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_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_multi_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 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);
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* o0 = (uint32_t*) output;
44 uint32_t* o1 = (uint32_t*) ((uintptr_t) o0 + output_stride);
45 uint32_t* o2 = (uint32_t*) ((uintptr_t) o1 + output_stride);
46 uint32_t* o3 = (uint32_t*) ((uintptr_t) o2 + output_stride);
47
48 do {
49 if XNN_UNPREDICTABLE(block_width < 2) {
50 o1 = o0;
51 }
52 if XNN_UNPREDICTABLE(block_width <= 2) {
53 o2 = o0;
54 }
55 if XNN_UNPREDICTABLE(block_width < 4) {
56 o3 = o0;
57 }
58 size_t bh = block_height;
59 for (; bh >= 4; bh -= 4) {
60 const __m128i v2_0 = _mm_loadu_si128((const __m128i*) i0);
61 i0 = (uint32_t*) ((uintptr_t) i0 + input_offset);
62 const __m128i v2_1 = _mm_loadu_si128((const __m128i*) i1);
63 i1 = (uint32_t*) ((uintptr_t) i1 + input_offset);
64 const __m128i v2_2 = _mm_loadu_si128((const __m128i*) i2);
65 i2 = (uint32_t*) ((uintptr_t) i2 + input_offset);
66 const __m128i v2_3 = _mm_loadu_si128((const __m128i*) i3);
67 i3 = (uint32_t*) ((uintptr_t) i3 + input_offset);
68
69 const __m128i v1_0 = _mm_unpacklo_epi32(v2_0, v2_1);
70 const __m128i v1_1 = _mm_unpackhi_epi32(v2_0, v2_1);
71 const __m128i v1_2 = _mm_unpacklo_epi32(v2_2, v2_3);
72 const __m128i v1_3 = _mm_unpackhi_epi32(v2_2, v2_3);
73
74 const __m128i v0_0 = _mm_unpacklo_epi64(v1_0, v1_2);
75 const __m128i v0_1 = _mm_unpackhi_epi64(v1_0, v1_2);
76 const __m128i v0_2 = _mm_unpacklo_epi64(v1_1, v1_3);
77 const __m128i v0_3 = _mm_unpackhi_epi64(v1_1, v1_3);
78
79
80
81 _mm_storeu_si128((__m128i*) o3, v0_3);
82 o3 = (uint32_t*) ((uintptr_t) o3 + tile_hbytes);
83 _mm_storeu_si128((__m128i*) o2, v0_2);
84 o2 = (uint32_t*) ((uintptr_t) o2 + tile_hbytes);
85 _mm_storeu_si128((__m128i*) o1, v0_1);
86 o1 = (uint32_t*) ((uintptr_t) o1 + tile_hbytes);
87 _mm_storeu_si128((__m128i*) o0, v0_0);
88 o0 = (uint32_t*) ((uintptr_t) o0 + tile_hbytes);
89 }
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 _mm_storel_epi64((__m128i*) o3, v0_3);
116 o3 += 2;
117 _mm_storel_epi64((__m128i*) o2, v0_2);
118 o2 += 2;
119 _mm_storel_epi64((__m128i*) o1, v0_1);
120 o1 += 2;
121 _mm_storel_epi64((__m128i*) o0, v0_0);
122 o0 += 2;
123 v0_0 = _mm_unpackhi_epi64(v0_0, v0_0);
124 v0_1 = _mm_unpackhi_epi64(v0_1, v0_1);
125 v0_2 = _mm_unpackhi_epi64(v0_2, v0_2);
126 v0_3 = _mm_unpackhi_epi64(v0_3, v0_3);
127 }
128
129 if (bh & 1) {
130 unaligned_store_u32(o3, (uint32_t) _mm_cvtsi128_si32(v0_3));
131 unaligned_store_u32(o2, (uint32_t) _mm_cvtsi128_si32(v0_2));
132 unaligned_store_u32(o1, (uint32_t) _mm_cvtsi128_si32(v0_1));
133 unaligned_store_u32(o0, (uint32_t) _mm_cvtsi128_si32(v0_0));
134 }
135 }
136
137 i0 = (const uint32_t*) ((uintptr_t) i0 + input_reset);
138 i1 = (const uint32_t*) ((uintptr_t) i0 + input_stride);
139 i2 = (const uint32_t*) ((uintptr_t) i1 + input_stride);
140 i3 = (const uint32_t*) ((uintptr_t) i2 + input_stride);
141 o0 = (uint32_t*) ((uintptr_t) o0 + output_reset);
142 o1 = (uint32_t*) ((uintptr_t) o1 + output_reset);
143 o2 = (uint32_t*) ((uintptr_t) o2 + output_reset);
144 o3 = (uint32_t*) ((uintptr_t) o3 + output_reset);
145 block_width = doz(block_width, tile_width);
146 } while (block_width != 0);
147 }
148