FLA_Gemm_nn.h File Reference

(r)

Go to the source code of this file.

Functions

FLA_Error FLA_Gemm_nn_blk_var1 (FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C, fla_gemm_t *cntl)
FLA_Error FLA_Gemm_nn_blk_var2 (FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C, fla_gemm_t *cntl)
FLA_Error FLA_Gemm_nn_blk_var3 (FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C, fla_gemm_t *cntl)
FLA_Error FLA_Gemm_nn_blk_var4 (FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C, fla_gemm_t *cntl)
FLA_Error FLA_Gemm_nn_blk_var5 (FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C, fla_gemm_t *cntl)
FLA_Error FLA_Gemm_nn_blk_var6 (FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C, fla_gemm_t *cntl)
FLA_Error FLA_Gemm_nn_unb_var1 (FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C)
FLA_Error FLA_Gemm_nn_unb_var2 (FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C)
FLA_Error FLA_Gemm_nn_unb_var3 (FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C)
FLA_Error FLA_Gemm_nn_unb_var4 (FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C)
FLA_Error FLA_Gemm_nn_unb_var5 (FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C)
FLA_Error FLA_Gemm_nn_unb_var6 (FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C)


Function Documentation

FLA_Error FLA_Gemm_nn_blk_var1 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  B,
FLA_Obj  beta,
FLA_Obj  C,
fla_gemm_t cntl 
)

References FLA_Cont_with_3x1_to_2x1(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_Obj_length(), FLA_Part_2x1(), and FLA_Repart_2x1_to_3x1().

Referenced by FLA_Gemm_nn().

00036 {
00037   FLA_Obj AT,              A0,
00038           AB,              A1,
00039                            A2;
00040 
00041   FLA_Obj CT,              C0,
00042           CB,              C1,
00043                            C2;
00044 
00045   dim_t b;
00046 
00047   FLA_Part_2x1( A,    &AT, 
00048                       &AB,            0, FLA_TOP );
00049 
00050   FLA_Part_2x1( C,    &CT, 
00051                       &CB,            0, FLA_TOP );
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_2x1_to_3x1( CT,                &C0, 
00063                         /* ** */            /* ** */
00064                                               &C1, 
00065                            CB,                &C2,        b, FLA_BOTTOM );
00066 
00067     /*------------------------------------------------------------*/
00068 
00069     /* C1 = alpha * A1 * B + C1; */
00070     FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, 
00071                        alpha, A1, B, beta, C1, 
00072                        FLA_Cntl_sub_gemm( cntl ) );
00073 
00074     /*------------------------------------------------------------*/
00075 
00076     FLA_Cont_with_3x1_to_2x1( &AT,                A0, 
00077                                                   A1, 
00078                             /* ** */           /* ** */
00079                               &AB,                A2,     FLA_TOP );
00080 
00081     FLA_Cont_with_3x1_to_2x1( &CT,                C0, 
00082                                                   C1, 
00083                             /* ** */           /* ** */
00084                               &CB,                C2,     FLA_TOP );
00085 
00086   }
00087 
00088   return FLA_SUCCESS;
00089 }

FLA_Error FLA_Gemm_nn_blk_var2 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  B,
FLA_Obj  beta,
FLA_Obj  C,
fla_gemm_t cntl 
)

References FLA_Cont_with_3x1_to_2x1(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_Obj_length(), FLA_Part_2x1(), and FLA_Repart_2x1_to_3x1().

Referenced by FLA_Gemm_nn().

