Gitiles
Code Review Sign In
review.mlplatform.org / ml / ComputeLibrary / 4ee8b1599dbaf7634d25607fa5ac96ba3dc6b0f2 / . / src / core / NEON / kernels / arm_gemm / transforms / sve_transpose_interleave_8VL_2x4.hpp
blob: 891e3abeb00f0566192a2398fc0cca65e9db1757 [file] [log] [blame]
Georgios Pinitas4ee8b152021-07-16 16:16:43 +0100[diff] [blame^]1/*
2 * Copyright (c) 2021 Arm Limited.
3 *
4 * SPDX-License-Identifier: MIT
5 *
6 * Permission is hereby granted, free of charge, to any person obtaining a copy
7 * of this software and associated documentation files (the "Software"), to
8 * deal in the Software without restriction, including without limitation the
9 * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
10 * sell copies of the Software, and to permit persons to whom the Software is
11 * furnished to do so, subject to the following conditions:
12 *
13 * The above copyright notice and this permission notice shall be included in
14 * all copies or substantial portions of the Software.
15 *
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
19 * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
21 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
22 * IN THE SOFTWARE.
23 */
24
25#pragma once
26
27#ifdef __ARM_FEATURE_SVE
28
29
30namespace {
31
32void sve_transpose_interleave_8VL_2x4(uint16_t *out, const uint16_t *in, size_t width, size_t in_stride, size_t height)
33{
34 uint16_t *pad_row = reinterpret_cast<uint16_t *>(alloca(width * sizeof(uint16_t)));
35
36 if (height % 4) {
37 memset(pad_row, 0, width * sizeof(uint16_t));
38 }
39
40 size_t out_stride = 8 * roundup<size_t>(height, 4) * get_vector_length<uint32_t>();
41
42 __asm__ __volatile__(
43 "ptrue p2.b\n"
44 "cmp %x[height], #0x8\n"
45 "blt 6f\n"
46 "1:" // Main row loop: Head
47 "mov x11, %x[in]\n"
48 "mov x10, %x[out]\n"
49 "add x9, x11, %x[in_stride]\n"
50 "add x28, x9, %x[in_stride]\n"
51 "add x27, x28, %x[in_stride]\n"
52 "add x26, x27, %x[in_stride]\n"
53 "add x25, x26, %x[in_stride]\n"
54 "add x24, x25, %x[in_stride]\n"
55 "add x23, x24, %x[in_stride]\n"
56 "add %x[in], x23, %x[in_stride]\n"
57 "sub %x[height], %x[height], #0x8\n"
58 "mov x22, %x[width]\n"
59 "cnth x21, ALL, MUL #4\n"
60 "cmp x22, x21\n"
61 "blt 3f\n"
62 "2:" // Main row loop: Unroll column loop
63 "ld1h { z21.h }, p2/Z, [x11]\n"
64 "mov x20, x10\n"
65 "ld1h { z24.h }, p2/Z, [x11, #1, MUL VL]\n"
