27 #if defined(__aarch64__)
31 void a64_transpose_interleave_12_2x4(uint16_t *out,
const uint16_t *in,
size_t width,
size_t in_stride,
size_t height)
33 uint16_t *pad_row =
reinterpret_cast<uint16_t *
>(alloca(width *
sizeof(uint16_t)));
36 memset(pad_row, 0, width *
sizeof(uint16_t));
39 size_t out_stride = 12 * roundup<size_t>(height, 4) *
sizeof(uint16_t);
42 "cmp %x[height], #0x8\n"
46 "add x28, x9, %x[in_stride]\n"
47 "add x27, x28, %x[in_stride]\n"
48 "add x26, x27, %x[in_stride]\n"
49 "add x25, x26, %x[in_stride]\n"
50 "mov x24, %x[width]\n"
51 "add x23, x25, %x[in_stride]\n"
52 "add x22, x23, %x[in_stride]\n"
53 "add x20, x22, %x[in_stride]\n"
55 "add %x[in], x20, %x[in_stride]\n"
57 "sub %x[height], %x[height], #0x8\n"
60 "ldr q21, [x9], #0x10\n"
61 "ldr q20, [x28], #0x10\n"
62 "sub x24, x24, #0x18\n"
64 "ldr q17, [x27], #0x10\n"
65 "ldr q16, [x26], #0x10\n"
66 "zip1 v13.8h, v21.8h, v17.8h\n"
67 "zip1 v12.8h, v20.8h, v16.8h\n"
68 "ldr q19, [x25], #0x10\n"
69 "ldr q18, [x23], #0x10\n"
70 "zip2 v11.8h, v21.8h, v17.8h\n"
71 "zip2 v10.8h, v20.8h, v16.8h\n"
72 "ldr q17, [x22], #0x10\n"
73 "ldr q16, [x20], #0x10\n"
74 "zip1 v9.8h, v19.8h, v17.8h\n"
75 "zip1 v8.8h, v18.8h, v16.8h\n"
76 "ldr q21, [x9], #0x10\n"
77 "ldr q20, [x28], #0x10\n"
78 "zip2 v7.8h, v19.8h, v17.8h\n"
79 "zip2 v6.8h, v18.8h, v16.8h\n"
80 "ldr q17, [x27], #0x10\n"
81 "ldr q16, [x26], #0x10\n"
82 "zip1 v27.8h, v21.8h, v17.8h\n"
83 "zip1 v22.8h, v20.8h, v16.8h\n"
84 "ldr q19, [x25], #0x10\n"
85 "ldr q18, [x23], #0x10\n"
86 "zip2 v5.8h, v21.8h, v17.8h\n"
87 "zip2 v4.8h, v20.8h, v16.8h\n"
88 "ldr q17, [x22], #0x10\n"
89 "ldr q16, [x20], #0x10\n"
90 "zip1 v26.8h, v19.8h, v17.8h\n"
91 "zip1 v25.8h, v18.8h, v16.8h\n"
92 "ldr q21, [x9], #0x10\n"
93 "ldr q20, [x28], #0x10\n"
94 "zip2 v3.8h, v19.8h, v17.8h\n"
95 "zip2 v2.8h, v18.8h, v16.8h\n"
96 "ldr q17, [x27], #0x10\n"
97 "ldr q16, [x26], #0x10\n"
98 "zip1 v24.8h, v21.8h, v17.8h\n"
99 "zip1 v23.8h, v20.8h, v16.8h\n"
100 "ldr q19, [x25], #0x10\n"
101 "ldr q18, [x23], #0x10\n"
102 "zip2 v1.8h, v21.8h, v17.8h\n"
103 "zip2 v0.8h, v20.8h, v16.8h\n"
104 "ldr q17, [x22], #0x10\n"
105 "ldr q16, [x20], #0x10\n"
106 "zip1 v31.8h, v19.8h, v17.8h\n"
107 "zip1 v30.8h, v18.8h, v16.8h\n"
108 "zip2 v29.8h, v19.8h, v17.8h\n"
109 "zip2 v28.8h, v18.8h, v16.8h\n"
