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_x8(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_x8(
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_DISABLE_TSAN
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 >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
30 const float32x4_t vx0123 = vld1q_f32(x); x += 4;
31 const float32x4_t vx4567 = vld1q_f32(x); x += 4;
32
33 float32x4_t vacc0123 = vmulq_f32(vx0123, vslope);
34 const uint32x4_t vmask0123 = vcltq_s32(vreinterpretq_s32_f32(vx0123), vmovq_n_s32(0));
35 float32x4_t vacc4567 = vmulq_f32(vx4567, vslope);
36 const uint32x4_t vmask4567 = vcltq_s32(vreinterpretq_s32_f32(vx4567), vmovq_n_s32(0));
37
38 vacc0123 = vbslq_f32(vmask0123, vacc0123, vx0123);
39 vacc4567 = vbslq_f32(vmask4567, vacc4567, vx4567);
40
41 vst1q_f32(y, vacc0123); y += 4;
42 vst1q_f32(y, vacc4567); y += 4;
43 }
44 for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
45 const float32x4_t vx = vld1q_f32(x); x += 4;
46 float32x4_t vacc = vmulq_f32(vx, vslope);
47 const uint32x4_t vmask = vcltq_s32(vreinterpretq_s32_f32(vx), vmovq_n_s32(0));
48 vacc = vbslq_f32(vmask, vacc, vx);
49 vst1q_f32(y, vacc); y += 4;
50 }
51 if XNN_UNLIKELY(n != 0) {
52 const float32x4_t vx = vld1q_f32(x);
53 float32x4_t vacc = vmulq_f32(vx, vslope);
54 const uint32x4_t vmask = vcltq_s32(vreinterpretq_s32_f32(vx), vmovq_n_s32(0));
55 vacc = vbslq_f32(vmask, vacc, vx);
56
57 float32x2_t vacc_lo = vget_low_f32(vacc);
58 if (n & (2 * sizeof(float))) {
59 vst1_f32(y, vacc_lo); y += 2;
60 vacc_lo = vget_high_f32(vacc);
61 }
62 if (n & (1 * sizeof(float))) {
63 vst1_lane_f32(y, vacc_lo, 0);
64 }
65 }
66 }
67