libflame
revision_anchor
|
Go to the source code of this file.
Functions | |
FLA_Error | FLA_Apply_Q2_UT_lnfc_blk_var1 (FLA_Obj D, FLA_Obj T, FLA_Obj W1, FLA_Obj C, FLA_Obj E, fla_apq2ut_t *cntl) |
FLA_Error | FLA_Apply_Q2_UT_lnfc_blk_var2 (FLA_Obj D, FLA_Obj T, FLA_Obj W1, FLA_Obj C, FLA_Obj E, fla_apq2ut_t *cntl) |
FLA_Error | FLA_Apply_Q2_UT_lnfc_blk_var3 (FLA_Obj D, FLA_Obj T, FLA_Obj W, FLA_Obj C, FLA_Obj E, fla_apq2ut_t *cntl) |
FLA_Error FLA_Apply_Q2_UT_lnfc_blk_var1 | ( | FLA_Obj | D, |
FLA_Obj | T, | ||
FLA_Obj | W1, | ||
FLA_Obj | C, | ||
FLA_Obj | E, | ||
fla_apq2ut_t * | cntl | ||
) |
References FLA_Axpyt_internal(), FLA_Cont_with_1x3_to_1x2(), FLA_Cont_with_3x1_to_2x1(), FLA_Copyt_internal(), FLA_Gemm_internal(), FLA_MINUS_ONE, FLA_Obj_length(), FLA_Obj_width(), FLA_ONE, FLA_Part_1x2(), FLA_Part_2x1(), FLA_Part_2x2(), FLA_Repart_1x2_to_1x3(), FLA_Repart_2x1_to_3x1(), and FLA_Trsm_internal().
Referenced by FLA_Apply_Q2_UT_lnfc().
{ FLA_Obj DL, DR, D0, D1, D2; FLA_Obj TL, TR, T0, T1, T2; FLA_Obj T1T, T2B; FLA_Obj CT, C0, CB, C1, C2; FLA_Obj W1TL, W1TR, W1BL, W1BR; dim_t b_alg, b; // Query the algorithmic blocksize by inspecting the length of T. b_alg = FLA_Obj_length( T ); FLA_Part_1x2( D, &DL, &DR, 0, FLA_RIGHT ); FLA_Part_1x2( T, &TL, &TR, 0, FLA_RIGHT ); FLA_Part_2x1( C, &CT, &CB, 0, FLA_BOTTOM ); while ( FLA_Obj_width( DR ) < FLA_Obj_width( D ) ){ b = min( b_alg, FLA_Obj_width( DL ) ); // Since T was filled from left to right, and since we need to access them // in reverse order, we need to handle the case where the last block is // smaller than the other b x b blocks. if ( FLA_Obj_width( TR ) == 0 && FLA_Obj_width( T ) % b_alg > 0 ) b = FLA_Obj_width( T ) % b_alg; FLA_Repart_1x2_to_1x3( DL, /**/ DR, &D0, &D1, /**/ &D2, b, FLA_LEFT ); FLA_Repart_1x2_to_1x3( TL, /**/ TR, &T0, &T1, /**/ &T2, b, FLA_LEFT ); FLA_Repart_2x1_to_3x1( CT, &C0, &C1, /* ** */ /* ** */ CB, &C2, b, FLA_TOP ); /*------------------------------------------------------------*/ FLA_Part_2x1( T1, &T1T, &T2B, b, FLA_TOP ); FLA_Part_2x2( W1, &W1TL, &W1TR, &W1BL, &W1BR, b, FLA_Obj_width( C1 ), FLA_TL ); // W1TL = C1; FLA_Copyt_internal( FLA_NO_TRANSPOSE, C1, W1TL, FLA_Cntl_sub_copyt( cntl ) ); // W1TL = inv( triu( T1T ) ) * ( C1 + D1' * E ); FLA_Gemm_internal( FLA_CONJ_TRANSPOSE, FLA_NO_TRANSPOSE, FLA_ONE, D1, E, FLA_ONE, W1TL, FLA_Cntl_sub_gemm1( cntl ) ); FLA_Trsm_internal( FLA_LEFT, FLA_UPPER_TRIANGULAR, FLA_NO_TRANSPOSE, FLA_NONUNIT_DIAG, FLA_ONE, T1T, W1TL, FLA_Cntl_sub_trsm( cntl ) ); // C1 = C1 - W1TL; // E = E - D1 * W1TL; FLA_Axpyt_internal( FLA_NO_TRANSPOSE, FLA_MINUS_ONE, W1TL, C1, FLA_Cntl_sub_axpyt( cntl ) ); FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, FLA_MINUS_ONE, D1, W1TL, FLA_ONE, E, FLA_Cntl_sub_gemm2( cntl ) ); /*------------------------------------------------------------*/ FLA_Cont_with_1x3_to_1x2( &DL, /**/ &DR, D0, /**/ D1, D2, FLA_RIGHT ); FLA_Cont_with_1x3_to_1x2( &TL, /**/ &TR, T0, /**/ T1, T2, FLA_RIGHT ); FLA_Cont_with_3x1_to_2x1( &CT, C0, /* ** */ /* ** */ C1, &CB, C2, FLA_BOTTOM ); } return FLA_SUCCESS; }
FLA_Error FLA_Apply_Q2_UT_lnfc_blk_var2 | ( | FLA_Obj | D, |
FLA_Obj | T, | ||
FLA_Obj | W1, | ||
FLA_Obj | C, | ||
FLA_Obj | E, | ||
fla_apq2ut_t * | cntl | ||
) |
References FLA_Apply_Q2_UT_internal(), FLA_Cont_with_3x1_to_2x1(), FLA_Determine_blocksize(), FLA_Obj_length(), FLA_Part_2x1(), and FLA_Repart_2x1_to_3x1().
