BTL: - patch from Victor (add ACML support)

- fix overflow issues

bench/btl/actions/action_cholesky.hh

@@ -43,7 +43,7 @@ public :
     init_matrix<pseudo_random>(tmp,_size);
     init_matrix<null_function>(X_stl,_size);
     STL_interface<typename Interface::real_type>::ata_product(tmp,X_stl,_size);
-    
+
     init_matrix<null_function>(C_stl,_size);
     init_matrix<null_function>(resu_stl,_size);
 
@@ -55,7 +55,7 @@ public :
     _cost = 0;
     for (int j=0; j<_size; ++j)
     {
-      int r = std::max(_size - j -1,0);
+      double r = std::max(_size - j -1,0);
       _cost += 2*(r*j+r+j);
     }
   }
@@ -104,10 +104,10 @@ public :
     Interface::matrix_to_stl(C,resu_stl);
 
 //     STL_interface<typename Interface::real_type>::cholesky(X_stl,C_stl,_size);
-// 
+//
 //     typename Interface::real_type error=
 //       STL_interface<typename Interface::real_type>::norm_diff(C_stl,resu_stl);
-// 
+//
 //     if (error>1.e-6){
 //       INFOS("WRONG CALCULATION...residual=" << error);
 //       exit(0);
@@ -126,7 +126,7 @@ private :
   typename Interface::gene_matrix C;
 
   int _size;
-  int _cost;
+  double _cost;
 };
 
 #endif

bench/btl/actions/action_hessenberg.hh

@@ -52,8 +52,8 @@ public :
     _cost = 0;
     for (int j=0; j<_size-2; ++j)
     {
-      int r = std::max(0,_size-j-1);
+      double r = std::max(0,_size-j-1);
-      int b = std::max(0,_size-j-2);
+      double b = std::max(0,_size-j-2);
       _cost += 6 + 3*b + r*r*4 + r*_size*4;
     }
   }

bench/btl/actions/action_lu_decomp.hh

@@ -40,7 +40,7 @@ public :
 
     // STL vector initialization
     init_matrix<pseudo_random>(X_stl,_size);
-    
+
     init_matrix<null_function>(C_stl,_size);
     init_matrix<null_function>(resu_stl,_size);
 
@@ -96,10 +96,10 @@ public :
     Interface::matrix_to_stl(C,resu_stl);
 
 //     STL_interface<typename Interface::real_type>::lu_decomp(X_stl,C_stl,_size);
-// 
+//
 //     typename Interface::real_type error=
 //       STL_interface<typename Interface::real_type>::norm_diff(C_stl,resu_stl);
-// 
+//
 //     if (error>1.e-6){
 //       INFOS("WRONG CALCULATION...residual=" << error);
 //       exit(0);
@@ -118,7 +118,7 @@ private :
   typename Interface::gene_matrix C;
 
   int _size;
-  int _cost;
+  double _cost;
 };
 
 #endif

bench/btl/actions/action_trisolve.hh

@@ -130,7 +130,7 @@ private :
   typename Interface::gene_vector B;
 
   int _size;
-  int _cost;
+  double _cost;
 };
 
 #endif

bench/btl/actions/action_trisolve_matrix.hh

@@ -125,12 +125,12 @@ public :
     // calculation check
 
 //     Interface::matrix_to_stl(X,resu_stl);
-// 
+//
 //     STL_interface<typename Interface::real_type>::matrix_matrix_product(A_stl,B_stl,X_stl,_size);
-// 
+//
 //     typename Interface::real_type error=
 //       STL_interface<typename Interface::real_type>::norm_diff(X_stl,resu_stl);
-// 
+//
 //     if (error>1.e-6){
 //       INFOS("WRONG CALCULATION...residual=" << error);
 // //       exit(1);

