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 x13, x12, [%x[outptrs], #0x0]\n" 44 "add x11, %x[clamps], #0x4\n" 45 "ldp x10, x9, [%x[outptrs], #0x10]\n" 47 "ldp x27, x26, [%x[outptrs], #0x20]\n" 49 "ldp x24, x23, [%x[outptrs], #0x30]\n" 50 "lsr x22, %x[channel_multiplier], #0x2\n" 51 "ldr x21, [%x[inptrs], #0x0]\n" 52 "ldr x20, [%x[inptrs], #0x8]\n" 53 "ldr x19, [%x[inptrs], #0x10]\n" 54 "ldr q0, [x21, #0x0]\n" 55 "ldr q1, [x21, #0x10]\n" 56 "ldr q2, [x20, #0x0]\n" 57 "ldr q3, [x20, #0x10]\n" 58 "ldr q4, [x19, #0x0]\n" 59 "ldr q5, [x19, #0x10]\n" 60 "ldr x21, [%x[inptrs], #0x18]\n" 61 "ldr x20, [%x[inptrs], #0x20]\n" 62 "ldr x19, [%x[inptrs], #0x28]\n" 63 "ldr q6, [x21, #0x0]\n" 64 "ldr q7, [x21, #0x10]\n" 65 "ldr q8, [x20, #0x0]\n" 66 "ldr q9, [x20, #0x10]\n" 67 "ldr q10, [x19, #0x0]\n" 68 "ldr q11, [x19, #0x10]\n" 69 "ld1r { v21.4s }, [%x[clamps]]\n" 70 "ld1r { v20.4s }, [x11]\n" 72 "ldr q12, [%x[params], #0x0]\n" 73 "mov v13.16b, v12.16b\n" 74 "ldr q31, [%x[params], #0x10]\n" 75 "subs x22, x22, #0x1\n" 76 "mov v14.16b, v12.16b\n" 77 "ldr q30, [%x[params], #0x20]\n" 78 "mov v15.16b, v12.16b\n" 79 "ldr q29, [%x[params], #0x30]\n" 80 "mov v16.16b, v12.16b\n" 81 "ldr q28, [%x[params], #0x40]\n" 82 "mov v17.16b, v12.16b\n" 83 "ldr q27, [%x[params], #0x50]\n" 84 "add %x[params], %x[params], #0x60\n" 85 "mov v18.16b, v12.16b\n" 86 "mov v19.16b, v12.16b\n" 89 "fmla v12.4s, v31.4s, v0.s[0]\n" 90 "add x28, x28, #0x4\n" 91 "fmla v13.4s, v31.4s, v0.s[1]\n" 92 "subs x22, x22, #0x1\n" 93 "fmla v14.4s, v31.4s, v0.s[2]\n" 94 "fmla v15.4s, v31.4s, v0.s[3]\n" 95 "fmla v16.4s, v31.4s, v2.s[0]\n" 96 "fmla v17.4s, v31.4s, v2.s[1]\n" 97 "fmla v18.4s, v31.4s, v2.s[2]\n" 98 "fmla v19.4s, v31.4s, v2.s[3]\n" 99 "ldr q31, [%x[params], #0x0]\n" 100 "fmla v12.4s, v30.4s, v0.s[1]\n" 101 "fmla v13.4s, v30.4s, v0.s[2]\n" 102 "fmla v14.4s, v30.4s, v0.s[3]\n" 103 "fmla v15.4s, v30.4s, v1.s[0]\n" 104 "fmla v16.4s, v30.4s, v2.s[1]\n" 105 "fmla v17.4s, v30.4s, v2.s[2]\n" 106 "fmla v18.4s, v30.4s, v2.s[3]\n" 107 "fmla v19.4s, v30.4s, v3.s[0]\n" 108 "ldr q30, [%x[params], #0x10]\n" 109 "fmla v12.4s, v29.4s, v0.s[2]\n" 110 "fmla v13.4s, v29.4s, v0.s[3]\n" 111 "fmla v14.4s, v29.4s, v1.s[0]\n" 112 "fmla v15.4s, v29.4s, v1.s[1]\n" 113 "fmla v16.4s, v29.4s, v2.s[2]\n" 114 "fmla v17.4s, v29.4s, v2.s[3]\n" 115 "fmla v18.4s, v29.4s, v3.s[0]\n" 116 "fmla v19.4s, v29.4s, v3.s[1]\n" 117 "ldr q29, [%x[params], #0x20]\n" 118 "fmla v12.4s, v28.4s, v0.s[3]\n" 119 "fmla v13.4s, v28.4s, v1.s[0]\n" 120 "fmla v14.4s, v28.4s, v1.s[1]\n" 121 "fmla v15.4s, v28.4s, v1.s[2]\n" 122 "fmla v16.4s, v28.4s, v2.s[3]\n" 123 "fmla v17.4s, v28.4s, v3.s[0]\n" 124 "fmla v18.4s, v28.4s, v3.s[1]\n" 125 "fmla v19.4s, v28.4s, v3.s[2]\n" 126 "ldr q28, [%x[params], #0x30]\n" 127 "fmla v12.4s, v27.4s, v1.s[0]\n" 128 "fmla v13.4s, v27.4s, v1.s[1]\n" 129 "fmla v14.4s, v27.4s, v1.s[2]\n" 130 "fmla v15.4s, v27.4s, v1.s[3]\n" 131 "fmla v16.4s, v27.4s, v3.s[0]\n" 132 "fmla v17.4s, v27.4s, v3.s[1]\n" 133 "fmla v18.4s, v27.4s, v3.s[2]\n" 134 "fmla v19.4s, v27.4s, v3.s[3]\n" 135 "ldr q27, [%x[params], #0x40]\n" 136 "fmla v12.4s, v31.4s, v2.s[0]\n" 137 "fmla v13.4s, v31.4s, v2.s[1]\n" 138 "fmla v14.4s, v31.4s, v2.s[2]\n" 139 "fmla v15.4s, v31.4s, v2.s[3]\n" 140 "fmla v16.4s, v31.4s, v4.s[0]\n" 141 "fmla v17.4s, v31.4s, v4.s[1]\n" 142 "fmla v18.4s, v31.4s, v4.s[2]\n" 143 "fmla v19.4s, v31.4s, v4.s[3]\n" 144 "ldr q31, [%x[params], #0x50]\n" 145 "fmla v12.4s, v30.4s, v2.s[1]\n" 146 "fmla v13.4s, v30.4s, v2.s[2]\n" 147 "fmla v14.4s, v30.4s, v2.s[3]\n" 148 "fmla v15.4s, v30.4s, v3.s[0]\n" 149 "fmla v16.4s, v30.4s, v4.s[1]\n" 150 "fmla v17.4s, v30.4s, v4.s[2]\n" 151 "fmla v18.4s, v30.4s, v4.s[3]\n" 152 "fmla v19.4s, v30.4s, v5.s[0]\n" 153 "ldr q30, [%x[params], #0x60]\n" 154 "fmla v12.4s, v29.4s, v2.s[2]\n" 155 "fmla v13.4s, v29.4s, v2.s[3]\n" 156 "fmla v14.4s, v29.4s, v3.s[0]\n" 157 "fmla v15.4s, v29.4s, v3.s[1]\n" 