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_4VL_2x2(uint16_t * out,const uint16_t * in,size_t width,size_t in_stride,size_t height)32 void sve_transpose_interleave_4VL_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 = 4 * roundup<size_t>(height, 2) * get_vector_length<uint16_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 { z19.h }, p2/Z, [x11, #1, MUL VL]\n"
66       "add x10, x10, %x[out_stride]\n"
67       "ld1h { z23.h }, p2/Z, [x11, #2, MUL VL]\n"
68       "mov x19, x10\n"
69       "ld1h { z31.h }, p2/Z, [x11, #3, MUL VL]\n"
70       "addvl x11, x11, #4\n"
71       "ld1h { z29.h }, p2/Z, [x9]\n"
72       "zip1 z0.h, z21.h, z29.h\n"
73       "ld1h { z17.h }, p2/Z, [x9, #1, MUL VL]\n"
74       "add x10, x10, %x[out_stride]\n"
75       "zip2 z22.h, z21.h, z29.h\n"
76       "ld1h { z15.h }, p2/Z, [x9, #2, MUL VL]\n"
77       "sub x22, x22, x21\n"
78       "zip1 z13.h, z19.h, z17.h\n"
79       "ld1h { z6.h }, p2/Z, [x9, #3, MUL VL]\n"
80       "addvl x9, x9, #4\n"
81       "zip2 z12.h, z19.h, z17.h\n"
82       "ld1h { z20.h }, p2/Z, [x28]\n"
83       "cmp x22, x21\n"
84       "zip1 z14.h, z23.h, z15.h\n"
85       "ld1h { z1.h }, p2/Z, [x28, #1, MUL VL]\n"
86       "zip2 z3.h, z23.h, z15.h\n"
87       "ld1h { z19.h }, p2/Z, [x28, #2, MUL VL]\n"
88       "zip1 z16.h, z31.h, z6.h\n"
89       "ld1h { z11.h }, p2/Z, [x28, #3, MUL VL]\n"
90       "addvl x28, x28, #4\n"
91       "zip2 z10.h, z31.h, z6.h\n"
92       "ld1h { z27.h }, p2/Z, [x27]\n"
93       "ld1h { z18.h }, p2/Z, [x27, #1, MUL VL]\n"
94       "zip1 z9.h, z20.h, z27.h\n"
95       "ld1h { z2.h }, p2/Z, [x27, #2, MUL VL]\n"
96       "zip2 z24.h, z20.h, z27.h\n"
97       "ld1h { z5.h }, p2/Z, [x27, #3, MUL VL]\n"
98       "addvl x27, x27, #4\n"
99       "zip1 z8.h, z1.h, z18.h\n"
100       "ld1h { z30.h }, p2/Z, [x26]\n"
101       "zip2 z17.h, z1.h, z18.h\n"
102       "ld1h { z28.h }, p2/Z, [x26, #1, MUL VL]\n"
103       "zip1 z6.h, z19.h, z2.h\n"
104       "ld1h { z23.h }, p2/Z, [x26, #2, MUL VL]\n"
105       "zip2 z1.h, z19.h, z2.h\n"
106       "ld1h { z25.h }, p2/Z, [x26, #3, MUL VL]\n"
107       "addvl x26, x26, #4\n"
108       "zip1 z31.h, z11.h, z5.h\n"
109       "ld1h { z21.h }, p2/Z, [x25]\n"
110       "zip2 z11.h, z11.h, z5.h\n"
111       "ld1h { z19.h }, p2/Z, [x25, #1, MUL VL]\n"
112       "ld1h { z18.h }, p2/Z, [x25, #2, MUL VL]\n"
113       "zip1 z29.h, z30.h, z21.h\n"
114       "ld1h { z26.h }, p2/Z, [x25, #3, MUL VL]\n"
115       "addvl x25, x25, #4\n"
116       "zip2 z30.h, z30.h, z21.h\n"
117       "ld1h { z21.h }, p2/Z, [x24]\n"
118       "zip1 z27.h, z28.h, z19.h\n"
119       "ld1h { z20.h }, p2/Z, [x24, #1, MUL VL]\n"
120       "zip2 z28.h, z28.h, z19.h\n"
121       "ld1h { z4.h }, p2/Z, [x24, #2, MUL VL]\n"
122       "zip1 z2.h, z23.h, z18.h\n"
123       "ld1h { z15.h }, p2/Z, [x24, #3, MUL VL]\n"
124       "addvl x24, x24, #4\n"
125       "zip2 z5.h, z23.h, z18.h\n"
126       "ld1h { z23.h }, p2/Z, [x23]\n"
127       "zip1 z7.h, z25.h, z26.h\n"
128       "ld1h { z19.h }, p2/Z, [x23, #1, MUL VL]\n"
