31 #if defined(__aarch64__)
39 int32_t multiplier, shift;
42 constexpr RescaleParams rescale_params[8] = {
54 void a64_u8_nhwc_avg_generic_depthfirst_impl(
55 const uint64_t window_cells,
56 const uint64_t n_valid_cells,
58 const uint8_t *
const *
const inptrs,
62 if (n_valid_cells == 1 && window_cells == 1)
65 std::memcpy(outptr, *inptrs, n_channels);
70 int32_t shift_value = 0, rescale_value = 0;
71 if (2 <= window_cells && window_cells <= 9)
73 auto ¶ms = rescale_params[window_cells - 2];
74 rescale_value = params.multiplier;
75 shift_value = params.shift;
79 auto f_rescale_value = 1.0f /
static_cast<float>(window_cells);
82 while (f_rescale_value < 0.5f)
85 f_rescale_value *= 2.0f;
88 int64_t long_rescale_value =
round(f_rescale_value *
static_cast<float>(1ll << 31));
89 if (long_rescale_value == (1ll << 31))
92 long_rescale_value >>= 1;
94 rescale_value =
static_cast<int32_t
>(long_rescale_value);
98 "cmp %x[n_channels], #0x40\n"
105 "lsr x23, %x[n_valid_cells], #0x1\n"
106 "movi v15.4s, #0x0\n"
107 "movi v14.4s, #0x0\n"
108 "mov x22, %x[inptrs]\n"
109 "movi v13.4s, #0x0\n"
110 "movi v12.4s, #0x0\n"
111 "movi v11.4s, #0x0\n"
112 "movi v10.4s, #0x0\n"
124 "ldp x21, x20, [x22, #0x0]\n"
125 "ldr q31, [x21, x27]\n"
126 "subs x23, x23, #0x1\n"
127 "add x22, x22, #0x10\n"
128 "ldr q30, [x20, x27]\n"
129 "ldr q29, [x21, x26]\n"
130 "ldr q28, [x20, x26]\n"
131 "ldr q27, [x21, x25]\n"
132 "ldr q26, [x20, x25]\n"
133 "ldr q25, [x21, x24]\n"
134 "ldr q24, [x20, x24]\n"
137 "uaddl v23.8h, v31.8b, v30.8b\n"
138 "uaddl2 v22.8h, v31.16b, v30.16b\n"
139 "ldp x21, x20, [x22, #0x0]\n"
140 "ldr q31, [x21, x27]\n"
141 "ldr q30, [x20, x27]\n"
142 "uaddl v21.8h, v29.8b, v28.8b\n"
143 "uaddl2 v20.8h, v29.16b, v28.16b\n"
144 "ldr q29, [x21, x26]\n"
145 "ldr q28, [x20, x26]\n"
146 "uaddl v19.8h, v27.8b, v26.8b\n"
147 "uaddl2 v18.8h, v27.16b, v26.16b\n"
148 "ldr q27, [x21, x25]\n"
149 "ldr q26, [x20, x25]\n"
150 "uaddl v17.8h, v25.8b, v24.8b\n"
151 "uaddl2 v16.8h, v25.16b, v24.16b\n"
152 "ldr q25, [x21, x24]\n"
153 "ldr q24, [x20, x24]\n"
154 "subs x23, x23, #0x1\n"
155 "uaddw v15.4s, v15.4s, v23.4h\n"
156 "uaddw2 v14.4s, v14.4s, v23.8h\n"
157 "uaddw v13.4s, v13.4s, v22.4h\n"
158 "uaddw2 v12.4s, v12.4s, v22.8h\n"
159 "add x22, x22, #0x10\n"
160 "uaddw v11.4s, v11.4s, v21.4h\n"
161 "uaddw2 v10.4s, v10.4s, v21.8h\n"
