Functions | |
FLA_Error | FLA_Sylv_tn_blk_var17 (FLA_Obj isgn, FLA_Obj A, FLA_Obj B, FLA_Obj C, FLA_Obj scale, fla_sylv_t *cntl) |
FLA_Error FLA_Sylv_tn_blk_var17 | ( | FLA_Obj | isgn, | |
FLA_Obj | A, | |||
FLA_Obj | B, | |||
FLA_Obj | C, | |||
FLA_Obj | scale, | |||
fla_sylv_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_Sylv_internal().
Referenced by FLA_Sylv_tn().
00036 { 00037 FLA_Obj BTL, BTR, B00, B01, B02, 00038 BBL, BBR, B10, B11, B12, 00039 B20, B21, B22; 00040 00041 FLA_Obj CL, CR, C0, C1, C2; 00042 00043 dim_t b; 00044 00045 FLA_Part_2x2( B, &BTL, &BTR, 00046 &BBL, &BBR, 0, 0, FLA_TL ); 00047 00048 FLA_Part_1x2( C, &CL, &CR, 0, FLA_LEFT ); 00049 00050 while ( FLA_Obj_length( BTL ) < FLA_Obj_length( B ) ){ 00051 00052 b = FLA_Determine_blocksize( CR, FLA_RIGHT, FLA_Cntl_blocksize( cntl ) ); 00053 00054 FLA_Repart_2x2_to_3x3( BTL, /**/ BTR, &B00, /**/ &B01, &B02, 00055 /* ************* */ /* ******************** */ 00056 &B10, /**/ &B11, &B12, 00057 BBL, /**/ BBR, &B20, /**/ &B21, &B22, 00058 b, b, FLA_BR ); 00059 00060 FLA_Repart_1x2_to_1x3( CL, /**/ CR, &C0, /**/ &C1, &C2, 00061 b, FLA_RIGHT ); 00062 00063 // Loop Invariant: 00064 // CL = 00065 // CR = 00066 00067 /*------------------------------------------------------------*/ 00068 00069 // C1 = sylv( A', B11, C1 -/+ C0 * B01 ); 00070 FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, 00071 FLA_NEGATE( isgn ), C0, B01, FLA_ONE, C1, 00072 FLA_Cntl_sub_gemm1( cntl ) ); 00073 00074 FLA_Sylv_internal( FLA_TRANSPOSE, FLA_NO_TRANSPOSE, 00075 isgn, A, B11, C1, scale, 00076 FLA_Cntl_sub_sylv1( cntl ) ); 00077 00078 /*------------------------------------------------------------*/ 00079 00080 FLA_Cont_with_3x3_to_2x2( &BTL, /**/ &BTR, B00, B01, /**/ B02, 00081 B10, B11, /**/ B12, 00082 /* ************** */ /* ****************** */ 00083 &BBL, /**/ &BBR, B20, B21, /**/ B22, 00084 FLA_TL ); 00085 00086 FLA_Cont_with_1x3_to_1x2( &CL, /**/ &CR, C0, C1, /**/ C2, 00087 FLA_LEFT ); 00088 } 00089 00090 return FLA_SUCCESS; 00091 }