IVGCVSW-4770 Fix segmentation fault in FileOnlyProfilingDecoratorTests
Signed-off-by: Jan Eilers <jan.eilers@arm.com>
Change-Id: I5725aa418b50fc14ce4e6638fe29a74d762cc304
diff --git a/src/profiling/test/FileOnlyProfilingDecoratorTests.cpp b/src/profiling/test/FileOnlyProfilingDecoratorTests.cpp
index 79da17a..69ebe33 100644
--- a/src/profiling/test/FileOnlyProfilingDecoratorTests.cpp
+++ b/src/profiling/test/FileOnlyProfilingDecoratorTests.cpp
@@ -38,212 +38,206 @@
BOOST_AUTO_TEST_CASE(TestFileOnlyProfiling)
{
- // This test requires at least one backend registry to be enabled
- // which can execute a NormalizationLayer
- if (!HasSuitableBackendRegistered())
+ // Get all registered backends
+ std::vector<BackendId> suitableBackends = GetSuitableBackendRegistered();
+
+ // Run test for each backend separately
+ for (auto const& backend : suitableBackends)
{
- return;
+ // Enable m_FileOnly but also provide ILocalPacketHandler which should consume the packets.
+ // This won't dump anything to file.
+ armnn::Runtime::CreationOptions creationOptions;
+ creationOptions.m_ProfilingOptions.m_EnableProfiling = true;
+ creationOptions.m_ProfilingOptions.m_FileOnly = true;
+ creationOptions.m_ProfilingOptions.m_CapturePeriod = 100;
+ creationOptions.m_ProfilingOptions.m_TimelineEnabled = true;
+ ILocalPacketHandlerSharedPtr localPacketHandlerPtr = std::make_shared<TestTimelinePacketHandler>();
+ creationOptions.m_ProfilingOptions.m_LocalPacketHandlers.push_back(localPacketHandlerPtr);
+
+ armnn::Runtime runtime(creationOptions);
+ // ensure the GUID generator is reset to zero
+ GetProfilingService(&runtime).ResetGuidGenerator();
+
+ // Load a simple network
+ // build up the structure of the network
+ INetworkPtr net(INetwork::Create());
+
+ IConnectableLayer* input = net->AddInputLayer(0, "input");
+
+ ElementwiseUnaryDescriptor descriptor(UnaryOperation::Rsqrt);
+ IConnectableLayer* Rsqrt = net->AddElementwiseUnaryLayer(descriptor, "Rsqrt");
+
+ IConnectableLayer* output = net->AddOutputLayer(0, "output");
+
+ input->GetOutputSlot(0).Connect(Rsqrt->GetInputSlot(0));
+ Rsqrt->GetOutputSlot(0).Connect(output->GetInputSlot(0));
+
+ input->GetOutputSlot(0).SetTensorInfo(TensorInfo({ 1, 1, 4, 4 }, DataType::Float32));
+ Rsqrt->GetOutputSlot(0).SetTensorInfo(TensorInfo({ 1, 1, 4, 4 }, DataType::Float32));
+
+ std::vector<armnn::BackendId> backendsVec {backend};
+ IOptimizedNetworkPtr optNet = Optimize(*net, backendsVec, runtime.GetDeviceSpec());
+
+ // Load it into the runtime. It should succeed.
+ armnn::NetworkId netId;
+ BOOST_TEST(runtime.LoadNetwork(netId, std::move(optNet)) == Status::Success);
+
+ // Creates structures for input & output.
+ std::vector<float> inputData(16);
+ std::vector<float> outputData(16);
+ for (unsigned int i = 0; i < 16; ++i) {
+ inputData[i] = 9.0;
+ outputData[i] = 3.0;
+ }
+
+ InputTensors inputTensors
+ {
+ {0, ConstTensor(runtime.GetInputTensorInfo(netId, 0), inputData.data())}
+ };
+ OutputTensors outputTensors
+ {
+ {0, Tensor(runtime.GetOutputTensorInfo(netId, 0), outputData.data())}
+ };
+
+ // Does the inference.
