libflame
12600
|
Go to the source code of this file.
FLA_Error FLA_Gemm_hc_blk_var1 | ( | FLA_Obj | alpha, |
FLA_Obj | A, | ||
FLA_Obj | B, | ||
FLA_Obj | beta, | ||
FLA_Obj | C, | ||
fla_gemm_t * | cntl | ||
) |
References FLA_Cont_with_1x3_to_1x2(), FLA_Cont_with_3x1_to_2x1(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_Obj_width(), FLA_Part_1x2(), FLA_Part_2x1(), FLA_Repart_1x2_to_1x3(), and FLA_Repart_2x1_to_3x1().
Referenced by FLA_Gemm_hc().
{ FLA_Obj AL, AR, A0, A1, A2; FLA_Obj CT, C0, CB, C1, C2; dim_t b; FLA_Part_1x2( A, &AL, &AR, 0, FLA_LEFT ); FLA_Part_2x1( C, &CT, &CB, 0, FLA_TOP ); while ( FLA_Obj_width( AL ) < FLA_Obj_width( A ) ){ b = FLA_Determine_blocksize( AR, FLA_RIGHT, FLA_Cntl_blocksize( cntl ) ); FLA_Repart_1x2_to_1x3( AL, /**/ AR, &A0, /**/ &A1, &A2, b, FLA_RIGHT ); FLA_Repart_2x1_to_3x1( CT, &C0, /* ** */ /* ** */ &C1, CB, &C2, b, FLA_BOTTOM ); /*------------------------------------------------------------*/ /* C1 = alpha * A1' * B + C1; */ FLA_Gemm_internal( FLA_CONJ_TRANSPOSE, FLA_CONJ_NO_TRANSPOSE, alpha, A1, B, beta, C1, FLA_Cntl_sub_gemm( cntl ) ); /*------------------------------------------------------------*/ FLA_Cont_with_1x3_to_1x2( &AL, /**/ &AR, A0, A1, /**/ A2, FLA_LEFT ); FLA_Cont_with_3x1_to_2x1( &CT, C0, C1, /* ** */ /* ** */ &CB, C2, FLA_TOP ); } return FLA_SUCCESS; }
FLA_Error FLA_Gemm_hc_blk_var2 | ( | FLA_Obj | alpha, |
FLA_Obj | A, | ||
FLA_Obj | B, | ||
FLA_Obj | beta, | ||
FLA_Obj | C, | ||
fla_gemm_t * | cntl | ||
) |
References FLA_Cont_with_1x3_to_1x2(), FLA_Cont_with_3x1_to_2x1(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_Obj_width(), FLA_Part_1x2(), FLA_Part_2x1(), FLA_Repart_1x2_to_1x3(), and FLA_Repart_2x1_to_3x1().
Referenced by FLA_Gemm_hc().
{ FLA_Obj AL, AR, A0, A1, A2; FLA_Obj CT, C0, CB, C1, C2; dim_t b; FLA_Part_1x2( A, &AL, &AR, 0, FLA_RIGHT ); FLA_Part_2x1( C, &CT, &CB, 0, FLA_BOTTOM ); while ( FLA_Obj_width( AR ) < FLA_Obj_width( A ) ){ b = FLA_Determine_blocksize( AL, FLA_LEFT, FLA_Cntl_blocksize( cntl ) ); FLA_Repart_1x2_to_1x3( AL, /**/ AR, &A0, &A1, /**/ &A2, b, FLA_LEFT ); FLA_Repart_2x1_to_3x1( CT, &C0, &C1, /* ** */ /* ** */ CB, &C2, b, FLA_TOP ); /*------------------------------------------------------------*/ /* C1 = alpha * A1' * B + C1; */ FLA_Gemm_internal( FLA_CONJ_TRANSPOSE, FLA_CONJ_NO_TRANSPOSE, alpha, A1, B, beta, C1, FLA_Cntl_sub_gemm( cntl ) ); /*------------------------------------------------------------*/ FLA_Cont_with_1x3_to_1x2( &AL, /**/ &AR, A0, /**/ A1, A2, FLA_RIGHT ); FLA_Cont_with_3x1_to_2x1( &CT, C0, /* ** */ /* ** */ C1, &CB, C2, FLA_BOTTOM ); } return FLA_SUCCESS; }
FLA_Error FLA_Gemm_hc_blk_var3 | ( | FLA_Obj | alpha, |
FLA_Obj | A, | ||
FLA_Obj | B, | ||
FLA_Obj | beta, | ||
FLA_Obj | C, | ||
fla_gemm_t * | cntl | ||
) |
References FLA_Cont_with_1x3_to_1x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_Obj_width(), FLA_Part_1x2(), and FLA_Repart_1x2_to_1x3().
