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_6VL_4x2(uint32_t * out,const uint32_t * in,size_t width,size_t in_stride,size_t height)32 void sve_transpose_interleave_6VL_4x2(uint32_t *out, const uint32_t *in, size_t width, size_t in_stride, size_t height)
33 {
34 uint32_t *pad_row = reinterpret_cast<uint32_t *>(alloca(width * sizeof(uint32_t)));
35
36 if (height % 2) {
37 memset(pad_row, 0, width * sizeof(uint32_t));
38 }
39
40 size_t out_stride = 6 * roundup<size_t>(height, 2) * get_vector_length<uint16_t>();
41
42 __asm__ __volatile__(
43 "ptrue p2.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 "cntw x21, ALL, MUL #6\n"
56 "cmp x22, x21\n"
57 "blt 3f\n"
58 "2:" // Main row loop: Unroll column loop
59 "ld1w { z19.s }, p2/Z, [x27]\n"
60 "mov x20, x26\n"
61 "ld1w { z18.s }, p2/Z, [x27, #1, MUL VL]\n"
62 "add x26, x26, %x[out_stride]\n"
63 "ld1w { z21.s }, p2/Z, [x27, #2, MUL VL]\n"
64 "mov x19, x26\n"
65 "ld1w { z26.s }, p2/Z, [x27, #3, MUL VL]\n"
66 "add x26, x26, %x[out_stride]\n"
67 "ld1w { z25.s }, p2/Z, [x27, #4, MUL VL]\n"
68 "sub x22, x22, x21\n"
69 "ld1w { z24.s }, p2/Z, [x27, #5, MUL VL]\n"
70 "addvl x27, x27, #6\n"
71 "ld1w { z16.s }, p2/Z, [x25]\n"
72 "zip1 z23.s, z19.s, z16.s\n"
73 "ld1w { z17.s }, p2/Z, [x25, #1, MUL VL]\n"
74 "cmp x22, x21\n"
75 "zip2 z9.s, z19.s, z16.s\n"
76 "ld1w { z20.s }, p2/Z, [x25, #2, MUL VL]\n"
77 "ld1w { z19.s }, p2/Z, [x25, #3, MUL VL]\n"
78 "zip1 z8.s, z18.s, z17.s\n"
79 "ld1w { z16.s }, p2/Z, [x25, #4, MUL VL]\n"
80 "zip2 z7.s, z18.s, z17.s\n"
81 "ld1w { z18.s }, p2/Z, [x25, #5, MUL VL]\n"
82 "addvl x25, x25, #6\n"
83 "zip1 z6.s, z21.s, z20.s\n"
84 "ld1w { z17.s }, p2/Z, [x24]\n"
85 "zip2 z5.s, z21.s, z20.s\n"
86 "ld1w { z22.s }, p2/Z, [x24, #1, MUL VL]\n"
87 "zip1 z4.s, z26.s, z19.s\n"
88 "ld1w { z21.s }, p2/Z, [x24, #2, MUL VL]\n"
89 "zip2 z3.s, z26.s, z19.s\n"
90 "ld1w { z2.s }, p2/Z, [x24, #3, MUL VL]\n"
91 "zip1 z1.s, z25.s, z16.s\n"
92 "ld1w { z0.s }, p2/Z, [x24, #4, MUL VL]\n"
93 "zip2 z31.s, z25.s, z16.s\n"
94 "ld1w { z30.s }, p2/Z, [x24, #5, MUL VL]\n"
95 "addvl x24, x24, #6\n"
96 "zip1 z29.s, z24.s, z18.s\n"
97 "ld1w { z16.s }, p2/Z, [x23]\n"
98 "zip2 z28.s, z24.s, z18.s\n"
99 "ld1w { z20.s }, p2/Z, [x23, #1, MUL VL]\n"
100 "ld1w { z19.s }, p2/Z, [x23, #2, MUL VL]\n"
101 "zip1 z27.s, z17.s, z16.s\n"
102 "ld1w { z18.s }, p2/Z, [x23, #3, MUL VL]\n"
103 "zip2 z26.s, z17.s, z16.s\n"
104 "ld1w { z17.s }, p2/Z, [x23, #4, MUL VL]\n"
105 "zip1 z25.s, z22.s, z20.s\n"
106 "ld1w { z16.s }, p2/Z, [x23, #5, MUL VL]\n"
107 "addvl x23, x23, #6\n"
108 "zip2 z24.s, z22.s, z20.s\n"
109 "st1w { z23.s }, p2, [x20]\n"
