Revert "MLCE-1093 Reshape and concat invalid results"

This reverts commit 4980e21193f0a14fef084a7f4b4197392f3c0845.

Reason for revert: Android Build for v82a failed due to schema not re-generating

Change-Id: Ic19cf471b487f321c97ff837d36526512fb12fa4

1 files changed, 282 insertions, 1 deletions

diff --git a/src/backends/neon/test/NeonTensorHandleTests.cpp b/src/backends/neon/test/NeonTensorHandleTests.cpp
index a94e4dd187..c8e781b71d 100644
--- a/src/backends/neon/test/NeonTensorHandleTests.cpp
+++ b/src/backends/neon/test/NeonTensorHandleTests.cpp
@@ -1,5 +1,5 @@
 //
-// Copyright © 2020-2021,2023 Arm Ltd and Contributors. All rights reserved.
+// Copyright © 2020 Arm Ltd and Contributors. All rights reserved.
 // SPDX-License-Identifier: MIT
 //
 #include <Graph.hpp>
@@ -89,6 +89,81 @@ TEST_CASE("NeonTensorHandleGetCapabilitiesPadding")
     CHECK(capabilities[0].m_Value);
 }
 
+TEST_CASE("ConcatOnXorYSubTensorsNoPaddingRequiredTest")
+{
+    armnn::INetworkPtr net(armnn::INetwork::Create());
+
+    // Set up tensor infos
+    const armnn::TensorInfo inputInfo = armnn::TensorInfo({2, 3, 2, 2}, armnn::DataType::Float32);
+    const armnn::TensorInfo intermediateInfo = armnn::TensorInfo({2, 3, 2, 2}, armnn::DataType::Float32);
+    const armnn::TensorInfo outputInfo = armnn::TensorInfo({2, 3, 4, 2}, armnn::DataType::Float32);
+
+    armnn::ElementwiseUnaryDescriptor descriptor(armnn::UnaryOperation::Abs);
+
+    // Create the network
+    armnn::IConnectableLayer* const input0Layer = net->AddInputLayer(0, "input_0");
+    input0Layer->GetOutputSlot(0).SetTensorInfo(inputInfo);
+    armnn::IConnectableLayer* elementwiseUnaryLayer0 = net->AddElementwiseUnaryLayer(descriptor, "elementwiseUnary_0");
+    elementwiseUnaryLayer0->GetOutputSlot(0).SetTensorInfo(intermediateInfo);
+    input0Layer->GetOutputSlot(0).Connect(elementwiseUnaryLayer0->GetInputSlot(0));
+
+    armnn::IConnectableLayer* const input1Layer = net->AddInputLayer(1, "input_1");
+    input1Layer->GetOutputSlot(0).SetTensorInfo(inputInfo);
+    armnn::IConnectableLayer* elementwiseUnaryLayer1 = net->AddElementwiseUnaryLayer(descriptor, "elementwiseUnary_1");
+    elementwiseUnaryLayer1->GetOutputSlot(0).SetTensorInfo(intermediateInfo);
+    input1Layer->GetOutputSlot(0).Connect(elementwiseUnaryLayer1->GetInputSlot(0));
+
+    std::array<armnn::TensorShape, 2> concatInputShapes = { intermediateInfo.GetShape(), intermediateInfo.GetShape() };
+    armnn::IConnectableLayer* const concatLayer = net->AddConcatLayer(armnn::CreateDescriptorForConcatenation(
+        concatInputShapes.begin(), concatInputShapes.end(), 2), "concatenation");
+    concatLayer->GetOutputSlot(0).SetTensorInfo(outputInfo);
+    elementwiseUnaryLayer0->GetOutputSlot(0).Connect(concatLayer->GetInputSlot(0));
+    elementwiseUnaryLayer1->GetOutputSlot(0).Connect(concatLayer->GetInputSlot(1));
+
+    armnn::IConnectableLayer* const outputLayer = net->AddOutputLayer(0, "output");
+    concatLayer->GetOutputSlot(0).Connect(outputLayer->GetInputSlot(0));
+
+    armnn::IRuntime::CreationOptions options;
+    armnn::IRuntimePtr runtime(armnn::IRuntime::Create(options));
+
+    std::vector<armnn::BackendId> backends = { armnn::Compute::CpuAcc };
+    armnn::IOptimizedNetworkPtr optimizedNet = armnn::Optimize(*net, backends, runtime->GetDeviceSpec());
+
+    const armnn::Graph& theGraph = GetGraphForTesting(optimizedNet.get());
+
+    // Load graph into runtime
+    armnn::NetworkId networkIdentifier;
+    runtime->LoadNetwork(networkIdentifier, std::move(optimizedNet));
+
+    // now check the concat how many sub-tensors it is using..
+    auto TraceSubTensorHandleAncestry = [](armnn::ITensorHandle* const subTensorHandle)
+    {
+        if (subTensorHandle && subTensorHandle->GetParent())
+        {
+            return true;
+        }
+        return false;
+    };
+
+    for (auto&& layer : theGraph)
+    {
+        if(layer->GetType() == armnn::LayerType::Concat)
+        {
+            unsigned int numberOfSubTensors = 0;
+            for (unsigned int i = 0; i < layer->GetNumInputSlots(); ++i)
+            {
+                const armnn::OutputSlot* slot = layer->GetInputSlot(i).GetConnectedOutputSlot();
+                if (TraceSubTensorHandleAncestry(slot->GetOutputHandler().GetData()))
+                {
+                    ++numberOfSubTensors;
+                }
+            }
+            // sub-tensors should be supported in this configuration
+            ARMNN_ASSERT(numberOfSubTensors > 0);
+        }
+    }
+}
+
 TEST_CASE("ConcatonXorYPaddingRequiredTest")
 {
     armnn::INetworkPtr net(armnn::INetwork::Create());
@@ -172,6 +247,212 @@ TEST_CASE("ConcatonXorYPaddingRequiredTest")
     ARMNN_ASSERT(numberOfSubTensors == 0);
 }
 
