Skip to content
Branch: master
Find file History
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Type Name Latest commit message Commit time
Failed to load latest commit information.

JavaCPP Presets for TensorRT

License Agreements

By downloading these archives, you agree to the terms of the license agreements for NVIDIA software included in the archives.


To view the license for TensorRT included in these archives, click here


This directory contains the JavaCPP Presets module for:

Please refer to the parent file for more detailed information about the JavaCPP Presets.


Java API documentation is available here:

Sample Usage

Here is a simple example of TensorRT ported to Java from the sampleGoogleNet.cpp sample file included in TensorRT- available at:

We can use Maven 3 to download and install automatically all the class files as well as the native binaries. To run this sample code, after creating the pom.xml and source files below, simply execute on the command line:

 $ mvn compile exec:java

The pom.xml build file


        <!-- Additional dependencies to use bundled CUDA, cuDNN, NCCL, and TensorRT -->


The source file

import java.util.*;
import org.bytedeco.javacpp.*;

import org.bytedeco.cuda.cudart.*;
import org.bytedeco.tensorrt.nvinfer.*;
import org.bytedeco.tensorrt.nvparsers.*;
import static*;
import static*;
import static*;

public class SampleGoogleNet {
    static void CHECK(int status)
        if (status != 0)
            System.out.println("Cuda failure: " + status);

    // Logger for GIE info/warning/errors
    static class Logger extends ILogger
        @Override public void log(Severity severity, String msg)
            severity = severity.intern();

            // suppress info-level messages
            if (severity == Severity.kINFO) return;

            switch (severity)
                case kINTERNAL_ERROR: System.err.print("INTERNAL_ERROR: "); break;
                case kERROR: System.err.print("ERROR: "); break;
                case kWARNING: System.err.print("WARNING: "); break;
                case kINFO: System.err.print("INFO: "); break;
                default: System.err.print("UNKNOWN: "); break;
    static Logger gLogger = new Logger();

    static String locateFile(String input, String[] directories)
        String file = "";
        int MAX_DEPTH = 10;
        boolean found = false;
        for (String dir : directories)
            file = dir + input;
            for (int i = 0; i < MAX_DEPTH && !found; i++)
                File checkFile = new File(file);
                found = checkFile.exists();
                if (found) break;
                file = "../" + file;
            if (found) break;
            file = "";

        if (file.isEmpty())
            System.err.println("Could not find a file due to it not existing in the data directory.");
        return file;

    // stuff we know about the network and the caffe input/output blobs

    static int BATCH_SIZE = 4;
    static int TIMING_ITERATIONS = 1000;

    static String INPUT_BLOB_NAME = "data";
    static String OUTPUT_BLOB_NAME = "prob";

    static String locateFile(String input)
        String[] dirs = {"data/samples/googlenet/", "data/googlenet/"};
        return locateFile(input, dirs);

    static class Profiler extends IProfiler
        LinkedHashMap<String, Float> mProfile = new LinkedHashMap<String, Float>();

        @Override public void reportLayerTime(String layerName, float ms)
            Float time = mProfile.get(layerName);
            mProfile.put(layerName, (time != null ? time : 0) + ms);

        public void printLayerTimes()
            float totalTime = 0;
            for (Map.Entry<String,Float> e : mProfile.entrySet())
                System.out.printf("%-40.40s %4.3fms\n", e.getKey(), e.getValue() / TIMING_ITERATIONS);
                totalTime += e.getValue();
            System.out.printf("Time over all layers: %4.3f\n", totalTime / TIMING_ITERATIONS);

    static Profiler gProfiler = new Profiler();

    static void caffeToGIEModel(String deployFile,     // name for caffe prototxt
                         String modelFile,             // name for model 
                         String[] outputs,             // network outputs
                         int maxBatchSize,             // batch size - NB must be at least as large as the batch we want to run with)
                         IHostMemory[] gieModelStream)
        // create API root class - must span the lifetime of the engine usage
        IBuilder builder = createInferBuilder(gLogger);
        INetworkDefinition network = builder.createNetwork();

