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_3VL_2x2(uint16_t * out,const uint16_t * in,size_t width,size_t in_stride,size_t height)32 void sve_transpose_interleave_3VL_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 = 3 * 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 #3\n"
60 "cmp x22, x21\n"
61 "blt 3f\n"
62 "2:" // Main row loop: Unroll column loop
63 "ld1h { z19.h }, p2/Z, [x11]\n"
64 "mov x20, x10\n"
65 "ld1h { z18.h }, p2/Z, [x11, #1, MUL VL]\n"
66 "add x10, x10, %x[out_stride]\n"
67 "ld1h { z21.h }, p2/Z, [x11, #2, MUL VL]\n"
68 "addvl x11, x11, #3\n"
69 "ld1h { z16.h }, p2/Z, [x9]\n"
70 "zip1 z9.h, z19.h, z16.h\n"
71 "ld1h { z17.h }, p2/Z, [x9, #1, MUL VL]\n"
72 "mov x19, x10\n"
73 "zip2 z8.h, z19.h, z16.h\n"
74 "ld1h { z16.h }, p2/Z, [x9, #2, MUL VL]\n"
75 "addvl x9, x9, #3\n"
76 "zip1 z7.h, z18.h, z17.h\n"
77 "ld1h { z19.h }, p2/Z, [x28]\n"
78 "add x10, x10, %x[out_stride]\n"
79 "zip2 z6.h, z18.h, z17.h\n"
80 "ld1h { z20.h }, p2/Z, [x28, #1, MUL VL]\n"
81 "sub x22, x22, x21\n"
82 "zip1 z5.h, z21.h, z16.h\n"
83 "ld1h { z18.h }, p2/Z, [x28, #2, MUL VL]\n"
84 "addvl x28, x28, #3\n"
85 "zip2 z4.h, z21.h, z16.h\n"
86 "ld1h { z16.h }, p2/Z, [x27]\n"
87 "cmp x22, x21\n"
88 "zip1 z3.h, z19.h, z16.h\n"
89 "ld1h { z17.h }, p2/Z, [x27, #1, MUL VL]\n"
90 "zip2 z2.h, z19.h, z16.h\n"
91 "ld1h { z16.h }, p2/Z, [x27, #2, MUL VL]\n"
92 "addvl x27, x27, #3\n"
93 "zip1 z1.h, z20.h, z17.h\n"
94 "ld1h { z19.h }, p2/Z, [x26]\n"
95 "zip2 z0.h, z20.h, z17.h\n"
96 "ld1h { z21.h }, p2/Z, [x26, #1, MUL VL]\n"
97 "zip1 z31.h, z18.h, z16.h\n"
98 "ld1h { z20.h }, p2/Z, [x26, #2, MUL VL]\n"
99 "addvl x26, x26, #3\n"
100 "zip2 z30.h, z18.h, z16.h\n"
101 "ld1h { z18.h }, p2/Z, [x25]\n"
102 "ld1h { z17.h }, p2/Z, [x25, #1, MUL VL]\n"
103 "zip1 z29.h, z19.h, z18.h\n"
104 "ld1h { z16.h }, p2/Z, [x25, #2, MUL VL]\n"
105 "addvl x25, x25, #3\n"
106 "zip2 z28.h, z19.h, z18.h\n"
107 "ld1h { z19.h }, p2/Z, [x24]\n"
108 "zip1 z27.h, z21.h, z17.h\n"
109 "ld1h { z26.h }, p2/Z, [x24, #1, MUL VL]\n"
110 "zip2 z25.h, z21.h, z17.h\n"
111 "ld1h { z24.h }, p2/Z, [x24, #2, MUL VL]\n"
112 "addvl x24, x24, #3\n"
113 "zip1 z23.h, z20.h, z16.h\n"
114 "ld1h { z18.h }, p2/Z, [x23]\n"
115 "zip2 z22.h, z20.h, z16.h\n"
116 "ld1h { z17.h }, p2/Z, [x23, #1, MUL VL]\n"
117 "ld1h { z16.h }, p2/Z, [x23, #2, MUL VL]\n"
118 "zip1 z21.h, z19.h, z18.h\n"
