27 #if defined(__aarch64__)
31 void a64_transpose_interleave_16_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 = 16 * 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 q21, [x17], #0x10\n"
69 "ldr q20, [x16], #0x10\n"
70 "sub x24, x24, #0x10\n"
72 "ldr q17, [x15], #0x10\n"
73 "ldr q16, [x14], #0x10\n"
74 "zip1 v3.16b, v21.16b, v17.16b\n"
75 "zip1 v2.16b, v20.16b, v16.16b\n"
76 "ldr q19, [x13], #0x10\n"
77 "ldr q18, [x12], #0x10\n"
78 "zip2 v1.16b, v21.16b, v17.16b\n"
79 "zip2 v0.16b, v20.16b, v16.16b\n"
80 "ldr q17, [x11], #0x10\n"
81 "ldr q16, [x10], #0x10\n"
82 "zip1 v31.16b, v19.16b, v17.16b\n"
83 "zip1 v30.16b, v18.16b, v16.16b\n"
84 "ldr q25, [x9], #0x10\n"
85 "ldr q20, [x28], #0x10\n"
86 "zip2 v24.16b, v19.16b, v17.16b\n"
87 "zip2 v23.16b, v18.16b, v16.16b\n"
88 "ldr q17, [x27], #0x10\n"
89 "ldr q16, [x26], #0x10\n"
90 "zip1 v22.16b, v25.16b, v17.16b\n"
91 "zip1 v21.16b, v20.16b, v16.16b\n"
92 "ldr q19, [x25], #0x10\n"
93 "ldr q18, [x23], #0x10\n"
94 "zip2 v29.16b, v25.16b, v17.16b\n"
95 "zip2 v20.16b, v20.16b, v16.16b\n"
96 "ldr q17, [x22], #0x10\n"
97 "ldr q16, [x20], #0x10\n"
98 "zip1 v28.16b, v19.16b, v17.16b\n"
99 "zip1 v27.16b, v18.16b, v16.16b\n"
100 "zip2 v26.16b, v19.16b, v17.16b\n"
101 "zip2 v25.16b, v18.16b, v16.16b\n"
102 "zip1 v16.16b, v3.16b, v2.16b\n"
103 "zip2 v17.16b, v3.16b, v2.16b\n"
104 "str q16, [x21, #0x0]\n"
105 "zip1 v16.16b, v1.16b, v0.16b\n"
106 "zip2 v19.16b, v1.16b, v0.16b\n"
107 "str q17, [x21, #0x10]\n"
108 "zip1 v18.16b, v31.16b, v30.16b\n"
109 "zip2 v17.16b, v31.16b, v30.16b\n"
110 "str q16, [x21, #0x20]\n"
111 "zip1 v16.16b, v24.16b, v23.16b\n"
112 "zip2 v24.16b, v24.16b, v23.16b\n"
113 "str q19, [x21, #0x30]\n"
114 "zip1 v23.16b, v22.16b, v21.16b\n"
115 "zip2 v22.16b, v22.16b, v21.16b\n"
116 "str q18, [x21, #0x40]\n"
117 "zip1 v21.16b, v29.16b, v20.16b\n"
118 "zip2 v20.16b, v29.16b, v20.16b\n"
119 "str q17, [x21, #0x50]\n"
120 "zip1 v19.16b, v28.16b, v27.16b\n"
121 "zip2 v18.16b, v28.16b, v27.16b\n"
122 "str q16, [x21, #0x60]\n"
123 "zip1 v17.16b, v26.16b, v25.16b\n"
124 "zip2 v16.16b, v26.16b, v25.16b\n"
125 "str q24, [x21, #0x70]\n"
126 "str q23, [x21, #0x80]\n"
127 "str q22, [x21, #0x90]\n"
128 "str q21, [x21, #0xa0]\n"
129 "str q20, [x21, #0xb0]\n"
130 "str q19, [x21, #0xc0]\n"
131 "str q18, [x21, #0xd0]\n"
132 "str q17, [x21, #0xe0]\n"
133 "str q16, [x21, #0xf0]\n"
134 "add x21, x21, %x[out_stride]\n"
140 "ldr s19, [x17], #0x4\n"
141 "ldr s18, [x16], #0x4\n"