00038 {
00039   FLA_Obj AT,              A0,
00040           AB,              A1,
00041                            A2;
00042 
00043   FLA_Obj CT,              C0,
00044           CB,              C1,
00045                            C2;
00046 
00047   dim_t b;
00048 
00049   FLA_Part_2x1( A,    &AT, 
00050                       &AB,            0, FLA_BOTTOM );
00051 
00052   FLA_Part_2x1( C,    &CT, 
00053                       &CB,            0, FLA_BOTTOM );
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_2x1_to_3x1( CT,                &C0, 
00065                                               &C1, 
00066                         /* ** */            /* ** */
00067                            CB,                &C2,        b, FLA_TOP );
00068 
00069     /*------------------------------------------------------------*/
00070 
00071     /* C1 = alpha * A1 * B + C1; */
00072     FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, 
00073                        alpha, A1, B, beta, C1, 
00074                        FLA_Cntl_sub_gemm( cntl ) );
00075 
00076     /*------------------------------------------------------------*/
00077 
00078     FLA_Cont_with_3x1_to_2x1( &AT,                A0, 
00079                             /* ** */           /* ** */
00080                                                   A1, 
00081                               &AB,                A2,     FLA_BOTTOM );
00082 
00083     FLA_Cont_with_3x1_to_2x1( &CT,                C0, 
00084                             /* ** */           /* ** */
00085                                                   C1, 
00086                               &CB,                C2,     FLA_BOTTOM );
00087 
00088   }
00089 
00090   return FLA_SUCCESS;
00091 }

FLA_Error FLA_Gemm_nn_blk_var3 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  B,
FLA_Obj  beta,
FLA_Obj  C,
fla_gemm_t cntl 
)

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

Referenced by FLA_Gemm_nn().

00036 {
00037   FLA_Obj BL,    BR,       B0,  B1,  B2;
00038 
00039   FLA_Obj CL,    CR,       C0,  C1,  C2;
00040 
00041   dim_t b;
00042 
00043   FLA_Part_1x2( B,    &BL,  &BR,      0, FLA_LEFT );
00044 
00045   FLA_Part_1x2( C,    &CL,  &CR,      0, FLA_LEFT );
00046 
00047   while ( FLA_Obj_width( BL ) < FLA_Obj_width( B ) ){
00048 
00049     b = FLA_Determine_blocksize( BR, FLA_RIGHT, FLA_Cntl_blocksize( cntl ) );
00050 
00051     FLA_Repart_1x2_to_1x3( BL,  /**/ BR,        &B0, /**/ &B1, &B2,
00052                            b, FLA_RIGHT );
00053 
00054     FLA_Repart_1x2_to_1x3( CL,  /**/ CR,        &C0, /**/ &C1, &C2,
00055                            b, FLA_RIGHT );
00056 
00057     /*------------------------------------------------------------*/
00058 
00059     /* C1 = alpha * A * B1 + C1; */
00060     FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, 
00061                        alpha, A, B1, beta, C1, 
00062                        FLA_Cntl_sub_gemm( cntl ) );
00063 
00064     /*------------------------------------------------------------*/
00065 
00066     FLA_Cont_with_1x3_to_1x2( &BL,  /**/ &BR,        B0, B1, /**/ B2,
00067                               FLA_LEFT );
00068 
00069     FLA_Cont_with_1x3_to_1x2( &CL,  /**/ &CR,        C0, C1, /**/ C2,
00070                               FLA_LEFT );
00071 
00072   }
00073 
00074   return FLA_SUCCESS;
00075 }

FLA_Error FLA_Gemm_nn_blk_var4 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  B,
FLA_Obj  beta,
FLA_Obj  C,
fla_gemm_t cntl 
)

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

Referenced by FLA_Gemm_nn().

00038 {
00039   FLA_Obj BL,    BR,       B0,  B1,  B2;
00040 
00041   FLA_Obj CL,    CR,       C0,  C1,  C2;
00042 
00043   dim_t b;
00044 
00045   FLA_Part_1x2( B,    &BL,  &BR,      0, FLA_RIGHT );
00046 
00047   FLA_Part_1x2( C,    &CL,  &CR,      0, FLA_RIGHT );
00048 
00049   while ( FLA_Obj_width( BR ) < FLA_Obj_width( B ) ){
00050 
00051     b = FLA_Determine_blocksize( BL, FLA_LEFT, FLA_Cntl_blocksize( cntl ) );
00052 
00053     FLA_Repart_1x2_to_1x3( BL,  /**/ BR,        &B0, &B1, /**/ &B2,
00054                            b, FLA_LEFT );
00055 
00056     FLA_Repart_1x2_to_1x3( CL,  /**/ CR,        &C0, &C1, /**/ &C2,
00057                            b, FLA_LEFT );
00058 
00059     /*------------------------------------------------------------*/
00060 
00061     /* C1 = alpha * A * B1 + C1; */
00062     FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, 
00063                        alpha, A, B1, beta, C1, 
00064                        FLA_Cntl_sub_gemm( cntl ) );
00065 
00066     /*------------------------------------------------------------*/
00067 
00068     FLA_Cont_with_1x3_to_1x2( &BL,  /**/ &BR,        B0, /**/ B1, B2,
00069                               FLA_RIGHT );
00070 
00071     FLA_Cont_with_1x3_to_1x2( &CL,  /**/ &CR,        C0, /**/ C1, C2,
00072                               FLA_RIGHT );
00073 
00074   }
00075 
00076   return FLA_SUCCESS;
00077 }

