - 1) : half_width_weights_start;
+    const int half_height_weights_start = height_weights / 2;
+    const int half_height_weights_end   = ((height_weights % 2) == 0) ? (half_height_weights_start - 1) : half_height_weights_start;
+
+    using namespace fixed_point_arithmetic;
+    using promoted_type = fixed_point_arithmetic::traits::promote_t<T>;
+
+    // Reset accumulator
+    fixed_point<promoted_type> acc(0, fixed_point_position);
+
+    // Compute a 2D convolution for each IFM and accumulate the result
+    for(int ifm = 0; ifm < depth_in; ++ifm)
+    {
+        // Compute the offset for the input slice
+        const int offset_slice_in = xi + yi * width_in + ifm * width_in * height_in;
+
+        // Compute 2D convolution
+        for(int yk = -half_height_weights_start; yk <= half_height_weights_end; ++yk)
+        {
+            for(int xk = -half_width_weights_start; xk <= half_width_weights_end; ++xk)
+            {
+                // Check if the pixel is out-of-bound
+                if(is_valid_pixel(xi + xk, 0, width_in) && is_valid_pixel(yi + yk, 0, height_in))
+                {
+                    const int idx = xk + half_width_weights_start;
+                    const int idy = yk + half_height_weights_start;
+
+                    const fixed_point<promoted_type> i_value(in_ptr[offset_slice_in + xk + yk * width_in], fixed_point_position, true);
+                    const fixed_point<promoted_type> w_value(w_ptr[idx + idy * width_weights + ifm * width_weights * height_weights], fixed_point_position, true);
+                    const fixed_point<promoted_type> iw = i_value * w_value;
+                    acc                                 = iw + acc;
+                }
+            }
+        }
+    }
+
+    // Get the bias
+    const fixed_point<promoted_type> b(*b_ptr, fixed_point_position, true);
+
+    // Accumulate the bias and covert back
+    acc = acc + b;
+    fixed_point<T> res(acc);
+    *out_ptr = res.raw();
+}
+
+// 3D convolution for QASYMM8 type
+template <>
+inline void convolution3d(const SimpleTensor<uint8_t> &in, const SimpleTensor<uint8_t> &weights, const SimpleTensor<int32_t> &bias, SimpleTensor<uint8_t> &out,
+                          int i_offset, int w_offset, int b_offset, int o_offset,
+                          int xi, int yi, int width_in, int height_in, int depth_in, int width_weights, int height_weights)
+{
+    const uint8_t *in_ptr  = in.data() + i_offset;
+    const uint8_t *w_ptr   = weights.data() + w_offset;
+    const int32_t *b_ptr   = bias.data() + b_offset;
+    uint8_t       *out_ptr = out.data() + o_offset;
+
+    const int   input_offset   = -in.quantization_info().offset;
+    const float input_scale    = in.quantization_info().scale;
+    const int   weights_offset = -weights.quantization_info().offset;
+    const float weights_scale  = weights.quantization_info().scale;
+    const int   output_offset  = out.quantization_info().offset;
+    const float output_scale   = out.quantization_info().scale;
+
+    int         output_multiplier = 0;
+    int         output_shift      = 0;
+    const float multiplier        = input_scale * weights_scale / output_scale;
+    arm_compute::quantization::calculate_quantized_multiplier_less_than_one(multiplier, &output_multiplier, &output_shift);
+
+    const int half_width_weights_start  = width_weights / 2;
+    const int half_width_weights_end    = ((width_weights % 2) == 0) ? (half_width_weights_start - 1) : half_width_weights_start;
+    const int half_height_weights_start = height_weights / 2;
+    const int half_height_weights_end   = ((height_weights % 2) == 0) ? (half_height_weights_start - 1) : half_height_weights_start;
+
+    // Reset accumulator
+    int32_t acc(0);
+
+    // Compute a 2D convolution for each IFM and accumulate the result
+    for(int ifm = 0; ifm < depth_in; ++ifm)
+    {
+        // Compute the offset for the input slice
+        const int offset_slice_in = xi + yi * width_in + ifm * width_in * height_in;
+
+        // Compute 2D convolution
+        for(int yk = -half_height_weights_start; yk <= half_height_weights_end; ++yk)
