FLA_Syrk_ut.h File Reference

(r)

Go to the source code of this file.

Functions

FLA_Error FLA_Syrk_ut_blk_var1 (FLA_Obj alpha, FLA_Obj A, FLA_Obj beta, FLA_Obj C, fla_syrk_t *cntl)
FLA_Error FLA_Syrk_ut_blk_var2 (FLA_Obj alpha, FLA_Obj A, FLA_Obj beta, FLA_Obj C, fla_syrk_t *cntl)
FLA_Error FLA_Syrk_ut_blk_var3 (FLA_Obj alpha, FLA_Obj A, FLA_Obj beta, FLA_Obj C, fla_syrk_t *cntl)
FLA_Error FLA_Syrk_ut_blk_var4 (FLA_Obj alpha, FLA_Obj A, FLA_Obj beta, FLA_Obj C, fla_syrk_t *cntl)
FLA_Error FLA_Syrk_ut_blk_var5 (FLA_Obj alpha, FLA_Obj A, FLA_Obj beta, FLA_Obj C, fla_syrk_t *cntl)
FLA_Error FLA_Syrk_ut_blk_var6 (FLA_Obj alpha, FLA_Obj A, FLA_Obj beta, FLA_Obj C, fla_syrk_t *cntl)
FLA_Error FLA_Syrk_ut_unb_var1 (FLA_Obj alpha, FLA_Obj A, FLA_Obj beta, FLA_Obj C)
FLA_Error FLA_Syrk_ut_unb_var2 (FLA_Obj alpha, FLA_Obj A, FLA_Obj beta, FLA_Obj C)
FLA_Error FLA_Syrk_ut_unb_var3 (FLA_Obj alpha, FLA_Obj A, FLA_Obj beta, FLA_Obj C)
FLA_Error FLA_Syrk_ut_unb_var4 (FLA_Obj alpha, FLA_Obj A, FLA_Obj beta, FLA_Obj C)
FLA_Error FLA_Syrk_ut_unb_var5 (FLA_Obj alpha, FLA_Obj A, FLA_Obj beta, FLA_Obj C)
FLA_Error FLA_Syrk_ut_unb_var6 (FLA_Obj alpha, FLA_Obj A, FLA_Obj beta, FLA_Obj C)


Function Documentation

FLA_Error FLA_Syrk_ut_blk_var1 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  beta,
FLA_Obj  C,
fla_syrk_t cntl 
)

References FLA_Cont_with_1x3_to_1x2(), FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_Obj_width(), FLA_Part_1x2(), FLA_Part_2x2(), FLA_Repart_1x2_to_1x3(), FLA_Repart_2x2_to_3x3(), and FLA_Syrk_internal().

Referenced by FLA_Syrk_ut().

00038 {
00039   FLA_Obj AL,    AR,       A0,  A1,  A2;
00040 
00041   FLA_Obj CTL,   CTR,      C00, C01, C02, 
00042           CBL,   CBR,      C10, C11, C12,
00043                            C20, C21, C22;
00044 
00045   dim_t b;
00046 
00047   FLA_Part_1x2( A,    &AL,  &AR,      0, FLA_LEFT );
00048 
00049   FLA_Part_2x2( C,    &CTL, &CTR,
00050                       &CBL, &CBR,     0, 0, FLA_TL );
00051 
00052   while ( FLA_Obj_width( AL ) < FLA_Obj_width( A ) ){
00053 
00054     b = FLA_Determine_blocksize( AR, FLA_RIGHT, FLA_Cntl_blocksize( cntl ) );
00055 
00056     FLA_Repart_1x2_to_1x3( AL,  /**/ AR,        &A0, /**/ &A1, &A2,
00057                            b, FLA_RIGHT );
00058 
00059     FLA_Repart_2x2_to_3x3( CTL, /**/ CTR,       &C00, /**/ &C01, &C02,
00060                         /* ************* */   /* ******************** */
00061                                                 &C10, /**/ &C11, &C12,
00062                            CBL, /**/ CBR,       &C20, /**/ &C21, &C22,
00063                            b, b, FLA_BR );
00064 
00065     /*------------------------------------------------------------*/
00066 
00067     /* C12 = C12 + A1' * A2 */
00068     FLA_Gemm_internal( FLA_TRANSPOSE, FLA_NO_TRANSPOSE, 
00069                        alpha, A1, A2, beta, C12,
00070                        FLA_Cntl_sub_gemm( cntl ) );
00071 
00072     /* C11 = C11 + A1' * A1 */
00073     FLA_Syrk_internal( FLA_UPPER_TRIANGULAR, FLA_TRANSPOSE, 
00074                        alpha, A1, beta, C11,
00075                        FLA_Cntl_sub_syrk( cntl ) );
00076 
00077     /*------------------------------------------------------------*/
00078 
00079     FLA_Cont_with_1x3_to_1x2( &AL,  /**/ &AR,        A0, A1, /**/ A2,
00080                               FLA_LEFT );
00081 
00082     FLA_Cont_with_3x3_to_2x2( &CTL, /**/ &CTR,       C00, C01, /**/ C02,
00083                                                      C10, C11, /**/ C12,
00084                             /* ************** */  /* ****************** */
00085                               &CBL, /**/ &CBR,       C20, C21, /**/ C22,
00086                               FLA_TL );
00087 
00088   }
00089 
00090   return FLA_SUCCESS;
00091 }

