28 #include "../../asmlib.hpp"
32 void a64_sgemm_asimd_8x12_a55r1(
const float *Apanel,
const float *Bpanel,
float *Cpanel,
const int ablocks,
const int bblocks,
const int K) {
33 const float *a_ptr = Apanel;
34 float *c_ptr = Cpanel;
39 int k_iters = ((
K+1)/2) - 1;
41 for (
int yb=0; yb<ablocks; yb++) {
42 const float *a_ptr0 = a_ptr;
43 const float *b_ptr = Bpanel;
45 for (
int xb=0; xb<bblocks; xb++) {
49 register float32x4_t a0
asm(
"v0");
50 register float32x4_t a1
asm(
"v1");
51 register float32x4_t b0
asm(
"v2");
52 register float32x4_t b1
asm(
"v3");
53 register float32x4_t b2
asm(
"v4");
54 register float32x4_t a0a
asm(
"v5");
55 register float32x4_t a1a
asm(
"v6");
60 "ldr %q[a0], [%[a_ptr]]\n"
62 "ldr %q[b0], [%[b_ptr]]\n"
64 "ldr %q[a1], [%[a_ptr], #16]\n"
66 "ldr %q[b1], [%[b_ptr], #16]\n"
103 "fmla v8.4s , %[b0].4s, %[a0].s[0]\n"
104 "ldr %d[b2], [%[b_ptr], #32]\n"
111 "fmla v9.4s , %[b0].4s, %[a0].s[1]\n"
112 "ldr x20, [%[b_ptr], #40]\n"
113 "fmla v10.4s, %[b0].4s, %[a0].s[2]\n"
114 "subs %w[k], %w[k], #1\n"
115 "fmla v11.4s, %[b0].4s, %[a0].s[3]\n"
116 "ldr %d[a0a], [%[a_ptr], #32]\n"
118 "fmla v12.4s, %[b0].4s, %[a1].s[0]\n"
119 "ins %[b2].d[1], x20\n"
120 "fmla v13.4s, %[b0].4s, %[a1].s[1]\n"
121 "ldr x20, [%[a_ptr], #40]\n"
122 "fmla v14.4s, %[b0].4s, %[a1].s[2]\n"
123 "fmla v15.4s, %[b0].4s, %[a1].s[3]\n"
124 "ldr %d[a1a], [%[a_ptr], #48]\n"
126 "fmla v16.4s, %[b1].4s, %[a0].s[0]\n"
127 "ins %[a0a].d[1], x20\n"
128 "fmla v17.4s, %[b1].4s, %[a0].s[1]\n"
129 "ldr x20, [%[a_ptr], #56]\n"
130 "fmla v18.4s, %[b1].4s, %[a0].s[2]\n"
131 "fmla v19.4s, %[b1].4s, %[a0].s[3]\n"
132 "ldr %d[b0], [%[b_ptr], #48]\n"
134 "fmla v20.4s, %[b1].4s, %[a1].s[0]\n"
135 "ins %[a1a].d[1], x20\n"
136 "fmla v21.4s, %[b1].4s, %[a1].s[1]\n"
137 "ldr x20, [%[b_ptr], #56]\n"
138 "fmla v22.4s, %[b1].4s, %[a1].s[2]\n"
139 "fmla v23.4s, %[b1].4s, %[a1].s[3]\n"
140 "ldr %d[b1], [%[b_ptr], #64]\n"
142 "fmla v24.4s, %[b2].4s, %[a0].s[0]\n"
143 "ins %[b0].d[1], x20\n"
144 "fmla v25.4s, %[b2].4s, %[a0].s[1]\n"
145 "ldr x20, [%[b_ptr], #72]\n"
146 "fmla v26.4s, %[b2].4s, %[a0].s[2]\n"
147 "fmla v27.4s, %[b2].4s, %[a0].s[3]\n"
150 "fmla v28.4s, %[b2].4s, %[a1].s[0]\n"
151 "fmla v29.4s, %[b2].4s, %[a1].s[1]\n"
153 "fmla v30.4s, %[b2].4s, %[a1].s[2]\n"
154 "fmla v31.4s, %[b2].4s, %[a1].s[3]\n"
157 "ldr %d[b2], [%[b_ptr], #80]\n"