158 "fmla v16.4s, v29.4s, v4.s[2]\n" 159 "fmla v17.4s, v29.4s, v4.s[3]\n" 160 "fmla v18.4s, v29.4s, v5.s[0]\n" 161 "fmla v19.4s, v29.4s, v5.s[1]\n" 162 "ldr q29, [%x[params], #0x70]\n" 163 "fmla v12.4s, v28.4s, v2.s[3]\n" 164 "fmla v13.4s, v28.4s, v3.s[0]\n" 165 "fmla v14.4s, v28.4s, v3.s[1]\n" 166 "fmla v15.4s, v28.4s, v3.s[2]\n" 167 "fmla v16.4s, v28.4s, v4.s[3]\n" 168 "fmla v17.4s, v28.4s, v5.s[0]\n" 169 "fmla v18.4s, v28.4s, v5.s[1]\n" 170 "fmla v19.4s, v28.4s, v5.s[2]\n" 171 "ldr q28, [%x[params], #0x80]\n" 172 "fmla v12.4s, v27.4s, v3.s[0]\n" 173 "fmla v13.4s, v27.4s, v3.s[1]\n" 174 "fmla v14.4s, v27.4s, v3.s[2]\n" 175 "fmla v15.4s, v27.4s, v3.s[3]\n" 176 "fmla v16.4s, v27.4s, v5.s[0]\n" 177 "fmla v17.4s, v27.4s, v5.s[1]\n" 178 "fmla v18.4s, v27.4s, v5.s[2]\n" 179 "fmla v19.4s, v27.4s, v5.s[3]\n" 180 "ldr q27, [%x[params], #0x90]\n" 181 "fmla v12.4s, v31.4s, v4.s[0]\n" 182 "fmla v13.4s, v31.4s, v4.s[1]\n" 183 "fmla v14.4s, v31.4s, v4.s[2]\n" 184 "fmla v15.4s, v31.4s, v4.s[3]\n" 185 "fmla v16.4s, v31.4s, v6.s[0]\n" 186 "fmla v17.4s, v31.4s, v6.s[1]\n" 187 "fmla v18.4s, v31.4s, v6.s[2]\n" 188 "fmla v19.4s, v31.4s, v6.s[3]\n" 189 "ldr q31, [%x[params], #0xa0]\n" 190 "fmla v12.4s, v30.4s, v4.s[1]\n" 191 "fmla v13.4s, v30.4s, v4.s[2]\n" 192 "fmla v14.4s, v30.4s, v4.s[3]\n" 193 "fmla v15.4s, v30.4s, v5.s[0]\n" 194 "fmla v16.4s, v30.4s, v6.s[1]\n" 195 "fmla v17.4s, v30.4s, v6.s[2]\n" 196 "fmla v18.4s, v30.4s, v6.s[3]\n" 197 "fmla v19.4s, v30.4s, v7.s[0]\n" 198 "ldr q30, [%x[params], #0xb0]\n" 199 "fmla v12.4s, v29.4s, v4.s[2]\n" 200 "fmla v13.4s, v29.4s, v4.s[3]\n" 201 "fmla v14.4s, v29.4s, v5.s[0]\n" 202 "fmla v15.4s, v29.4s, v5.s[1]\n" 203 "fmla v16.4s, v29.4s, v6.s[2]\n" 204 "fmla v17.4s, v29.4s, v6.s[3]\n" 205 "fmla v18.4s, v29.4s, v7.s[0]\n" 206 "fmla v19.4s, v29.4s, v7.s[1]\n" 207 "ldr q29, [%x[params], #0xc0]\n" 208 "fmla v12.4s, v28.4s, v4.s[3]\n" 209 "fmla v13.4s, v28.4s, v5.s[0]\n" 210 "fmla v14.4s, v28.4s, v5.s[1]\n" 211 "fmla v15.4s, v28.4s, v5.s[2]\n" 212 "fmla v16.4s, v28.4s, v6.s[3]\n" 213 "fmla v17.4s, v28.4s, v7.s[0]\n" 214 "fmla v18.4s, v28.4s, v7.s[1]\n" 215 "fmla v19.4s, v28.4s, v7.s[2]\n" 216 "ldr q28, [%x[params], #0xd0]\n" 217 "fmla v12.4s, v27.4s, v5.s[0]\n" 218 "fmla v13.4s, v27.4s, v5.s[1]\n" 219 "fmla v14.4s, v27.4s, v5.s[2]\n" 220 "fmla v15.4s, v27.4s, v5.s[3]\n" 221 "fmla v16.4s, v27.4s, v7.s[0]\n" 222 "fmla v17.4s, v27.4s, v7.s[1]\n" 223 "fmla v18.4s, v27.4s, v7.s[2]\n" 224 "fmla v19.4s, v27.4s, v7.s[3]\n" 225 "ldr q27, [%x[params], #0xe0]\n" 226 "fmla v12.4s, v31.4s, v6.s[0]\n" 227 "fmla v13.4s, v31.4s, v6.s[1]\n" 228 "fmla v14.4s, v31.4s, v6.s[2]\n" 229 "fmla v15.4s, v31.4s, v6.s[3]\n" 230 "fmla v16.4s, v31.4s, v8.s[0]\n" 231 "fmla v17.4s, v31.4s, v8.s[1]\n" 232 "fmla v18.4s, v31.4s, v8.s[2]\n" 233 "fmla v19.4s, v31.4s, v8.s[3]\n" 234 "ldr q31, [%x[params], #0xf0]\n" 235 "fmla v12.4s, v30.4s, v6.s[1]\n" 236 "fmla v13.4s, v30.4s, v6.s[2]\n" 237 "fmla v14.4s, v30.4s, v6.s[3]\n" 238 "fmla v15.4s, v30.4s, v7.s[0]\n" 239 "fmla v16.4s, v30.4s, v8.s[1]\n" 240 "fmla v17.4s, v30.4s, v8.s[2]\n" 241 "fmla v18.4s, v30.4s, v8.s[3]\n" 242 "fmla v19.4s, v30.4s, v9.s[0]\n" 243 "ldr q30, [%x[params], #0x100]\n" 244 "fmla v12.4s, v29.4s, v6.s[2]\n" 245 "fmla v13.4s, v29.4s, v6.s[3]\n" 246 "fmla v14.4s, v29.4s, v7.s[0]\n" 247 "fmla v15.4s, v29.4s, v7.s[1]\n" 248 "fmla v16.4s, v29.4s, v8.s[2]\n" 249 "fmla v17.4s, v29.4s, v8.s[3]\n" 250 "fmla v18.4s, v29.4s, v9.s[0]\n" 251 "fmla v19.4s, v29.4s, v9.s[1]\n" 252 "ldr q29, [%x[params], #0x110]\n" 253 "fmla v12.4s, v28.4s, v6.s[3]\n" 254 "fmla v13.4s, v28.4s, v7.s[0]\n" 255 "fmla v14.4s, v28.4s, v7.s[1]\n" 256 "fmla v15.4s, v28.4s, v7.s[2]\n" 257 "fmla v16.4s, v28.4s, v8.s[3]\n" 258 "fmla v17.4s, v28.4s, v9.s[0]\n" 259 "fmla v18.4s, v28.4s, v9.s[1]\n" 260 "fmla v19.4s, v28.4s, v9.s[2]\n" 261 "ldr q28, [%x[params], #0x120]\n" 262 "fmla v12.4s, v27.4s, v7.s[0]\n" 263 "fmla