1 // Copyright 2019 Google LLC
2 //
3 // This source code is licensed under the BSD-style license found in the
4 // LICENSE file in the root directory of this source tree.
5
6 #include <assert.h>
7
8 #include <emmintrin.h>
9
10 #include <xnnpack/zip.h>
11
12
xnn_x32_zip_x4_ukernel__sse2(size_t n,const uint32_t * input,uint32_t * output)13 void xnn_x32_zip_x4_ukernel__sse2(
14 size_t n,
15 const uint32_t* input,
16 uint32_t* output)
17 {
18 assert(n != 0);
19 assert(n % 4 == 0);
20
21 const uint32_t* x = input;
22 const uint32_t* y = (const uint32_t*) ((uintptr_t) x + n);
23 const uint32_t* z = (const uint32_t*) ((uintptr_t) y + n);
24 const uint32_t* w = (const uint32_t*) ((uintptr_t) z + n);
25 uint32_t* o = output;
26
27 while (n >= 16) {
28 const __m128i vx = _mm_loadu_si128((const __m128i*) x);
29 x += 4;
30 const __m128i vy = _mm_loadu_si128((const __m128i*) y);
31 y += 4;
32 const __m128i vz = _mm_loadu_si128((const __m128i*) z);
33 z += 4;
34 const __m128i vw = _mm_loadu_si128((const __m128i*) w);
35 w += 4;
36
37 const __m128i vxy_lo = _mm_unpacklo_epi32(vx, vy);
38 const __m128i vxy_hi = _mm_unpackhi_epi32(vx, vy);
39 const __m128i vzw_lo = _mm_unpacklo_epi32(vz, vw);
40 const __m128i vzw_hi = _mm_unpackhi_epi32(vz, vw);
41
42 const __m128i vxyzw0 = _mm_unpacklo_epi64(vxy_lo, vzw_lo);
43 const __m128i vxyzw1 = _mm_unpackhi_epi64(vxy_lo, vzw_lo);
44 const __m128i vxyzw2 = _mm_unpacklo_epi64(vxy_hi, vzw_hi);
45 const __m128i vxyzw3 = _mm_unpackhi_epi64(vxy_hi, vzw_hi);
46
47 _mm_storeu_si128((__m128i*) o, vxyzw0);
48 _mm_storeu_si128((__m128i*) (o + 4), vxyzw1);
49 _mm_storeu_si128((__m128i*) (o + 8), vxyzw2);
50 _mm_storeu_si128((__m128i*) (o + 12), vxyzw3);
51 o += 16;
52 n -= 16;
53 }
54 if XNN_UNLIKELY(n != 0) {
55 if (n & 8) {
56 const __m128i vx = _mm_loadl_epi64((const __m128i*) x);
57 x += 2;
58 const __m128i vy = _mm_loadl_epi64((const __m128i*) y);
59 y += 2;
60 const __m128i vz = _mm_loadl_epi64((const __m128i*) z);
61 z += 2;
62 const __m128i vw = _mm_loadl_epi64((const __m128i*) w);
63 w += 2;
64
65 const __m128i vxy = _mm_unpacklo_epi32(vx, vy);
66 const __m128i vzw = _mm_unpacklo_epi32(vz, vw);
67
68 const __m128i vxyzw_lo = _mm_unpacklo_epi64(vxy, vzw);
69 const __m128i vxyzw_hi = _mm_unpackhi_epi64(vxy, vzw);
70
71 _mm_storeu_si128((__m128i*) o, vxyzw_lo);
72 _mm_storeu_si128((__m128i*) (o + 4), vxyzw_hi);
73 o += 8;
74 }
75 if (n & 4) {
76 const uint32_t vx = *x;
77 const uint32_t vy = *y;
78 const uint32_t vz = *z;
79 const uint32_t vw = *w;
80 o[0] = vx;
81 o[1] = vy;
82 o[2] = vz;
83 o[3] = vw;
84 }
85 }
86 }
87