1 /*********************************************************************/
2 /* Copyright 2009, 2010 The University of Texas at Austin. */
3 /* All rights reserved. */
5 /* Redistribution and use in source and binary forms, with or */
6 /* without modification, are permitted provided that the following */
7 /* conditions are met: */
9 /* 1. Redistributions of source code must retain the above */
10 /* copyright notice, this list of conditions and the following */
13 /* 2. Redistributions in binary form must reproduce the above */
14 /* copyright notice, this list of conditions and the following */
15 /* disclaimer in the documentation and/or other materials */
16 /* provided with the distribution. */
18 /* THIS SOFTWARE IS PROVIDED BY THE UNIVERSITY OF TEXAS AT */
19 /* AUSTIN ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, */
20 /* INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF */
21 /* MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE */
22 /* DISCLAIMED. IN NO EVENT SHALL THE UNIVERSITY OF TEXAS AT */
23 /* AUSTIN OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, */
24 /* INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES */
25 /* (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE */
26 /* GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR */
27 /* BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF */
28 /* LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT */
29 /* (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT */
30 /* OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE */
31 /* POSSIBILITY OF SUCH DAMAGE. */
33 /* The views and conclusions contained in the software and */
34 /* documentation are those of the authors and should not be */
35 /* interpreted as representing official policies, either expressed */
36 /* or implied, of The University of Texas at Austin. */
37 /*********************************************************************/
39 #ifndef COMMON_PARAM_H
40 #define COMMON_PARAM_H
48 int offsetA, offsetB, align;
51 int sbgemm_p, sbgemm_q, sbgemm_r;
52 int sbgemm_unroll_m, sbgemm_unroll_n, sbgemm_unroll_mn;
54 void (*sbstobf16_k) (BLASLONG, float *, BLASLONG, bfloat16 *, BLASLONG);
55 void (*sbdtobf16_k) (BLASLONG, double *, BLASLONG, bfloat16 *, BLASLONG);
56 void (*sbf16tos_k) (BLASLONG, bfloat16 *, BLASLONG, float *, BLASLONG);
57 void (*dbf16tod_k) (BLASLONG, bfloat16 *, BLASLONG, double *, BLASLONG);
59 float (*sbamax_k) (BLASLONG, float *, BLASLONG);
60 float (*sbamin_k) (BLASLONG, float *, BLASLONG);
61 float (*sbmax_k) (BLASLONG, float *, BLASLONG);
62 float (*sbmin_k) (BLASLONG, float *, BLASLONG);
63 BLASLONG (*isbamax_k)(BLASLONG, float *, BLASLONG);
64 BLASLONG (*isbamin_k)(BLASLONG, float *, BLASLONG);
65 BLASLONG (*isbmax_k) (BLASLONG, float *, BLASLONG);
66 BLASLONG (*isbmin_k) (BLASLONG, float *, BLASLONG);
68 float (*sbnrm2_k) (BLASLONG, float *, BLASLONG);
69 float (*sbasum_k) (BLASLONG, float *, BLASLONG);
70 float (*sbsum_k) (BLASLONG, float *, BLASLONG);
71 int (*sbcopy_k) (BLASLONG, float *, BLASLONG, float *, BLASLONG);
72 float (*sbdot_k) (BLASLONG, bfloat16 *, BLASLONG, bfloat16 *, BLASLONG);
73 double (*dsbdot_k) (BLASLONG, float *, BLASLONG, float *, BLASLONG);
75 int (*sbrot_k) (BLASLONG, float *, BLASLONG, float *, BLASLONG, float, float);
77 int (*sbaxpy_k) (BLASLONG, BLASLONG, BLASLONG, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG);
78 int (*sbscal_k) (BLASLONG, BLASLONG, BLASLONG, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG);
79 int (*sbswap_k) (BLASLONG, BLASLONG, BLASLONG, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG);
81 int (*sbgemv_n) (BLASLONG, BLASLONG, float, bfloat16 *, BLASLONG, bfloat16 *, BLASLONG, float, float *, BLASLONG);
82 int (*sbgemv_t) (BLASLONG, BLASLONG, float, bfloat16 *, BLASLONG, bfloat16 *, BLASLONG, float, float *, BLASLONG);
83 int (*sbger_k) (BLASLONG, BLASLONG, BLASLONG, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
85 int (*sbsymv_L) (BLASLONG, BLASLONG, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
86 int (*sbsymv_U) (BLASLONG, BLASLONG, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
88 int (*sbgemm_kernel )(BLASLONG, BLASLONG, BLASLONG, float, bfloat16 *, bfloat16 *, float *, BLASLONG);
89 int (*sbgemm_beta )(BLASLONG, BLASLONG, BLASLONG, float, bfloat16 *, BLASLONG, bfloat16 *, BLASLONG, float *, BLASLONG);
91 int (*sbgemm_incopy )(BLASLONG, BLASLONG, bfloat16 *, BLASLONG, bfloat16 *);
92 int (*sbgemm_itcopy )(BLASLONG, BLASLONG, bfloat16 *, BLASLONG, bfloat16 *);
93 int (*sbgemm_oncopy )(BLASLONG, BLASLONG, bfloat16 *, BLASLONG, bfloat16 *);
94 int (*sbgemm_otcopy )(BLASLONG, BLASLONG, bfloat16 *, BLASLONG, bfloat16 *);
96 int (*sbtrsm_kernel_LN)(BLASLONG, BLASLONG, BLASLONG, float, float *, float *, float *, BLASLONG, BLASLONG);
97 int (*sbtrsm_kernel_LT)(BLASLONG, BLASLONG, BLASLONG, float, float *, float *, float *, BLASLONG, BLASLONG);
98 int (*sbtrsm_kernel_RN)(BLASLONG, BLASLONG, BLASLONG, float, float *, float *, float *, BLASLONG, BLASLONG);
99 int (*sbtrsm_kernel_RT)(BLASLONG, BLASLONG, BLASLONG, float, float *, float *, float *, BLASLONG, BLASLONG);
101 int (*sbtrsm_iunucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
102 int (*sbtrsm_iunncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
103 int (*sbtrsm_iutucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
104 int (*sbtrsm_iutncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
105 int (*sbtrsm_ilnucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
106 int (*sbtrsm_ilnncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
107 int (*sbtrsm_iltucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
108 int (*sbtrsm_iltncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
109 int (*sbtrsm_ounucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
110 int (*sbtrsm_ounncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
111 int (*sbtrsm_outucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
112 int (*sbtrsm_outncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
113 int (*sbtrsm_olnucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
114 int (*sbtrsm_olnncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
115 int (*sbtrsm_oltucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
116 int (*sbtrsm_oltncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
118 int (*sbtrmm_kernel_RN)(BLASLONG, BLASLONG, BLASLONG, float, float *, float *, float *, BLASLONG, BLASLONG);
119 int (*sbtrmm_kernel_RT)(BLASLONG, BLASLONG, BLASLONG, float, float *, float *, float *, BLASLONG, BLASLONG);
120 int (*sbtrmm_kernel_LN)(BLASLONG, BLASLONG, BLASLONG, float, float *, float *, float *, BLASLONG, BLASLONG);
121 int (*sbtrmm_kernel_LT)(BLASLONG, BLASLONG, BLASLONG, float, float *, float *, float *, BLASLONG, BLASLONG);
123 int (*sbtrmm_iunucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
124 int (*sbtrmm_iunncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
125 int (*sbtrmm_iutucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
126 int (*sbtrmm_iutncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
127 int (*sbtrmm_ilnucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
128 int (*sbtrmm_ilnncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
129 int (*sbtrmm_iltucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
130 int (*sbtrmm_iltncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
131 int (*sbtrmm_ounucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
132 int (*sbtrmm_ounncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
133 int (*sbtrmm_outucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
134 int (*sbtrmm_outncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
135 int (*sbtrmm_olnucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
136 int (*sbtrmm_olnncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
137 int (*sbtrmm_oltucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
138 int (*sbtrmm_oltncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
140 int (*sbsymm_iutcopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
141 int (*sbsymm_iltcopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
142 int (*sbsymm_outcopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
143 int (*sbsymm_oltcopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
145 int (*sbneg_tcopy) (BLASLONG, BLASLONG, float *, BLASLONG, float *);
146 int (*sblaswp_ncopy) (BLASLONG, BLASLONG, BLASLONG, float *, BLASLONG, blasint *, float *);
150 #if defined(BUILD_SINGLE) || defined(BUILD_COMPLEX)
151 int sgemm_p, sgemm_q, sgemm_r;
152 int sgemm_unroll_m, sgemm_unroll_n, sgemm_unroll_mn;
157 #if defined(BUILD_SINGLE) || defined(BUILD_COMPLEX)
158 float (*samax_k) (BLASLONG, float *, BLASLONG);
159 float (*samin_k) (BLASLONG, float *, BLASLONG);
160 float (*smax_k) (BLASLONG, float *, BLASLONG);
161 float (*smin_k) (BLASLONG, float *, BLASLONG);
163 BLASLONG (*isamax_k)(BLASLONG, float *, BLASLONG);
164 BLASLONG (*isamin_k)(BLASLONG, float *, BLASLONG);
165 BLASLONG (*ismax_k) (BLASLONG, float *, BLASLONG);
166 BLASLONG (*ismin_k) (BLASLONG, float *, BLASLONG);
167 float (*snrm2_k) (BLASLONG, float *, BLASLONG);
168 float (*sasum_k) (BLASLONG, float *, BLASLONG);
171 float (*ssum_k) (BLASLONG, float *, BLASLONG);
173 #if defined(BUILD_SINGLE) || defined(BUILD_COMPLEX)
174 int (*scopy_k) (BLASLONG, float *, BLASLONG, float *, BLASLONG);
175 float (*sdot_k) (BLASLONG, float *, BLASLONG, float *, BLASLONG);
176 //double (*dsdot_k) (BLASLONG, float *, BLASLONG, float *, BLASLONG);
178 int (*srot_k) (BLASLONG, float *, BLASLONG, float *, BLASLONG, float, float);
180 int (*saxpy_k) (BLASLONG, BLASLONG, BLASLONG, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG);
181 int (*sscal_k) (BLASLONG, BLASLONG, BLASLONG, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG);
182 int (*sswap_k) (BLASLONG, BLASLONG, BLASLONG, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG);
184 int (*sgemv_n) (BLASLONG, BLASLONG, BLASLONG, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
185 int (*sgemv_t) (BLASLONG, BLASLONG, BLASLONG, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
188 int (*sger_k) (BLASLONG, BLASLONG, BLASLONG, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
190 int (*ssymv_L) (BLASLONG, BLASLONG, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
191 int (*ssymv_U) (BLASLONG, BLASLONG, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
194 #if defined(BUILD_SINGLE) || defined(BUILD_COMPLEX)
196 void (*sgemm_direct) (BLASLONG, BLASLONG, BLASLONG, float *, BLASLONG , float *, BLASLONG , float * , BLASLONG);
197 int (*sgemm_direct_performant) (BLASLONG M, BLASLONG N, BLASLONG K);
200 int (*sgemm_kernel )(BLASLONG, BLASLONG, BLASLONG, float, float *, float *, float *, BLASLONG);
201 int (*sgemm_beta )(BLASLONG, BLASLONG, BLASLONG, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG);
204 int (*sgemm_incopy )(BLASLONG, BLASLONG, float *, BLASLONG, float *);
205 int (*sgemm_itcopy )(BLASLONG, BLASLONG, float *, BLASLONG, float *);
206 int (*sgemm_oncopy )(BLASLONG, BLASLONG, float *, BLASLONG, float *);
207 int (*sgemm_otcopy )(BLASLONG, BLASLONG, float *, BLASLONG, float *);
210 #ifdef SMALL_MATRIX_OPT
211 int (*sgemm_small_matrix_permit)(int transa, int transb, BLASLONG m, BLASLONG n, BLASLONG k, float alpha, float beta);
