IVGCVSW-6249 Add ProfilingDetails Macros to all workloads in Ref, Neon, CL

 * Add functionality to only output network details in ExNet

Signed-off-by: Keith Davis <keith.davis@arm.com>
Change-Id: I0c45e67193f308ce7b86f1bb1a918a266fefba2e

10 files changed, 61 insertions, 23 deletions

diff --git a/include/armnn/IProfiler.hpp b/include/armnn/IProfiler.hpp
index 5fbb67000a..1b450d0151 100644
--- a/include/armnn/IProfiler.hpp
+++ b/include/armnn/IProfiler.hpp
@@ -40,8 +40,8 @@ public:
     void Print(std::ostream& outStream) const;
 
     /// Print out details of each layer within the network that possesses a descriptor.
-    /// Also outputs tensor info.
-    void EnableNetworkDetailsToStdOut();
+    /// Also outputs tensor info. This will be part of the profiling json output
+    void EnableNetworkDetailsToStdOut(ProfilingDetailsMethod detailsMethod);
 
     ~IProfiler();
     IProfiler();
diff --git a/include/armnn/IRuntime.hpp b/include/armnn/IRuntime.hpp
index 97a9c2889e..345cdeb10a 100644
--- a/include/armnn/IRuntime.hpp
+++ b/include/armnn/IRuntime.hpp
@@ -40,7 +40,7 @@ struct INetworkProperties
           m_ExportEnabled(exportEnabled),
           m_AsyncEnabled(asyncEnabled),
           m_ProfilingEnabled(profilingEnabled),
-          m_OutputNetworkDetails(false),
+          m_OutputNetworkDetailsMethod(ProfilingDetailsMethod::Undefined),
           m_InputSource(m_ImportEnabled ? MemorySource::Malloc : MemorySource::Undefined),
           m_OutputSource(m_ExportEnabled ? MemorySource::Malloc : MemorySource::Undefined)
     {}
@@ -55,7 +55,7 @@ struct INetworkProperties
           m_ExportEnabled(outputSource != MemorySource::Undefined),
           m_AsyncEnabled(asyncEnabled),
           m_ProfilingEnabled(profilingEnabled),
-          m_OutputNetworkDetails(false),
+          m_OutputNetworkDetailsMethod(ProfilingDetailsMethod::Undefined),
           m_InputSource(inputSource),
           m_OutputSource(outputSource)
     {
@@ -66,12 +66,12 @@ struct INetworkProperties
                        MemorySource inputSource,
                        MemorySource outputSource,
                        bool profilingEnabled = false,
-                       bool outputDetails = false)
+                       ProfilingDetailsMethod detailsMethod = ProfilingDetailsMethod::Undefined)
         : m_ImportEnabled(inputSource != MemorySource::Undefined),
           m_ExportEnabled(outputSource != MemorySource::Undefined),
           m_AsyncEnabled(asyncEnabled),
           m_ProfilingEnabled(profilingEnabled),
-          m_OutputNetworkDetails(outputDetails),
+          m_OutputNetworkDetailsMethod(detailsMethod),
           m_InputSource(inputSource),
           m_OutputSource(outputSource)
     {}
@@ -85,7 +85,7 @@ struct INetworkProperties
 
     const bool m_ProfilingEnabled;
 
-    const bool m_OutputNetworkDetails;
+    const ProfilingDetailsMethod m_OutputNetworkDetailsMethod;
 
     const MemorySource m_InputSource;
     const MemorySource m_OutputSource;
diff --git a/include/armnn/Types.hpp b/include/armnn/Types.hpp
index 056aa83d2f..5e00026eba 100644
--- a/include/armnn/Types.hpp
+++ b/include/armnn/Types.hpp
@@ -56,6 +56,15 @@ enum class DataLayout
     NHWC = 2
 };
 
