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_x16_transposec_ukernel__8x8_multi_switch_sse2(const uint16_t * input,uint16_t * output,size_t input_stride,size_t output_stride,size_t block_width,size_t block_height)20 void xnn_x16_transposec_ukernel__8x8_multi_switch_sse2(
21 const uint16_t* input,
22 uint16_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(uint16_t));
29 assert(input_stride >= block_width * sizeof(uint16_t));
30
31 const size_t tile_height = 8;
32 const size_t tile_width = 8;
33 const size_t tile_hbytes = tile_height * sizeof(uint16_t);
34 const size_t tile_wbytes = tile_width * sizeof(uint16_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(uint16_t);
38
39 const uint16_t* i0 = input;
40 const uint16_t* i1 = (const uint16_t*) ((uintptr_t) i0 + input_stride);
41 const uint16_t* i2 = (const uint16_t*) ((uintptr_t) i1 + input_stride);
42 const uint16_t* i3 = (const uint16_t*) ((uintptr_t) i2 + input_stride);
43 const uint16_t* i4 = (const uint16_t*) ((uintptr_t) i3 + input_stride);
44 const uint16_t* i5 = (const uint16_t*) ((uintptr_t) i4 + input_stride);
45 const uint16_t* i6 = (const uint16_t*) ((uintptr_t) i5 + input_stride);
46 const uint16_t* i7 = (const uint16_t*) ((uintptr_t) i6 + input_stride);
47 uint16_t* o = (uint16_t*) output;
48 const size_t minus_output_stride = -output_stride;
49
50 do {
51 const size_t rem = min(block_width - 1, 7);
52 const size_t oN_stride = rem * output_stride;
53 size_t bh = block_height;
54 for (; bh >= 8; bh -= 8) {
55 const __m128i v3_0 = _mm_loadu_si128((const __m128i*) i0);
56 i0 = (uint16_t*) ((uintptr_t) i0 + input_offset);
57 const __m128i v3_1 = _mm_loadu_si128((const __m128i*) i1);
58 i1 = (uint16_t*) ((uintptr_t) i1 + input_offset);
59 const __m128i v3_2 = _mm_loadu_si128((const __m128i*) i2);
60 i2 = (uint16_t*) ((uintptr_t) i2 + input_offset);
61 const __m128i v3_3 = _mm_loadu_si128((const __m128i*) i3);
62 i3 = (uint16_t*) ((uintptr_t) i3 + input_offset);
63 const __m128i v3_4 = _mm_loadu_si128((const __m128i*) i4);
64 i4 = (uint16_t*) ((uintptr_t) i4 + input_offset);
65 const __m128i v3_5 = _mm_loadu_si128((const __m128i*) i5);
66 i5 = (uint16_t*) ((uintptr_t) i5 + input_offset);
67 const __m128i v3_6 = _mm_loadu_si128((const __m128i*) i6);
68 i6 = (uint16_t*) ((uintptr_t) i6 + input_offset);
69 const __m128i v3_7 = _mm_loadu_si128((const __m128i*) i7);
70 i7 = (uint16_t*) ((uintptr_t) i7 + input_offset);
71
72 const __m128i v2_0 = _mm_unpacklo_epi16(v3_0, v3_1);
73 const __m128i v2_1 = _mm_unpackhi_epi16(v3_0, v3_1);
74 const __m128i v2_2 = _mm_unpacklo_epi16(v3_2, v3_3);
75 const __m128i v2_3 = _mm_unpackhi_epi16(v3_2, v3_3);
76 const __m128i v2_4 = _mm_unpacklo_epi16(v3_4, v3_5);
77 const __m128i v2_5 = _mm_unpackhi_epi16(v3_4, v3_5);
78 const __m128i v2_6 = _mm_unpacklo_epi16(v3_6, v3_7);
79 const __m128i v2_7 = _mm_unpackhi_epi16(v3_6, v3_7);
