IVGCVSW-4452 Add QLstm EndToEndTest

* QLstm EndToEndTest added for Ref, NEON
  and CL.

Signed-off-by: James Conroy <james.conroy@arm.com>
Change-Id: Icd2c878541f6304d726202a93d71ff3d79f6f054

7 files changed, 303 insertions, 0 deletions

diff --git a/src/backends/backendsCommon/common.mk b/src/backends/backendsCommon/common.mk
index 09783ee7ea..450313f8d7 100644
--- a/src/backends/backendsCommon/common.mk
+++ b/src/backends/backendsCommon/common.mk
@@ -32,6 +32,7 @@ COMMON_TEST_SOURCES := \
     test/InstanceNormalizationEndToEndTestImpl.cpp \
     test/JsonPrinterTestImpl.cpp \
     test/LogSoftmaxEndToEndTestImpl.cpp \
+    test/QLstmEndToEndTestImpl.cpp \
     test/QuantizedLstmEndToEndTestImpl.cpp \
     test/SpaceToDepthEndToEndTestImpl.cpp \
     test/TensorCopyUtils.cpp \
diff --git a/src/backends/backendsCommon/test/CMakeLists.txt b/src/backends/backendsCommon/test/CMakeLists.txt
index 0241a60715..0679050c16 100644
--- a/src/backends/backendsCommon/test/CMakeLists.txt
+++ b/src/backends/backendsCommon/test/CMakeLists.txt
@@ -39,6 +39,8 @@ list(APPEND armnnBackendsCommonUnitTests_sources
     OptimizeSubgraphViewTests.cpp
     OptimizationViewsTests.cpp
     PreluEndToEndTestImpl.hpp
+    QLstmEndToEndTestImpl.cpp
+    QLstmEndToEndTestImpl.hpp
     QuantizedLstmEndToEndTestImpl.cpp
     QuantizedLstmEndToEndTestImpl.hpp
     ResizeEndToEndTestImpl.hpp
diff --git a/src/backends/backendsCommon/test/QLstmEndToEndTestImpl.cpp b/src/backends/backendsCommon/test/QLstmEndToEndTestImpl.cpp
new file mode 100644
index 0000000000..9949824c2e
--- /dev/null
+++ b/src/backends/backendsCommon/test/QLstmEndToEndTestImpl.cpp
@@ -0,0 +1,270 @@
+//
+// Copyright © 2020 Arm Ltd. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+
+#include "QLstmEndToEndTestImpl.hpp"
+
+#include "CommonTestUtils.hpp"
+#include "EndToEndTestImpl.hpp"
+
+#include <armnn/INetwork.hpp>
+#include <armnn/LstmParams.hpp>
+
+#include <boost/test/unit_test.hpp>
+
+namespace
+{
+
+// Checks if two values of an arithmetic type are close enough to each other
+// with regard to a given tolerance value.
+template<typename T>
+typename std::enable_if<std::is_arithmetic<T>::value, bool>::type
+IsCloseEnough(T value1, T value2, T tolerance)
+{
+    if (tolerance < 0)
+    {
+        throw armnn::InvalidArgumentException("Tolerance cannot be < 0");
+    }
+
+    T diff = value1 >= value2 ? static_cast<T>(value1 - value2) : static_cast<T>(value2 - value1);
+    return diff <= tolerance;
+}
+
+} // anonymous namespace
+
+void QLstmEndToEnd(const std::vector<armnn::BackendId>& backends)
+{
+    const unsigned int numBatches = 2;
+    const unsigned int inputSize  = 5;
+    const unsigned int outputSize = 4;
+    const unsigned int numUnits   = 4;
+
+    bool cifgEnabled       = true;
+    bool peepholeEnabled   = false;
+    bool projectionEnabled = false;
+    bool layerNormEnabled  = true;
+
+    // Scale/Offset quantization info
+    const float inputScale    = 0.0078125f;
+    const int32_t inputOffset = 0;
+
+    const int32_t hiddenStateZeroPoint = 0;
+    const float hiddenStateScale       = 0.007f;
+
+    // if (!projectionEnabled) outputScale == hiddenStateScale
+    const float outputScale    = hiddenStateScale;
+    const int32_t outputOffset = hiddenStateZeroPoint;
+
+    const float cellStateScale    = 3.05176e-05f;
+    const int32_t cellStateOffset = 0;
+
+    const float weightsScale    = 0.00784314f;
+    const int32_t weightsOffset = 0;
+
+    const float layerNormScale    = 3.05182e-05f;
