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_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_reuse_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 output_reset = tile_width * output_stride - round_down_po2(block_height, 2) * sizeof(uint32_t) - tile_hbytes;
37
38 const uint32_t* i0 = input;
39 uint32_t* o = (uint32_t*) ((uintptr_t) output - tile_hbytes);
40 const size_t minus_output_stride = -output_stride;
41
42 do {
43 const size_t rem = min(block_width - 1, 3);
44 const size_t oN_stride = rem * output_stride;
45 const size_t oN_offset = oN_stride + tile_hbytes;
46 size_t bh = block_height;
47 for (; bh >= 4; bh -= 4) {
48 const __m128i v2_0 = _mm_loadu_si128((const __m128i*) i0);
49 i0 = (uint32_t*) ((uintptr_t) i0 + input_stride);
50 const __m128i v2_1 = _mm_loadu_si128((const __m128i*) i0);
51 i0 = (uint32_t*) ((uintptr_t) i0 + input_stride);
52 const __m128i v2_2 = _mm_loadu_si128((const __m128i*) i0);
53 i0 = (uint32_t*) ((uintptr_t) i0 + input_stride);
54 const __m128i v2_3 = _mm_loadu_si128((const __m128i*) i0);
55 i0 = (uint32_t*) ((uintptr_t) i0 + input_stride);
56
57 const __m128i v1_0 = _mm_unpacklo_epi32(v2_0, v2_1);
58 const __m128i v1_1 = _mm_unpackhi_epi32(v2_0, v2_1);
59 const __m128i v1_2 = _mm_unpacklo_epi32(v2_2, v2_3);
60 const __m128i v1_3 = _mm_unpackhi_epi32(v2_2, v2_3);
61
62 const __m128i v0_0 = _mm_unpacklo_epi64(v1_0, v1_2);
63 const __m128i v0_1 = _mm_unpackhi_epi64(v1_0, v1_2);
64 const __m128i v0_2 = _mm_unpacklo_epi64(v1_1, v1_3);
65 const __m128i v0_3 = _mm_unpackhi_epi64(v1_1, v1_3);
66
67
68
69 o = (uint32_t*) ((uintptr_t) o + oN_offset);
70 _mm_storeu_si128((__m128i*) o, v0_3);
71 uint32_t *oN = (uint32_t*) ((uintptr_t) o + minus_output_stride);
72 if XNN_UNPREDICTABLE(block_width > 3) {
73 o = oN;
74 }
75 _mm_storeu_si128((__m128i*) o, v0_2);
76 oN = (uint32_t*) ((uintptr_t) o + minus_output_stride);
77 if XNN_UNPREDICTABLE(block_width >= 3) {
78 o = oN;
79 }
80 _mm_storeu_si128((__m128i*) o, v0_1);
81 oN = (uint32_t*) ((uintptr_t) o + minus_output_stride);
82 if XNN_UNPREDICTABLE(block_width > 1) {
83 o = oN;
84 }
85 _mm_storeu_si128((__m128i*) o, v0_0);
86 }
87 o = (uint32_t*) ((uintptr_t) o + tile_hbytes);
88 if (bh != 0) {
89 const __m128i v2_0 = _mm_loadu_si128((const __m128i*) i0);
90 const uint32_t *i1 = (const uint32_t*) ((uintptr_t) i0 + input_stride);
91 if XNN_UNPREDICTABLE(bh < 2) {
92 i1 = i0;
93 }
94 const __m128i v2_1 = _mm_loadu_si128((const __m128i*) i1);
95 const uint32_t *i2 = (const uint32_t*) ((uintptr_t) i1 + input_stride);
96 if XNN_UNPREDICTABLE(bh <= 2) {
97 i2 = i1;
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 unaligned_store_u32(o, (uint32_t) _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 unaligned_store_u32(o, (uint32_t) _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 unaligned_store_u32(o, (uint32_t) _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 unaligned_store_u32(o, (uint32_t) _mm_cvtsi128_si32(v0_0));
157 }
158 }
159
160 i0 = (const uint32_t*) ((uintptr_t) i0 + input_reset);
161 o = (uint32_t*) ((uintptr_t) o + output_reset);
162 block_width = doz(block_width, tile_width);
163 } while (block_width != 0);
164 }
165