FLA_Syrk_ln.h File Reference

(r)

Go to the source code of this file.

Functions

FLA_Error FLA_Syrk_ln_blk_var1 (FLA_Obj alpha, FLA_Obj A, FLA_Obj beta, FLA_Obj C, fla_syrk_t *cntl)
FLA_Error FLA_Syrk_ln_blk_var2 (FLA_Obj alpha, FLA_Obj A, FLA_Obj beta, FLA_Obj C, fla_syrk_t *cntl)
FLA_Error FLA_Syrk_ln_blk_var3 (FLA_Obj alpha, FLA_Obj A, FLA_Obj beta, FLA_Obj C, fla_syrk_t *cntl)
FLA_Error FLA_Syrk_ln_blk_var4 (FLA_Obj alpha, FLA_Obj A, FLA_Obj beta, FLA_Obj C, fla_syrk_t *cntl)
FLA_Error FLA_Syrk_ln_blk_var5 (FLA_Obj alpha, FLA_Obj A, FLA_Obj beta, FLA_Obj C, fla_syrk_t *cntl)
FLA_Error FLA_Syrk_ln_blk_var6 (FLA_Obj alpha, FLA_Obj A, FLA_Obj beta, FLA_Obj C, fla_syrk_t *cntl)
FLA_Error FLA_Syrk_ln_unb_var1 (FLA_Obj alpha, FLA_Obj A, FLA_Obj beta, FLA_Obj C)
FLA_Error FLA_Syrk_ln_unb_var2 (FLA_Obj alpha, FLA_Obj A, FLA_Obj beta, FLA_Obj C)
FLA_Error FLA_Syrk_ln_unb_var3 (FLA_Obj alpha, FLA_Obj A, FLA_Obj beta, FLA_Obj C)
FLA_Error FLA_Syrk_ln_unb_var4 (FLA_Obj alpha, FLA_Obj A, FLA_Obj beta, FLA_Obj C)
FLA_Error FLA_Syrk_ln_unb_var5 (FLA_Obj alpha, FLA_Obj A, FLA_Obj beta, FLA_Obj C)
FLA_Error FLA_Syrk_ln_unb_var6 (FLA_Obj alpha, FLA_Obj A, FLA_Obj beta, FLA_Obj C)


Function Documentation

FLA_Error FLA_Syrk_ln_blk_var1 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  beta,
FLA_Obj  C,
fla_syrk_t cntl 
)

References FLA_Cont_with_3x1_to_2x1(), FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_Obj_length(), FLA_Part_2x1(), FLA_Part_2x2(), FLA_Repart_2x1_to_3x1(), FLA_Repart_2x2_to_3x3(), and FLA_Syrk_internal().

Referenced by FLA_Syrk_ln().

00038 {
00039   FLA_Obj AT,              A0,
00040           AB,              A1,
00041                            A2;
00042 
00043   FLA_Obj CTL,   CTR,      C00, C01, C02, 
00044           CBL,   CBR,      C10, C11, C12,
00045                            C20, C21, C22;
00046 
00047   dim_t b;
00048 
00049   FLA_Part_2x1( A,    &AT, 
00050                       &AB,            0, FLA_TOP );
00051 
00052   FLA_Part_2x2( C,    &CTL, &CTR,
00053                       &CBL, &CBR,     0, 0, FLA_TL );
00054 
00055   while ( FLA_Obj_length( AT ) < FLA_Obj_length( A ) ){
00056 
00057     b = FLA_Determine_blocksize( AB, FLA_BOTTOM, FLA_Cntl_blocksize( cntl ) );
00058 
00059     FLA_Repart_2x1_to_3x1( AT,                &A0, 
00060                         /* ** */            /* ** */
00061                                               &A1, 
00062                            AB,                &A2,        b, FLA_BOTTOM );
00063 
00064     FLA_Repart_2x2_to_3x3( CTL, /**/ CTR,       &C00, /**/ &C01, &C02,
00065                         /* ************* */   /* ******************** */
00066                                                 &C10, /**/ &C11, &C12,
00067                            CBL, /**/ CBR,       &C20, /**/ &C21, &C22,
00068                            b, b, FLA_BR );
00069 
00070     /*------------------------------------------------------------*/
00071 
00072     /* C10 = C10 + A1 * A0' */
00073     FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_TRANSPOSE, 
00074                        alpha, A1, A0, beta, C10,
00075                        FLA_Cntl_sub_gemm( cntl ) );
00076 
00077     /* C11 = C11 + A1 * A1' */
00078     FLA_Syrk_internal( FLA_LOWER_TRIANGULAR, FLA_NO_TRANSPOSE, 
00079                        alpha, A1, beta, C11,
00080                        FLA_Cntl_sub_syrk( cntl ) );
00081 
00082     /*------------------------------------------------------------*/
00083 
00084     FLA_Cont_with_3x1_to_2x1( &AT,                A0, 
00085                                                   A1, 
00086                             /* ** */           /* ** */
00087                               &AB,                A2,     FLA_TOP );
00088 
00089     FLA_Cont_with_3x3_to_2x2( &CTL, /**/ &CTR,       C00, C01, /**/ C02,
00090                                                      C10, C11, /**/ C12,
00091                             /* ************** */  /* ****************** */
00092                               &CBL, /**/ &CBR,       C20, C21, /**/ C22,
00093                               FLA_TL );
00094 
00095   }
00096 
00097   return FLA_SUCCESS;
00098 }

