mirror of
https://gitlab.com/libeigen/eigen.git
synced 2024-12-21 07:19:46 +08:00
194 lines
6.1 KiB
C++
194 lines
6.1 KiB
C++
|
|
// g++-4.4 bench_gemm.cpp -I .. -O2 -DNDEBUG -lrt -fopenmp && OMP_NUM_THREADS=2 ./a.out
|
|
// icpc bench_gemm.cpp -I .. -O3 -DNDEBUG -lrt -openmp && OMP_NUM_THREADS=2 ./a.out
|
|
|
|
#include <iostream>
|
|
#include <Eigen/Core>
|
|
#include <bench/BenchTimer.h>
|
|
|
|
using namespace std;
|
|
using namespace Eigen;
|
|
|
|
#ifndef SCALAR
|
|
#define SCALAR std::complex<double>
|
|
// #define SCALAR double
|
|
#endif
|
|
|
|
typedef SCALAR Scalar;
|
|
typedef NumTraits<Scalar>::Real RealScalar;
|
|
typedef Matrix<RealScalar,Dynamic,Dynamic> A;
|
|
typedef Matrix<Scalar,Dynamic,Dynamic> B;
|
|
typedef Matrix<Scalar,Dynamic,Dynamic> C;
|
|
|
|
#ifdef HAVE_BLAS
|
|
|
|
extern "C" {
|
|
#include <bench/btl/libs/C_BLAS/blas.h>
|
|
}
|
|
|
|
static float fone = 1;
|
|
static float fzero = 0;
|
|
static double done = 1;
|
|
static double szero = 0;
|
|
static std::complex<float> cfone = 1;
|
|
static std::complex<float> cfzero = 0;
|
|
static std::complex<double> cdone = 1;
|
|
static std::complex<double> cdzero = 0;
|
|
static char notrans = 'N';
|
|
static char trans = 'T';
|
|
static char nonunit = 'N';
|
|
static char lower = 'L';
|
|
static char right = 'R';
|
|
static int intone = 1;
|
|
|
|
void blas_gemm(const MatrixXf& a, const MatrixXf& b, MatrixXf& c)
|
|
{
|
|
int M = c.rows(); int N = c.cols(); int K = a.cols();
|
|
int lda = a.rows(); int ldb = b.rows(); int ldc = c.rows();
|
|
|
|
sgemm_(¬rans,¬rans,&M,&N,&K,&fone,
|
|
const_cast<float*>(a.data()),&lda,
|
|
const_cast<float*>(b.data()),&ldb,&fone,
|
|
c.data(),&ldc);
|
|
}
|
|
|
|
void blas_gemm(const MatrixXcf& a, const MatrixXcf& b, MatrixXcf& c)
|
|
{
|
|
int M = c.rows(); int N = c.cols(); int K = a.cols();
|
|
int lda = a.rows(); int ldb = b.rows(); int ldc = c.rows();
|
|
|
|
cgemm_(¬rans,¬rans,&M,&N,&K,(float*)&cfone,
|
|
const_cast<float*>((const float*)a.data()),&lda,
|
|
const_cast<float*>((const float*)b.data()),&ldb,(float*)&cfone,
|
|
(float*)c.data(),&ldc);
|
|
}
|
|
|
|
void blas_gemm(const MatrixXcd& a, const MatrixXcd& b, MatrixXcd& c)
|
|
{
|
|
int M = c.rows(); int N = c.cols(); int K = a.cols();
|
|
int lda = a.rows(); int ldb = b.rows(); int ldc = c.rows();
|
|
|
|
zgemm_(¬rans,¬rans,&M,&N,&K,(double*)&cdone,
|
|
const_cast<double*>((const double*)a.data()),&lda,
|
|
const_cast<double*>((const double*)b.data()),&ldb,(double*)&cdone,
|
|
(double*)c.data(),&ldc);
|
|
}
|
|
|
|
void blas_gemm(const MatrixXd& a, const MatrixXd& b, MatrixXd& c)
|
|
{
|
|
int M = c.rows(); int N = c.cols(); int K = a.cols();
|
|
int lda = a.rows(); int ldb = b.rows(); int ldc = c.rows();
|
|
|
|
dgemm_(¬rans,¬rans,&M,&N,&K,&done,
|
|
const_cast<double*>(a.data()),&lda,
|
|
const_cast<double*>(b.data()),&ldb,&done,
|
|
c.data(),&ldc);
|
|
}
|
|
|
|
#endif
|
|
|
|
template<typename A, typename B, typename C>
|
|
EIGEN_DONT_INLINE void gemm(const A& a, const B& b, C& c)
|
|
{
|
|
c.noalias() += a * b;
|
|
}
|
|
|
|
int main(int argc, char ** argv)
|
|
{
|
|
std::ptrdiff_t l1 = ei_queryL1CacheSize();
|
|
std::ptrdiff_t l2 = ei_queryTopLevelCacheSize();
|
|
std::cout << "L1 cache size = " << (l1>0 ? l1/1024 : -1) << " KB\n";
|
|
std::cout << "L2/L3 cache size = " << (l2>0 ? l2/1024 : -1) << " KB\n";
