28 #include "../../asmlib.hpp"
32 void a64_sgemm_asimd_8x12_a53(
const float *Apanel,
const float *Bpanel,
float *Cpanel,
int ablocks,
int bblocks,
int K) {
33 const float *a_ptr = Apanel;
34 float *c_ptr = Cpanel;
36 for (
int yb=0; yb<ablocks; yb++) {
37 const float *a_ptr0 = a_ptr;
38 const float *b_ptr = Bpanel;
40 for (
int xb=0; xb<bblocks; xb++) {
45 int k = ((
K+1)/2) - 1;
47 register float32x4_t a0
asm(
"v0");
48 register float32x4_t a1
asm(
"v1");
49 register float32x4_t b0
asm(
"v2");
50 register float32x4_t b1
asm(
"v3");
51 register float32x4_t b2
asm(
"v4");
52 register float32x4_t a0a
asm(
"v5");
53 register float32x4_t a1a
asm(
"v6");
58 "ldr %q[a0], [%[a_ptr]]\n"
60 "ldr %q[b0], [%[b_ptr]]\n"
62 "ldr %q[a1], [%[a_ptr], #16]\n"
64 "ldr %q[b1], [%[b_ptr], #16]\n"
101 "ldr %d[b2], [%[b_ptr], #32]\n"
103 "fmla v8.4s , %[b0].4s, %[a0].s[0]\n"
104 "ldr x20, [%[b_ptr], #40]\n"
105 "fmla v9.4s , %[b0].4s, %[a0].s[1]\n"
106 "subs %w[k], %w[k], #1\n"
107 "fmla v10.4s, %[b0].4s, %[a0].s[2]\n"
109 "ldr %d[a0a], [%[a_ptr], #32]\n"
110 "ins %[b2].d[1], x20\n"
111 "fmla v11.4s, %[b0].4s, %[a0].s[3]\n"
112 "ldr x20, [%[a_ptr], #40]\n"
113 "fmla v12.4s, %[b0].4s, %[a1].s[0]\n"
114 "fmla v13.4s, %[b0].4s, %[a1].s[1]\n"
116 "ldr %d[a1a], [%[a_ptr], #48]\n"
117 "ins %[a0a].d[1], x20\n"
118 "fmla v14.4s, %[b0].4s, %[a1].s[2]\n"
119 "ldr x20, [%[a_ptr], #56]\n"
120 "fmla v15.4s, %[b0].4s, %[a1].s[3]\n"
121 "fmla v16.4s, %[b1].4s, %[a0].s[0]\n"
123 "ldr %d[b0], [%[b_ptr], #48]\n"
124 "ins %[a1a].d[1], x20\n"
125 "fmla v17.4s, %[b1].4s, %[a0].s[1]\n"
126 "ldr x20, [%[b_ptr], #56]\n"
127 "fmla v18.4s, %[b1].4s, %[a0].s[2]\n"
128 "fmla v19.4s, %[b1].4s, %[a0].s[3]\n"
131 "ins %[b0].d[1], x20\n"
132 "fmla v20.4s, %[b1].4s, %[a1].s[0]\n"
133 "fmla v21.4s, %[b1].4s, %[a1].s[1]\n"
134 "fmla v22.4s, %[b1].4s, %[a1].s[2]\n"
138 "fmla v23.4s, %[b1].4s, %[a1].s[3]\n"
139 "fmla v24.4s, %[b2].4s, %[a0].s[0]\n"
140 "fmla v25.4s, %[b2].4s, %[a0].s[1]\n"
142 "ldr %d[b1], [%[b_ptr], #64]\n"
144 "fmla v26.4s, %[b2].4s, %[a0].s[2]\n"
145 "ldr x20, [%[b_ptr], #72]\n"
146 "fmla v27.4s, %[b2].4s, %[a0].s[3]\n"
147 "fmla v28.4s, %[b2].4s, %[a1].s[0]\n"
150 "ins %[b1].d[1], x20\n"
151 "fmla v29.4s, %[b2].4s, %[a1].s[1]\n"
