blasfeo-0.1.2/benchmarks/benchmark_d_blasfeo_api.c

87 void dgemm_nn_1_1_1(double alpha, double *A, int sda, double *B, int sdb, double beta, double *C, i…
93 void dgemm_nn_2_2_2(double alpha, double *A, int sda, double *B, int sdb, double beta, double *C, i…
99 void dgemm_nn_3_3_3(double alpha, double *A, int sda, double *B, int sdb, double beta, double *C, i…
105 void dgemm_nn_4_4_4(double alpha, double *A, int sda, double *B, int sdb, double beta, double *C, i…
111 void dgemm_nn_5_5_5(double alpha, double *A, int sda, double *B, int sdb, double beta, double *C, i…
113 	kernel_dgemm_nn_6x6_vs_lib4(5, &alpha, A, sda, 0, B, sdb, &beta, C, sdc, D, sdd, 5, 5);
117 void dgemm_nn_6_6_6(double alpha, double *A, int sda, double *B, int sdb, double beta, double *C, i…
119 	kernel_dgemm_nn_6x6_lib4(6, &alpha, A, sda, 0, B, sdb, &beta, C, sdc, D, sdd);
123 void dgemm_nn_7_7_7(double alpha, double *A, int sda, double *B, int sdb, double beta, double *C, i…
125 	kernel_dgemm_nn_8x4_vs_lib4(7, &alpha, A, sda, 0, B, sdb, &beta, C, sdd, D, sdd, 7, 4);
126 	kernel_dgemm_nn_8x4_vs_lib4(7, &alpha, A, sda, 0, B+4*4, sdb, &beta, C+4*4, sdd, D+4*4, sdd, 7, 3);
130 void dgemm_nn_8_8_8(double alpha, double *A, int sda, double *B, int sdb, double beta, double *C, i…
132 	kernel_dgemm_nn_8x4_lib4(8, &alpha, A, sda, 0, B, sdb, &beta, C, sdd, D, sdd);
133 	kernel_dgemm_nn_8x4_lib4(8, &alpha, A, sda, 0, B+4*4, sdb, &beta, C+4*4, sdd, D+4*4, sdd);
137 void dgemm_nn_9_9_9(double alpha, double *A, int sda, double *B, int sdb, double beta, double *C, i…
139 	kernel_dgemm_nn_10x4_vs_lib4(9, &alpha, A, sda, 0, B, sdb, &beta, C, sdd, D, sdd, 9, 4);
140 …kernel_dgemm_nn_10x4_vs_lib4(9, &alpha, A, sda, 0, B+4*4, sdb, &beta, C+4*4, sdd, D+4*4, sdd, 9, 4…
141 …kernel_dgemm_nn_10x2_vs_lib4(9, &alpha, A, sda, 0, B+8*4, sdb, &beta, C+8*4, sdd, D+8*4, sdd, 9, 1…
145 void dgemm_nn_10_10_10(double alpha, double *A, int sda, double *B, int sdb, double beta, double *C…
147 	kernel_dgemm_nn_10x4_lib4(10, &alpha, A, sda, 0, B, sdb, &beta, C, sdd, D, sdd);
148 	kernel_dgemm_nn_10x4_lib4(10, &alpha, A, sda, 0, B+4*4, sdb, &beta, C+4*4, sdd, D+4*4, sdd);
149 	kernel_dgemm_nn_10x2_lib4(10, &alpha, A, sda, 0, B+8*4, sdb, &beta, C+8*4, sdd, D+8*4, sdd);
153 void dgemm_nn_11_11_11(double alpha, double *A, int sda, double *B, int sdb, double beta, double *C…
155 	kernel_dgemm_nn_12x4_vs_lib4(11, &alpha, A, sda, 0, B, sdb, &beta, C, sdd, D, sdd, 11, 4);
156 …kernel_dgemm_nn_12x4_vs_lib4(11, &alpha, A, sda, 0, B+4*4, sdb, &beta, C+4*4, sdd, D+4*4, sdd, 11,…
157 …kernel_dgemm_nn_12x4_vs_lib4(11, &alpha, A, sda, 0, B+8*4, sdb, &beta, C+8*4, sdd, D+8*4, sdd, 11,…
161 void dgemm_nn_12_12_12(double alpha, double *A, int sda, double *B, int sdb, double beta, double *C…
163 	kernel_dgemm_nn_12x4_lib4(12, &alpha, A, sda, 0, B, sdb, &beta, C, sdd, D, sdd);
164 	kernel_dgemm_nn_12x4_lib4(12, &alpha, A, sda, 0, B+4*4, sdb, &beta, C+4*4, sdd, D+4*4, sdd);
165 	kernel_dgemm_nn_12x4_lib4(12, &alpha, A, sda, 0, B+8*4, sdb, &beta, C+8*4, sdd, D+8*4, sdd);
