libflame
12600
|
Go to the source code of this file.
FLA_Error FLA_Hemm_lu_blk_var1 | ( | FLA_Obj | alpha, |
FLA_Obj | A, | ||
FLA_Obj | B, | ||
FLA_Obj | beta, | ||
FLA_Obj | C, | ||
fla_hemm_t * | cntl | ||
) |
References FLA_Cont_with_3x1_to_2x1(), FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_Hemm_internal(), FLA_Obj_length(), FLA_ONE, FLA_Part_2x1(), FLA_Part_2x2(), FLA_Repart_2x1_to_3x1(), FLA_Repart_2x2_to_3x3(), and FLA_Scal_internal().
Referenced by FLA_Hemm_lu().
{ FLA_Obj ATL, ATR, A00, A01, A02, ABL, ABR, A10, A11, A12, A20, A21, A22; FLA_Obj BT, B0, BB, B1, B2; FLA_Obj CT, C0, CB, C1, C2; dim_t b; FLA_Scal_internal( beta, C, FLA_Cntl_sub_scal( cntl ) ); FLA_Part_2x2( A, &ATL, &ATR, &ABL, &ABR, 0, 0, FLA_TL ); FLA_Part_2x1( B, &BT, &BB, 0, FLA_TOP ); FLA_Part_2x1( C, &CT, &CB, 0, FLA_TOP ); while ( FLA_Obj_length( ATL ) < FLA_Obj_length( A ) ){ b = FLA_Determine_blocksize( ABR, FLA_BR, 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_BR ); FLA_Repart_2x1_to_3x1( BT, &B0, /* ** */ /* ** */ &B1, BB, &B2, b, FLA_BOTTOM ); FLA_Repart_2x1_to_3x1( CT, &C0, /* ** */ /* ** */ &C1, CB, &C2, b, FLA_BOTTOM ); /*------------------------------------------------------------*/ /* C0 = C0 + A01 * B1 */ FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, alpha, A01, B1, FLA_ONE, C0, FLA_Cntl_sub_gemm1( cntl ) ); /* C1 = C1 + A11 * B1 */ FLA_Hemm_internal( FLA_LEFT, FLA_UPPER_TRIANGULAR, alpha, A11, B1, FLA_ONE, C1, FLA_Cntl_sub_hemm( cntl ) ); /* C1 = C1 + A01' * B0 */ FLA_Gemm_internal( FLA_CONJ_TRANSPOSE, FLA_NO_TRANSPOSE, alpha, A01, B0, FLA_ONE, C1, FLA_Cntl_sub_gemm2( cntl ) ); /*------------------------------------------------------------*/ FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, A01, /**/ A02, A10, A11, /**/ A12, /* ************** */ /* ****************** */ &ABL, /**/ &ABR, A20, A21, /**/ A22, FLA_TL ); FLA_Cont_with_3x1_to_2x1( &BT, B0, B1, /* ** */ /* ** */ &BB, B2, FLA_TOP ); FLA_Cont_with_3x1_to_2x1( &CT, C0, C1, /* ** */ /* ** */ &CB, C2, FLA_TOP ); } return FLA_SUCCESS; }
FLA_Error FLA_Hemm_lu_blk_var10 | ( | FLA_Obj | alpha, |
FLA_Obj | A, | ||
FLA_Obj | B, | ||
FLA_Obj | beta, | ||
FLA_Obj | C, | ||
fla_hemm_t * | cntl | ||
) |
References FLA_Cont_with_1x3_to_1x2(), FLA_Determine_blocksize(), FLA_Hemm_internal(), FLA_Obj_width(), FLA_Part_1x2(), and FLA_Repart_1x2_to_1x3().
Referenced by FLA_Hemm_lu().
{ FLA_Obj BL, BR, B0, B1, B2; FLA_Obj CL, CR, C0, C1, C2; dim_t b; FLA_Part_1x2( B, &BL, &BR, 0, FLA_RIGHT ); FLA_Part_1x2( C, &CL, &CR, 0, FLA_RIGHT ); while ( FLA_Obj_width( BR ) < FLA_Obj_width( B ) ){ b = FLA_Determine_blocksize( BL, FLA_LEFT, FLA_Cntl_blocksize( cntl ) ); FLA_Repart_1x2_to_1x3( BL, /**/ BR, &B0, &B1, /**/ &B2, b, FLA_LEFT ); FLA_Repart_1x2_to_1x3( CL, /**/ CR, &C0, &C1, /**/ &C2, b, FLA_LEFT ); /*------------------------------------------------------------*/ /* C1 = C1 + A * B1 */ FLA_Hemm_internal( FLA_LEFT, FLA_UPPER_TRIANGULAR, alpha, A, B1, beta, C1, FLA_Cntl_sub_hemm( cntl ) ); /*------------------------------------------------------------*/ FLA_Cont_with_1x3_to_1x2( &BL, /**/ &BR, B0, /**/ B1, B2, FLA_RIGHT ); FLA_Cont_with_1x3_to_1x2( &CL, /**/ &CR, C0, /**/ C1, C2, FLA_RIGHT ); } return FLA_SUCCESS; }
FLA_Error FLA_Hemm_lu_blk_var2 | ( | FLA_Obj | alpha, |
FLA_Obj | A, | ||
FLA_Obj | B, | ||
FLA_Obj | beta, | ||
FLA_Obj | C, | ||
fla_hemm_t * | cntl | ||
) |
References FLA_Cont_with_3x1_to_2x1(), FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_Hemm_internal(), FLA_Obj_length(), FLA_ONE, FLA_Part_2x1(), FLA_Part_2x2(), FLA_Repart_2x1_to_3x1(), FLA_Repart_2x2_to_3x3(), and FLA_Scal_internal().