66 "add x10, x10, %x[out_stride]\n"
67 "ld1h { z8.h }, p2/Z, [x11, #2, MUL VL]\n"
68 "mov x19, x10\n"
69 "ld1h { z11.h }, p2/Z, [x11, #3, MUL VL]\n"
70 "addvl x11, x11, #4\n"
71 "ld1h { z3.h }, p2/Z, [x9]\n"
72 "add x10, x10, %x[out_stride]\n"
73 "ld1h { z0.h }, p2/Z, [x9, #1, MUL VL]\n"
74 "sub x22, x22, x21\n"
75 "ld1h { z18.h }, p2/Z, [x9, #2, MUL VL]\n"
76 "cmp x22, x21\n"
77 "ld1h { z12.h }, p2/Z, [x9, #3, MUL VL]\n"
78 "addvl x9, x9, #4\n"
79 "ld1h { z16.h }, p2/Z, [x28]\n"
80 "zip1 z22.h, z21.h, z16.h\n"
81 "ld1h { z19.h }, p2/Z, [x28, #1, MUL VL]\n"
82 "zip2 z21.h, z21.h, z16.h\n"
83 "ld1h { z26.h }, p2/Z, [x28, #2, MUL VL]\n"
84 "ld1h { z13.h }, p2/Z, [x28, #3, MUL VL]\n"
85 "zip1 z14.h, z24.h, z19.h\n"
86 "ld1h { z16.h }, p2/Z, [x27]\n"
87 "addvl x28, x28, #4\n"
88 "zip2 z24.h, z24.h, z19.h\n"
89 "ld1h { z27.h }, p2/Z, [x27, #1, MUL VL]\n"
90 "zip1 z17.h, z8.h, z26.h\n"
91 "ld1h { z15.h }, p2/Z, [x27, #2, MUL VL]\n"
92 "zip2 z9.h, z8.h, z26.h\n"
93 "ld1h { z5.h }, p2/Z, [x27, #3, MUL VL]\n"
94 "addvl x27, x27, #4\n"
95 "zip1 z2.h, z3.h, z16.h\n"
96 "ld1h { z4.h }, p2/Z, [x26]\n"
97 "zip2 z16.h, z3.h, z16.h\n"
98 "ld1h { z23.h }, p2/Z, [x26, #1, MUL VL]\n"
99 "zip1 z20.h, z22.h, z2.h\n"
100 "ld1h { z10.h }, p2/Z, [x26, #2, MUL VL]\n"
101 "zip2 z28.h, z22.h, z2.h\n"
102 "ld1h { z8.h }, p2/Z, [x26, #3, MUL VL]\n"
103 "addvl x26, x26, #4\n"
104 "zip1 z26.h, z21.h, z16.h\n"
105 "ld1h { z25.h }, p2/Z, [x25]\n"
106 "zip2 z7.h, z21.h, z16.h\n"
107 "ld1h { z31.h }, p2/Z, [x25, #1, MUL VL]\n"
108 "zip1 z3.h, z0.h, z27.h\n"
109 "ld1h { z16.h }, p2/Z, [x25, #2, MUL VL]\n"
110 "zip1 z22.h, z14.h, z3.h\n"
111 "ld1h { z6.h }, p2/Z, [x25, #3, MUL VL]\n"
112 "addvl x25, x25, #4\n"
113 "zip2 z19.h, z14.h, z3.h\n"
114 "ld1h { z2.h }, p2/Z, [x24]\n"
115 "zip2 z14.h, z0.h, z27.h\n"
116 "ld1h { z21.h }, p2/Z, [x24, #1, MUL VL]\n"
117 "zip1 z29.h, z24.h, z14.h\n"
118 "ld1h { z0.h }, p2/Z, [x24, #2, MUL VL]\n"
119 "zip2 z27.h, z24.h, z14.h\n"
120 "ld1h { z1.h }, p2/Z, [x24, #3, MUL VL]\n"
121 "addvl x24, x24, #4\n"
122 "zip1 z30.h, z4.h, z2.h\n"
123 "ld1h { z3.h }, p2/Z, [x23]\n"
124 "zip2 z14.h, z4.h, z2.h\n"
125 "ld1h { z4.h }, p2/Z, [x23, #1, MUL VL]\n"
126 "zip1 z2.h, z23.h, z21.h\n"
127 "ld1h { z24.h }, p2/Z, [x23, #2, MUL VL]\n"
128 "zip2 z21.h, z23.h, z21.h\n"
129 "ld1h { z23.h }, p2/Z, [x23, #3, MUL VL]\n"
130 "addvl x23, x23, #4\n"
131 "st1h { z20.h }, p2, [x20]\n"
