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 <emmintrin.h>
10
11 #include <xnnpack/zip.h>
12
13
xnn_x8_zip_x2_ukernel__sse2(size_t n,const uint8_t * input,uint8_t * output)14 void xnn_x8_zip_x2_ukernel__sse2(
15 size_t n,
16 const uint8_t* input,
17 uint8_t* output)
18 {
19 const uint8_t* x = input;
20 const uint8_t* y = (const uint8_t*) ((uintptr_t) x + n);
21 uint8_t* o = output;
22
23 if (n >= 16) {
24 do {
25 const __m128i vx = _mm_loadu_si128((const __m128i*) x);
26 x += 16;
27 const __m128i vy = _mm_loadu_si128((const __m128i*) y);
28 y += 16;
29 const __m128i vxy_lo = _mm_unpacklo_epi8(vx, vy);
30 const __m128i vxy_hi = _mm_unpackhi_epi8(vx, vy);
31 _mm_storeu_si128((__m128i*) o, vxy_lo);
32 _mm_storeu_si128((__m128i*) (o + 16), vxy_hi);
33 o = (void*) ((uintptr_t) o + 32);
34 n -= 16;
35 } while (n >= 16);
36 if (n != 0) {
37 const size_t address_increment = n - 16;
38 const __m128i vx = _mm_loadu_si128((const __m128i*) ((uintptr_t) x + address_increment));
39 const __m128i vy = _mm_loadu_si128((const __m128i*) ((uintptr_t) y + address_increment));
40 const __m128i vxy_lo = _mm_unpacklo_epi8(vx, vy);
41 const __m128i vxy_hi = _mm_unpackhi_epi8(vx, vy);
42 o = (void*) ((uintptr_t) o + address_increment * 2);
43 _mm_storeu_si128((__m128i*) o, vxy_lo);
44 _mm_storeu_si128((__m128i*) o + 1, vxy_hi);
45 }
46 } else {
47 do {
48 const uint8_t vx = *x++;
49 const uint8_t vy = *y++;
50 o[0] = vx;
51 o[1] = vy;
52 o += 2;
53 } while (--n != 0);
54 }
55 }
56