28 #if defined(ARM_COMPUTE_ENABLE_SME2)
33 void sme2_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,
90 "ldr x8, [%x[params_struct], %[offsetof_args_params]]\n"
91 ".inst 0xd503477f // SMSTART ZA\n"
92 "add x17, %x[params_struct], %[offsetof_Args_inptrs]\n"
94 ".inst 0x25207810 // ptrue pn8.b\n"
95 "ld1w { z20.s }, p3/Z, [x8]\n"
97 "ldp x24, x23, [x17, #0x0]\n"
98 "ldp x22, x21, [x17, #0x10]\n"
100 ".inst 0xa040c100 // ld1w { z0.s-z3.s }, pn8.b/Z, [x8]\n"
102 "ldr x20, [x17, #0x20]\n"
104 "whilelt p2.s, XZR, %x[n_channels]\n"
105 ".inst 0xa040c104 // ld1w { z4.s-z7.s }, pn8.b/Z, [x8]\n"
107 "cmp x16, %x[n_channels]\n"
108 "ldr x14, [%x[params_struct], %[offsetof_args_outptrs]]\n"
109 "ld1rw { z22.s }, p3/Z, [%x[params_struct], %[offsetof_args_min]]\n"
110 "ld1rw { z14.s }, p3/Z, [%x[params_struct], %[offsetof_args_max]]\n"
111 "sub x13, XZR, x16\n"
112 "ld1w { z8.s }, p3/Z, [x8]\n"
114 "ld1w { z9.s }, p2/Z, [x24, x15, LSL #2]\n"
115 "ld1w { z10.s }, p2/Z, [x23, x15, LSL #2]\n"
116 "ld1w { z11.s }, p2/Z, [x22, x15, LSL #2]\n"
117 "ld1w { z12.s }, p2/Z, [x21, x15, LSL #2]\n"
118 "ld1w { z13.s }, p2/Z, [x20, x15, LSL #2]\n"
121 "movprfx z21, z20\n fmla z21.s, p3/M, z8.s, z9.s\n"
122 "movprfx z24, z20\n fmla z24.s, p3/M, z7.s, z9.s\n"
123 "ldr x22, [x17, #0x30]\n"
125 "movprfx z25, z20\n fmla z25.s, p3/M, z6.s, z9.s\n"
126 "fmla z21.s, p3/M, z0.s, z10.s\n"
127 "ldr x25, [x17, #0x38]\n"
129 "fmla z24.s, p3/M, z4.s, z13.s\n"
130 "movprfx z26, z20\n fmla z26.s, p3/M, z5.s, z9.s\n"
131 "ldr x21, [x17, #0x28]\n"
132 "whilelt p0.s, x16, %x[n_channels]\n"
133 "movprfx z27, z20\n fmla z27.s, p3/M, z4.s, z9.s\n"
134 "movprfx z28, z20\n fmla z28.s, p3/M, z3.s, z9.s\n"
135 "ldr x20, [x17, #0x48]\n"
136 "ld1w { z19.s }, p2/Z, [x20, x15, LSL #2]\n"
137 "fmla z25.s, p3/M, z2.s, z11.s\n"
138 "ld1w { z23.s }, p2/Z, [x22, x15, LSL #2]\n"
139 "movprfx z29, z20\n fmla z29.s, p3/M, z2.s, z9.s\n"
140 "ldr x20, [x17, #0x40]\n"
141 "fmla z21.s, p3/M, z5.s, z13.s\n"
142 "fmla z24.s, p3/M, z6.s, z23.s\n"
143 "ldr x24, [x17, #0x50]\n"
144 "movprfx z31, z20\n fmla z31.s, p3/M, z0.s, z9.s\n"
145 "fmla z25.s, p3/M, z3.s, z13.s\n"
146 "ldr x23, [x17, #0x58]\n"
147 "fmla z26.s, p3/M, z2.s, z13.s\n"
148 "fmla z27.s, p3/M, z1.s, z13.s\n"
149 "ldr x22, [x17, #0x60]\n"
150 "fmla z28.s, p3/M, z0.s, z13.s\n"