213 int (*sgemm_small_kernel_nn )(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha, float * B, BLASLONG ldb, float beta, float * C, BLASLONG ldc);
214 int (*sgemm_small_kernel_nt )(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha, float * B, BLASLONG ldb, float beta, float * C, BLASLONG ldc);
215 int (*sgemm_small_kernel_tn )(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha, float * B, BLASLONG ldb, float beta, float * C, BLASLONG ldc);
216 int (*sgemm_small_kernel_tt )(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha, float * B, BLASLONG ldb, float beta, float * C, BLASLONG ldc);
218 int (*sgemm_small_kernel_b0_nn )(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
219 int (*sgemm_small_kernel_b0_nt )(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
220 int (*sgemm_small_kernel_b0_tn )(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
221 int (*sgemm_small_kernel_b0_tt )(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
224 int (*strsm_kernel_LN)(BLASLONG, BLASLONG, BLASLONG, float, float *, float *, float *, BLASLONG, BLASLONG);
225 int (*strsm_kernel_LT)(BLASLONG, BLASLONG, BLASLONG, float, float *, float *, float *, BLASLONG, BLASLONG);
226 int (*strsm_kernel_RN)(BLASLONG, BLASLONG, BLASLONG, float, float *, float *, float *, BLASLONG, BLASLONG);
227 int (*strsm_kernel_RT)(BLASLONG, BLASLONG, BLASLONG, float, float *, float *, float *, BLASLONG, BLASLONG);
229 int (*strsm_iunucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
230 int (*strsm_iunncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
231 int (*strsm_iutucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
232 int (*strsm_iutncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
233 int (*strsm_ilnucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
234 int (*strsm_ilnncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
235 int (*strsm_iltucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
236 int (*strsm_iltncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
237 int (*strsm_ounucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
238 int (*strsm_ounncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
239 int (*strsm_outucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
240 int (*strsm_outncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
241 int (*strsm_olnucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
242 int (*strsm_olnncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
243 int (*strsm_oltucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
244 int (*strsm_oltncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
246 int (*strmm_kernel_RN)(BLASLONG, BLASLONG, BLASLONG, float, float *, float *, float *, BLASLONG, BLASLONG);
247 int (*strmm_kernel_RT)(BLASLONG, BLASLONG, BLASLONG, float, float *, float *, float *, BLASLONG, BLASLONG);
248 int (*strmm_kernel_LN)(BLASLONG, BLASLONG, BLASLONG, float, float *, float *, float *, BLASLONG, BLASLONG);
249 int (*strmm_kernel_LT)(BLASLONG, BLASLONG, BLASLONG, float, float *, float *, float *, BLASLONG, BLASLONG);
251 int (*strmm_iunucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
252 int (*strmm_iunncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
253 int (*strmm_iutucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
254 int (*strmm_iutncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
255 int (*strmm_ilnucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
256 int (*strmm_ilnncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
257 int (*strmm_iltucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
258 int (*strmm_iltncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
259 int (*strmm_ounucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
260 int (*strmm_ounncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
261 int (*strmm_outucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
262 int (*strmm_outncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
263 int (*strmm_olnucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
264 int (*strmm_olnncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
265 int (*strmm_oltucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
266 int (*strmm_oltncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
268 int (*ssymm_iutcopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
269 int (*ssymm_iltcopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
270 int (*ssymm_outcopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
271 int (*ssymm_oltcopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
273 int (*sneg_tcopy) (BLASLONG, BLASLONG, float *, BLASLONG, float *);
274 int (*slaswp_ncopy) (BLASLONG, BLASLONG, BLASLONG, float *, BLASLONG, blasint *, float *);
277 #if defined(BUILD_DOUBLE) || defined(BUILD_COMPLEX16)
278 int dgemm_p, dgemm_q, dgemm_r;
279 int dgemm_unroll_m, dgemm_unroll_n, dgemm_unroll_mn;
282 #if defined(BUILD_DOUBLE) || defined(BUILD_COMPLEX16)
283 double (*damax_k) (BLASLONG, double *, BLASLONG);
284 double (*damin_k) (BLASLONG, double *, BLASLONG);
285 double (*dmax_k) (BLASLONG, double *, BLASLONG);
286 double (*dmin_k) (BLASLONG, double *, BLASLONG);
287 BLASLONG (*idamax_k)(BLASLONG, double *, BLASLONG);
288 BLASLONG (*idamin_k)(BLASLONG, double *, BLASLONG);
289 BLASLONG (*idmax_k) (BLASLONG, double *, BLASLONG);
290 BLASLONG (*idmin_k) (BLASLONG, double *, BLASLONG);
293 #if defined(BUILD_DOUBLE) || defined(BUILD_COMPLEX16)
294 double (*dnrm2_k) (BLASLONG, double *, BLASLONG);
295 double (*dasum_k) (BLASLONG, double *, BLASLONG);
298 double (*dsum_k) (BLASLONG, double *, BLASLONG);
300 #if defined(BUILD_DOUBLE) || defined(BUILD_COMPLEX16)
301 int (*dcopy_k) (BLASLONG, double *, BLASLONG, double *, BLASLONG);
302 double (*ddot_k) (BLASLONG, double *, BLASLONG, double *, BLASLONG);
304 #if defined (BUILD_SINGLE) || defined(BUILD_DOUBLE)
305 double (*dsdot_k) (BLASLONG, float *, BLASLONG, float *, BLASLONG);
307 #if defined(BUILD_DOUBLE) || defined(BUILD_COMPLEX16)
308 int (*drot_k) (BLASLONG, double *, BLASLONG, double *, BLASLONG, double, double);
309 int (*daxpy_k) (BLASLONG, BLASLONG, BLASLONG, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG);
310 int (*dscal_k) (BLASLONG, BLASLONG, BLASLONG, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG);
311 int (*dswap_k) (BLASLONG, BLASLONG, BLASLONG, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG);
312 int (*dgemv_n) (BLASLONG, BLASLONG, BLASLONG, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, double *);
313 int (*dgemv_t) (BLASLONG, BLASLONG, BLASLONG, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, double *);
316 int (*dger_k) (BLASLONG, BLASLONG, BLASLONG, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, double *);
318 int (*dsymv_L) (BLASLONG, BLASLONG, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, double *);
319 int (*dsymv_U) (BLASLONG, BLASLONG, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, double *);
321 #if defined(BUILD_DOUBLE) || defined(BUILD_COMPLEX16)
322 int (*dgemm_kernel )(BLASLONG, BLASLONG, BLASLONG, double, double *, double *, double *, BLASLONG);
323 int (*dgemm_beta )(BLASLONG, BLASLONG, BLASLONG, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG);
325 int (*dgemm_incopy )(BLASLONG, BLASLONG, double *, BLASLONG, double *);
326 int (*dgemm_itcopy )(BLASLONG, BLASLONG, double *, BLASLONG, double *);
327 int (*dgemm_oncopy )(BLASLONG, BLASLONG, double *, BLASLONG, double *);
328 int (*dgemm_otcopy )(BLASLONG, BLASLONG, double *, BLASLONG, double *);
331 #ifdef SMALL_MATRIX_OPT
332 int (*dgemm_small_matrix_permit)(int transa, int transb, BLASLONG m, BLASLONG n, BLASLONG k, double alpha, double beta);
334 int (*dgemm_small_kernel_nn )(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha, double * B, BLASLONG ldb, double beta, double * C, BLASLONG ldc);
335 int (*dgemm_small_kernel_nt )(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha, double * B, BLASLONG ldb, double beta, double * C, BLASLONG ldc);
336 int (*dgemm_small_kernel_tn )(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha, double * B, BLASLONG ldb, double beta, double * C, BLASLONG ldc);
337 int (*dgemm_small_kernel_tt )(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha, double * B, BLASLONG ldb, double beta, double * C, BLASLONG ldc);
339 int (*dgemm_small_kernel_b0_nn )(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
340 int (*dgemm_small_kernel_b0_nt )(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
341 int (*dgemm_small_kernel_b0_tn )(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
342 int (*dgemm_small_kernel_b0_tt )(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
344 int (*dtrsm_kernel_LN)(BLASLONG, BLASLONG, BLASLONG, double, double *, double *, double *, BLASLONG, BLASLONG);
345 int (*dtrsm_kernel_LT)(BLASLONG, BLASLONG, BLASLONG, double, double *, double *, double *, BLASLONG, BLASLONG);
346 int (*dtrsm_kernel_RN)(BLASLONG, BLASLONG, BLASLONG, double, double *, double *, double *, BLASLONG, BLASLONG);
347 int (*dtrsm_kernel_RT)(BLASLONG, BLASLONG, BLASLONG, double, double *, double *, double *, BLASLONG, BLASLONG);
349 int (*dtrsm_iunucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
350 int (*dtrsm_iunncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
351 int (*dtrsm_iutucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
352 int (*dtrsm_iutncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
353 int (*dtrsm_ilnucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
354 int (*dtrsm_ilnncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
355 int (*dtrsm_iltucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
356 int (*dtrsm_iltncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
357 int (*dtrsm_ounucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
358 int (*dtrsm_ounncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
359 int (*dtrsm_outucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
360 int (*dtrsm_outncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
361 int (*dtrsm_olnucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
362 int (*dtrsm_olnncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
363 int (*dtrsm_oltucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
364 int (*dtrsm_oltncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
366 int (*dtrmm_kernel_RN)(BLASLONG, BLASLONG, BLASLONG, double, double *, double *, double *, BLASLONG, BLASLONG);
367 int (*dtrmm_kernel_RT)(BLASLONG, BLASLONG, BLASLONG, double, double *, double *, double *, BLASLONG, BLASLONG);
368 int (*dtrmm_kernel_LN)(BLASLONG, BLASLONG, BLASLONG, double, double *, double *, double *, BLASLONG, BLASLONG);
369 int (*dtrmm_kernel_LT)(BLASLONG, BLASLONG, BLASLONG, double, double *, double *, double *, BLASLONG, BLASLONG);
371 int (*dtrmm_iunucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
372 int (*dtrmm_iunncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
373 int (*dtrmm_iutucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
374 int (*dtrmm_iutncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
375 int (*dtrmm_ilnucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
376 int (*dtrmm_ilnncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
377 int (*dtrmm_iltucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
