libflame
revision_anchor
|
Go to the source code of this file.
FLA_Error FLA_Sylv_nn_blk_var1 | ( | FLA_Obj | isgn, |
FLA_Obj | A, | ||
FLA_Obj | B, | ||
FLA_Obj | C, | ||
FLA_Obj | scale, | ||
fla_sylv_t * | cntl | ||
) |
References FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_MINUS_ONE, FLA_Obj_length(), FLA_ONE, FLA_Part_2x2(), FLA_Repart_2x2_to_3x3(), and FLA_Sylv_internal().
Referenced by FLA_Sylv_nn().
{ FLA_Obj ATL, ATR, A00, A01, A02, ABL, ABR, A10, A11, A12, A20, A21, A22; FLA_Obj BTL, BTR, B00, B01, B02, BBL, BBR, B10, B11, B12, B20, B21, B22; FLA_Obj CTL, CTR, C00, C01, C02, CBL, CBR, C10, C11, C12, C20, C21, C22; dim_t b; FLA_Part_2x2( A, &ATL, &ATR, &ABL, &ABR, 0, 0, FLA_BR ); FLA_Part_2x2( B, &BTL, &BTR, &BBL, &BBR, 0, 0, FLA_TL ); FLA_Part_2x2( C, &CTL, &CTR, &CBL, &CBR, 0, 0, FLA_BL ); while ( FLA_Obj_length( ABR ) < FLA_Obj_length( A ) ){ b = FLA_Determine_blocksize( CTR, FLA_TR, FLA_Cntl_blocksize( cntl ) ); FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, &A01, /**/ &A02, &A10, &A11, /**/ &A12, /* ************* */ /* ******************** */ ABL, /**/ ABR, &A20, &A21, /**/ &A22, b, b, FLA_TL ); FLA_Repart_2x2_to_3x3( BTL, /**/ BTR, &B00, /**/ &B01, &B02, /* ************* */ /* ******************** */ &B10, /**/ &B11, &B12, BBL, /**/ BBR, &B20, /**/ &B21, &B22, b, b, FLA_BR ); FLA_Repart_2x2_to_3x3( CTL, /**/ CTR, &C00, /**/ &C01, &C02, &C10, /**/ &C11, &C12, /* ************* */ /* ******************** */ CBL, /**/ CBR, &C20, /**/ &C21, &C22, b, b, FLA_TR ); // Loop Invariant: // CTL = CTL // CTR = CTR // CBL = sylv( ABR, BTL, CBL ) // CBR = CBR /*------------------------------------------------------------*/ // C10 = sylv( A11, B00, C10 - A12 * C20 ); FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, FLA_MINUS_ONE, A12, C20, FLA_ONE, C10, FLA_Cntl_sub_gemm1( cntl ) ); FLA_Sylv_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, isgn, A11, B00, C10, scale, FLA_Cntl_sub_sylv1( cntl ) ); // C21 = sylv( A22, B11, C21 -/+ C20 * B01 ); FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, FLA_NEGATE( isgn ), C20, B01, FLA_ONE, C21, FLA_Cntl_sub_gemm2( cntl ) ); FLA_Sylv_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, isgn, A22, B11, C21, scale, FLA_Cntl_sub_sylv2( cntl ) ); // C11 = sylv( A11, B11, C11 - A12 * C21 -/+ C10 * B01 ); FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, FLA_NEGATE( isgn ), C10, B01, FLA_ONE, C11, FLA_Cntl_sub_gemm3( cntl ) ); FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, FLA_MINUS_ONE, A12, C21, FLA_ONE, C11, FLA_Cntl_sub_gemm4( cntl ) ); FLA_Sylv_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, isgn, A11, B11, C11, scale, FLA_Cntl_sub_sylv3( cntl ) ); /*------------------------------------------------------------*/ FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, /**/ A01, A02, /* ************** */ /* ****************** */ A10, /**/ A11, A12, &ABL, /**/ &ABR, A20, /**/ A21, A22, FLA_BR ); FLA_Cont_with_3x3_to_2x2( &BTL, /**/ &BTR, B00, B01, /**/ B02, B10, B11, /**/ B12, /* ************** */ /* ****************** */ &BBL, /**/ &BBR, B20, B21, /**/ B22, FLA_TL ); FLA_Cont_with_3x3_to_2x2( &CTL, /**/ &CTR, C00, C01, /**/ C02, /* ************** */ /* ****************** */ C10, C11, /**/ C12, &CBL, /**/ &CBR, C20, C21, /**/ C22, FLA_BL ); } return FLA_SUCCESS; }
FLA_Error FLA_Sylv_nn_blk_var10 | ( | FLA_Obj | isgn, |
FLA_Obj | A, | ||
FLA_Obj | B, | ||
FLA_Obj | C, | ||
FLA_Obj | scale, | ||
fla_sylv_t * | cntl | ||
) |
References FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_MINUS_ONE, FLA_Obj_length(), FLA_ONE, FLA_Part_2x2(), FLA_Repart_2x2_to_3x3(), and FLA_Sylv_internal().
Referenced by FLA_Sylv_nn().
{ FLA_Obj ATL, ATR, A00, A01, A02, ABL, ABR, A10, A11, A12, A20, A21, A22; FLA_Obj BTL, BTR, B00, B01, B02, BBL, BBR, B10, B11, B12, B20, B21, B22; FLA_Obj CTL, CTR, C00, C01, C02, CBL, CBR, C10, C11, C12, C20, C21, C22; dim_t b; FLA_Part_2x2( A, &ATL, &ATR, &ABL, &ABR, 0, 0, FLA_BR ); FLA_Part_2x2( B, &BTL, &BTR, &BBL, &BBR, 0, 0, FLA_TL ); FLA_Part_2x2( C, &CTL, &CTR, &CBL, &CBR, 0, 0, FLA_BL ); while ( FLA_Obj_length( ABR ) < FLA_Obj_length( A ) ){ b = FLA_Determine_blocksize( CTR, FLA_TR, FLA_Cntl_blocksize( cntl ) ); FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, &A01, /**/ &A02, &A10, &A11, /**/ &A12, /* ************* */ /* ******************** */ ABL, /**/ ABR, &A20, &A21, /**/ &A22, b, b, FLA_TL ); FLA_Repart_2x2_to_3x3( BTL, /**/ BTR, &B00, /**/ &B01, &B02, /* ************* */ /* ******************** */ &B10, /**/ &B11, &B12, BBL, /**/ BBR, &B20, /**/ &B21, &B22, b, b, FLA_BR ); FLA_Repart_2x2_to_3x3( CTL, /**/ CTR, &C00, /**/ &C01, &C02, &C10, /**/ &C11, &C12, /* ************* */ /* ******************** */ CBL, /**/ CBR, &C20, /**/ &C21, &C22, b, b, FLA_TR ); // Loop Invariant: // CTL = sylv( ATL, BTL, CTL - ATR * sylv( ABR, BTL, CBL ) ) // CTR = CTR - sylv( ATL, BTL, CTL - ATR * sylv( ABR, BTL, CBL ) ) * BTR // CBL = sylv( ABR, BTL, CBL ) // CBR = CBR - sylv( ABR, BTL, CBL ) * BTR /*------------------------------------------------------------*/ // C21 = sylv( A22, B11, C21 ); FLA_Sylv_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, isgn, A22, B11, C21, scale, FLA_Cntl_sub_sylv1( cntl ) ); // C11 = sylv( A11, B11, C11 - A12 * C21 ); FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, FLA_MINUS_ONE, A12, C21, FLA_ONE, C11, FLA_Cntl_sub_gemm1( cntl ) ); FLA_Sylv_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, isgn, A11, B11, C11, scale, FLA_Cntl_sub_sylv2( cntl ) ); // C01 = sylv( A00, B11, C01 - A01 * C11 - A02 * C21 ); FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, FLA_MINUS_ONE, A02, C21, FLA_ONE, C01, FLA_Cntl_sub_gemm2( cntl ) ); FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, FLA_MINUS_ONE, A01, C11, FLA_ONE, C01, FLA_Cntl_sub_gemm3( cntl ) ); FLA_Sylv_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, isgn, A00, B11, C01, scale, FLA_Cntl_sub_sylv3( cntl ) ); // C02 = C02 -/+ C01 * B12; FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, FLA_NEGATE( isgn ), C01, B12, FLA_ONE, C02, FLA_Cntl_sub_gemm4( cntl ) ); // C12 = C12 -/+ C11 * B12; FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, FLA_NEGATE( isgn ), C11, B12, FLA_ONE, C12, FLA_Cntl_sub_gemm5( cntl ) ); // C22 = C22 -/+ C21 * B12; FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, FLA_NEGATE( isgn ), C21, B12, FLA_ONE, C22, FLA_Cntl_sub_gemm6( cntl ) ); /*------------------------------------------------------------*/ FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, /**/ A01, A02, /* ************** */ /* ****************** */ A10, /**/ A11, A12, &ABL, /**/ &ABR, A20, /**/ A21, A22, FLA_BR ); FLA_Cont_with_3x3_to_2x2( &BTL, /**/ &BTR, B00, B01, /**/ B02, B10, B11, /**/ B12, /* ************** */ /* ****************** */ &BBL, /**/ &BBR, B20, B21, /**/ B22, FLA_TL ); FLA_Cont_with_3x3_to_2x2( &CTL, /**/ &CTR, C00, C01, /**/ C02, /* ************** */ /* ****************** */ C10, C11, /**/ C12, &CBL, /**/ &CBR, C20, C21, /**/ C22, FLA_BL ); } return FLA_SUCCESS; }
FLA_Error FLA_Sylv_nn_blk_var11 | ( | FLA_Obj | isgn, |
FLA_Obj | A, | ||
FLA_Obj | B, | ||
FLA_Obj | C, | ||
FLA_Obj | scale, | ||
fla_sylv_t * | cntl | ||
) |
References FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_MINUS_ONE, FLA_Obj_length(), FLA_ONE, FLA_Part_2x2(), FLA_Repart_2x2_to_3x3(), and FLA_Sylv_internal().
Referenced by FLA_Sylv_nn().