129       "zip2 z25.h, z25.h, z26.h\n"
130       "ld1h { z18.h }, p2/Z, [x23, #2, MUL VL]\n"
131       "ld1h { z26.h }, p2/Z, [x23, #3, MUL VL]\n"
132       "addvl x23, x23, #4\n"
133       "st1h { z0.h }, p2, [x20]\n"
134       "zip1 z0.h, z21.h, z23.h\n"
135       "zip2 z23.h, z21.h, z23.h\n"
136       "st1h { z22.h }, p2, [x20, #1, MUL VL]\n"
137       "zip1 z22.h, z20.h, z19.h\n"
138       "st1h { z13.h }, p2, [x20, #2, MUL VL]\n"
139       "zip2 z21.h, z20.h, z19.h\n"
140       "st1h { z12.h }, p2, [x20, #3, MUL VL]\n"
141       "zip1 z20.h, z4.h, z18.h\n"
142       "st1h { z9.h }, p2, [x20, #4, MUL VL]\n"
143       "zip2 z19.h, z4.h, z18.h\n"
144       "st1h { z24.h }, p2, [x20, #5, MUL VL]\n"
145       "zip1 z18.h, z15.h, z26.h\n"
146       "st1h { z8.h }, p2, [x20, #6, MUL VL]\n"
147       "zip2 z9.h, z15.h, z26.h\n"
148       "st1h { z17.h }, p2, [x20, #7, MUL VL]\n"
149       "addvl x20, x20, #16\n"
150       "st1h { z29.h }, p2, [x20, #-8, MUL VL]\n"
151       "st1h { z30.h }, p2, [x20, #-7, MUL VL]\n"
152       "st1h { z27.h }, p2, [x20, #-6, MUL VL]\n"
153       "st1h { z28.h }, p2, [x20, #-5, MUL VL]\n"
154       "st1h { z0.h }, p2, [x20, #-4, MUL VL]\n"
155       "st1h { z23.h }, p2, [x20, #-3, MUL VL]\n"
156       "st1h { z22.h }, p2, [x20, #-2, MUL VL]\n"
157       "st1h { z21.h }, p2, [x20, #-1, MUL VL]\n"
158       "st1h { z14.h }, p2, [x19]\n"
159       "st1h { z3.h }, p2, [x19, #1, MUL VL]\n"
160       "st1h { z16.h }, p2, [x19, #2, MUL VL]\n"
161       "st1h { z10.h }, p2, [x19, #3, MUL VL]\n"
162       "st1h { z6.h }, p2, [x19, #4, MUL VL]\n"
163       "st1h { z1.h }, p2, [x19, #5, MUL VL]\n"
164       "st1h { z31.h }, p2, [x19, #6, MUL VL]\n"
165       "st1h { z11.h }, p2, [x19, #7, MUL VL]\n"
166       "addvl x19, x19, #16\n"
167       "st1h { z2.h }, p2, [x19, #-8, MUL VL]\n"
168       "st1h { z5.h }, p2, [x19, #-7, MUL VL]\n"
169       "st1h { z7.h }, p2, [x19, #-6, MUL VL]\n"
170       "st1h { z25.h }, p2, [x19, #-5, MUL VL]\n"
171       "st1h { z20.h }, p2, [x19, #-4, MUL VL]\n"
172       "st1h { z19.h }, p2, [x19, #-3, MUL VL]\n"
173       "st1h { z18.h }, p2, [x19, #-2, MUL VL]\n"
174       "st1h { z9.h }, p2, [x19, #-1, MUL VL]\n"
175       "bge 2b\n"
176       "3:"  // Main row loop: Unroll column loop skip
177       "cbz x22, 5f\n"
178       "4:"  // Main row loop: Column loop
179       "mov x20, x22\n"
180       "mov x19, x10\n"
181       "whilelt p1.h, XZR, x20\n"
182       "ld1h { z18.h }, p1/Z, [x11]\n"
183       "ld1h { z16.h }, p1/Z, [x9]\n"
184       "zip1 z0.h, z18.h, z16.h\n"
185       "ld1h { z17.h }, p1/Z, [x28]\n"
186       "dech x20\n"
187       "zip2 z31.h, z18.h, z16.h\n"
188       "ld1h { z16.h }, p1/Z, [x27]\n"
189       "whilelt p0.h, XZR, x20\n"
190       "zip1 z30.h, z17.h, z16.h\n"
191       "ld1h { z18.h }, p0/Z, [x11, #1, MUL VL]\n"
192       "addvl x11, x11, #2\n"
193       "zip2 z29.h, z17.h, z16.h\n"
194       "ld1h { z16.h }, p0/Z, [x9, #1, MUL VL]\n"
195       "addvl x9, x9, #2\n"
196       "zip1 z28.h, z18.h, z16.h\n"
197       "ld1h { z17.h }, p0/Z, [x28, #1, MUL VL]\n"
