File size: 19,804 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 |
// Auto-generated file. Do not edit!
// Template: src/f16-gemm/6x16-aarch64-neonfp16arith-cortex-a55r0.S.in
// Generator: tools/xngen
//
// Copyright 2020 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>
# void xnn_f16_gemm_minmax_ukernel_6x16__asm_aarch64_neonfp16arith_cortex_a55r0(
# size_t mr, x0
# size_t nc, x1
# size_t kc, x2 / x0
# const void* restrict a, x3
# size_t a_stride, x4
# const void* restrict w, x5
# void* restrict c, x6
# size_t cm_stride, x7
# size_t cn_stride, [sp] -> (x0)
# const union xnn_f16_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) [sp + 8] -> (x8)
# d8-d15, x19-x30 need to be preserved if used. x18 is reserved by the OS.
// Register usage
// A0 x3 v0
// A1 x9 v1
// A2 x10 v2
// A3 x11 v3
// A4 x12 v4
// A5 x4 v5
// B x5 v16 v17 v18 v19
// C0 x6 v20 v21
// C1 x16 v22 v23
// C2 x17 v24 v25
// C3 x14 v26 v27
// C4 x13 v28 v29
// C5 x7 v30 v31
// clamp v6, (v4), (v5)
// unused v7
// unused A v8 v9 v10 v11
// unused B v12 v13 v14 v15
// x8 temporary vector shadow register
BEGIN_FUNCTION xnn_f16_gemm_minmax_ukernel_6x16__asm_aarch64_neonfp16arith_cortex_a55r0
# Load params pointer
LDR x8, [sp, 8]
# Clamp A and C pointers
CMP x0, 2 // if mr < 2
ADD x9, x3, x4 // a1 = a0 + a_stride
ADD x16, x6, x7 // c1 = c0 + cm_stride
CSEL x9, x3, x9, LO // a1 = a0
CSEL x16, x6, x16, LO // c1 = c0
# Load params
LDR s6, [x8]
ADD x10, x9, x4 // a2 = a1 + a_stride
ADD x17, x16, x7 // c2 = c1 + cm_stride
// if mr <= 2
CSEL x10, x9, x10, LS // a2 = a1
CSEL x17, x16, x17, LS // c2 = c1
CMP x0, 4 // if mr < 4
ADD x11, x10, x4 // a3 = a2 + a_stride
ADD x14, x17, x7 // c3 = c2 + cm_stride
CSEL x11, x10, x11, LO // a3 = a2
CSEL x14, x17, x14, LO // c3 = c2
ADD x12, x11, x4 // a4 = a3 + a_stride
ADD x13, x14, x7 // c4 = c3 + cm_stride
// if mr <= 4
CSEL x12, x11, x12, LS // a4 = a3
CSEL x13, x14, x13, LS // c4 = c3
CMP x0, 6 // if mr < 6
ADD x4, x12, x4 // a5 = a4 + a_stride
ADD x7, x13, x7 // c5 = c4 + cm_stride
CSEL x4, x12, x4, LO // a5 = a4
CSEL x7, x13, x7, LO // c5 = c4
# Save d12-d15 on stack
STP d12, d13, [sp, -32]!
STP d14, d15, [sp, 16]
0:
