devtools/eigen
1
0
Fork 0
mirror of https://gitlab.com/libeigen/eigen.git synced 2026-04-10 11:34:33 +08:00
Code Issues Packages Projects Releases Wiki Activity

implement a smarter parallelization strategy for gemm avoiding multiple

Browse Source 
paking of the same data
This commit is contained in:
Gael Guennebaud
2010-02-26 12:32:00 +01:00
parent a1e1103328
commit 3ac2b96a2f
5 changed files with 246 additions and 138 deletions
Download Patch File Download Diff File Expand all files Collapse all files

79
bench/bench_gemm.cpp
View File

@@ -15,6 +15,52 @@ using namespace Eigen;
typedef SCALAR Scalar;
typedef Matrix<Scalar,Dynamic,Dynamic> M;
#ifdef HAVE_BLAS
extern "C" {
#include <bench/btl/libs/C_BLAS/blas.h>
void sgemm_kernel(int actual_mc, int cols, int actual_kc, float alpha,
float* blockA, float* blockB, float* res, int resStride);
void sgemm_oncopy(int actual_kc, int cols, const float* rhs, int rhsStride, float* blockB);
void sgemm_itcopy(int actual_kc, int cols, const float* rhs, int rhsStride, float* blockB);
}
static float fone = 1;
static float fzero = 0;
static double done = 1;
static double szero = 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_(&notrans,&notrans,&M,&N,&K,&fone,
const_cast<float*>(a.data()),&lda,
const_cast<float*>(b.data()),&ldb,&fone,
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_(&notrans,&notrans,&M,&N,&K,&done,
const_cast<double*>(a.data()),&lda,
const_cast<double*>(b.data()),&ldb,&done,
c.data(),&ldc);
}
#endif
void gemm(const M& a, const M& b, M& c)
{
c.noalias() += a * b;
@@ -22,21 +68,42 @@ void gemm(const M& a, const M& b, M& c)
int main(int argc, char ** argv)
{
int rep = 1;
int rep = 1; // number of repetitions per try
int tries = 5; // number of tries, we keep the best
int s = 2048;
int m = s;
int n = s;
int p = s;
M a(m,n); a.setOnes();
M b(n,p); b.setOnes();
M a(m,n); a.setRandom();
M b(n,p); b.setRandom();
M c(m,p); c.setOnes();
BenchTimer t;
BENCH(t, 5, rep, gemm(a,b,c));
M r = c;
std::cerr << "cpu " << t.best(CPU_TIMER)/rep << "s \t" << (double(m)*n*p*rep*2/t.best(CPU_TIMER))*1e-9 << " GFLOPS \t(" << t.total(CPU_TIMER) << "s)\n";
std::cerr << "real " << t.best(REAL_TIMER)/rep << "s \t" << (double(m)*n*p*rep*2/t.best(REAL_TIMER))*1e-9 << " GFLOPS \t(" << t.total(REAL_TIMER) << "s)\n";
// check the parallel product is correct
#ifdef HAVE_BLAS
blas_gemm(a,b,r);
#else
int procs = omp_get_max_threads();
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";
#ifdef HAVE_BLAS
BENCH(t, tries, rep, blas_gemm(a,b,c));
std::cerr << "blas cpu " << t.best(CPU_TIMER)/rep << "s \t" << (double(m)*n*p*rep*2/t.best(CPU_TIMER))*1e-9 << " GFLOPS \t(" << t.total(CPU_TIMER) << "s)\n";
std::cerr << "blas real " << t.best(REAL_TIMER)/rep << "s \t" << (double(m)*n*p*rep*2/t.best(REAL_TIMER))*1e-9 << " GFLOPS \t(" << t.total(REAL_TIMER) << "s)\n";
#endif
BENCH(t, tries, rep, gemm(a,b,c));
std::cerr << "eigen cpu " << t.best(CPU_TIMER)/rep << "s \t" << (double(m)*n*p*rep*2/t.best(CPU_TIMER))*1e-9 << " GFLOPS \t(" << t.total(CPU_TIMER) << "s)\n";
std::cerr << "eigen real " << t.best(REAL_TIMER)/rep << "s \t" << (double(m)*n*p*rep*2/t.best(REAL_TIMER))*1e-9 << " GFLOPS \t(" << t.total(REAL_TIMER) << "s)\n";
return 0;
}