FLA_Error FLA_Syrk_ln_blk_var2 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  beta,
FLA_Obj  C,
fla_syrk_t cntl 
)

References FLA_Cont_with_3x1_to_2x1(), FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_Obj_length(), FLA_Part_2x1(), FLA_Part_2x2(), FLA_Repart_2x1_to_3x1(), FLA_Repart_2x2_to_3x3(), and FLA_Syrk_internal().

Referenced by FLA_Syrk_ln().

00036 {
00037   FLA_Obj AT,              A0,
00038           AB,              A1,
00039                            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_2x1( A,    &AT, 
00048                       &AB,            0, FLA_TOP );
00049 
00050   FLA_Part_2x2( C,    &CTL, &CTR,
00051                       &CBL, &CBR,     0, 0, FLA_TL );
00052 
00053   while ( FLA_Obj_length( AT ) < FLA_Obj_length( A ) ){
00054 
00055     b = FLA_Determine_blocksize( AB, FLA_BOTTOM, FLA_Cntl_blocksize( cntl ) );
00056 
00057     FLA_Repart_2x1_to_3x1( AT,                &A0, 
00058                         /* ** */            /* ** */
00059                                               &A1, 
00060                            AB,                &A2,        b, FLA_BOTTOM );
00061 
00062     FLA_Repart_2x2_to_3x3( CTL, /**/ CTR,       &C00, /**/ &C01, &C02,
00063                         /* ************* */   /* ******************** */
00064                                                 &C10, /**/ &C11, &C12,
00065                            CBL, /**/ CBR,       &C20, /**/ &C21, &C22,
00066                            b, b, FLA_BR );
00067 
00068     /*------------------------------------------------------------*/
00069 
00070     /* C21 = C21 + A2 * A1' */
00071     FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_TRANSPOSE, 
00072                        alpha, A2, A1, beta, C21,
00073                        FLA_Cntl_sub_gemm( cntl ) );
00074 
00075     /* C11 = C11 + A1 * A1' */
00076     FLA_Syrk_internal( FLA_LOWER_TRIANGULAR, FLA_NO_TRANSPOSE, 
00077                        alpha, A1, beta, C11,
00078                        FLA_Cntl_sub_syrk( cntl ) );
00079 
00080     /*------------------------------------------------------------*/
00081 
00082     FLA_Cont_with_3x1_to_2x1( &AT,                A0, 
00083                                                   A1, 
00084                             /* ** */           /* ** */
00085                               &AB,                A2,     FLA_TOP );
00086 
00087     FLA_Cont_with_3x3_to_2x2( &CTL, /**/ &CTR,       C00, C01, /**/ C02,
00088                                                      C10, C11, /**/ C12,
00089                             /* ************** */  /* ****************** */
00090                               &CBL, /**/ &CBR,       C20, C21, /**/ C22,
00091                               FLA_TL );
00092 
00093   }
00094 
00095   return FLA_SUCCESS;
00096 }

FLA_Error FLA_Syrk_ln_blk_var3 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  beta,
FLA_Obj  C,
fla_syrk_t cntl 
)

References FLA_Cont_with_3x1_to_2x1(), FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_Obj_length(), FLA_Part_2x1(), FLA_Part_2x2(), FLA_Repart_2x1_to_3x1(), FLA_Repart_2x2_to_3x3(), and FLA_Syrk_internal().