{ FLA_Obj ATL, ATR, A00, A01, A02, ABL, ABR, A10, A11, A12, A20, A21, A22; FLA_Obj BTL, BTR, B00, B01, B02, BBL, BBR, B10, B11, B12, B20, B21, B22; FLA_Obj CTL, CTR, C00, C01, C02, CBL, CBR, C10, C11, C12, C20, C21, C22; dim_t b; FLA_Part_2x2( A, &ATL, &ATR, &ABL, &ABR, 0, 0, FLA_BR ); FLA_Part_2x2( B, &BTL, &BTR, &BBL, &BBR, 0, 0, FLA_TL ); FLA_Part_2x2( C, &CTL, &CTR, &CBL, &CBR, 0, 0, FLA_BL ); while ( FLA_Obj_length( ABR ) < FLA_Obj_length( A ) ){ b = FLA_Determine_blocksize( CTR, FLA_TR, FLA_Cntl_blocksize( cntl ) ); FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, &A01, /**/ &A02, &A10, &A11, /**/ &A12, /* ************* */ /* ******************** */ ABL, /**/ ABR, &A20, &A21, /**/ &A22, b, b, FLA_TL ); FLA_Repart_2x2_to_3x3( BTL, /**/ BTR, &B00, /**/ &B01, &B02, /* ************* */ /* ******************** */ &B10, /**/ &B11, &B12, BBL, /**/ BBR, &B20, /**/ &B21, &B22, b, b, FLA_BR ); FLA_Repart_2x2_to_3x3( CTL, /**/ CTR, &C00, /**/ &C01, &C02, &C10, /**/ &C11, &C12, /* ************* */ /* ******************** */ CBL, /**/ CBR, &C20, /**/ &C21, &C22, b, b, FLA_TR ); // Loop Invariant: // CTL = CTL - ATR * sylv( ABR, BTL, CBL ) // CTR = CTR - ATR * sylv( ABR, BBR, CBR - sylv( ABR, BTL, CBL ) * BTR ) // CBL = sylv( ABR, BTL, CBL ) // CBR = sylv( ABR, BBR, CBR - sylv( ABR, BTL, CBL ) * BTR ) /*------------------------------------------------------------*/ // C10 = sylv( A11, B00, C10 ); FLA_Sylv_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, isgn, A11, B00, C10, scale, FLA_Cntl_sub_sylv1( cntl ) ); // C00 = C00 - A01 * C10; FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, FLA_MINUS_ONE, A01, C10, FLA_ONE, C00, FLA_Cntl_sub_gemm1( cntl ) ); // C11 = sylv( A11, B11, C11 -/+ C10 * B01 ); FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, FLA_NEGATE( isgn ), C10, B01, FLA_ONE, C11, FLA_Cntl_sub_gemm2( cntl ) ); FLA_Sylv_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, isgn, A11, B11, C11, scale, FLA_Cntl_sub_sylv2( cntl ) ); // C01 = C01 - A01 * C11; FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, FLA_MINUS_ONE, A01, C11, FLA_ONE, C01, FLA_Cntl_sub_gemm3( cntl ) ); // C12 = sylv( A11, B22, C12 -/+ C10 * B02 -/+ C11 * B12 ); FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, FLA_NEGATE( isgn ), C11, B12, FLA_ONE, C12, FLA_Cntl_sub_gemm4( cntl ) ); FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, FLA_NEGATE( isgn ), C10, B02, FLA_ONE, C12, FLA_Cntl_sub_gemm5( cntl ) ); FLA_Sylv_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, isgn, A11, B22, C12, scale, FLA_Cntl_sub_sylv3( cntl ) ); // C02 = C02 - A01 * C12; FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, FLA_MINUS_ONE, A01, C12, FLA_ONE, C02, FLA_Cntl_sub_gemm6( cntl ) ); /*------------------------------------------------------------*/ FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, /**/ A01, A02, /* ************** */ /* ****************** */ A10, /**/ A11, A12, &ABL, /**/ &ABR, A20, /**/ A21, A22, FLA_BR ); FLA_Cont_with_3x3_to_2x2( &BTL, /**/ &BTR, B00, B01, /**/ B02, B10, B11, /**/ B12, /* ************** */ /* ****************** */ &BBL, /**/ &BBR, B20, B21, /**/ B22, FLA_TL ); FLA_Cont_with_3x3_to_2x2( &CTL, /**/ &CTR, C00, C01, /**/ C02, /* ************** */ /* ****************** */ C10, C11, /**/ C12, &CBL, /**/ &CBR, C20, C21, /**/ C22, FLA_BL ); } return FLA_SUCCESS; }
FLA_Error FLA_Sylv_nn_blk_var12 | ( | FLA_Obj | isgn, |
FLA_Obj | A, | ||
FLA_Obj | B, | ||
FLA_Obj | C, | ||
FLA_Obj | scale, | ||
fla_sylv_t * | cntl | ||
) |
References FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_MINUS_ONE, FLA_Obj_length(), FLA_ONE, FLA_Part_2x2(), FLA_Repart_2x2_to_3x3(), and FLA_Sylv_internal().
Referenced by FLA_Sylv_nn().
{ FLA_Obj ATL, ATR, A00, A01, A02, ABL, ABR, A10, A11, A12, A20, A21, A22; FLA_Obj BTL, BTR, B00, B01, B02, BBL, BBR, B10, B11, B12, B20, B21, B22; FLA_Obj CTL, CTR, C00, C01, C02, CBL, CBR, C10, C11, C12, C20, C21, C22; dim_t b; FLA_Part_2x2( A, &ATL, &ATR, &ABL, &ABR, 0, 0, FLA_BR ); FLA_Part_2x2( B, &BTL, &BTR, &BBL, &BBR, 0, 0, FLA_TL ); FLA_Part_2x2( C, &CTL, &CTR, &CBL, &CBR, 0, 0, FLA_BL ); while ( FLA_Obj_length( ABR ) < FLA_Obj_length( A ) ){ b = FLA_Determine_blocksize( CTR, FLA_TR, FLA_Cntl_blocksize( cntl ) ); FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, &A01, /**/ &A02, &A10, &A11, /**/ &A12, /* ************* */ /* ******************** */ ABL, /**/ ABR, &A20, &A21, /**/ &A22, b, b, FLA_TL ); FLA_Repart_2x2_to_3x3( BTL, /**/ BTR, &B00, /**/ &B01, &B02, /* ************* */ /* ******************** */ &B10, /**/ &B11, &B12, BBL, /**/ BBR, &B20, /**/ &B21, &B22, b, b, FLA_BR ); FLA_Repart_2x2_to_3x3( CTL, /**/ CTR, &C00, /**/ &C01, &C02, &C10, /**/ &C11, &C12, /* ************* */ /* ******************** */ CBL, /**/ CBR, &C20, /**/ &C21, &C22, b, b, FLA_TR ); // Loop Invariant: // CTL = sylv( ATL, BTL, CTL - ATR * sylv( ABR, BTL, CBL ) ) // CTR = CTR - sylv( ATL, BTL, CTL - ATR * sylv( ABR, BTL, CBL ) ) * BTR // CBL = sylv( ABR, BTL, CBL ) // CBR = sylv( ABR, BBR, CBR - sylv( ABR, BTL, CBL ) * BTR ) /*------------------------------------------------------------*/ // C11 = sylv( A11, B11, C11 - A12 * C21 ); FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, FLA_MINUS_ONE, A12, C21, FLA_ONE, C11, FLA_Cntl_sub_gemm1( cntl ) ); FLA_Sylv_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, isgn, A11, B11, C11, scale, FLA_Cntl_sub_sylv1( cntl ) ); // C01 = sylv( A00, B11, C01 - A01 * C11 - A02 * C21 ); FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, FLA_MINUS_ONE, A02, C21, FLA_ONE, C01, FLA_Cntl_sub_gemm2( cntl ) ); FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, FLA_MINUS_ONE, A01, C11, FLA_ONE, C01, FLA_Cntl_sub_gemm3( cntl ) ); FLA_Sylv_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, isgn, A00, B11, C01, scale, FLA_Cntl_sub_sylv2( cntl ) ); // C02 = C02 -/+ C01 * B12; FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, FLA_NEGATE( isgn ), C01, B12, FLA_ONE, C02, FLA_Cntl_sub_gemm4( cntl ) ); // C12 = sylv( A11, B22, C12 - A12 * C22 -/+ C11 * B12 ); FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, FLA_NEGATE( isgn ), C11, B12, FLA_ONE, C12, FLA_Cntl_sub_gemm5( cntl ) ); FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, FLA_MINUS_ONE, A12, C22, FLA_ONE, C12, FLA_Cntl_sub_gemm6( cntl ) ); FLA_Sylv_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, isgn, A11, B22, C12, scale, FLA_Cntl_sub_sylv3( cntl ) ); /*------------------------------------------------------------*/ FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, /**/ A01, A02, /* ************** */ /* ****************** */ A10, /**/ A11, A12, &ABL, /**/ &ABR, A20, /**/ A21, A22, FLA_BR ); FLA_Cont_with_3x3_to_2x2( &BTL, /**/ &BTR, B00, B01, /**/ B02, B10, B11, /**/ B12, /* ************** */ /* ****************** */ &BBL, /**/ &BBR, B20, B21, /**/ B22, FLA_TL ); FLA_Cont_with_3x3_to_2x2( &CTL, /**/ &CTR, C00, C01, /**/ C02, /* ************** */ /* ****************** */ C10, C11, /**/ C12, &CBL, /**/ &CBR, C20, C21, /**/ C22, FLA_BL ); } return FLA_SUCCESS; }
FLA_Error FLA_Sylv_nn_blk_var13 | ( | FLA_Obj | isgn, |
FLA_Obj | A, | ||
FLA_Obj | B, | ||
FLA_Obj | C, | ||
FLA_Obj | scale, | ||
fla_sylv_t * | cntl | ||
) |
References FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_MINUS_ONE, FLA_Obj_length(), FLA_ONE, FLA_Part_2x2(), FLA_Repart_2x2_to_3x3(), and FLA_Sylv_internal().
Referenced by FLA_Sylv_nn().
{ FLA_Obj ATL, ATR, A00, A01, A02, ABL, ABR, A10, A11, A12, A20, A21, A22; FLA_Obj BTL, BTR, B00, B01, B02, BBL, BBR, B10, B11, B12, B20, B21, B22; FLA_Obj CTL, CTR, C00, C01, C02, CBL, CBR, C10, C11, C12, C20, C21, C22; dim_t b; FLA_Part_2x2( A, &ATL, &ATR, &ABL, &ABR, 0, 0, FLA_BR ); FLA_Part_2x2( B, &BTL, &BTR, &BBL, &BBR, 0, 0, FLA_TL ); FLA_Part_2x2( C, &CTL, &CTR, &CBL, &CBR, 0, 0, FLA_BL ); while ( FLA_Obj_length( ABR ) < FLA_Obj_length( A ) ){ b = FLA_Determine_blocksize( CTR, FLA_TR, FLA_Cntl_blocksize( cntl ) ); FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, &A01, /**/ &A02, &A10, &A11, /**/ &A12, /* ************* */ /* ******************** */ ABL, /**/ ABR, &A20, &A21, /**/ &A22, b, b, FLA_TL ); FLA_Repart_2x2_to_3x3( BTL, /**/ BTR, &B00, /**/ &B01, &B02, /* ************* */ /* ******************** */ &B10, /**/ &B11, &B12, BBL, /**/ BBR, &B20, /**/ &B21, &B22, b, b, FLA_BR ); FLA_Repart_2x2_to_3x3( CTL, /**/ CTR, &C00, /**/ &C01, &C02, &C10, /**/ &C11, &C12, /* ************* */ /* ******************** */ CBL, /**/ CBR, &C20, /**/ &C21, &C22, b, b, FLA_TR ); // Loop Invariant: // CTL = sylv( ATL, BTL, CTL - ATR * sylv( ABR, BTL, CBL ) ) // CTR = CTR - ATR * sylv( ABR, BBR, CBR - sylv( ABR, BTL, CBL ) * BTR ) // CBL = sylv( ABR, BTL, CBL ) // CBR = sylv( ABR, BBR, CBR - sylv( ABR, BTL, CBL ) * BTR ) /*------------------------------------------------------------*/ // C11 = sylv( A11, B11, C11 -/+ C10 * B01 ); FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, FLA_NEGATE( isgn ), C10, B01, FLA_ONE, C11, FLA_Cntl_sub_gemm1( cntl ) ); FLA_Sylv_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, isgn, A11, B11, C11, scale, FLA_Cntl_sub_sylv1( cntl ) ); // C01 = sylv( A00, B11, C01 - A01 * C11 -/+ C00 * B01 ); FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, FLA_NEGATE( isgn ), C00, B01, FLA_ONE, C01, FLA_Cntl_sub_gemm2( cntl ) ); FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, FLA_MINUS_ONE, A01, C11, FLA_ONE, C01, FLA_Cntl_sub_gemm3( cntl ) ); FLA_Sylv_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, isgn, A00, B11, C01, scale, FLA_Cntl_sub_sylv2( cntl ) ); // C12 = sylv( A11, B22, C12 -/+ C10 * B02 -/+ C11 * B12 ); FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, FLA_NEGATE( isgn ), C11, B12, FLA_ONE, C12, FLA_Cntl_sub_gemm4( cntl ) ); FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, FLA_NEGATE( isgn ), C10, B02, FLA_ONE, C12, FLA_Cntl_sub_gemm5( cntl ) ); FLA_Sylv_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, isgn, A11, B22, C12, scale, FLA_Cntl_sub_sylv3( cntl ) ); // C02 = C02 - A01 * C12; FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, FLA_MINUS_ONE, A01, C12, FLA_ONE, C02, FLA_Cntl_sub_gemm6( cntl ) ); /*------------------------------------------------------------*/ FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, /**/ A01, A02, /* ************** */ /* ****************** */ A10, /**/ A11, A12, &ABL, /**/ &ABR, A20, /**/ A21, A22, FLA_BR ); FLA_Cont_with_3x3_to_2x2( &BTL, /**/ &BTR, B00, B01, /**/ B02, B10, B11, /**/ B12, /* ************** */ /* ****************** */ &BBL, /**/ &BBR, B20, B21, /**/ B22, FLA_TL ); FLA_Cont_with_3x3_to_2x2( &CTL, /**/ &CTR, C00, C01, /**/ C02, /* ************** */ /* ****************** */ C10, C11, /**/ C12, &CBL, /**/ &CBR, C20, C21, /**/ C22, FLA_BL ); } return FLA_SUCCESS; }
FLA_Error FLA_Sylv_nn_blk_var14 | ( | FLA_Obj | isgn, |
FLA_Obj | A, | ||
FLA_Obj | B, | ||
FLA_Obj | C, | ||
FLA_Obj | scale, | ||
fla_sylv_t * | cntl | ||
) |
References FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_MINUS_ONE, FLA_Obj_length(), FLA_ONE, FLA_Part_2x2(), FLA_Repart_2x2_to_3x3(), and FLA_Sylv_internal().