119 "st1h { z9.h }, p2, [x20]\n"
120 "addvl x23, x23, #3\n"
121 "zip2 z20.h, z19.h, z18.h\n"
122 "st1h { z8.h }, p2, [x20, #1, MUL VL]\n"
123 "zip1 z19.h, z26.h, z17.h\n"
124 "st1h { z7.h }, p2, [x20, #2, MUL VL]\n"
125 "zip2 z18.h, z26.h, z17.h\n"
126 "st1h { z3.h }, p2, [x20, #3, MUL VL]\n"
127 "zip1 z17.h, z24.h, z16.h\n"
128 "st1h { z2.h }, p2, [x20, #4, MUL VL]\n"
129 "zip2 z16.h, z24.h, z16.h\n"
130 "st1h { z1.h }, p2, [x20, #5, MUL VL]\n"
131 "st1h { z29.h }, p2, [x20, #6, MUL VL]\n"
132 "st1h { z28.h }, p2, [x20, #7, MUL VL]\n"
133 "addvl x20, x20, #12\n"
134 "st1h { z27.h }, p2, [x20, #-4, MUL VL]\n"
135 "st1h { z21.h }, p2, [x20, #-3, MUL VL]\n"
136 "st1h { z20.h }, p2, [x20, #-2, MUL VL]\n"
137 "st1h { z19.h }, p2, [x20, #-1, MUL VL]\n"
138 "st1h { z6.h }, p2, [x19]\n"
139 "st1h { z5.h }, p2, [x19, #1, MUL VL]\n"
140 "st1h { z4.h }, p2, [x19, #2, MUL VL]\n"
141 "st1h { z0.h }, p2, [x19, #3, MUL VL]\n"
142 "st1h { z31.h }, p2, [x19, #4, MUL VL]\n"
143 "st1h { z30.h }, p2, [x19, #5, MUL VL]\n"
144 "st1h { z25.h }, p2, [x19, #6, MUL VL]\n"
145 "st1h { z23.h }, p2, [x19, #7, MUL VL]\n"
146 "addvl x19, x19, #12\n"
147 "st1h { z22.h }, p2, [x19, #-4, MUL VL]\n"
148 "st1h { z18.h }, p2, [x19, #-3, MUL VL]\n"
149 "st1h { z17.h }, p2, [x19, #-2, MUL VL]\n"
150 "st1h { z16.h }, p2, [x19, #-1, MUL VL]\n"
151 "bge 2b\n"
152 "3:" // Main row loop: Unroll column loop skip
153 "cbz x22, 5f\n"
154 "4:" // Main row loop: Column loop
155 "mov x20, x22\n"
156 "mov x19, x10\n"
157 "whilelt p1.h, XZR, x20\n"
158 "ld1h { z17.h }, p1/Z, [x11]\n"
159 "ld1h { z16.h }, p1/Z, [x9]\n"
160 "zip1 z29.h, z17.h, z16.h\n"
161 "ld1h { z18.h }, p1/Z, [x28]\n"
162 "dech x20\n"
163 "zip2 z28.h, z17.h, z16.h\n"
164 "ld1h { z16.h }, p1/Z, [x27]\n"
165 "whilelt p0.h, XZR, x20\n"
166 "zip1 z27.h, z18.h, z16.h\n"
167 "ld1h { z17.h }, p0/Z, [x11, #1, MUL VL]\n"
168 "addvl x11, x11, #1\n"
169 "zip2 z26.h, z18.h, z16.h\n"
170 "ld1h { z16.h }, p0/Z, [x9, #1, MUL VL]\n"
171 "incd x11, ALL, MUL #4\n"
172 "zip1 z25.h, z17.h, z16.h\n"
173 "ld1h { z17.h }, p0/Z, [x28, #1, MUL VL]\n"
174 "addvl x9, x9, #1\n"
175 "ld1h { z16.h }, p0/Z, [x27, #1, MUL VL]\n"
176 "zip1 z24.h, z17.h, z16.h\n"
177 "ld1h { z19.h }, p1/Z, [x26]\n"
178 "incd x9, ALL, MUL #4\n"
179 "ld1h { z18.h }, p0/Z, [x26, #1, MUL VL]\n"
180 "addvl x28, x28, #1\n"
181 "ld1h { z17.h }, p1/Z, [x25]\n"