+        {
+            for(int xk = -half_width_weights_start; xk <= half_width_weights_end; ++xk)
+            {
+                // Check if the pixel is out-of-bound
+                if(is_valid_pixel(xi + xk, 0, width_in) && is_valid_pixel(yi + yk, 0, height_in))
+                {
+                    const int idx = xk + half_width_weights_start;
+                    const int idy = yk + half_height_weights_start;
+
+                    const uint8_t i_value = in_ptr[offset_slice_in + xk + yk * width_in];
+                    const uint8_t w_value = w_ptr[idx + idy * width_weights + ifm * width_weights * height_weights];
+
+                    acc += (i_value + input_offset) * (w_value + weights_offset);
+                }
+            }
+        }
+    }
+
+    // Accumulate the bias
+    acc += (*b_ptr);
+
+    acc = validation::asymm_rounding_divide_by_pow2(validation::asymm_int_mult(acc, output_multiplier), output_shift);
+    acc += output_offset;
+    acc = utility::clamp<int32_t>(acc, 0, 255);
+
+    // Store the result
+    *out_ptr = acc;
+}
+} // namespace detail
+} // namespace convolution_3d
+} // namespace test
+} // namespace arm_compute
+#endif /*__ARM_COMPUTE_TEST_VALIDATION_CONVOLUTION_H__ */
diff --git a/tests/validation/reference/ConvolutionLayer.cpp b/tests/validation/reference/ConvolutionLayer.cpp
index b7ed2f56c0..24bbf32a30 100644
--- a/tests/validation/reference/ConvolutionLayer.cpp
+++ b/tests/validation/reference/ConvolutionLayer.cpp
@@ -25,6 +25,7 @@
 
 #include "tests/validation/FixedPoint.h"
 #include "tests/validation/Helpers.h"
+#include "tests/validation/reference/Convolution3d.h"
 #include "tests/validation/reference/Utils.h"
 #include "tests/validation/reference/UtilsQuantizedAsymm.h"
 
@@ -42,185 +43,6 @@ namespace reference
 {
 namespace
 {
-inline bool is_valid_pixel(int i, int min, int max)
-{
-    return (i >= min && i < max);
-}
-
-// 3D convolution for floating point type
-template < typename T, typename TB, typename std::enable_if < is_floating_point<T>::value &&is_floating_point<TB>::value, int >::type = 0 >
-void convolution3d(const SimpleTensor<T> &in, const SimpleTensor<T> &weights, const SimpleTensor<TB> &bias, SimpleTensor<T> &out,
-                   int i_offset, int w_offset, int b_offset, int o_offset,
-                   int xi, int yi, int width_in, int height_in, int depth_in, int width_weights, int height_weights)
-{
-    const T *in_ptr  = in.data() + i_offset;
-    const T *w_ptr   = weights.data() + w_offset;
-    const TB *b_ptr   = bias.data() + b_offset;
-    T        *out_ptr = out.data() + o_offset;
-
-    const int half_width_weights_start  = width_weights / 2;
-    const int half_width_weights_end    = ((width_weights % 2) == 0) ? (half_width_weights_start - 1) : half_width_weights_start;
-    const int half_height_weights_start = height_weights / 2;
-    const int half_height_weights_end   = ((height_weights % 2) == 0) ? (half_height_weights_start - 1) : half_height_weights_start;
-
-    // Reset accumulator
-    T acc(0);
-
-    // Compute a 2D convolution for each IFM and accumulate the result
-    for(int ifm = 0; ifm < depth_in; ++ifm)
-    {
-        // Compute the offset for the input slice
-        const int offset_slice_in = xi + yi * width_in + ifm * width_in * height_in;
-
-        // Compute 2D convolution
-        for(int yk = -half_height_weights_start; yk <= half_height_weights_end; ++yk)
-        {
-            for(int xk = -half_width_weights_start; xk <= half_width_weights_end; ++xk)
-            {
-                // Check if the pixel is out-of-bound
-                if(is_valid_pixel(xi + xk, 0, width_in) && is_valid_pixel(yi + yk, 0, height_in))
-                {
-                    const int idx = xk + half_width_weights_start;
-                    const int idy = yk + half_height_weights_start;
-
-                    const T i_value = in_ptr[offset_slice_in + xk + yk * width_in];
-                    const T w_value = w_ptr[idx + idy * width_weights + ifm * width_weights * height_weights];
-
-                    acc += i_value * w_value;
