1 // Copyright 2021 Google LLC
2 //
3 // This source code is licensed under the BSD-style license found in the
4 // LICENSE file in the root directory of this source tree.
5
6 #include <assert.h>
7
8 #include <emmintrin.h>
9
10 #include <xnnpack/common.h>
11 #include <xnnpack/math.h>
12 #include <xnnpack/transpose.h>
13 #include <xnnpack/unaligned.h>
14
15
xnn_x16_transposec_ukernel__4x8_sse2(const uint16_t * input,uint16_t * output,size_t input_stride,size_t output_stride,size_t block_width,size_t block_height)16 void xnn_x16_transposec_ukernel__4x8_sse2(
17 const uint16_t* input,
18 uint16_t* output,
19 size_t input_stride,
20 size_t output_stride,
21 size_t block_width,
22 size_t block_height) XNN_OOB_READS
23 {
24 assert(output_stride >= block_height * sizeof(uint16_t));
25 assert(input_stride >= block_width * sizeof(uint16_t));
26
27 const size_t tile_height = 4;
28 const size_t tile_width = 8;
29 const size_t tile_hbytes = tile_height * sizeof(uint16_t);
30 const size_t tile_wbytes = tile_width * sizeof(uint16_t);
31 const size_t input_reset = tile_wbytes - round_down_po2(block_height, tile_height) * input_stride;
32 const size_t output_reset = tile_width * output_stride - round_down_po2(block_height, 2) * sizeof(uint16_t);
33 const size_t input_offset = tile_height * input_stride;
34
35 const uint16_t* i0 = (const uint16_t*) input;
36 const uint16_t* i1 = (const uint16_t*) ((uintptr_t) i0 + input_stride);
37 const uint16_t* i2 = (const uint16_t*) ((uintptr_t) i1 + input_stride);
38 const uint16_t* i3 = (const uint16_t*) ((uintptr_t) i2 + input_stride);
39
40 uint16_t* o0 = (uint16_t*) output;
41 uint16_t* o1 = (uint16_t*) ((uintptr_t) o0 + output_stride);
42 uint16_t* o2 = (uint16_t*) ((uintptr_t) o1 + output_stride);
43 uint16_t* o3 = (uint16_t*) ((uintptr_t) o2 + output_stride);
44 uint16_t* o4 = (uint16_t*) ((uintptr_t) o3 + output_stride);
45 uint16_t* o5 = (uint16_t*) ((uintptr_t) o4 + output_stride);
46 uint16_t* o6 = (uint16_t*) ((uintptr_t) o5 + output_stride);
47 uint16_t* o7 = (uint16_t*) ((uintptr_t) o6 + output_stride);
48
49 do {
50 if XNN_UNPREDICTABLE(block_width < 2) {
51 o1 = o0;
52 }
53 if XNN_UNPREDICTABLE(block_width <= 2) {
54 o2 = o0;
55 }
56 if XNN_UNPREDICTABLE(block_width < 4) {
57 o3 = o0;
58 }
59 if XNN_UNPREDICTABLE(block_width <= 4) {
60 o4 = o0;
61 }
62 if XNN_UNPREDICTABLE(block_width < 6) {
63 o5 = o0;
64 }
65 if XNN_UNPREDICTABLE(block_width <= 6) {
66 o6 = o0;
67 }
68 if XNN_UNPREDICTABLE(block_width < 8) {
69 o7 = o0;
70 }
71 size_t bh = block_height;
72 for (; bh >= 4; bh -= 4) {
73 __m128i v0 = _mm_loadu_si128((const __m128i*) i0);
74 i0 = (const uint16_t*) ((uintptr_t) i0 + input_offset);
75 __m128i v1 = _mm_loadu_si128((const __m128i*) i1);
76 i1 = (const uint16_t*) ((uintptr_t) i1 + input_offset);
77 __m128i v2 = _mm_loadu_si128((const __m128i*) i2);
78 i2 = (const uint16_t*) ((uintptr_t) i2 + input_offset);
79 __m128i v3 = _mm_loadu_si128((const __m128i*) i3);
80 i3 = (const uint16_t*) ((uintptr_t) i3 + input_offset);
81
82 __m128i vtmp0 = _mm_unpacklo_epi16(v0, v2);
83 __m128i vtmp1 = _mm_unpacklo_epi16(v1, v3);
84 __m128i vtmp2 = _mm_unpackhi_epi16(v0, v2);
85 __m128i vtmp3 = _mm_unpackhi_epi16(v1, v3);
86
87 v0 = _mm_unpacklo_epi16(vtmp0, vtmp1);
88 v1 = _mm_unpacklo_epi16(vtmp2, vtmp3);
89 v2 = _mm_unpackhi_epi16(vtmp0, vtmp1);
90 v3 = _mm_unpackhi_epi16(vtmp2, vtmp3);
91
92 _mm_storeh_pi((__m64*) o7, _mm_castsi128_ps(v3));
