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_48(uint16_t * out,const uint16_t * in,size_t width,size_t in_stride,size_t height)31 void a64_transpose_interleave_48(uint16_t *out, const uint16_t *in, size_t width, size_t in_stride, size_t height)
32 {
33 size_t out_stride = 24 * 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: Column loop
50 "ldr q27, [x24], #0x10\n"
51 "sub x19, x19, #0x18\n"
52 "ldr q26, [x22], #0x10\n"
53 "cmp x19, #0x18\n"
54 "ldr q25, [x21], #0x10\n"
55 "ldr q24, [x20], #0x10\n"
56 "ldr q23, [x24], #0x10\n"
57 "ldr q22, [x22], #0x10\n"
58 "ldr q21, [x21], #0x10\n"
59 "ldr q20, [x20], #0x10\n"
60 "ldr q19, [x24], #0x10\n"
61 "ldr q18, [x22], #0x10\n"
62 "ldr q17, [x21], #0x10\n"
63 "ldr q16, [x20], #0x10\n"
64 "str q27, [x23, #0x0]\n"
65 "str q23, [x23, #0x10]\n"
66 "str q19, [x23, #0x20]\n"
67 "str q26, [x23, #0x30]\n"
68 "str q22, [x23, #0x40]\n"
69 "str q18, [x23, #0x50]\n"
70 "str q25, [x23, #0x60]\n"
71 "str q21, [x23, #0x70]\n"
72 "str q17, [x23, #0x80]\n"
73 "str q24, [x23, #0x90]\n"
74 "str q20, [x23, #0xa0]\n"
75 "str q16, [x23, #0xb0]\n"
76 "add x23, x23, %x[out_stride]\n"
77 "bge 2b\n"
78 "3:" // Main row loop: Column loop skip
79 "cmp x19, #0x10\n"
80 "blt 5f\n"
81 "4:" // Main row loop: width 16 loop: loop
82 "ldr q23, [x24], #0x10\n"
83 "sub x19, x19, #0x10\n"
84 "ldr q22, [x22], #0x10\n"
85 "cmp x19, #0x10\n"
86 "ldr q21, [x21], #0x10\n"
87 "ldr q20, [x20], #0x10\n"
88 "ldr q19, [x24], #0x10\n"
89 "ldr q18, [x22], #0x10\n"
90 "ldr q17, [x21], #0x10\n"
91 "ldr q16, [x20], #0x10\n"
92 "str q23, [x23, #0x0]\n"
93 "str q19, [x23, #0x10]\n"
94 "str q22, [x23, #0x30]\n"
95 "str q18, [x23, #0x40]\n"
96 "str q21, [x23, #0x60]\n"
97 "str q17, [x23, #0x70]\n"
98 "str q20, [x23, #0x90]\n"
99 "str q16, [x23, #0xa0]\n"
100 "add x23, x23, #0x20\n"
101 "bge 4b\n"
102 "5:" // Main row loop: width 16 loop: skip
103 "cmp x19, #0x4\n"
104 "blt 7f\n"
105 "6:" // Main row loop: width 4 loop: loop
106 "ldr d19, [x24], #0x8\n"
107 "sub x19, x19, #0x4\n"
108 "ldr d18, [x22], #0x8\n"
109 "cmp x19, #0x4\n"
110 "ldr d17, [x21], #0x8\n"
111 "ldr d16, [x20], #0x8\n"
112 "str d19, [x23, #0x0]\n"
113 "str d18, [x23, #0x30]\n"
114 "str d17, [x23, #0x60]\n"
115 "str d16, [x23, #0x90]\n"
116 "add x23, x23, #0x8\n"
117 "bge 6b\n"
118 "7:" // Main row loop: width 4 loop: skip
119 "cmp x19, #0x1\n"
120 "blt 9f\n"
121 "8:" // Main row loop: width 1 loop: loop
122 "ldr h19, [x24], #0x2\n"
123 "sub x19, x19, #0x1\n"
124 "ldr h18, [x22], #0x2\n"
125 "cmp x19, #0x1\n"
126 "ldr h17, [x21], #0x2\n"
127 "ldr h16, [x20], #0x2\n"
128 "str h19, [x23, #0x0]\n"
129 "str h18, [x23, #0x30]\n"
130 "str h17, [x23, #0x60]\n"
131 "str h16, [x23, #0x90]\n"
132 "add x23, x23, #0x2\n"
133 "bge 8b\n"
134 "9:" // Main row loop: width 1 loop: skip
135 "add %x[out], %x[out], #0xc0\n"
136 "cmp %x[height], #0x4\n"
137 "bge 1b\n"
138 "cbz %x[height], 20f\n"
139 "10:" // Main loop skip
140
141 "11:" // Tail row loop: Head
142 "mov x24, %x[in]\n"
143 "mov x23, %x[out]\n"
144 "add %x[in], x24, %x[in_stride]\n"
145 "sub %x[height], %x[height], #0x1\n"
146 "mov x19, %x[width]\n"
147 "cmp x19, #0x18\n"
