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 __aarch64__
28
29 namespace {
30
a64_transpose_interleave_128(uint32_t * out,const uint32_t * in,size_t width,size_t in_stride,size_t height)31 void a64_transpose_interleave_128(uint32_t *out, const uint32_t *in, size_t width, size_t in_stride, size_t height)
32 {
33 size_t out_stride = 32 * height * sizeof(uint32_t);
34
35 __asm__ __volatile__(
36 "cmp %x[height], #0x4\n"
37 "blt 10f\n"
38 "1:" // Main row loop: Head
39 "mov x24, %x[in]\n"
40 "mov x23, %x[out]\n"
41 "add x22, x24, %x[in_stride]\n"
42 "add x21, x22, %x[in_stride]\n"
43 "add x20, x21, %x[in_stride]\n"
44 "add %x[in], x20, %x[in_stride]\n"
45 "sub %x[height], %x[height], #0x4\n"
46 "mov x19, %x[width]\n"
47 "cmp x19, #0x20\n"
48 "blt 3f\n"
49 "2:" // Main row loop: Column loop
50 "ldr q15, [x24], #0x10\n"
51 "sub x19, x19, #0x20\n"
52 "ldr q14, [x22], #0x10\n"
53 "cmp x19, #0x20\n"
54 "ldr q13, [x21], #0x10\n"
55 "ldr q12, [x20], #0x10\n"
56 "ldr q11, [x24], #0x10\n"
57 "ldr q10, [x22], #0x10\n"
58 "ldr q9, [x21], #0x10\n"
59 "ldr q8, [x20], #0x10\n"
60 "ldr q7, [x24], #0x10\n"
61 "ldr q6, [x22], #0x10\n"
62 "ldr q5, [x21], #0x10\n"
63 "ldr q4, [x20], #0x10\n"
64 "ldr q3, [x24], #0x10\n"
65 "ldr q2, [x22], #0x10\n"
66 "ldr q1, [x21], #0x10\n"
67 "ldr q0, [x20], #0x10\n"
68 "ldr q31, [x24], #0x10\n"
69 "ldr q30, [x22], #0x10\n"
70 "ldr q29, [x21], #0x10\n"
71 "ldr q28, [x20], #0x10\n"
72 "ldr q27, [x24], #0x10\n"
73 "ldr q26, [x22], #0x10\n"
74 "ldr q25, [x21], #0x10\n"
75 "ldr q24, [x20], #0x10\n"
76 "ldr q23, [x24], #0x10\n"
77 "ldr q22, [x22], #0x10\n"
78 "ldr q21, [x21], #0x10\n"
79 "ldr q20, [x20], #0x10\n"
80 "ldr q19, [x24], #0x10\n"
81 "ldr q18, [x22], #0x10\n"
82 "ldr q17, [x21], #0x10\n"
83 "ldr q16, [x20], #0x10\n"
84 "str q15, [x23, #0x0]\n"
85 "str q11, [x23, #0x10]\n"
86 "str q7, [x23, #0x20]\n"
87 "str q3, [x23, #0x30]\n"
88 "str q31, [x23, #0x40]\n"
89 "str q27, [x23, #0x50]\n"
90 "str q23, [x23, #0x60]\n"
91 "str q19, [x23, #0x70]\n"
92 "str q14, [x23, #0x80]\n"
93 "str q10, [x23, #0x90]\n"
94 "str q6, [x23, #0xa0]\n"
95 "str q2, [x23, #0xb0]\n"
96 "str q30, [x23, #0xc0]\n"
97 "str q26, [x23, #0xd0]\n"
98 "str q22, [x23, #0xe0]\n"
99 "str q18, [x23, #0xf0]\n"
100 "str q13, [x23, #0x100]\n"
101 "str q9, [x23, #0x110]\n"
102 "str q5, [x23, #0x120]\n"
103 "str q1, [x23, #0x130]\n"
104 "str q29, [x23, #0x140]\n"
105 "str q25, [x23, #0x150]\n"
106 "str q21, [x23, #0x160]\n"
107 "str q17, [x23, #0x170]\n"
108 "str q12, [x23, #0x180]\n"
109 "str q8, [x23, #0x190]\n"
110 "str q4, [x23, #0x1a0]\n"
111 "str q0, [x23, #0x1b0]\n"
112 "str q28, [x23, #0x1c0]\n"
