25 #include "pooling.hpp"
32 #if defined(ARM_COMPUTE_ENABLE_SVE)
40 int32_t multiplier, shift;
43 constexpr RescaleParams rescale_params[8] = {
55 void sve_s8q_nhwc_avg_generic_depthfirst_impl(
56 const uint64_t window_cells,
57 const uint64_t n_valid_cells,
59 const int8_t *
const *
const inptrs,
61 const Requantize32 &qp
64 if (n_valid_cells == 1 && window_cells == 1)
67 std::memcpy(outptr, *inptrs, n_channels);
72 int32_t shift_value = 0, rescale_value = 0;
73 if (2 <= window_cells && window_cells <= 9)
75 auto ¶ms = rescale_params[window_cells - 2];
76 rescale_value = params.multiplier;
77 shift_value = params.shift;
81 auto f_rescale_value = 1.0f /
static_cast<float>(window_cells);
84 while (f_rescale_value < 0.5f)
87 f_rescale_value *= 2.0f;
90 int64_t long_rescale_value =
round(f_rescale_value *
static_cast<float>(1ll << 31));
91 if (long_rescale_value == (1ll << 31))
94 long_rescale_value >>= 1;
96 rescale_value =
static_cast<int32_t
>(long_rescale_value);
101 const int32_t shift = qp.per_layer_left_shift - qp.per_layer_right_shift + shift_value;
102 const int32_t left_shift = shift > 0 ? shift : 0;
103 const int32_t right_shift = shift <= 0 ? shift : 0;
105 int32_t combined_rescale_value = 0;
106 __asm__ __volatile__ (
107 "mov v16.s[0], %w[per_layer_mul]\n"
108 "mov v17.s[0], %w[rescale_value]\n"
109 "sqrdmulh s18, s16, s17\n"
110 "mov %w[combined_rescale_value], v18.s[0]\n"
111 : [combined_rescale_value]
"=r" (combined_rescale_value)
112 : [per_layer_mul]
"r" (qp.per_layer_mul), [rescale_value]
"r" (rescale_value)
113 :
"v16",
"v17",
"v18"
116 __asm__ __volatile__(
119 "cntb x25, ALL, MUL #2\n"
120 "cntb x24, ALL, MUL #3\n"
121 "whilelt p4.b, x27, %x[n_channels]\n"
122 "whilelt p3.b, x26, %x[n_channels]\n"
123 "whilelt p2.b, x25, %x[n_channels]\n"
124 "whilelt p1.b, x24, %x[n_channels]\n"
128 "lsr x23, %x[n_valid_cells], #0x1\n"
131 "mov x22, %x[inptrs]\n"
147 "ldp x21, x20, [x22, #0x0]\n"
148 "subs x23, x23, #0x1\n"
149 "add x22, x22, #0x10\n"
150 "ld1b { z31.b }, p4/Z, [x21, x27]\n"
151 "ld1b { z30.b }, p4/Z, [x20, x27]\n"
152 "ld1b { z29.b }, p3/Z, [x21, x26]\n"
153 "ld1b { z28.b }, p3/Z, [x20, x26]\n"
154 "ld1b { z27.b }, p2/Z, [x21, x25]\n"
155 "ld1b { z26.b }, p2/Z, [x20, x25]\n"
156 "ld1b { z25.b }, p1/Z, [x21, x24]\n"
157 "ld1b { z24.b }, p1/Z, [x20, x24]\n"
160 ".inst 0x455e03f7 // saddlb z23.h, z31.b, z30.b\n"
161 ".inst 0x455e07f6 // saddlt z22.h, z31.b, z30.b\n"
162 "ldp x21, x20, [x22, #0x0]\n"
163 "subs x23, x23, #0x1\n"
164 ".inst 0x455c03b5 // saddlb z21.h, z29.b, z28.b\n"
