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 #ifdef __aarch64__
26 
27 template<>
interleave_block(bfloat16 * & out_ptr,const float * const * in,size_t width,size_t height,size_t row_offset,bool)28 void interleave_block<8, 4, VLType::None, false>(
29   bfloat16 * &out_ptr, const float * const * in, size_t width, size_t height,
30   size_t row_offset, bool
31 )
32 {
33   __asm__ __volatile__(
34       "ldr x27, [%x[in], #0x0]\n"
35       "cmp %x[height], #0x8\n"
36       "ldr x26, [%x[in], #0x8]\n"
37       "add x27, x27, %x[row_offset], LSL #2\n"
38       "ldr x25, [%x[in], #0x10]\n"
39       "ldr x24, [%x[in], #0x18]\n"
40       "add x26, x26, %x[row_offset], LSL #2\n"
41       "ldr x23, [%x[in], #0x20]\n"
42       "add x25, x25, %x[row_offset], LSL #2\n"
43       "ldr x22, [%x[in], #0x28]\n"
44       "ldr x21, [%x[in], #0x30]\n"
45       "add x24, x24, %x[row_offset], LSL #2\n"
46       "ldr x20, [%x[in], #0x38]\n"
47       "add x23, x23, %x[row_offset], LSL #2\n"
48       "add x22, x22, %x[row_offset], LSL #2\n"
49       "add x21, x21, %x[row_offset], LSL #2\n"
50       "add x20, x20, %x[row_offset], LSL #2\n"
51       "beq 1f\n"
52       "mov x20, x27\n"
53       "cmp %x[height], #0x2\n"
54       "csel x26, x26, x27, GE\n"
55       "csel x25, x25, x27, GT\n"
56       "cmp %x[height], #0x4\n"
57       "csel x24, x24, x27, GE\n"
58       "csel x23, x23, x27, GT\n"
59       "cmp %x[height], #0x6\n"
60       "csel x22, x22, x27, GE\n"
61       "csel x21, x21, x27, GT\n"
62       "1:"  // no_pointer_adj
63       "prfm pldl1keep, [x27, #0x0]\n"
64       "cmp %x[width], #0x4\n"
65       "prfm pldl1keep, [x26, #0x0]\n"
66       "prfm pldl1keep, [x25, #0x0]\n"
67       "prfm pldl1keep, [x24, #0x0]\n"
68       "prfm pldl1keep, [x23, #0x0]\n"
69       "prfm pldl1keep, [x22, #0x0]\n"
70       "prfm pldl1keep, [x21, #0x0]\n"
71       "prfm pldl1keep, [x20, #0x0]\n"
72       "prfm pldl1keep, [x27, #0x40]\n"
73       "prfm pldl1keep, [x26, #0x40]\n"
74       "prfm pldl1keep, [x25, #0x40]\n"
75       "prfm pldl1keep, [x24, #0x40]\n"
76       "prfm pldl1keep, [x23, #0x40]\n"
77       "prfm pldl1keep, [x22, #0x40]\n"
78       "prfm pldl1keep, [x21, #0x40]\n"
79       "prfm pldl1keep, [x20, #0x40]\n"
80       "blt 3f\n"
81       "2:"  // Main loop head
82       "ldr q23, [x27], #0x10\n"
83       ".inst 0x0ea16af7  // bfcvtn v23.4h, v23.4s\n"
84       "ldr q22, [x26], #0x10\n"
85       "subs %x[width], %x[width], #0x4\n"
86       ".inst 0x4ea16ad7  // bfcvtn2 v23.8h, v22.4s\n"
87       "ldr q21, [x25], #0x10\n"
88       "cmp %x[width], #0x4\n"
89       ".inst 0x0ea16ab5  // bfcvtn v21.4h, v21.4s\n"
90       "ldr q20, [x24], #0x10\n"
91       "ldr q18, [x23], #0x10\n"
92       ".inst 0x4ea16a95  // bfcvtn2 v21.8h, v20.4s\n"
93       "ldr q19, [x22], #0x10\n"
94       ".inst 0x0ea16a52  // bfcvtn v18.4h, v18.4s\n"
95       "ldr q16, [x21], #0x10\n"
96       "ldr q17, [x20], #0x10\n"
