diff --git a/src/armnn/BackendHelper.cpp b/src/armnn/BackendHelper.cpp
index 056fbb0..e2aa672 100644
--- a/src/armnn/BackendHelper.cpp
+++ b/src/armnn/BackendHelper.cpp
@@ -1332,16 +1332,14 @@
 bool LayerSupportHandle::IsUnidirectionalSequenceLstmSupported(const TensorInfo& input,
                                                                const TensorInfo& outputStateIn,
                                                                const TensorInfo& cellStateIn,
+                                                               const TensorInfo& outputStateOut,
+                                                               const TensorInfo& cellStateOut,
                                                                const TensorInfo& output,
-                                                               const Optional<TensorInfo>& hiddenStateOutput,
-                                                               const Optional<TensorInfo>& cellStateOutput,
                                                                const LstmDescriptor& descriptor,
                                                                const LstmInputParamsInfo& paramsInfo,
                                                                Optional<std::string&> reasonIfUnsupported)
 {
-    TensorInfo hiddenStateOutputVal =  hiddenStateOutput.has_value() ? hiddenStateOutput.value() : TensorInfo();
-    TensorInfo cellStateOutputVal   =  cellStateOutput.has_value() ? cellStateOutput.value() : TensorInfo();
-    TensorInfos infos{input, outputStateIn, cellStateIn, hiddenStateOutputVal, cellStateOutputVal, output};
+    TensorInfos infos{input, outputStateIn, cellStateIn, outputStateOut, cellStateOut, output};
 
     return m_LayerSupport->IsLayerSupported(LayerType::UnidirectionalSequenceLstm,
                                             infos,
diff --git a/src/armnn/ILayerSupport.cpp b/src/armnn/ILayerSupport.cpp
index bf54223..5366b13 100644
--- a/src/armnn/ILayerSupport.cpp
+++ b/src/armnn/ILayerSupport.cpp
@@ -488,57 +488,15 @@
                                                "hiddenStateOutputVal, cellStateOutputVal, output}");
             }
             auto desc = *(PolymorphicDowncast<const UnidirectionalSequenceLstmDescriptor*>(&descriptor));
-
-            bool isHiddenStateOutputOptional = (infos[4] == TensorInfo());
-            bool isCellStateOutput = (infos[5] == TensorInfo());
-            if (isHiddenStateOutputOptional && isCellStateOutput)
-            {
-                return IsUnidirectionalSequenceLstmSupported(infos[0],
-                                                             infos[1],
-                                                             infos[2],
-                                                             infos[3],
-                                                             EmptyOptional(),
-                                                             EmptyOptional(),
-                                                             desc,
-                                                             lstmParamsInfo.value(),
-                                                             reasonIfUnsupported);
-            }
-            else if (isHiddenStateOutputOptional)
-            {
-                return IsUnidirectionalSequenceLstmSupported(infos[0],
-                                                             infos[1],
-                                                             infos[2],
-                                                             infos[3],
-                                                             EmptyOptional(),
-                                                             infos[5],
-                                                             desc,
-                                                             lstmParamsInfo.value(),
-                                                             reasonIfUnsupported);
-            }
-            else if (isCellStateOutput)
-            {
-                return IsUnidirectionalSequenceLstmSupported(infos[0],
-                                                             infos[1],
-                                                             infos[2],
-                                                             infos[3],
-                                                             infos[4],
-                                                             EmptyOptional(),
-                                                             desc,
-                                                             lstmParamsInfo.value(),
-                                                             reasonIfUnsupported);
-            }
-            else
-            {
-                return IsUnidirectionalSequenceLstmSupported(infos[0],
-                                                             infos[1],
-                                                             infos[2],
-                                                             infos[3],
-                                                             infos[4],
-                                                             infos[5],
-                                                             desc,
-                                                             lstmParamsInfo.value(),
-                                                             reasonIfUnsupported);
-            }
+            return IsUnidirectionalSequenceLstmSupported(infos[0],
+                                                         infos[1],
+                                                         infos[2],
+                                                         infos[3],
+                                                         infos[4],
+                                                         infos[5],
+                                                         desc,
+                                                         lstmParamsInfo.value(),
+                                                         reasonIfUnsupported);
         }
         case LayerType::ChannelShuffle:
             return IsChannelShuffleSupported(infos[0],
@@ -1285,9 +1243,9 @@
         const TensorInfo& input,
         const TensorInfo& outputStateIn,
         const TensorInfo& cellStateIn,
+        const TensorInfo& outputStateOut,
+        const TensorInfo& cellStateOut,
         const TensorInfo& output,
-        const Optional<TensorInfo>& hiddenStateOutput,
-        const Optional<TensorInfo>& cellStateOutput,
         const LstmDescriptor& descriptor,
         const LstmInputParamsInfo& paramsInfo,
         Optional<std::string&> reasonIfUnsupported) const
@@ -1295,9 +1253,9 @@
     IgnoreUnused(input,
                  outputStateIn,
                  cellStateIn,
+                 outputStateOut,
+                 cellStateOut,
                  output,
-                 hiddenStateOutput,
-                 cellStateOutput,
                  descriptor,
                  paramsInfo,
                  reasonIfUnsupported);
diff --git a/src/armnn/layers/UnidirectionalSequenceLstmLayer.cpp b/src/armnn/layers/UnidirectionalSequenceLstmLayer.cpp
index 1999614..e5f89bd 100644
--- a/src/armnn/layers/UnidirectionalSequenceLstmLayer.cpp
+++ b/src/armnn/layers/UnidirectionalSequenceLstmLayer.cpp
@@ -15,7 +15,7 @@
 {
 
 UnidirectionalSequenceLstmLayer::UnidirectionalSequenceLstmLayer(const LstmDescriptor& param, const char* name)
-        : LayerWithParameters(3, 1, LayerType::UnidirectionalSequenceLstm, param, name)
+        : LayerWithParameters(3, 3, LayerType::UnidirectionalSequenceLstm, param, name)
 {
 }
 
@@ -171,7 +171,7 @@
 {
     VerifyLayerConnections(3, CHECK_LOCATION());
 
-    const TensorShape& outputShape = GetOutputSlot(0).GetTensorInfo().GetShape();
+    const TensorShape& outputShape = GetOutputSlot(2).GetTensorInfo().GetShape();
 
     VerifyShapeInferenceType(outputShape, m_ShapeInferenceMethod);
 
diff --git a/src/armnnDeserializer/Deserializer.cpp b/src/armnnDeserializer/Deserializer.cpp
index 75c60cc..93fa99d 100644
--- a/src/armnnDeserializer/Deserializer.cpp
+++ b/src/armnnDeserializer/Deserializer.cpp
@@ -3616,7 +3616,7 @@
     CHECK_VALID_SIZE(inputs.size(), 3);
 
     auto outputs = GetOutputs(graph, layerIndex);
-    CHECK_VALID_SIZE(outputs.size(), 1);
+    CHECK_VALID_SIZE(outputs.size(), 3);
 
     auto flatBufferLayer = graph->layers()->Get(layerIndex)->layer_as_UnidirectionalSequenceLstmLayer();
     auto layerName = GetLayerName(graph, layerIndex);
@@ -3714,8 +3714,14 @@
                                                                              lstmInputParams,
                                                                              layerName.c_str());
 
-    armnn::TensorInfo outputTensorInfo1 = ToTensorInfo(outputs[0]);
-    layer->GetOutputSlot(0).SetTensorInfo(outputTensorInfo1);
+    armnn::TensorInfo outputTensorInfo0 = ToTensorInfo(outputs[0]);
+    layer->GetOutputSlot(0).SetTensorInfo(outputTensorInfo0);
+
+    armnn::TensorInfo outputTensorInfo1 = ToTensorInfo(outputs[1]);
+    layer->GetOutputSlot(1).SetTensorInfo(outputTensorInfo1);
+
+    armnn::TensorInfo outputTensorInfo2 = ToTensorInfo(outputs[2]);
+    layer->GetOutputSlot(2).SetTensorInfo(outputTensorInfo2);
 
     RegisterInputSlots(graph, layerIndex, layer);
     RegisterOutputSlots(graph, layerIndex, layer);
diff --git a/src/armnnSerializer/test/LstmSerializationTests.cpp b/src/armnnSerializer/test/LstmSerializationTests.cpp
index d8f8967..ae2d813 100644
--- a/src/armnnSerializer/test/LstmSerializationTests.cpp
+++ b/src/armnnSerializer/test/LstmSerializationTests.cpp
@@ -2299,6 +2299,8 @@
     armnn::TensorInfo inputTensorInfo({ batchSize, timeSize, inputSize }, armnn::DataType::Float32);
     armnn::TensorInfo cellStateTensorInfo({ batchSize, numUnits}, armnn::DataType::Float32);
     armnn::TensorInfo outputStateTensorInfo({ batchSize, outputSize }, armnn::DataType::Float32);
+    armnn::TensorInfo outputStateOutTensorInfo({ batchSize, timeSize, outputSize }, armnn::DataType::Float32);
+    armnn::TensorInfo cellStateOutTensorInfo({ batchSize, outputSize }, armnn::DataType::Float32);
     armnn::TensorInfo outputTensorInfo({ batchSize, timeSize, outputSize }, armnn::DataType::Float32);
 
     inputLayer->GetOutputSlot(0).Connect(unidirectionalSequenceLstmLayer->GetInputSlot(0));
@@ -2310,8 +2312,10 @@
     cellStateIn->GetOutputSlot(0).Connect(unidirectionalSequenceLstmLayer->GetInputSlot(2));
     cellStateIn->GetOutputSlot(0).SetTensorInfo(cellStateTensorInfo);
 
-    unidirectionalSequenceLstmLayer->GetOutputSlot(0).Connect(outputLayer->GetInputSlot(0));
-    unidirectionalSequenceLstmLayer->GetOutputSlot(0).SetTensorInfo(outputTensorInfo);
+    unidirectionalSequenceLstmLayer->GetOutputSlot(0).SetTensorInfo(outputStateOutTensorInfo);
+    unidirectionalSequenceLstmLayer->GetOutputSlot(1).SetTensorInfo(cellStateOutTensorInfo);
+    unidirectionalSequenceLstmLayer->GetOutputSlot(2).Connect(outputLayer->GetInputSlot(0));
+    unidirectionalSequenceLstmLayer->GetOutputSlot(2).SetTensorInfo(outputTensorInfo);
 
     armnn::INetworkPtr deserializedNetwork = DeserializeNetwork(SerializeNetwork(*network));
     CHECK(deserializedNetwork);
@@ -2319,7 +2323,7 @@
     VerifyLstmLayer<armnn::UnidirectionalSequenceLstmDescriptor> checker(
         layerName,
         {inputTensorInfo, outputStateTensorInfo, cellStateTensorInfo},
-        {outputTensorInfo},
+        {outputStateOutTensorInfo, cellStateOutTensorInfo, outputTensorInfo},
         descriptor,
         params);
     deserializedNetwork->ExecuteStrategy(checker);
@@ -2436,6 +2440,8 @@
     armnn::TensorInfo inputTensorInfo({ batchSize, timeSize, inputSize }, armnn::DataType::Float32);
     armnn::TensorInfo cellStateTensorInfo({ batchSize, numUnits}, armnn::DataType::Float32);
     armnn::TensorInfo outputStateTensorInfo({ batchSize, outputSize }, armnn::DataType::Float32);
+    armnn::TensorInfo outputStateOutTensorInfo({ batchSize, timeSize, outputSize }, armnn::DataType::Float32);
+    armnn::TensorInfo cellStateOutTensorInfo({ batchSize, outputSize }, armnn::DataType::Float32);
     armnn::TensorInfo outputTensorInfo({ batchSize, timeSize, outputSize }, armnn::DataType::Float32);
 
     inputLayer->GetOutputSlot(0).Connect(unidirectionalSequenceLstmLayer->GetInputSlot(0));
@@ -2447,8 +2453,10 @@
     cellStateIn->GetOutputSlot(0).Connect(unidirectionalSequenceLstmLayer->GetInputSlot(2));
     cellStateIn->GetOutputSlot(0).SetTensorInfo(cellStateTensorInfo);
 
-    unidirectionalSequenceLstmLayer->GetOutputSlot(0).Connect(outputLayer->GetInputSlot(0));
-    unidirectionalSequenceLstmLayer->GetOutputSlot(0).SetTensorInfo(outputTensorInfo);
+    unidirectionalSequenceLstmLayer->GetOutputSlot(0).SetTensorInfo(outputStateOutTensorInfo);
+    unidirectionalSequenceLstmLayer->GetOutputSlot(1).SetTensorInfo(cellStateOutTensorInfo);
+    unidirectionalSequenceLstmLayer->GetOutputSlot(2).Connect(outputLayer->GetInputSlot(0));
+    unidirectionalSequenceLstmLayer->GetOutputSlot(2).SetTensorInfo(outputTensorInfo);
 
     armnn::INetworkPtr deserializedNetwork = DeserializeNetwork(SerializeNetwork(*network));
     CHECK(deserializedNetwork);
@@ -2456,7 +2464,7 @@
     VerifyLstmLayer<armnn::UnidirectionalSequenceLstmDescriptor> checker(
         layerName,
         {inputTensorInfo, outputStateTensorInfo, cellStateTensorInfo},
-        {outputTensorInfo},
+        {outputStateOutTensorInfo, cellStateOutTensorInfo, outputTensorInfo},
         descriptor,
         params);
     deserializedNetwork->ExecuteStrategy(checker);
@@ -2592,6 +2600,8 @@
     armnn::TensorInfo inputTensorInfo({ batchSize, timeSize, inputSize }, armnn::DataType::Float32);
     armnn::TensorInfo cellStateTensorInfo({ batchSize, numUnits}, armnn::DataType::Float32);
     armnn::TensorInfo outputStateTensorInfo({ batchSize, outputSize }, armnn::DataType::Float32);
+    armnn::TensorInfo outputStateOutTensorInfo({batchSize, timeSize, outputSize}, armnn::DataType::Float32);
+    armnn::TensorInfo cellStateOutTensorInfo({batchSize, timeSize, outputSize}, armnn::DataType::Float32);
     armnn::TensorInfo outputTensorInfo({ batchSize, timeSize, outputSize }, armnn::DataType::Float32);
 
     inputLayer->GetOutputSlot(0).Connect(unidirectionalSequenceLstmLayer->GetInputSlot(0));
@@ -2603,8 +2613,10 @@
     cellStateIn->GetOutputSlot(0).Connect(unidirectionalSequenceLstmLayer->GetInputSlot(2));
     cellStateIn->GetOutputSlot(0).SetTensorInfo(cellStateTensorInfo);
 
-    unidirectionalSequenceLstmLayer->GetOutputSlot(0).Connect(outputLayer->GetInputSlot(0));
-    unidirectionalSequenceLstmLayer->GetOutputSlot(0).SetTensorInfo(outputTensorInfo);
+    unidirectionalSequenceLstmLayer->GetOutputSlot(0).SetTensorInfo(outputStateOutTensorInfo);
+    unidirectionalSequenceLstmLayer->GetOutputSlot(1).SetTensorInfo(cellStateOutTensorInfo);
+    unidirectionalSequenceLstmLayer->GetOutputSlot(2).Connect(outputLayer->GetInputSlot(0));
+    unidirectionalSequenceLstmLayer->GetOutputSlot(2).SetTensorInfo(outputTensorInfo);
 
     armnn::INetworkPtr deserializedNetwork = DeserializeNetwork(SerializeNetwork(*network));
     CHECK(deserializedNetwork);
@@ -2612,7 +2624,7 @@
     VerifyLstmLayer<armnn::UnidirectionalSequenceLstmDescriptor> checker(
             layerName,
             {inputTensorInfo, outputStateTensorInfo, cellStateTensorInfo},
-            {outputTensorInfo},
+            {outputStateOutTensorInfo, cellStateOutTensorInfo, outputTensorInfo},
             descriptor,
             params);
     deserializedNetwork->ExecuteStrategy(checker);
@@ -2697,6 +2709,8 @@
     armnn::TensorInfo inputTensorInfo({ timeSize, batchSize, inputSize }, armnn::DataType::Float32);
     armnn::TensorInfo cellStateTensorInfo({ batchSize, numUnits}, armnn::DataType::Float32);
     armnn::TensorInfo outputStateTensorInfo({ batchSize, outputSize }, armnn::DataType::Float32);
+    armnn::TensorInfo outputStateOutTensorInfo({batchSize, timeSize, outputSize}, armnn::DataType::Float32);
+    armnn::TensorInfo cellStateOutTensorInfo({batchSize, timeSize, outputSize}, armnn::DataType::Float32);
     armnn::TensorInfo outputTensorInfo({ timeSize, batchSize, outputSize }, armnn::DataType::Float32);
 
     inputLayer->GetOutputSlot(0).Connect(unidirectionalSequenceLstmLayer->GetInputSlot(0));
@@ -2708,8 +2722,10 @@
     cellStateIn->GetOutputSlot(0).Connect(unidirectionalSequenceLstmLayer->GetInputSlot(2));
     cellStateIn->GetOutputSlot(0).SetTensorInfo(cellStateTensorInfo);
 
