Go to the source code of this file.
Functions | |
FLA_Error | FLA_Ttmm_u_blk_var1 (FLA_Obj A, fla_ttmm_t *cntl) |
FLA_Error | FLA_Ttmm_u_blk_var2 (FLA_Obj A, fla_ttmm_t *cntl) |
FLA_Error | FLA_Ttmm_u_blk_var3 (FLA_Obj A, fla_ttmm_t *cntl) |
FLA_Error | FLA_Ttmm_u_unb_var1 (FLA_Obj A) |
FLA_Error | FLA_Ttmm_u_unb_var2 (FLA_Obj A) |
FLA_Error | FLA_Ttmm_u_unb_var3 (FLA_Obj A) |
FLA_Error FLA_Ttmm_u_blk_var1 | ( | FLA_Obj | A, | |
fla_ttmm_t * | cntl | |||
) |
References FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Herk_internal(), FLA_Obj_is_real(), FLA_Obj_length(), FLA_ONE, FLA_Part_2x2(), FLA_Repart_2x2_to_3x3(), FLA_Syrk_internal(), FLA_Trmm_internal(), and FLA_Ttmm_internal().
Referenced by FLA_Ttmm_u().
00036 { 00037 FLA_Obj ATL, ATR, A00, A01, A02, 00038 ABL, ABR, A10, A11, A12, 00039 A20, A21, A22; 00040 00041 dim_t b; 00042 00043 FLA_Part_2x2( A, &ATL, &ATR, 00044 &ABL, &ABR, 0, 0, FLA_TL ); 00045 00046 while ( FLA_Obj_length( ATL ) < FLA_Obj_length( A ) ){ 00047 00048 b = FLA_Determine_blocksize( ABR, FLA_BR, FLA_Cntl_blocksize( cntl ) ); 00049 00050 FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, /**/ &A01, &A02, 00051 /* ************* */ /* ******************** */ 00052 &A10, /**/ &A11, &A12, 00053 ABL, /**/ ABR, &A20, /**/ &A21, &A22, 00054 b, b, FLA_BR ); 00055 00056 /*------------------------------------------------------------*/ 00057 00058 if ( FLA_Obj_is_real( A ) ) 00059 { 00060 /* A00 = A00 + A01 * A01'; */ 00061 FLA_Syrk_internal( FLA_UPPER_TRIANGULAR, FLA_NO_TRANSPOSE, 00062 FLA_ONE, A01, FLA_ONE, A00, 00063 FLA_Cntl_sub_syrk( cntl ) ); 00064 00065 /* A01 = A01 * triu( A11 )'; */ 00066 FLA_Trmm_internal( FLA_RIGHT, FLA_UPPER_TRIANGULAR, FLA_TRANSPOSE, FLA_NONUNIT_DIAG, 00067 FLA_ONE, A11, A01, 00068 FLA_Cntl_sub_trmm( cntl ) ); 00069 00070 /* A11 = triu( A11 ) * triu( A11 )'; */ 00071 FLA_Ttmm_internal( FLA_UPPER_TRIANGULAR, A11, 00072 FLA_Cntl_sub_ttmm( cntl ) ); 00073 } 00074 else 00075 { 00076 /* A00 = A00 + A01 * A01'; */ 00077 FLA_Herk_internal( FLA_UPPER_TRIANGULAR, FLA_NO_TRANSPOSE, 00078 FLA_ONE, A01, FLA_ONE, A00, 00079 FLA_Cntl_sub_herk( cntl ) ); 00080 00081 /* A01 = A01 * triu( A11 )'; */ 00082 FLA_Trmm_internal( FLA_RIGHT, FLA_UPPER_TRIANGULAR, FLA_CONJ_TRANSPOSE, FLA_NONUNIT_DIAG, 00083 FLA_ONE, A11, A01, 00084 FLA_Cntl_sub_trmm( cntl ) ); 00085 00086 /* A11 = triu( A11 ) * triu( A11 )'; */ 00087 FLA_Ttmm_internal( FLA_UPPER_TRIANGULAR, A11, 00088 FLA_Cntl_sub_ttmm( cntl ) ); 00089 } 00090 00091 /*------------------------------------------------------------*/ 00092 00093 FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, A01, /**/ A02, 00094 A10, A11, /**/ A12, 00095 /* ************** */ /* ****************** */ 00096 &ABL, /**/ &ABR, A20, A21, /**/ A22, 00097 FLA_TL ); 00098 00099 } 00100 00101 return FLA_SUCCESS; 00102 }
FLA_Error FLA_Ttmm_u_blk_var2 | ( | FLA_Obj | A, | |
fla_ttmm_t * | cntl | |||
) |
References FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_Herk_internal(), FLA_Obj_is_real(), FLA_Obj_length(), FLA_ONE, FLA_Part_2x2(), FLA_Repart_2x2_to_3x3(), FLA_Syrk_internal(), FLA_Trmm_internal(), and FLA_Ttmm_internal().
