27 #if defined(__aarch64__)
31 void a64_transpose_interleave_64(uint16_t *out,
const uint16_t *in,
size_t width,
size_t in_stride,
size_t height)
33 size_t out_stride = 32 * 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 q31, [x25], #0x10\n"
51 "ldr q30, [x23], #0x10\n"
52 "sub x24, x24, #0x20\n"
54 "ldr q29, [x22], #0x10\n"
55 "ldr q28, [x20], #0x10\n"
56 "ldr q27, [x25], #0x10\n"
57 "ldr q26, [x23], #0x10\n"
58 "ldr q25, [x22], #0x10\n"
59 "ldr q24, [x20], #0x10\n"
60 "ldr q23, [x25], #0x10\n"
61 "ldr q22, [x23], #0x10\n"
62 "ldr q21, [x22], #0x10\n"
63 "ldr q20, [x20], #0x10\n"
64 "ldr q19, [x25], #0x10\n"
65 "ldr q18, [x23], #0x10\n"
66 "ldr q17, [x22], #0x10\n"
67 "ldr q16, [x20], #0x10\n"
68 "str q31, [x21, #0x0]\n"
69 "str q27, [x21, #0x10]\n"
70 "str q23, [x21, #0x20]\n"
71 "str q19, [x21, #0x30]\n"
72 "str q30, [x21, #0x40]\n"
73 "str q26, [x21, #0x50]\n"
74 "str q22, [x21, #0x60]\n"
75 "str q18, [x21, #0x70]\n"
76 "str q29, [x21, #0x80]\n"
77 "str q25, [x21, #0x90]\n"
78 "str q21, [x21, #0xa0]\n"
79 "str q17, [x21, #0xb0]\n"
80 "str q28, [x21, #0xc0]\n"
81 "str q24, [x21, #0xd0]\n"
82 "str q20, [x21, #0xe0]\n"
83 "str q16, [x21, #0xf0]\n"
84 "add x21, x21, %x[out_stride]\n"
90 "ldr q23, [x25], #0x10\n"
91 "ldr q22, [x23], #0x10\n"
92 "sub x24, x24, #0x10\n"
94 "ldr q21, [x22], #0x10\n"
95 "ldr q20, [x20], #0x10\n"
96 "ldr q19, [x25], #0x10\n"
97 "ldr q18, [x23], #0x10\n"
98 "ldr q17, [x22], #0x10\n"
99 "ldr q16, [x20], #0x10\n"
100 "str q23, [x21, #0x0]\n"
101 "str q19, [x21, #0x10]\n"
102 "str q22, [x21, #0x40]\n"
103 "str q18, [x21, #0x50]\n"
104 "str q21, [x21, #0x80]\n"
105 "str q17, [x21, #0x90]\n"
106 "str q20, [x21, #0xc0]\n"
107 "str q16, [x21, #0xd0]\n"
108 "add x21, x21, #0x20\n"
114 "ldr d19, [x25], #0x8\n"
115 "ldr d18, [x23], #0x8\n"
116 "sub x24, x24, #0x4\n"
118 "ldr d17, [x22], #0x8\n"
119 "ldr d16, [x20], #0x8\n"
120 "str d19, [x21, #0x0]\n"
121 "str d18, [x21, #0x40]\n"
122 "str d17, [x21, #0x80]\n"
123 "str d16, [x21, #0xc0]\n"
124 "add x21, x21, #0x8\n"
130 "ldr h19, [x25], #0x2\n"
131 "ldr h18, [x23], #0x2\n"
132 "sub x24, x24, #0x1\n"
134 "ldr h17, [x22], #0x2\n"
135 "ldr h16, [x20], #0x2\n"
136 "str h19, [x21, #0x0]\n"
137 "str h18, [x21, #0x40]\n"
138 "str h17, [x21, #0x80]\n"
139 "str h16, [x21, #0xc0]\n"
140 "add x21, x21, #0x2\n"
143 "cmp %x[height], #0x4\n"
144 "add %x[out], %x[out], #0x100\n"
146 "cbz %x[height], 20f\n"
149 "mov x20, %x[width]\n"
152 "add %x[in], x25, %x[in_stride]\n"
154 "sub %x[height], %x[height], #0x1\n"
157 "ldr q19, [x25], #0x10\n"
158 "ldr q18, [x25], #0x10\n"
159 "sub x20, x20, #0x20\n"
161 "ldr q17, [x25], #0x10\n"
162 "ldr q16, [x25], #0x10\n"
163 "str q19, [x21, #0x0]\n"
164 "str q18, [x21, #0x10]\n"
165 "str q17, [x21, #0x20]\n"
166 "str q16, [x21, #0x30]\n"
167 "add x21, x21, %x[out_stride]\n"
173 "ldr q17, [x25], #0x10\n"
174 "ldr q16, [x25], #0x10\n"
175 "sub x20, x20, #0x10\n"
177 "str q17, [x21, #0x0]\n"
178 "str q16, [x21, #0x10]\n"
179 "add x21, x21, #0x20\n"
185 "ldr d16, [x25], #0x8\n"
186 "sub x20, x20, #0x4\n"
188 "str d16, [x21, #0x0]\n"
189 "add x21, x21, #0x8\n"
195 "ldr h16, [x25], #0x2\n"
196 "sub x20, x20, #0x1\n"
198 "str h16, [x21, #0x0]\n"
199 "add x21, x21, #0x2\n"
202 "cmp %x[height], #0x1\n"
203 "add %x[out], %x[out], #0x40\n"
206 : [height]
"+&r" (height), [in]
"+&r" (in), [out]
"+&r" (out)
207 : [in_stride]
"r" (in_stride), [out_stride]
"r" (out_stride), [width]
"r" (width)
208 :
"cc",
"memory",
"v16",
"v17",
"v18",
"v19",
"v20",
"v21",
"v22",
"v23",
"v24",
"v25",
"v26",
"v27",
"v28",
"v29",
"v30",
"v31",
"x20",
"x21",
"x22",
"x23",
"x24",
"x25"
215 void Transform<16, 1, true, VLType::None>(
216 float *out,
const float *in,
int stride,
int x0,
int xmax,
int k0,
int kmax)
218 a64_transpose_interleave_64(
219 reinterpret_cast<uint16_t *
>(out),
220 reinterpret_cast<const uint16_t *
>(in + k0 * stride + x0),
221 (xmax-x0) *
sizeof(
float) / 2,
222 stride *
sizeof(
float),
228 void Transform<32, 1, true, VLType::None>(
229 __fp16 *out,
const __fp16 *in,
int stride,
int x0,
int xmax,
int k0,
int kmax)
231 a64_transpose_interleave_64(
232 reinterpret_cast<uint16_t *
>(out),
233 reinterpret_cast<const uint16_t *
>(in + k0 * stride + x0),
234 (xmax-x0) *
sizeof(__fp16) / 2,
235 stride *
sizeof(__fp16),
241 void Transform<32, 1, true, VLType::None>(
242 uint16_t *out,
const uint16_t *in,
int stride,
int x0,
int xmax,
int k0,
int kmax)
244 a64_transpose_interleave_64(
245 reinterpret_cast<uint16_t *
>(out),
246 reinterpret_cast<const uint16_t *
>(in + k0 * stride + x0),
247 (xmax-x0) *
sizeof(uint16_t) / 2,
248 stride *
sizeof(uint16_t),
254 #endif // defined(__aarch64__)