        // parse the caffe model to populate the network, then set the outputs
        ICaffeParser parser = createCaffeParser();

        boolean useFp16 = builder.platformHasFastFp16();

        DataType modelDataType = useFp16 ? DataType.kHALF : DataType.kFLOAT; // create a 16-bit model if it's natively supported
        IBlobNameToTensor blobNameToTensor =
            parser.parse(locateFile(deployFile),                // caffe deploy file
                                     locateFile(modelFile),     // caffe model file
                                     network,                   // network definition that the parser will populate

        assert blobNameToTensor != null;
        // the caffe file has no notion of outputs, so we need to manually say which tensors the engine should generate    
        for (String s : outputs)

        // Build the engine
        builder.setMaxWorkspaceSize(16 << 20);

        // set up the network for paired-fp16 format if available

        ICudaEngine engine = builder.buildCudaEngine(network);
        assert engine != null;

        // we don't need the network any more, and we can destroy the parser

        // serialize the engine, then close everything down
        gieModelStream[0] = engine.serialize();

    static void timeInference(ICudaEngine engine, int batchSize)
        // input and output buffer pointers that we pass to the engine - the engine requires exactly ICudaEngine::getNbBindings(),
        // of these, but in this case we know that there is exactly one input and one output.
        assert engine.getNbBindings() == 2;
        PointerPointer buffers = new PointerPointer(2);

        // In order to bind the buffers, we need to know the names of the input and output tensors.
        // note that indices are guaranteed to be less than ICudaEngine::getNbBindings()
        int inputIndex = engine.getBindingIndex(INPUT_BLOB_NAME), outputIndex = engine.getBindingIndex(OUTPUT_BLOB_NAME);

        // allocate GPU buffers
        DimsCHW inputDims = new DimsCHW(engine.getBindingDimensions(inputIndex)), outputDims = new DimsCHW(engine.getBindingDimensions(outputIndex));
        long inputSize = batchSize * inputDims.c().get() * inputDims.h().get() * inputDims.w().get() * Float.SIZE / 8;
        long outputSize = batchSize * outputDims.c().get() * outputDims.h().get() * outputDims.w().get() * Float.SIZE / 8;

        CHECK(cudaMalloc(buffers.position(inputIndex), inputSize));
        CHECK(cudaMalloc(buffers.position(outputIndex), outputSize));

        IExecutionContext context = engine.createExecutionContext();

        // zero the input buffer
        CHECK(cudaMemset(buffers.position(inputIndex).get(), 0, inputSize));

        for (int i = 0; i < TIMING_ITERATIONS;i++)
            context.execute(batchSize, buffers.position(0));

        // release the context and buffers

    public static void main(String[] args)
        System.out.println("Building and running a GPU inference engine for GoogleNet, N=4...");

        // parse the caffe model and the mean file
        IHostMemory[] gieModelStream = { null };
        caffeToGIEModel("googlenet.prototxt", "googlenet.caffemodel", new String[] { OUTPUT_BLOB_NAME }, BATCH_SIZE, gieModelStream);

        // create an engine
        IRuntime infer = createInferRuntime(gLogger);
        ICudaEngine engine = infer.deserializeCudaEngine(gieModelStream[0].data(), gieModelStream[0].size(), null);

        System.out.println("Bindings after deserializing:"); 
        for (int bi = 0; bi < engine.getNbBindings(); bi++) { 
            if (engine.bindingIsInput(bi)) { 
                System.out.printf("Binding %d (%s): Input.\n",  bi, engine.getBindingName(bi));
            } else { 
                System.out.printf("Binding %d (%s): Output.\n", bi, engine.getBindingName(bi));

        // run inference with null data to time network performance
        timeInference(engine, BATCH_SIZE);




You can’t perform that action at this time.