Referenced by FLA_Syrk_ln().

00038 {
00039   FLA_Obj AT,              A0,
00040           AB,              A1,
00041                            A2;
00042 
00043   FLA_Obj CTL,   CTR,      C00, C01, C02, 
00044           CBL,   CBR,      C10, C11, C12,
00045                            C20, C21, C22;
00046 
00047   dim_t b;
00048 
00049   FLA_Part_2x1( A,    &AT, 
00050                       &AB,            0, FLA_BOTTOM );
00051 
00052   FLA_Part_2x2( C,    &CTL, &CTR,
00053                       &CBL, &CBR,     0, 0, FLA_BR );
00054 
00055   while ( FLA_Obj_length( AB ) < FLA_Obj_length( A ) ){
00056 
00057     b = FLA_Determine_blocksize( AT, FLA_TOP, FLA_Cntl_blocksize( cntl ) );
00058 
00059     FLA_Repart_2x1_to_3x1( AT,                &A0, 
00060                                               &A1, 
00061                         /* ** */            /* ** */
00062                            AB,                &A2,        b, FLA_TOP );
00063 
00064     FLA_Repart_2x2_to_3x3( CTL, /**/ CTR,       &C00, &C01, /**/ &C02,
00065                                                 &C10, &C11, /**/ &C12,
00066                         /* ************* */   /* ******************** */
00067                            CBL, /**/ CBR,       &C20, &C21, /**/ &C22,
00068                            b, b, FLA_TL );
00069 
00070     /*------------------------------------------------------------*/
00071 
00072     /* C21 = C21 + A2 * A1' */
00073     FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_TRANSPOSE, 
00074                        alpha, A2, A1, beta, C21,
00075                        FLA_Cntl_sub_gemm( cntl ) );
00076 
00077     /* C11 = C11 + A1 * A1' */
00078     FLA_Syrk_internal( FLA_LOWER_TRIANGULAR, FLA_NO_TRANSPOSE, 
00079                        alpha, A1, beta, C11,
00080                        FLA_Cntl_sub_syrk( cntl ) );
00081 
00082     /*------------------------------------------------------------*/
00083 
00084     FLA_Cont_with_3x1_to_2x1( &AT,                A0, 
00085                             /* ** */           /* ** */
00086                                                   A1, 
00087                               &AB,                A2,     FLA_BOTTOM );
00088 
00089     FLA_Cont_with_3x3_to_2x2( &CTL, /**/ &CTR,       C00, /**/ C01, C02,
00090                             /* ************** */  /* ****************** */
00091                                                      C10, /**/ C11, C12,
00092                               &CBL, /**/ &CBR,       C20, /**/ C21, C22,
00093                               FLA_BR );
00094 
00095   }
00096 
00097   return FLA_SUCCESS;
00098 }

FLA_Error FLA_Syrk_ln_blk_var4 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  beta,
FLA_Obj  C,
fla_syrk_t cntl 
)

References FLA_Cont_with_3x1_to_2x1(), FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_Obj_length(), FLA_Part_2x1(), FLA_Part_2x2(), FLA_Repart_2x1_to_3x1(), FLA_Repart_2x2_to_3x3(), and FLA_Syrk_internal().

Referenced by FLA_Syrk_ln().