182 "zip1 z23.h, z19.h, z17.h\n"
183 "ld1h { z16.h }, p0/Z, [x25, #1, MUL VL]\n"
184 "incd x28, ALL, MUL #4\n"
185 "zip2 z22.h, z19.h, z17.h\n"
186 "ld1h { z21.h }, p1/Z, [x24]\n"
187 "addvl x27, x27, #1\n"
188 "zip1 z20.h, z18.h, z16.h\n"
189 "ld1h { z19.h }, p0/Z, [x24, #1, MUL VL]\n"
190 "incd x27, ALL, MUL #4\n"
191 "ld1h { z17.h }, p1/Z, [x23]\n"
192 "zip1 z18.h, z21.h, z17.h\n"
193 "ld1h { z16.h }, p0/Z, [x23, #1, MUL VL]\n"
194 "addvl x26, x26, #1\n"
195 "zip2 z17.h, z21.h, z17.h\n"
196 "st1h { z29.h }, p2, [x19]\n"
197 "incd x26, ALL, MUL #4\n"
198 "zip1 z16.h, z19.h, z16.h\n"
199 "st1h { z28.h }, p2, [x19, #1, MUL VL]\n"
200 "addvl x25, x25, #1\n"
201 "st1h { z25.h }, p2, [x19, #2, MUL VL]\n"
202 "incd x25, ALL, MUL #4\n"
203 "st1h { z27.h }, p2, [x19, #3, MUL VL]\n"
204 "addvl x24, x24, #1\n"
205 "st1h { z26.h }, p2, [x19, #4, MUL VL]\n"
206 "incd x24, ALL, MUL #4\n"
207 "st1h { z24.h }, p2, [x19, #5, MUL VL]\n"
208 "addvl x23, x23, #1\n"
209 "st1h { z23.h }, p2, [x19, #6, MUL VL]\n"
210 "incd x23, ALL, MUL #4\n"
211 "st1h { z22.h }, p2, [x19, #7, MUL VL]\n"
212 "addvl x19, x19, #12\n"
213 "add x10, x10, %x[out_stride]\n"
214 "st1h { z20.h }, p2, [x19, #-4, MUL VL]\n"
215 "st1h { z18.h }, p2, [x19, #-3, MUL VL]\n"
216 "decw x22, ALL, MUL #3\n"
217 "st1h { z17.h }, p2, [x19, #-2, MUL VL]\n"
218 "cmp x22, #0x0\n"
219 "st1h { z16.h }, p2, [x19, #-1, MUL VL]\n"
220 "bgt 4b\n"
221 "5:" // Main row loop: Column loop skip
222 "addvl %x[out], %x[out], #12\n"
223 "cmp %x[height], #0x8\n"
224 "bge 1b\n"
225 "cbz %x[height], 12f\n"
226 "6:" // Main loop skip
227
228 "7:" // Tail row loop: Head
229 "mov x11, %x[in]\n"
230 "mov x10, %x[out]\n"
231 "add x9, x11, %x[in_stride]\n"
232 "add %x[in], x9, %x[in_stride]\n"
233 "cmp %x[height], #0x1\n"
234 "csel x9, x9, %x[pad_row], GT\n"
235 "sub %x[height], %x[height], #0x2\n"
236 "mov x20, %x[width]\n"
237 "cnth x19, ALL, MUL #3\n"
238 "cmp x20, x19\n"
239 "blt 9f\n"
240 "8:" // Tail row loop: Unroll column loop
241 "ld1h { z17.h }, p2/Z, [x11]\n"
242 "sub x20, x20, x19\n"
243 "ld1h { z22.h }, p2/Z, [x11, #1, MUL VL]\n"
244 "cmp x20, x19\n"
245 "ld1h { z21.h }, p2/Z, [x11, #2, MUL VL]\n"
246 "addvl x11, x11, #3\n"
247 "ld1h { z16.h }, p2/Z, [x9]\n"
248 "zip1 z20.h, z17.h, z16.h\n"
249 "ld1h { z18.h }, p2/Z, [x9, #1, MUL VL]\n"
250 "zip2 z17.h, z17.h, z16.h\n"