169 void dgemm_nn_13_13_13(double alpha, double *A, int sda, double *B, int sdb, double beta, double *C…
171 	kernel_dgemm_nn_8x4_lib4(13, &alpha, A, sda, 0, B, sdb, &beta, C, sdd, D, sdd);
172 	kernel_dgemm_nn_8x4_lib4(13, &alpha, A, sda, 0, B+4*4, sdb, &beta, C+4*4, sdd, D+4*4, sdd);
173 …kernel_dgemm_nn_8x6_vs_lib4(13, &alpha, A, sda, 0, B+8*4, sdb, &beta, C+8*4, sdd, D+8*4, sdd, 8, 5…
175 …kernel_dgemm_nn_6x8_vs_lib4(13, &alpha, A+8*sda, sda, 0, B, sdb, &beta, C+8*sdc, sdd, D+8*sdd, sdd…
176 …kernel_dgemm_nn_6x6_vs_lib4(13, &alpha, A+8*sda, sda, 0, B+8*4, sdb, &beta, C+8*sdc+8*4, sdd, D+8*…
180 void dgemm_nn_14_14_14(double alpha, double *A, int sda, double *B, int sdb, double beta, double *C…
182 	kernel_dgemm_nn_8x4_lib4(14, &alpha, A, sda, 0, B, sdb, &beta, C, sdd, D, sdd);
183 	kernel_dgemm_nn_8x4_lib4(14, &alpha, A, sda, 0, B+4*4, sdb, &beta, C+4*4, sdd, D+4*4, sdd);
184 	kernel_dgemm_nn_8x6_lib4(14, &alpha, A, sda, 0, B+8*4, sdb, &beta, C+8*4, sdd, D+8*4, sdd);
186 	kernel_dgemm_nn_6x8_lib4(14, &alpha, A+8*sda, sda, 0, B, sdb, &beta, C+8*sdc, sdd, D+8*sdd, sdd);
187 …kernel_dgemm_nn_6x6_lib4(14, &alpha, A+8*sda, sda, 0, B+8*4, sdb, &beta, C+8*sdc+8*4, sdd, D+8*sdd…
191 void dgemm_nn_15_15_15(double alpha, double *A, int sda, double *B, int sdb, double beta, double *C…
194 	kernel_dgemm_nn_8x6_lib4(15, &alpha, A, sda, 0, B, sdb, &beta, C, sdd, D, sdd);
195 	kernel_dgemm_nn_8x6_lib4(15, &alpha, A, sda, 0, B+6*4, sdb, &beta, C+6*4, sdd, D+6*4, sdd);
196 …kernel_dgemm_nn_8x4_vs_lib4(15, &alpha, A, sda, 0, B+12*4, sdb, &beta, C+12*4, sdd, D+12*4, sdd, 8…
198 …kernel_dgemm_nn_8x6_vs_lib4(15, &alpha, A+8*sda, sda, 0, B, sdb, &beta, C+8*sdd, sdd, D+8*sdd, sdd…
199 …kernel_dgemm_nn_8x6_vs_lib4(15, &alpha, A+8*sda, sda, 0, B+6*4, sdb, &beta, C+8*sdd+6*4, sdd, D+8*…
200 …kernel_dgemm_nn_8x4_vs_lib4(15, &alpha, A+8*sda, sda, 0, B+12*4, sdb, &beta, C+8*sdd+12*4, sdd, D+…
202 	kernel_dgemm_nn_8x4_lib4(15, &alpha, A, sda, 0, B, sdb, &beta, C, sdd, D, sdd);
203 	kernel_dgemm_nn_8x4_lib4(15, &alpha, A, sda, 0, B+4*4, sdb, &beta, C+4*4, sdd, D+4*4, sdd);
204 	kernel_dgemm_nn_8x4_lib4(15, &alpha, A, sda, 0, B+8*4, sdb, &beta, C+8*4, sdd, D+8*4, sdd);
205 …kernel_dgemm_nn_8x4_vs_lib4(15, &alpha, A, sda, 0, B+12*4, sdb, &beta, C+12*4, sdd, D+12*4, sdd, 8…
207 …kernel_dgemm_nn_8x4_vs_lib4(15, &alpha, A+8*sda, sda, 0, B, sdb, &beta, C+8*sdd, sdd, D+8*sdd, sdd…
208 …kernel_dgemm_nn_8x4_vs_lib4(15, &alpha, A+8*sda, sda, 0, B+4*4, sdb, &beta, C+8*sdd+4*4, sdd, D+8*…
209 …kernel_dgemm_nn_8x4_vs_lib4(15, &alpha, A+8*sda, sda, 0, B+8*4, sdb, &beta, C+8*sdd+8*4, sdd, D+8*…
210 …kernel_dgemm_nn_8x4_vs_lib4(15, &alpha, A+8*sda, sda, 0, B+12*4, sdb, &beta, C+8*sdd+12*4, sdd, D+…
215 void dgemm_nn_16_16_16(double alpha, double *A, int sda, double *B, int sdb, double beta, double *C…
218 	kernel_dgemm_nn_8x6_lib4(16, &alpha, A, sda, 0, B, sdb, &beta, C, sdd, D, sdd);
219 	kernel_dgemm_nn_8x6_lib4(16, &alpha, A, sda, 0, B+6*4, sdb, &beta, C+6*4, sdd, D+6*4, sdd);
220 	kernel_dgemm_nn_8x4_lib4(16, &alpha, A, sda, 0, B+12*4, sdb, &beta, C+12*4, sdd, D+12*4, sdd);
