28 #if defined(ARM_COMPUTE_ENABLE_SME2)
33 void sme2_fp32_nhwc_3x3_s1_output4x4_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[36];
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[14];
58 inptrs[1] = input_ptrs[0];
59 inptrs[2] = input_ptrs[5];
60 inptrs[3] = input_ptrs[15];
61 inptrs[4] = input_ptrs[30];
62 inptrs[5] = input_ptrs[35];
63 inptrs[6] = input_ptrs[20];
64 inptrs[7] = input_ptrs[1];
65 inptrs[8] = input_ptrs[4];
66 inptrs[9] = input_ptrs[21];
67 inptrs[10] = input_ptrs[6];
68 inptrs[11] = input_ptrs[11];
69 inptrs[12] = input_ptrs[24];
70 inptrs[13] = input_ptrs[8];
71 inptrs[14] = input_ptrs[29];
72 inptrs[15] = input_ptrs[9];
73 inptrs[16] = input_ptrs[31];
74 inptrs[17] = input_ptrs[13];
75 inptrs[18] = input_ptrs[34];
76 inptrs[19] = input_ptrs[16];
77 inptrs[20] = input_ptrs[2];
78 inptrs[21] = input_ptrs[19];
79 inptrs[22] = input_ptrs[3];
80 inptrs[23] = input_ptrs[12];
81 inptrs[24] = input_ptrs[22];
82 inptrs[25] = input_ptrs[17];
83 inptrs[26] = input_ptrs[18];
84 inptrs[27] = input_ptrs[26];
85 inptrs[28] = input_ptrs[23];
86 inptrs[29] = input_ptrs[32];
87 inptrs[30] = input_ptrs[27];
88 inptrs[31] = input_ptrs[33];
89 inptrs[32] = input_ptrs[7];
90 inptrs[33] = input_ptrs[10];
91 inptrs[34] = input_ptrs[25];
92 inptrs[35] = input_ptrs[28];
97 Args params_struct(input_ptrs, outptrs, params,
100 __asm__ __volatile__(
101 "ldr x8, [%x[params_struct], %[offsetof_args_params]]\n"
102 ".inst 0xd503477f // SMSTART ZA\n"
103 "add x17, %x[params_struct], %[offsetof_Args_inptrs]\n"
105 ".inst 0x25207810 // ptrue pn8.b\n"
106 "ld1w { z13.s }, p3/Z, [x8]\n"
108 "ldp x23, x22, [x17, #0x0]\n"
109 "ldp x21, x20, [x17, #0x10]\n"
111 ".inst 0xa040c100 // ld1w { z0.s-z3.s }, pn8.b/Z, [x8]\n"
114 "whilelt p2.s, XZR, %x[n_channels]\n"
115 ".inst 0xa040c104 // ld1w { z4.s-z7.s }, pn8.b/Z, [x8]\n"
116 "ldr x14, [%x[params_struct], %[offsetof_args_outptrs]]\n"
118 "cmp x16, %x[n_channels]\n"
119 "ld1rw { z14.s }, p3/Z, [%x[params_struct], %[offsetof_args_min]]\n"
120 "ld1rw { z15.s }, p3/Z, [%x[params_struct], %[offsetof_args_max]]\n"
121 "sub x13, XZR, x16\n"
122 "ld1w { z8.s }, p3/Z, [x8]\n"
124 "ld1w { z9.s }, p2/Z, [x23, x15, LSL #2]\n"
125 "ld1w { z10.s }, p2/Z, [x22, x15, LSL #2]\n"
126 "ld1w { z11.s }, p2/Z, [x21, x15, LSL #2]\n"
127 "ld1w { z12.s }, p2/Z, [x20, x15, LSL #2]\n"
130 "movprfx z29, z13\n fmla z29.s, p3/M, z4.s, z9.s\n"
131 "movprfx z16, z13\n fmla z16.s, p3/M, z8.s, z9.s\n"
132 "ldr x24, [x17, #0x20]\n"
134 "movprfx z30, z13\n fmla z30.s, p3/M, z3.s, z9.s\n"
135 "movprfx z25, z13\n fmla z25.s, p3/M, z1.s, z9.s\n"
136 "ldr x20, [x17, #0x30]\n"
138 "movprfx z26, z13\n fmla z26.s, p3/M, z0.s, z9.s\n"
139 "ldr x21, [x17, #0x28]\n"
140 "movprfx z17, z13\n fmla z17.s, p3/M, z7.s, z9.s\n"
141 "whilelt p0.s, x16, %x[n_channels]\n"