378 int (*dtrmm_iltncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
379 int (*dtrmm_ounucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
380 int (*dtrmm_ounncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
381 int (*dtrmm_outucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
382 int (*dtrmm_outncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
383 int (*dtrmm_olnucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
384 int (*dtrmm_olnncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
385 int (*dtrmm_oltucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
386 int (*dtrmm_oltncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
388 int (*dsymm_iutcopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
389 int (*dsymm_iltcopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
390 int (*dsymm_outcopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
391 int (*dsymm_oltcopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
393 int (*dneg_tcopy) (BLASLONG, BLASLONG, double *, BLASLONG, double *);
394 int (*dlaswp_ncopy) (BLASLONG, BLASLONG, BLASLONG, double *, BLASLONG, blasint *, double *);
398 int qgemm_p, qgemm_q, qgemm_r;
399 int qgemm_unroll_m, qgemm_unroll_n, qgemm_unroll_mn;
401 xdouble (*qamax_k) (BLASLONG, xdouble *, BLASLONG);
402 xdouble (*qamin_k) (BLASLONG, xdouble *, BLASLONG);
403 xdouble (*qmax_k) (BLASLONG, xdouble *, BLASLONG);
404 xdouble (*qmin_k) (BLASLONG, xdouble *, BLASLONG);
405 BLASLONG (*iqamax_k)(BLASLONG, xdouble *, BLASLONG);
406 BLASLONG (*iqamin_k)(BLASLONG, xdouble *, BLASLONG);
407 BLASLONG (*iqmax_k) (BLASLONG, xdouble *, BLASLONG);
408 BLASLONG (*iqmin_k) (BLASLONG, xdouble *, BLASLONG);
410 xdouble (*qnrm2_k) (BLASLONG, xdouble *, BLASLONG);
411 xdouble (*qasum_k) (BLASLONG, xdouble *, BLASLONG);
412 xdouble (*qsum_k) (BLASLONG, xdouble *, BLASLONG);
413 int (*qcopy_k) (BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG);
414 xdouble (*qdot_k) (BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG);
415 int (*qrot_k) (BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble, xdouble);
417 int (*qaxpy_k) (BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG);
418 int (*qscal_k) (BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG);
419 int (*qswap_k) (BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG);
421 int (*qgemv_n) (BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *);
422 int (*qgemv_t) (BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *);
423 int (*qger_k) (BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *);
425 int (*qsymv_L) (BLASLONG, BLASLONG, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *);
426 int (*qsymv_U) (BLASLONG, BLASLONG, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *);
428 int (*qgemm_kernel )(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG);
429 int (*qgemm_beta )(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG);
431 int (*qgemm_incopy )(BLASLONG, BLASLONG, xdouble *, BLASLONG, xdouble *);
432 int (*qgemm_itcopy )(BLASLONG, BLASLONG, xdouble *, BLASLONG, xdouble *);
433 int (*qgemm_oncopy )(BLASLONG, BLASLONG, xdouble *, BLASLONG, xdouble *);
434 int (*qgemm_otcopy )(BLASLONG, BLASLONG, xdouble *, BLASLONG, xdouble *);
436 int (*qtrsm_kernel_LN)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
437 int (*qtrsm_kernel_LT)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
438 int (*qtrsm_kernel_RN)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
439 int (*qtrsm_kernel_RT)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
441 int (*qtrsm_iunucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
442 int (*qtrsm_iunncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
443 int (*qtrsm_iutucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
444 int (*qtrsm_iutncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
445 int (*qtrsm_ilnucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
446 int (*qtrsm_ilnncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
447 int (*qtrsm_iltucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
448 int (*qtrsm_iltncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
449 int (*qtrsm_ounucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
450 int (*qtrsm_ounncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
451 int (*qtrsm_outucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
452 int (*qtrsm_outncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
453 int (*qtrsm_olnucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
454 int (*qtrsm_olnncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
455 int (*qtrsm_oltucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
456 int (*qtrsm_oltncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
458 int (*qtrmm_kernel_RN)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
459 int (*qtrmm_kernel_RT)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
460 int (*qtrmm_kernel_LN)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
461 int (*qtrmm_kernel_LT)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
463 int (*qtrmm_iunucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
464 int (*qtrmm_iunncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
465 int (*qtrmm_iutucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
466 int (*qtrmm_iutncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
467 int (*qtrmm_ilnucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
468 int (*qtrmm_ilnncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
469 int (*qtrmm_iltucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
470 int (*qtrmm_iltncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
471 int (*qtrmm_ounucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
472 int (*qtrmm_ounncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
473 int (*qtrmm_outucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
474 int (*qtrmm_outncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
475 int (*qtrmm_olnucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
476 int (*qtrmm_olnncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
477 int (*qtrmm_oltucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
478 int (*qtrmm_oltncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
480 int (*qsymm_iutcopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
481 int (*qsymm_iltcopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
482 int (*qsymm_outcopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
483 int (*qsymm_oltcopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
485 int (*qneg_tcopy) (BLASLONG, BLASLONG, xdouble *, BLASLONG, xdouble *);
486 int (*qlaswp_ncopy) (BLASLONG, BLASLONG, BLASLONG, xdouble *, BLASLONG, blasint *, xdouble *);
491 int cgemm_p, cgemm_q, cgemm_r;
492 int cgemm_unroll_m, cgemm_unroll_n, cgemm_unroll_mn;
494 float (*camax_k) (BLASLONG, float *, BLASLONG);
495 float (*camin_k) (BLASLONG, float *, BLASLONG);
496 BLASLONG (*icamax_k)(BLASLONG, float *, BLASLONG);
497 BLASLONG (*icamin_k)(BLASLONG, float *, BLASLONG);
499 float (*cnrm2_k) (BLASLONG, float *, BLASLONG);
500 float (*casum_k) (BLASLONG, float *, BLASLONG);
501 float (*csum_k) (BLASLONG, float *, BLASLONG);
502 int (*ccopy_k) (BLASLONG, float *, BLASLONG, float *, BLASLONG);
503 openblas_complex_float (*cdotu_k) (BLASLONG, float *, BLASLONG, float *, BLASLONG);
504 openblas_complex_float (*cdotc_k) (BLASLONG, float *, BLASLONG, float *, BLASLONG);
505 int (*csrot_k) (BLASLONG, float *, BLASLONG, float *, BLASLONG, float, float);
507 int (*caxpy_k) (BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG);
508 int (*caxpyc_k)(BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG);
509 int (*cscal_k) (BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG);
510 int (*cswap_k) (BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG);
512 int (*cgemv_n) (BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
513 int (*cgemv_t) (BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
514 int (*cgemv_r) (BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
515 int (*cgemv_c) (BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
516 int (*cgemv_o) (BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
517 int (*cgemv_u) (BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
518 int (*cgemv_s) (BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
519 int (*cgemv_d) (BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
520 int (*cgeru_k) (BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
521 int (*cgerc_k) (BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
522 int (*cgerv_k) (BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
523 int (*cgerd_k) (BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
525 int (*csymv_L) (BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
526 int (*csymv_U) (BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
527 int (*chemv_L) (BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
528 int (*chemv_U) (BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
529 int (*chemv_M) (BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
530 int (*chemv_V) (BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *);
532 int (*cgemm_kernel_n )(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG);
533 int (*cgemm_kernel_l )(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG);
534 int (*cgemm_kernel_r )(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG);
535 int (*cgemm_kernel_b )(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG);
536 int (*cgemm_beta )(BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG);
538 int (*cgemm_incopy )(BLASLONG, BLASLONG, float *, BLASLONG, float *);
539 int (*cgemm_itcopy )(BLASLONG, BLASLONG, float *, BLASLONG, float *);
540 int (*cgemm_oncopy )(BLASLONG, BLASLONG, float *, BLASLONG, float *);
541 int (*cgemm_otcopy )(BLASLONG, BLASLONG, float *, BLASLONG, float *);
543 #ifdef SMALL_MATRIX_OPT
544 int (*cgemm_small_matrix_permit)(int transa, int transb, BLASLONG m, BLASLONG n, BLASLONG k, float alpha0, float alpha1, float beta0, float beta1);
546 int (*cgemm_small_kernel_nn )(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float beta0, float beta1, float * C, BLASLONG ldc);
547 int (*cgemm_small_kernel_nt )(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float beta0, float beta1, float * C, BLASLONG ldc);
548 int (*cgemm_small_kernel_nr )(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float beta0, float beta1, float * C, BLASLONG ldc);
549 int (*cgemm_small_kernel_nc )(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float beta0, float beta1, float * C, BLASLONG ldc);
551 int (*cgemm_small_kernel_tn )(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float beta0, float beta1, float * C, BLASLONG ldc);
552 int (*cgemm_small_kernel_tt )(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float beta0, float beta1, float * C, BLASLONG ldc);
553 int (*cgemm_small_kernel_tr )(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float beta0, float beta1, float * C, BLASLONG ldc);
554 int (*cgemm_small_kernel_tc )(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float beta0, float beta1, float * C, BLASLONG ldc);
556 int (*cgemm_small_kernel_rn )(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float beta0, float beta1, float * C, BLASLONG ldc);