132 "zip1 z20.h, z25.h, z3.h\n"
133 "zip2 z3.h, z25.h, z3.h\n"
134 "st1h { z28.h }, p2, [x20, #1, MUL VL]\n"
135 "zip1 z28.h, z30.h, z20.h\n"
136 "st1h { z26.h }, p2, [x20, #2, MUL VL]\n"
137 "zip2 z20.h, z30.h, z20.h\n"
138 "st1h { z7.h }, p2, [x20, #3, MUL VL]\n"
139 "zip1 z25.h, z14.h, z3.h\n"
140 "st1h { z22.h }, p2, [x20, #4, MUL VL]\n"
141 "zip2 z7.h, z14.h, z3.h\n"
142 "st1h { z19.h }, p2, [x20, #5, MUL VL]\n"
143 "zip1 z14.h, z31.h, z4.h\n"
144 "st1h { z29.h }, p2, [x20, #6, MUL VL]\n"
145 "zip1 z19.h, z2.h, z14.h\n"
146 "st1h { z27.h }, p2, [x20, #7, MUL VL]\n"
147 "addvl x20, x20, #16\n"
148 "zip2 z29.h, z2.h, z14.h\n"
149 "st1h { z28.h }, p2, [x20, #-8, MUL VL]\n"
150 "zip2 z27.h, z31.h, z4.h\n"
151 "st1h { z20.h }, p2, [x20, #-7, MUL VL]\n"
152 "zip1 z30.h, z21.h, z27.h\n"
153 "st1h { z25.h }, p2, [x20, #-6, MUL VL]\n"
154 "zip2 z20.h, z21.h, z27.h\n"
155 "st1h { z7.h }, p2, [x20, #-5, MUL VL]\n"
156 "zip1 z14.h, z18.h, z15.h\n"
157 "st1h { z19.h }, p2, [x20, #-4, MUL VL]\n"
158 "zip1 z19.h, z17.h, z14.h\n"
159 "st1h { z29.h }, p2, [x20, #-3, MUL VL]\n"
160 "zip2 z7.h, z17.h, z14.h\n"
161 "st1h { z30.h }, p2, [x20, #-2, MUL VL]\n"
162 "zip2 z14.h, z18.h, z15.h\n"
163 "st1h { z20.h }, p2, [x20, #-1, MUL VL]\n"
164 "zip1 z17.h, z9.h, z14.h\n"
165 "st1h { z19.h }, p2, [x19]\n"
166 "zip2 z27.h, z9.h, z14.h\n"
167 "st1h { z7.h }, p2, [x19, #1, MUL VL]\n"
168 "zip1 z18.h, z11.h, z13.h\n"
169 "st1h { z17.h }, p2, [x19, #2, MUL VL]\n"
170 "zip1 z17.h, z12.h, z5.h\n"
171 "st1h { z27.h }, p2, [x19, #3, MUL VL]\n"
172 "zip1 z20.h, z18.h, z17.h\n"
173 "st1h { z20.h }, p2, [x19, #4, MUL VL]\n"
174 "zip2 z18.h, z18.h, z17.h\n"
175 "st1h { z18.h }, p2, [x19, #5, MUL VL]\n"
176 "zip2 z18.h, z11.h, z13.h\n"
177 "zip2 z17.h, z12.h, z5.h\n"
178 "zip1 z29.h, z18.h, z17.h\n"
179 "st1h { z29.h }, p2, [x19, #6, MUL VL]\n"
180 "zip2 z17.h, z18.h, z17.h\n"
181 "st1h { z17.h }, p2, [x19, #7, MUL VL]\n"
182 "addvl x19, x19, #16\n"
183 "zip1 z18.h, z10.h, z0.h\n"
184 "zip1 z17.h, z16.h, z24.h\n"
185 "zip1 z30.h, z18.h, z17.h\n"
186 "st1h { z30.h }, p2, [x19, #-8, MUL VL]\n"
187 "zip2 z30.h, z18.h, z17.h\n"
188 "st1h { z30.h }, p2, [x19, #-7, MUL VL]\n"
189 "zip2 z18.h, z10.h, z0.h\n"
190 "zip2 z17.h, z16.h, z24.h\n"
191 "zip1 z16.h, z18.h, z17.h\n"
192 "st1h { z16.h }, p2, [x19, #-6, MUL VL]\n"
193 "zip2 z16.h, z18.h, z17.h\n"
