Go to the source code of this file.
Functions | |
FLA_Error | FLA_Trsm_run_blk_var1 (FLA_Diag diagA, FLA_Obj alpha, FLA_Obj A, FLA_Obj B, fla_trsm_t *cntl) |
FLA_Error | FLA_Trsm_run_blk_var2 (FLA_Diag diagA, FLA_Obj alpha, FLA_Obj A, FLA_Obj B, fla_trsm_t *cntl) |
FLA_Error | FLA_Trsm_run_blk_var3 (FLA_Diag diagA, FLA_Obj alpha, FLA_Obj A, FLA_Obj B, fla_trsm_t *cntl) |
FLA_Error | FLA_Trsm_run_blk_var4 (FLA_Diag diagA, FLA_Obj alpha, FLA_Obj A, FLA_Obj B, fla_trsm_t *cntl) |
FLA_Error | FLA_Trsm_run_unb_var1 (FLA_Diag diagA, FLA_Obj alpha, FLA_Obj A, FLA_Obj B) |
FLA_Error | FLA_Trsm_run_unb_var2 (FLA_Diag diagA, FLA_Obj alpha, FLA_Obj A, FLA_Obj B) |
FLA_Error | FLA_Trsm_run_unb_var3 (FLA_Diag diagA, FLA_Obj alpha, FLA_Obj A, FLA_Obj B) |
FLA_Error | FLA_Trsm_run_unb_var4 (FLA_Diag diagA, FLA_Obj alpha, FLA_Obj A, FLA_Obj B) |
FLA_Error FLA_Trsm_run_blk_var1 | ( | FLA_Diag | diagA, | |
FLA_Obj | alpha, | |||
FLA_Obj | A, | |||
FLA_Obj | B, | |||
fla_trsm_t * | cntl | |||
) |
References FLA_Cont_with_1x3_to_1x2(), FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_MINUS_ONE, FLA_Obj_length(), FLA_ONE, FLA_Part_1x2(), FLA_Part_2x2(), FLA_Repart_1x2_to_1x3(), FLA_Repart_2x2_to_3x3(), and FLA_Trsm_internal().
Referenced by FLA_Trsm_run().
00038 { 00039 FLA_Obj ATL, ATR, A00, A01, A02, 00040 ABL, ABR, A10, A11, A12, 00041 A20, A21, A22; 00042 00043 FLA_Obj BL, BR, B0, B1, B2; 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 while ( FLA_Obj_length( ATL ) < FLA_Obj_length( A ) ){ 00053 00054 b = FLA_Determine_blocksize( ABR, FLA_BR, FLA_Cntl_blocksize( cntl ) ); 00055 00056 FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, /**/ &A01, &A02, 00057 /* ************* */ /* ******************** */ 00058 &A10, /**/ &A11, &A12, 00059 ABL, /**/ ABR, &A20, /**/ &A21, &A22, 00060 b, b, FLA_BR ); 00061 00062 FLA_Repart_1x2_to_1x3( BL, /**/ BR, &B0, /**/ &B1, &B2, 00063 b, FLA_RIGHT ); 00064 00065 /*------------------------------------------------------------*/ 00066 00067 /* B1 = B1 - B0 * A01; */ 00068 FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, 00069 FLA_MINUS_ONE, B0, A01, FLA_ONE, B1, 00070 FLA_Cntl_sub_gemm( cntl ) ); 00071 00072 /* B1 = B1 / triu( A11 ); */ 00073 FLA_Trsm_internal( FLA_RIGHT, FLA_UPPER_TRIANGULAR, FLA_NO_TRANSPOSE, diagA, 00074 alpha, A11, B1, 00075 FLA_Cntl_sub_trsm( 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 FLA_Cont_with_1x3_to_1x2( &BL, /**/ &BR, B0, B1, /**/ B2, 00086 FLA_LEFT ); 00087 00088 } 00089 00090 return FLA_SUCCESS; 00091 }
FLA_Error FLA_Trsm_run_blk_var2 | ( | FLA_Diag | diagA, | |
FLA_Obj | alpha, | |||
FLA_Obj | A, | |||
FLA_Obj | B, | |||
fla_trsm_t * | cntl | |||
) |
References FLA_Cont_with_1x3_to_1x2(), FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_MINUS_ONE, FLA_Obj_length(), FLA_ONE, FLA_Part_1x2(), FLA_Part_2x2(), FLA_Repart_1x2_to_1x3(), FLA_Repart_2x2_to_3x3(), and FLA_Trsm_internal().
