Jungfraujoch/tools/jfjoch_resonet_test.cpp

// SPDX-FileCopyrightText: 2025 Filip Leonarski, Paul Scherrer Institute <filip.leonarski@psi.ch>
// SPDX-License-Identifier: GPL-3.0-only

#include <fstream>
#include <chrono>
#include <future>
#include <cmath>

#include <getopt.h>
#include <httplib/httplib.h>

#include "../image_analysis/NeuralNetInferenceClient.h"
#include "../writer/HDF5Objects.h"
#include "../common/print_license.h"

bool verbose = false;
int64_t niter = 1000;
size_t nthreads = 1;
int nquads = 4;

Logger logger("jfjoch_resonet_test");

void print_usage() {
    logger.Info("Usage ./jfjoch_resonet_test {options}");
    logger.Info("");
    logger.Info("Available options:");
    logger.Info("--verbose/-v                       Verbose");
    logger.Info("--iterations=<int>/-I<int>         Number of iterations (default: 1000)");
    logger.Info("--threads=<int>/-N<int>            Number of threads (default: 1)");
    logger.Info("--quads=<int>/-q<int>              Number of quads 1-4 (default: 4)");
    logger.Info("");
}

int parse_options(int argc, char **argv) {

    int c;
    static struct option long_options[] = {
            {"verbose",    no_argument, 0, 'v'},
            {"iterations",    required_argument, 0, 'I'},
            {"threads",    required_argument, 0, 'N'},
            {"quads",    required_argument, 0, 'q'},
            {"help",    no_argument, 0, 'h'},
            {0, 0, 0, 0}
    };

    int option_index = 0;
    int opt;
    while ((opt = getopt_long(argc, argv, "q:I:N:vh",long_options, &option_index)) != -1 ) {
        switch (opt) {
            case 'q':
                nquads = atol(optarg);
                if ((nquads < 1) || (nquads > 4)){
                    logger.Info("Number of quads must be in range from 1 to 4");
                    exit(EXIT_FAILURE);
                }
                break;
            case 'I':
                niter = atol(optarg);
                if (niter <= 0) {
                    logger.Info("Number of iterations requires positive number");
                    exit(EXIT_FAILURE);
                }
                break;
            case 'N':
                nthreads = atol(optarg);
                if (nthreads <= 0) {
                    logger.Info("Number of threads requires positive number");
                    exit(EXIT_FAILURE);
                }
                break;
            case 'v':
                verbose = true;
                break;
            case 'h':
                print_usage();
                exit(EXIT_SUCCESS);
            default:
                logger.Error("Unknown option {}", opt);
                exit(EXIT_FAILURE);
        }
    }
    return optind;
}


int main(int argc, char **argv) {
    print_license("jfjoch_resonet_test");

    RegisterHDF5Filter();

    size_t first_argc = parse_options(argc, argv);

    logger.Verbose(verbose);

    DiffractionExperiment experiment(DetJF4M());
    experiment.DetectorDistance_mm(75).IncidentEnergy_keV(12.4).BeamY_pxl(1136).BeamX_pxl(1090);

    NeuralNetInferenceClient predictor;
    for (size_t i = first_argc; i < argc; i++) {
        try {
            predictor.AddHost(argv[i]);
            logger.Info("Added host {}", argv[i]);
        } catch (std::exception &e) {
            logger.ErrorException(e);
            exit(EXIT_FAILURE);
        }
    }

    predictor.AddLogger(&logger);

    if (predictor.GetHostCount() == 0) {
        logger.Info("Default localhost:8000 for inference.");
        predictor.AddHost("localhost", 8000);
    }

    HDF5ReadOnlyFile data("../../tests/test_data/compression_benchmark.h5");
    HDF5DataSet dataset(data, "/entry/data/data");
    HDF5DataSpace file_space(dataset);

    size_t xpixel = file_space.GetDimensions()[2];
    size_t ypixel = file_space.GetDimensions()[1];
    size_t nimages = file_space.GetDimensions()[0];
    std::vector<int16_t> image_conv (nimages * xpixel * ypixel);

    std::vector<hsize_t> start = {0,0,0};
    std::vector<hsize_t> file_size = {nimages, ypixel, xpixel};
    dataset.ReadVector(image_conv, start, file_size);

    logger.Info("Dimension {:d}x{:d} pxl, max pooling factor {:d}",
                xpixel, ypixel, predictor.GetMaxPoolFactor(experiment));

    // Start time
    auto start_time = std::chrono::high_resolution_clock::now();

    std::atomic<int64_t> done = 0;

    std::vector<std::future<void>> futures;
    for (int t = 0; t < nthreads; t++) {
        futures.emplace_back(std::async(std::launch::async, [&] {
            for (int i = 0; i < niter; i += nthreads) {
                auto iter_start_time = std::chrono::high_resolution_clock::now();

                size_t image_number = i % nimages;
                const auto image = image_conv.data() + image_number * xpixel * ypixel;
                auto val = predictor.Inference(experiment, image, nquads);

                auto iter_end_time = std::chrono::high_resolution_clock::now();
                std::chrono::duration<double> iter_diff = iter_end_time - iter_start_time;
                auto iter_diff_ms = std::lround(iter_diff.count() * 1000.0);

                if (val) {
                    done++;
                    logger.Debug("Res: {:2.f} Duration {:d} ms", val.value(), iter_diff_ms);
                } else
                    logger.Warning("Missing results");
            }
        }));
    }

    for (auto &f: futures)
        f.get();

    // End time
    auto end = std::chrono::high_resolution_clock::now();

    // Calculate duration
    std::chrono::duration<double> diff = end - start_time;

    // Output time taken
    logger.Info("Total performance {:.1f} Hz (iteration: {:d}; {:.1f}%)", niter / diff.count(), niter,
                std::floor(done / static_cast<double>(niter)) * 100.0);
}