28 #if defined(__aarch64__)
33 void a64_fp32_nhwc_3x3_s1_output3x3_mla_depthfirst_indirect_impl(
34 const float *
const *
const input_ptrs,
35 float *
const *
const outptrs,
37 unsigned int n_channels,
44 float *
const *outptrs;
47 const float *inptrs[25];
50 const float *
const *
const input_ptrs,
51 float *
const *
const outptrs,
52 const void *
const params,
55 ) : outptrs(outptrs), params(params), min(min), max(max)
57 inptrs[0] = input_ptrs[12];
58 inptrs[1] = input_ptrs[0];
59 inptrs[2] = input_ptrs[4];
60 inptrs[3] = input_ptrs[20];
61 inptrs[4] = input_ptrs[7];
62 inptrs[5] = input_ptrs[24];
63 inptrs[6] = input_ptrs[11];
64 inptrs[7] = input_ptrs[1];
65 inptrs[8] = input_ptrs[3];
66 inptrs[9] = input_ptrs[13];
67 inptrs[10] = input_ptrs[5];
68 inptrs[11] = input_ptrs[9];
69 inptrs[12] = input_ptrs[15];
70 inptrs[13] = input_ptrs[17];
71 inptrs[14] = input_ptrs[19];
72 inptrs[15] = input_ptrs[21];
73 inptrs[16] = input_ptrs[6];
74 inptrs[17] = input_ptrs[8];
75 inptrs[18] = input_ptrs[23];
76 inptrs[19] = input_ptrs[16];
77 inptrs[20] = input_ptrs[2];
78 inptrs[21] = input_ptrs[18];
79 inptrs[22] = input_ptrs[10];
80 inptrs[23] = input_ptrs[14];
81 inptrs[24] = input_ptrs[22];
86 Args params_struct(input_ptrs, outptrs, params,
91 "lsr x8, %x[n_channels], #0x2\n"
92 "ldr x17, [%x[params_struct], %[offsetof_args_outptrs]]\n"
93 "ldr x16, [%x[params_struct], %[offsetof_args_params]]\n"
94 "add x20, %x[params_struct], %[offsetof_args_min]\n"
95 "ld1r { v15.4s }, [x20]\n"
96 "add x20, %x[params_struct], %[offsetof_args_max]\n"
97 "ld1r { v14.4s }, [x20]\n"
98 "add x15, %x[params_struct], %[offsetof_Args_inptrs]\n"
102 "ldr q31, [x16, #0x0]\n"
103 "ldr q0, [x16, #0x10]\n"
104 "cmp x7, x8, LSL #4\n"
105 "ldr q1, [x16, #0x20]\n"
106 "ldr q2, [x16, #0x30]\n"
107 "ldr q3, [x16, #0x40]\n"
108 "ldr q4, [x16, #0x50]\n"
109 "ldr q5, [x16, #0x60]\n"
110 "ldr q6, [x16, #0x70]\n"
111 "ldr q7, [x16, #0x80]\n"
112 "ldr q8, [x16, #0x90]\n"
113 "add x16, x16, #0xa0\n"
114 "ldp x21, x20, [x15, #0x0]\n"
115 "ldr q9, [x21, x14]\n"
116 "ldr q10, [x20, x14]\n"
117 "ldp x21, x20, [x15, #0x10]\n"
118 "ldr q11, [x21, x14]\n"
119 "ldr q12, [x20, x14]\n"
120 "ldr x20, [x15, #0x20]\n"
121 "ldr q13, [x20, x14]\n"
124 "mov v29.16b, v31.16b\n fmla v29.4s, v8.4s, v9.4s\n"
125 "mov v28.16b, v31.16b\n fmla v28.4s, v7.4s, v9.4s\n"
126 "ldr x26, [x15, #0x30]\n"
127 "ldr x23, [x15, #0x38]\n"
128 "mov v27.16b, v31.16b\n fmla v27.4s, v6.4s, v9.4s\n"
129 "fmla v29.4s, v0.4s, v10.4s\n"
130 "ldr x22, [x15, #0x28]\n"
131 "ldr x20, [x15, #0x48]\n"
132 "ldr q19, [x20, x14]\n"
133 "fmla v28.4s, v4.4s, v13.4s\n"
134 "mov v26.16b, v31.16b\n fmla v26.4s, v5.4s, v9.4s\n"
135 "ldr x21, [x15, #0x40]\n"
136 "mov v25.16b, v31.16b\n fmla v25.4s, v4.4s, v9.4s\n"
137 "mov v24.16b, v31.16b\n fmla v24.4s, v3.4s, v9.4s\n"
138 "ldr x25, [x15, #0x50]\n"
139 "ldr x24, [x15, #0x58]\n"
140 "fmla v27.4s, v2.4s, v11.4s\n"
141 "ldr q17, [x26, x14]\n"
142 "mov v23.16b, v31.16b\n fmla v23.4s, v2.4s, v9.4s\n"
143 "ldr x20, [x15, #0x60]\n"
144 "fmla v29.4s, v5.4s, v13.4s\n"
145 "fmla v28.4s, v6.4s, v17.4s\n"
146 "ldr x12, [x15, #0x70]\n"
147 "ldr x11, [x15, #0x88]\n"
148 "mov v22.16b, v31.16b\n fmla v22.4s, v0.4s, v9.4s\n"
149 "fmla v27.4s, v3.4s, v13.4s\n"
150 "ldr x10, [x17, #0x0]\n"
151 "add x13, x13, #0x10\n"
152 "fmla v26.4s, v2.4s, v13.4s\n"
