xref: /aosp_15_r20/external/XNNPACK/src/f16-gemm/gen/1x16-minmax-aarch64-neonfp16arith-ld32.S (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1*4bdc9457SAndroid Build Coastguard Worker// Auto-generated file. Do not edit!
2*4bdc9457SAndroid Build Coastguard Worker//   Template: src/f16-gemm/1x16-aarch64-neonfp16arith-ld32.S.in
3*4bdc9457SAndroid Build Coastguard Worker//   Generator: tools/xngen
4*4bdc9457SAndroid Build Coastguard Worker//
5*4bdc9457SAndroid Build Coastguard Worker// Copyright 2020 Google LLC
6*4bdc9457SAndroid Build Coastguard Worker//
7*4bdc9457SAndroid Build Coastguard Worker// This source code is licensed under the BSD-style license found in the
8*4bdc9457SAndroid Build Coastguard Worker// LICENSE file in the root directory of this source tree.
9*4bdc9457SAndroid Build Coastguard Worker
10*4bdc9457SAndroid Build Coastguard Worker#include <xnnpack/assembly.h>
11*4bdc9457SAndroid Build Coastguard Worker
12*4bdc9457SAndroid Build Coastguard Worker# void xnn_f16_gemm_minmax_ukernel_1x16__aarch64_neonfp16arith_ld32(
13*4bdc9457SAndroid Build Coastguard Worker#     size_t mr,                (x0) - unused.  mr = 1
14*4bdc9457SAndroid Build Coastguard Worker#     size_t nc,                x1
15*4bdc9457SAndroid Build Coastguard Worker#     size_t kc,                x2 / x0
16*4bdc9457SAndroid Build Coastguard Worker#     const void*restrict a,    x3
17*4bdc9457SAndroid Build Coastguard Worker#     size_t a_stride,          (x4) - unused
18*4bdc9457SAndroid Build Coastguard Worker#     const void*restrict w,    x5
19*4bdc9457SAndroid Build Coastguard Worker#     void*restrict c,          x6
20*4bdc9457SAndroid Build Coastguard Worker#     size_t cm_stride,         (x7) - unused
21*4bdc9457SAndroid Build Coastguard Worker#     size_t cn_stride,         [sp] -> x14
22*4bdc9457SAndroid Build Coastguard Worker#     const union xnn_f16_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])  [sp + 8] -> (x8)
23*4bdc9457SAndroid Build Coastguard Worker
24*4bdc9457SAndroid Build Coastguard Worker# d8-d15, x19-x30 need to be preserved if used. x18 is reserved by the OS.
25*4bdc9457SAndroid Build Coastguard Worker
26*4bdc9457SAndroid Build Coastguard Worker# A pointer
27*4bdc9457SAndroid Build Coastguard Worker# x3  a0
28*4bdc9457SAndroid Build Coastguard Worker
29*4bdc9457SAndroid Build Coastguard Worker# C pointer
30*4bdc9457SAndroid Build Coastguard Worker# x6  c0
31*4bdc9457SAndroid Build Coastguard Worker
32*4bdc9457SAndroid Build Coastguard Worker# Clamp v4, v5
33*4bdc9457SAndroid Build Coastguard Worker
34*4bdc9457SAndroid Build Coastguard WorkerBEGIN_FUNCTION xnn_f16_gemm_minmax_ukernel_1x16__aarch64_neonfp16arith_ld32
35*4bdc9457SAndroid Build Coastguard Worker
36*4bdc9457SAndroid Build Coastguard Worker        # Load cn_stride, params pointer
37*4bdc9457SAndroid Build Coastguard Worker        LDP     x14, x8, [sp]
38*4bdc9457SAndroid Build Coastguard Worker
39*4bdc9457SAndroid Build Coastguard Worker        # Load params values