Referenced by FLA_Trsm_run().
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 dim_t b; 00044 00045 FLA_Part_2x2( A, &ATL, &ATR, 00046 &ABL, &ABR, 0, 0, FLA_TL ); 00047 00048 FLA_Part_1x2( B, &BL, &BR, 0, FLA_LEFT ); 00049 00050 while ( FLA_Obj_length( ATL ) < FLA_Obj_length( A ) ){ 00051 00052 b = FLA_Determine_blocksize( ABR, FLA_BR, FLA_Cntl_blocksize( cntl ) ); 00053 00054 FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, /**/ &A01, &A02, 00055 /* ************* */ /* ******************** */ 00056 &A10, /**/ &A11, &A12, 00057 ABL, /**/ ABR, &A20, /**/ &A21, &A22, 00058 b, b, FLA_BR ); 00059 00060 FLA_Repart_1x2_to_1x3( BL, /**/ BR, &B0, /**/ &B1, &B2, 00061 b, FLA_RIGHT ); 00062 00063 /*------------------------------------------------------------*/ 00064 00065 /* B1 = B1 / triu( A11 ); */ 00066 FLA_Trsm_internal( FLA_RIGHT, FLA_UPPER_TRIANGULAR, FLA_NO_TRANSPOSE, diagA, 00067 alpha, A11, B1, 00068 FLA_Cntl_sub_trsm( cntl ) ); 00069 00070 /* B2 = B2 - B1 * A12; */ 00071 FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, 00072 FLA_MINUS_ONE, B1, A12, FLA_ONE, B2, 00073 FLA_Cntl_sub_gemm( cntl ) ); 00074 00075 /*------------------------------------------------------------*/ 00076 00077 FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, A01, /**/ A02, 00078 A10, A11, /**/ A12, 00079 /* ************** */ /* ****************** */ 00080 &ABL, /**/ &ABR, A20, A21, /**/ A22, 00081 FLA_TL ); 00082 00083 FLA_Cont_with_1x3_to_1x2( &BL, /**/ &BR, B0, B1, /**/ B2, 00084 FLA_LEFT ); 00085 00086 } 00087 00088 return FLA_SUCCESS; 00089 }
FLA_Error FLA_Trsm_run_blk_var3 | ( | FLA_Diag | diagA, | |
FLA_Obj | alpha, | |||
FLA_Obj | A, | |||
FLA_Obj | B, | |||
fla_trsm_t * | cntl | |||
) |
References FLA_Cont_with_3x1_to_2x1(), FLA_Determine_blocksize(), FLA_Obj_length(), FLA_Part_2x1(), FLA_Repart_2x1_to_3x1(), and FLA_Trsm_internal().
Referenced by FLA_Trsm_run().
00036 { 00037 FLA_Obj BT, B0, 00038 BB, B1, 00039 B2; 00040 00041 dim_t b; 00042 00043 FLA_Part_2x1( B, &BT, 00044 &BB, 0, FLA_TOP ); 00045 00046 while ( FLA_Obj_length( BT ) < FLA_Obj_length( B ) ){ 00047 00048 b = FLA_Determine_blocksize( BB, FLA_BOTTOM, FLA_Cntl_blocksize( cntl ) ); 00049 00050 FLA_Repart_2x1_to_3x1( BT, &B0, 00051 /* ** */ /* ** */ 00052 &B1, 00053 BB, &B2, b, FLA_BOTTOM ); 00054 00055 /*------------------------------------------------------------*/ 00056 00057 /* B1 = B1 / triu( A); */ 00058 FLA_Trsm_internal( FLA_RIGHT, FLA_UPPER_TRIANGULAR, FLA_NO_TRANSPOSE, diagA, 00059 alpha, A, B1, 00060 FLA_Cntl_sub_trsm( cntl ) ); 00061 00062 /*------------------------------------------------------------*/ 00063 00064 FLA_Cont_with_3x1_to_2x1( &BT, B0, 00065 B1, 00066 /* ** */ /* ** */ 00067 &BB, B2, FLA_TOP ); 00068 00069 } 00070 00071 return FLA_SUCCESS; 00072 }
FLA_Error FLA_Trsm_run_blk_var4 | ( | FLA_Diag | diagA, | |
FLA_Obj | alpha, | |||
FLA_Obj | A, | |||
FLA_Obj | B, | |||
fla_trsm_t * | cntl | |||
) |
References FLA_Cont_with_3x1_to_2x1(), FLA_Determine_blocksize(), FLA_Obj_length(), FLA_Part_2x1(), FLA_Repart_2x1_to_3x1(), and FLA_Trsm_internal().