153 "fmla v25.4s, v1.4s, v13.4s\n"
154 "ldr x9, [x17, #0x8]\n"
155 "ldr x28, [x17, #0x10]\n"
156 "fmla v24.4s, v0.4s, v13.4s\n"
157 "ldr q18, [x23, x14]\n"
158 "fmla v23.4s, v6.4s, v12.4s\n"
159 "ldr q16, [x22, x14]\n"
160 "mov v21.16b, v31.16b\n fmla v21.4s, v1.4s, v9.4s\n"
161 "ldr q31, [x16, #0x0]\n"
162 "fmla v29.4s, v7.4s, v17.4s\n"
163 "ldr x23, [x15, #0x68]\n"
164 "fmla v28.4s, v0.4s, v18.4s\n"
165 "fmla v22.4s, v8.4s, v16.4s\n"
166 "ldr q16, [x21, x14]\n"
167 "ldr x22, [x15, #0x78]\n"
168 "fmla v26.4s, v4.4s, v17.4s\n"
169 "fmla v25.4s, v3.4s, v17.4s\n"
170 "ldr x21, [x15, #0x80]\n"
171 "ldr x27, [x17, #0x18]\n"
172 "fmla v21.4s, v0.4s, v17.4s\n"
173 "fmla v24.4s, v4.4s, v19.4s\n"
174 "fmla v23.4s, v1.4s, v17.4s\n"
175 "ldr q17, [x25, x14]\n"
176 "fmla v29.4s, v1.4s, v18.4s\n"
177 "ldr q20, [x24, x14]\n"
178 "fmla v28.4s, v2.4s, v16.4s\n"
179 "fmla v27.4s, v1.4s, v16.4s\n"
180 "ldr q16, [x20, x14]\n"
181 "ldr x26, [x15, #0x90]\n"
182 "fmla v25.4s, v5.4s, v19.4s\n"
183 "fmla v21.4s, v2.4s, v19.4s\n"
184 "ldr x25, [x15, #0xa0]\n"
185 "ldr x20, [x15, #0x98]\n"
186 "fmla v26.4s, v0.4s, v17.4s\n"
187 "fmla v24.4s, v2.4s, v20.4s\n"
188 "fmla v28.4s, v8.4s, v19.4s\n"
189 "fmla v27.4s, v7.4s, v19.4s\n"
190 "fmla v22.4s, v1.4s, v19.4s\n"
191 "ldr q19, [x23, x14]\n"
192 "fmla v23.4s, v3.4s, v16.4s\n"
193 "ldr x24, [x15, #0xa8]\n"
194 "fmla v26.4s, v6.4s, v16.4s\n"
195 "ldr q18, [x21, x14]\n"
196 "fmla v25.4s, v7.4s, v19.4s\n"
197 "ldr x23, [x15, #0xc0]\n"
198 "fmla v24.4s, v6.4s, v19.4s\n"
199 "fmla v21.4s, v4.4s, v19.4s\n"
200 "fmla v29.4s, v3.4s, v17.4s\n"
201 "ldr q17, [x12, x14]\n"
202 "fmla v27.4s, v5.4s, v20.4s\n"
203 "ldr q16, [x22, x14]\n"
204 "fmla v23.4s, v5.4s, v19.4s\n"
205 "fmla v22.4s, v3.4s, v19.4s\n"
206 "ldr x22, [x15, #0xb0]\n"
207 "ldr x21, [x15, #0xb8]\n"
208 "fmla v26.4s, v8.4s, v19.4s\n"
209 "fmla v24.4s, v8.4s, v17.4s\n"
210 "fmla v21.4s, v6.4s, v16.4s\n"
211 "fmla v28.4s, v3.4s, v18.4s\n"
212 "fmla v25.4s, v0.4s, v18.4s\n"
213 "fmla v22.4s, v5.4s, v17.4s\n"
214 "ldr q17, [x11, x14]\n"
215 "fmla v23.4s, v7.4s, v16.4s\n"
216 "ldr q16, [x26, x14]\n"
217 "fmla v29.4s, v4.4s, v18.4s\n"
218 "fmla v26.4s, v1.4s, v18.4s\n"
219 "ldr q18, [x20, x14]\n"
220 "fmla v28.4s, v5.4s, v17.4s\n"
221 "fmla v27.4s, v4.4s, v17.4s\n"
222 "fmla v25.4s, v2.4s, v17.4s\n"
223 "fmla v24.4s, v1.4s, v17.4s\n"
224 "ldr q17, [x25, x14]\n"
225 "fmla v21.4s, v8.4s, v16.4s\n"
226 "ldr x20, [x15, #0x20]\n"
227 "fmla v22.4s, v7.4s, v16.4s\n"
228 "ldr q16, [x24, x14]\n"
229 "fmla v29.4s, v2.4s, v17.4s\n"
230 "fmla v26.4s, v7.4s, v18.4s\n"
231 "fmla v25.4s, v6.4s, v18.4s\n"
232 "fmla v23.4s, v4.4s, v18.4s\n"
233 "fmla v21.4s, v3.4s, v18.4s\n"
234 "ldr q18, [x22, x14]\n"
235 "fmla v22.4s, v4.4s, v16.4s\n"
236 "ldr q4, [x16, #0x50]\n"
237 "fmla v28.4s, v1.4s, v17.4s\n"
238 "ldr q1, [x16, #0x20]\n"
239 "fmla v27.4s, v0.4s, v17.4s\n"
240 "ldr q17, [x21, x14]\n"
241 "fmla v29.4s, v6.4s, v18.4s\n"
242 "fmax v29.4s, v29.4s, v15.4s\n"
243 "fmla v24.4s, v7.4s, v16.4s\n"
244 "fmla v21.4s, v5.4s, v16.4s\n"
245 "fmin v29.4s, v29.4s, v14.4s\n"
246 "str q29, [x10, x13]\n"
247 "fmla v23.4s, v0.4s, v18.4s\n"
248 "ldr q0, [x16, #0x10]\n"
249 "fmla v22.4s, v2.4s, v17.4s\n"
250 "ldr q2, [x16, #0x30]\n"
251 "fmla v25.4s, v8.4s, v16.4s\n"
252 "ldr q16, [x23, x14]\n"
253 "fmla v26.4s, v3.4s, v18.4s\n"
