libflame
revision_anchor
|
Functions | |
FLA_Error | FLA_Trinv_ln_opt_var4 (FLA_Obj A) |
FLA_Error | FLA_Trinv_ln_ops_var4 (int mn_A, float *buff_A, int rs_A, int cs_A) |
FLA_Error | FLA_Trinv_ln_opd_var4 (int mn_A, double *buff_A, int rs_A, int cs_A) |
FLA_Error | FLA_Trinv_ln_opc_var4 (int mn_A, scomplex *buff_A, int rs_A, int cs_A) |
FLA_Error | FLA_Trinv_ln_opz_var4 (int mn_A, dcomplex *buff_A, int rs_A, int cs_A) |
FLA_Error FLA_Trinv_ln_opc_var4 | ( | int | mn_A, |
scomplex * | buff_A, | ||
int | rs_A, | ||
int | cs_A | ||
) |
References bli_cger(), bli_cinverts(), bli_cscalv(), bli_ctrmv(), bli_ctrsv(), BLIS_LOWER_TRIANGULAR, BLIS_NO_CONJUGATE, BLIS_NO_TRANSPOSE, BLIS_NONUNIT_DIAG, BLIS_TRANSPOSE, and FLA_MINUS_ONE.
Referenced by FLA_Trinv_ln_opt_var4().
{ scomplex* buff_m1 = FLA_COMPLEX_PTR( FLA_MINUS_ONE ); int i; for ( i = 0; i < mn_A; ++i ) { scomplex* A00 = buff_A + (0 )*cs_A + (0 )*rs_A; scomplex* a10t = buff_A + (0 )*cs_A + (i )*rs_A; scomplex* A20 = buff_A + (0 )*cs_A + (i+1)*rs_A; scomplex* alpha11 = buff_A + (i )*cs_A + (i )*rs_A; scomplex* a21 = buff_A + (i )*cs_A + (i+1)*rs_A; scomplex* A22 = buff_A + (i+1)*cs_A + (i+1)*rs_A; int mn_ahead = mn_A - i - 1; int mn_behind = i; /*------------------------------------------------------------*/ // FLA_Scal_external( FLA_MINUS_ONE, a21 ); // FLA_Trsv_external( FLA_LOWER_TRIANGULAR, FLA_NO_TRANSPOSE, FLA_NONUNIT_DIAG, A22, a21 ); bli_cscalv( BLIS_NO_CONJUGATE, mn_ahead, buff_m1, a21, rs_A ); bli_ctrsv( BLIS_LOWER_TRIANGULAR, BLIS_NO_TRANSPOSE, BLIS_NONUNIT_DIAG, mn_ahead, A22, rs_A, cs_A, a21, rs_A ); // FLA_Ger_external( FLA_MINUS_ONE, a21, a10t, A20 ); bli_cger( BLIS_NO_CONJUGATE, BLIS_NO_CONJUGATE, mn_ahead, mn_behind, buff_m1, a21, rs_A, a10t, cs_A, A20, rs_A, cs_A ); // FLA_Trmv_external( FLA_LOWER_TRIANGULAR, FLA_TRANSPOSE, FLA_NONUNIT_DIAG, A00, a10t ); bli_ctrmv( BLIS_LOWER_TRIANGULAR, BLIS_TRANSPOSE, BLIS_NONUNIT_DIAG, mn_behind, A00, rs_A, cs_A, a10t, cs_A ); // FLA_Invert( FLA_NO_CONJUGATE, alpha11 ); bli_cinverts( BLIS_NO_CONJUGATE, alpha11 ); /*------------------------------------------------------------*/ } return FLA_SUCCESS; }
FLA_Error FLA_Trinv_ln_opd_var4 | ( | int | mn_A, |
double * | buff_A, | ||
int | rs_A, | ||
int | cs_A | ||
) |
References bli_dger(), bli_dinverts(), bli_dscalv(), bli_dtrmv(), bli_dtrsv(), BLIS_LOWER_TRIANGULAR, BLIS_NO_CONJUGATE, BLIS_NO_TRANSPOSE, BLIS_NONUNIT_DIAG, BLIS_TRANSPOSE, and FLA_MINUS_ONE.
Referenced by FLA_Trinv_ln_opt_var4().
