Sync from Head.

11 files changed, 116 insertions, 32 deletions

diff --git a/bench/btl/actions/basic_actions.hh b/bench/btl/actions/basic_actions.hh
index a3333ea26..62442f01f 100644
--- a/bench/btl/actions/basic_actions.hh
+++ b/bench/btl/actions/basic_actions.hh
@@ -6,7 +6,7 @@
 #include "action_atv_product.hh"
 
 #include "action_matrix_matrix_product.hh"
-// #include "action_ata_product.hh"
+#include "action_ata_product.hh"
 #include "action_aat_product.hh"
 
 #include "action_trisolve.hh"
diff --git a/bench/btl/libs/BLAS/blas_interface_impl.hh b/bench/btl/libs/BLAS/blas_interface_impl.hh
index fc4ba2a1f..9e0a64905 100644
--- a/bench/btl/libs/BLAS/blas_interface_impl.hh
+++ b/bench/btl/libs/BLAS/blas_interface_impl.hh
@@ -46,9 +46,9 @@ public :
     BLAS_FUNC(gemm)(&notrans,&notrans,&N,&N,&N,&fone,A,&N,B,&N,&fzero,X,&N);
   }
 
-//   static inline void ata_product(gene_matrix & A, gene_matrix & X, int N){
-//     ssyrk_(&lower,&trans,&N,&N,&fone,A,&N,&fzero,X,&N);
-//   }
+  static inline void ata_product(gene_matrix & A, gene_matrix & X, int N){
+    BLAS_FUNC(syrk)(&lower,&trans,&N,&N,&fone,A,&N,&fzero,X,&N);
+  }
 
   static inline void aat_product(gene_matrix & A, gene_matrix & X, int N){
     BLAS_FUNC(syrk)(&lower,&notrans,&N,&N,&fone,A,&N,&fzero,X,&N);
diff --git a/bench/btl/libs/BLAS/main.cpp b/bench/btl/libs/BLAS/main.cpp
index 564d55ef2..fd991490a 100644
--- a/bench/btl/libs/BLAS/main.cpp
+++ b/bench/btl/libs/BLAS/main.cpp
@@ -48,7 +48,7 @@ int main()
   bench<Action_rot<blas_interface<REAL_TYPE> > >(MIN_AXPY,MAX_AXPY,NB_POINT);
 
   bench<Action_matrix_matrix_product<blas_interface<REAL_TYPE> > >(MIN_MM,MAX_MM,NB_POINT);
-//   bench<Action_ata_product<blas_interface<REAL_TYPE> > >(MIN_MM,MAX_MM,NB_POINT);
+  bench<Action_ata_product<blas_interface<REAL_TYPE> > >(MIN_MM,MAX_MM,NB_POINT);
   bench<Action_aat_product<blas_interface<REAL_TYPE> > >(MIN_MM,MAX_MM,NB_POINT);
 
   bench<Action_trisolve<blas_interface<REAL_TYPE> > >(MIN_MM,MAX_MM,NB_POINT);
diff --git a/bench/btl/libs/STL/STL_interface.hh b/bench/btl/libs/STL/STL_interface.hh
index ef4cc9233..16658c4ba 100644
--- a/bench/btl/libs/STL/STL_interface.hh
+++ b/bench/btl/libs/STL/STL_interface.hh
@@ -78,18 +78,18 @@ public :
         cible[i][j]=source[i][j];
   }
 
-//   static inline void ata_product(const gene_matrix & A, gene_matrix & X, int N)
-//   {
-//     real somme;
-//     for (int j=0;j<N;j++){
-//       for (int i=0;i<N;i++){
-//         somme=0.0;
-//         for (int k=0;k<N;k++)
-//           somme += A[i][k]*A[j][k];
-//         X[j][i]=somme;
-//       }
-//     }
-//   }
+  static inline void ata_product(const gene_matrix & A, gene_matrix & X, int N)
+  {
+    real somme;
+    for (int j=0;j<N;j++){
+      for (int i=0;i<N;i++){
+        somme=0.0;
+        for (int k=0;k<N;k++)
+          somme += A[i][k]*A[j][k];
+        X[j][i]=somme;
+      }
+    }
+  }
 
   static inline void aat_product(const gene_matrix & A, gene_matrix & X, int N)
   {
diff --git a/bench/btl/libs/blaze/blaze_interface.hh b/bench/btl/libs/blaze/blaze_interface.hh
index ee1523944..3635339ef 100644
--- a/bench/btl/libs/blaze/blaze_interface.hh
+++ b/bench/btl/libs/blaze/blaze_interface.hh
@@ -80,35 +80,35 @@ public :
     }
   }
 
