libflame
revision_anchor
|
00001 /* 00002 00003 Copyright (C) 2014, The University of Texas at Austin 00004 00005 This file is part of libflame and is available under the 3-Clause 00006 BSD license, which can be found in the LICENSE file at the top-level 00007 directory, or at http://opensource.org/licenses/BSD-3-Clause 00008 00009 */ 00010 00011 FLA_Error FLA_Tridiag_UT_l_blk_var1( FLA_Obj A, FLA_Obj T ); 00012 FLA_Error FLA_Tridiag_UT_l_unb_var1( FLA_Obj A, FLA_Obj T ); 00013 FLA_Error FLA_Tridiag_UT_l_step_unb_var1( FLA_Obj A, FLA_Obj T ); 00014 00015 FLA_Error FLA_Tridiag_UT_l_blk_var2( FLA_Obj A, FLA_Obj T ); 00016 FLA_Error FLA_Tridiag_UT_l_blf_var2( FLA_Obj A, FLA_Obj T ); 00017 FLA_Error FLA_Tridiag_UT_l_unb_var2( FLA_Obj A, FLA_Obj T ); 00018 FLA_Error FLA_Tridiag_UT_l_step_unb_var2( FLA_Obj A, FLA_Obj T ); 00019 00020 FLA_Error FLA_Tridiag_UT_l_blk_var3( FLA_Obj A, FLA_Obj T ); 00021 FLA_Error FLA_Tridiag_UT_l_blf_var3( FLA_Obj A, FLA_Obj T ); 00022 FLA_Error FLA_Tridiag_UT_l_unb_var3( FLA_Obj A, FLA_Obj T ); 00023 FLA_Error FLA_Tridiag_UT_l_step_unb_var3( FLA_Obj A, FLA_Obj Z, FLA_Obj T ); 00024 00025 FLA_Error FLA_Tridiag_UT_l_opt_var1( FLA_Obj A, FLA_Obj T ); 00026 FLA_Error FLA_Tridiag_UT_l_step_opt_var1( FLA_Obj A, FLA_Obj T ); 00027 FLA_Error FLA_Tridiag_UT_l_step_ops_var1( int m_A, 00028 int m_T, 00029 float* buff_A, int rs_A, int cs_A, 00030 float* buff_T, int rs_T, int cs_T ); 00031 FLA_Error FLA_Tridiag_UT_l_step_opd_var1( int m_A, 00032 int m_T, 00033 double* buff_A, int rs_A, int cs_A, 00034 double* buff_T, int rs_T, int cs_T ); 00035 FLA_Error FLA_Tridiag_UT_l_step_opc_var1( int m_A, 00036 int m_T, 00037 scomplex* buff_A, int rs_A, int cs_A, 00038 scomplex* buff_T, int rs_T, int cs_T ); 00039 FLA_Error FLA_Tridiag_UT_l_step_opz_var1( int m_A, 00040 int m_T, 00041 dcomplex* buff_A, int rs_A, int cs_A, 00042 dcomplex* buff_T, int rs_T, int cs_T ); 00043 00044 FLA_Error FLA_Tridiag_UT_l_opt_var2( FLA_Obj A, FLA_Obj T ); 00045 FLA_Error FLA_Tridiag_UT_l_step_opt_var2( FLA_Obj A, FLA_Obj T ); 00046 FLA_Error FLA_Tridiag_UT_l_step_ops_var2( int m_A, 00047 int m_T, 00048 float* buff_A, int rs_A, int cs_A, 00049 float* buff_T, int rs_T, int cs_T ); 00050 FLA_Error FLA_Tridiag_UT_l_step_opd_var2( int m_A, 00051 int m_T, 00052 double* buff_A, int rs_A, int cs_A, 00053 double* buff_T, int rs_T, int cs_T ); 00054 FLA_Error FLA_Tridiag_UT_l_step_opc_var2( int m_A, 00055 int m_T, 00056 scomplex* buff_A, int rs_A, int cs_A, 00057 scomplex* buff_T, int rs_T, int cs_T ); 00058 FLA_Error FLA_Tridiag_UT_l_step_opz_var2( int m_A, 00059 int m_T, 00060 dcomplex* buff_A, int rs_A, int cs_A, 00061 dcomplex* buff_T, int rs_T, int cs_T ); 00062 00063 FLA_Error FLA_Tridiag_UT_l_opt_var3( FLA_Obj A, FLA_Obj T ); 00064 FLA_Error FLA_Tridiag_UT_l_step_opt_var3( FLA_Obj A, FLA_Obj Z, FLA_Obj T ); 00065 FLA_Error FLA_Tridiag_UT_l_step_ops_var3( int