backend_performance.cpp

// ----------------------------------------------------------------------------
// Copyright 2017 Nervana Systems Inc.
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
//      http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// ----------------------------------------------------------------------------

#include <sstream>
#include <string>
#include <vector>

#include "gtest/gtest.h"

#include "ngraph/codegen/compiler.hpp"
#include "ngraph/codegen/execution_engine.hpp"
#include "ngraph/file_util.hpp"
#include "ngraph/log.hpp"
#include "ngraph/ops/concatenate.hpp"
#include "ngraph/runtime/backend.hpp"
#include "ngraph/runtime/call_frame.hpp"
#include "ngraph/runtime/cpu/cpu_call_frame.hpp"
#include "ngraph/runtime/manager.hpp"
#include "ngraph/serializer.hpp"
#include "ngraph/util.hpp"
#include "util/random.hpp"
#include "util/test_tools.hpp"

using namespace std;
using namespace ngraph;

static multimap<size_t, string>
    agregate_timing(const vector<runtime::PerformanceCounter>& perf_data)
{
    unordered_map<string, size_t> timing;
    for (const runtime::PerformanceCounter& p : perf_data)
    {
        string op = p.name().substr(0, p.name().find('_'));
        timing[op] += p.microseconds();
    }

    multimap<size_t, string> rc;
    for (const pair<string, size_t>& t : timing)
    {
        rc.insert({t.second, t.first});
    }
    return rc;
}

void run_benchmark(const string& json_path, const string& backend_name, size_t iterations)
{
    string env_var_name = "NGRAPH_" + backend_name + "_EMIT_TIMING";
    bool emit_timing = (std::getenv(env_var_name.c_str()) != nullptr);
    if (!emit_timing)
    {
        cout << "To get per-op timing set the environment variable " << env_var_name << "\n";
    }

    test::Uniform<float> rng{-1, 1, 0};
    const string json_string = file_util::read_file_to_string(json_path);
    stringstream ss(json_string);
    shared_ptr<Function> f = ngraph::deserialize(ss);

    stopwatch build_time;
    build_time.start();
    auto manager = runtime::Manager::get(backend_name);
    auto external = manager->compile(f);
    auto backend = manager->allocate_backend();
    auto cf = backend->make_call_frame(external);
    build_time.stop();
    cout << "build_time " << build_time.get_milliseconds() << "ms" << endl;

    vector<shared_ptr<runtime::TensorView>> args;
    for (shared_ptr<op::Parameter> param : f->get_parameters())
    {
        auto tensor =
            backend->make_primary_tensor_view(param->get_element_type(), param->get_shape());
        rng.initialize(tensor);
        args.push_back(tensor);
    }
    vector<shared_ptr<runtime::TensorView>> results;
    for (shared_ptr<Node> out : f->get_results())
    {
        auto result = backend->make_primary_tensor_view(out->get_element_type(), out->get_shape());
        results.push_back(result);
    }

    stopwatch t1;
    t1.start();
    for (size_t i = 0; i < static_cast<size_t>(iterations); i++)
    {
        cf->tensor_call(args, results);
    }
    t1.stop();
    float time = t1.get_milliseconds();
    cout << time / iterations << "ms per iteration" << endl;

    vector<runtime::PerformanceCounter> perf_data = cf->get_performance_data();
    sort(perf_data.begin(),
         perf_data.end(),
         [](const runtime::PerformanceCounter& p1, const runtime::PerformanceCounter& p2) {
             return p1.total_microseconds() > p2.total_microseconds();
         });
    multimap<size_t, string> timing = agregate_timing(perf_data);
    for (auto it = timing.rbegin(); it != timing.rend(); it++)
    {
        cout.imbue(locale(""));
        cout << setw(15) << left << it->second << " " << setw(10) << right << it->first << "us\n";
    }
}

TEST(benchmark, mxnet_mnist_mlp_forward)
{
    const string json_path = file_util::path_join(SERIALIZED_ZOO, "mxnet/mnist_mlp_forward.json");
    run_benchmark(json_path, "CPU", 1000);
}

TEST(benchmark, mxnet_10_bucket_lstm)
{
    const string json_path = file_util::path_join(SERIALIZED_ZOO, "mxnet/10_bucket_LSTM.json");
    run_benchmark(json_path, "CPU", 10);
}

TEST(benchmark, mxnet_lstm_backward)
{
    const string json_path = file_util::path_join(SERIALIZED_ZOO, "mxnet/LSTM_backward.json");
    run_benchmark(json_path, "CPU", 10);
}