-                }
-            }
-        }
-    }
-
-    // Accumulate the bias and store the result
-    *out_ptr = acc + (*b_ptr);
-}
-
-// 3D convolution for fixed point type
-template < typename T, typename TB, typename std::enable_if < std::is_integral<T>::value &&std::is_integral<TB>::value, int >::type = 0 >
-void convolution3d(const SimpleTensor<T> &in, const SimpleTensor<T> &weights, const SimpleTensor<TB> &bias, SimpleTensor<T> &out,
-                   int i_offset, int w_offset, int b_offset, int o_offset,
-                   int xi, int yi, int width_in, int height_in, int depth_in, int width_weights, int height_weights)
-{
-    const T *in_ptr               = in.data() + i_offset;
-    const T *w_ptr                = weights.data() + w_offset;
-    const T *b_ptr                = bias.data() + b_offset;
-    T       *out_ptr              = out.data() + o_offset;
-    int      fixed_point_position = in.fixed_point_position();
-
-    const int half_width_weights_start  = width_weights / 2;
-    const int half_width_weights_end    = ((width_weights % 2) == 0) ? (half_width_weights_start - 1) : half_width_weights_start;
-    const int half_height_weights_start = height_weights / 2;
-    const int half_height_weights_end   = ((height_weights % 2) == 0) ? (half_height_weights_start - 1) : half_height_weights_start;
-
-    using namespace fixed_point_arithmetic;
-    using promoted_type = fixed_point_arithmetic::traits::promote_t<T>;
-
-    // Reset accumulator
-    fixed_point<promoted_type> acc(0, fixed_point_position);
-
-    // Compute a 2D convolution for each IFM and accumulate the result
-    for(int ifm = 0; ifm < depth_in; ++ifm)
-    {
-        // Compute the offset for the input slice
-        const int offset_slice_in = xi + yi * width_in + ifm * width_in * height_in;
-
-        // Compute 2D convolution
-        for(int yk = -half_height_weights_start; yk <= half_height_weights_end; ++yk)
-        {
-            for(int xk = -half_width_weights_start; xk <= half_width_weights_end; ++xk)
-            {
-                // Check if the pixel is out-of-bound
-                if(is_valid_pixel(xi + xk, 0, width_in) && is_valid_pixel(yi + yk, 0, height_in))
-                {
-                    const int idx = xk + half_width_weights_start;
-                    const int idy = yk + half_height_weights_start;
-
-                    const fixed_point<promoted_type> i_value(in_ptr[offset_slice_in + xk + yk * width_in], fixed_point_position, true);
-                    const fixed_point<promoted_type> w_value(w_ptr[idx + idy * width_weights + ifm * width_weights * height_weights], fixed_point_position, true);
-                    const fixed_point<promoted_type> iw = i_value * w_value;
-                    acc                                 = iw + acc;
-                }
-            }
-        }
-    }
-
-    // Get the bias
-    const fixed_point<promoted_type> b(*b_ptr, fixed_point_position, true);
-
-    // Accumulate the bias and covert back
-    acc = acc + b;
-    fixed_point<T> res(acc);
-    *out_ptr = res.raw();
-}
-
-// 3D convolution for QASYMM8 type
-template <>
-void convolution3d(const SimpleTensor<uint8_t> &in, const SimpleTensor<uint8_t> &weights, const SimpleTensor<int32_t> &bias, SimpleTensor<uint8_t> &out,
-                   int i_offset, int w_offset, int b_offset, int o_offset,
-                   int xi, int yi, int width_in, int height_in, int depth_in, int width_weights, int height_weights)
-{
-    const uint8_t *in_ptr  = in.data() + i_offset;
-    const uint8_t *w_ptr   = weights.data() + w_offset;
-    const int32_t *b_ptr   = bias.data() + b_offset;
-    uint8_t       *out_ptr = out.data() + o_offset;
-
-    const int   input_offset   = -in.quantization_info().offset;
-    const float input_scale    = in.quantization_info().scale;
-    const int   weights_offset = -weights.quantization_info().offset;
-    const float weights_scale  = weights.quantization_info().scale;
-    const int   output_offset  = out.quantization_info().offset;
-    const float output_scale   = out.quantization_info().scale;
-