142 "movprfx z18, z13\n fmla z18.s, p3/M, z6.s, z9.s\n"
143 "fmla z29.s, p3/M, z5.s, z12.s\n"
144 "ldr x23, [x17, #0x38]\n"
145 "movprfx z28, z13\n fmla z28.s, p3/M, z5.s, z9.s\n"
146 "movprfx z24, z13\n fmla z24.s, p3/M, z2.s, z9.s\n"
147 "ld1w { z9.s }, p2/Z, [x20, x15, LSL #2]\n"
148 "ldr x22, [x17, #0x40]\n"
149 "fmla z16.s, p3/M, z0.s, z10.s\n"
150 "movprfx z19, z13\n fmla z19.s, p3/M, z2.s, z11.s\n"
151 "ld1w { z22.s }, p2/Z, [x24, x15, LSL #2]\n"
152 "ldr x20, [x17, #0x48]\n"
153 "fmla z30.s, p3/M, z4.s, z12.s\n"
154 "fmla z25.s, p3/M, z2.s, z12.s\n"
155 "ld1w { z21.s }, p2/Z, [x21, x15, LSL #2]\n"
156 "ldr x27, [x17, #0x50]\n"
157 "fmla z26.s, p3/M, z1.s, z12.s\n"
158 "fmla z17.s, p3/M, z8.s, z12.s\n"
159 "ldr x26, [x17, #0x60]\n"
160 "fmla z18.s, p3/M, z7.s, z12.s\n"
161 "movprfx z20, z13\n fmla z20.s, p3/M, z6.s, z22.s\n"
162 "ld1w { z11.s }, p2/Z, [x20, x15, LSL #2]\n"
163 "ldr x25, [x17, #0x68]\n"
164 "fmla z29.s, p3/M, z7.s, z9.s\n"
165 "fmla z19.s, p3/M, z6.s, z12.s\n"
166 "ldr x21, [x17, #0x58]\n"
167 "movprfx z31, z13\n fmla z31.s, p3/M, z3.s, z12.s\n"
168 "movprfx z27, z13\n fmla z27.s, p3/M, z0.s, z12.s\n"
169 "ld1w { z10.s }, p2/Z, [x23, x15, LSL #2]\n"
170 "ldr x24, [x17, #0x70]\n"
171 "movprfx z23, z13\n fmla z23.s, p3/M, z8.s, z21.s\n"
172 "fmla z30.s, p3/M, z6.s, z9.s\n"
173 "ld1w { z12.s }, p2/Z, [x22, x15, LSL #2]\n"
174 "ldr x23, [x17, #0x78]\n"
175 "fmla z25.s, p3/M, z4.s, z9.s\n"
176 "fmla z26.s, p3/M, z3.s, z9.s\n"
177 "ldr x22, [x17, #0x80]\n"
178 "movprfx z21, z13\n fmla z21.s, p3/M, z1.s, z9.s\n"
179 "movprfx z22, z13\n fmla z22.s, p3/M, z0.s, z9.s\n"
180 "ldr x20, [x17, #0x88]\n"
181 "ld1w { z13.s }, p3/Z, [x8]\n"
182 "fmla z28.s, p3/M, z8.s, z9.s\n"
183 "fmla z24.s, p3/M, z5.s, z9.s\n"
184 "ldr x12, [x14, #0x0]\n"
186 "fmla z20.s, p3/M, z2.s, z9.s\n"
187 "fmla z16.s, p3/M, z1.s, z10.s\n"
188 "ld1w { z9.s }, p2/Z, [x27, x15, LSL #2]\n"
189 "ldr x27, [x17, #0x90]\n"
190 "fmla z17.s, p3/M, z0.s, z10.s\n"
191 "fmla z18.s, p3/M, z2.s, z12.s\n"
192 "ld1w { z10.s }, p2/Z, [x21, x15, LSL #2]\n"
193 "ldr x21, [x17, #0x98]\n"
194 "fmla z29.s, p3/M, z8.s, z11.s\n"
195 "fmla z19.s, p3/M, z1.s, z12.s\n"
196 "ld1w { z12.s }, p2/Z, [x26, x15, LSL #2]\n"
197 "ldr x26, [x17, #0xa0]\n"
198 "fmla z30.s, p3/M, z7.s, z11.s\n"
199 "fmla z31.s, p3/M, z6.s, z11.s\n"
200 "ldr x11, [x14, #0x8]\n"
201 "fmla z25.s, p3/M, z5.s, z11.s\n"
202 "fmla z26.s, p3/M, z4.s, z11.s\n"
203 "ldr x10, [x14, #0x10]\n"
204 "fmla z27.s, p3/M, z3.s, z11.s\n"
205 "fmla z21.s, p3/M, z2.s, z11.s\n"
206 "ldr x9, [x14, #0x18]\n"
207 "fmla z22.s, p3/M, z1.s, z11.s\n"
208 "fmla z23.s, p3/M, z0.s, z11.s\n"
209 "ld1w { z11.s }, p2/Z, [x25, x15, LSL #2]\n"
210 "ldr x25, [x17, #0xa8]\n"
211 "fmla z16.s, p3/M, z3.s, z9.s\n"
212 "fmla z28.s, p3/M, z0.s, z9.s\n"
213 "fmla z24.s, p3/M, z6.s, z12.s\n"
214 "fmla z20.s, p3/M, z3.s, z12.s\n"