162 "uaddw v9.4s, v9.4s, v20.4h\n"
163 "uaddw2 v8.4s, v8.4s, v20.8h\n"
164 "uaddw v7.4s, v7.4s, v19.4h\n"
165 "uaddw2 v6.4s, v6.4s, v19.8h\n"
166 "uaddw v5.4s, v5.4s, v18.4h\n"
167 "uaddw2 v4.4s, v4.4s, v18.8h\n"
168 "uaddw v3.4s, v3.4s, v17.4h\n"
169 "uaddw2 v2.4s, v2.4s, v17.8h\n"
170 "uaddw v1.4s, v1.4s, v16.4h\n"
171 "uaddw2 v0.4s, v0.4s, v16.8h\n"
174 "uaddl v23.8h, v31.8b, v30.8b\n"
175 "uaddl2 v22.8h, v31.16b, v30.16b\n"
176 "uaddl v21.8h, v29.8b, v28.8b\n"
177 "uaddl2 v20.8h, v29.16b, v28.16b\n"
178 "uaddl v19.8h, v27.8b, v26.8b\n"
179 "uaddl2 v18.8h, v27.16b, v26.16b\n"
180 "uaddl v17.8h, v25.8b, v24.8b\n"
181 "uaddl2 v16.8h, v25.16b, v24.16b\n"
182 "uaddw v15.4s, v15.4s, v23.4h\n"
183 "uaddw2 v14.4s, v14.4s, v23.8h\n"
184 "uaddw v13.4s, v13.4s, v22.4h\n"
185 "uaddw2 v12.4s, v12.4s, v22.8h\n"
186 "uaddw v11.4s, v11.4s, v21.4h\n"
187 "uaddw2 v10.4s, v10.4s, v21.8h\n"
188 "uaddw v9.4s, v9.4s, v20.4h\n"
189 "uaddw2 v8.4s, v8.4s, v20.8h\n"
190 "uaddw v7.4s, v7.4s, v19.4h\n"
191 "uaddw2 v6.4s, v6.4s, v19.8h\n"
192 "uaddw v5.4s, v5.4s, v18.4h\n"
193 "uaddw2 v4.4s, v4.4s, v18.8h\n"
194 "uaddw v3.4s, v3.4s, v17.4h\n"
195 "uaddw2 v2.4s, v2.4s, v17.8h\n"
196 "uaddw v1.4s, v1.4s, v16.4h\n"
197 "uaddw2 v0.4s, v0.4s, v16.8h\n"
199 "ands x23, %x[n_valid_cells], #0x1\n"
202 "ldr x20, [x22], #0x8\n"
203 "ldr q16, [x20, x27]\n"
204 "uxtl v23.8h, v16.8b\n"
205 "uxtl2 v22.8h, v16.16b\n"
206 "ldr q16, [x20, x26]\n"
207 "ldr q17, [x20, x25]\n"
208 "uxtl v21.8h, v16.8b\n"
209 "uxtl2 v20.8h, v16.16b\n"
210 "ldr q16, [x20, x24]\n"
211 "uxtl v19.8h, v17.8b\n"
212 "uxtl2 v18.8h, v17.16b\n"
213 "subs x23, x23, #0x1\n"
214 "uxtl v17.8h, v16.8b\n"
215 "uxtl2 v16.8h, v16.16b\n"
216 "uaddw v15.4s, v15.4s, v23.4h\n"
217 "uaddw2 v14.4s, v14.4s, v23.8h\n"
218 "uaddw v13.4s, v13.4s, v22.4h\n"
219 "uaddw2 v12.4s, v12.4s, v22.8h\n"
220 "uaddw v11.4s, v11.4s, v21.4h\n"
221 "uaddw2 v10.4s, v10.4s, v21.8h\n"
222 "uaddw v9.4s, v9.4s, v20.4h\n"
223 "uaddw2 v8.4s, v8.4s, v20.8h\n"
224 "uaddw v7.4s, v7.4s, v19.4h\n"
225 "uaddw2 v6.4s, v6.4s, v19.8h\n"
226 "uaddw v5.4s, v5.4s, v18.4h\n"
227 "uaddw2 v4.4s, v4.4s, v18.8h\n"
228 "uaddw v3.4s, v3.4s, v17.4h\n"
229 "uaddw2 v2.4s, v2.4s, v17.8h\n"
230 "uaddw v1.4s, v1.4s, v16.4h\n"
231 "uaddw2 v0.4s, v0.4s, v16.8h\n"
234 "ld1r { v17.4s }, [%x[rescale_ptr]]\n"