TEST(benchmark, mxnet_lstm_forward)
{
    const string json_path = file_util::path_join(SERIALIZED_ZOO, "mxnet/LSTM_forward.json");
    run_benchmark(json_path, "CPU", 10);
}

TEST(benchmark, mxnet_seq2seq_forward)
{
    const string json_path = file_util::path_join(SERIALIZED_ZOO, "mxnet/Seq2Seq_forward.json");
    run_benchmark(json_path, "CPU", 10);
}

TEST(benchmark, mxnet_seq2seq_backward)
{
    const string json_path = file_util::path_join(SERIALIZED_ZOO, "mxnet/Seq2Seq_backward.json");
    run_benchmark(json_path, "CPU", 10);
}

//
// Benchmarks a graph that concatenates six 32x1x200 arrays along the middle axis.
//
TEST(benchmark, concat_32x1x200_axis1_6)
{
    const size_t n_arrays = 6;
    Shape shape_of_each_array = Shape{32, 1, 200};
    size_t concatenation_axis = 1;

    Shape result_shape;
    result_shape = shape_of_each_array;
    result_shape[concatenation_axis] *= n_arrays;

    size_t elements_per_array = 1;
    for (size_t d : shape_of_each_array)
    {
        elements_per_array *= d;
    }

    vector<vector<float>> data_arrays(n_arrays);
    for (size_t i = 0; i < n_arrays; i++)
    {
        data_arrays[i] = vector<float>(elements_per_array);
        for (size_t j = 0; j < elements_per_array; j++)
        {
            data_arrays[i][j] = float(j + 1);
        }
    }

    bool using_ref_kernels = (std::getenv("NGRAPH_CPU_USE_REF_KERNELS") != nullptr);

    vector<std::string> backend_names{"INTERPRETER", "CPU"};
    vector<int> n_runs{200, 200, using_ref_kernels ? 200 : 200000}; // one for each backend
    vector<std::function<void()>> test_callbacks;                   // one for each backend
    vector<std::shared_ptr<runtime::TensorView>> result_tvs;        // one for each backend

    for (std::string backend_name : backend_names)
    {
        vector<std::shared_ptr<op::Parameter>> params(n_arrays);
        vector<std::shared_ptr<Node>> params_as_nodes(n_arrays);
        for (size_t i = 0; i < n_arrays; i++)
        {
            auto param = make_shared<op::Parameter>(element::f32, shape_of_each_array);
            params[i] = param;
            params_as_nodes[i] = param;
        }

        auto concat = make_shared<op::Concat>(params_as_nodes, concatenation_axis);
        auto f = make_shared<Function>(concat, params);

        auto manager = runtime::Manager::get(backend_name);
        auto external = manager->compile(f);
        auto backend = manager->allocate_backend();
        auto cf = backend->make_call_frame(external);

        vector<shared_ptr<runtime::TensorView>> input_vals;

        for (size_t i = 0; i < n_arrays; i++)
        {
            auto tv = backend->make_primary_tensor_view(element::f32, shape_of_each_array);
            copy_data(tv, data_arrays[i]);
            input_vals.push_back(tv);
        }

        auto result_tv = backend->make_primary_tensor_view(element::f32, result_shape);
        result_tvs.push_back(result_tv);

        std::function<void()> cb = [input_vals, result_tv, cf]() {
            cf->call(input_vals, {result_tv});
        };

        test_callbacks.push_back(cb);
    }

    for (size_t i = 0; i < backend_names.size(); i++)
    {
        std::cout << backend_names[i] << ": " << n_runs[i] << " tests in " << std::flush;

        stopwatch sw;
        std::function<void()> cb = test_callbacks[i];

        sw.start();
        for (int j = 0; j < n_runs[i]; j++)
        {
            cb();
        }
        sw.stop();

        std::cout << sw.get_milliseconds() << "ms (" << (sw.get_microseconds() / n_runs[i])
                  << " us/test)" << std::endl;
    }

    for (size_t i = 1; i < backend_names.size(); i++)
    {
        std::cout << "Verifying " << backend_names[i] << " result against " << backend_names[0]
                  << "..." << std::flush;

        if (read_vector<float>(result_tvs[i]) == read_vector<float>(result_tvs[0]))
        {
            std::cout << " OK" << std::endl;
        }
        else
        {
            std::cout << " FAILED" << std::endl;
            ADD_FAILURE();
        }
    }
}