{ double* buff_m1 = FLA_DOUBLE_PTR( FLA_MINUS_ONE ); int i; for ( i = 0; i < mn_A; ++i ) { double* A00 = buff_A + (0 )*cs_A + (0 )*rs_A; double* a10t = buff_A + (0 )*cs_A + (i )*rs_A; double* A20 = buff_A + (0 )*cs_A + (i+1)*rs_A; double* alpha11 = buff_A + (i )*cs_A + (i )*rs_A; double* a21 = buff_A + (i )*cs_A + (i+1)*rs_A; double* A22 = buff_A + (i+1)*cs_A + (i+1)*rs_A; int mn_ahead = mn_A - i - 1; int mn_behind = i; /*------------------------------------------------------------*/ // FLA_Scal_external( FLA_MINUS_ONE, a21 ); // FLA_Trsv_external( FLA_LOWER_TRIANGULAR, FLA_NO_TRANSPOSE, FLA_NONUNIT_DIAG, A22, a21 ); bli_dscalv( BLIS_NO_CONJUGATE, mn_ahead, buff_m1, a21, rs_A ); bli_dtrsv( BLIS_LOWER_TRIANGULAR, BLIS_NO_TRANSPOSE, BLIS_NONUNIT_DIAG, mn_ahead, A22, rs_A, cs_A, a21, rs_A ); // FLA_Ger_external( FLA_MINUS_ONE, a21, a10t, A20 ); bli_dger( BLIS_NO_CONJUGATE, BLIS_NO_CONJUGATE, mn_ahead, mn_behind, buff_m1, a21, rs_A, a10t, cs_A, A20, rs_A, cs_A ); // FLA_Trmv_external( FLA_LOWER_TRIANGULAR, FLA_TRANSPOSE, FLA_NONUNIT_DIAG, A00, a10t ); bli_dtrmv( BLIS_LOWER_TRIANGULAR, BLIS_TRANSPOSE, BLIS_NONUNIT_DIAG, mn_behind, A00, rs_A, cs_A, a10t, cs_A ); // FLA_Invert( FLA_NO_CONJUGATE, alpha11 ); bli_dinverts( BLIS_NO_CONJUGATE, alpha11 ); /*------------------------------------------------------------*/ } return FLA_SUCCESS; }
FLA_Error FLA_Trinv_ln_ops_var4 | ( | int | mn_A, |
float * | buff_A, | ||
int | rs_A, | ||
int | cs_A | ||
) |
References bli_sger(), bli_sinverts(), bli_sscalv(), bli_strmv(), bli_strsv(), BLIS_LOWER_TRIANGULAR, BLIS_NO_CONJUGATE, BLIS_NO_TRANSPOSE, BLIS_NONUNIT_DIAG, BLIS_TRANSPOSE, and FLA_MINUS_ONE.
Referenced by FLA_Trinv_ln_opt_var4().
{ float* buff_m1 = FLA_FLOAT_PTR( FLA_MINUS_ONE ); int i; for ( i = 0; i < mn_A; ++i ) { float* A00 = buff_A + (0 )*cs_A + (0 )*rs_A; float* a10t = buff_A + (0 )*cs_A + (i )*rs_A; float* A20 = buff_A + (0 )*cs_A + (i+1)*rs_A; float* alpha11 = buff_A + (i )*cs_A + (i )*rs_A; float* a21 = buff_A + (i )*cs_A + (i+1)*rs_A; float* A22 = buff_A + (i+1)*cs_A + (i+1)*rs_A; int mn_ahead = mn_A - i - 1; int mn_behind = i; /*------------------------------------------------------------*/ // FLA_Scal_external( FLA_MINUS_ONE, a21 ); // FLA_Trsv_external( FLA_LOWER_TRIANGULAR, FLA_NO_TRANSPOSE, FLA_NONUNIT_DIAG, A22, a21 ); bli_sscalv( BLIS_NO_CONJUGATE, mn_ahead, buff_m1, a21, rs_A ); bli_strsv( BLIS_LOWER_TRIANGULAR, BLIS_NO_TRANSPOSE, BLIS_NONUNIT_DIAG, mn_ahead, A22, rs_A, cs_A, a21, rs_A ); // FLA_Ger_external( FLA_MINUS_ONE, a21, a10t, A20 ); bli_sger( BLIS_NO_CONJUGATE, BLIS_NO_CONJUGATE, mn_ahead, mn_behind, buff_m1, a21, rs_A, a10t, cs_A, A20, rs_A, cs_A ); // FLA_Trmv_external( FLA_LOWER_TRIANGULAR, FLA_TRANSPOSE, FLA_NONUNIT_DIAG, A00, a10t ); bli_strmv( BLIS_LOWER_TRIANGULAR, BLIS_TRANSPOSE, BLIS_NONUNIT_DIAG, mn_behind, A00, rs_A, cs_A, a10t, cs_A ); // FLA_Invert( FLA_NO_CONJUGATE, alpha11 ); bli_sinverts( BLIS_NO_CONJUGATE, alpha11 ); /*------------------------------------------------------------*/ } return FLA_SUCCESS; }
References FLA_Obj_col_stride(), FLA_Obj_datatype(), FLA_Obj_length(), FLA_Obj_row_stride(), FLA_Trinv_ln_opc_var4(), FLA_Trinv_ln_opd_var4(), FLA_Trinv_ln_ops_var4(), and FLA_Trinv_ln_opz_var4().