v13.4s, v27.4s, v7.s[1]\n" 264 "fmla v14.4s, v27.4s, v7.s[2]\n" 265 "fmla v15.4s, v27.4s, v7.s[3]\n" 266 "fmla v16.4s, v27.4s, v9.s[0]\n" 267 "fmla v17.4s, v27.4s, v9.s[1]\n" 268 "fmla v18.4s, v27.4s, v9.s[2]\n" 269 "fmla v19.4s, v27.4s, v9.s[3]\n" 270 "ldr q27, [%x[params], #0x130]\n" 271 "fmla v12.4s, v31.4s, v8.s[0]\n" 272 "fmla v13.4s, v31.4s, v8.s[1]\n" 273 "fmla v14.4s, v31.4s, v8.s[2]\n" 274 "fmla v15.4s, v31.4s, v8.s[3]\n" 275 "fmla v16.4s, v31.4s, v10.s[0]\n" 276 "fmla v17.4s, v31.4s, v10.s[1]\n" 277 "fmla v18.4s, v31.4s, v10.s[2]\n" 278 "fmla v19.4s, v31.4s, v10.s[3]\n" 279 "ldr q31, [%x[params], #0x150]\n" 280 "fmla v12.4s, v30.4s, v8.s[1]\n" 281 "fmla v13.4s, v30.4s, v8.s[2]\n" 282 "fmla v14.4s, v30.4s, v8.s[3]\n" 283 "fmla v15.4s, v30.4s, v9.s[0]\n" 284 "fmla v16.4s, v30.4s, v10.s[1]\n" 285 "fmla v17.4s, v30.4s, v10.s[2]\n" 286 "fmla v18.4s, v30.4s, v10.s[3]\n" 287 "fmla v19.4s, v30.4s, v11.s[0]\n" 288 "ldr q30, [%x[params], #0x160]\n" 289 "fmla v12.4s, v29.4s, v8.s[2]\n" 290 "fmla v13.4s, v29.4s, v8.s[3]\n" 291 "fmla v14.4s, v29.4s, v9.s[0]\n" 292 "fmla v15.4s, v29.4s, v9.s[1]\n" 293 "fmla v16.4s, v29.4s, v10.s[2]\n" 294 "fmla v17.4s, v29.4s, v10.s[3]\n" 295 "fmla v18.4s, v29.4s, v11.s[0]\n" 296 "fmla v19.4s, v29.4s, v11.s[1]\n" 297 "ldr q29, [%x[params], #0x170]\n" 298 "fmla v12.4s, v28.4s, v8.s[3]\n" 299 "fmla v13.4s, v28.4s, v9.s[0]\n" 300 "fmla v14.4s, v28.4s, v9.s[1]\n" 301 "fmla v15.4s, v28.4s, v9.s[2]\n" 302 "fmla v16.4s, v28.4s, v10.s[3]\n" 303 "fmla v17.4s, v28.4s, v11.s[0]\n" 304 "fmla v18.4s, v28.4s, v11.s[1]\n" 305 "fmla v19.4s, v28.4s, v11.s[2]\n" 306 "ldr q28, [%x[params], #0x180]\n" 307 "fmla v12.4s, v27.4s, v9.s[0]\n" 308 "fmla v13.4s, v27.4s, v9.s[1]\n" 309 "fmla v14.4s, v27.4s, v9.s[2]\n" 310 "fmla v15.4s, v27.4s, v9.s[3]\n" 311 "fmla v16.4s, v27.4s, v11.s[0]\n" 312 "fmla v17.4s, v27.4s, v11.s[1]\n" 313 "fmla v18.4s, v27.4s, v11.s[2]\n" 314 "fmla v19.4s, v27.4s, v11.s[3]\n" 315 "ldr q27, [%x[params], #0x190]\n" 316 "fmin v12.4s, v12.4s, v20.4s\n" 317 "fmin v13.4s, v13.4s, v20.4s\n" 318 "fmin v14.4s, v14.4s, v20.4s\n" 319 "fmax v12.4s, v12.4s, v21.4s\n" 320 "str q12, [x13, x25]\n" 321 "fmax v13.4s, v13.4s, v21.4s\n" 322 "fmax v14.4s, v14.4s, v21.4s\n" 323 "ldr q12, [%x[params], #0x140]\n" 324 "add %x[params], %x[params], #0x1a0\n" 325 "fmin v15.4s, v15.4s, v20.4s\n" 326 "str q13, [x12, x25]\n" 327 "fmin v16.4s, v16.4s, v20.4s\n" 328 "fmin v17.4s, v17.4s, v20.4s\n" 329 "str q14, [x10, x25]\n" 330 "fmin v18.4s, v18.4s, v20.4s\n" 331 "fmax v15.4s, v15.4s, v21.4s\n" 332 "str q15, [x9, x25]\n" 333 "fmax v16.4s, v16.4s, v21.4s\n" 334 "fmax v17.4s, v17.4s, v21.4s\n" 335 "str q16, [x27, x25]\n" 336 "fmax v18.4s, v18.4s, v21.4s\n" 337 "fmin v19.4s, v19.4s, v20.4s\n" 338 "str q17, [x26, x25]\n" 339 "mov v13.16b, v12.16b\n" 340 "str q18, [x24, x25]\n" 341 "fmax v19.4s, v19.4s, v21.4s\n" 342 "mov v14.16b, v12.16b\n" 343 "str q19, [x23, x25]\n" 344 "mov v15.16b, v12.16b\n" 345 "add x25, x25, #0x10\n" 346 "mov v16.16b, v12.16b\n" 347 "mov v17.16b, v12.16b\n" 348 "mov v18.16b, v12.16b\n" 349 "mov v19.16b, v12.16b\n" 352 "fmla v12.4s, v31.4s, v0.s[0]\n" 353 "fmla v13.4s, v31.4s, v0.s[1]\n" 354 "fmla v14.4s, v31.4s, v0.s[2]\n" 355 "fmla v15.4s, v31.4s, v0.s[3]\n" 356 "fmla v16.4s, v31.4s, v2.s[0]\n" 357 "fmla v17.4s, v31.4s, v2.s[1]\n" 358 "fmla v18.4s, v31.4s, v2.s[2]\n" 359 "fmla v19.4s, v31.4s, v2.s[3]\n" 360 "ldr q31, [%x[params], #0x0]\n" 361 "fmla v12.4s, v30.4s, v0.s[1]\n" 362 "fmla v13.4s, v30.4s, v0.s[2]\n" 363 "fmla v14.4s, v30.4s, v0.s[3]\n" 364 "fmla v15.4s, v30.4s, v1.s[0]\n" 365 "fmla v16.4s, v30.4s, v2.s[1]\n" 366 "fmla v17.4s, v30.4s, v2.s[2]\n" 367 "fmla v18.4s, v30.4s, v2.s[3]\n" 368 "fmla v19.4s, v30.4s, v3.s[0]\n" 369 "ldr q30, [%x[params], #0x10]\n" 370 "fmla v12.4s, v29.4s, v0.s[2]\n" 371 "fmla v13.4s, v29.4s, v0.s[3]\n" 372 "fmla v14.4s, v29.4s, v1.s[0]\n" 373 "fmla v15.4s, v29.4s, v1.s[1]\n" 374 "fmla v16.4s, v29.4s, v2.s[2]\n" 375 "fmla v17.4s, v29.4s, v2.s[3]\n" 376 "fmla v18.4s, v29.4s, v3.s[0]\n" 377 "fmla v19.4s, v29.4s, v3.s[1]\n" 378 "ldr q29, [%x[params], #0x20]\n" 379 "fmla v12.4s, v28.4s, v0.s[3]\n" 380 "fmla v13.4s, v28.4s, v1.s[0]\n" 381 "fmla v14.4s, v28.4s, v1.s[1]\n" 382 "fmla v15.4s, v28.4s, v1.s[2]\n" 383 "fmla v16.4s, v28.4s, v2.s[3]\n" 384 "fmla v17.4s, v28.4s, v3.s[0]\n" 385 "fmla v18.4s, v28.4s, v3.s[1]\n" 386 "fmla v19.4s, v28.4s, v3.s[2]\n" 387 "ldr q28, [%x[params], #0x30]\n" 388 "fmla v12.4s, v27.4s, v1.s[0]\n" 389 "fmla v13.4s, v27.4s, v1.s[1]\n" 390 "fmla v14.4s, v27.4s, v1.s[2]\n" 391 "fmla v15.4s, v27.4s, v1.s[3]\n" 392 "fmla v16.4s, v27.4s, v3.s[0]\n" 393 "fmla v17.4s, v27.4s, v3.s[1]\n" 394 "fmla v18.4s, v27.4s, v3.s[2]\n" 395 "fmla v19.4s, v27.4s, v3.s[3]\n" 396 "ldr q27, [%x[params], #0x40]\n" 397 "fmla v12.4s, v31.4s, v2.s[0]\n" 398 "fmla v13.4s, v31.4s, v2.s[1]\n" 399 "fmla v14.4s, v31.4s, v2.s[2]\n" 400 "fmla v15.4s, v31.4s, v2.s[3]\n" 401 "fmla v16.4s, v31.4s, v4.s[0]\n" 402 "fmla v17.4s, v31.4s, v4.s[1]\n" 403 "fmla v18.4s, v31.4s, v4.s[2]\n" 404 "fmla v19.4s, v31.4s, v4.s[3]\n" 405 "ldr q31, [%x[params], #0x50]\n" 406 "fmla v12.4s, v30.4s, v2.s[1]\n" 407 "fmla v13.4s, v30.4s, v2.s[2]\n" 408 "fmla v14.4s, v30.4s, v2.s[3]\n" 409 "fmla v15.4s, v30.4s, v3.s[0]\n" 410 "fmla v16.4s, v30.4s, v4.s[1]\n" 411 "fmla v17.4s, v30.4s, v4.s[2]\n" 412 "fmla v18.4s, v30.4s, v4.s[3]\n" 413 "fmla v19.4s, v30.4s, v5.s[0]\n" 414 "ldr q30, [%x[params], #0x60]\n" 415 "fmla v12.4s, v29.4s, v2.s[2]\n" 416 "fmla v13.4s, v29.4s, v2.s[3]\n" 417 "fmla v14.4s, v29.4s, v3.s[0]\n" 418 "fmla v15.4s, v29.4s, v3.s[1]\n" 419 "fmla v16.4s, v29.4s, v4.s[2]\n" 420 "fmla v17.4s, v29.4s, v4.s[3]\n" 421 "fmla v18.4s, v29.4s, v5.s[0]\n" 422 "fmla v19.4s, v29.4s, v5.s[1]\n" 423 "ldr q29, [%x[params], #0x70]\n" 424 "fmla v12.4s, v28.4s, v2.s[3]\n" 425 "fmla v13.4s, v28.4s, v3.s[0]\n" 426 "fmla v14.4s, v28.4s, v3.s[1]\n" 427 "fmla v15.4s, v28.4s, v3.s[2]\n" 428 "fmla v16.4s, v28.4s, v4.s[3]\n" 429 "fmla v17.4s, v28.4s, v5.s[0]\n" 430 "fmla v18.4s, v28.4s, v5.s[1]\n" 431 "fmla v19.4s, v28.4s, v5.s[2]\n" 432 "ldr q28, [%x[params], #0x80]\n" 433 "fmla v12.4s, v27.4s, v3.s[0]\n" 434 "fmla v13.4s, v27.4s, v3.s[1]\n" 435 "fmla v14.4s, v27.4s, v3.s[2]\n" 436 "fmla v15.4s, v27.4s, v3.s[3]\n" 437 "fmla v16.4s, v27.4s, v5.s[0]\n" 438 "fmla v17.4s, v27.4s, v5.s[1]\n" 439 "fmla v18.4s, v27.4s, v5.s[2]\n" 440 "fmla v19.4s, v27.4s, v5.s[3]\n" 441 "ldr q27, [%x[params], #0x90]\n" 442 "fmla v12.4s, v31.4s, v4.s[0]\n" 443 "fmla v13.4s, v31.4s, v4.s[1]\n" 444 "fmla v14.4s, v31.4s, v4.s[2]\n" 445 "fmla v15.4s, v31.4s, v4.s[3]\n" 446 "fmla v16.4s, v31.4s, v6.s[0]\n" 447 "fmla v17.4s, v31.4s, v6.s[1]\n" 448 "fmla v18.4s, v31.4s, v6.s[2]\n" 449 "fmla v19.4s, v31.4s, v6.s[3]\n" 450 "ldr q31, [%x[params], #0xa0]\n" 451 "fmla v12.4s, v30.4s, v4.s[1]\n" 452 "fmla v13.4s, v30.4s, v4.s[2]\n" 453 "fmla v14.4s, v30.4s, v4.s[3]\n" 454 "fmla v15.4s, v30.4s, v5.s[0]\n" 455 "fmla v16.4s, v30.4s, v6.s[1]\n" 456 "fmla v17.4s, v30.4s, v6.s[2]\n" 457 "fmla v18.4s, v30.4s, v6.s[3]\n" 458 "fmla v19.4s, v30.4s, v7.s[0]\n" 459 "ldr q30, [%x[params], #0xb0]\n" 460 "fmla v12.4s, v29.4s, v4.s[2]\n" 461 "fmla v13.4s, v29.4s, v4.s[3]\n" 462 "fmla v14.4s, v29.4s, v5.s[0]\n" 463 "fmla v15.4s, v29.4s, v5.s[1]\n" 464 "fmla v16.4s, v29.4s, v6.s[2]\n" 465 "fmla v17.4s, v29.4s, v6.s[3]\n" 466 "fmla v18.4s, v29.4s, v7.s[0]\n" 467 "fmla v19.4s, v29.4s, v7.s[1]\n" 468 "ldr q29, [%x[params], #0xc0]\n" 469 "fmla v12.4s, v28.4s, v4.s[3]\n" 470 "fmla v13.4s, v28.4s, v5.s[0]\n" 471 "fmla v14.4s, v28.4s, v5.s[1]\n" 472 "fmla v15.4s, v28.4s, v5.s[2]\n" 473 "fmla v16.4s, v28.4s, v6.s[3]\n" 474 "fmla v17.4s, v28.4s, v7.s[0]\n" 475 "fmla v18.4s, v28.4s, v7.s[1]\n" 476 "fmla v19.4s, v28.4s, v7.s[2]\n" 477 "ldr q28, [%x[params], #0xd0]\n" 478 "fmla v12.4s, v27.4s, v5.s[0]\n" 479 "fmla v13.4s, v27.4s, v5.s[1]\n" 