165 ".inst 0x455c07b4 // saddlt z20.h, z29.b, z28.b\n"
166 "add x22, x22, #0x10\n"
167 "ld1b { z31.b }, p4/Z, [x21, x27]\n"
168 ".inst 0x455a0373 // saddlb z19.h, z27.b, z26.b\n"
169 ".inst 0x455a0772 // saddlt z18.h, z27.b, z26.b\n"
170 "ld1b { z30.b }, p4/Z, [x20, x27]\n"
171 "ld1b { z29.b }, p3/Z, [x21, x26]\n"
172 ".inst 0x45580331 // saddlb z17.h, z25.b, z24.b\n"
173 ".inst 0x45580730 // saddlt z16.h, z25.b, z24.b\n"
174 "ld1b { z28.b }, p3/Z, [x20, x26]\n"
175 "ld1b { z27.b }, p2/Z, [x21, x25]\n"
176 ".inst 0x459741ef // saddwb z15.s, z15.s, z23.h\n"
177 ".inst 0x459745ce // saddwt z14.s, z14.s, z23.h\n"
178 "ld1b { z26.b }, p2/Z, [x20, x25]\n"
179 "ld1b { z25.b }, p1/Z, [x21, x24]\n"
180 ".inst 0x459641ad // saddwb z13.s, z13.s, z22.h\n"
181 ".inst 0x4596458c // saddwt z12.s, z12.s, z22.h\n"
182 "ld1b { z24.b }, p1/Z, [x20, x24]\n"
183 ".inst 0x4595416b // saddwb z11.s, z11.s, z21.h\n"
184 ".inst 0x4595454a // saddwt z10.s, z10.s, z21.h\n"
185 ".inst 0x45944129 // saddwb z9.s, z9.s, z20.h\n"
186 ".inst 0x45944508 // saddwt z8.s, z8.s, z20.h\n"
187 ".inst 0x459340e7 // saddwb z7.s, z7.s, z19.h\n"
188 ".inst 0x459344c6 // saddwt z6.s, z6.s, z19.h\n"
189 ".inst 0x459240a5 // saddwb z5.s, z5.s, z18.h\n"
190 ".inst 0x45924484 // saddwt z4.s, z4.s, z18.h\n"
191 ".inst 0x45914063 // saddwb z3.s, z3.s, z17.h\n"
192 ".inst 0x45914442 // saddwt z2.s, z2.s, z17.h\n"
193 ".inst 0x45904021 // saddwb z1.s, z1.s, z16.h\n"
194 ".inst 0x45904400 // saddwt z0.s, z0.s, z16.h\n"
197 ".inst 0x455e03f7 // saddlb z23.h, z31.b, z30.b\n"
198 ".inst 0x455e07f6 // saddlt z22.h, z31.b, z30.b\n"
199 ".inst 0x455c03b5 // saddlb z21.h, z29.b, z28.b\n"
200 ".inst 0x455c07b4 // saddlt z20.h, z29.b, z28.b\n"
201 ".inst 0x455a0373 // saddlb z19.h, z27.b, z26.b\n"
202 ".inst 0x455a0772 // saddlt z18.h, z27.b, z26.b\n"
203 ".inst 0x45580331 // saddlb z17.h, z25.b, z24.b\n"
204 ".inst 0x45580730 // saddlt z16.h, z25.b, z24.b\n"
205 ".inst 0x459741ef // saddwb z15.s, z15.s, z23.h\n"
206 ".inst 0x459745ce // saddwt z14.s, z14.s, z23.h\n"
207 ".inst 0x459641ad // saddwb z13.s, z13.s, z22.h\n"
208 ".inst 0x4596458c // saddwt z12.s, z12.s, z22.h\n"
209 ".inst 0x4595416b // saddwb z11.s, z11.s, z21.h\n"
210 ".inst 0x4595454a // saddwt z10.s, z10.s, z21.h\n"
211 ".inst 0x45944129 // saddwb z9.s, z9.s, z20.h\n"
212 ".inst 0x45944508 // saddwt z8.s, z8.s, z20.h\n"
213 ".inst 0x459340e7 // saddwb z7.s, z7.s, z19.h\n"
214 ".inst 0x459344c6 // saddwt z6.s, z6.s, z19.h\n"
215 ".inst 0x459240a5 // saddwb z5.s, z5.s, z18.h\n"
