27 #if defined(__aarch64__)
31 void a64_transpose_interleave_12_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 = 12 * roundup<size_t>(height, 4) *
sizeof(uint8_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, #0x30\n"
64 "ldr q17, [x27], #0x10\n"
65 "ldr q16, [x26], #0x10\n"
66 "zip1 v31.16b, v21.16b, v17.16b\n"
67 "zip1 v22.16b, v20.16b, v16.16b\n"
68 "ldr q19, [x25], #0x10\n"
69 "ldr q18, [x23], #0x10\n"
70 "zip2 v14.16b, v21.16b, v17.16b\n"
71 "zip2 v13.16b, v20.16b, v16.16b\n"
72 "ldr q17, [x22], #0x10\n"
73 "ldr q16, [x20], #0x10\n"
74 "zip1 v30.16b, v19.16b, v17.16b\n"
75 "zip1 v29.16b, v18.16b, v16.16b\n"
76 "ldr q21, [x9], #0x10\n"
77 "ldr q20, [x28], #0x10\n"
78 "zip2 v12.16b, v19.16b, v17.16b\n"
79 "zip2 v11.16b, v18.16b, v16.16b\n"
80 "ldr q17, [x27], #0x10\n"
81 "ldr q16, [x26], #0x10\n"
82 "zip1 v10.16b, v21.16b, v17.16b\n"
83 "zip1 v9.16b, v20.16b, v16.16b\n"
84 "ldr q19, [x25], #0x10\n"
85 "ldr q18, [x23], #0x10\n"
86 "zip2 v8.16b, v21.16b, v17.16b\n"
87 "zip2 v7.16b, v20.16b, v16.16b\n"
88 "ldr q17, [x22], #0x10\n"
89 "ldr q16, [x20], #0x10\n"
90 "zip1 v6.16b, v19.16b, v17.16b\n"
91 "zip1 v5.16b, v18.16b, v16.16b\n"
92 "ldr q28, [x9], #0x10\n"
93 "ldr q27, [x28], #0x10\n"
94 "zip2 v4.16b, v19.16b, v17.16b\n"
95 "zip2 v3.16b, v18.16b, v16.16b\n"
96 "ldr q26, [x27], #0x10\n"
97 "ldr q25, [x26], #0x10\n"
98 "zip1 v2.16b, v28.16b, v26.16b\n"
99 "zip1 v1.16b, v27.16b, v25.16b\n"
100 "ldr q24, [x25], #0x10\n"
101 "ldr q23, [x23], #0x10\n"
102 "zip1 v16.16b, v31.16b, v22.16b\n"
103 "zip2 v22.16b, v31.16b, v22.16b\n"
104 "ldr q21, [x22], #0x10\n"
105 "ldr q20, [x20], #0x10\n"
106 "zip1 v0.16b, v24.16b, v21.16b\n"
107 "zip1 v31.16b, v23.16b, v20.16b\n"
108 "zip1 v19.16b, v14.16b, v13.16b\n"
109 "zip1 v18.16b, v30.16b, v29.16b\n"
110 "str q16, [x21, #0x0]\n"
111 "zip2 v16.16b, v30.16b, v29.16b\n"
112 "zip1 v17.16b, v12.16b, v11.16b\n"
113 "str q22, [x21, #0x10]\n"
114 "str q19, [x21, #0x20]\n"
115 "zip2 v30.16b, v28.16b, v26.16b\n"
116 "zip2 v29.16b, v27.16b, v25.16b\n"
117 "str q18, [x21, #0x30]\n"
118 "zip2 v28.16b, v24.16b, v21.16b\n"
119 "zip2 v27.16b, v23.16b, v20.16b\n"
120 "str q16, [x21, #0x40]\n"
121 "zip2 v21.16b, v14.16b, v13.16b\n"
122 "zip1 v16.16b, v10.16b, v9.16b\n"
123 "str q17, [x21, #0x50]\n"
124 "add x21, x21, %x[out_stride]\n"
125 "zip2 v20.16b, v10.16b, v9.16b\n"
126 "zip2 v19.16b, v12.16b, v11.16b\n"
127 "zip1 v18.16b, v6.16b, v5.16b\n"
128 "zip2 v17.16b, v6.16b, v5.16b\n"
129 "str q21, [x21, #0x0]\n"
130 "str q16, [x21, #0x10]\n"
131 "zip1 v16.16b, v8.16b, v7.16b\n"
132 "zip2 v26.16b, v8.16b, v7.16b\n"
133 "str q20, [x21, #0x20]\n"
134 "zip1 v25.16b, v2.16b, v1.16b\n"
135 "zip1 v24.16b, v4.16b, v3.16b\n"