00038 {
00039   FLA_Obj AT,              A0,
00040           AB,              A1,
00041                            A2;
00042 
00043   FLA_Obj CTL,   CTR,      C00, C01, C02, 
00044           CBL,   CBR,      C10, C11, C12,
00045                            C20, C21, C22;
00046 
00047   dim_t b;
00048 
00049   FLA_Part_2x1( A,    &AT, 
00050                       &AB,            0, FLA_BOTTOM );
00051 
00052   FLA_Part_2x2( C,    &CTL, &CTR,
00053                       &CBL, &CBR,     0, 0, FLA_BR );
00054 
00055   while ( FLA_Obj_length( AB ) < FLA_Obj_length( A ) ){
00056 
00057     b = FLA_Determine_blocksize( AT, FLA_TOP, FLA_Cntl_blocksize( cntl ) );
00058 
00059     FLA_Repart_2x1_to_3x1( AT,                &A0, 
00060                                               &A1, 
00061                         /* ** */            /* ** */
00062                            AB,                &A2,        b, FLA_TOP );
00063 
00064     FLA_Repart_2x2_to_3x3( CTL, /**/ CTR,       &C00, &C01, /**/ &C02,
00065                                                 &C10, &C11, /**/ &C12,
00066                         /* ************* */   /* ******************** */
00067                            CBL, /**/ CBR,       &C20, &C21, /**/ &C22,
00068                            b, b, FLA_TL );
00069 
00070     /*------------------------------------------------------------*/
00071 
00072     /* C10 = C10 + A1 * A0' */
00073     FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_TRANSPOSE, 
00074                        alpha, A1, A0, beta, C10,
00075                        FLA_Cntl_sub_gemm( cntl ) );
00076 
00077     /* C11 = C11 + A1 * A1' */
00078     FLA_Syrk_internal( FLA_LOWER_TRIANGULAR, FLA_NO_TRANSPOSE, 
00079                        alpha, A1, beta, C11,
00080                        FLA_Cntl_sub_syrk( cntl ) );
00081 
00082     /*------------------------------------------------------------*/
00083 
00084     FLA_Cont_with_3x1_to_2x1( &AT,                A0, 
00085                             /* ** */           /* ** */
00086                                                   A1, 
00087                               &AB,                A2,     FLA_BOTTOM );
00088 
00089     FLA_Cont_with_3x3_to_2x2( &CTL, /**/ &CTR,       C00, /**/ C01, C02,
00090                             /* ************** */  /* ****************** */
00091                                                      C10, /**/ C11, C12,
00092                               &CBL, /**/ &CBR,       C20, /**/ C21, C22,
00093                               FLA_BR );
00094 
00095   }
00096 
00097   return FLA_SUCCESS;
00098 }

FLA_Error FLA_Syrk_ln_blk_var5 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  beta,
FLA_Obj  C,
fla_syrk_t cntl 
)

References FLA_Cont_with_1x3_to_1x2(), FLA_Determine_blocksize(), FLA_Obj_width(), FLA_Part_1x2(), FLA_Repart_1x2_to_1x3(), and FLA_Syrk_internal().

Referenced by FLA_Syrk_ln().

00036 {
00037   FLA_Obj AL,    AR,       A0,  A1,  A2;
00038 
00039   dim_t b;
00040 
00041   FLA_Part_1x2( A,    &AL,  &AR,      0, FLA_LEFT );
00042 
00043   while ( FLA_Obj_width( AL ) < FLA_Obj_width( A ) ){
00044 
00045     b = FLA_Determine_blocksize( AR, FLA_RIGHT, FLA_Cntl_blocksize( cntl ) );
00046 
00047     FLA_Repart_1x2_to_1x3( AL,  /**/ AR,        &A0, /**/ &A1, &A2,
00048                            b, FLA_RIGHT );
00049 
00050     /*------------------------------------------------------------*/
00051 
00052     /* C = C + A1 * A1' */
00053     FLA_Syrk_internal( FLA_LOWER_TRIANGULAR, FLA_NO_TRANSPOSE, 
00054                        alpha, A1, beta, C,
00055                        FLA_Cntl_sub_syrk( cntl ) );
00056 
00057     /*------------------------------------------------------------*/
00058 
00059     FLA_Cont_with_1x3_to_1x2( &AL,  /**/ &AR,        A0, A1, /**/ A2,
00060                               FLA_LEFT );
00061   }
00062 
00063   return FLA_SUCCESS;
00064 }

FLA_Error FLA_Syrk_ln_blk_var6 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  beta,
FLA_Obj  C,
fla_syrk_t cntl 
)

References FLA_Cont_with_1x3_to_1x2(), FLA_Determine_blocksize(), FLA_Obj_width(), FLA_Part_1x2(), FLA_Repart_1x2_to_1x3(), and FLA_Syrk_internal().

