File size: 18,805 Bytes
8b7c501 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 |
// Copyright 2019 Google LLC
//
// This source code is licensed under the BSD-style license found in the
// LICENSE file in the root directory of this source tree.
#include <xnnpack/assembly.h>
# LINT.IfChange
# void xnn_f32_gemm${"inc" if INC else ""}_minmax_ukernel_4x8__asm_aarch64_neonfma_cortex_a75${"_prfm" if PREFETCH else ""}(
# size_t mr, x0
# size_t nc, x1
# size_t kc, x2 / x0
# const float* a, x3
# size_t a_stride, x4
# const float* w, x5
# float* c, x6
# size_t cm_stride, x7
# size_t cn_stride, [sp] -> x14
$if INC:
# const float* acc, [sp + 8] -> x15
# const xnn_f32_minmax_params* params) [sp + 16] -> x8
$else:
# const xnn_f32_minmax_params* params) [sp + 8] -> x8
# d8-d15, x19-x30 need to be preserved if used. x18 is reserved by the OS.
# Vector register usage
# A0 x3 v0 v4
# A1 x11 v1 v5
# A2 x12 v2 v6
# A3 x4 v3 v7
# B x5 v8 v9 v10 v11
# B v12 v13 v14 v15
# B v16 v17 v18 v19
# B v20 v21 v22 v23
# C x6 v24 v25
# C x9 v26 v27
# C x10 v28 v29
# C x7 v30 v31
# Clamp v4 v5
BEGIN_FUNCTION xnn_f32_gemm${"inc" if INC else ""}_minmax_ukernel_4x8__asm_aarch64_neonfma_cortex_a75${"_prfm" if PREFETCH else ""}
$if INC:
# Load cn_stride, acc
LDP x14, x15, [sp]
# Load params pointer
LDR x8, [sp, 16]
$else:
# Load cn_stride, params pointer
LDP x14, x8, [sp]
# Load min/max values
LD2R {v4.4s, v5.4s}, [x8]
# Save d8-d15 on stack
STP d8, d9, [sp, -64]!
STP d10, d11, [sp, 16]
STP d12, d13, [sp, 32]
STP d14, d15, [sp, 48]
# Clamp A and C pointers
CMP x0, 2 // if mr < 2
ADD x11, x3, x4 // a1 = a0 + a_stride
ADD x9, x6, x7 // c1 = c0 + cm_stride
CSEL x11, x3, x11, LO // a1 = a0
CSEL x9, x6, x9, LO // c1 = c0
ADD x12, x11, x4 // a2 = a1 + a_stride
ADD x10, x9, x7 // c2 = c1 + cm_stride
// if mr <= 2
CSEL x12, x11, x12, LS // a2 = a1
CSEL x10, x9, x10, LS // c2 = c1
CMP x0, 4 // if mr < 4
ADD x4, x12, x4 // a3 = a2 + a_stride
ADD x7, x10, x7 // c3 = c2 + cm_stride
CSEL x4, x12, x4, LO // a3 = a2
CSEL x7, x10, x7, LO // c3 = c2
0:
$if INC:
# Load initial accumulators
LDP q24, q25, [x15], 32
LDP q26, q27, [x15], 32
LDP q28, q29, [x15], 32
LDP q30, q31, [x15], 32
$else:
# Load initial bias from w into accumulators
LDP q24, q25, [x5], 32
MOV v26.16b, v24.16b
MOV v27.16b, v25.16b
MOV v28.16b, v24.16b
MOV v29.16b, v25.16b
MOV v30.16b, v24.16b
MOV v31.16b, v25.16b
# Is there at least 8 floats (32 bytes) for prologue + epilogue?
SUBS x0, x2, 32 // k = kc - 32
B.LO 3f
# 16 prologue
# Read first block of 4 A and B.
LDR q0, [x3], 16
LDP q16, q17, [x5], 32
LDR q1, [x11], 16
LDR q2, [x12], 16
LDR q3, [x4], 16
LDP q18, q19, [x5], 32
LDP q20, q21, [x5], 32
LDP q22, q23, [x5], 32
# Is there at least 32. yes do main loop
SUBS x0, x0, 32
B.LO 2f
# Main loop - 8 floats of A (32 bytes)
1:
# First block of 4. FMA for first 4, loads for 2nd block of 4.
FMLA v24.4s, v16.4s, v0.s[0]
LDP q8, q9, [x5], 32
FMLA v25.4s, v17.4s, v0.s[0]
FMLA v26.4s, v16.4s, v1.s[0]
LDP q10, q11, [x5], 32
FMLA v27.4s, v17.4s, v1.s[0]
FMLA v28.4s, v16.4s, v2.s[0]
LDP q12, q13, [x5], 32
FMLA v29.4s, v17.4s, v2.s[0]
FMLA v30.4s, v16.4s, v3.s[0]
LDP q14, q15, [x5], 32
FMLA v31.4s, v17.4s, v3.s[0]
FMLA v24.4s, v18.4s, v0.s[1]
LDR q4, [x3], 16
FMLA v25.4s, v19.4s, v0.s[1]
FMLA v26.4s, v18.4s, v1.s[1]
LDR q5, [x11], 16
FMLA v27.4s, v19.4s, v1.s[1]
FMLA v28.4s, v18.4s, v2.s[1]
LDR q6, [x12], 16
FMLA v29.4s, v19.4s, v2.s[1]
FMLA v30.4s, v18.4s, v3.s[1]
LDR q7, [x4], 16
FMLA v31.4s, v19.4s, v3.s[1]
FMLA v24.4s, v20.4s, v0.s[2]
$if PREFETCH:
PRFM PLDL1KEEP, [x5, 128]
FMLA v25.4s, v21.4s, v0.s[2]
FMLA v26.4s, v20.4s, v1.s[2]
$if PREFETCH:
PRFM PLDL1KEEP, [x5, 192]
FMLA v27.4s, v21.4s, v1.s[2]
FMLA v28.4s, v20.4s, v2.s[2]
$if PREFETCH:
PRFM PLDL1KEEP, [x5, 256]
FMLA v29.4s, v21.4s, v2.s[2]
FMLA v30.4s, v20.4s, v3.s[2]
$if PREFETCH:
PRFM PLDL1KEEP, [x5, 320]
FMLA v31.4s, v21.4s, v3.s[2]
FMLA v24.4s, v22.4s, v0.s[3]
FMLA v25.4s, v23.4s, v0.s[3]
FMLA v26.4s, v22.4s, v1.s[3]
FMLA v27.4s, v23.4s, v1.s[3]
FMLA v28.4s, v22.4s, v2.s[3]
FMLA v29.4s, v23.4s, v2.s[3]
FMLA v30.4s, v22.4s, v3.s[3]
FMLA v31.4s, v23.4s, v3.s[3]
# Second block of 4. FMA for second 4, loads for 1st block of 4.
FMLA v24.4s, v8.4s, v4.s[0]
LDP q16, q17, [x5], 32
FMLA v25.4s, v9.4s, v4.s[0]
FMLA v26.4s, v8.4s, v5.s[0]
LDP q18, q19, [x5], 32
FMLA v27.4s, v9.4s, v5.s[0]
FMLA v28.4s, v8.4s, v6.s[0]
LDP q20, q21, [x5], 32
FMLA v29.4s, v9.4s, v6.s[0]
FMLA v30.4s, v8.4s, v7.s[0]
LDP q22, q23, [x5], 32
FMLA v31.4s, v9.4s, v7.s[0]
FMLA v24.4s, v10.4s, v4.s[1]
LDR q0, [x3], 16
FMLA v25.4s, v11.4s, v4.s[1]
FMLA v26.4s, v10.4s, v5.s[1]
LDR q1, [x11], 16
FMLA v27.4s, v11.4s, v5.s[1]
FMLA v28.4s, v10.4s, v6.s[1]
LDR q2, [x12], 16
FMLA v29.4s, v11.4s, v6.s[1]
FMLA v30.4s, v10.4s, v7.s[1]
LDR q3, [x4], 16
FMLA v31.4s, v11.4s, v7.s[1]
FMLA v24.4s, v12.4s, v4.s[2]
FMLA v25.4s, v13.4s, v4.s[2]
FMLA v26.4s, v12.4s, v5.s[2]
FMLA v27.4s, v13.4s, v5.s[2]
FMLA v28.4s, v12.4s, v6.s[2]
FMLA v29.4s, v13.4s, v6.s[2]
FMLA v30.4s, v12.4s, v7.s[2]
FMLA v31.4s, v13.4s, v7.s[2]
FMLA v24.4s, v14.4s, v4.s[3]
FMLA v25.4s, v15.4s, v4.s[3]
FMLA v26.4s, v14.4s, v5.s[3]
FMLA v27.4s, v15.4s, v5.s[3]
FMLA v28.4s, v14.4s, v6.s[3]
FMLA v29.4s, v15.4s, v6.s[3]
SUBS x0, x0, 32
FMLA v30.4s, v14.4s, v7.s[3]
FMLA v31.4s, v15.4s, v7.s[3]
B.HS 1b
2:
# Epilogue
# First block of 4. FMA for first 4, loads for 2nd block of 4.
FMLA v24.4s, v16.4s, v0.s[0]
LDP q8, q9, [x5], 32
FMLA v25.4s, v17.4s, v0.s[0]
FMLA v26.4s, v16.4s, v1.s[0]
LDP q10, q11, [x5], 32
FMLA v27.4s, v17.4s, v1.s[0]
FMLA v28.4s, v16.4s, v2.s[0]
LDP q12, q13, [x5], 32
FMLA v29.4s, v17.4s, v2.s[0]
FMLA v30.4s, v16.4s, v3.s[0]
LDP q14, q15, [x5], 32
FMLA v31.4s, v17.4s, v3.s[0]
FMLA v24.4s, v18.4s, v0.s[1]
LDR q4, [x3], 16
FMLA v25.4s, v19.4s, v0.s[1]
FMLA v26.4s, v18.4s, v1.s[1]
LDR q5, [x11], 16
FMLA v27.4s, v19.4s, v1.s[1]
FMLA v28.4s, v18.4s, v2.s[1]
LDR q6, [x12], 16
FMLA v29.4s, v19.4s, v2.s[1]
FMLA v30.4s, v18.4s, v3.s[1]
LDR q7, [x4], 16
FMLA v31.4s, v19.4s, v3.s[1]
FMLA v24.4s, v20.4s, v0.s[2]
FMLA v25.4s, v21.4s, v0.s[2]
FMLA v26.4s, v20.4s, v1.s[2]
FMLA v27.4s, v21.4s, v1.s[2]
FMLA v28.4s, v20.4s, v2.s[2]
FMLA v29.4s, v21.4s, v2.s[2]
FMLA v30.4s, v20.4s, v3.s[2]
FMLA v31.4s, v21.4s, v3.s[2]
FMLA v24.4s, v22.4s, v0.s[3]
FMLA v25.4s, v23.4s, v0.s[3]
FMLA v26.4s, v22.4s, v1.s[3]
FMLA v27.4s, v23.4s, v1.s[3]
FMLA v28.4s, v22.4s, v2.s[3]
FMLA v29.4s, v23.4s, v2.s[3]
FMLA v30.4s, v22.4s, v3.s[3]
FMLA v31.4s, v23.4s, v3.s[3]
# Second block of 4. FMA for second 4, noloads
FMLA v24.4s, v8.4s, v4.s[0]
FMLA v25.4s, v9.4s, v4.s[0]
FMLA v26.4s, v8.4s, v5.s[0]
FMLA v27.4s, v9.4s, v5.s[0]
FMLA v28.4s, v8.4s, v6.s[0]
FMLA v29.4s, v9.4s, v6.s[0]
FMLA v30.4s, v8.4s, v7.s[0]
FMLA v31.4s, v9.4s, v7.s[0]
FMLA v24.4s, v10.4s, v4.s[1]
FMLA v25.4s, v11.4s, v4.s[1]
FMLA v26.4s, v10.4s, v5.s[1]
FMLA v27.4s, v11.4s, v5.s[1]
FMLA v28.4s, v10.4s, v6.s[1]
FMLA v29.4s, v11.4s, v6.s[1]
FMLA v30.4s, v10.4s, v7.s[1]
FMLA v31.4s, v11.4s, v7.s[1]
FMLA v24.4s, v12.4s, v4.s[2]
FMLA v25.4s, v13.4s, v4.s[2]
FMLA v26.4s, v12.4s, v5.s[2]
FMLA v27.4s, v13.4s, v5.s[2]
FMLA v28.4s, v12.4s, v6.s[2]
FMLA v29.4s, v13.4s, v6.s[2]
FMLA v30.4s, v12.4s, v7.s[2]
FMLA v31.4s, v13.4s, v7.s[2]
FMLA v24.4s, v14.4s, v4.s[3]
FMLA v25.4s, v15.4s, v4.s[3]
FMLA v26.4s, v14.4s, v5.s[3]
FMLA v27.4s, v15.4s, v5.s[3]
# Load min/max values
LD2R {v4.4s, v5.4s}, [x8]
FMLA v28.4s, v14.4s, v6.s[3]
FMLA v29.4s, v15.4s, v6.s[3]
FMLA v30.4s, v14.4s, v7.s[3]
FMLA v31.4s, v15.4s, v7.s[3]
3:
# Remainder- 4 floats of A (16 bytes)
TBZ x0, 4, 4f
