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_reuse_switch_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_reuse_switch_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 output_reset = tile_width * output_stride - round_down_po2(block_height, 2) * sizeof(uint64_t);
35
36 const uint64_t* i0 = input;
37 uint64_t* o = (uint64_t*) output;
38
39 do {
40 const size_t rem = min(block_width - 1, 1);
41 const size_t oN_stride = rem * output_stride;
42 size_t bh = block_height;
43 for (; bh >= 2; bh -= 2) {
44 const __m128i v1_0 = _mm_loadu_si128((const __m128i*) i0);
45 i0 = (uint64_t*) ((uintptr_t) i0 + input_stride);
46 const __m128i v1_1 = _mm_loadu_si128((const __m128i*) i0);
47 i0 = (uint64_t*) ((uintptr_t) i0 + input_stride);
48
49 const __m128i v0_0 = _mm_unpacklo_epi64(v1_0, v1_1);
50 const __m128i v0_1 = _mm_unpackhi_epi64(v1_0, v1_1);
51
52
53
54
55 uint64_t* oN = (uint64_t*) ((uintptr_t) o + oN_stride);
56 switch (rem) {
57 case 1:
58 _mm_storeu_si128((__m128i*) oN, v0_1);
59 case 0:
60 _mm_storeu_si128((__m128i*) o, v0_0);
61 o = (uint64_t*) ((uintptr_t) o + tile_hbytes);
62 break;
63 default:
64 XNN_UNREACHABLE;
65 }
66 }
67 if (bh != 0) {
68 const __m128i v1_0 = _mm_loadu_si128((const __m128i*) i0);
69 const __m128i v1_1 = _mm_undefined_si128();
70
71 __m128i v0_0 = _mm_unpacklo_epi64(v1_0, v1_1);
72 __m128i v0_1 = _mm_unpackhi_epi64(v1_0, v1_1);
73
74
75
76
77 if (bh & 1) {
78 uint64_t* oN = (uint64_t*) ((uintptr_t) o + oN_stride);
79 switch (rem) {
80 case 1:
81 _mm_storel_epi64((__m128i*) oN, v0_1);
82 case 0:
83 _mm_storel_epi64((__m128i*) o, v0_0);
84 break;
85 default:
86 XNN_UNREACHABLE;
87 }
88 }
89
90 }
91
92 i0 = (const uint64_t*) ((uintptr_t) i0 + input_reset);
93 o = (uint64_t*) ((uintptr_t) o + output_reset);
94 block_width = doz(block_width, tile_width);
95 } while (block_width != 0);
96 }
97