194 "st1h { z16.h }, p2, [x19, #-5, MUL VL]\n"
195 "zip1 z18.h, z8.h, z1.h\n"
196 "zip1 z17.h, z6.h, z23.h\n"
197 "zip1 z16.h, z18.h, z17.h\n"
198 "st1h { z16.h }, p2, [x19, #-4, MUL VL]\n"
199 "zip2 z16.h, z18.h, z17.h\n"
200 "st1h { z16.h }, p2, [x19, #-3, MUL VL]\n"
201 "zip2 z18.h, z8.h, z1.h\n"
202 "zip2 z17.h, z6.h, z23.h\n"
203 "zip1 z16.h, z18.h, z17.h\n"
204 "st1h { z16.h }, p2, [x19, #-2, MUL VL]\n"
205 "zip2 z16.h, z18.h, z17.h\n"
206 "st1h { z16.h }, p2, [x19, #-1, MUL VL]\n"
207 "bge 2b\n"
208 "3:" // Main row loop: Unroll column loop skip
209 "cbz x22, 5f\n"
210 "4:" // Main row loop: Column loop
211 "mov x20, x22\n"
212 "mov x19, x10\n"
213 "whilelt p1.h, XZR, x20\n"
214 "ld1h { z17.h }, p1/Z, [x11]\n"
215 "ld1h { z22.h }, p1/Z, [x9]\n"
216 "dech x20\n"
217 "ld1h { z16.h }, p1/Z, [x28]\n"
218 "zip1 z19.h, z17.h, z16.h\n"
219 "ld1h { z18.h }, p1/Z, [x27]\n"
220 "whilelt p0.h, XZR, x20\n"
221 "zip2 z17.h, z17.h, z16.h\n"
222 "ld1h { z21.h }, p0/Z, [x11, #1, MUL VL]\n"
223 "addvl x11, x11, #2\n"
224 "zip1 z16.h, z22.h, z18.h\n"
225 "ld1h { z2.h }, p0/Z, [x9, #1, MUL VL]\n"
226 "addvl x9, x9, #2\n"
227 "zip1 z1.h, z19.h, z16.h\n"
228 "ld1h { z20.h }, p0/Z, [x28, #1, MUL VL]\n"
229 "addvl x28, x28, #2\n"
230 "zip2 z0.h, z19.h, z16.h\n"
231 "ld1h { z19.h }, p0/Z, [x27, #1, MUL VL]\n"
232 "addvl x27, x27, #2\n"
233 "zip2 z16.h, z22.h, z18.h\n"
234 "ld1h { z31.h }, p1/Z, [x26]\n"
235 "add x10, x10, %x[out_stride]\n"
236 "zip1 z30.h, z17.h, z16.h\n"
237 "ld1h { z29.h }, p0/Z, [x26, #1, MUL VL]\n"
238 "addvl x26, x26, #2\n"
239 "zip2 z28.h, z17.h, z16.h\n"
240 "ld1h { z27.h }, p1/Z, [x25]\n"
241 "decd x22, ALL, MUL #8\n"
242 "zip1 z17.h, z21.h, z20.h\n"
243 "ld1h { z26.h }, p0/Z, [x25, #1, MUL VL]\n"
244 "addvl x25, x25, #2\n"
245 "zip1 z16.h, z2.h, z19.h\n"
246 "ld1h { z25.h }, p1/Z, [x24]\n"
247 "cmp x22, #0x0\n"
248 "zip1 z18.h, z17.h, z16.h\n"
249 "ld1h { z24.h }, p0/Z, [x24, #1, MUL VL]\n"
250 "addvl x24, x24, #2\n"
251 "zip2 z23.h, z17.h, z16.h\n"
252 "ld1h { z22.h }, p1/Z, [x23]\n"
253 "zip2 z17.h, z21.h, z20.h\n"
254 "ld1h { z21.h }, p0/Z, [x23, #1, MUL VL]\n"
255 "addvl x23, x23, #2\n"
256 "zip2 z16.h, z2.h, z19.h\n"
257 "st1h { z1.h }, p2, [x19]\n"
258 "zip1 z20.h, z31.h, z25.h\n"
259 "st1h { z0.h }, p2, [x19, #1, MUL VL]\n"
260 "zip1 z19.h, z17.h, z16.h\n"
261 "st1h { z30.h }, p2, [x19, #2, MUL VL]\n"
262 "zip2 z17.h, z17.h, z16.h\n"