110 "zip1 z23.s, z21.s, z19.s\n"
111 "st1w { z9.s }, p2, [x20, #1, MUL VL]\n"
112 "zip2 z22.s, z21.s, z19.s\n"
113 "st1w { z8.s }, p2, [x20, #2, MUL VL]\n"
114 "zip1 z21.s, z2.s, z18.s\n"
115 "st1w { z7.s }, p2, [x20, #3, MUL VL]\n"
116 "zip2 z20.s, z2.s, z18.s\n"
117 "st1w { z6.s }, p2, [x20, #4, MUL VL]\n"
118 "zip1 z19.s, z0.s, z17.s\n"
119 "st1w { z5.s }, p2, [x20, #5, MUL VL]\n"
120 "zip2 z18.s, z0.s, z17.s\n"
121 "st1w { z27.s }, p2, [x20, #6, MUL VL]\n"
122 "zip1 z17.s, z30.s, z16.s\n"
123 "st1w { z26.s }, p2, [x20, #7, MUL VL]\n"
124 "addvl x20, x20, #12\n"
125 "zip2 z16.s, z30.s, z16.s\n"
126 "st1w { z25.s }, p2, [x20, #-4, MUL VL]\n"
127 "st1w { z24.s }, p2, [x20, #-3, MUL VL]\n"
128 "st1w { z23.s }, p2, [x20, #-2, MUL VL]\n"
129 "st1w { z22.s }, p2, [x20, #-1, MUL VL]\n"
130 "st1w { z4.s }, p2, [x19]\n"
131 "st1w { z3.s }, p2, [x19, #1, MUL VL]\n"
132 "st1w { z1.s }, p2, [x19, #2, MUL VL]\n"
133 "st1w { z31.s }, p2, [x19, #3, MUL VL]\n"
134 "st1w { z29.s }, p2, [x19, #4, MUL VL]\n"
135 "st1w { z28.s }, p2, [x19, #5, MUL VL]\n"
136 "st1w { z21.s }, p2, [x19, #6, MUL VL]\n"
137 "st1w { z20.s }, p2, [x19, #7, MUL VL]\n"
138 "addvl x19, x19, #12\n"
139 "st1w { z19.s }, p2, [x19, #-4, MUL VL]\n"
140 "st1w { z18.s }, p2, [x19, #-3, MUL VL]\n"
141 "st1w { z17.s }, p2, [x19, #-2, MUL VL]\n"
142 "st1w { z16.s }, p2, [x19, #-1, MUL VL]\n"
143 "bge 2b\n"
144 "3:" // Main row loop: Unroll column loop skip
145 "cbz x22, 5f\n"
146 "4:" // Main row loop: Column loop
147 "mov x20, x22\n"
148 "mov x19, x26\n"
149 "whilelt p0.s, XZR, x20\n"
150 "ld1w { z18.s }, p0/Z, [x27]\n"
151 "ld1w { z16.s }, p0/Z, [x25]\n"
152 "zip1 z28.s, z18.s, z16.s\n"
153 "ld1w { z17.s }, p0/Z, [x24]\n"
154 "decw x20\n"
155 "zip2 z27.s, z18.s, z16.s\n"
156 "ld1w { z16.s }, p0/Z, [x23]\n"
157 "whilelt p1.s, XZR, x20\n"
158 "zip1 z26.s, z17.s, z16.s\n"
159 "ld1w { z18.s }, p1/Z, [x27, #1, MUL VL]\n"
160 "decw x20\n"
161 "zip2 z25.s, z17.s, z16.s\n"
162 "ld1w { z16.s }, p1/Z, [x25, #1, MUL VL]\n"
163 "whilelt p0.s, XZR, x20\n"
164 "zip1 z24.s, z18.s, z16.s\n"
165 "ld1w { z17.s }, p0/Z, [x27, #2, MUL VL]\n"
166 "addvl x27, x27, #3\n"
167 "zip2 z23.s, z18.s, z16.s\n"
168 "ld1w { z16.s }, p0/Z, [x25, #2, MUL VL]\n"
169 "addvl x25, x25, #3\n"
170 "zip1 z22.s, z17.s, z16.s\n"
171 "ld1w { z18.s }, p1/Z, [x24, #1, MUL VL]\n"
172 "add x26, x26, %x[out_stride]\n"
173 "zip2 z21.s, z17.s, z16.s\n"
174 "ld1w { z20.s }, p0/Z, [x24, #2, MUL VL]\n"
175 "addvl x24, x24, #3\n"
176 "ld1w { z17.s }, p1/Z, [x23, #1, MUL VL]\n"
177 "zip1 z19.s, z18.s, z17.s\n"
178 "ld1w { z16.s }, p0/Z, [x23, #2, MUL VL]\n"