+/// Define the behaviour of the internal profiler when outputting network details
+enum class ProfilingDetailsMethod
+{
+    Undefined = 0,
+    DetailsWithEvents = 1,
+    DetailsOnly = 2
+};
+
+
 enum class QosExecPriority
 {
     Low    = 0,
diff --git a/src/armnn/LoadedNetwork.cpp b/src/armnn/LoadedNetwork.cpp
index 70b5f4e89b..531d43dbb5 100644
--- a/src/armnn/LoadedNetwork.cpp
+++ b/src/armnn/LoadedNetwork.cpp
@@ -127,7 +127,7 @@ LoadedNetwork::LoadedNetwork(std::unique_ptr<IOptimizedNetwork> net,
 
     m_Profiler->EnableProfiling(networkProperties.m_ProfilingEnabled);
 
-    if (networkProperties.m_OutputNetworkDetails) m_Profiler->EnableNetworkDetailsToStdOut();
+    m_Profiler->EnableNetworkDetailsToStdOut(networkProperties.m_OutputNetworkDetailsMethod);
 
     Graph& order = m_OptimizedNetwork->pOptimizedNetworkImpl->GetGraph().TopologicalSort();
     //First create tensor handlers, backends and workload factories.
diff --git a/src/armnn/Profiling.cpp b/src/armnn/Profiling.cpp
index 8208fd9e0e..7de602fc40 100644
--- a/src/armnn/Profiling.cpp
+++ b/src/armnn/Profiling.cpp
@@ -163,7 +163,7 @@ void ProfilerImpl::AnalyzeEventSequenceAndWriteResults(ItertType first, ItertTyp
 
 ProfilerImpl::ProfilerImpl()
     : m_ProfilingEnabled(false),
-      m_EnableDetailsToStdOut(false)
+      m_DetailsToStdOutMethod(ProfilingDetailsMethod::Undefined)
 {
     m_EventSequence.reserve(g_ProfilingEventCountHint);
 
@@ -197,9 +197,9 @@ void ProfilerImpl::EnableProfiling(bool enableProfiling)
     m_ProfilingEnabled = enableProfiling;
 }
 
-void ProfilerImpl::EnableNetworkDetailsToStdOut()
+void ProfilerImpl::EnableNetworkDetailsToStdOut(ProfilingDetailsMethod details)
 {
-    m_EnableDetailsToStdOut = true;
+    m_DetailsToStdOutMethod = details;
 }
 
 Event* ProfilerImpl::BeginEvent(armnn::IProfiler* profiler,
@@ -384,7 +384,9 @@ void ProfilerImpl::Print(std::ostream& outStream) const
     printer.PrintHeader();
     printer.PrintArmNNHeader();
 
-    if (m_ProfilingDetails.get()->DetailsExist() && m_EnableDetailsToStdOut)
+    if (m_ProfilingDetails.get()->DetailsExist() &&
+        (m_DetailsToStdOutMethod == ProfilingDetailsMethod::DetailsOnly
+         || m_DetailsToStdOutMethod == ProfilingDetailsMethod::DetailsWithEvents))
     {
         JsonChildObject detailsObject{ "layer_details" };
         ConfigureDetailsObject(detailsObject, m_ProfilingDetails.get()->GetProfilingDetails());
@@ -395,8 +397,10 @@ void ProfilerImpl::Print(std::ostream& outStream) const
 
     // print inference object, also prints child layer and kernel measurements
     size_t id = 0;
-    printer.PrintJsonChildObject(inferenceObject, id);
-
+    if (m_DetailsToStdOutMethod != ProfilingDetailsMethod::DetailsOnly)
+    {
+        printer.PrintJsonChildObject(inferenceObject, id);
+    }
     // end of ArmNN
     printer.PrintNewLine();
     printer.PrintFooter();
@@ -409,6 +413,11 @@ void ProfilerImpl::Print(std::ostream& outStream) const
     // Restores previous precision settings.
     outStream.flags(oldFlags);
     outStream.precision(oldPrecision);
+
+    if (m_DetailsToStdOutMethod == ProfilingDetailsMethod::DetailsOnly)
+    {
+        exit(0);
+    }
 }
 
 void ProfilerImpl::AnalyzeEventsAndWriteResults(std::ostream& outStream) const
@@ -545,9 +554,9 @@ void IProfiler::EnableProfiling(bool enableProfiling)
     pProfilerImpl->EnableProfiling(enableProfiling);
 }
 
