32 const unsigned int n_tile_rows,
33 const unsigned int n_tile_cols,
38 int64_t ld_output_row,
39 int64_t ld_output_col,
41 unsigned int n_channels,
42 const float activation_min,
43 const float activation_max
48 const uint64_t n_tile_rows, n_tile_cols;
50 const uint64_t ld_input_row;
51 const uint64_t ld_input_col;
53 const uint64_t ld_output_row;
54 const uint64_t ld_output_col;
58 uint64_t tile_i = 0, tile_j = 0;
61 const unsigned int n_tile_rows,
62 const unsigned int n_tile_cols,
67 int64_t ld_output_row,
68 int64_t ld_output_col,
70 const float activation_min,
71 const float activation_max
72 ) : n_tile_rows(n_tile_rows), n_tile_cols(n_tile_cols), inptr(inptr),
73 ld_input_row(ld_input_row), ld_input_col(ld_input_col), outptr(outptr),
74 ld_output_row(ld_output_row), ld_output_col(ld_output_col),
75 params(params), min(activation_min), max(activation_max)
81 n_tile_rows, n_tile_cols,
82 inptr, ld_input_row, ld_input_col,
83 outptr, ld_output_row, ld_output_col,
84 params, activation_min, activation_max
91 "str x26, [%x[params_struct], %[offsetof_args_tile_i]]\n" 94 "str x25, [%x[params_struct], %[offsetof_args_tile_j]]\n" 95 "ldr x24, [%x[params_struct], %[offsetof_args_ld_input_row]]\n" 96 "ldr x3, [%x[params_struct], %[offsetof_args_ld_input_col]]\n" 98 "ldr x23, [%x[params_struct], %[offsetof_args_ld_output_row]]\n" 99 "madd x20, x25, x3, x20\n" 100 "ldr x4, [%x[params_struct], %[offsetof_args_ld_output_col]]\n" 102 "mul x19, x26, x23\n" 103 "ldr x5, [%x[params_struct], %[offsetof_args_inptr]]\n" 104 "ldr x6, [%x[params_struct], %[offsetof_args_outptr]]\n" 106 "mul x20, x20, x22\n" 107 "add x5, x5, x20, LSL #2\n" 108 "add x8, x5, x24, LSL #2\n" 109 "ldr x17, [%x[params_struct], %[offsetof_args_params]]\n" 110 "madd x19, x25, x4, x19\n" 111 "add x16, x8, x24, LSL #2\n" 113 "mul x19, x19, x21\n" 114 "lsr x21, %x[n_channels], #0x2\n" 115 "add x15, x16, x24, LSL #2\n" 117 "add x13, x15, x24, LSL #2\n" 119 "add x6, x6, x19, LSL #2\n" 120 "add x20, %x[params_struct], %[offsetof_args_min]\n" 121 "add x19, %x[params_struct], %[offsetof_args_max]\n" 122 "ld1r { v18.4s }, [x20]\n" 123 "ld1r { v17.4s }, [x19]\n" 124 "add x11, x13, x24, LSL #2\n" 126 "add x9, x6, x23, LSL #2\n" 129 "sub x19, XZR, x22\n" 131 "ldr q16, [x17, #0x0]\n" 132 "cmp x22, x21, LSL #4\n" 133 "ldr q0, [x17, #0x10]\n" 134 "ldr q1, [x17, #0x20]\n" 135 "ldr q2, [x17, #0x30]\n" 136 "ldr q3, [x17, #0x40]\n" 137 "ldr q4, [x17, #0x50]\n" 138 "ld1 { v5.4s }, [x5]\n" 139 "add x17, x17, #0x60\n" 141 "ld1 { v7.4s }, [x8]\n" 144 "ldr q13, [x8, x7]\n" 145 "ldr q11, [x5, x14]\n" 146 "ldr q12, [x5, x12]\n" 147 "ldr q10, [x8, x10]\n" 148 "ld1 { v14.4s }, [x16]\n" 151 "mov v28.16b, v16.16b\n fmla v28.4s, v0.4s, v5.4s\n" 152 "mov v29.16b, v16.16b\n fmla v29.4s, v0.4s, v6.4s\n" 153 "ldr q5, [x8, x14]\n" 154 "add x22, x22, #0x10\n" 155 "mov v30.16b, v16.16b\n fmla v30.4s, v0.4s, v7.4s\n" 156 "mov v31.16b, v16.16b\n fmla v31.4s, v0.4s, v8.4s\n" 157 "ldr q0, [x17, #0x0]\n" 158 "cmp x22, x21, LSL #4\n" 159 "fmla v28.4s, v1.4s, v6.4s\n" 160 "fmla v29.4s, v1.4s, v9.4s\n" 161 "ldr q6, [x8, x12]\n" 162 "add x8, x8, #0x10\n" 163 "fmla v30.4s, v1.4s, v8.4s\n" 164 "fmla v31.4s, v1.4s, v13.4s\n" 165 "ldr q1, [x17, #0x10]\n" 166 "add x19, x19, #0x10\n" 167 "fmla v28.4s, v2.4s, v9.4s\n" 168 "fmla v29.4s, v2.4s, v11.4s\n" 169 "ldr q9, [x5, x10]\n" 170 "add x5, x5, #0x10\n" 171 "fmla v30.4s, v2.4s, v13.4s\n" 172 "fmla v31.4s, v2.4s, v5.4s\n" 173 "ldr q2, [x17, #0x20]\n" 174 "add x20, x20, #0x10\n" 175 "fmla v28.4s, v3.4s, v11.4s\n" 176 "fmla v29.4s, v3.4s, v12.4s\n" 177 "ldr q11, [x16, x3]\n" 178 "ldr q16, [x17, #0x140]\n" 179 "fmla v30.4s, v3.4s, v5.4s\n" 180 "fmla v31.4s, v3.4s, v6.4s\n" 181 "ldr q3, [x17, #0x30]\n" 182 "fmla v28.4s, v4.4s, v12.4s\n" 183 "fmla v29.4s, v4.4s, v9.4s\n" 184 "ldr q12, [x16, x7]\n" 185 "ldr q9, [x16, x14]\n" 186 "fmla v30.4s, v4.4s, v6.4s\n" 187 "fmla v31.4s, v4.4s, v10.4s\n" 188 "ldr q4, [x17, #0x40]\n" 189 "fmla v28.4s, v0.4s, v7.4s\n" 190 "fmla v29.4s, v0.4s, v8.4s\n" 191 "ld1 { v7.4s }, [x8]\n" 192 "fmla v30.4s, v0.4s, v14.4s\n" 193 "fmla v31.4s, v0.4s, v11.4s\n" 194 "ldr q0, [x17, #0x50]\n" 195 "fmla v28.4s, v1.4s, v8.4s\n" 196 "fmla v29.4s, v1.4s, v13.4s\n" 197 "ldr q8, [x16, x10]\n" 198 "fmla v30.4s, v1.4s, v11.4s\n" 199 "fmla v31.4s, v1.4s, v12.4s\n" 200 "ldr q1, [x17, #0x60]\n" 201 "fmla v28.4s, v2.4s, v13.4s\n" 202 "fmla v29.4s, v2.4s, v5.4s\n" 203 "ldr q13, [x16, x12]\n" 204 "add x16, x16, #0x10\n" 205 "fmla v30.4s, v2.4s, v12.4s\n" 206 "fmla v31.4s, v2.4s, v9.4s\n" 207 "ldr q2, [x17, #0x70]\n" 208 "fmla v28.4s, v3.4s, v5.4s\n" 209 "fmla v29.4s, v3.4s, v6.4s\n" 210 "ld1 { v5.4s }, [x15]\n" 211 "fmla v30.4s, v3.4s, v9.4s\n" 212 "fmla v31.4s, v3.4s, v13.4s\n" 213 "ldr q3, [x17, #0x80]\n" 214 "fmla v28.4s, v4.4s, v6.4s\n" 215 "fmla v29.4s, v4.4s, v10.4s\n" 216 "ldr q6, [x15, x3]\n" 217 "ldr q10, [x15, x7]\n" 218 "fmla v30.4s, v4.4s, v13.4s\n" 219 "fmla v31.4s, v4.4s, v8.4s\n" 220 "ldr q4, [x17, #0x90]\n" 221 "fmla v28.4s, v0.4s, v14.4s\n" 222 "fmla v29.4s, v0.4s, v11.4s\n" 223 "ldr q14, [x15, x10]\n" 224 "fmla v30.4s, v0.4s, v5.4s\n" 225 "fmla v31.4s, v0.4s, v6.4s\n" 226 "ldr q0, [x17, #0xa0]\n" 227 "fmla v28.4s, v1.4s, v11.4s\n" 228 "fmla v29.4s, v1.4s, v12.4s\n" 229 "ldr q11, [x15, x14]\n" 230 "fmla v30.4s, v1.4s, v6.4s\n" 231 "fmla v31.4s, v1.4s, v10.4s\n" 232 "ldr q1, [x17, #0xb0]\n" 233 "fmla v28.4s, v2.4s, v12.4s\n" 234 "fmla v29.4s, v2.4s, v9.4s\n" 235 "ldr q12, [x15, x12]\n" 236 "add x15, x15, #0x10\n" 237 "fmla v30.4s, v2.4s, v10.4s\n" 238 "fmla v31.4s, v2.4s, v11.4s\n" 239 "ldr q2, [x17, #0xc0]\n" 240 "fmla v28.4s, v3.4s, v9.4s\n" 241 "fmla v29.4s, v3.4s, v13.4s\n" 242 "ld1 { v9.4s }, [x13]\n" 243 "fmla v30.4s, v3.4s, v11.4s\n" 244 "fmla v31.4s, v3.4s, v12.4s\n" 245 "ldr q3, [x17, #0xd0]\n" 246 "fmla v28.4s, v4.4s, v13.4s\n" 247 "fmla v29.4s, v4.4s, v8.4s\n" 248 "ldr q13, [x13, x3]\n" 249 "ldr q8, [x13, x12]\n" 250 "fmla v30.4s, v4.4s, v12.4s\n" 251 "fmla v31.4s, v4.4s, v14.4s\n" 252 "ldr q4, [x17, #0xe0]\n" 253 "fmla v28.4s, v0.4s, v5.4s\n" 254 "fmla v29.4s, v0.4s, v6.4s\n" 255 "ldr q5, [x13, x7]\n" 256 "fmla v30.4s, v0.4s, v9.4s\n" 257 "fmla v31.4s, v0.4s, v13.4s\n" 258 "ldr q0, [x17, #0xf0]\n" 259 "fmla v28.4s, v1.4s, v6.4s\n" 260 "fmla v29.4s, v1.4s, v10.4s\n" 261 "ldr q6, [x13, x14]\n" 262 "fmla v30.4s, v1.4s, v13.4s\n" 263 "fmla v31.4s, v1.4s, v5.4s\n" 264 "ldr q1, [x17, #0x100]\n" 265 "fmla v28.4s, v2.4s, v10.4s\n" 266 "fmla v29.4s, v2.4s, v11.4s\n" 267 "ldr q10, [x13, x10]\n" 268 "add x13, x13, #0x10\n" 269 "fmla v30.4s, v2.4s, v5.4s\n" 270 "fmla v31.4s, v2.4s, v6.4s\n" 271 "ldr q2, [x17, #0x110]\n" 272 "fmla v28.4s, v3.4s, v11.4s\n" 273 "fmla v29.4s, v3.4s, v12.4s\n" 274 "ld1 { v11.4s }, [x11]\n" 275 "fmla v30.4s, v3.4s, v6.4s\n" 276 "fmla v31.4s, v3.4s, v8.4s\n" 277 "ldr q3, [x17, #0x120]\n" 278 "fmla v28.4s, v4.4s, v12.4s\n" 279 "fmla v29.4s, v4.4s, v14.4s\n" 280 "ldr q12, [x11, x3]\n" 281 "ld1 { v14.4s }, [x16]\n" 282 "fmla v30.4s, v4.4s, v8.4s\n" 283 "fmla v31.4s, v4.4s, v10.4s\n" 284 "ldr q4, [x17, #0x130]\n" 285 "fmla v28.4s, v0.4s, v9.4s\n" 286 "fmla v29.4s, v0.4s, v13.4s\n" 287 "ldr q9, [x11, x7]\n" 288 "fmla v30.4s, v0.4s, v11.4s\n" 289 "fmla v31.4s, v0.4s, v12.4s\n" 290 "ldr q11, [x11, x14]\n" 291 "ldr q0, [x17, #0x150]\n" 292 "fmla v28.4s, v1.4s, v13.4s\n" 293 "fmla v29.4s, v1.4s, v5.4s\n" 294 "ldr q13, [x8, x7]\n" 295 "fmla v30.4s, v1.4s, v12.4s\n" 296 "fmla v31.4s, v1.4s, v9.4s\n" 297 "ldr q12, [x11, x12]\n" 298 "ldr q1, [x17, #0x160]\n" 299 "fmla v28.4s, v2.4s, v5.4s\n" 300 "fmla v29.4s, v2.4s, v6.4s\n" 301 "ld1 { v5.4s }, [x5]\n" 302 "fmla v30.4s, v2.4s, v9.4s\n" 303 "fmla v31.4s, v2.4s, v11.4s\n" 304 "ldr q9, [x11, x10]\n" 305 "add x11, x11, #0x10\n" 306 "fmla v28.4s, v3.4s, v6.4s\n" 307 "fmla v29.4s, v3.4s, v8.4s\n" 309 "ldr q2, [x17, #0x170]\n" 310 "fmla v30.4s, v3.4s, v11.4s\n" 311 "fmla v31.4s, v3.4s, v12.4s\n" 312 "ldr q11, [x5, x14]\n" 313 "ldr q3, [x17, #0x180]\n" 314 "fmla v28.4s, v4.4s, v8.4s\n" 315 "fmla v29.4s, v4.4s, v10.4s\n" 316 "fmax v28.4s, v28.4s, v18.4s\n" 318 "fmla v30.4s, v4.4s, v12.4s\n" 319 "fmla v31.4s, v4.4s, v9.4s\n" 320 "fmax v29.4s, v29.4s, v18.4s\n" 322 "fmax v30.4s, v30.4s, v18.4s\n" 323 "fmax v31.4s, v31.4s, v18.4s\n" 324 "ldr q12, [x5, x12]\n" 325 "ldr q10, [x8, x10]\n" 326 "fmin v28.4s, v28.4s, v17.4s\n" 327 "fmin v29.4s, v29.4s, v17.4s\n" 328 "st1 { v28.4s }, [x6]\n" 329 "ldr q4, [x17, #0x190]\n" 330 "fmin v30.4s, v30.4s, v17.4s\n" 331 "fmin v31.4s, v31.4s, v17.4s\n" 332 "str q29, [x6, x4]\n" 333 "add x6, x6, #0x10\n" 334 "st1 { v30.4s }, [x9]\n" 335 "add x17, x17, #0x1a0\n" 336 "str q31, [x9, x4]\n" 337 "add x9, x9, #0x10\n" 340 "mov v28.16b, v16.16b\n fmla v28.4s, v0.4s, v5.4s\n" 341 "mov v29.16b, v16.16b\n fmla v29.4s, v0.4s, v6.4s\n" 342 "ldr q5, [x8, x14]\n" 343 "mov v30.16b, v16.16b\n fmla v30.4s, v0.4s, v7.4s\n" 344 "mov v31.16b, v16.16b\n fmla v31.4s, v0.4s, v8.4s\n" 345 "ldr q0, [x17, #0x0]\n" 346 "fmla v28.4s, v1.4s, v6.4s\n" 347 "fmla v29.4s, v1.4s, v9.4s\n" 348 "ldr q6, [x8, x12]\n" 349 "add x8, x8, #0x10\n" 350 "fmla v30.4s, v1.4s, v8.4s\n" 351 "fmla v31.4s, v1.4s, v13.4s\n" 352 "ldr q1, [x17, #0x10]\n" 353 "fmla v28.4s, v2.4s, v9.4s\n" 354 "fmla v29.4s, v2.4s, v11.4s\n" 355 "ldr q9, [x5, x10]\n" 356 "add x5, x5, #0x10\n" 357 "fmla v30.4s, v2.4s, v13.4s\n" 358 "fmla v31.4s, v2.4s, v5.4s\n" 359 "ldr q2, [x17, #0x20]\n" 360 "fmla v28.4s, v3.4s, v11.4s\n" 361 "fmla v29.4s, v3.4s, v12.4s\n" 362 "ldr q11, [x16, x3]\n" 363 "fmla v30.4s, v3.4s, v5.4s\n" 364 "fmla v31.4s, v3.4s, v6.4s\n" 365 "ldr q3, [x17, #0x30]\n" 366 "fmla v28.4s, v4.4s, v12.4s\n" 367 "fmla v29.4s, v4.4s, v9.4s\n" 368 "ldr q12, [x16, x7]\n" 369 "ldr q9, [x16, x14]\n" 370 "fmla v30.4s, v4.4s, v6.4s\n" 371 "fmla v31.4s, v4.4s, v10.4s\n" 372 "ldr q4, [x17, #0x40]\n" 373 "fmla v28.4s, v0.4s, v7.4s\n" 374 "fmla v29.4s, v0.4s, v8.4s\n" 375 "fmla v30.4s, v0.4s, v14.4s\n" 376 "fmla v31.4s, v0.4s, v11.4s\n" 377 "ldr q0, [x17, #0x50]\n" 378 "fmla v28.4s, v1.4s, v8.4s\n" 379 "fmla v29.4s, v1.4s, v13.4s\n" 380 "ldr q8, [x16, x10]\n" 381 "fmla v30.4s, v1.4s, v11.4s\n" 382 "fmla v31.4s, v1.4s, v12.4s\n" 383 "ldr q1, [x17, #0x60]\n" 384 "fmla v28.4s, v2.4s, v13.4s\n" 385 "fmla v29.4s, v2.4s, v5.4s\n" 386 "ldr q13, [x16, x12]\n" 387 "add x16, x16, #0x10\n" 388 "fmla v30.4s, v2.4s, v12.4s\n" 389 "fmla v31.4s, v2.4s, v9.4s\n" 390 "ldr q2, [x17, #0x70]\n" 391 "fmla v28.4s, v3.4s, v5.4s\n" 392 "fmla v29.4s, v3.4s, v6.4s\n" 393 "ld1 { v5.4s }, [x15]\n" 394 "fmla v30.4s, v3.4s, v9.4s\n" 395 "fmla v31.4s, v3.4s, v13.4s\n" 396 "ldr q3, [x17, #0x80]\n" 397 "fmla v28.4s, v4.4s, v6.4s\n" 398 "fmla v29.4s, v4.4s, v10.4s\n" 399 "ldr q6, [x15, x3]\n" 400 "ldr q10, [x15, x7]\n" 401 "fmla v30.4s, v4.4s, v13.4s\n" 402 "fmla v31.4s, v4.4s, v8.4s\n" 403 "ldr q4, [x17, #0x90]\n" 404 "fmla v28.4s, v0.4s, v14.4s\n" 405 "fmla v29.4s, v0.4s, v11.4s\n" 406 "ldr q14, [x15, x10]\n" 407 "fmla v30.4s, v0.4s, v5.4s\n" 408 "fmla v31.4s, v0.4s, v6.4s\n" 409 "ldr q0, [x17, #0xa0]\n" 410 "fmla v28.4s, v1.4s, v11.4s\n" 411 "fmla v29.4s, v1.4s, v12.4s\n" 412 "ldr q11, [x15, x14]\n" 413 "fmla v30.4s, v1.4s, v6.4s\n" 414 "fmla v31.4s, v1.4s, v10.4s\n" 415 "ldr q1, [x17, #0xb0]\n" 416 "fmla v28.4s, v2.4s, v12.4s\n" 417 "fmla v29.4s, v2.4s, v9.4s\n" 418 "ldr q12, [x15, x12]\n" 419 "add x15, x15, #0x10\n" 420 "fmla v30.4s, v2.4s, v10.4s\n" 421 "fmla v31.4s, v2.4s, v11.4s\n" 422 "ldr q2, [x17, #0xc0]\n" 423 "fmla v28.4s, v3.4s, v9.4s\n" 424 "fmla v29.4s, v3.4s, v13.4s\n" 425 "ld1 { v9.4s }, [x13]\n" 426 "fmla v30.4s, v3.4s, v11.4s\n" 427 "fmla v31.4s, v3.4s, v12.4s\n" 428 "ldr q3, [x17, #0xd0]\n" 429 "fmla v28.4s, v4.4s, v13.4s\n" 430 "fmla v29.4s, v4.4s, v8.4s\n" 431 "ldr q13, [x13, x3]\n" 432 "ldr q8, [x13, x12]\n" 433 "fmla v30.4s, v4.4s, v12.4s\n" 434 "fmla v31.4s, v4.4s, v14.4s\n" 435 "ldr q4, [x17, #0xe0]\n" 436 "fmla v28.4s, v0.4s, v5.4s\n" 437 "fmla v29.4s, v0.4s, v6.4s\n" 438 "ldr q5, [x13, x7]\n" 439 "fmla v30.4s, v0.4s, v9.4s\n" 440 "fmla v31.4s, v0.4s, v13.4s\n" 441 "ldr q0, [x17, #0xf0]\n" 442 "fmla v28.4s, v1.4s, v6.4s\n" 443 "fmla v29.4s, v1.4s, v10.4s\n" 444 "ldr q6, [x13, x14]\n" 445 "fmla v30.4s, v1.4s, v13.4s\n" 446 "fmla v31.4s, v1.4s, v5.4s\n" 447 "ldr q1, [x17, #0x100]\n" 448 "fmla v28.4s, v2.4s, v10.4s\n" 449 "fmla v29.4s, v2.4s, v11.4s\n" 450 "ldr q10, [x13, x10]\n" 451 "add x13, x13, #0x10\n" 452 "fmla v30.4s, v2.4s, v5.4s\n" 453 "fmla v31.4s, v2.4s, v6.4s\n" 454 "ldr q2, [x17, #0x110]\n" 455 "fmla v28.4s, v3.4s, v11.4s\n" 456 "fmla v29.4s, v3.4s, v12.4s\n" 457 "ld1 { v11.4s }, [x11]\n" 458 "fmla v30.4s, v3.4s, v6.4s\n" 459 "fmla v31.4s, v3.4s, v8.4s\n" 460 "ldr q3, [x17, #0x120]\n" 461 "fmla v28.4s, v4.4s, v12.4s\n" 462 "fmla v29.4s, v4.4s, v14.4s\n" 463 "ldr q12, [x11, x3]\n" 464 "fmla v30.4s, v4.4s, v8.4s\n" 465 "fmla v31.4s, v4.4s, v10.4s\n" 466 "ldr q4, [x17, #0x130]\n" 467 "add x17, x17, #0x140\n" 468 "fmla v28.4s, v0.4s, v9.4s\n" 469 "fmla v29.4s, v0.4s, v13.4s\n" 470 "ldr q9, [x11, x7]\n" 471 "fmla v30.4s, v0.4s, v11.4s\n" 472 "fmla v31.4s, v0.4s, v12.4s\n" 473 "ldr q11, [x11, x14]\n" 474 "fmla v28.4s, v1.4s, v13.4s\n" 475 "fmla v29.4s, v1.4s, v5.4s\n" 476 "fmla v30.4s, v1.4s, v12.4s\n" 477 "fmla v31.4s, v1.4s, v9.4s\n" 478 "ldr q12, [x11, x12]\n" 479 "fmla v28.4s, v2.4s, v5.4s\n" 480 "fmla v29.4s, v2.4s, v6.4s\n" 481 "fmla v30.4s, v2.4s, v9.4s\n" 482 "fmla v31.4s, v2.4s, v11.4s\n" 483 "ldr q9, [x11, x10]\n" 484 "add x11, x11, #0x10\n" 485 "fmla v28.4s, v3.4s, v6.4s\n" 486 "fmla v29.4s, v3.4s, v8.4s\n" 487 "fmla v30.4s, v3.4s, v11.4s\n" 488 "fmla v31.4s, v3.4s, v12.4s\n" 489 "fmla v28.4s, v4.4s, v8.4s\n" 490 "fmla v29.4s, v4.4s, v10.4s\n" 491 "fmax v28.4s, v28.4s, v18.4s\n" 492 "fmla