27 #if defined(__aarch64__)
31 void a64_transpose_interleave_32_2x2(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 = 32 * roundup<size_t>(height, 2) *
sizeof(uint16_t);
42 "cmp %x[height], #0x4\n"
46 "mov x24, %x[width]\n"
47 "add x23, x25, %x[in_stride]\n"
48 "add x22, x23, %x[in_stride]\n"
49 "add x20, x22, %x[in_stride]\n"
51 "add %x[in], x20, %x[in_stride]\n"
53 "sub %x[height], %x[height], #0x4\n"
56 "ldr q14, [x25], #0x10\n"
57 "ldr q10, [x23], #0x10\n"
58 "sub x24, x24, #0x40\n"
59 "zip1 v12.8h, v14.8h, v10.8h\n"
60 "ldr q5, [x22], #0x10\n"
61 "ldr q3, [x20], #0x10\n"
62 "zip2 v31.8h, v14.8h, v10.8h\n"
63 "zip1 v19.8h, v5.8h, v3.8h\n"
64 "ldr q27, [x25], #0x10\n"
65 "ldr q25, [x23], #0x10\n"
66 "zip1 v11.8h, v27.8h, v25.8h\n"
67 "zip2 v24.8h, v27.8h, v25.8h\n"
68 "ldr q6, [x22], #0x10\n"
69 "ldr q29, [x20], #0x10\n"
70 "zip2 v15.8h, v5.8h, v3.8h\n"
71 "zip1 v18.8h, v6.8h, v29.8h\n"
72 "ldr q17, [x25], #0x10\n"
73 "ldr q9, [x23], #0x10\n"
74 "zip1 v0.8h, v17.8h, v9.8h\n"
75 "zip2 v9.8h, v17.8h, v9.8h\n"
76 "ldr q21, [x22], #0x10\n"
77 "ldr q20, [x20], #0x10\n"
78 "zip2 v8.8h, v6.8h, v29.8h\n"
79 "zip1 v30.8h, v21.8h, v20.8h\n"
80 "ldr q17, [x25], #0x10\n"
81 "ldr q5, [x23], #0x10\n"
82 "zip1 v13.8h, v17.8h, v5.8h\n"
83 "zip2 v25.8h, v17.8h, v5.8h\n"
84 "ldr q7, [x22], #0x10\n"
85 "ldr q29, [x20], #0x10\n"
86 "zip2 v27.8h, v21.8h, v20.8h\n"
87 "zip1 v14.8h, v7.8h, v29.8h\n"
88 "ldr q28, [x25], #0x10\n"
89 "ldr q17, [x23], #0x10\n"
90 "zip2 v1.8h, v7.8h, v29.8h\n"
92 "ldr q10, [x22], #0x10\n"
93 "ldr q21, [x20], #0x10\n"
94 "zip1 v16.8h, v28.8h, v17.8h\n"
95 "zip2 v17.8h, v28.8h, v17.8h\n"
96 "ldr q5, [x25], #0x10\n"
97 "ldr q20, [x23], #0x10\n"
98 "zip1 v3.8h, v5.8h, v20.8h\n"
99 "zip2 v7.8h, v5.8h, v20.8h\n"
100 "ldr q22, [x22], #0x10\n"
101 "ldr q29, [x20], #0x10\n"
102 "zip1 v2.8h, v10.8h, v21.8h\n"
103 "zip2 v5.8h, v10.8h, v21.8h\n"
104 "ldr q21, [x25], #0x10\n"
105 "ldr q20, [x23], #0x10\n"
106 "zip1 v4.8h, v21.8h, v20.8h\n"
107 "zip2 v28.8h, v21.8h, v20.8h\n"
108 "ldr q6, [x22], #0x10\n"
109 "ldr q10, [x20], #0x10\n"
110 "zip1 v26.8h, v22.8h, v29.8h\n"
111 "zip2 v20.8h, v22.8h, v29.8h\n"
112 "ldr q29, [x25], #0x10\n"
113 "ldr q23, [x23], #0x10\n"
114 "zip1 v21.8h, v29.8h, v23.8h\n"
115 "zip2 v23.8h, v29.8h, v23.8h\n"
116 "ldr q22, [x22], #0x10\n"
117 "ldr q29, [x20], #0x10\n"
118 "str q12, [x21, #0x0]\n"
119 "zip1 v12.8h, v6.8h, v10.8h\n"
120 "str q31, [x21, #0x10]\n"
121 "zip2 v6.8h, v6.8h, v10.8h\n"
122 "zip1 v31.8h, v22.8h, v29.8h\n"
123 "str q11, [x21, #0x20]\n"
124 "zip2 v11.8h, v22.8h, v29.8h\n"
125 "str q24, [x21, #0x30]\n"
126 "str q0, [x21, #0x40]\n"
127 "str q9, [x21, #0x50]\n"
128 "str q13, [x21, #0x60]\n"
