| 1 |
// Auto-generated file. Do not edit! |
| 2 |
// Template: src/qs8-igemm/2x8c8-aarch64-neon-mlal-cortex-a53.S.in |
| 3 |
// Generator: tools/xngen |
| 4 |
// |
| 5 |
// Copyright 2021 Google LLC |
| 6 |
// |
| 7 |
// This source code is licensed under the BSD-style license found in the |
| 8 |
// LICENSE file in the root directory of this source tree. |
| 9 |
|
| 10 |
|
| 11 |
#include "xnnpack/assembly.h" |
| 12 |
|
| 13 |
# void xnn_qs8_qc8w_igemm_minmax_fp32_ukernel_2x8c8__asm_aarch64_neon_mlal_cortex_a53( |
| 14 |
# size_t mr, x0 |
| 15 |
# size_t nc, x1 |
| 16 |
# size_t kc, x2 / x0 |
| 17 |
# size_t ks, x3 / x9 |
| 18 |
# const int8_t** restrict a, x4 |
| 19 |
# const int8_t* restrict w, x5 |
| 20 |
# int8_t* restrict c, x6 |
| 21 |
# size_t cm_stride, x7 |
| 22 |
# size_t cn_stride, [sp] -> x10 |
| 23 |
# size_t a_offset, [sp + 8] -> x8 |
| 24 |
# const int8_t* zero, [sp + 16] -> x12 |
| 25 |
# const union xnn_qs8_qc8w_conv_minmax_params params [sp + 24] -> x11 |
| 26 |
|
| 27 |
# d8-d15, x19-x30 need to be preserved if used. x18 is reserved by the OS. |
| 28 |
|
| 29 |
// Register usage |
| 30 |
// A0 x13 v0 v6 |
| 31 |
// A1 x15 v1 v7 |
| 32 |
// B x5 v4 v5 v8 v9 |
| 33 |
// C0 x6 v16 v18 v20 v22 v24 v26 v28 v30 |
| 34 |
// C1 x7 v17 v19 v21 v23 v25 v27 v29 v31 |
| 35 |
// temp0 v2 v10 v12 v14 |
| 36 |
// temp1 v3 v11 v13 v15 |
| 37 |
// x16, x17, x20, x21 tenporary a53 gpr load data |
| 38 |
|
| 39 |
|
| 40 |
BEGIN_FUNCTION xnn_qs8_qc8w_igemm_minmax_fp32_ukernel_2x8c8__asm_aarch64_neon_mlal_cortex_a53 |
| 41 |
|
| 42 |
# Clamp C pointers |
| 43 |
LDP x10, x8, [sp] // Load cn_stride, a_offset |
| 44 |
CMP x0, 2 // if mr < 2 |
| 45 |
LDP x12, x11, [sp, 16] // Load zero, params pointer |
| 46 |
ADD x7, x6, x7 // c1 = c0 + cm_stride |
| 47 |
STP d8, d9, [sp, -80]! |
| 48 |
ADD x2, x2, 7 // kc = (kc + 7) & ~7 |
| 49 |
STP d10, d11, [sp, 16] |
| 50 |
CSEL x7, x6, x7, LO // c1 = c0 |
| 51 |
STP d12, d13, [sp, 32] |
| 52 |
BIC x2, x2, 7 |
| 53 |
STP d14, d15, [sp, 48] |
| 54 |
STP x20, x21, [sp, 64] // Save x20,x21 on stack |
| 55 |
|
| 56 |
.p2align 3 |
| 57 |
0: |
| 58 |
# Load initial bias from w into accumulators |
| 59 |
LDP s16, s18, [x5], 8 |
| 60 |
MOV v17.16b, v16.16b |
| 61 |
MOV v19.16b, v18.16b |
| 62 |
LDP s20, s22, [x5], 8 |
