| 1 |
// Auto-generated file. Do not edit! |
| 2 |
// Template: src/qs8-igemm/1x8c8-aarch64-neon-mlal.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_1x8c8__asm_aarch64_neon_mlal( |
| 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 |
// B x5 v4 v5 v2 v3 |
| 32 |
// C0 x6 v16 v18 v20 v22 v24 v26 v28 v30 |
| 33 |
// temp0 v17 v19 v21 v23 |
| 34 |
|
| 35 |
BEGIN_FUNCTION xnn_qs8_qc8w_igemm_minmax_fp32_ukernel_1x8c8__asm_aarch64_neon_mlal |
| 36 |
# Clamp C pointers |
| 37 |
LDP x10, x8, [sp] // Load cn_stride, a_offset |
| 38 |
ADD x2, x2, 7 // kc = (kc + 7) & ~7 |
| 39 |
LDP x12, x11, [sp, 16] // Load zero, params pointer |
| 40 |
BIC x2, x2, 7 |
| 41 |
|
| 42 |
.p2align 3 |
| 43 |
0: |
| 44 |
# Load initial bias from w into accumulators |
| 45 |
LDP s16, s18, [x5], 8 |
| 46 |
LDP s20, s22, [x5], 8 |
| 47 |
LDP s24, s26, [x5], 8 |
| 48 |
LDP s28, s30, [x5], 8 |
| 49 |
MOV x9, x3 // p = ks |
| 50 |
|
| 51 |
.p2align 3 |
| 52 |
1: |
| 53 |
# Load next A pointer |
| 54 |
LDR x13, [x4], 8 |
| 55 |
CMP x13, x12 // if a0 == zero |
| 56 |
ADD x13, x13, x8 // a0 += a_offset |
| 57 |
CSEL x13, x12, x13, EQ // a0 = zero, else += a0 + a_offset |
| 58 |
|
| 59 |
# Is there at least 16 bytes for epilogue? |
| 60 |
SUBS x0, x2, 16 // k = kc - 16 |
| 61 |
B.LO 5f |
| 62 |
|
| 63 |
# Prologue: load A0 and 4 B's |
| 64 |
LDP d0, d6, [x13], 16 // Read A0 |
| 65 |
LDP d4, d5, [x5] // Read B |
| 66 |
LDP d2, d3, [x5, 64] // Read B |
| 67 |
|
| 68 |
# Is there at least 16 bytes for main loop? |
| 69 |
SUBS x0, x0, 16 // k = k - 16 |
| 70 |
B.LO 3f |
| 71 |
|
| 72 |
# Main loop - 16 bytes of A |
| 73 |
# 4 groups of 2 mul/mla/adap = 6 cycles. |
| 74 |
# 2 load for A0, A1 = +4 cycle. Total 36 cycles. |
| 75 |
|
| 76 |
.p2align 3 |
| 77 |
2: |
| 78 |
# BLOCK 0 - 4 cycles |
| 79 |
SMULL v17.8h, v4.8b, v0.8b |
| 80 |
SMULL v19.8h, v5.8b, v0.8b |
| 81 |
LDP d4, d5, [x5, 16] |
| 82 |
SMLAL v17.8h, v2.8b, v6.8b |
| 83 |
SMLAL v19.8h, v3.8b, v6.8b |
| 84 |
LDP d2, d3, [x5, 80] |
| 85 |
|
| 86 |
# BLOCK 1 - 6 cycles |
| 87 |
SMULL v21.8h, v4.8b, v0.8b |
| 88 |
SMULL v23.8h, v5.8b, v0.8b |
| 89 |
SADALP v16.4s, v17.8h |
| 90 |
SADALP v18.4s, v19.8h |
| 91 |
LDP d4, d5, [x5, 32] |
| 92 |
SMLAL v21.8h, v2.8b, v6.8b |
| 93 |
SMLAL v23.8h, v3.8b, v6.8b |
| 94 |
LDP d2, d3, [x5, 96] |
| 95 |
|
| 96 |
# BLOCK 2 - 6 cycles |
| 97 |
SMULL v17.8h, v4.8b, v0.8b |
| 98 |
SMULL v19.8h, v5.8b, v0.8b |
| 99 |
SADALP v20.4s, v21.8h |