Referenced by FLA_Sylv_nn().
{ FLA_Obj ATL, ATR, A00, A01, A02, ABL, ABR, A10, A11, A12, A20, A21, A22; FLA_Obj BTL, BTR, B00, B01, B02, BBL, BBR, B10, B11, B12, B20, B21, B22; FLA_Obj CTL, CTR, C00, C01, C02, CBL, CBR, C10, C11, C12, C20, C21, C22; dim_t b; FLA_Part_2x2( A, &ATL, &ATR, &ABL, &ABR, 0, 0, FLA_BR ); FLA_Part_2x2( B, &BTL, &BTR, &BBL, &BBR, 0, 0, FLA_TL ); FLA_Part_2x2( C, &CTL, &CTR, &CBL, &CBR, 0, 0, FLA_BL ); while ( FLA_Obj_length( ABR ) < FLA_Obj_length( A ) ){ b = FLA_Determine_blocksize( CTR, FLA_TR, FLA_Cntl_blocksize( cntl ) ); FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, &A01, /**/ &A02, &A10, &A11, /**/ &A12, /* ************* */ /* ******************** */ ABL, /**/ ABR, &A20, &A21, /**/ &A22, b, b, FLA_TL ); FLA_Repart_2x2_to_3x3( BTL, /**/ BTR, &B00, /**/ &B01, &B02, /* ************* */ /* ******************** */ &B10, /**/ &B11, &B12, BBL, /**/ BBR, &B20, /**/ &B21, &B22, b, b, FLA_BR ); FLA_Repart_2x2_to_3x3( CTL, /**/ CTR, &C00, /**/ &C01, &C02, &C10, /**/ &C11, &C12, /* ************* */ /* ******************** */ CBL, /**/ CBR, &C20, /**/ &C21, &C22, b, b, FLA_TR ); // Loop Invariant: // CTL = sylv( ATL, BTL, CTL - ATR * sylv( ABR, BTL, CBL ) ) // CTR = CTR - ATR * sylv( ABR, BBR, CBR - sylv( ABR, BTL, CBL ) * BTR ) // - sylv( ATL, BTL, CTL - ATR * sylv( ABR, BTL, CBL ) ) * BTR // CBL = sylv( ABR, BTL, CBL ) // CBR = sylv( ABR, BBR, CBR - sylv( ABR, BTL, CBL ) * BTR ) /*------------------------------------------------------------*/ // C11 = sylv( A11, B11, C11 ); FLA_Sylv_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, isgn, A11, B11, C11, scale, FLA_Cntl_sub_sylv1( cntl ) ); // C01 = sylv( A00, B11, C01 - A01 * C11 ); FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, FLA_MINUS_ONE, A01, C11, FLA_ONE, C01, FLA_Cntl_sub_gemm1( cntl ) ); FLA_Sylv_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, isgn, A00, B11, C01, scale, FLA_Cntl_sub_sylv2( cntl ) ); // C12 = sylv( A11, B22, C12 -/+ C11 * B12 ); FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, FLA_NEGATE( isgn ), C11, B12, FLA_ONE, C12, FLA_Cntl_sub_gemm2( cntl ) ); FLA_Sylv_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, isgn, A11, B22, C12, scale, FLA_Cntl_sub_sylv3( cntl ) ); // C02 = C02 - A01 * C12 -/+ C01 * B12; FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, FLA_NEGATE( isgn ), C01, B12, FLA_ONE, C02, FLA_Cntl_sub_gemm3( cntl ) ); FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, FLA_MINUS_ONE, A01, C12, FLA_ONE, C02, FLA_Cntl_sub_gemm4( cntl ) ); /*------------------------------------------------------------*/ FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, /**/ A01, A02, /* ************** */ /* ****************** */ A10, /**/ A11, A12, &ABL, /**/ &ABR, A20, /**/ A21, A22, FLA_BR ); FLA_Cont_with_3x3_to_2x2( &BTL, /**/ &BTR, B00, B01, /**/ B02, B10, B11, /**/ B12, /* ************** */ /* ****************** */ &BBL, /**/ &BBR, B20, B21, /**/ B22, FLA_TL ); FLA_Cont_with_3x3_to_2x2( &CTL, /**/ &CTR, C00, C01, /**/ C02, /* ************** */ /* ****************** */ C10, C11, /**/ C12, &CBL, /**/ &CBR, C20, C21, /**/ C22, FLA_BL ); } return FLA_SUCCESS; }
FLA_Error FLA_Sylv_nn_blk_var15 | ( | FLA_Obj | isgn, |
FLA_Obj | A, | ||
FLA_Obj | B, | ||
FLA_Obj | C, | ||
FLA_Obj | scale, | ||
fla_sylv_t * | cntl | ||
) |
References FLA_Cont_with_3x1_to_2x1(), FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_MINUS_ONE, FLA_Obj_length(), FLA_ONE, FLA_Part_2x1(), FLA_Part_2x2(), FLA_Repart_2x1_to_3x1(), FLA_Repart_2x2_to_3x3(), and FLA_Sylv_internal().
Referenced by FLA_Sylv_nn().
{ FLA_Obj ATL, ATR, A00, A01, A02, ABL, ABR, A10, A11, A12, A20, A21, A22; FLA_Obj CT, C0, CB, C1, C2; dim_t b; FLA_Part_2x2( A, &ATL, &ATR, &ABL, &ABR, 0, 0, FLA_BR ); FLA_Part_2x1( C, &CT, &CB, 0, FLA_BOTTOM ); while ( FLA_Obj_length( ABR ) < FLA_Obj_length( A ) ){ b = FLA_Determine_blocksize( CT, FLA_TOP, FLA_Cntl_blocksize( cntl ) ); FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, &A01, /**/ &A02, &A10, &A11, /**/ &A12, /* ************* */ /* ******************** */ ABL, /**/ ABR, &A20, &A21, /**/ &A22, b, b, FLA_TL ); FLA_Repart_2x1_to_3x1( CT, &C0, &C1, /* ** */ /* ** */ CB, &C2, b, FLA_TOP ); // Loop Invariant: // CT = CT // CB = sylv( ABR, B, CB ) /*------------------------------------------------------------*/ // C1 = sylv( A11, B, C1 - A12 * C2 ); FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, FLA_MINUS_ONE, A12, C2, FLA_ONE, C1, FLA_Cntl_sub_gemm1( cntl ) ); FLA_Sylv_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, isgn, A11, B, C1, scale, FLA_Cntl_sub_sylv1( cntl ) ); /*------------------------------------------------------------*/ FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, /**/ A01, A02, /* ************** */ /* ****************** */ A10, /**/ A11, A12, &ABL, /**/ &ABR, A20, /**/ A21, A22, FLA_BR ); FLA_Cont_with_3x1_to_2x1( &CT, C0, /* ** */ /* ** */ C1, &CB, C2, FLA_BOTTOM ); } return FLA_SUCCESS; }
FLA_Error FLA_Sylv_nn_blk_var16 | ( | FLA_Obj | isgn, |
FLA_Obj | A, | ||
FLA_Obj | B, | ||
FLA_Obj | C, | ||
FLA_Obj | scale, | ||
fla_sylv_t * | cntl | ||
) |
References FLA_Cont_with_3x1_to_2x1(), FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_MINUS_ONE, FLA_Obj_length(), FLA_ONE, FLA_Part_2x1(), FLA_Part_2x2(), FLA_Repart_2x1_to_3x1(), FLA_Repart_2x2_to_3x3(), and FLA_Sylv_internal().
Referenced by FLA_Sylv_nn().
{ FLA_Obj ATL, ATR, A00, A01, A02, ABL, ABR, A10, A11, A12, A20, A21, A22; FLA_Obj CT, C0, CB, C1, C2; dim_t b; FLA_Part_2x2( A, &ATL, &ATR, &ABL, &ABR, 0, 0, FLA_BR ); FLA_Part_2x1( C, &CT, &CB, 0, FLA_BOTTOM ); while ( FLA_Obj_length( ABR ) < FLA_Obj_length( A ) ){ b = FLA_Determine_blocksize( CT, FLA_TOP, FLA_Cntl_blocksize( cntl ) ); FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, &A01, /**/ &A02, &A10, &A11, /**/ &A12, /* ************* */ /* ******************** */ ABL, /**/ ABR, &A20, &A21, /**/ &A22, b, b, FLA_TL ); FLA_Repart_2x1_to_3x1( CT, &C0, &C1, /* ** */ /* ** */ CB, &C2, b, FLA_TOP ); // Loop Invariant: // CT = CT - ATR * sylv( ABR, B, CB ) // CB = sylv( ABR, B, CB ) /*------------------------------------------------------------*/ // C1 = sylv( A11, B, C1 ); FLA_Sylv_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, isgn, A11, B, C1, scale, FLA_Cntl_sub_sylv1( cntl ) ); // C0 = C0 - A01 * C1; FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, FLA_MINUS_ONE, A01, C1, FLA_ONE, C0, FLA_Cntl_sub_gemm1( cntl ) ); /*------------------------------------------------------------*/ FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, /**/ A01, A02, /* ************** */ /* ****************** */ A10, /**/ A11, A12, &ABL, /**/ &ABR, A20, /**/ A21, A22, FLA_BR ); FLA_Cont_with_3x1_to_2x1( &CT, C0, /* ** */ /* ** */ C1, &CB, C2, FLA_BOTTOM ); } return FLA_SUCCESS; }
FLA_Error FLA_Sylv_nn_blk_var17 | ( | FLA_Obj | isgn, |
FLA_Obj | A, | ||
FLA_Obj | B, | ||
FLA_Obj | C, | ||
FLA_Obj | scale, | ||
fla_sylv_t * | cntl | ||
) |
References FLA_Cont_with_1x3_to_1x2(), FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_Obj_length(), FLA_ONE, FLA_Part_1x2(), FLA_Part_2x2(), FLA_Repart_1x2_to_1x3(), FLA_Repart_2x2_to_3x3(), and FLA_Sylv_internal().
Referenced by FLA_Sylv_nn().
