1// Copyright 2020 Google LLC
2//
3// This source code is licensed under the BSD-style license found in the
4// LICENSE file in the root directory of this source tree.
5
6$assert BATCH_TILE % 8 == 0
7$assert BATCH_TILE >= 8
8$ABC = "01234567456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
9$assert OP in ["ADD", "DIV", "MAX", "MIN", "MUL", "SUB", "SQRDIFF"]
10$assert ACTIVATION in ["LINEAR", "MINMAX"]
11#include <assert.h>
12
13#include <arm_neon.h>
14
15#include <xnnpack/common.h>
16#include <xnnpack/vbinary.h>
17
18
19$VOPQ_f16 = {
20$  "ADD": lambda x, y: "vaddq_f16(%s, %s)" % (x, y),
21$  "DIV": lambda x, y: "vdivq_f16(%s, %s)" % (x, y),
22$  "MAX": lambda x, y: "vmaxq_f16(%s, %s)" % (x, y),
23$  "MIN": lambda x, y: "vminq_f16(%s, %s)" % (x, y),
24$  "MUL": lambda x, y: "vmulq_f16(%s, %s)" % (x, y),
25$  "SUB": lambda x, y: "vsubq_f16(%s, %s)" % (x, y),
26$  "SQRDIFF": lambda x, y: "vsubq_f16(%s, %s)" % (x, y),
27$}[OP]
28$SUFFIX = {"LINEAR": "", "MINMAX": "_minmax"}[ACTIVATION]
29$PARAMS = {"LINEAR": "xnn_f16_default_params", "MINMAX": "xnn_f16_minmax_params"}[ACTIVATION]
30void xnn_f16_v${OP.lower()}${SUFFIX}_ukernel__neonfp16arith_x${BATCH_TILE}(
31    size_t n,
32    const void* restrict a_ptr,
33    const void* restrict b_ptr,
34    void* restrict y_ptr,
35    const struct ${PARAMS} params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
36{
37  assert(n != 0);
38  assert(n % sizeof(__fp16) == 0);
39  assert(a_ptr != NULL);
40  assert(b_ptr != NULL);
41  assert(y_ptr != NULL);
42
43  const __fp16* a = (const __fp16*) a_ptr;
44  const __fp16* b = (const __fp16*) b_ptr;
45  __fp16* y = (__fp16*) y_ptr;
46
47  $if ACTIVATION == "MINMAX":
48    const float16x8_t vy_min = vld1q_dup_f16(&params->min);
49    const float16x8_t vy_max = vld1q_dup_f16(&params->max);
50
51  for (; n >= ${BATCH_TILE} * sizeof(__fp16); n -= ${BATCH_TILE} * sizeof(__fp16)) {
52    $for N in range(0, BATCH_TILE, 8):
53      const float16x8_t va${ABC[N:N+8]} = vld1q_f16(a); a += 8;
54      const float16x8_t vb${ABC[N:N+8]} = vld1q_f16(b); b += 8;
55
56    $for N in range(0, BATCH_TILE, 8):
57      float16x8_t vy${ABC[N:N+8]} = ${VOPQ_f16("va" + ABC[N:N+8], "vb" + ABC[N:N+8])};
58
59    $if OP == "SQRDIFF":
60      $for N in range(0, BATCH_TILE, 8):
61        vy${ABC[N:N+8]} = vmulq_f16(vy${ABC[N:N+8]}, vy${ABC[N:N+8]});
62
63    $if ACTIVATION == "MINMAX":
64      $for N in range(0, BATCH_TILE, 8):
65        vy${ABC[N:N+8]} = vmaxq_f16(vy${ABC[N:N+8]}, vy_min);
66
67      $for N in range(0, BATCH_TILE, 8):
68        vy${ABC[N:N+8]} = vminq_f16(vy${ABC[N:N+8]}, vy_max);
69
70    $for N in range(0, BATCH_TILE, 8):
71      vst1q_f16(y, vy${ABC[N:N+8]}); y += 8;
72  }
73  $if BATCH_TILE > 8:
74    for (; n >= 8 * sizeof(__fp16); n -= 8 * sizeof(__fp16)) {
75      const float16x8_t va01234567 = vld1q_f16(a); a += 8;
76      const float16x8_t vb01234567 = vld1q_f16(b); b += 8;
77
78      float16x8_t vy01234567 = ${VOPQ_f16("va01234567", "vb01234567")};
79      $if OP == "SQRDIFF":
80        vy01234567 = vmulq_f16(vy01234567, vy01234567);
81      $if ACTIVATION == "MINMAX":
82        vy01234567 = vmaxq_f16(vy01234567, vy_min);
83        vy01234567 = vminq_f16(vy01234567, vy_max);
84      vst1q_f16(y, vy01234567); y += 8;
85    }
86  if XNN_UNLIKELY(n != 0) {
87    const float16x8_t va01234567 = vld1q_f16(a);
88    const float16x8_t vb01234567 = vld1q_f16(b);
89
90    float16x8_t vy01234567 = ${VOPQ_f16("va01234567", "vb01234567")};
91    $if OP == "SQRDIFF":
92      vy01234567 = vmulq_f16(vy01234567, vy01234567);
93    $if ACTIVATION == "MINMAX":
94      vy01234567 = vmaxq_f16(vy01234567, vy_min);
95      vy01234567 = vminq_f16(vy01234567, vy_max);
96
97    float16x4_t vy0123 = vget_low_f16(vy01234567);
98    if (n & (4 * sizeof(__fp16))) {
99      vst1_f16(y, vy0123); y += 4;
100      vy0123 = vget_high_f16(vy01234567);
101    }
102
103    if (n & (2 * sizeof(__fp16))) {
104      vst1_lane_u32(__builtin_assume_aligned(y, 1), vreinterpret_u32_f16(vy0123), 0); y += 2;
105      vy0123 = vext_f16(vy0123, vy0123, 2);
106    }
107
108    if (n & (1 * sizeof(__fp16))) {
109      vst1_lane_f16(y, vy0123, 0);
110    }
111  }
112}
113