148 "blt 13f\n"
149 "12:" // Tail row loop: Column loop
150 "ldr q18, [x24], #0x10\n"
151 "sub x19, x19, #0x18\n"
152 "cmp x19, #0x18\n"
153 "ldr q17, [x24], #0x10\n"
154 "ldr q16, [x24], #0x10\n"
155 "str q18, [x23, #0x0]\n"
156 "str q17, [x23, #0x10]\n"
157 "str q16, [x23, #0x20]\n"
158 "add x23, x23, %x[out_stride]\n"
159 "bge 12b\n"
160 "13:" // Tail row loop: Column loop skip
161 "cmp x19, #0x10\n"
162 "blt 15f\n"
163 "14:" // Tail row loop: width 16 loop: loop
164 "ldr q17, [x24], #0x10\n"
165 "sub x19, x19, #0x10\n"
166 "cmp x19, #0x10\n"
167 "ldr q16, [x24], #0x10\n"
168 "str q17, [x23, #0x0]\n"
169 "str q16, [x23, #0x10]\n"
170 "add x23, x23, #0x20\n"
171 "bge 14b\n"
172 "15:" // Tail row loop: width 16 loop: skip
173 "cmp x19, #0x4\n"
174 "blt 17f\n"
175 "16:" // Tail row loop: width 4 loop: loop
176 "ldr d16, [x24], #0x8\n"
177 "sub x19, x19, #0x4\n"
178 "cmp x19, #0x4\n"
179 "str d16, [x23, #0x0]\n"
180 "add x23, x23, #0x8\n"
181 "bge 16b\n"
182 "17:" // Tail row loop: width 4 loop: skip
183 "cmp x19, #0x1\n"
184 "blt 19f\n"
185 "18:" // Tail row loop: width 1 loop: loop
186 "ldr h16, [x24], #0x2\n"
187 "sub x19, x19, #0x1\n"
188 "cmp x19, #0x1\n"
189 "str h16, [x23, #0x0]\n"
190 "add x23, x23, #0x2\n"
191 "bge 18b\n"
192 "19:" // Tail row loop: width 1 loop: skip
193 "add %x[out], %x[out], #0x30\n"
194 "cmp %x[height], #0x1\n"
195 "bge 11b\n"
196 "20:" // Done
197
198 : [height] "+&r" (height), [in] "+&r" (in), [out] "+&r" (out)
199 : [in_stride] "r" (in_stride), [out_stride] "r" (out_stride), [width] "r" (width)
200 : "cc", "memory", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "x19", "x20", "x21", "x22", "x23", "x24"
201 );
202 }
203
204 } // anonymous namespace
205
206 template<>
Transform(float * out,const float * in,int stride,int x0,int xmax,int k0,int kmax)207 void Transform<12, 1, true, VLType::None>(
208 float *out, const float *in, int stride, int x0, int xmax, int k0, int kmax)
209 {
210 a64_transpose_interleave_48(
211 reinterpret_cast<uint16_t *>(out),
212 reinterpret_cast<const uint16_t *>(in + k0 * stride + x0),
213 (xmax-x0) * sizeof(float) / 2,
214 stride * sizeof(float),
215 (kmax-k0)
216 );
217 }
218
219 template<>
Transform(__fp16 * out,const __fp16 * in,int stride,int x0,int xmax,int k0,int kmax)220 void Transform<24, 1, true, VLType::None>(
221 __fp16 *out, const __fp16 *in, int stride, int x0, int xmax, int k0, int kmax)
222 {
223 a64_transpose_interleave_48(
224 reinterpret_cast<uint16_t *>(out),
225 reinterpret_cast<const uint16_t *>(in + k0 * stride + x0),
226 (xmax-x0) * sizeof(__fp16) / 2,
227 stride * sizeof(__fp16),
228 (kmax-k0)
229 );
230 }
231
232 template<>
Transform(double * out,const double * in,int stride,int x0,int xmax,int k0,int kmax)233 void Transform<6, 1, true, VLType::None>(
234 double *out, const double *in, int stride, int x0, int xmax, int k0, int kmax)
235 {
236 a64_transpose_interleave_48(
237 reinterpret_cast<uint16_t *>(out),
238 reinterpret_cast<const uint16_t *>(in + k0 * stride + x0),
239 (xmax-x0) * sizeof(double) / 2,
240 stride * sizeof(double),
241 (kmax-k0)
242 );
243 }
244
245 #endif
246