32 const float *
const *
const inptrs,
33 float *
const *
const outptrs,
35 const unsigned int n_output_channels,
36 const float activation_min,
37 const float activation_max
40 const float minmax_vals[2] = { activation_min, activation_max };
43 "ldp x14, x13, [%x[outptrs], #0x0]\n" 44 "add x12, %x[clamps], #0x4\n" 45 "ldp x11, x10, [%x[outptrs], #0x10]\n" 47 "ldp x28, x27, [%x[outptrs], #0x20]\n" 49 "ldp x25, x24, [%x[outptrs], #0x30]\n" 50 "lsr x23, %x[channel_multiplier], #0x2\n" 51 "ldr x22, [%x[outptrs], #0x40]\n" 52 "ldr x21, [%x[inptrs], #0x0]\n" 53 "ldr x20, [%x[inptrs], #0x8]\n" 54 "ldr x19, [%x[inptrs], #0x10]\n" 55 "ldr q0, [x21, #0x0]\n" 56 "ldr q1, [x21, #0x10]\n" 57 "ldr q2, [x20, #0x0]\n" 58 "ldr q3, [x20, #0x10]\n" 59 "ldr q4, [x19, #0x0]\n" 60 "ldr q5, [x19, #0x10]\n" 61 "ldr x21, [%x[inptrs], #0x18]\n" 62 "ldr x20, [%x[inptrs], #0x20]\n" 63 "ldr x19, [%x[inptrs], #0x28]\n" 64 "ldr q6, [x21, #0x0]\n" 65 "ldr q7, [x21, #0x10]\n" 66 "ldr q8, [x20, #0x0]\n" 67 "ldr q9, [x20, #0x10]\n" 68 "ldr q10, [x19, #0x0]\n" 69 "ldr q11, [x19, #0x10]\n" 70 "ldr x19, [%x[inptrs], #0x30]\n" 71 "ld1r { v24.4s }, [%x[clamps]]\n" 72 "ld1r { v23.4s }, [x12]\n" 73 "ldr q12, [x19, #0x0]\n" 74 "ldr q13, [x19, #0x10]\n" 76 "ldr q14, [%x[params], #0x0]\n" 77 "mov v15.16b, v14.16b\n" 78 "ldr q31, [%x[params], #0x10]\n" 79 "subs x23, x23, #0x1\n" 80 "mov v16.16b, v14.16b\n" 81 "ldr q30, [%x[params], #0x20]\n" 82 "mov v17.16b, v14.16b\n" 83 "ldr q29, [%x[params], #0x30]\n" 84 "add %x[params], %x[params], #0x40\n" 85 "mov v18.16b, v14.16b\n" 86 "mov v19.16b, v14.16b\n" 87 "mov v20.16b, v14.16b\n" 88 "mov v21.16b, v14.16b\n" 89 "mov v22.16b, v14.16b\n" 92 "fmla v14.4s, v31.4s, v0.s[0]\n" 94 "fmla v15.4s, v31.4s, v0.s[2]\n" 95 "subs x23, x23, #0x1\n" 96 "fmla v16.4s, v31.4s, v1.s[0]\n" 97 "fmla v17.4s, v31.4s, v4.s[0]\n" 98 "fmla v18.4s, v31.4s, v4.s[2]\n" 99 "fmla v19.4s, v31.4s, v5.s[0]\n" 100 "fmla v20.4s, v31.4s, v8.s[0]\n" 101 "fmla v21.4s, v31.4s, v8.s[2]\n" 102 "fmla v22.4s, v31.4s, v9.s[0]\n" 103 "ldr q31, [%x[params], #0x0]\n" 104 "fmla v14.4s, v30.4s, v0.s[1]\n" 105 "fmla v15.4s, v30.4s, v0.s[3]\n" 106 "fmla v16.4s, v30.4s, v1.s[1]\n" 107 "fmla v17.4s, v30.4s, v4.s[1]\n" 108 "fmla v18.4s, v30.4s, v4.s[3]\n" 109 "fmla v19.4s, v30.4s, v5.s[1]\n" 110 "fmla v20.4s, v30.4s, v8.s[1]\n" 111 "fmla v21.4s, v30.4s, v8.s[3]\n" 112 "fmla v22.4s, v30.4s, v9.s[1]\n" 113 "ldr q30, [%x[params], #0x10]\n" 114 "fmla v14.4s, v29.4s, v0.s[2]\n" 115 "fmla v15.4s, v29.4s, v1.s[0]\n" 116 "fmla v16.4s, v29.4s, v1.s[2]\n" 117 "fmla v17.4s, v29.4s, v4.s[2]\n" 118 "fmla v18.4s, v29.4s, v5.s[0]\n" 119 "fmla v19.4s, v29.4s, v5.s[2]\n" 120 "fmla v20.4s, v29.4s, v8.s[2]\n" 121 "fmla v21.4s, v29.4s, v9.s[0]\n" 122 "fmla v22.4s, v29.4s, v9.s[2]\n" 123 "ldr q29, [%x[params], #0x20]\n" 124 "fmla v14.4s, v31.4s, v2.s[0]\n" 125 "fmla v15.4s, v31.4s, v2.s[2]\n" 126 "fmla v16.4s, v31.4s, v3.s[0]\n" 127 "fmla v17.4s, v31.4s, v6.s[0]\n" 128 "fmla v18.4s, v31.4s, v6.s[2]\n" 129 "fmla v19.4s, v31.4s, v7.s[0]\n" 130 "fmla v20.4s, v31.4s, v10.s[0]\n" 131 "fmla v21.4s, v31.4s, v10.s[2]\n" 132 "fmla v22.4s, v31.4s, v11.s[0]\n" 133 "ldr q31, [%x[params], #0x30]\n" 134 "fmla v14.4s, v30.4s, v2.s[1]\n" 135 "fmla v15.4s, v30.4s, v2.s[3]\n" 136 "fmla v16.4s, v30.4s, v3.s[1]\n" 137 "fmla v17.4s, v30.4s, v6.s[1]\n" 138 "fmla v18.4s, v30.4s, v6.s[3]\n" 139 "fmla v19.4s, v30.4s, v7.s[1]\n" 140 "fmla v20.4s, v30.4s, v10.s[1]\n" 141 "fmla v21.4s, v30.4s, v10.s[3]\n" 142 "fmla v22.4s, v30.4s, v11.s[1]\n" 143 "ldr q30, [%x[params], #0x40]\n" 144 "fmla v14.4s, v29.4s, v2.s[2]\n" 145 "fmla v15.4s, v29.4s, v3.s[0]\n" 146 "fmla v16.4s, v29.4s, v3.s[2]\n" 147 "fmla v17.4s, v29.4s, v6.s[2]\n" 148 "fmla v18.4s, v29.4s, v7.s[0]\n" 149 "fmla v19.4s, v29.4s, v7.s[2]\n" 150 "fmla v20.4s, v29.4s, v10.s[2]\n" 151 "fmla v21.4s, v29.4s, v11.s[0]\n" 152 "fmla v22.4s, v29.4s, v11.s[2]\n" 153 "ldr q29, [%x[params], #0x50]\n" 154 "fmla v14.4s, v31.4s, v4.s[0]\n" 155 "fmla v15.4s, v31.4s, v4.s[2]\n" 156 "fmla v16.4s, v31.4s, v5.s[0]\n" 157 "fmla v17.4s, v31.4s, v8.s[0]\n" 158 "fmla v18.4s, v31.4s, v8.s[2]\n" 159 "fmla v19.4s, v31.4s, v9.s[0]\n" 160 "fmla v20.4s, v31.4s, v12.s[0]\n" 161 "fmla v21.4s, v31.4s, v12.s[2]\n" 162 "fmla v22.4s, v31.4s, v13.s[0]\n" 163 "ldr q31, [%x[params], #0x70]\n" 164 "fmla v14.4s, v30.4s, v4.s[1]\n" 165 "fmla