Referenced by FLA_Ttmm_u().
00038 { 00039 FLA_Obj ATL, ATR, A00, A01, A02, 00040 ABL, ABR, A10, A11, A12, 00041 A20, A21, A22; 00042 00043 dim_t b; 00044 00045 FLA_Part_2x2( A, &ATL, &ATR, 00046 &ABL, &ABR, 0, 0, FLA_TL ); 00047 00048 while ( FLA_Obj_length( ATL ) < FLA_Obj_length( A ) ){ 00049 00050 b = FLA_Determine_blocksize( ABR, FLA_BR, FLA_Cntl_blocksize( cntl ) ); 00051 00052 FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, /**/ &A01, &A02, 00053 /* ************* */ /* ******************** */ 00054 &A10, /**/ &A11, &A12, 00055 ABL, /**/ ABR, &A20, /**/ &A21, &A22, 00056 b, b, FLA_BR ); 00057 00058 /*------------------------------------------------------------*/ 00059 00060 if ( FLA_Obj_is_real( A ) ) 00061 { 00062 /* A01 = A01 * triu( A11 )'; */ 00063 FLA_Trmm_internal( FLA_RIGHT, FLA_UPPER_TRIANGULAR, FLA_TRANSPOSE, FLA_NONUNIT_DIAG, 00064 FLA_ONE, A11, A01, 00065 FLA_Cntl_sub_trmm( cntl ) ); 00066 00067 /* A01 = A01 + A02 * A12'; */ 00068 FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_TRANSPOSE, 00069 FLA_ONE, A02, A12, FLA_ONE, A01, 00070 FLA_Cntl_sub_gemm( cntl ) ); 00071 00072 /* A11 = triu( A11 ) * triu( A11 )'; */ 00073 FLA_Ttmm_internal( FLA_UPPER_TRIANGULAR, A11, 00074 FLA_Cntl_sub_ttmm( cntl ) ); 00075 00076 /* A11 = A11 + A12 * A12'; */ 00077 FLA_Syrk_internal( FLA_UPPER_TRIANGULAR, FLA_NO_TRANSPOSE, 00078 FLA_ONE, A12, FLA_ONE, A11, 00079 FLA_Cntl_sub_syrk( cntl ) ); 00080 } 00081 else 00082 { 00083 /* A01 = A01 * triu( A11 )'; */ 00084 FLA_Trmm_internal( FLA_RIGHT, FLA_UPPER_TRIANGULAR, FLA_CONJ_TRANSPOSE, FLA_NONUNIT_DIAG, 00085 FLA_ONE, A11, A01, 00086 FLA_Cntl_sub_trmm( cntl ) ); 00087 00088 /* A01 = A01 + A02 * A12'; */ 00089 FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_CONJ_TRANSPOSE, 00090 FLA_ONE, A02, A12, FLA_ONE, A01, 00091 FLA_Cntl_sub_gemm( cntl ) ); 00092 00093 /* A11 = triu( A11 ) * triu( A11 )'; */ 00094 FLA_Ttmm_internal( FLA_UPPER_TRIANGULAR, A11, 00095 FLA_Cntl_sub_ttmm( cntl ) ); 00096 00097 /* A11 = A11 + A12 * A12'; */ 00098 FLA_Herk_internal( FLA_UPPER_TRIANGULAR, FLA_NO_TRANSPOSE, 00099 FLA_ONE, A12, FLA_ONE, A11, 00100 FLA_Cntl_sub_herk( cntl ) ); 00101 } 00102 00103 /*------------------------------------------------------------*/ 00104 00105 FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, A01, /**/ A02, 00106 A10, A11, /**/ A12, 00107 /* ************** */ /* ****************** */ 00108 &ABL, /**/ &ABR, A20, A21, /**/ A22, 00109 FLA_TL ); 00110 00111 } 00112 00113 return FLA_SUCCESS; 00114 }
FLA_Error FLA_Ttmm_u_blk_var3 | ( | FLA_Obj | A, | |
fla_ttmm_t * | cntl | |||
) |
References FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Herk_internal(), FLA_Obj_is_real(), FLA_Obj_length(), FLA_ONE, FLA_Part_2x2(), FLA_Repart_2x2_to_3x3(), FLA_Syrk_internal(), FLA_Trmm_internal(), and FLA_Ttmm_internal().
