Go to the source code of this file.
Functions | |
FLA_Error | FLA_Gemv_n_blk_var1 (FLA_Obj alpha, FLA_Obj A, FLA_Obj x, FLA_Obj beta, FLA_Obj y, fla_gemv_t *cntl) |
FLA_Error | FLA_Gemv_n_blk_var2 (FLA_Obj alpha, FLA_Obj A, FLA_Obj x, FLA_Obj beta, FLA_Obj y, fla_gemv_t *cntl) |
FLA_Error | FLA_Gemv_n_blk_var5 (FLA_Obj alpha, FLA_Obj A, FLA_Obj x, FLA_Obj beta, FLA_Obj y, fla_gemv_t *cntl) |
FLA_Error | FLA_Gemv_n_blk_var6 (FLA_Obj alpha, FLA_Obj A, FLA_Obj x, FLA_Obj beta, FLA_Obj y, fla_gemv_t *cntl) |
FLA_Error FLA_Gemv_n_blk_var1 | ( | FLA_Obj | alpha, | |
FLA_Obj | A, | |||
FLA_Obj | x, | |||
FLA_Obj | beta, | |||
FLA_Obj | y, | |||
fla_gemv_t * | cntl | |||
) |
References FLA_Cont_with_3x1_to_2x1(), FLA_Determine_blocksize(), FLA_Gemv_internal(), FLA_Obj_length(), FLA_Part_2x1(), and FLA_Repart_2x1_to_3x1().
Referenced by FLA_Gemv_n().
00036 { 00037 FLA_Obj AT, A0, 00038 AB, A1, 00039 A2; 00040 00041 FLA_Obj yT, y0, 00042 yB, y1, 00043 y2; 00044 00045 dim_t b; 00046 00047 FLA_Part_2x1( A, &AT, 00048 &AB, 0, FLA_TOP ); 00049 00050 FLA_Part_2x1( y, &yT, 00051 &yB, 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( yT, &y0, 00063 /* ** */ /* ** */ 00064 &y1, 00065 yB, &y2, b, FLA_BOTTOM ); 00066 00067 /*------------------------------------------------------------*/ 00068 00069 /* y1 = alpha * A1 * x + y1; */ 00070 FLA_Gemv_internal( FLA_NO_TRANSPOSE, 00071 alpha, A1, x, beta, y1, 00072 FLA_Cntl_sub_gemv( 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( &yT, y0, 00082 y1, 00083 /* ** */ /* ** */ 00084 &yB, y2, FLA_TOP ); 00085 00086 } 00087 00088 return FLA_SUCCESS; 00089 }
FLA_Error FLA_Gemv_n_blk_var2 | ( | FLA_Obj | alpha, | |
FLA_Obj | A, | |||
FLA_Obj | x, | |||
FLA_Obj | beta, | |||
FLA_Obj | y, | |||
fla_gemv_t * | cntl | |||
) |
References FLA_Cont_with_3x1_to_2x1(), FLA_Determine_blocksize(), FLA_Gemv_internal(), FLA_Obj_length(), FLA_Part_2x1(), and FLA_Repart_2x1_to_3x1().
Referenced by FLA_Gemv_n().