93 o7 = (uint16_t*) ((uintptr_t) o7 + tile_hbytes);
94 _mm_storel_epi64((__m128i*) o6, v3);
95 o6 = (uint16_t*) ((uintptr_t) o6 + tile_hbytes);
96 _mm_storeh_pi((__m64*) o5, _mm_castsi128_ps(v1));
97 o5 = (uint16_t*) ((uintptr_t) o5 + tile_hbytes);
98 _mm_storel_epi64((__m128i*) o4, v1);
99 o4 = (uint16_t*) ((uintptr_t) o4 + tile_hbytes);
100 _mm_storeh_pi((__m64*) o3, _mm_castsi128_ps(v2));
101 o3 = (uint16_t*) ((uintptr_t) o3 + tile_hbytes);
102 _mm_storel_epi64((__m128i*) o2, v2);
103 o2 = (uint16_t*) ((uintptr_t) o2 + tile_hbytes);
104 _mm_storeh_pi((__m64*) o1, _mm_castsi128_ps(v0));
105 o1 = (uint16_t*) ((uintptr_t) o1 + tile_hbytes);
106 _mm_storel_epi64((__m128i*) o0, v0);
107 o0 = (uint16_t*) ((uintptr_t) o0 + tile_hbytes);
108 }
109
110 if (bh != 0) {
111 if XNN_UNPREDICTABLE(bh <= 2) {
112 i2 = i0;
113 }
114 if XNN_UNPREDICTABLE(bh < 2) {
115 i1 = i0;
116 }
117 __m128i v0 = _mm_loadu_si128((const __m128i*) i0);
118 __m128i v1 = _mm_loadu_si128((const __m128i*) i1);
119 __m128i v2 = _mm_loadu_si128((const __m128i*) i2);
120
121 __m128i vtmp0 = _mm_unpacklo_epi16(v0, v2);
122 __m128i vtmp1 = _mm_unpacklo_epi16(v1, v1);
123 __m128i vtmp2 = _mm_unpackhi_epi16(v0, v2);
124 __m128i vtmp3 = _mm_unpackhi_epi16(v1, v1);
125
126 v0 = _mm_unpacklo_epi16(vtmp0, vtmp1);
127 v1 = _mm_unpacklo_epi16(vtmp2, vtmp3);
128 v2 = _mm_unpackhi_epi16(vtmp0, vtmp1);
129 __m128i v3 = _mm_unpackhi_epi16(vtmp2, vtmp3);
130
131 if (bh & 2) {
132 unaligned_store_u32(o7, (uint32_t) _mm_cvtsi128_si32(_mm_shuffle_epi32(v3, 0xE)));
133 o7 += 2;
134 unaligned_store_u32(o6, (uint32_t) _mm_cvtsi128_si32(v3));
135 o6 += 2;
136 unaligned_store_u32(o5, (uint32_t) _mm_cvtsi128_si32(_mm_shuffle_epi32(v1, 0xE)));
137 o5 += 2;
138 unaligned_store_u32(o4, (uint32_t) _mm_cvtsi128_si32(v1));
139 o4 += 2;
140 unaligned_store_u32(o3, (uint32_t) _mm_cvtsi128_si32(_mm_shuffle_epi32(v2, 0xE)));
141 o3 += 2;
142 unaligned_store_u32(o2, (uint32_t) _mm_cvtsi128_si32(v2));
143 o2 += 2;
144 unaligned_store_u32(o1, (uint32_t) _mm_cvtsi128_si32(_mm_shuffle_epi32(v0, 0xE)));
145 o1 += 2;
146 unaligned_store_u32(o0, (uint32_t) _mm_cvtsi128_si32(v0));
147 o0 += 2;
148 v0 = _mm_srli_epi64(v0, 32);
149 v1 = _mm_srli_epi64(v1, 32);
150 v2 = _mm_srli_epi64(v2, 32);
151 v3 = _mm_srli_epi64(v3, 32);
152 }
153 if (bh & 1) {
154 *o7 = (uint16_t) _mm_extract_epi16(v3, 4);
155 *o6 = (uint16_t) _mm_cvtsi128_si32(v3);
156 *o5 = (uint16_t) _mm_extract_epi16(v1, 4);
157 *o4 = (uint16_t) _mm_cvtsi128_si32(v1);
158 *o3 = (uint16_t) _mm_extract_epi16(v2, 4);
159 *o2 = (uint16_t) _mm_cvtsi128_si32(v2);
160 *o1 = (uint16_t) _mm_extract_epi16(v0, 4);
161 *o0 = (uint16_t) _mm_cvtsi128_si32(v0);
162 }
163 }
164 i0 = (const uint16_t*) ((uintptr_t) i0 + input_reset);
165 i1 = (const uint16_t*) ((uintptr_t) i0 + input_stride);
166 i2 = (const uint16_t*) ((uintptr_t) i1 + input_stride);
167 i3 = (const uint16_t*) ((uintptr_t) i2 + input_stride);
168 o0 = (uint16_t*) ((uintptr_t) o0 + output_reset);
169 o1 = (uint16_t*) ((uintptr_t) o1 + output_reset);
170 o2 = (uint16_t*) ((uintptr_t) o2 + output_reset);
171 o3 = (uint16_t*) ((uintptr_t) o3 + output_reset);
172 o4 = (uint16_t*) ((uintptr_t) o4 + output_reset);
173 o5 = (uint16_t*) ((uintptr_t) o5 + output_reset);
174 o6 = (uint16_t*) ((uintptr_t) o6 + output_reset);
175 o7 = (uint16_t*) ((uintptr_t) o7 + output_reset);
176 block_width = doz(block_width, tile_width);
177 } while (block_width != 0);
178 }
179