v15.4s, v30.4s, v4.s[3]\n" 166 "fmla v16.4s, v30.4s, v5.s[1]\n" 167 "fmla v17.4s, v30.4s, v8.s[1]\n" 168 "fmla v18.4s, v30.4s, v8.s[3]\n" 169 "fmla v19.4s, v30.4s, v9.s[1]\n" 170 "fmla v20.4s, v30.4s, v12.s[1]\n" 171 "fmla v21.4s, v30.4s, v12.s[3]\n" 172 "fmla v22.4s, v30.4s, v13.s[1]\n" 173 "ldr q30, [%x[params], #0x80]\n" 174 "fmla v14.4s, v29.4s, v4.s[2]\n" 175 "fmla v15.4s, v29.4s, v5.s[0]\n" 176 "fmla v16.4s, v29.4s, v5.s[2]\n" 177 "fmla v17.4s, v29.4s, v8.s[2]\n" 178 "fmla v18.4s, v29.4s, v9.s[0]\n" 179 "fmla v19.4s, v29.4s, v9.s[2]\n" 180 "fmla v20.4s, v29.4s, v12.s[2]\n" 181 "fmla v21.4s, v29.4s, v13.s[0]\n" 182 "fmla v22.4s, v29.4s, v13.s[2]\n" 183 "ldr q29, [%x[params], #0x90]\n" 184 "fmin v14.4s, v14.4s, v23.4s\n" 185 "fmin v15.4s, v15.4s, v23.4s\n" 186 "fmin v16.4s, v16.4s, v23.4s\n" 187 "fmax v14.4s, v14.4s, v24.4s\n" 188 "str q14, [x14, x26]\n" 189 "fmax v15.4s, v15.4s, v24.4s\n" 190 "fmax v16.4s, v16.4s, v24.4s\n" 191 "ldr q14, [%x[params], #0x60]\n" 192 "add %x[params], %x[params], #0xa0\n" 193 "fmin v17.4s, v17.4s, v23.4s\n" 194 "str q15, [x13, x26]\n" 195 "fmin v18.4s, v18.4s, v23.4s\n" 196 "fmin v19.4s, v19.4s, v23.4s\n" 197 "str q16, [x11, x26]\n" 198 "fmin v20.4s, v20.4s, v23.4s\n" 199 "fmax v17.4s, v17.4s, v24.4s\n" 200 "str q17, [x10, x26]\n" 201 "fmax v18.4s, v18.4s, v24.4s\n" 202 "fmax v19.4s, v19.4s, v24.4s\n" 203 "str q18, [x28, x26]\n" 204 "fmax v20.4s, v20.4s, v24.4s\n" 205 "fmin v21.4s, v21.4s, v23.4s\n" 206 "str q19, [x27, x26]\n" 207 "fmin v22.4s, v22.4s, v23.4s\n" 208 "str q20, [x25, x26]\n" 209 "fmax v21.4s, v21.4s, v24.4s\n" 210 "mov v15.16b, v14.16b\n" 211 "str q21, [x24, x26]\n" 212 "fmax v22.4s, v22.4s, v24.4s\n" 213 "mov v16.16b, v14.16b\n" 214 "str q22, [x22, x26]\n" 215 "mov v17.16b, v14.16b\n" 216 "add x26, x26, #0x10\n" 217 "mov v18.16b, v14.16b\n" 218 "mov v19.16b, v14.16b\n" 219 "mov v20.16b, v14.16b\n" 220 "mov v21.16b, v14.16b\n" 221 "mov v22.16b, v14.16b\n" 224 "fmla v14.4s, v31.4s, v0.s[0]\n" 225 "fmla v15.4s, v31.4s, v0.s[2]\n" 226 "fmla v16.4s, v31.4s, v1.s[0]\n" 227 "fmla v17.4s, v31.4s, v4.s[0]\n" 228 "fmla v18.4s, v31.4s, v4.s[2]\n" 229 "fmla v19.4s, v31.4s, v5.s[0]\n" 230 "fmla