Go to the source code of this file.
Functions | |
FLA_Error | FLA_Trinv_l_blk_var1 (FLA_Diag diag, FLA_Obj A, fla_trinv_t *cntl) |
FLA_Error | FLA_Trinv_l_blk_var2 (FLA_Diag diag, FLA_Obj A, fla_trinv_t *cntl) |
FLA_Error | FLA_Trinv_l_blk_var3 (FLA_Diag diag, FLA_Obj A, fla_trinv_t *cntl) |
FLA_Error | FLA_Trinv_l_blk_var4 (FLA_Diag diag, FLA_Obj A, fla_trinv_t *cntl) |
FLA_Error | FLA_Trinv_l_unb_var1 (FLA_Diag diag, FLA_Obj A) |
FLA_Error | FLA_Trinv_l_unb_var2 (FLA_Diag diag, FLA_Obj A) |
FLA_Error | FLA_Trinv_l_unb_var3 (FLA_Diag diag, FLA_Obj A) |
FLA_Error | FLA_Trinv_l_unb_var4 (FLA_Diag diag, FLA_Obj A) |
FLA_Error FLA_Trinv_l_blk_var1 | ( | FLA_Diag | diag, | |
FLA_Obj | A, | |||
fla_trinv_t * | cntl | |||
) |
References FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_MINUS_ONE, FLA_Obj_length(), FLA_ONE, FLA_Part_2x2(), FLA_Repart_2x2_to_3x3(), FLA_Trinv_internal(), FLA_Trmm_internal(), and FLA_Trsm_internal().
Referenced by FLA_Trinv_l().
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 /* A10 = A10 * A00 */ 00061 FLA_Trmm_internal( FLA_RIGHT, FLA_LOWER_TRIANGULAR, FLA_NO_TRANSPOSE, diag, 00062 FLA_ONE, A00, A10, 00063 FLA_Cntl_sub_trmm( cntl ) ); 00064 00065 /* A10 = -triu( A11 ) \ A10 */ 00066 FLA_Trsm_internal( FLA_LEFT, FLA_LOWER_TRIANGULAR, FLA_NO_TRANSPOSE, diag, 00067 FLA_MINUS_ONE, A11, A10, 00068 FLA_Cntl_sub_trsm1( cntl ) ); 00069 00070 /* A11 = inv( A11 ) */ 00071 FLA_Trinv_internal( FLA_LOWER_TRIANGULAR, diag, A11, 00072 FLA_Cntl_sub_trinv( cntl ) ); 00073 00074 /*------------------------------------------------------------*/ 00075 00076 FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, A01, /**/ A02, 00077 A10, A11, /**/ A12, 00078 /* ************** */ /* ****************** */ 00079 &ABL, /**/ &ABR, A20, A21, /**/ A22, 00080 FLA_TL ); 00081 00082 } 00083 00084 return FLA_SUCCESS; 00085 }
FLA_Error FLA_Trinv_l_blk_var2 | ( | FLA_Diag | diag, | |
FLA_Obj | A, | |||
fla_trinv_t * | cntl | |||
) |
References FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_MINUS_ONE, FLA_Obj_length(), FLA_ONE, FLA_Part_2x2(), FLA_Repart_2x2_to_3x3(), FLA_Trinv_internal(), and FLA_Trsm_internal().
Referenced by FLA_Trinv_l().
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 /* A21 = triu( A22 ) \ A21 */ 00061 FLA_Trsm_internal( FLA_LEFT, FLA_LOWER_TRIANGULAR, FLA_NO_TRANSPOSE, diag, 00062 FLA_ONE, A22, A21, 00063 FLA_Cntl_sub_trsm1( cntl ) ); 00064 00065 /* A21 = -A21 / triu( A11 ) */ 00066 FLA_Trsm_internal( FLA_RIGHT, FLA_LOWER_TRIANGULAR, FLA_NO_TRANSPOSE, diag, 00067 FLA_MINUS_ONE, A11, A21, 00068 FLA_Cntl_sub_trsm2( cntl ) ); 00069 00070 /* A11 = inv( A11 ) */ 00071 FLA_Trinv_internal( FLA_LOWER_TRIANGULAR, diag, A11, 00072 FLA_Cntl_sub_trinv( cntl ) ); 00073 00074 /*------------------------------------------------------------*/ 00075 00076 FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, A01, /**/ A02, 00077 A10, A11, /**/ A12, 00078 /* ************** */ /* ****************** */ 00079 &ABL, /**/ &ABR, A20, A21, /**/ A22, 00080 FLA_TL ); 00081 00082 } 00083 00084 return FLA_SUCCESS; 00085 }
FLA_Error FLA_Trinv_l_blk_var3 | ( | FLA_Diag | diag, | |
FLA_Obj | A, | |||
fla_trinv_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(), FLA_Trinv_internal(), and FLA_Trsm_internal().