+ runtime.EnqueueWorkload(netId, inputTensors, outputTensors);
+
+ static_cast<TestTimelinePacketHandler *>(localPacketHandlerPtr.get())->WaitOnInferenceCompletion(3000);
+
+ const TimelineModel &model =
+ static_cast<TestTimelinePacketHandler *>(localPacketHandlerPtr.get())->GetTimelineModel();
+
+ for (auto &error : model.GetErrors()) {
+ std::cout << error.what() << std::endl;
+ }
+ BOOST_TEST(model.GetErrors().empty());
+ std::vector<std::string> desc = GetModelDescription(model);
+ std::vector<std::string> expectedOutput;
+ expectedOutput.push_back("Entity [0] name = input type = layer");
+ expectedOutput.push_back(" connection [14] from entity [0] to entity [1]");
+ expectedOutput.push_back(" child: Entity [23] backendId = " + backend.Get() + " type = workload");
+ expectedOutput.push_back("Entity [1] name = Rsqrt type = layer");
+ expectedOutput.push_back(" connection [22] from entity [1] to entity [2]");
+ expectedOutput.push_back(" child: Entity [15] backendId = " + backend.Get() + " type = workload");
+ expectedOutput.push_back("Entity [2] name = output type = layer");
+ expectedOutput.push_back(" child: Entity [27] backendId = " + backend.Get() + " type = workload");
+ expectedOutput.push_back("Entity [6] type = network");
+ expectedOutput.push_back(" child: Entity [0] name = input type = layer");
+ expectedOutput.push_back(" child: Entity [1] name = Rsqrt type = layer");
+ expectedOutput.push_back(" child: Entity [2] name = output type = layer");
+ expectedOutput.push_back(" execution: Entity [31] type = inference");
+ expectedOutput.push_back("Entity [15] backendId = " + backend.Get() + " type = workload");
+ expectedOutput.push_back(" execution: Entity [44] type = workload_execution");
+ expectedOutput.push_back("Entity [23] backendId = " + backend.Get() + " type = workload");
+ expectedOutput.push_back(" execution: Entity [36] type = workload_execution");
+ expectedOutput.push_back("Entity [27] backendId = " + backend.Get() + " type = workload");
+ expectedOutput.push_back(" execution: Entity [52] type = workload_execution");
+ expectedOutput.push_back("Entity [31] type = inference");
+ expectedOutput.push_back(" child: Entity [36] type = workload_execution");
+ expectedOutput.push_back(" child: Entity [44] type = workload_execution");
+ expectedOutput.push_back(" child: Entity [52] type = workload_execution");
+ expectedOutput.push_back(" event: [34] class [start_of_life]");
+ expectedOutput.push_back(" event: [60] class [end_of_life]");
+ expectedOutput.push_back("Entity [36] type = workload_execution");
+ expectedOutput.push_back(" event: [40] class [start_of_life]");
+ expectedOutput.push_back(" event: [42] class [end_of_life]");
+ expectedOutput.push_back("Entity [44] type = workload_execution");
+ expectedOutput.push_back(" event: [48] class [start_of_life]");
+ expectedOutput.push_back(" event: [50] class [end_of_life]");
+ expectedOutput.push_back("Entity [52] type = workload_execution");
+ expectedOutput.push_back(" event: [56] class [start_of_life]");
+ expectedOutput.push_back(" event: [58] class [end_of_life]");
+ BOOST_TEST(CompareOutput(desc, expectedOutput));
}
-
- // Enable m_FileOnly but also provide ILocalPacketHandler which should consume the packets.
- // This won't dump anything to file.