+TEST_CASE("SplitteronXorYNoPaddingRequiredTest")
+{
+    using namespace armnn;
+
+    unsigned int splitAxis = 2;
+    unsigned int numSplit = 2;
+
+    const TensorShape& inputShape = { 2, 3, 4, 2 };
+    const armnn::TensorInfo intermediateInfo = armnn::TensorInfo({ 2, 3, 2, 2 }, armnn::DataType::Float32);
+    const std::vector<TensorShape> outputShapes{{ 2, 3, 2, 2 },
+                                                { 2, 3, 2, 2 }};
+    const float qScale = 1.0f;
+    const int32_t qOffset = 0;
+
+    // Creates structures for input & output.
+    std::vector<float> inputData{
+            1, 2,
+            3, 4,
+            5, 6,
+            7, 8,
+            9, 10,
+            11, 12,
+            13, 14,
+            15, 16,
+            17, 18,
+            19, 20,
+            21, 22,
+            23, 24,
+            25, 26,
+            27, 28,
+            29, 30,
+            31, 32,
+            33, 34,
+            35, 36,
+            37, 38,
+            39, 40,
+            41, 42,
+            43, 44,
+            45, 46,
+            47, 48
+    };
+
+    std::vector<float> expectedOutput0{
+            1, 2,
+            3, 4,
+            9, 10,
+            11, 12,
+            17, 18,
+            19, 20,
+            25, 26,
+            27, 28,
+            33, 34,
+            35, 36,
+            41, 42,
+            43, 44
+    };
+
+    std::vector<float> expectedOutput1{
+            5, 6,
+            7, 8,
+            13, 14,
+            15, 16,
+            21, 22,
+            23, 24,
+            29, 30,
+            31, 32,
+            37, 38,
+            39, 40,
+            45, 46,
+            47, 48
+    };
+
+    // Builds up the structure of the network.
+    INetworkPtr net(INetwork::Create());
+
+    TensorInfo inputTensorInfo(inputShape, armnn::DataType::Float32, qScale, qOffset);
+
+    armnn::ElementwiseUnaryDescriptor descriptor(armnn::UnaryOperation::Abs);
+
+    // Splitter
+    std::vector<unsigned int> splitterDimSizes(inputShape.GetNumDimensions());
+
+    // Add current input shape to splitterDimSizes
+    for (unsigned int i = 0; i < inputShape.GetNumDimensions(); ++i)
+    {
+        splitterDimSizes[i] = inputTensorInfo.GetShape()[i];
+    }
+
+    if (splitterDimSizes[splitAxis] % numSplit != 0)
+    {
+        throw ParseException("Number of splits must evenly divide the dimension");
+    }
+
+    splitterDimSizes[splitAxis] /= numSplit;
+
+    SplitterDescriptor splitDesc(numSplit, inputShape.GetNumDimensions());
+
+    for (unsigned int g = 0; g < numSplit; ++g)
+    {
+        // Set the size of the views.
+        for (unsigned int dimIdx = 0; dimIdx < splitterDimSizes.size(); ++dimIdx)
+        {
+            splitDesc.SetViewSize(g, dimIdx, splitterDimSizes[dimIdx]);
+        }
+        splitDesc.SetViewOriginCoord(g, splitAxis, splitterDimSizes[splitAxis] * g);
+    }
+    IConnectableLayer* input = net->AddInputLayer(0, "input");
+    IConnectableLayer* elementWiseUnary0 = net->AddElementwiseUnaryLayer(descriptor, "elementwiseunary_0");
+    IConnectableLayer* elementWiseUnary1 = net->AddElementwiseUnaryLayer(descriptor, "elementwiseunary_0");
+    IConnectableLayer* splitter = net->AddSplitterLayer(splitDesc, "splitter");
+
+    // Connections
+    Connect(input, splitter, inputTensorInfo, 0, 0);
+    Connect(splitter, elementWiseUnary0, intermediateInfo, 0, 0);
+    Connect(splitter, elementWiseUnary1, intermediateInfo, 1, 0);
+
+    std::vector<IConnectableLayer*> pooling2dLayers{elementWiseUnary0, elementWiseUnary1};
+
