32 const float *
const *
const input_ptrs,
33 float *
const *
const outptrs,
35 unsigned int n_channels,
36 const float activation_min,
37 const float activation_max
42 float *
const *outptrs;
45 const float *inptrs[25];
48 const float *
const *
const input_ptrs,
49 float *
const *
const outptrs,
50 const void *
const params,
53 ) : outptrs(outptrs), params(params), min(min), max(max)
55 inptrs[0] = input_ptrs[12];
56 inptrs[1] = input_ptrs[0];
57 inptrs[2] = input_ptrs[1];
58 inptrs[3] = input_ptrs[3];
59 inptrs[4] = input_ptrs[4];
60 inptrs[5] = input_ptrs[5];
61 inptrs[6] = input_ptrs[6];
62 inptrs[7] = input_ptrs[2];
63 inptrs[8] = input_ptrs[8];
64 inptrs[9] = input_ptrs[9];
65 inptrs[10] = input_ptrs[7];
66 inptrs[11] = input_ptrs[15];
67 inptrs[12] = input_ptrs[10];
68 inptrs[13] = input_ptrs[16];
69 inptrs[14] = input_ptrs[11];
70 inptrs[15] = input_ptrs[18];
71 inptrs[16] = input_ptrs[13];
72 inptrs[17] = input_ptrs[19];
73 inptrs[18] = input_ptrs[20];
74 inptrs[19] = input_ptrs[14];
75 inptrs[20] = input_ptrs[21];
76 inptrs[21] = input_ptrs[17];
77 inptrs[22] = input_ptrs[23];
78 inptrs[23] = input_ptrs[22];
79 inptrs[24] = input_ptrs[24];
84 Args params_struct(input_ptrs, outptrs, params,
85 activation_min, activation_max);
88 "ldr x21, [%x[params_struct], %[offsetof_args_outptrs]]\n" 89 "add x16, %x[params_struct], %[offsetof_Args_inptrs]\n" 90 "ldr x15, [%x[params_struct], %[offsetof_args_params]]\n" 91 "add x20, %x[params_struct], %[offsetof_args_min]\n" 92 "add x19, %x[params_struct], %[offsetof_args_max]\n" 93 "ld1r { v19.4s }, [x20]\n" 94 "ld1r { v18.4s }, [x19]\n" 96 "ldp x13, x12, [x21, #0x0]\n" 98 "ldp x10, x9, [x21, #0x10]\n" 100 "lsr x27, %x[n_channels], #0x2\n" 102 "ldr q17, [x15, #0x0]\n" 103 "ldr q0, [x15, #0x10]\n" 104 "cmp x11, x27, LSL #4\n" 105 "ldr q1, [x15, #0x20]\n" 106 "ldr q2, [x15, #0x30]\n" 107 "ldr q3, [x15, #0x40]\n" 108 "ldr q4, [x15, #0x50]\n" 109 "ldr q5, [x15, #0x60]\n" 110 "ldr q6, [x15, #0x70]\n" 111 "ldr q7, [x15, #0x80]\n" 112 "ldr q8, [x15, #0x90]\n" 113 "add x15, x15, #0xa0\n" 114 "ldp x26, x25, [x16, #0x0]\n" 115 "ldp x24, x23, [x16, #0x10]\n" 116 "ldp x22, x21, [x16, #0x20]\n" 117 "ldr q9, [x26, x14]\n" 118 "ldr q10, [x25, x14]\n" 119 "ldr q11, [x24, x14]\n" 120 "ldr q12, [x23, x14]\n" 121 "ldr q13, [x22, x14]\n" 122 "ldr q14, [x21, x14]\n" 123 "ldp x20, x19, [x16, #0x30]\n" 124 "ldr q15, [x20, x14]\n" 125 "ldr q16, [x19, x14]\n" 128 "mov v31.16b, v17.16b\n fmla v31.4s, v8.4s, v9.4s\n" 129 "ldr x26, [x16, #0x40]\n" 130 "add x28, x28, #0x10\n" 131 "mov v30.16b, v17.16b\n fmla v30.4s, v6.4s, v9.4s\n" 132 "ldr x25, [x16, #0x48]\n" 133 "mov v29.16b, v17.16b\n fmla v29.4s, v2.4s, v9.4s\n" 134 "ldr x24, [x16, #0x50]\n" 135 "mov v28.16b, v17.16b\n fmla v28.4s, v0.4s, v9.4s\n" 136 "ldr x23, [x16, #0x58]\n" 137 "ldr x22, [x16, #0x60]\n" 138 "fmla v31.4s, v0.4s, v10.4s\n" 139 "ldr x21, [x16, #0x68]\n" 140 "fmla v30.4s, v1.4s, v12.4s\n" 141 "ldr q12, [x25, x14]\n" 142 "fmla v31.4s, v1.4s, v11.4s\n" 143 "ldr q11, [x26, x14]\n" 144 "ldr x20, [x16, #0x70]\n" 145 "fmla v30.4s, v2.4s, v13.4s\n" 146 "ldr q13, [x24, x14]\n" 147 "fmla v31.4s, v3.4s, v14.4s\n" 148 "ldr q14, [x23, x14]\n" 149 "ldr x19, [x16, #0x78]\n" 150 "fmla v30.4s, v0.4s, v16.4s\n" 151 "ldr x26, [x16, #0x80]\n" 152 "fmla v31.4s, v4.4s, v15.4s\n" 153 "ldr q15, [x22, x14]\n" 154 "fmla v29.4s, v3.4s, v14.4s\n" 155 "ldr x25, [x16, #0x88]\n" 156 "fmla v30.4s, v4.4s, v11.4s\n" 157 "ldr q11, [x21, x14]\n" 158 "ldr x24, [x16, #0x90]\n" 159 "fmla v31.4s, v2.4s, v16.4s\n" 160 "ldr q16, [x20, x14]\n" 161 "fmla v29.4s, v0.4s, v15.4s\n" 162 "ldr q14, [x25, x14]\n" 163 "fmla v30.4s, v5.4s, v12.4s\n" 164 "ldr q12, [x26, x14]\n" 165 "ldr x23, [x16, #0x98]\n" 166 "fmla v31.4s, v5.4s, v13.4s\n" 167 "ldr x22, [x16, #0xa0]\n" 168 "fmla v29.4s, v4.4s, v11.4s\n" 169 "ldr q11, [x23, x14]\n" 170 "fmla v30.4s, v3.4s, v13.4s\n" 171 "ldr q13, [x19, x14]\n" 172 "ldr x21, [x16, #0xa8]\n" 173 "fmla v31.4s, v6.4s, v15.4s\n" 174 "ldr q15, [x24, x14]\n" 175 "fmla v29.4s, v1.4s, v16.4s\n" 176 "ldr x20, [x16, #0xb0]\n" 177 "fmla v30.4s, v7.4s, v12.4s\n" 178 "ldr x19, [x16, #0xb8]\n" 179 "fmla v28.4s, v4.4s, v13.4s\n" 180 "ldr q13, [x22, x14]\n" 181 "ldr x26, [x16, #0xc0]\n" 182 "fmla v31.4s, v7.4s, v16.4s\n" 183 "fmla v29.4s, v6.4s, v15.4s\n" 184 "ldr q16, [x21, x14]\n" 185 "fmla v30.4s, v8.4s, v11.4s\n" 186 "ldr q15, [x19, x14]\n" 187 "fmla v28.4s, v1.4s, v12.4s\n" 188 "ldr q17, [x15, #0x0]\n" 189 "ldr