Referenced by FLA_Syrk_ln().

00038 {
00039   FLA_Obj AL,    AR,       A0,  A1,  A2;
00040 
00041   dim_t b;
00042 
00043   FLA_Part_1x2( A,    &AL,  &AR,      0, FLA_RIGHT );
00044 
00045   while ( FLA_Obj_width( AR ) < FLA_Obj_width( A ) ){
00046 
00047     b = FLA_Determine_blocksize( AL, FLA_LEFT, FLA_Cntl_blocksize( cntl ) );
00048 
00049     FLA_Repart_1x2_to_1x3( AL,  /**/ AR,        &A0, &A1, /**/ &A2,
00050                            b, FLA_LEFT );
00051 
00052     /*------------------------------------------------------------*/
00053 
00054     /* C = C + A1 * A1' */
00055     FLA_Syrk_internal( FLA_LOWER_TRIANGULAR, FLA_NO_TRANSPOSE, 
00056                        alpha, A1, beta, C,
00057                        FLA_Cntl_sub_syrk( cntl ) );
00058 
00059     /*------------------------------------------------------------*/
00060 
00061     FLA_Cont_with_1x3_to_1x2( &AL,  /**/ &AR,        A0, /**/ A1, A2,
00062                               FLA_RIGHT );
00063 
00064   }
00065 
00066   return FLA_SUCCESS;
00067 }

FLA_Error FLA_Syrk_ln_unb_var1 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  beta,
FLA_Obj  C 
)

References FLA_Cont_with_3x1_to_2x1(), FLA_Cont_with_3x3_to_2x2(), FLA_Dots_external(), FLA_Gemv_external(), FLA_Obj_length(), FLA_ONE, FLA_Part_2x1(), FLA_Part_2x2(), FLA_Repart_2x1_to_3x1(), FLA_Repart_2x2_to_3x3(), and FLA_Scalr_external().

00038 {
00039   FLA_Obj AT,              A0,
00040           AB,              a1t,
00041                            A2;
00042 
00043   FLA_Obj CTL,   CTR,      C00,  c01,     C02, 
00044           CBL,   CBR,      c10t, gamma11, c12t,
00045                            C20,  c21,     C22;
00046 
00047   FLA_Scalr_external( FLA_LOWER_TRIANGULAR, beta, C );
00048 
00049   FLA_Part_2x1( A,    &AT, 
00050                       &AB,            0, FLA_TOP );
00051 
00052   FLA_Part_2x2( C,    &CTL, &CTR,
00053                       &CBL, &CBR,     0, 0, FLA_TL );
00054 
00055   while ( FLA_Obj_length( AT ) < FLA_Obj_length( A ) ){
00056 
00057     FLA_Repart_2x1_to_3x1( AT,                &A0, 
00058                         /* ** */            /* *** */
00059                                               &a1t, 
00060                            AB,                &A2,        1, FLA_BOTTOM );
00061 
00062     FLA_Repart_2x2_to_3x3( CTL, /**/ CTR,       &C00,  /**/ &c01,     &C02,
00063                         /* ************* */   /* ************************** */
00064                                                 &c10t, /**/ &gamma11, &c12t,
00065                            CBL, /**/ CBR,       &C20,  /**/ &c21,     &C22,
00066                            1, 1, FLA_BR );
00067 
00068     /*------------------------------------------------------------*/
00069 
00070     /* c10t    = c10t    + A0  * a1t' */
00071     FLA_Gemv_external( FLA_NO_TRANSPOSE, alpha, A0, a1t, FLA_ONE, c10t );
00072 
00073     /* gamma11 = gamma11 + a1t * a1t' */
00074     FLA_Dots_external( alpha, a1t, a1t, FLA_ONE, gamma11 );
00075     
00076     /*------------------------------------------------------------*/
00077 
00078     FLA_Cont_with_3x1_to_2x1( &AT,                A0, 
00079                                                   a1t, 
00080                             /* ** */           /* *** */
00081                               &AB,                A2,     FLA_TOP );
00082 
00083     FLA_Cont_with_3x3_to_2x2( &CTL, /**/ &CTR,       C00,  c01,     /**/ C02,
00084                                                      c10t, gamma11, /**/ c12t,
00085                             /* ************** */  /* ************************ */
00086                               &CBL, /**/ &CBR,       C20,  c21,     /**/ C22,
00087                               FLA_TL );
00088 
00089   }
00090 
00091   return FLA_SUCCESS;
00092 }

