27 #if defined(__aarch64__)
31 void a64_transpose_interleave_16_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 = 16 * roundup<size_t>(height, 2) *
sizeof(uint16_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 q17, [x9], #0x10\n"
61 "ldr q16, [x28], #0x10\n"
62 "sub x24, x24, #0x10\n"
64 "ldr q19, [x27], #0x10\n"
65 "ldr q18, [x26], #0x10\n"
66 "zip1 v1.8h, v17.8h, v16.8h\n"
67 "zip2 v0.8h, v17.8h, v16.8h\n"
68 "ldr q17, [x25], #0x10\n"
69 "ldr q16, [x23], #0x10\n"
70 "zip1 v31.8h, v19.8h, v18.8h\n"
71 "zip2 v30.8h, v19.8h, v18.8h\n"
72 "ldr q29, [x22], #0x10\n"
73 "ldr q18, [x20], #0x10\n"
74 "zip1 v28.8h, v17.8h, v16.8h\n"
75 "zip2 v27.8h, v17.8h, v16.8h\n"
76 "ldr q17, [x9], #0x10\n"
77 "ldr q16, [x28], #0x10\n"
78 "zip1 v26.8h, v17.8h, v16.8h\n"
79 "zip2 v25.8h, v17.8h, v16.8h\n"
80 "ldr q17, [x27], #0x10\n"
81 "ldr q16, [x26], #0x10\n"
82 "zip1 v24.8h, v17.8h, v16.8h\n"
83 "zip2 v23.8h, v17.8h, v16.8h\n"
84 "ldr q17, [x25], #0x10\n"
85 "ldr q16, [x23], #0x10\n"
86 "zip1 v22.8h, v17.8h, v16.8h\n"
87 "zip2 v21.8h, v17.8h, v16.8h\n"
88 "ldr q20, [x22], #0x10\n"
89 "ldr q16, [x20], #0x10\n"
90 "zip1 v19.8h, v29.8h, v18.8h\n"
91 "zip2 v18.8h, v29.8h, v18.8h\n"
92 "zip1 v17.8h, v20.8h, v16.8h\n"
93 "zip2 v16.8h, v20.8h, v16.8h\n"
94 "str q1, [x21, #0x0]\n"
95 "str q0, [x21, #0x10]\n"
96 "str q26, [x21, #0x20]\n"
97 "str q25, [x21, #0x30]\n"
98 "str q31, [x21, #0x40]\n"
99 "str q30, [x21, #0x50]\n"
100 "str q24, [x21, #0x60]\n"
101 "str q23, [x21, #0x70]\n"
102 "str q28, [x21, #0x80]\n"
103 "str q27, [x21, #0x90]\n"
104 "str q22, [x21, #0xa0]\n"
105 "str q21, [x21, #0xb0]\n"
106 "str q19, [x21, #0xc0]\n"
107 "str q18, [x21, #0xd0]\n"
108 "str q17, [x21, #0xe0]\n"
109 "str q16, [x21, #0xf0]\n"
110 "add x21, x21, %x[out_stride]\n"
116 "ldr d19, [x9], #0x8\n"
117 "ldr d18, [x28], #0x8\n"
118 "sub x24, x24, #0x4\n"
120 "ldr d17, [x27], #0x8\n"
121 "ldr d16, [x26], #0x8\n"
122 "zip1 v20.8h, v19.8h, v18.8h\n"
123 "zip1 v19.8h, v17.8h, v16.8h\n"
124 "ldr d17, [x25], #0x8\n"
125 "ldr d16, [x23], #0x8\n"
126 "zip1 v18.8h, v17.8h, v16.8h\n"
127 "ldr d17, [x22], #0x8\n"
128 "ldr d16, [x20], #0x8\n"
129 "str q20, [x21, #0x0]\n"
130 "zip1 v16.8h, v17.8h, v16.8h\n"
131 "str q19, [x21, #0x40]\n"
132 "str q18, [x21, #0x80]\n"
133 "str q16, [x21, #0xc0]\n"
134 "add x21, x21, #0x10\n"