# Load initial bias from w into accumulators
LDP q20, q21, [x5], 32
MOV v22.16b, v20.16b
MOV v23.16b, v21.16b
MOV v24.16b, v20.16b
MOV v25.16b, v21.16b
MOV v26.16b, v20.16b
MOV v27.16b, v21.16b
MOV v28.16b, v20.16b
MOV v29.16b, v21.16b
MOV v30.16b, v20.16b
MOV v31.16b, v21.16b
# Is there at least 4 halffloats (8 bytes) for prologue + epilogue?
SUBS x0, x2, 8 // k = kc - 8
B.LO 4f
# Prologue - First group loads, no FMA
LDR s0, [x3], 4 // A0
LDP q16, q17, [x5], 32 // B
LDR s1, [x10], 4 // A2
LDR s2, [x12], 4 // A4
LD1 {v0.s}[2], [x9], 4 // A1
LD1 {v1.s}[2], [x11], 4 // A3
LD1 {v2.s}[2], [x4], 4 // A5
LDR q18, [x5], 16
LDR d19, [x5], 8
LDR x8, [x5], 8 // ins is in BLOCK 0
SUBS x0, x0, 8
# Is there at least 4 halffloats (8 bytes) for main loop?
B.LO 2f
# Main loop - 4 halffloats of A (8 bytes)
# 48 FMA + 12 LD32 A + 8 LDR B
1:
# First group of 24 FMA, Second group loads
# BLOCK 0
LDR s3, [x3], 4 // A0
INS v19.d[1], x8 // B from second group
FMLA v20.8h, v16.8h, v0.h[0]
LDR w8, [x9], 4 // A1
FMLA v22.8h, v16.8h, v0.h[4]
FMLA v24.8h, v16.8h, v1.h[0]
# BLOCK 1
LDR d12, [x5]
INS v3.d[1], x8 // A1 ins
FMLA v26.8h, v16.8h, v1.h[4]
LDR x8, [x5, 8] // B
FMLA v28.8h, v16.8h, v2.h[0]
FMLA v30.8h, v16.8h, v2.h[4]
# BLOCK 2
LDR s4, [x10], 4 // A2
INS v12.d[1], x8 // B ins
FMLA v21.8h, v17.8h, v0.h[0]
LDR w8, [x11], 4 // A3
FMLA v23.8h, v17.8h, v0.h[4]
FMLA v25.8h, v17.8h, v1.h[0]
# BLOCK 3
LDR s5, [x12], 4 // A4
INS v4.d[1], x8 // A3 ins
FMLA v27.8h, v17.8h, v1.h[4]
LDR w8, [x4], 4 // A5
FMLA v29.8h, v17.8h, v2.h[0]
FMLA v31.8h, v17.8h, v2.h[4]
# BLOCK 4
LDR d13, [x5, 16]
INS v5.d[1], x8 // A5 ins
FMLA v20.8h, v18.8h, v0.h[1]
LDR x8, [x5, 24]
FMLA v22.8h, v18.8h, v0.h[5]
FMLA v24.8h, v18.8h, v1.h[1]
# BLOCK 5
LDR d14, [x5, 32]
INS v13.d[1], x8 // B
FMLA v26.8h, v18.8h, v1.h[5]
LDR x8, [x5, 40]
FMLA v28.8h, v18.8h, v2.h[1]
FMLA v30.8h, v18.8h, v2.h[5]
# BLOCK 6
LDR d15, [x5, 48]
INS v14.d[1], x8 // B
FMLA v21.8h, v19.8h, v0.h[1]
LDR x8, [x5, 56]
FMLA v23.8h, v19.8h, v0.h[5]
FMLA v25.8h, v19.8h, v1.h[1]
# BLOCK 7
INS v15.d[1], x8
FMLA v27.8h, v19.8h, v1.h[5]
FMLA v29.8h, v19.8h, v2.h[1]
FMLA v31.8h, v19.8h, v2.h[5]
# Second group of 24 FMA, First group of loads
# BLOCK 0
LDR s0, [x3], 4 // A0
FMLA v20.8h, v12.8h, v3.h[0]
LDR w8, [x9], 4 // A1
FMLA v22.8h, v12.8h, v3.h[4]
FMLA v24.8h, v12.8h, v4.h[0]
# BLOCK 1
LDR d16, [x5, 64]
INS v0.d[1], x8 // A1 ins
FMLA v26.8h, v12.8h, v4.h[4]
LDR x8, [x5, 72] // B