00038 { 00039 FLA_Obj AT, A0, 00040 AB, A1, 00041 A2; 00042 00043 FLA_Obj yT, y0, 00044 yB, y1, 00045 y2; 00046 00047 dim_t b; 00048 00049 FLA_Part_2x1( A, &AT, 00050 &AB, 0, FLA_BOTTOM ); 00051 00052 FLA_Part_2x1( y, &yT, 00053 &yB, 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( yT, &y0, 00065 &y1, 00066 /* ** */ /* ** */ 00067 yB, &y2, b, FLA_TOP ); 00068 00069 /*------------------------------------------------------------*/ 00070 00071 /* y1 = alpha * A1 * x + y1; */ 00072 FLA_Gemv_internal( FLA_NO_TRANSPOSE, 00073 alpha, A1, x, beta, y1, 00074 FLA_Cntl_sub_gemv( 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( &yT, y0, 00084 /* ** */ /* ** */ 00085 y1, 00086 &yB, y2, FLA_BOTTOM ); 00087 00088 } 00089 00090 return FLA_SUCCESS; 00091 }
FLA_Error FLA_Gemv_n_blk_var5 | ( | FLA_Obj | alpha, | |
FLA_Obj | A, | |||
FLA_Obj | x, | |||
FLA_Obj | beta, | |||
FLA_Obj | y, | |||
fla_gemv_t * | cntl | |||
) |
References FLA_Cont_with_1x3_to_1x2(), FLA_Cont_with_3x1_to_2x1(), FLA_Determine_blocksize(), FLA_Gemv_internal(), FLA_Obj_width(), FLA_Part_1x2(), FLA_Part_2x1(), FLA_Repart_1x2_to_1x3(), and FLA_Repart_2x1_to_3x1().
Referenced by FLA_Gemv_n().
00036 { 00037 FLA_Obj AL, AR, A0, A1, A2; 00038 00039 FLA_Obj xT, x0, 00040 xB, x1, 00041 x2; 00042 00043 dim_t b; 00044 00045 FLA_Part_1x2( A, &AL, &AR, 0, FLA_LEFT ); 00046 00047 FLA_Part_2x1( x, &xT, 00048 &xB, 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( xT, &x0, 00058 /* ** */ /* ** */ 00059 &x1, 00060 xB, &x2, b, FLA_BOTTOM ); 00061 00062 /*------------------------------------------------------------*/ 00063 00064 /* y = alpha * A1 * x1 + y; */ 00065 FLA_Gemv_internal( FLA_NO_TRANSPOSE, 00066 alpha, A1, x1, beta, y, 00067 FLA_Cntl_sub_gemv( 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( &xT, x0, 00075 x1, 00076 /* ** */ /* ** */ 00077 &xB, x2, FLA_TOP ); 00078 00079 } 00080 00081 return FLA_SUCCESS; 00082 }
FLA_Error FLA_Gemv_n_blk_var6 | ( | FLA_Obj | alpha, | |
FLA_Obj | A, | |||
FLA_Obj | x, | |||
FLA_Obj | beta, | |||
FLA_Obj | y, | |||
fla_gemv_t * | cntl | |||
) |
References FLA_Cont_with_1x3_to_1x2(), FLA_Cont_with_3x1_to_2x1(), FLA_Determine_blocksize(), FLA_Gemv_internal(), FLA_Obj_width(), FLA_Part_1x2(), FLA_Part_2x1(), FLA_Repart_1x2_to_1x3(), and FLA_Repart_2x1_to_3x1().
Referenced by FLA_Gemv_n().
00038 { 00039 FLA_Obj AL, AR, A0, A1, A2; 00040 00041 FLA_Obj xT, x0, 00042 xB, x1, 00043 x2; 00044 00045 dim_t b; 00046 00047 FLA_Part_1x2( A, &AL, &AR, 0, FLA_RIGHT ); 00048 00049 FLA_Part_2x1( x, &xT, 00050 &xB, 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( xT, &x0, 00060 &x1, 00061 /* ** */ /* ** */ 00062 xB, &x2, b, FLA_TOP ); 00063 00064 /*------------------------------------------------------------*/ 00065 00066 /* y = alpha * A1 * x1 + y; */ 00067 FLA_Gemv_internal( FLA_NO_TRANSPOSE, 00068 alpha, A1, x1, beta, y, 00069 FLA_Cntl_sub_gemv( 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( &xT, x0, 00077 /* ** */ /* ** */ 00078 x1, 00079 &xB, x2, FLA_BOTTOM ); 00080 00081 } 00082 00083 return FLA_SUCCESS; 00084 }