| 1 |
// Auto-generated file. Do not edit! |
| 2 |
// Template: src/qs8-gemm/4x16c4-aarch64-neondot-ld64.S.in |
| 3 |
// Generator: tools/xngen |
| 4 |
// |
| 5 |
// Copyright 2020 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_qd8_f32_qc8w_gemm_minmax__ukernel_4x16c4__asm_aarch64_neondot_ld64( |
| 14 |
# size_t mr, x0 |
| 15 |
# size_t nc, x1 |
| 16 |
# size_t kc, x2 / x0 |
| 17 |
# const int8_t* restrict a, x3 |
| 18 |
# size_t a_stride, x4 |
| 19 |
# const void* restrict w, x5 |
| 20 |
# int8_t* restrict c, x6 |
| 21 |
# size_t cm_stride, x7 |
| 22 |
# size_t cn_stride, [sp] -> x12 |
| 23 |
# const union xnn_f32_minmax_params *params, [sp + 8] -> x11 |
| 24 |
# const struct xnn_qd8_quantization_params *quantization_params) [sp + 16] -> x16 |
| 25 |
|
| 26 |
# d8-d15, x19-x30 need to be preserved if used. x18 is reserved by the OS. |
| 27 |
|
| 28 |
// Register usage |
| 29 |
// A0 x3 v0 |
| 30 |
// A1 x15 v1 |
| 31 |
// A2 x13 v2 |
| 32 |
// A3 x4 v3 |
| 33 |
// B x5 v4 v5 v6 v7 |
| 34 |
// C0 x6 v16 v20 v24 v28 |
| 35 |
// C1 x8 v17 v21 v25 v29 |
| 36 |
// C2 x9 v18 v22 v26 v30 |
| 37 |
// C3 x7 v19 v23 v27 v31 |
| 38 |
// unused v14 v15 |
| 39 |
|
| 40 |
BEGIN_FUNCTION xnn_qd8_f32_qc8w_gemm_minmax_ukernel_4x16c4__asm_aarch64_neondot_ld64 |
| 41 |
|
| 42 |
# Clamp A and C pointers |
| 43 |
CMP x0, 2 // if mr < 2 |
| 44 |
ADD x2, x2, 3 // kc = (kc + 3) & ~3 |
| 45 |
ADD x15, x3, x4 // a1 = a0 + a_stride |
| 46 |
ADD x8, x6, x7 // c1 = c0 + cm_stride |
| 47 |
CSEL x15, x3, x15, LO // a1 = a0 |
| 48 |
CSEL x8, x6, x8, LO // c1 = c0 |
| 49 |
BIC x2, x2, 3 |
| 50 |
|
| 51 |
ADD x13, x15, x4 // a2 = a1 + a_stride |
| 52 |
ADD x9, x8, x7 // c2 = c1 + cm_stride |
| 53 |
// if mr <= 2 |
| 54 |
CSEL x13, x15, x13, LS // a2 = a1 |
| 55 |
CSEL x9, x8, x9, LS // c2 = c1 |
| 56 |
|
| 57 |
LDP x12, x11, [sp] // cn_stride, params |
| 58 |
|
| 59 |
LDR x16, [sp, 16] // &quantization_params[0].zero_point |
| 60 |
STP d8, d9, [sp, -48]! |
| 61 |
STP d10, d11, [sp, 16] |
| 62 |
STP d12, d13, [sp, 32] |
| 63 |
LDP q12, q13, [x16] // v12 & v13 interleaved zero_point & scale |
| 64 |
|
| 65 |
CMP x0, 4 // if mr < 4 |
| 66 |
ADD x4, x13, x4 // a3 = a2 + a_stride |
| 67 |
ADD x7, x9, x7 // c3 = c2 + cm_stride |
| 68 |
CSEL x4, x13, x4, LO // a3 = a2 |
| 69 |
CSEL x7, x9, x7, LO // c3 = c2 |
| 70 |
|
| 71 |
.p2align 3 |
| 72 |
0: |
| 73 |
