28 #if defined(__aarch64__)
33 void a64_fp32_nhwc_3x3_s2_output2x2_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[1];
60 inptrs[3] = input_ptrs[3];
61 inptrs[4] = input_ptrs[4];
62 inptrs[5] = input_ptrs[5];
63 inptrs[6] = input_ptrs[6];
64 inptrs[7] = input_ptrs[2];
65 inptrs[8] = input_ptrs[8];
66 inptrs[9] = input_ptrs[9];
67 inptrs[10] = input_ptrs[7];
68 inptrs[11] = input_ptrs[15];
69 inptrs[12] = input_ptrs[10];
70 inptrs[13] = input_ptrs[16];
71 inptrs[14] = input_ptrs[11];
72 inptrs[15] = input_ptrs[18];
73 inptrs[16] = input_ptrs[13];
74 inptrs[17] = input_ptrs[19];
75 inptrs[18] = input_ptrs[20];
76 inptrs[19] = input_ptrs[14];
77 inptrs[20] = input_ptrs[21];
78 inptrs[21] = input_ptrs[17];
79 inptrs[22] = input_ptrs[23];
80 inptrs[23] = input_ptrs[22];
81 inptrs[24] = input_ptrs[24];
86 Args params_struct(input_ptrs, outptrs, params,
90 "ldr x21, [%x[params_struct], %[offsetof_args_outptrs]]\n"
92 "lsr x24, %x[n_channels], #0x2\n"
93 "ldr x23, [%x[params_struct], %[offsetof_args_params]]\n"
94 "add x20, %x[params_struct], %[offsetof_args_min]\n"
95 "ld1r { v26.4s }, [x20]\n"
96 "add x20, %x[params_struct], %[offsetof_args_max]\n"
97 "ld1r { v27.4s }, [x20]\n"
98 "add x13, %x[params_struct], %[offsetof_Args_inptrs]\n"
99 "ldp x12, x11, [x21, #0x0]\n"
100 "ldp x10, x9, [x21, #0x10]\n"
102 "sub x22, XZR, x25\n"
104 "ldr q31, [x23, #0x0]\n"
105 "ldr q0, [x23, #0x10]\n"
106 "cmp x25, x24, LSL #4\n"
107 "ldr q1, [x23, #0x20]\n"
108 "ldr q2, [x23, #0x30]\n"
109 "ldr q3, [x23, #0x40]\n"
110 "ldr q4, [x23, #0x50]\n"
111 "ldr q5, [x23, #0x60]\n"
112 "ldr q6, [x23, #0x70]\n"
113 "ldr q7, [x23, #0x80]\n"
114 "ldr q8, [x23, #0x90]\n"
115 "add x23, x23, #0xa0\n"
116 "ldp x21, x20, [x13, #0x0]\n"
117 "ldr q9, [x21, x28]\n"
118 "ldr q10, [x20, x28]\n"
119 "ldp x21, x20, [x13, #0x10]\n"
120 "ldr q11, [x21, x28]\n"
121 "ldr q12, [x20, x28]\n"
122 "ldp x21, x20, [x13, #0x20]\n"
123 "ldr q13, [x21, x28]\n"
124 "ldr q14, [x20, x28]\n"
125 "ldp x21, x20, [x13, #0x30]\n"
126 "ldr q15, [x21, x28]\n"
127 "ldr q16, [x20, x28]\n"
130 "mov v24.16b, v31.16b\n fmla v24.4s, v8.4s, v9.4s\n"
131 "mov v23.16b, v31.16b\n fmla v23.4s, v6.4s, v9.4s\n"
132 "ldr x21, [x13, #0x40]\n"
133 "ldr x20, [x13, #0x48]\n"
134 "fmla v24.4s, v0.4s, v10.4s\n"
135 "fmla v23.4s, v1.4s, v12.4s\n"
136 "ldr q20, [x20, x28]\n"
137 "ldr x20, [x13, #0x50]\n"
138 "fmla v24.4s, v1.4s, v11.4s\n"
