27 #if defined(__aarch64__)
31 void a64_transpose_interleave_24_bf16fp32(
float *out,
const bfloat16 *in,
size_t width,
size_t in_stride,
size_t height)
33 size_t out_stride = 12 * height *
sizeof(float);
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 q18, [x25], #0x10\n"
51 "ldr q17, [x23], #0x10\n"
52 "sub x24, x24, #0x18\n"
53 "shll v26.4s, v18.4h, #0x10\n"
54 "ldr q16, [x22], #0x10\n"
55 "ldr q25, [x20], #0x10\n"
56 "shll2 v24.4s, v18.8h, #0x10\n"
57 "shll v5.4s, v17.4h, #0x10\n"
58 "ldr q23, [x25], #0x10\n"
59 "ldr q22, [x23], #0x10\n"
60 "shll v21.4s, v23.4h, #0x10\n"
61 "shll2 v4.4s, v17.8h, #0x10\n"
62 "ldr q20, [x22], #0x10\n"
63 "ldr q3, [x20], #0x10\n"
64 "shll v2.4s, v22.4h, #0x10\n"
65 "shll v1.4s, v16.4h, #0x10\n"
66 "ldr q19, [x25], #0x10\n"
67 "ldr q18, [x23], #0x10\n"
68 "shll2 v0.4s, v16.8h, #0x10\n"
69 "shll v31.4s, v20.4h, #0x10\n"
70 "ldr q17, [x22], #0x10\n"
71 "ldr q16, [x20], #0x10\n"
72 "shll v30.4s, v25.4h, #0x10\n"
73 "shll2 v29.4s, v25.8h, #0x10\n"
74 "shll v28.4s, v3.4h, #0x10\n"
75 "str q26, [x21, #0x0]\n"
77 "shll2 v27.4s, v23.8h, #0x10\n"
78 "str q24, [x21, #0x10]\n"
79 "shll v26.4s, v19.4h, #0x10\n"
80 "shll2 v25.4s, v19.8h, #0x10\n"
81 "str q21, [x21, #0x20]\n"
82 "shll2 v24.4s, v22.8h, #0x10\n"
83 "shll v23.4s, v18.4h, #0x10\n"
84 "str q5, [x21, #0x30]\n"
85 "shll2 v22.4s, v18.8h, #0x10\n"
86 "shll2 v21.4s, v20.8h, #0x10\n"
87 "str q4, [x21, #0x40]\n"
88 "shll v20.4s, v17.4h, #0x10\n"
89 "shll2 v19.4s, v17.8h, #0x10\n"
90 "str q2, [x21, #0x50]\n"
91 "shll2 v18.4s, v3.8h, #0x10\n"
92 "shll v17.4s, v16.4h, #0x10\n"
93 "str q1, [x21, #0x60]\n"
94 "shll2 v16.4s, v16.8h, #0x10\n"
95 "str q0, [x21, #0x70]\n"
96 "str q31, [x21, #0x80]\n"
97 "str q30, [x21, #0x90]\n"
98 "str q29, [x21, #0xa0]\n"
99 "str q28, [x21, #0xb0]\n"
100 "add x21, x21, %x[out_stride]\n"
101 "str q27, [x21, #0x0]\n"
102 "str q26, [x21, #0x10]\n"
103 "str q25, [x21, #0x20]\n"
104 "str q24, [x21, #0x30]\n"
105 "str q23, [x21, #0x40]\n"
106 "str q22, [x21, #0x50]\n"
107 "str q21, [x21, #0x60]\n"
108 "str q20, [x21, #0x70]\n"
109 "str q19, [x21, #0x80]\n"
110 "str q18, [x21, #0x90]\n"
111 "str q17, [x21, #0xa0]\n"
112 "str q16, [x21, #0xb0]\n"
113 "add x21, x21, %x[out_stride]\n"
119 "ldr q16, [x25], #0x10\n"
120 "ldr q21, [x23], #0x10\n"
121 "sub x24, x24, #0xc\n"
123 "ldr q20, [x22], #0x10\n"
124 "ldr q27, [x20], #0x10\n"
125 "shll v19.4s, v16.4h, #0x10\n"
126 "shll2 v26.4s, v16.8h, #0x10\n"
127 "ldr d16, [x25], #0x8\n"
128 "ldr d18, [x23], #0x8\n"
129 "shll v25.4s, v16.4h, #0x10\n"
130 "shll v24.4s, v21.4h, #0x10\n"
131 "ldr d17, [x22], #0x8\n"
132 "ldr d16, [x20], #0x8\n"
133 "shll2 v23.4s, v21.8h, #0x10\n"
134 "shll v22.4s, v18.4h, #0x10\n"
135 "shll v21.4s, v20.4h, #0x10\n"
136 "shll2 v20.4s, v20.8h, #0x10\n"
137 "str q19, [x21, #0x0]\n"
138 "shll v19.4s, v17.4h, #0x10\n"
139 "shll v18.4s, v27.4h, #0x10\n"
140 "str q26, [x21, #0x10]\n"
141 "shll2 v17.4s, v27.8h, #0x10\n"
142 "shll v16.4s, v16.4h, #0x10\n"
143 "str q25, [x21, #0x20]\n"
144 "str q24, [x21, #0x30]\n"
145 "str q23, [x21, #0x40]\n"
146 "str q22, [x21, #0x50]\n"
147 "str q21, [x21, #0x60]\n"