v30.4s, v4.4s, v12.4s\n" 493 "fmla v31.4s, v4.4s, v9.4s\n" 494 "fmax v29.4s, v29.4s, v18.4s\n" 495 "fmax v30.4s, v30.4s, v18.4s\n" 496 "fmax v31.4s, v31.4s, v18.4s\n" 497 "fmin v28.4s, v28.4s, v17.4s\n" 498 "fmin v29.4s, v29.4s, v17.4s\n" 499 "st1 { v28.4s }, [x6]\n" 500 "fmin v30.4s, v30.4s, v17.4s\n" 501 "fmin v31.4s, v31.4s, v17.4s\n" 502 "str q29, [x6, x4]\n" 503 "add x6, x6, #0x10\n" 504 "st1 { v30.4s }, [x9]\n" 505 "str q31, [x9, x4]\n" 506 "add x9, x9, #0x10\n" 508 "tst %x[n_channels], #0x3\n" 510 "ldr q16, [x17, #0x0]\n" 511 "ldr q0, [x17, #0x10]\n" 512 "ldr q1, [x17, #0x20]\n" 513 "ldr q2, [x17, #0x30]\n" 516 "ldr q3, [x17, #0x40]\n" 517 "ldr q4, [x17, #0x50]\n" 525 "add x19, x16, XZR\n" 526 "add x17, x17, #0x60\n" 527 "tbz %x[n_channels], #1, 5f\n" 528 "ldr d5, [x28], #0x8\n" 529 "ldr d6, [x27], #0x8\n" 530 "ldr d7, [x26], #0x8\n" 531 "ldr d8, [x25], #0x8\n" 532 "ldr d9, [x24], #0x8\n" 533 "ldr d13, [x23], #0x8\n" 534 "ldr d11, [x22], #0x8\n" 535 "ldr d12, [x21], #0x8\n" 536 "ldr d10, [x20], #0x8\n" 537 "ldr d14, [x19], #0x8\n" 538 "tbz %x[n_channels], #0, 6f\n" 539 "ld1 { v5.s }[2], [x28]\n" 540 "ld1 { v6.s }[2], [x27]\n" 541 "ld1 { v7.s }[2], [x26]\n" 542 "ld1 { v8.s }[2], [x25]\n" 543 "ld1 { v9.s }[2], [x24]\n" 544 "ld1 { v13.s }[2], [x23]\n" 545 "ld1 { v11.s }[2], [x22]\n" 546 "ld1 { v12.s }[2], [x21]\n" 547 "ld1 { v10.s }[2], [x20]\n" 548 "ld1 { v14.s }[2], [x19]\n" 551 "ldr s5, [x28, #0x0]\n" 552 "ldr s6, [x27, #0x0]\n" 553 "ldr s7, [x26, #0x0]\n" 554 "ldr s8, [x25, #0x0]\n" 555 "ldr s9, [x24, #0x0]\n" 556 "ldr s13, [x23, #0x0]\n" 557 "ldr s11, [x22, #0x0]\n" 558 "ldr s12, [x21, #0x0]\n" 559 "ldr s10, [x20, #0x0]\n" 560 "ldr s14, [x19, #0x0]\n" 562 "mov v28.16b, v16.16b\n fmla v28.4s, v0.4s, v5.4s\n" 563 "mov v29.16b, v16.16b\n fmla v29.4s, v0.4s, v6.4s\n" 565 "mov v30.16b, v16.16b\n fmla v30.4s, v0.4s, v7.4s\n" 566 "mov v31.16b, v16.16b\n fmla v31.4s, v0.4s, v8.4s\n" 567 "fmla v28.4s, v1.4s, v6.4s\n" 568 "fmla v29.4s, v1.4s, v9.4s\n" 569 "fmla v30.4s, v1.4s, v8.4s\n" 570 "fmla v31.4s, v1.4s, v13.4s\n" 571 "fmla v28.4s, v2.4s, v9.4s\n" 572 "fmla v29.4s, v2.4s, v11.4s\n" 573 "fmla v30.4s, v2.4s, v13.4s\n" 574 "tbz %x[n_channels], #1, 7f\n" 575 "ldr d5, [x19], #0x8\n" 576 "tbz %x[n_channels], #0, 8f\n" 577 "ld1 { v5.s }[2], [x19]\n" 580 "ldr s5, [x19, #0x0]\n" 582 "fmla v31.4s, v2.4s, v5.4s\n" 583 "fmla v28.4s, v3.4s, v11.4s\n" 585 "fmla v29.4s, v3.4s, v12.4s\n" 586 "fmla v30.4s, v3.4s, v5.4s\n" 587 "tbz %x[n_channels], #1, 9f\n" 588 "ldr d6, [x19], #0x8\n" 589 "tbz %x[n_channels], #0, 10f\n" 590 "ld1 { v6.s }[2], [x19]\n" 593 "ldr s6, [x19, #0x0]\n" 595 "fmla v31.4s, v3.4s, v6.4s\n" 596 "fmla v28.4s, v4.4s, v12.4s\n" 598 "tbz %x[n_channels], #1, 11f\n" 599 "ldr d9, [x19], #0x8\n" 600 "tbz %x[n_channels], #0, 12f\n" 601 "ld1 { v9.s }[2], [x19]\n" 604 "ldr s9, [x19, #0x0]\n" 606 "fmla v29.4s, v4.4s, v9.4s\n" 607 "fmla v30.4s, v4.4s, v6.4s\n" 608 "ldr q0, [x17, #0x0]\n" 610 "fmla v31.4s, v4.4s, v10.4s\n" 611 "fmla v28.4s, v0.4s, v7.4s\n" 612 "add x17, x17, #0x10\n" 613 "fmla v29.4s, v0.4s, v8.4s\n" 614 "fmla v30.4s, v0.4s, v14.4s\n" 615 "tbz %x[n_channels], #1, 13f\n" 616 "ldr d11, [x19], #0x8\n" 617 "tbz %x[n_channels], #0, 14f\n" 618 "ld1 { v11.s }[2], [x19]\n" 621 "ldr s11, [x19, #0x0]\n" 623 "ldr q1, [x17, #0x0]\n" 624 "fmla v31.4s, v0.4s, v11.4s\n" 625 "fmla v28.4s, v1.4s, v8.4s\n" 627 "fmla v29.4s, v1.4s, v13.4s\n" 628 "fmla