FLA_Error FLA_Syrk_ln_unb_var2 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  beta,
FLA_Obj  C 
)

References FLA_Cont_with_3x1_to_2x1(), FLA_Cont_with_3x3_to_2x2(), FLA_Dots_external(), FLA_Gemv_external(), FLA_Obj_length(), FLA_ONE, FLA_Part_2x1(), FLA_Part_2x2(), FLA_Repart_2x1_to_3x1(), FLA_Repart_2x2_to_3x3(), and FLA_Scalr_external().

00038 {
00039   FLA_Obj AT,              A0,
00040           AB,              a1t,
00041                            A2;
00042 
00043   FLA_Obj CTL,   CTR,      C00,  c01,     C02, 
00044           CBL,   CBR,      c10t, gamma11, c12t,
00045                            C20,  c21,     C22;
00046 
00047   FLA_Scalr_external( FLA_LOWER_TRIANGULAR, beta, C );
00048 
00049   FLA_Part_2x1( A,    &AT, 
00050                       &AB,            0, FLA_TOP );
00051 
00052   FLA_Part_2x2( C,    &CTL, &CTR,
00053                       &CBL, &CBR,     0, 0, FLA_TL );
00054 
00055   while ( FLA_Obj_length( AT ) < FLA_Obj_length( A ) ){
00056 
00057     FLA_Repart_2x1_to_3x1( AT,                &A0, 
00058                         /* ** */            /* *** */
00059                                               &a1t, 
00060                            AB,                &A2,        1, FLA_BOTTOM );
00061 
00062     FLA_Repart_2x2_to_3x3( CTL, /**/ CTR,       &C00,  /**/ &c01,     &C02,
00063                         /* ************* */   /* ************************** */
00064                                                 &c10t, /**/ &gamma11, &c12t,
00065                            CBL, /**/ CBR,       &C20,  /**/ &c21,     &C22,
00066                            1, 1, FLA_BR );
00067 
00068     /*------------------------------------------------------------*/
00069 
00070     /* c21     = c21     + A2  * a1t' */
00071     FLA_Gemv_external( FLA_NO_TRANSPOSE, alpha, A2, a1t, FLA_ONE, c21 );
00072 
00073     /* gamma11 = gamma11 + a1t * a1t' */
00074     FLA_Dots_external( alpha, a1t, a1t, FLA_ONE, gamma11 );
00075 
00076     /*------------------------------------------------------------*/
00077 
00078     FLA_Cont_with_3x1_to_2x1( &AT,                A0, 
00079                                                   a1t, 
00080                             /* ** */           /* *** */
00081                               &AB,                A2,     FLA_TOP );
00082 
00083     FLA_Cont_with_3x3_to_2x2( &CTL, /**/ &CTR,       C00,  c01,     /**/ C02,
00084                                                      c10t, gamma11, /**/ c12t,
00085                             /* ************** */  /* ************************ */
00086                               &CBL, /**/ &CBR,       C20,  c21,     /**/ C22,
00087                               FLA_TL );
00088 
00089   }
00090 
00091   return FLA_SUCCESS;
00092 }

FLA_Error FLA_Syrk_ln_unb_var3 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  beta,
FLA_Obj  C 
)

References FLA_Cont_with_3x1_to_2x1(), FLA_Cont_with_3x3_to_2x2(), FLA_Dots_external(), FLA_Gemv_external(), FLA_Obj_length(), FLA_ONE, FLA_Part_2x1(), FLA_Part_2x2(), FLA_Repart_2x1_to_3x1(), FLA_Repart_2x2_to_3x3(), and FLA_Scalr_external().

