libflame revision_anchor
Functions
FLA_Ttmm_u_opt_var2.c File Reference

(r)

Functions

FLA_Error FLA_Ttmm_u_opt_var2 (FLA_Obj A)
 
FLA_Error FLA_Ttmm_u_ops_var2 (int mn_A, float *buff_A, int rs_A, int cs_A)
 
FLA_Error FLA_Ttmm_u_opd_var2 (int mn_A, double *buff_A, int rs_A, int cs_A)
 
FLA_Error FLA_Ttmm_u_opc_var2 (int mn_A, scomplex *buff_A, int rs_A, int cs_A)
 
FLA_Error FLA_Ttmm_u_opz_var2 (int mn_A, dcomplex *buff_A, int rs_A, int cs_A)
 

Function Documentation

◆ FLA_Ttmm_u_opc_var2()

FLA_Error FLA_Ttmm_u_opc_var2 ( int  mn_A,
scomplex *  buff_A,
int  rs_A,
int  cs_A 
)
188{
189 scomplex* buff_1 = FLA_COMPLEX_PTR( FLA_ONE );
190 int i;
191
192 for ( i = 0; i < mn_A; ++i )
193 {
194 scomplex* a01 = buff_A + (i )*cs_A + (0 )*rs_A;
195 scomplex* alpha11 = buff_A + (i )*cs_A + (i )*rs_A;
196 scomplex* A02 = buff_A + (i+1)*cs_A + (0 )*rs_A;
197 scomplex* a12t = buff_A + (i+1)*cs_A + (i )*rs_A;
198
199 int mn_ahead = mn_A - i - 1;
200 int mn_behind = i;
201
202 /*------------------------------------------------------------*/
203
204 // FLA_Scal_external( alpha11, a01 );
205 bl1_cscalv( BLIS1_NO_CONJUGATE,
206 mn_behind,
207 alpha11,
208 a01, rs_A );
209
210 // FLA_Gemvc_external( FLA_NO_TRANSPOSE, FLA_CONJUGATE, FLA_ONE, A02, a12t, FLA_ONE, a01 );
211 bl1_cgemv( BLIS1_NO_TRANSPOSE,
212 BLIS1_CONJUGATE,
213 mn_behind,
214 mn_ahead,
215 buff_1,
216 A02, rs_A, cs_A,
217 a12t, cs_A,
218 buff_1,
219 a01, rs_A );
220
221 // FLA_Absolute_square( alpha11 );
222 bl1_cabsqr( alpha11 );
223
224 // FLA_Dotcs_external( FLA_CONJUGATE, FLA_ONE, a21, a21, FLA_ONE, alpha11 );
225 bl1_cdots( BLIS1_CONJUGATE,
226 mn_ahead,
227 buff_1,
228 a12t, cs_A,
229 a12t, cs_A,
230 buff_1,
231 alpha11 );
232
233 /*------------------------------------------------------------*/
234
235 }
236
237 return FLA_SUCCESS;
238}
FLA_ONE
FLA_Obj FLA_ONE
Definition FLA_Init.c:18
i
int i
Definition bl1_axmyv2.c:145
bl1_cdots
void bl1_cdots(conj1_t conj, int n, scomplex *alpha, scomplex *x, int incx, scomplex *y, int incy, scomplex *beta, scomplex *rho)
Definition bl1_dots.c:39
bl1_cgemv
void bl1_cgemv(trans1_t transa, conj1_t conjx, int m, int n, scomplex *alpha, scomplex *a, int a_rs, int a_cs, scomplex *x, int incx, scomplex *beta, scomplex *y, int incy)
Definition bl1_gemv.c:125
bl1_cscalv
void bl1_cscalv(conj1_t conj, int n, scomplex *alpha, scomplex *x, int incx)
Definition bl1_scalv.c:46
BLIS1_NO_TRANSPOSE
@ BLIS1_NO_TRANSPOSE
Definition blis_type_defs.h:54
BLIS1_CONJUGATE
@ BLIS1_CONJUGATE
Definition blis_type_defs.h:82
BLIS1_NO_CONJUGATE
@ BLIS1_NO_CONJUGATE
Definition blis_type_defs.h:81
scomplex
Definition blis_type_defs.h:133

References bl1_cdots(), bl1_cgemv(), bl1_cscalv(), BLIS1_CONJUGATE, BLIS1_NO_CONJUGATE, BLIS1_NO_TRANSPOSE, FLA_ONE, and i.

