27 #if defined(__aarch64__)
31 void a64_transpose_interleave_32_1x4(uint8_t *out,
const uint8_t *in,
size_t width,
size_t in_stride,
size_t height)
33 uint8_t *pad_row =
reinterpret_cast<uint8_t *
>(alloca(width *
sizeof(uint8_t)));
36 memset(pad_row, 0, width *
sizeof(uint8_t));
39 size_t out_stride = 32 * roundup<size_t>(height, 4) *
sizeof(uint8_t);
42 "cmp %x[height], #0x10\n"
46 "add x16, x17, %x[in_stride]\n"
47 "add x15, x16, %x[in_stride]\n"
48 "add x14, x15, %x[in_stride]\n"
49 "add x13, x14, %x[in_stride]\n"
50 "add x12, x13, %x[in_stride]\n"
51 "add x11, x12, %x[in_stride]\n"
52 "add x10, x11, %x[in_stride]\n"
53 "add x9, x10, %x[in_stride]\n"
54 "add x28, x9, %x[in_stride]\n"
55 "add x27, x28, %x[in_stride]\n"
56 "add x26, x27, %x[in_stride]\n"
57 "add x25, x26, %x[in_stride]\n"
58 "mov x24, %x[width]\n"
59 "add x23, x25, %x[in_stride]\n"
60 "add x22, x23, %x[in_stride]\n"
61 "add x20, x22, %x[in_stride]\n"
63 "add %x[in], x20, %x[in_stride]\n"
65 "sub %x[height], %x[height], #0x10\n"
68 "ldr q6, [x17], #0x10\n"
69 "ldr q31, [x16], #0x10\n"
70 "sub x24, x24, #0x20\n"
72 "ldr q7, [x15], #0x10\n"
73 "ldr q0, [x14], #0x10\n"
74 "zip1 v9.16b, v6.16b, v7.16b\n"
75 "zip1 v20.16b, v31.16b, v0.16b\n"
76 "ldr q24, [x13], #0x10\n"
77 "ldr q19, [x12], #0x10\n"
78 "zip2 v30.16b, v6.16b, v7.16b\n"
79 "zip2 v12.16b, v31.16b, v0.16b\n"
80 "ldr q23, [x11], #0x10\n"
81 "ldr q17, [x10], #0x10\n"
82 "zip1 v13.16b, v24.16b, v23.16b\n"
83 "zip1 v16.16b, v19.16b, v17.16b\n"
84 "ldr q0, [x9], #0x10\n"
85 "ldr q31, [x28], #0x10\n"
86 "zip2 v15.16b, v24.16b, v23.16b\n"
87 "zip2 v11.16b, v19.16b, v17.16b\n"
88 "ldr q17, [x27], #0x10\n"
89 "ldr q4, [x26], #0x10\n"
90 "zip1 v1.16b, v0.16b, v17.16b\n"
91 "zip1 v21.16b, v31.16b, v4.16b\n"
92 "ldr q28, [x25], #0x10\n"
93 "ldr q18, [x23], #0x10\n"
94 "zip2 v0.16b, v0.16b, v17.16b\n"
95 "zip2 v26.16b, v31.16b, v4.16b\n"
96 "ldr q17, [x22], #0x10\n"
97 "ldr q19, [x20], #0x10\n"
98 "zip1 v23.16b, v28.16b, v17.16b\n"
99 "zip1 v25.16b, v18.16b, v19.16b\n"
100 "ldr q2, [x17], #0x10\n"
101 "ldr q3, [x16], #0x10\n"
102 "zip2 v7.16b, v28.16b, v17.16b\n"
103 "zip2 v8.16b, v18.16b, v19.16b\n"
104 "ldr q22, [x15], #0x10\n"
105 "ldr q27, [x14], #0x10\n"
106 "zip1 v19.16b, v2.16b, v22.16b\n"
107 "zip1 v17.16b, v3.16b, v27.16b\n"
108 "ldr q6, [x13], #0x10\n"
