1 /*
2  * Copyright (c) 2022 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 #if defined(__ARM_FEATURE_SVE)
26 
27 template <>
interleave_block(__fp16 * & out,const __fp16 * const * in,size_t width,size_t height,size_t row_offset,bool first)28 void interleave_block<1, 1, VLType::SME, false>(
29   __fp16 * &out, const __fp16 * const *in,
30   size_t width, size_t height, size_t row_offset, bool first
31 )
32 {
33   ARM_COMPUTE_UNUSED(first);
34 
35   __asm__ __volatile__(
36       ".inst 0xd503477f  // SMSTART ZA\n"
37       "mov x20, %x[width]\n"
38       "inch x20\n"
39       "cnth x10\n"
40       "sub x20, x20, #0x1\n"
41       "udiv x20, x20, x10\n"  // n_passes = ceildiv(width, VL<T>)
42       "mov x19, %x[width]\n"
43       "sub x9, x10, #0x1\n"
44       "sub x28, x20, #0x1\n"
45       "ands x9, x19, x9\n"
46       "sub x27, x10, #0x2\n"
47       "lsl x19, %x[height], #0x1\n"  // height * 2
48       "mov x26, #0x0\n"
49       "mov x25, %x[in]\n"
50       "lsr x28, x28, #0x1\n"  // n_loops = (n_passes - 1) / 2
51       "ldr x24, [x25, #0x0]\n"
52       "and x23, x20, #0x1\n"  // odd_tail = bool(n_passes & 0x1)
53       "csel x9, x9, x10, NE\n"
54       "ldr x22, [x25, #0x8]\n"
55       "ptrue p11.h\n"
56       "whilelt p10.h, XZR, x19\n"
57       "mov x21, %x[row_offset]\n"
58       "mov x20, %x[out]\n"
59       "whilelt p9.h, x26, %x[width]\n"
60       "whilelt p8.h, x26, %x[width]\n"
61       "add x25, x25, #0x10\n"
62       "mov x12, #0x0\n"
63       "cbz x27, 2f\n"
64       "1:"  // K loop: Charge: Loop
65       ".inst 0x25286140  // psel p0.h, p8.h/Z, p10.h[w12]\n"
66       ".inst 0xe0550300  // ld1h { za0h.h[x12] }, p0/Z, [x24, x21, LSL #1]\n"
67       ".inst 0x25386140  // psel p0.h, p8.h/Z, p10.h[w12, #1]\n"
68       "ldr x24, [x25, #0x0]\n"
69       ".inst 0xe05502c1  // ld1h { za0h.h[x12, #1] }, p0/Z, [x22, x21, LSL #1]\n"
70       "add x12, x12, #0x2\n"
71       "cmp x12, x27\n"
72       "ldr x22, [x25, #0x8]\n"
73       "add x25, x25, #0x10\n"
74       "blt 1b\n"
75       "2:"  // K loop: Charge: End
76       ".inst 0x25286140  // psel p0.h, p8.h/Z, p10.h[w12]\n"
77       ".inst 0xe0550300  // ld1h { za0h.h[x12] }, p0/Z, [x24, x21, LSL #1]\n"
78       ".inst 0x25386140  // psel p0.h, p8.h/Z, p10.h[w12, #1]\n"
79       "mov x25, %x[in]\n"
80       ".inst 0xe05502c1  // ld1h { za0h.h[x12, #1] }, p0/Z, [x22, x21, LSL #1]\n"
81       "ldr x24, [x25, #0x0]\n"
82       "inch x21\n"
83       "ldr x22, [x25, #0x8]\n"
84       "add x25, x25, #0x10\n"
85       "inch x26\n"
86       "cbz x28, 8f\n"
87       "mov x19, x28\n"
88       "3:"  // K loop: Main loop
89       "whilelt p8.h, x26, %x[width]\n"
90       "mov x12, #0x0\n"
91       "cbz x27, 5f\n"
92       "4:"  // K loop: Main loop: First: Loop
93       ".inst 0x25286140  // psel p0.h, p8.h/Z, p10.h[w12]\n"