FLA_Error FLA_Syrk_ut_blk_var2 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  beta,
FLA_Obj  C,
fla_syrk_t cntl 
)

References FLA_Cont_with_1x3_to_1x2(), FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_Obj_width(), FLA_Part_1x2(), FLA_Part_2x2(), FLA_Repart_1x2_to_1x3(), FLA_Repart_2x2_to_3x3(), and FLA_Syrk_internal().

Referenced by FLA_Syrk_ut().

00036 {
00037   FLA_Obj AL,    AR,       A0,  A1,  A2;
00038 
00039   FLA_Obj CTL,   CTR,      C00, C01, C02, 
00040           CBL,   CBR,      C10, C11, C12,
00041                            C20, C21, C22;
00042 
00043   dim_t b;
00044 
00045   FLA_Part_1x2( A,    &AL,  &AR,      0, FLA_LEFT );
00046 
00047   FLA_Part_2x2( C,    &CTL, &CTR,
00048                       &CBL, &CBR,     0, 0, FLA_TL );
00049 
00050   while ( FLA_Obj_width( AL ) < FLA_Obj_width( A ) ){
00051 
00052     b = FLA_Determine_blocksize( AR, FLA_RIGHT, FLA_Cntl_blocksize( cntl ) );
00053 
00054     FLA_Repart_1x2_to_1x3( AL,  /**/ AR,        &A0, /**/ &A1, &A2,
00055                            b, FLA_RIGHT );
00056 
00057     FLA_Repart_2x2_to_3x3( CTL, /**/ CTR,       &C00, /**/ &C01, &C02,
00058                         /* ************* */   /* ******************** */
00059                                                 &C10, /**/ &C11, &C12,
00060                            CBL, /**/ CBR,       &C20, /**/ &C21, &C22,
00061                            b, b, FLA_BR );
00062 
00063     /*------------------------------------------------------------*/
00064 
00065     /* C01 = C01 + A0' * A1 */
00066     FLA_Gemm_internal( FLA_TRANSPOSE, FLA_NO_TRANSPOSE, 
00067                        alpha, A0, A1, beta, C01,
00068                        FLA_Cntl_sub_gemm( cntl ) );
00069 
00070     /* C11 = C11 + A1' * A1 */
00071     FLA_Syrk_internal( FLA_UPPER_TRIANGULAR, FLA_TRANSPOSE, 
00072                        alpha, A1, beta, C11,
00073                        FLA_Cntl_sub_syrk( cntl ) );
00074 
00075     /*------------------------------------------------------------*/
00076 
00077     FLA_Cont_with_1x3_to_1x2( &AL,  /**/ &AR,        A0, A1, /**/ A2,
00078                               FLA_LEFT );
00079 
00080     FLA_Cont_with_3x3_to_2x2( &CTL, /**/ &CTR,       C00, C01, /**/ C02,
00081                                                      C10, C11, /**/ C12,
00082                             /* ************** */  /* ****************** */
00083                               &CBL, /**/ &CBR,       C20, C21, /**/ C22,
00084                               FLA_TL );
00085 
00086   }
00087 
00088   return FLA_SUCCESS;
00089 }

