ViennaCL - The Vienna Computing Library: /data/development/ViennaCL/dev/viennacl/linalg/kernels/matrix_col

Go to the documentation of this file.
00001 #ifndef _VIENNACL_MATRIX_COL_KERNELS_HPP_
00002 #define _VIENNACL_MATRIX_COL_KERNELS_HPP_
00003 #include "viennacl/tools/tools.hpp"
00004 #include "viennacl/ocl/kernel.hpp"
00005 #include "viennacl/ocl/platform.hpp"
00006 #include "viennacl/ocl/utils.hpp"
00007 #include "viennacl/linalg/kernels/matrix_col_source.h"
00008 
00009 //Automatically generated file from aux-directory, do not edit manually!
00010 namespace viennacl
00011 {
00012  namespace linalg
00013  {
00014   namespace kernels
00015   {
00016    template<class TYPE, unsigned int alignment>
00017    struct matrix_col;
00018 
00019 
00021    template <>
00022    struct matrix_col<float, 16>
00023    {
00024     static std::string program_name()
00025     {
00026       return "f_matrix_col_16";
00027     }
00028     static void init()
00029     {
00030       viennacl::ocl::DOUBLE_PRECISION_CHECKER<float>::apply();
00031       static std::map<cl_context, bool> init_done;
00032       viennacl::ocl::context & context_ = viennacl::ocl::current_context();
00033       if (!init_done[context_.handle()])
00034       {
00035         std::string source;
00036         source.append(matrix_col_align1_unit_lower_triangular_substitute_inplace);
00037         source.append(matrix_col_align1_inplace_sub);
00038         source.append(matrix_col_align1_lower_triangular_substitute_inplace);
00039         source.append(matrix_col_align1_trans_vec_mul);
00040         source.append(matrix_col_align1_rank1_update);
00041         source.append(matrix_col_align1_sub);
00042         source.append(matrix_col_align1_trans_unit_upper_triangular_substitute_inplace);
00043         source.append(matrix_col_align1_lu_factorize);
00044         source.append(matrix_col_align1_add);
00045         source.append(matrix_col_align1_fft_direct);
00046         source.append(matrix_col_align1_vec_mul);
00047         source.append(matrix_col_align1_fft_radix2_local);
00048         source.append(matrix_col_align1_trans_lower_triangular_substitute_inplace);
00049         source.append(matrix_col_align1_inplace_divide);
00050         source.append(matrix_col_align1_trans_upper_triangular_substitute_inplace);
00051         source.append(matrix_col_align1_unit_upper_triangular_substitute_inplace);
00052         source.append(matrix_col_align1_inplace_add);
00053         source.append(matrix_col_align1_trans_unit_lower_triangular_substitute_inplace);
00054         source.append(matrix_col_align1_scaled_rank1_update);
00055         source.append(matrix_col_align1_clear);
00056         source.append(matrix_col_align1_fft_radix2);
00057         source.append(matrix_col_align1_cpu_inplace_mult);
00058         source.append(matrix_col_align1_fft_reorder);
00059         source.append(matrix_col_align1_inplace_mult);
00060         source.append(matrix_col_align1_upper_triangular_substitute_inplace);
00061         std::string prog_name = program_name();
00062         #ifdef VIENNACL_BUILD_INFO
00063         std::cout << "Creating program " << prog_name << std::endl;
00064         #endif
00065         context_.add_program(source, prog_name);
00066         viennacl::ocl::program & prog_ = context_.get_program(prog_name);
00067         prog_.add_kernel("unit_lower_triangular_substitute_inplace");
00068         prog_.add_kernel("inplace_sub");
00069         prog_.add_kernel("lower_triangular_substitute_inplace");
00070         prog_.add_kernel("trans_vec_mul");
00071         prog_.add_kernel("rank1_update");
00072         prog_.add_kernel("sub");
00073         prog_.add_kernel("trans_unit_upper_triangular_substitute_inplace");
00074         prog_.add_kernel("lu_factorize");
00075         prog_.add_kernel("add");
00076         prog_.add_kernel("fft_direct");
00077         prog_.add_kernel("vec_mul");
00078         prog_.add_kernel("fft_radix2_local");
