Functions | |
FLA_Error | FLA_Trmm_ruh_blk_var2 (FLA_Diag diagA, FLA_Obj alpha, FLA_Obj A, FLA_Obj B, fla_trmm_t *cntl) |
FLA_Error FLA_Trmm_ruh_blk_var2 | ( | FLA_Diag | diagA, | |
FLA_Obj | alpha, | |||
FLA_Obj | A, | |||
FLA_Obj | B, | |||
fla_trmm_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_Trmm_internal().
Referenced by FLA_Trmm_ruh().
00036 { 00037 FLA_Obj ATL, ATR, A00, A01, A02, 00038 ABL, ABR, A10, A11, A12, 00039 A20, A21, A22; 00040 00041 FLA_Obj BL, BR, B0, B1, B2; 00042 00043 dim_t b; 00044 00045 FLA_Part_2x2( A, &ATL, &ATR, 00046 &ABL, &ABR, 0, 0, FLA_TL ); 00047 00048 FLA_Part_1x2( B, &BL, &BR, 0, FLA_LEFT ); 00049 00050 while ( FLA_Obj_length( ATL ) < FLA_Obj_length( A ) ){ 00051 00052 b = FLA_Determine_blocksize( ABR, FLA_BR, FLA_Cntl_blocksize( cntl ) ); 00053 00054 FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, /**/ &A01, &A02, 00055 /* ************* */ /* ******************** */ 00056 &A10, /**/ &A11, &A12, 00057 ABL, /**/ ABR, &A20, /**/ &A21, &A22, 00058 b, b, FLA_BR ); 00059 00060 FLA_Repart_1x2_to_1x3( BL, /**/ BR, &B0, /**/ &B1, &B2, 00061 b, FLA_RIGHT ); 00062 00063 /*------------------------------------------------------------*/ 00064 00065 /* B0 = B0 + B1 * A01'; */ 00066 FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_CONJ_TRANSPOSE, 00067 FLA_ONE, B1, A01, FLA_ONE, B0, 00068 FLA_Cntl_sub_gemm( cntl ) ); 00069 00070 /* B1 = B1 * triu( A11 )'; */ 00071 FLA_Trmm_internal( FLA_RIGHT, FLA_UPPER_TRIANGULAR, FLA_CONJ_TRANSPOSE, diagA, 00072 alpha, A11, B1, 00073 FLA_Cntl_sub_trmm( cntl ) ); 00074 00075 /*------------------------------------------------------------*/ 00076 00077 FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, A01, /**/ A02, 00078 A10, A11, /**/ A12, 00079 /* ************** */ /* ****************** */ 00080 &ABL, /**/ &ABR, A20, A21, /**/ A22, 00081 FLA_TL ); 00082 00083 FLA_Cont_with_1x3_to_1x2( &BL, /**/ &BR, B0, B1, /**/ B2, 00084 FLA_LEFT ); 00085 00086 } 00087 00088 return FLA_SUCCESS; 00089 }