FLA_Error FLA_Gemm_nn_blk_var5 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  B,
FLA_Obj  beta,
FLA_Obj  C,
fla_gemm_t cntl 
)

References FLA_Cont_with_1x3_to_1x2(), FLA_Cont_with_3x1_to_2x1(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_Obj_width(), FLA_Part_1x2(), FLA_Part_2x1(), FLA_Repart_1x2_to_1x3(), and FLA_Repart_2x1_to_3x1().

Referenced by FLA_Gemm_nn().

00036 {
00037   FLA_Obj AL,    AR,       A0,  A1,  A2;
00038 
00039   FLA_Obj BT,              B0,
00040           BB,              B1,
00041                            B2;
00042 
00043   dim_t b;
00044 
00045   FLA_Part_1x2( A,    &AL,  &AR,      0, FLA_LEFT );
00046 
00047   FLA_Part_2x1( B,    &BT, 
00048                       &BB,            0, FLA_TOP );
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_2x1_to_3x1( BT,                &B0, 
00058                         /* ** */            /* ** */
00059                                               &B1, 
00060                            BB,                &B2,        b, FLA_BOTTOM );
00061 
00062     /*------------------------------------------------------------*/
00063 
00064     /* C = alpha * A1 * B1 + C; */
00065     FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, 
00066                        alpha, A1, B1, beta, C, 
00067                        FLA_Cntl_sub_gemm( cntl ) );
00068 
00069     /*------------------------------------------------------------*/
00070 
00071     FLA_Cont_with_1x3_to_1x2( &AL,  /**/ &AR,        A0, A1, /**/ A2,
00072                               FLA_LEFT );
00073 
00074     FLA_Cont_with_3x1_to_2x1( &BT,                B0, 
00075                                                   B1, 
00076                             /* ** */           /* ** */
00077                               &BB,                B2,     FLA_TOP );
00078 
00079   }
00080 
00081   return FLA_SUCCESS;
00082 }

FLA_Error FLA_Gemm_nn_blk_var6 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  B,
FLA_Obj  beta,
FLA_Obj  C,
fla_gemm_t cntl 
)

References FLA_Cont_with_1x3_to_1x2(), FLA_Cont_with_3x1_to_2x1(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_Obj_width(), FLA_Part_1x2(), FLA_Part_2x1(), FLA_Repart_1x2_to_1x3(), and FLA_Repart_2x1_to_3x1().

Referenced by FLA_Gemm_nn().

00038 {
00039   FLA_Obj AL,    AR,       A0,  A1,  A2;
00040 
00041   FLA_Obj BT,              B0,
00042           BB,              B1,
00043                            B2;
00044 
00045   dim_t b;
00046 
00047   FLA_Part_1x2( A,    &AL,  &AR,      0, FLA_RIGHT );
00048 
00049   FLA_Part_2x1( B,    &BT, 
00050                       &BB,            0, FLA_BOTTOM );
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_2x1_to_3x1( BT,                &B0, 
00060                                               &B1, 
00061                         /* ** */            /* ** */
00062                            BB,                &B2,        b, FLA_TOP );
00063 
00064     /*------------------------------------------------------------*/
00065 
00066     /* C = alpha * A1 * B1 + C; */
00067     FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, 
00068                        alpha, A1, B1, beta, C, 
00069                        FLA_Cntl_sub_gemm( cntl ) );
00070 
00071     /*------------------------------------------------------------*/
00072 
00073     FLA_Cont_with_1x3_to_1x2( &AL,  /**/ &AR,        A0, /**/ A1, A2,
00074                               FLA_RIGHT );
00075 
00076     FLA_Cont_with_3x1_to_2x1( &BT,                B0, 
00077                             /* ** */           /* ** */
00078                                                   B1, 
00079                               &BB,                B2,     FLA_BOTTOM );
00080 
00081   }
00082 
00083   return FLA_SUCCESS;
00084 }

