add a partial LU bench in BTL

6 files changed, 23 insertions, 4 deletions

diff --git a/bench/btl/data/action_settings.txt b/bench/btl/data/action_settings.txt
index ac27a482c..ab1880e84 100644
--- a/bench/btl/data/action_settings.txt
+++ b/bench/btl/data/action_settings.txt
@@ -8,7 +8,8 @@ matrix_vector ; "{/*1.5 matrix vector product}" ; "matrix size" ; 4:2048
 trisolve ; "{/*1.5 triangular solver (X = inv(L) X)}" ; "size" ; 4:2048
 matrix_trisolve ; "{/*1.5 matrix triangular solver (M = inv(L) M)}" ; "size" ; 4:2048
 cholesky ; "{/*1.5 Cholesky decomposition}" ; "matrix size" ; 4:2048
-lu_decomp ; "{/*1.5 LU decomposition}" ; "matrix size" ; 4:2048
+lu_decomp ; "{/*1.5 Complete LU decomposition}" ; "matrix size" ; 4:2048
+partial_lu_decomp ; "{/*1.5 Partial LU decomposition}" ; "matrix size" ; 4:2048
 tridiagonalization ; "{/*1.5 Tridiagonalization}" ; "matrix size" ; 4:2048
 hessenberg ; "{/*1.5 Hessenberg decomposition}" ; "matrix size" ; 4:2048
 symv ; "{/*1.5 symmetric matrix vector product}" ; "matrix size" ; 4:2048
diff --git a/bench/btl/data/go_mean b/bench/btl/data/go_mean
index 5e0529b8a..38c50d226 100755
--- a/bench/btl/data/go_mean
+++ b/bench/btl/data/go_mean
@@ -1,4 +1,4 @@
-#! /bin/bash
+#!/bin/bash
 
 if [ $# < 1 ]; then
   echo "Usage: $0 working_directory [tiny|large [prefix]]"
@@ -42,6 +42,7 @@ source mk_mean_script.sh trisolve $1 11 100 300 1000 $mode $prefix
 source mk_mean_script.sh matrix_trisolve $1 11 100 300 1000 $mode $prefix
 source mk_mean_script.sh cholesky $1 11 100 300 1000 $mode $prefix
 source mk_mean_script.sh lu_decomp $1 11 100 300 1000 $mode $prefix
+source mk_mean_script.sh partial_lu_decomp $1 11 100 300 1000 $mode $prefix
 source mk_mean_script.sh tridiagonalization $1 11 100 300 1000 $mode $prefix
 source mk_mean_script.sh hessenberg $1 11 100 300 1000 $mode $prefix
 source mk_mean_script.sh symv $1 11 50 300 1000 $mode $prefix
diff --git a/bench/btl/libs/C_BLAS/C_BLAS_interface.hh b/bench/btl/libs/C_BLAS/C_BLAS_interface.hh
index db02c2e2e..c85c4e818 100644
--- a/bench/btl/libs/C_BLAS/C_BLAS_interface.hh
+++ b/bench/btl/libs/C_BLAS/C_BLAS_interface.hh
@@ -61,6 +61,7 @@ extern "C"
   void sgehrd_( const int *n, int *ilo, int *ihi, float *a, const int *lda, float *tau, float *work, int *lwork, int *info );
 
   // LU row pivoting
+//   void dgetrf_( int *m, int *n, double *a, int *lda, int *ipiv, int *info );
 //   void sgetrf_(const int* m, const int* n, float *a, const int* ld, int* ipivot, int* info);
   // LU full pivoting
   void sgetc2_(const int* n, float *a, const int *lda, int *ipiv, int *jpiv, int*info );
@@ -160,7 +161,7 @@ public :
     cblas_ssymv(CblasColMajor,CblasLower,N,1.0,A,N,B,1,0.0,X,1);
     #endif
   }
-  
+
   static inline void syr2(gene_matrix & A, gene_vector & B, gene_vector & X, int N){
     #ifdef PUREBLAS
     ssyr2_(&lower,&N,&fone,B,&intone,X,&intone,A,&N);
@@ -247,6 +248,15 @@ public :
     sgetc2_(&N, C, &N, ipiv, jpiv, &info);
   }
 
