27 #if defined(__aarch64__)
31 void a64_transpose_interleave_4_1x16(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 = 4 * roundup<size_t>(height, 16) *
sizeof(uint8_t);
44 "add x16, x17, %x[in_stride]\n"
45 "add x15, x16, %x[in_stride]\n"
46 "add x14, x15, %x[in_stride]\n"
47 "add x13, x14, %x[in_stride]\n"
48 "add x12, x13, %x[in_stride]\n"
49 "add x11, x12, %x[in_stride]\n"
50 "add x10, x11, %x[in_stride]\n"
51 "add x9, x10, %x[in_stride]\n"
52 "add x28, x9, %x[in_stride]\n"
53 "add x27, x28, %x[in_stride]\n"
54 "add x26, x27, %x[in_stride]\n"
55 "add x25, x26, %x[in_stride]\n"
56 "add x24, x25, %x[in_stride]\n"
57 "add x23, x24, %x[in_stride]\n"
58 "add x22, x23, %x[in_stride]\n"
59 "cmp %x[height], #0xf\n"
60 "add %x[in], x22, %x[in_stride]\n"
61 "csel x22, x22, %x[pad_row], GT\n"
62 "csel x23, x23, %x[pad_row], GE\n"
63 "cmp %x[height], #0xd\n"
64 "csel x24, x24, %x[pad_row], GT\n"
65 "csel x25, x25, %x[pad_row], GE\n"
66 "cmp %x[height], #0xb\n"
67 "csel x26, x26, %x[pad_row], GT\n"
68 "csel x27, x27, %x[pad_row], GE\n"
69 "cmp %x[height], #0x9\n"
70 "csel x28, x28, %x[pad_row], GT\n"
71 "csel x9, x9, %x[pad_row], GE\n"
72 "cmp %x[height], #0x7\n"
73 "csel x10, x10, %x[pad_row], GT\n"
74 "csel x11, x11, %x[pad_row], GE\n"
75 "cmp %x[height], #0x5\n"
76 "mov x21, %x[width]\n"
77 "csel x12, x12, %x[pad_row], GT\n"
78 "csel x13, x13, %x[pad_row], GE\n"
79 "cmp %x[height], #0x3\n"
80 "csel x14, x14, %x[pad_row], GT\n"
81 "csel x15, x15, %x[pad_row], GE\n"
82 "cmp %x[height], #0x1\n"
83 "csel x16, x16, %x[pad_row], GT\n"
86 "sub %x[height], %x[height], #0x10\n"
89 "ldr q3, [x17], #0x10\n"
90 "ldr q9, [x16], #0x10\n"
91 "sub x21, x21, #0x10\n"
93 "ldr q2, [x15], #0x10\n"
94 "ldr q8, [x14], #0x10\n"
95 "ldr q0, [x13], #0x10\n"
96 "ldr q31, [x12], #0x10\n"
97 "ldr q30, [x11], #0x10\n"
98 "ldr q7, [x10], #0x10\n"
99 "ldr q29, [x9], #0x10\n"
100 "ldr q28, [x28], #0x10\n"
101 "zip1 v27.16b, v3.16b, v29.16b\n"
102 "zip1 v6.16b, v9.16b, v28.16b\n"
103 "ldr q25, [x27], #0x10\n"
104 "ldr q24, [x26], #0x10\n"
105 "zip1 v26.16b, v2.16b, v25.16b\n"
106 "zip1 v1.16b, v8.16b, v24.16b\n"
107 "ldr q23, [x25], #0x10\n"
108 "ldr q22, [x24], #0x10\n"
109 "zip1 v21.16b, v0.16b, v23.16b\n"
110 "zip1 v20.16b, v31.16b, v22.16b\n"
111 "ldr q19, [x23], #0x10\n"
112 "ldr q18, [x22], #0x10\n"
113 "zip1 v17.16b, v30.16b, v19.16b\n"
114 "zip1 v16.16b, v7.16b, v18.16b\n"
115 "zip2 v5.16b, v3.16b, v29.16b\n"
116 "zip2 v0.16b, v0.16b, v23.16b\n"
117 "zip2 v4.16b, v2.16b, v25.16b\n"
118 "zip2 v3.16b, v30.16b, v19.16b\n"
119 "zip2 v2.16b, v9.16b, v28.16b\n"
120 "zip2 v31.16b, v31.16b, v22.16b\n"
121 "zip2 v30.16b, v8.16b, v24.16b\n"
122 "zip2 v29.16b, v7.16b, v18.16b\n"
123 "zip1 v25.16b, v27.16b, v21.16b\n"
124 "zip1 v24.16b, v26.16b, v17.16b\n"
125 "zip1 v23.16b, v6.16b, v20.16b\n"
126 "zip1 v22.16b, v1.16b, v16.16b\n"
127 "zip2 v28.16b, v27.16b, v21.16b\n"
128 "zip2 v27.16b, v26.16b, v17.16b\n"
