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 #include <assert.h>
7
8 #include <arm_neon.h>
9
10 #include <xnnpack/clamp.h>
11 #include <xnnpack/common.h>
12
13
xnn_f32_clamp_ukernel__neon(size_t n,const float * x,float * y,const union xnn_f32_output_params params[restrict static1])14 void xnn_f32_clamp_ukernel__neon(
15 size_t n,
16 const float* x,
17 float* y,
18 const union xnn_f32_output_params params[restrict static 1])
19 {
20 assert(n != 0);
21 assert(n % sizeof(float) == 0);
22
23 #if XNN_ARCH_ARM64
24 const float32x4x2_t voutput_clamp = vld2q_dup_f32(¶ms->scalar.max);
25 const float32x4_t voutput_max = voutput_clamp.val[0];
26 const float32x4_t voutput_min = voutput_clamp.val[1];
27 #else
28 const float32x4_t voutput_max = vld1q_dup_f32(¶ms->scalar.max);
29 const float32x4_t voutput_min = vld1q_dup_f32(¶ms->scalar.min);
30 #endif
31
32 for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
33 const float32x4_t vx = vld1q_f32(x); x += 4;
34
35 float32x4_t vy = vminq_f32(vx, voutput_max);
36 vy = vmaxq_f32(vy, voutput_min);
37
38 vst1q_f32(y, vy); y += 4;
39 }
40 if (n != 0) {
41 const float32x4_t vx = vld1q_f32(x);
42
43 float32x4_t vy = vminq_f32(vx, voutput_max);
44 vy = vmaxq_f32(vy, voutput_min);
45
46 float32x2_t vy_lo = vget_low_f32(vy);
47 if (n & 2 * sizeof(float)) {
48 vst1_f32(y, vy_lo); y += 2;
49 vy_lo = vget_high_f32(vy);
50 }
51 if (n & 1 * sizeof(float)) {
52 vst1_lane_f32(y, vy_lo, 0);
53 }
54 }
55 }
56