• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 
2 // g++-4.4 bench_gemm.cpp -I .. -O2 -DNDEBUG -lrt -fopenmp && OMP_NUM_THREADS=2  ./a.out
3 // icpc bench_gemm.cpp -I .. -O3 -DNDEBUG -lrt -openmp  && OMP_NUM_THREADS=2  ./a.out
4 
5 #include <iostream>
6 #include <Eigen/Core>
7 #include <bench/BenchTimer.h>
8 
9 using namespace std;
10 using namespace Eigen;
11 
12 #ifndef SCALAR
13 // #define SCALAR std::complex<float>
14 #define SCALAR float
15 #endif
16 
17 typedef SCALAR Scalar;
18 typedef NumTraits<Scalar>::Real RealScalar;
19 typedef Matrix<RealScalar,Dynamic,Dynamic> A;
20 typedef Matrix</*Real*/Scalar,Dynamic,Dynamic> B;
21 typedef Matrix<Scalar,Dynamic,Dynamic> C;
22 typedef Matrix<RealScalar,Dynamic,Dynamic> M;
23 
24 #ifdef HAVE_BLAS
25 
26 extern "C" {
27   #include <Eigen/src/misc/blas.h>
28 }
29 
30 static float fone = 1;
31 static float fzero = 0;
32 static double done = 1;
33 static double szero = 0;
34 static std::complex<float> cfone = 1;
35 static std::complex<float> cfzero = 0;
36 static std::complex<double> cdone = 1;
37 static std::complex<double> cdzero = 0;
38 static char notrans = 'N';
39 static char trans = 'T';
40 static char nonunit = 'N';
41 static char lower = 'L';
42 static char right = 'R';
43 static int intone = 1;
44 
blas_gemm(const MatrixXf & a,const MatrixXf & b,MatrixXf & c)45 void blas_gemm(const MatrixXf& a, const MatrixXf& b, MatrixXf& c)
46 {
47   int M = c.rows(); int N = c.cols(); int K = a.cols();
48   int lda = a.rows(); int ldb = b.rows(); int ldc = c.rows();
49 
50   sgemm_(&notrans,&notrans,&M,&N,&K,&fone,
51          const_cast<float*>(a.data()),&lda,
52          const_cast<float*>(b.data()),&ldb,&fone,
53          c.data(),&ldc);
54 }
55 
blas_gemm(const MatrixXd & a,const MatrixXd & b,MatrixXd & c)56 EIGEN_DONT_INLINE void blas_gemm(const MatrixXd& a, const MatrixXd& b, MatrixXd& c)
57 {
58   int M = c.rows(); int N = c.cols(); int K = a.cols();
59   int lda = a.rows(); int ldb = b.rows(); int ldc = c.rows();
60 
61   dgemm_(&notrans,&notrans,&M,&N,&K,&done,
62          const_cast<double*>(a.data()),&lda,
63          const_cast<double*>(b.data()),&ldb,&done,
64          c.data(),&ldc);
65 }
66 
blas_gemm(const MatrixXcf & a,const MatrixXcf & b,MatrixXcf & c)67 void blas_gemm(const MatrixXcf& a, const MatrixXcf& b, MatrixXcf& c)
68 {
69   int M = c.rows(); int N = c.cols(); int K = a.cols();
70   int lda = a.rows(); int ldb = b.rows(); int ldc = c.rows();
71 
72   cgemm_(&notrans,&notrans,&M,&N,&K,(float*)&cfone,
73          const_cast<float*>((const float*)a.data()),&lda,
74          const_cast<float*>((const float*)b.data()),&ldb,(float*)&cfone,
75          (float*)c.data(),&ldc);
76 }
77 
blas_gemm(const MatrixXcd & a,const MatrixXcd & b,MatrixXcd & c)78 void blas_gemm(const MatrixXcd& a, const MatrixXcd& b, MatrixXcd& c)
79 {
80   int M = c.rows(); int N = c.cols(); int K = a.cols();
81   int lda = a.rows(); int ldb = b.rows(); int ldc = c.rows();
82 
83   zgemm_(&notrans,&notrans,&M,&N,&K,(double*)&cdone,
84          const_cast<double*>((const double*)a.data()),&lda,
85          const_cast<double*>((const double*)b.data()),&ldb,(double*)&cdone,
86          (double*)c.data(),&ldc);
87 }
88 
89 
90 
91 #endif
92 
matlab_cplx_cplx(const M & ar,const M & ai,const M & br,const M & bi,M & cr,M & ci)93 void matlab_cplx_cplx(const M& ar, const M& ai, const M& br, const M& bi, M& cr, M& ci)
94 {
95   cr.noalias() += ar * br;
96   cr.noalias() -= ai * bi;
97   ci.noalias() += ar * bi;
98   ci.noalias() += ai * br;
99 }
100 
matlab_real_cplx(const M & a,const M & br,const M & bi,M & cr,M & ci)101 void matlab_real_cplx(const M& a, const M& br, const M& bi, M& cr, M& ci)
102 {
103   cr.noalias() += a * br;
104   ci.noalias() += a * bi;
105 }
106 
matlab_cplx_real(const M & ar,const M & ai,const M & b,M & cr,M & ci)107 void matlab_cplx_real(const M& ar, const M& ai, const M& b, M& cr, M& ci)
108 {
109   cr.noalias() += ar * b;
110   ci.noalias() += ai * b;
111 }
112 
113 template<typename A, typename B, typename C>
gemm(const A & a,const B & b,C & c)114 EIGEN_DONT_INLINE void gemm(const A& a, const B& b, C& c)
115 {
116  c.noalias() += a * b;
117 }
118 
main(int argc,char ** argv)119 int main(int argc, char ** argv)
120 {
121   std::ptrdiff_t l1 = internal::queryL1CacheSize();
122   std::ptrdiff_t l2 = internal::queryTopLevelCacheSize();
123   std::cout << "L1 cache size     = " << (l1>0 ? l1/1024 : -1) << " KB\n";
124   std::cout << "L2/L3 cache size  = " << (l2>0 ? l2/1024 : -1) << " KB\n";
125   typedef internal::gebp_traits<Scalar,Scalar> Traits;