FMLA v28.8h, v12.8h, v5.h[0]
FMLA v30.8h, v12.8h, v5.h[4]
# BLOCK 2
LDR s1, [x10], 4 // A2
INS v16.d[1], x8 // B
FMLA v21.8h, v13.8h, v3.h[0]
LDR w8, [x11], 4 // A3
FMLA v23.8h, v13.8h, v3.h[4]
FMLA v25.8h, v13.8h, v4.h[0]
# BLOCK 3
LDR s2, [x12], 4 // A4
INS v1.d[1], x8 // A3 ins
FMLA v27.8h, v13.8h, v4.h[4]
LDR w8, [x4], 4 // A5
FMLA v29.8h, v13.8h, v5.h[0]
FMLA v31.8h, v13.8h, v5.h[4]
# BLOCK 4
LDR d17, [x5, 80]
INS v2.d[1], x8 // A5 ins
FMLA v20.8h, v14.8h, v3.h[1]
LDR x8, [x5, 88]
FMLA v22.8h, v14.8h, v3.h[5]
FMLA v24.8h, v14.8h, v4.h[1]
# BLOCK 5
LDR d18, [x5, 96]
INS v17.d[1], x8 // B
FMLA v26.8h, v14.8h, v4.h[5]
LDR x8, [x5, 104]
FMLA v28.8h, v14.8h, v5.h[1]
FMLA v30.8h, v14.8h, v5.h[5]
# BLOCK 6
LDR d19, [x5, 112]
INS v18.d[1], x8 // B
FMLA v21.8h, v15.8h, v3.h[1]
LDR x8, [x5, 120]
FMLA v23.8h, v15.8h, v3.h[5]
FMLA v25.8h, v15.8h, v4.h[1]
# BLOCK 7
SUBS x0, x0, 8 // LDR lands here
FMLA v27.8h, v15.8h, v4.h[5]
FMLA v29.8h, v15.8h, v5.h[1]
ADD x5, x5, 128
FMLA v31.8h, v15.8h, v5.h[5]
B.HS 1b
# Epilogue - 4 halffloats of A (8 bytes)
# 48 FMA + 12 LD32 A + 8 LDR B
2:
# First group of 24 FMA, Second group loads
# BLOCK 0
LDR s3, [x3], 4 // A0
INS v19.d[1], x8 // B from second group
FMLA v20.8h, v16.8h, v0.h[0]
LDR w8, [x9], 4 // A1
FMLA v22.8h, v16.8h, v0.h[4]
FMLA v24.8h, v16.8h, v1.h[0]
# BLOCK 1
LDR d12, [x5]
INS v3.d[1], x8 // A1 ins
FMLA v26.8h, v16.8h, v1.h[4]
LDR x8, [x5, 8] // B
FMLA v28.8h, v16.8h, v2.h[0]
FMLA v30.8h, v16.8h, v2.h[4]
# BLOCK 2
LDR s4, [x10], 4 // A2
INS v12.d[1], x8 // B ins
FMLA v21.8h, v17.8h, v0.h[0]
LDR w8, [x11], 4 // A3
FMLA v23.8h, v17.8h, v0.h[4]
FMLA v25.8h, v17.8h, v1.h[0]
# BLOCK 3
LDR s5, [x12], 4 // A4
INS v4.d[1], x8 // A3 ins
FMLA v27.8h, v17.8h, v1.h[4]
LDR w8, [x4], 4 // A5
FMLA v29.8h, v17.8h, v2.h[0]
FMLA v31.8h, v17.8h, v2.h[4]
# BLOCK 4
LDR d13, [x5, 16]
INS v5.d[1], x8 // A5 ins
FMLA v20.8h, v18.8h, v0.h[1]
LDR x8, [x5, 24]
FMLA v22.8h, v18.8h, v0.h[5]
FMLA v24.8h, v18.8h, v1.h[1]
# BLOCK 5
LDR d14, [x5, 32]
INS v13.d[1], x8 // B
FMLA v26.8h, v18.8h, v1.h[5]
LDR x8, [x5, 40]
FMLA v28.8h, v18.8h, v2.h[1]
FMLA v30.8h, v18.8h, v2.h[5]
# BLOCK 6
LDR d15, [x5, 48]
INS v14.d[1], x8 // B
FMLA v21.8h, v19.8h, v0.h[1]
LDR x8, [x5, 56]
FMLA v23.8h, v19.8h, v0.h[5]
FMLA v25.8h, v19.8h, v1.h[1]
# BLOCK 7
INS v15.d[1], x8 // B
FMLA v27.8h, v19.8h, v1.h[5]
FMLA v29.8h, v19.8h, v2.h[1]
FMLA v31.8h, v19.8h, v2.h[5]