|
|
typedef ei_product_blocking_traits<Scalar,Scalar> Blocking;
|
|
std::cout << "Register blocking = " << Blocking::mr << " x " << Blocking::nr << "\n";
|
|
|
|
int rep = 1; // number of repetitions per try
|
|
int tries = 2; // number of tries, we keep the best
|
|
|
|
int s = 2048;
|
|
int cache_size = -1;
|
|
|
|
bool need_help = false;
|
|
for (int i=1; i<argc; ++i)
|
|
{
|
|
if(argv[i][0]=='s')
|
|
s = atoi(argv[i]+1);
|
|
else if(argv[i][0]=='c')
|
|
cache_size = atoi(argv[i]+1);
|
|
else if(argv[i][0]=='t')
|
|
tries = atoi(argv[i]+1);
|
|
else if(argv[i][0]=='p')
|
|
rep = atoi(argv[i]+1);
|
|
else
|
|
need_help = true;
|
|
}
|
|
|
|
if(need_help)
|
|
{
|
|
std::cout << argv[0] << " s<matrix size> c<cache size> t<nb tries> p<nb repeats>\n";
|
|
return 1;
|
|
}
|
|
|
|
if(cache_size>0)
|
|
setCpuCacheSizes(cache_size,96*cache_size);
|
|
|
|
int m = s;
|
|
int n = s;
|
|
int p = s;
|
|
A a(m,n); a.setRandom();
|
|
B b(n,p); b.setRandom();
|
|
C c(m,p); c.setOnes();
|
|
|
|
std::cout << "Matrix sizes = " << m << "x" << p << " * " << p << "x" << n << "\n";
|
|
std::ptrdiff_t cm(m), cn(n), ck(p);
|
|
computeProductBlockingSizes<Scalar,Scalar>(ck, cm, cn);
|
|
std::cout << "blocking size = " << cm << " x " << ck << "\n";
|
|
|
|
C r = c;
|
|
|
|
// check the parallel product is correct
|
|
#ifdef EIGEN_HAS_OPENMP
|
|
int procs = omp_get_max_threads();
|
|
if(procs>1)
|
|
{
|
|
#ifdef HAVE_BLAS
|
|
blas_gemm(a,b,r);
|
|
#else
|
|
omp_set_num_threads(1);
|
|
r.noalias() += a * b;
|
|
omp_set_num_threads(procs);
|
|
#endif
|
|
c.noalias() += a * b;
|
|
if(!r.isApprox(c)) std::cerr << "Warning, your parallel product is crap!\n\n";
|
|
}
|
|
#endif
|
|
|
|
#ifdef HAVE_BLAS
|
|
BenchTimer tblas;
|
|
BENCH(tblas, tries, rep, blas_gemm(a,b,c));
|
|
std::cout << "blas cpu " << tblas.best(CPU_TIMER)/rep << "s \t" << (double(m)*n*p*rep*2/tblas.best(CPU_TIMER))*1e-9 << " GFLOPS \t(" << tblas.total(CPU_TIMER) << "s)\n";
|
|
std::cout << "blas real " << tblas.best(REAL_TIMER)/rep << "s \t" << (double(m)*n*p*rep*2/tblas.best(REAL_TIMER))*1e-9 << " GFLOPS \t(" << tblas.total(REAL_TIMER) << "s)\n";
|
|
#endif
|
|
|
|
BenchTimer tmt;
|
|
BENCH(tmt, tries, rep, gemm(a,b,c));
|
|
std::cout << "eigen cpu " << tmt.best(CPU_TIMER)/rep << "s \t" << (double(m)*n*p*rep*2/tmt.best(CPU_TIMER))*1e-9 << " GFLOPS \t(" << tmt.total(CPU_TIMER) << "s)\n";
|
|
std::cout << "eigen real " << tmt.best(REAL_TIMER)/rep << "s \t" << (double(m)*n*p*rep*2/tmt.best(REAL_TIMER))*1e-9 << " GFLOPS \t(" << tmt.total(REAL_TIMER) << "s)\n";
|
|
|
|
#ifdef EIGEN_HAS_OPENMP
|
|
if(procs>1)
|
|
{
|
|
BenchTimer tmono;
|
|
//omp_set_num_threads(1);
|
|
Eigen::setNbThreads(1);
|
|
BENCH(tmono, tries, rep, gemm(a,b,c));
|
|
std::cout << "eigen mono cpu " << tmono.best(CPU_TIMER)/rep << "s \t" << (double(m)*n*p*rep*2/tmono.best(CPU_TIMER))*1e-9 << " GFLOPS \t(" << tmono.total(CPU_TIMER) << "s)\n";
|
|
std::cout << "eigen mono real " << tmono.best(REAL_TIMER)/rep << "s \t" << (double(m)*n*p*rep*2/tmono.best(REAL_TIMER))*1e-9 << " GFLOPS \t(" << tmono.total(REAL_TIMER) << "s)\n";
|
|
std::cout << "mt speed up x" << tmono.best(CPU_TIMER) / tmt.best(REAL_TIMER) << " => " << (100.0*tmono.best(CPU_TIMER) / tmt.best(REAL_TIMER))/procs << "%\n";
|
|
}
|
|
#endif
|
|
|
|
return 0;
|
|
}
|
|
|