216 ".inst 0x45924484 // saddwt z4.s, z4.s, z18.h\n"
217 ".inst 0x45914063 // saddwb z3.s, z3.s, z17.h\n"
218 ".inst 0x45914442 // saddwt z2.s, z2.s, z17.h\n"
219 ".inst 0x45904021 // saddwb z1.s, z1.s, z16.h\n"
220 ".inst 0x45904400 // saddwt z0.s, z0.s, z16.h\n"
222 "ands x21, %x[n_valid_cells], #0x1\n"
225 "ldr x20, [x22], #0x8\n"
226 "ld1b { z16.b }, p4/Z, [x20, x27]\n"
227 ".inst 0x4508a217 // sshllb z23.h, z16.b, #0x0\n"
228 ".inst 0x4508a616 // sshllt z22.h, z16.b, #0x0\n"
229 "ld1b { z16.b }, p3/Z, [x20, x26]\n"
230 "ld1b { z17.b }, p2/Z, [x20, x25]\n"
231 ".inst 0x4508a215 // sshllb z21.h, z16.b, #0x0\n"
232 ".inst 0x4508a614 // sshllt z20.h, z16.b, #0x0\n"
233 "ld1b { z16.b }, p1/Z, [x20, x24]\n"
234 ".inst 0x4508a233 // sshllb z19.h, z17.b, #0x0\n"
235 ".inst 0x4508a632 // sshllt z18.h, z17.b, #0x0\n"
236 "subs x21, x21, #0x1\n"
237 ".inst 0x4508a211 // sshllb z17.h, z16.b, #0x0\n"
238 ".inst 0x4508a610 // sshllt z16.h, z16.b, #0x0\n"
239 ".inst 0x459741ef // saddwb z15.s, z15.s, z23.h\n"
240 ".inst 0x459745ce // saddwt z14.s, z14.s, z23.h\n"
241 ".inst 0x459641ad // saddwb z13.s, z13.s, z22.h\n"
242 ".inst 0x4596458c // saddwt z12.s, z12.s, z22.h\n"
243 ".inst 0x4595416b // saddwb z11.s, z11.s, z21.h\n"
244 ".inst 0x4595454a // saddwt z10.s, z10.s, z21.h\n"
245 ".inst 0x45944129 // saddwb z9.s, z9.s, z20.h\n"
246 ".inst 0x45944508 // saddwt z8.s, z8.s, z20.h\n"
247 ".inst 0x459340e7 // saddwb z7.s, z7.s, z19.h\n"
248 ".inst 0x459344c6 // saddwt z6.s, z6.s, z19.h\n"
249 ".inst 0x459240a5 // saddwb z5.s, z5.s, z18.h\n"
250 ".inst 0x45924484 // saddwt z4.s, z4.s, z18.h\n"
251 ".inst 0x45914063 // saddwb z3.s, z3.s, z17.h\n"
252 ".inst 0x45914442 // saddwt z2.s, z2.s, z17.h\n"
253 ".inst 0x45904021 // saddwb z1.s, z1.s, z16.h\n"
254 ".inst 0x45904400 // saddwt z0.s, z0.s, z16.h\n"
257 "ld1rw { z18.s }, p0/Z, [%x[left_shift]]\n"
258 "ld1rw { z17.s }, p0/Z, [%x[combined_rescale_value]]\n"
259 ".inst 0x4482824f // srshl z15.s, p0/M, z15.s, z18.s\n"
260 ".inst 0x4482824e // srshl z14.s, p0/M, z14.s, z18.s\n"
261 ".inst 0x4482824d // srshl z13.s, p0/M, z13.s, z18.s\n"
262 ".inst 0x4482824c // srshl z12.s, p0/M, z12.s, z18.s\n"
263 "ld1rw { z16.s }, p0/Z, [%x[right_shift]]\n"
264 ".inst 0x04b175ef // sqrdmulh z15.s, z15.s, z17.s\n"
265 ".inst 0x4482824b // srshl z11.s, p0/M, z11.s, z18.s\n"
266 ".inst 0x4482824a // srshl z10.s, p0/M, z10.s, z18.s\n"
267 ".inst 0x04b175ce // sqrdmulh z14.s, z14.s, z17.s\n"
268 ".inst 0x04b175ad // sqrdmulh z13.s, z13.s, z17.s\n"
269 ".inst 0x44828249 // srshl z9.s, p0/M, z9.s, z18.s\n"