254 "ldr q3, [x16, #0x40]\n"
255 "fmla v27.4s, v8.4s, v17.4s\n"
256 "fmla v24.4s, v5.4s, v17.4s\n"
257 "ldr q5, [x16, #0x60]\n"
258 "fmax v28.4s, v28.4s, v15.4s\n"
259 "fmla v23.4s, v8.4s, v16.4s\n"
260 "ldr q8, [x16, #0x90]\n"
261 "fmla v21.4s, v7.4s, v16.4s\n"
262 "ldr q7, [x16, #0x80]\n"
263 "fmla v22.4s, v6.4s, v16.4s\n"
264 "ldr q13, [x20, x7]\n"
265 "ldr q6, [x16, #0x70]\n"
266 "fmax v27.4s, v27.4s, v15.4s\n"
267 "fmax v26.4s, v26.4s, v15.4s\n"
268 "fmax v25.4s, v25.4s, v15.4s\n"
269 "ldr x24, [x17, #0x20]\n"
270 "ldp x21, x20, [x15, #0x0]\n"
271 "ldr q9, [x21, x7]\n"
272 "ldr q10, [x20, x7]\n"
273 "fmin v28.4s, v28.4s, v14.4s\n"
274 "fmin v27.4s, v27.4s, v14.4s\n"
275 "ldp x21, x20, [x15, #0x10]\n"
276 "ldr q11, [x21, x7]\n"
277 "fmin v26.4s, v26.4s, v14.4s\n"
278 "fmin v25.4s, v25.4s, v14.4s\n"
279 "ldr q12, [x20, x7]\n"
280 "fmax v24.4s, v24.4s, v15.4s\n"
281 "fmax v23.4s, v23.4s, v15.4s\n"
282 "str q28, [x9, x13]\n"
283 "fmax v21.4s, v21.4s, v15.4s\n"
284 "fmax v22.4s, v22.4s, v15.4s\n"
285 "str q27, [x28, x13]\n"
286 "ldr x23, [x17, #0x28]\n"
287 "str q26, [x27, x13]\n"
288 "ldr x22, [x17, #0x30]\n"
289 "ldr x21, [x17, #0x38]\n"
290 "add x7, x7, #0x10\n"
291 "str q25, [x24, x13]\n"
292 "ldr x20, [x17, #0x40]\n"
293 "cmp x7, x8, LSL #4\n"
294 "fmin v24.4s, v24.4s, v14.4s\n"
295 "fmin v23.4s, v23.4s, v14.4s\n"
296 "fmin v21.4s, v21.4s, v14.4s\n"
297 "add x14, x14, #0x10\n"
298 "str q24, [x23, x13]\n"
299 "fmin v22.4s, v22.4s, v14.4s\n"
300 "str q23, [x22, x13]\n"
301 "add x16, x16, #0xa0\n"
302 "str q21, [x21, x13]\n"
303 "str q22, [x20, x13]\n"
306 "mov v29.16b, v31.16b\n fmla v29.4s, v8.4s, v9.4s\n"
307 "mov v28.16b, v31.16b\n fmla v28.4s, v7.4s, v9.4s\n"
308 "ldr x23, [x15, #0x30]\n"
309 "ldr x22, [x15, #0x38]\n"
310 "mov v27.16b, v31.16b\n fmla v27.4s, v6.4s, v9.4s\n"
311 "fmla v29.4s, v0.4s, v10.4s\n"
312 "ldr x21, [x15, #0x28]\n"
313 "ldr x20, [x15, #0x48]\n"
314 "ldr q19, [x20, x14]\n"
315 "fmla v28.4s, v4.4s, v13.4s\n"
316 "mov v26.16b, v31.16b\n fmla v26.4s, v5.4s, v9.4s\n"
317 "ldr x20, [x15, #0x40]\n"
318 "mov v25.16b, v31.16b\n fmla v25.4s, v4.4s, v9.4s\n"
319 "mov v24.16b, v31.16b\n fmla v24.4s, v3.4s, v9.4s\n"
320 "ldr x25, [x15, #0x50]\n"
321 "ldr x24, [x15, #0x58]\n"
322 "fmla v27.4s, v2.4s, v11.4s\n"
323 "ldr q17, [x23, x14]\n"
324 "mov v23.16b, v31.16b\n fmla v23.4s, v2.4s, v9.4s\n"
325 "ldr x23, [x15, #0x60]\n"
326 "fmla v29.4s, v5.4s, v13.4s\n"
327 "fmla v28.4s, v6.4s, v17.4s\n"
328 "ldr x12, [x15, #0x70]\n"
329 "ldr x11, [x15, #0x88]\n"
330 "mov v22.16b, v31.16b\n fmla v22.4s, v0.4s, v9.4s\n"
331 "fmla v27.4s, v3.4s, v13.4s\n"
332 "ldr x10, [x17, #0x0]\n"
333 "add x13, x13, #0x10\n"
334 "fmla v26.4s, v2.4s, v13.4s\n"
335 "fmla v25.4s, v1.4s, v13.4s\n"
336 "ldr x9, [x17, #0x8]\n"
337 "ldr x28, [x17, #0x10]\n"
338 "fmla v24.4s, v0.4s, v13.4s\n"
339 "ldr q18, [x22, x14]\n"
340 "fmla v23.4s, v6.4s, v12.4s\n"
341 "ldr q16, [x21, x14]\n"
342 "mov v21.16b, v31.16b\n fmla v21.4s, v1.4s, v9.4s\n"
343 "fmla v29.4s, v7.4s, v17.4s\n"
344 "ldr x22, [x15, #0x68]\n"
345 "ldr x21, [x15, #0x78]\n"
346 "fmla v28.4s, v0.4s, v18.4s\n"
347 "fmla v22.4s, v8.4s, v16.4s\n"
348 "ldr q16, [x20, x14]\n"
349 "ldr x20, [x15, #0x80]\n"
350 "fmla v26.4s, v4.4s, v17.4s\n"
351 "fmla v25.4s, v3.4s, v17.4s\n"
352 "ldr x27, [x17, #0x18]\n"
353 "fmla v21.4s, v0.4s, v17.4s\n"