215 "ld1w { z9.s }, p2/Z, [x24, x15, LSL #2]\n"
216 "ldr x24, [x17, #0xb0]\n"
217 "fmla z17.s, p3/M, z4.s, z11.s\n"
218 "fmla z18.s, p3/M, z3.s, z11.s\n"
219 "fmla z29.s, p3/M, z1.s, z11.s\n"
220 "fmla z19.s, p3/M, z5.s, z10.s\n"
221 "fmla z31.s, p3/M, z2.s, z10.s\n"
222 "fmla z30.s, p3/M, z0.s, z11.s\n"
223 "ld1w { z12.s }, p2/Z, [x23, x15, LSL #2]\n"
224 "ldr x23, [x17, #0xb8]\n"
225 "fmla z27.s, p3/M, z8.s, z9.s\n"
226 "fmla z23.s, p3/M, z5.s, z9.s\n"
227 "ld1w { z10.s }, p2/Z, [x22, x15, LSL #2]\n"
228 "ldr x22, [x17, #0xc0]\n"
229 "fmla z16.s, p3/M, z5.s, z11.s\n"
230 "fmla z28.s, p3/M, z2.s, z11.s\n"
231 "ld1w { z11.s }, p2/Z, [x20, x15, LSL #2]\n"
232 "ldr x20, [x17, #0xc8]\n"
233 "fmla z17.s, p3/M, z5.s, z12.s\n"
234 "fmla z18.s, p3/M, z4.s, z12.s\n"
235 "fmla z29.s, p3/M, z2.s, z12.s\n"
236 "fmla z19.s, p3/M, z3.s, z12.s\n"
237 "fmla z30.s, p3/M, z1.s, z12.s\n"
238 "fmla z31.s, p3/M, z0.s, z12.s\n"
239 "ld1w { z9.s }, p2/Z, [x21, x15, LSL #2]\n"
240 "ldr x28, [x17, #0xd8]\n"
241 "fmla z20.s, p3/M, z7.s, z10.s\n"
242 "fmla z21.s, p3/M, z6.s, z10.s\n"
243 "ld1w { z10.s }, p2/Z, [x27, x15, LSL #2]\n"
244 "ldr x21, [x17, #0xd0]\n"
245 "fmla z16.s, p3/M, z7.s, z11.s\n"
246 "fmla z17.s, p3/M, z6.s, z11.s\n"
247 "fmla z28.s, p3/M, z4.s, z11.s\n"
248 "fmla z29.s, p3/M, z3.s, z11.s\n"
249 "fmla z24.s, p3/M, z1.s, z11.s\n"
250 "fmla z25.s, p3/M, z0.s, z11.s\n"
251 "ld1w { z11.s }, p2/Z, [x26, x15, LSL #2]\n"
252 "ldr x27, [x17, #0xe0]\n"
253 "fmla z18.s, p3/M, z8.s, z9.s\n"
254 "fmla z22.s, p3/M, z8.s, z10.s\n"
255 "fmla z23.s, p3/M, z7.s, z10.s\n"
256 "ld1w { z10.s }, p2/Z, [x25, x15, LSL #2]\n"
257 "fmla z27.s, p3/M, z1.s, z9.s\n"
258 "ldr x26, [x17, #0xe8]\n"
259 "fmla z19.s, p3/M, z7.s, z9.s\n"
260 "fmla z30.s, p3/M, z5.s, z9.s\n"
261 "fmla z31.s, p3/M, z4.s, z9.s\n"
262 "fmla z26.s, p3/M, z2.s, z9.s\n"
263 "ld1w { z9.s }, p2/Z, [x24, x15, LSL #2]\n"
264 "ldr x25, [x17, #0xf0]\n"
265 "fmla z16.s, p3/M, z2.s, z11.s\n"
266 "fmla z17.s, p3/M, z1.s, z11.s\n"
267 "fmla z18.s, p3/M, z0.s, z11.s\n"
268 "fmla z28.s, p3/M, z7.s, z10.s\n"
269 "ld1w { z11.s }, p2/Z, [x23, x15, LSL #2]\n"
270 "ldr x24, [x17, #0xf8]\n"
271 "fmla z29.s, p3/M, z6.s, z10.s\n"
272 "fmla z24.s, p3/M, z4.s, z10.s\n"
273 "fmla z25.s, p3/M, z3.s, z10.s\n"
274 "fmla z20.s, p3/M, z1.s, z10.s\n"
275 "fmla z21.s, p3/M, z0.s, z10.s\n"
276 "ld1w { z10.s }, p2/Z, [x22, x15, LSL #2]\n"
277 "fmla z27.s, p3/M, z4.s, z10.s\n"
278 "ldr x23, [x17, #0x100]\n"
279 "fmla z22.s, p3/M, z2.s, z10.s\n"
280 "fmla z17.s, p3/M, z2.s, z9.s\n"
281 "fmla z18.s, p3/M, z1.s, z9.s\n"
282 "fmla z19.s, p3/M, z0.s, z9.s\n"
283 "ld1w { z9.s }, p2/Z, [x20, x15, LSL #2]\n"
284 "ldr x20, [x17, #0x108]\n"
285 "fmla z16.s, p3/M, z6.s, z11.s\n"
286 "fmla z28.s, p3/M, z3.s, z11.s\n"
287 "fmla z24.s, p3/M, z0.s, z11.s\n"