FLA_Error FLA_Gemm_nn_unb_var1 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  B,
FLA_Obj  beta,
FLA_Obj  C 
)

References FLA_Cont_with_3x1_to_2x1(), FLA_Gemv_external(), FLA_Obj_length(), FLA_ONE, FLA_Part_2x1(), FLA_Repart_2x1_to_3x1(), and FLA_Scal_external().

00038 {
00039   FLA_Obj AT,              A0,
00040           AB,              a1t,
00041                            A2;
00042 
00043   FLA_Obj CT,              C0,
00044           CB,              c1t,
00045                            C2;
00046 
00047   FLA_Scal_external( beta, C );
00048 
00049   FLA_Part_2x1( A,    &AT, 
00050                       &AB,            0, FLA_TOP );
00051 
00052   FLA_Part_2x1( C,    &CT, 
00053                       &CB,            0, FLA_TOP );
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_2x1_to_3x1( CT,                &C0, 
00063                         /* ** */            /* *** */
00064                                               &c1t, 
00065                            CB,                &C2,        1, FLA_BOTTOM );
00066 
00067     /*------------------------------------------------------------*/
00068 
00069     /* c1t  = a1t * B + c1t    */
00070     /* c1t' = B' * a1t' + c1t' */
00071     FLA_Gemv_external( FLA_TRANSPOSE, alpha, B, a1t, FLA_ONE, c1t );
00072 
00073     /*------------------------------------------------------------*/
00074 
00075     FLA_Cont_with_3x1_to_2x1( &AT,                A0, 
00076                                                   a1t, 
00077                             /* ** */           /* *** */
00078                               &AB,                A2,     FLA_TOP );
00079 
00080     FLA_Cont_with_3x1_to_2x1( &CT,                C0, 
00081                                                   c1t, 
00082                             /* ** */           /* *** */
00083                               &CB,                C2,     FLA_TOP );
00084 
00085   }
00086 
00087   return FLA_SUCCESS;
00088 }

FLA_Error FLA_Gemm_nn_unb_var2 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  B,
FLA_Obj  beta,
FLA_Obj  C 
)

References FLA_Cont_with_3x1_to_2x1(), FLA_Gemv_external(), FLA_Obj_length(), FLA_ONE, FLA_Part_2x1(), FLA_Repart_2x1_to_3x1(), and FLA_Scal_external().

00038 {
00039   FLA_Obj AT,              A0,
00040           AB,              a1t,
00041                            A2;
00042 
00043   FLA_Obj CT,              C0,
00044           CB,              c1t,
00045                            C2;
00046 
00047   FLA_Scal_external( beta, C );
00048 
00049   FLA_Part_2x1( A,    &AT, 
00050                       &AB,            0, FLA_BOTTOM );
00051 
00052   FLA_Part_2x1( C,    &CT, 
00053                       &CB,            0, FLA_BOTTOM );
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_2x1_to_3x1( CT,                &C0, 
00063                                               &c1t, 
00064                         /* ** */            /* *** */
00065                            CB,                &C2,        1, FLA_TOP );
00066 
00067     /*------------------------------------------------------------*/
00068 
00069     /* c1t  = a1t * B + c1t    */
00070     /* c1t' = B' * a1t' + c1t' */
00071     FLA_Gemv_external( FLA_TRANSPOSE, alpha, B, a1t, FLA_ONE, c1t );
00072 
00073     /*------------------------------------------------------------*/
00074 
00075     FLA_Cont_with_3x1_to_2x1( &AT,                A0, 
00076                             /* ** */           /* *** */
00077                                                   a1t, 
00078                               &AB,                A2,     FLA_BOTTOM );
00079 
00080     FLA_Cont_with_3x1_to_2x1( &CT,                C0, 
00081                             /* ** */           /* *** */
00082                                                   c1t, 
00083                               &CB,                C2,     FLA_BOTTOM );
00084 
00085   }
00086 
00087   return FLA_SUCCESS;
00088 }

