guide for measuring performance (#676)

2021-08-17 10:50:43 +08:00 · 2021-08-17 10:50:43 +08:00 · ec20cead6f
commit ec20cead6f
parent ccdf0e6316
1 changed files with 113 additions and 0 deletions
--- a/tutorials/measure_performance.md
+++ b/tutorials/measure_performance.md
@ -0,0 +1,113 @@
+# Measure performance of TensorRT
+
+## 1. add some variables and structures
+
+see https://github.com/NVIDIA/TensorRT/tree/master/samples/sampleNMT for more detail.
+
+```c++
+// for rcnn, you can put these code into common.hpp
+#include "logging.h" // rcnn/logging.h
+static Logger gLogger{ Logger::Severity::kINFO };
+static LogStreamConsumer gLogInfo{ LOG_INFO(gLogger) };
+
+struct SimpleProfiler : public nvinfer1::IProfiler
+{
+    struct Record
+    {
+        float time{ 0 };
+        int count{ 0 };
+    };
+
+    virtual void reportLayerTime(const char* layerName, float ms)
+    {
+        mProfile[layerName].count++;
+        mProfile[layerName].time += ms;
+        if (std::find(mLayerNames.begin(), mLayerNames.end(), layerName) == mLayerNames.end())
+        {
+            mLayerNames.push_back(layerName);
+        }
+    }
+
+    SimpleProfiler(const char* name, const std::vector<SimpleProfiler>& srcProfilers = std::vector<SimpleProfiler>())
+        : mName(name)
+    {
+        for (const auto& srcProfiler : srcProfilers)
+        {
+            for (const auto& rec : srcProfiler.mProfile)
+            {
+                auto it = mProfile.find(rec.first);
+                if (it == mProfile.end())
+                {
+                    mProfile.insert(rec);
+                }
+                else
+                {
+                    it->second.time += rec.second.time;
+                    it->second.count += rec.second.count;
+                }
+            }
+        }
+    }
+
+    friend std::ostream& operator<<(std::ostream& out, const SimpleProfiler& value)
+    {
+        out << "========== " << value.mName << " profile ==========" << std::endl;
+        float totalTime = 0;
+        std::string layerNameStr = "TensorRT layer name";
+        int maxLayerNameLength = std::max(static_cast<int>(layerNameStr.size()), 70);
+        for (const auto& elem : value.mProfile)
+        {
+            totalTime += elem.second.time;
+            maxLayerNameLength = std::max(maxLayerNameLength, static_cast<int>(elem.first.size()));
+        }
+
+        auto old_settings = out.flags();
+        auto old_precision = out.precision();
+        // Output header
+        {
+            out << std::setw(maxLayerNameLength) << layerNameStr << " ";
+            out << std::setw(12) << "Runtime, "
+                << "%"
+                << " ";
+            out << std::setw(12) << "Invocations"
+                << " ";
+            out << std::setw(12) << "Runtime, ms" << std::endl;
+        }
+        for (size_t i = 0; i < value.mLayerNames.size(); i++)
+        {
+            const std::string layerName = value.mLayerNames[i];
+            auto elem = value.mProfile.at(layerName);
+            out << std::setw(maxLayerNameLength) << layerName << " ";
+            out << std::setw(12) << std::fixed << std::setprecision(1) << (elem.time * 100.0F / totalTime) << "%"
+                << " ";
+            out << std::setw(12) << elem.count << " ";
+            out << std::setw(12) << std::fixed << std::setprecision(2) << elem.time << std::endl;
+        }
+        out.flags(old_settings);
+        out.precision(old_precision);
+        out << "========== " << value.mName << " total runtime = " << totalTime << " ms ==========" << std::endl;
+
+        return out;
+    }
+
+private:
+    std::string mName;
+    std::vector<std::string> mLayerNames;
+    std::map<std::string, Record> mProfile;
+};
+```
+
+
+
+## 2. set profiler for context and print the log
+
+```c++
+// you'd better set name for every layers
+// build engine
+// build context
+auto sp = SimpleProfiler("test");
+context->setProfiler(&sp);
+context->enqueue(...);
+gLogInfo << sp << std::endl;
+```
+