• 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 // Compilation options:
6 //
7 // -DSCALAR=std::complex<double>
8 // -DSCALARA=double or -DSCALARB=double
9 // -DHAVE_BLAS
10 // -DDECOUPLED
11 //
12 
13 #include <iostream>
14 #include <Eigen/Core>
15 #include <bench/BenchTimer.h>
16 
17 using namespace std;
18 using namespace Eigen;
19 
20 #ifndef SCALAR
21 // #define SCALAR std::complex<float>
22 #define SCALAR float
23 #endif
24 
25 #ifndef SCALARA
26 #define SCALARA SCALAR
27 #endif
28 
29 #ifndef SCALARB
30 #define SCALARB SCALAR
31 #endif
32 
33 typedef SCALAR Scalar;
34 typedef NumTraits<Scalar>::Real RealScalar;
35 typedef Matrix<SCALARA,Dynamic,Dynamic> A;
36 typedef Matrix<SCALARB,Dynamic,Dynamic> B;
37 typedef Matrix<Scalar,Dynamic,Dynamic> C;
38 typedef Matrix<RealScalar,Dynamic,Dynamic> M;
39 
40 #ifdef HAVE_BLAS
41 
42 extern "C" {
43   #include <Eigen/src/misc/blas.h>
44 }
45 
46 static float fone = 1;
47 static float fzero = 0;
48 static double done = 1;
49 static double szero = 0;
50 static std::complex<float> cfone = 1;
51 static std::complex<float> cfzero = 0;
52 static std::complex<double> cdone = 1;
53 static std::complex<double> cdzero = 0;
54 static char notrans = 'N';
55 static char trans = 'T';
56 static char nonunit = 'N';
57 static char lower = 'L';
58 static char right = 'R';
59 static int intone = 1;
60 
blas_gemm(const MatrixXf & a,const MatrixXf & b,MatrixXf & c)61 void blas_gemm(const MatrixXf& a, const MatrixXf& b, MatrixXf& c)
62 {
63   int M = c.rows(); int N = c.cols(); int K = a.cols();
64   int lda = a.rows(); int ldb = b.rows(); int ldc = c.rows();
65 
66   sgemm_(&notrans,&notrans,&M,&N,&K,&fone,
67          const_cast<float*>(a.data()),&lda,
68          const_cast<float*>(b.data()),&ldb,&fone,
69          c.data(),&ldc);
70 }
71 
blas_gemm(const MatrixXd & a,const MatrixXd & b,MatrixXd & c)72 EIGEN_DONT_INLINE void blas_gemm(const MatrixXd& a, const MatrixXd& b, MatrixXd& c)
73 {
74   int M = c.rows(); int N = c.cols(); int K = a.cols();
75   int lda = a.rows(); int ldb = b.rows(); int ldc = c.rows();
76 
77   dgemm_(&notrans,&notrans,&M,&N,&K,&done,
78          const_cast<double*>(a.data()),&lda,
79          const_cast<double*>(b.data()),&ldb,&done,
80          c.data(),&ldc);
81 }
82 
blas_gemm(const MatrixXcf & a,const MatrixXcf & b,MatrixXcf & c)83 void blas_gemm(const MatrixXcf& a, const MatrixXcf& b, MatrixXcf& c)
84 {
85   int M = c.rows(); int N = c.cols(); int K = a.cols();
86   int lda = a.rows(); int ldb = b.rows(); int ldc = c.rows();
87 
88   cgemm_(&notrans,&notrans,&M,&N,&K,(float*)&cfone,
89          const_cast<float*>((const float*)a.data()),&lda,
90          const_cast<float*>((const float*)b.data()),&ldb,(float*)&cfone,
91          (float*)c.data(),&ldc);
92 }
93 
blas_gemm(const MatrixXcd & a,const MatrixXcd & b,MatrixXcd & c)94 void blas_gemm(const MatrixXcd& a, const MatrixXcd& b, MatrixXcd& c)
95 {
96   int M = c.rows(); int N = c.cols(); int K = a.cols();
97   int lda = a.rows(); int ldb = b.rows(); int ldc = c.rows();
98 
99   zgemm_(&notrans,&notrans,&M,&N,&K,(double*)&cdone,
100          const_cast<double*>((const double*)a.data()),&lda,
101          const_cast<double*>((const double*)b.data()),&ldb,(double*)&cdone,
102          (double*)c.data(),&ldc);
103 }
104 
105 
106 
107 #endif
108 
matlab_cplx_cplx(const M & ar,const M & ai,const M & br,const M & bi,M & cr,M & ci)109 void matlab_cplx_cplx(const M& ar, const M& ai, const M& br, const M& bi, M& cr, M& ci)
110 {
111   cr.noalias() += ar * br;
112   cr.noalias() -= ai * bi;