Referenced by FLA_Trinv_l().
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 /* A21 = -A21 / triu( A11 ) */ 00059 FLA_Trsm_internal( FLA_RIGHT, FLA_LOWER_TRIANGULAR, FLA_NO_TRANSPOSE, diag, 00060 FLA_MINUS_ONE, A11, A21, 00061 FLA_Cntl_sub_trsm1( cntl ) ); 00062 00063 /* A20 = A21 * A10 + A20 */ 00064 FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, 00065 FLA_ONE, A21, A10, FLA_ONE, A20, 00066 FLA_Cntl_sub_gemm( cntl ) ); 00067 00068 /* A10 = triu( A11 ) \ A10 */ 00069 FLA_Trsm_internal( FLA_LEFT, FLA_LOWER_TRIANGULAR, FLA_NO_TRANSPOSE, diag, 00070 FLA_ONE, A11, A10, 00071 FLA_Cntl_sub_trsm2( cntl ) ); 00072 00073 /* A11 = inv( A11 ) */ 00074 FLA_Trinv_internal( FLA_LOWER_TRIANGULAR, diag, A11, 00075 FLA_Cntl_sub_trinv( cntl ) ); 00076 00077 /*------------------------------------------------------------*/ 00078 00079 FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, A01, /**/ A02, 00080 A10, A11, /**/ A12, 00081 /* ************** */ /* ****************** */ 00082 &ABL, /**/ &ABR, A20, A21, /**/ A22, 00083 FLA_TL ); 00084 00085 } 00086 00087 return FLA_SUCCESS; 00088 }
FLA_Error FLA_Trinv_l_blk_var4 | ( | FLA_Diag | diag, | |
FLA_Obj | A, | |||
fla_trinv_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(), FLA_Trinv_internal(), FLA_Trmm_internal(), and FLA_Trsm_internal().
Referenced by FLA_Trinv_l().
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 /* A21 = -triu( A22 ) \ A21 */ 00061 FLA_Trsm_internal( FLA_LEFT, FLA_LOWER_TRIANGULAR, FLA_NO_TRANSPOSE, diag, 00062 FLA_MINUS_ONE, A22, A21, 00063 FLA_Cntl_sub_trsm1( cntl ) ); 00064 00065 /* A20 = -A21 * A10 + A20 */ 00066 FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, 00067 FLA_MINUS_ONE, A21, A10, FLA_ONE, A20, 00068 FLA_Cntl_sub_gemm( cntl ) ); 00069 00070 /* A10 = A10 * A00 */ 00071 FLA_Trmm_internal( FLA_RIGHT, FLA_LOWER_TRIANGULAR, FLA_NO_TRANSPOSE, diag, 00072 FLA_ONE, A00, A10, 00073 FLA_Cntl_sub_trmm( cntl ) ); 00074 00075 /* A11 = inv( A11 ) */ 00076 FLA_Trinv_internal( FLA_LOWER_TRIANGULAR, diag, A11, 00077 FLA_Cntl_sub_trinv( cntl ) ); 00078 00079 /*------------------------------------------------------------*/ 00080 00081 FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, A01, /**/ A02, 00082 A10, A11, /**/ A12, 00083 /* ************** */ /* ****************** */ 00084 &ABL, /**/ &ABR, A20, A21, /**/ A22, 00085 FLA_TL ); 00086 00087 } 00088 00089 return FLA_SUCCESS; 00090 }
References FLA_Cont_with_3x3_to_2x2(), FLA_Invert(), FLA_MINUS_ONE, FLA_Obj_length(), FLA_Part_2x2(), FLA_Repart_2x2_to_3x3(), FLA_Scal_external(), 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 /* alpha11 = 1.0 / alpha11 */ 00057 if ( diag == FLA_NONUNIT_DIAG ) 00058 FLA_Invert( alpha11 ); 00059 00060 /* a10t = -a10t * alpha11 */ 00061 FLA_Scal_external( FLA_MINUS_ONE, a10t ); 00062 if ( diag == FLA_NONUNIT_DIAG ) 00063 FLA_Scal_external( alpha11, a10t ); 00064 00065 /* a10t' = triu( A00' ) * a10t' */ 00066 FLA_Trmv_external( FLA_LOWER_TRIANGULAR, FLA_TRANSPOSE, diag, 00067 A00, a10t ); 00068 00069 /*------------------------------------------------------------*/ 00070 00071 FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, a01, /**/ A02, 00072 a10t, alpha11, /**/ a12t, 00073 /* ************** */ /* ************************ */ 00074 &ABL, /**/ &ABR, A20, a21, /**/ A22, 00075 FLA_TL ); 00076 00077 } 00078 00079 return FLA_SUCCESS; 00080 }