q0, [x15, #0x10]\n" 190 "fmla v29.4s, v7.4s, v13.4s\n" 191 "fmax v31.4s, v31.4s, v19.4s\n" 192 "ldr q1, [x15, #0x20]\n" 193 "fmax v30.4s, v30.4s, v19.4s\n" 194 "ldr q4, [x15, #0x50]\n" 195 "fmla v28.4s, v5.4s, v14.4s\n" 196 "ldr q14, [x20, x14]\n" 197 "fmin v31.4s, v31.4s, v18.4s\n" 198 "str q31, [x13, x28]\n" 199 "fmla v28.4s, v2.4s, v11.4s\n" 200 "fmla v29.4s, v5.4s, v16.4s\n" 201 "ldr q11, [x26, x14]\n" 202 "add x14, x14, #0x10\n" 203 "fmin v30.4s, v30.4s, v18.4s\n" 204 "ldp x26, x25, [x16, #0x0]\n" 205 "ldp x24, x23, [x16, #0x10]\n" 206 "fmla v28.4s, v3.4s, v16.4s\n" 207 "ldp x22, x21, [x16, #0x20]\n" 208 "fmla v29.4s, v8.4s, v15.4s\n" 209 "ldr q9, [x26, x11]\n" 210 "ldr q10, [x25, x11]\n" 211 "fmla v28.4s, v7.4s, v14.4s\n" 212 "ldr q12, [x23, x11]\n" 213 "fmax v29.4s, v29.4s, v19.4s\n" 214 "ldr q13, [x22, x11]\n" 215 "ldr q14, [x21, x11]\n" 216 "fmin v29.4s, v29.4s, v18.4s\n" 217 "ldp x20, x19, [x16, #0x30]\n" 218 "str q30, [x12, x28]\n" 219 "fmla v28.4s, v6.4s, v15.4s\n" 220 "ldr q2, [x15, #0x30]\n" 221 "fmla v28.4s, v8.4s, v11.4s\n" 222 "ldr q11, [x24, x11]\n" 223 "ldr q15, [x20, x11]\n" 224 "fmax v28.4s, v28.4s, v19.4s\n" 225 "ldr q16, [x19, x11]\n" 226 "add x11, x11, #0x10\n" 227 "fmin v28.4s, v28.4s, v18.4s\n" 228 "str q29, [x10, x28]\n" 229 "cmp x11, x27, LSL #4\n" 230 "ldr q3, [x15, #0x40]\n" 231 "ldr q5, [x15, #0x60]\n" 232 "ldr q6, [x15, #0x70]\n" 233 "str q28, [x9, x28]\n" 234 "ldr q7, [x15, #0x80]\n" 235 "ldr q8, [x15, #0x90]\n" 236 "add x15, x15, #0xa0\n" 239 "mov v31.16b, v17.16b\n fmla v31.4s, v8.4s, v9.4s\n" 240 "ldr x26, [x16, #0x40]\n" 241 "add x28, x28, #0x10\n" 242 "mov v30.16b, v17.16b\n fmla v30.4s, v6.4s, v9.4s\n" 243 "ldr x25, [x16, #0x48]\n" 244 "mov v29.16b, v17.16b\n fmla v29.4s, v2.4s, v9.4s\n" 245 "ldr x24, [x16, #0x50]\n" 246 "mov v28.16b, v17.16b\n fmla v28.4s, v0.4s, v9.4s\n" 247 "ldr x23, [x16, #0x58]\n" 248 "ldr x22, [x16, #0x60]\n" 249 "fmla v31.4s, v0.4s, v10.4s\n" 250 "ldr x21, [x16, #0x68]\n" 251 "fmla v30.4s, v1.4s, v12.4s\n" 252 "ldr q12, [x25, x14]\n" 253 "fmla v31.4s, v1.4s, v11.4s\n" 254 "ldr q11, [x26, x14]\n" 255 "ldr x20, [x16, #0x70]\n" 256 "fmla v30.4s, v2.4s, v13.4s\n" 257 "ldr q13, [x24, x14]\n" 258 "fmla v31.4s, v3.4s, v14.4s\n" 259 "ldr q14, [x23, x14]\n" 260 "ldr x19, [x16, #0x78]\n" 261 "fmla v30.4s, v0.4s, v16.4s\n" 262 "ldr x26, [x16, #0x80]\n" 263 "fmla v31.4s, v4.4s, v15.4s\n" 264 "ldr q15, [x22, x14]\n" 265 "fmla v29.4s, v3.4s, v14.4s\n" 266 "ldr x25, [x16, #0x88]\n" 267 "fmla v30.4s, v4.4s, v11.4s\n" 268 "ldr q11, [x21, x14]\n" 269 "ldr x24, [x16, #0x90]\n" 270 "fmla v31.4s, v2.4s, v16.4s\n" 271 "ldr q16, [x20, x14]\n" 272 "fmla v29.4s, v0.4s, v15.4s\n" 273 "ldr q14, [x25, x14]\n" 274 "fmla v30.4s, v5.4s, v12.4s\n" 275 "ldr q12, [x26, x14]\n" 276 "ldr x23, [x16, #0x98]\n" 277 "fmla v31.4s, v5.4s, v13.4s\n" 278 "ldr x22, [x16, #0xa0]\n" 279 "fmla v29.4s, v4.4s, v11.4s\n" 280 "ldr q11, [x23, x14]\n" 281 "fmla v30.4s, v3.4s, v13.4s\n" 282 "ldr q13, [x19, x14]\n" 283 "ldr x21, [x16, #0xa8]\n" 284 "fmla v31.4s, v6.4s, v15.4s\n" 285 "ldr q15, [x24, x14]\n" 286 "fmla v29.4s, v1.4s, v16.4s\n" 287 "ldr x20, [x16, #0xb0]\n" 288 "fmla v30.4s, v7.4s, v12.4s\n" 289 "ldr x19, [x16, #0xb8]\n" 290 "fmla v28.4s, v4.4s, v13.4s\n" 291 "ldr q13, [x22, x14]\n" 292 "ldr x26, [x16, #0xc0]\n" 293 "fmla v31.4s, v7.4s, v16.4s\n" 294 "fmla v29.4s, v6.4s, v15.4s\n" 295 "ldr q16, [x21, x14]\n" 296 "fmla v30.4s, v8.4s, v11.4s\n" 297 "ldr q15, [x19, x14]\n" 298 "fmla v28.4s, v1.4s, v12.4s\n" 299 "fmla v29.4s, v7.4s, v13.4s\n" 300 "fmax v31.4s, v31.4s, v19.4s\n" 301 "fmax v30.4s, v30.4s, v19.4s\n" 302 "fmla v28.4s, v5.4s, v14.4s\n" 303 "ldr q14, [x20, x14]\n" 304 "fmin v31.4s, v31.4s, v18.4s\n" 305 "str q31, [x13, x28]\n" 306 "fmla v28.4s, v2.4s, v11.4s\n" 307 "fmla v29.4s, v5.4s, v16.4s\n" 308 "ldr q11, [x26, x14]\n" 309 "add x14, x14, #0x10\n" 310 "fmin v30.4s, v30.4s, v18.4s\n" 311 "str q30, [x12, x28]\n" 312 "fmla v28.4s, v3.4s, v16.4s\n" 313 "fmla v29.4s, v8.4s, v15.4s\n" 314 "fmla v28.4s, v7.4s, v14.4s\n" 315 "fmax v29.4s, v29.4s, v19.4s\n" 316 "fmin v29.4s, v29.4s, v18.4s\n" 317 "str q29, [x10, x28]\n" 318 "fmla v28.4s, v6.4s, v15.4s\n" 319 "fmla v28.4s, v8.4s, v11.4s\n" 320 "fmax v28.4s, v28.4s, v19.4s\n" 321 "fmin v28.4s, v28.4s, v18.4s\n" 322 "str q28, [x9, x28]\n" 324 "tst %x[n_channels], #0x3\n" 326 "ldr q17, [x15, #0x0]\n" 327 "ldr q0, [x15, #0x10]\n" 329 "ldr q1, [x15, #0x20]\n" 330 "add x13, x13, x28\n" 331 "ldr