FLA_Error FLA_Syrk_ut_blk_var3 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  beta,
FLA_Obj  C,
fla_syrk_t cntl 
)

References FLA_Cont_with_1x3_to_1x2(), FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_Obj_width(), FLA_Part_1x2(), FLA_Part_2x2(), FLA_Repart_1x2_to_1x3(), FLA_Repart_2x2_to_3x3(), and FLA_Syrk_internal().

Referenced by FLA_Syrk_ut().

00038 {
00039   FLA_Obj AL,    AR,       A0,  A1,  A2;
00040 
00041   FLA_Obj CTL,   CTR,      C00, C01, C02, 
00042           CBL,   CBR,      C10, C11, C12,
00043                            C20, C21, C22;
00044 
00045   dim_t b;
00046 
00047   FLA_Part_1x2( A,    &AL,  &AR,      0, FLA_RIGHT );
00048 
00049   FLA_Part_2x2( C,    &CTL, &CTR,
00050                       &CBL, &CBR,     0, 0, FLA_BR );
00051 
00052   while ( FLA_Obj_width( AR ) < FLA_Obj_width( A ) ){
00053 
00054     b = FLA_Determine_blocksize( AL, FLA_LEFT, FLA_Cntl_blocksize( cntl ) );
00055 
00056     FLA_Repart_1x2_to_1x3( AL,  /**/ AR,        &A0, &A1, /**/ &A2,
00057                            b, FLA_LEFT );
00058 
00059     FLA_Repart_2x2_to_3x3( CTL, /**/ CTR,       &C00, &C01, /**/ &C02,
00060                                                 &C10, &C11, /**/ &C12,
00061                         /* ************* */   /* ******************** */
00062                            CBL, /**/ CBR,       &C20, &C21, /**/ &C22,
00063                            b, b, FLA_TL );
00064 
00065     /*------------------------------------------------------------*/
00066 
00067     /* C12 = C12 + A1' * A2 */
00068     FLA_Gemm_internal( FLA_TRANSPOSE, FLA_NO_TRANSPOSE, 
00069                        alpha, A1, A2, beta, C12,
00070                        FLA_Cntl_sub_gemm( cntl ) );
00071 
00072     /* C11 = C11 + A1' * A1 */
00073     FLA_Syrk_internal( FLA_UPPER_TRIANGULAR, FLA_TRANSPOSE, 
00074                        alpha, A1, beta, C11,
00075                        FLA_Cntl_sub_syrk( cntl ) );
00076 
00077     /*------------------------------------------------------------*/
00078 
00079     FLA_Cont_with_1x3_to_1x2( &AL,  /**/ &AR,        A0, /**/ A1, A2,
00080                               FLA_RIGHT );
00081 
00082     FLA_Cont_with_3x3_to_2x2( &CTL, /**/ &CTR,       C00, /**/ C01, C02,
00083                             /* ************** */  /* ****************** */
00084                                                      C10, /**/ C11, C12,
00085                               &CBL, /**/ &CBR,       C20, /**/ C21, C22,
00086                               FLA_BR );
00087 
00088   }
00089 
00090   return FLA_SUCCESS;
00091 }

FLA_Error FLA_Syrk_ut_blk_var4 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  beta,
FLA_Obj  C,
fla_syrk_t cntl 
)

References FLA_Cont_with_1x3_to_1x2(), FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_Obj_width(), FLA_Part_1x2(), FLA_Part_2x2(), FLA_Repart_1x2_to_1x3(), FLA_Repart_2x2_to_3x3(), and FLA_Syrk_internal().

Referenced by FLA_Syrk_ut().