113 "str q24, [x23, #0x1d0]\n"
114 "str q20, [x23, #0x1e0]\n"
115 "str q16, [x23, #0x1f0]\n"
116 "add x23, x23, %x[out_stride]\n"
117 "bge 2b\n"
118 "3:" // Main row loop: Column loop skip
119 "cmp x19, #0x10\n"
120 "blt 5f\n"
121 "4:" // Main row loop: width 16 loop: loop
122 "ldr q31, [x24], #0x10\n"
123 "sub x19, x19, #0x10\n"
124 "ldr q30, [x22], #0x10\n"
125 "cmp x19, #0x10\n"
126 "ldr q29, [x21], #0x10\n"
127 "ldr q28, [x20], #0x10\n"
128 "ldr q27, [x24], #0x10\n"
129 "ldr q26, [x22], #0x10\n"
130 "ldr q25, [x21], #0x10\n"
131 "ldr q24, [x20], #0x10\n"
132 "ldr q23, [x24], #0x10\n"
133 "ldr q22, [x22], #0x10\n"
134 "ldr q21, [x21], #0x10\n"
135 "ldr q20, [x20], #0x10\n"
136 "ldr q19, [x24], #0x10\n"
137 "ldr q18, [x22], #0x10\n"
138 "ldr q17, [x21], #0x10\n"
139 "ldr q16, [x20], #0x10\n"
140 "str q31, [x23, #0x0]\n"
141 "str q27, [x23, #0x10]\n"
142 "str q23, [x23, #0x20]\n"
143 "str q19, [x23, #0x30]\n"
144 "str q30, [x23, #0x80]\n"
145 "str q26, [x23, #0x90]\n"
146 "str q22, [x23, #0xa0]\n"
147 "str q18, [x23, #0xb0]\n"
148 "str q29, [x23, #0x100]\n"
149 "str q25, [x23, #0x110]\n"
150 "str q21, [x23, #0x120]\n"
151 "str q17, [x23, #0x130]\n"
152 "str q28, [x23, #0x180]\n"
153 "str q24, [x23, #0x190]\n"
154 "str q20, [x23, #0x1a0]\n"
155 "str q16, [x23, #0x1b0]\n"
156 "add x23, x23, #0x40\n"
157 "bge 4b\n"
158 "5:" // Main row loop: width 16 loop: skip
159 "cmp x19, #0x4\n"
160 "blt 7f\n"
161 "6:" // Main row loop: width 4 loop: loop
162 "ldr q19, [x24], #0x10\n"
163 "sub x19, x19, #0x4\n"
164 "ldr q18, [x22], #0x10\n"
165 "cmp x19, #0x4\n"
166 "ldr q17, [x21], #0x10\n"
167 "ldr q16, [x20], #0x10\n"
168 "str q19, [x23, #0x0]\n"
169 "str q18, [x23, #0x80]\n"
170 "str q17, [x23, #0x100]\n"
171 "str q16, [x23, #0x180]\n"
172 "add x23, x23, #0x10\n"
173 "bge 6b\n"
174 "7:" // Main row loop: width 4 loop: skip
175 "cmp x19, #0x1\n"
176 "blt 9f\n"
177 "8:" // Main row loop: width 1 loop: loop
178 "ldr s19, [x24], #0x4\n"
179 "sub x19, x19, #0x1\n"
180 "ldr s18, [x22], #0x4\n"
181 "cmp x19, #0x1\n"
182 "ldr s17, [x21], #0x4\n"
183 "ldr s16, [x20], #0x4\n"
184 "str s19, [x23, #0x0]\n"
185 "str s18, [x23, #0x80]\n"
186 "str s17, [x23, #0x100]\n"
187 "str s16, [x23, #0x180]\n"
188 "add x23, x23, #0x4\n"
189 "bge 8b\n"
190 "9:" // Main row loop: width 1 loop: skip
191 "add %x[out], %x[out], #0x200\n"
192 "cmp %x[height], #0x4\n"
193 "bge 1b\n"
194 "cbz %x[height], 20f\n"
195 "10:" // Main loop skip
196
197 "11:" // Tail row loop: Head
198 "mov x24, %x[in]\n"
199 "mov x23, %x[out]\n"
200 "add %x[in], x24, %x[in_stride]\n"
201 "sub %x[height], %x[height], #0x1\n"
202 "mov x19, %x[width]\n"
