27 #if defined(__aarch64__)
31 void a64_transpose_interleave_12_u8u16(uint16_t *out,
const uint8_t *in,
size_t width,
size_t in_stride,
size_t height)
33 size_t out_stride = 12 * height *
sizeof(uint16_t);
36 "cmp %x[height], #0x4\n"
40 "mov x24, %x[width]\n"
41 "add x23, x25, %x[in_stride]\n"
42 "add x22, x23, %x[in_stride]\n"
43 "add x20, x22, %x[in_stride]\n"
45 "add %x[in], x20, %x[in_stride]\n"
47 "sub %x[height], %x[height], #0x4\n"
50 "ldr q3, [x25], #0x10\n"
51 "ldr q21, [x23], #0x10\n"
52 "ushll2 v20.8h, v3.16b, #0x0\n"
53 "ushll v2.8h, v21.8b, #0x0\n"
54 "ldr q1, [x22], #0x10\n"
55 "ldr q19, [x20], #0x10\n"
56 "ushll2 v18.8h, v1.16b, #0x0\n"
57 "ushll v0.8h, v19.8b, #0x0\n"
58 "ldr d17, [x25], #0x8\n"
59 "ldr d16, [x22], #0x8\n"
60 "ushll v31.8h, v17.8b, #0x0\n"
61 "ushll v30.8h, v16.8b, #0x0\n"
62 "ldr d29, [x23], #0x8\n"
63 "ldr d28, [x20], #0x8\n"
64 "ushll2 v27.8h, v21.16b, #0x0\n"
65 "ushll2 v26.8h, v19.16b, #0x0\n"
66 "dup v25.2d, v20.d[0]\n"
67 "dup v24.2d, v2.d[1]\n"
68 "sub x24, x24, #0x18\n"
70 "dup v23.2d, v18.d[0]\n"
71 "dup v22.2d, v0.d[1]\n"
72 "dup v21.2d, v20.d[1]\n"
73 "dup v20.2d, v31.d[1]\n"
74 "dup v19.2d, v18.d[1]\n"
75 "dup v18.2d, v30.d[1]\n"
76 "ushll v17.8h, v3.8b, #0x0\n"
77 "ushll v16.8h, v1.8b, #0x0\n"
78 "str q17, [x21, #0x0]\n"
79 "mov v25.d[1], v2.d[0]\n"
80 "mov v24.d[1], v27.d[0]\n"
81 "str q25, [x21, #0x10]\n"
82 "mov v23.d[1], v0.d[0]\n"
83 "mov v22.d[1], v26.d[0]\n"
84 "str q24, [x21, #0x20]\n"
85 "str q16, [x21, #0x30]\n"
86 "ushll v17.8h, v29.8b, #0x0\n"
87 "ushll v16.8h, v28.8b, #0x0\n"
88 "str q23, [x21, #0x40]\n"
89 "mov v21.d[1], v31.d[0]\n"
90 "mov v20.d[1], v27.d[1]\n"
91 "str q22, [x21, #0x50]\n"
92 "add x21, x21, %x[out_stride]\n"
93 "mov v19.d[1], v30.d[0]\n"
94 "mov v18.d[1], v26.d[1]\n"
95 "str q21, [x21, #0x0]\n"
96 "str q20, [x21, #0x10]\n"
97 "str q17, [x21, #0x20]\n"
98 "str q19, [x21, #0x30]\n"
99 "str q18, [x21, #0x40]\n"
100 "str q16, [x21, #0x50]\n"
101 "add x21, x21, %x[out_stride]\n"
107 "ldr d19, [x23], #0x8\n"
108 "ldr d18, [x20], #0x8\n"
109 "sub x24, x24, #0xc\n"
111 "ld1 { v19.s }[2], [x23], #0x4\n"
112 "ld1 { v18.s }[2], [x20], #0x4\n"
113 "ushll v25.8h, v19.8b, #0x0\n"
114 "ushll v24.8h, v18.8b, #0x0\n"
115 "ldr d17, [x25], #0x8\n"
116 "ldr d16, [x22], #0x8\n"
117 "ushll2 v23.8h, v19.16b, #0x0\n"
118 "ushll2 v22.8h, v18.16b, #0x0\n"
119 "ld1 { v17.s }[2], [x25], #0x4\n"
120 "ld1 { v16.s }[2], [x22], #0x4\n"
121 "ushll2 v21.8h, v17.16b, #0x0\n"
122 "ushll2 v20.8h, v16.16b, #0x0\n"
123 "dup v19.2d, v25.d[1]\n"
124 "dup v18.2d, v24.d[1]\n"
125 "ushll v17.8h, v17.8b, #0x0\n"
126 "ushll v16.8h, v16.8b, #0x0\n"
127 "str q17, [x21, #0x0]\n"
128 "mov v21.d[1], v25.d[0]\n"
129 "mov v19.d[1], v23.d[0]\n"
130 "str q21, [x21, #0x10]\n"
131 "mov v20.d[1], v24.d[0]\n"
132 "mov v18.d[1], v22.d[0]\n"
133 "str q19, [x21, #0x20]\n"
134 "str q16, [x21, #0x30]\n"