557 int (*cgemm_small_kernel_rt )(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float beta0, float beta1, float * C, BLASLONG ldc);
558 int (*cgemm_small_kernel_rr )(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float beta0, float beta1, float * C, BLASLONG ldc);
559 int (*cgemm_small_kernel_rc )(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float beta0, float beta1, float * C, BLASLONG ldc);
561 int (*cgemm_small_kernel_cn )(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float beta0, float beta1, float * C, BLASLONG ldc);
562 int (*cgemm_small_kernel_ct )(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float beta0, float beta1, float * C, BLASLONG ldc);
563 int (*cgemm_small_kernel_cr )(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float beta0, float beta1, float * C, BLASLONG ldc);
564 int (*cgemm_small_kernel_cc )(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float beta0, float beta1, float * C, BLASLONG ldc);
566 int (*cgemm_small_kernel_b0_nn )(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
567 int (*cgemm_small_kernel_b0_nt )(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
568 int (*cgemm_small_kernel_b0_nr )(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
569 int (*cgemm_small_kernel_b0_nc )(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
571 int (*cgemm_small_kernel_b0_tn )(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
572 int (*cgemm_small_kernel_b0_tt )(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
573 int (*cgemm_small_kernel_b0_tr )(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
574 int (*cgemm_small_kernel_b0_tc )(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
576 int (*cgemm_small_kernel_b0_rn )(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
577 int (*cgemm_small_kernel_b0_rt )(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
578 int (*cgemm_small_kernel_b0_rr )(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
579 int (*cgemm_small_kernel_b0_rc )(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
581 int (*cgemm_small_kernel_b0_cn )(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
582 int (*cgemm_small_kernel_b0_ct )(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
583 int (*cgemm_small_kernel_b0_cr )(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
584 int (*cgemm_small_kernel_b0_cc )(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
587 int (*ctrsm_kernel_LN)(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG, BLASLONG);
588 int (*ctrsm_kernel_LT)(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG, BLASLONG);
589 int (*ctrsm_kernel_LR)(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG, BLASLONG);
590 int (*ctrsm_kernel_LC)(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG, BLASLONG);
591 int (*ctrsm_kernel_RN)(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG, BLASLONG);
592 int (*ctrsm_kernel_RT)(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG, BLASLONG);
593 int (*ctrsm_kernel_RR)(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG, BLASLONG);
594 int (*ctrsm_kernel_RC)(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG, BLASLONG);
596 int (*ctrsm_iunucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
597 int (*ctrsm_iunncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
598 int (*ctrsm_iutucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
599 int (*ctrsm_iutncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
600 int (*ctrsm_ilnucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
601 int (*ctrsm_ilnncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
602 int (*ctrsm_iltucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
603 int (*ctrsm_iltncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
604 int (*ctrsm_ounucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
605 int (*ctrsm_ounncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
606 int (*ctrsm_outucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
607 int (*ctrsm_outncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
608 int (*ctrsm_olnucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
609 int (*ctrsm_olnncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
610 int (*ctrsm_oltucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
611 int (*ctrsm_oltncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, float *);
613 int (*ctrmm_kernel_RN)(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG, BLASLONG);
614 int (*ctrmm_kernel_RT)(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG, BLASLONG);
615 int (*ctrmm_kernel_RR)(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG, BLASLONG);
616 int (*ctrmm_kernel_RC)(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG, BLASLONG);
617 int (*ctrmm_kernel_LN)(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG, BLASLONG);
618 int (*ctrmm_kernel_LT)(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG, BLASLONG);
619 int (*ctrmm_kernel_LR)(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG, BLASLONG);
620 int (*ctrmm_kernel_LC)(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG, BLASLONG);
622 int (*ctrmm_iunucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
623 int (*ctrmm_iunncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
624 int (*ctrmm_iutucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
625 int (*ctrmm_iutncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
626 int (*ctrmm_ilnucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
627 int (*ctrmm_ilnncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
628 int (*ctrmm_iltucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
629 int (*ctrmm_iltncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
630 int (*ctrmm_ounucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
631 int (*ctrmm_ounncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
632 int (*ctrmm_outucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
633 int (*ctrmm_outncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
634 int (*ctrmm_olnucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
635 int (*ctrmm_olnncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
636 int (*ctrmm_oltucopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
637 int (*ctrmm_oltncopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
639 int (*csymm_iutcopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
640 int (*csymm_iltcopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
641 int (*csymm_outcopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
642 int (*csymm_oltcopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
644 int (*chemm_iutcopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
645 int (*chemm_iltcopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
646 int (*chemm_outcopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
647 int (*chemm_oltcopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
649 int cgemm3m_p, cgemm3m_q, cgemm3m_r;
650 int cgemm3m_unroll_m, cgemm3m_unroll_n, cgemm3m_unroll_mn;
652 int (*cgemm3m_kernel)(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG);
654 int (*cgemm3m_incopyb)(BLASLONG, BLASLONG, float *, BLASLONG, float *);
655 int (*cgemm3m_incopyr)(BLASLONG, BLASLONG, float *, BLASLONG, float *);
656 int (*cgemm3m_incopyi)(BLASLONG, BLASLONG, float *, BLASLONG, float *);
657 int (*cgemm3m_itcopyb)(BLASLONG, BLASLONG, float *, BLASLONG, float *);
658 int (*cgemm3m_itcopyr)(BLASLONG, BLASLONG, float *, BLASLONG, float *);
659 int (*cgemm3m_itcopyi)(BLASLONG, BLASLONG, float *, BLASLONG, float *);
661 int (*cgemm3m_oncopyb)(BLASLONG, BLASLONG, float *, BLASLONG, float, float, float *);
662 int (*cgemm3m_oncopyr)(BLASLONG, BLASLONG, float *, BLASLONG, float, float, float *);
663 int (*cgemm3m_oncopyi)(BLASLONG, BLASLONG, float *, BLASLONG, float, float, float *);
664 int (*cgemm3m_otcopyb)(BLASLONG, BLASLONG, float *, BLASLONG, float, float, float *);
665 int (*cgemm3m_otcopyr)(BLASLONG, BLASLONG, float *, BLASLONG, float, float, float *);
666 int (*cgemm3m_otcopyi)(BLASLONG, BLASLONG, float *, BLASLONG, float, float, float *);
668 int (*csymm3m_iucopyb)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
669 int (*csymm3m_ilcopyb)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
670 int (*csymm3m_iucopyr)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
671 int (*csymm3m_ilcopyr)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
672 int (*csymm3m_iucopyi)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
673 int (*csymm3m_ilcopyi)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
675 int (*csymm3m_oucopyb)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float, float, float *);
676 int (*csymm3m_olcopyb)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float, float, float *);
677 int (*csymm3m_oucopyr)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float, float, float *);
678 int (*csymm3m_olcopyr)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float, float, float *);
679 int (*csymm3m_oucopyi)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float, float, float *);
680 int (*csymm3m_olcopyi)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float, float, float *);
682 int (*chemm3m_iucopyb)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
683 int (*chemm3m_ilcopyb)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
684 int (*chemm3m_iucopyr)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
685 int (*chemm3m_ilcopyr)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
686 int (*chemm3m_iucopyi)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
687 int (*chemm3m_ilcopyi)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
689 int (*chemm3m_oucopyb)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float, float, float *);
690 int (*chemm3m_olcopyb)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float, float, float *);
691 int (*chemm3m_oucopyr)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float, float, float *);
692 int (*chemm3m_olcopyr)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float, float, float *);
693 int (*chemm3m_oucopyi)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float, float, float *);
694 int (*chemm3m_olcopyi)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float, float, float *);
696 int (*cneg_tcopy) (BLASLONG, BLASLONG, float *, BLASLONG, float *);
697 int (*claswp_ncopy) (BLASLONG, BLASLONG, BLASLONG, float *, BLASLONG, blasint *, float *);
700 #ifdef BUILD_COMPLEX16
701 int zgemm_p, zgemm_q, zgemm_r;
702 int zgemm_unroll_m, zgemm_unroll_n, zgemm_unroll_mn;
704 double (*zamax_k) (BLASLONG, double *, BLASLONG);
705 double (*zamin_k) (BLASLONG, double *, BLASLONG);
706 BLASLONG (*izamax_k)(BLASLONG, double *, BLASLONG);
707 BLASLONG (*izamin_k)(BLASLONG, double *, BLASLONG);
709 double (*znrm2_k) (BLASLONG, double *, BLASLONG);
710 double (*zasum_k) (BLASLONG, double *, BLASLONG);
711 double (*zsum_k) (BLASLONG, double *, BLASLONG);
712 int (*zcopy_k) (BLASLONG, double *, BLASLONG, double *, BLASLONG);
713 openblas_complex_double (*zdotu_k) (BLASLONG, double *, BLASLONG, double *, BLASLONG);
714 openblas_complex_double (*zdotc_k) (BLASLONG, double *, BLASLONG, double *, BLASLONG);
715 int (*zdrot_k) (BLASLONG, double *, BLASLONG, double *, BLASLONG, double, double);