151 "ld1w { z17.s }, p2/Z, [x25, x15, LSL #2]\n"
152 "fmla z29.s, p3/M, z6.s, z12.s\n"
153 "ldr x12, [x17, #0x70]\n"
154 "ld1w { z16.s }, p2/Z, [x21, x15, LSL #2]\n"
155 "movprfx z30, z20\n fmla z30.s, p3/M, z1.s, z9.s\n"
156 "fmla z21.s, p3/M, z7.s, z23.s\n"
157 "ldr x21, [x17, #0x68]\n"
158 "fmla z24.s, p3/M, z0.s, z17.s\n"
159 "fmla z31.s, p3/M, z8.s, z16.s\n"
160 "ld1w { z16.s }, p2/Z, [x20, x15, LSL #2]\n"
161 "ldr x27, [x17, #0x78]\n"
162 "fmla z26.s, p3/M, z4.s, z23.s\n"
163 "fmla z27.s, p3/M, z3.s, z23.s\n"
164 "ldr x20, [x17, #0x80]\n"
165 "ld1w { z20.s }, p3/Z, [x8]\n"
166 "fmla z30.s, p3/M, z0.s, z23.s\n"
167 "fmla z28.s, p3/M, z4.s, z19.s\n"
168 "ldr x11, [x17, #0x88]\n"
170 "fmla z29.s, p3/M, z1.s, z23.s\n"
171 "fmla z21.s, p3/M, z1.s, z17.s\n"
172 "ld1w { z18.s }, p2/Z, [x24, x15, LSL #2]\n"
173 "ldr x26, [x17, #0x90]\n"
174 "fmla z24.s, p3/M, z2.s, z16.s\n"
175 "fmla z25.s, p3/M, z1.s, z16.s\n"
176 "ld1w { z11.s }, p2/Z, [x23, x15, LSL #2]\n"
177 "ldr x25, [x17, #0x98]\n"
178 "ld1w { z17.s }, p2/Z, [x22, x15, LSL #2]\n"
179 "fmla z27.s, p3/M, z5.s, z19.s\n"
180 "fmla z30.s, p3/M, z2.s, z19.s\n"
181 "ldr x24, [x17, #0xa0]\n"
182 "fmla z26.s, p3/M, z0.s, z18.s\n"
183 "fmla z28.s, p3/M, z2.s, z11.s\n"
184 "ldr x10, [x14, #0x0]\n"
185 "fmla z24.s, p3/M, z8.s, z19.s\n"
186 "fmla z25.s, p3/M, z7.s, z19.s\n"
187 "ldr x9, [x14, #0x8]\n"
188 "fmla z31.s, p3/M, z1.s, z19.s\n"
189 "fmla z29.s, p3/M, z3.s, z17.s\n"
190 "ld1w { z16.s }, p2/Z, [x21, x15, LSL #2]\n"
191 "ldr x23, [x17, #0xa8]\n"
192 "fmla z26.s, p3/M, z6.s, z17.s\n"
193 "fmla z27.s, p3/M, z7.s, z16.s\n"
194 "ld1w { z23.s }, p2/Z, [x20, x15, LSL #2]\n"
195 "ldr x22, [x17, #0xc0]\n"
196 "fmla z28.s, p3/M, z6.s, z16.s\n"
197 "fmla z30.s, p3/M, z4.s, z16.s\n"
198 "ldr x28, [x14, #0x10]\n"
199 "fmla z21.s, p3/M, z3.s, z18.s\n"
200 "fmla z25.s, p3/M, z5.s, z11.s\n"
201 "ld1w { z15.s }, p2/Z, [x12, x15, LSL #2]\n"
202 "ldr x21, [x17, #0xb0]\n"
203 "fmla z29.s, p3/M, z5.s, z16.s\n"
204 "fmla z31.s, p3/M, z3.s, z16.s\n"
205 "ld1w { z19.s }, p2/Z, [x27, x15, LSL #2]\n"
206 "ldr x20, [x17, #0xb8]\n"
207 "fmla z26.s, p3/M, z8.s, z16.s\n"
208 "fmla z28.s, p3/M, z8.s, z15.s\n"
209 "ldr x27, [x14, #0x18]\n"
210 "fmla z30.s, p3/M, z6.s, z19.s\n"
211 "fmla z24.s, p3/M, z3.s, z23.s\n"
212 "fmla z27.s, p3/M, z0.s, z23.s\n"