263 "st1h { z28.h }, p2, [x19, #3, MUL VL]\n"
264 "zip1 z16.h, z27.h, z22.h\n"
265 "st1h { z18.h }, p2, [x19, #4, MUL VL]\n"
266 "zip1 z18.h, z20.h, z16.h\n"
267 "st1h { z23.h }, p2, [x19, #5, MUL VL]\n"
268 "zip2 z16.h, z20.h, z16.h\n"
269 "st1h { z19.h }, p2, [x19, #6, MUL VL]\n"
270 "zip2 z19.h, z31.h, z25.h\n"
271 "st1h { z17.h }, p2, [x19, #7, MUL VL]\n"
272 "addvl x19, x19, #16\n"
273 "zip2 z17.h, z27.h, z22.h\n"
274 "st1h { z18.h }, p2, [x19, #-8, MUL VL]\n"
275 "zip1 z18.h, z29.h, z24.h\n"
276 "st1h { z16.h }, p2, [x19, #-7, MUL VL]\n"
277 "zip1 z16.h, z19.h, z17.h\n"
278 "st1h { z16.h }, p2, [x19, #-6, MUL VL]\n"
279 "zip2 z16.h, z19.h, z17.h\n"
280 "st1h { z16.h }, p2, [x19, #-5, MUL VL]\n"
281 "zip1 z17.h, z26.h, z21.h\n"
282 "zip1 z16.h, z18.h, z17.h\n"
283 "st1h { z16.h }, p2, [x19, #-4, MUL VL]\n"
284 "zip2 z16.h, z18.h, z17.h\n"
285 "st1h { z16.h }, p2, [x19, #-3, MUL VL]\n"
286 "zip2 z18.h, z29.h, z24.h\n"
287 "zip2 z17.h, z26.h, z21.h\n"
288 "zip1 z16.h, z18.h, z17.h\n"
289 "st1h { z16.h }, p2, [x19, #-2, MUL VL]\n"
290 "zip2 z16.h, z18.h, z17.h\n"
291 "st1h { z16.h }, p2, [x19, #-1, MUL VL]\n"
292 "bgt 4b\n"
293 "5:" // Main row loop: Column loop skip
294 "addvl %x[out], %x[out], #16\n"
295 "cmp %x[height], #0x8\n"
296 "bge 1b\n"
297 "cbz %x[height], 12f\n"
298 "6:" // Main loop skip
299
300 "7:" // Tail row loop: Head
301 "mov x11, %x[in]\n"
302 "mov x10, %x[out]\n"
303 "add x9, x11, %x[in_stride]\n"
304 "add x28, x9, %x[in_stride]\n"
305 "add x27, x28, %x[in_stride]\n"
306 "add %x[in], x27, %x[in_stride]\n"
307 "cmp %x[height], #0x3\n"
308 "csel x27, x27, %x[pad_row], GT\n"
309 "csel x28, x28, %x[pad_row], GE\n"
310 "cmp %x[height], #0x1\n"
311 "csel x9, x9, %x[pad_row], GT\n"
312 "sub %x[height], %x[height], #0x4\n"
313 "mov x20, %x[width]\n"
314 "cnth x19, ALL, MUL #4\n"
315 "cmp x20, x19\n"
316 "blt 9f\n"
317 "8:" // Tail row loop: Unroll column loop
318 "ld1h { z17.h }, p2/Z, [x11]\n"
319 "sub x20, x20, x19\n"
320 "ld1h { z20.h }, p2/Z, [x11, #1, MUL VL]\n"
321 "cmp x20, x19\n"
322 "ld1h { z19.h }, p2/Z, [x11, #2, MUL VL]\n"
323 "ld1h { z1.h }, p2/Z, [x11, #3, MUL VL]\n"
324 "addvl x11, x11, #4\n"
325 "ld1h { z0.h }, p2/Z, [x9]\n"
326 "ld1h { z31.h }, p2/Z, [x9, #1, MUL VL]\n"
327 "ld1h { z30.h }, p2/Z, [x9, #2, MUL VL]\n"
328 "ld1h { z29.h }, p2/Z, [x9, #3, MUL VL]\n"
329 "addvl x9, x9, #4\n"
330 "ld1h { z16.h }, p2/Z, [x28]\n"