222 	kernel_dgemm_nn_8x6_lib4(16, &alpha, A+8*sda, sda, 0, B, sdb, &beta, C+8*sdd, sdd, D+8*sdd, sdd);
223 …kernel_dgemm_nn_8x6_lib4(16, &alpha, A+8*sda, sda, 0, B+6*4, sdb, &beta, C+8*sdd+6*4, sdd, D+8*sdd…
224 …kernel_dgemm_nn_8x4_lib4(16, &alpha, A+8*sda, sda, 0, B+12*4, sdb, &beta, C+8*sdd+12*4, sdd, D+8*s…
226 	kernel_dgemm_nn_8x4_lib4(16, &alpha, A, sda, 0, B, sdb, &beta, C, sdd, D, sdd);
227 	kernel_dgemm_nn_8x4_lib4(16, &alpha, A, sda, 0, B+4*4, sdb, &beta, C+4*4, sdd, D+4*4, sdd);
228 	kernel_dgemm_nn_8x4_lib4(16, &alpha, A, sda, 0, B+8*4, sdb, &beta, C+8*4, sdd, D+8*4, sdd);
229 	kernel_dgemm_nn_8x4_lib4(16, &alpha, A, sda, 0, B+12*4, sdb, &beta, C+12*4, sdd, D+12*4, sdd);
231 	kernel_dgemm_nn_8x4_lib4(16, &alpha, A+8*sda, sda, 0, B, sdb, &beta, C+8*sdd, sdd, D+8*sdd, sdd);
232 …kernel_dgemm_nn_8x4_lib4(16, &alpha, A+8*sda, sda, 0, B+4*4, sdb, &beta, C+8*sdd+4*4, sdd, D+8*sdd…
233 …kernel_dgemm_nn_8x4_lib4(16, &alpha, A+8*sda, sda, 0, B+8*4, sdb, &beta, C+8*sdd+8*4, sdd, D+8*sdd…
234 …kernel_dgemm_nn_8x4_lib4(16, &alpha, A+8*sda, sda, 0, B+12*4, sdb, &beta, C+8*sdd+12*4, sdd, D+8*s…
239 void dgemm_nn_17_17_17(double alpha, double *A, int sda, double *B, int sdb, double beta, double *C…
242 	kernel_dgemm_nn_12x4_lib4(17, &alpha, A, sda, 0, B, sdb, &beta, C, sdd, D, sdd);
243 	kernel_dgemm_nn_12x4_lib4(17, &alpha, A, sda, 0, B+4*4, sdb, &beta, C+4*4, sdd, D+4*4, sdd);
244 	kernel_dgemm_nn_12x4_lib4(17, &alpha, A, sda, 0, B+8*4, sdb, &beta, C+8*4, sdd, D+8*4, sdd);
245 	kernel_dgemm_nn_12x4_lib4(17, &alpha, A, sda, 0, B+12*4, sdb, &beta, C+12*4, sdd, D+12*4, sdd);
246 …kernel_dgemm_nn_12x4_vs_lib4(17, &alpha, A, sda, 0, B+16*4, sdb, &beta, C+16*4, sdd, D+16*4, sdd, …
248 …kernel_dgemm_nn_6x8_vs_lib4(17, &alpha, A+12*sda, sda, 0, B, sdb, &beta, C+12*sdd, sdd, D+12*sdd, …
249 …kernel_dgemm_nn_6x8_vs_lib4(17, &alpha, A+12*sda, sda, 0, B+8*4, sdb, &beta, C+12*sdd+8*4, sdd, D+…
250 …kernel_dgemm_nn_6x2_vs_lib4(17, &alpha, A+12*sda, sda, 0, B+16*4, sdb, &beta, C+12*sdd+16*4, sdd, …
252 	kernel_dgemm_nn_8x4_lib4(17, &alpha, A, sda, 0, B, sdb, &beta, C, sdd, D, sdd);
253 	kernel_dgemm_nn_8x4_lib4(17, &alpha, A, sda, 0, B+4*4, sdb, &beta, C+4*4, sdd, D+4*4, sdd);
254 	kernel_dgemm_nn_8x4_lib4(17, &alpha, A, sda, 0, B+8*4, sdb, &beta, C+8*4, sdd, D+8*4, sdd);
255 …kernel_dgemm_nn_8x6_vs_lib4(17, &alpha, A, sda, 0, B+12*4, sdb, &beta, C+12*4, sdd, D+12*4, sdd, 8…
257 …kernel_dgemm_nn_10x4_vs_lib4(17, &alpha, A+8*sda, sda, 0, B, sdb, &beta, C+8*sdd, sdd, D+8*sdd, sd…
258 …kernel_dgemm_nn_10x4_vs_lib4(17, &alpha, A+8*sda, sda, 0, B+4*4, sdb, &beta, C+8*sdd+4*4, sdd, D+8…
259 …kernel_dgemm_nn_10x4_vs_lib4(17, &alpha, A+8*sda, sda, 0, B+8*4, sdb, &beta, C+8*sdd+8*4, sdd, D+8…
260 …kernel_dgemm_nn_10x4_vs_lib4(17, &alpha, A+8*sda, sda, 0, B+12*4, sdb, &beta, C+8*sdd+12*4, sdd, D…
261 …kernel_dgemm_nn_10x2_vs_lib4(17, &alpha, A+8*sda, sda, 0, B+16*4, sdb, &beta, C+8*sdd+16*4, sdd, D…
266 void dgemm_nn_18_18_18(double alpha, double *A, int sda, double *B, int sdb, double beta, double *C…
269 	kernel_dgemm_nn_12x4_lib4(18, &alpha, A, sda, 0, B, sdb, &beta, C, sdd, D, sdd);