129 "str q25, [x21, #0x70]\n"
130 "str q19, [x21, #0x80]\n"
131 "str q15, [x21, #0x90]\n"
132 "str q18, [x21, #0xa0]\n"
133 "str q8, [x21, #0xb0]\n"
134 "str q30, [x21, #0xc0]\n"
135 "str q27, [x21, #0xd0]\n"
136 "str q14, [x21, #0xe0]\n"
137 "str q1, [x21, #0xf0]\n"
138 "add x21, x21, %x[out_stride]\n"
139 "str q16, [x21, #0x0]\n"
140 "str q17, [x21, #0x10]\n"
141 "str q3, [x21, #0x20]\n"
142 "str q7, [x21, #0x30]\n"
143 "str q4, [x21, #0x40]\n"
144 "str q28, [x21, #0x50]\n"
145 "str q21, [x21, #0x60]\n"
146 "str q23, [x21, #0x70]\n"
147 "str q2, [x21, #0x80]\n"
148 "str q5, [x21, #0x90]\n"
149 "str q26, [x21, #0xa0]\n"
150 "str q20, [x21, #0xb0]\n"
151 "str q12, [x21, #0xc0]\n"
152 "str q6, [x21, #0xd0]\n"
153 "str q31, [x21, #0xe0]\n"
154 "str q11, [x21, #0xf0]\n"
155 "add x21, x21, %x[out_stride]\n"
161 "ldr q17, [x25], #0x10\n"
162 "ldr q16, [x23], #0x10\n"
163 "sub x24, x24, #0x20\n"
165 "ldr q21, [x22], #0x10\n"
166 "ldr q18, [x20], #0x10\n"
167 "zip1 v1.8h, v17.8h, v16.8h\n"
168 "zip2 v0.8h, v17.8h, v16.8h\n"
169 "ldr q17, [x25], #0x10\n"
170 "ldr q16, [x23], #0x10\n"
171 "zip1 v31.8h, v17.8h, v16.8h\n"
172 "zip2 v30.8h, v17.8h, v16.8h\n"
173 "ldr q20, [x22], #0x10\n"
174 "ldr q19, [x20], #0x10\n"
175 "zip1 v29.8h, v21.8h, v18.8h\n"
176 "zip2 v28.8h, v21.8h, v18.8h\n"
177 "ldr q17, [x25], #0x10\n"
178 "ldr q16, [x23], #0x10\n"
179 "zip1 v27.8h, v17.8h, v16.8h\n"
180 "zip2 v26.8h, v17.8h, v16.8h\n"
181 "ldr q25, [x22], #0x10\n"
182 "ldr q18, [x20], #0x10\n"
183 "zip1 v24.8h, v20.8h, v19.8h\n"
184 "zip2 v23.8h, v20.8h, v19.8h\n"
185 "ldr q17, [x25], #0x10\n"
186 "ldr q16, [x23], #0x10\n"
187 "zip1 v22.8h, v17.8h, v16.8h\n"
188 "zip2 v21.8h, v17.8h, v16.8h\n"
189 "ldr q20, [x22], #0x10\n"
190 "ldr q16, [x20], #0x10\n"
191 "zip1 v19.8h, v25.8h, v18.8h\n"
192 "zip2 v18.8h, v25.8h, v18.8h\n"
193 "zip1 v17.8h, v20.8h, v16.8h\n"
194 "zip2 v16.8h, v20.8h, v16.8h\n"
195 "str q1, [x21, #0x0]\n"
196 "str q0, [x21, #0x10]\n"
197 "str q31, [x21, #0x20]\n"
198 "str q30, [x21, #0x30]\n"
199 "str q27, [x21, #0x40]\n"
200 "str q26, [x21, #0x50]\n"
201 "str q22, [x21, #0x60]\n"
202 "str q21, [x21, #0x70]\n"
203 "str q29, [x21, #0x80]\n"
204 "str q28, [x21, #0x90]\n"
205 "str q24, [x21, #0xa0]\n"
206 "str q23, [x21, #0xb0]\n"
207 "str q19, [x21, #0xc0]\n"
208 "str q18, [x21, #0xd0]\n"
209 "str q17, [x21, #0xe0]\n"
210 "str q16, [x21, #0xf0]\n"
211 "add x21, x21, %x[out_stride]\n"
217 "ldr q17, [x25], #0x10\n"
218 "ldr q16, [x23], #0x10\n"
219 "sub x24, x24, #0x10\n"
221 "ldr q24, [x22], #0x10\n"
222 "ldr q23, [x20], #0x10\n"
223 "zip1 v19.8h, v17.8h, v16.8h\n"
224 "zip2 v18.8h, v17.8h, v16.8h\n"
225 "ldr q17, [x25], #0x10\n"