152 "fmla v30.4s, %[b2].4s, %[a1].s[2]\n"
153 "fmla v31.4s, %[b2].4s, %[a1].s[3]\n"
156 "ldr %d[b2], [%[b_ptr], #80]\n"
158 "fmla v8.4s , %[b0].4s, %[a0a].s[0]\n"
159 "ldr x20, [%[b_ptr], #88]\n"
160 "fmla v9.4s , %[b0].4s, %[a0a].s[1]\n"
161 "fmla v10.4s, %[b0].4s, %[a0a].s[2]\n"
163 "ldr %d[a0], [%[a_ptr], #64]\n"
164 "ins %[b2].d[1], x20\n"
165 "fmla v11.4s, %[b0].4s, %[a0a].s[3]\n"
166 "ldr x20, [%[a_ptr], #72]\n"
167 "fmla v12.4s, %[b0].4s, %[a1a].s[0]\n"
168 "fmla v13.4s, %[b0].4s, %[a1a].s[1]\n"
170 "ldr %d[a1], [%[a_ptr], #80]\n"
171 "ins %[a0].d[1], x20\n"
172 "fmla v14.4s, %[b0].4s, %[a1a].s[2]\n"
173 "ldr x20, [%[a_ptr], #88]\n"
174 "fmla v15.4s, %[b0].4s, %[a1a].s[3]\n"
175 "fmla v16.4s, %[b1].4s, %[a0a].s[0]\n"
177 "ldr %d[b0], [%[b_ptr], #96]\n"
178 "ins %[a1].d[1], x20\n"
179 "fmla v17.4s, %[b1].4s, %[a0a].s[1]\n"
180 "ldr x20, [%[b_ptr], #104]\n"
181 "fmla v18.4s, %[b1].4s, %[a0a].s[2]\n"
182 "fmla v19.4s, %[b1].4s, %[a0a].s[3]\n"
185 "ins %[b0].d[1], x20\n"
186 "fmla v20.4s, %[b1].4s, %[a1a].s[0]\n"
187 "fmla v21.4s, %[b1].4s, %[a1a].s[1]\n"
188 "fmla v22.4s, %[b1].4s, %[a1a].s[2]\n"
192 "fmla v23.4s, %[b1].4s, %[a1a].s[3]\n"
193 "fmla v24.4s, %[b2].4s, %[a0a].s[0]\n"
194 "fmla v25.4s, %[b2].4s, %[a0a].s[1]\n"
196 "ldr %d[b1], [%[b_ptr], #112]\n"
198 "fmla v26.4s, %[b2].4s, %[a0a].s[2]\n"
199 "ldr x20, [%[b_ptr], #120]\n"
200 "fmla v27.4s, %[b2].4s, %[a0a].s[3]\n"
201 "add %[a_ptr], %[a_ptr], #64\n"
202 "fmla v28.4s, %[b2].4s, %[a1a].s[0]\n"
203 "add %[b_ptr], %[b_ptr], #96\n"
206 "ins %[b1].d[1], x20\n"
207 "fmla v29.4s, %[b2].4s, %[a1a].s[1]\n"
208 "fmla v30.4s, %[b2].4s, %[a1a].s[2]\n"
209 "fmla v31.4s, %[b2].4s, %[a1a].s[3]\n"
215 "cbnz %w[oddk], 2f\n"
218 "ldr %d[b2], [%[b_ptr], #32]\n"
220 "fmla v8.4s , %[b0].4s, %[a0].s[0]\n"
221 "ldr x20, [%[b_ptr], #40]\n"
222 "fmla v9.4s , %[b0].4s, %[a0].s[1]\n"
223 "subs %w[k], %w[k], #1\n"
224 "fmla v10.4s, %[b0].4s, %[a0].s[2]\n"
226 "ldr %d[a0a], [%[a_ptr], #32]\n"
227 "ins %[b2].d[1], x20\n"
228 "fmla v11.4s, %[b0].4s, %[a0].s[3]\n"
229 "ldr x20, [%[a_ptr], #40]\n"
230 "fmla v12.4s, %[b0].4s, %[a1].s[0]\n"
231 "fmla v13.4s, %[b0].4s, %[a1].s[1]\n"
233 "ldr %d[a1a], [%[a_ptr], #48]\n"
234 "ins %[a0a].d[1], x20\n"
235 "fmla v14.4s, %[b0].4s, %[a1].s[2]\n"