110 "zip1 v17.8h, v13.8h, v12.8h\n"
111 "zip2 v16.8h, v13.8h, v12.8h\n"
112 "str q17, [x21, #0x0]\n"
113 "zip1 v18.8h, v11.8h, v10.8h\n"
114 "zip2 v17.8h, v11.8h, v10.8h\n"
115 "str q16, [x21, #0x10]\n"
116 "zip1 v16.8h, v27.8h, v22.8h\n"
117 "zip2 v22.8h, v27.8h, v22.8h\n"
118 "str q18, [x21, #0x20]\n"
119 "zip1 v21.8h, v9.8h, v8.8h\n"
120 "zip2 v20.8h, v9.8h, v8.8h\n"
121 "str q17, [x21, #0x30]\n"
122 "zip1 v19.8h, v7.8h, v6.8h\n"
123 "zip2 v18.8h, v7.8h, v6.8h\n"
124 "str q16, [x21, #0x40]\n"
125 "zip1 v17.8h, v26.8h, v25.8h\n"
126 "zip2 v16.8h, v26.8h, v25.8h\n"
127 "str q22, [x21, #0x50]\n"
128 "str q21, [x21, #0x60]\n"
129 "zip1 v27.8h, v5.8h, v4.8h\n"
130 "zip2 v26.8h, v5.8h, v4.8h\n"
131 "str q20, [x21, #0x70]\n"
132 "zip1 v25.8h, v24.8h, v23.8h\n"
133 "zip2 v24.8h, v24.8h, v23.8h\n"
134 "str q19, [x21, #0x80]\n"
135 "zip1 v23.8h, v1.8h, v0.8h\n"
136 "zip2 v22.8h, v1.8h, v0.8h\n"
137 "str q18, [x21, #0x90]\n"
138 "zip1 v21.8h, v3.8h, v2.8h\n"
139 "zip2 v20.8h, v3.8h, v2.8h\n"
140 "str q17, [x21, #0xa0]\n"
141 "zip1 v19.8h, v31.8h, v30.8h\n"
142 "zip2 v18.8h, v31.8h, v30.8h\n"
143 "str q16, [x21, #0xb0]\n"
144 "add x21, x21, %x[out_stride]\n"
145 "zip1 v17.8h, v29.8h, v28.8h\n"
146 "zip2 v16.8h, v29.8h, v28.8h\n"
147 "str q27, [x21, #0x0]\n"
148 "str q26, [x21, #0x10]\n"
149 "str q25, [x21, #0x20]\n"
150 "str q24, [x21, #0x30]\n"
151 "str q23, [x21, #0x40]\n"
152 "str q22, [x21, #0x50]\n"
153 "str q21, [x21, #0x60]\n"
154 "str q20, [x21, #0x70]\n"
155 "str q19, [x21, #0x80]\n"
156 "str q18, [x21, #0x90]\n"
157 "str q17, [x21, #0xa0]\n"
158 "str q16, [x21, #0xb0]\n"
159 "add x21, x21, %x[out_stride]\n"
165 "ldr q19, [x9], #0x10\n"
166 "ldr q18, [x28], #0x10\n"
167 "sub x24, x24, #0xc\n"
169 "ldr q17, [x27], #0x10\n"
170 "ldr q16, [x26], #0x10\n"
171 "zip1 v25.8h, v19.8h, v17.8h\n"
172 "zip1 v24.8h, v18.8h, v16.8h\n"
173 "ldr q21, [x25], #0x10\n"
174 "ldr q20, [x23], #0x10\n"
175 "zip2 v31.8h, v19.8h, v17.8h\n"
176 "zip2 v23.8h, v18.8h, v16.8h\n"
177 "ldr q17, [x22], #0x10\n"
178 "ldr q16, [x20], #0x10\n"
179 "zip1 v30.8h, v21.8h, v17.8h\n"
180 "zip1 v29.8h, v20.8h, v16.8h\n"
181 "ldr d19, [x9], #0x8\n"
182 "ldr d18, [x28], #0x8\n"
183 "zip2 v28.8h, v21.8h, v17.8h\n"
184 "zip2 v27.8h, v20.8h, v16.8h\n"
185 "ldr d17, [x27], #0x8\n"
186 "ldr d16, [x26], #0x8\n"