# Second group of 24 FMA, First group of loads
# BLOCK 0
FMLA v20.8h, v12.8h, v3.h[0]
FMLA v22.8h, v12.8h, v3.h[4]
FMLA v24.8h, v12.8h, v4.h[0]
# BLOCK 1
FMLA v26.8h, v12.8h, v4.h[4]
FMLA v28.8h, v12.8h, v5.h[0]
FMLA v30.8h, v12.8h, v5.h[4]
# BLOCK 2
FMLA v21.8h, v13.8h, v3.h[0]
FMLA v23.8h, v13.8h, v3.h[4]
FMLA v25.8h, v13.8h, v4.h[0]
# BLOCK 3
FMLA v27.8h, v13.8h, v4.h[4]
FMLA v29.8h, v13.8h, v5.h[0]
FMLA v31.8h, v13.8h, v5.h[4]
# BLOCK 4
FMLA v20.8h, v14.8h, v3.h[1]
FMLA v22.8h, v14.8h, v3.h[5]
FMLA v24.8h, v14.8h, v4.h[1]
# BLOCK 5
FMLA v26.8h, v14.8h, v4.h[5]
FMLA v28.8h, v14.8h, v5.h[1]
FMLA v30.8h, v14.8h, v5.h[5]
TST x0, 7
# BLOCK 6
FMLA v21.8h, v15.8h, v3.h[1]
FMLA v23.8h, v15.8h, v3.h[5]
FMLA v25.8h, v15.8h, v4.h[1]
ADD x5, x5, 64
# BLOCK 7
FMLA v27.8h, v15.8h, v4.h[5]
FMLA v29.8h, v15.8h, v5.h[1]
FMLA v31.8h, v15.8h, v5.h[5]
# Is there a remainder?- 2 halffloats of A (4 bytes) or less
B.NE 4f
3:
# Clamp
DUP v4.8h, v6.h[0]
DUP v5.8h, v6.h[1]
FMAX v20.8h, v20.8h, v4.8h
LDR x0, [sp, 32] // cn_stride
FMAX v21.8h, v21.8h, v4.8h
FMAX v22.8h, v22.8h, v4.8h
FMAX v23.8h, v23.8h, v4.8h
FMAX v24.8h, v24.8h, v4.8h
FMAX v25.8h, v25.8h, v4.8h
FMAX v26.8h, v26.8h, v4.8h
FMAX v27.8h, v27.8h, v4.8h
FMAX v28.8h, v28.8h, v4.8h
FMAX v29.8h, v29.8h, v4.8h
FMAX v30.8h, v30.8h, v4.8h
FMAX v31.8h, v31.8h, v4.8h
SUBS x1, x1, 16
FMIN v20.8h, v20.8h, v5.8h
FMIN v21.8h, v21.8h, v5.8h
FMIN v22.8h, v22.8h, v5.8h
FMIN v23.8h, v23.8h, v5.8h
FMIN v24.8h, v24.8h, v5.8h
FMIN v25.8h, v25.8h, v5.8h
FMIN v26.8h, v26.8h, v5.8h
FMIN v27.8h, v27.8h, v5.8h
FMIN v28.8h, v28.8h, v5.8h
FMIN v29.8h, v29.8h, v5.8h
FMIN v30.8h, v30.8h, v5.8h
FMIN v31.8h, v31.8h, v5.8h
# Store full 6 x 16
B.LO 6f
ST1 {v20.16b, v21.16b}, [x6], x0
SUB x3, x3, x2 // a0 -= kc
ST1 {v22.16b, v23.16b}, [x16], x0
SUB x9, x9, x2 // a1 -= kc
ST1 {v24.16b, v25.16b}, [x17], x0
SUB x10, x10, x2 // a2 -= kc
ST1 {v26.16b, v27.16b}, [x14], x0
SUB x11, x11, x2 // a3 -= kc
ST1 {v28.16b, v29.16b}, [x13], x0
SUB x12, x12, x2 // a4 -= kc
ST1 {v30.16b, v31.16b}, [x7], x0
SUB x4, x4, x2 // a5 -= kc
B.HI 0b
# Restore d12-d15 from stack
LDP d14, d15, [sp, 16]
LDP d12, d13, [sp], 32
RET
4:
# Is there a remainder?- 2 halffloats of A (4 bytes)
TBZ x0, 2, 5f
# Remainder- 2 halffloats of A (4 bytes)
LDR s0, [x3], 4 // A0
LDP q16, q17, [x5], 32 // B
LDR s1, [x10], 4 // A2
LDR s2, [x12], 4 // A4
LD1 {v0.s}[2], [x9], 4 // A1
LD1 {v1.s}[2], [x11], 4 // A3
LD1 {v2.s}[2], [x4], 4 // A5
LDR q18, [x5], 16
