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_24(uint16_t * out,const uint16_t * in,size_t width,size_t in_stride,size_t height)31 void a64_transpose_interleave_24(uint16_t *out, const uint16_t *in, size_t width, size_t in_stride, size_t height)
32 {
33 size_t out_stride = 12 * height * sizeof(uint16_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, #0x18\n"
48 "blt 3f\n"
49 "2:" // Main row loop: Unroll column loop
50 "ldr q29, [x24], #0x10\n"
51 "sub x19, x19, #0x18\n"
52 "ldr q18, [x22], #0x10\n"
53 "dup v28.2d, v18.d[1]\n"
54 "ldr q16, [x24], #0x10\n"
55 "cmp x19, #0x18\n"
56 "dup v27.2d, v16.d[0]\n"
57 "ldr q17, [x24], #0x10\n"
58 "dup v26.2d, v16.d[1]\n"
59 "ldr q16, [x22], #0x10\n"
60 "mov v27.d[1], v18.d[0]\n"
61 "ldr q25, [x21], #0x10\n"
62 "ldr q24, [x20], #0x10\n"
63 "mov v26.d[1], v17.d[0]\n"
64 "ldr q23, [x22], #0x10\n"
65 "mov v28.d[1], v16.d[0]\n"
66 "dup v22.2d, v17.d[1]\n"
67 "ldr q17, [x21], #0x10\n"
68 "dup v21.2d, v24.d[1]\n"
69 "ldr q20, [x20], #0x10\n"
70 "mov v22.d[1], v16.d[1]\n"
71 "ldr q16, [x21], #0x10\n"
72 "dup v19.2d, v17.d[0]\n"
73 "dup v18.2d, v17.d[1]\n"
74 "ldr q17, [x20], #0x10\n"
75 "mov v19.d[1], v24.d[0]\n"
76 "str q29, [x23, #0x0]\n"
77 "mov v21.d[1], v20.d[0]\n"
78 "str q27, [x23, #0x10]\n"
79 "str q28, [x23, #0x20]\n"
80 "mov v18.d[1], v16.d[0]\n"
81 "dup v16.2d, v16.d[1]\n"
82 "str q25, [x23, #0x30]\n"
83 "mov v16.d[1], v20.d[1]\n"
84 "str q19, [x23, #0x40]\n"
85 "str q21, [x23, #0x50]\n"
86 "add x23, x23, %x[out_stride]\n"
87 "str q26, [x23, #0x0]\n"
88 "str q22, [x23, #0x10]\n"
89 "str q23, [x23, #0x20]\n"
90 "str q18, [x23, #0x30]\n"
91 "str q16, [x23, #0x40]\n"
92 "str q17, [x23, #0x50]\n"
93 "add x23, x23, %x[out_stride]\n"
94 "bge 2b\n"
95 "3:" // Main row loop: Unroll column loop skip
96 "cmp x19, #0xc\n"
97 "blt 5f\n"
98 "4:" // Main row loop: Column loop
99 "ldr q22, [x24], #0x10\n"
100 "sub x19, x19, #0xc\n"
101 "ldr q16, [x22], #0x10\n"
102 "dup v21.2d, v16.d[1]\n"
103 "ldr d20, [x24], #0x8\n"
104 "cmp x19, #0xc\n"
105 "mov v20.d[1], v16.d[0]\n"
106 "ldr d16, [x22], #0x8\n"
107 "ldr q19, [x21], #0x10\n"
108 "mov v21.d[1], v16.d[0]\n"
109 "ldr d18, [x21], #0x8\n"
110 "ldr q16, [x20], #0x10\n"
111 "mov v18.d[1], v16.d[0]\n"
112 "ldr d17, [x20], #0x8\n"
113 "dup v16.2d, v16.d[1]\n"
114 "str q22, [x23, #0x0]\n"
115 "str q20, [x23, #0x10]\n"
116 "mov v16.d[1], v17.d[0]\n"
117 "str q21, [x23, #0x20]\n"
118 "str q19, [x23, #0x30]\n"
119 "str q18, [x23, #0x40]\n"
120 "str q16, [x23, #0x50]\n"
121 "add x23, x23, %x[out_stride]\n"
122 "bge 4b\n"
123 "5:" // Main row loop: Column loop skip
124 "cmp x19, #0x4\n"
125 "blt 7f\n"
126 "6:" // Main row loop: width 4 loop: loop
127 "ldr d19, [x24], #0x8\n"
128 "sub x19, x19, #0x4\n"
129 "ldr d18, [x22], #0x8\n"
130 "cmp x19, #0x4\n"
131 "ldr d17, [x21], #0x8\n"
132 "ldr d16, [x20], #0x8\n"
133 "str d19, [x23, #0x0]\n"
134 "str d18, [x23, #0x18]\n"
135 "str d17, [x23, #0x30]\n"
136 "str d16, [x23, #0x48]\n"
137 "add x23, x23, #0x8\n"
138 "bge 6b\n"
139 "7:" // Main row loop: width 4 loop: skip
140 "cmp x19, #0x1\n"
141 "blt 9f\n"
142 "8:" // Main row loop: width 1 loop: loop
143 "ldr h19, [x24], #0x2\n"
144 "sub x19, x19, #0x1\n"
145 "ldr h18, [x22], #0x2\n"
146 "cmp x19, #0x1\n"
147 "ldr h17, [x21], #0x2\n"
148 "ldr h16, [x20], #0x2\n"
149 "str h19, [x23, #0x0]\n"
150 "str h18, [x23, #0x18]\n"
151 "str h17, [x23, #0x30]\n"
152 "str h16, [x23, #0x48]\n"
153 "add x23, x23, #0x2\n"
154 "bge 8b\n"
155 "9:" // Main row loop: width 1 loop: skip
156 "add %x[out], %x[out], #0x60\n"
157 "cmp %x[height], #0x4\n"
158 "bge 1b\n"