m_A, 00066 int m_T, 00067 float* buff_A, int rs_A, int cs_A, 00068 float* buff_Z, int rs_Z, int cs_Z, 00069 float* buff_T, int rs_T, int cs_T ); 00070 FLA_Error FLA_Tridiag_UT_l_step_opd_var3( int m_A, 00071 int m_T, 00072 double* buff_A, int rs_A, int cs_A, 00073 double* buff_Z, int rs_Z, int cs_Z, 00074 double* buff_T, int rs_T, int cs_T ); 00075 FLA_Error FLA_Tridiag_UT_l_step_opc_var3( int m_A, 00076 int m_T, 00077 scomplex* buff_A, int rs_A, int cs_A, 00078 scomplex* buff_Z, int rs_Z, int cs_Z, 00079 scomplex* buff_T, int rs_T, int cs_T ); 00080 FLA_Error FLA_Tridiag_UT_l_step_opz_var3( int m_A, 00081 int m_T, 00082 dcomplex* buff_A, int rs_A, int cs_A, 00083 dcomplex* buff_Z, int rs_Z, int cs_Z, 00084 dcomplex* buff_T, int rs_T, int cs_T ); 00085 00086 FLA_Error FLA_Tridiag_UT_l_ofu_var1( FLA_Obj A, FLA_Obj T ); 00087 FLA_Error FLA_Tridiag_UT_l_step_ofu_var1( FLA_Obj A, FLA_Obj T ); 00088 FLA_Error FLA_Tridiag_UT_l_step_ofs_var1( int m_A, 00089 int m_T, 00090 float* buff_A, int rs_A, int cs_A, 00091 float* buff_T, int rs_T, int cs_T ); 00092 FLA_Error FLA_Tridiag_UT_l_step_ofd_var1( int m_A, 00093 int m_T, 00094 double* buff_A, int rs_A, int cs_A, 00095 double* buff_T, int rs_T, int cs_T ); 00096 FLA_Error FLA_Tridiag_UT_l_step_ofc_var1( int m_A, 00097 int m_T, 00098 scomplex* buff_A, int rs_A, int cs_A, 00099 scomplex* buff_T, int rs_T, int cs_T ); 00100 FLA_Error FLA_Tridiag_UT_l_step_ofz_var1( int m_A, 00101 int m_T, 00102 dcomplex* buff_A, int rs_A, int cs_A, 00103 dcomplex* buff_T, int rs_T, int cs_T ); 00104 00105 FLA_Error FLA_Tridiag_UT_l_ofu_var2( FLA_Obj A, FLA_Obj T ); 00106 FLA_Error FLA_Tridiag_UT_l_step_ofu_var2( FLA_Obj A, FLA_Obj T ); 00107 FLA_Error FLA_Tridiag_UT_l_step_ofs_var2( int m_A, 00108 int m_T, 00109 float* buff_A, int rs_A, int cs_A, 00110 float* buff_T, int rs_T, int cs_T ); 00111 FLA_Error FLA_Tridiag_UT_l_step_ofd_var2( int m_A, 00112 int m_T, 00113 double* buff_A, int rs_A, int cs_A, 00114 double* buff_T, int rs_T, int cs_T ); 00115 FLA_Error FLA_Tridiag_UT_l_step_ofc_var2( int m_A, 00116 int m_T, 00117 scomplex* buff_A, int rs_A, int cs_A, 00118 scomplex* buff_T, int rs_T, int cs_T ); 00119 FLA_Error FLA_Tridiag_UT_l_step_ofz_var2( int m_A, 00120 int m_T, 00121 dcomplex* buff_A, int rs_A, int cs_A, 00122 dcomplex* buff_T, int rs_T, int cs_T ); 00123 00124 FLA_Error FLA_Tridiag_UT_l_ofu_var3( FLA_Obj A, FLA_Obj T ); 00125 FLA_Error FLA_Tridiag_UT_l_step_ofu_var3( FLA_Obj A, FLA_Obj Z, FLA_Obj T ); 00126 FLA_Error FLA_Tridiag_UT_l_step_ofs_var3( int m_A, 00127 int m_T, 00128 float* buff_A, int rs_A, int cs_A, 00129 float* buff_Z, int rs_Z, int cs_Z, 00130 float* buff_T, int rs_T, int cs_T ); 00131 FLA_Error FLA_Tridiag_UT_l_step_ofd_var3( int m_A, 00132 int m_T, 00133 double* buff_A, int rs_A, int cs_A, 00134 double* buff_Z, int rs_Z, int cs_Z, 00135 double* buff_T, int rs_T, int cs_T ); 