80
81 const __m128i v1_0 = _mm_unpacklo_epi32(v2_0, v2_2);
82 const __m128i v1_1 = _mm_unpackhi_epi32(v2_0, v2_2);
83 const __m128i v1_2 = _mm_unpacklo_epi32(v2_1, v2_3);
84 const __m128i v1_3 = _mm_unpackhi_epi32(v2_1, v2_3);
85 const __m128i v1_4 = _mm_unpacklo_epi32(v2_4, v2_6);
86 const __m128i v1_5 = _mm_unpackhi_epi32(v2_4, v2_6);
87 const __m128i v1_6 = _mm_unpacklo_epi32(v2_5, v2_7);
88 const __m128i v1_7 = _mm_unpackhi_epi32(v2_5, v2_7);
89
90 const __m128i v0_0 = _mm_unpacklo_epi64(v1_0, v1_4);
91 const __m128i v0_1 = _mm_unpackhi_epi64(v1_0, v1_4);
92 const __m128i v0_2 = _mm_unpacklo_epi64(v1_1, v1_5);
93 const __m128i v0_3 = _mm_unpackhi_epi64(v1_1, v1_5);
94 const __m128i v0_4 = _mm_unpacklo_epi64(v1_2, v1_6);
95 const __m128i v0_5 = _mm_unpackhi_epi64(v1_2, v1_6);
96 const __m128i v0_6 = _mm_unpacklo_epi64(v1_3, v1_7);
97 const __m128i v0_7 = _mm_unpackhi_epi64(v1_3, v1_7);
98
99
100 uint16_t* oN = (uint16_t*) ((uintptr_t) o + oN_stride);
101 switch (rem) {
102 case 7:
103 _mm_storeu_si128((__m128i*) oN, v0_7);
104 oN = (uint16_t*) ((uintptr_t) oN + minus_output_stride);
105 case 6:
106 _mm_storeu_si128((__m128i*) oN, v0_6);
107 oN = (uint16_t*) ((uintptr_t) oN + minus_output_stride);
108 case 5:
109 _mm_storeu_si128((__m128i*) oN, v0_5);
110 oN = (uint16_t*) ((uintptr_t) oN + minus_output_stride);
111 case 4:
112 _mm_storeu_si128((__m128i*) oN, v0_4);
113 oN = (uint16_t*) ((uintptr_t) oN + minus_output_stride);
114 case 3:
115 _mm_storeu_si128((__m128i*) oN, v0_3);
116 oN = (uint16_t*) ((uintptr_t) oN + minus_output_stride);
117 case 2:
118 _mm_storeu_si128((__m128i*) oN, v0_2);
119 oN = (uint16_t*) ((uintptr_t) oN + minus_output_stride);
120 case 1:
121 _mm_storeu_si128((__m128i*) oN, v0_1);
122 case 0:
123 _mm_storeu_si128((__m128i*) o, v0_0);
124 o = (uint16_t*) ((uintptr_t) o + tile_hbytes);
125 break;
126 default:
127 XNN_UNREACHABLE;
128 }
129 }
130 if (bh != 0) {
131 const __m128i v3_0 = _mm_loadu_si128((const __m128i*) i0);
132 if XNN_UNPREDICTABLE(bh < 2) {
133 i1 = i0;
134 }
135 const __m128i v3_1 = _mm_loadu_si128((const __m128i*) i1);
136 if XNN_UNPREDICTABLE(bh <= 2) {
137 i2 = i0;
138 }
139 const __m128i v3_2 = _mm_loadu_si128((const __m128i*) i2);
140 if XNN_UNPREDICTABLE(bh < 4) {
141 i3 = i0;
142 }
143 const __m128i v3_3 = _mm_loadu_si128((const __m128i*) i3);
144 if XNN_UNPREDICTABLE(bh <= 4) {
145 i4 = i0;
146 }
147 const __m128i v3_4 = _mm_loadu_si128((const __m128i*) i4);
148 if XNN_UNPREDICTABLE(bh < 6) {
149 i5 = i0;
150 }
151 const __m128i v3_5 = _mm_loadu_si128((const __m128i*) i5);
152 if XNN_UNPREDICTABLE(bh <= 6) {
153 i6 = i0;
154 }
155 const __m128i v3_6 = _mm_loadu_si128((const __m128i*) i6);
156 const __m128i v3_7 = _mm_undefined_si128();
157
158 const __m128i v2_0 = _mm_unpacklo_epi16(v3_0, v3_1);
159 const __m128i v2_1 = _mm_unpackhi_epi16(v3_0, v3_1);