v30.4s, v1.4s, v11.4s\n" 629 "add x17, x17, #0x10\n" 630 "tbz %x[n_channels], #1, 15f\n" 631 "ldr d12, [x19], #0x8\n" 632 "tbz %x[n_channels], #0, 16f\n" 633 "ld1 { v12.s }[2], [x19]\n" 636 "ldr s12, [x19, #0x0]\n" 638 "ldr q2, [x17, #0x0]\n" 639 "fmla v31.4s, v1.4s, v12.4s\n" 640 "fmla v28.4s, v2.4s, v13.4s\n" 641 "add x19, x16, x14\n" 642 "fmla v29.4s, v2.4s, v5.4s\n" 643 "fmla v30.4s, v2.4s, v12.4s\n" 644 "add x17, x17, #0x10\n" 645 "tbz %x[n_channels], #1, 17f\n" 646 "ldr d9, [x19], #0x8\n" 647 "tbz %x[n_channels], #0, 18f\n" 648 "ld1 { v9.s }[2], [x19]\n" 651 "ldr s9, [x19, #0x0]\n" 653 "ldr q3, [x17, #0x0]\n" 654 "fmla v31.4s, v2.4s, v9.4s\n" 655 "fmla v28.4s, v3.4s, v5.4s\n" 656 "add x19, x16, x12\n" 657 "fmla v29.4s, v3.4s, v6.4s\n" 658 "fmla v30.4s, v3.4s, v9.4s\n" 659 "add x17, x17, #0x10\n" 660 "tbz %x[n_channels], #1, 19f\n" 661 "ldr d13, [x19], #0x8\n" 662 "tbz %x[n_channels], #0, 20f\n" 663 "ld1 { v13.s }[2], [x19]\n" 666 "ldr s13, [x19, #0x0]\n" 668 "ldr q4, [x17, #0x0]\n" 669 "fmla v31.4s, v3.4s, v13.4s\n" 670 "fmla v28.4s, v4.4s, v6.4s\n" 671 "add x19, x16, x10\n" 672 "fmla v29.4s, v4.4s, v10.4s\n" 673 "fmla v30.4s, v4.4s, v13.4s\n" 674 "add x17, x17, #0x10\n" 675 "tbz %x[n_channels], #1, 21f\n" 676 "ldr d8, [x19], #0x8\n" 677 "tbz %x[n_channels], #0, 22f\n" 678 "ld1 { v8.s }[2], [x19]\n" 681 "ldr s8, [x19, #0x0]\n" 683 "ldr q0, [x17, #0x0]\n" 684 "fmla v31.4s, v4.4s, v8.4s\n" 685 "fmla v28.4s, v0.4s, v14.4s\n" 686 "add x19, x15, XZR\n" 687 "fmla v29.4s, v0.4s, v11.4s\n" 688 "add x17, x17, #0x10\n" 689 "tbz %x[n_channels], #1, 23f\n" 690 "ldr d5, [x19], #0x8\n" 691 "tbz %x[n_channels], #0, 24f\n" 692 "ld1 { v5.s }[2], [x19]\n" 695 "ldr s5, [x19, #0x0]\n" 697 "fmla v30.4s, v0.4s, v5.4s\n" 699 "tbz %x[n_channels], #1, 25f\n" 700 "ldr d6, [x19], #0x8\n" 701 "tbz %x[n_channels], #0, 26f\n" 702 "ld1 { v6.s }[2], [x19]\n" 705 "ldr s6, [x19, #0x0]\n" 707 "ldr q1, [x17, #0x0]\n" 708 "fmla v31.4s, v0.4s, v6.4s\n" 709 "fmla v28.4s, v1.4s, v11.4s\n" 711 "fmla v29.4s, v1.4s, v12.4s\n" 712 "fmla v30.4s, v1.4s, v6.4s\n" 713 "add x17, x17, #0x10\n" 714 "tbz %x[n_channels], #1, 27f\n" 715 "ldr d10, [x19], #0x8\n" 716 "tbz %x[n_channels], #0, 28f\n" 717 "ld1 { v10.s }[2], [x19]\n" 720 "ldr s10, [x19, #0x0]\n" 722 "ldr q2, [x17, #0x0]\n" 723 "fmla v31.4s, v1.4s, v10.4s\n" 724 "fmla v28.4s, v2.4s, v12.4s\n" 725 "add x19, x15, x14\n" 726 "fmla v29.4s, v2.4s, v9.4s\n" 727 "fmla v30.4s, v2.4s, v10.4s\n" 728 "add x17, x17, #0x10\n" 729 "tbz %x[n_channels], #1, 29f\n" 730 "ldr d11, [x19], #0x8\n" 731 "tbz %x[n_channels], #0, 30f\n" 732 "ld1 { v11.s }[2], [x19]\n" 735 "ldr s11, [x19, #0x0]\n" 737 "ldr q3, [x17, #0x0]\n" 738 "fmla v31.4s, v2.4s, v11.4s\n" 739 "fmla v28.4s, v3.4s, v9.4s\n" 740 "add x19, x15, x12\n" 741 "fmla v29.4s, v3.4s, v13.4s\n" 742 "fmla v30.4s, v3.4s, v11.4s\n" 743 "add x17, x17, #0x10\n" 744 "tbz %x[n_channels], #1, 31f\n" 745 "ldr d12, [x19], #0x8\n" 746 "tbz %x[n_channels], #0, 32f\n" 747 "ld1 { v12.s }[2], [x19]\n" 750 "ldr s12, [x19, #0x0]\n" 752 "ldr q4, [x17, #0x0]\n" 753 "fmla v31.4s, v3.4s, v12.4s\n" 754 "fmla v28.4s, v4.4s, v13.4s\n" 755 "add x19, x15, x10\n" 756 "fmla v29.4s, v4.4s, v8.4s\n" 757 "fmla v30.4s, v4.4s, v12.4s\n" 758 "add x17, x17, #0x10\n" 759 "tbz %x[n_channels], #1, 33f\n" 760 "ldr d14, [x19], #0x8\n" 761 "tbz %x[n_channels], #0, 34f\n" 762 "ld1 { v14.s }[2], [x19]\n" 765 "ldr s14, [x19, #0x0]\n" 767 "ldr