354 "fmla v24.4s, v4.4s, v19.4s\n"
355 "fmla v23.4s, v1.4s, v17.4s\n"
356 "ldr q17, [x25, x14]\n"
357 "fmla v29.4s, v1.4s, v18.4s\n"
358 "ldr q20, [x24, x14]\n"
359 "fmla v28.4s, v2.4s, v16.4s\n"
360 "fmla v27.4s, v1.4s, v16.4s\n"
361 "ldr q16, [x23, x14]\n"
362 "ldr x26, [x15, #0x90]\n"
363 "fmla v25.4s, v5.4s, v19.4s\n"
364 "fmla v21.4s, v2.4s, v19.4s\n"
365 "ldr x25, [x15, #0xa0]\n"
366 "ldr x24, [x15, #0x98]\n"
367 "fmla v26.4s, v0.4s, v17.4s\n"
368 "fmla v24.4s, v2.4s, v20.4s\n"
369 "fmla v28.4s, v8.4s, v19.4s\n"
370 "fmla v27.4s, v7.4s, v19.4s\n"
371 "fmla v22.4s, v1.4s, v19.4s\n"
372 "ldr q19, [x22, x14]\n"
373 "fmla v23.4s, v3.4s, v16.4s\n"
374 "ldr x23, [x15, #0xa8]\n"
375 "fmla v26.4s, v6.4s, v16.4s\n"
376 "ldr q18, [x20, x14]\n"
377 "fmla v25.4s, v7.4s, v19.4s\n"
378 "ldr x22, [x15, #0xc0]\n"
379 "fmla v24.4s, v6.4s, v19.4s\n"
380 "fmla v21.4s, v4.4s, v19.4s\n"
381 "fmla v29.4s, v3.4s, v17.4s\n"
382 "ldr q17, [x12, x14]\n"
383 "fmla v27.4s, v5.4s, v20.4s\n"
384 "ldr q16, [x21, x14]\n"
385 "fmla v23.4s, v5.4s, v19.4s\n"
386 "fmla v22.4s, v3.4s, v19.4s\n"
387 "ldr x21, [x15, #0xb0]\n"
388 "ldr x20, [x15, #0xb8]\n"
389 "fmla v26.4s, v8.4s, v19.4s\n"
390 "fmla v24.4s, v8.4s, v17.4s\n"
391 "fmla v21.4s, v6.4s, v16.4s\n"
392 "fmla v28.4s, v3.4s, v18.4s\n"
393 "fmla v25.4s, v0.4s, v18.4s\n"
394 "fmla v22.4s, v5.4s, v17.4s\n"
395 "ldr q17, [x11, x14]\n"
396 "fmla v23.4s, v7.4s, v16.4s\n"
397 "ldr q16, [x26, x14]\n"
398 "fmla v29.4s, v4.4s, v18.4s\n"
399 "fmla v26.4s, v1.4s, v18.4s\n"
400 "ldr q18, [x24, x14]\n"
401 "fmla v28.4s, v5.4s, v17.4s\n"
402 "fmla v27.4s, v4.4s, v17.4s\n"
403 "fmla v25.4s, v2.4s, v17.4s\n"
404 "fmla v24.4s, v1.4s, v17.4s\n"
405 "ldr q17, [x25, x14]\n"
406 "fmla v21.4s, v8.4s, v16.4s\n"
407 "fmla v22.4s, v7.4s, v16.4s\n"
408 "ldr q16, [x23, x14]\n"
409 "fmla v29.4s, v2.4s, v17.4s\n"
410 "fmla v26.4s, v7.4s, v18.4s\n"
411 "fmla v25.4s, v6.4s, v18.4s\n"
412 "fmla v23.4s, v4.4s, v18.4s\n"
413 "fmla v21.4s, v3.4s, v18.4s\n"
414 "ldr q18, [x21, x14]\n"
415 "fmla v22.4s, v4.4s, v16.4s\n"
416 "fmla v28.4s, v1.4s, v17.4s\n"
417 "fmax v28.4s, v28.4s, v15.4s\n"
418 "fmla v27.4s, v0.4s, v17.4s\n"
419 "ldr q17, [x20, x14]\n"
420 "fmla v29.4s, v6.4s, v18.4s\n"
421 "fmax v29.4s, v29.4s, v15.4s\n"
422 "fmla v24.4s, v7.4s, v16.4s\n"
423 "fmla v21.4s, v5.4s, v16.4s\n"
424 "fmin v29.4s, v29.4s, v14.4s\n"
425 "str q29, [x10, x13]\n"
426 "fmla v23.4s, v0.4s, v18.4s\n"
427 "fmla v22.4s, v2.4s, v17.4s\n"
428 "ldr x20, [x17, #0x20]\n"
429 "fmin v28.4s, v28.4s, v14.4s\n"
430 "fmla v25.4s, v8.4s, v16.4s\n"
431 "ldr q16, [x22, x14]\n"
432 "fmla v26.4s, v3.4s, v18.4s\n"
433 "fmax v26.4s, v26.4s, v15.4s\n"
434 "fmla v27.4s, v8.4s, v17.4s\n"
435 "fmla v24.4s, v5.4s, v17.4s\n"
436 "fmax v27.4s, v27.4s, v15.4s\n"
437 "str q28, [x9, x13]\n"
438 "fmla v23.4s, v8.4s, v16.4s\n"
439 "fmla v21.4s, v7.4s, v16.4s\n"
440 "fmax v25.4s, v25.4s, v15.4s\n"
441 "ldr x23, [x17, #0x28]\n"
442 "fmla v22.4s, v6.4s, v16.4s\n"
443 "fmin v27.4s, v27.4s, v14.4s\n"
444 "str q27, [x28, x13]\n"
445 "ldr x22, [x17, #0x30]\n"
446 "fmin v26.4s, v26.4s, v14.4s\n"
447 "fmin v25.4s, v25.4s, v14.4s\n"
448 "str q26, [x27, x13]\n"
449 "ldr x21, [x17, #0x38]\n"
450 "fmax v24.4s, v24.4s, v15.4s\n"
451 "fmax v23.4s, v23.4s, v15.4s\n"
452 "str q25, [x20, x13]\n"
453 "ldr x20, [x17, #0x40]\n"