235 "ld1r { v16.4s }, [%x[shift_ptr]]\n"
236 "sqdmulh v15.4s, v15.4s, v17.4s\n"
237 "sqdmulh v14.4s, v14.4s, v17.4s\n"
238 "sqdmulh v13.4s, v13.4s, v17.4s\n"
239 "sqdmulh v12.4s, v12.4s, v17.4s\n"
240 "sub %x[n_channels], %x[n_channels], #0x40\n"
241 "cmp %x[n_channels], #0x40\n"
242 "sqdmulh v11.4s, v11.4s, v17.4s\n"
243 "sqdmulh v10.4s, v10.4s, v17.4s\n"
244 "sqdmulh v9.4s, v9.4s, v17.4s\n"
245 "sqdmulh v8.4s, v8.4s, v17.4s\n"
246 "sqdmulh v7.4s, v7.4s, v17.4s\n"
247 "sqdmulh v6.4s, v6.4s, v17.4s\n"
248 "sqdmulh v5.4s, v5.4s, v17.4s\n"
249 "sqdmulh v4.4s, v4.4s, v17.4s\n"
250 "sqdmulh v3.4s, v3.4s, v17.4s\n"
251 "sqdmulh v2.4s, v2.4s, v17.4s\n"
252 "sqdmulh v1.4s, v1.4s, v17.4s\n"
253 "sqdmulh v0.4s, v0.4s, v17.4s\n"
254 "srshl v15.4s, v15.4s, v16.4s\n"
255 "srshl v14.4s, v14.4s, v16.4s\n"
256 "srshl v13.4s, v13.4s, v16.4s\n"
257 "srshl v12.4s, v12.4s, v16.4s\n"
258 "srshl v11.4s, v11.4s, v16.4s\n"
259 "srshl v10.4s, v10.4s, v16.4s\n"
260 "srshl v9.4s, v9.4s, v16.4s\n"
261 "srshl v8.4s, v8.4s, v16.4s\n"
262 "srshl v7.4s, v7.4s, v16.4s\n"
263 "srshl v6.4s, v6.4s, v16.4s\n"
264 "srshl v5.4s, v5.4s, v16.4s\n"
265 "srshl v4.4s, v4.4s, v16.4s\n"
266 "srshl v3.4s, v3.4s, v16.4s\n"
267 "srshl v2.4s, v2.4s, v16.4s\n"
268 "srshl v1.4s, v1.4s, v16.4s\n"
269 "srshl v0.4s, v0.4s, v16.4s\n"
270 "movi v16.4s, #0x0\n"
271 "smax v15.4s, v15.4s, v16.4s\n"
272 "smax v14.4s, v14.4s, v16.4s\n"
273 "smax v13.4s, v13.4s, v16.4s\n"
274 "smax v12.4s, v12.4s, v16.4s\n"
275 "smax v11.4s, v11.4s, v16.4s\n"
276 "smax v10.4s, v10.4s, v16.4s\n"
277 "smax v9.4s, v9.4s, v16.4s\n"
278 "smax v8.4s, v8.4s, v16.4s\n"
279 "smax v7.4s, v7.4s, v16.4s\n"
280 "smax v6.4s, v6.4s, v16.4s\n"
281 "smax v5.4s, v5.4s, v16.4s\n"
282 "smax v4.4s, v4.4s, v16.4s\n"
283 "smax v3.4s, v3.4s, v16.4s\n"
284 "smax v2.4s, v2.4s, v16.4s\n"
285 "smax v1.4s, v1.4s, v16.4s\n"
286 "smax v0.4s, v0.4s, v16.4s\n"
287 "movi v16.4s, #0xff\n"
288 "smin v15.4s, v15.4s, v16.4s\n"
289 "smin v14.4s, v14.4s, v16.4s\n"
290 "smin v13.4s, v13.4s, v16.4s\n"
291 "smin v12.4s, v12.4s, v16.4s\n"
292 "smin v11.4s, v11.4s, v16.4s\n"
293 "smin v10.4s, v10.4s, v16.4s\n"
294 "smin v9.4s, v9.4s, v16.4s\n"
295 "smin v8.4s, v8.4s, v16.4s\n"
296 "smin v7.4s, v7.4s, v16.4s\n"
297 "smin v6.4s, v6.4s, v16.4s\n"
298 "smin v5.4s, v5.4s, v16.4s\n"
299 "smin v4.4s, v4.4s, v16.4s\n"
