libicr: bench_gemm.cpp Source File

Go to the documentation of this file.
00001 
00002 // g++-4.4 bench_gemm.cpp -I .. -O2 -DNDEBUG -lrt -fopenmp && OMP_NUM_THREADS=2  ./a.out
00003 // icpc bench_gemm.cpp -I .. -O3 -DNDEBUG -lrt -openmp  && OMP_NUM_THREADS=2  ./a.out
00004 
00005 #include <iostream>
00006 #include <Eigen/Core>
00007 #include <bench/BenchTimer.h>
00008 
00009 using namespace std;
00010 using namespace Eigen;
00011 
00012 #ifndef SCALAR
00013 // #define SCALAR std::complex<float>
00014 #define SCALAR float
00015 #endif
00016 
00017 typedef SCALAR Scalar;
00018 typedef NumTraits<Scalar>::Real RealScalar;
00019 typedef Matrix<RealScalar,Dynamic,Dynamic> A;
00020 typedef Matrix</*Real*/Scalar,Dynamic,Dynamic> B;
00021 typedef Matrix<Scalar,Dynamic,Dynamic> C;
00022 typedef Matrix<RealScalar,Dynamic,Dynamic> M;
00023 
00024 #ifdef HAVE_BLAS
00025 
00026 extern "C" {
00027   #include <bench/btl/libs/C_BLAS/blas.h>
00028 }
00029 
00030 static float fone = 1;
00031 static float fzero = 0;
00032 static double done = 1;
00033 static double szero = 0;
00034 static std::complex<float> cfone = 1;
00035 static std::complex<float> cfzero = 0;
00036 static std::complex<double> cdone = 1;
00037 static std::complex<double> cdzero = 0;
00038 static char notrans = 'N';
00039 static char trans = 'T';  
00040 static char nonunit = 'N';
00041 static char lower = 'L';
00042 static char right = 'R';
00043 static int intone = 1;
00044 
00045 void blas_gemm(const MatrixXf& a, const MatrixXf& b, MatrixXf& c)
00046 {
00047   int M = c.rows(); int N = c.cols(); int K = a.cols();
00048   int lda = a.rows(); int ldb = b.rows(); int ldc = c.rows();
00049 
00050   sgemm_(&notrans,&notrans,&M,&N,&K,&fone,
00051          const_cast<float*>(a.data()),&lda,
00052          const_cast<float*>(b.data()),&ldb,&fone,
00053          c.data(),&ldc);
00054 }
00055 
00056 EIGEN_DONT_INLINE void blas_gemm(const MatrixXd& a, const MatrixXd& b, MatrixXd& c)
00057 {
00058   int M = c.rows(); int N = c.cols(); int K = a.cols();
00059   int lda = a.rows(); int ldb = b.rows(); int ldc = c.rows();
00060 
00061   dgemm_(&notrans,&notrans,&M,&N,&K,&done,
00062          const_cast<double*>(a.data()),&lda,
00063          const_cast<double*>(b.data()),&ldb,&done,
00064          c.data(),&ldc);
00065 }
00066 
00067 void blas_gemm(const MatrixXcf& a, const MatrixXcf& b, MatrixXcf& c)
00068 {
00069   int M = c.rows(); int N = c.cols(); int K = a.cols();
00070   int lda = a.rows(); int ldb = b.rows(); int ldc = c.rows();
00071 
00072   cgemm_(&notrans,&notrans,&M,&N,&K,(float*)&cfone,
00073          const_cast<float*>((const float*)a.data()),&lda,
00074          const_cast<float*>((const float*)b.data()),&ldb,(float*)&cfone,
00075          (float*)c.data(),&ldc);
00076 }
00077 
00078 void blas_gemm(const MatrixXcd& a, const MatrixXcd& b, MatrixXcd& c)
00079 {
00080   int M = c.rows(); int N = c.cols(); int K = a.cols();
00081   int lda = a.rows(); int ldb = b.rows(); int ldc = c.rows();
00082 
00083   zgemm_(&notrans,&notrans,&M,&N,&K,(double*)&cdone,
00084          const_cast<double*>((const double*)a.data()),&lda,
00085          const_cast<double*>((const double*)b.data()),&ldb,(double*)&cdone,
