File size: 8,020 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 |
// 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.
$assert DATATYPE in ["F32", "QC8", "QC4"]
#include <xnnpack/assembly.h>
$DATATYPE_SPEC = {"F32": "f32", "QC8": "f32_qc8w", "QC4": "f32_qc4w"}[DATATYPE]
# void xnn_${DATATYPE_SPEC}_gemm_minmax_ukernel_1x8__asm_aarch64_neonfma_ld128_acc4${"_prfm" if PREFETCH else ""}(
# size_t mr, (x0) - unused. mr = 1
# size_t nc, x1
# size_t kc, x2 / x0
# const float* a, x3
# size_t a_stride, (x4) - unused
# const void* w, x5
# float* c, x6
# size_t cm_stride, (x7) - unused
# size_t cn_stride, [sp] -> x14
$if DATATYPE == "QC4":
# const xnn_f32_qc4w_minmax_params* params) [sp + 8] -> (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.
# Register usage
# A0 x3 v0
# B x5 v20 v21 v22 v23
# C0 x6 v16 v17 v18 v19 v26 v27 v28 v29
# Clamp v4 v5
$if DATATYPE == "QC4":
# zerop/mask v6 v7
BEGIN_FUNCTION xnn_${DATATYPE_SPEC}_gemm_minmax_ukernel_1x8__asm_aarch64_neonfma_ld128_acc4${"_prfm" if PREFETCH else ""}
# Load cn_stride, params pointer
LDP x14, x8, [sp]
$if DATATYPE == "QC4":
# Load min/max/zerop values
LD3R {v4.4s, v5.4s, v6.4s}, [x8]
MOVI v7.8b, 15
$else:
# Load min/max values
LD2R {v4.4s, v5.4s}, [x8]
0:
# Load initial bias from w into accumulators
LDP q16, q17, [x5], 32
SUBS x0, x2, 16 // k = kc - 16
MOVI v18.4s, 0 // four sets of C for pipelining FMLA
MOVI v19.4s, 0
# Is there at least 4 floats (16 bytes)
B.LO 3f
MOVI v26.4s, 0
$if PREFETCH:
PRFM PLDL1KEEP, [x5]
MOVI v27.4s, 0
$if PREFETCH:
PRFM PLDL1KEEP, [x5, 64]
MOVI v28.4s, 0
$if PREFETCH:
PRFM PLDL1KEEP, [x5, 128]
MOVI v29.4s, 0
# Main loop - 4 floats of A (16 bytes)
1:
$if DATATYPE == "F32":
LDR q0, [x3], 16
LDP q20, q21, [x5], 32
LDP q22, q23, [x5], 32
$else:
LDR q22, [x5], 16
SXTL v21.8h, v22.8b
SXTL2 v23.8h, v22.16b
LDR q0, [x3], 16
SXTL v20.4s, v21.4h
SXTL v22.4s, v23.4h
SXTL2 v21.4s, v21.8h
SXTL2 v23.4s, v23.8h
SCVTF v20.4s, v20.4s
SCVTF v21.4s, v21.4s
SCVTF v22.4s, v22.4s
SCVTF v23.4s, v23.4s
FMLA v16.4s, v20.4s, v0.s[0]
FMLA v17.4s, v21.4s, v0.s[0]
$if PREFETCH:
PRFM PLDL1KEEP, [x5, 128]
FMLA v18.4s, v22.4s, v0.s[1]
FMLA v19.4s, v23.4s, v0.s[1]
$if DATATYPE == "F32":
LDP q20, q21, [x5], 32
LDP q22, q23, [x5], 32
$else:
LDR q22, [x5], 16
SXTL v21.8h, v22.8b
SXTL2 v23.8h, v22.16b
SXTL v20.4s, v21.4h
SXTL v22.4s, v23.4h
SXTL2 v21.4s, v21.8h
SXTL2 v23.4s, v23.8h
SCVTF v20.4s, v20.4s
SCVTF v21.4s, v21.4s
SCVTF v22.4s, v22.4s
SCVTF v23.4s, v23.4s
SUBS x0, x0, 16
FMLA v26.4s, v20.4s, v0.s[2]
FMLA v27.4s, v21.4s, v0.s[2]
$if PREFETCH and (DATATYPE == "F32"):
