xref: /aosp_15_r20/external/XNNPACK/src/f16-gemm/gen/1x16-minmax-aarch64-neonfp16arith-ld32.S (revision 4bdc94577ba0e567308109d787f7fec7b531ce36) 
1// Auto-generated file. Do not edit!
2//   Template: src/f16-gemm/1x16-aarch64-neonfp16arith-ld32.S.in
3//   Generator: tools/xngen
4//
5// Copyright 2020 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#include <xnnpack/assembly.h>
11
12# void xnn_f16_gemm_minmax_ukernel_1x16__aarch64_neonfp16arith_ld32(
13#     size_t mr,                (x0) - unused.  mr = 1
14#     size_t nc,                x1
15#     size_t kc,                x2 / x0
16#     const void*restrict a,    x3
17#     size_t a_stride,          (x4) - unused
18#     const void*restrict w,    x5
19#     void*restrict c,          x6
20#     size_t cm_stride,         (x7) - unused
21#     size_t cn_stride,         [sp] -> x14
22#     const union xnn_f16_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])  [sp + 8] -> (x8)
23
24# d8-d15, x19-x30 need to be preserved if used. x18 is reserved by the OS.
25
26# A pointer
27# x3  a0
28
29# C pointer
30# x6  c0
31
32# Clamp v4, v5
33
34BEGIN_FUNCTION xnn_f16_gemm_minmax_ukernel_1x16__aarch64_neonfp16arith_ld32
35
36        # Load cn_stride, params pointer
37        LDP     x14, x8, [sp]
38
39        # Load params values
40        LD2R    {v4.8h, v5.8h}, [x8]
410:
42        # Load initial bias from w into accumulators
43        LDP     q16, q17, [x5], 32
44
45        MOVI    v18.8h, 0               // second set of C for pipelining FMLA
46        MOVI    v19.8h, 0
47
48        # Is there at least 2 halffloats (4 bytes)
49        SUBS    x0, x2, 4               // k = kc - 4
50
51        B.LO    3f
52
53        # Main loop - 2 halffloats of A (4 bytes)
541:
55        LDR     s0, [x3], 4
56        LDR     q20, [x5, 0]
57        LDR     q21, [x5, 16]
58        LDR     q22, [x5, 32]
59        LDR     q23, [x5, 48]
60        SUBS    x0, x0, 4
61        FMLA    v16.8h, v20.8h, v0.h[0]
62        FMLA    v17.8h, v21.8h, v0.h[0]
63        FMLA    v18.8h, v22.8h, v0.h[1]
64        FMLA    v19.8h, v23.8h, v0.h[1]
65        ADD     x5, x5, 64
66        B.HS    1b
67
68        # Is there a remainder?- 1 halffloat of A (2 bytes)
69        TBNZ    x0, 1, 3f
70
712:
72        FADD    v16.8h, v16.8h, v18.8h
73        FADD    v17.8h, v17.8h, v19.8h
74        SUBS    x1, x1, 16
75
76        # Clamp
77        FMAX    v16.8h, v16.8h, v4.8h
78        FMAX    v17.8h, v17.8h, v4.8h
79        FMIN    v16.8h, v16.8h, v5.8h
80        FMIN    v17.8h, v17.8h, v5.8h
81
82        # Store full 1 x 16
83        B.LO    4f
84
85        STP     q16, q17, [x6]
86        ADD     x6, x6, x14
87
88        SUB     x3,  x3, x2             // a0 -= kc
89
90        B.HI    0b
91
92        RET
93
943:
95        # Remainder- 1 halffloat of A (2 bytes)
96        LDR     q20, [x5], 16
97        LDR     q21, [x5], 16
98        LDR     h0, [x3], 2
99        FMLA    v16.8h, v20.8h, v0.h[0]
100        FMLA    v17.8h, v21.8h, v0.h[0]
101        B       2b
102
103        # Store odd channels
1044:
105        TBZ     x1, 3, 5f
106        STR     q16, [x6], 16
107        MOV     v16.16b, v17.16b
108
1095:
110        TBZ     x1, 2, 6f
111        STR     d16, [x6], 8
112        DUP     d16, v16.d[1]
113
1146:
115        TBZ     x1, 1, 7f
116        STR     s16, [x6], 4
117        DUP     s16, v16.s[1]
118
1197:
120        TBZ     x1, 0, 8f
121        STR     h16, [x6]
1228:
123        RET
124
125END_FUNCTION xnn_f16_gemm_minmax_ukernel_1x16__aarch64_neonfp16arith_ld32
126
127#ifdef __ELF__
128.section ".note.GNU-stack","",%progbits
129#endif
130