109 "ldr q4, [x12], #0x10\n"
110 "zip2 v24.16b, v2.16b, v22.16b\n"
111 "zip2 v22.16b, v3.16b, v27.16b\n"
112 "ldr q14, [x11], #0x10\n"
113 "ldr q18, [x10], #0x10\n"
114 "zip1 v29.16b, v6.16b, v14.16b\n"
115 "zip1 v31.16b, v4.16b, v18.16b\n"
116 "ldr q2, [x9], #0x10\n"
117 "ldr q10, [x28], #0x10\n"
118 "zip2 v28.16b, v6.16b, v14.16b\n"
119 "zip2 v27.16b, v4.16b, v18.16b\n"
120 "ldr q6, [x27], #0x10\n"
121 "ldr q5, [x26], #0x10\n"
122 "zip1 v14.16b, v2.16b, v6.16b\n"
123 "zip1 v4.16b, v10.16b, v5.16b\n"
124 "ldr q3, [x25], #0x10\n"
125 "ldr q18, [x23], #0x10\n"
126 "zip2 v6.16b, v2.16b, v6.16b\n"
127 "zip2 v10.16b, v10.16b, v5.16b\n"
128 "ldr q5, [x22], #0x10\n"
129 "zip1 v2.16b, v3.16b, v5.16b\n"
130 "zip2 v3.16b, v3.16b, v5.16b\n"
131 "zip1 v5.16b, v9.16b, v20.16b\n"
132 "zip2 v20.16b, v9.16b, v20.16b\n"
133 "ldr q9, [x20], #0x10\n"
134 "str q5, [x21, #0x0]\n"
135 "zip1 v5.16b, v18.16b, v9.16b\n"
136 "zip2 v9.16b, v18.16b, v9.16b\n"
137 "str q20, [x21, #0x10]\n"
138 "zip1 v18.16b, v30.16b, v12.16b\n"
139 "zip2 v30.16b, v30.16b, v12.16b\n"
140 "str q18, [x21, #0x20]\n"
141 "zip1 v20.16b, v19.16b, v17.16b\n"
142 "zip2 v12.16b, v19.16b, v17.16b\n"
143 "str q30, [x21, #0x30]\n"
144 "zip1 v18.16b, v24.16b, v22.16b\n"
145 "zip2 v17.16b, v24.16b, v22.16b\n"
146 "str q20, [x21, #0x40]\n"
147 "zip1 v30.16b, v13.16b, v16.16b\n"
148 "zip2 v24.16b, v13.16b, v16.16b\n"
149 "str q12, [x21, #0x50]\n"
150 "zip1 v22.16b, v15.16b, v11.16b\n"
151 "zip2 v20.16b, v15.16b, v11.16b\n"
152 "str q18, [x21, #0x60]\n"
153 "zip1 v19.16b, v29.16b, v31.16b\n"
154 "zip2 v18.16b, v29.16b, v31.16b\n"
155 "str q17, [x21, #0x70]\n"
156 "zip1 v17.16b, v28.16b, v27.16b\n"
157 "zip2 v16.16b, v28.16b, v27.16b\n"
158 "str q30, [x21, #0x80]\n"
159 "zip1 v31.16b, v1.16b, v21.16b\n"
160 "zip2 v1.16b, v1.16b, v21.16b\n"
161 "str q24, [x21, #0x90]\n"
162 "zip1 v30.16b, v0.16b, v26.16b\n"
163 "zip2 v29.16b, v0.16b, v26.16b\n"
164 "str q22, [x21, #0xa0]\n"
165 "zip1 v28.16b, v14.16b, v4.16b\n"
166 "zip2 v27.16b, v14.16b, v4.16b\n"
167 "str q20, [x21, #0xb0]\n"
168 "zip1 v26.16b, v6.16b, v10.16b\n"
169 "zip2 v24.16b, v6.16b, v10.16b\n"
170 "str q19, [x21, #0xc0]\n"
171 "zip1 v14.16b, v23.16b, v25.16b\n"
172 "zip2 v22.16b, v23.16b, v25.16b\n"
173 "str q18, [x21, #0xd0]\n"
174 "zip1 v21.16b, v7.16b, v8.16b\n"