179 "addvl x23, x23, #3\n"
180 "zip2 z18.s, z18.s, z17.s\n"
181 "st1w { z28.s }, p2, [x19]\n"
182 "decd x22, ALL, MUL #6\n"
183 "zip1 z17.s, z20.s, z16.s\n"
184 "st1w { z27.s }, p2, [x19, #1, MUL VL]\n"
185 "cmp x22, #0x0\n"
186 "zip2 z16.s, z20.s, z16.s\n"
187 "st1w { z24.s }, p2, [x19, #2, MUL VL]\n"
188 "st1w { z23.s }, p2, [x19, #3, MUL VL]\n"
189 "st1w { z22.s }, p2, [x19, #4, MUL VL]\n"
190 "st1w { z21.s }, p2, [x19, #5, MUL VL]\n"
191 "st1w { z26.s }, p2, [x19, #6, MUL VL]\n"
192 "st1w { z25.s }, p2, [x19, #7, MUL VL]\n"
193 "addvl x19, x19, #12\n"
194 "st1w { z19.s }, p2, [x19, #-4, MUL VL]\n"
195 "st1w { z18.s }, p2, [x19, #-3, MUL VL]\n"
196 "st1w { z17.s }, p2, [x19, #-2, MUL VL]\n"
197 "st1w { z16.s }, p2, [x19, #-1, MUL VL]\n"
198 "bgt 4b\n"
199 "5:" // Main row loop: Column loop skip
200 "addvl %x[out], %x[out], #12\n"
201 "cmp %x[height], #0x4\n"
202 "bge 1b\n"
203 "cbz %x[height], 12f\n"
204 "6:" // Main loop skip
205
206 "7:" // Tail row loop: Head
207 "mov x27, %x[in]\n"
208 "mov x26, %x[out]\n"
209 "add x25, x27, %x[in_stride]\n"
210 "add %x[in], x25, %x[in_stride]\n"
211 "cmp %x[height], #0x1\n"
212 "csel x25, x25, %x[pad_row], GT\n"
213 "sub %x[height], %x[height], #0x2\n"
214 "mov x20, %x[width]\n"
215 "cntw x19, ALL, MUL #6\n"
216 "cmp x20, x19\n"
217 "blt 9f\n"
218 "8:" // Tail row loop: Unroll column loop
219 "ld1w { z19.s }, p2/Z, [x27]\n"
220 "sub x20, x20, x19\n"
221 "ld1w { z18.s }, p2/Z, [x27, #1, MUL VL]\n"
222 "cmp x20, x19\n"
223 "ld1w { z29.s }, p2/Z, [x27, #2, MUL VL]\n"
224 "ld1w { z28.s }, p2/Z, [x27, #3, MUL VL]\n"
225 "ld1w { z27.s }, p2/Z, [x27, #4, MUL VL]\n"
226 "ld1w { z26.s }, p2/Z, [x27, #5, MUL VL]\n"
227 "addvl x27, x27, #6\n"
228 "ld1w { z16.s }, p2/Z, [x25]\n"
229 "zip1 z25.s, z19.s, z16.s\n"
230 "ld1w { z17.s }, p2/Z, [x25, #1, MUL VL]\n"
231 "zip2 z24.s, z19.s, z16.s\n"
232 "ld1w { z16.s }, p2/Z, [x25, #2, MUL VL]\n"
233 "ld1w { z23.s }, p2/Z, [x25, #3, MUL VL]\n"
234 "zip1 z20.s, z18.s, z17.s\n"
235 "ld1w { z22.s }, p2/Z, [x25, #4, MUL VL]\n"
236 "zip2 z19.s, z18.s, z17.s\n"
237 "ld1w { z21.s }, p2/Z, [x25, #5, MUL VL]\n"
238 "addvl x25, x25, #6\n"
239 "zip1 z18.s, z29.s, z16.s\n"
240 "st1w { z25.s }, p2, [x26]\n"
241 "zip2 z17.s, z29.s, z16.s\n"
242 "st1w { z24.s }, p2, [x26, #1, MUL VL]\n"
243 "zip1 z16.s, z28.s, z23.s\n"
244 "st1w { z20.s }, p2, [x26, #2, MUL VL]\n"
245 "zip2 z20.s, z28.s, z23.s\n"
246 "st1w { z19.s }, p2, [x26, #3, MUL VL]\n"
247 "zip1 z19.s, z27.s, z22.s\n"
248 "st1w { z18.s }, p2, [x26, #4, MUL VL]\n"
249 "zip2 z18.s, z27.s, z22.s\n"