198       "addvl x28, x28, #2\n"
199       "zip2 z27.h, z18.h, z16.h\n"
200       "ld1h { z16.h }, p0/Z, [x27, #1, MUL VL]\n"
201       "addvl x27, x27, #2\n"
202       "zip1 z26.h, z17.h, z16.h\n"
203       "ld1h { z18.h }, p1/Z, [x26]\n"
204       "add x10, x10, %x[out_stride]\n"
205       "zip2 z25.h, z17.h, z16.h\n"
206       "ld1h { z19.h }, p0/Z, [x26, #1, MUL VL]\n"
207       "addvl x26, x26, #2\n"
208       "ld1h { z17.h }, p1/Z, [x25]\n"
209       "zip1 z24.h, z18.h, z17.h\n"
210       "ld1h { z16.h }, p0/Z, [x25, #1, MUL VL]\n"
211       "addvl x25, x25, #2\n"
212       "zip2 z23.h, z18.h, z17.h\n"
213       "ld1h { z18.h }, p1/Z, [x24]\n"
214       "decw x22, ALL, MUL #4\n"
215       "zip1 z22.h, z19.h, z16.h\n"
216       "ld1h { z21.h }, p0/Z, [x24, #1, MUL VL]\n"
217       "addvl x24, x24, #2\n"
218       "zip2 z20.h, z19.h, z16.h\n"
219       "ld1h { z17.h }, p1/Z, [x23]\n"
220       "cmp x22, #0x0\n"
221       "zip1 z19.h, z18.h, z17.h\n"
222       "ld1h { z16.h }, p0/Z, [x23, #1, MUL VL]\n"
223       "addvl x23, x23, #2\n"
224       "zip2 z18.h, z18.h, z17.h\n"
225       "st1h { z0.h }, p2, [x19]\n"
226       "st1h { z31.h }, p2, [x19, #1, MUL VL]\n"
227       "zip1 z17.h, z21.h, z16.h\n"
228       "st1h { z28.h }, p2, [x19, #2, MUL VL]\n"
229       "zip2 z16.h, z21.h, z16.h\n"
230       "st1h { z27.h }, p2, [x19, #3, MUL VL]\n"
231       "st1h { z30.h }, p2, [x19, #4, MUL VL]\n"
232       "st1h { z29.h }, p2, [x19, #5, MUL VL]\n"
233       "st1h { z26.h }, p2, [x19, #6, MUL VL]\n"
234       "st1h { z25.h }, p2, [x19, #7, MUL VL]\n"
235       "addvl x19, x19, #16\n"
236       "st1h { z24.h }, p2, [x19, #-8, MUL VL]\n"
237       "st1h { z23.h }, p2, [x19, #-7, MUL VL]\n"
238       "st1h { z22.h }, p2, [x19, #-6, MUL VL]\n"
239       "st1h { z20.h }, p2, [x19, #-5, MUL VL]\n"
240       "st1h { z19.h }, p2, [x19, #-4, MUL VL]\n"
241       "st1h { z18.h }, p2, [x19, #-3, MUL VL]\n"
242       "st1h { z17.h }, p2, [x19, #-2, MUL VL]\n"
243       "st1h { z16.h }, p2, [x19, #-1, MUL VL]\n"
244       "bgt 4b\n"
245       "5:"  // Main row loop: Column loop skip
246       "addvl %x[out], %x[out], #16\n"
247       "cmp %x[height], #0x8\n"
248       "bge 1b\n"
249       "cbz %x[height], 12f\n"
250       "6:"  // Main loop skip
251 
252       "7:"  // Tail row loop: Head
253       "mov x11, %x[in]\n"
254       "mov x10, %x[out]\n"
255       "add x9, x11, %x[in_stride]\n"
256       "add %x[in], x9, %x[in_stride]\n"
257       "cmp %x[height], #0x1\n"
258       "csel x9, x9, %x[pad_row], GT\n"
259       "sub %x[height], %x[height], #0x2\n"
260       "mov x20, %x[width]\n"
261       "cnth x19, ALL, MUL #4\n"
262       "cmp x20, x19\n"
263       "blt 9f\n"
264       "8:"  // Tail row loop: Unroll column loop
265       "ld1h { z18.h }, p2/Z, [x11]\n"
266       "sub x20, x20, x19\n"
267       "ld1h { z24.h }, p2/Z, [x11, #1, MUL VL]\n"
268       "cmp x20, x19\n"
269       "ld1h { z23.h }, p2/Z, [x11, #2, MUL VL]\n"
270       "ld1h { z22.h }, p2/Z, [x11, #3, MUL VL]\n"
271       "addvl x11, x11, #4\n"
