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 
30 namespace {
31 
sve_transpose_interleave_8VL_2x2(uint16_t * out,const uint16_t * in,size_t width,size_t in_stride,size_t height)32 void sve_transpose_interleave_8VL_2x2(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 % 2) {
37         memset(pad_row, 0, width * sizeof(uint16_t));
38     }
39 
40     size_t out_stride = 8 * roundup<size_t>(height, 2) * get_vector_length<uint16_t>();
41 
42     __asm__ __volatile__(
43       "ptrue p3.b\n"
44       "cmp %x[height], #0x4\n"
45       "blt 6f\n"
46       "1:"  // Main row loop: Head
47       "mov x27, %x[in]\n"
48       "mov x26, %x[out]\n"
49       "add x25, x27, %x[in_stride]\n"
50       "add x24, x25, %x[in_stride]\n"
51       "add x23, x24, %x[in_stride]\n"
52       "add %x[in], x23, %x[in_stride]\n"
53       "sub %x[height], %x[height], #0x4\n"
54       "mov x22, %x[width]\n"
55       "cnth x21, ALL, MUL #8\n"
56       "cmp x22, x21\n"
57       "blt 3f\n"
58       "2:"  // Main row loop: Unroll column loop
59       "ld1h { z8.h }, p3/Z, [x27]\n"
60       "mov x20, x26\n"
61       "ld1h { z3.h }, p3/Z, [x27, #1, MUL VL]\n"
62       "add x26, x26, %x[out_stride]\n"
63       "ld1h { z22.h }, p3/Z, [x27, #2, MUL VL]\n"
64       "mov x19, x26\n"
65       "ld1h { z12.h }, p3/Z, [x27, #3, MUL VL]\n"
66       "add x26, x26, %x[out_stride]\n"
67       "ld1h { z4.h }, p3/Z, [x27, #4, MUL VL]\n"
68       "sub x22, x22, x21\n"
69       "ld1h { z25.h }, p3/Z, [x27, #5, MUL VL]\n"
70       "cmp x22, x21\n"
71       "ld1h { z15.h }, p3/Z, [x27, #6, MUL VL]\n"
72       "ld1h { z2.h }, p3/Z, [x27, #7, MUL VL]\n"
73       "addvl x27, x27, #8\n"
74       "ld1h { z16.h }, p3/Z, [x25]\n"
75       "zip1 z21.h, z8.h, z16.h\n"
76       "ld1h { z27.h }, p3/Z, [x25, #1, MUL VL]\n"
77       "zip2 z7.h, z8.h, z16.h\n"
78       "ld1h { z18.h }, p3/Z, [x25, #2, MUL VL]\n"
79       "ld1h { z30.h }, p3/Z, [x25, #3, MUL VL]\n"
80       "zip1 z19.h, z3.h, z27.h\n"
81       "ld1h { z0.h }, p3/Z, [x25, #4, MUL VL]\n"
82       "zip2 z16.h, z3.h, z27.h\n"
83       "ld1h { z27.h }, p3/Z, [x25, #5, MUL VL]\n"
84       "zip1 z13.h, z22.h, z18.h\n"
85       "ld1h { z26.h }, p3/Z, [x25, #6, MUL VL]\n"
86       "zip2 z29.h, z22.h, z18.h\n"
87       "ld1h { z24.h }, p3/Z, [x25, #7, MUL VL]\n"
88       "addvl x25, x25, #8\n"
89       "zip1 z20.h, z12.h, z30.h\n"
90       "ld1h { z9.h }, p3/Z, [x24]\n"
91       "zip2 z14.h, z12.h, z30.h\n"
92       "ld1h { z12.h }, p3/Z, [x24, #1, MUL VL]\n"
93       "zip1 z5.h, z4.h, z0.h\n"
94       "ld1h { z31.h }, p3/Z, [x24, #2, MUL VL]\n"
95       "zip2 z1.h, z4.h, z0.h\n"
96       "ld1h { z22.h }, p3/Z, [x24, #3, MUL VL]\n"
97       "zip1 z10.h, z25.h, z27.h\n"
98       "ld1h { z3.h }, p3/Z, [x24, #4, MUL VL]\n"
99       "zip2 z6.h, z25.h, z27.h\n"
100       "ld1h { z4.h }, p3/Z, [x24, #5, MUL VL]\n"
101       "zip1 z8.h, z15.h, z26.h\n"
102       "ld1h { z25.h }, p3/Z, [x24, #6, MUL VL]\n"
103       "zip2 z11.h, z15.h, z26.h\n"