q0, [x17, #0x0]\n" 768 "fmla v31.4s, v4.4s, v14.4s\n" 769 "fmla v28.4s, v0.4s, v5.4s\n" 770 "add x19, x13, XZR\n" 771 "fmla v29.4s, v0.4s, v6.4s\n" 772 "add x17, x17, #0x10\n" 773 "tbz %x[n_channels], #1, 35f\n" 774 "ldr d9, [x19], #0x8\n" 775 "tbz %x[n_channels], #0, 36f\n" 776 "ld1 { v9.s }[2], [x19]\n" 779 "ldr s9, [x19, #0x0]\n" 781 "fmla v30.4s, v0.4s, v9.4s\n" 783 "tbz %x[n_channels], #1, 37f\n" 784 "ldr d13, [x19], #0x8\n" 785 "tbz %x[n_channels], #0, 38f\n" 786 "ld1 { v13.s }[2], [x19]\n" 789 "ldr s13, [x19, #0x0]\n" 791 "ldr q1, [x17, #0x0]\n" 792 "fmla v31.4s, v0.4s, v13.4s\n" 793 "fmla v28.4s, v1.4s, v6.4s\n" 795 "fmla v29.4s, v1.4s, v10.4s\n" 796 "fmla v30.4s, v1.4s, v13.4s\n" 797 "add x17, x17, #0x10\n" 798 "tbz %x[n_channels], #1, 39f\n" 799 "ldr d5, [x19], #0x8\n" 800 "tbz %x[n_channels], #0, 40f\n" 801 "ld1 { v5.s }[2], [x19]\n" 804 "ldr s5, [x19, #0x0]\n" 806 "ldr q2, [x17, #0x0]\n" 807 "fmla v31.4s, v1.4s, v5.4s\n" 808 "fmla v28.4s, v2.4s, v10.4s\n" 809 "add x19, x13, x14\n" 810 "fmla v29.4s, v2.4s, v11.4s\n" 811 "fmla v30.4s, v2.4s, v5.4s\n" 812 "add x17, x17, #0x10\n" 813 "tbz %x[n_channels], #1, 41f\n" 814 "ldr d6, [x19], #0x8\n" 815 "tbz %x[n_channels], #0, 42f\n" 816 "ld1 { v6.s }[2], [x19]\n" 819 "ldr s6, [x19, #0x0]\n" 821 "ldr q3, [x17, #0x0]\n" 822 "fmla v31.4s, v2.4s, v6.4s\n" 823 "fmla v28.4s, v3.4s, v11.4s\n" 824 "add x19, x13, x12\n" 825 "fmla v29.4s, v3.4s, v12.4s\n" 826 "fmla v30.4s, v3.4s, v6.4s\n" 827 "add x17, x17, #0x10\n" 828 "tbz %x[n_channels], #1, 43f\n" 829 "ldr d8, [x19], #0x8\n" 830 "tbz %x[n_channels], #0, 44f\n" 831 "ld1 { v8.s }[2], [x19]\n" 834 "ldr s8, [x19, #0x0]\n" 836 "ldr q4, [x17, #0x0]\n" 837 "fmla v31.4s, v3.4s, v8.4s\n" 838 "fmla v28.4s, v4.4s, v12.4s\n" 839 "add x19, x13, x10\n" 840 "fmla v29.4s, v4.4s, v14.4s\n" 841 "fmla v30.4s, v4.4s, v8.4s\n" 842 "add x17, x17, #0x10\n" 843 "tbz %x[n_channels], #1, 45f\n" 844 "ldr d10, [x19], #0x8\n" 845 "tbz %x[n_channels], #0, 46f\n" 846 "ld1 { v10.s }[2], [x19]\n" 849 "ldr s10, [x19, #0x0]\n" 851 "ldr q0, [x17, #0x0]\n" 852 "fmla v31.4s, v4.4s, v10.4s\n" 853 "fmla v28.4s, v0.4s, v9.4s\n" 854 "add x19, x11, XZR\n" 855 "fmla v29.4s, v0.4s, v13.4s\n" 856 "add x17, x17, #0x10\n" 857 "tbz %x[n_channels], #1, 47f\n" 858 "ldr d11, [x19], #0x8\n" 859 "tbz %x[n_channels], #0, 48f\n" 860 "ld1 { v11.s }[2], [x19]\n" 863 "ldr s11, [x19, #0x0]\n" 865 "fmla v30.4s, v0.4s, v11.4s\n" 867 "tbz %x[n_channels], #1, 49f\n" 868 "ldr d12, [x19], #0x8\n" 869 "tbz %x[n_channels], #0, 50f\n" 870 "ld1 { v12.s }[2], [x19]\n" 873 "ldr s12, [x19, #0x0]\n" 875 "ldr q1, [x17, #0x0]\n" 876 "fmla v31.4s, v0.4s, v12.4s\n" 877 "fmla v28.4s, v1.4s, v13.4s\n" 879 "fmla v29.4s, v1.4s, v5.4s\n" 880 "fmla v30.4s, v1.4s, v12.4s\n" 881 "add x17, x17, #0x10\n" 882 "tbz %x[n_channels], #1, 51f\n" 883 "ldr d9, [x19], #0x8\n" 884 "tbz %x[n_channels], #0, 52f\n" 885 "ld1 { v9.s }[2], [x19]\n" 888 "ldr s9, [x19, #0x0]\n" 890 "ldr q2, [x17, #0x0]\n" 891 "fmla v31.4s, v1.4s, v9.4s\n" 892 "fmla v28.4s, v2.4s, v5.4s\n" 893 "add x19, x11, x14\n" 894 "fmla v29.4s, v2.4s, v6.4s\n" 895 "fmla v30.4s, v2.4s, v9.4s\n" 896 "add x17, x17, #0x10\n" 897 "tbz %x[n_channels], #1, 53f\n" 898 "ldr d11, [x19], #0x8\n" 899 "tbz %x[n_channels], #0, 54f\n" 900 "ld1 { v11.s }[2], [x19]\n" 903 "ldr s11, [x19, #0x0]\n" 905 "ldr q3, [x17, #0x0]\n" 906 "fmla v31.4s, v2.4s, v11.4s\n" 907 "fmla v28.4s, v3.4s, v6.4s\n" 908 "add