175 "zip2 v20.16b, v7.16b, v8.16b\n"
176 "str q17, [x21, #0xe0]\n"
177 "zip1 v19.16b, v2.16b, v5.16b\n"
178 "zip2 v18.16b, v2.16b, v5.16b\n"
179 "str q16, [x21, #0xf0]\n"
180 "zip1 v17.16b, v3.16b, v9.16b\n"
181 "zip2 v16.16b, v3.16b, v9.16b\n"
182 "str q31, [x21, #0x100]\n"
183 "str q1, [x21, #0x110]\n"
184 "str q30, [x21, #0x120]\n"
185 "str q29, [x21, #0x130]\n"
186 "str q28, [x21, #0x140]\n"
187 "str q27, [x21, #0x150]\n"
188 "str q26, [x21, #0x160]\n"
189 "str q24, [x21, #0x170]\n"
190 "str q14, [x21, #0x180]\n"
191 "str q22, [x21, #0x190]\n"
192 "str q21, [x21, #0x1a0]\n"
193 "str q20, [x21, #0x1b0]\n"
194 "str q19, [x21, #0x1c0]\n"
195 "str q18, [x21, #0x1d0]\n"
196 "str q17, [x21, #0x1e0]\n"
197 "str q16, [x21, #0x1f0]\n"
198 "add x21, x21, %x[out_stride]\n"
204 "ldr q21, [x17], #0x10\n"
205 "ldr q20, [x16], #0x10\n"
206 "sub x24, x24, #0x10\n"
208 "ldr q17, [x15], #0x10\n"
209 "ldr q16, [x14], #0x10\n"
210 "zip1 v3.16b, v21.16b, v17.16b\n"
211 "zip1 v2.16b, v20.16b, v16.16b\n"
212 "ldr q19, [x13], #0x10\n"
213 "ldr q18, [x12], #0x10\n"
214 "zip2 v1.16b, v21.16b, v17.16b\n"
215 "zip2 v0.16b, v20.16b, v16.16b\n"
216 "ldr q17, [x11], #0x10\n"
217 "ldr q16, [x10], #0x10\n"
218 "zip1 v31.16b, v19.16b, v17.16b\n"
219 "zip1 v30.16b, v18.16b, v16.16b\n"
220 "ldr q24, [x9], #0x10\n"
221 "ldr q20, [x28], #0x10\n"
222 "zip2 v29.16b, v19.16b, v17.16b\n"
223 "zip2 v23.16b, v18.16b, v16.16b\n"
224 "ldr q17, [x27], #0x10\n"
225 "ldr q16, [x26], #0x10\n"
226 "zip1 v22.16b, v24.16b, v17.16b\n"
227 "zip1 v21.16b, v20.16b, v16.16b\n"
228 "ldr q19, [x25], #0x10\n"
229 "ldr q18, [x23], #0x10\n"
230 "zip2 v28.16b, v24.16b, v17.16b\n"
231 "zip2 v20.16b, v20.16b, v16.16b\n"
232 "ldr q17, [x22], #0x10\n"
233 "ldr q16, [x20], #0x10\n"
234 "zip1 v27.16b, v19.16b, v17.16b\n"
235 "zip1 v26.16b, v18.16b, v16.16b\n"
236 "zip2 v25.16b, v19.16b, v17.16b\n"
237 "zip2 v24.16b, v18.16b, v16.16b\n"
238 "zip1 v16.16b, v3.16b, v2.16b\n"
239 "zip2 v18.16b, v3.16b, v2.16b\n"
240 "str q16, [x21, #0x0]\n"
241 "zip1 v17.16b, v1.16b, v0.16b\n"
242 "zip2 v16.16b, v1.16b, v0.16b\n"
243 "str q18, [x21, #0x10]\n"
244 "str q17, [x21, #0x20]\n"
245 "zip1 v19.16b, v31.16b, v30.16b\n"
246 "zip2 v18.16b, v31.16b, v30.16b\n"
247 "str q16, [x21, #0x30]\n"
248 "zip1 v17.16b, v29.16b, v23.16b\n"
249 "zip2 v16.16b, v29.16b, v23.16b\n"