Referenced by FLA_Ttmm_u().
00038 { 00039 FLA_Obj ATL, ATR, A00, A01, A02, 00040 ABL, ABR, A10, A11, A12, 00041 A20, A21, A22; 00042 00043 dim_t b; 00044 00045 FLA_Part_2x2( A, &ATL, &ATR, 00046 &ABL, &ABR, 0, 0, FLA_TL ); 00047 00048 while ( FLA_Obj_length( ATL ) < FLA_Obj_length( A ) ){ 00049 00050 b = FLA_Determine_blocksize( ABR, FLA_BR, FLA_Cntl_blocksize( cntl ) ); 00051 00052 FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, /**/ &A01, &A02, 00053 /* ************* */ /* ******************** */ 00054 &A10, /**/ &A11, &A12, 00055 ABL, /**/ ABR, &A20, /**/ &A21, &A22, 00056 b, b, FLA_BR ); 00057 00058 /*------------------------------------------------------------*/ 00059 00060 if ( FLA_Obj_is_real( A ) ) 00061 { 00062 /* A11 = triu( A11 ) * triu( A11 )'; */ 00063 FLA_Ttmm_internal( FLA_UPPER_TRIANGULAR, A11, 00064 FLA_Cntl_sub_ttmm( cntl ) ); 00065 00066 /* A11 = A11 + A12 * A12'; */ 00067 FLA_Syrk_internal( FLA_UPPER_TRIANGULAR, FLA_NO_TRANSPOSE, 00068 FLA_ONE, A12, FLA_ONE, A11, 00069 FLA_Cntl_sub_syrk( cntl ) ); 00070 00071 /* A12 = A12 * triu( A22 )'; */ 00072 FLA_Trmm_internal( FLA_RIGHT, FLA_UPPER_TRIANGULAR, FLA_TRANSPOSE, FLA_NONUNIT_DIAG, 00073 FLA_ONE, A22, A12, 00074 FLA_Cntl_sub_trmm( cntl ) ); 00075 } 00076 else 00077 { 00078 /* A11 = triu( A11 ) * triu( A11 )'; */ 00079 FLA_Ttmm_internal( FLA_UPPER_TRIANGULAR, A11, 00080 FLA_Cntl_sub_ttmm( cntl ) ); 00081 00082 /* A11 = A11 + A12 * A12'; */ 00083 FLA_Herk_internal( FLA_UPPER_TRIANGULAR, FLA_NO_TRANSPOSE, 00084 FLA_ONE, A12, FLA_ONE, A11, 00085 FLA_Cntl_sub_herk( cntl ) ); 00086 00087 /* A12 = A12 * triu( A22 )'; */ 00088 FLA_Trmm_internal( FLA_RIGHT, FLA_UPPER_TRIANGULAR, FLA_CONJ_TRANSPOSE, FLA_NONUNIT_DIAG, 00089 FLA_ONE, A22, A12, 00090 FLA_Cntl_sub_trmm( cntl ) ); 00091 } 00092 00093 /*------------------------------------------------------------*/ 00094 00095 FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, A01, /**/ A02, 00096 A10, A11, /**/ A12, 00097 /* ************** */ /* ****************** */ 00098 &ABL, /**/ &ABR, A20, A21, /**/ A22, 00099 FLA_TL ); 00100 00101 } 00102 00103 return FLA_SUCCESS; 00104 }
References FLA_Absolute_square(), FLA_Cont_with_3x3_to_2x2(), FLA_Her_external(), FLA_Obj_is_real(), FLA_Obj_length(), FLA_ONE, FLA_Part_2x2(), FLA_Repart_2x2_to_3x3(), FLA_Scal_external(), and FLA_Syr_external().
