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_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)20 void xnn_x32_transposec_ukernel__4x4_multi_mov_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 input_offset = tile_height * input_stride;
37 const size_t output_reset = tile_width * output_stride - round_down_po2(block_height, 2) * sizeof(uint32_t) - tile_hbytes;
38
39 const uint32_t* i0 = input;
40 const uint32_t* i1 = (const uint32_t*) ((uintptr_t) i0 + input_stride);
41 const uint32_t* i2 = (const uint32_t*) ((uintptr_t) i1 + input_stride);
42 const uint32_t* i3 = (const uint32_t*) ((uintptr_t) i2 + input_stride);
43 uint32_t* o = (uint32_t*) ((uintptr_t) output - tile_hbytes);
44 const size_t minus_output_stride = -output_stride;
45
46 do {
47 const size_t rem = min(block_width - 1, 3);
48 const size_t oN_stride = rem * output_stride;
49 const size_t oN_offset = oN_stride + tile_hbytes;
50 size_t bh = block_height;
51 for (; bh >= 4; bh -= 4) {
52 const __m128i v2_0 = _mm_loadu_si128((const __m128i*) i0);
53 i0 = (uint32_t*) ((uintptr_t) i0 + input_offset);
54 const __m128i v2_1 = _mm_loadu_si128((const __m128i*) i1);
55 i1 = (uint32_t*) ((uintptr_t) i1 + input_offset);
56 const __m128i v2_2 = _mm_loadu_si128((const __m128i*) i2);
57 i2 = (uint32_t*) ((uintptr_t) i2 + input_offset);
58 const __m128i v2_3 = _mm_loadu_si128((const __m128i*) i3);
59 i3 = (uint32_t*) ((uintptr_t) i3 + input_offset);
60
61 const __m128i v1_0 = _mm_unpacklo_epi32(v2_0, v2_1);
62 const __m128i v1_1 = _mm_unpackhi_epi32(v2_0, v2_1);
63 const __m128i v1_2 = _mm_unpacklo_epi32(v2_2, v2_3);
64 const __m128i v1_3 = _mm_unpackhi_epi32(v2_2, v2_3);
65
66 const __m128i v0_0 = _mm_unpacklo_epi64(v1_0, v1_2);
67 const __m128i v0_1 = _mm_unpackhi_epi64(v1_0, v1_2);
68 const __m128i v0_2 = _mm_unpacklo_epi64(v1_1, v1_3);
69 const __m128i v0_3 = _mm_unpackhi_epi64(v1_1, v1_3);
70
71
72
73 o = (uint32_t*) ((uintptr_t) o + oN_offset);
74 _mm_storeu_si128((__m128i*) o, v0_3);
75 uint32_t *oN = (uint32_t*) ((uintptr_t) o + minus_output_stride);
76 if XNN_UNPREDICTABLE(block_width > 3) {
77 o = oN;
78 }
79 _mm_storeu_si128((__m128i*) o, v0_2);
80 oN = (uint32_t*) ((uintptr_t) o + minus_output_stride);
81 if XNN_UNPREDICTABLE(block_width >= 3) {
82 o = oN;
83 }
84 _mm_storeu_si128((__m128i*) o, v0_1);
85 oN = (uint32_t*) ((uintptr_t) o + minus_output_stride);
86 if XNN_UNPREDICTABLE(block_width > 1) {
87 o = oN;
88 }
89 _mm_storeu_si128((__m128i*) o, v0_0);
90 }
91 o = (uint32_t*) ((uintptr_t) o + tile_hbytes);
92 if (bh != 0) {
93 const __m128i v2_0 = _mm_loadu_si128((const __m128i*) i0);
