1 // Auto-generated file. Do not edit!
2 // Template: src/f32-relu/neon.c.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 <assert.h>
11
12 #include <arm_neon.h>
13
14 #include <xnnpack/vunary.h>
15 #include <xnnpack/common.h>
16
17
xnn_f32_relu_ukernel__neon_x8(size_t n,const float * x,float * y,const union xnn_f32_relu_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_f32_relu_ukernel__neon_x8(
19 size_t n,
20 const float* x,
21 float* y,
22 const union xnn_f32_relu_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
23 {
24 assert(n != 0);
25 assert(n % sizeof(float) == 0);
26 assert(x != NULL);
27 assert(y != NULL);
28
29 const float32x4_t vzero = vmovq_n_f32(0.0f);
30
31 for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
32 float32x4_t vacc0123 = vld1q_f32(x); x += 4;
33 float32x4_t vacc4567 = vld1q_f32(x); x += 4;
34
35 vacc0123 = vmaxq_f32(vacc0123, vzero);
36 vacc4567 = vmaxq_f32(vacc4567, vzero);
37
38 vst1q_f32(y, vacc0123); y += 4;
39 vst1q_f32(y, vacc4567); y += 4;
40 }
41 for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
42 float32x4_t vacc = vld1q_f32(x); x += 4;
43 vacc = vmaxq_f32(vacc, vzero);
44 vst1q_f32(y, vacc); y += 4;
45 }
46 if XNN_UNLIKELY(n != 0) {
47 float32x4_t vacc = vld1q_f32(x);
48 vacc = vmaxq_f32(vacc, vzero);
49
50 float32x2_t vacc_lo = vget_low_f32(vacc);
51 if (n & (2 * sizeof(float))) {
52 vst1_f32(y, vacc_lo); y += 2;
53 vacc_lo = vget_high_f32(vacc);
54 }
55 if (n & (1 * sizeof(float))) {
56 vst1_lane_f32(y, vacc_lo, 0);
57 }
58 }
59 }
60