140 "ldr h19, [x9], #0x2\n"
141 "ldr h18, [x28], #0x2\n"
142 "sub x24, x24, #0x1\n"
144 "ldr h17, [x27], #0x2\n"
145 "ldr h16, [x26], #0x2\n"
146 "zip1 v20.8h, v19.8h, v18.8h\n"
147 "zip1 v19.8h, v17.8h, v16.8h\n"
148 "ldr h17, [x25], #0x2\n"
149 "ldr h16, [x23], #0x2\n"
150 "zip1 v18.8h, v17.8h, v16.8h\n"
151 "ldr h17, [x22], #0x2\n"
152 "ldr h16, [x20], #0x2\n"
153 "str s20, [x21, #0x0]\n"
154 "zip1 v16.8h, v17.8h, v16.8h\n"
155 "str s19, [x21, #0x40]\n"
156 "str s18, [x21, #0x80]\n"
157 "str s16, [x21, #0xc0]\n"
158 "add x21, x21, #0x4\n"
161 "cmp %x[height], #0x8\n"
162 "add %x[out], %x[out], #0x100\n"
164 "cbz %x[height], 16f\n"
168 "mov x20, %x[width]\n"
169 "add x28, x9, %x[in_stride]\n"
170 "cmp %x[height], #0x1\n"
171 "add %x[in], x28, %x[in_stride]\n"
172 "csel x28, x28, %x[pad_row], GT\n"
175 "sub %x[height], %x[height], #0x2\n"
178 "ldr q18, [x9], #0x10\n"
179 "ldr q17, [x28], #0x10\n"
180 "sub x20, x20, #0x10\n"
182 "ldr q20, [x9], #0x10\n"
183 "ldr q16, [x28], #0x10\n"
184 "zip1 v19.8h, v18.8h, v17.8h\n"
185 "zip2 v18.8h, v18.8h, v17.8h\n"
186 "zip1 v17.8h, v20.8h, v16.8h\n"
187 "zip2 v16.8h, v20.8h, v16.8h\n"
188 "str q19, [x21, #0x0]\n"
189 "str q18, [x21, #0x10]\n"
190 "str q17, [x21, #0x20]\n"
191 "str q16, [x21, #0x30]\n"
192 "add x21, x21, %x[out_stride]\n"
198 "ldr d17, [x9], #0x8\n"
199 "ldr d16, [x28], #0x8\n"
200 "sub x20, x20, #0x4\n"
202 "zip1 v16.8h, v17.8h, v16.8h\n"
203 "str q16, [x21, #0x0]\n"
204 "add x21, x21, #0x10\n"
210 "ldr h17, [x9], #0x2\n"
211 "ldr h16, [x28], #0x2\n"
212 "sub x20, x20, #0x1\n"
214 "zip1 v16.8h, v17.8h, v16.8h\n"
215 "str s16, [x21, #0x0]\n"
216 "add x21, x21, #0x4\n"
219 "cmp %x[height], #0x1\n"
220 "add %x[out], %x[out], #0x40\n"
223 : [height]
"+&r" (height), [in]
"+&r" (in), [out]
"+&r" (out)
224 : [in_stride]
"r" (in_stride), [out_stride]
"r" (out_stride), [pad_row]
"r" (pad_row), [width]
"r" (width)
225 :
"cc",
"memory",
"v0",
"v1",
"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"
232 void Transform<16, 2, true, VLType::None>(
233 bfloat16 *out,
const bfloat16 *in,
int stride,
int x0,
int xmax,
int k0,
int kmax)
235 a64_transpose_interleave_16_2x2(
236 reinterpret_cast<uint16_t *
>(out),
237 reinterpret_cast<const uint16_t *
>(in + k0 * stride + x0),
245 #endif // defined(__aarch64__)