159 "fmla v8.4s , %[b0].4s, %[a0a].s[0]\n"
160 "ins %[b1].d[1], x20\n"
161 "fmla v9.4s , %[b0].4s, %[a0a].s[1]\n"
162 "ldr x20, [%[b_ptr], #88]\n"
163 "fmla v10.4s, %[b0].4s, %[a0a].s[2]\n"
164 "fmla v11.4s, %[b0].4s, %[a0a].s[3]\n"
165 "ldr %d[a0], [%[a_ptr], #64]\n"
167 "fmla v12.4s, %[b0].4s, %[a1a].s[0]\n"
168 "ins %[b2].d[1], x20\n"
169 "fmla v13.4s, %[b0].4s, %[a1a].s[1]\n"
170 "ldr x20, [%[a_ptr], #72]\n"
171 "fmla v14.4s, %[b0].4s, %[a1a].s[2]\n"
172 "fmla v15.4s, %[b0].4s, %[a1a].s[3]\n"
173 "ldr %d[a1], [%[a_ptr], #80]\n"
175 "fmla v16.4s, %[b1].4s, %[a0a].s[0]\n"
176 "ins %[a0].d[1], x20\n"
177 "fmla v17.4s, %[b1].4s, %[a0a].s[1]\n"
178 "ldr x20, [%[a_ptr], #88]\n"
179 "fmla v18.4s, %[b1].4s, %[a0a].s[2]\n"
180 "fmla v19.4s, %[b1].4s, %[a0a].s[3]\n"
181 "ldr %d[b0], [%[b_ptr], #96]\n"
183 "fmla v20.4s, %[b1].4s, %[a1a].s[0]\n"
184 "ins %[a1].d[1], x20\n"
185 "fmla v21.4s, %[b1].4s, %[a1a].s[1]\n"
186 "ldr x20, [%[b_ptr], #104]\n"
187 "fmla v22.4s, %[b1].4s, %[a1a].s[2]\n"
188 "fmla v23.4s, %[b1].4s, %[a1a].s[3]\n"
189 "ldr %d[b1], [%[b_ptr], #112]\n"
191 "fmla v24.4s, %[b2].4s, %[a0a].s[0]\n"
192 "ins %[b0].d[1], x20\n"
193 "fmla v25.4s, %[b2].4s, %[a0a].s[1]\n"
194 "ldr x20, [%[b_ptr], #120]\n"
195 "fmla v26.4s, %[b2].4s, %[a0a].s[2]\n"
197 "fmla v27.4s, %[b2].4s, %[a0a].s[3]\n"
198 "add %[a_ptr], %[a_ptr], #64\n"
200 "fmla v28.4s, %[b2].4s, %[a1a].s[0]\n"
202 "fmla v29.4s, %[b2].4s, %[a1a].s[1]\n"
203 "add %[b_ptr], %[b_ptr], #96\n"
204 "fmla v30.4s, %[b2].4s, %[a1a].s[2]\n"
205 "ins %[b1].d[1], x20\n"
206 "fmla v31.4s, %[b2].4s, %[a1a].s[3]\n"
207 "ldr %d[b2], [%[b_ptr], #32]\n"
209 "fmla v8.4s , %[b0].4s, %[a0].s[0]\n"
216 "fmla v9.4s , %[b0].4s, %[a0].s[1]\n"
217 "ldr x20, [%[b_ptr], #40]\n"
218 "fmla v10.4s, %[b0].4s, %[a0].s[2]\n"
219 "cbnz %w[oddk], 2f\n"
222 "fmla v11.4s, %[b0].4s, %[a0].s[3]\n"
223 "ldr %d[a0a], [%[a_ptr], #32]\n"
225 "fmla v12.4s, %[b0].4s, %[a1].s[0]\n"
226 "ins %[b2].d[1], x20\n"
227 "fmla v13.4s, %[b0].4s, %[a1].s[1]\n"
228 "ldr x20, [%[a_ptr], #40]\n"
229 "fmla v14.4s, %[b0].4s, %[a1].s[2]\n"
231 "fmla v15.4s, %[b0].4s, %[a1].s[3]\n"
232 "ldr %d[a1a], [%[a_ptr], #48]\n"
234 "fmla v16.4s, %[b1].4s, %[a0].s[0]\n"
235 "ins %[a0a].d[1], x20\n"
236 "fmla v17.4s, %[b1].4s, %[a0].s[1]\n"
237 "ldr x20, [%[a_ptr], #56]\n"
238 "fmla v18.4s, %[b1].4s, %[a0].s[2]\n"