v20.4s, v31.4s, v8.s[0]\n" 231 "fmla v21.4s, v31.4s, v8.s[2]\n" 232 "fmla v22.4s, v31.4s, v9.s[0]\n" 233 "ldr q31, [%x[params], #0x0]\n" 234 "fmla v14.4s, v30.4s, v0.s[1]\n" 235 "fmla v15.4s, v30.4s, v0.s[3]\n" 236 "fmla v16.4s, v30.4s, v1.s[1]\n" 237 "fmla v17.4s, v30.4s, v4.s[1]\n" 238 "fmla v18.4s, v30.4s, v4.s[3]\n" 239 "fmla v19.4s, v30.4s, v5.s[1]\n" 240 "fmla v20.4s, v30.4s, v8.s[1]\n" 241 "fmla v21.4s, v30.4s, v8.s[3]\n" 242 "fmla v22.4s, v30.4s, v9.s[1]\n" 243 "ldr q30, [%x[params], #0x10]\n" 244 "fmla v14.4s, v29.4s, v0.s[2]\n" 245 "fmla v15.4s, v29.4s, v1.s[0]\n" 246 "fmla v16.4s, v29.4s, v1.s[2]\n" 247 "fmla v17.4s, v29.4s, v4.s[2]\n" 248 "fmla v18.4s, v29.4s, v5.s[0]\n" 249 "fmla v19.4s, v29.4s, v5.s[2]\n" 250 "fmla v20.4s, v29.4s, v8.s[2]\n" 251 "fmla v21.4s, v29.4s, v9.s[0]\n" 252 "fmla v22.4s, v29.4s, v9.s[2]\n" 253 "ldr q29, [%x[params], #0x20]\n" 254 "fmla v14.4s, v31.4s, v2.s[0]\n" 255 "fmla v15.4s, v31.4s, v2.s[2]\n" 256 "fmla v16.4s, v31.4s, v3.s[0]\n" 257 "fmla v17.4s, v31.4s, v6.s[0]\n" 258 "fmla v18.4s, v31.4s, v6.s[2]\n" 259 "fmla v19.4s, v31.4s, v7.s[0]\n" 260 "fmla v20.4s, v31.4s, v10.s[0]\n" 261 "fmla v21.4s, v31.4s, v10.s[2]\n" 262 "fmla v22.4s, v31.4s, v11.s[0]\n" 263 "ldr q31, [%x[params], #0x30]\n" 264 "fmla v14.4s, v30.4s, v2.s[1]\n" 265 "fmla v15.4s, v30.4s, v2.s[3]\n" 266 "fmla v16.4s, v30.4s, v3.s[1]\n" 267 "fmla v17.4s, v30.4s, v6.s[1]\n" 268 "fmla v18.4s, v30.4s, v6.s[3]\n" 269 "fmla v19.4s, v30.4s, v7.s[1]\n" 270 "fmla v20.4s, v30.4s, v10.s[1]\n" 271 "fmla v21.4s, v30.4s, v10.s[3]\n" 272 "fmla v22.4s, v30.4s, v11.s[1]\n" 273 "ldr q30, [%x[params], #0x40]\n" 274 "fmla v14.4s, v29.4s, v2.s[2]\n" 275 "fmla v15.4s, v29.4s, v3.s[0]\n" 276 "fmla v16.4s, v29.4s, v3.s[2]\n" 277 "fmla v17.4s, v29.4s, v6.s[2]\n" 278 "fmla v18.4s, v29.4s, v7.s[0]\n" 279 "fmla v19.4s, v29.4s, v7.s[2]\n" 280 "fmla v20.4s, v29.4s, v10.s[2]\n" 281 "fmla v21.4s, v29.4s, v11.s[0]\n" 282 "fmla v22.4s, v29.4s, v11.s[2]\n" 283 "ldr q29, [%x[params], #0x50]\n" 284 "add %x[params], %x[params], #0x60\n" 285 "fmla v14.4s, v31.4s, v4.s[0]\n" 286 "fmla v15.4s, v31.4s, v4.s[2]\n" 287 "fmla v16.4s, v31.4s, v5.s[0]\n" 288 "fmla v17.4s, v31.4s, v8.s[0]\n" 