40*4bdc9457SAndroid Build Coastguard Worker        LD2R    {v4.8h, v5.8h}, [x8]
41*4bdc9457SAndroid Build Coastguard Worker0:
42*4bdc9457SAndroid Build Coastguard Worker        # Load initial bias from w into accumulators
43*4bdc9457SAndroid Build Coastguard Worker        LDP     q16, q17, [x5], 32
44*4bdc9457SAndroid Build Coastguard Worker
45*4bdc9457SAndroid Build Coastguard Worker        MOVI    v18.8h, 0               // second set of C for pipelining FMLA
46*4bdc9457SAndroid Build Coastguard Worker        MOVI    v19.8h, 0
47*4bdc9457SAndroid Build Coastguard Worker
48*4bdc9457SAndroid Build Coastguard Worker        # Is there at least 2 halffloats (4 bytes)
49*4bdc9457SAndroid Build Coastguard Worker        SUBS    x0, x2, 4               // k = kc - 4
50*4bdc9457SAndroid Build Coastguard Worker
51*4bdc9457SAndroid Build Coastguard Worker        B.LO    3f
52*4bdc9457SAndroid Build Coastguard Worker
53*4bdc9457SAndroid Build Coastguard Worker        # Main loop - 2 halffloats of A (4 bytes)
54*4bdc9457SAndroid Build Coastguard Worker1:
55*4bdc9457SAndroid Build Coastguard Worker        LDR     s0, [x3], 4
56*4bdc9457SAndroid Build Coastguard Worker        LDR     q20, [x5, 0]
57*4bdc9457SAndroid Build Coastguard Worker        LDR     q21, [x5, 16]
58*4bdc9457SAndroid Build Coastguard Worker        LDR     q22, [x5, 32]
59*4bdc9457SAndroid Build Coastguard Worker        LDR     q23, [x5, 48]
60*4bdc9457SAndroid Build Coastguard Worker        SUBS    x0, x0, 4
61*4bdc9457SAndroid Build Coastguard Worker        FMLA    v16.8h, v20.8h, v0.h[0]
62*4bdc9457SAndroid Build Coastguard Worker        FMLA    v17.8h, v21.8h, v0.h[0]
63*4bdc9457SAndroid Build Coastguard Worker        FMLA    v18.8h, v22.8h, v0.h[1]
64*4bdc9457SAndroid Build Coastguard Worker        FMLA    v19.8h, v23.8h, v0.h[1]
65*4bdc9457SAndroid Build Coastguard Worker        ADD     x5, x5, 64
66*4bdc9457SAndroid Build Coastguard Worker        B.HS    1b
67*4bdc9457SAndroid Build Coastguard Worker
68*4bdc9457SAndroid Build Coastguard Worker        # Is there a remainder?- 1 halffloat of A (2 bytes)
69*4bdc9457SAndroid Build Coastguard Worker        TBNZ    x0, 1, 3f
70*4bdc9457SAndroid Build Coastguard Worker
71*4bdc9457SAndroid Build Coastguard Worker2:
72*4bdc9457SAndroid Build Coastguard Worker        FADD    v16.8h, v16.8h, v18.8h
73*4bdc9457SAndroid Build Coastguard Worker        FADD    v17.8h, v17.8h, v19.8h
74*4bdc9457SAndroid Build Coastguard Worker        SUBS    x1, x1, 16
75*4bdc9457SAndroid Build Coastguard Worker
76*4bdc9457SAndroid Build Coastguard Worker        # Clamp
77*4bdc9457SAndroid Build Coastguard Worker        FMAX    v16.8h, v16.8h, v4.8h
78*4bdc9457SAndroid Build Coastguard Worker        FMAX    v17.8h, v17.8h, v4.8h
79*4bdc9457SAndroid Build Coastguard Worker        FMIN    v16.8h, v16.8h, v5.8h
80*4bdc9457SAndroid Build Coastguard Worker        FMIN    v17.8h, v17.8h, v5.8h
81*4bdc9457SAndroid Build Coastguard Worker
82*4bdc9457SAndroid Build Coastguard Worker        # Store full 1 x 16