270 	kernel_dgemm_nn_12x4_lib4(18, &alpha, A, sda, 0, B+4*4, sdb, &beta, C+4*4, sdd, D+4*4, sdd);
271 	kernel_dgemm_nn_12x4_lib4(18, &alpha, A, sda, 0, B+8*4, sdb, &beta, C+8*4, sdd, D+8*4, sdd);
272 	kernel_dgemm_nn_12x4_lib4(18, &alpha, A, sda, 0, B+12*4, sdb, &beta, C+12*4, sdd, D+12*4, sdd);
273 …kernel_dgemm_nn_12x4_vs_lib4(18, &alpha, A, sda, 0, B+16*4, sdb, &beta, C+16*4, sdd, D+16*4, sdd, …
275 …kernel_dgemm_nn_6x8_lib4(18, &alpha, A+12*sda, sda, 0, B, sdb, &beta, C+12*sdd, sdd, D+12*sdd, sdd…
276 …kernel_dgemm_nn_6x8_lib4(18, &alpha, A+12*sda, sda, 0, B+8*4, sdb, &beta, C+12*sdd+8*4, sdd, D+12*…
277 …kernel_dgemm_nn_6x2_lib4(18, &alpha, A+12*sda, sda, 0, B+16*4, sdb, &beta, C+12*sdd+16*4, sdd, D+1…
279 	kernel_dgemm_nn_8x4_lib4(18, &alpha, A, sda, 0, B, sdb, &beta, C, sdd, D, sdd);
280 	kernel_dgemm_nn_8x4_lib4(18, &alpha, A, sda, 0, B+4*4, sdb, &beta, C+4*4, sdd, D+4*4, sdd);
281 	kernel_dgemm_nn_8x4_lib4(18, &alpha, A, sda, 0, B+8*4, sdb, &beta, C+8*4, sdd, D+8*4, sdd);
282 	kernel_dgemm_nn_8x6_lib4(18, &alpha, A, sda, 0, B+12*4, sdb, &beta, C+12*4, sdd, D+12*4, sdd);
284 	kernel_dgemm_nn_10x4_lib4(18, &alpha, A+8*sda, sda, 0, B, sdb, &beta, C+8*sdd, sdd, D+8*sdd, sdd);
285 …kernel_dgemm_nn_10x4_lib4(18, &alpha, A+8*sda, sda, 0, B+4*4, sdb, &beta, C+8*sdd+4*4, sdd, D+8*sd…
286 …kernel_dgemm_nn_10x4_lib4(18, &alpha, A+8*sda, sda, 0, B+8*4, sdb, &beta, C+8*sdd+8*4, sdd, D+8*sd…
287 …kernel_dgemm_nn_10x4_lib4(18, &alpha, A+8*sda, sda, 0, B+12*4, sdb, &beta, C+8*sdd+12*4, sdd, D+8*…
288 …kernel_dgemm_nn_10x2_lib4(18, &alpha, A+8*sda, sda, 0, B+16*4, sdb, &beta, C+8*sdd+16*4, sdd, D+8*…
293 void dgemm_nn_19_19_19(double alpha, double *A, int sda, double *B, int sdb, double beta, double *C…
295 	kernel_dgemm_nn_8x4_lib4(19, &alpha, A, sda, 0, B, sdb, &beta, C, sdd, D, sdd);
296 	kernel_dgemm_nn_8x4_lib4(19, &alpha, A, sda, 0, B+4*4, sdb, &beta, C+4*4, sdd, D+4*4, sdd);
297 	kernel_dgemm_nn_8x4_lib4(19, &alpha, A, sda, 0, B+8*4, sdb, &beta, C+8*4, sdd, D+8*4, sdd);
298 	kernel_dgemm_nn_8x4_lib4(19, &alpha, A, sda, 0, B+12*4, sdb, &beta, C+12*4, sdd, D+12*4, sdd);
299 …kernel_dgemm_nn_8x4_vs_lib4(19, &alpha, A, sda, 0, B+16*4, sdb, &beta, C+16*4, sdd, D+16*4, sdd, 8…
301 …kernel_dgemm_nn_12x4_vs_lib4(19, &alpha, A+8*sda, sda, 0, B, sdb, &beta, C+8*sdd, sdd, D+8*sdd, sd…
302 …kernel_dgemm_nn_12x4_vs_lib4(19, &alpha, A+8*sda, sda, 0, B+4*4, sdb, &beta, C+8*sdd+4*4, sdd, D+8…
303 …kernel_dgemm_nn_12x4_vs_lib4(19, &alpha, A+8*sda, sda, 0, B+8*4, sdb, &beta, C+8*sdd+8*4, sdd, D+8…
304 …kernel_dgemm_nn_12x4_vs_lib4(19, &alpha, A+8*sda, sda, 0, B+12*4, sdb, &beta, C+8*sdd+12*4, sdd, D…
305 …kernel_dgemm_nn_12x4_vs_lib4(19, &alpha, A+8*sda, sda, 0, B+16*4, sdb, &beta, C+8*sdd+16*4, sdd, D…
309 void dgemm_nn_20_20_20(double alpha, double *A, int sda, double *B, int sdb, double beta, double *C…
311 	kernel_dgemm_nn_8x4_lib4(20, &alpha, A, sda, 0, B, sdb, &beta, C, sdd, D, sdd);
312 	kernel_dgemm_nn_8x4_lib4(20, &alpha, A, sda, 0, B+4*4, sdb, &beta, C+4*4, sdd, D+4*4, sdd);
313 	kernel_dgemm_nn_8x4_lib4(20, &alpha, A, sda, 0, B+8*4, sdb, &beta, C+8*4, sdd, D+8*4, sdd);