270 ".inst 0x44828248 // srshl z8.s, p0/M, z8.s, z18.s\n"
271 ".inst 0x04b1758c // sqrdmulh z12.s, z12.s, z17.s\n"
272 ".inst 0x04b1756b // sqrdmulh z11.s, z11.s, z17.s\n"
273 ".inst 0x44828247 // srshl z7.s, p0/M, z7.s, z18.s\n"
274 ".inst 0x44828246 // srshl z6.s, p0/M, z6.s, z18.s\n"
275 ".inst 0x04b1754a // sqrdmulh z10.s, z10.s, z17.s\n"
276 ".inst 0x04b17529 // sqrdmulh z9.s, z9.s, z17.s\n"
277 ".inst 0x44828245 // srshl z5.s, p0/M, z5.s, z18.s\n"
278 ".inst 0x44828244 // srshl z4.s, p0/M, z4.s, z18.s\n"
279 ".inst 0x04b17508 // sqrdmulh z8.s, z8.s, z17.s\n"
280 ".inst 0x04b174e7 // sqrdmulh z7.s, z7.s, z17.s\n"
281 ".inst 0x44828243 // srshl z3.s, p0/M, z3.s, z18.s\n"
282 ".inst 0x44828242 // srshl z2.s, p0/M, z2.s, z18.s\n"
283 ".inst 0x04b174c6 // sqrdmulh z6.s, z6.s, z17.s\n"
284 ".inst 0x04b174a5 // sqrdmulh z5.s, z5.s, z17.s\n"
285 ".inst 0x44828241 // srshl z1.s, p0/M, z1.s, z18.s\n"
286 ".inst 0x44828240 // srshl z0.s, p0/M, z0.s, z18.s\n"
287 ".inst 0x04b17484 // sqrdmulh z4.s, z4.s, z17.s\n"
288 ".inst 0x04b17463 // sqrdmulh z3.s, z3.s, z17.s\n"
289 ".inst 0x04b17442 // sqrdmulh z2.s, z2.s, z17.s\n"
290 ".inst 0x04b17421 // sqrdmulh z1.s, z1.s, z17.s\n"
291 ".inst 0x4482820f // srshl z15.s, p0/M, z15.s, z16.s\n"
292 ".inst 0x4482820e // srshl z14.s, p0/M, z14.s, z16.s\n"
293 ".inst 0x04b17400 // sqrdmulh z0.s, z0.s, z17.s\n"
295 ".inst 0x4482820d // srshl z13.s, p0/M, z13.s, z16.s\n"
296 ".inst 0x4482820c // srshl z12.s, p0/M, z12.s, z16.s\n"
297 ".inst 0x4482820b // srshl z11.s, p0/M, z11.s, z16.s\n"
298 ".inst 0x4482820a // srshl z10.s, p0/M, z10.s, z16.s\n"
299 ".inst 0x44828209 // srshl z9.s, p0/M, z9.s, z16.s\n"
300 ".inst 0x44828208 // srshl z8.s, p0/M, z8.s, z16.s\n"
301 ".inst 0x44828207 // srshl z7.s, p0/M, z7.s, z16.s\n"
302 ".inst 0x44828206 // srshl z6.s, p0/M, z6.s, z16.s\n"
303 ".inst 0x44828205 // srshl z5.s, p0/M, z5.s, z16.s\n"
304 ".inst 0x44828204 // srshl z4.s, p0/M, z4.s, z16.s\n"
305 ".inst 0x44828203 // srshl z3.s, p0/M, z3.s, z16.s\n"
306 ".inst 0x44828202 // srshl z2.s, p0/M, z2.s, z16.s\n"
307 ".inst 0x44828201 // srshl z1.s, p0/M, z1.s, z16.s\n"
308 ".inst 0x44828200 // srshl z0.s, p0/M, z0.s, z16.s\n"
309 "not z16.s, p0/M, z18.s\n"
310 "smax z15.s, p0/M, z15.s, z16.s\n"
311 "smax z14.s, p0/M, z14.s, z16.s\n"
312 "smax z13.s, p0/M, z13.s, z16.s\n"
313 "smax z12.s, p0/M, z12.s, z16.s\n"
314 "smax z11.s, p0/M, z11.s, z16.s\n"
315 "smax z10.s, p0/M, z10.s, z16.s\n"
316 "smax z9.s, p0/M, z9.s, z16.s\n"
317 "smax z8.s, p0/M, z8.s, z16.s\n"
