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_x4_ukernel__sse2(size_t n,const uint8_t * input,uint8_t * output)14 void xnn_x8_zip_x4_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 const uint8_t* z = (const uint8_t*) ((uintptr_t) y + n);
22 const uint8_t* w = (const uint8_t*) ((uintptr_t) z + n);
23 uint8_t* o = output;
24
25 if (n >= 16) {
26 do {
27 const __m128i vx = _mm_loadu_si128((const __m128i*) x);
28 x += 16;
29 const __m128i vy = _mm_loadu_si128((const __m128i*) y);
30 y += 16;
31 const __m128i vz = _mm_loadu_si128((const __m128i*) z);
32 z += 16;
33 const __m128i vw = _mm_loadu_si128((const __m128i*) w);
34 w += 16;
35 const __m128i vxy_lo = _mm_unpacklo_epi8(vx, vy);
36 const __m128i vxy_hi = _mm_unpackhi_epi8(vx, vy);
37 const __m128i vzw_lo = _mm_unpacklo_epi8(vz, vw);
38 const __m128i vzw_hi = _mm_unpackhi_epi8(vz, vw);
39 const __m128i vxyzw0 = _mm_unpacklo_epi16(vxy_lo, vzw_lo);
40 const __m128i vxyzw1 = _mm_unpackhi_epi16(vxy_lo, vzw_lo);
41 const __m128i vxyzw2 = _mm_unpacklo_epi16(vxy_hi, vzw_hi);
42 const __m128i vxyzw3 = _mm_unpackhi_epi16(vxy_hi, vzw_hi);
43 _mm_storeu_si128((__m128i*) o, vxyzw0);
44 _mm_storeu_si128((__m128i*) o + 1, vxyzw1);
45 _mm_storeu_si128((__m128i*) o + 2, vxyzw2);
46 _mm_storeu_si128((__m128i*) o + 3, vxyzw3);
47 o = (void*) ((uintptr_t) o + 64);
48 n -= 16;
49 } while (n >= 16);
50 if (n != 0) {
51 const size_t address_increment = n - 16;
52 const __m128i vx = _mm_loadu_si128((const __m128i*) ((uintptr_t) x + address_increment));
53 const __m128i vy = _mm_loadu_si128((const __m128i*) ((uintptr_t) y + address_increment));
54 const __m128i vz = _mm_loadu_si128((const __m128i*) ((uintptr_t) z + address_increment));
55 const __m128i vw = _mm_loadu_si128((const __m128i*) ((uintptr_t) w + address_increment));
56 const __m128i vxy_lo = _mm_unpacklo_epi8(vx, vy);
57 const __m128i vxy_hi = _mm_unpackhi_epi8(vx, vy);
58 const __m128i vzw_lo = _mm_unpacklo_epi8(vz, vw);
59 const __m128i vzw_hi = _mm_unpackhi_epi8(vz, vw);
60 const __m128i vxyzw0 = _mm_unpacklo_epi16(vxy_lo, vzw_lo);
61 const __m128i vxyzw1 = _mm_unpackhi_epi16(vxy_lo, vzw_lo);
62 const __m128i vxyzw2 = _mm_unpacklo_epi16(vxy_hi, vzw_hi);
63 const __m128i vxyzw3 = _mm_unpackhi_epi16(vxy_hi, vzw_hi);
64 o = (void*) ((uintptr_t) o + address_increment * 4);
65 _mm_storeu_si128((__m128i*) o, vxyzw0);
66 _mm_storeu_si128((__m128i*) o + 1, vxyzw1);
67 _mm_storeu_si128((__m128i*) o + 2, vxyzw2);
68 _mm_storeu_si128((__m128i*) o + 3, vxyzw3);
69 }
70 } else {
71 do {
72 const uint8_t vx = *x++;
73 const uint8_t vy = *y++;
74 const uint8_t vz = *z++;
75 const uint8_t vw = *w++;
76 o[0] = vx;
77 o[1] = vy;
78 o[2] = vz;
79 o[3] = vw;
80 o += 4;
81 } while (--n != 0);
82 }
83 }
84