Referenced by FLA_Hemm_lu().
{ FLA_Obj ATL, ATR, A00, A01, A02, ABL, ABR, A10, A11, A12, A20, A21, A22; FLA_Obj BT, B0, BB, B1, B2; FLA_Obj CT, C0, CB, C1, C2; dim_t b; FLA_Scal_internal( beta, C, FLA_Cntl_sub_scal( cntl ) ); FLA_Part_2x2( A, &ATL, &ATR, &ABL, &ABR, 0, 0, FLA_TL ); FLA_Part_2x1( B, &BT, &BB, 0, FLA_TOP ); FLA_Part_2x1( C, &CT, &CB, 0, FLA_TOP ); while ( FLA_Obj_length( ATL ) < FLA_Obj_length( A ) ){ b = FLA_Determine_blocksize( ABR, FLA_BR, 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_BR ); FLA_Repart_2x1_to_3x1( BT, &B0, /* ** */ /* ** */ &B1, BB, &B2, b, FLA_BOTTOM ); FLA_Repart_2x1_to_3x1( CT, &C0, /* ** */ /* ** */ &C1, CB, &C2, b, FLA_BOTTOM ); /*------------------------------------------------------------*/ /* C1 = C1 + A01' * B0 */ FLA_Gemm_internal( FLA_CONJ_TRANSPOSE, FLA_NO_TRANSPOSE, alpha, A01, B0, FLA_ONE, C1, FLA_Cntl_sub_gemm1( cntl ) ); /* C1 = C1 + A12 * B2 */ FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, alpha, A12, B2, FLA_ONE, C1, FLA_Cntl_sub_gemm2( cntl ) ); /* C1 = C1 + A11 * B1 */ FLA_Hemm_internal( FLA_LEFT, FLA_UPPER_TRIANGULAR, alpha, A11, B1, FLA_ONE, C1, FLA_Cntl_sub_hemm( cntl ) ); /*------------------------------------------------------------*/ FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, A01, /**/ A02, A10, A11, /**/ A12, /* ************** */ /* ****************** */ &ABL, /**/ &ABR, A20, A21, /**/ A22, FLA_TL ); FLA_Cont_with_3x1_to_2x1( &BT, B0, B1, /* ** */ /* ** */ &BB, B2, FLA_TOP ); FLA_Cont_with_3x1_to_2x1( &CT, C0, C1, /* ** */ /* ** */ &CB, C2, FLA_TOP ); } return FLA_SUCCESS; }
FLA_Error FLA_Hemm_lu_blk_var3 | ( | FLA_Obj | alpha, |
FLA_Obj | A, | ||
FLA_Obj | B, | ||
FLA_Obj | beta, | ||
FLA_Obj | C, | ||
fla_hemm_t * | cntl | ||
) |
References FLA_Cont_with_3x1_to_2x1(), FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_Hemm_internal(), FLA_Obj_length(), FLA_ONE, FLA_Part_2x1(), FLA_Part_2x2(), FLA_Repart_2x1_to_3x1(), FLA_Repart_2x2_to_3x3(), and FLA_Scal_internal().
Referenced by FLA_Hemm_lu().
