Functions | |
FLA_Error | FLA_Ttmm_l_blk_var2 (FLA_Obj A, fla_ttmm_t *cntl) |
FLA_Error FLA_Ttmm_l_blk_var2 | ( | FLA_Obj | A, | |
fla_ttmm_t * | cntl | |||
) |
References FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_Herk_internal(), FLA_Obj_is_real(), FLA_Obj_length(), FLA_ONE, FLA_Part_2x2(), FLA_Repart_2x2_to_3x3(), FLA_Syrk_internal(), FLA_Trmm_internal(), and FLA_Ttmm_internal().
Referenced by FLA_Ttmm_l().
00038 { 00039 FLA_Obj ATL, ATR, A00, A01, A02, 00040 ABL, ABR, A10, A11, A12, 00041 A20, A21, A22; 00042 00043 dim_t b; 00044 00045 FLA_Part_2x2( A, &ATL, &ATR, 00046 &ABL, &ABR, 0, 0, FLA_TL ); 00047 00048 while ( FLA_Obj_length( ATL ) < FLA_Obj_length( A ) ){ 00049 00050 b = FLA_Determine_blocksize( ABR, FLA_BR, FLA_Cntl_blocksize( cntl ) ); 00051 00052 FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, /**/ &A01, &A02, 00053 /* ************* */ /* ******************** */ 00054 &A10, /**/ &A11, &A12, 00055 ABL, /**/ ABR, &A20, /**/ &A21, &A22, 00056 b, b, FLA_BR ); 00057 00058 /*------------------------------------------------------------*/ 00059 00060 if ( FLA_Obj_is_real( A ) ) 00061 { 00062 /* A10 = tril( A11 )' * A10; */ 00063 FLA_Trmm_internal( FLA_LEFT, FLA_LOWER_TRIANGULAR, FLA_TRANSPOSE, FLA_NONUNIT_DIAG, 00064 FLA_ONE, A11, A10, 00065 FLA_Cntl_sub_trmm( cntl ) ); 00066 00067 /* A10 = A10 + A21' * A20; */ 00068 FLA_Gemm_internal( FLA_TRANSPOSE, FLA_NO_TRANSPOSE, 00069 FLA_ONE, A21, A20, FLA_ONE, A10, 00070 FLA_Cntl_sub_gemm( cntl ) ); 00071 00072 /* A11 = tril( A11 )' * tril( A11 ); */ 00073 FLA_Ttmm_internal( FLA_LOWER_TRIANGULAR, A11, 00074 FLA_Cntl_sub_ttmm( cntl ) ); 00075 00076 /* A11 = A11 + A21' * A21; */ 00077 FLA_Syrk_internal( FLA_LOWER_TRIANGULAR, FLA_TRANSPOSE, 00078 FLA_ONE, A21, FLA_ONE, A11, 00079 FLA_Cntl_sub_syrk( cntl ) ); 00080 } 00081 else 00082 { 00083 /* A10 = tril( A11 )' * A10; */ 00084 FLA_Trmm_internal( FLA_LEFT, FLA_LOWER_TRIANGULAR, FLA_CONJ_TRANSPOSE, FLA_NONUNIT_DIAG, 00085 FLA_ONE, A11, A10, 00086 FLA_Cntl_sub_trmm( cntl ) ); 00087 00088 /* A10 = A10 + A21' * A20; */ 00089 FLA_Gemm_internal( FLA_CONJ_TRANSPOSE, FLA_NO_TRANSPOSE, 00090 FLA_ONE, A21, A20, FLA_ONE, A10, 00091 FLA_Cntl_sub_gemm( cntl ) ); 00092 00093 /* A11 = tril( A11 )' * tril( A11 ); */ 00094 FLA_Ttmm_internal( FLA_LOWER_TRIANGULAR, A11, 00095 FLA_Cntl_sub_ttmm( cntl ) ); 00096 00097 /* A11 = A11 + A21' * A21; */ 00098 FLA_Herk_internal( FLA_LOWER_TRIANGULAR, FLA_CONJ_TRANSPOSE, 00099 FLA_ONE, A21, FLA_ONE, A11, 00100 FLA_Cntl_sub_herk( cntl ) ); 00101 } 00102 00103 /*------------------------------------------------------------*/ 00104 00105 FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, A01, /**/ A02, 00106 A10, A11, /**/ A12, 00107 /* ************** */ /* ****************** */ 00108 &ABL, /**/ &ABR, A20, A21, /**/ A22, 00109 FLA_TL ); 00110 00111 } 00112 00113 return FLA_SUCCESS; 00114 }