113   ci.noalias() += ar * bi;
114   ci.noalias() += ai * br;
115 }
116 
matlab_real_cplx(const M & a,const M & br,const M & bi,M & cr,M & ci)117 void matlab_real_cplx(const M& a, const M& br, const M& bi, M& cr, M& ci)
118 {
119   cr.noalias() += a * br;
120   ci.noalias() += a * bi;
121 }
122 
matlab_cplx_real(const M & ar,const M & ai,const M & b,M & cr,M & ci)123 void matlab_cplx_real(const M& ar, const M& ai, const M& b, M& cr, M& ci)
124 {
125   cr.noalias() += ar * b;
126   ci.noalias() += ai * b;
127 }
128 
129 template<typename A, typename B, typename C>
gemm(const A & a,const B & b,C & c)130 EIGEN_DONT_INLINE void gemm(const A& a, const B& b, C& c)
131 {
132  c.noalias() += a * b;
133 }
134 
main(int argc,char ** argv)135 int main(int argc, char ** argv)
136 {
137   std::ptrdiff_t l1 = internal::queryL1CacheSize();
138   std::ptrdiff_t l2 = internal::queryTopLevelCacheSize();
139   std::cout << "L1 cache size     = " << (l1>0 ? l1/1024 : -1) << " KB\n";
140   std::cout << "L2/L3 cache size  = " << (l2>0 ? l2/1024 : -1) << " KB\n";
141   typedef internal::gebp_traits<Scalar,Scalar> Traits;
142   std::cout << "Register blocking = " << Traits::mr << " x " << Traits::nr << "\n";
143 
144   int rep = 1;    // number of repetitions per try
145   int tries = 2;  // number of tries, we keep the best
146 
147   int s = 2048;
148   int m = s;
149   int n = s;
150   int p = s;
151   int cache_size1=-1, cache_size2=l2, cache_size3 = 0;
152 
153   bool need_help = false;
154   for (int i=1; i<argc;)
155   {
156     if(argv[i][0]=='-')
157     {
158       if(argv[i][1]=='s')
159       {
160         ++i;
161         s = atoi(argv[i++]);
162         m = n = p = s;
163         if(argv[i][0]!='-')
164         {
165           n = atoi(argv[i++]);
166           p = atoi(argv[i++]);
167         }
168       }
169       else if(argv[i][1]=='c')
170       {
171         ++i;
172         cache_size1 = atoi(argv[i++]);
173         if(argv[i][0]!='-')
174         {
175           cache_size2 = atoi(argv[i++]);
176           if(argv[i][0]!='-')
177             cache_size3 = atoi(argv[i++]);
178         }
179       }
180       else if(argv[i][1]=='t')
181       {
182         ++i;
183         tries = atoi(argv[i++]);
184       }
185       else if(argv[i][1]=='p')
186       {
187         ++i;
188         rep = atoi(argv[i++]);
189       }
190     }
191     else
192     {
193       need_help = true;
194       break;
195     }
196   }
197 
198   if(need_help)
199   {
200     std::cout << argv[0] << " -s <matrix sizes> -c <cache sizes> -t <nb tries> -p <nb repeats>\n";
201     std::cout << "   <matrix sizes> : size\n";
202     std::cout << "   <matrix sizes> : rows columns depth\n";
203     return 1;
204   }
205 
206 #if EIGEN_VERSION_AT_LEAST(3,2,90)
207   if(cache_size1>0)
208     setCpuCacheSizes(cache_size1,cache_size2,cache_size3);
209 #endif
210 
211   A a(m,p); a.setRandom();
212   B b(p,n); b.setRandom();
213   C c(m,n); c.setOnes();
214   C rc = c;
215 
216   std::cout << "Matrix sizes = " << m << "x" << p << " * " << p << "x" << n << "\n";
217   std::ptrdiff_t mc(m), nc(n), kc(p);
218   internal::computeProductBlockingSizes<Scalar,Scalar>(kc, mc, nc);
219   std::cout << "blocking size (mc x kc) = " << mc << " x " << kc << "\n";
220 
221   C r = c;
222 
223   // check the parallel product is correct
224   #if defined EIGEN_HAS_OPENMP
225   Eigen::initParallel();
226   int procs = omp_get_max_threads();
227   if(procs>1)
228   {
229     #ifdef HAVE_BLAS
230     blas_gemm(a,b,r);
231     #else
232     omp_set_num_threads(1);
