perftest.cpp

#define OPENCL_PLATFORM_ID 0
#define OPENCL_DEVICE_ID 0
#define CL_USE_DEPRECATED_OPENCL_1_1_APIS
//#define USE_TIMING_OUTPUTS

//#include <CL/cl.hpp>
#include <Eigen/Dense>
#include <stan/math.hpp>
#include <vector>
#include <iostream>
#include <chrono>


using Eigen::Matrix;
using Eigen::Dynamic;

using namespace std;
using namespace stan::math;


//max ~ 40k * 1k
const int CASES = 20000;
const int ATTRS = 5000;
const int REPEATS = 10;

template<typename T>
void perf_test(T func, int cases = CASES, int attrs = ATTRS, int repeats = REPEATS) {
    volatile Matrix<double, Dynamic, 1> y_ = Matrix<double, Dynamic, 1>::Random(cases);
    volatile double alpha_ = -2;
    volatile Matrix<double, Dynamic, 1> beta_ = Matrix<double, Dynamic, 1>::Random(attrs);
    volatile Matrix<double, Dynamic, Dynamic> x_ = Matrix<double, Dynamic, Dynamic>::Random(cases, attrs);
    volatile double sigma_ = 2;


    Matrix<double, Dynamic, 1> y = *const_cast<Matrix<double, Dynamic, 1>*>(&y_);
    Matrix<double, Dynamic, Dynamic> x = *const_cast<Matrix<double, Dynamic, Dynamic>*>(&x_);
    auto start = std::chrono::steady_clock::now();

    for (int i = 0; i < repeats; i++) {

        var alpha = /*reinterpret_cast<var>*/(alpha_);
        Matrix<var, Dynamic, 1> beta = *const_cast<Matrix<double, Dynamic, 1>*>(&beta_);
        var sigma = /*reinterpret_cast<var>*/(sigma_);
        vector<var> grad_args(beta.data(), beta.data() + beta.size());
        grad_args.push_back(alpha);
        grad_args.push_back(sigma);

        var target = func(y, x, alpha, beta, sigma);
        volatile double no_opt = target.val();
        vector<double> grad;
        target.grad(grad_args, grad);
        for (double val : grad) {
            no_opt = val;
        }
        recover_memory();
    }
    auto end = std::chrono::steady_clock::now();

    cout << "Total time of execution: "
         << std::chrono::duration_cast<std::chrono::milliseconds>(end - start).count()
         << "ms" << endl;

    recover_memory();
    ChainableStack::instance().memalloc_.free_all();

}

int main() {
    using Eigen::Matrix;
    using Eigen::Dynamic;

    /*
    cout << "naive implementation:" << endl;
    perf_test(
            normal_id_glm_lpdf_basic<false, Matrix<double, Dynamic, 1>, Matrix<double, Dynamic, Dynamic>, var, Matrix<var, Dynamic, 1>, var>);
    cout << ":" << endl;
    perf_test(
            normal_id_glm_lpdf<false, Matrix<double, Dynamic, 1>, Matrix<double, Dynamic, Dynamic>, var, Matrix<var, Dynamic, 1>, var>);
    cout << "typed implementation with operands_and_partials:" << endl;
    perf_test(normal_id_glm_lpdf_opt_typed);
    /*cout << "opt:" << endl;
    perf_test(
            normal_id_glm_lpdf_opt<false, Matrix<double, Dynamic, 1>, Matrix<double, Dynamic, Dynamic>, var, Matrix<var, Dynamic, 1>, var>);
*/
    cout << "library implementation:" << endl;
    perf_test(
            normal_id_glm_lpdf<false, Matrix<double, Dynamic, 1>, Matrix<double, Dynamic, Dynamic>, var, Matrix<var, Dynamic, 1>, var>);
    //    cout << "opt_typed:" << endl;
    //    perf_test(normal_id_glm_lpdf_opt_typed);
    /*cout << "opt:" << endl;
    perf_test(
            normal_id_glm_lpdf_opt<false, Matrix<double, Dynamic, 1>, Matrix<double, Dynamic, Dynamic>, var, Matrix<var, Dynamic, 1>, var>);
*/
};