136 "str q19, [x21, #0x30]\n"
137 "zip2 v23.16b, v4.16b, v3.16b\n"
138 "zip1 v22.16b, v0.16b, v31.16b\n"
139 "str q18, [x21, #0x40]\n"
140 "zip2 v21.16b, v2.16b, v1.16b\n"
141 "zip1 v20.16b, v30.16b, v29.16b\n"
142 "str q17, [x21, #0x50]\n"
143 "add x21, x21, %x[out_stride]\n"
144 "zip2 v19.16b, v30.16b, v29.16b\n"
145 "zip2 v18.16b, v0.16b, v31.16b\n"
146 "str q16, [x21, #0x0]\n"
147 "zip1 v17.16b, v28.16b, v27.16b\n"
148 "zip2 v16.16b, v28.16b, v27.16b\n"
149 "str q26, [x21, #0x10]\n"
150 "str q25, [x21, #0x20]\n"
151 "str q24, [x21, #0x30]\n"
152 "str q23, [x21, #0x40]\n"
153 "str q22, [x21, #0x50]\n"
154 "add x21, x21, %x[out_stride]\n"
155 "str q21, [x21, #0x0]\n"
156 "str q20, [x21, #0x10]\n"
157 "str q19, [x21, #0x20]\n"
158 "str q18, [x21, #0x30]\n"
159 "str q17, [x21, #0x40]\n"
160 "str q16, [x21, #0x50]\n"
161 "add x21, x21, %x[out_stride]\n"
167 "ldr d23, [x9], #0x8\n"
168 "ldr d22, [x28], #0x8\n"
169 "sub x24, x24, #0xc\n"
171 "ldr d19, [x27], #0x8\n"
172 "ldr d18, [x26], #0x8\n"
173 "ldr d21, [x25], #0x8\n"
174 "ldr d25, [x23], #0x8\n"
175 "ldr d20, [x22], #0x8\n"
176 "ldr d17, [x20], #0x8\n"
177 "ld1 { v23.s }[2], [x9], #0x4\n"
178 "ld1 { v22.s }[2], [x28], #0x4\n"
179 "ld1 { v19.s }[2], [x27], #0x4\n"
180 "ld1 { v18.s }[2], [x26], #0x4\n"
181 "zip1 v24.16b, v23.16b, v19.16b\n"
182 "zip1 v16.16b, v22.16b, v18.16b\n"
183 "ld1 { v21.s }[2], [x25], #0x4\n"
184 "ld1 { v25.s }[2], [x23], #0x4\n"
185 "zip2 v19.16b, v23.16b, v19.16b\n"
186 "zip2 v18.16b, v22.16b, v18.16b\n"
187 "ld1 { v20.s }[2], [x22], #0x4\n"
188 "ld1 { v17.s }[2], [x20], #0x4\n"
189 "zip1 v23.16b, v21.16b, v20.16b\n"
190 "zip1 v22.16b, v25.16b, v17.16b\n"
191 "zip2 v21.16b, v21.16b, v20.16b\n"
192 "zip2 v20.16b, v25.16b, v17.16b\n"
193 "zip1 v17.16b, v24.16b, v16.16b\n"
194 "zip2 v16.16b, v24.16b, v16.16b\n"
195 "str q17, [x21, #0x0]\n"
196 "zip1 v19.16b, v19.16b, v18.16b\n"
197 "zip1 v18.16b, v23.16b, v22.16b\n"
198 "str q16, [x21, #0x10]\n"
199 "zip2 v17.16b, v23.16b, v22.16b\n"
200 "zip1 v16.16b, v21.16b, v20.16b\n"
201 "str q19, [x21, #0x20]\n"
202 "str q18, [x21, #0x30]\n"
203 "str q17, [x21, #0x40]\n"
204 "str q16, [x21, #0x50]\n"
205 "add x21, x21, %x[out_stride]\n"
211 "ldr s19, [x9], #0x4\n"
212 "ldr s18, [x28], #0x4\n"
213 "sub x24, x24, #0x4\n"
215 "ldr s17, [x27], #0x4\n"
216 "ldr s16, [x26], #0x4\n"
217 "zip1 v17.16b, v19.16b, v17.16b\n"
218 "zip1 v16.16b, v18.16b, v16.16b\n"
219 "ldr s20, [x25], #0x4\n"
220 "ldr s19, [x23], #0x4\n"
221 "zip1 v18.16b, v17.16b, v16.16b\n"
222 "ldr s17, [x22], #0x4\n"
223 "ldr s16, [x20], #0x4\n"
224 "zip1 v17.16b, v20.16b, v17.16b\n"
225 "zip1 v16.16b, v19.16b, v16.16b\n"
226 "str q18, [x21, #0x0]\n"
227 "zip1 v16.16b, v17.16b, v16.16b\n"