Referenced by FLA_Ttmm_u_opt_var2().

◆ FLA_Ttmm_u_opd_var2()

FLA_Error FLA_Ttmm_u_opd_var2 ( int  mn_A,
double *  buff_A,
int  rs_A,
int  cs_A 
)
132{
133 double* buff_1 = FLA_DOUBLE_PTR( FLA_ONE );
134 int i;
135
136 for ( i = 0; i < mn_A; ++i )
137 {
138 double* a01 = buff_A + (i )*cs_A + (0 )*rs_A;
139 double* alpha11 = buff_A + (i )*cs_A + (i )*rs_A;
140 double* A02 = buff_A + (i+1)*cs_A + (0 )*rs_A;
141 double* a12t = buff_A + (i+1)*cs_A + (i )*rs_A;
142
143 int mn_ahead = mn_A - i - 1;
144 int mn_behind = i;
145
146 /*------------------------------------------------------------*/
147
148 // FLA_Scal_external( alpha11, a01 );
149 bl1_dscalv( BLIS1_NO_CONJUGATE,
150 mn_behind,
151 alpha11,
152 a01, rs_A );
153
154 // FLA_Gemvc_external( FLA_NO_TRANSPOSE, FLA_CONJUGATE, FLA_ONE, A02, a12t, FLA_ONE, a01 );
155 bl1_dgemv( BLIS1_NO_TRANSPOSE,
156 BLIS1_CONJUGATE,
157 mn_behind,
158 mn_ahead,
159 buff_1,
160 A02, rs_A, cs_A,
161 a12t, cs_A,
162 buff_1,
163 a01, rs_A );
164
165 // FLA_Absolute_square( alpha11 );
166 bl1_dabsqr( alpha11 );
167
168 // FLA_Dotcs_external( FLA_CONJUGATE, FLA_ONE, a21, a21, FLA_ONE, alpha11 );
169 bl1_ddots( BLIS1_CONJUGATE,
170 mn_ahead,
171 buff_1,
172 a12t, cs_A,
173 a12t, cs_A,
174 buff_1,
175 alpha11 );
176
177 /*------------------------------------------------------------*/
178
179 }
180
181 return FLA_SUCCESS;
182}
bl1_ddots
void bl1_ddots(conj1_t conj, int n, double *alpha, double *x, int incx, double *y, int incy, double *beta, double *rho)
Definition bl1_dots.c:26
bl1_dgemv
void bl1_dgemv(trans1_t transa, conj1_t conjx, int m, int n, double *alpha, double *a, int a_rs, int a_cs, double *x, int incx, double *beta, double *y, int incy)
Definition bl1_gemv.c:69
bl1_dscalv
void bl1_dscalv(conj1_t conj, int n, double *alpha, double *x, int incx)
Definition bl1_scalv.c:24

References bl1_ddots(), bl1_dgemv(), bl1_dscalv(), BLIS1_CONJUGATE, BLIS1_NO_CONJUGATE, BLIS1_NO_TRANSPOSE, FLA_ONE, and i.

Referenced by FLA_Ttmm_u_opt_var2().

◆ FLA_Ttmm_u_ops_var2()

