1 // Copyright 2022 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 <tmmintrin.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_ALIGN(16) static const uint8_t pos0[16] = {0, 4, 8, 2, 6, 10, 1, 5, 9, 3, 7, 11, -1, -1, -1, -1};
16 XNN_ALIGN(16) static const uint8_t pos1[16] = {4, 8, 12, 6, 10, 14, 5, 9, 13, 7, 11, 15, -1, -1, -1, -1};
17 XNN_ALIGN(16) static const uint8_t pos2[16] = {12, -1, -1, 14, -1, -1, 13, -1, -1, 15, -1, -1, -1, -1, -1, -1};
18 XNN_ALIGN(16) static const uint8_t pos3[16] = {-1, 0, 4, -1, 2, 6, -1, 1, 5, -1, 3, 7, -1, -1, -1, -1};
19 XNN_ALIGN(16) static const uint8_t pos4[16] = {8, 12, -1, 10, 14, -1, 9, 13, -1, 11, 15, -1, -1, -1, -1, -1};
20 XNN_ALIGN(16) static const uint8_t pos5[16] = {-1, -1, 0, -1, -1, 2, -1, -1, 1, -1, -1, 3, -1, -1, -1, -1};
21
xnn_x24_transposec_ukernel__4x4_ssse3(const void * input,void * output,size_t input_stride,size_t output_stride,size_t block_width,size_t block_height)22 void xnn_x24_transposec_ukernel__4x4_ssse3(
23 const void *input,
24 void * output,
25 size_t input_stride,
26 size_t output_stride,
27 size_t block_width,
28 size_t block_height)
29 {
30 assert(output_stride >= block_height * 3);
31 assert(input_stride >= block_width * 3);
32
33 assert(output_stride >= block_height * 3);
34 assert(input_stride >= block_width * 3);
35
36 const size_t tile_height = 4;
37 const size_t tile_width = 4;
38 const size_t tile_wbytes = tile_width * 3;
39 const size_t input_reset = tile_wbytes - round_down_po2(block_height, tile_height) * input_stride;
40 const size_t output_reset = tile_width * output_stride - block_height * 3;
41 const size_t tile_stride = tile_height * input_stride;
42
43 const uint8_t* i0 = (const uint8_t*) input;
44 const uint8_t* i1 = (const uint8_t*) ((uintptr_t) i0 + input_stride);
45 const uint8_t* i2 = (const uint8_t*) ((uintptr_t) i1 + input_stride);
46 const uint8_t* i3 = (const uint8_t*) ((uintptr_t) i2 + input_stride);
47
48 uint8_t* o0 = (uint8_t*) output;
49 uint8_t* o1 = (uint8_t*) ((uintptr_t) o0 + output_stride);
50 uint8_t* o2 = (uint8_t*) ((uintptr_t) o1 + output_stride);
51 uint8_t* o3 = (uint8_t*) ((uintptr_t) o2 + output_stride);
52
53 const __m128i vperm0 = _mm_load_si128((const __m128i*) pos0);
54 const __m128i vperm1 = _mm_load_si128((const __m128i*) pos1);
55 const __m128i vperm2 = _mm_load_si128((const __m128i*) pos2);
56 const __m128i vperm3 = _mm_load_si128((const __m128i*) pos3);
57 const __m128i vperm4 = _mm_load_si128((const __m128i*) pos4);
58 const __m128i vperm5 = _mm_load_si128((const __m128i*) pos5);
59 do {