126   std::cout << "Register blocking = " << Traits::mr << " x " << Traits::nr << "\n";
127 
128   int rep = 1;    // number of repetitions per try
129   int tries = 2;  // number of tries, we keep the best
130 
131   int s = 2048;
132   int cache_size = -1;
133 
134   bool need_help = false;
135   for (int i=1; i<argc; ++i)
136   {
137     if(argv[i][0]=='s')
138       s = atoi(argv[i]+1);
139     else if(argv[i][0]=='c')
140       cache_size = atoi(argv[i]+1);
141     else if(argv[i][0]=='t')
142       tries = atoi(argv[i]+1);
143     else if(argv[i][0]=='p')
144       rep = atoi(argv[i]+1);
145     else
146       need_help = true;
147   }
148 
149   if(need_help)
150   {
151     std::cout << argv[0] << " s<matrix size> c<cache size> t<nb tries> p<nb repeats>\n";
152     return 1;
153   }
154 
155   if(cache_size>0)
156     setCpuCacheSizes(cache_size,96*cache_size);
157 
158   int m = s;
159   int n = s;
160   int p = s;
161   A a(m,p); a.setRandom();
162   B b(p,n); b.setRandom();
163   C c(m,n); c.setOnes();
164   C rc = c;
165 
166   std::cout << "Matrix sizes = " << m << "x" << p << " * " << p << "x" << n << "\n";
167   std::ptrdiff_t mc(m), nc(n), kc(p);
168   internal::computeProductBlockingSizes<Scalar,Scalar>(kc, mc, nc);
169   std::cout << "blocking size (mc x kc) = " << mc << " x " << kc << "\n";
170 
171   C r = c;
172 
173   // check the parallel product is correct
174   #if defined EIGEN_HAS_OPENMP
175   int procs = omp_get_max_threads();
176   if(procs>1)
177   {
178     #ifdef HAVE_BLAS
179     blas_gemm(a,b,r);
180     #else
181     omp_set_num_threads(1);
182     r.noalias() += a * b;
183     omp_set_num_threads(procs);
184     #endif
185     c.noalias() += a * b;
186     if(!r.isApprox(c)) std::cerr << "Warning, your parallel product is crap!\n\n";
187   }
188   #elif defined HAVE_BLAS
189     blas_gemm(a,b,r);
190     c.noalias() += a * b;
191     if(!r.isApprox(c)) std::cerr << "Warning, your product is crap!\n\n";
192   #else
193     gemm(a,b,c);
194     r.noalias() += a.cast<Scalar>() * b.cast<Scalar>();
195     if(!r.isApprox(c)) std::cerr << "Warning, your product is crap!\n\n";
196   #endif
197 
198   #ifdef HAVE_BLAS
199   BenchTimer tblas;
200   c = rc;
201   BENCH(tblas, tries, rep, blas_gemm(a,b,c));
202   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";
203   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";
204   #endif
205 
206   BenchTimer tmt;
207   c = rc;
208   BENCH(tmt, tries, rep, gemm(a,b,c));
209   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";
210   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";
211 
212   #ifdef EIGEN_HAS_OPENMP
213   if(procs>1)
214   {
215     BenchTimer tmono;
216     omp_set_num_threads(1);
217     Eigen::internal::setNbThreads(1);
218     c = rc;
219     BENCH(tmono, tries, rep, gemm(a,b,c));
220     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";
221     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";
222     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";
223   }
224   #endif
225 
226   #ifdef DECOUPLED
227   if((NumTraits<A::Scalar>::IsComplex) && (NumTraits<B::Scalar>::IsComplex))
228   {
229     M ar(m,p); ar.setRandom();
230     M ai(m,p); ai.setRandom();
231     M br(p,n); br.setRandom();
232     M bi(p,n); bi.setRandom();
233     M cr(m,n); cr.setRandom();
234     M ci(m,n); ci.setRandom();
235 
236     BenchTimer t;
237     BENCH(t, tries, rep, matlab_cplx_cplx(ar,ai,br,bi,cr,ci));
238     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";
239     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";
240   }
241   if((!NumTraits<A::Scalar>::IsComplex) && (NumTraits<B::Scalar>::IsComplex))
242   {
243     M a(m,p);  a.setRandom();
244     M br(p,n); br.setRandom();
245     M bi(p,n); bi.setRandom();
246     M cr(m,n); cr.setRandom();
247     M ci(m,n); ci.setRandom();
248 
249     BenchTimer t;
250     BENCH(t, tries, rep, matlab_real_cplx(a,br,bi,cr,ci));
251     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";
252     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";
253   }
254   if((NumTraits<A::Scalar>::IsComplex) && (!NumTraits<B::Scalar>::IsComplex))
255   {
256     M ar(m,p); ar.setRandom();
257     M ai(m,p); ai.setRandom();
258     M b(p,n);  b.setRandom();
259     M cr(m,n); cr.setRandom();
260     M ci(m,n); ci.setRandom();
261 
262     BenchTimer t;
263     BENCH(t, tries, rep, matlab_cplx_real(ar,ai,b,cr,ci));
264     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";
265     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";
266   }
267   #endif
268 
269   return 0;
270 }
271 
272