300 "smin v3.4s, v3.4s, v16.4s\n"
301 "smin v2.4s, v2.4s, v16.4s\n"
302 "smin v1.4s, v1.4s, v16.4s\n"
303 "smin v0.4s, v0.4s, v16.4s\n"
304 "uzp1 v23.16b, v15.16b, v14.16b\n"
305 "uzp1 v16.16b, v13.16b, v12.16b\n"
306 "uzp1 v22.16b, v11.16b, v10.16b\n"
307 "uzp1 v18.16b, v9.16b, v8.16b\n"
308 "uzp1 v21.16b, v7.16b, v6.16b\n"
309 "uzp1 v17.16b, v5.16b, v4.16b\n"
310 "uzp1 v20.16b, v3.16b, v2.16b\n"
311 "uzp1 v19.16b, v1.16b, v0.16b\n"
312 "uzp1 v16.16b, v23.16b, v16.16b\n"
313 "uzp1 v18.16b, v22.16b, v18.16b\n"
314 "str q16, [%x[outptr], x27]\n"
315 "add x27, x27, #0x40\n"
316 "uzp1 v17.16b, v21.16b, v17.16b\n"
317 "uzp1 v16.16b, v20.16b, v19.16b\n"
318 "str q18, [%x[outptr], x26]\n"
319 "add x26, x26, #0x40\n"
320 "str q17, [%x[outptr], x25]\n"
321 "add x25, x25, #0x40\n"
322 "str q16, [%x[outptr], x24]\n"
323 "add x24, x24, #0x40\n"
325 "cbz %x[n_channels], 43f\n"
327 "cmp %x[n_channels], #0x10\n"
330 "lsr x23, %x[n_valid_cells], #0x1\n"
331 "movi v15.4s, #0x0\n"
332 "movi v14.4s, #0x0\n"
333 "mov x22, %x[inptrs]\n"
334 "movi v13.4s, #0x0\n"
335 "movi v12.4s, #0x0\n"
337 "ldp x21, x20, [x22, #0x0]\n"
338 "ldr q31, [x21, x27]\n"
339 "subs x23, x23, #0x1\n"
340 "add x22, x22, #0x10\n"
341 "ldr q30, [x20, x27]\n"
344 "uaddl v17.8h, v31.8b, v30.8b\n"
345 "uaddl2 v16.8h, v31.16b, v30.16b\n"
346 "ldp x21, x20, [x22, #0x0]\n"
347 "ldr q31, [x21, x27]\n"
348 "ldr q30, [x20, x27]\n"
349 "subs x23, x23, #0x1\n"
350 "uaddw v15.4s, v15.4s, v17.4h\n"
351 "uaddw2 v14.4s, v14.4s, v17.8h\n"
352 "uaddw v13.4s, v13.4s, v16.4h\n"
353 "uaddw2 v12.4s, v12.4s, v16.8h\n"
354 "add x22, x22, #0x10\n"
357 "uaddl v17.8h, v31.8b, v30.8b\n"
358 "uaddl2 v16.8h, v31.16b, v30.16b\n"
359 "uaddw v15.4s, v15.4s, v17.4h\n"
360 "uaddw2 v14.4s, v14.4s, v17.8h\n"
361 "uaddw v13.4s, v13.4s, v16.4h\n"
362 "uaddw2 v12.4s, v12.4s, v16.8h\n"
364 "ands x23, %x[n_valid_cells], #0x1\n"
367 "ldr x20, [x22], #0x8\n"
368 "ldr q16, [x20, x27]\n"
369 "uxtl v17.8h, v16.8b\n"
370 "uxtl2 v16.8h, v16.16b\n"
371 "subs x23, x23, #0x1\n"
372 "uaddw v15.4s, v15.4s, v17.4h\n"
373 "uaddw2 v14.4s, v14.4s, v17.8h\n"
374 "uaddw v13.4s, v13.4s, v16.4h\n"
375 "uaddw2 v12.4s, v12.4s, v16.8h\n"
378 "ld1r { v17.4s }, [%x[rescale_ptr]]\n"
379 "ld1r { v16.4s }, [%x[shift_ptr]]\n"
380 "sqdmulh v15.4s, v15.4s, v17.4s\n"
381 "sqdmulh v14.4s, v14.4s, v17.4s\n"