00038 { 00039 FLA_Obj ATL, ATR, A00, a01, A02, 00040 ABL, ABR, a10t, alpha11, a12t, 00041 A20, a21, A22; 00042 00043 FLA_Part_2x2( A, &ATL, &ATR, 00044 &ABL, &ABR, 0, 0, FLA_TL ); 00045 00046 while ( FLA_Obj_length( ATL ) < FLA_Obj_length( A ) ){ 00047 00048 FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, /**/ &a01, &A02, 00049 /* ************* */ /* ************************** */ 00050 &a10t, /**/ &alpha11, &a12t, 00051 ABL, /**/ ABR, &A20, /**/ &a21, &A22, 00052 1, 1, FLA_BR ); 00053 00054 /*------------------------------------------------------------*/ 00055 00056 if ( FLA_Obj_is_real( A ) ) 00057 { 00058 /* A00 = A00 + a01 * a01'; */ 00059 FLA_Syr_external( FLA_UPPER_TRIANGULAR, FLA_ONE, a01, A00 ); 00060 00061 /* a01 = alpha11 * a01; */ 00062 FLA_Scal_external( alpha11, a01 ); 00063 00064 /* alpha11 = alpha11 * alpha11; */ 00065 FLA_Absolute_square( alpha11 ); 00066 } 00067 else 00068 { 00069 /* A00 = A00 + a01 * a01'; */ 00070 FLA_Her_external( FLA_UPPER_TRIANGULAR, FLA_ONE, a01, A00 ); 00071 00072 /* a10t = alpha11 * a01; */ 00073 FLA_Scal_external( alpha11, a01 ); 00074 00075 /* alpha11 = alpha11 * alpha11'; */ 00076 FLA_Absolute_square( alpha11 ); 00077 } 00078 00079 /*------------------------------------------------------------*/ 00080 00081 FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, a01, /**/ A02, 00082 a10t, alpha11, /**/ a12t, 00083 /* ************** */ /* ************************ */ 00084 &ABL, /**/ &ABR, A20, a21, /**/ A22, 00085 FLA_TL ); 00086 00087 } 00088 00089 return FLA_SUCCESS; 00090 }
References FLA_Absolute_square(), FLA_Cont_with_3x3_to_2x2(), FLA_Dotcs_external(), FLA_Dots_external(), FLA_Gemv_external(), FLA_Gemvc_external(), FLA_Obj_is_real(), FLA_Obj_length(), FLA_ONE, FLA_Part_2x2(), FLA_Repart_2x2_to_3x3(), and FLA_Scal_external().
00038 { 00039 FLA_Obj ATL, ATR, A00, a01, A02, 00040 ABL, ABR, a10t, alpha11, a12t, 00041 A20, a21, A22; 00042 00043 FLA_Part_2x2( A, &ATL, &ATR, 00044 &ABL, &ABR, 0, 0, FLA_TL ); 00045 00046 while ( FLA_Obj_length( ATL ) < FLA_Obj_length( A ) ){ 00047 00048 FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, /**/ &a01, &A02, 00049 /* ************* */ /* ************************** */ 00050 &a10t, /**/ &alpha11, &a12t, 00051 ABL, /**/ ABR, &A20, /**/ &a21, &A22, 00052 1, 1, FLA_BR ); 00053 00054 /*------------------------------------------------------------*/ 00055 00056 if ( FLA_Obj_is_real( A ) ) 00057 { 00058 /* a10t = alpha11 * a01; */ 00059 FLA_Scal_external( alpha11, a01 ); 00060 00061 /* a01 = a01 + A02 * a12t'; */ 00062 FLA_Gemv_external( FLA_NO_TRANSPOSE, FLA_ONE, A02, a12t, FLA_ONE, a01 ); 00063 00064 /* alpha11 = alpha11 * alpha11; */ 00065 FLA_Absolute_square( alpha11 ); 00066 00067 /* alpha11 = alpha11 + a12t * a12t'; */ 00068 FLA_Dots_external( FLA_ONE, a12t, a12t, FLA_ONE, alpha11 ); 00069 } 00070 