- armnn::Runtime::CreationOptions creationOptions;
- creationOptions.m_ProfilingOptions.m_EnableProfiling = true;
- creationOptions.m_ProfilingOptions.m_FileOnly = true;
- creationOptions.m_ProfilingOptions.m_CapturePeriod = 100;
- creationOptions.m_ProfilingOptions.m_TimelineEnabled = true;
- ILocalPacketHandlerSharedPtr localPacketHandlerPtr = std::make_shared<TestTimelinePacketHandler>();
- creationOptions.m_ProfilingOptions.m_LocalPacketHandlers.push_back(localPacketHandlerPtr);
-
- armnn::Runtime runtime(creationOptions);
- // ensure the GUID generator is reset to zero
- GetProfilingService(&runtime).ResetGuidGenerator();
-
- // Load a simple network
- // build up the structure of the network
- INetworkPtr net(INetwork::Create());
-
- IConnectableLayer* input = net->AddInputLayer(0, "input");
-
- ElementwiseUnaryDescriptor descriptor(UnaryOperation::Sqrt);
- IConnectableLayer* normalize = net->AddElementwiseUnaryLayer(descriptor, "normalization");
-
- IConnectableLayer* output = net->AddOutputLayer(0, "output");
-
- input->GetOutputSlot(0).Connect(normalize->GetInputSlot(0));
- normalize->GetOutputSlot(0).Connect(output->GetInputSlot(0));
-
- input->GetOutputSlot(0).SetTensorInfo(TensorInfo({ 1, 1, 4, 4 }, DataType::Float32));
- normalize->GetOutputSlot(0).SetTensorInfo(TensorInfo({ 1, 1, 4, 4 }, DataType::Float32));
-
- // optimize the network
- std::vector<armnn::BackendId> backends =
- { armnn::Compute::CpuRef, armnn::Compute::CpuAcc, armnn::Compute::GpuAcc };
- IOptimizedNetworkPtr optNet = Optimize(*net, backends, runtime.GetDeviceSpec());
-
- // Load it into the runtime. It should succeed.
- armnn::NetworkId netId;
- BOOST_TEST(runtime.LoadNetwork(netId, std::move(optNet)) == Status::Success);
-
- // Creates structures for input & output.
- std::vector<float> inputData(16);
- std::vector<float> outputData(16);
- for (unsigned int i = 0; i < 16; ++i)
- {
- inputData[i] = 9.0;
- outputData[i] = 3.0;
- }
-
- InputTensors inputTensors
- {
- {0, ConstTensor(runtime.GetInputTensorInfo(netId, 0), inputData.data())}
- };
- OutputTensors outputTensors
- {
- {0, Tensor(runtime.GetOutputTensorInfo(netId, 0), outputData.data())}
- };
-
- // Does the inference.
- runtime.EnqueueWorkload(netId, inputTensors, outputTensors);
-
- static_cast<TestTimelinePacketHandler*>(localPacketHandlerPtr.get())->WaitOnInferenceCompletion(3000);
-
- const TimelineModel& model =
- static_cast<TestTimelinePacketHandler*>(localPacketHandlerPtr.get())->GetTimelineModel();
-
- for (auto& error : model.GetErrors())
- {
- std::cout << error.what() << std::endl;
- }
- BOOST_TEST(model.GetErrors().empty());
- std::vector<std::string> desc = GetModelDescription(model);
- std::vector<std::string> expectedOutput;
- expectedOutput.push_back("Entity [0] name = input type = layer");
- expectedOutput.push_back(" connection [14] from entity [0] to entity [1]");
- expectedOutput.push_back(" child: Entity [23] backendId = CpuRef type = workload");
- expectedOutput.push_back("Entity [1] name = normalization type = layer");
- expectedOutput.push_back(" connection [22] from entity [1] to entity [2]");
- expectedOutput.push_back(" child: Entity [15] backendId = CpuRef type = workload");
- expectedOutput.push_back("Entity [2] name = output type = layer");
- expectedOutput.push_back(" child: Entity [27] backendId = CpuRef type = workload");
- expectedOutput.push_back("Entity [6] type = network");
- expectedOutput.push_back(" child: Entity [0] name = input type = layer");
- expectedOutput.push_back(" child: Entity [1] name = normalization type = layer");
- expectedOutput.push_back(" child: Entity [2] name = output type = layer");
- expectedOutput.push_back(" execution: Entity [31] type = inference");
- expectedOutput.push_back("Entity [15] backendId = CpuRef type = workload");