480 "fmla v14.4s, v27.4s, v5.s[2]\n" 481 "fmla v15.4s, v27.4s, v5.s[3]\n" 482 "fmla v16.4s, v27.4s, v7.s[0]\n" 483 "fmla v17.4s, v27.4s, v7.s[1]\n" 484 "fmla v18.4s, v27.4s, v7.s[2]\n" 485 "fmla v19.4s, v27.4s, v7.s[3]\n" 486 "ldr q27, [%x[params], #0xe0]\n" 487 "fmla v12.4s, v31.4s, v6.s[0]\n" 488 "fmla v13.4s, v31.4s, v6.s[1]\n" 489 "fmla v14.4s, v31.4s, v6.s[2]\n" 490 "fmla v15.4s, v31.4s, v6.s[3]\n" 491 "fmla v16.4s, v31.4s, v8.s[0]\n" 492 "fmla v17.4s, v31.4s, v8.s[1]\n" 493 "fmla v18.4s, v31.4s, v8.s[2]\n" 494 "fmla v19.4s, v31.4s, v8.s[3]\n" 495 "ldr q31, [%x[params], #0xf0]\n" 496 "fmla v12.4s, v30.4s, v6.s[1]\n" 497 "fmla v13.4s, v30.4s, v6.s[2]\n" 498 "fmla v14.4s, v30.4s, v6.s[3]\n" 499 "fmla v15.4s, v30.4s, v7.s[0]\n" 500 "fmla v16.4s, v30.4s, v8.s[1]\n" 501 "fmla v17.4s, v30.4s, v8.s[2]\n" 502 "fmla v18.4s, v30.4s, v8.s[3]\n" 503 "fmla v19.4s, v30.4s, v9.s[0]\n" 504 "ldr q30, [%x[params], #0x100]\n" 505 "fmla v12.4s, v29.4s, v6.s[2]\n" 506 "fmla v13.4s, v29.4s, v6.s[3]\n" 507 "fmla v14.4s, v29.4s, v7.s[0]\n" 508 "fmla v15.4s, v29.4s, v7.s[1]\n" 509 "fmla v16.4s, v29.4s, v8.s[2]\n" 510 "fmla v17.4s, v29.4s, v8.s[3]\n" 511 "fmla v18.4s, v29.4s, v9.s[0]\n" 512 "fmla v19.4s, v29.4s, v9.s[1]\n" 513 "ldr q29, [%x[params], #0x110]\n" 514 "fmla v12.4s, v28.4s, v6.s[3]\n" 515 "fmla v13.4s, v28.4s, v7.s[0]\n" 516 "fmla v14.4s, v28.4s, v7.s[1]\n" 517 "fmla v15.4s, v28.4s, v7.s[2]\n" 518 "fmla v16.4s, v28.4s, v8.s[3]\n" 519 "fmla v17.4s, v28.4s, v9.s[0]\n" 520 "fmla v18.4s, v28.4s, v9.s[1]\n" 521 "fmla v19.4s, v28.4s, v9.s[2]\n" 522 "ldr q28, [%x[params], #0x120]\n" 523 "fmla v12.4s, v27.4s, v7.s[0]\n" 524 "fmla v13.4s, v27.4s, v7.s[1]\n" 525 "fmla v14.4s, v27.4s, v7.s[2]\n" 526 "fmla v15.4s, v27.4s, v7.s[3]\n" 527 "fmla v16.4s, v27.4s, v9.s[0]\n" 528 "fmla v17.4s, v27.4s, v9.s[1]\n" 529 "fmla v18.4s, v27.4s, v9.s[2]\n" 530 "fmla v19.4s, v27.4s, v9.s[3]\n" 531 "ldr q27, [%x[params], #0x130]\n" 532 "add %x[params], %x[params], #0x140\n" 533 "fmla v12.4s, v31.4s, v8.s[0]\n" 534 "fmla v13.4s, v31.4s, v8.s[1]\n" 535 "fmla v14.4s, v31.4s, v8.s[2]\n" 536 "fmla v15.4s, v31.4s, v8.s[3]\n" 537 "fmla v16.4s, v31.4s, v10.s[0]\n" 538 "fmla v17.4s, v31.4s, v10.s[1]\n" 539 "fmla v18.4s, v31.4s, v10.s[2]\n" 540 "fmla v19.4s, v31.4s, v10.s[3]\n" 541 "fmla v12.4s, v30.4s, v8.s[1]\n" 542 "fmla v13.4s, v30.4s, v8.s[2]\n" 543 "fmla v14.4s, v30.4s, v8.s[3]\n" 544 "fmla v15.4s, v30.4s, v9.s[0]\n" 545 "fmla v16.4s, v30.4s, v10.s[1]\n" 546 "fmla v17.4s, v30.4s, v10.s[2]\n" 547 "fmla v18.4s, v30.4s, v10.s[3]\n" 548 "fmla v19.4s, v30.4s, v11.s[0]\n" 549 "fmla v12.4s, v29.4s, v8.s[2]\n" 550 "fmla v13.4s, v29.4s, v8.s[3]\n" 551 "fmla v14.4s, v29.4s, v9.s[0]\n" 552 "fmla v15.4s, v29.4s, v9.s[1]\n" 553 "fmla v16.4s, v29.4s, v10.s[2]\n" 554 "fmla v17.4s, v29.4s, v10.s[3]\n" 555 "fmla v18.4s, v29.4s, v11.s[0]\n" 556 "fmla v19.4s, v29.4s, v11.s[1]\n" 557 "fmla v12.4s, v28.4s, v8.s[3]\n" 558 "fmla v13.4s, v28.4s, v9.s[0]\n" 559 "fmla v14.4s, v28.4s, v9.s[1]\n" 560 "fmla v15.4s, v28.4s, v9.s[2]\n" 561 "fmla v16.4s, v28.4s, v10.s[3]\n" 562 "fmla v17.4s, v28.4s, v11.s[0]\n" 563 "fmla v18.4s, v28.4s, v11.s[1]\n" 564 "fmla v19.4s, v28.4s, v11.s[2]\n" 565 "fmla v12.4s, v27.4s, v9.s[0]\n" 566 "fmla v13.4s, v27.4s, v9.s[1]\n" 567 "fmla v14.4s, v27.4s, v9.s[2]\n" 568 "fmla v15.4s, v27.4s, v9.s[3]\n" 569 "fmla v16.4s, v27.4s, v11.s[0]\n" 570 "fmla v17.4s, v27.4s, v11.s[1]\n" 571 "fmla v18.4s, v27.4s, v11.s[2]\n" 572 "fmla v19.4s, v27.4s, v11.s[3]\n" 573 "fmin v12.4s, v12.4s, v20.4s\n" 574 "fmin v13.4s, v13.4s, v20.4s\n" 575 "fmin v14.4s, v14.4s, v20.4s\n" 576 "fmax v12.4s, v12.4s, v21.4s\n" 577 "str q12, [x13, x25]\n" 578 "fmax v13.4s, v13.4s, v21.4s\n" 579 "fmax v14.4s, v14.4s, v21.4s\n" 580 "str q13, [x12, x25]\n" 581 "fmin v15.4s, v15.4s, v20.4s\n" 582 "fmin v16.4s, v16.4s, v20.4s\n" 583 "str q14, [x10, x25]\n" 584 "fmin v17.4s, v17.4s, v20.4s\n" 585 "fmin