60 if XNN_UNPREDICTABLE(block_width < 2) {
61 o1 = o0;
62 }
63 if XNN_UNPREDICTABLE(block_width <= 2) {
64 o2 = o0;
65 }
66 if XNN_UNPREDICTABLE(block_width < 4) {
67 o3 = o0;
68 }
69 size_t bh = block_height;
70 for (; bh >= 4; bh -= 4) {
71 const __m128i v0 = _mm_loadu_si128((const __m128i*) i0);
72 const __m128i v1 = _mm_loadu_si128((const __m128i*) i1);
73 const __m128i v2 = _mm_loadu_si128((const __m128i*) i2);
74 const __m128i v3 = _mm_loadu_si128((const __m128i*) i3);
75 i0 = (const uint8_t*) ((uintptr_t) i0 + tile_stride);
76 i1 = (const uint8_t*) ((uintptr_t) i1 + tile_stride);
77 i2 = (const uint8_t*) ((uintptr_t) i2 + tile_stride);
78 i3 = (const uint8_t*) ((uintptr_t) i3 + tile_stride);
79
80 const __m128i v1_0 = _mm_unpacklo_epi8(v0, v1);
81 const __m128i v1_1 = _mm_unpackhi_epi8(v0, v1);
82 const __m128i v1_2 = _mm_unpacklo_epi8(v2, v3);
83 const __m128i v1_3 = _mm_unpackhi_epi8(v2, v3);
84
85 const __m128i v3_0 = _mm_unpacklo_epi8(v1_0, v1_2);
86 const __m128i v3_1 = _mm_unpackhi_epi8(v1_0, v1_2);
87 const __m128i v3_2 = _mm_unpacklo_epi8(v1_1, v1_3);
88
89 __m128i v4_0 = _mm_shuffle_epi8(v3_0, vperm0);
90 __m128i v4_1 = _mm_or_si128(_mm_shuffle_epi8(v3_0, vperm2), _mm_shuffle_epi8(v3_1, vperm3));
91 __m128i v4_2 = _mm_or_si128(_mm_shuffle_epi8(v3_1, vperm4), _mm_shuffle_epi8(v3_2, vperm5));
92 __m128i v4_3 = _mm_shuffle_epi8(v3_2, vperm1);
93
94 _mm_storel_epi64((__m128i*) o3, v4_3);
95 _mm_storel_epi64((__m128i*) o2, v4_2);
96 _mm_storel_epi64((__m128i*) o1, v4_1);
97 _mm_storel_epi64((__m128i*) o0, v4_0);
98 o3 += 8;
99 o2 += 8;
100 o1 += 8;
101 o0 += 8;
102
103 v4_3 = _mm_unpackhi_epi64(v4_3, v4_3);
104 unaligned_store_u32(o3, (uint32_t) _mm_cvtsi128_si32(v4_3));
105 v4_2 = _mm_unpackhi_epi64(v4_2, v4_2);
106 unaligned_store_u32(o2, (uint32_t) _mm_cvtsi128_si32(v4_2));
107 v4_1 = _mm_unpackhi_epi64(v4_1, v4_1);
108 unaligned_store_u32(o1, (uint32_t) _mm_cvtsi128_si32(v4_1));
109 v4_0 = _mm_unpackhi_epi64(v4_0, v4_0);
110 unaligned_store_u32(o0, (uint32_t) _mm_cvtsi128_si32(v4_0));
111 o3 += 4;
112 o2 += 4;
113 o1 += 4;
114 o0 += 4;
115 }
116
117 if (bh != 0) {
118 if XNN_UNPREDICTABLE(bh <= 2) {
119 i2 = i0;
120 }
121 if XNN_UNPREDICTABLE(bh < 2) {
122 i1 = i0;
123 }
124 const __m128i v0 = _mm_loadu_si128((const __m128i*) i0);
125 const __m128i v1 = _mm_loadu_si128((const __m128i*) i1);
126 const __m128i v2 = _mm_loadu_si128((const __m128i*) i2);
127
128 const __m128i v1_0 = _mm_unpacklo_epi8(v0, v1);
129 const __m128i v1_1 = _mm_unpackhi_epi8(v0, v1);
130 const __m128i v1_2 = _mm_unpacklo_epi8(v2, v2);