187 "zip1 v26.8h, v19.8h, v17.8h\n"
188 "zip1 v22.8h, v18.8h, v16.8h\n"
189 "ldr d21, [x25], #0x8\n"
190 "ldr d20, [x23], #0x8\n"
191 "zip1 v19.8h, v25.8h, v24.8h\n"
192 "zip2 v18.8h, v25.8h, v24.8h\n"
193 "ldr d17, [x22], #0x8\n"
194 "ldr d16, [x20], #0x8\n"
195 "zip1 v25.8h, v21.8h, v17.8h\n"
196 "zip1 v24.8h, v20.8h, v16.8h\n"
197 "zip1 v17.8h, v31.8h, v23.8h\n"
198 "zip2 v16.8h, v31.8h, v23.8h\n"
199 "str q19, [x21, #0x0]\n"
200 "zip1 v23.8h, v26.8h, v22.8h\n"
201 "zip2 v22.8h, v26.8h, v22.8h\n"
202 "str q18, [x21, #0x10]\n"
203 "zip1 v21.8h, v30.8h, v29.8h\n"
204 "zip2 v20.8h, v30.8h, v29.8h\n"
205 "str q17, [x21, #0x20]\n"
206 "zip1 v19.8h, v28.8h, v27.8h\n"
207 "zip2 v18.8h, v28.8h, v27.8h\n"
208 "str q16, [x21, #0x30]\n"
209 "zip1 v17.8h, v25.8h, v24.8h\n"
210 "zip2 v16.8h, v25.8h, v24.8h\n"
211 "str q23, [x21, #0x40]\n"
212 "str q22, [x21, #0x50]\n"
213 "str q21, [x21, #0x60]\n"
214 "str q20, [x21, #0x70]\n"
215 "str q19, [x21, #0x80]\n"
216 "str q18, [x21, #0x90]\n"
217 "str q17, [x21, #0xa0]\n"
218 "str q16, [x21, #0xb0]\n"
219 "add x21, x21, %x[out_stride]\n"
225 "ldr d19, [x9], #0x8\n"
226 "ldr d18, [x28], #0x8\n"
227 "sub x24, x24, #0x4\n"
229 "ldr d17, [x27], #0x8\n"
230 "ldr d16, [x26], #0x8\n"
231 "zip1 v17.8h, v19.8h, v17.8h\n"
232 "zip1 v16.8h, v18.8h, v16.8h\n"
233 "ldr d18, [x25], #0x8\n"
234 "ldr d21, [x23], #0x8\n"
235 "zip1 v20.8h, v17.8h, v16.8h\n"
236 "zip2 v19.8h, v17.8h, v16.8h\n"
237 "ldr d17, [x22], #0x8\n"
238 "ldr d16, [x20], #0x8\n"
239 "zip1 v18.8h, v18.8h, v17.8h\n"
240 "zip1 v16.8h, v21.8h, v16.8h\n"
241 "str q20, [x21, #0x0]\n"
242 "zip1 v17.8h, v18.8h, v16.8h\n"
243 "zip2 v16.8h, v18.8h, v16.8h\n"
244 "str q19, [x21, #0x10]\n"
245 "str q17, [x21, #0x60]\n"
246 "str q16, [x21, #0x70]\n"
247 "add x21, x21, #0x20\n"
253 "ldr h19, [x9], #0x2\n"
254 "ldr h18, [x28], #0x2\n"
255 "sub x24, x24, #0x1\n"
257 "ldr h17, [x27], #0x2\n"
258 "ldr h16, [x26], #0x2\n"
259 "zip1 v17.8h, v19.8h, v17.8h\n"
260 "zip1 v16.8h, v18.8h, v16.8h\n"
261 "ldr h20, [x25], #0x2\n"
262 "ldr h19, [x23], #0x2\n"
263 "zip1 v18.8h, v17.8h, v16.8h\n"
264 "ldr h17, [x22], #0x2\n"
265 "ldr h16, [x20], #0x2\n"
266 "zip1 v17.8h, v20.8h, v17.8h\n"
267 "zip1 v16.8h, v19.8h, v16.8h\n"
268 "str d18, [x21, #0x0]\n"
269 "zip1 v16.8h, v17.8h, v16.8h\n"
270 "str d16, [x21, #0x60]\n"
