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_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)20 void xnn_x64_transposec_ukernel__2x2_reuse_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 output_reset = tile_width * output_stride - round_down_po2(block_height, 2) * sizeof(uint64_t);
37
38 const uint64_t* i0 = input;
39 uint64_t* o = (uint64_t*) output;
40
41 do {
42 const size_t rem = min(block_width - 1, 1);
43 const size_t oN_stride = rem * output_stride;
44 size_t bh = block_height;
45 for (; bh >= 2; bh -= 2) {
46 const __m128i v1_0 = _mm_loadu_si128((const __m128i*) i0);
47 i0 = (uint64_t*) ((uintptr_t) i0 + input_stride);
48 const __m128i v1_1 = _mm_loadu_si128((const __m128i*) i0);
49 i0 = (uint64_t*) ((uintptr_t) i0 + input_stride);
50
51 const __m128i v0_0 = _mm_unpacklo_epi64(v1_0, v1_1);
52 const __m128i v0_1 = _mm_unpackhi_epi64(v1_0, v1_1);
53
54
55
56
57 uint64_t* oN = (uint64_t*) ((uintptr_t) o + oN_stride);
58 switch (rem) {
59 case 1:
60 _mm_storeu_si128((__m128i*) oN, v0_1);
61 case 0:
62 _mm_storeu_si128((__m128i*) o, v0_0);
63 o = (uint64_t*) ((uintptr_t) o + tile_hbytes);
64 break;
65 default:
66 XNN_UNREACHABLE;
67 }
68 }
69 if (bh != 0) {
70 const __m128i v1_0 = _mm_loadu_si128((const __m128i*) i0);
71 const __m128i v1_1 = _mm_undefined_si128();
72
73 __m128i v0_0 = _mm_unpacklo_epi64(v1_0, v1_1);
74 __m128i v0_1 = _mm_unpackhi_epi64(v1_0, v1_1);
75
76
77
78
79 if (bh & 1) {
80 uint64_t* oN = (uint64_t*) ((uintptr_t) o + oN_stride);
81 switch (rem) {
82 case 1:
83 _mm_storel_epi64((__m128i*) oN, v0_1);
84 case 0:
85 _mm_storel_epi64((__m128i*) o, v0_0);
86 break;
87 default:
88 XNN_UNREACHABLE;
89 }
90 }
91
92 }
93
94 i0 = (const uint64_t*) ((uintptr_t) i0 + input_reset);
95 o = (uint64_t*) ((uintptr_t) o + output_reset);
96 block_width = doz(block_width, tile_width);
97 } while (block_width != 0);
98 }
99