228 "str q16, [x21, #0x30]\n"
229 "add x21, x21, #0x10\n"
235 "ldr b19, [x9], #0x1\n"
236 "ldr b18, [x28], #0x1\n"
237 "sub x24, x24, #0x1\n"
239 "ldr b17, [x27], #0x1\n"
240 "ldr b16, [x26], #0x1\n"
241 "zip1 v17.16b, v19.16b, v17.16b\n"
242 "zip1 v16.16b, v18.16b, v16.16b\n"
243 "ldr b20, [x25], #0x1\n"
244 "ldr b19, [x23], #0x1\n"
245 "zip1 v18.16b, v17.16b, v16.16b\n"
246 "ldr b17, [x22], #0x1\n"
247 "ldr b16, [x20], #0x1\n"
248 "zip1 v17.16b, v20.16b, v17.16b\n"
249 "zip1 v16.16b, v19.16b, v16.16b\n"
250 "str s18, [x21, #0x0]\n"
251 "zip1 v16.16b, v17.16b, v16.16b\n"
252 "str s16, [x21, #0x30]\n"
253 "add x21, x21, #0x4\n"
256 "cmp %x[height], #0x8\n"
257 "add %x[out], %x[out], #0x60\n"
259 "cbz %x[height], 20f\n"
263 "add x28, x9, %x[in_stride]\n"
264 "add x27, x28, %x[in_stride]\n"
265 "mov x20, %x[width]\n"
266 "add x26, x27, %x[in_stride]\n"
267 "cmp %x[height], #0x3\n"
268 "add %x[in], x26, %x[in_stride]\n"
269 "csel x26, x26, %x[pad_row], GT\n"
270 "csel x27, x27, %x[pad_row], GE\n"
271 "cmp %x[height], #0x1\n"
272 "csel x28, x28, %x[pad_row], GT\n"
275 "sub %x[height], %x[height], #0x4\n"
278 "ldr q21, [x9], #0x10\n"
279 "ldr q20, [x28], #0x10\n"
280 "sub x20, x20, #0x30\n"
282 "ldr q17, [x27], #0x10\n"
283 "ldr q16, [x26], #0x10\n"
284 "zip1 v31.16b, v21.16b, v17.16b\n"
285 "zip1 v30.16b, v20.16b, v16.16b\n"
286 "ldr q19, [x9], #0x10\n"
287 "ldr q18, [x28], #0x10\n"
288 "zip2 v29.16b, v21.16b, v17.16b\n"
289 "zip2 v28.16b, v20.16b, v16.16b\n"
290 "ldr q17, [x27], #0x10\n"
291 "ldr q16, [x26], #0x10\n"
292 "zip1 v27.16b, v19.16b, v17.16b\n"
293 "zip1 v26.16b, v18.16b, v16.16b\n"
294 "ldr q22, [x9], #0x10\n"
295 "ldr q21, [x28], #0x10\n"
296 "zip2 v25.16b, v19.16b, v17.16b\n"
297 "zip2 v20.16b, v18.16b, v16.16b\n"
298 "ldr q19, [x27], #0x10\n"
299 "ldr q18, [x26], #0x10\n"
300 "zip1 v24.16b, v22.16b, v19.16b\n"
301 "zip1 v23.16b, v21.16b, v18.16b\n"
302 "zip1 v16.16b, v31.16b, v30.16b\n"
303 "zip2 v17.16b, v31.16b, v30.16b\n"
304 "str q16, [x21, #0x0]\n"
305 "zip1 v16.16b, v29.16b, v28.16b\n"
306 "str q17, [x21, #0x10]\n"
307 "zip2 v22.16b, v22.16b, v19.16b\n"
308 "str q16, [x21, #0x20]\n"
309 "add x21, x21, %x[out_stride]\n"
310 "zip2 v21.16b, v21.16b, v18.16b\n"
311 "zip2 v18.16b, v29.16b, v28.16b\n"
312 "zip1 v16.16b, v27.16b, v26.16b\n"
313 "zip2 v17.16b, v27.16b, v26.16b\n"
314 "str q18, [x21, #0x0]\n"
315 "str q16, [x21, #0x10]\n"
316 "zip1 v16.16b, v25.16b, v20.16b\n"
317 "zip2 v20.16b, v25.16b, v20.16b\n"
318 "str q17, [x21, #0x20]\n"
319 "add x21, x21, %x[out_stride]\n"
320 "zip1 v19.16b, v24.16b, v23.16b\n"
321 "zip2 v18.16b, v24.16b, v23.16b\n"
322 "str q16, [x21, #0x0]\n"
323 "zip1 v17.16b, v22.16b, v21.16b\n"