-    unidirectionalSequenceLstmLayer->GetOutputSlot(0).Connect(outputLayer->GetInputSlot(0));
-    unidirectionalSequenceLstmLayer->GetOutputSlot(0).SetTensorInfo(outputTensorInfo);
+    unidirectionalSequenceLstmLayer->GetOutputSlot(0).SetTensorInfo(outputStateOutTensorInfo);
+    unidirectionalSequenceLstmLayer->GetOutputSlot(1).SetTensorInfo(cellStateOutTensorInfo);
+    unidirectionalSequenceLstmLayer->GetOutputSlot(2).Connect(outputLayer->GetInputSlot(0));
+    unidirectionalSequenceLstmLayer->GetOutputSlot(2).SetTensorInfo(outputTensorInfo);
 
     armnn::INetworkPtr deserializedNetwork = DeserializeNetwork(SerializeNetwork(*network));
     CHECK(deserializedNetwork);
@@ -2717,7 +2733,7 @@
     VerifyLstmLayer<armnn::UnidirectionalSequenceLstmDescriptor> checker(
         layerName,
         {inputTensorInfo, outputStateTensorInfo, cellStateTensorInfo},
-        {outputTensorInfo},
+        {outputStateOutTensorInfo, cellStateOutTensorInfo, outputTensorInfo},
         descriptor,
         params);
     deserializedNetwork->ExecuteStrategy(checker);
diff --git a/src/backends/backendsCommon/LayerSupportBase.cpp b/src/backends/backendsCommon/LayerSupportBase.cpp
index 89a0772..0010379 100644
--- a/src/backends/backendsCommon/LayerSupportBase.cpp
+++ b/src/backends/backendsCommon/LayerSupportBase.cpp
@@ -680,9 +680,9 @@
 bool LayerSupportBase::IsUnidirectionalSequenceLstmSupported(const TensorInfo&, // input
                                                              const TensorInfo&, // outputStateIn
                                                              const TensorInfo&, // cellStateIn
+                                                             const TensorInfo&, // outputStateOut
+                                                             const TensorInfo&, // cellStateOut
                                                              const TensorInfo&, // output
-                                                             const Optional<TensorInfo>&, // hiddenStateOut
-                                                             const Optional<TensorInfo>&, // cellStateOut
                                                              const LstmDescriptor&, // descriptor
                                                              const LstmInputParamsInfo&, // paramsInfo
                                                              Optional<std::string&> reasonIfUnsupported) const
diff --git a/src/backends/backendsCommon/LayerSupportBase.hpp b/src/backends/backendsCommon/LayerSupportBase.hpp
index e911c00..618d21e 100644
--- a/src/backends/backendsCommon/LayerSupportBase.hpp
+++ b/src/backends/backendsCommon/LayerSupportBase.hpp
@@ -465,9 +465,9 @@
         const TensorInfo& input,
         const TensorInfo& outputStateIn,
         const TensorInfo& cellStateIn,
+        const TensorInfo& outputStateOut,
+        const TensorInfo& cellStateOut,
         const TensorInfo& output,
-        const Optional<TensorInfo>& hiddenStateOutput,
-        const Optional<TensorInfo>& cellStateOutput,
         const LstmDescriptor& descriptor,
         const LstmInputParamsInfo& paramsInfo,
         Optional<std::string&> reasonIfUnsupported = EmptyOptional()) const override;
diff --git a/src/backends/backendsCommon/WorkloadData.cpp b/src/backends/backendsCommon/WorkloadData.cpp
index 70d7641..a2dcd63 100644
--- a/src/backends/backendsCommon/WorkloadData.cpp
+++ b/src/backends/backendsCommon/WorkloadData.cpp
@@ -3860,38 +3860,20 @@
     {
         throw InvalidArgumentException(descriptorName + ": Invalid number of inputs.");
     }
-    if (workloadInfo.m_OutputTensorInfos.size() != 1)
+    if (workloadInfo.m_OutputTensorInfos.size() != 3)
     {
         throw InvalidArgumentException(descriptorName + ": Invalid number of outputs.");
     }
 
     std::vector<DataType> supportedTypes =
     {
-        DataType::Float32
+        DataType::Float32,
+        DataType::QAsymmS8
     };
 
     // check for supported type of one input and match them with all the other input and output
     ValidateDataTypes(workloadInfo.m_InputTensorInfos[0], supportedTypes, descriptorName);
 
-    // type matches all other inputs
-    for (uint32_t i = 1u; i < workloadInfo.m_InputTensorInfos.size(); ++i)
-    {
-        ValidateTensorDataTypesMatch(workloadInfo.m_InputTensorInfos[0],
-                                     workloadInfo.m_InputTensorInfos[i],
-                                     descriptorName,
-                                     "input_0",
-                                     "input_" + std::to_string(i));
-    }
-    // type matches all other outputs
-    for (uint32_t i = 0u; i < workloadInfo.m_OutputTensorInfos.size(); ++i)
-    {
-        ValidateTensorDataTypesMatch(workloadInfo.m_InputTensorInfos[0],
-                                     workloadInfo.m_OutputTensorInfos[i],
-                                     "LstmQueueDescriptor",
-                                     "input_0",
-                                     "output_" + std::to_string(i));
-    }
-
     // Making sure clipping parameters have valid values.
     // == 0 means no clipping
     //  > 0 means clipping
@@ -3936,7 +3918,7 @@
                                 descriptorName + " input_2");
 
     // outputTensor
-    ValidateTensorNumDimNumElem(workloadInfo.m_OutputTensorInfos[0], 3, (timeStep * n_batch * n_output),
+    ValidateTensorNumDimNumElem(workloadInfo.m_OutputTensorInfos[2], 3, (timeStep * n_batch * n_output),
                                 descriptorName + " output_0");
 
     // check that dimensions of inputs/outputs and QueueDescriptor data match with each other
diff --git a/src/backends/backendsCommon/WorkloadFactory.cpp b/src/backends/backendsCommon/WorkloadFactory.cpp
index f955aec..5847e8c 100644
--- a/src/backends/backendsCommon/WorkloadFactory.cpp
+++ b/src/backends/backendsCommon/WorkloadFactory.cpp
@@ -1367,7 +1367,9 @@
             const TensorInfo& cellStateIn = OverrideDataType(layer.GetInputSlot(2).GetConnection()->GetTensorInfo(),
                                                              dataType);
             // Outputs
-            const TensorInfo&  output = OverrideDataType(layer.GetOutputSlot(0).GetTensorInfo(), dataType);
+            const TensorInfo& outputStateOut = OverrideDataType(layer.GetOutputSlot(0).GetTensorInfo(), dataType);
+            const TensorInfo& cellStateOut = OverrideDataType(layer.GetOutputSlot(1).GetTensorInfo(), dataType);
+            const TensorInfo& output = OverrideDataType(layer.GetOutputSlot(2).GetTensorInfo(), dataType);
 
             // Basic parameters
             const TensorInfo& inputToForgetWeights
@@ -1481,15 +1483,12 @@
                 paramsInfo.m_OutputLayerNormWeights = &optOutputLayerNormWeights;
             }
 
-            Optional<TensorInfo> hiddenStateOut;
-            Optional<TensorInfo> cellStateOut;
-
             result = layerSupportObject.IsUnidirectionalSequenceLstmSupported(input,
                                                                               outputStateIn,
                                                                               cellStateIn,
-                                                                              output,
-                                                                              hiddenStateOut,
+                                                                              outputStateOut,
                                                                               cellStateOut,
+                                                                              output,
                                                                               descriptor,
                                                                               paramsInfo,
                                                                               reason);
diff --git a/src/backends/backendsCommon/test/layerTests/UnidirectionalSequenceLstmTestImpl.cpp b/src/backends/backendsCommon/test/layerTests/UnidirectionalSequenceLstmTestImpl.cpp
index c719472..6effa9c 100644
--- a/src/backends/backendsCommon/test/layerTests/UnidirectionalSequenceLstmTestImpl.cpp
+++ b/src/backends/backendsCommon/test/layerTests/UnidirectionalSequenceLstmTestImpl.cpp
@@ -31,7 +31,7 @@
     armnn::DataType constantDataType = armnn::DataType::Float32)
 {
     IgnoreUnused(memoryManager);
-    unsigned int batchSize = armnn::numeric_cast<unsigned int>(inputShape[1]);
+    unsigned int batchSize = armnn::numeric_cast<unsigned int>(inputShape[0]);
     unsigned int inputSize = armnn::numeric_cast<unsigned int>(inputShape[2]);
     unsigned int outputSize = armnn::numeric_cast<unsigned int>(outputExpectedShape[2]);
     unsigned numUnits = outputSize;
@@ -39,7 +39,8 @@
     armnn::TensorInfo inputTensorInfo({1, batchSize , inputSize}, ArmnnType,  qScale, qOffset );
     armnn::TensorInfo cellStateInTensorInfo({batchSize , numUnits}, ArmnnType, qScale, qOffset);
     armnn::TensorInfo outputStateInTensorInfo({batchSize , outputSize}, ArmnnType, qScale, qOffset);
-
+    armnn::TensorInfo outputStateOutTensorInfo({ batchSize, 1, outputSize }, ArmnnType, qScale, qOffset);
+    armnn::TensorInfo cellStateOutTensorInfo({ batchSize, 1, outputSize }, ArmnnType, qScale, qOffset);
     armnn::TensorInfo outputTensorInfo({1, batchSize, outputSize}, ArmnnType, qScale, qOffset);
 
     std::vector<T> inputVector;
@@ -48,6 +49,8 @@
     std::vector<T> cellStateInVector(batchSize * numUnits, T());
     std::vector<T> outputStateInVector(batchSize * outputSize, T());
 
+    std::vector<T> actualOutputStateOut(outputStateOutTensorInfo.GetNumElements());
+    std::vector<T> actualCellStateOut(cellStateOutTensorInfo.GetNumElements());
     std::vector<T> actualOutput(outputTensorInfo.GetNumElements());
 
     std::vector<T> outputVector;
@@ -59,6 +62,10 @@
     std::unique_ptr<armnn::ITensorHandle> outputStateInHandle =
                                               tensorHandleFactory.CreateTensorHandle(outputStateInTensorInfo);
 
+    std::unique_ptr<armnn::ITensorHandle> outputStateOutHandle =
+            tensorHandleFactory.CreateTensorHandle(outputStateOutTensorInfo);
+    std::unique_ptr<armnn::ITensorHandle> cellStateOutHandle =
+            tensorHandleFactory.CreateTensorHandle(cellStateOutTensorInfo);
     std::unique_ptr<armnn::ITensorHandle> outputHandle = tensorHandleFactory.CreateTensorHandle(outputTensorInfo);
 
     armnn::UnidirectionalSequenceLstmQueueDescriptor data;
@@ -68,6 +75,8 @@
     AddInputToWorkload(data, info, outputStateInTensorInfo, outputStateInHandle.get());
     AddInputToWorkload(data, info, cellStateInTensorInfo, cellStateInHandle.get());
 
+    AddOutputToWorkload(data, info, outputStateOutTensorInfo, outputStateOutHandle.get());
+    AddOutputToWorkload(data, info, cellStateOutTensorInfo, cellStateOutHandle.get());
     AddOutputToWorkload(data, info, outputTensorInfo, outputHandle.get());
 
     armnn::TensorInfo tensorInfo4({numUnits}, constantDataType , qScale, qOffset);
@@ -184,6 +193,8 @@
     outputStateInHandle->Allocate();
     cellStateInHandle->Allocate();
 
+    outputStateOutHandle->Allocate();
+    cellStateOutHandle->Allocate();
     outputHandle->Allocate();
 
     CopyDataToITensorHandle(inputHandle.get(), inputVector.data());
@@ -192,6 +203,8 @@
 
     workload->Execute();
 
+    CopyDataFromITensorHandle(actualOutputStateOut.data(), outputStateOutHandle.get());
+    CopyDataFromITensorHandle(actualCellStateOut.data(), cellStateOutHandle.get());
     CopyDataFromITensorHandle(actualOutput.data(), outputHandle.get());
 
     return LayerTestResult<T, 3>(actualOutput,
@@ -222,7 +235,8 @@
     armnn::TensorInfo inputTensorInfo({batchSize, timeSize, inputSize}, ArmnnType, qScale, qOffset);
     armnn::TensorInfo cellStateInTensorInfo({batchSize, numUnits}, ArmnnType, qScale, qOffset);
     armnn::TensorInfo outputStateInTensorInfo({batchSize, outputSize}, ArmnnType, qScale, qOffset);
-
+    armnn::TensorInfo outputStateOutTensorInfo({batchSize, timeSize, outputSize}, ArmnnType, qScale, qOffset);
+    armnn::TensorInfo cellStateOutTensorInfo({batchSize, timeSize, outputSize}, ArmnnType, qScale, qOffset);
     armnn::TensorInfo outputTensorInfo({batchSize, timeSize, outputSize}, ArmnnType, qScale, qOffset);
 
     std::vector<T> inputVector;
@@ -231,6 +245,8 @@
     std::vector<T> cellStateInVector(batchSize * numUnits, T());
     std::vector<T> outputStateInVector(batchSize * outputSize, T());
 
+    std::vector<T> actualOutputStateOut(outputStateOutTensorInfo.GetNumElements());
+    std::vector<T> actualCellStateOut(cellStateOutTensorInfo.GetNumElements());
     std::vector<T> actualOutput(outputTensorInfo.GetNumElements());
 
     std::vector<T> outputVector;
@@ -242,6 +258,10 @@
     std::unique_ptr<armnn::ITensorHandle> outputStateInHandle =
         tensorHandleFactory.CreateTensorHandle(outputStateInTensorInfo);
 
+    std::unique_ptr<armnn::ITensorHandle> outputStateOutHandle =
+            tensorHandleFactory.CreateTensorHandle(outputStateOutTensorInfo);
+    std::unique_ptr<armnn::ITensorHandle> cellStateOutHandle =
+            tensorHandleFactory.CreateTensorHandle(cellStateOutTensorInfo);
     std::unique_ptr<armnn::ITensorHandle> outputHandle = tensorHandleFactory.CreateTensorHandle(outputTensorInfo);
 
     armnn::UnidirectionalSequenceLstmQueueDescriptor data;
@@ -251,6 +271,8 @@
     AddInputToWorkload(data, info, outputStateInTensorInfo, outputStateInHandle.get());
     AddInputToWorkload(data, info, cellStateInTensorInfo, cellStateInHandle.get());
 
+    AddOutputToWorkload(data, info, outputStateOutTensorInfo, outputStateOutHandle.get());
+    AddOutputToWorkload(data, info, cellStateOutTensorInfo, cellStateOutHandle.get());
     AddOutputToWorkload(data, info, outputTensorInfo, outputHandle.get());
 
     armnn::TensorInfo tensorInfo4({numUnits}, constantDataType, qScale, qOffset);
@@ -359,6 +381,8 @@
     outputStateInHandle->Allocate();
     cellStateInHandle->Allocate();
 
+    outputStateOutHandle->Allocate();
+    cellStateOutHandle->Allocate();
     outputHandle->Allocate();
 
     CopyDataToITensorHandle(inputHandle.get(), inputVector.data());
@@ -367,6 +391,8 @@
 
     workload->Execute();
 
+    CopyDataFromITensorHandle(actualOutputStateOut.data(), outputStateOutHandle.get());
+    CopyDataFromITensorHandle(actualCellStateOut.data(), cellStateOutHandle.get());
     CopyDataFromITensorHandle(actualOutput.data(), outputHandle.get());
 
     return LayerTestResult<T, 3>(actualOutput,
@@ -398,7 +424,8 @@
     armnn::TensorInfo inputTensorInfo({timeSize, batchSize, inputSize}, ArmnnType, qScale, qOffset);
     armnn::TensorInfo cellStateInTensorInfo({batchSize, numUnits}, ArmnnType, qScale, qOffset);
     armnn::TensorInfo outputStateInTensorInfo({batchSize, outputSize}, ArmnnType, qScale, qOffset);
-
+    armnn::TensorInfo outputStateOutTensorInfo({batchSize, timeSize, outputSize}, armnn::DataType::Float32);
+    armnn::TensorInfo cellStateOutTensorInfo({batchSize, timeSize, outputSize}, armnn::DataType::Float32);
     armnn::TensorInfo outputTensorInfo({timeSize, batchSize, outputSize}, ArmnnType, qScale, qOffset);
 
     std::vector<T> inputVector;
@@ -407,6 +434,8 @@
     std::vector<T> cellStateInVector(batchSize * numUnits, T());
     std::vector<T> outputStateInVector(batchSize * outputSize, T());
 
+    std::vector<T> actualOutputStateOut(outputStateOutTensorInfo.GetNumElements());
+    std::vector<T> actualCellStateOut(cellStateOutTensorInfo.GetNumElements());
     std::vector<T> actualOutput(outputTensorInfo.GetNumElements());
 
     std::vector<T> outputVector;
@@ -418,6 +447,10 @@
     std::unique_ptr<armnn::ITensorHandle> outputStateInHandle =
         tensorHandleFactory.CreateTensorHandle(outputStateInTensorInfo);
 
+    std::unique_ptr<armnn::ITensorHandle> outputStateOutHandle =
+        tensorHandleFactory.CreateTensorHandle(outputStateOutTensorInfo);
+    std::unique_ptr<armnn::ITensorHandle> cellStateOutHandle =
+        tensorHandleFactory.CreateTensorHandle(cellStateOutTensorInfo);
     std::unique_ptr<armnn::ITensorHandle> outputHandle = tensorHandleFactory.CreateTensorHandle(outputTensorInfo);
 
     armnn::UnidirectionalSequenceLstmQueueDescriptor data;
@@ -427,6 +460,8 @@
     AddInputToWorkload(data, info, outputStateInTensorInfo, outputStateInHandle.get());
     AddInputToWorkload(data, info, cellStateInTensorInfo, cellStateInHandle.get());
 
+    AddOutputToWorkload(data, info, outputStateOutTensorInfo, outputStateOutHandle.get());
+    AddOutputToWorkload(data, info, cellStateOutTensorInfo, cellStateOutHandle.get());
     AddOutputToWorkload(data, info, outputTensorInfo, outputHandle.get());
 
     armnn::TensorInfo tensorInfo4({numUnits}, constantDataType, qScale, qOffset);
@@ -535,6 +570,8 @@
     outputStateInHandle->Allocate();
     cellStateInHandle->Allocate();
 
+    outputStateOutHandle->Allocate();
+    cellStateOutHandle->Allocate();
     outputHandle->Allocate();
 
     CopyDataToITensorHandle(inputHandle.get(), inputVector.data());
@@ -543,6 +580,8 @@
 
     workload->Execute();
 
+    CopyDataFromITensorHandle(actualOutputStateOut.data(), outputStateOutHandle.get());
+    CopyDataFromITensorHandle(actualCellStateOut.data(), cellStateOutHandle.get());
     CopyDataFromITensorHandle(actualOutput.data(), outputHandle.get());
 
     return LayerTestResult<T, 3>(actualOutput,
@@ -644,6 +683,8 @@
     armnn::TensorInfo inputTensorInfo({batchSize, timeSize, inputSize}, armnn::DataType::Float32);
     armnn::TensorInfo cellStateInTensorInfo({batchSize , numUnits}, armnn::DataType::Float32);
     armnn::TensorInfo outputStateInTensorInfo({batchSize , outputSize}, armnn::DataType::Float32);
+    armnn::TensorInfo outputStateOutTensorInfo({batchSize, timeSize, outputSize}, armnn::DataType::Float32);
+    armnn::TensorInfo cellStateOutTensorInfo({batchSize, timeSize, outputSize}, armnn::DataType::Float32);
     armnn::TensorInfo outputTensorInfo({batchSize, timeSize, outputSize}, armnn::DataType::Float32);
 
     const std::vector<float> inputVector = { 1., 2., 3., 4., 5., 4.,
@@ -654,6 +695,8 @@
     std::vector<float> cellStateInVector(batchSize * numUnits, 0.f);
     std::vector<float> outputStateInVector(batchSize * outputSize, 0.f);
 