{ FLA_Obj BTL, BTR, B00, B01, B02, BBL, BBR, B10, B11, B12, B20, B21, B22; FLA_Obj CL, CR, C0, C1, C2; dim_t b; FLA_Part_2x2( B, &BTL, &BTR, &BBL, &BBR, 0, 0, FLA_TL ); FLA_Part_1x2( C, &CL, &CR, 0, FLA_LEFT ); while ( FLA_Obj_length( BTL ) < FLA_Obj_length( B ) ){ b = FLA_Determine_blocksize( CR, FLA_RIGHT, FLA_Cntl_blocksize( cntl ) ); FLA_Repart_2x2_to_3x3( BTL, /**/ BTR, &B00, /**/ &B01, &B02, /* ************* */ /* ******************** */ &B10, /**/ &B11, &B12, BBL, /**/ BBR, &B20, /**/ &B21, &B22, b, b, FLA_BR ); FLA_Repart_1x2_to_1x3( CL, /**/ CR, &C0, /**/ &C1, &C2, b, FLA_RIGHT ); // Loop Invariant: // CL = // CR = /*------------------------------------------------------------*/ // C1 = sylv( A, B11, C1 -/+ C0 * B01 ); FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, FLA_NEGATE( isgn ), C0, B01, FLA_ONE, C1, FLA_Cntl_sub_gemm1( cntl ) ); FLA_Sylv_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, isgn, A, B11, C1, scale, FLA_Cntl_sub_sylv1( cntl ) ); /*------------------------------------------------------------*/ FLA_Cont_with_3x3_to_2x2( &BTL, /**/ &BTR, B00, B01, /**/ B02, B10, B11, /**/ B12, /* ************** */ /* ****************** */ &BBL, /**/ &BBR, B20, B21, /**/ B22, FLA_TL ); FLA_Cont_with_1x3_to_1x2( &CL, /**/ &CR, C0, C1, /**/ C2, FLA_LEFT ); } return FLA_SUCCESS; }
FLA_Error FLA_Sylv_nn_blk_var18 | ( | FLA_Obj | isgn, |
FLA_Obj | A, | ||
FLA_Obj | B, | ||
FLA_Obj | C, | ||
FLA_Obj | scale, | ||
fla_sylv_t * | cntl | ||
) |
References FLA_Cont_with_1x3_to_1x2(), FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_Obj_length(), FLA_ONE, FLA_Part_1x2(), FLA_Part_2x2(), FLA_Repart_1x2_to_1x3(), FLA_Repart_2x2_to_3x3(), and FLA_Sylv_internal().
Referenced by FLA_Sylv_nn().
{ FLA_Obj BTL, BTR, B00, B01, B02, BBL, BBR, B10, B11, B12, B20, B21, B22; FLA_Obj CL, CR, C0, C1, C2; dim_t b; FLA_Part_2x2( B, &BTL, &BTR, &BBL, &BBR, 0, 0, FLA_TL ); FLA_Part_1x2( C, &CL, &CR, 0, FLA_LEFT ); while ( FLA_Obj_length( BTL ) < FLA_Obj_length( B ) ){ b = FLA_Determine_blocksize( CR, FLA_RIGHT, FLA_Cntl_blocksize( cntl ) ); FLA_Repart_2x2_to_3x3( BTL, /**/ BTR, &B00, /**/ &B01, &B02, /* ************* */ /* ******************** */ &B10, /**/ &B11, &B12, BBL, /**/ BBR, &B20, /**/ &B21, &B22, b, b, FLA_BR ); FLA_Repart_1x2_to_1x3( CL, /**/ CR, &C0, /**/ &C1, &C2, b, FLA_RIGHT ); // Loop Invariant: // CL = // CR = /*------------------------------------------------------------*/ // C1 = sylv( A, B11, C1 ); FLA_Sylv_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, isgn, A, B11, C1, scale, FLA_Cntl_sub_sylv1( cntl ) ); // C2 = C2 -/+ C1 * B12; FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, FLA_NEGATE( isgn ), C1, B12, FLA_ONE, C2, FLA_Cntl_sub_gemm1( cntl ) ); /*------------------------------------------------------------*/ FLA_Cont_with_3x3_to_2x2( &BTL, /**/ &BTR, B00, B01, /**/ B02, B10, B11, /**/ B12, /* ************** */ /* ****************** */ &BBL, /**/ &BBR, B20, B21, /**/ B22, FLA_TL ); FLA_Cont_with_1x3_to_1x2( &CL, /**/ &CR, C0, C1, /**/ C2, FLA_LEFT ); } return FLA_SUCCESS; }
FLA_Error FLA_Sylv_nn_blk_var2 | ( | FLA_Obj | isgn, |
FLA_Obj | A, | ||
FLA_Obj | B, | ||
FLA_Obj | C, | ||
FLA_Obj | scale, | ||
fla_sylv_t * | cntl | ||
) |
References FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_MINUS_ONE, FLA_Obj_length(), FLA_ONE, FLA_Part_2x2(), FLA_Repart_2x2_to_3x3(), and FLA_Sylv_internal().
Referenced by FLA_Sylv_nn().
{ FLA_Obj ATL, ATR, A00, A01, A02, ABL, ABR, A10, A11, A12, A20, A21, A22; FLA_Obj BTL, BTR, B00, B01, B02, BBL, BBR, B10, B11, B12, B20, B21, B22; FLA_Obj CTL, CTR, C00, C01, C02, CBL, CBR, C10, C11, C12, C20, C21, C22; dim_t b; FLA_Part_2x2( A, &ATL, &ATR, &ABL, &ABR, 0, 0, FLA_BR ); FLA_Part_2x2( B, &BTL, &BTR, &BBL, &BBR, 0, 0, FLA_TL ); FLA_Part_2x2( C, &CTL, &CTR, &CBL, &CBR, 0, 0, FLA_BL ); while ( FLA_Obj_length( ABR ) < FLA_Obj_length( A ) ){ b = FLA_Determine_blocksize( CTR, FLA_TR, FLA_Cntl_blocksize( cntl ) ); FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, &A01, /**/ &A02, &A10, &A11, /**/ &A12, /* ************* */ /* ******************** */ ABL, /**/ ABR, &A20, &A21, /**/ &A22, b, b, FLA_TL ); FLA_Repart_2x2_to_3x3( BTL, /**/ BTR, &B00, /**/ &B01, &B02, /* ************* */ /* ******************** */ &B10, /**/ &B11, &B12, BBL, /**/ BBR, &B20, /**/ &B21, &B22, b, b, FLA_BR ); FLA_Repart_2x2_to_3x3( CTL, /**/ CTR, &C00, /**/ &C01, &C02, &C10, /**/ &C11, &C12, /* ************* */ /* ******************** */ CBL, /**/ CBR, &C20, /**/ &C21, &C22, b, b, FLA_TR ); // Loop Invariant: // CTL = CTL - ATR * sylv( ABR, BTL, CBL) // CTR = CTR // CBL = sylv( ABR, BTL, CBL ) // CBR = CBR /*------------------------------------------------------------*/ // C10 = sylv( A11, B00, C10 ); FLA_Sylv_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, isgn, A11, B00, C10, scale, FLA_Cntl_sub_sylv1( cntl ) ); // C00 = C00 - A01 * C10; FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, FLA_MINUS_ONE, A01, C10, FLA_ONE, C00, FLA_Cntl_sub_gemm1( cntl ) ); // C21 = sylv( A22, B11, C21 -/+ C20 * B01 ); FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, FLA_NEGATE( isgn ), C20, B01, FLA_ONE, C21, FLA_Cntl_sub_gemm2( cntl ) ); FLA_Sylv_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, isgn, A22, B11, C21, scale, FLA_Cntl_sub_sylv2( cntl ) ); // C11 = sylv( A11, B11, C11 - A12 * C21 -/+ C10 * B01 ); FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, FLA_NEGATE( isgn ), C10, B01, FLA_ONE, C11, FLA_Cntl_sub_gemm3( cntl ) ); FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, FLA_MINUS_ONE, A12, C21, FLA_ONE, C11, FLA_Cntl_sub_gemm4( cntl ) ); FLA_Sylv_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, isgn, A11, B11, C11, scale, FLA_Cntl_sub_sylv3( cntl ) ); // C01 = C01 - A01 * C11 - A02 * C21; FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, FLA_MINUS_ONE, A02, C21, FLA_ONE, C01, FLA_Cntl_sub_gemm5( cntl ) ); FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, FLA_MINUS_ONE, A01, C11, FLA_ONE, C01, FLA_Cntl_sub_gemm6( cntl ) ); /*------------------------------------------------------------*/ FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, /**/ A01, A02, /* ************** */ /* ****************** */ A10, /**/ A11, A12, &ABL, /**/ &ABR, A20, /**/ A21, A22, FLA_BR ); FLA_Cont_with_3x3_to_2x2( &BTL, /**/ &BTR, B00, B01, /**/ B02, B10, B11, /**/ B12, /* ************** */ /* ****************** */ &BBL, /**/ &BBR, B20, B21, /**/ B22, FLA_TL ); FLA_Cont_with_3x3_to_2x2( &CTL, /**/ &CTR, C00, C01, /**/ C02, /* ************** */ /* ****************** */ C10, C11, /**/ C12, &CBL, /**/ &CBR, C20, C21, /**/ C22, FLA_BL ); } return FLA_SUCCESS; }
FLA_Error FLA_Sylv_nn_blk_var3 | ( | FLA_Obj | isgn, |
FLA_Obj | A, | ||
FLA_Obj | B, | ||
FLA_Obj | C, | ||
FLA_Obj | scale, | ||
fla_sylv_t * | cntl | ||
) |
References FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_MINUS_ONE, FLA_Obj_length(), FLA_ONE, FLA_Part_2x2(), FLA_Repart_2x2_to_3x3(), and FLA_Sylv_internal().
Referenced by FLA_Sylv_nn().
{ FLA_Obj ATL, ATR, A00, A01, A02, ABL, ABR, A10, A11, A12, A20, A21, A22; FLA_Obj BTL, BTR, B00, B01, B02, BBL, BBR, B10, B11, B12, B20, B21, B22; FLA_Obj CTL, CTR, C00, C01, C02, CBL, CBR, C10, C11, C12, C20, C21, C22; dim_t b; FLA_Part_2x2( A, &ATL, &ATR, &ABL, &ABR, 0, 0, FLA_BR ); FLA_Part_2x2( B, &BTL, &BTR, &BBL, &BBR, 0, 0, FLA_TL ); FLA_Part_2x2( C, &CTL, &CTR, &CBL, &CBR, 0, 0, FLA_BL ); while ( FLA_Obj_length( ABR ) < FLA_Obj_length( A ) ){ b = FLA_Determine_blocksize( CTR, FLA_TR, FLA_Cntl_blocksize( cntl ) ); FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, &A01, /**/ &A02, &A10, &A11, /**/ &A12, /* ************* */ /* ******************** */ ABL, /**/ ABR, &A20, &A21, /**/ &A22, b, b, FLA_TL ); FLA_Repart_2x2_to_3x3( BTL, /**/ BTR, &B00, /**/ &B01, &B02, /* ************* */ /* ******************** */ &B10, /**/ &B11, &B12, BBL, /**/ BBR, &B20, /**/ &B21, &B22, b, b, FLA_BR ); FLA_Repart_2x2_to_3x3( CTL, /**/ CTR, &C00, /**/ &C01, &C02, &C10, /**/ &C11, &C12, /* ************* */ /* ******************** */ CBL, /**/ CBR, &C20, /**/ &C21, &C22, b, b, FLA_TR ); // Loop Invariant: // CTL = sylv( ATL, BTL, CTL - ATR * sylv( ABR, BTL, CBL ) ) // CTR = CTR // CBL = sylv( ABR, BTL, CBL ) // CBR = CBR /*------------------------------------------------------------*/ // C21 = sylv( A22, B11, C21 -/+ C20 * B01 ); FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, FLA_NEGATE( isgn ), C20, B01, FLA_ONE, C21, FLA_Cntl_sub_gemm1( cntl ) ); FLA_Sylv_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, isgn, A22, B11, C21, scale, FLA_Cntl_sub_sylv1( cntl ) ); // C11 = sylv( A11, B11, C11 - A12 * C21 -/+ C10 * B01 ); FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, FLA_NEGATE( isgn ), C10, B01, FLA_ONE, C11, FLA_Cntl_sub_gemm2( cntl ) ); FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, FLA_MINUS_ONE, A12, C21, FLA_ONE, C11, FLA_Cntl_sub_gemm3( cntl ) ); FLA_Sylv_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, isgn, A11, B11, C11, scale, FLA_Cntl_sub_sylv2( cntl ) ); // C01 = sylv( A00, B11, C01 - A01 * C11 - A02 * C21 -/+ C00 * B01 ); FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, FLA_NEGATE( isgn ), C00, B01, FLA_ONE, C01, FLA_Cntl_sub_gemm4( cntl ) ); FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, FLA_MINUS_ONE, A02, C21, FLA_ONE, C01, FLA_Cntl_sub_gemm5( cntl ) ); FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, FLA_MINUS_ONE, A01, C11, FLA_ONE, C01, FLA_Cntl_sub_gemm6( cntl ) ); FLA_Sylv_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, isgn, A00, B11, C01, scale, FLA_Cntl_sub_sylv3( cntl ) ); /*------------------------------------------------------------*/ FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, /**/ A01, A02, /* ************** */ /* ****************** */ A10, /**/ A11, A12, &ABL, /**/ &ABR, A20, /**/ A21, A22, FLA_BR ); FLA_Cont_with_3x3_to_2x2( &BTL, /**/ &BTR, B00, B01, /**/ B02, B10, B11, /**/ B12, /* ************** */ /* ****************** */ &BBL, /**/ &BBR, B20, B21, /**/ B22, FLA_TL ); FLA_Cont_with_3x3_to_2x2( &CTL, /**/ &CTR, C00, C01, /**/ C02, /* ************** */ /* ****************** */ C10, C11, /**/ C12, &CBL, /**/ &CBR, C20, C21, /**/ C22, FLA_BL ); } return FLA_SUCCESS; }
FLA_Error FLA_Sylv_nn_blk_var4 | ( | FLA_Obj | isgn, |
FLA_Obj | A, | ||
FLA_Obj | B, | ||
FLA_Obj | C, | ||
FLA_Obj | scale, | ||
fla_sylv_t * | cntl | ||
) |
References FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_MINUS_ONE, FLA_Obj_length(), FLA_ONE, FLA_Part_2x2(), FLA_Repart_2x2_to_3x3(), and FLA_Sylv_internal().