213 "fmla z31.s, p3/M, z5.s, z15.s\n"
214 "ld1w { z17.s }, p2/Z, [x11, x15, LSL #2]\n"
215 "fmla z29.s, p3/M, z7.s, z19.s\n"
216 "ld1w { z19.s }, p2/Z, [x26, x15, LSL #2]\n"
217 "fmla z21.s, p3/M, z4.s, z23.s\n"
218 "fmla z26.s, p3/M, z1.s, z23.s\n"
219 "fmla z24.s, p3/M, z5.s, z17.s\n"
220 "ld1w { z16.s }, p2/Z, [x25, x15, LSL #2]\n"
221 "fmla z25.s, p3/M, z4.s, z17.s\n"
222 "fmla z27.s, p3/M, z2.s, z17.s\n"
223 "fmla z28.s, p3/M, z1.s, z17.s\n"
224 "fmla z30.s, p3/M, z8.s, z19.s\n"
225 "ld1w { z17.s }, p2/Z, [x24, x15, LSL #2]\n"
226 "ldr x26, [x17, #0x20]\n"
227 "fmla z21.s, p3/M, z2.s, z17.s\n"
228 "fmla z26.s, p3/M, z7.s, z16.s\n"
229 "fmla z27.s, p3/M, z6.s, z16.s\n"
230 "fmla z29.s, p3/M, z4.s, z16.s\n"
231 "fmla z30.s, p3/M, z3.s, z16.s\n"
232 "ld1w { z18.s }, p2/Z, [x21, x15, LSL #2]\n"
233 "fmla z31.s, p3/M, z7.s, z19.s\n"
234 "ld1w { z16.s }, p2/Z, [x23, x15, LSL #2]\n"
235 "fmla z21.s, p3/M, z6.s, z18.s\n"
236 "fmla z31.s, p3/M, z4.s, z16.s\n"
237 "fmla z24.s, p3/M, z1.s, z17.s\n"
238 "fmla z25.s, p3/M, z0.s, z17.s\n"
239 "ld1w { z17.s }, p2/Z, [x20, x15, LSL #2]\n"
240 "fmax z21.s, p3/M, z21.s, z22.s\n"
241 "fmla z28.s, p3/M, z7.s, z16.s\n"
242 "fmla z30.s, p3/M, z5.s, z16.s\n"
243 "fmla z29.s, p3/M, z0.s, z18.s\n"
244 "fmla z31.s, p3/M, z2.s, z17.s\n"
245 "fmla z27.s, p3/M, z8.s, z16.s\n"
246 "ld1w { z16.s }, p2/Z, [x22, x15, LSL #2]\n"
247 "ldp x22, x21, [x17, #0x0]\n"
248 "fmla z26.s, p3/M, z3.s, z18.s\n"
249 "fmla z25.s, p3/M, z8.s, z17.s\n"
250 "ldp x25, x24, [x17, #0x10]\n"
252 "fmin z21.s, p3/M, z21.s, z14.s\n"
253 "st1w { z21.s }, p1, [x10, x13, LSL #2]\n"
254 "ldr x20, [x14, #0x20]\n"
255 "fmla z28.s, p3/M, z5.s, z17.s\n"
256 "fmla z29.s, p3/M, z8.s, z16.s\n"
257 "fmla z30.s, p3/M, z7.s, z16.s\n"
258 "ld1w { z9.s }, p0/Z, [x22, x16, LSL #2]\n"
259 "whilelt p2.s, x15, %x[n_channels]\n"
260 "fmla z31.s, p3/M, z6.s, z16.s\n"
261 ".inst 0xc1aecad8 // fclamp { z24.s-z27.s }, z22.s, z14.s\n"
262 "st1w { z24.s }, p1, [x9, x13, LSL #2]\n"
263 "ldr x23, [x14, #0x28]\n"
264 "st1w { z25.s }, p1, [x28, x13, LSL #2]\n"
265 "ldr x22, [x14, #0x30]\n"
266 "ld1w { z10.s }, p0/Z, [x21, x16, LSL #2]\n"
267 ".inst 0xc1aecadc // fclamp { z28.s-z31.s }, z22.s, z14.s\n"
268 "st1w { z26.s }, p1, [x27, x13, LSL #2]\n"
269 "ldr x21, [x14, #0x38]\n"