00086          (double*)c.data(),&ldc);
00087 }
00088 
00089 
00090 
00091 #endif
00092 
00093 void matlab_cplx_cplx(const M& ar, const M& ai, const M& br, const M& bi, M& cr, M& ci)
00094 {
00095   cr.noalias() += ar * br;
00096   cr.noalias() -= ai * bi;
00097   ci.noalias() += ar * bi;
00098   ci.noalias() += ai * br;
00099 }
00100 
00101 void matlab_real_cplx(const M& a, const M& br, const M& bi, M& cr, M& ci)
00102 {
00103   cr.noalias() += a * br;
00104   ci.noalias() += a * bi;
00105 }
00106 
00107 void matlab_cplx_real(const M& ar, const M& ai, const M& b, M& cr, M& ci)
00108 {
00109   cr.noalias() += ar * b;
00110   ci.noalias() += ai * b;
00111 }
00112 
00113 template<typename A, typename B, typename C>
00114 EIGEN_DONT_INLINE void gemm(const A& a, const B& b, C& c)
00115 {
00116  c.noalias() += a * b;
00117 }
00118 
00119 int main(int argc, char ** argv)
00120 {
00121   std::ptrdiff_t l1 = internal::queryL1CacheSize();
00122   std::ptrdiff_t l2 = internal::queryTopLevelCacheSize();
00123   std::cout << "L1 cache size     = " << (l1>0 ? l1/1024 : -1) << " KB\n";
00124   std::cout << "L2/L3 cache size  = " << (l2>0 ? l2/1024 : -1) << " KB\n";
00125   typedef internal::gebp_traits<Scalar,Scalar> Traits;
00126   std::cout << "Register blocking = " << Traits::mr << " x " << Traits::nr << "\n";
00127 
00128   int rep = 1;    // number of repetitions per try
00129   int tries = 2;  // number of tries, we keep the best
00130 
00131   int s = 2048;
00132   int cache_size = -1;
00133 
00134   bool need_help = false;
00135   for (int i=1; i<argc; ++i)
00136   {
00137     if(argv[i][0]=='s')
00138       s = atoi(argv[i]+1);
00139     else if(argv[i][0]=='c')
00140       cache_size = atoi(argv[i]+1);
00141     else if(argv[i][0]=='t')
00142       tries = atoi(argv[i]+1);
00143     else if(argv[i][0]=='p')
00144       rep = atoi(argv[i]+1);
00145     else
00146       need_help = true;
00147   }
00148 
00149   if(need_help)
00150   {
00151     std::cout << argv[0] << " s<matrix size> c<cache size> t<nb tries> p<nb repeats>\n";
00152     return 1;
00153   }
00154 
00155   if(cache_size>0)
00156     setCpuCacheSizes(cache_size,96*cache_size);
00157 
00158   int m = s;
00159   int n = s;
00160   int p = s;
00161   A a(m,p); a.setRandom();
00162   B b(p,n); b.setRandom();
00163   C c(m,n); c.setOnes();
00164 
00165   std::cout << "Matrix sizes = " << m << "x" << p << " * " << p << "x" << n << "\n";
00166   std::ptrdiff_t mc(m), nc(n), kc(p);
00167   computeProductBlockingSizes<Scalar,Scalar>(kc, mc, nc);
00168   std::cout << "blocking size (mc x kc) = " << mc << " x " << kc << "\n";
00169 
00170   C r = c;
00171 
00172   // check the parallel product is correct
00173   #if defined EIGEN_HAS_OPENMP
00174   int procs = omp_get_max_threads();
00175   if(procs>1)
00176   {
00177     #ifdef HAVE_BLAS
00178     blas_gemm(a,b,r);
00179     #else
00180     omp_set_num_threads(1);
00181     r.noalias() += a * b;
00182     omp_set_num_threads(procs);
00183     #endif
00184     c.noalias() += a * b;
00185     if(!r.isApprox(c)) std::cerr << "Warning, your parallel product is crap!\n\n";
00186   }
00187   #elif defined HAVE_BLAS