00038 {
00039   FLA_Obj AT,              A0,
00040           AB,              a1t,
00041                            A2;
00042 
00043   FLA_Obj CTL,   CTR,      C00,  c01,     C02, 
00044           CBL,   CBR,      c10t, gamma11, c12t,
00045                            C20,  c21,     C22;
00046 
00047   FLA_Scalr_external( FLA_LOWER_TRIANGULAR, beta, C );
00048 
00049   FLA_Part_2x1( A,    &AT, 
00050                       &AB,            0, FLA_BOTTOM );
00051 
00052   FLA_Part_2x2( C,    &CTL, &CTR,
00053                       &CBL, &CBR,     0, 0, FLA_BR );
00054 
00055   while ( FLA_Obj_length( AB ) < FLA_Obj_length( A ) ){
00056 
00057     FLA_Repart_2x1_to_3x1( AT,                &A0, 
00058                                               &a1t, 
00059                         /* ** */            /* *** */
00060                            AB,                &A2,        1, FLA_TOP );
00061 
00062     FLA_Repart_2x2_to_3x3( CTL, /**/ CTR,       &C00,  &c01,     /**/ &C02,
00063                                                 &c10t, &gamma11, /**/ &c12t,
00064                         /* ************* */   /* ************************** */
00065                            CBL, /**/ CBR,       &C20,  &c21,     /**/ &C22,
00066                            1, 1, FLA_TL );
00067 
00068     /*------------------------------------------------------------*/
00069 
00070     /* c21     = c21     + A2  * a1t' */
00071     FLA_Gemv_external( FLA_NO_TRANSPOSE, alpha, A2, a1t, FLA_ONE, c21 );
00072 
00073     /* gamma11 = gamma11 + a1t * a1t' */
00074     FLA_Dots_external( alpha, a1t, a1t, FLA_ONE, gamma11 );
00075 
00076     /*------------------------------------------------------------*/
00077 
00078     FLA_Cont_with_3x1_to_2x1( &AT,                A0, 
00079                             /* ** */           /* *** */
00080                                                   a1t, 
00081                               &AB,                A2,     FLA_BOTTOM );
00082 
00083     FLA_Cont_with_3x3_to_2x2( &CTL, /**/ &CTR,       C00,  /**/ c01,     C02,
00084                             /* ************** */  /* ************************ */
00085                                                      c10t, /**/ gamma11, c12t,
00086                               &CBL, /**/ &CBR,       C20,  /**/ c21,     C22,
00087                               FLA_BR );
00088 
00089   }
00090 
00091   return FLA_SUCCESS;
00092 }

FLA_Error FLA_Syrk_ln_unb_var4 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  beta,
FLA_Obj  C 
)

References FLA_Cont_with_3x1_to_2x1(), FLA_Cont_with_3x3_to_2x2(), FLA_Dots_external(), FLA_Gemv_external(), FLA_Obj_length(), FLA_ONE, FLA_Part_2x1(), FLA_Part_2x2(), FLA_Repart_2x1_to_3x1(), FLA_Repart_2x2_to_3x3(), and FLA_Scalr_external().