203 "cmp x19, #0x20\n"
204 "blt 13f\n"
205 "12:" // Tail row loop: Column loop
206 "ldr q23, [x24], #0x10\n"
207 "sub x19, x19, #0x20\n"
208 "cmp x19, #0x20\n"
209 "ldr q22, [x24], #0x10\n"
210 "ldr q21, [x24], #0x10\n"
211 "ldr q20, [x24], #0x10\n"
212 "ldr q19, [x24], #0x10\n"
213 "ldr q18, [x24], #0x10\n"
214 "ldr q17, [x24], #0x10\n"
215 "ldr q16, [x24], #0x10\n"
216 "str q23, [x23, #0x0]\n"
217 "str q22, [x23, #0x10]\n"
218 "str q21, [x23, #0x20]\n"
219 "str q20, [x23, #0x30]\n"
220 "str q19, [x23, #0x40]\n"
221 "str q18, [x23, #0x50]\n"
222 "str q17, [x23, #0x60]\n"
223 "str q16, [x23, #0x70]\n"
224 "add x23, x23, %x[out_stride]\n"
225 "bge 12b\n"
226 "13:" // Tail row loop: Column loop skip
227 "cmp x19, #0x10\n"
228 "blt 15f\n"
229 "14:" // Tail row loop: width 16 loop: loop
230 "ldr q19, [x24], #0x10\n"
231 "sub x19, x19, #0x10\n"
232 "cmp x19, #0x10\n"
233 "ldr q18, [x24], #0x10\n"
234 "ldr q17, [x24], #0x10\n"
235 "ldr q16, [x24], #0x10\n"
236 "str q19, [x23, #0x0]\n"
237 "str q18, [x23, #0x10]\n"
238 "str q17, [x23, #0x20]\n"
239 "str q16, [x23, #0x30]\n"
240 "add x23, x23, #0x40\n"
241 "bge 14b\n"
242 "15:" // Tail row loop: width 16 loop: skip
243 "cmp x19, #0x4\n"
244 "blt 17f\n"
245 "16:" // Tail row loop: width 4 loop: loop
246 "ldr q16, [x24], #0x10\n"
247 "sub x19, x19, #0x4\n"
248 "cmp x19, #0x4\n"
249 "str q16, [x23, #0x0]\n"
250 "add x23, x23, #0x10\n"
251 "bge 16b\n"
252 "17:" // Tail row loop: width 4 loop: skip
253 "cmp x19, #0x1\n"
254 "blt 19f\n"
255 "18:" // Tail row loop: width 1 loop: loop
256 "ldr s16, [x24], #0x4\n"
257 "sub x19, x19, #0x1\n"
258 "cmp x19, #0x1\n"
259 "str s16, [x23, #0x0]\n"
260 "add x23, x23, #0x4\n"
261 "bge 18b\n"
262 "19:" // Tail row loop: width 1 loop: skip
263 "add %x[out], %x[out], #0x80\n"
264 "cmp %x[height], #0x1\n"
265 "bge 11b\n"
266 "20:" // Done
267
268 : [height] "+&r" (height), [in] "+&r" (in), [out] "+&r" (out)
269 : [in_stride] "r" (in_stride), [out_stride] "r" (out_stride), [width] "r" (width)
270 : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x19", "x20", "x21", "x22", "x23", "x24"
271 );
272 }
273
274 } // anonymous namespace
275
276 template<>
Transform(float * out,const float * in,int stride,int x0,int xmax,int k0,int kmax)277 void Transform<32, 1, true, VLType::None>(
278 float *out, const float *in, int stride, int x0, int xmax, int k0, int kmax)
279 {
280 a64_transpose_interleave_128(
281 reinterpret_cast<uint32_t *>(out),
282 reinterpret_cast<const uint32_t *>(in + k0 * stride + x0),
283 (xmax-x0) * sizeof(float) / 4,
284 stride * sizeof(float),
285 (kmax-k0)
286 );
287 }
288
289 #endif
290