314 	kernel_dgemm_nn_8x4_lib4(20, &alpha, A, sda, 0, B+12*4, sdb, &beta, C+12*4, sdd, D+12*4, sdd);
315 	kernel_dgemm_nn_8x4_lib4(20, &alpha, A, sda, 0, B+16*4, sdb, &beta, C+16*4, sdd, D+16*4, sdd);
317 	kernel_dgemm_nn_12x4_lib4(20, &alpha, A+8*sda, sda, 0, B, sdb, &beta, C+8*sdd, sdd, D+8*sdd, sdd);
318 …kernel_dgemm_nn_12x4_lib4(20, &alpha, A+8*sda, sda, 0, B+4*4, sdb, &beta, C+8*sdd+4*4, sdd, D+8*sd…
319 …kernel_dgemm_nn_12x4_lib4(20, &alpha, A+8*sda, sda, 0, B+8*4, sdb, &beta, C+8*sdd+8*4, sdd, D+8*sd…
320 …kernel_dgemm_nn_12x4_lib4(20, &alpha, A+8*sda, sda, 0, B+12*4, sdb, &beta, C+8*sdd+12*4, sdd, D+8*…
321 …kernel_dgemm_nn_12x4_lib4(20, &alpha, A+8*sda, sda, 0, B+16*4, sdb, &beta, C+8*sdd+16*4, sdd, D+8*…
325 void dgemm_nn_21_21_21(double alpha, double *A, int sda, double *B, int sdb, double beta, double *C…
328 	kernel_dgemm_nn_12x4_lib4(21, &alpha, A, sda, 0, B, sdb, &beta, C, sdd, D, sdd);
329 	kernel_dgemm_nn_12x4_lib4(21, &alpha, A, sda, 0, B+4*4, sdb, &beta, C+4*4, sdd, D+4*4, sdd);
330 	kernel_dgemm_nn_12x4_lib4(21, &alpha, A, sda, 0, B+8*4, sdb, &beta, C+8*4, sdd, D+8*4, sdd);
331 	kernel_dgemm_nn_12x4_lib4(21, &alpha, A, sda, 0, B+12*4, sdb, &beta, C+12*4, sdd, D+12*4, sdd);
332 	kernel_dgemm_nn_12x4_lib4(21, &alpha, A, sda, 0, B+16*4, sdb, &beta, C+16*4, sdd, D+16*4, sdd);
333 …kernel_dgemm_nn_12x4_vs_lib4(21, &alpha, A, sda, 0, B+20*4, sdb, &beta, C+20*4, sdd, D+20*4, sdd, …
335 …kernel_dgemm_nn_10x4_vs_lib4(21, &alpha, A+12*sda, sda, 0, B, sdb, &beta, C+12*sdd, sdd, D+12*sdd,…
336 …kernel_dgemm_nn_10x4_vs_lib4(21, &alpha, A+12*sda, sda, 0, B+4*4, sdb, &beta, C+12*sdd+4*4, sdd, D…
337 …kernel_dgemm_nn_10x4_vs_lib4(21, &alpha, A+12*sda, sda, 0, B+8*4, sdb, &beta, C+12*sdd+8*4, sdd, D…
338 …kernel_dgemm_nn_10x4_vs_lib4(21, &alpha, A+12*sda, sda, 0, B+12*4, sdb, &beta, C+12*sdd+12*4, sdd,…
339 …kernel_dgemm_nn_10x4_vs_lib4(21, &alpha, A+12*sda, sda, 0, B+16*4, sdb, &beta, C+12*sdd+16*4, sdd,…
340 …kernel_dgemm_nn_10x4_vs_lib4(21, &alpha, A+12*sda, sda, 0, B+20*4, sdb, &beta, C+12*sdd+20*4, sdd,…
342 	kernel_dgemm_nn_8x4_lib4(21, &alpha, A, sda, 0, B, sdb, &beta, C, sdd, D, sdd);
343 	kernel_dgemm_nn_8x4_lib4(21, &alpha, A, sda, 0, B+4*4, sdb, &beta, C+4*4, sdd, D+4*4, sdd);
344 	kernel_dgemm_nn_8x4_lib4(21, &alpha, A, sda, 0, B+8*4, sdb, &beta, C+8*4, sdd, D+8*4, sdd);
345 	kernel_dgemm_nn_8x4_lib4(21, &alpha, A, sda, 0, B+12*4, sdb, &beta, C+12*4, sdd, D+12*4, sdd);
346 …kernel_dgemm_nn_8x6_vs_lib4(21, &alpha, A, sda, 0, B+16*4, sdb, &beta, C+16*4, sdd, D+16*4, sdd, 8…
348 	kernel_dgemm_nn_8x4_lib4(21, &alpha, A+8*sda, sda, 0, B, sdb, &beta, C+8*sdd, sdd, D+8*sdd, sdd);
349 …kernel_dgemm_nn_8x4_lib4(21, &alpha, A+8*sda, sda, 0, B+4*4, sdb, &beta, C+8*sdd+4*4, sdd, D+8*sdd…
350 …kernel_dgemm_nn_8x4_lib4(21, &alpha, A+8*sda, sda, 0, B+8*4, sdb, &beta, C+8*sdd+8*4, sdd, D+8*sdd…
351 …kernel_dgemm_nn_8x4_lib4(21, &alpha, A+8*sda, sda, 0, B+12*4, sdb, &beta, C+8*sdd+12*4, sdd, D+8*s…
352 …kernel_dgemm_nn_8x6_vs_lib4(21, &alpha, A+8*sda, sda, 0, B+16*4, sdb, &beta, C+8*sdd+16*4, sdd, D+…