142 "sub x24, x24, #0x4\n"
144 "ldr s17, [x15], #0x4\n"
145 "ldr s16, [x14], #0x4\n"
146 "zip1 v17.16b, v19.16b, v17.16b\n"
147 "zip1 v16.16b, v18.16b, v16.16b\n"
148 "ldr s19, [x13], #0x4\n"
149 "ldr s18, [x12], #0x4\n"
150 "zip1 v22.16b, v17.16b, v16.16b\n"
151 "ldr s17, [x11], #0x4\n"
152 "ldr s16, [x10], #0x4\n"
153 "zip1 v17.16b, v19.16b, v17.16b\n"
154 "zip1 v16.16b, v18.16b, v16.16b\n"
155 "ldr s19, [x9], #0x4\n"
156 "ldr s18, [x28], #0x4\n"
157 "zip1 v21.16b, v17.16b, v16.16b\n"
158 "ldr s17, [x27], #0x4\n"
159 "ldr s16, [x26], #0x4\n"
160 "zip1 v17.16b, v19.16b, v17.16b\n"
161 "zip1 v16.16b, v18.16b, v16.16b\n"
162 "ldr s20, [x25], #0x4\n"
163 "ldr s19, [x23], #0x4\n"
164 "zip1 v18.16b, v17.16b, v16.16b\n"
165 "ldr s17, [x22], #0x4\n"
166 "ldr s16, [x20], #0x4\n"
167 "zip1 v17.16b, v20.16b, v17.16b\n"
168 "zip1 v16.16b, v19.16b, v16.16b\n"
169 "str q22, [x21, #0x0]\n"
170 "zip1 v16.16b, v17.16b, v16.16b\n"
171 "str q21, [x21, #0x40]\n"
172 "str q18, [x21, #0x80]\n"
173 "str q16, [x21, #0xc0]\n"
174 "add x21, x21, #0x10\n"
180 "ldr b19, [x17], #0x1\n"
181 "ldr b18, [x16], #0x1\n"
182 "sub x24, x24, #0x1\n"
184 "ldr b17, [x15], #0x1\n"
185 "ldr b16, [x14], #0x1\n"
186 "zip1 v17.16b, v19.16b, v17.16b\n"
187 "zip1 v16.16b, v18.16b, v16.16b\n"
188 "ldr b19, [x13], #0x1\n"
189 "ldr b18, [x12], #0x1\n"
190 "zip1 v22.16b, v17.16b, v16.16b\n"
191 "ldr b17, [x11], #0x1\n"
192 "ldr b16, [x10], #0x1\n"
193 "zip1 v17.16b, v19.16b, v17.16b\n"
194 "zip1 v16.16b, v18.16b, v16.16b\n"
195 "ldr b19, [x9], #0x1\n"
196 "ldr b18, [x28], #0x1\n"
197 "zip1 v21.16b, v17.16b, v16.16b\n"
198 "ldr b17, [x27], #0x1\n"
199 "ldr b16, [x26], #0x1\n"
200 "zip1 v17.16b, v19.16b, v17.16b\n"
201 "zip1 v16.16b, v18.16b, v16.16b\n"
202 "ldr b20, [x25], #0x1\n"
203 "ldr b19, [x23], #0x1\n"
204 "zip1 v18.16b, v17.16b, v16.16b\n"
205 "ldr b17, [x22], #0x1\n"
206 "ldr b16, [x20], #0x1\n"
207 "zip1 v17.16b, v20.16b, v17.16b\n"
208 "zip1 v16.16b, v19.16b, v16.16b\n"
209 "str s22, [x21, #0x0]\n"
210 "zip1 v16.16b, v17.16b, v16.16b\n"
211 "str s21, [x21, #0x40]\n"
212 "str s18, [x21, #0x80]\n"
213 "str s16, [x21, #0xc0]\n"
214 "add x21, x21, #0x4\n"
217 "cmp %x[height], #0x10\n"
218 "add %x[out], %x[out], #0x100\n"
220 "cbz %x[height], 16f\n"
224 "add x16, x17, %x[in_stride]\n"
225 "add x15, x16, %x[in_stride]\n"
226 "mov x20, %x[width]\n"
227 "add x14, x15, %x[in_stride]\n"
228 "cmp %x[height], #0x3\n"
229 "add %x[in], x14, %x[in_stride]\n"
230 "csel x14, x14, %x[pad_row], GT\n"
