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_multi_mov_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_multi_mov_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 input_offset = tile_height * input_stride;
35 const size_t output_reset = tile_width * output_stride - round_down_po2(block_height, 2) * sizeof(uint32_t) - tile_hbytes;
36
37 const uint32_t* i0 = input;
38 const uint32_t* i1 = (const uint32_t*) ((uintptr_t) i0 + input_stride);
39 const uint32_t* i2 = (const uint32_t*) ((uintptr_t) i1 + input_stride);
40 const uint32_t* i3 = (const uint32_t*) ((uintptr_t) i2 + input_stride);
41 uint32_t* o = (uint32_t*) ((uintptr_t) output - tile_hbytes);
42 const size_t minus_output_stride = -output_stride;
43
44 do {
45 const size_t rem = min(block_width - 1, 3);
46 const size_t oN_stride = rem * output_stride;
47 const size_t oN_offset = oN_stride + tile_hbytes;
48 size_t bh = block_height;
49 for (; bh >= 4; bh -= 4) {
50 const __m128i v2_0 = _mm_loadu_si128((const __m128i*) i0);
51 i0 = (uint32_t*) ((uintptr_t) i0 + input_offset);
52 const __m128i v2_1 = _mm_loadu_si128((const __m128i*) i1);
53 i1 = (uint32_t*) ((uintptr_t) i1 + input_offset);
54 const __m128i v2_2 = _mm_loadu_si128((const __m128i*) i2);
55 i2 = (uint32_t*) ((uintptr_t) i2 + input_offset);
56 const __m128i v2_3 = _mm_loadu_si128((const __m128i*) i3);
57 i3 = (uint32_t*) ((uintptr_t) i3 + input_offset);
58
59 const __m128i v1_0 = _mm_unpacklo_epi32(v2_0, v2_1);
60 const __m128i v1_1 = _mm_unpackhi_epi32(v2_0, v2_1);
61 const __m128i v1_2 = _mm_unpacklo_epi32(v2_2, v2_3);
62 const __m128i v1_3 = _mm_unpackhi_epi32(v2_2, v2_3);
63
64 const __m128i v0_0 = _mm_unpacklo_epi64(v1_0, v1_2);
65 const __m128i v0_1 = _mm_unpackhi_epi64(v1_0, v1_2);
66 const __m128i v0_2 = _mm_unpacklo_epi64(v1_1, v1_3);
67 const __m128i v0_3 = _mm_unpackhi_epi64(v1_1, v1_3);
68
69
70
71 o = (uint32_t*) ((uintptr_t) o + oN_offset);
72 _mm_storeu_si128((__m128i*) o, v0_3);
73 uint32_t *oN = (uint32_t*) ((uintptr_t) o + minus_output_stride);
74 if XNN_UNPREDICTABLE(block_width > 3) {
75 o = oN;
76 }
77 _mm_storeu_si128((__m128i*) o, v0_2);
78 oN = (uint32_t*) ((uintptr_t) o + minus_output_stride);
79 if XNN_UNPREDICTABLE(block_width >= 3) {
80 o = oN;
81 }
82 _mm_storeu_si128((__m128i*) o, v0_1);
83 oN = (uint32_t*) ((uintptr_t) o + minus_output_stride);
84 if XNN_UNPREDICTABLE(block_width > 1) {
85 o = oN;
86 }
87 _mm_storeu_si128((__m128i*) o, v0_0);
88 }
89 o = (uint32_t*) ((uintptr_t) o + tile_hbytes);
90 if (bh != 0) {
91 const __m128i v2_0 = _mm_loadu_si128((const __m128i*) i0);
92 if XNN_UNPREDICTABLE(bh < 2) {
93 i1 = i0;
94 }
95 const __m128i v2_1 = _mm_loadu_si128((const __m128i*) i1);
96 if XNN_UNPREDICTABLE(bh <= 2) {
97 i2 = i0;
98 }
99 const __m128i v2_2 = _mm_loadu_si128((const __m128i*) i2);
100 const __m128i v2_3 = _mm_undefined_si128();
101
102 const __m128i v1_0 = _mm_unpacklo_epi32(v2_0, v2_1);
103 const __m128i v1_1 = _mm_unpackhi_epi32(v2_0, v2_1);
104 const __m128i v1_2 = _mm_unpacklo_epi32(v2_2, v2_3);
105 const __m128i v1_3 = _mm_unpackhi_epi32(v2_2, v2_3);
106
107 __m128i v0_0 = _mm_unpacklo_epi64(v1_0, v1_2);
108 __m128i v0_1 = _mm_unpackhi_epi64(v1_0, v1_2);
109 __m128i v0_2 = _mm_unpacklo_epi64(v1_1, v1_3);
110 __m128i v0_3 = _mm_unpackhi_epi64(v1_1, v1_3);
111
112
113
114 if (bh & 2) {
115 o = (uint32_t*) ((uintptr_t) o + oN_stride);
116 _mm_storel_epi64((__m128i*) o, v0_3);
117 uint32_t *oN = (uint32_t*) ((uintptr_t) o + minus_output_stride);
118 if XNN_UNPREDICTABLE(block_width > 3) {
119 o = oN;
120 }
121 _mm_storel_epi64((__m128i*) o, v0_2);
122 oN = (uint32_t*) ((uintptr_t) o + minus_output_stride);
123 if XNN_UNPREDICTABLE(block_width >= 3) {
124 o = oN;
125 }
126 _mm_storel_epi64((__m128i*) o, v0_1);
127 oN = (uint32_t*) ((uintptr_t) o + minus_output_stride);
128 if XNN_UNPREDICTABLE(block_width > 1) {
129 o = oN;
130 }
131 _mm_storel_epi64((__m128i*) o, v0_0);
132 o += 2;
133 v0_0 = _mm_unpackhi_epi64(v0_0, v0_0);
134 v0_1 = _mm_unpackhi_epi64(v0_1, v0_1);
135 v0_2 = _mm_unpackhi_epi64(v0_2, v0_2);
136 v0_3 = _mm_unpackhi_epi64(v0_3, v0_3);
137 }
138
139 if (bh & 1) {
140 o = (uint32_t*) ((uintptr_t) o + oN_stride);
141 *((int*) o) = _mm_cvtsi128_si32(v0_3);
142 uint32_t *oN = (uint32_t*) ((uintptr_t) o + minus_output_stride);
143 if XNN_UNPREDICTABLE(block_width > 3) {
144 o = oN;
145 }
146 *((int*) o) = _mm_cvtsi128_si32(v0_2);
147 oN = (uint32_t*) ((uintptr_t) o + minus_output_stride);
148 if XNN_UNPREDICTABLE(block_width >= 3) {
149 o = oN;
150 }
151 *((int*) o) = _mm_cvtsi128_si32(v0_1);
152 oN = (uint32_t*) ((uintptr_t) o + minus_output_stride);
153 if XNN_UNPREDICTABLE(block_width > 1) {
154 o = oN;
155 }
156 *((int*) o) = _mm_cvtsi128_si32(v0_0);
157 }
158 }
159
160 i0 = (const uint32_t*) ((uintptr_t) i0 + input_reset);
161 i1 = (const uint32_t*) ((uintptr_t) i0 + input_stride);
162 i2 = (const uint32_t*) ((uintptr_t) i1 + input_stride);
163 i3 = (const uint32_t*) ((uintptr_t) i2 + input_stride);
164 o = (uint32_t*) ((uintptr_t) o + output_reset);
165 block_width = doz(block_width, tile_width);
166 } while (block_width != 0);
167 }
168