LDR q19, [x5], 16
FMLA v20.8h, v16.8h, v0.h[0]
FMLA v22.8h, v16.8h, v0.h[4]
FMLA v24.8h, v16.8h, v1.h[0]
FMLA v26.8h, v16.8h, v1.h[4]
FMLA v28.8h, v16.8h, v2.h[0]
FMLA v30.8h, v16.8h, v2.h[4]
FMLA v21.8h, v17.8h, v0.h[0]
FMLA v23.8h, v17.8h, v0.h[4]
FMLA v25.8h, v17.8h, v1.h[0]
FMLA v27.8h, v17.8h, v1.h[4]
FMLA v29.8h, v17.8h, v2.h[0]
FMLA v31.8h, v17.8h, v2.h[4]
FMLA v20.8h, v18.8h, v0.h[1]
FMLA v22.8h, v18.8h, v0.h[5]
FMLA v24.8h, v18.8h, v1.h[1]
FMLA v26.8h, v18.8h, v1.h[5]
FMLA v28.8h, v18.8h, v2.h[1]
FMLA v30.8h, v18.8h, v2.h[5]
FMLA v21.8h, v19.8h, v0.h[1]
FMLA v23.8h, v19.8h, v0.h[5]
FMLA v25.8h, v19.8h, v1.h[1]
FMLA v27.8h, v19.8h, v1.h[5]
FMLA v29.8h, v19.8h, v2.h[1]
FMLA v31.8h, v19.8h, v2.h[5]
# Is there a remainder?- 1 halffloat of A (2 bytes)
TBZ x0, 1, 3b
5:
# Remainder- 1 halffloat of A (2 bytes)
LDR h0, [x3], 2 // A0
LDP q16, q17, [x5], 32 // B
LDR h1, [x10], 2 // A2
LDR h2, [x12], 2 // A4
LD1 {v0.h}[4], [x9], 2 // A1
LD1 {v1.h}[4], [x11], 2 // A3
LD1 {v2.h}[4], [x4], 2 // A5
FMLA v20.8h, v16.8h, v0.h[0]
FMLA v22.8h, v16.8h, v0.h[4]
FMLA v24.8h, v16.8h, v1.h[0]
FMLA v26.8h, v16.8h, v1.h[4]
FMLA v28.8h, v16.8h, v2.h[0]
FMLA v30.8h, v16.8h, v2.h[4]
FMLA v21.8h, v17.8h, v0.h[0]
FMLA v23.8h, v17.8h, v0.h[4]
FMLA v25.8h, v17.8h, v1.h[0]
FMLA v27.8h, v17.8h, v1.h[4]
FMLA v29.8h, v17.8h, v2.h[0]
FMLA v31.8h, v17.8h, v2.h[4]
B 3b
# Store odd width
6:
TBZ x1, 3, 7f
STR q20, [x6], 16
MOV v20.16b, v21.16b
STR q22, [x16], 16
MOV v22.16b, v23.16b
STR q24, [x17], 16
MOV v24.16b, v25.16b
STR q26, [x14], 16
MOV v26.16b, v27.16b
STR q28, [x13], 16
MOV v28.16b, v29.16b
STR q30, [x7], 16
MOV v30.16b, v31.16b
7:
TBZ x1, 2, 8f
STR d20, [x6], 8
STR d22, [x16], 8
DUP d20, v20.d[1]
DUP d22, v22.d[1]
STR d24, [x17], 8
STR d26, [x14], 8
DUP d24, v24.d[1]
DUP d26, v26.d[1]
STR d28, [x13], 8
STR d30, [x7], 8
DUP d28, v28.d[1]
DUP d30, v30.d[1]
8:
TBZ x1, 1, 9f
STR s20, [x6], 4
STR s22, [x16], 4
DUP s20, v20.s[1]
DUP s22, v22.s[1]
STR s24, [x17], 4
STR s26, [x14], 4
DUP s24, v24.s[1]
DUP s26, v26.s[1]
STR s28, [x13], 4
STR s30, [x7], 4
DUP s28, v28.s[1]
DUP s30, v30.s[1]
9:
TBZ x1, 0, 10f
STR h20, [x6]
STR h22, [x16]
STR h24, [x17]
STR h26, [x14]
STR h28, [x13]
STR h30, [x7]
10:
# Restore d12-d15 from stack
LDP d14, d15, [sp, 16]
LDP d12, d13, [sp], 32
RET
END_FUNCTION xnn_f16_gemm_minmax_ukernel_6x16__asm_aarch64_neonfp16arith_cortex_a55r0
#ifdef __ELF__
.section ".note.GNU-stack","",%progbits
#endif
|