00079         prog_.add_kernel("trans_lower_triangular_substitute_inplace");
00080         prog_.add_kernel("inplace_divide");
00081         prog_.add_kernel("trans_upper_triangular_substitute_inplace");
00082         prog_.add_kernel("unit_upper_triangular_substitute_inplace");
00083         prog_.add_kernel("inplace_add");
00084         prog_.add_kernel("trans_unit_lower_triangular_substitute_inplace");
00085         prog_.add_kernel("scaled_rank1_update");
00086         prog_.add_kernel("clear");
00087         prog_.add_kernel("fft_radix2");
00088         prog_.add_kernel("cpu_inplace_mult");
00089         prog_.add_kernel("fft_reorder");
00090         prog_.add_kernel("inplace_mult");
00091         prog_.add_kernel("upper_triangular_substitute_inplace");
00092         init_done[context_.handle()] = true;
00093        } //if
00094      } //init
00095     }; // struct
00096 
00097    template <>
00098    struct matrix_col<float, 1>
00099    {
00100     static std::string program_name()
00101     {
00102       return "f_matrix_col_1";
00103     }
00104     static void init()
00105     {
00106       viennacl::ocl::DOUBLE_PRECISION_CHECKER<float>::apply();
00107       static std::map<cl_context, bool> init_done;
00108       viennacl::ocl::context & context_ = viennacl::ocl::current_context();
00109       if (!init_done[context_.handle()])
00110       {
00111         std::string source;
00112         source.append(matrix_col_align1_unit_lower_triangular_substitute_inplace);
00113         source.append(matrix_col_align1_inplace_sub);
00114         source.append(matrix_col_align1_lower_triangular_substitute_inplace);
00115         source.append(matrix_col_align1_trans_vec_mul);
00116         source.append(matrix_col_align1_rank1_update);
00117         source.append(matrix_col_align1_sub);
00118         source.append(matrix_col_align1_trans_unit_upper_triangular_substitute_inplace);
00119         source.append(matrix_col_align1_lu_factorize);
00120         source.append(matrix_col_align1_add);
00121         source.append(matrix_col_align1_fft_direct);
00122         source.append(matrix_col_align1_vec_mul);
00123         source.append(matrix_col_align1_fft_radix2_local);
00124         source.append(matrix_col_align1_trans_lower_triangular_substitute_inplace);
00125         source.append(matrix_col_align1_inplace_divide);
00126         source.append(matrix_col_align1_trans_upper_triangular_substitute_inplace);
00127         source.append(matrix_col_align1_unit_upper_triangular_substitute_inplace);
00128         source.append(matrix_col_align1_inplace_add);
00129         source.append(matrix_col_align1_trans_unit_lower_triangular_substitute_inplace);
00130         source.append(matrix_col_align1_scaled_rank1_update);
00131         source.append(matrix_col_align1_clear);
00132         source.append(matrix_col_align1_fft_radix2);
00133         source.append(matrix_col_align1_cpu_inplace_mult);
00134         source.append(matrix_col_align1_fft_reorder);
00135         source.append(matrix_col_align1_inplace_mult);
00136         source.append(matrix_col_align1_upper_triangular_substitute_inplace);
00137         std::string prog_name = program_name();
00138         #ifdef VIENNACL_BUILD_INFO
00139         std::cout << "Creating program " << prog_name << std::endl;
00140         #endif
00141         context_.add_program(source, prog_name);
00142         viennacl::ocl::program & prog_ = context_.get_program(prog_name);
00143         prog_.add_kernel("unit_lower_triangular_substitute_inplace");
00144         prog_.add_kernel("inplace_sub");
00145         prog_.add_kernel("lower_triangular_substitute_inplace");
00146         prog_.add_kernel("trans_vec_mul");
00147         prog_.add_kernel("rank1_update");
00148         prog_.add_kernel("sub");
00149         prog_.add_kernel("trans_unit_upper_triangular_substitute_inplace");
00150         prog_.add_kernel("lu_factorize");
00151         prog_.add_kernel("add");
00152         prog_.add_kernel("fft_direct");