# Load initial bias from w into accumulators |
| 74 |
SUBS x0, x2, 8 // k = kc - 8 |
| 75 |
LDP q0, q1, [x5], 32 |
| 76 |
MUL v16.4s, v0.4s, v12.s[0] |
| 77 |
MUL v17.4s, v0.4s, v12.s[2] |
| 78 |
MUL v18.4s, v0.4s, v13.s[0] |
| 79 |
LDP q2, q3, [x5], 32 |
| 80 |
MUL v19.4s, v0.4s, v13.s[2] |
| 81 |
MUL v20.4s, v1.4s, v12.s[0] |
| 82 |
MUL v21.4s, v1.4s, v12.s[2] |
| 83 |
MUL v22.4s, v1.4s, v13.s[0] |
| 84 |
MUL v23.4s, v1.4s, v13.s[2] |
| 85 |
MUL v24.4s, v2.4s, v12.s[0] |
| 86 |
MUL v25.4s, v2.4s, v12.s[2] |
| 87 |
MUL v26.4s, v2.4s, v13.s[0] |
| 88 |
MUL v27.4s, v2.4s, v13.s[2] |
| 89 |
MUL v28.4s, v3.4s, v12.s[0] |
| 90 |
MUL v29.4s, v3.4s, v12.s[2] |
| 91 |
MUL v30.4s, v3.4s, v13.s[0] |
| 92 |
MUL v31.4s, v3.4s, v13.s[2] |
| 93 |
|
| 94 |
# Is there at least 8 bytes? |
| 95 |
B.LO 3f |
| 96 |
|
| 97 |
# Main loop - 8 bytes of A |
| 98 |
.p2align 3 |
| 99 |
1: |
| 100 |
LDR d0, [x3], 8 |
| 101 |
LDR q4, [x5], 16 |
| 102 |
LDR d1, [x15], 8 |
| 103 |
LDR d2, [x13], 8 |
| 104 |
LDR d3, [x4], 8 |
| 105 |
LDR q5, [x5], 16 |
| 106 |
SDOT v16.4s, v4.16b, v0.4b[0] |
| 107 |
SDOT v17.4s, v4.16b, v1.4b[0] |
| 108 |
LDP q6, q7, [x5], 32 |
| 109 |
SDOT v18.4s, v4.16b, v2.4b[0] |
| 110 |
SDOT v19.4s, v4.16b, v3.4b[0] |
| 111 |
SDOT v20.4s, v5.16b, v0.4b[0] |
| 112 |
SDOT v21.4s, v5.16b, v1.4b[0] |
| 113 |
SDOT v22.4s, v5.16b, v2.4b[0] |
| 114 |
SDOT v23.4s, v5.16b, v3.4b[0] |
| 115 |
SDOT v24.4s, v6.16b, v0.4b[0] |
| 116 |
SDOT v25.4s, v6.16b, v1.4b[0] |
| 117 |
LDP q4, q5, [x5], 32 |
| 118 |
SDOT v26.4s, v6.16b, v2.4b[0] |
| 119 |
SDOT v27.4s, v6.16b, v3.4b[0] |
| 120 |
SDOT v28.4s, v7.16b, v0.4b[0] |
| 121 |
SDOT v29.4s, v7.16b, v1.4b[0] |
| 122 |
SDOT v30.4s, v7.16b, v2.4b[0] |
| 123 |
SDOT v31.4s, v7.16b, v3.4b[0] |
| 124 |
SDOT v16.4s, v4.16b, v0.4b[1] |
| 125 |
SDOT v17.4s, v4.16b, v1.4b[1] |
| 126 |
LDP q6, q7, [x5], 32 |
| 127 |
SDOT v18.4s, v4.16b, v2.4b[1] |
| 128 |
SDOT v19.4s, v4.16b, v3.4b[1] |
| 129 |
SDOT v20.4s, v5.16b, v0.4b[1] |
| 130 |
SDOT v21.4s, v5.16b, v1.4b[1] |
| 131 |
SDOT v22.4s, v5.16b, v2.4b[1] |
| 132 |
SDOT v23.4s, v5.16b, v3.4b[1] |
| 133 |
SDOT v24.4s, v6.16b, v0.4b[1] |
| 134 |
SDOT v25.4s, v6.16b, v1.4b[1] |
| 135 |
SDOT v26.4s, v6.16b, v2.4b[1] |
| 136 |
SDOT v27.4s, v6.16b, v3.4b[1] |
| 137 |
SDOT v28.4s, v7.16b, v0.4b[1] |
| 138 |
SDOT v29.4s, v7.16b, v1.4b[1] |
| 139 |
SDOT v30.4s, v7.16b, v2.4b[1] |
| 140 |
SUBS x0, x0, 8 |
| 141 |