454 "fmax v21.4s, v21.4s, v15.4s\n"
455 "fmax v22.4s, v22.4s, v15.4s\n"
456 "add x14, x14, #0x10\n"
457 "fmin v24.4s, v24.4s, v14.4s\n"
458 "fmin v23.4s, v23.4s, v14.4s\n"
459 "str q24, [x23, x13]\n"
460 "fmin v21.4s, v21.4s, v14.4s\n"
461 "fmin v22.4s, v22.4s, v14.4s\n"
462 "str q23, [x22, x13]\n"
463 "str q21, [x21, x13]\n"
464 "str q22, [x20, x13]\n"
466 "tst %x[n_channels], #0x3\n"
468 "ldr q31, [x16, #0x0]\n"
469 "ldr q0, [x16, #0x10]\n"
471 "ldr q1, [x16, #0x20]\n"
472 "ldr q2, [x16, #0x30]\n"
473 "ldr q3, [x16, #0x40]\n"
474 "ldr q4, [x16, #0x50]\n"
475 "ldr q5, [x16, #0x60]\n"
476 "ldr q6, [x16, #0x70]\n"
477 "ldr q7, [x16, #0x80]\n"
478 "ldr q8, [x16, #0x90]\n"
479 "ldr x24, [x15, #0x0]\n"
480 "ldr x23, [x15, #0x8]\n"
481 "add x24, x24, x14\n"
482 "add x23, x23, x14\n"
483 "ldr x22, [x15, #0x10]\n"
484 "ldr x21, [x15, #0x18]\n"
485 "add x22, x22, x14\n"
486 "add x21, x21, x14\n"
487 "ldr x20, [x15, #0x20]\n"
488 "add x20, x20, x14\n"
489 "tbz %x[n_channels], #1, 4f\n"
490 "ld1 { v9.d }[0], [x24], #0x8\n"
491 "ld1 { v10.d }[0], [x23], #0x8\n"
492 "ld1 { v11.d }[0], [x22], #0x8\n"
493 "ld1 { v12.d }[0], [x21], #0x8\n"
494 "ld1 { v13.d }[0], [x20], #0x8\n"
495 "tbz %x[n_channels], #0, 5f\n"
496 "ld1 { v9.s }[2], [x24], #0x4\n"
497 "ld1 { v10.s }[2], [x23], #0x4\n"
498 "ld1 { v11.s }[2], [x22], #0x4\n"
499 "ld1 { v12.s }[2], [x21], #0x4\n"
500 "ld1 { v13.s }[2], [x20], #0x4\n"
503 "ld1 { v9.s }[0], [x24], #0x4\n"
504 "ld1 { v10.s }[0], [x23], #0x4\n"
505 "ld1 { v11.s }[0], [x22], #0x4\n"
506 "ld1 { v12.s }[0], [x21], #0x4\n"
507 "ld1 { v13.s }[0], [x20], #0x4\n"
509 "mov v23.16b, v31.16b\n fmla v23.4s, v8.4s, v9.4s\n"
510 "mov v25.16b, v31.16b\n fmla v25.4s, v6.4s, v9.4s\n"
511 "ldr x20, [x15, #0x28]\n"
512 "add x20, x20, x14\n"
513 "mov v24.16b, v31.16b\n fmla v24.4s, v7.4s, v9.4s\n"
514 "mov v26.16b, v31.16b\n fmla v26.4s, v5.4s, v9.4s\n"
515 "mov v27.16b, v31.16b\n fmla v27.4s, v4.4s, v9.4s\n"
516 "mov v28.16b, v31.16b\n fmla v28.4s, v3.4s, v9.4s\n"
517 "mov v29.16b, v31.16b\n fmla v29.4s, v2.4s, v9.4s\n"
518 "fmla v23.4s, v0.4s, v10.4s\n"
519 "fmla v25.4s, v2.4s, v11.4s\n"
520 "mov v30.16b, v31.16b\n fmla v30.4s, v1.4s, v9.4s\n"
521 "fmla v31.4s, v0.4s, v9.4s\n"
522 "fmla v29.4s, v6.4s, v12.4s\n"
523 "fmla v23.4s, v5.4s, v13.4s\n"
524 "fmla v24.4s, v4.4s, v13.4s\n"
525 "fmla v25.4s, v3.4s, v13.4s\n"
526 "fmla v26.4s, v2.4s, v13.4s\n"
527 "fmla v27.4s, v1.4s, v13.4s\n"
528 "fmla v28.4s, v0.4s, v13.4s\n"
529 "tbz %x[n_channels], #1, 6f\n"
530 "ld1 { v12.d }[0], [x20], #0x8\n"
531 "tbz %x[n_channels], #0, 7f\n"
532 "ld1 { v12.s }[2], [x20], #0x4\n"
535 "ld1 { v12.s }[0], [x20], #0x4\n"
537 "ldr x20, [x15, #0x30]\n"
538 "fmla v31.4s, v8.4s, v12.4s\n"
539 "add x20, x20, x14\n"
540 "tbz %x[n_channels], #1, 8f\n"
541 "ld1 { v11.d }[0], [x20], #0x8\n"
542 "tbz %x[n_channels], #0, 9f\n"
543 "ld1 { v11.s }[2], [x20], #0x4\n"
546 "ld1 { v11.s }[0], [x20], #0x4\n"
548 "ldr x20, [x15, #0x38]\n"
549 "fmla v23.4s, v7.4s, v11.4s\n"
550 "fmla v24.4s, v6.4s, v11.4s\n"
551 "add x20, x20, x14\n"
552 "fmla v26.4s, v4.4s, v11.4s\n"
553 "fmla v27.4s, v3.4s, v11.4s\n"
554 "fmla v29.4s, v1.4s, v11.4s\n"
555 "fmla v30.4s, v0.4s, v11.4s\n"
556 "tbz %x[n_channels], #1, 10f\n"
557 "ld1 { v13.d }[0], [x20], #0x8\n"
558 "tbz %x[n_channels], #0, 11f\n"
559 "ld1 { v13.s }[2], [x20], #0x4\n"