226 "ldr q16, [x23], #0x10\n"
227 "zip1 v22.8h, v17.8h, v16.8h\n"
228 "zip2 v21.8h, v17.8h, v16.8h\n"
229 "ldr q20, [x22], #0x10\n"
230 "ldr q16, [x20], #0x10\n"
231 "str q19, [x21, #0x0]\n"
232 "zip1 v19.8h, v24.8h, v23.8h\n"
233 "str q18, [x21, #0x10]\n"
234 "zip2 v18.8h, v24.8h, v23.8h\n"
235 "zip1 v17.8h, v20.8h, v16.8h\n"
236 "str q22, [x21, #0x20]\n"
237 "zip2 v16.8h, v20.8h, v16.8h\n"
238 "str q21, [x21, #0x30]\n"
239 "str q19, [x21, #0x80]\n"
240 "str q18, [x21, #0x90]\n"
241 "str q17, [x21, #0xa0]\n"
242 "str q16, [x21, #0xb0]\n"
243 "add x21, x21, #0x40\n"
249 "ldr d19, [x25], #0x8\n"
250 "ldr d16, [x23], #0x8\n"
251 "sub x24, x24, #0x4\n"
253 "ldr d18, [x22], #0x8\n"
254 "ldr d17, [x20], #0x8\n"
255 "zip1 v16.8h, v19.8h, v16.8h\n"
256 "str q16, [x21, #0x0]\n"
257 "zip1 v16.8h, v18.8h, v17.8h\n"
258 "str q16, [x21, #0x80]\n"
259 "add x21, x21, #0x10\n"
265 "ldr h19, [x25], #0x2\n"
266 "ldr h16, [x23], #0x2\n"
267 "sub x24, x24, #0x1\n"
269 "ldr h18, [x22], #0x2\n"
270 "ldr h17, [x20], #0x2\n"
271 "zip1 v16.8h, v19.8h, v16.8h\n"
272 "str s16, [x21, #0x0]\n"
273 "zip1 v16.8h, v18.8h, v17.8h\n"
274 "str s16, [x21, #0x80]\n"
275 "add x21, x21, #0x4\n"
278 "cmp %x[height], #0x4\n"
279 "add %x[out], %x[out], #0x100\n"
281 "cbz %x[height], 24f\n"
285 "mov x20, %x[width]\n"
286 "add x23, x25, %x[in_stride]\n"
287 "cmp %x[height], #0x1\n"
288 "add %x[in], x23, %x[in_stride]\n"
289 "csel x23, x23, %x[pad_row], GT\n"
292 "sub %x[height], %x[height], #0x2\n"
295 "ldr q18, [x25], #0x10\n"
296 "ldr q17, [x23], #0x10\n"
297 "sub x20, x20, #0x40\n"
298 "zip1 v0.8h, v18.8h, v17.8h\n"
299 "ldr q19, [x25], #0x10\n"
300 "ldr q16, [x23], #0x10\n"
301 "zip2 v31.8h, v18.8h, v17.8h\n"
302 "zip1 v30.8h, v19.8h, v16.8h\n"
303 "ldr q18, [x25], #0x10\n"
304 "ldr q17, [x23], #0x10\n"
305 "zip2 v29.8h, v19.8h, v16.8h\n"
306 "zip1 v28.8h, v18.8h, v17.8h\n"
307 "ldr q19, [x25], #0x10\n"
308 "ldr q16, [x23], #0x10\n"
309 "zip2 v27.8h, v18.8h, v17.8h\n"
310 "zip1 v26.8h, v19.8h, v16.8h\n"
311 "ldr q18, [x25], #0x10\n"
312 "ldr q17, [x23], #0x10\n"
313 "zip2 v25.8h, v19.8h, v16.8h\n"
315 "ldr q19, [x25], #0x10\n"
316 "ldr q16, [x23], #0x10\n"
317 "zip1 v24.8h, v18.8h, v17.8h\n"
318 "zip2 v23.8h, v18.8h, v17.8h\n"
319 "ldr q18, [x25], #0x10\n"
320 "ldr q17, [x23], #0x10\n"
321 "zip1 v22.8h, v19.8h, v16.8h\n"
322 "zip2 v21.8h, v19.8h, v16.8h\n"
323 "ldr q20, [x25], #0x10\n"
324 "ldr q16, [x23], #0x10\n"
325 "str q0, [x21, #0x0]\n"
326 "zip1 v19.8h, v18.8h, v17.8h\n"
327 "str q31, [x21, #0x10]\n"
328 "zip2 v18.8h, v18.8h, v17.8h\n"
329 "zip1 v17.8h, v20.8h, v16.8h\n"
330 "str q30, [x21, #0x20]\n"
331 "zip2 v16.8h, v20.8h, v16.8h\n"
332 "str q29, [x21, #0x30]\n"