239 "fmla v19.4s, %[b1].4s, %[a0].s[3]\n"
240 "ldr %d[b0], [%[b_ptr], #48]\n"
242 "fmla v20.4s, %[b1].4s, %[a1].s[0]\n"
243 "ins %[a1a].d[1], x20\n"
244 "fmla v21.4s, %[b1].4s, %[a1].s[1]\n"
245 "ldr x20, [%[b_ptr], #56]\n"
246 "fmla v22.4s, %[b1].4s, %[a1].s[2]\n"
248 "fmla v23.4s, %[b1].4s, %[a1].s[3]\n"
250 "fmla v24.4s, %[b2].4s, %[a0].s[0]\n"
251 "fmla v25.4s, %[b2].4s, %[a0].s[1]\n"
253 "fmla v26.4s, %[b2].4s, %[a0].s[2]\n"
254 "fmla v27.4s, %[b2].4s, %[a0].s[3]\n"
255 "ldr %d[b1], [%[b_ptr], #64]\n"
257 "fmla v28.4s, %[b2].4s, %[a1].s[0]\n"
258 "ins %[b0].d[1], x20\n"
259 "fmla v29.4s, %[b2].4s, %[a1].s[1]\n"
260 "ldr x20, [%[b_ptr], #72]\n"
261 "fmla v30.4s, %[b2].4s, %[a1].s[2]\n"
263 "fmla v31.4s, %[b2].4s, %[a1].s[3]\n"
264 "ldr %d[b2], [%[b_ptr], #80]\n"
266 "fmla v8.4s , %[b0].4s, %[a0a].s[0]\n"
267 "ins %[b1].d[1], x20\n"
268 "fmla v9.4s , %[b0].4s, %[a0a].s[1]\n"
269 "ldr x20, [%[b_ptr], #88]\n"
270 "fmla v10.4s, %[b0].4s, %[a0a].s[2]\n"
271 "ins %[b2].d[1], x20\n"
273 "fmla v11.4s, %[b0].4s, %[a0a].s[3]\n"
275 "fmla v12.4s, %[b0].4s, %[a1a].s[0]\n"
276 "fmla v13.4s, %[b0].4s, %[a1a].s[1]\n"
277 "fmla v14.4s, %[b0].4s, %[a1a].s[2]\n"
279 "fmla v15.4s, %[b0].4s, %[a1a].s[3]\n"
280 "fmla v16.4s, %[b1].4s, %[a0a].s[0]\n"
281 ASM_PREFETCHWL2(
"[%[c_ptr], #384]")
282 "fmla v17.4s, %[b1].4s, %[a0a].s[1]\n"
283 "fmla v18.4s, %[b1].4s, %[a0a].s[2]\n"
284 ASM_PREFETCHWL2(
"[%[c_ptr], #448]")
285 "fmla v19.4s, %[b1].4s, %[a0a].s[3]\n"
286 "fmla v20.4s, %[b1].4s, %[a1a].s[0]\n"
287 "fmla v21.4s, %[b1].4s, %[a1a].s[1]\n"
288 ASM_PREFETCHWL2(
"[%[c_ptr], #512]")
289 "fmla v22.4s, %[b1].4s, %[a1a].s[2]\n"
290 "fmla v23.4s, %[b1].4s, %[a1a].s[3]\n"
291 ASM_PREFETCHWL2(
"[%[c_ptr], #576]")
292 "fmla v24.4s, %[b2].4s, %[a0a].s[0]\n"
293 "fmla v25.4s, %[b2].4s, %[a0a].s[1]\n"
294 "fmla v26.4s, %[b2].4s, %[a0a].s[2]\n"
295 ASM_PREFETCHWL2(
"[%[c_ptr], #640]")
296 "fmla v27.4s, %[b2].4s, %[a0a].s[3]\n"
297 "fmla v28.4s, %[b2].4s, %[a1a].s[0]\n"
298 ASM_PREFETCHWL2(
"[%[c_ptr], #704]")
299 "fmla v29.4s, %[b2].4s, %[a1a].s[1]\n"
300 "add %[a_ptr], %[a_ptr], #64\n"
301 "fmla v30.4s, %[b2].4s, %[a1a].s[2]\n"
302 "add %[b_ptr], %[b_ptr], #96\n"
303 "fmla v31.4s, %[b2].4s, %[a1a].s[3]\n"
308 "fmla v11.4s, %[b0].4s, %[a0].s[3]\n"
310 "fmla v12.4s, %[b0].4s, %[a1].s[0]\n"
311 "ins %[b2].d[1], x20\n"