00038 {
00039   FLA_Obj AL,    AR,       A0,  A1,  A2;
00040 
00041   FLA_Obj CTL,   CTR,      C00, C01, C02, 
00042           CBL,   CBR,      C10, C11, C12,
00043                            C20, C21, C22;
00044 
00045   dim_t b;
00046 
00047   FLA_Part_1x2( A,    &AL,  &AR,      0, FLA_RIGHT );
00048 
00049   FLA_Part_2x2( C,    &CTL, &CTR,
00050                       &CBL, &CBR,     0, 0, FLA_BR );
00051 
00052   while ( FLA_Obj_width( AR ) < FLA_Obj_width( A ) ){
00053 
00054     b = FLA_Determine_blocksize( AL, FLA_LEFT, FLA_Cntl_blocksize( cntl ) );
00055 
00056     FLA_Repart_1x2_to_1x3( AL,  /**/ AR,        &A0, &A1, /**/ &A2,
00057                            b, FLA_LEFT );
00058 
00059     FLA_Repart_2x2_to_3x3( CTL, /**/ CTR,       &C00, &C01, /**/ &C02,
00060                                                 &C10, &C11, /**/ &C12,
00061                         /* ************* */   /* ******************** */
00062                            CBL, /**/ CBR,       &C20, &C21, /**/ &C22,
00063                            b, b, FLA_TL );
00064 
00065     /*------------------------------------------------------------*/
00066 
00067     /* C01 = C01 + A0' * A1 */
00068     FLA_Gemm_internal( FLA_TRANSPOSE, FLA_NO_TRANSPOSE, 
00069                        alpha, A0, A1, beta, C01,
00070                        FLA_Cntl_sub_gemm( cntl ) );
00071 
00072     /* C11 = C11 + A1' * A1 */
00073     FLA_Syrk_internal( FLA_UPPER_TRIANGULAR, FLA_TRANSPOSE, 
00074                        alpha, A1, beta, C11,
00075                        FLA_Cntl_sub_syrk( cntl ) );
00076 
00077     /*------------------------------------------------------------*/
00078 
00079     FLA_Cont_with_1x3_to_1x2( &AL,  /**/ &AR,        A0, /**/ A1, A2,
00080                               FLA_RIGHT );
00081 
00082     FLA_Cont_with_3x3_to_2x2( &CTL, /**/ &CTR,       C00, /**/ C01, C02,
00083                             /* ************** */  /* ****************** */
00084                                                      C10, /**/ C11, C12,
00085                               &CBL, /**/ &CBR,       C20, /**/ C21, C22,
00086                               FLA_BR );
00087 
00088   }
00089 
00090   return FLA_SUCCESS;
00091 }

FLA_Error FLA_Syrk_ut_blk_var5 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  beta,
FLA_Obj  C,
fla_syrk_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_Syrk_internal().

Referenced by FLA_Syrk_ut().

00036 {
00037   FLA_Obj AT,              A0,
00038           AB,              A1,
00039                            A2;
00040 
00041   dim_t b;
00042 
00043   FLA_Part_2x1( A,    &AT, 
00044                       &AB,            0, FLA_TOP );
00045 
00046   while ( FLA_Obj_length( AT ) < FLA_Obj_length( A ) ){
00047 
00048     b = FLA_Determine_blocksize( AB, FLA_BOTTOM, FLA_Cntl_blocksize( cntl ) );
00049 
00050     FLA_Repart_2x1_to_3x1( AT,                &A0, 
00051                         /* ** */            /* ** */
00052                                               &A1, 
00053                            AB,                &A2,        b, FLA_BOTTOM );
00054 
00055     /*------------------------------------------------------------*/
00056 
00057     /* C = C + A1' * A1 */
00058     FLA_Syrk_internal( FLA_UPPER_TRIANGULAR, FLA_TRANSPOSE, 
00059                        alpha, A1, beta, C,
00060                        FLA_Cntl_sub_syrk( cntl ) );
00061 
00062     /*------------------------------------------------------------*/
00063 
00064     FLA_Cont_with_3x1_to_2x1( &AT,                A0, 
00065                                                   A1, 
00066                             /* ** */           /* ** */
00067                               &AB,                A2,     FLA_TOP );
00068 
00069   }
00070 
00071   return FLA_SUCCESS;
00072 }

FLA_Error FLA_Syrk_ut_blk_var6 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  beta,
FLA_Obj  C,
fla_syrk_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_Syrk_internal().

