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