| 100 |
SADALP v22.4s, v23.8h |
| 101 |
LDP d4, d5, [x5, 48] |
| 102 |
SMLAL v17.8h, v2.8b, v6.8b |
| 103 |
SMLAL v19.8h, v3.8b, v6.8b |
| 104 |
LDP d2, d3, [x5, 112] |
| 105 |
|
| 106 |
# BLOCK 3 - 14 cycles |
| 107 |
SMULL v21.8h, v4.8b, v0.8b |
| 108 |
ADD x5, x5, 128 |
| 109 |
SMULL v23.8h, v5.8b, v0.8b |
| 110 |
SADALP v24.4s, v17.8h |
| 111 |
SUBS x0, x0, 16 |
| 112 |
SADALP v26.4s, v19.8h |
| 113 |
LDP d4, d5, [x5] // Read B |
| 114 |
SMLAL v21.8h, v2.8b, v6.8b |
| 115 |
SMLAL v23.8h, v3.8b, v6.8b |
| 116 |
LDP d0, d6, [x13], 16 // Read A0 |
| 117 |
SADALP v28.4s, v21.8h |
| 118 |
LDP d2, d3, [x5, 64] // Read B |
| 119 |
SADALP v30.4s, v23.8h |
| 120 |
B.HS 2b |
| 121 |
|
| 122 |
# Epilogue |
| 123 |
# Same as main loop except no loads at end of loop |
| 124 |
|
| 125 |
.p2align 3 |
| 126 |
3: |
| 127 |
# BLOCK 0 - 4 cycles |
| 128 |
SMULL v17.8h, v4.8b, v0.8b |
| 129 |
SMULL v19.8h, v5.8b, v0.8b |
| 130 |
LDP d4, d5, [x5, 16] |
| 131 |
SMLAL v17.8h, v2.8b, v6.8b |
| 132 |
SMLAL v19.8h, v3.8b, v6.8b |
| 133 |
LDP d2, d3, [x5, 80] |
| 134 |
|
| 135 |
# BLOCK 1 - 6 cycles |
| 136 |
SMULL v21.8h, v4.8b, v0.8b |
| 137 |
SMULL v23.8h, v5.8b, v0.8b |
| 138 |
SADALP v16.4s, v17.8h |
| 139 |
SADALP v18.4s, v19.8h |
| 140 |
LDP d4, d5, [x5, 32] |
| 141 |
SMLAL v21.8h, v2.8b, v6.8b |
| 142 |
SMLAL v23.8h, v3.8b, v6.8b |
| 143 |
LDP d2, d3, [x5, 96] |
| 144 |
|
| 145 |
# BLOCK 2 - 6 cycles |
| 146 |
SMULL v17.8h, v4.8b, v0.8b |
| 147 |
SMULL v19.8h, v5.8b, v0.8b |
| 148 |
SADALP v20.4s, v21.8h |
| 149 |
SADALP v22.4s, v23.8h |
| 150 |
LDP d4, d5, [x5, 48] |
| 151 |
SMLAL v17.8h, v2.8b, v6.8b |
| 152 |
SMLAL v19.8h, v3.8b, v6.8b |
| 153 |
LDP d2, d3, [x5, 112] |
| 154 |
|
| 155 |
# BLOCK 3 - 8 cycles |
| 156 |
SMULL v21.8h, v4.8b, v0.8b |
| 157 |
ADD x5, x5, 128 |
| 158 |
SMULL v23.8h, v5.8b, v0.8b |
| 159 |
SADALP v24.4s, v17.8h |
| 160 |
SUBS x0, x0, 16 |
| 161 |
SADALP v26.4s, v19.8h |
| 162 |
SMLAL v21.8h, v2.8b, v6.8b |
| 163 |
SMLAL v23.8h, v3.8b, v6.8b |
| 164 |
SADALP v28.4s, v21.8h |
| 165 |
SADALP v30.4s, v23.8h |
| 166 |
|
| 167 |
# Is there a remainder?- 8 bytes of A |
| 168 |
TBNZ x0, 3, 5f |
| 169 |
|
| 170 |
# ks loop |
| 171 |
SUBS x9, x9, 8 // ks -= MR * sizeof(int8_t*) |
| 172 |
B.HI 1b |
| 173 |
|
| 174 |
4: |
| 175 |
# Add columns |
| 176 |
ADDP v16.4s, v16.4s, v18.4s |
| 177 |
ADDP v20.4s, v20.4s, v22.4s |
| 178 |
ADDP v24.4s, v24.4s, v26.4s |
| 179 |