+    const int32_t layerNormOffset = 0;
+
+    const float biasScale    = layerNormScale / 1024;
+    const int32_t biasOffset = 0;
+
+    const float inputIntermediateScale  = 0.007059f;
+    const float forgetIntermediateScale = 0.007812f;
+    const float cellIntermediateScale   = inputIntermediateScale;
+    const float outputIntermediateScale = forgetIntermediateScale;
+
+    const float cellClip       = 0.0f;
+    const float projectionClip = 0.0f;
+
+    // Weights and bias tensor info
+    const armnn::TensorInfo inputWeightsInfo({outputSize, inputSize},
+                                             armnn::DataType::QSymmS8,
+                                             weightsScale,
+                                             weightsOffset);
+
+    const armnn::TensorInfo recurrentWeightsInfo({outputSize, outputSize},
+                                                 armnn::DataType::QSymmS8,
+                                                 weightsScale,
+                                                 weightsOffset);
+
+    const armnn::TensorInfo biasInfo({outputSize},
+                                     armnn::DataType::Signed32,
+                                     biasScale,
+                                     biasOffset);
+
+    const armnn::TensorInfo layerNormWeightsInfo({numUnits},
+                                                 armnn::DataType::QSymmS16,
+                                                 layerNormScale,
+                                                 layerNormOffset);
+
+    // Mandatory params
+    const std::vector<int8_t> inputToForgetWeightsVector =
+            {-77, -13, 38, 25, 115, -64, -25, -51, 38, -102, -51, 38, -64, -51, -77, 38, -51, -77, -64, -64};
+    const std::vector<int8_t> inputToCellWeightsTensorVector =
+            {-51, -38, -25, -13, -64, 64, -25, -38, -25, -77, 77, -13, -51, -38, -89, 89, -115, -64, 102, 77};
+    const std::vector<int8_t> inputToOutputWeightsTensorVector =
+            {-102, -51, -25, -115, -13, -89, 38, -38, -102, -25, 77, -25, 51, -89, -38, -64, 13, 64, -77, -51};
+
+    armnn::ConstTensor inputToForgetWeightsTensor(inputWeightsInfo, inputToForgetWeightsVector.data());
+    armnn::ConstTensor inputToCellWeightsTensor(inputWeightsInfo, inputToCellWeightsTensorVector.data());
+    armnn::ConstTensor inputToOutputWeightsTensor(inputWeightsInfo, inputToOutputWeightsTensorVector.data());
+
+    const std::vector<int8_t> recurrentToForgetWeightsTensorVector =
+            {-64, -38, -64, -25, 77, 51, 115, 38, -13, 25, 64, 25, 25, 38, -13, 51};
+    const std::vector<int8_t> recurrentToCellWeightsTensorVector =
+            {-38, 25, 13, -38, 102, -10, -25, 38, 102, -77, -13, 25, 38, -13, 25, 64};
+    const std::vector<int8_t> recurrentToOutputWeightsTensorVector =
+            {38, -13, 13, -25, -64, -89, -25, -77, -13, -51, -89, -25, 13, 64, 25, -38};
+
+    armnn::ConstTensor recurrentToForgetWeightsTensor(recurrentWeightsInfo,
+                                                      recurrentToForgetWeightsTensorVector.data());
+    armnn::ConstTensor recurrentToCellWeightsTensor(recurrentWeightsInfo,
+                                                    recurrentToCellWeightsTensorVector.data());
+    armnn::ConstTensor recurrentToOutputWeightsTensor(recurrentWeightsInfo,
+                                                      recurrentToOutputWeightsTensorVector.data());
+
+    const std::vector<int32_t> forgetGateBiasTensorVector = {2147484, -6442451, -4294968, 2147484};
