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 <emmintrin.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__sse2(size_t n,const uint8_t * x,uint8_t * y)16*4bdc9457SAndroid Build Coastguard Worker void xnn_u8_rmax_ukernel__sse2(
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 __m128i vmax = _mm_setzero_si128();
25*4bdc9457SAndroid Build Coastguard Worker do {
26*4bdc9457SAndroid Build Coastguard Worker const __m128i vx = _mm_loadu_si128((const __m128i*) x);
27*4bdc9457SAndroid Build Coastguard Worker x += 16;
28*4bdc9457SAndroid Build Coastguard Worker vmax = _mm_max_epu8(vmax, vx);
29*4bdc9457SAndroid Build Coastguard Worker n -= 16;
30*4bdc9457SAndroid Build Coastguard Worker } while (n >= 16);
31*4bdc9457SAndroid Build Coastguard Worker if (n != 0) {
32*4bdc9457SAndroid Build Coastguard Worker const size_t x_increment = n - 16;
33*4bdc9457SAndroid Build Coastguard Worker x = (const uint8_t*) ((uintptr_t) x + x_increment);
34*4bdc9457SAndroid Build Coastguard Worker const __m128i vx = _mm_loadu_si128((const __m128i*) x);
35*4bdc9457SAndroid Build Coastguard Worker vmax = _mm_max_epu8(vmax, vx);
36*4bdc9457SAndroid Build Coastguard Worker }
37*4bdc9457SAndroid Build Coastguard Worker vmax = _mm_max_epu8(vmax, _mm_unpackhi_epi64(vmax, vmax));
38*4bdc9457SAndroid Build Coastguard Worker vmax = _mm_max_epu8(vmax, _mm_srli_epi64(vmax, 32));
39*4bdc9457SAndroid Build Coastguard Worker vmax = _mm_max_epu8(vmax, _mm_srli_epi32(vmax, 16));
40*4bdc9457SAndroid Build Coastguard Worker vmax = _mm_max_epu8(vmax, _mm_srli_epi16(vmax, 8));
41*4bdc9457SAndroid Build Coastguard Worker *y = (uint8_t) _mm_cvtsi128_si32(vmax);
42*4bdc9457SAndroid Build Coastguard Worker } else {
43*4bdc9457SAndroid Build Coastguard Worker uint8_t vmax = 0;
44*4bdc9457SAndroid Build Coastguard Worker do {
45*4bdc9457SAndroid Build Coastguard Worker const uint8_t vx = *x++;
46*4bdc9457SAndroid Build Coastguard Worker vmax = vx > vmax ? vx : vmax;
47*4bdc9457SAndroid Build Coastguard Worker } while (--n != 0);
48*4bdc9457SAndroid Build Coastguard Worker *y = vmax;
49*4bdc9457SAndroid Build Coastguard Worker }
50*4bdc9457SAndroid Build Coastguard Worker }
51