Functions | |
FLA_Error | FLA_Hemm_ll_blk_var8 (FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C, fla_hemm_t *cntl) |
FLA_Error FLA_Hemm_ll_blk_var8 | ( | FLA_Obj | alpha, | |
FLA_Obj | A, | |||
FLA_Obj | B, | |||
FLA_Obj | beta, | |||
FLA_Obj | C, | |||
fla_hemm_t * | cntl | |||
) |
References FLA_Cont_with_3x1_to_2x1(), FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_Hemm_internal(), FLA_Obj_length(), FLA_ONE, FLA_Part_2x1(), FLA_Part_2x2(), FLA_Repart_2x1_to_3x1(), and FLA_Repart_2x2_to_3x3().
Referenced by FLA_Hemm_ll().
00038 { 00039 FLA_Obj ATL, ATR, A00, A01, A02, 00040 ABL, ABR, A10, A11, A12, 00041 A20, A21, A22; 00042 00043 FLA_Obj BT, B0, 00044 BB, B1, 00045 B2; 00046 00047 FLA_Obj CT, C0, 00048 CB, C1, 00049 C2; 00050 00051 dim_t b; 00052 00053 FLA_Part_2x2( A, &ATL, &ATR, 00054 &ABL, &ABR, 0, 0, FLA_BR ); 00055 00056 FLA_Part_2x1( B, &BT, 00057 &BB, 0, FLA_BOTTOM ); 00058 00059 FLA_Part_2x1( C, &CT, 00060 &CB, 0, FLA_BOTTOM ); 00061 00062 while ( FLA_Obj_length( ABR ) < FLA_Obj_length( A ) ){ 00063 00064 b = FLA_Determine_blocksize( ATL, FLA_TL, FLA_Cntl_blocksize( cntl ) ); 00065 00066 FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, &A01, /**/ &A02, 00067 &A10, &A11, /**/ &A12, 00068 /* ************* */ /* ******************** */ 00069 ABL, /**/ ABR, &A20, &A21, /**/ &A22, 00070 b, b, FLA_TL ); 00071 00072 FLA_Repart_2x1_to_3x1( BT, &B0, 00073 &B1, 00074 /* ** */ /* ** */ 00075 BB, &B2, b, FLA_TOP ); 00076 00077 FLA_Repart_2x1_to_3x1( CT, &C0, 00078 &C1, 00079 /* ** */ /* ** */ 00080 CB, &C2, b, FLA_TOP ); 00081 00082 /*------------------------------------------------------------*/ 00083 00084 /* C0 = C0 + A10' * B1 */ 00085 FLA_Gemm_internal( FLA_CONJ_TRANSPOSE, FLA_NO_TRANSPOSE, 00086 alpha, A10, B1, FLA_ONE, C0, 00087 FLA_Cntl_sub_gemm1( cntl ) ); 00088 00089 /* C1 = C1 + A11 * B1 */ 00090 FLA_Hemm_internal( FLA_LEFT, FLA_LOWER_TRIANGULAR, 00091 alpha, A11, B1, beta, C1, 00092 FLA_Cntl_sub_hemm( cntl ) ); 00093 00094 /* C2 = C2 + A21 * B1 */ 00095 FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, 00096 alpha, A21, B1, FLA_ONE, C2, 00097 FLA_Cntl_sub_gemm2( cntl ) ); 00098 00099 /*------------------------------------------------------------*/ 00100 00101 FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, /**/ A01, A02, 00102 /* ************** */ /* ****************** */ 00103 A10, /**/ A11, A12, 00104 &ABL, /**/ &ABR, A20, /**/ A21, A22, 00105 FLA_BR ); 00106 00107 FLA_Cont_with_3x1_to_2x1( &BT, B0, 00108 /* ** */ /* ** */ 00109 B1, 00110 &BB, B2, FLA_BOTTOM ); 00111 00112 FLA_Cont_with_3x1_to_2x1( &CT, C0, 00113 /* ** */ /* ** */ 00114 C1, 00115 &CB, C2, FLA_BOTTOM ); 00116 00117 } 00118 00119 return FLA_SUCCESS; 00120 }