1 // Copyright (c) Facebook, Inc. and its affiliates.
2 // All rights reserved.
3 //
4 // Copyright 2019 Google LLC
5 //
6 // This source code is licensed under the BSD-style license found in the
7 // LICENSE file in the root directory of this source tree.
8
9 #include <assert.h>
10
11 #include <arm_neon.h>
12
13 #include <xnnpack/clamp.h>
14
15
xnn_u8_clamp_ukernel__neon(size_t n,const uint8_t * x,uint8_t * y,const union xnn_u8_output_params params[restrict static1])16 void xnn_u8_clamp_ukernel__neon(
17 size_t n,
18 const uint8_t* x,
19 uint8_t* y,
20 const union xnn_u8_output_params params[restrict static 1])
21 {
22 assert(n != 0);
23
24 const uint8x16_t voutput_max = vld1q_dup_u8(¶ms->neon.max);
25 const uint8x16_t voutput_min = vld1q_dup_u8(¶ms->neon.min);
26
27 for (; n >= 64; n -= 64) {
28 const uint8x16_t vx0 = vld1q_u8(x); x += 16;
29 const uint8x16_t vx1 = vld1q_u8(x); x += 16;
30 const uint8x16_t vx2 = vld1q_u8(x); x += 16;
31 const uint8x16_t vx3 = vld1q_u8(x); x += 16;
32
33 const uint8x16_t vy0 = vminq_u8(vmaxq_u8(vx0, voutput_min), voutput_max);
34 const uint8x16_t vy1 = vminq_u8(vmaxq_u8(vx1, voutput_min), voutput_max);
35 const uint8x16_t vy2 = vminq_u8(vmaxq_u8(vx2, voutput_min), voutput_max);
36 const uint8x16_t vy3 = vminq_u8(vmaxq_u8(vx3, voutput_min), voutput_max);
37
38 __builtin_prefetch(x + 640);
39
40 vst1q_u8(y, vy0); y += 16;
41 vst1q_u8(y, vy1); y += 16;
42 vst1q_u8(y, vy2); y += 16;
43 vst1q_u8(y, vy3); y += 16;
44 }
45 for (; n >= 8; n -= 8) {
46 uint8x8_t vout = vld1_u8(x); x += 8;
47 vout = vmin_u8(vout, vget_low_u8(voutput_max));
48 vout = vmax_u8(vout, vget_low_u8(voutput_min));
49 vst1_u8(y, vout); y += 8;
50 }
51 if XNN_UNLIKELY(n != 0) {
52 uint8x8_t vout = vld1_u8(x);
53 vout = vmin_u8(vout, vget_low_u8(voutput_max));
54 vout = vmax_u8(vout, vget_low_u8(voutput_min));
55
56 if (n & 4) {
57 vst1_lane_u32(__builtin_assume_aligned(y, 1), vreinterpret_u32_u8(vout), 0); y += 4;
58 vout = vext_u8(vout, vout, 4);
59 }
60 if (n & 2) {
61 vst1_lane_u16(__builtin_assume_aligned(y, 1), vreinterpret_u16_u8(vout), 0); y += 2;
62 vout = vext_u8(vout, vout, 2);
63 }
64 if (n & 1) {
65 vst1_lane_u8(y, vout, 0);
66 }
67 }
68 }
69