| 63 |
MOV v21.16b, v20.16b |
| 64 |
MOV v23.16b, v22.16b |
| 65 |
LDP s24, s26, [x5], 8 |
| 66 |
MOV v25.16b, v24.16b |
| 67 |
MOV v27.16b, v26.16b |
| 68 |
LDP s28, s30, [x5], 8 |
| 69 |
MOV v29.16b, v28.16b |
| 70 |
MOV v31.16b, v30.16b |
| 71 |
MOV x9, x3 // p = ks |
| 72 |
|
| 73 |
.p2align 3 |
| 74 |
1: |
| 75 |
# Load next 2 A pointers |
| 76 |
LDP x13, x15, [x4], 16 |
| 77 |
CMP x13, x12 // if a0 == zero |
| 78 |
ADD x13, x13, x8 // a0 += a_offset |
| 79 |
CSEL x13, x12, x13, EQ // a0 = zero, else += a0 + a_offset |
| 80 |
CMP x15, x12 // if a1 == zero |
| 81 |
ADD x15, x15, x8 // a1 += a_offset |
| 82 |
CSEL x15, x12, x15, EQ // a1 = zero, else += a1 + a_offset |
| 83 |
|
| 84 |
# Is there at least 16 bytes for epilogue? |
| 85 |
SUBS x0, x2, 16 // k = kc - 16 |
| 86 |
B.LO 5f |
| 87 |
|
| 88 |
# Prologue: load A0, A1 and 2 B's |
| 89 |
LDP d4, d5, [x5] // Read B |
| 90 |
LDP d0, d6, [x13], 16 |
| 91 |
LDP d1, d7, [x15], 16 |
| 92 |
// LDP d8, d9, [x5, 64] |
| 93 |
LDR x17, [x5, 64] // Read B |
| 94 |
LDR x16, [x5, 16] |
| 95 |
|
| 96 |
# Is there at least 16 bytes for main loop? |
| 97 |
SUBS x0, x0, 16 // k = k - 16 |
| 98 |
B.LO 3f |
| 99 |
|
| 100 |
# Main loop - 16 bytes of A |
| 101 |
# 4 groups of 4 mul/mla/adap + 2 load = 18 cycles. |
| 102 |
# 2 loads for A0 = +2 cycles. Total 18 * 4 + 2 = 74 cycles. |
| 103 |
|
| 104 |
.p2align 3 |
| 105 |
2: |
| 106 |
# BLOCK 0 - 18 cycles - includes prfm |
| 107 |
LDR d9, [x5, 72] // Read B |
| 108 |
INS v8.d[0], x17 |
| 109 |
SMULL v2.8h, v4.8b, v0.8b |
| 110 |
SMULL v3.8h, v4.8b, v1.8b |
| 111 |
LDR x17, [x5, 80] |
| 112 |
SMULL v10.8h, v5.8b, v0.8b |
| 113 |
SMULL v11.8h, v5.8b, v1.8b |
| 114 |
LDR d5, [x5, 24] |
| 115 |
INS v4.d[0], x16 |
| 116 |
SMLAL v2.8h, v8.8b, v6.8b |
| 117 |
SMLAL v3.8h, v8.8b, v7.8b |
| 118 |
LDR x16, [x5, 32] |
| 119 |
SMLAL v10.8h, v9.8b, v6.8b |
| 120 |
SMLAL v11.8h, v9.8b, v7.8b |
| 121 |
SADALP v16.4s, v2.8h |
| 122 |
SADALP v17.4s, v3.8h |
| 123 |
SADALP v18.4s, v10.8h |
| 124 |
SADALP v19.4s, v11.8h |
| 125 |
|
| 126 |
# BLOCK 1- 18 cycles |
| 127 |
LDR d9, [x5, 88] |
| 128 |
INS v8.d[0], x17 |
| 129 |
SMULL v12.8h, v4.8b, v0.8b |
| 130 |
SMULL v13.8h, v4.8b, v1.8b |
| 131 |
LDR x17, [x5, 96] |
| 132 |
SMULL v14.8h, v5.8b, v0.8b |
| 133 |