129 "zip2 v26.16b, v6.16b, v20.16b\n"
130 "zip2 v21.16b, v1.16b, v16.16b\n"
131 "zip1 v1.16b, v5.16b, v0.16b\n"
132 "zip1 v20.16b, v4.16b, v3.16b\n"
133 "zip1 v19.16b, v2.16b, v31.16b\n"
134 "zip1 v16.16b, v30.16b, v29.16b\n"
135 "zip1 v18.16b, v25.16b, v24.16b\n"
136 "zip1 v17.16b, v23.16b, v22.16b\n"
137 "zip2 v25.16b, v25.16b, v24.16b\n"
138 "zip2 v24.16b, v23.16b, v22.16b\n"
139 "zip2 v0.16b, v5.16b, v0.16b\n"
140 "zip2 v23.16b, v4.16b, v3.16b\n"
141 "zip2 v31.16b, v2.16b, v31.16b\n"
142 "zip2 v22.16b, v30.16b, v29.16b\n"
143 "zip1 v30.16b, v28.16b, v27.16b\n"
144 "zip1 v29.16b, v26.16b, v21.16b\n"
145 "zip2 v28.16b, v28.16b, v27.16b\n"
146 "zip2 v27.16b, v26.16b, v21.16b\n"
147 "zip1 v26.16b, v1.16b, v20.16b\n"
148 "zip1 v21.16b, v19.16b, v16.16b\n"
149 "zip2 v20.16b, v1.16b, v20.16b\n"
150 "zip2 v19.16b, v19.16b, v16.16b\n"
151 "zip1 v16.16b, v18.16b, v17.16b\n"
152 "zip2 v18.16b, v18.16b, v17.16b\n"
153 "str q16, [x20, #0x0]\n"
154 "zip1 v17.16b, v25.16b, v24.16b\n"
155 "zip2 v16.16b, v25.16b, v24.16b\n"
156 "str q18, [x20, #0x10]\n"
157 "str q17, [x20, #0x20]\n"
158 "zip1 v25.16b, v0.16b, v23.16b\n"
159 "zip1 v24.16b, v31.16b, v22.16b\n"
160 "str q16, [x20, #0x30]\n"
161 "add x20, x20, %x[out_stride]\n"
162 "zip2 v23.16b, v0.16b, v23.16b\n"
163 "zip2 v22.16b, v31.16b, v22.16b\n"
164 "zip1 v16.16b, v30.16b, v29.16b\n"
165 "zip2 v17.16b, v30.16b, v29.16b\n"
166 "str q16, [x20, #0x0]\n"
167 "zip1 v16.16b, v28.16b, v27.16b\n"
168 "zip2 v18.16b, v28.16b, v27.16b\n"
169 "str q17, [x20, #0x10]\n"
170 "str q16, [x20, #0x20]\n"
171 "zip1 v17.16b, v26.16b, v21.16b\n"
172 "zip2 v16.16b, v26.16b, v21.16b\n"
173 "str q18, [x20, #0x30]\n"
174 "add x20, x20, %x[out_stride]\n"
175 "zip1 v21.16b, v20.16b, v19.16b\n"
176 "zip2 v20.16b, v20.16b, v19.16b\n"
177 "str q17, [x20, #0x0]\n"
178 "zip1 v19.16b, v25.16b, v24.16b\n"
179 "zip2 v18.16b, v25.16b, v24.16b\n"
180 "str q16, [x20, #0x10]\n"
181 "zip1 v17.16b, v23.16b, v22.16b\n"
182 "zip2 v16.16b, v23.16b, v22.16b\n"
183 "str q21, [x20, #0x20]\n"
184 "str q20, [x20, #0x30]\n"
185 "add x20, x20, %x[out_stride]\n"
186 "str q19, [x20, #0x0]\n"
187 "str q18, [x20, #0x10]\n"
188 "str q17, [x20, #0x20]\n"
189 "str q16, [x20, #0x30]\n"
190 "add x20, x20, %x[out_stride]\n"
196 "ldr s21, [x17], #0x4\n"
197 "ldr s23, [x16], #0x4\n"
198 "sub x21, x21, #0x4\n"
200 "ldr s20, [x15], #0x4\n"
201 "ldr s22, [x14], #0x4\n"
202 "ldr s19, [x13], #0x4\n"
203 "ldr s18, [x12], #0x4\n"
204 "ldr s25, [x11], #0x4\n"
205 "ldr s24, [x10], #0x4\n"
206 "ldr s17, [x9], #0x4\n"
207 "ldr s16, [x28], #0x4\n"
208 "zip1 v21.16b, v21.16b, v17.16b\n"
209 "zip1 v23.16b, v23.16b, v16.16b\n"
210 "ldr s17, [x27], #0x4\n"
211 "ldr s16, [x26], #0x4\n"
212 "zip1 v20.16b, v20.16b, v17.16b\n"
213 "zip1 v22.16b, v22.16b, v16.16b\n"
214 "ldr s17, [x25], #0x4\n"
215 "ldr s16, [x24], #0x4\n"
216 "zip1 v19.16b, v19.16b, v17.16b\n"
217 "zip1 v18.16b, v18.16b, v16.16b\n"
218 "ldr s17, [x23], #0x4\n"
