COMPMID-992 Implement CL RNN function

Change-Id: I8dbada5fabedbb8523e433ba73d504bd15b81466
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/125787
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
Tested-by: Jenkins <bsgcomp@arm.com>

3 files changed, 424 insertions, 0 deletions

diff --git a/tests/datasets/RNNLayerDataset.h b/tests/datasets/RNNLayerDataset.h
new file mode 100644
index 0000000000..616a69e213
--- /dev/null
+++ b/tests/datasets/RNNLayerDataset.h
@@ -0,0 +1,141 @@
+/*
+ * Copyright (c) 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_RNN_LAYER_DATASET
+#define ARM_COMPUTE_TEST_RNN_LAYER_DATASET
+
+#include "utils/TypePrinter.h"
+
+#include "arm_compute/core/TensorShape.h"
+#include "arm_compute/core/Types.h"
+
+namespace arm_compute
+{
+namespace test
+{
+namespace datasets
+{
+class RNNLayerDataset
+{
+public:
+    using type = std::tuple<TensorShape, TensorShape, TensorShape, TensorShape, TensorShape, ActivationLayerInfo>;
+
+    struct iterator
+    {
+        iterator(std::vector<TensorShape>::const_iterator         src_it,
+                 std::vector<TensorShape>::const_iterator         weights_it,
+                 std::vector<TensorShape>::const_iterator         recurrent_weights_it,
+                 std::vector<TensorShape>::const_iterator         biases_it,
+                 std::vector<TensorShape>::const_iterator         dst_it,
+                 std::vector<ActivationLayerInfo>::const_iterator infos_it)
+            : _src_it{ std::move(src_it) },
+              _weights_it{ std::move(weights_it) },
+              _recurrent_weights_it{ std::move(recurrent_weights_it) },
+              _biases_it{ std::move(biases_it) },
+              _dst_it{ std::move(dst_it) },
+              _infos_it{ std::move(infos_it) }
+        {
+        }
+
+        std::string description() const
+        {
+            std::stringstream description;
+            description << "In=" << *_src_it << ":";
+            description << "Weights=" << *_weights_it << ":";
+            description << "Biases=" << *_biases_it << ":";
+            description << "Out=" << *_dst_it;
+            return description.str();
+        }
+
+        RNNLayerDataset::type operator*() const
+        {
+            return std::make_tuple(*_src_it, *_weights_it, *_recurrent_weights_it, *_biases_it, *_dst_it, *_infos_it);
+        }
+
+        iterator &operator++()
+        {
+            ++_src_it;
+            ++_weights_it;
+            ++_recurrent_weights_it;
+            ++_biases_it;
+            ++_dst_it;
+            ++_infos_it;
+
+            return *this;
+        }
+
+    private:
+        std::vector<TensorShape>::const_iterator         _src_it;
+        std::vector<TensorShape>::const_iterator         _weights_it;
+        std::vector<TensorShape>::const_iterator         _recurrent_weights_it;
+        std::vector<TensorShape>::const_iterator         _biases_it;
+        std::vector<TensorShape>::const_iterator         _dst_it;
+        std::vector<ActivationLayerInfo>::const_iterator _infos_it;
+    };
+
+    iterator begin() const
+    {
+        return iterator(_src_shapes.begin(), _weight_shapes.begin(), _recurrent_weight_shapes.begin(), _bias_shapes.begin(), _dst_shapes.begin(), _infos.begin());
+    }
+
+    int size() const
+    {
+        return std::min(_src_shapes.size(), std::min(_weight_shapes.size(), std::min(_recurrent_weight_shapes.size(), std::min(_bias_shapes.size(), std::min(_dst_shapes.size(), _infos.size())))));
+    }
+
+    void add_config(TensorShape src, TensorShape weights, TensorShape recurrent_weights, TensorShape biases, TensorShape dst, ActivationLayerInfo info)
+    {
+        _src_shapes.emplace_back(std::move(src));
+        _weight_shapes.emplace_back(std::move(weights));
+        _recurrent_weight_shapes.emplace_back(std::move(recurrent_weights));
+        _bias_shapes.emplace_back(std::move(biases));
+        _dst_shapes.emplace_back(std::move(dst));
+        _infos.emplace_back(std::move(info));
+    }
+
+protected:
+    RNNLayerDataset()                   = default;
+    RNNLayerDataset(RNNLayerDataset &&) = default;
+
+private:
+    std::vector<TensorShape>         _src_shapes{};
+    std::vector<TensorShape>         _weight_shapes{};
+    std::vector<TensorShape>         _recurrent_weight_shapes{};
+    std::vector<TensorShape>         _bias_shapes{};
+    std::vector<TensorShape>         _dst_shapes{};
+    std::vector<ActivationLayerInfo> _infos{};
+};
+
+class SmallRNNLayerDataset final : public RNNLayerDataset
+{
+public:
+    SmallRNNLayerDataset()
+    {
+        add_config(TensorShape(8U, 2U), TensorShape(8U, 16U), TensorShape(16U, 16U), TensorShape(16U), TensorShape(16U, 2U), ActivationLayerInfo());