+    std::vector<float> actualOutputStateOut(outputStateOutTensorInfo.GetNumElements());
+    std::vector<float> actualCellStateOut(cellStateOutTensorInfo.GetNumElements());
     std::vector<float> actualOutput(outputTensorInfo.GetNumElements());
 
     const std::vector<float> expectedOutput = { -0.0135612f, -0.0263441f, 0.0314008f, -0.00883455f, 0.00763052f,
@@ -668,6 +711,11 @@
             tensorHandleFactory.CreateTensorHandle(cellStateInTensorInfo);
     std::unique_ptr<armnn::ITensorHandle> outputStateInHandle =
             tensorHandleFactory.CreateTensorHandle(outputStateInTensorInfo);
+
+    std::unique_ptr<armnn::ITensorHandle> outputStateOutHandle =
+            tensorHandleFactory.CreateTensorHandle(outputStateOutTensorInfo);
+    std::unique_ptr<armnn::ITensorHandle> cellStateOutHandle =
+            tensorHandleFactory.CreateTensorHandle(cellStateOutTensorInfo);
     std::unique_ptr<armnn::ITensorHandle> outputHandle = tensorHandleFactory.CreateTensorHandle(outputTensorInfo);
 
     armnn::UnidirectionalSequenceLstmQueueDescriptor data;
@@ -676,6 +724,9 @@
     AddInputToWorkload(data, info, inputTensorInfo, inputHandle.get());
     AddInputToWorkload(data, info, outputStateInTensorInfo, outputStateInHandle.get());
     AddInputToWorkload(data, info, cellStateInTensorInfo, cellStateInHandle.get());
+
+    AddOutputToWorkload(data, info, outputStateOutTensorInfo, outputStateOutHandle.get());
+    AddOutputToWorkload(data, info, cellStateOutTensorInfo, cellStateOutHandle.get());
     AddOutputToWorkload(data, info, outputTensorInfo, outputHandle.get());
 
     armnn::TensorInfo tensorInfo5({outputSize}, armnn::DataType::Float32);
@@ -849,6 +900,9 @@
     inputHandle->Allocate();
     outputStateInHandle->Allocate();
     cellStateInHandle->Allocate();
+
+    outputStateOutHandle->Allocate();
+    cellStateOutHandle->Allocate();
     outputHandle->Allocate();
 
     CopyDataToITensorHandle(inputHandle.get(), inputVector.data());
@@ -857,6 +911,8 @@
 
     workload->Execute();
 
+    CopyDataFromITensorHandle(actualOutputStateOut.data(), outputStateOutHandle.get());
+    CopyDataFromITensorHandle(actualCellStateOut.data(), cellStateOutHandle.get());
     CopyDataFromITensorHandle(actualOutput.data(), outputHandle.get());
 
     return LayerTestResult<float, 3>(actualOutput,
@@ -880,6 +936,8 @@
     armnn::TensorInfo inputTensorInfo({batchSize, timeSize, inputSize}, armnn::DataType::Float32);
     armnn::TensorInfo cellStateInTensorInfo({batchSize , numUnits}, armnn::DataType::Float32);
     armnn::TensorInfo outputStateInTensorInfo({batchSize , outputSize}, armnn::DataType::Float32);
+    armnn::TensorInfo outputStateOutTensorInfo({batchSize, timeSize, outputSize}, armnn::DataType::Float32);
+    armnn::TensorInfo cellStateOutTensorInfo({batchSize, timeSize, outputSize}, armnn::DataType::Float32);
     armnn::TensorInfo outputTensorInfo({batchSize, timeSize, outputSize}, armnn::DataType::Float32);
 
     const std::vector<float> inputVector = { 1., 2., 3., 4., 5., 4.,
@@ -889,6 +947,8 @@
     std::vector<float> cellStateInVector(batchSize * numUnits, 0.f);
     std::vector<float> outputStateInVector(batchSize * outputSize, 0.f);
 
+    std::vector<float> actualOutputStateOut(outputStateOutTensorInfo.GetNumElements());
+    std::vector<float> actualCellStateOut(cellStateOutTensorInfo.GetNumElements());
     std::vector<float> actualOutput(outputTensorInfo.GetNumElements());
 
     const std::vector<float> expectedOutput = { 0.0642256f, 0.0343966f, 0.184122f, 0.114717f,
@@ -904,6 +964,10 @@
     std::unique_ptr<armnn::ITensorHandle> outputStateInHandle =
             tensorHandleFactory.CreateTensorHandle(outputStateInTensorInfo);
 
+    std::unique_ptr<armnn::ITensorHandle> outputStateOutHandle =
+            tensorHandleFactory.CreateTensorHandle(outputStateOutTensorInfo);
+    std::unique_ptr<armnn::ITensorHandle> cellStateOutHandle =
+            tensorHandleFactory.CreateTensorHandle(cellStateOutTensorInfo);
     std::unique_ptr<armnn::ITensorHandle> outputHandle = tensorHandleFactory.CreateTensorHandle(outputTensorInfo);
 
     armnn::UnidirectionalSequenceLstmQueueDescriptor data;
@@ -913,6 +977,8 @@
     AddInputToWorkload(data, info, outputStateInTensorInfo, outputStateInHandle.get());
     AddInputToWorkload(data, info, cellStateInTensorInfo, cellStateInHandle.get());
 
+    AddOutputToWorkload(data, info, outputStateOutTensorInfo, outputStateOutHandle.get());
+    AddOutputToWorkload(data, info, cellStateOutTensorInfo, cellStateOutHandle.get());
     AddOutputToWorkload(data, info, outputTensorInfo, outputHandle.get());
 
     armnn::TensorInfo tensorInfo4({outputSize}, armnn::DataType::Float32);
@@ -1074,6 +1140,9 @@
     inputHandle->Allocate();
     outputStateInHandle->Allocate();
     cellStateInHandle->Allocate();
+
+    outputStateOutHandle->Allocate();
+    cellStateOutHandle->Allocate();
     outputHandle->Allocate();
 
     CopyDataToITensorHandle(inputHandle.get(), inputVector.data());
@@ -1082,6 +1151,8 @@
 
     workload->Execute();
 
+    CopyDataFromITensorHandle(actualOutputStateOut.data(), outputStateOutHandle.get());
+    CopyDataFromITensorHandle(actualCellStateOut.data(), cellStateOutHandle.get());
     CopyDataFromITensorHandle(actualOutput.data(), outputHandle.get());
 
     return LayerTestResult<float, 3>(actualOutput,
@@ -1105,7 +1176,8 @@
     armnn::TensorInfo inputTensorInfo({batchSize, timeSize, inputSize}, armnn::DataType::Float32);
     armnn::TensorInfo cellStateInTensorInfo({batchSize, numUnits}, armnn::DataType::Float32);
     armnn::TensorInfo outputStateInTensorInfo({batchSize, outputSize}, armnn::DataType::Float32);
-
+    armnn::TensorInfo outputStateOutTensorInfo({batchSize, timeSize, outputSize}, armnn::DataType::Float32);
+    armnn::TensorInfo cellStateOutTensorInfo({batchSize, timeSize, outputSize}, armnn::DataType::Float32);
     armnn::TensorInfo outputTensorInfo({batchSize, timeSize, outputSize}, armnn::DataType::Float32);
 
     std::vector<float> inputVector = { 1., 2., 3., 4., 5., 4.,
@@ -1115,6 +1187,8 @@
     std::vector<float> cellStateInVector(batchSize * numUnits, 0.f);
     std::vector<float> outputStateInVector(batchSize * outputSize, 0.f);
 
+    std::vector<float> actualOutputStateOut(outputStateOutTensorInfo.GetNumElements());
+    std::vector<float> actualCellStateOut(cellStateOutTensorInfo.GetNumElements());
     std::vector<float> actualOutput(outputTensorInfo.GetNumElements());
 
     std::vector<float> outputVector = { -0.0129257f, -0.070531f, -0.153508f, -0.0392391f,
@@ -1130,6 +1204,10 @@
     std::unique_ptr<armnn::ITensorHandle> outputStateInHandle =
         tensorHandleFactory.CreateTensorHandle(outputStateInTensorInfo);
 
+    std::unique_ptr<armnn::ITensorHandle> outputStateOutHandle =
+        tensorHandleFactory.CreateTensorHandle(outputStateOutTensorInfo);
+    std::unique_ptr<armnn::ITensorHandle> cellStateOutHandle =
+        tensorHandleFactory.CreateTensorHandle(cellStateOutTensorInfo);
     std::unique_ptr<armnn::ITensorHandle> outputHandle = tensorHandleFactory.CreateTensorHandle(outputTensorInfo);
 
     armnn::UnidirectionalSequenceLstmQueueDescriptor data;
@@ -1139,6 +1217,8 @@
     AddInputToWorkload(data, info, outputStateInTensorInfo, outputStateInHandle.get());
     AddInputToWorkload(data, info, cellStateInTensorInfo, cellStateInHandle.get());
 
+    AddOutputToWorkload(data, info, outputStateOutTensorInfo, outputStateOutHandle.get());
+    AddOutputToWorkload(data, info, cellStateOutTensorInfo, cellStateOutHandle.get());
     AddOutputToWorkload(data, info, outputTensorInfo, outputHandle.get());
 
     armnn::TensorInfo tensorInfo4({numUnits}, armnn::DataType::Float32);
@@ -1236,6 +1316,8 @@
     outputStateInHandle->Allocate();
     cellStateInHandle->Allocate();
 
+    outputStateOutHandle->Allocate();
+    cellStateOutHandle->Allocate();
     outputHandle->Allocate();
 
     CopyDataToITensorHandle(inputHandle.get(), inputVector.data());
@@ -1244,6 +1326,8 @@
 
     workload->Execute();
 
+    CopyDataFromITensorHandle(actualOutputStateOut.data(), outputStateOutHandle.get());
+    CopyDataFromITensorHandle(actualCellStateOut.data(), cellStateOutHandle.get());
     CopyDataFromITensorHandle(actualOutput.data(), outputHandle.get());
 
     return LayerTestResult<float, 3>(actualOutput,
@@ -1267,7 +1351,8 @@
     armnn::TensorInfo inputTensorInfo({batchSize, timeSize, inputSize}, armnn::DataType::Float32);
     armnn::TensorInfo cellStateInTensorInfo({batchSize, numUnits}, armnn::DataType::Float32);
     armnn::TensorInfo outputStateInTensorInfo({batchSize, outputSize}, armnn::DataType::Float32);
-
+    armnn::TensorInfo outputStateOutTensorInfo({batchSize, timeSize, outputSize}, armnn::DataType::Float32);
+    armnn::TensorInfo cellStateOutTensorInfo({batchSize, timeSize, outputSize}, armnn::DataType::Float32);
     armnn::TensorInfo outputTensorInfo({batchSize, timeSize, outputSize}, armnn::DataType::Float32);
 
     const std::vector<float> inputVector = { 0.1f, 0.2f, 0.3f, 0.4f, 0.5f, 0.4f,
@@ -1277,6 +1362,8 @@
     std::vector<float> cellStateInVector(batchSize * numUnits, 0.f);
     std::vector<float> outputStateInVector(batchSize * outputSize, 0.f);
 
+    std::vector<float> actualOutputStateOut(outputStateOutTensorInfo.GetNumElements());
+    std::vector<float> actualCellStateOut(cellStateOutTensorInfo.GetNumElements());
     std::vector<float> actualOutput(outputTensorInfo.GetNumElements());
 
     const std::vector<float> outputVector = { -0.0142517f, -0.0198845f, -0.0120569f, -0.0116868f,
@@ -1292,8 +1379,13 @@
     std::unique_ptr<armnn::ITensorHandle> outputStateInHandle =
         tensorHandleFactory.CreateTensorHandle(outputStateInTensorInfo);
 
+    std::unique_ptr<armnn::ITensorHandle> outputStateOutHandle =
+        tensorHandleFactory.CreateTensorHandle(outputStateOutTensorInfo);
+    std::unique_ptr<armnn::ITensorHandle> cellStateOutHandle =
+        tensorHandleFactory.CreateTensorHandle(cellStateOutTensorInfo);
     std::unique_ptr<armnn::ITensorHandle> outputHandle = tensorHandleFactory.CreateTensorHandle(outputTensorInfo);
 
+
     armnn::UnidirectionalSequenceLstmQueueDescriptor data;
     armnn::WorkloadInfo info;
 
@@ -1301,6 +1393,8 @@
     AddInputToWorkload(data, info, outputStateInTensorInfo, outputStateInHandle.get());
     AddInputToWorkload(data, info, cellStateInTensorInfo, cellStateInHandle.get());
 
+    AddOutputToWorkload(data, info, outputStateOutTensorInfo, outputStateOutHandle.get());
+    AddOutputToWorkload(data, info, cellStateOutTensorInfo, cellStateOutHandle.get());
     AddOutputToWorkload(data, info, outputTensorInfo, outputHandle.get());
 
     armnn::TensorInfo tensorInfoNumFp({numUnits}, armnn::DataType::Float32);
@@ -1376,6 +1470,8 @@
     outputStateInHandle->Allocate();
     cellStateInHandle->Allocate();
 
+    outputStateOutHandle->Allocate();
+    cellStateOutHandle->Allocate();
     outputHandle->Allocate();
 
     CopyDataToITensorHandle(inputHandle.get(), inputVector.data());
@@ -1384,6 +1480,8 @@
 
     workload->Execute();
 
+    CopyDataFromITensorHandle(actualOutputStateOut.data(), outputStateOutHandle.get());
+    CopyDataFromITensorHandle(actualCellStateOut.data(), cellStateOutHandle.get());
     CopyDataFromITensorHandle(actualOutput.data(), outputHandle.get());
 
     return LayerTestResult<float, 3>(actualOutput,
@@ -1407,7 +1505,8 @@
     armnn::TensorInfo inputTensorInfo({timeSize, batchSize, inputSize}, armnn::DataType::Float32);
     armnn::TensorInfo cellStateInTensorInfo({batchSize, numUnits}, armnn::DataType::Float32);
     armnn::TensorInfo outputStateInTensorInfo({batchSize, outputSize}, armnn::DataType::Float32);
-
+    armnn::TensorInfo outputStateOutTensorInfo({batchSize, timeSize, outputSize}, armnn::DataType::Float32);
+    armnn::TensorInfo cellStateOutTensorInfo({batchSize, timeSize, outputSize}, armnn::DataType::Float32);
     armnn::TensorInfo outputTensorInfo({timeSize, batchSize, outputSize}, armnn::DataType::Float32);
 
     const std::vector<float> inputVector = { 0.1f, 0.2f, 0.3f, 0.4f, 0.5f, 0.4f,
@@ -1417,6 +1516,8 @@
     std::vector<float> cellStateInVector(batchSize * numUnits, 0.f);
     std::vector<float> outputStateInVector(batchSize * outputSize, 0.f);
 
+    std::vector<float> actualOutputStateOut(outputStateOutTensorInfo.GetNumElements());
+    std::vector<float> actualCellStateOut(cellStateOutTensorInfo.GetNumElements());
     std::vector<float> actualOutput(outputTensorInfo.GetNumElements());
 
     const std::vector<float> outputVector = { -0.0142517f, -0.0198845f, -0.0120122f, -0.0116868f,
@@ -1431,8 +1532,13 @@
     std::unique_ptr<armnn::ITensorHandle> outputStateInHandle =
         tensorHandleFactory.CreateTensorHandle(outputStateInTensorInfo);
 
+    std::unique_ptr<armnn::ITensorHandle> outputStateOutHandle =
+        tensorHandleFactory.CreateTensorHandle(outputStateOutTensorInfo);
+    std::unique_ptr<armnn::ITensorHandle> cellStateOutHandle =
+        tensorHandleFactory.CreateTensorHandle(cellStateOutTensorInfo);
     std::unique_ptr<armnn::ITensorHandle> outputHandle = tensorHandleFactory.CreateTensorHandle(outputTensorInfo);
 
+
     armnn::UnidirectionalSequenceLstmQueueDescriptor data;
     armnn::WorkloadInfo info;
 
@@ -1440,6 +1546,8 @@
     AddInputToWorkload(data, info, outputStateInTensorInfo, outputStateInHandle.get());
     AddInputToWorkload(data, info, cellStateInTensorInfo, cellStateInHandle.get());
 
+    AddOutputToWorkload(data, info, outputStateOutTensorInfo, outputStateOutHandle.get());
+    AddOutputToWorkload(data, info, cellStateOutTensorInfo, cellStateOutHandle.get());
     AddOutputToWorkload(data, info, outputTensorInfo, outputHandle.get());
 
     armnn::TensorInfo tensorInfoNumFp({numUnits}, armnn::DataType::Float32);
@@ -1516,6 +1624,8 @@
     outputStateInHandle->Allocate();
     cellStateInHandle->Allocate();
 
+    outputStateOutHandle->Allocate();
+    cellStateOutHandle->Allocate();
     outputHandle->Allocate();
 
     CopyDataToITensorHandle(inputHandle.get(), inputVector.data());
@@ -1524,6 +1634,8 @@
 
     workload->Execute();
 
+    CopyDataFromITensorHandle(actualOutputStateOut.data(), outputStateOutHandle.get());
+    CopyDataFromITensorHandle(actualCellStateOut.data(), cellStateOutHandle.get());
     CopyDataFromITensorHandle(actualOutput.data(), outputHandle.get());
 
     return LayerTestResult<float, 3>(actualOutput,
@@ -1547,6 +1659,8 @@
     armnn::TensorInfo inputTensorInfo({batchSize, timeSize, inputSize}, armnn::DataType::Float32);
     armnn::TensorInfo cellStateInTensorInfo({batchSize , numUnits}, armnn::DataType::Float32);
     armnn::TensorInfo outputStateInTensorInfo({batchSize , outputSize}, armnn::DataType::Float32);
+    armnn::TensorInfo outputStateOutTensorInfo({batchSize, timeSize, outputSize}, armnn::DataType::Float32);
+    armnn::TensorInfo cellStateOutTensorInfo({batchSize, timeSize, outputSize}, armnn::DataType::Float32);
     armnn::TensorInfo outputTensorInfo({batchSize, timeSize, outputSize}, armnn::DataType::Float32);
 
     const std::vector<float> inputVector = { 0.1f, 0.2f, 0.3f, 0.4f, 0.5f, 0.4f,
@@ -1556,6 +1670,8 @@
     std::vector<float> cellStateInVector(batchSize * numUnits, 0.f);
     std::vector<float> outputStateInVector(batchSize * outputSize, 0.f);
 
+    std::vector<float> actualOutputStateOut(outputStateOutTensorInfo.GetNumElements());
+    std::vector<float> actualCellStateOut(cellStateOutTensorInfo.GetNumElements());
     std::vector<float> actualOutput(outputTensorInfo.GetNumElements());
 
     const std::vector<float> expectedOutput = { 0.612103f, 1.56788f, 0.31966f, 1.42956f,
@@ -1570,6 +1686,11 @@
             tensorHandleFactory.CreateTensorHandle(cellStateInTensorInfo);
     std::unique_ptr<armnn::ITensorHandle> outputStateInHandle =
             tensorHandleFactory.CreateTensorHandle(outputStateInTensorInfo);
+
+    std::unique_ptr<armnn::ITensorHandle> outputStateOutHandle =
+            tensorHandleFactory.CreateTensorHandle(outputStateOutTensorInfo);
+    std::unique_ptr<armnn::ITensorHandle> cellStateOutHandle =
+            tensorHandleFactory.CreateTensorHandle(cellStateOutTensorInfo);
     std::unique_ptr<armnn::ITensorHandle> outputHandle = tensorHandleFactory.CreateTensorHandle(outputTensorInfo);
 
     armnn::UnidirectionalSequenceLstmQueueDescriptor data;
@@ -1578,6 +1699,9 @@
     AddInputToWorkload(data, info, inputTensorInfo, inputHandle.get());
     AddInputToWorkload(data, info, outputStateInTensorInfo, outputStateInHandle.get());
     AddInputToWorkload(data, info, cellStateInTensorInfo, cellStateInHandle.get());
+
+    AddOutputToWorkload(data, info, outputStateOutTensorInfo, outputStateOutHandle.get());
+    AddOutputToWorkload(data, info, cellStateOutTensorInfo, cellStateOutHandle.get());
     AddOutputToWorkload(data, info, outputTensorInfo, outputHandle.get());
 
     armnn::TensorInfo tensorInfoOut({outputSize}, armnn::DataType::Float32);
@@ -1679,6 +1803,9 @@
     inputHandle->Allocate();
     outputStateInHandle->Allocate();
     cellStateInHandle->Allocate();
+
+    outputStateOutHandle->Allocate();
+    cellStateOutHandle->Allocate();
     outputHandle->Allocate();
 