104       "ld1h { z30.h }, p3/Z, [x24, #7, MUL VL]\n"
105       "addvl x24, x24, #8\n"
106       "zip1 z17.h, z2.h, z24.h\n"
107       "ld1h { z23.h }, p3/Z, [x23]\n"
108       "zip2 z0.h, z2.h, z24.h\n"
109       "ld1h { z28.h }, p3/Z, [x23, #1, MUL VL]\n"
110       "ld1h { z15.h }, p3/Z, [x23, #2, MUL VL]\n"
111       "zip1 z18.h, z9.h, z23.h\n"
112       "ld1h { z26.h }, p3/Z, [x23, #3, MUL VL]\n"
113       "zip2 z27.h, z9.h, z23.h\n"
114       "ld1h { z2.h }, p3/Z, [x23, #4, MUL VL]\n"
115       "zip1 z9.h, z12.h, z28.h\n"
116       "ld1h { z24.h }, p3/Z, [x23, #5, MUL VL]\n"
117       "zip2 z12.h, z12.h, z28.h\n"
118       "ld1h { z23.h }, p3/Z, [x23, #6, MUL VL]\n"
119       "zip1 z28.h, z31.h, z15.h\n"
120       "zip2 z31.h, z31.h, z15.h\n"
121       "ld1h { z15.h }, p3/Z, [x23, #7, MUL VL]\n"
122       "addvl x23, x23, #8\n"
123       "st1h { z21.h }, p3, [x20]\n"
124       "zip1 z21.h, z22.h, z26.h\n"
125       "zip2 z26.h, z22.h, z26.h\n"
126       "st1h { z7.h }, p3, [x20, #1, MUL VL]\n"
127       "zip1 z7.h, z3.h, z2.h\n"
128       "st1h { z19.h }, p3, [x20, #2, MUL VL]\n"
129       "zip2 z22.h, z3.h, z2.h\n"
130       "st1h { z16.h }, p3, [x20, #3, MUL VL]\n"
131       "zip1 z2.h, z4.h, z24.h\n"
132       "st1h { z13.h }, p3, [x20, #4, MUL VL]\n"
133       "zip2 z3.h, z4.h, z24.h\n"
134       "st1h { z29.h }, p3, [x20, #5, MUL VL]\n"
135       "zip1 z4.h, z25.h, z23.h\n"
136       "st1h { z20.h }, p3, [x20, #6, MUL VL]\n"
137       "zip2 z20.h, z25.h, z23.h\n"
138       "st1h { z14.h }, p3, [x20, #7, MUL VL]\n"
139       "addvl x20, x20, #16\n"
140       "zip1 z25.h, z30.h, z15.h\n"
141       "st1h { z18.h }, p3, [x20, #-8, MUL VL]\n"
142       "zip2 z18.h, z30.h, z15.h\n"
143       "st1h { z27.h }, p3, [x20, #-7, MUL VL]\n"
144       "st1h { z9.h }, p3, [x20, #-6, MUL VL]\n"
145       "st1h { z12.h }, p3, [x20, #-5, MUL VL]\n"
146       "st1h { z28.h }, p3, [x20, #-4, MUL VL]\n"
147       "st1h { z31.h }, p3, [x20, #-3, MUL VL]\n"
148       "st1h { z21.h }, p3, [x20, #-2, MUL VL]\n"
149       "st1h { z26.h }, p3, [x20, #-1, MUL VL]\n"
150       "st1h { z5.h }, p3, [x19]\n"
151       "st1h { z1.h }, p3, [x19, #1, MUL VL]\n"
152       "st1h { z10.h }, p3, [x19, #2, MUL VL]\n"
153       "st1h { z6.h }, p3, [x19, #3, MUL VL]\n"
154       "st1h { z8.h }, p3, [x19, #4, MUL VL]\n"
155       "st1h { z11.h }, p3, [x19, #5, MUL VL]\n"
156       "st1h { z17.h }, p3, [x19, #6, MUL VL]\n"
157       "st1h { z0.h }, p3, [x19, #7, MUL VL]\n"
158       "addvl x19, x19, #16\n"
159       "st1h { z7.h }, p3, [x19, #-8, MUL VL]\n"
160       "st1h { z22.h }, p3, [x19, #-7, MUL VL]\n"
161       "st1h { z2.h }, p3, [x19, #-6, MUL VL]\n"
162       "st1h { z3.h }, p3, [x19, #-5, MUL VL]\n"
163       "st1h { z4.h }, p3, [x19, #-4, MUL VL]\n"
164       "st1h { z20.h }, p3, [x19, #-3, MUL VL]\n"
165       "st1h { z25.h }, p3, [x19, #-2, MUL VL]\n"
166       "st1h { z18.h }, p3, [x19, #-1, MUL VL]\n"
167       "bge 2b\n"
168       "3:"  // Main row loop: Unroll column loop skip
