File size: 5,265 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
// Copyright 2022 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_igemm_minmax_ukernel_1x16__asm_aarch64_neonfp16arith_ld64(
#     size_t mr,                         (x0) - unused.  mr = 1
#     size_t nc,                         x1
#     size_t kc,                         x2 / x0
#     size_t ks,                         x3 / x9
#     const void** restrict a,            x4
#     const void* restrict w,             x5
#     void* restrict c,                   x6
#     size_t cm_stride,                  (x7) - unused
#     size_t cn_stride,                  [sp] -> x10
#     size_t a_offset,                   [sp + 8] -> x11
#     const void* zero,                  [sp + 16] -> x12
#     const xnn_f16_minmax_params params [sp + 24] -> (x8)

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

// Register usage
// A0  x8 v0
// B   x5 v24 v25 v26 v27 v28 v29 v30 v31
// C0  x6 v16 v17 v18 v19 v20 v21 v22 v23
// clamp  v4, v5

BEGIN_FUNCTION xnn_f16_igemm_minmax_ukernel_1x16__asm_aarch64_neonfp16arith_ld64
        # Load cn_stride, a_offset
        LDP         x10, x11, [sp]

        # Load zero, params pointer
        LDP         x12, x8, [sp, 16]

        # Load params values
        LD2R        {v4.8h, v5.8h}, [x8]

0:
        # Load initial bias from w into accumulators
        LDR         q16, [x5], 16
        LDR         q17, [x5], 16
        MOVI        v18.8h, 0               // 4 sets of C for pipelining FMLA
        MOVI        v19.8h, 0
        MOVI        v20.8h, 0
        MOVI        v21.8h, 0
        MOVI        v22.8h, 0
        MOVI        v23.8h, 0

        MOV         x9, x3                  // p = ks

1:
        # Load next A pointer
        LDR         x8, [x4], 8

        CMP         x8, x12                 // if a0 == zero
        ADD         x8, x8, x11             // a0 += a_offset
        CSEL        x8, x12, x8, EQ         //   a0 = zero, else += a0 + a_offset

        # Is there at least 4 halffloats (8 bytes)?
        SUBS        x0, x2, 8               // k = kc - 8
        B.LO        4f

       .p2align 3
        # Main loop - 2 halffloats of A (4 bytes)
2:
        LDR         d0,  [x8], 8
        LDR         q24, [x5, 0]
        LDR         q25, [x5, 16]
        LDR         q26, [x5, 32]
        LDR         q27, [x5, 48]
        LDR         q28, [x5, 64]
        LDR         q29, [x5, 80]
        LDR         q30, [x5, 96]
        LDR         q31, [x5, 112]
        SUBS        x0, x0, 8
        FMLA        v16.8h, v24.8h, v0.h[0]
        FMLA        v17.8h, v25.8h, v0.h[0]
        FMLA        v18.8h, v26.8h, v0.h[1]
        FMLA        v19.8h, v27.8h, v0.h[1]
        FMLA        v20.8h, v28.8h, v0.h[2]
        FMLA        v21.8h, v29.8h, v0.h[2]
        FMLA        v22.8h, v30.8h, v0.h[3]
        FMLA        v23.8h, v31.8h, v0.h[3]
        ADD         x5, x5, 128
        B.HS        2b

        # Is there a remainder?- 1 halffloat of A (2 bytes)
        ANDS        x0, x0, 7
        B.NE        4f

3:
        # ks loop
        SUBS        x9, x9, 8               // ks -= MR * sizeof(void*)
        B.HI        1b

        FADD        v16.8h, v16.8h, v18.8h
        FADD        v17.8h, v17.8h, v19.8h
        FADD        v20.8h, v20.8h, v22.8h
        FADD        v21.8h, v21.8h, v23.8h
        FADD        v16.8h, v16.8h, v20.8h
        FADD        v17.8h, v17.8h, v21.8h

        # 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
        SUBS        x1, x1, 16
        B.LO        6f

        STP         q16, q17,  [x6]
        ADD         x6,  x6, x10

        SUB         x4, x4, x3              // a -= ks

        # nc loop
        B.HI        0b
        RET


        # Remainder- 1 to 3 halffloats of A (2 to 6 bytes)
4:
        TBZ         x0, 2, 5f
        LDR         s0,  [x8], 4
        LDR         q24, [x5, 0]
        LDR         q25, [x5, 16]
        LDR         q26, [x5, 32]
        LDR         q27, [x5, 48]
        FMLA        v16.8h, v24.8h, v0.h[0]
        FMLA        v17.8h, v25.8h, v0.h[0]
        FMLA        v18.8h, v26.8h, v0.h[1]
        FMLA        v19.8h, v27.8h, v0.h[1]
        ADD         x5, x5, 64
        TBZ         x0, 1, 3b
5:
        LDR         h0, [x8], 2
        LDR         q24, [x5, 0]
        LDR         q25, [x5, 16]
        FMLA        v16.8h, v24.8h, v0.h[0]
        FMLA        v17.8h, v25.8h, v0.h[0]
        ADD         x5, x5, 32
        B           3b

        # Store odd width
6:
        TBZ         x1, 3, 7f
        STR         q16, [x6], 16
        MOV         v16.16b, v17.16b
7:
        TBZ         x1, 2, 8f
        STR         d16, [x6], 8
        DUP         d16, v16.d[1]
8:
        TBZ         x1, 1, 9f
        STR         s16,  [x6], 4
        DUP         s16, v16.s[1]
9:
        TBZ         x1, 0, 10f
        STR         h16, [x6]
10:
        RET

END_FUNCTION xnn_f16_igemm_minmax_ukernel_1x16__asm_aarch64_neonfp16arith_ld64

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