Go to the source code of this file.
Functions | |
FLA_Error | FLA_Axpy_blk_var1 (FLA_Obj alpha, FLA_Obj A, FLA_Obj B, fla_axpy_t *cntl) |
FLA_Error | FLA_Axpy_blk_var2 (FLA_Obj alpha, FLA_Obj A, FLA_Obj B, fla_axpy_t *cntl) |
FLA_Error | FLA_Axpy_blk_var3 (FLA_Obj alpha, FLA_Obj A, FLA_Obj B, fla_axpy_t *cntl) |
FLA_Error | FLA_Axpy_blk_var4 (FLA_Obj alpha, FLA_Obj A, FLA_Obj B, fla_axpy_t *cntl) |
FLA_Error FLA_Axpy_blk_var1 | ( | FLA_Obj | alpha, | |
FLA_Obj | A, | |||
FLA_Obj | B, | |||
fla_axpy_t * | cntl | |||
) |
References FLA_Axpy_internal(), FLA_Cont_with_3x1_to_2x1(), FLA_Determine_blocksize(), FLA_Obj_length(), FLA_Part_2x1(), and FLA_Repart_2x1_to_3x1().
Referenced by FLA_Axpy_internal().
00036 { 00037 FLA_Obj AT, A0, 00038 AB, A1, 00039 A2; 00040 00041 FLA_Obj BT, B0, 00042 BB, B1, 00043 B2; 00044 00045 dim_t b; 00046 00047 FLA_Part_2x1( A, &AT, 00048 &AB, 0, FLA_TOP ); 00049 00050 FLA_Part_2x1( B, &BT, 00051 &BB, 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( BT, &B0, 00063 /* ** */ /* ** */ 00064 &B1, 00065 BB, &B2, b, FLA_BOTTOM ); 00066 00067 /*------------------------------------------------------------*/ 00068 00069 FLA_Axpy_internal( alpha, A1, B1, 00070 FLA_Cntl_sub_axpy( cntl ) ); 00071 00072 /*------------------------------------------------------------*/ 00073 00074 FLA_Cont_with_3x1_to_2x1( &AT, A0, 00075 A1, 00076 /* ** */ /* ** */ 00077 &AB, A2, FLA_TOP ); 00078 00079 FLA_Cont_with_3x1_to_2x1( &BT, B0, 00080 B1, 00081 /* ** */ /* ** */ 00082 &BB, B2, FLA_TOP ); 00083 } 00084 00085 return FLA_SUCCESS; 00086 }
FLA_Error FLA_Axpy_blk_var2 | ( | FLA_Obj | alpha, | |
FLA_Obj | A, | |||
FLA_Obj | B, | |||
fla_axpy_t * | cntl | |||
) |
References FLA_Axpy_internal(), FLA_Cont_with_3x1_to_2x1(), FLA_Determine_blocksize(), FLA_Obj_length(), FLA_Part_2x1(), and FLA_Repart_2x1_to_3x1().
Referenced by FLA_Axpy_internal().
00036 { 00037 FLA_Obj AT, A0, 00038 AB, A1, 00039 A2; 00040 00041 FLA_Obj BT, B0, 00042 BB, B1, 00043 B2; 00044 00045 dim_t b; 00046 00047 FLA_Part_2x1( A, &AT, 00048 &AB, 0, FLA_BOTTOM ); 00049 00050 FLA_Part_2x1( B, &BT, 00051 &BB, 0, FLA_BOTTOM ); 00052 00053 while ( FLA_Obj_length( AB ) < FLA_Obj_length( A ) ){ 00054 00055 b = FLA_Determine_blocksize( AT, FLA_TOP, FLA_Cntl_blocksize( cntl ) ); 00056 00057 FLA_Repart_2x1_to_3x1( AT, &A0, 00058 &A1, 00059 /* ** */ /* ** */ 00060 AB, &A2, b, FLA_TOP ); 00061 00062 FLA_Repart_2x1_to_3x1( BT, &B0, 00063 &B1, 00064 /* ** */ /* ** */ 00065 BB, &B2, b, FLA_TOP ); 00066 00067 /*------------------------------------------------------------*/ 00068 00069 FLA_Axpy_internal( alpha, A1, B1, 00070 FLA_Cntl_sub_axpy( cntl ) ); 00071 00072 /*------------------------------------------------------------*/ 00073 00074 FLA_Cont_with_3x1_to_2x1( &AT, A0, 00075 /* ** */ /* ** */ 00076 A1, 00077 &AB, A2, FLA_BOTTOM ); 00078 00079 FLA_Cont_with_3x1_to_2x1( &BT, B0, 00080 /* ** */ /* ** */ 00081 B1, 00082 &BB, B2, FLA_BOTTOM ); 00083 } 00084 00085 return FLA_SUCCESS; 00086 }
FLA_Error FLA_Axpy_blk_var3 | ( | FLA_Obj | alpha, | |
FLA_Obj | A, | |||
FLA_Obj | B, | |||
fla_axpy_t * | cntl | |||
) |
References FLA_Axpy_internal(), FLA_Cont_with_1x3_to_1x2(), FLA_Determine_blocksize(), FLA_Obj_width(), FLA_Part_1x2(), and FLA_Repart_1x2_to_1x3().