LDR q0, [x3], 16
LDP q16, q17, [x5], 32
LDR q1, [x11], 16
LDR q2, [x12], 16
LDR q3, [x4], 16
FMLA v24.4s, v16.4s, v0.s[0]
FMLA v25.4s, v17.4s, v0.s[0]
LDP q18, q19, [x5], 32
FMLA v26.4s, v16.4s, v1.s[0]
FMLA v27.4s, v17.4s, v1.s[0]
LDP q20, q21, [x5], 32
FMLA v28.4s, v16.4s, v2.s[0]
FMLA v29.4s, v17.4s, v2.s[0]
LDP q22, q23, [x5], 32
FMLA v30.4s, v16.4s, v3.s[0]
FMLA v31.4s, v17.4s, v3.s[0]
FMLA v24.4s, v18.4s, v0.s[1]
FMLA v25.4s, v19.4s, v0.s[1]
FMLA v26.4s, v18.4s, v1.s[1]
FMLA v27.4s, v19.4s, v1.s[1]
FMLA v28.4s, v18.4s, v2.s[1]
FMLA v29.4s, v19.4s, v2.s[1]
FMLA v30.4s, v18.4s, v3.s[1]
FMLA v31.4s, v19.4s, v3.s[1]
FMLA v24.4s, v20.4s, v0.s[2]
FMLA v25.4s, v21.4s, v0.s[2]
FMLA v26.4s, v20.4s, v1.s[2]
FMLA v27.4s, v21.4s, v1.s[2]
FMLA v28.4s, v20.4s, v2.s[2]
FMLA v29.4s, v21.4s, v2.s[2]
FMLA v30.4s, v20.4s, v3.s[2]
FMLA v31.4s, v21.4s, v3.s[2]
FMLA v24.4s, v22.4s, v0.s[3]
FMLA v25.4s, v23.4s, v0.s[3]
FMLA v26.4s, v22.4s, v1.s[3]
FMLA v27.4s, v23.4s, v1.s[3]
FMLA v28.4s, v22.4s, v2.s[3]
FMLA v29.4s, v23.4s, v2.s[3]
FMLA v30.4s, v22.4s, v3.s[3]
FMLA v31.4s, v23.4s, v3.s[3]
4:
# Remainder- 2 floats of A (8 bytes)
TBZ x0, 3, 5f
LDR d0, [x3], 8
LDP q16, q17, [x5], 32
LDR d1, [x11], 8
LDR d2, [x12], 8
LDR d3, [x4], 8
FMLA v24.4s, v16.4s, v0.s[0]
FMLA v25.4s, v17.4s, v0.s[0]
LDP q18, q19, [x5], 32
FMLA v26.4s, v16.4s, v1.s[0]
FMLA v27.4s, v17.4s, v1.s[0]
FMLA v28.4s, v16.4s, v2.s[0]
FMLA v29.4s, v17.4s, v2.s[0]
FMLA v30.4s, v16.4s, v3.s[0]
FMLA v31.4s, v17.4s, v3.s[0]
FMLA v24.4s, v18.4s, v0.s[1]
FMLA v25.4s, v19.4s, v0.s[1]
FMLA v26.4s, v18.4s, v1.s[1]
FMLA v27.4s, v19.4s, v1.s[1]
FMLA v28.4s, v18.4s, v2.s[1]
FMLA v29.4s, v19.4s, v2.s[1]
FMLA v30.4s, v18.4s, v3.s[1]
FMLA v31.4s, v19.4s, v3.s[1]
5:
# Remainder- 1 float of A (4 bytes)
TBZ x0, 2, 6f
LDR s0, [x3], 4
LDP q16, q17, [x5], 32
LDR s1, [x11], 4
LDR s2, [x12], 4
LDR s3, [x4], 4
FMLA v24.4s, v16.4s, v0.s[0]
FMLA v25.4s, v17.4s, v0.s[0]
FMLA v26.4s, v16.4s, v1.s[0]
FMLA v27.4s, v17.4s, v1.s[0]
FMLA v28.4s, v16.4s, v2.s[0]
FMLA v29.4s, v17.4s, v2.s[0]
FMLA v30.4s, v16.4s, v3.s[0]
FMLA v31.4s, v17.4s, v3.s[0]
6:
# Clamp
FMAX v24.4s, v24.4s, v4.4s
SUBS x1, x1, 8
FMAX v25.4s, v25.4s, v4.4s
FMAX v26.4s, v26.4s, v4.4s
FMAX v27.4s, v27.4s, v4.4s