00136 FLA_Error FLA_Tridiag_UT_l_step_ofc_var3( int m_A, 00137 int m_T, 00138 scomplex* buff_A, int rs_A, int cs_A, 00139 scomplex* buff_Z, int rs_Z, int cs_Z, 00140 scomplex* buff_T, int rs_T, int cs_T ); 00141 FLA_Error FLA_Tridiag_UT_l_step_ofz_var3( int m_A, 00142 int m_T, 00143 dcomplex* buff_A, int rs_A, int cs_A, 00144 dcomplex* buff_Z, int rs_Z, int cs_Z, 00145 dcomplex* buff_T, int rs_T, int cs_T ); 00146 00147 // --- Fused operations --- 00148 00149 FLA_Error FLA_Fused_Her2_Ax_l_opt_var1( FLA_Obj alpha, FLA_Obj u, FLA_Obj z, FLA_Obj A, FLA_Obj x, FLA_Obj w ); 00150 FLA_Error FLA_Fused_Her2_Ax_l_ops_var1( int m_A, 00151 float* buff_alpha, 00152 float* buff_u, int inc_u, 00153 float* buff_z, int inc_z, 00154 float* buff_A, int rs_A, int cs_A, 00155 float* buff_x, int inc_x, 00156 float* buff_w, int inc_w ); 00157 FLA_Error FLA_Fused_Her2_Ax_l_opd_var1( int m_A, 00158 double* buff_alpha, 00159 double* buff_u, int inc_u, 00160 double* buff_z, int inc_z, 00161 double* buff_A, int rs_A, int cs_A, 00162 double* buff_x, int inc_x, 00163 double* buff_w, int inc_w ); 00164 FLA_Error FLA_Fused_Her2_Ax_l_opc_var1( int m_A, 00165 scomplex* buff_alpha, 00166 scomplex* buff_u, int inc_u, 00167 scomplex* buff_z, int inc_z, 00168 scomplex* buff_A, int rs_A, int cs_A, 00169 scomplex* buff_x, int inc_x, 00170 scomplex* buff_w, int inc_w ); 00171 FLA_Error FLA_Fused_Her2_Ax_l_opz_var1( int m_A, 00172 dcomplex* buff_alpha, 00173 dcomplex* buff_u, int inc_u, 00174 dcomplex* buff_z, int inc_z, 00175 dcomplex* buff_A, int rs_A, int cs_A, 00176 dcomplex* buff_x, int inc_x, 00177 dcomplex* buff_w, int inc_w ); 00178 00179 FLA_Error FLA_Fused_UZhu_ZUhu_opt_var1( FLA_Obj delta, FLA_Obj U, FLA_Obj Z, FLA_Obj t, FLA_Obj u, FLA_Obj w ); 00180 FLA_Error FLA_Fused_UZhu_ZUhu_ops_var1( int m_U, 00181 int n_U, 00182 float* buff_delta, 00183 float* buff_U, int rs_U, int cs_U, 00184 float* buff_Z, int rs_Z, int cs_Z, 00185 float* buff_t, int inc_t, 00186 float* buff_u, int inc_u, 00187 float* buff_w, int inc_w ); 00188 FLA_Error FLA_Fused_UZhu_ZUhu_opd_var1( int m_U, 00189 int n_U, 00190 double* buff_delta, 00191 double* buff_U, int rs_U, int cs_U, 00192 double* buff_Z, int rs_Z, int cs_Z, 00193 double* buff_t, int inc_t, 00194 double* buff_u, int inc_u, 00195 double* buff_w, int inc_w ); 00196 FLA_Error FLA_Fused_UZhu_ZUhu_opc_var1( int m_U, 00197 int n_U, 00198 scomplex* buff_delta, 00199 scomplex* buff_U, int rs_U, int cs_U, 00200 scomplex* buff_Z, int rs_Z, int cs_Z, 00201 scomplex* buff_t, int inc_t, 00202 scomplex* buff_u, int inc_u, 00203 scomplex* buff_w, int inc_w ); 00204 FLA_Error FLA_Fused_UZhu_ZUhu_opz_var1( int m_U, 00205 int n_U, 00206 dcomplex* buff_delta, 00207 dcomplex* buff_U, int rs_U, int cs_U, 00208 dcomplex* buff_Z, int rs_Z, int cs_Z, 00209 dcomplex* buff_t, int inc_t, 00210 dcomplex* buff_u, int inc_u, 00211 dcomplex* buff_w, int inc_w );