289 "fmla v18.4s, v31.4s, v8.s[2]\n" 290 "fmla v19.4s, v31.4s, v9.s[0]\n" 291 "fmla v20.4s, v31.4s, v12.s[0]\n" 292 "fmla v21.4s, v31.4s, v12.s[2]\n" 293 "fmla v22.4s, v31.4s, v13.s[0]\n" 294 "fmla v14.4s, v30.4s, v4.s[1]\n" 295 "fmla v15.4s, v30.4s, v4.s[3]\n" 296 "fmla v16.4s, v30.4s, v5.s[1]\n" 297 "fmla v17.4s, v30.4s, v8.s[1]\n" 298 "fmla v18.4s, v30.4s, v8.s[3]\n" 299 "fmla v19.4s, v30.4s, v9.s[1]\n" 300 "fmla v20.4s, v30.4s, v12.s[1]\n" 301 "fmla v21.4s, v30.4s, v12.s[3]\n" 302 "fmla v22.4s, v30.4s, v13.s[1]\n" 303 "fmla v14.4s, v29.4s, v4.s[2]\n" 304 "fmla v15.4s, v29.4s, v5.s[0]\n" 305 "fmla v16.4s, v29.4s, v5.s[2]\n" 306 "fmla v17.4s, v29.4s, v8.s[2]\n" 307 "fmla v18.4s, v29.4s, v9.s[0]\n" 308 "fmla v19.4s, v29.4s, v9.s[2]\n" 309 "fmla v20.4s, v29.4s, v12.s[2]\n" 310 "fmla v21.4s, v29.4s, v13.s[0]\n" 311 "fmla v22.4s, v29.4s, v13.s[2]\n" 312 "fmin v14.4s, v14.4s, v23.4s\n" 313 "fmin v15.4s, v15.4s, v23.4s\n" 314 "fmin v16.4s, v16.4s, v23.4s\n" 315 "fmax v14.4s, v14.4s, v24.4s\n" 316 "str q14, [x14, x26]\n" 317 "fmax v15.4s, v15.4s, v24.4s\n" 318 "fmax v16.4s, v16.4s, v24.4s\n" 319 "str q15, [x13, x26]\n" 320 "fmin v17.4s, v17.4s, v23.4s\n" 321 "fmin v18.4s, v18.4s, v23.4s\n" 322 "str q16, [x11, x26]\n" 323 "fmin v19.4s, v19.4s, v23.4s\n" 324 "fmin v20.4s, v20.4s, v23.4s\n" 325 "fmax v17.4s, v17.4s, v24.4s\n" 326 "str q17, [x10, x26]\n" 327 "fmax v18.4s, v18.4s, v24.4s\n" 328 "fmax v19.4s, v19.4s, v24.4s\n" 329 "str q18, [x28, x26]\n" 330 "fmax v20.4s, v20.4s, v24.4s\n" 331 "fmin v21.4s, v21.4s, v23.4s\n" 332 "str q19, [x27, x26]\n" 333 "fmin v22.4s, v22.4s, v23.4s\n" 334 "str q20, [x25, x26]\n" 335 "fmax v21.4s, v21.4s, v24.4s\n" 336 "fmax v22.4s, v22.4s, v24.4s\n" 337 "str q21, [x24, x26]\n" 338 "str q22, [x22, x26]\n" 339 "add x26, x26, #0x10\n" 341 "tst %x[channel_multiplier], #0x3\n" 343 "ldr q14, [%x[params], #0x0]\n" 344 "mov v15.16b, v14.16b\n" 345 "ldr q31, [%x[params], #0x10]\n" 346 "mov v16.16b, v14.16b\n" 347 "ldr q30, [%x[params], #0x20]\n" 348 "mov v17.16b, v14.16b\n" 349 "ldr q29, [%x[params], #0x30]\n" 350 "mov v18.16b, v14.16b\n" 351 "mov v19.16b, v14.16b\n" 352 "mov v20.16b, v14.16b\n" 353 "mov