- expectedOutput.push_back(" execution: Entity [44] type = workload_execution");
- expectedOutput.push_back("Entity [23] backendId = CpuRef type = workload");
- expectedOutput.push_back(" execution: Entity [36] type = workload_execution");
- expectedOutput.push_back("Entity [27] backendId = CpuRef type = workload");
- expectedOutput.push_back(" execution: Entity [52] type = workload_execution");
- expectedOutput.push_back("Entity [31] type = inference");
- expectedOutput.push_back(" child: Entity [36] type = workload_execution");
- expectedOutput.push_back(" child: Entity [44] type = workload_execution");
- expectedOutput.push_back(" child: Entity [52] type = workload_execution");
- expectedOutput.push_back(" event: [34] class [start_of_life]");
- expectedOutput.push_back(" event: [60] class [end_of_life]");
- expectedOutput.push_back("Entity [36] type = workload_execution");
- expectedOutput.push_back(" event: [40] class [start_of_life]");
- expectedOutput.push_back(" event: [42] class [end_of_life]");
- expectedOutput.push_back("Entity [44] type = workload_execution");
- expectedOutput.push_back(" event: [48] class [start_of_life]");
- expectedOutput.push_back(" event: [50] class [end_of_life]");
- expectedOutput.push_back("Entity [52] type = workload_execution");
- expectedOutput.push_back(" event: [56] class [start_of_life]");
- expectedOutput.push_back(" event: [58] class [end_of_life]");
- BOOST_TEST(CompareOutput(desc, expectedOutput));
}
BOOST_AUTO_TEST_CASE(DumpOutgoingValidFileEndToEnd)
{
- // This test requires at least one backend registry to be enabled
- // which can execute a NormalizationLayer
- if (!HasSuitableBackendRegistered())
+ // Get all registered backends
+ std::vector<BackendId> suitableBackends = GetSuitableBackendRegistered();
+
+ // Run test for each backend separately
+ for (auto const& backend : suitableBackends)
{
- return;
+ // Create a temporary file name.
+ fs::path tempPath = armnnUtils::Filesystem::NamedTempFile("DumpOutgoingValidFileEndToEnd_CaptureFile.txt");
+ // Make sure the file does not exist at this point
+ BOOST_CHECK(!fs::exists(tempPath));
+
+ armnn::Runtime::CreationOptions options;
+ options.m_ProfilingOptions.m_EnableProfiling = true;
+ options.m_ProfilingOptions.m_FileOnly = true;
+ options.m_ProfilingOptions.m_IncomingCaptureFile = "";
+ options.m_ProfilingOptions.m_OutgoingCaptureFile = tempPath.string();
+ options.m_ProfilingOptions.m_CapturePeriod = 100;
+ options.m_ProfilingOptions.m_TimelineEnabled = true;
+
+ ILocalPacketHandlerSharedPtr localPacketHandlerPtr = std::make_shared<TestTimelinePacketHandler>();
+ options.m_ProfilingOptions.m_LocalPacketHandlers.push_back(localPacketHandlerPtr);
+
+ armnn::Runtime runtime(options);
+ // ensure the GUID generator is reset to zero
+ GetProfilingService(&runtime).ResetGuidGenerator();
+
+ // Load a simple network
+ // build up the structure of the network
+ INetworkPtr net(INetwork::Create());
+
+ IConnectableLayer* input = net->AddInputLayer(0, "input");
+
+ ElementwiseUnaryDescriptor descriptor(UnaryOperation::Rsqrt);
+ IConnectableLayer* Rsqrt = net->AddElementwiseUnaryLayer(descriptor, "Rsqrt");
+
+ IConnectableLayer* output = net->AddOutputLayer(0, "output");
+
+ input->GetOutputSlot(0).Connect(Rsqrt->GetInputSlot(0));
+ Rsqrt->GetOutputSlot(0).Connect(output->GetInputSlot(0));
+
+ input->GetOutputSlot(0).SetTensorInfo(TensorInfo({ 1, 1, 4, 4 }, DataType::Float32));
+ Rsqrt->GetOutputSlot(0).SetTensorInfo(TensorInfo({ 1, 1, 4, 4 }, DataType::Float32));
+
+
+ std::vector<BackendId> backendsVec{backend};
+ IOptimizedNetworkPtr optNet = Optimize(*net, backendsVec, runtime.GetDeviceSpec());
+
+ // Load it into the runtime. It should succeed.
+ armnn::NetworkId netId;
+ BOOST_TEST(runtime.LoadNetwork(netId, std::move(optNet)) == Status::Success);
+
+ // Creates structures for input & output.