717 int (*zaxpy_k) (BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG);
718 int (*zaxpyc_k)(BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG);
719 int (*zscal_k) (BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG);
720 int (*zswap_k) (BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG);
722 int (*zgemv_n) (BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, double *);
723 int (*zgemv_t) (BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, double *);
724 int (*zgemv_r) (BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, double *);
725 int (*zgemv_c) (BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, double *);
726 int (*zgemv_o) (BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, double *);
727 int (*zgemv_u) (BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, double *);
728 int (*zgemv_s) (BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, double *);
729 int (*zgemv_d) (BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, double *);
730 int (*zgeru_k) (BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, double *);
731 int (*zgerc_k) (BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, double *);
732 int (*zgerv_k) (BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, double *);
733 int (*zgerd_k) (BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, double *);
735 int (*zsymv_L) (BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, double *);
736 int (*zsymv_U) (BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, double *);
737 int (*zhemv_L) (BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, double *);
738 int (*zhemv_U) (BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, double *);
739 int (*zhemv_M) (BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, double *);
740 int (*zhemv_V) (BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, double *);
742 int (*zgemm_kernel_n )(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG);
743 int (*zgemm_kernel_l )(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG);
744 int (*zgemm_kernel_r )(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG);
745 int (*zgemm_kernel_b )(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG);
746 int (*zgemm_beta )(BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG);
748 int (*zgemm_incopy )(BLASLONG, BLASLONG, double *, BLASLONG, double *);
749 int (*zgemm_itcopy )(BLASLONG, BLASLONG, double *, BLASLONG, double *);
750 int (*zgemm_oncopy )(BLASLONG, BLASLONG, double *, BLASLONG, double *);
751 int (*zgemm_otcopy )(BLASLONG, BLASLONG, double *, BLASLONG, double *);
753 #ifdef SMALL_MATRIX_OPT
754 int (*zgemm_small_matrix_permit)(int transa, int transb, BLASLONG m, BLASLONG n, BLASLONG k, double alpha0, double alpha1, double beta0, double beta1);
756 int (*zgemm_small_kernel_nn )(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double beta0, double beta1, double * C, BLASLONG ldc);
757 int (*zgemm_small_kernel_nt )(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double beta0, double beta1, double * C, BLASLONG ldc);
758 int (*zgemm_small_kernel_nr )(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double beta0, double beta1, double * C, BLASLONG ldc);
759 int (*zgemm_small_kernel_nc )(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double beta0, double beta1, double * C, BLASLONG ldc);
761 int (*zgemm_small_kernel_tn )(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double beta0, double beta1, double * C, BLASLONG ldc);
762 int (*zgemm_small_kernel_tt )(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double beta0, double beta1, double * C, BLASLONG ldc);
763 int (*zgemm_small_kernel_tr )(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double beta0, double beta1, double * C, BLASLONG ldc);
764 int (*zgemm_small_kernel_tc )(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double beta0, double beta1, double * C, BLASLONG ldc);
766 int (*zgemm_small_kernel_rn )(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double beta0, double beta1, double * C, BLASLONG ldc);
767 int (*zgemm_small_kernel_rt )(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double beta0, double beta1, double * C, BLASLONG ldc);
768 int (*zgemm_small_kernel_rr )(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double beta0, double beta1, double * C, BLASLONG ldc);
769 int (*zgemm_small_kernel_rc )(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double beta0, double beta1, double * C, BLASLONG ldc);
771 int (*zgemm_small_kernel_cn )(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double beta0, double beta1, double * C, BLASLONG ldc);
772 int (*zgemm_small_kernel_ct )(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double beta0, double beta1, double * C, BLASLONG ldc);
773 int (*zgemm_small_kernel_cr )(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double beta0, double beta1, double * C, BLASLONG ldc);
774 int (*zgemm_small_kernel_cc )(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double beta0, double beta1, double * C, BLASLONG ldc);
776 int (*zgemm_small_kernel_b0_nn )(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
777 int (*zgemm_small_kernel_b0_nt )(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
778 int (*zgemm_small_kernel_b0_nr )(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
779 int (*zgemm_small_kernel_b0_nc )(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
781 int (*zgemm_small_kernel_b0_tn )(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
782 int (*zgemm_small_kernel_b0_tt )(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
783 int (*zgemm_small_kernel_b0_tr )(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
784 int (*zgemm_small_kernel_b0_tc )(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
786 int (*zgemm_small_kernel_b0_rn )(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
787 int (*zgemm_small_kernel_b0_rt )(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
788 int (*zgemm_small_kernel_b0_rr )(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
789 int (*zgemm_small_kernel_b0_rc )(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
791 int (*zgemm_small_kernel_b0_cn )(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
792 int (*zgemm_small_kernel_b0_ct )(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
793 int (*zgemm_small_kernel_b0_cr )(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
794 int (*zgemm_small_kernel_b0_cc )(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
797 int (*ztrsm_kernel_LN)(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG, BLASLONG);
798 int (*ztrsm_kernel_LT)(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG, BLASLONG);
799 int (*ztrsm_kernel_LR)(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG, BLASLONG);
800 int (*ztrsm_kernel_LC)(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG, BLASLONG);
801 int (*ztrsm_kernel_RN)(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG, BLASLONG);
802 int (*ztrsm_kernel_RT)(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG, BLASLONG);
803 int (*ztrsm_kernel_RR)(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG, BLASLONG);
804 int (*ztrsm_kernel_RC)(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG, BLASLONG);
806 int (*ztrsm_iunucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
807 int (*ztrsm_iunncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
808 int (*ztrsm_iutucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
809 int (*ztrsm_iutncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
810 int (*ztrsm_ilnucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
811 int (*ztrsm_ilnncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
812 int (*ztrsm_iltucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
813 int (*ztrsm_iltncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
814 int (*ztrsm_ounucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
815 int (*ztrsm_ounncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
816 int (*ztrsm_outucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
817 int (*ztrsm_outncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
818 int (*ztrsm_olnucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
819 int (*ztrsm_olnncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
820 int (*ztrsm_oltucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
821 int (*ztrsm_oltncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, double *);
823 int (*ztrmm_kernel_RN)(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG, BLASLONG);
824 int (*ztrmm_kernel_RT)(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG, BLASLONG);
825 int (*ztrmm_kernel_RR)(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG, BLASLONG);
826 int (*ztrmm_kernel_RC)(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG, BLASLONG);
827 int (*ztrmm_kernel_LN)(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG, BLASLONG);
828 int (*ztrmm_kernel_LT)(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG, BLASLONG);
829 int (*ztrmm_kernel_LR)(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG, BLASLONG);
830 int (*ztrmm_kernel_LC)(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG, BLASLONG);
832 int (*ztrmm_iunucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
833 int (*ztrmm_iunncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
834 int (*ztrmm_iutucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
835 int (*ztrmm_iutncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
836 int (*ztrmm_ilnucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
837 int (*ztrmm_ilnncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
838 int (*ztrmm_iltucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
839 int (*ztrmm_iltncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
840 int (*ztrmm_ounucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
841 int (*ztrmm_ounncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
842 int (*ztrmm_outucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
843 int (*ztrmm_outncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
844 int (*ztrmm_olnucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
845 int (*ztrmm_olnncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
846 int (*ztrmm_oltucopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
847 int (*ztrmm_oltncopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
849 int (*zsymm_iutcopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
850 int (*zsymm_iltcopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
851 int (*zsymm_outcopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
852 int (*zsymm_oltcopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
854 int (*zhemm_iutcopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
855 int (*zhemm_iltcopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
856 int (*zhemm_outcopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
857 int (*zhemm_oltcopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
859 int zgemm3m_p, zgemm3m_q, zgemm3m_r;
860 int zgemm3m_unroll_m, zgemm3m_unroll_n, zgemm3m_unroll_mn;
862 int (*zgemm3m_kernel)(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG);
864 int (*zgemm3m_incopyb)(BLASLONG, BLASLONG, double *, BLASLONG, double *);
865 int (*zgemm3m_incopyr)(BLASLONG, BLASLONG, double *, BLASLONG, double *);