v21.16b, v14.16b\n" 354 "mov v22.16b, v14.16b\n" 355 "fmla v14.4s, v31.4s, v0.s[0]\n" 356 "fmla v15.4s, v31.4s, v0.s[2]\n" 357 "fmla v16.4s, v31.4s, v1.s[0]\n" 358 "fmla v17.4s, v31.4s, v4.s[0]\n" 359 "fmla v18.4s, v31.4s, v4.s[2]\n" 360 "fmla v19.4s, v31.4s, v5.s[0]\n" 361 "fmla v20.4s, v31.4s, v8.s[0]\n" 362 "fmla v21.4s, v31.4s, v8.s[2]\n" 363 "fmla v22.4s, v31.4s, v9.s[0]\n" 364 "ldr q31, [%x[params], #0x40]\n" 365 "fmla v14.4s, v30.4s, v0.s[1]\n" 366 "fmla v15.4s, v30.4s, v0.s[3]\n" 367 "fmla v16.4s, v30.4s, v1.s[1]\n" 368 "fmla v17.4s, v30.4s, v4.s[1]\n" 369 "fmla v18.4s, v30.4s, v4.s[3]\n" 370 "fmla v19.4s, v30.4s, v5.s[1]\n" 371 "fmla v20.4s, v30.4s, v8.s[1]\n" 372 "fmla v21.4s, v30.4s, v8.s[3]\n" 373 "fmla v22.4s, v30.4s, v9.s[1]\n" 374 "ldr q30, [%x[params], #0x50]\n" 375 "fmla v14.4s, v29.4s, v0.s[2]\n" 376 "fmla v15.4s, v29.4s, v1.s[0]\n" 377 "fmla v16.4s, v29.4s, v1.s[2]\n" 378 "fmla v17.4s, v29.4s, v4.s[2]\n" 379 "fmla v18.4s, v29.4s, v5.s[0]\n" 380 "fmla v19.4s, v29.4s, v5.s[2]\n" 381 "fmla v20.4s, v29.4s, v8.s[2]\n" 382 "fmla v21.4s, v29.4s, v9.s[0]\n" 383 "fmla v22.4s, v29.4s, v9.s[2]\n" 384 "ldr q29, [%x[params], #0x60]\n" 385 "fmla v14.4s, v31.4s, v2.s[0]\n" 386 "fmla v15.4s, v31.4s, v2.s[2]\n" 387 "fmla v16.4s, v31.4s, v3.s[0]\n" 388 "fmla v17.4s, v31.4s, v6.s[0]\n" 389 "fmla v18.4s, v31.4s, v6.s[2]\n" 390 "fmla v19.4s, v31.4s, v7.s[0]\n" 391 "fmla v20.4s, v31.4s, v10.s[0]\n" 392 "fmla v21.4s, v31.4s, v10.s[2]\n" 393 "fmla v22.4s, v31.4s, v11.s[0]\n" 394 "ldr q31, [%x[params], #0x70]\n" 395 "fmla v14.4s, v30.4s, v2.s[1]\n" 396 "fmla v15.4s, v30.4s, v2.s[3]\n" 397 "fmla v16.4s, v30.4s, v3.s[1]\n" 398 "fmla v17.4s, v30.4s, v6.s[1]\n" 399 "fmla v18.4s, v30.4s, v6.s[3]\n" 400 "fmla v19.4s, v30.4s, v7.s[1]\n" 401 "fmla v20.4s, v30.4s, v10.s[1]\n" 402 "fmla v21.4s, v30.4s, v10.s[3]\n" 403 "fmla v22.4s, v30.4s, v11.s[1]\n" 404 "ldr q30, [%x[params], #0x80]\n" 405 "fmla v14.4s, v29.4s, v2.s[2]\n" 406 "fmla v15.4s, v29.4s, v3.s[0]\n" 407 "fmla v16.4s, v29.4s, v3.s[2]\n" 408 "fmla v17.4s, v29.4s, v6.s[2]\n" 409 "fmla v18.4s, v29.4s, v7.s[0]\n" 410 "fmla v19.4s, v29.4s, v7.s[2]\n" 411 "fmla v20.4s, v29.4s, v10.s[2]\n" 412 "fmla