Referenced by FLA_Apply_Q2_UT_lnfc().
{ FLA_Obj DT, D0, DB, D1, D2; FLA_Obj TT, T0, TB, T1, T2; FLA_Obj ET, E0, EB, E1, E2; dim_t b; FLA_Part_2x1( D, &DT, &DB, 0, FLA_BOTTOM ); FLA_Part_2x1( T, &TT, &TB, 0, FLA_BOTTOM ); FLA_Part_2x1( E, &ET, &EB, 0, FLA_BOTTOM ); while ( FLA_Obj_length( DB ) < FLA_Obj_length( D ) ){ b = FLA_Determine_blocksize( DT, FLA_TOP, FLA_Cntl_blocksize( cntl ) ); FLA_Repart_2x1_to_3x1( DT, &D0, &D1, /* ** */ /* ** */ DB, &D2, b, FLA_TOP ); FLA_Repart_2x1_to_3x1( TT, &T0, &T1, /* ** */ /* ** */ TB, &T2, b, FLA_TOP ); FLA_Repart_2x1_to_3x1( ET, &E0, &E1, /* ** */ /* ** */ EB, &E2, b, FLA_TOP ); /*------------------------------------------------------------*/ // / C \ = Q / C \ // \ E1 / \ E1 / // // where Q is formed from D1 and T1. FLA_Apply_Q2_UT_internal( FLA_LEFT, FLA_NO_TRANSPOSE, FLA_FORWARD, FLA_COLUMNWISE, D1, T1, W1, C, E1, FLA_Cntl_sub_apq2ut( cntl ) ); /*------------------------------------------------------------*/ FLA_Cont_with_3x1_to_2x1( &DT, D0, /* ** */ /* ** */ D1, &DB, D2, FLA_BOTTOM ); FLA_Cont_with_3x1_to_2x1( &TT, T0, /* ** */ /* ** */ T1, &TB, T2, FLA_BOTTOM ); FLA_Cont_with_3x1_to_2x1( &ET, E0, /* ** */ /* ** */ E1, &EB, E2, FLA_BOTTOM ); } return FLA_SUCCESS; }
FLA_Error FLA_Apply_Q2_UT_lnfc_blk_var3 | ( | FLA_Obj | D, |
FLA_Obj | T, | ||
FLA_Obj | W, | ||
FLA_Obj | C, | ||
FLA_Obj | E, | ||
fla_apq2ut_t * | cntl | ||
) |
References FLA_Apply_Q2_UT_internal(), FLA_Cont_with_1x3_to_1x2(), FLA_Determine_blocksize(), FLA_Obj_width(), FLA_Part_1x2(), and FLA_Repart_1x2_to_1x3().
Referenced by FLA_Apply_Q2_UT_lnfc().
{ FLA_Obj WL, WR, W0, W1, W2; FLA_Obj CL, CR, C0, C1, C2; FLA_Obj EL, ER, E0, E1, E2; dim_t b; FLA_Part_1x2( W, &WL, &WR, 0, FLA_RIGHT ); FLA_Part_1x2( C, &CL, &CR, 0, FLA_RIGHT ); FLA_Part_1x2( E, &EL, &ER, 0, FLA_RIGHT ); while ( FLA_Obj_width( CR ) < FLA_Obj_width( C ) ){ b = FLA_Determine_blocksize( CL, FLA_LEFT, FLA_Cntl_blocksize( cntl ) ); FLA_Repart_1x2_to_1x3( WL, /**/ WR, &W0, &W1, /**/ &W2, b, FLA_LEFT ); FLA_Repart_1x2_to_1x3( CL, /**/ CR, &C0, &C1, /**/ &C2, b, FLA_LEFT ); FLA_Repart_1x2_to_1x3( EL, /**/ ER, &E0, &E1, /**/ &E2, b, FLA_LEFT ); /*------------------------------------------------------------*/ // / C1 \ = Q / C1 \ // \ E1 / \ E1 / // // where Q is formed from D and T. FLA_Apply_Q2_UT_internal( FLA_LEFT, FLA_NO_TRANSPOSE, FLA_FORWARD, FLA_COLUMNWISE, D, T, W1, C1, E1, FLA_Cntl_sub_apq2ut( cntl ) ); /*------------------------------------------------------------*/ FLA_Cont_with_1x3_to_1x2( &WL, /**/ &WR, W0, /**/ W1, W2, FLA_RIGHT ); FLA_Cont_with_1x3_to_1x2( &CL, /**/ &CR, C0, /**/ C1, C2, FLA_RIGHT ); FLA_Cont_with_1x3_to_1x2( &EL, /**/ &ER, E0, /**/ E1, E2, FLA_RIGHT ); } return FLA_SUCCESS; }