272       "ld1h { z16.h }, p2/Z, [x9]\n"
273       "zip1 z21.h, z18.h, z16.h\n"
274       "ld1h { z17.h }, p2/Z, [x9, #1, MUL VL]\n"
275       "zip2 z16.h, z18.h, z16.h\n"
276       "ld1h { z20.h }, p2/Z, [x9, #2, MUL VL]\n"
277       "ld1h { z19.h }, p2/Z, [x9, #3, MUL VL]\n"
278       "zip1 z18.h, z24.h, z17.h\n"
279       "st1h { z21.h }, p2, [x10]\n"
280       "addvl x9, x9, #4\n"
281       "zip2 z17.h, z24.h, z17.h\n"
282       "st1h { z16.h }, p2, [x10, #1, MUL VL]\n"
283       "zip1 z16.h, z23.h, z20.h\n"
284       "st1h { z18.h }, p2, [x10, #2, MUL VL]\n"
285       "zip2 z18.h, z23.h, z20.h\n"
286       "st1h { z17.h }, p2, [x10, #3, MUL VL]\n"
287       "add x10, x10, %x[out_stride]\n"
288       "zip1 z17.h, z22.h, z19.h\n"
289       "st1h { z16.h }, p2, [x10]\n"
290       "zip2 z16.h, z22.h, z19.h\n"
291       "st1h { z18.h }, p2, [x10, #1, MUL VL]\n"
292       "st1h { z17.h }, p2, [x10, #2, MUL VL]\n"
293       "st1h { z16.h }, p2, [x10, #3, MUL VL]\n"
294       "add x10, x10, %x[out_stride]\n"
295       "bge 8b\n"
296       "9:"  // Tail row loop: Unroll column loop skip
297       "cbz x20, 11f\n"
298       "10:"  // Tail row loop: Column loop
299       "mov x19, x20\n"
300       "decw x20, ALL, MUL #4\n"
301       "whilelt p0.h, XZR, x19\n"
302       "ld1h { z17.h }, p0/Z, [x11]\n"
303       "ld1h { z16.h }, p0/Z, [x9]\n"
304       "zip1 z20.h, z17.h, z16.h\n"
305       "dech x19\n"
306       "zip2 z19.h, z17.h, z16.h\n"
307       "whilelt p0.h, XZR, x19\n"
308       "ld1h { z18.h }, p0/Z, [x11, #1, MUL VL]\n"
309       "addvl x11, x11, #2\n"
310       "ld1h { z16.h }, p0/Z, [x9, #1, MUL VL]\n"
311       "zip1 z17.h, z18.h, z16.h\n"
312       "st1h { z20.h }, p2, [x10]\n"
313       "addvl x9, x9, #2\n"
314       "zip2 z16.h, z18.h, z16.h\n"
315       "st1h { z19.h }, p2, [x10, #1, MUL VL]\n"
316       "cmp x20, #0x0\n"
317       "st1h { z17.h }, p2, [x10, #2, MUL VL]\n"
318       "st1h { z16.h }, p2, [x10, #3, MUL VL]\n"
319       "add x10, x10, %x[out_stride]\n"
320       "bgt 10b\n"
321       "11:"  // Tail row loop: Column loop skip
322       "addvl %x[out], %x[out], #4\n"
323       "cmp %x[height], #0x1\n"
324       "bge 7b\n"
325       "12:"  // Done
326 
327       : [height] "+&r" (height), [in] "+&r" (in), [out] "+&r" (out)
328       : [in_stride] "r" (in_stride), [out_stride] "r" (out_stride), [pad_row] "r" (pad_row), [width] "r" (width)
329       : "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"
330     );
331 }
332 
333 } // anonymous namespace
334 
335 template<>
Transform(bfloat16 * out,const bfloat16 * in,int stride,int x0,int xmax,int k0,int kmax)336 void Transform<4, 2, true, VLType::SVE>(
337     bfloat16 *out, const bfloat16 *in, int stride, int x0, int xmax, int k0, int kmax)
338 {
339     sve_transpose_interleave_4VL_2x2(
340         reinterpret_cast<uint16_t *>(out),
341         reinterpret_cast<const uint16_t *>(in + k0 * stride + x0),
342         (xmax-x0) * sizeof(bfloat16) / 2,
343         stride * sizeof(bfloat16),
344         (kmax-k0)
345     );
346 }
347 
348 #endif
349