+    const std::vector<int32_t> cellBiasTensorVector       = {-1073742, 15461883, 5368709, 1717987};
+    const std::vector<int32_t> outputGateBiasTensorVector = {1073742, -214748, 4294968, 2147484};
+
+    armnn::ConstTensor forgetGateBiasTensor(biasInfo, forgetGateBiasTensorVector.data());
+    armnn::ConstTensor cellBiasTensor(biasInfo, cellBiasTensorVector.data());
+    armnn::ConstTensor outputGateBiasTensor(biasInfo, outputGateBiasTensorVector.data());
+
+    // Layer Norm
+    const std::vector<int16_t> forgetLayerNormWeightsVector = {6553, 6553, 13107, 9830};
+    const std::vector<int16_t> cellLayerNormWeightsVector   = {22937, 6553, 9830, 26214};
+    const std::vector<int16_t> outputLayerNormWeightsVector = {19660, 6553, 6553, 16384};
+
+    armnn::ConstTensor forgetLayerNormWeights(layerNormWeightsInfo, forgetLayerNormWeightsVector.data());
+    armnn::ConstTensor cellLayerNormWeights(layerNormWeightsInfo, cellLayerNormWeightsVector.data());
+    armnn::ConstTensor outputLayerNormWeights(layerNormWeightsInfo, outputLayerNormWeightsVector.data());
+
+    // Set up params
+    armnn::LstmInputParams params;
+    params.m_InputToForgetWeights = &inputToForgetWeightsTensor;
+    params.m_InputToCellWeights   = &inputToCellWeightsTensor;
+    params.m_InputToOutputWeights = &inputToOutputWeightsTensor;
+
+    params.m_RecurrentToForgetWeights = &recurrentToForgetWeightsTensor;
+    params.m_RecurrentToCellWeights   = &recurrentToCellWeightsTensor;
+    params.m_RecurrentToOutputWeights = &recurrentToOutputWeightsTensor;
+
+    params.m_ForgetGateBias = &forgetGateBiasTensor;
+    params.m_CellBias       = &cellBiasTensor;
+    params.m_OutputGateBias = &outputGateBiasTensor;
+
+    params.m_ForgetLayerNormWeights = &forgetLayerNormWeights;
+    params.m_CellLayerNormWeights   = &cellLayerNormWeights;
+    params.m_OutputLayerNormWeights = &outputLayerNormWeights;
+
+    QLstmDescriptor descriptor;
+    descriptor.m_CifgEnabled       = cifgEnabled;
+    descriptor.m_PeepholeEnabled   = peepholeEnabled;
+    descriptor.m_ProjectionEnabled = projectionEnabled;
+    descriptor.m_LayerNormEnabled  = layerNormEnabled;
+
+    descriptor.m_CellClip       = cellClip;
+    descriptor.m_ProjectionClip = projectionClip;
+
+    descriptor.m_HiddenStateZeroPoint = hiddenStateZeroPoint;
+    descriptor.m_HiddenStateScale     = hiddenStateScale;
+
+    descriptor.m_InputIntermediateScale  = inputIntermediateScale;
+    descriptor.m_ForgetIntermediateScale = forgetIntermediateScale;
+    descriptor.m_CellIntermediateScale   = cellIntermediateScale;
+    descriptor.m_OutputIntermediateScale = outputIntermediateScale;
+
+    // Input/Output tensor info
+    const armnn::TensorInfo inputInfo({numBatches , inputSize},
+                                      armnn::DataType::QAsymmS8,
+                                      inputScale,
+                                      inputOffset);
+
+    const armnn::TensorInfo cellStateInfo({numBatches , numUnits},
+                                          armnn::DataType::QSymmS16,
+                                          cellStateScale,
+                                          cellStateOffset);
+
+    const armnn::TensorInfo outputStateInfo({numBatches , outputSize},
+                                            armnn::DataType::QAsymmS8,
+                                            outputScale,
+                                            outputOffset);
+
+    // Input tensor data
+    const std::vector<int8_t> inputVector         = {90, 102, 13, 26, 38, 102, 13, 26, 51, 64};