324 "zip2 v16.16b, v22.16b, v21.16b\n"
325 "str q20, [x21, #0x10]\n"
326 "str q19, [x21, #0x20]\n"
327 "add x21, x21, %x[out_stride]\n"
328 "str q18, [x21, #0x0]\n"
329 "str q17, [x21, #0x10]\n"
330 "str q16, [x21, #0x20]\n"
331 "add x21, x21, %x[out_stride]\n"
337 "ldr d19, [x9], #0x8\n"
338 "ldr d21, [x28], #0x8\n"
339 "sub x20, x20, #0xc\n"
341 "ldr d18, [x27], #0x8\n"
342 "ldr d16, [x26], #0x8\n"
343 "ld1 { v19.s }[2], [x9], #0x4\n"
344 "ld1 { v21.s }[2], [x28], #0x4\n"
345 "ld1 { v18.s }[2], [x27], #0x4\n"
346 "ld1 { v16.s }[2], [x26], #0x4\n"
347 "zip1 v20.16b, v19.16b, v18.16b\n"
348 "zip1 v17.16b, v21.16b, v16.16b\n"
349 "zip2 v19.16b, v19.16b, v18.16b\n"
350 "zip2 v18.16b, v21.16b, v16.16b\n"
351 "zip1 v16.16b, v20.16b, v17.16b\n"
352 "zip2 v17.16b, v20.16b, v17.16b\n"
353 "str q16, [x21, #0x0]\n"
354 "zip1 v16.16b, v19.16b, v18.16b\n"
355 "str q17, [x21, #0x10]\n"
356 "str q16, [x21, #0x20]\n"
357 "add x21, x21, %x[out_stride]\n"
363 "ldr s19, [x9], #0x4\n"
364 "ldr s18, [x28], #0x4\n"
365 "sub x20, x20, #0x4\n"
367 "ldr s17, [x27], #0x4\n"
368 "ldr s16, [x26], #0x4\n"
369 "zip1 v17.16b, v19.16b, v17.16b\n"
370 "zip1 v16.16b, v18.16b, v16.16b\n"
371 "zip1 v16.16b, v17.16b, v16.16b\n"
372 "str q16, [x21, #0x0]\n"
373 "add x21, x21, #0x10\n"
379 "ldr b19, [x9], #0x1\n"
380 "ldr b18, [x28], #0x1\n"
381 "sub x20, x20, #0x1\n"
383 "ldr b17, [x27], #0x1\n"
384 "ldr b16, [x26], #0x1\n"
385 "zip1 v17.16b, v19.16b, v17.16b\n"
386 "zip1 v16.16b, v18.16b, v16.16b\n"
387 "zip1 v16.16b, v17.16b, v16.16b\n"
388 "str s16, [x21, #0x0]\n"
389 "add x21, x21, #0x4\n"
392 "cmp %x[height], #0x1\n"
393 "add %x[out], %x[out], #0x30\n"
396 : [height]
"+&r" (height), [in]
"+&r" (in), [out]
"+&r" (out)
397 : [in_stride]
"r" (in_stride), [out_stride]
"r" (out_stride), [pad_row]
"r" (pad_row), [width]
"r" (width)
398 :
"cc",
"memory",
"v0",
"v1",
"v2",
"v3",
"v4",
"v5",
"v6",
"v7",
"v8",
"v9",
"v10",
"v11",
"v12",
"v13",
"v14",
"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"
405 void Transform<12, 4, true, VLType::None>(
406 uint8_t *out,
const uint8_t *in,
int stride,
int x0,
int xmax,
int k0,
int kmax)
408 a64_transpose_interleave_12_1x4(
409 reinterpret_cast<uint8_t *
>(out),
410 reinterpret_cast<const uint8_t *
>(in + k0 * stride + x0),
411 (xmax-x0) *
sizeof(uint8_t) / 1,
412 stride *
sizeof(uint8_t),
418 void Transform<12, 4, true, VLType::None>(
419 int8_t *out,
const int8_t *in,
int stride,
int x0,
int xmax,
int k0,
int kmax)
421 a64_transpose_interleave_12_1x4(
422 reinterpret_cast<uint8_t *
>(out),
423 reinterpret_cast<const uint8_t *
>(in + k0 * stride + x0),
424 (xmax-x0) *
sizeof(int8_t) / 1,
425 stride *
sizeof(int8_t),
431 #endif // defined(__aarch64__)