866 int (*zgemm3m_incopyi)(BLASLONG, BLASLONG, double *, BLASLONG, double *);
867 int (*zgemm3m_itcopyb)(BLASLONG, BLASLONG, double *, BLASLONG, double *);
868 int (*zgemm3m_itcopyr)(BLASLONG, BLASLONG, double *, BLASLONG, double *);
869 int (*zgemm3m_itcopyi)(BLASLONG, BLASLONG, double *, BLASLONG, double *);
871 int (*zgemm3m_oncopyb)(BLASLONG, BLASLONG, double *, BLASLONG, double, double, double *);
872 int (*zgemm3m_oncopyr)(BLASLONG, BLASLONG, double *, BLASLONG, double, double, double *);
873 int (*zgemm3m_oncopyi)(BLASLONG, BLASLONG, double *, BLASLONG, double, double, double *);
874 int (*zgemm3m_otcopyb)(BLASLONG, BLASLONG, double *, BLASLONG, double, double, double *);
875 int (*zgemm3m_otcopyr)(BLASLONG, BLASLONG, double *, BLASLONG, double, double, double *);
876 int (*zgemm3m_otcopyi)(BLASLONG, BLASLONG, double *, BLASLONG, double, double, double *);
878 int (*zsymm3m_iucopyb)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
879 int (*zsymm3m_ilcopyb)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
880 int (*zsymm3m_iucopyr)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
881 int (*zsymm3m_ilcopyr)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
882 int (*zsymm3m_iucopyi)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
883 int (*zsymm3m_ilcopyi)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
885 int (*zsymm3m_oucopyb)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double, double, double *);
886 int (*zsymm3m_olcopyb)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double, double, double *);
887 int (*zsymm3m_oucopyr)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double, double, double *);
888 int (*zsymm3m_olcopyr)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double, double, double *);
889 int (*zsymm3m_oucopyi)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double, double, double *);
890 int (*zsymm3m_olcopyi)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double, double, double *);
892 int (*zhemm3m_iucopyb)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
893 int (*zhemm3m_ilcopyb)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
894 int (*zhemm3m_iucopyr)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
895 int (*zhemm3m_ilcopyr)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
896 int (*zhemm3m_iucopyi)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
897 int (*zhemm3m_ilcopyi)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
899 int (*zhemm3m_oucopyb)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double, double, double *);
900 int (*zhemm3m_olcopyb)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double, double, double *);
901 int (*zhemm3m_oucopyr)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double, double, double *);
902 int (*zhemm3m_olcopyr)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double, double, double *);
903 int (*zhemm3m_oucopyi)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double, double, double *);
904 int (*zhemm3m_olcopyi)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double, double, double *);
906 int (*zneg_tcopy) (BLASLONG, BLASLONG, double *, BLASLONG, double *);
907 int (*zlaswp_ncopy) (BLASLONG, BLASLONG, BLASLONG, double *, BLASLONG, blasint *, double *);
912 int xgemm_p, xgemm_q, xgemm_r;
913 int xgemm_unroll_m, xgemm_unroll_n, xgemm_unroll_mn;
915 xdouble (*xamax_k) (BLASLONG, xdouble *, BLASLONG);
916 xdouble (*xamin_k) (BLASLONG, xdouble *, BLASLONG);
917 BLASLONG (*ixamax_k)(BLASLONG, xdouble *, BLASLONG);
918 BLASLONG (*ixamin_k)(BLASLONG, xdouble *, BLASLONG);
920 xdouble (*xnrm2_k) (BLASLONG, xdouble *, BLASLONG);
921 xdouble (*xasum_k) (BLASLONG, xdouble *, BLASLONG);
922 xdouble (*xsum_k) (BLASLONG, xdouble *, BLASLONG);
923 int (*xcopy_k) (BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG);
924 openblas_complex_xdouble (*xdotu_k) (BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG);
925 openblas_complex_xdouble (*xdotc_k) (BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG);
926 int (*xqrot_k) (BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble, xdouble);
928 int (*xaxpy_k) (BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG);
929 int (*xaxpyc_k)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG);
930 int (*xscal_k) (BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG);
931 int (*xswap_k) (BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG);
933 int (*xgemv_n) (BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *);
934 int (*xgemv_t) (BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *);
935 int (*xgemv_r) (BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *);
936 int (*xgemv_c) (BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *);
937 int (*xgemv_o) (BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *);
938 int (*xgemv_u) (BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *);
939 int (*xgemv_s) (BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *);
940 int (*xgemv_d) (BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *);
941 int (*xgeru_k) (BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *);
942 int (*xgerc_k) (BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *);
943 int (*xgerv_k) (BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *);
944 int (*xgerd_k) (BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *);
946 int (*xsymv_L) (BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *);
947 int (*xsymv_U) (BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *);
948 int (*xhemv_L) (BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *);
949 int (*xhemv_U) (BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *);
950 int (*xhemv_M) (BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *);
951 int (*xhemv_V) (BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *);
953 int (*xgemm_kernel_n )(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG);
954 int (*xgemm_kernel_l )(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG);
955 int (*xgemm_kernel_r )(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG);
956 int (*xgemm_kernel_b )(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG);
957 int (*xgemm_beta )(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG);
959 int (*xgemm_incopy )(BLASLONG, BLASLONG, xdouble *, BLASLONG, xdouble *);
960 int (*xgemm_itcopy )(BLASLONG, BLASLONG, xdouble *, BLASLONG, xdouble *);
961 int (*xgemm_oncopy )(BLASLONG, BLASLONG, xdouble *, BLASLONG, xdouble *);
962 int (*xgemm_otcopy )(BLASLONG, BLASLONG, xdouble *, BLASLONG, xdouble *);
964 int (*xtrsm_kernel_LN)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
965 int (*xtrsm_kernel_LT)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
966 int (*xtrsm_kernel_LR)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
967 int (*xtrsm_kernel_LC)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
968 int (*xtrsm_kernel_RN)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
969 int (*xtrsm_kernel_RT)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
970 int (*xtrsm_kernel_RR)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
971 int (*xtrsm_kernel_RC)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
973 int (*xtrsm_iunucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
974 int (*xtrsm_iunncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
975 int (*xtrsm_iutucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
976 int (*xtrsm_iutncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
977 int (*xtrsm_ilnucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
978 int (*xtrsm_ilnncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
979 int (*xtrsm_iltucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
980 int (*xtrsm_iltncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
981 int (*xtrsm_ounucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
982 int (*xtrsm_ounncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
983 int (*xtrsm_outucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
984 int (*xtrsm_outncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
985 int (*xtrsm_olnucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
986 int (*xtrsm_olnncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
987 int (*xtrsm_oltucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
988 int (*xtrsm_oltncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, xdouble *);
990 int (*xtrmm_kernel_RN)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
991 int (*xtrmm_kernel_RT)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
992 int (*xtrmm_kernel_RR)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
993 int (*xtrmm_kernel_RC)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
994 int (*xtrmm_kernel_LN)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
995 int (*xtrmm_kernel_LT)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
996 int (*xtrmm_kernel_LR)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
997 int (*xtrmm_kernel_LC)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG, BLASLONG);
999 int (*xtrmm_iunucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
1000 int (*xtrmm_iunncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
1001 int (*xtrmm_iutucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
1002 int (*xtrmm_iutncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
1003 int (*xtrmm_ilnucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
1004 int (*xtrmm_ilnncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
1005 int (*xtrmm_iltucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
1006 int (*xtrmm_iltncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
1007 int (*xtrmm_ounucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
1008 int (*xtrmm_ounncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
1009 int (*xtrmm_outucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
1010 int (*xtrmm_outncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
1011 int (*xtrmm_olnucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
1012 int (*xtrmm_olnncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
1013 int (*xtrmm_oltucopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
1014 int (*xtrmm_oltncopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
1016 int (*xsymm_iutcopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
1017 int (*xsymm_iltcopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
1018 int (*xsymm_outcopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
1019 int (*xsymm_oltcopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
1021 int (*xhemm_iutcopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
1022 int (*xhemm_iltcopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
1023 int (*xhemm_outcopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
1024 int (*xhemm_oltcopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
1026 int xgemm3m_p, xgemm3m_q, xgemm3m_r;
1027 int xgemm3m_unroll_m, xgemm3m_unroll_n, xgemm3m_unroll_mn;
1029 int (*xgemm3m_kernel)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG);
1031 int (*xgemm3m_incopyb)(BLASLONG, BLASLONG, xdouble *, BLASLONG, xdouble *);
1032 int (*xgemm3m_incopyr)(BLASLONG, BLASLONG, xdouble *, BLASLONG, xdouble *);
1033 int (*xgemm3m_incopyi)(BLASLONG, BLASLONG, xdouble *, BLASLONG, xdouble *);