+  static inline void partial_lu_decomp(const gene_matrix & X, gene_matrix & C, int N){
+    int N2 = N*N;
+    scopy_(&N2, X, &intone, C, &intone);
+    char uplo = 'L';
+    int info = 0;
+    int * ipiv = (int*)alloca(sizeof(int)*N);
+    sgetrf_(&N, &N, C, &N, ipiv, &info);
+  }
+
   static inline void hessenberg(const gene_matrix & X, gene_matrix & C, int N){
 #ifdef PUREBLAS
     {
diff --git a/bench/btl/libs/C_BLAS/main.cpp b/bench/btl/libs/C_BLAS/main.cpp
index 57cb9930e..81cac6c29 100644
--- a/bench/btl/libs/C_BLAS/main.cpp
+++ b/bench/btl/libs/C_BLAS/main.cpp
@@ -24,6 +24,7 @@
 
 #include "action_cholesky.hh"
 #include "action_lu_decomp.hh"
+#include "action_partial_lu_decomp.hh"
 #include "action_trisolve_matrix.hh"
 
 #ifdef HAS_LAPACK
@@ -54,6 +55,7 @@ int main()
 
   #ifdef HAS_LAPACK
   bench<Action_lu_decomp<C_BLAS_interface<REAL_TYPE> > >(MIN_MM,MAX_MM,NB_POINT);
+  bench<Action_partial_lu_decomp<C_BLAS_interface<REAL_TYPE> > >(MIN_MM,MAX_MM,NB_POINT);
   bench<Action_hessenberg<C_BLAS_interface<REAL_TYPE> > >(MIN_MM,MAX_MM,NB_POINT);
   bench<Action_tridiagonalization<C_BLAS_interface<REAL_TYPE> > >(MIN_MM,MAX_MM,NB_POINT);
   #endif
diff --git a/bench/btl/libs/eigen2/eigen2_interface.hh b/bench/btl/libs/eigen2/eigen2_interface.hh
index 5d70279ca..186a485cf 100644
--- a/bench/btl/libs/eigen2/eigen2_interface.hh
+++ b/bench/btl/libs/eigen2/eigen2_interface.hh
@@ -202,7 +202,10 @@ public :
 
   static inline void lu_decomp(const gene_matrix & X, gene_matrix & C, int N){
     C = X.lu().matrixLU();
-//     C = X.inverse();
+  }
+
+  static inline void partial_lu_decomp(const gene_matrix & X, gene_matrix & C, int N){
+    C = X.partialLu().matrixLU();
   }
 
   static inline void tridiagonalization(const gene_matrix & X, gene_matrix & C, int N){
diff --git a/bench/btl/libs/eigen2/main_adv.cpp b/bench/btl/libs/eigen2/main_adv.cpp
index 6dd9d3884..e6c7fe7cd 100644
--- a/bench/btl/libs/eigen2/main_adv.cpp
+++ b/bench/btl/libs/eigen2/main_adv.cpp
@@ -23,6 +23,7 @@
 #include "action_cholesky.hh"
 #include "action_hessenberg.hh"
 #include "action_lu_decomp.hh"
+#include "action_partial_lu_decomp.hh"
 
 BTL_MAIN;
 
@@ -32,6 +33,7 @@ int main()
   bench<Action_trisolve_matrix<eigen2_interface<REAL_TYPE> > >(MIN_MM,MAX_MM,NB_POINT);
   bench<Action_cholesky<eigen2_interface<REAL_TYPE> > >(MIN_MM,MAX_MM,NB_POINT);
   bench<Action_lu_decomp<eigen2_interface<REAL_TYPE> > >(MIN_MM,MAX_MM,NB_POINT);
+  bench<Action_partial_lu_decomp<eigen2_interface<REAL_TYPE> > >(MIN_MM,MAX_MM,NB_POINT);
 
   bench<Action_hessenberg<eigen2_interface<REAL_TYPE> > >(MIN_MM,MAX_MM,NB_POINT);
   bench<Action_tridiagonalization<eigen2_interface<REAL_TYPE> > >(MIN_MM,MAX_MM,NB_POINT);