562 "ld1 { v13.s }[0], [x20], #0x4\n"
564 "ldr x20, [x15, #0x40]\n"
565 "fmla v23.4s, v1.4s, v13.4s\n"
566 "fmla v24.4s, v0.4s, v13.4s\n"
567 "add x20, x20, x14\n"
568 "tbz %x[n_channels], #1, 12f\n"
569 "ld1 { v12.d }[0], [x20], #0x8\n"
570 "tbz %x[n_channels], #0, 13f\n"
571 "ld1 { v12.s }[2], [x20], #0x4\n"
574 "ld1 { v12.s }[0], [x20], #0x4\n"
576 "ldr x20, [x15, #0x48]\n"
577 "fmla v24.4s, v2.4s, v12.4s\n"
578 "fmla v25.4s, v1.4s, v12.4s\n"
579 "add x20, x20, x14\n"
580 "tbz %x[n_channels], #1, 14f\n"
581 "ld1 { v10.d }[0], [x20], #0x8\n"
582 "tbz %x[n_channels], #0, 15f\n"
583 "ld1 { v10.s }[2], [x20], #0x4\n"
586 "ld1 { v10.s }[0], [x20], #0x4\n"
588 "ldr x20, [x15, #0x50]\n"
589 "fmla v24.4s, v8.4s, v10.4s\n"
590 "fmla v25.4s, v7.4s, v10.4s\n"
591 "add x20, x20, x14\n"
592 "fmla v27.4s, v5.4s, v10.4s\n"
593 "fmla v28.4s, v4.4s, v10.4s\n"
594 "fmla v30.4s, v2.4s, v10.4s\n"
595 "fmla v31.4s, v1.4s, v10.4s\n"
596 "tbz %x[n_channels], #1, 16f\n"
597 "ld1 { v11.d }[0], [x20], #0x8\n"
598 "tbz %x[n_channels], #0, 17f\n"
599 "ld1 { v11.s }[2], [x20], #0x4\n"
602 "ld1 { v11.s }[0], [x20], #0x4\n"
604 "ldr x20, [x15, #0x58]\n"
605 "fmla v23.4s, v3.4s, v11.4s\n"
606 "fmla v26.4s, v0.4s, v11.4s\n"
607 "add x20, x20, x14\n"
608 "tbz %x[n_channels], #1, 18f\n"
609 "ld1 { v13.d }[0], [x20], #0x8\n"
610 "tbz %x[n_channels], #0, 19f\n"
611 "ld1 { v13.s }[2], [x20], #0x4\n"
614 "ld1 { v13.s }[0], [x20], #0x4\n"
616 "ldr x20, [x15, #0x60]\n"
617 "fmla v25.4s, v5.4s, v13.4s\n"
618 "fmla v28.4s, v2.4s, v13.4s\n"
619 "add x20, x20, x14\n"
620 "tbz %x[n_channels], #1, 20f\n"
621 "ld1 { v12.d }[0], [x20], #0x8\n"
622 "tbz %x[n_channels], #0, 21f\n"
623 "ld1 { v12.s }[2], [x20], #0x4\n"
626 "ld1 { v12.s }[0], [x20], #0x4\n"
628 "ldr x20, [x15, #0x68]\n"
629 "fmla v26.4s, v6.4s, v12.4s\n"
630 "fmla v29.4s, v3.4s, v12.4s\n"
631 "add x20, x20, x14\n"
632 "tbz %x[n_channels], #1, 22f\n"
633 "ld1 { v10.d }[0], [x20], #0x8\n"
634 "tbz %x[n_channels], #0, 23f\n"
635 "ld1 { v10.s }[2], [x20], #0x4\n"
638 "ld1 { v10.s }[0], [x20], #0x4\n"
640 "ldr x20, [x15, #0x70]\n"
641 "fmla v26.4s, v8.4s, v10.4s\n"
642 "fmla v27.4s, v7.4s, v10.4s\n"
643 "add x20, x20, x14\n"
644 "fmla v28.4s, v6.4s, v10.4s\n"
645 "fmla v29.4s, v5.4s, v10.4s\n"
646 "fmla v30.4s, v4.4s, v10.4s\n"
647 "fmla v31.4s, v3.4s, v10.4s\n"
648 "tbz %x[n_channels], #1, 24f\n"
649 "ld1 { v11.d }[0], [x20], #0x8\n"
650 "tbz %x[n_channels], #0, 25f\n"
651 "ld1 { v11.s }[2], [x20], #0x4\n"
654 "ld1 { v11.s }[0], [x20], #0x4\n"
656 "ldr x20, [x15, #0x78]\n"
657 "fmla v28.4s, v8.4s, v11.4s\n"
658 "fmla v31.4s, v5.4s, v11.4s\n"
659 "add x20, x20, x14\n"
660 "tbz %x[n_channels], #1, 26f\n"
661 "ld1 { v13.d }[0], [x20], #0x8\n"
662 "tbz %x[n_channels], #0, 27f\n"
663 "ld1 { v13.s }[2], [x20], #0x4\n"
666 "ld1 { v13.s }[0], [x20], #0x4\n"
668 "ldr x20, [x15, #0x80]\n"
669 "fmla v29.4s, v7.4s, v13.4s\n"
670 "fmla v30.4s, v6.4s, v13.4s\n"
671 "add x20, x20, x14\n"
672 "tbz %x[n_channels], #1, 28f\n"
673 "ld1 { v12.d }[0], [x20], #0x8\n"
674 "tbz %x[n_channels], #0, 29f\n"
675 "ld1 { v12.s }[2], [x20], #0x4\n"
678 "ld1 { v12.s }[0], [x20], #0x4\n"
680 "ldr x20, [x15, #0x88]\n"
681 "fmla v23.4s, v4.4s, v12.4s\n"
682 "fmla v24.4s, v3.4s, v12.4s\n"
683 "add x20, x20, x14\n"
684 "fmla v26.4s, v1.4s, v12.4s\n"
685 "fmla v27.4s, v0.4s, v12.4s\n"