159 "cbz %x[height], 20f\n"
160 "10:" // Main loop skip
161
162 "11:" // Tail row loop: Head
163 "mov x24, %x[in]\n"
164 "mov x23, %x[out]\n"
165 "add %x[in], x24, %x[in_stride]\n"
166 "sub %x[height], %x[height], #0x1\n"
167 "mov x19, %x[width]\n"
168 "cmp x19, #0x18\n"
169 "blt 13f\n"
170 "12:" // Tail row loop: Unroll column loop
171 "ldr q19, [x24], #0x10\n"
172 "sub x19, x19, #0x18\n"
173 "cmp x19, #0x18\n"
174 "ldr q16, [x24], #0x10\n"
175 "dup v18.2d, v16.d[0]\n"
176 "ldr q17, [x24], #0x10\n"
177 "dup v16.2d, v16.d[1]\n"
178 "str q19, [x23, #0x0]\n"
179 "str d18, [x23, #0x10]\n"
180 "mov v16.d[1], v17.d[0]\n"
181 "add x23, x23, %x[out_stride]\n"
182 "str q16, [x23, #0x0]\n"
183 "dup v16.2d, v17.d[1]\n"
184 "str d16, [x23, #0x10]\n"
185 "add x23, x23, %x[out_stride]\n"
186 "bge 12b\n"
187 "13:" // Tail row loop: Unroll column loop skip
188 "cmp x19, #0xc\n"
189 "blt 15f\n"
190 "14:" // Tail row loop: Column loop
191 "ldr q17, [x24], #0x10\n"
192 "sub x19, x19, #0xc\n"
193 "cmp x19, #0xc\n"
194 "ldr d16, [x24], #0x8\n"
195 "str q17, [x23, #0x0]\n"
196 "str d16, [x23, #0x10]\n"
197 "add x23, x23, %x[out_stride]\n"
198 "bge 14b\n"
199 "15:" // Tail row loop: Column loop skip
200 "cmp x19, #0x4\n"
201 "blt 17f\n"
202 "16:" // Tail row loop: width 4 loop: loop
203 "ldr d16, [x24], #0x8\n"
204 "sub x19, x19, #0x4\n"
205 "cmp x19, #0x4\n"
206 "str d16, [x23, #0x0]\n"
207 "add x23, x23, #0x8\n"
208 "bge 16b\n"
209 "17:" // Tail row loop: width 4 loop: skip
210 "cmp x19, #0x1\n"
211 "blt 19f\n"
212 "18:" // Tail row loop: width 1 loop: loop
213 "ldr h16, [x24], #0x2\n"
214 "sub x19, x19, #0x1\n"
215 "cmp x19, #0x1\n"
216 "str h16, [x23, #0x0]\n"
217 "add x23, x23, #0x2\n"
218 "bge 18b\n"
219 "19:" // Tail row loop: width 1 loop: skip
220 "add %x[out], %x[out], #0x18\n"
221 "cmp %x[height], #0x1\n"
222 "bge 11b\n"
223 "20:" // Done
224
225 : [height] "+&r" (height), [in] "+&r" (in), [out] "+&r" (out)
226 : [in_stride] "r" (in_stride), [out_stride] "r" (out_stride), [width] "r" (width)
227 : "cc", "memory", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "x19", "x20", "x21", "x22", "x23", "x24"
228 );
229 }
230
231 } // anonymous namespace
232
233 template<>
Transform(float * out,const float * in,int stride,int x0,int xmax,int k0,int kmax)234 void Transform<6, 1, true, VLType::None>(
235 float *out, const float *in, int stride, int x0, int xmax, int k0, int kmax)
236 {
237 a64_transpose_interleave_24(
238 reinterpret_cast<uint16_t *>(out),
239 reinterpret_cast<const uint16_t *>(in + k0 * stride + x0),
240 (xmax-x0) * sizeof(float) / 2,
241 stride * sizeof(float),
242 (kmax-k0)
243 );
244 }
245
246 template<>
Transform(int16_t * out,const int16_t * in,int stride,int x0,int xmax,int k0,int kmax)247 void Transform<12, 1, true, VLType::None>(
248 int16_t *out, const int16_t *in, int stride, int x0, int xmax, int k0, int kmax)
249 {
250 a64_transpose_interleave_24(
251 reinterpret_cast<uint16_t *>(out),
252 reinterpret_cast<const uint16_t *>(in + k0 * stride + x0),
253 (xmax-x0) * sizeof(int16_t) / 2,
254 stride * sizeof(int16_t),
255 (kmax-k0)
256 );
257 }
258
259 template<>
Transform(uint16_t * out,const uint16_t * in,int stride,int x0,int xmax,int k0,int kmax)260 void Transform<12, 1, true, VLType::None>(
261 uint16_t *out, const uint16_t *in, int stride, int x0, int xmax, int k0, int kmax)
262 {
263 a64_transpose_interleave_24(
264 reinterpret_cast<uint16_t *>(out),
265 reinterpret_cast<const uint16_t *>(in + k0 * stride + x0),
266 (xmax-x0) * sizeof(uint16_t) / 2,
267 stride * sizeof(uint16_t),
268 (kmax-k0)
269 );
270 }
271
272 #endif
273