     CopyDataToITensorHandle(inputHandle.get(), inputVector.data());
@@ -1687,6 +1814,8 @@
 
     workload->Execute();
 
+    CopyDataFromITensorHandle(actualOutputStateOut.data(), outputStateOutHandle.get());
+    CopyDataFromITensorHandle(actualCellStateOut.data(), cellStateOutHandle.get());
     CopyDataFromITensorHandle(actualOutput.data(), outputHandle.get());
 
     return LayerTestResult<float, 3>(actualOutput,
@@ -1710,6 +1839,8 @@
     armnn::TensorInfo inputTensorInfo({batchSize, timeSize, inputSize}, armnn::DataType::Float32);
     armnn::TensorInfo cellStateInTensorInfo({batchSize , numUnits}, armnn::DataType::Float32);
     armnn::TensorInfo outputStateInTensorInfo({batchSize , outputSize}, armnn::DataType::Float32);
+    armnn::TensorInfo outputStateOutTensorInfo({batchSize, timeSize, outputSize}, armnn::DataType::Float32);
+    armnn::TensorInfo cellStateOutTensorInfo({batchSize, timeSize, outputSize}, armnn::DataType::Float32);
     armnn::TensorInfo outputTensorInfo({batchSize, timeSize, outputSize}, armnn::DataType::Float32);
 
     const std::vector<float> inputVector = { 1., 8., 3., 4., 5., 4.,
@@ -1719,6 +1850,8 @@
     std::vector<float> cellStateInVector(batchSize * numUnits, 0.f);
     std::vector<float> outputStateInVector(batchSize * outputSize, 0.f);
 
+    std::vector<float> actualOutputStateOut(outputStateOutTensorInfo.GetNumElements());
+    std::vector<float> actualCellStateOut(cellStateOutTensorInfo.GetNumElements());
     std::vector<float> actualOutput(outputTensorInfo.GetNumElements());
 
     const std::vector<float> expectedOutput = { 0.0471276f, 0.0168155f, 0.0789885f, 0.16550f,
@@ -1734,6 +1867,10 @@
     std::unique_ptr<armnn::ITensorHandle> outputStateInHandle =
             tensorHandleFactory.CreateTensorHandle(outputStateInTensorInfo);
 
+    std::unique_ptr<armnn::ITensorHandle> outputStateOutHandle =
+            tensorHandleFactory.CreateTensorHandle(outputStateOutTensorInfo);
+    std::unique_ptr<armnn::ITensorHandle> cellStateOutHandle =
+            tensorHandleFactory.CreateTensorHandle(cellStateOutTensorInfo);
     std::unique_ptr<armnn::ITensorHandle> outputHandle = tensorHandleFactory.CreateTensorHandle(outputTensorInfo);
 
     armnn::UnidirectionalSequenceLstmQueueDescriptor data;
@@ -1743,6 +1880,8 @@
     AddInputToWorkload(data, info, outputStateInTensorInfo, outputStateInHandle.get());
     AddInputToWorkload(data, info, cellStateInTensorInfo, cellStateInHandle.get());
 
+    AddOutputToWorkload(data, info, outputStateOutTensorInfo, outputStateOutHandle.get());
+    AddOutputToWorkload(data, info, cellStateOutTensorInfo, cellStateOutHandle.get());
     AddOutputToWorkload(data, info, outputTensorInfo, outputHandle.get());
 
     armnn::TensorInfo tensorInfoOut({outputSize}, armnn::DataType::Float32);
@@ -1871,6 +2010,9 @@
     inputHandle->Allocate();
     outputStateInHandle->Allocate();
     cellStateInHandle->Allocate();
+
+    outputStateOutHandle->Allocate();
+    cellStateOutHandle->Allocate();
     outputHandle->Allocate();
 
     CopyDataToITensorHandle(inputHandle.get(), inputVector.data());
@@ -1879,6 +2021,8 @@
 
     workload->Execute();
 
+    CopyDataFromITensorHandle(actualOutputStateOut.data(), outputStateOutHandle.get());
+    CopyDataFromITensorHandle(actualCellStateOut.data(), cellStateOutHandle.get());
     CopyDataFromITensorHandle(actualOutput.data(), outputHandle.get());
 
     return LayerTestResult<float, 3>(actualOutput,
@@ -1902,7 +2046,8 @@
     armnn::TensorInfo inputTensorInfo({batchSize, timeSize, inputSize}, armnn::DataType::Float32);
     armnn::TensorInfo cellStateInTensorInfo({batchSize, numUnits}, armnn::DataType::Float32);
     armnn::TensorInfo outputStateInTensorInfo({batchSize, outputSize}, armnn::DataType::Float32);
-
+    armnn::TensorInfo outputStateOutTensorInfo({batchSize, timeSize, outputSize}, armnn::DataType::Float32);
+    armnn::TensorInfo cellStateOutTensorInfo({batchSize, timeSize, outputSize}, armnn::DataType::Float32);
     armnn::TensorInfo outputTensorInfo({batchSize, timeSize, outputSize}, armnn::DataType::Float32);
 
     const std::vector<float> inputVector = { 0.1f, 0.2f, 0.3f, 0.4f, 0.5f, 0.4f,
@@ -1912,6 +2057,8 @@
     std::vector<float> cellStateInVector(batchSize * numUnits, 0.f);
     std::vector<float> outputStateInVector(batchSize * outputSize, 0.f);
 
+    std::vector<float> actualOutputStateOut(outputStateOutTensorInfo.GetNumElements());
+    std::vector<float> actualCellStateOut(cellStateOutTensorInfo.GetNumElements());
     std::vector<float> actualOutput(outputTensorInfo.GetNumElements());
 
     const std::vector<float> outputVector = { -0.0072104f, -0.00991171f, -0.00650478f, -0.00713055f,
@@ -1927,6 +2074,10 @@
     std::unique_ptr<armnn::ITensorHandle> outputStateInHandle =
         tensorHandleFactory.CreateTensorHandle(outputStateInTensorInfo);
 
+    std::unique_ptr<armnn::ITensorHandle> outputStateOutHandle =
+        tensorHandleFactory.CreateTensorHandle(outputStateOutTensorInfo);
+    std::unique_ptr<armnn::ITensorHandle> cellStateOutHandle =
+        tensorHandleFactory.CreateTensorHandle(cellStateOutTensorInfo);
     std::unique_ptr<armnn::ITensorHandle> outputHandle = tensorHandleFactory.CreateTensorHandle(outputTensorInfo);
 
     armnn::UnidirectionalSequenceLstmQueueDescriptor data;
@@ -1936,6 +2087,8 @@
     AddInputToWorkload(data, info, outputStateInTensorInfo, outputStateInHandle.get());
     AddInputToWorkload(data, info, cellStateInTensorInfo, cellStateInHandle.get());
 
+    AddOutputToWorkload(data, info, outputStateOutTensorInfo, outputStateOutHandle.get());
+    AddOutputToWorkload(data, info, cellStateOutTensorInfo, cellStateOutHandle.get());
     AddOutputToWorkload(data, info, outputTensorInfo, outputHandle.get());
 
     armnn::TensorInfo tensorInfoNumFp({numUnits}, armnn::DataType::Float32);
@@ -2009,6 +2162,8 @@
     outputStateInHandle->Allocate();
     cellStateInHandle->Allocate();
 
+    outputStateOutHandle->Allocate();
+    cellStateOutHandle->Allocate();
     outputHandle->Allocate();
 
     CopyDataToITensorHandle(inputHandle.get(), inputVector.data());
@@ -2017,10 +2172,12 @@
 
     workload->Execute();
 
+    CopyDataFromITensorHandle(actualOutputStateOut.data(), outputStateOutHandle.get());
+    CopyDataFromITensorHandle(actualCellStateOut.data(), cellStateOutHandle.get());
     CopyDataFromITensorHandle(actualOutput.data(), outputHandle.get());
 
     return LayerTestResult<float, 3>(actualOutput,
                                      outputVector,
                                      outputHandle->GetShape(),
                                      outputTensorInfo.GetShape());
-}
\ No newline at end of file
+}
diff --git a/src/backends/cl/ClLayerSupport.cpp b/src/backends/cl/ClLayerSupport.cpp
index 4dcaca9..09b3f43 100644
--- a/src/backends/cl/ClLayerSupport.cpp
+++ b/src/backends/cl/ClLayerSupport.cpp
@@ -1444,9 +1444,9 @@
 bool ClLayerSupport::IsUnidirectionalSequenceLstmSupported(const TensorInfo& input,
                                                            const TensorInfo& outputStateIn,
                                                            const TensorInfo& cellStateIn,
+                                                           const TensorInfo& outputStateOut,
+                                                           const TensorInfo& cellStateOut,
                                                            const TensorInfo& output,
-                                                           const Optional<TensorInfo>& hiddenStateOutput,
-                                                           const Optional<TensorInfo>& cellStateOutput,
                                                            const UnidirectionalSequenceLstmDescriptor& descriptor,
                                                            const LstmInputParamsInfo& paramsInfo,
                                                            Optional<std::string&> reasonIfUnsupported) const
@@ -1456,9 +1456,9 @@
                                    input,
                                    outputStateIn,
                                    cellStateIn,
+                                   outputStateOut,
+                                   cellStateOut,
                                    output,
-                                   hiddenStateOutput,
-                                   cellStateOutput,
                                    descriptor,
                                    paramsInfo);
 }
diff --git a/src/backends/cl/ClLayerSupport.hpp b/src/backends/cl/ClLayerSupport.hpp
index b4d0e82..4f4e64e 100644
--- a/src/backends/cl/ClLayerSupport.hpp
+++ b/src/backends/cl/ClLayerSupport.hpp
@@ -334,9 +334,9 @@
     bool IsUnidirectionalSequenceLstmSupported(const TensorInfo& input,
                                                const TensorInfo& outputStateIn,
                                                const TensorInfo& cellStateIn,
+                                               const TensorInfo& outputStateOut,
+                                               const TensorInfo& cellStateOut,
                                                const TensorInfo& output,
-                                               const Optional<TensorInfo>& hiddenStateOutput,
-                                               const Optional<TensorInfo>& cellStateOutput,
                                                const UnidirectionalSequenceLstmDescriptor& descriptor,
                                                const LstmInputParamsInfo& paramsInfo,
                                                Optional<std::string&> reasonIfUnsupported) const override;
diff --git a/src/backends/cl/workloads/ClUnidirectionalSequenceLstmFloatWorkload.cpp b/src/backends/cl/workloads/ClUnidirectionalSequenceLstmFloatWorkload.cpp
index cc9aea8..ac24120 100644
--- a/src/backends/cl/workloads/ClUnidirectionalSequenceLstmFloatWorkload.cpp
+++ b/src/backends/cl/workloads/ClUnidirectionalSequenceLstmFloatWorkload.cpp
@@ -41,17 +41,17 @@
                                          GetGuid());
 
     const arm_compute::ICLTensor& input = static_cast<IClTensorHandle*>(m_Data.m_Inputs[0])->GetTensor();
-    arm_compute::ICLTensor& output = static_cast<IClTensorHandle*>(m_Data.m_Outputs[0])->GetTensor();
+    arm_compute::ICLTensor& output = static_cast<IClTensorHandle*>(m_Data.m_Outputs[2])->GetTensor();
 
     TensorInfo inputInfo = info.m_InputTensorInfos[0];
-    TensorInfo outputInfo = info.m_OutputTensorInfos[0];
+    TensorInfo outputInfo = info.m_OutputTensorInfos[2];
 
     arm_compute::DataType armComputeDataType = static_cast<IClTensorHandle*>(m_Data.m_Inputs[0])->GetDataType();
     armnn::DataType armnnDataType = GetArmNNDataType(armComputeDataType);
 
     TensorShape inputLayerShape = static_cast<IClTensorHandle*>(m_Data.m_Inputs[0])->GetShape();
     TensorShape cellStateLayerShape = static_cast<IClTensorHandle*>(m_Data.m_Inputs[2])->GetShape();
-    TensorShape outputLayerShape = static_cast<IClTensorHandle*>(m_Data.m_Outputs[0])->GetShape();
+    TensorShape outputLayerShape = static_cast<IClTensorHandle*>(m_Data.m_Outputs[2])->GetShape();
 
     unsigned int maxTime = m_Data.m_Parameters.m_TimeMajor ? inputLayerShape[0] : inputLayerShape[1];
     unsigned int batchSize = m_Data.m_Parameters.m_TimeMajor ? inputLayerShape[1] : inputLayerShape[0];
diff --git a/src/backends/neon/NeonLayerSupport.cpp b/src/backends/neon/NeonLayerSupport.cpp
index e2098a3..2105355 100644
--- a/src/backends/neon/NeonLayerSupport.cpp
+++ b/src/backends/neon/NeonLayerSupport.cpp
@@ -77,6 +77,7 @@
 #include "workloads/NeonTransposeConvolution2dWorkload.hpp"
 #include "workloads/NeonTransposeWorkload.hpp"
 #include "workloads/NeonUnidirectionalSequenceLstmFloatWorkload.hpp"
+#include "workloads/NeonUnidirectionalSequenceLstmWorkload.hpp"
 #endif
 
 namespace armnn
@@ -1436,23 +1437,44 @@
 bool NeonLayerSupport::IsUnidirectionalSequenceLstmSupported(const TensorInfo& input,
                                                              const TensorInfo& outputStateIn,
                                                              const TensorInfo& cellStateIn,
+                                                             const TensorInfo& outputStateOut,
+                                                             const TensorInfo& cellStateOut,
                                                              const TensorInfo& output,
-                                                             const Optional<TensorInfo>& hiddenStateOutput,
-                                                             const Optional<TensorInfo>& cellStateOutput,
                                                              const UnidirectionalSequenceLstmDescriptor& descriptor,
                                                              const LstmInputParamsInfo& paramsInfo,
                                                              Optional<std::string&> reasonIfUnsupported) const
 {
-    FORWARD_WORKLOAD_VALIDATE_FUNC(NeonUnidirectionalSequenceLstmFloatWorkloadValidate,
-                                   reasonIfUnsupported,
-                                   input,
-                                   outputStateIn,
-                                   cellStateIn,
-                                   output,
-                                   hiddenStateOutput,
-                                   cellStateOutput,
-                                   descriptor,
-                                   paramsInfo);
+    if (input.GetDataType() == armnn::DataType::QAsymmS8 &&
+        outputStateIn.GetDataType() == armnn::DataType::QAsymmS8 &&
+        cellStateIn.GetDataType() == armnn::DataType::QSymmS16 &&
+        outputStateOut.GetDataType() == armnn::DataType::QAsymmS8 &&
+        cellStateOut.GetDataType() == armnn::DataType::QSymmS16 &&
+        output.GetDataType() == armnn::DataType::QAsymmS8)
+    {
+        FORWARD_WORKLOAD_VALIDATE_FUNC(NeonUnidirectionalSequenceLstmWorkloadValidate,
+                                       reasonIfUnsupported,
+                                       input,
+                                       outputStateIn,
+                                       cellStateIn,
+                                       outputStateOut,
+                                       cellStateOut,
+                                       output,
+                                       descriptor,
+                                       paramsInfo);
+    }
+    else
+    {
+        FORWARD_WORKLOAD_VALIDATE_FUNC(NeonUnidirectionalSequenceLstmFloatWorkloadValidate,
+                                       reasonIfUnsupported,
+                                       input,
+                                       outputStateIn,
+                                       cellStateIn,
+                                       outputStateOut,
+                                       cellStateOut,
+                                       output,
+                                       descriptor,
+                                       paramsInfo);
+    }
 }
 
 } // namespace armnn
diff --git a/src/backends/neon/NeonLayerSupport.hpp b/src/backends/neon/NeonLayerSupport.hpp
index 1eef41f..511bb03 100644
--- a/src/backends/neon/NeonLayerSupport.hpp
+++ b/src/backends/neon/NeonLayerSupport.hpp
@@ -339,9 +339,9 @@
     bool IsUnidirectionalSequenceLstmSupported(const TensorInfo& input,
                                                const TensorInfo& outputStateIn,
                                                const TensorInfo& cellStateIn,
+                                               const TensorInfo& outputStateOut,
+                                               const TensorInfo& cellStateOut,
                                                const TensorInfo& output,
-                                               const Optional<TensorInfo>& hiddenStateOutput,
-                                               const Optional<TensorInfo>& cellStateOutput,
                                                const UnidirectionalSequenceLstmDescriptor& descriptor,
                                                const LstmInputParamsInfo& paramsInfo,
                                                Optional<std::string&> reasonIfUnsupported) const override;
diff --git a/src/backends/neon/NeonWorkloadFactory.cpp b/src/backends/neon/NeonWorkloadFactory.cpp
index 7d94daf..c83e8b3 100644
--- a/src/backends/neon/NeonWorkloadFactory.cpp
+++ b/src/backends/neon/NeonWorkloadFactory.cpp
@@ -558,7 +558,20 @@
         case LayerType::UnidirectionalSequenceLstm :
         {
             auto desc = PolymorphicDowncast<const UnidirectionalSequenceLstmQueueDescriptor*>(&descriptor);
-            return MakeWorkloadHelper<NeonUnidirectionalSequenceLstmFloatWorkload, NullWorkload>(*desc, info);
+
+            if ((info.m_InputTensorInfos[0].GetDataType() == armnn::DataType::Float32) &&
+                (info.m_InputTensorInfos[1].GetDataType() == armnn::DataType::Float32) &&
+                (info.m_InputTensorInfos[2].GetDataType() == armnn::DataType::Float32) &&
+                (info.m_OutputTensorInfos[0].GetDataType() == armnn::DataType::Float32) &&
+                (info.m_OutputTensorInfos[1].GetDataType() == armnn::DataType::Float32) &&
+                (info.m_OutputTensorInfos[2].GetDataType() == armnn::DataType::Float32))
+            {
+                return std::make_unique<NeonUnidirectionalSequenceLstmFloatWorkload>(*desc, info);
+            }
+            else
+            {
+                return std::make_unique<NeonUnidirectionalSequenceLstmWorkload>(*desc, info);
+            }
         }
         default:
             return nullptr;
diff --git a/src/backends/neon/backend.mk b/src/backends/neon/backend.mk
index d43426f..0d6fd80 100644
--- a/src/backends/neon/backend.mk
+++ b/src/backends/neon/backend.mk
@@ -85,7 +85,8 @@
         workloads/NeonSubtractionWorkload.cpp \
         workloads/NeonTransposeConvolution2dWorkload.cpp \
         workloads/NeonTransposeWorkload.cpp \
-        workloads/NeonUnidirectionalSequenceLstmFloatWorkload.cpp
+        workloads/NeonUnidirectionalSequenceLstmFloatWorkload.cpp \
+        workloads/NeonUnidirectionalSequenceLstmWorkload.cpp
 
 else
 
diff --git a/src/backends/neon/workloads/CMakeLists.txt b/src/backends/neon/workloads/CMakeLists.txt
index 41c5f5a..33a18e3 100644
--- a/src/backends/neon/workloads/CMakeLists.txt
+++ b/src/backends/neon/workloads/CMakeLists.txt
@@ -133,6 +133,8 @@
     NeonTransposeWorkload.hpp
     NeonUnidirectionalSequenceLstmFloatWorkload.cpp
     NeonUnidirectionalSequenceLstmFloatWorkload.hpp
+    NeonUnidirectionalSequenceLstmWorkload.cpp
+    NeonUnidirectionalSequenceLstmWorkload.hpp
     NeonWorkloads.hpp
     NeonWorkloadUtils.hpp
 )
diff --git a/src/backends/neon/workloads/NeonUnidirectionalSequenceLstmFloatWorkload.cpp b/src/backends/neon/workloads/NeonUnidirectionalSequenceLstmFloatWorkload.cpp
index c911afb..8dba719 100644
--- a/src/backends/neon/workloads/NeonUnidirectionalSequenceLstmFloatWorkload.cpp
+++ b/src/backends/neon/workloads/NeonUnidirectionalSequenceLstmFloatWorkload.cpp
@@ -39,7 +39,7 @@
                                          GetGuid());
 
     const arm_compute::ITensor& input = static_cast<IAclTensorHandle*>(m_Data.m_Inputs[0])->GetTensor();
-    arm_compute::ITensor& output = static_cast<IAclTensorHandle*>(m_Data.m_Outputs[0])->GetTensor();
+    arm_compute::ITensor& output = static_cast<IAclTensorHandle*>(m_Data.m_Outputs[2])->GetTensor();
 