686 "tbz %x[n_channels], #1, 30f\n"
687 "ld1 { v11.d }[0], [x20], #0x8\n"
688 "tbz %x[n_channels], #0, 31f\n"
689 "ld1 { v11.s }[2], [x20], #0x4\n"
692 "ld1 { v11.s }[0], [x20], #0x4\n"
694 "ldr x20, [x15, #0x90]\n"
695 "fmla v24.4s, v5.4s, v11.4s\n"
696 "fmla v25.4s, v4.4s, v11.4s\n"
697 "add x20, x20, x14\n"
698 "fmla v27.4s, v2.4s, v11.4s\n"
699 "fmla v28.4s, v1.4s, v11.4s\n"
700 "tbz %x[n_channels], #1, 32f\n"
701 "ld1 { v13.d }[0], [x20], #0x8\n"
702 "tbz %x[n_channels], #0, 33f\n"
703 "ld1 { v13.s }[2], [x20], #0x4\n"
706 "ld1 { v13.s }[0], [x20], #0x4\n"
708 "ldr x20, [x15, #0x98]\n"
709 "fmla v30.4s, v8.4s, v13.4s\n"
710 "fmla v31.4s, v7.4s, v13.4s\n"
711 "add x20, x20, x14\n"
712 "tbz %x[n_channels], #1, 34f\n"
713 "ld1 { v12.d }[0], [x20], #0x8\n"
714 "tbz %x[n_channels], #0, 35f\n"
715 "ld1 { v12.s }[2], [x20], #0x4\n"
718 "ld1 { v12.s }[0], [x20], #0x4\n"
720 "ldr x20, [x15, #0xa0]\n"
721 "fmla v26.4s, v7.4s, v12.4s\n"
722 "fmla v27.4s, v6.4s, v12.4s\n"
723 "add x20, x20, x14\n"
724 "fmla v29.4s, v4.4s, v12.4s\n"
725 "fmla v30.4s, v3.4s, v12.4s\n"
726 "tbz %x[n_channels], #1, 36f\n"
727 "ld1 { v11.d }[0], [x20], #0x8\n"
728 "tbz %x[n_channels], #0, 37f\n"
729 "ld1 { v11.s }[2], [x20], #0x4\n"
732 "ld1 { v11.s }[0], [x20], #0x4\n"
734 "ldr x20, [x15, #0xa8]\n"
735 "fmla v23.4s, v2.4s, v11.4s\n"
736 "fmla v24.4s, v1.4s, v11.4s\n"
737 "add x20, x20, x14\n"
738 "fmla v25.4s, v0.4s, v11.4s\n"
739 "tbz %x[n_channels], #1, 38f\n"
740 "ld1 { v13.d }[0], [x20], #0x8\n"
741 "tbz %x[n_channels], #0, 39f\n"
742 "ld1 { v13.s }[2], [x20], #0x4\n"
745 "ld1 { v13.s }[0], [x20], #0x4\n"
747 "ldr x20, [x15, #0xb0]\n"
748 "fmla v27.4s, v8.4s, v13.4s\n"
749 "fmla v28.4s, v7.4s, v13.4s\n"
750 "add x20, x20, x14\n"
751 "fmla v30.4s, v5.4s, v13.4s\n"
752 "fmla v31.4s, v4.4s, v13.4s\n"
753 "tbz %x[n_channels], #1, 40f\n"
754 "ld1 { v12.d }[0], [x20], #0x8\n"
755 "tbz %x[n_channels], #0, 41f\n"
756 "ld1 { v12.s }[2], [x20], #0x4\n"
759 "ld1 { v12.s }[0], [x20], #0x4\n"
761 "ldr x20, [x15, #0xb8]\n"
762 "fmla v23.4s, v6.4s, v12.4s\n"
763 "fmla v26.4s, v3.4s, v12.4s\n"
764 "add x20, x20, x14\n"
765 "fmla v29.4s, v0.4s, v12.4s\n"
766 "tbz %x[n_channels], #1, 42f\n"
767 "ld1 { v11.d }[0], [x20], #0x8\n"
768 "tbz %x[n_channels], #0, 43f\n"
769 "ld1 { v11.s }[2], [x20], #0x4\n"
772 "ld1 { v11.s }[0], [x20], #0x4\n"
774 "ldr x20, [x15, #0xc0]\n"
775 "fmla v25.4s, v8.4s, v11.4s\n"
776 "fmla v28.4s, v5.4s, v11.4s\n"
777 "add x20, x20, x14\n"
778 "fmla v31.4s, v2.4s, v11.4s\n"
779 "tbz %x[n_channels], #1, 44f\n"
780 "ld1 { v13.d }[0], [x20], #0x8\n"
781 "tbz %x[n_channels], #0, 45f\n"
782 "ld1 { v13.s }[2], [x20], #0x4\n"
785 "ld1 { v13.s }[0], [x20], #0x4\n"
787 "fmla v29.4s, v8.4s, v13.4s\n"
788 "fmla v30.4s, v7.4s, v13.4s\n"
789 "fmax v23.4s, v23.4s, v15.4s\n"
790 "fmla v31.4s, v6.4s, v13.4s\n"
791 "fmax v24.4s, v24.4s, v15.4s\n"
792 "fmax v25.4s, v25.4s, v15.4s\n"
793 "fmax v26.4s, v26.4s, v15.4s\n"
794 "fmax v27.4s, v27.4s, v15.4s\n"
795 "fmax v28.4s, v28.4s, v15.4s\n"
796 "fmax v29.4s, v29.4s, v15.4s\n"
797 "fmax v30.4s, v30.4s, v15.4s\n"
798 "fmax v31.4s, v31.4s, v15.4s\n"
799 "fmin v23.4s, v23.4s, v14.4s\n"
800 "fmin v24.4s, v24.4s, v14.4s\n"
801 "fmin v25.4s, v25.4s, v14.4s\n"
802 "fmin v26.4s, v26.4s, v14.4s\n"
803 "fmin v27.4s, v27.4s, v14.4s\n"
804 "fmin v28.4s, v28.4s, v14.4s\n"