382 "sqdmulh v13.4s, v13.4s, v17.4s\n"
383 "sqdmulh v12.4s, v12.4s, v17.4s\n"
384 "sub %x[n_channels], %x[n_channels], #0x10\n"
385 "cmp %x[n_channels], #0x10\n"
386 "srshl v15.4s, v15.4s, v16.4s\n"
387 "srshl v14.4s, v14.4s, v16.4s\n"
388 "srshl v13.4s, v13.4s, v16.4s\n"
389 "srshl v12.4s, v12.4s, v16.4s\n"
390 "movi v16.4s, #0x0\n"
391 "smax v15.4s, v15.4s, v16.4s\n"
392 "smax v14.4s, v14.4s, v16.4s\n"
393 "smax v13.4s, v13.4s, v16.4s\n"
394 "smax v12.4s, v12.4s, v16.4s\n"
395 "movi v16.4s, #0xff\n"
396 "smin v15.4s, v15.4s, v16.4s\n"
397 "smin v14.4s, v14.4s, v16.4s\n"
398 "smin v13.4s, v13.4s, v16.4s\n"
399 "smin v12.4s, v12.4s, v16.4s\n"
400 "uzp1 v17.16b, v15.16b, v14.16b\n"
401 "uzp1 v16.16b, v13.16b, v12.16b\n"
402 "uzp1 v16.16b, v17.16b, v16.16b\n"
403 "str q16, [%x[outptr], x27]\n"
404 "add x27, x27, #0x10\n"
406 "cbz %x[n_channels], 43f\n"
408 "lsr x23, %x[n_valid_cells], #0x1\n"
409 "add %x[outptr], %x[outptr], x27\n"
410 "movi v15.4s, #0x0\n"
411 "movi v14.4s, #0x0\n"
412 "movi v13.4s, #0x0\n"
413 "movi v12.4s, #0x0\n"
414 "mov x22, %x[inptrs]\n"
417 "ldp x21, x20, [x22, #0x0]\n"
418 "add x22, x22, #0x10\n"
419 "add x21, x21, x27\n"
420 "movi v31.16b, #0x0\n"
421 "add x20, x20, x27\n"
422 "movi v30.16b, #0x0\n"
423 "tbz %x[n_channels], #3, 19f\n"
424 "ldr d31, [x21], #0x8\n"
425 "ldr d30, [x20], #0x8\n"
426 "tbz %x[n_channels], #2, 17f\n"
427 "ld1 { v31.s }[2], [x21], #0x4\n"
428 "ld1 { v30.s }[2], [x20], #0x4\n"
429 "tbz %x[n_channels], #1, 16f\n"
430 "ld1 { v31.h }[6], [x21], #0x2\n"
431 "ld1 { v30.h }[6], [x20], #0x2\n"
432 "tbz %x[n_channels], #0, 23f\n"
433 "ld1 { v31.b }[14], [x21], #0x1\n"
434 "ld1 { v30.b }[14], [x20], #0x1\n"
437 "tbz %x[n_channels], #0, 23f\n"
438 "ld1 { v31.b }[12], [x21], #0x1\n"
439 "ld1 { v30.b }[12], [x20], #0x1\n"
442 "tbz %x[n_channels], #1, 18f\n"
443 "ld1 { v31.h }[4], [x21], #0x2\n"
444 "ld1 { v30.h }[4], [x20], #0x2\n"
445 "tbz %x[n_channels], #0, 23f\n"
446 "ld1 { v31.b }[10], [x21], #0x1\n"
447 "ld1 { v30.b }[10], [x20], #0x1\n"
450 "tbz %x[n_channels], #0, 23f\n"
451 "ld1 { v31.b }[8], [x21], #0x1\n"
452 "ld1 { v30.b }[8], [x20], #0x1\n"
455 "tbz %x[n_channels], #2, 21f\n"
456 "ldr s31, [x21], #0x4\n"
457 "ldr s30, [x20], #0x4\n"
458 "tbz %x[n_channels], #1, 20f\n"
459 "ld1 { v31.h }[2], [x21], #0x2\n"
460 "ld1 { v30.h }[2], [x20], #0x2\n"