00153         prog_.add_kernel("vec_mul");
00154         prog_.add_kernel("fft_radix2_local");
00155         prog_.add_kernel("trans_lower_triangular_substitute_inplace");
00156         prog_.add_kernel("inplace_divide");
00157         prog_.add_kernel("trans_upper_triangular_substitute_inplace");
00158         prog_.add_kernel("unit_upper_triangular_substitute_inplace");
00159         prog_.add_kernel("inplace_add");
00160         prog_.add_kernel("trans_unit_lower_triangular_substitute_inplace");
00161         prog_.add_kernel("scaled_rank1_update");
00162         prog_.add_kernel("clear");
00163         prog_.add_kernel("fft_radix2");
00164         prog_.add_kernel("cpu_inplace_mult");
00165         prog_.add_kernel("fft_reorder");
00166         prog_.add_kernel("inplace_mult");
00167         prog_.add_kernel("upper_triangular_substitute_inplace");
00168         init_done[context_.handle()] = true;
00169        } //if
00170      } //init
00171     }; // struct
00172 
00173 
00174 
00176    template <>
00177    struct matrix_col<double, 16>
00178    {
00179     static std::string program_name()
00180     {
00181       return "d_matrix_col_16";
00182     }
00183     static void init()
00184     {
00185       viennacl::ocl::DOUBLE_PRECISION_CHECKER<double>::apply();
00186       static std::map<cl_context, bool> init_done;
00187       viennacl::ocl::context & context_ = viennacl::ocl::current_context();
00188       if (!init_done[context_.handle()])
00189       {
00190         std::string source;
00191         std::string fp64_ext = viennacl::ocl::current_device().double_support_extension();
00192         source.append(viennacl::tools::make_double_kernel(matrix_col_align1_unit_lower_triangular_substitute_inplace, fp64_ext));
00193         source.append(viennacl::tools::make_double_kernel(matrix_col_align1_inplace_sub, fp64_ext));
00194         source.append(viennacl::tools::make_double_kernel(matrix_col_align1_lower_triangular_substitute_inplace, fp64_ext));
00195         source.append(viennacl::tools::make_double_kernel(matrix_col_align1_trans_vec_mul, fp64_ext));
00196         source.append(viennacl::tools::make_double_kernel(matrix_col_align1_rank1_update, fp64_ext));
00197         source.append(viennacl::tools::make_double_kernel(matrix_col_align1_sub, fp64_ext));
00198         source.append(viennacl::tools::make_double_kernel(matrix_col_align1_trans_unit_upper_triangular_substitute_inplace, fp64_ext));
00199         source.append(viennacl::tools::make_double_kernel(matrix_col_align1_lu_factorize, fp64_ext));
00200         source.append(viennacl::tools::make_double_kernel(matrix_col_align1_add, fp64_ext));
00201         source.append(viennacl::tools::make_double_kernel(matrix_col_align1_fft_direct, fp64_ext));
00202         source.append(viennacl::tools::make_double_kernel(matrix_col_align1_vec_mul, fp64_ext));
00203         source.append(viennacl::tools::make_double_kernel(matrix_col_align1_fft_radix2_local, fp64_ext));
00204         source.append(viennacl::tools::make_double_kernel(matrix_col_align1_trans_lower_triangular_substitute_inplace, fp64_ext));
00205         source.append(viennacl::tools::make_double_kernel(matrix_col_align1_inplace_divide, fp64_ext));
00206         source.append(viennacl::tools::make_double_kernel(matrix_col_align1_trans_upper_triangular_substitute_inplace, fp64_ext));
00207         source.append(viennacl::tools::make_double_kernel(matrix_col_align1_unit_upper_triangular_substitute_inplace, fp64_ext));
00208         source.append(viennacl::tools::make_double_kernel(matrix_col_align1_inplace_add, fp64_ext));
00209         source.append(viennacl::tools::make_double_kernel(matrix_col_align1_trans_unit_lower_triangular_substitute_inplace, fp64_ext));
00210         source.append(viennacl::tools::make_double_kernel(matrix_col_align1_scaled_rank1_update, fp64_ext));
00211         source.append(viennacl::tools::make_double_kernel(matrix_col_align1_clear, fp64_ext));