231 "csel x15, x15, %x[pad_row], GE\n"
232 "cmp %x[height], #0x1\n"
233 "csel x16, x16, %x[pad_row], GT\n"
236 "sub %x[height], %x[height], #0x4\n"
239 "ldr q20, [x17], #0x10\n"
240 "ldr q21, [x16], #0x10\n"
241 "sub x20, x20, #0x10\n"
243 "ldr q19, [x15], #0x10\n"
244 "ldr q16, [x14], #0x10\n"
245 "zip1 v18.16b, v20.16b, v19.16b\n"
246 "zip1 v17.16b, v21.16b, v16.16b\n"
247 "zip2 v20.16b, v20.16b, v19.16b\n"
248 "zip2 v19.16b, v21.16b, v16.16b\n"
249 "zip1 v16.16b, v18.16b, v17.16b\n"
250 "zip2 v18.16b, v18.16b, v17.16b\n"
251 "str q16, [x21, #0x0]\n"
252 "zip1 v17.16b, v20.16b, v19.16b\n"
253 "zip2 v16.16b, v20.16b, v19.16b\n"
254 "str q18, [x21, #0x10]\n"
255 "str q17, [x21, #0x20]\n"
256 "str q16, [x21, #0x30]\n"
257 "add x21, x21, %x[out_stride]\n"
263 "ldr s19, [x17], #0x4\n"
264 "ldr s18, [x16], #0x4\n"
265 "sub x20, x20, #0x4\n"
267 "ldr s17, [x15], #0x4\n"
268 "ldr s16, [x14], #0x4\n"
269 "zip1 v17.16b, v19.16b, v17.16b\n"
270 "zip1 v16.16b, v18.16b, v16.16b\n"
271 "zip1 v16.16b, v17.16b, v16.16b\n"
272 "str q16, [x21, #0x0]\n"
273 "add x21, x21, #0x10\n"
279 "ldr b19, [x17], #0x1\n"
280 "ldr b18, [x16], #0x1\n"
281 "sub x20, x20, #0x1\n"
283 "ldr b17, [x15], #0x1\n"
284 "ldr b16, [x14], #0x1\n"
285 "zip1 v17.16b, v19.16b, v17.16b\n"
286 "zip1 v16.16b, v18.16b, v16.16b\n"
287 "zip1 v16.16b, v17.16b, v16.16b\n"
288 "str s16, [x21, #0x0]\n"
289 "add x21, x21, #0x4\n"
292 "cmp %x[height], #0x1\n"
293 "add %x[out], %x[out], #0x40\n"
296 : [height]
"+&r" (height), [in]
"+&r" (in), [out]
"+&r" (out)
297 : [in_stride]
"r" (in_stride), [out_stride]
"r" (out_stride), [pad_row]
"r" (pad_row), [width]
"r" (width)
298 :
"cc",
"memory",
"v0",
"v1",
"v2",
"v3",
"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"
305 void Transform<16, 4, true, VLType::None>(
306 uint8_t *out,
const uint8_t *in,
int stride,
int x0,
int xmax,
int k0,
int kmax)
308 a64_transpose_interleave_16_1x4(
309 reinterpret_cast<uint8_t *
>(out),
310 reinterpret_cast<const uint8_t *
>(in + k0 * stride + x0),
311 (xmax-x0) *
sizeof(uint8_t) / 1,
312 stride *
sizeof(uint8_t),
318 void Transform<16, 4, true, VLType::None>(
319 int8_t *out,
const int8_t *in,
int stride,
int x0,
int xmax,
int k0,
int kmax)
321 a64_transpose_interleave_16_1x4(
322 reinterpret_cast<uint8_t *
>(out),
323 reinterpret_cast<const uint8_t *
>(in + k0 * stride + x0),
324 (xmax-x0) *
sizeof(int8_t) / 1,
325 stride *
sizeof(int8_t),
331 #endif // defined(__aarch64__)