00038 {
00039   FLA_Obj AT,              A0,
00040           AB,              a1t,
00041                            A2;
00042 
00043   FLA_Obj CTL,   CTR,      C00,  c01,     C02, 
00044           CBL,   CBR,      c10t, gamma11, c12t,
00045                            C20,  c21,     C22;
00046 
00047   FLA_Scalr_external( FLA_LOWER_TRIANGULAR, beta, C );
00048 
00049   FLA_Part_2x1( A,    &AT, 
00050                       &AB,            0, FLA_BOTTOM );
00051 
00052   FLA_Part_2x2( C,    &CTL, &CTR,
00053                       &CBL, &CBR,     0, 0, FLA_BR );
00054 
00055   while ( FLA_Obj_length( AB ) < FLA_Obj_length( A ) ){
00056 
00057     FLA_Repart_2x1_to_3x1( AT,                &A0, 
00058                                               &a1t, 
00059                         /* ** */            /* *** */
00060                            AB,                &A2,        1, FLA_TOP );
00061 
00062     FLA_Repart_2x2_to_3x3( CTL, /**/ CTR,       &C00,  &c01,     /**/ &C02,
00063                                                 &c10t, &gamma11, /**/ &c12t,
00064                         /* ************* */   /* ************************** */
00065                            CBL, /**/ CBR,       &C20,  &c21,     /**/ &C22,
00066                            1, 1, FLA_TL );
00067 
00068     /*------------------------------------------------------------*/
00069 
00070     /* c10t    = c10t    + A0  * a1t' */
00071     FLA_Gemv_external( FLA_NO_TRANSPOSE, alpha, A0, a1t, FLA_ONE, c10t );
00072 
00073     /* gamma11 = gamma11 + a1t * a1t' */
00074     FLA_Dots_external( alpha, a1t, a1t, FLA_ONE, gamma11 );
00075 
00076     /*------------------------------------------------------------*/
00077 
00078     FLA_Cont_with_3x1_to_2x1( &AT,                A0, 
00079                             /* ** */           /* *** */
00080                                                   a1t, 
00081                               &AB,                A2,     FLA_BOTTOM );
00082 
00083     FLA_Cont_with_3x3_to_2x2( &CTL, /**/ &CTR,       C00,  /**/ c01,     C02,
00084                             /* ************** */  /* ************************ */
00085                                                      c10t, /**/ gamma11, c12t,
00086                               &CBL, /**/ &CBR,       C20,  /**/ c21,     C22,
00087                               FLA_BR );
00088 
00089   }
00090 
00091   return FLA_SUCCESS;
00092 }

FLA_Error FLA_Syrk_ln_unb_var5 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  beta,
FLA_Obj  C 
)

References FLA_Cont_with_1x3_to_1x2(), FLA_Obj_width(), FLA_Part_1x2(), FLA_Repart_1x2_to_1x3(), FLA_Scalr_external(), and FLA_Syr_external().

00038 {
00039   FLA_Obj AL,    AR,       A0,  a1,  A2;
00040 
00041   FLA_Scalr_external( FLA_LOWER_TRIANGULAR, beta, C );
00042 
00043   FLA_Part_1x2( A,    &AL,  &AR,      0, FLA_LEFT );
00044 
00045   while ( FLA_Obj_width( AL ) < FLA_Obj_width( A ) ){
00046 
00047     FLA_Repart_1x2_to_1x3( AL,  /**/ AR,        &A0, /**/ &a1, &A2,
00048                            1, FLA_RIGHT );
00049 
00050     /*------------------------------------------------------------*/
00051 
00052     /* C := C + a1 * a1' */
00053     FLA_Syr_external( FLA_LOWER_TRIANGULAR, alpha, a1, C );
00054 
00055     /*------------------------------------------------------------*/
00056 
00057     FLA_Cont_with_1x3_to_1x2( &AL,  /**/ &AR,        A0, a1, /**/ A2,
00058                               FLA_LEFT );
00059 
00060   }
00061 
00062   return FLA_SUCCESS;
00063 }

FLA_Error FLA_Syrk_ln_unb_var6 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  beta,
FLA_Obj  C 
)

References FLA_Cont_with_1x3_to_1x2(), FLA_Obj_width(), FLA_Part_1x2(), FLA_Repart_1x2_to_1x3(), FLA_Scalr_external(), and FLA_Syr_external().

00038 {
00039   FLA_Obj AL,    AR,       A0,  a1,  A2;
00040 
00041   FLA_Scalr_external( FLA_LOWER_TRIANGULAR, beta, C );
00042 
00043   FLA_Part_1x2( A,    &AL,  &AR,      0, FLA_RIGHT );
00044 
00045   while ( FLA_Obj_width( AR ) < FLA_Obj_width( A ) ){
00046 
00047     FLA_Repart_1x2_to_1x3( AL,  /**/ AR,        &A0, &a1, /**/ &A2,
00048                            1, FLA_LEFT );
00049 
00050     /*------------------------------------------------------------*/
00051 
00052     /* C := C + a1 * a1' */
00053     FLA_Syr_external( FLA_LOWER_TRIANGULAR, alpha, a1, C );
00054 
00055     /*------------------------------------------------------------*/
00056 
00057     FLA_Cont_with_1x3_to_1x2( &AL,  /**/ &AR,        A0, /**/ a1, A2,
00058                               FLA_RIGHT );
00059 
00060   }
00061 
00062   return FLA_SUCCESS;
00063 }


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