1 // Auto-generated file. Do not edit!
2 // Template: src/f32-vlrelu/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/common.h>
15 #include <xnnpack/vunary.h>
16
17
xnn_f32_vlrelu_ukernel__neon_x4(size_t n,const float * x,float * y,const union xnn_f32_lrelu_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_f32_vlrelu_ukernel__neon_x4(
19 size_t n,
20 const float* x,
21 float* y,
22 const union xnn_f32_lrelu_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
23 {
24 assert(n != 0);
25 assert(n % sizeof(float) == 0);
26
27 const float32x4_t vslope = vld1q_dup_f32(¶ms->scalar.slope);
28
29 for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
30 const float32x4_t vx0123 = vld1q_f32(x); x += 4;
31
32 float32x4_t vacc0123 = vmulq_f32(vx0123, vslope);
33 const uint32x4_t vmask0123 = vcltq_s32(vreinterpretq_s32_f32(vx0123), vmovq_n_s32(0));
34
35 vacc0123 = vbslq_f32(vmask0123, vacc0123, vx0123);
36
37 vst1q_f32(y, vacc0123); y += 4;
38 }
39 if XNN_UNLIKELY(n != 0) {
40 const float32x4_t vx = vld1q_f32(x);
41 float32x4_t vacc = vmulq_f32(vx, vslope);
42 const uint32x4_t vmask = vcltq_s32(vreinterpretq_s32_f32(vx), vmovq_n_s32(0));
43 vacc = vbslq_f32(vmask, vacc, vx);
44
45 float32x2_t vacc_lo = vget_low_f32(vacc);
46 if (n & (2 * sizeof(float))) {
47 vst1_f32(y, vacc_lo); y += 2;
48 vacc_lo = vget_high_f32(vacc);
49 }
50 if (n & (1 * sizeof(float))) {
51 vst1_lane_f32(y, vacc_lo, 0);
52 }
53 }
54 }
55