SDOT v31.4s, v7.16b, v3.4b[1] |
| 142 |
B.HS 1b |
| 143 |
|
| 144 |
# Is there a remainder?- 4 bytes of A |
| 145 |
TBNZ x0, 2, 3f |
| 146 |
|
| 147 |
2: |
| 148 |
LDP q0, q1, [x5], 32 // kernel_scale |
| 149 |
LDP q2, q3, [x5], 32 |
| 150 |
SCVTF v19.4s, v19.4s |
| 151 |
SCVTF v23.4s, v23.4s |
| 152 |
SCVTF v27.4s, v27.4s |
| 153 |
SCVTF v31.4s, v31.4s |
| 154 |
SCVTF v18.4s, v18.4s |
| 155 |
SCVTF v22.4s, v22.4s |
| 156 |
SCVTF v26.4s, v26.4s |
| 157 |
FMUL v8.4s, v0.4s, v13.s[3] // kernel_scale * scale |
| 158 |
FMUL v9.4s, v1.4s, v13.s[3] |
| 159 |
FMUL v10.4s, v2.4s, v13.s[3] |
| 160 |
FMUL v11.4s, v3.4s, v13.s[3] |
| 161 |
FMUL v4.4s, v0.4s, v13.s[1] |
| 162 |
FMUL v5.4s, v1.4s, v13.s[1] |
| 163 |
FMUL v6.4s, v2.4s, v13.s[1] |
| 164 |
FMUL v7.4s, v3.4s, v13.s[1] |
| 165 |
SCVTF v30.4s, v30.4s |
| 166 |
SCVTF v17.4s, v17.4s |
| 167 |
SCVTF v21.4s, v21.4s |
| 168 |
SCVTF v25.4s, v25.4s |
| 169 |
SCVTF v29.4s, v29.4s |
| 170 |
SCVTF v16.4s, v16.4s |
| 171 |
SCVTF v20.4s, v20.4s |
| 172 |
SCVTF v24.4s, v24.4s |
| 173 |
SCVTF v28.4s, v28.4s |
| 174 |
FMUL v19.4s, v19.4s, v8.4s |
| 175 |
FMUL v8.4s, v0.4s, v12.s[3] |
| 176 |
FMUL v23.4s, v23.4s, v9.4s |
| 177 |
FMUL v9.4s, v1.4s, v12.s[3] |
| 178 |
FMUL v27.4s, v27.4s, v10.4s |
| 179 |
FMUL v10.4s, v2.4s, v12.s[3] |
| 180 |
FMUL v31.4s, v31.4s, v11.4s |
| 181 |
FMUL v11.4s, v3.4s, v12.s[3] |
| 182 |
FMUL v18.4s, v18.4s, v4.4s |
| 183 |
FMUL v4.4s, v0.4s, v12.s[1] |
| 184 |
FMUL v22.4s, v22.4s, v5.4s |
| 185 |
FMUL v5.4s, v1.4s, v12.s[1] |
| 186 |
FMUL v26.4s, v26.4s, v6.4s |
| 187 |
FMUL v6.4s, v2.4s, v12.s[1] |
| 188 |
FMUL v30.4s, v30.4s, v7.4s |
| 189 |
FMUL v7.4s, v3.4s, v12.s[1] |
| 190 |
LDP q0, q1, [x5], 32 // bias |
| 191 |
LDP q2, q3, [x5], 32 |
| 192 |
FMUL v17.4s, v17.4s, v8.4s |
| 193 |
FMUL v21.4s, v21.4s, v9.4s |
| 194 |
FMUL v25.4s, v25.4s, v10.4s |
| 195 |
FMUL v29.4s, v29.4s, v11.4s |
| 196 |
FMUL v16.4s, v16.4s, v4.4s |
| 197 |
FMUL v20.4s, v20.4s, v5.4s |
| 198 |
FMUL v24.4s, v24.4s, v6.4s |
| 199 |
FMUL v28.4s, v28.4s, v7.4s |
| 200 |
LD2R {v4.4s, v5.4s}, [x11] // min max |
| 201 |
FADD v19.4s, v19.4s, v0.4s |
| 202 |
FADD v23.4s, v23.4s, v1.4s |
| 203 |
FADD v27.4s, v27.4s, v2.4s |
| 204 |
FADD v31.4s, v31.4s, v3.4s |
| 205 |
FADD v18.4s, v18.4s, v0.4s |
| 206 |
FADD v22.4s, v22.4s, v1.4s |
| 207 |
FADD v26.4s, v26.4s, v2.4s |
| 208 |
FADD v30.4s, v30.4s, v3.4s |
| 209 |
FADD v17.4s, v17.4s, v0.4s |