x19, x11, x12\n" 909 "fmla v29.4s, v3.4s, v8.4s\n" 910 "fmla v30.4s, v3.4s, v11.4s\n" 911 "add x17, x17, #0x10\n" 912 "tbz %x[n_channels], #1, 55f\n" 913 "ldr d12, [x19], #0x8\n" 914 "tbz %x[n_channels], #0, 56f\n" 915 "ld1 { v12.s }[2], [x19]\n" 918 "ldr s12, [x19, #0x0]\n" 920 "ldr q4, [x17, #0x0]\n" 921 "fmla v31.4s, v3.4s, v12.4s\n" 922 "fmla v28.4s, v4.4s, v8.4s\n" 923 "add x19, x11, x10\n" 924 "fmla v29.4s, v4.4s, v10.4s\n" 925 "fmla v30.4s, v4.4s, v12.4s\n" 926 "tbz %x[n_channels], #1, 57f\n" 927 "ldr d9, [x19], #0x8\n" 928 "tbz %x[n_channels], #0, 58f\n" 929 "ld1 { v9.s }[2], [x19]\n" 932 "ldr s9, [x19, #0x0]\n" 934 "fmla v31.4s, v4.4s, v9.4s\n" 935 "fmax v28.4s, v28.4s, v18.4s\n" 936 "fmax v29.4s, v29.4s, v18.4s\n" 937 "fmax v30.4s, v30.4s, v18.4s\n" 938 "fmax v31.4s, v31.4s, v18.4s\n" 939 "fmin v28.4s, v28.4s, v17.4s\n" 940 "fmin v29.4s, v29.4s, v17.4s\n" 941 "fmin v30.4s, v30.4s, v17.4s\n" 942 "fmin v31.4s, v31.4s, v17.4s\n" 943 "tbz %x[n_channels], #1, 59f\n" 946 "st1 { v28.d }[0], [x20], x4\n" 949 "st1 { v30.d }[0], [x19], x4\n" 950 "st1 { v29.d }[0], [x20]\n" 951 "st1 { v31.d }[0], [x19]\n" 952 "tbz %x[n_channels], #0, 60f\n" 955 "st1 { v28.s }[2], [x20], x4\n" 956 "st1 { v30.s }[2], [x19], x4\n" 957 "st1 { v29.s }[2], [x20]\n" 958 "st1 { v31.s }[2], [x19]\n" 963 "st1 { v28.s }[0], [x20], x4\n" 964 "st1 { v30.s }[0], [x19], x4\n" 965 "st1 { v29.s }[0], [x20]\n" 966 "st1 { v31.s }[0], [x19]\n" 970 "ldr x25, [%x[params_struct], %[offsetof_args_tile_j]]\n" 971 "ldr x26, [%x[params_struct], %[offsetof_args_tile_i]]\n" 972 "add x25, x25, #0x1\n" 973 "add x20, x26, #0x1\n" 974 "ldr x19, [%x[params_struct], %[offsetof_args_n_tile_cols]]\n" 976 "ldr x19, [%x[params_struct], %[offsetof_args_n_tile_rows]]\n" 977 "csel x26, x26, x20, LT\n" 978 "csel x25, x25, XZR, LT\n" 982 : [n_channels]
"r" ((
unsigned long) n_channels), [offsetof_args_inptr]
"I" (offsetof(Args, inptr)), [offsetof_args_ld_input_col]
"I" (offsetof(Args, ld_input_col)), [offsetof_args_ld_input_row]
"I" (offsetof(Args, ld_input_row)), [offsetof_args_ld_output_col]
"I" (offsetof(Args, ld_output_col)), [offsetof_args_ld_output_row]
"I" (offsetof(Args, ld_output_row)), [offsetof_args_max]
"I" (offsetof(Args, max)), [offsetof_args_min]
"I" (offsetof(Args, min)), [offsetof_args_n_tile_cols]
"I" (offsetof(Args, n_tile_cols)), [offsetof_args_n_tile_rows]
"I" (offsetof(Args, n_tile_rows)), [offsetof_args_outptr]
"I" (offsetof(Args, outptr)), [offsetof_args_params]
"I" (offsetof(Args, params)), [offsetof_args_tile_i]
"I" (offsetof(Args, tile_i)), [offsetof_args_tile_j]
"I" (offsetof(Args, tile_j)), [params_struct]
"r" (¶ms_struct)
983 :
"cc",
"memory",
"v0",
"v1",
"v2",
"v3",
"v4",
"v5",
"v6",
"v7",
"v8",
"v9",
"v10",
"v11",
"v12",
"v13",
"v14",
"v16",
"v17",
"v18",
"v28",
"v29",
"v30",
"v31",
"x3",
"x4",
"x5",
"x6",
"x7",
"x8",
"x9",
"x10",
"x11",
"x12",
"x13",
"x14",
"x15",
"x16",
"x17",
"x19",
"x20",
"x21",
"x22",
"x23",
"x24",
"x25",
"x26",
"x27",
"x28"
template UniqueDepthwiseCommon< float > depthwise(const DepthwiseArgs &, const Nothing &)
void a64_fp32_nhwc_5x5_s1_output2x2_mla_depthfirst_direct_impl(const unsigned int n_tile_rows, const unsigned int n_tile_cols, const float *inptr, int64_t ld_input_row, int64_t ld_input_col, float *outptr, int64_t ld_output_row, int64_t ld_output_col, const void *params, unsigned int n_channels, const float activation_min, const float activation_max)