FLA_Error FLA_Gemm_nn_unb_var3 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  B,
FLA_Obj  beta,
FLA_Obj  C 
)

References FLA_Cont_with_1x3_to_1x2(), FLA_Gemv_external(), FLA_Obj_width(), FLA_ONE, FLA_Part_1x2(), FLA_Repart_1x2_to_1x3(), and FLA_Scal_external().

00038 {
00039   FLA_Obj BL,    BR,       B0,  b1,  B2;
00040 
00041   FLA_Obj CL,    CR,       C0,  c1,  C2;
00042 
00043   FLA_Scal_external( beta, C );
00044 
00045   FLA_Part_1x2( B,    &BL,  &BR,      0, FLA_LEFT );
00046 
00047   FLA_Part_1x2( C,    &CL,  &CR,      0, FLA_LEFT );
00048 
00049   while ( FLA_Obj_width( BL ) < FLA_Obj_width( B ) ){
00050 
00051     FLA_Repart_1x2_to_1x3( BL,  /**/ BR,        &B0, /**/ &b1, &B2,
00052                            1, FLA_RIGHT );
00053 
00054     FLA_Repart_1x2_to_1x3( CL,  /**/ CR,        &C0, /**/ &c1, &C2,
00055                            1, FLA_RIGHT );
00056 
00057     /*------------------------------------------------------------*/
00058 
00059     /* c1 = A * b1 + c1 */
00060     FLA_Gemv_external( FLA_NO_TRANSPOSE, alpha, A, b1, FLA_ONE, c1 );
00061 
00062     /*------------------------------------------------------------*/
00063 
00064     FLA_Cont_with_1x3_to_1x2( &BL,  /**/ &BR,        B0, b1, /**/ B2,
00065                               FLA_LEFT );
00066 
00067     FLA_Cont_with_1x3_to_1x2( &CL,  /**/ &CR,        C0, c1, /**/ C2,
00068                               FLA_LEFT );
00069 
00070   }
00071 
00072   return FLA_SUCCESS;
00073 }

FLA_Error FLA_Gemm_nn_unb_var4 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  B,
FLA_Obj  beta,
FLA_Obj  C 
)

References FLA_Cont_with_1x3_to_1x2(), FLA_Gemv_external(), FLA_Obj_width(), FLA_ONE, FLA_Part_1x2(), FLA_Repart_1x2_to_1x3(), and FLA_Scal_external().

00038 {
00039   FLA_Obj BL,    BR,       B0,  b1,  B2;
00040 
00041   FLA_Obj CL,    CR,       C0,  c1,  C2;
00042 
00043   FLA_Scal_external( beta, C );
00044 
00045   FLA_Part_1x2( B,    &BL,  &BR,      0, FLA_RIGHT );
00046 
00047   FLA_Part_1x2( C,    &CL,  &CR,      0, FLA_RIGHT );
00048 
00049   while ( FLA_Obj_width( BR ) < FLA_Obj_width( B ) ){
00050 
00051     FLA_Repart_1x2_to_1x3( BL,  /**/ BR,        &B0, &b1, /**/ &B2,
00052                            1, FLA_LEFT );
00053 
00054     FLA_Repart_1x2_to_1x3( CL,  /**/ CR,        &C0, &c1, /**/ &C2,
00055                            1, FLA_LEFT );
00056 
00057     /*------------------------------------------------------------*/
00058 
00059     /* c1 = A * b1 + c1 */
00060     FLA_Gemv_external( FLA_NO_TRANSPOSE, alpha, A, b1, FLA_ONE, c1 );
00061 
00062     /*------------------------------------------------------------*/
00063 
00064     FLA_Cont_with_1x3_to_1x2( &BL,  /**/ &BR,        B0, /**/ b1, B2,
00065                               FLA_RIGHT );
00066 
00067     FLA_Cont_with_1x3_to_1x2( &CL,  /**/ &CR,        C0, /**/ c1, C2,
00068                               FLA_RIGHT );
00069 
00070   }
00071 
00072   return FLA_SUCCESS;
00073 }