250 "str q19, [x21, #0x80]\n"
251 "zip1 v23.16b, v22.16b, v21.16b\n"
252 "zip2 v22.16b, v22.16b, v21.16b\n"
253 "str q18, [x21, #0x90]\n"
254 "zip1 v21.16b, v28.16b, v20.16b\n"
255 "zip2 v20.16b, v28.16b, v20.16b\n"
256 "str q17, [x21, #0xa0]\n"
257 "zip1 v19.16b, v27.16b, v26.16b\n"
258 "zip2 v18.16b, v27.16b, v26.16b\n"
259 "str q16, [x21, #0xb0]\n"
260 "zip1 v17.16b, v25.16b, v24.16b\n"
261 "zip2 v16.16b, v25.16b, v24.16b\n"
262 "str q23, [x21, #0x100]\n"
263 "str q22, [x21, #0x110]\n"
264 "str q21, [x21, #0x120]\n"
265 "str q20, [x21, #0x130]\n"
266 "str q19, [x21, #0x180]\n"
267 "str q18, [x21, #0x190]\n"
268 "str q17, [x21, #0x1a0]\n"
269 "str q16, [x21, #0x1b0]\n"
270 "add x21, x21, #0x40\n"
276 "ldr s19, [x17], #0x4\n"
277 "ldr s18, [x16], #0x4\n"
278 "sub x24, x24, #0x4\n"
280 "ldr s17, [x15], #0x4\n"
281 "ldr s16, [x14], #0x4\n"
282 "zip1 v17.16b, v19.16b, v17.16b\n"
283 "zip1 v16.16b, v18.16b, v16.16b\n"
284 "ldr s19, [x13], #0x4\n"
285 "ldr s18, [x12], #0x4\n"
286 "zip1 v22.16b, v17.16b, v16.16b\n"
287 "ldr s17, [x11], #0x4\n"
288 "ldr s16, [x10], #0x4\n"
289 "zip1 v17.16b, v19.16b, v17.16b\n"
290 "zip1 v16.16b, v18.16b, v16.16b\n"
291 "ldr s19, [x9], #0x4\n"
292 "ldr s18, [x28], #0x4\n"
293 "zip1 v21.16b, v17.16b, v16.16b\n"
294 "ldr s17, [x27], #0x4\n"
295 "ldr s16, [x26], #0x4\n"
296 "zip1 v17.16b, v19.16b, v17.16b\n"
297 "zip1 v16.16b, v18.16b, v16.16b\n"
298 "ldr s20, [x25], #0x4\n"
299 "ldr s19, [x23], #0x4\n"
300 "zip1 v18.16b, v17.16b, v16.16b\n"
301 "ldr s17, [x22], #0x4\n"
302 "ldr s16, [x20], #0x4\n"
303 "zip1 v17.16b, v20.16b, v17.16b\n"
304 "zip1 v16.16b, v19.16b, v16.16b\n"
305 "str q22, [x21, #0x0]\n"
306 "zip1 v16.16b, v17.16b, v16.16b\n"
307 "str q21, [x21, #0x80]\n"
308 "str q18, [x21, #0x100]\n"
309 "str q16, [x21, #0x180]\n"
310 "add x21, x21, #0x10\n"
316 "ldr b19, [x17], #0x1\n"
317 "ldr b18, [x16], #0x1\n"
318 "sub x24, x24, #0x1\n"
320 "ldr b17, [x15], #0x1\n"
321 "ldr b16, [x14], #0x1\n"
322 "zip1 v17.16b, v19.16b, v17.16b\n"
323 "zip1 v16.16b, v18.16b, v16.16b\n"
324 "ldr b19, [x13], #0x1\n"
325 "ldr b18, [x12], #0x1\n"
326 "zip1 v22.16b, v17.16b, v16.16b\n"
327 "ldr b17, [x11], #0x1\n"
328 "ldr b16, [x10], #0x1\n"
329 "zip1 v17.16b, v19.16b, v17.16b\n"
330 "zip1 v16.16b, v18.16b, v16.16b\n"
331 "ldr b19, [x9], #0x1\n"
