File size: 3,619 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
// Auto-generated file. Do not edit!
//   Template: src/f16-gemm/1x16-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_gemminc_minmax_ukernel_1x16__asm_aarch64_neonfp16arith_ld32(
#     size_t mr,                (x0) - unused.  mr = 1
#     size_t nc,                x1
#     size_t kc,                x2 / x0
#     const void* restrict a,    x3
#     size_t a_stride,          (x4) - unused
#     const void* restrict w,    x5
#     void* restrict c,          x6
#     size_t cm_stride,         (x7) - unused
#     size_t cn_stride,         [sp] -> x14
#     const float* restrict acc,  [sp + 8] -> x15
#     const union xnn_f16_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])  [sp + 16] -> (x8)

# d8-d15, x19-x30 need to be preserved if used. x18 is reserved by the OS.

# Register usage
# A0  x3  v0 v1
# B   x5 v20 v21 v22 v23
# C0  x6 v16 v17
# Clamp v4, v5

BEGIN_FUNCTION xnn_f16_gemminc_minmax_ukernel_1x16__asm_aarch64_neonfp16arith_ld32

        # Load cn_stride, acc
        LDP         x14, x15, [sp]
        # Load params pointer
        LDR         x8, [sp, 16]

        # Load params values
        LD2R        {v4.8h, v5.8h}, [x8]
0:
        # Load initial accumulators
        LDP         q16, q17, [x15], 32

        MOVI        v18.8h, 0               // second set of C for pipelining FMLA
        MOVI        v19.8h, 0

        # Is there at least 2 halffloats (4 bytes)
        SUBS        x0, x2, 4               // k = kc - 4

        B.LO        3f

        # Main loop - 2 halffloats of A (4 bytes)
1:
        LDR         s0, [x3], 4
        LDR         q20, [x5, 0]
        LDR         q21, [x5, 16]
        LDR         q22, [x5, 32]
        LDR         q23, [x5, 48]
        SUBS        x0, x0, 4
        FMLA        v16.8h, v20.8h, v0.h[0]
        FMLA        v17.8h, v21.8h, v0.h[0]
        FMLA        v18.8h, v22.8h, v0.h[1]
        FMLA        v19.8h, v23.8h, v0.h[1]
        ADD         x5, x5, 64
        B.HS        1b

        # Is there a remainder?- 1 halffloat of A (2 bytes)
        TBNZ        x0, 1, 3f

2:
        FADD        v16.8h, v16.8h, v18.8h
        FADD        v17.8h, v17.8h, v19.8h
        SUBS        x1, x1, 16

        # Clamp
        FMAX        v16.8h, v16.8h, v4.8h
        FMAX        v17.8h, v17.8h, v4.8h
        FMIN        v16.8h, v16.8h, v5.8h
        FMIN        v17.8h, v17.8h, v5.8h

        # Store full 1 x 16
        B.LO        4f

        STP         q16, q17, [x6]
        ADD         x6, x6, x14

        SUB         x3,  x3, x2             // a0 -= kc

        B.HI        0b

        RET

3:
        # Remainder- 1 halffloat of A (2 bytes)
        LDR         q20, [x5], 16
        LDR         q21, [x5], 16
        LDR         h0, [x3], 2
        FMLA        v16.8h, v20.8h, v0.h[0]
        FMLA        v17.8h, v21.8h, v0.h[0]
        B           2b

        # Store odd channels
4:
        TBZ         x1, 3, 5f
        STR         q16, [x6], 16
        MOV         v16.16b, v17.16b

5:
        TBZ         x1, 2, 6f
        STR         d16, [x6], 8
        DUP         d16, v16.d[1]

6:
        TBZ         x1, 1, 7f
        STR         s16, [x6], 4
        DUP         s16, v16.s[1]

7:
        TBZ         x1, 0, 8f
        STR         h16, [x6]
8:
        RET

END_FUNCTION xnn_f16_gemminc_minmax_ukernel_1x16__asm_aarch64_neonfp16arith_ld32

#ifdef __ELF__
.section ".note.GNU-stack","",%progbits
#endif