354 …kernel_dgemm_nn_6x8_vs_lib4(21, &alpha, A+16*sda, sda, 0, B, sdb, &beta, C+16*sdd, sdd, D+16*sdd, …
355 …kernel_dgemm_nn_6x8_vs_lib4(21, &alpha, A+16*sda, sda, 0, B+8*4, sdb, &beta, C+16*sdd+8*4, sdd, D+…
356 …kernel_dgemm_nn_6x6_vs_lib4(21, &alpha, A+16*sda, sda, 0, B+16*4, sdb, &beta, C+16*sdd+16*4, sdd, …
361 void dgemm_nn_22_22_22(double alpha, double *A, int sda, double *B, int sdb, double beta, double *C…
364 	kernel_dgemm_nn_12x4_lib4(22, &alpha, A, sda, 0, B, sdb, &beta, C, sdd, D, sdd);
365 	kernel_dgemm_nn_12x4_lib4(22, &alpha, A, sda, 0, B+4*4, sdb, &beta, C+4*4, sdd, D+4*4, sdd);
366 	kernel_dgemm_nn_12x4_lib4(22, &alpha, A, sda, 0, B+8*4, sdb, &beta, C+8*4, sdd, D+8*4, sdd);
367 	kernel_dgemm_nn_12x4_lib4(22, &alpha, A, sda, 0, B+12*4, sdb, &beta, C+12*4, sdd, D+12*4, sdd);
368 	kernel_dgemm_nn_12x4_lib4(22, &alpha, A, sda, 0, B+16*4, sdb, &beta, C+16*4, sdd, D+16*4, sdd);
369 …kernel_dgemm_nn_12x4_vs_lib4(22, &alpha, A, sda, 0, B+20*4, sdb, &beta, C+20*4, sdd, D+20*4, sdd, …
371 …kernel_dgemm_nn_10x4_lib4(22, &alpha, A+12*sda, sda, 0, B, sdb, &beta, C+12*sdd, sdd, D+12*sdd, sd…
372 …kernel_dgemm_nn_10x4_lib4(22, &alpha, A+12*sda, sda, 0, B+4*4, sdb, &beta, C+12*sdd+4*4, sdd, D+12…
373 …kernel_dgemm_nn_10x4_lib4(22, &alpha, A+12*sda, sda, 0, B+8*4, sdb, &beta, C+12*sdd+8*4, sdd, D+12…
374 …kernel_dgemm_nn_10x4_lib4(22, &alpha, A+12*sda, sda, 0, B+12*4, sdb, &beta, C+12*sdd+12*4, sdd, D+…
375 …kernel_dgemm_nn_10x4_lib4(22, &alpha, A+12*sda, sda, 0, B+16*4, sdb, &beta, C+12*sdd+16*4, sdd, D+…
376 …kernel_dgemm_nn_10x4_vs_lib4(22, &alpha, A+12*sda, sda, 0, B+20*4, sdb, &beta, C+12*sdd+20*4, sdd,…
378 	kernel_dgemm_nn_8x4_lib4(22, &alpha, A, sda, 0, B, sdb, &beta, C, sdd, D, sdd);
379 	kernel_dgemm_nn_8x4_lib4(22, &alpha, A, sda, 0, B+4*4, sdb, &beta, C+4*4, sdd, D+4*4, sdd);
380 	kernel_dgemm_nn_8x4_lib4(22, &alpha, A, sda, 0, B+8*4, sdb, &beta, C+8*4, sdd, D+8*4, sdd);
381 	kernel_dgemm_nn_8x4_lib4(22, &alpha, A, sda, 0, B+12*4, sdb, &beta, C+12*4, sdd, D+12*4, sdd);
382 	kernel_dgemm_nn_8x6_lib4(22, &alpha, A, sda, 0, B+16*4, sdb, &beta, C+16*4, sdd, D+16*4, sdd);
384 	kernel_dgemm_nn_8x4_lib4(22, &alpha, A+8*sda, sda, 0, B, sdb, &beta, C+8*sdd, sdd, D+8*sdd, sdd);
385 …kernel_dgemm_nn_8x4_lib4(22, &alpha, A+8*sda, sda, 0, B+4*4, sdb, &beta, C+8*sdd+4*4, sdd, D+8*sdd…
386 …kernel_dgemm_nn_8x4_lib4(22, &alpha, A+8*sda, sda, 0, B+8*4, sdb, &beta, C+8*sdd+8*4, sdd, D+8*sdd…
387 …kernel_dgemm_nn_8x4_lib4(22, &alpha, A+8*sda, sda, 0, B+12*4, sdb, &beta, C+8*sdd+12*4, sdd, D+8*s…
388 …kernel_dgemm_nn_8x6_lib4(22, &alpha, A+8*sda, sda, 0, B+16*4, sdb, &beta, C+8*sdd+16*4, sdd, D+8*s…
390 …kernel_dgemm_nn_6x8_lib4(22, &alpha, A+16*sda, sda, 0, B, sdb, &beta, C+16*sdd, sdd, D+16*sdd, sdd…
391 …kernel_dgemm_nn_6x8_lib4(22, &alpha, A+16*sda, sda, 0, B+8*4, sdb, &beta, C+16*sdd+8*4, sdd, D+16*…
392 …kernel_dgemm_nn_6x6_lib4(22, &alpha, A+16*sda, sda, 0, B+16*4, sdb, &beta, C+16*sdd+16*4, sdd, D+1…
397 void dgemm_nn_23_23_23(double alpha, double *A, int sda, double *B, int sdb, double beta, double *C…