805 "fmin v29.4s, v29.4s, v14.4s\n"
806 "fmin v30.4s, v30.4s, v14.4s\n"
807 "fmin v31.4s, v31.4s, v14.4s\n"
808 "tbz %x[n_channels], #1, 46f\n"
809 "ldr x20, [x17, #0x0]\n"
810 "add x20, x20, x13\n"
811 "st1 { v23.d }[0], [x20]\n"
812 "ldr x23, [x17, #0x8]\n"
813 "ldr x22, [x17, #0x10]\n"
814 "ldr x21, [x17, #0x18]\n"
815 "add x23, x23, x13\n"
816 "add x22, x22, x13\n"
817 "ldr x20, [x17, #0x20]\n"
818 "add x21, x21, x13\n"
819 "add x20, x20, x13\n"
820 "st1 { v24.d }[0], [x23]\n"
821 "st1 { v25.d }[0], [x22]\n"
822 "ldr x23, [x17, #0x28]\n"
823 "ldr x22, [x17, #0x30]\n"
824 "add x23, x23, x13\n"
825 "st1 { v26.d }[0], [x21]\n"
826 "ldr x21, [x17, #0x38]\n"
827 "add x22, x22, x13\n"
828 "add x21, x21, x13\n"
829 "st1 { v27.d }[0], [x20]\n"
830 "ldr x20, [x17, #0x40]\n"
831 "add x20, x20, x13\n"
832 "add x13, x13, #0x8\n"
833 "st1 { v28.d }[0], [x23]\n"
834 "st1 { v29.d }[0], [x22]\n"
835 "st1 { v30.d }[0], [x21]\n"
836 "st1 { v31.d }[0], [x20]\n"
837 "tbz %x[n_channels], #0, 47f\n"
838 "ldr x20, [x17, #0x0]\n"
839 "add x20, x20, x13\n"
840 "st1 { v23.s }[2], [x20]\n"
841 "ldr x23, [x17, #0x8]\n"
842 "ldr x22, [x17, #0x10]\n"
843 "ldr x21, [x17, #0x18]\n"
844 "add x23, x23, x13\n"
845 "add x22, x22, x13\n"
846 "ldr x20, [x17, #0x20]\n"
847 "add x21, x21, x13\n"
848 "add x20, x20, x13\n"
849 "st1 { v24.s }[2], [x23]\n"
850 "st1 { v25.s }[2], [x22]\n"
851 "ldr x23, [x17, #0x28]\n"
852 "ldr x22, [x17, #0x30]\n"
853 "add x23, x23, x13\n"
854 "st1 { v26.s }[2], [x21]\n"
855 "ldr x21, [x17, #0x38]\n"
856 "add x22, x22, x13\n"
857 "add x21, x21, x13\n"
858 "st1 { v27.s }[2], [x20]\n"
859 "ldr x20, [x17, #0x40]\n"
860 "add x20, x20, x13\n"
861 "st1 { v28.s }[2], [x23]\n"
862 "st1 { v29.s }[2], [x22]\n"
863 "st1 { v30.s }[2], [x21]\n"
864 "st1 { v31.s }[2], [x20]\n"
867 "ldr x20, [x17, #0x0]\n"
868 "add x20, x20, x13\n"
869 "st1 { v23.s }[0], [x20]\n"
870 "ldr x23, [x17, #0x8]\n"
871 "ldr x22, [x17, #0x10]\n"
872 "ldr x21, [x17, #0x18]\n"
873 "add x23, x23, x13\n"
874 "add x22, x22, x13\n"
875 "ldr x20, [x17, #0x20]\n"
876 "add x21, x21, x13\n"
877 "add x20, x20, x13\n"
878 "st1 { v24.s }[0], [x23]\n"
879 "st1 { v25.s }[0], [x22]\n"
880 "ldr x23, [x17, #0x28]\n"
881 "ldr x22, [x17, #0x30]\n"
882 "add x23, x23, x13\n"
883 "st1 { v26.s }[0], [x21]\n"
884 "ldr x21, [x17, #0x38]\n"
885 "add x22, x22, x13\n"
886 "add x21, x21, x13\n"
887 "st1 { v27.s }[0], [x20]\n"
888 "ldr x20, [x17, #0x40]\n"
889 "add x20, x20, x13\n"
890 "st1 { v28.s }[0], [x23]\n"
891 "st1 { v29.s }[0], [x22]\n"
892 "st1 { v30.s }[0], [x21]\n"
893 "st1 { v31.s }[0], [x20]\n"
897 : [n_channels]
"r" ((
unsigned long) n_channels), [offsetof_Args_inptrs]
"I" (offsetof(Args, inptrs)), [offsetof_args_max]
"I" (offsetof(Args, max)), [offsetof_args_min]
"I" (offsetof(Args, min)), [offsetof_args_outptrs]
"I" (offsetof(Args, outptrs)), [offsetof_args_params]
"I" (offsetof(Args, params)), [params_struct]
"r" (¶ms_struct)
898 :
"cc",
"memory",
"v0",
"v1",
"v2",
"v3",
"v4",
"v5",
"v6",
"v7",
"v8",
"v9",
"v10",
"v11",
"v12",
"v13",
"v14",
"v15",
"v16",
"v17",
"v18",
"v19",
"v20",
"v21",
"v22",
"v23",
"v24",
"v25",
"v26",
"v27",
"v28",
"v29",
"v30",
"v31",
"x7",
"x8",
"x9",
"x10",
"x11",
"x12",
"x13",
"x14",
"x15",
"x16",
"x17",
"x20",
"x21",
"x22",
"x23",
"x24",
"x25",
"x26",
"x27",
"x28"
905 #endif // defined(__aarch64__)