| 210 |
FADD v21.4s, v21.4s, v1.4s |
| 211 |
FADD v25.4s, v25.4s, v2.4s |
| 212 |
FADD v29.4s, v29.4s, v3.4s |
| 213 |
FADD v16.4s, v16.4s, v0.4s |
| 214 |
FADD v20.4s, v20.4s, v1.4s |
| 215 |
FADD v24.4s, v24.4s, v2.4s |
| 216 |
FADD v28.4s, v28.4s, v3.4s |
| 217 |
FMAX v19.4s, v19.4s, v4.4s |
| 218 |
FMAX v23.4s, v23.4s, v4.4s |
| 219 |
FMAX v27.4s, v27.4s, v4.4s |
| 220 |
FMAX v31.4s, v31.4s, v4.4s |
| 221 |
FMAX v18.4s, v18.4s, v4.4s |
| 222 |
FMAX v22.4s, v22.4s, v4.4s |
| 223 |
FMAX v26.4s, v26.4s, v4.4s |
| 224 |
FMAX v30.4s, v30.4s, v4.4s |
| 225 |
FMAX v17.4s, v17.4s, v4.4s |
| 226 |
FMAX v21.4s, v21.4s, v4.4s |
| 227 |
FMAX v25.4s, v25.4s, v4.4s |
| 228 |
FMAX v29.4s, v29.4s, v4.4s |
| 229 |
FMAX v16.4s, v16.4s, v4.4s |
| 230 |
FMAX v20.4s, v20.4s, v4.4s |
| 231 |
FMAX v24.4s, v24.4s, v4.4s |
| 232 |
FMAX v28.4s, v28.4s, v4.4s |
| 233 |
FMIN v19.4s, v19.4s, v5.4s |
| 234 |
FMIN v23.4s, v23.4s, v5.4s |
| 235 |
FMIN v27.4s, v27.4s, v5.4s |
| 236 |
FMIN v31.4s, v31.4s, v5.4s |
| 237 |
FMIN v18.4s, v18.4s, v5.4s |
| 238 |
FMIN v22.4s, v22.4s, v5.4s |
| 239 |
FMIN v26.4s, v26.4s, v5.4s |
| 240 |
FMIN v30.4s, v30.4s, v5.4s |
| 241 |
FMIN v17.4s, v17.4s, v5.4s |
| 242 |
FMIN v21.4s, v21.4s, v5.4s |
| 243 |
FMIN v25.4s, v25.4s, v5.4s |
| 244 |
FMIN v29.4s, v29.4s, v5.4s |
| 245 |
FMIN v16.4s, v16.4s, v5.4s |
| 246 |
FMIN v20.4s, v20.4s, v5.4s |
| 247 |
FMIN v24.4s, v24.4s, v5.4s |
| 248 |
FMIN v28.4s, v28.4s, v5.4s |
| 249 |
SUBS x1, x1, 16 |
| 250 |
B.LO 4f |
| 251 |
STP q19, q23, [x7] |
| 252 |
STP q27, q31, [x7, #32] |
| 253 |
ADD x7, x7, x12 |
| 254 |
STP q18, q22, [x9] |
| 255 |
STP q26, q30, [x9, #32] |
| 256 |
ADD x9, x9, x12 |
| 257 |
STP q17, q21, [x8] |
| 258 |
STP q25, q29, [x8, #32] |
| 259 |
ADD x8, x8, x12 |
| 260 |
STP q16, q20, [x6] |
| 261 |
STP q24, q28, [x6, #32] |
| 262 |
ADD x6, x6, x12 |
| 263 |
SUB x3, x3, x2 // a0 -= kc |
| 264 |
SUB x15, x15, x2 // a1 -= kc |
| 265 |
SUB x13, x13, x2 // a2 -= kc |
| 266 |
SUB x4, x4, x2 // a3 -= kc |
| 267 |
B.NE 0b |
| 268 |
|
| 269 |
# Restore d8-d13 from stack |
| 270 |
LDP d12, d13, [sp, 32] |
| 271 |
LDP d10, d11, [sp, 16] |
| 272 |
LDP d8, d9, [sp], 48 |
| 273 |
RET |
| 274 |
|
| 275 |
# Remainder- 4 bytes of A |
| 276 |
.p2align 3 |
| 277 |
3: |
| 278 |
LDR s0, [x3], 4 |
| 279 |
LDR q4, [x5], 16 |
| 280 |
LDR s1, [x15], 4 |
| 281 |
LDR s2, [x13], 4 |
| 282 |
LDR s3, [x4], 4 |