-    int         output_multiplier = 0;
-    int         output_shift      = 0;
-    const float multiplier        = input_scale * weights_scale / output_scale;
-    arm_compute::quantization::calculate_quantized_multiplier_less_than_one(multiplier, &output_multiplier, &output_shift);
-
-    const int half_width_weights_start  = width_weights / 2;
-    const int half_width_weights_end    = ((width_weights % 2) == 0) ? (half_width_weights_start - 1) : half_width_weights_start;
-    const int half_height_weights_start = height_weights / 2;
-    const int half_height_weights_end   = ((height_weights % 2) == 0) ? (half_height_weights_start - 1) : half_height_weights_start;
-
-    // Reset accumulator
-    int32_t acc(0);
-
-    // Compute a 2D convolution for each IFM and accumulate the result
-    for(int ifm = 0; ifm < depth_in; ++ifm)
-    {
-        // Compute the offset for the input slice
-        const int offset_slice_in = xi + yi * width_in + ifm * width_in * height_in;
-
-        // Compute 2D convolution
-        for(int yk = -half_height_weights_start; yk <= half_height_weights_end; ++yk)
-        {
-            for(int xk = -half_width_weights_start; xk <= half_width_weights_end; ++xk)
-            {
-                // Check if the pixel is out-of-bound
-                if(is_valid_pixel(xi + xk, 0, width_in) && is_valid_pixel(yi + yk, 0, height_in))
-                {
-                    const int idx = xk + half_width_weights_start;
-                    const int idy = yk + half_height_weights_start;
-
-                    const uint8_t i_value = in_ptr[offset_slice_in + xk + yk * width_in];
-                    const uint8_t w_value = w_ptr[idx + idy * width_weights + ifm * width_weights * height_weights];
-
-                    acc += (i_value + input_offset) * (w_value + weights_offset);
-                }
-            }
-        }
-    }
-
-    // Accumulate the bias
-    acc += (*b_ptr);
-
-    acc = asymm_rounding_divide_by_pow2(asymm_int_mult(acc, output_multiplier), output_shift);
-    acc += output_offset;
-    acc = utility::clamp<int32_t>(acc, 0, 255);
-
-    // Store the result
-    *out_ptr = acc;
-}
 } // namespace
 
 template <typename T, typename TB>
@@ -270,11 +92,11 @@ SimpleTensor<T> convolution_layer(const SimpleTensor<T> &src, const SimpleTensor
                     ARM_COMPUTE_ASSERT(yo < height_out);
 
                     // Compute 3D convolution
-                    convolution3d(src, weights, bias, dst,
-                                  offset_in, ofm * width_weights * height_weights * depth_weights, ofm, offset_out,
-                                  xi, yi,
-                                  width_in, height_in, depth_in,
-                                  width_weights, height_weights);
+                    convolution_3d::detail::convolution3d(src, weights, bias, dst,
+                                                          offset_in, ofm * width_weights * height_weights * depth_weights, ofm, offset_out,
+                                                          xi, yi,
+                                                          width_in, height_in, depth_in,
+                                                          width_weights, height_weights);
                 }
             }
         }
diff --git a/tests/validation/reference/LocallyConnected.cpp b/tests/validation/reference/LocallyConnected.cpp
new file mode 100644
index 0000000000..08e3f02761
--- /dev/null
+++ b/tests/validation/reference/LocallyConnected.cpp
@@ -0,0 +1,111 @@
+/*
+ * Copyright (c) 2017-2018 ARM Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#include "LocallyConnected.h"
+
+#include "tests/validation/Helpers.h"
+#include "tests/validation/reference/Convolution3d.h"
+#include "tests/validation/reference/Utils.h"
+
+#include "tests/framework/Asserts.h"
+
+namespace arm_compute
+{
+namespace test
+{
+namespace validation
+{
+namespace reference
+{
+template <typename T, typename TB>
+SimpleTensor<T> locally_connected(const SimpleTensor<T> &src, const SimpleTensor<T> &weights, const SimpleTensor<TB> &bias, const TensorShape &output_shape, const PadStrideInfo &info)