Referenced by FLA_Gemm_hc().
{ 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_LEFT ); FLA_Part_1x2( C, &CL, &CR, 0, FLA_LEFT ); while ( FLA_Obj_width( BL ) < FLA_Obj_width( B ) ){ b = FLA_Determine_blocksize( BR, FLA_RIGHT, FLA_Cntl_blocksize( cntl ) ); FLA_Repart_1x2_to_1x3( BL, /**/ BR, &B0, /**/ &B1, &B2, b, FLA_RIGHT ); FLA_Repart_1x2_to_1x3( CL, /**/ CR, &C0, /**/ &C1, &C2, b, FLA_RIGHT ); /*------------------------------------------------------------*/ /* C1 = alpha * A' * B1 + C1; */ FLA_Gemm_internal( FLA_CONJ_TRANSPOSE, FLA_CONJ_NO_TRANSPOSE, alpha, A, B1, beta, C1, FLA_Cntl_sub_gemm( cntl ) ); /*------------------------------------------------------------*/ FLA_Cont_with_1x3_to_1x2( &BL, /**/ &BR, B0, B1, /**/ B2, FLA_LEFT ); FLA_Cont_with_1x3_to_1x2( &CL, /**/ &CR, C0, C1, /**/ C2, FLA_LEFT ); } return FLA_SUCCESS; }
FLA_Error FLA_Gemm_hc_blk_var4 | ( | FLA_Obj | alpha, |
FLA_Obj | A, | ||
FLA_Obj | B, | ||
FLA_Obj | beta, | ||
FLA_Obj | C, | ||
fla_gemm_t * | cntl | ||
) |
References FLA_Cont_with_1x3_to_1x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_Obj_width(), FLA_Part_1x2(), and FLA_Repart_1x2_to_1x3().
Referenced by FLA_Gemm_hc().
{ 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 = alpha * A' * B1 + C1; */ FLA_Gemm_internal( FLA_CONJ_TRANSPOSE, FLA_CONJ_NO_TRANSPOSE, alpha, A, B1, beta, C1, FLA_Cntl_sub_gemm( 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_Gemm_hc_blk_var5 | ( | FLA_Obj | alpha, |
FLA_Obj | A, | ||
FLA_Obj | B, | ||
FLA_Obj | beta, | ||
FLA_Obj | C, | ||
fla_gemm_t * | cntl | ||
) |
References FLA_Cont_with_3x1_to_2x1(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_Obj_length(), FLA_ONE, FLA_Part_2x1(), FLA_Repart_2x1_to_3x1(), and FLA_Scal_internal().
Referenced by FLA_Gemm_hc().