139 "ldr q19, [x21, x28]\n"
140 "fmla v23.4s, v2.4s, v13.4s\n"
141 "ldr q18, [x20, x28]\n"
142 "fmla v24.4s, v3.4s, v14.4s\n"
143 "fmla v23.4s, v0.4s, v16.4s\n"
144 "ldr x20, [x13, #0x58]\n"
145 "ldr q17, [x20, x28]\n"
146 "fmla v24.4s, v4.4s, v15.4s\n"
147 "fmla v23.4s, v4.4s, v19.4s\n"
148 "ldr x21, [x13, #0x78]\n"
149 "ldr x20, [x13, #0x60]\n"
150 "ldr q22, [x20, x28]\n"
151 "fmla v24.4s, v2.4s, v16.4s\n"
152 "fmla v23.4s, v5.4s, v20.4s\n"
153 "ldr x20, [x13, #0x80]\n"
154 "ldr q21, [x20, x28]\n"
155 "mov v20.16b, v31.16b\n fmla v20.4s, v2.4s, v9.4s\n"
156 "mov v19.16b, v31.16b\n fmla v19.4s, v0.4s, v9.4s\n"
157 "ldr q31, [x23, #0x0]\n"
158 "fmla v24.4s, v5.4s, v18.4s\n"
159 "fmla v23.4s, v3.4s, v18.4s\n"
160 "ldr q16, [x21, x28]\n"
161 "ldr x20, [x13, #0x68]\n"
162 "ldr q18, [x20, x28]\n"
163 "fmla v20.4s, v3.4s, v17.4s\n"
164 "fmla v19.4s, v4.4s, v16.4s\n"
165 "ldr x20, [x13, #0x88]\n"
166 "ldr q16, [x20, x28]\n"
167 "fmla v20.4s, v0.4s, v22.4s\n"
168 "ldr q0, [x23, #0x10]\n"
169 "fmla v19.4s, v1.4s, v21.4s\n"
170 "ldr x20, [x13, #0x70]\n"
171 "ldr q17, [x20, x28]\n"
172 "fmla v20.4s, v4.4s, v18.4s\n"
173 "fmla v19.4s, v5.4s, v16.4s\n"
174 "ldr q4, [x23, #0x50]\n"
175 "ldr x20, [x13, #0x98]\n"
176 "fmla v24.4s, v6.4s, v22.4s\n"
177 "fmla v20.4s, v1.4s, v17.4s\n"
178 "ldr q16, [x20, x28]\n"
179 "ldr q1, [x23, #0x20]\n"
180 "fmla v19.4s, v2.4s, v16.4s\n"
181 "fmla v24.4s, v7.4s, v17.4s\n"
182 "ldr q2, [x23, #0x30]\n"
183 "ldr x20, [x13, #0x90]\n"
184 "fmla v23.4s, v7.4s, v21.4s\n"
185 "fmla v23.4s, v8.4s, v16.4s\n"
186 "ldr q16, [x20, x28]\n"
187 "ldr x20, [x13, #0xa8]\n"
188 "fmla v20.4s, v6.4s, v16.4s\n"
189 "fmax v24.4s, v24.4s, v26.4s\n"
190 "ldr q17, [x20, x28]\n"
191 "ldr x20, [x13, #0xa0]\n"
192 "fmla v19.4s, v3.4s, v17.4s\n"
193 "fmax v23.4s, v23.4s, v26.4s\n"
194 "ldr q16, [x20, x28]\n"
195 "ldr q3, [x23, #0x40]\n"
196 "fmla v20.4s, v7.4s, v16.4s\n"
197 "fmla v20.4s, v5.4s, v17.4s\n"
198 "ldr q5, [x23, #0x60]\n"
199 "ldr x20, [x13, #0xb0]\n"
200 "add x22, x22, #0x10\n"
201 "fmin v24.4s, v24.4s, v27.4s\n"
202 "ldr q16, [x20, x28]\n"
203 "ldr x20, [x13, #0xb8]\n"
204 "fmla v19.4s, v7.4s, v16.4s\n"
205 "fmin v23.4s, v23.4s, v27.4s\n"
206 "ldr q16, [x20, x28]\n"
207 "ldr q7, [x23, #0x80]\n"
208 "fmla v19.4s, v6.4s, v16.4s\n"
209 "fmla v20.4s, v8.4s, v16.4s\n"
210 "ldr q6, [x23, #0x70]\n"
211 "ldr x20, [x13, #0xc0]\n"