131 const __m128i v1_3 = _mm_unpackhi_epi8(v2, v2);
132
133 const __m128i v3_0 = _mm_unpacklo_epi8(v1_0, v1_2);
134 const __m128i v3_1 = _mm_unpackhi_epi8(v1_0, v1_2);
135 const __m128i v3_2 = _mm_unpacklo_epi8(v1_1, v1_3);
136
137 __m128i v4_0 = _mm_shuffle_epi8(v3_0, vperm0);
138 __m128i v4_1 = _mm_or_si128(_mm_shuffle_epi8(v3_0, vperm2), _mm_shuffle_epi8(v3_1, vperm3));
139 __m128i v4_2 = _mm_or_si128(_mm_shuffle_epi8(v3_1, vperm4), _mm_shuffle_epi8(v3_2, vperm5));
140 __m128i v4_3 = _mm_shuffle_epi8(v3_2, vperm1);
141
142 if (bh & 2) {
143 unaligned_store_u32(o3, (uint32_t) _mm_cvtsi128_si32(v4_3));
144 unaligned_store_u32(o2, (uint32_t) _mm_cvtsi128_si32(v4_2));
145 unaligned_store_u32(o1, (uint32_t) _mm_cvtsi128_si32(v4_1));
146 unaligned_store_u32(o0, (uint32_t) _mm_cvtsi128_si32(v4_0));
147 o3 += 4;
148 o2 += 4;
149 o1 += 4;
150 o0 += 4;
151 unaligned_store_u16(o3, (uint16_t) _mm_extract_epi16(v4_3, 2));
152 unaligned_store_u16(o2, (uint16_t) _mm_extract_epi16(v4_2, 2));
153 unaligned_store_u16(o1, (uint16_t) _mm_extract_epi16(v4_1, 2));
154 unaligned_store_u16(o0, (uint16_t) _mm_extract_epi16(v4_0, 2));
155 o3 += 2;
156 o2 += 2;
157 o1 += 2;
158 o0 += 2;
159 v4_3 = _mm_bsrli_si128(v4_3, 6);
160 v4_2 = _mm_bsrli_si128(v4_2, 6);
161 v4_1 = _mm_bsrli_si128(v4_1, 6);
162 v4_0 = _mm_bsrli_si128(v4_0, 6);
163 }
164 if (bh & 1) {
165 unaligned_store_u16(o3, (uint16_t) _mm_cvtsi128_si32(v4_3));
166 unaligned_store_u16(o2, (uint16_t) _mm_cvtsi128_si32(v4_2));
167 unaligned_store_u16(o1, (uint16_t) _mm_cvtsi128_si32(v4_1));
168 unaligned_store_u16(o0, (uint16_t) _mm_cvtsi128_si32(v4_0));
169 o3 += 2;
170 o2 += 2;
171 o1 += 2;
172 o0 += 2;
173 *((uint8_t*) o3) = (uint8_t) _mm_cvtsi128_si32(_mm_bsrli_si128(v4_3, 2));
174 *((uint8_t*) o2) = (uint8_t) _mm_cvtsi128_si32(_mm_bsrli_si128(v4_2, 2));
175 *((uint8_t*) o1) = (uint8_t) _mm_cvtsi128_si32(_mm_bsrli_si128(v4_1, 2));
176 *((uint8_t*) o0) = (uint8_t) _mm_cvtsi128_si32(_mm_bsrli_si128(v4_0, 2));
177 o3 += 1;
178 o2 += 1;
179 o1 += 1;
180 o0 += 1;
181 }
182 }
183 i0 = (const uint8_t*) ((uintptr_t) i0 + input_reset);
184 i1 = (const uint8_t*) ((uintptr_t) i0 + input_stride);
185 i2 = (const uint8_t*) ((uintptr_t) i1 + input_stride);
186 i3 = (const uint8_t*) ((uintptr_t) i2 + input_stride);
187 o0 = (uint8_t*) ((uintptr_t) o0 + output_reset);
188 o1 = (uint8_t*) ((uintptr_t) o1 + output_reset);
189 o2 = (uint8_t*) ((uintptr_t) o2 + output_reset);
190 o3 = (uint8_t*) ((uintptr_t) o3 + output_reset);
191 block_width = doz(block_width, tile_width);
192 } while (block_width != 0);
193 }
194