+ std::vector<float> inputData(16);
+ std::vector<float> outputData(16);
+ for (unsigned int i = 0; i < 16; ++i) {
+ inputData[i] = 9.0;
+ outputData[i] = 3.0;
+ }
+
+ InputTensors inputTensors
+ {
+ {0, ConstTensor(runtime.GetInputTensorInfo(netId, 0), inputData.data())}
+ };
+ OutputTensors outputTensors
+ {
+ {0, Tensor(runtime.GetOutputTensorInfo(netId, 0), outputData.data())}
+ };
+
+ // Does the inference.
+ runtime.EnqueueWorkload(netId, inputTensors, outputTensors);
+
+ static_cast<TestTimelinePacketHandler *>(localPacketHandlerPtr.get())->WaitOnInferenceCompletion(3000);
+
+ // In order to flush the files we need to gracefully close the profiling service.
+ options.m_ProfilingOptions.m_EnableProfiling = false;
+ GetProfilingService(&runtime).ResetExternalProfilingOptions(options.m_ProfilingOptions, true);
+
+ // The output file size should be greater than 0.
+ BOOST_CHECK(fs::file_size(tempPath) > 0);
+
+ // NOTE: would be an interesting exercise to take this file and decode it
+
+ // Delete the tmp file.
+ BOOST_CHECK(fs::remove(tempPath));
}
-
- // Create a temporary file name.
- fs::path tempPath = armnnUtils::Filesystem::NamedTempFile("DumpOutgoingValidFileEndToEnd_CaptureFile.txt");
- armnn::Runtime::CreationOptions options;
- options.m_ProfilingOptions.m_EnableProfiling = true;
- options.m_ProfilingOptions.m_FileOnly = true;
- options.m_ProfilingOptions.m_IncomingCaptureFile = "";
- options.m_ProfilingOptions.m_OutgoingCaptureFile = tempPath.string();
- options.m_ProfilingOptions.m_CapturePeriod = 100;
- options.m_ProfilingOptions.m_TimelineEnabled = true;
-
- ILocalPacketHandlerSharedPtr localPacketHandlerPtr = std::make_shared<TestTimelinePacketHandler>();
- options.m_ProfilingOptions.m_LocalPacketHandlers.push_back(localPacketHandlerPtr);
-
- // Make sure the file does not exist at this point
- BOOST_CHECK(!fs::exists(tempPath));
-
- armnn::Runtime runtime(options);
- // ensure the GUID generator is reset to zero
- GetProfilingService(&runtime).ResetGuidGenerator();
-
- // Load a simple network
- // build up the structure of the network
- INetworkPtr net(INetwork::Create());
-
- IConnectableLayer* input = net->AddInputLayer(0, "input");
-
- ElementwiseUnaryDescriptor descriptor(UnaryOperation::Sqrt);
- IConnectableLayer* normalize = net->AddElementwiseUnaryLayer(descriptor, "normalization");
-
- IConnectableLayer* output = net->AddOutputLayer(0, "output");
-
- input->GetOutputSlot(0).Connect(normalize->GetInputSlot(0));
- normalize->GetOutputSlot(0).Connect(output->GetInputSlot(0));
-
- input->GetOutputSlot(0).SetTensorInfo(TensorInfo({ 1, 1, 4, 4 }, DataType::Float32));
- normalize->GetOutputSlot(0).SetTensorInfo(TensorInfo({ 1, 1, 4, 4 }, DataType::Float32));
-
- // optimize the network
- std::vector<armnn::BackendId> backends =
- { armnn::Compute::CpuRef, armnn::Compute::CpuAcc, armnn::Compute::GpuAcc };
- IOptimizedNetworkPtr optNet = Optimize(*net, backends, runtime.GetDeviceSpec());
-
- // Load it into the runtime. It should succeed.
- armnn::NetworkId netId;
- BOOST_TEST(runtime.LoadNetwork(netId, std::move(optNet)) == Status::Success);
-
- // Creates structures for input & output.
- std::vector<float> inputData(16);
- std::vector<float> outputData(16);
- for (unsigned int i = 0; i < 16; ++i)
- {
- inputData[i] = 9.0;
- outputData[i] = 3.0;
- }
-
- InputTensors inputTensors
- {
- {0, ConstTensor(runtime.GetInputTensorInfo(netId, 0), inputData.data())}
- };
- OutputTensors outputTensors
- {
- {0, Tensor(runtime.GetOutputTensorInfo(netId, 0), outputData.data())}
- };
-
- // Does the inference.