00188     blas_gemm(a,b,r);
00189     c.noalias() += a * b;
00190     if(!r.isApprox(c)) std::cerr << "Warning, your product is crap!\n\n";
00191 //     std::cerr << r << "\n\n" << c << "\n\n";
00192   #else
00193     gemm(a,b,c);
00194     r.noalias() += a.cast<Scalar>() * b.cast<Scalar>();
00195     if(!r.isApprox(c)) std::cerr << "Warning, your product is crap!\n\n";
00196 //     std::cerr << c << "\n\n";
00197 //     std::cerr << r << "\n\n";
00198   #endif
00199 
00200   #ifdef HAVE_BLAS
00201   BenchTimer tblas;
00202   BENCH(tblas, tries, rep, blas_gemm(a,b,c));
00203   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";
00204   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";
00205   #endif
00206 
00207   BenchTimer tmt;
00208   BENCH(tmt, tries, rep, gemm(a,b,c));
00209   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";
00210   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";
00211 
00212   #ifdef EIGEN_HAS_OPENMP
00213   if(procs>1)
00214   {
00215     BenchTimer tmono;
00216     //omp_set_num_threads(1);
00217     Eigen::setNbThreads(1);
00218     BENCH(tmono, tries, rep, gemm(a,b,c));
00219     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";
00220     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";
00221     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";
00222   }
00223   #endif
00224   
00225   #ifdef DECOUPLED
00226   if((NumTraits<A::Scalar>::IsComplex) && (NumTraits<B::Scalar>::IsComplex))
00227   {
00228     M ar(m,p); ar.setRandom();
00229     M ai(m,p); ai.setRandom();
00230     M br(p,n); br.setRandom();
00231     M bi(p,n); bi.setRandom();
00232     M cr(m,n); cr.setRandom();
00233     M ci(m,n); ci.setRandom();
00234     
00235     BenchTimer t;
00236     BENCH(t, tries, rep, matlab_cplx_cplx(ar,ai,br,bi,cr,ci));
00237     std::cout << "\"matlab\" 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";
00238     std::cout << "\"matlab\" 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";
00239   }
00240   if((!NumTraits<A::Scalar>::IsComplex) && (NumTraits<B::Scalar>::IsComplex))
00241   {
00242     M a(m,p);  a.setRandom();
00243     M br(p,n); br.setRandom();
00244     M bi(p,n); bi.setRandom();
00245     M cr(m,n); cr.setRandom();
00246     M ci(m,n); ci.setRandom();
00247     
00248     BenchTimer t;
00249     BENCH(t, tries, rep, matlab_real_cplx(a,br,bi,cr,ci));
00250     std::cout << "\"matlab\" 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";
00251     std::cout << "\"matlab\" 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";
00252   }
00253   if((NumTraits<A::Scalar>::IsComplex) && (!NumTraits<B::Scalar>::IsComplex))
00254   {
00255     M ar(m,p); ar.setRandom();
00256     M ai(m,p); ai.setRandom();
00257     M b(p,n);  b.setRandom();
00258     M cr(m,n); cr.setRandom();
00259     M ci(m,n); ci.setRandom();
00260     
00261     BenchTimer t;
00262     BENCH(t, tries, rep, matlab_cplx_real(ar,ai,b,cr,ci));
00263     std::cout << "\"matlab\" 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";
00264     std::cout << "\"matlab\" 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";
00265   }
00266   #endif
00267 
00268   return 0;
00269 }
00270