FMAX v28.4s, v28.4s, v4.4s
FMAX v29.4s, v29.4s, v4.4s
FMAX v30.4s, v30.4s, v4.4s
FMAX v31.4s, v31.4s, v4.4s
FMIN v24.4s, v24.4s, v5.4s
FMIN v25.4s, v25.4s, v5.4s
FMIN v26.4s, v26.4s, v5.4s
FMIN v27.4s, v27.4s, v5.4s
FMIN v28.4s, v28.4s, v5.4s
FMIN v29.4s, v29.4s, v5.4s
FMIN v30.4s, v30.4s, v5.4s
FMIN v31.4s, v31.4s, v5.4s
# Store full 4 x 8
B.LO 7f
$if INC:
STP q30, q31, [x7]
SUB x3, x3, x2 // a0 -= kc
ADD x7, x7, x14
STP q28, q29, [x10]
SUB x11, x11, x2 // a1 -= kc
ADD x10, x10, x14
STP q26, q27, [x9]
SUB x12, x12, x2 // a2 -= kc
ADD x9, x9, x14
STP q24, q25, [x6]
SUB x4, x4, x2 // a3 -= kc
ADD x6, x6, x14
$else:
STP q24, q25, [x6]
SUB x3, x3, x2 // a0 -= kc
ADD x6, x6, x14
STP q26, q27, [x9]
SUB x11, x11, x2 // a1 -= kc
ADD x9, x9, x14
STP q28, q29, [x10]
SUB x12, x12, x2 // a2 -= kc
ADD x10, x10, x14
STP q30, q31, [x7]
SUB x4, x4, x2 // a3 -= kc
ADD x7, x7, x14
B.HI 0b
# Restore d8-d15 from stack
LDP d14, d15, [sp, 48]
LDP d12, d13, [sp, 32]
LDP d10, d11, [sp, 16]
LDP d8, d9, [sp], 64
RET
# Store odd width
7:
TBZ x1, 2, 8f
$if INC:
STR q30, [x7], 16
MOV v30.16b, v31.16b
STR q28, [x10], 16
MOV v28.16b, v29.16b
STR q26, [x9], 16
MOV v26.16b, v27.16b
STR q24, [x6], 16
MOV v24.16b, v25.16b
$else:
STR q24, [x6], 16
MOV v24.16b, v25.16b
STR q26, [x9], 16
MOV v26.16b, v27.16b
STR q28, [x10], 16
MOV v28.16b, v29.16b
STR q30, [x7], 16
MOV v30.16b, v31.16b
8:
TBZ x1, 1, 9f
$if INC:
STR d30, [x7], 8
STR d28, [x10], 8
DUP d30, v30.d[1]
DUP d28, v28.d[1]
STR d26, [x9], 8
STR d24, [x6], 8
DUP d26, v26.d[1]
DUP d24, v24.d[1]
$else:
STR d24, [x6], 8
STR d26, [x9], 8
DUP d24, v24.d[1]
DUP d26, v26.d[1]
STR d28, [x10], 8
STR d30, [x7], 8
DUP d28, v28.d[1]
DUP d30, v30.d[1]
9:
TBZ x1, 0, 10f
$if INC:
STR s30, [x7]
STR s28, [x10]
STR s26, [x9]
STR s24, [x6]
$else:
STR s24, [x6]
STR s26, [x9]
STR s28, [x10]
STR s30, [x7]
10:
# Restore d8-d15 from stack
LDP d14, d15, [sp, 48]
LDP d12, d13, [sp, 32]
LDP d10, d11, [sp, 16]
LDP d8, d9, [sp], 64
RET
END_FUNCTION xnn_f32_gemm${"inc" if INC else ""}_minmax_ukernel_4x8__asm_aarch64_neonfma_cortex_a75${"_prfm" if PREFETCH else ""}
# LINT.ThenChange(gen/f32-gemm-4x8-aarch64-neonfma-cortex-a75.cc)
#ifdef __ELF__
.section ".note.GNU-stack","",%progbits
#endif
|