Referenced by FLA_Syrk_ut().

00038 {
00039   FLA_Obj AT,              A0,
00040           AB,              A1,
00041                            A2;
00042 
00043   dim_t b;
00044 
00045   FLA_Part_2x1( A,    &AT, 
00046                       &AB,            0, FLA_BOTTOM );
00047 
00048   while ( FLA_Obj_length( AB ) < FLA_Obj_length( A ) ){
00049 
00050     b = FLA_Determine_blocksize( AT, FLA_TOP, FLA_Cntl_blocksize( cntl ) );
00051 
00052     FLA_Repart_2x1_to_3x1( AT,                &A0, 
00053                                               &A1, 
00054                         /* ** */            /* ** */
00055                            AB,                &A2,        b, FLA_TOP );
00056 
00057     /*------------------------------------------------------------*/
00058 
00059     /* C = C + A1' * A1 */
00060     FLA_Syrk_internal( FLA_UPPER_TRIANGULAR, FLA_TRANSPOSE, 
00061                        alpha, A1, beta, C,
00062                        FLA_Cntl_sub_syrk( cntl ) );
00063 
00064     /*------------------------------------------------------------*/
00065 
00066     FLA_Cont_with_3x1_to_2x1( &AT,                A0, 
00067                             /* ** */           /* ** */
00068                                                   A1, 
00069                               &AB,                A2,     FLA_BOTTOM );
00070 
00071   }
00072 
00073   return FLA_SUCCESS;
00074 }

FLA_Error FLA_Syrk_ut_unb_var1 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  beta,
FLA_Obj  C 
)

References FLA_Cont_with_1x3_to_1x2(), FLA_Cont_with_3x3_to_2x2(), FLA_Dots_external(), FLA_Gemv_external(), FLA_Obj_width(), FLA_ONE, FLA_Part_1x2(), FLA_Part_2x2(), FLA_Repart_1x2_to_1x3(), FLA_Repart_2x2_to_3x3(), and FLA_Scalr_external().

00038 {
00039   FLA_Obj AL,    AR,       A0,  a1,  A2;
00040 
00041   FLA_Obj CTL,   CTR,      C00,  c01,     C02, 
00042           CBL,   CBR,      c10t, gamma11, c12t,
00043                            C20,  c21,     C22;
00044 
00045   FLA_Scalr_external( FLA_UPPER_TRIANGULAR, beta, C );
00046 
00047   FLA_Part_1x2( A,    &AL,  &AR,      0, FLA_LEFT );
00048 
00049   FLA_Part_2x2( C,    &CTL, &CTR,
00050                       &CBL, &CBR,     0, 0, FLA_TL );
00051 
00052   while ( FLA_Obj_width( AL ) < FLA_Obj_width( A ) ){
00053 
00054     FLA_Repart_1x2_to_1x3( AL,  /**/ AR,        &A0, /**/ &a1, &A2,
00055                            1, FLA_RIGHT );
00056 
00057     FLA_Repart_2x2_to_3x3( CTL, /**/ CTR,       &C00,  /**/ &c01,     &C02,
00058                         /* ************* */   /* ************************** */
00059                                                 &c10t, /**/ &gamma11, &c12t,
00060                            CBL, /**/ CBR,       &C20,  /**/ &c21,     &C22,
00061                            1, 1, FLA_BR );
00062 
00063     /*------------------------------------------------------------*/
00064 
00065     /* c12t    = c12t    + A2' * a1 */
00066     FLA_Gemv_external( FLA_TRANSPOSE, alpha, A2, a1, FLA_ONE, c12t );
00067 
00068     /* gamma11 = gamma11 + a1 * a1' */
00069     FLA_Dots_external( alpha, a1, a1, FLA_ONE, gamma11 );
00070 
00071     /*------------------------------------------------------------*/
00072 
00073     FLA_Cont_with_1x3_to_1x2( &AL,  /**/ &AR,        A0, a1, /**/ A2,
00074                               FLA_LEFT );
00075 
00076     FLA_Cont_with_3x3_to_2x2( &CTL, /**/ &CTR,       C00,  c01,     /**/ C02,
00077                                                      c10t, gamma11, /**/ c12t,
00078                             /* ************** */  /* ************************ */
00079                               &CBL, /**/ &CBR,       C20,  c21,     /**/ C22,
00080                               FLA_TL );
00081 
00082   }
00083 
00084   return FLA_SUCCESS;
00085 }