- runtime.EnqueueWorkload(netId, inputTensors, outputTensors);
-
- static_cast<TestTimelinePacketHandler*>(localPacketHandlerPtr.get())->WaitOnInferenceCompletion(3000);
-
- // In order to flush the files we need to gracefully close the profiling service.
- options.m_ProfilingOptions.m_EnableProfiling = false;
- GetProfilingService(&runtime).ResetExternalProfilingOptions(options.m_ProfilingOptions, true);
-
- // The output file size should be greater than 0.
- BOOST_CHECK(fs::file_size(tempPath) > 0);
-
- // NOTE: would be an interesting exercise to take this file and decode it
-
- // Delete the tmp file.
- BOOST_CHECK(fs::remove(tempPath));
}
BOOST_AUTO_TEST_SUITE_END()
diff --git a/src/profiling/test/ProfilingTestUtils.cpp b/src/profiling/test/ProfilingTestUtils.cpp
index 0f43728..9f6bc43 100644
--- a/src/profiling/test/ProfilingTestUtils.cpp
+++ b/src/profiling/test/ProfilingTestUtils.cpp
@@ -34,6 +34,24 @@
return headerSize + bodySize + payloadSize;
}
+std::vector<BackendId> GetSuitableBackendRegistered()
+{
+ std::vector<BackendId> suitableBackends;
+ if (BackendRegistryInstance().IsBackendRegistered(GetComputeDeviceAsCString(armnn::Compute::CpuRef)))
+ {
+ suitableBackends.push_back(armnn::Compute::CpuRef);
+ }
+ if (BackendRegistryInstance().IsBackendRegistered(GetComputeDeviceAsCString(armnn::Compute::CpuAcc)))
+ {
+ suitableBackends.push_back(armnn::Compute::CpuAcc);
+ }
+ if (BackendRegistryInstance().IsBackendRegistered(GetComputeDeviceAsCString(armnn::Compute::GpuAcc)))
+ {
+ suitableBackends.push_back(armnn::Compute::GpuAcc);
+ }
+ return suitableBackends;
+}
+
inline unsigned int OffsetToNextWord(unsigned int numberOfBytes)
{
unsigned int uint32_t_size = sizeof(uint32_t);
@@ -1199,17 +1217,6 @@
bufferManager.MarkRead(inferenceReadableBuffer);
}
-bool HasSuitableBackendRegistered()
-{
- // Only run the file only profiling unit tests on CpuRef until failure on build system can be debugged
- if (BackendRegistryInstance().GetBackendIds().size() == 1 &&
- BackendRegistryInstance().IsBackendRegistered(GetComputeDeviceAsCString(armnn::Compute::CpuRef)))
- {
- return true;
- }
- return false;
-}
-
bool CompareOutput(std::vector<std::string> output, std::vector<std::string> expectedOutput)
{
if (output.size() != expectedOutput.size())
diff --git a/src/profiling/test/ProfilingTestUtils.hpp b/src/profiling/test/ProfilingTestUtils.hpp
index a9a6921..8f138bb 100644
--- a/src/profiling/test/ProfilingTestUtils.hpp
+++ b/src/profiling/test/ProfilingTestUtils.hpp
@@ -21,6 +21,9 @@
uint32_t GetStreamMetaDataPacketSize();
+/// Returns a vector of CpuRef, CpuAcc or GpuAcc backends if they where registered
+std::vector<BackendId> GetSuitableBackendRegistered();
+
inline unsigned int OffsetToNextWord(unsigned int numberOfBytes);
void VerifyTimelineHeaderBinary(const unsigned char* readableData,
@@ -57,10 +60,6 @@
void VerifyPostOptimisationStructureTestImpl(armnn::BackendId backendId);
-// returns true if a CpuRef, CpuAcc or GpuAcc
-// backend is registered
-bool HasSuitableBackendRegistered();
-
bool CompareOutput(std::vector<std::string> output, std::vector<std::string> expectedOutput);
namespace armnn