Functions | |
FLA_Error | FLA_Herk_uh_blk_var1 (FLA_Obj alpha, FLA_Obj A, FLA_Obj beta, FLA_Obj C, fla_herk_t *cntl) |
FLA_Error FLA_Herk_uh_blk_var1 | ( | FLA_Obj | alpha, | |
FLA_Obj | A, | |||
FLA_Obj | beta, | |||
FLA_Obj | C, | |||
fla_herk_t * | cntl | |||
) |
References FLA_Cont_with_1x3_to_1x2(), FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_Herk_internal(), FLA_Obj_width(), FLA_Part_1x2(), FLA_Part_2x2(), FLA_Repart_1x2_to_1x3(), and FLA_Repart_2x2_to_3x3().
Referenced by FLA_Herk_uh().
00038 { 00039 FLA_Obj AL, AR, A0, A1, A2; 00040 00041 FLA_Obj CTL, CTR, C00, C01, C02, 00042 CBL, CBR, C10, C11, C12, 00043 C20, C21, C22; 00044 00045 dim_t b; 00046 00047 FLA_Part_1x2( A, &AL, &AR, 0, FLA_LEFT ); 00048 00049 FLA_Part_2x2( C, &CTL, &CTR, 00050 &CBL, &CBR, 0, 0, FLA_TL ); 00051 00052 while ( FLA_Obj_width( AL ) < FLA_Obj_width( A ) ){ 00053 00054 b = FLA_Determine_blocksize( AR, FLA_RIGHT, FLA_Cntl_blocksize( cntl ) ); 00055 00056 FLA_Repart_1x2_to_1x3( AL, /**/ AR, &A0, /**/ &A1, &A2, 00057 b, FLA_RIGHT ); 00058 00059 FLA_Repart_2x2_to_3x3( CTL, /**/ CTR, &C00, /**/ &C01, &C02, 00060 /* ************* */ /* ******************** */ 00061 &C10, /**/ &C11, &C12, 00062 CBL, /**/ CBR, &C20, /**/ &C21, &C22, 00063 b, b, FLA_BR ); 00064 00065 /*------------------------------------------------------------*/ 00066 00067 /* C12 = C12 + A1' * A2 */ 00068 FLA_Gemm_internal( FLA_CONJ_TRANSPOSE, FLA_NO_TRANSPOSE, 00069 alpha, A1, A2, beta, C12, 00070 FLA_Cntl_sub_gemm( cntl ) ); 00071 00072 /* C11 = C11 + A1' * A1 */ 00073 FLA_Herk_internal( FLA_UPPER_TRIANGULAR, FLA_CONJ_TRANSPOSE, 00074 alpha, A1, beta, C11, 00075 FLA_Cntl_sub_herk( cntl ) ); 00076 00077 /*------------------------------------------------------------*/ 00078 00079 FLA_Cont_with_1x3_to_1x2( &AL, /**/ &AR, A0, A1, /**/ A2, 00080 FLA_LEFT ); 00081 00082 FLA_Cont_with_3x3_to_2x2( &CTL, /**/ &CTR, C00, C01, /**/ C02, 00083 C10, C11, /**/ C12, 00084 /* ************** */ /* ****************** */ 00085 &CBL, /**/ &CBR, C20, C21, /**/ C22, 00086 FLA_TL ); 00087 00088 } 00089 00090 return FLA_SUCCESS; 00091 }