SMULL v15.8h, v5.8b, v1.8b |
| 134 |
LDR d5, [x5, 40] |
| 135 |
INS v4.d[0], x16 |
| 136 |
SMLAL v12.8h, v8.8b, v6.8b |
| 137 |
SMLAL v13.8h, v8.8b, v7.8b |
| 138 |
LDR x16, [x5, 48] |
| 139 |
SMLAL v14.8h, v9.8b, v6.8b |
| 140 |
SMLAL v15.8h, v9.8b, v7.8b |
| 141 |
SADALP v20.4s, v12.8h |
| 142 |
SADALP v21.4s, v13.8h |
| 143 |
SADALP v22.4s, v14.8h |
| 144 |
SADALP v23.4s, v15.8h |
| 145 |
|
| 146 |
# BLOCK 2 - 18 cycles |
| 147 |
LDR d9, [x5, 104] |
| 148 |
INS v8.d[0], x17 |
| 149 |
SMULL v2.8h, v4.8b, v0.8b |
| 150 |
SMULL v3.8h, v4.8b, v1.8b |
| 151 |
LDR x17, [x5, 112] |
| 152 |
SMULL v10.8h, v5.8b, v0.8b |
| 153 |
SMULL v11.8h, v5.8b, v1.8b |
| 154 |
LDR d5, [x5, 56] |
| 155 |
INS v4.d[0], x16 |
| 156 |
SMLAL v2.8h, v8.8b, v6.8b |
| 157 |
SMLAL v3.8h, v8.8b, v7.8b |
| 158 |
LDR x16, [x5, 128] |
| 159 |
SMLAL v10.8h, v9.8b, v6.8b |
| 160 |
SMLAL v11.8h, v9.8b, v7.8b |
| 161 |
SADALP v24.4s, v2.8h |
| 162 |
LDR x20, [x13], 8 // Read A0 |
| 163 |
SADALP v25.4s, v3.8h |
| 164 |
LDR x21, [x15], 8 // Read A1 |
| 165 |
SADALP v26.4s, v10.8h |
| 166 |
SADALP v27.4s, v11.8h |
| 167 |
SUBS x0, x0, 16 |
| 168 |
|
| 169 |
# BLOCK 3 - includes 2 cycles to read A0, A1 = 20 cycles |
| 170 |
LDR d9, [x5, 120] |
| 171 |
INS v8.d[0], x17 |
| 172 |
SMULL v12.8h, v4.8b, v0.8b |
| 173 |
SMULL v13.8h, v4.8b, v1.8b |
| 174 |
LDR x17, [x5, 192] // Read B |
| 175 |
SMULL v14.8h, v5.8b, v0.8b |
| 176 |
SMULL v15.8h, v5.8b, v1.8b |
| 177 |
LDR d5, [x5, 136] // Read B |
| 178 |
INS v4.d[0], x16 |
| 179 |
SMLAL v12.8h, v8.8b, v6.8b |
| 180 |
SMLAL v13.8h, v8.8b, v7.8b |
| 181 |
LDR x16, [x5, 144] |
| 182 |
SMLAL v14.8h, v9.8b, v6.8b |
| 183 |
SMLAL v15.8h, v9.8b, v7.8b |
| 184 |
LDR d6, [x13], 8 // Read A0 |
| 185 |
INS v0.d[0], x20 |
| 186 |
LDR d7, [x15], 8 // Read A1 |
| 187 |
INS v1.d[0], x21 |
| 188 |
SADALP v28.4s, v12.8h |
| 189 |
SADALP v29.4s, v13.8h |
| 190 |
ADD x5, x5, 128 |
| 191 |
SADALP v30.4s, v14.8h |
| 192 |
SADALP v31.4s, v15.8h |
| 193 |
B.HS 2b |
| 194 |
|
| 195 |
# Epilogue |
| 196 |
# Same as main loop except no loads at end of loop |
| 197 |
.p2align 3 |
| 198 |
3: |
| 199 |
# BLOCK 0 - 18 cycles |
| 200 |
LDR d9, [x5, 72] // Read B |
| 201 |
INS v8.d[0], x17 |
| 202 |
SMULL v2.8h, v4.8b, v0.8b |
| 203 |
SMULL v3.8h, v4.8b, v1.8b |
| 204 |