+    const std::vector<int8_t> outputStateInVector = {0, 0, 0, 0, 0, 0, 0, 0};
+    const std::vector<int16_t> cellStateInVector  = {0, 0, 0, 0, 0, 0, 0, 0};
+
+    // Expected output tensor data
+    const std::vector<int8_t> outputStateOutVector = {-15, 21, 14, 20, -15, 15, 5, 27};
+    const std::vector<int16_t> cellStateOutVector  = {-11692, 9960, 5491, 8861, -9422, 7726, 2056, 13149};
+    const std::vector<int8_t> outputVector         = {-15, 21, 14, 20, -15, 15, 5, 27};
+
+    // Build network
+    armnn::INetworkPtr net(armnn::INetwork::Create());
+
+    armnn::IConnectableLayer* const input         = net->AddInputLayer(0);
+    armnn::IConnectableLayer* const outputStateIn = net->AddInputLayer(1);
+    armnn::IConnectableLayer* const cellStateIn   = net->AddInputLayer(2);
+
+    armnn::IConnectableLayer* const qLstmLayer = net->AddQLstmLayer(descriptor, params, "qLstm");
+
+    armnn::IConnectableLayer* const outputStateOut = net->AddOutputLayer(0);
+    armnn::IConnectableLayer* const cellStateOut   = net->AddOutputLayer(1);
+    armnn::IConnectableLayer* const output         = net->AddOutputLayer(2);
+
+    // Connect input/output slots
+    Connect(input, qLstmLayer, inputInfo, 0, 0);
+    Connect(outputStateIn, qLstmLayer, outputStateInfo, 0, 1);
+    Connect(cellStateIn, qLstmLayer, cellStateInfo, 0, 2);
+
+    Connect(qLstmLayer, outputStateOut, outputStateInfo, 0, 0);
+    Connect(qLstmLayer, cellStateOut, cellStateInfo, 1, 0);
+    Connect(qLstmLayer, output, outputStateInfo, 2, 0);
+
+    // Create runtime
+    IRuntime::CreationOptions options;
+    IRuntimePtr runtime(IRuntime::Create(options));
+
+    // Optimize the network
+    IOptimizedNetworkPtr optNet = Optimize(*net, backends, runtime->GetDeviceSpec());
+
+    // Loads network into runtime
+    NetworkId netId;
+    runtime->LoadNetwork(netId, std::move(optNet));
+
+    // Push back input tensors
+    InputTensors inputTensors;
+    inputTensors.reserve(3);
+
+    inputTensors.push_back({0, ConstTensor(runtime->GetInputTensorInfo(netId, 0), inputVector.data())});
+    inputTensors.push_back({1, ConstTensor(runtime->GetInputTensorInfo(netId, 1), outputStateInVector.data())});
+    inputTensors.push_back({2, ConstTensor(runtime->GetInputTensorInfo(netId, 2), cellStateInVector.data())});
+
+    // Push back output tensors
+    OutputTensors outputTensors;
+    outputTensors.reserve(3);
+
+    std::vector<int8_t> outputStateOutResult(outputStateOutVector.size());
+    std::vector<int16_t> cellStateOutResult(cellStateOutVector.size());
+    std::vector<int8_t> outputResult(outputStateOutVector.size());
+
+    outputTensors.push_back({0, Tensor(runtime->GetOutputTensorInfo(netId, 0), outputStateOutResult.data())});
+    outputTensors.push_back({1, Tensor(runtime->GetOutputTensorInfo(netId, 1), cellStateOutResult.data())});
+    outputTensors.push_back({2, Tensor(runtime->GetOutputTensorInfo(netId, 2), outputResult.data())});
+
+    // Execute inference
+    runtime->EnqueueWorkload(netId, inputTensors, outputTensors);
+
+    constexpr int8_t toleranceInt8 = 1;
+    for (unsigned int i = 0u; i < outputStateOutResult.size(); ++i)
+    {
+        BOOST_TEST(IsCloseEnough(outputStateOutVector[i], outputStateOutResult[i], toleranceInt8));
+    }
+
+    for (unsigned int i = 0u; i < outputResult.size(); ++i)
+    {
+        BOOST_TEST(IsCloseEnough(outputVector[i], outputResult[i], toleranceInt8));
+    }
+}
\ No newline at end of file
diff --git a/src/backends/backendsCommon/test/QLstmEndToEndTestImpl.hpp b/src/backends/backendsCommon/test/QLstmEndToEndTestImpl.hpp