160 const __m128i v2_2 = _mm_unpacklo_epi16(v3_2, v3_3);
161 const __m128i v2_3 = _mm_unpackhi_epi16(v3_2, v3_3);
162 const __m128i v2_4 = _mm_unpacklo_epi16(v3_4, v3_5);
163 const __m128i v2_5 = _mm_unpackhi_epi16(v3_4, v3_5);
164 const __m128i v2_6 = _mm_unpacklo_epi16(v3_6, v3_7);
165 const __m128i v2_7 = _mm_unpackhi_epi16(v3_6, v3_7);
166
167 const __m128i v1_0 = _mm_unpacklo_epi32(v2_0, v2_2);
168 const __m128i v1_1 = _mm_unpackhi_epi32(v2_0, v2_2);
169 const __m128i v1_2 = _mm_unpacklo_epi32(v2_1, v2_3);
170 const __m128i v1_3 = _mm_unpackhi_epi32(v2_1, v2_3);
171 const __m128i v1_4 = _mm_unpacklo_epi32(v2_4, v2_6);
172 const __m128i v1_5 = _mm_unpackhi_epi32(v2_4, v2_6);
173 const __m128i v1_6 = _mm_unpacklo_epi32(v2_5, v2_7);
174 const __m128i v1_7 = _mm_unpackhi_epi32(v2_5, v2_7);
175
176 __m128i v0_0 = _mm_unpacklo_epi64(v1_0, v1_4);
177 __m128i v0_1 = _mm_unpackhi_epi64(v1_0, v1_4);
178 __m128i v0_2 = _mm_unpacklo_epi64(v1_1, v1_5);
179 __m128i v0_3 = _mm_unpackhi_epi64(v1_1, v1_5);
180 __m128i v0_4 = _mm_unpacklo_epi64(v1_2, v1_6);
181 __m128i v0_5 = _mm_unpackhi_epi64(v1_2, v1_6);
182 __m128i v0_6 = _mm_unpacklo_epi64(v1_3, v1_7);
183 __m128i v0_7 = _mm_unpackhi_epi64(v1_3, v1_7);
184
185
186 if (bh & 4) {
187 uint16_t* oN = (uint16_t*) ((uintptr_t) o + oN_stride);
188 switch (rem) {
189 case 7:
190 _mm_storel_epi64((__m128i*) oN, v0_7);
191 oN = (uint16_t*) ((uintptr_t) oN + minus_output_stride);
192 case 6:
193 _mm_storel_epi64((__m128i*) oN, v0_6);
194 oN = (uint16_t*) ((uintptr_t) oN + minus_output_stride);
195 case 5:
196 _mm_storel_epi64((__m128i*) oN, v0_5);
197 oN = (uint16_t*) ((uintptr_t) oN + minus_output_stride);
198 case 4:
199 _mm_storel_epi64((__m128i*) oN, v0_4);
200 oN = (uint16_t*) ((uintptr_t) oN + minus_output_stride);
201 case 3:
202 _mm_storel_epi64((__m128i*) oN, v0_3);
203 oN = (uint16_t*) ((uintptr_t) oN + minus_output_stride);
204 case 2:
205 _mm_storel_epi64((__m128i*) oN, v0_2);
206 oN = (uint16_t*) ((uintptr_t) oN + minus_output_stride);
207 case 1:
208 _mm_storel_epi64((__m128i*) oN, v0_1);
209 case 0:
210 _mm_storel_epi64((__m128i*) o, v0_0);
211 break;
212 default:
213 XNN_UNREACHABLE;
214 }
215 o += 4;
216 v0_0 = _mm_unpackhi_epi64(v0_0, v0_0);
217 v0_1 = _mm_unpackhi_epi64(v0_1, v0_1);
218 v0_2 = _mm_unpackhi_epi64(v0_2, v0_2);
219 v0_3 = _mm_unpackhi_epi64(v0_3, v0_3);
220 v0_4 = _mm_unpackhi_epi64(v0_4, v0_4);
221 v0_5 = _mm_unpackhi_epi64(v0_5, v0_5);
222 v0_6 = _mm_unpackhi_epi64(v0_6, v0_6);
223 v0_7 = _mm_unpackhi_epi64(v0_7, v0_7);
224 }
225
226 if (bh & 2) {
227 uint16_t* oN = (uint16_t*) ((uintptr_t) o + oN_stride);
228 switch (rem) {
229 case 7:
230 unaligned_store_u32(oN, (uint32_t) _mm_cvtsi128_si32(v0_7));
231 oN = (uint16_t*) ((uintptr_t) oN + minus_output_stride);
232 case 6:
233 unaligned_store_u32(oN, (uint32_t) _mm_cvtsi128_si32(v0_6));
