Functions | |
FLA_Error | FLA_Sylv_nn_blk_var4 (FLA_Obj isgn, FLA_Obj A, FLA_Obj B, FLA_Obj C, FLA_Obj scale, fla_sylv_t *cntl) |
FLA_Error FLA_Sylv_nn_blk_var4 | ( | 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_nn().
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_TL ); 00058 00059 FLA_Part_2x2( C, &CTL, &CTR, 00060 &CBL, &CBR, 0, 0, FLA_BL ); 00061 00062 while ( FLA_Obj_length( ABR ) < FLA_Obj_length( A ) ){ 00063 00064 b = FLA_Determine_blocksize( CTR, FLA_TR, 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 /* ************* */ /* ******************** */ 00074 &B10, /**/ &B11, &B12, 00075 BBL, /**/ BBR, &B20, /**/ &B21, &B22, 00076 b, b, FLA_BR ); 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_TR ); 00083 00084 // Loop Invariant: 00085 // CTL = CTL 00086 // CTR = CTR 00087 // CBL = sylv( ABR, BTL, CBL ) 00088 // CBR = CBR - sylv( ABR, BTL, CBL ) * BTR 00089 00090 /*------------------------------------------------------------*/ 00091 00092 // C10 = sylv( A11, B00, C10 - A12 * C20 ); 00093 FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, 00094 FLA_MINUS_ONE, A12, C20, FLA_ONE, C10, 00095 FLA_Cntl_sub_gemm1( cntl ) ); 00096 00097 FLA_Sylv_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, 00098 isgn, A11, B00, C10, scale, 00099 FLA_Cntl_sub_sylv1( cntl ) ); 00100 00101 // C21 = sylv( A22, B11, C21 ); 00102 FLA_Sylv_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, 00103 isgn, A22, B11, C21, scale, 00104 FLA_Cntl_sub_sylv2( cntl ) ); 00105 00106 // C11 = sylv( A11, B11, C11 - A12 * C21 -/+ C10 * B01 ); 00107 FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, 00108 FLA_NEGATE( isgn ), C10, B01, FLA_ONE, C11, 00109 FLA_Cntl_sub_gemm2( cntl ) ); 00110 00111 FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, 00112 FLA_MINUS_ONE, A12, C21, FLA_ONE, C11, 00113 FLA_Cntl_sub_gemm3( cntl ) ); 00114 00115 FLA_Sylv_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, 00116 isgn, A11, B11, C11, scale, 00117 FLA_Cntl_sub_sylv3( cntl ) ); 00118 00119 // C12 = C12 -/+ C10 * B02 -/+ C11 * B12; 00120 FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, 00121 FLA_NEGATE( isgn ), C11, B12, FLA_ONE, C12, 00122 FLA_Cntl_sub_gemm4( cntl ) ); 00123 00124 FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, 00125 FLA_NEGATE( isgn ), C10, B02, FLA_ONE, C12, 00126 FLA_Cntl_sub_gemm5( cntl ) ); 00127 00128 // C22 = C22 -/+ C21 * B12; 00129 FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, 00130 FLA_NEGATE( isgn ), C21, B12, FLA_ONE, C22, 00131 FLA_Cntl_sub_gemm6( cntl ) ); 00132 00133 00134 /*------------------------------------------------------------*/ 00135 00136 FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, /**/ A01, A02, 00137 /* ************** */ /* ****************** */ 00138 A10, /**/ A11, A12, 00139 &ABL, /**/ &ABR, A20, /**/ A21, A22, 00140 FLA_BR ); 00141 00142 FLA_Cont_with_3x3_to_2x2( &BTL, /**/ &BTR, B00, B01, /**/ B02, 00143 B10, B11, /**/ B12, 00144 /* ************** */ /* ****************** */ 00145 &BBL, /**/ &BBR, B20, B21, /**/ B22, 00146 FLA_TL ); 00147 00148 FLA_Cont_with_3x3_to_2x2( &CTL, /**/ &CTR, C00, C01, /**/ C02, 00149 /* ************** */ /* ****************** */ 00150 C10, C11, /**/ C12, 00151 &CBL, /**/ &CBR, C20, C21, /**/ C22, 00152 FLA_BL ); 00153 00154 } 00155 00156 return FLA_SUCCESS; 00157 }