FLA_Error FLA_Syrk_ut_unb_var2 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  beta,
FLA_Obj  C 
)

References FLA_Cont_with_1x3_to_1x2(), FLA_Cont_with_3x3_to_2x2(), FLA_Dots_external(), FLA_Gemv_external(), FLA_Obj_width(), FLA_ONE, FLA_Part_1x2(), FLA_Part_2x2(), FLA_Repart_1x2_to_1x3(), FLA_Repart_2x2_to_3x3(), and FLA_Scalr_external().

00038 {
00039   FLA_Obj AL,    AR,       A0,  a1,  A2;
00040 
00041   FLA_Obj CTL,   CTR,      C00,  c01,     C02, 
00042           CBL,   CBR,      c10t, gamma11, c12t,
00043                            C20,  c21,     C22;
00044 
00045   FLA_Scalr_external( FLA_UPPER_TRIANGULAR, beta, C );
00046 
00047   FLA_Part_1x2( A,    &AL,  &AR,      0, FLA_LEFT );
00048 
00049   FLA_Part_2x2( C,    &CTL, &CTR,
00050                       &CBL, &CBR,     0, 0, FLA_TL );
00051 
00052   while ( FLA_Obj_width( AL ) < FLA_Obj_width( A ) ){
00053 
00054     FLA_Repart_1x2_to_1x3( AL,  /**/ AR,        &A0, /**/ &a1, &A2,
00055                            1, FLA_RIGHT );
00056 
00057     FLA_Repart_2x2_to_3x3( CTL, /**/ CTR,       &C00,  /**/ &c01,     &C02,
00058                         /* ************* */   /* ************************** */
00059                                                 &c10t, /**/ &gamma11, &c12t,
00060                            CBL, /**/ CBR,       &C20,  /**/ &c21,     &C22,
00061                            1, 1, FLA_BR );
00062 
00063     /*------------------------------------------------------------*/
00064 
00065     /* c01     = c01     + A0' * a1 */
00066     FLA_Gemv_external( FLA_TRANSPOSE, alpha, A0, a1, FLA_ONE, c01 );
00067 
00068     /* gamma11 = gamma11 + a1 * a1' */
00069     FLA_Dots_external( alpha, a1, a1, FLA_ONE, gamma11 );
00070 
00071     /*------------------------------------------------------------*/
00072 
00073     FLA_Cont_with_1x3_to_1x2( &AL,  /**/ &AR,        A0, a1, /**/ A2,
00074                               FLA_LEFT );
00075 
00076     FLA_Cont_with_3x3_to_2x2( &CTL, /**/ &CTR,       C00,  c01,     /**/ C02,
00077                                                      c10t, gamma11, /**/ c12t,
00078                             /* ************** */  /* ************************ */
00079                               &CBL, /**/ &CBR,       C20,  c21,     /**/ C22,
00080                               FLA_TL );
00081 
00082   }
00083 
00084   return FLA_SUCCESS;
00085 }

FLA_Error FLA_Syrk_ut_unb_var3 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  beta,
FLA_Obj  C 
)

References FLA_Cont_with_1x3_to_1x2(), FLA_Cont_with_3x3_to_2x2(), FLA_Dots_external(), FLA_Gemv_external(), FLA_Obj_width(), FLA_ONE, FLA_Part_1x2(), FLA_Part_2x2(), FLA_Repart_1x2_to_1x3(), FLA_Repart_2x2_to_3x3(), and FLA_Scalr_external().