234 oN = (uint16_t*) ((uintptr_t) oN + minus_output_stride);
235 case 5:
236 unaligned_store_u32(oN, (uint32_t) _mm_cvtsi128_si32(v0_5));
237 oN = (uint16_t*) ((uintptr_t) oN + minus_output_stride);
238 case 4:
239 unaligned_store_u32(oN, (uint32_t) _mm_cvtsi128_si32(v0_4));
240 oN = (uint16_t*) ((uintptr_t) oN + minus_output_stride);
241 case 3:
242 unaligned_store_u32(oN, (uint32_t) _mm_cvtsi128_si32(v0_3));
243 oN = (uint16_t*) ((uintptr_t) oN + minus_output_stride);
244 case 2:
245 unaligned_store_u32(oN, (uint32_t) _mm_cvtsi128_si32(v0_2));
246 oN = (uint16_t*) ((uintptr_t) oN + minus_output_stride);
247 case 1:
248 unaligned_store_u32(oN, (uint32_t) _mm_cvtsi128_si32(v0_1));
249 case 0:
250 unaligned_store_u32(o, (uint32_t) _mm_cvtsi128_si32(v0_0));
251 break;
252 default:
253 XNN_UNREACHABLE;
254 }
255 o += 2;
256 v0_0 = _mm_srli_epi64(v0_0, 32);
257 v0_1 = _mm_srli_epi64(v0_1, 32);
258 v0_2 = _mm_srli_epi64(v0_2, 32);
259 v0_3 = _mm_srli_epi64(v0_3, 32);
260 v0_4 = _mm_srli_epi64(v0_4, 32);
261 v0_5 = _mm_srli_epi64(v0_5, 32);
262 v0_6 = _mm_srli_epi64(v0_6, 32);
263 v0_7 = _mm_srli_epi64(v0_7, 32);
264 }
265 if (bh & 1) {
266 uint16_t* oN = (uint16_t*) ((uintptr_t) o + oN_stride);
267 switch (rem) {
268 case 7:
269 unaligned_store_u16(oN, (uint16_t) _mm_cvtsi128_si32(v0_7));
270 oN = (uint16_t*) ((uintptr_t) oN + minus_output_stride);
271 case 6:
272 unaligned_store_u16(oN, (uint16_t) _mm_cvtsi128_si32(v0_6));
273 oN = (uint16_t*) ((uintptr_t) oN + minus_output_stride);
274 case 5:
275 unaligned_store_u16(oN, (uint16_t) _mm_cvtsi128_si32(v0_5));
276 oN = (uint16_t*) ((uintptr_t) oN + minus_output_stride);
277 case 4:
278 unaligned_store_u16(oN, (uint16_t) _mm_cvtsi128_si32(v0_4));
279 oN = (uint16_t*) ((uintptr_t) oN + minus_output_stride);
280 case 3:
281 unaligned_store_u16(oN, (uint16_t) _mm_cvtsi128_si32(v0_3));
282 oN = (uint16_t*) ((uintptr_t) oN + minus_output_stride);
283 case 2:
284 unaligned_store_u16(oN, (uint16_t) _mm_cvtsi128_si32(v0_2));
285 oN = (uint16_t*) ((uintptr_t) oN + minus_output_stride);
286 case 1:
287 unaligned_store_u16(oN, (uint16_t) _mm_cvtsi128_si32(v0_1));
288 case 0:
289 unaligned_store_u16(o, (uint16_t) _mm_cvtsi128_si32(v0_0));
290 break;
291 default:
292 XNN_UNREACHABLE;
293 }
294 }
295 }
296
297 i0 = (const uint16_t*) ((uintptr_t) i0 + input_reset);
298 i1 = (const uint16_t*) ((uintptr_t) i0 + input_stride);
299 i2 = (const uint16_t*) ((uintptr_t) i1 + input_stride);
300 i3 = (const uint16_t*) ((uintptr_t) i2 + input_stride);
301 i4 = (const uint16_t*) ((uintptr_t) i3 + input_stride);
302 i5 = (const uint16_t*) ((uintptr_t) i4 + input_stride);
303 i6 = (const uint16_t*) ((uintptr_t) i5 + input_stride);
304 i7 = (const uint16_t*) ((uintptr_t) i6 + input_stride);
305 o = (uint16_t*) ((uintptr_t) o + output_reset);
306 block_width = doz(block_width, tile_width);
307 } while (block_width != 0);
308 }
309