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_x2_ukernel__sse2(size_t n,const uint32_t * input,uint32_t * output)13 void xnn_x32_zip_x2_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 uint32_t* o = output;
24
25 while (n >= 16) {
26 const __m128i vx = _mm_loadu_si128((const __m128i*) x);
27 x += 4;
28 const __m128i vy = _mm_loadu_si128((const __m128i*) y);
29 y += 4;
30 const __m128i vxy_lo = _mm_unpacklo_epi32(vx, vy);
31 const __m128i vxy_hi = _mm_unpackhi_epi32(vx, vy);
32 _mm_storeu_si128((__m128i*) o, vxy_lo);
33 _mm_storeu_si128((__m128i*) (o + 4), vxy_hi);
34 o += 8;
35 n -= 16;
36 }
37 if XNN_UNLIKELY(n != 0) {
38 if (n & 8) {
39 const __m128i vx = _mm_loadl_epi64((const __m128i*) x);
40 x += 2;
41 const __m128i vy = _mm_loadl_epi64((const __m128i*) y);
42 y += 2;
43 const __m128i vxy = _mm_unpacklo_epi32(vx, vy);
44 _mm_storeu_si128((__m128i*) o, vxy);
45 o += 4;
46 }
47 if (n & 4) {
48 const uint32_t vx = *x;
49 const uint32_t vy = *y;
50 o[0] = vx;
51 o[1] = vy;
52 }
53 }
54 }
55