Referenced by FLA_Sylv_nn().
{ FLA_Obj ATL, ATR, A00, A01, A02, ABL, ABR, A10, A11, A12, A20, A21, A22; FLA_Obj BTL, BTR, B00, B01, B02, BBL, BBR, B10, B11, B12, B20, B21, B22; FLA_Obj CTL, CTR, C00, C01, C02, CBL, CBR, C10, C11, C12, C20, C21, C22; dim_t b; FLA_Part_2x2( A, &ATL, &ATR, &ABL, &ABR, 0, 0, FLA_BR ); FLA_Part_2x2( B, &BTL, &BTR, &BBL, &BBR, 0, 0, FLA_TL ); FLA_Part_2x2( C, &CTL, &CTR, &CBL, &CBR, 0, 0, FLA_BL ); while ( FLA_Obj_length( ABR ) < FLA_Obj_length( A ) ){ b = FLA_Determine_blocksize( CTR, FLA_TR, FLA_Cntl_blocksize( cntl ) ); FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, &A01, /**/ &A02, &A10, &A11, /**/ &A12, /* ************* */ /* ******************** */ ABL, /**/ ABR, &A20, &A21, /**/ &A22, b, b, FLA_TL ); FLA_Repart_2x2_to_3x3( BTL, /**/ BTR, &B00, /**/ &B01, &B02, /* ************* */ /* ******************** */ &B10, /**/ &B11, &B12, BBL, /**/ BBR, &B20, /**/ &B21, &B22, b, b, FLA_BR ); FLA_Repart_2x2_to_3x3( CTL, /**/ CTR, &C00, /**/ &C01, &C02, &C10, /**/ &C11, &C12, /* ************* */ /* ******************** */ CBL, /**/ CBR, &C20, /**/ &C21, &C22, b, b, FLA_TR ); // Loop Invariant: // CTL = CTL // CTR = CTR // CBL = sylv( ABR, BTL, CBL ) // CBR = CBR - sylv( ABR, BTL, CBL ) * BTR /*------------------------------------------------------------*/ // C10 = sylv( A11, B00, C10 - A12 * C20 ); FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, FLA_MINUS_ONE, A12, C20, FLA_ONE, C10, FLA_Cntl_sub_gemm1( cntl ) ); FLA_Sylv_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, isgn, A11, B00, C10, scale, FLA_Cntl_sub_sylv1( cntl ) ); // C21 = sylv( A22, B11, C21 ); FLA_Sylv_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, isgn, A22, B11, C21, scale, FLA_Cntl_sub_sylv2( cntl ) ); // C11 = sylv( A11, B11, C11 - A12 * C21 -/+ C10 * B01 ); FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, FLA_NEGATE( isgn ), C10, B01, FLA_ONE, C11, FLA_Cntl_sub_gemm2( cntl ) ); FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, FLA_MINUS_ONE, A12, C21, FLA_ONE, C11, FLA_Cntl_sub_gemm3( cntl ) ); FLA_Sylv_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, isgn, A11, B11, C11, scale, FLA_Cntl_sub_sylv3( cntl ) ); // C12 = C12 -/+ C10 * B02 -/+ C11 * B12; FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, FLA_NEGATE( isgn ), C11, B12, FLA_ONE, C12, FLA_Cntl_sub_gemm4( cntl ) ); FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, FLA_NEGATE( isgn ), C10, B02, FLA_ONE, C12, FLA_Cntl_sub_gemm5( cntl ) ); // C22 = C22 -/+ C21 * B12; FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, FLA_NEGATE( isgn ), C21, B12, FLA_ONE, C22, FLA_Cntl_sub_gemm6( cntl ) ); /*------------------------------------------------------------*/ FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, /**/ A01, A02, /* ************** */ /* ****************** */ A10, /**/ A11, A12, &ABL, /**/ &ABR, A20, /**/ A21, A22, FLA_BR ); FLA_Cont_with_3x3_to_2x2( &BTL, /**/ &BTR, B00, B01, /**/ B02, B10, B11, /**/ B12, /* ************** */ /* ****************** */ &BBL, /**/ &BBR, B20, B21, /**/ B22, FLA_TL ); FLA_Cont_with_3x3_to_2x2( &CTL, /**/ &CTR, C00, C01, /**/ C02, /* ************** */ /* ****************** */ C10, C11, /**/ C12, &CBL, /**/ &CBR, C20, C21, /**/ C22, FLA_BL ); } return FLA_SUCCESS; }
FLA_Error FLA_Sylv_nn_blk_var5 | ( | FLA_Obj | isgn, |
FLA_Obj | A, | ||
FLA_Obj | B, | ||
FLA_Obj | C, | ||
FLA_Obj | scale, | ||
fla_sylv_t * | cntl | ||
) |
References FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_MINUS_ONE, FLA_Obj_length(), FLA_ONE, FLA_Part_2x2(), FLA_Repart_2x2_to_3x3(), and FLA_Sylv_internal().
Referenced by FLA_Sylv_nn().
{ FLA_Obj ATL, ATR, A00, A01, A02, ABL, ABR, A10, A11, A12, A20, A21, A22; FLA_Obj BTL, BTR, B00, B01, B02, BBL, BBR, B10, B11, B12, B20, B21, B22; FLA_Obj CTL, CTR, C00, C01, C02, CBL, CBR, C10, C11, C12, C20, C21, C22; dim_t b; FLA_Part_2x2( A, &ATL, &ATR, &ABL, &ABR, 0, 0, FLA_BR ); FLA_Part_2x2( B, &BTL, &BTR, &BBL, &BBR, 0, 0, FLA_TL ); FLA_Part_2x2( C, &CTL, &CTR, &CBL, &CBR, 0, 0, FLA_BL ); while ( FLA_Obj_length( ABR ) < FLA_Obj_length( A ) ){ b = FLA_Determine_blocksize( CTR, FLA_TR, FLA_Cntl_blocksize( cntl ) ); FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, &A01, /**/ &A02, &A10, &A11, /**/ &A12, /* ************* */ /* ******************** */ ABL, /**/ ABR, &A20, &A21, /**/ &A22, b, b, FLA_TL ); FLA_Repart_2x2_to_3x3( BTL, /**/ BTR, &B00, /**/ &B01, &B02, /* ************* */ /* ******************** */ &B10, /**/ &B11, &B12, BBL, /**/ BBR, &B20, /**/ &B21, &B22, b, b, FLA_BR ); FLA_Repart_2x2_to_3x3( CTL, /**/ CTR, &C00, /**/ &C01, &C02, &C10, /**/ &C11, &C12, /* ************* */ /* ******************** */ CBL, /**/ CBR, &C20, /**/ &C21, &C22, b, b, FLA_TR ); // Loop Invariant: // CTL = CTL // CTR = CTR // CBL = sylv( ABR, BTL, CBL ) // CBR = sylv( ABR, BBR, CBR - sylv( ABR, BTL, CBL ) * BTR ) /*------------------------------------------------------------*/ // C10 = sylv( A11, B00, C10 - A12 * C20 ); FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, FLA_MINUS_ONE, A12, C20, FLA_ONE, C10, FLA_Cntl_sub_gemm1( cntl ) ); FLA_Sylv_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, isgn, A11, B00, C10, scale, FLA_Cntl_sub_sylv1( cntl ) ); // C11 = sylv( A11, B11, C11 - A12 * C21 -/+ C10 * B01 ); FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, FLA_NEGATE( isgn ), C10, B01, FLA_ONE, C11, FLA_Cntl_sub_gemm2( cntl ) ); FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, FLA_MINUS_ONE, A12, C21, FLA_ONE, C11, FLA_Cntl_sub_gemm3( cntl ) ); FLA_Sylv_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, isgn, A11, B11, C11, scale, FLA_Cntl_sub_sylv2( cntl ) ); // C12 = sylv( A11, B22, C12 - A12 * C22 -/+ C10 * B02 -/+ C11 * B12 ); FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, FLA_NEGATE( isgn ), C11, B12, FLA_ONE, C12, FLA_Cntl_sub_gemm4( cntl ) ); FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, FLA_NEGATE( isgn ), C10, B02, FLA_ONE, C12, FLA_Cntl_sub_gemm5( cntl ) ); FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, FLA_MINUS_ONE, A12, C22, FLA_ONE, C12, FLA_Cntl_sub_gemm6( cntl ) ); FLA_Sylv_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, isgn, A11, B22, C12, scale, FLA_Cntl_sub_sylv3( cntl ) ); /*------------------------------------------------------------*/ FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, /**/ A01, A02, /* ************** */ /* ****************** */ A10, /**/ A11, A12, &ABL, /**/ &ABR, A20, /**/ A21, A22, FLA_BR ); FLA_Cont_with_3x3_to_2x2( &BTL, /**/ &BTR, B00, B01, /**/ B02, B10, B11, /**/ B12, /* ************** */ /* ****************** */ &BBL, /**/ &BBR, B20, B21, /**/ B22, FLA_TL ); FLA_Cont_with_3x3_to_2x2( &CTL, /**/ &CTR, C00, C01, /**/ C02, /* ************** */ /* ****************** */ C10, C11, /**/ C12, &CBL, /**/ &CBR, C20, C21, /**/ C22, FLA_BL ); } return FLA_SUCCESS; }
FLA_Error FLA_Sylv_nn_blk_var6 | ( | FLA_Obj | isgn, |
FLA_Obj | A, | ||
FLA_Obj | B, | ||
FLA_Obj | C, | ||
FLA_Obj | scale, | ||
fla_sylv_t * | cntl | ||
) |
References FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_MINUS_ONE, FLA_Obj_length(), FLA_ONE, FLA_Part_2x2(), FLA_Repart_2x2_to_3x3(), and FLA_Sylv_internal().
