Functions | |
FLA_Error | FLA_Hemm_ru_blk_var4 (FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C, fla_hemm_t *cntl) |
FLA_Error FLA_Hemm_ru_blk_var4 | ( | FLA_Obj | alpha, | |
FLA_Obj | A, | |||
FLA_Obj | B, | |||
FLA_Obj | beta, | |||
FLA_Obj | C, | |||
fla_hemm_t * | cntl | |||
) |
References FLA_Cont_with_1x3_to_1x2(), FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_Hemm_internal(), FLA_Obj_length(), FLA_ONE, FLA_Part_1x2(), FLA_Part_2x2(), FLA_Repart_1x2_to_1x3(), and FLA_Repart_2x2_to_3x3().
Referenced by FLA_Hemm_ru().
00036 { 00037 FLA_Obj ATL, ATR, A00, A01, A02, 00038 ABL, ABR, A10, A11, A12, 00039 A20, A21, A22; 00040 00041 FLA_Obj BL, BR, B0, B1, B2; 00042 00043 FLA_Obj CL, CR, C0, C1, C2; 00044 00045 dim_t b; 00046 00047 FLA_Part_2x2( A, &ATL, &ATR, 00048 &ABL, &ABR, 0, 0, FLA_TL ); 00049 00050 FLA_Part_1x2( B, &BL, &BR, 0, FLA_LEFT ); 00051 00052 FLA_Part_1x2( C, &CL, &CR, 0, FLA_LEFT ); 00053 00054 while ( FLA_Obj_length( ATL ) < FLA_Obj_length( A ) ){ 00055 00056 b = FLA_Determine_blocksize( ABR, FLA_BR, FLA_Cntl_blocksize( cntl ) ); 00057 00058 FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, /**/ &A01, &A02, 00059 /* ************* */ /* ******************** */ 00060 &A10, /**/ &A11, &A12, 00061 ABL, /**/ ABR, &A20, /**/ &A21, &A22, 00062 b, b, FLA_BR ); 00063 00064 FLA_Repart_1x2_to_1x3( BL, /**/ BR, &B0, /**/ &B1, &B2, 00065 b, FLA_RIGHT ); 00066 00067 FLA_Repart_1x2_to_1x3( CL, /**/ CR, &C0, /**/ &C1, &C2, 00068 b, FLA_RIGHT ); 00069 00070 /*------------------------------------------------------------*/ 00071 00072 /* C0 = C0 + B1 * A01' */ 00073 FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_CONJ_TRANSPOSE, 00074 alpha, B1, A01, FLA_ONE, C0, 00075 FLA_Cntl_sub_gemm1( cntl ) ); 00076 00077 /* C1 = C1 + B1 * A11 */ 00078 FLA_Hemm_internal( FLA_RIGHT, FLA_UPPER_TRIANGULAR, 00079 alpha, A11, B1, beta, C1, 00080 FLA_Cntl_sub_hemm( cntl ) ); 00081 00082 /* C2 = C2 + B1 * A12 */ 00083 FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, 00084 alpha, B1, A12, FLA_ONE, C2, 00085 FLA_Cntl_sub_gemm2( cntl ) ); 00086 00087 /*------------------------------------------------------------*/ 00088 00089 FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, A01, /**/ A02, 00090 A10, A11, /**/ A12, 00091 /* ************** */ /* ****************** */ 00092 &ABL, /**/ &ABR, A20, A21, /**/ A22, 00093 FLA_TL ); 00094 00095 FLA_Cont_with_1x3_to_1x2( &BL, /**/ &BR, B0, B1, /**/ B2, 00096 FLA_LEFT ); 00097 00098 FLA_Cont_with_1x3_to_1x2( &CL, /**/ &CR, C0, C1, /**/ C2, 00099 FLA_LEFT ); 00100 00101 } 00102 00103 return FLA_SUCCESS; 00104 }