288 "fmla z30.s, p3/M, z8.s, z10.s\n"
289 "ld1w { z11.s }, p2/Z, [x21, x15, LSL #2]\n"
290 "ldr x22, [x17, #0x110]\n"
291 "fmla z31.s, p3/M, z7.s, z10.s\n"
292 "fmla z26.s, p3/M, z5.s, z10.s\n"
293 "fmla z23.s, p3/M, z1.s, z10.s\n"
294 "ld1w { z10.s }, p2/Z, [x28, x15, LSL #2]\n"
295 "fmla z27.s, p3/M, z2.s, z9.s\n"
296 "ldr x21, [x17, #0x118]\n"
297 "fmla z20.s, p3/M, z0.s, z11.s\n"
298 "fmla z21.s, p3/M, z4.s, z10.s\n"
299 "fmla z22.s, p3/M, z3.s, z10.s\n"
300 "fmla z19.s, p3/M, z8.s, z9.s\n"
301 "fmla z31.s, p3/M, z5.s, z9.s\n"
302 "fmla z28.s, p3/M, z6.s, z11.s\n"
303 "ld1w { z9.s }, p2/Z, [x27, x15, LSL #2]\n"
304 "fmla z24.s, p3/M, z3.s, z11.s\n"
305 "ld1w { z12.s }, p2/Z, [x26, x15, LSL #2]\n"
306 "fmla z25.s, p3/M, z7.s, z10.s\n"
307 "fmla z26.s, p3/M, z6.s, z10.s\n"
308 "fmla z20.s, p3/M, z5.s, z10.s\n"
309 "fmla z27.s, p3/M, z5.s, z9.s\n"
310 "fmla z23.s, p3/M, z2.s, z9.s\n"
311 "fmla z21.s, p3/M, z7.s, z12.s\n"
312 "fmla z22.s, p3/M, z6.s, z12.s\n"
313 "fmla z24.s, p3/M, z8.s, z10.s\n"
314 "ld1w { z11.s }, p2/Z, [x25, x15, LSL #2]\n"
315 "fmla z20.s, p3/M, z8.s, z12.s\n"
316 "fmla z25.s, p3/M, z8.s, z11.s\n"
317 "fmla z26.s, p3/M, z7.s, z11.s\n"
318 "ld1w { z10.s }, p2/Z, [x23, x15, LSL #2]\n"
319 "fmla z27.s, p3/M, z6.s, z11.s\n"
320 "fmla z21.s, p3/M, z5.s, z11.s\n"
321 "fmla z22.s, p3/M, z4.s, z11.s\n"
322 "fmla z23.s, p3/M, z3.s, z11.s\n"
323 "ld1w { z11.s }, p2/Z, [x20, x15, LSL #2]\n"
324 "ldp x20, x25, [x17, #0x0]\n"
325 "fmla z31.s, p3/M, z8.s, z9.s\n"
326 "ld1w { z12.s }, p2/Z, [x24, x15, LSL #2]\n"
327 "fmla z16.s, p3/M, z4.s, z10.s\n"
328 "fmla z17.s, p3/M, z3.s, z10.s\n"
329 "fmla z18.s, p3/M, z5.s, z11.s\n"
330 "ld1w { z9.s }, p0/Z, [x20, x16, LSL #2]\n"
331 "fmla z19.s, p3/M, z4.s, z11.s\n"
332 "fmla z21.s, p3/M, z8.s, z12.s\n"
333 "fmla z22.s, p3/M, z7.s, z12.s\n"
334 "fmla z23.s, p3/M, z6.s, z12.s\n"
335 "ld1w { z12.s }, p2/Z, [x22, x15, LSL #2]\n"
336 "fmla z28.s, p3/M, z1.s, z10.s\n"
337 "fmla z29.s, p3/M, z0.s, z10.s\n"
338 "ld1w { z0.s }, p2/Z, [x21, x15, LSL #2]\n"
339 "ldp x20, x24, [x17, #0x10]\n"
340 "fmla z30.s, p3/M, z2.s, z11.s\n"
341 "fmla z31.s, p3/M, z1.s, z11.s\n"
343 "ld1w { z11.s }, p0/Z, [x20, x16, LSL #2]\n"
344 ".inst 0xc1afc9d0 // fclamp { z16.s-z19.s }, z14.s, z15.s\n"
345 "st1w { z16.s }, p1, [x12, x13, LSL #2]\n"
346 "ldr x23, [x14, #0x20]\n"
347 "fmla z24.s, p3/M, z7.s, z12.s\n"
348 "st1w { z17.s }, p1, [x11, x13, LSL #2]\n"
349 "ldr x22, [x14, #0x28]\n"
350 "fmla z25.s, p3/M, z6.s, z12.s\n"
351 "fmla z26.s, p3/M, z8.s, z0.s\n"
352 "st1w { z18.s }, p1, [x10, x13, LSL #2]\n"
353 "ldr x21, [x14, #0x30]\n"
354 "fmla z27.s, p3/M, z7.s, z0.s\n"
355 ".inst 0xc1afc9dc // fclamp { z28.s-z31.s }, z14.s, z15.s\n"
356 "st1w { z19.s }, p1, [x9, x13, LSL #2]\n"
357 "ldr x20, [x14, #0x38]\n"
358 "fmla z20.s, p3/M, z4.s, z12.s\n"
359 "fmla z21.s, p3/M, z3.s, z12.s\n"