212 "fmax v20.4s, v20.4s, v26.4s\n"
213 "fmin v20.4s, v20.4s, v27.4s\n"
214 "ldr q16, [x20, x28]\n"
215 "fmla v19.4s, v8.4s, v16.4s\n"
216 "ldr q8, [x23, #0x90]\n"
217 "fmax v19.4s, v19.4s, v26.4s\n"
218 "ldp x21, x20, [x13, #0x0]\n"
219 "ldr q9, [x21, x25]\n"
220 "fmin v19.4s, v19.4s, v27.4s\n"
221 "add x28, x28, #0x10\n"
222 "ldr q10, [x20, x25]\n"
223 "ldp x21, x20, [x13, #0x10]\n"
224 "str q24, [x12, x22]\n"
225 "add x23, x23, #0xa0\n"
226 "ldr q11, [x21, x25]\n"
227 "ldr q12, [x20, x25]\n"
228 "str q23, [x11, x22]\n"
229 "ldp x21, x20, [x13, #0x20]\n"
230 "ldr q13, [x21, x25]\n"
231 "str q20, [x10, x22]\n"
232 "ldr q14, [x20, x25]\n"
233 "ldp x21, x20, [x13, #0x30]\n"
234 "str q19, [x9, x22]\n"
235 "ldr q15, [x21, x25]\n"
236 "ldr q16, [x20, x25]\n"
237 "add x25, x25, #0x10\n"
238 "cmp x25, x24, LSL #4\n"
241 "mov v25.16b, v31.16b\n fmla v25.4s, v8.4s, v9.4s\n"
242 "mov v24.16b, v31.16b\n fmla v24.4s, v6.4s, v9.4s\n"
243 "ldr x21, [x13, #0x40]\n"
244 "ldr x20, [x13, #0x48]\n"
245 "fmla v25.4s, v0.4s, v10.4s\n"
246 "fmla v24.4s, v1.4s, v12.4s\n"
247 "ldr q20, [x20, x28]\n"
248 "ldr x20, [x13, #0x50]\n"
249 "fmla v25.4s, v1.4s, v11.4s\n"
250 "ldr q18, [x21, x28]\n"
251 "fmla v24.4s, v2.4s, v13.4s\n"
252 "ldr q19, [x20, x28]\n"
253 "fmla v25.4s, v3.4s, v14.4s\n"
254 "fmla v24.4s, v0.4s, v16.4s\n"
255 "ldr x20, [x13, #0x58]\n"
256 "ldr q17, [x20, x28]\n"
257 "fmla v25.4s, v4.4s, v15.4s\n"
258 "fmla v24.4s, v4.4s, v18.4s\n"
259 "ldr x21, [x13, #0x78]\n"
260 "ldr x20, [x13, #0x60]\n"
261 "ldr q23, [x20, x28]\n"
262 "fmla v25.4s, v2.4s, v16.4s\n"
263 "fmla v24.4s, v5.4s, v20.4s\n"
264 "ldr x20, [x13, #0x80]\n"
265 "ldr q22, [x20, x28]\n"
266 "mov v21.16b, v31.16b\n fmla v21.4s, v2.4s, v9.4s\n"
267 "mov v20.16b, v31.16b\n fmla v20.4s, v0.4s, v9.4s\n"
268 "ldr x20, [x13, #0x68]\n"
269 "ldr q18, [x20, x28]\n"
270 "fmla v25.4s, v5.4s, v19.4s\n"
271 "fmla v24.4s, v3.4s, v19.4s\n"
272 "ldr q16, [x21, x28]\n"
273 "fmla v21.4s, v3.4s, v17.4s\n"
274 "fmla v20.4s, v4.4s, v16.4s\n"
275 "ldr x20, [x13, #0x88]\n"
276 "ldr q16, [x20, x28]\n"
277 "fmla v21.4s, v0.4s, v23.4s\n"
278 "fmla v20.4s, v1.4s, v22.4s\n"
279 "ldr x20, [x13, #0x70]\n"
280 "ldr q17, [x20, x28]\n"
281 "ldr x20, [x13, #0x98]\n"
282 "fmla v21.4s, v4.4s, v18.4s\n"
283 "ldr q19, [x20, x28]\n"
284 "fmla v20.4s, v5.4s, v16.4s\n"
285 "fmla v25.4s, v6.4s, v23.4s\n"
286 "ldr x20, [x13, #0x90]\n"