     TensorInfo inputInfo = info.m_InputTensorInfos[0];
     TensorInfo outputInfo = info.m_OutputTensorInfos[0];
@@ -49,7 +49,7 @@
 
     TensorShape inputLayerShape = static_cast<IAclTensorHandle*>(m_Data.m_Inputs[0])->GetShape();
     TensorShape cellStateLayerShape = static_cast<IAclTensorHandle*>(m_Data.m_Inputs[2])->GetShape();
-    TensorShape outputLayerShape = static_cast<IAclTensorHandle*>(m_Data.m_Outputs[0])->GetShape();
+    TensorShape outputLayerShape = static_cast<IAclTensorHandle*>(m_Data.m_Outputs[2])->GetShape();
 
     unsigned int maxTime = m_Data.m_Parameters.m_TimeMajor ? inputLayerShape[0] : inputLayerShape[1];
     unsigned int batchSize = m_Data.m_Parameters.m_TimeMajor ? inputLayerShape[1] : inputLayerShape[0];
@@ -288,7 +288,7 @@
         // LSTM input/output cannot be > 2 dimensions so need to resize its TensorInfo.
         if (maxTime == 1 && m_Data.m_Parameters.m_TimeMajor)
         {
-            TensorShape inputShape = GetTensorShape((&input)->info()->tensor_shape(), 1U);
+            TensorShape inputShape = GetTensorShape(input.info()->tensor_shape(), 1U);
             TensorShape outputShape = GetTensorShape((&output)->info()->tensor_shape(), 1U);
 
             TensorShape inputShapeShrink({inputShape[1], inputShape[2]});
@@ -297,10 +297,10 @@
             auto acl_input_shape_shrink = BuildArmComputeTensorShape(inputShapeShrink);
             auto acl_output_shape_shrink = BuildArmComputeTensorShape(outputShapeShrink);
 
-            (&input)->info()->set_tensor_shape(acl_input_shape_shrink);
+            input.info()->set_tensor_shape(acl_input_shape_shrink);
             inputLSTM = const_cast<arm_compute::ITensor*>(&input);
 
-            (&output)->info()->set_tensor_shape(acl_output_shape_shrink);
+            output.info()->set_tensor_shape(acl_output_shape_shrink);
             outputLSTM = &output;
         }
         // If there is only one LSTM batch major batch, we will not concat, only permute.
@@ -432,9 +432,9 @@
         unsigned int aclAxisConcat = CalcAclAxis(concatDescriptor.GetNumDimensions(), concatDescriptor.GetConcatAxis());
         if (!m_Data.m_Parameters.m_TimeMajor)
         {
-            TensorInfo concatOuputTensorInfo = outputInfo;
-            concatOuputTensorInfo.SetShape(timeMajorShapeOutput);
-            BuildArmComputeTensor(concat_out, concatOuputTensorInfo);
+            TensorInfo concatOutputTensorInfo = outputInfo;
+            concatOutputTensorInfo.SetShape(timeMajorShapeOutput);
+            BuildArmComputeTensor(concat_out, concatOutputTensorInfo);
             armcomputetensorutils::InitialiseArmComputeTensorEmpty(concat_out);
 
             m_Concat->configure(m_ConcatInputs, &concat_out, aclAxisConcat);
@@ -452,11 +452,11 @@
     {
         if (!m_Data.m_Parameters.m_TimeMajor)
         {
-            (&output)->info()->set_tensor_shape(BuildArmComputeTensorShape(shapeExpandBatchMajor));
+            output.info()->set_tensor_shape(BuildArmComputeTensorShape(shapeExpandBatchMajor));
         }
         else
         {
-            (&output)->info()->set_tensor_shape(BuildArmComputeTensorShape(shapeExpandTimeMajor));
+            output.info()->set_tensor_shape(BuildArmComputeTensorShape(shapeExpandTimeMajor));
         }
     }
 
@@ -510,14 +510,12 @@
 NeonUnidirectionalSequenceLstmFloatWorkloadValidate(const TensorInfo& input,
                                                     const TensorInfo& outputStateIn,
                                                     const TensorInfo& cellStateIn,
+                                                    const TensorInfo& outputStateOut,
+                                                    const TensorInfo& cellStateOut,
                                                     const TensorInfo& output,
-                                                    const Optional<TensorInfo>& hiddenStateOutput,
-                                                    const Optional<TensorInfo>& cellStateOutput,
                                                     const UnidirectionalSequenceLstmDescriptor& descriptor,
                                                     const LstmInputParamsInfo& paramsInfo)
 {
-    IgnoreUnused(hiddenStateOutput, cellStateOutput);
-
     TensorShape inputLayerShape = input.GetShape();
     TensorShape outputLayerShape = outputStateIn.GetShape();
 
@@ -612,8 +610,6 @@
     arm_compute::LSTMParams<arm_compute::ITensorInfo> lstm_params_info;
 
     const TensorInfo& scratchBuffer = TensorInfo(cellStateIn.GetShape(), input.GetDataType());
-    const TensorInfo& outputStateOut = TensorInfo(outputStateIn.GetShape(), input.GetDataType());
-    const TensorInfo& cellStateOut = TensorInfo(cellStateIn.GetShape(), input.GetDataType());
 
     // The inputs and outputs
     const arm_compute::TensorInfo aclOutputStateInInfo = BuildArmComputeTensorInfo(outputStateIn);
@@ -704,7 +700,7 @@
         aclOutputLayerNormWeightsInfo = BuildArmComputeTensorInfo(paramsInfo.GetOutputLayerNormWeights());
 
         lstm_params_info.set_layer_normalization_params(descriptor.m_CifgEnabled ? nullptr :
-                                                            &aclInputLayerNormWeightsInfo,
+                                                        &aclInputLayerNormWeightsInfo,
                                                         &aclForgetLayerNormWeightsInfo,
                                                         &aclCellLayerNormWeightsInfo,
                                                         &aclOutputLayerNormWeightsInfo);
@@ -803,9 +799,9 @@
     TensorShape shapeExpandTimeMajor({1, shape[0], shape[1]});
     TensorShape shapeExpandBatchMajor({shape[0], 1, shape[1]});
 
-    TensorInfo concatOuputTensorInfo = TensorInfo(output);
-    concatOuputTensorInfo.SetShape(timeMajorShapeOutput);
-    arm_compute::TensorInfo aclConcatOuputTensorInfo= BuildArmComputeTensorInfo(concatOuputTensorInfo);
+    TensorInfo concatOutputTensorInfo = TensorInfo(output);
+    concatOutputTensorInfo.SetShape(timeMajorShapeOutput);
+    arm_compute::TensorInfo aclConcatOutputTensorInfo= BuildArmComputeTensorInfo(concatOutputTensorInfo);
 
     if (maxTime != 1) // ACL concat does not work with only one element to concatenate.
     {
@@ -819,7 +815,7 @@
         if (!descriptor.m_TimeMajor)
         {
             statusConcat = arm_compute::NEConcatenateLayer::validate(concatInputsTensorInfosPtr,
-                                                                     &aclConcatOuputTensorInfo,
+                                                                     &aclConcatOutputTensorInfo,
                                                                      aclAxisConcat);
         }
         else
@@ -853,7 +849,7 @@
         // Output now time major. Permute output back to batch major.
         if (maxTime != 1)
         {
-            statusPermute2 = arm_compute::NEPermute::validate(&aclConcatOuputTensorInfo,
+            statusPermute2 = arm_compute::NEPermute::validate(&aclConcatOutputTensorInfo,
                                                               &aclOutputInfo,
                                                               arm_compute::PermutationVector(0U, 2U, 1U));
         }
diff --git a/src/backends/neon/workloads/NeonUnidirectionalSequenceLstmFloatWorkload.hpp b/src/backends/neon/workloads/NeonUnidirectionalSequenceLstmFloatWorkload.hpp
index 776afd3..48cf7dc 100644
--- a/src/backends/neon/workloads/NeonUnidirectionalSequenceLstmFloatWorkload.hpp
+++ b/src/backends/neon/workloads/NeonUnidirectionalSequenceLstmFloatWorkload.hpp
@@ -10,7 +10,6 @@
 #include <armnn/backends/Workload.hpp>
 #include <armnn/backends/WorkloadData.hpp>
 
-#include "arm_compute/graph/Tensor.h"
 #include "arm_compute/runtime/NEON/functions/NELSTMLayer.h"
 #include "arm_compute/runtime/NEON/functions/NEPermute.h"
 #include "arm_compute/runtime/NEON/functions/NESplit.h"
@@ -86,9 +85,9 @@
 NeonUnidirectionalSequenceLstmFloatWorkloadValidate(const TensorInfo& input,
                                                     const TensorInfo& outputStateIn,
                                                     const TensorInfo& cellStateIn,
+                                                    const TensorInfo& outputStateOut,
+                                                    const TensorInfo& cellStateOut,
                                                     const TensorInfo& output,
-                                                    const Optional<TensorInfo>& hiddenStateOutput,
-                                                    const Optional<TensorInfo>& cellStateOutput,
                                                     const UnidirectionalSequenceLstmDescriptor& descriptor,
                                                     const LstmInputParamsInfo& paramsInfo);
 
