Functions | |
FLA_Error | FLA_Trmm_run_blk_var1 (FLA_Diag diagA, FLA_Obj alpha, FLA_Obj A, FLA_Obj B, fla_trmm_t *cntl) |
FLA_Error FLA_Trmm_run_blk_var1 | ( | 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_run().
00038 { 00039 FLA_Obj ATL, ATR, A00, A01, A02, 00040 ABL, ABR, A10, A11, A12, 00041 A20, A21, A22; 00042 00043 FLA_Obj BL, BR, B0, B1, B2; 00044 00045 dim_t b; 00046 00047 FLA_Part_2x2( A, &ATL, &ATR, 00048 &ABL, &ABR, 0, 0, FLA_BR ); 00049 00050 FLA_Part_1x2( B, &BL, &BR, 0, FLA_RIGHT ); 00051 00052 while ( FLA_Obj_length( ABR ) < FLA_Obj_length( A ) ){ 00053 00054 b = FLA_Determine_blocksize( ATL, FLA_TL, FLA_Cntl_blocksize( cntl ) ); 00055 00056 FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, &A01, /**/ &A02, 00057 &A10, &A11, /**/ &A12, 00058 /* ************* */ /* ******************** */ 00059 ABL, /**/ ABR, &A20, &A21, /**/ &A22, 00060 b, b, FLA_TL ); 00061 00062 FLA_Repart_1x2_to_1x3( BL, /**/ BR, &B0, &B1, /**/ &B2, 00063 b, FLA_LEFT ); 00064 00065 /*------------------------------------------------------------*/ 00066 00067 /* B1 = B1 * triu( A11 ); */ 00068 FLA_Trmm_internal( FLA_RIGHT, FLA_UPPER_TRIANGULAR, FLA_NO_TRANSPOSE, diagA, 00069 alpha, A11, B1, 00070 FLA_Cntl_sub_trmm( cntl ) ); 00071 00072 /* B1 = B1 + B0 * A01; */ 00073 FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, 00074 FLA_ONE, B0, A01, FLA_ONE, B1, 00075 FLA_Cntl_sub_gemm( cntl ) ); 00076 00077 /*------------------------------------------------------------*/ 00078 00079 FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, /**/ A01, A02, 00080 /* ************** */ /* ****************** */ 00081 A10, /**/ A11, A12, 00082 &ABL, /**/ &ABR, A20, /**/ A21, A22, 00083 FLA_BR ); 00084 00085 FLA_Cont_with_1x3_to_1x2( &BL, /**/ &BR, B0, /**/ B1, B2, 00086 FLA_RIGHT ); 00087 00088 } 00089 00090 return FLA_SUCCESS; 00091 }