libflame
revision_anchor
|
Functions | |
FLA_Error | FLA_LU_nopiv_opt_var3 (FLA_Obj A) |
FLA_Error | FLA_LU_nopiv_ops_var3 (int m_A, int n_A, float *buff_A, int rs_A, int cs_A) |
FLA_Error | FLA_LU_nopiv_opd_var3 (int m_A, int n_A, double *buff_A, int rs_A, int cs_A) |
FLA_Error | FLA_LU_nopiv_opc_var3 (int m_A, int n_A, scomplex *buff_A, int rs_A, int cs_A) |
FLA_Error | FLA_LU_nopiv_opz_var3 (int m_A, int n_A, dcomplex *buff_A, int rs_A, int cs_A) |
FLA_Error FLA_LU_nopiv_opc_var3 | ( | int | m_A, |
int | n_A, | ||
scomplex * | buff_A, | ||
int | rs_A, | ||
int | cs_A | ||
) |
References bli_cdots(), bli_cgemv(), bli_cinvscalv(), bli_ctrsm(), bli_ctrsv(), BLIS_LEFT, BLIS_LOWER_TRIANGULAR, BLIS_NO_CONJUGATE, BLIS_NO_TRANSPOSE, BLIS_UNIT_DIAG, FLA_MINUS_ONE, and FLA_ONE.
Referenced by FLA_LU_nopiv_opt_var3().
{ scomplex* buff_1 = FLA_COMPLEX_PTR( FLA_ONE ); scomplex* buff_m1 = FLA_COMPLEX_PTR( FLA_MINUS_ONE ); int min_m_n = min( m_A, n_A ); int i; for ( i = 0; i < min_m_n; ++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* a01 = buff_A + (i )*cs_A + (0 )*rs_A; scomplex* alpha11 = buff_A + (i )*cs_A + (i )*rs_A; scomplex* a21 = buff_A + (i )*cs_A + (i+1)*rs_A; int m_ahead = m_A - i - 1; int mn_behind = i; /*------------------------------------------------------------*/ // FLA_Trsv_external( FLA_LOWER_TRIANGULAR, FLA_NO_TRANSPOSE, FLA_UNIT_DIAG, A00, a01 ); bli_ctrsv( BLIS_LOWER_TRIANGULAR, BLIS_NO_TRANSPOSE, BLIS_UNIT_DIAG, mn_behind, A00, rs_A, cs_A, a01, rs_A ); // FLA_Dots_external( FLA_MINUS_ONE, a10t, a01, FLA_ONE, alpha11 ); bli_cdots( BLIS_NO_CONJUGATE, mn_behind, buff_m1, a10t, cs_A, a01, rs_A, buff_1, alpha11 ); // FLA_Gemv_external( FLA_NO_TRANSPOSE, FLA_MINUS_ONE, A20, a01, FLA_ONE, a21 ); bli_cgemv( BLIS_NO_TRANSPOSE, BLIS_NO_CONJUGATE, m_ahead, mn_behind, buff_m1, A20, rs_A, cs_A, a01, rs_A, buff_1, a21, rs_A ); // FLA_Inv_scal_external( alpha11, a21 ); bli_cinvscalv( BLIS_NO_CONJUGATE, m_ahead, alpha11, a21, rs_A ); /*------------------------------------------------------------*/ } if ( m_A < n_A ) { scomplex* ATL = buff_A; scomplex* ATR = buff_A + m_A*cs_A; // FLA_Trsm_external( FLA_LEFT, FLA_LOWER_TRIANGULAR, // FLA_NO_TRANSPOSE, FLA_UNIT_DIAG, // FLA_ONE, ATL, ATR ); bli_ctrsm( BLIS_LEFT, BLIS_LOWER_TRIANGULAR, BLIS_NO_TRANSPOSE, BLIS_UNIT_DIAG, m_A, n_A - m_A, buff_1, ATL, rs_A, cs_A, ATR, rs_A, cs_A ); } return FLA_SUCCESS; }
FLA_Error FLA_LU_nopiv_opd_var3 | ( | int | m_A, |
int | n_A, | ||
double * | buff_A, | ||
int | rs_A, | ||
int | cs_A | ||
) |
References bli_ddots(), bli_dgemv(), bli_dinvscalv(), bli_dtrsm(), bli_dtrsv(), BLIS_LEFT, BLIS_LOWER_TRIANGULAR, BLIS_NO_CONJUGATE, BLIS_NO_TRANSPOSE, BLIS_UNIT_DIAG, FLA_MINUS_ONE, and FLA_ONE.
Referenced by FLA_LU_nopiv_opt_var3().
