Functions | |
FLA_Error | FLA_Sylv_nt_blk_var6 (FLA_Obj isgn, FLA_Obj A, FLA_Obj B, FLA_Obj C, FLA_Obj scale, fla_sylv_t *cntl) |
FLA_Error FLA_Sylv_nt_blk_var6 | ( | FLA_Obj | isgn, | |
FLA_Obj | A, | |||
FLA_Obj | B, | |||
FLA_Obj | C, | |||
FLA_Obj | scale, | |||
fla_sylv_t * | cntl | |||
) |
References FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_MINUS_ONE, FLA_Obj_length(), FLA_ONE, FLA_Part_2x2(), FLA_Repart_2x2_to_3x3(), and FLA_Sylv_internal().
Referenced by FLA_Sylv_nt().
00038 { 00039 FLA_Obj ATL, ATR, A00, A01, A02, 00040 ABL, ABR, A10, A11, A12, 00041 A20, A21, A22; 00042 00043 FLA_Obj BTL, BTR, B00, B01, B02, 00044 BBL, BBR, B10, B11, B12, 00045 B20, B21, B22; 00046 00047 FLA_Obj CTL, CTR, C00, C01, C02, 00048 CBL, CBR, C10, C11, C12, 00049 C20, C21, C22; 00050 00051 dim_t b; 00052 00053 FLA_Part_2x2( A, &ATL, &ATR, 00054 &ABL, &ABR, 0, 0, FLA_BR ); 00055 00056 FLA_Part_2x2( B, &BTL, &BTR, 00057 &BBL, &BBR, 0, 0, FLA_BR ); 00058 00059 FLA_Part_2x2( C, &CTL, &CTR, 00060 &CBL, &CBR, 0, 0, FLA_BR ); 00061 00062 while ( FLA_Obj_length( ABR ) < FLA_Obj_length( A ) ){ 00063 00064 b = FLA_Determine_blocksize( CTL, 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_2x2_to_3x3( BTL, /**/ BTR, &B00, &B01, /**/ &B02, 00073 &B10, &B11, /**/ &B12, 00074 /* ************* */ /* ******************** */ 00075 BBL, /**/ BBR, &B20, &B21, /**/ &B22, 00076 b, b, FLA_TL ); 00077 00078 FLA_Repart_2x2_to_3x3( CTL, /**/ CTR, &C00, &C01, /**/ &C02, 00079 &C10, &C11, /**/ &C12, 00080 /* ************* */ /* ******************** */ 00081 CBL, /**/ CBR, &C20, &C21, /**/ &C22, 00082 b, b, FLA_TL ); 00083 00084 // Loop Invariant: 00085 // CTL = 00086 // CTR = 00087 // CBL = 00088 // CBR = 00089 00090 /*------------------------------------------------------------*/ 00091 00092 // C12 = sylv( A11, B22', C12 ); 00093 FLA_Sylv_internal( FLA_NO_TRANSPOSE, FLA_TRANSPOSE, 00094 isgn, A11, B22, C12, scale, 00095 FLA_Cntl_sub_sylv1( cntl ) ); 00096 00097 // C02 = C02 - A01 * C12; 00098 FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, 00099 FLA_MINUS_ONE, A01, C12, FLA_ONE, C02, 00100 FLA_Cntl_sub_gemm1( cntl ) ); 00101 00102 // C21 = sylv( A22, B11', C21 ); 00103 FLA_Sylv_internal( FLA_NO_TRANSPOSE, FLA_TRANSPOSE, 00104 isgn, A22, B11, C21, scale, 00105 FLA_Cntl_sub_sylv2( cntl ) ); 00106 00107 // C11 = sylv( A11, B11', C11 - A12 * C21 -/+ C12 * B12' ); 00108 FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_TRANSPOSE, 00109 FLA_NEGATE( isgn ), C12, B12, FLA_ONE, C11, 00110 FLA_Cntl_sub_gemm2( cntl ) ); 00111 00112 FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, 00113 FLA_MINUS_ONE, A12, C21, FLA_ONE, C11, 00114 FLA_Cntl_sub_gemm3( cntl ) ); 00115 00116 FLA_Sylv_internal( FLA_NO_TRANSPOSE, FLA_TRANSPOSE, 00117 isgn, A11, B11, C11, scale, 00118 FLA_Cntl_sub_sylv3( cntl ) ); 00119 00120 // C01 = C01 - A01 * C11 - A02 * C21; 00121 FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, 00122 FLA_MINUS_ONE, A02, C21, FLA_ONE, C01, 00123 FLA_Cntl_sub_gemm4( cntl ) ); 00124 00125 FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, 00126 FLA_MINUS_ONE, A01, C11, FLA_ONE, C01, 00127 FLA_Cntl_sub_gemm5( cntl ) ); 00128 00129 // C10 = C10 -/+ C12 * B02' -/+ C11 * B01'; 00130 FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_TRANSPOSE, 00131 FLA_NEGATE( isgn ), C11, B01, FLA_ONE, C10, 00132 FLA_Cntl_sub_gemm6( cntl ) ); 00133 00134 FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_TRANSPOSE, 00135 FLA_NEGATE( isgn ), C12, B02, FLA_ONE, C10, 00136 FLA_Cntl_sub_gemm7( cntl ) ); 00137 00138 // C20 = C20 -/+ C21 * B01'; 00139 FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_TRANSPOSE, 00140 FLA_NEGATE( isgn ), C21, B01, FLA_ONE, C20, 00141 FLA_Cntl_sub_gemm8( cntl ) ); 00142 00143 /*------------------------------------------------------------*/ 00144 00145 FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, /**/ A01, A02, 00146 /* ************** */ /* ****************** */ 00147 A10, /**/ A11, A12, 00148 &ABL, /**/ &ABR, A20, /**/ A21, A22, 00149 FLA_BR ); 00150 00151 FLA_Cont_with_3x3_to_2x2( &BTL, /**/ &BTR, B00, /**/ B01, B02, 00152 /* ************** */ /* ****************** */ 00153 B10, /**/ B11, B12, 00154 &BBL, /**/ &BBR, B20, /**/ B21, B22, 00155 FLA_BR ); 00156 00157 FLA_Cont_with_3x3_to_2x2( &CTL, /**/ &CTR, C00, /**/ C01, C02, 00158 /* ************** */ /* ****************** */ 00159 C10, /**/ C11, C12, 00160 &CBL, /**/ &CBR, C20, /**/ C21, C22, 00161 FLA_BR ); 00162 00163 } 00164 00165 return FLA_SUCCESS; 00166 }