diff --git a/src/backends/neon/workloads/NeonUnidirectionalSequenceLstmWorkload.cpp b/src/backends/neon/workloads/NeonUnidirectionalSequenceLstmWorkload.cpp
new file mode 100644
index 0000000..dfbbb3c
--- /dev/null
+++ b/src/backends/neon/workloads/NeonUnidirectionalSequenceLstmWorkload.cpp
@@ -0,0 +1,879 @@
+//
+// Copyright © 2022 Arm Ltd and Contributors. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+
+#include "NeonUnidirectionalSequenceLstmWorkload.hpp"
+#include "NeonWorkloadUtils.hpp"
+
+#include <aclCommon/ArmComputeUtils.hpp>
+#include <aclCommon/ArmComputeTensorUtils.hpp>
+
+#include <armnn/utility/NumericCast.hpp>
+#include <armnnUtils/Permute.hpp>
+#include <neon/test/NeonWorkloadFactoryHelper.hpp>
+#include <backendsCommon/WorkloadUtils.hpp>
+
+#include "neon/NeonTensorHandle.hpp"
+
+namespace
+{
+
+unsigned int CalcAclAxis(unsigned int numDimensions, unsigned int axis)
+{
+    return (numDimensions - axis) - 1;
+}
+} //namespace
+
+namespace armnn
+{
+using namespace armcomputetensorutils;
+
+NeonUnidirectionalSequenceLstmWorkload::NeonUnidirectionalSequenceLstmWorkload
+    (const UnidirectionalSequenceLstmQueueDescriptor& descriptor, const WorkloadInfo& info)
+    : NeonBaseWorkload<UnidirectionalSequenceLstmQueueDescriptor>(descriptor, info)
+{
+    // Report Profiling Details
+    ARMNN_REPORT_PROFILING_WORKLOAD_DESC("NeonUnidirectionalSequenceLstmWorkload_Construct",
+                                         descriptor.m_Parameters,
+                                         info,
+                                         GetGuid());
+
+    // Input/Output tensors
+    const arm_compute::ITensor& input         = static_cast<IAclTensorHandle*>(m_Data.m_Inputs[0])->GetTensor();
+    arm_compute::ITensor& outputStateIn       = static_cast<IAclTensorHandle*>(m_Data.m_Inputs[1])->GetTensor();
+    const arm_compute::ITensor& cellStateIn   = static_cast<IAclTensorHandle*>(m_Data.m_Inputs[2])->GetTensor();
+
+    arm_compute::ITensor& outputStateOut = static_cast<IAclTensorHandle*>(m_Data.m_Outputs[0])->GetTensor();
+    arm_compute::ITensor& cellStateOut   = static_cast<IAclTensorHandle*>(m_Data.m_Outputs[1])->GetTensor();
+    arm_compute::ITensor& output         = static_cast<IAclTensorHandle*>(m_Data.m_Outputs[2])->GetTensor();
+
+    TensorInfo inputInfo = info.m_InputTensorInfos[0];
+    TensorInfo outputInfo = info.m_OutputTensorInfos[2];
+
+    TensorShape inputLayerShape = static_cast<IAclTensorHandle*>(m_Data.m_Inputs[0])->GetShape();
+    TensorShape outputLayerShape = static_cast<IAclTensorHandle*>(m_Data.m_Outputs[2])->GetShape();
+
+    unsigned int maxTime = m_Data.m_Parameters.m_TimeMajor ? inputLayerShape[0] : inputLayerShape[1];
+    unsigned int batchSize = m_Data.m_Parameters.m_TimeMajor ? inputLayerShape[1] : inputLayerShape[0];
+    unsigned int inputSize = inputLayerShape[2];
+    unsigned int outputSize = outputLayerShape[2];
+
+    const TensorShape timeMajorShapeInput({maxTime, batchSize, inputSize});
+    const TensorShape timeMajorShapeOutput({maxTime, batchSize, outputSize});
+
+    //
+    // Permute: performed if Unidirectional Sequence Layer inputs/outputs are in batch major format.
+    //
+    if (!m_Data.m_Parameters.m_TimeMajor)
+    {
+        std::unique_ptr<arm_compute::NEPermute> layer(new arm_compute::NEPermute());
+
+        TensorInfo permuteOutInfo = inputInfo;
+        permuteOutInfo.SetShape(timeMajorShapeInput);
+        BuildArmComputeTensor(m_PermuteFirstOut, permuteOutInfo);
+        armcomputetensorutils::InitialiseArmComputeTensorEmpty(m_PermuteFirstOut);
+
+        // Permute to time major format.
+        layer->configure(&input, &m_PermuteFirstOut, arm_compute::PermutationVector(0U,2U,1U));
+        m_Permute1.reset(layer.release());
+    }
+
+    //
+    // Split and Concat Tensors
+    //
+    for (unsigned int i = 0; i < maxTime; ++i)
+    {
+        arm_compute::Tensor splitter_out;
+        arm_compute::Tensor concat_in;
+
+        auto splitterTensorInfo = inputInfo;
+        auto concatTensorInfo = outputInfo;
+        splitterTensorInfo.SetShape({batchSize, inputSize});
+        concatTensorInfo.SetShape({batchSize, outputSize});
+        BuildArmComputeTensor(splitter_out, splitterTensorInfo);
+        BuildArmComputeTensor(concat_in, concatTensorInfo);
+
+        armcomputetensorutils::InitialiseArmComputeTensorEmpty(splitter_out);
+        armcomputetensorutils::InitialiseArmComputeTensorEmpty(concat_in);
+
+        // append to std::vector<arm_compute::Tensor>
+        m_SplitterOutputsTensors.push_back(std::move(splitter_out));
+        m_ConcatInputsTensors.push_back(std::move(concat_in));
+    }
+
+    for (unsigned int i = 0; i < maxTime; ++i)
+    {
+        // append to std::vector<arm_compute::ITensor*>
+        m_SplitterOutputs.push_back(&m_SplitterOutputsTensors[i]);
+        m_ConcatInputs.push_back(&m_ConcatInputsTensors[i]);
+    }
+
+    //
+    // Split
+    //
+    unsigned int numberDimensions = 3;
+    unsigned int dimension = 0; // splitting on 0-dimension (i.e. maxTime dimension)
+
+    if (maxTime != 1) // ACL split does not work with only one element to split.
+    {
+        ViewsDescriptor splitterDesc(maxTime, numberDimensions);
+        unsigned int splitterDimSizes[3] = {1, batchSize, inputSize};
+        for (unsigned int outputIdx = 0u; outputIdx < maxTime; ++outputIdx)
+        {
+            splitterDesc.SetViewOriginCoord(outputIdx, dimension, splitterDimSizes[dimension] * outputIdx);
+            for (unsigned int dimIdx = 0u; dimIdx < numberDimensions; ++dimIdx)
+            {
+                splitterDesc.SetViewSize(outputIdx, dimIdx, splitterDimSizes[dimIdx]);
+            }
+        }
+
+        std::set<unsigned int> splitAxis = ComputeSplitAxis(splitterDesc, timeMajorShapeInput);
+
+        std::unique_ptr<arm_compute::NESplit> split_layer(new arm_compute::NESplit());
+        unsigned int                          aclAxisSplit = CalcAclAxis(splitterDesc.GetNumDimensions(),
+                                                                         *splitAxis.begin());
+        if (!m_Data.m_Parameters.m_TimeMajor)
+        {
+            split_layer->configure(&m_PermuteFirstOut, m_SplitterOutputs, aclAxisSplit);
+        } else
+        {
+            split_layer->configure(&input, m_SplitterOutputs, aclAxisSplit);
+        }
+
+        split_layer->prepare();
+        m_Splitter.reset(split_layer.release());
+    }
+
+    //
+    // Lstm
+    //
+    arm_compute::LSTMParams<arm_compute::ITensor> lstm_param;
+
+    lstm_param.set_cell_clip_params(descriptor.m_Parameters.m_ClippingThresCell);
+    lstm_param.set_projection_clip_params(descriptor.m_Parameters.m_ClippingThresProj);
+
+    lstm_param.set_matmul_scale_params(descriptor.m_Parameters.m_InputIntermediateScale,
+                                       descriptor.m_Parameters.m_ForgetIntermediateScale,
+                                       descriptor.m_Parameters.m_CellIntermediateScale,
+                                       descriptor.m_Parameters.m_OutputIntermediateScale);
+
+    lstm_param.set_hidden_state_params(descriptor.m_Parameters.m_HiddenStateZeroPoint,
+                                       descriptor.m_Parameters.m_HiddenStateScale);
+
+    m_InputToForgetWeightsTensor = std::make_unique<arm_compute::Tensor>();
+    BuildArmComputeTensor(*m_InputToForgetWeightsTensor, m_Data.m_InputToForgetWeights->GetTensorInfo());
+
+    m_InputToCellWeightsTensor = std::make_unique<arm_compute::Tensor>();
+    BuildArmComputeTensor(*m_InputToCellWeightsTensor, m_Data.m_InputToCellWeights->GetTensorInfo());
+
+    m_InputToOutputWeightsTensor = std::make_unique<arm_compute::Tensor>();
+    BuildArmComputeTensor(*m_InputToOutputWeightsTensor, m_Data.m_InputToOutputWeights->GetTensorInfo());
+
+    m_RecurrentToForgetWeightsTensor = std::make_unique<arm_compute::Tensor>();
+    BuildArmComputeTensor(*m_RecurrentToForgetWeightsTensor, m_Data.m_RecurrentToForgetWeights->GetTensorInfo());
+
+    m_RecurrentToCellWeightsTensor = std::make_unique<arm_compute::Tensor>();
+    BuildArmComputeTensor(*m_RecurrentToCellWeightsTensor, m_Data.m_RecurrentToCellWeights->GetTensorInfo());
+
+    m_RecurrentToOutputWeightsTensor = std::make_unique<arm_compute::Tensor>();
+    BuildArmComputeTensor(*m_RecurrentToOutputWeightsTensor, m_Data.m_RecurrentToOutputWeights->GetTensorInfo());
+
+    m_ForgetGateBiasTensor = std::make_unique<arm_compute::Tensor>();
+    BuildArmComputeTensor(*m_ForgetGateBiasTensor, m_Data.m_ForgetGateBias->GetTensorInfo());
+
+    m_CellBiasTensor = std::make_unique<arm_compute::Tensor>();
+    BuildArmComputeTensor(*m_CellBiasTensor, m_Data.m_CellBias->GetTensorInfo());
+
+    m_OutputGateBiasTensor = std::make_unique<arm_compute::Tensor>();
+    BuildArmComputeTensor(*m_OutputGateBiasTensor, m_Data.m_OutputGateBias->GetTensorInfo());
+
+    // for future reference: check the AndroidNN API for the logic here
+    if (!m_Data.m_Parameters.m_CifgEnabled)
+    {
+        m_InputToInputWeightsTensor = std::make_unique<arm_compute::Tensor>();
+        BuildArmComputeTensor(*m_InputToInputWeightsTensor, m_Data.m_InputToInputWeights->GetTensorInfo());
+
+        m_RecurrentToInputWeightsTensor = std::make_unique<arm_compute::Tensor>();
+        BuildArmComputeTensor(*m_RecurrentToInputWeightsTensor, m_Data.m_RecurrentToInputWeights->GetTensorInfo());
+
+        m_CellToInputWeightsTensor = std::make_unique<arm_compute::Tensor>();
+        if (m_Data.m_CellToInputWeights != nullptr)
+        {
+            BuildArmComputeTensor(*m_CellToInputWeightsTensor, m_Data.m_CellToInputWeights->GetTensorInfo());
+        }
+
+        m_InputGateBiasTensor = std::make_unique<arm_compute::Tensor>();
+        BuildArmComputeTensor(*m_InputGateBiasTensor, m_Data.m_InputGateBias->GetTensorInfo());
+        lstm_param.set_cifg_params(m_InputToInputWeightsTensor.get(),
+                                   m_RecurrentToInputWeightsTensor.get(),
+                                   m_Data.m_CellToInputWeights ? m_CellToInputWeightsTensor.get() : nullptr,
+                                   m_InputGateBiasTensor.get());
+    }
+
+    if (m_Data.m_Parameters.m_ProjectionEnabled)
+    {
+        m_ProjectionWeightsTensor = std::make_unique<arm_compute::Tensor>();
+        BuildArmComputeTensor(*m_ProjectionWeightsTensor, m_Data.m_ProjectionWeights->GetTensorInfo());
+
+        m_ProjectionBiasTensor = std::make_unique<arm_compute::Tensor>();
+        if (m_Data.m_ProjectionBias != nullptr)
+        {
+            BuildArmComputeTensor(*m_ProjectionBiasTensor, m_Data.m_ProjectionBias->GetTensorInfo());
+        }
+
+        lstm_param.set_projection_params(m_ProjectionWeightsTensor.get(),
+                                         m_Data.m_ProjectionBias ? m_ProjectionBiasTensor.get() : nullptr);
+    }
+
+    if (m_Data.m_Parameters.m_PeepholeEnabled)
+    {
+        m_CellToForgetWeightsTensor = std::make_unique<arm_compute::Tensor>();
+        BuildArmComputeTensor(*m_CellToForgetWeightsTensor, m_Data.m_CellToForgetWeights->GetTensorInfo());
+
+        m_CellToOutputWeightsTensor = std::make_unique<arm_compute::Tensor>();
+        BuildArmComputeTensor(*m_CellToOutputWeightsTensor, m_Data.m_CellToOutputWeights->GetTensorInfo());
+
+        lstm_param.set_peephole_params(m_CellToForgetWeightsTensor.get(), m_CellToOutputWeightsTensor.get());
+    }
+
+    if (m_Data.m_Parameters.m_LayerNormEnabled)
+    {
+        m_InputLayerNormWeightsTensor = std::make_unique<arm_compute::Tensor>();
+        if (!m_Data.m_Parameters.m_CifgEnabled)
+        {
+            BuildArmComputeTensor(*m_InputLayerNormWeightsTensor, m_Data.m_InputLayerNormWeights->GetTensorInfo());
+        }
+
+        m_ForgetLayerNormWeightsTensor = std::make_unique<arm_compute::Tensor>();
+        BuildArmComputeTensor(*m_ForgetLayerNormWeightsTensor, m_Data.m_ForgetLayerNormWeights->GetTensorInfo());
+
+        m_CellLayerNormWeightsTensor = std::make_unique<arm_compute::Tensor>();
+        BuildArmComputeTensor(*m_CellLayerNormWeightsTensor, m_Data.m_CellLayerNormWeights->GetTensorInfo());
+
+        m_OutputLayerNormWeightsTensor = std::make_unique<arm_compute::Tensor>();
+        BuildArmComputeTensor(*m_OutputLayerNormWeightsTensor, m_Data.m_OutputLayerNormWeights->GetTensorInfo());
+
+        auto inputNormWeightTensor = m_Data.m_Parameters.m_CifgEnabled ? nullptr : m_InputLayerNormWeightsTensor.get();
+        lstm_param.set_layer_normalization_params(inputNormWeightTensor,
+                                                  m_ForgetLayerNormWeightsTensor.get(),
+                                                  m_CellLayerNormWeightsTensor.get(),
+                                                  m_OutputLayerNormWeightsTensor.get());
+    }
+
+    for (unsigned int i = 0; i != maxTime; ++i)
+    {
+        // Set LSTM input and output ITensors depending on:
+        // input format (timeMajor) & number of LSTM batches (maxTime).
+        arm_compute::ITensor* outputLSTM;
+        arm_compute::ITensor* inputLSTM;
+
+        // If there is only one LSTM time major batch, we will not concat OR permute.
+        // Set input of LSTM to be first input ITensor.
+        // Set output of LSTM to be final output ITensor.
+        // LSTM input/output cannot be > 2 dimensions so need to resize its TensorInfo.
+        if (maxTime == 1 && m_Data.m_Parameters.m_TimeMajor)
+        {
+            TensorShape inputShape = GetTensorShape(input.info()->tensor_shape(), 1U);
+            TensorShape outputShape = GetTensorShape(output.info()->tensor_shape(), 1U);
+
+            TensorShape inputShapeShrink({inputShape[1], inputShape[2]});
+            TensorShape outputShapeShrink({outputShape[1], outputShape[2]});
+
+            auto acl_input_shape_shrink = BuildArmComputeTensorShape(inputShapeShrink);
+            auto acl_output_shape_shrink = BuildArmComputeTensorShape(outputShapeShrink);
+
+            input.info()->set_tensor_shape(acl_input_shape_shrink);
+            inputLSTM = const_cast<arm_compute::ITensor*>(&input);
+
+            output.info()->set_tensor_shape(acl_output_shape_shrink);
+            outputLSTM = &output;
+        }
+        // If there is only one LSTM batch major batch, we will not concat, only permute.
+        // Set input of LSTM to be output of initial permute.
+        // Set output of LSTM to be first element of m_ConcatInputs & use that value later in permute.
+        // LSTM output cannot be > 2 dimensions so need to resize its TensorInfo.
+        else if (maxTime == 1 && !m_Data.m_Parameters.m_TimeMajor)
+        {
+            TensorShape inputShape = GetTensorShape(m_PermuteFirstOut.info()->tensor_shape(), 1U);
+            TensorShape inputShapeShrink({inputShape[1], inputShape[2]});
+            auto acl_input_shape_shrink = BuildArmComputeTensorShape(inputShapeShrink);
+            m_PermuteFirstOut.info()->set_tensor_shape(acl_input_shape_shrink);
+            inputLSTM = &m_PermuteFirstOut;
+
+            outputLSTM = const_cast<arm_compute::ITensor*>(m_ConcatInputs[i]);
+        }
+        // Batch major AND/OR 2+ LSTM batches so will use concat AND/OR permute later on.
+        else
+        {
+            inputLSTM = m_SplitterOutputs[i];
+            outputLSTM = const_cast<arm_compute::ITensor*>(m_ConcatInputs[i]);
+        }
+
+        std::unique_ptr<arm_compute::NEQLSTMLayer> lstm_layer(new arm_compute::NEQLSTMLayer());
+
+        lstm_layer->configure(inputLSTM,
+                              m_InputToForgetWeightsTensor.get(),
+                              m_InputToCellWeightsTensor.get(),
+                              m_InputToOutputWeightsTensor.get(),
+                              m_RecurrentToForgetWeightsTensor.get(),
+                              m_RecurrentToCellWeightsTensor.get(),
+                              m_RecurrentToOutputWeightsTensor.get(),
+                              m_ForgetGateBiasTensor.get(),
+                              m_CellBiasTensor.get(),
+                              m_OutputGateBiasTensor.get(),
+                              &cellStateIn,
+                              &outputStateIn,
+                              &cellStateOut,
+                              &outputStateOut,
+                              outputLSTM,
+                              lstm_param);
+
+        m_Layers.emplace_back(std::move(lstm_layer));
+    }
+
+    InitializeArmComputeTensorData(*m_InputToForgetWeightsTensor, m_Data.m_InputToForgetWeights);
+    InitializeArmComputeTensorData(*m_InputToCellWeightsTensor, m_Data.m_InputToCellWeights);
+    InitializeArmComputeTensorData(*m_InputToOutputWeightsTensor, m_Data.m_InputToOutputWeights);
+    InitializeArmComputeTensorData(*m_RecurrentToForgetWeightsTensor, m_Data.m_RecurrentToForgetWeights);
+    InitializeArmComputeTensorData(*m_RecurrentToCellWeightsTensor, m_Data.m_RecurrentToCellWeights);
+    InitializeArmComputeTensorData(*m_RecurrentToOutputWeightsTensor, m_Data.m_RecurrentToOutputWeights);
+    InitializeArmComputeTensorData(*m_ForgetGateBiasTensor, m_Data.m_ForgetGateBias);
+    InitializeArmComputeTensorData(*m_CellBiasTensor, m_Data.m_CellBias);
+    InitializeArmComputeTensorData(*m_OutputGateBiasTensor, m_Data.m_OutputGateBias);
+
+    if (!m_Data.m_Parameters.m_CifgEnabled)
+    {
+        InitializeArmComputeTensorData(*m_InputToInputWeightsTensor, m_Data.m_InputToInputWeights);
+        InitializeArmComputeTensorData(*m_RecurrentToInputWeightsTensor, m_Data.m_RecurrentToInputWeights);
+        if (m_Data.m_CellToInputWeights != nullptr)
+        {
+            InitializeArmComputeTensorData(*m_CellToInputWeightsTensor, m_Data.m_CellToInputWeights);
+        }
+        InitializeArmComputeTensorData(*m_InputGateBiasTensor, m_Data.m_InputGateBias);
+    }
+
+    if (m_Data.m_Parameters.m_ProjectionEnabled)
+    {
+        InitializeArmComputeTensorData(*m_ProjectionWeightsTensor, m_Data.m_ProjectionWeights);
+        if (m_Data.m_ProjectionBias != nullptr)
+        {
+            InitializeArmComputeTensorData(*m_ProjectionBiasTensor, m_Data.m_ProjectionBias);
+        }
+    }
+
+    if (m_Data.m_Parameters.m_PeepholeEnabled)
+    {
+        InitializeArmComputeTensorData(*m_CellToForgetWeightsTensor, m_Data.m_CellToForgetWeights);
+        InitializeArmComputeTensorData(*m_CellToOutputWeightsTensor, m_Data.m_CellToOutputWeights);
+    }
+
+    if (m_Data.m_Parameters.m_LayerNormEnabled)
+    {
+        if (!m_Data.m_Parameters.m_CifgEnabled)
+        {
+            InitializeArmComputeTensorData(*m_InputLayerNormWeightsTensor, m_Data.m_InputLayerNormWeights);
+        }
+        InitializeArmComputeTensorData(*m_ForgetLayerNormWeightsTensor, m_Data.m_ForgetLayerNormWeights);
+        InitializeArmComputeTensorData(*m_CellLayerNormWeightsTensor, m_Data.m_CellLayerNormWeights);
+        InitializeArmComputeTensorData(*m_OutputLayerNormWeightsTensor, m_Data.m_OutputLayerNormWeights);
+    }
+
+    // Force Compute Library to perform the necessary copying and reshaping.
+    // After which delete all the input tensors that will no longer be needed.
+    for (uint32_t i = 0; i < m_Layers.size(); ++i)
+    {
+        m_Layers[i]->prepare();
+    }
+
+    //
+    // Concat
+    //
+
+    // Expand dimensions of LSTM outputs adding one empty dimension to fit concatenate inputs.
+    TensorShape shape = GetTensorShape(m_ConcatInputs[0]->info()->tensor_shape(), 1U);
+    TensorShape shapeExpandTimeMajor({1, shape[0], shape[1]});
+    TensorShape shapeExpandBatchMajor({shape[0], 1, shape[1]});
+
+    if (maxTime != 1) // ACL concat does not work with only one element to concatenate.
+    {
+        for (unsigned int i = 0; i < maxTime; ++i)
+        {
+            m_ConcatInputs[i]->info()->set_tensor_shape(BuildArmComputeTensorShape(shapeExpandTimeMajor));
+        }
+        ConcatDescriptor  concatDescriptor(maxTime, numberDimensions);  // maxTime = num inputs (aka. number of views).
+
+        for (unsigned int inputIdx = 0u; inputIdx < maxTime; ++inputIdx)
+        {
+            concatDescriptor.SetViewOriginCoord(inputIdx, dimension, inputIdx);
+            concatDescriptor.SetConcatAxis(dimension);
+        }
+        m_Concat.reset(new arm_compute::NEConcatenateLayer());
+
+        unsigned int aclAxisConcat = CalcAclAxis(concatDescriptor.GetNumDimensions(), concatDescriptor.GetConcatAxis());
+        if (!m_Data.m_Parameters.m_TimeMajor)
+        {
+            TensorInfo concatOutputTensorInfo = outputInfo;
+            concatOutputTensorInfo.SetShape(timeMajorShapeOutput);
+            BuildArmComputeTensor(concat_out, concatOutputTensorInfo);
+            armcomputetensorutils::InitialiseArmComputeTensorEmpty(concat_out);
+
+            m_Concat->configure(m_ConcatInputs, &concat_out, aclAxisConcat);
+        }
+        else
+        {
+            m_Concat->configure(m_ConcatInputs, &output, aclAxisConcat);
+        }
+
+        m_Concat->prepare();
+    }
+    // If only one LSTM batch, we do not concat and/or permute.
+    // Must ensure final output info is expanded to correct batch major dimensions.
+    else
+    {
+        if (!m_Data.m_Parameters.m_TimeMajor)
+        {
+            output.info()->set_tensor_shape(BuildArmComputeTensorShape(shapeExpandBatchMajor));
+        }
+        else
+        {
+            output.info()->set_tensor_shape(BuildArmComputeTensorShape(shapeExpandTimeMajor));
+        }
+    }
+
+    //
+    // Permute: only done if input/output are in batch major format.
+    //
+    if (!m_Data.m_Parameters.m_TimeMajor)
+    {
+        // Output now time major. Permute output back to batch major.
+        std::unique_ptr<arm_compute::NEPermute> layer(new arm_compute::NEPermute());
+        if (maxTime != 1)
+        {
+            layer->configure(&concat_out, &output, arm_compute::PermutationVector(0U, 2U, 1U));
+        }
+        else
+        {
+            layer->configure(m_ConcatInputs[0], &output, arm_compute::PermutationVector(0U, 2U, 1U));
+        }
+        m_Permute2.reset(layer.release());
+    }
+
+    FreeUnusedTensors();
+}
+
+void NeonUnidirectionalSequenceLstmWorkload::Execute() const
+{
+    ARMNN_SCOPED_PROFILING_EVENT_NEON_GUID("NeonUnidirectionalSequenceLstmWorkload_Execute", GetGuid());
+    if (m_Permute1)
+    {
+        m_Permute1->run();
+    }
+    if (m_Splitter)
+    {
+        m_Splitter->run();
+    }
+    for (uint32_t i = 0; i < m_Layers.size(); ++i)
+    {
+        m_Layers[i]->run();
+    }
+    if (m_Concat)
+    {
+        m_Concat->run();
+    }
+    if (m_Permute2)
+    {
+        m_Permute2->run();
+    }
+}
+
+arm_compute::Status
+NeonUnidirectionalSequenceLstmWorkloadValidate(const TensorInfo& input,
+                                               const TensorInfo& outputStateIn,
+                                               const TensorInfo& cellStateIn,
+                                               const TensorInfo& outputStateOut,
+                                               const TensorInfo& cellStateOut,
+                                               const TensorInfo& output,
+                                               const UnidirectionalSequenceLstmDescriptor& descriptor,
+                                               const LstmInputParamsInfo& paramsInfo)
+{
+    TensorShape inputLayerShape = input.GetShape();
+    TensorShape outputLayerShape = output.GetShape();
+
+    unsigned int maxTime = descriptor.m_TimeMajor ? inputLayerShape[0] : inputLayerShape[1];
+    unsigned int batchSize = descriptor.m_TimeMajor ? inputLayerShape[1] : inputLayerShape[0];
+    unsigned int inputSize = inputLayerShape[2];
+    unsigned int outputSize = outputLayerShape[2];
+
+    const TensorShape timeMajorShapeInput({maxTime, batchSize, inputSize});
+    const TensorShape timeMajorShapeOutput({maxTime, batchSize, outputSize});
+
+    arm_compute::Status statusPermute1 = arm_compute::Status(arm_compute::ErrorCode::OK,
+                                                             "Permute1 status");
+    arm_compute::Status statusSplit = arm_compute::Status(arm_compute::ErrorCode::OK,
+                                                          "Split status");
+    arm_compute::Status statusLSTM = arm_compute::Status(arm_compute::ErrorCode::OK,
+                                                         "LSTM status");
+    arm_compute::Status statusConcat = arm_compute::Status(arm_compute::ErrorCode::OK,
+                                                           "Concat status");
+    arm_compute::Status statusPermute2 = arm_compute::Status(arm_compute::ErrorCode::OK,
+                                                             "Permute2 status");
+
+    const arm_compute::TensorInfo aclInputInfo  = armcomputetensorutils::BuildArmComputeTensorInfo(input);
+    const arm_compute::TensorInfo aclOutputInfo  = armcomputetensorutils::BuildArmComputeTensorInfo(output);
+
+    //
+    // Permute validate
+    //
+    TensorInfo permuteOutInfo = TensorInfo(input);
+    arm_compute::TensorInfo aclPermuteOutInfo = armcomputetensorutils::BuildArmComputeTensorInfo(permuteOutInfo);
+    if (!descriptor.m_TimeMajor)
+    {
+        statusPermute1 =  arm_compute::NEPermute::validate(&aclInputInfo,
+                                                           &aclPermuteOutInfo,
+                                                           arm_compute::PermutationVector(0U, 2U, 1U));
+    }
+
+    //
+    // Split and Concat Tensors validate
+    //
+    std::vector<arm_compute::TensorInfo> splitterOutputsTensorInfos;
+    std::vector<arm_compute::TensorInfo> concatInputsTensorInfos;
+    std::vector<arm_compute::ITensorInfo*> splitterOutputsTensorInfosPtr;
+    std::vector<const arm_compute::ITensorInfo*> concatInputsTensorInfosPtr;
+    splitterOutputsTensorInfos.reserve(maxTime);
+    concatInputsTensorInfos.reserve(maxTime);
+    for (unsigned int i = 0; i < maxTime; ++i)
+    {
+        arm_compute::TensorInfo splitter_out;
+        arm_compute::TensorInfo concat_in;
+
+        auto splitterTensorInfo = TensorInfo(input);
+        auto concatTensorInfo   = TensorInfo(output);
+        splitterTensorInfo.SetShape({batchSize, inputSize});
+        concatTensorInfo.SetShape({batchSize, outputSize});
+
+        arm_compute::TensorInfo aclSplitterTensorInfo
+            = armcomputetensorutils::BuildArmComputeTensorInfo(splitterTensorInfo);
+        arm_compute::TensorInfo aclConcatTensorInfo
