xref: /aosp_15_r20/external/eigen/bench/benchCholesky.cpp (revision bf2c37156dfe67e5dfebd6d394bad8b2ab5804d4)
1 // g++ -DNDEBUG -O3 -I.. benchCholesky.cpp  -o benchCholesky && ./benchCholesky
2 // options:
3 //  -DBENCH_GSL -lgsl /usr/lib/libcblas.so.3
4 //  -DEIGEN_DONT_VECTORIZE
5 //  -msse2
6 //  -DREPEAT=100
7 //  -DTRIES=10
8 //  -DSCALAR=double
9 
10 #include <iostream>
11 
12 #include <Eigen/Core>
13 #include <Eigen/Cholesky>
14 #include <bench/BenchUtil.h>
15 using namespace Eigen;
16 
17 #ifndef REPEAT
18 #define REPEAT 10000
19 #endif
20 
21 #ifndef TRIES
22 #define TRIES 10
23 #endif
24 
25 typedef float Scalar;
26 
27 template <typename MatrixType>
benchLLT(const MatrixType & m)28 __attribute__ ((noinline)) void benchLLT(const MatrixType& m)
29 {
30   int rows = m.rows();
31   int cols = m.cols();
32 
33   double cost = 0;
34   for (int j=0; j<rows; ++j)
35   {
36     int r = std::max(rows - j -1,0);
37     cost += 2*(r*j+r+j);
38   }
39 
40   int repeats = (REPEAT*1000)/(rows*rows);
41 
42   typedef typename MatrixType::Scalar Scalar;
43   typedef Matrix<Scalar, MatrixType::RowsAtCompileTime, MatrixType::RowsAtCompileTime> SquareMatrixType;
44 
45   MatrixType a = MatrixType::Random(rows,cols);
46   SquareMatrixType covMat =  a * a.adjoint();
47 
48   BenchTimer timerNoSqrt, timerSqrt;
49 
50   Scalar acc = 0;
51   int r = internal::random<int>(0,covMat.rows()-1);
52   int c = internal::random<int>(0,covMat.cols()-1);
53   for (int t=0; t<TRIES; ++t)
54   {
55     timerNoSqrt.start();
56     for (int k=0; k<repeats; ++k)
57     {
58       LDLT<SquareMatrixType> cholnosqrt(covMat);
59       acc += cholnosqrt.matrixL().coeff(r,c);
60     }
61     timerNoSqrt.stop();
62   }
63 
64   for (int t=0; t<TRIES; ++t)
65   {
66     timerSqrt.start();
67     for (int k=0; k<repeats; ++k)
68     {
69       LLT<SquareMatrixType> chol(covMat);
70       acc += chol.matrixL().coeff(r,c);
71     }
72     timerSqrt.stop();
73   }
74 
75   if (MatrixType::RowsAtCompileTime==Dynamic)
76     std::cout << "dyn   ";
77   else
78     std::cout << "fixed ";
79   std::cout << covMat.rows() << " \t"
80             << (timerNoSqrt.best()) / repeats << "s "
81             << "(" << 1e-9 * cost*repeats/timerNoSqrt.best() << " GFLOPS)\t"
82             << (timerSqrt.best()) / repeats << "s "
83             << "(" << 1e-9 * cost*repeats/timerSqrt.best() << " GFLOPS)\n";
84 
85 
86   #ifdef BENCH_GSL
87   if (MatrixType::RowsAtCompileTime==Dynamic)
88   {
89     timerSqrt.reset();
90 
91     gsl_matrix* gslCovMat = gsl_matrix_alloc(covMat.rows(),covMat.cols());
92     gsl_matrix* gslCopy = gsl_matrix_alloc(covMat.rows(),covMat.cols());
93 
94     eiToGsl(covMat, &gslCovMat);
95     for (int t=0; t<TRIES; ++t)
96     {
97       timerSqrt.start();
98       for (int k=0; k<repeats; ++k)
99       {
100         gsl_matrix_memcpy(gslCopy,gslCovMat);
101         gsl_linalg_cholesky_decomp(gslCopy);
102         acc += gsl_matrix_get(gslCopy,r,c);
103       }
104       timerSqrt.stop();
105     }
106 
107     std::cout << " | \t"
108               << timerSqrt.value() * REPEAT / repeats << "s";
109 
110     gsl_matrix_free(gslCovMat);
111   }
112   #endif
113   std::cout << "\n";
114   // make sure the compiler does not optimize too much
115   if (acc==123)
116     std::cout << acc;
117 }
118 
main(int argc,char * argv[])119 int main(int argc, char* argv[])
120 {
121   const int dynsizes[] = {4,6,8,16,24,32,49,64,128,256,512,900,1500,0};
122   std::cout << "size            LDLT                            LLT";
123 //   #ifdef BENCH_GSL
124 //   std::cout << "       GSL (standard + double + ATLAS)  ";
125 //   #endif
126   std::cout << "\n";
127   for (int i=0; dynsizes[i]>0; ++i)
128     benchLLT(Matrix<Scalar,Dynamic,Dynamic>(dynsizes[i],dynsizes[i]));
129 
130   benchLLT(Matrix<Scalar,2,2>());
131   benchLLT(Matrix<Scalar,3,3>());
132   benchLLT(Matrix<Scalar,4,4>());
133   benchLLT(Matrix<Scalar,5,5>());
134   benchLLT(Matrix<Scalar,6,6>());
135   benchLLT(Matrix<Scalar,7,7>());
136   benchLLT(Matrix<Scalar,8,8>());
137   benchLLT(Matrix<Scalar,12,12>());
138   benchLLT(Matrix<Scalar,16,16>());
139   return 0;
140 }
141 
142