Functions | |
FLA_Error | FLA_LU_nopiv_blk_var4 (FLA_Obj A, fla_lu_t *cntl) |
References FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_LU_nopiv_internal(), FLA_MINUS_ONE, FLA_Obj_length(), FLA_Obj_width(), FLA_ONE, FLA_Part_2x2(), FLA_Repart_2x2_to_3x3(), and FLA_Trsm_internal().
Referenced by FLA_LU_nopiv_internal().
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 FLA_Obj_width( ATL ) < FLA_Obj_width( A )){ 00050 00051 b = FLA_Determine_blocksize( ABR, FLA_BR, FLA_Cntl_blocksize( cntl ) ); 00052 00053 FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, /**/ &A01, &A02, 00054 /* ************* */ /* ******************** */ 00055 &A10, /**/ &A11, &A12, 00056 ABL, /**/ ABR, &A20, /**/ &A21, &A22, 00057 b, b, FLA_BR ); 00058 00059 /*------------------------------------------------------------*/ 00060 00061 /* A11 = LU( A11 - A10 * A01 ) */ 00062 FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, 00063 FLA_MINUS_ONE, A10, A01, FLA_ONE, A11, 00064 FLA_Cntl_sub_gemm1( cntl ) ); 00065 00066 FLA_LU_nopiv_internal( A11, FLA_Cntl_sub_lu( cntl ) ); 00067 00068 /* A12 = A12 - A10 * A02 */ 00069 FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, 00070 FLA_MINUS_ONE, A10, A02, FLA_ONE, A12, 00071 FLA_Cntl_sub_gemm2( cntl ) ); 00072 00073 /* A12 = trilu( A11 ) \ A12 */ 00074 FLA_Trsm_internal( FLA_LEFT, FLA_LOWER_TRIANGULAR, 00075 FLA_NO_TRANSPOSE, FLA_UNIT_DIAG, 00076 FLA_ONE, A11, A12, 00077 FLA_Cntl_sub_trsm1( cntl ) ); 00078 00079 /* A21 = A21 - A20 * A01 */ 00080 FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, 00081 FLA_MINUS_ONE, A20, A01, FLA_ONE, A21, 00082 FLA_Cntl_sub_gemm3( cntl ) ); 00083 00084 /* A21 = A21 / triu( A11 ) */ 00085 FLA_Trsm_internal( FLA_RIGHT, FLA_UPPER_TRIANGULAR, 00086 FLA_NO_TRANSPOSE, FLA_NONUNIT_DIAG, 00087 FLA_ONE, A11, A21, 00088 FLA_Cntl_sub_trsm2( cntl ) ); 00089 00090 /*------------------------------------------------------------*/ 00091 00092 FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, A01, /**/ A02, 00093 A10, A11, /**/ A12, 00094 /* ************** */ /* ****************** */ 00095 &ABL, /**/ &ABR, A20, A21, /**/ A22, 00096 FLA_TL ); 00097 } 00098 00099 return FLA_SUCCESS; 00100 }