LDR x17, [x5, 80] |
| 205 |
SMULL v10.8h, v5.8b, v0.8b |
| 206 |
SMULL v11.8h, v5.8b, v1.8b |
| 207 |
LDR d5, [x5, 24] |
| 208 |
INS v4.d[0], x16 |
| 209 |
SMLAL v2.8h, v8.8b, v6.8b |
| 210 |
SMLAL v3.8h, v8.8b, v7.8b |
| 211 |
LDR x16, [x5, 32] |
| 212 |
SMLAL v10.8h, v9.8b, v6.8b |
| 213 |
SMLAL v11.8h, v9.8b, v7.8b |
| 214 |
SADALP v16.4s, v2.8h |
| 215 |
SADALP v17.4s, v3.8h |
| 216 |
SADALP v18.4s, v10.8h |
| 217 |
SADALP v19.4s, v11.8h |
| 218 |
|
| 219 |
# BLOCK 1- 18 cycles |
| 220 |
LDR d9, [x5, 88] |
| 221 |
INS v8.d[0], x17 |
| 222 |
SMULL v12.8h, v4.8b, v0.8b |
| 223 |
SMULL v13.8h, v4.8b, v1.8b |
| 224 |
LDR x17, [x5, 96] |
| 225 |
SMULL v14.8h, v5.8b, v0.8b |
| 226 |
SMULL v15.8h, v5.8b, v1.8b |
| 227 |
LDR d5, [x5, 40] |
| 228 |
INS v4.d[0], x16 |
| 229 |
SMLAL v12.8h, v8.8b, v6.8b |
| 230 |
SMLAL v13.8h, v8.8b, v7.8b |
| 231 |
LDR x16, [x5, 48] |
| 232 |
SMLAL v14.8h, v9.8b, v6.8b |
| 233 |
SMLAL v15.8h, v9.8b, v7.8b |
| 234 |
SADALP v20.4s, v12.8h |
| 235 |
SADALP v21.4s, v13.8h |
| 236 |
SADALP v22.4s, v14.8h |
| 237 |
SADALP v23.4s, v15.8h |
| 238 |
|
| 239 |
# BLOCK 2 - 18 cycles |
| 240 |
LDR d9, [x5, 104] |
| 241 |
INS v8.d[0], x17 |
| 242 |
SMULL v2.8h, v4.8b, v0.8b |
| 243 |
SMULL v3.8h, v4.8b, v1.8b |
| 244 |
LDR x17, [x5, 112] |
| 245 |
SMULL v10.8h, v5.8b, v0.8b |
| 246 |
SMULL v11.8h, v5.8b, v1.8b |
| 247 |
LDR d5, [x5, 56] |
| 248 |
INS v4.d[0], x16 |
| 249 |
SMLAL v2.8h, v8.8b, v6.8b |
| 250 |
SMLAL v3.8h, v8.8b, v7.8b |
| 251 |
SMLAL v10.8h, v9.8b, v6.8b |
| 252 |
SMLAL v11.8h, v9.8b, v7.8b |
| 253 |
SADALP v24.4s, v2.8h |
| 254 |
SADALP v25.4s, v3.8h |
| 255 |
SADALP v26.4s, v10.8h |
| 256 |
SADALP v27.4s, v11.8h |
| 257 |
|
| 258 |
# BLOCK 3 - 17 cycles |
| 259 |
LDR d9, [x5, 120] |
| 260 |
INS v8.d[0], x17 |
| 261 |
SMULL v12.8h, v4.8b, v0.8b |
| 262 |
SMULL v13.8h, v4.8b, v1.8b |
| 263 |
SMULL v14.8h, v5.8b, v0.8b |
| 264 |
SMULL v15.8h, v5.8b, v1.8b |
| 265 |
SMLAL v12.8h, v8.8b, v6.8b |
| 266 |
SMLAL v13.8h, v8.8b, v7.8b |
| 267 |
SMLAL v14.8h, v9.8b, v6.8b |
| 268 |
SMLAL v15.8h, v9.8b, v7.8b |
| 269 |
SADALP v28.4s, v12.8h |
| 270 |
SADALP v29.4s, v13.8h |
| 271 |
ADD x5, x5, 128 |
| 272 |
SADALP v30.4s, v14.8h |
| 273 |
SADALP v31.4s, v15.8h |
| 274 |
|
| 275 |
# Is there a remainder?- 8 bytes of A |
| 276 |