+    for (unsigned int i = 0; i < outputShapes.size(); ++i)
+    {
+        TensorInfo outputTensorInfo(outputShapes[i], armnn::DataType::Float32, qScale, qOffset);
+        IConnectableLayer* output = net->AddOutputLayer(armnn::numeric_cast<LayerBindingId>(i));
+        Connect(pooling2dLayers[i], output, outputTensorInfo, 0, 0);
+    }
+
+    std::map<int, std::vector<float>> inputTensorData = {{ 0,inputData }};
+    std::map<int, std::vector<float>> expectedOutputData = {{ 0, expectedOutput0 }, { 1, expectedOutput1 }};
+
+    armnn::IRuntime::CreationOptions options;
+    armnn::IRuntimePtr runtime(armnn::IRuntime::Create(options));
+
+    std::vector<armnn::BackendId> backends = { armnn::Compute::CpuAcc };
+    armnn::IOptimizedNetworkPtr optimizedNet = armnn::Optimize(*net, backends, runtime->GetDeviceSpec());
+
+    const armnn::Graph& theGraph = GetGraphForTesting(optimizedNet.get());
+
+    // Load graph into runtime
+    armnn::NetworkId networkIdentifier;
+    runtime->LoadNetwork(networkIdentifier, std::move(optimizedNet));
+
+    // now check the concat how many sub-tensors it is using..
+    auto TraceSubTensorHandleAncestry = [](armnn::ITensorHandle* const subTensorHandle)
+    {
+        if (subTensorHandle && subTensorHandle->GetParent())
+        {
+            return true;
+        }
+        return false;
+    };
+
+    for (auto&& layer : theGraph)
+    {
+        if(layer->GetType() == armnn::LayerType::ElementwiseUnary)
+        {
+            unsigned int numberOfSubTensors = 0;
+            for (unsigned int i = 0; i < layer->GetNumInputSlots(); ++i)
+            {
+                const armnn::OutputSlot* slot = layer->GetInputSlot(i).GetConnectedOutputSlot();
+                if (TraceSubTensorHandleAncestry(slot->GetOutputHandler().GetData()))
+                {
+                    ++numberOfSubTensors;
+                }
+            }
+            // sub-tensors should be supported in this configuration
+            ARMNN_ASSERT(numberOfSubTensors > 0);
+        }
+    }
+
+    InputTensors inputTensors;
+    inputTensors.reserve(inputTensorData.size());
+    for (auto&& it : inputTensorData)
+    {
+        TensorInfo inputTensorInfo = runtime->GetInputTensorInfo(networkIdentifier, it.first);
+        inputTensorInfo.SetConstant(true);
+        inputTensors.push_back({it.first,
+                                ConstTensor(inputTensorInfo, it.second.data())});
+    }
+    OutputTensors outputTensors;
+    outputTensors.reserve(expectedOutputData.size());
+    std::map<int, std::vector<float>> outputStorage;
+    for (auto&& it : expectedOutputData)
+    {
+        std::vector<float> out(it.second.size());
+        outputStorage.emplace(it.first, out);
+        outputTensors.push_back({it.first,
+                                 Tensor(runtime->GetOutputTensorInfo(networkIdentifier, it.first),
+                                               outputStorage.at(it.first).data())});
+    }
+
+    // Does the inference.
+    runtime->EnqueueWorkload(networkIdentifier, inputTensors, outputTensors);
+
+    // Checks the results.
+    float tolerance = 0.000001f;
+    for (auto&& it : expectedOutputData)
+    {
+        std::vector<float> out = outputStorage.at(it.first);
+        for (unsigned int i = 0; i < out.size(); ++i)
+        {
+            CHECK_MESSAGE(Compare<armnn::DataType::Float32>(it.second[i], out[i], tolerance) == true,
+                    "Actual output: " << out[i] << ". Expected output:" << it.second[i]);
+
+        }
+    }
+}
+
 TEST_CASE("SplitteronXorYPaddingRequiredTest")
 {
     using namespace armnn;