360 "st1w { z28.s }, p1, [x23, x13, LSL #2]\n"
361 "ldr x23, [x14, #0x40]\n"
362 "fmla z22.s, p3/M, z5.s, z0.s\n"
363 "fmla z23.s, p3/M, z4.s, z0.s\n"
364 "st1w { z29.s }, p1, [x22, x13, LSL #2]\n"
365 "ldr x22, [x14, #0x48]\n"
366 ".inst 0xc1afc9d8 // fclamp { z24.s-z27.s }, z14.s, z15.s\n"
367 "ld1w { z10.s }, p0/Z, [x25, x16, LSL #2]\n"
368 "st1w { z30.s }, p1, [x21, x13, LSL #2]\n"
369 "ldr x21, [x14, #0x50]\n"
370 "ld1w { z12.s }, p0/Z, [x24, x16, LSL #2]\n"
372 "st1w { z31.s }, p1, [x20, x13, LSL #2]\n"
373 "ldr x20, [x14, #0x58]\n"
374 ".inst 0xa040c100 // ld1w { z0.s-z3.s }, pn8.b/Z, [x8]\n"
376 "st1w { z24.s }, p1, [x23, x13, LSL #2]\n"
377 "ldr x23, [x14, #0x60]\n"
378 "whilelt p2.s, x15, %x[n_channels]\n"
379 ".inst 0xa040c104 // ld1w { z4.s-z7.s }, pn8.b/Z, [x8]\n"
380 "st1w { z25.s }, p1, [x22, x13, LSL #2]\n"
381 "ldr x22, [x14, #0x68]\n"
383 "cmp x16, %x[n_channels]\n"
384 "st1w { z26.s }, p1, [x21, x13, LSL #2]\n"
385 "ldr x21, [x14, #0x70]\n"
386 ".inst 0xc1afc9d4 // fclamp { z20.s-z23.s }, z14.s, z15.s\n"
387 "ld1w { z8.s }, p3/Z, [x8]\n"
388 "st1w { z27.s }, p1, [x20, x13, LSL #2]\n"
389 "ldr x20, [x14, #0x78]\n"
391 "st1w { z20.s }, p1, [x23, x13, LSL #2]\n"
392 "st1w { z21.s }, p1, [x22, x13, LSL #2]\n"
393 "st1w { z22.s }, p1, [x21, x13, LSL #2]\n"
394 "st1w { z23.s }, p1, [x20, x13, LSL #2]\n"
397 "movprfx z29, z13\n fmla z29.s, p3/M, z4.s, z9.s\n"
398 "movprfx z20, z13\n fmla z20.s, p3/M, z8.s, z9.s\n"
399 "ldr x24, [x17, #0x20]\n"
401 "movprfx z30, z13\n fmla z30.s, p3/M, z3.s, z9.s\n"
402 "movprfx z25, z13\n fmla z25.s, p3/M, z1.s, z9.s\n"
403 "ldr x20, [x17, #0x30]\n"
405 "movprfx z26, z13\n fmla z26.s, p3/M, z0.s, z9.s\n"
406 "ldr x23, [x17, #0x28]\n"
407 "movprfx z21, z13\n fmla z21.s, p3/M, z7.s, z9.s\n"
408 "movprfx z22, z13\n fmla z22.s, p3/M, z6.s, z9.s\n"
409 "fmla z29.s, p3/M, z5.s, z12.s\n"
410 "ldr x22, [x17, #0x38]\n"
411 "movprfx z28, z13\n fmla z28.s, p3/M, z5.s, z9.s\n"
412 "movprfx z24, z13\n fmla z24.s, p3/M, z2.s, z9.s\n"
413 "ld1w { z9.s }, p2/Z, [x20, x15, LSL #2]\n"
414 "ldr x21, [x17, #0x40]\n"
415 "fmla z20.s, p3/M, z0.s, z10.s\n"
416 "movprfx z23, z13\n fmla z23.s, p3/M, z2.s, z11.s\n"
417 "ld1w { z19.s }, p2/Z, [x24, x15, LSL #2]\n"
418 "ldr x20, [x17, #0x48]\n"
419 "fmla z30.s, p3/M, z4.s, z12.s\n"
420 "fmla z25.s, p3/M, z2.s, z12.s\n"
421 "ld1w { z17.s }, p2/Z, [x23, x15, LSL #2]\n"
422 "ldr x27, [x17, #0x50]\n"
423 "fmla z26.s, p3/M, z1.s, z12.s\n"
424 "fmla z21.s, p3/M, z8.s, z12.s\n"
425 "ldr x26, [x17, #0x60]\n"
426 "fmla z22.s, p3/M, z7.s, z12.s\n"
427 "movprfx z16, z13\n fmla z16.s, p3/M, z6.s, z19.s\n"
428 "ld1w { z11.s }, p2/Z, [x20, x15, LSL #2]\n"
429 "ldr x25, [x17, #0x68]\n"
430 "fmla z29.s, p3/M, z7.s, z9.s\n"
431 "fmla z23.s, p3/M, z6.s, z12.s\n"
432 "ldr x20, [x17, #0x58]\n"
