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_x32_transposec_ukernel__4x4_reuse_multi_sse2(const uint32_t * input,uint32_t * output,size_t input_stride,size_t output_stride,size_t block_width,size_t block_height)20 void xnn_x32_transposec_ukernel__4x4_reuse_multi_sse2(
21 const uint32_t* input,
22 uint32_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(uint32_t));
29 assert(input_stride >= block_width * sizeof(uint32_t));
30
31 const size_t tile_height = 4;
32 const size_t tile_width = 4;
33 const size_t tile_hbytes = tile_height * sizeof(uint32_t);
34 const size_t tile_wbytes = tile_width * sizeof(uint32_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(uint32_t);
37
38 const uint32_t* i0 = input;
39 uint32_t* o0 = (uint32_t*) output;
40 uint32_t* o1 = (uint32_t*) ((uintptr_t) o0 + output_stride);
41 uint32_t* o2 = (uint32_t*) ((uintptr_t) o1 + output_stride);
42 uint32_t* o3 = (uint32_t*) ((uintptr_t) o2 + output_stride);
43
44 do {
45 if XNN_UNPREDICTABLE(block_width < 2) {
46 o1 = o0;
47 }
48 if XNN_UNPREDICTABLE(block_width <= 2) {
49 o2 = o0;
50 }
51 if XNN_UNPREDICTABLE(block_width < 4) {
52 o3 = o0;
53 }
54 size_t bh = block_height;
55 for (; bh >= 4; bh -= 4) {
56 const __m128i v2_0 = _mm_loadu_si128((const __m128i*) i0);
57 i0 = (uint32_t*) ((uintptr_t) i0 + input_stride);
58 const __m128i v2_1 = _mm_loadu_si128((const __m128i*) i0);
59 i0 = (uint32_t*) ((uintptr_t) i0 + input_stride);
60 const __m128i v2_2 = _mm_loadu_si128((const __m128i*) i0);
61 i0 = (uint32_t*) ((uintptr_t) i0 + input_stride);
62 const __m128i v2_3 = _mm_loadu_si128((const __m128i*) i0);
63 i0 = (uint32_t*) ((uintptr_t) i0 + input_stride);
64
65 const __m128i v1_0 = _mm_unpacklo_epi32(v2_0, v2_1);
66 const __m128i v1_1 = _mm_unpackhi_epi32(v2_0, v2_1);
67 const __m128i v1_2 = _mm_unpacklo_epi32(v2_2, v2_3);
68 const __m128i v1_3 = _mm_unpackhi_epi32(v2_2, v2_3);
69
70 const __m128i v0_0 = _mm_unpacklo_epi64(v1_0, v1_2);
71 const __m128i v0_1 = _mm_unpackhi_epi64(v1_0, v1_2);
72 const __m128i v0_2 = _mm_unpacklo_epi64(v1_1, v1_3);
73 const __m128i v0_3 = _mm_unpackhi_epi64(v1_1, v1_3);
74
75
76
77 _mm_storeu_si128((__m128i*) o3, v0_3);
78 o3 = (uint32_t*) ((uintptr_t) o3 + tile_hbytes);
79 _mm_storeu_si128((__m128i*) o2, v0_2);
80 o2 = (uint32_t*) ((uintptr_t) o2 + tile_hbytes);
81 _mm_storeu_si128((__m128i*) o1, v0_1);
82 o1 = (uint32_t*) ((uintptr_t) o1 + tile_hbytes);
83 _mm_storeu_si128((__m128i*) o0, v0_0);
84 o0 = (uint32_t*) ((uintptr_t) o0 + tile_hbytes);
85 }
86 if (bh != 0) {
87 const __m128i v2_0 = _mm_loadu_si128((const __m128i*) i0);
88 const uint32_t *i1 = (const uint32_t*) ((uintptr_t) i0 + input_stride);
89 if XNN_UNPREDICTABLE(bh < 2) {
90 i1 = i0;
91 }
92 const __m128i v2_1 = _mm_loadu_si128((const __m128i*) i1);
93 const uint32_t *i2 = (const uint32_t*) ((uintptr_t) i1 + input_stride);
94 if XNN_UNPREDICTABLE(bh <= 2) {
95 i2 = i1;
96 }
97 const __m128i v2_2 = _mm_loadu_si128((const __m128i*) i2);
98 const __m128i v2_3 = _mm_undefined_si128();
99
100 const __m128i v1_0 = _mm_unpacklo_epi32(v2_0, v2_1);
101 const __m128i v1_1 = _mm_unpackhi_epi32(v2_0, v2_1);
102 const __m128i v1_2 = _mm_unpacklo_epi32(v2_2, v2_3);
103 const __m128i v1_3 = _mm_unpackhi_epi32(v2_2, v2_3);
104
105 __m128i v0_0 = _mm_unpacklo_epi64(v1_0, v1_2);
106 __m128i v0_1 = _mm_unpackhi_epi64(v1_0, v1_2);
107 __m128i v0_2 = _mm_unpacklo_epi64(v1_1, v1_3);
108 __m128i v0_3 = _mm_unpackhi_epi64(v1_1, v1_3);
109
110
111
112 if (bh & 2) {
113 _mm_storel_epi64((__m128i*) o3, v0_3);
114 o3 += 2;
115 _mm_storel_epi64((__m128i*) o2, v0_2);
116 o2 += 2;
117 _mm_storel_epi64((__m128i*) o1, v0_1);
118 o1 += 2;
119 _mm_storel_epi64((__m128i*) o0, v0_0);
120 o0 += 2;
121 v0_0 = _mm_unpackhi_epi64(v0_0, v0_0);
122 v0_1 = _mm_unpackhi_epi64(v0_1, v0_1);
123 v0_2 = _mm_unpackhi_epi64(v0_2, v0_2);
124 v0_3 = _mm_unpackhi_epi64(v0_3, v0_3);
125 }
126
127 if (bh & 1) {
128 unaligned_store_u32(o3, (uint32_t) _mm_cvtsi128_si32(v0_3));
129 unaligned_store_u32(o2, (uint32_t) _mm_cvtsi128_si32(v0_2));
130 unaligned_store_u32(o1, (uint32_t) _mm_cvtsi128_si32(v0_1));
131 unaligned_store_u32(o0, (uint32_t) _mm_cvtsi128_si32(v0_0));
132 }
133 }
134
135 i0 = (const uint32_t*) ((uintptr_t) i0 + input_reset);
136 o0 = (uint32_t*) ((uintptr_t) o0 + output_reset);
137 o1 = (uint32_t*) ((uintptr_t) o1 + output_reset);
138 o2 = (uint32_t*) ((uintptr_t) o2 + output_reset);
139 o3 = (uint32_t*) ((uintptr_t) o3 + output_reset);
140 block_width = doz(block_width, tile_width);
141 } while (block_width != 0);
142 }
143