Referenced by FLA_Trinv_ln().
{ FLA_Datatype datatype; int mn_A; int rs_A, cs_A; datatype = FLA_Obj_datatype( A ); mn_A = FLA_Obj_length( A ); rs_A = FLA_Obj_row_stride( A ); cs_A = FLA_Obj_col_stride( A ); switch ( datatype ) { case FLA_FLOAT: { float* buff_A = FLA_FLOAT_PTR( A ); FLA_Trinv_ln_ops_var4( mn_A, buff_A, rs_A, cs_A ); break; } case FLA_DOUBLE: { double* buff_A = FLA_DOUBLE_PTR( A ); FLA_Trinv_ln_opd_var4( mn_A, buff_A, rs_A, cs_A ); break; } case FLA_COMPLEX: { scomplex* buff_A = FLA_COMPLEX_PTR( A ); FLA_Trinv_ln_opc_var4( mn_A, buff_A, rs_A, cs_A ); break; } case FLA_DOUBLE_COMPLEX: { dcomplex* buff_A = FLA_DOUBLE_COMPLEX_PTR( A ); FLA_Trinv_ln_opz_var4( mn_A, buff_A, rs_A, cs_A ); break; } } return FLA_SUCCESS; }
FLA_Error FLA_Trinv_ln_opz_var4 | ( | int | mn_A, |
dcomplex * | buff_A, | ||
int | rs_A, | ||
int | cs_A | ||
) |
References bli_zger(), bli_zinverts(), bli_zscalv(), bli_ztrmv(), bli_ztrsv(), BLIS_LOWER_TRIANGULAR, BLIS_NO_CONJUGATE, BLIS_NO_TRANSPOSE, BLIS_NONUNIT_DIAG, BLIS_TRANSPOSE, and FLA_MINUS_ONE.
Referenced by FLA_Trinv_ln_opt_var4().
{ dcomplex* buff_m1 = FLA_DOUBLE_COMPLEX_PTR( FLA_MINUS_ONE ); int i; for ( i = 0; i < mn_A; ++i ) { dcomplex* A00 = buff_A + (0 )*cs_A + (0 )*rs_A; dcomplex* a10t = buff_A + (0 )*cs_A + (i )*rs_A; dcomplex* A20 = buff_A + (0 )*cs_A + (i+1)*rs_A; dcomplex* alpha11 = buff_A + (i )*cs_A + (i )*rs_A; dcomplex* a21 = buff_A + (i )*cs_A + (i+1)*rs_A; dcomplex* A22 = buff_A + (i+1)*cs_A + (i+1)*rs_A; int mn_ahead = mn_A - i - 1; int mn_behind = i; /*------------------------------------------------------------*/ // FLA_Scal_external( FLA_MINUS_ONE, a21 ); // FLA_Trsv_external( FLA_LOWER_TRIANGULAR, FLA_NO_TRANSPOSE, FLA_NONUNIT_DIAG, A22, a21 ); bli_zscalv( BLIS_NO_CONJUGATE, mn_ahead, buff_m1, a21, rs_A ); bli_ztrsv( BLIS_LOWER_TRIANGULAR, BLIS_NO_TRANSPOSE, BLIS_NONUNIT_DIAG, mn_ahead, A22, rs_A, cs_A, a21, rs_A ); // FLA_Ger_external( FLA_MINUS_ONE, a21, a10t, A20 ); bli_zger( BLIS_NO_CONJUGATE, BLIS_NO_CONJUGATE, mn_ahead, mn_behind, buff_m1, a21, rs_A, a10t, cs_A, A20, rs_A, cs_A ); // FLA_Trmv_external( FLA_LOWER_TRIANGULAR, FLA_TRANSPOSE, FLA_NONUNIT_DIAG, A00, a10t ); bli_ztrmv( BLIS_LOWER_TRIANGULAR, BLIS_TRANSPOSE, BLIS_NONUNIT_DIAG, mn_behind, A00, rs_A, cs_A, a10t, cs_A ); // FLA_Invert( FLA_NO_CONJUGATE, alpha11 ); bli_zinverts( BLIS_NO_CONJUGATE, alpha11 ); /*------------------------------------------------------------*/ } return FLA_SUCCESS; }