433 "movprfx z31, z13\n fmla z31.s, p3/M, z3.s, z12.s\n"
434 "movprfx z27, z13\n fmla z27.s, p3/M, z0.s, z12.s\n"
435 "ld1w { z12.s }, p2/Z, [x22, x15, LSL #2]\n"
436 "ldr x24, [x17, #0x70]\n"
437 "movprfx z19, z13\n fmla z19.s, p3/M, z8.s, z17.s\n"
438 "fmla z30.s, p3/M, z6.s, z9.s\n"
439 "ld1w { z10.s }, p2/Z, [x21, x15, LSL #2]\n"
440 "ldr x23, [x17, #0x78]\n"
441 "fmla z25.s, p3/M, z4.s, z9.s\n"
442 "fmla z26.s, p3/M, z3.s, z9.s\n"
443 "ldr x22, [x17, #0x80]\n"
444 "movprfx z17, z13\n fmla z17.s, p3/M, z1.s, z9.s\n"
445 "movprfx z18, z13\n fmla z18.s, p3/M, z0.s, z9.s\n"
446 "ldr x21, [x17, #0x88]\n"
447 "fmla z28.s, p3/M, z8.s, z9.s\n"
448 "fmla z24.s, p3/M, z5.s, z9.s\n"
449 "ldr x12, [x14, #0x0]\n"
450 "fmla z16.s, p3/M, z2.s, z9.s\n"
451 "fmla z20.s, p3/M, z1.s, z12.s\n"
452 "ld1w { z9.s }, p2/Z, [x27, x15, LSL #2]\n"
453 "ldr x27, [x17, #0x90]\n"
454 "fmla z21.s, p3/M, z0.s, z12.s\n"
455 "fmla z22.s, p3/M, z2.s, z10.s\n"
456 "ld1w { z13.s }, p2/Z, [x20, x15, LSL #2]\n"
457 "ldr x20, [x17, #0x98]\n"
458 "fmla z29.s, p3/M, z8.s, z11.s\n"
459 "fmla z23.s, p3/M, z1.s, z10.s\n"
460 "ld1w { z12.s }, p2/Z, [x26, x15, LSL #2]\n"
461 "ldr x26, [x17, #0xa0]\n"
462 "fmla z30.s, p3/M, z7.s, z11.s\n"
463 "fmla z31.s, p3/M, z6.s, z11.s\n"
464 "ldr x11, [x14, #0x8]\n"
465 "fmla z25.s, p3/M, z5.s, z11.s\n"
466 "fmla z26.s, p3/M, z4.s, z11.s\n"
467 "ldr x10, [x14, #0x10]\n"
468 "fmla z27.s, p3/M, z3.s, z11.s\n"
469 "fmla z17.s, p3/M, z2.s, z11.s\n"
470 "ldr x9, [x14, #0x18]\n"
471 "fmla z18.s, p3/M, z1.s, z11.s\n"
472 "fmla z19.s, p3/M, z0.s, z11.s\n"
473 "ld1w { z10.s }, p2/Z, [x25, x15, LSL #2]\n"
474 "ldr x25, [x17, #0xa8]\n"
475 "fmla z20.s, p3/M, z3.s, z9.s\n"
476 "fmla z28.s, p3/M, z0.s, z9.s\n"
477 "fmla z24.s, p3/M, z6.s, z12.s\n"
478 "fmla z16.s, p3/M, z3.s, z12.s\n"
479 "ld1w { z12.s }, p2/Z, [x24, x15, LSL #2]\n"
480 "ldr x24, [x17, #0xb0]\n"
481 "fmla z21.s, p3/M, z4.s, z10.s\n"
482 "fmla z22.s, p3/M, z3.s, z10.s\n"
483 "fmla z29.s, p3/M, z1.s, z10.s\n"
484 "fmla z23.s, p3/M, z5.s, z13.s\n"
485 "fmla z31.s, p3/M, z2.s, z13.s\n"
486 "fmla z30.s, p3/M, z0.s, z10.s\n"
487 "ld1w { z13.s }, p2/Z, [x23, x15, LSL #2]\n"
488 "ldr x23, [x17, #0xb8]\n"
489 "fmla z27.s, p3/M, z8.s, z12.s\n"
490 "fmla z19.s, p3/M, z5.s, z12.s\n"
491 "ld1w { z9.s }, p2/Z, [x22, x15, LSL #2]\n"
492 "ldr x22, [x17, #0xc0]\n"
493 "fmla z20.s, p3/M, z5.s, z10.s\n"
494 "fmla z28.s, p3/M, z2.s, z10.s\n"
495 "ld1w { z12.s }, p2/Z, [x21, x15, LSL #2]\n"
496 "ldr x21, [x17, #0xc8]\n"
497 "fmla z21.s, p3/M, z5.s, z13.s\n"
498 "fmla z22.s, p3/M, z4.s, z13.s\n"
499 "fmla z29.s, p3/M, z2.s, z13.s\n"
500 "fmla z23.s, p3/M, z3.s, z13.s\n"
501 "fmla z30.s, p3/M, z1.s, z13.s\n"
502 "fmla z31.s, p3/M, z0.s, z13.s\n"
503 "ld1w { z10.s }, p2/Z, [x20, x15, LSL #2]\n"
504 "ldr x28, [x17, #0xd8]\n"
505 "fmla z16.s, p3/M, z7.s, z9.s\n"