94       ".inst 0xe0550308  // ld1h { za1h.h[x12] }, p0/Z, [x24, x21, LSL #1]\n"
95       ".inst 0x25386141  // psel p1.h, p8.h/Z, p10.h[w12, #1]\n"
96       "ldr x24, [x25, #0x0]\n"
97       ".inst 0x25286d20  // psel p0.h, p11.h/Z, p9.h[w12]\n"
98       ".inst 0xe05506c9  // ld1h { za1h.h[x12, #1] }, p1/Z, [x22, x21, LSL #1]\n"
99       "ldr x22, [x25, #0x8]\n"
100       ".inst 0xe07f8280  // st1h { za0v.h[x12] }, p0/Z, [x20, XZR, LSL #1]\n"
101       ".inst 0x25386d20  // psel p0.h, p11.h/Z, p9.h[w12, #1]\n"
102       ".inst 0xe06a8281  // st1h { za0v.h[x12, #1] }, p0/Z, [x20, x10, LSL #1]\n"
103       "add x12, x12, #0x2\n"
104       "cmp x12, x27\n"
105       "add x25, x25, #0x10\n"
106       "addvl x20, x20, #2\n"
107       "blt 4b\n"
108       "5:"  // K loop: Main loop: First: Tail
109       ".inst 0x25286140  // psel p0.h, p8.h/Z, p10.h[w12]\n"
110       ".inst 0xe0550308  // ld1h { za1h.h[x12] }, p0/Z, [x24, x21, LSL #1]\n"
111       "mov x25, %x[in]\n"
112       "ldr x24, [x25, #0x0]\n"
113       ".inst 0x25386141  // psel p1.h, p8.h/Z, p10.h[w12, #1]\n"
114       ".inst 0x25286d20  // psel p0.h, p11.h/Z, p9.h[w12]\n"
115       ".inst 0xe05506c9  // ld1h { za1h.h[x12, #1] }, p1/Z, [x22, x21, LSL #1]\n"
116       "ldr x22, [x25, #0x8]\n"
117       ".inst 0xe07f8280  // st1h { za0v.h[x12] }, p0/Z, [x20, XZR, LSL #1]\n"
118       ".inst 0x25386d20  // psel p0.h, p11.h/Z, p9.h[w12, #1]\n"
119       "whilelt p9.h, x26, %x[width]\n"
120       "inch x26\n"
121       "add x25, x25, #0x10\n"
122       ".inst 0xe06a8281  // st1h { za0v.h[x12, #1] }, p0/Z, [x20, x10, LSL #1]\n"
123       "addvl x20, x20, #2\n"
124       "inch x21\n"
125       "whilelt p8.h, x26, %x[width]\n"
126       "mov x12, #0x0\n"
127       "cbz x27, 7f\n"
128       "6:"  // K loop: Main loop: Second: Loop
129       ".inst 0x25286140  // psel p0.h, p8.h/Z, p10.h[w12]\n"
130       ".inst 0xe0550300  // ld1h { za0h.h[x12] }, p0/Z, [x24, x21, LSL #1]\n"
131       ".inst 0x25386141  // psel p1.h, p8.h/Z, p10.h[w12, #1]\n"
132       "ldr x24, [x25, #0x0]\n"
133       ".inst 0x25286d20  // psel p0.h, p11.h/Z, p9.h[w12]\n"
134       ".inst 0xe05506c1  // ld1h { za0h.h[x12, #1] }, p1/Z, [x22, x21, LSL #1]\n"
135       "ldr x22, [x25, #0x8]\n"
136       ".inst 0xe07f8288  // st1h { za1v.h[x12] }, p0/Z, [x20, XZR, LSL #1]\n"
137       ".inst 0x25386d20  // psel p0.h, p11.h/Z, p9.h[w12, #1]\n"
138       ".inst 0xe06a8289  // st1h { za1v.h[x12, #1] }, p0/Z, [x20, x10, LSL #1]\n"
139       "add x12, x12, #0x2\n"
140       "cmp x12, x27\n"
141       "add x25, x25, #0x10\n"
142       "addvl x20, x20, #2\n"
143       "blt 6b\n"
144       "7:"  // K loop: Main loop: Second: Tail
145       ".inst 0x25286140  // psel p0.h, p8.h/Z, p10.h[w12]\n"
146       ".inst 0xe0550300  // ld1h { za0h.h[x12] }, p0/Z, [x24, x21, LSL #1]\n"