TBNZ x0, 3, 5f |
| 277 |
|
| 278 |
# ks loop |
| 279 |
SUBS x9, x9, 16 // ks -= MR * sizeof(int8_t*) |
| 280 |
B.HI 1b |
| 281 |
|
| 282 |
4: |
| 283 |
# Add columns |
| 284 |
ADDP v16.4s, v16.4s, v18.4s |
| 285 |
ADDP v20.4s, v20.4s, v22.4s |
| 286 |
ADDP v24.4s, v24.4s, v26.4s |
| 287 |
ADDP v28.4s, v28.4s, v30.4s |
| 288 |
ADDP v17.4s, v17.4s, v19.4s |
| 289 |
ADDP v21.4s, v21.4s, v23.4s |
| 290 |
ADDP v25.4s, v25.4s, v27.4s |
| 291 |
ADDP v29.4s, v29.4s, v31.4s |
| 292 |
ADDP v0.4s, v16.4s, v20.4s |
| 293 |
ADDP v1.4s, v24.4s, v28.4s |
| 294 |
ADDP v2.4s, v17.4s, v21.4s |
| 295 |
ADDP v3.4s, v25.4s, v29.4s |
| 296 |
|
| 297 |
# Load per channel scale values from weights |
| 298 |
SCVTF v0.4s, v0.4s |
| 299 |
LDR q4, [x5], 16 |
| 300 |
SCVTF v1.4s, v1.4s |
| 301 |
LDR q5, [x5], 16 |
| 302 |
SCVTF v2.4s, v2.4s |
| 303 |
SCVTF v3.4s, v3.4s |
| 304 |
FMUL v0.4s, v0.4s, v4.4s |
| 305 |
FMUL v1.4s, v1.4s, v5.4s |
| 306 |
FMUL v2.4s, v2.4s, v4.4s |
| 307 |
FMUL v3.4s, v3.4s, v5.4s |
| 308 |
|
| 309 |
FCVTNS v0.4s, v0.4s |
| 310 |
FCVTNS v1.4s, v1.4s |
| 311 |
FCVTNS v2.4s, v2.4s |
| 312 |
FCVTNS v3.4s, v3.4s |
| 313 |
|
| 314 |
LD1R {v5.8h}, [x11], 2 |
| 315 |
SQXTN v0.4h, v0.4s |
| 316 |
SQXTN v2.4h, v2.4s |
| 317 |
SQXTN2 v0.8h, v1.4s |
| 318 |
SQXTN2 v2.8h, v3.4s |
| 319 |
SUBS x1, x1, 8 |
| 320 |
SQADD v0.8h, v0.8h, v5.8h |
| 321 |
SQADD v1.8h, v2.8h, v5.8h |
| 322 |
SQXTN v0.8b, v0.8h |
| 323 |
SQXTN2 v0.16b, v1.8h |
| 324 |
LD1R {v1.16b}, [x11], 1 |
| 325 |
LD1R {v2.16b}, [x11] |
| 326 |
SMAX v0.16b, v0.16b, v1.16b |
| 327 |
SUB x11, x11, 3 // rewind params pointer |
| 328 |
SMIN v0.16b, v0.16b, v2.16b |
| 329 |
B.LO 6f |
| 330 |
|
| 331 |
# Store full 2 x 8 |
| 332 |
ST1 {v0.d}[1], [x7], x10 |
| 333 |
ST1 {v0.8b}, [x6], x10 |
| 334 |
|
| 335 |
SUB x4, x4, x3 // a -= ks |
| 336 |
|
| 337 |
# nc loop |
| 338 |
B.HI 0b |
| 339 |
|
| 340 |
# Restore x20,x21 from stack |
| 341 |
LDP x20, x21, [sp, 64] |
| 342 |
|
| 343 |
# Restore d8-d15 from stack |
| 344 |
LDP d14, d15, [sp, 48] |
| 345 |
LDP d12, d13, [sp, 32] |
| 346 |
LDP d10, d11, [sp, 16] |
| 347 |
LDP d8, d9, [sp], 80 |
| 348 |
RET |
| 349 |
|
| 350 |
# Remainder - 8 bytes of A |
| 351 |
.p2align 3 |
| 352 |
5: |
| 353 |
LDR d0, [x13], 8 |
| 354 |
LDP d4, d5, [x5] |