219 "ldr s16, [x22], #0x4\n"
220 "zip1 v17.16b, v25.16b, v17.16b\n"
221 "zip1 v16.16b, v24.16b, v16.16b\n"
222 "zip1 v21.16b, v21.16b, v19.16b\n"
223 "zip1 v20.16b, v20.16b, v17.16b\n"
224 "zip1 v19.16b, v23.16b, v18.16b\n"
225 "zip1 v16.16b, v22.16b, v16.16b\n"
226 "zip1 v18.16b, v21.16b, v20.16b\n"
227 "zip1 v17.16b, v19.16b, v16.16b\n"
228 "zip2 v20.16b, v21.16b, v20.16b\n"
229 "zip2 v19.16b, v19.16b, v16.16b\n"
230 "zip1 v16.16b, v18.16b, v17.16b\n"
231 "zip2 v18.16b, v18.16b, v17.16b\n"
232 "str q16, [x20, #0x0]\n"
233 "zip1 v17.16b, v20.16b, v19.16b\n"
234 "zip2 v16.16b, v20.16b, v19.16b\n"
235 "str q18, [x20, #0x10]\n"
236 "str q17, [x20, #0x20]\n"
237 "str q16, [x20, #0x30]\n"
238 "add x20, x20, %x[out_stride]\n"
244 "ldr b23, [x17], #0x1\n"
245 "ldr b22, [x16], #0x1\n"
246 "sub x21, x21, #0x1\n"
248 "ldr b21, [x15], #0x1\n"
249 "ldr b20, [x14], #0x1\n"
250 "ldr b19, [x13], #0x1\n"
251 "ldr b18, [x12], #0x1\n"
252 "ldr b25, [x11], #0x1\n"
253 "ldr b24, [x10], #0x1\n"
254 "ldr b17, [x9], #0x1\n"
255 "ldr b16, [x28], #0x1\n"
256 "zip1 v23.16b, v23.16b, v17.16b\n"
257 "zip1 v22.16b, v22.16b, v16.16b\n"
258 "ldr b17, [x27], #0x1\n"
259 "ldr b16, [x26], #0x1\n"
260 "zip1 v21.16b, v21.16b, v17.16b\n"
261 "zip1 v20.16b, v20.16b, v16.16b\n"
262 "ldr b17, [x25], #0x1\n"
263 "ldr b16, [x24], #0x1\n"
264 "zip1 v19.16b, v19.16b, v17.16b\n"
265 "zip1 v18.16b, v18.16b, v16.16b\n"
266 "ldr b17, [x23], #0x1\n"
267 "ldr b16, [x22], #0x1\n"
268 "zip1 v17.16b, v25.16b, v17.16b\n"
269 "zip1 v16.16b, v24.16b, v16.16b\n"
270 "zip1 v19.16b, v23.16b, v19.16b\n"
271 "zip1 v17.16b, v21.16b, v17.16b\n"
272 "zip1 v18.16b, v22.16b, v18.16b\n"
273 "zip1 v16.16b, v20.16b, v16.16b\n"
274 "zip1 v17.16b, v19.16b, v17.16b\n"
275 "zip1 v16.16b, v18.16b, v16.16b\n"
276 "zip1 v16.16b, v17.16b, v16.16b\n"
277 "str q16, [x20, #0x0]\n"
278 "add x20, x20, #0x10\n"
281 "cmp %x[height], #0x1\n"
282 "add %x[out], %x[out], #0x40\n"
284 : [height]
"+&r" (height), [in]
"+&r" (in), [out]
"+&r" (out)
285 : [in_stride]
"r" (in_stride), [out_stride]
"r" (out_stride), [pad_row]
"r" (pad_row), [width]
"r" (width)
286 :
"cc",
"memory",
"v0",
"v1",
"v2",
"v3",
"v4",
"v5",
"v6",
"v7",
"v8",
"v9",
"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"
293 void Transform<4, 16, true, VLType::None>(
294 uint8_t *out,
const uint8_t *in,
int stride,
int x0,
int xmax,
int k0,
int kmax)
296 a64_transpose_interleave_4_1x16(
297 reinterpret_cast<uint8_t *
>(out),
298 reinterpret_cast<const uint8_t *
>(in + k0 * stride + x0),
299 (xmax-x0) *
sizeof(uint8_t) / 1,
300 stride *
sizeof(uint8_t),
306 void Transform<4, 16, true, VLType::None>(
307 int8_t *out,
const int8_t *in,
int stride,
int x0,
int xmax,
int k0,
int kmax)
309 a64_transpose_interleave_4_1x16(
310 reinterpret_cast<uint8_t *
>(out),
311 reinterpret_cast<const uint8_t *
>(in + k0 * stride + x0),
312 (xmax-x0) *
sizeof(int8_t) / 1,
313 stride *
sizeof(int8_t),
319 #endif // defined(__aarch64__)