+            = armcomputetensorutils::BuildArmComputeTensorInfo(concatTensorInfo);
+
+        splitterOutputsTensorInfos.emplace_back(aclSplitterTensorInfo);
+        concatInputsTensorInfos.emplace_back(aclConcatTensorInfo);
+        splitterOutputsTensorInfosPtr.emplace_back(&splitterOutputsTensorInfos[i]);
+        concatInputsTensorInfosPtr.emplace_back(&concatInputsTensorInfos[i]);
+    }
+
+    //
+    // Split validate
+    //
+    unsigned int numberDimensions = 3;
+    unsigned int dimension = 0; // splitting on 0-dimension (i.e. maxTime dimension)
+    unsigned int aclAxisSplit = CalcAclAxis(numberDimensions, dimension);
+
+    if (maxTime != 1) // ACL split does not work with only one element to split.
+    {
+        if (!descriptor.m_TimeMajor)
+        {
+            statusSplit = arm_compute::NESplit::validate(&aclPermuteOutInfo,
+                                                         splitterOutputsTensorInfosPtr,
+                                                         aclAxisSplit);
+        } else
+        {
+            statusSplit = arm_compute::NESplit::validate(&aclInputInfo, splitterOutputsTensorInfosPtr, aclAxisSplit);
+        }
+    }
+
+    //
+    // LSTM validate
+    //
+
+    arm_compute::LSTMParams<arm_compute::ITensorInfo> lstm_params_info;
+
+    const TensorInfo& scratchBuffer = TensorInfo(cellStateIn.GetShape(), input.GetDataType());
+
+    lstm_params_info.set_cell_clip_params(descriptor.m_ClippingThresCell);
+    lstm_params_info.set_projection_clip_params(descriptor.m_ClippingThresProj);
+    // The inputs and outputs
+    const arm_compute::TensorInfo aclOutputStateInInfo = BuildArmComputeTensorInfo(outputStateIn);
+    const arm_compute::TensorInfo aclCellStateInInfo = BuildArmComputeTensorInfo(cellStateIn);
+    const arm_compute::TensorInfo aclScratchBufferInfo = BuildArmComputeTensorInfo(scratchBuffer);
+    const arm_compute::TensorInfo aclOutputStateOutInfo = BuildArmComputeTensorInfo(outputStateOut);
+    const arm_compute::TensorInfo aclCellStateOutInfo = BuildArmComputeTensorInfo(cellStateOut);
+
+    // Basic parameters
+    const arm_compute::TensorInfo aclInputToForgetWeightsInfo
+                                      = BuildArmComputeTensorInfo(paramsInfo.GetInputToForgetWeights());
+    const arm_compute::TensorInfo aclInputToCellWeightsInfo
+                                      = BuildArmComputeTensorInfo(paramsInfo.GetInputToCellWeights());
+    const arm_compute::TensorInfo aclInputToOutputWeightsInfo
+                                      = BuildArmComputeTensorInfo(paramsInfo.GetInputToOutputWeights());
+    const arm_compute::TensorInfo aclRecurrentToForgetWeightsInfo
+                                      = BuildArmComputeTensorInfo(paramsInfo.GetRecurrentToForgetWeights());
+    const arm_compute::TensorInfo aclRecurrentToCellWeightsInfo
+                                      = BuildArmComputeTensorInfo(paramsInfo.GetRecurrentToCellWeights());
+    const arm_compute::TensorInfo aclRecurrentToOutputWeightsInfo
+                                      = BuildArmComputeTensorInfo(paramsInfo.GetRecurrentToOutputWeights());
+    const arm_compute::TensorInfo aclForgetGateBiasInfo
+                                      = BuildArmComputeTensorInfo(paramsInfo.GetForgetGateBias());
+    const arm_compute::TensorInfo aclCellBiasInfo
+                                      = BuildArmComputeTensorInfo(paramsInfo.GetCellBias());
+    const arm_compute::TensorInfo aclOutputGateBiasInfo
+                                      = BuildArmComputeTensorInfo(paramsInfo.GetOutputGateBias());
+
+    arm_compute::TensorInfo aclInputToInputWeightsInfo;
+    arm_compute::TensorInfo aclRecurrentToInputWeightsInfo;
+    arm_compute::TensorInfo aclCellToInputWeightsInfo;
+    arm_compute::TensorInfo aclInputGateBiasInfo;
+    arm_compute::TensorInfo aclProjectionWeightsInfo;
+    arm_compute::TensorInfo aclProjectionBiasInfo;
+    arm_compute::TensorInfo aclCellToForgetWeightsInfo;
+    arm_compute::TensorInfo aclCellToOutputWeightsInfo;
+
+    arm_compute::TensorInfo aclInputLayerNormWeightsInfo;
+    arm_compute::TensorInfo aclForgetLayerNormWeightsInfo;
+    arm_compute::TensorInfo aclCellLayerNormWeightsInfo;
+    arm_compute::TensorInfo aclOutputLayerNormWeightsInfo;
+
+    if (!descriptor.m_CifgEnabled)
+    {
+        if (descriptor.m_PeepholeEnabled)
+        {
+            aclCellToInputWeightsInfo = BuildArmComputeTensorInfo(paramsInfo.GetCellToInputWeights());
+        }
+        aclInputToInputWeightsInfo = BuildArmComputeTensorInfo(paramsInfo.GetInputToInputWeights());
+        aclRecurrentToInputWeightsInfo = BuildArmComputeTensorInfo(paramsInfo.GetRecurrentToInputWeights());
+        aclInputGateBiasInfo = BuildArmComputeTensorInfo(paramsInfo.GetInputGateBias());
+
+        lstm_params_info.set_cifg_params(&aclInputToInputWeightsInfo,
+                                         &aclRecurrentToInputWeightsInfo,
+                                         descriptor.m_PeepholeEnabled ? &aclCellToInputWeightsInfo : nullptr,
+                                         &aclInputGateBiasInfo);
+    }
+
+    if (descriptor.m_ProjectionEnabled)
+    {
+        if (paramsInfo.m_ProjectionBias != nullptr)
+        {
+            aclProjectionBiasInfo = BuildArmComputeTensorInfo(paramsInfo.GetProjectionBias());
+        }
+        aclProjectionWeightsInfo = BuildArmComputeTensorInfo(paramsInfo.GetProjectionWeights());
+
+        lstm_params_info.set_projection_params(&aclProjectionWeightsInfo,
+                                               paramsInfo.m_ProjectionBias ? &aclProjectionBiasInfo : nullptr);
+    }
+
+    if (descriptor.m_PeepholeEnabled)
+    {
+        aclCellToForgetWeightsInfo = BuildArmComputeTensorInfo(paramsInfo.GetCellToForgetWeights());
+        aclCellToOutputWeightsInfo = BuildArmComputeTensorInfo(paramsInfo.GetCellToOutputWeights());
+
+        lstm_params_info.set_peephole_params(&aclCellToForgetWeightsInfo, &aclCellToOutputWeightsInfo);
+    }
+
+    if (descriptor.m_LayerNormEnabled)
+    {
+        if (!descriptor.m_CifgEnabled)
+        {
+            aclInputLayerNormWeightsInfo = BuildArmComputeTensorInfo(paramsInfo.GetInputLayerNormWeights());
+        }
+        aclForgetLayerNormWeightsInfo = BuildArmComputeTensorInfo(paramsInfo.GetForgetLayerNormWeights());
+        aclCellLayerNormWeightsInfo = BuildArmComputeTensorInfo(paramsInfo.GetCellLayerNormWeights());
+        aclOutputLayerNormWeightsInfo = BuildArmComputeTensorInfo(paramsInfo.GetOutputLayerNormWeights());
+
+        lstm_params_info.set_layer_normalization_params(descriptor.m_CifgEnabled ? nullptr :
+                                                        &aclInputLayerNormWeightsInfo,
+                                                        &aclForgetLayerNormWeightsInfo,
+                                                        &aclCellLayerNormWeightsInfo,
+                                                        &aclOutputLayerNormWeightsInfo);
+    }
+
+    lstm_params_info.set_matmul_scale_params(descriptor.m_InputIntermediateScale,
+                                             descriptor.m_ForgetIntermediateScale,
+                                             descriptor.m_CellIntermediateScale,
+                                             descriptor.m_OutputIntermediateScale);
+
+    lstm_params_info.set_hidden_state_params(descriptor.m_HiddenStateZeroPoint, descriptor.m_HiddenStateScale);
+
+    for (unsigned int i = 0; i != maxTime; ++i)
+    {
+
+        // Set LSTM input and output ITensors depending on:
+        // input format (timeMajor) & number of LSTM batches (maxTime).
+        arm_compute::ITensorInfo* outputLSTM;
+        arm_compute::ITensorInfo* inputLSTM;
+
+        // If there is only one LSTM time major batch, we will not concat OR permute.
+        // Set input of LSTM to be first input ITensor.
+        // Set output of LSTM to be final output ITensor.
+        // LSTM input/output cannot be > 2 dimensions so need to resize its TensorInfo.
+        if (maxTime == 1 && !descriptor.m_TimeMajor)
+        {
+            TensorShape inputShape = GetTensorShape(aclInputInfo.tensor_shape(), 1U);
+            TensorShape outputShape = GetTensorShape(aclOutputInfo.tensor_shape(), 1U);
+
+            TensorShape inputShapeShrink({inputShape[1], inputShape[2]});
+            TensorShape outputShapeShrink({outputShape[1], outputShape[2]});
+
+            auto acl_input_shape_shrink = BuildArmComputeTensorShape(inputShapeShrink);
+            auto acl_output_shape_shrink = BuildArmComputeTensorShape(outputShapeShrink);
+
+            const_cast<arm_compute::TensorInfo*>(&aclInputInfo)->set_tensor_shape(acl_input_shape_shrink);
+            inputLSTM = const_cast<arm_compute::TensorInfo*>(&aclInputInfo);
+
+            const_cast<arm_compute::TensorInfo*>(&aclOutputInfo)->set_tensor_shape(acl_output_shape_shrink);
+            outputLSTM = const_cast<arm_compute::TensorInfo*>(&aclOutputInfo);
+        }
+        // If there is only one LSTM batch major batch, we will not concat, only permute.
+        // Set input of LSTM to be output of initial permute.
+        // Set output of LSTM to be first element of m_ConcatInputs & use that value later in permute.
+        // LSTM output cannot be > 2 dimensions so need to resize its TensorInfo.
+        else if (maxTime == 1 && !descriptor.m_TimeMajor)
+        {
+            TensorShape inputShape = GetTensorShape(aclPermuteOutInfo.tensor_shape(), 1U);
+            TensorShape inputShapeShrink({inputShape[1], inputShape[2]});
+            auto acl_input_shape_shrink = BuildArmComputeTensorShape(inputShapeShrink);
+            aclPermuteOutInfo.set_tensor_shape(acl_input_shape_shrink);
+            inputLSTM = &aclPermuteOutInfo;
+
+            outputLSTM = const_cast<arm_compute::ITensorInfo*>(concatInputsTensorInfosPtr[i]);
+        }
+        // Batch major AND/OR 2+ LSTM batches so will use concat AND/OR permute later on.
+        else
+        {
+            inputLSTM = splitterOutputsTensorInfosPtr[i];
+            outputLSTM = const_cast<arm_compute::ITensorInfo*>(concatInputsTensorInfosPtr[i]);
+        }
+
+        statusLSTM = arm_compute::NEQLSTMLayer::validate(inputLSTM,
+                                                         &aclInputToForgetWeightsInfo,
+                                                         &aclInputToCellWeightsInfo,
+                                                         &aclInputToOutputWeightsInfo,
+                                                         &aclRecurrentToForgetWeightsInfo,
+                                                         &aclRecurrentToCellWeightsInfo,
+                                                         &aclRecurrentToOutputWeightsInfo,
+                                                         &aclForgetGateBiasInfo,
+                                                         &aclCellBiasInfo,
+                                                         &aclOutputGateBiasInfo,
+                                                         &aclCellStateInInfo,
+                                                         &aclOutputStateInInfo,
+                                                         &aclCellStateOutInfo,
+                                                         &aclOutputStateOutInfo,
+                                                         outputLSTM,
+                                                         lstm_params_info);
+    }
+
+    //
+    // Concat validate
+    //
+
+    // Expand dimensions of LSTM outputs adding one empty dimension to fit concatenate inputs.
+    TensorShape shape = GetTensorShape(concatInputsTensorInfosPtr[0]->tensor_shape(), 1U);
+    TensorShape shapeExpandTimeMajor({1, shape[0], shape[1]});
+    TensorShape shapeExpandBatchMajor({shape[0], 1, shape[1]});
+
+    TensorInfo concatOutputTensorInfo = TensorInfo(output);
+    concatOutputTensorInfo.SetShape(timeMajorShapeOutput);
+    arm_compute::TensorInfo aclConcatOutputTensorInfo= BuildArmComputeTensorInfo(concatOutputTensorInfo);
+
+    if (maxTime != 1) // ACL concat does not work with only one element to concatenate.
+    {
+        for (unsigned int i = 0; i < maxTime; ++i)
+        {
+            auto acl_shape_expand = BuildArmComputeTensorShape(shapeExpandTimeMajor);
+            concatInputsTensorInfos[i].set_tensor_shape(acl_shape_expand);
+        }
+
+        unsigned int aclAxisConcat = CalcAclAxis(numberDimensions, dimension);
+        if (!descriptor.m_TimeMajor)
+        {
+            statusConcat = arm_compute::NEConcatenateLayer::validate(concatInputsTensorInfosPtr,
+                                                                     &aclConcatOutputTensorInfo,
+                                                                     aclAxisConcat);
+        }
+        else
+        {
+            statusConcat = arm_compute::NEConcatenateLayer::validate(concatInputsTensorInfosPtr,
+                                                                     &aclOutputInfo,
+                                                                     aclAxisConcat);
+        }
+    }
+    // If only one LSTM batch, we do not concat and/or permute.
+    // Must ensure final output info is expanded to correct batch major dimensions.
+    else
+    {
+        if (!descriptor.m_TimeMajor)
+        {
+            const_cast<arm_compute::TensorInfo*>(&aclInputInfo)->set_tensor_shape(
+                BuildArmComputeTensorShape(shapeExpandBatchMajor));
+        }
+        else
+        {
+            const_cast<arm_compute::TensorInfo*>(&aclInputInfo)->set_tensor_shape(
+                BuildArmComputeTensorShape(shapeExpandTimeMajor));
+        }
+    }
+
+    //
+    // Permute validate
+    //
+    if (!descriptor.m_TimeMajor)
+    {
+        // Output now time major. Permute output back to batch major.
+        if (maxTime != 1)
+        {
+            statusPermute2 = arm_compute::NEPermute::validate(&aclConcatOutputTensorInfo,
+                                                              &aclOutputInfo,
+                                                              arm_compute::PermutationVector(0U, 2U, 1U));
+        }
+        else
+        {
+            statusPermute2 = arm_compute::NEPermute::validate(concatInputsTensorInfosPtr[0],
+                                                              &aclOutputInfo,
+                                                              arm_compute::PermutationVector(0U, 2U, 1U));
+        }
+    }
+
+    auto okCode = arm_compute::ErrorCode::OK;
+    if (statusPermute1.error_code() == okCode &&
+        statusSplit.error_code()    == okCode &&
+        statusLSTM .error_code()    == okCode &&
+        statusConcat.error_code()   == okCode &&
+        statusPermute2.error_code() == okCode)
+    {
+        return arm_compute::Status(arm_compute::ErrorCode::OK,
+                                   "All Unidirectional Sequence LSTM layer validate status OK.");
+    }
+    else
+    {
+        return arm_compute::Status(arm_compute::ErrorCode::RUNTIME_ERROR,
+                                   "Unidirectional Sequence LSTM layer validate status failed.");
+    }
+}
+
+void NeonUnidirectionalSequenceLstmWorkload::FreeUnusedTensors()
+{
+    FreeTensorIfUnused(m_InputToInputWeightsTensor);
+    FreeTensorIfUnused(m_InputToForgetWeightsTensor);
+    FreeTensorIfUnused(m_InputToCellWeightsTensor);
+    FreeTensorIfUnused(m_InputToOutputWeightsTensor);
+    FreeTensorIfUnused(m_RecurrentToInputWeightsTensor);
+    FreeTensorIfUnused(m_RecurrentToForgetWeightsTensor);
+    FreeTensorIfUnused(m_RecurrentToCellWeightsTensor);
+    FreeTensorIfUnused(m_RecurrentToOutputWeightsTensor);
+    FreeTensorIfUnused(m_CellToInputWeightsTensor);
+    FreeTensorIfUnused(m_CellToForgetWeightsTensor);
+    FreeTensorIfUnused(m_CellToOutputWeightsTensor);
+    FreeTensorIfUnused(m_InputGateBiasTensor);
+    FreeTensorIfUnused(m_ForgetGateBiasTensor);
+    FreeTensorIfUnused(m_CellBiasTensor);
+    FreeTensorIfUnused(m_OutputGateBiasTensor);
+    FreeTensorIfUnused(m_ProjectionWeightsTensor);
+    FreeTensorIfUnused(m_ProjectionBiasTensor);
+    FreeTensorIfUnused(m_InputLayerNormWeightsTensor);
+    FreeTensorIfUnused(m_ForgetLayerNormWeightsTensor);
+    FreeTensorIfUnused(m_CellLayerNormWeightsTensor);
+    FreeTensorIfUnused(m_OutputLayerNormWeightsTensor);
+}
+
+} //namespace armnn
diff --git a/src/backends/neon/workloads/NeonUnidirectionalSequenceLstmWorkload.hpp b/src/backends/neon/workloads/NeonUnidirectionalSequenceLstmWorkload.hpp
new file mode 100644
index 0000000..f012258
--- /dev/null
+++ b/src/backends/neon/workloads/NeonUnidirectionalSequenceLstmWorkload.hpp
@@ -0,0 +1,90 @@
+//
+// Copyright © 2022 Arm Ltd and Contributors. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+
+#pragma once
+
+#include <armnn/Descriptors.hpp>
+#include <armnn/LstmParams.hpp>
+#include <armnn/backends/Workload.hpp>
+#include <armnn/backends/WorkloadData.hpp>
+#include "NeonBaseWorkload.hpp"
+
+#include "arm_compute/runtime/NEON/functions/NEQLSTMLayer.h"
+#include "arm_compute/runtime/NEON/functions/NEPermute.h"
+#include "arm_compute/runtime/NEON/functions/NESplit.h"
+#include "arm_compute/runtime/NEON/functions/NEConcatenateLayer.h"
+
+namespace armnn
+{
+
+class NeonUnidirectionalSequenceLstmWorkload : public NeonBaseWorkload<UnidirectionalSequenceLstmQueueDescriptor>
+{
+public:
+    NeonUnidirectionalSequenceLstmWorkload(const UnidirectionalSequenceLstmQueueDescriptor& descriptor,
+                                           const WorkloadInfo& info);
+    virtual void Execute() const override;
+
+private:
+
+    //
+    // ACL layers required to fully form a Unidirectional Sequence LSTM layer.
+    //
+    mutable std::unique_ptr<arm_compute::NEPermute> m_Permute1;
+    mutable std::unique_ptr<arm_compute::IFunction> m_Splitter;
+    mutable std::vector<std::unique_ptr<arm_compute::NEQLSTMLayer>> m_Layers;
+    mutable std::unique_ptr<arm_compute::NEConcatenateLayer> m_Concat;
+    mutable std::unique_ptr<arm_compute::NEPermute> m_Permute2;
+
+    //
+    // ACL LSTM arm_compute::Tensors.
+    //
+    std::unique_ptr<arm_compute::Tensor> m_InputToInputWeightsTensor;
+    std::unique_ptr<arm_compute::Tensor> m_InputToForgetWeightsTensor;
+    std::unique_ptr<arm_compute::Tensor> m_InputToCellWeightsTensor;
+    std::unique_ptr<arm_compute::Tensor> m_InputToOutputWeightsTensor;
+    std::unique_ptr<arm_compute::Tensor> m_RecurrentToInputWeightsTensor;
+    std::unique_ptr<arm_compute::Tensor> m_RecurrentToForgetWeightsTensor;
+    std::unique_ptr<arm_compute::Tensor> m_RecurrentToCellWeightsTensor;
+    std::unique_ptr<arm_compute::Tensor> m_RecurrentToOutputWeightsTensor;
+    std::unique_ptr<arm_compute::Tensor> m_CellToInputWeightsTensor;
+    std::unique_ptr<arm_compute::Tensor> m_CellToForgetWeightsTensor;
+    std::unique_ptr<arm_compute::Tensor> m_CellToOutputWeightsTensor;
+    std::unique_ptr<arm_compute::Tensor> m_InputGateBiasTensor;
+    std::unique_ptr<arm_compute::Tensor> m_ForgetGateBiasTensor;
+    std::unique_ptr<arm_compute::Tensor> m_CellBiasTensor;
+    std::unique_ptr<arm_compute::Tensor> m_OutputGateBiasTensor;
+    std::unique_ptr<arm_compute::Tensor> m_ProjectionWeightsTensor;
+    std::unique_ptr<arm_compute::Tensor> m_ProjectionBiasTensor;
+
+    std::unique_ptr<arm_compute::Tensor> m_InputLayerNormWeightsTensor;
+    std::unique_ptr<arm_compute::Tensor> m_ForgetLayerNormWeightsTensor;
+    std::unique_ptr<arm_compute::Tensor> m_CellLayerNormWeightsTensor;
+    std::unique_ptr<arm_compute::Tensor> m_OutputLayerNormWeightsTensor;
+
+    //
+    // Additional ACL arm_compute::Tensors and std::vector<arm_compute::Tensor>.
+    // Required to perform splitting, concatenation and permutations.
+    //
+    arm_compute::Tensor m_PermuteFirstOut;
+    std::vector<arm_compute::Tensor> m_SplitterOutputsTensors;
+    std::vector<arm_compute::Tensor> m_ConcatInputsTensors;
+    std::vector<arm_compute::ITensor*> m_SplitterOutputs;
+    std::vector<const arm_compute::ITensor*> m_ConcatInputs;
+    arm_compute::Tensor concat_out;
+
+    void FreeUnusedTensors();
+};
+
+arm_compute::Status
+NeonUnidirectionalSequenceLstmWorkloadValidate(const TensorInfo& input,
+                                               const TensorInfo& outputStateIn,
+                                               const TensorInfo& cellStateIn,
+                                               const TensorInfo& outputStateOut,
+                                               const TensorInfo& cellStateOut,
+                                               const TensorInfo& output,
+                                               const UnidirectionalSequenceLstmDescriptor& descriptor,
+                                               const LstmInputParamsInfo& paramsInfo);
+
+} //namespace armnn
diff --git a/src/backends/neon/workloads/NeonWorkloads.hpp b/src/backends/neon/workloads/NeonWorkloads.hpp
index 4f5ba2d..8b99f03 100644
--- a/src/backends/neon/workloads/NeonWorkloads.hpp
+++ b/src/backends/neon/workloads/NeonWorkloads.hpp
@@ -69,3 +69,4 @@
 #include "NeonTransposeConvolution2dWorkload.hpp"
 #include "NeonTransposeWorkload.hpp"
 #include "NeonUnidirectionalSequenceLstmFloatWorkload.hpp"
+#include "NeonUnidirectionalSequenceLstmWorkload.hpp"
\ No newline at end of file
diff --git a/src/backends/reference/RefLayerSupport.cpp b/src/backends/reference/RefLayerSupport.cpp
index 66661cb..919c6db 100644
--- a/src/backends/reference/RefLayerSupport.cpp
+++ b/src/backends/reference/RefLayerSupport.cpp
@@ -465,57 +465,15 @@
                                                "hiddenStateOutputVal, cellStateOutputVal, output}");
             }
             auto desc = *(PolymorphicDowncast<const UnidirectionalSequenceLstmDescriptor*>(&descriptor));
-
-            bool isHiddenStateOutputOptional = (infos[4] == TensorInfo());
-            bool isCellStateOutput = (infos[5] == TensorInfo());
-            if (isHiddenStateOutputOptional && isCellStateOutput)
-            {
-                return IsUnidirectionalSequenceLstmSupported(infos[0],
-                                                             infos[1],
-                                                             infos[2],
-                                                             infos[3],
-                                                             EmptyOptional(),
-                                                             EmptyOptional(),
-                                                             desc,
-                                                             lstmParamsInfo.value(),
-                                                             reasonIfUnsupported);
-            }
-            else if (isHiddenStateOutputOptional)
-            {
-                return IsUnidirectionalSequenceLstmSupported(infos[0],
-                                                             infos[1],
-                                                             infos[2],
-                                                             infos[3],
-                                                             EmptyOptional(),
-                                                             infos[5],
-                                                             desc,
-                                                             lstmParamsInfo.value(),
-                                                             reasonIfUnsupported);
-            }
-            else if (isCellStateOutput)
-            {
-                return IsUnidirectionalSequenceLstmSupported(infos[0],
-                                                             infos[1],
-                                                             infos[2],
-                                                             infos[3],
-                                                             infos[4],
-                                                             EmptyOptional(),
-                                                             desc,
-                                                             lstmParamsInfo.value(),
-                                                             reasonIfUnsupported);
-            }
-            else
-            {
-                return IsUnidirectionalSequenceLstmSupported(infos[0],
-                                                             infos[1],
-                                                             infos[2],
-                                                             infos[3],
-                                                             infos[4],
-                                                             infos[5],
-                                                             desc,
-                                                             lstmParamsInfo.value(),
-                                                             reasonIfUnsupported);
-            }
+            return IsUnidirectionalSequenceLstmSupported(infos[0],
+                                                         infos[1],
+                                                         infos[2],
+                                                         infos[3],
+                                                         infos[4],
+                                                         infos[5],
+                                                         desc,
+                                                         lstmParamsInfo.value(),
+                                                         reasonIfUnsupported);
         }
         case LayerType::Pooling3d:
             return IsPooling3dSupported(infos[0],
@@ -2841,9 +2799,9 @@
         const TensorInfo& input,
         const TensorInfo& outputStateIn,
         const TensorInfo& cellStateIn,
+        const TensorInfo& outputStateOut,
+        const TensorInfo& cellStateOut,
         const TensorInfo& output,
-        const Optional<TensorInfo>& hiddenStateOutput,
-        const Optional<TensorInfo>& cellStateOutput,
         const UnidirectionalSequenceLstmDescriptor& descriptor,
         const LstmInputParamsInfo& paramsInfo,
         Optional<std::string&> reasonIfUnsupported) const
@@ -2852,17 +2810,14 @@
     IgnoreUnused(paramsInfo);
     IgnoreUnused(outputStateIn);
     IgnoreUnused(cellStateIn);
+    IgnoreUnused(outputStateOut);
+    IgnoreUnused(cellStateOut);
     bool supported = true;
 