00038 {
00039   FLA_Obj AL,    AR,       A0,  a1,  A2;
00040 
00041   FLA_Obj CTL,   CTR,      C00,  c01,     C02, 
00042           CBL,   CBR,      c10t, gamma11, c12t,
00043                            C20,  c21,     C22;
00044 
00045   FLA_Scalr_external( FLA_UPPER_TRIANGULAR, beta, C );
00046 
00047   FLA_Part_1x2( A,    &AL,  &AR,      0, FLA_RIGHT );
00048 
00049   FLA_Part_2x2( C,    &CTL, &CTR,
00050                       &CBL, &CBR,     0, 0, FLA_BR );
00051 
00052   while ( FLA_Obj_width( AR ) < FLA_Obj_width( A ) ){
00053 
00054     FLA_Repart_1x2_to_1x3( AL,  /**/ AR,        &A0, &a1, /**/ &A2,
00055                            1, FLA_LEFT );
00056 
00057     FLA_Repart_2x2_to_3x3( CTL, /**/ CTR,       &C00,  &c01,     /**/ &C02,
00058                                                 &c10t, &gamma11, /**/ &c12t,
00059                         /* ************* */   /* ************************** */
00060                            CBL, /**/ CBR,       &C20,  &c21,     /**/ &C22,
00061                            1, 1, FLA_TL );
00062 
00063     /*------------------------------------------------------------*/
00064 
00065     /* c12t    = c12t    + A2' * a1 */
00066     FLA_Gemv_external( FLA_TRANSPOSE, alpha, A2, a1, FLA_ONE, c12t );
00067 
00068     /* gamma11 = gamma11 + a1 * a1' */
00069     FLA_Dots_external( alpha, a1, a1, FLA_ONE, gamma11 );
00070 
00071     /*------------------------------------------------------------*/
00072 
00073     FLA_Cont_with_1x3_to_1x2( &AL,  /**/ &AR,        A0, /**/ a1, A2,
00074                               FLA_RIGHT );
00075 
00076     FLA_Cont_with_3x3_to_2x2( &CTL, /**/ &CTR,       C00,  /**/ c01,     C02,
00077                             /* ************** */  /* ************************ */
00078                                                      c10t, /**/ gamma11, c12t,
00079                               &CBL, /**/ &CBR,       C20,  /**/ c21,     C22,
00080                               FLA_BR );
00081 
00082   }
00083 
00084   return FLA_SUCCESS;
00085 }

FLA_Error FLA_Syrk_ut_unb_var4 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  beta,
FLA_Obj  C 
)

References FLA_Cont_with_1x3_to_1x2(), FLA_Cont_with_3x3_to_2x2(), FLA_Dots_external(), FLA_Gemv_external(), FLA_Obj_width(), FLA_ONE, FLA_Part_1x2(), FLA_Part_2x2(), FLA_Repart_1x2_to_1x3(), FLA_Repart_2x2_to_3x3(), and FLA_Scalr_external().

00038 {
00039   FLA_Obj AL,    AR,       A0,  a1,  A2;
00040 
00041   FLA_Obj CTL,   CTR,      C00,  c01,     C02, 
00042           CBL,   CBR,      c10t, gamma11, c12t,
00043                            C20,  c21,     C22;
00044 
00045   FLA_Scalr_external( FLA_UPPER_TRIANGULAR, beta, C );
00046 
00047   FLA_Part_1x2( A,    &AL,  &AR,      0, FLA_RIGHT );
00048 
00049   FLA_Part_2x2( C,    &CTL, &CTR,
00050                       &CBL, &CBR,     0, 0, FLA_BR );
00051 
00052   while ( FLA_Obj_width( AR ) < FLA_Obj_width( A ) ){
00053 
00054     FLA_Repart_1x2_to_1x3( AL,  /**/ AR,        &A0, &a1, /**/ &A2,
00055                            1, FLA_LEFT );
00056 
00057     FLA_Repart_2x2_to_3x3( CTL, /**/ CTR,       &C00,  &c01,     /**/ &C02,
00058                                                 &c10t, &gamma11, /**/ &c12t,
00059                         /* ************* */   /* ************************** */
00060                            CBL, /**/ CBR,       &C20,  &c21,     /**/ &C22,
00061                            1, 1, FLA_TL );
00062 
00063     /*------------------------------------------------------------*/
00064 
00065     /* c01     = c01     + A0' * a1 */
00066     FLA_Gemv_external( FLA_TRANSPOSE, alpha, A0, a1, FLA_ONE, c01 );
00067 
00068     /* gamma11 = gamma11 + a1 * a1' */
00069     FLA_Dots_external( alpha, a1, a1, FLA_ONE, gamma11 );
00070 
00071     /*------------------------------------------------------------*/
00072 
00073     FLA_Cont_with_1x3_to_1x2( &AL,  /**/ &AR,        A0, /**/ a1, A2,
00074                               FLA_RIGHT );
00075 
00076     FLA_Cont_with_3x3_to_2x2( &CTL, /**/ &CTR,       C00,  /**/ c01,     C02,
00077                             /* ************** */  /* ************************ */
00078                                                      c10t, /**/ gamma11, c12t,
00079                               &CBL, /**/ &CBR,       C20,  /**/ c21,     C22,
00080                               FLA_BR );
00081 
00082   }
00083 
00084   return FLA_SUCCESS;
00085 }