else 00071 { 00072 /* a10t = alpha11 * a01; */ 00073 FLA_Scal_external( alpha11, a01 ); 00074 00075 /* a01 = a01 + A02 * a12t'; */ 00076 FLA_Gemvc_external( FLA_NO_TRANSPOSE, FLA_CONJUGATE, FLA_ONE, A02, a12t, FLA_ONE, a01 ); 00077 00078 /* alpha11 = alpha11 * alpha11'; */ 00079 FLA_Absolute_square( alpha11 ); 00080 00081 /* alpha11 = alpha11 + a12t * a12t'; */ 00082 FLA_Dotcs_external( FLA_CONJUGATE, FLA_ONE, a12t, a12t, FLA_ONE, alpha11 ); 00083 } 00084 00085 /*------------------------------------------------------------*/ 00086 00087 FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, a01, /**/ A02, 00088 a10t, alpha11, /**/ a12t, 00089 /* ************** */ /* ************************ */ 00090 &ABL, /**/ &ABR, A20, a21, /**/ A22, 00091 FLA_TL ); 00092 00093 } 00094 00095 return FLA_SUCCESS; 00096 }
References FLA_Absolute_square(), FLA_Cont_with_3x3_to_2x2(), FLA_Dotcs_external(), FLA_Dots_external(), FLA_Obj_is_real(), FLA_Obj_length(), FLA_ONE, FLA_Part_2x2(), FLA_Repart_2x2_to_3x3(), and FLA_Trmv_external().
00038 { 00039 FLA_Obj ATL, ATR, A00, a01, A02, 00040 ABL, ABR, a10t, alpha11, a12t, 00041 A20, a21, A22; 00042 00043 FLA_Part_2x2( A, &ATL, &ATR, 00044 &ABL, &ABR, 0, 0, FLA_TL ); 00045 00046 while ( FLA_Obj_length( ATL ) < FLA_Obj_length( A ) ){ 00047 00048 FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, /**/ &a01, &A02, 00049 /* ************* */ /* ************************** */ 00050 &a10t, /**/ &alpha11, &a12t, 00051 ABL, /**/ ABR, &A20, /**/ &a21, &A22, 00052 1, 1, FLA_BR ); 00053 00054 /*------------------------------------------------------------*/ 00055 00056 if ( FLA_Obj_is_real( A ) ) 00057 { 00058 /* alpha11 = alpha11 * alpha11; */ 00059 FLA_Absolute_square( alpha11 ); 00060 00061 /* alpha11 = alpha11 + a12t * a12t' */ 00062 FLA_Dots_external( FLA_ONE, a12t, a12t, FLA_ONE, alpha11 ); 00063 00064 /* a12t = a12t * triu( A22 )'; */ 00065 /* a12t' = triu( A22 ) * a12t'; */ 00066 FLA_Trmv_external( FLA_UPPER_TRIANGULAR, FLA_NO_TRANSPOSE, FLA_NONUNIT_DIAG, 00067 A22, a12t ); 00068 } 00069 else 00070 { 00071 /* alpha11 = alpha11 * alpha11'; */ 00072 FLA_Absolute_square( alpha11 ); 00073 00074 /* alpha11 = alpha11 + a12t * a12t'; */ 00075 FLA_Dotcs_external( FLA_CONJUGATE, FLA_ONE, a12t, a12t, FLA_ONE, alpha11 ); 00076 00077 /* a12t = a12t * triu( A22 )'; */ 00078 /* a12t' = triu( A22 ) * a12t'; */ 00079 FLA_Trmv_external( FLA_UPPER_TRIANGULAR, FLA_CONJ_NO_TRANSPOSE, FLA_NONUNIT_DIAG, 00080 A22, a12t ); 00081 } 00082 00083 /*------------------------------------------------------------*/ 00084 00085 FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, a01, /**/ A02, 00086 a10t, alpha11, /**/ a12t, 00087 /* ************** */ /* ************************ */ 00088 &ABL, /**/ &ABR, A20, a21, /**/ A22, 00089 FLA_TL ); 00090 00091 } 00092 00093 return FLA_SUCCESS; 00094 }