-void IProfiler::EnableNetworkDetailsToStdOut()
+void IProfiler::EnableNetworkDetailsToStdOut(ProfilingDetailsMethod detailsMethod)
 {
-    pProfilerImpl->EnableNetworkDetailsToStdOut();
+    pProfilerImpl->EnableNetworkDetailsToStdOut(detailsMethod);
 }
 
 bool IProfiler::IsProfilingEnabled()
diff --git a/src/armnn/Profiling.hpp b/src/armnn/Profiling.hpp
index 372b489abf..42d7f4d638 100644
--- a/src/armnn/Profiling.hpp
+++ b/src/armnn/Profiling.hpp
@@ -60,7 +60,7 @@ public:
     bool IsProfilingEnabled();
 
     // Enables outputting the layer descriptors and infos to stdout
-    void EnableNetworkDetailsToStdOut();
+    void EnableNetworkDetailsToStdOut(ProfilingDetailsMethod detailsMethod);
 
     // Increments the event tag, allowing grouping of events in a user-defined manner (e.g. per inference).
     void UpdateEventTag();
@@ -102,7 +102,8 @@ public:
     std::vector<EventPtr> m_EventSequence;
     DescPtr m_ProfilingDetails = std::make_unique<ProfilingDetails>();
     bool m_ProfilingEnabled;
-    bool m_EnableDetailsToStdOut;
+    ProfilingDetailsMethod m_DetailsToStdOutMethod;
+
 };
 
 // Singleton profiler manager.
diff --git a/tests/ExecuteNetwork/ExecuteNetwork.cpp b/tests/ExecuteNetwork/ExecuteNetwork.cpp
index 64296d31b7..9a4864542f 100644
--- a/tests/ExecuteNetwork/ExecuteNetwork.cpp
+++ b/tests/ExecuteNetwork/ExecuteNetwork.cpp
@@ -325,6 +325,7 @@ int MainImpl(const ExecuteNetworkParams& params,
         inferenceModelParams.m_AsyncEnabled                   = params.m_Concurrent;
         inferenceModelParams.m_ThreadPoolSize                 = params.m_ThreadPoolSize;
         inferenceModelParams.m_OutputDetailsToStdOut          = params.m_OutputDetailsToStdOut;
+        inferenceModelParams.m_OutputDetailsOnlyToStdOut      = params.m_OutputDetailsOnlyToStdOut;
 
         for(const std::string& inputName: params.m_InputNames)
         {
@@ -769,7 +770,9 @@ int main(int argc, const char* argv[])
         return EXIT_FAILURE;
     }
 
-    if (ProgramOptions.m_ExNetParams.m_OutputDetailsToStdOut && !ProgramOptions.m_ExNetParams.m_EnableProfiling)
+    if ((ProgramOptions.m_ExNetParams.m_OutputDetailsToStdOut ||
+         ProgramOptions.m_ExNetParams.m_OutputDetailsOnlyToStdOut)
+         && !ProgramOptions.m_ExNetParams.m_EnableProfiling)
     {
         ARMNN_LOG(fatal) << "You must enable profiling if you would like to output layer details";
         return EXIT_FAILURE;
diff --git a/tests/ExecuteNetwork/ExecuteNetworkParams.hpp b/tests/ExecuteNetwork/ExecuteNetworkParams.hpp
index 97c605b0a7..e519b028a0 100644
--- a/tests/ExecuteNetwork/ExecuteNetworkParams.hpp
+++ b/tests/ExecuteNetwork/ExecuteNetworkParams.hpp
@@ -44,6 +44,7 @@ struct ExecuteNetworkParams
     std::string                   m_ModelPath;
     unsigned int                  m_NumberOfThreads;
     bool                          m_OutputDetailsToStdOut;
+    bool                          m_OutputDetailsOnlyToStdOut;
     std::vector<std::string>      m_OutputNames;
     std::vector<std::string>      m_OutputTensorFiles;
     std::vector<std::string>      m_OutputTypes;
diff --git a/tests/ExecuteNetwork/ExecuteNetworkProgramOptions.cpp b/tests/ExecuteNetwork/ExecuteNetworkProgramOptions.cpp
index 1fd4b3d96d..927d804725 100644
--- a/tests/ExecuteNetwork/ExecuteNetworkProgramOptions.cpp
+++ b/tests/ExecuteNetwork/ExecuteNetworkProgramOptions.cpp
@@ -410,9 +410,14 @@ ProgramOptions::ProgramOptions() : m_CxxOptions{"ExecuteNetwork",
                  cxxopts::value<uint32_t>(m_RuntimeOptions.m_ProfilingOptions.m_CapturePeriod)->default_value("150"))
 
