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_x32_transpose_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)18 void xnn_x32_transpose_ukernel__4x4_reuse_multi_sse2(
19 const uint32_t* input,
20 uint32_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(uint32_t));
27 assert(input_stride >= block_width * sizeof(uint32_t));
28
29 const size_t tile_height = 4;
30 const size_t tile_width = 4;
31 const size_t tile_hbytes = tile_height * sizeof(uint32_t);
32 const size_t tile_wbytes = tile_width * sizeof(uint32_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(uint32_t);
35
36 const uint32_t* i0 = input;
37 uint32_t* o0 = (uint32_t*) output;
38 uint32_t* o1 = (uint32_t*) ((uintptr_t) o0 + output_stride);
39 uint32_t* o2 = (uint32_t*) ((uintptr_t) o1 + output_stride);
40 uint32_t* o3 = (uint32_t*) ((uintptr_t) o2 + output_stride);
41
42 do {
43 if XNN_UNPREDICTABLE(block_width < 2) {
44 o1 = o0;
45 }
46 if XNN_UNPREDICTABLE(block_width <= 2) {
47 o2 = o0;
48 }
49 if XNN_UNPREDICTABLE(block_width < 4) {
50 o3 = o0;
51 }
52 size_t bh = block_height;
53 for (; bh >= 4; bh -= 4) {
54 const __m128i v2_0 = _mm_loadu_si128((const __m128i*) i0);
55 i0 = (uint32_t*) ((uintptr_t) i0 + input_stride);
56 const __m128i v2_1 = _mm_loadu_si128((const __m128i*) i0);
57 i0 = (uint32_t*) ((uintptr_t) i0 + input_stride);
58 const __m128i v2_2 = _mm_loadu_si128((const __m128i*) i0);
59 i0 = (uint32_t*) ((uintptr_t) i0 + input_stride);
60 const __m128i v2_3 = _mm_loadu_si128((const __m128i*) i0);
61 i0 = (uint32_t*) ((uintptr_t) i0 + input_stride);
62
63 const __m128i v1_0 = _mm_unpacklo_epi32(v2_0, v2_1);
64 const __m128i v1_1 = _mm_unpackhi_epi32(v2_0, v2_1);
65 const __m128i v1_2 = _mm_unpacklo_epi32(v2_2, v2_3);
66 const __m128i v1_3 = _mm_unpackhi_epi32(v2_2, v2_3);
67
68 const __m128i v0_0 = _mm_unpacklo_epi64(v1_0, v1_2);
69 const __m128i v0_1 = _mm_unpackhi_epi64(v1_0, v1_2);
70 const __m128i v0_2 = _mm_unpacklo_epi64(v1_1, v1_3);
71 const __m128i v0_3 = _mm_unpackhi_epi64(v1_1, v1_3);
72
73
74
75 _mm_storeu_si128((__m128i*) o3, v0_3);
76 o3 = (uint32_t*) ((uintptr_t) o3 + tile_hbytes);
77 _mm_storeu_si128((__m128i*) o2, v0_2);
78 o2 = (uint32_t*) ((uintptr_t) o2 + tile_hbytes);
79 _mm_storeu_si128((__m128i*) o1, v0_1);
80 o1 = (uint32_t*) ((uintptr_t) o1 + tile_hbytes);
81 _mm_storeu_si128((__m128i*) o0, v0_0);
82 o0 = (uint32_t*) ((uintptr_t) o0 + tile_hbytes);
83 }
84 if (bh != 0) {
85 const __m128i v2_0 = _mm_loadu_si128((const __m128i*) i0);
86 const uint32_t *i1 = (const uint32_t*) ((uintptr_t) i0 + input_stride);
87 if XNN_UNPREDICTABLE(bh < 2) {
88 i1 = i0;
89 }
90 const __m128i v2_1 = _mm_loadu_si128((const __m128i*) i1);
91 const uint32_t *i2 = (const uint32_t*) ((uintptr_t) i1 + input_stride);
92 if XNN_UNPREDICTABLE(bh <= 2) {
93 i2 = i1;
94 }
95 const __m128i v2_2 = _mm_loadu_si128((const __m128i*) i2);
96 const __m128i v2_3 = _mm_undefined_si128();
97
98 const __m128i v1_0 = _mm_unpacklo_epi32(v2_0, v2_1);
99 const __m128i v1_1 = _mm_unpackhi_epi32(v2_0, v2_1);
100 const __m128i v1_2 = _mm_unpacklo_epi32(v2_2, v2_3);
101 const __m128i v1_3 = _mm_unpackhi_epi32(v2_2, v2_3);
102
103 __m128i v0_0 = _mm_unpacklo_epi64(v1_0, v1_2);
104 __m128i v0_1 = _mm_unpackhi_epi64(v1_0, v1_2);
105 __m128i v0_2 = _mm_unpacklo_epi64(v1_1, v1_3);
106 __m128i v0_3 = _mm_unpackhi_epi64(v1_1, v1_3);
107
108
109
110 if (bh & 2) {
111 _mm_storel_epi64((__m128i*) o3, v0_3);
112 o3 += 2;
113 _mm_storel_epi64((__m128i*) o2, v0_2);
114 o2 += 2;
115 _mm_storel_epi64((__m128i*) o1, v0_1);
116 o1 += 2;
117 _mm_storel_epi64((__m128i*) o0, v0_0);
118 o0 += 2;
119 v0_0 = _mm_unpackhi_epi64(v0_0, v0_0);
120 v0_1 = _mm_unpackhi_epi64(v0_1, v0_1);
121 v0_2 = _mm_unpackhi_epi64(v0_2, v0_2);
122 v0_3 = _mm_unpackhi_epi64(v0_3, v0_3);
123 }
124
125 if (bh & 1) {
126 *((int*) o3) = _mm_cvtsi128_si32(v0_3);
127 *((int*) o2) = _mm_cvtsi128_si32(v0_2);
128 *((int*) o1) = _mm_cvtsi128_si32(v0_1);
129 *((int*) o0) = _mm_cvtsi128_si32(v0_0);
130 }
131 }
132
133 i0 = (const uint32_t*) ((uintptr_t) i0 + input_reset);
134 o0 = (uint32_t*) ((uintptr_t) o0 + output_reset);
135 o1 = (uint32_t*) ((uintptr_t) o1 + output_reset);
136 o2 = (uint32_t*) ((uintptr_t) o2 + output_reset);
137 o3 = (uint32_t*) ((uintptr_t) o3 + output_reset);
138 block_width = doz(block_width, tile_width);
139 } while (block_width != 0);
140 }
141