148 "str q20, [x21, #0x70]\n"
149 "str q19, [x21, #0x80]\n"
150 "str q18, [x21, #0x90]\n"
151 "str q17, [x21, #0xa0]\n"
152 "str q16, [x21, #0xb0]\n"
153 "add x21, x21, %x[out_stride]\n"
159 "ldr d19, [x25], #0x8\n"
160 "ldr d18, [x23], #0x8\n"
161 "sub x24, x24, #0x4\n"
163 "ldr d17, [x22], #0x8\n"
164 "ldr d16, [x20], #0x8\n"
165 "shll v19.4s, v19.4h, #0x10\n"
166 "shll v18.4s, v18.4h, #0x10\n"
167 "shll v17.4s, v17.4h, #0x10\n"
168 "shll v16.4s, v16.4h, #0x10\n"
169 "str q19, [x21, #0x0]\n"
170 "str q18, [x21, #0x30]\n"
171 "str q17, [x21, #0x60]\n"
172 "str q16, [x21, #0x90]\n"
173 "add x21, x21, #0x10\n"
179 "ldr h19, [x25], #0x2\n"
180 "ldr h18, [x23], #0x2\n"
181 "sub x24, x24, #0x1\n"
183 "ldr h17, [x22], #0x2\n"
184 "ldr h16, [x20], #0x2\n"
185 "shll v19.4s, v19.4h, #0x10\n"
186 "shll v18.4s, v18.4h, #0x10\n"
187 "shll v17.4s, v17.4h, #0x10\n"
188 "shll v16.4s, v16.4h, #0x10\n"
189 "str s19, [x21, #0x0]\n"
190 "str s18, [x21, #0x30]\n"
191 "str s17, [x21, #0x60]\n"
192 "str s16, [x21, #0x90]\n"
193 "add x21, x21, #0x4\n"
196 "cmp %x[height], #0x4\n"
197 "add %x[out], %x[out], #0xc0\n"
199 "cbz %x[height], 20f\n"
202 "mov x20, %x[width]\n"
205 "add %x[in], x25, %x[in_stride]\n"
207 "sub %x[height], %x[height], #0x1\n"
210 "ldr q16, [x25], #0x10\n"
211 "ldr q20, [x25], #0x10\n"
212 "sub x20, x20, #0x18\n"
213 "shll v18.4s, v16.4h, #0x10\n"
214 "ldr q19, [x25], #0x10\n"
215 "shll2 v17.4s, v16.8h, #0x10\n"
216 "shll v16.4s, v20.4h, #0x10\n"
217 "str q18, [x21, #0x0]\n"
218 "str q17, [x21, #0x10]\n"
220 "shll2 v18.4s, v20.8h, #0x10\n"
221 "shll v17.4s, v19.4h, #0x10\n"
222 "str q16, [x21, #0x20]\n"
223 "add x21, x21, %x[out_stride]\n"
224 "shll2 v16.4s, v19.8h, #0x10\n"
225 "str q18, [x21, #0x0]\n"
226 "str q17, [x21, #0x10]\n"
227 "str q16, [x21, #0x20]\n"
228 "add x21, x21, %x[out_stride]\n"
234 "ldr q17, [x25], #0x10\n"
235 "ldr d18, [x25], #0x8\n"
236 "sub x20, x20, #0xc\n"
238 "shll v16.4s, v17.4h, #0x10\n"
239 "shll2 v17.4s, v17.8h, #0x10\n"
240 "str q16, [x21, #0x0]\n"
241 "shll v16.4s, v18.4h, #0x10\n"
242 "str q17, [x21, #0x10]\n"
243 "str q16, [x21, #0x20]\n"
244 "add x21, x21, %x[out_stride]\n"
250 "ldr d16, [x25], #0x8\n"
251 "sub x20, x20, #0x4\n"
253 "shll v16.4s, v16.4h, #0x10\n"
254 "str q16, [x21, #0x0]\n"
255 "add x21, x21, #0x10\n"
261 "ldr h16, [x25], #0x2\n"
262 "sub x20, x20, #0x1\n"
264 "shll v16.4s, v16.4h, #0x10\n"
265 "str s16, [x21, #0x0]\n"
266 "add x21, x21, #0x4\n"
269 "cmp %x[height], #0x1\n"
270 "add %x[out], %x[out], #0x30\n"
273 : [height]
"+&r" (height), [in]
"+&r" (in), [out]
"+&r" (out)
274 : [in_stride]
"r" (in_stride), [out_stride]
"r" (out_stride), [width]
"r" (width)
275 :
"cc",
"memory",
"v0",
"v1",
"v2",
"v3",
"v4",
"v5",
"v16",
"v17",
"v18",
"v19",
"v20",
"v21",
"v22",
"v23",
"v24",
"v25",
"v26",
"v27",
"v28",
"v29",
"v30",
"v31",
"x20",
"x21",
"x22",
"x23",
"x24",
"x25"
281 void Transform<12, 1, true, VLType::None>(
282 float *out,
const bfloat16 *in,
int stride,
int x0,
int xmax,
int k0,
int kmax)
284 a64_transpose_interleave_24_bf16fp32(
286 in + k0 * stride + x0,
294 #endif // defined(__aarch64__)