{ FLA_Obj AT, A0, AB, A1, A2; FLA_Obj BT, B0, BB, B1, B2; dim_t b; FLA_Scal_internal( beta, C, FLA_Cntl_sub_scal( cntl ) ); FLA_Part_2x1( A, &AT, &AB, 0, FLA_TOP ); FLA_Part_2x1( B, &BT, &BB, 0, FLA_TOP ); while ( FLA_Obj_length( AT ) < FLA_Obj_length( A ) ){ b = FLA_Determine_blocksize( AB, FLA_BOTTOM, FLA_Cntl_blocksize( cntl ) ); FLA_Repart_2x1_to_3x1( AT, &A0, /* ** */ /* ** */ &A1, AB, &A2, b, FLA_BOTTOM ); FLA_Repart_2x1_to_3x1( BT, &B0, /* ** */ /* ** */ &B1, BB, &B2, b, FLA_BOTTOM ); /*------------------------------------------------------------*/ /* C = alpha * A1' * B1 + C; */ FLA_Gemm_internal( FLA_CONJ_TRANSPOSE, FLA_CONJ_NO_TRANSPOSE, alpha, A1, B1, FLA_ONE, C, FLA_Cntl_sub_gemm( cntl ) ); /*------------------------------------------------------------*/ FLA_Cont_with_3x1_to_2x1( &AT, A0, A1, /* ** */ /* ** */ &AB, A2, FLA_TOP ); FLA_Cont_with_3x1_to_2x1( &BT, B0, B1, /* ** */ /* ** */ &BB, B2, FLA_TOP ); } return FLA_SUCCESS; }
FLA_Error FLA_Gemm_hc_blk_var6 | ( | FLA_Obj | alpha, |
FLA_Obj | A, | ||
FLA_Obj | B, | ||
FLA_Obj | beta, | ||
FLA_Obj | C, | ||
fla_gemm_t * | cntl | ||
) |
References FLA_Cont_with_3x1_to_2x1(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_Obj_length(), FLA_ONE, FLA_Part_2x1(), FLA_Repart_2x1_to_3x1(), and FLA_Scal_internal().
Referenced by FLA_Gemm_hc().
{ FLA_Obj AT, A0, AB, A1, A2; FLA_Obj BT, B0, BB, B1, B2; dim_t b; FLA_Scal_internal( beta, C, FLA_Cntl_sub_scal( cntl ) ); FLA_Part_2x1( A, &AT, &AB, 0, FLA_BOTTOM ); FLA_Part_2x1( B, &BT, &BB, 0, FLA_BOTTOM ); while ( FLA_Obj_length( AB ) < FLA_Obj_length( A ) ){ b = FLA_Determine_blocksize( AT, FLA_TOP, FLA_Cntl_blocksize( cntl ) ); FLA_Repart_2x1_to_3x1( AT, &A0, &A1, /* ** */ /* ** */ AB, &A2, b, FLA_TOP ); FLA_Repart_2x1_to_3x1( BT, &B0, &B1, /* ** */ /* ** */ BB, &B2, b, FLA_TOP ); /*------------------------------------------------------------*/ /* C = alpha * A1' * B1 + C; */ FLA_Gemm_internal( FLA_CONJ_TRANSPOSE, FLA_CONJ_NO_TRANSPOSE, alpha, A1, B1, FLA_ONE, C, FLA_Cntl_sub_gemm( cntl ) ); /*------------------------------------------------------------*/ FLA_Cont_with_3x1_to_2x1( &AT, A0, /* ** */ /* ** */ A1, &AB, A2, FLA_BOTTOM ); FLA_Cont_with_3x1_to_2x1( &BT, B0, /* ** */ /* ** */ B1, &BB, B2, FLA_BOTTOM ); } return FLA_SUCCESS; }
References FLA_Cont_with_1x3_to_1x2(), FLA_Cont_with_3x1_to_2x1(), FLA_Gemvc_external(), FLA_Obj_width(), FLA_ONE, FLA_Part_1x2(), FLA_Part_2x1(), FLA_Repart_1x2_to_1x3(), FLA_Repart_2x1_to_3x1(), and FLA_Scal_external().
Referenced by FLA_Gemm_hc().