{ double* buff_1 = FLA_DOUBLE_PTR( FLA_ONE ); double* buff_m1 = FLA_DOUBLE_PTR( FLA_MINUS_ONE ); int min_m_n = min( m_A, n_A ); int i; for ( i = 0; i < min_m_n; ++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* a01 = buff_A + (i )*cs_A + (0 )*rs_A; double* alpha11 = buff_A + (i )*cs_A + (i )*rs_A; double* a21 = buff_A + (i )*cs_A + (i+1)*rs_A; int m_ahead = m_A - i - 1; int mn_behind = i; /*------------------------------------------------------------*/ // FLA_Trsv_external( FLA_LOWER_TRIANGULAR, FLA_NO_TRANSPOSE, FLA_UNIT_DIAG, A00, a01 ); bli_dtrsv( BLIS_LOWER_TRIANGULAR, BLIS_NO_TRANSPOSE, BLIS_UNIT_DIAG, mn_behind, A00, rs_A, cs_A, a01, rs_A ); // FLA_Dots_external( FLA_MINUS_ONE, a10t, a01, FLA_ONE, alpha11 ); bli_ddots( BLIS_NO_CONJUGATE, mn_behind, buff_m1, a10t, cs_A, a01, rs_A, buff_1, alpha11 ); // FLA_Gemv_external( FLA_NO_TRANSPOSE, FLA_MINUS_ONE, A20, a01, FLA_ONE, a21 ); bli_dgemv( BLIS_NO_TRANSPOSE, BLIS_NO_CONJUGATE, m_ahead, mn_behind, buff_m1, A20, rs_A, cs_A, a01, rs_A, buff_1, a21, rs_A ); // FLA_Inv_scal_external( alpha11, a21 ); bli_dinvscalv( BLIS_NO_CONJUGATE, m_ahead, alpha11, a21, rs_A ); /*------------------------------------------------------------*/ } if ( m_A < n_A ) { double* ATL = buff_A; double* ATR = buff_A + m_A*cs_A; // FLA_Trsm_external( FLA_LEFT, FLA_LOWER_TRIANGULAR, // FLA_NO_TRANSPOSE, FLA_UNIT_DIAG, // FLA_ONE, ATL, ATR ); bli_dtrsm( BLIS_LEFT, BLIS_LOWER_TRIANGULAR, BLIS_NO_TRANSPOSE, BLIS_UNIT_DIAG, m_A, n_A - m_A, buff_1, ATL, rs_A, cs_A, ATR, rs_A, cs_A ); } return FLA_SUCCESS; }
FLA_Error FLA_LU_nopiv_ops_var3 | ( | int | m_A, |
int | n_A, | ||
float * | buff_A, | ||
int | rs_A, | ||
int | cs_A | ||
) |
References bli_sdots(), bli_sgemv(), bli_sinvscalv(), bli_strsm(), bli_strsv(), BLIS_LEFT, BLIS_LOWER_TRIANGULAR, BLIS_NO_CONJUGATE, BLIS_NO_TRANSPOSE, BLIS_UNIT_DIAG, FLA_MINUS_ONE, and FLA_ONE.
Referenced by FLA_LU_nopiv_opt_var3().
{ float* buff_1 = FLA_FLOAT_PTR( FLA_ONE ); float* buff_m1 = FLA_FLOAT_PTR( FLA_MINUS_ONE ); int min_m_n = min( m_A, n_A ); int i; for ( i = 0; i < min_m_n; ++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* a01 = buff_A + (i )*cs_A + (0 )*rs_A; float* alpha11 = buff_A + (i )*cs_A + (i )*rs_A; float* a21 = buff_A + (i )*cs_A + (i+1)*rs_A; int m_ahead = m_A - i - 1; int mn_behind = i; /*------------------------------------------------------------*/ // FLA_Trsv_external( FLA_LOWER_TRIANGULAR, FLA_NO_TRANSPOSE, FLA_UNIT_DIAG, A00, a01 ); bli_strsv( BLIS_LOWER_TRIANGULAR, BLIS_NO_TRANSPOSE, BLIS_UNIT_DIAG, mn_behind, A00, rs_A, cs_A, a01, rs_A ); // FLA_Dots_external( FLA_MINUS_ONE, a10t, a01, FLA_ONE, alpha11 ); bli_sdots( BLIS_NO_CONJUGATE, mn_behind, buff_m1, a10t, cs_A, a01, rs_A, buff_1, alpha11 ); // FLA_Gemv_external( FLA_NO_TRANSPOSE, FLA_MINUS_ONE, A20, a01, FLA_ONE, a21 ); bli_sgemv( BLIS_NO_TRANSPOSE, BLIS_NO_CONJUGATE, m_ahead, mn_behind, buff_m1, A20, rs_A, cs_A, a01, rs_A, buff_1, a21, rs_A ); // FLA_Inv_scal_external( alpha11, a21 ); bli_sinvscalv( BLIS_NO_CONJUGATE, m_ahead, alpha11, a21, rs_A ); /*------------------------------------------------------------*/ } if ( m_A < n_A ) { float* ATL = buff_A; float* ATR = buff_A + m_A*cs_A; // FLA_Trsm_external( FLA_LEFT, FLA_LOWER_TRIANGULAR, // FLA_NO_TRANSPOSE, FLA_UNIT_DIAG, // FLA_ONE, ATL, ATR ); bli_strsm( BLIS_LEFT, BLIS_LOWER_TRIANGULAR, BLIS_NO_TRANSPOSE, BLIS_UNIT_DIAG, m_A, n_A - m_A, buff_1, ATL, rs_A, cs_A, ATR, rs_A, cs_A ); } return FLA_SUCCESS; }
References FLA_LU_nopiv_opc_var3(), FLA_LU_nopiv_opd_var3(), FLA_LU_nopiv_ops_var3(), FLA_LU_nopiv_opz_var3(), FLA_Obj_col_stride(), FLA_Obj_datatype(), FLA_Obj_length(), FLA_Obj_row_stride(), and FLA_Obj_width().