236 "ldr x20, [%[a_ptr], #56]\n"
237 "fmla v15.4s, %[b0].4s, %[a1].s[3]\n"
238 "fmla v16.4s, %[b1].4s, %[a0].s[0]\n"
240 "ldr %d[b0], [%[b_ptr], #48]\n"
241 "ins %[a1a].d[1], x20\n"
242 "fmla v17.4s, %[b1].4s, %[a0].s[1]\n"
243 "ldr x20, [%[b_ptr], #56]\n"
244 "fmla v18.4s, %[b1].4s, %[a0].s[2]\n"
245 "fmla v19.4s, %[b1].4s, %[a0].s[3]\n"
247 "ins %[b0].d[1], x20\n"
248 "fmla v20.4s, %[b1].4s, %[a1].s[0]\n"
249 "fmla v21.4s, %[b1].4s, %[a1].s[1]\n"
250 "fmla v22.4s, %[b1].4s, %[a1].s[2]\n"
253 "fmla v23.4s, %[b1].4s, %[a1].s[3]\n"
254 "fmla v24.4s, %[b2].4s, %[a0].s[0]\n"
255 "fmla v25.4s, %[b2].4s, %[a0].s[1]\n"
257 "ldr %d[b1], [%[b_ptr], #64]\n"
259 "fmla v26.4s, %[b2].4s, %[a0].s[2]\n"
260 "ldr x20, [%[b_ptr], #72]\n"
261 "fmla v27.4s, %[b2].4s, %[a0].s[3]\n"
262 "fmla v28.4s, %[b2].4s, %[a1].s[0]\n"
264 "ins %[b1].d[1], x20\n"
265 "fmla v29.4s, %[b2].4s, %[a1].s[1]\n"
266 "fmla v30.4s, %[b2].4s, %[a1].s[2]\n"
267 "fmla v31.4s, %[b2].4s, %[a1].s[3]\n"
269 "ldr %d[b2], [%[b_ptr], #80]\n"
271 "fmla v8.4s , %[b0].4s, %[a0a].s[0]\n"
272 "ldr x20, [%[b_ptr], #88]\n"
273 "fmla v9.4s , %[b0].4s, %[a0a].s[1]\n"
274 "fmla v10.4s, %[b0].4s, %[a0a].s[2]\n"
276 "ins %[b2].d[1], x20\n"
277 "fmla v11.4s, %[b0].4s, %[a0a].s[3]\n"
278 "fmla v12.4s, %[b0].4s, %[a1a].s[0]\n"
279 "fmla v13.4s, %[b0].4s, %[a1a].s[1]\n"
280 "fmla v14.4s, %[b0].4s, %[a1a].s[2]\n"
281 "fmla v15.4s, %[b0].4s, %[a1a].s[3]\n"
282 "fmla v16.4s, %[b1].4s, %[a0a].s[0]\n"
283 "fmla v17.4s, %[b1].4s, %[a0a].s[1]\n"
284 "fmla v18.4s, %[b1].4s, %[a0a].s[2]\n"
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 "fmla v22.4s, %[b1].4s, %[a1a].s[2]\n"
289 "fmla v23.4s, %[b1].4s, %[a1a].s[3]\n"
290 "fmla v24.4s, %[b2].4s, %[a0a].s[0]\n"
291 "fmla v25.4s, %[b2].4s, %[a0a].s[1]\n"
292 "fmla v26.4s, %[b2].4s, %[a0a].s[2]\n"
293 "fmla v27.4s, %[b2].4s, %[a0a].s[3]\n"
294 "fmla v28.4s, %[b2].4s, %[a1a].s[0]\n"
295 "fmla v29.4s, %[b2].4s, %[a1a].s[1]\n"
296 "add %[a_ptr], %[a_ptr], #64\n"
297 "fmla v30.4s, %[b2].4s, %[a1a].s[2]\n"
298 "add %[b_ptr], %[b_ptr], #96\n"
299 "fmla v31.4s, %[b2].4s, %[a1a].s[3]\n"
304 "ldr %d[b2], [%[b_ptr], #32]\n"
306 "fmla v8.4s , %[b0].4s, %[a0].s[0]\n"
307 "ldr x20, [%[b_ptr], #40]\n"
308 "fmla v9.4s , %[b0].4s, %[a0].s[1]\n"