-  static inline void matrix_matrix_product(const gene_matrix & A, const gene_matrix & B, gene_matrix & X, int N){
+  static EIGEN_DONT_INLINE void matrix_matrix_product(const gene_matrix & A, const gene_matrix & B, gene_matrix & X, int N){
     X = (A*B);
   }
 
-  static inline void transposed_matrix_matrix_product(const gene_matrix & A, const gene_matrix & B, gene_matrix & X, int N){
+  static EIGEN_DONT_INLINE void transposed_matrix_matrix_product(const gene_matrix & A, const gene_matrix & B, gene_matrix & X, int N){
     X = (trans(A)*trans(B));
   }
 
-  static inline void ata_product(const gene_matrix & A, gene_matrix & X, int N){
+  static EIGEN_DONT_INLINE void ata_product(const gene_matrix & A, gene_matrix & X, int N){
     X = (trans(A)*A);
   }
 
-  static inline void aat_product(const gene_matrix & A, gene_matrix & X, int N){
+  static EIGEN_DONT_INLINE void aat_product(const gene_matrix & A, gene_matrix & X, int N){
     X = (A*trans(A));
   }
 
-  static inline void matrix_vector_product(gene_matrix & A, gene_vector & B, gene_vector & X, int N){
+  static EIGEN_DONT_INLINE void matrix_vector_product(gene_matrix & A, gene_vector & B, gene_vector & X, int N){
     X = (A*B);
   }
 
-  static inline void atv_product(gene_matrix & A, gene_vector & B, gene_vector & X, int N){
+  static EIGEN_DONT_INLINE void atv_product(gene_matrix & A, gene_vector & B, gene_vector & X, int N){
     X = (trans(A)*B);
   }
 
-  static inline void axpy(const real coef, const gene_vector & X, gene_vector & Y, int N){
+  static EIGEN_DONT_INLINE void axpy(const real coef, const gene_vector & X, gene_vector & Y, int N){
     Y += coef * X;
   }
 
-  static inline void axpby(real a, const gene_vector & X, real b, gene_vector & Y, int N){
+  static EIGEN_DONT_INLINE void axpby(real a, const gene_vector & X, real b, gene_vector & Y, int N){
     Y = a*X + b*Y;
   }
 
diff --git a/bench/btl/libs/blaze/main.cpp b/bench/btl/libs/blaze/main.cpp
index 80e8f4eaa..ccae0cbd5 100644
--- a/bench/btl/libs/blaze/main.cpp
+++ b/bench/btl/libs/blaze/main.cpp
@@ -30,9 +30,9 @@ int main()
 
   bench<Action_matrix_vector_product<blaze_interface<REAL_TYPE> > >(MIN_MV,MAX_MV,NB_POINT);
   bench<Action_atv_product<blaze_interface<REAL_TYPE> > >(MIN_MV,MAX_MV,NB_POINT);
-//   bench<Action_matrix_matrix_product<blaze_interface<REAL_TYPE> > >(MIN_MM,MAX_MM,NB_POINT);
-//   bench<Action_ata_product<blaze_interface<REAL_TYPE> > >(MIN_MM,MAX_MM,NB_POINT);
-//   bench<Action_aat_product<blaze_interface<REAL_TYPE> > >(MIN_MM,MAX_MM,NB_POINT);
+  bench<Action_matrix_matrix_product<blaze_interface<REAL_TYPE> > >(MIN_MM,MAX_MM,NB_POINT);
+  bench<Action_ata_product<blaze_interface<REAL_TYPE> > >(MIN_MM,MAX_MM,NB_POINT);
+  bench<Action_aat_product<blaze_interface<REAL_TYPE> > >(MIN_MM,MAX_MM,NB_POINT);
 
   return 0;
 }
diff --git a/bench/btl/libs/eigen3/eigen3_interface.hh b/bench/btl/libs/eigen3/eigen3_interface.hh
index b821fd721..2e302d072 100644
--- a/bench/btl/libs/eigen3/eigen3_interface.hh
+++ b/bench/btl/libs/eigen3/eigen3_interface.hh
@@ -92,9 +92,11 @@ public :
     X.noalias() = A.transpose()*B.transpose();
   }
 