169       "cbz x22, 5f\n"
170       "4:"  // Main row loop: Column loop
171       "mov x20, x22\n"
172       "mov x19, x26\n"
173       "whilelt p0.h, XZR, x20\n"
174       "ld1h { z17.h }, p0/Z, [x27]\n"
175       "ld1h { z16.h }, p0/Z, [x25]\n"
176       "zip1 z0.h, z17.h, z16.h\n"
177       "ld1h { z18.h }, p0/Z, [x24]\n"
178       "dech x20\n"
179       "zip2 z31.h, z17.h, z16.h\n"
180       "ld1h { z16.h }, p0/Z, [x23]\n"
181       "whilelt p2.h, XZR, x20\n"
182       "zip1 z30.h, z18.h, z16.h\n"
183       "ld1h { z17.h }, p2/Z, [x27, #1, MUL VL]\n"
184       "dech x20\n"
185       "zip2 z29.h, z18.h, z16.h\n"
186       "ld1h { z16.h }, p2/Z, [x25, #1, MUL VL]\n"
187       "whilelt p1.h, XZR, x20\n"
188       "zip1 z28.h, z17.h, z16.h\n"
189       "ld1h { z18.h }, p1/Z, [x27, #2, MUL VL]\n"
190       "dech x20\n"
191       "zip2 z27.h, z17.h, z16.h\n"
192       "ld1h { z16.h }, p1/Z, [x25, #2, MUL VL]\n"
193       "whilelt p0.h, XZR, x20\n"
194       "zip1 z26.h, z18.h, z16.h\n"
195       "ld1h { z17.h }, p0/Z, [x27, #3, MUL VL]\n"
196       "addvl x27, x27, #4\n"
197       "zip2 z25.h, z18.h, z16.h\n"
198       "ld1h { z16.h }, p0/Z, [x25, #3, MUL VL]\n"
199       "addvl x25, x25, #4\n"
200       "zip1 z24.h, z17.h, z16.h\n"
201       "ld1h { z19.h }, p2/Z, [x24, #1, MUL VL]\n"
202       "add x26, x26, %x[out_stride]\n"
203       "zip2 z23.h, z17.h, z16.h\n"
204       "ld1h { z18.h }, p1/Z, [x24, #2, MUL VL]\n"
205       "decw x22, ALL, MUL #8\n"
206       "ld1h { z22.h }, p0/Z, [x24, #3, MUL VL]\n"
207       "addvl x24, x24, #4\n"
208       "ld1h { z16.h }, p2/Z, [x23, #1, MUL VL]\n"
209       "zip1 z21.h, z19.h, z16.h\n"
210       "ld1h { z17.h }, p1/Z, [x23, #2, MUL VL]\n"
211       "cmp x22, #0x0\n"
212       "zip2 z20.h, z19.h, z16.h\n"
213       "ld1h { z16.h }, p0/Z, [x23, #3, MUL VL]\n"
214       "addvl x23, x23, #4\n"
215       "zip1 z19.h, z18.h, z17.h\n"
216       "st1h { z0.h }, p3, [x19]\n"
217       "zip2 z18.h, z18.h, z17.h\n"
218       "st1h { z31.h }, p3, [x19, #1, MUL VL]\n"
219       "zip1 z17.h, z22.h, z16.h\n"
220       "st1h { z28.h }, p3, [x19, #2, MUL VL]\n"
221       "zip2 z16.h, z22.h, z16.h\n"
222       "st1h { z27.h }, p3, [x19, #3, MUL VL]\n"
223       "st1h { z26.h }, p3, [x19, #4, MUL VL]\n"
224       "st1h { z25.h }, p3, [x19, #5, MUL VL]\n"
225       "st1h { z24.h }, p3, [x19, #6, MUL VL]\n"
226       "st1h { z23.h }, p3, [x19, #7, MUL VL]\n"
227       "addvl x19, x19, #16\n"
228       "st1h { z30.h }, p3, [x19, #-8, MUL VL]\n"
229       "st1h { z29.h }, p3, [x19, #-7, MUL VL]\n"
230       "st1h { z21.h }, p3, [x19, #-6, MUL VL]\n"
231       "st1h { z20.h }, p3, [x19, #-5, MUL VL]\n"
232       "st1h { z19.h }, p3, [x19, #-4, MUL VL]\n"
233       "st1h { z18.h }, p3, [x19, #-3, MUL VL]\n"
234       "st1h { z17.h }, p3, [x19, #-2, MUL VL]\n"
235       "st1h { z16.h }, p3, [x19, #-1, MUL VL]\n"
236       "bgt 4b\n"
237       "5:"  // Main row loop: Column loop skip