331 "zip1 z28.h, z17.h, z16.h\n"
332 "ld1h { z18.h }, p2/Z, [x28, #1, MUL VL]\n"
333 "zip2 z27.h, z17.h, z16.h\n"
334 "ld1h { z17.h }, p2/Z, [x28, #2, MUL VL]\n"
335 "ld1h { z26.h }, p2/Z, [x28, #3, MUL VL]\n"
336 "zip1 z25.h, z20.h, z18.h\n"
337 "ld1h { z16.h }, p2/Z, [x27]\n"
338 "addvl x28, x28, #4\n"
339 "zip2 z24.h, z20.h, z18.h\n"
340 "ld1h { z23.h }, p2/Z, [x27, #1, MUL VL]\n"
341 "zip1 z22.h, z19.h, z17.h\n"
342 "ld1h { z21.h }, p2/Z, [x27, #2, MUL VL]\n"
343 "zip2 z20.h, z19.h, z17.h\n"
344 "ld1h { z19.h }, p2/Z, [x27, #3, MUL VL]\n"
345 "addvl x27, x27, #4\n"
346 "zip1 z18.h, z0.h, z16.h\n"
347 "zip2 z17.h, z0.h, z16.h\n"
348 "zip1 z16.h, z28.h, z18.h\n"
349 "st1h { z16.h }, p2, [x10]\n"
350 "zip2 z16.h, z28.h, z18.h\n"
351 "st1h { z16.h }, p2, [x10, #1, MUL VL]\n"
352 "zip1 z16.h, z27.h, z17.h\n"
353 "st1h { z16.h }, p2, [x10, #2, MUL VL]\n"
354 "zip2 z16.h, z27.h, z17.h\n"
355 "st1h { z16.h }, p2, [x10, #3, MUL VL]\n"
356 "zip1 z17.h, z31.h, z23.h\n"
357 "zip1 z16.h, z25.h, z17.h\n"
358 "st1h { z16.h }, p2, [x10, #4, MUL VL]\n"
359 "zip2 z16.h, z25.h, z17.h\n"
360 "st1h { z16.h }, p2, [x10, #5, MUL VL]\n"
361 "zip2 z17.h, z31.h, z23.h\n"
362 "zip1 z16.h, z24.h, z17.h\n"
363 "st1h { z16.h }, p2, [x10, #6, MUL VL]\n"
364 "zip2 z16.h, z24.h, z17.h\n"
365 "st1h { z16.h }, p2, [x10, #7, MUL VL]\n"
366 "add x10, x10, %x[out_stride]\n"
367 "zip1 z18.h, z30.h, z21.h\n"
368 "zip2 z17.h, z30.h, z21.h\n"
369 "zip1 z16.h, z22.h, z18.h\n"
370 "st1h { z16.h }, p2, [x10]\n"
371 "zip2 z16.h, z22.h, z18.h\n"
372 "st1h { z16.h }, p2, [x10, #1, MUL VL]\n"
373 "zip1 z16.h, z20.h, z17.h\n"
374 "st1h { z16.h }, p2, [x10, #2, MUL VL]\n"
375 "zip2 z16.h, z20.h, z17.h\n"
376 "st1h { z16.h }, p2, [x10, #3, MUL VL]\n"
377 "zip1 z18.h, z1.h, z26.h\n"
378 "zip1 z17.h, z29.h, z19.h\n"
379 "zip1 z16.h, z18.h, z17.h\n"
380 "st1h { z16.h }, p2, [x10, #4, MUL VL]\n"
381 "zip2 z16.h, z18.h, z17.h\n"
382 "st1h { z16.h }, p2, [x10, #5, MUL VL]\n"
383 "zip2 z18.h, z1.h, z26.h\n"
384 "zip2 z17.h, z29.h, z19.h\n"
385 "zip1 z16.h, z18.h, z17.h\n"
386 "st1h { z16.h }, p2, [x10, #6, MUL VL]\n"
387 "zip2 z16.h, z18.h, z17.h\n"
388 "st1h { z16.h }, p2, [x10, #7, MUL VL]\n"
389 "add x10, x10, %x[out_stride]\n"
390 "bge 8b\n"
391 "9:" // Tail row loop: Unroll column loop skip
392 "cbz x20, 11f\n"
393 "10:" // Tail row loop: Column loop
394 "mov x19, x20\n"