400 	kernel_dgemm_nn_12x4_lib4(23, &alpha, A, sda, 0, B, sdb, &beta, C, sdd, D, sdd);
401 	kernel_dgemm_nn_12x4_lib4(23, &alpha, A, sda, 0, B+4*4, sdb, &beta, C+4*4, sdd, D+4*4, sdd);
402 	kernel_dgemm_nn_12x4_lib4(23, &alpha, A, sda, 0, B+8*4, sdb, &beta, C+8*4, sdd, D+8*4, sdd);
403 	kernel_dgemm_nn_12x4_lib4(23, &alpha, A, sda, 0, B+12*4, sdb, &beta, C+12*4, sdd, D+12*4, sdd);
404 	kernel_dgemm_nn_12x4_lib4(23, &alpha, A, sda, 0, B+16*4, sdb, &beta, C+16*4, sdd, D+16*4, sdd);
405 …kernel_dgemm_nn_12x4_vs_lib4(23, &alpha, A, sda, 0, B+20*4, sdb, &beta, C+20*4, sdd, D+20*4, sdd, …
407 …kernel_dgemm_nn_12x4_vs_lib4(23, &alpha, A+12*sda, sda, 0, B, sdb, &beta, C+12*sdd, sdd, D+12*sdd,…
408 …kernel_dgemm_nn_12x4_vs_lib4(23, &alpha, A+12*sda, sda, 0, B+4*4, sdb, &beta, C+12*sdd+4*4, sdd, D…
409 …kernel_dgemm_nn_12x4_vs_lib4(23, &alpha, A+12*sda, sda, 0, B+8*4, sdb, &beta, C+12*sdd+8*4, sdd, D…
410 …kernel_dgemm_nn_12x4_vs_lib4(23, &alpha, A+12*sda, sda, 0, B+12*4, sdb, &beta, C+12*sdd+12*4, sdd,…
411 …kernel_dgemm_nn_12x4_vs_lib4(23, &alpha, A+12*sda, sda, 0, B+16*4, sdb, &beta, C+12*sdd+16*4, sdd,…
412 …kernel_dgemm_nn_12x4_vs_lib4(23, &alpha, A+12*sda, sda, 0, B+20*4, sdb, &beta, C+12*sdd+20*4, sdd,…
414 	kernel_dgemm_nn_8x4_lib4(23, &alpha, A, sda, 0, B, sdb, &beta, C, sdd, D, sdd);
415 	kernel_dgemm_nn_8x4_lib4(23, &alpha, A, sda, 0, B+4*4, sdb, &beta, C+4*4, sdd, D+4*4, sdd);
416 	kernel_dgemm_nn_8x4_lib4(23, &alpha, A, sda, 0, B+8*4, sdb, &beta, C+8*4, sdd, D+8*4, sdd);
417 	kernel_dgemm_nn_8x4_lib4(23, &alpha, A, sda, 0, B+12*4, sdb, &beta, C+12*4, sdd, D+12*4, sdd);
418 	kernel_dgemm_nn_8x4_lib4(23, &alpha, A, sda, 0, B+16*4, sdb, &beta, C+16*4, sdd, D+16*4, sdd);
419 …kernel_dgemm_nn_8x4_vs_lib4(23, &alpha, A, sda, 0, B+20*4, sdb, &beta, C+20*4, sdd, D+20*4, sdd, 8…
421 	kernel_dgemm_nn_8x4_lib4(23, &alpha, A+8*sda, sda, 0, B, sdb, &beta, C+8*sdd, sdd, D+8*sdd, sdd);
422 …kernel_dgemm_nn_8x4_lib4(23, &alpha, A+8*sda, sda, 0, B+4*4, sdb, &beta, C+8*sdd+4*4, sdd, D+8*sdd…
423 …kernel_dgemm_nn_8x4_lib4(23, &alpha, A+8*sda, sda, 0, B+8*4, sdb, &beta, C+8*sdd+8*4, sdd, D+8*sdd…
424 …kernel_dgemm_nn_8x4_lib4(23, &alpha, A+8*sda, sda, 0, B+12*4, sdb, &beta, C+8*sdd+12*4, sdd, D+8*s…
425 …kernel_dgemm_nn_8x4_lib4(23, &alpha, A+8*sda, sda, 0, B+16*4, sdb, &beta, C+8*sdd+16*4, sdd, D+8*s…
426 …kernel_dgemm_nn_8x4_vs_lib4(23, &alpha, A+8*sda, sda, 0, B+20*4, sdb, &beta, C+8*sdd+20*4, sdd, D+…
428 …kernel_dgemm_nn_8x4_vs_lib4(23, &alpha, A+16*sda, sda, 0, B, sdb, &beta, C+16*sdd, sdd, D+16*sdd, …
429 …kernel_dgemm_nn_8x4_vs_lib4(23, &alpha, A+16*sda, sda, 0, B+4*4, sdb, &beta, C+16*sdd+4*4, sdd, D+…
430 …kernel_dgemm_nn_8x4_vs_lib4(23, &alpha, A+16*sda, sda, 0, B+8*4, sdb, &beta, C+16*sdd+8*4, sdd, D+…
431 …kernel_dgemm_nn_8x4_vs_lib4(23, &alpha, A+16*sda, sda, 0, B+12*4, sdb, &beta, C+16*sdd+12*4, sdd, …
432 …kernel_dgemm_nn_8x4_vs_lib4(23, &alpha, A+16*sda, sda, 0, B+16*4, sdb, &beta, C+16*sdd+16*4, sdd, …
433 …kernel_dgemm_nn_8x4_vs_lib4(23, &alpha, A+16*sda, sda, 0, B+20*4, sdb, &beta, C+16*sdd+20*4, sdd, …
438 void dgemm_nn_24_24_24(double alpha, double *A, int sda, double *B, int sdb, double beta, double *C…