v21.4s, v29.4s, v11.s[0]\n" 413 "fmla v22.4s, v29.4s, v11.s[2]\n" 414 "ldr q29, [%x[params], #0x90]\n" 415 "add %x[params], %x[params], #0xa0\n" 416 "fmla v14.4s, v31.4s, v4.s[0]\n" 417 "fmla v15.4s, v31.4s, v4.s[2]\n" 418 "fmla v16.4s, v31.4s, v5.s[0]\n" 419 "fmla v17.4s, v31.4s, v8.s[0]\n" 420 "fmla v18.4s, v31.4s, v8.s[2]\n" 421 "fmla v19.4s, v31.4s, v9.s[0]\n" 422 "fmla v20.4s, v31.4s, v12.s[0]\n" 423 "fmla v21.4s, v31.4s, v12.s[2]\n" 424 "fmla v22.4s, v31.4s, v13.s[0]\n" 425 "fmla v14.4s, v30.4s, v4.s[1]\n" 426 "fmla v15.4s, v30.4s, v4.s[3]\n" 427 "fmla v16.4s, v30.4s, v5.s[1]\n" 428 "fmla v17.4s, v30.4s, v8.s[1]\n" 429 "fmla v18.4s, v30.4s, v8.s[3]\n" 430 "fmla v19.4s, v30.4s, v9.s[1]\n" 431 "fmla v20.4s, v30.4s, v12.s[1]\n" 432 "fmla v21.4s, v30.4s, v12.s[3]\n" 433 "fmla v22.4s, v30.4s, v13.s[1]\n" 434 "fmla v14.4s, v29.4s, v4.s[2]\n" 435 "fmla v15.4s, v29.4s, v5.s[0]\n" 436 "fmla v16.4s, v29.4s, v5.s[2]\n" 437 "fmla v17.4s, v29.4s, v8.s[2]\n" 438 "fmla v18.4s, v29.4s, v9.s[0]\n" 439 "fmla v19.4s, v29.4s, v9.s[2]\n" 440 "fmla v20.4s, v29.4s, v12.s[2]\n" 441 "fmla v21.4s, v29.4s, v13.s[0]\n" 442 "fmla v22.4s, v29.4s, v13.s[2]\n" 443 "fmin v14.4s, v14.4s, v23.4s\n" 444 "fmin v15.4s, v15.4s, v23.4s\n" 445 "fmin v16.4s, v16.4s, v23.4s\n" 446 "fmax v14.4s, v14.4s, v24.4s\n" 447 "fmax v15.4s, v15.4s, v24.4s\n" 448 "fmax v16.4s, v16.4s, v24.4s\n" 449 "fmin v17.4s, v17.4s, v23.4s\n" 450 "fmin v18.4s, v18.4s, v23.4s\n" 451 "fmin v19.4s, v19.4s, v23.4s\n" 452 "fmax v17.4s, v17.4s, v24.4s\n" 453 "fmax v18.4s, v18.4s, v24.4s\n" 454 "fmax v19.4s, v19.4s, v24.4s\n" 455 "fmin v20.4s, v20.4s, v23.4s\n" 456 "fmin v21.4s, v21.4s, v23.4s\n" 457 "fmin v22.4s, v22.4s, v23.4s\n" 458 "fmax v20.4s, v20.4s, v24.4s\n" 459 "fmax v21.4s, v21.4s, v24.4s\n" 460 "fmax v22.4s, v22.4s, v24.4s\n" 461 "tbz %x[channel_multiplier], #1, 4f\n" 462 "add x19, x14, x26\n" 463 "st1 { v14.d }[0], [x19]\n" 464 "add x19, x13, x26\n" 465 "st1 { v15.d }[0], [x19]\n" 466 "add x19, x11, x26\n" 467 "st1 { v16.d }[0], [x19]\n" 468 "add x19, x10, x26\n" 469 "st1 { v17.d }[0], [x19]\n" 470 "add x19, x28, x26\n" 471 "st1 { v18.d }[0], [x19]\n" 472 "add x19, x27, x26\n" 473 "st1 { v19.d }[0], [x19]\n" 474 "add x19, x25, x26\n" 475 "st1 { v20.d }[0], [x19]\n" 476 "add x19, x24, x26\n" 477 "st1 { v21.d }[0], [x19]\n" 478 "add x19, x22, x26\n" 479 "st1 { v22.d }[0], [x19]\n" 480 "add x26, x26, #0x8\n" 481 "tbz %x[channel_multiplier], #0, 5f\n" 482 "add x19, x14, x26\n" 483 "st1 { v14.s }[2], [x19]\n" 484 "add x19, x13, x26\n" 485 "st1 { v15.s }[2], [x19]\n" 486 "add x19, x11, x26\n" 487 "st1 { v16.s }[2], [x19]\n" 488 "add x19, x10, x26\n" 489 "st1 { v17.s }[2], [x19]\n" 490 "add x19, x28, x26\n" 491 "st1 { v18.s }[2], [x19]\n" 492 "add x19, x27, x26\n" 493 "st1 { v19.s }[2], [x19]\n" 494 "add x19, x25, x26\n" 495 "st1 { v20.s }[2], [x19]\n" 496 "add x19, x24, x26\n" 497 "st1 { v21.s }[2], [x19]\n" 498 "add x19, x22, x26\n" 499 "st1 { v22.s }[2], [x19]\n" 502 "tbz %x[channel_multiplier], #0, 5f\n" 503 "add x19, x14, x26\n" 504 "st1 { v14.s }[0], [x19]\n" 505 "add x19, x13, x26\n" 506 "st1 { v15.s }[0], [x19]\n" 507 "add x19, x11, x26\n" 508 "st1 { v16.s }[0], [x19]\n" 509 "add x19, x10, x26\n" 510 "st1 { v17.s }[0], [x19]\n" 511 "add x19, x28, x26\n" 512 "st1 { v18.s }[0], [x19]\n" 513 "add x19, x27, x26\n" 514 "st1 { v19.s }[0], [x19]\n" 515 "add x19, x25, x26\n" 516 "st1 { v20.s }[0], [x19]\n" 517 "add x19, x24, x26\n" 518 "st1 { v21.s }[0], [x19]\n" 519 "add x19, x22, x26\n" 520 "st1 { v22.s }[0], [x19]\n" 525 : [params]
"+&r" (params)
526 : [channel_multiplier]
"r" (n_output_channels), [clamps]
"r" (minmax_vals), [inptrs]
"r" (inptrs), [outptrs]
"r" (outptrs)
527 :
"cc",
"memory",
"v0",
"v1",
"v2",
"v3",
"v4",
"v5",
"v6",
"v7",
"v8",
"v9",
"v10",
"v11",
"v12",
"v13",
"v14",
"v15",
"v16",
"v17",
"v18",
"v19",
"v20",
"v21",
"v22",
"v23",
"v24",
"v29",
"v30",
"v31",
"x9",
"x10",
"x11",
"x12",
"x13",
"x14",
"x19",
"x20",
"x21",
"x22",
"x23",
"x24",
"x25",
"x26",
"x27",
"x28"
void a64_fp32_packed_to_nhwc_3x3_s2_with_multiplier_output3x3_mla_depthfirst_impl(const float *const *const inptrs, float *const *const outptrs, const void *params, const unsigned int n_output_channels, const float activation_min, const float activation_max)
template UniqueDepthwiseCommon< float > depthwise(const DepthwiseArgs &, const Nothing &)