FLA_Error FLA_Ttmm_u_ops_var2 ( int  mn_A,
float *  buff_A,
int  rs_A,
int  cs_A 
)
76{
77 float* buff_1 = FLA_FLOAT_PTR( FLA_ONE );
78 int i;
79
80 for ( i = 0; i < mn_A; ++i )
81 {
82 float* a01 = buff_A + (i )*cs_A + (0 )*rs_A;
83 float* alpha11 = buff_A + (i )*cs_A + (i )*rs_A;
84 float* A02 = buff_A + (i+1)*cs_A + (0 )*rs_A;
85 float* a12t = buff_A + (i+1)*cs_A + (i )*rs_A;
86
87 int mn_ahead = mn_A - i - 1;
88 int mn_behind = i;
89
90 /*------------------------------------------------------------*/
91
92 // FLA_Scal_external( alpha11, a01 );
93 bl1_sscalv( BLIS1_NO_CONJUGATE,
94 mn_behind,
95 alpha11,
96 a01, rs_A );
97
98 // FLA_Gemvc_external( FLA_NO_TRANSPOSE, FLA_CONJUGATE, FLA_ONE, A02, a12t, FLA_ONE, a01 );
99 bl1_sgemv( BLIS1_NO_TRANSPOSE,
100 BLIS1_CONJUGATE,
101 mn_behind,
102 mn_ahead,
103 buff_1,
104 A02, rs_A, cs_A,
105 a12t, cs_A,
106 buff_1,
107 a01, rs_A );
108
109 // FLA_Absolute_square( alpha11 );
110 bl1_sabsqr( alpha11 );
111
112 // FLA_Dotcs_external( FLA_CONJUGATE, FLA_ONE, a21, a21, FLA_ONE, alpha11 );
113 bl1_sdots( BLIS1_CONJUGATE,
114 mn_ahead,
115 buff_1,
116 a12t, cs_A,
117 a12t, cs_A,
118 buff_1,
119 alpha11 );
120
121 /*------------------------------------------------------------*/
122
123 }
124
125 return FLA_SUCCESS;
126}
bl1_sdots
void bl1_sdots(conj1_t conj, int n, float *alpha, float *x, int incx, float *y, int incy, float *beta, float *rho)
Definition bl1_dots.c:13
bl1_sgemv
void bl1_sgemv(trans1_t transa, conj1_t conjx, int m, int n, float *alpha, float *a, int a_rs, int a_cs, float *x, int incx, float *beta, float *y, int incy)
Definition bl1_gemv.c:13
bl1_sscalv
void bl1_sscalv(conj1_t conj, int n, float *alpha, float *x, int incx)
Definition bl1_scalv.c:13

References bl1_sdots(), bl1_sgemv(), bl1_sscalv(), BLIS1_CONJUGATE, BLIS1_NO_CONJUGATE, BLIS1_NO_TRANSPOSE, FLA_ONE, and i.

Referenced by FLA_Ttmm_u_opt_var2().

◆ FLA_Ttmm_u_opt_var2()

FLA_Error FLA_Ttmm_u_opt_var2 ( FLA_Obj  A)
14{
15 FLA_Datatype datatype;
16 int mn_A;
17 int rs_A, cs_A;
18
19 datatype = FLA_Obj_datatype( A );
20
21 mn_A = FLA_Obj_length( A );
22 rs_A = FLA_Obj_row_stride( A );
23 cs_A = FLA_Obj_col_stride( A );
24
25
26 switch ( datatype )
27 {
28 case FLA_FLOAT:
29 {
30 float* buff_A = FLA_FLOAT_PTR( A );
31
32 FLA_Ttmm_u_ops_var2( mn_A,
33 buff_A, rs_A, cs_A );
34
35 break;
36 }
37
38 case FLA_DOUBLE:
39 {
40 double* buff_A = FLA_DOUBLE_PTR( A );
41
42 FLA_Ttmm_u_opd_var2( mn_A,
43 buff_A, rs_A, cs_A );
44
45 break;
46 }
47
48 case FLA_COMPLEX:
49 {
50 scomplex* buff_A = FLA_COMPLEX_PTR( A );
51
52 FLA_Ttmm_u_opc_var2( mn_A,
53 buff_A, rs_A, cs_A );
54
55 break;
56 }
57
58 case FLA_DOUBLE_COMPLEX:
59 {
60 dcomplex* buff_A = FLA_DOUBLE_COMPLEX_PTR( A );
61
62 FLA_Ttmm_u_opz_var2( mn_A,
63 buff_A, rs_A, cs_A );
64
65 break;
66 }
67 }
68
69 return FLA_SUCCESS;
70}
FLA_Ttmm_u_ops_var2
FLA_Error FLA_Ttmm_u_ops_var2(int mn_A, float *buff_A, int rs_A, int cs_A)
Definition FLA_Ttmm_u_opt_var2.c:74
FLA_Ttmm_u_opc_var2
FLA_Error FLA_Ttmm_u_opc_var2(int mn_A, scomplex *buff_A, int rs_A, int cs_A)
Definition FLA_Ttmm_u_opt_var2.c:186
FLA_Ttmm_u_opz_var2
FLA_Error FLA_Ttmm_u_opz_var2(int mn_A, dcomplex *buff_A, int rs_A, int cs_A)
Definition FLA_Ttmm_u_opt_var2.c:242
FLA_Ttmm_u_opd_var2
FLA_Error FLA_Ttmm_u_opd_var2(int mn_A, double *buff_A, int rs_A, int cs_A)
Definition FLA_Ttmm_u_opt_var2.c:130
FLA_Obj_row_stride
dim_t FLA_Obj_row_stride(FLA_Obj obj)
Definition FLA_Query.c:167
FLA_Obj_length
dim_t FLA_Obj_length(FLA_Obj obj)
Definition FLA_Query.c:116
FLA_Obj_col_stride
dim_t FLA_Obj_col_stride(FLA_Obj obj)
Definition FLA_Query.c:174
FLA_Obj_datatype
FLA_Datatype FLA_Obj_datatype(FLA_Obj obj)
Definition FLA_Query.c:13
FLA_Datatype
int FLA_Datatype
Definition FLA_type_defs.h:49
dcomplex
Definition blis_type_defs.h:138