v18.4s, v18.4s, v20.4s\n" 586 "fmax v15.4s, v15.4s, v21.4s\n" 587 "str q15, [x9, x25]\n" 588 "fmax v16.4s, v16.4s, v21.4s\n" 589 "fmax v17.4s, v17.4s, v21.4s\n" 590 "str q16, [x27, x25]\n" 591 "fmax v18.4s, v18.4s, v21.4s\n" 592 "fmin v19.4s, v19.4s, v20.4s\n" 593 "str q17, [x26, x25]\n" 594 "fmax v19.4s, v19.4s, v21.4s\n" 595 "str q18, [x24, x25]\n" 596 "str q19, [x23, x25]\n" 597 "add x25, x25, #0x10\n" 599 "tst %x[channel_multiplier], #0x3\n" 601 "ldr q12, [%x[params], #0x0]\n" 602 "mov v13.16b, v12.16b\n" 603 "ldr q31, [%x[params], #0x10]\n" 604 "mov v14.16b, v12.16b\n" 605 "ldr q30, [%x[params], #0x20]\n" 606 "mov v15.16b, v12.16b\n" 607 "ldr q29, [%x[params], #0x30]\n" 608 "mov v16.16b, v12.16b\n" 609 "ldr q28, [%x[params], #0x40]\n" 610 "mov v17.16b, v12.16b\n" 611 "ldr q27, [%x[params], #0x50]\n" 612 "mov v18.16b, v12.16b\n" 613 "mov v19.16b, v12.16b\n" 614 "fmla v12.4s, v31.4s, v0.s[0]\n" 615 "fmla v13.4s, v31.4s, v0.s[1]\n" 616 "fmla v14.4s, v31.4s, v0.s[2]\n" 617 "fmla v15.4s, v31.4s, v0.s[3]\n" 618 "fmla v16.4s, v31.4s, v2.s[0]\n" 619 "fmla v17.4s, v31.4s, v2.s[1]\n" 620 "fmla v18.4s, v31.4s, v2.s[2]\n" 621 "fmla v19.4s, v31.4s, v2.s[3]\n" 622 "ldr q31, [%x[params], #0x60]\n" 623 "fmla v12.4s, v30.4s, v0.s[1]\n" 624 "fmla v13.4s, v30.4s, v0.s[2]\n" 625 "fmla v14.4s, v30.4s, v0.s[3]\n" 626 "fmla v15.4s, v30.4s, v1.s[0]\n" 627 "fmla v16.4s, v30.4s, v2.s[1]\n" 628 "fmla v17.4s, v30.4s, v2.s[2]\n" 629 "fmla v18.4s, v30.4s, v2.s[3]\n" 630 "fmla v19.4s, v30.4s, v3.s[0]\n" 631 "ldr q30, [%x[params], #0x70]\n" 632 "fmla v12.4s, v29.4s, v0.s[2]\n" 633 "fmla v13.4s, v29.4s, v0.s[3]\n" 634 "fmla v14.4s, v29.4s, v1.s[0]\n" 635 "fmla v15.4s, v29.4s, v1.s[1]\n" 636 "fmla v16.4s, v29.4s, v2.s[2]\n" 637 "fmla v17.4s, v29.4s, v2.s[3]\n" 638 "fmla v18.4s, v29.4s, v3.s[0]\n" 639 "fmla v19.4s, v29.4s, v3.s[1]\n" 640 "ldr q29, [%x[params], #0x80]\n" 641 "fmla v12.4s, v28.4s, v0.s[3]\n" 642 "fmla v13.4s, v28.4s, v1.s[0]\n" 643 "fmla v14.4s, v28.4s, v1.s[1]\n" 644 "fmla v15.4s, v28.4s, v1.s[2]\n" 645 "fmla v16.4s, v28.4s, v2.s[3]\n" 646 "fmla v17.4s, v28.4s, v3.s[0]\n" 647 "fmla v18.4s, v28.4s, v3.s[1]\n" 648 "fmla v19.4s, v28.4s, v3.s[2]\n" 649 "ldr q28, [%x[params], #0x90]\n" 650 "fmla v12.4s, v27.4s, v1.s[0]\n" 651 "fmla v13.4s, v27.4s, v1.s[1]\n" 652 "fmla v14.4s, v27.4s, v1.s[2]\n" 653 "fmla v15.4s, v27.4s, v1.s[3]\n" 654 "fmla v16.4s, v27.4s, v3.s[0]\n" 655 "fmla v17.4s, v27.4s, v3.s[1]\n" 656 "fmla v18.4s, v27.4s, v3.s[2]\n" 657 "fmla v19.4s, v27.4s, v3.s[3]\n" 658 "ldr q27, [%x[params], #0xa0]\n" 659 "fmla v12.4s, v31.4s, v2.s[0]\n" 660 "fmla v13.4s, v31.4s, v2.s[1]\n" 661 "fmla v14.4s, v31.4s, v2.s[2]\n" 662 "fmla v15.4s, v31.4s, v2.s[3]\n" 663 "fmla v16.4s, v31.4s, v4.s[0]\n" 664 "fmla v17.4s, v31.4s, v4.s[1]\n" 665 "fmla v18.4s, v31.4s, v4.s[2]\n" 666 "fmla v19.4s, v31.4s, v4.s[3]\n" 667 "ldr q31, [%x[params], #0xb0]\n" 668 "fmla v12.4s, v30.4s, v2.s[1]\n" 669 "fmla v13.4s, v30.4s, v2.s[2]\n" 670 "fmla v14.4s, v30.4s, v2.s[3]\n" 671 "fmla v15.4s, v30.4s, v3.s[0]\n" 672 "fmla v16.4s, v30.4s, v4.s[1]\n" 673 "fmla v17.4s, v30.4s, v4.s[2]\n" 674 "fmla v18.4s, v30.4s, v4.s[3]\n" 675 "fmla v19.4s, v30.4s, v5.s[0]\n" 676 "ldr q30, [%x[params], #0xc0]\n" 677 "fmla v12.4s, v29.4s, v2.s[2]\n" 678 "fmla v13.4s, v29.4s, v2.s[3]\n" 679 "fmla v14.4s, v29.4s, v3.s[0]\n" 680 "fmla v15.4s, v29.4s, v3.s[1]\n" 681 "fmla v16.4s, v29.4s, v4.s[2]\n" 682 "fmla v17.4s, v29.4s, v4.s[3]\n" 683 "fmla v18.4s, v29.4s, v5.s[0]\n" 684 "fmla v19.4s, v29.4s, v5.s[1]\n" 685 "ldr q29, [%x[params], #0xd0]\n" 686 "fmla v12.4s, v28.4s, v2.s[3]\n" 687 "fmla v13.4s, v28.4s, v3.s[0]\n" 688 "fmla v14.4s, v28.4s, v3.s[1]\n" 689 "fmla v15.4s, v28.4s, v3.s[2]\n" 690 "fmla v16.4s, v28.4s, v4.s[3]\n" 691 "fmla v17.4s, v28.4s, v5.s[0]\n" 692 "fmla v18.4s, v28.4s, v5.s[1]\n" 693 "fmla v19.4s, v28.4s, v5.s[2]\n" 694 "ldr q28, [%x[params], #0xe0]\n" 695 "fmla v12.4s, v27.4s, v3.s[0]\n" 