FLA_Error FLA_Syrk_ut_unb_var5 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  beta,
FLA_Obj  C 
)

References FLA_Cont_with_3x1_to_2x1(), FLA_Obj_length(), FLA_Part_2x1(), FLA_Repart_2x1_to_3x1(), FLA_Scalr_external(), and FLA_Syr_external().

00038 {
00039   FLA_Obj AT,              A0,
00040           AB,              a1t,
00041                            A2;
00042 
00043   FLA_Scalr_external( FLA_UPPER_TRIANGULAR, beta, C );
00044 
00045   FLA_Part_2x1( A,    &AT, 
00046                       &AB,            0, FLA_TOP );
00047 
00048   while ( FLA_Obj_length( AT ) < FLA_Obj_length( A ) ){
00049 
00050     FLA_Repart_2x1_to_3x1( AT,                &A0, 
00051                         /* ** */            /* *** */
00052                                               &a1t, 
00053                            AB,                &A2,        1, FLA_BOTTOM );
00054 
00055     /*------------------------------------------------------------*/
00056 
00057     /* C := C + a1t' * a1t */ 
00058     FLA_Syr_external( FLA_UPPER_TRIANGULAR, alpha, a1t, C );
00059 
00060     /*------------------------------------------------------------*/
00061 
00062     FLA_Cont_with_3x1_to_2x1( &AT,                A0, 
00063                                                   a1t, 
00064                             /* ** */           /* *** */
00065                               &AB,                A2,     FLA_TOP );
00066 
00067   }
00068 
00069   return FLA_SUCCESS;
00070 }

FLA_Error FLA_Syrk_ut_unb_var6 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  beta,
FLA_Obj  C 
)

References FLA_Cont_with_3x1_to_2x1(), FLA_Obj_length(), FLA_Part_2x1(), FLA_Repart_2x1_to_3x1(), FLA_Scalr_external(), and FLA_Syr_external().

00038 {
00039   FLA_Obj AT,              A0,
00040           AB,              a1t,
00041                            A2;
00042 
00043   FLA_Scalr_external( FLA_UPPER_TRIANGULAR, beta, C );
00044 
00045   FLA_Part_2x1( A,    &AT, 
00046                       &AB,            0, FLA_BOTTOM );
00047 
00048   while ( FLA_Obj_length( AB ) < FLA_Obj_length( A ) ){
00049 
00050     FLA_Repart_2x1_to_3x1( AT,                &A0, 
00051                                               &a1t, 
00052                         /* ** */            /* *** */
00053                            AB,                &A2,        1, FLA_TOP );
00054 
00055     /*------------------------------------------------------------*/
00056 
00057     /* C := C + a1t' * a1t */ 
00058     FLA_Syr_external( FLA_UPPER_TRIANGULAR, alpha, a1t, C );
00059 
00060     /*------------------------------------------------------------*/
00061 
00062     FLA_Cont_with_3x1_to_2x1( &AT,                A0, 
00063                             /* ** */           /* *** */
00064                                                   a1t, 
00065                               &AB,                A2,     FLA_BOTTOM );
00066 
00067   }
00068 
00069   return FLA_SUCCESS;
00070 }


Generated on Mon Jul 6 05:45:55 2009 for libflame by  doxygen 1.5.9