332 "ldr b18, [x28], #0x1\n"
333 "zip1 v21.16b, v17.16b, v16.16b\n"
334 "ldr b17, [x27], #0x1\n"
335 "ldr b16, [x26], #0x1\n"
336 "zip1 v17.16b, v19.16b, v17.16b\n"
337 "zip1 v16.16b, v18.16b, v16.16b\n"
338 "ldr b20, [x25], #0x1\n"
339 "ldr b19, [x23], #0x1\n"
340 "zip1 v18.16b, v17.16b, v16.16b\n"
341 "ldr b17, [x22], #0x1\n"
342 "ldr b16, [x20], #0x1\n"
343 "zip1 v17.16b, v20.16b, v17.16b\n"
344 "zip1 v16.16b, v19.16b, v16.16b\n"
345 "str s22, [x21, #0x0]\n"
346 "zip1 v16.16b, v17.16b, v16.16b\n"
347 "str s21, [x21, #0x80]\n"
348 "str s18, [x21, #0x100]\n"
349 "str s16, [x21, #0x180]\n"
350 "add x21, x21, #0x4\n"
353 "cmp %x[height], #0x10\n"
354 "add %x[out], %x[out], #0x200\n"
356 "cbz %x[height], 20f\n"
360 "add x16, x17, %x[in_stride]\n"
361 "add x15, x16, %x[in_stride]\n"
362 "mov x20, %x[width]\n"
363 "add x14, x15, %x[in_stride]\n"
364 "cmp %x[height], #0x3\n"
365 "add %x[in], x14, %x[in_stride]\n"
366 "csel x14, x14, %x[pad_row], GT\n"
367 "csel x15, x15, %x[pad_row], GE\n"
368 "cmp %x[height], #0x1\n"
369 "csel x16, x16, %x[pad_row], GT\n"
372 "sub %x[height], %x[height], #0x4\n"
375 "ldr q19, [x17], #0x10\n"
376 "ldr q18, [x16], #0x10\n"
377 "sub x20, x20, #0x20\n"
379 "ldr q17, [x15], #0x10\n"
380 "ldr q16, [x14], #0x10\n"
381 "zip1 v25.16b, v19.16b, v17.16b\n"
382 "zip1 v24.16b, v18.16b, v16.16b\n"
383 "ldr q22, [x17], #0x10\n"
384 "ldr q21, [x16], #0x10\n"
385 "zip2 v20.16b, v19.16b, v17.16b\n"
386 "zip2 v19.16b, v18.16b, v16.16b\n"
387 "ldr q17, [x15], #0x10\n"
388 "ldr q16, [x14], #0x10\n"
389 "zip1 v23.16b, v22.16b, v17.16b\n"
390 "zip1 v18.16b, v21.16b, v16.16b\n"
391 "zip2 v22.16b, v22.16b, v17.16b\n"
392 "zip2 v21.16b, v21.16b, v16.16b\n"
393 "zip1 v16.16b, v25.16b, v24.16b\n"
394 "zip2 v17.16b, v25.16b, v24.16b\n"
395 "str q16, [x21, #0x0]\n"
396 "zip1 v16.16b, v20.16b, v19.16b\n"
397 "zip2 v20.16b, v20.16b, v19.16b\n"
398 "str q17, [x21, #0x10]\n"
399 "zip1 v19.16b, v23.16b, v18.16b\n"
400 "zip2 v18.16b, v23.16b, v18.16b\n"
401 "str q16, [x21, #0x20]\n"
402 "zip1 v17.16b, v22.16b, v21.16b\n"
403 "zip2 v16.16b, v22.16b, v21.16b\n"
404 "str q20, [x21, #0x30]\n"
405 "str q19, [x21, #0x40]\n"
406 "str q18, [x21, #0x50]\n"
407 "str q17, [x21, #0x60]\n"
408 "str q16, [x21, #0x70]\n"
409 "add x21, x21, %x[out_stride]\n"
415 "ldr q20, [x17], #0x10\n"