135 "str q20, [x21, #0x40]\n"
136 "str q18, [x21, #0x50]\n"
137 "add x21, x21, %x[out_stride]\n"
143 "ldr s19, [x25], #0x4\n"
144 "ldr s18, [x23], #0x4\n"
145 "sub x24, x24, #0x4\n"
147 "ldr s17, [x22], #0x4\n"
148 "ldr s16, [x20], #0x4\n"
149 "ushll v19.8h, v19.8b, #0x0\n"
150 "ushll v18.8h, v18.8b, #0x0\n"
151 "ushll v17.8h, v17.8b, #0x0\n"
152 "ushll v16.8h, v16.8b, #0x0\n"
153 "str d19, [x21, #0x0]\n"
154 "str d18, [x21, #0x18]\n"
155 "str d17, [x21, #0x30]\n"
156 "str d16, [x21, #0x48]\n"
157 "add x21, x21, #0x8\n"
163 "ldr b19, [x25], #0x1\n"
164 "ldr b18, [x23], #0x1\n"
165 "sub x24, x24, #0x1\n"
167 "ldr b17, [x22], #0x1\n"
168 "ldr b16, [x20], #0x1\n"
169 "ushll v19.8h, v19.8b, #0x0\n"
170 "ushll v18.8h, v18.8b, #0x0\n"
171 "ushll v17.8h, v17.8b, #0x0\n"
172 "ushll v16.8h, v16.8b, #0x0\n"
173 "str h19, [x21, #0x0]\n"
174 "str h18, [x21, #0x18]\n"
175 "str h17, [x21, #0x30]\n"
176 "str h16, [x21, #0x48]\n"
177 "add x21, x21, #0x2\n"
180 "cmp %x[height], #0x4\n"
181 "add %x[out], %x[out], #0x60\n"
183 "cbz %x[height], 20f\n"
186 "mov x20, %x[width]\n"
189 "add %x[in], x25, %x[in_stride]\n"
191 "sub %x[height], %x[height], #0x1\n"
194 "ldr q20, [x25], #0x10\n"
195 "ldr d16, [x25], #0x8\n"
196 "ushll2 v19.8h, v20.16b, #0x0\n"
197 "ushll v18.8h, v16.8b, #0x0\n"
198 "dup v17.2d, v19.d[1]\n"
199 "sub x20, x20, #0x18\n"
200 "ushll v16.8h, v20.8b, #0x0\n"
201 "str q16, [x21, #0x0]\n"
202 "dup v16.2d, v19.d[0]\n"
203 "str d16, [x21, #0x10]\n"
204 "add x21, x21, %x[out_stride]\n"
206 "mov v17.d[1], v18.d[0]\n"
207 "dup v16.2d, v18.d[1]\n"
208 "str q17, [x21, #0x0]\n"
209 "str d16, [x21, #0x10]\n"
210 "add x21, x21, %x[out_stride]\n"
216 "ldr d16, [x25], #0x8\n"
217 "ld1 { v16.s }[2], [x25], #0x4\n"
218 "sub x20, x20, #0xc\n"
220 "ushll v17.8h, v16.8b, #0x0\n"
221 "ushll2 v16.8h, v16.16b, #0x0\n"
222 "str q17, [x21, #0x0]\n"
223 "str d16, [x21, #0x10]\n"
224 "add x21, x21, %x[out_stride]\n"
230 "ldr s16, [x25], #0x4\n"
231 "sub x20, x20, #0x4\n"
233 "ushll v16.8h, v16.8b, #0x0\n"
234 "str d16, [x21, #0x0]\n"
235 "add x21, x21, #0x8\n"
241 "ldr b16, [x25], #0x1\n"
242 "sub x20, x20, #0x1\n"
244 "ushll v16.8h, v16.8b, #0x0\n"
245 "str h16, [x21, #0x0]\n"
246 "add x21, x21, #0x2\n"
249 "cmp %x[height], #0x1\n"
250 "add %x[out], %x[out], #0x18\n"
253 : [height]
"+&r" (height), [in]
"+&r" (in), [out]
"+&r" (out)
254 : [in_stride]
"r" (in_stride), [out_stride]
"r" (out_stride), [width]
"r" (width)
255 :
"cc",
"memory",
"v0",
"v1",
"v2",
"v3",
"v16",
"v17",
"v18",
"v19",
"v20",
"v21",
"v22",
"v23",
"v24",
"v25",
"v26",
"v27",
"v28",
"v29",
"v30",
"v31",
"x20",
"x21",
"x22",
"x23",
"x24",
"x25"
261 void Transform<12, 1, true, VLType::None>(
262 uint16_t *out,
const uint8_t *in,
int stride,
int x0,
int xmax,
int k0,
int kmax)
264 a64_transpose_interleave_12_u8u16(
266 in + k0 * stride + x0,
268 stride *
sizeof(uint8_t),
274 #endif // defined(__aarch64__)