xref: /aosp_15_r20/external/XNNPACK/src/u8-rmax/neon.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1*4bdc9457SAndroid Build Coastguard Worker // Copyright (c) Facebook, Inc. and its affiliates.
2*4bdc9457SAndroid Build Coastguard Worker // All rights reserved.
3*4bdc9457SAndroid Build Coastguard Worker //
4*4bdc9457SAndroid Build Coastguard Worker // Copyright 2019 Google LLC
5*4bdc9457SAndroid Build Coastguard Worker //
6*4bdc9457SAndroid Build Coastguard Worker // This source code is licensed under the BSD-style license found in the
7*4bdc9457SAndroid Build Coastguard Worker // LICENSE file in the root directory of this source tree.
8*4bdc9457SAndroid Build Coastguard Worker 
9*4bdc9457SAndroid Build Coastguard Worker #include <assert.h>
10*4bdc9457SAndroid Build Coastguard Worker 
11*4bdc9457SAndroid Build Coastguard Worker #include <arm_neon.h>
12*4bdc9457SAndroid Build Coastguard Worker 
13*4bdc9457SAndroid Build Coastguard Worker #include <xnnpack/rmax.h>
14*4bdc9457SAndroid Build Coastguard Worker 
15*4bdc9457SAndroid Build Coastguard Worker 
xnn_u8_rmax_ukernel__neon(size_t n,const uint8_t * x,uint8_t * y)16*4bdc9457SAndroid Build Coastguard Worker void xnn_u8_rmax_ukernel__neon(
17*4bdc9457SAndroid Build Coastguard Worker     size_t n,
18*4bdc9457SAndroid Build Coastguard Worker     const uint8_t* x,
19*4bdc9457SAndroid Build Coastguard Worker     uint8_t* y)
20*4bdc9457SAndroid Build Coastguard Worker {
21*4bdc9457SAndroid Build Coastguard Worker   assert(n != 0);
22*4bdc9457SAndroid Build Coastguard Worker 
23*4bdc9457SAndroid Build Coastguard Worker   if XNN_LIKELY(n >= 16) {
24*4bdc9457SAndroid Build Coastguard Worker     uint8x16_t vmax = vmovq_n_u8(0);
25*4bdc9457SAndroid Build Coastguard Worker     do {
26*4bdc9457SAndroid Build Coastguard Worker       const uint8x16_t vx = vld1q_u8(x); x += 16;
27*4bdc9457SAndroid Build Coastguard Worker       vmax = vmaxq_u8(vmax, vx);
28*4bdc9457SAndroid Build Coastguard Worker       n -= 16;
29*4bdc9457SAndroid Build Coastguard Worker     } while (n >= 16);
30*4bdc9457SAndroid Build Coastguard Worker     if (n != 0) {
31*4bdc9457SAndroid Build Coastguard Worker       const size_t x_increment = n - 16;
32*4bdc9457SAndroid Build Coastguard Worker       x = (const uint8_t*) ((uintptr_t) x + x_increment);
33*4bdc9457SAndroid Build Coastguard Worker       const uint8x16_t vx = vld1q_u8(x);
34*4bdc9457SAndroid Build Coastguard Worker       vmax = vmaxq_u8(vmax, vx);
35*4bdc9457SAndroid Build Coastguard Worker     }
36*4bdc9457SAndroid Build Coastguard Worker     uint8x8_t vmax8 = vmax_u8(vget_low_u8(vmax), vget_high_u8(vmax));
37*4bdc9457SAndroid Build Coastguard Worker     const uint8x8_t vmax4 = vpmax_u8(vmax8, vmax8);
38*4bdc9457SAndroid Build Coastguard Worker     const uint8x8_t vmax2 = vpmax_u8(vmax4, vmax4);
39*4bdc9457SAndroid Build Coastguard Worker     const uint8x8_t vmax1 = vpmax_u8(vmax2, vmax2);
40*4bdc9457SAndroid Build Coastguard Worker     vst1_lane_u8(y, vmax1, 0);
41*4bdc9457SAndroid Build Coastguard Worker   } else {
42*4bdc9457SAndroid Build Coastguard Worker     uint8x8_t vmax = vmov_n_u8(0);
43*4bdc9457SAndroid Build Coastguard Worker     do {
44*4bdc9457SAndroid Build Coastguard Worker       const uint8x8_t vx = vld1_dup_u8(x); x += 1;
45*4bdc9457SAndroid Build Coastguard Worker       vmax = vmax_u8(vmax, vx);
46*4bdc9457SAndroid Build Coastguard Worker     } while (--n != 0);
47*4bdc9457SAndroid Build Coastguard Worker     vst1_lane_u8(y, vmax, 0);
48*4bdc9457SAndroid Build Coastguard Worker   }
49*4bdc9457SAndroid Build Coastguard Worker }
50