416 "ldr q21, [x16], #0x10\n"
417 "sub x20, x20, #0x10\n"
419 "ldr q19, [x15], #0x10\n"
420 "ldr q16, [x14], #0x10\n"
421 "zip1 v18.16b, v20.16b, v19.16b\n"
422 "zip1 v17.16b, v21.16b, v16.16b\n"
423 "zip2 v20.16b, v20.16b, v19.16b\n"
424 "zip2 v19.16b, v21.16b, v16.16b\n"
425 "zip1 v16.16b, v18.16b, v17.16b\n"
426 "zip2 v18.16b, v18.16b, v17.16b\n"
427 "str q16, [x21, #0x0]\n"
428 "zip1 v17.16b, v20.16b, v19.16b\n"
429 "zip2 v16.16b, v20.16b, v19.16b\n"
430 "str q18, [x21, #0x10]\n"
431 "str q17, [x21, #0x20]\n"
432 "str q16, [x21, #0x30]\n"
433 "add x21, x21, #0x40\n"
439 "ldr s19, [x17], #0x4\n"
440 "ldr s18, [x16], #0x4\n"
441 "sub x20, x20, #0x4\n"
443 "ldr s17, [x15], #0x4\n"
444 "ldr s16, [x14], #0x4\n"
445 "zip1 v17.16b, v19.16b, v17.16b\n"
446 "zip1 v16.16b, v18.16b, v16.16b\n"
447 "zip1 v16.16b, v17.16b, v16.16b\n"
448 "str q16, [x21, #0x0]\n"
449 "add x21, x21, #0x10\n"
455 "ldr b19, [x17], #0x1\n"
456 "ldr b18, [x16], #0x1\n"
457 "sub x20, x20, #0x1\n"
459 "ldr b17, [x15], #0x1\n"
460 "ldr b16, [x14], #0x1\n"
461 "zip1 v17.16b, v19.16b, v17.16b\n"
462 "zip1 v16.16b, v18.16b, v16.16b\n"
463 "zip1 v16.16b, v17.16b, v16.16b\n"
464 "str s16, [x21, #0x0]\n"
465 "add x21, x21, #0x4\n"
468 "cmp %x[height], #0x1\n"
469 "add %x[out], %x[out], #0x80\n"
472 : [height]
"+&r" (height), [in]
"+&r" (in), [out]
"+&r" (out)
473 : [in_stride]
"r" (in_stride), [out_stride]
"r" (out_stride), [pad_row]
"r" (pad_row), [width]
"r" (width)
474 :
"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",
"x14",
"x15",
"x16",
"x17",
"x20",
"x21",
"x22",
"x23",
"x24",
"x25",
"x26",
"x27",
"x28"
481 void Transform<32, 4, true, VLType::None>(
482 uint8_t *out,
const uint8_t *in,
int stride,
int x0,
int xmax,
int k0,
int kmax)
484 a64_transpose_interleave_32_1x4(
485 reinterpret_cast<uint8_t *
>(out),
486 reinterpret_cast<const uint8_t *
>(in + k0 * stride + x0),
487 (xmax-x0) *
sizeof(uint8_t) / 1,
488 stride *
sizeof(uint8_t),
494 void Transform<32, 4, true, VLType::None>(
495 int8_t *out,
const int8_t *in,
int stride,
int x0,
int xmax,
int k0,
int kmax)
497 a64_transpose_interleave_32_1x4(
498 reinterpret_cast<uint8_t *
>(out),
499 reinterpret_cast<const uint8_t *
>(in + k0 * stride + x0),
500 (xmax-x0) *
sizeof(int8_t) / 1,
501 stride *
sizeof(int8_t),
507 #endif // defined(__aarch64__)