287 "ldr q16, [x20, x28]\n"
288 "fmla v21.4s, v1.4s, v17.4s\n"
289 "ldr x20, [x13, #0xa8]\n"
290 "fmla v20.4s, v2.4s, v19.4s\n"
291 "fmla v25.4s, v7.4s, v17.4s\n"
292 "ldr q18, [x20, x28]\n"
293 "ldr x20, [x13, #0xa0]\n"
294 "ldr q17, [x20, x28]\n"
295 "fmla v21.4s, v6.4s, v16.4s\n"
296 "fmla v20.4s, v3.4s, v18.4s\n"
297 "ldr x20, [x13, #0xb0]\n"
298 "ldr q16, [x20, x28]\n"
299 "fmla v21.4s, v7.4s, v17.4s\n"
300 "fmla v20.4s, v7.4s, v16.4s\n"
301 "ldr x20, [x13, #0xb8]\n"
302 "ldr q17, [x20, x28]\n"
303 "fmla v24.4s, v7.4s, v22.4s\n"
304 "fmla v21.4s, v5.4s, v18.4s\n"
305 "ldr x20, [x13, #0xc0]\n"
306 "fmla v20.4s, v6.4s, v17.4s\n"
307 "fmla v24.4s, v8.4s, v19.4s\n"
308 "ldr q16, [x20, x28]\n"
309 "fmla v21.4s, v8.4s, v17.4s\n"
310 "fmla v20.4s, v8.4s, v16.4s\n"
311 "fmax v25.4s, v25.4s, v26.4s\n"
312 "add x22, x22, #0x10\n"
313 "fmax v24.4s, v24.4s, v26.4s\n"
314 "fmax v21.4s, v21.4s, v26.4s\n"
315 "add x28, x28, #0x10\n"
316 "fmax v20.4s, v20.4s, v26.4s\n"
317 "fmin v25.4s, v25.4s, v27.4s\n"
318 "str q25, [x12, x22]\n"
319 "fmin v24.4s, v24.4s, v27.4s\n"
320 "fmin v21.4s, v21.4s, v27.4s\n"
321 "str q24, [x11, x22]\n"
322 "fmin v20.4s, v20.4s, v27.4s\n"
323 "str q21, [x10, x22]\n"
324 "str q20, [x9, x22]\n"
326 "tst %x[n_channels], #0x3\n"
328 "ldr q31, [x23, #0x0]\n"
329 "ldr q0, [x23, #0x10]\n"
331 "add x12, x12, x20\n"
332 "ldr q1, [x23, #0x20]\n"
333 "ldr q2, [x23, #0x30]\n"
334 "add x11, x11, x20\n"
335 "add x10, x10, x20\n"
336 "ldr q3, [x23, #0x40]\n"
337 "ldr q4, [x23, #0x50]\n"
339 "ldr q5, [x23, #0x60]\n"
340 "ldr q6, [x23, #0x70]\n"
341 "ldr q7, [x23, #0x80]\n"
342 "ldr q8, [x23, #0x90]\n"
343 "ldr x27, [x13, #0x0]\n"
344 "ldr x26, [x13, #0x8]\n"
345 "add x27, x27, x28\n"
346 "add x26, x26, x28\n"
347 "ldr x25, [x13, #0x10]\n"
348 "ldr x24, [x13, #0x18]\n"
349 "add x25, x25, x28\n"
350 "add x24, x24, x28\n"
351 "ldr x23, [x13, #0x20]\n"
352 "ldr x22, [x13, #0x28]\n"
353 "add x23, x23, x28\n"
354 "add x22, x22, x28\n"
355 "ldr x21, [x13, #0x30]\n"
356 "ldr x20, [x13, #0x38]\n"
357 "add x21, x21, x28\n"
358 "add x20, x20, x28\n"
359 "tbz %x[n_channels], #1, 4f\n"
360 "ld1 { v9.d }[0], [x27], #0x8\n"
361 "ld1 { v10.d }[0], [x26], #0x8\n"
362 "ld1 { v11.d }[0], [x25], #0x8\n"
363 "ld1 { v12.d }[0], [x24], #0x8\n"
364 "ld1 { v13.d }[0], [x23], #0x8\n"
365 "ld1 { v14.d }[0], [x22], #0x8\n"
366 "ld1 { v15.d }[0], [x21], #0x8\n"
367 "ld1 { v16.d }[0], [x20], #0x8\n"