309 "fmla v10.4s, %[b0].4s, %[a0].s[2]\n"
311 "ins %[b2].d[1], x20\n"
312 "fmla v11.4s, %[b0].4s, %[a0].s[3]\n"
313 "fmla v12.4s, %[b0].4s, %[a1].s[0]\n"
314 "fmla v13.4s, %[b0].4s, %[a1].s[1]\n"
315 "fmla v14.4s, %[b0].4s, %[a1].s[2]\n"
316 "fmla v15.4s, %[b0].4s, %[a1].s[3]\n"
317 "fmla v16.4s, %[b1].4s, %[a0].s[0]\n"
318 "fmla v17.4s, %[b1].4s, %[a0].s[1]\n"
319 "fmla v18.4s, %[b1].4s, %[a0].s[2]\n"
320 "fmla v19.4s, %[b1].4s, %[a0].s[3]\n"
321 "fmla v20.4s, %[b1].4s, %[a1].s[0]\n"
322 "fmla v21.4s, %[b1].4s, %[a1].s[1]\n"
323 "fmla v22.4s, %[b1].4s, %[a1].s[2]\n"
324 "fmla v23.4s, %[b1].4s, %[a1].s[3]\n"
325 "fmla v24.4s, %[b2].4s, %[a0].s[0]\n"
326 "fmla v25.4s, %[b2].4s, %[a0].s[1]\n"
327 "fmla v26.4s, %[b2].4s, %[a0].s[2]\n"
328 "fmla v27.4s, %[b2].4s, %[a0].s[3]\n"
329 "fmla v28.4s, %[b2].4s, %[a1].s[0]\n"
330 "fmla v29.4s, %[b2].4s, %[a1].s[1]\n"
331 "add %[a_ptr], %[a_ptr], #32\n"
332 "fmla v30.4s, %[b2].4s, %[a1].s[2]\n"
333 "add %[b_ptr], %[b_ptr], #48\n"
334 "fmla v31.4s, %[b2].4s, %[a1].s[3]\n"
338 "str q8, [%[c_ptr]]\n"
339 "str q16, [%[c_ptr], #16]\n"
340 "str q24, [%[c_ptr], #32]\n"
341 "str q9, [%[c_ptr], #48]\n"
342 "str q17, [%[c_ptr], #64]\n"
343 "str q25, [%[c_ptr], #80]\n"
344 "str q10, [%[c_ptr], #96]\n"
345 "str q18, [%[c_ptr], #112]\n"
346 "str q26, [%[c_ptr], #128]\n"
347 "str q11, [%[c_ptr], #144]\n"
348 "str q19, [%[c_ptr], #160]\n"
349 "str q27, [%[c_ptr], #176]\n"
350 "str q12, [%[c_ptr], #192]\n"
351 "str q20, [%[c_ptr], #208]\n"
352 "str q28, [%[c_ptr], #224]\n"
353 "str q13, [%[c_ptr], #240]\n"
354 "str q21, [%[c_ptr], #256]\n"
355 "str q29, [%[c_ptr], #272]\n"
356 "str q14, [%[c_ptr], #288]\n"
357 "str q22, [%[c_ptr], #304]\n"
358 "str q30, [%[c_ptr], #320]\n"
359 "str q15, [%[c_ptr], #336]\n"
360 "str q23, [%[c_ptr], #352]\n"
361 "str q31, [%[c_ptr], #368]\n"
362 "add %[c_ptr], %[c_ptr], #384\n"
364 [a_ptr]
"+r" (a_ptr), [b_ptr]
"+r" (b_ptr), [c_ptr]
"+r" (c_ptr),
365 [a0]
"+w" (a0), [a1]
"+w" (a1), [a0a]
"+w" (a0a), [a1a]
"+w" (a1a),
366 [b0]
"+w" (b0), [b1]
"+w" (b1), [b2]
"+w" (b2), [k]
"+r" (k)
368 :
"x20",
"x21",
"v8",
"v9",
"v10",
"v11",
"v12",
"v13",
"v14",
"v15",
"v16",
"v17",
"v18",
369 "v19",
"v20",
"v21",
"v22",
"v23",
"v24",
"v25",
"v26",
"v27",
"v28",
"v29",
"v30",
"v31",
"cc"