506 "fmla z17.s, p3/M, z6.s, z9.s\n"
507 "ld1w { z11.s }, p2/Z, [x27, x15, LSL #2]\n"
508 "ldr x20, [x17, #0xd0]\n"
509 "fmla z20.s, p3/M, z7.s, z12.s\n"
510 "fmla z21.s, p3/M, z6.s, z12.s\n"
511 "fmla z28.s, p3/M, z4.s, z12.s\n"
512 "fmla z29.s, p3/M, z3.s, z12.s\n"
513 "fmla z24.s, p3/M, z1.s, z12.s\n"
514 "fmla z25.s, p3/M, z0.s, z12.s\n"
515 "ld1w { z12.s }, p2/Z, [x26, x15, LSL #2]\n"
516 "ldr x27, [x17, #0xe0]\n"
517 "fmla z22.s, p3/M, z8.s, z10.s\n"
518 "fmla z18.s, p3/M, z8.s, z11.s\n"
519 "fmla z19.s, p3/M, z7.s, z11.s\n"
520 "ld1w { z11.s }, p2/Z, [x25, x15, LSL #2]\n"
521 "fmla z27.s, p3/M, z1.s, z10.s\n"
522 "ldr x26, [x17, #0xe8]\n"
523 "fmla z23.s, p3/M, z7.s, z10.s\n"
524 "fmla z30.s, p3/M, z5.s, z10.s\n"
525 "fmla z31.s, p3/M, z4.s, z10.s\n"
526 "fmla z26.s, p3/M, z2.s, z10.s\n"
527 "ld1w { z9.s }, p2/Z, [x24, x15, LSL #2]\n"
528 "ldr x25, [x17, #0xf0]\n"
529 "fmla z20.s, p3/M, z2.s, z12.s\n"
530 "fmla z21.s, p3/M, z1.s, z12.s\n"
531 "fmla z22.s, p3/M, z0.s, z12.s\n"
532 "fmla z28.s, p3/M, z7.s, z11.s\n"
533 "ld1w { z12.s }, p2/Z, [x23, x15, LSL #2]\n"
534 "ldr x24, [x17, #0xf8]\n"
535 "fmla z29.s, p3/M, z6.s, z11.s\n"
536 "fmla z24.s, p3/M, z4.s, z11.s\n"
537 "fmla z25.s, p3/M, z3.s, z11.s\n"
538 "fmla z16.s, p3/M, z1.s, z11.s\n"
539 "fmla z17.s, p3/M, z0.s, z11.s\n"
540 "ld1w { z10.s }, p2/Z, [x22, x15, LSL #2]\n"
541 "fmla z27.s, p3/M, z4.s, z10.s\n"
542 "ldr x23, [x17, #0x100]\n"
543 "fmla z18.s, p3/M, z2.s, z10.s\n"
544 "fmla z21.s, p3/M, z2.s, z9.s\n"
545 "fmla z22.s, p3/M, z1.s, z9.s\n"
546 "fmla z23.s, p3/M, z0.s, z9.s\n"
547 "ld1w { z11.s }, p2/Z, [x21, x15, LSL #2]\n"
548 "ldr x22, [x17, #0x108]\n"
549 "fmla z20.s, p3/M, z6.s, z12.s\n"
550 "fmla z28.s, p3/M, z3.s, z12.s\n"
551 "fmla z24.s, p3/M, z0.s, z12.s\n"
552 "fmla z30.s, p3/M, z8.s, z10.s\n"
553 "ld1w { z12.s }, p2/Z, [x20, x15, LSL #2]\n"
554 "ldr x21, [x17, #0x110]\n"
555 "fmla z31.s, p3/M, z7.s, z10.s\n"
556 "fmla z26.s, p3/M, z5.s, z10.s\n"
557 "fmla z19.s, p3/M, z1.s, z10.s\n"
558 "ld1w { z9.s }, p2/Z, [x28, x15, LSL #2]\n"
559 "fmla z27.s, p3/M, z2.s, z11.s\n"
560 "ldr x20, [x17, #0x118]\n"
561 "fmla z16.s, p3/M, z0.s, z12.s\n"
562 "fmla z17.s, p3/M, z4.s, z9.s\n"
563 "fmla z18.s, p3/M, z3.s, z9.s\n"
564 "fmla z23.s, p3/M, z8.s, z11.s\n"
565 "fmla z31.s, p3/M, z5.s, z11.s\n"
566 "fmla z28.s, p3/M, z6.s, z12.s\n"
567 "ld1w { z10.s }, p2/Z, [x27, x15, LSL #2]\n"
568 "fmla z24.s, p3/M, z3.s, z12.s\n"
569 "ld1w { z12.s }, p2/Z, [x26, x15, LSL #2]\n"
570 "fmla z25.s, p3/M, z7.s, z9.s\n"
571 "fmla z26.s, p3/M, z6.s, z9.s\n"
572 "fmla z16.s, p3/M, z5.s, z9.s\n"
573 "fmla z27.s, p3/M, z5.s, z10.s\n"
574 "fmla z19.s, p3/M, z2.s, z10.s\n"
575 "fmla z17.s, p3/M, z7.s, z12.s\n"
576 "fmla z18.s, p3/M, z6.s, z12.s\n"
577 "fmla z24.s, p3/M, z8.s, z9.s\n"
578 "ld1w { z11.s }, p2/Z, [x25, x15, LSL #2]\n"