238       "addvl %x[out], %x[out], #16\n"
239       "cmp %x[height], #0x4\n"
240       "bge 1b\n"
241       "cbz %x[height], 12f\n"
242       "6:"  // Main loop skip
243 
244       "7:"  // Tail row loop: Head
245       "mov x27, %x[in]\n"
246       "mov x26, %x[out]\n"
247       "add x25, x27, %x[in_stride]\n"
248       "add %x[in], x25, %x[in_stride]\n"
249       "cmp %x[height], #0x1\n"
250       "csel x25, x25, %x[pad_row], GT\n"
251       "sub %x[height], %x[height], #0x2\n"
252       "mov x20, %x[width]\n"
253       "cnth x19, ALL, MUL #8\n"
254       "cmp x20, x19\n"
255       "blt 9f\n"
256       "8:"  // Tail row loop: Unroll column loop
257       "ld1h { z17.h }, p3/Z, [x27]\n"
258       "sub x20, x20, x19\n"
259       "ld1h { z20.h }, p3/Z, [x27, #1, MUL VL]\n"
260       "cmp x20, x19\n"
261       "ld1h { z19.h }, p3/Z, [x27, #2, MUL VL]\n"
262       "ld1h { z1.h }, p3/Z, [x27, #3, MUL VL]\n"
263       "ld1h { z0.h }, p3/Z, [x27, #4, MUL VL]\n"
264       "ld1h { z31.h }, p3/Z, [x27, #5, MUL VL]\n"
265       "ld1h { z30.h }, p3/Z, [x27, #6, MUL VL]\n"
266       "ld1h { z29.h }, p3/Z, [x27, #7, MUL VL]\n"
267       "addvl x27, x27, #8\n"
268       "ld1h { z16.h }, p3/Z, [x25]\n"
269       "zip1 z28.h, z17.h, z16.h\n"
270       "ld1h { z18.h }, p3/Z, [x25, #1, MUL VL]\n"
271       "zip2 z27.h, z17.h, z16.h\n"
272       "ld1h { z17.h }, p3/Z, [x25, #2, MUL VL]\n"
273       "ld1h { z16.h }, p3/Z, [x25, #3, MUL VL]\n"
274       "zip1 z26.h, z20.h, z18.h\n"
275       "ld1h { z22.h }, p3/Z, [x25, #4, MUL VL]\n"
276       "zip2 z21.h, z20.h, z18.h\n"
277       "ld1h { z25.h }, p3/Z, [x25, #5, MUL VL]\n"
278       "zip1 z20.h, z19.h, z17.h\n"
279       "ld1h { z24.h }, p3/Z, [x25, #6, MUL VL]\n"
280       "zip2 z19.h, z19.h, z17.h\n"
281       "ld1h { z23.h }, p3/Z, [x25, #7, MUL VL]\n"
282       "addvl x25, x25, #8\n"
283       "zip1 z18.h, z1.h, z16.h\n"
284       "st1h { z28.h }, p3, [x26]\n"
285       "zip2 z17.h, z1.h, z16.h\n"
286       "st1h { z27.h }, p3, [x26, #1, MUL VL]\n"
287       "zip1 z16.h, z0.h, z22.h\n"
288       "st1h { z26.h }, p3, [x26, #2, MUL VL]\n"
289       "zip2 z22.h, z0.h, z22.h\n"
290       "st1h { z21.h }, p3, [x26, #3, MUL VL]\n"
291       "zip1 z21.h, z31.h, z25.h\n"
292       "st1h { z20.h }, p3, [x26, #4, MUL VL]\n"
293       "zip2 z20.h, z31.h, z25.h\n"
294       "st1h { z19.h }, p3, [x26, #5, MUL VL]\n"
295       "zip1 z19.h, z30.h, z24.h\n"
296       "st1h { z18.h }, p3, [x26, #6, MUL VL]\n"
297       "zip2 z18.h, z30.h, z24.h\n"
298       "st1h { z17.h }, p3, [x26, #7, MUL VL]\n"
299       "add x26, x26, %x[out_stride]\n"
300       "zip1 z17.h, z29.h, z23.h\n"
301       "st1h { z16.h }, p3, [x26]\n"
302       "zip2 z16.h, z29.h, z23.h\n"
303       "st1h { z22.h }, p3, [x26, #1, MUL VL]\n"
304       "st1h { z21.h }, p3, [x26, #2, MUL VL]\n"
305       "st1h { z20.h }, p3, [x26, #3, MUL VL]\n"
306       "st1h { z19.h }, p3, [x26, #4, MUL VL]\n"
307       "st1h { z18.h }, p3, [x26, #5, MUL VL]\n"