FLA_Error FLA_Gemm_nn_unb_var5 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  B,
FLA_Obj  beta,
FLA_Obj  C 
)

References FLA_Cont_with_1x3_to_1x2(), FLA_Cont_with_3x1_to_2x1(), FLA_Ger_external(), FLA_Obj_width(), FLA_Part_1x2(), FLA_Part_2x1(), FLA_Repart_1x2_to_1x3(), FLA_Repart_2x1_to_3x1(), and FLA_Scal_external().

00038 {
00039   FLA_Obj AL,    AR,       A0,  a1,  A2;
00040 
00041   FLA_Obj BT,              B0,
00042           BB,              b1t,
00043                            B2;
00044 
00045   FLA_Scal_external( beta, C );
00046 
00047   FLA_Part_1x2( A,    &AL,  &AR,      0, FLA_LEFT );
00048 
00049   FLA_Part_2x1( B,    &BT, 
00050                       &BB,            0, FLA_TOP );
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_2x1_to_3x1( BT,                &B0, 
00058                         /* ** */            /* *** */
00059                                               &b1t, 
00060                            BB,                &B2,        1, FLA_BOTTOM );
00061 
00062     /*------------------------------------------------------------*/
00063 
00064     /* C = a1 * b1t + C */
00065     FLA_Ger_external( alpha, a1, b1t, C );
00066 
00067     /*------------------------------------------------------------*/
00068 
00069     FLA_Cont_with_1x3_to_1x2( &AL,  /**/ &AR,        A0, a1, /**/ A2,
00070                               FLA_LEFT );
00071 
00072     FLA_Cont_with_3x1_to_2x1( &BT,                B0, 
00073                                                   b1t, 
00074                             /* ** */           /* *** */
00075                               &BB,                B2,     FLA_TOP );
00076 
00077   }
00078 
00079   return FLA_SUCCESS;
00080 }

FLA_Error FLA_Gemm_nn_unb_var6 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  B,
FLA_Obj  beta,
FLA_Obj  C 
)

References FLA_Cont_with_1x3_to_1x2(), FLA_Cont_with_3x1_to_2x1(), FLA_Ger_external(), FLA_Obj_width(), FLA_Part_1x2(), FLA_Part_2x1(), FLA_Repart_1x2_to_1x3(), FLA_Repart_2x1_to_3x1(), and FLA_Scal_external().

00038 {
00039   FLA_Obj AL,    AR,       A0,  a1,  A2;
00040 
00041   FLA_Obj BT,              B0,
00042           BB,              b1t,
00043                            B2;
00044 
00045   FLA_Scal_external( beta, C );
00046 
00047   FLA_Part_1x2( A,    &AL,  &AR,      0, FLA_RIGHT );
00048 
00049   FLA_Part_2x1( B,    &BT, 
00050                       &BB,            0, FLA_BOTTOM );
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_2x1_to_3x1( BT,                &B0, 
00058                                               &b1t, 
00059                         /* ** */            /* *** */
00060                            BB,                &B2,        1, FLA_TOP );
00061 
00062     /*------------------------------------------------------------*/
00063 
00064     /* C = a1 * b1t + C */
00065     FLA_Ger_external( alpha, a1, b1t, C );
00066 
00067     /*------------------------------------------------------------*/
00068 
00069     FLA_Cont_with_1x3_to_1x2( &AL,  /**/ &AR,        A0, /**/ a1, A2,
00070                               FLA_RIGHT );
00071 
00072     FLA_Cont_with_3x1_to_2x1( &BT,                B0, 
00073                             /* ** */           /* *** */
00074                                                   b1t, 
00075                               &BB,                B2,     FLA_BOTTOM );
00076 
00077   }
00078 
00079   return FLA_SUCCESS;
00080 }


Generated on Mon Jul 6 05:45:54 2009 for libflame by  doxygen 1.5.9