+    }
+};
+
+} // namespace datasets
+} // namespace test
+} // namespace arm_compute
+#endif /* ARM_COMPUTE_TEST_RNN_LAYER_DATASET */
diff --git a/tests/validation/CL/RNNLayer.cpp b/tests/validation/CL/RNNLayer.cpp
new file mode 100644
index 0000000000..0af6f8ea00
--- /dev/null
+++ b/tests/validation/CL/RNNLayer.cpp
@@ -0,0 +1,138 @@
+/*
+ * Copyright (c) 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/runtime/CL/functions/CLRNNLayer.h"
+#include "tests/CL/CLAccessor.h"
+#include "tests/PaddingCalculator.h"
+#include "tests/datasets/RNNLayerDataset.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/RNNLayerFixture.h"
+
+namespace arm_compute
+{
+namespace test
+{
+namespace validation
+{
+namespace
+{
+RelativeTolerance<float> tolerance_f32(0.001f);
+RelativeTolerance<half>  tolerance_f16(half(0.1));
+} // namespace
+
+TEST_SUITE(CL)
+TEST_SUITE(RNNLayer)
+
+// *INDENT-OFF*
+// clang-format off
+DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(zip(zip(zip(zip(
+               framework::dataset::make("InputInfo", { TensorInfo(TensorShape(27U, 13U), 1, DataType::U8, 0),      // Wrong data type
+                                                       TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32, 0), // Wrong input size
+                                                       TensorInfo(TensorShape(27U, 13U), 1, DataType::F32, 0),     // Wrong weights size
+                                                       TensorInfo(TensorShape(27U, 13U), 1, DataType::F32, 0),     // Wrong recurrent weights size
+                                                       TensorInfo(TensorShape(27U, 13U), 1, DataType::F32, 0),     // Wrong bias size
+                                                       TensorInfo(TensorShape(27U, 13U), 1, DataType::F32, 0),     // Wrong output size
+                                                       TensorInfo(TensorShape(27U, 13U), 1, DataType::F32, 0),     // Wrong hidden output size
+               }),
+               framework::dataset::make("WeightsInfo", { TensorInfo(TensorShape(27U, 11U), 1, DataType::F32, 0),
+                                                       TensorInfo(TensorShape(27U, 11U), 1, DataType::F32, 0),
+                                                       TensorInfo(TensorShape(27U, 11U, 2U), 1, DataType::F32, 0),
+                                                       TensorInfo(TensorShape(27U, 11U), 1, DataType::F32, 0),
+                                                       TensorInfo(TensorShape(27U, 11U), 1, DataType::F32, 0),
+                                                       TensorInfo(TensorShape(27U, 11U), 1, DataType::F32, 0),
+                                                       TensorInfo(TensorShape(27U, 11U), 1, DataType::F32, 0),
+               })),
+               framework::dataset::make("RecurrentWeightsInfo", { TensorInfo(TensorShape(11U, 11U), 1, DataType::F32, 0),
+                                                                  TensorInfo(TensorShape(11U, 11U), 1, DataType::F32, 0),
+                                                                  TensorInfo(TensorShape(11U, 11U), 1, DataType::F32, 0),
+                                                                  TensorInfo(TensorShape(25U, 11U, 2U), 1, DataType::F32, 0),
+                                                                  TensorInfo(TensorShape(11U, 11U), 1, DataType::F32, 0),
+                                                                  TensorInfo(TensorShape(11U, 11U), 1, DataType::F32, 0),
+                                                                  TensorInfo(TensorShape(11U, 11U), 1, DataType::F32, 0),
+               })),
+               framework::dataset::make("BiasInfo", { TensorInfo(TensorShape(11U), 1, DataType::F32, 0),
+                                                      TensorInfo(TensorShape(11U), 1, DataType::F32, 0),
+                                                      TensorInfo(TensorShape(11U), 1, DataType::F32, 0),
+                                                      TensorInfo(TensorShape(11U), 1, DataType::F32, 0),
+                                                      TensorInfo(TensorShape(30U), 1, DataType::F32, 0),
+                                                      TensorInfo(TensorShape(11U), 1, DataType::F32, 0),
+                                                      TensorInfo(TensorShape(11U), 1, DataType::F32, 0),
+               })),
+               framework::dataset::make("OutputInfo", { TensorInfo(TensorShape(11U, 13U), 1, DataType::F32, 0),
+                                                        TensorInfo(TensorShape(11U, 13U), 1, DataType::F32, 0),