Referenced by FLA_Sylv_nn().
{ FLA_Obj ATL, ATR, A00, A01, A02, ABL, ABR, A10, A11, A12, A20, A21, A22; FLA_Obj BTL, BTR, B00, B01, B02, BBL, BBR, B10, B11, B12, B20, B21, B22; FLA_Obj CTL, CTR, C00, C01, C02, CBL, CBR, C10, C11, C12, C20, C21, C22; dim_t b; FLA_Part_2x2( A, &ATL, &ATR, &ABL, &ABR, 0, 0, FLA_BR ); FLA_Part_2x2( B, &BTL, &BTR, &BBL, &BBR, 0, 0, FLA_TL ); FLA_Part_2x2( C, &CTL, &CTR, &CBL, &CBR, 0, 0, FLA_BL ); while ( FLA_Obj_length( ABR ) < FLA_Obj_length( A ) ){ b = FLA_Determine_blocksize( CTR, FLA_TR, FLA_Cntl_blocksize( cntl ) ); FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, &A01, /**/ &A02, &A10, &A11, /**/ &A12, /* ************* */ /* ******************** */ ABL, /**/ ABR, &A20, &A21, /**/ &A22, b, b, FLA_TL ); FLA_Repart_2x2_to_3x3( BTL, /**/ BTR, &B00, /**/ &B01, &B02, /* ************* */ /* ******************** */ &B10, /**/ &B11, &B12, BBL, /**/ BBR, &B20, /**/ &B21, &B22, b, b, FLA_BR ); FLA_Repart_2x2_to_3x3( CTL, /**/ CTR, &C00, /**/ &C01, &C02, &C10, /**/ &C11, &C12, /* ************* */ /* ******************** */ CBL, /**/ CBR, &C20, /**/ &C21, &C22, b, b, FLA_TR ); // Loop Invariant: // CTL = CTL - ATR * sylv( ABR, BTL, CBL ) // CTR = CTR // CBL = sylv( ABR, BTL, CBL ) // CBR = CBR - sylv( ABR, BTL, CBL ) * BTR /*------------------------------------------------------------*/ // C10 = sylv( A11, B00, C10 ); FLA_Sylv_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, isgn, A11, B00, C10, scale, FLA_Cntl_sub_sylv1( cntl ) ); // C00 = C00 - A01 * C10; FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, FLA_MINUS_ONE, A01, C10, FLA_ONE, C00, FLA_Cntl_sub_gemm1( cntl ) ); // C21 = sylv( A22, B11, C21 ); FLA_Sylv_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, isgn, A22, B11, C21, scale, FLA_Cntl_sub_sylv2( cntl ) ); // C11 = sylv( A11, B11, C11 - A12 * C21 -/+ C10 * B01 ); FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, FLA_NEGATE( isgn ), C10, B01, FLA_ONE, C11, FLA_Cntl_sub_gemm2( cntl ) ); FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, FLA_MINUS_ONE, A12, C21, FLA_ONE, C11, FLA_Cntl_sub_gemm3( cntl ) ); FLA_Sylv_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, isgn, A11, B11, C11, scale, FLA_Cntl_sub_sylv3( cntl ) ); // C01 = C01 - A01 * C11 - A02 * C21; FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, FLA_MINUS_ONE, A02, C21, FLA_ONE, C01, FLA_Cntl_sub_gemm4( cntl ) ); FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, FLA_MINUS_ONE, A01, C11, FLA_ONE, C01, FLA_Cntl_sub_gemm5( cntl ) ); // C12 = C12 -/+ C10 * B02 -/+ C11 * B12; FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, FLA_NEGATE( isgn ), C11, B12, FLA_ONE, C12, FLA_Cntl_sub_gemm6( cntl ) ); FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, FLA_NEGATE( isgn ), C10, B02, FLA_ONE, C12, FLA_Cntl_sub_gemm7( cntl ) ); // C22 = C22 -/+ C21 * B12; FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, FLA_NEGATE( isgn ), C21, B12, FLA_ONE, C22, FLA_Cntl_sub_gemm8( cntl ) ); /*------------------------------------------------------------*/ FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, /**/ A01, A02, /* ************** */ /* ****************** */ A10, /**/ A11, A12, &ABL, /**/ &ABR, A20, /**/ A21, A22, FLA_BR ); FLA_Cont_with_3x3_to_2x2( &BTL, /**/ &BTR, B00, B01, /**/ B02, B10, B11, /**/ B12, /* ************** */ /* ****************** */ &BBL, /**/ &BBR, B20, B21, /**/ B22, FLA_TL ); FLA_Cont_with_3x3_to_2x2( &CTL, /**/ &CTR, C00, C01, /**/ C02, /* ************** */ /* ****************** */ C10, C11, /**/ C12, &CBL, /**/ &CBR, C20, C21, /**/ C22, FLA_BL ); } return FLA_SUCCESS; }
FLA_Error FLA_Sylv_nn_blk_var7 | ( | FLA_Obj | isgn, |
FLA_Obj | A, | ||
FLA_Obj | B, | ||
FLA_Obj | C, | ||
FLA_Obj | scale, | ||
fla_sylv_t * | cntl | ||
) |
References FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_MINUS_ONE, FLA_Obj_length(), FLA_ONE, FLA_Part_2x2(), FLA_Repart_2x2_to_3x3(), and FLA_Sylv_internal().
Referenced by FLA_Sylv_nn().
{ FLA_Obj ATL, ATR, A00, A01, A02, ABL, ABR, A10, A11, A12, A20, A21, A22; FLA_Obj BTL, BTR, B00, B01, B02, BBL, BBR, B10, B11, B12, B20, B21, B22; FLA_Obj CTL, CTR, C00, C01, C02, CBL, CBR, C10, C11, C12, C20, C21, C22; dim_t b; FLA_Part_2x2( A, &ATL, &ATR, &ABL, &ABR, 0, 0, FLA_BR ); FLA_Part_2x2( B, &BTL, &BTR, &BBL, &BBR, 0, 0, FLA_TL ); FLA_Part_2x2( C, &CTL, &CTR, &CBL, &CBR, 0, 0, FLA_BL ); while ( FLA_Obj_length( ABR ) < FLA_Obj_length( A ) ){ b = FLA_Determine_blocksize( CTR, FLA_TR, FLA_Cntl_blocksize( cntl ) ); FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, &A01, /**/ &A02, &A10, &A11, /**/ &A12, /* ************* */ /* ******************** */ ABL, /**/ ABR, &A20, &A21, /**/ &A22, b, b, FLA_TL ); FLA_Repart_2x2_to_3x3( BTL, /**/ BTR, &B00, /**/ &B01, &B02, /* ************* */ /* ******************** */ &B10, /**/ &B11, &B12, BBL, /**/ BBR, &B20, /**/ &B21, &B22, b, b, FLA_BR ); FLA_Repart_2x2_to_3x3( CTL, /**/ CTR, &C00, /**/ &C01, &C02, &C10, /**/ &C11, &C12, /* ************* */ /* ******************** */ CBL, /**/ CBR, &C20, /**/ &C21, &C22, b, b, FLA_TR ); // Loop Invariant: // CTL = sylv( ATL, BTL, CTL - ATR * sylv( ABR, BTL, CBL ) ) // CTR = CTR // CBL = sylv( ABR, BTL, CBL ) // CBR = CBR - sylv( ABR, BTL, CBL ) * BTR /*------------------------------------------------------------*/ // C21 = sylv( A22, B11, C21 ); FLA_Sylv_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, isgn, A22, B11, C21, scale, FLA_Cntl_sub_sylv1( cntl ) ); // C11 = sylv( A11, B11, C11 - A12 * C21 -/+ C10 * B01 ); FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, FLA_NEGATE( isgn ), C10, B01, FLA_ONE, C11, FLA_Cntl_sub_gemm1( cntl ) ); FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, FLA_MINUS_ONE, A12, C21, FLA_ONE, C11, FLA_Cntl_sub_gemm2( cntl ) ); FLA_Sylv_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, isgn, A11, B11, C11, scale, FLA_Cntl_sub_sylv2( cntl ) ); // C01 = sylv( A00, B11, C01 - A01 * C11 - A02 * C21 -/+ C00 * B01 ); FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, FLA_NEGATE( isgn ), C00, B01, FLA_ONE, C01, FLA_Cntl_sub_gemm3( cntl ) ); FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, FLA_MINUS_ONE, A02, C21, FLA_ONE, C01, FLA_Cntl_sub_gemm4( cntl ) ); FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, FLA_MINUS_ONE, A01, C11, FLA_ONE, C01, FLA_Cntl_sub_gemm5( cntl ) ); FLA_Sylv_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, isgn, A00, B11, C01, scale, FLA_Cntl_sub_sylv3( cntl ) ); // C12 = C12 -/+ C10 * B02 -/+ C11 * B12; FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, FLA_NEGATE( isgn ), C11, B12, FLA_ONE, C12, FLA_Cntl_sub_gemm6( cntl ) ); FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, FLA_NEGATE( isgn ), C10, B02, FLA_ONE, C12, FLA_Cntl_sub_gemm7( cntl ) ); // C22 = C22 -/+ C21 * B12; FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, FLA_NEGATE( isgn ), C21, B12, FLA_ONE, C22, FLA_Cntl_sub_gemm8( cntl ) ); /*------------------------------------------------------------*/ FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, /**/ A01, A02, /* ************** */ /* ****************** */ A10, /**/ A11, A12, &ABL, /**/ &ABR, A20, /**/ A21, A22, FLA_BR ); FLA_Cont_with_3x3_to_2x2( &BTL, /**/ &BTR, B00, B01, /**/ B02, B10, B11, /**/ B12, /* ************** */ /* ****************** */ &BBL, /**/ &BBR, B20, B21, /**/ B22, FLA_TL ); FLA_Cont_with_3x3_to_2x2( &CTL, /**/ &CTR, C00, C01, /**/ C02, /* ************** */ /* ****************** */ C10, C11, /**/ C12, &CBL, /**/ &CBR, C20, C21, /**/ C22, FLA_BL ); } return FLA_SUCCESS; }
FLA_Error FLA_Sylv_nn_blk_var8 | ( | FLA_Obj | isgn, |
FLA_Obj | A, | ||
FLA_Obj | B, | ||
FLA_Obj | C, | ||
FLA_Obj | scale, | ||
fla_sylv_t * | cntl | ||
) |
References FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_MINUS_ONE, FLA_Obj_length(), FLA_ONE, FLA_Part_2x2(), FLA_Repart_2x2_to_3x3(), and FLA_Sylv_internal().
Referenced by FLA_Sylv_nn().