                 ("output-network-details",
-                 "Outputs layer tensor infos and descriptors to std out. Defaults to off.",
+                 "Outputs layer tensor infos and descriptors to std out along with profiling events. Defaults to off.",
                  cxxopts::value<bool>(m_ExNetParams.m_OutputDetailsToStdOut)->default_value("false")
-                                                                            ->implicit_value("true"));
+                                                                            ->implicit_value("true"))
+                ("output-network-details-only",
+                 "Outputs layer tensor infos and descriptors to std out without profiling events. Defaults to off.",
+                 cxxopts::value<bool>(m_ExNetParams.m_OutputDetailsOnlyToStdOut)->default_value("false")
+                                                                                ->implicit_value("true"));
+
     }
     catch (const std::exception& e)
     {
diff --git a/tests/InferenceModel.hpp b/tests/InferenceModel.hpp
index 1db287f453..b982df396d 100644
--- a/tests/InferenceModel.hpp
+++ b/tests/InferenceModel.hpp
@@ -101,6 +101,7 @@ struct Params
     bool                            m_EnableFastMath;
     bool                            m_SaveCachedNetwork;
     bool                            m_OutputDetailsToStdOut;
+    bool                            m_OutputDetailsOnlyToStdOut;
     std::string                     m_CachedNetworkFilePath;
     unsigned int                    m_NumberOfThreads;
     std::string                     m_MLGOTuningFilePath;
@@ -121,6 +122,7 @@ struct Params
         , m_EnableFastMath(false)
         , m_SaveCachedNetwork(false)
         , m_OutputDetailsToStdOut(false)
+        , m_OutputDetailsOnlyToStdOut(false)
         , m_CachedNetworkFilePath("")
         , m_NumberOfThreads(0)
         , m_MLGOTuningFilePath("")
@@ -406,7 +408,8 @@ public:
                    bool enableProfiling,
                    const std::string& dynamicBackendsPath,
                    const std::shared_ptr<armnn::IRuntime>& runtime = nullptr)
-        : m_EnableProfiling(enableProfiling)
+        : m_EnableProfiling(enableProfiling),
+          m_ProfilingDetailsMethod(armnn::ProfilingDetailsMethod::Undefined)
         , m_DynamicBackendsPath(dynamicBackendsPath)
     {
         if (runtime)
@@ -421,6 +424,12 @@ public:
             m_Runtime = armnn::IRuntime::Create(options);
         }
 
+        // Configure the Profiler if the the profiling details are opted for
+        if (params.m_OutputDetailsOnlyToStdOut)
+            m_ProfilingDetailsMethod = armnn::ProfilingDetailsMethod::DetailsOnly;
+        else if (params.m_OutputDetailsToStdOut)
+            m_ProfilingDetailsMethod = armnn::ProfilingDetailsMethod::DetailsWithEvents;
+
         std::string invalidBackends;
         if (!CheckRequestedBackendsAreValid(params.m_ComputeDevices, armnn::Optional<std::string&>(invalidBackends)))
         {
@@ -492,7 +501,7 @@ public:
                                                         armnn::MemorySource::Undefined,
                                                         armnn::MemorySource::Undefined,
                                                         enableProfiling,
-                                                        params.m_OutputDetailsToStdOut);
+                                                        m_ProfilingDetailsMethod);
             std::string errorMessage;
             ret = m_Runtime->LoadNetwork(m_NetworkIdentifier, std::move(optNet), errorMessage, networkProperties);
 
@@ -744,6 +753,7 @@ private:
     std::vector<armnn::BindingPointInfo> m_InputBindings;
     std::vector<armnn::BindingPointInfo> m_OutputBindings;
     bool m_EnableProfiling;
+    armnn::ProfilingDetailsMethod m_ProfilingDetailsMethod;
     std::string m_DynamicBackendsPath;
 
     template<typename TContainer>