97       ".inst 0x4ea16a72  // bfcvtn2 v18.8h, v19.4s\n"
98       "prfm pldl1keep, [x27, #0x70]\n"
99       "prfm pldl1keep, [x26, #0x70]\n"
100       ".inst 0x0ea16a10  // bfcvtn v16.4h, v16.4s\n"
101       "prfm pldl1keep, [x25, #0x70]\n"
102       ".inst 0x4ea16a30  // bfcvtn2 v16.8h, v17.4s\n"
103       "prfm pldl1keep, [x24, #0x70]\n"
104       "prfm pldl1keep, [x23, #0x70]\n"
105       "prfm pldl1keep, [x22, #0x70]\n"
106       "prfm pldl1keep, [x21, #0x70]\n"
107       "prfm pldl1keep, [x20, #0x70]\n"
108       "str q23, [%x[out_ptr], #0x0]\n"
109       "str q21, [%x[out_ptr], #0x10]\n"
110       "str q18, [%x[out_ptr], #0x20]\n"
111       "str q16, [%x[out_ptr], #0x30]\n"
112       "add %x[out_ptr], %x[out_ptr], #0x40\n"
113       "bge 2b\n"
114       "3:"  // Main loop skip
115       "cbz %x[width], 6f\n"
116       "tbz %x[width], #1, 4f\n"
117       "ldr d23, [x27], #0x8\n"
118       "ldr d22, [x26], #0x8\n"
119       "mov x19, #0x1\n"
120       "ldr d21, [x25], #0x8\n"
121       "ldr d20, [x24], #0x8\n"
122       "ldr d18, [x23], #0x8\n"
123       "ldr d19, [x22], #0x8\n"
124       "ldr d16, [x21], #0x8\n"
125       "ldr d17, [x20], #0x8\n"
126       "tbz %x[width], #0, 5f\n"
127       "ld1 { v23.s }[2], [x27]\n"
128       "ld1 { v22.s }[2], [x26]\n"
129       "ld1 { v21.s }[2], [x25]\n"
130       "ld1 { v20.s }[2], [x24]\n"
131       "ld1 { v18.s }[2], [x23]\n"
132       "ld1 { v19.s }[2], [x22]\n"
133       "ld1 { v16.s }[2], [x21]\n"
134       "ld1 { v17.s }[2], [x20]\n"
135       "b 5f\n"
136       "4:"  // odd_loads_1_0
137       "ldr s23, [x27, #0x0]\n"
138       "mov x19, #0x1\n"
139       "ldr s22, [x26, #0x0]\n"
140       "ldr s21, [x25, #0x0]\n"
141       "ldr s20, [x24, #0x0]\n"
142       "ldr s18, [x23, #0x0]\n"
143       "ldr s19, [x22, #0x0]\n"
144       "ldr s16, [x21, #0x0]\n"
145       "ldr s17, [x20, #0x0]\n"
146       "5:"  // Odd load end
147       ".inst 0x0ea16af7  // bfcvtn v23.4h, v23.4s\n"
148       ".inst 0x0ea16ab5  // bfcvtn v21.4h, v21.4s\n"
149       ".inst 0x0ea16a52  // bfcvtn v18.4h, v18.4s\n"
150       ".inst 0x0ea16a10  // bfcvtn v16.4h, v16.4s\n"
151       ".inst 0x4ea16ad7  // bfcvtn2 v23.8h, v22.4s\n"
152       "str q23, [%x[out_ptr], #0x0]\n"
153       ".inst 0x4ea16a95  // bfcvtn2 v21.8h, v20.4s\n"
154       ".inst 0x4ea16a72  // bfcvtn2 v18.8h, v19.4s\n"
155       "str q21, [%x[out_ptr], #0x10]\n"
156       ".inst 0x4ea16a30  // bfcvtn2 v16.8h, v17.4s\n"
157       "str q18, [%x[out_ptr], #0x20]\n"
158       "str q16, [%x[out_ptr], #0x30]\n"
159       "add %x[out_ptr], %x[out_ptr], #0x40\n"
160       "6:"  // Odds skip
161 
162       : [out_ptr] "+&r" (out_ptr), [width] "+&r" (width)
163       : [height] "r" (height), [in] "r" (in), [row_offset] "r" (row_offset)
164       : "cc", "memory", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "x19", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27"
165     );
166 }
167 
168 
169 #endif // __aarch64__
170