q2, [x15, #0x30]\n" 332 "add x12, x12, x28\n" 333 "ldr q3, [x15, #0x40]\n" 334 "add x10, x10, x28\n" 335 "ldr q4, [x15, #0x50]\n" 337 "ldr q5, [x15, #0x60]\n" 338 "ldr q6, [x15, #0x70]\n" 339 "ldr q7, [x15, #0x80]\n" 340 "ldr q8, [x15, #0x90]\n" 341 "ldr x26, [x16, #0x0]\n" 342 "ldr x25, [x16, #0x8]\n" 343 "ldr x24, [x16, #0x10]\n" 344 "add x26, x26, x14\n" 345 "ldr x23, [x16, #0x18]\n" 346 "add x25, x25, x14\n" 347 "ldr x22, [x16, #0x20]\n" 348 "add x24, x24, x14\n" 349 "ldr x21, [x16, #0x28]\n" 350 "add x23, x23, x14\n" 351 "ldr x20, [x16, #0x30]\n" 352 "add x22, x22, x14\n" 353 "ldr x19, [x16, #0x38]\n" 354 "add x21, x21, x14\n" 355 "add x20, x20, x14\n" 356 "add x19, x19, x14\n" 357 "tbz %x[n_channels], #1, 4f\n" 358 "ld1 { v9.d }[0], [x26], #0x8\n" 359 "ld1 { v10.d }[0], [x25], #0x8\n" 360 "ld1 { v11.d }[0], [x24], #0x8\n" 361 "ld1 { v12.d }[0], [x23], #0x8\n" 362 "ld1 { v13.d }[0], [x22], #0x8\n" 363 "ld1 { v14.d }[0], [x21], #0x8\n" 364 "ld1 { v15.d }[0], [x20], #0x8\n" 365 "ld1 { v16.d }[0], [x19], #0x8\n" 366 "tbz %x[n_channels], #0, 5f\n" 367 "ld1 { v9.s }[2], [x26], #0x4\n" 368 "ld1 { v10.s }[2], [x25], #0x4\n" 369 "ld1 { v11.s }[2], [x24], #0x4\n" 370 "ld1 { v12.s }[2], [x23], #0x4\n" 371 "ld1 { v13.s }[2], [x22], #0x4\n" 372 "ld1 { v14.s }[2], [x21], #0x4\n" 373 "ld1 { v15.s }[2], [x20], #0x4\n" 374 "ld1 { v16.s }[2], [x19], #0x4\n" 377 "ld1 { v9.s }[0], [x26], #0x4\n" 378 "ld1 { v10.s }[0], [x25], #0x4\n" 379 "ld1 { v11.s }[0], [x24], #0x4\n" 380 "ld1 { v12.s }[0], [x23], #0x4\n" 381 "ld1 { v13.s }[0], [x22], #0x4\n" 382 "ld1 { v14.s }[0], [x21], #0x4\n" 383 "ld1 { v15.s }[0], [x20], #0x4\n" 384 "ld1 { v16.s }[0], [x19], #0x4\n" 386 "mov v31.16b, v17.16b\n fmla v31.4s, v8.4s, v9.4s\n" 387 "ldr x26, [x16, #0x40]\n" 388 "add x26, x26, x14\n" 389 "mov v30.16b, v17.16b\n fmla v30.4s, v6.4s, v9.4s\n" 390 "mov v29.16b, v17.16b\n fmla v29.4s, v2.4s, v9.4s\n" 391 "mov v28.16b, v17.16b\n fmla v28.4s, v0.4s, v9.4s\n" 392 "fmla v31.4s, v0.4s, v10.4s\n" 393 "fmla v30.4s, v1.4s, v12.4s\n" 394 "fmla v31.4s, v1.4s, v11.4s\n" 395 "fmla v30.4s, v2.4s, v13.4s\n" 396 "fmla v31.4s, v3.4s, v14.4s\n" 397 "fmla v30.4s, v0.4s, v16.4s\n" 398 "fmla v31.4s, v4.4s, v15.4s\n" 399 "fmla v31.4s, v2.4s, v16.4s\n" 