250 "st1w { z17.s }, p2, [x26, #5, MUL VL]\n"
251 "add x26, x26, %x[out_stride]\n"
252 "zip1 z17.s, z26.s, z21.s\n"
253 "st1w { z16.s }, p2, [x26]\n"
254 "zip2 z16.s, z26.s, z21.s\n"
255 "st1w { z20.s }, p2, [x26, #1, MUL VL]\n"
256 "st1w { z19.s }, p2, [x26, #2, MUL VL]\n"
257 "st1w { z18.s }, p2, [x26, #3, MUL VL]\n"
258 "st1w { z17.s }, p2, [x26, #4, MUL VL]\n"
259 "st1w { z16.s }, p2, [x26, #5, MUL VL]\n"
260 "add x26, x26, %x[out_stride]\n"
261 "bge 8b\n"
262 "9:" // Tail row loop: Unroll column loop skip
263 "cbz x20, 11f\n"
264 "10:" // Tail row loop: Column loop
265 "mov x19, x20\n"
266 "decd x20, ALL, MUL #6\n"
267 "whilelt p0.s, XZR, x19\n"
268 "ld1w { z17.s }, p0/Z, [x27]\n"
269 "ld1w { z16.s }, p0/Z, [x25]\n"
270 "zip1 z22.s, z17.s, z16.s\n"
271 "decw x19\n"
272 "zip2 z21.s, z17.s, z16.s\n"
273 "whilelt p0.s, XZR, x19\n"
274 "ld1w { z17.s }, p0/Z, [x27, #1, MUL VL]\n"
275 "decw x19\n"
276 "ld1w { z16.s }, p0/Z, [x25, #1, MUL VL]\n"
277 "zip1 z20.s, z17.s, z16.s\n"
278 "whilelt p0.s, XZR, x19\n"
279 "ld1w { z19.s }, p0/Z, [x27, #2, MUL VL]\n"
280 "zip2 z18.s, z17.s, z16.s\n"
281 "addvl x27, x27, #3\n"
282 "ld1w { z16.s }, p0/Z, [x25, #2, MUL VL]\n"
283 "zip1 z17.s, z19.s, z16.s\n"
284 "st1w { z22.s }, p2, [x26]\n"
285 "addvl x25, x25, #3\n"
286 "zip2 z16.s, z19.s, z16.s\n"
287 "st1w { z21.s }, p2, [x26, #1, MUL VL]\n"
288 "cmp x20, #0x0\n"
289 "st1w { z20.s }, p2, [x26, #2, MUL VL]\n"
290 "st1w { z18.s }, p2, [x26, #3, MUL VL]\n"
291 "st1w { z17.s }, p2, [x26, #4, MUL VL]\n"
292 "st1w { z16.s }, p2, [x26, #5, MUL VL]\n"
293 "add x26, x26, %x[out_stride]\n"
294 "bgt 10b\n"
295 "11:" // Tail row loop: Column loop skip
296 "addvl %x[out], %x[out], #6\n"
297 "cmp %x[height], #0x1\n"
298 "bge 7b\n"
299 "12:" // Done
300
301 : [height] "+&r" (height), [in] "+&r" (in), [out] "+&r" (out)
302 : [in_stride] "r" (in_stride), [out_stride] "r" (out_stride), [pad_row] "r" (pad_row), [width] "r" (width)
303 : "cc", "memory", "p0", "p1", "p2", "x19", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "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"
304 );
305 }
306
307 } // anonymous namespace
308
309 template<>
Transform(float * out,const float * in,int stride,int x0,int xmax,int k0,int kmax)310 void Transform<6, 2, true, VLType::SVE>(
311 float *out, const float *in, int stride, int x0, int xmax, int k0, int kmax)
312 {
313 sve_transpose_interleave_6VL_4x2(
314 reinterpret_cast<uint32_t *>(out),
315 reinterpret_cast<const uint32_t *>(in + k0 * stride + x0),
316 (xmax-x0) * sizeof(float) / 4,
317 stride * sizeof(float),
318 (kmax-k0)
319 );
320 }
321
322 #endif
323