251 "ld1h { z19.h }, p2/Z, [x9, #2, MUL VL]\n"
252 "addvl x9, x9, #3\n"
253 "zip1 z16.h, z22.h, z18.h\n"
254 "st1h { z20.h }, p2, [x10]\n"
255 "zip2 z18.h, z22.h, z18.h\n"
256 "st1h { z17.h }, p2, [x10, #1, MUL VL]\n"
257 "zip1 z17.h, z21.h, z19.h\n"
258 "st1h { z16.h }, p2, [x10, #2, MUL VL]\n"
259 "add x10, x10, %x[out_stride]\n"
260 "zip2 z16.h, z21.h, z19.h\n"
261 "st1h { z18.h }, p2, [x10]\n"
262 "st1h { z17.h }, p2, [x10, #1, MUL VL]\n"
263 "st1h { z16.h }, p2, [x10, #2, MUL VL]\n"
264 "add x10, x10, %x[out_stride]\n"
265 "bge 8b\n"
266 "9:" // Tail row loop: Unroll column loop skip
267 "cbz x20, 11f\n"
268 "10:" // Tail row loop: Column loop
269 "mov x19, x20\n"
270 "decw x20, ALL, MUL #3\n"
271 "whilelt p0.h, XZR, x19\n"
272 "ld1h { z17.h }, p0/Z, [x11]\n"
273 "ld1h { z16.h }, p0/Z, [x9]\n"
274 "zip1 z19.h, z17.h, z16.h\n"
275 "dech x19\n"
276 "zip2 z18.h, z17.h, z16.h\n"
277 "whilelt p0.h, XZR, x19\n"
278 "ld1h { z17.h }, p0/Z, [x11, #1, MUL VL]\n"
279 "addvl x11, x11, #1\n"
280 "ld1h { z16.h }, p0/Z, [x9, #1, MUL VL]\n"
281 "zip1 z16.h, z17.h, z16.h\n"
282 "st1h { z19.h }, p2, [x10]\n"
283 "incd x11, ALL, MUL #4\n"
284 "st1h { z18.h }, p2, [x10, #1, MUL VL]\n"
285 "addvl x9, x9, #1\n"
286 "st1h { z16.h }, p2, [x10, #2, MUL VL]\n"
287 "incd x9, ALL, MUL #4\n"
288 "add x10, x10, %x[out_stride]\n"
289 "cmp x20, #0x0\n"
290 "bgt 10b\n"
291 "11:" // Tail row loop: Column loop skip
292 "addvl %x[out], %x[out], #3\n"
293 "cmp %x[height], #0x1\n"
294 "bge 7b\n"
295 "12:" // Done
296
297 : [height] "+&r" (height), [in] "+&r" (in), [out] "+&r" (out)
298 : [in_stride] "r" (in_stride), [out_stride] "r" (out_stride), [pad_row] "r" (pad_row), [width] "r" (width)
299 : "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", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
300 );
301 }
302
303 } // anonymous namespace
304
305 template<>
Transform(bfloat16 * out,const bfloat16 * in,int stride,int x0,int xmax,int k0,int kmax)306 void Transform<3, 2, true, VLType::SVE>(
307 bfloat16 *out, const bfloat16 *in, int stride, int x0, int xmax, int k0, int kmax)
308 {
309 sve_transpose_interleave_3VL_2x2(
310 reinterpret_cast<uint16_t *>(out),
311 reinterpret_cast<const uint16_t *>(in + k0 * stride + x0),
312 (xmax-x0) * sizeof(bfloat16) / 2,
313 stride * sizeof(bfloat16),
314 (kmax-k0)
315 );
316 }
317
318 #endif
319