147       "mov x25, %x[in]\n"
148       "ldr x24, [x25, #0x0]\n"
149       ".inst 0x25386141  // psel p1.h, p8.h/Z, p10.h[w12, #1]\n"
150       ".inst 0x25286d20  // psel p0.h, p11.h/Z, p9.h[w12]\n"
151       ".inst 0xe05506c1  // ld1h { za0h.h[x12, #1] }, p1/Z, [x22, x21, LSL #1]\n"
152       "ldr x22, [x25, #0x8]\n"
153       ".inst 0xe07f8288  // st1h { za1v.h[x12] }, p0/Z, [x20, XZR, LSL #1]\n"
154       ".inst 0x25386d20  // psel p0.h, p11.h/Z, p9.h[w12, #1]\n"
155       "whilelt p9.h, x26, %x[width]\n"
156       "subs x19, x19, #0x1\n"
157       "add x25, x25, #0x10\n"
158       ".inst 0xe06a8289  // st1h { za1v.h[x12, #1] }, p0/Z, [x20, x10, LSL #1]\n"
159       "addvl x20, x20, #2\n"
160       "inch x26\n"
161       "inch x21\n"
162       "bgt 3b\n"
163       "8:"  // K loop: Tails
164       "cbnz x23, 11f\n"
165       "mov x25, %x[in]\n"
166       "whilelt p8.h, x26, %x[width]\n"
167       "mov x12, #0x0\n"
168       "9:"  // K loop: Tails: Even: First
169       ".inst 0x25286d20  // psel p0.h, p11.h/Z, p9.h[w12]\n"
170       ".inst 0xe07f8280  // st1h { za0v.h[x12] }, p0/Z, [x20, XZR, LSL #1]\n"
171       "ldr x24, [x25, #0x0]\n"
172       ".inst 0x25286140  // psel p0.h, p8.h/Z, p10.h[w12]\n"
173       ".inst 0xe0550308  // ld1h { za1h.h[x12] }, p0/Z, [x24, x21, LSL #1]\n"
174       "add x12, x12, #0x1\n"
175       "cmp x12, x10\n"
176       "add x25, x25, #0x8\n"
177       "addvl x20, x20, #1\n"
178       "blt 9b\n"
179       "whilelt p9.h, x26, %x[width]\n"
180       "whilelt p8.h, x26, %x[width]\n"
181       "mov x12, #0x0\n"
182       "10:"  // K loop: Tails: Even: Second
183       ".inst 0x25286d20  // psel p0.h, p11.h/Z, p9.h[w12]\n"
184       ".inst 0xe07f8288  // st1h { za1v.h[x12] }, p0/Z, [x20, XZR, LSL #1]\n"
185       "add x12, x12, #0x1\n"
186       "cmp x12, x9\n"
187       "addvl x20, x20, #1\n"
188       "blt 10b\n"
189       "whilelt p9.h, x26, %x[width]\n"
190       "b 13f\n"
191       "11:"  // K loop: Tails: Odd
192       "mov x12, #0x0\n"
193       "12:"  // K loop: Tails: Odd: Loop
194       ".inst 0x25286d20  // psel p0.h, p11.h/Z, p9.h[w12]\n"
195       ".inst 0xe07f8280  // st1h { za0v.h[x12] }, p0/Z, [x20, XZR, LSL #1]\n"
196       "add x12, x12, #0x1\n"
197       "cmp x12, x9\n"
198       "addvl x20, x20, #1\n"
199       "blt 12b\n"
200       "13:"  // K loop: End
201       "mov %x[out], x20\n"
202       ".inst 0xd503467f  // SMSTOP\n"
203       : [out] "+&r" (out)
204       : [height] "r" (height), [in] "r" (in), [row_offset] "r" (row_offset), [width] "r" (width)
205       : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "p5", "p6", "p7", "p8", "p9", "p10", "p11", "p12", "p13", "p14", "p15", "x9", "x10", "x12", "x19", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
206     );
207 }
208 
209 #endif  // defined(__ARM_FEATURE_SVE)
210