368 "tbz %x[n_channels], #0, 5f\n"
369 "ld1 { v9.s }[2], [x27], #0x4\n"
370 "ld1 { v10.s }[2], [x26], #0x4\n"
371 "ld1 { v11.s }[2], [x25], #0x4\n"
372 "ld1 { v12.s }[2], [x24], #0x4\n"
373 "ld1 { v13.s }[2], [x23], #0x4\n"
374 "ld1 { v14.s }[2], [x22], #0x4\n"
375 "ld1 { v15.s }[2], [x21], #0x4\n"
376 "ld1 { v16.s }[2], [x20], #0x4\n"
379 "ld1 { v9.s }[0], [x27], #0x4\n"
380 "ld1 { v10.s }[0], [x26], #0x4\n"
381 "ld1 { v11.s }[0], [x25], #0x4\n"
382 "ld1 { v12.s }[0], [x24], #0x4\n"
383 "ld1 { v13.s }[0], [x23], #0x4\n"
384 "ld1 { v14.s }[0], [x22], #0x4\n"
385 "ld1 { v15.s }[0], [x21], #0x4\n"
386 "ld1 { v16.s }[0], [x20], #0x4\n"
388 "mov v28.16b, v31.16b\n fmla v28.4s, v8.4s, v9.4s\n"
389 "fmla v28.4s, v0.4s, v10.4s\n"
390 "ldr x20, [x13, #0x40]\n"
391 "add x20, x20, x28\n"
392 "mov v29.16b, v31.16b\n fmla v29.4s, v6.4s, v9.4s\n"
393 "fmla v28.4s, v1.4s, v11.4s\n"
394 "fmla v29.4s, v1.4s, v12.4s\n"
395 "fmla v28.4s, v3.4s, v14.4s\n"
396 "fmla v29.4s, v2.4s, v13.4s\n"
397 "fmla v28.4s, v4.4s, v15.4s\n"
398 "mov v30.16b, v31.16b\n fmla v30.4s, v2.4s, v9.4s\n"
399 "fmla v31.4s, v0.4s, v9.4s\n"
400 "fmla v28.4s, v2.4s, v16.4s\n"
401 "fmla v29.4s, v0.4s, v16.4s\n"
402 "tbz %x[n_channels], #1, 6f\n"
403 "ld1 { v11.d }[0], [x20], #0x8\n"
404 "tbz %x[n_channels], #0, 7f\n"
405 "ld1 { v11.s }[2], [x20], #0x4\n"
408 "ld1 { v11.s }[0], [x20], #0x4\n"
410 "ldr x20, [x13, #0x48]\n"
411 "fmla v29.4s, v4.4s, v11.4s\n"
412 "add x20, x20, x28\n"
413 "tbz %x[n_channels], #1, 8f\n"
414 "ld1 { v12.d }[0], [x20], #0x8\n"
415 "tbz %x[n_channels], #0, 9f\n"
416 "ld1 { v12.s }[2], [x20], #0x4\n"
419 "ld1 { v12.s }[0], [x20], #0x4\n"
421 "ldr x20, [x13, #0x50]\n"
422 "fmla v29.4s, v5.4s, v12.4s\n"
423 "add x20, x20, x28\n"
424 "tbz %x[n_channels], #1, 10f\n"
425 "ld1 { v13.d }[0], [x20], #0x8\n"
426 "tbz %x[n_channels], #0, 11f\n"
427 "ld1 { v13.s }[2], [x20], #0x4\n"
430 "ld1 { v13.s }[0], [x20], #0x4\n"
432 "ldr x20, [x13, #0x58]\n"
433 "fmla v28.4s, v5.4s, v13.4s\n"
434 "fmla v29.4s, v3.4s, v13.4s\n"
435 "add x20, x20, x28\n"
436 "tbz %x[n_channels], #1, 12f\n"
437 "ld1 { v14.d }[0], [x20], #0x8\n"
438 "tbz %x[n_channels], #0, 13f\n"
439 "ld1 { v14.s }[2], [x20], #0x4\n"
442 "ld1 { v14.s }[0], [x20], #0x4\n"
444 "ldr x20, [x13, #0x60]\n"
445 "fmla v30.4s, v3.4s, v14.4s\n"
446 "add x20, x20, x28\n"
447 "tbz %x[n_channels], #1, 14f\n"