{ FLA_Obj AL, AR, A0, a1, A2; FLA_Obj CT, C0, CB, c1t, C2; FLA_Scal_external( beta, C ); FLA_Part_1x2( A, &AL, &AR, 0, FLA_LEFT ); FLA_Part_2x1( C, &CT, &CB, 0, FLA_TOP ); while ( FLA_Obj_width( AL ) < FLA_Obj_width( A ) ){ FLA_Repart_1x2_to_1x3( AL, /**/ AR, &A0, /**/ &a1, &A2, 1, FLA_RIGHT ); FLA_Repart_2x1_to_3x1( CT, &C0, /* ** */ /* *** */ &c1t, CB, &C2, 1, FLA_BOTTOM ); /*------------------------------------------------------------*/ /* c1t = a1' * B + c1t */ /* c1t' = B' * a1 + c1t' */ FLA_Gemvc_external( FLA_CONJ_TRANSPOSE, FLA_CONJUGATE, alpha, B, a1, FLA_ONE, c1t ); /*------------------------------------------------------------*/ FLA_Cont_with_1x3_to_1x2( &AL, /**/ &AR, A0, a1, /**/ A2, FLA_LEFT ); FLA_Cont_with_3x1_to_2x1( &CT, C0, c1t, /* ** */ /* *** */ &CB, C2, FLA_TOP ); } return FLA_SUCCESS; }
References FLA_Cont_with_1x3_to_1x2(), FLA_Cont_with_3x1_to_2x1(), FLA_Gemvc_external(), FLA_Obj_width(), FLA_ONE, FLA_Part_1x2(), FLA_Part_2x1(), FLA_Repart_1x2_to_1x3(), FLA_Repart_2x1_to_3x1(), and FLA_Scal_external().
Referenced by FLA_Gemm_hc().
{ FLA_Obj AL, AR, A0, a1, A2; FLA_Obj CT, C0, CB, c1t, C2; FLA_Scal_external( beta, C ); FLA_Part_1x2( A, &AL, &AR, 0, FLA_RIGHT ); FLA_Part_2x1( C, &CT, &CB, 0, FLA_BOTTOM ); while ( FLA_Obj_width( AR ) < FLA_Obj_width( A ) ){ FLA_Repart_1x2_to_1x3( AL, /**/ AR, &A0, &a1, /**/ &A2, 1, FLA_LEFT ); FLA_Repart_2x1_to_3x1( CT, &C0, &c1t, /* ** */ /* *** */ CB, &C2, 1, FLA_TOP ); /*------------------------------------------------------------*/ /* c1t = a1' * B + c1t */ /* c1t' = B' * a1 + c1t' */ FLA_Gemvc_external( FLA_CONJ_TRANSPOSE, FLA_CONJUGATE, alpha, B, a1, FLA_ONE, c1t ); /*------------------------------------------------------------*/ FLA_Cont_with_1x3_to_1x2( &AL, /**/ &AR, A0, /**/ a1, A2, FLA_RIGHT ); FLA_Cont_with_3x1_to_2x1( &CT, C0, /* ** */ /* *** */ c1t, &CB, C2, FLA_BOTTOM ); } return FLA_SUCCESS; }
References FLA_Cont_with_1x3_to_1x2(), FLA_Gemvc_external(), FLA_Obj_width(), FLA_ONE, FLA_Part_1x2(), FLA_Repart_1x2_to_1x3(), and FLA_Scal_external().
Referenced by FLA_Gemm_hc().
{ FLA_Obj BL, BR, B0, b1, B2; FLA_Obj CL, CR, C0, c1, C2; FLA_Scal_external( beta, C ); FLA_Part_1x2( B, &BL, &BR, 0, FLA_LEFT ); FLA_Part_1x2( C, &CL, &CR, 0, FLA_LEFT ); while ( FLA_Obj_width( BL ) < FLA_Obj_width( B ) ){ FLA_Repart_1x2_to_1x3( BL, /**/ BR, &B0, /**/ &b1, &B2, 1, FLA_RIGHT ); FLA_Repart_1x2_to_1x3( CL, /**/ CR, &C0, /**/ &c1, &C2, 1, FLA_RIGHT ); /*------------------------------------------------------------*/ /* c1 = A' * b1 + c1 */ FLA_Gemvc_external( FLA_CONJ_TRANSPOSE, FLA_CONJUGATE, alpha, A, b1, FLA_ONE, c1 ); /*------------------------------------------------------------*/ FLA_Cont_with_1x3_to_1x2( &BL, /**/ &BR, B0, b1, /**/ B2, FLA_LEFT ); FLA_Cont_with_1x3_to_1x2( &CL, /**/ &CR, C0, c1, /**/ C2, FLA_LEFT ); } return FLA_SUCCESS; }
References FLA_Cont_with_1x3_to_1x2(), FLA_Gemvc_external(), FLA_Obj_width(), FLA_ONE, FLA_Part_1x2(), FLA_Repart_1x2_to_1x3(), and FLA_Scal_external().
