blob: 73a2c9e5a483403e1ae415bf9acb2c433e7b1e88 [file] [log] [blame]
Marat Dukhan403b7d42019-12-05 12:49:11 -08001// Auto-generated file. Do not edit!
2// Template: src/f32-vbinary/vop-neon.c.in
3// Generator: tools/xngen
4//
5// Copyright 2019 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 <assert.h>
11
12#include <arm_neon.h>
13
14#include <xnnpack/common.h>
15#include <xnnpack/vbinary.h>
16
17
18void xnn_f32_vmax_ukernel__neon_x4(
19 size_t n,
20 const float* a,
21 const float* b,
22 float* y,
Marat Dukhanb2217dd2020-05-28 17:30:28 -070023 const union xnn_f32_default_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
Marat Dukhan403b7d42019-12-05 12:49:11 -080024{
25 assert(n != 0);
26 assert(n % sizeof(float) == 0);
Frank Barchard0822dde2020-07-04 12:47:24 -070027 assert(a != NULL);
28 assert(b != NULL);
29 assert(y != NULL);
Marat Dukhan403b7d42019-12-05 12:49:11 -080030
Marat Dukhan403b7d42019-12-05 12:49:11 -080031
32 for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
33 const float32x4_t va0123 = vld1q_f32(a); a += 4;
34 const float32x4_t vb0123 = vld1q_f32(b); b += 4;
35
36 float32x4_t vy0123 = vmaxq_f32(va0123, vb0123);
37
Marat Dukhan403b7d42019-12-05 12:49:11 -080038
Marat Dukhan13bafb02020-06-05 00:43:11 -070039
Marat Dukhan403b7d42019-12-05 12:49:11 -080040 vst1q_f32(y, vy0123); y += 4;
41 }
Marat Dukhan403b7d42019-12-05 12:49:11 -080042 if XNN_UNLIKELY(n != 0) {
43 const float32x4_t va0123 = vld1q_f32(a);
44 const float32x4_t vb0123 = vld1q_f32(b);
45
46 float32x4_t vy0123 = vmaxq_f32(va0123, vb0123);
Marat Dukhan403b7d42019-12-05 12:49:11 -080047
48 float32x2_t vy01 = vget_low_f32(vy0123);
49 if (n & (2 * sizeof(float))) {
50 vst1_f32(y, vy01); y += 2;
51 vy01 = vget_high_f32(vy0123);
52 }
53 if (n & (1 * sizeof(float))) {
54 vst1_lane_f32(y, vy01, 0);
55 }
56 }
57}