1034 int (*xgemm3m_itcopyb)(BLASLONG, BLASLONG, xdouble *, BLASLONG, xdouble *);
1035 int (*xgemm3m_itcopyr)(BLASLONG, BLASLONG, xdouble *, BLASLONG, xdouble *);
1036 int (*xgemm3m_itcopyi)(BLASLONG, BLASLONG, xdouble *, BLASLONG, xdouble *);
1038 int (*xgemm3m_oncopyb)(BLASLONG, BLASLONG, xdouble *, BLASLONG, xdouble, xdouble, xdouble *);
1039 int (*xgemm3m_oncopyr)(BLASLONG, BLASLONG, xdouble *, BLASLONG, xdouble, xdouble, xdouble *);
1040 int (*xgemm3m_oncopyi)(BLASLONG, BLASLONG, xdouble *, BLASLONG, xdouble, xdouble, xdouble *);
1041 int (*xgemm3m_otcopyb)(BLASLONG, BLASLONG, xdouble *, BLASLONG, xdouble, xdouble, xdouble *);
1042 int (*xgemm3m_otcopyr)(BLASLONG, BLASLONG, xdouble *, BLASLONG, xdouble, xdouble, xdouble *);
1043 int (*xgemm3m_otcopyi)(BLASLONG, BLASLONG, xdouble *, BLASLONG, xdouble, xdouble, xdouble *);
1045 int (*xsymm3m_iucopyb)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
1046 int (*xsymm3m_ilcopyb)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
1047 int (*xsymm3m_iucopyr)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
1048 int (*xsymm3m_ilcopyr)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
1049 int (*xsymm3m_iucopyi)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
1050 int (*xsymm3m_ilcopyi)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
1052 int (*xsymm3m_oucopyb)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *);
1053 int (*xsymm3m_olcopyb)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *);
1054 int (*xsymm3m_oucopyr)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *);
1055 int (*xsymm3m_olcopyr)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *);
1056 int (*xsymm3m_oucopyi)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *);
1057 int (*xsymm3m_olcopyi)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *);
1059 int (*xhemm3m_iucopyb)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
1060 int (*xhemm3m_ilcopyb)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
1061 int (*xhemm3m_iucopyr)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
1062 int (*xhemm3m_ilcopyr)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
1063 int (*xhemm3m_iucopyi)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
1064 int (*xhemm3m_ilcopyi)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
1066 int (*xhemm3m_oucopyb)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *);
1067 int (*xhemm3m_olcopyb)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *);
1068 int (*xhemm3m_oucopyr)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *);
1069 int (*xhemm3m_olcopyr)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *);
1070 int (*xhemm3m_oucopyi)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *);
1071 int (*xhemm3m_olcopyi)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *);
1073 int (*xneg_tcopy) (BLASLONG, BLASLONG, xdouble *, BLASLONG, xdouble *);
1074 int (*xlaswp_ncopy) (BLASLONG, BLASLONG, BLASLONG, xdouble *, BLASLONG, blasint *, xdouble *);
1081 int snum_opt, dnum_opt, qnum_opt;
1083 int (*saxpby_k) (BLASLONG, float, float*, BLASLONG,float, float*, BLASLONG);
1086 int (*daxpby_k) (BLASLONG, double, double*, BLASLONG,double, double*, BLASLONG);
1088 #ifdef BUILD_COMPLEX
1089 int (*caxpby_k) (BLASLONG, float, float, float*, BLASLONG,float,float, float*, BLASLONG);
1091 #ifdef BUILD_COMPLEX16
1092 int (*zaxpby_k) (BLASLONG, double, double, double*, BLASLONG,double,double, double*, BLASLONG);
1096 int (*somatcopy_k_cn) (BLASLONG, BLASLONG, float, float*, BLASLONG, float*, BLASLONG);
1097 int (*somatcopy_k_ct) (BLASLONG, BLASLONG, float, float*, BLASLONG, float*, BLASLONG);
1098 int (*somatcopy_k_rn) (BLASLONG, BLASLONG, float, float*, BLASLONG, float*, BLASLONG);
1099 int (*somatcopy_k_rt) (BLASLONG, BLASLONG, float, float*, BLASLONG, float*, BLASLONG);
1103 int (*domatcopy_k_cn) (BLASLONG, BLASLONG, double, double*, BLASLONG, double*, BLASLONG);
1104 int (*domatcopy_k_ct) (BLASLONG, BLASLONG, double, double*, BLASLONG, double*, BLASLONG);
1105 int (*domatcopy_k_rn) (BLASLONG, BLASLONG, double, double*, BLASLONG, double*, BLASLONG);
1106 int (*domatcopy_k_rt) (BLASLONG, BLASLONG, double, double*, BLASLONG, double*, BLASLONG);
1109 #ifdef BUILD_COMPLEX
1110 int (*comatcopy_k_cn) (BLASLONG, BLASLONG, float, float, float*, BLASLONG, float*, BLASLONG);
1111 int (*comatcopy_k_ct) (BLASLONG, BLASLONG, float, float, float*, BLASLONG, float*, BLASLONG);
1112 int (*comatcopy_k_rn) (BLASLONG, BLASLONG, float, float, float*, BLASLONG, float*, BLASLONG);
1113 int (*comatcopy_k_rt) (BLASLONG, BLASLONG, float, float, float*, BLASLONG, float*, BLASLONG);
1115 int (*comatcopy_k_cnc) (BLASLONG, BLASLONG, float, float, float*, BLASLONG, float*, BLASLONG);
1116 int (*comatcopy_k_ctc) (BLASLONG, BLASLONG, float, float, float*, BLASLONG, float*, BLASLONG);
1117 int (*comatcopy_k_rnc) (BLASLONG, BLASLONG, float, float, float*, BLASLONG, float*, BLASLONG);
1118 int (*comatcopy_k_rtc) (BLASLONG, BLASLONG, float, float, float*, BLASLONG, float*, BLASLONG);
1121 #ifdef BUILD_COMPLEX16
1122 int (*zomatcopy_k_cn) (BLASLONG, BLASLONG, double, double, double*, BLASLONG, double*, BLASLONG);
1123 int (*zomatcopy_k_ct) (BLASLONG, BLASLONG, double, double, double*, BLASLONG, double*, BLASLONG);
1124 int (*zomatcopy_k_rn) (BLASLONG, BLASLONG, double, double, double*, BLASLONG, double*, BLASLONG);
1125 int (*zomatcopy_k_rt) (BLASLONG, BLASLONG, double, double, double*, BLASLONG, double*, BLASLONG);
1127 int (*zomatcopy_k_cnc) (BLASLONG, BLASLONG, double, double, double*, BLASLONG, double*, BLASLONG);
1128 int (*zomatcopy_k_ctc) (BLASLONG, BLASLONG, double, double, double*, BLASLONG, double*, BLASLONG);
1129 int (*zomatcopy_k_rnc) (BLASLONG, BLASLONG, double, double, double*, BLASLONG, double*, BLASLONG);
1130 int (*zomatcopy_k_rtc) (BLASLONG, BLASLONG, double, double, double*, BLASLONG, double*, BLASLONG);
1134 int (*simatcopy_k_cn) (BLASLONG, BLASLONG, float, float*, BLASLONG);
1135 int (*simatcopy_k_ct) (BLASLONG, BLASLONG, float, float*, BLASLONG);
1136 int (*simatcopy_k_rn) (BLASLONG, BLASLONG, float, float*, BLASLONG);
1137 int (*simatcopy_k_rt) (BLASLONG, BLASLONG, float, float*, BLASLONG);
1141 int (*dimatcopy_k_cn) (BLASLONG, BLASLONG, double, double*, BLASLONG);
1142 int (*dimatcopy_k_ct) (BLASLONG, BLASLONG, double, double*, BLASLONG);
1143 int (*dimatcopy_k_rn) (BLASLONG, BLASLONG, double, double*, BLASLONG);
1144 int (*dimatcopy_k_rt) (BLASLONG, BLASLONG, double, double*, BLASLONG);
1147 #ifdef BUILD_COMPLEX
1148 int (*cimatcopy_k_cn) (BLASLONG, BLASLONG, float, float, float*, BLASLONG);
1149 int (*cimatcopy_k_ct) (BLASLONG, BLASLONG, float, float, float*, BLASLONG);
1150 int (*cimatcopy_k_rn) (BLASLONG, BLASLONG, float, float, float*, BLASLONG);
1151 int (*cimatcopy_k_rt) (BLASLONG, BLASLONG, float, float, float*, BLASLONG);
1153 int (*cimatcopy_k_cnc) (BLASLONG, BLASLONG, float, float, float*, BLASLONG);
1154 int (*cimatcopy_k_ctc) (BLASLONG, BLASLONG, float, float, float*, BLASLONG);
1155 int (*cimatcopy_k_rnc) (BLASLONG, BLASLONG, float, float, float*, BLASLONG);
1156 int (*cimatcopy_k_rtc) (BLASLONG, BLASLONG, float, float, float*, BLASLONG);
1159 #ifdef BUILD_COMPLEX16
1160 int (*zimatcopy_k_cn) (BLASLONG, BLASLONG, double, double, double*, BLASLONG);
1161 int (*zimatcopy_k_ct) (BLASLONG, BLASLONG, double, double, double*, BLASLONG);
1162 int (*zimatcopy_k_rn) (BLASLONG, BLASLONG, double, double, double*, BLASLONG);
1163 int (*zimatcopy_k_rt) (BLASLONG, BLASLONG, double, double, double*, BLASLONG);
1165 int (*zimatcopy_k_cnc) (BLASLONG, BLASLONG, double, double, double*, BLASLONG);
1166 int (*zimatcopy_k_ctc) (BLASLONG, BLASLONG, double, double, double*, BLASLONG);
1167 int (*zimatcopy_k_rnc) (BLASLONG, BLASLONG, double, double, double*, BLASLONG);
1168 int (*zimatcopy_k_rtc) (BLASLONG, BLASLONG, double, double, double*, BLASLONG);
1172 int (*sgeadd_k) (BLASLONG, BLASLONG, float, float *, BLASLONG, float, float *, BLASLONG);
1175 int (*dgeadd_k) (BLASLONG, BLASLONG, double, double *, BLASLONG, double, double *, BLASLONG);
1177 #ifdef BUILD_COMPLEX
1178 int (*cgeadd_k) (BLASLONG, BLASLONG, float, float, float *, BLASLONG, float, float, float *, BLASLONG);
1180 #ifdef BUILD_COMPLEX16
1181 int (*zgeadd_k) (BLASLONG, BLASLONG, double, double, double *, BLASLONG, double, double, double *, BLASLONG);
1185 extern gotoblas_t *gotoblas;
1187 #define FUNC_OFFSET(func) (size_t)(&((gotoblas_t *)NULL)->func)
1189 #define DTB_ENTRIES gotoblas -> dtb_entries
1190 #define GEMM_OFFSET_A gotoblas -> offsetA
1191 #define GEMM_OFFSET_B gotoblas -> offsetB
1192 #define GEMM_ALIGN gotoblas -> align
1194 #define HAVE_EX_L2 gotoblas -> exclusive_cache
1196 #ifdef BUILD_BFLOAT16
1197 #define SBGEMM_P gotoblas -> sbgemm_p
1198 #define SBGEMM_Q gotoblas -> sbgemm_q
1199 #define SBGEMM_R gotoblas -> sbgemm_r
1200 #define SBGEMM_UNROLL_M gotoblas -> sbgemm_unroll_m
1201 #define SBGEMM_UNROLL_N gotoblas -> sbgemm_unroll_n
1202 #define SBGEMM_UNROLL_MN gotoblas -> sbgemm_unroll_mn
1205 #if defined (BUILD_SINGLE)
1206 #define SGEMM_P gotoblas -> sgemm_p
1207 #define SGEMM_Q gotoblas -> sgemm_q
1208 #define SGEMM_R gotoblas -> sgemm_r
1209 #define SGEMM_UNROLL_M gotoblas -> sgemm_unroll_m
1210 #define SGEMM_UNROLL_N gotoblas -> sgemm_unroll_n
1211 #define SGEMM_UNROLL_MN gotoblas -> sgemm_unroll_mn
1214 #if defined (BUILD_DOUBLE)
1215 #define DGEMM_P gotoblas -> dgemm_p
1216 #define DGEMM_Q gotoblas -> dgemm_q
1217 #define DGEMM_R gotoblas -> dgemm_r
1218 #define DGEMM_UNROLL_M gotoblas -> dgemm_unroll_m
1219 #define DGEMM_UNROLL_N gotoblas -> dgemm_unroll_n
1220 #define DGEMM_UNROLL_MN gotoblas -> dgemm_unroll_mn
1223 #define QGEMM_P gotoblas -> qgemm_p
1224 #define QGEMM_Q gotoblas -> qgemm_q
1225 #define QGEMM_R gotoblas -> qgemm_r
1226 #define QGEMM_UNROLL_M gotoblas -> qgemm_unroll_m
1227 #define QGEMM_UNROLL_N gotoblas -> qgemm_unroll_n
1228 #define QGEMM_UNROLL_MN gotoblas -> qgemm_unroll_mn
1230 #ifdef BUILD_COMPLEX
1231 #define CGEMM_P gotoblas -> cgemm_p
1232 #define CGEMM_Q gotoblas -> cgemm_q
1233 #define CGEMM_R gotoblas -> cgemm_r
1234 #define CGEMM_UNROLL_M gotoblas -> cgemm_unroll_m
1235 #define CGEMM_UNROLL_N gotoblas -> cgemm_unroll_n
1236 #define CGEMM_UNROLL_MN gotoblas -> cgemm_unroll_mn
1237 #ifndef BUILD_SINGLE
1238 #define SGEMM_P gotoblas -> sgemm_p
1239 #define SGEMM_Q gotoblas -> sgemm_q
1240 #define SGEMM_R 1024
1241 #define SGEMM_UNROLL_M gotoblas -> sgemm_unroll_m
1242 #define SGEMM_UNROLL_N gotoblas -> sgemm_unroll_n
1243 #define SGEMM_UNROLL_MN gotoblas -> sgemm_unroll_mn
1247 #ifdef BUILD_COMPLEX16
1248 #define ZGEMM_P gotoblas -> zgemm_p
1249 #define ZGEMM_Q gotoblas -> zgemm_q
1250 #define ZGEMM_R gotoblas -> zgemm_r
1251 #define ZGEMM_UNROLL_M gotoblas -> zgemm_unroll_m
1252 #define ZGEMM_UNROLL_N gotoblas -> zgemm_unroll_n
1253 #define ZGEMM_UNROLL_MN gotoblas -> zgemm_unroll_mn
1254 #ifndef BUILD_DOUBLE
1255 #define DGEMM_P gotoblas -> dgemm_p
1256 #define DGEMM_Q gotoblas -> dgemm_q
1257 #define DGEMM_R 1024
1258 #define DGEMM_UNROLL_M gotoblas -> dgemm_unroll_m
1259 #define DGEMM_UNROLL_N gotoblas -> dgemm_unroll_n
1260 #define DGEMM_UNROLL_MN gotoblas -> dgemm_unroll_mn
1264 #define XGEMM_P gotoblas -> xgemm_p
1265 #define XGEMM_Q gotoblas -> xgemm_q
1266 #define XGEMM_R gotoblas -> xgemm_r
1267 #define XGEMM_UNROLL_M gotoblas -> xgemm_unroll_m