270 "ld1w { z11.s }, p0/Z, [x25, x16, LSL #2]\n"
271 "st1w { z27.s }, p1, [x20, x13, LSL #2]\n"
272 "ldr x20, [x14, #0x40]\n"
273 "ld1w { z12.s }, p0/Z, [x24, x16, LSL #2]\n"
274 "ld1w { z13.s }, p0/Z, [x26, x16, LSL #2]\n"
276 "cmp x16, %x[n_channels]\n"
277 "st1w { z28.s }, p1, [x23, x13, LSL #2]\n"
278 ".inst 0xa040c100 // ld1w { z0.s-z3.s }, pn8.b/Z, [x8]\n"
280 "st1w { z29.s }, p1, [x22, x13, LSL #2]\n"
281 ".inst 0xa040c104 // ld1w { z4.s-z7.s }, pn8.b/Z, [x8]\n"
283 "st1w { z30.s }, p1, [x21, x13, LSL #2]\n"
284 "st1w { z31.s }, p1, [x20, x13, LSL #2]\n"
285 "ld1w { z8.s }, p3/Z, [x8]\n"
289 "movprfx z21, z20\n fmla z21.s, p3/M, z8.s, z9.s\n"
290 "movprfx z24, z20\n fmla z24.s, p3/M, z7.s, z9.s\n"
291 "ldr x23, [x17, #0x30]\n"
293 "movprfx z25, z20\n fmla z25.s, p3/M, z6.s, z9.s\n"
294 "fmla z21.s, p3/M, z0.s, z10.s\n"
295 "ldr x22, [x17, #0x38]\n"
297 "fmla z24.s, p3/M, z4.s, z13.s\n"
298 "movprfx z26, z20\n fmla z26.s, p3/M, z5.s, z9.s\n"
299 "ldr x21, [x17, #0x28]\n"
300 "movprfx z27, z20\n fmla z27.s, p3/M, z4.s, z9.s\n"
301 "movprfx z28, z20\n fmla z28.s, p3/M, z3.s, z9.s\n"
302 "ldr x20, [x17, #0x48]\n"
303 "ld1w { z19.s }, p2/Z, [x20, x15, LSL #2]\n"
304 "fmla z25.s, p3/M, z2.s, z11.s\n"
305 "ld1w { z18.s }, p2/Z, [x23, x15, LSL #2]\n"
306 "movprfx z29, z20\n fmla z29.s, p3/M, z2.s, z9.s\n"
307 "ldr x20, [x17, #0x40]\n"
308 "fmla z21.s, p3/M, z5.s, z13.s\n"
309 "fmla z24.s, p3/M, z6.s, z18.s\n"
310 "ldr x25, [x17, #0x50]\n"
311 "movprfx z31, z20\n fmla z31.s, p3/M, z0.s, z9.s\n"
312 "fmla z25.s, p3/M, z3.s, z13.s\n"
313 "ldr x24, [x17, #0x58]\n"
314 "fmla z26.s, p3/M, z2.s, z13.s\n"
315 "fmla z27.s, p3/M, z1.s, z13.s\n"
316 "ldr x23, [x17, #0x60]\n"
317 "fmla z28.s, p3/M, z0.s, z13.s\n"
318 "ld1w { z17.s }, p2/Z, [x22, x15, LSL #2]\n"
319 "fmla z29.s, p3/M, z6.s, z12.s\n"
320 "ldr x12, [x17, #0x70]\n"
321 "ld1w { z16.s }, p2/Z, [x21, x15, LSL #2]\n"
322 "movprfx z30, z20\n fmla z30.s, p3/M, z1.s, z9.s\n"
323 "fmla z21.s, p3/M, z7.s, z18.s\n"
324 "ldr x22, [x17, #0x68]\n"
325 "fmla z24.s, p3/M, z0.s, z17.s\n"
326 "fmla z31.s, p3/M, z8.s, z16.s\n"
327 "ld1w { z16.s }, p2/Z, [x20, x15, LSL #2]\n"
328 "ldr x21, [x17, #0x78]\n"
329 "fmla z26.s, p3/M, z4.s, z18.s\n"
330 "fmla z27.s, p3/M, z3.s, z18.s\n"
331 "ldr x20, [x17, #0x80]\n"
332 "fmla z30.s, p3/M, z0.s, z18.s\n"