| 283 |
SDOT v16.4s, v4.16b, v0.4b[0] |
| 284 |
LDR q5, [x5], 16 |
| 285 |
SDOT v17.4s, v4.16b, v1.4b[0] |
| 286 |
SDOT v18.4s, v4.16b, v2.4b[0] |
| 287 |
SDOT v19.4s, v4.16b, v3.4b[0] |
| 288 |
SDOT v20.4s, v5.16b, v0.4b[0] |
| 289 |
LDP q6, q7, [x5], 32 |
| 290 |
SDOT v21.4s, v5.16b, v1.4b[0] |
| 291 |
SDOT v22.4s, v5.16b, v2.4b[0] |
| 292 |
SDOT v23.4s, v5.16b, v3.4b[0] |
| 293 |
SDOT v24.4s, v6.16b, v0.4b[0] |
| 294 |
SDOT v25.4s, v6.16b, v1.4b[0] |
| 295 |
SDOT v26.4s, v6.16b, v2.4b[0] |
| 296 |
SDOT v27.4s, v6.16b, v3.4b[0] |
| 297 |
SDOT v28.4s, v7.16b, v0.4b[0] |
| 298 |
SDOT v29.4s, v7.16b, v1.4b[0] |
| 299 |
SDOT v30.4s, v7.16b, v2.4b[0] |
| 300 |
SDOT v31.4s, v7.16b, v3.4b[0] |
| 301 |
B 2b |
| 302 |
|
| 303 |
# Store odd width |
| 304 |
.p2align 3 |
| 305 |
4: |
| 306 |
TBZ x1, 3, 5f |
| 307 |
STP q19, q23, [x7] |
| 308 |
STP q18, q22, [x9] |
| 309 |
MOV v19.16b, v27.16b |
| 310 |
MOV v23.16b, v31.16b |
| 311 |
MOV v18.16b, v26.16b |
| 312 |
MOV v22.16b, v30.16b |
| 313 |
STP q17, q21, [x8] |
| 314 |
STP q16, q20, [x6] |
| 315 |
MOV v17.16b, v25.16b |
| 316 |
MOV v21.16b, v29.16b |
| 317 |
MOV v16.16b, v24.16b |
| 318 |
MOV v20.16b, v28.16b |
| 319 |
ADD x6, x6, #32 |
| 320 |
ADD x7, x7, #32 |
| 321 |
ADD x8, x8, #32 |
| 322 |
ADD x9, x9, #32 |
| 323 |
5: |
| 324 |
TBZ x1, 2, 6f |
| 325 |
STR q19, [x7] |
| 326 |
STR q18, [x9] |
| 327 |
MOV v19.16b, v23.16b |
| 328 |
MOV v18.16b, v22.16b |
| 329 |
STR q17, [x8] |
| 330 |
STR q16, [x6] |
| 331 |
MOV v17.16b, v21.16b |
| 332 |
MOV v16.16b, v20.16b |
| 333 |
ADD x6, x6, #16 |
| 334 |
ADD x7, x7, #16 |
| 335 |
ADD x8, x8, #16 |
| 336 |
ADD x9, x9, #16 |
| 337 |
6: |
| 338 |
TBZ x1, 1, 7f |
| 339 |
STR d19, [x7], 8 |
| 340 |
STR d18, [x9], 8 |
| 341 |
DUP d19, v19.d[1] |
| 342 |
DUP d18, v18.d[1] |
| 343 |
STR d17, [x8], 8 |
| 344 |
STR d16, [x6], 8 |
| 345 |
DUP d17, v17.d[1] |
| 346 |
DUP d16, v16.d[1] |
| 347 |
7: |
| 348 |
TBZ x1, 0, 8f |
| 349 |
STR s19, [x7] |
| 350 |
STR s18, [x9] |
| 351 |
STR s17, [x8] |
| 352 |
STR s16, [x6] |
| 353 |
8: |
| 354 |
# Restore d8-d13 from stack |
| 355 |
LDP d12, d13, [sp, 32] |
| 356 |
LDP d10, d11, [sp, 16] |
| 357 |
LDP d8, d9, [sp], 48 |
| 358 |
RET |
| 359 |
|
| 360 |
END_FUNCTION xnn_qd8_f32_qc8w_gemm_minmax_ukernel_4x16c4__asm_aarch64_neondot_ld64 |
| 361 |
|
| 362 |
#ifdef __ELF__ |
| 363 |
.section ".note.GNU-stack","",%progbits |
| 364 |
#endif |