1 // Auto-generated file. Do not edit!
2 // Template: src/x32-transpose/sse2.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2021 Google LLC
6 //
7 // This source code is licensed under the BSD-style license found in the
8 // LICENSE file in the root directory of this source tree.
9
10 #include <immintrin.h>
11
12 #include <assert.h>
13
14 #include <xnnpack/common.h>
15 #include <xnnpack/math.h>
16 #include <xnnpack/transpose.h>
17
xnn_x64_transpose_ukernel__2x2_multi_multi_sse2(const uint64_t * input,uint64_t * output,size_t input_stride,size_t output_stride,size_t block_width,size_t block_height)18 void xnn_x64_transpose_ukernel__2x2_multi_multi_sse2(
19 const uint64_t* input,
20 uint64_t* output,
21 size_t input_stride,
22 size_t output_stride,
23 size_t block_width,
24 size_t block_height)
25 {
26 assert(output_stride >= block_height * sizeof(uint64_t));
27 assert(input_stride >= block_width * sizeof(uint64_t));
28
29 const size_t tile_height = 2;
30 const size_t tile_width = 2;
31 const size_t tile_hbytes = tile_height * sizeof(uint64_t);
32 const size_t tile_wbytes = tile_width * sizeof(uint64_t);
33 const size_t input_reset = tile_wbytes - round_down_po2(block_height, tile_height) * input_stride;
34 const size_t input_offset = tile_height * input_stride;
35 const size_t output_reset = tile_width * output_stride - round_down_po2(block_height, 2) * sizeof(uint64_t);
36
37 const uint64_t* i0 = input;
38 const uint64_t* i1 = (const uint64_t*) ((uintptr_t) i0 + input_stride);
39 uint64_t* o0 = (uint64_t*) output;
40 uint64_t* o1 = (uint64_t*) ((uintptr_t) o0 + output_stride);
41
42 do {
43 if XNN_UNPREDICTABLE(block_width < 2) {
44 o1 = o0;
45 }
46 size_t bh = block_height;
47 for (; bh >= 2; bh -= 2) {
48 const __m128i v1_0 = _mm_loadu_si128((const __m128i*) i0);
49 i0 = (uint64_t*) ((uintptr_t) i0 + input_offset);
50 const __m128i v1_1 = _mm_loadu_si128((const __m128i*) i1);
51 i1 = (uint64_t*) ((uintptr_t) i1 + input_offset);
52
53 const __m128i v0_0 = _mm_unpacklo_epi64(v1_0, v1_1);
54 const __m128i v0_1 = _mm_unpackhi_epi64(v1_0, v1_1);
55
56
57
58
59 _mm_storeu_si128((__m128i*) o1, v0_1);
60 o1 = (uint64_t*) ((uintptr_t) o1 + tile_hbytes);
61 _mm_storeu_si128((__m128i*) o0, v0_0);
62 o0 = (uint64_t*) ((uintptr_t) o0 + tile_hbytes);
63 }
64 if (bh != 0) {
65 const __m128i v1_0 = _mm_loadu_si128((const __m128i*) i0);
66 const __m128i v1_1 = _mm_undefined_si128();
67
68 __m128i v0_0 = _mm_unpacklo_epi64(v1_0, v1_1);
69 __m128i v0_1 = _mm_unpackhi_epi64(v1_0, v1_1);
70
71
72
73
74 if (bh & 1) {
75 _mm_storel_epi64((__m128i*) o1, v0_1);
76 _mm_storel_epi64((__m128i*) o0, v0_0);
77 }
78
79 }
80
81 i0 = (const uint64_t*) ((uintptr_t) i0 + input_reset);
82 i1 = (const uint64_t*) ((uintptr_t) i0 + input_stride);
83 o0 = (uint64_t*) ((uintptr_t) o0 + output_reset);
84 o1 = (uint64_t*) ((uintptr_t) o1 + output_reset);
85 block_width = doz(block_width, tile_width);
86 } while (block_width != 0);
87 }
88