94 if XNN_UNPREDICTABLE(bh < 2) {
95 i1 = i0;
96 }
97 const __m128i v2_1 = _mm_loadu_si128((const __m128i*) i1);
98 if XNN_UNPREDICTABLE(bh <= 2) {
99 i2 = i0;
100 }
101 const __m128i v2_2 = _mm_loadu_si128((const __m128i*) i2);
102 const __m128i v2_3 = _mm_undefined_si128();
103
104 const __m128i v1_0 = _mm_unpacklo_epi32(v2_0, v2_1);
105 const __m128i v1_1 = _mm_unpackhi_epi32(v2_0, v2_1);
106 const __m128i v1_2 = _mm_unpacklo_epi32(v2_2, v2_3);
107 const __m128i v1_3 = _mm_unpackhi_epi32(v2_2, v2_3);
108
109 __m128i v0_0 = _mm_unpacklo_epi64(v1_0, v1_2);
110 __m128i v0_1 = _mm_unpackhi_epi64(v1_0, v1_2);
111 __m128i v0_2 = _mm_unpacklo_epi64(v1_1, v1_3);
112 __m128i v0_3 = _mm_unpackhi_epi64(v1_1, v1_3);
113
114
115
116 if (bh & 2) {
117 o = (uint32_t*) ((uintptr_t) o + oN_stride);
118 _mm_storel_epi64((__m128i*) o, v0_3);
119 uint32_t *oN = (uint32_t*) ((uintptr_t) o + minus_output_stride);
120 if XNN_UNPREDICTABLE(block_width > 3) {
121 o = oN;
122 }
123 _mm_storel_epi64((__m128i*) o, v0_2);
124 oN = (uint32_t*) ((uintptr_t) o + minus_output_stride);
125 if XNN_UNPREDICTABLE(block_width >= 3) {
126 o = oN;
127 }
128 _mm_storel_epi64((__m128i*) o, v0_1);
129 oN = (uint32_t*) ((uintptr_t) o + minus_output_stride);
130 if XNN_UNPREDICTABLE(block_width > 1) {
131 o = oN;
132 }
133 _mm_storel_epi64((__m128i*) o, v0_0);
134 o += 2;
135 v0_0 = _mm_unpackhi_epi64(v0_0, v0_0);
136 v0_1 = _mm_unpackhi_epi64(v0_1, v0_1);
137 v0_2 = _mm_unpackhi_epi64(v0_2, v0_2);
138 v0_3 = _mm_unpackhi_epi64(v0_3, v0_3);
139 }
140
141 if (bh & 1) {
142 o = (uint32_t*) ((uintptr_t) o + oN_stride);
143 unaligned_store_u32(o, (uint32_t) _mm_cvtsi128_si32(v0_3));
144 uint32_t *oN = (uint32_t*) ((uintptr_t) o + minus_output_stride);
145 if XNN_UNPREDICTABLE(block_width > 3) {
146 o = oN;
147 }
148 unaligned_store_u32(o, (uint32_t) _mm_cvtsi128_si32(v0_2));
149 oN = (uint32_t*) ((uintptr_t) o + minus_output_stride);
150 if XNN_UNPREDICTABLE(block_width >= 3) {
151 o = oN;
152 }
153 unaligned_store_u32(o, (uint32_t) _mm_cvtsi128_si32(v0_1));
154 oN = (uint32_t*) ((uintptr_t) o + minus_output_stride);
155 if XNN_UNPREDICTABLE(block_width > 1) {
156 o = oN;
157 }
158 unaligned_store_u32(o, (uint32_t) _mm_cvtsi128_si32(v0_0));
159 }
160 }
161
162 i0 = (const uint32_t*) ((uintptr_t) i0 + input_reset);
163 i1 = (const uint32_t*) ((uintptr_t) i0 + input_stride);
164 i2 = (const uint32_t*) ((uintptr_t) i1 + input_stride);
165 i3 = (const uint32_t*) ((uintptr_t) i2 + input_stride);
166 o = (uint32_t*) ((uintptr_t) o + output_reset);
167 block_width = doz(block_width, tile_width);
168 } while (block_width != 0);
169 }
170