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_switch_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_switch_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* o = (uint32_t*) output;
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 size_t bh = block_height;
50 for (; bh >= 4; bh -= 4) {
51 const __m128i v2_0 = _mm_loadu_si128((const __m128i*) i0);
52 i0 = (uint32_t*) ((uintptr_t) i0 + input_offset);
53 const __m128i v2_1 = _mm_loadu_si128((const __m128i*) i1);
54 i1 = (uint32_t*) ((uintptr_t) i1 + input_offset);
55 const __m128i v2_2 = _mm_loadu_si128((const __m128i*) i2);
56 i2 = (uint32_t*) ((uintptr_t) i2 + input_offset);
57 const __m128i v2_3 = _mm_loadu_si128((const __m128i*) i3);
58 i3 = (uint32_t*) ((uintptr_t) i3 + input_offset);
59
60 const __m128i v1_0 = _mm_unpacklo_epi32(v2_0, v2_1);
61 const __m128i v1_1 = _mm_unpackhi_epi32(v2_0, v2_1);
62 const __m128i v1_2 = _mm_unpacklo_epi32(v2_2, v2_3);
63 const __m128i v1_3 = _mm_unpackhi_epi32(v2_2, v2_3);
64
65 const __m128i v0_0 = _mm_unpacklo_epi64(v1_0, v1_2);
66 const __m128i v0_1 = _mm_unpackhi_epi64(v1_0, v1_2);
67 const __m128i v0_2 = _mm_unpacklo_epi64(v1_1, v1_3);
68 const __m128i v0_3 = _mm_unpackhi_epi64(v1_1, v1_3);
69
70
71
72 uint32_t* oN = (uint32_t*) ((uintptr_t) o + oN_stride);
73 switch (rem) {
74 case 3:
75 _mm_storeu_si128((__m128i*) oN, v0_3);
76 oN = (uint32_t*) ((uintptr_t) oN + minus_output_stride);
77 case 2:
78 _mm_storeu_si128((__m128i*) oN, v0_2);
79 oN = (uint32_t*) ((uintptr_t) oN + minus_output_stride);
80 case 1:
81 _mm_storeu_si128((__m128i*) oN, v0_1);
82 case 0:
83 _mm_storeu_si128((__m128i*) o, v0_0);
84 o = (uint32_t*) ((uintptr_t) o + tile_hbytes);
85 break;
86 default:
87 XNN_UNREACHABLE;
88 }
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 uint32_t* oN = (uint32_t*) ((uintptr_t) o + oN_stride);
116 switch (rem) {
117 case 3:
118 _mm_storel_epi64((__m128i*) oN, v0_3);
119 oN = (uint32_t*) ((uintptr_t) oN + minus_output_stride);
120 case 2:
121 _mm_storel_epi64((__m128i*) oN, v0_2);
122 oN = (uint32_t*) ((uintptr_t) oN + minus_output_stride);
123 case 1:
124 _mm_storel_epi64((__m128i*) oN, v0_1);
125 case 0:
126 _mm_storel_epi64((__m128i*) o, v0_0);
127 break;
128 default:
129 XNN_UNREACHABLE;
130 }
131 o += 2;
132 v0_0 = _mm_unpackhi_epi64(v0_0, v0_0);
133 v0_1 = _mm_unpackhi_epi64(v0_1, v0_1);
134 v0_2 = _mm_unpackhi_epi64(v0_2, v0_2);
135 v0_3 = _mm_unpackhi_epi64(v0_3, v0_3);
136 }
137
138 if (bh & 1) {
139 uint32_t* oN = (uint32_t*) ((uintptr_t) o + oN_stride);
140 switch (rem) {
141 case 3:
142 unaligned_store_u32(oN, (uint32_t) _mm_cvtsi128_si32(v0_3));
143 oN = (uint32_t*) ((uintptr_t) oN + minus_output_stride);
144 case 2:
145 unaligned_store_u32(oN, (uint32_t) _mm_cvtsi128_si32(v0_2));
146 oN = (uint32_t*) ((uintptr_t) oN + minus_output_stride);
147 case 1:
148 unaligned_store_u32(oN, (uint32_t) _mm_cvtsi128_si32(v0_1));
149 case 0:
150 unaligned_store_u32(o, (uint32_t) _mm_cvtsi128_si32(v0_0));
151 break;
152 default:
153 XNN_UNREACHABLE;
154 }
155 }
156 }
157
158 i0 = (const uint32_t*) ((uintptr_t) i0 + input_reset);
159 i1 = (const uint32_t*) ((uintptr_t) i0 + input_stride);
160 i2 = (const uint32_t*) ((uintptr_t) i1 + input_stride);
161 i3 = (const uint32_t*) ((uintptr_t) i2 + input_stride);
162 o = (uint32_t*) ((uintptr_t) o + output_reset);
163 block_width = doz(block_width, tile_width);
164 } while (block_width != 0);
165 }
166