333 "str q28, [x21, #0x40]\n"
334 "str q27, [x21, #0x50]\n"
335 "str q26, [x21, #0x60]\n"
336 "str q25, [x21, #0x70]\n"
337 "add x21, x21, %x[out_stride]\n"
338 "str q24, [x21, #0x0]\n"
339 "str q23, [x21, #0x10]\n"
340 "str q22, [x21, #0x20]\n"
341 "str q21, [x21, #0x30]\n"
342 "str q19, [x21, #0x40]\n"
343 "str q18, [x21, #0x50]\n"
344 "str q17, [x21, #0x60]\n"
345 "str q16, [x21, #0x70]\n"
346 "add x21, x21, %x[out_stride]\n"
352 "ldr q18, [x25], #0x10\n"
353 "ldr q17, [x23], #0x10\n"
354 "sub x20, x20, #0x20\n"
356 "ldr q19, [x25], #0x10\n"
357 "ldr q16, [x23], #0x10\n"
358 "zip1 v24.8h, v18.8h, v17.8h\n"
359 "zip2 v23.8h, v18.8h, v17.8h\n"
360 "ldr q18, [x25], #0x10\n"
361 "ldr q17, [x23], #0x10\n"
362 "zip1 v22.8h, v19.8h, v16.8h\n"
363 "zip2 v21.8h, v19.8h, v16.8h\n"
364 "ldr q20, [x25], #0x10\n"
365 "ldr q16, [x23], #0x10\n"
366 "zip1 v19.8h, v18.8h, v17.8h\n"
367 "zip2 v18.8h, v18.8h, v17.8h\n"
368 "zip1 v17.8h, v20.8h, v16.8h\n"
369 "zip2 v16.8h, v20.8h, v16.8h\n"
370 "str q24, [x21, #0x0]\n"
371 "str q23, [x21, #0x10]\n"
372 "str q22, [x21, #0x20]\n"
373 "str q21, [x21, #0x30]\n"
374 "str q19, [x21, #0x40]\n"
375 "str q18, [x21, #0x50]\n"
376 "str q17, [x21, #0x60]\n"
377 "str q16, [x21, #0x70]\n"
378 "add x21, x21, %x[out_stride]\n"
384 "ldr q18, [x25], #0x10\n"
385 "ldr q17, [x23], #0x10\n"
386 "sub x20, x20, #0x10\n"
388 "ldr q20, [x25], #0x10\n"
389 "ldr q16, [x23], #0x10\n"
390 "zip1 v19.8h, v18.8h, v17.8h\n"
391 "zip2 v18.8h, v18.8h, v17.8h\n"
392 "zip1 v17.8h, v20.8h, v16.8h\n"
393 "zip2 v16.8h, v20.8h, v16.8h\n"
394 "str q19, [x21, #0x0]\n"
395 "str q18, [x21, #0x10]\n"
396 "str q17, [x21, #0x20]\n"
397 "str q16, [x21, #0x30]\n"
398 "add x21, x21, #0x40\n"
404 "ldr d17, [x25], #0x8\n"
405 "ldr d16, [x23], #0x8\n"
406 "sub x20, x20, #0x4\n"
408 "zip1 v16.8h, v17.8h, v16.8h\n"
409 "str q16, [x21, #0x0]\n"
410 "add x21, x21, #0x10\n"
416 "ldr h17, [x25], #0x2\n"
417 "ldr h16, [x23], #0x2\n"
418 "sub x20, x20, #0x1\n"
420 "zip1 v16.8h, v17.8h, v16.8h\n"
421 "str s16, [x21, #0x0]\n"
422 "add x21, x21, #0x4\n"
425 "cmp %x[height], #0x1\n"
426 "add %x[out], %x[out], #0x80\n"
429 : [height]
"+&r" (height), [in]
"+&r" (in), [out]
"+&r" (out)
430 : [in_stride]
"r" (in_stride), [out_stride]
"r" (out_stride), [pad_row]
"r" (pad_row), [width]
"r" (width)
431 :
"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"
438 void Transform<32, 2, true, VLType::None>(
439 bfloat16 *out,
const bfloat16 *in,
int stride,
int x0,
int xmax,
int k0,
int kmax)
441 a64_transpose_interleave_32_2x2(
442 reinterpret_cast<uint16_t *
>(out),
443 reinterpret_cast<const uint16_t *
>(in + k0 * stride + x0),
451 #endif // defined(__aarch64__)