83*4bdc9457SAndroid Build Coastguard Worker        B.LO    4f
84*4bdc9457SAndroid Build Coastguard Worker
85*4bdc9457SAndroid Build Coastguard Worker        STP     q16, q17, [x6]
86*4bdc9457SAndroid Build Coastguard Worker        ADD     x6, x6, x14
87*4bdc9457SAndroid Build Coastguard Worker
88*4bdc9457SAndroid Build Coastguard Worker        SUB     x3,  x3, x2             // a0 -= kc
89*4bdc9457SAndroid Build Coastguard Worker
90*4bdc9457SAndroid Build Coastguard Worker        B.HI    0b
91*4bdc9457SAndroid Build Coastguard Worker
92*4bdc9457SAndroid Build Coastguard Worker        RET
93*4bdc9457SAndroid Build Coastguard Worker
94*4bdc9457SAndroid Build Coastguard Worker3:
95*4bdc9457SAndroid Build Coastguard Worker        # Remainder- 1 halffloat of A (2 bytes)
96*4bdc9457SAndroid Build Coastguard Worker        LDR     q20, [x5], 16
97*4bdc9457SAndroid Build Coastguard Worker        LDR     q21, [x5], 16
98*4bdc9457SAndroid Build Coastguard Worker        LDR     h0, [x3], 2
99*4bdc9457SAndroid Build Coastguard Worker        FMLA    v16.8h, v20.8h, v0.h[0]
100*4bdc9457SAndroid Build Coastguard Worker        FMLA    v17.8h, v21.8h, v0.h[0]
101*4bdc9457SAndroid Build Coastguard Worker        B       2b
102*4bdc9457SAndroid Build Coastguard Worker
103*4bdc9457SAndroid Build Coastguard Worker        # Store odd channels
104*4bdc9457SAndroid Build Coastguard Worker4:
105*4bdc9457SAndroid Build Coastguard Worker        TBZ     x1, 3, 5f
106*4bdc9457SAndroid Build Coastguard Worker        STR     q16, [x6], 16
107*4bdc9457SAndroid Build Coastguard Worker        MOV     v16.16b, v17.16b
108*4bdc9457SAndroid Build Coastguard Worker
109*4bdc9457SAndroid Build Coastguard Worker5:
110*4bdc9457SAndroid Build Coastguard Worker        TBZ     x1, 2, 6f
111*4bdc9457SAndroid Build Coastguard Worker        STR     d16, [x6], 8
112*4bdc9457SAndroid Build Coastguard Worker        DUP     d16, v16.d[1]
113*4bdc9457SAndroid Build Coastguard Worker
114*4bdc9457SAndroid Build Coastguard Worker6:
115*4bdc9457SAndroid Build Coastguard Worker        TBZ     x1, 1, 7f
116*4bdc9457SAndroid Build Coastguard Worker        STR     s16, [x6], 4
117*4bdc9457SAndroid Build Coastguard Worker        DUP     s16, v16.s[1]
118*4bdc9457SAndroid Build Coastguard Worker
119*4bdc9457SAndroid Build Coastguard Worker7:
120*4bdc9457SAndroid Build Coastguard Worker        TBZ     x1, 0, 8f
121*4bdc9457SAndroid Build Coastguard Worker        STR     h16, [x6]
122*4bdc9457SAndroid Build Coastguard Worker8:
123*4bdc9457SAndroid Build Coastguard Worker        RET
124*4bdc9457SAndroid Build Coastguard Worker
125*4bdc9457SAndroid Build Coastguard WorkerEND_FUNCTION xnn_f16_gemm_minmax_ukernel_1x16__aarch64_neonfp16arith_ld32
126*4bdc9457SAndroid Build Coastguard Worker
127*4bdc9457SAndroid Build Coastguard Worker#ifdef __ELF__
128*4bdc9457SAndroid Build Coastguard Worker.section ".note.GNU-stack","",%progbits
129*4bdc9457SAndroid Build Coastguard Worker#endif
130