441 	kernel_dgemm_nn_12x4_lib4(24, &alpha, A, sda, 0, B, sdb, &beta, C, sdd, D, sdd);
442 	kernel_dgemm_nn_12x4_lib4(24, &alpha, A, sda, 0, B+4*4, sdb, &beta, C+4*4, sdd, D+4*4, sdd);
443 	kernel_dgemm_nn_12x4_lib4(24, &alpha, A, sda, 0, B+8*4, sdb, &beta, C+8*4, sdd, D+8*4, sdd);
444 	kernel_dgemm_nn_12x4_lib4(24, &alpha, A, sda, 0, B+12*4, sdb, &beta, C+12*4, sdd, D+12*4, sdd);
445 	kernel_dgemm_nn_12x4_lib4(24, &alpha, A, sda, 0, B+16*4, sdb, &beta, C+16*4, sdd, D+16*4, sdd);
446 	kernel_dgemm_nn_12x4_lib4(24, &alpha, A, sda, 0, B+20*4, sdb, &beta, C+20*4, sdd, D+20*4, sdd);
448 …kernel_dgemm_nn_12x4_lib4(24, &alpha, A+12*sda, sda, 0, B, sdb, &beta, C+12*sdd, sdd, D+12*sdd, sd…
449 …kernel_dgemm_nn_12x4_lib4(24, &alpha, A+12*sda, sda, 0, B+4*4, sdb, &beta, C+12*sdd+4*4, sdd, D+12…
450 …kernel_dgemm_nn_12x4_lib4(24, &alpha, A+12*sda, sda, 0, B+8*4, sdb, &beta, C+12*sdd+8*4, sdd, D+12…
451 …kernel_dgemm_nn_12x4_lib4(24, &alpha, A+12*sda, sda, 0, B+12*4, sdb, &beta, C+12*sdd+12*4, sdd, D+…
452 …kernel_dgemm_nn_12x4_lib4(24, &alpha, A+12*sda, sda, 0, B+16*4, sdb, &beta, C+12*sdd+16*4, sdd, D+…
453 …kernel_dgemm_nn_12x4_lib4(24, &alpha, A+12*sda, sda, 0, B+20*4, sdb, &beta, C+12*sdd+20*4, sdd, D+…
455 	kernel_dgemm_nn_8x4_lib4(24, &alpha, A, sda, 0, B, sdb, &beta, C, sdd, D, sdd);
456 	kernel_dgemm_nn_8x4_lib4(24, &alpha, A, sda, 0, B+4*4, sdb, &beta, C+4*4, sdd, D+4*4, sdd);
457 	kernel_dgemm_nn_8x4_lib4(24, &alpha, A, sda, 0, B+8*4, sdb, &beta, C+8*4, sdd, D+8*4, sdd);
458 	kernel_dgemm_nn_8x4_lib4(24, &alpha, A, sda, 0, B+12*4, sdb, &beta, C+12*4, sdd, D+12*4, sdd);
459 	kernel_dgemm_nn_8x4_lib4(24, &alpha, A, sda, 0, B+16*4, sdb, &beta, C+16*4, sdd, D+16*4, sdd);
460 	kernel_dgemm_nn_8x4_lib4(24, &alpha, A, sda, 0, B+20*4, sdb, &beta, C+20*4, sdd, D+20*4, sdd);
462 	kernel_dgemm_nn_8x4_lib4(24, &alpha, A+8*sda, sda, 0, B, sdb, &beta, C+8*sdd, sdd, D+8*sdd, sdd);
463 …kernel_dgemm_nn_8x4_lib4(24, &alpha, A+8*sda, sda, 0, B+4*4, sdb, &beta, C+8*sdd+4*4, sdd, D+8*sdd…
464 …kernel_dgemm_nn_8x4_lib4(24, &alpha, A+8*sda, sda, 0, B+8*4, sdb, &beta, C+8*sdd+8*4, sdd, D+8*sdd…
465 …kernel_dgemm_nn_8x4_lib4(24, &alpha, A+8*sda, sda, 0, B+12*4, sdb, &beta, C+8*sdd+12*4, sdd, D+8*s…
466 …kernel_dgemm_nn_8x4_lib4(24, &alpha, A+8*sda, sda, 0, B+16*4, sdb, &beta, C+8*sdd+16*4, sdd, D+8*s…
467 …kernel_dgemm_nn_8x4_lib4(24, &alpha, A+8*sda, sda, 0, B+20*4, sdb, &beta, C+8*sdd+20*4, sdd, D+8*s…
469 …kernel_dgemm_nn_8x4_lib4(24, &alpha, A+16*sda, sda, 0, B, sdb, &beta, C+16*sdd, sdd, D+16*sdd, sdd…
470 …kernel_dgemm_nn_8x4_lib4(24, &alpha, A+16*sda, sda, 0, B+4*4, sdb, &beta, C+16*sdd+4*4, sdd, D+16*…
471 …kernel_dgemm_nn_8x4_lib4(24, &alpha, A+16*sda, sda, 0, B+8*4, sdb, &beta, C+16*sdd+8*4, sdd, D+16*…
472 …kernel_dgemm_nn_8x4_lib4(24, &alpha, A+16*sda, sda, 0, B+12*4, sdb, &beta, C+16*sdd+12*4, sdd, D+1…
473 …kernel_dgemm_nn_8x4_lib4(24, &alpha, A+16*sda, sda, 0, B+16*4, sdb, &beta, C+16*sdd+16*4, sdd, D+1…
474 …kernel_dgemm_nn_8x4_lib4(24, &alpha, A+16*sda, sda, 0, B+20*4, sdb, &beta, C+16*sdd+20*4, sdd, D+1…