References FLA_Obj_col_stride(), FLA_Obj_datatype(), FLA_Obj_length(), FLA_Obj_row_stride(), FLA_Ttmm_u_opc_var2(), FLA_Ttmm_u_opd_var2(), FLA_Ttmm_u_ops_var2(), FLA_Ttmm_u_opz_var2(), and i.

Referenced by FLA_Ttmm_u().

◆ FLA_Ttmm_u_opz_var2()

FLA_Error FLA_Ttmm_u_opz_var2 ( int  mn_A,
dcomplex *  buff_A,
int  rs_A,
int  cs_A 
)
244{
245 dcomplex* buff_1 = FLA_DOUBLE_COMPLEX_PTR( FLA_ONE );
246 int i;
247
248 for ( i = 0; i < mn_A; ++i )
249 {
250 dcomplex* a01 = buff_A + (i )*cs_A + (0 )*rs_A;
251 dcomplex* alpha11 = buff_A + (i )*cs_A + (i )*rs_A;
252 dcomplex* A02 = buff_A + (i+1)*cs_A + (0 )*rs_A;
253 dcomplex* a12t = buff_A + (i+1)*cs_A + (i )*rs_A;
254
255 int mn_ahead = mn_A - i - 1;
256 int mn_behind = i;
257
258 /*------------------------------------------------------------*/
259
260 // FLA_Scal_external( alpha11, a01 );
261 bl1_zscalv( BLIS1_NO_CONJUGATE,
262 mn_behind,
263 alpha11,
264 a01, rs_A );
265
266 // FLA_Gemvc_external( FLA_NO_TRANSPOSE, FLA_CONJUGATE, FLA_ONE, A02, a12t, FLA_ONE, a01 );
267 bl1_zgemv( BLIS1_NO_TRANSPOSE,
268 BLIS1_CONJUGATE,
269 mn_behind,
270 mn_ahead,
271 buff_1,
272 A02, rs_A, cs_A,
273 a12t, cs_A,
274 buff_1,
275 a01, rs_A );
276
277 // FLA_Absolute_square( alpha11 );
278 bl1_zabsqr( alpha11 );
279
280 // FLA_Dotcs_external( FLA_CONJUGATE, FLA_ONE, a21, a21, FLA_ONE, alpha11 );
281 bl1_zdots( BLIS1_CONJUGATE,
282 mn_ahead,
283 buff_1,
284 a12t, cs_A,
285 a12t, cs_A,
286 buff_1,
287 alpha11 );
288
289 /*------------------------------------------------------------*/
290
291 }
292
293 return FLA_SUCCESS;
294}
bl1_zdots
void bl1_zdots(conj1_t conj, int n, dcomplex *alpha, dcomplex *x, int incx, dcomplex *y, int incy, dcomplex *beta, dcomplex *rho)
Definition bl1_dots.c:56
bl1_zgemv
void bl1_zgemv(trans1_t transa, conj1_t conjx, int m, int n, dcomplex *alpha, dcomplex *a, int a_rs, int a_cs, dcomplex *x, int incx, dcomplex *beta, dcomplex *y, int incy)
Definition bl1_gemv.c:255
bl1_zscalv
void bl1_zscalv(conj1_t conj, int n, dcomplex *alpha, dcomplex *x, int incx)
Definition bl1_scalv.c:72

References bl1_zdots(), bl1_zgemv(), bl1_zscalv(), BLIS1_CONJUGATE, BLIS1_NO_CONJUGATE, BLIS1_NO_TRANSPOSE, FLA_ONE, and i.

Referenced by FLA_Ttmm_u_opt_var2().

Generated on Thu Sep 23 2021 17:46:17 for libflame by doxygen 1.9.8