{ FLA_Obj ATL, ATR, A00, A01, A02, ABL, ABR, A10, A11, A12, A20, A21, A22; FLA_Obj BTL, BTR, B00, B01, B02, BBL, BBR, B10, B11, B12, B20, B21, B22; FLA_Obj CTL, CTR, C00, C01, C02, CBL, CBR, C10, C11, C12, C20, C21, C22; dim_t b; FLA_Part_2x2( A, &ATL, &ATR, &ABL, &ABR, 0, 0, FLA_BR ); FLA_Part_2x2( B, &BTL, &BTR, &BBL, &BBR, 0, 0, FLA_TL ); FLA_Part_2x2( C, &CTL, &CTR, &CBL, &CBR, 0, 0, FLA_BL ); while ( FLA_Obj_length( ABR ) < FLA_Obj_length( A ) ){ b = FLA_Determine_blocksize( CTR, FLA_TR, FLA_Cntl_blocksize( cntl ) ); FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, &A01, /**/ &A02, &A10, &A11, /**/ &A12, /* ************* */ /* ******************** */ ABL, /**/ ABR, &A20, &A21, /**/ &A22, b, b, FLA_TL ); FLA_Repart_2x2_to_3x3( BTL, /**/ BTR, &B00, /**/ &B01, &B02, /* ************* */ /* ******************** */ &B10, /**/ &B11, &B12, BBL, /**/ BBR, &B20, /**/ &B21, &B22, b, b, FLA_BR ); FLA_Repart_2x2_to_3x3( CTL, /**/ CTR, &C00, /**/ &C01, &C02, &C10, /**/ &C11, &C12, /* ************* */ /* ******************** */ CBL, /**/ CBR, &C20, /**/ &C21, &C22, b, b, FLA_TR ); // Loop Invariant: // CTL = CTL - ATR * sylv( ABR, BTL, CBL ) // CTR = CTR // CBL = sylv( ABR, BTL, CBL ) // CBR = sylv( ABR, BBR, CBR - sylv( ABR, BTL, CBL ) * BTR ) /*------------------------------------------------------------*/ // C10 = sylv( A11, B00, C10 ); FLA_Sylv_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, isgn, A11, B00, C10, scale, FLA_Cntl_sub_sylv1( cntl ) ); // C00 = C00 - A01 * C10; FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, FLA_MINUS_ONE, A01, C10, FLA_ONE, C00, FLA_Cntl_sub_gemm1( cntl ) ); // C11 = sylv( A11, B11, C11 - A12 * C21 -/+ C10 * B01 ); FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, FLA_NEGATE( isgn ), C10, B01, FLA_ONE, C11, FLA_Cntl_sub_gemm2( cntl ) ); FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, FLA_MINUS_ONE, A12, C21, FLA_ONE, C11, FLA_Cntl_sub_gemm3( cntl ) ); FLA_Sylv_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, isgn, A11, B11, C11, scale, FLA_Cntl_sub_sylv2( cntl ) ); // C01 = C01 - A01 * C11 - A02 * C21; FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, FLA_MINUS_ONE, A02, C21, FLA_ONE, C01, FLA_Cntl_sub_gemm4( cntl ) ); FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, FLA_MINUS_ONE, A01, C11, FLA_ONE, C01, FLA_Cntl_sub_gemm5( cntl ) ); // C12 = sylv( A11, B22, C12 - A12 * C22 -/+ C10 * B02 -/+ C11 * B12 ); FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, FLA_NEGATE( isgn ), C11, B12, FLA_ONE, C12, FLA_Cntl_sub_gemm6( cntl ) ); FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, FLA_NEGATE( isgn ), C10, B02, FLA_ONE, C12, FLA_Cntl_sub_gemm7( cntl ) ); FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, FLA_MINUS_ONE, A12, C22, FLA_ONE, C12, FLA_Cntl_sub_gemm8( cntl ) ); FLA_Sylv_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, isgn, A11, B22, C12, scale, FLA_Cntl_sub_sylv3( cntl ) ); /*------------------------------------------------------------*/ FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, /**/ A01, A02, /* ************** */ /* ****************** */ A10, /**/ A11, A12, &ABL, /**/ &ABR, A20, /**/ A21, A22, FLA_BR ); FLA_Cont_with_3x3_to_2x2( &BTL, /**/ &BTR, B00, B01, /**/ B02, B10, B11, /**/ B12, /* ************** */ /* ****************** */ &BBL, /**/ &BBR, B20, B21, /**/ B22, FLA_TL ); FLA_Cont_with_3x3_to_2x2( &CTL, /**/ &CTR, C00, C01, /**/ C02, /* ************** */ /* ****************** */ C10, C11, /**/ C12, &CBL, /**/ &CBR, C20, C21, /**/ C22, FLA_BL ); } return FLA_SUCCESS; }
FLA_Error FLA_Sylv_nn_blk_var9 | ( | FLA_Obj | isgn, |
FLA_Obj | A, | ||
FLA_Obj | B, | ||
FLA_Obj | C, | ||
FLA_Obj | scale, | ||
fla_sylv_t * | cntl | ||
) |
References FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_MINUS_ONE, FLA_Obj_length(), FLA_ONE, FLA_Part_2x2(), FLA_Repart_2x2_to_3x3(), and FLA_Sylv_internal().
Referenced by FLA_Sylv_nn().
{ FLA_Obj ATL, ATR, A00, A01, A02, ABL, ABR, A10, A11, A12, A20, A21, A22; FLA_Obj BTL, BTR, B00, B01, B02, BBL, BBR, B10, B11, B12, B20, B21, B22; FLA_Obj CTL, CTR, C00, C01, C02, CBL, CBR, C10, C11, C12, C20, C21, C22; dim_t b; FLA_Part_2x2( A, &ATL, &ATR, &ABL, &ABR, 0, 0, FLA_BR ); FLA_Part_2x2( B, &BTL, &BTR, &BBL, &BBR, 0, 0, FLA_TL ); FLA_Part_2x2( C, &CTL, &CTR, &CBL, &CBR, 0, 0, FLA_BL ); while ( FLA_Obj_length( ABR ) < FLA_Obj_length( A ) ){ b = FLA_Determine_blocksize( CTR, FLA_TR, FLA_Cntl_blocksize( cntl ) ); FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, &A01, /**/ &A02, &A10, &A11, /**/ &A12, /* ************* */ /* ******************** */ ABL, /**/ ABR, &A20, &A21, /**/ &A22, b, b, FLA_TL ); FLA_Repart_2x2_to_3x3( BTL, /**/ BTR, &B00, /**/ &B01, &B02, /* ************* */ /* ******************** */ &B10, /**/ &B11, &B12, BBL, /**/ BBR, &B20, /**/ &B21, &B22, b, b, FLA_BR ); FLA_Repart_2x2_to_3x3( CTL, /**/ CTR, &C00, /**/ &C01, &C02, &C10, /**/ &C11, &C12, /* ************* */ /* ******************** */ CBL, /**/ CBR, &C20, /**/ &C21, &C22, b, b, FLA_TR ); // Loop Invariant: // CTL = sylv( ATL, BTL, CTL - ATR * sylv( ABR, BTL, CBL ) ) // CTR = CTR // CBL = sylv( ABR, BTL, CBL) // CBR = sylv( ABR, BBR, CBR - sylv( ABR, BTL, CBL ) * BTR ) /*------------------------------------------------------------*/ // C11 = sylv( A11, B11, C11 - A12 * C21 -/+ C10 * B01 ); FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, FLA_NEGATE( isgn ), C10, B01, FLA_ONE, C11, FLA_Cntl_sub_gemm1( cntl ) ); FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, FLA_MINUS_ONE, A12, C21, FLA_ONE, C11, FLA_Cntl_sub_gemm2( cntl ) ); FLA_Sylv_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, isgn, A11, B11, C11, scale, FLA_Cntl_sub_sylv1( cntl ) ); // C01 = sylv( A00, B11, C01 - A01 * C11 - A02 * C21 -/+ C00 * B01 ); FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, FLA_NEGATE( isgn ), C00, B01, FLA_ONE, C01, FLA_Cntl_sub_gemm3( cntl ) ); FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, FLA_MINUS_ONE, A02, C21, FLA_ONE, C01, FLA_Cntl_sub_gemm4( cntl ) ); FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, FLA_MINUS_ONE, A01, C11, FLA_ONE, C01, FLA_Cntl_sub_gemm5( cntl ) ); FLA_Sylv_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, isgn, A00, B11, C01, scale, FLA_Cntl_sub_sylv2( cntl ) ); // C12 = sylv( A11, B22, C12 - A12 * C22 -/+ C10 * B02 -/+ C11 * B12 ); FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, FLA_NEGATE( isgn ), C11, B12, FLA_ONE, C12, FLA_Cntl_sub_gemm6( cntl ) ); FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, FLA_NEGATE( isgn ), C10, B02, FLA_ONE, C12, FLA_Cntl_sub_gemm7( cntl ) ); FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, FLA_MINUS_ONE, A12, C22, FLA_ONE, C12, FLA_Cntl_sub_gemm8( cntl ) ); FLA_Sylv_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, isgn, A11, B22, C12, scale, FLA_Cntl_sub_sylv3( cntl ) ); /*------------------------------------------------------------*/ FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, /**/ A01, A02, /* ************** */ /* ****************** */ A10, /**/ A11, A12, &ABL, /**/ &ABR, A20, /**/ A21, A22, FLA_BR ); FLA_Cont_with_3x3_to_2x2( &BTL, /**/ &BTR, B00, B01, /**/ B02, B10, B11, /**/ B12, /* ************** */ /* ****************** */ &BBL, /**/ &BBR, B20, B21, /**/ B22, FLA_TL ); FLA_Cont_with_3x3_to_2x2( &CTL, /**/ &CTR, C00, C01, /**/ C02, /* ************** */ /* ****************** */ C10, C11, /**/ C12, &CBL, /**/ &CBR, C20, C21, /**/ C22, FLA_BL ); } return FLA_SUCCESS; }
FLA_Error FLA_Sylv_nn_opc_var1 | ( | float | sgn, |
int | m_C, | ||
int | n_C, | ||
scomplex * | buff_A, | ||
int | rs_A, | ||
int | cs_A, | ||
scomplex * | buff_B, | ||
int | rs_B, | ||
int | cs_B, | ||
scomplex * | buff_C, | ||
int | rs_C, | ||
int | cs_C, | ||
scomplex * | buff_scale, | ||
int * | info | ||
) |
References bl1_cdot(), BLIS1_NO_CONJUGATE, scomplex::imag, and scomplex::real.
Referenced by FLA_Sylv_nn_opt_var1().
{ int l, k; for ( l = 0; l < n_C; l++ ) { for ( k = m_C - 1; k >= 0; k-- ) { scomplex* a12t = buff_A + (k+1)*cs_A + (k )*rs_A; scomplex* b01 = buff_B + (l )*cs_B + (0 )*rs_B; scomplex* c10t = buff_C + (0 )*cs_C + (k )*rs_C; scomplex* c21 = buff_C + (l )*cs_C + (k+1)*rs_C; scomplex* alpha11 = buff_A + (k )*cs_A + (k )*rs_A; scomplex* beta11 = buff_B + (l )*cs_B + (l )*rs_B; scomplex* ckl = buff_C + (l )*cs_C + (k )*rs_C; scomplex suml; scomplex sumr; scomplex vec; scomplex a11; scomplex x11; int m_behind = m_C - k - 1; int n_behind = l; /*------------------------------------------------------------*/ bl1_cdot( BLIS1_NO_CONJUGATE, m_behind, a12t, cs_A, c21, rs_C, &suml ); bl1_cdot( BLIS1_NO_CONJUGATE, n_behind, c10t, cs_C, b01, rs_B, &sumr ); vec.real = ckl->real - ( suml.real + sgn * sumr.real ); vec.imag = ckl->imag - ( suml.imag + sgn * sumr.imag ); a11.real = alpha11->real + sgn * beta11->real; a11.imag = alpha11->imag + sgn * beta11->imag; bl1_cdiv3( &vec, &a11, &x11 ); *ckl = x11; /*------------------------------------------------------------*/ } } return FLA_SUCCESS; }
FLA_Error FLA_Sylv_nn_opd_var1 | ( | double | sgn, |
int | m_C, | ||
int | n_C, | ||
double * | buff_A, | ||
int | rs_A, | ||
int | cs_A, | ||
double * | buff_B, | ||
int | rs_B, | ||
int | cs_B, | ||
double * | buff_C, | ||
int | rs_C, | ||
int | cs_C, | ||
double * | buff_scale, | ||
int * | info | ||
) |
References bl1_ddot(), and BLIS1_NO_CONJUGATE.
Referenced by FLA_Sylv_nn_opt_var1().