00212         source.append(viennacl::tools::make_double_kernel(matrix_col_align1_fft_radix2, fp64_ext));
00213         source.append(viennacl::tools::make_double_kernel(matrix_col_align1_cpu_inplace_mult, fp64_ext));
00214         source.append(viennacl::tools::make_double_kernel(matrix_col_align1_fft_reorder, fp64_ext));
00215         source.append(viennacl::tools::make_double_kernel(matrix_col_align1_inplace_mult, fp64_ext));
00216         source.append(viennacl::tools::make_double_kernel(matrix_col_align1_upper_triangular_substitute_inplace, fp64_ext));
00217         std::string prog_name = program_name();
00218         #ifdef VIENNACL_BUILD_INFO
00219         std::cout << "Creating program " << prog_name << std::endl;
00220         #endif
00221         context_.add_program(source, prog_name);
00222         viennacl::ocl::program & prog_ = context_.get_program(prog_name);
00223         prog_.add_kernel("unit_lower_triangular_substitute_inplace");
00224         prog_.add_kernel("inplace_sub");
00225         prog_.add_kernel("lower_triangular_substitute_inplace");
00226         prog_.add_kernel("trans_vec_mul");
00227         prog_.add_kernel("rank1_update");
00228         prog_.add_kernel("sub");
00229         prog_.add_kernel("trans_unit_upper_triangular_substitute_inplace");
00230         prog_.add_kernel("lu_factorize");
00231         prog_.add_kernel("add");
00232         prog_.add_kernel("fft_direct");
00233         prog_.add_kernel("vec_mul");
00234         prog_.add_kernel("fft_radix2_local");
00235         prog_.add_kernel("trans_lower_triangular_substitute_inplace");
00236         prog_.add_kernel("inplace_divide");
00237         prog_.add_kernel("trans_upper_triangular_substitute_inplace");
00238         prog_.add_kernel("unit_upper_triangular_substitute_inplace");
00239         prog_.add_kernel("inplace_add");
00240         prog_.add_kernel("trans_unit_lower_triangular_substitute_inplace");
00241         prog_.add_kernel("scaled_rank1_update");
00242         prog_.add_kernel("clear");
00243         prog_.add_kernel("fft_radix2");
00244         prog_.add_kernel("cpu_inplace_mult");
00245         prog_.add_kernel("fft_reorder");
00246         prog_.add_kernel("inplace_mult");
00247         prog_.add_kernel("upper_triangular_substitute_inplace");
00248         init_done[context_.handle()] = true;
00249        } //if
00250      } //init
00251     }; // struct
00252 
00253    template <>
00254    struct matrix_col<double, 1>
00255    {
00256     static std::string program_name()
00257     {
00258       return "d_matrix_col_1";
00259     }
00260     static void init()
00261     {
00262       viennacl::ocl::DOUBLE_PRECISION_CHECKER<double>::apply();
00263       static std::map<cl_context, bool> init_done;
00264       viennacl::ocl::context & context_ = viennacl::ocl::current_context();
00265       if (!init_done[context_.handle()])
00266       {
00267         std::string source;
00268         std::string fp64_ext = viennacl::ocl::current_device().double_support_extension();
00269         source.append(viennacl::tools::make_double_kernel(matrix_col_align1_unit_lower_triangular_substitute_inplace, fp64_ext));
00270         source.append(viennacl::tools::make_double_kernel(matrix_col_align1_inplace_sub, fp64_ext));
00271         source.append(viennacl::tools::make_double_kernel(matrix_col_align1_lower_triangular_substitute_inplace, fp64_ext));
00272         source.append(viennacl::tools::make_double_kernel(matrix_col_align1_trans_vec_mul, fp64_ext));
00273         source.append(viennacl::tools::make_double_kernel(matrix_col_align1_rank1_update, fp64_ext));
00274         source.append(viennacl::tools::make_double_kernel(matrix_col_align1_sub, fp64_ext));
00275         source.append(viennacl::tools::make_double_kernel(matrix_col_align1_trans_unit_upper_triangular_substitute_inplace, fp64_ext));
00276         source.append(viennacl::tools::make_double_kernel(matrix_col_align1_lu_factorize, fp64_ext));