271 "add x21, x21, #0x8\n"
274 "cmp %x[height], #0x8\n"
275 "add %x[out], %x[out], #0xc0\n"
277 "cbz %x[height], 20f\n"
281 "add x28, x9, %x[in_stride]\n"
282 "add x27, x28, %x[in_stride]\n"
283 "mov x20, %x[width]\n"
284 "add x26, x27, %x[in_stride]\n"
285 "cmp %x[height], #0x3\n"
286 "add %x[in], x26, %x[in_stride]\n"
287 "csel x26, x26, %x[pad_row], GT\n"
288 "csel x27, x27, %x[pad_row], GE\n"
289 "cmp %x[height], #0x1\n"
290 "csel x28, x28, %x[pad_row], GT\n"
293 "sub %x[height], %x[height], #0x4\n"
296 "ldr q19, [x9], #0x10\n"
297 "ldr q18, [x28], #0x10\n"
298 "sub x20, x20, #0x18\n"
300 "ldr q17, [x27], #0x10\n"
301 "ldr q16, [x26], #0x10\n"
302 "zip1 v31.8h, v19.8h, v17.8h\n"
303 "zip1 v30.8h, v18.8h, v16.8h\n"
304 "ldr q22, [x9], #0x10\n"
305 "ldr q20, [x28], #0x10\n"
306 "zip2 v29.8h, v19.8h, v17.8h\n"
307 "zip2 v28.8h, v18.8h, v16.8h\n"
308 "ldr q17, [x27], #0x10\n"
309 "ldr q16, [x26], #0x10\n"
310 "zip1 v27.8h, v22.8h, v17.8h\n"
311 "zip1 v21.8h, v20.8h, v16.8h\n"
312 "ldr q19, [x9], #0x10\n"
313 "ldr q18, [x28], #0x10\n"
314 "zip2 v26.8h, v22.8h, v17.8h\n"
315 "zip2 v20.8h, v20.8h, v16.8h\n"
316 "ldr q17, [x27], #0x10\n"
317 "ldr q16, [x26], #0x10\n"
318 "zip1 v25.8h, v19.8h, v17.8h\n"
319 "zip1 v24.8h, v18.8h, v16.8h\n"
320 "zip2 v23.8h, v19.8h, v17.8h\n"
321 "zip2 v22.8h, v18.8h, v16.8h\n"
322 "zip1 v17.8h, v31.8h, v30.8h\n"
323 "zip2 v16.8h, v31.8h, v30.8h\n"
324 "str q17, [x21, #0x0]\n"
325 "zip1 v19.8h, v29.8h, v28.8h\n"
326 "zip2 v18.8h, v29.8h, v28.8h\n"
327 "str q16, [x21, #0x10]\n"
328 "zip1 v17.8h, v27.8h, v21.8h\n"
329 "zip2 v16.8h, v27.8h, v21.8h\n"
330 "str q19, [x21, #0x20]\n"
331 "str q18, [x21, #0x30]\n"
332 "zip1 v21.8h, v26.8h, v20.8h\n"
333 "zip2 v20.8h, v26.8h, v20.8h\n"
334 "str q17, [x21, #0x40]\n"
335 "zip1 v19.8h, v25.8h, v24.8h\n"
336 "zip2 v18.8h, v25.8h, v24.8h\n"
337 "str q16, [x21, #0x50]\n"
338 "add x21, x21, %x[out_stride]\n"
339 "zip1 v17.8h, v23.8h, v22.8h\n"
340 "zip2 v16.8h, v23.8h, v22.8h\n"
341 "str q21, [x21, #0x0]\n"
342 "str q20, [x21, #0x10]\n"
343 "str q19, [x21, #0x20]\n"
344 "str q18, [x21, #0x30]\n"
345 "str q17, [x21, #0x40]\n"
346 "str q16, [x21, #0x50]\n"
347 "add x21, x21, %x[out_stride]\n"
353 "ldr q21, [x9], #0x10\n"
354 "ldr q18, [x28], #0x10\n"
355 "sub x20, x20, #0xc\n"
357 "ldr q17, [x27], #0x10\n"