{ int l, k; for ( l = 0; l < n_C; l++ ) { for ( k = m_C - 1; k >= 0; k-- ) { double* a12t = buff_A + (k+1)*cs_A + (k )*rs_A; double* b01 = buff_B + (l )*cs_B + (0 )*rs_B; double* c10t = buff_C + (0 )*cs_C + (k )*rs_C; double* c21 = buff_C + (l )*cs_C + (k+1)*rs_C; double* alpha11 = buff_A + (k )*cs_A + (k )*rs_A; double* beta11 = buff_B + (l )*cs_B + (l )*rs_B; double* ckl = buff_C + (l )*cs_C + (k )*rs_C; double suml; double sumr; double vec; double a11; double x11; int m_behind = m_C - k - 1; int n_behind = l; /*------------------------------------------------------------*/ bl1_ddot( BLIS1_NO_CONJUGATE, m_behind, a12t, cs_A, c21, rs_C, &suml ); bl1_ddot( BLIS1_NO_CONJUGATE, n_behind, c10t, cs_C, b01, rs_B, &sumr ); vec = (*ckl) - ( suml + sgn * sumr ); a11 = (*alpha11) + sgn * (*beta11); bl1_ddiv3( &vec, &a11, &x11 ); *ckl = x11; /*------------------------------------------------------------*/ } } return FLA_SUCCESS; }
FLA_Error FLA_Sylv_nn_ops_var1 | ( | float | sgn, |
int | m_C, | ||
int | n_C, | ||
float * | buff_A, | ||
int | rs_A, | ||
int | cs_A, | ||
float * | buff_B, | ||
int | rs_B, | ||
int | cs_B, | ||
float * | buff_C, | ||
int | rs_C, | ||
int | cs_C, | ||
float * | buff_scale, | ||
int * | info | ||
) |
References bl1_sdot(), and BLIS1_NO_CONJUGATE.
Referenced by FLA_Sylv_nn_opt_var1().
{ int l, k; for ( l = 0; l < n_C; l++ ) { for ( k = m_C - 1; k >= 0; k-- ) { float* a12t = buff_A + (k+1)*cs_A + (k )*rs_A; float* b01 = buff_B + (l )*cs_B + (0 )*rs_B; float* c10t = buff_C + (0 )*cs_C + (k )*rs_C; float* c21 = buff_C + (l )*cs_C + (k+1)*rs_C; float* alpha11 = buff_A + (k )*cs_A + (k )*rs_A; float* beta11 = buff_B + (l )*cs_B + (l )*rs_B; float* ckl = buff_C + (l )*cs_C + (k )*rs_C; float suml; float sumr; float vec; float a11; float x11; int m_behind = m_C - k - 1; int n_behind = l; /*------------------------------------------------------------*/ bl1_sdot( BLIS1_NO_CONJUGATE, m_behind, a12t, cs_A, c21, rs_C, &suml ); bl1_sdot( BLIS1_NO_CONJUGATE, n_behind, c10t, cs_C, b01, rs_B, &sumr ); vec = (*ckl) - ( suml + sgn * sumr ); a11 = (*alpha11) + sgn * (*beta11); bl1_sdiv3( &vec, &a11, &x11 ); *ckl = x11; /*------------------------------------------------------------*/ } } return FLA_SUCCESS; }
References FLA_Obj_col_stride(), FLA_Obj_datatype(), FLA_Obj_length(), FLA_Obj_row_stride(), FLA_Obj_width(), FLA_Sylv_nn_opc_var1(), FLA_Sylv_nn_opd_var1(), FLA_Sylv_nn_ops_var1(), and FLA_Sylv_nn_opz_var1().
Referenced by FLA_Sylv_nn(), FLA_Sylv_nn_opt_var10(), FLA_Sylv_nn_opt_var11(), FLA_Sylv_nn_opt_var12(), FLA_Sylv_nn_opt_var13(), FLA_Sylv_nn_opt_var14(), FLA_Sylv_nn_opt_var15(), FLA_Sylv_nn_opt_var16(), FLA_Sylv_nn_opt_var17(), FLA_Sylv_nn_opt_var18(), FLA_Sylv_nn_opt_var2(), FLA_Sylv_nn_opt_var3(), FLA_Sylv_nn_opt_var4(), FLA_Sylv_nn_opt_var5(), FLA_Sylv_nn_opt_var6(), FLA_Sylv_nn_opt_var7(), FLA_Sylv_nn_opt_var8(), and FLA_Sylv_nn_opt_var9().
{ FLA_Datatype datatype; int m_C, n_C; int rs_A, cs_A; int rs_B, cs_B; int rs_C, cs_C; int info; datatype = FLA_Obj_datatype( A ); rs_A = FLA_Obj_row_stride( A ); cs_A = FLA_Obj_col_stride( A ); rs_B = FLA_Obj_row_stride( B ); cs_B = FLA_Obj_col_stride( B ); m_C = FLA_Obj_length( C ); n_C = FLA_Obj_width( C ); rs_C = FLA_Obj_row_stride( C ); cs_C = FLA_Obj_col_stride( C ); switch ( datatype ) { case FLA_FLOAT: { int* buff_isgn = FLA_INT_PTR( isgn ); float* buff_A = FLA_FLOAT_PTR( A ); float* buff_B = FLA_FLOAT_PTR( B ); float* buff_C = FLA_FLOAT_PTR( C ); float* buff_scale = FLA_FLOAT_PTR( scale ); float sgn = ( float ) *buff_isgn; FLA_Sylv_nn_ops_var1( sgn, m_C, n_C, buff_A, rs_A, cs_A, buff_B, rs_B, cs_B, buff_C, rs_C, cs_C, buff_scale, &info ); break; } case FLA_DOUBLE: { int* buff_isgn = FLA_INT_PTR( isgn ); double* buff_A = FLA_DOUBLE_PTR( A ); double* buff_B = FLA_DOUBLE_PTR( B ); double* buff_C = FLA_DOUBLE_PTR( C ); double* buff_scale = FLA_DOUBLE_PTR( scale ); double sgn = ( double ) *buff_isgn; FLA_Sylv_nn_opd_var1( sgn, m_C, n_C, buff_A, rs_A, cs_A, buff_B, rs_B, cs_B, buff_C, rs_C, cs_C, buff_scale, &info ); break; } case FLA_COMPLEX: { int* buff_isgn = FLA_INT_PTR( isgn ); scomplex* buff_A = FLA_COMPLEX_PTR( A ); scomplex* buff_B = FLA_COMPLEX_PTR( B ); scomplex* buff_C = FLA_COMPLEX_PTR( C ); scomplex* buff_scale = FLA_COMPLEX_PTR( scale ); float sgn = ( float ) *buff_isgn; FLA_Sylv_nn_opc_var1( sgn, m_C, n_C, buff_A, rs_A, cs_A, buff_B, rs_B, cs_B, buff_C, rs_C, cs_C, buff_scale, &info ); break; } case FLA_DOUBLE_COMPLEX: { int* buff_isgn = FLA_INT_PTR( isgn ); dcomplex* buff_A = FLA_DOUBLE_COMPLEX_PTR( A ); dcomplex* buff_B = FLA_DOUBLE_COMPLEX_PTR( B ); dcomplex* buff_C = FLA_DOUBLE_COMPLEX_PTR( C ); dcomplex* buff_scale = FLA_DOUBLE_COMPLEX_PTR( scale ); double sgn = ( double ) *buff_isgn; FLA_Sylv_nn_opz_var1( sgn, m_C, n_C, buff_A, rs_A, cs_A, buff_B, rs_B, cs_B, buff_C, rs_C, cs_C, buff_scale, &info ); break; } } return FLA_SUCCESS; }
References FLA_Sylv_nn_opt_var1().
{ return FLA_Sylv_nn_opt_var1( isgn, A, B, C, scale ); }
References FLA_Sylv_nn_opt_var1().
{ return FLA_Sylv_nn_opt_var1( isgn, A, B, C, scale ); }
References FLA_Sylv_nn_opt_var1().
{ return FLA_Sylv_nn_opt_var1( isgn, A, B, C, scale ); }
References FLA_Sylv_nn_opt_var1().
{ return FLA_Sylv_nn_opt_var1( isgn, A, B, C, scale ); }
References FLA_Sylv_nn_opt_var1().
{ return FLA_Sylv_nn_opt_var1( isgn, A, B, C, scale ); }
References FLA_Sylv_nn_opt_var1().
{ return FLA_Sylv_nn_opt_var1( isgn, A, B, C, scale ); }
References FLA_Sylv_nn_opt_var1().
{ return FLA_Sylv_nn_opt_var1( isgn, A, B, C, scale ); }
References FLA_Sylv_nn_opt_var1().
{ return FLA_Sylv_nn_opt_var1( isgn, A, B, C, scale ); }
References FLA_Sylv_nn_opt_var1().
{ return FLA_Sylv_nn_opt_var1( isgn, A, B, C, scale ); }
References FLA_Sylv_nn_opt_var1().
{ return FLA_Sylv_nn_opt_var1( isgn, A, B, C, scale ); }
References FLA_Sylv_nn_opt_var1().
{ return FLA_Sylv_nn_opt_var1( isgn, A, B, C, scale ); }
References FLA_Sylv_nn_opt_var1().
{ return FLA_Sylv_nn_opt_var1( isgn, A, B, C, scale ); }
References FLA_Sylv_nn_opt_var1().
{ return FLA_Sylv_nn_opt_var1( isgn, A, B, C, scale ); }
References FLA_Sylv_nn_opt_var1().
{ return FLA_Sylv_nn_opt_var1( isgn, A, B, C, scale ); }
References FLA_Sylv_nn_opt_var1().
{ return FLA_Sylv_nn_opt_var1( isgn, A, B, C, scale ); }
References FLA_Sylv_nn_opt_var1().
{ return FLA_Sylv_nn_opt_var1( isgn, A, B, C, scale ); }
References FLA_Sylv_nn_opt_var1().
{ return FLA_Sylv_nn_opt_var1( isgn, A, B, C, scale ); }
FLA_Error FLA_Sylv_nn_opz_var1 | ( | double | sgn, |
int | m_C, | ||
int | n_C, | ||
dcomplex * | buff_A, | ||
int | rs_A, | ||
int | cs_A, | ||
dcomplex * | buff_B, | ||
int | rs_B, | ||
int | cs_B, | ||
dcomplex * | buff_C, | ||
int | rs_C, | ||
int | cs_C, | ||
dcomplex * | buff_scale, | ||
int * | info | ||
) |
References bl1_zdot(), BLIS1_NO_CONJUGATE, dcomplex::imag, and dcomplex::real.
Referenced by FLA_Sylv_nn_opt_var1().
{ int l, k; for ( l = 0; l < n_C; l++ ) { for ( k = m_C - 1; k >= 0; k-- ) { dcomplex* a12t = buff_A + (k+1)*cs_A + (k )*rs_A; dcomplex* b01 = buff_B + (l )*cs_B + (0 )*rs_B; dcomplex* c10t = buff_C + (0 )*cs_C + (k )*rs_C; dcomplex* c21 = buff_C + (l )*cs_C + (k+1)*rs_C; dcomplex* alpha11 = buff_A + (k )*cs_A + (k )*rs_A; dcomplex* beta11 = buff_B + (l )*cs_B + (l )*rs_B; dcomplex* ckl = buff_C + (l )*cs_C + (k )*rs_C; dcomplex suml; dcomplex sumr; dcomplex vec; dcomplex a11; dcomplex x11; int m_behind = m_C - k - 1; int n_behind = l; /*------------------------------------------------------------*/ bl1_zdot( BLIS1_NO_CONJUGATE, m_behind, a12t, cs_A, c21, rs_C, &suml ); bl1_zdot( BLIS1_NO_CONJUGATE, n_behind, c10t, cs_C, b01, rs_B, &sumr ); vec.real = ckl->real - ( suml.real + sgn * sumr.real ); vec.imag = ckl->imag - ( suml.imag + sgn * sumr.imag ); a11.real = alpha11->real + sgn * beta11->real; a11.imag = alpha11->imag + sgn * beta11->imag; bl1_zdiv3( &vec, &a11, &x11 ); *ckl = x11; /*------------------------------------------------------------*/ } } return FLA_SUCCESS; }