233     r.noalias() += a * b;
234     omp_set_num_threads(procs);
235     #endif
236     c.noalias() += a * b;
237     if(!r.isApprox(c)) std::cerr << "Warning, your parallel product is crap!\n\n";
238   }
239   #elif defined HAVE_BLAS
240     blas_gemm(a,b,r);
241     c.noalias() += a * b;
242     if(!r.isApprox(c)) {
243       std::cout << r  - c << "\n";
244       std::cerr << "Warning, your product is crap!\n\n";
245     }
246   #else
247     if(1.*m*n*p<2000.*2000*2000)
248     {
249       gemm(a,b,c);
250       r.noalias() += a.cast<Scalar>() .lazyProduct( b.cast<Scalar>() );
251       if(!r.isApprox(c)) {
252         std::cout << r - c << "\n";
253         std::cerr << "Warning, your product is crap!\n\n";
254       }
255     }
256   #endif
257 
258   #ifdef HAVE_BLAS
259   BenchTimer tblas;
260   c = rc;
261   BENCH(tblas, tries, rep, blas_gemm(a,b,c));
262   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";
263   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";
264   #endif
265 
266   BenchTimer tmt;
267   c = rc;
268   BENCH(tmt, tries, rep, gemm(a,b,c));
269   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";
270   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";
271 
272   #ifdef EIGEN_HAS_OPENMP
273   if(procs>1)
274   {
275     BenchTimer tmono;
276     omp_set_num_threads(1);
277     Eigen::setNbThreads(1);
278     c = rc;
279     BENCH(tmono, tries, rep, gemm(a,b,c));
280     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";
281     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";
282     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";
283   }
284   #endif
285 
286   if(1.*m*n*p<30*30*30)
287   {
288       BenchTimer tmt;
289       c = rc;
290       BENCH(tmt, tries, rep, c.noalias()+=a.lazyProduct(b));
291       std::cout << "lazy 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";
292       std::cout << "lazy 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";
293   }
294 
295   #ifdef DECOUPLED
296   if((NumTraits<A::Scalar>::IsComplex) && (NumTraits<B::Scalar>::IsComplex))
297   {
298     M ar(m,p); ar.setRandom();
299     M ai(m,p); ai.setRandom();
300     M br(p,n); br.setRandom();
301     M bi(p,n); bi.setRandom();
302     M cr(m,n); cr.setRandom();
303     M ci(m,n); ci.setRandom();
304 
305     BenchTimer t;
306     BENCH(t, tries, rep, matlab_cplx_cplx(ar,ai,br,bi,cr,ci));
307     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";
308     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";
309   }
310   if((!NumTraits<A::Scalar>::IsComplex) && (NumTraits<B::Scalar>::IsComplex))
311   {
312     M a(m,p);  a.setRandom();
313     M br(p,n); br.setRandom();
314     M bi(p,n); bi.setRandom();
315     M cr(m,n); cr.setRandom();
316     M ci(m,n); ci.setRandom();
317 
318     BenchTimer t;
319     BENCH(t, tries, rep, matlab_real_cplx(a,br,bi,cr,ci));
320     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";
321     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";
322   }
323   if((NumTraits<A::Scalar>::IsComplex) && (!NumTraits<B::Scalar>::IsComplex))
324   {
325     M ar(m,p); ar.setRandom();
326     M ai(m,p); ai.setRandom();
327     M b(p,n);  b.setRandom();
328     M cr(m,n); cr.setRandom();
329     M ci(m,n); ci.setRandom();
330 
331     BenchTimer t;
332     BENCH(t, tries, rep, matlab_cplx_real(ar,ai,b,cr,ci));
333     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";
334     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";
335   }
336   #endif
337 
338   return 0;
339 }
340 
341