我正在比較各種線性代數運算中Eigen(v3.2.8)和GSL(v2.1)庫的性能。雖然在大多數操作中,Eigen大幅度獲勝(少數因素),但在計算奇異值分解時,它滯後。我使用SVD的兩個實現 - JacobiSVD和BDCSVD(後者仍處於「不支持」模塊,這個版本),並測試了下面的代碼在幾個平臺:本徵庫:與GSL相比,SVD較慢
#include <gsl/gsl_rng.h>
#include <gsl/gsl_linalg.h>
#include <unsupported/Eigen/SVD>
#include <ctime>
#include <iostream>
void testSVD(const int size)
{
clock_t tbegin;
gsl_rng* rng = gsl_rng_alloc(gsl_rng_default);
gsl_rng_set(rng, 42);
Eigen::MatrixXd mat(size, size);
for(int i=0; i<size; i++)
for(int j=0; j<size; j++)
mat(i, j) = gsl_rng_uniform(rng);
gsl_matrix_view matv = gsl_matrix_view_array(mat.data(), size, size);
gsl_matrix* mat1 = &matv.matrix;
gsl_matrix* mat2 = gsl_matrix_alloc(size, size);
gsl_vector* vec1 = gsl_vector_alloc(size);
gsl_vector* vec2 = gsl_vector_alloc(size);
tbegin = std::clock();
Eigen::JacobiSVD<Eigen::MatrixXd> jac(mat, Eigen::ComputeThinU | Eigen::ComputeThinV);
double tjac = (std::clock() - tbegin)*1.0/CLOCKS_PER_SEC;
tbegin = std::clock();
Eigen::BDCSVD<Eigen::MatrixXd> bdc(mat, Eigen::ComputeThinU | Eigen::ComputeThinV);
double tbdc = (std::clock() - tbegin)*1.0/CLOCKS_PER_SEC;
tbegin = std::clock();
gsl_linalg_SV_decomp(mat1, mat2, vec1, vec2);
double tgsl = (std::clock() - tbegin)*1.0/CLOCKS_PER_SEC;
std::cout << "Matrix size: "<<size<<", GSL SVD: "<<tgsl<<" s, JacobiSVD: "<<tjac<<" s, BDCSVD: "<<tbdc<<" s\n";
gsl_vector_free(vec2);
gsl_vector_free(vec1);
gsl_matrix_free(mat2);
}
機1(舊,GCC 4.2 0.1,編譯選項-O3):
Matrix size: 125, GSL SVD: 0.054497 s, JacobiSVD: 0.143428 s, BDCSVD: 0.177061 s Matrix size: 250, GSL SVD: 0.457591 s, JacobiSVD: 1.16189 s, BDCSVD: 1.291 s Matrix size: 500, GSL SVD: 6.01799 s, JacobiSVD: 15.2217 s, BDCSVD: 12.1353 s Matrix size: 1000, GSL SVD: 67.3727 s, JacobiSVD: 155.501 s, BDCSVD: 114.372 s
機2(GCC 4.8.3,-O3):
Matrix size: 125, GSL SVD: 0.008666 s, JacobiSVD: 0.029863 s, BDCSVD: 0.026779 s Matrix size: 250, GSL SVD: 0.082332 s, JacobiSVD: 0.318304 s, BDCSVD: 0.235657 s Matrix size: 500, GSL SVD: 0.807378 s, JacobiSVD: 2.93758 s, BDCSVD: 1.98003 s Matrix size: 1000, GSL SVD: 23.1846 s, JacobiSVD: 59.1015 s, BDCSVD: 27.8348 s
機3(ICC 13.1.3,-O3):
Matrix size: 125, GSL SVD: 0.02 s, JacobiSVD: 0.05 s, BDCSVD: 0.04 s Matrix size: 250, GSL SVD: 0.15 s, JacobiSVD: 0.4 s, BDCSVD: 0.28 s Matrix size: 500, GSL SVD: 2.18 s, JacobiSVD: 3.51 s, BDCSVD: 2.41 s Matrix size: 1000, GSL SVD: 10 s, JacobiSVD: 42.87 s, BDCSVD: 23.4 s
問題是,爲什麼Eigen不擊敗GSL?
如果一些答案幫你解決問題,不要忘了「接受」它。 –