696 "fmla v13.4s, v27.4s, v3.s[1]\n" 697 "fmla v14.4s, v27.4s, v3.s[2]\n" 698 "fmla v15.4s, v27.4s, v3.s[3]\n" 699 "fmla v16.4s, v27.4s, v5.s[0]\n" 700 "fmla v17.4s, v27.4s, v5.s[1]\n" 701 "fmla v18.4s, v27.4s, v5.s[2]\n" 702 "fmla v19.4s, v27.4s, v5.s[3]\n" 703 "ldr q27, [%x[params], #0xf0]\n" 704 "fmla v12.4s, v31.4s, v4.s[0]\n" 705 "fmla v13.4s, v31.4s, v4.s[1]\n" 706 "fmla v14.4s, v31.4s, v4.s[2]\n" 707 "fmla v15.4s, v31.4s, v4.s[3]\n" 708 "fmla v16.4s, v31.4s, v6.s[0]\n" 709 "fmla v17.4s, v31.4s, v6.s[1]\n" 710 "fmla v18.4s, v31.4s, v6.s[2]\n" 711 "fmla v19.4s, v31.4s, v6.s[3]\n" 712 "ldr q31, [%x[params], #0x100]\n" 713 "fmla v12.4s, v30.4s, v4.s[1]\n" 714 "fmla v13.4s, v30.4s, v4.s[2]\n" 715 "fmla v14.4s, v30.4s, v4.s[3]\n" 716 "fmla v15.4s, v30.4s, v5.s[0]\n" 717 "fmla v16.4s, v30.4s, v6.s[1]\n" 718 "fmla v17.4s, v30.4s, v6.s[2]\n" 719 "fmla v18.4s, v30.4s, v6.s[3]\n" 720 "fmla v19.4s, v30.4s, v7.s[0]\n" 721 "ldr q30, [%x[params], #0x110]\n" 722 "fmla v12.4s, v29.4s, v4.s[2]\n" 723 "fmla v13.4s, v29.4s, v4.s[3]\n" 724 "fmla v14.4s, v29.4s, v5.s[0]\n" 725 "fmla v15.4s, v29.4s, v5.s[1]\n" 726 "fmla v16.4s, v29.4s, v6.s[2]\n" 727 "fmla v17.4s, v29.4s, v6.s[3]\n" 728 "fmla v18.4s, v29.4s, v7.s[0]\n" 729 "fmla v19.4s, v29.4s, v7.s[1]\n" 730 "ldr q29, [%x[params], #0x120]\n" 731 "fmla v12.4s, v28.4s, v4.s[3]\n" 732 "fmla v13.4s, v28.4s, v5.s[0]\n" 733 "fmla v14.4s, v28.4s, v5.s[1]\n" 734 "fmla v15.4s, v28.4s, v5.s[2]\n" 735 "fmla v16.4s, v28.4s, v6.s[3]\n" 736 "fmla v17.4s, v28.4s, v7.s[0]\n" 737 "fmla v18.4s, v28.4s, v7.s[1]\n" 738 "fmla v19.4s, v28.4s, v7.s[2]\n" 739 "ldr q28, [%x[params], #0x130]\n" 740 "fmla v12.4s, v27.4s, v5.s[0]\n" 741 "fmla v13.4s, v27.4s, v5.s[1]\n" 742 "fmla v14.4s, v27.4s, v5.s[2]\n" 743 "fmla v15.4s, v27.4s, v5.s[3]\n" 744 "fmla v16.4s, v27.4s, v7.s[0]\n" 745 "fmla v17.4s, v27.4s, v7.s[1]\n" 746 "fmla v18.4s, v27.4s, v7.s[2]\n" 747 "fmla v19.4s, v27.4s, v7.s[3]\n" 748 "ldr q27, [%x[params], #0x140]\n" 749 "fmla v12.4s, v31.4s, v6.s[0]\n" 750 "fmla v13.4s, v31.4s, v6.s[1]\n" 751 "fmla v14.4s, v31.4s, v6.s[2]\n" 752 "fmla v15.4s, v31.4s, v6.s[3]\n" 753 "fmla v16.4s, v31.4s, v8.s[0]\n" 754 "fmla v17.4s, v31.4s, v8.s[1]\n" 755 "fmla v18.4s, v31.4s, v8.s[2]\n" 756 "fmla v19.4s, v31.4s, v8.s[3]\n" 757 "ldr q31, [%x[params], #0x150]\n" 758 "fmla v12.4s, v30.4s, v6.s[1]\n" 759 "fmla v13.4s, v30.4s, v6.s[2]\n" 760 "fmla v14.4s, v30.4s, v6.s[3]\n" 761 "fmla v15.4s, v30.4s, v7.s[0]\n" 762 "fmla v16.4s, v30.4s, v8.s[1]\n" 763 "fmla v17.4s, v30.4s, v8.s[2]\n" 764 "fmla v18.4s, v30.4s, v8.s[3]\n" 765 "fmla v19.4s, v30.4s, v9.s[0]\n" 766 "ldr q30, [%x[params], #0x160]\n" 767 "fmla v12.4s, v29.4s, v6.s[2]\n" 768 "fmla v13.4s, v29.4s, v6.s[3]\n" 769 "fmla v14.4s, v29.4s, v7.s[0]\n" 770 "fmla v15.4s, v29.4s, v7.s[1]\n" 771 "fmla v16.4s, v29.4s, v8.s[2]\n" 772 "fmla v17.4s, v29.4s, v8.s[3]\n" 773 "fmla v18.4s, v29.4s, v9.s[0]\n" 774 "fmla v19.4s, v29.4s, v9.s[1]\n" 775 "ldr q29, [%x[params], #0x170]\n" 776 "fmla v12.4s, v28.4s, v6.s[3]\n" 777 "fmla v13.4s, v28.4s, v7.s[0]\n" 778 "fmla v14.4s, v28.4s, v7.s[1]\n" 779 "fmla v15.4s, v28.4s, v7.s[2]\n" 780 "fmla v16.4s, v28.4s, v8.s[3]\n" 781 "fmla v17.4s, v28.4s, v9.s[0]\n" 782 "fmla v18.4s, v28.4s, v9.s[1]\n" 783 "fmla v19.4s, v28.4s, v9.s[2]\n" 784 "ldr q28, [%x[params], #0x180]\n" 785 "fmla v12.4s, v27.4s, v7.s[0]\n" 786 "fmla v13.4s, v27.4s, v7.s[1]\n" 787 "fmla v14.4s, v27.4s, v7.s[2]\n" 788 "fmla v15.4s, v27.4s, v7.s[3]\n" 789 "fmla v16.4s, v27.4s, v9.s[0]\n" 790 "fmla v17.4s, v27.4s, v9.s[1]\n" 791 "fmla v18.4s, v27.4s, v9.s[2]\n" 792 "fmla v19.4s, v27.4s, v9.s[3]\n" 793 "ldr q27, [%x[params], #0x190]\n" 794 "add %x[params], %x[params], #0x1a0\n" 795 "fmla v12.4s, v31.4s, v8.s[0]\n" 796 "fmla v13.4s, v31.4s, v8.s[1]\n" 797 "fmla v14.4s, v31.4s, v8.s[2]\n" 798 "fmla v15.4s, v31.4s, v8.s[3]\n" 799 "fmla v16.4s, v31.4s, v10.s[0]\n" 800 "fmla v17.4s, v31.4s, v10.s[1]\n" 