448 "ld1 { v15.d }[0], [x20], #0x8\n"
449 "tbz %x[n_channels], #0, 15f\n"
450 "ld1 { v15.s }[2], [x20], #0x4\n"
453 "ld1 { v15.s }[0], [x20], #0x4\n"
455 "ldr x20, [x13, #0x68]\n"
456 "fmla v28.4s, v6.4s, v15.4s\n"
457 "fmla v30.4s, v0.4s, v15.4s\n"
458 "add x20, x20, x28\n"
459 "tbz %x[n_channels], #1, 16f\n"
460 "ld1 { v11.d }[0], [x20], #0x8\n"
461 "tbz %x[n_channels], #0, 17f\n"
462 "ld1 { v11.s }[2], [x20], #0x4\n"
465 "ld1 { v11.s }[0], [x20], #0x4\n"
467 "ldr x20, [x13, #0x70]\n"
468 "fmla v30.4s, v4.4s, v11.4s\n"
469 "add x20, x20, x28\n"
470 "tbz %x[n_channels], #1, 18f\n"
471 "ld1 { v16.d }[0], [x20], #0x8\n"
472 "tbz %x[n_channels], #0, 19f\n"
473 "ld1 { v16.s }[2], [x20], #0x4\n"
476 "ld1 { v16.s }[0], [x20], #0x4\n"
478 "ldr x20, [x13, #0x78]\n"
479 "fmla v28.4s, v7.4s, v16.4s\n"
480 "fmla v30.4s, v1.4s, v16.4s\n"
481 "add x20, x20, x28\n"
482 "tbz %x[n_channels], #1, 20f\n"
483 "ld1 { v13.d }[0], [x20], #0x8\n"
484 "tbz %x[n_channels], #0, 21f\n"
485 "ld1 { v13.s }[2], [x20], #0x4\n"
488 "ld1 { v13.s }[0], [x20], #0x4\n"
490 "ldr x20, [x13, #0x80]\n"
491 "fmla v31.4s, v4.4s, v13.4s\n"
492 "add x20, x20, x28\n"
493 "tbz %x[n_channels], #1, 22f\n"
494 "ld1 { v12.d }[0], [x20], #0x8\n"
495 "tbz %x[n_channels], #0, 23f\n"
496 "ld1 { v12.s }[2], [x20], #0x4\n"
499 "ld1 { v12.s }[0], [x20], #0x4\n"
501 "ldr x20, [x13, #0x88]\n"
502 "fmla v29.4s, v7.4s, v12.4s\n"
503 "fmla v31.4s, v1.4s, v12.4s\n"
504 "add x20, x20, x28\n"
505 "tbz %x[n_channels], #1, 24f\n"
506 "ld1 { v14.d }[0], [x20], #0x8\n"
507 "tbz %x[n_channels], #0, 25f\n"
508 "ld1 { v14.s }[2], [x20], #0x4\n"
511 "ld1 { v14.s }[0], [x20], #0x4\n"
513 "ldr x20, [x13, #0x90]\n"
514 "fmla v31.4s, v5.4s, v14.4s\n"
515 "add x20, x20, x28\n"
516 "tbz %x[n_channels], #1, 26f\n"
517 "ld1 { v15.d }[0], [x20], #0x8\n"
518 "tbz %x[n_channels], #0, 27f\n"
519 "ld1 { v15.s }[2], [x20], #0x4\n"
522 "ld1 { v15.s }[0], [x20], #0x4\n"
524 "ldr x20, [x13, #0x98]\n"
525 "fmla v30.4s, v6.4s, v15.4s\n"
526 "add x20, x20, x28\n"
527 "tbz %x[n_channels], #1, 28f\n"
528 "ld1 { v11.d }[0], [x20], #0x8\n"
529 "tbz %x[n_channels], #0, 29f\n"
530 "ld1 { v11.s }[2], [x20], #0x4\n"
533 "ld1 { v11.s }[0], [x20], #0x4\n"
535 "ldr x20, [x13, #0xa0]\n"
536 "fmla v29.4s, v8.4s, v11.4s\n"
537 "fmla v31.4s, v2.4s, v11.4s\n"
538 "add x20, x20, x28\n"
539 "tbz %x[n_channels], #1, 30f\n"
540 "ld1 { v13.d }[0], [x20], #0x8\n"
