File size: 7,076 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 |
// Auto-generated file. Do not edit!
// Template: src/f16-gemm/4x16-aarch64-neonfp16arith-ld32.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_4x16__asm_aarch64_neonfp16arith_ld32(
# 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] -> x14
# 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 x11 v1
// A2 x12 v2
// A3 x4 v3
// B x5 v20 v21 v22 v23
// C0 x6 v24 v25
// C1 x9 v26 v27
// C2 x10 v28 v29
// C3 x7 v30 v31
// clamp v4, v5
BEGIN_FUNCTION xnn_f16_gemm_minmax_ukernel_4x16__asm_aarch64_neonfp16arith_ld32
# Load cn_stride, params pointer
LDP x14, x8, [sp]
# Load params values
LD2R {v4.8h, v5.8h}, [x8]
# Clamp A and C pointers
CMP x0, 2 // if mr < 2
ADD x11, x3, x4 // a1 = a0 + a_stride
ADD x9, x6, x7 // c1 = c0 + cm_stride
CSEL x11, x3, x11, LO // a1 = a0
CSEL x9, x6, x9, LO // c1 = c0
ADD x12, x11, x4 // a2 = a1 + a_stride
ADD x10, x9, x7 // c2 = c1 + cm_stride
// if mr <= 2
CSEL x12, x11, x12, LS // a2 = a1
CSEL x10, x9, x10, LS // c2 = c1
CMP x0, 4 // if mr < 4
ADD x4, x12, x4 // a3 = a2 + a_stride
ADD x7, x10, x7 // c3 = c2 + cm_stride
CSEL x4, x12, x4, LO // a3 = a2
CSEL x7, x10, x7, LO // c3 = c2
0:
# Load initial bias from w into accumulators
LDR q24, [x5], 16
LDR q25, [x5], 16
MOV v26.16b, v24.16b
MOV v28.16b, v24.16b
MOV v30.16b, v24.16b
MOV v27.16b, v25.16b
MOV v29.16b, v25.16b
MOV v31.16b, v25.16b
# Is there at least 2 halffloats (4 bytes)?
SUBS x0, x2, 4 // k = kc - 4
B.LO 3f
.p2align 3
# Main loop - 2 halffloats of A (4 bytes)
1:
LDR s0, [x3], 4
LDR q20, [x5], 16
LDR q21, [x5], 16
LDR s1, [x11], 4
LDR s2, [x12], 4
LDR s3, [x4], 4
LDR q22, [x5], 16
LDR q23, [x5], 16
SUBS x0, x0, 4
FMLA v24.8h, v20.8h, v0.h[0]
FMLA v25.8h, v21.8h, v0.h[0]
FMLA v26.8h, v20.8h, v1.h[0]
FMLA v27.8h, v21.8h, v1.h[0]
FMLA v28.8h, v20.8h, v2.h[0]
FMLA v29.8h, v21.8h, v2.h[0]
FMLA v30.8h, v20.8h, v3.h[0]
FMLA v31.8h, v21.8h, v3.h[0]
FMLA v24.8h, v22.8h, v0.h[1]
FMLA v25.8h, v23.8h, v0.h[1]
FMLA v26.8h, v22.8h, v1.h[1]
FMLA v27.8h, v23.8h, v1.h[1]
FMLA v28.8h, v22.8h, v2.h[1]
FMLA v29.8h, v23.8h, v2.h[1]
FMLA v30.8h, v22.8h, v3.h[1]
FMLA v31.8h, v23.8h, v3.h[1]
B.HS 1b
# Is there a remainder?- 1 halffloat of A (2 bytes)
TBNZ x0, 1, 3f
2:
# Clamp
FMAX v24.8h, v24.8h, v4.8h
SUBS x1, x1, 16
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
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 4 x 16
B.LO 4f
ST1 {v24.16b, v25.16b}, [x6], x14
SUB x3, x3, x2 // a0 -= kc
ST1 {v26.16b, v27.16b}, [x9], x14
SUB x11, x11, x2 // a1 -= kc
ST1 {v28.16b, v29.16b}, [x10], x14
SUB x12, x12, x2 // a2 -= kc
ST1 {v30.16b, v31.16b}, [x7], x14
SUB x4, x4, x2 // a3 -= kc
B.HI 0b
RET
# Remainder- 1 halffloat of A (2 bytes)
3:
LDR h0, [x3], 2
LDR q20, [x5], 16
LDR q21, [x5], 16
LDR h1, [x11], 2
LDR h2, [x12], 2
LDR h3, [x4], 2
FMLA v24.8h, v20.8h, v0.h[0]
FMLA v25.8h, v21.8h, v0.h[0]
FMLA v26.8h, v20.8h, v1.h[0]
FMLA v27.8h, v21.8h, v1.h[0]
FMLA v28.8h, v20.8h, v2.h[0]
FMLA v29.8h, v21.8h, v2.h[0]
FMLA v30.8h, v20.8h, v3.h[0]
FMLA v31.8h, v21.8h, v3.h[0]
B 2b
# Store odd width
4:
TBZ x1, 3, 5f
STR q24, [x6], 16
MOV v24.16b, v25.16b
STR q26, [x9], 16
MOV v26.16b, v27.16b
STR q28, [x10], 16
MOV v28.16b, v29.16b
STR q30, [x7], 16
MOV v30.16b, v31.16b
5:
TBZ x1, 2, 6f
STR d24, [x6], 8
STR d26, [x9], 8
DUP d24, v24.d[1]
DUP d26, v26.d[1]
STR d28, [x10], 8
STR d30, [x7], 8
DUP d28, v28.d[1]
DUP d30, v30.d[1]
6:
TBZ x1, 1, 7f
STR s24, [x6], 4
STR s26, [x9], 4
DUP s24, v24.s[1]
DUP s26, v26.s[1]
STR s28, [x10], 4
STR s30, [x7], 4
DUP s28, v28.s[1]
DUP s30, v30.s[1]
7:
TBZ x1, 0, 8f
STR h24, [x6]
STR h26, [x9]
STR h28, [x10]
STR h30, [x7]
8:
RET
END_FUNCTION xnn_f16_gemm_minmax_ukernel_4x16__asm_aarch64_neonfp16arith_ld32
#ifdef __ELF__
.section ".note.GNU-stack","",%progbits
#endif
|