400 "tbz %x[n_channels], #1, 6f\n" 401 "ld1 { v11.d }[0], [x26], #0x8\n" 402 "tbz %x[n_channels], #0, 7f\n" 403 "ld1 { v11.s }[2], [x26], #0x4\n" 406 "ld1 { v11.s }[0], [x26], #0x4\n" 408 "fmla v30.4s, v4.4s, v11.4s\n" 409 "ldr x25, [x16, #0x48]\n" 410 "add x25, x25, x14\n" 411 "tbz %x[n_channels], #1, 8f\n" 412 "ld1 { v12.d }[0], [x25], #0x8\n" 413 "tbz %x[n_channels], #0, 9f\n" 414 "ld1 { v12.s }[2], [x25], #0x4\n" 417 "ld1 { v12.s }[0], [x25], #0x4\n" 419 "fmla v30.4s, v5.4s, v12.4s\n" 420 "ldr x24, [x16, #0x50]\n" 421 "add x24, x24, x14\n" 422 "tbz %x[n_channels], #1, 10f\n" 423 "ld1 { v13.d }[0], [x24], #0x8\n" 424 "tbz %x[n_channels], #0, 11f\n" 425 "ld1 { v13.s }[2], [x24], #0x4\n" 428 "ld1 { v13.s }[0], [x24], #0x4\n" 430 "fmla v31.4s, v5.4s, v13.4s\n" 431 "ldr x23, [x16, #0x58]\n" 432 "fmla v30.4s, v3.4s, v13.4s\n" 433 "add x23, x23, x14\n" 434 "tbz %x[n_channels], #1, 12f\n" 435 "ld1 { v14.d }[0], [x23], #0x8\n" 436 "tbz %x[n_channels], #0, 13f\n" 437 "ld1 { v14.s }[2], [x23], #0x4\n" 440 "ld1 { v14.s }[0], [x23], #0x4\n" 442 "fmla v29.4s, v3.4s, v14.4s\n" 443 "ldr x22, [x16, #0x60]\n" 444 "add x22, x22, x14\n" 445 "tbz %x[n_channels], #1, 14f\n" 446 "ld1 { v15.d }[0], [x22], #0x8\n" 447 "tbz %x[n_channels], #0, 15f\n" 448 "ld1 { v15.s }[2], [x22], #0x4\n" 451 "ld1 { v15.s }[0], [x22], #0x4\n" 453 "fmla v31.4s, v6.4s, v15.4s\n" 454 "ldr x21, [x16, #0x68]\n" 455 "fmla v29.4s, v0.4s, v15.4s\n" 456 "add x21, x21, x14\n" 457 "tbz %x[n_channels], #1, 16f\n" 458 "ld1 { v11.d }[0], [x21], #0x8\n" 459 "tbz %x[n_channels], #0, 17f\n" 460 "ld1 { v11.s }[2], [x21], #0x4\n" 463 "ld1 { v11.s }[0], [x21], #0x4\n" 465 "fmla v29.4s, v4.4s, v11.4s\n" 466 "ldr x20, [x16, #0x70]\n" 467 "add x20, x20, x14\n" 468 "tbz %x[n_channels], #1, 18f\n" 469 "ld1 { v16.d }[0], [x20], #0x8\n" 470 "tbz %x[n_channels], #0, 19f\n" 471 "ld1 { v16.s }[2], [x20], #0x4\n" 474 "ld1 { v16.s }[0], [x20], #0x4\n" 476 "fmla v31.4s, v7.4s, v16.4s\n" 477 "ldr x19, [x16, #0x78]\n" 478 "fmla v29.4s, v1.4s, v16.4s\n" 479 "add x19, x19, x14\n" 480 "tbz %x[n_channels], #1, 20f\n" 481 "ld1 { v13.d }[0], [x19], #0x8\n" 482 "tbz %x[n_channels], #0, 21f\n" 483 "ld1 { v13.s }[2], [x19], #0x4\n" 486 "ld1 { v13.s }[0], [x19], #0x4\n" 