579 "fmla z16.s, p3/M, z8.s, z12.s\n"
580 "fmla z25.s, p3/M, z8.s, z11.s\n"
581 "fmla z26.s, p3/M, z7.s, z11.s\n"
582 "ld1w { z9.s }, p2/Z, [x23, x15, LSL #2]\n"
583 "fmla z27.s, p3/M, z6.s, z11.s\n"
584 "fmla z17.s, p3/M, z5.s, z11.s\n"
585 "fmla z18.s, p3/M, z4.s, z11.s\n"
586 "fmla z19.s, p3/M, z3.s, z11.s\n"
587 "ld1w { z11.s }, p2/Z, [x22, x15, LSL #2]\n"
588 "fmla z31.s, p3/M, z8.s, z10.s\n"
589 "ld1w { z12.s }, p2/Z, [x24, x15, LSL #2]\n"
590 "fmla z20.s, p3/M, z4.s, z9.s\n"
591 "fmla z21.s, p3/M, z3.s, z9.s\n"
592 "fmla z22.s, p3/M, z5.s, z11.s\n"
593 "fmla z23.s, p3/M, z4.s, z11.s\n"
594 "fmla z17.s, p3/M, z8.s, z12.s\n"
595 "fmla z18.s, p3/M, z7.s, z12.s\n"
596 "fmla z19.s, p3/M, z6.s, z12.s\n"
597 "ld1w { z13.s }, p2/Z, [x21, x15, LSL #2]\n"
598 "fmla z28.s, p3/M, z1.s, z9.s\n"
599 "fmla z29.s, p3/M, z0.s, z9.s\n"
600 "ld1w { z0.s }, p2/Z, [x20, x15, LSL #2]\n"
601 "fmla z30.s, p3/M, z2.s, z11.s\n"
602 "fmla z31.s, p3/M, z1.s, z11.s\n"
603 ".inst 0xc1afc9d4 // fclamp { z20.s-z23.s }, z14.s, z15.s\n"
604 "st1w { z20.s }, p0, [x12, x13, LSL #2]\n"
605 "ldr x23, [x14, #0x20]\n"
606 "fmla z24.s, p3/M, z7.s, z13.s\n"
607 "st1w { z21.s }, p0, [x11, x13, LSL #2]\n"
608 "ldr x22, [x14, #0x28]\n"
609 "fmla z25.s, p3/M, z6.s, z13.s\n"
610 "fmla z26.s, p3/M, z8.s, z0.s\n"
611 "st1w { z22.s }, p0, [x10, x13, LSL #2]\n"
612 "ldr x21, [x14, #0x30]\n"
613 "fmla z27.s, p3/M, z7.s, z0.s\n"
614 ".inst 0xc1afc9dc // fclamp { z28.s-z31.s }, z14.s, z15.s\n"
615 "st1w { z23.s }, p0, [x9, x13, LSL #2]\n"
616 "ldr x20, [x14, #0x38]\n"
617 "fmla z16.s, p3/M, z4.s, z13.s\n"
618 "fmla z17.s, p3/M, z3.s, z13.s\n"
619 "st1w { z28.s }, p0, [x23, x13, LSL #2]\n"
620 "ldr x23, [x14, #0x40]\n"
621 "fmla z18.s, p3/M, z5.s, z0.s\n"
622 "fmla z19.s, p3/M, z4.s, z0.s\n"
623 "st1w { z29.s }, p0, [x22, x13, LSL #2]\n"
624 "ldr x22, [x14, #0x48]\n"
625 ".inst 0xc1afc9d8 // fclamp { z24.s-z27.s }, z14.s, z15.s\n"
626 ".inst 0xc1afc9d0 // fclamp { z16.s-z19.s }, z14.s, z15.s\n"
627 "st1w { z30.s }, p0, [x21, x13, LSL #2]\n"
628 "ldr x21, [x14, #0x50]\n"
629 "st1w { z31.s }, p0, [x20, x13, LSL #2]\n"
630 "ldr x20, [x14, #0x58]\n"
631 "st1w { z24.s }, p0, [x23, x13, LSL #2]\n"
632 "ldr x23, [x14, #0x60]\n"
633 "st1w { z25.s }, p0, [x22, x13, LSL #2]\n"
634 "ldr x22, [x14, #0x68]\n"
635 "st1w { z26.s }, p0, [x21, x13, LSL #2]\n"
636 "ldr x21, [x14, #0x70]\n"
637 "st1w { z27.s }, p0, [x20, x13, LSL #2]\n"
638 "ldr x20, [x14, #0x78]\n"
639 "st1w { z16.s }, p0, [x23, x13, LSL #2]\n"
640 "st1w { z17.s }, p0, [x22, x13, LSL #2]\n"
641 "st1w { z18.s }, p0, [x21, x13, LSL #2]\n"
642 "st1w { z19.s }, p0, [x20, x13, LSL #2]\n"
643 ".inst 0xd503467f // SMSTOP\n"
645 : [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)
646 :
"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"
653 #endif // defined(ARM_COMPUTE_ENABLE_SME2)