Referenced by FLA_Gemm_hc().
{ FLA_Obj BL, BR, B0, b1, B2; FLA_Obj CL, CR, C0, c1, C2; FLA_Scal_external( beta, C ); 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 ) ){ FLA_Repart_1x2_to_1x3( BL, /**/ BR, &B0, &b1, /**/ &B2, 1, FLA_LEFT ); FLA_Repart_1x2_to_1x3( CL, /**/ CR, &C0, &c1, /**/ &C2, 1, FLA_LEFT ); /*------------------------------------------------------------*/ /* c1 = A' * b1 + c1 */ FLA_Gemvc_external( FLA_CONJ_TRANSPOSE, FLA_CONJUGATE, alpha, A, b1, FLA_ONE, c1 ); /*------------------------------------------------------------*/ 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; }
References FLA_Cont_with_3x1_to_2x1(), FLA_Gerc_external(), FLA_Obj_length(), FLA_Part_2x1(), FLA_Repart_2x1_to_3x1(), and FLA_Scal_external().
Referenced by FLA_Gemm_hc().
{ FLA_Obj AT, A0, AB, a1t, A2; FLA_Obj BT, B0, BB, b1t, B2; FLA_Scal_external( beta, C ); FLA_Part_2x1( A, &AT, &AB, 0, FLA_TOP ); FLA_Part_2x1( B, &BT, &BB, 0, FLA_TOP ); while ( FLA_Obj_length( AT ) < FLA_Obj_length( A ) ){ FLA_Repart_2x1_to_3x1( AT, &A0, /* ** */ /* *** */ &a1t, AB, &A2, 1, FLA_BOTTOM ); FLA_Repart_2x1_to_3x1( BT, &B0, /* ** */ /* *** */ &b1t, BB, &B2, 1, FLA_BOTTOM ); /*------------------------------------------------------------*/ /* C = a1t' * b1t + C */ FLA_Gerc_external( FLA_CONJUGATE, FLA_CONJUGATE, alpha, a1t, b1t, C ); /*------------------------------------------------------------*/ FLA_Cont_with_3x1_to_2x1( &AT, A0, a1t, /* ** */ /* *** */ &AB, A2, FLA_TOP ); FLA_Cont_with_3x1_to_2x1( &BT, B0, b1t, /* ** */ /* *** */ &BB, B2, FLA_TOP ); } return FLA_SUCCESS; }
References FLA_Cont_with_3x1_to_2x1(), FLA_Gerc_external(), FLA_Obj_length(), FLA_Part_2x1(), FLA_Repart_2x1_to_3x1(), and FLA_Scal_external().
Referenced by FLA_Gemm_hc().
{ FLA_Obj AT, A0, AB, a1t, A2; FLA_Obj BT, B0, BB, b1t, B2; FLA_Scal_external( beta, C ); FLA_Part_2x1( A, &AT, &AB, 0, FLA_BOTTOM ); FLA_Part_2x1( B, &BT, &BB, 0, FLA_BOTTOM ); while ( FLA_Obj_length( AB ) < FLA_Obj_length( A ) ){ FLA_Repart_2x1_to_3x1( AT, &A0, &a1t, /* ** */ /* *** */ AB, &A2, 1, FLA_TOP ); FLA_Repart_2x1_to_3x1( BT, &B0, &b1t, /* ** */ /* *** */ BB, &B2, 1, FLA_TOP ); /*------------------------------------------------------------*/ /* C = a1t' * b1t + C */ FLA_Gerc_external( FLA_CONJUGATE, FLA_CONJUGATE, alpha, a1t, b1t, C ); /*------------------------------------------------------------*/ FLA_Cont_with_3x1_to_2x1( &AT, A0, /* ** */ /* *** */ a1t, &AB, A2, FLA_BOTTOM ); FLA_Cont_with_3x1_to_2x1( &BT, B0, /* ** */ /* *** */ b1t, &BB, B2, FLA_BOTTOM ); } return FLA_SUCCESS; }