-//   static inline void ata_product(const gene_matrix & A, gene_matrix & X, int  /*N*/){
-//     X.noalias() = A.transpose()*A;
-//   }
+  static inline void ata_product(const gene_matrix & A, gene_matrix & X, int  /*N*/){
+    //X.noalias() = A.transpose()*A;
+    X.template triangularView<Lower>().setZero();
+    X.template selfadjointView<Lower>().rankUpdate(A.transpose());
+  }
 
   static inline void aat_product(const gene_matrix & A, gene_matrix & X, int  /*N*/){
     X.template triangularView<Lower>().setZero();
diff --git a/bench/btl/libs/eigen3/main_matmat.cpp b/bench/btl/libs/eigen3/main_matmat.cpp
index 926fa2b01..052810a16 100644
--- a/bench/btl/libs/eigen3/main_matmat.cpp
+++ b/bench/btl/libs/eigen3/main_matmat.cpp
@@ -25,7 +25,7 @@ BTL_MAIN;
 int main()
 {
   bench<Action_matrix_matrix_product<eigen3_interface<REAL_TYPE> > >(MIN_MM,MAX_MM,NB_POINT);
-//   bench<Action_ata_product<eigen3_interface<REAL_TYPE> > >(MIN_MM,MAX_MM,NB_POINT);
+  bench<Action_ata_product<eigen3_interface<REAL_TYPE> > >(MIN_MM,MAX_MM,NB_POINT);
   bench<Action_aat_product<eigen3_interface<REAL_TYPE> > >(MIN_MM,MAX_MM,NB_POINT);
   bench<Action_trmm<eigen3_interface<REAL_TYPE> > >(MIN_MM,MAX_MM,NB_POINT);
 
diff --git a/bench/perf_monitoring/gemm/changesets.txt b/bench/perf_monitoring/gemm/changesets.txt
index af8eb9b8f..4bff3cc4a 100644
--- a/bench/perf_monitoring/gemm/changesets.txt
+++ b/bench/perf_monitoring/gemm/changesets.txt
@@ -59,3 +59,6 @@ before-evaluators
 9174:d228bc282ac9   # merge
 9212:c90098affa7b   # Fix performance regression introduced in changeset 8aad8f35c955
 9213:9f1c14e4694b   # Fix performance regression in dgemm introduced by changeset 81d53c711775
+3.3-beta2
+3.3-rc1
+3.3.0
diff --git a/bench/perf_monitoring/gemm/gemv.cpp b/bench/perf_monitoring/gemm/gemv.cpp
new file mode 100644
index 000000000..b7441a357
--- /dev/null
+++ b/bench/perf_monitoring/gemm/gemv.cpp
@@ -0,0 +1,68 @@
+#include <iostream>
+#include <fstream>
+#include <vector>
+#include <Eigen/Core>
+#include "../../BenchTimer.h"
+using namespace Eigen;
+
+#ifndef SCALAR
+#error SCALAR must be defined
+#endif
+
+typedef SCALAR Scalar;
+
+typedef Matrix<Scalar,Dynamic,Dynamic> Mat;
+typedef Matrix<Scalar,Dynamic,1>       Vec;
+
+EIGEN_DONT_INLINE
+void gemv(const Mat &A, const Vec &B, Vec &C)
+{
+  C.noalias() += A * B;
+}
+
+EIGEN_DONT_INLINE
+double bench(long m, long n)
+{
+  Mat A(m,n);
+  Vec B(n);
+  Vec C(m);
+  A.setRandom();
+  B.setRandom();
+  C.setZero();
+  
+  BenchTimer t;
+  
+  double up = 1e9*4/sizeof(Scalar);
+  double tm0 = 4, tm1 = 10;
+  if(NumTraits<Scalar>::IsComplex)
+  {
+    up /= 4;
+    tm0 = 2;
+    tm1 = 4;
+  }
+  
+  double flops = 2. * m * n;
+  long rep = std::max(1., std::min(100., up/flops) );
+  long tries = std::max(tm0, std::min(tm1, up/flops) );
+  
+  BENCH(t, tries, rep, gemv(A,B,C));
+  
+  return 1e-9 * rep * flops / t.best();
+}
+
+int main(int argc, char **argv)
+{
+  std::vector<double> results;
+  
+  std::ifstream settings("gemv_settings.txt");
+  long m, n;
+  while(settings >> m >> n)
+  {
+    //std::cerr << "  Testing " << m << " " << n << " " << k << std::endl;
+    results.push_back( bench(m, n) );
+  }
+  
+  std::cout << RowVectorXd::Map(results.data(), results.size());
+  
+  return 0;
+}
diff --git a/bench/perf_monitoring/gemm/gemv_settings.txt b/bench/perf_monitoring/gemm/gemv_settings.txt
new file mode 100644
index 000000000..21a5ee051
--- /dev/null
+++ b/bench/perf_monitoring/gemm/gemv_settings.txt
@@ -0,0 +1,11 @@
+8 8
+9 9
+24 24
+239 239
+240 240
+2400 24
+24 2400
+24 240
+2400 2400
+4800 23
+23 4800