| 355 |
LDR d1, [x15], 8 |
| 356 |
LDP d6, d7, [x5, 16] |
| 357 |
SMULL v2.8h, v4.8b, v0.8b |
| 358 |
SMULL v3.8h, v4.8b, v1.8b |
| 359 |
SMULL v10.8h, v5.8b, v0.8b |
| 360 |
SMULL v11.8h, v5.8b, v1.8b |
| 361 |
SMULL v12.8h, v6.8b, v0.8b |
| 362 |
SADALP v16.4s, v2.8h |
| 363 |
SMULL v13.8h, v6.8b, v1.8b |
| 364 |
SADALP v17.4s, v3.8h |
| 365 |
SMULL v14.8h, v7.8b, v0.8b |
| 366 |
SADALP v18.4s, v10.8h |
| 367 |
SMULL v15.8h, v7.8b, v1.8b |
| 368 |
SADALP v19.4s, v11.8h |
| 369 |
LDP d4, d5, [x5, 32] |
| 370 |
SMULL v2.8h, v4.8b, v0.8b |
| 371 |
SADALP v20.4s, v12.8h |
| 372 |
SMULL v3.8h, v4.8b, v1.8b |
| 373 |
SADALP v21.4s, v13.8h |
| 374 |
SMULL v10.8h, v5.8b, v0.8b |
| 375 |
SADALP v22.4s, v14.8h |
| 376 |
SMULL v11.8h, v5.8b, v1.8b |
| 377 |
SADALP v23.4s, v15.8h |
| 378 |
LDP d6, d7, [x5, 48] |
| 379 |
SMULL v12.8h, v6.8b, v0.8b |
| 380 |
SADALP v24.4s, v2.8h |
| 381 |
SMULL v13.8h, v6.8b, v1.8b |
| 382 |
SADALP v25.4s, v3.8h |
| 383 |
SMULL v14.8h, v7.8b, v0.8b |
| 384 |
SADALP v26.4s, v10.8h |
| 385 |
SMULL v15.8h, v7.8b, v1.8b |
| 386 |
SADALP v27.4s, v11.8h |
| 387 |
ADD x5, x5, 64 |
| 388 |
SADALP v28.4s, v12.8h |
| 389 |
SADALP v29.4s, v13.8h |
| 390 |
SADALP v30.4s, v14.8h |
| 391 |
SADALP v31.4s, v15.8h |
| 392 |
|
| 393 |
# ks loop |
| 394 |
SUBS x9, x9, 16 // ks -= MR * sizeof(int8_t*) |
| 395 |
B.HI 1b |
| 396 |
B 4b |
| 397 |
|
| 398 |
# Store odd width |
| 399 |
.p2align 3 |
| 400 |
6: |
| 401 |
TBZ x1, 2, 7f |
| 402 |
ST1 {v0.s}[2], [x7], 4 |
| 403 |
STR s0, [x6], 4 |
| 404 |
EXT v0.16b, v0.16b, v0.16b, 4 |
| 405 |
|
| 406 |
7: |
| 407 |
TBZ x1, 1, 8f |
| 408 |
ST1 {v0.h}[4], [x7], 2 |
| 409 |
STR h0, [x6], 2 |
| 410 |
EXT v0.16b, v0.16b, v0.16b, 2 |
| 411 |
8: |
| 412 |
TBZ x1, 0, 9f |
| 413 |
ST1 {v0.b}[8], [x7] |
| 414 |
STR b0, [x6] |
| 415 |
9: |
| 416 |
# Restore x20,x21 from stack |
| 417 |
LDP x20, x21, [sp, 64] |
| 418 |
|
| 419 |
# Restore d8-d15 from stack |
| 420 |
LDP d14, d15, [sp, 48] |
| 421 |
LDP d12, d13, [sp, 32] |
| 422 |
LDP d10, d11, [sp, 16] |
| 423 |
LDP d8, d9, [sp], 80 |
| 424 |
RET |
| 425 |
|
| 426 |
END_FUNCTION xnn_qs8_qc8w_igemm_minmax_fp32_ukernel_2x8c8__asm_aarch64_neon_mlal_cortex_a53 |
| 427 |
|
| 428 |
#ifdef __ELF__ |
| 429 |
.section ".note.GNU-stack","",%progbits |
| 430 |
#endif |