References FLA_Cont_with_3x3_to_2x2(), FLA_Invert(), FLA_MINUS_ONE, FLA_Obj_length(), FLA_Part_2x2(), FLA_Repart_2x2_to_3x3(), FLA_Scal_external(), and FLA_Trsv_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 /* alpha11 = 1.0 / alpha11 */ 00057 if ( diag == FLA_NONUNIT_DIAG ) 00058 FLA_Invert( alpha11 ); 00059 00060 /* a21 = -a21 * alpha11 */ 00061 FLA_Scal_external( FLA_MINUS_ONE, a21 ); 00062 if ( diag == FLA_NONUNIT_DIAG ) 00063 FLA_Scal_external( alpha11, a21 ); 00064 00065 /* a21 = triu( A22 ) \ a21 */ 00066 FLA_Trsv_external( FLA_LOWER_TRIANGULAR, FLA_NO_TRANSPOSE, diag, 00067 A22, a21 ); 00068 00069 /*------------------------------------------------------------*/ 00070 00071 FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, a01, /**/ A02, 00072 a10t, alpha11, /**/ a12t, 00073 /* ************** */ /* ************************ */ 00074 &ABL, /**/ &ABR, A20, a21, /**/ A22, 00075 FLA_TL ); 00076 00077 } 00078 00079 return FLA_SUCCESS; 00080 }
References FLA_Cont_with_3x3_to_2x2(), FLA_Ger_external(), FLA_Invert(), FLA_MINUS_ONE, 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 /* alpha11 = 1.0 / alpha11 */ 00057 if ( diag == FLA_NONUNIT_DIAG ) 00058 FLA_Invert( alpha11 ); 00059 00060 /* a21 = -a21 * alpha11 */ 00061 FLA_Scal_external( FLA_MINUS_ONE, a21 ); 00062 if ( diag == FLA_NONUNIT_DIAG ) 00063 FLA_Scal_external( alpha11, a21 ); 00064 00065 /* A20 = a21 * a10t + A20 */ 00066 FLA_Ger_external( FLA_ONE, a21, a10t, A20 ); 00067 00068 /* a10t = alpha11 * a10t */ 00069 if ( diag == FLA_NONUNIT_DIAG ) 00070 FLA_Scal_external( alpha11, a10t ); 00071 00072 /*------------------------------------------------------------*/ 00073 00074 FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, a01, /**/ A02, 00075 a10t, alpha11, /**/ a12t, 00076 /* ************** */ /* ************************ */ 00077 &ABL, /**/ &ABR, A20, a21, /**/ A22, 00078 FLA_TL ); 00079 00080 } 00081 00082 return FLA_SUCCESS; 00083 }
References FLA_Cont_with_3x3_to_2x2(), FLA_Ger_external(), FLA_Invert(), FLA_MINUS_ONE, FLA_Obj_length(), FLA_Part_2x2(), FLA_Repart_2x2_to_3x3(), FLA_Scal_external(), FLA_Trmv_external(), and FLA_Trsv_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 /* alpha11 = 1.0 / alpha11 */ 00057 if ( diag == FLA_NONUNIT_DIAG ) 00058 FLA_Invert( alpha11 ); 00059 00060 /* a21 = -triu( A22 ) \ a21 */ 00061 FLA_Scal_external( FLA_MINUS_ONE, a21 ); 00062 FLA_Trsv_external( FLA_LOWER_TRIANGULAR, FLA_NO_TRANSPOSE, diag, 00063 A22, a21 ); 00064 00065 /* A20 = -a21 * a10t + A20 */ 00066 FLA_Ger_external( FLA_MINUS_ONE, a21, a10t, A20 ); 00067 00068 /* a10t' = triu( A00' ) * a10t' */ 00069 FLA_Trmv_external( FLA_LOWER_TRIANGULAR, FLA_TRANSPOSE, diag, 00070 A00, a10t ); 00071 00072 /*------------------------------------------------------------*/ 00073 00074 FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, a01, /**/ A02, 00075 a10t, alpha11, /**/ a12t, 00076 /* ************** */ /* ************************ */ 00077 &ABL, /**/ &ABR, A20, a21, /**/ A22, 00078 FLA_TL ); 00079 00080 } 00081 00082 return FLA_SUCCESS; 00083 }