333 "fmla z28.s, p3/M, z4.s, z19.s\n"
334 "ldr x11, [x17, #0x88]\n"
335 "fmla z29.s, p3/M, z1.s, z18.s\n"
336 "fmla z21.s, p3/M, z1.s, z17.s\n"
337 "ld1w { z20.s }, p2/Z, [x25, x15, LSL #2]\n"
338 "ldr x10, [x17, #0x90]\n"
339 "fmla z24.s, p3/M, z2.s, z16.s\n"
340 "fmla z25.s, p3/M, z1.s, z16.s\n"
341 "ld1w { z17.s }, p2/Z, [x24, x15, LSL #2]\n"
342 "ldr x9, [x17, #0x98]\n"
343 "ld1w { z16.s }, p2/Z, [x23, x15, LSL #2]\n"
344 "fmla z27.s, p3/M, z5.s, z19.s\n"
345 "fmla z30.s, p3/M, z2.s, z19.s\n"
346 "ldr x28, [x17, #0xa0]\n"
347 "fmla z26.s, p3/M, z0.s, z20.s\n"
348 "fmla z28.s, p3/M, z2.s, z17.s\n"
349 "ldr x27, [x14, #0x0]\n"
350 "fmla z24.s, p3/M, z8.s, z19.s\n"
351 "fmla z25.s, p3/M, z7.s, z19.s\n"
352 "ldr x26, [x14, #0x8]\n"
353 "fmla z31.s, p3/M, z1.s, z19.s\n"
354 "fmla z29.s, p3/M, z3.s, z16.s\n"
355 "ld1w { z19.s }, p2/Z, [x22, x15, LSL #2]\n"
356 "ldr x25, [x17, #0xa8]\n"
357 "fmla z26.s, p3/M, z6.s, z16.s\n"
358 "fmla z27.s, p3/M, z7.s, z19.s\n"
359 "ld1w { z18.s }, p2/Z, [x20, x15, LSL #2]\n"
360 "ldr x23, [x17, #0xc0]\n"
361 "fmla z28.s, p3/M, z6.s, z19.s\n"
362 "fmla z30.s, p3/M, z4.s, z19.s\n"
363 "ldr x24, [x14, #0x10]\n"
364 "fmla z21.s, p3/M, z3.s, z20.s\n"
365 "fmla z25.s, p3/M, z5.s, z17.s\n"
366 "ld1w { z17.s }, p2/Z, [x12, x15, LSL #2]\n"
367 "ldr x22, [x17, #0xb0]\n"
368 "fmla z29.s, p3/M, z5.s, z19.s\n"
369 "fmla z31.s, p3/M, z3.s, z19.s\n"
370 "ld1w { z16.s }, p2/Z, [x21, x15, LSL #2]\n"
371 "ldr x20, [x17, #0xb8]\n"
372 "fmla z26.s, p3/M, z8.s, z19.s\n"
373 "fmla z28.s, p3/M, z8.s, z17.s\n"
374 "ldr x21, [x14, #0x18]\n"
375 "fmla z30.s, p3/M, z6.s, z16.s\n"
376 "fmla z24.s, p3/M, z3.s, z18.s\n"
377 "fmla z27.s, p3/M, z0.s, z18.s\n"
378 "fmla z31.s, p3/M, z5.s, z17.s\n"
379 "ld1w { z17.s }, p2/Z, [x11, x15, LSL #2]\n"
380 "fmla z29.s, p3/M, z7.s, z16.s\n"
381 "ld1w { z19.s }, p2/Z, [x10, x15, LSL #2]\n"
382 "fmla z21.s, p3/M, z4.s, z18.s\n"
383 "fmla z26.s, p3/M, z1.s, z18.s\n"
384 "fmla z24.s, p3/M, z5.s, z17.s\n"
385 "ld1w { z16.s }, p2/Z, [x9, x15, LSL #2]\n"
386 "fmla z25.s, p3/M, z4.s, z17.s\n"
387 "fmla z27.s, p3/M, z2.s, z17.s\n"
388 "fmla z28.s, p3/M, z1.s, z17.s\n"
389 "fmla z30.s, p3/M, z8.s, z19.s\n"
390 "ld1w { z17.s }, p2/Z, [x28, x15, LSL #2]\n"
391 "fmla z21.s, p3/M, z2.s, z17.s\n"
392 "fmla z26.s, p3/M, z7.s, z16.s\n"