+                                                        TensorInfo(TensorShape(11U, 13U), 1, DataType::F32, 0),
+                                                        TensorInfo(TensorShape(11U, 13U), 1, DataType::F32, 0),
+                                                        TensorInfo(TensorShape(11U, 13U), 1, DataType::F32, 0),
+                                                        TensorInfo(TensorShape(11U), 1, DataType::F32, 0),
+                                                        TensorInfo(TensorShape(11U, 13U), 1, DataType::F32, 0),
+               })),
+               framework::dataset::make("HiddenStateInfo", { TensorInfo(TensorShape(11U, 13U), 1, DataType::F32, 0),
+                                                             TensorInfo(TensorShape(11U, 13U), 1, DataType::F32, 0),
+                                                             TensorInfo(TensorShape(11U, 13U), 1, DataType::F32, 0),
+                                                             TensorInfo(TensorShape(11U, 13U), 1, DataType::F32, 0),
+                                                             TensorInfo(TensorShape(11U, 13U), 1, DataType::F32, 0),
+                                                             TensorInfo(TensorShape(11U, 13U), 1, DataType::F32, 0),
+                                                             TensorInfo(TensorShape(11U, 13U, 2U), 1, DataType::F32, 0),
+               })),
+               framework::dataset::make("ActivationInfo", { ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU),
+                                                            ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU),
+                                                            ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU),
+                                                            ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU),
+                                                            ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU),
+                                                            ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU),
+                                                            ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU),
+               })),
+               framework::dataset::make("Expected", { false, false, false, false, false, false, false })),
+               input_info, weights_info, recurrent_weights_info, bias_info, output_info, hidden_output_info, info, expected)
+{
+    ARM_COMPUTE_EXPECT(bool(CLRNNLayer::validate(&input_info.clone()->set_is_resizable(false), &weights_info.clone()->set_is_resizable(false), &recurrent_weights_info.clone()->set_is_resizable(false), &bias_info.clone()->set_is_resizable(false), &output_info.clone()->set_is_resizable(false), &hidden_output_info.clone()->set_is_resizable(false), info)) == expected, framework::LogLevel::ERRORS);
+}
+// clang-format on
+// *INDENT-ON*
+
+template <typename T>
+using CLRNNLayerFixture = RNNLayerValidationFixture<CLTensor, CLAccessor, CLRNNLayer, T>;
+
+TEST_SUITE(FP32)
+FIXTURE_DATA_TEST_CASE(RunSmall, CLRNNLayerFixture<float>, framework::DatasetMode::ALL, combine(datasets::SmallRNNLayerDataset(), framework::dataset::make("DataType", DataType::F32)))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference, tolerance_f32);
+}
+TEST_SUITE_END() // FP32
+
+TEST_SUITE(FP16)
+FIXTURE_DATA_TEST_CASE(RunSmall, CLRNNLayerFixture<half>, framework::DatasetMode::ALL, combine(datasets::SmallRNNLayerDataset(), framework::dataset::make("DataType", DataType::F16)))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference, tolerance_f16);
+}
+TEST_SUITE_END() // FP16
+TEST_SUITE_END() // RNNLayer
+TEST_SUITE_END() // CL
+} // namespace validation
+} // namespace test
+} // namespace arm_compute
diff --git a/tests/validation/fixtures/RNNLayerFixture.h b/tests/validation/fixtures/RNNLayerFixture.h
new file mode 100644
index 0000000000..42b99cce1c
--- /dev/null
+++ b/tests/validation/fixtures/RNNLayerFixture.h
@@ -0,0 +1,145 @@
+/*
+ * Copyright (c) 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_RNN_LAYER_FIXTURE
+#define ARM_COMPUTE_TEST_RNN_LAYER_FIXTURE
+
+#include "tests/Globals.h"
+#include "tests/framework/Asserts.h"
+#include "tests/framework/Fixture.h"
+#include "tests/validation/reference/ActivationLayer.h"
+#include "tests/validation/reference/ArithmeticAddition.h"
+#include "tests/validation/reference/FullyConnectedLayer.h"
+#include "tests/validation/reference/GEMM.h"
+
+namespace arm_compute
+{
+namespace test
+{
+namespace validation
+{
+template <typename TensorType, typename AccessorType, typename FunctionType, typename T>