461 "tbz %x[n_channels], #0, 23f\n"
462 "ld1 { v31.b }[6], [x21], #0x1\n"
463 "ld1 { v30.b }[6], [x20], #0x1\n"
466 "tbz %x[n_channels], #0, 23f\n"
467 "ld1 { v31.b }[4], [x21], #0x1\n"
468 "ld1 { v30.b }[4], [x20], #0x1\n"
471 "tbz %x[n_channels], #1, 22f\n"
472 "ldr h31, [x21], #0x2\n"
473 "ldr h30, [x20], #0x2\n"
474 "tbz %x[n_channels], #0, 23f\n"
475 "ld1 { v31.b }[2], [x21], #0x1\n"
476 "ld1 { v30.b }[2], [x20], #0x1\n"
479 "tbz %x[n_channels], #0, 23f\n"
480 "ldr b31, [x21], #0x1\n"
481 "ldr b30, [x20], #0x1\n"
483 "uaddl v17.8h, v31.8b, v30.8b\n"
484 "uaddl2 v16.8h, v31.16b, v30.16b\n"
485 "subs x23, x23, #0x1\n"
486 "uaddw v15.4s, v15.4s, v17.4h\n"
487 "uaddw2 v14.4s, v14.4s, v17.8h\n"
488 "uaddw v13.4s, v13.4s, v16.4h\n"
489 "uaddw2 v12.4s, v12.4s, v16.8h\n"
492 "ands x23, %x[n_valid_cells], #0x1\n"
495 "ldr x21, [x22], #0x8\n"
496 "add x21, x21, x27\n"
497 "movi v31.16b, #0x0\n"
498 "tbz %x[n_channels], #3, 29f\n"
499 "ldr d31, [x21], #0x8\n"
500 "tbz %x[n_channels], #2, 27f\n"
501 "ld1 { v31.s }[2], [x21], #0x4\n"
502 "tbz %x[n_channels], #1, 26f\n"
503 "ld1 { v31.h }[6], [x21], #0x2\n"
504 "tbz %x[n_channels], #0, 33f\n"
505 "ld1 { v31.b }[14], [x21], #0x1\n"
508 "tbz %x[n_channels], #0, 33f\n"
509 "ld1 { v31.b }[12], [x21], #0x1\n"
512 "tbz %x[n_channels], #1, 28f\n"
513 "ld1 { v31.h }[4], [x21], #0x2\n"
514 "tbz %x[n_channels], #0, 33f\n"
515 "ld1 { v31.b }[10], [x21], #0x1\n"
518 "tbz %x[n_channels], #0, 33f\n"
519 "ld1 { v31.b }[8], [x21], #0x1\n"
522 "tbz %x[n_channels], #2, 31f\n"
523 "ldr s31, [x21], #0x4\n"
524 "tbz %x[n_channels], #1, 30f\n"
525 "ld1 { v31.h }[2], [x21], #0x2\n"
526 "tbz %x[n_channels], #0, 33f\n"
527 "ld1 { v31.b }[6], [x21], #0x1\n"
530 "tbz %x[n_channels], #0, 33f\n"
531 "ld1 { v31.b }[4], [x21], #0x1\n"
534 "tbz %x[n_channels], #1, 32f\n"
535 "ldr h31, [x21], #0x2\n"
536 "tbz %x[n_channels], #0, 33f\n"
537 "ld1 { v31.b }[2], [x21], #0x1\n"
540 "tbz %x[n_channels], #0, 33f\n"
541 "ldr b31, [x21], #0x1\n"
543 "uxtl v17.8h, v31.8b\n"
544 "uxtl2 v16.8h, v31.16b\n"
545 "subs x23, x23, #0x1\n"
546 "uaddw v15.4s, v15.4s, v17.4h\n"
547 "uaddw2 v14.4s, v14.4s, v17.8h\n"
548 "uaddw v13.4s, v13.4s, v16.4h\n"
549 "uaddw2 v12.4s, v12.4s, v16.8h\n"
552 "ld1r { v17.4s }, [%x[rescale_ptr]]\n"
553 "ld1r { v16.4s }, [%x[shift_ptr]]\n"