PRFM PLDL1KEEP, [x5, 128]
FMLA v28.4s, v22.4s, v0.s[3]
FMLA v29.4s, v23.4s, v0.s[3]
B.HS 1b
FADD v16.4s, v16.4s, v26.4s
FADD v18.4s, v18.4s, v28.4s
FADD v17.4s, v17.4s, v27.4s
FADD v19.4s, v19.4s, v29.4s
# Is there a remainder?- 2 float of A (8 bytes)
TBNZ x0, 3, 4f
# Is there a remainder?- 1 float of A (4 bytes)
TBNZ x0, 2, 5f
2:
$if DATATYPE in ["QC8", "QC4"]:
# Load Scale
LDP q24, q25, [x5], 32
FADD v16.4s, v16.4s, v18.4s
FADD v17.4s, v17.4s, v19.4s
$if DATATYPE in ["QC8", "QC4"]:
# Scale
FMUL v16.4s, v16.4s, v24.4s
FMUL v17.4s, v17.4s, v25.4s
SUBS x1, x1, 8
# Clamp
FMAX v16.4s, v16.4s, v4.4s
FMAX v17.4s, v17.4s, v4.4s
FMIN v16.4s, v16.4s, v5.4s
FMIN v17.4s, v17.4s, v5.4s
# Store full 1 x 8
B.LO 6f
STP q16, q17, [x6]
ADD x6, x6, x14
SUB x3, x3, x2 // a0 -= kc
B.HI 0b
RET
3:
TBZ x0, 3, 5f
# Remainder- 2 float of A (4 bytes)
4:
LDR d0, [x3], 8
$if DATATYPE == "F32":
LDP q20, q21, [x5], 32 // 16 F32 weights
LDP q22, q23, [x5], 32
$elif DATATYPE == "QC4":
LDR d22, [x5], 8 // 16 QC4 weights
AND v21.8b, v22.8b, v7.8b // first set of 8 weights
USHR v23.8b, v22.8b, 4 // second set of 8 weights
SADDW v21.8h, v6.8h, v21.8b
SADDW v23.8h, v6.8h, v23.8b
SXTL v20.4s, v21.4h
SXTL v22.4s, v23.4h
SXTL2 v21.4s, v21.8h
SXTL2 v23.4s, v23.8h
SCVTF v20.4s, v20.4s
SCVTF v21.4s, v21.4s
SCVTF v22.4s, v22.4s
SCVTF v23.4s, v23.4s
$else:
LDR q22, [x5], 16 // 16 QC8 weights
SXTL v21.8h, v22.8b
SXTL2 v23.8h, v22.16b
SXTL v20.4s, v21.4h
SXTL v22.4s, v23.4h
SXTL2 v21.4s, v21.8h
SXTL2 v23.4s, v23.8h
SCVTF v20.4s, v20.4s
SCVTF v21.4s, v21.4s
SCVTF v22.4s, v22.4s
SCVTF v23.4s, v23.4s
FMLA v16.4s, v20.4s, v0.s[0]
FMLA v17.4s, v21.4s, v0.s[0]
FMLA v18.4s, v22.4s, v0.s[1]
FMLA v19.4s, v23.4s, v0.s[1]
TBZ x0, 2, 2b
5:
# Remainder- 1 float of A (4 bytes)
LDR s0, [x3], 4
$if DATATYPE == "F32":
LDP q20, q21, [x5], 32 // 8 F32 weights
$elif DATATYPE == "QC4":
LDR d21, [x5], 8 // 8 QC4 weights
SADDW v21.8h, v6.8h, v21.8b
SXTL v20.4s, v21.4h
SXTL2 v21.4s, v21.8h
SCVTF v20.4s, v20.4s
SCVTF v21.4s, v21.4s
$else:
LDR d21, [x5], 8 // 8 QC8 weights
SXTL v21.8h, v21.8b
SXTL v20.4s, v21.4h
SXTL2 v21.4s, v21.8h
SCVTF v20.4s, v20.4s
SCVTF v21.4s, v21.4s
FMLA v16.4s, v20.4s, v0.s[0]
FMLA v17.4s, v21.4s, v0.s[0]
B 2b
# Store odd channels
6:
TBZ x1, 2, 7f
STR q16, [x6], 16
MOV v16.16b, v17.16b
7:
TBZ x1, 1, 8f
STR d16, [x6], 8
DUP d16, v16.d[1]
8:
TBZ x1, 0, 9f
STR s16, [x6]
9:
RET
END_FUNCTION xnn_${DATATYPE_SPEC}_gemm_minmax_ukernel_1x8__asm_aarch64_neonfma_ld128_acc4${"_prfm" if PREFETCH else ""}
#ifdef __ELF__
.section ".note.GNU-stack","",%progbits
#endif
|