• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1// Copyright 2019 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$assert BATCH_TILE % 4 == 0
7$assert BATCH_TILE >= 4
8$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
9$assert OP in ["ADD", "DIV", "RDIV", "MAX", "MIN", "MUL", "SUB", "RSUB", "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_F32 = {
20$  "ADD": lambda x: "vaddq_f32(%s, vb)" % x,
21$  "DIV": lambda x: "vdivq_f32(%s, vb)" % x,
22$  "RDIV": lambda x: "vdivq_f32(vb, %s)" % x,
23$  "MAX": lambda x: "vmaxq_f32(%s, vb)" % x,
24$  "MIN": lambda x: "vminq_f32(%s, vb)" % x,
25$  "MUL": lambda x: "vmulq_f32(%s, vb)" % x,
26$  "SUB": lambda x: "vsubq_f32(%s, vb)" % x,
27$  "RSUB": lambda x: "vsubq_f32(vb, %s)" % x,
28$  "SQRDIFF": lambda x: "vsubq_f32(%s, vb)" % x,
29$}[OP]
30$SUFFIX = {"LINEAR": "", "MINMAX": "_minmax"}[ACTIVATION]
31$PARAMS = {"LINEAR": "xnn_f32_default_params", "MINMAX": "xnn_f32_minmax_params"}[ACTIVATION]
32void xnn_f32_v${OP.lower()}c${SUFFIX}_ukernel__neon_x${BATCH_TILE}(
33    size_t n,
34    const float* a,
35    const float* b,
36    float* y,
37    const union ${PARAMS} params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
38{
39  assert(n != 0);
40  assert(n % sizeof(float) == 0);
41  assert(a != NULL);
42  assert(b != NULL);
43  assert(y != NULL);
44
45  $if ACTIVATION == "MINMAX":
46    const float32x4_t vy_min = vld1q_dup_f32(&params->scalar.min);
47    const float32x4_t vy_max = vld1q_dup_f32(&params->scalar.max);
48
49  const float32x4_t vb = vld1q_dup_f32(b);
50  for (; n >= ${BATCH_TILE} * sizeof(float); n -= ${BATCH_TILE} * sizeof(float)) {
51    $for N in range(0, BATCH_TILE, 4):
52      const float32x4_t va${ABC[N:N+4]} = vld1q_f32(a); a += 4;
53
54    $for N in range(0, BATCH_TILE, 4):
55      float32x4_t vy${ABC[N:N+4]} = ${VOPQ_F32("va" + ABC[N:N+4])};
56
57    $if OP == "SQRDIFF":
58      $for N in range(0, BATCH_TILE, 4):
59        vy${ABC[N:N+4]} = vmulq_f32(vy${ABC[N:N+4]}, vy${ABC[N:N+4]});
60
61    $if ACTIVATION == "MINMAX":
62      $for N in range(0, BATCH_TILE, 4):
63        vy${ABC[N:N+4]} = vmaxq_f32(vy${ABC[N:N+4]}, vy_min);
64
65      $for N in range(0, BATCH_TILE, 4):
66        vy${ABC[N:N+4]} = vminq_f32(vy${ABC[N:N+4]}, vy_max);
67
68    $for N in range(0, BATCH_TILE, 4):
69      vst1q_f32(y, vy${ABC[N:N+4]}); y += 4;
70  }
71  $if BATCH_TILE > 4:
72    for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
73      const float32x4_t va0123 = vld1q_f32(a); a += 4;
74
75      float32x4_t vy0123 = ${VOPQ_F32("va0123")};
76      $if OP == "SQRDIFF":
77        vy0123 = vmulq_f32(vy0123, vy0123);
78      $if ACTIVATION == "MINMAX":
79        vy0123 = vmaxq_f32(vy0123, vy_min);
80        vy0123 = vminq_f32(vy0123, vy_max);
81      vst1q_f32(y, vy0123); y += 4;
82    }
83  if XNN_UNLIKELY(n != 0) {
84    const float32x4_t va0123 = vld1q_f32(a);
85
86    float32x4_t vy0123 = ${VOPQ_F32("va0123")};
87    $if OP == "SQRDIFF":
88      vy0123 = vmulq_f32(vy0123, vy0123);
89    $if ACTIVATION == "MINMAX":
90      vy0123 = vmaxq_f32(vy0123, vy_min);
91      vy0123 = vminq_f32(vy0123, vy_max);
92
93    float32x2_t vy01 = vget_low_f32(vy0123);
94    if (n & (2 * sizeof(float))) {
95      vst1_f32(y, vy01); y += 2;
96      vy01 = vget_high_f32(vy0123);
97    }
98    if (n & (1 * sizeof(float))) {
99      vst1_lane_f32(y, vy01, 0);
100    }
101  }
102}
103