00277         source.append(viennacl::tools::make_double_kernel(matrix_col_align1_add, fp64_ext));
00278         source.append(viennacl::tools::make_double_kernel(matrix_col_align1_fft_direct, fp64_ext));
00279         source.append(viennacl::tools::make_double_kernel(matrix_col_align1_vec_mul, fp64_ext));
00280         source.append(viennacl::tools::make_double_kernel(matrix_col_align1_fft_radix2_local, fp64_ext));
00281         source.append(viennacl::tools::make_double_kernel(matrix_col_align1_trans_lower_triangular_substitute_inplace, fp64_ext));
00282         source.append(viennacl::tools::make_double_kernel(matrix_col_align1_inplace_divide, fp64_ext));
00283         source.append(viennacl::tools::make_double_kernel(matrix_col_align1_trans_upper_triangular_substitute_inplace, fp64_ext));
00284         source.append(viennacl::tools::make_double_kernel(matrix_col_align1_unit_upper_triangular_substitute_inplace, fp64_ext));
00285         source.append(viennacl::tools::make_double_kernel(matrix_col_align1_inplace_add, fp64_ext));
00286         source.append(viennacl::tools::make_double_kernel(matrix_col_align1_trans_unit_lower_triangular_substitute_inplace, fp64_ext));
00287         source.append(viennacl::tools::make_double_kernel(matrix_col_align1_scaled_rank1_update, fp64_ext));
00288         source.append(viennacl::tools::make_double_kernel(matrix_col_align1_clear, fp64_ext));
00289         source.append(viennacl::tools::make_double_kernel(matrix_col_align1_fft_radix2, fp64_ext));
00290         source.append(viennacl::tools::make_double_kernel(matrix_col_align1_cpu_inplace_mult, fp64_ext));
00291         source.append(viennacl::tools::make_double_kernel(matrix_col_align1_fft_reorder, fp64_ext));
00292         source.append(viennacl::tools::make_double_kernel(matrix_col_align1_inplace_mult, fp64_ext));
00293         source.append(viennacl::tools::make_double_kernel(matrix_col_align1_upper_triangular_substitute_inplace, fp64_ext));
00294         std::string prog_name = program_name();
00295         #ifdef VIENNACL_BUILD_INFO
00296         std::cout << "Creating program " << prog_name << std::endl;
00297         #endif
00298         context_.add_program(source, prog_name);
00299         viennacl::ocl::program & prog_ = context_.get_program(prog_name);
00300         prog_.add_kernel("unit_lower_triangular_substitute_inplace");
00301         prog_.add_kernel("inplace_sub");
00302         prog_.add_kernel("lower_triangular_substitute_inplace");
00303         prog_.add_kernel("trans_vec_mul");
00304         prog_.add_kernel("rank1_update");
00305         prog_.add_kernel("sub");
00306         prog_.add_kernel("trans_unit_upper_triangular_substitute_inplace");
00307         prog_.add_kernel("lu_factorize");
00308         prog_.add_kernel("add");
00309         prog_.add_kernel("fft_direct");
00310         prog_.add_kernel("vec_mul");
00311         prog_.add_kernel("fft_radix2_local");
00312         prog_.add_kernel("trans_lower_triangular_substitute_inplace");
00313         prog_.add_kernel("inplace_divide");
00314         prog_.add_kernel("trans_upper_triangular_substitute_inplace");
00315         prog_.add_kernel("unit_upper_triangular_substitute_inplace");
00316         prog_.add_kernel("inplace_add");
00317         prog_.add_kernel("trans_unit_lower_triangular_substitute_inplace");
00318         prog_.add_kernel("scaled_rank1_update");
00319         prog_.add_kernel("clear");
00320         prog_.add_kernel("fft_radix2");
00321         prog_.add_kernel("cpu_inplace_mult");
00322         prog_.add_kernel("fft_reorder");
00323         prog_.add_kernel("inplace_mult");
00324         prog_.add_kernel("upper_triangular_substitute_inplace");
00325         init_done[context_.handle()] = true;
00326        } //if
00327      } //init
00328     }; // struct
00329 
00330 
00331   }  //namespace kernels
00332  }  //namespace linalg
00333 }  //namespace viennacl
00334 #endif
/data/development/ViennaCL/dev/viennacl/linalg/kernels/matrix_col_kernels.h