1268 #define XGEMM_UNROLL_N gotoblas -> xgemm_unroll_n
1269 #define XGEMM_UNROLL_MN gotoblas -> xgemm_unroll_mn
1271 #define CGEMM3M_P gotoblas -> cgemm3m_p
1272 #define CGEMM3M_Q gotoblas -> cgemm3m_q
1273 #define CGEMM3M_R gotoblas -> cgemm3m_r
1274 #define CGEMM3M_UNROLL_M gotoblas -> cgemm3m_unroll_m
1275 #define CGEMM3M_UNROLL_N gotoblas -> cgemm3m_unroll_n
1276 #define CGEMM3M_UNROLL_MN gotoblas -> cgemm3m_unroll_mn
1278 #define ZGEMM3M_P gotoblas -> zgemm3m_p
1279 #define ZGEMM3M_Q gotoblas -> zgemm3m_q
1280 #define ZGEMM3M_R gotoblas -> zgemm3m_r
1281 #define ZGEMM3M_UNROLL_M gotoblas -> zgemm3m_unroll_m
1282 #define ZGEMM3M_UNROLL_N gotoblas -> zgemm3m_unroll_n
1283 #define ZGEMM3M_UNROLL_MN gotoblas -> zgemm3m_unroll_mn
1285 #define XGEMM3M_P gotoblas -> xgemm3m_p
1286 #define XGEMM3M_Q gotoblas -> xgemm3m_q
1287 #define XGEMM3M_R gotoblas -> xgemm3m_r
1288 #define XGEMM3M_UNROLL_M gotoblas -> xgemm3m_unroll_m
1289 #define XGEMM3M_UNROLL_N gotoblas -> xgemm3m_unroll_n
1290 #define XGEMM3M_UNROLL_MN gotoblas -> xgemm3m_unroll_mn
1294 #define FUNC_OFFSET(func) (size_t)(func)
1296 #define DTB_ENTRIES DTB_DEFAULT_ENTRIES
1298 #define GEMM_OFFSET_A GEMM_DEFAULT_OFFSET_A
1299 #define GEMM_OFFSET_B GEMM_DEFAULT_OFFSET_B
1300 #define GEMM_ALIGN GEMM_DEFAULT_ALIGN
1302 #ifdef HAVE_EXCLUSIVE_CACHE
1303 #define HAVE_EX_L2 1
1305 #define HAVE_EX_L2 0
1308 #ifdef BUILD_BFLOAT16
1309 #define SBGEMM_P SBGEMM_DEFAULT_P
1310 #define SBGEMM_Q SBGEMM_DEFAULT_Q
1311 #define SBGEMM_R SBGEMM_DEFAULT_R
1312 #define SBGEMM_UNROLL_M SBGEMM_DEFAULT_UNROLL_M
1313 #define SBGEMM_UNROLL_N SBGEMM_DEFAULT_UNROLL_N
1314 #ifdef SBGEMM_DEFAULT_UNROLL_MN
1315 #define SBGEMM_UNROLL_MN SBGEMM_DEFAULT_UNROLL_MN
1317 #define SBGEMM_UNROLL_MN MAX((SBGEMM_UNROLL_M), (SBGEMM_UNROLL_N))
1321 #define SGEMM_P SGEMM_DEFAULT_P
1322 #define SGEMM_Q SGEMM_DEFAULT_Q
1323 #define SGEMM_R SGEMM_DEFAULT_R
1324 #define SGEMM_UNROLL_M SGEMM_DEFAULT_UNROLL_M
1325 #define SGEMM_UNROLL_N SGEMM_DEFAULT_UNROLL_N
1326 #ifdef SGEMM_DEFAULT_UNROLL_MN
1327 #define SGEMM_UNROLL_MN SGEMM_DEFAULT_UNROLL_MN
1329 #define SGEMM_UNROLL_MN MAX((SGEMM_UNROLL_M), (SGEMM_UNROLL_N))
1332 #define DGEMM_P DGEMM_DEFAULT_P
1333 #define DGEMM_Q DGEMM_DEFAULT_Q
1334 #define DGEMM_R DGEMM_DEFAULT_R
1335 #define DGEMM_UNROLL_M DGEMM_DEFAULT_UNROLL_M
1336 #define DGEMM_UNROLL_N DGEMM_DEFAULT_UNROLL_N
1337 #ifdef DGEMM_DEFAULT_UNROLL_MN
1338 #define DGEMM_UNROLL_MN DGEMM_DEFAULT_UNROLL_MN
1340 #define DGEMM_UNROLL_MN MAX((DGEMM_UNROLL_M), (DGEMM_UNROLL_N))
1343 #define QGEMM_P QGEMM_DEFAULT_P
1344 #define QGEMM_Q QGEMM_DEFAULT_Q
1345 #define QGEMM_R QGEMM_DEFAULT_R
1346 #define QGEMM_UNROLL_M QGEMM_DEFAULT_UNROLL_M
1347 #define QGEMM_UNROLL_N QGEMM_DEFAULT_UNROLL_N
1348 #define QGEMM_UNROLL_MN MAX((QGEMM_UNROLL_M), (QGEMM_UNROLL_N))
1350 #define CGEMM_P CGEMM_DEFAULT_P
1351 #define CGEMM_Q CGEMM_DEFAULT_Q
1352 #define CGEMM_R CGEMM_DEFAULT_R
1353 #define CGEMM_UNROLL_M CGEMM_DEFAULT_UNROLL_M
1354 #define CGEMM_UNROLL_N CGEMM_DEFAULT_UNROLL_N
1355 #ifdef CGEMM_DEFAULT_UNROLL_MN
1356 #define CGEMM_UNROLL_MN CGEMM_DEFAULT_UNROLL_MN
1358 #define CGEMM_UNROLL_MN MAX((CGEMM_UNROLL_M), (CGEMM_UNROLL_N))
1361 #define ZGEMM_P ZGEMM_DEFAULT_P
1362 #define ZGEMM_Q ZGEMM_DEFAULT_Q
1363 #define ZGEMM_R ZGEMM_DEFAULT_R
1364 #define ZGEMM_UNROLL_M ZGEMM_DEFAULT_UNROLL_M
1365 #define ZGEMM_UNROLL_N ZGEMM_DEFAULT_UNROLL_N
1366 #ifdef ZGEMM_DEFAULT_UNROLL_MN
1367 #define ZGEMM_UNROLL_MN ZGEMM_DEFAULT_UNROLL_MN
1369 #define ZGEMM_UNROLL_MN MAX((ZGEMM_UNROLL_M), (ZGEMM_UNROLL_N))
1372 #define XGEMM_P XGEMM_DEFAULT_P
1373 #define XGEMM_Q XGEMM_DEFAULT_Q
1374 #define XGEMM_R XGEMM_DEFAULT_R
1375 #define XGEMM_UNROLL_M XGEMM_DEFAULT_UNROLL_M
1376 #define XGEMM_UNROLL_N XGEMM_DEFAULT_UNROLL_N
1377 #define XGEMM_UNROLL_MN MAX((XGEMM_UNROLL_M), (XGEMM_UNROLL_N))
1379 #ifdef CGEMM3M_DEFAULT_UNROLL_N
1381 #define CGEMM3M_P CGEMM3M_DEFAULT_P
1382 #define CGEMM3M_Q CGEMM3M_DEFAULT_Q
1383 #define CGEMM3M_R CGEMM3M_DEFAULT_R
1384 #define CGEMM3M_UNROLL_M CGEMM3M_DEFAULT_UNROLL_M
1385 #define CGEMM3M_UNROLL_N CGEMM3M_DEFAULT_UNROLL_N
1386 #define CGEMM3M_UNROLL_MN MAX((CGEMM3M_UNROLL_M), (CGEMM3M_UNROLL_N))
1390 #define CGEMM3M_P SGEMM_DEFAULT_P
1391 #define CGEMM3M_Q SGEMM_DEFAULT_Q
1392 #define CGEMM3M_R SGEMM_DEFAULT_R
1393 #define CGEMM3M_UNROLL_M SGEMM_DEFAULT_UNROLL_M
1394 #define CGEMM3M_UNROLL_N SGEMM_DEFAULT_UNROLL_N
1395 #define CGEMM3M_UNROLL_MN MAX((CGEMM_UNROLL_M), (CGEMM_UNROLL_N))
1400 #ifdef ZGEMM3M_DEFAULT_UNROLL_N
1402 #define ZGEMM3M_P ZGEMM3M_DEFAULT_P
1403 #define ZGEMM3M_Q ZGEMM3M_DEFAULT_Q
1404 #define ZGEMM3M_R ZGEMM3M_DEFAULT_R
1405 #define ZGEMM3M_UNROLL_M ZGEMM3M_DEFAULT_UNROLL_M
1406 #define ZGEMM3M_UNROLL_N ZGEMM3M_DEFAULT_UNROLL_N
1407 #define ZGEMM3M_UNROLL_MN MAX((ZGEMM_UNROLL_M), (ZGEMM_UNROLL_N))
1411 #define ZGEMM3M_P DGEMM_DEFAULT_P
1412 #define ZGEMM3M_Q DGEMM_DEFAULT_Q
1413 #define ZGEMM3M_R DGEMM_DEFAULT_R
1414 #define ZGEMM3M_UNROLL_M DGEMM_DEFAULT_UNROLL_M
1415 #define ZGEMM3M_UNROLL_N DGEMM_DEFAULT_UNROLL_N
1416 #define ZGEMM3M_UNROLL_MN MAX((ZGEMM_UNROLL_M), (ZGEMM_UNROLL_N))
1420 #define XGEMM3M_P QGEMM_DEFAULT_P
1421 #define XGEMM3M_Q QGEMM_DEFAULT_Q
1422 #define XGEMM3M_R QGEMM_DEFAULT_R
1423 #define XGEMM3M_UNROLL_M QGEMM_DEFAULT_UNROLL_M
1424 #define XGEMM3M_UNROLL_N QGEMM_DEFAULT_UNROLL_N
1425 #define XGEMM3M_UNROLL_MN MAX((QGEMM_UNROLL_M), (QGEMM_UNROLL_N))
1432 #if defined(XDOUBLE)
1433 #define GEMM_P QGEMM_P
1434 #define GEMM_Q QGEMM_Q
1435 #define GEMM_R QGEMM_R
1436 #define GEMM_UNROLL_M QGEMM_UNROLL_M
1437 #define GEMM_UNROLL_N QGEMM_UNROLL_N
1438 #define GEMM_UNROLL_MN QGEMM_UNROLL_MN
1439 #define GEMM_DEFAULT_P QGEMM_DEFAULT_P
1440 #define GEMM_DEFAULT_Q QGEMM_DEFAULT_Q
1441 #define GEMM_DEFAULT_R QGEMM_DEFAULT_R
1442 #define GEMM_DEFAULT_UNROLL_M QGEMM_DEFAULT_UNROLL_M
1443 #define GEMM_DEFAULT_UNROLL_N QGEMM_DEFAULT_UNROLL_N
1444 #elif defined(DOUBLE)
1445 #define GEMM_P DGEMM_P
1446 #define GEMM_Q DGEMM_Q
1447 #define GEMM_R DGEMM_R
1448 #define GEMM_UNROLL_M DGEMM_UNROLL_M
1449 #define GEMM_UNROLL_N DGEMM_UNROLL_N
1450 #define GEMM_UNROLL_MN DGEMM_UNROLL_MN
1451 #define GEMM_DEFAULT_P DGEMM_DEFAULT_P
1452 #define GEMM_DEFAULT_Q DGEMM_DEFAULT_Q
1453 #define GEMM_DEFAULT_R DGEMM_DEFAULT_R
1454 #define GEMM_DEFAULT_UNROLL_M DGEMM_DEFAULT_UNROLL_M
1455 #define GEMM_DEFAULT_UNROLL_N DGEMM_DEFAULT_UNROLL_N
1456 #elif defined(BFLOAT16)
1457 #define GEMM_P SBGEMM_P
1458 #define GEMM_Q SBGEMM_Q
1459 #define GEMM_R SBGEMM_R
1460 #define GEMM_UNROLL_M SBGEMM_UNROLL_M
1461 #define GEMM_UNROLL_N SBGEMM_UNROLL_N
1462 #define GEMM_UNROLL_MN SBGEMM_UNROLL_MN
1463 #define GEMM_DEFAULT_P SBGEMM_DEFAULT_P
1464 #define GEMM_DEFAULT_Q SBGEMM_DEFAULT_Q
1465 #define GEMM_DEFAULT_R SBGEMM_DEFAULT_R
1466 #define GEMM_DEFAULT_UNROLL_M SBGEMM_DEFAULT_UNROLL_M
1467 #define GEMM_DEFAULT_UNROLL_N SBGEMM_DEFAULT_UNROLL_N
1469 #define GEMM_P SGEMM_P
1470 #define GEMM_Q SGEMM_Q
1471 #define GEMM_R SGEMM_R
1472 #define GEMM_UNROLL_M SGEMM_UNROLL_M
1473 #define GEMM_UNROLL_N SGEMM_UNROLL_N
1474 #define GEMM_UNROLL_MN SGEMM_UNROLL_MN
1475 #define GEMM_DEFAULT_P SGEMM_DEFAULT_P
1476 #define GEMM_DEFAULT_Q SGEMM_DEFAULT_Q
1477 #define GEMM_DEFAULT_R SGEMM_DEFAULT_R
1478 #define GEMM_DEFAULT_UNROLL_M SGEMM_DEFAULT_UNROLL_M
1479 #define GEMM_DEFAULT_UNROLL_N SGEMM_DEFAULT_UNROLL_N
1482 #if defined(XDOUBLE)
1483 #define GEMM_P XGEMM_P
1484 #define GEMM_Q XGEMM_Q
1485 #define GEMM_R XGEMM_R
1486 #define GEMM_UNROLL_M XGEMM_UNROLL_M
1487 #define GEMM_UNROLL_N XGEMM_UNROLL_N
1488 #define GEMM_UNROLL_MN XGEMM_UNROLL_MN
1489 #define GEMM_DEFAULT_P XGEMM_DEFAULT_P
1490 #define GEMM_DEFAULT_Q XGEMM_DEFAULT_Q
1491 #define GEMM_DEFAULT_R XGEMM_DEFAULT_R
1492 #define GEMM_DEFAULT_UNROLL_M XGEMM_DEFAULT_UNROLL_M
1493 #define GEMM_DEFAULT_UNROLL_N XGEMM_DEFAULT_UNROLL_N
1494 #elif defined(DOUBLE)
1495 #define GEMM_P ZGEMM_P
1496 #define GEMM_Q ZGEMM_Q
1497 #define GEMM_R ZGEMM_R
1498 #define GEMM_UNROLL_M ZGEMM_UNROLL_M
1499 #define GEMM_UNROLL_N ZGEMM_UNROLL_N
1500 #define GEMM_UNROLL_MN ZGEMM_UNROLL_MN
1501 #define GEMM_DEFAULT_P ZGEMM_DEFAULT_P
1502 #define GEMM_DEFAULT_Q ZGEMM_DEFAULT_Q
1503 #define GEMM_DEFAULT_R ZGEMM_DEFAULT_R
1504 #define GEMM_DEFAULT_UNROLL_M ZGEMM_DEFAULT_UNROLL_M
1505 #define GEMM_DEFAULT_UNROLL_N ZGEMM_DEFAULT_UNROLL_N
1507 #define GEMM_P CGEMM_P
1508 #define GEMM_Q CGEMM_Q
1509 #define GEMM_R CGEMM_R
1510 #define GEMM_UNROLL_M CGEMM_UNROLL_M
1511 #define GEMM_UNROLL_N CGEMM_UNROLL_N
1512 #define GEMM_UNROLL_MN CGEMM_UNROLL_MN
1513 #define GEMM_DEFAULT_P CGEMM_DEFAULT_P
1514 #define GEMM_DEFAULT_Q CGEMM_DEFAULT_Q
1515 #define GEMM_DEFAULT_R CGEMM_DEFAULT_R
1516 #define GEMM_DEFAULT_UNROLL_M CGEMM_DEFAULT_UNROLL_M
1517 #define GEMM_DEFAULT_UNROLL_N CGEMM_DEFAULT_UNROLL_N
1522 #define GEMM3M_UNROLL_M XGEMM3M_UNROLL_M
1523 #define GEMM3M_UNROLL_N XGEMM3M_UNROLL_N
1524 #elif defined(DOUBLE)
1525 #define GEMM3M_UNROLL_M ZGEMM3M_UNROLL_M
1526 #define GEMM3M_UNROLL_N ZGEMM3M_UNROLL_N
1528 #define GEMM3M_UNROLL_M CGEMM3M_UNROLL_M
1529 #define GEMM3M_UNROLL_N CGEMM3M_UNROLL_N
1533 #ifndef QGEMM_DEFAULT_UNROLL_M
1534 #define QGEMM_DEFAULT_UNROLL_M 2
1537 #ifndef QGEMM_DEFAULT_UNROLL_N
1538 #define QGEMM_DEFAULT_UNROLL_N 2
1541 #ifndef XGEMM_DEFAULT_UNROLL_M
1542 #define XGEMM_DEFAULT_UNROLL_M 2
1545 #ifndef XGEMM_DEFAULT_UNROLL_N
1546 #define XGEMM_DEFAULT_UNROLL_N 2
1550 #define GEMM_THREAD gemm_thread_n
1553 #ifndef SBGEMM_DEFAULT_R
1554 #define SBGEMM_DEFAULT_R (((BUFFER_SIZE - ((SBGEMM_DEFAULT_P * SBGEMM_DEFAULT_Q * 4 + GEMM_DEFAULT_OFFSET_A + GEMM_DEFAULT_ALIGN) & ~GEMM_DEFAULT_ALIGN)) / (SBGEMM_DEFAULT_Q * 4) - 15) & ~15UL)
1557 #ifndef SGEMM_DEFAULT_R
1558 #define SGEMM_DEFAULT_R (((BUFFER_SIZE - ((SGEMM_DEFAULT_P * SGEMM_DEFAULT_Q * 4 + GEMM_DEFAULT_OFFSET_A + GEMM_DEFAULT_ALIGN) & ~GEMM_DEFAULT_ALIGN)) / (SGEMM_DEFAULT_Q * 4) - 15) & ~15UL)
1561 #ifndef DGEMM_DEFAULT_R
1562 #define DGEMM_DEFAULT_R (((BUFFER_SIZE - ((DGEMM_DEFAULT_P * DGEMM_DEFAULT_Q * 8 + GEMM_DEFAULT_OFFSET_A + GEMM_DEFAULT_ALIGN) & ~GEMM_DEFAULT_ALIGN)) / (DGEMM_DEFAULT_Q * 8) - 15) & ~15UL)
1565 #ifndef QGEMM_DEFAULT_R
1566 #define QGEMM_DEFAULT_R (((BUFFER_SIZE - ((QGEMM_DEFAULT_P * QGEMM_DEFAULT_Q * 16 + GEMM_DEFAULT_OFFSET_A + GEMM_DEFAULT_ALIGN) & ~GEMM_DEFAULT_ALIGN)) / (QGEMM_DEFAULT_Q * 16) - 15) & ~15UL)
1569 #ifndef CGEMM_DEFAULT_R
1570 #define CGEMM_DEFAULT_R (((BUFFER_SIZE - ((CGEMM_DEFAULT_P * CGEMM_DEFAULT_Q * 8 + GEMM_DEFAULT_OFFSET_A + GEMM_DEFAULT_ALIGN) & ~GEMM_DEFAULT_ALIGN)) / (CGEMM_DEFAULT_Q * 8) - 15) & ~15UL)
1573 #ifndef ZGEMM_DEFAULT_R
1574 #define ZGEMM_DEFAULT_R (((BUFFER_SIZE - ((ZGEMM_DEFAULT_P * ZGEMM_DEFAULT_Q * 16 + GEMM_DEFAULT_OFFSET_A + GEMM_DEFAULT_ALIGN) & ~GEMM_DEFAULT_ALIGN)) / (ZGEMM_DEFAULT_Q * 16) - 15) & ~15UL)
1577 #ifndef XGEMM_DEFAULT_R
1578 #define XGEMM_DEFAULT_R (((BUFFER_SIZE - ((XGEMM_DEFAULT_P * XGEMM_DEFAULT_Q * 32 + GEMM_DEFAULT_OFFSET_A + GEMM_DEFAULT_ALIGN) & ~GEMM_DEFAULT_ALIGN)) / (XGEMM_DEFAULT_Q * 32) - 15) & ~15UL)
1595 #define GEMM3M_P XGEMM3M_P
1596 #elif defined(DOUBLE)
1597 #define GEMM3M_P ZGEMM3M_P
1599 #define GEMM3M_P CGEMM3M_P
1605 #define GEMM3M_Q XGEMM3M_Q
1606 #elif defined(DOUBLE)
1607 #define GEMM3M_Q ZGEMM3M_Q
1609 #define GEMM3M_Q CGEMM3M_Q
1615 #define GEMM3M_R XGEMM3M_R
1616 #elif defined(DOUBLE)
1617 #define GEMM3M_R ZGEMM3M_R
1619 #define GEMM3M_R CGEMM3M_R