27 #if defined(__aarch64__)
31 void a64_transpose_interleave_48(uint16_t *out,
const uint16_t *in,
size_t width,
size_t in_stride,
size_t height)
33 size_t out_stride = 24 * 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 q27, [x25], #0x10\n"
51 "ldr q26, [x23], #0x10\n"
52 "sub x24, x24, #0x18\n"
54 "ldr q25, [x22], #0x10\n"
55 "ldr q24, [x20], #0x10\n"
56 "ldr q23, [x25], #0x10\n"
57 "ldr q22, [x23], #0x10\n"
58 "ldr q21, [x22], #0x10\n"
59 "ldr q20, [x20], #0x10\n"
60 "ldr q19, [x25], #0x10\n"
61 "ldr q18, [x23], #0x10\n"
62 "ldr q17, [x22], #0x10\n"
63 "ldr q16, [x20], #0x10\n"
64 "str q27, [x21, #0x0]\n"
65 "str q23, [x21, #0x10]\n"
66 "str q19, [x21, #0x20]\n"
67 "str q26, [x21, #0x30]\n"
68 "str q22, [x21, #0x40]\n"
69 "str q18, [x21, #0x50]\n"
70 "str q25, [x21, #0x60]\n"
71 "str q21, [x21, #0x70]\n"
72 "str q17, [x21, #0x80]\n"
73 "str q24, [x21, #0x90]\n"
74 "str q20, [x21, #0xa0]\n"
75 "str q16, [x21, #0xb0]\n"
76 "add x21, x21, %x[out_stride]\n"
82 "ldr q23, [x25], #0x10\n"
83 "ldr q22, [x23], #0x10\n"
84 "sub x24, x24, #0x10\n"
86 "ldr q21, [x22], #0x10\n"
87 "ldr q20, [x20], #0x10\n"
88 "ldr q19, [x25], #0x10\n"
89 "ldr q18, [x23], #0x10\n"
90 "ldr q17, [x22], #0x10\n"
91 "ldr q16, [x20], #0x10\n"
92 "str q23, [x21, #0x0]\n"
93 "str q19, [x21, #0x10]\n"
94 "str q22, [x21, #0x30]\n"
95 "str q18, [x21, #0x40]\n"
96 "str q21, [x21, #0x60]\n"
97 "str q17, [x21, #0x70]\n"
98 "str q20, [x21, #0x90]\n"
99 "str q16, [x21, #0xa0]\n"
100 "add x21, x21, #0x20\n"
106 "ldr d19, [x25], #0x8\n"
107 "ldr d18, [x23], #0x8\n"
108 "sub x24, x24, #0x4\n"
110 "ldr d17, [x22], #0x8\n"
111 "ldr d16, [x20], #0x8\n"
112 "str d19, [x21, #0x0]\n"
113 "str d18, [x21, #0x30]\n"
114 "str d17, [x21, #0x60]\n"
115 "str d16, [x21, #0x90]\n"
116 "add x21, x21, #0x8\n"
122 "ldr h19, [x25], #0x2\n"
123 "ldr h18, [x23], #0x2\n"
124 "sub x24, x24, #0x1\n"
126 "ldr h17, [x22], #0x2\n"
127 "ldr h16, [x20], #0x2\n"
128 "str h19, [x21, #0x0]\n"
129 "str h18, [x21, #0x30]\n"
130 "str h17, [x21, #0x60]\n"
131 "str h16, [x21, #0x90]\n"
132 "add x21, x21, #0x2\n"
135 "cmp %x[height], #0x4\n"
136 "add %x[out], %x[out], #0xc0\n"
138 "cbz %x[height], 20f\n"
141 "mov x20, %x[width]\n"
144 "add %x[in], x25, %x[in_stride]\n"
146 "sub %x[height], %x[height], #0x1\n"
149 "ldr q18, [x25], #0x10\n"
150 "ldr q17, [x25], #0x10\n"
151 "sub x20, x20, #0x18\n"
153 "ldr q16, [x25], #0x10\n"
154 "str q18, [x21, #0x0]\n"
155 "str q17, [x21, #0x10]\n"
156 "str q16, [x21, #0x20]\n"
157 "add x21, x21, %x[out_stride]\n"
163 "ldr q17, [x25], #0x10\n"
164 "ldr q16, [x25], #0x10\n"
165 "sub x20, x20, #0x10\n"
167 "str q17, [x21, #0x0]\n"
168 "str q16, [x21, #0x10]\n"
169 "add x21, x21, #0x20\n"
175 "ldr d16, [x25], #0x8\n"
176 "sub x20, x20, #0x4\n"
178 "str d16, [x21, #0x0]\n"
179 "add x21, x21, #0x8\n"
185 "ldr h16, [x25], #0x2\n"
186 "sub x20, x20, #0x1\n"
188 "str h16, [x21, #0x0]\n"
189 "add x21, x21, #0x2\n"
192 "cmp %x[height], #0x1\n"
193 "add %x[out], %x[out], #0x30\n"
196 : [height]
"+&r" (height), [in]
"+&r" (in), [out]
"+&r" (out)
197 : [in_stride]
"r" (in_stride), [out_stride]
"r" (out_stride), [width]
"r" (width)
198 :
"cc",
"memory",
"v16",
"v17",
"v18",
"v19",
"v20",
"v21",
"v22",
"v23",
"v24",
"v25",
"v26",
"v27",
"x20",
"x21",
"x22",
"x23",
"x24",
"x25"
205 void Transform<12, 1, true, VLType::None>(
206 float *out,
const float *in,
int stride,
int x0,
int xmax,
int k0,
int kmax)
208 a64_transpose_interleave_48(
209 reinterpret_cast<uint16_t *
>(out),
210 reinterpret_cast<const uint16_t *
>(in + k0 * stride + x0),
211 (xmax-x0) *
sizeof(
float) / 2,
212 stride *
sizeof(
float),
218 void Transform<24, 1, true, VLType::None>(
219 __fp16 *out,
const __fp16 *in,
int stride,
int x0,
int xmax,
int k0,
int kmax)
221 a64_transpose_interleave_48(
222 reinterpret_cast<uint16_t *
>(out),
223 reinterpret_cast<const uint16_t *
>(in + k0 * stride + x0),
224 (xmax-x0) *
sizeof(__fp16) / 2,
225 stride *
sizeof(__fp16),
231 void Transform<6, 1, true, VLType::None>(
232 double *out,
const double *in,
int stride,
int x0,
int xmax,
int k0,
int kmax)
234 a64_transpose_interleave_48(
235 reinterpret_cast<uint16_t *
>(out),
236 reinterpret_cast<const uint16_t *
>(in + k0 * stride + x0),
237 (xmax-x0) *
sizeof(
double) / 2,
238 stride *
sizeof(
double),
244 #endif // defined(__aarch64__)