new file mode 100644
index 0000000000..4c3daa04f9
--- /dev/null
+++ b/src/backends/backendsCommon/test/QLstmEndToEndTestImpl.hpp
@@ -0,0 +1,12 @@
+//
+// Copyright © 2020 Arm Ltd. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+
+#pragma once
+
+#include <armnn/BackendId.hpp>
+
+#include <vector>
+
+void QLstmEndToEnd(const std::vector<armnn::BackendId>& backends);
\ No newline at end of file
diff --git a/src/backends/cl/test/ClEndToEndTests.cpp b/src/backends/cl/test/ClEndToEndTests.cpp
index 5a9cf391a0..84aa9138ee 100644
--- a/src/backends/cl/test/ClEndToEndTests.cpp
+++ b/src/backends/cl/test/ClEndToEndTests.cpp
@@ -14,6 +14,7 @@
 #include <backendsCommon/test/ElementwiseUnaryEndToEndTestImpl.hpp>
 #include <backendsCommon/test/InstanceNormalizationEndToEndTestImpl.hpp>
 #include <backendsCommon/test/PreluEndToEndTestImpl.hpp>
+#include <backendsCommon/test/QLstmEndToEndTestImpl.hpp>
 #include <backendsCommon/test/QuantizedLstmEndToEndTestImpl.hpp>
 #include <backendsCommon/test/SpaceToDepthEndToEndTestImpl.hpp>
 #include <backendsCommon/test/SplitterEndToEndTestImpl.hpp>
@@ -490,4 +491,9 @@ BOOST_AUTO_TEST_CASE(ClArgMinAxis3TestQAsymmU8)
     ArgMinAxis3EndToEnd<armnn::DataType::QAsymmU8>(defaultBackends);
 }
 
+BOOST_AUTO_TEST_CASE(ClQLstmEndToEndTest)
+{
+    QLstmEndToEnd(defaultBackends);
+}
+
 BOOST_AUTO_TEST_SUITE_END()
diff --git a/src/backends/neon/test/NeonEndToEndTests.cpp b/src/backends/neon/test/NeonEndToEndTests.cpp
index 487aaeb0c3..9444ee0e96 100644
--- a/src/backends/neon/test/NeonEndToEndTests.cpp
+++ b/src/backends/neon/test/NeonEndToEndTests.cpp
@@ -15,6 +15,7 @@
 #include <backendsCommon/test/ElementwiseUnaryEndToEndTestImpl.hpp>
 #include <backendsCommon/test/InstanceNormalizationEndToEndTestImpl.hpp>
 #include <backendsCommon/test/PreluEndToEndTestImpl.hpp>
+#include <backendsCommon/test/QLstmEndToEndTestImpl.hpp>
 #include <backendsCommon/test/QuantizedLstmEndToEndTestImpl.hpp>
 #include <backendsCommon/test/SpaceToDepthEndToEndTestImpl.hpp>
 #include <backendsCommon/test/SplitterEndToEndTestImpl.hpp>
@@ -714,4 +715,9 @@ BOOST_AUTO_TEST_CASE(NeonDetectionPostProcessFastNmsUint8Test, * boost::unit_tes
                                                                           1.0f, 1, 0.01f, 0, 0.5f, 0);
 }
 
+BOOST_AUTO_TEST_CASE(NeonQLstmEndToEndTest)
+{
+    QLstmEndToEnd(defaultBackends);
+}
+
 BOOST_AUTO_TEST_SUITE_END()
diff --git a/src/backends/reference/test/RefEndToEndTests.cpp b/src/backends/reference/test/RefEndToEndTests.cpp
index 9c10a68018..45605e0fc6 100644
--- a/src/backends/reference/test/RefEndToEndTests.cpp
+++ b/src/backends/reference/test/RefEndToEndTests.cpp
@@ -18,6 +18,7 @@
 #include <backendsCommon/test/InstanceNormalizationEndToEndTestImpl.hpp>
 #include <backendsCommon/test/LogSoftmaxEndToEndTestImpl.hpp>
 #include <backendsCommon/test/PreluEndToEndTestImpl.hpp>
+#include <backendsCommon/test/QLstmEndToEndTestImpl.hpp>
 #include <backendsCommon/test/ResizeEndToEndTestImpl.hpp>
 #include <backendsCommon/test/SpaceToDepthEndToEndTestImpl.hpp>
 #include <backendsCommon/test/SplitterEndToEndTestImpl.hpp>
@@ -1235,6 +1236,11 @@ BOOST_AUTO_TEST_CASE(RefArgMinAxis3Uint8Test)
     ArgMinAxis3EndToEnd<armnn::DataType::QAsymmU8>(defaultBackends);
 }
 
+BOOST_AUTO_TEST_CASE(RefQLstmEndToEndTest)
+{
+    QLstmEndToEnd(defaultBackends);
+}
+
 #if !defined(__ANDROID__)
 // Only run these tests on non Android platforms
 BOOST_AUTO_TEST_CASE(RefImportNonAlignedPointerTest)