Referenced by FLA_Axpy_internal().
00036 { 00037 FLA_Obj AL, AR, A0, A1, A2; 00038 00039 FLA_Obj BL, BR, B0, B1, B2; 00040 00041 dim_t b; 00042 00043 FLA_Part_1x2( A, &AL, &AR, 0, FLA_LEFT ); 00044 00045 FLA_Part_1x2( B, &BL, &BR, 0, FLA_LEFT ); 00046 00047 while ( FLA_Obj_width( AL ) < FLA_Obj_width( A ) ){ 00048 00049 b = FLA_Determine_blocksize( AR, FLA_RIGHT, FLA_Cntl_blocksize( cntl ) ); 00050 00051 FLA_Repart_1x2_to_1x3( AL, /**/ AR, &A0, /**/ &A1, &A2, 00052 b, FLA_RIGHT ); 00053 00054 FLA_Repart_1x2_to_1x3( BL, /**/ BR, &B0, /**/ &B1, &B2, 00055 b, FLA_RIGHT ); 00056 00057 /*------------------------------------------------------------*/ 00058 00059 FLA_Axpy_internal( alpha, A1, B1, 00060 FLA_Cntl_sub_axpy( cntl ) ); 00061 00062 /*------------------------------------------------------------*/ 00063 00064 FLA_Cont_with_1x3_to_1x2( &AL, /**/ &AR, A0, A1, /**/ A2, 00065 FLA_LEFT ); 00066 00067 FLA_Cont_with_1x3_to_1x2( &BL, /**/ &BR, B0, B1, /**/ B2, 00068 FLA_LEFT ); 00069 } 00070 00071 return FLA_SUCCESS; 00072 }
FLA_Error FLA_Axpy_blk_var4 | ( | FLA_Obj | alpha, | |
FLA_Obj | A, | |||
FLA_Obj | B, | |||
fla_axpy_t * | cntl | |||
) |
References FLA_Axpy_internal(), FLA_Cont_with_1x3_to_1x2(), FLA_Determine_blocksize(), FLA_Obj_width(), FLA_Part_1x2(), and FLA_Repart_1x2_to_1x3().
Referenced by FLA_Axpy_internal().
00036 { 00037 FLA_Obj AL, AR, A0, A1, A2; 00038 00039 FLA_Obj BL, BR, B0, B1, B2; 00040 00041 dim_t b; 00042 00043 FLA_Part_1x2( A, &AL, &AR, 0, FLA_RIGHT ); 00044 00045 FLA_Part_1x2( B, &BL, &BR, 0, FLA_RIGHT ); 00046 00047 while ( FLA_Obj_width( AR ) < FLA_Obj_width( A ) ){ 00048 00049 b = FLA_Determine_blocksize( AL, FLA_LEFT, FLA_Cntl_blocksize( cntl ) ); 00050 00051 FLA_Repart_1x2_to_1x3( AL, /**/ AR, &A0, &A1, /**/ &A2, 00052 b, FLA_LEFT ); 00053 00054 FLA_Repart_1x2_to_1x3( BL, /**/ BR, &B0, &B1, /**/ &B2, 00055 b, FLA_LEFT ); 00056 00057 /*------------------------------------------------------------*/ 00058 00059 FLA_Axpy_internal( alpha, A1, B1, 00060 FLA_Cntl_sub_axpy( cntl ) ); 00061 00062 /*------------------------------------------------------------*/ 00063 00064 FLA_Cont_with_1x3_to_1x2( &AL, /**/ &AR, A0, /**/ A1, A2, 00065 FLA_RIGHT ); 00066 00067 FLA_Cont_with_1x3_to_1x2( &BL, /**/ &BR, B0, /**/ B1, B2, 00068 FLA_RIGHT ); 00069 } 00070 00071 return FLA_SUCCESS; 00072 }