+{
+    // Create reference
+    SimpleTensor<T> dst{ output_shape, src.data_type(), 1, src.fixed_point_position(), src.quantization_info() };
+
+    // Compute reference
+    const int width_in  = src.shape().x();
+    const int height_in = src.shape().y();
+    const int depth_in  = src.shape().z();
+
+    const int width_out  = dst.shape().x();
+    const int height_out = dst.shape().y();
+    const int depth_out  = dst.shape().z();
+
+    const int width_weights  = weights.shape().x();
+    const int height_weights = weights.shape().y();
+    const int depth_weights  = weights.shape().z();
+
+    const int pad_left  = info.pad_left();
+    const int pad_top   = info.pad_top();
+    const int stride_xi = info.stride().first;
+    const int stride_yi = info.stride().second;
+
+    auto output_wh = scaled_dimensions(width_in, height_in, width_weights, height_weights, info);
+
+    const int start_xi    = width_weights / 2 - pad_left;
+    const int start_yi    = height_weights / 2 - pad_top;
+    const int end_xi      = output_wh.first * stride_xi;
+    const int end_yi      = output_wh.second * stride_yi;
+    const int num_batches = src.shape().total_size() / (width_in * height_in * depth_in);
+
+    for(int r = 0; r < num_batches; ++r)
+    {
+        int count = 0;
+        for(int yi = start_yi; yi < start_yi + end_yi; yi += stride_yi)
+        {
+            for(int xi = start_xi; xi < start_xi + end_xi; xi += stride_xi)
+            {
+                for(int ofm = 0; ofm < depth_out; ++ofm)
+                {
+                    // Compute input and output offsets
+                    const int offset_in  = r * width_in * height_in * depth_in;
+                    const int xo         = (xi - start_xi) / stride_xi;
+                    const int yo         = (yi - start_yi) / stride_yi;
+                    const int offset_out = xo + yo * width_out + ofm * width_out * height_out + r * width_out * height_out * depth_out;
+
+                    ARM_COMPUTE_ASSERT(xo < width_out);
+                    ARM_COMPUTE_ASSERT(yo < height_out);
+
+                    // Compute 3D convolution
+                    convolution_3d::detail::convolution3d(src, weights, bias, dst,
+                                                          offset_in, count * width_weights * height_weights * depth_weights, count, offset_out,
+                                                          xi, yi,
+                                                          width_in, height_in, depth_in,
+                                                          width_weights, height_weights);
+                    count++;
+                }
+            }
+        }
+    }
+
+    return dst;
+}
+
+// Locally Connected only supports F32
+template SimpleTensor<float> locally_connected(const SimpleTensor<float> &src, const SimpleTensor<float> &weights, const SimpleTensor<float> &bias, const TensorShape &output_shape,
+                                               const PadStrideInfo &info);
+} // namespace reference
+} // namespace validation
+} // namespace test
+} // namespace arm_compute
diff --git a/tests/validation/reference/LocallyConnected.h b/tests/validation/reference/LocallyConnected.h
new file mode 100644
index 0000000000..bf78d2c02a
--- /dev/null
+++ b/tests/validation/reference/LocallyConnected.h
@@ -0,0 +1,44 @@
+/*
+ * Copyright (c) 2017-2018 ARM Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#ifndef __ARM_COMPUTE_TEST_LOCALLY_CONNECTED_H__
+#define __ARM_COMPUTE_TEST_LOCALLY_CONNECTED_H__
+
+#include "tests/SimpleTensor.h"
+#include "tests/validation/Helpers.h"
+
+namespace arm_compute
+{
+namespace test
+{
+namespace validation
+{
+namespace reference
+{
+template <typename T, typename TB>
+SimpleTensor<T> locally_connected(const SimpleTensor<T> &src, const SimpleTensor<T> &weights, const SimpleTensor<TB> &bias, const TensorShape &output_shape, const PadStrideInfo &info);
+} // namespace reference
+} // namespace validation
+} // namespace test
+} // namespace arm_compute
+#endif /* __ARM_COMPUTE_TEST_LOCALLY_CONNECTED_H__ */