-    if (hiddenStateOutput.has_value() || cellStateOutput.has_value())
+    std::array<DataType, 2> supportedTypes =
     {
-        reasonIfUnsupported.value() += "Reference UnidirectionalSequenceLstm: hidden state output "
-                                       "and cell state output are not supported at the moment.";
-    }
-
-    std::array<DataType, 1> supportedTypes =
-    {
-        DataType::Float32
+        DataType::Float32,
+        DataType::QAsymmS8
     };
 
     std::array<DataType, 2> supportedWeightTypes =
@@ -2871,16 +2826,19 @@
         DataType::QAsymmS8
     };
 
+    std::array<DataType, 3> supportedBiasTypes =
+    {
+        DataType::Float32,
+        DataType::QAsymmS8,
+        DataType::Signed32
+    };
+
     // check inputs and outputs
     supported &= CheckSupportRule(TypeAnyOf(input, supportedTypes), reasonIfUnsupported,
                                   "Reference UnidirectionalSequenceLstm: input is not a supported type.");
-    supported &= CheckSupportRule(TypesAreEqual(input, outputStateIn), reasonIfUnsupported,
-                                  "Reference UnidirectionalSequenceLstm: input and outputStateIn types are mismatched");
-    supported &= CheckSupportRule(TypesAreEqual(input, cellStateIn), reasonIfUnsupported,
-                                  "Reference UnidirectionalSequenceLstm: input and cellStateIn types are mismatched");
+    supported &= CheckSupportRule(TypeAnyOf(output, supportedTypes), reasonIfUnsupported,
+                                  "Reference UnidirectionalSequenceLstm: output is not a supported type.");
 
-    supported &= CheckSupportRule(TypesAreEqual(input, output), reasonIfUnsupported,
-                                  "Reference UnidirectionalSequenceLstm: input and output types are mismatched");
     // check layer parameters
     supported &= CheckSupportRule(TypeAnyOf(paramsInfo.GetInputToForgetWeights(), supportedWeightTypes),
                                   reasonIfUnsupported,
@@ -2905,14 +2863,13 @@
                                   reasonIfUnsupported,
                                   "Reference UnidirectionalSequenceLstm: RecurrentToOutputWeights "
                                   "is not a supported type.");
-    supported &= CheckSupportRule(TypesAreEqual(input, paramsInfo.GetForgetGateBias()), reasonIfUnsupported,
-                                  "Reference UnidirectionalSequenceLstm: input and ForgetGateBias types "
-                                  "are mismatched");
-    supported &= CheckSupportRule(TypesAreEqual(input, paramsInfo.GetCellBias()), reasonIfUnsupported,
-                                  "Reference UnidirectionalSequenceLstm: input and CellBias types are mismatched");
-    supported &= CheckSupportRule(TypesAreEqual(input, paramsInfo.GetOutputGateBias()), reasonIfUnsupported,
-                                  "Reference UnidirectionalSequenceLstm: input and OutputGateBias types "
-                                  "are mismatched");
+
+    supported &= CheckSupportRule(TypeAnyOf(paramsInfo.GetForgetGateBias(), supportedBiasTypes), reasonIfUnsupported,
+                                  "Reference UnidirectionalSequenceLstm: ForgetGateBias is not a supported type.");
+    supported &= CheckSupportRule(TypeAnyOf(paramsInfo.GetCellBias(), supportedBiasTypes), reasonIfUnsupported,
+                                  "Reference UnidirectionalSequenceLstm: CellBias is not a supported type.");
+    supported &= CheckSupportRule(TypeAnyOf(paramsInfo.GetOutputGateBias(), supportedBiasTypes), reasonIfUnsupported,
+                                  "Reference UnidirectionalSequenceLstm: OutputGateBias is not a supported type.");
     if (!descriptor.m_CifgEnabled)
     {
         supported &= CheckSupportRule(TypeAnyOf(paramsInfo.GetInputToInputWeights(), supportedWeightTypes),
@@ -2923,9 +2880,8 @@
                                       reasonIfUnsupported,
                                       "Reference UnidirectionalSequenceLstm: RecurrentToInputWeights "
                                       "is not a supported type.");
-        supported &= CheckSupportRule(TypesAreEqual(input, paramsInfo.GetInputGateBias()), reasonIfUnsupported,
-                                      "Reference UnidirectionalSequenceLstm: input and InputGateBias types "
-                                      "are mismatched");
+        supported &= CheckSupportRule(TypeAnyOf(paramsInfo.GetInputGateBias(), supportedBiasTypes), reasonIfUnsupported,
+                                      "Reference UnidirectionalSequenceLstm: InputGateBias is not a supported type.");
         if (descriptor.m_PeepholeEnabled)
         {
             supported &= CheckSupportRule(TypeAnyOf(paramsInfo.GetCellToInputWeights(), supportedWeightTypes),
diff --git a/src/backends/reference/RefLayerSupport.hpp b/src/backends/reference/RefLayerSupport.hpp
index 98770ad..aa8bd8d 100644
--- a/src/backends/reference/RefLayerSupport.hpp
+++ b/src/backends/reference/RefLayerSupport.hpp
@@ -367,9 +367,9 @@
         const TensorInfo& input,
         const TensorInfo& outputStateIn,
         const TensorInfo& cellStateIn,
+        const TensorInfo& outputStateOut,
+        const TensorInfo& cellStateOut,
         const TensorInfo& output,
-        const Optional<TensorInfo>& hiddenStateOutput,
-        const Optional<TensorInfo>& cellStateOutput,
         const UnidirectionalSequenceLstmDescriptor& descriptor,
         const LstmInputParamsInfo& paramsInfo,
         Optional<std::string&> reasonIfUnsupported = EmptyOptional()) const override;
diff --git a/src/backends/reference/workloads/RefUnidirectionalSequenceLstmWorkload.cpp b/src/backends/reference/workloads/RefUnidirectionalSequenceLstmWorkload.cpp
index d447a46..c4345d4 100644
--- a/src/backends/reference/workloads/RefUnidirectionalSequenceLstmWorkload.cpp
+++ b/src/backends/reference/workloads/RefUnidirectionalSequenceLstmWorkload.cpp
@@ -59,7 +59,9 @@
     TensorInfo inputInfo = GetTensorInfo(inputs[0]);
     const TensorInfo& outputStateInfo = GetTensorInfo(inputs[1]);
     const TensorInfo& cellStateInfo = GetTensorInfo(inputs[2]);
-    TensorInfo outputInfo = GetTensorInfo(outputs[0]);
+    TensorInfo outputStateOutInfo = GetTensorInfo(outputs[0]);
+    TensorInfo cellStateOutInfo = GetTensorInfo(outputs[1]);
+    TensorInfo outputInfo = GetTensorInfo(outputs[2]);
     TensorShape& inputShape = inputInfo.GetShape();
     TensorShape& outputShape= outputInfo.GetShape();
     auto inputTensor = reinterpret_cast<float*>(inputs[0]->Map());
@@ -140,7 +142,7 @@
 
     auto currentInputData = reinterpret_cast<float*>(inputs[0]->Map());
     std::unique_ptr<Decoder<float>> inputData = MakeDecoder<float>(lstmInputInfo, currentInputData);
-    auto currentOutputData = reinterpret_cast<float*>(outputs[0]->Map());
+    auto currentOutputData = reinterpret_cast<float*>(outputs[2]->Map());
     std::unique_ptr<Encoder<float>> output = MakeEncoder<float>(lstmOutputInfo, currentOutputData);
     std::unique_ptr<Decoder<float>> outputDecoder = MakeDecoder<float>(lstmOutputInfo, currentOutputData);
 
@@ -296,7 +298,7 @@
     {
         // Permute Output back to batch major
         const PermutationVector& mappings = {1U, 0U, 2U};
-        auto outputData = reinterpret_cast<float*>(outputs[0]->Map());
+        auto outputData = reinterpret_cast<float*>(outputs[2]->Map());
         std::vector<float> outputValue(outputData, outputData + outputInfo.GetNumElements());
         outputShape = armnnUtils::Permuted(outputInfo.GetShape(), mappings);
         outputInfo.SetShape(outputShape);