Referenced by FLA_Trsm_run().
00038 { 00039 FLA_Obj BT, B0, 00040 BB, B1, 00041 B2; 00042 00043 dim_t b; 00044 00045 FLA_Part_2x1( B, &BT, 00046 &BB, 0, FLA_BOTTOM ); 00047 00048 while ( FLA_Obj_length( BB ) < FLA_Obj_length( B ) ){ 00049 00050 b = FLA_Determine_blocksize( BT, FLA_TOP, FLA_Cntl_blocksize( cntl ) ); 00051 00052 FLA_Repart_2x1_to_3x1( BT, &B0, 00053 &B1, 00054 /* ** */ /* ** */ 00055 BB, &B2, b, FLA_TOP ); 00056 00057 /*------------------------------------------------------------*/ 00058 00059 /* B1 = B1 / triu( A ); */ 00060 FLA_Trsm_internal( FLA_RIGHT, FLA_UPPER_TRIANGULAR, FLA_NO_TRANSPOSE, diagA, 00061 alpha, A, B1, 00062 FLA_Cntl_sub_trsm( cntl ) ); 00063 00064 /*------------------------------------------------------------*/ 00065 00066 FLA_Cont_with_3x1_to_2x1( &BT, B0, 00067 /* ** */ /* ** */ 00068 B1, 00069 &BB, B2, FLA_BOTTOM ); 00070 00071 } 00072 00073 return FLA_SUCCESS; 00074 }
References FLA_Cont_with_1x3_to_1x2(), FLA_Cont_with_3x3_to_2x2(), FLA_Gemv_external(), FLA_Inv_scal_external(), FLA_MINUS_ONE, FLA_Obj_length(), FLA_ONE, FLA_Part_1x2(), FLA_Part_2x2(), FLA_Repart_1x2_to_1x3(), 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_Obj BL, BR, B0, b1, B2; 00044 00045 FLA_Scal_external( alpha, 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 while ( FLA_Obj_length( ATL ) < FLA_Obj_length( A ) ){ 00053 00054 FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, /**/ &a01, &A02, 00055 /* ************* */ /* ************************** */ 00056 &a10t, /**/ &alpha11, &a12t, 00057 ABL, /**/ ABR, &A20, /**/ &a21, &A22, 00058 1, 1, FLA_BR ); 00059 00060 FLA_Repart_1x2_to_1x3( BL, /**/ BR, &B0, /**/ &b1, &B2, 00061 1, FLA_RIGHT ); 00062 00063 /*------------------------------------------------------------*/ 00064 00065 /* b1 = b1 - B0 * a01; */ 00066 FLA_Gemv_external( FLA_NO_TRANSPOSE, FLA_MINUS_ONE, B0, a01, FLA_ONE, b1 ); 00067 00068 /* b1 = b1 / alpha11; */ 00069 if ( diagA != FLA_UNIT_DIAG ) 00070 FLA_Inv_scal_external( alpha11, b1 ); 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 FLA_Cont_with_1x3_to_1x2( &BL, /**/ &BR, B0, b1, /**/ B2, 00081 FLA_LEFT ); 00082 00083 } 00084 00085 return FLA_SUCCESS; 00086 }
References FLA_Cont_with_1x3_to_1x2(), FLA_Cont_with_3x3_to_2x2(), FLA_Ger_external(), FLA_Inv_scal_external(), FLA_MINUS_ONE, FLA_Obj_length(), FLA_Part_1x2(), FLA_Part_2x2(), FLA_Repart_1x2_to_1x3(), 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_Obj BL, BR, B0, b1, B2; 00044 00045 FLA_Scal_external( alpha, 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 while ( FLA_Obj_length( ATL ) < FLA_Obj_length( A ) ){ 00053 00054 FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, /**/ &a01, &A02, 00055 /* ************* */ /* ************************** */ 00056 &a10t, /**/ &alpha11, &a12t, 00057 ABL, /**/ ABR, &A20, /**/ &a21, &A22, 00058 1, 1, FLA_BR ); 00059 00060 FLA_Repart_1x2_to_1x3( BL, /**/ BR, &B0, /**/ &b1, &B2, 00061 1, FLA_RIGHT ); 00062 00063 /*------------------------------------------------------------*/ 00064 00065 /* b1 = b1 / alpha11; */ 00066 if ( diagA != FLA_UNIT_DIAG ) 00067 FLA_Inv_scal_external( alpha11, b1 ); 00068 00069 /* B2 = B2 - b1 * a12t; */ 00070 FLA_Ger_external( FLA_MINUS_ONE, b1, a12t, B2 ); 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 FLA_Cont_with_1x3_to_1x2( &BL, /**/ &BR, B0, b1, /**/ B2, 00081 FLA_LEFT ); 00082 00083 } 00084 00085 return FLA_SUCCESS; 00086 }
References FLA_Cont_with_3x1_to_2x1(), FLA_Obj_length(), FLA_Part_2x1(), FLA_Repart_2x1_to_3x1(), FLA_Scal_external(), and FLA_Trsv_external().
00038 { 00039 FLA_Obj BT, B0, 00040 BB, b1t, 00041 B2; 00042 00043 FLA_Scal_external( alpha, B ); 00044 00045 FLA_Part_2x1( B, &BT, 00046 &BB, 0, FLA_TOP ); 00047 00048 while ( FLA_Obj_length( BT ) < FLA_Obj_length( B ) ){ 00049 00050 FLA_Repart_2x1_to_3x1( BT, &B0, 00051 /* ** */ /* *** */ 00052 &b1t, 00053 BB, &B2, 1, FLA_BOTTOM ); 00054 00055 /*------------------------------------------------------------*/ 00056 00057 /* b1t = b1t / triu( A); */ 00058 FLA_Trsv_external( FLA_UPPER_TRIANGULAR, FLA_TRANSPOSE, diagA, A, b1t ); 00059 00060 /*------------------------------------------------------------*/ 00061 00062 FLA_Cont_with_3x1_to_2x1( &BT, B0, 00063 b1t, 00064 /* ** */ /* *** */ 00065 &BB, B2, FLA_TOP ); 00066 00067 } 00068 00069 return FLA_SUCCESS; 00070 }
References FLA_Cont_with_3x1_to_2x1(), FLA_Obj_length(), FLA_Part_2x1(), FLA_Repart_2x1_to_3x1(), FLA_Scal_external(), and FLA_Trsv_external().
00038 { 00039 FLA_Obj BT, B0, 00040 BB, b1t, 00041 B2; 00042 00043 FLA_Scal_external( alpha, B ); 00044 00045 FLA_Part_2x1( B, &BT, 00046 &BB, 0, FLA_BOTTOM ); 00047 00048 while ( FLA_Obj_length( BB ) < FLA_Obj_length( B ) ){ 00049 00050 FLA_Repart_2x1_to_3x1( BT, &B0, 00051 &b1t, 00052 /* ** */ /* *** */ 00053 BB, &B2, 1, FLA_TOP ); 00054 00055 /*------------------------------------------------------------*/ 00056 00057 /* b1t = b1t / triu( A); */ 00058 FLA_Trsv_external( FLA_UPPER_TRIANGULAR, FLA_TRANSPOSE, diagA, A, b1t ); 00059 00060 /*------------------------------------------------------------*/ 00061 00062 FLA_Cont_with_3x1_to_2x1( &BT, B0, 00063 /* ** */ /* *** */ 00064 b1t, 00065 &BB, B2, FLA_BOTTOM ); 00066 00067 } 00068 00069 return FLA_SUCCESS; 00070 }