488 "fmla v28.4s, v4.4s, v13.4s\n" 489 "ldr x26, [x16, #0x80]\n" 490 "add x26, x26, x14\n" 491 "tbz %x[n_channels], #1, 22f\n" 492 "ld1 { v12.d }[0], [x26], #0x8\n" 493 "tbz %x[n_channels], #0, 23f\n" 494 "ld1 { v12.s }[2], [x26], #0x4\n" 497 "ld1 { v12.s }[0], [x26], #0x4\n" 499 "fmla v30.4s, v7.4s, v12.4s\n" 500 "ldr x25, [x16, #0x88]\n" 501 "fmla v28.4s, v1.4s, v12.4s\n" 502 "add x25, x25, x14\n" 503 "tbz %x[n_channels], #1, 24f\n" 504 "ld1 { v14.d }[0], [x25], #0x8\n" 505 "tbz %x[n_channels], #0, 25f\n" 506 "ld1 { v14.s }[2], [x25], #0x4\n" 509 "ld1 { v14.s }[0], [x25], #0x4\n" 511 "fmla v28.4s, v5.4s, v14.4s\n" 512 "ldr x24, [x16, #0x90]\n" 513 "add x24, x24, x14\n" 514 "tbz %x[n_channels], #1, 26f\n" 515 "ld1 { v15.d }[0], [x24], #0x8\n" 516 "tbz %x[n_channels], #0, 27f\n" 517 "ld1 { v15.s }[2], [x24], #0x4\n" 520 "ld1 { v15.s }[0], [x24], #0x4\n" 522 "fmla v29.4s, v6.4s, v15.4s\n" 523 "ldr x23, [x16, #0x98]\n" 524 "add x23, x23, x14\n" 525 "tbz %x[n_channels], #1, 28f\n" 526 "ld1 { v11.d }[0], [x23], #0x8\n" 527 "tbz %x[n_channels], #0, 29f\n" 528 "ld1 { v11.s }[2], [x23], #0x4\n" 531 "ld1 { v11.s }[0], [x23], #0x4\n" 533 "fmla v30.4s, v8.4s, v11.4s\n" 534 "ldr x22, [x16, #0xa0]\n" 535 "fmla v28.4s, v2.4s, v11.4s\n" 536 "add x22, x22, x14\n" 537 "tbz %x[n_channels], #1, 30f\n" 538 "ld1 { v13.d }[0], [x22], #0x8\n" 539 "tbz %x[n_channels], #0, 31f\n" 540 "ld1 { v13.s }[2], [x22], #0x4\n" 543 "ld1 { v13.s }[0], [x22], #0x4\n" 545 "fmla v29.4s, v7.4s, v13.4s\n" 546 "ldr x21, [x16, #0xa8]\n" 547 "add x21, x21, x14\n" 548 "tbz %x[n_channels], #1, 32f\n" 549 "ld1 { v16.d }[0], [x21], #0x8\n" 550 "tbz %x[n_channels], #0, 33f\n" 551 "ld1 { v16.s }[2], [x21], #0x4\n" 554 "ld1 { v16.s }[0], [x21], #0x4\n" 556 "fmla v29.4s, v5.4s, v16.4s\n" 557 "ldr x20, [x16, #0xb0]\n" 558 "fmla v28.4s, v3.4s, v16.4s\n" 559 "add x20, x20, x14\n" 560 "tbz %x[n_channels], #1, 34f\n" 561 "ld1 { v14.d }[0], [x20], #0x8\n" 562 "tbz %x[n_channels], #0, 35f\n" 563 "ld1 { v14.s }[2], [x20], #0x4\n" 566 "ld1 { v14.s }[0], [x20], #0x4\n" 568 "fmla v28.4s, v7.4s, v14.4s\n" 569 "ldr x19, [x16, #0xb8]\n" 570 "add x19, x19, x14\n" 571 "tbz %x[n_channels], #1, 36f\n" 572 "ld1 { v15.d }[0], [x19], #0x8\n" 573 "tbz %x[n_channels], #0, 37f\n" 574 "ld1 { v15.s }[2], [x19], #0x4\n" 577 "ld1 { v15.s }[0], [x19], #0x4\n" 579 "fmla v29.4s, v8.4s, v15.4s\n" 580 "ldr x26, [x16, #0xc0]\n" 581 "fmla v28.4s, v6.4s, v15.4s\n" 582 "add x26, x26, x14\n" 583 "tbz %x[n_channels], #1, 38f\n" 584 "ld1 { v11.d }[0], [x26], #0x8\n" 585 "tbz %x[n_channels], #0, 39f\n" 586 "ld1 { v11.s }[2], [x26], #0x4\n" 589 "ld1 { v11.s }[0], [x26], #0x4\n" 591 "fmla v28.4s, v8.4s, v11.4s\n" 592 "fmax v31.4s, v31.4s, v19.4s\n" 593 "fmax v30.4s, v30.4s, v19.4s\n" 594 "fmax v29.4s, v29.4s, v19.4s\n" 595 "fmin v31.4s, v31.4s, v18.4s\n" 596 "fmin v30.4s, v30.4s, v18.4s\n" 597 "fmin v29.4s, v29.4s, v18.4s\n" 598 "fmax v28.4s, v28.4s, v19.4s\n" 599 "fmin v28.4s, v28.4s, v18.4s\n" 600 "tbz %x[n_channels], #1, 40f\n" 601 "st1 { v31.d }[0], [x13], #0x8\n" 602 "st1 { v30.d }[0], [x12], #0x8\n" 603 "st1 { v29.d }[0], [x10], #0x8\n" 604 "st1 { v28.d }[0], [x9], #0x8\n" 605 "tbz %x[n_channels], #0, 41f\n" 606 "st1 { v31.s }[2], [x13], #0x4\n" 607 "st1 { v30.s }[2], [x12], #0x4\n" 608 "st1 { v29.s }[2], [x10], #0x4\n" 609 "st1 { v28.s }[2], [x9], #0x4\n" 612 "st1 { v31.s }[0], [x13], #0x4\n" 613 "st1 { v30.s }[0], [x12], #0x4\n" 614 "st1 { v29.s }[0], [x10], #0x4\n" 615 "st1 { v28.s }[0], [x9], #0x4\n" 621 : [n_channels]
"r" ((
unsigned long) n_channels), [offsetof_Args_inptrs]
"I" (offsetof(Args, inptrs)), [offsetof_args_max]
"I" (offsetof(Args, max)), [offsetof_args_min]
"I" (offsetof(Args, min)), [offsetof_args_outptrs]
"I" (offsetof(Args, outptrs)), [offsetof_args_params]
"I" (offsetof(Args, params)), [params_struct]
"r" (¶ms_struct)
622 :
"cc",
"memory",
"v0",
"v1",
"v2",
"v3",
"v4",
"v5",
"v6",
"v7",
"v8",
"v9",
"v10",
"v11",
"v12",
"v13",
"v14",
"v15",
"v16",
"v17",
"v18",
"v19",
"v28",
"v29",
"v30",
"v31",
"x9",
"x10",
"x11",
"x12",
"x13",
"x14",
"x15",
"x16",
"x19",
"x20",
"x21",
"x22",
"x23",
"x24",
"x25",
"x26",
"x27",
"x28"
void a64_fp32_nhwc_3x3_s2_output2x2_mla_depthfirst_indirect_impl(const float *const *const input_ptrs, float *const *const outptrs, const void *params, unsigned int n_channels, const float activation_min, const float activation_max)
template UniqueDepthwiseCommon< float > depthwise(const DepthwiseArgs &, const Nothing &)