801 "fmla v18.4s, v31.4s, v10.s[2]\n" 802 "fmla v19.4s, v31.4s, v10.s[3]\n" 803 "fmla v12.4s, v30.4s, v8.s[1]\n" 804 "fmla v13.4s, v30.4s, v8.s[2]\n" 805 "fmla v14.4s, v30.4s, v8.s[3]\n" 806 "fmla v15.4s, v30.4s, v9.s[0]\n" 807 "fmla v16.4s, v30.4s, v10.s[1]\n" 808 "fmla v17.4s, v30.4s, v10.s[2]\n" 809 "fmla v18.4s, v30.4s, v10.s[3]\n" 810 "fmla v19.4s, v30.4s, v11.s[0]\n" 811 "fmla v12.4s, v29.4s, v8.s[2]\n" 812 "fmla v13.4s, v29.4s, v8.s[3]\n" 813 "fmla v14.4s, v29.4s, v9.s[0]\n" 814 "fmla v15.4s, v29.4s, v9.s[1]\n" 815 "fmla v16.4s, v29.4s, v10.s[2]\n" 816 "fmla v17.4s, v29.4s, v10.s[3]\n" 817 "fmla v18.4s, v29.4s, v11.s[0]\n" 818 "fmla v19.4s, v29.4s, v11.s[1]\n" 819 "fmla v12.4s, v28.4s, v8.s[3]\n" 820 "fmla v13.4s, v28.4s, v9.s[0]\n" 821 "fmla v14.4s, v28.4s, v9.s[1]\n" 822 "fmla v15.4s, v28.4s, v9.s[2]\n" 823 "fmla v16.4s, v28.4s, v10.s[3]\n" 824 "fmla v17.4s, v28.4s, v11.s[0]\n" 825 "fmla v18.4s, v28.4s, v11.s[1]\n" 826 "fmla v19.4s, v28.4s, v11.s[2]\n" 827 "fmla v12.4s, v27.4s, v9.s[0]\n" 828 "fmla v13.4s, v27.4s, v9.s[1]\n" 829 "fmla v14.4s, v27.4s, v9.s[2]\n" 830 "fmla v15.4s, v27.4s, v9.s[3]\n" 831 "fmla v16.4s, v27.4s, v11.s[0]\n" 832 "fmla v17.4s, v27.4s, v11.s[1]\n" 833 "fmla v18.4s, v27.4s, v11.s[2]\n" 834 "fmla v19.4s, v27.4s, v11.s[3]\n" 835 "fmin v12.4s, v12.4s, v20.4s\n" 836 "fmin v13.4s, v13.4s, v20.4s\n" 837 "fmin v14.4s, v14.4s, v20.4s\n" 838 "fmax v12.4s, v12.4s, v21.4s\n" 839 "fmax v13.4s, v13.4s, v21.4s\n" 840 "fmax v14.4s, v14.4s, v21.4s\n" 841 "fmin v15.4s, v15.4s, v20.4s\n" 842 "fmin v16.4s, v16.4s, v20.4s\n" 843 "fmin v17.4s, v17.4s, v20.4s\n" 844 "fmax v15.4s, v15.4s, v21.4s\n" 845 "fmax v16.4s, v16.4s, v21.4s\n" 846 "fmax v17.4s, v17.4s, v21.4s\n" 847 "fmin v18.4s, v18.4s, v20.4s\n" 848 "fmin v19.4s, v19.4s, v20.4s\n" 849 "fmax v18.4s, v18.4s, v21.4s\n" 850 "fmax v19.4s, v19.4s, v21.4s\n" 851 "tbz %x[channel_multiplier], #1, 4f\n" 852 "add x19, x13, x25\n" 853 "st1 { v12.d }[0], [x19]\n" 854 "add x19, x12, x25\n" 855 "st1 { v13.d }[0], [x19]\n" 856 "add x19, x10, x25\n" 857 "st1 { v14.d }[0], [x19]\n" 859 "st1 { v15.d }[0], [x19]\n" 860 "add x19, x27, x25\n" 861 "st1 { v16.d }[0], [x19]\n" 862 "add x19, x26, x25\n" 863 "st1 { v17.d }[0], [x19]\n" 864 "add x19, x24, x25\n" 865 "st1 { v18.d }[0], [x19]\n" 866 "add x19, x23, x25\n" 867 "st1 { v19.d }[0], [x19]\n" 868 "add x25, x25, #0x8\n" 869 "tbz %x[channel_multiplier], #0, 5f\n" 870 "add x19, x13, x25\n" 871 "st1 { v12.s }[2], [x19]\n" 872 "add x19, x12, x25\n" 873 "st1 { v13.s }[2], [x19]\n" 874 "add x19, x10, x25\n" 875 "st1 { v14.s }[2], [x19]\n" 877 "st1 { v15.s }[2], [x19]\n" 878 "add x19, x27, x25\n" 879 "st1 { v16.s }[2], [x19]\n" 880 "add x19, x26, x25\n" 881 "st1 { v17.s }[2], [x19]\n" 882 "add x19, x24, x25\n" 883 "st1 { v18.s }[2], [x19]\n" 884 "add x19, x23, x25\n" 885 "st1 { v19.s }[2], [x19]\n" 888 "tbz %x[channel_multiplier], #0, 5f\n" 889 "add x19, x13, x25\n" 890 "st1 { v12.s }[0], [x19]\n" 891 "add x19, x12, x25\n" 892 "st1 { v13.s }[0], [x19]\n" 893 "add x19, x10, x25\n" 894 "st1 { v14.s }[0], [x19]\n" 896 "st1 { v15.s }[0], [x19]\n" 897 "add x19, x27, x25\n" 898 "st1 { v16.s }[0], [x19]\n" 899 "add x19, x26, x25\n" 900 "st1 { v17.s }[0], [x19]\n" 901 "add x19, x24, x25\n" 902 "st1 { v18.s }[0], [x19]\n" 903 "add x19, x23, x25\n" 904 "st1 { v19.s }[0], [x19]\n" 909 : [params]
"+&r" (params)
910 : [channel_multiplier]
"r" (n_output_channels), [clamps]
"r" (minmax_vals), [inptrs]
"r" (inptrs), [outptrs]
"r" (outptrs)
911 :
"cc",
"memory",
"v0",
"v1",
"v2",
"v3",
"v4",
"v5",
"v6",
"v7",
"v8",
"v9",
"v10",
"v11",
"v12",
"v13",
"v14",
"v15",
"v16",
"v17",
"v18",
"v19",
"v20",
"v21",
"v27",
"v28",
"v29",
"v30",
"v31",
"x9",
"x10",
"x11",
"x12",
"x13",
"x19",
"x20",
"x21",
"x22",
"x23",
"x24",
"x25",
"x26",
"x27",
"x28"
void a64_fp32_packed_to_nhwc_5x5_s1_with_multiplier_output2x4_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 &)