541 "tbz %x[n_channels], #0, 31f\n"
542 "ld1 { v13.s }[2], [x20], #0x4\n"
545 "ld1 { v13.s }[0], [x20], #0x4\n"
547 "ldr x20, [x13, #0xa8]\n"
548 "fmla v30.4s, v7.4s, v13.4s\n"
549 "add x20, x20, x28\n"
550 "tbz %x[n_channels], #1, 32f\n"
551 "ld1 { v16.d }[0], [x20], #0x8\n"
552 "tbz %x[n_channels], #0, 33f\n"
553 "ld1 { v16.s }[2], [x20], #0x4\n"
556 "ld1 { v16.s }[0], [x20], #0x4\n"
558 "ldr x20, [x13, #0xb0]\n"
559 "fmla v30.4s, v5.4s, v16.4s\n"
560 "fmla v31.4s, v3.4s, v16.4s\n"
561 "add x20, x20, x28\n"
562 "tbz %x[n_channels], #1, 34f\n"
563 "ld1 { v14.d }[0], [x20], #0x8\n"
564 "tbz %x[n_channels], #0, 35f\n"
565 "ld1 { v14.s }[2], [x20], #0x4\n"
568 "ld1 { v14.s }[0], [x20], #0x4\n"
570 "ldr x20, [x13, #0xb8]\n"
571 "fmla v31.4s, v7.4s, v14.4s\n"
572 "add x20, x20, x28\n"
573 "tbz %x[n_channels], #1, 36f\n"
574 "ld1 { v15.d }[0], [x20], #0x8\n"
575 "tbz %x[n_channels], #0, 37f\n"
576 "ld1 { v15.s }[2], [x20], #0x4\n"
579 "ld1 { v15.s }[0], [x20], #0x4\n"
581 "ldr x20, [x13, #0xc0]\n"
582 "fmla v30.4s, v8.4s, v15.4s\n"
583 "fmla v31.4s, v6.4s, v15.4s\n"
584 "add x20, x20, x28\n"
585 "tbz %x[n_channels], #1, 38f\n"
586 "ld1 { v11.d }[0], [x20], #0x8\n"
587 "tbz %x[n_channels], #0, 39f\n"
588 "ld1 { v11.s }[2], [x20], #0x4\n"
591 "ld1 { v11.s }[0], [x20], #0x4\n"
593 "fmla v31.4s, v8.4s, v11.4s\n"
594 "fmax v28.4s, v28.4s, v26.4s\n"
595 "fmax v29.4s, v29.4s, v26.4s\n"
596 "fmax v30.4s, v30.4s, v26.4s\n"
597 "fmax v31.4s, v31.4s, v26.4s\n"
598 "fmin v28.4s, v28.4s, v27.4s\n"
599 "fmin v29.4s, v29.4s, v27.4s\n"
600 "fmin v30.4s, v30.4s, v27.4s\n"
601 "fmin v31.4s, v31.4s, v27.4s\n"
602 "tbz %x[n_channels], #1, 40f\n"
603 "st1 { v28.d }[0], [x12], #0x8\n"
604 "st1 { v29.d }[0], [x11], #0x8\n"
605 "st1 { v30.d }[0], [x10], #0x8\n"
606 "st1 { v31.d }[0], [x9], #0x8\n"
607 "tbz %x[n_channels], #0, 41f\n"
608 "st1 { v28.s }[2], [x12], #0x4\n"
609 "st1 { v29.s }[2], [x11], #0x4\n"
610 "st1 { v30.s }[2], [x10], #0x4\n"
611 "st1 { v31.s }[2], [x9], #0x4\n"
614 "st1 { v28.s }[0], [x12], #0x4\n"
615 "st1 { v29.s }[0], [x11], #0x4\n"
616 "st1 { v30.s }[0], [x10], #0x4\n"
617 "st1 { v31.s }[0], [x9], #0x4\n"
621 : [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)
622 :
"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",
"x9",
"x10",
"x11",
"x12",
"x13",
"x20",
"x21",
"x22",
"x23",
"x24",
"x25",
"x26",
"x27",
"x28"
629 #endif // defined(__aarch64__)