312 "fmla v13.4s, %[b0].4s, %[a1].s[1]\n"
314 "fmla v14.4s, %[b0].4s, %[a1].s[2]\n"
315 "add %[a_ptr], %[a_ptr], #32\n"
316 "fmla v15.4s, %[b0].4s, %[a1].s[3]\n"
318 "fmla v16.4s, %[b1].4s, %[a0].s[0]\n"
319 "add %[b_ptr], %[b_ptr], #48\n"
320 "fmla v17.4s, %[b1].4s, %[a0].s[1]\n"
322 "fmla v18.4s, %[b1].4s, %[a0].s[2]\n"
323 "fmla v19.4s, %[b1].4s, %[a0].s[3]\n"
325 "fmla v20.4s, %[b1].4s, %[a1].s[0]\n"
326 "fmla v21.4s, %[b1].4s, %[a1].s[1]\n"
328 "fmla v22.4s, %[b1].4s, %[a1].s[2]\n"
329 "fmla v23.4s, %[b1].4s, %[a1].s[3]\n"
330 ASM_PREFETCHWL2(
"[%[c_ptr], #384]")
331 "fmla v24.4s, %[b2].4s, %[a0].s[0]\n"
332 "fmla v25.4s, %[b2].4s, %[a0].s[1]\n"
333 ASM_PREFETCHWL2(
"[%[c_ptr], #448]")
334 "fmla v26.4s, %[b2].4s, %[a0].s[2]\n"
335 "fmla v27.4s, %[b2].4s, %[a0].s[3]\n"
336 ASM_PREFETCHWL2(
"[%[c_ptr], #512]")
337 "fmla v28.4s, %[b2].4s, %[a1].s[0]\n"
338 ASM_PREFETCHWL2(
"[%[c_ptr], #576]")
339 "fmla v29.4s, %[b2].4s, %[a1].s[1]\n"
340 ASM_PREFETCHWL2(
"[%[c_ptr], #640]")
341 "fmla v30.4s, %[b2].4s, %[a1].s[2]\n"
342 ASM_PREFETCHWL2(
"[%[c_ptr], #704]")
343 "fmla v31.4s, %[b2].4s, %[a1].s[3]\n"
347 "str q8, [%[c_ptr]]\n"
348 "str q16, [%[c_ptr], #16]\n"
349 "str q24, [%[c_ptr], #32]\n"
350 "str q9, [%[c_ptr], #48]\n"
351 "str q17, [%[c_ptr], #64]\n"
352 "str q25, [%[c_ptr], #80]\n"
353 "str q10, [%[c_ptr], #96]\n"
354 "str q18, [%[c_ptr], #112]\n"
355 "str q26, [%[c_ptr], #128]\n"
356 "str q11, [%[c_ptr], #144]\n"
357 "str q19, [%[c_ptr], #160]\n"
358 "str q27, [%[c_ptr], #176]\n"
359 "str q12, [%[c_ptr], #192]\n"
360 "str q20, [%[c_ptr], #208]\n"
361 "str q28, [%[c_ptr], #224]\n"
362 "str q13, [%[c_ptr], #240]\n"
363 "str q21, [%[c_ptr], #256]\n"
364 "str q29, [%[c_ptr], #272]\n"
365 "str q14, [%[c_ptr], #288]\n"
366 "str q22, [%[c_ptr], #304]\n"
367 "str q30, [%[c_ptr], #320]\n"
368 "str q15, [%[c_ptr], #336]\n"
369 "str q23, [%[c_ptr], #352]\n"
370 "str q31, [%[c_ptr], #368]\n"
371 "add %[c_ptr], %[c_ptr], #384\n"
373 [a_ptr]
"+r" (a_ptr), [b_ptr]
"+r" (b_ptr), [c_ptr]
"+r" (c_ptr),
374 [a0]
"+w" (a0), [a1]
"+w" (a1), [a0a]
"+w" (a0a), [a1a]
"+w" (a1a),
375 [b0]
"+w" (b0), [b1]
"+w" (b1), [b2]
"+w" (b2), [k]
"+r" (k)
377 :
"x20",
"x21",
"v8",
"v9",
"v10",
"v11",
"v12",
"v13",
"v14",
"v15",
"v16",
"v17",
"v18",
378 "v19",
"v20",
"v21",
"v22",
"v23",
"v24",
"v25",
"v26",
"v27",
"v28",
"v29",
"v30",
"v31",
"cc"