ADDP v28.4s, v28.4s, v30.4s |
| 180 |
ADDP v0.4s, v16.4s, v20.4s |
| 181 |
ADDP v1.4s, v24.4s, v28.4s |
| 182 |
|
| 183 |
# Load per channel scale values from weights |
| 184 |
SCVTF v0.4s, v0.4s |
| 185 |
LDR q4, [x5], 16 |
| 186 |
SCVTF v1.4s, v1.4s |
| 187 |
LDR q5, [x5], 16 |
| 188 |
FMUL v0.4s, v0.4s, v4.4s |
| 189 |
FMUL v1.4s, v1.4s, v5.4s |
| 190 |
|
| 191 |
FCVTNS v0.4s, v0.4s |
| 192 |
FCVTNS v1.4s, v1.4s |
| 193 |
|
| 194 |
LD1R {v5.8h}, [x11], 2 |
| 195 |
SQXTN v0.4h, v0.4s |
| 196 |
SQXTN2 v0.8h, v1.4s |
| 197 |
SUBS x1, x1, 8 |
| 198 |
SQADD v0.8h, v0.8h, v5.8h |
| 199 |
LD1R {v1.16b}, [x11], 1 |
| 200 |
SQXTN v0.8b, v0.8h |
| 201 |
LD1R {v17.16b}, [x11] |
| 202 |
SMAX v0.8b, v0.8b, v1.8b |
| 203 |
SUB x11, x11, 3 // rewind params pointer |
| 204 |
SMIN v0.8b, v0.8b, v17.8b |
| 205 |
B.LO 6f |
| 206 |
|
| 207 |
# Store full 1 x 8 |
| 208 |
ST1 {v0.8b}, [x6], x10 |
| 209 |
SUB x4, x4, x3 // a -= ks |
| 210 |
B.HI 0b |
| 211 |
RET |
| 212 |
|
| 213 |
# Remainder - 8 bytes of A |
| 214 |
.p2align 3 |
| 215 |
5: |
| 216 |
LDR d0, [x13], 8 |
| 217 |
LDP d4, d5, [x5] |
| 218 |
LDP d6, d7, [x5, 16] |
| 219 |
SMULL v17.8h, v4.8b, v0.8b |
| 220 |
SMULL v19.8h, v5.8b, v0.8b |
| 221 |
SMULL v21.8h, v6.8b, v0.8b |
| 222 |
SMULL v23.8h, v7.8b, v0.8b |
| 223 |
LDP d4, d5, [x5, 32] |
| 224 |
LDP d6, d7, [x5, 48] |
| 225 |
SADALP v16.4s, v17.8h |
| 226 |
SADALP v18.4s, v19.8h |
| 227 |
SADALP v20.4s, v21.8h |
| 228 |
SADALP v22.4s, v23.8h |
| 229 |
SMULL v17.8h, v4.8b, v0.8b |
| 230 |
SMULL v19.8h, v5.8b, v0.8b |
| 231 |
SMULL v21.8h, v6.8b, v0.8b |
| 232 |
SMULL v23.8h, v7.8b, v0.8b |
| 233 |
ADD x5, x5, 64 |
| 234 |
SADALP v24.4s, v17.8h |
| 235 |
SADALP v26.4s, v19.8h |
| 236 |
SADALP v28.4s, v21.8h |
| 237 |
SADALP v30.4s, v23.8h |
| 238 |
|
| 239 |
# ks loop |
| 240 |
SUBS x9, x9, 8 // ks -= MR * sizeof(int8_t*) |
| 241 |
B.HI 1b |
| 242 |
B 4b |
| 243 |
|
| 244 |
# Store odd width |
| 245 |
.p2align 3 |
| 246 |
6: |
| 247 |
TBZ x1, 2, 7f |
| 248 |
STR s0, [x6], 4 |
| 249 |
EXT v0.16b, v0.16b, v0.16b, 4 |
| 250 |
|
| 251 |
7: |
| 252 |
TBZ x1, 1, 8f |
| 253 |
STR h0, [x6], 2 |
| 254 |
EXT v0.16b, v0.16b, v0.16b, 2 |
| 255 |
8: |
| 256 |
TBZ x1, 0, 9f |
| 257 |
STR b0, [x6] |
| 258 |
9: |
| 259 |
RET |
| 260 |
|
| 261 |
END_FUNCTION xnn_qs8_qc8w_igemm_minmax_fp32_ukernel_1x8c8__asm_aarch64_neon_mlal |
| 262 |
|
| 263 |
#ifdef __ELF__ |
| 264 |
.section ".note.GNU-stack","",%progbits |
| 265 |
#endif |