bench/btl/cmake/FindACML.cmake

@@ -0,0 +1,60 @@
+# include(FindLibraryWithDebug)
+
+if (ACML_INCLUDES AND ACML_LIBRARIES)
+  set(ACML_FIND_QUIETLY TRUE)
+endif (ACML_INCLUDES AND ACML_LIBRARIES)
+
+find_path(ACML_INCLUDES
+  NAMES
+  acml.h
+  PATHS
+  $ENV{ACMLDIR}/include
+  $ENV{ACML_DIR}/include
+  ${INCLUDE_INSTALL_DIR}
+)
+
+find_library(ACML_LIBRARIES
+  NAMES
+  acml_mp acml_mv
+  PATHS
+  $ENV{ACMLDIR}/lib
+  $ENV{ACML_DIR}/lib
+  ${LIB_INSTALL_DIR}
+)
+
+find_file(ACML_LIBRARIES
+  NAMES
+  libacml_mp.so
+  PATHS
+  /usr/lib
+  $ENV{ACMLDIR}/lib
+  ${LIB_INSTALL_DIR}
+)
+
+if(NOT ACML_LIBRARIES)
+    message(STATUS "Multi-threaded library not found, looking for single-threaded")
+    find_library(ACML_LIBRARIES
+        NAMES
+        acml acml_mv
+        PATHS
+        $ENV{ACMLDIR}/lib
+        $ENV{ACML_DIR}/lib
+        ${LIB_INSTALL_DIR}
+        )
+    find_file(ACML_LIBRARIES
+        libacml.so libacml_mv.so
+        PATHS
+        /usr/lib
+        $ENV{ACMLDIR}/lib
+        ${LIB_INSTALL_DIR}
+        )
+endif()
+
+
+
+
+include(FindPackageHandleStandardArgs)
+find_package_handle_standard_args(ACML DEFAULT_MSG
+                                  ACML_INCLUDES ACML_LIBRARIES)
+
+mark_as_advanced(ACML_INCLUDES ACML_LIBRARIES)

bench/btl/libs/C_BLAS/CMakeLists.txt

@@ -28,3 +28,14 @@ if (GOTO_FOUND)
     set_target_properties(btl_goto PROPERTIES COMPILE_FLAGS "-DCBLASNAME=GOTO -DPUREBLAS")
   endif(BUILD_btl_goto)
 endif (GOTO_FOUND)
+
+
+find_package(ACML)
+if (ACML_FOUND)
+  include_directories(${ACML_INCLUDES} ${PROJECT_SOURCE_DIR}/libs/f77)
+  btl_add_bench(btl_acml main.cpp)
+  if(BUILD_btl_acml)
+    target_link_libraries(btl_acml ${ACML_LIBRARIES} )
+    set_target_properties(btl_acml PROPERTIES COMPILE_FLAGS "-DCBLASNAME=ACML -DHAS_LAPACK=1 -DPUREBLAS")
+  endif(BUILD_btl_acml)
+endif (ACML_FOUND)

bench/btl/libs/C_BLAS/C_BLAS_interface.hh

@@ -248,7 +248,15 @@ public :
   }
 
   static inline void hessenberg(const gene_matrix & X, gene_matrix & C, int N){
+#ifdef PUREBLAS
+    {
+    int N2 = N*N;
+    int inc = 1;
+    scopy_(&N2, X, &inc, C, &inc);
+    }
+#else
     cblas_scopy(N*N, X, 1, C, 1);
+#endif
     int info = 0;
     int ilo = 1;
     int ihi = N;
@@ -260,7 +268,15 @@ public :
   }
 
   static inline void tridiagonalization(const gene_matrix & X, gene_matrix & C, int N){
+#ifdef PUREBLAS
+    {
+    int N2 = N*N;
+    int inc = 1;
+    scopy_(&N2, X, &inc, C, &inc);
+    }
+#else
     cblas_scopy(N*N, X, 1, C, 1);
+#endif
     char uplo = 'U';
     int info = 0;
     int ilo = 1;