{ FLA_Obj ATL, ATR, A00, A01, A02, ABL, ABR, A10, A11, A12, A20, A21, A22; FLA_Obj BT, B0, BB, B1, B2; FLA_Obj CT, C0, CB, C1, C2; dim_t b; FLA_Scal_internal( beta, C, FLA_Cntl_sub_scal( cntl ) ); FLA_Part_2x2( A, &ATL, &ATR, &ABL, &ABR, 0, 0, FLA_TL ); FLA_Part_2x1( B, &BT, &BB, 0, FLA_TOP ); FLA_Part_2x1( C, &CT, &CB, 0, FLA_TOP ); while ( FLA_Obj_length( ATL ) < FLA_Obj_length( A ) ){ b = FLA_Determine_blocksize( ABR, FLA_BR, 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_BR ); FLA_Repart_2x1_to_3x1( BT, &B0, /* ** */ /* ** */ &B1, BB, &B2, b, FLA_BOTTOM ); FLA_Repart_2x1_to_3x1( CT, &C0, /* ** */ /* ** */ &C1, CB, &C2, b, FLA_BOTTOM ); /*------------------------------------------------------------*/ /* C1 = C1 + A11 * B1 */ FLA_Hemm_internal( FLA_LEFT, FLA_UPPER_TRIANGULAR, alpha, A11, B1, FLA_ONE, C1, FLA_Cntl_sub_hemm( cntl ) ); /* C1 = C1 + A12 * B2 */ FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, alpha, A12, B2, FLA_ONE, C1, FLA_Cntl_sub_gemm1( cntl ) ); /* C2 = C2 + A12' * B1 */ FLA_Gemm_internal( FLA_CONJ_TRANSPOSE, FLA_NO_TRANSPOSE, alpha, A12, B1, FLA_ONE, C2, FLA_Cntl_sub_gemm2( cntl ) ); /*------------------------------------------------------------*/ FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, A01, /**/ A02, A10, A11, /**/ A12, /* ************** */ /* ****************** */ &ABL, /**/ &ABR, A20, A21, /**/ A22, FLA_TL ); FLA_Cont_with_3x1_to_2x1( &BT, B0, B1, /* ** */ /* ** */ &BB, B2, FLA_TOP ); FLA_Cont_with_3x1_to_2x1( &CT, C0, C1, /* ** */ /* ** */ &CB, C2, FLA_TOP ); } return FLA_SUCCESS; }
FLA_Error FLA_Hemm_lu_blk_var4 | ( | FLA_Obj | alpha, |
FLA_Obj | A, | ||
FLA_Obj | B, | ||
FLA_Obj | beta, | ||
FLA_Obj | C, | ||
fla_hemm_t * | cntl | ||
) |
References FLA_Cont_with_3x1_to_2x1(), FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_Hemm_internal(), FLA_Obj_length(), FLA_ONE, FLA_Part_2x1(), FLA_Part_2x2(), FLA_Repart_2x1_to_3x1(), FLA_Repart_2x2_to_3x3(), and FLA_Scal_internal().
Referenced by FLA_Hemm_lu().
{ FLA_Obj ATL, ATR, A00, A01, A02, ABL, ABR, A10, A11, A12, A20, A21, A22; FLA_Obj BT, B0, BB, B1, B2; FLA_Obj CT, C0, CB, C1, C2; dim_t b; FLA_Scal_internal( beta, C, FLA_Cntl_sub_scal( cntl ) ); FLA_Part_2x2( A, &ATL, &ATR, &ABL, &ABR, 0, 0, FLA_TL ); FLA_Part_2x1( B, &BT, &BB, 0, FLA_TOP ); FLA_Part_2x1( C, &CT, &CB, 0, FLA_TOP ); while ( FLA_Obj_length( ATL ) < FLA_Obj_length( A ) ){ b = FLA_Determine_blocksize( ABR, FLA_BR, 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_BR ); FLA_Repart_2x1_to_3x1( BT, &B0, /* ** */ /* ** */ &B1, BB, &B2, b, FLA_BOTTOM ); FLA_Repart_2x1_to_3x1( CT, &C0, /* ** */ /* ** */ &C1, CB, &C2, b, FLA_BOTTOM ); /*------------------------------------------------------------*/ /* C0 = C0 + A01 * B1 */ FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, alpha, A01, B1, FLA_ONE, C0, FLA_Cntl_sub_gemm1( cntl ) ); /* C1 = C1 + A11 * B1 */ FLA_Hemm_internal( FLA_LEFT, FLA_UPPER_TRIANGULAR, alpha, A11, B1, FLA_ONE, C1, FLA_Cntl_sub_hemm( cntl ) ); /* C2 = C2 + A12' * B1 */ FLA_Gemm_internal( FLA_CONJ_TRANSPOSE, FLA_NO_TRANSPOSE, alpha, A12, B1, FLA_ONE, C2, FLA_Cntl_sub_gemm2( cntl ) ); /*------------------------------------------------------------*/ FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, A01, /**/ A02, A10, A11, /**/ A12, /* ************** */ /* ****************** */ &ABL, /**/ &ABR, A20, A21, /**/ A22, FLA_TL ); FLA_Cont_with_3x1_to_2x1( &BT, B0, B1, /* ** */ /* ** */ &BB, B2, FLA_TOP ); FLA_Cont_with_3x1_to_2x1( &CT, C0, C1, /* ** */ /* ** */ &CB, C2, FLA_TOP ); } return FLA_SUCCESS; }
FLA_Error FLA_Hemm_lu_blk_var5 | ( | FLA_Obj | alpha, |