1 // Auto-generated file. Do not edit!
2 // Template: src/x32-transposec/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 #include <xnnpack/unaligned.h>
18
19
xnn_x64_transposec_ukernel__2x2_multi_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)20 void xnn_x64_transposec_ukernel__2x2_multi_switch_sse2(
21 const uint64_t* input,
22 uint64_t* output,
23 size_t input_stride,
24 size_t output_stride,
25 size_t block_width,
26 size_t block_height) XNN_OOB_READS
27 {
28 assert(output_stride >= block_height * sizeof(uint64_t));
29 assert(input_stride >= block_width * sizeof(uint64_t));
30
31 const size_t tile_height = 2;
32 const size_t tile_width = 2;
33 const size_t tile_hbytes = tile_height * sizeof(uint64_t);
34 const size_t tile_wbytes = tile_width * sizeof(uint64_t);
35 const size_t input_reset = tile_wbytes - round_down_po2(block_height, tile_height) * input_stride;
36 const size_t input_offset = tile_height * input_stride;
37 const size_t output_reset = tile_width * output_stride - round_down_po2(block_height, 2) * sizeof(uint64_t);
38
39 const uint64_t* i0 = input;
40 const uint64_t* i1 = (const uint64_t*) ((uintptr_t) i0 + input_stride);
41 uint64_t* o = (uint64_t*) output;
42
43 do {
44 const size_t rem = min(block_width - 1, 1);
45 const size_t oN_stride = rem * output_stride;
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 uint64_t* oN = (uint64_t*) ((uintptr_t) o + oN_stride);
60 switch (rem) {
61 case 1:
62 _mm_storeu_si128((__m128i*) oN, v0_1);
63 case 0:
64 _mm_storeu_si128((__m128i*) o, v0_0);
65 o = (uint64_t*) ((uintptr_t) o + tile_hbytes);
66 break;
67 default:
68 XNN_UNREACHABLE;
69 }
70 }
71 if (bh != 0) {
72 const __m128i v1_0 = _mm_loadu_si128((const __m128i*) i0);
73 const __m128i v1_1 = _mm_undefined_si128();
74
75 __m128i v0_0 = _mm_unpacklo_epi64(v1_0, v1_1);
76 __m128i v0_1 = _mm_unpackhi_epi64(v1_0, v1_1);
77
78
79
80
81 if (bh & 1) {
82 uint64_t* oN = (uint64_t*) ((uintptr_t) o + oN_stride);
83 switch (rem) {
84 case 1:
85 _mm_storel_epi64((__m128i*) oN, v0_1);
86 case 0:
87 _mm_storel_epi64((__m128i*) o, v0_0);
88 break;
89 default:
90 XNN_UNREACHABLE;
91 }
92 }
93
94 }
95
96 i0 = (const uint64_t*) ((uintptr_t) i0 + input_reset);
97 i1 = (const uint64_t*) ((uintptr_t) i0 + input_stride);
98 o = (uint64_t*) ((uintptr_t) o + output_reset);
99 block_width = doz(block_width, tile_width);
100 } while (block_width != 0);
101 }
102