395 "decd x20, ALL, MUL #8\n"
396 "whilelt p0.h, XZR, x19\n"
397 "ld1h { z17.h }, p0/Z, [x11]\n"
398 "ld1h { z25.h }, p0/Z, [x9]\n"
399 "dech x19\n"
400 "ld1h { z16.h }, p0/Z, [x28]\n"
401 "zip1 z18.h, z17.h, z16.h\n"
402 "ld1h { z24.h }, p0/Z, [x27]\n"
403 "whilelt p0.h, XZR, x19\n"
404 "zip2 z23.h, z17.h, z16.h\n"
405 "ld1h { z22.h }, p0/Z, [x11, #1, MUL VL]\n"
406 "addvl x11, x11, #2\n"
407 "zip1 z16.h, z25.h, z24.h\n"
408 "ld1h { z21.h }, p0/Z, [x9, #1, MUL VL]\n"
409 "addvl x9, x9, #2\n"
410 "zip1 z17.h, z18.h, z16.h\n"
411 "ld1h { z20.h }, p0/Z, [x28, #1, MUL VL]\n"
412 "addvl x28, x28, #2\n"
413 "zip2 z18.h, z18.h, z16.h\n"
414 "ld1h { z19.h }, p0/Z, [x27, #1, MUL VL]\n"
415 "addvl x27, x27, #2\n"
416 "zip2 z16.h, z25.h, z24.h\n"
417 "st1h { z17.h }, p2, [x10]\n"
418 "cmp x20, #0x0\n"
419 "zip1 z17.h, z23.h, z16.h\n"
420 "st1h { z18.h }, p2, [x10, #1, MUL VL]\n"
421 "zip2 z16.h, z23.h, z16.h\n"
422 "st1h { z17.h }, p2, [x10, #2, MUL VL]\n"
423 "zip1 z18.h, z22.h, z20.h\n"
424 "st1h { z16.h }, p2, [x10, #3, MUL VL]\n"
425 "zip1 z17.h, z21.h, z19.h\n"
426 "zip1 z16.h, z18.h, z17.h\n"
427 "st1h { z16.h }, p2, [x10, #4, MUL VL]\n"
428 "zip2 z16.h, z18.h, z17.h\n"
429 "st1h { z16.h }, p2, [x10, #5, MUL VL]\n"
430 "zip2 z18.h, z22.h, z20.h\n"
431 "zip2 z17.h, z21.h, z19.h\n"
432 "zip1 z16.h, z18.h, z17.h\n"
433 "st1h { z16.h }, p2, [x10, #6, MUL VL]\n"
434 "zip2 z16.h, z18.h, z17.h\n"
435 "st1h { z16.h }, p2, [x10, #7, MUL VL]\n"
436 "add x10, x10, %x[out_stride]\n"
437 "bgt 10b\n"
438 "11:" // Tail row loop: Column loop skip
439 "addvl %x[out], %x[out], #8\n"
440 "cmp %x[height], #0x1\n"
441 "bge 7b\n"
442 "12:" // Done
443
444 : [height] "+&r" (height), [in] "+&r" (in), [out] "+&r" (out)
445 : [in_stride] "r" (in_stride), [out_stride] "r" (out_stride), [pad_row] "r" (pad_row), [width] "r" (width)
446 : "cc", "memory", "p0", "p1", "p2", "x9", "x10", "x11", "x19", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
447 );
448}
449
450} // anonymous namespace
451
452template<>
453void Transform<8, 4, true, VLType::SVE>(
454 bfloat16 *out, const bfloat16 *in, int stride, int x0, int xmax, int k0, int kmax)
455{
456 sve_transpose_interleave_8VL_2x4(
457 reinterpret_cast<uint16_t *>(out),
458 reinterpret_cast<const uint16_t *>(in + k0 * stride + x0),
459 (xmax-x0) * sizeof(bfloat16) / 2,
460 stride * sizeof(bfloat16),
461 (kmax-k0)
462 );
463}
464
465#endif