554 "sqdmulh v15.4s, v15.4s, v17.4s\n"
555 "sqdmulh v14.4s, v14.4s, v17.4s\n"
556 "sqdmulh v13.4s, v13.4s, v17.4s\n"
557 "sqdmulh v12.4s, v12.4s, v17.4s\n"
558 "srshl v15.4s, v15.4s, v16.4s\n"
559 "srshl v14.4s, v14.4s, v16.4s\n"
560 "srshl v13.4s, v13.4s, v16.4s\n"
561 "srshl v12.4s, v12.4s, v16.4s\n"
562 "movi v16.4s, #0x0\n"
563 "smax v15.4s, v15.4s, v16.4s\n"
564 "smax v14.4s, v14.4s, v16.4s\n"
565 "smax v13.4s, v13.4s, v16.4s\n"
566 "smax v12.4s, v12.4s, v16.4s\n"
567 "movi v16.4s, #0xff\n"
568 "smin v15.4s, v15.4s, v16.4s\n"
569 "smin v14.4s, v14.4s, v16.4s\n"
570 "smin v13.4s, v13.4s, v16.4s\n"
571 "smin v12.4s, v12.4s, v16.4s\n"
572 "uzp1 v17.16b, v15.16b, v14.16b\n"
573 "uzp1 v16.16b, v13.16b, v12.16b\n"
574 "uzp1 v16.16b, v17.16b, v16.16b\n"
575 "tbz %x[n_channels], #3, 38f\n"
576 "st1 { v16.d }[0], [%x[outptr]], #0x8\n"
577 "tbz %x[n_channels], #2, 36f\n"
578 "st1 { v16.s }[2], [%x[outptr]], #0x4\n"
579 "tbz %x[n_channels], #1, 35f\n"
580 "st1 { v16.h }[6], [%x[outptr]], #0x2\n"
581 "tbz %x[n_channels], #0, 42f\n"
582 "st1 { v16.b }[14], [%x[outptr]], #0x1\n"
585 "tbz %x[n_channels], #0, 42f\n"
586 "st1 { v16.b }[12], [%x[outptr]], #0x1\n"
589 "tbz %x[n_channels], #1, 37f\n"
590 "st1 { v16.h }[4], [%x[outptr]], #0x2\n"
591 "tbz %x[n_channels], #0, 42f\n"
592 "st1 { v16.b }[10], [%x[outptr]], #0x1\n"
595 "tbz %x[n_channels], #0, 42f\n"
596 "st1 { v16.b }[8], [%x[outptr]], #0x1\n"
599 "tbz %x[n_channels], #2, 40f\n"
600 "st1 { v16.s }[0], [%x[outptr]], #0x4\n"
601 "tbz %x[n_channels], #1, 39f\n"
602 "st1 { v16.h }[2], [%x[outptr]], #0x2\n"
603 "tbz %x[n_channels], #0, 42f\n"
604 "st1 { v16.b }[6], [%x[outptr]], #0x1\n"
607 "tbz %x[n_channels], #0, 42f\n"
608 "st1 { v16.b }[4], [%x[outptr]], #0x1\n"
611 "tbz %x[n_channels], #1, 41f\n"
612 "st1 { v16.h }[0], [%x[outptr]], #0x2\n"
613 "tbz %x[n_channels], #0, 42f\n"
614 "st1 { v16.b }[2], [%x[outptr]], #0x1\n"
617 "tbz %x[n_channels], #0, 42f\n"
618 "st1 { v16.b }[0], [%x[outptr]], #0x1\n"
621 : [n_channels]
"+&r" (n_channels), [outptr]
"+&r" (outptr)
622 : [inptrs]
"r" (inptrs), [n_valid_cells]
"r" (n_valid_cells), [rescale_ptr]
"r" (&rescale_value), [shift_ptr]
"r" (&shift_value)
623 :
"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",
"x20",
"x21",
"x22",
"x23",
"x24",
"x25",
"x26",
"x27"
630 #endif // defined(__aarch64__)