318 "smax z7.s, p0/M, z7.s, z16.s\n"
319 "smax z6.s, p0/M, z6.s, z16.s\n"
320 "smax z5.s, p0/M, z5.s, z16.s\n"
321 "smax z4.s, p0/M, z4.s, z16.s\n"
322 "smax z3.s, p0/M, z3.s, z16.s\n"
323 "smax z2.s, p0/M, z2.s, z16.s\n"
324 "smax z1.s, p0/M, z1.s, z16.s\n"
325 "smax z0.s, p0/M, z0.s, z16.s\n"
326 "smin z15.s, p0/M, z15.s, z18.s\n"
327 "smin z14.s, p0/M, z14.s, z18.s\n"
328 "smin z13.s, p0/M, z13.s, z18.s\n"
329 "trn1 z17.h, z15.h, z14.h\n"
330 "smin z12.s, p0/M, z12.s, z18.s\n"
331 "smin z11.s, p0/M, z11.s, z18.s\n"
332 "trn1 z16.h, z13.h, z12.h\n"
333 "trn1 z16.b, z17.b, z16.b\n"
334 "smin z10.s, p0/M, z10.s, z18.s\n"
335 "smin z9.s, p0/M, z9.s, z18.s\n"
336 "trn1 z17.h, z11.h, z10.h\n"
337 "st1b { z16.b }, p4, [%x[outptr], x27]\n"
338 "smin z8.s, p0/M, z8.s, z18.s\n"
339 "smin z7.s, p0/M, z7.s, z18.s\n"
340 "trn1 z16.h, z9.h, z8.h\n"
341 "trn1 z16.b, z17.b, z16.b\n"
342 "smin z6.s, p0/M, z6.s, z18.s\n"
343 "smin z5.s, p0/M, z5.s, z18.s\n"
344 "trn1 z17.h, z7.h, z6.h\n"
345 "st1b { z16.b }, p3, [%x[outptr], x26]\n"
346 "smin z4.s, p0/M, z4.s, z18.s\n"
347 "smin z3.s, p0/M, z3.s, z18.s\n"
348 "trn1 z16.h, z5.h, z4.h\n"
349 "trn1 z16.b, z17.b, z16.b\n"
350 "smin z2.s, p0/M, z2.s, z18.s\n"
351 "smin z1.s, p0/M, z1.s, z18.s\n"
352 "trn1 z17.h, z3.h, z2.h\n"
353 "st1b { z16.b }, p2, [%x[outptr], x25]\n"
354 "smin z0.s, p0/M, z0.s, z18.s\n"
355 "trn1 z16.h, z1.h, z0.h\n"
356 "trn1 z16.b, z17.b, z16.b\n"
357 "st1b { z16.b }, p1, [%x[outptr], x24]\n"
358 "incb x24, ALL, MUL #4\n"
359 "whilelt p1.b, x24, %x[n_channels]\n"
360 "incb x27, ALL, MUL #4\n"
361 "incb x26, ALL, MUL #4\n"
362 "incb x25, ALL, MUL #4\n"
365 "whilelt p4.b, x27, %x[n_channels]\n"
368 "lsr x23, %x[n_valid_cells], #0x1\n"
371 "mov x22, %x[inptrs]\n"
375 "ldp x21, x20, [x22, #0x0]\n"
376 "subs x23, x23, #0x1\n"
377 "add x22, x22, #0x10\n"
378 "ld1b { z31.b }, p4/Z, [x21, x27]\n"
379 "ld1b { z30.b }, p4/Z, [x20, x27]\n"
382 ".inst 0x455e03f1 // saddlb z17.h, z31.b, z30.b\n"
383 ".inst 0x455e07f0 // saddlt z16.h, z31.b, z30.b\n"
384 "ldp x21, x20, [x22, #0x0]\n"
385 "subs x23, x23, #0x1\n"
386 ".inst 0x459141ef // saddwb z15.s, z15.s, z17.h\n"
387 ".inst 0x459145ce // saddwt z14.s, z14.s, z17.h\n"
388 "add x22, x22, #0x10\n"
389 "ld1b { z31.b }, p4/Z, [x21, x27]\n"
390 ".inst 0x459041ad // saddwb z13.s, z13.s, z16.h\n"
391 ".inst 0x4590458c // saddwt z12.s, z12.s, z16.h\n"
392 "ld1b { z30.b }, p4/Z, [x20, x27]\n"
395 ".inst 0x455e03f1 // saddlb z17.h, z31.b, z30.b\n"
396 ".inst 0x455e07f0 // saddlt z16.h, z31.b, z30.b\n"