Referenced by FLA_LU_nopiv_internal().
{ FLA_Datatype datatype; int m_A, n_A; int rs_A, cs_A; datatype = FLA_Obj_datatype( A ); m_A = FLA_Obj_length( A ); n_A = FLA_Obj_width( 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_LU_nopiv_ops_var3( m_A, n_A, buff_A, rs_A, cs_A ); break; } case FLA_DOUBLE: { double* buff_A = FLA_DOUBLE_PTR( A ); FLA_LU_nopiv_opd_var3( m_A, n_A, buff_A, rs_A, cs_A ); break; } case FLA_COMPLEX: { scomplex* buff_A = FLA_COMPLEX_PTR( A ); FLA_LU_nopiv_opc_var3( m_A, n_A, buff_A, rs_A, cs_A ); break; } case FLA_DOUBLE_COMPLEX: { dcomplex* buff_A = FLA_DOUBLE_COMPLEX_PTR( A ); FLA_LU_nopiv_opz_var3( m_A, n_A, buff_A, rs_A, cs_A ); break; } } return FLA_SUCCESS; }
FLA_Error FLA_LU_nopiv_opz_var3 | ( | int | m_A, |
int | n_A, | ||
dcomplex * | buff_A, | ||
int | rs_A, | ||
int | cs_A | ||
) |
References bli_zdots(), bli_zgemv(), bli_zinvscalv(), bli_ztrsm(), bli_ztrsv(), BLIS_LEFT, BLIS_LOWER_TRIANGULAR, BLIS_NO_CONJUGATE, BLIS_NO_TRANSPOSE, BLIS_UNIT_DIAG, FLA_MINUS_ONE, and FLA_ONE.
Referenced by FLA_LU_nopiv_opt_var3().
{ dcomplex* buff_1 = FLA_DOUBLE_COMPLEX_PTR( FLA_ONE ); dcomplex* buff_m1 = FLA_DOUBLE_COMPLEX_PTR( FLA_MINUS_ONE ); int min_m_n = min( m_A, n_A ); int i; for ( i = 0; i < min_m_n; ++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* a01 = buff_A + (i )*cs_A + (0 )*rs_A; dcomplex* alpha11 = buff_A + (i )*cs_A + (i )*rs_A; dcomplex* a21 = buff_A + (i )*cs_A + (i+1)*rs_A; int m_ahead = m_A - i - 1; int mn_behind = i; /*------------------------------------------------------------*/ // FLA_Trsv_external( FLA_LOWER_TRIANGULAR, FLA_NO_TRANSPOSE, FLA_UNIT_DIAG, A00, a01 ); bli_ztrsv( BLIS_LOWER_TRIANGULAR, BLIS_NO_TRANSPOSE, BLIS_UNIT_DIAG, mn_behind, A00, rs_A, cs_A, a01, rs_A ); // FLA_Dots_external( FLA_MINUS_ONE, a10t, a01, FLA_ONE, alpha11 ); bli_zdots( BLIS_NO_CONJUGATE, mn_behind, buff_m1, a10t, cs_A, a01, rs_A, buff_1, alpha11 ); // FLA_Gemv_external( FLA_NO_TRANSPOSE, FLA_MINUS_ONE, A20, a01, FLA_ONE, a21 ); bli_zgemv( BLIS_NO_TRANSPOSE, BLIS_NO_CONJUGATE, m_ahead, mn_behind, buff_m1, A20, rs_A, cs_A, a01, rs_A, buff_1, a21, rs_A ); // FLA_Inv_scal_external( alpha11, a21 ); bli_zinvscalv( BLIS_NO_CONJUGATE, m_ahead, alpha11, a21, rs_A ); /*------------------------------------------------------------*/ } if ( m_A < n_A ) { dcomplex* ATL = buff_A; dcomplex* ATR = buff_A + m_A*cs_A; // FLA_Trsm_external( FLA_LEFT, FLA_LOWER_TRIANGULAR, // FLA_NO_TRANSPOSE, FLA_UNIT_DIAG, // FLA_ONE, ATL, ATR ); bli_ztrsm( BLIS_LEFT, BLIS_LOWER_TRIANGULAR, BLIS_NO_TRANSPOSE, BLIS_UNIT_DIAG, m_A, n_A - m_A, buff_1, ATL, rs_A, cs_A, ATR, rs_A, cs_A ); } return FLA_SUCCESS; }