+class RNNLayerValidationFixture : public framework::Fixture
+{
+public:
+    template <typename...>
+    void setup(TensorShape input_shape, TensorShape weights_shape, TensorShape recurrent_weights_shape, TensorShape bias_shape, TensorShape output_shape, ActivationLayerInfo info,
+               DataType data_type)
+    {
+        _target    = compute_target(input_shape, weights_shape, recurrent_weights_shape, bias_shape, output_shape, info, data_type);
+        _reference = compute_reference(input_shape, weights_shape, recurrent_weights_shape, bias_shape, output_shape, info, data_type);
+    }
+
+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 &recurrent_weights_shape, const TensorShape &bias_shape, const TensorShape &output_shape,
+                              ActivationLayerInfo info, DataType data_type)
+    {
+        // Create tensors
+        TensorType input             = create_tensor<TensorType>(input_shape, data_type);
+        TensorType weights           = create_tensor<TensorType>(weights_shape, data_type);
+        TensorType recurrent_weights = create_tensor<TensorType>(recurrent_weights_shape, data_type);
+        TensorType bias              = create_tensor<TensorType>(bias_shape, data_type);
+        TensorType hidden_state      = create_tensor<TensorType>(output_shape, data_type);
+        TensorType output            = create_tensor<TensorType>(output_shape, data_type);
+
+        // Create and configure function
+        FunctionType rnn;
+        rnn.configure(&input, &weights, &recurrent_weights, &bias, &hidden_state, &output, info);
+
+        ARM_COMPUTE_EXPECT(input.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(weights.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(recurrent_weights.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(bias.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(hidden_state.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(output.info()->is_resizable(), framework::LogLevel::ERRORS);
+
+        // Allocate tensors
+        input.allocator()->allocate();
+        weights.allocator()->allocate();
+        recurrent_weights.allocator()->allocate();
+        bias.allocator()->allocate();
+        hidden_state.allocator()->allocate();
+        output.allocator()->allocate();
+
+        ARM_COMPUTE_EXPECT(!input.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(!weights.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(!recurrent_weights.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(!bias.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(!hidden_state.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(!output.info()->is_resizable(), framework::LogLevel::ERRORS);
+
+        // Fill tensors
+        fill(AccessorType(input), 0);
+        fill(AccessorType(weights), 0);
+        fill(AccessorType(recurrent_weights), 0);
+        fill(AccessorType(bias), 0);
+        fill(AccessorType(hidden_state), 0);
+
+        // Compute function
+        rnn.run();
+
+        return output;
+    }
+
+    SimpleTensor<T> compute_reference(const TensorShape &input_shape, const TensorShape &weights_shape, const TensorShape &recurrent_weights_shape, const TensorShape &bias_shape,
+                                      const TensorShape &output_shape, ActivationLayerInfo info, DataType data_type)
+    {
+        // Create reference
+        SimpleTensor<T> input{ input_shape, data_type };
+        SimpleTensor<T> weights{ weights_shape, data_type };
+        SimpleTensor<T> recurrent_weights{ recurrent_weights_shape, data_type };
+        SimpleTensor<T> bias{ bias_shape, data_type };
+        SimpleTensor<T> hidden_state{ output_shape, data_type };
+
+        // Fill reference
+        fill(input, 0);
+        fill(weights, 0);
+        fill(recurrent_weights, 0);
+        fill(bias, 0);
+        fill(hidden_state, 0);
+
+        TensorShape out_shape = recurrent_weights_shape;
+        out_shape.set(1, output_shape.y());
+
+        // Compute reference
+        SimpleTensor<T> out_w{ out_shape, data_type };
+        SimpleTensor<T> fully_connected = reference::fully_connected_layer(input, weights, bias, out_shape);
+        SimpleTensor<T> gemm            = reference::gemm(hidden_state, recurrent_weights, out_w, 1.f, 0.f);
+        SimpleTensor<T> add_res         = reference::arithmetic_addition(fully_connected, gemm, data_type, ConvertPolicy::SATURATE);
+        return reference::activation_layer(add_res, info);
+    }
+
+    TensorType      _target{};
+    SimpleTensor<T> _reference{};
+};
+} // namespace validation
+} // namespace test
+} // namespace arm_compute
+#endif /* ARM_COMPUTE_TEST_RNN_LAYER_FIXTURE */