393 "fmla z27.s, p3/M, z6.s, z16.s\n"
394 "fmla z29.s, p3/M, z4.s, z16.s\n"
395 "fmla z30.s, p3/M, z3.s, z16.s\n"
396 "ld1w { z18.s }, p2/Z, [x22, x15, LSL #2]\n"
397 "fmla z31.s, p3/M, z7.s, z19.s\n"
398 "ld1w { z16.s }, p2/Z, [x25, x15, LSL #2]\n"
399 "fmla z21.s, p3/M, z6.s, z18.s\n"
400 "fmla z31.s, p3/M, z4.s, z16.s\n"
401 "fmla z24.s, p3/M, z1.s, z17.s\n"
402 "fmla z25.s, p3/M, z0.s, z17.s\n"
403 "ld1w { z17.s }, p2/Z, [x20, x15, LSL #2]\n"
404 "fmax z21.s, p3/M, z21.s, z22.s\n"
405 "fmla z28.s, p3/M, z7.s, z16.s\n"
406 "fmla z30.s, p3/M, z5.s, z16.s\n"
407 "fmla z29.s, p3/M, z0.s, z18.s\n"
408 "fmla z31.s, p3/M, z2.s, z17.s\n"
409 "fmla z27.s, p3/M, z8.s, z16.s\n"
410 "ld1w { z16.s }, p2/Z, [x23, x15, LSL #2]\n"
411 "fmla z26.s, p3/M, z3.s, z18.s\n"
412 "fmla z25.s, p3/M, z8.s, z17.s\n"
413 "fmin z21.s, p3/M, z21.s, z14.s\n"
414 "st1w { z21.s }, p0, [x27, x13, LSL #2]\n"
415 "ldr x20, [x14, #0x20]\n"
416 "fmla z28.s, p3/M, z5.s, z17.s\n"
417 "fmla z29.s, p3/M, z8.s, z16.s\n"
418 "fmla z30.s, p3/M, z7.s, z16.s\n"
419 "fmla z31.s, p3/M, z6.s, z16.s\n"
420 ".inst 0xc1aecad8 // fclamp { z24.s-z27.s }, z22.s, z14.s\n"
421 "st1w { z24.s }, p0, [x26, x13, LSL #2]\n"
422 "ldr x23, [x14, #0x28]\n"
423 "st1w { z25.s }, p0, [x24, x13, LSL #2]\n"
424 "ldr x22, [x14, #0x30]\n"
425 ".inst 0xc1aecadc // fclamp { z28.s-z31.s }, z22.s, z14.s\n"
426 "st1w { z26.s }, p0, [x21, x13, LSL #2]\n"
427 "ldr x21, [x14, #0x38]\n"
428 "st1w { z27.s }, p0, [x20, x13, LSL #2]\n"
429 "ldr x20, [x14, #0x40]\n"
430 "st1w { z28.s }, p0, [x23, x13, LSL #2]\n"
431 "st1w { z29.s }, p0, [x22, x13, LSL #2]\n"
432 "st1w { z30.s }, p0, [x21, x13, LSL #2]\n"
433 "st1w { z31.s }, p0, [x20, x13, LSL #2]\n"
434 ".inst 0xd503467f // SMSTOP\n"
436 : [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)
437 :
"cc",
"memory",
"p0",
"p1",
"p2",
"p3",
"p4",
"p5",
"p6",
"p7",
"p8",
"p9",
"p10",
"p11",
"p12",
"p13",
"p14",
"p15",
"x8",
"x9",
"x10",
"x11",
"x12",
"x13",
"x14",
"x15",
"x16",
"x17",
"x20",
"x21",
"x22",
"x23",
"x24",
"x25",
"x26",
"x27",
"x28",
"z0",
"z1",
"z2",
"z3",
"z4",
"z5",
"z6",
"z7",
"z8",
"z9",
"z10",
"z11",
"z12",
"z13",
"z14",
"z15",
"z16",
"z17",
"z18",
"z19",
"z20",
"z21",
"z22",
"z23",
"z24",
"z25",
"z26",
"z27",
"z28",
"z29",
"z30",
"z31"
444 #endif // defined(ARM_COMPUTE_ENABLE_SME2)