397 ".inst 0x459141ef // saddwb z15.s, z15.s, z17.h\n"
398 ".inst 0x459145ce // saddwt z14.s, z14.s, z17.h\n"
399 ".inst 0x459041ad // saddwb z13.s, z13.s, z16.h\n"
400 ".inst 0x4590458c // saddwt z12.s, z12.s, z16.h\n"
402 "ands x21, %x[n_valid_cells], #0x1\n"
405 "ldr x20, [x22], #0x8\n"
406 "ld1b { z16.b }, p4/Z, [x20, x27]\n"
407 ".inst 0x4508a211 // sshllb z17.h, z16.b, #0x0\n"
408 ".inst 0x4508a610 // sshllt z16.h, z16.b, #0x0\n"
409 "subs x21, x21, #0x1\n"
410 ".inst 0x459141ef // saddwb z15.s, z15.s, z17.h\n"
411 ".inst 0x459145ce // saddwt z14.s, z14.s, z17.h\n"
412 ".inst 0x459041ad // saddwb z13.s, z13.s, z16.h\n"
413 ".inst 0x4590458c // saddwt z12.s, z12.s, z16.h\n"
416 "ld1rw { z16.s }, p0/Z, [%x[left_shift]]\n"
417 "ld1rw { z17.s }, p0/Z, [%x[combined_rescale_value]]\n"
418 ".inst 0x4482820f // srshl z15.s, p0/M, z15.s, z16.s\n"
419 ".inst 0x4482820e // srshl z14.s, p0/M, z14.s, z16.s\n"
420 ".inst 0x4482820d // srshl z13.s, p0/M, z13.s, z16.s\n"
421 ".inst 0x4482820c // srshl z12.s, p0/M, z12.s, z16.s\n"
422 "ld1rw { z16.s }, p0/Z, [%x[right_shift]]\n"
423 ".inst 0x04b175ef // sqrdmulh z15.s, z15.s, z17.s\n"
424 ".inst 0x04b175ce // sqrdmulh z14.s, z14.s, z17.s\n"
425 ".inst 0x04b175ad // sqrdmulh z13.s, z13.s, z17.s\n"
426 ".inst 0x4482820f // srshl z15.s, p0/M, z15.s, z16.s\n"
427 ".inst 0x4482820e // srshl z14.s, p0/M, z14.s, z16.s\n"
428 ".inst 0x04b1758c // sqrdmulh z12.s, z12.s, z17.s\n"
430 ".inst 0x4482820d // srshl z13.s, p0/M, z13.s, z16.s\n"
431 ".inst 0x4482820c // srshl z12.s, p0/M, z12.s, z16.s\n"
432 "not z16.s, p0/M, z18.s\n"
433 "smax z15.s, p0/M, z15.s, z16.s\n"
434 "smax z14.s, p0/M, z14.s, z16.s\n"
435 "smax z13.s, p0/M, z13.s, z16.s\n"
436 "smax z12.s, p0/M, z12.s, z16.s\n"
437 "smin z15.s, p0/M, z15.s, z18.s\n"
438 "smin z14.s, p0/M, z14.s, z18.s\n"
439 "smin z13.s, p0/M, z13.s, z18.s\n"
440 "trn1 z17.h, z15.h, z14.h\n"
441 "smin z12.s, p0/M, z12.s, z18.s\n"
442 "trn1 z16.h, z13.h, z12.h\n"
443 "trn1 z16.b, z17.b, z16.b\n"
444 "st1b { z16.b }, p4, [%x[outptr], x27]\n"
446 "whilelt p4.b, x27, %x[n_channels]\n"
450 : [combined_rescale_value]
"r" (&combined_rescale_value), [inptrs]
"r" (inptrs), [left_shift]
"r" (&left_shift), [n_channels]
"r" (n_channels), [n_valid_cells]
"r" (n_valid_cells), [outptr]
"r" (outptr), [right_shift]
"r" (&right_shift)
451 :
"cc",
"memory",
"p0",
"p1",
"p2",
"p3",
"p4",
"x20",
"x21",
"x22",
"x23",
"x24",
"x25",
"x26",
"x27",
"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"
458 #endif // defined(ARM_COMPUTE_ENABLE_SVE)