358 "ldr q16, [x26], #0x10\n"
359 "zip1 v24.8h, v21.8h, v17.8h\n"
360 "zip1 v23.8h, v18.8h, v16.8h\n"
361 "ldr d20, [x9], #0x8\n"
362 "ldr d19, [x28], #0x8\n"
363 "zip2 v22.8h, v21.8h, v17.8h\n"
364 "zip2 v18.8h, v18.8h, v16.8h\n"
365 "ldr d17, [x27], #0x8\n"
366 "ldr d16, [x26], #0x8\n"
367 "zip1 v21.8h, v20.8h, v17.8h\n"
368 "zip1 v20.8h, v19.8h, v16.8h\n"
369 "zip1 v17.8h, v24.8h, v23.8h\n"
370 "zip2 v16.8h, v24.8h, v23.8h\n"
371 "str q17, [x21, #0x0]\n"
372 "zip1 v19.8h, v22.8h, v18.8h\n"
373 "zip2 v18.8h, v22.8h, v18.8h\n"
374 "str q16, [x21, #0x10]\n"
375 "zip1 v17.8h, v21.8h, v20.8h\n"
376 "zip2 v16.8h, v21.8h, v20.8h\n"
377 "str q19, [x21, #0x20]\n"
378 "str q18, [x21, #0x30]\n"
379 "str q17, [x21, #0x40]\n"
380 "str q16, [x21, #0x50]\n"
381 "add x21, x21, %x[out_stride]\n"
387 "ldr d18, [x9], #0x8\n"
388 "ldr d19, [x28], #0x8\n"
389 "sub x20, x20, #0x4\n"
391 "ldr d17, [x27], #0x8\n"
392 "ldr d16, [x26], #0x8\n"
393 "zip1 v18.8h, v18.8h, v17.8h\n"
394 "zip1 v16.8h, v19.8h, v16.8h\n"
395 "zip1 v17.8h, v18.8h, v16.8h\n"
396 "zip2 v16.8h, v18.8h, v16.8h\n"
397 "str q17, [x21, #0x0]\n"
398 "str q16, [x21, #0x10]\n"
399 "add x21, x21, #0x20\n"
405 "ldr h19, [x9], #0x2\n"
406 "ldr h18, [x28], #0x2\n"
407 "sub x20, x20, #0x1\n"
409 "ldr h17, [x27], #0x2\n"
410 "ldr h16, [x26], #0x2\n"
411 "zip1 v17.8h, v19.8h, v17.8h\n"
412 "zip1 v16.8h, v18.8h, v16.8h\n"
413 "zip1 v16.8h, v17.8h, v16.8h\n"
414 "str d16, [x21, #0x0]\n"
415 "add x21, x21, #0x8\n"
418 "cmp %x[height], #0x1\n"
419 "add %x[out], %x[out], #0x60\n"
422 : [height]
"+&r" (height), [in]
"+&r" (in), [out]
"+&r" (out)
423 : [in_stride]
"r" (in_stride), [out_stride]
"r" (out_stride), [pad_row]
"r" (pad_row), [width]
"r" (width)
424 :
"cc",
"memory",
"v0",
"v1",
"v2",
"v3",
"v4",
"v5",
"v6",
"v7",
"v8",
"v9",
"v10",
"v11",
"v12",
"v13",
"v16",
"v17",
"v18",
"v19",
"v20",
"v21",
"v22",
"v23",
"v24",
"v25",
"v26",
"v27",
"v28",
"v29",
"v30",
"v31",
"x9",
"x20",
"x21",
"x22",
"x23",
"x24",
"x25",
"x26",
"x27",
"x28"
431 void Transform<12, 4, true, VLType::None>(
432 bfloat16 *out,
const bfloat16 *in,
int stride,
int x0,
int xmax,
int k0,
int kmax)
434 a64_transpose_interleave_12_2x4(
435 reinterpret_cast<uint16_t *
>(out),
436 reinterpret_cast<const uint16_t *
>(in + k0 * stride + x0),
444 #endif // defined(__aarch64__)