308       "st1h { z17.h }, p3, [x26, #6, MUL VL]\n"
309       "st1h { z16.h }, p3, [x26, #7, MUL VL]\n"
310       "add x26, x26, %x[out_stride]\n"
311       "bge 8b\n"
312       "9:"  // Tail row loop: Unroll column loop skip
313       "cbz x20, 11f\n"
314       "10:"  // Tail row loop: Column loop
315       "mov x19, x20\n"
316       "decw x20, ALL, MUL #8\n"
317       "whilelt p0.h, XZR, x19\n"
318       "ld1h { z17.h }, p0/Z, [x27]\n"
319       "ld1h { z16.h }, p0/Z, [x25]\n"
320       "zip1 z24.h, z17.h, z16.h\n"
321       "dech x19\n"
322       "zip2 z23.h, z17.h, z16.h\n"
323       "whilelt p0.h, XZR, x19\n"
324       "ld1h { z18.h }, p0/Z, [x27, #1, MUL VL]\n"
325       "dech x19\n"
326       "ld1h { z16.h }, p0/Z, [x25, #1, MUL VL]\n"
327       "zip1 z22.h, z18.h, z16.h\n"
328       "whilelt p0.h, XZR, x19\n"
329       "ld1h { z17.h }, p0/Z, [x27, #2, MUL VL]\n"
330       "zip2 z21.h, z18.h, z16.h\n"
331       "dech x19\n"
332       "ld1h { z16.h }, p0/Z, [x25, #2, MUL VL]\n"
333       "zip1 z20.h, z17.h, z16.h\n"
334       "whilelt p0.h, XZR, x19\n"
335       "ld1h { z19.h }, p0/Z, [x27, #3, MUL VL]\n"
336       "zip2 z18.h, z17.h, z16.h\n"
337       "addvl x27, x27, #4\n"
338       "ld1h { z16.h }, p0/Z, [x25, #3, MUL VL]\n"
339       "zip1 z17.h, z19.h, z16.h\n"
340       "st1h { z24.h }, p3, [x26]\n"
341       "addvl x25, x25, #4\n"
342       "zip2 z16.h, z19.h, z16.h\n"
343       "st1h { z23.h }, p3, [x26, #1, MUL VL]\n"
344       "cmp x20, #0x0\n"
345       "st1h { z22.h }, p3, [x26, #2, MUL VL]\n"
346       "st1h { z21.h }, p3, [x26, #3, MUL VL]\n"
347       "st1h { z20.h }, p3, [x26, #4, MUL VL]\n"
348       "st1h { z18.h }, p3, [x26, #5, MUL VL]\n"
349       "st1h { z17.h }, p3, [x26, #6, MUL VL]\n"
350       "st1h { z16.h }, p3, [x26, #7, MUL VL]\n"
351       "add x26, x26, %x[out_stride]\n"
352       "bgt 10b\n"
353       "11:"  // Tail row loop: Column loop skip
354       "addvl %x[out], %x[out], #8\n"
355       "cmp %x[height], #0x1\n"
356       "bge 7b\n"
357       "12:"  // Done
358 
359       : [height] "+&r" (height), [in] "+&r" (in), [out] "+&r" (out)
360       : [in_stride] "r" (in_stride), [out_stride] "r" (out_stride), [pad_row] "r" (pad_row), [width] "r" (width)
361       : "cc", "memory", "p0", "p1", "p2", "p3", "x19", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "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"
362     );
363 }
364 
365 } // anonymous namespace
366 
367 template<>
Transform(bfloat16 * out,const bfloat16 * in,int stride,int x0,int xmax,int k0,int kmax)368 void Transform<8, 2, true, VLType::SVE>(
369     bfloat16 *out, const bfloat16 *in, int stride, int x0, int xmax, int k0, int kmax)
370 {
371     sve_transpose_interleave_8VL_2x2(
372         reinterpret_cast<uint16_t *>(out),
373         reinterpret_cast<const uint16_t *>(in + k0 * stride + x0),
374         (xmax-x0) * sizeof(bfloat16) / 2,
375         stride * sizeof(bfloat16),
376         (kmax-k0)
377     );
378 }
379 
380 #endif
381